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 lustre_packed_msg_size(req->rq_reqmsg) ==
441 req_capsule_fmt_size(req->rq_reqmsg->lm_magic,
442 &RQF_MDS_STATFS_NEW, RCL_CLIENT)) {
443 req_capsule_extend(info->mti_pill, &RQF_MDS_STATFS_NEW);
444 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
447 if (reqbody && reqbody->mbo_valid & OBD_MD_FLAGSTATFS)
448 msf = &mdt->mdt_sum_osfs;
450 msf = &mdt->mdt_osfs;
452 if (msf->msf_age + OBD_STATFS_CACHE_SECONDS <= ktime_get_seconds()) {
453 /** statfs data is too old, get up-to-date one */
454 if (reqbody && reqbody->mbo_valid & OBD_MD_FLAGSTATFS)
455 rc = next->md_ops->mdo_statfs(info->mti_env,
458 rc = dt_statfs(info->mti_env, mdt->mdt_bottom,
462 spin_lock(&mdt->mdt_lock);
463 msf->msf_osfs = *osfs;
464 msf->msf_age = ktime_get_seconds();
465 spin_unlock(&mdt->mdt_lock);
467 /** use cached statfs data */
468 spin_lock(&mdt->mdt_lock);
469 *osfs = msf->msf_osfs;
470 spin_unlock(&mdt->mdt_lock);
473 /* at least try to account for cached pages. its still racy and
474 * might be under-reporting if clients haven't announced their
475 * caches with brw recently */
476 CDEBUG(D_SUPER | D_CACHE, "blocks cached %llu granted %llu"
477 " pending %llu free %llu avail %llu\n",
478 tgd->tgd_tot_dirty, tgd->tgd_tot_granted,
479 tgd->tgd_tot_pending,
480 osfs->os_bfree << tgd->tgd_blockbits,
481 osfs->os_bavail << tgd->tgd_blockbits);
483 osfs->os_bavail -= min_t(u64, osfs->os_bavail,
484 ((tgd->tgd_tot_dirty + tgd->tgd_tot_pending +
485 osfs->os_bsize - 1) >> tgd->tgd_blockbits));
487 tgt_grant_sanity_check(mdt->mdt_lu_dev.ld_obd, __func__);
488 CDEBUG(D_CACHE, "%llu blocks: %llu free, %llu avail; "
489 "%llu objects: %llu free; state %x\n",
490 osfs->os_blocks, osfs->os_bfree, osfs->os_bavail,
491 osfs->os_files, osfs->os_ffree, osfs->os_state);
493 if (!exp_grant_param_supp(tsi->tsi_exp) &&
494 tgd->tgd_blockbits > COMPAT_BSIZE_SHIFT) {
495 /* clients which don't support OBD_CONNECT_GRANT_PARAM
496 * should not see a block size > page size, otherwise
497 * cl_lost_grant goes mad. Therefore, we emulate a 4KB (=2^12)
498 * block size which is the biggest block size known to work
499 * with all client's page size. */
500 osfs->os_blocks <<= tgd->tgd_blockbits - COMPAT_BSIZE_SHIFT;
501 osfs->os_bfree <<= tgd->tgd_blockbits - COMPAT_BSIZE_SHIFT;
502 osfs->os_bavail <<= tgd->tgd_blockbits - COMPAT_BSIZE_SHIFT;
503 osfs->os_bsize = 1 << COMPAT_BSIZE_SHIFT;
506 mdt_counter_incr(req, LPROC_MDT_STATFS);
508 mdt_thread_info_fini(info);
513 * Pack size attributes into the reply.
515 int mdt_pack_size2body(struct mdt_thread_info *info,
516 const struct lu_fid *fid, struct lustre_handle *lh)
519 struct md_attr *ma = &info->mti_attr;
521 bool dom_lock = false;
525 LASSERT(ma->ma_attr.la_valid & LA_MODE);
527 if (!S_ISREG(ma->ma_attr.la_mode) ||
528 !(ma->ma_valid & MA_LOV && ma->ma_lmm != NULL))
531 dom_stripe = mdt_lmm_dom_entry(ma->ma_lmm);
532 /* no DoM stripe, no size in reply */
533 if (dom_stripe == LMM_NO_DOM)
536 if (lustre_handle_is_used(lh)) {
537 struct ldlm_lock *lock;
539 lock = ldlm_handle2lock(lh);
541 dom_lock = ldlm_has_dom(lock);
546 /* no DoM lock, no size in reply */
550 /* Either DoM lock exists or LMM has only DoM stripe then
551 * return size on body. */
552 b = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
554 mdt_dom_object_size(info->mti_env, info->mti_mdt, fid, b, dom_lock);
558 #ifdef CONFIG_LUSTRE_FS_POSIX_ACL
560 * Pack ACL data into the reply. UIDs/GIDs are mapped and filtered by nodemap.
562 * \param info thread info object
563 * \param repbody reply to pack ACLs into
564 * \param o mdt object of file to examine
565 * \param nodemap nodemap of client to reply to
567 * \retval -errno error getting or parsing ACL from disk
569 int mdt_pack_acl2body(struct mdt_thread_info *info, struct mdt_body *repbody,
570 struct mdt_object *o, struct lu_nodemap *nodemap)
572 const struct lu_env *env = info->mti_env;
573 struct md_object *next = mdt_object_child(o);
574 struct lu_buf *buf = &info->mti_buf;
575 struct mdt_device *mdt = info->mti_mdt;
576 struct req_capsule *pill = info->mti_pill;
581 buf->lb_buf = req_capsule_server_get(pill, &RMF_ACL);
582 buf->lb_len = req_capsule_get_size(pill, &RMF_ACL, RCL_SERVER);
583 if (buf->lb_len == 0)
587 rc = mo_xattr_get(env, next, buf, XATTR_NAME_ACL_ACCESS);
589 if (rc == -ENODATA) {
590 repbody->mbo_aclsize = 0;
591 repbody->mbo_valid |= OBD_MD_FLACL;
593 } else if (rc == -EOPNOTSUPP) {
597 exp_connect_large_acl(info->mti_exp) &&
598 buf->lb_buf != info->mti_big_acl) {
599 if (info->mti_big_acl == NULL) {
600 info->mti_big_aclsize =
602 mdt->mdt_max_ea_size,
604 OBD_ALLOC_LARGE(info->mti_big_acl,
605 info->mti_big_aclsize);
606 if (info->mti_big_acl == NULL) {
607 info->mti_big_aclsize = 0;
608 CERROR("%s: unable to grow "
611 PFID(mdt_object_fid(o)));
616 CDEBUG(D_INODE, "%s: grow the "DFID
617 " ACL buffer to size %d\n",
619 PFID(mdt_object_fid(o)),
620 info->mti_big_aclsize);
622 buf->lb_buf = info->mti_big_acl;
623 buf->lb_len = info->mti_big_aclsize;
628 CERROR("%s: unable to read "DFID" ACL: rc = %d\n",
629 mdt_obd_name(mdt), PFID(mdt_object_fid(o)), rc);
638 acl_buflen = req_capsule_get_size(pill, &RMF_ACL, RCL_SERVER);
639 if (acl_buflen >= rc)
642 /* If LOV/LMA EA is small, we can reuse part of their buffer */
643 client = ptlrpc_req_get_repsize(pill->rc_req);
644 server = lustre_packed_msg_size(pill->rc_req->rq_repmsg);
645 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER)) {
646 lmm_buflen = req_capsule_get_size(pill, &RMF_MDT_MD,
648 lmmsize = repbody->mbo_eadatasize;
651 if (client < server - acl_buflen - lmm_buflen + rc + lmmsize) {
652 CDEBUG(D_INODE, "%s: client prepared buffer size %d "
653 "is not big enough with the ACL size %d (%d)\n",
654 mdt_obd_name(mdt), client, rc,
655 server - acl_buflen - lmm_buflen + rc + lmmsize);
656 repbody->mbo_aclsize = 0;
657 repbody->mbo_valid &= ~OBD_MD_FLACL;
662 if (buf->lb_buf == info->mti_big_acl)
663 info->mti_big_acl_used = 1;
665 rc = nodemap_map_acl(nodemap, buf->lb_buf,
666 rc, NODEMAP_FS_TO_CLIENT);
667 /* if all ACLs mapped out, rc is still >= 0 */
669 CERROR("%s: nodemap_map_acl unable to parse "DFID
670 " ACL: rc = %d\n", mdt_obd_name(mdt),
671 PFID(mdt_object_fid(o)), rc);
672 repbody->mbo_aclsize = 0;
673 repbody->mbo_valid &= ~OBD_MD_FLACL;
675 repbody->mbo_aclsize = rc;
676 repbody->mbo_valid |= OBD_MD_FLACL;
685 /* XXX Look into layout in MDT layer. */
686 static inline bool mdt_hsm_is_released(struct lov_mds_md *lmm)
688 struct lov_comp_md_v1 *comp_v1;
689 struct lov_mds_md *v1;
692 if (lmm->lmm_magic == LOV_MAGIC_COMP_V1) {
693 comp_v1 = (struct lov_comp_md_v1 *)lmm;
695 for (i = 0; i < comp_v1->lcm_entry_count; i++) {
696 v1 = (struct lov_mds_md *)((char *)comp_v1 +
697 comp_v1->lcm_entries[i].lcme_offset);
698 /* We don't support partial release for now */
699 if (!(v1->lmm_pattern & LOV_PATTERN_F_RELEASED))
704 return (lmm->lmm_pattern & LOV_PATTERN_F_RELEASED) ?
709 void mdt_pack_attr2body(struct mdt_thread_info *info, struct mdt_body *b,
710 const struct lu_attr *attr, const struct lu_fid *fid)
712 struct md_attr *ma = &info->mti_attr;
713 struct obd_export *exp = info->mti_exp;
714 struct lu_nodemap *nodemap = NULL;
716 LASSERT(ma->ma_valid & MA_INODE);
718 if (attr->la_valid & LA_ATIME) {
719 b->mbo_atime = attr->la_atime;
720 b->mbo_valid |= OBD_MD_FLATIME;
722 if (attr->la_valid & LA_MTIME) {
723 b->mbo_mtime = attr->la_mtime;
724 b->mbo_valid |= OBD_MD_FLMTIME;
726 if (attr->la_valid & LA_CTIME) {
727 b->mbo_ctime = attr->la_ctime;
728 b->mbo_valid |= OBD_MD_FLCTIME;
730 if (attr->la_valid & LA_FLAGS) {
731 b->mbo_flags = attr->la_flags;
732 b->mbo_valid |= OBD_MD_FLFLAGS;
734 if (attr->la_valid & LA_NLINK) {
735 b->mbo_nlink = attr->la_nlink;
736 b->mbo_valid |= OBD_MD_FLNLINK;
738 if (attr->la_valid & (LA_UID|LA_GID)) {
739 nodemap = nodemap_get_from_exp(exp);
743 if (attr->la_valid & LA_UID) {
744 b->mbo_uid = nodemap_map_id(nodemap, NODEMAP_UID,
745 NODEMAP_FS_TO_CLIENT,
747 b->mbo_valid |= OBD_MD_FLUID;
749 if (attr->la_valid & LA_GID) {
750 b->mbo_gid = nodemap_map_id(nodemap, NODEMAP_GID,
751 NODEMAP_FS_TO_CLIENT,
753 b->mbo_valid |= OBD_MD_FLGID;
756 if (attr->la_valid & LA_PROJID) {
757 /* TODO, nodemap for project id */
758 b->mbo_projid = attr->la_projid;
759 b->mbo_valid |= OBD_MD_FLPROJID;
762 b->mbo_mode = attr->la_mode;
763 if (attr->la_valid & LA_MODE)
764 b->mbo_valid |= OBD_MD_FLMODE;
765 if (attr->la_valid & LA_TYPE)
766 b->mbo_valid |= OBD_MD_FLTYPE;
770 b->mbo_valid |= OBD_MD_FLID;
771 CDEBUG(D_INODE, DFID": nlink=%d, mode=%o, valid=%#llx\n",
772 PFID(fid), b->mbo_nlink, b->mbo_mode, b->mbo_valid);
775 if (!(attr->la_valid & LA_TYPE))
778 b->mbo_rdev = attr->la_rdev;
779 b->mbo_size = attr->la_size;
780 b->mbo_blocks = attr->la_blocks;
782 if (!S_ISREG(attr->la_mode)) {
783 b->mbo_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS | OBD_MD_FLRDEV;
784 } else if (ma->ma_need & MA_LOV && !(ma->ma_valid & MA_LOV)) {
785 /* means no objects are allocated on osts. */
786 LASSERT(!(ma->ma_valid & MA_LOV));
787 /* just ignore blocks occupied by extend attributes on MDS */
789 /* if no object is allocated on osts, the size on mds is valid.
791 b->mbo_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
792 } else if ((ma->ma_valid & MA_LOV) && ma->ma_lmm != NULL) {
793 if (mdt_hsm_is_released(ma->ma_lmm)) {
794 /* A released file stores its size on MDS. */
795 /* But return 1 block for released file, unless tools
796 * like tar will consider it fully sparse. (LU-3864)
798 if (unlikely(b->mbo_size == 0))
802 b->mbo_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
803 } else if (info->mti_som_valid) { /* som is valid */
804 b->mbo_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
805 } else if (ma->ma_valid & MA_SOM) { /* lsom is valid */
806 b->mbo_valid |= OBD_MD_FLLAZYSIZE | OBD_MD_FLLAZYBLOCKS;
807 b->mbo_size = ma->ma_som.ms_size;
808 b->mbo_blocks = ma->ma_som.ms_blocks;
812 if (fid != NULL && (b->mbo_valid & OBD_MD_FLSIZE ||
813 b->mbo_valid & OBD_MD_FLLAZYSIZE))
814 CDEBUG(D_VFSTRACE, DFID": returning size %llu\n",
815 PFID(fid), (unsigned long long)b->mbo_size);
818 if (!IS_ERR_OR_NULL(nodemap))
819 nodemap_putref(nodemap);
822 static inline int mdt_body_has_lov(const struct lu_attr *la,
823 const struct mdt_body *body)
825 return (S_ISREG(la->la_mode) && (body->mbo_valid & OBD_MD_FLEASIZE)) ||
826 (S_ISDIR(la->la_mode) && (body->mbo_valid & OBD_MD_FLDIREA));
829 void mdt_client_compatibility(struct mdt_thread_info *info)
831 struct mdt_body *body;
832 struct ptlrpc_request *req = mdt_info_req(info);
833 struct obd_export *exp = req->rq_export;
834 struct md_attr *ma = &info->mti_attr;
835 struct lu_attr *la = &ma->ma_attr;
838 if (exp_connect_layout(exp))
839 /* the client can deal with 16-bit lmm_stripe_count */
842 body = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
844 if (!mdt_body_has_lov(la, body))
847 /* now we have a reply with a lov for a client not compatible with the
848 * layout lock so we have to clean the layout generation number */
849 if (S_ISREG(la->la_mode))
850 ma->ma_lmm->lmm_layout_gen = 0;
854 static int mdt_attr_get_eabuf_size(struct mdt_thread_info *info,
855 struct mdt_object *o)
857 const struct lu_env *env = info->mti_env;
860 rc = mo_xattr_get(env, mdt_object_child(o), &LU_BUF_NULL,
869 /* Is it a directory? Let's check for the LMV as well */
870 if (S_ISDIR(lu_object_attr(&mdt_object_child(o)->mo_lu))) {
871 rc2 = mo_xattr_get(env, mdt_object_child(o), &LU_BUF_NULL,
875 rc2 = mo_xattr_get(env, mdt_object_child(o),
877 XATTR_NAME_DEFAULT_LMV);
879 if ((rc2 < 0 && rc2 != -ENODATA) || (rc2 > rc))
887 int mdt_big_xattr_get(struct mdt_thread_info *info, struct mdt_object *o,
890 const struct lu_env *env = info->mti_env;
894 LASSERT(info->mti_big_lmm_used == 0);
895 rc = mo_xattr_get(env, mdt_object_child(o), &LU_BUF_NULL, name);
899 /* big_lmm may need to be grown */
900 if (info->mti_big_lmmsize < rc) {
901 int size = size_roundup_power2(rc);
903 if (info->mti_big_lmmsize > 0) {
904 /* free old buffer */
905 LASSERT(info->mti_big_lmm);
906 OBD_FREE_LARGE(info->mti_big_lmm,
907 info->mti_big_lmmsize);
908 info->mti_big_lmm = NULL;
909 info->mti_big_lmmsize = 0;
912 OBD_ALLOC_LARGE(info->mti_big_lmm, size);
913 if (info->mti_big_lmm == NULL)
915 info->mti_big_lmmsize = size;
917 LASSERT(info->mti_big_lmmsize >= rc);
919 info->mti_buf.lb_buf = info->mti_big_lmm;
920 info->mti_buf.lb_len = info->mti_big_lmmsize;
921 rc = mo_xattr_get(env, mdt_object_child(o), &info->mti_buf, name);
926 int mdt_stripe_get(struct mdt_thread_info *info, struct mdt_object *o,
927 struct md_attr *ma, const char *name)
929 struct md_object *next = mdt_object_child(o);
930 struct lu_buf *buf = &info->mti_buf;
933 if (strcmp(name, XATTR_NAME_LOV) == 0) {
934 buf->lb_buf = ma->ma_lmm;
935 buf->lb_len = ma->ma_lmm_size;
936 LASSERT(!(ma->ma_valid & MA_LOV));
937 } else if (strcmp(name, XATTR_NAME_LMV) == 0) {
938 buf->lb_buf = ma->ma_lmv;
939 buf->lb_len = ma->ma_lmv_size;
940 LASSERT(!(ma->ma_valid & MA_LMV));
941 } else if (strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0) {
942 buf->lb_buf = ma->ma_default_lmv;
943 buf->lb_len = ma->ma_default_lmv_size;
944 LASSERT(!(ma->ma_valid & MA_LMV_DEF));
949 LASSERT(buf->lb_buf);
951 rc = mo_xattr_get(info->mti_env, next, buf, name);
955 if (strcmp(name, XATTR_NAME_LOV) == 0) {
956 if (info->mti_big_lmm_used)
957 ma->ma_lmm = info->mti_big_lmm;
959 /* NOT return LOV EA with hole to old client. */
960 if (unlikely(le32_to_cpu(ma->ma_lmm->lmm_pattern) &
961 LOV_PATTERN_F_HOLE) &&
962 !(exp_connect_flags(info->mti_exp) &
963 OBD_CONNECT_LFSCK)) {
966 ma->ma_lmm_size = rc;
967 ma->ma_valid |= MA_LOV;
969 } else if (strcmp(name, XATTR_NAME_LMV) == 0) {
970 if (info->mti_big_lmm_used)
971 ma->ma_lmv = info->mti_big_lmm;
973 ma->ma_lmv_size = rc;
974 ma->ma_valid |= MA_LMV;
975 } else if (strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0) {
976 ma->ma_default_lmv_size = rc;
977 ma->ma_valid |= MA_LMV_DEF;
980 /* Update mdt_max_mdsize so all clients will be aware that */
981 if (info->mti_mdt->mdt_max_mdsize < rc)
982 info->mti_mdt->mdt_max_mdsize = rc;
985 } else if (rc == -ENODATA) {
988 } else if (rc == -ERANGE) {
989 /* Default LMV has fixed size, so it must be able to fit
990 * in the original buffer */
991 if (strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0)
993 rc = mdt_big_xattr_get(info, o, name);
995 info->mti_big_lmm_used = 1;
1003 int mdt_attr_get_pfid(struct mdt_thread_info *info, struct mdt_object *o,
1004 struct lu_fid *pfid)
1006 struct lu_buf *buf = &info->mti_buf;
1007 struct link_ea_header *leh;
1008 struct link_ea_entry *lee;
1012 buf->lb_buf = info->mti_big_lmm;
1013 buf->lb_len = info->mti_big_lmmsize;
1014 rc = mo_xattr_get(info->mti_env, mdt_object_child(o),
1015 buf, XATTR_NAME_LINK);
1016 /* ignore errors, MA_PFID won't be set and it is
1017 * up to the caller to treat this as an error */
1018 if (rc == -ERANGE || buf->lb_len == 0) {
1019 rc = mdt_big_xattr_get(info, o, XATTR_NAME_LINK);
1020 buf->lb_buf = info->mti_big_lmm;
1021 buf->lb_len = info->mti_big_lmmsize;
1026 if (rc < sizeof(*leh)) {
1027 CERROR("short LinkEA on "DFID": rc = %d\n",
1028 PFID(mdt_object_fid(o)), rc);
1032 leh = (struct link_ea_header *) buf->lb_buf;
1033 lee = (struct link_ea_entry *)(leh + 1);
1034 if (leh->leh_magic == __swab32(LINK_EA_MAGIC)) {
1035 leh->leh_magic = LINK_EA_MAGIC;
1036 leh->leh_reccount = __swab32(leh->leh_reccount);
1037 leh->leh_len = __swab64(leh->leh_len);
1039 if (leh->leh_magic != LINK_EA_MAGIC)
1041 if (leh->leh_reccount == 0)
1044 memcpy(pfid, &lee->lee_parent_fid, sizeof(*pfid));
1045 fid_be_to_cpu(pfid, pfid);
1050 int mdt_attr_get_complex(struct mdt_thread_info *info,
1051 struct mdt_object *o, struct md_attr *ma)
1053 const struct lu_env *env = info->mti_env;
1054 struct md_object *next = mdt_object_child(o);
1055 struct lu_buf *buf = &info->mti_buf;
1056 int need = ma->ma_need;
1063 if (mdt_object_exists(o) == 0)
1064 GOTO(out, rc = -ENOENT);
1065 mode = lu_object_attr(&next->mo_lu);
1067 if (need & MA_INODE) {
1068 ma->ma_need = MA_INODE;
1069 rc = mo_attr_get(env, next, ma);
1074 (void) mdt_get_som(info, o, ma);
1075 ma->ma_valid |= MA_INODE;
1078 if (need & MA_PFID) {
1079 rc = mdt_attr_get_pfid(info, o, &ma->ma_pfid);
1081 ma->ma_valid |= MA_PFID;
1082 /* ignore this error, parent fid is not mandatory */
1086 if (need & MA_LOV && (S_ISREG(mode) || S_ISDIR(mode))) {
1087 rc = mdt_stripe_get(info, o, ma, XATTR_NAME_LOV);
1092 if (need & MA_LMV && S_ISDIR(mode)) {
1093 rc = mdt_stripe_get(info, o, ma, XATTR_NAME_LMV);
1098 if (need & MA_LMV_DEF && S_ISDIR(mode)) {
1099 rc = mdt_stripe_get(info, o, ma, XATTR_NAME_DEFAULT_LMV);
1105 * In the handle of MA_INODE, we may already get the SOM attr.
1107 if (need & MA_SOM && S_ISREG(mode) && !(ma->ma_valid & MA_SOM)) {
1108 rc = mdt_get_som(info, o, ma);
1113 if (need & MA_HSM && S_ISREG(mode)) {
1114 buf->lb_buf = info->mti_xattr_buf;
1115 buf->lb_len = sizeof(info->mti_xattr_buf);
1116 BUILD_BUG_ON(sizeof(struct hsm_attrs) >
1117 sizeof(info->mti_xattr_buf));
1118 rc2 = mo_xattr_get(info->mti_env, next, buf, XATTR_NAME_HSM);
1119 rc2 = lustre_buf2hsm(info->mti_xattr_buf, rc2, &ma->ma_hsm);
1121 ma->ma_valid |= MA_HSM;
1122 else if (rc2 < 0 && rc2 != -ENODATA)
1123 GOTO(out, rc = rc2);
1126 #ifdef CONFIG_LUSTRE_FS_POSIX_ACL
1127 if (need & MA_ACL_DEF && S_ISDIR(mode)) {
1128 buf->lb_buf = ma->ma_acl;
1129 buf->lb_len = ma->ma_acl_size;
1130 rc2 = mo_xattr_get(env, next, buf, XATTR_NAME_ACL_DEFAULT);
1132 ma->ma_acl_size = rc2;
1133 ma->ma_valid |= MA_ACL_DEF;
1134 } else if (rc2 == -ENODATA) {
1136 ma->ma_acl_size = 0;
1138 GOTO(out, rc = rc2);
1143 CDEBUG(D_INODE, "after getattr rc = %d, ma_valid = %#llx ma_lmm=%p\n",
1144 rc, ma->ma_valid, ma->ma_lmm);
1148 static int mdt_getattr_internal(struct mdt_thread_info *info,
1149 struct mdt_object *o, int ma_need)
1151 struct md_object *next = mdt_object_child(o);
1152 const struct mdt_body *reqbody = info->mti_body;
1153 struct ptlrpc_request *req = mdt_info_req(info);
1154 struct md_attr *ma = &info->mti_attr;
1155 struct lu_attr *la = &ma->ma_attr;
1156 struct req_capsule *pill = info->mti_pill;
1157 const struct lu_env *env = info->mti_env;
1158 struct mdt_body *repbody;
1159 struct lu_buf *buffer = &info->mti_buf;
1160 struct obd_export *exp = info->mti_exp;
1164 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK))
1165 RETURN(err_serious(-ENOMEM));
1167 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
1171 if (mdt_object_remote(o)) {
1172 /* This object is located on remote node.*/
1173 /* Return -ENOTSUPP for old client */
1174 if (!mdt_is_dne_client(req->rq_export))
1175 GOTO(out, rc = -ENOTSUPP);
1177 repbody->mbo_fid1 = *mdt_object_fid(o);
1178 repbody->mbo_valid = OBD_MD_FLID | OBD_MD_MDS;
1182 if (reqbody->mbo_eadatasize > 0) {
1183 buffer->lb_buf = req_capsule_server_get(pill, &RMF_MDT_MD);
1184 if (buffer->lb_buf == NULL)
1185 GOTO(out, rc = -EPROTO);
1186 buffer->lb_len = req_capsule_get_size(pill, &RMF_MDT_MD,
1189 buffer->lb_buf = NULL;
1191 ma_need &= ~(MA_LOV | MA_LMV);
1192 CDEBUG(D_INFO, "%s: RPC from %s: does not need LOVEA.\n",
1193 mdt_obd_name(info->mti_mdt),
1194 req->rq_export->exp_client_uuid.uuid);
1197 /* from 2.12.58 intent_getattr pack default LMV in reply */
1198 if (S_ISDIR(lu_object_attr(&next->mo_lu)) &&
1199 ((reqbody->mbo_valid & (OBD_MD_MEA | OBD_MD_DEFAULT_MEA)) ==
1200 (OBD_MD_MEA | OBD_MD_DEFAULT_MEA)) &&
1201 req_capsule_has_field(&req->rq_pill, &RMF_DEFAULT_MDT_MD,
1203 ma->ma_lmv = buffer->lb_buf;
1204 ma->ma_lmv_size = buffer->lb_len;
1205 ma->ma_default_lmv = req_capsule_server_get(pill,
1206 &RMF_DEFAULT_MDT_MD);
1207 ma->ma_default_lmv_size = req_capsule_get_size(pill,
1208 &RMF_DEFAULT_MDT_MD,
1210 ma->ma_need = MA_INODE;
1211 if (ma->ma_lmv_size > 0)
1212 ma->ma_need |= MA_LMV;
1213 if (ma->ma_default_lmv_size > 0)
1214 ma->ma_need |= MA_LMV_DEF;
1215 } else if (S_ISDIR(lu_object_attr(&next->mo_lu)) &&
1216 (reqbody->mbo_valid & (OBD_MD_MEA | OBD_MD_DEFAULT_MEA))) {
1217 /* If it is dir and client require MEA, then we got MEA */
1218 /* Assumption: MDT_MD size is enough for lmv size. */
1219 ma->ma_lmv = buffer->lb_buf;
1220 ma->ma_lmv_size = buffer->lb_len;
1221 ma->ma_need = MA_INODE;
1222 if (ma->ma_lmv_size > 0) {
1223 if (reqbody->mbo_valid & OBD_MD_MEA) {
1224 ma->ma_need |= MA_LMV;
1225 } else if (reqbody->mbo_valid & OBD_MD_DEFAULT_MEA) {
1226 ma->ma_need |= MA_LMV_DEF;
1227 ma->ma_default_lmv = buffer->lb_buf;
1229 ma->ma_default_lmv_size = buffer->lb_len;
1230 ma->ma_lmv_size = 0;
1234 ma->ma_lmm = buffer->lb_buf;
1235 ma->ma_lmm_size = buffer->lb_len;
1236 ma->ma_need = MA_INODE | MA_HSM;
1237 if (ma->ma_lmm_size > 0) {
1238 ma->ma_need |= MA_LOV;
1239 /* Older clients may crash if they getattr overstriped
1242 if (!exp_connect_overstriping(exp) &&
1243 mdt_lmm_is_overstriping(ma->ma_lmm))
1244 RETURN(-EOPNOTSUPP);
1248 if (S_ISDIR(lu_object_attr(&next->mo_lu)) &&
1249 reqbody->mbo_valid & OBD_MD_FLDIREA &&
1250 lustre_msg_get_opc(req->rq_reqmsg) == MDS_GETATTR) {
1251 /* get default stripe info for this dir. */
1252 ma->ma_need |= MA_LOV_DEF;
1254 ma->ma_need |= ma_need;
1256 rc = mdt_attr_get_complex(info, o, ma);
1258 CDEBUG(rc == -ENOENT ? D_OTHER : D_ERROR,
1259 "%s: getattr error for "DFID": rc = %d\n",
1260 mdt_obd_name(info->mti_mdt),
1261 PFID(mdt_object_fid(o)), rc);
1265 /* if file is released, check if a restore is running */
1266 if (ma->ma_valid & MA_HSM) {
1267 repbody->mbo_valid |= OBD_MD_TSTATE;
1268 if ((ma->ma_hsm.mh_flags & HS_RELEASED) &&
1269 mdt_hsm_restore_is_running(info, mdt_object_fid(o)))
1270 repbody->mbo_t_state = MS_RESTORE;
1273 if (likely(ma->ma_valid & MA_INODE))
1274 mdt_pack_attr2body(info, repbody, la, mdt_object_fid(o));
1278 if (mdt_body_has_lov(la, reqbody)) {
1279 if (ma->ma_valid & MA_LOV) {
1280 LASSERT(ma->ma_lmm_size);
1281 repbody->mbo_eadatasize = ma->ma_lmm_size;
1282 if (S_ISDIR(la->la_mode))
1283 repbody->mbo_valid |= OBD_MD_FLDIREA;
1285 repbody->mbo_valid |= OBD_MD_FLEASIZE;
1286 mdt_dump_lmm(D_INFO, ma->ma_lmm, repbody->mbo_valid);
1288 if (ma->ma_valid & MA_LMV) {
1289 /* Return -ENOTSUPP for old client */
1290 if (!mdt_is_striped_client(req->rq_export))
1293 LASSERT(S_ISDIR(la->la_mode));
1294 mdt_dump_lmv(D_INFO, ma->ma_lmv);
1295 repbody->mbo_eadatasize = ma->ma_lmv_size;
1296 repbody->mbo_valid |= (OBD_MD_FLDIREA|OBD_MD_MEA);
1298 if (ma->ma_valid & MA_LMV_DEF) {
1299 /* Return -ENOTSUPP for old client */
1300 if (!mdt_is_striped_client(req->rq_export))
1302 LASSERT(S_ISDIR(la->la_mode));
1304 * when ll_dir_getstripe() gets default LMV, it
1305 * checks mbo_eadatasize.
