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/
31 * lustre/mdt/mdt_handler.c
33 * Lustre Metadata Target (mdt) request handler
35 * Author: Peter Braam <braam@clusterfs.com>
36 * Author: Andreas Dilger <adilger@clusterfs.com>
37 * Author: Phil Schwan <phil@clusterfs.com>
38 * Author: Mike Shaver <shaver@clusterfs.com>
39 * Author: Nikita Danilov <nikita@clusterfs.com>
40 * Author: Huang Hua <huanghua@clusterfs.com>
41 * Author: Yury Umanets <umka@clusterfs.com>
44 #define DEBUG_SUBSYSTEM S_MDS
46 #include <linux/module.h>
47 #include <linux/pagemap.h>
49 #include <dt_object.h>
50 #include <lustre_acl.h>
51 #include <lustre_export.h>
52 #include <uapi/linux/lustre/lustre_ioctl.h>
53 #include <lustre_lfsck.h>
54 #include <lustre_log.h>
55 #include <lustre_nodemap.h>
56 #include <lustre_mds.h>
57 #include <uapi/linux/lustre/lustre_param.h>
58 #include <lustre_quota.h>
59 #include <lustre_swab.h>
60 #include <lustre_lmv.h>
62 #include <obd_support.h>
63 #include <lustre_barrier.h>
64 #include <obd_cksum.h>
65 #include <llog_swab.h>
67 #include "mdt_internal.h"
69 static unsigned int max_mod_rpcs_per_client = 8;
70 module_param(max_mod_rpcs_per_client, uint, 0644);
71 MODULE_PARM_DESC(max_mod_rpcs_per_client, "maximum number of modify RPCs in flight allowed per client");
73 mdl_mode_t mdt_mdl_lock_modes[] = {
74 [LCK_MINMODE] = MDL_MINMODE,
81 [LCK_GROUP] = MDL_GROUP
84 enum ldlm_mode mdt_dlm_lock_modes[] = {
85 [MDL_MINMODE] = LCK_MINMODE,
92 [MDL_GROUP] = LCK_GROUP
95 static struct mdt_device *mdt_dev(struct lu_device *d);
97 static const struct lu_object_operations mdt_obj_ops;
99 /* Slab for MDT object allocation */
100 static struct kmem_cache *mdt_object_kmem;
102 /* For HSM restore handles */
103 struct kmem_cache *mdt_hsm_cdt_kmem;
105 /* For HSM request handles */
106 struct kmem_cache *mdt_hsm_car_kmem;
108 static struct lu_kmem_descr mdt_caches[] = {
110 .ckd_cache = &mdt_object_kmem,
111 .ckd_name = "mdt_obj",
112 .ckd_size = sizeof(struct mdt_object)
115 .ckd_cache = &mdt_hsm_cdt_kmem,
116 .ckd_name = "mdt_cdt_restore_handle",
117 .ckd_size = sizeof(struct cdt_restore_handle)
120 .ckd_cache = &mdt_hsm_car_kmem,
121 .ckd_name = "mdt_cdt_agent_req",
122 .ckd_size = sizeof(struct cdt_agent_req)
129 __u64 mdt_get_disposition(struct ldlm_reply *rep, __u64 op_flag)
133 return rep->lock_policy_res1 & op_flag;
136 void mdt_clear_disposition(struct mdt_thread_info *info,
137 struct ldlm_reply *rep, __u64 op_flag)
140 info->mti_opdata &= ~op_flag;
141 tgt_opdata_clear(info->mti_env, op_flag);
144 rep->lock_policy_res1 &= ~op_flag;
147 void mdt_set_disposition(struct mdt_thread_info *info,
148 struct ldlm_reply *rep, __u64 op_flag)
151 info->mti_opdata |= op_flag;
152 tgt_opdata_set(info->mti_env, op_flag);
155 rep->lock_policy_res1 |= op_flag;
158 void mdt_lock_reg_init(struct mdt_lock_handle *lh, enum ldlm_mode lm)
160 lh->mlh_pdo_hash = 0;
161 lh->mlh_reg_mode = lm;
162 lh->mlh_rreg_mode = lm;
163 lh->mlh_type = MDT_REG_LOCK;
166 void mdt_lh_reg_init(struct mdt_lock_handle *lh, struct ldlm_lock *lock)
168 mdt_lock_reg_init(lh, lock->l_req_mode);
169 if (lock->l_req_mode == LCK_GROUP)
170 lh->mlh_gid = lock->l_policy_data.l_inodebits.li_gid;
173 void mdt_lock_pdo_init(struct mdt_lock_handle *lh, enum ldlm_mode lock_mode,
174 const struct lu_name *lname)
176 lh->mlh_reg_mode = lock_mode;
177 lh->mlh_pdo_mode = LCK_MINMODE;
178 lh->mlh_rreg_mode = lock_mode;
179 lh->mlh_type = MDT_PDO_LOCK;
181 if (lu_name_is_valid(lname)) {
182 lh->mlh_pdo_hash = ll_full_name_hash(NULL, lname->ln_name,
184 /* XXX Workaround for LU-2856
186 * Zero is a valid return value of full_name_hash, but
187 * several users of mlh_pdo_hash assume a non-zero
188 * hash value. We therefore map zero onto an
189 * arbitrary, but consistent value (1) to avoid
190 * problems further down the road. */
191 if (unlikely(lh->mlh_pdo_hash == 0))
192 lh->mlh_pdo_hash = 1;
194 lh->mlh_pdo_hash = 0;
198 static void mdt_lock_pdo_mode(struct mdt_thread_info *info, struct mdt_object *o,
199 struct mdt_lock_handle *lh)
205 * Any dir access needs couple of locks:
207 * 1) on part of dir we gonna take lookup/modify;
209 * 2) on whole dir to protect it from concurrent splitting and/or to
210 * flush client's cache for readdir().
