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_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 =
601 MIN(mdt->mdt_max_ea_size,
603 OBD_ALLOC_LARGE(info->mti_big_acl,
604 info->mti_big_aclsize);
605 if (info->mti_big_acl == NULL) {
606 info->mti_big_aclsize = 0;
607 CERROR("%s: unable to grow "
610 PFID(mdt_object_fid(o)));
615 CDEBUG(D_INODE, "%s: grow the "DFID
616 " ACL buffer to size %d\n",
618 PFID(mdt_object_fid(o)),
619 info->mti_big_aclsize);
621 buf->lb_buf = info->mti_big_acl;
622 buf->lb_len = info->mti_big_aclsize;
627 CERROR("%s: unable to read "DFID" ACL: rc = %d\n",
628 mdt_obd_name(mdt), PFID(mdt_object_fid(o)), rc);
637 acl_buflen = req_capsule_get_size(pill, &RMF_ACL, RCL_SERVER);
638 if (acl_buflen >= rc)
641 /* If LOV/LMA EA is small, we can reuse part of their buffer */
642 client = ptlrpc_req_get_repsize(pill->rc_req);
643 server = lustre_packed_msg_size(pill->rc_req->rq_repmsg);
644 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER)) {
645 lmm_buflen = req_capsule_get_size(pill, &RMF_MDT_MD,
647 lmmsize = repbody->mbo_eadatasize;
650 if (client < server - acl_buflen - lmm_buflen + rc + lmmsize) {
651 CDEBUG(D_INODE, "%s: client prepared buffer size %d "
652 "is not big enough with the ACL size %d (%d)\n",
653 mdt_obd_name(mdt), client, rc,
654 server - acl_buflen - lmm_buflen + rc + lmmsize);
655 repbody->mbo_aclsize = 0;
656 repbody->mbo_valid &= ~OBD_MD_FLACL;
661 if (buf->lb_buf == info->mti_big_acl)
662 info->mti_big_acl_used = 1;
664 rc = nodemap_map_acl(nodemap, buf->lb_buf,
665 rc, NODEMAP_FS_TO_CLIENT);
666 /* if all ACLs mapped out, rc is still >= 0 */
668 CERROR("%s: nodemap_map_acl unable to parse "DFID
669 " ACL: rc = %d\n", mdt_obd_name(mdt),
670 PFID(mdt_object_fid(o)), rc);
671 repbody->mbo_aclsize = 0;
672 repbody->mbo_valid &= ~OBD_MD_FLACL;
674 repbody->mbo_aclsize = rc;
675 repbody->mbo_valid |= OBD_MD_FLACL;
684 /* XXX Look into layout in MDT layer. */
685 static inline bool mdt_hsm_is_released(struct lov_mds_md *lmm)
687 struct lov_comp_md_v1 *comp_v1;
688 struct lov_mds_md *v1;
691 if (lmm->lmm_magic == LOV_MAGIC_COMP_V1) {
692 comp_v1 = (struct lov_comp_md_v1 *)lmm;
694 for (i = 0; i < comp_v1->lcm_entry_count; i++) {
695 v1 = (struct lov_mds_md *)((char *)comp_v1 +
696 comp_v1->lcm_entries[i].lcme_offset);
697 /* We don't support partial release for now */
698 if (!(v1->lmm_pattern & LOV_PATTERN_F_RELEASED))
703 return (lmm->lmm_pattern & LOV_PATTERN_F_RELEASED) ?
708 void mdt_pack_attr2body(struct mdt_thread_info *info, struct mdt_body *b,
709 const struct lu_attr *attr, const struct lu_fid *fid)
711 struct md_attr *ma = &info->mti_attr;
712 struct obd_export *exp = info->mti_exp;
713 struct lu_nodemap *nodemap = NULL;
715 LASSERT(ma->ma_valid & MA_INODE);
717 if (attr->la_valid & LA_ATIME) {
718 b->mbo_atime = attr->la_atime;
719 b->mbo_valid |= OBD_MD_FLATIME;
721 if (attr->la_valid & LA_MTIME) {
722 b->mbo_mtime = attr->la_mtime;
723 b->mbo_valid |= OBD_MD_FLMTIME;
725 if (attr->la_valid & LA_CTIME) {
726 b->mbo_ctime = attr->la_ctime;
727 b->mbo_valid |= OBD_MD_FLCTIME;
729 if (attr->la_valid & LA_FLAGS) {
730 b->mbo_flags = attr->la_flags;
731 b->mbo_valid |= OBD_MD_FLFLAGS;
733 if (attr->la_valid & LA_NLINK) {
734 b->mbo_nlink = attr->la_nlink;
735 b->mbo_valid |= OBD_MD_FLNLINK;
737 if (attr->la_valid & (LA_UID|LA_GID)) {
738 nodemap = nodemap_get_from_exp(exp);
742 if (attr->la_valid & LA_UID) {
743 b->mbo_uid = nodemap_map_id(nodemap, NODEMAP_UID,
744 NODEMAP_FS_TO_CLIENT,
746 b->mbo_valid |= OBD_MD_FLUID;
748 if (attr->la_valid & LA_GID) {
749 b->mbo_gid = nodemap_map_id(nodemap, NODEMAP_GID,
750 NODEMAP_FS_TO_CLIENT,
752 b->mbo_valid |= OBD_MD_FLGID;
755 if (attr->la_valid & LA_PROJID) {
756 /* TODO, nodemap for project id */
757 b->mbo_projid = attr->la_projid;
758 b->mbo_valid |= OBD_MD_FLPROJID;
761 b->mbo_mode = attr->la_mode;
762 if (attr->la_valid & LA_MODE)
763 b->mbo_valid |= OBD_MD_FLMODE;
764 if (attr->la_valid & LA_TYPE)
765 b->mbo_valid |= OBD_MD_FLTYPE;
769 b->mbo_valid |= OBD_MD_FLID;
770 CDEBUG(D_INODE, DFID": nlink=%d, mode=%o, valid=%#llx\n",
771 PFID(fid), b->mbo_nlink, b->mbo_mode, b->mbo_valid);
774 if (!(attr->la_valid & LA_TYPE))
777 b->mbo_rdev = attr->la_rdev;
778 b->mbo_size = attr->la_size;
779 b->mbo_blocks = attr->la_blocks;
781 if (!S_ISREG(attr->la_mode)) {
782 b->mbo_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS | OBD_MD_FLRDEV;
783 } else if (ma->ma_need & MA_LOV && !(ma->ma_valid & MA_LOV)) {
784 /* means no objects are allocated on osts. */
785 LASSERT(!(ma->ma_valid & MA_LOV));
786 /* just ignore blocks occupied by extend attributes on MDS */
788 /* if no object is allocated on osts, the size on mds is valid.
790 b->mbo_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
791 } else if ((ma->ma_valid & MA_LOV) && ma->ma_lmm != NULL) {
792 if (mdt_hsm_is_released(ma->ma_lmm)) {
793 /* A released file stores its size on MDS. */
794 /* But return 1 block for released file, unless tools
795 * like tar will consider it fully sparse. (LU-3864)
797 if (unlikely(b->mbo_size == 0))
801 b->mbo_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
802 } else if (info->mti_som_valid) { /* som is valid */
803 b->mbo_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
807 if (fid != NULL && (b->mbo_valid & OBD_MD_FLSIZE))
808 CDEBUG(D_VFSTRACE, DFID": returning size %llu\n",
809 PFID(fid), (unsigned long long)b->mbo_size);
812 if (!IS_ERR_OR_NULL(nodemap))
813 nodemap_putref(nodemap);
816 static inline int mdt_body_has_lov(const struct lu_attr *la,
817 const struct mdt_body *body)
819 return (S_ISREG(la->la_mode) && (body->mbo_valid & OBD_MD_FLEASIZE)) ||
820 (S_ISDIR(la->la_mode) && (body->mbo_valid & OBD_MD_FLDIREA));
823 void mdt_client_compatibility(struct mdt_thread_info *info)
825 struct mdt_body *body;
826 struct ptlrpc_request *req = mdt_info_req(info);
827 struct obd_export *exp = req->rq_export;
828 struct md_attr *ma = &info->mti_attr;
829 struct lu_attr *la = &ma->ma_attr;
832 if (exp_connect_layout(exp))
833 /* the client can deal with 16-bit lmm_stripe_count */
836 body = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
838 if (!mdt_body_has_lov(la, body))
841 /* now we have a reply with a lov for a client not compatible with the
842 * layout lock so we have to clean the layout generation number */
843 if (S_ISREG(la->la_mode))
844 ma->ma_lmm->lmm_layout_gen = 0;
848 static int mdt_attr_get_eabuf_size(struct mdt_thread_info *info,
849 struct mdt_object *o)
851 const struct lu_env *env = info->mti_env;
854 rc = mo_xattr_get(env, mdt_object_child(o), &LU_BUF_NULL,
863 /* Is it a directory? Let's check for the LMV as well */
864 if (S_ISDIR(lu_object_attr(&mdt_object_child(o)->mo_lu))) {
865 rc2 = mo_xattr_get(env, mdt_object_child(o), &LU_BUF_NULL,
869 rc2 = mo_xattr_get(env, mdt_object_child(o),
871 XATTR_NAME_DEFAULT_LMV);
873 if ((rc2 < 0 && rc2 != -ENODATA) || (rc2 > rc))
881 int mdt_big_xattr_get(struct mdt_thread_info *info, struct mdt_object *o,
884 const struct lu_env *env = info->mti_env;
888 LASSERT(info->mti_big_lmm_used == 0);
889 rc = mo_xattr_get(env, mdt_object_child(o), &LU_BUF_NULL, name);
893 /* big_lmm may need to be grown */
894 if (info->mti_big_lmmsize < rc) {
895 int size = size_roundup_power2(rc);
897 if (info->mti_big_lmmsize > 0) {
898 /* free old buffer */
899 LASSERT(info->mti_big_lmm);
900 OBD_FREE_LARGE(info->mti_big_lmm,
901 info->mti_big_lmmsize);
902 info->mti_big_lmm = NULL;
903 info->mti_big_lmmsize = 0;
906 OBD_ALLOC_LARGE(info->mti_big_lmm, size);
907 if (info->mti_big_lmm == NULL)
909 info->mti_big_lmmsize = size;
911 LASSERT(info->mti_big_lmmsize >= rc);
913 info->mti_buf.lb_buf = info->mti_big_lmm;
914 info->mti_buf.lb_len = info->mti_big_lmmsize;
915 rc = mo_xattr_get(env, mdt_object_child(o), &info->mti_buf, name);
920 int mdt_stripe_get(struct mdt_thread_info *info, struct mdt_object *o,
921 struct md_attr *ma, const char *name)
923 struct md_object *next = mdt_object_child(o);
924 struct lu_buf *buf = &info->mti_buf;
927 if (strcmp(name, XATTR_NAME_LOV) == 0) {
928 buf->lb_buf = ma->ma_lmm;
929 buf->lb_len = ma->ma_lmm_size;
930 LASSERT(!(ma->ma_valid & MA_LOV));
931 } else if (strcmp(name, XATTR_NAME_LMV) == 0) {
932 buf->lb_buf = ma->ma_lmv;
933 buf->lb_len = ma->ma_lmv_size;
934 LASSERT(!(ma->ma_valid & MA_LMV));
935 } else if (strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0) {
936 buf->lb_buf = ma->ma_lmv;
937 buf->lb_len = ma->ma_lmv_size;
938 LASSERT(!(ma->ma_valid & MA_LMV_DEF));
943 LASSERT(buf->lb_buf);
945 rc = mo_xattr_get(info->mti_env, next, buf, name);
949 if (strcmp(name, XATTR_NAME_LOV) == 0) {
950 if (info->mti_big_lmm_used)
951 ma->ma_lmm = info->mti_big_lmm;
953 /* NOT return LOV EA with hole to old client. */
954 if (unlikely(le32_to_cpu(ma->ma_lmm->lmm_pattern) &
955 LOV_PATTERN_F_HOLE) &&
956 !(exp_connect_flags(info->mti_exp) &
957 OBD_CONNECT_LFSCK)) {
960 ma->ma_lmm_size = rc;
961 ma->ma_valid |= MA_LOV;
963 } else if (strcmp(name, XATTR_NAME_LMV) == 0) {
964 if (info->mti_big_lmm_used)
965 ma->ma_lmv = info->mti_big_lmm;
967 ma->ma_lmv_size = rc;
968 ma->ma_valid |= MA_LMV;
969 } else if (strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0) {
970 ma->ma_lmv_size = rc;
971 ma->ma_valid |= MA_LMV_DEF;
974 /* Update mdt_max_mdsize so all clients will be aware that */
975 if (info->mti_mdt->mdt_max_mdsize < rc)
976 info->mti_mdt->mdt_max_mdsize = rc;
979 } else if (rc == -ENODATA) {
982 } else if (rc == -ERANGE) {
983 /* Default LMV has fixed size, so it must be able to fit
984 * in the original buffer */
985 if (strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0)
987 rc = mdt_big_xattr_get(info, o, name);
989 info->mti_big_lmm_used = 1;
997 int mdt_attr_get_pfid(struct mdt_thread_info *info, struct mdt_object *o,
1000 struct lu_buf *buf = &info->mti_buf;
1001 struct link_ea_header *leh;
1002 struct link_ea_entry *lee;
1006 buf->lb_buf = info->mti_big_lmm;
1007 buf->lb_len = info->mti_big_lmmsize;
1008 rc = mo_xattr_get(info->mti_env, mdt_object_child(o),
1009 buf, XATTR_NAME_LINK);
1010 /* ignore errors, MA_PFID won't be set and it is
1011 * up to the caller to treat this as an error */
1012 if (rc == -ERANGE || buf->lb_len == 0) {
1013 rc = mdt_big_xattr_get(info, o, XATTR_NAME_LINK);
1014 buf->lb_buf = info->mti_big_lmm;
1015 buf->lb_len = info->mti_big_lmmsize;
1020 if (rc < sizeof(*leh)) {
1021 CERROR("short LinkEA on "DFID": rc = %d\n",
1022 PFID(mdt_object_fid(o)), rc);
1026 leh = (struct link_ea_header *) buf->lb_buf;
1027 lee = (struct link_ea_entry *)(leh + 1);
1028 if (leh->leh_magic == __swab32(LINK_EA_MAGIC)) {
1029 leh->leh_magic = LINK_EA_MAGIC;
1030 leh->leh_reccount = __swab32(leh->leh_reccount);
1031 leh->leh_len = __swab64(leh->leh_len);
1033 if (leh->leh_magic != LINK_EA_MAGIC)
1035 if (leh->leh_reccount == 0)
1038 memcpy(pfid, &lee->lee_parent_fid, sizeof(*pfid));
1039 fid_be_to_cpu(pfid, pfid);
1044 int mdt_attr_get_complex(struct mdt_thread_info *info,
1045 struct mdt_object *o, struct md_attr *ma)
1047 const struct lu_env *env = info->mti_env;
1048 struct md_object *next = mdt_object_child(o);
1049 struct lu_buf *buf = &info->mti_buf;
1050 int need = ma->ma_need;
1057 if (mdt_object_exists(o) == 0)
1058 GOTO(out, rc = -ENOENT);
1059 mode = lu_object_attr(&next->mo_lu);
1061 if (need & MA_INODE) {
1062 ma->ma_need = MA_INODE;
1063 rc = mo_attr_get(env, next, ma);
1068 (void) mdt_get_som(info, o, ma);
1069 ma->ma_valid |= MA_INODE;
1072 if (need & MA_PFID) {
1073 rc = mdt_attr_get_pfid(info, o, &ma->ma_pfid);
1075 ma->ma_valid |= MA_PFID;
1076 /* ignore this error, parent fid is not mandatory */
1080 if (need & MA_LOV && (S_ISREG(mode) || S_ISDIR(mode))) {
1081 rc = mdt_stripe_get(info, o, ma, XATTR_NAME_LOV);
1086 if (need & MA_LMV && S_ISDIR(mode)) {
1087 rc = mdt_stripe_get(info, o, ma, XATTR_NAME_LMV);
1092 if (need & MA_LMV_DEF && S_ISDIR(mode)) {
1093 rc = mdt_stripe_get(info, o, ma, XATTR_NAME_DEFAULT_LMV);
1099 * In the handle of MA_INODE, we may already get the SOM attr.
1101 if (need & MA_SOM && S_ISREG(mode) && !(ma->ma_valid & MA_SOM)) {
1102 rc = mdt_get_som(info, o, ma);
1107 if (need & MA_HSM && S_ISREG(mode)) {
1108 buf->lb_buf = info->mti_xattr_buf;
1109 buf->lb_len = sizeof(info->mti_xattr_buf);
1110 CLASSERT(sizeof(struct hsm_attrs) <=
1111 sizeof(info->mti_xattr_buf));
1112 rc2 = mo_xattr_get(info->mti_env, next, buf, XATTR_NAME_HSM);
1113 rc2 = lustre_buf2hsm(info->mti_xattr_buf, rc2, &ma->ma_hsm);
1115 ma->ma_valid |= MA_HSM;
1116 else if (rc2 < 0 && rc2 != -ENODATA)
1117 GOTO(out, rc = rc2);
1120 #ifdef CONFIG_FS_POSIX_ACL
1121 if (need & MA_ACL_DEF && S_ISDIR(mode)) {
1122 buf->lb_buf = ma->ma_acl;
1123 buf->lb_len = ma->ma_acl_size;
1124 rc2 = mo_xattr_get(env, next, buf, XATTR_NAME_ACL_DEFAULT);
1126 ma->ma_acl_size = rc2;
1127 ma->ma_valid |= MA_ACL_DEF;
1128 } else if (rc2 == -ENODATA) {
1130 ma->ma_acl_size = 0;
1132 GOTO(out, rc = rc2);
1137 CDEBUG(D_INODE, "after getattr rc = %d, ma_valid = %#llx ma_lmm=%p\n",
1138 rc, ma->ma_valid, ma->ma_lmm);
1142 static int mdt_getattr_internal(struct mdt_thread_info *info,
1143 struct mdt_object *o, int ma_need)
1145 struct md_object *next = mdt_object_child(o);
1146 const struct mdt_body *reqbody = info->mti_body;
1147 struct ptlrpc_request *req = mdt_info_req(info);
1148 struct md_attr *ma = &info->mti_attr;
1149 struct lu_attr *la = &ma->ma_attr;
1150 struct req_capsule *pill = info->mti_pill;
1151 const struct lu_env *env = info->mti_env;
1152 struct mdt_body *repbody;
1153 struct lu_buf *buffer = &info->mti_buf;
1154 struct obd_export *exp = info->mti_exp;
1158 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK))
1159 RETURN(err_serious(-ENOMEM));
1161 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
1165 if (mdt_object_remote(o)) {
1166 /* This object is located on remote node.*/
1167 /* Return -ENOTSUPP for old client */
1168 if (!mdt_is_dne_client(req->rq_export))
1169 GOTO(out, rc = -ENOTSUPP);
1171 repbody->mbo_fid1 = *mdt_object_fid(o);
1172 repbody->mbo_valid = OBD_MD_FLID | OBD_MD_MDS;
1176 if (reqbody->mbo_eadatasize > 0) {
1177 buffer->lb_buf = req_capsule_server_get(pill, &RMF_MDT_MD);
1178 if (buffer->lb_buf == NULL)
1179 GOTO(out, rc = -EPROTO);
1180 buffer->lb_len = req_capsule_get_size(pill, &RMF_MDT_MD,
1183 buffer->lb_buf = NULL;
1185 ma_need &= ~(MA_LOV | MA_LMV);
1186 CDEBUG(D_INFO, "%s: RPC from %s: does not need LOVEA.\n",
1187 mdt_obd_name(info->mti_mdt),
1188 req->rq_export->exp_client_uuid.uuid);
1191 /* If it is dir object and client require MEA, then we got MEA */
1192 if (S_ISDIR(lu_object_attr(&next->mo_lu)) &&
1193 (reqbody->mbo_valid & (OBD_MD_MEA | OBD_MD_DEFAULT_MEA))) {
1194 /* Assumption: MDT_MD size is enough for lmv size. */
1195 ma->ma_lmv = buffer->lb_buf;
1196 ma->ma_lmv_size = buffer->lb_len;
1197 ma->ma_need = MA_INODE;
1198 if (ma->ma_lmv_size > 0) {
1199 if (reqbody->mbo_valid & OBD_MD_MEA)
1200 ma->ma_need |= MA_LMV;
1201 else if (reqbody->mbo_valid & OBD_MD_DEFAULT_MEA)
1202 ma->ma_need |= MA_LMV_DEF;
1205 ma->ma_lmm = buffer->lb_buf;
1206 ma->ma_lmm_size = buffer->lb_len;
1207 ma->ma_need = MA_INODE | MA_HSM;
1208 if (ma->ma_lmm_size > 0) {
1209 ma->ma_need |= MA_LOV;
1210 /* Older clients may crash if they getattr overstriped
1213 if (!exp_connect_overstriping(exp) &&
1214 mdt_lmm_is_overstriping(ma->ma_lmm))
1215 RETURN(-EOPNOTSUPP);
1219 if (S_ISDIR(lu_object_attr(&next->mo_lu)) &&
1220 reqbody->mbo_valid & OBD_MD_FLDIREA &&
1221 lustre_msg_get_opc(req->rq_reqmsg) == MDS_GETATTR) {
1222 /* get default stripe info for this dir. */
1223 ma->ma_need |= MA_LOV_DEF;
1225 ma->ma_need |= ma_need;
1227 rc = mdt_attr_get_complex(info, o, ma);
1229 CDEBUG(rc == -ENOENT ? D_OTHER : D_ERROR,
1230 "%s: getattr error for "DFID": rc = %d\n",
1231 mdt_obd_name(info->mti_mdt),
1232 PFID(mdt_object_fid(o)), rc);
1236 /* if file is released, check if a restore is running */
1237 if (ma->ma_valid & MA_HSM) {
1238 repbody->mbo_valid |= OBD_MD_TSTATE;
1239 if ((ma->ma_hsm.mh_flags & HS_RELEASED) &&
1240 mdt_hsm_restore_is_running(info, mdt_object_fid(o)))
1241 repbody->mbo_t_state = MS_RESTORE;
1244 if (likely(ma->ma_valid & MA_INODE))
1245 mdt_pack_attr2body(info, repbody, la, mdt_object_fid(o));
1249 if (mdt_body_has_lov(la, reqbody)) {
1250 if (ma->ma_valid & MA_LOV) {
1251 LASSERT(ma->ma_lmm_size);
1252 repbody->mbo_eadatasize = ma->ma_lmm_size;
1253 if (S_ISDIR(la->la_mode))
1254 repbody->mbo_valid |= OBD_MD_FLDIREA;
1256 repbody->mbo_valid |= OBD_MD_FLEASIZE;
1257 mdt_dump_lmm(D_INFO, ma->ma_lmm, repbody->mbo_valid);
1259 if (ma->ma_valid & MA_LMV) {
1260 /* Return -ENOTSUPP for old client */
1261 if (!mdt_is_striped_client(req->rq_export))
1264 LASSERT(S_ISDIR(la->la_mode));
1265 mdt_dump_lmv(D_INFO, ma->ma_lmv);
1266 repbody->mbo_eadatasize = ma->ma_lmv_size;
1267 repbody->mbo_valid |= (OBD_MD_FLDIREA|OBD_MD_MEA);
1269 if (ma->ma_valid & MA_LMV_DEF) {
1270 /* Return -ENOTSUPP for old client */
1271 if (!mdt_is_striped_client(req->rq_export))
1273 LASSERT(S_ISDIR(la->la_mode));
1274 mdt_dump_lmv(D_INFO, ma->ma_lmv);
1275 repbody->mbo_eadatasize = ma->ma_lmv_size;
1276 repbody->mbo_valid |= (OBD_MD_FLDIREA |
1277 OBD_MD_DEFAULT_MEA);
1279 } else if (S_ISLNK(la->la_mode) &&
1280 reqbody->mbo_valid & OBD_MD_LINKNAME) {
1281 buffer->lb_buf = ma->ma_lmm;
1282 /* eadatasize from client includes NULL-terminator, so
1283 * there is no need to read it */
1284 buffer->lb_len = reqbody->mbo_eadatasize - 1;
1285 rc = mo_readlink(env, next, buffer);
1286 if (unlikely(rc <= 0)) {
1287 CERROR("%s: readlink failed for "DFID": rc = %d\n",
1288 mdt_obd_name(info->mti_mdt),
1289 PFID(mdt_object_fid(o)), rc);
1292 int print_limit = min_t(int, PAGE_SIZE - 128, rc);
1294 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_READLINK_EPROTO))
1296 repbody->mbo_valid |= OBD_MD_LINKNAME;
1297 /* we need to report back size with NULL-terminator
1298 * because client expects that */
1299 repbody->mbo_eadatasize = rc + 1;
1300 if (repbody->mbo_eadatasize != reqbody->mbo_eadatasize)
1301 CDEBUG(D_INODE, "%s: Read shorter symlink %d "
1302 "on "DFID ", expected %d\n",
1303 mdt_obd_name(info->mti_mdt),
1304 rc, PFID(mdt_object_fid(o)),
1305 reqbody->mbo_eadatasize - 1);
1306 /* NULL terminate */
1307 ((char *)ma->ma_lmm)[rc] = 0;
1309 /* If the total CDEBUG() size is larger than a page, it
1310 * will print a warning to the console, avoid this by
1311 * printing just the last part of the symlink. */
1312 CDEBUG(D_INODE, "symlink dest %s%.*s, len = %d\n",
1313 print_limit < rc ? "..." : "", print_limit,
1314 (char *)ma->ma_lmm + rc - print_limit, rc);
1319 if (reqbody->mbo_valid & OBD_MD_FLMODEASIZE) {
1320 repbody->mbo_max_mdsize = info->mti_mdt->mdt_max_mdsize;
1321 repbody->mbo_valid |= OBD_MD_FLMODEASIZE;
1322 CDEBUG(D_INODE, "changing the max MD size to %u\n",
1323 repbody->mbo_max_mdsize);
1326 #ifdef CONFIG_FS_POSIX_ACL
1327 if ((exp_connect_flags(req->rq_export) & OBD_CONNECT_ACL) &&
1328 (reqbody->mbo_valid & OBD_MD_FLACL)) {
1329 struct lu_nodemap *nodemap = nodemap_get_from_exp(exp);
1330 if (IS_ERR(nodemap))
1331 RETURN(PTR_ERR(nodemap));
1333 rc = mdt_pack_acl2body(info, repbody, o, nodemap);
1334 nodemap_putref(nodemap);
1340 mdt_counter_incr(req, LPROC_MDT_GETATTR);
1345 static int mdt_getattr(struct tgt_session_info *tsi)
1347 struct mdt_thread_info *info = tsi2mdt_info(tsi);
1348 struct mdt_object *obj = info->mti_object;
1349 struct req_capsule *pill = info->mti_pill;
1350 struct mdt_body *reqbody;
1351 struct mdt_body *repbody;
1355 reqbody = req_capsule_client_get(pill, &RMF_MDT_BODY);
1357 LASSERT(obj != NULL);
1358 LASSERT(lu_object_assert_exists(&obj->mot_obj));
1360 /* Special case for Data-on-MDT files to get data version */
1361 if (unlikely(reqbody->mbo_valid & OBD_MD_FLDATAVERSION)) {
1362 rc = mdt_data_version_get(tsi);
1366 /* Unlike intent case where we need to pre-fill out buffers early on
1367 * in intent policy for ldlm reasons, here we can have a much better
1368 * guess at EA size by just reading it from disk.