1307 if (!(ma->ma_valid & MA_LMV))
1308 repbody->mbo_eadatasize =
1309 ma->ma_default_lmv_size;
1310 repbody->mbo_valid |= (OBD_MD_FLDIREA |
1311 OBD_MD_DEFAULT_MEA);
1313 } else if (S_ISLNK(la->la_mode) &&
1314 reqbody->mbo_valid & OBD_MD_LINKNAME) {
1315 buffer->lb_buf = ma->ma_lmm;
1316 /* eadatasize from client includes NULL-terminator, so
1317 * there is no need to read it */
1318 buffer->lb_len = reqbody->mbo_eadatasize - 1;
1319 rc = mo_readlink(env, next, buffer);
1320 if (unlikely(rc <= 0)) {
1321 CERROR("%s: readlink failed for "DFID": rc = %d\n",
1322 mdt_obd_name(info->mti_mdt),
1323 PFID(mdt_object_fid(o)), rc);
1326 int print_limit = min_t(int, PAGE_SIZE - 128, rc);
1328 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_READLINK_EPROTO))
1330 repbody->mbo_valid |= OBD_MD_LINKNAME;
1331 /* we need to report back size with NULL-terminator
1332 * because client expects that */
1333 repbody->mbo_eadatasize = rc + 1;
1334 if (repbody->mbo_eadatasize != reqbody->mbo_eadatasize)
1335 CDEBUG(D_INODE, "%s: Read shorter symlink %d "
1336 "on "DFID ", expected %d\n",
1337 mdt_obd_name(info->mti_mdt),
1338 rc, PFID(mdt_object_fid(o)),
1339 reqbody->mbo_eadatasize - 1);
1340 /* NULL terminate */
1341 ((char *)ma->ma_lmm)[rc] = 0;
1343 /* If the total CDEBUG() size is larger than a page, it
1344 * will print a warning to the console, avoid this by
1345 * printing just the last part of the symlink. */
1346 CDEBUG(D_INODE, "symlink dest %s%.*s, len = %d\n",
1347 print_limit < rc ? "..." : "", print_limit,
1348 (char *)ma->ma_lmm + rc - print_limit, rc);
1353 if (reqbody->mbo_valid & OBD_MD_FLMODEASIZE) {
1354 repbody->mbo_max_mdsize = info->mti_mdt->mdt_max_mdsize;
1355 repbody->mbo_valid |= OBD_MD_FLMODEASIZE;
1356 CDEBUG(D_INODE, "changing the max MD size to %u\n",
1357 repbody->mbo_max_mdsize);
1360 #ifdef CONFIG_LUSTRE_FS_POSIX_ACL
1361 if ((exp_connect_flags(req->rq_export) & OBD_CONNECT_ACL) &&
1362 (reqbody->mbo_valid & OBD_MD_FLACL)) {
1363 struct lu_nodemap *nodemap = nodemap_get_from_exp(exp);
1364 if (IS_ERR(nodemap))
1365 RETURN(PTR_ERR(nodemap));
1367 rc = mdt_pack_acl2body(info, repbody, o, nodemap);
1368 nodemap_putref(nodemap);
1374 mdt_counter_incr(req, LPROC_MDT_GETATTR);
1379 static int mdt_getattr(struct tgt_session_info *tsi)
1381 struct mdt_thread_info *info = tsi2mdt_info(tsi);
1382 struct mdt_object *obj = info->mti_object;
1383 struct req_capsule *pill = info->mti_pill;
1384 struct mdt_body *reqbody;
1385 struct mdt_body *repbody;
1389 if (unlikely(info->mti_object == NULL))
1392 reqbody = req_capsule_client_get(pill, &RMF_MDT_BODY);
1394 LASSERT(lu_object_assert_exists(&obj->mot_obj));
1396 /* Special case for Data-on-MDT files to get data version */
1397 if (unlikely(reqbody->mbo_valid & OBD_MD_FLDATAVERSION)) {
1398 rc = mdt_data_version_get(tsi);
1402 /* Unlike intent case where we need to pre-fill out buffers early on
1403 * in intent policy for ldlm reasons, here we can have a much better
1404 * guess at EA size by just reading it from disk.
1405 * Exceptions are readdir and (missing) directory striping */
1407 if (reqbody->mbo_valid & OBD_MD_LINKNAME) {
1408 /* No easy way to know how long is the symlink, but it cannot
1409 * be more than PATH_MAX, so we allocate +1 */
1411 /* A special case for fs ROOT: getattr there might fetch
1412 * default EA for entire fs, not just for this dir!
1414 } else if (lu_fid_eq(mdt_object_fid(obj),
1415 &info->mti_mdt->mdt_md_root_fid) &&
1416 (reqbody->mbo_valid & OBD_MD_FLDIREA) &&
1417 (lustre_msg_get_opc(mdt_info_req(info)->rq_reqmsg) ==
1419 /* Should the default strping be bigger, mdt_fix_reply
1420 * will reallocate */
1421 rc = DEF_REP_MD_SIZE;
1423 /* Read the actual EA size from disk */
1424 rc = mdt_attr_get_eabuf_size(info, obj);
1428 GOTO(out, rc = err_serious(rc));
1430 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER, rc);
1432 /* Set ACL reply buffer size as LUSTRE_POSIX_ACL_MAX_SIZE_OLD
1433 * by default. If the target object has more ACL entries, then
1434 * enlarge the buffer when necessary. */
1435 req_capsule_set_size(pill, &RMF_ACL, RCL_SERVER,
1436 LUSTRE_POSIX_ACL_MAX_SIZE_OLD);
1438 rc = req_capsule_server_pack(pill);
1439 if (unlikely(rc != 0))
1440 GOTO(out, rc = err_serious(rc));
1442 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
1443 LASSERT(repbody != NULL);
1444 repbody->mbo_eadatasize = 0;
1445 repbody->mbo_aclsize = 0;
1447 rc = mdt_check_ucred(info);
1449 GOTO(out_shrink, rc);
1451 info->mti_cross_ref = !!(reqbody->mbo_valid & OBD_MD_FLCROSSREF);
1453 rc = mdt_getattr_internal(info, obj, 0);
1456 mdt_client_compatibility(info);
1457 rc2 = mdt_fix_reply(info);
1461 mdt_thread_info_fini(info);
1466 * Handler of layout intent RPC requiring the layout modification
1468 * \param[in] info thread environment
1469 * \param[in] obj object
1470 * \param[out] lhc object ldlm lock handle
1471 * \param[in] layout layout change descriptor
1473 * \retval 0 on success
1474 * \retval < 0 error code
1476 int mdt_layout_change(struct mdt_thread_info *info, struct mdt_object *obj,
1477 struct mdt_lock_handle *lhc,
1478 struct md_layout_change *layout)
1484 if (!mdt_object_exists(obj))
1487 if (!S_ISREG(lu_object_attr(&obj->mot_obj)))
1490 rc = mo_permission(info->mti_env, NULL, mdt_object_child(obj), NULL,
1495 rc = mdt_check_resent_lock(info, obj, lhc);
1501 mdt_lock_handle_init(lhc);
1502 mdt_lock_reg_init(lhc, LCK_EX);
1503 rc = mdt_reint_object_lock(info, obj, lhc, MDS_INODELOCK_LAYOUT,
1509 mutex_lock(&obj->mot_som_mutex);
1510 rc = mo_layout_change(info->mti_env, mdt_object_child(obj), layout);
1511 mutex_unlock(&obj->mot_som_mutex);
1514 mdt_object_unlock(info, obj, lhc, 1);
1520 * Exchange MOF_LOV_CREATED flags between two objects after a
1521 * layout swap. No assumption is made on whether o1 or o2 have
1522 * created objects or not.
1524 * \param[in,out] o1 First swap layout object
1525 * \param[in,out] o2 Second swap layout object
1527 static void mdt_swap_lov_flag(struct mdt_object *o1, struct mdt_object *o2)
1529 unsigned int o1_lov_created = o1->mot_lov_created;
1531 mutex_lock(&o1->mot_lov_mutex);
1532 mutex_lock(&o2->mot_lov_mutex);
1534 o1->mot_lov_created = o2->mot_lov_created;
1535 o2->mot_lov_created = o1_lov_created;
1537 mutex_unlock(&o2->mot_lov_mutex);
1538 mutex_unlock(&o1->mot_lov_mutex);
1541 static int mdt_swap_layouts(struct tgt_session_info *tsi)
1543 struct mdt_thread_info *info;
1544 struct ptlrpc_request *req = tgt_ses_req(tsi);
1545 struct obd_export *exp = req->rq_export;
1546 struct mdt_object *o1, *o2, *o;
1547 struct mdt_lock_handle *lh1, *lh2;
1548 struct mdc_swap_layouts *msl;
1552 /* client does not support layout lock, so layout swaping
1554 * FIXME: there is a problem for old clients which don't support
1555 * layout lock yet. If those clients have already opened the file
1556 * they won't be notified at all so that old layout may still be
1557 * used to do IO. This can be fixed after file release is landed by
1558 * doing exclusive open and taking full EX ibits lock. - Jinshan */
1559 if (!exp_connect_layout(exp))
1560 RETURN(-EOPNOTSUPP);
1562 info = tsi2mdt_info(tsi);
1563 if (unlikely(info->mti_object == NULL))
1566 if (info->mti_dlm_req != NULL)
1567 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1569 o1 = info->mti_object;
1570 o = o2 = mdt_object_find(info->mti_env, info->mti_mdt,
1571 &info->mti_body->mbo_fid2);
1573 GOTO(out, rc = PTR_ERR(o));
1575 if (mdt_object_remote(o) || !mdt_object_exists(o)) /* remote object */
1576 GOTO(put, rc = -ENOENT);
1578 rc = lu_fid_cmp(&info->mti_body->mbo_fid1, &info->mti_body->mbo_fid2);
1579 if (unlikely(rc == 0)) /* same file, you kidding me? no-op. */
1585 /* permission check. Make sure the calling process having permission
1586 * to write both files. */
1587 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o1), NULL,
1592 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o2), NULL,
1597 msl = req_capsule_client_get(info->mti_pill, &RMF_SWAP_LAYOUTS);
1599 GOTO(put, rc = -EPROTO);
1601 lh1 = &info->mti_lh[MDT_LH_NEW];
1602 mdt_lock_reg_init(lh1, LCK_EX);
1603 lh2 = &info->mti_lh[MDT_LH_OLD];
1604 mdt_lock_reg_init(lh2, LCK_EX);
1606 rc = mdt_object_lock(info, o1, lh1, MDS_INODELOCK_LAYOUT |
1607 MDS_INODELOCK_XATTR);
1611 rc = mdt_object_lock(info, o2, lh2, MDS_INODELOCK_LAYOUT |
1612 MDS_INODELOCK_XATTR);
1616 rc = mo_swap_layouts(info->mti_env, mdt_object_child(o1),
1617 mdt_object_child(o2), msl->msl_flags);
1621 mdt_swap_lov_flag(o1, o2);
1624 mdt_object_unlock(info, o2, lh2, rc);
1626 mdt_object_unlock(info, o1, lh1, rc);
1628 mdt_object_put(info->mti_env, o);
1630 mdt_thread_info_fini(info);
1634 static int mdt_raw_lookup(struct mdt_thread_info *info,
1635 struct mdt_object *parent,
1636 const struct lu_name *lname,
1637 struct ldlm_reply *ldlm_rep)
1639 struct lu_fid *child_fid = &info->mti_tmp_fid1;
1643 LASSERT(!info->mti_cross_ref);
1645 /* Only got the fid of this obj by name */
1646 fid_zero(child_fid);
1647 rc = mdo_lookup(info->mti_env, mdt_object_child(info->mti_object),
1648 lname, child_fid, &info->mti_spec);
1650 struct mdt_body *repbody;
1652 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1653 repbody->mbo_fid1 = *child_fid;
1654 repbody->mbo_valid = OBD_MD_FLID;
1655 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1656 } else if (rc == -ENOENT) {
1657 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1664 * UPDATE lock should be taken against parent, and be released before exit;
1665 * child_bits lock should be taken against child, and be returned back:
1666 * (1)normal request should release the child lock;
1667 * (2)intent request will grant the lock to client.
1669 static int mdt_getattr_name_lock(struct mdt_thread_info *info,
1670 struct mdt_lock_handle *lhc,
1672 struct ldlm_reply *ldlm_rep)
1674 struct ptlrpc_request *req = mdt_info_req(info);
1675 struct mdt_body *reqbody = NULL;
1676 struct mdt_object *parent = info->mti_object;
1677 struct mdt_object *child;
1678 struct lu_fid *child_fid = &info->mti_tmp_fid1;
1679 struct lu_name *lname = NULL;
1680 struct mdt_lock_handle *lhp = NULL;
1681 struct ldlm_lock *lock;
1682 struct req_capsule *pill = info->mti_pill;
1690 is_resent = lustre_handle_is_used(&lhc->mlh_reg_lh);
1691 LASSERT(ergo(is_resent,
1692 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT));
1697 if (info->mti_cross_ref) {
1698 /* Only getattr on the child. Parent is on another node. */
1699 mdt_set_disposition(info, ldlm_rep,
1700 DISP_LOOKUP_EXECD | DISP_LOOKUP_POS);
1702 CDEBUG(D_INODE, "partial getattr_name child_fid = "DFID", "
1704 PFID(mdt_object_fid(child)), ldlm_rep);
1706 rc = mdt_check_resent_lock(info, child, lhc);
1709 } else if (rc > 0) {
1710 mdt_lock_handle_init(lhc);
1711 mdt_lock_reg_init(lhc, LCK_PR);
1714 * Object's name entry is on another MDS, it will
1715 * request PERM lock only because LOOKUP lock is owned
1716 * by the MDS where name entry resides.
1718 * TODO: it should try layout lock too. - Jinshan
1720 child_bits &= ~(MDS_INODELOCK_LOOKUP |
1721 MDS_INODELOCK_LAYOUT);
1722 child_bits |= MDS_INODELOCK_PERM;
1724 rc = mdt_object_lock(info, child, lhc, child_bits);
1729 /* Finally, we can get attr for child. */
1730 if (!mdt_object_exists(child)) {
1731 LU_OBJECT_DEBUG(D_INFO, info->mti_env,
1733 "remote object doesn't exist.");
1734 mdt_object_unlock(info, child, lhc, 1);
1738 rc = mdt_getattr_internal(info, child, 0);
1739 if (unlikely(rc != 0)) {
1740 mdt_object_unlock(info, child, lhc, 1);
1744 rc = mdt_pack_secctx_in_reply(info, child);
1746 mdt_object_unlock(info, child, lhc, 1);
1750 lname = &info->mti_name;
1751 mdt_name_unpack(pill, &RMF_NAME, lname, MNF_FIX_ANON);
1753 if (lu_name_is_valid(lname)) {
1754 CDEBUG(D_INODE, "getattr with lock for "DFID"/"DNAME", "
1755 "ldlm_rep = %p\n", PFID(mdt_object_fid(parent)),
1756 PNAME(lname), ldlm_rep);
1758 reqbody = req_capsule_client_get(pill, &RMF_MDT_BODY);
1759 if (unlikely(reqbody == NULL))
1760 RETURN(err_serious(-EPROTO));
1762 *child_fid = reqbody->mbo_fid2;
1764 if (unlikely(!fid_is_sane(child_fid)))
1765 RETURN(err_serious(-EINVAL));
1767 CDEBUG(D_INODE, "getattr with lock for "DFID"/"DFID", "
1769 PFID(mdt_object_fid(parent)),
1770 PFID(&reqbody->mbo_fid2), ldlm_rep);
1773 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_EXECD);
1775 if (unlikely(!mdt_object_exists(parent)) && lu_name_is_valid(lname)) {
1776 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
1778 "Parent doesn't exist!");
1782 if (mdt_object_remote(parent)) {
1783 CERROR("%s: parent "DFID" is on remote target\n",
1784 mdt_obd_name(info->mti_mdt),
1785 PFID(mdt_object_fid(parent)));
1789 if (lu_name_is_valid(lname)) {
1790 /* Always allow to lookup ".." */
1791 if (unlikely(lname->ln_namelen == 2 &&
1792 lname->ln_name[0] == '.' &&
1793 lname->ln_name[1] == '.'))
1794 info->mti_spec.sp_permitted = 1;
1796 if (info->mti_body->mbo_valid == OBD_MD_FLID) {
1797 rc = mdt_raw_lookup(info, parent, lname, ldlm_rep);
1802 /* step 1: lock parent only if parent is a directory */
1803 if (S_ISDIR(lu_object_attr(&parent->mot_obj))) {
1804 lhp = &info->mti_lh[MDT_LH_PARENT];
1805 mdt_lock_pdo_init(lhp, LCK_PR, lname);
1806 rc = mdt_object_lock(info, parent, lhp,
1807 MDS_INODELOCK_UPDATE);
1808 if (unlikely(rc != 0))
1812 /* step 2: lookup child's fid by name */
1813 fid_zero(child_fid);
1814 rc = mdo_lookup(info->mti_env, mdt_object_child(parent), lname,
1815 child_fid, &info->mti_spec);
1817 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1820 GOTO(out_parent, rc);
1823 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1826 *step 3: find the child object by fid & lock it.
1827 * regardless if it is local or remote.
1829 *Note: LU-3240 (commit 762f2114d282a98ebfa4dbbeea9298a8088ad24e)
1830 * set parent dir fid the same as child fid in getattr by fid case
1831 * we should not lu_object_find() the object again, could lead
1832 * to hung if there is a concurrent unlink destroyed the object.
1834 if (lu_fid_eq(mdt_object_fid(parent), child_fid)) {
1835 mdt_object_get(info->mti_env, parent);
1838 child = mdt_object_find(info->mti_env, info->mti_mdt,
1842 if (unlikely(IS_ERR(child)))
1843 GOTO(out_parent, rc = PTR_ERR(child));
1845 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RESEND, obd_timeout * 2);
1846 if (!mdt_object_exists(child)) {
1847 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
1849 "Object doesn't exist!");
1850 GOTO(out_child, rc = -ENOENT);
1853 rc = mdt_check_resent_lock(info, child, lhc);
1855 GOTO(out_child, rc);
1856 } else if (rc > 0) {
1857 mdt_lock_handle_init(lhc);
1858 mdt_lock_reg_init(lhc, LCK_PR);
1860 if (!(child_bits & MDS_INODELOCK_UPDATE) &&
1861 !mdt_object_remote(child)) {
1862 struct md_attr *ma = &info->mti_attr;
1865 ma->ma_need = MA_INODE;
1866 rc = mdt_attr_get_complex(info, child, ma);
1867 if (unlikely(rc != 0))
1868 GOTO(out_child, rc);
1870 /* If the file has not been changed for some time, we
1871 * return not only a LOOKUP lock, but also an UPDATE
1872 * lock and this might save us RPC on later STAT. For
1873 * directories, it also let negative dentry cache start
1874 * working for this dir. */
1875 if (ma->ma_valid & MA_INODE &&
1876 ma->ma_attr.la_valid & LA_CTIME &&
1877 info->mti_mdt->mdt_namespace->ns_ctime_age_limit +
1878 ma->ma_attr.la_ctime < ktime_get_real_seconds())
1879 child_bits |= MDS_INODELOCK_UPDATE;
1882 /* layout lock must be granted in a best-effort way
1883 * for IT operations */
1884 LASSERT(!(child_bits & MDS_INODELOCK_LAYOUT));
1885 if (S_ISREG(lu_object_attr(&child->mot_obj)) &&
1886 !mdt_object_remote(child) && ldlm_rep != NULL) {
1887 if (!OBD_FAIL_CHECK(OBD_FAIL_MDS_NO_LL_GETATTR) &&
1888 exp_connect_layout(info->mti_exp)) {
1889 /* try to grant layout lock for regular file. */
1890 try_bits = MDS_INODELOCK_LAYOUT;
1892 /* Acquire DOM lock in advance for data-on-mdt file */
1893 if (child != parent)
1894 try_bits |= MDS_INODELOCK_DOM;
1897 if (try_bits != 0) {
1898 /* try layout lock, it may fail to be granted due to
1899 * contention at LOOKUP or UPDATE */
1900 rc = mdt_object_lock_try(info, child, lhc, &child_bits,
1902 if (child_bits & MDS_INODELOCK_LAYOUT)
1905 /* Do not enqueue the UPDATE lock from MDT(cross-MDT),
1906 * client will enqueue the lock to the remote MDT */
1907 if (mdt_object_remote(child))
1908 child_bits &= ~MDS_INODELOCK_UPDATE;
1909 rc = mdt_object_lock(info, child, lhc, child_bits);
1911 if (unlikely(rc != 0))
1912 GOTO(out_child, rc);
1915 /* finally, we can get attr for child. */
1916 rc = mdt_getattr_internal(info, child, ma_need);
1917 if (unlikely(rc != 0)) {
1918 mdt_object_unlock(info, child, lhc, 1);
1919 GOTO(out_child, rc);
1922 rc = mdt_pack_secctx_in_reply(info, child);
1924 mdt_object_unlock(info, child, lhc, 1);
1925 GOTO(out_child, rc);
1928 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
1930 /* Debugging code. */
1931 LDLM_DEBUG(lock, "Returning lock to client");
1932 LASSERTF(fid_res_name_eq(mdt_object_fid(child),
1933 &lock->l_resource->lr_name),
1934 "Lock res_id: "DLDLMRES", fid: "DFID"\n",
1935 PLDLMRES(lock->l_resource),
1936 PFID(mdt_object_fid(child)));
1938 if (S_ISREG(lu_object_attr(&child->mot_obj)) &&
1939 !mdt_object_remote(child) && child != parent) {
1940 mdt_object_put(info->mti_env, child);
1941 rc = mdt_pack_size2body(info, child_fid,
1943 if (rc != 0 && child_bits & MDS_INODELOCK_DOM) {
1944 /* DOM lock was taken in advance but this is
1945 * not DoM file. Drop the lock.
