4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2010, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/mdt/mdt_handler.c
34 * Lustre Metadata Target (mdt) request handler
36 * Author: Peter Braam <braam@clusterfs.com>
37 * Author: Andreas Dilger <adilger@clusterfs.com>
38 * Author: Phil Schwan <phil@clusterfs.com>
39 * Author: Mike Shaver <shaver@clusterfs.com>
40 * Author: Nikita Danilov <nikita@clusterfs.com>
41 * Author: Huang Hua <huanghua@clusterfs.com>
42 * Author: Yury Umanets <umka@clusterfs.com>
45 #define DEBUG_SUBSYSTEM S_MDS
47 #include <linux/module.h>
48 #include <linux/pagemap.h>
50 #include <dt_object.h>
51 #include <lustre_acl.h>
52 #include <lustre_export.h>
53 #include <uapi/linux/lustre/lustre_ioctl.h>
54 #include <lustre_lfsck.h>
55 #include <lustre_log.h>
56 #include <lustre_nodemap.h>
57 #include <lustre_mds.h>
58 #include <uapi/linux/lustre/lustre_param.h>
59 #include <lustre_quota.h>
60 #include <lustre_swab.h>
62 #include <obd_support.h>
63 #include <lustre_barrier.h>
64 #include <obd_cksum.h>
65 #include <llog_swab.h>
67 #include "mdt_internal.h"
69 static unsigned int max_mod_rpcs_per_client = 8;
70 module_param(max_mod_rpcs_per_client, uint, 0644);
71 MODULE_PARM_DESC(max_mod_rpcs_per_client, "maximum number of modify RPCs in flight allowed per client");
73 mdl_mode_t mdt_mdl_lock_modes[] = {
74 [LCK_MINMODE] = MDL_MINMODE,
81 [LCK_GROUP] = MDL_GROUP
84 enum ldlm_mode mdt_dlm_lock_modes[] = {
85 [MDL_MINMODE] = LCK_MINMODE,
92 [MDL_GROUP] = LCK_GROUP
95 static struct mdt_device *mdt_dev(struct lu_device *d);
97 static const struct lu_object_operations mdt_obj_ops;
99 /* Slab for MDT object allocation */
100 static struct kmem_cache *mdt_object_kmem;
102 /* For HSM restore handles */
103 struct kmem_cache *mdt_hsm_cdt_kmem;
105 /* For HSM request handles */
106 struct kmem_cache *mdt_hsm_car_kmem;
108 static struct lu_kmem_descr mdt_caches[] = {
110 .ckd_cache = &mdt_object_kmem,
111 .ckd_name = "mdt_obj",
112 .ckd_size = sizeof(struct mdt_object)
115 .ckd_cache = &mdt_hsm_cdt_kmem,
116 .ckd_name = "mdt_cdt_restore_handle",
117 .ckd_size = sizeof(struct cdt_restore_handle)
120 .ckd_cache = &mdt_hsm_car_kmem,
121 .ckd_name = "mdt_cdt_agent_req",
122 .ckd_size = sizeof(struct cdt_agent_req)
129 __u64 mdt_get_disposition(struct ldlm_reply *rep, __u64 op_flag)
133 return rep->lock_policy_res1 & op_flag;
136 void mdt_clear_disposition(struct mdt_thread_info *info,
137 struct ldlm_reply *rep, __u64 op_flag)
140 info->mti_opdata &= ~op_flag;
141 tgt_opdata_clear(info->mti_env, op_flag);
144 rep->lock_policy_res1 &= ~op_flag;
147 void mdt_set_disposition(struct mdt_thread_info *info,
148 struct ldlm_reply *rep, __u64 op_flag)
151 info->mti_opdata |= op_flag;
152 tgt_opdata_set(info->mti_env, op_flag);
155 rep->lock_policy_res1 |= op_flag;
158 void mdt_lock_reg_init(struct mdt_lock_handle *lh, enum ldlm_mode lm)
160 lh->mlh_pdo_hash = 0;
161 lh->mlh_reg_mode = lm;
162 lh->mlh_rreg_mode = lm;
163 lh->mlh_type = MDT_REG_LOCK;
166 void mdt_lock_pdo_init(struct mdt_lock_handle *lh, enum ldlm_mode lock_mode,
167 const struct lu_name *lname)
169 lh->mlh_reg_mode = lock_mode;
170 lh->mlh_pdo_mode = LCK_MINMODE;
171 lh->mlh_rreg_mode = lock_mode;
172 lh->mlh_type = MDT_PDO_LOCK;
174 if (lu_name_is_valid(lname)) {
175 lh->mlh_pdo_hash = ll_full_name_hash(NULL, lname->ln_name,
177 /* XXX Workaround for LU-2856
179 * Zero is a valid return value of full_name_hash, but
180 * several users of mlh_pdo_hash assume a non-zero
181 * hash value. We therefore map zero onto an
182 * arbitrary, but consistent value (1) to avoid
183 * problems further down the road. */
184 if (unlikely(lh->mlh_pdo_hash == 0))
185 lh->mlh_pdo_hash = 1;
187 lh->mlh_pdo_hash = 0;
191 static void mdt_lock_pdo_mode(struct mdt_thread_info *info, struct mdt_object *o,
192 struct mdt_lock_handle *lh)
198 * Any dir access needs couple of locks:
200 * 1) on part of dir we gonna take lookup/modify;
202 * 2) on whole dir to protect it from concurrent splitting and/or to
203 * flush client's cache for readdir().
205 * so, for a given mode and object this routine decides what lock mode
206 * to use for lock #2:
208 * 1) if caller's gonna lookup in dir then we need to protect dir from
209 * being splitted only - LCK_CR
211 * 2) if caller's gonna modify dir then we need to protect dir from
212 * being splitted and to flush cache - LCK_CW
214 * 3) if caller's gonna modify dir and that dir seems ready for
215 * splitting then we need to protect it from any type of access
216 * (lookup/modify/split) - LCK_EX --bzzz
219 LASSERT(lh->mlh_reg_mode != LCK_MINMODE);
220 LASSERT(lh->mlh_pdo_mode == LCK_MINMODE);
223 * Ask underlaying level its opinion about preferable PDO lock mode
224 * having access type passed as regular lock mode:
226 * - MDL_MINMODE means that lower layer does not want to specify lock
229 * - MDL_NL means that no PDO lock should be taken. This is used in some
230 * cases. Say, for non-splittable directories no need to use PDO locks
233 mode = mdo_lock_mode(info->mti_env, mdt_object_child(o),
234 mdt_dlm_mode2mdl_mode(lh->mlh_reg_mode));
236 if (mode != MDL_MINMODE) {
237 lh->mlh_pdo_mode = mdt_mdl_mode2dlm_mode(mode);
240 * Lower layer does not want to specify locking mode. We do it
241 * our selves. No special protection is needed, just flush
242 * client's cache on modification and allow concurrent
245 switch (lh->mlh_reg_mode) {
247 lh->mlh_pdo_mode = LCK_EX;
250 lh->mlh_pdo_mode = LCK_CR;
253 lh->mlh_pdo_mode = LCK_CW;
256 CERROR("Not expected lock type (0x%x)\n",
257 (int)lh->mlh_reg_mode);
262 LASSERT(lh->mlh_pdo_mode != LCK_MINMODE);
266 static int mdt_lookup_fileset(struct mdt_thread_info *info, const char *fileset,
269 struct mdt_device *mdt = info->mti_mdt;
270 struct lu_name *lname = &info->mti_name;
271 char *filename = info->mti_filename;
272 struct mdt_object *parent;
276 LASSERT(!info->mti_cross_ref);
279 * We may want to allow this to mount a completely separate
280 * fileset from the MDT in the future, but keeping it to
281 * ROOT/ only for now avoid potential security issues.
283 *fid = mdt->mdt_md_root_fid;
285 while (rc == 0 && fileset != NULL && *fileset != '\0') {
286 const char *s1 = fileset;
292 while (*s2 != '/' && *s2 != '\0')
300 lname->ln_namelen = s2 - s1;
301 if (lname->ln_namelen > NAME_MAX) {
306 /* reject .. as a path component */
307 if (lname->ln_namelen == 2 &&
308 strncmp(s1, "..", 2) == 0) {
313 strncpy(filename, s1, lname->ln_namelen);
314 filename[lname->ln_namelen] = '\0';
315 lname->ln_name = filename;
317 parent = mdt_object_find(info->mti_env, mdt, fid);
318 if (IS_ERR(parent)) {
319 rc = PTR_ERR(parent);
322 /* Only got the fid of this obj by name */
324 rc = mdo_lookup(info->mti_env, mdt_object_child(parent), lname,
325 fid, &info->mti_spec);
326 mdt_object_put(info->mti_env, parent);
329 parent = mdt_object_find(info->mti_env, mdt, fid);
331 rc = PTR_ERR(parent);
333 mode = lu_object_attr(&parent->mot_obj);
334 mdt_object_put(info->mti_env, parent);
343 static int mdt_get_root(struct tgt_session_info *tsi)
345 struct mdt_thread_info *info = tsi2mdt_info(tsi);
346 struct mdt_device *mdt = info->mti_mdt;
347 struct mdt_body *repbody;
348 char *fileset = NULL, *buffer = NULL;
350 struct obd_export *exp = info->mti_exp;
351 char *nodemap_fileset;
355 rc = mdt_check_ucred(info);
357 GOTO(out, rc = err_serious(rc));
359 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GET_ROOT_PACK))
360 GOTO(out, rc = err_serious(-ENOMEM));
362 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
363 if (req_capsule_get_size(info->mti_pill, &RMF_NAME, RCL_CLIENT) > 0) {
364 fileset = req_capsule_client_get(info->mti_pill, &RMF_NAME);
366 GOTO(out, rc = err_serious(-EFAULT));
369 nodemap_fileset = nodemap_get_fileset(exp->exp_target_data.ted_nodemap);
370 if (nodemap_fileset && nodemap_fileset[0]) {
371 CDEBUG(D_INFO, "nodemap fileset is %s\n", nodemap_fileset);
373 /* consider fileset from client as a sub-fileset
374 * of the nodemap one */
375 OBD_ALLOC(buffer, PATH_MAX + 1);
377 GOTO(out, rc = err_serious(-ENOMEM));
378 if (snprintf(buffer, PATH_MAX + 1, "%s/%s",
379 nodemap_fileset, fileset) >= PATH_MAX + 1)
380 GOTO(out, rc = err_serious(-EINVAL));
383 /* enforce fileset as specified in the nodemap */
384 fileset = nodemap_fileset;
389 CDEBUG(D_INFO, "Getting fileset %s\n", fileset);
390 rc = mdt_lookup_fileset(info, fileset, &repbody->mbo_fid1);
392 GOTO(out, rc = err_serious(rc));
394 repbody->mbo_fid1 = mdt->mdt_md_root_fid;
396 repbody->mbo_valid |= OBD_MD_FLID;
400 mdt_thread_info_fini(info);
402 OBD_FREE(buffer, PATH_MAX+1);
406 static int mdt_statfs(struct tgt_session_info *tsi)
408 struct ptlrpc_request *req = tgt_ses_req(tsi);
409 struct mdt_thread_info *info = tsi2mdt_info(tsi);
410 struct mdt_device *mdt = info->mti_mdt;
411 struct tg_grants_data *tgd = &mdt->mdt_lut.lut_tgd;
412 struct md_device *next = mdt->mdt_child;
413 struct ptlrpc_service_part *svcpt;
414 struct obd_statfs *osfs;
415 struct mdt_body *reqbody = NULL;
416 struct mdt_statfs_cache *msf;
421 svcpt = req->rq_rqbd->rqbd_svcpt;
423 /* This will trigger a watchdog timeout */
424 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_STATFS_LCW_SLEEP,
425 (MDT_SERVICE_WATCHDOG_FACTOR *
426 at_get(&svcpt->scp_at_estimate)) + 1);
428 rc = mdt_check_ucred(info);
430 GOTO(out, rc = err_serious(rc));
432 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK))
433 GOTO(out, rc = err_serious(-ENOMEM));
435 osfs = req_capsule_server_get(info->mti_pill, &RMF_OBD_STATFS);
437 GOTO(out, rc = -EPROTO);
439 if (mdt_is_sum_statfs_client(req->rq_export) &&
440 lustre_packed_msg_size(req->rq_reqmsg) ==
441 req_capsule_fmt_size(req->rq_reqmsg->lm_magic,
442 &RQF_MDS_STATFS_NEW, RCL_CLIENT)) {
443 req_capsule_extend(info->mti_pill, &RQF_MDS_STATFS_NEW);
444 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
447 if (reqbody && reqbody->mbo_valid & OBD_MD_FLAGSTATFS)
448 msf = &mdt->mdt_sum_osfs;
450 msf = &mdt->mdt_osfs;
452 if (msf->msf_age + OBD_STATFS_CACHE_SECONDS <= ktime_get_seconds()) {
453 /** statfs data is too old, get up-to-date one */
454 if (reqbody && reqbody->mbo_valid & OBD_MD_FLAGSTATFS)
455 rc = next->md_ops->mdo_statfs(info->mti_env,
458 rc = dt_statfs(info->mti_env, mdt->mdt_bottom,
462 spin_lock(&mdt->mdt_lock);
463 msf->msf_osfs = *osfs;
464 msf->msf_age = ktime_get_seconds();
465 spin_unlock(&mdt->mdt_lock);
467 /** use cached statfs data */
468 spin_lock(&mdt->mdt_lock);
469 *osfs = msf->msf_osfs;
470 spin_unlock(&mdt->mdt_lock);
473 /* at least try to account for cached pages. its still racy and
474 * might be under-reporting if clients haven't announced their
475 * caches with brw recently */
476 CDEBUG(D_SUPER | D_CACHE, "blocks cached %llu granted %llu"
477 " pending %llu free %llu avail %llu\n",
478 tgd->tgd_tot_dirty, tgd->tgd_tot_granted,
479 tgd->tgd_tot_pending,
480 osfs->os_bfree << tgd->tgd_blockbits,
481 osfs->os_bavail << tgd->tgd_blockbits);
483 osfs->os_bavail -= min_t(u64, osfs->os_bavail,
484 ((tgd->tgd_tot_dirty + tgd->tgd_tot_pending +
485 osfs->os_bsize - 1) >> tgd->tgd_blockbits));
487 tgt_grant_sanity_check(mdt->mdt_lu_dev.ld_obd, __func__);
488 CDEBUG(D_CACHE, "%llu blocks: %llu free, %llu avail; "
489 "%llu objects: %llu free; state %x\n",
490 osfs->os_blocks, osfs->os_bfree, osfs->os_bavail,
491 osfs->os_files, osfs->os_ffree, osfs->os_state);
493 if (!exp_grant_param_supp(tsi->tsi_exp) &&
494 tgd->tgd_blockbits > COMPAT_BSIZE_SHIFT) {
495 /* clients which don't support OBD_CONNECT_GRANT_PARAM
496 * should not see a block size > page size, otherwise
497 * cl_lost_grant goes mad. Therefore, we emulate a 4KB (=2^12)
498 * block size which is the biggest block size known to work
499 * with all client's page size. */
500 osfs->os_blocks <<= tgd->tgd_blockbits - COMPAT_BSIZE_SHIFT;
501 osfs->os_bfree <<= tgd->tgd_blockbits - COMPAT_BSIZE_SHIFT;
502 osfs->os_bavail <<= tgd->tgd_blockbits - COMPAT_BSIZE_SHIFT;
503 osfs->os_bsize = 1 << COMPAT_BSIZE_SHIFT;
506 mdt_counter_incr(req, LPROC_MDT_STATFS);
508 mdt_thread_info_fini(info);
513 * Pack size attributes into the reply.
515 int mdt_pack_size2body(struct mdt_thread_info *info,
516 const struct lu_fid *fid, struct lustre_handle *lh)
519 struct md_attr *ma = &info->mti_attr;
521 bool dom_lock = false;
525 LASSERT(ma->ma_attr.la_valid & LA_MODE);
527 if (!S_ISREG(ma->ma_attr.la_mode) ||
528 !(ma->ma_valid & MA_LOV && ma->ma_lmm != NULL))
531 dom_stripe = mdt_lmm_dom_entry(ma->ma_lmm);
532 /* no DoM stripe, no size in reply */
533 if (dom_stripe == LMM_NO_DOM)
536 if (lustre_handle_is_used(lh)) {
537 struct ldlm_lock *lock;
539 lock = ldlm_handle2lock(lh);
541 dom_lock = ldlm_has_dom(lock);
546 /* no DoM lock, no size in reply */
550 /* Either DoM lock exists or LMM has only DoM stripe then
551 * return size on body. */
552 b = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
554 mdt_dom_object_size(info->mti_env, info->mti_mdt, fid, b, dom_lock);
558 #ifdef CONFIG_LUSTRE_FS_POSIX_ACL
560 * Pack ACL data into the reply. UIDs/GIDs are mapped and filtered by nodemap.
562 * \param info thread info object
563 * \param repbody reply to pack ACLs into
564 * \param o mdt object of file to examine
565 * \param nodemap nodemap of client to reply to
567 * \retval -errno error getting or parsing ACL from disk
569 int mdt_pack_acl2body(struct mdt_thread_info *info, struct mdt_body *repbody,
570 struct mdt_object *o, struct lu_nodemap *nodemap)
572 const struct lu_env *env = info->mti_env;
573 struct md_object *next = mdt_object_child(o);
574 struct lu_buf *buf = &info->mti_buf;
575 struct mdt_device *mdt = info->mti_mdt;
576 struct req_capsule *pill = info->mti_pill;
581 buf->lb_buf = req_capsule_server_get(pill, &RMF_ACL);
582 buf->lb_len = req_capsule_get_size(pill, &RMF_ACL, RCL_SERVER);
583 if (buf->lb_len == 0)
587 rc = mo_xattr_get(env, next, buf, XATTR_NAME_ACL_ACCESS);
589 if (rc == -ENODATA) {
590 repbody->mbo_aclsize = 0;
591 repbody->mbo_valid |= OBD_MD_FLACL;
593 } else if (rc == -EOPNOTSUPP) {
597 exp_connect_large_acl(info->mti_exp) &&
598 buf->lb_buf != info->mti_big_acl) {
599 if (info->mti_big_acl == NULL) {
600 info->mti_big_aclsize =
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;
804 } else if (ma->ma_valid & MA_SOM) { /* lsom is valid */
805 b->mbo_valid |= OBD_MD_FLLAZYSIZE | OBD_MD_FLLAZYBLOCKS;
806 b->mbo_size = ma->ma_som.ms_size;
807 b->mbo_blocks = ma->ma_som.ms_blocks;
811 if (fid != NULL && (b->mbo_valid & OBD_MD_FLSIZE ||
812 b->mbo_valid & OBD_MD_FLLAZYSIZE))
813 CDEBUG(D_VFSTRACE, DFID": returning size %llu\n",
814 PFID(fid), (unsigned long long)b->mbo_size);
817 if (!IS_ERR_OR_NULL(nodemap))
818 nodemap_putref(nodemap);
821 static inline int mdt_body_has_lov(const struct lu_attr *la,
822 const struct mdt_body *body)
824 return (S_ISREG(la->la_mode) && (body->mbo_valid & OBD_MD_FLEASIZE)) ||
825 (S_ISDIR(la->la_mode) && (body->mbo_valid & OBD_MD_FLDIREA));
828 void mdt_client_compatibility(struct mdt_thread_info *info)
830 struct mdt_body *body;
831 struct ptlrpc_request *req = mdt_info_req(info);
832 struct obd_export *exp = req->rq_export;
833 struct md_attr *ma = &info->mti_attr;
834 struct lu_attr *la = &ma->ma_attr;
837 if (exp_connect_layout(exp))
838 /* the client can deal with 16-bit lmm_stripe_count */
841 body = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
843 if (!mdt_body_has_lov(la, body))
846 /* now we have a reply with a lov for a client not compatible with the
847 * layout lock so we have to clean the layout generation number */
848 if (S_ISREG(la->la_mode))
849 ma->ma_lmm->lmm_layout_gen = 0;
853 static int mdt_attr_get_eabuf_size(struct mdt_thread_info *info,
854 struct mdt_object *o)
856 const struct lu_env *env = info->mti_env;
859 rc = mo_xattr_get(env, mdt_object_child(o), &LU_BUF_NULL,
868 /* Is it a directory? Let's check for the LMV as well */
869 if (S_ISDIR(lu_object_attr(&mdt_object_child(o)->mo_lu))) {
870 rc2 = mo_xattr_get(env, mdt_object_child(o), &LU_BUF_NULL,
874 rc2 = mo_xattr_get(env, mdt_object_child(o),
876 XATTR_NAME_DEFAULT_LMV);
878 if ((rc2 < 0 && rc2 != -ENODATA) || (rc2 > rc))
886 int mdt_big_xattr_get(struct mdt_thread_info *info, struct mdt_object *o,
889 const struct lu_env *env = info->mti_env;
893 LASSERT(info->mti_big_lmm_used == 0);
894 rc = mo_xattr_get(env, mdt_object_child(o), &LU_BUF_NULL, name);
898 /* big_lmm may need to be grown */
899 if (info->mti_big_lmmsize < rc) {
900 int size = size_roundup_power2(rc);
902 if (info->mti_big_lmmsize > 0) {
903 /* free old buffer */
904 LASSERT(info->mti_big_lmm);
905 OBD_FREE_LARGE(info->mti_big_lmm,
906 info->mti_big_lmmsize);
907 info->mti_big_lmm = NULL;
908 info->mti_big_lmmsize = 0;
911 OBD_ALLOC_LARGE(info->mti_big_lmm, size);
912 if (info->mti_big_lmm == NULL)
914 info->mti_big_lmmsize = size;
916 LASSERT(info->mti_big_lmmsize >= rc);
918 info->mti_buf.lb_buf = info->mti_big_lmm;
919 info->mti_buf.lb_len = info->mti_big_lmmsize;
920 rc = mo_xattr_get(env, mdt_object_child(o), &info->mti_buf, name);
925 int mdt_stripe_get(struct mdt_thread_info *info, struct mdt_object *o,
926 struct md_attr *ma, const char *name)
928 struct md_object *next = mdt_object_child(o);
929 struct lu_buf *buf = &info->mti_buf;
932 if (strcmp(name, XATTR_NAME_LOV) == 0) {
933 buf->lb_buf = ma->ma_lmm;
934 buf->lb_len = ma->ma_lmm_size;
935 LASSERT(!(ma->ma_valid & MA_LOV));
936 } else if (strcmp(name, XATTR_NAME_LMV) == 0) {
937 buf->lb_buf = ma->ma_lmv;
938 buf->lb_len = ma->ma_lmv_size;
939 LASSERT(!(ma->ma_valid & MA_LMV));
940 } else if (strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0) {
941 buf->lb_buf = ma->ma_default_lmv;
942 buf->lb_len = ma->ma_default_lmv_size;
943 LASSERT(!(ma->ma_valid & MA_LMV_DEF));
948 LASSERT(buf->lb_buf);
950 rc = mo_xattr_get(info->mti_env, next, buf, name);
954 if (strcmp(name, XATTR_NAME_LOV) == 0) {
955 if (info->mti_big_lmm_used)
956 ma->ma_lmm = info->mti_big_lmm;
958 /* NOT return LOV EA with hole to old client. */
959 if (unlikely(le32_to_cpu(ma->ma_lmm->lmm_pattern) &
960 LOV_PATTERN_F_HOLE) &&
961 !(exp_connect_flags(info->mti_exp) &
962 OBD_CONNECT_LFSCK)) {
965 ma->ma_lmm_size = rc;
966 ma->ma_valid |= MA_LOV;
968 } else if (strcmp(name, XATTR_NAME_LMV) == 0) {
969 if (info->mti_big_lmm_used)
970 ma->ma_lmv = info->mti_big_lmm;
972 ma->ma_lmv_size = rc;
973 ma->ma_valid |= MA_LMV;
974 } else if (strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0) {
975 ma->ma_default_lmv_size = rc;
976 ma->ma_valid |= MA_LMV_DEF;
979 /* Update mdt_max_mdsize so all clients will be aware that */
980 if (info->mti_mdt->mdt_max_mdsize < rc)
981 info->mti_mdt->mdt_max_mdsize = rc;
984 } else if (rc == -ENODATA) {
987 } else if (rc == -ERANGE) {
988 /* Default LMV has fixed size, so it must be able to fit
989 * in the original buffer */
990 if (strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0)
992 rc = mdt_big_xattr_get(info, o, name);
994 info->mti_big_lmm_used = 1;
1002 int mdt_attr_get_pfid(struct mdt_thread_info *info, struct mdt_object *o,
1003 struct lu_fid *pfid)
1005 struct lu_buf *buf = &info->mti_buf;
1006 struct link_ea_header *leh;
1007 struct link_ea_entry *lee;
1011 buf->lb_buf = info->mti_big_lmm;
1012 buf->lb_len = info->mti_big_lmmsize;
1013 rc = mo_xattr_get(info->mti_env, mdt_object_child(o),
1014 buf, XATTR_NAME_LINK);
1015 /* ignore errors, MA_PFID won't be set and it is
1016 * up to the caller to treat this as an error */
1017 if (rc == -ERANGE || buf->lb_len == 0) {
1018 rc = mdt_big_xattr_get(info, o, XATTR_NAME_LINK);
1019 buf->lb_buf = info->mti_big_lmm;
1020 buf->lb_len = info->mti_big_lmmsize;
1025 if (rc < sizeof(*leh)) {
1026 CERROR("short LinkEA on "DFID": rc = %d\n",
1027 PFID(mdt_object_fid(o)), rc);
1031 leh = (struct link_ea_header *) buf->lb_buf;
1032 lee = (struct link_ea_entry *)(leh + 1);
1033 if (leh->leh_magic == __swab32(LINK_EA_MAGIC)) {
1034 leh->leh_magic = LINK_EA_MAGIC;
1035 leh->leh_reccount = __swab32(leh->leh_reccount);
1036 leh->leh_len = __swab64(leh->leh_len);
1038 if (leh->leh_magic != LINK_EA_MAGIC)
1040 if (leh->leh_reccount == 0)
1043 memcpy(pfid, &lee->lee_parent_fid, sizeof(*pfid));
1044 fid_be_to_cpu(pfid, pfid);
1049 int mdt_attr_get_complex(struct mdt_thread_info *info,
1050 struct mdt_object *o, struct md_attr *ma)
1052 const struct lu_env *env = info->mti_env;
1053 struct md_object *next = mdt_object_child(o);
1054 struct lu_buf *buf = &info->mti_buf;
1055 int need = ma->ma_need;
1062 if (mdt_object_exists(o) == 0)
1063 GOTO(out, rc = -ENOENT);
1064 mode = lu_object_attr(&next->mo_lu);
1066 if (need & MA_INODE) {
1067 ma->ma_need = MA_INODE;
1068 rc = mo_attr_get(env, next, ma);
1073 (void) mdt_get_som(info, o, ma);
1074 ma->ma_valid |= MA_INODE;
1077 if (need & MA_PFID) {
1078 rc = mdt_attr_get_pfid(info, o, &ma->ma_pfid);
1080 ma->ma_valid |= MA_PFID;
1081 /* ignore this error, parent fid is not mandatory */
1085 if (need & MA_LOV && (S_ISREG(mode) || S_ISDIR(mode))) {
1086 rc = mdt_stripe_get(info, o, ma, XATTR_NAME_LOV);
1091 if (need & MA_LMV && S_ISDIR(mode)) {
1092 rc = mdt_stripe_get(info, o, ma, XATTR_NAME_LMV);
1097 if (need & MA_LMV_DEF && S_ISDIR(mode)) {
1098 rc = mdt_stripe_get(info, o, ma, XATTR_NAME_DEFAULT_LMV);
1104 * In the handle of MA_INODE, we may already get the SOM attr.
