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"
70 static unsigned int max_mod_rpcs_per_client = 8;
71 module_param(max_mod_rpcs_per_client, uint, 0644);
72 MODULE_PARM_DESC(max_mod_rpcs_per_client, "maximum number of modify RPCs in flight allowed per client");
74 mdl_mode_t mdt_mdl_lock_modes[] = {
75 [LCK_MINMODE] = MDL_MINMODE,
82 [LCK_GROUP] = MDL_GROUP
85 enum ldlm_mode mdt_dlm_lock_modes[] = {
86 [MDL_MINMODE] = LCK_MINMODE,
93 [MDL_GROUP] = LCK_GROUP
96 static struct mdt_device *mdt_dev(struct lu_device *d);
98 static const struct lu_object_operations mdt_obj_ops;
100 /* Slab for MDT object allocation */
101 static struct kmem_cache *mdt_object_kmem;
103 /* For HSM restore handles */
104 struct kmem_cache *mdt_hsm_cdt_kmem;
106 /* For HSM request handles */
107 struct kmem_cache *mdt_hsm_car_kmem;
109 static struct lu_kmem_descr mdt_caches[] = {
111 .ckd_cache = &mdt_object_kmem,
112 .ckd_name = "mdt_obj",
113 .ckd_size = sizeof(struct mdt_object)
116 .ckd_cache = &mdt_hsm_cdt_kmem,
117 .ckd_name = "mdt_cdt_restore_handle",
118 .ckd_size = sizeof(struct cdt_restore_handle)
121 .ckd_cache = &mdt_hsm_car_kmem,
122 .ckd_name = "mdt_cdt_agent_req",
123 .ckd_size = sizeof(struct cdt_agent_req)
130 __u64 mdt_get_disposition(struct ldlm_reply *rep, __u64 op_flag)
134 return rep->lock_policy_res1 & op_flag;
137 void mdt_clear_disposition(struct mdt_thread_info *info,
138 struct ldlm_reply *rep, __u64 op_flag)
141 info->mti_opdata &= ~op_flag;
142 tgt_opdata_clear(info->mti_env, op_flag);
145 rep->lock_policy_res1 &= ~op_flag;
148 void mdt_set_disposition(struct mdt_thread_info *info,
149 struct ldlm_reply *rep, __u64 op_flag)
152 info->mti_opdata |= op_flag;
153 tgt_opdata_set(info->mti_env, op_flag);
156 rep->lock_policy_res1 |= op_flag;
159 void mdt_lock_reg_init(struct mdt_lock_handle *lh, enum ldlm_mode lm)
161 lh->mlh_pdo_hash = 0;
162 lh->mlh_reg_mode = lm;
163 lh->mlh_rreg_mode = lm;
164 lh->mlh_type = MDT_REG_LOCK;
167 void mdt_lock_pdo_init(struct mdt_lock_handle *lh, enum ldlm_mode lock_mode,
168 const struct lu_name *lname)
170 lh->mlh_reg_mode = lock_mode;
171 lh->mlh_pdo_mode = LCK_MINMODE;
172 lh->mlh_rreg_mode = lock_mode;
173 lh->mlh_type = MDT_PDO_LOCK;
175 if (lu_name_is_valid(lname)) {
176 lh->mlh_pdo_hash = ll_full_name_hash(NULL, lname->ln_name,
178 /* XXX Workaround for LU-2856
180 * Zero is a valid return value of full_name_hash, but
181 * several users of mlh_pdo_hash assume a non-zero
182 * hash value. We therefore map zero onto an
183 * arbitrary, but consistent value (1) to avoid
184 * problems further down the road. */
185 if (unlikely(lh->mlh_pdo_hash == 0))
186 lh->mlh_pdo_hash = 1;
188 lh->mlh_pdo_hash = 0;
192 static void mdt_lock_pdo_mode(struct mdt_thread_info *info, struct mdt_object *o,
193 struct mdt_lock_handle *lh)
199 * Any dir access needs couple of locks:
201 * 1) on part of dir we gonna take lookup/modify;
203 * 2) on whole dir to protect it from concurrent splitting and/or to
204 * flush client's cache for readdir().
206 * so, for a given mode and object this routine decides what lock mode
207 * to use for lock #2:
209 * 1) if caller's gonna lookup in dir then we need to protect dir from
210 * being splitted only - LCK_CR
212 * 2) if caller's gonna modify dir then we need to protect dir from
213 * being splitted and to flush cache - LCK_CW
215 * 3) if caller's gonna modify dir and that dir seems ready for
216 * splitting then we need to protect it from any type of access
217 * (lookup/modify/split) - LCK_EX --bzzz
220 LASSERT(lh->mlh_reg_mode != LCK_MINMODE);
221 LASSERT(lh->mlh_pdo_mode == LCK_MINMODE);
224 * Ask underlaying level its opinion about preferable PDO lock mode
225 * having access type passed as regular lock mode:
227 * - MDL_MINMODE means that lower layer does not want to specify lock
230 * - MDL_NL means that no PDO lock should be taken. This is used in some
231 * cases. Say, for non-splittable directories no need to use PDO locks
234 mode = mdo_lock_mode(info->mti_env, mdt_object_child(o),
235 mdt_dlm_mode2mdl_mode(lh->mlh_reg_mode));
237 if (mode != MDL_MINMODE) {
238 lh->mlh_pdo_mode = mdt_mdl_mode2dlm_mode(mode);
241 * Lower layer does not want to specify locking mode. We do it
242 * our selves. No special protection is needed, just flush
243 * client's cache on modification and allow concurrent
246 switch (lh->mlh_reg_mode) {
248 lh->mlh_pdo_mode = LCK_EX;
251 lh->mlh_pdo_mode = LCK_CR;
254 lh->mlh_pdo_mode = LCK_CW;
257 CERROR("Not expected lock type (0x%x)\n",
258 (int)lh->mlh_reg_mode);
263 LASSERT(lh->mlh_pdo_mode != LCK_MINMODE);
267 static int mdt_lookup_fileset(struct mdt_thread_info *info, const char *fileset,
270 struct mdt_device *mdt = info->mti_mdt;
271 struct lu_name *lname = &info->mti_name;
272 char *filename = info->mti_filename;
273 struct mdt_object *parent;
277 LASSERT(!info->mti_cross_ref);
280 * We may want to allow this to mount a completely separate
281 * fileset from the MDT in the future, but keeping it to
282 * ROOT/ only for now avoid potential security issues.
284 *fid = mdt->mdt_md_root_fid;
286 while (rc == 0 && fileset != NULL && *fileset != '\0') {
287 const char *s1 = fileset;
293 while (*s2 != '/' && *s2 != '\0')
301 lname->ln_namelen = s2 - s1;
302 if (lname->ln_namelen > NAME_MAX) {
307 /* reject .. as a path component */
308 if (lname->ln_namelen == 2 &&
309 strncmp(s1, "..", 2) == 0) {
314 strncpy(filename, s1, lname->ln_namelen);
315 filename[lname->ln_namelen] = '\0';
316 lname->ln_name = filename;
318 parent = mdt_object_find(info->mti_env, mdt, fid);
319 if (IS_ERR(parent)) {
320 rc = PTR_ERR(parent);
323 /* Only got the fid of this obj by name */
325 rc = mdo_lookup(info->mti_env, mdt_object_child(parent), lname,
326 fid, &info->mti_spec);
327 mdt_object_put(info->mti_env, parent);
330 parent = mdt_object_find(info->mti_env, mdt, fid);
332 rc = PTR_ERR(parent);
334 mode = lu_object_attr(&parent->mot_obj);
335 mdt_object_put(info->mti_env, parent);
344 static int mdt_get_root(struct tgt_session_info *tsi)
346 struct mdt_thread_info *info = tsi2mdt_info(tsi);
347 struct mdt_device *mdt = info->mti_mdt;
348 struct mdt_body *repbody;
349 char *fileset = NULL, *buffer = NULL;
351 struct obd_export *exp = info->mti_exp;
352 char *nodemap_fileset;
356 rc = mdt_check_ucred(info);
358 GOTO(out, rc = err_serious(rc));
360 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GET_ROOT_PACK))
361 GOTO(out, rc = err_serious(-ENOMEM));
363 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
364 if (req_capsule_get_size(info->mti_pill, &RMF_NAME, RCL_CLIENT) > 0) {
365 fileset = req_capsule_client_get(info->mti_pill, &RMF_NAME);
367 GOTO(out, rc = err_serious(-EFAULT));
370 nodemap_fileset = nodemap_get_fileset(exp->exp_target_data.ted_nodemap);
371 if (nodemap_fileset && nodemap_fileset[0]) {
372 CDEBUG(D_INFO, "nodemap fileset is %s\n", nodemap_fileset);
374 /* consider fileset from client as a sub-fileset
375 * of the nodemap one */
376 OBD_ALLOC(buffer, PATH_MAX + 1);
378 GOTO(out, rc = err_serious(-ENOMEM));
379 if (snprintf(buffer, PATH_MAX + 1, "%s/%s",
380 nodemap_fileset, fileset) >= PATH_MAX + 1)
381 GOTO(out, rc = err_serious(-EINVAL));
384 /* enforce fileset as specified in the nodemap */
385 fileset = nodemap_fileset;
390 CDEBUG(D_INFO, "Getting fileset %s\n", fileset);
391 rc = mdt_lookup_fileset(info, fileset, &repbody->mbo_fid1);
393 GOTO(out, rc = err_serious(rc));
395 repbody->mbo_fid1 = mdt->mdt_md_root_fid;
397 repbody->mbo_valid |= OBD_MD_FLID;
401 mdt_thread_info_fini(info);
403 OBD_FREE(buffer, PATH_MAX+1);
407 static int mdt_statfs(struct tgt_session_info *tsi)
409 struct ptlrpc_request *req = tgt_ses_req(tsi);
410 struct mdt_thread_info *info = tsi2mdt_info(tsi);
411 struct mdt_device *mdt = info->mti_mdt;
412 struct tg_grants_data *tgd = &mdt->mdt_lut.lut_tgd;
413 struct md_device *next = mdt->mdt_child;
414 struct ptlrpc_service_part *svcpt;
415 struct obd_statfs *osfs;
416 struct mdt_body *reqbody = NULL;
417 struct mdt_statfs_cache *msf;
422 svcpt = req->rq_rqbd->rqbd_svcpt;
424 /* This will trigger a watchdog timeout */
425 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_STATFS_LCW_SLEEP,
426 (MDT_SERVICE_WATCHDOG_FACTOR *
427 at_get(&svcpt->scp_at_estimate)) + 1);
429 rc = mdt_check_ucred(info);
431 GOTO(out, rc = err_serious(rc));
433 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK))
434 GOTO(out, rc = err_serious(-ENOMEM));
436 osfs = req_capsule_server_get(info->mti_pill, &RMF_OBD_STATFS);
438 GOTO(out, rc = -EPROTO);
440 if (mdt_is_sum_statfs_client(req->rq_export))
441 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
443 if (reqbody && reqbody->mbo_valid & OBD_MD_FLAGSTATFS)
444 msf = &mdt->mdt_sum_osfs;
446 msf = &mdt->mdt_osfs;
448 if (msf->msf_age + OBD_STATFS_CACHE_SECONDS <= ktime_get_seconds()) {
449 /** statfs data is too old, get up-to-date one */
450 if (reqbody && reqbody->mbo_valid & OBD_MD_FLAGSTATFS)
451 rc = next->md_ops->mdo_statfs(info->mti_env,
454 rc = dt_statfs(info->mti_env, mdt->mdt_bottom,
458 spin_lock(&mdt->mdt_lock);
459 msf->msf_osfs = *osfs;
460 msf->msf_age = ktime_get_seconds();
461 spin_unlock(&mdt->mdt_lock);
463 /** use cached statfs data */
464 spin_lock(&mdt->mdt_lock);
465 *osfs = msf->msf_osfs;
466 spin_unlock(&mdt->mdt_lock);
469 /* at least try to account for cached pages. its still racy and
470 * might be under-reporting if clients haven't announced their
471 * caches with brw recently */
472 CDEBUG(D_SUPER | D_CACHE, "blocks cached %llu granted %llu"
473 " pending %llu free %llu avail %llu\n",
474 tgd->tgd_tot_dirty, tgd->tgd_tot_granted,
475 tgd->tgd_tot_pending,
476 osfs->os_bfree << tgd->tgd_blockbits,
477 osfs->os_bavail << tgd->tgd_blockbits);
479 osfs->os_bavail -= min_t(u64, osfs->os_bavail,
480 ((tgd->tgd_tot_dirty + tgd->tgd_tot_pending +
481 osfs->os_bsize - 1) >> tgd->tgd_blockbits));
483 tgt_grant_sanity_check(mdt->mdt_lu_dev.ld_obd, __func__);
484 CDEBUG(D_CACHE, "%llu blocks: %llu free, %llu avail; "
485 "%llu objects: %llu free; state %x\n",
486 osfs->os_blocks, osfs->os_bfree, osfs->os_bavail,
487 osfs->os_files, osfs->os_ffree, osfs->os_state);
489 if (!exp_grant_param_supp(tsi->tsi_exp) &&
490 tgd->tgd_blockbits > COMPAT_BSIZE_SHIFT) {
491 /* clients which don't support OBD_CONNECT_GRANT_PARAM
492 * should not see a block size > page size, otherwise
493 * cl_lost_grant goes mad. Therefore, we emulate a 4KB (=2^12)
494 * block size which is the biggest block size known to work
495 * with all client's page size. */
496 osfs->os_blocks <<= tgd->tgd_blockbits - COMPAT_BSIZE_SHIFT;
497 osfs->os_bfree <<= tgd->tgd_blockbits - COMPAT_BSIZE_SHIFT;
498 osfs->os_bavail <<= tgd->tgd_blockbits - COMPAT_BSIZE_SHIFT;
499 osfs->os_bsize = 1 << COMPAT_BSIZE_SHIFT;
502 mdt_counter_incr(req, LPROC_MDT_STATFS);
504 mdt_thread_info_fini(info);
509 * Pack size attributes into the reply.
511 int mdt_pack_size2body(struct mdt_thread_info *info,
512 const struct lu_fid *fid, struct lustre_handle *lh)
515 struct md_attr *ma = &info->mti_attr;
517 bool dom_lock = false;
521 LASSERT(ma->ma_attr.la_valid & LA_MODE);
523 if (!S_ISREG(ma->ma_attr.la_mode) ||
524 !(ma->ma_valid & MA_LOV && ma->ma_lmm != NULL))
527 dom_stripe = mdt_lmm_dom_entry(ma->ma_lmm);
528 /* no DoM stripe, no size in reply */
529 if (dom_stripe == LMM_NO_DOM)
532 if (lustre_handle_is_used(lh)) {
533 struct ldlm_lock *lock;
535 lock = ldlm_handle2lock(lh);
537 dom_lock = ldlm_has_dom(lock);
542 /* no DoM lock, no size in reply */
546 /* Either DoM lock exists or LMM has only DoM stripe then
547 * return size on body. */
548 b = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
550 mdt_dom_object_size(info->mti_env, info->mti_mdt, fid, b, dom_lock);
554 #ifdef CONFIG_FS_POSIX_ACL
556 * Pack ACL data into the reply. UIDs/GIDs are mapped and filtered by nodemap.
558 * \param info thread info object
559 * \param repbody reply to pack ACLs into
560 * \param o mdt object of file to examine
561 * \param nodemap nodemap of client to reply to
563 * \retval -errno error getting or parsing ACL from disk
565 int mdt_pack_acl2body(struct mdt_thread_info *info, struct mdt_body *repbody,
566 struct mdt_object *o, struct lu_nodemap *nodemap)
568 const struct lu_env *env = info->mti_env;
569 struct md_object *next = mdt_object_child(o);
570 struct lu_buf *buf = &info->mti_buf;
571 struct mdt_device *mdt = info->mti_mdt;
572 struct req_capsule *pill = info->mti_pill;
577 buf->lb_buf = req_capsule_server_get(pill, &RMF_ACL);
578 buf->lb_len = req_capsule_get_size(pill, &RMF_ACL, RCL_SERVER);
579 if (buf->lb_len == 0)
583 rc = mo_xattr_get(env, next, buf, XATTR_NAME_ACL_ACCESS);
585 if (rc == -ENODATA) {
586 repbody->mbo_aclsize = 0;
587 repbody->mbo_valid |= OBD_MD_FLACL;
589 } else if (rc == -EOPNOTSUPP) {
593 exp_connect_large_acl(info->mti_exp) &&
594 buf->lb_buf != info->mti_big_acl) {
595 if (info->mti_big_acl == NULL) {
596 OBD_ALLOC_LARGE(info->mti_big_acl,
597 mdt->mdt_max_ea_size);
598 if (info->mti_big_acl == NULL) {
599 CERROR("%s: unable to grow "
602 PFID(mdt_object_fid(o)));
606 info->mti_big_aclsize =
607 mdt->mdt_max_ea_size;
610 CDEBUG(D_INODE, "%s: grow the "DFID
611 " ACL buffer to size %d\n",
613 PFID(mdt_object_fid(o)),
614 mdt->mdt_max_ea_size);
616 buf->lb_buf = info->mti_big_acl;
617 buf->lb_len = info->mti_big_aclsize;
622 CERROR("%s: unable to read "DFID" ACL: rc = %d\n",
623 mdt_obd_name(mdt), PFID(mdt_object_fid(o)), rc);
632 acl_buflen = req_capsule_get_size(pill, &RMF_ACL, RCL_SERVER);
633 if (acl_buflen >= rc)
636 /* If LOV/LMA EA is small, we can reuse part of their buffer */
637 client = ptlrpc_req_get_repsize(pill->rc_req);
638 server = lustre_packed_msg_size(pill->rc_req->rq_repmsg);
639 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER)) {
640 lmm_buflen = req_capsule_get_size(pill, &RMF_MDT_MD,
642 lmmsize = repbody->mbo_eadatasize;
645 if (client < server - acl_buflen - lmm_buflen + rc + lmmsize) {
646 CDEBUG(D_INODE, "%s: client prepared buffer size %d "
647 "is not big enough with the ACL size %d (%d)\n",
648 mdt_obd_name(mdt), client, rc,
649 server - acl_buflen - lmm_buflen + rc + lmmsize);
650 repbody->mbo_aclsize = 0;
651 repbody->mbo_valid &= ~OBD_MD_FLACL;
656 if (buf->lb_buf == info->mti_big_acl)
657 info->mti_big_acl_used = 1;
659 rc = nodemap_map_acl(nodemap, buf->lb_buf,
660 rc, NODEMAP_FS_TO_CLIENT);
661 /* if all ACLs mapped out, rc is still >= 0 */
663 CERROR("%s: nodemap_map_acl unable to parse "DFID
664 " ACL: rc = %d\n", mdt_obd_name(mdt),
665 PFID(mdt_object_fid(o)), rc);
666 repbody->mbo_aclsize = 0;
667 repbody->mbo_valid &= ~OBD_MD_FLACL;
669 repbody->mbo_aclsize = rc;
670 repbody->mbo_valid |= OBD_MD_FLACL;
679 /* XXX Look into layout in MDT layer. */
680 static inline bool mdt_hsm_is_released(struct lov_mds_md *lmm)
682 struct lov_comp_md_v1 *comp_v1;
683 struct lov_mds_md *v1;
686 if (lmm->lmm_magic == LOV_MAGIC_COMP_V1) {
687 comp_v1 = (struct lov_comp_md_v1 *)lmm;
689 for (i = 0; i < comp_v1->lcm_entry_count; i++) {
690 v1 = (struct lov_mds_md *)((char *)comp_v1 +
691 comp_v1->lcm_entries[i].lcme_offset);
692 /* We don't support partial release for now */
693 if (!(v1->lmm_pattern & LOV_PATTERN_F_RELEASED))
698 return (lmm->lmm_pattern & LOV_PATTERN_F_RELEASED) ?
703 void mdt_pack_attr2body(struct mdt_thread_info *info, struct mdt_body *b,
704 const struct lu_attr *attr, const struct lu_fid *fid)
706 struct md_attr *ma = &info->mti_attr;
707 struct obd_export *exp = info->mti_exp;
708 struct lu_nodemap *nodemap = NULL;
710 LASSERT(ma->ma_valid & MA_INODE);
712 if (attr->la_valid & LA_ATIME) {
713 b->mbo_atime = attr->la_atime;
714 b->mbo_valid |= OBD_MD_FLATIME;
716 if (attr->la_valid & LA_MTIME) {
717 b->mbo_mtime = attr->la_mtime;
718 b->mbo_valid |= OBD_MD_FLMTIME;
720 if (attr->la_valid & LA_CTIME) {
721 b->mbo_ctime = attr->la_ctime;
722 b->mbo_valid |= OBD_MD_FLCTIME;
724 if (attr->la_valid & LA_FLAGS) {
725 b->mbo_flags = attr->la_flags;
726 b->mbo_valid |= OBD_MD_FLFLAGS;
728 if (attr->la_valid & LA_NLINK) {
729 b->mbo_nlink = attr->la_nlink;
730 b->mbo_valid |= OBD_MD_FLNLINK;
732 if (attr->la_valid & (LA_UID|LA_GID)) {
733 nodemap = nodemap_get_from_exp(exp);
737 if (attr->la_valid & LA_UID) {
738 b->mbo_uid = nodemap_map_id(nodemap, NODEMAP_UID,
739 NODEMAP_FS_TO_CLIENT,
741 b->mbo_valid |= OBD_MD_FLUID;
743 if (attr->la_valid & LA_GID) {
744 b->mbo_gid = nodemap_map_id(nodemap, NODEMAP_GID,
745 NODEMAP_FS_TO_CLIENT,
747 b->mbo_valid |= OBD_MD_FLGID;
750 if (attr->la_valid & LA_PROJID) {
751 /* TODO, nodemap for project id */
752 b->mbo_projid = attr->la_projid;
753 b->mbo_valid |= OBD_MD_FLPROJID;
756 b->mbo_mode = attr->la_mode;
757 if (attr->la_valid & LA_MODE)
758 b->mbo_valid |= OBD_MD_FLMODE;
759 if (attr->la_valid & LA_TYPE)
760 b->mbo_valid |= OBD_MD_FLTYPE;
764 b->mbo_valid |= OBD_MD_FLID;
765 CDEBUG(D_INODE, DFID": nlink=%d, mode=%o, valid=%#llx\n",
766 PFID(fid), b->mbo_nlink, b->mbo_mode, b->mbo_valid);
769 if (!(attr->la_valid & LA_TYPE))
772 b->mbo_rdev = attr->la_rdev;
773 b->mbo_size = attr->la_size;
774 b->mbo_blocks = attr->la_blocks;
776 if (!S_ISREG(attr->la_mode)) {
777 b->mbo_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS | OBD_MD_FLRDEV;
778 } else if (ma->ma_need & MA_LOV && !(ma->ma_valid & MA_LOV)) {
779 /* means no objects are allocated on osts. */
780 LASSERT(!(ma->ma_valid & MA_LOV));
781 /* just ignore blocks occupied by extend attributes on MDS */
783 /* if no object is allocated on osts, the size on mds is valid.