1947 lock_res_and_lock(lock);
1948 ldlm_inodebits_drop(lock, MDS_INODELOCK_DOM);
1949 unlock_res_and_lock(lock);
1951 LDLM_LOCK_PUT(lock);
1952 GOTO(out_parent, rc = 0);
1954 LDLM_LOCK_PUT(lock);
1959 mdt_object_put(info->mti_env, child);
1962 mdt_object_unlock(info, parent, lhp, 1);
1966 /* normal handler: should release the child lock */
1967 static int mdt_getattr_name(struct tgt_session_info *tsi)
1969 struct mdt_thread_info *info = tsi2mdt_info(tsi);
1970 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_CHILD];
1971 struct mdt_body *reqbody;
1972 struct mdt_body *repbody;
1976 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
1977 LASSERT(reqbody != NULL);
1978 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1979 LASSERT(repbody != NULL);
1981 info->mti_cross_ref = !!(reqbody->mbo_valid & OBD_MD_FLCROSSREF);
1982 repbody->mbo_eadatasize = 0;
1983 repbody->mbo_aclsize = 0;
1985 rc = mdt_init_ucred_intent_getattr(info, reqbody);
1987 GOTO(out_shrink, rc);
1989 rc = mdt_getattr_name_lock(info, lhc, MDS_INODELOCK_UPDATE, NULL);
1990 if (lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1991 ldlm_lock_decref(&lhc->mlh_reg_lh, lhc->mlh_reg_mode);
1992 lhc->mlh_reg_lh.cookie = 0;
1994 mdt_exit_ucred(info);
1997 mdt_client_compatibility(info);
1998 rc2 = mdt_fix_reply(info);
2001 mdt_thread_info_fini(info);
2005 static int mdt_rmfid_unlink(struct mdt_thread_info *info,
2006 const struct lu_fid *pfid,
2007 const struct lu_name *name,
2008 struct mdt_object *obj, s64 ctime)
2010 struct lu_fid *child_fid = &info->mti_tmp_fid1;
2011 struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
2012 struct mdt_device *mdt = info->mti_mdt;
2013 struct md_attr *ma = &info->mti_attr;
2014 struct mdt_lock_handle *parent_lh;
2015 struct mdt_lock_handle *child_lh;
2016 struct mdt_object *pobj;
2017 bool cos_incompat = false;
2021 pobj = mdt_object_find(info->mti_env, mdt, pfid);
2023 GOTO(out, rc = PTR_ERR(pobj));
2025 parent_lh = &info->mti_lh[MDT_LH_PARENT];
2026 mdt_lock_pdo_init(parent_lh, LCK_PW, name);
2027 rc = mdt_object_lock(info, pobj, parent_lh, MDS_INODELOCK_UPDATE);
2029 GOTO(put_parent, rc);
2031 if (mdt_object_remote(pobj))
2032 cos_incompat = true;
2034 rc = mdo_lookup(info->mti_env, mdt_object_child(pobj),
2035 name, child_fid, &info->mti_spec);
2037 GOTO(unlock_parent, rc);
2039 if (!lu_fid_eq(child_fid, mdt_object_fid(obj)))
2040 GOTO(unlock_parent, rc = -EREMCHG);
2042 child_lh = &info->mti_lh[MDT_LH_CHILD];
2043 mdt_lock_reg_init(child_lh, LCK_EX);
2044 rc = mdt_reint_striped_lock(info, obj, child_lh,
2045 MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE,
2046 einfo, cos_incompat);
2048 GOTO(unlock_parent, rc);
2050 if (atomic_read(&obj->mot_open_count)) {
2051 CDEBUG(D_OTHER, "object "DFID" open, skip\n",
2052 PFID(mdt_object_fid(obj)));
2053 GOTO(unlock_child, rc = -EBUSY);
2057 ma->ma_valid = MA_INODE;
2058 ma->ma_attr.la_valid = LA_CTIME;
2059 ma->ma_attr.la_ctime = ctime;
2061 mutex_lock(&obj->mot_lov_mutex);
2063 rc = mdo_unlink(info->mti_env, mdt_object_child(pobj),
2064 mdt_object_child(obj), name, ma, 0);
2066 mutex_unlock(&obj->mot_lov_mutex);
2069 mdt_reint_striped_unlock(info, obj, child_lh, einfo, 1);
2071 mdt_object_unlock(info, pobj, parent_lh, 1);
2073 mdt_object_put(info->mti_env, pobj);
2078 static int mdt_rmfid_check_permission(struct mdt_thread_info *info,
2079 struct mdt_object *obj)
2081 struct lu_ucred *uc = lu_ucred(info->mti_env);
2082 struct md_attr *ma = &info->mti_attr;
2083 struct lu_attr *la = &ma->ma_attr;
2087 ma->ma_need = MA_INODE;
2088 rc = mo_attr_get(info->mti_env, mdt_object_child(obj), ma);
2092 if (la->la_flags & LUSTRE_IMMUTABLE_FL)
2095 if (md_capable(uc, CFS_CAP_DAC_OVERRIDE))
2097 if (uc->uc_fsuid == la->la_uid) {
2098 if ((la->la_mode & S_IWUSR) == 0)
2100 } else if (uc->uc_fsgid == la->la_gid) {
2101 if ((la->la_mode & S_IWGRP) == 0)
2103 } else if ((la->la_mode & S_IWOTH) == 0) {
2111 static int mdt_rmfid_one(struct mdt_thread_info *info, struct lu_fid *fid,
2114 struct mdt_device *mdt = info->mti_mdt;
2115 struct mdt_object *obj = NULL;
2116 struct linkea_data ldata = { NULL };
2117 struct lu_buf *buf = &info->mti_big_buf;
2118 struct lu_name *name = &info->mti_name;
2119 struct lu_fid *pfid = &info->mti_tmp_fid1;
2120 struct link_ea_header *leh;
2121 struct link_ea_entry *lee;
2122 int reclen, count, rc = 0;
2125 if (!fid_is_sane(fid))
2126 GOTO(out, rc = -EINVAL);
2128 if (!fid_is_namespace_visible(fid))
2129 GOTO(out, rc = -EINVAL);
2131 obj = mdt_object_find(info->mti_env, mdt, fid);
2133 GOTO(out, rc = PTR_ERR(obj));
2135 if (mdt_object_remote(obj))
2136 GOTO(out, rc = -EREMOTE);
2137 if (!mdt_object_exists(obj) || lu_object_is_dying(&obj->mot_header))
2138 GOTO(out, rc = -ENOENT);
2140 rc = mdt_rmfid_check_permission(info, obj);
2145 buf = lu_buf_check_and_alloc(buf, PATH_MAX);
2147 GOTO(out, rc = -ENOMEM);
2150 rc = mdt_links_read(info, obj, &ldata);
2155 lee = (struct link_ea_entry *)(leh + 1);
2156 for (count = 0; count < leh->leh_reccount; count++) {
2157 /* remove every hardlink */
2158 linkea_entry_unpack(lee, &reclen, name, pfid);
2159 lee = (struct link_ea_entry *) ((char *)lee + reclen);
2160 rc = mdt_rmfid_unlink(info, pfid, name, obj, ctime);
2166 if (obj && !IS_ERR(obj))
2167 mdt_object_put(info->mti_env, obj);
2168 if (info->mti_big_buf.lb_buf)
2169 lu_buf_free(&info->mti_big_buf);
2174 static int mdt_rmfid(struct tgt_session_info *tsi)
2176 struct mdt_thread_info *mti = tsi2mdt_info(tsi);
2177 struct mdt_body *reqbody;
2178 struct lu_fid *fids, *rfids;
2184 reqbody = req_capsule_client_get(tsi->tsi_pill, &RMF_MDT_BODY);
2185 if (reqbody == NULL)
2187 bufsize = req_capsule_get_size(tsi->tsi_pill, &RMF_FID_ARRAY,
2189 nr = bufsize / sizeof(struct lu_fid);
2190 if (nr * sizeof(struct lu_fid) != bufsize)
2192 req_capsule_set_size(tsi->tsi_pill, &RMF_RCS,
2193 RCL_SERVER, nr * sizeof(__u32));
2194 req_capsule_set_size(tsi->tsi_pill, &RMF_FID_ARRAY,
2195 RCL_SERVER, nr * sizeof(struct lu_fid));
2196 rc = req_capsule_server_pack(tsi->tsi_pill);
2198 GOTO(out, rc = err_serious(rc));
2199 fids = req_capsule_client_get(tsi->tsi_pill, &RMF_FID_ARRAY);
2202 rcs = req_capsule_server_get(tsi->tsi_pill, &RMF_RCS);
2204 rfids = req_capsule_server_get(tsi->tsi_pill, &RMF_FID_ARRAY);
2207 mdt_init_ucred(mti, reqbody);
2208 for (i = 0; i < nr; i++) {
2210 rcs[i] = mdt_rmfid_one(mti, fids + i, reqbody->mbo_ctime);
2212 mdt_exit_ucred(mti);
2218 static int mdt_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2219 void *karg, void __user *uarg);
2221 static int mdt_set_info(struct tgt_session_info *tsi)
2223 struct ptlrpc_request *req = tgt_ses_req(tsi);
2226 int keylen, vallen, rc = 0;
2230 key = req_capsule_client_get(tsi->tsi_pill, &RMF_SETINFO_KEY);
2232 DEBUG_REQ(D_HA, req, "no set_info key");
2233 RETURN(err_serious(-EFAULT));
2236 keylen = req_capsule_get_size(tsi->tsi_pill, &RMF_SETINFO_KEY,
2239 val = req_capsule_client_get(tsi->tsi_pill, &RMF_SETINFO_VAL);
2241 DEBUG_REQ(D_HA, req, "no set_info val");
2242 RETURN(err_serious(-EFAULT));
2245 vallen = req_capsule_get_size(tsi->tsi_pill, &RMF_SETINFO_VAL,
2248 /* Swab any part of val you need to here */
2249 if (KEY_IS(KEY_READ_ONLY)) {
2250 spin_lock(&req->rq_export->exp_lock);
2252 *exp_connect_flags_ptr(req->rq_export) |=
2255 *exp_connect_flags_ptr(req->rq_export) &=
2256 ~OBD_CONNECT_RDONLY;
2257 spin_unlock(&req->rq_export->exp_lock);
2258 } else if (KEY_IS(KEY_CHANGELOG_CLEAR)) {
2259 struct changelog_setinfo *cs = val;
2261 if (vallen != sizeof(*cs)) {
2262 CERROR("%s: bad changelog_clear setinfo size %d\n",
2263 tgt_name(tsi->tsi_tgt), vallen);
2266 if (ptlrpc_req_need_swab(req)) {
2267 __swab64s(&cs->cs_recno);
2268 __swab32s(&cs->cs_id);
2271 rc = mdt_iocontrol(OBD_IOC_CHANGELOG_CLEAR, req->rq_export,
2273 } else if (KEY_IS(KEY_EVICT_BY_NID)) {
2275 obd_export_evict_by_nid(req->rq_export->exp_obd, val);
2282 static int mdt_readpage(struct tgt_session_info *tsi)
2284 struct mdt_thread_info *info = mdt_th_info(tsi->tsi_env);
2285 struct mdt_object *object = mdt_obj(tsi->tsi_corpus);
2286 struct lu_rdpg *rdpg = &info->mti_u.rdpg.mti_rdpg;
2287 const struct mdt_body *reqbody = tsi->tsi_mdt_body;
2288 struct mdt_body *repbody;
2294 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK))
2295 RETURN(err_serious(-ENOMEM));
2297 repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_MDT_BODY);
2298 if (repbody == NULL || reqbody == NULL)
2299 RETURN(err_serious(-EFAULT));
2302 * prepare @rdpg before calling lower layers and transfer itself. Here
2303 * reqbody->size contains offset of where to start to read and
2304 * reqbody->nlink contains number bytes to read.
2306 rdpg->rp_hash = reqbody->mbo_size;
2307 if (rdpg->rp_hash != reqbody->mbo_size) {
2308 CERROR("Invalid hash: %#llx != %#llx\n",
2309 rdpg->rp_hash, reqbody->mbo_size);
2313 rdpg->rp_attrs = reqbody->mbo_mode;
2314 if (exp_connect_flags(tsi->tsi_exp) & OBD_CONNECT_64BITHASH)
2315 rdpg->rp_attrs |= LUDA_64BITHASH;
2316 rdpg->rp_count = min_t(unsigned int, reqbody->mbo_nlink,
2317 exp_max_brw_size(tsi->tsi_exp));
2318 rdpg->rp_npages = (rdpg->rp_count + PAGE_SIZE - 1) >>
2320 OBD_ALLOC(rdpg->rp_pages, rdpg->rp_npages * sizeof rdpg->rp_pages[0]);
2321 if (rdpg->rp_pages == NULL)
2324 for (i = 0; i < rdpg->rp_npages; ++i) {
2325 rdpg->rp_pages[i] = alloc_page(GFP_NOFS);
2326 if (rdpg->rp_pages[i] == NULL)
2327 GOTO(free_rdpg, rc = -ENOMEM);
2330 /* call lower layers to fill allocated pages with directory data */
2331 rc = mo_readpage(tsi->tsi_env, mdt_object_child(object), rdpg);
2333 GOTO(free_rdpg, rc);
2335 /* send pages to client */
2336 rc = tgt_sendpage(tsi, rdpg, rc);
2341 for (i = 0; i < rdpg->rp_npages; i++)
2342 if (rdpg->rp_pages[i] != NULL)
2343 __free_page(rdpg->rp_pages[i]);
2344 OBD_FREE(rdpg->rp_pages, rdpg->rp_npages * sizeof rdpg->rp_pages[0]);
2346 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE))
2352 static int mdt_fix_attr_ucred(struct mdt_thread_info *info, __u32 op)
2354 struct lu_ucred *uc = mdt_ucred_check(info);
2355 struct lu_attr *attr = &info->mti_attr.ma_attr;
2360 if (op != REINT_SETATTR) {
2361 if ((attr->la_valid & LA_UID) && (attr->la_uid != -1))
2362 attr->la_uid = uc->uc_fsuid;
2363 /* for S_ISGID, inherit gid from his parent, such work will be
2364 * done in cmm/mdd layer, here set all cases as uc->uc_fsgid. */
2365 if ((attr->la_valid & LA_GID) && (attr->la_gid != -1))
2366 attr->la_gid = uc->uc_fsgid;
2372 static inline bool mdt_is_readonly_open(struct mdt_thread_info *info, __u32 op)
2374 return op == REINT_OPEN &&
2375 !(info->mti_spec.sp_cr_flags & (MDS_FMODE_WRITE | MDS_OPEN_CREAT));
2378 static void mdt_preset_secctx_size(struct mdt_thread_info *info)
2380 struct req_capsule *pill = info->mti_pill;
2382 if (req_capsule_has_field(pill, &RMF_FILE_SECCTX,
2384 req_capsule_has_field(pill, &RMF_FILE_SECCTX_NAME,
2386 if (req_capsule_get_size(pill, &RMF_FILE_SECCTX_NAME,
2388 /* pre-set size in server part with max size */
2389 req_capsule_set_size(pill, &RMF_FILE_SECCTX,
2391 info->mti_mdt->mdt_max_ea_size);
2393 req_capsule_set_size(pill, &RMF_FILE_SECCTX,
2400 static int mdt_reint_internal(struct mdt_thread_info *info,
2401 struct mdt_lock_handle *lhc,
2404 struct req_capsule *pill = info->mti_pill;
2405 struct mdt_body *repbody;
2410 rc = mdt_reint_unpack(info, op);
2412 CERROR("Can't unpack reint, rc %d\n", rc);
2413 RETURN(err_serious(rc));
2417 /* check if the file system is set to readonly. O_RDONLY open
2418 * is still allowed even the file system is set to readonly mode */
2419 if (mdt_rdonly(info->mti_exp) && !mdt_is_readonly_open(info, op))
2420 RETURN(err_serious(-EROFS));
2422 /* for replay (no_create) lmm is not needed, client has it already */
2423 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
2424 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER,
2427 /* llog cookies are always 0, the field is kept for compatibility */
2428 if (req_capsule_has_field(pill, &RMF_LOGCOOKIES, RCL_SERVER))
2429 req_capsule_set_size(pill, &RMF_LOGCOOKIES, RCL_SERVER, 0);
2431 /* Set ACL reply buffer size as LUSTRE_POSIX_ACL_MAX_SIZE_OLD
2432 * by default. If the target object has more ACL entries, then
2433 * enlarge the buffer when necessary. */
2434 if (req_capsule_has_field(pill, &RMF_ACL, RCL_SERVER))
2435 req_capsule_set_size(pill, &RMF_ACL, RCL_SERVER,
2436 LUSTRE_POSIX_ACL_MAX_SIZE_OLD);
2438 mdt_preset_secctx_size(info);
2440 rc = req_capsule_server_pack(pill);
2442 CERROR("Can't pack response, rc %d\n", rc);
2443 RETURN(err_serious(rc));
2446 if (req_capsule_has_field(pill, &RMF_MDT_BODY, RCL_SERVER)) {
2447 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
2449 repbody->mbo_eadatasize = 0;
2450 repbody->mbo_aclsize = 0;
2453 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_REINT_DELAY, 10);
2455 /* for replay no cookkie / lmm need, because client have this already */
2456 if (info->mti_spec.no_create)
2457 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
2458 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER, 0);
2460 rc = mdt_init_ucred_reint(info);
2462 GOTO(out_shrink, rc);
2464 rc = mdt_fix_attr_ucred(info, op);
2466 GOTO(out_ucred, rc = err_serious(rc));
2468 rc = mdt_check_resent(info, mdt_reconstruct, lhc);
2470 GOTO(out_ucred, rc);
2471 } else if (rc == 1) {
2472 DEBUG_REQ(D_INODE, mdt_info_req(info), "resent opt");
2473 rc = lustre_msg_get_status(mdt_info_req(info)->rq_repmsg);
2474 GOTO(out_ucred, rc);
2476 rc = mdt_reint_rec(info, lhc);
2479 mdt_exit_ucred(info);
2481 mdt_client_compatibility(info);
2483 rc2 = mdt_fix_reply(info);
2488 * Data-on-MDT optimization - read data along with OPEN and return it
2489 * in reply. Do that only if we have both DOM and LAYOUT locks.
2491 if (rc == 0 && op == REINT_OPEN && !req_is_replay(pill->rc_req) &&
2492 info->mti_attr.ma_lmm != NULL &&
2493 mdt_lmm_dom_entry(info->mti_attr.ma_lmm) == LMM_DOM_ONLY) {
2494 rc = mdt_dom_read_on_open(info, info->mti_mdt,
2501 static long mdt_reint_opcode(struct ptlrpc_request *req,
2502 const struct req_format **fmt)
2504 struct mdt_device *mdt;
2505 struct mdt_rec_reint *rec;
2508 rec = req_capsule_client_get(&req->rq_pill, &RMF_REC_REINT);
2510 opc = rec->rr_opcode;
2511 DEBUG_REQ(D_INODE, req, "reint opt = %ld", opc);
2512 if (opc < REINT_MAX && fmt[opc] != NULL)
2513 req_capsule_extend(&req->rq_pill, fmt[opc]);
2515 mdt = mdt_exp2dev(req->rq_export);
2516 CERROR("%s: Unsupported opcode '%ld' from client '%s':"
2517 " rc = %d\n", req->rq_export->exp_obd->obd_name,
2518 opc, mdt->mdt_ldlm_client->cli_name, -EFAULT);
2519 opc = err_serious(-EFAULT);
2522 opc = err_serious(-EFAULT);
2527 static int mdt_reint(struct tgt_session_info *tsi)
2531 static const struct req_format *reint_fmts[REINT_MAX] = {
2532 [REINT_SETATTR] = &RQF_MDS_REINT_SETATTR,
2533 [REINT_CREATE] = &RQF_MDS_REINT_CREATE,
2534 [REINT_LINK] = &RQF_MDS_REINT_LINK,
2535 [REINT_UNLINK] = &RQF_MDS_REINT_UNLINK,
2536 [REINT_RENAME] = &RQF_MDS_REINT_RENAME,
2537 [REINT_OPEN] = &RQF_MDS_REINT_OPEN,
2538 [REINT_SETXATTR] = &RQF_MDS_REINT_SETXATTR,
2539 [REINT_RMENTRY] = &RQF_MDS_REINT_UNLINK,
2540 [REINT_MIGRATE] = &RQF_MDS_REINT_MIGRATE,
2541 [REINT_RESYNC] = &RQF_MDS_REINT_RESYNC,
2546 opc = mdt_reint_opcode(tgt_ses_req(tsi), reint_fmts);
2548 struct mdt_thread_info *info = tsi2mdt_info(tsi);
2550 * No lock possible here from client to pass it to reint code
2553 rc = mdt_reint_internal(info, NULL, opc);
2554 mdt_thread_info_fini(info);
2559 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_REINT_NET_REP;
2563 /* this should sync the whole device */
2564 int mdt_device_sync(const struct lu_env *env, struct mdt_device *mdt)
2566 struct dt_device *dt = mdt->mdt_bottom;
2570 rc = dt->dd_ops->dt_sync(env, dt);
2574 /* this should sync this object */
2575 static int mdt_object_sync(const struct lu_env *env, struct obd_export *exp,
2576 struct mdt_object *mo)
2582 if (!mdt_object_exists(mo)) {
2583 CWARN("%s: non existing object "DFID": rc = %d\n",
2584 exp->exp_obd->obd_name, PFID(mdt_object_fid(mo)),
2589 if (S_ISREG(lu_object_attr(&mo->mot_obj))) {
2590 struct lu_target *tgt = tgt_ses_info(env)->tsi_tgt;
2591 dt_obj_version_t version;
2593 version = dt_version_get(env, mdt_obj2dt(mo));
2594 if (version > tgt->lut_obd->obd_last_committed)
2595 rc = mo_object_sync(env, mdt_object_child(mo));
2597 rc = mo_object_sync(env, mdt_object_child(mo));
2603 static int mdt_sync(struct tgt_session_info *tsi)
2605 struct ptlrpc_request *req = tgt_ses_req(tsi);
2606 struct req_capsule *pill = tsi->tsi_pill;
2607 struct mdt_body *body;
2612 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SYNC_PACK))
2613 RETURN(err_serious(-ENOMEM));
2615 if (fid_seq(&tsi->tsi_mdt_body->mbo_fid1) == 0) {
2616 rc = mdt_device_sync(tsi->tsi_env, mdt_exp2dev(tsi->tsi_exp));
2618 struct mdt_thread_info *info = tsi2mdt_info(tsi);
2620 if (unlikely(info->mti_object == NULL))
2623 /* sync an object */
2624 rc = mdt_object_sync(tsi->tsi_env, tsi->tsi_exp,
2627 const struct lu_fid *fid;
2628 struct lu_attr *la = &info->mti_attr.ma_attr;
2630 info->mti_attr.ma_need = MA_INODE;
2631 info->mti_attr.ma_valid = 0;
2632 rc = mdt_attr_get_complex(info, info->mti_object,
2635 body = req_capsule_server_get(pill,
2637 fid = mdt_object_fid(info->mti_object);
2638 mdt_pack_attr2body(info, body, la, fid);
2641 mdt_thread_info_fini(info);
2644 mdt_counter_incr(req, LPROC_MDT_SYNC);
2649 static int mdt_data_sync(struct tgt_session_info *tsi)
2651 struct mdt_thread_info *info;
2652 struct mdt_device *mdt = mdt_exp2dev(tsi->tsi_exp);
2653 struct ost_body *body = tsi->tsi_ost_body;
2654 struct ost_body *repbody;
2655 struct mdt_object *mo = NULL;
2661 repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_OST_BODY);
2663 /* if no fid is specified then do nothing,
2664 * device sync is done via MDS_SYNC */
2665 if (fid_is_zero(&tsi->tsi_fid))
2668 mo = mdt_object_find(tsi->tsi_env, mdt, &tsi->tsi_fid);
2670 RETURN(PTR_ERR(mo));
2672 rc = mdt_object_sync(tsi->tsi_env, tsi->tsi_exp, mo);
2676 repbody->oa.o_oi = body->oa.o_oi;
2677 repbody->oa.o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
2679 info = tsi2mdt_info(tsi);
2680 ma = &info->mti_attr;
2681 ma->ma_need = MA_INODE;
2683 rc = mdt_attr_get_complex(info, mo, ma);
2685 obdo_from_la(&repbody->oa, &ma->ma_attr, VALID_FLAGS);
2688 mdt_thread_info_fini(info);
2693 mdt_object_put(tsi->tsi_env, mo);
2698 * Handle quota control requests to consult current usage/limit, but also
2699 * to configure quota enforcement
2701 static int mdt_quotactl(struct tgt_session_info *tsi)
2703 struct obd_export *exp = tsi->tsi_exp;
2704 struct req_capsule *pill = tsi->tsi_pill;
2705 struct obd_quotactl *oqctl, *repoqc;
2707 struct mdt_device *mdt = mdt_exp2dev(exp);
2708 struct lu_device *qmt = mdt->mdt_qmt_dev;
2709 struct lu_nodemap *nodemap;
2712 oqctl = req_capsule_client_get(pill, &RMF_OBD_QUOTACTL);
2714 RETURN(err_serious(-EPROTO));
2716 rc = req_capsule_server_pack(pill);
2718 RETURN(err_serious(rc));
2720 nodemap = nodemap_get_from_exp(exp);
2721 if (IS_ERR(nodemap))
2722 RETURN(PTR_ERR(nodemap));
2724 switch (oqctl->qc_cmd) {
2725 /* master quotactl */
2728 case LUSTRE_Q_SETDEFAULT:
2729 if (!nodemap_can_setquota(nodemap))
2730 GOTO(out_nodemap, rc = -EPERM);
2734 case LUSTRE_Q_GETDEFAULT:
2736 GOTO(out_nodemap, rc = -EOPNOTSUPP);
2737 /* slave quotactl */
2743 CERROR("Unsupported quotactl command: %d\n", oqctl->qc_cmd);
2744 GOTO(out_nodemap, rc = -EFAULT);
2748 switch (oqctl->qc_type) {
2750 id = nodemap_map_id(nodemap, NODEMAP_UID,
2751 NODEMAP_CLIENT_TO_FS, id);
2754 id = nodemap_map_id(nodemap, NODEMAP_GID,
2755 NODEMAP_CLIENT_TO_FS, id);
2758 /* todo: check/map project id */
2762 GOTO(out_nodemap, rc = -EOPNOTSUPP);
2764 repoqc = req_capsule_server_get(pill, &RMF_OBD_QUOTACTL);
2766 GOTO(out_nodemap, rc = err_serious(-EFAULT));
2768 if (oqctl->qc_cmd == Q_SETINFO || oqctl->qc_cmd == Q_SETQUOTA)
2769 barrier_exit(tsi->tsi_tgt->lut_bottom);
2771 if (oqctl->qc_id != id)
2772 swap(oqctl->qc_id, id);
2774 if (oqctl->qc_cmd == Q_SETINFO || oqctl->qc_cmd == Q_SETQUOTA) {
2775 if (unlikely(!barrier_entry(tsi->tsi_tgt->lut_bottom)))
2776 RETURN(-EINPROGRESS);
2779 switch (oqctl->qc_cmd) {
2785 case LUSTRE_Q_SETDEFAULT:
2786 case LUSTRE_Q_GETDEFAULT:
2787 /* forward quotactl request to QMT */
2788 rc = qmt_hdls.qmth_quotactl(tsi->tsi_env, qmt, oqctl);
2793 /* slave quotactl */
2794 rc = lquotactl_slv(tsi->tsi_env, tsi->tsi_tgt->lut_bottom,
2799 CERROR("Unsupported quotactl command: %d\n", oqctl->qc_cmd);
2800 GOTO(out_nodemap, rc = -EFAULT);
2803 if (oqctl->qc_id != id)
2804 swap(oqctl->qc_id, id);
2811 nodemap_putref(nodemap);
2816 /** clone llog ctxt from child (mdd)
2817 * This allows remote llog (replicator) access.
2818 * We can either pass all llog RPCs (eg mdt_llog_create) on to child where the
2819 * context was originally set up, or we can handle them directly.
2820 * I choose the latter, but that means I need any llog
2821 * contexts set up by child to be accessable by the mdt. So we clone the
2822 * context into our context list here.
2824 static int mdt_llog_ctxt_clone(const struct lu_env *env, struct mdt_device *mdt,
2827 struct md_device *next = mdt->mdt_child;
2828 struct llog_ctxt *ctxt;
2831 if (!llog_ctxt_null(mdt2obd_dev(mdt), idx))
2834 rc = next->md_ops->mdo_llog_ctxt_get(env, next, idx, (void **)&ctxt);
2835 if (rc || ctxt == NULL) {
2839 rc = llog_group_set_ctxt(&mdt2obd_dev(mdt)->obd_olg, ctxt, idx);
2841 CERROR("Can't set mdt ctxt %d\n", rc);
2846 static int mdt_llog_ctxt_unclone(const struct lu_env *env,
2847 struct mdt_device *mdt, int idx)
2849 struct llog_ctxt *ctxt;
2851 ctxt = llog_get_context(mdt2obd_dev(mdt), idx);
2854 /* Put once for the get we just did, and once for the clone */
2855 llog_ctxt_put(ctxt);
2856 llog_ctxt_put(ctxt);
2861 * sec context handlers
2863 static int mdt_sec_ctx_handle(struct tgt_session_info *tsi)
2865 CFS_FAIL_TIMEOUT(OBD_FAIL_SEC_CTX_HDL_PAUSE, cfs_fail_val);
2871 * quota request handlers
2873 static int mdt_quota_dqacq(struct tgt_session_info *tsi)
2875 struct mdt_device *mdt = mdt_exp2dev(tsi->tsi_exp);
2876 struct lu_device *qmt = mdt->mdt_qmt_dev;
2881 RETURN(err_serious(-EOPNOTSUPP));
2883 rc = qmt_hdls.qmth_dqacq(tsi->tsi_env, qmt, tgt_ses_req(tsi));
2887 struct mdt_object *mdt_object_new(const struct lu_env *env,
2888 struct mdt_device *d,
2889 const struct lu_fid *f)
2891 struct lu_object_conf conf = { .loc_flags = LOC_F_NEW };
2892 struct lu_object *o;
2893 struct mdt_object *m;
2896 CDEBUG(D_INFO, "Allocate object for "DFID"\n", PFID(f));
2897 o = lu_object_find(env, &d->mdt_lu_dev, f, &conf);
2898 if (unlikely(IS_ERR(o)))
2899 m = (struct mdt_object *)o;
2905 struct mdt_object *mdt_object_find(const struct lu_env *env,
2906 struct mdt_device *d,
2907 const struct lu_fid *f)
2909 struct lu_object *o;
2910 struct mdt_object *m;
2913 CDEBUG(D_INFO, "Find object for "DFID"\n", PFID(f));
2914 o = lu_object_find(env, &d->mdt_lu_dev, f, NULL);
2915 if (unlikely(IS_ERR(o)))
2916 m = (struct mdt_object *)o;
2924 * Asyncronous commit for mdt device.