1106 if (need & MA_SOM && S_ISREG(mode) && !(ma->ma_valid & MA_SOM)) {
1107 rc = mdt_get_som(info, o, ma);
1112 if (need & MA_HSM && S_ISREG(mode)) {
1113 buf->lb_buf = info->mti_xattr_buf;
1114 buf->lb_len = sizeof(info->mti_xattr_buf);
1115 BUILD_BUG_ON(sizeof(struct hsm_attrs) >
1116 sizeof(info->mti_xattr_buf));
1117 rc2 = mo_xattr_get(info->mti_env, next, buf, XATTR_NAME_HSM);
1118 rc2 = lustre_buf2hsm(info->mti_xattr_buf, rc2, &ma->ma_hsm);
1120 ma->ma_valid |= MA_HSM;
1121 else if (rc2 < 0 && rc2 != -ENODATA)
1122 GOTO(out, rc = rc2);
1125 #ifdef CONFIG_LUSTRE_FS_POSIX_ACL
1126 if (need & MA_ACL_DEF && S_ISDIR(mode)) {
1127 buf->lb_buf = ma->ma_acl;
1128 buf->lb_len = ma->ma_acl_size;
1129 rc2 = mo_xattr_get(env, next, buf, XATTR_NAME_ACL_DEFAULT);
1131 ma->ma_acl_size = rc2;
1132 ma->ma_valid |= MA_ACL_DEF;
1133 } else if (rc2 == -ENODATA) {
1135 ma->ma_acl_size = 0;
1137 GOTO(out, rc = rc2);
1142 CDEBUG(D_INODE, "after getattr rc = %d, ma_valid = %#llx ma_lmm=%p\n",
1143 rc, ma->ma_valid, ma->ma_lmm);
1147 static int mdt_getattr_internal(struct mdt_thread_info *info,
1148 struct mdt_object *o, int ma_need)
1150 struct md_object *next = mdt_object_child(o);
1151 const struct mdt_body *reqbody = info->mti_body;
1152 struct ptlrpc_request *req = mdt_info_req(info);
1153 struct md_attr *ma = &info->mti_attr;
1154 struct lu_attr *la = &ma->ma_attr;
1155 struct req_capsule *pill = info->mti_pill;
1156 const struct lu_env *env = info->mti_env;
1157 struct mdt_body *repbody;
1158 struct lu_buf *buffer = &info->mti_buf;
1159 struct obd_export *exp = info->mti_exp;
1163 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK))
1164 RETURN(err_serious(-ENOMEM));
1166 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
1170 if (mdt_object_remote(o)) {
1171 /* This object is located on remote node.*/
1172 /* Return -ENOTSUPP for old client */
1173 if (!mdt_is_dne_client(req->rq_export))
1174 GOTO(out, rc = -ENOTSUPP);
1176 repbody->mbo_fid1 = *mdt_object_fid(o);
1177 repbody->mbo_valid = OBD_MD_FLID | OBD_MD_MDS;
1181 if (reqbody->mbo_eadatasize > 0) {
1182 buffer->lb_buf = req_capsule_server_get(pill, &RMF_MDT_MD);
1183 if (buffer->lb_buf == NULL)
1184 GOTO(out, rc = -EPROTO);
1185 buffer->lb_len = req_capsule_get_size(pill, &RMF_MDT_MD,
1188 buffer->lb_buf = NULL;
1190 ma_need &= ~(MA_LOV | MA_LMV);
1191 CDEBUG(D_INFO, "%s: RPC from %s: does not need LOVEA.\n",
1192 mdt_obd_name(info->mti_mdt),
1193 req->rq_export->exp_client_uuid.uuid);
1196 /* from 2.12.58 intent_getattr pack default LMV in reply */
1197 if (S_ISDIR(lu_object_attr(&next->mo_lu)) &&
1198 ((reqbody->mbo_valid & (OBD_MD_MEA | OBD_MD_DEFAULT_MEA)) ==
1199 (OBD_MD_MEA | OBD_MD_DEFAULT_MEA)) &&
1200 req_capsule_has_field(&req->rq_pill, &RMF_DEFAULT_MDT_MD,
1202 ma->ma_lmv = buffer->lb_buf;
1203 ma->ma_lmv_size = buffer->lb_len;
1204 ma->ma_default_lmv = req_capsule_server_get(pill,
1205 &RMF_DEFAULT_MDT_MD);
1206 ma->ma_default_lmv_size = req_capsule_get_size(pill,
1207 &RMF_DEFAULT_MDT_MD,
1209 ma->ma_need = MA_INODE;
1210 if (ma->ma_lmv_size > 0)
1211 ma->ma_need |= MA_LMV;
1212 if (ma->ma_default_lmv_size > 0)
1213 ma->ma_need |= MA_LMV_DEF;
1214 } else if (S_ISDIR(lu_object_attr(&next->mo_lu)) &&
1215 (reqbody->mbo_valid & (OBD_MD_MEA | OBD_MD_DEFAULT_MEA))) {
1216 /* If it is dir and client require MEA, then we got MEA */
1217 /* Assumption: MDT_MD size is enough for lmv size. */
1218 ma->ma_lmv = buffer->lb_buf;
1219 ma->ma_lmv_size = buffer->lb_len;
1220 ma->ma_need = MA_INODE;
1221 if (ma->ma_lmv_size > 0) {
1222 if (reqbody->mbo_valid & OBD_MD_MEA) {
1223 ma->ma_need |= MA_LMV;
1224 } else if (reqbody->mbo_valid & OBD_MD_DEFAULT_MEA) {
1225 ma->ma_need |= MA_LMV_DEF;
1226 ma->ma_default_lmv = buffer->lb_buf;
1228 ma->ma_default_lmv_size = buffer->lb_len;
1229 ma->ma_lmv_size = 0;
1233 ma->ma_lmm = buffer->lb_buf;
1234 ma->ma_lmm_size = buffer->lb_len;
1235 ma->ma_need = MA_INODE | MA_HSM;
1236 if (ma->ma_lmm_size > 0) {
1237 ma->ma_need |= MA_LOV;
1238 /* Older clients may crash if they getattr overstriped
1241 if (!exp_connect_overstriping(exp) &&
1242 mdt_lmm_is_overstriping(ma->ma_lmm))
1243 RETURN(-EOPNOTSUPP);
1247 if (S_ISDIR(lu_object_attr(&next->mo_lu)) &&
1248 reqbody->mbo_valid & OBD_MD_FLDIREA &&
1249 lustre_msg_get_opc(req->rq_reqmsg) == MDS_GETATTR) {
1250 /* get default stripe info for this dir. */
1251 ma->ma_need |= MA_LOV_DEF;
1253 ma->ma_need |= ma_need;
1255 rc = mdt_attr_get_complex(info, o, ma);
1257 CDEBUG(rc == -ENOENT ? D_OTHER : D_ERROR,
1258 "%s: getattr error for "DFID": rc = %d\n",
1259 mdt_obd_name(info->mti_mdt),
1260 PFID(mdt_object_fid(o)), rc);
1264 /* if file is released, check if a restore is running */
1265 if (ma->ma_valid & MA_HSM) {
1266 repbody->mbo_valid |= OBD_MD_TSTATE;
1267 if ((ma->ma_hsm.mh_flags & HS_RELEASED) &&
1268 mdt_hsm_restore_is_running(info, mdt_object_fid(o)))
1269 repbody->mbo_t_state = MS_RESTORE;
1272 if (likely(ma->ma_valid & MA_INODE))
1273 mdt_pack_attr2body(info, repbody, la, mdt_object_fid(o));
1277 if (mdt_body_has_lov(la, reqbody)) {
1278 if (ma->ma_valid & MA_LOV) {
1279 LASSERT(ma->ma_lmm_size);
1280 repbody->mbo_eadatasize = ma->ma_lmm_size;
1281 if (S_ISDIR(la->la_mode))
1282 repbody->mbo_valid |= OBD_MD_FLDIREA;
1284 repbody->mbo_valid |= OBD_MD_FLEASIZE;
1285 mdt_dump_lmm(D_INFO, ma->ma_lmm, repbody->mbo_valid);
1287 if (ma->ma_valid & MA_LMV) {
1288 /* Return -ENOTSUPP for old client */
1289 if (!mdt_is_striped_client(req->rq_export))
1292 LASSERT(S_ISDIR(la->la_mode));
1293 mdt_dump_lmv(D_INFO, ma->ma_lmv);
1294 repbody->mbo_eadatasize = ma->ma_lmv_size;
1295 repbody->mbo_valid |= (OBD_MD_FLDIREA|OBD_MD_MEA);
1297 if (ma->ma_valid & MA_LMV_DEF) {
1298 /* Return -ENOTSUPP for old client */
1299 if (!mdt_is_striped_client(req->rq_export))
1301 LASSERT(S_ISDIR(la->la_mode));
1303 * when ll_dir_getstripe() gets default LMV, it
1304 * checks mbo_eadatasize.
1306 if (!(ma->ma_valid & MA_LMV))
1307 repbody->mbo_eadatasize =
1308 ma->ma_default_lmv_size;
1309 repbody->mbo_valid |= (OBD_MD_FLDIREA |
1310 OBD_MD_DEFAULT_MEA);
1312 } else if (S_ISLNK(la->la_mode) &&
1313 reqbody->mbo_valid & OBD_MD_LINKNAME) {
1314 buffer->lb_buf = ma->ma_lmm;
1315 /* eadatasize from client includes NULL-terminator, so
1316 * there is no need to read it */
1317 buffer->lb_len = reqbody->mbo_eadatasize - 1;
1318 rc = mo_readlink(env, next, buffer);
1319 if (unlikely(rc <= 0)) {
1320 CERROR("%s: readlink failed for "DFID": rc = %d\n",
1321 mdt_obd_name(info->mti_mdt),
1322 PFID(mdt_object_fid(o)), rc);
1325 int print_limit = min_t(int, PAGE_SIZE - 128, rc);
1327 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_READLINK_EPROTO))
1329 repbody->mbo_valid |= OBD_MD_LINKNAME;
1330 /* we need to report back size with NULL-terminator
1331 * because client expects that */
1332 repbody->mbo_eadatasize = rc + 1;
1333 if (repbody->mbo_eadatasize != reqbody->mbo_eadatasize)
1334 CDEBUG(D_INODE, "%s: Read shorter symlink %d "
1335 "on "DFID ", expected %d\n",
1336 mdt_obd_name(info->mti_mdt),
1337 rc, PFID(mdt_object_fid(o)),
1338 reqbody->mbo_eadatasize - 1);
1339 /* NULL terminate */
1340 ((char *)ma->ma_lmm)[rc] = 0;
1342 /* If the total CDEBUG() size is larger than a page, it
1343 * will print a warning to the console, avoid this by
1344 * printing just the last part of the symlink. */
1345 CDEBUG(D_INODE, "symlink dest %s%.*s, len = %d\n",
1346 print_limit < rc ? "..." : "", print_limit,
1347 (char *)ma->ma_lmm + rc - print_limit, rc);
1352 if (reqbody->mbo_valid & OBD_MD_FLMODEASIZE) {
1353 repbody->mbo_max_mdsize = info->mti_mdt->mdt_max_mdsize;
1354 repbody->mbo_valid |= OBD_MD_FLMODEASIZE;
1355 CDEBUG(D_INODE, "changing the max MD size to %u\n",
1356 repbody->mbo_max_mdsize);
1359 #ifdef CONFIG_LUSTRE_FS_POSIX_ACL
1360 if ((exp_connect_flags(req->rq_export) & OBD_CONNECT_ACL) &&
1361 (reqbody->mbo_valid & OBD_MD_FLACL)) {
1362 struct lu_nodemap *nodemap = nodemap_get_from_exp(exp);
1363 if (IS_ERR(nodemap))
1364 RETURN(PTR_ERR(nodemap));
1366 rc = mdt_pack_acl2body(info, repbody, o, nodemap);
1367 nodemap_putref(nodemap);
1373 mdt_counter_incr(req, LPROC_MDT_GETATTR);
1378 static int mdt_getattr(struct tgt_session_info *tsi)
1380 struct mdt_thread_info *info = tsi2mdt_info(tsi);
1381 struct mdt_object *obj = info->mti_object;
1382 struct req_capsule *pill = info->mti_pill;
1383 struct mdt_body *reqbody;
1384 struct mdt_body *repbody;
1388 if (unlikely(info->mti_object == NULL))
1391 reqbody = req_capsule_client_get(pill, &RMF_MDT_BODY);
1393 LASSERT(lu_object_assert_exists(&obj->mot_obj));
1395 /* Special case for Data-on-MDT files to get data version */
1396 if (unlikely(reqbody->mbo_valid & OBD_MD_FLDATAVERSION)) {
1397 rc = mdt_data_version_get(tsi);
1401 /* Unlike intent case where we need to pre-fill out buffers early on
1402 * in intent policy for ldlm reasons, here we can have a much better
1403 * guess at EA size by just reading it from disk.
1404 * Exceptions are readdir and (missing) directory striping */
1406 if (reqbody->mbo_valid & OBD_MD_LINKNAME) {
1407 /* No easy way to know how long is the symlink, but it cannot
1408 * be more than PATH_MAX, so we allocate +1 */
1410 /* A special case for fs ROOT: getattr there might fetch
1411 * default EA for entire fs, not just for this dir!
1413 } else if (lu_fid_eq(mdt_object_fid(obj),
1414 &info->mti_mdt->mdt_md_root_fid) &&
1415 (reqbody->mbo_valid & OBD_MD_FLDIREA) &&
1416 (lustre_msg_get_opc(mdt_info_req(info)->rq_reqmsg) ==
1418 /* Should the default strping be bigger, mdt_fix_reply
1419 * will reallocate */
1420 rc = DEF_REP_MD_SIZE;
1422 /* Read the actual EA size from disk */
1423 rc = mdt_attr_get_eabuf_size(info, obj);
1427 GOTO(out, rc = err_serious(rc));
1429 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER, rc);
1431 /* Set ACL reply buffer size as LUSTRE_POSIX_ACL_MAX_SIZE_OLD
1432 * by default. If the target object has more ACL entries, then
1433 * enlarge the buffer when necessary. */
1434 req_capsule_set_size(pill, &RMF_ACL, RCL_SERVER,
1435 LUSTRE_POSIX_ACL_MAX_SIZE_OLD);
1437 rc = req_capsule_server_pack(pill);
1438 if (unlikely(rc != 0))
1439 GOTO(out, rc = err_serious(rc));
1441 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
1442 LASSERT(repbody != NULL);
1443 repbody->mbo_eadatasize = 0;
1444 repbody->mbo_aclsize = 0;
1446 rc = mdt_check_ucred(info);
1448 GOTO(out_shrink, rc);
1450 info->mti_cross_ref = !!(reqbody->mbo_valid & OBD_MD_FLCROSSREF);
1452 rc = mdt_getattr_internal(info, obj, 0);
1455 mdt_client_compatibility(info);
1456 rc2 = mdt_fix_reply(info);
1460 mdt_thread_info_fini(info);
1465 * Handler of layout intent RPC requiring the layout modification
1467 * \param[in] info thread environment
1468 * \param[in] obj object
1469 * \param[in] layout layout change descriptor
1471 * \retval 0 on success
1472 * \retval < 0 error code
1474 int mdt_layout_change(struct mdt_thread_info *info, struct mdt_object *obj,
1475 struct md_layout_change *layout)
1477 struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LOCAL];
1481 if (!mdt_object_exists(obj))
1482 GOTO(out, rc = -ENOENT);
1484 if (!S_ISREG(lu_object_attr(&obj->mot_obj)))
1485 GOTO(out, rc = -EINVAL);
1487 rc = mo_permission(info->mti_env, NULL, mdt_object_child(obj), NULL,
1492 /* take layout lock to prepare layout change */
1493 mdt_lock_reg_init(lh, LCK_EX);
1494 rc = mdt_object_lock(info, obj, lh, MDS_INODELOCK_LAYOUT);
1498 mutex_lock(&obj->mot_som_mutex);
1499 rc = mo_layout_change(info->mti_env, mdt_object_child(obj), layout);
1500 mutex_unlock(&obj->mot_som_mutex);
1501 mdt_object_unlock(info, obj, lh, 1);
1507 * Exchange MOF_LOV_CREATED flags between two objects after a
1508 * layout swap. No assumption is made on whether o1 or o2 have
1509 * created objects or not.
1511 * \param[in,out] o1 First swap layout object
1512 * \param[in,out] o2 Second swap layout object
1514 static void mdt_swap_lov_flag(struct mdt_object *o1, struct mdt_object *o2)
1516 unsigned int o1_lov_created = o1->mot_lov_created;
1518 mutex_lock(&o1->mot_lov_mutex);
1519 mutex_lock(&o2->mot_lov_mutex);
1521 o1->mot_lov_created = o2->mot_lov_created;
1522 o2->mot_lov_created = o1_lov_created;
1524 mutex_unlock(&o2->mot_lov_mutex);
1525 mutex_unlock(&o1->mot_lov_mutex);
1528 static int mdt_swap_layouts(struct tgt_session_info *tsi)
1530 struct mdt_thread_info *info;
1531 struct ptlrpc_request *req = tgt_ses_req(tsi);
1532 struct obd_export *exp = req->rq_export;
1533 struct mdt_object *o1, *o2, *o;
1534 struct mdt_lock_handle *lh1, *lh2;
1535 struct mdc_swap_layouts *msl;
1539 /* client does not support layout lock, so layout swaping
1541 * FIXME: there is a problem for old clients which don't support
1542 * layout lock yet. If those clients have already opened the file
1543 * they won't be notified at all so that old layout may still be
1544 * used to do IO. This can be fixed after file release is landed by
1545 * doing exclusive open and taking full EX ibits lock. - Jinshan */
1546 if (!exp_connect_layout(exp))
1547 RETURN(-EOPNOTSUPP);
1549 info = tsi2mdt_info(tsi);
1550 if (unlikely(info->mti_object == NULL))
1553 if (info->mti_dlm_req != NULL)
1554 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1556 o1 = info->mti_object;
1557 o = o2 = mdt_object_find(info->mti_env, info->mti_mdt,
1558 &info->mti_body->mbo_fid2);
1560 GOTO(out, rc = PTR_ERR(o));
1562 if (mdt_object_remote(o) || !mdt_object_exists(o)) /* remote object */
1563 GOTO(put, rc = -ENOENT);
1565 rc = lu_fid_cmp(&info->mti_body->mbo_fid1, &info->mti_body->mbo_fid2);
1566 if (unlikely(rc == 0)) /* same file, you kidding me? no-op. */
1572 /* permission check. Make sure the calling process having permission
1573 * to write both files. */
1574 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o1), NULL,
1579 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o2), NULL,
1584 msl = req_capsule_client_get(info->mti_pill, &RMF_SWAP_LAYOUTS);
1586 GOTO(put, rc = -EPROTO);
1588 lh1 = &info->mti_lh[MDT_LH_NEW];
1589 mdt_lock_reg_init(lh1, LCK_EX);
1590 lh2 = &info->mti_lh[MDT_LH_OLD];
1591 mdt_lock_reg_init(lh2, LCK_EX);
1593 rc = mdt_object_lock(info, o1, lh1, MDS_INODELOCK_LAYOUT |
1594 MDS_INODELOCK_XATTR);
1598 rc = mdt_object_lock(info, o2, lh2, MDS_INODELOCK_LAYOUT |
1599 MDS_INODELOCK_XATTR);
1603 rc = mo_swap_layouts(info->mti_env, mdt_object_child(o1),
1604 mdt_object_child(o2), msl->msl_flags);
1608 mdt_swap_lov_flag(o1, o2);
1611 mdt_object_unlock(info, o2, lh2, rc);
1613 mdt_object_unlock(info, o1, lh1, rc);
1615 mdt_object_put(info->mti_env, o);
1617 mdt_thread_info_fini(info);
1621 static int mdt_raw_lookup(struct mdt_thread_info *info,
1622 struct mdt_object *parent,
1623 const struct lu_name *lname,
1624 struct ldlm_reply *ldlm_rep)
1626 struct lu_fid *child_fid = &info->mti_tmp_fid1;
1630 LASSERT(!info->mti_cross_ref);
1632 /* Only got the fid of this obj by name */
1633 fid_zero(child_fid);
1634 rc = mdo_lookup(info->mti_env, mdt_object_child(info->mti_object),
1635 lname, child_fid, &info->mti_spec);
1637 struct mdt_body *repbody;
1639 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1640 repbody->mbo_fid1 = *child_fid;
1641 repbody->mbo_valid = OBD_MD_FLID;
1642 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1643 } else if (rc == -ENOENT) {
1644 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1651 * UPDATE lock should be taken against parent, and be released before exit;
1652 * child_bits lock should be taken against child, and be returned back:
1653 * (1)normal request should release the child lock;
1654 * (2)intent request will grant the lock to client.
1656 static int mdt_getattr_name_lock(struct mdt_thread_info *info,
1657 struct mdt_lock_handle *lhc,
1659 struct ldlm_reply *ldlm_rep)
1661 struct ptlrpc_request *req = mdt_info_req(info);
1662 struct mdt_body *reqbody = NULL;
1663 struct mdt_object *parent = info->mti_object;
1664 struct mdt_object *child;
1665 struct lu_fid *child_fid = &info->mti_tmp_fid1;
1666 struct lu_name *lname = NULL;
1667 struct mdt_lock_handle *lhp = NULL;
1668 struct ldlm_lock *lock;
1669 struct req_capsule *pill = info->mti_pill;
1677 is_resent = lustre_handle_is_used(&lhc->mlh_reg_lh);
1678 LASSERT(ergo(is_resent,
1679 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT));
1684 if (info->mti_cross_ref) {
1685 /* Only getattr on the child. Parent is on another node. */
1686 mdt_set_disposition(info, ldlm_rep,
1687 DISP_LOOKUP_EXECD | DISP_LOOKUP_POS);
1689 CDEBUG(D_INODE, "partial getattr_name child_fid = "DFID", "
1691 PFID(mdt_object_fid(child)), ldlm_rep);
1693 rc = mdt_check_resent_lock(info, child, lhc);
1696 } else if (rc > 0) {
1697 mdt_lock_handle_init(lhc);
1698 mdt_lock_reg_init(lhc, LCK_PR);
1701 * Object's name entry is on another MDS, it will
1702 * request PERM lock only because LOOKUP lock is owned
1703 * by the MDS where name entry resides.
1705 * TODO: it should try layout lock too. - Jinshan
1707 child_bits &= ~(MDS_INODELOCK_LOOKUP |
1708 MDS_INODELOCK_LAYOUT);
1709 child_bits |= MDS_INODELOCK_PERM;
1711 rc = mdt_object_lock(info, child, lhc, child_bits);
1716 /* Finally, we can get attr for child. */
1717 if (!mdt_object_exists(child)) {
1718 LU_OBJECT_DEBUG(D_INFO, info->mti_env,
1720 "remote object doesn't exist.");
1721 mdt_object_unlock(info, child, lhc, 1);
1725 rc = mdt_getattr_internal(info, child, 0);
1726 if (unlikely(rc != 0))
1727 mdt_object_unlock(info, child, lhc, 1);
1729 mdt_pack_secctx_in_reply(info, child);
1734 lname = &info->mti_name;
1735 mdt_name_unpack(pill, &RMF_NAME, lname, MNF_FIX_ANON);
1737 if (lu_name_is_valid(lname)) {
1738 CDEBUG(D_INODE, "getattr with lock for "DFID"/"DNAME", "
1739 "ldlm_rep = %p\n", PFID(mdt_object_fid(parent)),
1740 PNAME(lname), ldlm_rep);
1742 reqbody = req_capsule_client_get(pill, &RMF_MDT_BODY);
1743 if (unlikely(reqbody == NULL))
1744 RETURN(err_serious(-EPROTO));
1746 *child_fid = reqbody->mbo_fid2;
1748 if (unlikely(!fid_is_sane(child_fid)))
1749 RETURN(err_serious(-EINVAL));
1751 CDEBUG(D_INODE, "getattr with lock for "DFID"/"DFID", "
1753 PFID(mdt_object_fid(parent)),
1754 PFID(&reqbody->mbo_fid2), ldlm_rep);
1757 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_EXECD);
1759 if (unlikely(!mdt_object_exists(parent)) && lu_name_is_valid(lname)) {
1760 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
1762 "Parent doesn't exist!");
1766 if (mdt_object_remote(parent)) {
1767 CERROR("%s: parent "DFID" is on remote target\n",
1768 mdt_obd_name(info->mti_mdt),
1769 PFID(mdt_object_fid(parent)));
1773 if (lu_name_is_valid(lname)) {
1774 /* Always allow to lookup ".." */
1775 if (unlikely(lname->ln_namelen == 2 &&
1776 lname->ln_name[0] == '.' &&
1777 lname->ln_name[1] == '.'))
1778 info->mti_spec.sp_permitted = 1;
1780 if (info->mti_body->mbo_valid == OBD_MD_FLID) {
1781 rc = mdt_raw_lookup(info, parent, lname, ldlm_rep);
1786 /* step 1: lock parent only if parent is a directory */
1787 if (S_ISDIR(lu_object_attr(&parent->mot_obj))) {
1788 lhp = &info->mti_lh[MDT_LH_PARENT];
1789 mdt_lock_pdo_init(lhp, LCK_PR, lname);
1790 rc = mdt_object_lock(info, parent, lhp,
1791 MDS_INODELOCK_UPDATE);
1792 if (unlikely(rc != 0))
1796 /* step 2: lookup child's fid by name */
1797 fid_zero(child_fid);
1798 rc = mdo_lookup(info->mti_env, mdt_object_child(parent), lname,
1799 child_fid, &info->mti_spec);
1801 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1804 GOTO(out_parent, rc);
1807 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1810 *step 3: find the child object by fid & lock it.
1811 * regardless if it is local or remote.
1813 *Note: LU-3240 (commit 762f2114d282a98ebfa4dbbeea9298a8088ad24e)
1814 * set parent dir fid the same as child fid in getattr by fid case
1815 * we should not lu_object_find() the object again, could lead
1816 * to hung if there is a concurrent unlink destroyed the object.
1818 if (lu_fid_eq(mdt_object_fid(parent), child_fid)) {
1819 mdt_object_get(info->mti_env, parent);
1822 child = mdt_object_find(info->mti_env, info->mti_mdt,
1826 if (unlikely(IS_ERR(child)))
1827 GOTO(out_parent, rc = PTR_ERR(child));
1829 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RESEND, obd_timeout * 2);
1830 if (!mdt_object_exists(child)) {
1831 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
1833 "Object doesn't exist!");
1834 GOTO(out_child, rc = -ENOENT);
1837 rc = mdt_check_resent_lock(info, child, lhc);
1839 GOTO(out_child, rc);
1840 } else if (rc > 0) {
1841 mdt_lock_handle_init(lhc);
1842 mdt_lock_reg_init(lhc, LCK_PR);
1844 if (!(child_bits & MDS_INODELOCK_UPDATE) &&
1845 mdt_object_exists(child) && !mdt_object_remote(child)) {
1846 struct md_attr *ma = &info->mti_attr;
1849 ma->ma_need = MA_INODE;
1850 rc = mdt_attr_get_complex(info, child, ma);
1851 if (unlikely(rc != 0))
1852 GOTO(out_child, rc);
1854 /* If the file has not been changed for some time, we
1855 * return not only a LOOKUP lock, but also an UPDATE
1856 * lock and this might save us RPC on later STAT. For
1857 * directories, it also let negative dentry cache start
1858 * working for this dir. */
1859 if (ma->ma_valid & MA_INODE &&
1860 ma->ma_attr.la_valid & LA_CTIME &&
1861 info->mti_mdt->mdt_namespace->ns_ctime_age_limit +
1862 ma->ma_attr.la_ctime < ktime_get_real_seconds())
1863 child_bits |= MDS_INODELOCK_UPDATE;
1866 /* layout lock must be granted in a best-effort way
1867 * for IT operations */
1868 LASSERT(!(child_bits & MDS_INODELOCK_LAYOUT));
1869 if (S_ISREG(lu_object_attr(&child->mot_obj)) &&
1870 !mdt_object_remote(child) && ldlm_rep != NULL) {
1871 if (!OBD_FAIL_CHECK(OBD_FAIL_MDS_NO_LL_GETATTR) &&
1872 exp_connect_layout(info->mti_exp)) {
1873 /* try to grant layout lock for regular file. */
1874 try_bits = MDS_INODELOCK_LAYOUT;
1876 /* Acquire DOM lock in advance for data-on-mdt file */
1877 if (child != parent)
1878 try_bits |= MDS_INODELOCK_DOM;
1881 if (try_bits != 0) {
1882 /* try layout lock, it may fail to be granted due to
1883 * contention at LOOKUP or UPDATE */
1884 rc = mdt_object_lock_try(info, child, lhc, &child_bits,
1886 if (child_bits & MDS_INODELOCK_LAYOUT)
1889 /* Do not enqueue the UPDATE lock from MDT(cross-MDT),
1890 * client will enqueue the lock to the remote MDT */
1891 if (mdt_object_remote(child))
1892 child_bits &= ~MDS_INODELOCK_UPDATE;
1893 rc = mdt_object_lock(info, child, lhc, child_bits);
1895 if (unlikely(rc != 0))
1896 GOTO(out_child, rc);
1899 /* finally, we can get attr for child. */
1900 rc = mdt_getattr_internal(info, child, ma_need);
1901 if (unlikely(rc != 0)) {
1902 mdt_object_unlock(info, child, lhc, 1);
1903 GOTO(out_child, rc);
1906 mdt_pack_secctx_in_reply(info, child);
1908 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
1910 /* Debugging code. */
1911 LDLM_DEBUG(lock, "Returning lock to client");
1912 LASSERTF(fid_res_name_eq(mdt_object_fid(child),
1913 &lock->l_resource->lr_name),
1914 "Lock res_id: "DLDLMRES", fid: "DFID"\n",
1915 PLDLMRES(lock->l_resource),
1916 PFID(mdt_object_fid(child)));
1918 if (S_ISREG(lu_object_attr(&child->mot_obj)) &&
1919 mdt_object_exists(child) && !mdt_object_remote(child) &&
1921 mdt_object_put(info->mti_env, child);
1922 rc = mdt_pack_size2body(info, child_fid,
1924 if (rc != 0 && child_bits & MDS_INODELOCK_DOM) {
1925 /* DOM lock was taken in advance but this is
1926 * not DoM file. Drop the lock.