1369 * Exceptions are readdir and (missing) directory striping */
1371 if (reqbody->mbo_valid & OBD_MD_LINKNAME) {
1372 /* No easy way to know how long is the symlink, but it cannot
1373 * be more than PATH_MAX, so we allocate +1 */
1375 /* A special case for fs ROOT: getattr there might fetch
1376 * default EA for entire fs, not just for this dir!
1378 } else if (lu_fid_eq(mdt_object_fid(obj),
1379 &info->mti_mdt->mdt_md_root_fid) &&
1380 (reqbody->mbo_valid & OBD_MD_FLDIREA) &&
1381 (lustre_msg_get_opc(mdt_info_req(info)->rq_reqmsg) ==
1383 /* Should the default strping be bigger, mdt_fix_reply
1384 * will reallocate */
1385 rc = DEF_REP_MD_SIZE;
1387 /* Read the actual EA size from disk */
1388 rc = mdt_attr_get_eabuf_size(info, obj);
1392 GOTO(out, rc = err_serious(rc));
1394 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER, rc);
1396 /* Set ACL reply buffer size as LUSTRE_POSIX_ACL_MAX_SIZE_OLD
1397 * by default. If the target object has more ACL entries, then
1398 * enlarge the buffer when necessary. */
1399 req_capsule_set_size(pill, &RMF_ACL, RCL_SERVER,
1400 LUSTRE_POSIX_ACL_MAX_SIZE_OLD);
1402 rc = req_capsule_server_pack(pill);
1403 if (unlikely(rc != 0))
1404 GOTO(out, rc = err_serious(rc));
1406 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
1407 LASSERT(repbody != NULL);
1408 repbody->mbo_eadatasize = 0;
1409 repbody->mbo_aclsize = 0;
1411 rc = mdt_check_ucred(info);
1413 GOTO(out_shrink, rc);
1415 info->mti_cross_ref = !!(reqbody->mbo_valid & OBD_MD_FLCROSSREF);
1417 rc = mdt_getattr_internal(info, obj, 0);
1420 mdt_client_compatibility(info);
1421 rc2 = mdt_fix_reply(info);
1425 mdt_thread_info_fini(info);
1430 * Handler of layout intent RPC requiring the layout modification
1432 * \param[in] info thread environment
1433 * \param[in] obj object
1434 * \param[in] layout layout change descriptor
1436 * \retval 0 on success
1437 * \retval < 0 error code
1439 int mdt_layout_change(struct mdt_thread_info *info, struct mdt_object *obj,
1440 struct md_layout_change *layout)
1442 struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LOCAL];
1446 if (!mdt_object_exists(obj))
1447 GOTO(out, rc = -ENOENT);
1449 if (!S_ISREG(lu_object_attr(&obj->mot_obj)))
1450 GOTO(out, rc = -EINVAL);
1452 rc = mo_permission(info->mti_env, NULL, mdt_object_child(obj), NULL,
1457 /* take layout lock to prepare layout change */
1458 mdt_lock_reg_init(lh, LCK_EX);
1459 rc = mdt_object_lock(info, obj, lh, MDS_INODELOCK_LAYOUT);
1463 mutex_lock(&obj->mot_som_mutex);
1464 rc = mo_layout_change(info->mti_env, mdt_object_child(obj), layout);
1465 mutex_unlock(&obj->mot_som_mutex);
1466 mdt_object_unlock(info, obj, lh, 1);
1472 * Exchange MOF_LOV_CREATED flags between two objects after a
1473 * layout swap. No assumption is made on whether o1 or o2 have
1474 * created objects or not.
1476 * \param[in,out] o1 First swap layout object
1477 * \param[in,out] o2 Second swap layout object
1479 static void mdt_swap_lov_flag(struct mdt_object *o1, struct mdt_object *o2)
1481 unsigned int o1_lov_created = o1->mot_lov_created;
1483 mutex_lock(&o1->mot_lov_mutex);
1484 mutex_lock(&o2->mot_lov_mutex);
1486 o1->mot_lov_created = o2->mot_lov_created;
1487 o2->mot_lov_created = o1_lov_created;
1489 mutex_unlock(&o2->mot_lov_mutex);
1490 mutex_unlock(&o1->mot_lov_mutex);
1493 static int mdt_swap_layouts(struct tgt_session_info *tsi)
1495 struct mdt_thread_info *info;
1496 struct ptlrpc_request *req = tgt_ses_req(tsi);
1497 struct obd_export *exp = req->rq_export;
1498 struct mdt_object *o1, *o2, *o;
1499 struct mdt_lock_handle *lh1, *lh2;
1500 struct mdc_swap_layouts *msl;
1504 /* client does not support layout lock, so layout swaping
1506 * FIXME: there is a problem for old clients which don't support
1507 * layout lock yet. If those clients have already opened the file
1508 * they won't be notified at all so that old layout may still be
1509 * used to do IO. This can be fixed after file release is landed by
1510 * doing exclusive open and taking full EX ibits lock. - Jinshan */
1511 if (!exp_connect_layout(exp))
1512 RETURN(-EOPNOTSUPP);
1514 info = tsi2mdt_info(tsi);
1516 if (info->mti_dlm_req != NULL)
1517 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1519 o1 = info->mti_object;
1520 o = o2 = mdt_object_find(info->mti_env, info->mti_mdt,
1521 &info->mti_body->mbo_fid2);
1523 GOTO(out, rc = PTR_ERR(o));
1525 if (mdt_object_remote(o) || !mdt_object_exists(o)) /* remote object */
1526 GOTO(put, rc = -ENOENT);
1528 rc = lu_fid_cmp(&info->mti_body->mbo_fid1, &info->mti_body->mbo_fid2);
1529 if (unlikely(rc == 0)) /* same file, you kidding me? no-op. */
1535 /* permission check. Make sure the calling process having permission
1536 * to write both files. */
1537 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o1), NULL,
1542 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o2), NULL,
1547 msl = req_capsule_client_get(info->mti_pill, &RMF_SWAP_LAYOUTS);
1549 GOTO(put, rc = -EPROTO);
1551 lh1 = &info->mti_lh[MDT_LH_NEW];
1552 mdt_lock_reg_init(lh1, LCK_EX);
1553 lh2 = &info->mti_lh[MDT_LH_OLD];
1554 mdt_lock_reg_init(lh2, LCK_EX);
1556 rc = mdt_object_lock(info, o1, lh1, MDS_INODELOCK_LAYOUT |
1557 MDS_INODELOCK_XATTR);
1561 rc = mdt_object_lock(info, o2, lh2, MDS_INODELOCK_LAYOUT |
1562 MDS_INODELOCK_XATTR);
1566 rc = mo_swap_layouts(info->mti_env, mdt_object_child(o1),
1567 mdt_object_child(o2), msl->msl_flags);
1571 mdt_swap_lov_flag(o1, o2);
1574 mdt_object_unlock(info, o2, lh2, rc);
1576 mdt_object_unlock(info, o1, lh1, rc);
1578 mdt_object_put(info->mti_env, o);
1580 mdt_thread_info_fini(info);
1584 static int mdt_raw_lookup(struct mdt_thread_info *info,
1585 struct mdt_object *parent,
1586 const struct lu_name *lname,
1587 struct ldlm_reply *ldlm_rep)
1589 struct lu_fid *child_fid = &info->mti_tmp_fid1;
1593 LASSERT(!info->mti_cross_ref);
1595 /* Only got the fid of this obj by name */
1596 fid_zero(child_fid);
1597 rc = mdo_lookup(info->mti_env, mdt_object_child(info->mti_object),
1598 lname, child_fid, &info->mti_spec);
1600 struct mdt_body *repbody;
1602 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1603 repbody->mbo_fid1 = *child_fid;
1604 repbody->mbo_valid = OBD_MD_FLID;
1605 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1606 } else if (rc == -ENOENT) {
1607 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1614 * UPDATE lock should be taken against parent, and be released before exit;
1615 * child_bits lock should be taken against child, and be returned back:
1616 * (1)normal request should release the child lock;
1617 * (2)intent request will grant the lock to client.
1619 static int mdt_getattr_name_lock(struct mdt_thread_info *info,
1620 struct mdt_lock_handle *lhc,
1622 struct ldlm_reply *ldlm_rep)
1624 struct ptlrpc_request *req = mdt_info_req(info);
1625 struct mdt_body *reqbody = NULL;
1626 struct mdt_object *parent = info->mti_object;
1627 struct mdt_object *child;
1628 struct lu_fid *child_fid = &info->mti_tmp_fid1;
1629 struct lu_name *lname = NULL;
1630 struct mdt_lock_handle *lhp = NULL;
1631 struct ldlm_lock *lock;
1632 struct req_capsule *pill = info->mti_pill;
1640 is_resent = lustre_handle_is_used(&lhc->mlh_reg_lh);
1641 LASSERT(ergo(is_resent,
1642 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT));
1647 if (info->mti_cross_ref) {
1648 /* Only getattr on the child. Parent is on another node. */
1649 mdt_set_disposition(info, ldlm_rep,
1650 DISP_LOOKUP_EXECD | DISP_LOOKUP_POS);
1652 CDEBUG(D_INODE, "partial getattr_name child_fid = "DFID", "
1654 PFID(mdt_object_fid(child)), ldlm_rep);
1656 rc = mdt_check_resent_lock(info, child, lhc);
1659 } else if (rc > 0) {
1660 mdt_lock_handle_init(lhc);
1661 mdt_lock_reg_init(lhc, LCK_PR);
1664 * Object's name entry is on another MDS, it will
1665 * request PERM lock only because LOOKUP lock is owned
1666 * by the MDS where name entry resides.
1668 * TODO: it should try layout lock too. - Jinshan
1670 child_bits &= ~(MDS_INODELOCK_LOOKUP |
1671 MDS_INODELOCK_LAYOUT);
1672 child_bits |= MDS_INODELOCK_PERM;
1674 rc = mdt_object_lock(info, child, lhc, child_bits);
1679 /* Finally, we can get attr for child. */
1680 if (!mdt_object_exists(child)) {
1681 LU_OBJECT_DEBUG(D_INFO, info->mti_env,
1683 "remote object doesn't exist.");
1684 mdt_object_unlock(info, child, lhc, 1);
1688 rc = mdt_getattr_internal(info, child, 0);
1689 if (unlikely(rc != 0))
1690 mdt_object_unlock(info, child, lhc, 1);
1692 mdt_pack_secctx_in_reply(info, child);
1697 lname = &info->mti_name;
1698 mdt_name_unpack(pill, &RMF_NAME, lname, MNF_FIX_ANON);
1700 if (lu_name_is_valid(lname)) {
1701 CDEBUG(D_INODE, "getattr with lock for "DFID"/"DNAME", "
1702 "ldlm_rep = %p\n", PFID(mdt_object_fid(parent)),
1703 PNAME(lname), ldlm_rep);
1705 reqbody = req_capsule_client_get(pill, &RMF_MDT_BODY);
1706 if (unlikely(reqbody == NULL))
1707 RETURN(err_serious(-EPROTO));
1709 *child_fid = reqbody->mbo_fid2;
1711 if (unlikely(!fid_is_sane(child_fid)))
1712 RETURN(err_serious(-EINVAL));
1714 CDEBUG(D_INODE, "getattr with lock for "DFID"/"DFID", "
1716 PFID(mdt_object_fid(parent)),
1717 PFID(&reqbody->mbo_fid2), ldlm_rep);
1720 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_EXECD);
1722 if (unlikely(!mdt_object_exists(parent)) && lu_name_is_valid(lname)) {
1723 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
1725 "Parent doesn't exist!");
1729 if (mdt_object_remote(parent)) {
1730 CERROR("%s: parent "DFID" is on remote target\n",
1731 mdt_obd_name(info->mti_mdt),
1732 PFID(mdt_object_fid(parent)));
1736 if (lu_name_is_valid(lname)) {
1737 /* Always allow to lookup ".." */
1738 if (unlikely(lname->ln_namelen == 2 &&
1739 lname->ln_name[0] == '.' &&
1740 lname->ln_name[1] == '.'))
1741 info->mti_spec.sp_permitted = 1;
1743 if (info->mti_body->mbo_valid == OBD_MD_FLID) {
1744 rc = mdt_raw_lookup(info, parent, lname, ldlm_rep);
1749 /* step 1: lock parent only if parent is a directory */
1750 if (S_ISDIR(lu_object_attr(&parent->mot_obj))) {
1751 lhp = &info->mti_lh[MDT_LH_PARENT];
1752 mdt_lock_pdo_init(lhp, LCK_PR, lname);
1753 rc = mdt_object_lock(info, parent, lhp,
1754 MDS_INODELOCK_UPDATE);
1755 if (unlikely(rc != 0))
1759 /* step 2: lookup child's fid by name */
1760 fid_zero(child_fid);
1761 rc = mdo_lookup(info->mti_env, mdt_object_child(parent), lname,
1762 child_fid, &info->mti_spec);
1764 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1767 GOTO(out_parent, rc);
1770 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1773 *step 3: find the child object by fid & lock it.
1774 * regardless if it is local or remote.
1776 *Note: LU-3240 (commit 762f2114d282a98ebfa4dbbeea9298a8088ad24e)
1777 * set parent dir fid the same as child fid in getattr by fid case
1778 * we should not lu_object_find() the object again, could lead
1779 * to hung if there is a concurrent unlink destroyed the object.
1781 if (lu_fid_eq(mdt_object_fid(parent), child_fid)) {
1782 mdt_object_get(info->mti_env, parent);
1785 child = mdt_object_find(info->mti_env, info->mti_mdt,
1789 if (unlikely(IS_ERR(child)))
1790 GOTO(out_parent, rc = PTR_ERR(child));
1792 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RESEND, obd_timeout * 2);
1793 if (!mdt_object_exists(child)) {
1794 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
1796 "Object doesn't exist!");
1797 GOTO(out_child, rc = -ENOENT);
1800 rc = mdt_check_resent_lock(info, child, lhc);
1802 GOTO(out_child, rc);
1803 } else if (rc > 0) {
1804 mdt_lock_handle_init(lhc);
1805 mdt_lock_reg_init(lhc, LCK_PR);
1807 if (!(child_bits & MDS_INODELOCK_UPDATE) &&
1808 mdt_object_exists(child) && !mdt_object_remote(child)) {
1809 struct md_attr *ma = &info->mti_attr;
1812 ma->ma_need = MA_INODE;
1813 rc = mdt_attr_get_complex(info, child, ma);
1814 if (unlikely(rc != 0))
1815 GOTO(out_child, rc);
1817 /* If the file has not been changed for some time, we
1818 * return not only a LOOKUP lock, but also an UPDATE
1819 * lock and this might save us RPC on later STAT. For
1820 * directories, it also let negative dentry cache start
1821 * working for this dir. */
1822 if (ma->ma_valid & MA_INODE &&
1823 ma->ma_attr.la_valid & LA_CTIME &&
1824 info->mti_mdt->mdt_namespace->ns_ctime_age_limit +
1825 ma->ma_attr.la_ctime < ktime_get_real_seconds())
1826 child_bits |= MDS_INODELOCK_UPDATE;
1829 /* layout lock must be granted in a best-effort way
1830 * for IT operations */
1831 LASSERT(!(child_bits & MDS_INODELOCK_LAYOUT));
1832 if (S_ISREG(lu_object_attr(&child->mot_obj)) &&
1833 !mdt_object_remote(child) && ldlm_rep != NULL) {
1834 if (!OBD_FAIL_CHECK(OBD_FAIL_MDS_NO_LL_GETATTR) &&
1835 exp_connect_layout(info->mti_exp)) {
1836 /* try to grant layout lock for regular file. */
1837 try_bits = MDS_INODELOCK_LAYOUT;
1839 /* Acquire DOM lock in advance for data-on-mdt file */
1840 if (child != parent)
1841 try_bits |= MDS_INODELOCK_DOM;
1844 if (try_bits != 0) {
1845 /* try layout lock, it may fail to be granted due to
1846 * contention at LOOKUP or UPDATE */
1847 rc = mdt_object_lock_try(info, child, lhc, &child_bits,
1849 if (child_bits & MDS_INODELOCK_LAYOUT)
1852 /* Do not enqueue the UPDATE lock from MDT(cross-MDT),
1853 * client will enqueue the lock to the remote MDT */
1854 if (mdt_object_remote(child))
1855 child_bits &= ~MDS_INODELOCK_UPDATE;
1856 rc = mdt_object_lock(info, child, lhc, child_bits);
1858 if (unlikely(rc != 0))
1859 GOTO(out_child, rc);
1862 /* finally, we can get attr for child. */
1863 rc = mdt_getattr_internal(info, child, ma_need);
1864 if (unlikely(rc != 0)) {
1865 mdt_object_unlock(info, child, lhc, 1);
1866 GOTO(out_child, rc);
1869 mdt_pack_secctx_in_reply(info, child);
1871 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
1873 /* Debugging code. */
1874 LDLM_DEBUG(lock, "Returning lock to client");
1875 LASSERTF(fid_res_name_eq(mdt_object_fid(child),
1876 &lock->l_resource->lr_name),
1877 "Lock res_id: "DLDLMRES", fid: "DFID"\n",
1878 PLDLMRES(lock->l_resource),
1879 PFID(mdt_object_fid(child)));
1881 if (S_ISREG(lu_object_attr(&child->mot_obj)) &&
1882 mdt_object_exists(child) && !mdt_object_remote(child) &&
1884 mdt_object_put(info->mti_env, child);
1885 rc = mdt_pack_size2body(info, child_fid,
1887 if (rc != 0 && child_bits & MDS_INODELOCK_DOM) {
1888 /* DOM lock was taken in advance but this is
1889 * not DoM file. Drop the lock.
1891 lock_res_and_lock(lock);
1892 ldlm_inodebits_drop(lock, MDS_INODELOCK_DOM);
1893 unlock_res_and_lock(lock);
1895 LDLM_LOCK_PUT(lock);
1896 GOTO(out_parent, rc = 0);
1898 LDLM_LOCK_PUT(lock);
1903 mdt_object_put(info->mti_env, child);
1906 mdt_object_unlock(info, parent, lhp, 1);
1910 /* normal handler: should release the child lock */
1911 static int mdt_getattr_name(struct tgt_session_info *tsi)
1913 struct mdt_thread_info *info = tsi2mdt_info(tsi);
1914 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_CHILD];
1915 struct mdt_body *reqbody;
1916 struct mdt_body *repbody;
1920 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
1921 LASSERT(reqbody != NULL);
1922 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1923 LASSERT(repbody != NULL);
1925 info->mti_cross_ref = !!(reqbody->mbo_valid & OBD_MD_FLCROSSREF);
1926 repbody->mbo_eadatasize = 0;
1927 repbody->mbo_aclsize = 0;
1929 rc = mdt_init_ucred_intent_getattr(info, reqbody);
1931 GOTO(out_shrink, rc);
1933 rc = mdt_getattr_name_lock(info, lhc, MDS_INODELOCK_UPDATE, NULL);
1934 if (lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1935 ldlm_lock_decref(&lhc->mlh_reg_lh, lhc->mlh_reg_mode);
1936 lhc->mlh_reg_lh.cookie = 0;
1938 mdt_exit_ucred(info);
1941 mdt_client_compatibility(info);
1942 rc2 = mdt_fix_reply(info);
1945 mdt_thread_info_fini(info);
1949 static int mdt_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
1950 void *karg, void __user *uarg);
1952 static int mdt_set_info(struct tgt_session_info *tsi)
1954 struct ptlrpc_request *req = tgt_ses_req(tsi);
1957 int keylen, vallen, rc = 0;
1961 key = req_capsule_client_get(tsi->tsi_pill, &RMF_SETINFO_KEY);
1963 DEBUG_REQ(D_HA, req, "no set_info key");
1964 RETURN(err_serious(-EFAULT));
1967 keylen = req_capsule_get_size(tsi->tsi_pill, &RMF_SETINFO_KEY,
1970 val = req_capsule_client_get(tsi->tsi_pill, &RMF_SETINFO_VAL);
1972 DEBUG_REQ(D_HA, req, "no set_info val");
1973 RETURN(err_serious(-EFAULT));
1976 vallen = req_capsule_get_size(tsi->tsi_pill, &RMF_SETINFO_VAL,
1979 /* Swab any part of val you need to here */
1980 if (KEY_IS(KEY_READ_ONLY)) {
1981 spin_lock(&req->rq_export->exp_lock);
1983 *exp_connect_flags_ptr(req->rq_export) |=
1986 *exp_connect_flags_ptr(req->rq_export) &=
1987 ~OBD_CONNECT_RDONLY;
1988 spin_unlock(&req->rq_export->exp_lock);
1989 } else if (KEY_IS(KEY_CHANGELOG_CLEAR)) {
1990 struct changelog_setinfo *cs = val;
1992 if (vallen != sizeof(*cs)) {
1993 CERROR("%s: bad changelog_clear setinfo size %d\n",
1994 tgt_name(tsi->tsi_tgt), vallen);
1997 if (ptlrpc_req_need_swab(req)) {
1998 __swab64s(&cs->cs_recno);
1999 __swab32s(&cs->cs_id);
2002 rc = mdt_iocontrol(OBD_IOC_CHANGELOG_CLEAR, req->rq_export,
2004 } else if (KEY_IS(KEY_EVICT_BY_NID)) {
2006 obd_export_evict_by_nid(req->rq_export->exp_obd, val);
2013 static int mdt_readpage(struct tgt_session_info *tsi)
2015 struct mdt_thread_info *info = mdt_th_info(tsi->tsi_env);
2016 struct mdt_object *object = mdt_obj(tsi->tsi_corpus);
2017 struct lu_rdpg *rdpg = &info->mti_u.rdpg.mti_rdpg;
2018 const struct mdt_body *reqbody = tsi->tsi_mdt_body;
2019 struct mdt_body *repbody;
2025 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK))
2026 RETURN(err_serious(-ENOMEM));
2028 repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_MDT_BODY);
2029 if (repbody == NULL || reqbody == NULL)
2030 RETURN(err_serious(-EFAULT));
2033 * prepare @rdpg before calling lower layers and transfer itself. Here
2034 * reqbody->size contains offset of where to start to read and
2035 * reqbody->nlink contains number bytes to read.