785 b->mbo_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
786 } else if ((ma->ma_valid & MA_LOV) && ma->ma_lmm != NULL) {
787 if (mdt_hsm_is_released(ma->ma_lmm)) {
788 /* A released file stores its size on MDS. */
789 /* But return 1 block for released file, unless tools
790 * like tar will consider it fully sparse. (LU-3864)
792 if (unlikely(b->mbo_size == 0))
796 b->mbo_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
797 } else if (info->mti_som_valid) { /* som is valid */
798 b->mbo_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
802 if (fid != NULL && (b->mbo_valid & OBD_MD_FLSIZE))
803 CDEBUG(D_VFSTRACE, DFID": returning size %llu\n",
804 PFID(fid), (unsigned long long)b->mbo_size);
807 if (!IS_ERR_OR_NULL(nodemap))
808 nodemap_putref(nodemap);
811 static inline int mdt_body_has_lov(const struct lu_attr *la,
812 const struct mdt_body *body)
814 return (S_ISREG(la->la_mode) && (body->mbo_valid & OBD_MD_FLEASIZE)) ||
815 (S_ISDIR(la->la_mode) && (body->mbo_valid & OBD_MD_FLDIREA));
818 void mdt_client_compatibility(struct mdt_thread_info *info)
820 struct mdt_body *body;
821 struct ptlrpc_request *req = mdt_info_req(info);
822 struct obd_export *exp = req->rq_export;
823 struct md_attr *ma = &info->mti_attr;
824 struct lu_attr *la = &ma->ma_attr;
827 if (exp_connect_layout(exp))
828 /* the client can deal with 16-bit lmm_stripe_count */
831 body = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
833 if (!mdt_body_has_lov(la, body))
836 /* now we have a reply with a lov for a client not compatible with the
837 * layout lock so we have to clean the layout generation number */
838 if (S_ISREG(la->la_mode))
839 ma->ma_lmm->lmm_layout_gen = 0;
843 static int mdt_attr_get_eabuf_size(struct mdt_thread_info *info,
844 struct mdt_object *o)
846 const struct lu_env *env = info->mti_env;
849 rc = mo_xattr_get(env, mdt_object_child(o), &LU_BUF_NULL,
858 /* Is it a directory? Let's check for the LMV as well */
859 if (S_ISDIR(lu_object_attr(&mdt_object_child(o)->mo_lu))) {
860 rc2 = mo_xattr_get(env, mdt_object_child(o), &LU_BUF_NULL,
864 rc2 = mo_xattr_get(env, mdt_object_child(o),
866 XATTR_NAME_DEFAULT_LMV);
868 if ((rc2 < 0 && rc2 != -ENODATA) || (rc2 > rc))
876 int mdt_big_xattr_get(struct mdt_thread_info *info, struct mdt_object *o,
879 const struct lu_env *env = info->mti_env;
883 LASSERT(info->mti_big_lmm_used == 0);
884 rc = mo_xattr_get(env, mdt_object_child(o), &LU_BUF_NULL, name);
888 /* big_lmm may need to be grown */
889 if (info->mti_big_lmmsize < rc) {
890 int size = size_roundup_power2(rc);
892 if (info->mti_big_lmmsize > 0) {
893 /* free old buffer */
894 LASSERT(info->mti_big_lmm);
895 OBD_FREE_LARGE(info->mti_big_lmm,
896 info->mti_big_lmmsize);
897 info->mti_big_lmm = NULL;
898 info->mti_big_lmmsize = 0;
901 OBD_ALLOC_LARGE(info->mti_big_lmm, size);
902 if (info->mti_big_lmm == NULL)
904 info->mti_big_lmmsize = size;
906 LASSERT(info->mti_big_lmmsize >= rc);
908 info->mti_buf.lb_buf = info->mti_big_lmm;
909 info->mti_buf.lb_len = info->mti_big_lmmsize;
910 rc = mo_xattr_get(env, mdt_object_child(o), &info->mti_buf, name);
915 int mdt_stripe_get(struct mdt_thread_info *info, struct mdt_object *o,
916 struct md_attr *ma, const char *name)
918 struct md_object *next = mdt_object_child(o);
919 struct lu_buf *buf = &info->mti_buf;
922 if (strcmp(name, XATTR_NAME_LOV) == 0) {
923 buf->lb_buf = ma->ma_lmm;
924 buf->lb_len = ma->ma_lmm_size;
925 LASSERT(!(ma->ma_valid & MA_LOV));
926 } else if (strcmp(name, XATTR_NAME_LMV) == 0) {
927 buf->lb_buf = ma->ma_lmv;
928 buf->lb_len = ma->ma_lmv_size;
929 LASSERT(!(ma->ma_valid & MA_LMV));
930 } else if (strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0) {
931 buf->lb_buf = ma->ma_lmv;
932 buf->lb_len = ma->ma_lmv_size;
933 LASSERT(!(ma->ma_valid & MA_LMV_DEF));
938 LASSERT(buf->lb_buf);
940 rc = mo_xattr_get(info->mti_env, next, buf, name);
944 if (strcmp(name, XATTR_NAME_LOV) == 0) {
945 if (info->mti_big_lmm_used)
946 ma->ma_lmm = info->mti_big_lmm;
948 /* NOT return LOV EA with hole to old client. */
949 if (unlikely(le32_to_cpu(ma->ma_lmm->lmm_pattern) &
950 LOV_PATTERN_F_HOLE) &&
951 !(exp_connect_flags(info->mti_exp) &
952 OBD_CONNECT_LFSCK)) {
955 ma->ma_lmm_size = rc;
956 ma->ma_valid |= MA_LOV;
958 } else if (strcmp(name, XATTR_NAME_LMV) == 0) {
959 if (info->mti_big_lmm_used)
960 ma->ma_lmv = info->mti_big_lmm;
962 ma->ma_lmv_size = rc;
963 ma->ma_valid |= MA_LMV;
964 } else if (strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0) {
965 ma->ma_lmv_size = rc;
966 ma->ma_valid |= MA_LMV_DEF;
969 /* Update mdt_max_mdsize so all clients will be aware that */
970 if (info->mti_mdt->mdt_max_mdsize < rc)
971 info->mti_mdt->mdt_max_mdsize = rc;
974 } else if (rc == -ENODATA) {
977 } else if (rc == -ERANGE) {
978 /* Default LMV has fixed size, so it must be able to fit
979 * in the original buffer */
980 if (strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0)
982 rc = mdt_big_xattr_get(info, o, name);
984 info->mti_big_lmm_used = 1;
992 int mdt_attr_get_pfid(struct mdt_thread_info *info, struct mdt_object *o,
995 struct lu_buf *buf = &info->mti_buf;
996 struct link_ea_header *leh;
997 struct link_ea_entry *lee;
1001 buf->lb_buf = info->mti_big_lmm;
1002 buf->lb_len = info->mti_big_lmmsize;
1003 rc = mo_xattr_get(info->mti_env, mdt_object_child(o),
1004 buf, XATTR_NAME_LINK);
1005 /* ignore errors, MA_PFID won't be set and it is
1006 * up to the caller to treat this as an error */
1007 if (rc == -ERANGE || buf->lb_len == 0) {
1008 rc = mdt_big_xattr_get(info, o, XATTR_NAME_LINK);
1009 buf->lb_buf = info->mti_big_lmm;
1010 buf->lb_len = info->mti_big_lmmsize;
1015 if (rc < sizeof(*leh)) {
1016 CERROR("short LinkEA on "DFID": rc = %d\n",
1017 PFID(mdt_object_fid(o)), rc);
1021 leh = (struct link_ea_header *) buf->lb_buf;
1022 lee = (struct link_ea_entry *)(leh + 1);
1023 if (leh->leh_magic == __swab32(LINK_EA_MAGIC)) {
1024 leh->leh_magic = LINK_EA_MAGIC;
1025 leh->leh_reccount = __swab32(leh->leh_reccount);
1026 leh->leh_len = __swab64(leh->leh_len);
1028 if (leh->leh_magic != LINK_EA_MAGIC)
1030 if (leh->leh_reccount == 0)
1033 memcpy(pfid, &lee->lee_parent_fid, sizeof(*pfid));
1034 fid_be_to_cpu(pfid, pfid);
1039 int mdt_attr_get_complex(struct mdt_thread_info *info,
1040 struct mdt_object *o, struct md_attr *ma)
1042 const struct lu_env *env = info->mti_env;
1043 struct md_object *next = mdt_object_child(o);
1044 struct lu_buf *buf = &info->mti_buf;
1045 int need = ma->ma_need;
1052 if (mdt_object_exists(o) == 0)
1053 GOTO(out, rc = -ENOENT);
1054 mode = lu_object_attr(&next->mo_lu);
1056 if (need & MA_INODE) {
1057 ma->ma_need = MA_INODE;
1058 rc = mo_attr_get(env, next, ma);
1063 (void) mdt_get_som(info, o, ma);
1064 ma->ma_valid |= MA_INODE;
1067 if (need & MA_PFID) {
1068 rc = mdt_attr_get_pfid(info, o, &ma->ma_pfid);
1070 ma->ma_valid |= MA_PFID;
1071 /* ignore this error, parent fid is not mandatory */
1075 if (need & MA_LOV && (S_ISREG(mode) || S_ISDIR(mode))) {
1076 rc = mdt_stripe_get(info, o, ma, XATTR_NAME_LOV);
1081 if (need & MA_LMV && S_ISDIR(mode)) {
1082 rc = mdt_stripe_get(info, o, ma, XATTR_NAME_LMV);
1087 if (need & MA_LMV_DEF && S_ISDIR(mode)) {
1088 rc = mdt_stripe_get(info, o, ma, XATTR_NAME_DEFAULT_LMV);
1094 * In the handle of MA_INODE, we may already get the SOM attr.
1096 if (need & MA_SOM && S_ISREG(mode) && !(ma->ma_valid & MA_SOM)) {
1097 rc = mdt_get_som(info, o, ma);
1102 if (need & MA_HSM && S_ISREG(mode)) {
1103 buf->lb_buf = info->mti_xattr_buf;
1104 buf->lb_len = sizeof(info->mti_xattr_buf);
1105 CLASSERT(sizeof(struct hsm_attrs) <=
1106 sizeof(info->mti_xattr_buf));
1107 rc2 = mo_xattr_get(info->mti_env, next, buf, XATTR_NAME_HSM);
1108 rc2 = lustre_buf2hsm(info->mti_xattr_buf, rc2, &ma->ma_hsm);
1110 ma->ma_valid |= MA_HSM;
1111 else if (rc2 < 0 && rc2 != -ENODATA)
1112 GOTO(out, rc = rc2);
1115 #ifdef CONFIG_FS_POSIX_ACL
1116 if (need & MA_ACL_DEF && S_ISDIR(mode)) {
1117 buf->lb_buf = ma->ma_acl;
1118 buf->lb_len = ma->ma_acl_size;
1119 rc2 = mo_xattr_get(env, next, buf, XATTR_NAME_ACL_DEFAULT);
1121 ma->ma_acl_size = rc2;
1122 ma->ma_valid |= MA_ACL_DEF;
1123 } else if (rc2 == -ENODATA) {
1125 ma->ma_acl_size = 0;
1127 GOTO(out, rc = rc2);
1132 CDEBUG(D_INODE, "after getattr rc = %d, ma_valid = %#llx ma_lmm=%p\n",
1133 rc, ma->ma_valid, ma->ma_lmm);
1137 static int mdt_getattr_internal(struct mdt_thread_info *info,
1138 struct mdt_object *o, int ma_need)
1140 struct md_object *next = mdt_object_child(o);
1141 const struct mdt_body *reqbody = info->mti_body;
1142 struct ptlrpc_request *req = mdt_info_req(info);
1143 struct md_attr *ma = &info->mti_attr;
1144 struct lu_attr *la = &ma->ma_attr;
1145 struct req_capsule *pill = info->mti_pill;
1146 const struct lu_env *env = info->mti_env;
1147 struct mdt_body *repbody;
1148 struct lu_buf *buffer = &info->mti_buf;
1149 struct obd_export *exp = info->mti_exp;
1153 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK))
1154 RETURN(err_serious(-ENOMEM));
1156 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
1160 if (mdt_object_remote(o)) {
1161 /* This object is located on remote node.*/
1162 /* Return -ENOTSUPP for old client */
1163 if (!mdt_is_dne_client(req->rq_export))
1164 GOTO(out, rc = -ENOTSUPP);
1166 repbody->mbo_fid1 = *mdt_object_fid(o);
1167 repbody->mbo_valid = OBD_MD_FLID | OBD_MD_MDS;
1171 if (reqbody->mbo_eadatasize > 0) {
1172 buffer->lb_buf = req_capsule_server_get(pill, &RMF_MDT_MD);
1173 if (buffer->lb_buf == NULL)
1174 GOTO(out, rc = -EPROTO);
1175 buffer->lb_len = req_capsule_get_size(pill, &RMF_MDT_MD,
1178 buffer->lb_buf = NULL;
1180 ma_need &= ~(MA_LOV | MA_LMV);
1181 CDEBUG(D_INFO, "%s: RPC from %s: does not need LOVEA.\n",
1182 mdt_obd_name(info->mti_mdt),
1183 req->rq_export->exp_client_uuid.uuid);
1186 /* If it is dir object and client require MEA, then we got MEA */
1187 if (S_ISDIR(lu_object_attr(&next->mo_lu)) &&
1188 (reqbody->mbo_valid & (OBD_MD_MEA | OBD_MD_DEFAULT_MEA))) {
1189 /* Assumption: MDT_MD size is enough for lmv size. */
1190 ma->ma_lmv = buffer->lb_buf;
1191 ma->ma_lmv_size = buffer->lb_len;
1192 ma->ma_need = MA_INODE;
1193 if (ma->ma_lmv_size > 0) {
1194 if (reqbody->mbo_valid & OBD_MD_MEA)
1195 ma->ma_need |= MA_LMV;
1196 else if (reqbody->mbo_valid & OBD_MD_DEFAULT_MEA)
1197 ma->ma_need |= MA_LMV_DEF;
1200 ma->ma_lmm = buffer->lb_buf;
1201 ma->ma_lmm_size = buffer->lb_len;
1202 ma->ma_need = MA_INODE | MA_HSM;
1203 if (ma->ma_lmm_size > 0)
1204 ma->ma_need |= MA_LOV;
1207 if (S_ISDIR(lu_object_attr(&next->mo_lu)) &&
1208 reqbody->mbo_valid & OBD_MD_FLDIREA &&
1209 lustre_msg_get_opc(req->rq_reqmsg) == MDS_GETATTR) {
1210 /* get default stripe info for this dir. */
1211 ma->ma_need |= MA_LOV_DEF;
1213 ma->ma_need |= ma_need;
1215 rc = mdt_attr_get_complex(info, o, ma);
1217 CDEBUG(rc == -ENOENT ? D_OTHER : D_ERROR,
1218 "%s: getattr error for "DFID": rc = %d\n",
1219 mdt_obd_name(info->mti_mdt),
1220 PFID(mdt_object_fid(o)), rc);
1224 /* if file is released, check if a restore is running */
1225 if (ma->ma_valid & MA_HSM) {
1226 repbody->mbo_valid |= OBD_MD_TSTATE;
1227 if ((ma->ma_hsm.mh_flags & HS_RELEASED) &&
1228 mdt_hsm_restore_is_running(info, mdt_object_fid(o)))
1229 repbody->mbo_t_state = MS_RESTORE;
1232 if (likely(ma->ma_valid & MA_INODE))
1233 mdt_pack_attr2body(info, repbody, la, mdt_object_fid(o));
1237 if (mdt_body_has_lov(la, reqbody)) {
1238 if (ma->ma_valid & MA_LOV) {
1239 LASSERT(ma->ma_lmm_size);
1240 repbody->mbo_eadatasize = ma->ma_lmm_size;
1241 if (S_ISDIR(la->la_mode))
1242 repbody->mbo_valid |= OBD_MD_FLDIREA;
1244 repbody->mbo_valid |= OBD_MD_FLEASIZE;
1245 mdt_dump_lmm(D_INFO, ma->ma_lmm, repbody->mbo_valid);
1247 if (ma->ma_valid & MA_LMV) {
1248 /* Return -ENOTSUPP for old client */
1249 if (!mdt_is_striped_client(req->rq_export))
1252 LASSERT(S_ISDIR(la->la_mode));
1253 mdt_dump_lmv(D_INFO, ma->ma_lmv);
1254 repbody->mbo_eadatasize = ma->ma_lmv_size;
1255 repbody->mbo_valid |= (OBD_MD_FLDIREA|OBD_MD_MEA);
1257 if (ma->ma_valid & MA_LMV_DEF) {
1258 /* Return -ENOTSUPP for old client */
1259 if (!mdt_is_striped_client(req->rq_export))
1261 LASSERT(S_ISDIR(la->la_mode));
1262 mdt_dump_lmv(D_INFO, ma->ma_lmv);
1263 repbody->mbo_eadatasize = ma->ma_lmv_size;
1264 repbody->mbo_valid |= (OBD_MD_FLDIREA |
1265 OBD_MD_DEFAULT_MEA);
1267 } else if (S_ISLNK(la->la_mode) &&
1268 reqbody->mbo_valid & OBD_MD_LINKNAME) {
1269 buffer->lb_buf = ma->ma_lmm;
1270 /* eadatasize from client includes NULL-terminator, so
1271 * there is no need to read it */
1272 buffer->lb_len = reqbody->mbo_eadatasize - 1;
1273 rc = mo_readlink(env, next, buffer);
1274 if (unlikely(rc <= 0)) {
1275 CERROR("%s: readlink failed for "DFID": rc = %d\n",
1276 mdt_obd_name(info->mti_mdt),
1277 PFID(mdt_object_fid(o)), rc);
1280 int print_limit = min_t(int, PAGE_SIZE - 128, rc);
1282 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_READLINK_EPROTO))
1284 repbody->mbo_valid |= OBD_MD_LINKNAME;
1285 /* we need to report back size with NULL-terminator
1286 * because client expects that */
1287 repbody->mbo_eadatasize = rc + 1;
1288 if (repbody->mbo_eadatasize != reqbody->mbo_eadatasize)
1289 CDEBUG(D_INODE, "%s: Read shorter symlink %d "
1290 "on "DFID ", expected %d\n",
1291 mdt_obd_name(info->mti_mdt),
1292 rc, PFID(mdt_object_fid(o)),
1293 reqbody->mbo_eadatasize - 1);
1294 /* NULL terminate */
1295 ((char *)ma->ma_lmm)[rc] = 0;
1297 /* If the total CDEBUG() size is larger than a page, it
1298 * will print a warning to the console, avoid this by
1299 * printing just the last part of the symlink. */
1300 CDEBUG(D_INODE, "symlink dest %s%.*s, len = %d\n",
1301 print_limit < rc ? "..." : "", print_limit,
1302 (char *)ma->ma_lmm + rc - print_limit, rc);
1307 if (reqbody->mbo_valid & OBD_MD_FLMODEASIZE) {
1308 repbody->mbo_max_mdsize = info->mti_mdt->mdt_max_mdsize;
1309 repbody->mbo_valid |= OBD_MD_FLMODEASIZE;
1310 CDEBUG(D_INODE, "changing the max MD size to %u\n",
1311 repbody->mbo_max_mdsize);
1314 #ifdef CONFIG_FS_POSIX_ACL
1315 if ((exp_connect_flags(req->rq_export) & OBD_CONNECT_ACL) &&
1316 (reqbody->mbo_valid & OBD_MD_FLACL)) {
1317 struct lu_nodemap *nodemap = nodemap_get_from_exp(exp);
1318 if (IS_ERR(nodemap))
1319 RETURN(PTR_ERR(nodemap));
1321 rc = mdt_pack_acl2body(info, repbody, o, nodemap);
1322 nodemap_putref(nodemap);
1328 mdt_counter_incr(req, LPROC_MDT_GETATTR);
1333 static int mdt_getattr(struct tgt_session_info *tsi)
1335 struct mdt_thread_info *info = tsi2mdt_info(tsi);
1336 struct mdt_object *obj = info->mti_object;
1337 struct req_capsule *pill = info->mti_pill;
1338 struct mdt_body *reqbody;
1339 struct mdt_body *repbody;
1343 reqbody = req_capsule_client_get(pill, &RMF_MDT_BODY);
1345 LASSERT(obj != NULL);
1346 LASSERT(lu_object_assert_exists(&obj->mot_obj));
1348 /* Special case for Data-on-MDT files to get data version */
1349 if (unlikely(reqbody->mbo_valid & OBD_MD_FLDATAVERSION)) {
1350 rc = mdt_data_version_get(tsi);
1354 /* Unlike intent case where we need to pre-fill out buffers early on
1355 * in intent policy for ldlm reasons, here we can have a much better
1356 * guess at EA size by just reading it from disk.
1357 * Exceptions are readdir and (missing) directory striping */
1359 if (reqbody->mbo_valid & OBD_MD_LINKNAME) {
1360 /* No easy way to know how long is the symlink, but it cannot
1361 * be more than PATH_MAX, so we allocate +1 */
1363 /* A special case for fs ROOT: getattr there might fetch
1364 * default EA for entire fs, not just for this dir!
1366 } else if (lu_fid_eq(mdt_object_fid(obj),
1367 &info->mti_mdt->mdt_md_root_fid) &&
1368 (reqbody->mbo_valid & OBD_MD_FLDIREA) &&
1369 (lustre_msg_get_opc(mdt_info_req(info)->rq_reqmsg) ==
1371 /* Should the default strping be bigger, mdt_fix_reply
1372 * will reallocate */
1373 rc = DEF_REP_MD_SIZE;
1375 /* Read the actual EA size from disk */
1376 rc = mdt_attr_get_eabuf_size(info, obj);
1380 GOTO(out, rc = err_serious(rc));
1382 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER, rc);
1384 /* Set ACL reply buffer size as LUSTRE_POSIX_ACL_MAX_SIZE_OLD
1385 * by default. If the target object has more ACL entries, then
1386 * enlarge the buffer when necessary. */
1387 req_capsule_set_size(pill, &RMF_ACL, RCL_SERVER,
1388 LUSTRE_POSIX_ACL_MAX_SIZE_OLD);
1390 rc = req_capsule_server_pack(pill);
1391 if (unlikely(rc != 0))
1392 GOTO(out, rc = err_serious(rc));
1394 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
1395 LASSERT(repbody != NULL);
1396 repbody->mbo_eadatasize = 0;
1397 repbody->mbo_aclsize = 0;
1399 rc = mdt_check_ucred(info);
1401 GOTO(out_shrink, rc);
1403 info->mti_cross_ref = !!(reqbody->mbo_valid & OBD_MD_FLCROSSREF);
1405 rc = mdt_getattr_internal(info, obj, 0);
1408 mdt_client_compatibility(info);
1409 rc2 = mdt_fix_reply(info);
1413 mdt_thread_info_fini(info);
1418 * Handler of layout intent RPC requiring the layout modification
1420 * \param[in] info thread environment
1421 * \param[in] obj object
1422 * \param[in] layout layout change descriptor
1424 * \retval 0 on success
1425 * \retval < 0 error code
1427 int mdt_layout_change(struct mdt_thread_info *info, struct mdt_object *obj,
1428 struct md_layout_change *layout)
1430 struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LOCAL];
1434 if (!mdt_object_exists(obj))
1435 GOTO(out, rc = -ENOENT);
1437 if (!S_ISREG(lu_object_attr(&obj->mot_obj)))
1438 GOTO(out, rc = -EINVAL);
1440 rc = mo_permission(info->mti_env, NULL, mdt_object_child(obj), NULL,
1445 /* take layout lock to prepare layout change */
1446 mdt_lock_reg_init(lh, LCK_EX);
1447 rc = mdt_object_lock(info, obj, lh, MDS_INODELOCK_LAYOUT);
1451 mutex_lock(&obj->mot_som_mutex);
1452 rc = mo_layout_change(info->mti_env, mdt_object_child(obj), layout);
1453 mutex_unlock(&obj->mot_som_mutex);
1454 mdt_object_unlock(info, obj, lh, 1);
1460 * Exchange MOF_LOV_CREATED flags between two objects after a
1461 * layout swap. No assumption is made on whether o1 or o2 have
1462 * created objects or not.
1464 * \param[in,out] o1 First swap layout object
1465 * \param[in,out] o2 Second swap layout object
1467 static void mdt_swap_lov_flag(struct mdt_object *o1, struct mdt_object *o2)
1469 unsigned int o1_lov_created = o1->mot_lov_created;
1471 mutex_lock(&o1->mot_lov_mutex);
1472 mutex_lock(&o2->mot_lov_mutex);
1474 o1->mot_lov_created = o2->mot_lov_created;
1475 o2->mot_lov_created = o1_lov_created;
1477 mutex_unlock(&o2->mot_lov_mutex);
1478 mutex_unlock(&o1->mot_lov_mutex);
1481 static int mdt_swap_layouts(struct tgt_session_info *tsi)
1483 struct mdt_thread_info *info;
1484 struct ptlrpc_request *req = tgt_ses_req(tsi);
1485 struct obd_export *exp = req->rq_export;
1486 struct mdt_object *o1, *o2, *o;
1487 struct mdt_lock_handle *lh1, *lh2;
1488 struct mdc_swap_layouts *msl;
1492 /* client does not support layout lock, so layout swaping
1494 * FIXME: there is a problem for old clients which don't support
1495 * layout lock yet. If those clients have already opened the file
1496 * they won't be notified at all so that old layout may still be
1497 * used to do IO. This can be fixed after file release is landed by
1498 * doing exclusive open and taking full EX ibits lock. - Jinshan */
1499 if (!exp_connect_layout(exp))
1500 RETURN(-EOPNOTSUPP);
1502 info = tsi2mdt_info(tsi);
1504 if (info->mti_dlm_req != NULL)
1505 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1507 o1 = info->mti_object;
1508 o = o2 = mdt_object_find(info->mti_env, info->mti_mdt,
1509 &info->mti_body->mbo_fid2);
1511 GOTO(out, rc = PTR_ERR(o));
1513 if (mdt_object_remote(o) || !mdt_object_exists(o)) /* remote object */
1514 GOTO(put, rc = -ENOENT);
1516 rc = lu_fid_cmp(&info->mti_body->mbo_fid1, &info->mti_body->mbo_fid2);
1517 if (unlikely(rc == 0)) /* same file, you kidding me? no-op. */
1523 /* permission check. Make sure the calling process having permission
1524 * to write both files. */
1525 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o1), NULL,
1530 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o2), NULL,
1535 msl = req_capsule_client_get(info->mti_pill, &RMF_SWAP_LAYOUTS);
1537 GOTO(put, rc = -EPROTO);
1539 lh1 = &info->mti_lh[MDT_LH_NEW];
1540 mdt_lock_reg_init(lh1, LCK_EX);
1541 lh2 = &info->mti_lh[MDT_LH_OLD];
1542 mdt_lock_reg_init(lh2, LCK_EX);
1544 rc = mdt_object_lock(info, o1, lh1, MDS_INODELOCK_LAYOUT |
1545 MDS_INODELOCK_XATTR);
1549 rc = mdt_object_lock(info, o2, lh2, MDS_INODELOCK_LAYOUT |
1550 MDS_INODELOCK_XATTR);
1554 rc = mo_swap_layouts(info->mti_env, mdt_object_child(o1),
1555 mdt_object_child(o2), msl->msl_flags);
1559 mdt_swap_lov_flag(o1, o2);
1562 mdt_object_unlock(info, o2, lh2, rc);
1564 mdt_object_unlock(info, o1, lh1, rc);
1566 mdt_object_put(info->mti_env, o);
1568 mdt_thread_info_fini(info);
1572 static int mdt_raw_lookup(struct mdt_thread_info *info,
1573 struct mdt_object *parent,
1574 const struct lu_name *lname,
1575 struct ldlm_reply *ldlm_rep)
1577 struct lu_fid *child_fid = &info->mti_tmp_fid1;
1581 LASSERT(!info->mti_cross_ref);
1583 /* Only got the fid of this obj by name */
1584 fid_zero(child_fid);
1585 rc = mdo_lookup(info->mti_env, mdt_object_child(info->mti_object),
1586 lname, child_fid, &info->mti_spec);
1588 struct mdt_body *repbody;
1590 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1591 repbody->mbo_fid1 = *child_fid;
1592 repbody->mbo_valid = OBD_MD_FLID;
1593 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1594 } else if (rc == -ENOENT) {
1595 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1602 * UPDATE lock should be taken against parent, and be released before exit;
1603 * child_bits lock should be taken against child, and be returned back:
1604 * (1)normal request should release the child lock;
1605 * (2)intent request will grant the lock to client.
1607 static int mdt_getattr_name_lock(struct mdt_thread_info *info,
1608 struct mdt_lock_handle *lhc,
1610 struct ldlm_reply *ldlm_rep)
1612 struct ptlrpc_request *req = mdt_info_req(info);
1613 struct mdt_body *reqbody = NULL;
1614 struct mdt_object *parent = info->mti_object;
1615 struct mdt_object *child;
1616 struct lu_fid *child_fid = &info->mti_tmp_fid1;
1617 struct lu_name *lname = NULL;
1618 struct mdt_lock_handle *lhp = NULL;
1619 struct ldlm_lock *lock;
1627 is_resent = lustre_handle_is_used(&lhc->mlh_reg_lh);
1628 LASSERT(ergo(is_resent,
1629 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT));
1634 if (info->mti_cross_ref) {
1635 /* Only getattr on the child. Parent is on another node. */
1636 mdt_set_disposition(info, ldlm_rep,
1637 DISP_LOOKUP_EXECD | DISP_LOOKUP_POS);
1639 CDEBUG(D_INODE, "partial getattr_name child_fid = "DFID", "
1641 PFID(mdt_object_fid(child)), ldlm_rep);
1643 rc = mdt_check_resent_lock(info, child, lhc);
1646 } else if (rc > 0) {
1647 mdt_lock_handle_init(lhc);
1648 mdt_lock_reg_init(lhc, LCK_PR);
1651 * Object's name is on another MDS, no lookup or layout
1652 * lock is needed here but update lock is.
1654 child_bits &= ~(MDS_INODELOCK_LOOKUP |
1655 MDS_INODELOCK_LAYOUT);
1656 child_bits |= MDS_INODELOCK_PERM | MDS_INODELOCK_UPDATE;
1658 rc = mdt_object_lock(info, child, lhc, child_bits);
1663 /* Finally, we can get attr for child. */
1664 if (!mdt_object_exists(child)) {
1665 LU_OBJECT_DEBUG(D_INFO, info->mti_env,
1667 "remote object doesn't exist.");
1668 mdt_object_unlock(info, child, lhc, 1);
1672 rc = mdt_getattr_internal(info, child, 0);
1673 if (unlikely(rc != 0))
1674 mdt_object_unlock(info, child, lhc, 1);
1679 lname = &info->mti_name;
1680 mdt_name_unpack(info->mti_pill, &RMF_NAME, lname, MNF_FIX_ANON);
1682 if (lu_name_is_valid(lname)) {
1683 CDEBUG(D_INODE, "getattr with lock for "DFID"/"DNAME", "
1684 "ldlm_rep = %p\n", PFID(mdt_object_fid(parent)),
1685 PNAME(lname), ldlm_rep);
1687 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
1688 if (unlikely(reqbody == NULL))
1689 RETURN(err_serious(-EPROTO));
1691 *child_fid = reqbody->mbo_fid2;
1693 if (unlikely(!fid_is_sane(child_fid)))
1694 RETURN(err_serious(-EINVAL));
1696 CDEBUG(D_INODE, "getattr with lock for "DFID"/"DFID", "
1698 PFID(mdt_object_fid(parent)),
1699 PFID(&reqbody->mbo_fid2), ldlm_rep);
1702 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_EXECD);
1704 if (unlikely(!mdt_object_exists(parent)) && lu_name_is_valid(lname)) {
1705 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
1707 "Parent doesn't exist!");
1711 if (mdt_object_remote(parent)) {
1712 CERROR("%s: parent "DFID" is on remote target\n",
1713 mdt_obd_name(info->mti_mdt),
1714 PFID(mdt_object_fid(parent)));
1718 if (lu_name_is_valid(lname)) {
1719 /* Always allow to lookup ".." */
1720 if (unlikely(lname->ln_namelen == 2 &&
1721 lname->ln_name[0] == '.' &&
1722 lname->ln_name[1] == '.'))