1928 lock_res_and_lock(lock);
1929 ldlm_inodebits_drop(lock, MDS_INODELOCK_DOM);
1930 unlock_res_and_lock(lock);
1932 LDLM_LOCK_PUT(lock);
1933 GOTO(out_parent, rc = 0);
1935 LDLM_LOCK_PUT(lock);
1940 mdt_object_put(info->mti_env, child);
1943 mdt_object_unlock(info, parent, lhp, 1);
1947 /* normal handler: should release the child lock */
1948 static int mdt_getattr_name(struct tgt_session_info *tsi)
1950 struct mdt_thread_info *info = tsi2mdt_info(tsi);
1951 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_CHILD];
1952 struct mdt_body *reqbody;
1953 struct mdt_body *repbody;
1957 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
1958 LASSERT(reqbody != NULL);
1959 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1960 LASSERT(repbody != NULL);
1962 info->mti_cross_ref = !!(reqbody->mbo_valid & OBD_MD_FLCROSSREF);
1963 repbody->mbo_eadatasize = 0;
1964 repbody->mbo_aclsize = 0;
1966 rc = mdt_init_ucred_intent_getattr(info, reqbody);
1968 GOTO(out_shrink, rc);
1970 rc = mdt_getattr_name_lock(info, lhc, MDS_INODELOCK_UPDATE, NULL);
1971 if (lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1972 ldlm_lock_decref(&lhc->mlh_reg_lh, lhc->mlh_reg_mode);
1973 lhc->mlh_reg_lh.cookie = 0;
1975 mdt_exit_ucred(info);
1978 mdt_client_compatibility(info);
1979 rc2 = mdt_fix_reply(info);
1982 mdt_thread_info_fini(info);
1986 static int mdt_rmfid_unlink(struct mdt_thread_info *info,
1987 const struct lu_fid *pfid,
1988 const struct lu_name *name,
1989 struct mdt_object *obj, s64 ctime)
1991 struct lu_fid *child_fid = &info->mti_tmp_fid1;
1992 struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
1993 struct mdt_device *mdt = info->mti_mdt;
1994 struct md_attr *ma = &info->mti_attr;
1995 struct mdt_lock_handle *parent_lh;
1996 struct mdt_lock_handle *child_lh;
1997 struct mdt_object *pobj;
1998 bool cos_incompat = false;
2002 pobj = mdt_object_find(info->mti_env, mdt, pfid);
2004 GOTO(out, rc = PTR_ERR(pobj));
2006 parent_lh = &info->mti_lh[MDT_LH_PARENT];
2007 mdt_lock_pdo_init(parent_lh, LCK_PW, name);
2008 rc = mdt_object_lock(info, pobj, parent_lh, MDS_INODELOCK_UPDATE);
2010 GOTO(put_parent, rc);
2012 if (mdt_object_remote(pobj))
2013 cos_incompat = true;
2015 rc = mdo_lookup(info->mti_env, mdt_object_child(pobj),
2016 name, child_fid, &info->mti_spec);
2018 GOTO(unlock_parent, rc);
2020 if (!lu_fid_eq(child_fid, mdt_object_fid(obj)))
2021 GOTO(unlock_parent, rc = -EREMCHG);
2023 child_lh = &info->mti_lh[MDT_LH_CHILD];
2024 mdt_lock_reg_init(child_lh, LCK_EX);
2025 rc = mdt_reint_striped_lock(info, obj, child_lh,
2026 MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE,
2027 einfo, cos_incompat);
2029 GOTO(unlock_parent, rc);
2031 if (atomic_read(&obj->mot_open_count)) {
2032 CDEBUG(D_OTHER, "object "DFID" open, skip\n",
2033 PFID(mdt_object_fid(obj)));
2034 GOTO(unlock_child, rc = -EBUSY);
2038 ma->ma_valid = MA_INODE;
2039 ma->ma_attr.la_valid = LA_CTIME;
2040 ma->ma_attr.la_ctime = ctime;
2042 mutex_lock(&obj->mot_lov_mutex);
2044 rc = mdo_unlink(info->mti_env, mdt_object_child(pobj),
2045 mdt_object_child(obj), name, ma, 0);
2047 mutex_unlock(&obj->mot_lov_mutex);
2050 mdt_reint_striped_unlock(info, obj, child_lh, einfo, 1);
2052 mdt_object_unlock(info, pobj, parent_lh, 1);
2054 mdt_object_put(info->mti_env, pobj);
2059 static int mdt_rmfid_check_permission(struct mdt_thread_info *info,
2060 struct mdt_object *obj)
2062 struct lu_ucred *uc = lu_ucred(info->mti_env);
2063 struct md_attr *ma = &info->mti_attr;
2064 struct lu_attr *la = &ma->ma_attr;
2068 ma->ma_need = MA_INODE;
2069 rc = mo_attr_get(info->mti_env, mdt_object_child(obj), ma);
2073 if (la->la_flags & LUSTRE_IMMUTABLE_FL)
2076 if (md_capable(uc, CFS_CAP_DAC_OVERRIDE))
2078 if (uc->uc_fsuid == la->la_uid) {
2079 if ((la->la_mode & S_IWUSR) == 0)
2081 } else if (uc->uc_fsgid == la->la_gid) {
2082 if ((la->la_mode & S_IWGRP) == 0)
2084 } else if ((la->la_mode & S_IWOTH) == 0) {
2092 static int mdt_rmfid_one(struct mdt_thread_info *info, struct lu_fid *fid,
2095 struct mdt_device *mdt = info->mti_mdt;
2096 struct mdt_object *obj = NULL;
2097 struct linkea_data ldata = { NULL };
2098 struct lu_buf *buf = &info->mti_big_buf;
2099 struct lu_name *name = &info->mti_name;
2100 struct lu_fid *pfid = &info->mti_tmp_fid1;
2101 struct link_ea_header *leh;
2102 struct link_ea_entry *lee;
2103 int reclen, count, rc = 0;
2106 if (!fid_is_sane(fid))
2107 GOTO(out, rc = -EINVAL);
2109 if (!fid_is_namespace_visible(fid))
2110 GOTO(out, rc = -EINVAL);
2112 obj = mdt_object_find(info->mti_env, mdt, fid);
2114 GOTO(out, rc = PTR_ERR(obj));
2116 if (mdt_object_remote(obj))
2117 GOTO(out, rc = -EREMOTE);
2118 if (!mdt_object_exists(obj) || lu_object_is_dying(&obj->mot_header))
2119 GOTO(out, rc = -ENOENT);
2121 rc = mdt_rmfid_check_permission(info, obj);
2126 buf = lu_buf_check_and_alloc(buf, PATH_MAX);
2128 GOTO(out, rc = -ENOMEM);
2131 rc = mdt_links_read(info, obj, &ldata);
2136 lee = (struct link_ea_entry *)(leh + 1);
2137 for (count = 0; count < leh->leh_reccount; count++) {
2138 /* remove every hardlink */
2139 linkea_entry_unpack(lee, &reclen, name, pfid);
2140 lee = (struct link_ea_entry *) ((char *)lee + reclen);
2141 rc = mdt_rmfid_unlink(info, pfid, name, obj, ctime);
2147 if (obj && !IS_ERR(obj))
2148 mdt_object_put(info->mti_env, obj);
2149 if (info->mti_big_buf.lb_buf)
2150 lu_buf_free(&info->mti_big_buf);
2155 static int mdt_rmfid(struct tgt_session_info *tsi)
2157 struct mdt_thread_info *mti = tsi2mdt_info(tsi);
2158 struct mdt_body *reqbody;
2159 struct lu_fid *fids, *rfids;
2165 reqbody = req_capsule_client_get(tsi->tsi_pill, &RMF_MDT_BODY);
2166 if (reqbody == NULL)
2168 bufsize = req_capsule_get_size(tsi->tsi_pill, &RMF_FID_ARRAY,
2170 nr = bufsize / sizeof(struct lu_fid);
2171 if (nr * sizeof(struct lu_fid) != bufsize)
2173 req_capsule_set_size(tsi->tsi_pill, &RMF_RCS,
2174 RCL_SERVER, nr * sizeof(__u32));
2175 req_capsule_set_size(tsi->tsi_pill, &RMF_FID_ARRAY,
2176 RCL_SERVER, nr * sizeof(struct lu_fid));
2177 rc = req_capsule_server_pack(tsi->tsi_pill);
2179 GOTO(out, rc = err_serious(rc));
2180 fids = req_capsule_client_get(tsi->tsi_pill, &RMF_FID_ARRAY);
2183 rcs = req_capsule_server_get(tsi->tsi_pill, &RMF_RCS);
2185 rfids = req_capsule_server_get(tsi->tsi_pill, &RMF_FID_ARRAY);
2188 mdt_init_ucred(mti, reqbody);
2189 for (i = 0; i < nr; i++) {
2191 rcs[i] = mdt_rmfid_one(mti, fids + i, reqbody->mbo_ctime);
2193 mdt_exit_ucred(mti);
2199 static int mdt_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2200 void *karg, void __user *uarg);
2202 static int mdt_set_info(struct tgt_session_info *tsi)
2204 struct ptlrpc_request *req = tgt_ses_req(tsi);
2207 int keylen, vallen, rc = 0;
2211 key = req_capsule_client_get(tsi->tsi_pill, &RMF_SETINFO_KEY);
2213 DEBUG_REQ(D_HA, req, "no set_info key");
2214 RETURN(err_serious(-EFAULT));
2217 keylen = req_capsule_get_size(tsi->tsi_pill, &RMF_SETINFO_KEY,
2220 val = req_capsule_client_get(tsi->tsi_pill, &RMF_SETINFO_VAL);
2222 DEBUG_REQ(D_HA, req, "no set_info val");
2223 RETURN(err_serious(-EFAULT));
2226 vallen = req_capsule_get_size(tsi->tsi_pill, &RMF_SETINFO_VAL,
2229 /* Swab any part of val you need to here */
2230 if (KEY_IS(KEY_READ_ONLY)) {
2231 spin_lock(&req->rq_export->exp_lock);
2233 *exp_connect_flags_ptr(req->rq_export) |=
2236 *exp_connect_flags_ptr(req->rq_export) &=
2237 ~OBD_CONNECT_RDONLY;
2238 spin_unlock(&req->rq_export->exp_lock);
2239 } else if (KEY_IS(KEY_CHANGELOG_CLEAR)) {
2240 struct changelog_setinfo *cs = val;
2242 if (vallen != sizeof(*cs)) {
2243 CERROR("%s: bad changelog_clear setinfo size %d\n",
2244 tgt_name(tsi->tsi_tgt), vallen);
2247 if (ptlrpc_req_need_swab(req)) {
2248 __swab64s(&cs->cs_recno);
2249 __swab32s(&cs->cs_id);
2252 rc = mdt_iocontrol(OBD_IOC_CHANGELOG_CLEAR, req->rq_export,
2254 } else if (KEY_IS(KEY_EVICT_BY_NID)) {
2256 obd_export_evict_by_nid(req->rq_export->exp_obd, val);
2263 static int mdt_readpage(struct tgt_session_info *tsi)
2265 struct mdt_thread_info *info = mdt_th_info(tsi->tsi_env);
2266 struct mdt_object *object = mdt_obj(tsi->tsi_corpus);
2267 struct lu_rdpg *rdpg = &info->mti_u.rdpg.mti_rdpg;
2268 const struct mdt_body *reqbody = tsi->tsi_mdt_body;
2269 struct mdt_body *repbody;
2275 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK))
2276 RETURN(err_serious(-ENOMEM));
2278 repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_MDT_BODY);
2279 if (repbody == NULL || reqbody == NULL)
2280 RETURN(err_serious(-EFAULT));
2283 * prepare @rdpg before calling lower layers and transfer itself. Here
2284 * reqbody->size contains offset of where to start to read and
2285 * reqbody->nlink contains number bytes to read.
2287 rdpg->rp_hash = reqbody->mbo_size;
2288 if (rdpg->rp_hash != reqbody->mbo_size) {
2289 CERROR("Invalid hash: %#llx != %#llx\n",
2290 rdpg->rp_hash, reqbody->mbo_size);
2294 rdpg->rp_attrs = reqbody->mbo_mode;
2295 if (exp_connect_flags(tsi->tsi_exp) & OBD_CONNECT_64BITHASH)
2296 rdpg->rp_attrs |= LUDA_64BITHASH;
2297 rdpg->rp_count = min_t(unsigned int, reqbody->mbo_nlink,
2298 exp_max_brw_size(tsi->tsi_exp));
2299 rdpg->rp_npages = (rdpg->rp_count + PAGE_SIZE - 1) >>
2301 OBD_ALLOC(rdpg->rp_pages, rdpg->rp_npages * sizeof rdpg->rp_pages[0]);
2302 if (rdpg->rp_pages == NULL)
2305 for (i = 0; i < rdpg->rp_npages; ++i) {
2306 rdpg->rp_pages[i] = alloc_page(GFP_NOFS);
2307 if (rdpg->rp_pages[i] == NULL)
2308 GOTO(free_rdpg, rc = -ENOMEM);
2311 /* call lower layers to fill allocated pages with directory data */
2312 rc = mo_readpage(tsi->tsi_env, mdt_object_child(object), rdpg);
2314 GOTO(free_rdpg, rc);
2316 /* send pages to client */
2317 rc = tgt_sendpage(tsi, rdpg, rc);
2322 for (i = 0; i < rdpg->rp_npages; i++)
2323 if (rdpg->rp_pages[i] != NULL)
2324 __free_page(rdpg->rp_pages[i]);
2325 OBD_FREE(rdpg->rp_pages, rdpg->rp_npages * sizeof rdpg->rp_pages[0]);
2327 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE))
2333 static int mdt_fix_attr_ucred(struct mdt_thread_info *info, __u32 op)
2335 struct lu_ucred *uc = mdt_ucred_check(info);
2336 struct lu_attr *attr = &info->mti_attr.ma_attr;
2341 if (op != REINT_SETATTR) {
2342 if ((attr->la_valid & LA_UID) && (attr->la_uid != -1))
2343 attr->la_uid = uc->uc_fsuid;
2344 /* for S_ISGID, inherit gid from his parent, such work will be
2345 * done in cmm/mdd layer, here set all cases as uc->uc_fsgid. */
2346 if ((attr->la_valid & LA_GID) && (attr->la_gid != -1))
2347 attr->la_gid = uc->uc_fsgid;
2353 static inline bool mdt_is_readonly_open(struct mdt_thread_info *info, __u32 op)
2355 return op == REINT_OPEN &&
2356 !(info->mti_spec.sp_cr_flags & (MDS_FMODE_WRITE | MDS_OPEN_CREAT));
2359 static void mdt_preset_secctx_size(struct mdt_thread_info *info)
2361 struct req_capsule *pill = info->mti_pill;
2363 if (req_capsule_has_field(pill, &RMF_FILE_SECCTX,
2365 req_capsule_has_field(pill, &RMF_FILE_SECCTX_NAME,
2367 if (req_capsule_get_size(pill, &RMF_FILE_SECCTX_NAME,
2369 /* pre-set size in server part with max size */
2370 req_capsule_set_size(pill, &RMF_FILE_SECCTX,
2372 info->mti_mdt->mdt_max_ea_size);
2374 req_capsule_set_size(pill, &RMF_FILE_SECCTX,
2381 static int mdt_reint_internal(struct mdt_thread_info *info,
2382 struct mdt_lock_handle *lhc,
2385 struct req_capsule *pill = info->mti_pill;
2386 struct mdt_body *repbody;
2391 rc = mdt_reint_unpack(info, op);
2393 CERROR("Can't unpack reint, rc %d\n", rc);
2394 RETURN(err_serious(rc));
2398 /* check if the file system is set to readonly. O_RDONLY open
2399 * is still allowed even the file system is set to readonly mode */
2400 if (mdt_rdonly(info->mti_exp) && !mdt_is_readonly_open(info, op))
2401 RETURN(err_serious(-EROFS));
2403 /* for replay (no_create) lmm is not needed, client has it already */
2404 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
2405 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER,
2408 /* llog cookies are always 0, the field is kept for compatibility */
2409 if (req_capsule_has_field(pill, &RMF_LOGCOOKIES, RCL_SERVER))
2410 req_capsule_set_size(pill, &RMF_LOGCOOKIES, RCL_SERVER, 0);
2412 /* Set ACL reply buffer size as LUSTRE_POSIX_ACL_MAX_SIZE_OLD
2413 * by default. If the target object has more ACL entries, then
2414 * enlarge the buffer when necessary. */
2415 if (req_capsule_has_field(pill, &RMF_ACL, RCL_SERVER))
2416 req_capsule_set_size(pill, &RMF_ACL, RCL_SERVER,
2417 LUSTRE_POSIX_ACL_MAX_SIZE_OLD);
2419 mdt_preset_secctx_size(info);
2421 rc = req_capsule_server_pack(pill);
2423 CERROR("Can't pack response, rc %d\n", rc);
2424 RETURN(err_serious(rc));
2427 if (req_capsule_has_field(pill, &RMF_MDT_BODY, RCL_SERVER)) {
2428 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
2430 repbody->mbo_eadatasize = 0;
2431 repbody->mbo_aclsize = 0;
2434 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_REINT_DELAY, 10);
2436 /* for replay no cookkie / lmm need, because client have this already */
2437 if (info->mti_spec.no_create)
2438 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
2439 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER, 0);
2441 rc = mdt_init_ucred_reint(info);
2443 GOTO(out_shrink, rc);
2445 rc = mdt_fix_attr_ucred(info, op);
2447 GOTO(out_ucred, rc = err_serious(rc));
2449 rc = mdt_check_resent(info, mdt_reconstruct, lhc);
2451 GOTO(out_ucred, rc);
2452 } else if (rc == 1) {
2453 DEBUG_REQ(D_INODE, mdt_info_req(info), "resent opt");
2454 rc = lustre_msg_get_status(mdt_info_req(info)->rq_repmsg);
2455 GOTO(out_ucred, rc);
2457 rc = mdt_reint_rec(info, lhc);
2460 mdt_exit_ucred(info);
2462 mdt_client_compatibility(info);
2464 rc2 = mdt_fix_reply(info);
2469 * Data-on-MDT optimization - read data along with OPEN and return it
2470 * in reply. Do that only if we have both DOM and LAYOUT locks.
2472 if (rc == 0 && op == REINT_OPEN && !req_is_replay(pill->rc_req) &&
2473 info->mti_attr.ma_lmm != NULL &&
2474 mdt_lmm_dom_entry(info->mti_attr.ma_lmm) == LMM_DOM_ONLY) {
2475 rc = mdt_dom_read_on_open(info, info->mti_mdt,
2482 static long mdt_reint_opcode(struct ptlrpc_request *req,
2483 const struct req_format **fmt)
2485 struct mdt_device *mdt;
2486 struct mdt_rec_reint *rec;
2489 rec = req_capsule_client_get(&req->rq_pill, &RMF_REC_REINT);
2491 opc = rec->rr_opcode;
2492 DEBUG_REQ(D_INODE, req, "reint opt = %ld", opc);
2493 if (opc < REINT_MAX && fmt[opc] != NULL)
2494 req_capsule_extend(&req->rq_pill, fmt[opc]);
2496 mdt = mdt_exp2dev(req->rq_export);
2497 CERROR("%s: Unsupported opcode '%ld' from client '%s':"
2498 " rc = %d\n", req->rq_export->exp_obd->obd_name,
2499 opc, mdt->mdt_ldlm_client->cli_name, -EFAULT);
2500 opc = err_serious(-EFAULT);
2503 opc = err_serious(-EFAULT);
2508 static int mdt_reint(struct tgt_session_info *tsi)
2512 static const struct req_format *reint_fmts[REINT_MAX] = {
2513 [REINT_SETATTR] = &RQF_MDS_REINT_SETATTR,
2514 [REINT_CREATE] = &RQF_MDS_REINT_CREATE,
2515 [REINT_LINK] = &RQF_MDS_REINT_LINK,
2516 [REINT_UNLINK] = &RQF_MDS_REINT_UNLINK,
2517 [REINT_RENAME] = &RQF_MDS_REINT_RENAME,
2518 [REINT_OPEN] = &RQF_MDS_REINT_OPEN,
2519 [REINT_SETXATTR] = &RQF_MDS_REINT_SETXATTR,
2520 [REINT_RMENTRY] = &RQF_MDS_REINT_UNLINK,
2521 [REINT_MIGRATE] = &RQF_MDS_REINT_MIGRATE,
2522 [REINT_RESYNC] = &RQF_MDS_REINT_RESYNC,
2527 opc = mdt_reint_opcode(tgt_ses_req(tsi), reint_fmts);
2529 struct mdt_thread_info *info = tsi2mdt_info(tsi);
2531 * No lock possible here from client to pass it to reint code
2534 rc = mdt_reint_internal(info, NULL, opc);
2535 mdt_thread_info_fini(info);
2540 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_REINT_NET_REP;
2544 /* this should sync the whole device */
2545 int mdt_device_sync(const struct lu_env *env, struct mdt_device *mdt)
2547 struct dt_device *dt = mdt->mdt_bottom;
2551 rc = dt->dd_ops->dt_sync(env, dt);
2555 /* this should sync this object */
2556 static int mdt_object_sync(const struct lu_env *env, struct obd_export *exp,
2557 struct mdt_object *mo)
2563 if (!mdt_object_exists(mo)) {
2564 CWARN("%s: non existing object "DFID": rc = %d\n",
2565 exp->exp_obd->obd_name, PFID(mdt_object_fid(mo)),
2570 if (S_ISREG(lu_object_attr(&mo->mot_obj))) {
2571 struct lu_target *tgt = tgt_ses_info(env)->tsi_tgt;
2572 dt_obj_version_t version;
2574 version = dt_version_get(env, mdt_obj2dt(mo));
2575 if (version > tgt->lut_obd->obd_last_committed)
2576 rc = mo_object_sync(env, mdt_object_child(mo));
2578 rc = mo_object_sync(env, mdt_object_child(mo));
2584 static int mdt_sync(struct tgt_session_info *tsi)
2586 struct ptlrpc_request *req = tgt_ses_req(tsi);
2587 struct req_capsule *pill = tsi->tsi_pill;
2588 struct mdt_body *body;
2593 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SYNC_PACK))
2594 RETURN(err_serious(-ENOMEM));
2596 if (fid_seq(&tsi->tsi_mdt_body->mbo_fid1) == 0) {
2597 rc = mdt_device_sync(tsi->tsi_env, mdt_exp2dev(tsi->tsi_exp));
2599 struct mdt_thread_info *info = tsi2mdt_info(tsi);
2601 if (unlikely(info->mti_object == NULL))
2604 /* sync an object */
2605 rc = mdt_object_sync(tsi->tsi_env, tsi->tsi_exp,
2608 const struct lu_fid *fid;
2609 struct lu_attr *la = &info->mti_attr.ma_attr;
2611 info->mti_attr.ma_need = MA_INODE;
2612 info->mti_attr.ma_valid = 0;
2613 rc = mdt_attr_get_complex(info, info->mti_object,
2616 body = req_capsule_server_get(pill,
2618 fid = mdt_object_fid(info->mti_object);
2619 mdt_pack_attr2body(info, body, la, fid);
2622 mdt_thread_info_fini(info);
2625 mdt_counter_incr(req, LPROC_MDT_SYNC);
2630 static int mdt_data_sync(struct tgt_session_info *tsi)
2632 struct mdt_thread_info *info;
2633 struct mdt_device *mdt = mdt_exp2dev(tsi->tsi_exp);
2634 struct ost_body *body = tsi->tsi_ost_body;
2635 struct ost_body *repbody;
2636 struct mdt_object *mo = NULL;
2642 repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_OST_BODY);
2644 /* if no fid is specified then do nothing,
2645 * device sync is done via MDS_SYNC */
2646 if (fid_is_zero(&tsi->tsi_fid))
2649 mo = mdt_object_find(tsi->tsi_env, mdt, &tsi->tsi_fid);
2651 RETURN(PTR_ERR(mo));
2653 rc = mdt_object_sync(tsi->tsi_env, tsi->tsi_exp, mo);
2657 repbody->oa.o_oi = body->oa.o_oi;
2658 repbody->oa.o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
2660 info = tsi2mdt_info(tsi);
2661 ma = &info->mti_attr;
2662 ma->ma_need = MA_INODE;
2664 rc = mdt_attr_get_complex(info, mo, ma);
2666 obdo_from_la(&repbody->oa, &ma->ma_attr, VALID_FLAGS);
2669 mdt_thread_info_fini(info);
2674 mdt_object_put(tsi->tsi_env, mo);
2679 * Handle quota control requests to consult current usage/limit, but also
2680 * to configure quota enforcement
2682 static int mdt_quotactl(struct tgt_session_info *tsi)
2684 struct obd_export *exp = tsi->tsi_exp;
2685 struct req_capsule *pill = tsi->tsi_pill;
2686 struct obd_quotactl *oqctl, *repoqc;
2688 struct mdt_device *mdt = mdt_exp2dev(exp);
2689 struct lu_device *qmt = mdt->mdt_qmt_dev;
2690 struct lu_nodemap *nodemap;
2693 oqctl = req_capsule_client_get(pill, &RMF_OBD_QUOTACTL);
2695 RETURN(err_serious(-EPROTO));
2697 rc = req_capsule_server_pack(pill);
2699 RETURN(err_serious(rc));
2701 nodemap = nodemap_get_from_exp(exp);
2702 if (IS_ERR(nodemap))
2703 RETURN(PTR_ERR(nodemap));
2705 switch (oqctl->qc_cmd) {
2706 /* master quotactl */
2709 case LUSTRE_Q_SETDEFAULT:
2710 if (!nodemap_can_setquota(nodemap))
2711 GOTO(out_nodemap, rc = -EPERM);
2715 case LUSTRE_Q_GETDEFAULT:
2717 GOTO(out_nodemap, rc = -EOPNOTSUPP);
2718 /* slave quotactl */
2724 CERROR("Unsupported quotactl command: %d\n", oqctl->qc_cmd);
2725 GOTO(out_nodemap, rc = -EFAULT);
2729 switch (oqctl->qc_type) {
2731 id = nodemap_map_id(nodemap, NODEMAP_UID,
2732 NODEMAP_CLIENT_TO_FS, id);
2735 id = nodemap_map_id(nodemap, NODEMAP_GID,
2736 NODEMAP_CLIENT_TO_FS, id);
2739 /* todo: check/map project id */
2743 GOTO(out_nodemap, rc = -EOPNOTSUPP);
2745 repoqc = req_capsule_server_get(pill, &RMF_OBD_QUOTACTL);
2747 GOTO(out_nodemap, rc = err_serious(-EFAULT));
2749 if (oqctl->qc_cmd == Q_SETINFO || oqctl->qc_cmd == Q_SETQUOTA)
2750 barrier_exit(tsi->tsi_tgt->lut_bottom);
2752 if (oqctl->qc_id != id)
2753 swap(oqctl->qc_id, id);
2755 if (oqctl->qc_cmd == Q_SETINFO || oqctl->qc_cmd == Q_SETQUOTA) {
2756 if (unlikely(!barrier_entry(tsi->tsi_tgt->lut_bottom)))
2757 RETURN(-EINPROGRESS);
2760 switch (oqctl->qc_cmd) {
2766 case LUSTRE_Q_SETDEFAULT:
2767 case LUSTRE_Q_GETDEFAULT:
2768 /* forward quotactl request to QMT */
2769 rc = qmt_hdls.qmth_quotactl(tsi->tsi_env, qmt, oqctl);
2774 /* slave quotactl */
2775 rc = lquotactl_slv(tsi->tsi_env, tsi->tsi_tgt->lut_bottom,
2780 CERROR("Unsupported quotactl command: %d\n", oqctl->qc_cmd);
2781 GOTO(out_nodemap, rc = -EFAULT);
2784 if (oqctl->qc_id != id)
2785 swap(oqctl->qc_id, id);
2792 nodemap_putref(nodemap);
2797 /** clone llog ctxt from child (mdd)
2798 * This allows remote llog (replicator) access.