2926 * Pass asynchonous commit call down the MDS stack.
2928 * \param env environment
2929 * \param mdt the mdt device
2931 static void mdt_device_commit_async(const struct lu_env *env,
2932 struct mdt_device *mdt)
2934 struct dt_device *dt = mdt->mdt_bottom;
2938 rc = dt->dd_ops->dt_commit_async(env, dt);
2939 if (unlikely(rc != 0))
2940 CWARN("%s: async commit start failed: rc = %d\n",
2941 mdt_obd_name(mdt), rc);
2942 atomic_inc(&mdt->mdt_async_commit_count);
2947 * Mark the lock as "synchonous".
2949 * Mark the lock to deffer transaction commit to the unlock time.
2951 * \param lock the lock to mark as "synchonous"
2953 * \see mdt_is_lock_sync
2954 * \see mdt_save_lock
2956 static inline void mdt_set_lock_sync(struct ldlm_lock *lock)
2958 lock->l_ast_data = (void*)1;
2962 * Check whehter the lock "synchonous" or not.
2964 * \param lock the lock to check
2965 * \retval 1 the lock is "synchonous"
2966 * \retval 0 the lock isn't "synchronous"
2968 * \see mdt_set_lock_sync
2969 * \see mdt_save_lock
2971 static inline int mdt_is_lock_sync(struct ldlm_lock *lock)
2973 return lock->l_ast_data != NULL;
2977 * Blocking AST for mdt locks.
2979 * Starts transaction commit if in case of COS lock conflict or
2980 * deffers such a commit to the mdt_save_lock.
2982 * \param lock the lock which blocks a request or cancelling lock
2983 * \param desc unused
2984 * \param data unused
2985 * \param flag indicates whether this cancelling or blocking callback
2987 * \see ldlm_blocking_ast_nocheck
2989 int mdt_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
2990 void *data, int flag)
2992 struct obd_device *obd = ldlm_lock_to_ns(lock)->ns_obd;
2993 struct mdt_device *mdt = mdt_dev(obd->obd_lu_dev);
2994 struct ldlm_cb_set_arg *arg = data;
2995 bool commit_async = false;
2999 if (flag == LDLM_CB_CANCELING)
3002 lock_res_and_lock(lock);
3003 if (lock->l_blocking_ast != mdt_blocking_ast) {
3004 unlock_res_and_lock(lock);
3008 /* A blocking ast may be sent from ldlm_lock_decref_internal
3009 * when the last reference to a local lock was released and
3010 * during blocking event from ldlm_work_bl_ast_lock().
3011 * The 'data' parameter is l_ast_data in the first case and
3012 * callback arguments in the second one. Distinguish them by that.
3014 if (!data || data == lock->l_ast_data || !arg->bl_desc)
3015 goto skip_cos_checks;
3017 if (lock->l_req_mode & (LCK_PW | LCK_EX)) {
3018 if (mdt_cos_is_enabled(mdt)) {
3019 if (!arg->bl_desc->bl_same_client)
3020 mdt_set_lock_sync(lock);
3021 } else if (mdt_slc_is_enabled(mdt) &&
3022 arg->bl_desc->bl_cos_incompat) {
3023 mdt_set_lock_sync(lock);
3025 * we may do extra commit here, but there is a small
3026 * window to miss a commit: lock was unlocked (saved),
3027 * then a conflict lock queued and we come here, but
3028 * REP-ACK not received, so lock was not converted to
3030 * Fortunately this window is quite small, so the
3031 * extra commit should be rare (not to say distributed
3032 * operation is rare too).
3034 commit_async = true;
3036 } else if (lock->l_req_mode == LCK_COS) {
3037 commit_async = true;
3041 rc = ldlm_blocking_ast_nocheck(lock);
3046 rc = lu_env_init(&env, LCT_LOCAL);
3047 if (unlikely(rc != 0))
3048 CWARN("%s: lu_env initialization failed, cannot "
3049 "start asynchronous commit: rc = %d\n",
3052 mdt_device_commit_async(&env, mdt);
3059 * Blocking AST for cross-MDT lock
3061 * Discard lock from uncommitted_slc_locks and cancel it.
3063 * \param lock the lock which blocks a request or cancelling lock
3064 * \param desc unused
3065 * \param data unused
3066 * \param flag indicates whether this cancelling or blocking callback
3067 * \retval 0 on success
3068 * \retval negative number on error
3070 int mdt_remote_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
3071 void *data, int flag)
3077 case LDLM_CB_BLOCKING: {
3078 struct lustre_handle lockh;
3080 ldlm_lock2handle(lock, &lockh);
3081 rc = ldlm_cli_cancel(&lockh,
3082 ldlm_is_atomic_cb(lock) ? 0 : LCF_ASYNC);
3084 CDEBUG(D_INODE, "ldlm_cli_cancel: %d\n", rc);
3089 case LDLM_CB_CANCELING: {
3090 struct obd_device *obd = ldlm_lock_to_ns(lock)->ns_obd;
3091 struct mdt_device *mdt =
3092 mdt_dev(obd->obd_lu_dev->ld_site->ls_top_dev);
3094 LDLM_DEBUG(lock, "Revoke remote lock\n");
3096 /* discard slc lock here so that it can be cleaned anytime,
3097 * especially for cleanup_resource() */
3098 tgt_discard_slc_lock(&mdt->mdt_lut, lock);
3100 /* once we cache lock, l_ast_data is set to mdt_object */
3101 if (lock->l_ast_data != NULL) {
3102 struct mdt_object *mo = lock->l_ast_data;
3105 rc = lu_env_init(&env, LCT_MD_THREAD);
3106 if (unlikely(rc != 0)) {
3107 CWARN("%s: lu_env initialization failed, object %p "DFID" is leaked!: rc = %d\n",
3109 PFID(mdt_object_fid(mo)), rc);
3113 if (lock->l_policy_data.l_inodebits.bits &
3114 (MDS_INODELOCK_XATTR | MDS_INODELOCK_UPDATE)) {
3115 rc = mo_invalidate(&env, mdt_object_child(mo));
3116 mo->mot_cache_attr = 0;
3118 mdt_object_put(&env, mo);
3130 int mdt_check_resent_lock(struct mdt_thread_info *info,
3131 struct mdt_object *mo,
3132 struct mdt_lock_handle *lhc)
3134 /* the lock might already be gotten in ldlm_handle_enqueue() */
3135 if (unlikely(lustre_handle_is_used(&lhc->mlh_reg_lh))) {
3136 struct ptlrpc_request *req = mdt_info_req(info);
3137 struct ldlm_lock *lock;
3139 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
3140 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
3142 /* Lock is pinned by ldlm_handle_enqueue0() as it is
3143 * a resend case, however, it could be already destroyed
3144 * due to client eviction or a raced cancel RPC. */
3145 LDLM_DEBUG_NOLOCK("Invalid lock handle %#llx",
3146 lhc->mlh_reg_lh.cookie);
3150 if (!fid_res_name_eq(mdt_object_fid(mo),
3151 &lock->l_resource->lr_name)) {
3152 CWARN("%s: Although resent, but still not "
3153 "get child lock:"DFID"\n",
3154 info->mti_exp->exp_obd->obd_name,
3155 PFID(mdt_object_fid(mo)));
3156 LDLM_LOCK_PUT(lock);
3159 LDLM_LOCK_PUT(lock);
3165 static void mdt_remote_object_lock_created_cb(struct ldlm_lock *lock)
3167 mdt_object_get(NULL, lock->l_ast_data);
3170 int mdt_remote_object_lock_try(struct mdt_thread_info *mti,
3171 struct mdt_object *o, const struct lu_fid *fid,
3172 struct lustre_handle *lh, enum ldlm_mode mode,
3173 __u64 *ibits, __u64 trybits, bool cache)
3175 struct ldlm_enqueue_info *einfo = &mti->mti_remote_einfo;
3176 union ldlm_policy_data *policy = &mti->mti_policy;
3177 struct ldlm_res_id *res_id = &mti->mti_res_id;
3181 LASSERT(mdt_object_remote(o));
3183 fid_build_reg_res_name(fid, res_id);
3185 memset(einfo, 0, sizeof(*einfo));
3186 einfo->ei_type = LDLM_IBITS;
3187 einfo->ei_mode = mode;
3188 einfo->ei_cb_bl = mdt_remote_blocking_ast;
3189 einfo->ei_cb_cp = ldlm_completion_ast;
3190 einfo->ei_enq_slave = 0;
3191 einfo->ei_res_id = res_id;
3195 * if we cache lock, couple lock with mdt_object, so that object
3196 * can be easily found in lock ASTs.
3198 einfo->ei_cbdata = o;
3199 einfo->ei_cb_created = mdt_remote_object_lock_created_cb;
3202 memset(policy, 0, sizeof(*policy));
3203 policy->l_inodebits.bits = *ibits;
3204 policy->l_inodebits.try_bits = trybits;
3206 rc = mo_object_lock(mti->mti_env, mdt_object_child(o), lh, einfo,
3209 /* Return successfully acquired bits to a caller */
3211 struct ldlm_lock *lock = ldlm_handle2lock(lh);
3214 *ibits = lock->l_policy_data.l_inodebits.bits;
3215 LDLM_LOCK_PUT(lock);
3220 int mdt_remote_object_lock(struct mdt_thread_info *mti, struct mdt_object *o,
3221 const struct lu_fid *fid, struct lustre_handle *lh,
3222 enum ldlm_mode mode, __u64 ibits, bool cache)
3224 return mdt_remote_object_lock_try(mti, o, fid, lh, mode, &ibits, 0,
3228 static int mdt_object_local_lock(struct mdt_thread_info *info,
3229 struct mdt_object *o,
3230 struct mdt_lock_handle *lh, __u64 *ibits,
3231 __u64 trybits, bool cos_incompat)
3233 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
3234 union ldlm_policy_data *policy = &info->mti_policy;
3235 struct ldlm_res_id *res_id = &info->mti_res_id;
3236 __u64 dlmflags = 0, *cookie = NULL;
3240 LASSERT(!lustre_handle_is_used(&lh->mlh_reg_lh));
3241 LASSERT(!lustre_handle_is_used(&lh->mlh_pdo_lh));
3242 LASSERT(lh->mlh_reg_mode != LCK_MINMODE);
3243 LASSERT(lh->mlh_type != MDT_NUL_LOCK);
3246 LASSERT(lh->mlh_reg_mode == LCK_PW ||
3247 lh->mlh_reg_mode == LCK_EX);
3248 dlmflags |= LDLM_FL_COS_INCOMPAT;
3249 } else if (mdt_cos_is_enabled(info->mti_mdt)) {
3250 dlmflags |= LDLM_FL_COS_ENABLED;
3253 /* Only enqueue LOOKUP lock for remote object */
3254 LASSERT(ergo(mdt_object_remote(o), *ibits == MDS_INODELOCK_LOOKUP));
3256 /* Lease lock are granted with LDLM_FL_CANCEL_ON_BLOCK */
3257 if (lh->mlh_type == MDT_REG_LOCK && lh->mlh_reg_mode == LCK_EX &&
3258 *ibits == MDS_INODELOCK_OPEN)
3259 dlmflags |= LDLM_FL_CANCEL_ON_BLOCK;
3261 if (lh->mlh_type == MDT_PDO_LOCK) {
3262 /* check for exists after object is locked */
3263 if (mdt_object_exists(o) == 0) {
3264 /* Non-existent object shouldn't have PDO lock */
3267 /* Non-dir object shouldn't have PDO lock */
3268 if (!S_ISDIR(lu_object_attr(&o->mot_obj)))
3273 fid_build_reg_res_name(mdt_object_fid(o), res_id);
3274 dlmflags |= LDLM_FL_ATOMIC_CB;
3277 cookie = &info->mti_exp->exp_handle.h_cookie;
3280 * Take PDO lock on whole directory and build correct @res_id for lock
3281 * on part of directory.
3283 if (lh->mlh_pdo_hash != 0) {
3284 LASSERT(lh->mlh_type == MDT_PDO_LOCK);
3285 mdt_lock_pdo_mode(info, o, lh);
3286 if (lh->mlh_pdo_mode != LCK_NL) {
3288 * Do not use LDLM_FL_LOCAL_ONLY for parallel lock, it
3289 * is never going to be sent to client and we do not
3290 * want it slowed down due to possible cancels.
3292 policy->l_inodebits.bits =
3293 *ibits & MDS_INODELOCK_UPDATE;
3294 policy->l_inodebits.try_bits =
3295 trybits & MDS_INODELOCK_UPDATE;
3296 /* at least one of them should be set */
3297 LASSERT(policy->l_inodebits.bits |
3298 policy->l_inodebits.try_bits);
3299 rc = mdt_fid_lock(info->mti_env, ns, &lh->mlh_pdo_lh,
3300 lh->mlh_pdo_mode, policy, res_id,
3302 if (unlikely(rc != 0))
3303 GOTO(out_unlock, rc);
3307 * Finish res_id initializing by name hash marking part of
3308 * directory which is taking modification.
3310 res_id->name[LUSTRE_RES_ID_HSH_OFF] = lh->mlh_pdo_hash;
3313 policy->l_inodebits.bits = *ibits;
3314 policy->l_inodebits.try_bits = trybits;
3317 * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
3318 * going to be sent to client. If it is - mdt_intent_policy() path will
3319 * fix it up and turn FL_LOCAL flag off.
3321 rc = mdt_fid_lock(info->mti_env, ns, &lh->mlh_reg_lh, lh->mlh_reg_mode,
3322 policy, res_id, LDLM_FL_LOCAL_ONLY | dlmflags,
3326 mdt_object_unlock(info, o, lh, 1);
3327 else if (unlikely(OBD_FAIL_PRECHECK(OBD_FAIL_MDS_PDO_LOCK)) &&
3328 lh->mlh_pdo_hash != 0 &&
3329 (lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX))
3330 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK, 15);
3332 /* Return successfully acquired bits to a caller */
3334 struct ldlm_lock *lock = ldlm_handle2lock(&lh->mlh_reg_lh);
3337 *ibits = lock->l_policy_data.l_inodebits.bits;
3338 LDLM_LOCK_PUT(lock);
3344 mdt_object_lock_internal(struct mdt_thread_info *info, struct mdt_object *o,
3345 struct mdt_lock_handle *lh, __u64 *ibits,
3346 __u64 trybits, bool cos_incompat)
3348 struct mdt_lock_handle *local_lh = NULL;
3352 if (!mdt_object_remote(o)) {
3353 rc = mdt_object_local_lock(info, o, lh, ibits, trybits,
3358 /* XXX do not support PERM/LAYOUT/XATTR lock for remote object yet */
3359 *ibits &= ~(MDS_INODELOCK_PERM | MDS_INODELOCK_LAYOUT |
3360 MDS_INODELOCK_XATTR);
3362 /* Only enqueue LOOKUP lock for remote object */
3363 if (*ibits & MDS_INODELOCK_LOOKUP) {
3364 __u64 local = MDS_INODELOCK_LOOKUP;
3366 rc = mdt_object_local_lock(info, o, lh, &local, 0,
3374 if ((*ibits | trybits) & MDS_INODELOCK_UPDATE) {
3375 /* Sigh, PDO needs to enqueue 2 locks right now, but
3376 * enqueue RPC can only request 1 lock, to avoid extra
3377 * RPC, so it will instead enqueue EX lock for remote
3378 * object anyway XXX*/
3379 if (lh->mlh_type == MDT_PDO_LOCK &&
3380 lh->mlh_pdo_hash != 0) {
3382 "%s: "DFID" convert PDO lock to EX lock.\n",
3383 mdt_obd_name(info->mti_mdt),
3384 PFID(mdt_object_fid(o)));
3385 lh->mlh_pdo_hash = 0;
3386 lh->mlh_rreg_mode = LCK_EX;
3387 lh->mlh_type = MDT_REG_LOCK;
3390 rc = mdt_remote_object_lock_try(info, o, mdt_object_fid(o),
3393 ibits, trybits, false);
3394 if (rc != ELDLM_OK) {
3395 if (local_lh != NULL)
3396 mdt_object_unlock(info, o, local_lh, rc);
3401 /* other components like LFSCK can use lockless access
3402 * and populate cache, so we better invalidate it */
3403 mo_invalidate(info->mti_env, mdt_object_child(o));
3408 int mdt_object_lock(struct mdt_thread_info *info, struct mdt_object *o,
3409 struct mdt_lock_handle *lh, __u64 ibits)
3411 return mdt_object_lock_internal(info, o, lh, &ibits, 0, false);
3414 int mdt_reint_object_lock(struct mdt_thread_info *info, struct mdt_object *o,
3415 struct mdt_lock_handle *lh, __u64 ibits,
3418 LASSERT(lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX);
3419 return mdt_object_lock_internal(info, o, lh, &ibits, 0,
3423 int mdt_object_lock_try(struct mdt_thread_info *info, struct mdt_object *o,
3424 struct mdt_lock_handle *lh, __u64 *ibits,
3425 __u64 trybits, bool cos_incompat)
3427 bool trylock_only = *ibits == 0;
3430 LASSERT(!(*ibits & trybits));
3431 rc = mdt_object_lock_internal(info, o, lh, ibits, trybits,
3433 if (rc && trylock_only) { /* clear error for try ibits lock only */
3434 LASSERT(*ibits == 0);
3441 * Save a lock within request object.
3443 * Keep the lock referenced until whether client ACK or transaction
3444 * commit happens or release the lock immediately depending on input
3445 * parameters. If COS is ON, a write lock is converted to COS lock
3448 * \param info thead info object
3449 * \param h lock handle
3450 * \param mode lock mode
3451 * \param decref force immediate lock releasing
3453 void mdt_save_lock(struct mdt_thread_info *info, struct lustre_handle *h,
3454 enum ldlm_mode mode, int decref)
3458 if (lustre_handle_is_used(h)) {
3459 if (decref || !info->mti_has_trans ||
3460 !(mode & (LCK_PW | LCK_EX))) {
3461 mdt_fid_unlock(h, mode);
3463 struct mdt_device *mdt = info->mti_mdt;
3464 struct ldlm_lock *lock = ldlm_handle2lock(h);
3465 struct ptlrpc_request *req = mdt_info_req(info);
3466 bool cos = mdt_cos_is_enabled(mdt);
3467 bool convert_lock = !cos && mdt_slc_is_enabled(mdt);
3469 LASSERTF(lock != NULL, "no lock for cookie %#llx\n",
3472 /* there is no request if mdt_object_unlock() is called
3473 * from mdt_export_cleanup()->mdt_add_dirty_flag() */
3474 if (likely(req != NULL)) {
3475 LDLM_DEBUG(lock, "save lock request %p reply "
3476 "state %p transno %lld\n", req,
3477 req->rq_reply_state, req->rq_transno);
3479 ldlm_lock_mode_downgrade(lock, LCK_COS);
3482 if (req->rq_export->exp_disconnected)
3483 mdt_fid_unlock(h, mode);
3485 ptlrpc_save_lock(req, h, mode, cos,
3488 mdt_fid_unlock(h, mode);
3490 if (mdt_is_lock_sync(lock)) {
3491 CDEBUG(D_HA, "found sync-lock,"
3492 " async commit started\n");
3493 mdt_device_commit_async(info->mti_env,
3496 LDLM_LOCK_PUT(lock);
3505 * Save cross-MDT lock in uncommitted_slc_locks
3507 * Keep the lock referenced until transaction commit happens or release the lock
3508 * immediately depending on input parameters.
3510 * \param info thead info object
3511 * \param h lock handle
3512 * \param mode lock mode
3513 * \param decref force immediate lock releasing
3515 static void mdt_save_remote_lock(struct mdt_thread_info *info,
3516 struct mdt_object *o, struct lustre_handle *h,
3517 enum ldlm_mode mode, int decref)
3521 if (lustre_handle_is_used(h)) {
3522 struct ldlm_lock *lock = ldlm_handle2lock(h);
3525 (lock->l_policy_data.l_inodebits.bits &
3526 (MDS_INODELOCK_XATTR | MDS_INODELOCK_UPDATE)))
3527 mo_invalidate(info->mti_env, mdt_object_child(o));
3529 if (decref || !info->mti_has_trans ||
3530 !(mode & (LCK_PW | LCK_EX))) {
3531 ldlm_lock_decref_and_cancel(h, mode);
3532 LDLM_LOCK_PUT(lock);
3534 struct ptlrpc_request *req = mdt_info_req(info);
3536 LASSERT(req != NULL);
3537 tgt_save_slc_lock(&info->mti_mdt->mdt_lut, lock,
3539 ldlm_lock_decref(h, mode);
3548 * Unlock mdt object.
3550 * Immeditely release the regular lock and the PDO lock or save the
3551 * lock in request and keep them referenced until client ACK or
3552 * transaction commit.
3554 * \param info thread info object
3555 * \param o mdt object
3556 * \param lh mdt lock handle referencing regular and PDO locks
3557 * \param decref force immediate lock releasing
3559 * XXX o is not used and may be NULL, see hsm_cdt_request_completed().
3561 void mdt_object_unlock(struct mdt_thread_info *info, struct mdt_object *o,
3562 struct mdt_lock_handle *lh, int decref)
3566 mdt_save_lock(info, &lh->mlh_pdo_lh, lh->mlh_pdo_mode, decref);
3567 mdt_save_lock(info, &lh->mlh_reg_lh, lh->mlh_reg_mode, decref);
3568 mdt_save_remote_lock(info, o, &lh->mlh_rreg_lh, lh->mlh_rreg_mode,
3574 struct mdt_object *mdt_object_find_lock(struct mdt_thread_info *info,
3575 const struct lu_fid *f,
3576 struct mdt_lock_handle *lh,
3579 struct mdt_object *o;
3581 o = mdt_object_find(info->mti_env, info->mti_mdt, f);
3585 rc = mdt_object_lock(info, o, lh, ibits);
3587 mdt_object_put(info->mti_env, o);
3594 void mdt_object_unlock_put(struct mdt_thread_info * info,
3595 struct mdt_object * o,
3596 struct mdt_lock_handle *lh,
3599 mdt_object_unlock(info, o, lh, decref);
3600 mdt_object_put(info->mti_env, o);
3604 * Generic code handling requests that have struct mdt_body passed in:
3606 * - extract mdt_body from request and save it in @info, if present;
3608 * - create lu_object, corresponding to the fid in mdt_body, and save it in
3611 * - if HAS_BODY flag is set for this request type check whether object
3612 * actually exists on storage (lu_object_exists()).
3615 static int mdt_body_unpack(struct mdt_thread_info *info,
3616 enum tgt_handler_flags flags)
3618 const struct mdt_body *body;
3619 struct mdt_object *obj;
3620 const struct lu_env *env;
3621 struct req_capsule *pill;
3625 env = info->mti_env;
3626 pill = info->mti_pill;
3628 body = info->mti_body = req_capsule_client_get(pill, &RMF_MDT_BODY);
3632 if (!(body->mbo_valid & OBD_MD_FLID))
3635 if (!fid_is_sane(&body->mbo_fid1)) {
3636 CERROR("Invalid fid: "DFID"\n", PFID(&body->mbo_fid1));
3640 obj = mdt_object_find(env, info->mti_mdt, &body->mbo_fid1);
3642 if ((flags & HAS_BODY) && !mdt_object_exists(obj)) {
3643 mdt_object_put(env, obj);
3646 info->mti_object = obj;
3655 static int mdt_unpack_req_pack_rep(struct mdt_thread_info *info,
3656 enum tgt_handler_flags flags)
3658 struct req_capsule *pill = info->mti_pill;
3663 if (req_capsule_has_field(pill, &RMF_MDT_BODY, RCL_CLIENT))
3664 rc = mdt_body_unpack(info, flags);
3668 if (rc == 0 && (flags & HAS_REPLY)) {
3670 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
3671 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER,
3673 if (req_capsule_has_field(pill, &RMF_LOGCOOKIES, RCL_SERVER))
3674 req_capsule_set_size(pill, &RMF_LOGCOOKIES,
3677 /* Set ACL reply buffer size as LUSTRE_POSIX_ACL_MAX_SIZE_OLD
3678 * by default. If the target object has more ACL entries, then
3679 * enlarge the buffer when necessary. */
3680 if (req_capsule_has_field(pill, &RMF_ACL, RCL_SERVER))
3681 req_capsule_set_size(pill, &RMF_ACL, RCL_SERVER,
3682 LUSTRE_POSIX_ACL_MAX_SIZE_OLD);
3684 mdt_preset_secctx_size(info);
3686 rc = req_capsule_server_pack(pill);
3688 CWARN("%s: cannot pack response: rc = %d\n",
3689 mdt_obd_name(info->mti_mdt), rc);
3694 void mdt_lock_handle_init(struct mdt_lock_handle *lh)
3696 lh->mlh_type = MDT_NUL_LOCK;
3697 lh->mlh_reg_lh.cookie = 0ull;
3698 lh->mlh_reg_mode = LCK_MINMODE;
3699 lh->mlh_pdo_lh.cookie = 0ull;
3700 lh->mlh_pdo_mode = LCK_MINMODE;
3701 lh->mlh_rreg_lh.cookie = 0ull;
3702 lh->mlh_rreg_mode = LCK_MINMODE;
3705 void mdt_lock_handle_fini(struct mdt_lock_handle *lh)
3707 LASSERT(!lustre_handle_is_used(&lh->mlh_reg_lh));
3708 LASSERT(!lustre_handle_is_used(&lh->mlh_pdo_lh));
3712 * Initialize fields of struct mdt_thread_info. Other fields are left in
3713 * uninitialized state, because it's too expensive to zero out whole
3714 * mdt_thread_info (> 1K) on each request arrival.