1723 info->mti_spec.sp_permitted = 1;
1725 if (info->mti_body->mbo_valid == OBD_MD_FLID) {
1726 rc = mdt_raw_lookup(info, parent, lname, ldlm_rep);
1731 /* step 1: lock parent only if parent is a directory */
1732 if (S_ISDIR(lu_object_attr(&parent->mot_obj))) {
1733 lhp = &info->mti_lh[MDT_LH_PARENT];
1734 mdt_lock_pdo_init(lhp, LCK_PR, lname);
1735 rc = mdt_object_lock(info, parent, lhp,
1736 MDS_INODELOCK_UPDATE);
1737 if (unlikely(rc != 0))
1741 /* step 2: lookup child's fid by name */
1742 fid_zero(child_fid);
1743 rc = mdo_lookup(info->mti_env, mdt_object_child(parent), lname,
1744 child_fid, &info->mti_spec);
1746 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1749 GOTO(out_parent, rc);
1752 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1755 *step 3: find the child object by fid & lock it.
1756 * regardless if it is local or remote.
1758 *Note: LU-3240 (commit 762f2114d282a98ebfa4dbbeea9298a8088ad24e)
1759 * set parent dir fid the same as child fid in getattr by fid case
1760 * we should not lu_object_find() the object again, could lead
1761 * to hung if there is a concurrent unlink destroyed the object.
1763 if (lu_fid_eq(mdt_object_fid(parent), child_fid)) {
1764 mdt_object_get(info->mti_env, parent);
1767 child = mdt_object_find(info->mti_env, info->mti_mdt,
1771 if (unlikely(IS_ERR(child)))
1772 GOTO(out_parent, rc = PTR_ERR(child));
1774 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RESEND, obd_timeout * 2);
1775 if (!mdt_object_exists(child)) {
1776 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
1778 "Object doesn't exist!");
1779 GOTO(out_child, rc = -ENOENT);
1782 rc = mdt_check_resent_lock(info, child, lhc);
1784 GOTO(out_child, rc);
1785 } else if (rc > 0) {
1786 mdt_lock_handle_init(lhc);
1787 mdt_lock_reg_init(lhc, LCK_PR);
1789 if (!(child_bits & MDS_INODELOCK_UPDATE) &&
1790 mdt_object_exists(child) && !mdt_object_remote(child)) {
1791 struct md_attr *ma = &info->mti_attr;
1794 ma->ma_need = MA_INODE;
1795 rc = mdt_attr_get_complex(info, child, ma);
1796 if (unlikely(rc != 0))
1797 GOTO(out_child, rc);
1799 /* If the file has not been changed for some time, we
1800 * return not only a LOOKUP lock, but also an UPDATE
1801 * lock and this might save us RPC on later STAT. For
1802 * directories, it also let negative dentry cache start
1803 * working for this dir. */
1804 if (ma->ma_valid & MA_INODE &&
1805 ma->ma_attr.la_valid & LA_CTIME &&
1806 info->mti_mdt->mdt_namespace->ns_ctime_age_limit +
1807 ma->ma_attr.la_ctime < ktime_get_real_seconds())
1808 child_bits |= MDS_INODELOCK_UPDATE;
1811 /* layout lock must be granted in a best-effort way
1812 * for IT operations */
1813 LASSERT(!(child_bits & MDS_INODELOCK_LAYOUT));
1814 if (S_ISREG(lu_object_attr(&child->mot_obj)) &&
1815 !mdt_object_remote(child) && ldlm_rep != NULL) {
1816 if (!OBD_FAIL_CHECK(OBD_FAIL_MDS_NO_LL_GETATTR) &&
1817 exp_connect_layout(info->mti_exp)) {
1818 /* try to grant layout lock for regular file. */
1819 try_bits = MDS_INODELOCK_LAYOUT;
1821 /* Acquire DOM lock in advance for data-on-mdt file */
1822 if (child != parent)
1823 try_bits |= MDS_INODELOCK_DOM;
1826 if (try_bits != 0) {
1827 /* try layout lock, it may fail to be granted due to
1828 * contention at LOOKUP or UPDATE */
1829 rc = mdt_object_lock_try(info, child, lhc, &child_bits,
1831 if (child_bits & MDS_INODELOCK_LAYOUT)
1834 /* Do not enqueue the UPDATE lock from MDT(cross-MDT),
1835 * client will enqueue the lock to the remote MDT */
1836 if (mdt_object_remote(child))
1837 child_bits &= ~MDS_INODELOCK_UPDATE;
1838 rc = mdt_object_lock(info, child, lhc, child_bits);
1840 if (unlikely(rc != 0))
1841 GOTO(out_child, rc);
1844 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
1846 /* finally, we can get attr for child. */
1847 rc = mdt_getattr_internal(info, child, ma_need);
1848 if (unlikely(rc != 0)) {
1849 mdt_object_unlock(info, child, lhc, 1);
1851 /* Debugging code. */
1852 LDLM_DEBUG(lock, "Returning lock to client");
1853 LASSERTF(fid_res_name_eq(mdt_object_fid(child),
1854 &lock->l_resource->lr_name),
1855 "Lock res_id: "DLDLMRES", fid: "DFID"\n",
1856 PLDLMRES(lock->l_resource),
1857 PFID(mdt_object_fid(child)));
1859 if (S_ISREG(lu_object_attr(&child->mot_obj)) &&
1860 mdt_object_exists(child) && !mdt_object_remote(child) &&
1862 LDLM_LOCK_PUT(lock);
1863 mdt_object_put(info->mti_env, child);
1864 /* NB: call the mdt_pack_size2body always after
1865 * mdt_object_put(), that is why this special
1866 * exit path is used. */
1867 rc = mdt_pack_size2body(info, child_fid,
1869 if (rc != 0 && child_bits & MDS_INODELOCK_DOM) {
1870 /* DOM lock was taken in advance but this is
1871 * not DoM file. Drop the lock. */
1872 lock_res_and_lock(lock);
1873 ldlm_inodebits_drop(lock, MDS_INODELOCK_DOM);
1874 unlock_res_and_lock(lock);
1877 GOTO(out_parent, rc = 0);
1881 LDLM_LOCK_PUT(lock);
1885 mdt_object_put(info->mti_env, child);
1888 mdt_object_unlock(info, parent, lhp, 1);
1892 /* normal handler: should release the child lock */
1893 static int mdt_getattr_name(struct tgt_session_info *tsi)
1895 struct mdt_thread_info *info = tsi2mdt_info(tsi);
1896 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_CHILD];
1897 struct mdt_body *reqbody;
1898 struct mdt_body *repbody;
1902 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
1903 LASSERT(reqbody != NULL);
1904 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1905 LASSERT(repbody != NULL);
1907 info->mti_cross_ref = !!(reqbody->mbo_valid & OBD_MD_FLCROSSREF);
1908 repbody->mbo_eadatasize = 0;
1909 repbody->mbo_aclsize = 0;
1911 rc = mdt_init_ucred_intent_getattr(info, reqbody);
1913 GOTO(out_shrink, rc);
1915 rc = mdt_getattr_name_lock(info, lhc, MDS_INODELOCK_UPDATE, NULL);
1916 if (lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1917 ldlm_lock_decref(&lhc->mlh_reg_lh, lhc->mlh_reg_mode);
1918 lhc->mlh_reg_lh.cookie = 0;
1920 mdt_exit_ucred(info);
1923 mdt_client_compatibility(info);
1924 rc2 = mdt_fix_reply(info);
1927 mdt_thread_info_fini(info);
1931 static int mdt_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
1932 void *karg, void __user *uarg);
1934 static int mdt_set_info(struct tgt_session_info *tsi)
1936 struct ptlrpc_request *req = tgt_ses_req(tsi);
1939 int keylen, vallen, rc = 0;
1943 key = req_capsule_client_get(tsi->tsi_pill, &RMF_SETINFO_KEY);
1945 DEBUG_REQ(D_HA, req, "no set_info key");
1946 RETURN(err_serious(-EFAULT));
1949 keylen = req_capsule_get_size(tsi->tsi_pill, &RMF_SETINFO_KEY,
1952 val = req_capsule_client_get(tsi->tsi_pill, &RMF_SETINFO_VAL);
1954 DEBUG_REQ(D_HA, req, "no set_info val");
1955 RETURN(err_serious(-EFAULT));
1958 vallen = req_capsule_get_size(tsi->tsi_pill, &RMF_SETINFO_VAL,
1961 /* Swab any part of val you need to here */
1962 if (KEY_IS(KEY_READ_ONLY)) {
1963 spin_lock(&req->rq_export->exp_lock);
1965 *exp_connect_flags_ptr(req->rq_export) |=
1968 *exp_connect_flags_ptr(req->rq_export) &=
1969 ~OBD_CONNECT_RDONLY;
1970 spin_unlock(&req->rq_export->exp_lock);
1971 } else if (KEY_IS(KEY_CHANGELOG_CLEAR)) {
1972 struct changelog_setinfo *cs = val;
1974 if (vallen != sizeof(*cs)) {
1975 CERROR("%s: bad changelog_clear setinfo size %d\n",
1976 tgt_name(tsi->tsi_tgt), vallen);
1979 if (ptlrpc_req_need_swab(req)) {
1980 __swab64s(&cs->cs_recno);
1981 __swab32s(&cs->cs_id);
1984 rc = mdt_iocontrol(OBD_IOC_CHANGELOG_CLEAR, req->rq_export,
1986 } else if (KEY_IS(KEY_EVICT_BY_NID)) {
1988 obd_export_evict_by_nid(req->rq_export->exp_obd, val);
1995 static int mdt_readpage(struct tgt_session_info *tsi)
1997 struct mdt_thread_info *info = mdt_th_info(tsi->tsi_env);
1998 struct mdt_object *object = mdt_obj(tsi->tsi_corpus);
1999 struct lu_rdpg *rdpg = &info->mti_u.rdpg.mti_rdpg;
2000 const struct mdt_body *reqbody = tsi->tsi_mdt_body;
2001 struct mdt_body *repbody;
2007 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK))
2008 RETURN(err_serious(-ENOMEM));
2010 repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_MDT_BODY);
2011 if (repbody == NULL || reqbody == NULL)
2012 RETURN(err_serious(-EFAULT));
2015 * prepare @rdpg before calling lower layers and transfer itself. Here
2016 * reqbody->size contains offset of where to start to read and
2017 * reqbody->nlink contains number bytes to read.
2019 rdpg->rp_hash = reqbody->mbo_size;
2020 if (rdpg->rp_hash != reqbody->mbo_size) {
2021 CERROR("Invalid hash: %#llx != %#llx\n",
2022 rdpg->rp_hash, reqbody->mbo_size);
2026 rdpg->rp_attrs = reqbody->mbo_mode;
2027 if (exp_connect_flags(tsi->tsi_exp) & OBD_CONNECT_64BITHASH)
2028 rdpg->rp_attrs |= LUDA_64BITHASH;
2029 rdpg->rp_count = min_t(unsigned int, reqbody->mbo_nlink,
2030 exp_max_brw_size(tsi->tsi_exp));
2031 rdpg->rp_npages = (rdpg->rp_count + PAGE_SIZE - 1) >>
2033 OBD_ALLOC(rdpg->rp_pages, rdpg->rp_npages * sizeof rdpg->rp_pages[0]);
2034 if (rdpg->rp_pages == NULL)
2037 for (i = 0; i < rdpg->rp_npages; ++i) {
2038 rdpg->rp_pages[i] = alloc_page(GFP_NOFS);
2039 if (rdpg->rp_pages[i] == NULL)
2040 GOTO(free_rdpg, rc = -ENOMEM);
2043 /* call lower layers to fill allocated pages with directory data */
2044 rc = mo_readpage(tsi->tsi_env, mdt_object_child(object), rdpg);
2046 GOTO(free_rdpg, rc);
2048 /* send pages to client */
2049 rc = tgt_sendpage(tsi, rdpg, rc);
2054 for (i = 0; i < rdpg->rp_npages; i++)
2055 if (rdpg->rp_pages[i] != NULL)
2056 __free_page(rdpg->rp_pages[i]);
2057 OBD_FREE(rdpg->rp_pages, rdpg->rp_npages * sizeof rdpg->rp_pages[0]);
2059 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE))
2065 static int mdt_fix_attr_ucred(struct mdt_thread_info *info, __u32 op)
2067 struct lu_ucred *uc = mdt_ucred_check(info);
2068 struct lu_attr *attr = &info->mti_attr.ma_attr;
2073 if (op != REINT_SETATTR) {
2074 if ((attr->la_valid & LA_UID) && (attr->la_uid != -1))
2075 attr->la_uid = uc->uc_fsuid;
2076 /* for S_ISGID, inherit gid from his parent, such work will be
2077 * done in cmm/mdd layer, here set all cases as uc->uc_fsgid. */
2078 if ((attr->la_valid & LA_GID) && (attr->la_gid != -1))
2079 attr->la_gid = uc->uc_fsgid;
2085 static int mdt_reint_internal(struct mdt_thread_info *info,
2086 struct mdt_lock_handle *lhc,
2089 struct req_capsule *pill = info->mti_pill;
2090 struct mdt_body *repbody;
2095 rc = mdt_reint_unpack(info, op);
2097 CERROR("Can't unpack reint, rc %d\n", rc);
2098 RETURN(err_serious(rc));
2101 /* for replay (no_create) lmm is not needed, client has it already */
2102 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
2103 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER,
2106 /* llog cookies are always 0, the field is kept for compatibility */
2107 if (req_capsule_has_field(pill, &RMF_LOGCOOKIES, RCL_SERVER))
2108 req_capsule_set_size(pill, &RMF_LOGCOOKIES, RCL_SERVER, 0);
2110 /* Set ACL reply buffer size as LUSTRE_POSIX_ACL_MAX_SIZE_OLD
2111 * by default. If the target object has more ACL entries, then
2112 * enlarge the buffer when necessary. */
2113 if (req_capsule_has_field(pill, &RMF_ACL, RCL_SERVER))
2114 req_capsule_set_size(pill, &RMF_ACL, RCL_SERVER,
2115 LUSTRE_POSIX_ACL_MAX_SIZE_OLD);
2117 rc = req_capsule_server_pack(pill);
2119 CERROR("Can't pack response, rc %d\n", rc);
2120 RETURN(err_serious(rc));
2123 if (req_capsule_has_field(pill, &RMF_MDT_BODY, RCL_SERVER)) {
2124 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
2126 repbody->mbo_eadatasize = 0;
2127 repbody->mbo_aclsize = 0;
2130 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_REINT_DELAY, 10);
2132 /* for replay no cookkie / lmm need, because client have this already */
2133 if (info->mti_spec.no_create)
2134 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
2135 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER, 0);
2137 rc = mdt_init_ucred_reint(info);
2139 GOTO(out_shrink, rc);
2141 rc = mdt_fix_attr_ucred(info, op);
2143 GOTO(out_ucred, rc = err_serious(rc));
2145 rc = mdt_check_resent(info, mdt_reconstruct, lhc);
2147 GOTO(out_ucred, rc);
2148 } else if (rc == 1) {
2149 DEBUG_REQ(D_INODE, mdt_info_req(info), "resent opt.");
2150 rc = lustre_msg_get_status(mdt_info_req(info)->rq_repmsg);
2151 GOTO(out_ucred, rc);
2153 rc = mdt_reint_rec(info, lhc);
2156 mdt_exit_ucred(info);
2158 mdt_client_compatibility(info);
2160 rc2 = mdt_fix_reply(info);
2165 * Data-on-MDT optimization - read data along with OPEN and return it
2166 * in reply. Do that only if we have both DOM and LAYOUT locks.
2168 if (rc == 0 && op == REINT_OPEN &&
2169 info->mti_attr.ma_lmm != NULL &&
2170 mdt_lmm_dom_entry(info->mti_attr.ma_lmm) == LMM_DOM_ONLY) {
2171 rc = mdt_dom_read_on_open(info, info->mti_mdt,
2178 static long mdt_reint_opcode(struct ptlrpc_request *req,
2179 const struct req_format **fmt)
2181 struct mdt_device *mdt;
2182 struct mdt_rec_reint *rec;
2185 rec = req_capsule_client_get(&req->rq_pill, &RMF_REC_REINT);
2187 opc = rec->rr_opcode;
2188 DEBUG_REQ(D_INODE, req, "reint opt = %ld", opc);
2189 if (opc < REINT_MAX && fmt[opc] != NULL)
2190 req_capsule_extend(&req->rq_pill, fmt[opc]);
2192 mdt = mdt_exp2dev(req->rq_export);
2193 CERROR("%s: Unsupported opcode '%ld' from client '%s':"
2194 " rc = %d\n", req->rq_export->exp_obd->obd_name,
2195 opc, mdt->mdt_ldlm_client->cli_name, -EFAULT);
2196 opc = err_serious(-EFAULT);
2199 opc = err_serious(-EFAULT);
2204 static int mdt_reint(struct tgt_session_info *tsi)
2208 static const struct req_format *reint_fmts[REINT_MAX] = {
2209 [REINT_SETATTR] = &RQF_MDS_REINT_SETATTR,
2210 [REINT_CREATE] = &RQF_MDS_REINT_CREATE,
2211 [REINT_LINK] = &RQF_MDS_REINT_LINK,
2212 [REINT_UNLINK] = &RQF_MDS_REINT_UNLINK,
2213 [REINT_RENAME] = &RQF_MDS_REINT_RENAME,
2214 [REINT_OPEN] = &RQF_MDS_REINT_OPEN,
2215 [REINT_SETXATTR] = &RQF_MDS_REINT_SETXATTR,
2216 [REINT_RMENTRY] = &RQF_MDS_REINT_UNLINK,
2217 [REINT_MIGRATE] = &RQF_MDS_REINT_MIGRATE,
2218 [REINT_RESYNC] = &RQF_MDS_REINT_RESYNC,
2223 opc = mdt_reint_opcode(tgt_ses_req(tsi), reint_fmts);
2225 struct mdt_thread_info *info = tsi2mdt_info(tsi);
2227 * No lock possible here from client to pass it to reint code
2230 rc = mdt_reint_internal(info, NULL, opc);
2231 mdt_thread_info_fini(info);
2236 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_REINT_NET_REP;
2240 /* this should sync the whole device */
2241 int mdt_device_sync(const struct lu_env *env, struct mdt_device *mdt)
2243 struct dt_device *dt = mdt->mdt_bottom;
2247 rc = dt->dd_ops->dt_sync(env, dt);
2251 /* this should sync this object */
2252 static int mdt_object_sync(const struct lu_env *env, struct obd_export *exp,
2253 struct mdt_object *mo)
2259 if (!mdt_object_exists(mo)) {
2260 CWARN("%s: non existing object "DFID": rc = %d\n",
2261 exp->exp_obd->obd_name, PFID(mdt_object_fid(mo)),
2266 rc = mo_object_sync(env, mdt_object_child(mo));
2271 static int mdt_sync(struct tgt_session_info *tsi)
2273 struct ptlrpc_request *req = tgt_ses_req(tsi);
2274 struct req_capsule *pill = tsi->tsi_pill;
2275 struct mdt_body *body;
2280 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SYNC_PACK))
2281 RETURN(err_serious(-ENOMEM));
2283 if (fid_seq(&tsi->tsi_mdt_body->mbo_fid1) == 0) {
2284 rc = mdt_device_sync(tsi->tsi_env, mdt_exp2dev(tsi->tsi_exp));
2286 struct mdt_thread_info *info = tsi2mdt_info(tsi);
2288 /* sync an object */
2289 rc = mdt_object_sync(tsi->tsi_env, tsi->tsi_exp,
2292 const struct lu_fid *fid;
2293 struct lu_attr *la = &info->mti_attr.ma_attr;
2295 info->mti_attr.ma_need = MA_INODE;
2296 info->mti_attr.ma_valid = 0;
2297 rc = mdt_attr_get_complex(info, info->mti_object,
2300 body = req_capsule_server_get(pill,
2302 fid = mdt_object_fid(info->mti_object);
2303 mdt_pack_attr2body(info, body, la, fid);
2306 mdt_thread_info_fini(info);
2309 mdt_counter_incr(req, LPROC_MDT_SYNC);
2314 static int mdt_data_sync(struct tgt_session_info *tsi)
2316 struct mdt_thread_info *info;
2317 struct mdt_device *mdt = mdt_exp2dev(tsi->tsi_exp);
2318 struct ost_body *body = tsi->tsi_ost_body;
2319 struct ost_body *repbody;
2320 struct mdt_object *mo = NULL;
2326 repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_OST_BODY);
2328 /* if no fid is specified then do nothing,
2329 * device sync is done via MDS_SYNC */
2330 if (fid_is_zero(&tsi->tsi_fid))
2333 mo = mdt_object_find(tsi->tsi_env, mdt, &tsi->tsi_fid);
2335 RETURN(PTR_ERR(mo));
2337 rc = mdt_object_sync(tsi->tsi_env, tsi->tsi_exp, mo);
2341 repbody->oa.o_oi = body->oa.o_oi;
2342 repbody->oa.o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
2344 info = tsi2mdt_info(tsi);
2345 ma = &info->mti_attr;
2346 ma->ma_need = MA_INODE;
2348 rc = mdt_attr_get_complex(info, mo, ma);
2350 obdo_from_la(&repbody->oa, &ma->ma_attr, VALID_FLAGS);
2353 mdt_thread_info_fini(info);
2358 mdt_object_put(tsi->tsi_env, mo);
2363 * Handle quota control requests to consult current usage/limit, but also
2364 * to configure quota enforcement
2366 static int mdt_quotactl(struct tgt_session_info *tsi)
2368 struct obd_export *exp = tsi->tsi_exp;
2369 struct req_capsule *pill = tsi->tsi_pill;
2370 struct obd_quotactl *oqctl, *repoqc;
2372 struct mdt_device *mdt = mdt_exp2dev(exp);
2373 struct lu_device *qmt = mdt->mdt_qmt_dev;
2374 struct lu_nodemap *nodemap;
2377 oqctl = req_capsule_client_get(pill, &RMF_OBD_QUOTACTL);
2379 RETURN(err_serious(-EPROTO));
2381 rc = req_capsule_server_pack(pill);
2383 RETURN(err_serious(rc));
2385 nodemap = nodemap_get_from_exp(exp);
2386 if (IS_ERR(nodemap))
2387 RETURN(PTR_ERR(nodemap));
2389 switch (oqctl->qc_cmd) {
2390 /* master quotactl */
2393 case LUSTRE_Q_SETDEFAULT:
2394 if (!nodemap_can_setquota(nodemap))
2395 GOTO(out_nodemap, rc = -EPERM);
2398 case LUSTRE_Q_GETDEFAULT:
2400 GOTO(out_nodemap, rc = -EOPNOTSUPP);
2401 /* slave quotactl */
2406 CERROR("Unsupported quotactl command: %d\n", oqctl->qc_cmd);
2407 GOTO(out_nodemap, rc = -EFAULT);
2411 switch (oqctl->qc_type) {
2413 id = nodemap_map_id(nodemap, NODEMAP_UID,
2414 NODEMAP_CLIENT_TO_FS, id);
2417 id = nodemap_map_id(nodemap, NODEMAP_GID,
2418 NODEMAP_CLIENT_TO_FS, id);
2421 /* todo: check/map project id */
2425 GOTO(out_nodemap, rc = -EOPNOTSUPP);
2427 repoqc = req_capsule_server_get(pill, &RMF_OBD_QUOTACTL);
2429 GOTO(out_nodemap, rc = err_serious(-EFAULT));
2431 if (oqctl->qc_cmd == Q_SETINFO || oqctl->qc_cmd == Q_SETQUOTA)
2432 barrier_exit(tsi->tsi_tgt->lut_bottom);
2434 if (oqctl->qc_id != id)
2435 swap(oqctl->qc_id, id);
2437 if (oqctl->qc_cmd == Q_SETINFO || oqctl->qc_cmd == Q_SETQUOTA) {
2438 if (unlikely(!barrier_entry(tsi->tsi_tgt->lut_bottom)))
2439 RETURN(-EINPROGRESS);
2442 switch (oqctl->qc_cmd) {
2448 case LUSTRE_Q_SETDEFAULT:
2449 case LUSTRE_Q_GETDEFAULT:
2450 /* forward quotactl request to QMT */
2451 rc = qmt_hdls.qmth_quotactl(tsi->tsi_env, qmt, oqctl);
2456 /* slave quotactl */
2457 rc = lquotactl_slv(tsi->tsi_env, tsi->tsi_tgt->lut_bottom,
2462 CERROR("Unsupported quotactl command: %d\n", oqctl->qc_cmd);
2463 GOTO(out_nodemap, rc = -EFAULT);
2466 if (oqctl->qc_id != id)
2467 swap(oqctl->qc_id, id);
2474 nodemap_putref(nodemap);
2479 /** clone llog ctxt from child (mdd)
2480 * This allows remote llog (replicator) access.
2481 * We can either pass all llog RPCs (eg mdt_llog_create) on to child where the
2482 * context was originally set up, or we can handle them directly.
2483 * I choose the latter, but that means I need any llog
2484 * contexts set up by child to be accessable by the mdt. So we clone the
2485 * context into our context list here.
2487 static int mdt_llog_ctxt_clone(const struct lu_env *env, struct mdt_device *mdt,
2490 struct md_device *next = mdt->mdt_child;
2491 struct llog_ctxt *ctxt;
2494 if (!llog_ctxt_null(mdt2obd_dev(mdt), idx))
2497 rc = next->md_ops->mdo_llog_ctxt_get(env, next, idx, (void **)&ctxt);
2498 if (rc || ctxt == NULL) {
2502 rc = llog_group_set_ctxt(&mdt2obd_dev(mdt)->obd_olg, ctxt, idx);
2504 CERROR("Can't set mdt ctxt %d\n", rc);
2509 static int mdt_llog_ctxt_unclone(const struct lu_env *env,
2510 struct mdt_device *mdt, int idx)
2512 struct llog_ctxt *ctxt;
2514 ctxt = llog_get_context(mdt2obd_dev(mdt), idx);
2517 /* Put once for the get we just did, and once for the clone */
2518 llog_ctxt_put(ctxt);
2519 llog_ctxt_put(ctxt);
2524 * sec context handlers
2526 static int mdt_sec_ctx_handle(struct tgt_session_info *tsi)
2528 CFS_FAIL_TIMEOUT(OBD_FAIL_SEC_CTX_HDL_PAUSE, cfs_fail_val);
2534 * quota request handlers
2536 static int mdt_quota_dqacq(struct tgt_session_info *tsi)
2538 struct mdt_device *mdt = mdt_exp2dev(tsi->tsi_exp);
2539 struct lu_device *qmt = mdt->mdt_qmt_dev;
2544 RETURN(err_serious(-EOPNOTSUPP));
2546 rc = qmt_hdls.qmth_dqacq(tsi->tsi_env, qmt, tgt_ses_req(tsi));
2550 struct mdt_object *mdt_object_new(const struct lu_env *env,
2551 struct mdt_device *d,
2552 const struct lu_fid *f)
2554 struct lu_object_conf conf = { .loc_flags = LOC_F_NEW };
2555 struct lu_object *o;
2556 struct mdt_object *m;
2559 CDEBUG(D_INFO, "Allocate object for "DFID"\n", PFID(f));
2560 o = lu_object_find(env, &d->mdt_lu_dev, f, &conf);
2561 if (unlikely(IS_ERR(o)))
2562 m = (struct mdt_object *)o;
2568 struct mdt_object *mdt_object_find(const struct lu_env *env,
2569 struct mdt_device *d,
2570 const struct lu_fid *f)
2572 struct lu_object *o;
2573 struct mdt_object *m;
2576 CDEBUG(D_INFO, "Find object for "DFID"\n", PFID(f));
2577 o = lu_object_find(env, &d->mdt_lu_dev, f, NULL);
2578 if (unlikely(IS_ERR(o)))
2579 m = (struct mdt_object *)o;
2587 * Asyncronous commit for mdt device.