2037 rdpg->rp_hash = reqbody->mbo_size;
2038 if (rdpg->rp_hash != reqbody->mbo_size) {
2039 CERROR("Invalid hash: %#llx != %#llx\n",
2040 rdpg->rp_hash, reqbody->mbo_size);
2044 rdpg->rp_attrs = reqbody->mbo_mode;
2045 if (exp_connect_flags(tsi->tsi_exp) & OBD_CONNECT_64BITHASH)
2046 rdpg->rp_attrs |= LUDA_64BITHASH;
2047 rdpg->rp_count = min_t(unsigned int, reqbody->mbo_nlink,
2048 exp_max_brw_size(tsi->tsi_exp));
2049 rdpg->rp_npages = (rdpg->rp_count + PAGE_SIZE - 1) >>
2051 OBD_ALLOC(rdpg->rp_pages, rdpg->rp_npages * sizeof rdpg->rp_pages[0]);
2052 if (rdpg->rp_pages == NULL)
2055 for (i = 0; i < rdpg->rp_npages; ++i) {
2056 rdpg->rp_pages[i] = alloc_page(GFP_NOFS);
2057 if (rdpg->rp_pages[i] == NULL)
2058 GOTO(free_rdpg, rc = -ENOMEM);
2061 /* call lower layers to fill allocated pages with directory data */
2062 rc = mo_readpage(tsi->tsi_env, mdt_object_child(object), rdpg);
2064 GOTO(free_rdpg, rc);
2066 /* send pages to client */
2067 rc = tgt_sendpage(tsi, rdpg, rc);
2072 for (i = 0; i < rdpg->rp_npages; i++)
2073 if (rdpg->rp_pages[i] != NULL)
2074 __free_page(rdpg->rp_pages[i]);
2075 OBD_FREE(rdpg->rp_pages, rdpg->rp_npages * sizeof rdpg->rp_pages[0]);
2077 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE))
2083 static int mdt_fix_attr_ucred(struct mdt_thread_info *info, __u32 op)
2085 struct lu_ucred *uc = mdt_ucred_check(info);
2086 struct lu_attr *attr = &info->mti_attr.ma_attr;
2091 if (op != REINT_SETATTR) {
2092 if ((attr->la_valid & LA_UID) && (attr->la_uid != -1))
2093 attr->la_uid = uc->uc_fsuid;
2094 /* for S_ISGID, inherit gid from his parent, such work will be
2095 * done in cmm/mdd layer, here set all cases as uc->uc_fsgid. */
2096 if ((attr->la_valid & LA_GID) && (attr->la_gid != -1))
2097 attr->la_gid = uc->uc_fsgid;
2103 static inline bool mdt_is_readonly_open(struct mdt_thread_info *info, __u32 op)
2105 return op == REINT_OPEN &&
2106 !(info->mti_spec.sp_cr_flags & (MDS_FMODE_WRITE | MDS_OPEN_CREAT));
2109 static void mdt_preset_secctx_size(struct mdt_thread_info *info)
2111 struct req_capsule *pill = info->mti_pill;
2113 if (req_capsule_has_field(pill, &RMF_FILE_SECCTX,
2115 req_capsule_has_field(pill, &RMF_FILE_SECCTX_NAME,
2117 if (req_capsule_get_size(pill, &RMF_FILE_SECCTX_NAME,
2119 /* pre-set size in server part with max size */
2120 req_capsule_set_size(pill, &RMF_FILE_SECCTX,
2122 info->mti_mdt->mdt_max_ea_size);
2124 req_capsule_set_size(pill, &RMF_FILE_SECCTX,
2131 static int mdt_reint_internal(struct mdt_thread_info *info,
2132 struct mdt_lock_handle *lhc,
2135 struct req_capsule *pill = info->mti_pill;
2136 struct mdt_body *repbody;
2141 rc = mdt_reint_unpack(info, op);
2143 CERROR("Can't unpack reint, rc %d\n", rc);
2144 RETURN(err_serious(rc));
2148 /* check if the file system is set to readonly. O_RDONLY open
2149 * is still allowed even the file system is set to readonly mode */
2150 if (mdt_rdonly(info->mti_exp) && !mdt_is_readonly_open(info, op))
2151 RETURN(err_serious(-EROFS));
2153 /* for replay (no_create) lmm is not needed, client has it already */
2154 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
2155 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER,
2158 /* llog cookies are always 0, the field is kept for compatibility */
2159 if (req_capsule_has_field(pill, &RMF_LOGCOOKIES, RCL_SERVER))
2160 req_capsule_set_size(pill, &RMF_LOGCOOKIES, RCL_SERVER, 0);
2162 /* Set ACL reply buffer size as LUSTRE_POSIX_ACL_MAX_SIZE_OLD
2163 * by default. If the target object has more ACL entries, then
2164 * enlarge the buffer when necessary. */
2165 if (req_capsule_has_field(pill, &RMF_ACL, RCL_SERVER))
2166 req_capsule_set_size(pill, &RMF_ACL, RCL_SERVER,
2167 LUSTRE_POSIX_ACL_MAX_SIZE_OLD);
2169 mdt_preset_secctx_size(info);
2171 rc = req_capsule_server_pack(pill);
2173 CERROR("Can't pack response, rc %d\n", rc);
2174 RETURN(err_serious(rc));
2177 if (req_capsule_has_field(pill, &RMF_MDT_BODY, RCL_SERVER)) {
2178 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
2180 repbody->mbo_eadatasize = 0;
2181 repbody->mbo_aclsize = 0;
2184 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_REINT_DELAY, 10);
2186 /* for replay no cookkie / lmm need, because client have this already */
2187 if (info->mti_spec.no_create)
2188 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
2189 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER, 0);
2191 rc = mdt_init_ucred_reint(info);
2193 GOTO(out_shrink, rc);
2195 rc = mdt_fix_attr_ucred(info, op);
2197 GOTO(out_ucred, rc = err_serious(rc));
2199 rc = mdt_check_resent(info, mdt_reconstruct, lhc);
2201 GOTO(out_ucred, rc);
2202 } else if (rc == 1) {
2203 DEBUG_REQ(D_INODE, mdt_info_req(info), "resent opt.");
2204 rc = lustre_msg_get_status(mdt_info_req(info)->rq_repmsg);
2205 GOTO(out_ucred, rc);
2207 rc = mdt_reint_rec(info, lhc);
2210 mdt_exit_ucred(info);
2212 mdt_client_compatibility(info);
2214 rc2 = mdt_fix_reply(info);
2219 * Data-on-MDT optimization - read data along with OPEN and return it
2220 * in reply. Do that only if we have both DOM and LAYOUT locks.
2222 if (rc == 0 && op == REINT_OPEN && !req_is_replay(pill->rc_req) &&
2223 info->mti_attr.ma_lmm != NULL &&
2224 mdt_lmm_dom_entry(info->mti_attr.ma_lmm) == LMM_DOM_ONLY) {
2225 rc = mdt_dom_read_on_open(info, info->mti_mdt,
2232 static long mdt_reint_opcode(struct ptlrpc_request *req,
2233 const struct req_format **fmt)
2235 struct mdt_device *mdt;
2236 struct mdt_rec_reint *rec;
2239 rec = req_capsule_client_get(&req->rq_pill, &RMF_REC_REINT);
2241 opc = rec->rr_opcode;
2242 DEBUG_REQ(D_INODE, req, "reint opt = %ld", opc);
2243 if (opc < REINT_MAX && fmt[opc] != NULL)
2244 req_capsule_extend(&req->rq_pill, fmt[opc]);
2246 mdt = mdt_exp2dev(req->rq_export);
2247 CERROR("%s: Unsupported opcode '%ld' from client '%s':"
2248 " rc = %d\n", req->rq_export->exp_obd->obd_name,
2249 opc, mdt->mdt_ldlm_client->cli_name, -EFAULT);
2250 opc = err_serious(-EFAULT);
2253 opc = err_serious(-EFAULT);
2258 static int mdt_reint(struct tgt_session_info *tsi)
2262 static const struct req_format *reint_fmts[REINT_MAX] = {
2263 [REINT_SETATTR] = &RQF_MDS_REINT_SETATTR,
2264 [REINT_CREATE] = &RQF_MDS_REINT_CREATE,
2265 [REINT_LINK] = &RQF_MDS_REINT_LINK,
2266 [REINT_UNLINK] = &RQF_MDS_REINT_UNLINK,
2267 [REINT_RENAME] = &RQF_MDS_REINT_RENAME,
2268 [REINT_OPEN] = &RQF_MDS_REINT_OPEN,
2269 [REINT_SETXATTR] = &RQF_MDS_REINT_SETXATTR,
2270 [REINT_RMENTRY] = &RQF_MDS_REINT_UNLINK,
2271 [REINT_MIGRATE] = &RQF_MDS_REINT_MIGRATE,
2272 [REINT_RESYNC] = &RQF_MDS_REINT_RESYNC,
2277 opc = mdt_reint_opcode(tgt_ses_req(tsi), reint_fmts);
2279 struct mdt_thread_info *info = tsi2mdt_info(tsi);
2281 * No lock possible here from client to pass it to reint code
2284 rc = mdt_reint_internal(info, NULL, opc);
2285 mdt_thread_info_fini(info);
2290 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_REINT_NET_REP;
2294 /* this should sync the whole device */
2295 int mdt_device_sync(const struct lu_env *env, struct mdt_device *mdt)
2297 struct dt_device *dt = mdt->mdt_bottom;
2301 rc = dt->dd_ops->dt_sync(env, dt);
2305 /* this should sync this object */
2306 static int mdt_object_sync(const struct lu_env *env, struct obd_export *exp,
2307 struct mdt_object *mo)
2313 if (!mdt_object_exists(mo)) {
2314 CWARN("%s: non existing object "DFID": rc = %d\n",
2315 exp->exp_obd->obd_name, PFID(mdt_object_fid(mo)),
2320 rc = mo_object_sync(env, mdt_object_child(mo));
2325 static int mdt_sync(struct tgt_session_info *tsi)
2327 struct ptlrpc_request *req = tgt_ses_req(tsi);
2328 struct req_capsule *pill = tsi->tsi_pill;
2329 struct mdt_body *body;
2334 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SYNC_PACK))
2335 RETURN(err_serious(-ENOMEM));
2337 if (fid_seq(&tsi->tsi_mdt_body->mbo_fid1) == 0) {
2338 rc = mdt_device_sync(tsi->tsi_env, mdt_exp2dev(tsi->tsi_exp));
2340 struct mdt_thread_info *info = tsi2mdt_info(tsi);
2342 /* sync an object */
2343 rc = mdt_object_sync(tsi->tsi_env, tsi->tsi_exp,
2346 const struct lu_fid *fid;
2347 struct lu_attr *la = &info->mti_attr.ma_attr;
2349 info->mti_attr.ma_need = MA_INODE;
2350 info->mti_attr.ma_valid = 0;
2351 rc = mdt_attr_get_complex(info, info->mti_object,
2354 body = req_capsule_server_get(pill,
2356 fid = mdt_object_fid(info->mti_object);
2357 mdt_pack_attr2body(info, body, la, fid);
2360 mdt_thread_info_fini(info);
2363 mdt_counter_incr(req, LPROC_MDT_SYNC);
2368 static int mdt_data_sync(struct tgt_session_info *tsi)
2370 struct mdt_thread_info *info;
2371 struct mdt_device *mdt = mdt_exp2dev(tsi->tsi_exp);
2372 struct ost_body *body = tsi->tsi_ost_body;
2373 struct ost_body *repbody;
2374 struct mdt_object *mo = NULL;
2380 repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_OST_BODY);
2382 /* if no fid is specified then do nothing,
2383 * device sync is done via MDS_SYNC */
2384 if (fid_is_zero(&tsi->tsi_fid))
2387 mo = mdt_object_find(tsi->tsi_env, mdt, &tsi->tsi_fid);
2389 RETURN(PTR_ERR(mo));
2391 rc = mdt_object_sync(tsi->tsi_env, tsi->tsi_exp, mo);
2395 repbody->oa.o_oi = body->oa.o_oi;
2396 repbody->oa.o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
2398 info = tsi2mdt_info(tsi);
2399 ma = &info->mti_attr;
2400 ma->ma_need = MA_INODE;
2402 rc = mdt_attr_get_complex(info, mo, ma);
2404 obdo_from_la(&repbody->oa, &ma->ma_attr, VALID_FLAGS);
2407 mdt_thread_info_fini(info);
2412 mdt_object_put(tsi->tsi_env, mo);
2417 * Handle quota control requests to consult current usage/limit, but also
2418 * to configure quota enforcement
2420 static int mdt_quotactl(struct tgt_session_info *tsi)
2422 struct obd_export *exp = tsi->tsi_exp;
2423 struct req_capsule *pill = tsi->tsi_pill;
2424 struct obd_quotactl *oqctl, *repoqc;
2426 struct mdt_device *mdt = mdt_exp2dev(exp);
2427 struct lu_device *qmt = mdt->mdt_qmt_dev;
2428 struct lu_nodemap *nodemap;
2431 oqctl = req_capsule_client_get(pill, &RMF_OBD_QUOTACTL);
2433 RETURN(err_serious(-EPROTO));
2435 rc = req_capsule_server_pack(pill);
2437 RETURN(err_serious(rc));
2439 nodemap = nodemap_get_from_exp(exp);
2440 if (IS_ERR(nodemap))
2441 RETURN(PTR_ERR(nodemap));
2443 switch (oqctl->qc_cmd) {
2444 /* master quotactl */
2447 case LUSTRE_Q_SETDEFAULT:
2448 if (!nodemap_can_setquota(nodemap))
2449 GOTO(out_nodemap, rc = -EPERM);
2452 case LUSTRE_Q_GETDEFAULT:
2454 GOTO(out_nodemap, rc = -EOPNOTSUPP);
2455 /* slave quotactl */
2460 CERROR("Unsupported quotactl command: %d\n", oqctl->qc_cmd);
2461 GOTO(out_nodemap, rc = -EFAULT);
2465 switch (oqctl->qc_type) {
2467 id = nodemap_map_id(nodemap, NODEMAP_UID,
2468 NODEMAP_CLIENT_TO_FS, id);
2471 id = nodemap_map_id(nodemap, NODEMAP_GID,
2472 NODEMAP_CLIENT_TO_FS, id);
2475 /* todo: check/map project id */
2479 GOTO(out_nodemap, rc = -EOPNOTSUPP);
2481 repoqc = req_capsule_server_get(pill, &RMF_OBD_QUOTACTL);
2483 GOTO(out_nodemap, rc = err_serious(-EFAULT));
2485 if (oqctl->qc_cmd == Q_SETINFO || oqctl->qc_cmd == Q_SETQUOTA)
2486 barrier_exit(tsi->tsi_tgt->lut_bottom);
2488 if (oqctl->qc_id != id)
2489 swap(oqctl->qc_id, id);
2491 if (oqctl->qc_cmd == Q_SETINFO || oqctl->qc_cmd == Q_SETQUOTA) {
2492 if (unlikely(!barrier_entry(tsi->tsi_tgt->lut_bottom)))
2493 RETURN(-EINPROGRESS);
2496 switch (oqctl->qc_cmd) {
2502 case LUSTRE_Q_SETDEFAULT:
2503 case LUSTRE_Q_GETDEFAULT:
2504 /* forward quotactl request to QMT */
2505 rc = qmt_hdls.qmth_quotactl(tsi->tsi_env, qmt, oqctl);
2510 /* slave quotactl */
2511 rc = lquotactl_slv(tsi->tsi_env, tsi->tsi_tgt->lut_bottom,
2516 CERROR("Unsupported quotactl command: %d\n", oqctl->qc_cmd);
2517 GOTO(out_nodemap, rc = -EFAULT);
2520 if (oqctl->qc_id != id)
2521 swap(oqctl->qc_id, id);
2528 nodemap_putref(nodemap);
2533 /** clone llog ctxt from child (mdd)
2534 * This allows remote llog (replicator) access.
2535 * We can either pass all llog RPCs (eg mdt_llog_create) on to child where the
2536 * context was originally set up, or we can handle them directly.
2537 * I choose the latter, but that means I need any llog
2538 * contexts set up by child to be accessable by the mdt. So we clone the
2539 * context into our context list here.
2541 static int mdt_llog_ctxt_clone(const struct lu_env *env, struct mdt_device *mdt,
2544 struct md_device *next = mdt->mdt_child;
2545 struct llog_ctxt *ctxt;
2548 if (!llog_ctxt_null(mdt2obd_dev(mdt), idx))
2551 rc = next->md_ops->mdo_llog_ctxt_get(env, next, idx, (void **)&ctxt);
2552 if (rc || ctxt == NULL) {
2556 rc = llog_group_set_ctxt(&mdt2obd_dev(mdt)->obd_olg, ctxt, idx);
2558 CERROR("Can't set mdt ctxt %d\n", rc);
2563 static int mdt_llog_ctxt_unclone(const struct lu_env *env,
2564 struct mdt_device *mdt, int idx)
2566 struct llog_ctxt *ctxt;
2568 ctxt = llog_get_context(mdt2obd_dev(mdt), idx);
2571 /* Put once for the get we just did, and once for the clone */
2572 llog_ctxt_put(ctxt);
2573 llog_ctxt_put(ctxt);
2578 * sec context handlers
2580 static int mdt_sec_ctx_handle(struct tgt_session_info *tsi)
2582 CFS_FAIL_TIMEOUT(OBD_FAIL_SEC_CTX_HDL_PAUSE, cfs_fail_val);
2588 * quota request handlers
2590 static int mdt_quota_dqacq(struct tgt_session_info *tsi)
2592 struct mdt_device *mdt = mdt_exp2dev(tsi->tsi_exp);
2593 struct lu_device *qmt = mdt->mdt_qmt_dev;
2598 RETURN(err_serious(-EOPNOTSUPP));
2600 rc = qmt_hdls.qmth_dqacq(tsi->tsi_env, qmt, tgt_ses_req(tsi));
2604 struct mdt_object *mdt_object_new(const struct lu_env *env,
2605 struct mdt_device *d,
2606 const struct lu_fid *f)
2608 struct lu_object_conf conf = { .loc_flags = LOC_F_NEW };
2609 struct lu_object *o;
2610 struct mdt_object *m;
2613 CDEBUG(D_INFO, "Allocate object for "DFID"\n", PFID(f));
2614 o = lu_object_find(env, &d->mdt_lu_dev, f, &conf);
2615 if (unlikely(IS_ERR(o)))
2616 m = (struct mdt_object *)o;
2622 struct mdt_object *mdt_object_find(const struct lu_env *env,
2623 struct mdt_device *d,
2624 const struct lu_fid *f)
2626 struct lu_object *o;
2627 struct mdt_object *m;
2630 CDEBUG(D_INFO, "Find object for "DFID"\n", PFID(f));
2631 o = lu_object_find(env, &d->mdt_lu_dev, f, NULL);
2632 if (unlikely(IS_ERR(o)))
2633 m = (struct mdt_object *)o;
2641 * Asyncronous commit for mdt device.
2643 * Pass asynchonous commit call down the MDS stack.
2645 * \param env environment
2646 * \param mdt the mdt device
2648 static void mdt_device_commit_async(const struct lu_env *env,
2649 struct mdt_device *mdt)
2651 struct dt_device *dt = mdt->mdt_bottom;
2655 rc = dt->dd_ops->dt_commit_async(env, dt);
2656 if (unlikely(rc != 0))
2657 CWARN("%s: async commit start failed: rc = %d\n",
2658 mdt_obd_name(mdt), rc);
2659 atomic_inc(&mdt->mdt_async_commit_count);
2664 * Mark the lock as "synchonous".
2666 * Mark the lock to deffer transaction commit to the unlock time.
2668 * \param lock the lock to mark as "synchonous"
2670 * \see mdt_is_lock_sync
2671 * \see mdt_save_lock
2673 static inline void mdt_set_lock_sync(struct ldlm_lock *lock)
2675 lock->l_ast_data = (void*)1;
2679 * Check whehter the lock "synchonous" or not.
2681 * \param lock the lock to check
2682 * \retval 1 the lock is "synchonous"
2683 * \retval 0 the lock isn't "synchronous"
2685 * \see mdt_set_lock_sync
2686 * \see mdt_save_lock
2688 static inline int mdt_is_lock_sync(struct ldlm_lock *lock)
2690 return lock->l_ast_data != NULL;
2694 * Blocking AST for mdt locks.
2696 * Starts transaction commit if in case of COS lock conflict or
2697 * deffers such a commit to the mdt_save_lock.
2699 * \param lock the lock which blocks a request or cancelling lock
2700 * \param desc unused
2701 * \param data unused
2702 * \param flag indicates whether this cancelling or blocking callback
2704 * \see ldlm_blocking_ast_nocheck
2706 int mdt_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
2707 void *data, int flag)
2709 struct obd_device *obd = ldlm_lock_to_ns(lock)->ns_obd;
2710 struct mdt_device *mdt = mdt_dev(obd->obd_lu_dev);
2711 struct ldlm_cb_set_arg *arg = data;
2712 bool commit_async = false;
2716 if (flag == LDLM_CB_CANCELING)
2719 lock_res_and_lock(lock);
2720 if (lock->l_blocking_ast != mdt_blocking_ast) {
2721 unlock_res_and_lock(lock);
2725 /* A blocking ast may be sent from ldlm_lock_decref_internal
2726 * when the last reference to a local lock was released and
2727 * during blocking event from ldlm_work_bl_ast_lock().
2728 * The 'data' parameter is l_ast_data in the first case and
2729 * callback arguments in the second one. Distinguish them by that.
2731 if (!data || data == lock->l_ast_data || !arg->bl_desc)
2732 goto skip_cos_checks;
2734 if (lock->l_req_mode & (LCK_PW | LCK_EX)) {
2735 if (mdt_cos_is_enabled(mdt)) {
2736 if (!arg->bl_desc->bl_same_client)
2737 mdt_set_lock_sync(lock);
2738 } else if (mdt_slc_is_enabled(mdt) &&
2739 arg->bl_desc->bl_cos_incompat) {
2740 mdt_set_lock_sync(lock);
2742 * we may do extra commit here, but there is a small
2743 * window to miss a commit: lock was unlocked (saved),
2744 * then a conflict lock queued and we come here, but
2745 * REP-ACK not received, so lock was not converted to
2747 * Fortunately this window is quite small, so the
2748 * extra commit should be rare (not to say distributed
2749 * operation is rare too).
2751 commit_async = true;
2753 } else if (lock->l_req_mode == LCK_COS) {
2754 commit_async = true;
2758 rc = ldlm_blocking_ast_nocheck(lock);
2763 rc = lu_env_init(&env, LCT_LOCAL);
2764 if (unlikely(rc != 0))
2765 CWARN("%s: lu_env initialization failed, cannot "
2766 "start asynchronous commit: rc = %d\n",
2769 mdt_device_commit_async(&env, mdt);
2776 * Blocking AST for cross-MDT lock
2778 * Discard lock from uncommitted_slc_locks and cancel it.
2780 * \param lock the lock which blocks a request or cancelling lock
2781 * \param desc unused
2782 * \param data unused
2783 * \param flag indicates whether this cancelling or blocking callback
2784 * \retval 0 on success
2785 * \retval negative number on error
2787 int mdt_remote_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
2788 void *data, int flag)
2794 case LDLM_CB_BLOCKING: {
2795 struct lustre_handle lockh;
2797 ldlm_lock2handle(lock, &lockh);
2798 rc = ldlm_cli_cancel(&lockh,
2799 ldlm_is_atomic_cb(lock) ? 0 : LCF_ASYNC);
2801 CDEBUG(D_INODE, "ldlm_cli_cancel: %d\n", rc);
2806 case LDLM_CB_CANCELING: {
2807 struct obd_device *obd = ldlm_lock_to_ns(lock)->ns_obd;
2808 struct mdt_device *mdt =
2809 mdt_dev(obd->obd_lu_dev->ld_site->ls_top_dev);
2811 LDLM_DEBUG(lock, "Revoke remote lock\n");
2813 /* discard slc lock here so that it can be cleaned anytime,
2814 * especially for cleanup_resource() */
2815 tgt_discard_slc_lock(&mdt->mdt_lut, lock);
2817 /* once we cache lock, l_ast_data is set to mdt_object */
2818 if (lock->l_ast_data != NULL) {
2819 struct mdt_object *mo = lock->l_ast_data;
2822 rc = lu_env_init(&env, LCT_MD_THREAD);
2823 if (unlikely(rc != 0)) {
2824 CWARN("%s: lu_env initialization failed, object"
2825 "%p "DFID" is leaked!\n",
2827 PFID(mdt_object_fid(mo)));
2831 if (lock->l_policy_data.l_inodebits.bits &
2832 (MDS_INODELOCK_XATTR | MDS_INODELOCK_UPDATE)) {
2833 rc = mo_invalidate(&env, mdt_object_child(mo));
2834 mo->mot_cache_attr = 0;
2836 mdt_object_put(&env, mo);
2848 int mdt_check_resent_lock(struct mdt_thread_info *info,
2849 struct mdt_object *mo,
2850 struct mdt_lock_handle *lhc)
2852 /* the lock might already be gotten in ldlm_handle_enqueue() */
2853 if (unlikely(lustre_handle_is_used(&lhc->mlh_reg_lh))) {
2854 struct ptlrpc_request *req = mdt_info_req(info);
2855 struct ldlm_lock *lock;
2857 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
2858 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
2860 /* Lock is pinned by ldlm_handle_enqueue0() as it is
2861 * a resend case, however, it could be already destroyed
2862 * due to client eviction or a raced cancel RPC. */
2863 LDLM_DEBUG_NOLOCK("Invalid lock handle %#llx",
2864 lhc->mlh_reg_lh.cookie);
2868 if (!fid_res_name_eq(mdt_object_fid(mo),
2869 &lock->l_resource->lr_name)) {
2870 CWARN("%s: Although resent, but still not "
2871 "get child lock:"DFID"\n",
2872 info->mti_exp->exp_obd->obd_name,
2873 PFID(mdt_object_fid(mo)));
2874 LDLM_LOCK_PUT(lock);
2877 LDLM_LOCK_PUT(lock);
2883 static void mdt_remote_object_lock_created_cb(struct ldlm_lock *lock)
2885 mdt_object_get(NULL, lock->l_ast_data);
2888 int mdt_remote_object_lock_try(struct mdt_thread_info *mti,
2889 struct mdt_object *o, const struct lu_fid *fid,
2890 struct lustre_handle *lh, enum ldlm_mode mode,
2891 __u64 *ibits, __u64 trybits, bool cache)
2893 struct ldlm_enqueue_info *einfo = &mti->mti_remote_einfo;
2894 union ldlm_policy_data *policy = &mti->mti_policy;
2895 struct ldlm_res_id *res_id = &mti->mti_res_id;
2899 LASSERT(mdt_object_remote(o));
2901 fid_build_reg_res_name(fid, res_id);
2903 memset(einfo, 0, sizeof(*einfo));
2904 einfo->ei_type = LDLM_IBITS;
2905 einfo->ei_mode = mode;
2906 einfo->ei_cb_bl = mdt_remote_blocking_ast;
2907 einfo->ei_cb_cp = ldlm_completion_ast;
2908 einfo->ei_enq_slave = 0;
2909 einfo->ei_res_id = res_id;
2913 * if we cache lock, couple lock with mdt_object, so that object
2914 * can be easily found in lock ASTs.