3716 void mdt_thread_info_init(struct ptlrpc_request *req,
3717 struct mdt_thread_info *info)
3721 info->mti_pill = &req->rq_pill;
3724 for (i = 0; i < ARRAY_SIZE(info->mti_lh); i++)
3725 mdt_lock_handle_init(&info->mti_lh[i]);
3727 /* mdt device: it can be NULL while CONNECT */
3728 if (req->rq_export) {
3729 info->mti_mdt = mdt_dev(req->rq_export->exp_obd->obd_lu_dev);
3730 info->mti_exp = req->rq_export;
3732 info->mti_mdt = NULL;
3733 info->mti_env = req->rq_svc_thread->t_env;
3734 info->mti_transno = lustre_msg_get_transno(req->rq_reqmsg);
3736 memset(&info->mti_attr, 0, sizeof(info->mti_attr));
3737 info->mti_big_buf = LU_BUF_NULL;
3738 info->mti_body = NULL;
3739 info->mti_object = NULL;
3740 info->mti_dlm_req = NULL;
3741 info->mti_has_trans = 0;
3742 info->mti_cross_ref = 0;
3743 info->mti_opdata = 0;
3744 info->mti_big_lmm_used = 0;
3745 info->mti_big_acl_used = 0;
3746 info->mti_som_valid = 0;
3748 info->mti_spec.no_create = 0;
3749 info->mti_spec.sp_rm_entry = 0;
3750 info->mti_spec.sp_permitted = 0;
3751 info->mti_spec.sp_migrate_close = 0;
3753 info->mti_spec.u.sp_ea.eadata = NULL;
3754 info->mti_spec.u.sp_ea.eadatalen = 0;
3757 void mdt_thread_info_fini(struct mdt_thread_info *info)
3761 if (info->mti_object != NULL) {
3762 mdt_object_put(info->mti_env, info->mti_object);
3763 info->mti_object = NULL;
3766 for (i = 0; i < ARRAY_SIZE(info->mti_lh); i++)
3767 mdt_lock_handle_fini(&info->mti_lh[i]);
3768 info->mti_env = NULL;
3769 info->mti_pill = NULL;
3770 info->mti_exp = NULL;
3771 info->mti_mdt = NULL;
3773 if (unlikely(info->mti_big_buf.lb_buf != NULL))
3774 lu_buf_free(&info->mti_big_buf);
3777 struct mdt_thread_info *tsi2mdt_info(struct tgt_session_info *tsi)
3779 struct mdt_thread_info *mti;
3781 mti = mdt_th_info(tsi->tsi_env);
3782 LASSERT(mti != NULL);
3784 mdt_thread_info_init(tgt_ses_req(tsi), mti);
3785 if (tsi->tsi_corpus != NULL) {
3786 mti->mti_object = mdt_obj(tsi->tsi_corpus);
3787 lu_object_get(tsi->tsi_corpus);
3789 mti->mti_body = tsi->tsi_mdt_body;
3790 mti->mti_dlm_req = tsi->tsi_dlm_req;
3795 static int mdt_tgt_connect(struct tgt_session_info *tsi)
3797 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_DELAY_CONDITIONAL) &&
3799 tsi2mdt_info(tsi)->mti_mdt->mdt_seq_site.ss_node_id)
3800 schedule_timeout_uninterruptible(cfs_time_seconds(3));
3802 return tgt_connect(tsi);
3805 static int mdt_intent_glimpse(enum ldlm_intent_flags it_opc,
3806 struct mdt_thread_info *info,
3807 struct ldlm_lock **lockp, __u64 flags)
3809 return mdt_glimpse_enqueue(info, info->mti_mdt->mdt_namespace,
3812 static int mdt_intent_brw(enum ldlm_intent_flags it_opc,
3813 struct mdt_thread_info *info,
3814 struct ldlm_lock **lockp, __u64 flags)
3816 return mdt_brw_enqueue(info, info->mti_mdt->mdt_namespace,
3820 int mdt_intent_lock_replace(struct mdt_thread_info *info,
3821 struct ldlm_lock **lockp,
3822 struct mdt_lock_handle *lh,
3823 __u64 flags, int result)
3825 struct ptlrpc_request *req = mdt_info_req(info);
3826 struct ldlm_lock *lock = *lockp;
3827 struct ldlm_lock *new_lock;
3829 /* If possible resent found a lock, @lh is set to its handle */
3830 new_lock = ldlm_handle2lock_long(&lh->mlh_reg_lh, 0);
3832 if (new_lock == NULL && (flags & LDLM_FL_INTENT_ONLY)) {
3833 lh->mlh_reg_lh.cookie = 0;
3837 if (new_lock == NULL && (flags & LDLM_FL_RESENT)) {
3838 /* Lock is pinned by ldlm_handle_enqueue0() as it is
3839 * a resend case, however, it could be already destroyed
3840 * due to client eviction or a raced cancel RPC. */
3841 LDLM_DEBUG_NOLOCK("Invalid lock handle %#llx\n",
3842 lh->mlh_reg_lh.cookie);
3843 lh->mlh_reg_lh.cookie = 0;
3847 LASSERTF(new_lock != NULL,
3848 "lockh %#llx flags %#llx : rc = %d\n",
3849 lh->mlh_reg_lh.cookie, flags, result);
3852 * If we've already given this lock to a client once, then we should
3853 * have no readers or writers. Otherwise, we should have one reader
3854 * _or_ writer ref (which will be zeroed below) before returning the
3857 if (new_lock->l_export == req->rq_export) {
3858 LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
3860 LASSERT(new_lock->l_export == NULL);
3861 LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
3866 if (new_lock->l_export == req->rq_export) {
3868 * Already gave this to the client, which means that we
3869 * reconstructed a reply.
3871 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
3874 LDLM_LOCK_RELEASE(new_lock);
3875 lh->mlh_reg_lh.cookie = 0;
3876 RETURN(ELDLM_LOCK_REPLACED);
3880 * Fixup the lock to be given to the client.
3882 lock_res_and_lock(new_lock);
3883 /* Zero new_lock->l_readers and new_lock->l_writers without triggering
3884 * possible blocking AST. */
3885 while (new_lock->l_readers > 0) {
3886 lu_ref_del(&new_lock->l_reference, "reader", new_lock);
3887 lu_ref_del(&new_lock->l_reference, "user", new_lock);
3888 new_lock->l_readers--;
3890 while (new_lock->l_writers > 0) {
3891 lu_ref_del(&new_lock->l_reference, "writer", new_lock);
3892 lu_ref_del(&new_lock->l_reference, "user", new_lock);
3893 new_lock->l_writers--;
3896 new_lock->l_export = class_export_lock_get(req->rq_export, new_lock);
3897 new_lock->l_blocking_ast = lock->l_blocking_ast;
3898 new_lock->l_completion_ast = lock->l_completion_ast;
3899 if (ldlm_has_dom(new_lock))
3900 new_lock->l_glimpse_ast = ldlm_server_glimpse_ast;
3901 new_lock->l_remote_handle = lock->l_remote_handle;
3902 new_lock->l_flags &= ~LDLM_FL_LOCAL;
3904 unlock_res_and_lock(new_lock);
3906 cfs_hash_add(new_lock->l_export->exp_lock_hash,
3907 &new_lock->l_remote_handle,
3908 &new_lock->l_exp_hash);
3910 LDLM_LOCK_RELEASE(new_lock);
3911 lh->mlh_reg_lh.cookie = 0;
3913 RETURN(ELDLM_LOCK_REPLACED);
3916 void mdt_intent_fixup_resent(struct mdt_thread_info *info,
3917 struct ldlm_lock *new_lock,
3918 struct mdt_lock_handle *lh, __u64 flags)
3920 struct ptlrpc_request *req = mdt_info_req(info);
3921 struct ldlm_request *dlmreq;
3923 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
3926 dlmreq = req_capsule_client_get(info->mti_pill, &RMF_DLM_REQ);
3928 /* Check if this is a resend case (MSG_RESENT is set on RPC) and a
3929 * lock was found by ldlm_handle_enqueue(); if so @lh must be
3931 if (flags & LDLM_FL_RESENT) {
3932 lh->mlh_reg_lh.cookie = new_lock->l_handle.h_cookie;
3933 lh->mlh_reg_mode = new_lock->l_granted_mode;
3935 LDLM_DEBUG(new_lock, "Restoring lock cookie");
3936 DEBUG_REQ(D_DLMTRACE, req, "restoring lock cookie %#llx",
3937 lh->mlh_reg_lh.cookie);
3942 * If the xid matches, then we know this is a resent request, and allow
3943 * it. (It's probably an OPEN, for which we don't send a lock.
3945 if (req_can_reconstruct(req, NULL) != 0)
3949 * This remote handle isn't enqueued, so we never received or processed
3950 * this request. Clear MSG_RESENT, because it can be handled like any
3951 * normal request now.
3953 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
3955 DEBUG_REQ(D_DLMTRACE, req, "no existing lock with rhandle %#llx",
3956 dlmreq->lock_handle[0].cookie);
3959 static int mdt_intent_getxattr(enum ldlm_intent_flags it_opc,
3960 struct mdt_thread_info *info,
3961 struct ldlm_lock **lockp,
3964 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
3965 struct ldlm_reply *ldlm_rep = NULL;
3970 * Initialize lhc->mlh_reg_lh either from a previously granted lock
3971 * (for the resend case) or a new lock. Below we will use it to
3972 * replace the original lock.
3974 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
3975 if (!lustre_handle_is_used(&lhc->mlh_reg_lh)) {
3976 mdt_lock_reg_init(lhc, (*lockp)->l_req_mode);
3977 rc = mdt_object_lock(info, info->mti_object, lhc,
3978 MDS_INODELOCK_XATTR);
3983 rc = mdt_getxattr(info);
3985 if (mdt_info_req(info)->rq_repmsg != NULL)
3986 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
3988 if (ldlm_rep == NULL ||
3989 OBD_FAIL_CHECK(OBD_FAIL_MDS_XATTR_REP)) {
3990 mdt_object_unlock(info, info->mti_object, lhc, 1);
3994 RETURN(err_serious(-EFAULT));
3997 ldlm_rep->lock_policy_res2 = clear_serious(rc);
3999 /* This is left for interop instead of adding a new interop flag.
4001 #if LUSTRE_VERSION_CODE > OBD_OCD_VERSION(3, 0, 0, 0)
4002 if (ldlm_rep->lock_policy_res2) {
4003 mdt_object_unlock(info, info->mti_object, lhc, 1);
4004 RETURN(ELDLM_LOCK_ABORTED);
4008 rc = mdt_intent_lock_replace(info, lockp, lhc, flags, rc);
4012 static int mdt_intent_getattr(enum ldlm_intent_flags it_opc,
4013 struct mdt_thread_info *info,
4014 struct ldlm_lock **lockp,
4017 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
4019 struct ldlm_reply *ldlm_rep;
4020 struct mdt_body *reqbody;
4021 struct mdt_body *repbody;
4025 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
4028 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
4031 info->mti_cross_ref = !!(reqbody->mbo_valid & OBD_MD_FLCROSSREF);
4032 repbody->mbo_eadatasize = 0;
4033 repbody->mbo_aclsize = 0;
4037 child_bits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
4040 child_bits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE |
4044 CERROR("%s: unsupported intent %#x\n",
4045 mdt_obd_name(info->mti_mdt), (unsigned int)it_opc);
4046 GOTO(out_shrink, rc = -EINVAL);
4049 rc = mdt_init_ucred_intent_getattr(info, reqbody);
4051 GOTO(out_shrink, rc);
4053 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
4054 mdt_set_disposition(info, ldlm_rep, DISP_IT_EXECD);
4056 /* Get lock from request for possible resent case. */
4057 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
4059 rc = mdt_getattr_name_lock(info, lhc, child_bits, ldlm_rep);
4060 ldlm_rep->lock_policy_res2 = clear_serious(rc);
4062 if (mdt_get_disposition(ldlm_rep, DISP_LOOKUP_NEG))
4063 ldlm_rep->lock_policy_res2 = 0;
4064 if (!mdt_get_disposition(ldlm_rep, DISP_LOOKUP_POS) ||
4065 ldlm_rep->lock_policy_res2) {
4066 lhc->mlh_reg_lh.cookie = 0ull;
4067 GOTO(out_ucred, rc = ELDLM_LOCK_ABORTED);
4070 rc = mdt_intent_lock_replace(info, lockp, lhc, flags, rc);
4073 mdt_exit_ucred(info);
4075 mdt_client_compatibility(info);
4076 rc2 = mdt_fix_reply(info);
4082 static int mdt_intent_layout(enum ldlm_intent_flags it_opc,
4083 struct mdt_thread_info *info,
4084 struct ldlm_lock **lockp,
4087 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
4088 struct md_layout_change layout = { .mlc_opc = MD_LAYOUT_NOP };
4089 struct layout_intent *intent;
4090 struct ldlm_reply *ldlm_rep;
4091 struct lu_fid *fid = &info->mti_tmp_fid2;
4092 struct mdt_object *obj = NULL;
4093 int layout_size = 0;
4094 struct lu_buf *buf = &layout.mlc_buf;
4099 fid_extract_from_res_name(fid, &(*lockp)->l_resource->lr_name);
4101 intent = req_capsule_client_get(info->mti_pill, &RMF_LAYOUT_INTENT);
4105 CDEBUG(D_INFO, DFID "got layout change request from client: "
4106 "opc:%u flags:%#x extent "DEXT"\n",
4107 PFID(fid), intent->li_opc, intent->li_flags,
4108 PEXT(&intent->li_extent));
4110 switch (intent->li_opc) {
4111 case LAYOUT_INTENT_TRUNC:
4112 case LAYOUT_INTENT_WRITE:
4113 layout.mlc_opc = MD_LAYOUT_WRITE;
4114 layout.mlc_intent = intent;
4116 case LAYOUT_INTENT_ACCESS:
4118 case LAYOUT_INTENT_READ:
4119 case LAYOUT_INTENT_GLIMPSE:
4120 case LAYOUT_INTENT_RELEASE:
4121 case LAYOUT_INTENT_RESTORE:
4122 CERROR("%s: Unsupported layout intent opc %d\n",
4123 mdt_obd_name(info->mti_mdt), intent->li_opc);
4126 CERROR("%s: Unknown layout intent opc %d\n",
4127 mdt_obd_name(info->mti_mdt), intent->li_opc);
4131 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
4133 RETURN(PTR_ERR(obj));
4135 if (mdt_object_exists(obj) && !mdt_object_remote(obj)) {
4136 /* if layout is going to be changed don't use the current EA
4137 * size but the maximum one. That buffer will be shrinked
4138 * to the actual size in req_capsule_shrink() before reply.
4140 if (layout.mlc_opc == MD_LAYOUT_WRITE) {
4141 layout_size = info->mti_mdt->mdt_max_mdsize;
4143 layout_size = mdt_attr_get_eabuf_size(info, obj);
4144 if (layout_size < 0)
4145 GOTO(out, rc = layout_size);
4147 if (layout_size > info->mti_mdt->mdt_max_mdsize)
4148 info->mti_mdt->mdt_max_mdsize = layout_size;
4150 CDEBUG(D_INFO, "%s: layout_size %d\n",
4151 mdt_obd_name(info->mti_mdt), layout_size);
4155 * set reply buffer size, so that ldlm_handle_enqueue0()->
4156 * ldlm_lvbo_fill() will fill the reply buffer with lovea.
4158 req_capsule_set_size(info->mti_pill, &RMF_DLM_LVB, RCL_SERVER,
4160 rc = req_capsule_server_pack(info->mti_pill);
4164 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
4166 GOTO(out, rc = -EPROTO);
4168 mdt_set_disposition(info, ldlm_rep, DISP_IT_EXECD);
4170 /* take lock in ldlm_lock_enqueue() for LAYOUT_INTENT_ACCESS */
4171 if (layout.mlc_opc == MD_LAYOUT_NOP)
4174 rc = mdt_check_resent(info, mdt_reconstruct_generic, lhc);
4178 DEBUG_REQ(D_INODE, mdt_info_req(info), "resent opt.");
4179 rc = lustre_msg_get_status(mdt_info_req(info)->rq_repmsg);
4185 if (unlikely(req_is_replay(mdt_info_req(info)))) {
4186 buf->lb_buf = req_capsule_client_get(info->mti_pill,
4188 buf->lb_len = req_capsule_get_size(info->mti_pill,
4189 &RMF_EADATA, RCL_CLIENT);
4191 * If it's a replay of layout write intent RPC, the client has
4192 * saved the extended lovea when it get reply then.
4194 if (buf->lb_len > 0)
4195 mdt_fix_lov_magic(info, buf->lb_buf);
4198 /* Get lock from request for possible resent case. */
4199 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
4200 (*lockp)->l_lvb_type = LVB_T_LAYOUT;
4203 * Instantiate some layout components, if @buf contains lovea, then it's
4204 * a replay of the layout intent write RPC.
4206 rc = mdt_layout_change(info, obj, lhc, &layout);
4207 ldlm_rep->lock_policy_res2 = clear_serious(rc);
4209 if (lustre_handle_is_used(&lhc->mlh_reg_lh)) {
4210 rc = mdt_intent_lock_replace(info, lockp, lhc, flags, rc);
4211 if (rc == ELDLM_LOCK_REPLACED &&
4212 (*lockp)->l_granted_mode == LCK_EX)
4213 ldlm_lock_mode_downgrade(*lockp, LCK_CR);
4218 mdt_object_put(info->mti_env, obj);
4222 static int mdt_intent_open(enum ldlm_intent_flags it_opc,
4223 struct mdt_thread_info *info,
4224 struct ldlm_lock **lockp,
4227 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
4228 struct ldlm_reply *rep = NULL;
4232 static const struct req_format *intent_fmts[REINT_MAX] = {
4233 [REINT_CREATE] = &RQF_LDLM_INTENT_CREATE,
4234 [REINT_OPEN] = &RQF_LDLM_INTENT_OPEN
4239 opc = mdt_reint_opcode(mdt_info_req(info), intent_fmts);
4243 /* Get lock from request for possible resent case. */
4244 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
4246 rc = mdt_reint_internal(info, lhc, opc);
4248 /* Check whether the reply has been packed successfully. */
4249 if (mdt_info_req(info)->rq_repmsg != NULL)
4250 rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
4255 RETURN(err_serious(-EFAULT));
4258 /* MDC expects this in any case */
4260 mdt_set_disposition(info, rep, DISP_LOOKUP_EXECD);
4262 /* the open lock or the lock for cross-ref object should be
4263 * returned to the client */
4264 if (lustre_handle_is_used(&lhc->mlh_reg_lh) &&
4265 (rc == 0 || rc == -MDT_EREMOTE_OPEN)) {
4266 rep->lock_policy_res2 = 0;
4267 rc = mdt_intent_lock_replace(info, lockp, lhc, flags, rc);
4271 rep->lock_policy_res2 = clear_serious(rc);
4273 if (rep->lock_policy_res2 == -ENOENT &&
4274 mdt_get_disposition(rep, DISP_LOOKUP_NEG) &&
4275 !mdt_get_disposition(rep, DISP_OPEN_CREATE))
4276 rep->lock_policy_res2 = 0;
4278 lhc->mlh_reg_lh.cookie = 0ull;
4279 if (rc == -ENOTCONN || rc == -ENODEV ||
4280 rc == -EOVERFLOW) { /**< if VBR failure then return error */
4282 * If it is the disconnect error (ENODEV & ENOCONN), the error
4283 * will be returned by rq_status, and client at ptlrpc layer
4284 * will detect this, then disconnect, reconnect the import
4285 * immediately, instead of impacting the following the rpc.
4290 * For other cases, the error will be returned by intent, and client
4291 * will retrieve the result from intent.
4293 RETURN(ELDLM_LOCK_ABORTED);
4296 static int mdt_intent_opc(enum ldlm_intent_flags it_opc,
4297 struct mdt_thread_info *info,
4298 struct ldlm_lock **lockp,
4299 u64 flags /* LDLM_FL_* */)
4301 struct req_capsule *pill = info->mti_pill;
4302 struct ptlrpc_request *req = mdt_info_req(info);
4303 const struct req_format *it_format;
4304 int (*it_handler)(enum ldlm_intent_flags,
4305 struct mdt_thread_info *,
4306 struct ldlm_lock **,
4308 enum tgt_handler_flags it_handler_flags = 0;
4309 struct ldlm_reply *rep;
4310 bool check_mdt_object = false;
4316 case IT_OPEN|IT_CREAT:
4318 * OCREAT is not a IS_MUTABLE request since the file may
4319 * already exist. We do the extra check of
4320 * OBD_CONNECT_RDONLY in mdt_reint_open() when we
4321 * really need to create the object.
4323 it_format = &RQF_LDLM_INTENT;
4324 it_handler = &mdt_intent_open;
4327 check_mdt_object = true;
4330 it_format = &RQF_LDLM_INTENT_GETATTR;
4331 it_handler = &mdt_intent_getattr;
4332 it_handler_flags = HAS_REPLY;
4335 check_mdt_object = true;
4336 it_format = &RQF_LDLM_INTENT_GETXATTR;
4337 it_handler = &mdt_intent_getxattr;
4338 it_handler_flags = HAS_BODY;
4341 it_format = &RQF_LDLM_INTENT_LAYOUT;
4342 it_handler = &mdt_intent_layout;
4345 it_format = &RQF_LDLM_INTENT;
4346 it_handler = &mdt_intent_glimpse;
4349 it_format = &RQF_LDLM_INTENT;
4350 it_handler = &mdt_intent_brw;
4352 case IT_QUOTA_DQACQ:
4353 case IT_QUOTA_CONN: {
4354 struct lu_device *qmt = info->mti_mdt->mdt_qmt_dev;
4357 RETURN(-EOPNOTSUPP);
4359 if (mdt_rdonly(req->rq_export))
4362 (*lockp)->l_lvb_type = LVB_T_LQUOTA;
4363 /* pass the request to quota master */
4364 rc = qmt_hdls.qmth_intent_policy(info->mti_env, qmt,
4365 mdt_info_req(info), lockp,
4370 CERROR("%s: unknown intent code %#x\n",
4371 mdt_obd_name(info->mti_mdt), it_opc);
4375 req_capsule_extend(pill, it_format);
4377 rc = mdt_unpack_req_pack_rep(info, it_handler_flags);
4381 if (unlikely(info->mti_object == NULL && check_mdt_object))
4384 if (it_handler_flags & IS_MUTABLE && mdt_rdonly(req->rq_export))
4387 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_INTENT_DELAY, 10);
4389 /* execute policy */
4390 rc = (*it_handler)(it_opc, info, lockp, flags);
4392 /* Check whether the reply has been packed successfully. */
4393 if (req->rq_repmsg != NULL) {
4394 rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
4395 rep->lock_policy_res2 =
4396 ptlrpc_status_hton(rep->lock_policy_res2);
4402 static void mdt_ptlrpc_stats_update(struct ptlrpc_request *req,
4403 enum ldlm_intent_flags it_opc)
4405 struct lprocfs_stats *srv_stats = ptlrpc_req2svc(req)->srv_stats;
4407 /* update stats when IT code is known */
4408 if (srv_stats != NULL)
4409 lprocfs_counter_incr(srv_stats,
4410 PTLRPC_LAST_CNTR + (it_opc == IT_GLIMPSE ?
4411 LDLM_GLIMPSE_ENQUEUE : LDLM_IBITS_ENQUEUE));
4414 static int mdt_intent_policy(const struct lu_env *env,
4415 struct ldlm_namespace *ns,
4416 struct ldlm_lock **lockp,
4418 enum ldlm_mode mode,
4419 __u64 flags, void *data)
4421 struct tgt_session_info *tsi;
4422 struct mdt_thread_info *info;
4423 struct ptlrpc_request *req = req_cookie;
4424 struct ldlm_intent *it;
4425 struct req_capsule *pill;
4426 const struct ldlm_lock_desc *ldesc;
4431 LASSERT(req != NULL);
4433 tsi = tgt_ses_info(env);
4435 info = tsi2mdt_info(tsi);
4436 LASSERT(info != NULL);
4437 pill = info->mti_pill;
4438 LASSERT(pill->rc_req == req);
4439 ldesc = &info->mti_dlm_req->lock_desc;
4441 if (req->rq_reqmsg->lm_bufcount > DLM_INTENT_IT_OFF) {
4442 req_capsule_extend(pill, &RQF_LDLM_INTENT_BASIC);
4443 it = req_capsule_client_get(pill, &RMF_LDLM_INTENT);
4445 mdt_ptlrpc_stats_update(req, it->opc);
4446 rc = mdt_intent_opc(it->opc, info, lockp, flags);
4450 /* Lock without inodebits makes no sense and will oops
4451 * later in ldlm. Let's check it now to see if we have
4452 * ibits corrupted somewhere in mdt_intent_opc().
4453 * The case for client miss to set ibits has been
4454 * processed by others. */
4455 LASSERT(ergo(ldesc->l_resource.lr_type == LDLM_IBITS,
4456 ldesc->l_policy_data.l_inodebits.bits != 0));
4458 rc = err_serious(-EFAULT);
4461 /* No intent was provided */
4462 req_capsule_set_size(pill, &RMF_DLM_LVB, RCL_SERVER, 0);
4463 rc = req_capsule_server_pack(pill);
4465 rc = err_serious(rc);
4467 mdt_thread_info_fini(info);
4471 static void mdt_deregister_seq_exp(struct mdt_device *mdt)
4473 struct seq_server_site *ss = mdt_seq_site(mdt);
4475 if (ss->ss_node_id == 0)
4478 if (ss->ss_client_seq != NULL) {
4479 lustre_deregister_lwp_item(&ss->ss_client_seq->lcs_exp);
4480 ss->ss_client_seq->lcs_exp = NULL;
4483 if (ss->ss_server_fld != NULL) {
4484 lustre_deregister_lwp_item(&ss->ss_server_fld->lsf_control_exp);
4485 ss->ss_server_fld->lsf_control_exp = NULL;
4489 static void mdt_seq_fini_cli(struct mdt_device *mdt)
4491 struct seq_server_site *ss = mdt_seq_site(mdt);
4496 if (ss->ss_server_seq != NULL)
4497 seq_server_set_cli(NULL, ss->ss_server_seq, NULL);
4500 static int mdt_seq_fini(const struct lu_env *env, struct mdt_device *mdt)
4502 mdt_seq_fini_cli(mdt);
4503 mdt_deregister_seq_exp(mdt);
4505 return seq_site_fini(env, mdt_seq_site(mdt));
4509 * It will retrieve its FLDB entries from MDT0, and it only happens
4510 * when upgrading existent FS to 2.6 or when local FLDB is corrupted,
4511 * and it needs to refresh FLDB from the MDT0.
4513 static int mdt_register_lwp_callback(void *data)
4516 struct mdt_device *mdt = data;
4517 struct lu_server_fld *fld = mdt_seq_site(mdt)->ss_server_fld;
4521 LASSERT(mdt_seq_site(mdt)->ss_node_id != 0);
4523 rc = lu_env_init(&env, LCT_MD_THREAD);
4525 CERROR("%s: cannot init env: rc = %d\n", mdt_obd_name(mdt), rc);
4529 /* Allocate new sequence now to avoid creating local transaction
4530 * in the normal transaction process */
4531 rc = seq_server_check_and_alloc_super(&env,
4532 mdt_seq_site(mdt)->ss_server_seq);
4537 rc = fld_update_from_controller(&env, fld);
4539 CERROR("%s: cannot update controller: rc = %d\n",
4540 mdt_obd_name(mdt), rc);
4549 static int mdt_register_seq_exp(struct mdt_device *mdt)
4551 struct seq_server_site *ss = mdt_seq_site(mdt);
4552 char *lwp_name = NULL;
4555 if (ss->ss_node_id == 0)
4558 OBD_ALLOC(lwp_name, MAX_OBD_NAME);
4559 if (lwp_name == NULL)
4560 GOTO(out_free, rc = -ENOMEM);
4562 rc = tgt_name2lwp_name(mdt_obd_name(mdt), lwp_name, MAX_OBD_NAME, 0);
4566 rc = lustre_register_lwp_item(lwp_name, &ss->ss_client_seq->lcs_exp,
4571 rc = lustre_register_lwp_item(lwp_name,
4572 &ss->ss_server_fld->lsf_control_exp,
4573 mdt_register_lwp_callback, mdt);
4575 lustre_deregister_lwp_item(&ss->ss_client_seq->lcs_exp);
4576 ss->ss_client_seq->lcs_exp = NULL;
4580 if (lwp_name != NULL)
4581 OBD_FREE(lwp_name, MAX_OBD_NAME);
4587 * Init client sequence manager which is used by local MDS to talk to sequence
4588 * controller on remote node.