2589 * Pass asynchonous commit call down the MDS stack.
2591 * \param env environment
2592 * \param mdt the mdt device
2594 static void mdt_device_commit_async(const struct lu_env *env,
2595 struct mdt_device *mdt)
2597 struct dt_device *dt = mdt->mdt_bottom;
2601 rc = dt->dd_ops->dt_commit_async(env, dt);
2602 if (unlikely(rc != 0))
2603 CWARN("%s: async commit start failed: rc = %d\n",
2604 mdt_obd_name(mdt), rc);
2605 atomic_inc(&mdt->mdt_async_commit_count);
2610 * Mark the lock as "synchonous".
2612 * Mark the lock to deffer transaction commit to the unlock time.
2614 * \param lock the lock to mark as "synchonous"
2616 * \see mdt_is_lock_sync
2617 * \see mdt_save_lock
2619 static inline void mdt_set_lock_sync(struct ldlm_lock *lock)
2621 lock->l_ast_data = (void*)1;
2625 * Check whehter the lock "synchonous" or not.
2627 * \param lock the lock to check
2628 * \retval 1 the lock is "synchonous"
2629 * \retval 0 the lock isn't "synchronous"
2631 * \see mdt_set_lock_sync
2632 * \see mdt_save_lock
2634 static inline int mdt_is_lock_sync(struct ldlm_lock *lock)
2636 return lock->l_ast_data != NULL;
2640 * Blocking AST for mdt locks.
2642 * Starts transaction commit if in case of COS lock conflict or
2643 * deffers such a commit to the mdt_save_lock.
2645 * \param lock the lock which blocks a request or cancelling lock
2646 * \param desc unused
2647 * \param data unused
2648 * \param flag indicates whether this cancelling or blocking callback
2650 * \see ldlm_blocking_ast_nocheck
2652 int mdt_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
2653 void *data, int flag)
2655 struct obd_device *obd = ldlm_lock_to_ns(lock)->ns_obd;
2656 struct mdt_device *mdt = mdt_dev(obd->obd_lu_dev);
2657 bool commit_async = false;
2661 if (flag == LDLM_CB_CANCELING)
2664 lock_res_and_lock(lock);
2665 if (lock->l_blocking_ast != mdt_blocking_ast) {
2666 unlock_res_and_lock(lock);
2669 /* There is no lock conflict if l_blocking_lock == NULL,
2670 * it indicates a blocking ast sent from ldlm_lock_decref_internal
2671 * when the last reference to a local lock was released */
2672 if (lock->l_req_mode & (LCK_PW | LCK_EX) &&
2673 lock->l_blocking_lock != NULL) {
2674 if (mdt_cos_is_enabled(mdt)) {
2675 if (lock->l_client_cookie !=
2676 lock->l_blocking_lock->l_client_cookie)
2677 mdt_set_lock_sync(lock);
2678 } else if (mdt_slc_is_enabled(mdt) &&
2679 ldlm_is_cos_incompat(lock->l_blocking_lock)) {
2680 mdt_set_lock_sync(lock);
2682 * we may do extra commit here, but there is a small
2683 * window to miss a commit: lock was unlocked (saved),
2684 * then a conflict lock queued and we come here, but
2685 * REP-ACK not received, so lock was not converted to
2687 * Fortunately this window is quite small, so the
2688 * extra commit should be rare (not to say distributed
2689 * operation is rare too).
2691 commit_async = true;
2693 } else if (lock->l_req_mode == LCK_COS &&
2694 lock->l_blocking_lock != NULL) {
2695 commit_async = true;
2698 rc = ldlm_blocking_ast_nocheck(lock);
2703 rc = lu_env_init(&env, LCT_LOCAL);
2704 if (unlikely(rc != 0))
2705 CWARN("%s: lu_env initialization failed, cannot "
2706 "start asynchronous commit: rc = %d\n",
2709 mdt_device_commit_async(&env, mdt);
2716 * Blocking AST for cross-MDT lock
2718 * Discard lock from uncommitted_slc_locks and cancel it.
2720 * \param lock the lock which blocks a request or cancelling lock
2721 * \param desc unused
2722 * \param data unused
2723 * \param flag indicates whether this cancelling or blocking callback
2724 * \retval 0 on success
2725 * \retval negative number on error
2727 int mdt_remote_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
2728 void *data, int flag)
2734 case LDLM_CB_BLOCKING: {
2735 struct lustre_handle lockh;
2737 ldlm_lock2handle(lock, &lockh);
2738 rc = ldlm_cli_cancel(&lockh,
2739 ldlm_is_atomic_cb(lock) ? 0 : LCF_ASYNC);
2741 CDEBUG(D_INODE, "ldlm_cli_cancel: %d\n", rc);
2746 case LDLM_CB_CANCELING: {
2747 struct obd_device *obd = ldlm_lock_to_ns(lock)->ns_obd;
2748 struct mdt_device *mdt =
2749 mdt_dev(obd->obd_lu_dev->ld_site->ls_top_dev);
2751 LDLM_DEBUG(lock, "Revoke remote lock\n");
2753 /* discard slc lock here so that it can be cleaned anytime,
2754 * especially for cleanup_resource() */
2755 tgt_discard_slc_lock(&mdt->mdt_lut, lock);
2757 /* once we cache lock, l_ast_data is set to mdt_object */
2758 if (lock->l_ast_data != NULL) {
2759 struct mdt_object *mo = lock->l_ast_data;
2762 rc = lu_env_init(&env, LCT_MD_THREAD);
2763 if (unlikely(rc != 0)) {
2764 CWARN("%s: lu_env initialization failed, object"
2765 "%p "DFID" is leaked!\n",
2767 PFID(mdt_object_fid(mo)));
2771 if (lock->l_policy_data.l_inodebits.bits &
2772 (MDS_INODELOCK_XATTR | MDS_INODELOCK_UPDATE)) {
2773 rc = mo_invalidate(&env, mdt_object_child(mo));
2774 mo->mot_cache_attr = 0;
2776 mdt_object_put(&env, mo);
2788 int mdt_check_resent_lock(struct mdt_thread_info *info,
2789 struct mdt_object *mo,
2790 struct mdt_lock_handle *lhc)
2792 /* the lock might already be gotten in ldlm_handle_enqueue() */
2793 if (unlikely(lustre_handle_is_used(&lhc->mlh_reg_lh))) {
2794 struct ptlrpc_request *req = mdt_info_req(info);
2795 struct ldlm_lock *lock;
2797 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
2798 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
2800 /* Lock is pinned by ldlm_handle_enqueue0() as it is
2801 * a resend case, however, it could be already destroyed
2802 * due to client eviction or a raced cancel RPC. */
2803 LDLM_DEBUG_NOLOCK("Invalid lock handle %#llx",
2804 lhc->mlh_reg_lh.cookie);
2808 if (!fid_res_name_eq(mdt_object_fid(mo),
2809 &lock->l_resource->lr_name)) {
2810 CWARN("%s: Although resent, but still not "
2811 "get child lock:"DFID"\n",
2812 info->mti_exp->exp_obd->obd_name,
2813 PFID(mdt_object_fid(mo)));
2814 LDLM_LOCK_PUT(lock);
2817 LDLM_LOCK_PUT(lock);
2823 int mdt_remote_object_lock_try(struct mdt_thread_info *mti,
2824 struct mdt_object *o, const struct lu_fid *fid,
2825 struct lustre_handle *lh, enum ldlm_mode mode,
2826 __u64 *ibits, __u64 trybits, bool cache)
2828 struct ldlm_enqueue_info *einfo = &mti->mti_remote_einfo;
2829 union ldlm_policy_data *policy = &mti->mti_policy;
2830 struct ldlm_res_id *res_id = &mti->mti_res_id;
2834 LASSERT(mdt_object_remote(o));
2836 fid_build_reg_res_name(fid, res_id);
2838 memset(einfo, 0, sizeof(*einfo));
2839 einfo->ei_type = LDLM_IBITS;
2840 einfo->ei_mode = mode;
2841 einfo->ei_cb_bl = mdt_remote_blocking_ast;
2842 einfo->ei_cb_cp = ldlm_completion_ast;
2843 einfo->ei_enq_slave = 0;
2844 einfo->ei_res_id = res_id;
2848 * if we cache lock, couple lock with mdt_object, so that object
2849 * can be easily found in lock ASTs.
2851 mdt_object_get(mti->mti_env, o);
2852 einfo->ei_cbdata = o;
2855 memset(policy, 0, sizeof(*policy));
2856 policy->l_inodebits.bits = *ibits;
2857 policy->l_inodebits.try_bits = trybits;
2859 rc = mo_object_lock(mti->mti_env, mdt_object_child(o), lh, einfo,
2861 if (rc < 0 && cache)
2862 mdt_object_put(mti->mti_env, o);
2864 /* Return successfully acquired bits to a caller */
2866 struct ldlm_lock *lock = ldlm_handle2lock(lh);
2869 *ibits = lock->l_policy_data.l_inodebits.bits;
2870 LDLM_LOCK_PUT(lock);
2875 int mdt_remote_object_lock(struct mdt_thread_info *mti, struct mdt_object *o,
2876 const struct lu_fid *fid, struct lustre_handle *lh,
2877 enum ldlm_mode mode, __u64 ibits, bool cache)
2879 return mdt_remote_object_lock_try(mti, o, fid, lh, mode, &ibits, 0,
2883 static int mdt_object_local_lock(struct mdt_thread_info *info,
2884 struct mdt_object *o,
2885 struct mdt_lock_handle *lh, __u64 *ibits,
2886 __u64 trybits, bool cos_incompat)
2888 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
2889 union ldlm_policy_data *policy = &info->mti_policy;
2890 struct ldlm_res_id *res_id = &info->mti_res_id;
2895 LASSERT(!lustre_handle_is_used(&lh->mlh_reg_lh));
2896 LASSERT(!lustre_handle_is_used(&lh->mlh_pdo_lh));
2897 LASSERT(lh->mlh_reg_mode != LCK_MINMODE);
2898 LASSERT(lh->mlh_type != MDT_NUL_LOCK);
2901 LASSERT(lh->mlh_reg_mode == LCK_PW ||
2902 lh->mlh_reg_mode == LCK_EX);
2903 dlmflags |= LDLM_FL_COS_INCOMPAT;
2904 } else if (mdt_cos_is_enabled(info->mti_mdt)) {
2905 dlmflags |= LDLM_FL_COS_ENABLED;
2908 /* Only enqueue LOOKUP lock for remote object */
2909 LASSERT(ergo(mdt_object_remote(o), *ibits == MDS_INODELOCK_LOOKUP));
2911 if (lh->mlh_type == MDT_PDO_LOCK) {
2912 /* check for exists after object is locked */
2913 if (mdt_object_exists(o) == 0) {
2914 /* Non-existent object shouldn't have PDO lock */
2917 /* Non-dir object shouldn't have PDO lock */
2918 if (!S_ISDIR(lu_object_attr(&o->mot_obj)))
2924 fid_build_reg_res_name(mdt_object_fid(o), res_id);
2925 dlmflags |= LDLM_FL_ATOMIC_CB;
2928 * Take PDO lock on whole directory and build correct @res_id for lock
2929 * on part of directory.
2931 if (lh->mlh_pdo_hash != 0) {
2932 LASSERT(lh->mlh_type == MDT_PDO_LOCK);
2933 mdt_lock_pdo_mode(info, o, lh);
2934 if (lh->mlh_pdo_mode != LCK_NL) {
2936 * Do not use LDLM_FL_LOCAL_ONLY for parallel lock, it
2937 * is never going to be sent to client and we do not
2938 * want it slowed down due to possible cancels.
2940 policy->l_inodebits.bits =
2941 *ibits & MDS_INODELOCK_UPDATE;
2942 policy->l_inodebits.try_bits =
2943 trybits & MDS_INODELOCK_UPDATE;
2944 /* at least one of them should be set */
2945 LASSERT(policy->l_inodebits.bits |
2946 policy->l_inodebits.try_bits);
2947 rc = mdt_fid_lock(ns, &lh->mlh_pdo_lh, lh->mlh_pdo_mode,
2948 policy, res_id, dlmflags,
2949 info->mti_exp == NULL ? NULL :
2950 &info->mti_exp->exp_handle.h_cookie);
2951 if (unlikely(rc != 0))
2952 GOTO(out_unlock, rc);
2956 * Finish res_id initializing by name hash marking part of
2957 * directory which is taking modification.
2959 res_id->name[LUSTRE_RES_ID_HSH_OFF] = lh->mlh_pdo_hash;
2962 policy->l_inodebits.bits = *ibits;
2963 policy->l_inodebits.try_bits = trybits;
2966 * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
2967 * going to be sent to client. If it is - mdt_intent_policy() path will
2968 * fix it up and turn FL_LOCAL flag off.
2970 rc = mdt_fid_lock(ns, &lh->mlh_reg_lh, lh->mlh_reg_mode, policy,
2971 res_id, LDLM_FL_LOCAL_ONLY | dlmflags,
2972 info->mti_exp == NULL ? NULL :
2973 &info->mti_exp->exp_handle.h_cookie);
2976 mdt_object_unlock(info, o, lh, 1);
2977 else if (unlikely(OBD_FAIL_PRECHECK(OBD_FAIL_MDS_PDO_LOCK)) &&
2978 lh->mlh_pdo_hash != 0 &&
2979 (lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX))
2980 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK, 15);
2982 /* Return successfully acquired bits to a caller */
2984 struct ldlm_lock *lock = ldlm_handle2lock(&lh->mlh_reg_lh);
2987 *ibits = lock->l_policy_data.l_inodebits.bits;
2988 LDLM_LOCK_PUT(lock);
2994 mdt_object_lock_internal(struct mdt_thread_info *info, struct mdt_object *o,
2995 struct mdt_lock_handle *lh, __u64 *ibits,
2996 __u64 trybits, bool cos_incompat)
2998 struct mdt_lock_handle *local_lh = NULL;
3002 if (!mdt_object_remote(o)) {
3003 rc = mdt_object_local_lock(info, o, lh, ibits, trybits,
3008 /* XXX do not support PERM/LAYOUT/XATTR lock for remote object yet */
3009 *ibits &= ~(MDS_INODELOCK_PERM | MDS_INODELOCK_LAYOUT |
3010 MDS_INODELOCK_XATTR);
3012 /* Only enqueue LOOKUP lock for remote object */
3013 if (*ibits & MDS_INODELOCK_LOOKUP) {
3014 __u64 local = MDS_INODELOCK_LOOKUP;
3016 rc = mdt_object_local_lock(info, o, lh, &local, 0,
3024 if ((*ibits | trybits) & MDS_INODELOCK_UPDATE) {
3025 /* Sigh, PDO needs to enqueue 2 locks right now, but
3026 * enqueue RPC can only request 1 lock, to avoid extra
3027 * RPC, so it will instead enqueue EX lock for remote
3028 * object anyway XXX*/
3029 if (lh->mlh_type == MDT_PDO_LOCK &&
3030 lh->mlh_pdo_hash != 0) {
3031 CDEBUG(D_INFO, "%s: "DFID" convert PDO lock to"
3032 "EX lock.\n", mdt_obd_name(info->mti_mdt),
3033 PFID(mdt_object_fid(o)));
3034 lh->mlh_pdo_hash = 0;
3035 lh->mlh_rreg_mode = LCK_EX;
3036 lh->mlh_type = MDT_REG_LOCK;
3039 rc = mdt_remote_object_lock_try(info, o, mdt_object_fid(o),
3042 ibits, trybits, false);
3043 if (rc != ELDLM_OK) {
3044 if (local_lh != NULL)
3045 mdt_object_unlock(info, o, local_lh, rc);
3053 int mdt_object_lock(struct mdt_thread_info *info, struct mdt_object *o,
3054 struct mdt_lock_handle *lh, __u64 ibits)
3056 return mdt_object_lock_internal(info, o, lh, &ibits, 0, false);
3059 int mdt_reint_object_lock(struct mdt_thread_info *info, struct mdt_object *o,
3060 struct mdt_lock_handle *lh, __u64 ibits,
3063 LASSERT(lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX);
3064 return mdt_object_lock_internal(info, o, lh, &ibits, 0,
3068 int mdt_object_lock_try(struct mdt_thread_info *info, struct mdt_object *o,
3069 struct mdt_lock_handle *lh, __u64 *ibits,
3070 __u64 trybits, bool cos_incompat)
3072 bool trylock_only = *ibits == 0;
3075 LASSERT(!(*ibits & trybits));
3076 rc = mdt_object_lock_internal(info, o, lh, ibits, trybits,
3078 if (rc && trylock_only) { /* clear error for try ibits lock only */
3079 LASSERT(*ibits == 0);
3086 * Save a lock within request object.
3088 * Keep the lock referenced until whether client ACK or transaction
3089 * commit happens or release the lock immediately depending on input
3090 * parameters. If COS is ON, a write lock is converted to COS lock
3093 * \param info thead info object
3094 * \param h lock handle
3095 * \param mode lock mode
3096 * \param decref force immediate lock releasing
3098 void mdt_save_lock(struct mdt_thread_info *info, struct lustre_handle *h,
3099 enum ldlm_mode mode, int decref)
3103 if (lustre_handle_is_used(h)) {
3104 if (decref || !info->mti_has_trans ||
3105 !(mode & (LCK_PW | LCK_EX))) {
3106 mdt_fid_unlock(h, mode);
3108 struct mdt_device *mdt = info->mti_mdt;
3109 struct ldlm_lock *lock = ldlm_handle2lock(h);
3110 struct ptlrpc_request *req = mdt_info_req(info);
3111 bool cos = mdt_cos_is_enabled(mdt);
3112 bool convert_lock = !cos && mdt_slc_is_enabled(mdt);
3114 LASSERTF(lock != NULL, "no lock for cookie %#llx\n",
3117 /* there is no request if mdt_object_unlock() is called
3118 * from mdt_export_cleanup()->mdt_add_dirty_flag() */
3119 if (likely(req != NULL)) {
3120 LDLM_DEBUG(lock, "save lock request %p reply "
3121 "state %p transno %lld\n", req,
3122 req->rq_reply_state, req->rq_transno);
3124 ldlm_lock_mode_downgrade(lock, LCK_COS);
3127 if (req->rq_export->exp_disconnected)
3128 mdt_fid_unlock(h, mode);
3130 ptlrpc_save_lock(req, h, mode, cos,
3133 mdt_fid_unlock(h, mode);
3135 if (mdt_is_lock_sync(lock)) {
3136 CDEBUG(D_HA, "found sync-lock,"
3137 " async commit started\n");
3138 mdt_device_commit_async(info->mti_env,
3141 LDLM_LOCK_PUT(lock);
3150 * Save cross-MDT lock in uncommitted_slc_locks
3152 * Keep the lock referenced until transaction commit happens or release the lock
3153 * immediately depending on input parameters.
3155 * \param info thead info object
3156 * \param h lock handle
3157 * \param mode lock mode
3158 * \param decref force immediate lock releasing
3160 static void mdt_save_remote_lock(struct mdt_thread_info *info,
3161 struct mdt_object *o, struct lustre_handle *h,
3162 enum ldlm_mode mode, int decref)
3166 if (lustre_handle_is_used(h)) {
3167 struct ldlm_lock *lock = ldlm_handle2lock(h);
3170 (lock->l_policy_data.l_inodebits.bits &
3171 (MDS_INODELOCK_XATTR | MDS_INODELOCK_UPDATE)))
3172 mo_invalidate(info->mti_env, mdt_object_child(o));
3174 if (decref || !info->mti_has_trans ||
3175 !(mode & (LCK_PW | LCK_EX))) {
3176 ldlm_lock_decref_and_cancel(h, mode);
3177 LDLM_LOCK_PUT(lock);
3179 struct ptlrpc_request *req = mdt_info_req(info);
3181 LASSERT(req != NULL);
3182 tgt_save_slc_lock(&info->mti_mdt->mdt_lut, lock,
3184 ldlm_lock_decref(h, mode);
3193 * Unlock mdt object.
3195 * Immeditely release the regular lock and the PDO lock or save the
3196 * lock in request and keep them referenced until client ACK or
3197 * transaction commit.
3199 * \param info thread info object
3200 * \param o mdt object
3201 * \param lh mdt lock handle referencing regular and PDO locks
3202 * \param decref force immediate lock releasing
3204 * XXX o is not used and may be NULL, see hsm_cdt_request_completed().
3206 void mdt_object_unlock(struct mdt_thread_info *info, struct mdt_object *o,
3207 struct mdt_lock_handle *lh, int decref)
3211 mdt_save_lock(info, &lh->mlh_pdo_lh, lh->mlh_pdo_mode, decref);
3212 mdt_save_lock(info, &lh->mlh_reg_lh, lh->mlh_reg_mode, decref);
3213 mdt_save_remote_lock(info, o, &lh->mlh_rreg_lh, lh->mlh_rreg_mode,
3219 struct mdt_object *mdt_object_find_lock(struct mdt_thread_info *info,
3220 const struct lu_fid *f,
3221 struct mdt_lock_handle *lh,
3224 struct mdt_object *o;
3226 o = mdt_object_find(info->mti_env, info->mti_mdt, f);
3230 rc = mdt_object_lock(info, o, lh, ibits);
3232 mdt_object_put(info->mti_env, o);
3239 void mdt_object_unlock_put(struct mdt_thread_info * info,
3240 struct mdt_object * o,
3241 struct mdt_lock_handle *lh,
3244 mdt_object_unlock(info, o, lh, decref);
3245 mdt_object_put(info->mti_env, o);
3249 * Generic code handling requests that have struct mdt_body passed in:
3251 * - extract mdt_body from request and save it in @info, if present;
3253 * - create lu_object, corresponding to the fid in mdt_body, and save it in
3256 * - if HABEO_CORPUS flag is set for this request type check whether object
3257 * actually exists on storage (lu_object_exists()).
3260 static int mdt_body_unpack(struct mdt_thread_info *info,
3261 enum tgt_handler_flags flags)
3263 const struct mdt_body *body;
3264 struct mdt_object *obj;
3265 const struct lu_env *env;
3266 struct req_capsule *pill;
3270 env = info->mti_env;
3271 pill = info->mti_pill;
3273 body = info->mti_body = req_capsule_client_get(pill, &RMF_MDT_BODY);
3277 if (!(body->mbo_valid & OBD_MD_FLID))
3280 if (!fid_is_sane(&body->mbo_fid1)) {
3281 CERROR("Invalid fid: "DFID"\n", PFID(&body->mbo_fid1));
3285 obj = mdt_object_find(env, info->mti_mdt, &body->mbo_fid1);
3287 if ((flags & HABEO_CORPUS) && !mdt_object_exists(obj)) {
3288 mdt_object_put(env, obj);
3291 info->mti_object = obj;
3300 static int mdt_unpack_req_pack_rep(struct mdt_thread_info *info,
3301 enum tgt_handler_flags flags)
3303 struct req_capsule *pill = info->mti_pill;
3307 if (req_capsule_has_field(pill, &RMF_MDT_BODY, RCL_CLIENT))
3308 rc = mdt_body_unpack(info, flags);
3312 if (rc == 0 && (flags & HABEO_REFERO)) {
3314 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
3315 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER,
3317 if (req_capsule_has_field(pill, &RMF_LOGCOOKIES, RCL_SERVER))
3318 req_capsule_set_size(pill, &RMF_LOGCOOKIES,
3321 /* Set ACL reply buffer size as LUSTRE_POSIX_ACL_MAX_SIZE_OLD
3322 * by default. If the target object has more ACL entries, then
3323 * enlarge the buffer when necessary. */
3324 if (req_capsule_has_field(pill, &RMF_ACL, RCL_SERVER))
3325 req_capsule_set_size(pill, &RMF_ACL, RCL_SERVER,
3326 LUSTRE_POSIX_ACL_MAX_SIZE_OLD);
3328 rc = req_capsule_server_pack(pill);
3333 void mdt_lock_handle_init(struct mdt_lock_handle *lh)
3335 lh->mlh_type = MDT_NUL_LOCK;
3336 lh->mlh_reg_lh.cookie = 0ull;
3337 lh->mlh_reg_mode = LCK_MINMODE;
3338 lh->mlh_pdo_lh.cookie = 0ull;
3339 lh->mlh_pdo_mode = LCK_MINMODE;
3340 lh->mlh_rreg_lh.cookie = 0ull;
3341 lh->mlh_rreg_mode = LCK_MINMODE;
3344 void mdt_lock_handle_fini(struct mdt_lock_handle *lh)
3346 LASSERT(!lustre_handle_is_used(&lh->mlh_reg_lh));
3347 LASSERT(!lustre_handle_is_used(&lh->mlh_pdo_lh));
3351 * Initialize fields of struct mdt_thread_info. Other fields are left in
3352 * uninitialized state, because it's too expensive to zero out whole
3353 * mdt_thread_info (> 1K) on each request arrival.