2799 * We can either pass all llog RPCs (eg mdt_llog_create) on to child where the
2800 * context was originally set up, or we can handle them directly.
2801 * I choose the latter, but that means I need any llog
2802 * contexts set up by child to be accessable by the mdt. So we clone the
2803 * context into our context list here.
2805 static int mdt_llog_ctxt_clone(const struct lu_env *env, struct mdt_device *mdt,
2808 struct md_device *next = mdt->mdt_child;
2809 struct llog_ctxt *ctxt;
2812 if (!llog_ctxt_null(mdt2obd_dev(mdt), idx))
2815 rc = next->md_ops->mdo_llog_ctxt_get(env, next, idx, (void **)&ctxt);
2816 if (rc || ctxt == NULL) {
2820 rc = llog_group_set_ctxt(&mdt2obd_dev(mdt)->obd_olg, ctxt, idx);
2822 CERROR("Can't set mdt ctxt %d\n", rc);
2827 static int mdt_llog_ctxt_unclone(const struct lu_env *env,
2828 struct mdt_device *mdt, int idx)
2830 struct llog_ctxt *ctxt;
2832 ctxt = llog_get_context(mdt2obd_dev(mdt), idx);
2835 /* Put once for the get we just did, and once for the clone */
2836 llog_ctxt_put(ctxt);
2837 llog_ctxt_put(ctxt);
2842 * sec context handlers
2844 static int mdt_sec_ctx_handle(struct tgt_session_info *tsi)
2846 CFS_FAIL_TIMEOUT(OBD_FAIL_SEC_CTX_HDL_PAUSE, cfs_fail_val);
2852 * quota request handlers
2854 static int mdt_quota_dqacq(struct tgt_session_info *tsi)
2856 struct mdt_device *mdt = mdt_exp2dev(tsi->tsi_exp);
2857 struct lu_device *qmt = mdt->mdt_qmt_dev;
2862 RETURN(err_serious(-EOPNOTSUPP));
2864 rc = qmt_hdls.qmth_dqacq(tsi->tsi_env, qmt, tgt_ses_req(tsi));
2868 struct mdt_object *mdt_object_new(const struct lu_env *env,
2869 struct mdt_device *d,
2870 const struct lu_fid *f)
2872 struct lu_object_conf conf = { .loc_flags = LOC_F_NEW };
2873 struct lu_object *o;
2874 struct mdt_object *m;
2877 CDEBUG(D_INFO, "Allocate object for "DFID"\n", PFID(f));
2878 o = lu_object_find(env, &d->mdt_lu_dev, f, &conf);
2879 if (unlikely(IS_ERR(o)))
2880 m = (struct mdt_object *)o;
2886 struct mdt_object *mdt_object_find(const struct lu_env *env,
2887 struct mdt_device *d,
2888 const struct lu_fid *f)
2890 struct lu_object *o;
2891 struct mdt_object *m;
2894 CDEBUG(D_INFO, "Find object for "DFID"\n", PFID(f));
2895 o = lu_object_find(env, &d->mdt_lu_dev, f, NULL);
2896 if (unlikely(IS_ERR(o)))
2897 m = (struct mdt_object *)o;
2905 * Asyncronous commit for mdt device.
2907 * Pass asynchonous commit call down the MDS stack.
2909 * \param env environment
2910 * \param mdt the mdt device
2912 static void mdt_device_commit_async(const struct lu_env *env,
2913 struct mdt_device *mdt)
2915 struct dt_device *dt = mdt->mdt_bottom;
2919 rc = dt->dd_ops->dt_commit_async(env, dt);
2920 if (unlikely(rc != 0))
2921 CWARN("%s: async commit start failed: rc = %d\n",
2922 mdt_obd_name(mdt), rc);
2923 atomic_inc(&mdt->mdt_async_commit_count);
2928 * Mark the lock as "synchonous".
2930 * Mark the lock to deffer transaction commit to the unlock time.
2932 * \param lock the lock to mark as "synchonous"
2934 * \see mdt_is_lock_sync
2935 * \see mdt_save_lock
2937 static inline void mdt_set_lock_sync(struct ldlm_lock *lock)
2939 lock->l_ast_data = (void*)1;
2943 * Check whehter the lock "synchonous" or not.
2945 * \param lock the lock to check
2946 * \retval 1 the lock is "synchonous"
2947 * \retval 0 the lock isn't "synchronous"
2949 * \see mdt_set_lock_sync
2950 * \see mdt_save_lock
2952 static inline int mdt_is_lock_sync(struct ldlm_lock *lock)
2954 return lock->l_ast_data != NULL;
2958 * Blocking AST for mdt locks.
2960 * Starts transaction commit if in case of COS lock conflict or
2961 * deffers such a commit to the mdt_save_lock.
2963 * \param lock the lock which blocks a request or cancelling lock
2964 * \param desc unused
2965 * \param data unused
2966 * \param flag indicates whether this cancelling or blocking callback
2968 * \see ldlm_blocking_ast_nocheck
2970 int mdt_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
2971 void *data, int flag)
2973 struct obd_device *obd = ldlm_lock_to_ns(lock)->ns_obd;
2974 struct mdt_device *mdt = mdt_dev(obd->obd_lu_dev);
2975 struct ldlm_cb_set_arg *arg = data;
2976 bool commit_async = false;
2980 if (flag == LDLM_CB_CANCELING)
2983 lock_res_and_lock(lock);
2984 if (lock->l_blocking_ast != mdt_blocking_ast) {
2985 unlock_res_and_lock(lock);
2989 /* A blocking ast may be sent from ldlm_lock_decref_internal
2990 * when the last reference to a local lock was released and
2991 * during blocking event from ldlm_work_bl_ast_lock().
2992 * The 'data' parameter is l_ast_data in the first case and
2993 * callback arguments in the second one. Distinguish them by that.
2995 if (!data || data == lock->l_ast_data || !arg->bl_desc)
2996 goto skip_cos_checks;
2998 if (lock->l_req_mode & (LCK_PW | LCK_EX)) {
2999 if (mdt_cos_is_enabled(mdt)) {
3000 if (!arg->bl_desc->bl_same_client)
3001 mdt_set_lock_sync(lock);
3002 } else if (mdt_slc_is_enabled(mdt) &&
3003 arg->bl_desc->bl_cos_incompat) {
3004 mdt_set_lock_sync(lock);
3006 * we may do extra commit here, but there is a small
3007 * window to miss a commit: lock was unlocked (saved),
3008 * then a conflict lock queued and we come here, but
3009 * REP-ACK not received, so lock was not converted to
3011 * Fortunately this window is quite small, so the
3012 * extra commit should be rare (not to say distributed
3013 * operation is rare too).
3015 commit_async = true;
3017 } else if (lock->l_req_mode == LCK_COS) {
3018 commit_async = true;
3022 rc = ldlm_blocking_ast_nocheck(lock);
3027 rc = lu_env_init(&env, LCT_LOCAL);
3028 if (unlikely(rc != 0))
3029 CWARN("%s: lu_env initialization failed, cannot "
3030 "start asynchronous commit: rc = %d\n",
3033 mdt_device_commit_async(&env, mdt);
3040 * Blocking AST for cross-MDT lock
3042 * Discard lock from uncommitted_slc_locks and cancel it.
3044 * \param lock the lock which blocks a request or cancelling lock
3045 * \param desc unused
3046 * \param data unused
3047 * \param flag indicates whether this cancelling or blocking callback
3048 * \retval 0 on success
3049 * \retval negative number on error
3051 int mdt_remote_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
3052 void *data, int flag)
3058 case LDLM_CB_BLOCKING: {
3059 struct lustre_handle lockh;
3061 ldlm_lock2handle(lock, &lockh);
3062 rc = ldlm_cli_cancel(&lockh,
3063 ldlm_is_atomic_cb(lock) ? 0 : LCF_ASYNC);
3065 CDEBUG(D_INODE, "ldlm_cli_cancel: %d\n", rc);
3070 case LDLM_CB_CANCELING: {
3071 struct obd_device *obd = ldlm_lock_to_ns(lock)->ns_obd;
3072 struct mdt_device *mdt =
3073 mdt_dev(obd->obd_lu_dev->ld_site->ls_top_dev);
3075 LDLM_DEBUG(lock, "Revoke remote lock\n");
3077 /* discard slc lock here so that it can be cleaned anytime,
3078 * especially for cleanup_resource() */
3079 tgt_discard_slc_lock(&mdt->mdt_lut, lock);
3081 /* once we cache lock, l_ast_data is set to mdt_object */
3082 if (lock->l_ast_data != NULL) {
3083 struct mdt_object *mo = lock->l_ast_data;
3086 rc = lu_env_init(&env, LCT_MD_THREAD);
3087 if (unlikely(rc != 0)) {
3088 CWARN("%s: lu_env initialization failed, object"
3089 "%p "DFID" is leaked!\n",
3091 PFID(mdt_object_fid(mo)));
3095 if (lock->l_policy_data.l_inodebits.bits &
3096 (MDS_INODELOCK_XATTR | MDS_INODELOCK_UPDATE)) {
3097 rc = mo_invalidate(&env, mdt_object_child(mo));
3098 mo->mot_cache_attr = 0;
3100 mdt_object_put(&env, mo);
3112 int mdt_check_resent_lock(struct mdt_thread_info *info,
3113 struct mdt_object *mo,
3114 struct mdt_lock_handle *lhc)
3116 /* the lock might already be gotten in ldlm_handle_enqueue() */
3117 if (unlikely(lustre_handle_is_used(&lhc->mlh_reg_lh))) {
3118 struct ptlrpc_request *req = mdt_info_req(info);
3119 struct ldlm_lock *lock;
3121 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
3122 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
3124 /* Lock is pinned by ldlm_handle_enqueue0() as it is
3125 * a resend case, however, it could be already destroyed
3126 * due to client eviction or a raced cancel RPC. */
3127 LDLM_DEBUG_NOLOCK("Invalid lock handle %#llx",
3128 lhc->mlh_reg_lh.cookie);
3132 if (!fid_res_name_eq(mdt_object_fid(mo),
3133 &lock->l_resource->lr_name)) {
3134 CWARN("%s: Although resent, but still not "
3135 "get child lock:"DFID"\n",
3136 info->mti_exp->exp_obd->obd_name,
3137 PFID(mdt_object_fid(mo)));
3138 LDLM_LOCK_PUT(lock);
3141 LDLM_LOCK_PUT(lock);
3147 static void mdt_remote_object_lock_created_cb(struct ldlm_lock *lock)
3149 mdt_object_get(NULL, lock->l_ast_data);
3152 int mdt_remote_object_lock_try(struct mdt_thread_info *mti,
3153 struct mdt_object *o, const struct lu_fid *fid,
3154 struct lustre_handle *lh, enum ldlm_mode mode,
3155 __u64 *ibits, __u64 trybits, bool cache)
3157 struct ldlm_enqueue_info *einfo = &mti->mti_remote_einfo;
3158 union ldlm_policy_data *policy = &mti->mti_policy;
3159 struct ldlm_res_id *res_id = &mti->mti_res_id;
3163 LASSERT(mdt_object_remote(o));
3165 fid_build_reg_res_name(fid, res_id);
3167 memset(einfo, 0, sizeof(*einfo));
3168 einfo->ei_type = LDLM_IBITS;
3169 einfo->ei_mode = mode;
3170 einfo->ei_cb_bl = mdt_remote_blocking_ast;
3171 einfo->ei_cb_cp = ldlm_completion_ast;
3172 einfo->ei_enq_slave = 0;
3173 einfo->ei_res_id = res_id;
3177 * if we cache lock, couple lock with mdt_object, so that object
3178 * can be easily found in lock ASTs.
3180 einfo->ei_cbdata = o;
3181 einfo->ei_cb_created = mdt_remote_object_lock_created_cb;
3184 memset(policy, 0, sizeof(*policy));
3185 policy->l_inodebits.bits = *ibits;
3186 policy->l_inodebits.try_bits = trybits;
3188 rc = mo_object_lock(mti->mti_env, mdt_object_child(o), lh, einfo,
3191 /* Return successfully acquired bits to a caller */
3193 struct ldlm_lock *lock = ldlm_handle2lock(lh);
3196 *ibits = lock->l_policy_data.l_inodebits.bits;
3197 LDLM_LOCK_PUT(lock);
3202 int mdt_remote_object_lock(struct mdt_thread_info *mti, struct mdt_object *o,
3203 const struct lu_fid *fid, struct lustre_handle *lh,
3204 enum ldlm_mode mode, __u64 ibits, bool cache)
3206 return mdt_remote_object_lock_try(mti, o, fid, lh, mode, &ibits, 0,
3210 static int mdt_object_local_lock(struct mdt_thread_info *info,
3211 struct mdt_object *o,
3212 struct mdt_lock_handle *lh, __u64 *ibits,
3213 __u64 trybits, bool cos_incompat)
3215 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
3216 union ldlm_policy_data *policy = &info->mti_policy;
3217 struct ldlm_res_id *res_id = &info->mti_res_id;
3218 __u64 dlmflags = 0, *cookie = NULL;
3222 LASSERT(!lustre_handle_is_used(&lh->mlh_reg_lh));
3223 LASSERT(!lustre_handle_is_used(&lh->mlh_pdo_lh));
3224 LASSERT(lh->mlh_reg_mode != LCK_MINMODE);
3225 LASSERT(lh->mlh_type != MDT_NUL_LOCK);
3228 LASSERT(lh->mlh_reg_mode == LCK_PW ||
3229 lh->mlh_reg_mode == LCK_EX);
3230 dlmflags |= LDLM_FL_COS_INCOMPAT;
3231 } else if (mdt_cos_is_enabled(info->mti_mdt)) {
3232 dlmflags |= LDLM_FL_COS_ENABLED;
3235 /* Only enqueue LOOKUP lock for remote object */
3236 LASSERT(ergo(mdt_object_remote(o), *ibits == MDS_INODELOCK_LOOKUP));
3238 /* Lease lock are granted with LDLM_FL_CANCEL_ON_BLOCK */
3239 if (lh->mlh_type == MDT_REG_LOCK && lh->mlh_reg_mode == LCK_EX &&
3240 *ibits == MDS_INODELOCK_OPEN)
3241 dlmflags |= LDLM_FL_CANCEL_ON_BLOCK;
3243 if (lh->mlh_type == MDT_PDO_LOCK) {
3244 /* check for exists after object is locked */
3245 if (mdt_object_exists(o) == 0) {
3246 /* Non-existent object shouldn't have PDO lock */
3249 /* Non-dir object shouldn't have PDO lock */
3250 if (!S_ISDIR(lu_object_attr(&o->mot_obj)))
3255 fid_build_reg_res_name(mdt_object_fid(o), res_id);
3256 dlmflags |= LDLM_FL_ATOMIC_CB;
3259 cookie = &info->mti_exp->exp_handle.h_cookie;
3262 * Take PDO lock on whole directory and build correct @res_id for lock
3263 * on part of directory.
3265 if (lh->mlh_pdo_hash != 0) {
3266 LASSERT(lh->mlh_type == MDT_PDO_LOCK);
3267 mdt_lock_pdo_mode(info, o, lh);
3268 if (lh->mlh_pdo_mode != LCK_NL) {
3270 * Do not use LDLM_FL_LOCAL_ONLY for parallel lock, it
3271 * is never going to be sent to client and we do not
3272 * want it slowed down due to possible cancels.
3274 policy->l_inodebits.bits =
3275 *ibits & MDS_INODELOCK_UPDATE;
3276 policy->l_inodebits.try_bits =
3277 trybits & MDS_INODELOCK_UPDATE;
3278 /* at least one of them should be set */
3279 LASSERT(policy->l_inodebits.bits |
3280 policy->l_inodebits.try_bits);
3281 rc = mdt_fid_lock(info->mti_env, ns, &lh->mlh_pdo_lh,
3282 lh->mlh_pdo_mode, policy, res_id,
3284 if (unlikely(rc != 0))
3285 GOTO(out_unlock, rc);
3289 * Finish res_id initializing by name hash marking part of
3290 * directory which is taking modification.
3292 res_id->name[LUSTRE_RES_ID_HSH_OFF] = lh->mlh_pdo_hash;
3295 policy->l_inodebits.bits = *ibits;
3296 policy->l_inodebits.try_bits = trybits;
3299 * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
3300 * going to be sent to client. If it is - mdt_intent_policy() path will
3301 * fix it up and turn FL_LOCAL flag off.
3303 rc = mdt_fid_lock(info->mti_env, ns, &lh->mlh_reg_lh, lh->mlh_reg_mode,
3304 policy, res_id, LDLM_FL_LOCAL_ONLY | dlmflags,
3308 mdt_object_unlock(info, o, lh, 1);
3309 else if (unlikely(OBD_FAIL_PRECHECK(OBD_FAIL_MDS_PDO_LOCK)) &&
3310 lh->mlh_pdo_hash != 0 &&
3311 (lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX))
3312 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK, 15);
3314 /* Return successfully acquired bits to a caller */
3316 struct ldlm_lock *lock = ldlm_handle2lock(&lh->mlh_reg_lh);
3319 *ibits = lock->l_policy_data.l_inodebits.bits;
3320 LDLM_LOCK_PUT(lock);
3326 mdt_object_lock_internal(struct mdt_thread_info *info, struct mdt_object *o,
3327 struct mdt_lock_handle *lh, __u64 *ibits,
3328 __u64 trybits, bool cos_incompat)
3330 struct mdt_lock_handle *local_lh = NULL;
3334 if (!mdt_object_remote(o)) {
3335 rc = mdt_object_local_lock(info, o, lh, ibits, trybits,
3340 /* XXX do not support PERM/LAYOUT/XATTR lock for remote object yet */
3341 *ibits &= ~(MDS_INODELOCK_PERM | MDS_INODELOCK_LAYOUT |
3342 MDS_INODELOCK_XATTR);
3344 /* Only enqueue LOOKUP lock for remote object */
3345 if (*ibits & MDS_INODELOCK_LOOKUP) {
3346 __u64 local = MDS_INODELOCK_LOOKUP;
3348 rc = mdt_object_local_lock(info, o, lh, &local, 0,
3356 if ((*ibits | trybits) & MDS_INODELOCK_UPDATE) {
3357 /* Sigh, PDO needs to enqueue 2 locks right now, but
3358 * enqueue RPC can only request 1 lock, to avoid extra
3359 * RPC, so it will instead enqueue EX lock for remote
3360 * object anyway XXX*/
3361 if (lh->mlh_type == MDT_PDO_LOCK &&
3362 lh->mlh_pdo_hash != 0) {
3363 CDEBUG(D_INFO, "%s: "DFID" convert PDO lock to"
3364 "EX lock.\n", mdt_obd_name(info->mti_mdt),
3365 PFID(mdt_object_fid(o)));
3366 lh->mlh_pdo_hash = 0;
3367 lh->mlh_rreg_mode = LCK_EX;
3368 lh->mlh_type = MDT_REG_LOCK;
3371 rc = mdt_remote_object_lock_try(info, o, mdt_object_fid(o),
3374 ibits, trybits, false);
3375 if (rc != ELDLM_OK) {
3376 if (local_lh != NULL)
3377 mdt_object_unlock(info, o, local_lh, rc);
3382 /* other components like LFSCK can use lockless access
3383 * and populate cache, so we better invalidate it */
3384 mo_invalidate(info->mti_env, mdt_object_child(o));
3389 int mdt_object_lock(struct mdt_thread_info *info, struct mdt_object *o,
3390 struct mdt_lock_handle *lh, __u64 ibits)
3392 return mdt_object_lock_internal(info, o, lh, &ibits, 0, false);
3395 int mdt_reint_object_lock(struct mdt_thread_info *info, struct mdt_object *o,
3396 struct mdt_lock_handle *lh, __u64 ibits,
3399 LASSERT(lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX);
3400 return mdt_object_lock_internal(info, o, lh, &ibits, 0,
3404 int mdt_object_lock_try(struct mdt_thread_info *info, struct mdt_object *o,
3405 struct mdt_lock_handle *lh, __u64 *ibits,
3406 __u64 trybits, bool cos_incompat)
3408 bool trylock_only = *ibits == 0;
3411 LASSERT(!(*ibits & trybits));
3412 rc = mdt_object_lock_internal(info, o, lh, ibits, trybits,
3414 if (rc && trylock_only) { /* clear error for try ibits lock only */
3415 LASSERT(*ibits == 0);
3422 * Save a lock within request object.
3424 * Keep the lock referenced until whether client ACK or transaction
3425 * commit happens or release the lock immediately depending on input
3426 * parameters. If COS is ON, a write lock is converted to COS lock
3429 * \param info thead info object
3430 * \param h lock handle
3431 * \param mode lock mode
3432 * \param decref force immediate lock releasing
3434 void mdt_save_lock(struct mdt_thread_info *info, struct lustre_handle *h,
3435 enum ldlm_mode mode, int decref)
3439 if (lustre_handle_is_used(h)) {
3440 if (decref || !info->mti_has_trans ||
3441 !(mode & (LCK_PW | LCK_EX))) {
3442 mdt_fid_unlock(h, mode);
3444 struct mdt_device *mdt = info->mti_mdt;
3445 struct ldlm_lock *lock = ldlm_handle2lock(h);
3446 struct ptlrpc_request *req = mdt_info_req(info);
3447 bool cos = mdt_cos_is_enabled(mdt);
3448 bool convert_lock = !cos && mdt_slc_is_enabled(mdt);
3450 LASSERTF(lock != NULL, "no lock for cookie %#llx\n",
3453 /* there is no request if mdt_object_unlock() is called
3454 * from mdt_export_cleanup()->mdt_add_dirty_flag() */
3455 if (likely(req != NULL)) {
3456 LDLM_DEBUG(lock, "save lock request %p reply "
3457 "state %p transno %lld\n", req,
3458 req->rq_reply_state, req->rq_transno);
3460 ldlm_lock_mode_downgrade(lock, LCK_COS);
3463 if (req->rq_export->exp_disconnected)
3464 mdt_fid_unlock(h, mode);
3466 ptlrpc_save_lock(req, h, mode, cos,
3469 mdt_fid_unlock(h, mode);
3471 if (mdt_is_lock_sync(lock)) {
3472 CDEBUG(D_HA, "found sync-lock,"
3473 " async commit started\n");
3474 mdt_device_commit_async(info->mti_env,
3477 LDLM_LOCK_PUT(lock);
3486 * Save cross-MDT lock in uncommitted_slc_locks
3488 * Keep the lock referenced until transaction commit happens or release the lock
3489 * immediately depending on input parameters.
3491 * \param info thead info object
3492 * \param h lock handle
3493 * \param mode lock mode
3494 * \param decref force immediate lock releasing
3496 static void mdt_save_remote_lock(struct mdt_thread_info *info,
3497 struct mdt_object *o, struct lustre_handle *h,
3498 enum ldlm_mode mode, int decref)
3502 if (lustre_handle_is_used(h)) {
3503 struct ldlm_lock *lock = ldlm_handle2lock(h);
3506 (lock->l_policy_data.l_inodebits.bits &
3507 (MDS_INODELOCK_XATTR | MDS_INODELOCK_UPDATE)))
3508 mo_invalidate(info->mti_env, mdt_object_child(o));
3510 if (decref || !info->mti_has_trans ||
3511 !(mode & (LCK_PW | LCK_EX))) {
3512 ldlm_lock_decref_and_cancel(h, mode);
3513 LDLM_LOCK_PUT(lock);
3515 struct ptlrpc_request *req = mdt_info_req(info);
3517 LASSERT(req != NULL);
3518 tgt_save_slc_lock(&info->mti_mdt->mdt_lut, lock,
3520 ldlm_lock_decref(h, mode);
3529 * Unlock mdt object.
3531 * Immeditely release the regular lock and the PDO lock or save the
3532 * lock in request and keep them referenced until client ACK or
3533 * transaction commit.
3535 * \param info thread info object
3536 * \param o mdt object
3537 * \param lh mdt lock handle referencing regular and PDO locks
3538 * \param decref force immediate lock releasing
3540 * XXX o is not used and may be NULL, see hsm_cdt_request_completed().
3542 void mdt_object_unlock(struct mdt_thread_info *info, struct mdt_object *o,
3543 struct mdt_lock_handle *lh, int decref)
3547 mdt_save_lock(info, &lh->mlh_pdo_lh, lh->mlh_pdo_mode, decref);
3548 mdt_save_lock(info, &lh->mlh_reg_lh, lh->mlh_reg_mode, decref);
3549 mdt_save_remote_lock(info, o, &lh->mlh_rreg_lh, lh->mlh_rreg_mode,
3555 struct mdt_object *mdt_object_find_lock(struct mdt_thread_info *info,
3556 const struct lu_fid *f,
3557 struct mdt_lock_handle *lh,
3560 struct mdt_object *o;
3562 o = mdt_object_find(info->mti_env, info->mti_mdt, f);
3566 rc = mdt_object_lock(info, o, lh, ibits);
3568 mdt_object_put(info->mti_env, o);
3575 void mdt_object_unlock_put(struct mdt_thread_info * info,
3576 struct mdt_object * o,
3577 struct mdt_lock_handle *lh,
3580 mdt_object_unlock(info, o, lh, decref);
3581 mdt_object_put(info->mti_env, o);
3585 * Generic code handling requests that have struct mdt_body passed in:
3587 * - extract mdt_body from request and save it in @info, if present;
3589 * - create lu_object, corresponding to the fid in mdt_body, and save it in
3592 * - if HAS_BODY flag is set for this request type check whether object
3593 * actually exists on storage (lu_object_exists()).
3596 static int mdt_body_unpack(struct mdt_thread_info *info,
3597 enum tgt_handler_flags flags)
3599 const struct mdt_body *body;
3600 struct mdt_object *obj;
3601 const struct lu_env *env;
3602 struct req_capsule *pill;
3606 env = info->mti_env;
3607 pill = info->mti_pill;
3609 body = info->mti_body = req_capsule_client_get(pill, &RMF_MDT_BODY);
3613 if (!(body->mbo_valid & OBD_MD_FLID))
3616 if (!fid_is_sane(&body->mbo_fid1)) {
3617 CERROR("Invalid fid: "DFID"\n", PFID(&body->mbo_fid1));
3621 obj = mdt_object_find(env, info->mti_mdt, &body->mbo_fid1);
3623 if ((flags & HAS_BODY) && !mdt_object_exists(obj)) {
3624 mdt_object_put(env, obj);
3627 info->mti_object = obj;
3636 static int mdt_unpack_req_pack_rep(struct mdt_thread_info *info,
3637 enum tgt_handler_flags flags)
3639 struct req_capsule *pill = info->mti_pill;
3644 if (req_capsule_has_field(pill, &RMF_MDT_BODY, RCL_CLIENT))
3645 rc = mdt_body_unpack(info, flags);
3649 if (rc == 0 && (flags & HAS_REPLY)) {
3651 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
3652 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER,
3654 if (req_capsule_has_field(pill, &RMF_LOGCOOKIES, RCL_SERVER))
3655 req_capsule_set_size(pill, &RMF_LOGCOOKIES,
3658 /* Set ACL reply buffer size as LUSTRE_POSIX_ACL_MAX_SIZE_OLD
3659 * by default. If the target object has more ACL entries, then
3660 * enlarge the buffer when necessary. */
3661 if (req_capsule_has_field(pill, &RMF_ACL, RCL_SERVER))
3662 req_capsule_set_size(pill, &RMF_ACL, RCL_SERVER,
3663 LUSTRE_POSIX_ACL_MAX_SIZE_OLD);
3665 mdt_preset_secctx_size(info);
3667 rc = req_capsule_server_pack(pill);
3669 CWARN("%s: cannot pack response: rc = %d\n",
3670 mdt_obd_name(info->mti_mdt), rc);
3675 void mdt_lock_handle_init(struct mdt_lock_handle *lh)
3677 lh->mlh_type = MDT_NUL_LOCK;
3678 lh->mlh_reg_lh.cookie = 0ull;
3679 lh->mlh_reg_mode = LCK_MINMODE;
3680 lh->mlh_pdo_lh.cookie = 0ull;
3681 lh->mlh_pdo_mode = LCK_MINMODE;
3682 lh->mlh_rreg_lh.cookie = 0ull;
3683 lh->mlh_rreg_mode = LCK_MINMODE;
3686 void mdt_lock_handle_fini(struct mdt_lock_handle *lh)
3688 LASSERT(!lustre_handle_is_used(&lh->mlh_reg_lh));
3689 LASSERT(!lustre_handle_is_used(&lh->mlh_pdo_lh));
3693 * Initialize fields of struct mdt_thread_info. Other fields are left in
3694 * uninitialized state, because it's too expensive to zero out whole
3695 * mdt_thread_info (> 1K) on each request arrival.