2916 einfo->ei_cbdata = o;
2917 einfo->ei_cb_created = mdt_remote_object_lock_created_cb;
2920 memset(policy, 0, sizeof(*policy));
2921 policy->l_inodebits.bits = *ibits;
2922 policy->l_inodebits.try_bits = trybits;
2924 rc = mo_object_lock(mti->mti_env, mdt_object_child(o), lh, einfo,
2927 /* Return successfully acquired bits to a caller */
2929 struct ldlm_lock *lock = ldlm_handle2lock(lh);
2932 *ibits = lock->l_policy_data.l_inodebits.bits;
2933 LDLM_LOCK_PUT(lock);
2938 int mdt_remote_object_lock(struct mdt_thread_info *mti, struct mdt_object *o,
2939 const struct lu_fid *fid, struct lustre_handle *lh,
2940 enum ldlm_mode mode, __u64 ibits, bool cache)
2942 return mdt_remote_object_lock_try(mti, o, fid, lh, mode, &ibits, 0,
2946 static int mdt_object_local_lock(struct mdt_thread_info *info,
2947 struct mdt_object *o,
2948 struct mdt_lock_handle *lh, __u64 *ibits,
2949 __u64 trybits, bool cos_incompat)
2951 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
2952 union ldlm_policy_data *policy = &info->mti_policy;
2953 struct ldlm_res_id *res_id = &info->mti_res_id;
2954 __u64 dlmflags = 0, *cookie = NULL;
2958 LASSERT(!lustre_handle_is_used(&lh->mlh_reg_lh));
2959 LASSERT(!lustre_handle_is_used(&lh->mlh_pdo_lh));
2960 LASSERT(lh->mlh_reg_mode != LCK_MINMODE);
2961 LASSERT(lh->mlh_type != MDT_NUL_LOCK);
2964 LASSERT(lh->mlh_reg_mode == LCK_PW ||
2965 lh->mlh_reg_mode == LCK_EX);
2966 dlmflags |= LDLM_FL_COS_INCOMPAT;
2967 } else if (mdt_cos_is_enabled(info->mti_mdt)) {
2968 dlmflags |= LDLM_FL_COS_ENABLED;
2971 /* Only enqueue LOOKUP lock for remote object */
2972 LASSERT(ergo(mdt_object_remote(o), *ibits == MDS_INODELOCK_LOOKUP));
2974 /* Lease lock are granted with LDLM_FL_CANCEL_ON_BLOCK */
2975 if (lh->mlh_type == MDT_REG_LOCK && lh->mlh_reg_mode == LCK_EX &&
2976 *ibits == MDS_INODELOCK_OPEN)
2977 dlmflags |= LDLM_FL_CANCEL_ON_BLOCK;
2979 if (lh->mlh_type == MDT_PDO_LOCK) {
2980 /* check for exists after object is locked */
2981 if (mdt_object_exists(o) == 0) {
2982 /* Non-existent object shouldn't have PDO lock */
2985 /* Non-dir object shouldn't have PDO lock */
2986 if (!S_ISDIR(lu_object_attr(&o->mot_obj)))
2991 fid_build_reg_res_name(mdt_object_fid(o), res_id);
2992 dlmflags |= LDLM_FL_ATOMIC_CB;
2995 cookie = &info->mti_exp->exp_handle.h_cookie;
2998 * Take PDO lock on whole directory and build correct @res_id for lock
2999 * on part of directory.
3001 if (lh->mlh_pdo_hash != 0) {
3002 LASSERT(lh->mlh_type == MDT_PDO_LOCK);
3003 mdt_lock_pdo_mode(info, o, lh);
3004 if (lh->mlh_pdo_mode != LCK_NL) {
3006 * Do not use LDLM_FL_LOCAL_ONLY for parallel lock, it
3007 * is never going to be sent to client and we do not
3008 * want it slowed down due to possible cancels.
3010 policy->l_inodebits.bits =
3011 *ibits & MDS_INODELOCK_UPDATE;
3012 policy->l_inodebits.try_bits =
3013 trybits & MDS_INODELOCK_UPDATE;
3014 /* at least one of them should be set */
3015 LASSERT(policy->l_inodebits.bits |
3016 policy->l_inodebits.try_bits);
3017 rc = mdt_fid_lock(info->mti_env, ns, &lh->mlh_pdo_lh,
3018 lh->mlh_pdo_mode, policy, res_id,
3020 if (unlikely(rc != 0))
3021 GOTO(out_unlock, rc);
3025 * Finish res_id initializing by name hash marking part of
3026 * directory which is taking modification.
3028 res_id->name[LUSTRE_RES_ID_HSH_OFF] = lh->mlh_pdo_hash;
3031 policy->l_inodebits.bits = *ibits;
3032 policy->l_inodebits.try_bits = trybits;
3035 * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
3036 * going to be sent to client. If it is - mdt_intent_policy() path will
3037 * fix it up and turn FL_LOCAL flag off.
3039 rc = mdt_fid_lock(info->mti_env, ns, &lh->mlh_reg_lh, lh->mlh_reg_mode,
3040 policy, res_id, LDLM_FL_LOCAL_ONLY | dlmflags,
3044 mdt_object_unlock(info, o, lh, 1);
3045 else if (unlikely(OBD_FAIL_PRECHECK(OBD_FAIL_MDS_PDO_LOCK)) &&
3046 lh->mlh_pdo_hash != 0 &&
3047 (lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX))
3048 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK, 15);
3050 /* Return successfully acquired bits to a caller */
3052 struct ldlm_lock *lock = ldlm_handle2lock(&lh->mlh_reg_lh);
3055 *ibits = lock->l_policy_data.l_inodebits.bits;
3056 LDLM_LOCK_PUT(lock);
3062 mdt_object_lock_internal(struct mdt_thread_info *info, struct mdt_object *o,
3063 struct mdt_lock_handle *lh, __u64 *ibits,
3064 __u64 trybits, bool cos_incompat)
3066 struct mdt_lock_handle *local_lh = NULL;
3070 if (!mdt_object_remote(o)) {
3071 rc = mdt_object_local_lock(info, o, lh, ibits, trybits,
3076 /* XXX do not support PERM/LAYOUT/XATTR lock for remote object yet */
3077 *ibits &= ~(MDS_INODELOCK_PERM | MDS_INODELOCK_LAYOUT |
3078 MDS_INODELOCK_XATTR);
3080 /* Only enqueue LOOKUP lock for remote object */
3081 if (*ibits & MDS_INODELOCK_LOOKUP) {
3082 __u64 local = MDS_INODELOCK_LOOKUP;
3084 rc = mdt_object_local_lock(info, o, lh, &local, 0,
3092 if ((*ibits | trybits) & MDS_INODELOCK_UPDATE) {
3093 /* Sigh, PDO needs to enqueue 2 locks right now, but
3094 * enqueue RPC can only request 1 lock, to avoid extra
3095 * RPC, so it will instead enqueue EX lock for remote
3096 * object anyway XXX*/
3097 if (lh->mlh_type == MDT_PDO_LOCK &&
3098 lh->mlh_pdo_hash != 0) {
3099 CDEBUG(D_INFO, "%s: "DFID" convert PDO lock to"
3100 "EX lock.\n", mdt_obd_name(info->mti_mdt),
3101 PFID(mdt_object_fid(o)));
3102 lh->mlh_pdo_hash = 0;
3103 lh->mlh_rreg_mode = LCK_EX;
3104 lh->mlh_type = MDT_REG_LOCK;
3107 rc = mdt_remote_object_lock_try(info, o, mdt_object_fid(o),
3110 ibits, trybits, false);
3111 if (rc != ELDLM_OK) {
3112 if (local_lh != NULL)
3113 mdt_object_unlock(info, o, local_lh, rc);
3118 /* other components like LFSCK can use lockless access
3119 * and populate cache, so we better invalidate it */
3120 mo_invalidate(info->mti_env, mdt_object_child(o));
3125 int mdt_object_lock(struct mdt_thread_info *info, struct mdt_object *o,
3126 struct mdt_lock_handle *lh, __u64 ibits)
3128 return mdt_object_lock_internal(info, o, lh, &ibits, 0, false);
3131 int mdt_reint_object_lock(struct mdt_thread_info *info, struct mdt_object *o,
3132 struct mdt_lock_handle *lh, __u64 ibits,
3135 LASSERT(lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX);
3136 return mdt_object_lock_internal(info, o, lh, &ibits, 0,
3140 int mdt_object_lock_try(struct mdt_thread_info *info, struct mdt_object *o,
3141 struct mdt_lock_handle *lh, __u64 *ibits,
3142 __u64 trybits, bool cos_incompat)
3144 bool trylock_only = *ibits == 0;
3147 LASSERT(!(*ibits & trybits));
3148 rc = mdt_object_lock_internal(info, o, lh, ibits, trybits,
3150 if (rc && trylock_only) { /* clear error for try ibits lock only */
3151 LASSERT(*ibits == 0);
3158 * Save a lock within request object.
3160 * Keep the lock referenced until whether client ACK or transaction
3161 * commit happens or release the lock immediately depending on input
3162 * parameters. If COS is ON, a write lock is converted to COS lock
3165 * \param info thead info object
3166 * \param h lock handle
3167 * \param mode lock mode
3168 * \param decref force immediate lock releasing
3170 void mdt_save_lock(struct mdt_thread_info *info, struct lustre_handle *h,
3171 enum ldlm_mode mode, int decref)
3175 if (lustre_handle_is_used(h)) {
3176 if (decref || !info->mti_has_trans ||
3177 !(mode & (LCK_PW | LCK_EX))) {
3178 mdt_fid_unlock(h, mode);
3180 struct mdt_device *mdt = info->mti_mdt;
3181 struct ldlm_lock *lock = ldlm_handle2lock(h);
3182 struct ptlrpc_request *req = mdt_info_req(info);
3183 bool cos = mdt_cos_is_enabled(mdt);
3184 bool convert_lock = !cos && mdt_slc_is_enabled(mdt);
3186 LASSERTF(lock != NULL, "no lock for cookie %#llx\n",
3189 /* there is no request if mdt_object_unlock() is called
3190 * from mdt_export_cleanup()->mdt_add_dirty_flag() */
3191 if (likely(req != NULL)) {
3192 LDLM_DEBUG(lock, "save lock request %p reply "
3193 "state %p transno %lld\n", req,
3194 req->rq_reply_state, req->rq_transno);
3196 ldlm_lock_mode_downgrade(lock, LCK_COS);
3199 if (req->rq_export->exp_disconnected)
3200 mdt_fid_unlock(h, mode);
3202 ptlrpc_save_lock(req, h, mode, cos,
3205 mdt_fid_unlock(h, mode);
3207 if (mdt_is_lock_sync(lock)) {
3208 CDEBUG(D_HA, "found sync-lock,"
3209 " async commit started\n");
3210 mdt_device_commit_async(info->mti_env,
3213 LDLM_LOCK_PUT(lock);
3222 * Save cross-MDT lock in uncommitted_slc_locks
3224 * Keep the lock referenced until transaction commit happens or release the lock
3225 * immediately depending on input parameters.
3227 * \param info thead info object
3228 * \param h lock handle
3229 * \param mode lock mode
3230 * \param decref force immediate lock releasing
3232 static void mdt_save_remote_lock(struct mdt_thread_info *info,
3233 struct mdt_object *o, struct lustre_handle *h,
3234 enum ldlm_mode mode, int decref)
3238 if (lustre_handle_is_used(h)) {
3239 struct ldlm_lock *lock = ldlm_handle2lock(h);
3242 (lock->l_policy_data.l_inodebits.bits &
3243 (MDS_INODELOCK_XATTR | MDS_INODELOCK_UPDATE)))
3244 mo_invalidate(info->mti_env, mdt_object_child(o));
3246 if (decref || !info->mti_has_trans ||
3247 !(mode & (LCK_PW | LCK_EX))) {
3248 ldlm_lock_decref_and_cancel(h, mode);
3249 LDLM_LOCK_PUT(lock);
3251 struct ptlrpc_request *req = mdt_info_req(info);
3253 LASSERT(req != NULL);
3254 tgt_save_slc_lock(&info->mti_mdt->mdt_lut, lock,
3256 ldlm_lock_decref(h, mode);
3265 * Unlock mdt object.
3267 * Immeditely release the regular lock and the PDO lock or save the
3268 * lock in request and keep them referenced until client ACK or
3269 * transaction commit.
3271 * \param info thread info object
3272 * \param o mdt object
3273 * \param lh mdt lock handle referencing regular and PDO locks
3274 * \param decref force immediate lock releasing
3276 * XXX o is not used and may be NULL, see hsm_cdt_request_completed().
3278 void mdt_object_unlock(struct mdt_thread_info *info, struct mdt_object *o,
3279 struct mdt_lock_handle *lh, int decref)
3283 mdt_save_lock(info, &lh->mlh_pdo_lh, lh->mlh_pdo_mode, decref);
3284 mdt_save_lock(info, &lh->mlh_reg_lh, lh->mlh_reg_mode, decref);
3285 mdt_save_remote_lock(info, o, &lh->mlh_rreg_lh, lh->mlh_rreg_mode,
3291 struct mdt_object *mdt_object_find_lock(struct mdt_thread_info *info,
3292 const struct lu_fid *f,
3293 struct mdt_lock_handle *lh,
3296 struct mdt_object *o;
3298 o = mdt_object_find(info->mti_env, info->mti_mdt, f);
3302 rc = mdt_object_lock(info, o, lh, ibits);
3304 mdt_object_put(info->mti_env, o);
3311 void mdt_object_unlock_put(struct mdt_thread_info * info,
3312 struct mdt_object * o,
3313 struct mdt_lock_handle *lh,
3316 mdt_object_unlock(info, o, lh, decref);
3317 mdt_object_put(info->mti_env, o);
3321 * Generic code handling requests that have struct mdt_body passed in:
3323 * - extract mdt_body from request and save it in @info, if present;
3325 * - create lu_object, corresponding to the fid in mdt_body, and save it in
3328 * - if HAS_BODY flag is set for this request type check whether object
3329 * actually exists on storage (lu_object_exists()).
3332 static int mdt_body_unpack(struct mdt_thread_info *info,
3333 enum tgt_handler_flags flags)
3335 const struct mdt_body *body;
3336 struct mdt_object *obj;
3337 const struct lu_env *env;
3338 struct req_capsule *pill;
3342 env = info->mti_env;
3343 pill = info->mti_pill;
3345 body = info->mti_body = req_capsule_client_get(pill, &RMF_MDT_BODY);
3349 if (!(body->mbo_valid & OBD_MD_FLID))
3352 if (!fid_is_sane(&body->mbo_fid1)) {
3353 CERROR("Invalid fid: "DFID"\n", PFID(&body->mbo_fid1));
3357 obj = mdt_object_find(env, info->mti_mdt, &body->mbo_fid1);
3359 if ((flags & HAS_BODY) && !mdt_object_exists(obj)) {
3360 mdt_object_put(env, obj);
3363 info->mti_object = obj;
3372 static int mdt_unpack_req_pack_rep(struct mdt_thread_info *info,
3373 enum tgt_handler_flags flags)
3375 struct req_capsule *pill = info->mti_pill;
3380 if (req_capsule_has_field(pill, &RMF_MDT_BODY, RCL_CLIENT))
3381 rc = mdt_body_unpack(info, flags);
3385 if (rc == 0 && (flags & HAS_REPLY)) {
3387 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
3388 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER,
3390 if (req_capsule_has_field(pill, &RMF_LOGCOOKIES, RCL_SERVER))
3391 req_capsule_set_size(pill, &RMF_LOGCOOKIES,
3394 /* Set ACL reply buffer size as LUSTRE_POSIX_ACL_MAX_SIZE_OLD
3395 * by default. If the target object has more ACL entries, then
3396 * enlarge the buffer when necessary. */
3397 if (req_capsule_has_field(pill, &RMF_ACL, RCL_SERVER))
3398 req_capsule_set_size(pill, &RMF_ACL, RCL_SERVER,
3399 LUSTRE_POSIX_ACL_MAX_SIZE_OLD);
3401 mdt_preset_secctx_size(info);
3403 rc = req_capsule_server_pack(pill);
3405 CWARN("%s: cannot pack response: rc = %d\n",
3406 mdt_obd_name(info->mti_mdt), rc);
3411 void mdt_lock_handle_init(struct mdt_lock_handle *lh)
3413 lh->mlh_type = MDT_NUL_LOCK;
3414 lh->mlh_reg_lh.cookie = 0ull;
3415 lh->mlh_reg_mode = LCK_MINMODE;
3416 lh->mlh_pdo_lh.cookie = 0ull;
3417 lh->mlh_pdo_mode = LCK_MINMODE;
3418 lh->mlh_rreg_lh.cookie = 0ull;
3419 lh->mlh_rreg_mode = LCK_MINMODE;
3422 void mdt_lock_handle_fini(struct mdt_lock_handle *lh)
3424 LASSERT(!lustre_handle_is_used(&lh->mlh_reg_lh));
3425 LASSERT(!lustre_handle_is_used(&lh->mlh_pdo_lh));
3429 * Initialize fields of struct mdt_thread_info. Other fields are left in
3430 * uninitialized state, because it's too expensive to zero out whole
3431 * mdt_thread_info (> 1K) on each request arrival.
3433 void mdt_thread_info_init(struct ptlrpc_request *req,
3434 struct mdt_thread_info *info)
3438 info->mti_pill = &req->rq_pill;
3441 for (i = 0; i < ARRAY_SIZE(info->mti_lh); i++)
3442 mdt_lock_handle_init(&info->mti_lh[i]);
3444 /* mdt device: it can be NULL while CONNECT */
3445 if (req->rq_export) {
3446 info->mti_mdt = mdt_dev(req->rq_export->exp_obd->obd_lu_dev);
3447 info->mti_exp = req->rq_export;
3449 info->mti_mdt = NULL;
3450 info->mti_env = req->rq_svc_thread->t_env;
3451 info->mti_transno = lustre_msg_get_transno(req->rq_reqmsg);
3453 memset(&info->mti_attr, 0, sizeof(info->mti_attr));
3454 info->mti_big_buf = LU_BUF_NULL;
3455 info->mti_body = NULL;
3456 info->mti_object = NULL;
3457 info->mti_dlm_req = NULL;
3458 info->mti_has_trans = 0;
3459 info->mti_cross_ref = 0;
3460 info->mti_opdata = 0;
3461 info->mti_big_lmm_used = 0;
3462 info->mti_big_acl_used = 0;
3463 info->mti_som_valid = 0;
3465 info->mti_spec.no_create = 0;
3466 info->mti_spec.sp_rm_entry = 0;
3467 info->mti_spec.sp_permitted = 0;
3468 info->mti_spec.sp_migrate_close = 0;
3470 info->mti_spec.u.sp_ea.eadata = NULL;
3471 info->mti_spec.u.sp_ea.eadatalen = 0;
3474 void mdt_thread_info_fini(struct mdt_thread_info *info)
3478 if (info->mti_object != NULL) {
3479 mdt_object_put(info->mti_env, info->mti_object);
3480 info->mti_object = NULL;
3483 for (i = 0; i < ARRAY_SIZE(info->mti_lh); i++)
3484 mdt_lock_handle_fini(&info->mti_lh[i]);
3485 info->mti_env = NULL;
3486 info->mti_pill = NULL;
3487 info->mti_exp = NULL;
3489 if (unlikely(info->mti_big_buf.lb_buf != NULL))
3490 lu_buf_free(&info->mti_big_buf);
3493 struct mdt_thread_info *tsi2mdt_info(struct tgt_session_info *tsi)
3495 struct mdt_thread_info *mti;
3497 mti = mdt_th_info(tsi->tsi_env);
3498 LASSERT(mti != NULL);
3500 mdt_thread_info_init(tgt_ses_req(tsi), mti);
3501 if (tsi->tsi_corpus != NULL) {
3502 mti->mti_object = mdt_obj(tsi->tsi_corpus);
3503 lu_object_get(tsi->tsi_corpus);
3505 mti->mti_body = tsi->tsi_mdt_body;
3506 mti->mti_dlm_req = tsi->tsi_dlm_req;
3511 static int mdt_tgt_connect(struct tgt_session_info *tsi)
3513 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_DELAY_CONDITIONAL) &&
3515 tsi2mdt_info(tsi)->mti_mdt->mdt_seq_site.ss_node_id) {
3516 set_current_state(TASK_UNINTERRUPTIBLE);
3517 schedule_timeout(msecs_to_jiffies(3 * MSEC_PER_SEC));
3520 return tgt_connect(tsi);
3523 static int mdt_intent_glimpse(enum ldlm_intent_flags it_opc,
3524 struct mdt_thread_info *info,
3525 struct ldlm_lock **lockp, __u64 flags)
3527 return mdt_glimpse_enqueue(info, info->mti_mdt->mdt_namespace,
3530 static int mdt_intent_brw(enum ldlm_intent_flags it_opc,
3531 struct mdt_thread_info *info,
3532 struct ldlm_lock **lockp, __u64 flags)
3534 return mdt_brw_enqueue(info, info->mti_mdt->mdt_namespace,
3538 int mdt_intent_lock_replace(struct mdt_thread_info *info,
3539 struct ldlm_lock **lockp,
3540 struct mdt_lock_handle *lh,
3541 __u64 flags, int result)
3543 struct ptlrpc_request *req = mdt_info_req(info);
3544 struct ldlm_lock *lock = *lockp;
3545 struct ldlm_lock *new_lock;
3547 /* If possible resent found a lock, @lh is set to its handle */
3548 new_lock = ldlm_handle2lock_long(&lh->mlh_reg_lh, 0);
3550 if (new_lock == NULL && (flags & LDLM_FL_INTENT_ONLY)) {
3551 lh->mlh_reg_lh.cookie = 0;
3555 if (new_lock == NULL && (flags & LDLM_FL_RESENT)) {
3556 /* Lock is pinned by ldlm_handle_enqueue0() as it is
3557 * a resend case, however, it could be already destroyed
3558 * due to client eviction or a raced cancel RPC. */
3559 LDLM_DEBUG_NOLOCK("Invalid lock handle %#llx\n",
3560 lh->mlh_reg_lh.cookie);
3561 lh->mlh_reg_lh.cookie = 0;
3565 LASSERTF(new_lock != NULL,
3566 "lockh %#llx flags %#llx : rc = %d\n",
3567 lh->mlh_reg_lh.cookie, flags, result);
3570 * If we've already given this lock to a client once, then we should
3571 * have no readers or writers. Otherwise, we should have one reader
3572 * _or_ writer ref (which will be zeroed below) before returning the
3575 if (new_lock->l_export == req->rq_export) {
3576 LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
3578 LASSERT(new_lock->l_export == NULL);
3579 LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
3584 if (new_lock->l_export == req->rq_export) {
3586 * Already gave this to the client, which means that we
3587 * reconstructed a reply.
3589 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
3592 LDLM_LOCK_RELEASE(new_lock);
3593 lh->mlh_reg_lh.cookie = 0;
3594 RETURN(ELDLM_LOCK_REPLACED);
3598 * Fixup the lock to be given to the client.
3600 lock_res_and_lock(new_lock);
3601 /* Zero new_lock->l_readers and new_lock->l_writers without triggering
3602 * possible blocking AST. */
3603 while (new_lock->l_readers > 0) {
3604 lu_ref_del(&new_lock->l_reference, "reader", new_lock);
3605 lu_ref_del(&new_lock->l_reference, "user", new_lock);
3606 new_lock->l_readers--;
3608 while (new_lock->l_writers > 0) {
3609 lu_ref_del(&new_lock->l_reference, "writer", new_lock);
3610 lu_ref_del(&new_lock->l_reference, "user", new_lock);
3611 new_lock->l_writers--;
3614 new_lock->l_export = class_export_lock_get(req->rq_export, new_lock);
3615 new_lock->l_blocking_ast = lock->l_blocking_ast;
3616 new_lock->l_completion_ast = lock->l_completion_ast;
3617 if (ldlm_has_dom(new_lock))
3618 new_lock->l_glimpse_ast = ldlm_server_glimpse_ast;
3619 new_lock->l_remote_handle = lock->l_remote_handle;
3620 new_lock->l_flags &= ~LDLM_FL_LOCAL;
3622 unlock_res_and_lock(new_lock);
3624 cfs_hash_add(new_lock->l_export->exp_lock_hash,
3625 &new_lock->l_remote_handle,
3626 &new_lock->l_exp_hash);
3628 LDLM_LOCK_RELEASE(new_lock);
3629 lh->mlh_reg_lh.cookie = 0;
3631 RETURN(ELDLM_LOCK_REPLACED);
3634 void mdt_intent_fixup_resent(struct mdt_thread_info *info,
3635 struct ldlm_lock *new_lock,
3636 struct mdt_lock_handle *lh, __u64 flags)
3638 struct ptlrpc_request *req = mdt_info_req(info);
3639 struct ldlm_request *dlmreq;
3641 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
3644 dlmreq = req_capsule_client_get(info->mti_pill, &RMF_DLM_REQ);
3646 /* Check if this is a resend case (MSG_RESENT is set on RPC) and a
3647 * lock was found by ldlm_handle_enqueue(); if so @lh must be
3649 if (flags & LDLM_FL_RESENT) {
3650 lh->mlh_reg_lh.cookie = new_lock->l_handle.h_cookie;
3651 lh->mlh_reg_mode = new_lock->l_granted_mode;
3653 LDLM_DEBUG(new_lock, "Restoring lock cookie");
3654 DEBUG_REQ(D_DLMTRACE, req, "restoring lock cookie %#llx",
3655 lh->mlh_reg_lh.cookie);
3660 * If the xid matches, then we know this is a resent request, and allow
3661 * it. (It's probably an OPEN, for which we don't send a lock.
3663 if (req_can_reconstruct(req, NULL))
3667 * This remote handle isn't enqueued, so we never received or processed
3668 * this request. Clear MSG_RESENT, because it can be handled like any
3669 * normal request now.
3671 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
3673 DEBUG_REQ(D_DLMTRACE, req, "no existing lock with rhandle %#llx",
3674 dlmreq->lock_handle[0].cookie);
3677 static int mdt_intent_getxattr(enum ldlm_intent_flags it_opc,
3678 struct mdt_thread_info *info,
3679 struct ldlm_lock **lockp,
3682 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
3683 struct ldlm_reply *ldlm_rep = NULL;
3688 * Initialize lhc->mlh_reg_lh either from a previously granted lock
3689 * (for the resend case) or a new lock. Below we will use it to
3690 * replace the original lock.