3355 void mdt_thread_info_init(struct ptlrpc_request *req,
3356 struct mdt_thread_info *info)
3360 info->mti_pill = &req->rq_pill;
3363 for (i = 0; i < ARRAY_SIZE(info->mti_lh); i++)
3364 mdt_lock_handle_init(&info->mti_lh[i]);
3366 /* mdt device: it can be NULL while CONNECT */
3367 if (req->rq_export) {
3368 info->mti_mdt = mdt_dev(req->rq_export->exp_obd->obd_lu_dev);
3369 info->mti_exp = req->rq_export;
3371 info->mti_mdt = NULL;
3372 info->mti_env = req->rq_svc_thread->t_env;
3373 info->mti_transno = lustre_msg_get_transno(req->rq_reqmsg);
3375 memset(&info->mti_attr, 0, sizeof(info->mti_attr));
3376 info->mti_big_buf = LU_BUF_NULL;
3377 info->mti_body = NULL;
3378 info->mti_object = NULL;
3379 info->mti_dlm_req = NULL;
3380 info->mti_has_trans = 0;
3381 info->mti_cross_ref = 0;
3382 info->mti_opdata = 0;
3383 info->mti_big_lmm_used = 0;
3384 info->mti_big_acl_used = 0;
3385 info->mti_som_valid = 0;
3387 info->mti_spec.no_create = 0;
3388 info->mti_spec.sp_rm_entry = 0;
3389 info->mti_spec.sp_permitted = 0;
3390 info->mti_spec.sp_migrate_close = 0;
3392 info->mti_spec.u.sp_ea.eadata = NULL;
3393 info->mti_spec.u.sp_ea.eadatalen = 0;
3396 void mdt_thread_info_fini(struct mdt_thread_info *info)
3400 if (info->mti_object != NULL) {
3401 mdt_object_put(info->mti_env, info->mti_object);
3402 info->mti_object = NULL;
3405 for (i = 0; i < ARRAY_SIZE(info->mti_lh); i++)
3406 mdt_lock_handle_fini(&info->mti_lh[i]);
3407 info->mti_env = NULL;
3408 info->mti_pill = NULL;
3409 info->mti_exp = NULL;
3411 if (unlikely(info->mti_big_buf.lb_buf != NULL))
3412 lu_buf_free(&info->mti_big_buf);
3415 struct mdt_thread_info *tsi2mdt_info(struct tgt_session_info *tsi)
3417 struct mdt_thread_info *mti;
3419 mti = mdt_th_info(tsi->tsi_env);
3420 LASSERT(mti != NULL);
3422 mdt_thread_info_init(tgt_ses_req(tsi), mti);
3423 if (tsi->tsi_corpus != NULL) {
3424 mti->mti_object = mdt_obj(tsi->tsi_corpus);
3425 lu_object_get(tsi->tsi_corpus);
3427 mti->mti_body = tsi->tsi_mdt_body;
3428 mti->mti_dlm_req = tsi->tsi_dlm_req;
3433 static int mdt_tgt_connect(struct tgt_session_info *tsi)
3435 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_DELAY_CONDITIONAL) &&
3437 tsi2mdt_info(tsi)->mti_mdt->mdt_seq_site.ss_node_id) {
3438 set_current_state(TASK_UNINTERRUPTIBLE);
3439 schedule_timeout(msecs_to_jiffies(3 * MSEC_PER_SEC));
3442 return tgt_connect(tsi);
3445 static int mdt_intent_glimpse(enum ldlm_intent_flags it_opc,
3446 struct mdt_thread_info *info,
3447 struct ldlm_lock **lockp, __u64 flags)
3449 return mdt_glimpse_enqueue(info, info->mti_mdt->mdt_namespace,
3452 static int mdt_intent_brw(enum ldlm_intent_flags it_opc,
3453 struct mdt_thread_info *info,
3454 struct ldlm_lock **lockp, __u64 flags)
3456 return mdt_brw_enqueue(info, info->mti_mdt->mdt_namespace,
3460 int mdt_intent_lock_replace(struct mdt_thread_info *info,
3461 struct ldlm_lock **lockp,
3462 struct mdt_lock_handle *lh,
3463 __u64 flags, int result)
3465 struct ptlrpc_request *req = mdt_info_req(info);
3466 struct ldlm_lock *lock = *lockp;
3467 struct ldlm_lock *new_lock;
3469 /* If possible resent found a lock, @lh is set to its handle */
3470 new_lock = ldlm_handle2lock_long(&lh->mlh_reg_lh, 0);
3472 if (new_lock == NULL && (flags & LDLM_FL_INTENT_ONLY)) {
3473 lh->mlh_reg_lh.cookie = 0;
3477 if (new_lock == NULL && (flags & LDLM_FL_RESENT)) {
3478 /* Lock is pinned by ldlm_handle_enqueue0() as it is
3479 * a resend case, however, it could be already destroyed
3480 * due to client eviction or a raced cancel RPC. */
3481 LDLM_DEBUG_NOLOCK("Invalid lock handle %#llx\n",
3482 lh->mlh_reg_lh.cookie);
3483 lh->mlh_reg_lh.cookie = 0;
3487 LASSERTF(new_lock != NULL,
3488 "lockh %#llx flags %#llx : rc = %d\n",
3489 lh->mlh_reg_lh.cookie, flags, result);
3492 * If we've already given this lock to a client once, then we should
3493 * have no readers or writers. Otherwise, we should have one reader
3494 * _or_ writer ref (which will be zeroed below) before returning the
3497 if (new_lock->l_export == req->rq_export) {
3498 LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
3500 LASSERT(new_lock->l_export == NULL);
3501 LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
3506 if (new_lock->l_export == req->rq_export) {
3508 * Already gave this to the client, which means that we
3509 * reconstructed a reply.
3511 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
3514 LDLM_LOCK_RELEASE(new_lock);
3515 lh->mlh_reg_lh.cookie = 0;
3516 RETURN(ELDLM_LOCK_REPLACED);
3520 * Fixup the lock to be given to the client.
3522 lock_res_and_lock(new_lock);
3523 /* Zero new_lock->l_readers and new_lock->l_writers without triggering
3524 * possible blocking AST. */
3525 while (new_lock->l_readers > 0) {
3526 lu_ref_del(&new_lock->l_reference, "reader", new_lock);
3527 lu_ref_del(&new_lock->l_reference, "user", new_lock);
3528 new_lock->l_readers--;
3530 while (new_lock->l_writers > 0) {
3531 lu_ref_del(&new_lock->l_reference, "writer", new_lock);
3532 lu_ref_del(&new_lock->l_reference, "user", new_lock);
3533 new_lock->l_writers--;
3536 new_lock->l_export = class_export_lock_get(req->rq_export, new_lock);
3537 new_lock->l_blocking_ast = lock->l_blocking_ast;
3538 new_lock->l_completion_ast = lock->l_completion_ast;
3539 if (ldlm_has_dom(new_lock))
3540 new_lock->l_glimpse_ast = ldlm_server_glimpse_ast;
3541 new_lock->l_remote_handle = lock->l_remote_handle;
3542 new_lock->l_flags &= ~LDLM_FL_LOCAL;
3544 unlock_res_and_lock(new_lock);
3546 cfs_hash_add(new_lock->l_export->exp_lock_hash,
3547 &new_lock->l_remote_handle,
3548 &new_lock->l_exp_hash);
3550 LDLM_LOCK_RELEASE(new_lock);
3551 lh->mlh_reg_lh.cookie = 0;
3553 RETURN(ELDLM_LOCK_REPLACED);
3556 void mdt_intent_fixup_resent(struct mdt_thread_info *info,
3557 struct ldlm_lock *new_lock,
3558 struct mdt_lock_handle *lh, __u64 flags)
3560 struct ptlrpc_request *req = mdt_info_req(info);
3561 struct ldlm_request *dlmreq;
3563 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
3566 dlmreq = req_capsule_client_get(info->mti_pill, &RMF_DLM_REQ);
3568 /* Check if this is a resend case (MSG_RESENT is set on RPC) and a
3569 * lock was found by ldlm_handle_enqueue(); if so @lh must be
3571 if (flags & LDLM_FL_RESENT) {
3572 lh->mlh_reg_lh.cookie = new_lock->l_handle.h_cookie;
3573 lh->mlh_reg_mode = new_lock->l_granted_mode;
3575 LDLM_DEBUG(new_lock, "Restoring lock cookie");
3576 DEBUG_REQ(D_DLMTRACE, req, "restoring lock cookie %#llx",
3577 lh->mlh_reg_lh.cookie);
3582 * If the xid matches, then we know this is a resent request, and allow
3583 * it. (It's probably an OPEN, for which we don't send a lock.
3585 if (req_can_reconstruct(req, NULL))
3589 * This remote handle isn't enqueued, so we never received or processed
3590 * this request. Clear MSG_RESENT, because it can be handled like any
3591 * normal request now.
3593 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
3595 DEBUG_REQ(D_DLMTRACE, req, "no existing lock with rhandle %#llx",
3596 dlmreq->lock_handle[0].cookie);
3599 static int mdt_intent_getxattr(enum ldlm_intent_flags it_opc,
3600 struct mdt_thread_info *info,
3601 struct ldlm_lock **lockp,
3604 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
3605 struct ldlm_reply *ldlm_rep = NULL;
3610 * Initialize lhc->mlh_reg_lh either from a previously granted lock
3611 * (for the resend case) or a new lock. Below we will use it to
3612 * replace the original lock.
3614 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
3615 if (!lustre_handle_is_used(&lhc->mlh_reg_lh)) {
3616 mdt_lock_reg_init(lhc, (*lockp)->l_req_mode);
3617 rc = mdt_object_lock(info, info->mti_object, lhc,
3618 MDS_INODELOCK_XATTR);
3623 rc = mdt_getxattr(info);
3625 if (mdt_info_req(info)->rq_repmsg != NULL)
3626 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
3628 if (ldlm_rep == NULL ||
3629 OBD_FAIL_CHECK(OBD_FAIL_MDS_XATTR_REP)) {
3630 mdt_object_unlock(info, info->mti_object, lhc, 1);
3631 RETURN(err_serious(-EFAULT));
3634 ldlm_rep->lock_policy_res2 = clear_serious(rc);
3636 /* This is left for interop instead of adding a new interop flag.
3638 #if LUSTRE_VERSION_CODE > OBD_OCD_VERSION(3, 0, 0, 0)
3639 if (ldlm_rep->lock_policy_res2) {
3640 mdt_object_unlock(info, info->mti_object, lhc, 1);
3641 RETURN(ELDLM_LOCK_ABORTED);
3645 rc = mdt_intent_lock_replace(info, lockp, lhc, flags, rc);
3649 static int mdt_intent_getattr(enum ldlm_intent_flags it_opc,
3650 struct mdt_thread_info *info,
3651 struct ldlm_lock **lockp,
3654 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
3656 struct ldlm_reply *ldlm_rep;
3657 struct mdt_body *reqbody;
3658 struct mdt_body *repbody;
3662 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
3665 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
3668 info->mti_cross_ref = !!(reqbody->mbo_valid & OBD_MD_FLCROSSREF);
3669 repbody->mbo_eadatasize = 0;
3670 repbody->mbo_aclsize = 0;
3674 child_bits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
3677 child_bits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE |
3681 CERROR("%s: unsupported intent %#x\n",
3682 mdt_obd_name(info->mti_mdt), (unsigned int)it_opc);
3683 GOTO(out_shrink, rc = -EINVAL);
3686 rc = mdt_init_ucred_intent_getattr(info, reqbody);
3688 GOTO(out_shrink, rc);
3690 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
3691 mdt_set_disposition(info, ldlm_rep, DISP_IT_EXECD);
3693 /* Get lock from request for possible resent case. */
3694 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
3696 rc = mdt_getattr_name_lock(info, lhc, child_bits, ldlm_rep);
3697 ldlm_rep->lock_policy_res2 = clear_serious(rc);
3699 if (mdt_get_disposition(ldlm_rep, DISP_LOOKUP_NEG))
3700 ldlm_rep->lock_policy_res2 = 0;
3701 if (!mdt_get_disposition(ldlm_rep, DISP_LOOKUP_POS) ||
3702 ldlm_rep->lock_policy_res2) {
3703 lhc->mlh_reg_lh.cookie = 0ull;
3704 GOTO(out_ucred, rc = ELDLM_LOCK_ABORTED);
3707 rc = mdt_intent_lock_replace(info, lockp, lhc, flags, rc);
3710 mdt_exit_ucred(info);
3712 mdt_client_compatibility(info);
3713 rc2 = mdt_fix_reply(info);
3719 static int mdt_intent_layout(enum ldlm_intent_flags it_opc,
3720 struct mdt_thread_info *info,
3721 struct ldlm_lock **lockp,
3724 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_LAYOUT];
3725 struct md_layout_change layout = { .mlc_opc = MD_LAYOUT_NOP };
3726 struct layout_intent *intent;
3727 struct lu_fid *fid = &info->mti_tmp_fid2;
3728 struct mdt_object *obj = NULL;
3729 int layout_size = 0;
3733 fid_extract_from_res_name(fid, &(*lockp)->l_resource->lr_name);
3735 intent = req_capsule_client_get(info->mti_pill, &RMF_LAYOUT_INTENT);
3739 CDEBUG(D_INFO, DFID "got layout change request from client: "
3740 "opc:%u flags:%#x extent "DEXT"\n",
3741 PFID(fid), intent->li_opc, intent->li_flags,
3742 PEXT(&intent->li_extent));
3744 switch (intent->li_opc) {
3745 case LAYOUT_INTENT_TRUNC:
3746 case LAYOUT_INTENT_WRITE:
3747 layout.mlc_opc = MD_LAYOUT_WRITE;
3748 layout.mlc_intent = intent;
3750 case LAYOUT_INTENT_ACCESS:
3752 case LAYOUT_INTENT_READ:
3753 case LAYOUT_INTENT_GLIMPSE:
3754 case LAYOUT_INTENT_RELEASE:
3755 case LAYOUT_INTENT_RESTORE:
3756 CERROR("%s: Unsupported layout intent opc %d\n",
3757 mdt_obd_name(info->mti_mdt), intent->li_opc);
3761 CERROR("%s: Unknown layout intent opc %d\n",
3762 mdt_obd_name(info->mti_mdt), intent->li_opc);
3769 /* Get lock from request for possible resent case. */
3770 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
3772 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
3774 GOTO(out, rc = PTR_ERR(obj));
3777 if (mdt_object_exists(obj) && !mdt_object_remote(obj)) {
3778 /* if layout is going to be changed don't use the current EA
3779 * size but the maximum one. That buffer will be shrinked
3780 * to the actual size in req_capsule_shrink() before reply.
3782 if (layout.mlc_opc == MD_LAYOUT_WRITE) {
3783 layout_size = info->mti_mdt->mdt_max_mdsize;
3785 layout_size = mdt_attr_get_eabuf_size(info, obj);
3786 if (layout_size < 0)
3787 GOTO(out_obj, rc = layout_size);
3789 if (layout_size > info->mti_mdt->mdt_max_mdsize)
3790 info->mti_mdt->mdt_max_mdsize = layout_size;
3795 * set reply buffer size, so that ldlm_handle_enqueue0()->
3796 * ldlm_lvbo_fill() will fill the reply buffer with lovea.
3798 (*lockp)->l_lvb_type = LVB_T_LAYOUT;
3799 req_capsule_set_size(info->mti_pill, &RMF_DLM_LVB, RCL_SERVER,
3801 rc = req_capsule_server_pack(info->mti_pill);
3806 if (layout.mlc_opc != MD_LAYOUT_NOP) {
3807 struct lu_buf *buf = &layout.mlc_buf;
3810 * mdt_layout_change is a reint operation, when the request
3811 * is resent, layout write shouldn't reprocess it again.
3813 rc = mdt_check_resent(info, mdt_reconstruct_generic, lhc);
3815 GOTO(out_obj, rc = rc < 0 ? rc : 0);
3818 * There is another resent case: the client's job has been
3819 * done by another client, referring lod_declare_layout_change
3820 * -EALREADY case, and it became a operation w/o transaction,
3821 * so we should not do the layout change, otherwise
3822 * mdt_layout_change() will try to cancel the granted server
3823 * CR lock whose remote counterpart is still in hold on the
3824 * client, and a deadlock ensues.
3826 rc = mdt_check_resent_lock(info, obj, lhc);
3832 if (unlikely(req_is_replay(mdt_info_req(info)))) {
3833 buf->lb_buf = req_capsule_client_get(info->mti_pill,
3835 buf->lb_len = req_capsule_get_size(info->mti_pill,
3836 &RMF_EADATA, RCL_CLIENT);
3838 * If it's a replay of layout write intent RPC, the
3839 * client has saved the extended lovea when
3840 * it get reply then.
3842 if (buf->lb_len > 0)
3843 mdt_fix_lov_magic(info, buf->lb_buf);
3846 * Instantiate some layout components, if @buf contains
3847 * lovea, then it's a replay of the layout intent write
3850 rc = mdt_layout_change(info, obj, &layout);
3855 mdt_object_put(info->mti_env, obj);
3857 if (rc == 0 && lustre_handle_is_used(&lhc->mlh_reg_lh))
3858 rc = mdt_intent_lock_replace(info, lockp, lhc, flags, rc);
3861 lhc->mlh_reg_lh.cookie = 0;
3866 static int mdt_intent_open(enum ldlm_intent_flags it_opc,
3867 struct mdt_thread_info *info,
3868 struct ldlm_lock **lockp,
3871 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
3872 struct ldlm_reply *rep = NULL;
3876 static const struct req_format *intent_fmts[REINT_MAX] = {
3877 [REINT_CREATE] = &RQF_LDLM_INTENT_CREATE,
3878 [REINT_OPEN] = &RQF_LDLM_INTENT_OPEN
3883 opc = mdt_reint_opcode(mdt_info_req(info), intent_fmts);
3887 /* Get lock from request for possible resent case. */
3888 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
3890 rc = mdt_reint_internal(info, lhc, opc);
3892 /* Check whether the reply has been packed successfully. */
3893 if (mdt_info_req(info)->rq_repmsg != NULL)
3894 rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
3896 RETURN(err_serious(-EFAULT));
3898 /* MDC expects this in any case */
3900 mdt_set_disposition(info, rep, DISP_LOOKUP_EXECD);
3902 /* the open lock or the lock for cross-ref object should be
3903 * returned to the client */
3904 if (lustre_handle_is_used(&lhc->mlh_reg_lh) &&
3905 (rc == 0 || rc == -MDT_EREMOTE_OPEN)) {
3906 rep->lock_policy_res2 = 0;
3907 rc = mdt_intent_lock_replace(info, lockp, lhc, flags, rc);
3911 rep->lock_policy_res2 = clear_serious(rc);
3913 if (rep->lock_policy_res2 == -ENOENT &&
3914 mdt_get_disposition(rep, DISP_LOOKUP_NEG) &&
3915 !mdt_get_disposition(rep, DISP_OPEN_CREATE))
3916 rep->lock_policy_res2 = 0;
3918 lhc->mlh_reg_lh.cookie = 0ull;
3919 if (rc == -ENOTCONN || rc == -ENODEV ||
3920 rc == -EOVERFLOW) { /**< if VBR failure then return error */
3922 * If it is the disconnect error (ENODEV & ENOCONN), the error
3923 * will be returned by rq_status, and client at ptlrpc layer
3924 * will detect this, then disconnect, reconnect the import
3925 * immediately, instead of impacting the following the rpc.
3930 * For other cases, the error will be returned by intent, and client
3931 * will retrieve the result from intent.
3933 RETURN(ELDLM_LOCK_ABORTED);
3936 static int mdt_intent_opc(enum ldlm_intent_flags it_opc,
3937 struct mdt_thread_info *info,
3938 struct ldlm_lock **lockp,
3939 u64 flags /* LDLM_FL_* */)
3941 struct req_capsule *pill = info->mti_pill;
3942 struct ptlrpc_request *req = mdt_info_req(info);
3943 const struct req_format *it_format;
3944 int (*it_handler)(enum ldlm_intent_flags,
3945 struct mdt_thread_info *,
3946 struct ldlm_lock **,
3948 enum tgt_handler_flags it_handler_flags = 0;
3949 struct ldlm_reply *rep;
3955 case IT_OPEN|IT_CREAT:
3957 * OCREAT is not a MUTABOR request since the file may
3958 * already exist. We do the extra check of
3959 * OBD_CONNECT_RDONLY in mdt_reint_open() when we
3960 * really need to create the object.
3962 it_format = &RQF_LDLM_INTENT;
3963 it_handler = &mdt_intent_open;
3967 it_format = &RQF_LDLM_INTENT_GETATTR;
3968 it_handler = &mdt_intent_getattr;
3969 it_handler_flags = HABEO_REFERO;
3972 it_format = &RQF_LDLM_INTENT_GETXATTR;
3973 it_handler = &mdt_intent_getxattr;
3974 it_handler_flags = HABEO_CORPUS;
3977 it_format = &RQF_LDLM_INTENT_LAYOUT;
3978 it_handler = &mdt_intent_layout;
3981 it_format = &RQF_LDLM_INTENT;
3982 it_handler = &mdt_intent_glimpse;
3985 it_format = &RQF_LDLM_INTENT;
3986 it_handler = &mdt_intent_brw;
3988 case IT_QUOTA_DQACQ:
3989 case IT_QUOTA_CONN: {
3990 struct lu_device *qmt = info->mti_mdt->mdt_qmt_dev;
3993 RETURN(-EOPNOTSUPP);
3995 if (mdt_rdonly(req->rq_export))
3998 (*lockp)->l_lvb_type = LVB_T_LQUOTA;
3999 /* pass the request to quota master */
4000 rc = qmt_hdls.qmth_intent_policy(info->mti_env, qmt,
4001 mdt_info_req(info), lockp,
4006 CERROR("%s: unknown intent code %#x\n",
4007 mdt_obd_name(info->mti_mdt), it_opc);
4011 req_capsule_extend(pill, it_format);
4013 rc = mdt_unpack_req_pack_rep(info, it_handler_flags);
4017 if (it_handler_flags & MUTABOR && mdt_rdonly(req->rq_export))
4020 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_INTENT_DELAY, 10);
4022 /* execute policy */
4023 rc = (*it_handler)(it_opc, info, lockp, flags);
4025 /* Check whether the reply has been packed successfully. */
4026 if (req->rq_repmsg != NULL) {
4027 rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
4028 rep->lock_policy_res2 =
4029 ptlrpc_status_hton(rep->lock_policy_res2);
4035 static void mdt_ptlrpc_stats_update(struct ptlrpc_request *req,
4036 enum ldlm_intent_flags it_opc)
4038 struct lprocfs_stats *srv_stats = ptlrpc_req2svc(req)->srv_stats;
4040 /* update stats when IT code is known */
4041 if (srv_stats != NULL)
4042 lprocfs_counter_incr(srv_stats,
4043 PTLRPC_LAST_CNTR + (it_opc == IT_GLIMPSE ?
4044 LDLM_GLIMPSE_ENQUEUE : LDLM_IBITS_ENQUEUE));
4047 static int mdt_intent_policy(struct ldlm_namespace *ns,
4048 struct ldlm_lock **lockp, void *req_cookie,
4049 enum ldlm_mode mode, __u64 flags, void *data)
4051 struct tgt_session_info *tsi;
4052 struct mdt_thread_info *info;
4053 struct ptlrpc_request *req = req_cookie;
4054 struct ldlm_intent *it;
4055 struct req_capsule *pill;
4056 const struct ldlm_lock_desc *ldesc;
4061 LASSERT(req != NULL);
4063 tsi = tgt_ses_info(req->rq_svc_thread->t_env);
4065 info = tsi2mdt_info(tsi);
4066 LASSERT(info != NULL);
4067 pill = info->mti_pill;
4068 LASSERT(pill->rc_req == req);
4069 ldesc = &info->mti_dlm_req->lock_desc;
4071 if (req->rq_reqmsg->lm_bufcount > DLM_INTENT_IT_OFF) {
4072 req_capsule_extend(pill, &RQF_LDLM_INTENT_BASIC);
4073 it = req_capsule_client_get(pill, &RMF_LDLM_INTENT);
4075 mdt_ptlrpc_stats_update(req, it->opc);
4076 rc = mdt_intent_opc(it->opc, info, lockp, flags);
4080 /* Lock without inodebits makes no sense and will oops
4081 * later in ldlm. Let's check it now to see if we have
4082 * ibits corrupted somewhere in mdt_intent_opc().
4083 * The case for client miss to set ibits has been
4084 * processed by others. */
4085 LASSERT(ergo(ldesc->l_resource.lr_type == LDLM_IBITS,
4086 ldesc->l_policy_data.l_inodebits.bits != 0));
4088 rc = err_serious(-EFAULT);
4091 /* No intent was provided */
4092 req_capsule_set_size(pill, &RMF_DLM_LVB, RCL_SERVER, 0);
4093 rc = req_capsule_server_pack(pill);
4095 rc = err_serious(rc);
4097 mdt_thread_info_fini(info);
4101 static void mdt_deregister_seq_exp(struct mdt_device *mdt)
4103 struct seq_server_site *ss = mdt_seq_site(mdt);
4105 if (ss->ss_node_id == 0)
4108 if (ss->ss_client_seq != NULL) {
4109 lustre_deregister_lwp_item(&ss->ss_client_seq->lcs_exp);
4110 ss->ss_client_seq->lcs_exp = NULL;
4113 if (ss->ss_server_fld != NULL) {
4114 lustre_deregister_lwp_item(&ss->ss_server_fld->lsf_control_exp);
4115 ss->ss_server_fld->lsf_control_exp = NULL;
4119 static void mdt_seq_fini_cli(struct mdt_device *mdt)
4121 struct seq_server_site *ss = mdt_seq_site(mdt);
4126 if (ss->ss_server_seq != NULL)
4127 seq_server_set_cli(NULL, ss->ss_server_seq, NULL);
4130 static int mdt_seq_fini(const struct lu_env *env, struct mdt_device *mdt)
4132 mdt_seq_fini_cli(mdt);
4133 mdt_deregister_seq_exp(mdt);
4135 return seq_site_fini(env, mdt_seq_site(mdt));
4139 * It will retrieve its FLDB entries from MDT0, and it only happens
4140 * when upgrading existent FS to 2.6 or when local FLDB is corrupted,
4141 * and it needs to refresh FLDB from the MDT0.
4143 static int mdt_register_lwp_callback(void *data)
4146 struct mdt_device *mdt = data;
4147 struct lu_server_fld *fld = mdt_seq_site(mdt)->ss_server_fld;
4151 LASSERT(mdt_seq_site(mdt)->ss_node_id != 0);
4153 rc = lu_env_init(&env, LCT_MD_THREAD);
4155 CERROR("%s: cannot init env: rc = %d\n", mdt_obd_name(mdt), rc);
4159 /* Allocate new sequence now to avoid creating local transaction
4160 * in the normal transaction process */
4161 rc = seq_server_check_and_alloc_super(&env,
4162 mdt_seq_site(mdt)->ss_server_seq);
4167 rc = fld_update_from_controller(&env, fld);
4169 CERROR("%s: cannot update controller: rc = %d\n",
4170 mdt_obd_name(mdt), rc);
4179 static int mdt_register_seq_exp(struct mdt_device *mdt)
4181 struct seq_server_site *ss = mdt_seq_site(mdt);
4182 char *lwp_name = NULL;
4185 if (ss->ss_node_id == 0)
4188 OBD_ALLOC(lwp_name, MAX_OBD_NAME);
4189 if (lwp_name == NULL)
4190 GOTO(out_free, rc = -ENOMEM);
4192 rc = tgt_name2lwp_name(mdt_obd_name(mdt), lwp_name, MAX_OBD_NAME, 0);
4196 rc = lustre_register_lwp_item(lwp_name, &ss->ss_client_seq->lcs_exp,
4201 rc = lustre_register_lwp_item(lwp_name,
4202 &ss->ss_server_fld->lsf_control_exp,
4203 mdt_register_lwp_callback, mdt);
4205 lustre_deregister_lwp_item(&ss->ss_client_seq->lcs_exp);
4206 ss->ss_client_seq->lcs_exp = NULL;
4210 if (lwp_name != NULL)
4211 OBD_FREE(lwp_name, MAX_OBD_NAME);
4217 * Init client sequence manager which is used by local MDS to talk to sequence
4218 * controller on remote node.