212 * so, for a given mode and object this routine decides what lock mode
213 * to use for lock #2:
215 * 1) if caller's gonna lookup in dir then we need to protect dir from
216 * being splitted only - LCK_CR
218 * 2) if caller's gonna modify dir then we need to protect dir from
219 * being splitted and to flush cache - LCK_CW
221 * 3) if caller's gonna modify dir and that dir seems ready for
222 * splitting then we need to protect it from any type of access
223 * (lookup/modify/split) - LCK_EX --bzzz
226 LASSERT(lh->mlh_reg_mode != LCK_MINMODE);
227 LASSERT(lh->mlh_pdo_mode == LCK_MINMODE);
230 * Ask underlaying level its opinion about preferable PDO lock mode
231 * having access type passed as regular lock mode:
233 * - MDL_MINMODE means that lower layer does not want to specify lock
236 * - MDL_NL means that no PDO lock should be taken. This is used in some
237 * cases. Say, for non-splittable directories no need to use PDO locks
240 mode = mdo_lock_mode(info->mti_env, mdt_object_child(o),
241 mdt_dlm_mode2mdl_mode(lh->mlh_reg_mode));
243 if (mode != MDL_MINMODE) {
244 lh->mlh_pdo_mode = mdt_mdl_mode2dlm_mode(mode);
247 * Lower layer does not want to specify locking mode. We do it
248 * our selves. No special protection is needed, just flush
249 * client's cache on modification and allow concurrent
252 switch (lh->mlh_reg_mode) {
254 lh->mlh_pdo_mode = LCK_EX;
257 lh->mlh_pdo_mode = LCK_CR;
260 lh->mlh_pdo_mode = LCK_CW;
263 CERROR("Not expected lock type (0x%x)\n",
264 (int)lh->mlh_reg_mode);
269 LASSERT(lh->mlh_pdo_mode != LCK_MINMODE);
273 static int mdt_lookup_fileset(struct mdt_thread_info *info, const char *fileset,
276 struct mdt_device *mdt = info->mti_mdt;
277 struct lu_name *lname = &info->mti_name;
278 const char *start = fileset;
279 char *filename = info->mti_filename;
280 struct mdt_object *parent;
284 LASSERT(!info->mti_cross_ref);
287 * We may want to allow this to mount a completely separate
288 * fileset from the MDT in the future, but keeping it to
289 * ROOT/ only for now avoid potential security issues.
291 *fid = mdt->mdt_md_root_fid;
293 while (rc == 0 && start != NULL && *start != '\0') {
294 const char *s1 = start;
300 while (*s2 != '/' && *s2 != '\0')
308 lname->ln_namelen = s2 - s1;
309 if (lname->ln_namelen > NAME_MAX) {
314 /* reject .. as a path component */
315 if (lname->ln_namelen == 2 &&
316 strncmp(s1, "..", 2) == 0) {
321 strncpy(filename, s1, lname->ln_namelen);
322 filename[lname->ln_namelen] = '\0';
323 lname->ln_name = filename;
325 parent = mdt_object_find(info->mti_env, mdt, fid);
326 if (IS_ERR(parent)) {
327 rc = PTR_ERR(parent);
330 /* Only got the fid of this obj by name */
332 rc = mdo_lookup(info->mti_env, mdt_object_child(parent), lname,
333 fid, &info->mti_spec);
334 mdt_object_put(info->mti_env, parent);
337 parent = mdt_object_find(info->mti_env, mdt, fid);
339 rc = PTR_ERR(parent);
341 mode = lu_object_attr(&parent->mot_obj);
342 if (!S_ISDIR(mode)) {
344 } else if (mdt_is_remote_object(info, parent, parent)) {
345 if (!mdt->mdt_enable_remote_subdir_mount) {
347 LCONSOLE_WARN("%s: subdir mount '%s' refused because 'enable_remote_subdir_mount=0': rc = %d\n",
351 LCONSOLE_INFO("%s: subdir mount '%s' is remote and may be slow\n",
356 mdt_object_put(info->mti_env, parent);
363 static int mdt_get_root(struct tgt_session_info *tsi)
365 struct mdt_thread_info *info = tsi2mdt_info(tsi);
366 struct mdt_device *mdt = info->mti_mdt;
367 struct mdt_body *repbody;
368 char *fileset = NULL, *buffer = NULL;
370 struct obd_export *exp = info->mti_exp;
371 char *nodemap_fileset;
375 rc = mdt_check_ucred(info);
377 GOTO(out, rc = err_serious(rc));
379 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GET_ROOT_PACK))
380 GOTO(out, rc = err_serious(-ENOMEM));
382 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
383 if (req_capsule_get_size(info->mti_pill, &RMF_NAME, RCL_CLIENT) > 0) {
384 fileset = req_capsule_client_get(info->mti_pill, &RMF_NAME);
386 GOTO(out, rc = err_serious(-EFAULT));
389 nodemap_fileset = nodemap_get_fileset(exp->exp_target_data.ted_nodemap);
390 if (nodemap_fileset && nodemap_fileset[0]) {
391 CDEBUG(D_INFO, "nodemap fileset is %s\n", nodemap_fileset);
393 /* consider fileset from client as a sub-fileset
394 * of the nodemap one */
395 OBD_ALLOC(buffer, PATH_MAX + 1);
397 GOTO(out, rc = err_serious(-ENOMEM));
398 if (snprintf(buffer, PATH_MAX + 1, "%s/%s",
399 nodemap_fileset, fileset) >= PATH_MAX + 1)
400 GOTO(out, rc = err_serious(-EINVAL));
403 /* enforce fileset as specified in the nodemap */
404 fileset = nodemap_fileset;
409 CDEBUG(D_INFO, "Getting fileset %s\n", fileset);
410 rc = mdt_lookup_fileset(info, fileset, &repbody->mbo_fid1);
412 GOTO(out, rc = err_serious(rc));
414 repbody->mbo_fid1 = mdt->mdt_md_root_fid;
416 repbody->mbo_valid |= OBD_MD_FLID;
420 mdt_thread_info_fini(info);
422 OBD_FREE(buffer, PATH_MAX+1);
426 static int mdt_statfs(struct tgt_session_info *tsi)
428 struct ptlrpc_request *req = tgt_ses_req(tsi);
429 struct mdt_thread_info *info = tsi2mdt_info(tsi);