3692 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
3693 if (!lustre_handle_is_used(&lhc->mlh_reg_lh)) {
3694 mdt_lock_reg_init(lhc, (*lockp)->l_req_mode);
3695 rc = mdt_object_lock(info, info->mti_object, lhc,
3696 MDS_INODELOCK_XATTR);
3701 rc = mdt_getxattr(info);
3703 if (mdt_info_req(info)->rq_repmsg != NULL)
3704 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
3706 if (ldlm_rep == NULL ||
3707 OBD_FAIL_CHECK(OBD_FAIL_MDS_XATTR_REP)) {
3708 mdt_object_unlock(info, info->mti_object, lhc, 1);
3712 RETURN(err_serious(-EFAULT));
3715 ldlm_rep->lock_policy_res2 = clear_serious(rc);
3717 /* This is left for interop instead of adding a new interop flag.
3719 #if LUSTRE_VERSION_CODE > OBD_OCD_VERSION(3, 0, 0, 0)
3720 if (ldlm_rep->lock_policy_res2) {
3721 mdt_object_unlock(info, info->mti_object, lhc, 1);
3722 RETURN(ELDLM_LOCK_ABORTED);
3726 rc = mdt_intent_lock_replace(info, lockp, lhc, flags, rc);
3730 static int mdt_intent_getattr(enum ldlm_intent_flags it_opc,
3731 struct mdt_thread_info *info,
3732 struct ldlm_lock **lockp,
3735 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
3737 struct ldlm_reply *ldlm_rep;
3738 struct mdt_body *reqbody;
3739 struct mdt_body *repbody;
3743 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
3746 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
3749 info->mti_cross_ref = !!(reqbody->mbo_valid & OBD_MD_FLCROSSREF);
3750 repbody->mbo_eadatasize = 0;
3751 repbody->mbo_aclsize = 0;
3755 child_bits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
3758 child_bits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE |
3762 CERROR("%s: unsupported intent %#x\n",
3763 mdt_obd_name(info->mti_mdt), (unsigned int)it_opc);
3764 GOTO(out_shrink, rc = -EINVAL);
3767 rc = mdt_init_ucred_intent_getattr(info, reqbody);
3769 GOTO(out_shrink, rc);
3771 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
3772 mdt_set_disposition(info, ldlm_rep, DISP_IT_EXECD);
3774 /* Get lock from request for possible resent case. */
3775 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
3777 rc = mdt_getattr_name_lock(info, lhc, child_bits, ldlm_rep);
3778 ldlm_rep->lock_policy_res2 = clear_serious(rc);
3780 if (mdt_get_disposition(ldlm_rep, DISP_LOOKUP_NEG))
3781 ldlm_rep->lock_policy_res2 = 0;
3782 if (!mdt_get_disposition(ldlm_rep, DISP_LOOKUP_POS) ||
3783 ldlm_rep->lock_policy_res2) {
3784 lhc->mlh_reg_lh.cookie = 0ull;
3785 GOTO(out_ucred, rc = ELDLM_LOCK_ABORTED);
3788 rc = mdt_intent_lock_replace(info, lockp, lhc, flags, rc);
3791 mdt_exit_ucred(info);
3793 mdt_client_compatibility(info);
3794 rc2 = mdt_fix_reply(info);
3800 static int mdt_intent_layout(enum ldlm_intent_flags it_opc,
3801 struct mdt_thread_info *info,
3802 struct ldlm_lock **lockp,
3805 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_LAYOUT];
3806 struct md_layout_change layout = { .mlc_opc = MD_LAYOUT_NOP };
3807 struct layout_intent *intent;
3808 struct lu_fid *fid = &info->mti_tmp_fid2;
3809 struct mdt_object *obj = NULL;
3810 int layout_size = 0;
3814 fid_extract_from_res_name(fid, &(*lockp)->l_resource->lr_name);
3816 intent = req_capsule_client_get(info->mti_pill, &RMF_LAYOUT_INTENT);
3820 CDEBUG(D_INFO, DFID "got layout change request from client: "
3821 "opc:%u flags:%#x extent "DEXT"\n",
3822 PFID(fid), intent->li_opc, intent->li_flags,
3823 PEXT(&intent->li_extent));
3825 switch (intent->li_opc) {
3826 case LAYOUT_INTENT_TRUNC:
3827 case LAYOUT_INTENT_WRITE:
3828 layout.mlc_opc = MD_LAYOUT_WRITE;
3829 layout.mlc_intent = intent;
3831 case LAYOUT_INTENT_ACCESS:
3833 case LAYOUT_INTENT_READ:
3834 case LAYOUT_INTENT_GLIMPSE:
3835 case LAYOUT_INTENT_RELEASE:
3836 case LAYOUT_INTENT_RESTORE:
3837 CERROR("%s: Unsupported layout intent opc %d\n",
3838 mdt_obd_name(info->mti_mdt), intent->li_opc);
3842 CERROR("%s: Unknown layout intent opc %d\n",
3843 mdt_obd_name(info->mti_mdt), intent->li_opc);
3850 /* Get lock from request for possible resent case. */
3851 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
3853 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
3855 GOTO(out, rc = PTR_ERR(obj));
3858 if (mdt_object_exists(obj) && !mdt_object_remote(obj)) {
3859 /* if layout is going to be changed don't use the current EA
3860 * size but the maximum one. That buffer will be shrinked
3861 * to the actual size in req_capsule_shrink() before reply.
3863 if (layout.mlc_opc == MD_LAYOUT_WRITE) {
3864 layout_size = info->mti_mdt->mdt_max_mdsize;
3866 layout_size = mdt_attr_get_eabuf_size(info, obj);
3867 if (layout_size < 0)
3868 GOTO(out_obj, rc = layout_size);
3870 if (layout_size > info->mti_mdt->mdt_max_mdsize)
3871 info->mti_mdt->mdt_max_mdsize = layout_size;
3873 CDEBUG(D_INFO, "%s: layout_size %d\n",
3874 mdt_obd_name(info->mti_mdt), layout_size);
3878 * set reply buffer size, so that ldlm_handle_enqueue0()->
3879 * ldlm_lvbo_fill() will fill the reply buffer with lovea.
3881 (*lockp)->l_lvb_type = LVB_T_LAYOUT;
3882 req_capsule_set_size(info->mti_pill, &RMF_DLM_LVB, RCL_SERVER,
3884 rc = req_capsule_server_pack(info->mti_pill);
3889 if (layout.mlc_opc != MD_LAYOUT_NOP) {
3890 struct lu_buf *buf = &layout.mlc_buf;
3893 * mdt_layout_change is a reint operation, when the request
3894 * is resent, layout write shouldn't reprocess it again.
3896 rc = mdt_check_resent(info, mdt_reconstruct_generic, lhc);
3898 GOTO(out_obj, rc = rc < 0 ? rc : 0);
3901 * There is another resent case: the client's job has been
3902 * done by another client, referring lod_declare_layout_change
3903 * -EALREADY case, and it became a operation w/o transaction,
3904 * so we should not do the layout change, otherwise
3905 * mdt_layout_change() will try to cancel the granted server
3906 * CR lock whose remote counterpart is still in hold on the
3907 * client, and a deadlock ensues.
3909 rc = mdt_check_resent_lock(info, obj, lhc);
3915 if (unlikely(req_is_replay(mdt_info_req(info)))) {
3916 buf->lb_buf = req_capsule_client_get(info->mti_pill,
3918 buf->lb_len = req_capsule_get_size(info->mti_pill,
3919 &RMF_EADATA, RCL_CLIENT);
3921 * If it's a replay of layout write intent RPC, the
3922 * client has saved the extended lovea when
3923 * it get reply then.
3925 if (buf->lb_len > 0)
3926 mdt_fix_lov_magic(info, buf->lb_buf);
3929 * Instantiate some layout components, if @buf contains
3930 * lovea, then it's a replay of the layout intent write
3933 rc = mdt_layout_change(info, obj, &layout);
3938 mdt_object_put(info->mti_env, obj);
3940 if (rc == 0 && lustre_handle_is_used(&lhc->mlh_reg_lh))
3941 rc = mdt_intent_lock_replace(info, lockp, lhc, flags, rc);
3944 lhc->mlh_reg_lh.cookie = 0;
3949 static int mdt_intent_open(enum ldlm_intent_flags it_opc,
3950 struct mdt_thread_info *info,
3951 struct ldlm_lock **lockp,
3954 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
3955 struct ldlm_reply *rep = NULL;
3959 static const struct req_format *intent_fmts[REINT_MAX] = {
3960 [REINT_CREATE] = &RQF_LDLM_INTENT_CREATE,
3961 [REINT_OPEN] = &RQF_LDLM_INTENT_OPEN
3966 opc = mdt_reint_opcode(mdt_info_req(info), intent_fmts);
3970 /* Get lock from request for possible resent case. */
3971 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
3973 rc = mdt_reint_internal(info, lhc, opc);
3975 /* Check whether the reply has been packed successfully. */
3976 if (mdt_info_req(info)->rq_repmsg != NULL)
3977 rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
3982 RETURN(err_serious(-EFAULT));
3985 /* MDC expects this in any case */
3987 mdt_set_disposition(info, rep, DISP_LOOKUP_EXECD);
3989 /* the open lock or the lock for cross-ref object should be
3990 * returned to the client */
3991 if (lustre_handle_is_used(&lhc->mlh_reg_lh) &&
3992 (rc == 0 || rc == -MDT_EREMOTE_OPEN)) {
3993 rep->lock_policy_res2 = 0;
3994 rc = mdt_intent_lock_replace(info, lockp, lhc, flags, rc);
3998 rep->lock_policy_res2 = clear_serious(rc);
4000 if (rep->lock_policy_res2 == -ENOENT &&
4001 mdt_get_disposition(rep, DISP_LOOKUP_NEG) &&
4002 !mdt_get_disposition(rep, DISP_OPEN_CREATE))
4003 rep->lock_policy_res2 = 0;
4005 lhc->mlh_reg_lh.cookie = 0ull;
4006 if (rc == -ENOTCONN || rc == -ENODEV ||
4007 rc == -EOVERFLOW) { /**< if VBR failure then return error */
4009 * If it is the disconnect error (ENODEV & ENOCONN), the error
4010 * will be returned by rq_status, and client at ptlrpc layer
4011 * will detect this, then disconnect, reconnect the import
4012 * immediately, instead of impacting the following the rpc.
4017 * For other cases, the error will be returned by intent, and client
4018 * will retrieve the result from intent.
4020 RETURN(ELDLM_LOCK_ABORTED);
4023 static int mdt_intent_opc(enum ldlm_intent_flags it_opc,
4024 struct mdt_thread_info *info,
4025 struct ldlm_lock **lockp,
4026 u64 flags /* LDLM_FL_* */)
4028 struct req_capsule *pill = info->mti_pill;
4029 struct ptlrpc_request *req = mdt_info_req(info);
4030 const struct req_format *it_format;
4031 int (*it_handler)(enum ldlm_intent_flags,
4032 struct mdt_thread_info *,
4033 struct ldlm_lock **,
4035 enum tgt_handler_flags it_handler_flags = 0;
4036 struct ldlm_reply *rep;
4042 case IT_OPEN|IT_CREAT:
4044 * OCREAT is not a IS_MUTABLE request since the file may
4045 * already exist. We do the extra check of
4046 * OBD_CONNECT_RDONLY in mdt_reint_open() when we
4047 * really need to create the object.
4049 it_format = &RQF_LDLM_INTENT;
4050 it_handler = &mdt_intent_open;
4054 it_format = &RQF_LDLM_INTENT_GETATTR;
4055 it_handler = &mdt_intent_getattr;
4056 it_handler_flags = HAS_REPLY;
4059 it_format = &RQF_LDLM_INTENT_GETXATTR;
4060 it_handler = &mdt_intent_getxattr;
4061 it_handler_flags = HAS_BODY;
4064 it_format = &RQF_LDLM_INTENT_LAYOUT;
4065 it_handler = &mdt_intent_layout;
4068 it_format = &RQF_LDLM_INTENT;
4069 it_handler = &mdt_intent_glimpse;
4072 it_format = &RQF_LDLM_INTENT;
4073 it_handler = &mdt_intent_brw;
4075 case IT_QUOTA_DQACQ:
4076 case IT_QUOTA_CONN: {
4077 struct lu_device *qmt = info->mti_mdt->mdt_qmt_dev;
4080 RETURN(-EOPNOTSUPP);
4082 if (mdt_rdonly(req->rq_export))
4085 (*lockp)->l_lvb_type = LVB_T_LQUOTA;
4086 /* pass the request to quota master */
4087 rc = qmt_hdls.qmth_intent_policy(info->mti_env, qmt,
4088 mdt_info_req(info), lockp,
4093 CERROR("%s: unknown intent code %#x\n",
4094 mdt_obd_name(info->mti_mdt), it_opc);
4098 req_capsule_extend(pill, it_format);
4100 rc = mdt_unpack_req_pack_rep(info, it_handler_flags);
4104 if (it_handler_flags & IS_MUTABLE && mdt_rdonly(req->rq_export))
4107 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_INTENT_DELAY, 10);
4109 /* execute policy */
4110 rc = (*it_handler)(it_opc, info, lockp, flags);
4112 /* Check whether the reply has been packed successfully. */
4113 if (req->rq_repmsg != NULL) {
4114 rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
4115 rep->lock_policy_res2 =
4116 ptlrpc_status_hton(rep->lock_policy_res2);
4122 static void mdt_ptlrpc_stats_update(struct ptlrpc_request *req,
4123 enum ldlm_intent_flags it_opc)
4125 struct lprocfs_stats *srv_stats = ptlrpc_req2svc(req)->srv_stats;
4127 /* update stats when IT code is known */
4128 if (srv_stats != NULL)
4129 lprocfs_counter_incr(srv_stats,
4130 PTLRPC_LAST_CNTR + (it_opc == IT_GLIMPSE ?
4131 LDLM_GLIMPSE_ENQUEUE : LDLM_IBITS_ENQUEUE));
4134 static int mdt_intent_policy(const struct lu_env *env,
4135 struct ldlm_namespace *ns,
4136 struct ldlm_lock **lockp,
4138 enum ldlm_mode mode,
4139 __u64 flags, void *data)
4141 struct tgt_session_info *tsi;
4142 struct mdt_thread_info *info;
4143 struct ptlrpc_request *req = req_cookie;
4144 struct ldlm_intent *it;
4145 struct req_capsule *pill;
4146 const struct ldlm_lock_desc *ldesc;
4151 LASSERT(req != NULL);
4153 tsi = tgt_ses_info(env);
4155 info = tsi2mdt_info(tsi);
4156 LASSERT(info != NULL);
4157 pill = info->mti_pill;
4158 LASSERT(pill->rc_req == req);
4159 ldesc = &info->mti_dlm_req->lock_desc;
4161 if (req->rq_reqmsg->lm_bufcount > DLM_INTENT_IT_OFF) {
4162 req_capsule_extend(pill, &RQF_LDLM_INTENT_BASIC);
4163 it = req_capsule_client_get(pill, &RMF_LDLM_INTENT);
4165 mdt_ptlrpc_stats_update(req, it->opc);
4166 rc = mdt_intent_opc(it->opc, info, lockp, flags);
4170 /* Lock without inodebits makes no sense and will oops
4171 * later in ldlm. Let's check it now to see if we have
4172 * ibits corrupted somewhere in mdt_intent_opc().
4173 * The case for client miss to set ibits has been
4174 * processed by others. */
4175 LASSERT(ergo(ldesc->l_resource.lr_type == LDLM_IBITS,
4176 ldesc->l_policy_data.l_inodebits.bits != 0));
4178 rc = err_serious(-EFAULT);
4181 /* No intent was provided */
4182 req_capsule_set_size(pill, &RMF_DLM_LVB, RCL_SERVER, 0);
4183 rc = req_capsule_server_pack(pill);
4185 rc = err_serious(rc);
4187 mdt_thread_info_fini(info);
4191 static void mdt_deregister_seq_exp(struct mdt_device *mdt)
4193 struct seq_server_site *ss = mdt_seq_site(mdt);
4195 if (ss->ss_node_id == 0)
4198 if (ss->ss_client_seq != NULL) {
4199 lustre_deregister_lwp_item(&ss->ss_client_seq->lcs_exp);
4200 ss->ss_client_seq->lcs_exp = NULL;
4203 if (ss->ss_server_fld != NULL) {
4204 lustre_deregister_lwp_item(&ss->ss_server_fld->lsf_control_exp);
4205 ss->ss_server_fld->lsf_control_exp = NULL;
4209 static void mdt_seq_fini_cli(struct mdt_device *mdt)
4211 struct seq_server_site *ss = mdt_seq_site(mdt);
4216 if (ss->ss_server_seq != NULL)
4217 seq_server_set_cli(NULL, ss->ss_server_seq, NULL);
4220 static int mdt_seq_fini(const struct lu_env *env, struct mdt_device *mdt)
4222 mdt_seq_fini_cli(mdt);
4223 mdt_deregister_seq_exp(mdt);
4225 return seq_site_fini(env, mdt_seq_site(mdt));
4229 * It will retrieve its FLDB entries from MDT0, and it only happens
4230 * when upgrading existent FS to 2.6 or when local FLDB is corrupted,
4231 * and it needs to refresh FLDB from the MDT0.
4233 static int mdt_register_lwp_callback(void *data)
4236 struct mdt_device *mdt = data;
4237 struct lu_server_fld *fld = mdt_seq_site(mdt)->ss_server_fld;
4241 LASSERT(mdt_seq_site(mdt)->ss_node_id != 0);
4243 rc = lu_env_init(&env, LCT_MD_THREAD);
4245 CERROR("%s: cannot init env: rc = %d\n", mdt_obd_name(mdt), rc);
4249 /* Allocate new sequence now to avoid creating local transaction
4250 * in the normal transaction process */
4251 rc = seq_server_check_and_alloc_super(&env,
4252 mdt_seq_site(mdt)->ss_server_seq);
4257 rc = fld_update_from_controller(&env, fld);
4259 CERROR("%s: cannot update controller: rc = %d\n",
4260 mdt_obd_name(mdt), rc);
4269 static int mdt_register_seq_exp(struct mdt_device *mdt)
4271 struct seq_server_site *ss = mdt_seq_site(mdt);
4272 char *lwp_name = NULL;
4275 if (ss->ss_node_id == 0)
4278 OBD_ALLOC(lwp_name, MAX_OBD_NAME);
4279 if (lwp_name == NULL)
4280 GOTO(out_free, rc = -ENOMEM);
4282 rc = tgt_name2lwp_name(mdt_obd_name(mdt), lwp_name, MAX_OBD_NAME, 0);
4286 rc = lustre_register_lwp_item(lwp_name, &ss->ss_client_seq->lcs_exp,
4291 rc = lustre_register_lwp_item(lwp_name,
4292 &ss->ss_server_fld->lsf_control_exp,
4293 mdt_register_lwp_callback, mdt);
4295 lustre_deregister_lwp_item(&ss->ss_client_seq->lcs_exp);
4296 ss->ss_client_seq->lcs_exp = NULL;
4300 if (lwp_name != NULL)
4301 OBD_FREE(lwp_name, MAX_OBD_NAME);
4307 * Init client sequence manager which is used by local MDS to talk to sequence
4308 * controller on remote node.
4310 static int mdt_seq_init_cli(const struct lu_env *env, struct mdt_device *mdt)
4312 struct seq_server_site *ss = mdt_seq_site(mdt);
4317 /* check if this is adding the first MDC and controller is not yet
4319 OBD_ALLOC_PTR(ss->ss_client_seq);
4320 if (ss->ss_client_seq == NULL)
4323 OBD_ALLOC(prefix, MAX_OBD_NAME + 5);
4324 if (prefix == NULL) {
4325 OBD_FREE_PTR(ss->ss_client_seq);
4326 ss->ss_client_seq = NULL;
4330 /* Note: seq_client_fini will be called in seq_site_fini */
4331 snprintf(prefix, MAX_OBD_NAME + 5, "ctl-%s", mdt_obd_name(mdt));
4332 rc = seq_client_init(ss->ss_client_seq, NULL, LUSTRE_SEQ_METADATA,
4333 prefix, ss->ss_node_id == 0 ? ss->ss_control_seq :
4335 OBD_FREE(prefix, MAX_OBD_NAME + 5);
4337 OBD_FREE_PTR(ss->ss_client_seq);
4338 ss->ss_client_seq = NULL;
4342 rc = seq_server_set_cli(env, ss->ss_server_seq, ss->ss_client_seq);
4347 static int mdt_seq_init(const struct lu_env *env, struct mdt_device *mdt)
4349 struct seq_server_site *ss;
4353 ss = mdt_seq_site(mdt);
4354 /* init sequence controller server(MDT0) */
4355 if (ss->ss_node_id == 0) {
4356 OBD_ALLOC_PTR(ss->ss_control_seq);
4357 if (ss->ss_control_seq == NULL)
4360 rc = seq_server_init(env, ss->ss_control_seq, mdt->mdt_bottom,
4361 mdt_obd_name(mdt), LUSTRE_SEQ_CONTROLLER,
4364 GOTO(out_seq_fini, rc);
4367 /* Init normal sequence server */
4368 OBD_ALLOC_PTR(ss->ss_server_seq);
4369 if (ss->ss_server_seq == NULL)
4370 GOTO(out_seq_fini, rc = -ENOMEM);
4372 rc = seq_server_init(env, ss->ss_server_seq, mdt->mdt_bottom,
4373 mdt_obd_name(mdt), LUSTRE_SEQ_SERVER, ss);
4375 GOTO(out_seq_fini, rc);
4377 /* init seq client for seq server to talk to seq controller(MDT0) */
4378 rc = mdt_seq_init_cli(env, mdt);
4380 GOTO(out_seq_fini, rc);
4382 if (ss->ss_node_id != 0)
4383 /* register controller export through lwp */
4384 rc = mdt_register_seq_exp(mdt);
4389 mdt_seq_fini(env, mdt);
4397 static int mdt_fld_fini(const struct lu_env *env,
4398 struct mdt_device *m)
4400 struct seq_server_site *ss = mdt_seq_site(m);
4403 if (ss && ss->ss_server_fld) {
4404 fld_server_fini(env, ss->ss_server_fld);
4405 OBD_FREE_PTR(ss->ss_server_fld);
4406 ss->ss_server_fld = NULL;
4412 static int mdt_fld_init(const struct lu_env *env,
4414 struct mdt_device *m)
4416 struct seq_server_site *ss;
4420 ss = mdt_seq_site(m);
4422 OBD_ALLOC_PTR(ss->ss_server_fld);
4423 if (ss->ss_server_fld == NULL)
4424 RETURN(rc = -ENOMEM);
4426 rc = fld_server_init(env, ss->ss_server_fld, m->mdt_bottom, uuid,
4429 OBD_FREE_PTR(ss->ss_server_fld);
4430 ss->ss_server_fld = NULL;
4437 static void mdt_stack_pre_fini(const struct lu_env *env,
4438 struct mdt_device *m, struct lu_device *top)
4440 struct lustre_cfg_bufs *bufs;
4441 struct lustre_cfg *lcfg;
4442 struct mdt_thread_info *info;
4447 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
4448 LASSERT(info != NULL);
4450 bufs = &info->mti_u.bufs;
4452 LASSERT(m->mdt_child_exp);
4453 LASSERT(m->mdt_child_exp->exp_obd);
4455 /* process cleanup, pass mdt obd name to get obd umount flags */
4456 /* XXX: this is needed because all layers are referenced by
4457 * objects (some of them are pinned by osd, for example *
4458 * the proper solution should be a model where object used
4459 * by osd only doesn't have mdt/mdd slices -bzzz */
4460 lustre_cfg_bufs_reset(bufs, mdt_obd_name(m));
4461 lustre_cfg_bufs_set_string(bufs, 1, NULL);
4462 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
4465 lustre_cfg_init(lcfg, LCFG_PRE_CLEANUP, bufs);
4467 top->ld_ops->ldo_process_config(env, top, lcfg);
4468 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens));
4472 static void mdt_stack_fini(const struct lu_env *env,
4473 struct mdt_device *m, struct lu_device *top)
4475 struct obd_device *obd = mdt2obd_dev(m);
4476 struct lustre_cfg_bufs *bufs;
4477 struct lustre_cfg *lcfg;
4478 struct mdt_thread_info *info;
4482 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
4483 LASSERT(info != NULL);
4485 lu_dev_del_linkage(top->ld_site, top);
4487 lu_site_purge(env, top->ld_site, -1);
4489 bufs = &info->mti_u.bufs;
4490 /* process cleanup, pass mdt obd name to get obd umount flags */
4491 /* another purpose is to let all layers to release their objects */
4492 lustre_cfg_bufs_reset(bufs, mdt_obd_name(m));
4497 lustre_cfg_bufs_set_string(bufs, 1, flags);
4498 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
4501 lustre_cfg_init(lcfg, LCFG_CLEANUP, bufs);
4504 top->ld_ops->ldo_process_config(env, top, lcfg);
4505 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens));
4507 lu_site_purge(env, top->ld_site, -1);
4509 m->mdt_child = NULL;
4510 m->mdt_bottom = NULL;
4512 obd_disconnect(m->mdt_child_exp);
4513 m->mdt_child_exp = NULL;
4515 obd_disconnect(m->mdt_bottom_exp);
4516 m->mdt_child_exp = NULL;
4519 static int mdt_connect_to_next(const struct lu_env *env, struct mdt_device *m,
4520 const char *next, struct obd_export **exp)
4522 struct obd_connect_data *data = NULL;
4523 struct obd_device *obd;
4527 OBD_ALLOC_PTR(data);
4529 GOTO(out, rc = -ENOMEM);
4531 obd = class_name2obd(next);
4533 CERROR("%s: can't locate next device: %s\n",
4534 mdt_obd_name(m), next);
4535 GOTO(out, rc = -ENOTCONN);
4538 data->ocd_connect_flags = OBD_CONNECT_VERSION;
4539 data->ocd_version = LUSTRE_VERSION_CODE;
4541 rc = obd_connect(NULL, exp, obd, &obd->obd_uuid, data, NULL);
4543 CERROR("%s: cannot connect to next dev %s (%d)\n",
4544 mdt_obd_name(m), next, rc);
4554 static int mdt_stack_init(const struct lu_env *env, struct mdt_device *mdt,
4555 struct lustre_cfg *cfg)
4557 char *dev = lustre_cfg_string(cfg, 0);
4558 int rc, name_size, uuid_size;
4559 char *name, *uuid, *p;
4560 struct lustre_cfg_bufs *bufs;
4561 struct lustre_cfg *lcfg;
4562 struct obd_device *obd;
4563 struct lustre_profile *lprof;
4564 struct lu_site *site;
4567 /* in 1.8 we had the only device in the stack - MDS.