4220 static int mdt_seq_init_cli(const struct lu_env *env, struct mdt_device *mdt)
4222 struct seq_server_site *ss = mdt_seq_site(mdt);
4227 /* check if this is adding the first MDC and controller is not yet
4229 OBD_ALLOC_PTR(ss->ss_client_seq);
4230 if (ss->ss_client_seq == NULL)
4233 OBD_ALLOC(prefix, MAX_OBD_NAME + 5);
4234 if (prefix == NULL) {
4235 OBD_FREE_PTR(ss->ss_client_seq);
4236 ss->ss_client_seq = NULL;
4240 /* Note: seq_client_fini will be called in seq_site_fini */
4241 snprintf(prefix, MAX_OBD_NAME + 5, "ctl-%s", mdt_obd_name(mdt));
4242 rc = seq_client_init(ss->ss_client_seq, NULL, LUSTRE_SEQ_METADATA,
4243 prefix, ss->ss_node_id == 0 ? ss->ss_control_seq :
4245 OBD_FREE(prefix, MAX_OBD_NAME + 5);
4247 OBD_FREE_PTR(ss->ss_client_seq);
4248 ss->ss_client_seq = NULL;
4252 rc = seq_server_set_cli(env, ss->ss_server_seq, ss->ss_client_seq);
4257 static int mdt_seq_init(const struct lu_env *env, struct mdt_device *mdt)
4259 struct seq_server_site *ss;
4263 ss = mdt_seq_site(mdt);
4264 /* init sequence controller server(MDT0) */
4265 if (ss->ss_node_id == 0) {
4266 OBD_ALLOC_PTR(ss->ss_control_seq);
4267 if (ss->ss_control_seq == NULL)
4270 rc = seq_server_init(env, ss->ss_control_seq, mdt->mdt_bottom,
4271 mdt_obd_name(mdt), LUSTRE_SEQ_CONTROLLER,
4274 GOTO(out_seq_fini, rc);
4277 /* Init normal sequence server */
4278 OBD_ALLOC_PTR(ss->ss_server_seq);
4279 if (ss->ss_server_seq == NULL)
4280 GOTO(out_seq_fini, rc = -ENOMEM);
4282 rc = seq_server_init(env, ss->ss_server_seq, mdt->mdt_bottom,
4283 mdt_obd_name(mdt), LUSTRE_SEQ_SERVER, ss);
4285 GOTO(out_seq_fini, rc);
4287 /* init seq client for seq server to talk to seq controller(MDT0) */
4288 rc = mdt_seq_init_cli(env, mdt);
4290 GOTO(out_seq_fini, rc);
4292 if (ss->ss_node_id != 0)
4293 /* register controller export through lwp */
4294 rc = mdt_register_seq_exp(mdt);
4299 mdt_seq_fini(env, mdt);
4307 static int mdt_fld_fini(const struct lu_env *env,
4308 struct mdt_device *m)
4310 struct seq_server_site *ss = mdt_seq_site(m);
4313 if (ss && ss->ss_server_fld) {
4314 fld_server_fini(env, ss->ss_server_fld);
4315 OBD_FREE_PTR(ss->ss_server_fld);
4316 ss->ss_server_fld = NULL;
4322 static int mdt_fld_init(const struct lu_env *env,
4324 struct mdt_device *m)
4326 struct seq_server_site *ss;
4330 ss = mdt_seq_site(m);
4332 OBD_ALLOC_PTR(ss->ss_server_fld);
4333 if (ss->ss_server_fld == NULL)
4334 RETURN(rc = -ENOMEM);
4336 rc = fld_server_init(env, ss->ss_server_fld, m->mdt_bottom, uuid,
4339 OBD_FREE_PTR(ss->ss_server_fld);
4340 ss->ss_server_fld = NULL;
4347 static void mdt_stack_pre_fini(const struct lu_env *env,
4348 struct mdt_device *m, struct lu_device *top)
4350 struct lustre_cfg_bufs *bufs;
4351 struct lustre_cfg *lcfg;
4352 struct mdt_thread_info *info;
4357 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
4358 LASSERT(info != NULL);
4360 bufs = &info->mti_u.bufs;
4362 LASSERT(m->mdt_child_exp);
4363 LASSERT(m->mdt_child_exp->exp_obd);
4365 /* process cleanup, pass mdt obd name to get obd umount flags */
4366 /* XXX: this is needed because all layers are referenced by
4367 * objects (some of them are pinned by osd, for example *
4368 * the proper solution should be a model where object used
4369 * by osd only doesn't have mdt/mdd slices -bzzz */
4370 lustre_cfg_bufs_reset(bufs, mdt_obd_name(m));
4371 lustre_cfg_bufs_set_string(bufs, 1, NULL);
4372 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
4375 lustre_cfg_init(lcfg, LCFG_PRE_CLEANUP, bufs);
4377 top->ld_ops->ldo_process_config(env, top, lcfg);
4378 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens));
4382 static void mdt_stack_fini(const struct lu_env *env,
4383 struct mdt_device *m, struct lu_device *top)
4385 struct obd_device *obd = mdt2obd_dev(m);
4386 struct lustre_cfg_bufs *bufs;
4387 struct lustre_cfg *lcfg;
4388 struct mdt_thread_info *info;
4392 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
4393 LASSERT(info != NULL);
4395 lu_dev_del_linkage(top->ld_site, top);
4397 lu_site_purge(env, top->ld_site, -1);
4399 bufs = &info->mti_u.bufs;
4400 /* process cleanup, pass mdt obd name to get obd umount flags */
4401 /* another purpose is to let all layers to release their objects */
4402 lustre_cfg_bufs_reset(bufs, mdt_obd_name(m));
4407 lustre_cfg_bufs_set_string(bufs, 1, flags);
4408 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
4411 lustre_cfg_init(lcfg, LCFG_CLEANUP, bufs);
4414 top->ld_ops->ldo_process_config(env, top, lcfg);
4415 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens));
4417 lu_site_purge(env, top->ld_site, -1);
4419 m->mdt_child = NULL;
4420 m->mdt_bottom = NULL;
4422 obd_disconnect(m->mdt_child_exp);
4423 m->mdt_child_exp = NULL;
4425 obd_disconnect(m->mdt_bottom_exp);
4426 m->mdt_child_exp = NULL;
4429 static int mdt_connect_to_next(const struct lu_env *env, struct mdt_device *m,
4430 const char *next, struct obd_export **exp)
4432 struct obd_connect_data *data = NULL;
4433 struct obd_device *obd;
4437 OBD_ALLOC_PTR(data);
4439 GOTO(out, rc = -ENOMEM);
4441 obd = class_name2obd(next);
4443 CERROR("%s: can't locate next device: %s\n",
4444 mdt_obd_name(m), next);
4445 GOTO(out, rc = -ENOTCONN);
4448 data->ocd_connect_flags = OBD_CONNECT_VERSION;
4449 data->ocd_version = LUSTRE_VERSION_CODE;
4451 rc = obd_connect(NULL, exp, obd, &obd->obd_uuid, data, NULL);
4453 CERROR("%s: cannot connect to next dev %s (%d)\n",
4454 mdt_obd_name(m), next, rc);
4464 static int mdt_stack_init(const struct lu_env *env, struct mdt_device *mdt,
4465 struct lustre_cfg *cfg)
4467 char *dev = lustre_cfg_string(cfg, 0);
4468 int rc, name_size, uuid_size;
4469 char *name, *uuid, *p;
4470 struct lustre_cfg_bufs *bufs;
4471 struct lustre_cfg *lcfg;
4472 struct obd_device *obd;
4473 struct lustre_profile *lprof;
4474 struct lu_site *site;
4477 /* in 1.8 we had the only device in the stack - MDS.
4478 * 2.0 introduces MDT, MDD, OSD; MDT starts others internally.
4479 * in 2.3 OSD is instantiated by obd_mount.c, so we need
4480 * to generate names and setup MDT, MDD. MDT will be using
4481 * generated name to connect to MDD. for MDD the next device
4482 * will be LOD with name taken from so called "profile" which
4483 * is generated by mount_option line
4485 * 1.8 MGS generates config. commands like this:
4486 * #06 (104)mount_option 0: 1:lustre-MDT0000 2:lustre-mdtlov
4487 * #08 (120)setup 0:lustre-MDT0000 1:dev 2:type 3:lustre-MDT0000
4488 * 2.0 MGS generates config. commands like this:
4489 * #07 (112)mount_option 0: 1:lustre-MDT0000 2:lustre-MDT0000-mdtlov
4490 * #08 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
4491 * 3:lustre-MDT0000-mdtlov 4:f
4493 * we generate MDD name from MDT one, just replacing T with D
4495 * after all the preparations, the logical equivalent will be
4496 * #01 (160)setup 0:lustre-MDD0000 1:lustre-MDD0000_UUID 2:0
4497 * 3:lustre-MDT0000-mdtlov 4:f
4498 * #02 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
4499 * 3:lustre-MDD0000 4:f
4501 * notice we build the stack from down to top: MDD first, then MDT */
4503 name_size = MAX_OBD_NAME;
4504 uuid_size = MAX_OBD_NAME;
4506 OBD_ALLOC(name, name_size);
4507 OBD_ALLOC(uuid, uuid_size);
4508 if (name == NULL || uuid == NULL)
4509 GOTO(cleanup_mem, rc = -ENOMEM);
4511 OBD_ALLOC_PTR(bufs);
4513 GOTO(cleanup_mem, rc = -ENOMEM);
4516 p = strstr(name, "-MDT");
4518 GOTO(free_bufs, rc = -ENOMEM);
4521 snprintf(uuid, MAX_OBD_NAME, "%s_UUID", name);
4523 lprof = class_get_profile(lustre_cfg_string(cfg, 0));
4524 if (lprof == NULL || lprof->lp_dt == NULL) {
4525 CERROR("can't find the profile: %s\n",
4526 lustre_cfg_string(cfg, 0));
4527 GOTO(free_bufs, rc = -EINVAL);
4530 lustre_cfg_bufs_reset(bufs, name);
4531 lustre_cfg_bufs_set_string(bufs, 1, LUSTRE_MDD_NAME);
4532 lustre_cfg_bufs_set_string(bufs, 2, uuid);
4533 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
4535 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
4537 GOTO(put_profile, rc = -ENOMEM);
4538 lustre_cfg_init(lcfg, LCFG_ATTACH, bufs);
4540 rc = class_attach(lcfg);
4542 GOTO(lcfg_cleanup, rc);
4544 obd = class_name2obd(name);
4546 CERROR("Can not find obd %s (%s in config)\n",
4547 MDD_OBD_NAME, lustre_cfg_string(cfg, 0));
4548 GOTO(lcfg_cleanup, rc = -EINVAL);
4551 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens));
4553 lustre_cfg_bufs_reset(bufs, name);
4554 lustre_cfg_bufs_set_string(bufs, 1, uuid);
4555 lustre_cfg_bufs_set_string(bufs, 2, dev);
4556 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
4558 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
4560 GOTO(class_detach, rc = -ENOMEM);
4561 lustre_cfg_init(lcfg, LCFG_SETUP, bufs);
4563 rc = class_setup(obd, lcfg);
4565 GOTO(class_detach, rc);
4567 /* connect to MDD we just setup */
4568 rc = mdt_connect_to_next(env, mdt, name, &mdt->mdt_child_exp);
4570 GOTO(class_detach, rc);
4572 site = mdt->mdt_child_exp->exp_obd->obd_lu_dev->ld_site;
4574 LASSERT(mdt_lu_site(mdt) == NULL);
4575 mdt->mdt_lu_dev.ld_site = site;
4576 site->ls_top_dev = &mdt->mdt_lu_dev;
4577 mdt->mdt_child = lu2md_dev(mdt->mdt_child_exp->exp_obd->obd_lu_dev);
4579 /* now connect to bottom OSD */
4580 snprintf(name, MAX_OBD_NAME, "%s-osd", dev);
4581 rc = mdt_connect_to_next(env, mdt, name, &mdt->mdt_bottom_exp);
4583 GOTO(class_detach, rc);
4585 lu2dt_dev(mdt->mdt_bottom_exp->exp_obd->obd_lu_dev);
4587 rc = lu_env_refill((struct lu_env *)env);
4589 CERROR("Failure to refill session: '%d'\n", rc);
4591 lu_dev_add_linkage(site, &mdt->mdt_lu_dev);
4596 class_detach(obd, lcfg);
4598 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens));
4600 class_put_profile(lprof);
4605 OBD_FREE(name, name_size);
4607 OBD_FREE(uuid, uuid_size);
4611 /* setup quota master target on MDT0 */
4612 static int mdt_quota_init(const struct lu_env *env, struct mdt_device *mdt,
4613 struct lustre_cfg *cfg)
4615 struct obd_device *obd;
4616 char *dev = lustre_cfg_string(cfg, 0);
4617 char *qmtname, *uuid, *p;
4618 struct lustre_cfg_bufs *bufs;
4619 struct lustre_cfg *lcfg;
4620 struct lustre_profile *lprof;
4621 struct obd_connect_data *data;
4625 LASSERT(mdt->mdt_qmt_exp == NULL);
4626 LASSERT(mdt->mdt_qmt_dev == NULL);
4628 /* quota master is on MDT0 only for now */
4629 if (mdt->mdt_seq_site.ss_node_id != 0)
4632 /* MGS generates config commands which look as follows:
4633 * #01 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
4634 * 3:lustre-MDT0000-mdtlov 4:f
4636 * We generate the QMT name from the MDT one, just replacing MD with QM
4637 * after all the preparations, the logical equivalent will be:
4638 * #01 (160)setup 0:lustre-QMT0000 1:lustre-QMT0000_UUID 2:0
4639 * 3:lustre-MDT0000-osd 4:f */
4640 OBD_ALLOC(qmtname, MAX_OBD_NAME);
4641 OBD_ALLOC(uuid, UUID_MAX);
4642 OBD_ALLOC_PTR(bufs);
4643 OBD_ALLOC_PTR(data);
4644 if (qmtname == NULL || uuid == NULL || bufs == NULL || data == NULL)
4645 GOTO(cleanup_mem, rc = -ENOMEM);
4647 strcpy(qmtname, dev);
4648 p = strstr(qmtname, "-MDT");
4650 GOTO(cleanup_mem, rc = -ENOMEM);
4651 /* replace MD with QM */
4655 snprintf(uuid, UUID_MAX, "%s_UUID", qmtname);
4657 lprof = class_get_profile(lustre_cfg_string(cfg, 0));
4658 if (lprof == NULL || lprof->lp_dt == NULL) {
4659 CERROR("can't find profile for %s\n",
4660 lustre_cfg_string(cfg, 0));
4661 GOTO(cleanup_mem, rc = -EINVAL);
4664 lustre_cfg_bufs_reset(bufs, qmtname);
4665 lustre_cfg_bufs_set_string(bufs, 1, LUSTRE_QMT_NAME);
4666 lustre_cfg_bufs_set_string(bufs, 2, uuid);
4667 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
4669 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
4671 GOTO(put_profile, rc = -ENOMEM);
4672 lustre_cfg_init(lcfg, LCFG_ATTACH, bufs);
4674 rc = class_attach(lcfg);
4676 GOTO(lcfg_cleanup, rc);
4678 obd = class_name2obd(qmtname);
4680 CERROR("Can not find obd %s (%s in config)\n", qmtname,
4681 lustre_cfg_string(cfg, 0));
4682 GOTO(lcfg_cleanup, rc = -EINVAL);
4685 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens));
4687 lustre_cfg_bufs_reset(bufs, qmtname);
4688 lustre_cfg_bufs_set_string(bufs, 1, uuid);
4689 lustre_cfg_bufs_set_string(bufs, 2, dev);
4691 /* for quota, the next device should be the OSD device */
4692 lustre_cfg_bufs_set_string(bufs, 3,
4693 mdt->mdt_bottom->dd_lu_dev.ld_obd->obd_name);
4695 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
4697 GOTO(class_detach, rc = -ENOMEM);
4698 lustre_cfg_init(lcfg, LCFG_SETUP, bufs);
4700 rc = class_setup(obd, lcfg);
4702 GOTO(class_detach, rc);
4704 mdt->mdt_qmt_dev = obd->obd_lu_dev;
4706 /* configure local quota objects */
4707 rc = mdt->mdt_qmt_dev->ld_ops->ldo_prepare(env,
4711 GOTO(class_cleanup, rc);
4713 /* connect to quota master target */
4714 data->ocd_connect_flags = OBD_CONNECT_VERSION;
4715 data->ocd_version = LUSTRE_VERSION_CODE;
4716 rc = obd_connect(NULL, &mdt->mdt_qmt_exp, obd, &obd->obd_uuid,
4719 CERROR("cannot connect to quota master device %s (%d)\n",
4721 GOTO(class_cleanup, rc);
4727 class_manual_cleanup(obd);
4728 mdt->mdt_qmt_dev = NULL;
4732 class_detach(obd, lcfg);
4734 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens));
4736 class_put_profile(lprof);
4741 OBD_FREE(qmtname, MAX_OBD_NAME);
4743 OBD_FREE(uuid, UUID_MAX);
4749 /* Shutdown quota master target associated with mdt */
4750 static void mdt_quota_fini(const struct lu_env *env, struct mdt_device *mdt)
4754 if (mdt->mdt_qmt_exp == NULL)
4756 LASSERT(mdt->mdt_qmt_dev != NULL);
4758 /* the qmt automatically shuts down when the mdt disconnects */
4759 obd_disconnect(mdt->mdt_qmt_exp);
4760 mdt->mdt_qmt_exp = NULL;
4761 mdt->mdt_qmt_dev = NULL;
4765 /* mdt_getxattr() is used from mdt_intent_getxattr(), use this wrapper
4766 * for now. This will be removed along with converting rest of MDT code
4767 * to use tgt_session_info */
4768 static int mdt_tgt_getxattr(struct tgt_session_info *tsi)
4770 struct mdt_thread_info *info = tsi2mdt_info(tsi);
4773 rc = mdt_getxattr(info);
4775 mdt_thread_info_fini(info);
4779 #define OBD_FAIL_OST_READ_NET OBD_FAIL_OST_BRW_NET
4780 #define OBD_FAIL_OST_WRITE_NET OBD_FAIL_OST_BRW_NET
4781 #define OST_BRW_READ OST_READ
4782 #define OST_BRW_WRITE OST_WRITE
4784 static struct tgt_handler mdt_tgt_handlers[] = {
4785 TGT_RPC_HANDLER(MDS_FIRST_OPC,
4786 0, MDS_CONNECT, mdt_tgt_connect,
4787 &RQF_CONNECT, LUSTRE_OBD_VERSION),
4788 TGT_RPC_HANDLER(MDS_FIRST_OPC,
4789 0, MDS_DISCONNECT, tgt_disconnect,
4790 &RQF_MDS_DISCONNECT, LUSTRE_OBD_VERSION),
4791 TGT_RPC_HANDLER(MDS_FIRST_OPC,
4792 HABEO_REFERO, MDS_SET_INFO, mdt_set_info,
4793 &RQF_OBD_SET_INFO, LUSTRE_MDS_VERSION),
4794 TGT_MDT_HDL(0, MDS_GET_INFO, mdt_get_info),
4795 TGT_MDT_HDL(0 | HABEO_REFERO, MDS_GET_ROOT, mdt_get_root),
4796 TGT_MDT_HDL(HABEO_CORPUS, MDS_GETATTR, mdt_getattr),
4797 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_GETATTR_NAME,
4799 TGT_MDT_HDL(HABEO_CORPUS, MDS_GETXATTR, mdt_tgt_getxattr),
4800 TGT_MDT_HDL(0 | HABEO_REFERO, MDS_STATFS, mdt_statfs),
4801 TGT_MDT_HDL(0 | MUTABOR, MDS_REINT, mdt_reint),
4802 TGT_MDT_HDL(HABEO_CORPUS, MDS_CLOSE, mdt_close),
4803 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_READPAGE, mdt_readpage),
4804 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_SYNC, mdt_sync),
4805 TGT_MDT_HDL(0, MDS_QUOTACTL, mdt_quotactl),
4806 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO | MUTABOR, MDS_HSM_PROGRESS,
4808 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO | MUTABOR, MDS_HSM_CT_REGISTER,
4809 mdt_hsm_ct_register),
4810 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO | MUTABOR, MDS_HSM_CT_UNREGISTER,
4811 mdt_hsm_ct_unregister),
4812 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_HSM_STATE_GET,
4814 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO | MUTABOR, MDS_HSM_STATE_SET,
4816 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_HSM_ACTION, mdt_hsm_action),
4817 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_HSM_REQUEST,
4819 TGT_MDT_HDL(HABEO_CLAVIS | HABEO_CORPUS | HABEO_REFERO | MUTABOR,
4824 static struct tgt_handler mdt_io_ops[] = {
4825 TGT_OST_HDL_HP(HABEO_CORPUS | HABEO_REFERO, OST_BRW_READ, tgt_brw_read,
4827 TGT_OST_HDL_HP(HABEO_CORPUS | MUTABOR, OST_BRW_WRITE, tgt_brw_write,
4829 TGT_OST_HDL_HP(HABEO_CORPUS | HABEO_REFERO | MUTABOR,
4830 OST_PUNCH, mdt_punch_hdl,
4832 TGT_OST_HDL(HABEO_CORPUS | HABEO_REFERO, OST_SYNC, mdt_data_sync),
4835 static struct tgt_handler mdt_sec_ctx_ops[] = {
4836 TGT_SEC_HDL_VAR(0, SEC_CTX_INIT, mdt_sec_ctx_handle),
4837 TGT_SEC_HDL_VAR(0, SEC_CTX_INIT_CONT,mdt_sec_ctx_handle),
4838 TGT_SEC_HDL_VAR(0, SEC_CTX_FINI, mdt_sec_ctx_handle)
4841 static struct tgt_handler mdt_quota_ops[] = {
4842 TGT_QUOTA_HDL(HABEO_REFERO, QUOTA_DQACQ, mdt_quota_dqacq),
4845 static struct tgt_opc_slice mdt_common_slice[] = {
4847 .tos_opc_start = MDS_FIRST_OPC,
4848 .tos_opc_end = MDS_LAST_OPC,
4849 .tos_hs = mdt_tgt_handlers
4852 .tos_opc_start = OBD_FIRST_OPC,
4853 .tos_opc_end = OBD_LAST_OPC,
4854 .tos_hs = tgt_obd_handlers
4857 .tos_opc_start = LDLM_FIRST_OPC,
4858 .tos_opc_end = LDLM_LAST_OPC,
4859 .tos_hs = tgt_dlm_handlers
4862 .tos_opc_start = SEC_FIRST_OPC,
4863 .tos_opc_end = SEC_LAST_OPC,
4864 .tos_hs = mdt_sec_ctx_ops
4867 .tos_opc_start = OUT_UPDATE_FIRST_OPC,
4868 .tos_opc_end = OUT_UPDATE_LAST_OPC,
4869 .tos_hs = tgt_out_handlers
4872 .tos_opc_start = FLD_FIRST_OPC,
4873 .tos_opc_end = FLD_LAST_OPC,
4874 .tos_hs = fld_handlers
4877 .tos_opc_start = SEQ_FIRST_OPC,
4878 .tos_opc_end = SEQ_LAST_OPC,
4879 .tos_hs = seq_handlers
4882 .tos_opc_start = QUOTA_DQACQ,
4883 .tos_opc_end = QUOTA_LAST_OPC,
4884 .tos_hs = mdt_quota_ops
4887 .tos_opc_start = LLOG_FIRST_OPC,
4888 .tos_opc_end = LLOG_LAST_OPC,
4889 .tos_hs = tgt_llog_handlers
4892 .tos_opc_start = LFSCK_FIRST_OPC,
4893 .tos_opc_end = LFSCK_LAST_OPC,
4894 .tos_hs = tgt_lfsck_handlers
4897 .tos_opc_start = OST_FIRST_OPC,
4898 .tos_opc_end = OST_LAST_OPC,
4899 .tos_hs = mdt_io_ops
4906 static void mdt_fini(const struct lu_env *env, struct mdt_device *m)
4908 struct md_device *next = m->mdt_child;
4909 struct lu_device *d = &m->mdt_lu_dev;
4910 struct obd_device *obd = mdt2obd_dev(m);
4911 struct lfsck_stop stop;
4914 stop.ls_status = LS_PAUSED;
4916 next->md_ops->mdo_iocontrol(env, next, OBD_IOC_STOP_LFSCK, 0, &stop);
4918 mdt_stack_pre_fini(env, m, md2lu_dev(m->mdt_child));
4919 ping_evictor_stop();
4921 /* Remove the HSM /proc entry so the coordinator cannot be
4922 * restarted by a user while it's shutting down. */
4923 hsm_cdt_procfs_fini(m);
4924 mdt_hsm_cdt_stop(m);
4926 mdt_llog_ctxt_unclone(env, m, LLOG_AGENT_ORIG_CTXT);
4927 mdt_llog_ctxt_unclone(env, m, LLOG_CHANGELOG_ORIG_CTXT);
4929 if (m->mdt_namespace != NULL)
4930 ldlm_namespace_free_prior(m->mdt_namespace, NULL,
4931 d->ld_obd->obd_force);
4933 obd_exports_barrier(obd);
4934 obd_zombie_barrier();
4936 mdt_quota_fini(env, m);
4938 cfs_free_nidlist(&m->mdt_squash.rsi_nosquash_nids);
4940 /* Calling the cleanup functions in the same order as in the mdt_init0
4945 target_recovery_fini(obd);
4946 upcall_cache_cleanup(m->mdt_identity_cache);
4947 m->mdt_identity_cache = NULL;
4949 mdt_fs_cleanup(env, m);
4951 tgt_fini(env, &m->mdt_lut);
4953 mdt_hsm_cdt_fini(m);
4955 if (m->mdt_los != NULL) {
4956 local_oid_storage_fini(env, m->mdt_los);
4960 if (m->mdt_namespace != NULL) {
4961 ldlm_namespace_free_post(m->mdt_namespace);
4962 d->ld_obd->obd_namespace = m->mdt_namespace = NULL;
4965 if (m->mdt_md_root != NULL) {
4966 mdt_object_put(env, m->mdt_md_root);
4967 m->mdt_md_root = NULL;
4970 mdt_seq_fini(env, m);
4972 mdt_fld_fini(env, m);
4977 mdt_stack_fini(env, m, md2lu_dev(m->mdt_child));
4979 LASSERT(atomic_read(&d->ld_ref) == 0);
4981 server_put_mount(mdt_obd_name(m), true);
4986 static int mdt_postrecov(const struct lu_env *, struct mdt_device *);
4988 static int mdt_init0(const struct lu_env *env, struct mdt_device *m,
4989 struct lu_device_type *ldt, struct lustre_cfg *cfg)
4991 const struct dt_device_param *dt_conf;
4992 struct mdt_thread_info *info;
4993 struct obd_device *obd;
4994 const char *dev = lustre_cfg_string(cfg, 0);
4995 const char *num = lustre_cfg_string(cfg, 2);
4996 struct tg_grants_data *tgd = &m->mdt_lut.lut_tgd;
4997 struct lustre_mount_info *lmi = NULL;
4998 struct lustre_sb_info *lsi;
5000 struct seq_server_site *ss_site;
5001 const char *identity_upcall = "NONE";
5002 struct md_device *next;
5009 lu_device_init(&m->mdt_lu_dev, ldt);
5011 * Environment (env) might be missing mdt_thread_key values at that
5012 * point, if device is allocated when mdt_thread_key is in QUIESCENT
5015 * Usually device allocation path doesn't use module key values, but
5016 * mdt has to do a lot of work here, so allocate key value.