430 struct mdt_device *mdt = info->mti_mdt;
431 struct tg_grants_data *tgd = &mdt->mdt_lut.lut_tgd;
432 struct md_device *next = mdt->mdt_child;
433 struct ptlrpc_service_part *svcpt;
434 struct obd_statfs *osfs;
435 struct mdt_body *reqbody = NULL;
436 struct mdt_statfs_cache *msf;
437 ktime_t kstart = ktime_get();
438 int current_blockbits;
443 svcpt = req->rq_rqbd->rqbd_svcpt;
445 /* This will trigger a watchdog timeout */
446 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_STATFS_LCW_SLEEP,
447 (MDT_SERVICE_WATCHDOG_FACTOR *
448 at_get(&svcpt->scp_at_estimate)) + 1);
450 rc = mdt_check_ucred(info);
452 GOTO(out, rc = err_serious(rc));
454 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK))
455 GOTO(out, rc = err_serious(-ENOMEM));
457 osfs = req_capsule_server_get(info->mti_pill, &RMF_OBD_STATFS);
459 GOTO(out, rc = -EPROTO);
461 if (mdt_is_sum_statfs_client(req->rq_export) &&
462 lustre_packed_msg_size(req->rq_reqmsg) ==
463 req_capsule_fmt_size(req->rq_reqmsg->lm_magic,
464 &RQF_MDS_STATFS_NEW, RCL_CLIENT)) {
465 req_capsule_extend(info->mti_pill, &RQF_MDS_STATFS_NEW);
466 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
469 if (reqbody && reqbody->mbo_valid & OBD_MD_FLAGSTATFS)
470 msf = &mdt->mdt_sum_osfs;
472 msf = &mdt->mdt_osfs;
474 if (msf->msf_age + OBD_STATFS_CACHE_SECONDS <= ktime_get_seconds()) {
475 /** statfs data is too old, get up-to-date one */
476 if (reqbody && reqbody->mbo_valid & OBD_MD_FLAGSTATFS)
477 rc = next->md_ops->mdo_statfs(info->mti_env,
480 rc = dt_statfs(info->mti_env, mdt->mdt_bottom,
484 spin_lock(&mdt->mdt_lock);
485 msf->msf_osfs = *osfs;
486 msf->msf_age = ktime_get_seconds();
487 spin_unlock(&mdt->mdt_lock);
489 /** use cached statfs data */
490 spin_lock(&mdt->mdt_lock);
491 *osfs = msf->msf_osfs;
492 spin_unlock(&mdt->mdt_lock);
495 /* tgd_blockbit is recordsize bits set during mkfs.
496 * This once set does not change. However, 'zfs set'
497 * can be used to change the MDT blocksize. Instead
498 * of using cached value of 'tgd_blockbit' always
499 * calculate the blocksize bits which may have
502 current_blockbits = fls64(osfs->os_bsize) - 1;
504 /* at least try to account for cached pages. its still racy and
505 * might be under-reporting if clients haven't announced their
506 * caches with brw recently */
507 CDEBUG(D_SUPER | D_CACHE, "blocks cached %llu granted %llu"
508 " pending %llu free %llu avail %llu\n",
509 tgd->tgd_tot_dirty, tgd->tgd_tot_granted,
510 tgd->tgd_tot_pending,
511 osfs->os_bfree << current_blockbits,
512 osfs->os_bavail << current_blockbits);
514 osfs->os_bavail -= min_t(u64, osfs->os_bavail,
515 ((tgd->tgd_tot_dirty + tgd->tgd_tot_pending +
516 osfs->os_bsize - 1) >> current_blockbits));
518 tgt_grant_sanity_check(mdt->mdt_lu_dev.ld_obd, __func__);
519 CDEBUG(D_CACHE, "%llu blocks: %llu free, %llu avail; "
520 "%llu objects: %llu free; state %x\n",
521 osfs->os_blocks, osfs->os_bfree, osfs->os_bavail,
522 osfs->os_files, osfs->os_ffree, osfs->os_state);
524 if (!exp_grant_param_supp(tsi->tsi_exp) &&
525 current_blockbits > COMPAT_BSIZE_SHIFT) {
526 /* clients which don't support OBD_CONNECT_GRANT_PARAM
527 * should not see a block size > page size, otherwise
528 * cl_lost_grant goes mad. Therefore, we emulate a 4KB (=2^12)
529 * block size which is the biggest block size known to work
530 * with all client's page size. */
531 osfs->os_blocks <<= current_blockbits - COMPAT_BSIZE_SHIFT;
532 osfs->os_bfree <<= current_blockbits - COMPAT_BSIZE_SHIFT;
533 osfs->os_bavail <<= current_blockbits - COMPAT_BSIZE_SHIFT;
534 osfs->os_bsize = 1 << COMPAT_BSIZE_SHIFT;
537 mdt_counter_incr(req, LPROC_MDT_STATFS,
538 ktime_us_delta(ktime_get(), kstart));
540 mdt_thread_info_fini(info);
544 __u32 mdt_lmm_dom_entry_check(struct lov_mds_md *lmm, int *is_dom_only)
546 struct lov_comp_md_v1 *comp_v1;
547 struct lov_mds_md *v1;
549 __u32 dom_stripesize = 0;
551 bool has_ost_stripes = false;
558 if (le32_to_cpu(lmm->lmm_magic) != LOV_MAGIC_COMP_V1)
561 comp_v1 = (struct lov_comp_md_v1 *)lmm;
562 off = le32_to_cpu(comp_v1->lcm_entries[0].lcme_offset);
563 v1 = (struct lov_mds_md *)((char *)comp_v1 + off);
565 /* Fast check for DoM entry with no mirroring, should be the first */
566 if (le16_to_cpu(comp_v1->lcm_mirror_count) == 0 &&
567 lov_pattern(le32_to_cpu(v1->lmm_pattern)) != LOV_PATTERN_MDT)
570 /* check all entries otherwise */
571 for (i = 0; i < le16_to_cpu(comp_v1->lcm_entry_count); i++) {
572 struct lov_comp_md_entry_v1 *lcme;
574 lcme = &comp_v1->lcm_entries[i];
575 if (!(le32_to_cpu(lcme->lcme_flags) & LCME_FL_INIT))
578 off = le32_to_cpu(lcme->lcme_offset);
579 v1 = (struct lov_mds_md *)((char *)comp_v1 + off);
581 if (lov_pattern(le32_to_cpu(v1->lmm_pattern)) ==
583 dom_stripesize = le32_to_cpu(v1->lmm_stripe_size);
585 has_ost_stripes = true;
587 if (dom_stripesize && has_ost_stripes)
588 RETURN(dom_stripesize);
590 /* DoM-only case exits here */
591 if (is_dom_only && dom_stripesize)
593 RETURN(dom_stripesize);
597 * Pack size attributes into the reply.