4568 * 2.0 introduces MDT, MDD, OSD; MDT starts others internally.
4569 * in 2.3 OSD is instantiated by obd_mount.c, so we need
4570 * to generate names and setup MDT, MDD. MDT will be using
4571 * generated name to connect to MDD. for MDD the next device
4572 * will be LOD with name taken from so called "profile" which
4573 * is generated by mount_option line
4575 * 1.8 MGS generates config. commands like this:
4576 * #06 (104)mount_option 0: 1:lustre-MDT0000 2:lustre-mdtlov
4577 * #08 (120)setup 0:lustre-MDT0000 1:dev 2:type 3:lustre-MDT0000
4578 * 2.0 MGS generates config. commands like this:
4579 * #07 (112)mount_option 0: 1:lustre-MDT0000 2:lustre-MDT0000-mdtlov
4580 * #08 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
4581 * 3:lustre-MDT0000-mdtlov 4:f
4583 * we generate MDD name from MDT one, just replacing T with D
4585 * after all the preparations, the logical equivalent will be
4586 * #01 (160)setup 0:lustre-MDD0000 1:lustre-MDD0000_UUID 2:0
4587 * 3:lustre-MDT0000-mdtlov 4:f
4588 * #02 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
4589 * 3:lustre-MDD0000 4:f
4591 * notice we build the stack from down to top: MDD first, then MDT */
4593 name_size = MAX_OBD_NAME;
4594 uuid_size = MAX_OBD_NAME;
4596 OBD_ALLOC(name, name_size);
4597 OBD_ALLOC(uuid, uuid_size);
4598 if (name == NULL || uuid == NULL)
4599 GOTO(cleanup_mem, rc = -ENOMEM);
4601 OBD_ALLOC_PTR(bufs);
4603 GOTO(cleanup_mem, rc = -ENOMEM);
4606 p = strstr(name, "-MDT");
4608 GOTO(free_bufs, rc = -ENOMEM);
4611 snprintf(uuid, MAX_OBD_NAME, "%s_UUID", name);
4613 lprof = class_get_profile(lustre_cfg_string(cfg, 0));
4614 if (lprof == NULL || lprof->lp_dt == NULL) {
4615 CERROR("can't find the profile: %s\n",
4616 lustre_cfg_string(cfg, 0));
4617 GOTO(free_bufs, rc = -EINVAL);
4620 lustre_cfg_bufs_reset(bufs, name);
4621 lustre_cfg_bufs_set_string(bufs, 1, LUSTRE_MDD_NAME);
4622 lustre_cfg_bufs_set_string(bufs, 2, uuid);
4623 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
4625 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
4627 GOTO(put_profile, rc = -ENOMEM);
4628 lustre_cfg_init(lcfg, LCFG_ATTACH, bufs);
4630 rc = class_attach(lcfg);
4632 GOTO(lcfg_cleanup, rc);
4634 obd = class_name2obd(name);
4636 CERROR("Can not find obd %s (%s in config)\n",
4637 MDD_OBD_NAME, lustre_cfg_string(cfg, 0));
4638 GOTO(lcfg_cleanup, rc = -EINVAL);
4641 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens));
4643 lustre_cfg_bufs_reset(bufs, name);
4644 lustre_cfg_bufs_set_string(bufs, 1, uuid);
4645 lustre_cfg_bufs_set_string(bufs, 2, dev);
4646 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
4648 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
4650 GOTO(class_detach, rc = -ENOMEM);
4651 lustre_cfg_init(lcfg, LCFG_SETUP, bufs);
4653 rc = class_setup(obd, lcfg);
4655 GOTO(class_detach, rc);
4657 /* connect to MDD we just setup */
4658 rc = mdt_connect_to_next(env, mdt, name, &mdt->mdt_child_exp);
4660 GOTO(class_detach, rc);
4662 site = mdt->mdt_child_exp->exp_obd->obd_lu_dev->ld_site;
4664 LASSERT(mdt_lu_site(mdt) == NULL);
4665 mdt->mdt_lu_dev.ld_site = site;
4666 site->ls_top_dev = &mdt->mdt_lu_dev;
4667 mdt->mdt_child = lu2md_dev(mdt->mdt_child_exp->exp_obd->obd_lu_dev);
4669 /* now connect to bottom OSD */
4670 snprintf(name, MAX_OBD_NAME, "%s-osd", dev);
4671 rc = mdt_connect_to_next(env, mdt, name, &mdt->mdt_bottom_exp);
4673 GOTO(class_detach, rc);
4675 lu2dt_dev(mdt->mdt_bottom_exp->exp_obd->obd_lu_dev);
4677 rc = lu_env_refill((struct lu_env *)env);
4679 CERROR("Failure to refill session: '%d'\n", rc);
4681 lu_dev_add_linkage(site, &mdt->mdt_lu_dev);
4686 class_detach(obd, lcfg);
4688 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens));
4690 class_put_profile(lprof);
4695 OBD_FREE(name, name_size);
4697 OBD_FREE(uuid, uuid_size);
4701 /* setup quota master target on MDT0 */
4702 static int mdt_quota_init(const struct lu_env *env, struct mdt_device *mdt,
4703 struct lustre_cfg *cfg)
4705 struct obd_device *obd;
4706 char *dev = lustre_cfg_string(cfg, 0);
4707 char *qmtname, *uuid, *p;
4708 struct lustre_cfg_bufs *bufs;
4709 struct lustre_cfg *lcfg;
4710 struct lustre_profile *lprof;
4711 struct obd_connect_data *data;
4715 LASSERT(mdt->mdt_qmt_exp == NULL);
4716 LASSERT(mdt->mdt_qmt_dev == NULL);
4718 /* quota master is on MDT0 only for now */
4719 if (mdt->mdt_seq_site.ss_node_id != 0)
4722 /* MGS generates config commands which look as follows:
4723 * #01 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
4724 * 3:lustre-MDT0000-mdtlov 4:f
4726 * We generate the QMT name from the MDT one, just replacing MD with QM
4727 * after all the preparations, the logical equivalent will be:
4728 * #01 (160)setup 0:lustre-QMT0000 1:lustre-QMT0000_UUID 2:0
4729 * 3:lustre-MDT0000-osd 4:f */
4730 OBD_ALLOC(qmtname, MAX_OBD_NAME);
4731 OBD_ALLOC(uuid, UUID_MAX);
4732 OBD_ALLOC_PTR(bufs);
4733 OBD_ALLOC_PTR(data);
4734 if (qmtname == NULL || uuid == NULL || bufs == NULL || data == NULL)
4735 GOTO(cleanup_mem, rc = -ENOMEM);
4737 strcpy(qmtname, dev);
4738 p = strstr(qmtname, "-MDT");
4740 GOTO(cleanup_mem, rc = -ENOMEM);
4741 /* replace MD with QM */
4745 snprintf(uuid, UUID_MAX, "%s_UUID", qmtname);
4747 lprof = class_get_profile(lustre_cfg_string(cfg, 0));
4748 if (lprof == NULL || lprof->lp_dt == NULL) {
4749 CERROR("can't find profile for %s\n",
4750 lustre_cfg_string(cfg, 0));
4751 GOTO(cleanup_mem, rc = -EINVAL);
4754 lustre_cfg_bufs_reset(bufs, qmtname);
4755 lustre_cfg_bufs_set_string(bufs, 1, LUSTRE_QMT_NAME);
4756 lustre_cfg_bufs_set_string(bufs, 2, uuid);
4757 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
4759 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
4761 GOTO(put_profile, rc = -ENOMEM);
4762 lustre_cfg_init(lcfg, LCFG_ATTACH, bufs);
4764 rc = class_attach(lcfg);
4766 GOTO(lcfg_cleanup, rc);
4768 obd = class_name2obd(qmtname);
4770 CERROR("Can not find obd %s (%s in config)\n", qmtname,
4771 lustre_cfg_string(cfg, 0));
4772 GOTO(lcfg_cleanup, rc = -EINVAL);
4775 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens));
4777 lustre_cfg_bufs_reset(bufs, qmtname);
4778 lustre_cfg_bufs_set_string(bufs, 1, uuid);
4779 lustre_cfg_bufs_set_string(bufs, 2, dev);
4781 /* for quota, the next device should be the OSD device */
4782 lustre_cfg_bufs_set_string(bufs, 3,
4783 mdt->mdt_bottom->dd_lu_dev.ld_obd->obd_name);
4785 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
4787 GOTO(class_detach, rc = -ENOMEM);
4788 lustre_cfg_init(lcfg, LCFG_SETUP, bufs);
4790 rc = class_setup(obd, lcfg);
4792 GOTO(class_detach, rc);
4794 mdt->mdt_qmt_dev = obd->obd_lu_dev;
4796 /* configure local quota objects */
4797 if (OBD_FAIL_CHECK(OBD_FAIL_QUOTA_INIT))
4800 rc = mdt->mdt_qmt_dev->ld_ops->ldo_prepare(env,
4804 GOTO(class_cleanup, rc);
4806 /* connect to quota master target */
4807 data->ocd_connect_flags = OBD_CONNECT_VERSION;
4808 data->ocd_version = LUSTRE_VERSION_CODE;
4809 rc = obd_connect(NULL, &mdt->mdt_qmt_exp, obd, &obd->obd_uuid,
4812 CERROR("cannot connect to quota master device %s (%d)\n",
4814 GOTO(class_cleanup, rc);
4820 class_manual_cleanup(obd);
4821 mdt->mdt_qmt_dev = NULL;
4822 GOTO(lcfg_cleanup, rc);
4826 class_detach(obd, lcfg);
4828 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens));
4830 class_put_profile(lprof);
4835 OBD_FREE(qmtname, MAX_OBD_NAME);
4837 OBD_FREE(uuid, UUID_MAX);
4843 /* Shutdown quota master target associated with mdt */
4844 static void mdt_quota_fini(const struct lu_env *env, struct mdt_device *mdt)
4848 if (mdt->mdt_qmt_exp == NULL)
4850 LASSERT(mdt->mdt_qmt_dev != NULL);
4852 /* the qmt automatically shuts down when the mdt disconnects */
4853 obd_disconnect(mdt->mdt_qmt_exp);
4854 mdt->mdt_qmt_exp = NULL;
4855 mdt->mdt_qmt_dev = NULL;
4859 /* mdt_getxattr() is used from mdt_intent_getxattr(), use this wrapper
4860 * for now. This will be removed along with converting rest of MDT code
4861 * to use tgt_session_info */
4862 static int mdt_tgt_getxattr(struct tgt_session_info *tsi)
4864 struct mdt_thread_info *info = tsi2mdt_info(tsi);
4867 rc = mdt_getxattr(info);
4869 mdt_thread_info_fini(info);
4873 #define OBD_FAIL_OST_READ_NET OBD_FAIL_OST_BRW_NET
4874 #define OBD_FAIL_OST_WRITE_NET OBD_FAIL_OST_BRW_NET
4875 #define OST_BRW_READ OST_READ
4876 #define OST_BRW_WRITE OST_WRITE
4878 static struct tgt_handler mdt_tgt_handlers[] = {
4879 TGT_RPC_HANDLER(MDS_FIRST_OPC,
4880 0, MDS_CONNECT, mdt_tgt_connect,
4881 &RQF_CONNECT, LUSTRE_OBD_VERSION),
4882 TGT_RPC_HANDLER(MDS_FIRST_OPC,
4883 0, MDS_DISCONNECT, tgt_disconnect,
4884 &RQF_MDS_DISCONNECT, LUSTRE_OBD_VERSION),
4885 TGT_RPC_HANDLER(MDS_FIRST_OPC,
4886 HAS_REPLY, MDS_SET_INFO, mdt_set_info,
4887 &RQF_OBD_SET_INFO, LUSTRE_MDS_VERSION),
4888 TGT_MDT_HDL(0, MDS_GET_INFO, mdt_get_info),
4889 TGT_MDT_HDL(HAS_REPLY, MDS_GET_ROOT, mdt_get_root),
4890 TGT_MDT_HDL(HAS_BODY, MDS_GETATTR, mdt_getattr),
4891 TGT_MDT_HDL(HAS_BODY | HAS_REPLY, MDS_GETATTR_NAME,
4893 TGT_MDT_HDL(HAS_BODY, MDS_GETXATTR, mdt_tgt_getxattr),
4894 TGT_MDT_HDL(HAS_REPLY, MDS_STATFS, mdt_statfs),
4895 TGT_MDT_HDL(IS_MUTABLE, MDS_REINT, mdt_reint),
4896 TGT_MDT_HDL(HAS_BODY, MDS_CLOSE, mdt_close),
4897 TGT_MDT_HDL(HAS_BODY | HAS_REPLY, MDS_READPAGE, mdt_readpage),
4898 TGT_MDT_HDL(HAS_BODY | HAS_REPLY, MDS_SYNC, mdt_sync),
4899 TGT_MDT_HDL(0, MDS_QUOTACTL, mdt_quotactl),
4900 TGT_MDT_HDL(HAS_BODY | HAS_REPLY | IS_MUTABLE, MDS_HSM_PROGRESS,
4902 TGT_MDT_HDL(HAS_BODY | HAS_REPLY | IS_MUTABLE, MDS_HSM_CT_REGISTER,
4903 mdt_hsm_ct_register),
4904 TGT_MDT_HDL(HAS_BODY | HAS_REPLY | IS_MUTABLE, MDS_HSM_CT_UNREGISTER,
4905 mdt_hsm_ct_unregister),
4906 TGT_MDT_HDL(HAS_BODY | HAS_REPLY, MDS_HSM_STATE_GET,
4908 TGT_MDT_HDL(HAS_BODY | HAS_REPLY | IS_MUTABLE, MDS_HSM_STATE_SET,
4910 TGT_MDT_HDL(HAS_BODY | HAS_REPLY, MDS_HSM_ACTION, mdt_hsm_action),
4911 TGT_MDT_HDL(HAS_BODY | HAS_REPLY, MDS_HSM_REQUEST,
4913 TGT_MDT_HDL(HAS_KEY | HAS_BODY | HAS_REPLY | IS_MUTABLE,
4918 static struct tgt_handler mdt_io_ops[] = {
4919 TGT_OST_HDL_HP(HAS_BODY | HAS_REPLY, OST_BRW_READ, tgt_brw_read,
4921 TGT_OST_HDL_HP(HAS_BODY | IS_MUTABLE, OST_BRW_WRITE, tgt_brw_write,
4923 TGT_OST_HDL_HP(HAS_BODY | HAS_REPLY | IS_MUTABLE,
4924 OST_PUNCH, mdt_punch_hdl,
4926 TGT_OST_HDL(HAS_BODY | HAS_REPLY, OST_SYNC, mdt_data_sync),
4929 static struct tgt_handler mdt_sec_ctx_ops[] = {
4930 TGT_SEC_HDL_VAR(0, SEC_CTX_INIT, mdt_sec_ctx_handle),
4931 TGT_SEC_HDL_VAR(0, SEC_CTX_INIT_CONT,mdt_sec_ctx_handle),
4932 TGT_SEC_HDL_VAR(0, SEC_CTX_FINI, mdt_sec_ctx_handle)
4935 static struct tgt_handler mdt_quota_ops[] = {
4936 TGT_QUOTA_HDL(HAS_REPLY, QUOTA_DQACQ, mdt_quota_dqacq),
4939 static struct tgt_opc_slice mdt_common_slice[] = {
4941 .tos_opc_start = MDS_FIRST_OPC,
4942 .tos_opc_end = MDS_LAST_OPC,
4943 .tos_hs = mdt_tgt_handlers
4946 .tos_opc_start = OBD_FIRST_OPC,
4947 .tos_opc_end = OBD_LAST_OPC,
4948 .tos_hs = tgt_obd_handlers
4951 .tos_opc_start = LDLM_FIRST_OPC,
4952 .tos_opc_end = LDLM_LAST_OPC,
4953 .tos_hs = tgt_dlm_handlers
4956 .tos_opc_start = SEC_FIRST_OPC,
4957 .tos_opc_end = SEC_LAST_OPC,
4958 .tos_hs = mdt_sec_ctx_ops
4961 .tos_opc_start = OUT_UPDATE_FIRST_OPC,
4962 .tos_opc_end = OUT_UPDATE_LAST_OPC,
4963 .tos_hs = tgt_out_handlers
4966 .tos_opc_start = FLD_FIRST_OPC,
4967 .tos_opc_end = FLD_LAST_OPC,
4968 .tos_hs = fld_handlers
4971 .tos_opc_start = SEQ_FIRST_OPC,
4972 .tos_opc_end = SEQ_LAST_OPC,
4973 .tos_hs = seq_handlers
4976 .tos_opc_start = QUOTA_DQACQ,
4977 .tos_opc_end = QUOTA_LAST_OPC,
4978 .tos_hs = mdt_quota_ops
4981 .tos_opc_start = LLOG_FIRST_OPC,
4982 .tos_opc_end = LLOG_LAST_OPC,
4983 .tos_hs = tgt_llog_handlers
4986 .tos_opc_start = LFSCK_FIRST_OPC,
4987 .tos_opc_end = LFSCK_LAST_OPC,
4988 .tos_hs = tgt_lfsck_handlers
4991 .tos_opc_start = OST_FIRST_OPC,
4992 .tos_opc_end = OST_LAST_OPC,
4993 .tos_hs = mdt_io_ops
5000 static void mdt_fini(const struct lu_env *env, struct mdt_device *m)
5002 struct md_device *next = m->mdt_child;
5003 struct lu_device *d = &m->mdt_lu_dev;
5004 struct obd_device *obd = mdt2obd_dev(m);
5005 struct lfsck_stop stop;
5008 stop.ls_status = LS_PAUSED;
5010 next->md_ops->mdo_iocontrol(env, next, OBD_IOC_STOP_LFSCK, 0, &stop);
5012 mdt_stack_pre_fini(env, m, md2lu_dev(m->mdt_child));
5013 ping_evictor_stop();
5015 /* Remove the HSM /proc entry so the coordinator cannot be
5016 * restarted by a user while it's shutting down. */
5017 hsm_cdt_procfs_fini(m);
5018 mdt_hsm_cdt_stop(m);
5020 mdt_llog_ctxt_unclone(env, m, LLOG_AGENT_ORIG_CTXT);
5021 mdt_llog_ctxt_unclone(env, m, LLOG_CHANGELOG_ORIG_CTXT);
5023 if (m->mdt_namespace != NULL)
5024 ldlm_namespace_free_prior(m->mdt_namespace, NULL,
5025 d->ld_obd->obd_force);
5027 obd_exports_barrier(obd);
5028 obd_zombie_barrier();
5030 mdt_quota_fini(env, m);
5032 cfs_free_nidlist(&m->mdt_squash.rsi_nosquash_nids);
5034 /* Calling the cleanup functions in the same order as in the mdt_init0
5039 target_recovery_fini(obd);
5040 upcall_cache_cleanup(m->mdt_identity_cache);
5041 m->mdt_identity_cache = NULL;
5043 mdt_fs_cleanup(env, m);
5045 tgt_fini(env, &m->mdt_lut);
5047 mdt_hsm_cdt_fini(m);
5049 if (m->mdt_los != NULL) {
5050 local_oid_storage_fini(env, m->mdt_los);
5054 if (m->mdt_namespace != NULL) {
5055 ldlm_namespace_free_post(m->mdt_namespace);
5056 d->ld_obd->obd_namespace = m->mdt_namespace = NULL;
5059 if (m->mdt_md_root != NULL) {
5060 mdt_object_put(env, m->mdt_md_root);
5061 m->mdt_md_root = NULL;
5064 mdt_seq_fini(env, m);
5066 mdt_fld_fini(env, m);
5071 mdt_stack_fini(env, m, md2lu_dev(m->mdt_child));
5073 LASSERT(atomic_read(&d->ld_ref) == 0);
5075 server_put_mount(mdt_obd_name(m), true);
5080 static int mdt_postrecov(const struct lu_env *, struct mdt_device *);
5082 static int mdt_init0(const struct lu_env *env, struct mdt_device *m,
5083 struct lu_device_type *ldt, struct lustre_cfg *cfg)
5085 const struct dt_device_param *dt_conf;
5086 struct mdt_thread_info *info;
5087 struct obd_device *obd;
5088 const char *dev = lustre_cfg_string(cfg, 0);
5089 const char *num = lustre_cfg_string(cfg, 2);
5090 struct tg_grants_data *tgd = &m->mdt_lut.lut_tgd;
5091 struct lustre_mount_info *lmi = NULL;
5092 struct lustre_sb_info *lsi;
5094 struct seq_server_site *ss_site;
5095 const char *identity_upcall = "NONE";
5096 struct md_device *next;
5103 lu_device_init(&m->mdt_lu_dev, ldt);
5105 * Environment (env) might be missing mdt_thread_key values at that
5106 * point, if device is allocated when mdt_thread_key is in QUIESCENT
5109 * Usually device allocation path doesn't use module key values, but
5110 * mdt has to do a lot of work here, so allocate key value.
5112 rc = lu_env_refill((struct lu_env *)env);
5116 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
5117 LASSERT(info != NULL);
5119 obd = class_name2obd(dev);
5120 LASSERT(obd != NULL);
5122 m->mdt_max_mdsize = MAX_MD_SIZE_OLD;
5123 m->mdt_opts.mo_evict_tgt_nids = 1;
5124 m->mdt_opts.mo_cos = MDT_COS_DEFAULT;
5126 lmi = server_get_mount(dev);
5128 CERROR("Cannot get mount info for %s!\n", dev);
5131 lsi = s2lsi(lmi->lmi_sb);
5132 /* CMD is supported only in IAM mode */
5134 node_id = simple_strtol(num, NULL, 10);
5135 obd->u.obt.obt_magic = OBT_MAGIC;
5136 if (lsi->lsi_lmd != NULL &&
5137 lsi->lsi_lmd->lmd_flags & LMD_FLG_SKIP_LFSCK)
5138 m->mdt_skip_lfsck = 1;
5141 /* DoM files get IO lock at open by default */
5142 m->mdt_opts.mo_dom_lock = ALWAYS_DOM_LOCK_ON_OPEN;
5143 /* DoM files are read at open and data is packed in the reply */
5144 m->mdt_opts.mo_dom_read_open = 1;
5146 m->mdt_squash.rsi_uid = 0;
5147 m->mdt_squash.rsi_gid = 0;
5148 INIT_LIST_HEAD(&m->mdt_squash.rsi_nosquash_nids);
5149 init_rwsem(&m->mdt_squash.rsi_sem);
5150 spin_lock_init(&m->mdt_lock);
5151 m->mdt_enable_remote_dir = 1;
5152 m->mdt_enable_striped_dir = 1;
5153 m->mdt_enable_dir_migration = 1;
5154 m->mdt_enable_remote_dir_gid = 0;
5155 m->mdt_enable_remote_rename = 1;
5157 atomic_set(&m->mdt_mds_mds_conns, 0);
5158 atomic_set(&m->mdt_async_commit_count, 0);
5160 m->mdt_lu_dev.ld_ops = &mdt_lu_ops;
5161 m->mdt_lu_dev.ld_obd = obd;
5162 /* Set this lu_device to obd for error handling purposes. */
5163 obd->obd_lu_dev = &m->mdt_lu_dev;
5165 /* init the stack */
5166 rc = mdt_stack_init((struct lu_env *)env, m, cfg);
5168 CERROR("%s: Can't init device stack, rc %d\n",
5169 mdt_obd_name(m), rc);
5174 ss_site = mdt_seq_site(m);
5175 s->ld_seq_site = ss_site;
5178 /* set server index */
5179 ss_site->ss_node_id = node_id;
5181 /* failover is the default
5182 * FIXME: we do not failout mds0/mgs, which may cause some problems.