5018 rc = lu_env_refill((struct lu_env *)env);
5022 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
5023 LASSERT(info != NULL);
5025 obd = class_name2obd(dev);
5026 LASSERT(obd != NULL);
5028 m->mdt_max_mdsize = MAX_MD_SIZE; /* 4 stripes */
5029 m->mdt_opts.mo_evict_tgt_nids = 1;
5030 m->mdt_opts.mo_cos = MDT_COS_DEFAULT;
5032 lmi = server_get_mount(dev);
5034 CERROR("Cannot get mount info for %s!\n", dev);
5037 lsi = s2lsi(lmi->lmi_sb);
5038 /* CMD is supported only in IAM mode */
5040 node_id = simple_strtol(num, NULL, 10);
5041 obd->u.obt.obt_magic = OBT_MAGIC;
5042 if (lsi->lsi_lmd != NULL &&
5043 lsi->lsi_lmd->lmd_flags & LMD_FLG_SKIP_LFSCK)
5044 m->mdt_skip_lfsck = 1;
5047 /* DoM files get IO lock at open by default */
5048 m->mdt_opts.mo_dom_lock = ALWAYS_DOM_LOCK_ON_OPEN;
5049 /* DoM files are read at open and data is packed in the reply */
5050 m->mdt_opts.mo_dom_read_open = 1;
5052 m->mdt_squash.rsi_uid = 0;
5053 m->mdt_squash.rsi_gid = 0;
5054 INIT_LIST_HEAD(&m->mdt_squash.rsi_nosquash_nids);
5055 init_rwsem(&m->mdt_squash.rsi_sem);
5056 spin_lock_init(&m->mdt_lock);
5057 m->mdt_enable_remote_dir = 0;
5058 m->mdt_enable_remote_dir_gid = 0;
5060 atomic_set(&m->mdt_mds_mds_conns, 0);
5061 atomic_set(&m->mdt_async_commit_count, 0);
5063 m->mdt_lu_dev.ld_ops = &mdt_lu_ops;
5064 m->mdt_lu_dev.ld_obd = obd;
5065 /* Set this lu_device to obd for error handling purposes. */
5066 obd->obd_lu_dev = &m->mdt_lu_dev;
5068 /* init the stack */
5069 rc = mdt_stack_init((struct lu_env *)env, m, cfg);
5071 CERROR("%s: Can't init device stack, rc %d\n",
5072 mdt_obd_name(m), rc);
5077 ss_site = mdt_seq_site(m);
5078 s->ld_seq_site = ss_site;
5081 /* set server index */
5082 ss_site->ss_node_id = node_id;
5084 /* failover is the default
5085 * FIXME: we do not failout mds0/mgs, which may cause some problems.
5086 * assumed whose ss_node_id == 0 XXX
5088 obd->obd_replayable = 1;
5089 /* No connection accepted until configurations will finish */
5090 obd->obd_no_conn = 1;
5092 if (cfg->lcfg_bufcount > 4 && LUSTRE_CFG_BUFLEN(cfg, 4) > 0) {
5093 char *str = lustre_cfg_string(cfg, 4);
5094 if (strchr(str, 'n')) {
5095 CWARN("%s: recovery disabled\n", mdt_obd_name(m));
5096 obd->obd_replayable = 0;
5100 rc = mdt_fld_init(env, mdt_obd_name(m), m);
5102 GOTO(err_fini_stack, rc);
5104 rc = mdt_seq_init(env, m);
5106 GOTO(err_fini_fld, rc);
5108 snprintf(info->mti_u.ns_name, sizeof(info->mti_u.ns_name), "%s-%s",
5109 LUSTRE_MDT_NAME, obd->obd_uuid.uuid);
5110 m->mdt_namespace = ldlm_namespace_new(obd, info->mti_u.ns_name,
5111 LDLM_NAMESPACE_SERVER,
5112 LDLM_NAMESPACE_GREEDY,
5114 if (m->mdt_namespace == NULL)
5115 GOTO(err_fini_seq, rc = -ENOMEM);
5117 m->mdt_namespace->ns_lvbp = m;
5118 m->mdt_namespace->ns_lvbo = &mdt_lvbo;
5120 ldlm_register_intent(m->mdt_namespace, mdt_intent_policy);
5121 /* set obd_namespace for compatibility with old code */
5122 obd->obd_namespace = m->mdt_namespace;
5124 rc = tgt_init(env, &m->mdt_lut, obd, m->mdt_bottom, mdt_common_slice,
5125 OBD_FAIL_MDS_ALL_REQUEST_NET,
5126 OBD_FAIL_MDS_ALL_REPLY_NET);
5128 GOTO(err_free_ns, rc);
5130 /* Amount of available space excluded from granting and reserved
5131 * for metadata. It is in percentage and 50% is default value. */
5132 tgd->tgd_reserved_pcnt = 50;
5134 if (ONE_MB_BRW_SIZE < (1U << tgd->tgd_blockbits))
5135 m->mdt_brw_size = 1U << tgd->tgd_blockbits;
5137 m->mdt_brw_size = ONE_MB_BRW_SIZE;
5139 rc = mdt_fs_setup(env, m, obd, lsi);
5143 fid.f_seq = FID_SEQ_LOCAL_NAME;
5146 rc = local_oid_storage_init(env, m->mdt_bottom, &fid, &m->mdt_los);
5148 GOTO(err_fs_cleanup, rc);
5150 rc = mdt_hsm_cdt_init(m);
5152 CERROR("%s: error initializing coordinator, rc %d\n",
5153 mdt_obd_name(m), rc);
5154 GOTO(err_los_fini, rc);
5157 tgt_adapt_sptlrpc_conf(&m->mdt_lut);
5159 next = m->mdt_child;
5160 dt_conf = next->md_ops->mdo_dtconf_get(env, next);
5162 mntopts = dt_conf->ddp_mntopts;
5164 if (mntopts & MNTOPT_USERXATTR)
5165 m->mdt_opts.mo_user_xattr = 1;
5167 m->mdt_opts.mo_user_xattr = 0;
5169 m->mdt_max_ea_size = dt_conf->ddp_max_ea_size;
5171 if (mntopts & MNTOPT_ACL)
5172 m->mdt_opts.mo_acl = 1;
5174 m->mdt_opts.mo_acl = 0;
5176 /* XXX: to support suppgid for ACL, we enable identity_upcall
5177 * by default, otherwise, maybe got unexpected -EACCESS. */
5178 if (m->mdt_opts.mo_acl)
5179 identity_upcall = MDT_IDENTITY_UPCALL_PATH;
5181 m->mdt_identity_cache = upcall_cache_init(mdt_obd_name(m),
5183 &mdt_identity_upcall_cache_ops);
5184 if (IS_ERR(m->mdt_identity_cache)) {
5185 rc = PTR_ERR(m->mdt_identity_cache);
5186 m->mdt_identity_cache = NULL;
5187 GOTO(err_free_hsm, rc);
5190 rc = mdt_procfs_init(m, dev);
5192 CERROR("Can't init MDT lprocfs, rc %d\n", rc);
5193 GOTO(err_recovery, rc);
5196 rc = mdt_quota_init(env, m, cfg);
5198 GOTO(err_procfs, rc);
5200 m->mdt_ldlm_client = &mdt2obd_dev(m)->obd_ldlm_client;
5201 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
5202 "mdt_ldlm_client", m->mdt_ldlm_client);
5204 ping_evictor_start();
5206 /* recovery will be started upon mdt_prepare()
5207 * when the whole stack is complete and ready
5208 * to serve the requests */
5210 /* Reduce the initial timeout on an MDS because it doesn't need such
5211 * a long timeout as an OST does. Adaptive timeouts will adjust this
5212 * value appropriately. */
5213 if (ldlm_timeout == LDLM_TIMEOUT_DEFAULT)
5214 ldlm_timeout = MDS_LDLM_TIMEOUT_DEFAULT;
5220 target_recovery_fini(obd);
5221 upcall_cache_cleanup(m->mdt_identity_cache);
5222 m->mdt_identity_cache = NULL;
5224 mdt_hsm_cdt_fini(m);
5226 local_oid_storage_fini(env, m->mdt_los);
5229 mdt_fs_cleanup(env, m);
5231 tgt_fini(env, &m->mdt_lut);
5233 ldlm_namespace_free(m->mdt_namespace, NULL, 0);
5234 obd->obd_namespace = m->mdt_namespace = NULL;
5236 mdt_seq_fini(env, m);
5238 mdt_fld_fini(env, m);
5240 mdt_stack_fini(env, m, md2lu_dev(m->mdt_child));
5243 server_put_mount(dev, true);
5247 /* For interoperability, the left element is old parameter, the right one
5248 * is the new version of the parameter, if some parameter is deprecated,
5249 * the new version should be set as NULL. */
5250 static struct cfg_interop_param mdt_interop_param[] = {
5251 { "mdt.group_upcall", NULL },
5252 { "mdt.quota_type", NULL },
5253 { "mdd.quota_type", NULL },
5254 { "mdt.som", NULL },
5255 { "mdt.rootsquash", "mdt.root_squash" },
5256 { "mdt.nosquash_nid", "mdt.nosquash_nids" },
5260 /* used by MGS to process specific configurations */
5261 static int mdt_process_config(const struct lu_env *env,
5262 struct lu_device *d, struct lustre_cfg *cfg)
5264 struct mdt_device *m = mdt_dev(d);
5265 struct md_device *md_next = m->mdt_child;
5266 struct lu_device *next = md2lu_dev(md_next);
5270 switch (cfg->lcfg_command) {
5272 struct obd_device *obd = d->ld_obd;
5274 /* For interoperability */
5275 struct cfg_interop_param *ptr = NULL;
5276 struct lustre_cfg *old_cfg = NULL;
5279 param = lustre_cfg_string(cfg, 1);
5280 if (param == NULL) {
5281 CERROR("param is empty\n");
5286 ptr = class_find_old_param(param, mdt_interop_param);
5288 if (ptr->new_param == NULL) {
5290 CWARN("For interoperability, skip this %s."
5291 " It is obsolete.\n", ptr->old_param);
5295 CWARN("Found old param %s, changed it to %s.\n",
5296 ptr->old_param, ptr->new_param);
5299 cfg = lustre_cfg_rename(old_cfg, ptr->new_param);
5306 rc = class_process_proc_param(PARAM_MDT, obd->obd_vars,
5308 if (rc > 0 || rc == -ENOSYS) {
5309 /* is it an HSM var ? */
5310 rc = class_process_proc_param(PARAM_HSM,
5311 hsm_cdt_get_proc_vars(),
5313 if (rc > 0 || rc == -ENOSYS)
5314 /* we don't understand; pass it on */
5315 rc = next->ld_ops->ldo_process_config(env, next,
5320 OBD_FREE(cfg, lustre_cfg_len(cfg->lcfg_bufcount,
5321 cfg->lcfg_buflens));
5325 /* others are passed further */
5326 rc = next->ld_ops->ldo_process_config(env, next, cfg);
5332 static struct lu_object *mdt_object_alloc(const struct lu_env *env,
5333 const struct lu_object_header *hdr,
5334 struct lu_device *d)
5336 struct mdt_object *mo;
5340 OBD_SLAB_ALLOC_PTR_GFP(mo, mdt_object_kmem, GFP_NOFS);
5342 struct lu_object *o;
5343 struct lu_object_header *h;
5346 h = &mo->mot_header;
5347 lu_object_header_init(h);
5348 lu_object_init(o, h, d);
5349 lu_object_add_top(h, o);
5350 o->lo_ops = &mdt_obj_ops;
5351 spin_lock_init(&mo->mot_write_lock);
5352 mutex_init(&mo->mot_som_mutex);
5353 mutex_init(&mo->mot_lov_mutex);
5354 init_rwsem(&mo->mot_dom_sem);
5355 init_rwsem(&mo->mot_open_sem);
5356 atomic_set(&mo->mot_open_count, 0);
5362 static int mdt_object_init(const struct lu_env *env, struct lu_object *o,
5363 const struct lu_object_conf *unused)
5365 struct mdt_device *d = mdt_dev(o->lo_dev);
5366 struct lu_device *under;
5367 struct lu_object *below;
5371 CDEBUG(D_INFO, "object init, fid = "DFID"\n",
5372 PFID(lu_object_fid(o)));
5374 under = &d->mdt_child->md_lu_dev;
5375 below = under->ld_ops->ldo_object_alloc(env, o->lo_header, under);
5376 if (below != NULL) {
5377 lu_object_add(o, below);
5384 static void mdt_object_free(const struct lu_env *env, struct lu_object *o)
5386 struct mdt_object *mo = mdt_obj(o);
5387 struct lu_object_header *h;
5391 CDEBUG(D_INFO, "object free, fid = "DFID"\n",
5392 PFID(lu_object_fid(o)));
5394 LASSERT(atomic_read(&mo->mot_open_count) == 0);
5395 LASSERT(atomic_read(&mo->mot_lease_count) == 0);
5398 lu_object_header_fini(h);
5399 OBD_SLAB_FREE_PTR(mo, mdt_object_kmem);
5404 static int mdt_object_print(const struct lu_env *env, void *cookie,
5405 lu_printer_t p, const struct lu_object *o)
5407 struct mdt_object *mdto = mdt_obj((struct lu_object *)o);
5409 return (*p)(env, cookie,
5410 LUSTRE_MDT_NAME"-object@%p(%s %s, writecount=%d)",
5411 mdto, mdto->mot_lov_created ? "lov_created" : "",
5412 mdto->mot_cache_attr ? "cache_attr" : "",
5413 mdto->mot_write_count);
5416 static int mdt_prepare(const struct lu_env *env,
5417 struct lu_device *pdev,
5418 struct lu_device *cdev)
5420 struct mdt_device *mdt = mdt_dev(cdev);
5421 struct lu_device *next = &mdt->mdt_child->md_lu_dev;
5422 struct obd_device *obd = cdev->ld_obd;
5429 rc = next->ld_ops->ldo_prepare(env, cdev, next);
5433 rc = mdt_llog_ctxt_clone(env, mdt, LLOG_CHANGELOG_ORIG_CTXT);
5437 rc = mdt_llog_ctxt_clone(env, mdt, LLOG_AGENT_ORIG_CTXT);
5441 rc = lfsck_register_namespace(env, mdt->mdt_bottom, mdt->mdt_namespace);
5442 /* The LFSCK instance is registered just now, so it must be there when
5443 * register the namespace to such instance. */
5444 LASSERTF(rc == 0, "register namespace failed: rc = %d\n", rc);
5446 if (mdt->mdt_seq_site.ss_node_id == 0) {
5447 rc = mdt->mdt_child->md_ops->mdo_root_get(env, mdt->mdt_child,
5448 &mdt->mdt_md_root_fid);
5453 LASSERT(!test_bit(MDT_FL_CFGLOG, &mdt->mdt_state));
5455 target_recovery_init(&mdt->mdt_lut, tgt_request_handle);
5456 set_bit(MDT_FL_CFGLOG, &mdt->mdt_state);
5457 LASSERT(obd->obd_no_conn);
5458 spin_lock(&obd->obd_dev_lock);
5459 obd->obd_no_conn = 0;
5460 spin_unlock(&obd->obd_dev_lock);
5462 if (obd->obd_recovering == 0)
5463 mdt_postrecov(env, mdt);
5468 const struct lu_device_operations mdt_lu_ops = {
5469 .ldo_object_alloc = mdt_object_alloc,
5470 .ldo_process_config = mdt_process_config,
5471 .ldo_prepare = mdt_prepare,
5474 static const struct lu_object_operations mdt_obj_ops = {
5475 .loo_object_init = mdt_object_init,
5476 .loo_object_free = mdt_object_free,
5477 .loo_object_print = mdt_object_print
5480 static int mdt_obd_set_info_async(const struct lu_env *env,
5481 struct obd_export *exp,
5482 __u32 keylen, void *key,
5483 __u32 vallen, void *val,
5484 struct ptlrpc_request_set *set)
5490 if (KEY_IS(KEY_SPTLRPC_CONF)) {
5491 rc = tgt_adapt_sptlrpc_conf(class_exp2tgt(exp));
5499 * Match client and server connection feature flags.
5501 * Compute the compatibility flags for a connection request based on
5502 * features mutually supported by client and server.
5504 * The obd_export::exp_connect_data.ocd_connect_flags field in \a exp
5505 * must not be updated here, otherwise a partially initialized value may
5506 * be exposed. After the connection request is successfully processed,
5507 * the top-level MDT connect request handler atomically updates the export
5508 * connect flags from the obd_connect_data::ocd_connect_flags field of the
5509 * reply. \see mdt_connect().
5511 * Before 2.7.50 clients will send a struct obd_connect_data_v1 rather than a
5512 * full struct obd_connect_data. So care must be taken when accessing fields
5513 * that are not present in struct obd_connect_data_v1. See LU-16.
5515 * \param exp the obd_export associated with this client/target pair
5516 * \param mdt the target device for the connection
5517 * \param data stores data for this connect request
5520 * \retval -EPROTO \a data unexpectedly has zero obd_connect_data::ocd_brw_size
5521 * \retval -EBADE client and server feature requirements are incompatible
5523 static int mdt_connect_internal(const struct lu_env *env,
5524 struct obd_export *exp,
5525 struct mdt_device *mdt,
5526 struct obd_connect_data *data, bool reconnect)
5528 const char *obd_name = mdt_obd_name(mdt);
5529 LASSERT(data != NULL);
5531 data->ocd_connect_flags &= MDT_CONNECT_SUPPORTED;
5533 if (mdt->mdt_bottom->dd_rdonly &&
5534 !(data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
5535 !(data->ocd_connect_flags & OBD_CONNECT_RDONLY))
5538 if (data->ocd_connect_flags & OBD_CONNECT_FLAGS2)
5539 data->ocd_connect_flags2 &= MDT_CONNECT_SUPPORTED2;
5541 data->ocd_ibits_known &= MDS_INODELOCK_FULL;
5543 if (!mdt->mdt_opts.mo_acl)
5544 data->ocd_connect_flags &= ~OBD_CONNECT_ACL;
5546 if (!mdt->mdt_opts.mo_user_xattr)
5547 data->ocd_connect_flags &= ~OBD_CONNECT_XATTR;
5549 if (OCD_HAS_FLAG(data, BRW_SIZE)) {
5550 data->ocd_brw_size = min(data->ocd_brw_size,
5552 if (data->ocd_brw_size == 0) {
5553 CERROR("%s: cli %s/%p ocd_connect_flags: %#llx "
5554 "ocd_version: %x ocd_grant: %d ocd_index: %u "
5555 "ocd_brw_size unexpectedly zero, network data "
5556 "corruption? Refusing to connect this client\n",
5557 obd_name, exp->exp_client_uuid.uuid,
5558 exp, data->ocd_connect_flags, data->ocd_version,
5559 data->ocd_grant, data->ocd_index);
5564 if (OCD_HAS_FLAG(data, GRANT_PARAM)) {
5565 struct dt_device_param *ddp = &mdt->mdt_lut.lut_dt_conf;
5567 /* client is reporting its page size, for future use */
5568 exp->exp_target_data.ted_pagebits = data->ocd_grant_blkbits;
5569 data->ocd_grant_blkbits = mdt->mdt_lut.lut_tgd.tgd_blockbits;
5570 /* ddp_inodespace may not be power-of-two value, eg. for ldiskfs
5571 * it's LDISKFS_DIR_REC_LEN(20) = 28. */
5572 data->ocd_grant_inobits = fls(ddp->ddp_inodespace - 1);
5573 /* ocd_grant_tax_kb is in 1K byte blocks */
5574 data->ocd_grant_tax_kb = ddp->ddp_extent_tax >> 10;
5575 data->ocd_grant_max_blks = ddp->ddp_max_extent_blks;
5578 /* Save connect_data we have so far because tgt_grant_connect()
5579 * uses it to calculate grant, and we want to save the client
5580 * version before it is overwritten by LUSTRE_VERSION_CODE. */
5581 exp->exp_connect_data = *data;
5582 if (OCD_HAS_FLAG(data, GRANT))
5583 tgt_grant_connect(env, exp, data, !reconnect);
5585 if (OCD_HAS_FLAG(data, MAXBYTES))
5586 data->ocd_maxbytes = mdt->mdt_lut.lut_dt_conf.ddp_maxbytes;
5588 /* NB: Disregard the rule against updating
5589 * exp_connect_data.ocd_connect_flags in this case, since
5590 * tgt_client_new() needs to know if this is a lightweight
5591 * connection, and it is safe to expose this flag before
5592 * connection processing completes. */
5593 if (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) {
5594 spin_lock(&exp->exp_lock);
5595 *exp_connect_flags_ptr(exp) |= OBD_CONNECT_LIGHTWEIGHT;
5596 spin_unlock(&exp->exp_lock);
5599 data->ocd_version = LUSTRE_VERSION_CODE;
5601 if ((data->ocd_connect_flags & OBD_CONNECT_FID) == 0) {
5602 CWARN("%s: MDS requires FID support, but client not\n",
5607 if (OCD_HAS_FLAG(data, PINGLESS)) {
5608 if (ptlrpc_pinger_suppress_pings()) {
5609 spin_lock(&exp->exp_obd->obd_dev_lock);
5610 list_del_init(&exp->exp_obd_chain_timed);
5611 spin_unlock(&exp->exp_obd->obd_dev_lock);
5613 data->ocd_connect_flags &= ~OBD_CONNECT_PINGLESS;
5617 data->ocd_max_easize = mdt->mdt_max_ea_size;
5619 /* NB: Disregard the rule against updating
5620 * exp_connect_data.ocd_connect_flags in this case, since
5621 * tgt_client_new() needs to know if this is client supports
5622 * multiple modify RPCs, and it is safe to expose this flag before
5623 * connection processing completes. */
5624 if (data->ocd_connect_flags & OBD_CONNECT_MULTIMODRPCS) {
5625 data->ocd_maxmodrpcs = max_mod_rpcs_per_client;
5626 spin_lock(&exp->exp_lock);
5627 *exp_connect_flags_ptr(exp) |= OBD_CONNECT_MULTIMODRPCS;
5628 spin_unlock(&exp->exp_lock);
5631 if (OCD_HAS_FLAG(data, CKSUM)) {
5632 __u32 cksum_types = data->ocd_cksum_types;
5634 /* The client set in ocd_cksum_types the checksum types it
5635 * supports. We have to mask off the algorithms that we don't
5637 data->ocd_cksum_types &=
5638 obd_cksum_types_supported_server(obd_name);
5640 if (unlikely(data->ocd_cksum_types == 0)) {
5641 CERROR("%s: Connect with checksum support but no "
5642 "ocd_cksum_types is set\n",
5643 exp->exp_obd->obd_name);
5647 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
5648 "%x\n", exp->exp_obd->obd_name, obd_export_nid2str(exp),
5649 cksum_types, data->ocd_cksum_types);
5651 /* This client does not support OBD_CONNECT_CKSUM
5652 * fall back to CRC32 */
5653 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
5654 "OBD_CONNECT_CKSUM, CRC32 will be used\n",
5655 exp->exp_obd->obd_name, obd_export_nid2str(exp));
5661 static int mdt_ctxt_add_dirty_flag(struct lu_env *env,
5662 struct mdt_thread_info *info,
5663 struct mdt_file_data *mfd)
5665 struct lu_context ses;
5669 rc = lu_context_init(&ses, LCT_SERVER_SESSION);
5674 lu_context_enter(&ses);
5676 mdt_ucred(info)->uc_valid = UCRED_OLD;
5677 rc = mdt_add_dirty_flag(info, mfd->mfd_object, &info->mti_attr);
5679 lu_context_exit(&ses);
5680 lu_context_fini(&ses);
5686 static int mdt_export_cleanup(struct obd_export *exp)
5688 struct list_head closing_list;
5689 struct mdt_export_data *med = &exp->exp_mdt_data;
5690 struct obd_device *obd = exp->exp_obd;
5691 struct mdt_device *mdt;
5692 struct mdt_thread_info *info;
5694 struct mdt_file_data *mfd, *n;
5698 INIT_LIST_HEAD(&closing_list);
5699 spin_lock(&med->med_open_lock);
5700 while (!list_empty(&med->med_open_head)) {
5701 struct list_head *tmp = med->med_open_head.next;
5702 mfd = list_entry(tmp, struct mdt_file_data, mfd_list);
5704 /* Remove mfd handle so it can't be found again.
5705 * We are consuming the mfd_list reference here. */
5706 class_handle_unhash(&mfd->mfd_handle);
5707 list_move_tail(&mfd->mfd_list, &closing_list);
5709 spin_unlock(&med->med_open_lock);
5710 mdt = mdt_dev(obd->obd_lu_dev);
5711 LASSERT(mdt != NULL);
5713 rc = lu_env_init(&env, LCT_MD_THREAD);
5717 info = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
5718 LASSERT(info != NULL);
5719 memset(info, 0, sizeof *info);
5720 info->mti_env = &env;
5721 info->mti_mdt = mdt;
5722 info->mti_exp = exp;
5724 if (!list_empty(&closing_list)) {
5725 struct md_attr *ma = &info->mti_attr;
5727 /* Close any open files (which may also cause orphan
5729 list_for_each_entry_safe(mfd, n, &closing_list, mfd_list) {
5730 list_del_init(&mfd->mfd_list);
5731 ma->ma_need = ma->ma_valid = 0;
5733 /* This file is being closed due to an eviction, it
5734 * could have been modified and now dirty regarding to
5735 * HSM archive, check this!