599 int mdt_pack_size2body(struct mdt_thread_info *info,
600 const struct lu_fid *fid, struct lustre_handle *lh)
603 struct md_attr *ma = &info->mti_attr;
605 bool dom_lock = false;
609 LASSERT(ma->ma_attr.la_valid & LA_MODE);
611 if (!S_ISREG(ma->ma_attr.la_mode) ||
612 !(ma->ma_valid & MA_LOV && ma->ma_lmm != NULL))
615 dom_stripe = mdt_lmm_dom_stripesize(ma->ma_lmm);
616 /* no DoM stripe, no size in reply */
620 if (lustre_handle_is_used(lh)) {
621 struct ldlm_lock *lock;
623 lock = ldlm_handle2lock(lh);
625 dom_lock = ldlm_has_dom(lock);
630 /* no DoM lock, no size in reply */
634 /* Either DoM lock exists or LMM has only DoM stripe then
635 * return size on body. */
636 b = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
638 mdt_dom_object_size(info->mti_env, info->mti_mdt, fid, b, dom_lock);
642 #ifdef CONFIG_LUSTRE_FS_POSIX_ACL
644 * Pack ACL data into the reply. UIDs/GIDs are mapped and filtered by nodemap.
646 * \param info thread info object
647 * \param repbody reply to pack ACLs into
648 * \param o mdt object of file to examine
649 * \param nodemap nodemap of client to reply to
651 * \retval -errno error getting or parsing ACL from disk
653 int mdt_pack_acl2body(struct mdt_thread_info *info, struct mdt_body *repbody,
654 struct mdt_object *o, struct lu_nodemap *nodemap)
656 const struct lu_env *env = info->mti_env;
657 struct md_object *next = mdt_object_child(o);
658 struct lu_buf *buf = &info->mti_buf;
659 struct mdt_device *mdt = info->mti_mdt;
660 struct req_capsule *pill = info->mti_pill;
665 buf->lb_buf = req_capsule_server_get(pill, &RMF_ACL);
666 buf->lb_len = req_capsule_get_size(pill, &RMF_ACL, RCL_SERVER);
667 if (buf->lb_len == 0)
670 LASSERT(!info->mti_big_acl_used);
672 rc = mo_xattr_get(env, next, buf, XATTR_NAME_ACL_ACCESS);
674 if (rc == -ENODATA) {
675 repbody->mbo_aclsize = 0;
676 repbody->mbo_valid |= OBD_MD_FLACL;
678 } else if (rc == -EOPNOTSUPP) {
680 } else if (rc == -ERANGE) {
681 if (exp_connect_large_acl(info->mti_exp) &&
682 !info->mti_big_acl_used) {
683 if (info->mti_big_acl == NULL) {
684 info->mti_big_aclsize =
686 mdt->mdt_max_ea_size,
688 OBD_ALLOC_LARGE(info->mti_big_acl,
689 info->mti_big_aclsize);
690 if (info->mti_big_acl == NULL) {
691 info->mti_big_aclsize = 0;
692 CERROR("%s: unable to grow "
695 PFID(mdt_object_fid(o)));
700 CDEBUG(D_INODE, "%s: grow the "DFID
701 " ACL buffer to size %d\n",
703 PFID(mdt_object_fid(o)),
704 info->mti_big_aclsize);
706 buf->lb_buf = info->mti_big_acl;
707 buf->lb_len = info->mti_big_aclsize;
708 info->mti_big_acl_used = 1;
711 /* FS has ACL bigger that our limits */
712 CDEBUG(D_INODE, "%s: "DFID" ACL can't fit into %d\n",
713 mdt_obd_name(mdt), PFID(mdt_object_fid(o)),
714 info->mti_big_aclsize);
717 CERROR("%s: unable to read "DFID" ACL: rc = %d\n",
718 mdt_obd_name(mdt), PFID(mdt_object_fid(o)), rc);
721 rc = nodemap_map_acl(nodemap, buf->lb_buf,
722 rc, NODEMAP_FS_TO_CLIENT);
723 /* if all ACLs mapped out, rc is still >= 0 */
725 CERROR("%s: nodemap_map_acl unable to parse "DFID
726 " ACL: rc = %d\n", mdt_obd_name(mdt),
727 PFID(mdt_object_fid(o)), rc);
728 repbody->mbo_aclsize = 0;
729 repbody->mbo_valid &= ~OBD_MD_FLACL;
731 repbody->mbo_aclsize = rc;
732 repbody->mbo_valid |= OBD_MD_FLACL;
741 /* XXX Look into layout in MDT layer. */
742 static inline bool mdt_hsm_is_released(struct lov_mds_md *lmm)
744 struct lov_comp_md_v1 *comp_v1;
745 struct lov_mds_md *v1;
748 if (lmm->lmm_magic == LOV_MAGIC_COMP_V1) {
749 comp_v1 = (struct lov_comp_md_v1 *)lmm;
751 for (i = 0; i < comp_v1->lcm_entry_count; i++) {
752 v1 = (struct lov_mds_md *)((char *)comp_v1 +
753 comp_v1->lcm_entries[i].lcme_offset);
754 /* We don't support partial release for now */
755 if (!(v1->lmm_pattern & LOV_PATTERN_F_RELEASED))
760 return (lmm->lmm_pattern & LOV_PATTERN_F_RELEASED) ?