3697 void mdt_thread_info_init(struct ptlrpc_request *req,
3698 struct mdt_thread_info *info)
3702 info->mti_pill = &req->rq_pill;
3705 for (i = 0; i < ARRAY_SIZE(info->mti_lh); i++)
3706 mdt_lock_handle_init(&info->mti_lh[i]);
3708 /* mdt device: it can be NULL while CONNECT */
3709 if (req->rq_export) {
3710 info->mti_mdt = mdt_dev(req->rq_export->exp_obd->obd_lu_dev);
3711 info->mti_exp = req->rq_export;
3713 info->mti_mdt = NULL;
3714 info->mti_env = req->rq_svc_thread->t_env;
3715 info->mti_transno = lustre_msg_get_transno(req->rq_reqmsg);
3717 memset(&info->mti_attr, 0, sizeof(info->mti_attr));
3718 info->mti_big_buf = LU_BUF_NULL;
3719 info->mti_body = NULL;
3720 info->mti_object = NULL;
3721 info->mti_dlm_req = NULL;
3722 info->mti_has_trans = 0;
3723 info->mti_cross_ref = 0;
3724 info->mti_opdata = 0;
3725 info->mti_big_lmm_used = 0;
3726 info->mti_big_acl_used = 0;
3727 info->mti_som_valid = 0;
3729 info->mti_spec.no_create = 0;
3730 info->mti_spec.sp_rm_entry = 0;
3731 info->mti_spec.sp_permitted = 0;
3732 info->mti_spec.sp_migrate_close = 0;
3734 info->mti_spec.u.sp_ea.eadata = NULL;
3735 info->mti_spec.u.sp_ea.eadatalen = 0;
3738 void mdt_thread_info_fini(struct mdt_thread_info *info)
3742 if (info->mti_object != NULL) {
3743 mdt_object_put(info->mti_env, info->mti_object);
3744 info->mti_object = NULL;
3747 for (i = 0; i < ARRAY_SIZE(info->mti_lh); i++)
3748 mdt_lock_handle_fini(&info->mti_lh[i]);
3749 info->mti_env = NULL;
3750 info->mti_pill = NULL;
3751 info->mti_exp = NULL;
3753 if (unlikely(info->mti_big_buf.lb_buf != NULL))
3754 lu_buf_free(&info->mti_big_buf);
3757 struct mdt_thread_info *tsi2mdt_info(struct tgt_session_info *tsi)
3759 struct mdt_thread_info *mti;
3761 mti = mdt_th_info(tsi->tsi_env);
3762 LASSERT(mti != NULL);
3764 mdt_thread_info_init(tgt_ses_req(tsi), mti);
3765 if (tsi->tsi_corpus != NULL) {
3766 mti->mti_object = mdt_obj(tsi->tsi_corpus);
3767 lu_object_get(tsi->tsi_corpus);
3769 mti->mti_body = tsi->tsi_mdt_body;
3770 mti->mti_dlm_req = tsi->tsi_dlm_req;
3775 static int mdt_tgt_connect(struct tgt_session_info *tsi)
3777 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_DELAY_CONDITIONAL) &&
3779 tsi2mdt_info(tsi)->mti_mdt->mdt_seq_site.ss_node_id) {
3780 set_current_state(TASK_UNINTERRUPTIBLE);
3781 schedule_timeout(cfs_time_seconds(3));
3784 return tgt_connect(tsi);
3787 static int mdt_intent_glimpse(enum ldlm_intent_flags it_opc,
3788 struct mdt_thread_info *info,
3789 struct ldlm_lock **lockp, __u64 flags)
3791 return mdt_glimpse_enqueue(info, info->mti_mdt->mdt_namespace,
3794 static int mdt_intent_brw(enum ldlm_intent_flags it_opc,
3795 struct mdt_thread_info *info,
3796 struct ldlm_lock **lockp, __u64 flags)
3798 return mdt_brw_enqueue(info, info->mti_mdt->mdt_namespace,
3802 int mdt_intent_lock_replace(struct mdt_thread_info *info,
3803 struct ldlm_lock **lockp,
3804 struct mdt_lock_handle *lh,
3805 __u64 flags, int result)
3807 struct ptlrpc_request *req = mdt_info_req(info);
3808 struct ldlm_lock *lock = *lockp;
3809 struct ldlm_lock *new_lock;
3811 /* If possible resent found a lock, @lh is set to its handle */
3812 new_lock = ldlm_handle2lock_long(&lh->mlh_reg_lh, 0);
3814 if (new_lock == NULL && (flags & LDLM_FL_INTENT_ONLY)) {
3815 lh->mlh_reg_lh.cookie = 0;
3819 if (new_lock == NULL && (flags & LDLM_FL_RESENT)) {
3820 /* Lock is pinned by ldlm_handle_enqueue0() as it is
3821 * a resend case, however, it could be already destroyed
3822 * due to client eviction or a raced cancel RPC. */
3823 LDLM_DEBUG_NOLOCK("Invalid lock handle %#llx\n",
3824 lh->mlh_reg_lh.cookie);
3825 lh->mlh_reg_lh.cookie = 0;
3829 LASSERTF(new_lock != NULL,
3830 "lockh %#llx flags %#llx : rc = %d\n",
3831 lh->mlh_reg_lh.cookie, flags, result);
3834 * If we've already given this lock to a client once, then we should
3835 * have no readers or writers. Otherwise, we should have one reader
3836 * _or_ writer ref (which will be zeroed below) before returning the
3839 if (new_lock->l_export == req->rq_export) {
3840 LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
3842 LASSERT(new_lock->l_export == NULL);
3843 LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
3848 if (new_lock->l_export == req->rq_export) {
3850 * Already gave this to the client, which means that we
3851 * reconstructed a reply.
3853 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
3856 LDLM_LOCK_RELEASE(new_lock);
3857 lh->mlh_reg_lh.cookie = 0;
3858 RETURN(ELDLM_LOCK_REPLACED);
3862 * Fixup the lock to be given to the client.
3864 lock_res_and_lock(new_lock);
3865 /* Zero new_lock->l_readers and new_lock->l_writers without triggering
3866 * possible blocking AST. */
3867 while (new_lock->l_readers > 0) {
3868 lu_ref_del(&new_lock->l_reference, "reader", new_lock);
3869 lu_ref_del(&new_lock->l_reference, "user", new_lock);
3870 new_lock->l_readers--;
3872 while (new_lock->l_writers > 0) {
3873 lu_ref_del(&new_lock->l_reference, "writer", new_lock);
3874 lu_ref_del(&new_lock->l_reference, "user", new_lock);
3875 new_lock->l_writers--;
3878 new_lock->l_export = class_export_lock_get(req->rq_export, new_lock);
3879 new_lock->l_blocking_ast = lock->l_blocking_ast;
3880 new_lock->l_completion_ast = lock->l_completion_ast;
3881 if (ldlm_has_dom(new_lock))
3882 new_lock->l_glimpse_ast = ldlm_server_glimpse_ast;
3883 new_lock->l_remote_handle = lock->l_remote_handle;
3884 new_lock->l_flags &= ~LDLM_FL_LOCAL;
3886 unlock_res_and_lock(new_lock);
3888 cfs_hash_add(new_lock->l_export->exp_lock_hash,
3889 &new_lock->l_remote_handle,
3890 &new_lock->l_exp_hash);
3892 LDLM_LOCK_RELEASE(new_lock);
3893 lh->mlh_reg_lh.cookie = 0;
3895 RETURN(ELDLM_LOCK_REPLACED);
3898 void mdt_intent_fixup_resent(struct mdt_thread_info *info,
3899 struct ldlm_lock *new_lock,
3900 struct mdt_lock_handle *lh, __u64 flags)
3902 struct ptlrpc_request *req = mdt_info_req(info);
3903 struct ldlm_request *dlmreq;
3905 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
3908 dlmreq = req_capsule_client_get(info->mti_pill, &RMF_DLM_REQ);
3910 /* Check if this is a resend case (MSG_RESENT is set on RPC) and a
3911 * lock was found by ldlm_handle_enqueue(); if so @lh must be
3913 if (flags & LDLM_FL_RESENT) {
3914 lh->mlh_reg_lh.cookie = new_lock->l_handle.h_cookie;
3915 lh->mlh_reg_mode = new_lock->l_granted_mode;
3917 LDLM_DEBUG(new_lock, "Restoring lock cookie");
3918 DEBUG_REQ(D_DLMTRACE, req, "restoring lock cookie %#llx",
3919 lh->mlh_reg_lh.cookie);
3924 * If the xid matches, then we know this is a resent request, and allow
3925 * it. (It's probably an OPEN, for which we don't send a lock.
3927 if (req_can_reconstruct(req, NULL) == 1)
3931 * This remote handle isn't enqueued, so we never received or processed
3932 * this request. Clear MSG_RESENT, because it can be handled like any
3933 * normal request now.
3935 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
3937 DEBUG_REQ(D_DLMTRACE, req, "no existing lock with rhandle %#llx",
3938 dlmreq->lock_handle[0].cookie);
3941 static int mdt_intent_getxattr(enum ldlm_intent_flags it_opc,
3942 struct mdt_thread_info *info,
3943 struct ldlm_lock **lockp,
3946 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
3947 struct ldlm_reply *ldlm_rep = NULL;
3952 * Initialize lhc->mlh_reg_lh either from a previously granted lock
3953 * (for the resend case) or a new lock. Below we will use it to
3954 * replace the original lock.
3956 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
3957 if (!lustre_handle_is_used(&lhc->mlh_reg_lh)) {
3958 mdt_lock_reg_init(lhc, (*lockp)->l_req_mode);
3959 rc = mdt_object_lock(info, info->mti_object, lhc,
3960 MDS_INODELOCK_XATTR);
3965 rc = mdt_getxattr(info);
3967 if (mdt_info_req(info)->rq_repmsg != NULL)
3968 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
3970 if (ldlm_rep == NULL ||
3971 OBD_FAIL_CHECK(OBD_FAIL_MDS_XATTR_REP)) {
3972 mdt_object_unlock(info, info->mti_object, lhc, 1);
3976 RETURN(err_serious(-EFAULT));
3979 ldlm_rep->lock_policy_res2 = clear_serious(rc);
3981 /* This is left for interop instead of adding a new interop flag.
3983 #if LUSTRE_VERSION_CODE > OBD_OCD_VERSION(3, 0, 0, 0)
3984 if (ldlm_rep->lock_policy_res2) {
3985 mdt_object_unlock(info, info->mti_object, lhc, 1);
3986 RETURN(ELDLM_LOCK_ABORTED);
3990 rc = mdt_intent_lock_replace(info, lockp, lhc, flags, rc);
3994 static int mdt_intent_getattr(enum ldlm_intent_flags it_opc,
3995 struct mdt_thread_info *info,
3996 struct ldlm_lock **lockp,
3999 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
4001 struct ldlm_reply *ldlm_rep;
4002 struct mdt_body *reqbody;
4003 struct mdt_body *repbody;
4007 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
4010 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
4013 info->mti_cross_ref = !!(reqbody->mbo_valid & OBD_MD_FLCROSSREF);
4014 repbody->mbo_eadatasize = 0;
4015 repbody->mbo_aclsize = 0;
4019 child_bits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
4022 child_bits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE |
4026 CERROR("%s: unsupported intent %#x\n",
4027 mdt_obd_name(info->mti_mdt), (unsigned int)it_opc);
4028 GOTO(out_shrink, rc = -EINVAL);
4031 rc = mdt_init_ucred_intent_getattr(info, reqbody);
4033 GOTO(out_shrink, rc);
4035 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
4036 mdt_set_disposition(info, ldlm_rep, DISP_IT_EXECD);
4038 /* Get lock from request for possible resent case. */
4039 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
4041 rc = mdt_getattr_name_lock(info, lhc, child_bits, ldlm_rep);
4042 ldlm_rep->lock_policy_res2 = clear_serious(rc);
4044 if (mdt_get_disposition(ldlm_rep, DISP_LOOKUP_NEG))
4045 ldlm_rep->lock_policy_res2 = 0;
4046 if (!mdt_get_disposition(ldlm_rep, DISP_LOOKUP_POS) ||
4047 ldlm_rep->lock_policy_res2) {
4048 lhc->mlh_reg_lh.cookie = 0ull;
4049 GOTO(out_ucred, rc = ELDLM_LOCK_ABORTED);
4052 rc = mdt_intent_lock_replace(info, lockp, lhc, flags, rc);
4055 mdt_exit_ucred(info);
4057 mdt_client_compatibility(info);
4058 rc2 = mdt_fix_reply(info);
4064 static int mdt_intent_layout(enum ldlm_intent_flags it_opc,
4065 struct mdt_thread_info *info,
4066 struct ldlm_lock **lockp,
4069 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_LAYOUT];
4070 struct md_layout_change layout = { .mlc_opc = MD_LAYOUT_NOP };
4071 struct layout_intent *intent;
4072 struct lu_fid *fid = &info->mti_tmp_fid2;
4073 struct mdt_object *obj = NULL;
4074 int layout_size = 0;
4078 fid_extract_from_res_name(fid, &(*lockp)->l_resource->lr_name);
4080 intent = req_capsule_client_get(info->mti_pill, &RMF_LAYOUT_INTENT);
4084 CDEBUG(D_INFO, DFID "got layout change request from client: "
4085 "opc:%u flags:%#x extent "DEXT"\n",
4086 PFID(fid), intent->li_opc, intent->li_flags,
4087 PEXT(&intent->li_extent));
4089 switch (intent->li_opc) {
4090 case LAYOUT_INTENT_TRUNC:
4091 case LAYOUT_INTENT_WRITE:
4092 layout.mlc_opc = MD_LAYOUT_WRITE;
4093 layout.mlc_intent = intent;
4095 case LAYOUT_INTENT_ACCESS:
4097 case LAYOUT_INTENT_READ:
4098 case LAYOUT_INTENT_GLIMPSE:
4099 case LAYOUT_INTENT_RELEASE:
4100 case LAYOUT_INTENT_RESTORE:
4101 CERROR("%s: Unsupported layout intent opc %d\n",
4102 mdt_obd_name(info->mti_mdt), intent->li_opc);
4106 CERROR("%s: Unknown layout intent opc %d\n",
4107 mdt_obd_name(info->mti_mdt), intent->li_opc);
4114 /* Get lock from request for possible resent case. */
4115 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
4117 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
4119 GOTO(out, rc = PTR_ERR(obj));
4122 if (mdt_object_exists(obj) && !mdt_object_remote(obj)) {
4123 /* if layout is going to be changed don't use the current EA
4124 * size but the maximum one. That buffer will be shrinked
4125 * to the actual size in req_capsule_shrink() before reply.
4127 if (layout.mlc_opc == MD_LAYOUT_WRITE) {
4128 layout_size = info->mti_mdt->mdt_max_mdsize;
4130 layout_size = mdt_attr_get_eabuf_size(info, obj);
4131 if (layout_size < 0)
4132 GOTO(out_obj, rc = layout_size);
4134 if (layout_size > info->mti_mdt->mdt_max_mdsize)
4135 info->mti_mdt->mdt_max_mdsize = layout_size;
4137 CDEBUG(D_INFO, "%s: layout_size %d\n",
4138 mdt_obd_name(info->mti_mdt), layout_size);
4142 * set reply buffer size, so that ldlm_handle_enqueue0()->
4143 * ldlm_lvbo_fill() will fill the reply buffer with lovea.
4145 (*lockp)->l_lvb_type = LVB_T_LAYOUT;
4146 req_capsule_set_size(info->mti_pill, &RMF_DLM_LVB, RCL_SERVER,
4148 rc = req_capsule_server_pack(info->mti_pill);
4153 if (layout.mlc_opc != MD_LAYOUT_NOP) {
4154 struct lu_buf *buf = &layout.mlc_buf;
4157 * mdt_layout_change is a reint operation, when the request
4158 * is resent, layout write shouldn't reprocess it again.
4160 rc = mdt_check_resent(info, mdt_reconstruct_generic, lhc);
4162 GOTO(out_obj, rc = rc < 0 ? rc : 0);
4165 * There is another resent case: the client's job has been
4166 * done by another client, referring lod_declare_layout_change
4167 * -EALREADY case, and it became a operation w/o transaction,
4168 * so we should not do the layout change, otherwise
4169 * mdt_layout_change() will try to cancel the granted server
4170 * CR lock whose remote counterpart is still in hold on the
4171 * client, and a deadlock ensues.
4173 rc = mdt_check_resent_lock(info, obj, lhc);
4179 if (unlikely(req_is_replay(mdt_info_req(info)))) {
4180 buf->lb_buf = req_capsule_client_get(info->mti_pill,
4182 buf->lb_len = req_capsule_get_size(info->mti_pill,
4183 &RMF_EADATA, RCL_CLIENT);
4185 * If it's a replay of layout write intent RPC, the
4186 * client has saved the extended lovea when
4187 * it get reply then.
4189 if (buf->lb_len > 0)
4190 mdt_fix_lov_magic(info, buf->lb_buf);
4193 * Instantiate some layout components, if @buf contains
4194 * lovea, then it's a replay of the layout intent write
4197 rc = mdt_layout_change(info, obj, &layout);
4202 mdt_object_put(info->mti_env, obj);
4204 if (rc == 0 && lustre_handle_is_used(&lhc->mlh_reg_lh))
4205 rc = mdt_intent_lock_replace(info, lockp, lhc, flags, rc);
4208 lhc->mlh_reg_lh.cookie = 0;
4213 static int mdt_intent_open(enum ldlm_intent_flags it_opc,
4214 struct mdt_thread_info *info,
4215 struct ldlm_lock **lockp,
4218 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
4219 struct ldlm_reply *rep = NULL;
4223 static const struct req_format *intent_fmts[REINT_MAX] = {
4224 [REINT_CREATE] = &RQF_LDLM_INTENT_CREATE,
4225 [REINT_OPEN] = &RQF_LDLM_INTENT_OPEN
4230 opc = mdt_reint_opcode(mdt_info_req(info), intent_fmts);
4234 /* Get lock from request for possible resent case. */
4235 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
4237 rc = mdt_reint_internal(info, lhc, opc);
4239 /* Check whether the reply has been packed successfully. */
4240 if (mdt_info_req(info)->rq_repmsg != NULL)
4241 rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
4246 RETURN(err_serious(-EFAULT));
4249 /* MDC expects this in any case */
4251 mdt_set_disposition(info, rep, DISP_LOOKUP_EXECD);
4253 /* the open lock or the lock for cross-ref object should be
4254 * returned to the client */
4255 if (lustre_handle_is_used(&lhc->mlh_reg_lh) &&
4256 (rc == 0 || rc == -MDT_EREMOTE_OPEN)) {
4257 rep->lock_policy_res2 = 0;
4258 rc = mdt_intent_lock_replace(info, lockp, lhc, flags, rc);
4262 rep->lock_policy_res2 = clear_serious(rc);
4264 if (rep->lock_policy_res2 == -ENOENT &&
4265 mdt_get_disposition(rep, DISP_LOOKUP_NEG) &&
4266 !mdt_get_disposition(rep, DISP_OPEN_CREATE))
4267 rep->lock_policy_res2 = 0;
4269 lhc->mlh_reg_lh.cookie = 0ull;
4270 if (rc == -ENOTCONN || rc == -ENODEV ||
4271 rc == -EOVERFLOW) { /**< if VBR failure then return error */
4273 * If it is the disconnect error (ENODEV & ENOCONN), the error
4274 * will be returned by rq_status, and client at ptlrpc layer
4275 * will detect this, then disconnect, reconnect the import
4276 * immediately, instead of impacting the following the rpc.
4281 * For other cases, the error will be returned by intent, and client
4282 * will retrieve the result from intent.
4284 RETURN(ELDLM_LOCK_ABORTED);
4287 static int mdt_intent_opc(enum ldlm_intent_flags it_opc,
4288 struct mdt_thread_info *info,
4289 struct ldlm_lock **lockp,
4290 u64 flags /* LDLM_FL_* */)
4292 struct req_capsule *pill = info->mti_pill;
4293 struct ptlrpc_request *req = mdt_info_req(info);
4294 const struct req_format *it_format;
4295 int (*it_handler)(enum ldlm_intent_flags,
4296 struct mdt_thread_info *,
4297 struct ldlm_lock **,
4299 enum tgt_handler_flags it_handler_flags = 0;
4300 struct ldlm_reply *rep;
4301 bool check_mdt_object = false;
4307 case IT_OPEN|IT_CREAT:
4309 * OCREAT is not a IS_MUTABLE request since the file may
4310 * already exist. We do the extra check of
4311 * OBD_CONNECT_RDONLY in mdt_reint_open() when we
4312 * really need to create the object.
4314 it_format = &RQF_LDLM_INTENT;
4315 it_handler = &mdt_intent_open;
4318 check_mdt_object = true;
4320 it_format = &RQF_LDLM_INTENT_GETATTR;
4321 it_handler = &mdt_intent_getattr;
4322 it_handler_flags = HAS_REPLY;
4325 check_mdt_object = true;
4326 it_format = &RQF_LDLM_INTENT_GETXATTR;
4327 it_handler = &mdt_intent_getxattr;
4328 it_handler_flags = HAS_BODY;
4331 it_format = &RQF_LDLM_INTENT_LAYOUT;
4332 it_handler = &mdt_intent_layout;
4335 it_format = &RQF_LDLM_INTENT;
4336 it_handler = &mdt_intent_glimpse;
4339 it_format = &RQF_LDLM_INTENT;
4340 it_handler = &mdt_intent_brw;
4342 case IT_QUOTA_DQACQ:
4343 case IT_QUOTA_CONN: {
4344 struct lu_device *qmt = info->mti_mdt->mdt_qmt_dev;
4347 RETURN(-EOPNOTSUPP);
4349 if (mdt_rdonly(req->rq_export))
4352 (*lockp)->l_lvb_type = LVB_T_LQUOTA;
4353 /* pass the request to quota master */
4354 rc = qmt_hdls.qmth_intent_policy(info->mti_env, qmt,
4355 mdt_info_req(info), lockp,
4360 CERROR("%s: unknown intent code %#x\n",
4361 mdt_obd_name(info->mti_mdt), it_opc);
4365 req_capsule_extend(pill, it_format);
4367 rc = mdt_unpack_req_pack_rep(info, it_handler_flags);
4371 if (unlikely(info->mti_object == NULL && check_mdt_object))
4374 if (it_handler_flags & IS_MUTABLE && mdt_rdonly(req->rq_export))
4377 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_INTENT_DELAY, 10);
4379 /* execute policy */
4380 rc = (*it_handler)(it_opc, info, lockp, flags);
4382 /* Check whether the reply has been packed successfully. */
4383 if (req->rq_repmsg != NULL) {
4384 rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
4385 rep->lock_policy_res2 =
4386 ptlrpc_status_hton(rep->lock_policy_res2);
4392 static void mdt_ptlrpc_stats_update(struct ptlrpc_request *req,
4393 enum ldlm_intent_flags it_opc)
4395 struct lprocfs_stats *srv_stats = ptlrpc_req2svc(req)->srv_stats;
4397 /* update stats when IT code is known */
4398 if (srv_stats != NULL)
4399 lprocfs_counter_incr(srv_stats,
4400 PTLRPC_LAST_CNTR + (it_opc == IT_GLIMPSE ?
4401 LDLM_GLIMPSE_ENQUEUE : LDLM_IBITS_ENQUEUE));
4404 static int mdt_intent_policy(const struct lu_env *env,
4405 struct ldlm_namespace *ns,
4406 struct ldlm_lock **lockp,
4408 enum ldlm_mode mode,
4409 __u64 flags, void *data)
4411 struct tgt_session_info *tsi;
4412 struct mdt_thread_info *info;
4413 struct ptlrpc_request *req = req_cookie;
4414 struct ldlm_intent *it;
4415 struct req_capsule *pill;
4416 const struct ldlm_lock_desc *ldesc;
4421 LASSERT(req != NULL);
4423 tsi = tgt_ses_info(env);
4425 info = tsi2mdt_info(tsi);
4426 LASSERT(info != NULL);
4427 pill = info->mti_pill;
4428 LASSERT(pill->rc_req == req);
4429 ldesc = &info->mti_dlm_req->lock_desc;
4431 if (req->rq_reqmsg->lm_bufcount > DLM_INTENT_IT_OFF) {
4432 req_capsule_extend(pill, &RQF_LDLM_INTENT_BASIC);
4433 it = req_capsule_client_get(pill, &RMF_LDLM_INTENT);
4435 mdt_ptlrpc_stats_update(req, it->opc);
4436 rc = mdt_intent_opc(it->opc, info, lockp, flags);
4440 /* Lock without inodebits makes no sense and will oops
4441 * later in ldlm. Let's check it now to see if we have
4442 * ibits corrupted somewhere in mdt_intent_opc().
4443 * The case for client miss to set ibits has been
4444 * processed by others. */
4445 LASSERT(ergo(ldesc->l_resource.lr_type == LDLM_IBITS,
4446 ldesc->l_policy_data.l_inodebits.bits != 0));
4448 rc = err_serious(-EFAULT);
4451 /* No intent was provided */
4452 req_capsule_set_size(pill, &RMF_DLM_LVB, RCL_SERVER, 0);
4453 rc = req_capsule_server_pack(pill);
4455 rc = err_serious(rc);
4457 mdt_thread_info_fini(info);
4461 static void mdt_deregister_seq_exp(struct mdt_device *mdt)
4463 struct seq_server_site *ss = mdt_seq_site(mdt);
4465 if (ss->ss_node_id == 0)
4468 if (ss->ss_client_seq != NULL) {
4469 lustre_deregister_lwp_item(&ss->ss_client_seq->lcs_exp);
4470 ss->ss_client_seq->lcs_exp = NULL;
4473 if (ss->ss_server_fld != NULL) {
4474 lustre_deregister_lwp_item(&ss->ss_server_fld->lsf_control_exp);
4475 ss->ss_server_fld->lsf_control_exp = NULL;
4479 static void mdt_seq_fini_cli(struct mdt_device *mdt)
4481 struct seq_server_site *ss = mdt_seq_site(mdt);
4486 if (ss->ss_server_seq != NULL)
4487 seq_server_set_cli(NULL, ss->ss_server_seq, NULL);
4490 static int mdt_seq_fini(const struct lu_env *env, struct mdt_device *mdt)
4492 mdt_seq_fini_cli(mdt);
4493 mdt_deregister_seq_exp(mdt);
4495 return seq_site_fini(env, mdt_seq_site(mdt));
4499 * It will retrieve its FLDB entries from MDT0, and it only happens
4500 * when upgrading existent FS to 2.6 or when local FLDB is corrupted,
4501 * and it needs to refresh FLDB from the MDT0.
4503 static int mdt_register_lwp_callback(void *data)
4506 struct mdt_device *mdt = data;
4507 struct lu_server_fld *fld = mdt_seq_site(mdt)->ss_server_fld;
4511 LASSERT(mdt_seq_site(mdt)->ss_node_id != 0);
4513 rc = lu_env_init(&env, LCT_MD_THREAD);
4515 CERROR("%s: cannot init env: rc = %d\n", mdt_obd_name(mdt), rc);
4519 /* Allocate new sequence now to avoid creating local transaction
4520 * in the normal transaction process */
4521 rc = seq_server_check_and_alloc_super(&env,
4522 mdt_seq_site(mdt)->ss_server_seq);
4527 rc = fld_update_from_controller(&env, fld);
4529 CERROR("%s: cannot update controller: rc = %d\n",
4530 mdt_obd_name(mdt), rc);
4539 static int mdt_register_seq_exp(struct mdt_device *mdt)
4541 struct seq_server_site *ss = mdt_seq_site(mdt);
4542 char *lwp_name = NULL;
4545 if (ss->ss_node_id == 0)
4548 OBD_ALLOC(lwp_name, MAX_OBD_NAME);
4549 if (lwp_name == NULL)
4550 GOTO(out_free, rc = -ENOMEM);
4552 rc = tgt_name2lwp_name(mdt_obd_name(mdt), lwp_name, MAX_OBD_NAME, 0);
4556 rc = lustre_register_lwp_item(lwp_name, &ss->ss_client_seq->lcs_exp,
4561 rc = lustre_register_lwp_item(lwp_name,
4562 &ss->ss_server_fld->lsf_control_exp,
4563 mdt_register_lwp_callback, mdt);
4565 lustre_deregister_lwp_item(&ss->ss_client_seq->lcs_exp);
4566 ss->ss_client_seq->lcs_exp = NULL;
4570 if (lwp_name != NULL)
4571 OBD_FREE(lwp_name, MAX_OBD_NAME);
4577 * Init client sequence manager which is used by local MDS to talk to sequence
4578 * controller on remote node.