5183 * assumed whose ss_node_id == 0 XXX
5185 obd->obd_replayable = 1;
5186 /* No connection accepted until configurations will finish */
5187 obd->obd_no_conn = 1;
5189 if (cfg->lcfg_bufcount > 4 && LUSTRE_CFG_BUFLEN(cfg, 4) > 0) {
5190 char *str = lustre_cfg_string(cfg, 4);
5191 if (strchr(str, 'n')) {
5192 CWARN("%s: recovery disabled\n", mdt_obd_name(m));
5193 obd->obd_replayable = 0;
5197 rc = mdt_fld_init(env, mdt_obd_name(m), m);
5199 GOTO(err_fini_stack, rc);
5201 rc = mdt_seq_init(env, m);
5203 GOTO(err_fini_fld, rc);
5205 snprintf(info->mti_u.ns_name, sizeof(info->mti_u.ns_name), "%s-%s",
5206 LUSTRE_MDT_NAME, obd->obd_uuid.uuid);
5207 m->mdt_namespace = ldlm_namespace_new(obd, info->mti_u.ns_name,
5208 LDLM_NAMESPACE_SERVER,
5209 LDLM_NAMESPACE_GREEDY,
5211 if (m->mdt_namespace == NULL)
5212 GOTO(err_fini_seq, rc = -ENOMEM);
5214 m->mdt_namespace->ns_lvbp = m;
5215 m->mdt_namespace->ns_lvbo = &mdt_lvbo;
5217 ldlm_register_intent(m->mdt_namespace, mdt_intent_policy);
5218 /* set obd_namespace for compatibility with old code */
5219 obd->obd_namespace = m->mdt_namespace;
5221 rc = tgt_init(env, &m->mdt_lut, obd, m->mdt_bottom, mdt_common_slice,
5222 OBD_FAIL_MDS_ALL_REQUEST_NET,
5223 OBD_FAIL_MDS_ALL_REPLY_NET);
5225 GOTO(err_free_ns, rc);
5227 /* Amount of available space excluded from granting and reserved
5228 * for metadata. It is in percentage and 50% is default value. */
5229 tgd->tgd_reserved_pcnt = 50;
5231 if (ONE_MB_BRW_SIZE < (1U << tgd->tgd_blockbits))
5232 m->mdt_brw_size = 1U << tgd->tgd_blockbits;
5234 m->mdt_brw_size = ONE_MB_BRW_SIZE;
5236 rc = mdt_fs_setup(env, m, obd, lsi);
5240 fid.f_seq = FID_SEQ_LOCAL_NAME;
5243 rc = local_oid_storage_init(env, m->mdt_bottom, &fid, &m->mdt_los);
5245 GOTO(err_fs_cleanup, rc);
5247 rc = mdt_hsm_cdt_init(m);
5249 CERROR("%s: error initializing coordinator, rc %d\n",
5250 mdt_obd_name(m), rc);
5251 GOTO(err_los_fini, rc);
5254 tgt_adapt_sptlrpc_conf(&m->mdt_lut);
5256 next = m->mdt_child;
5257 dt_conf = next->md_ops->mdo_dtconf_get(env, next);
5259 mntopts = dt_conf->ddp_mntopts;
5261 if (mntopts & MNTOPT_USERXATTR)
5262 m->mdt_opts.mo_user_xattr = 1;
5264 m->mdt_opts.mo_user_xattr = 0;
5266 m->mdt_max_ea_size = dt_conf->ddp_max_ea_size;
5268 if (mntopts & MNTOPT_ACL)
5269 m->mdt_opts.mo_acl = 1;
5271 m->mdt_opts.mo_acl = 0;
5273 /* XXX: to support suppgid for ACL, we enable identity_upcall
5274 * by default, otherwise, maybe got unexpected -EACCESS. */
5275 if (m->mdt_opts.mo_acl)
5276 identity_upcall = MDT_IDENTITY_UPCALL_PATH;
5278 m->mdt_identity_cache = upcall_cache_init(mdt_obd_name(m),
5280 &mdt_identity_upcall_cache_ops);
5281 if (IS_ERR(m->mdt_identity_cache)) {
5282 rc = PTR_ERR(m->mdt_identity_cache);
5283 m->mdt_identity_cache = NULL;
5284 GOTO(err_free_hsm, rc);
5287 rc = mdt_procfs_init(m, dev);
5289 CERROR("Can't init MDT lprocfs, rc %d\n", rc);
5290 GOTO(err_recovery, rc);
5293 rc = mdt_quota_init(env, m, cfg);
5295 GOTO(err_procfs, rc);
5297 m->mdt_ldlm_client = &mdt2obd_dev(m)->obd_ldlm_client;
5298 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
5299 "mdt_ldlm_client", m->mdt_ldlm_client);
5301 ping_evictor_start();
5303 /* recovery will be started upon mdt_prepare()
5304 * when the whole stack is complete and ready
5305 * to serve the requests */
5307 /* Reduce the initial timeout on an MDS because it doesn't need such
5308 * a long timeout as an OST does. Adaptive timeouts will adjust this
5309 * value appropriately. */
5310 if (ldlm_timeout == LDLM_TIMEOUT_DEFAULT)
5311 ldlm_timeout = MDS_LDLM_TIMEOUT_DEFAULT;
5317 upcall_cache_cleanup(m->mdt_identity_cache);
5318 m->mdt_identity_cache = NULL;
5320 mdt_hsm_cdt_fini(m);
5322 local_oid_storage_fini(env, m->mdt_los);
5325 mdt_fs_cleanup(env, m);
5327 /* keep recoverable clients */
5329 target_recovery_fini(obd);
5330 obd_exports_barrier(obd);
5331 obd_zombie_barrier();
5332 tgt_fini(env, &m->mdt_lut);
5334 ldlm_namespace_free(m->mdt_namespace, NULL, 0);
5335 obd->obd_namespace = m->mdt_namespace = NULL;
5337 mdt_seq_fini(env, m);
5339 mdt_fld_fini(env, m);
5341 mdt_stack_fini(env, m, md2lu_dev(m->mdt_child));
5344 server_put_mount(dev, true);
5348 /* For interoperability, the left element is old parameter, the right one
5349 * is the new version of the parameter, if some parameter is deprecated,
5350 * the new version should be set as NULL. */
5351 static struct cfg_interop_param mdt_interop_param[] = {
5352 { "mdt.group_upcall", NULL },
5353 { "mdt.quota_type", NULL },
5354 { "mdd.quota_type", NULL },
5355 { "mdt.som", NULL },
5356 { "mdt.rootsquash", "mdt.root_squash" },
5357 { "mdt.nosquash_nid", "mdt.nosquash_nids" },
5361 /* used by MGS to process specific configurations */
5362 static int mdt_process_config(const struct lu_env *env,
5363 struct lu_device *d, struct lustre_cfg *cfg)
5365 struct mdt_device *m = mdt_dev(d);
5366 struct md_device *md_next = m->mdt_child;
5367 struct lu_device *next = md2lu_dev(md_next);
5371 switch (cfg->lcfg_command) {
5373 struct obd_device *obd = d->ld_obd;
5375 /* For interoperability */
5376 struct cfg_interop_param *ptr = NULL;
5377 struct lustre_cfg *old_cfg = NULL;
5380 param = lustre_cfg_string(cfg, 1);
5381 if (param == NULL) {
5382 CERROR("param is empty\n");
5387 ptr = class_find_old_param(param, mdt_interop_param);
5389 if (ptr->new_param == NULL) {
5391 CWARN("For interoperability, skip this %s."
5392 " It is obsolete.\n", ptr->old_param);
5396 CWARN("Found old param %s, changed it to %s.\n",
5397 ptr->old_param, ptr->new_param);
5400 cfg = lustre_cfg_rename(old_cfg, ptr->new_param);
5407 rc = class_process_proc_param(PARAM_MDT, obd->obd_vars,
5409 if (rc > 0 || rc == -ENOSYS) {
5410 /* is it an HSM var ? */
5411 rc = class_process_proc_param(PARAM_HSM,
5412 hsm_cdt_get_proc_vars(),
5414 if (rc > 0 || rc == -ENOSYS)
5415 /* we don't understand; pass it on */
5416 rc = next->ld_ops->ldo_process_config(env, next,
5421 OBD_FREE(cfg, lustre_cfg_len(cfg->lcfg_bufcount,
5422 cfg->lcfg_buflens));
5426 /* others are passed further */
5427 rc = next->ld_ops->ldo_process_config(env, next, cfg);
5433 static struct lu_object *mdt_object_alloc(const struct lu_env *env,
5434 const struct lu_object_header *hdr,
5435 struct lu_device *d)
5437 struct mdt_object *mo;
5441 OBD_SLAB_ALLOC_PTR_GFP(mo, mdt_object_kmem, GFP_NOFS);
5443 struct lu_object *o;
5444 struct lu_object_header *h;
5447 h = &mo->mot_header;
5448 lu_object_header_init(h);
5449 lu_object_init(o, h, d);
5450 lu_object_add_top(h, o);
5451 o->lo_ops = &mdt_obj_ops;
5452 spin_lock_init(&mo->mot_write_lock);
5453 mutex_init(&mo->mot_som_mutex);
5454 mutex_init(&mo->mot_lov_mutex);
5455 init_rwsem(&mo->mot_dom_sem);
5456 init_rwsem(&mo->mot_open_sem);
5457 atomic_set(&mo->mot_open_count, 0);
5463 static int mdt_object_init(const struct lu_env *env, struct lu_object *o,
5464 const struct lu_object_conf *unused)
5466 struct mdt_device *d = mdt_dev(o->lo_dev);
5467 struct lu_device *under;
5468 struct lu_object *below;
5472 CDEBUG(D_INFO, "object init, fid = "DFID"\n",
5473 PFID(lu_object_fid(o)));
5475 under = &d->mdt_child->md_lu_dev;
5476 below = under->ld_ops->ldo_object_alloc(env, o->lo_header, under);
5477 if (below != NULL) {
5478 lu_object_add(o, below);
5485 static void mdt_object_free(const struct lu_env *env, struct lu_object *o)
5487 struct mdt_object *mo = mdt_obj(o);
5488 struct lu_object_header *h;
5492 CDEBUG(D_INFO, "object free, fid = "DFID"\n",
5493 PFID(lu_object_fid(o)));
5495 LASSERT(atomic_read(&mo->mot_open_count) == 0);
5496 LASSERT(atomic_read(&mo->mot_lease_count) == 0);
5499 lu_object_header_fini(h);
5500 OBD_SLAB_FREE_PTR(mo, mdt_object_kmem);
5505 static int mdt_object_print(const struct lu_env *env, void *cookie,
5506 lu_printer_t p, const struct lu_object *o)
5508 struct mdt_object *mdto = mdt_obj((struct lu_object *)o);
5510 return (*p)(env, cookie,
5511 LUSTRE_MDT_NAME"-object@%p(%s %s, writecount=%d)",
5512 mdto, mdto->mot_lov_created ? "lov_created" : "",
5513 mdto->mot_cache_attr ? "cache_attr" : "",
5514 mdto->mot_write_count);
5517 static int mdt_prepare(const struct lu_env *env,
5518 struct lu_device *pdev,
5519 struct lu_device *cdev)
5521 struct mdt_device *mdt = mdt_dev(cdev);
5522 struct lu_device *next = &mdt->mdt_child->md_lu_dev;
5523 struct obd_device *obd = cdev->ld_obd;
5530 rc = next->ld_ops->ldo_prepare(env, cdev, next);
5534 rc = mdt_llog_ctxt_clone(env, mdt, LLOG_CHANGELOG_ORIG_CTXT);
5538 rc = mdt_llog_ctxt_clone(env, mdt, LLOG_AGENT_ORIG_CTXT);
5542 rc = lfsck_register_namespace(env, mdt->mdt_bottom, mdt->mdt_namespace);
5543 /* The LFSCK instance is registered just now, so it must be there when
5544 * register the namespace to such instance. */
5545 LASSERTF(rc == 0, "register namespace failed: rc = %d\n", rc);
5547 if (mdt->mdt_seq_site.ss_node_id == 0) {
5548 rc = mdt->mdt_child->md_ops->mdo_root_get(env, mdt->mdt_child,
5549 &mdt->mdt_md_root_fid);
5554 LASSERT(!test_bit(MDT_FL_CFGLOG, &mdt->mdt_state));
5556 target_recovery_init(&mdt->mdt_lut, tgt_request_handle);
5557 set_bit(MDT_FL_CFGLOG, &mdt->mdt_state);
5558 LASSERT(obd->obd_no_conn);
5559 spin_lock(&obd->obd_dev_lock);
5560 obd->obd_no_conn = 0;
5561 spin_unlock(&obd->obd_dev_lock);
5563 if (obd->obd_recovering == 0)
5564 mdt_postrecov(env, mdt);
5569 const struct lu_device_operations mdt_lu_ops = {
5570 .ldo_object_alloc = mdt_object_alloc,
5571 .ldo_process_config = mdt_process_config,
5572 .ldo_prepare = mdt_prepare,
5575 static const struct lu_object_operations mdt_obj_ops = {
5576 .loo_object_init = mdt_object_init,
5577 .loo_object_free = mdt_object_free,
5578 .loo_object_print = mdt_object_print
5581 static int mdt_obd_set_info_async(const struct lu_env *env,
5582 struct obd_export *exp,
5583 __u32 keylen, void *key,
5584 __u32 vallen, void *val,
5585 struct ptlrpc_request_set *set)
5591 if (KEY_IS(KEY_SPTLRPC_CONF)) {
5592 rc = tgt_adapt_sptlrpc_conf(class_exp2tgt(exp));
5600 * Match client and server connection feature flags.
5602 * Compute the compatibility flags for a connection request based on
5603 * features mutually supported by client and server.
5605 * The obd_export::exp_connect_data.ocd_connect_flags field in \a exp
5606 * must not be updated here, otherwise a partially initialized value may
5607 * be exposed. After the connection request is successfully processed,
5608 * the top-level MDT connect request handler atomically updates the export
5609 * connect flags from the obd_connect_data::ocd_connect_flags field of the
5610 * reply. \see mdt_connect().
5612 * Before 2.7.50 clients will send a struct obd_connect_data_v1 rather than a
5613 * full struct obd_connect_data. So care must be taken when accessing fields
5614 * that are not present in struct obd_connect_data_v1. See LU-16.
5616 * \param exp the obd_export associated with this client/target pair
5617 * \param mdt the target device for the connection
5618 * \param data stores data for this connect request
5621 * \retval -EPROTO \a data unexpectedly has zero obd_connect_data::ocd_brw_size
5622 * \retval -EBADE client and server feature requirements are incompatible
5624 static int mdt_connect_internal(const struct lu_env *env,
5625 struct obd_export *exp,
5626 struct mdt_device *mdt,
5627 struct obd_connect_data *data, bool reconnect)
5629 const char *obd_name = mdt_obd_name(mdt);
5630 LASSERT(data != NULL);
5632 data->ocd_connect_flags &= MDT_CONNECT_SUPPORTED;
5634 if (mdt->mdt_bottom->dd_rdonly &&
5635 !(data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
5636 !(data->ocd_connect_flags & OBD_CONNECT_RDONLY))
5639 if (data->ocd_connect_flags & OBD_CONNECT_FLAGS2)
5640 data->ocd_connect_flags2 &= MDT_CONNECT_SUPPORTED2;
5642 data->ocd_ibits_known &= MDS_INODELOCK_FULL;
5644 if (!mdt->mdt_opts.mo_acl)
5645 data->ocd_connect_flags &= ~OBD_CONNECT_ACL;
5647 if (!mdt->mdt_opts.mo_user_xattr)
5648 data->ocd_connect_flags &= ~OBD_CONNECT_XATTR;
5650 if (OCD_HAS_FLAG(data, BRW_SIZE)) {
5651 data->ocd_brw_size = min(data->ocd_brw_size,
5653 if (data->ocd_brw_size == 0) {
5654 CERROR("%s: cli %s/%p ocd_connect_flags: %#llx "
5655 "ocd_version: %x ocd_grant: %d ocd_index: %u "
5656 "ocd_brw_size unexpectedly zero, network data "
5657 "corruption? Refusing to connect this client\n",
5658 obd_name, exp->exp_client_uuid.uuid,
5659 exp, data->ocd_connect_flags, data->ocd_version,
5660 data->ocd_grant, data->ocd_index);
5665 if (OCD_HAS_FLAG(data, GRANT_PARAM)) {
5666 struct dt_device_param *ddp = &mdt->mdt_lut.lut_dt_conf;
5668 /* client is reporting its page size, for future use */
5669 exp->exp_target_data.ted_pagebits = data->ocd_grant_blkbits;
5670 data->ocd_grant_blkbits = mdt->mdt_lut.lut_tgd.tgd_blockbits;
5671 /* ddp_inodespace may not be power-of-two value, eg. for ldiskfs
5672 * it's LDISKFS_DIR_REC_LEN(20) = 28. */
5673 data->ocd_grant_inobits = fls(ddp->ddp_inodespace - 1);
5674 /* ocd_grant_tax_kb is in 1K byte blocks */
5675 data->ocd_grant_tax_kb = ddp->ddp_extent_tax >> 10;
5676 data->ocd_grant_max_blks = ddp->ddp_max_extent_blks;
5679 /* Save connect_data we have so far because tgt_grant_connect()
5680 * uses it to calculate grant, and we want to save the client
5681 * version before it is overwritten by LUSTRE_VERSION_CODE. */
5682 exp->exp_connect_data = *data;
5683 if (OCD_HAS_FLAG(data, GRANT))
5684 tgt_grant_connect(env, exp, data, !reconnect);
5686 if (OCD_HAS_FLAG(data, MAXBYTES))
5687 data->ocd_maxbytes = mdt->mdt_lut.lut_dt_conf.ddp_maxbytes;
5689 /* NB: Disregard the rule against updating
5690 * exp_connect_data.ocd_connect_flags in this case, since
5691 * tgt_client_new() needs to know if this is a lightweight
5692 * connection, and it is safe to expose this flag before
5693 * connection processing completes. */
5694 if (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) {
5695 spin_lock(&exp->exp_lock);
5696 *exp_connect_flags_ptr(exp) |= OBD_CONNECT_LIGHTWEIGHT;
5697 spin_unlock(&exp->exp_lock);
5700 data->ocd_version = LUSTRE_VERSION_CODE;
5702 if ((data->ocd_connect_flags & OBD_CONNECT_FID) == 0) {
5703 CWARN("%s: MDS requires FID support, but client not\n",
5708 if (OCD_HAS_FLAG(data, PINGLESS)) {
5709 if (ptlrpc_pinger_suppress_pings()) {
5710 spin_lock(&exp->exp_obd->obd_dev_lock);
5711 list_del_init(&exp->exp_obd_chain_timed);
5712 spin_unlock(&exp->exp_obd->obd_dev_lock);
5714 data->ocd_connect_flags &= ~OBD_CONNECT_PINGLESS;
5718 data->ocd_max_easize = mdt->mdt_max_ea_size;
5720 /* NB: Disregard the rule against updating
5721 * exp_connect_data.ocd_connect_flags in this case, since
5722 * tgt_client_new() needs to know if this is client supports
5723 * multiple modify RPCs, and it is safe to expose this flag before
5724 * connection processing completes. */
5725 if (data->ocd_connect_flags & OBD_CONNECT_MULTIMODRPCS) {
5726 data->ocd_maxmodrpcs = max_mod_rpcs_per_client;
5727 spin_lock(&exp->exp_lock);
5728 *exp_connect_flags_ptr(exp) |= OBD_CONNECT_MULTIMODRPCS;
5729 spin_unlock(&exp->exp_lock);
5732 if (OCD_HAS_FLAG(data, CKSUM)) {
5733 __u32 cksum_types = data->ocd_cksum_types;
5735 /* The client set in ocd_cksum_types the checksum types it
5736 * supports. We have to mask off the algorithms that we don't
5738 data->ocd_cksum_types &=
5739 obd_cksum_types_supported_server(obd_name);
5741 if (unlikely(data->ocd_cksum_types == 0)) {
5742 CERROR("%s: Connect with checksum support but no "
5743 "ocd_cksum_types is set\n",
5744 exp->exp_obd->obd_name);
5748 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
5749 "%x\n", exp->exp_obd->obd_name, obd_export_nid2str(exp),
5750 cksum_types, data->ocd_cksum_types);
5752 /* This client does not support OBD_CONNECT_CKSUM
5753 * fall back to CRC32 */
5754 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
5755 "OBD_CONNECT_CKSUM, CRC32 will be used\n",
5756 exp->exp_obd->obd_name, obd_export_nid2str(exp));
5762 static int mdt_ctxt_add_dirty_flag(struct lu_env *env,
5763 struct mdt_thread_info *info,
5764 struct mdt_file_data *mfd)
5766 struct lu_context ses;
5770 rc = lu_context_init(&ses, LCT_SERVER_SESSION);
5775 lu_context_enter(&ses);
5777 mdt_ucred(info)->uc_valid = UCRED_OLD;
5778 rc = mdt_add_dirty_flag(info, mfd->mfd_object, &info->mti_attr);
5780 lu_context_exit(&ses);
5781 lu_context_fini(&ses);
5787 static int mdt_export_cleanup(struct obd_export *exp)
5789 struct list_head closing_list;
5790 struct mdt_export_data *med = &exp->exp_mdt_data;
5791 struct obd_device *obd = exp->exp_obd;
5792 struct mdt_device *mdt;
5793 struct mdt_thread_info *info;
5795 struct mdt_file_data *mfd, *n;
5799 INIT_LIST_HEAD(&closing_list);
5800 spin_lock(&med->med_open_lock);
5801 while (!list_empty(&med->med_open_head)) {
5802 struct list_head *tmp = med->med_open_head.next;
5803 mfd = list_entry(tmp, struct mdt_file_data, mfd_list);
5805 /* Remove mfd handle so it can't be found again.
5806 * We are consuming the mfd_list reference here. */
5807 class_handle_unhash(&mfd->mfd_open_handle);
5808 list_move_tail(&mfd->mfd_list, &closing_list);
5810 spin_unlock(&med->med_open_lock);
5811 mdt = mdt_dev(obd->obd_lu_dev);
5812 LASSERT(mdt != NULL);
5814 rc = lu_env_init(&env, LCT_MD_THREAD);
5818 info = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
5819 LASSERT(info != NULL);
5820 memset(info, 0, sizeof *info);
5821 info->mti_env = &env;
5822 info->mti_mdt = mdt;
5823 info->mti_exp = exp;
5825 if (!list_empty(&closing_list)) {
5826 struct md_attr *ma = &info->mti_attr;
5828 /* Close any open files (which may also cause orphan
5830 list_for_each_entry_safe(mfd, n, &closing_list, mfd_list) {
5831 list_del_init(&mfd->mfd_list);
5832 ma->ma_need = ma->ma_valid = 0;
5834 /* This file is being closed due to an eviction, it
5835 * could have been modified and now dirty regarding to
5836 * HSM archive, check this!
5837 * The logic here is to mark a file dirty if there's a
5838 * chance it was dirtied before the client was evicted,
5839 * so that we don't have to wait for a release attempt
5840 * before finding out the file was actually dirty and
5841 * fail the release. Aggressively marking it dirty here
5842 * will cause the policy engine to attempt to
5843 * re-archive it; when rearchiving, we can compare the
5844 * current version to the HSM data_version and make the
5845 * archive request into a noop if it's not actually
5848 if (mfd->mfd_open_flags & MDS_FMODE_WRITE)
5849 rc = mdt_ctxt_add_dirty_flag(&env, info, mfd);
5851 /* Don't unlink orphan on failover umount, LU-184 */
5852 if (exp->exp_flags & OBD_OPT_FAILOVER) {
5853 ma->ma_valid = MA_FLAGS;
5854 ma->ma_attr_flags |= MDS_KEEP_ORPHAN;
5856 mdt_mfd_close(info, mfd);
5859 info->mti_mdt = NULL;
5860 /* cleanup client slot early */
5861 /* Do not erase record for recoverable client. */
5862 if (!(exp->exp_flags & OBD_OPT_FAILOVER) || exp->exp_failed)
5863 tgt_client_del(&env, exp);
5869 static inline void mdt_enable_slc(struct mdt_device *mdt)
5871 if (mdt->mdt_lut.lut_sync_lock_cancel == NEVER_SYNC_ON_CANCEL)
5872 mdt->mdt_lut.lut_sync_lock_cancel = BLOCKING_SYNC_ON_CANCEL;
5875 static inline void mdt_disable_slc(struct mdt_device *mdt)
5877 if (mdt->mdt_lut.lut_sync_lock_cancel == BLOCKING_SYNC_ON_CANCEL)
5878 mdt->mdt_lut.lut_sync_lock_cancel = NEVER_SYNC_ON_CANCEL;
5881 static int mdt_obd_disconnect(struct obd_export *exp)
5888 class_export_get(exp);
5890 if (!(exp->exp_flags & OBD_OPT_FORCE))
5891 tgt_grant_sanity_check(exp->exp_obd, __func__);
5893 if ((exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) &&
5894 !(exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)) {
5895 struct mdt_device *mdt = mdt_dev(exp->exp_obd->obd_lu_dev);
5897 if (atomic_dec_and_test(&mdt->mdt_mds_mds_conns))
5898 mdt_disable_slc(mdt);
5901 rc = server_disconnect_export(exp);
5903 CDEBUG(D_IOCTL, "server disconnect error: rc = %d\n", rc);
5905 tgt_grant_discard(exp);
5907 rc = mdt_export_cleanup(exp);
5908 nodemap_del_member(exp);
5909 class_export_put(exp);
5913 /* mds_connect copy */
5914 static int mdt_obd_connect(const struct lu_env *env,
5915 struct obd_export **exp, struct obd_device *obd,
5916 struct obd_uuid *cluuid,
5917 struct obd_connect_data *data,
5920 struct obd_export *lexp;
5921 struct lustre_handle conn = { 0 };
5922 struct mdt_device *mdt;
5924 lnet_nid_t *client_nid = localdata;
5927 LASSERT(env != NULL);
5928 LASSERT(data != NULL);
5930 if (!exp || !obd || !cluuid)
5933 mdt = mdt_dev(obd->obd_lu_dev);
5935 if ((data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
5936 !(data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT)) {
5937 atomic_inc(&mdt->mdt_mds_mds_conns);
5938 mdt_enable_slc(mdt);
5942 * first, check whether the stack is ready to handle requests
5943 * XXX: probably not very appropriate method is used now
5944 * at some point we should find a better one
5946 if (!test_bit(MDT_FL_SYNCED, &mdt->mdt_state) &&
5947 !(data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) &&
5948 !(data->ocd_connect_flags & OBD_CONNECT_MDS_MDS)) {
5949 rc = obd_get_info(env, mdt->mdt_child_exp,
5950 sizeof(KEY_OSP_CONNECTED),
5951 KEY_OSP_CONNECTED, NULL, NULL);
5954 set_bit(MDT_FL_SYNCED, &mdt->mdt_state);
5957 rc = class_connect(&conn, obd, cluuid);
5961 lexp = class_conn2export(&conn);
5962 LASSERT(lexp != NULL);
5964 rc = nodemap_add_member(*client_nid, lexp);
5965 if (rc != 0 && rc != -EEXIST)
5968 rc = mdt_connect_internal(env, lexp, mdt, data, false);
5970 struct lsd_client_data *lcd = lexp->exp_target_data.ted_lcd;
5973 memcpy(lcd->lcd_uuid, cluuid, sizeof lcd->lcd_uuid);
5974 rc = tgt_client_new(env, lexp);
5976 mdt_export_stats_init(obd, lexp, localdata);
5980 class_disconnect(lexp);
5981 nodemap_del_member(lexp);
5985 /* Because we do not want this export to be evicted by pinger,
5986 * let's not add this export to the timed chain list. */
5987 if (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) {
5988 spin_lock(&lexp->exp_obd->obd_dev_lock);
5989 list_del_init(&lexp->exp_obd_chain_timed);
5990 spin_unlock(&lexp->exp_obd->obd_dev_lock);
5997 static int mdt_obd_reconnect(const struct lu_env *env,
5998 struct obd_export *exp, struct obd_device *obd,
5999 struct obd_uuid *cluuid,
6000 struct obd_connect_data *data,
6003 lnet_nid_t *client_nid = localdata;
6007 if (exp == NULL || obd == NULL || cluuid == NULL)
6010 rc = nodemap_add_member(*client_nid, exp);
6011 if (rc != 0 && rc != -EEXIST)
6014 rc = mdt_connect_internal(env, exp, mdt_dev(obd->obd_lu_dev), data,
6017 mdt_export_stats_init(obd, exp, localdata);
6019 nodemap_del_member(exp);
6024 /* FIXME: Can we avoid using these two interfaces? */
6025 static int mdt_init_export(struct obd_export *exp)
6027 struct mdt_export_data *med = &exp->exp_mdt_data;
6031 INIT_LIST_HEAD(&med->med_open_head);
6032 spin_lock_init(&med->med_open_lock);
6033 spin_lock(&exp->exp_lock);
6034 exp->exp_connecting = 1;
6035 spin_unlock(&exp->exp_lock);
6037 /* self-export doesn't need client data and ldlm initialization */
6038 if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
6039 &exp->exp_client_uuid)))
6042 rc = tgt_client_alloc(exp);
6046 rc = ldlm_init_export(exp);
6053 tgt_client_free(exp);
6055 CERROR("%s: Failed to initialize export: rc = %d\n",
6056 exp->exp_obd->obd_name, rc);
6060 static int mdt_destroy_export(struct obd_export *exp)
6064 target_destroy_export(exp);
6065 /* destroy can be called from failed obd_setup, so
6066 * checking uuid is safer than obd_self_export */
6067 if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
6068 &exp->exp_client_uuid)))
6071 ldlm_destroy_export(exp);
6072 tgt_client_free(exp);
6074 LASSERT(list_empty(&exp->exp_outstanding_replies));
6075 LASSERT(list_empty(&exp->exp_mdt_data.med_open_head));
6078 * discard grants once we're sure no more
6079 * interaction with the client is possible
6081 tgt_grant_discard(exp);
6082 if (exp_connect_flags(exp) & OBD_CONNECT_GRANT)
6083 exp->exp_obd->u.obt.obt_lut->lut_tgd.tgd_tot_granted_clients--;
6085 if (!(exp->exp_flags & OBD_OPT_FORCE))
6086 tgt_grant_sanity_check(exp->exp_obd, __func__);
6091 int mdt_links_read(struct mdt_thread_info *info, struct mdt_object *mdt_obj,
6092 struct linkea_data *ldata)
6096 LASSERT(ldata->ld_buf->lb_buf != NULL);
6098 if (!mdt_object_exists(mdt_obj))
6101 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
6102 ldata->ld_buf, XATTR_NAME_LINK);
6103 if (rc == -ERANGE) {
6104 /* Buf was too small, figure out what we need. */
6105 lu_buf_free(ldata->ld_buf);
6106 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
6107 ldata->ld_buf, XATTR_NAME_LINK);
6110 ldata->ld_buf = lu_buf_check_and_alloc(ldata->ld_buf, rc);
6111 if (ldata->ld_buf->lb_buf == NULL)
6113 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
6114 ldata->ld_buf, XATTR_NAME_LINK);
6119 return linkea_init_with_rec(ldata);
6123 * Given an MDT object, try to look up the full path to the object.