765 void mdt_pack_attr2body(struct mdt_thread_info *info, struct mdt_body *b,
766 const struct lu_attr *attr, const struct lu_fid *fid)
768 struct md_attr *ma = &info->mti_attr;
769 struct obd_export *exp = info->mti_exp;
770 struct lu_nodemap *nodemap = NULL;
772 LASSERT(ma->ma_valid & MA_INODE);
774 if (attr->la_valid & LA_ATIME) {
775 b->mbo_atime = attr->la_atime;
776 b->mbo_valid |= OBD_MD_FLATIME;
778 if (attr->la_valid & LA_MTIME) {
779 b->mbo_mtime = attr->la_mtime;
780 b->mbo_valid |= OBD_MD_FLMTIME;
782 if (attr->la_valid & LA_CTIME) {
783 b->mbo_ctime = attr->la_ctime;
784 b->mbo_valid |= OBD_MD_FLCTIME;
786 if (attr->la_valid & LA_BTIME) {
787 b->mbo_btime = attr->la_btime;
788 b->mbo_valid |= OBD_MD_FLBTIME;
790 if (attr->la_valid & LA_FLAGS) {
791 b->mbo_flags = attr->la_flags;
792 b->mbo_valid |= OBD_MD_FLFLAGS;
794 if (attr->la_valid & LA_NLINK) {
795 b->mbo_nlink = attr->la_nlink;
796 b->mbo_valid |= OBD_MD_FLNLINK;
798 if (attr->la_valid & (LA_UID|LA_GID)) {
799 nodemap = nodemap_get_from_exp(exp);
803 if (attr->la_valid & LA_UID) {
804 b->mbo_uid = nodemap_map_id(nodemap, NODEMAP_UID,
805 NODEMAP_FS_TO_CLIENT,
807 b->mbo_valid |= OBD_MD_FLUID;
809 if (attr->la_valid & LA_GID) {
810 b->mbo_gid = nodemap_map_id(nodemap, NODEMAP_GID,
811 NODEMAP_FS_TO_CLIENT,
813 b->mbo_valid |= OBD_MD_FLGID;
816 if (attr->la_valid & LA_PROJID) {
817 /* TODO, nodemap for project id */
818 b->mbo_projid = attr->la_projid;
819 b->mbo_valid |= OBD_MD_FLPROJID;
822 b->mbo_mode = attr->la_mode;
823 if (attr->la_valid & LA_MODE)
824 b->mbo_valid |= OBD_MD_FLMODE;
825 if (attr->la_valid & LA_TYPE)
826 b->mbo_valid |= OBD_MD_FLTYPE;
830 b->mbo_valid |= OBD_MD_FLID;
831 CDEBUG(D_INODE, DFID": nlink=%d, mode=%o, valid=%#llx\n",
832 PFID(fid), b->mbo_nlink, b->mbo_mode, b->mbo_valid);
835 if (!(attr->la_valid & LA_TYPE))
838 b->mbo_rdev = attr->la_rdev;
839 b->mbo_size = attr->la_size;
840 b->mbo_blocks = attr->la_blocks;
842 if (!S_ISREG(attr->la_mode)) {
843 b->mbo_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS | OBD_MD_FLRDEV;
844 } else if (ma->ma_need & MA_LOV && !(ma->ma_valid & MA_LOV)) {
845 /* means no objects are allocated on osts. */
846 LASSERT(!(ma->ma_valid & MA_LOV));
847 /* just ignore blocks occupied by extend attributes on MDS */
849 /* if no object is allocated on osts, the size on mds is valid.
851 b->mbo_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
852 } else if ((ma->ma_valid & MA_LOV) && ma->ma_lmm != NULL) {
853 if (mdt_hsm_is_released(ma->ma_lmm)) {
854 /* A released file stores its size on MDS. */
855 /* But return 1 block for released file, unless tools
856 * like tar will consider it fully sparse. (LU-3864)
858 if (unlikely(b->mbo_size == 0))
862 b->mbo_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
863 } else if (info->mti_som_valid) { /* som is valid */
864 b->mbo_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
865 } else if (ma->ma_valid & MA_SOM) { /* lsom is valid */
866 b->mbo_valid |= OBD_MD_FLLAZYSIZE | OBD_MD_FLLAZYBLOCKS;
867 b->mbo_size = ma->ma_som.ms_size;
868 b->mbo_blocks = ma->ma_som.ms_blocks;
872 if (fid != NULL && (b->mbo_valid & OBD_MD_FLSIZE ||
873 b->mbo_valid & OBD_MD_FLLAZYSIZE))
874 CDEBUG(D_VFSTRACE, DFID": returning size %llu\n",
875 PFID(fid), (unsigned long long)b->mbo_size);
878 if (!IS_ERR_OR_NULL(nodemap))
879 nodemap_putref(nodemap);
882 static inline int mdt_body_has_lov(const struct lu_attr *la,
883 const struct mdt_body *body)
885 return (S_ISREG(la->la_mode) && (body->mbo_valid & OBD_MD_FLEASIZE)) ||
886 (S_ISDIR(la->la_mode) && (body->mbo_valid & OBD_MD_FLDIREA));
889 void mdt_client_compatibility(struct mdt_thread_info *info)
891 struct mdt_body *body;
892 struct ptlrpc_request *req = mdt_info_req(info);
893 struct obd_export *exp = req->rq_export;
894 struct md_attr *ma = &info->mti_attr;
895 struct lu_attr *la = &ma->ma_attr;
898 if (exp_connect_layout(exp))
899 /* the client can deal with 16-bit lmm_stripe_count */
902 body = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
904 if (!mdt_body_has_lov(la, body))
907 /* now we have a reply with a lov for a client not compatible with the
908 * layout lock so we have to clean the layout generation number */