4580 static int mdt_seq_init_cli(const struct lu_env *env, struct mdt_device *mdt)
4582 struct seq_server_site *ss = mdt_seq_site(mdt);
4587 /* check if this is adding the first MDC and controller is not yet
4589 OBD_ALLOC_PTR(ss->ss_client_seq);
4590 if (ss->ss_client_seq == NULL)
4593 OBD_ALLOC(prefix, MAX_OBD_NAME + 5);
4594 if (prefix == NULL) {
4595 OBD_FREE_PTR(ss->ss_client_seq);
4596 ss->ss_client_seq = NULL;
4600 /* Note: seq_client_fini will be called in seq_site_fini */
4601 snprintf(prefix, MAX_OBD_NAME + 5, "ctl-%s", mdt_obd_name(mdt));
4602 rc = seq_client_init(ss->ss_client_seq, NULL, LUSTRE_SEQ_METADATA,
4603 prefix, ss->ss_node_id == 0 ? ss->ss_control_seq :
4605 OBD_FREE(prefix, MAX_OBD_NAME + 5);
4607 OBD_FREE_PTR(ss->ss_client_seq);
4608 ss->ss_client_seq = NULL;
4612 rc = seq_server_set_cli(env, ss->ss_server_seq, ss->ss_client_seq);
4617 static int mdt_seq_init(const struct lu_env *env, struct mdt_device *mdt)
4619 struct seq_server_site *ss;
4623 ss = mdt_seq_site(mdt);
4624 /* init sequence controller server(MDT0) */
4625 if (ss->ss_node_id == 0) {
4626 OBD_ALLOC_PTR(ss->ss_control_seq);
4627 if (ss->ss_control_seq == NULL)
4630 rc = seq_server_init(env, ss->ss_control_seq, mdt->mdt_bottom,
4631 mdt_obd_name(mdt), LUSTRE_SEQ_CONTROLLER,
4634 GOTO(out_seq_fini, rc);
4637 /* Init normal sequence server */
4638 OBD_ALLOC_PTR(ss->ss_server_seq);
4639 if (ss->ss_server_seq == NULL)
4640 GOTO(out_seq_fini, rc = -ENOMEM);
4642 rc = seq_server_init(env, ss->ss_server_seq, mdt->mdt_bottom,
4643 mdt_obd_name(mdt), LUSTRE_SEQ_SERVER, ss);
4645 GOTO(out_seq_fini, rc);
4647 /* init seq client for seq server to talk to seq controller(MDT0) */
4648 rc = mdt_seq_init_cli(env, mdt);
4650 GOTO(out_seq_fini, rc);
4652 if (ss->ss_node_id != 0)
4653 /* register controller export through lwp */
4654 rc = mdt_register_seq_exp(mdt);
4659 mdt_seq_fini(env, mdt);
4667 static int mdt_fld_fini(const struct lu_env *env,
4668 struct mdt_device *m)
4670 struct seq_server_site *ss = mdt_seq_site(m);
4673 if (ss && ss->ss_server_fld) {
4674 fld_server_fini(env, ss->ss_server_fld);
4675 OBD_FREE_PTR(ss->ss_server_fld);
4676 ss->ss_server_fld = NULL;
4682 static int mdt_fld_init(const struct lu_env *env,
4684 struct mdt_device *m)
4686 struct seq_server_site *ss;
4690 ss = mdt_seq_site(m);
4692 OBD_ALLOC_PTR(ss->ss_server_fld);
4693 if (ss->ss_server_fld == NULL)
4694 RETURN(rc = -ENOMEM);
4696 rc = fld_server_init(env, ss->ss_server_fld, m->mdt_bottom, uuid,
4699 OBD_FREE_PTR(ss->ss_server_fld);
4700 ss->ss_server_fld = NULL;
4707 static void mdt_stack_pre_fini(const struct lu_env *env,
4708 struct mdt_device *m, struct lu_device *top)
4710 struct lustre_cfg_bufs *bufs;
4711 struct lustre_cfg *lcfg;
4712 struct mdt_thread_info *info;
4717 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
4718 LASSERT(info != NULL);
4720 bufs = &info->mti_u.bufs;
4722 LASSERT(m->mdt_child_exp);
4723 LASSERT(m->mdt_child_exp->exp_obd);
4725 /* process cleanup, pass mdt obd name to get obd umount flags */
4726 /* XXX: this is needed because all layers are referenced by
4727 * objects (some of them are pinned by osd, for example *
4728 * the proper solution should be a model where object used
4729 * by osd only doesn't have mdt/mdd slices -bzzz */
4730 lustre_cfg_bufs_reset(bufs, mdt_obd_name(m));
4731 lustre_cfg_bufs_set_string(bufs, 1, NULL);
4732 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
4735 lustre_cfg_init(lcfg, LCFG_PRE_CLEANUP, bufs);
4737 top->ld_ops->ldo_process_config(env, top, lcfg);
4738 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens));
4742 static void mdt_stack_fini(const struct lu_env *env,
4743 struct mdt_device *m, struct lu_device *top)
4745 struct obd_device *obd = mdt2obd_dev(m);
4746 struct lustre_cfg_bufs *bufs;
4747 struct lustre_cfg *lcfg;
4748 struct mdt_thread_info *info;
4752 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
4753 LASSERT(info != NULL);
4755 lu_dev_del_linkage(top->ld_site, top);
4757 lu_site_purge(env, top->ld_site, -1);
4759 bufs = &info->mti_u.bufs;
4760 /* process cleanup, pass mdt obd name to get obd umount flags */
4761 /* another purpose is to let all layers to release their objects */
4762 lustre_cfg_bufs_reset(bufs, mdt_obd_name(m));
4767 lustre_cfg_bufs_set_string(bufs, 1, flags);
4768 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
4771 lustre_cfg_init(lcfg, LCFG_CLEANUP, bufs);
4774 top->ld_ops->ldo_process_config(env, top, lcfg);
4775 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens));
4777 lu_site_purge(env, top->ld_site, -1);
4779 m->mdt_child = NULL;
4780 m->mdt_bottom = NULL;
4782 obd_disconnect(m->mdt_child_exp);
4783 m->mdt_child_exp = NULL;
4785 obd_disconnect(m->mdt_bottom_exp);
4786 m->mdt_child_exp = NULL;
4789 static int mdt_connect_to_next(const struct lu_env *env, struct mdt_device *m,
4790 const char *next, struct obd_export **exp)
4792 struct obd_connect_data *data = NULL;
4793 struct obd_device *obd;
4797 OBD_ALLOC_PTR(data);
4799 GOTO(out, rc = -ENOMEM);
4801 obd = class_name2obd(next);
4803 CERROR("%s: can't locate next device: %s\n",
4804 mdt_obd_name(m), next);
4805 GOTO(out, rc = -ENOTCONN);
4808 data->ocd_connect_flags = OBD_CONNECT_VERSION;
4809 data->ocd_version = LUSTRE_VERSION_CODE;
4811 rc = obd_connect(NULL, exp, obd, &obd->obd_uuid, data, NULL);
4813 CERROR("%s: cannot connect to next dev %s (%d)\n",
4814 mdt_obd_name(m), next, rc);
4824 static int mdt_stack_init(const struct lu_env *env, struct mdt_device *mdt,
4825 struct lustre_cfg *cfg)
4827 char *dev = lustre_cfg_string(cfg, 0);
4828 int rc, name_size, uuid_size;
4829 char *name, *uuid, *p;
4830 struct lustre_cfg_bufs *bufs;
4831 struct lustre_cfg *lcfg;
4832 struct obd_device *obd;
4833 struct lustre_profile *lprof;
4834 struct lu_site *site;
4837 /* in 1.8 we had the only device in the stack - MDS.
4838 * 2.0 introduces MDT, MDD, OSD; MDT starts others internally.
4839 * in 2.3 OSD is instantiated by obd_mount.c, so we need
4840 * to generate names and setup MDT, MDD. MDT will be using
4841 * generated name to connect to MDD. for MDD the next device
4842 * will be LOD with name taken from so called "profile" which
4843 * is generated by mount_option line
4845 * 1.8 MGS generates config. commands like this:
4846 * #06 (104)mount_option 0: 1:lustre-MDT0000 2:lustre-mdtlov
4847 * #08 (120)setup 0:lustre-MDT0000 1:dev 2:type 3:lustre-MDT0000
4848 * 2.0 MGS generates config. commands like this:
4849 * #07 (112)mount_option 0: 1:lustre-MDT0000 2:lustre-MDT0000-mdtlov
4850 * #08 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
4851 * 3:lustre-MDT0000-mdtlov 4:f
4853 * we generate MDD name from MDT one, just replacing T with D
4855 * after all the preparations, the logical equivalent will be
4856 * #01 (160)setup 0:lustre-MDD0000 1:lustre-MDD0000_UUID 2:0
4857 * 3:lustre-MDT0000-mdtlov 4:f
4858 * #02 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
4859 * 3:lustre-MDD0000 4:f
4861 * notice we build the stack from down to top: MDD first, then MDT */
4863 name_size = MAX_OBD_NAME;
4864 uuid_size = MAX_OBD_NAME;
4866 OBD_ALLOC(name, name_size);
4867 OBD_ALLOC(uuid, uuid_size);
4868 if (name == NULL || uuid == NULL)
4869 GOTO(cleanup_mem, rc = -ENOMEM);
4871 OBD_ALLOC_PTR(bufs);
4873 GOTO(cleanup_mem, rc = -ENOMEM);
4876 p = strstr(name, "-MDT");
4878 GOTO(free_bufs, rc = -ENOMEM);
4881 snprintf(uuid, MAX_OBD_NAME, "%s_UUID", name);
4883 lprof = class_get_profile(lustre_cfg_string(cfg, 0));
4884 if (lprof == NULL || lprof->lp_dt == NULL) {
4885 CERROR("can't find the profile: %s\n",
4886 lustre_cfg_string(cfg, 0));
4887 GOTO(free_bufs, rc = -EINVAL);
4890 lustre_cfg_bufs_reset(bufs, name);
4891 lustre_cfg_bufs_set_string(bufs, 1, LUSTRE_MDD_NAME);
4892 lustre_cfg_bufs_set_string(bufs, 2, uuid);
4893 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
4895 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
4897 GOTO(put_profile, rc = -ENOMEM);
4898 lustre_cfg_init(lcfg, LCFG_ATTACH, bufs);
4900 rc = class_attach(lcfg);
4902 GOTO(lcfg_cleanup, rc);
4904 obd = class_name2obd(name);
4906 CERROR("Can not find obd %s (%s in config)\n",
4907 MDD_OBD_NAME, lustre_cfg_string(cfg, 0));
4908 GOTO(lcfg_cleanup, rc = -EINVAL);
4911 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens));
4913 lustre_cfg_bufs_reset(bufs, name);
4914 lustre_cfg_bufs_set_string(bufs, 1, uuid);
4915 lustre_cfg_bufs_set_string(bufs, 2, dev);
4916 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
4918 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
4920 GOTO(class_detach, rc = -ENOMEM);
4921 lustre_cfg_init(lcfg, LCFG_SETUP, bufs);
4923 rc = class_setup(obd, lcfg);
4925 GOTO(class_detach, rc);
4927 /* connect to MDD we just setup */
4928 rc = mdt_connect_to_next(env, mdt, name, &mdt->mdt_child_exp);
4930 GOTO(class_detach, rc);
4932 site = mdt->mdt_child_exp->exp_obd->obd_lu_dev->ld_site;
4934 LASSERT(mdt_lu_site(mdt) == NULL);
4935 mdt->mdt_lu_dev.ld_site = site;
4936 site->ls_top_dev = &mdt->mdt_lu_dev;
4937 mdt->mdt_child = lu2md_dev(mdt->mdt_child_exp->exp_obd->obd_lu_dev);
4939 /* now connect to bottom OSD */
4940 snprintf(name, MAX_OBD_NAME, "%s-osd", dev);
4941 rc = mdt_connect_to_next(env, mdt, name, &mdt->mdt_bottom_exp);
4943 GOTO(class_detach, rc);
4945 lu2dt_dev(mdt->mdt_bottom_exp->exp_obd->obd_lu_dev);
4947 rc = lu_env_refill((struct lu_env *)env);
4949 CERROR("Failure to refill session: '%d'\n", rc);
4951 lu_dev_add_linkage(site, &mdt->mdt_lu_dev);
4956 class_detach(obd, lcfg);
4958 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens));
4960 class_put_profile(lprof);
4965 OBD_FREE(name, name_size);
4967 OBD_FREE(uuid, uuid_size);
4971 /* setup quota master target on MDT0 */
4972 static int mdt_quota_init(const struct lu_env *env, struct mdt_device *mdt,
4973 struct lustre_cfg *cfg)
4975 struct obd_device *obd;
4976 char *dev = lustre_cfg_string(cfg, 0);
4977 char *qmtname, *uuid, *p;
4978 struct lustre_cfg_bufs *bufs;
4979 struct lustre_cfg *lcfg;
4980 struct lustre_profile *lprof;
4981 struct obd_connect_data *data;
4985 LASSERT(mdt->mdt_qmt_exp == NULL);
4986 LASSERT(mdt->mdt_qmt_dev == NULL);
4988 /* quota master is on MDT0 only for now */
4989 if (mdt->mdt_seq_site.ss_node_id != 0)
4992 /* MGS generates config commands which look as follows:
4993 * #01 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
4994 * 3:lustre-MDT0000-mdtlov 4:f
4996 * We generate the QMT name from the MDT one, just replacing MD with QM
4997 * after all the preparations, the logical equivalent will be:
4998 * #01 (160)setup 0:lustre-QMT0000 1:lustre-QMT0000_UUID 2:0
4999 * 3:lustre-MDT0000-osd 4:f */
5000 OBD_ALLOC(qmtname, MAX_OBD_NAME);
5001 OBD_ALLOC(uuid, UUID_MAX);
5002 OBD_ALLOC_PTR(bufs);
5003 OBD_ALLOC_PTR(data);
5004 if (qmtname == NULL || uuid == NULL || bufs == NULL || data == NULL)
5005 GOTO(cleanup_mem, rc = -ENOMEM);
5007 strcpy(qmtname, dev);
5008 p = strstr(qmtname, "-MDT");
5010 GOTO(cleanup_mem, rc = -ENOMEM);
5011 /* replace MD with QM */
5015 snprintf(uuid, UUID_MAX, "%s_UUID", qmtname);
5017 lprof = class_get_profile(lustre_cfg_string(cfg, 0));
5018 if (lprof == NULL || lprof->lp_dt == NULL) {
5019 CERROR("can't find profile for %s\n",
5020 lustre_cfg_string(cfg, 0));
5021 GOTO(cleanup_mem, rc = -EINVAL);
5024 lustre_cfg_bufs_reset(bufs, qmtname);
5025 lustre_cfg_bufs_set_string(bufs, 1, LUSTRE_QMT_NAME);
5026 lustre_cfg_bufs_set_string(bufs, 2, uuid);
5027 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
5029 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
5031 GOTO(put_profile, rc = -ENOMEM);
5032 lustre_cfg_init(lcfg, LCFG_ATTACH, bufs);
5034 rc = class_attach(lcfg);
5036 GOTO(lcfg_cleanup, rc);
5038 obd = class_name2obd(qmtname);
5040 CERROR("Can not find obd %s (%s in config)\n", qmtname,
5041 lustre_cfg_string(cfg, 0));
5042 GOTO(lcfg_cleanup, rc = -EINVAL);
5045 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens));
5047 lustre_cfg_bufs_reset(bufs, qmtname);
5048 lustre_cfg_bufs_set_string(bufs, 1, uuid);
5049 lustre_cfg_bufs_set_string(bufs, 2, dev);
5051 /* for quota, the next device should be the OSD device */
5052 lustre_cfg_bufs_set_string(bufs, 3,
5053 mdt->mdt_bottom->dd_lu_dev.ld_obd->obd_name);
5055 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
5057 GOTO(class_detach, rc = -ENOMEM);
5058 lustre_cfg_init(lcfg, LCFG_SETUP, bufs);
5060 rc = class_setup(obd, lcfg);
5062 GOTO(class_detach, rc);
5064 mdt->mdt_qmt_dev = obd->obd_lu_dev;
5066 /* configure local quota objects */
5067 if (OBD_FAIL_CHECK(OBD_FAIL_QUOTA_INIT))
5070 rc = mdt->mdt_qmt_dev->ld_ops->ldo_prepare(env,
5074 GOTO(class_cleanup, rc);
5076 /* connect to quota master target */
5077 data->ocd_connect_flags = OBD_CONNECT_VERSION;
5078 data->ocd_version = LUSTRE_VERSION_CODE;
5079 rc = obd_connect(NULL, &mdt->mdt_qmt_exp, obd, &obd->obd_uuid,
5082 CERROR("cannot connect to quota master device %s (%d)\n",
5084 GOTO(class_cleanup, rc);
5090 class_manual_cleanup(obd);
5091 mdt->mdt_qmt_dev = NULL;
5092 GOTO(lcfg_cleanup, rc);
5096 class_detach(obd, lcfg);
5098 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens));
5100 class_put_profile(lprof);
5105 OBD_FREE(qmtname, MAX_OBD_NAME);
5107 OBD_FREE(uuid, UUID_MAX);
5113 /* Shutdown quota master target associated with mdt */
5114 static void mdt_quota_fini(const struct lu_env *env, struct mdt_device *mdt)
5118 if (mdt->mdt_qmt_exp == NULL)
5120 LASSERT(mdt->mdt_qmt_dev != NULL);
5122 /* the qmt automatically shuts down when the mdt disconnects */
5123 obd_disconnect(mdt->mdt_qmt_exp);
5124 mdt->mdt_qmt_exp = NULL;
5125 mdt->mdt_qmt_dev = NULL;
5129 /* mdt_getxattr() is used from mdt_intent_getxattr(), use this wrapper
5130 * for now. This will be removed along with converting rest of MDT code
5131 * to use tgt_session_info */
5132 static int mdt_tgt_getxattr(struct tgt_session_info *tsi)
5134 struct mdt_thread_info *info = tsi2mdt_info(tsi);
5137 if (unlikely(info->mti_object == NULL))
5140 rc = mdt_getxattr(info);
5142 mdt_thread_info_fini(info);
5146 #define OBD_FAIL_OST_READ_NET OBD_FAIL_OST_BRW_NET
5147 #define OBD_FAIL_OST_WRITE_NET OBD_FAIL_OST_BRW_NET
5148 #define OST_BRW_READ OST_READ
5149 #define OST_BRW_WRITE OST_WRITE
5151 static struct tgt_handler mdt_tgt_handlers[] = {
5152 TGT_RPC_HANDLER(MDS_FIRST_OPC,
5153 0, MDS_CONNECT, mdt_tgt_connect,
5154 &RQF_CONNECT, LUSTRE_OBD_VERSION),
5155 TGT_RPC_HANDLER(MDS_FIRST_OPC,
5156 0, MDS_DISCONNECT, tgt_disconnect,
5157 &RQF_MDS_DISCONNECT, LUSTRE_OBD_VERSION),
5158 TGT_RPC_HANDLER(MDS_FIRST_OPC,
5159 HAS_REPLY, MDS_SET_INFO, mdt_set_info,
5160 &RQF_OBD_SET_INFO, LUSTRE_MDS_VERSION),
5161 TGT_MDT_HDL(0, MDS_GET_INFO, mdt_get_info),
5162 TGT_MDT_HDL(HAS_REPLY, MDS_GET_ROOT, mdt_get_root),
5163 TGT_MDT_HDL(HAS_BODY, MDS_GETATTR, mdt_getattr),
5164 TGT_MDT_HDL(HAS_BODY | HAS_REPLY, MDS_GETATTR_NAME,
5166 TGT_MDT_HDL(HAS_BODY, MDS_GETXATTR, mdt_tgt_getxattr),
5167 TGT_MDT_HDL(HAS_REPLY, MDS_STATFS, mdt_statfs),
5168 TGT_MDT_HDL(IS_MUTABLE, MDS_REINT, mdt_reint),
5169 TGT_MDT_HDL(HAS_BODY, MDS_CLOSE, mdt_close),
5170 TGT_MDT_HDL(HAS_BODY | HAS_REPLY, MDS_READPAGE, mdt_readpage),
5171 TGT_MDT_HDL(HAS_BODY | HAS_REPLY, MDS_SYNC, mdt_sync),
5172 TGT_MDT_HDL(0, MDS_QUOTACTL, mdt_quotactl),
5173 TGT_MDT_HDL(HAS_BODY | HAS_REPLY | IS_MUTABLE, MDS_HSM_PROGRESS,
5175 TGT_MDT_HDL(HAS_BODY | HAS_REPLY | IS_MUTABLE, MDS_HSM_CT_REGISTER,
5176 mdt_hsm_ct_register),
5177 TGT_MDT_HDL(HAS_BODY | HAS_REPLY | IS_MUTABLE, MDS_HSM_CT_UNREGISTER,
5178 mdt_hsm_ct_unregister),
5179 TGT_MDT_HDL(HAS_BODY | HAS_REPLY, MDS_HSM_STATE_GET,
5181 TGT_MDT_HDL(HAS_BODY | HAS_REPLY | IS_MUTABLE, MDS_HSM_STATE_SET,
5183 TGT_MDT_HDL(HAS_BODY | HAS_REPLY, MDS_HSM_ACTION, mdt_hsm_action),
5184 TGT_MDT_HDL(HAS_BODY | HAS_REPLY, MDS_HSM_REQUEST,
5186 TGT_MDT_HDL(HAS_KEY | HAS_BODY | HAS_REPLY | IS_MUTABLE,
5189 TGT_MDT_HDL(IS_MUTABLE, MDS_RMFID, mdt_rmfid),
5192 static struct tgt_handler mdt_io_ops[] = {
5193 TGT_OST_HDL_HP(HAS_BODY | HAS_REPLY, OST_BRW_READ, tgt_brw_read,
5195 TGT_OST_HDL_HP(HAS_BODY | IS_MUTABLE, OST_BRW_WRITE, tgt_brw_write,
5197 TGT_OST_HDL_HP(HAS_BODY | HAS_REPLY | IS_MUTABLE,
5198 OST_PUNCH, mdt_punch_hdl,
5200 TGT_OST_HDL(HAS_BODY | HAS_REPLY, OST_SYNC, mdt_data_sync),
5203 static struct tgt_handler mdt_sec_ctx_ops[] = {
5204 TGT_SEC_HDL_VAR(0, SEC_CTX_INIT, mdt_sec_ctx_handle),
5205 TGT_SEC_HDL_VAR(0, SEC_CTX_INIT_CONT,mdt_sec_ctx_handle),
5206 TGT_SEC_HDL_VAR(0, SEC_CTX_FINI, mdt_sec_ctx_handle)
5209 static struct tgt_handler mdt_quota_ops[] = {
5210 TGT_QUOTA_HDL(HAS_REPLY, QUOTA_DQACQ, mdt_quota_dqacq),
5213 static struct tgt_opc_slice mdt_common_slice[] = {
5215 .tos_opc_start = MDS_FIRST_OPC,
5216 .tos_opc_end = MDS_LAST_OPC,
5217 .tos_hs = mdt_tgt_handlers
5220 .tos_opc_start = OBD_FIRST_OPC,
5221 .tos_opc_end = OBD_LAST_OPC,
5222 .tos_hs = tgt_obd_handlers
5225 .tos_opc_start = LDLM_FIRST_OPC,
5226 .tos_opc_end = LDLM_LAST_OPC,
5227 .tos_hs = tgt_dlm_handlers
5230 .tos_opc_start = SEC_FIRST_OPC,
5231 .tos_opc_end = SEC_LAST_OPC,
5232 .tos_hs = mdt_sec_ctx_ops
5235 .tos_opc_start = OUT_UPDATE_FIRST_OPC,
5236 .tos_opc_end = OUT_UPDATE_LAST_OPC,
5237 .tos_hs = tgt_out_handlers
5240 .tos_opc_start = FLD_FIRST_OPC,
5241 .tos_opc_end = FLD_LAST_OPC,
5242 .tos_hs = fld_handlers
5245 .tos_opc_start = SEQ_FIRST_OPC,
5246 .tos_opc_end = SEQ_LAST_OPC,
5247 .tos_hs = seq_handlers
5250 .tos_opc_start = QUOTA_DQACQ,
5251 .tos_opc_end = QUOTA_LAST_OPC,
5252 .tos_hs = mdt_quota_ops
5255 .tos_opc_start = LLOG_FIRST_OPC,
5256 .tos_opc_end = LLOG_LAST_OPC,
5257 .tos_hs = tgt_llog_handlers
5260 .tos_opc_start = LFSCK_FIRST_OPC,
5261 .tos_opc_end = LFSCK_LAST_OPC,
5262 .tos_hs = tgt_lfsck_handlers
5265 .tos_opc_start = OST_FIRST_OPC,
5266 .tos_opc_end = OST_LAST_OPC,
5267 .tos_hs = mdt_io_ops
5274 static void mdt_fini(const struct lu_env *env, struct mdt_device *m)
5276 struct md_device *next = m->mdt_child;
5277 struct lu_device *d = &m->mdt_lu_dev;
5278 struct obd_device *obd = mdt2obd_dev(m);
5279 struct lfsck_stop stop;
5282 stop.ls_status = LS_PAUSED;
5284 next->md_ops->mdo_iocontrol(env, next, OBD_IOC_STOP_LFSCK, 0, &stop);
5286 mdt_stack_pre_fini(env, m, md2lu_dev(m->mdt_child));
5287 ping_evictor_stop();
5289 /* Remove the HSM /proc entry so the coordinator cannot be
5290 * restarted by a user while it's shutting down.
5292 mdt_hsm_cdt_stop(m);
5294 mdt_llog_ctxt_unclone(env, m, LLOG_AGENT_ORIG_CTXT);
5295 mdt_llog_ctxt_unclone(env, m, LLOG_CHANGELOG_ORIG_CTXT);
5297 if (m->mdt_namespace != NULL)
5298 ldlm_namespace_free_prior(m->mdt_namespace, NULL,
5299 d->ld_obd->obd_force);
5301 obd_exports_barrier(obd);
5302 obd_zombie_barrier();
5304 mdt_quota_fini(env, m);
5306 cfs_free_nidlist(&m->mdt_squash.rsi_nosquash_nids);
5308 /* Calling the cleanup functions in the same order as in the mdt_init0
5311 mdt_tunables_fini(m);
5313 target_recovery_fini(obd);
5314 upcall_cache_cleanup(m->mdt_identity_cache);
5315 m->mdt_identity_cache = NULL;
5317 mdt_fs_cleanup(env, m);
5319 tgt_fini(env, &m->mdt_lut);
5321 mdt_hsm_cdt_fini(m);
5323 if (m->mdt_los != NULL) {
5324 local_oid_storage_fini(env, m->mdt_los);
5328 if (m->mdt_namespace != NULL) {
5329 ldlm_namespace_free_post(m->mdt_namespace);
5330 d->ld_obd->obd_namespace = m->mdt_namespace = NULL;
5333 if (m->mdt_md_root != NULL) {
5334 mdt_object_put(env, m->mdt_md_root);
5335 m->mdt_md_root = NULL;
5338 mdt_seq_fini(env, m);
5340 mdt_fld_fini(env, m);
5345 mdt_stack_fini(env, m, md2lu_dev(m->mdt_child));
5347 LASSERT(atomic_read(&d->ld_ref) == 0);
5349 server_put_mount(mdt_obd_name(m), true);
5354 static int mdt_postrecov(const struct lu_env *, struct mdt_device *);
5356 static int mdt_init0(const struct lu_env *env, struct mdt_device *m,
5357 struct lu_device_type *ldt, struct lustre_cfg *cfg)
5359 const struct dt_device_param *dt_conf;
5360 struct mdt_thread_info *info;
5361 struct obd_device *obd;
5362 const char *dev = lustre_cfg_string(cfg, 0);
5363 const char *num = lustre_cfg_string(cfg, 2);
5364 struct tg_grants_data *tgd = &m->mdt_lut.lut_tgd;
5365 struct lustre_mount_info *lmi = NULL;
5366 struct lustre_sb_info *lsi;
5368 struct seq_server_site *ss_site;
5369 const char *identity_upcall = "NONE";
5370 struct md_device *next;
5377 lu_device_init(&m->mdt_lu_dev, ldt);
5379 * Environment (env) might be missing mdt_thread_key values at that
5380 * point, if device is allocated when mdt_thread_key is in QUIESCENT
5383 * Usually device allocation path doesn't use module key values, but
5384 * mdt has to do a lot of work here, so allocate key value.