6124 * Part of the MDT layer implementation of lfs fid2path.
6126 * \param[in] info Per-thread common data shared by MDT level handlers.
6127 * \param[in] obj Object to do path lookup of
6128 * \param[in,out] fp User-provided struct to store path information
6129 * \param[in] root_fid Root FID of current path should reach
6131 * \retval 0 Lookup successful, path information stored in fp
6132 * \retval -EAGAIN Lookup failed, usually because object is being moved
6133 * \retval negative errno if there was a problem
6135 static int mdt_path_current(struct mdt_thread_info *info,
6136 struct mdt_object *obj,
6137 struct getinfo_fid2path *fp,
6138 struct lu_fid *root_fid)
6140 struct mdt_device *mdt = info->mti_mdt;
6141 struct mdt_object *mdt_obj;
6142 struct link_ea_header *leh;
6143 struct link_ea_entry *lee;
6144 struct lu_name *tmpname = &info->mti_name;
6145 struct lu_fid *tmpfid = &info->mti_tmp_fid1;
6146 struct lu_buf *buf = &info->mti_big_buf;
6149 struct linkea_data ldata = { NULL };
6154 /* temp buffer for path element, the buffer will be finally freed
6155 * in mdt_thread_info_fini */
6156 buf = lu_buf_check_and_alloc(buf, PATH_MAX);
6157 if (buf->lb_buf == NULL)
6161 ptr = fp->gf_u.gf_path + fp->gf_pathlen - 1;
6164 *tmpfid = fp->gf_fid = *mdt_object_fid(obj);
6166 while (!lu_fid_eq(root_fid, &fp->gf_fid)) {
6167 struct lu_buf lmv_buf;
6169 if (!lu_fid_eq(root_fid, &mdt->mdt_md_root_fid) &&
6170 lu_fid_eq(&mdt->mdt_md_root_fid, &fp->gf_fid))
6171 GOTO(out, rc = -ENOENT);
6173 if (lu_fid_eq(mdt_object_fid(obj), tmpfid)) {
6175 mdt_object_get(info->mti_env, mdt_obj);
6177 mdt_obj = mdt_object_find(info->mti_env, mdt, tmpfid);
6178 if (IS_ERR(mdt_obj))
6179 GOTO(out, rc = PTR_ERR(mdt_obj));
6182 if (!mdt_object_exists(mdt_obj)) {
6183 mdt_object_put(info->mti_env, mdt_obj);
6184 GOTO(out, rc = -ENOENT);
6187 if (mdt_object_remote(mdt_obj)) {
6188 mdt_object_put(info->mti_env, mdt_obj);
6189 GOTO(remote_out, rc = -EREMOTE);
6192 rc = mdt_links_read(info, mdt_obj, &ldata);
6194 mdt_object_put(info->mti_env, mdt_obj);
6199 lee = (struct link_ea_entry *)(leh + 1); /* link #0 */
6200 linkea_entry_unpack(lee, &reclen, tmpname, tmpfid);
6201 /* If set, use link #linkno for path lookup, otherwise use
6202 link #0. Only do this for the final path element. */
6203 if (first && fp->gf_linkno < leh->leh_reccount) {
6205 for (count = 0; count < fp->gf_linkno; count++) {
6206 lee = (struct link_ea_entry *)
6207 ((char *)lee + reclen);
6208 linkea_entry_unpack(lee, &reclen, tmpname,
6211 if (fp->gf_linkno < leh->leh_reccount - 1)
6212 /* indicate to user there are more links */
6216 lmv_buf.lb_buf = info->mti_xattr_buf;
6217 lmv_buf.lb_len = sizeof(info->mti_xattr_buf);
6218 /* Check if it is slave stripes */
6219 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
6220 &lmv_buf, XATTR_NAME_LMV);
6221 mdt_object_put(info->mti_env, mdt_obj);
6223 union lmv_mds_md *lmm = lmv_buf.lb_buf;
6225 /* For slave stripes, get its master */
6226 if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_STRIPE) {
6227 fp->gf_fid = *tmpfid;
6230 } else if (rc < 0 && rc != -ENODATA) {
6236 /* Pack the name in the end of the buffer */
6237 ptr -= tmpname->ln_namelen;
6238 if (ptr - 1 <= fp->gf_u.gf_path)
6239 GOTO(out, rc = -EOVERFLOW);
6240 strncpy(ptr, tmpname->ln_name, tmpname->ln_namelen);
6243 /* keep the last resolved fid to the client, so the
6244 * client will build the left path on another MDT for
6246 fp->gf_fid = *tmpfid;
6252 ptr++; /* skip leading / */
6253 memmove(fp->gf_u.gf_path, ptr,
6254 fp->gf_u.gf_path + fp->gf_pathlen - ptr);
6261 * Given an MDT object, use mdt_path_current to get the path.
6262 * Essentially a wrapper to retry mdt_path_current a set number of times
6263 * if -EAGAIN is returned (usually because an object is being moved).
6265 * Part of the MDT layer implementation of lfs fid2path.
6267 * \param[in] info Per-thread common data shared by mdt level handlers.
6268 * \param[in] obj Object to do path lookup of
6269 * \param[in,out] fp User-provided struct for arguments and to store path
6272 * \retval 0 Lookup successful, path information stored in fp
6273 * \retval negative errno if there was a problem
6275 static int mdt_path(struct mdt_thread_info *info, struct mdt_object *obj,
6276 struct getinfo_fid2path *fp, struct lu_fid *root_fid)
6278 struct mdt_device *mdt = info->mti_mdt;
6283 if (fp->gf_pathlen < 3)
6286 if (root_fid == NULL)
6287 root_fid = &mdt->mdt_md_root_fid;
6289 if (lu_fid_eq(root_fid, mdt_object_fid(obj))) {
6290 fp->gf_u.gf_path[0] = '\0';
6294 /* Retry multiple times in case file is being moved */
6295 while (tries-- && rc == -EAGAIN)
6296 rc = mdt_path_current(info, obj, fp, root_fid);
6302 * Get the full path of the provided FID, as of changelog record recno.
6304 * This checks sanity and looks up object for user provided FID
6305 * before calling the actual path lookup code.
6307 * Part of the MDT layer implementation of lfs fid2path.
6309 * \param[in] info Per-thread common data shared by mdt level handlers.
6310 * \param[in,out] fp User-provided struct for arguments and to store path
6313 * \retval 0 Lookup successful, path information and recno stored in fp
6314 * \retval -ENOENT, object does not exist
6315 * \retval negative errno if there was a problem
6317 static int mdt_fid2path(struct mdt_thread_info *info,
6318 struct lu_fid *root_fid,
6319 struct getinfo_fid2path *fp)
6321 struct mdt_device *mdt = info->mti_mdt;
6322 struct mdt_object *obj;
6326 CDEBUG(D_IOCTL, "path get "DFID" from %llu #%d\n",
6327 PFID(&fp->gf_fid), fp->gf_recno, fp->gf_linkno);
6329 if (!fid_is_sane(&fp->gf_fid))
6332 if (!fid_is_namespace_visible(&fp->gf_fid)) {
6333 CDEBUG(D_INFO, "%s: "DFID" is invalid, f_seq should be >= %#llx"
6334 ", or f_oid != 0, or f_ver == 0\n", mdt_obd_name(mdt),
6335 PFID(&fp->gf_fid), (__u64)FID_SEQ_NORMAL);
6339 obj = mdt_object_find(info->mti_env, mdt, &fp->gf_fid);
6342 CDEBUG(D_IOCTL, "cannot find "DFID": rc = %d\n",
6343 PFID(&fp->gf_fid), rc);
6347 if (mdt_object_remote(obj))
6349 else if (!mdt_object_exists(obj))
6355 mdt_object_put(info->mti_env, obj);
6356 CDEBUG(D_IOCTL, "nonlocal object "DFID": rc = %d\n",
6357 PFID(&fp->gf_fid), rc);
6361 rc = mdt_path(info, obj, fp, root_fid);
6363 CDEBUG(D_INFO, "fid "DFID", path %s recno %#llx linkno %u\n",
6364 PFID(&fp->gf_fid), fp->gf_u.gf_path,
6365 fp->gf_recno, fp->gf_linkno);
6367 mdt_object_put(info->mti_env, obj);
6372 static int mdt_rpc_fid2path(struct mdt_thread_info *info, void *key, int keylen,
6373 void *val, int vallen)
6375 struct getinfo_fid2path *fpout, *fpin;
6376 struct lu_fid *root_fid = NULL;
6379 fpin = key + cfs_size_round(sizeof(KEY_FID2PATH));
6382 if (ptlrpc_req_need_swab(info->mti_pill->rc_req))
6383 lustre_swab_fid2path(fpin);
6385 memcpy(fpout, fpin, sizeof(*fpin));
6386 if (fpout->gf_pathlen != vallen - sizeof(*fpin))
6389 if (keylen >= cfs_size_round(sizeof(KEY_FID2PATH)) + sizeof(*fpin) +
6390 sizeof(struct lu_fid)) {
6391 /* client sent its root FID, which is normally fileset FID */
6392 root_fid = fpin->gf_u.gf_root_fid;
6393 if (ptlrpc_req_need_swab(info->mti_pill->rc_req))
6394 lustre_swab_lu_fid(root_fid);
6396 if (root_fid != NULL && !fid_is_sane(root_fid))
6400 rc = mdt_fid2path(info, root_fid, fpout);
6404 int mdt_get_info(struct tgt_session_info *tsi)
6414 key = req_capsule_client_get(tsi->tsi_pill, &RMF_GETINFO_KEY);
6416 CDEBUG(D_IOCTL, "No GETINFO key\n");
6417 RETURN(err_serious(-EFAULT));
6419 keylen = req_capsule_get_size(tsi->tsi_pill, &RMF_GETINFO_KEY,
6422 vallen = req_capsule_client_get(tsi->tsi_pill, &RMF_GETINFO_VALLEN);
6423 if (vallen == NULL) {
6424 CDEBUG(D_IOCTL, "%s: cannot get RMF_GETINFO_VALLEN buffer\n",
6425 tgt_name(tsi->tsi_tgt));
6426 RETURN(err_serious(-EFAULT));
6429 req_capsule_set_size(tsi->tsi_pill, &RMF_GETINFO_VAL, RCL_SERVER,
6431 rc = req_capsule_server_pack(tsi->tsi_pill);
6433 RETURN(err_serious(rc));
6435 valout = req_capsule_server_get(tsi->tsi_pill, &RMF_GETINFO_VAL);
6436 if (valout == NULL) {
6437 CDEBUG(D_IOCTL, "%s: cannot get get-info RPC out buffer\n",
6438 tgt_name(tsi->tsi_tgt));
6439 RETURN(err_serious(-EFAULT));
6442 if (KEY_IS(KEY_FID2PATH)) {
6443 struct mdt_thread_info *info = tsi2mdt_info(tsi);
6445 rc = mdt_rpc_fid2path(info, key, keylen, valout, *vallen);
6446 mdt_thread_info_fini(info);
6453 static int mdt_ioc_version_get(struct mdt_thread_info *mti, void *karg)
6455 struct obd_ioctl_data *data = karg;
6458 struct mdt_object *obj;
6459 struct mdt_lock_handle *lh;
6463 if (data->ioc_inlbuf1 == NULL || data->ioc_inllen1 != sizeof(*fid) ||
6464 data->ioc_inlbuf2 == NULL || data->ioc_inllen2 != sizeof(version))
6467 fid = (struct lu_fid *)data->ioc_inlbuf1;
6469 if (!fid_is_sane(fid))
6472 CDEBUG(D_IOCTL, "getting version for "DFID"\n", PFID(fid));
6474 lh = &mti->mti_lh[MDT_LH_PARENT];
6475 mdt_lock_reg_init(lh, LCK_CR);
6477 obj = mdt_object_find_lock(mti, fid, lh, MDS_INODELOCK_UPDATE);
6479 RETURN(PTR_ERR(obj));
6481 if (mdt_object_remote(obj)) {
6484 * before calling version get the correct MDS should be
6485 * fid, this is error to find remote object here
6487 CERROR("nonlocal object "DFID"\n", PFID(fid));
6488 } else if (!mdt_object_exists(obj)) {
6489 *(__u64 *)data->ioc_inlbuf2 = ENOENT_VERSION;
6492 version = dt_version_get(mti->mti_env, mdt_obj2dt(obj));
6493 *(__u64 *)data->ioc_inlbuf2 = version;
6496 mdt_object_unlock_put(mti, obj, lh, 1);
6500 /* ioctls on obd dev */
6501 static int mdt_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
6502 void *karg, void __user *uarg)
6505 struct obd_device *obd = exp->exp_obd;
6506 struct mdt_device *mdt = mdt_dev(obd->obd_lu_dev);
6507 struct dt_device *dt = mdt->mdt_bottom;
6511 CDEBUG(D_IOCTL, "handling ioctl cmd %#x\n", cmd);
6512 rc = lu_env_init(&env, LCT_MD_THREAD);
6518 rc = mdt_device_sync(&env, mdt);
6520 case OBD_IOC_SET_READONLY:
6521 rc = dt_sync(&env, dt);
6523 rc = dt_ro(&env, dt);
6525 case OBD_IOC_ABORT_RECOVERY:
6526 CERROR("%s: Aborting recovery for device\n", mdt_obd_name(mdt));
6527 obd->obd_abort_recovery = 1;
6528 target_stop_recovery_thread(obd);
6531 case OBD_IOC_CHANGELOG_REG:
6532 case OBD_IOC_CHANGELOG_DEREG:
6533 case OBD_IOC_CHANGELOG_CLEAR:
6534 rc = mdt->mdt_child->md_ops->mdo_iocontrol(&env,
6538 case OBD_IOC_START_LFSCK: {
6539 struct md_device *next = mdt->mdt_child;
6540 struct obd_ioctl_data *data = karg;
6541 struct lfsck_start_param lsp;
6543 if (unlikely(data == NULL)) {
6548 lsp.lsp_start = (struct lfsck_start *)(data->ioc_inlbuf1);
6549 lsp.lsp_index_valid = 0;
6550 rc = next->md_ops->mdo_iocontrol(&env, next, cmd, 0, &lsp);
6553 case OBD_IOC_STOP_LFSCK: {
6554 struct md_device *next = mdt->mdt_child;
6555 struct obd_ioctl_data *data = karg;
6556 struct lfsck_stop stop;
6558 stop.ls_status = LS_STOPPED;
6559 /* Old lfsck utils may pass NULL @stop. */
6560 if (data->ioc_inlbuf1 == NULL)
6564 ((struct lfsck_stop *)(data->ioc_inlbuf1))->ls_flags;
6566 rc = next->md_ops->mdo_iocontrol(&env, next, cmd, 0, &stop);
6569 case OBD_IOC_QUERY_LFSCK: {
6570 struct md_device *next = mdt->mdt_child;
6571 struct obd_ioctl_data *data = karg;
6573 rc = next->md_ops->mdo_iocontrol(&env, next, cmd, 0,
6577 case OBD_IOC_GET_OBJ_VERSION: {
6578 struct mdt_thread_info *mti;
6579 mti = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
6580 memset(mti, 0, sizeof *mti);
6581 mti->mti_env = &env;
6585 rc = mdt_ioc_version_get(mti, karg);
6588 case OBD_IOC_CATLOGLIST: {
6589 struct mdt_thread_info *mti;
6591 mti = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
6592 lu_local_obj_fid(&mti->mti_tmp_fid1, LLOG_CATALOGS_OID);
6593 rc = llog_catalog_list(&env, mdt->mdt_bottom, 0, karg,
6594 &mti->mti_tmp_fid1);
6599 CERROR("%s: Not supported cmd = %d, rc = %d\n",
6600 mdt_obd_name(mdt), cmd, rc);
6607 static int mdt_postrecov(const struct lu_env *env, struct mdt_device *mdt)
6609 struct lu_device *ld = md2lu_dev(mdt->mdt_child);
6613 if (!mdt->mdt_skip_lfsck && !mdt->mdt_bottom->dd_rdonly) {
6614 struct lfsck_start_param lsp;
6616 lsp.lsp_start = NULL;
6617 lsp.lsp_index_valid = 0;
6618 rc = mdt->mdt_child->md_ops->mdo_iocontrol(env, mdt->mdt_child,
6619 OBD_IOC_START_LFSCK,
6621 if (rc != 0 && rc != -EALREADY)
6622 CWARN("%s: auto trigger paused LFSCK failed: rc = %d\n",
6623 mdt_obd_name(mdt), rc);
6626 rc = ld->ld_ops->ldo_recovery_complete(env, ld);
6630 static int mdt_obd_postrecov(struct obd_device *obd)
6635 rc = lu_env_init(&env, LCT_MD_THREAD);
6638 rc = mdt_postrecov(&env, mdt_dev(obd->obd_lu_dev));
6643 static struct obd_ops mdt_obd_device_ops = {
6644 .o_owner = THIS_MODULE,
6645 .o_set_info_async = mdt_obd_set_info_async,
6646 .o_connect = mdt_obd_connect,
6647 .o_reconnect = mdt_obd_reconnect,
6648 .o_disconnect = mdt_obd_disconnect,
6649 .o_init_export = mdt_init_export,
6650 .o_destroy_export = mdt_destroy_export,
6651 .o_iocontrol = mdt_iocontrol,
6652 .o_postrecov = mdt_obd_postrecov,
6653 /* Data-on-MDT IO methods */
6654 .o_preprw = mdt_obd_preprw,
6655 .o_commitrw = mdt_obd_commitrw,
6658 static struct lu_device* mdt_device_fini(const struct lu_env *env,
6659 struct lu_device *d)
6661 struct mdt_device *m = mdt_dev(d);
6668 static struct lu_device *mdt_device_free(const struct lu_env *env,
6669 struct lu_device *d)
6671 struct mdt_device *m = mdt_dev(d);
6674 lu_device_fini(&m->mdt_lu_dev);
6680 static struct lu_device *mdt_device_alloc(const struct lu_env *env,
6681 struct lu_device_type *t,
6682 struct lustre_cfg *cfg)
6684 struct lu_device *l;
6685 struct mdt_device *m;
6692 rc = mdt_init0(env, m, t, cfg);
6694 mdt_device_free(env, l);
6699 l = ERR_PTR(-ENOMEM);
6703 /* context key constructor/destructor: mdt_key_init, mdt_key_fini */
6704 LU_KEY_INIT(mdt, struct mdt_thread_info);
6706 static void mdt_key_fini(const struct lu_context *ctx,
6707 struct lu_context_key *key, void* data)
6709 struct mdt_thread_info *info = data;
6711 if (info->mti_big_lmm) {
6712 OBD_FREE_LARGE(info->mti_big_lmm, info->mti_big_lmmsize);
6713 info->mti_big_lmm = NULL;
6714 info->mti_big_lmmsize = 0;
6717 if (info->mti_big_acl) {
6718 OBD_FREE_LARGE(info->mti_big_acl, info->mti_big_aclsize);
6719 info->mti_big_acl = NULL;
6720 info->mti_big_aclsize = 0;
6726 /* context key: mdt_thread_key */
6727 LU_CONTEXT_KEY_DEFINE(mdt, LCT_MD_THREAD);
6729 struct lu_ucred *mdt_ucred(const struct mdt_thread_info *info)
6731 return lu_ucred(info->mti_env);
6734 struct lu_ucred *mdt_ucred_check(const struct mdt_thread_info *info)
6736 return lu_ucred_check(info->mti_env);
6740 * Enable/disable COS (Commit On Sharing).
6742 * Set/Clear the COS flag in mdt options.
6744 * \param mdt mdt device
6745 * \param val 0 disables COS, other values enable COS
6747 void mdt_enable_cos(struct mdt_device *mdt, bool val)
6752 mdt->mdt_opts.mo_cos = val;
6753 rc = lu_env_init(&env, LCT_LOCAL);
6754 if (unlikely(rc != 0)) {
6755 CWARN("%s: lu_env initialization failed, cannot "
6756 "sync: rc = %d\n", mdt_obd_name(mdt), rc);
6759 mdt_device_sync(&env, mdt);
6764 * Check COS (Commit On Sharing) status.
6766 * Return COS flag status.
6768 * \param mdt mdt device
6770 int mdt_cos_is_enabled(struct mdt_device *mdt)
6772 return mdt->mdt_opts.mo_cos != 0;
6775 static struct lu_device_type_operations mdt_device_type_ops = {
6776 .ldto_device_alloc = mdt_device_alloc,
6777 .ldto_device_free = mdt_device_free,
6778 .ldto_device_fini = mdt_device_fini
6781 static struct lu_device_type mdt_device_type = {
6782 .ldt_tags = LU_DEVICE_MD,
6783 .ldt_name = LUSTRE_MDT_NAME,
6784 .ldt_ops = &mdt_device_type_ops,
6785 .ldt_ctx_tags = LCT_MD_THREAD
6788 static int __init mdt_init(void)
6792 CLASSERT(sizeof("0x0123456789ABCDEF:0x01234567:0x01234567") ==
6793 FID_NOBRACE_LEN + 1);
6794 CLASSERT(sizeof("[0x0123456789ABCDEF:0x01234567:0x01234567]") ==
6796 rc = lu_kmem_init(mdt_caches);
6800 rc = mds_mod_init();
6804 rc = class_register_type(&mdt_obd_device_ops, NULL, true, NULL,
6805 LUSTRE_MDT_NAME, &mdt_device_type);
6810 lu_kmem_fini(mdt_caches);
6817 static void __exit mdt_exit(void)
6819 class_unregister_type(LUSTRE_MDT_NAME);
6821 lu_kmem_fini(mdt_caches);
6824 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
6825 MODULE_DESCRIPTION("Lustre Metadata Target ("LUSTRE_MDT_NAME")");
6826 MODULE_VERSION(LUSTRE_VERSION_STRING);
6827 MODULE_LICENSE("GPL");
6829 module_init(mdt_init);
6830 module_exit(mdt_exit);