909 if (S_ISREG(la->la_mode))
910 ma->ma_lmm->lmm_layout_gen = 0;
914 static int mdt_attr_get_eabuf_size(struct mdt_thread_info *info,
915 struct mdt_object *o)
917 const struct lu_env *env = info->mti_env;
920 rc = mo_xattr_get(env, mdt_object_child(o), &LU_BUF_NULL,
929 /* Is it a directory? Let's check for the LMV as well */
930 if (S_ISDIR(lu_object_attr(&mdt_object_child(o)->mo_lu))) {
931 rc2 = mo_xattr_get(env, mdt_object_child(o), &LU_BUF_NULL,
935 rc2 = mo_xattr_get(env, mdt_object_child(o),
937 XATTR_NAME_DEFAULT_LMV);
939 if ((rc2 < 0 && rc2 != -ENODATA) || (rc2 > rc))
947 int mdt_big_xattr_get(struct mdt_thread_info *info, struct mdt_object *o,
950 const struct lu_env *env = info->mti_env;
954 LASSERT(info->mti_big_lmm_used == 0);
955 rc = mo_xattr_get(env, mdt_object_child(o), &LU_BUF_NULL, name);
959 /* big_lmm may need to be grown */
960 if (info->mti_big_lmmsize < rc) {
961 int size = size_roundup_power2(rc);
963 if (info->mti_big_lmmsize > 0) {
964 /* free old buffer */
965 LASSERT(info->mti_big_lmm);
966 OBD_FREE_LARGE(info->mti_big_lmm,
967 info->mti_big_lmmsize);
968 info->mti_big_lmm = NULL;
969 info->mti_big_lmmsize = 0;
972 OBD_ALLOC_LARGE(info->mti_big_lmm, size);
973 if (info->mti_big_lmm == NULL)
975 info->mti_big_lmmsize = size;
977 LASSERT(info->mti_big_lmmsize >= rc);
979 info->mti_buf.lb_buf = info->mti_big_lmm;
980 info->mti_buf.lb_len = info->mti_big_lmmsize;
981 rc = mo_xattr_get(env, mdt_object_child(o), &info->mti_buf, name);
986 int __mdt_stripe_get(struct mdt_thread_info *info, struct mdt_object *o,
987 struct md_attr *ma, const char *name)
989 struct md_object *next = mdt_object_child(o);
990 struct lu_buf *buf = &info->mti_buf;
993 if (strcmp(name, XATTR_NAME_LOV) == 0) {
994 buf->lb_buf = ma->ma_lmm;
995 buf->lb_len = ma->ma_lmm_size;
996 LASSERT(!(ma->ma_valid & MA_LOV));
997 } else if (strcmp(name, XATTR_NAME_LMV) == 0) {
998 buf->lb_buf = ma->ma_lmv;
999 buf->lb_len = ma->ma_lmv_size;
1000 LASSERT(!(ma->ma_valid & MA_LMV));
1001 } else if (strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0) {
1002 buf->lb_buf = ma->ma_default_lmv;
1003 buf->lb_len = ma->ma_default_lmv_size;
1004 LASSERT(!(ma->ma_valid & MA_LMV_DEF));
1009 LASSERT(buf->lb_buf);
1011 rc = mo_xattr_get(info->mti_env, next, buf, name);
1015 if (strcmp(name, XATTR_NAME_LOV) == 0) {
1016 if (info->mti_big_lmm_used)
1017 ma->ma_lmm = info->mti_big_lmm;
1019 /* NOT return LOV EA with hole to old client. */
1020 if (unlikely(le32_to_cpu(ma->ma_lmm->lmm_pattern) &
1021 LOV_PATTERN_F_HOLE) &&
1022 !(exp_connect_flags(info->mti_exp) &
1023 OBD_CONNECT_LFSCK)) {
1026 ma->ma_lmm_size = rc;
1027 ma->ma_valid |= MA_LOV;
1029 } else if (strcmp(name, XATTR_NAME_LMV) == 0) {
1030 if (info->mti_big_lmm_used)
1031 ma->ma_lmv = info->mti_big_lmm;
1033 ma->ma_lmv_size = rc;
1034 ma->ma_valid |= MA_LMV;
1035 } else if (strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0) {
1036 ma->ma_default_lmv_size = rc;
1037 ma->ma_valid |= MA_LMV_DEF;
1040 /* Update mdt_max_mdsize so all clients will be aware that */
1041 if (info->mti_mdt->mdt_max_mdsize < rc)
1042 info->mti_mdt->mdt_max_mdsize = rc;
1045 } else if (rc == -ENODATA) {
1048 } else if (rc == -ERANGE) {
1049 /* Default LMV has fixed size, so it must be able to fit
1050 * in the original buffer */
1051 if (strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0)
1053 rc = mdt_big_xattr_get(info, o, name);
1055 info->mti_big_lmm_used = 1;
1063 int mdt_stripe_get(struct mdt_thread_info *info, struct mdt_object *o,
1064 struct md_attr *ma, const char *name)
1068 if (!info->mti_big_lmm) {
1069 OBD_ALLOC(info->mti_big_lmm, PAGE_SIZE);
1070 if (!info->mti_big_lmm)
1072 info->mti_big_lmmsize = PAGE_SIZE;
1075 if (strcmp(name, XATTR_NAME_LOV) == 0) {
1076 ma->ma_lmm = info->mti_big_lmm;
1077 ma->ma_lmm_size = info->mti_big_lmmsize;
1078 ma->ma_valid &= ~MA_LOV;
1079 } else if (strcmp(name, XATTR_NAME_LMV) == 0) {
1080 ma->ma_lmv = info->mti_big_lmm;
1081 ma->ma_lmv_size = info->mti_big_lmmsize;
1082 ma->ma_valid &= ~MA_LMV;
1087 LASSERT(!info->mti_big_lmm_used);
1088 rc = __mdt_stripe_get(info, o, ma, name);
1089 /* since big_lmm is always used here, clear 'used' flag to avoid
1090 * assertion in mdt_big_xattr_get().