4590 static int mdt_seq_init_cli(const struct lu_env *env, struct mdt_device *mdt)
4592 struct seq_server_site *ss = mdt_seq_site(mdt);
4596 /* check if this is adding the first MDC and controller is not yet
4598 OBD_ALLOC_PTR(ss->ss_client_seq);
4599 if (ss->ss_client_seq == NULL)
4602 OBD_ALLOC(prefix, MAX_OBD_NAME + 5);
4603 if (prefix == NULL) {
4604 OBD_FREE_PTR(ss->ss_client_seq);
4605 ss->ss_client_seq = NULL;
4609 /* Note: seq_client_fini will be called in seq_site_fini */
4610 snprintf(prefix, MAX_OBD_NAME + 5, "ctl-%s", mdt_obd_name(mdt));
4611 seq_client_init(ss->ss_client_seq, NULL, LUSTRE_SEQ_METADATA,
4612 prefix, ss->ss_node_id == 0 ? ss->ss_control_seq :
4614 OBD_FREE(prefix, MAX_OBD_NAME + 5);
4616 RETURN(seq_server_set_cli(env, ss->ss_server_seq, ss->ss_client_seq));
4619 static int mdt_seq_init(const struct lu_env *env, struct mdt_device *mdt)
4621 struct seq_server_site *ss;
4625 ss = mdt_seq_site(mdt);
4626 /* init sequence controller server(MDT0) */
4627 if (ss->ss_node_id == 0) {
4628 OBD_ALLOC_PTR(ss->ss_control_seq);
4629 if (ss->ss_control_seq == NULL)
4632 rc = seq_server_init(env, ss->ss_control_seq, mdt->mdt_bottom,
4633 mdt_obd_name(mdt), LUSTRE_SEQ_CONTROLLER,
4636 GOTO(out_seq_fini, rc);
4639 /* Init normal sequence server */
4640 OBD_ALLOC_PTR(ss->ss_server_seq);
4641 if (ss->ss_server_seq == NULL)
4642 GOTO(out_seq_fini, rc = -ENOMEM);
4644 rc = seq_server_init(env, ss->ss_server_seq, mdt->mdt_bottom,
4645 mdt_obd_name(mdt), LUSTRE_SEQ_SERVER, ss);
4647 GOTO(out_seq_fini, rc);
4649 /* init seq client for seq server to talk to seq controller(MDT0) */
4650 rc = mdt_seq_init_cli(env, mdt);
4652 GOTO(out_seq_fini, rc);
4654 if (ss->ss_node_id != 0)
4655 /* register controller export through lwp */
4656 rc = mdt_register_seq_exp(mdt);
4661 mdt_seq_fini(env, mdt);
4669 static int mdt_fld_fini(const struct lu_env *env,
4670 struct mdt_device *m)
4672 struct seq_server_site *ss = mdt_seq_site(m);
4675 if (ss && ss->ss_server_fld) {
4676 fld_server_fini(env, ss->ss_server_fld);
4677 OBD_FREE_PTR(ss->ss_server_fld);
4678 ss->ss_server_fld = NULL;
4684 static int mdt_fld_init(const struct lu_env *env,
4686 struct mdt_device *m)
4688 struct seq_server_site *ss;
4692 ss = mdt_seq_site(m);
4694 OBD_ALLOC_PTR(ss->ss_server_fld);
4695 if (ss->ss_server_fld == NULL)
4696 RETURN(rc = -ENOMEM);
4698 rc = fld_server_init(env, ss->ss_server_fld, m->mdt_bottom, uuid,
4701 OBD_FREE_PTR(ss->ss_server_fld);
4702 ss->ss_server_fld = NULL;
4709 static void mdt_stack_pre_fini(const struct lu_env *env,
4710 struct mdt_device *m, struct lu_device *top)
4712 struct lustre_cfg_bufs *bufs;
4713 struct lustre_cfg *lcfg;
4714 struct mdt_thread_info *info;
4719 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
4720 LASSERT(info != NULL);
4722 bufs = &info->mti_u.bufs;
4724 LASSERT(m->mdt_child_exp);
4725 LASSERT(m->mdt_child_exp->exp_obd);
4727 /* process cleanup, pass mdt obd name to get obd umount flags */
4728 /* XXX: this is needed because all layers are referenced by
4729 * objects (some of them are pinned by osd, for example *
4730 * the proper solution should be a model where object used
4731 * by osd only doesn't have mdt/mdd slices -bzzz */
4732 lustre_cfg_bufs_reset(bufs, mdt_obd_name(m));
4733 lustre_cfg_bufs_set_string(bufs, 1, NULL);
4734 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
4737 lustre_cfg_init(lcfg, LCFG_PRE_CLEANUP, bufs);
4739 top->ld_ops->ldo_process_config(env, top, lcfg);
4740 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens));
4744 static void mdt_stack_fini(const struct lu_env *env,
4745 struct mdt_device *m, struct lu_device *top)
4747 struct obd_device *obd = mdt2obd_dev(m);
4748 struct lustre_cfg_bufs *bufs;
4749 struct lustre_cfg *lcfg;
4750 struct mdt_thread_info *info;
4754 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
4755 LASSERT(info != NULL);
4757 lu_dev_del_linkage(top->ld_site, top);
4759 lu_site_purge(env, top->ld_site, -1);
4761 bufs = &info->mti_u.bufs;
4762 /* process cleanup, pass mdt obd name to get obd umount flags */
4763 /* another purpose is to let all layers to release their objects */
4764 lustre_cfg_bufs_reset(bufs, mdt_obd_name(m));
4769 lustre_cfg_bufs_set_string(bufs, 1, flags);
4770 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
4773 lustre_cfg_init(lcfg, LCFG_CLEANUP, bufs);
4776 top->ld_ops->ldo_process_config(env, top, lcfg);
4777 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens));
4779 lu_site_purge(env, top->ld_site, -1);
4781 m->mdt_child = NULL;
4782 m->mdt_bottom = NULL;
4784 obd_disconnect(m->mdt_child_exp);
4785 m->mdt_child_exp = NULL;
4787 obd_disconnect(m->mdt_bottom_exp);
4788 m->mdt_child_exp = NULL;
4791 static int mdt_connect_to_next(const struct lu_env *env, struct mdt_device *m,
4792 const char *next, struct obd_export **exp)
4794 struct obd_connect_data *data = NULL;
4795 struct obd_device *obd;
4799 OBD_ALLOC_PTR(data);
4801 GOTO(out, rc = -ENOMEM);
4803 obd = class_name2obd(next);
4805 CERROR("%s: can't locate next device: %s\n",
4806 mdt_obd_name(m), next);
4807 GOTO(out, rc = -ENOTCONN);
4810 data->ocd_connect_flags = OBD_CONNECT_VERSION;
4811 data->ocd_version = LUSTRE_VERSION_CODE;
4813 rc = obd_connect(NULL, exp, obd, &obd->obd_uuid, data, NULL);
4815 CERROR("%s: cannot connect to next dev %s (%d)\n",
4816 mdt_obd_name(m), next, rc);
4826 static int mdt_stack_init(const struct lu_env *env, struct mdt_device *mdt,
4827 struct lustre_cfg *cfg)
4829 char *dev = lustre_cfg_string(cfg, 0);
4830 int rc, name_size, uuid_size;
4831 char *name, *uuid, *p;
4832 struct lustre_cfg_bufs *bufs;
4833 struct lustre_cfg *lcfg;
4834 struct obd_device *obd;
4835 struct lustre_profile *lprof;
4836 struct lu_site *site;
4839 /* in 1.8 we had the only device in the stack - MDS.
4840 * 2.0 introduces MDT, MDD, OSD; MDT starts others internally.
4841 * in 2.3 OSD is instantiated by obd_mount.c, so we need
4842 * to generate names and setup MDT, MDD. MDT will be using
4843 * generated name to connect to MDD. for MDD the next device
4844 * will be LOD with name taken from so called "profile" which
4845 * is generated by mount_option line
4847 * 1.8 MGS generates config. commands like this:
4848 * #06 (104)mount_option 0: 1:lustre-MDT0000 2:lustre-mdtlov
4849 * #08 (120)setup 0:lustre-MDT0000 1:dev 2:type 3:lustre-MDT0000
4850 * 2.0 MGS generates config. commands like this:
4851 * #07 (112)mount_option 0: 1:lustre-MDT0000 2:lustre-MDT0000-mdtlov
4852 * #08 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
4853 * 3:lustre-MDT0000-mdtlov 4:f
4855 * we generate MDD name from MDT one, just replacing T with D
4857 * after all the preparations, the logical equivalent will be
4858 * #01 (160)setup 0:lustre-MDD0000 1:lustre-MDD0000_UUID 2:0
4859 * 3:lustre-MDT0000-mdtlov 4:f
4860 * #02 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
4861 * 3:lustre-MDD0000 4:f
4863 * notice we build the stack from down to top: MDD first, then MDT */
4865 name_size = MAX_OBD_NAME;
4866 uuid_size = MAX_OBD_NAME;
4868 OBD_ALLOC(name, name_size);
4869 OBD_ALLOC(uuid, uuid_size);
4870 if (name == NULL || uuid == NULL)
4871 GOTO(cleanup_mem, rc = -ENOMEM);
4873 OBD_ALLOC_PTR(bufs);
4875 GOTO(cleanup_mem, rc = -ENOMEM);
4878 p = strstr(name, "-MDT");
4880 GOTO(free_bufs, rc = -ENOMEM);
4883 snprintf(uuid, MAX_OBD_NAME, "%s_UUID", name);
4885 lprof = class_get_profile(lustre_cfg_string(cfg, 0));
4886 if (lprof == NULL || lprof->lp_dt == NULL) {
4887 CERROR("can't find the profile: %s\n",
4888 lustre_cfg_string(cfg, 0));
4889 GOTO(free_bufs, rc = -EINVAL);
4892 lustre_cfg_bufs_reset(bufs, name);
4893 lustre_cfg_bufs_set_string(bufs, 1, LUSTRE_MDD_NAME);
4894 lustre_cfg_bufs_set_string(bufs, 2, uuid);
4895 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
4897 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
4899 GOTO(put_profile, rc = -ENOMEM);
4900 lustre_cfg_init(lcfg, LCFG_ATTACH, bufs);
4902 rc = class_attach(lcfg);
4904 GOTO(lcfg_cleanup, rc);
4906 obd = class_name2obd(name);
4908 CERROR("Can not find obd %s (%s in config)\n",
4909 MDD_OBD_NAME, lustre_cfg_string(cfg, 0));
4910 GOTO(lcfg_cleanup, rc = -EINVAL);
4913 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens));
4915 lustre_cfg_bufs_reset(bufs, name);
4916 lustre_cfg_bufs_set_string(bufs, 1, uuid);
4917 lustre_cfg_bufs_set_string(bufs, 2, dev);
4918 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
4920 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
4922 GOTO(class_detach, rc = -ENOMEM);
4923 lustre_cfg_init(lcfg, LCFG_SETUP, bufs);
4925 rc = class_setup(obd, lcfg);
4927 GOTO(class_detach, rc);
4929 /* connect to MDD we just setup */
4930 rc = mdt_connect_to_next(env, mdt, name, &mdt->mdt_child_exp);
4932 GOTO(class_detach, rc);
4934 site = mdt->mdt_child_exp->exp_obd->obd_lu_dev->ld_site;
4936 LASSERT(mdt_lu_site(mdt) == NULL);
4937 mdt->mdt_lu_dev.ld_site = site;
4938 site->ls_top_dev = &mdt->mdt_lu_dev;
4939 mdt->mdt_child = lu2md_dev(mdt->mdt_child_exp->exp_obd->obd_lu_dev);
4941 /* now connect to bottom OSD */
4942 snprintf(name, MAX_OBD_NAME, "%s-osd", dev);
4943 rc = mdt_connect_to_next(env, mdt, name, &mdt->mdt_bottom_exp);
4945 GOTO(class_detach, rc);
4947 lu2dt_dev(mdt->mdt_bottom_exp->exp_obd->obd_lu_dev);
4949 rc = lu_env_refill((struct lu_env *)env);
4951 CERROR("Failure to refill session: '%d'\n", rc);
4953 lu_dev_add_linkage(site, &mdt->mdt_lu_dev);
4958 class_detach(obd, lcfg);
4960 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens));
4962 class_put_profile(lprof);
4967 OBD_FREE(name, name_size);
4969 OBD_FREE(uuid, uuid_size);
4973 /* setup quota master target on MDT0 */
4974 static int mdt_quota_init(const struct lu_env *env, struct mdt_device *mdt,
4975 struct lustre_cfg *cfg)
4977 struct obd_device *obd;
4978 char *dev = lustre_cfg_string(cfg, 0);
4979 char *qmtname, *uuid, *p;
4980 struct lustre_cfg_bufs *bufs;
4981 struct lustre_cfg *lcfg;
4982 struct lustre_profile *lprof;
4983 struct obd_connect_data *data;
4987 LASSERT(mdt->mdt_qmt_exp == NULL);
4988 LASSERT(mdt->mdt_qmt_dev == NULL);
4990 /* quota master is on MDT0 only for now */
4991 if (mdt->mdt_seq_site.ss_node_id != 0)
4994 /* MGS generates config commands which look as follows:
4995 * #01 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
4996 * 3:lustre-MDT0000-mdtlov 4:f
4998 * We generate the QMT name from the MDT one, just replacing MD with QM
4999 * after all the preparations, the logical equivalent will be:
5000 * #01 (160)setup 0:lustre-QMT0000 1:lustre-QMT0000_UUID 2:0
5001 * 3:lustre-MDT0000-osd 4:f */
5002 OBD_ALLOC(qmtname, MAX_OBD_NAME);
5003 OBD_ALLOC(uuid, UUID_MAX);
5004 OBD_ALLOC_PTR(bufs);
5005 OBD_ALLOC_PTR(data);
5006 if (qmtname == NULL || uuid == NULL || bufs == NULL || data == NULL)
5007 GOTO(cleanup_mem, rc = -ENOMEM);
5009 strcpy(qmtname, dev);
5010 p = strstr(qmtname, "-MDT");
5012 GOTO(cleanup_mem, rc = -ENOMEM);
5013 /* replace MD with QM */
5017 snprintf(uuid, UUID_MAX, "%s_UUID", qmtname);
5019 lprof = class_get_profile(lustre_cfg_string(cfg, 0));
5020 if (lprof == NULL || lprof->lp_dt == NULL) {
5021 CERROR("can't find profile for %s\n",
5022 lustre_cfg_string(cfg, 0));
5023 GOTO(cleanup_mem, rc = -EINVAL);
5026 lustre_cfg_bufs_reset(bufs, qmtname);
5027 lustre_cfg_bufs_set_string(bufs, 1, LUSTRE_QMT_NAME);
5028 lustre_cfg_bufs_set_string(bufs, 2, uuid);
5029 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
5031 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
5033 GOTO(put_profile, rc = -ENOMEM);
5034 lustre_cfg_init(lcfg, LCFG_ATTACH, bufs);
5036 rc = class_attach(lcfg);
5038 GOTO(lcfg_cleanup, rc);
5040 obd = class_name2obd(qmtname);
5042 CERROR("Can not find obd %s (%s in config)\n", qmtname,
5043 lustre_cfg_string(cfg, 0));
5044 GOTO(lcfg_cleanup, rc = -EINVAL);
5047 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens));
5049 lustre_cfg_bufs_reset(bufs, qmtname);
5050 lustre_cfg_bufs_set_string(bufs, 1, uuid);
5051 lustre_cfg_bufs_set_string(bufs, 2, dev);
5053 /* for quota, the next device should be the OSD device */
5054 lustre_cfg_bufs_set_string(bufs, 3,
5055 mdt->mdt_bottom->dd_lu_dev.ld_obd->obd_name);
5057 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
5059 GOTO(class_detach, rc = -ENOMEM);
5060 lustre_cfg_init(lcfg, LCFG_SETUP, bufs);
5062 rc = class_setup(obd, lcfg);
5064 GOTO(class_detach, rc);
5066 mdt->mdt_qmt_dev = obd->obd_lu_dev;
5068 /* configure local quota objects */
5069 if (OBD_FAIL_CHECK(OBD_FAIL_QUOTA_INIT))
5072 rc = mdt->mdt_qmt_dev->ld_ops->ldo_prepare(env,
5076 GOTO(class_cleanup, rc);
5078 /* connect to quota master target */
5079 data->ocd_connect_flags = OBD_CONNECT_VERSION;
5080 data->ocd_version = LUSTRE_VERSION_CODE;
5081 rc = obd_connect(NULL, &mdt->mdt_qmt_exp, obd, &obd->obd_uuid,
5084 CERROR("cannot connect to quota master device %s (%d)\n",
5086 GOTO(class_cleanup, rc);
5092 class_manual_cleanup(obd);
5093 mdt->mdt_qmt_dev = NULL;
5094 GOTO(lcfg_cleanup, rc);
5098 class_detach(obd, lcfg);
5100 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens));
5102 class_put_profile(lprof);
5107 OBD_FREE(qmtname, MAX_OBD_NAME);
5109 OBD_FREE(uuid, UUID_MAX);
5115 /* Shutdown quota master target associated with mdt */
5116 static void mdt_quota_fini(const struct lu_env *env, struct mdt_device *mdt)
5120 if (mdt->mdt_qmt_exp == NULL)
5122 LASSERT(mdt->mdt_qmt_dev != NULL);
5124 /* the qmt automatically shuts down when the mdt disconnects */
5125 obd_disconnect(mdt->mdt_qmt_exp);
5126 mdt->mdt_qmt_exp = NULL;
5127 mdt->mdt_qmt_dev = NULL;
5131 /* mdt_getxattr() is used from mdt_intent_getxattr(), use this wrapper
5132 * for now. This will be removed along with converting rest of MDT code
5133 * to use tgt_session_info */
5134 static int mdt_tgt_getxattr(struct tgt_session_info *tsi)
5136 struct mdt_thread_info *info = tsi2mdt_info(tsi);
5139 if (unlikely(info->mti_object == NULL))
5142 rc = mdt_getxattr(info);
5144 mdt_thread_info_fini(info);
5148 #define OBD_FAIL_OST_READ_NET OBD_FAIL_OST_BRW_NET
5149 #define OBD_FAIL_OST_WRITE_NET OBD_FAIL_OST_BRW_NET
5150 #define OST_BRW_READ OST_READ
5151 #define OST_BRW_WRITE OST_WRITE
5153 static struct tgt_handler mdt_tgt_handlers[] = {
5154 TGT_RPC_HANDLER(MDS_FIRST_OPC,
5155 0, MDS_CONNECT, mdt_tgt_connect,
5156 &RQF_CONNECT, LUSTRE_OBD_VERSION),
5157 TGT_RPC_HANDLER(MDS_FIRST_OPC,
5158 0, MDS_DISCONNECT, tgt_disconnect,
5159 &RQF_MDS_DISCONNECT, LUSTRE_OBD_VERSION),
5160 TGT_RPC_HANDLER(MDS_FIRST_OPC,
5161 HAS_REPLY, MDS_SET_INFO, mdt_set_info,
5162 &RQF_OBD_SET_INFO, LUSTRE_MDS_VERSION),
5163 TGT_MDT_HDL(0, MDS_GET_INFO, mdt_get_info),
5164 TGT_MDT_HDL(HAS_REPLY, MDS_GET_ROOT, mdt_get_root),
5165 TGT_MDT_HDL(HAS_BODY, MDS_GETATTR, mdt_getattr),
5166 TGT_MDT_HDL(HAS_BODY | HAS_REPLY, MDS_GETATTR_NAME,
5168 TGT_MDT_HDL(HAS_BODY, MDS_GETXATTR, mdt_tgt_getxattr),
5169 TGT_MDT_HDL(HAS_REPLY, MDS_STATFS, mdt_statfs),
5170 TGT_MDT_HDL(IS_MUTABLE, MDS_REINT, mdt_reint),
5171 TGT_MDT_HDL(HAS_BODY, MDS_CLOSE, mdt_close),
5172 TGT_MDT_HDL(HAS_BODY | HAS_REPLY, MDS_READPAGE, mdt_readpage),
5173 TGT_MDT_HDL(HAS_BODY | HAS_REPLY, MDS_SYNC, mdt_sync),
5174 TGT_MDT_HDL(0, MDS_QUOTACTL, mdt_quotactl),
5175 TGT_MDT_HDL(HAS_BODY | HAS_REPLY | IS_MUTABLE, MDS_HSM_PROGRESS,
5177 TGT_MDT_HDL(HAS_BODY | HAS_REPLY | IS_MUTABLE, MDS_HSM_CT_REGISTER,
5178 mdt_hsm_ct_register),
5179 TGT_MDT_HDL(HAS_BODY | HAS_REPLY | IS_MUTABLE, MDS_HSM_CT_UNREGISTER,
5180 mdt_hsm_ct_unregister),
5181 TGT_MDT_HDL(HAS_BODY | HAS_REPLY, MDS_HSM_STATE_GET,
5183 TGT_MDT_HDL(HAS_BODY | HAS_REPLY | IS_MUTABLE, MDS_HSM_STATE_SET,
5185 TGT_MDT_HDL(HAS_BODY | HAS_REPLY, MDS_HSM_ACTION, mdt_hsm_action),
5186 TGT_MDT_HDL(HAS_BODY | HAS_REPLY, MDS_HSM_REQUEST,
5188 TGT_MDT_HDL(HAS_KEY | HAS_BODY | HAS_REPLY | IS_MUTABLE,
5191 TGT_MDT_HDL(IS_MUTABLE, MDS_RMFID, mdt_rmfid),
5194 static struct tgt_handler mdt_io_ops[] = {
5195 TGT_OST_HDL_HP(HAS_BODY | HAS_REPLY, OST_BRW_READ, tgt_brw_read,
5197 TGT_OST_HDL_HP(HAS_BODY | IS_MUTABLE, OST_BRW_WRITE, tgt_brw_write,
5199 TGT_OST_HDL_HP(HAS_BODY | HAS_REPLY | IS_MUTABLE,
5200 OST_PUNCH, mdt_punch_hdl,
5202 TGT_OST_HDL(HAS_BODY | HAS_REPLY, OST_SYNC, mdt_data_sync),
5205 static struct tgt_handler mdt_sec_ctx_ops[] = {
5206 TGT_SEC_HDL_VAR(0, SEC_CTX_INIT, mdt_sec_ctx_handle),
5207 TGT_SEC_HDL_VAR(0, SEC_CTX_INIT_CONT,mdt_sec_ctx_handle),
5208 TGT_SEC_HDL_VAR(0, SEC_CTX_FINI, mdt_sec_ctx_handle)
5211 static struct tgt_handler mdt_quota_ops[] = {
5212 TGT_QUOTA_HDL(HAS_REPLY, QUOTA_DQACQ, mdt_quota_dqacq),
5215 static struct tgt_opc_slice mdt_common_slice[] = {
5217 .tos_opc_start = MDS_FIRST_OPC,
5218 .tos_opc_end = MDS_LAST_OPC,
5219 .tos_hs = mdt_tgt_handlers
5222 .tos_opc_start = OBD_FIRST_OPC,
5223 .tos_opc_end = OBD_LAST_OPC,
5224 .tos_hs = tgt_obd_handlers
5227 .tos_opc_start = LDLM_FIRST_OPC,
5228 .tos_opc_end = LDLM_LAST_OPC,
5229 .tos_hs = tgt_dlm_handlers
5232 .tos_opc_start = SEC_FIRST_OPC,
5233 .tos_opc_end = SEC_LAST_OPC,
5234 .tos_hs = mdt_sec_ctx_ops
5237 .tos_opc_start = OUT_UPDATE_FIRST_OPC,
5238 .tos_opc_end = OUT_UPDATE_LAST_OPC,
5239 .tos_hs = tgt_out_handlers
5242 .tos_opc_start = FLD_FIRST_OPC,
5243 .tos_opc_end = FLD_LAST_OPC,
5244 .tos_hs = fld_handlers
5247 .tos_opc_start = SEQ_FIRST_OPC,
5248 .tos_opc_end = SEQ_LAST_OPC,
5249 .tos_hs = seq_handlers
5252 .tos_opc_start = QUOTA_DQACQ,
5253 .tos_opc_end = QUOTA_LAST_OPC,
5254 .tos_hs = mdt_quota_ops
5257 .tos_opc_start = LLOG_FIRST_OPC,
5258 .tos_opc_end = LLOG_LAST_OPC,
5259 .tos_hs = tgt_llog_handlers
5262 .tos_opc_start = LFSCK_FIRST_OPC,
5263 .tos_opc_end = LFSCK_LAST_OPC,
5264 .tos_hs = tgt_lfsck_handlers
5267 .tos_opc_start = OST_FIRST_OPC,
5268 .tos_opc_end = OST_LAST_OPC,
5269 .tos_hs = mdt_io_ops
5276 static void mdt_fini(const struct lu_env *env, struct mdt_device *m)
5278 struct md_device *next = m->mdt_child;
5279 struct lu_device *d = &m->mdt_lu_dev;
5280 struct obd_device *obd = mdt2obd_dev(m);
5281 struct lfsck_stop stop;
5284 stop.ls_status = LS_PAUSED;
5286 next->md_ops->mdo_iocontrol(env, next, OBD_IOC_STOP_LFSCK, 0, &stop);
5288 mdt_stack_pre_fini(env, m, md2lu_dev(m->mdt_child));
5289 ping_evictor_stop();
5291 /* Remove the HSM /proc entry so the coordinator cannot be
5292 * restarted by a user while it's shutting down.
5294 mdt_hsm_cdt_stop(m);
5296 mdt_llog_ctxt_unclone(env, m, LLOG_AGENT_ORIG_CTXT);
5297 mdt_llog_ctxt_unclone(env, m, LLOG_CHANGELOG_ORIG_CTXT);
5299 if (m->mdt_namespace != NULL)
5300 ldlm_namespace_free_prior(m->mdt_namespace, NULL,
5301 d->ld_obd->obd_force);
5303 obd_exports_barrier(obd);
5304 obd_zombie_barrier();
5306 mdt_quota_fini(env, m);
5308 cfs_free_nidlist(&m->mdt_squash.rsi_nosquash_nids);
5310 /* Calling the cleanup functions in the same order as in the mdt_init0
5313 mdt_tunables_fini(m);
5315 target_recovery_fini(obd);
5316 upcall_cache_cleanup(m->mdt_identity_cache);
5317 m->mdt_identity_cache = NULL;
5319 mdt_fs_cleanup(env, m);
5321 tgt_fini(env, &m->mdt_lut);
5323 mdt_hsm_cdt_fini(m);
5325 if (m->mdt_los != NULL) {
5326 local_oid_storage_fini(env, m->mdt_los);
5330 if (m->mdt_namespace != NULL) {
5331 ldlm_namespace_free_post(m->mdt_namespace);
5332 d->ld_obd->obd_namespace = m->mdt_namespace = NULL;
5335 if (m->mdt_md_root != NULL) {
5336 mdt_object_put(env, m->mdt_md_root);
5337 m->mdt_md_root = NULL;
5340 mdt_seq_fini(env, m);
5342 mdt_fld_fini(env, m);
5347 mdt_stack_fini(env, m, md2lu_dev(m->mdt_child));
5349 LASSERT(atomic_read(&d->ld_ref) == 0);
5351 server_put_mount(mdt_obd_name(m), true);
5356 static int mdt_postrecov(const struct lu_env *, struct mdt_device *);
5358 static int mdt_init0(const struct lu_env *env, struct mdt_device *m,
5359 struct lu_device_type *ldt, struct lustre_cfg *cfg)
5361 const struct dt_device_param *dt_conf;
5362 struct mdt_thread_info *info;
5363 struct obd_device *obd;
5364 const char *dev = lustre_cfg_string(cfg, 0);
5365 const char *num = lustre_cfg_string(cfg, 2);
5366 struct tg_grants_data *tgd = &m->mdt_lut.lut_tgd;
5367 struct lustre_mount_info *lmi = NULL;
5368 struct lustre_sb_info *lsi;
5370 struct seq_server_site *ss_site;
5371 const char *identity_upcall = "NONE";
5372 struct md_device *next;
5379 lu_device_init(&m->mdt_lu_dev, ldt);
5381 * Environment (env) might be missing mdt_thread_key values at that
5382 * point, if device is allocated when mdt_thread_key is in QUIESCENT
5385 * Usually device allocation path doesn't use module key values, but
5386 * mdt has to do a lot of work here, so allocate key value.