5736 * The logic here is to mark a file dirty if there's a
5737 * chance it was dirtied before the client was evicted,
5738 * so that we don't have to wait for a release attempt
5739 * before finding out the file was actually dirty and
5740 * fail the release. Aggressively marking it dirty here
5741 * will cause the policy engine to attempt to
5742 * re-archive it; when rearchiving, we can compare the
5743 * current version to the HSM data_version and make the
5744 * archive request into a noop if it's not actually
5747 if (mfd->mfd_mode & MDS_FMODE_WRITE)
5748 rc = mdt_ctxt_add_dirty_flag(&env, info, mfd);
5750 /* Don't unlink orphan on failover umount, LU-184 */
5751 if (exp->exp_flags & OBD_OPT_FAILOVER) {
5752 ma->ma_valid = MA_FLAGS;
5753 ma->ma_attr_flags |= MDS_KEEP_ORPHAN;
5755 mdt_mfd_close(info, mfd);
5758 info->mti_mdt = NULL;
5759 /* cleanup client slot early */
5760 /* Do not erase record for recoverable client. */
5761 if (!(exp->exp_flags & OBD_OPT_FAILOVER) || exp->exp_failed)
5762 tgt_client_del(&env, exp);
5768 static inline void mdt_enable_slc(struct mdt_device *mdt)
5770 if (mdt->mdt_lut.lut_sync_lock_cancel == NEVER_SYNC_ON_CANCEL)
5771 mdt->mdt_lut.lut_sync_lock_cancel = BLOCKING_SYNC_ON_CANCEL;
5774 static inline void mdt_disable_slc(struct mdt_device *mdt)
5776 if (mdt->mdt_lut.lut_sync_lock_cancel == BLOCKING_SYNC_ON_CANCEL)
5777 mdt->mdt_lut.lut_sync_lock_cancel = NEVER_SYNC_ON_CANCEL;
5780 static int mdt_obd_disconnect(struct obd_export *exp)
5787 class_export_get(exp);
5789 if (!(exp->exp_flags & OBD_OPT_FORCE))
5790 tgt_grant_sanity_check(exp->exp_obd, __func__);
5792 if ((exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) &&
5793 !(exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)) {
5794 struct mdt_device *mdt = mdt_dev(exp->exp_obd->obd_lu_dev);
5796 if (atomic_dec_and_test(&mdt->mdt_mds_mds_conns))
5797 mdt_disable_slc(mdt);
5800 rc = server_disconnect_export(exp);
5802 CDEBUG(D_IOCTL, "server disconnect error: rc = %d\n", rc);
5804 tgt_grant_discard(exp);
5806 rc = mdt_export_cleanup(exp);
5807 nodemap_del_member(exp);
5808 class_export_put(exp);
5812 /* mds_connect copy */
5813 static int mdt_obd_connect(const struct lu_env *env,
5814 struct obd_export **exp, struct obd_device *obd,
5815 struct obd_uuid *cluuid,
5816 struct obd_connect_data *data,
5819 struct obd_export *lexp;
5820 struct lustre_handle conn = { 0 };
5821 struct mdt_device *mdt;
5823 lnet_nid_t *client_nid = localdata;
5826 LASSERT(env != NULL);
5827 LASSERT(data != NULL);
5829 if (!exp || !obd || !cluuid)
5832 mdt = mdt_dev(obd->obd_lu_dev);
5834 if ((data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
5835 !(data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT)) {
5836 atomic_inc(&mdt->mdt_mds_mds_conns);
5837 mdt_enable_slc(mdt);
5841 * first, check whether the stack is ready to handle requests
5842 * XXX: probably not very appropriate method is used now
5843 * at some point we should find a better one
5845 if (!test_bit(MDT_FL_SYNCED, &mdt->mdt_state) &&
5846 !(data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) &&
5847 !(data->ocd_connect_flags & OBD_CONNECT_MDS_MDS)) {
5848 rc = obd_get_info(env, mdt->mdt_child_exp,
5849 sizeof(KEY_OSP_CONNECTED),
5850 KEY_OSP_CONNECTED, NULL, NULL);
5853 set_bit(MDT_FL_SYNCED, &mdt->mdt_state);
5856 rc = class_connect(&conn, obd, cluuid);
5860 lexp = class_conn2export(&conn);
5861 LASSERT(lexp != NULL);
5863 rc = nodemap_add_member(*client_nid, lexp);
5864 if (rc != 0 && rc != -EEXIST)
5867 rc = mdt_connect_internal(env, lexp, mdt, data, false);
5869 struct lsd_client_data *lcd = lexp->exp_target_data.ted_lcd;
5872 memcpy(lcd->lcd_uuid, cluuid, sizeof lcd->lcd_uuid);
5873 rc = tgt_client_new(env, lexp);
5875 mdt_export_stats_init(obd, lexp, localdata);
5879 class_disconnect(lexp);
5880 nodemap_del_member(lexp);
5884 /* Because we do not want this export to be evicted by pinger,
5885 * let's not add this export to the timed chain list. */
5886 if (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) {
5887 spin_lock(&lexp->exp_obd->obd_dev_lock);
5888 list_del_init(&lexp->exp_obd_chain_timed);
5889 spin_unlock(&lexp->exp_obd->obd_dev_lock);
5896 static int mdt_obd_reconnect(const struct lu_env *env,
5897 struct obd_export *exp, struct obd_device *obd,
5898 struct obd_uuid *cluuid,
5899 struct obd_connect_data *data,
5902 lnet_nid_t *client_nid = localdata;
5906 if (exp == NULL || obd == NULL || cluuid == NULL)
5909 rc = nodemap_add_member(*client_nid, exp);
5910 if (rc != 0 && rc != -EEXIST)
5913 rc = mdt_connect_internal(env, exp, mdt_dev(obd->obd_lu_dev), data,
5916 mdt_export_stats_init(obd, exp, localdata);
5918 nodemap_del_member(exp);
5923 /* FIXME: Can we avoid using these two interfaces? */
5924 static int mdt_init_export(struct obd_export *exp)
5926 struct mdt_export_data *med = &exp->exp_mdt_data;
5930 INIT_LIST_HEAD(&med->med_open_head);
5931 spin_lock_init(&med->med_open_lock);
5932 spin_lock(&exp->exp_lock);
5933 exp->exp_connecting = 1;
5934 spin_unlock(&exp->exp_lock);
5936 /* self-export doesn't need client data and ldlm initialization */
5937 if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
5938 &exp->exp_client_uuid)))
5941 rc = tgt_client_alloc(exp);
5945 rc = ldlm_init_export(exp);
5952 tgt_client_free(exp);
5954 CERROR("%s: Failed to initialize export: rc = %d\n",
5955 exp->exp_obd->obd_name, rc);
5959 static int mdt_destroy_export(struct obd_export *exp)
5963 target_destroy_export(exp);
5964 /* destroy can be called from failed obd_setup, so
5965 * checking uuid is safer than obd_self_export */
5966 if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
5967 &exp->exp_client_uuid)))
5970 ldlm_destroy_export(exp);
5971 tgt_client_free(exp);
5973 LASSERT(list_empty(&exp->exp_outstanding_replies));
5974 LASSERT(list_empty(&exp->exp_mdt_data.med_open_head));
5977 * discard grants once we're sure no more
5978 * interaction with the client is possible
5980 tgt_grant_discard(exp);
5981 if (exp_connect_flags(exp) & OBD_CONNECT_GRANT)
5982 exp->exp_obd->u.obt.obt_lut->lut_tgd.tgd_tot_granted_clients--;
5984 if (!(exp->exp_flags & OBD_OPT_FORCE))
5985 tgt_grant_sanity_check(exp->exp_obd, __func__);
5990 int mdt_links_read(struct mdt_thread_info *info, struct mdt_object *mdt_obj,
5991 struct linkea_data *ldata)
5995 LASSERT(ldata->ld_buf->lb_buf != NULL);
5997 if (!mdt_object_exists(mdt_obj))
6000 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
6001 ldata->ld_buf, XATTR_NAME_LINK);
6002 if (rc == -ERANGE) {
6003 /* Buf was too small, figure out what we need. */
6004 lu_buf_free(ldata->ld_buf);
6005 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
6006 ldata->ld_buf, XATTR_NAME_LINK);
6009 ldata->ld_buf = lu_buf_check_and_alloc(ldata->ld_buf, rc);
6010 if (ldata->ld_buf->lb_buf == NULL)
6012 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
6013 ldata->ld_buf, XATTR_NAME_LINK);
6018 return linkea_init_with_rec(ldata);
6022 * Given an MDT object, try to look up the full path to the object.
6023 * Part of the MDT layer implementation of lfs fid2path.
6025 * \param[in] info Per-thread common data shared by MDT level handlers.
6026 * \param[in] obj Object to do path lookup of
6027 * \param[in,out] fp User-provided struct to store path information
6028 * \param[in] root_fid Root FID of current path should reach
6030 * \retval 0 Lookup successful, path information stored in fp
6031 * \retval -EAGAIN Lookup failed, usually because object is being moved
6032 * \retval negative errno if there was a problem
6034 static int mdt_path_current(struct mdt_thread_info *info,
6035 struct mdt_object *obj,
6036 struct getinfo_fid2path *fp,
6037 struct lu_fid *root_fid)
6039 struct mdt_device *mdt = info->mti_mdt;
6040 struct mdt_object *mdt_obj;
6041 struct link_ea_header *leh;
6042 struct link_ea_entry *lee;
6043 struct lu_name *tmpname = &info->mti_name;
6044 struct lu_fid *tmpfid = &info->mti_tmp_fid1;
6045 struct lu_buf *buf = &info->mti_big_buf;
6048 struct linkea_data ldata = { NULL };
6053 /* temp buffer for path element, the buffer will be finally freed
6054 * in mdt_thread_info_fini */
6055 buf = lu_buf_check_and_alloc(buf, PATH_MAX);
6056 if (buf->lb_buf == NULL)
6060 ptr = fp->gf_u.gf_path + fp->gf_pathlen - 1;
6063 *tmpfid = fp->gf_fid = *mdt_object_fid(obj);
6065 while (!lu_fid_eq(root_fid, &fp->gf_fid)) {
6066 struct lu_buf lmv_buf;
6068 if (!lu_fid_eq(root_fid, &mdt->mdt_md_root_fid) &&
6069 lu_fid_eq(&mdt->mdt_md_root_fid, &fp->gf_fid))
6070 GOTO(out, rc = -ENOENT);
6072 if (lu_fid_eq(mdt_object_fid(obj), tmpfid)) {
6074 mdt_object_get(info->mti_env, mdt_obj);
6076 mdt_obj = mdt_object_find(info->mti_env, mdt, tmpfid);
6077 if (IS_ERR(mdt_obj))
6078 GOTO(out, rc = PTR_ERR(mdt_obj));
6081 if (!mdt_object_exists(mdt_obj)) {
6082 mdt_object_put(info->mti_env, mdt_obj);
6083 GOTO(out, rc = -ENOENT);
6086 if (mdt_object_remote(mdt_obj)) {
6087 mdt_object_put(info->mti_env, mdt_obj);
6088 GOTO(remote_out, rc = -EREMOTE);
6091 rc = mdt_links_read(info, mdt_obj, &ldata);
6093 mdt_object_put(info->mti_env, mdt_obj);
6098 lee = (struct link_ea_entry *)(leh + 1); /* link #0 */
6099 linkea_entry_unpack(lee, &reclen, tmpname, tmpfid);
6100 /* If set, use link #linkno for path lookup, otherwise use
6101 link #0. Only do this for the final path element. */
6102 if (first && fp->gf_linkno < leh->leh_reccount) {
6104 for (count = 0; count < fp->gf_linkno; count++) {
6105 lee = (struct link_ea_entry *)
6106 ((char *)lee + reclen);
6107 linkea_entry_unpack(lee, &reclen, tmpname,
6110 if (fp->gf_linkno < leh->leh_reccount - 1)
6111 /* indicate to user there are more links */
6115 lmv_buf.lb_buf = info->mti_xattr_buf;
6116 lmv_buf.lb_len = sizeof(info->mti_xattr_buf);
6117 /* Check if it is slave stripes */
6118 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
6119 &lmv_buf, XATTR_NAME_LMV);
6120 mdt_object_put(info->mti_env, mdt_obj);
6122 union lmv_mds_md *lmm = lmv_buf.lb_buf;
6124 /* For slave stripes, get its master */
6125 if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_STRIPE) {
6126 fp->gf_fid = *tmpfid;
6129 } else if (rc < 0 && rc != -ENODATA) {
6135 /* Pack the name in the end of the buffer */
6136 ptr -= tmpname->ln_namelen;
6137 if (ptr - 1 <= fp->gf_u.gf_path)
6138 GOTO(out, rc = -EOVERFLOW);
6139 strncpy(ptr, tmpname->ln_name, tmpname->ln_namelen);
6142 /* keep the last resolved fid to the client, so the
6143 * client will build the left path on another MDT for
6145 fp->gf_fid = *tmpfid;
6151 ptr++; /* skip leading / */
6152 memmove(fp->gf_u.gf_path, ptr,
6153 fp->gf_u.gf_path + fp->gf_pathlen - ptr);
6160 * Given an MDT object, use mdt_path_current to get the path.
6161 * Essentially a wrapper to retry mdt_path_current a set number of times
6162 * if -EAGAIN is returned (usually because an object is being moved).
6164 * Part of the MDT layer implementation of lfs fid2path.
6166 * \param[in] info Per-thread common data shared by mdt level handlers.
6167 * \param[in] obj Object to do path lookup of
6168 * \param[in,out] fp User-provided struct for arguments and to store path
6171 * \retval 0 Lookup successful, path information stored in fp
6172 * \retval negative errno if there was a problem
6174 static int mdt_path(struct mdt_thread_info *info, struct mdt_object *obj,
6175 struct getinfo_fid2path *fp, struct lu_fid *root_fid)
6177 struct mdt_device *mdt = info->mti_mdt;
6182 if (fp->gf_pathlen < 3)
6185 if (root_fid == NULL)
6186 root_fid = &mdt->mdt_md_root_fid;
6188 if (lu_fid_eq(root_fid, mdt_object_fid(obj))) {
6189 fp->gf_u.gf_path[0] = '\0';
6193 /* Retry multiple times in case file is being moved */
6194 while (tries-- && rc == -EAGAIN)
6195 rc = mdt_path_current(info, obj, fp, root_fid);
6201 * Get the full path of the provided FID, as of changelog record recno.
6203 * This checks sanity and looks up object for user provided FID
6204 * before calling the actual path lookup code.
6206 * Part of the MDT layer implementation of lfs fid2path.
6208 * \param[in] info Per-thread common data shared by mdt level handlers.
6209 * \param[in,out] fp User-provided struct for arguments and to store path
6212 * \retval 0 Lookup successful, path information and recno stored in fp
6213 * \retval -ENOENT, object does not exist
6214 * \retval negative errno if there was a problem
6216 static int mdt_fid2path(struct mdt_thread_info *info,
6217 struct lu_fid *root_fid,
6218 struct getinfo_fid2path *fp)
6220 struct mdt_device *mdt = info->mti_mdt;
6221 struct mdt_object *obj;
6225 CDEBUG(D_IOCTL, "path get "DFID" from %llu #%d\n",
6226 PFID(&fp->gf_fid), fp->gf_recno, fp->gf_linkno);
6228 if (!fid_is_sane(&fp->gf_fid))
6231 if (!fid_is_namespace_visible(&fp->gf_fid)) {
6232 CDEBUG(D_INFO, "%s: "DFID" is invalid, f_seq should be >= %#llx"
6233 ", or f_oid != 0, or f_ver == 0\n", mdt_obd_name(mdt),
6234 PFID(&fp->gf_fid), (__u64)FID_SEQ_NORMAL);
6238 obj = mdt_object_find(info->mti_env, mdt, &fp->gf_fid);
6241 CDEBUG(D_IOCTL, "cannot find "DFID": rc = %d\n",
6242 PFID(&fp->gf_fid), rc);
6246 if (mdt_object_remote(obj))
6248 else if (!mdt_object_exists(obj))
6254 mdt_object_put(info->mti_env, obj);
6255 CDEBUG(D_IOCTL, "nonlocal object "DFID": rc = %d\n",
6256 PFID(&fp->gf_fid), rc);
6260 rc = mdt_path(info, obj, fp, root_fid);
6262 CDEBUG(D_INFO, "fid "DFID", path %s recno %#llx linkno %u\n",
6263 PFID(&fp->gf_fid), fp->gf_u.gf_path,
6264 fp->gf_recno, fp->gf_linkno);
6266 mdt_object_put(info->mti_env, obj);
6271 static int mdt_rpc_fid2path(struct mdt_thread_info *info, void *key, int keylen,
6272 void *val, int vallen)
6274 struct getinfo_fid2path *fpout, *fpin;
6275 struct lu_fid *root_fid = NULL;
6278 fpin = key + cfs_size_round(sizeof(KEY_FID2PATH));
6281 if (ptlrpc_req_need_swab(info->mti_pill->rc_req))
6282 lustre_swab_fid2path(fpin);
6284 memcpy(fpout, fpin, sizeof(*fpin));
6285 if (fpout->gf_pathlen != vallen - sizeof(*fpin))
6288 if (keylen >= cfs_size_round(sizeof(KEY_FID2PATH)) + sizeof(*fpin) +
6289 sizeof(struct lu_fid)) {
6290 /* client sent its root FID, which is normally fileset FID */
6291 root_fid = fpin->gf_u.gf_root_fid;
6292 if (ptlrpc_req_need_swab(info->mti_pill->rc_req))
6293 lustre_swab_lu_fid(root_fid);
6295 if (root_fid != NULL && !fid_is_sane(root_fid))
6299 rc = mdt_fid2path(info, root_fid, fpout);
6303 int mdt_get_info(struct tgt_session_info *tsi)
6313 key = req_capsule_client_get(tsi->tsi_pill, &RMF_GETINFO_KEY);
6315 CDEBUG(D_IOCTL, "No GETINFO key\n");
6316 RETURN(err_serious(-EFAULT));
6318 keylen = req_capsule_get_size(tsi->tsi_pill, &RMF_GETINFO_KEY,
6321 vallen = req_capsule_client_get(tsi->tsi_pill, &RMF_GETINFO_VALLEN);
6322 if (vallen == NULL) {
6323 CDEBUG(D_IOCTL, "%s: cannot get RMF_GETINFO_VALLEN buffer\n",
6324 tgt_name(tsi->tsi_tgt));
6325 RETURN(err_serious(-EFAULT));
6328 req_capsule_set_size(tsi->tsi_pill, &RMF_GETINFO_VAL, RCL_SERVER,
6330 rc = req_capsule_server_pack(tsi->tsi_pill);
6332 RETURN(err_serious(rc));
6334 valout = req_capsule_server_get(tsi->tsi_pill, &RMF_GETINFO_VAL);
6335 if (valout == NULL) {
6336 CDEBUG(D_IOCTL, "%s: cannot get get-info RPC out buffer\n",
6337 tgt_name(tsi->tsi_tgt));
6338 RETURN(err_serious(-EFAULT));
6341 if (KEY_IS(KEY_FID2PATH)) {
6342 struct mdt_thread_info *info = tsi2mdt_info(tsi);
6344 rc = mdt_rpc_fid2path(info, key, keylen, valout, *vallen);
6345 mdt_thread_info_fini(info);
6352 static int mdt_ioc_version_get(struct mdt_thread_info *mti, void *karg)
6354 struct obd_ioctl_data *data = karg;
6357 struct mdt_object *obj;
6358 struct mdt_lock_handle *lh;
6362 if (data->ioc_inlbuf1 == NULL || data->ioc_inllen1 != sizeof(*fid) ||
6363 data->ioc_inlbuf2 == NULL || data->ioc_inllen2 != sizeof(version))
6366 fid = (struct lu_fid *)data->ioc_inlbuf1;
6368 if (!fid_is_sane(fid))
6371 CDEBUG(D_IOCTL, "getting version for "DFID"\n", PFID(fid));
6373 lh = &mti->mti_lh[MDT_LH_PARENT];
6374 mdt_lock_reg_init(lh, LCK_CR);
6376 obj = mdt_object_find_lock(mti, fid, lh, MDS_INODELOCK_UPDATE);
6378 RETURN(PTR_ERR(obj));
6380 if (mdt_object_remote(obj)) {
6383 * before calling version get the correct MDS should be
6384 * fid, this is error to find remote object here
6386 CERROR("nonlocal object "DFID"\n", PFID(fid));
6387 } else if (!mdt_object_exists(obj)) {
6388 *(__u64 *)data->ioc_inlbuf2 = ENOENT_VERSION;
6391 version = dt_version_get(mti->mti_env, mdt_obj2dt(obj));
6392 *(__u64 *)data->ioc_inlbuf2 = version;
6395 mdt_object_unlock_put(mti, obj, lh, 1);
6399 /* ioctls on obd dev */
6400 static int mdt_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
6401 void *karg, void __user *uarg)
6404 struct obd_device *obd = exp->exp_obd;
6405 struct mdt_device *mdt = mdt_dev(obd->obd_lu_dev);
6406 struct dt_device *dt = mdt->mdt_bottom;
6410 CDEBUG(D_IOCTL, "handling ioctl cmd %#x\n", cmd);
6411 rc = lu_env_init(&env, LCT_MD_THREAD);
6417 rc = mdt_device_sync(&env, mdt);
6419 case OBD_IOC_SET_READONLY:
6420 rc = dt_sync(&env, dt);
6422 rc = dt_ro(&env, dt);
6424 case OBD_IOC_ABORT_RECOVERY:
6425 CERROR("%s: Aborting recovery for device\n", mdt_obd_name(mdt));
6426 obd->obd_abort_recovery = 1;
6427 target_stop_recovery_thread(obd);
6430 case OBD_IOC_CHANGELOG_REG:
6431 case OBD_IOC_CHANGELOG_DEREG:
6432 case OBD_IOC_CHANGELOG_CLEAR:
6433 rc = mdt->mdt_child->md_ops->mdo_iocontrol(&env,
6437 case OBD_IOC_START_LFSCK: {
6438 struct md_device *next = mdt->mdt_child;
6439 struct obd_ioctl_data *data = karg;
6440 struct lfsck_start_param lsp;
6442 if (unlikely(data == NULL)) {
6447 lsp.lsp_start = (struct lfsck_start *)(data->ioc_inlbuf1);
6448 lsp.lsp_index_valid = 0;
6449 rc = next->md_ops->mdo_iocontrol(&env, next, cmd, 0, &lsp);
6452 case OBD_IOC_STOP_LFSCK: {
6453 struct md_device *next = mdt->mdt_child;
6454 struct obd_ioctl_data *data = karg;
6455 struct lfsck_stop stop;
6457 stop.ls_status = LS_STOPPED;
6458 /* Old lfsck utils may pass NULL @stop. */
6459 if (data->ioc_inlbuf1 == NULL)
6463 ((struct lfsck_stop *)(data->ioc_inlbuf1))->ls_flags;
6465 rc = next->md_ops->mdo_iocontrol(&env, next, cmd, 0, &stop);
6468 case OBD_IOC_QUERY_LFSCK: {
6469 struct md_device *next = mdt->mdt_child;
6470 struct obd_ioctl_data *data = karg;
6472 rc = next->md_ops->mdo_iocontrol(&env, next, cmd, 0,
6476 case OBD_IOC_GET_OBJ_VERSION: {
6477 struct mdt_thread_info *mti;
6478 mti = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
6479 memset(mti, 0, sizeof *mti);
6480 mti->mti_env = &env;
6484 rc = mdt_ioc_version_get(mti, karg);
6487 case OBD_IOC_CATLOGLIST: {
6488 struct mdt_thread_info *mti;
6490 mti = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
6491 lu_local_obj_fid(&mti->mti_tmp_fid1, LLOG_CATALOGS_OID);
6492 rc = llog_catalog_list(&env, mdt->mdt_bottom, 0, karg,
6493 &mti->mti_tmp_fid1);
6498 CERROR("%s: Not supported cmd = %d, rc = %d\n",
6499 mdt_obd_name(mdt), cmd, rc);
6506 static int mdt_postrecov(const struct lu_env *env, struct mdt_device *mdt)
6508 struct lu_device *ld = md2lu_dev(mdt->mdt_child);
6512 if (!mdt->mdt_skip_lfsck && !mdt->mdt_bottom->dd_rdonly) {
6513 struct lfsck_start_param lsp;
6515 lsp.lsp_start = NULL;
6516 lsp.lsp_index_valid = 0;
6517 rc = mdt->mdt_child->md_ops->mdo_iocontrol(env, mdt->mdt_child,
6518 OBD_IOC_START_LFSCK,
6520 if (rc != 0 && rc != -EALREADY)
6521 CWARN("%s: auto trigger paused LFSCK failed: rc = %d\n",
6522 mdt_obd_name(mdt), rc);
6525 rc = ld->ld_ops->ldo_recovery_complete(env, ld);
6529 static int mdt_obd_postrecov(struct obd_device *obd)
6534 rc = lu_env_init(&env, LCT_MD_THREAD);
6537 rc = mdt_postrecov(&env, mdt_dev(obd->obd_lu_dev));
6542 static struct obd_ops mdt_obd_device_ops = {
6543 .o_owner = THIS_MODULE,
6544 .o_set_info_async = mdt_obd_set_info_async,
6545 .o_connect = mdt_obd_connect,
6546 .o_reconnect = mdt_obd_reconnect,
6547 .o_disconnect = mdt_obd_disconnect,
6548 .o_init_export = mdt_init_export,
6549 .o_destroy_export = mdt_destroy_export,
6550 .o_iocontrol = mdt_iocontrol,
6551 .o_postrecov = mdt_obd_postrecov,
6552 /* Data-on-MDT IO methods */
6553 .o_preprw = mdt_obd_preprw,
6554 .o_commitrw = mdt_obd_commitrw,
6557 static struct lu_device* mdt_device_fini(const struct lu_env *env,
6558 struct lu_device *d)
6560 struct mdt_device *m = mdt_dev(d);
6567 static struct lu_device *mdt_device_free(const struct lu_env *env,
6568 struct lu_device *d)
6570 struct mdt_device *m = mdt_dev(d);
6573 lu_device_fini(&m->mdt_lu_dev);
6579 static struct lu_device *mdt_device_alloc(const struct lu_env *env,
6580 struct lu_device_type *t,
6581 struct lustre_cfg *cfg)
6583 struct lu_device *l;
6584 struct mdt_device *m;
6591 rc = mdt_init0(env, m, t, cfg);
6593 mdt_device_free(env, l);
6598 l = ERR_PTR(-ENOMEM);
6602 /* context key constructor/destructor: mdt_key_init, mdt_key_fini */
6603 LU_KEY_INIT(mdt, struct mdt_thread_info);
6605 static void mdt_key_fini(const struct lu_context *ctx,
6606 struct lu_context_key *key, void* data)
6608 struct mdt_thread_info *info = data;
6610 if (info->mti_big_lmm) {
6611 OBD_FREE_LARGE(info->mti_big_lmm, info->mti_big_lmmsize);
6612 info->mti_big_lmm = NULL;
6613 info->mti_big_lmmsize = 0;
6616 if (info->mti_big_acl) {
6617 OBD_FREE_LARGE(info->mti_big_acl, info->mti_big_aclsize);
6618 info->mti_big_acl = NULL;
6619 info->mti_big_aclsize = 0;
6625 /* context key: mdt_thread_key */
6626 LU_CONTEXT_KEY_DEFINE(mdt, LCT_MD_THREAD);
6628 struct lu_ucred *mdt_ucred(const struct mdt_thread_info *info)
6630 return lu_ucred(info->mti_env);
6633 struct lu_ucred *mdt_ucred_check(const struct mdt_thread_info *info)
6635 return lu_ucred_check(info->mti_env);
6639 * Enable/disable COS (Commit On Sharing).
6641 * Set/Clear the COS flag in mdt options.
6643 * \param mdt mdt device
6644 * \param val 0 disables COS, other values enable COS
6646 void mdt_enable_cos(struct mdt_device *mdt, bool val)
6651 mdt->mdt_opts.mo_cos = val;
6652 rc = lu_env_init(&env, LCT_LOCAL);
6653 if (unlikely(rc != 0)) {
6654 CWARN("%s: lu_env initialization failed, cannot "
6655 "sync: rc = %d\n", mdt_obd_name(mdt), rc);
6658 mdt_device_sync(&env, mdt);
6663 * Check COS (Commit On Sharing) status.
6665 * Return COS flag status.
6667 * \param mdt mdt device
6669 int mdt_cos_is_enabled(struct mdt_device *mdt)
6671 return mdt->mdt_opts.mo_cos != 0;
6674 static struct lu_device_type_operations mdt_device_type_ops = {
6675 .ldto_device_alloc = mdt_device_alloc,
6676 .ldto_device_free = mdt_device_free,
6677 .ldto_device_fini = mdt_device_fini
6680 static struct lu_device_type mdt_device_type = {
6681 .ldt_tags = LU_DEVICE_MD,
6682 .ldt_name = LUSTRE_MDT_NAME,
6683 .ldt_ops = &mdt_device_type_ops,
6684 .ldt_ctx_tags = LCT_MD_THREAD
6687 static int __init mdt_init(void)
6691 CLASSERT(sizeof("0x0123456789ABCDEF:0x01234567:0x01234567") ==
6692 FID_NOBRACE_LEN + 1);
6693 CLASSERT(sizeof("[0x0123456789ABCDEF:0x01234567:0x01234567]") ==
6695 rc = lu_kmem_init(mdt_caches);
6699 rc = mds_mod_init();
6703 rc = class_register_type(&mdt_obd_device_ops, NULL, true, NULL,
6704 LUSTRE_MDT_NAME, &mdt_device_type);
6709 lu_kmem_fini(mdt_caches);
6716 static void __exit mdt_exit(void)
6718 class_unregister_type(LUSTRE_MDT_NAME);
6720 lu_kmem_fini(mdt_caches);
6723 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
6724 MODULE_DESCRIPTION("Lustre Metadata Target ("LUSTRE_MDT_NAME")");
6725 MODULE_VERSION(LUSTRE_VERSION_STRING);
6726 MODULE_LICENSE("GPL");
6728 module_init(mdt_init);
6729 module_exit(mdt_exit);