1092 info->mti_big_lmm_used = 0;
1097 int mdt_attr_get_pfid(struct mdt_thread_info *info, struct mdt_object *o,
1098 struct lu_fid *pfid)
1100 struct lu_buf *buf = &info->mti_buf;
1101 struct link_ea_header *leh;
1102 struct link_ea_entry *lee;
1106 buf->lb_buf = info->mti_big_lmm;
1107 buf->lb_len = info->mti_big_lmmsize;
1108 rc = mo_xattr_get(info->mti_env, mdt_object_child(o),
1109 buf, XATTR_NAME_LINK);
1110 /* ignore errors, MA_PFID won't be set and it is
1111 * up to the caller to treat this as an error */
1112 if (rc == -ERANGE || buf->lb_len == 0) {
1113 rc = mdt_big_xattr_get(info, o, XATTR_NAME_LINK);
1114 buf->lb_buf = info->mti_big_lmm;
1115 buf->lb_len = info->mti_big_lmmsize;
1120 if (rc < sizeof(*leh)) {
1121 CERROR("short LinkEA on "DFID": rc = %d\n",
1122 PFID(mdt_object_fid(o)), rc);
1126 leh = (struct link_ea_header *) buf->lb_buf;
1127 lee = (struct link_ea_entry *)(leh + 1);
1128 if (leh->leh_magic == __swab32(LINK_EA_MAGIC)) {
1129 leh->leh_magic = LINK_EA_MAGIC;
1130 leh->leh_reccount = __swab32(leh->leh_reccount);
1131 leh->leh_len = __swab64(leh->leh_len);
1133 if (leh->leh_magic != LINK_EA_MAGIC)
1135 if (leh->leh_reccount == 0)
1138 memcpy(pfid, &lee->lee_parent_fid, sizeof(*pfid));
1139 fid_be_to_cpu(pfid, pfid);
1144 int mdt_attr_get_pfid_name(struct mdt_thread_info *info, struct mdt_object *o,
1145 struct lu_fid *pfid, struct lu_name *lname)
1147 struct lu_buf *buf = &info->mti_buf;
1148 struct link_ea_header *leh;
1149 struct link_ea_entry *lee;
1153 buf->lb_buf = info->mti_xattr_buf;
1154 buf->lb_len = sizeof(info->mti_xattr_buf);
1155 rc = mo_xattr_get(info->mti_env, mdt_object_child(o), buf,
1157 if (rc == -ERANGE) {
1158 rc = mdt_big_xattr_get(info, o, XATTR_NAME_LINK);
1159 buf->lb_buf = info->mti_big_lmm;
1160 buf->lb_len = info->mti_big_lmmsize;
1165 if (rc < sizeof(*leh)) {
1166 CERROR("short LinkEA on "DFID": rc = %d\n",
1167 PFID(mdt_object_fid(o)), rc);
1171 leh = (struct link_ea_header *)buf->lb_buf;
1172 lee = (struct link_ea_entry *)(leh + 1);
1173 if (leh->leh_magic == __swab32(LINK_EA_MAGIC)) {
1174 leh->leh_magic = LINK_EA_MAGIC;
1175 leh->leh_reccount = __swab32(leh->leh_reccount);
1176 leh->leh_len = __swab64(leh->leh_len);
1178 if (leh->leh_magic != LINK_EA_MAGIC)
1181 if (leh->leh_reccount == 0)
1184 linkea_entry_unpack(lee, &reclen, lname, pfid);
1189 int mdt_attr_get_complex(struct mdt_thread_info *info,
1190 struct mdt_object *o, struct md_attr *ma)
1192 const struct lu_env *env = info->mti_env;
1193 struct md_object *next = mdt_object_child(o);
1194 struct lu_buf *buf = &info->mti_buf;
1195 int need = ma->ma_need;
1202 if (mdt_object_exists(o) == 0)
1203 GOTO(out, rc = -ENOENT);
1204 mode = lu_object_attr(&next->mo_lu);
1206 if (need & MA_INODE) {
1207 ma->ma_need = MA_INODE;
1208 rc = mo_attr_get(env, next, ma);
1213 (void) mdt_get_som(info, o, ma);
1214 ma->ma_valid |= MA_INODE;
1217 if (need & MA_PFID) {
1218 rc = mdt_attr_get_pfid(info, o, &ma->ma_pfid);
1220 ma->ma_valid |= MA_PFID;
1221 /* ignore this error, parent fid is not mandatory */
1225 if (need & MA_LOV && (S_ISREG(mode) || S_ISDIR(mode))) {
1226 rc = __mdt_stripe_get(info, o, ma, XATTR_NAME_LOV);
1231 if (need & MA_LMV && S_ISDIR(mode)) {
1232 rc = __mdt_stripe_get(info, o, ma, XATTR_NAME_LMV);
1237 if (need & MA_LMV_DEF && S_ISDIR(mode)) {
1238 rc = __mdt_stripe_get(info, o, ma, XATTR_NAME_DEFAULT_LMV);
1244 * In the handle of MA_INODE, we may already get the SOM attr.
1246 if (need & MA_SOM && S_ISREG(mode) && !(ma->ma_valid & MA_SOM)) {
1247 rc = mdt_get_som(info, o, ma);
1252 if (need & MA_HSM && S_ISREG(mode)) {
1253 buf->lb_buf = info->mti_xattr_buf;
1254 buf->lb_len = sizeof(info->mti_xattr_buf);
1255 BUILD_BUG_ON(sizeof(struct hsm_attrs) >
1256 sizeof(info->mti_xattr_buf));
1257 rc2 = mo_xattr_get(info->mti_env, next, buf, XATTR_NAME_HSM);
1258 rc2 = lustre_buf2hsm(info->mti_xattr_buf, rc2, &ma->ma_hsm);
1260 ma->ma_valid |= MA_HSM;
1261 else if (rc2 < 0 && rc2 != -ENODATA)
1262 GOTO(out, rc = rc2);
1265 #ifdef CONFIG_LUSTRE_FS_POSIX_ACL
1266 if (need & MA_ACL_DEF && S_ISDIR(mode)) {
1267 buf->lb_buf = ma->ma_acl;
1268 buf->lb_len = ma->ma_acl_size;
1269 rc2 = mo_xattr_get(env, next, buf, XATTR_NAME_ACL_DEFAULT);
1271 ma->ma_acl_size = rc2;
1272 ma->ma_valid |= MA_ACL_DEF;
1273 } else if (rc2 == -ENODATA) {
1275 ma->ma_acl_size = 0;
1277 GOTO(out, rc = rc2);
1282 CDEBUG(D_INODE, "after getattr rc = %d, ma_valid = %#llx ma_lmm=%p\n",
1283 rc, ma->ma_valid, ma->ma_lmm);