5388 rc = lu_env_refill((struct lu_env *)env);
5392 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
5393 LASSERT(info != NULL);
5395 obd = class_name2obd(dev);
5396 LASSERT(obd != NULL);
5398 m->mdt_max_mdsize = MAX_MD_SIZE_OLD;
5399 m->mdt_opts.mo_evict_tgt_nids = 1;
5400 m->mdt_opts.mo_cos = MDT_COS_DEFAULT;
5402 lmi = server_get_mount(dev);
5404 CERROR("Cannot get mount info for %s!\n", dev);
5407 lsi = s2lsi(lmi->lmi_sb);
5408 LASSERT(lsi->lsi_lmd);
5409 /* CMD is supported only in IAM mode */
5411 node_id = simple_strtol(num, NULL, 10);
5412 obd->u.obt.obt_magic = OBT_MAGIC;
5413 if (lsi->lsi_lmd->lmd_flags & LMD_FLG_SKIP_LFSCK)
5414 m->mdt_skip_lfsck = 1;
5417 /* DoM files get IO lock at open by default */
5418 m->mdt_opts.mo_dom_lock = ALWAYS_DOM_LOCK_ON_OPEN;
5419 /* DoM files are read at open and data is packed in the reply */
5420 m->mdt_opts.mo_dom_read_open = 1;
5422 m->mdt_squash.rsi_uid = 0;
5423 m->mdt_squash.rsi_gid = 0;
5424 INIT_LIST_HEAD(&m->mdt_squash.rsi_nosquash_nids);
5425 spin_lock_init(&m->mdt_squash.rsi_lock);
5426 spin_lock_init(&m->mdt_lock);
5427 m->mdt_enable_remote_dir = 1;
5428 m->mdt_enable_striped_dir = 1;
5429 m->mdt_enable_dir_migration = 1;
5430 m->mdt_enable_remote_dir_gid = 0;
5431 m->mdt_enable_chprojid_gid = 0;
5432 m->mdt_enable_remote_rename = 1;
5434 atomic_set(&m->mdt_mds_mds_conns, 0);
5435 atomic_set(&m->mdt_async_commit_count, 0);
5437 m->mdt_lu_dev.ld_ops = &mdt_lu_ops;
5438 m->mdt_lu_dev.ld_obd = obd;
5439 /* Set this lu_device to obd for error handling purposes. */
5440 obd->obd_lu_dev = &m->mdt_lu_dev;
5442 /* init the stack */
5443 rc = mdt_stack_init((struct lu_env *)env, m, cfg);
5445 CERROR("%s: Can't init device stack, rc %d\n",
5446 mdt_obd_name(m), rc);
5451 ss_site = mdt_seq_site(m);
5452 s->ld_seq_site = ss_site;
5455 /* set server index */
5456 ss_site->ss_node_id = node_id;
5458 /* failover is the default
5459 * FIXME: we do not failout mds0/mgs, which may cause some problems.
5460 * assumed whose ss_node_id == 0 XXX
5462 obd->obd_replayable = 1;
5463 /* No connection accepted until configurations will finish */
5464 obd->obd_no_conn = 1;
5466 if (cfg->lcfg_bufcount > 4 && LUSTRE_CFG_BUFLEN(cfg, 4) > 0) {
5467 char *str = lustre_cfg_string(cfg, 4);
5468 if (strchr(str, 'n')) {
5469 CWARN("%s: recovery disabled\n", mdt_obd_name(m));
5470 obd->obd_replayable = 0;
5474 rc = mdt_fld_init(env, mdt_obd_name(m), m);
5476 GOTO(err_fini_stack, rc);
5478 rc = mdt_seq_init(env, m);
5480 GOTO(err_fini_fld, rc);
5482 snprintf(info->mti_u.ns_name, sizeof(info->mti_u.ns_name), "%s-%s",
5483 LUSTRE_MDT_NAME, obd->obd_uuid.uuid);
5484 m->mdt_namespace = ldlm_namespace_new(obd, info->mti_u.ns_name,
5485 LDLM_NAMESPACE_SERVER,
5486 LDLM_NAMESPACE_GREEDY,
5488 if (m->mdt_namespace == NULL)
5489 GOTO(err_fini_seq, rc = -ENOMEM);
5491 m->mdt_namespace->ns_lvbp = m;
5492 m->mdt_namespace->ns_lvbo = &mdt_lvbo;
5494 ldlm_register_intent(m->mdt_namespace, mdt_intent_policy);
5495 /* set obd_namespace for compatibility with old code */
5496 obd->obd_namespace = m->mdt_namespace;
5498 rc = tgt_init(env, &m->mdt_lut, obd, m->mdt_bottom, mdt_common_slice,
5499 OBD_FAIL_MDS_ALL_REQUEST_NET,
5500 OBD_FAIL_MDS_ALL_REPLY_NET);
5502 GOTO(err_free_ns, rc);
5504 /* Amount of available space excluded from granting and reserved
5505 * for metadata. It is in percentage and 50% is default value. */
5506 tgd->tgd_reserved_pcnt = 50;
5508 if (ONE_MB_BRW_SIZE < (1U << tgd->tgd_blockbits))
5509 m->mdt_brw_size = 1U << tgd->tgd_blockbits;
5511 m->mdt_brw_size = ONE_MB_BRW_SIZE;
5513 rc = mdt_fs_setup(env, m, obd, lsi);
5517 fid.f_seq = FID_SEQ_LOCAL_NAME;
5520 rc = local_oid_storage_init(env, m->mdt_bottom, &fid, &m->mdt_los);
5522 GOTO(err_fs_cleanup, rc);
5524 rc = mdt_hsm_cdt_init(m);
5526 CERROR("%s: error initializing coordinator, rc %d\n",
5527 mdt_obd_name(m), rc);
5528 GOTO(err_los_fini, rc);
5531 tgt_adapt_sptlrpc_conf(&m->mdt_lut);
5533 next = m->mdt_child;
5534 dt_conf = next->md_ops->mdo_dtconf_get(env, next);
5536 mntopts = dt_conf->ddp_mntopts;
5538 if (mntopts & MNTOPT_USERXATTR)
5539 m->mdt_opts.mo_user_xattr = 1;
5541 m->mdt_opts.mo_user_xattr = 0;
5543 m->mdt_max_ea_size = dt_conf->ddp_max_ea_size;
5545 if (mntopts & MNTOPT_ACL)
5546 m->mdt_opts.mo_acl = 1;
5548 m->mdt_opts.mo_acl = 0;
5550 /* XXX: to support suppgid for ACL, we enable identity_upcall
5551 * by default, otherwise, maybe got unexpected -EACCESS. */
5552 if (m->mdt_opts.mo_acl)
5553 identity_upcall = MDT_IDENTITY_UPCALL_PATH;
5555 m->mdt_identity_cache = upcall_cache_init(mdt_obd_name(m),
5557 &mdt_identity_upcall_cache_ops);
5558 if (IS_ERR(m->mdt_identity_cache)) {
5559 rc = PTR_ERR(m->mdt_identity_cache);
5560 m->mdt_identity_cache = NULL;
5561 GOTO(err_free_hsm, rc);
5564 rc = mdt_tunables_init(m, dev);
5566 CERROR("Can't init MDT lprocfs, rc %d\n", rc);
5567 GOTO(err_recovery, rc);
5570 rc = mdt_quota_init(env, m, cfg);
5572 GOTO(err_procfs, rc);
5574 m->mdt_ldlm_client = &mdt2obd_dev(m)->obd_ldlm_client;
5575 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
5576 "mdt_ldlm_client", m->mdt_ldlm_client);
5578 ping_evictor_start();
5580 /* recovery will be started upon mdt_prepare()
5581 * when the whole stack is complete and ready
5582 * to serve the requests */
5584 /* Reduce the initial timeout on an MDS because it doesn't need such
5585 * a long timeout as an OST does. Adaptive timeouts will adjust this
5586 * value appropriately. */
5587 if (ldlm_timeout == LDLM_TIMEOUT_DEFAULT)
5588 ldlm_timeout = MDS_LDLM_TIMEOUT_DEFAULT;
5590 if ((lsi->lsi_lmd->lmd_flags & LMD_FLG_LOCAL_RECOV))
5591 m->mdt_lut.lut_local_recovery = 1;
5595 mdt_tunables_fini(m);
5597 upcall_cache_cleanup(m->mdt_identity_cache);
5598 m->mdt_identity_cache = NULL;
5600 mdt_hsm_cdt_fini(m);
5602 local_oid_storage_fini(env, m->mdt_los);
5605 mdt_fs_cleanup(env, m);
5607 /* keep recoverable clients */
5609 target_recovery_fini(obd);
5610 obd_exports_barrier(obd);
5611 obd_zombie_barrier();
5612 tgt_fini(env, &m->mdt_lut);
5614 ldlm_namespace_free(m->mdt_namespace, NULL, 0);
5615 obd->obd_namespace = m->mdt_namespace = NULL;
5617 mdt_seq_fini(env, m);
5619 mdt_fld_fini(env, m);
5621 mdt_stack_fini(env, m, md2lu_dev(m->mdt_child));
5624 server_put_mount(dev, true);
5628 /* For interoperability, the left element is old parameter, the right one
5629 * is the new version of the parameter, if some parameter is deprecated,
5630 * the new version should be set as NULL. */
5631 static struct cfg_interop_param mdt_interop_param[] = {
5632 { "mdt.group_upcall", NULL },
5633 { "mdt.quota_type", NULL },
5634 { "mdd.quota_type", NULL },
5635 { "mdt.som", NULL },
5636 { "mdt.rootsquash", "mdt.root_squash" },
5637 { "mdt.nosquash_nid", "mdt.nosquash_nids" },
5641 /* used by MGS to process specific configurations */
5642 static int mdt_process_config(const struct lu_env *env,
5643 struct lu_device *d, struct lustre_cfg *cfg)
5645 struct mdt_device *m = mdt_dev(d);
5646 struct md_device *md_next = m->mdt_child;
5647 struct lu_device *next = md2lu_dev(md_next);
5651 switch (cfg->lcfg_command) {
5653 struct obd_device *obd = d->ld_obd;
5654 /* For interoperability */
5655 struct cfg_interop_param *ptr = NULL;
5656 struct lustre_cfg *old_cfg = NULL;
5660 param = lustre_cfg_string(cfg, 1);
5661 if (param == NULL) {
5662 CERROR("param is empty\n");
5667 ptr = class_find_old_param(param, mdt_interop_param);
5669 if (ptr->new_param == NULL) {
5671 CWARN("For interoperability, skip this %s."
5672 " It is obsolete.\n", ptr->old_param);
5676 CWARN("Found old param %s, changed it to %s.\n",
5677 ptr->old_param, ptr->new_param);
5680 cfg = lustre_cfg_rename(old_cfg, ptr->new_param);
5687 count = class_modify_config(cfg, PARAM_MDT,
5688 &obd->obd_kset.kobj);
5690 struct coordinator *cdt = &m->mdt_coordinator;
5692 /* is it an HSM var ? */
5693 count = class_modify_config(cfg, PARAM_HSM,
5694 &cdt->cdt_hsm_kobj);
5696 /* we don't understand; pass it on */
5697 rc = next->ld_ops->ldo_process_config(env, next,
5700 rc = count > 0 ? 0 : count;
5702 rc = count > 0 ? 0 : count;
5706 OBD_FREE(cfg, lustre_cfg_len(cfg->lcfg_bufcount,
5707 cfg->lcfg_buflens));
5711 /* others are passed further */
5712 rc = next->ld_ops->ldo_process_config(env, next, cfg);
5718 static struct lu_object *mdt_object_alloc(const struct lu_env *env,
5719 const struct lu_object_header *hdr,
5720 struct lu_device *d)
5722 struct mdt_object *mo;
5726 OBD_SLAB_ALLOC_PTR_GFP(mo, mdt_object_kmem, GFP_NOFS);
5728 struct lu_object *o;
5729 struct lu_object_header *h;
5732 h = &mo->mot_header;
5733 lu_object_header_init(h);
5734 lu_object_init(o, h, d);
5735 lu_object_add_top(h, o);
5736 o->lo_ops = &mdt_obj_ops;
5737 spin_lock_init(&mo->mot_write_lock);
5738 mutex_init(&mo->mot_som_mutex);
5739 mutex_init(&mo->mot_lov_mutex);
5740 init_rwsem(&mo->mot_dom_sem);
5741 init_rwsem(&mo->mot_open_sem);
5742 atomic_set(&mo->mot_open_count, 0);
5748 static int mdt_object_init(const struct lu_env *env, struct lu_object *o,
5749 const struct lu_object_conf *unused)
5751 struct mdt_device *d = mdt_dev(o->lo_dev);
5752 struct lu_device *under;
5753 struct lu_object *below;
5757 CDEBUG(D_INFO, "object init, fid = "DFID"\n",
5758 PFID(lu_object_fid(o)));
5760 under = &d->mdt_child->md_lu_dev;
5761 below = under->ld_ops->ldo_object_alloc(env, o->lo_header, under);
5762 if (below != NULL) {
5763 lu_object_add(o, below);
5770 static void mdt_object_free_rcu(struct rcu_head *head)
5772 struct mdt_object *mo = container_of(head, struct mdt_object,
5773 mot_header.loh_rcu);
5775 kmem_cache_free(mdt_object_kmem, mo);
5778 static void mdt_object_free(const struct lu_env *env, struct lu_object *o)
5780 struct mdt_object *mo = mdt_obj(o);
5781 struct lu_object_header *h;
5785 CDEBUG(D_INFO, "object free, fid = "DFID"\n",
5786 PFID(lu_object_fid(o)));
5788 LASSERT(atomic_read(&mo->mot_open_count) == 0);
5789 LASSERT(atomic_read(&mo->mot_lease_count) == 0);
5792 lu_object_header_fini(h);
5793 OBD_FREE_PRE(mo, sizeof(*mo), "slab-freed");
5794 call_rcu(&mo->mot_header.loh_rcu, mdt_object_free_rcu);
5799 static int mdt_object_print(const struct lu_env *env, void *cookie,
5800 lu_printer_t p, const struct lu_object *o)
5802 struct mdt_object *mdto = mdt_obj((struct lu_object *)o);
5804 return (*p)(env, cookie,
5805 LUSTRE_MDT_NAME"-object@%p(%s %s, writecount=%d)",
5806 mdto, mdto->mot_lov_created ? "lov_created" : "",
5807 mdto->mot_cache_attr ? "cache_attr" : "",
5808 mdto->mot_write_count);
5811 static int mdt_prepare(const struct lu_env *env,
5812 struct lu_device *pdev,
5813 struct lu_device *cdev)
5815 struct mdt_device *mdt = mdt_dev(cdev);
5816 struct lu_device *next = &mdt->mdt_child->md_lu_dev;
5817 struct obd_device *obd = cdev->ld_obd;
5824 rc = next->ld_ops->ldo_prepare(env, cdev, next);
5828 rc = mdt_llog_ctxt_clone(env, mdt, LLOG_CHANGELOG_ORIG_CTXT);
5832 rc = mdt_llog_ctxt_clone(env, mdt, LLOG_AGENT_ORIG_CTXT);
5836 rc = lfsck_register_namespace(env, mdt->mdt_bottom, mdt->mdt_namespace);
5837 /* The LFSCK instance is registered just now, so it must be there when
5838 * register the namespace to such instance. */
5839 LASSERTF(rc == 0, "register namespace failed: rc = %d\n", rc);
5841 if (mdt->mdt_seq_site.ss_node_id == 0) {
5842 rc = mdt->mdt_child->md_ops->mdo_root_get(env, mdt->mdt_child,
5843 &mdt->mdt_md_root_fid);
5848 LASSERT(!test_bit(MDT_FL_CFGLOG, &mdt->mdt_state));
5850 target_recovery_init(&mdt->mdt_lut, tgt_request_handle);
5851 set_bit(MDT_FL_CFGLOG, &mdt->mdt_state);
5852 LASSERT(obd->obd_no_conn);
5853 spin_lock(&obd->obd_dev_lock);
5854 obd->obd_no_conn = 0;
5855 spin_unlock(&obd->obd_dev_lock);
5857 if (obd->obd_recovering == 0)
5858 mdt_postrecov(env, mdt);
5863 const struct lu_device_operations mdt_lu_ops = {
5864 .ldo_object_alloc = mdt_object_alloc,
5865 .ldo_process_config = mdt_process_config,
5866 .ldo_prepare = mdt_prepare,
5869 static const struct lu_object_operations mdt_obj_ops = {
5870 .loo_object_init = mdt_object_init,
5871 .loo_object_free = mdt_object_free,
5872 .loo_object_print = mdt_object_print
5875 static int mdt_obd_set_info_async(const struct lu_env *env,
5876 struct obd_export *exp,
5877 __u32 keylen, void *key,
5878 __u32 vallen, void *val,
5879 struct ptlrpc_request_set *set)
5885 if (KEY_IS(KEY_SPTLRPC_CONF)) {
5886 rc = tgt_adapt_sptlrpc_conf(class_exp2tgt(exp));
5894 * Match client and server connection feature flags.
5896 * Compute the compatibility flags for a connection request based on
5897 * features mutually supported by client and server.
5899 * The obd_export::exp_connect_data.ocd_connect_flags field in \a exp
5900 * must not be updated here, otherwise a partially initialized value may
5901 * be exposed. After the connection request is successfully processed,
5902 * the top-level MDT connect request handler atomically updates the export
5903 * connect flags from the obd_connect_data::ocd_connect_flags field of the
5904 * reply. \see mdt_connect().
5906 * Before 2.7.50 clients will send a struct obd_connect_data_v1 rather than a
5907 * full struct obd_connect_data. So care must be taken when accessing fields
5908 * that are not present in struct obd_connect_data_v1. See LU-16.
5910 * \param exp the obd_export associated with this client/target pair
5911 * \param mdt the target device for the connection
5912 * \param data stores data for this connect request
5915 * \retval -EPROTO \a data unexpectedly has zero obd_connect_data::ocd_brw_size
5916 * \retval -EBADE client and server feature requirements are incompatible
5918 static int mdt_connect_internal(const struct lu_env *env,
5919 struct obd_export *exp,
5920 struct mdt_device *mdt,
5921 struct obd_connect_data *data, bool reconnect)
5923 const char *obd_name = mdt_obd_name(mdt);
5924 LASSERT(data != NULL);
5926 data->ocd_connect_flags &= MDT_CONNECT_SUPPORTED;
5928 if (mdt->mdt_bottom->dd_rdonly &&
5929 !(data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
5930 !(data->ocd_connect_flags & OBD_CONNECT_RDONLY))
5933 if (data->ocd_connect_flags & OBD_CONNECT_FLAGS2)
5934 data->ocd_connect_flags2 &= MDT_CONNECT_SUPPORTED2;
5936 data->ocd_ibits_known &= MDS_INODELOCK_FULL;
5938 if (!mdt->mdt_opts.mo_acl)
5939 data->ocd_connect_flags &= ~OBD_CONNECT_ACL;
5941 if (!mdt->mdt_opts.mo_user_xattr)
5942 data->ocd_connect_flags &= ~OBD_CONNECT_XATTR;
5944 if (OCD_HAS_FLAG(data, BRW_SIZE)) {
5945 data->ocd_brw_size = min(data->ocd_brw_size,
5947 if (data->ocd_brw_size == 0) {
5948 CERROR("%s: cli %s/%p ocd_connect_flags: %#llx "
5949 "ocd_version: %x ocd_grant: %d ocd_index: %u "
5950 "ocd_brw_size unexpectedly zero, network data "
5951 "corruption? Refusing to connect this client\n",
5952 obd_name, exp->exp_client_uuid.uuid,
5953 exp, data->ocd_connect_flags, data->ocd_version,
5954 data->ocd_grant, data->ocd_index);
5959 if (OCD_HAS_FLAG(data, GRANT_PARAM)) {
5960 struct dt_device_param *ddp = &mdt->mdt_lut.lut_dt_conf;
5962 /* client is reporting its page size, for future use */
5963 exp->exp_target_data.ted_pagebits = data->ocd_grant_blkbits;
5964 data->ocd_grant_blkbits = mdt->mdt_lut.lut_tgd.tgd_blockbits;
5965 /* ddp_inodespace may not be power-of-two value, eg. for ldiskfs
5966 * it's LDISKFS_DIR_REC_LEN(20) = 28. */
5967 data->ocd_grant_inobits = fls(ddp->ddp_inodespace - 1);
5968 /* ocd_grant_tax_kb is in 1K byte blocks */
5969 data->ocd_grant_tax_kb = ddp->ddp_extent_tax >> 10;
5970 data->ocd_grant_max_blks = ddp->ddp_max_extent_blks;
5973 /* Save connect_data we have so far because tgt_grant_connect()
5974 * uses it to calculate grant, and we want to save the client
5975 * version before it is overwritten by LUSTRE_VERSION_CODE. */
5976 exp->exp_connect_data = *data;
5977 if (OCD_HAS_FLAG(data, GRANT))
5978 tgt_grant_connect(env, exp, data, !reconnect);
5980 if (OCD_HAS_FLAG(data, MAXBYTES))
5981 data->ocd_maxbytes = mdt->mdt_lut.lut_dt_conf.ddp_maxbytes;
5983 /* NB: Disregard the rule against updating
5984 * exp_connect_data.ocd_connect_flags in this case, since
5985 * tgt_client_new() needs to know if this is a lightweight
5986 * connection, and it is safe to expose this flag before
5987 * connection processing completes. */
5988 if (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) {
5989 spin_lock(&exp->exp_lock);
5990 *exp_connect_flags_ptr(exp) |= OBD_CONNECT_LIGHTWEIGHT;
5991 spin_unlock(&exp->exp_lock);
5994 data->ocd_version = LUSTRE_VERSION_CODE;
5996 if ((data->ocd_connect_flags & OBD_CONNECT_FID) == 0) {
5997 CWARN("%s: MDS requires FID support, but client not\n",
6002 if (OCD_HAS_FLAG(data, PINGLESS)) {
6003 if (ptlrpc_pinger_suppress_pings()) {
6004 spin_lock(&exp->exp_obd->obd_dev_lock);
6005 list_del_init(&exp->exp_obd_chain_timed);
6006 spin_unlock(&exp->exp_obd->obd_dev_lock);
6008 data->ocd_connect_flags &= ~OBD_CONNECT_PINGLESS;
6012 data->ocd_max_easize = mdt->mdt_max_ea_size;
6014 /* NB: Disregard the rule against updating
6015 * exp_connect_data.ocd_connect_flags in this case, since
6016 * tgt_client_new() needs to know if this is client supports
6017 * multiple modify RPCs, and it is safe to expose this flag before
6018 * connection processing completes. */
6019 if (data->ocd_connect_flags & OBD_CONNECT_MULTIMODRPCS) {
6020 data->ocd_maxmodrpcs = max_mod_rpcs_per_client;
6021 spin_lock(&exp->exp_lock);
6022 *exp_connect_flags_ptr(exp) |= OBD_CONNECT_MULTIMODRPCS;
6023 spin_unlock(&exp->exp_lock);
6026 if (OCD_HAS_FLAG(data, CKSUM)) {
6027 __u32 cksum_types = data->ocd_cksum_types;
6029 /* The client set in ocd_cksum_types the checksum types it
6030 * supports. We have to mask off the algorithms that we don't
6032 data->ocd_cksum_types &=
6033 obd_cksum_types_supported_server(obd_name);
6035 if (unlikely(data->ocd_cksum_types == 0)) {
6036 CERROR("%s: Connect with checksum support but no "
6037 "ocd_cksum_types is set\n",
6038 exp->exp_obd->obd_name);
6042 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
6043 "%x\n", exp->exp_obd->obd_name, obd_export_nid2str(exp),
6044 cksum_types, data->ocd_cksum_types);
6046 /* This client does not support OBD_CONNECT_CKSUM
6047 * fall back to CRC32 */
6048 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
6049 "OBD_CONNECT_CKSUM, CRC32 will be used\n",
6050 exp->exp_obd->obd_name, obd_export_nid2str(exp));
6056 static int mdt_ctxt_add_dirty_flag(struct lu_env *env,
6057 struct mdt_thread_info *info,
6058 struct mdt_file_data *mfd)
6060 struct lu_context ses;
6064 rc = lu_context_init(&ses, LCT_SERVER_SESSION);
6069 lu_context_enter(&ses);
6071 mdt_ucred(info)->uc_valid = UCRED_OLD;
6072 rc = mdt_add_dirty_flag(info, mfd->mfd_object, &info->mti_attr);
6074 lu_context_exit(&ses);
6075 lu_context_fini(&ses);
6081 static int mdt_export_cleanup(struct obd_export *exp)
6083 LIST_HEAD(closing_list);
6084 struct mdt_export_data *med = &exp->exp_mdt_data;
6085 struct obd_device *obd = exp->exp_obd;
6086 struct mdt_device *mdt;
6087 struct mdt_thread_info *info;
6089 struct mdt_file_data *mfd, *n;
6093 spin_lock(&med->med_open_lock);
6094 while (!list_empty(&med->med_open_head)) {
6095 struct list_head *tmp = med->med_open_head.next;
6096 mfd = list_entry(tmp, struct mdt_file_data, mfd_list);
6098 /* Remove mfd handle so it can't be found again.
6099 * We are consuming the mfd_list reference here. */
6100 class_handle_unhash(&mfd->mfd_open_handle);
6101 list_move_tail(&mfd->mfd_list, &closing_list);
6103 spin_unlock(&med->med_open_lock);
6104 mdt = mdt_dev(obd->obd_lu_dev);
6105 LASSERT(mdt != NULL);
6107 rc = lu_env_init(&env, LCT_MD_THREAD);
6111 info = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
6112 LASSERT(info != NULL);
6113 memset(info, 0, sizeof *info);
6114 info->mti_env = &env;
6115 info->mti_mdt = mdt;
6116 info->mti_exp = exp;
6118 if (!list_empty(&closing_list)) {
6119 struct md_attr *ma = &info->mti_attr;
6121 /* Close any open files (which may also cause orphan
6123 list_for_each_entry_safe(mfd, n, &closing_list, mfd_list) {
6124 list_del_init(&mfd->mfd_list);
6125 ma->ma_need = ma->ma_valid = 0;
6127 /* This file is being closed due to an eviction, it
6128 * could have been modified and now dirty regarding to
6129 * HSM archive, check this!