5386 rc = lu_env_refill((struct lu_env *)env);
5390 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
5391 LASSERT(info != NULL);
5393 obd = class_name2obd(dev);
5394 LASSERT(obd != NULL);
5396 m->mdt_max_mdsize = MAX_MD_SIZE_OLD;
5397 m->mdt_opts.mo_evict_tgt_nids = 1;
5398 m->mdt_opts.mo_cos = MDT_COS_DEFAULT;
5400 lmi = server_get_mount(dev);
5402 CERROR("Cannot get mount info for %s!\n", dev);
5405 lsi = s2lsi(lmi->lmi_sb);
5406 /* CMD is supported only in IAM mode */
5408 node_id = simple_strtol(num, NULL, 10);
5409 obd->u.obt.obt_magic = OBT_MAGIC;
5410 if (lsi->lsi_lmd != NULL &&
5411 lsi->lsi_lmd->lmd_flags & LMD_FLG_SKIP_LFSCK)
5412 m->mdt_skip_lfsck = 1;
5415 /* DoM files get IO lock at open by default */
5416 m->mdt_opts.mo_dom_lock = ALWAYS_DOM_LOCK_ON_OPEN;
5417 /* DoM files are read at open and data is packed in the reply */
5418 m->mdt_opts.mo_dom_read_open = 1;
5420 m->mdt_squash.rsi_uid = 0;
5421 m->mdt_squash.rsi_gid = 0;
5422 INIT_LIST_HEAD(&m->mdt_squash.rsi_nosquash_nids);
5423 spin_lock_init(&m->mdt_squash.rsi_lock);
5424 spin_lock_init(&m->mdt_lock);
5425 m->mdt_enable_remote_dir = 1;
5426 m->mdt_enable_striped_dir = 1;
5427 m->mdt_enable_dir_migration = 1;
5428 m->mdt_enable_remote_dir_gid = 0;
5429 m->mdt_enable_remote_rename = 1;
5431 atomic_set(&m->mdt_mds_mds_conns, 0);
5432 atomic_set(&m->mdt_async_commit_count, 0);
5434 m->mdt_lu_dev.ld_ops = &mdt_lu_ops;
5435 m->mdt_lu_dev.ld_obd = obd;
5436 /* Set this lu_device to obd for error handling purposes. */
5437 obd->obd_lu_dev = &m->mdt_lu_dev;
5439 /* init the stack */
5440 rc = mdt_stack_init((struct lu_env *)env, m, cfg);
5442 CERROR("%s: Can't init device stack, rc %d\n",
5443 mdt_obd_name(m), rc);
5448 ss_site = mdt_seq_site(m);
5449 s->ld_seq_site = ss_site;
5452 /* set server index */
5453 ss_site->ss_node_id = node_id;
5455 /* failover is the default
5456 * FIXME: we do not failout mds0/mgs, which may cause some problems.
5457 * assumed whose ss_node_id == 0 XXX
5459 obd->obd_replayable = 1;
5460 /* No connection accepted until configurations will finish */
5461 obd->obd_no_conn = 1;
5463 if (cfg->lcfg_bufcount > 4 && LUSTRE_CFG_BUFLEN(cfg, 4) > 0) {
5464 char *str = lustre_cfg_string(cfg, 4);
5465 if (strchr(str, 'n')) {
5466 CWARN("%s: recovery disabled\n", mdt_obd_name(m));
5467 obd->obd_replayable = 0;
5471 rc = mdt_fld_init(env, mdt_obd_name(m), m);
5473 GOTO(err_fini_stack, rc);
5475 rc = mdt_seq_init(env, m);
5477 GOTO(err_fini_fld, rc);
5479 snprintf(info->mti_u.ns_name, sizeof(info->mti_u.ns_name), "%s-%s",
5480 LUSTRE_MDT_NAME, obd->obd_uuid.uuid);
5481 m->mdt_namespace = ldlm_namespace_new(obd, info->mti_u.ns_name,
5482 LDLM_NAMESPACE_SERVER,
5483 LDLM_NAMESPACE_GREEDY,
5485 if (m->mdt_namespace == NULL)
5486 GOTO(err_fini_seq, rc = -ENOMEM);
5488 m->mdt_namespace->ns_lvbp = m;
5489 m->mdt_namespace->ns_lvbo = &mdt_lvbo;
5491 ldlm_register_intent(m->mdt_namespace, mdt_intent_policy);
5492 /* set obd_namespace for compatibility with old code */
5493 obd->obd_namespace = m->mdt_namespace;
5495 rc = tgt_init(env, &m->mdt_lut, obd, m->mdt_bottom, mdt_common_slice,
5496 OBD_FAIL_MDS_ALL_REQUEST_NET,
5497 OBD_FAIL_MDS_ALL_REPLY_NET);
5499 GOTO(err_free_ns, rc);
5501 /* Amount of available space excluded from granting and reserved
5502 * for metadata. It is in percentage and 50% is default value. */
5503 tgd->tgd_reserved_pcnt = 50;
5505 if (ONE_MB_BRW_SIZE < (1U << tgd->tgd_blockbits))
5506 m->mdt_brw_size = 1U << tgd->tgd_blockbits;
5508 m->mdt_brw_size = ONE_MB_BRW_SIZE;
5510 rc = mdt_fs_setup(env, m, obd, lsi);
5514 fid.f_seq = FID_SEQ_LOCAL_NAME;
5517 rc = local_oid_storage_init(env, m->mdt_bottom, &fid, &m->mdt_los);
5519 GOTO(err_fs_cleanup, rc);
5521 rc = mdt_hsm_cdt_init(m);
5523 CERROR("%s: error initializing coordinator, rc %d\n",
5524 mdt_obd_name(m), rc);
5525 GOTO(err_los_fini, rc);
5528 tgt_adapt_sptlrpc_conf(&m->mdt_lut);
5530 next = m->mdt_child;
5531 dt_conf = next->md_ops->mdo_dtconf_get(env, next);
5533 mntopts = dt_conf->ddp_mntopts;
5535 if (mntopts & MNTOPT_USERXATTR)
5536 m->mdt_opts.mo_user_xattr = 1;
5538 m->mdt_opts.mo_user_xattr = 0;
5540 m->mdt_max_ea_size = dt_conf->ddp_max_ea_size;
5542 if (mntopts & MNTOPT_ACL)
5543 m->mdt_opts.mo_acl = 1;
5545 m->mdt_opts.mo_acl = 0;
5547 /* XXX: to support suppgid for ACL, we enable identity_upcall
5548 * by default, otherwise, maybe got unexpected -EACCESS. */
5549 if (m->mdt_opts.mo_acl)
5550 identity_upcall = MDT_IDENTITY_UPCALL_PATH;
5552 m->mdt_identity_cache = upcall_cache_init(mdt_obd_name(m),
5554 &mdt_identity_upcall_cache_ops);
5555 if (IS_ERR(m->mdt_identity_cache)) {
5556 rc = PTR_ERR(m->mdt_identity_cache);
5557 m->mdt_identity_cache = NULL;
5558 GOTO(err_free_hsm, rc);
5561 rc = mdt_tunables_init(m, dev);
5563 CERROR("Can't init MDT lprocfs, rc %d\n", rc);
5564 GOTO(err_recovery, rc);
5567 rc = mdt_quota_init(env, m, cfg);
5569 GOTO(err_procfs, rc);
5571 m->mdt_ldlm_client = &mdt2obd_dev(m)->obd_ldlm_client;
5572 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
5573 "mdt_ldlm_client", m->mdt_ldlm_client);
5575 ping_evictor_start();
5577 /* recovery will be started upon mdt_prepare()
5578 * when the whole stack is complete and ready
5579 * to serve the requests */
5581 /* Reduce the initial timeout on an MDS because it doesn't need such
5582 * a long timeout as an OST does. Adaptive timeouts will adjust this
5583 * value appropriately. */
5584 if (ldlm_timeout == LDLM_TIMEOUT_DEFAULT)
5585 ldlm_timeout = MDS_LDLM_TIMEOUT_DEFAULT;
5589 mdt_tunables_fini(m);
5591 upcall_cache_cleanup(m->mdt_identity_cache);
5592 m->mdt_identity_cache = NULL;
5594 mdt_hsm_cdt_fini(m);
5596 local_oid_storage_fini(env, m->mdt_los);
5599 mdt_fs_cleanup(env, m);
5601 /* keep recoverable clients */
5603 target_recovery_fini(obd);
5604 obd_exports_barrier(obd);
5605 obd_zombie_barrier();
5606 tgt_fini(env, &m->mdt_lut);
5608 ldlm_namespace_free(m->mdt_namespace, NULL, 0);
5609 obd->obd_namespace = m->mdt_namespace = NULL;
5611 mdt_seq_fini(env, m);
5613 mdt_fld_fini(env, m);
5615 mdt_stack_fini(env, m, md2lu_dev(m->mdt_child));
5618 server_put_mount(dev, true);
5622 /* For interoperability, the left element is old parameter, the right one
5623 * is the new version of the parameter, if some parameter is deprecated,
5624 * the new version should be set as NULL. */
5625 static struct cfg_interop_param mdt_interop_param[] = {
5626 { "mdt.group_upcall", NULL },
5627 { "mdt.quota_type", NULL },
5628 { "mdd.quota_type", NULL },
5629 { "mdt.som", NULL },
5630 { "mdt.rootsquash", "mdt.root_squash" },
5631 { "mdt.nosquash_nid", "mdt.nosquash_nids" },
5635 /* used by MGS to process specific configurations */
5636 static int mdt_process_config(const struct lu_env *env,
5637 struct lu_device *d, struct lustre_cfg *cfg)
5639 struct mdt_device *m = mdt_dev(d);
5640 struct md_device *md_next = m->mdt_child;
5641 struct lu_device *next = md2lu_dev(md_next);
5645 switch (cfg->lcfg_command) {
5647 struct obd_device *obd = d->ld_obd;
5648 /* For interoperability */
5649 struct cfg_interop_param *ptr = NULL;
5650 struct lustre_cfg *old_cfg = NULL;
5654 param = lustre_cfg_string(cfg, 1);
5655 if (param == NULL) {
5656 CERROR("param is empty\n");
5661 ptr = class_find_old_param(param, mdt_interop_param);
5663 if (ptr->new_param == NULL) {
5665 CWARN("For interoperability, skip this %s."
5666 " It is obsolete.\n", ptr->old_param);
5670 CWARN("Found old param %s, changed it to %s.\n",
5671 ptr->old_param, ptr->new_param);
5674 cfg = lustre_cfg_rename(old_cfg, ptr->new_param);
5681 count = class_modify_config(cfg, PARAM_MDT,
5682 &obd->obd_kset.kobj);
5684 struct coordinator *cdt = &m->mdt_coordinator;
5686 /* is it an HSM var ? */
5687 count = class_modify_config(cfg, PARAM_HSM,
5688 &cdt->cdt_hsm_kobj);
5690 /* we don't understand; pass it on */
5691 rc = next->ld_ops->ldo_process_config(env, next,
5694 rc = count > 0 ? 0 : count;
5696 rc = count > 0 ? 0 : count;
5700 OBD_FREE(cfg, lustre_cfg_len(cfg->lcfg_bufcount,
5701 cfg->lcfg_buflens));
5705 /* others are passed further */
5706 rc = next->ld_ops->ldo_process_config(env, next, cfg);
5712 static struct lu_object *mdt_object_alloc(const struct lu_env *env,
5713 const struct lu_object_header *hdr,
5714 struct lu_device *d)
5716 struct mdt_object *mo;
5720 OBD_SLAB_ALLOC_PTR_GFP(mo, mdt_object_kmem, GFP_NOFS);
5722 struct lu_object *o;
5723 struct lu_object_header *h;
5726 h = &mo->mot_header;
5727 lu_object_header_init(h);
5728 lu_object_init(o, h, d);
5729 lu_object_add_top(h, o);
5730 o->lo_ops = &mdt_obj_ops;
5731 spin_lock_init(&mo->mot_write_lock);
5732 mutex_init(&mo->mot_som_mutex);
5733 mutex_init(&mo->mot_lov_mutex);
5734 init_rwsem(&mo->mot_dom_sem);
5735 init_rwsem(&mo->mot_open_sem);
5736 atomic_set(&mo->mot_open_count, 0);
5742 static int mdt_object_init(const struct lu_env *env, struct lu_object *o,
5743 const struct lu_object_conf *unused)
5745 struct mdt_device *d = mdt_dev(o->lo_dev);
5746 struct lu_device *under;
5747 struct lu_object *below;
5751 CDEBUG(D_INFO, "object init, fid = "DFID"\n",
5752 PFID(lu_object_fid(o)));
5754 under = &d->mdt_child->md_lu_dev;
5755 below = under->ld_ops->ldo_object_alloc(env, o->lo_header, under);
5756 if (below != NULL) {
5757 lu_object_add(o, below);
5764 static void mdt_object_free(const struct lu_env *env, struct lu_object *o)
5766 struct mdt_object *mo = mdt_obj(o);
5767 struct lu_object_header *h;
5771 CDEBUG(D_INFO, "object free, fid = "DFID"\n",
5772 PFID(lu_object_fid(o)));
5774 LASSERT(atomic_read(&mo->mot_open_count) == 0);
5775 LASSERT(atomic_read(&mo->mot_lease_count) == 0);
5778 lu_object_header_fini(h);
5779 OBD_SLAB_FREE_PTR(mo, mdt_object_kmem);
5784 static int mdt_object_print(const struct lu_env *env, void *cookie,
5785 lu_printer_t p, const struct lu_object *o)
5787 struct mdt_object *mdto = mdt_obj((struct lu_object *)o);
5789 return (*p)(env, cookie,
5790 LUSTRE_MDT_NAME"-object@%p(%s %s, writecount=%d)",
5791 mdto, mdto->mot_lov_created ? "lov_created" : "",
5792 mdto->mot_cache_attr ? "cache_attr" : "",
5793 mdto->mot_write_count);
5796 static int mdt_prepare(const struct lu_env *env,
5797 struct lu_device *pdev,
5798 struct lu_device *cdev)
5800 struct mdt_device *mdt = mdt_dev(cdev);
5801 struct lu_device *next = &mdt->mdt_child->md_lu_dev;
5802 struct obd_device *obd = cdev->ld_obd;
5809 rc = next->ld_ops->ldo_prepare(env, cdev, next);
5813 rc = mdt_llog_ctxt_clone(env, mdt, LLOG_CHANGELOG_ORIG_CTXT);
5817 rc = mdt_llog_ctxt_clone(env, mdt, LLOG_AGENT_ORIG_CTXT);
5821 rc = lfsck_register_namespace(env, mdt->mdt_bottom, mdt->mdt_namespace);
5822 /* The LFSCK instance is registered just now, so it must be there when
5823 * register the namespace to such instance. */
5824 LASSERTF(rc == 0, "register namespace failed: rc = %d\n", rc);
5826 if (mdt->mdt_seq_site.ss_node_id == 0) {
5827 rc = mdt->mdt_child->md_ops->mdo_root_get(env, mdt->mdt_child,
5828 &mdt->mdt_md_root_fid);
5833 LASSERT(!test_bit(MDT_FL_CFGLOG, &mdt->mdt_state));
5835 target_recovery_init(&mdt->mdt_lut, tgt_request_handle);
5836 set_bit(MDT_FL_CFGLOG, &mdt->mdt_state);
5837 LASSERT(obd->obd_no_conn);
5838 spin_lock(&obd->obd_dev_lock);
5839 obd->obd_no_conn = 0;
5840 spin_unlock(&obd->obd_dev_lock);
5842 if (obd->obd_recovering == 0)
5843 mdt_postrecov(env, mdt);
5848 const struct lu_device_operations mdt_lu_ops = {
5849 .ldo_object_alloc = mdt_object_alloc,
5850 .ldo_process_config = mdt_process_config,
5851 .ldo_prepare = mdt_prepare,
5854 static const struct lu_object_operations mdt_obj_ops = {
5855 .loo_object_init = mdt_object_init,
5856 .loo_object_free = mdt_object_free,
5857 .loo_object_print = mdt_object_print
5860 static int mdt_obd_set_info_async(const struct lu_env *env,
5861 struct obd_export *exp,
5862 __u32 keylen, void *key,
5863 __u32 vallen, void *val,
5864 struct ptlrpc_request_set *set)
5870 if (KEY_IS(KEY_SPTLRPC_CONF)) {
5871 rc = tgt_adapt_sptlrpc_conf(class_exp2tgt(exp));
5879 * Match client and server connection feature flags.
5881 * Compute the compatibility flags for a connection request based on
5882 * features mutually supported by client and server.
5884 * The obd_export::exp_connect_data.ocd_connect_flags field in \a exp
5885 * must not be updated here, otherwise a partially initialized value may
5886 * be exposed. After the connection request is successfully processed,
5887 * the top-level MDT connect request handler atomically updates the export
5888 * connect flags from the obd_connect_data::ocd_connect_flags field of the
5889 * reply. \see mdt_connect().
5891 * Before 2.7.50 clients will send a struct obd_connect_data_v1 rather than a
5892 * full struct obd_connect_data. So care must be taken when accessing fields
5893 * that are not present in struct obd_connect_data_v1. See LU-16.
5895 * \param exp the obd_export associated with this client/target pair
5896 * \param mdt the target device for the connection
5897 * \param data stores data for this connect request
5900 * \retval -EPROTO \a data unexpectedly has zero obd_connect_data::ocd_brw_size
5901 * \retval -EBADE client and server feature requirements are incompatible
5903 static int mdt_connect_internal(const struct lu_env *env,
5904 struct obd_export *exp,
5905 struct mdt_device *mdt,
5906 struct obd_connect_data *data, bool reconnect)
5908 const char *obd_name = mdt_obd_name(mdt);
5909 LASSERT(data != NULL);
5911 data->ocd_connect_flags &= MDT_CONNECT_SUPPORTED;
5913 if (mdt->mdt_bottom->dd_rdonly &&
5914 !(data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
5915 !(data->ocd_connect_flags & OBD_CONNECT_RDONLY))
5918 if (data->ocd_connect_flags & OBD_CONNECT_FLAGS2)
5919 data->ocd_connect_flags2 &= MDT_CONNECT_SUPPORTED2;
5921 data->ocd_ibits_known &= MDS_INODELOCK_FULL;
5923 if (!mdt->mdt_opts.mo_acl)
5924 data->ocd_connect_flags &= ~OBD_CONNECT_ACL;
5926 if (!mdt->mdt_opts.mo_user_xattr)
5927 data->ocd_connect_flags &= ~OBD_CONNECT_XATTR;
5929 if (OCD_HAS_FLAG(data, BRW_SIZE)) {
5930 data->ocd_brw_size = min(data->ocd_brw_size,
5932 if (data->ocd_brw_size == 0) {
5933 CERROR("%s: cli %s/%p ocd_connect_flags: %#llx "
5934 "ocd_version: %x ocd_grant: %d ocd_index: %u "
5935 "ocd_brw_size unexpectedly zero, network data "
5936 "corruption? Refusing to connect this client\n",
5937 obd_name, exp->exp_client_uuid.uuid,
5938 exp, data->ocd_connect_flags, data->ocd_version,
5939 data->ocd_grant, data->ocd_index);
5944 if (OCD_HAS_FLAG(data, GRANT_PARAM)) {
5945 struct dt_device_param *ddp = &mdt->mdt_lut.lut_dt_conf;
5947 /* client is reporting its page size, for future use */
5948 exp->exp_target_data.ted_pagebits = data->ocd_grant_blkbits;
5949 data->ocd_grant_blkbits = mdt->mdt_lut.lut_tgd.tgd_blockbits;
5950 /* ddp_inodespace may not be power-of-two value, eg. for ldiskfs
5951 * it's LDISKFS_DIR_REC_LEN(20) = 28. */
5952 data->ocd_grant_inobits = fls(ddp->ddp_inodespace - 1);
5953 /* ocd_grant_tax_kb is in 1K byte blocks */
5954 data->ocd_grant_tax_kb = ddp->ddp_extent_tax >> 10;
5955 data->ocd_grant_max_blks = ddp->ddp_max_extent_blks;
5958 /* Save connect_data we have so far because tgt_grant_connect()
5959 * uses it to calculate grant, and we want to save the client
5960 * version before it is overwritten by LUSTRE_VERSION_CODE. */
5961 exp->exp_connect_data = *data;
5962 if (OCD_HAS_FLAG(data, GRANT))
5963 tgt_grant_connect(env, exp, data, !reconnect);
5965 if (OCD_HAS_FLAG(data, MAXBYTES))
5966 data->ocd_maxbytes = mdt->mdt_lut.lut_dt_conf.ddp_maxbytes;
5968 /* NB: Disregard the rule against updating
5969 * exp_connect_data.ocd_connect_flags in this case, since
5970 * tgt_client_new() needs to know if this is a lightweight
5971 * connection, and it is safe to expose this flag before
5972 * connection processing completes. */
5973 if (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) {
5974 spin_lock(&exp->exp_lock);
5975 *exp_connect_flags_ptr(exp) |= OBD_CONNECT_LIGHTWEIGHT;
5976 spin_unlock(&exp->exp_lock);
5979 data->ocd_version = LUSTRE_VERSION_CODE;
5981 if ((data->ocd_connect_flags & OBD_CONNECT_FID) == 0) {
5982 CWARN("%s: MDS requires FID support, but client not\n",
5987 if (OCD_HAS_FLAG(data, PINGLESS)) {
5988 if (ptlrpc_pinger_suppress_pings()) {
5989 spin_lock(&exp->exp_obd->obd_dev_lock);
5990 list_del_init(&exp->exp_obd_chain_timed);
5991 spin_unlock(&exp->exp_obd->obd_dev_lock);
5993 data->ocd_connect_flags &= ~OBD_CONNECT_PINGLESS;
5997 data->ocd_max_easize = mdt->mdt_max_ea_size;
5999 /* NB: Disregard the rule against updating
6000 * exp_connect_data.ocd_connect_flags in this case, since
6001 * tgt_client_new() needs to know if this is client supports
6002 * multiple modify RPCs, and it is safe to expose this flag before
6003 * connection processing completes. */
6004 if (data->ocd_connect_flags & OBD_CONNECT_MULTIMODRPCS) {
6005 data->ocd_maxmodrpcs = max_mod_rpcs_per_client;
6006 spin_lock(&exp->exp_lock);
6007 *exp_connect_flags_ptr(exp) |= OBD_CONNECT_MULTIMODRPCS;
6008 spin_unlock(&exp->exp_lock);
6011 if (OCD_HAS_FLAG(data, CKSUM)) {
6012 __u32 cksum_types = data->ocd_cksum_types;
6014 /* The client set in ocd_cksum_types the checksum types it
6015 * supports. We have to mask off the algorithms that we don't
6017 data->ocd_cksum_types &=
6018 obd_cksum_types_supported_server(obd_name);
6020 if (unlikely(data->ocd_cksum_types == 0)) {
6021 CERROR("%s: Connect with checksum support but no "
6022 "ocd_cksum_types is set\n",
6023 exp->exp_obd->obd_name);
6027 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
6028 "%x\n", exp->exp_obd->obd_name, obd_export_nid2str(exp),
6029 cksum_types, data->ocd_cksum_types);
6031 /* This client does not support OBD_CONNECT_CKSUM
6032 * fall back to CRC32 */
6033 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
6034 "OBD_CONNECT_CKSUM, CRC32 will be used\n",
6035 exp->exp_obd->obd_name, obd_export_nid2str(exp));
6041 static int mdt_ctxt_add_dirty_flag(struct lu_env *env,
6042 struct mdt_thread_info *info,
6043 struct mdt_file_data *mfd)
6045 struct lu_context ses;
6049 rc = lu_context_init(&ses, LCT_SERVER_SESSION);
6054 lu_context_enter(&ses);
6056 mdt_ucred(info)->uc_valid = UCRED_OLD;
6057 rc = mdt_add_dirty_flag(info, mfd->mfd_object, &info->mti_attr);
6059 lu_context_exit(&ses);
6060 lu_context_fini(&ses);
6066 static int mdt_export_cleanup(struct obd_export *exp)
6068 struct list_head closing_list;
6069 struct mdt_export_data *med = &exp->exp_mdt_data;
6070 struct obd_device *obd = exp->exp_obd;
6071 struct mdt_device *mdt;
6072 struct mdt_thread_info *info;
6074 struct mdt_file_data *mfd, *n;
6078 INIT_LIST_HEAD(&closing_list);
6079 spin_lock(&med->med_open_lock);
6080 while (!list_empty(&med->med_open_head)) {
6081 struct list_head *tmp = med->med_open_head.next;
6082 mfd = list_entry(tmp, struct mdt_file_data, mfd_list);
6084 /* Remove mfd handle so it can't be found again.
6085 * We are consuming the mfd_list reference here. */
6086 class_handle_unhash(&mfd->mfd_open_handle);
6087 list_move_tail(&mfd->mfd_list, &closing_list);
6089 spin_unlock(&med->med_open_lock);
6090 mdt = mdt_dev(obd->obd_lu_dev);
6091 LASSERT(mdt != NULL);
6093 rc = lu_env_init(&env, LCT_MD_THREAD);
6097 info = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
6098 LASSERT(info != NULL);
6099 memset(info, 0, sizeof *info);
6100 info->mti_env = &env;
6101 info->mti_mdt = mdt;
6102 info->mti_exp = exp;
6104 if (!list_empty(&closing_list)) {
6105 struct md_attr *ma = &info->mti_attr;
6107 /* Close any open files (which may also cause orphan
6109 list_for_each_entry_safe(mfd, n, &closing_list, mfd_list) {
6110 list_del_init(&mfd->mfd_list);
6111 ma->ma_need = ma->ma_valid = 0;
6113 /* This file is being closed due to an eviction, it
6114 * could have been modified and now dirty regarding to
6115 * HSM archive, check this!
6116 * The logic here is to mark a file dirty if there's a
6117 * chance it was dirtied before the client was evicted,
6118 * so that we don't have to wait for a release attempt
6119 * before finding out the file was actually dirty and
6120 * fail the release. Aggressively marking it dirty here
6121 * will cause the policy engine to attempt to
6122 * re-archive it; when rearchiving, we can compare the
6123 * current version to the HSM data_version and make the
6124 * archive request into a noop if it's not actually
6127 if (mfd->mfd_open_flags & MDS_FMODE_WRITE)
6128 rc = mdt_ctxt_add_dirty_flag(&env, info, mfd);
6130 /* Don't unlink orphan on failover umount, LU-184 */
6131 if (exp->exp_flags & OBD_OPT_FAILOVER) {
6132 ma->ma_valid = MA_FLAGS;
6133 ma->ma_attr_flags |= MDS_KEEP_ORPHAN;
6135 ma->ma_valid |= MA_FORCE_LOG;
6136 mdt_mfd_close(info, mfd);
6139 info->mti_mdt = NULL;
6140 /* cleanup client slot early */
6141 /* Do not erase record for recoverable client. */
6142 if (!(exp->exp_flags & OBD_OPT_FAILOVER) || exp->exp_failed)
6143 tgt_client_del(&env, exp);
6149 static inline void mdt_enable_slc(struct mdt_device *mdt)
6151 if (mdt->mdt_lut.lut_sync_lock_cancel == NEVER_SYNC_ON_CANCEL)
6152 mdt->mdt_lut.lut_sync_lock_cancel = BLOCKING_SYNC_ON_CANCEL;
6155 static inline void mdt_disable_slc(struct mdt_device *mdt)
6157 if (mdt->mdt_lut.lut_sync_lock_cancel == BLOCKING_SYNC_ON_CANCEL)
6158 mdt->mdt_lut.lut_sync_lock_cancel = NEVER_SYNC_ON_CANCEL;
6161 static int mdt_obd_disconnect(struct obd_export *exp)
6168 class_export_get(exp);
6170 if (!(exp->exp_flags & OBD_OPT_FORCE))
6171 tgt_grant_sanity_check(exp->exp_obd, __func__);
6173 if ((exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) &&
6174 !(exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)) {
6175 struct mdt_device *mdt = mdt_dev(exp->exp_obd->obd_lu_dev);
6177 if (atomic_dec_and_test(&mdt->mdt_mds_mds_conns))
6178 mdt_disable_slc(mdt);
6181 rc = server_disconnect_export(exp);
6183 CDEBUG(D_IOCTL, "server disconnect error: rc = %d\n", rc);
6185 tgt_grant_discard(exp);
6187 rc = mdt_export_cleanup(exp);
6188 nodemap_del_member(exp);
6189 class_export_put(exp);
6193 /* mds_connect copy */
6194 static int mdt_obd_connect(const struct lu_env *env,
6195 struct obd_export **exp, struct obd_device *obd,
6196 struct obd_uuid *cluuid,
6197 struct obd_connect_data *data,
6200 struct obd_export *lexp;
6201 struct lustre_handle conn = { 0 };
6202 struct mdt_device *mdt;
6204 lnet_nid_t *client_nid = localdata;
6207 LASSERT(env != NULL);
6208 LASSERT(data != NULL);
6210 if (!exp || !obd || !cluuid)
6213 mdt = mdt_dev(obd->obd_lu_dev);
6215 if ((data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
6216 !(data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT)) {
6217 atomic_inc(&mdt->mdt_mds_mds_conns);
6218 mdt_enable_slc(mdt);
6222 * first, check whether the stack is ready to handle requests
6223 * XXX: probably not very appropriate method is used now
6224 * at some point we should find a better one
6226 if (!test_bit(MDT_FL_SYNCED, &mdt->mdt_state) &&
6227 !(data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) &&
6228 !(data->ocd_connect_flags & OBD_CONNECT_MDS_MDS)) {
6229 rc = obd_get_info(env, mdt->mdt_child_exp,
6230 sizeof(KEY_OSP_CONNECTED),
6231 KEY_OSP_CONNECTED, NULL, NULL);
6234 set_bit(MDT_FL_SYNCED, &mdt->mdt_state);
6237 rc = class_connect(&conn, obd, cluuid);
6241 lexp = class_conn2export(&conn);
6242 LASSERT(lexp != NULL);
6244 rc = nodemap_add_member(*client_nid, lexp);
6245 if (rc != 0 && rc != -EEXIST)
6248 rc = mdt_connect_internal(env, lexp, mdt, data, false);
6250 struct lsd_client_data *lcd = lexp->exp_target_data.ted_lcd;
6253 memcpy(lcd->lcd_uuid, cluuid, sizeof lcd->lcd_uuid);
6254 rc = tgt_client_new(env, lexp);
6256 mdt_export_stats_init(obd, lexp, localdata);
6260 class_disconnect(lexp);
6261 nodemap_del_member(lexp);
6265 /* Because we do not want this export to be evicted by pinger,
6266 * let's not add this export to the timed chain list. */
6267 if (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) {
6268 spin_lock(&lexp->exp_obd->obd_dev_lock);
6269 list_del_init(&lexp->exp_obd_chain_timed);
6270 spin_unlock(&lexp->exp_obd->obd_dev_lock);
6277 static int mdt_obd_reconnect(const struct lu_env *env,
6278 struct obd_export *exp, struct obd_device *obd,
6279 struct obd_uuid *cluuid,
6280 struct obd_connect_data *data,
6283 lnet_nid_t *client_nid = localdata;
6287 if (exp == NULL || obd == NULL || cluuid == NULL)
6290 rc = nodemap_add_member(*client_nid, exp);
6291 if (rc != 0 && rc != -EEXIST)
6294 rc = mdt_connect_internal(env, exp, mdt_dev(obd->obd_lu_dev), data,
6297 mdt_export_stats_init(obd, exp, localdata);
6299 nodemap_del_member(exp);
6304 /* FIXME: Can we avoid using these two interfaces? */
6305 static int mdt_init_export(struct obd_export *exp)
6307 struct mdt_export_data *med = &exp->exp_mdt_data;
6311 INIT_LIST_HEAD(&med->med_open_head);
6312 spin_lock_init(&med->med_open_lock);
6313 spin_lock(&exp->exp_lock);
6314 exp->exp_connecting = 1;
6315 spin_unlock(&exp->exp_lock);
6317 OBD_ALLOC(exp->exp_used_slots,
6318 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
6319 if (exp->exp_used_slots == NULL)
6322 /* self-export doesn't need client data and ldlm initialization */
6323 if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
6324 &exp->exp_client_uuid)))
6327 rc = tgt_client_alloc(exp);
6331 rc = ldlm_init_export(exp);
6338 tgt_client_free(exp);
6340 OBD_FREE(exp->exp_used_slots,
6341 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
6342 exp->exp_used_slots = NULL;
6344 CERROR("%s: Failed to initialize export: rc = %d\n",
6345 exp->exp_obd->obd_name, rc);
6349 static int mdt_destroy_export(struct obd_export *exp)
6353 target_destroy_export(exp);
6354 if (exp->exp_used_slots)
6355 OBD_FREE(exp->exp_used_slots,
6356 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
6358 /* destroy can be called from failed obd_setup, so
6359 * checking uuid is safer than obd_self_export */
6360 if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
6361 &exp->exp_client_uuid)))
6364 ldlm_destroy_export(exp);
6365 tgt_client_free(exp);
6367 LASSERT(list_empty(&exp->exp_outstanding_replies));
6368 LASSERT(list_empty(&exp->exp_mdt_data.med_open_head));
6371 * discard grants once we're sure no more
6372 * interaction with the client is possible
6374 tgt_grant_discard(exp);
6375 if (exp_connect_flags(exp) & OBD_CONNECT_GRANT)
6376 exp->exp_obd->u.obt.obt_lut->lut_tgd.tgd_tot_granted_clients--;
6378 if (!(exp->exp_flags & OBD_OPT_FORCE))
6379 tgt_grant_sanity_check(exp->exp_obd, __func__);
6384 int mdt_links_read(struct mdt_thread_info *info, struct mdt_object *mdt_obj,
6385 struct linkea_data *ldata)
6389 LASSERT(ldata->ld_buf->lb_buf != NULL);
6391 if (!mdt_object_exists(mdt_obj))
6394 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
6395 ldata->ld_buf, XATTR_NAME_LINK);
6396 if (rc == -ERANGE) {
6397 /* Buf was too small, figure out what we need. */
6398 lu_buf_free(ldata->ld_buf);
6399 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
6400 ldata->ld_buf, XATTR_NAME_LINK);
6403 ldata->ld_buf = lu_buf_check_and_alloc(ldata->ld_buf, rc);
6404 if (ldata->ld_buf->lb_buf == NULL)
6406 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
6407 ldata->ld_buf, XATTR_NAME_LINK);
6412 return linkea_init_with_rec(ldata);
6416 * Given an MDT object, try to look up the full path to the object.
6417 * Part of the MDT layer implementation of lfs fid2path.
6419 * \param[in] info Per-thread common data shared by MDT level handlers.
6420 * \param[in] obj Object to do path lookup of
6421 * \param[in,out] fp User-provided struct to store path information
6422 * \param[in] root_fid Root FID of current path should reach
6424 * \retval 0 Lookup successful, path information stored in fp
6425 * \retval -EAGAIN Lookup failed, usually because object is being moved
6426 * \retval negative errno if there was a problem
6428 static int mdt_path_current(struct mdt_thread_info *info,
6429 struct mdt_object *obj,
6430 struct getinfo_fid2path *fp,
6431 struct lu_fid *root_fid)
6433 struct mdt_device *mdt = info->mti_mdt;
6434 struct mdt_object *mdt_obj;
6435 struct link_ea_header *leh;
6436 struct link_ea_entry *lee;
6437 struct lu_name *tmpname = &info->mti_name;
6438 struct lu_fid *tmpfid = &info->mti_tmp_fid1;
6439 struct lu_buf *buf = &info->mti_big_buf;
6442 struct linkea_data ldata = { NULL };
6447 /* temp buffer for path element, the buffer will be finally freed
6448 * in mdt_thread_info_fini */
6449 buf = lu_buf_check_and_alloc(buf, PATH_MAX);
6450 if (buf->lb_buf == NULL)
6454 ptr = fp->gf_u.gf_path + fp->gf_pathlen - 1;
6457 *tmpfid = fp->gf_fid = *mdt_object_fid(obj);
6459 while (!lu_fid_eq(root_fid, &fp->gf_fid)) {
6460 struct lu_buf lmv_buf;
6462 if (!lu_fid_eq(root_fid, &mdt->mdt_md_root_fid) &&
6463 lu_fid_eq(&mdt->mdt_md_root_fid, &fp->gf_fid))
6464 GOTO(out, rc = -ENOENT);
6466 if (lu_fid_eq(mdt_object_fid(obj), tmpfid)) {
6468 mdt_object_get(info->mti_env, mdt_obj);
6470 mdt_obj = mdt_object_find(info->mti_env, mdt, tmpfid);
6471 if (IS_ERR(mdt_obj))
6472 GOTO(out, rc = PTR_ERR(mdt_obj));
6475 if (!mdt_object_exists(mdt_obj)) {
6476 mdt_object_put(info->mti_env, mdt_obj);
6477 GOTO(out, rc = -ENOENT);
6480 if (mdt_object_remote(mdt_obj)) {
6481 mdt_object_put(info->mti_env, mdt_obj);
6482 GOTO(remote_out, rc = -EREMOTE);
6485 rc = mdt_links_read(info, mdt_obj, &ldata);
6487 mdt_object_put(info->mti_env, mdt_obj);
6492 lee = (struct link_ea_entry *)(leh + 1); /* link #0 */
6493 linkea_entry_unpack(lee, &reclen, tmpname, tmpfid);
6494 /* If set, use link #linkno for path lookup, otherwise use
6495 link #0. Only do this for the final path element. */
6496 if (first && fp->gf_linkno < leh->leh_reccount) {
6498 for (count = 0; count < fp->gf_linkno; count++) {
6499 lee = (struct link_ea_entry *)
6500 ((char *)lee + reclen);
6501 linkea_entry_unpack(lee, &reclen, tmpname,
6504 if (fp->gf_linkno < leh->leh_reccount - 1)
6505 /* indicate to user there are more links */
6509 lmv_buf.lb_buf = info->mti_xattr_buf;
6510 lmv_buf.lb_len = sizeof(info->mti_xattr_buf);
6511 /* Check if it is slave stripes */
6512 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
6513 &lmv_buf, XATTR_NAME_LMV);
6514 mdt_object_put(info->mti_env, mdt_obj);
6516 union lmv_mds_md *lmm = lmv_buf.lb_buf;
6518 /* For slave stripes, get its master */
6519 if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_STRIPE) {
6520 fp->gf_fid = *tmpfid;
6523 } else if (rc < 0 && rc != -ENODATA) {
6529 /* Pack the name in the end of the buffer */
6530 ptr -= tmpname->ln_namelen;
6531 if (ptr - 1 <= fp->gf_u.gf_path)
6532 GOTO(out, rc = -EOVERFLOW);
6533 strncpy(ptr, tmpname->ln_name, tmpname->ln_namelen);
6536 /* keep the last resolved fid to the client, so the
6537 * client will build the left path on another MDT for
6539 fp->gf_fid = *tmpfid;
6545 ptr++; /* skip leading / */
6546 memmove(fp->gf_u.gf_path, ptr,
6547 fp->gf_u.gf_path + fp->gf_pathlen - ptr);
6554 * Given an MDT object, use mdt_path_current to get the path.
6555 * Essentially a wrapper to retry mdt_path_current a set number of times
6556 * if -EAGAIN is returned (usually because an object is being moved).
6558 * Part of the MDT layer implementation of lfs fid2path.
6560 * \param[in] info Per-thread common data shared by mdt level handlers.
6561 * \param[in] obj Object to do path lookup of
6562 * \param[in,out] fp User-provided struct for arguments and to store path
6565 * \retval 0 Lookup successful, path information stored in fp
6566 * \retval negative errno if there was a problem
6568 static int mdt_path(struct mdt_thread_info *info, struct mdt_object *obj,
6569 struct getinfo_fid2path *fp, struct lu_fid *root_fid)
6571 struct mdt_device *mdt = info->mti_mdt;
6576 if (fp->gf_pathlen < 3)
6579 if (root_fid == NULL)
6580 root_fid = &mdt->mdt_md_root_fid;
6582 if (lu_fid_eq(root_fid, mdt_object_fid(obj))) {
6583 fp->gf_u.gf_path[0] = '\0';
6587 /* Retry multiple times in case file is being moved */
6588 while (tries-- && rc == -EAGAIN)
6589 rc = mdt_path_current(info, obj, fp, root_fid);
6595 * Get the full path of the provided FID, as of changelog record recno.
6597 * This checks sanity and looks up object for user provided FID
6598 * before calling the actual path lookup code.
6600 * Part of the MDT layer implementation of lfs fid2path.
6602 * \param[in] info Per-thread common data shared by mdt level handlers.
6603 * \param[in,out] fp User-provided struct for arguments and to store path
6606 * \retval 0 Lookup successful, path information and recno stored in fp
6607 * \retval -ENOENT, object does not exist
6608 * \retval negative errno if there was a problem
6610 static int mdt_fid2path(struct mdt_thread_info *info,
6611 struct lu_fid *root_fid,
6612 struct getinfo_fid2path *fp)
6614 struct mdt_device *mdt = info->mti_mdt;
6615 struct mdt_object *obj;
6619 CDEBUG(D_IOCTL, "path get "DFID" from %llu #%d\n",
6620 PFID(&fp->gf_fid), fp->gf_recno, fp->gf_linkno);
6622 if (!fid_is_sane(&fp->gf_fid))
6625 if (!fid_is_namespace_visible(&fp->gf_fid)) {
6626 CDEBUG(D_INFO, "%s: "DFID" is invalid, f_seq should be >= %#llx"
6627 ", or f_oid != 0, or f_ver == 0\n", mdt_obd_name(mdt),
6628 PFID(&fp->gf_fid), (__u64)FID_SEQ_NORMAL);
6632 obj = mdt_object_find(info->mti_env, mdt, &fp->gf_fid);
6635 CDEBUG(D_IOCTL, "cannot find "DFID": rc = %d\n",
6636 PFID(&fp->gf_fid), rc);
6640 if (mdt_object_remote(obj))
6642 else if (!mdt_object_exists(obj))
6648 mdt_object_put(info->mti_env, obj);
6649 CDEBUG(D_IOCTL, "nonlocal object "DFID": rc = %d\n",
6650 PFID(&fp->gf_fid), rc);
6654 rc = mdt_path(info, obj, fp, root_fid);
6656 CDEBUG(D_INFO, "fid "DFID", path %s recno %#llx linkno %u\n",
6657 PFID(&fp->gf_fid), fp->gf_u.gf_path,
6658 fp->gf_recno, fp->gf_linkno);
6660 mdt_object_put(info->mti_env, obj);
6665 static int mdt_rpc_fid2path(struct mdt_thread_info *info, void *key, int keylen,
6666 void *val, int vallen)
6668 struct getinfo_fid2path *fpout, *fpin;
6669 struct lu_fid *root_fid = NULL;
6672 fpin = key + cfs_size_round(sizeof(KEY_FID2PATH));
6675 if (ptlrpc_req_need_swab(info->mti_pill->rc_req))
6676 lustre_swab_fid2path(fpin);
6678 memcpy(fpout, fpin, sizeof(*fpin));
6679 if (fpout->gf_pathlen != vallen - sizeof(*fpin))
6682 if (keylen >= cfs_size_round(sizeof(KEY_FID2PATH)) + sizeof(*fpin) +
6683 sizeof(struct lu_fid)) {
6684 /* client sent its root FID, which is normally fileset FID */
6685 root_fid = fpin->gf_u.gf_root_fid;
6686 if (ptlrpc_req_need_swab(info->mti_pill->rc_req))
6687 lustre_swab_lu_fid(root_fid);
6689 if (root_fid != NULL && !fid_is_sane(root_fid))
6693 rc = mdt_fid2path(info, root_fid, fpout);
6697 int mdt_get_info(struct tgt_session_info *tsi)
6707 key = req_capsule_client_get(tsi->tsi_pill, &RMF_GETINFO_KEY);
6709 CDEBUG(D_IOCTL, "No GETINFO key\n");
6710 RETURN(err_serious(-EFAULT));
6712 keylen = req_capsule_get_size(tsi->tsi_pill, &RMF_GETINFO_KEY,
6715 vallen = req_capsule_client_get(tsi->tsi_pill, &RMF_GETINFO_VALLEN);
6716 if (vallen == NULL) {
6717 CDEBUG(D_IOCTL, "%s: cannot get RMF_GETINFO_VALLEN buffer\n",
6718 tgt_name(tsi->tsi_tgt));
6719 RETURN(err_serious(-EFAULT));
6722 req_capsule_set_size(tsi->tsi_pill, &RMF_GETINFO_VAL, RCL_SERVER,
6724 rc = req_capsule_server_pack(tsi->tsi_pill);
6726 RETURN(err_serious(rc));
6728 valout = req_capsule_server_get(tsi->tsi_pill, &RMF_GETINFO_VAL);
6729 if (valout == NULL) {
6730 CDEBUG(D_IOCTL, "%s: cannot get get-info RPC out buffer\n",
6731 tgt_name(tsi->tsi_tgt));
6732 RETURN(err_serious(-EFAULT));
6735 if (KEY_IS(KEY_FID2PATH)) {
6736 struct mdt_thread_info *info = tsi2mdt_info(tsi);
6738 rc = mdt_rpc_fid2path(info, key, keylen, valout, *vallen);
6739 mdt_thread_info_fini(info);
6746 static int mdt_ioc_version_get(struct mdt_thread_info *mti, void *karg)
6748 struct obd_ioctl_data *data = karg;
6751 struct mdt_object *obj;
6752 struct mdt_lock_handle *lh;
6756 if (data->ioc_inlbuf1 == NULL || data->ioc_inllen1 != sizeof(*fid) ||
6757 data->ioc_inlbuf2 == NULL || data->ioc_inllen2 != sizeof(version))
6760 fid = (struct lu_fid *)data->ioc_inlbuf1;
6762 if (!fid_is_sane(fid))
6765 CDEBUG(D_IOCTL, "getting version for "DFID"\n", PFID(fid));
6767 lh = &mti->mti_lh[MDT_LH_PARENT];
6768 mdt_lock_reg_init(lh, LCK_CR);
6770 obj = mdt_object_find_lock(mti, fid, lh, MDS_INODELOCK_UPDATE);
6772 RETURN(PTR_ERR(obj));
6774 if (mdt_object_remote(obj)) {
6777 * before calling version get the correct MDS should be
6778 * fid, this is error to find remote object here
6780 CERROR("nonlocal object "DFID"\n", PFID(fid));
6781 } else if (!mdt_object_exists(obj)) {
6782 *(__u64 *)data->ioc_inlbuf2 = ENOENT_VERSION;
6785 version = dt_version_get(mti->mti_env, mdt_obj2dt(obj));
6786 *(__u64 *)data->ioc_inlbuf2 = version;
6789 mdt_object_unlock_put(mti, obj, lh, 1);
6793 /* ioctls on obd dev */
6794 static int mdt_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
6795 void *karg, void __user *uarg)
6798 struct obd_device *obd = exp->exp_obd;
6799 struct mdt_device *mdt = mdt_dev(obd->obd_lu_dev);
6800 struct dt_device *dt = mdt->mdt_bottom;
6804 CDEBUG(D_IOCTL, "handling ioctl cmd %#x\n", cmd);
6805 rc = lu_env_init(&env, LCT_MD_THREAD);
6811 rc = mdt_device_sync(&env, mdt);
6813 case OBD_IOC_SET_READONLY:
6814 rc = dt_sync(&env, dt);
6816 rc = dt_ro(&env, dt);
6818 case OBD_IOC_ABORT_RECOVERY:
6819 CERROR("%s: Aborting recovery for device\n", mdt_obd_name(mdt));
6820 obd->obd_abort_recovery = 1;
6821 target_stop_recovery_thread(obd);
6824 case OBD_IOC_CHANGELOG_REG:
6825 case OBD_IOC_CHANGELOG_DEREG:
6826 case OBD_IOC_CHANGELOG_CLEAR:
6827 rc = mdt->mdt_child->md_ops->mdo_iocontrol(&env,
6831 case OBD_IOC_START_LFSCK: {
6832 struct md_device *next = mdt->mdt_child;
6833 struct obd_ioctl_data *data = karg;
6834 struct lfsck_start_param lsp;
6836 if (unlikely(data == NULL)) {
6841 lsp.lsp_start = (struct lfsck_start *)(data->ioc_inlbuf1);
6842 lsp.lsp_index_valid = 0;
6843 rc = next->md_ops->mdo_iocontrol(&env, next, cmd, 0, &lsp);
6846 case OBD_IOC_STOP_LFSCK: {
6847 struct md_device *next = mdt->mdt_child;
6848 struct obd_ioctl_data *data = karg;
6849 struct lfsck_stop stop;
6851 stop.ls_status = LS_STOPPED;
6852 /* Old lfsck utils may pass NULL @stop. */
6853 if (data->ioc_inlbuf1 == NULL)
6857 ((struct lfsck_stop *)(data->ioc_inlbuf1))->ls_flags;
6859 rc = next->md_ops->mdo_iocontrol(&env, next, cmd, 0, &stop);
6862 case OBD_IOC_QUERY_LFSCK: {
6863 struct md_device *next = mdt->mdt_child;
6864 struct obd_ioctl_data *data = karg;
6866 rc = next->md_ops->mdo_iocontrol(&env, next, cmd, 0,
6870 case OBD_IOC_GET_OBJ_VERSION: {
6871 struct mdt_thread_info *mti;
6872 mti = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
6873 memset(mti, 0, sizeof *mti);
6874 mti->mti_env = &env;
6878 rc = mdt_ioc_version_get(mti, karg);
6881 case OBD_IOC_CATLOGLIST: {
6882 struct mdt_thread_info *mti;
6884 mti = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
6885 lu_local_obj_fid(&mti->mti_tmp_fid1, LLOG_CATALOGS_OID);
6886 rc = llog_catalog_list(&env, mdt->mdt_bottom, 0, karg,
6887 &mti->mti_tmp_fid1);
6892 CERROR("%s: Not supported cmd = %d, rc = %d\n",
6893 mdt_obd_name(mdt), cmd, rc);
6900 static int mdt_postrecov(const struct lu_env *env, struct mdt_device *mdt)
6902 struct lu_device *ld = md2lu_dev(mdt->mdt_child);
6906 if (!mdt->mdt_skip_lfsck && !mdt->mdt_bottom->dd_rdonly) {
6907 struct lfsck_start_param lsp;
6909 lsp.lsp_start = NULL;
6910 lsp.lsp_index_valid = 0;
6911 rc = mdt->mdt_child->md_ops->mdo_iocontrol(env, mdt->mdt_child,
6912 OBD_IOC_START_LFSCK,
6914 if (rc != 0 && rc != -EALREADY)
6915 CWARN("%s: auto trigger paused LFSCK failed: rc = %d\n",
6916 mdt_obd_name(mdt), rc);
6919 rc = ld->ld_ops->ldo_recovery_complete(env, ld);
6923 static int mdt_obd_postrecov(struct obd_device *obd)
6928 rc = lu_env_init(&env, LCT_MD_THREAD);
6931 rc = mdt_postrecov(&env, mdt_dev(obd->obd_lu_dev));
6936 static struct obd_ops mdt_obd_device_ops = {
6937 .o_owner = THIS_MODULE,
6938 .o_set_info_async = mdt_obd_set_info_async,
6939 .o_connect = mdt_obd_connect,
6940 .o_reconnect = mdt_obd_reconnect,
6941 .o_disconnect = mdt_obd_disconnect,
6942 .o_init_export = mdt_init_export,
6943 .o_destroy_export = mdt_destroy_export,
6944 .o_iocontrol = mdt_iocontrol,
6945 .o_postrecov = mdt_obd_postrecov,
6946 /* Data-on-MDT IO methods */
6947 .o_preprw = mdt_obd_preprw,
6948 .o_commitrw = mdt_obd_commitrw,
6951 static struct lu_device* mdt_device_fini(const struct lu_env *env,
6952 struct lu_device *d)
6954 struct mdt_device *m = mdt_dev(d);
6961 static struct lu_device *mdt_device_free(const struct lu_env *env,
6962 struct lu_device *d)
6964 struct mdt_device *m = mdt_dev(d);
6967 lu_device_fini(&m->mdt_lu_dev);
6973 static struct lu_device *mdt_device_alloc(const struct lu_env *env,
6974 struct lu_device_type *t,
6975 struct lustre_cfg *cfg)
6977 struct lu_device *l;
6978 struct mdt_device *m;
6985 rc = mdt_init0(env, m, t, cfg);
6987 mdt_device_free(env, l);
6992 l = ERR_PTR(-ENOMEM);
6996 /* context key constructor/destructor: mdt_key_init, mdt_key_fini */
6997 LU_KEY_INIT(mdt, struct mdt_thread_info);
6999 static void mdt_key_fini(const struct lu_context *ctx,
7000 struct lu_context_key *key, void* data)
7002 struct mdt_thread_info *info = data;
7004 if (info->mti_big_lmm) {
7005 OBD_FREE_LARGE(info->mti_big_lmm, info->mti_big_lmmsize);
7006 info->mti_big_lmm = NULL;
7007 info->mti_big_lmmsize = 0;
7010 if (info->mti_big_acl) {
7011 OBD_FREE_LARGE(info->mti_big_acl, info->mti_big_aclsize);
7012 info->mti_big_acl = NULL;
7013 info->mti_big_aclsize = 0;
7019 /* context key: mdt_thread_key */
7020 LU_CONTEXT_KEY_DEFINE(mdt, LCT_MD_THREAD);
7022 struct lu_ucred *mdt_ucred(const struct mdt_thread_info *info)
7024 return lu_ucred(info->mti_env);
7027 struct lu_ucred *mdt_ucred_check(const struct mdt_thread_info *info)
7029 return lu_ucred_check(info->mti_env);
7033 * Enable/disable COS (Commit On Sharing).
7035 * Set/Clear the COS flag in mdt options.
7037 * \param mdt mdt device
7038 * \param val 0 disables COS, other values enable COS
7040 void mdt_enable_cos(struct mdt_device *mdt, bool val)
7045 mdt->mdt_opts.mo_cos = val;
7046 rc = lu_env_init(&env, LCT_LOCAL);
7047 if (unlikely(rc != 0)) {
7048 CWARN("%s: lu_env initialization failed, cannot "
7049 "sync: rc = %d\n", mdt_obd_name(mdt), rc);
7052 mdt_device_sync(&env, mdt);
7057 * Check COS (Commit On Sharing) status.
7059 * Return COS flag status.
7061 * \param mdt mdt device
7063 int mdt_cos_is_enabled(struct mdt_device *mdt)
7065 return mdt->mdt_opts.mo_cos != 0;
7068 static struct lu_device_type_operations mdt_device_type_ops = {
7069 .ldto_device_alloc = mdt_device_alloc,
7070 .ldto_device_free = mdt_device_free,
7071 .ldto_device_fini = mdt_device_fini
7074 static struct lu_device_type mdt_device_type = {
7075 .ldt_tags = LU_DEVICE_MD,
7076 .ldt_name = LUSTRE_MDT_NAME,
7077 .ldt_ops = &mdt_device_type_ops,
7078 .ldt_ctx_tags = LCT_MD_THREAD
7081 static int __init mdt_init(void)
7085 BUILD_BUG_ON(sizeof("0x0123456789ABCDEF:0x01234567:0x01234567") !=
7086 FID_NOBRACE_LEN + 1);
7087 BUILD_BUG_ON(sizeof("[0x0123456789ABCDEF:0x01234567:0x01234567]") !=
7089 rc = lu_kmem_init(mdt_caches);
7093 rc = mds_mod_init();
7097 rc = class_register_type(&mdt_obd_device_ops, NULL, true, NULL,
7098 LUSTRE_MDT_NAME, &mdt_device_type);
7103 lu_kmem_fini(mdt_caches);
7110 static void __exit mdt_exit(void)
7112 class_unregister_type(LUSTRE_MDT_NAME);
7114 lu_kmem_fini(mdt_caches);
7117 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
7118 MODULE_DESCRIPTION("Lustre Metadata Target ("LUSTRE_MDT_NAME")");
7119 MODULE_VERSION(LUSTRE_VERSION_STRING);
7120 MODULE_LICENSE("GPL");
7122 module_init(mdt_init);
7123 module_exit(mdt_exit);