6130 * The logic here is to mark a file dirty if there's a
6131 * chance it was dirtied before the client was evicted,
6132 * so that we don't have to wait for a release attempt
6133 * before finding out the file was actually dirty and
6134 * fail the release. Aggressively marking it dirty here
6135 * will cause the policy engine to attempt to
6136 * re-archive it; when rearchiving, we can compare the
6137 * current version to the HSM data_version and make the
6138 * archive request into a noop if it's not actually
6141 if (mfd->mfd_open_flags & MDS_FMODE_WRITE)
6142 rc = mdt_ctxt_add_dirty_flag(&env, info, mfd);
6144 /* Don't unlink orphan on failover umount, LU-184 */
6145 if (exp->exp_flags & OBD_OPT_FAILOVER) {
6146 ma->ma_valid = MA_FLAGS;
6147 ma->ma_attr_flags |= MDS_KEEP_ORPHAN;
6149 ma->ma_valid |= MA_FORCE_LOG;
6150 mdt_mfd_close(info, mfd);
6153 info->mti_mdt = NULL;
6154 /* cleanup client slot early */
6155 /* Do not erase record for recoverable client. */
6156 if (!(exp->exp_flags & OBD_OPT_FAILOVER) || exp->exp_failed)
6157 tgt_client_del(&env, exp);
6163 static inline void mdt_enable_slc(struct mdt_device *mdt)
6165 if (mdt->mdt_lut.lut_sync_lock_cancel == SYNC_LOCK_CANCEL_NEVER)
6166 mdt->mdt_lut.lut_sync_lock_cancel = SYNC_LOCK_CANCEL_BLOCKING;
6169 static inline void mdt_disable_slc(struct mdt_device *mdt)
6171 if (mdt->mdt_lut.lut_sync_lock_cancel == SYNC_LOCK_CANCEL_BLOCKING)
6172 mdt->mdt_lut.lut_sync_lock_cancel = SYNC_LOCK_CANCEL_NEVER;
6175 static int mdt_obd_disconnect(struct obd_export *exp)
6182 class_export_get(exp);
6184 if (!(exp->exp_flags & OBD_OPT_FORCE))
6185 tgt_grant_sanity_check(exp->exp_obd, __func__);
6187 if ((exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) &&
6188 !(exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)) {
6189 struct mdt_device *mdt = mdt_dev(exp->exp_obd->obd_lu_dev);
6191 if (atomic_dec_and_test(&mdt->mdt_mds_mds_conns))
6192 mdt_disable_slc(mdt);
6195 rc = server_disconnect_export(exp);
6197 CDEBUG(D_IOCTL, "server disconnect error: rc = %d\n", rc);
6199 tgt_grant_discard(exp);
6201 rc = mdt_export_cleanup(exp);
6202 nodemap_del_member(exp);
6203 class_export_put(exp);
6207 /* mds_connect copy */
6208 static int mdt_obd_connect(const struct lu_env *env,
6209 struct obd_export **exp, struct obd_device *obd,
6210 struct obd_uuid *cluuid,
6211 struct obd_connect_data *data,
6214 struct obd_export *lexp;
6215 struct lustre_handle conn = { 0 };
6216 struct mdt_device *mdt;
6218 lnet_nid_t *client_nid = localdata;
6221 LASSERT(env != NULL);
6222 LASSERT(data != NULL);
6224 if (!exp || !obd || !cluuid)
6227 mdt = mdt_dev(obd->obd_lu_dev);
6229 if ((data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
6230 !(data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT)) {
6231 atomic_inc(&mdt->mdt_mds_mds_conns);
6232 mdt_enable_slc(mdt);
6236 * first, check whether the stack is ready to handle requests
6237 * XXX: probably not very appropriate method is used now
6238 * at some point we should find a better one
6240 if (!test_bit(MDT_FL_SYNCED, &mdt->mdt_state) &&
6241 !(data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) &&
6242 !(data->ocd_connect_flags & OBD_CONNECT_MDS_MDS)) {
6243 rc = obd_get_info(env, mdt->mdt_child_exp,
6244 sizeof(KEY_OSP_CONNECTED),
6245 KEY_OSP_CONNECTED, NULL, NULL);
6248 set_bit(MDT_FL_SYNCED, &mdt->mdt_state);
6251 rc = class_connect(&conn, obd, cluuid);
6255 lexp = class_conn2export(&conn);
6256 LASSERT(lexp != NULL);
6258 rc = nodemap_add_member(*client_nid, lexp);
6259 if (rc != 0 && rc != -EEXIST)
6262 rc = mdt_connect_internal(env, lexp, mdt, data, false);
6264 struct lsd_client_data *lcd = lexp->exp_target_data.ted_lcd;
6267 memcpy(lcd->lcd_uuid, cluuid, sizeof lcd->lcd_uuid);
6268 rc = tgt_client_new(env, lexp);
6270 mdt_export_stats_init(obd, lexp, localdata);
6274 class_disconnect(lexp);
6275 nodemap_del_member(lexp);
6279 /* Because we do not want this export to be evicted by pinger,
6280 * let's not add this export to the timed chain list. */
6281 if (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) {
6282 spin_lock(&lexp->exp_obd->obd_dev_lock);
6283 list_del_init(&lexp->exp_obd_chain_timed);
6284 spin_unlock(&lexp->exp_obd->obd_dev_lock);
6291 static int mdt_obd_reconnect(const struct lu_env *env,
6292 struct obd_export *exp, struct obd_device *obd,
6293 struct obd_uuid *cluuid,
6294 struct obd_connect_data *data,
6297 lnet_nid_t *client_nid = localdata;
6301 if (exp == NULL || obd == NULL || cluuid == NULL)
6304 rc = nodemap_add_member(*client_nid, exp);
6305 if (rc != 0 && rc != -EEXIST)
6308 rc = mdt_connect_internal(env, exp, mdt_dev(obd->obd_lu_dev), data,
6311 mdt_export_stats_init(obd, exp, localdata);
6313 nodemap_del_member(exp);
6318 /* FIXME: Can we avoid using these two interfaces? */
6319 static int mdt_init_export(struct obd_export *exp)
6321 struct mdt_export_data *med = &exp->exp_mdt_data;
6325 INIT_LIST_HEAD(&med->med_open_head);
6326 spin_lock_init(&med->med_open_lock);
6327 spin_lock(&exp->exp_lock);
6328 exp->exp_connecting = 1;
6329 spin_unlock(&exp->exp_lock);
6331 OBD_ALLOC(exp->exp_used_slots,
6332 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
6333 if (exp->exp_used_slots == NULL)
6336 /* self-export doesn't need client data and ldlm initialization */
6337 if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
6338 &exp->exp_client_uuid)))
6341 rc = tgt_client_alloc(exp);
6345 rc = ldlm_init_export(exp);
6352 tgt_client_free(exp);
6354 OBD_FREE(exp->exp_used_slots,
6355 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
6356 exp->exp_used_slots = NULL;
6358 CERROR("%s: Failed to initialize export: rc = %d\n",
6359 exp->exp_obd->obd_name, rc);
6363 static int mdt_destroy_export(struct obd_export *exp)
6367 target_destroy_export(exp);
6368 if (exp->exp_used_slots)
6369 OBD_FREE(exp->exp_used_slots,
6370 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
6372 /* destroy can be called from failed obd_setup, so
6373 * checking uuid is safer than obd_self_export */
6374 if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
6375 &exp->exp_client_uuid)))
6378 ldlm_destroy_export(exp);
6379 tgt_client_free(exp);
6381 LASSERT(list_empty(&exp->exp_outstanding_replies));
6382 LASSERT(list_empty(&exp->exp_mdt_data.med_open_head));
6385 * discard grants once we're sure no more
6386 * interaction with the client is possible
6388 tgt_grant_discard(exp);
6389 if (exp_connect_flags(exp) & OBD_CONNECT_GRANT)
6390 exp->exp_obd->u.obt.obt_lut->lut_tgd.tgd_tot_granted_clients--;
6392 if (!(exp->exp_flags & OBD_OPT_FORCE))
6393 tgt_grant_sanity_check(exp->exp_obd, __func__);
6398 int mdt_links_read(struct mdt_thread_info *info, struct mdt_object *mdt_obj,
6399 struct linkea_data *ldata)
6403 LASSERT(ldata->ld_buf->lb_buf != NULL);
6405 if (!mdt_object_exists(mdt_obj))
6408 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
6409 ldata->ld_buf, XATTR_NAME_LINK);
6410 if (rc == -ERANGE) {
6411 /* Buf was too small, figure out what we need. */
6412 lu_buf_free(ldata->ld_buf);
6413 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
6414 ldata->ld_buf, XATTR_NAME_LINK);
6417 ldata->ld_buf = lu_buf_check_and_alloc(ldata->ld_buf, rc);
6418 if (ldata->ld_buf->lb_buf == NULL)
6420 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
6421 ldata->ld_buf, XATTR_NAME_LINK);
6426 return linkea_init_with_rec(ldata);
6430 * Given an MDT object, try to look up the full path to the object.
6431 * Part of the MDT layer implementation of lfs fid2path.
6433 * \param[in] info Per-thread common data shared by MDT level handlers.
6434 * \param[in] obj Object to do path lookup of
6435 * \param[in,out] fp User-provided struct to store path information
6436 * \param[in] root_fid Root FID of current path should reach
6438 * \retval 0 Lookup successful, path information stored in fp
6439 * \retval -EAGAIN Lookup failed, usually because object is being moved
6440 * \retval negative errno if there was a problem
6442 static int mdt_path_current(struct mdt_thread_info *info,
6443 struct mdt_object *obj,
6444 struct getinfo_fid2path *fp,
6445 struct lu_fid *root_fid)
6447 struct mdt_device *mdt = info->mti_mdt;
6448 struct mdt_object *mdt_obj;
6449 struct link_ea_header *leh;
6450 struct link_ea_entry *lee;
6451 struct lu_name *tmpname = &info->mti_name;
6452 struct lu_fid *tmpfid = &info->mti_tmp_fid1;
6453 struct lu_buf *buf = &info->mti_big_buf;
6456 struct linkea_data ldata = { NULL };
6461 /* temp buffer for path element, the buffer will be finally freed
6462 * in mdt_thread_info_fini */
6463 buf = lu_buf_check_and_alloc(buf, PATH_MAX);
6464 if (buf->lb_buf == NULL)
6468 ptr = fp->gf_u.gf_path + fp->gf_pathlen - 1;
6471 *tmpfid = fp->gf_fid = *mdt_object_fid(obj);
6473 while (!lu_fid_eq(root_fid, &fp->gf_fid)) {
6474 struct lu_buf lmv_buf;
6476 if (!lu_fid_eq(root_fid, &mdt->mdt_md_root_fid) &&
6477 lu_fid_eq(&mdt->mdt_md_root_fid, &fp->gf_fid))
6478 GOTO(out, rc = -ENOENT);
6480 if (lu_fid_eq(mdt_object_fid(obj), tmpfid)) {
6482 mdt_object_get(info->mti_env, mdt_obj);
6484 mdt_obj = mdt_object_find(info->mti_env, mdt, tmpfid);
6485 if (IS_ERR(mdt_obj))
6486 GOTO(out, rc = PTR_ERR(mdt_obj));
6489 if (!mdt_object_exists(mdt_obj)) {
6490 mdt_object_put(info->mti_env, mdt_obj);
6491 GOTO(out, rc = -ENOENT);
6494 if (mdt_object_remote(mdt_obj)) {
6495 mdt_object_put(info->mti_env, mdt_obj);
6496 GOTO(remote_out, rc = -EREMOTE);
6499 rc = mdt_links_read(info, mdt_obj, &ldata);
6501 mdt_object_put(info->mti_env, mdt_obj);
6506 lee = (struct link_ea_entry *)(leh + 1); /* link #0 */
6507 linkea_entry_unpack(lee, &reclen, tmpname, tmpfid);
6508 /* If set, use link #linkno for path lookup, otherwise use
6509 link #0. Only do this for the final path element. */
6510 if (first && fp->gf_linkno < leh->leh_reccount) {
6512 for (count = 0; count < fp->gf_linkno; count++) {
6513 lee = (struct link_ea_entry *)
6514 ((char *)lee + reclen);
6515 linkea_entry_unpack(lee, &reclen, tmpname,
6518 if (fp->gf_linkno < leh->leh_reccount - 1)
6519 /* indicate to user there are more links */
6523 lmv_buf.lb_buf = info->mti_xattr_buf;
6524 lmv_buf.lb_len = sizeof(info->mti_xattr_buf);
6525 /* Check if it is slave stripes */
6526 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
6527 &lmv_buf, XATTR_NAME_LMV);
6528 mdt_object_put(info->mti_env, mdt_obj);
6530 union lmv_mds_md *lmm = lmv_buf.lb_buf;
6532 /* For slave stripes, get its master */
6533 if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_STRIPE) {
6534 fp->gf_fid = *tmpfid;
6537 } else if (rc < 0 && rc != -ENODATA) {
6543 /* Pack the name in the end of the buffer */
6544 ptr -= tmpname->ln_namelen;
6545 if (ptr - 1 <= fp->gf_u.gf_path)
6546 GOTO(out, rc = -EOVERFLOW);
6547 strncpy(ptr, tmpname->ln_name, tmpname->ln_namelen);
6550 /* keep the last resolved fid to the client, so the
6551 * client will build the left path on another MDT for
6553 fp->gf_fid = *tmpfid;
6559 ptr++; /* skip leading / */
6560 memmove(fp->gf_u.gf_path, ptr,
6561 fp->gf_u.gf_path + fp->gf_pathlen - ptr);
6568 * Given an MDT object, use mdt_path_current to get the path.
6569 * Essentially a wrapper to retry mdt_path_current a set number of times
6570 * if -EAGAIN is returned (usually because an object is being moved).
6572 * Part of the MDT layer implementation of lfs fid2path.
6574 * \param[in] info Per-thread common data shared by mdt level handlers.
6575 * \param[in] obj Object to do path lookup of
6576 * \param[in,out] fp User-provided struct for arguments and to store path
6579 * \retval 0 Lookup successful, path information stored in fp
6580 * \retval negative errno if there was a problem
6582 static int mdt_path(struct mdt_thread_info *info, struct mdt_object *obj,
6583 struct getinfo_fid2path *fp, struct lu_fid *root_fid)
6585 struct mdt_device *mdt = info->mti_mdt;
6590 if (fp->gf_pathlen < 3)
6593 if (root_fid == NULL)
6594 root_fid = &mdt->mdt_md_root_fid;
6596 if (lu_fid_eq(root_fid, mdt_object_fid(obj))) {
6597 fp->gf_u.gf_path[0] = '\0';
6601 /* Retry multiple times in case file is being moved */
6602 while (tries-- && rc == -EAGAIN)
6603 rc = mdt_path_current(info, obj, fp, root_fid);
6609 * Get the full path of the provided FID, as of changelog record recno.
6611 * This checks sanity and looks up object for user provided FID
6612 * before calling the actual path lookup code.
6614 * Part of the MDT layer implementation of lfs fid2path.
6616 * \param[in] info Per-thread common data shared by mdt level handlers.
6617 * \param[in,out] fp User-provided struct for arguments and to store path
6620 * \retval 0 Lookup successful, path information and recno stored in fp
6621 * \retval -ENOENT, object does not exist
6622 * \retval negative errno if there was a problem
6624 static int mdt_fid2path(struct mdt_thread_info *info,
6625 struct lu_fid *root_fid,
6626 struct getinfo_fid2path *fp)
6628 struct mdt_device *mdt = info->mti_mdt;
6629 struct mdt_object *obj;
6633 CDEBUG(D_IOCTL, "path get "DFID" from %llu #%d\n",
6634 PFID(&fp->gf_fid), fp->gf_recno, fp->gf_linkno);
6636 if (!fid_is_sane(&fp->gf_fid))
6639 if (!fid_is_namespace_visible(&fp->gf_fid)) {
6640 CDEBUG(D_INFO, "%s: "DFID" is invalid, f_seq should be >= %#llx"
6641 ", or f_oid != 0, or f_ver == 0\n", mdt_obd_name(mdt),
6642 PFID(&fp->gf_fid), (__u64)FID_SEQ_NORMAL);
6646 obj = mdt_object_find(info->mti_env, mdt, &fp->gf_fid);
6649 CDEBUG(D_IOCTL, "cannot find "DFID": rc = %d\n",
6650 PFID(&fp->gf_fid), rc);
6654 if (mdt_object_remote(obj))
6656 else if (!mdt_object_exists(obj))
6662 mdt_object_put(info->mti_env, obj);
6663 CDEBUG(D_IOCTL, "nonlocal object "DFID": rc = %d\n",
6664 PFID(&fp->gf_fid), rc);
6668 rc = mdt_path(info, obj, fp, root_fid);
6670 CDEBUG(D_INFO, "fid "DFID", path %s recno %#llx linkno %u\n",
6671 PFID(&fp->gf_fid), fp->gf_u.gf_path,
6672 fp->gf_recno, fp->gf_linkno);
6674 mdt_object_put(info->mti_env, obj);
6679 static int mdt_rpc_fid2path(struct mdt_thread_info *info, void *key, int keylen,
6680 void *val, int vallen)
6682 struct getinfo_fid2path *fpout, *fpin;
6683 struct lu_fid *root_fid = NULL;
6686 fpin = key + cfs_size_round(sizeof(KEY_FID2PATH));
6689 if (ptlrpc_req_need_swab(info->mti_pill->rc_req))
6690 lustre_swab_fid2path(fpin);
6692 memcpy(fpout, fpin, sizeof(*fpin));
6693 if (fpout->gf_pathlen != vallen - sizeof(*fpin))
6696 if (keylen >= cfs_size_round(sizeof(KEY_FID2PATH)) + sizeof(*fpin) +
6697 sizeof(struct lu_fid)) {
6698 /* client sent its root FID, which is normally fileset FID */
6699 root_fid = fpin->gf_u.gf_root_fid;
6700 if (ptlrpc_req_need_swab(info->mti_pill->rc_req))
6701 lustre_swab_lu_fid(root_fid);
6703 if (root_fid != NULL && !fid_is_sane(root_fid))
6707 rc = mdt_fid2path(info, root_fid, fpout);
6711 int mdt_get_info(struct tgt_session_info *tsi)
6721 key = req_capsule_client_get(tsi->tsi_pill, &RMF_GETINFO_KEY);
6723 CDEBUG(D_IOCTL, "No GETINFO key\n");
6724 RETURN(err_serious(-EFAULT));
6726 keylen = req_capsule_get_size(tsi->tsi_pill, &RMF_GETINFO_KEY,
6729 vallen = req_capsule_client_get(tsi->tsi_pill, &RMF_GETINFO_VALLEN);
6730 if (vallen == NULL) {
6731 CDEBUG(D_IOCTL, "%s: cannot get RMF_GETINFO_VALLEN buffer\n",
6732 tgt_name(tsi->tsi_tgt));
6733 RETURN(err_serious(-EFAULT));
6736 req_capsule_set_size(tsi->tsi_pill, &RMF_GETINFO_VAL, RCL_SERVER,
6738 rc = req_capsule_server_pack(tsi->tsi_pill);
6740 RETURN(err_serious(rc));
6742 valout = req_capsule_server_get(tsi->tsi_pill, &RMF_GETINFO_VAL);
6743 if (valout == NULL) {
6744 CDEBUG(D_IOCTL, "%s: cannot get get-info RPC out buffer\n",
6745 tgt_name(tsi->tsi_tgt));
6746 RETURN(err_serious(-EFAULT));
6749 if (KEY_IS(KEY_FID2PATH)) {
6750 struct mdt_thread_info *info = tsi2mdt_info(tsi);
6752 rc = mdt_rpc_fid2path(info, key, keylen, valout, *vallen);
6753 mdt_thread_info_fini(info);
6760 static int mdt_ioc_version_get(struct mdt_thread_info *mti, void *karg)
6762 struct obd_ioctl_data *data = karg;
6765 struct mdt_object *obj;
6766 struct mdt_lock_handle *lh;
6770 if (data->ioc_inlbuf1 == NULL || data->ioc_inllen1 != sizeof(*fid) ||
6771 data->ioc_inlbuf2 == NULL || data->ioc_inllen2 != sizeof(version))
6774 fid = (struct lu_fid *)data->ioc_inlbuf1;
6776 if (!fid_is_sane(fid))
6779 CDEBUG(D_IOCTL, "getting version for "DFID"\n", PFID(fid));
6781 lh = &mti->mti_lh[MDT_LH_PARENT];
6782 mdt_lock_reg_init(lh, LCK_CR);
6784 obj = mdt_object_find_lock(mti, fid, lh, MDS_INODELOCK_UPDATE);
6786 RETURN(PTR_ERR(obj));
6788 if (mdt_object_remote(obj)) {
6791 * before calling version get the correct MDS should be
6792 * fid, this is error to find remote object here
6794 CERROR("nonlocal object "DFID"\n", PFID(fid));
6795 } else if (!mdt_object_exists(obj)) {
6796 *(__u64 *)data->ioc_inlbuf2 = ENOENT_VERSION;
6799 version = dt_version_get(mti->mti_env, mdt_obj2dt(obj));
6800 *(__u64 *)data->ioc_inlbuf2 = version;
6803 mdt_object_unlock_put(mti, obj, lh, 1);
6807 /* ioctls on obd dev */
6808 static int mdt_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
6809 void *karg, void __user *uarg)
6812 struct obd_device *obd = exp->exp_obd;
6813 struct mdt_device *mdt = mdt_dev(obd->obd_lu_dev);
6814 struct dt_device *dt = mdt->mdt_bottom;
6818 CDEBUG(D_IOCTL, "handling ioctl cmd %#x\n", cmd);
6819 rc = lu_env_init(&env, LCT_MD_THREAD);
6825 rc = mdt_device_sync(&env, mdt);
6827 case OBD_IOC_SET_READONLY:
6828 rc = dt_sync(&env, dt);
6830 rc = dt_ro(&env, dt);
6832 case OBD_IOC_ABORT_RECOVERY:
6833 CERROR("%s: Aborting recovery for device\n", mdt_obd_name(mdt));
6834 obd->obd_abort_recovery = 1;
6835 target_stop_recovery_thread(obd);
6838 case OBD_IOC_CHANGELOG_REG:
6839 case OBD_IOC_CHANGELOG_DEREG:
6840 case OBD_IOC_CHANGELOG_CLEAR:
6841 rc = mdt->mdt_child->md_ops->mdo_iocontrol(&env,
6845 case OBD_IOC_START_LFSCK: {
6846 struct md_device *next = mdt->mdt_child;
6847 struct obd_ioctl_data *data = karg;
6848 struct lfsck_start_param lsp;
6850 if (unlikely(data == NULL)) {
6855 lsp.lsp_start = (struct lfsck_start *)(data->ioc_inlbuf1);
6856 lsp.lsp_index_valid = 0;
6857 rc = next->md_ops->mdo_iocontrol(&env, next, cmd, 0, &lsp);
6860 case OBD_IOC_STOP_LFSCK: {
6861 struct md_device *next = mdt->mdt_child;
6862 struct obd_ioctl_data *data = karg;
6863 struct lfsck_stop stop;
6865 stop.ls_status = LS_STOPPED;
6866 /* Old lfsck utils may pass NULL @stop. */
6867 if (data->ioc_inlbuf1 == NULL)
6871 ((struct lfsck_stop *)(data->ioc_inlbuf1))->ls_flags;
6873 rc = next->md_ops->mdo_iocontrol(&env, next, cmd, 0, &stop);
6876 case OBD_IOC_QUERY_LFSCK: {
6877 struct md_device *next = mdt->mdt_child;
6878 struct obd_ioctl_data *data = karg;
6880 rc = next->md_ops->mdo_iocontrol(&env, next, cmd, 0,
6884 case OBD_IOC_GET_OBJ_VERSION: {
6885 struct mdt_thread_info *mti;
6886 mti = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
6887 memset(mti, 0, sizeof *mti);
6888 mti->mti_env = &env;
6892 rc = mdt_ioc_version_get(mti, karg);
6895 case OBD_IOC_CATLOGLIST: {
6896 struct mdt_thread_info *mti;
6898 mti = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
6899 lu_local_obj_fid(&mti->mti_tmp_fid1, LLOG_CATALOGS_OID);
6900 rc = llog_catalog_list(&env, mdt->mdt_bottom, 0, karg,
6901 &mti->mti_tmp_fid1);
6906 CERROR("%s: Not supported cmd = %d, rc = %d\n",
6907 mdt_obd_name(mdt), cmd, rc);
6914 static int mdt_postrecov(const struct lu_env *env, struct mdt_device *mdt)
6916 struct lu_device *ld = md2lu_dev(mdt->mdt_child);
6920 if (!mdt->mdt_skip_lfsck && !mdt->mdt_bottom->dd_rdonly) {
6921 struct lfsck_start_param lsp;
6923 lsp.lsp_start = NULL;
6924 lsp.lsp_index_valid = 0;
6925 rc = mdt->mdt_child->md_ops->mdo_iocontrol(env, mdt->mdt_child,
6926 OBD_IOC_START_LFSCK,
6928 if (rc != 0 && rc != -EALREADY)
6929 CWARN("%s: auto trigger paused LFSCK failed: rc = %d\n",
6930 mdt_obd_name(mdt), rc);
6933 rc = ld->ld_ops->ldo_recovery_complete(env, ld);
6937 static int mdt_obd_postrecov(struct obd_device *obd)
6942 rc = lu_env_init(&env, LCT_MD_THREAD);
6945 rc = mdt_postrecov(&env, mdt_dev(obd->obd_lu_dev));
6950 static const struct obd_ops mdt_obd_device_ops = {
6951 .o_owner = THIS_MODULE,
6952 .o_set_info_async = mdt_obd_set_info_async,
6953 .o_connect = mdt_obd_connect,
6954 .o_reconnect = mdt_obd_reconnect,
6955 .o_disconnect = mdt_obd_disconnect,
6956 .o_init_export = mdt_init_export,
6957 .o_destroy_export = mdt_destroy_export,
6958 .o_iocontrol = mdt_iocontrol,
6959 .o_postrecov = mdt_obd_postrecov,
6960 /* Data-on-MDT IO methods */
6961 .o_preprw = mdt_obd_preprw,
6962 .o_commitrw = mdt_obd_commitrw,
6965 static struct lu_device* mdt_device_fini(const struct lu_env *env,
6966 struct lu_device *d)
6968 struct mdt_device *m = mdt_dev(d);
6975 static struct lu_device *mdt_device_free(const struct lu_env *env,
6976 struct lu_device *d)
6978 struct mdt_device *m = mdt_dev(d);
6981 lu_device_fini(&m->mdt_lu_dev);
6987 static struct lu_device *mdt_device_alloc(const struct lu_env *env,
6988 struct lu_device_type *t,
6989 struct lustre_cfg *cfg)
6991 struct lu_device *l;
6992 struct mdt_device *m;
6999 rc = mdt_init0(env, m, t, cfg);
7001 mdt_device_free(env, l);
7006 l = ERR_PTR(-ENOMEM);
7010 /* context key constructor/destructor: mdt_key_init, mdt_key_fini */
7011 LU_KEY_INIT(mdt, struct mdt_thread_info);
7013 static void mdt_key_fini(const struct lu_context *ctx,
7014 struct lu_context_key *key, void* data)
7016 struct mdt_thread_info *info = data;
7018 if (info->mti_big_lmm) {
7019 OBD_FREE_LARGE(info->mti_big_lmm, info->mti_big_lmmsize);
7020 info->mti_big_lmm = NULL;
7021 info->mti_big_lmmsize = 0;
7024 if (info->mti_big_acl) {
7025 OBD_FREE_LARGE(info->mti_big_acl, info->mti_big_aclsize);
7026 info->mti_big_acl = NULL;
7027 info->mti_big_aclsize = 0;
7033 /* context key: mdt_thread_key */
7034 LU_CONTEXT_KEY_DEFINE(mdt, LCT_MD_THREAD);
7036 struct lu_ucred *mdt_ucred(const struct mdt_thread_info *info)
7038 return lu_ucred(info->mti_env);
7041 struct lu_ucred *mdt_ucred_check(const struct mdt_thread_info *info)
7043 return lu_ucred_check(info->mti_env);
7047 * Enable/disable COS (Commit On Sharing).
7049 * Set/Clear the COS flag in mdt options.
7051 * \param mdt mdt device
7052 * \param val 0 disables COS, other values enable COS
7054 void mdt_enable_cos(struct mdt_device *mdt, bool val)
7059 mdt->mdt_opts.mo_cos = val;
7060 rc = lu_env_init(&env, LCT_LOCAL);
7061 if (unlikely(rc != 0)) {
7062 CWARN("%s: lu_env initialization failed, cannot "
7063 "sync: rc = %d\n", mdt_obd_name(mdt), rc);
7066 mdt_device_sync(&env, mdt);
7071 * Check COS (Commit On Sharing) status.
7073 * Return COS flag status.
7075 * \param mdt mdt device
7077 int mdt_cos_is_enabled(struct mdt_device *mdt)
7079 return mdt->mdt_opts.mo_cos != 0;
7082 static struct lu_device_type_operations mdt_device_type_ops = {
7083 .ldto_device_alloc = mdt_device_alloc,
7084 .ldto_device_free = mdt_device_free,
7085 .ldto_device_fini = mdt_device_fini
7088 static struct lu_device_type mdt_device_type = {
7089 .ldt_tags = LU_DEVICE_MD,
7090 .ldt_name = LUSTRE_MDT_NAME,
7091 .ldt_ops = &mdt_device_type_ops,
7092 .ldt_ctx_tags = LCT_MD_THREAD
7095 static int __init mdt_init(void)
7099 BUILD_BUG_ON(sizeof("0x0123456789ABCDEF:0x01234567:0x01234567") !=
7100 FID_NOBRACE_LEN + 1);
7101 BUILD_BUG_ON(sizeof("[0x0123456789ABCDEF:0x01234567:0x01234567]") !=
7103 rc = lu_kmem_init(mdt_caches);
7107 rc = mds_mod_init();
7111 rc = class_register_type(&mdt_obd_device_ops, NULL, true, NULL,
7112 LUSTRE_MDT_NAME, &mdt_device_type);
7117 lu_kmem_fini(mdt_caches);
7124 static void __exit mdt_exit(void)
7126 class_unregister_type(LUSTRE_MDT_NAME);
7128 lu_kmem_fini(mdt_caches);
7131 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
7132 MODULE_DESCRIPTION("Lustre Metadata Target ("LUSTRE_MDT_NAME")");
7133 MODULE_VERSION(LUSTRE_VERSION_STRING);
7134 MODULE_LICENSE("GPL");
7136 module_init(mdt_init);
7137 module_exit(mdt_exit);