1287 static int mdt_getattr_internal(struct mdt_thread_info *info,
1288 struct mdt_object *o, int ma_need)
1290 struct mdt_device *mdt = info->mti_mdt;
1291 struct md_object *next = mdt_object_child(o);
1292 const struct mdt_body *reqbody = info->mti_body;
1293 struct ptlrpc_request *req = mdt_info_req(info);
1294 struct md_attr *ma = &info->mti_attr;
1295 struct lu_attr *la = &ma->ma_attr;
1296 struct req_capsule *pill = info->mti_pill;
1297 const struct lu_env *env = info->mti_env;
1298 struct mdt_body *repbody;
1299 struct lu_buf *buffer = &info->mti_buf;
1300 struct obd_export *exp = info->mti_exp;
1301 ktime_t kstart = ktime_get();
1306 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK))
1307 RETURN(err_serious(-ENOMEM));
1309 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
1313 if (mdt_object_remote(o)) {
1314 /* This object is located on remote node.*/
1315 /* Return -ENOTSUPP for old client */
1316 if (!mdt_is_dne_client(req->rq_export))
1317 GOTO(out, rc = -ENOTSUPP);
1319 repbody->mbo_fid1 = *mdt_object_fid(o);
1320 repbody->mbo_valid = OBD_MD_FLID | OBD_MD_MDS;
1324 if (reqbody->mbo_eadatasize > 0) {
1325 buffer->lb_buf = req_capsule_server_get(pill, &RMF_MDT_MD);
1326 if (buffer->lb_buf == NULL)
1327 GOTO(out, rc = -EPROTO);
1328 buffer->lb_len = req_capsule_get_size(pill, &RMF_MDT_MD,
1331 buffer->lb_buf = NULL;
1333 ma_need &= ~(MA_LOV | MA_LMV);
1334 CDEBUG(D_INFO, "%s: RPC from %s: does not need LOVEA.\n",
1335 mdt_obd_name(info->mti_mdt),
1336 req->rq_export->exp_client_uuid.uuid);
1339 /* from 2.12.58 intent_getattr pack default LMV in reply */
1340 if (S_ISDIR(lu_object_attr(&next->mo_lu)) &&
1341 ((reqbody->mbo_valid & (OBD_MD_MEA | OBD_MD_DEFAULT_MEA)) ==
1342 (OBD_MD_MEA | OBD_MD_DEFAULT_MEA)) &&
1343 req_capsule_has_field(&req->rq_pill, &RMF_DEFAULT_MDT_MD,
1345 ma->ma_lmv = buffer->lb_buf;
1346 ma->ma_lmv_size = buffer->lb_len;
1347 ma->ma_default_lmv = req_capsule_server_get(pill,
1348 &RMF_DEFAULT_MDT_MD);
1349 ma->ma_default_lmv_size = req_capsule_get_size(pill,
1350 &RMF_DEFAULT_MDT_MD,
1352 ma->ma_need = MA_INODE;
1353 if (ma->ma_lmv_size > 0)
1354 ma->ma_need |= MA_LMV;
1355 if (ma->ma_default_lmv_size > 0)
1356 ma->ma_need |= MA_LMV_DEF;
1357 } else if (S_ISDIR(lu_object_attr(&next->mo_lu)) &&
1358 (reqbody->mbo_valid & (OBD_MD_MEA | OBD_MD_DEFAULT_MEA))) {
1359 /* If it is dir and client require MEA, then we got MEA */
1360 /* Assumption: MDT_MD size is enough for lmv size. */
1361 ma->ma_lmv = buffer->lb_buf;
1362 ma->ma_lmv_size = buffer->lb_len;
1363 ma->ma_need = MA_INODE;
1364 if (ma->ma_lmv_size > 0) {
1365 if (reqbody->mbo_valid & OBD_MD_MEA) {
1366 ma->ma_need |= MA_LMV;
1367 } else if (reqbody->mbo_valid & OBD_MD_DEFAULT_MEA) {
1368 ma->ma_need |= MA_LMV_DEF;
1369 ma->ma_default_lmv = buffer->lb_buf;
1371 ma->ma_default_lmv_size = buffer->lb_len;
1372 ma->ma_lmv_size = 0;
1376 ma->ma_lmm = buffer->lb_buf;
1377 ma->ma_lmm_size = buffer->lb_len;
1378 ma->ma_need = MA_INODE | MA_HSM;
1379 if (ma->ma_lmm_size > 0) {
1380 ma->ma_need |= MA_LOV;
1381 /* Older clients may crash if they getattr overstriped
1384 if (!exp_connect_overstriping(exp) &&
1385 mdt_lmm_is_overstriping(ma->ma_lmm))
1386 RETURN(-EOPNOTSUPP);
1390 if (S_ISDIR(lu_object_attr(&next->mo_lu)) &&
1391 reqbody->mbo_valid & OBD_MD_FLDIREA &&
1392 lustre_msg_get_opc(req->rq_reqmsg) == MDS_GETATTR) {
1393 /* get default stripe info for this dir. */
1394 ma->ma_need |= MA_LOV_DEF;
1396 ma->ma_need |= ma_need;
1398 rc = mdt_attr_get_complex(info, o, ma);
1400 CDEBUG_LIMIT(rc == -ENOENT ? D_OTHER : D_ERROR,
1401 "%s: getattr error for "DFID": rc = %d\n",
1402 mdt_obd_name(info->mti_mdt),
1403 PFID(mdt_object_fid(o)), rc);
1407 /* if file is released, check if a restore is running */
1408 if (ma->ma_valid & MA_HSM) {
1409 repbody->mbo_valid |= OBD_MD_TSTATE;
1410 if ((ma->ma_hsm.mh_flags & HS_RELEASED) &&
1411 mdt_hsm_restore_is_running(info, mdt_object_fid(o)))
1412 repbody->mbo_t_state = MS_RESTORE;
1415 if (unlikely(!(ma->ma_valid & MA_INODE)))
1418 mdt_pack_attr2body(info, repbody, la, mdt_object_fid(o));
1420 if (mdt_body_has_lov(la, reqbody)) {
1421 u32 stripe_count = 1;
1423 if (ma->ma_valid & MA_LOV) {
1424 LASSERT(ma->ma_lmm_size);
1425 repbody->mbo_eadatasize = ma->ma_lmm_size;
1426 if (S_ISDIR(la->la_mode))
1427 repbody->mbo_valid |= OBD_MD_FLDIREA;
1429 repbody->mbo_valid |= OBD_MD_FLEASIZE;
1430 mdt_dump_lmm(D_INFO, ma->ma_lmm, repbody->mbo_valid);
1432 if (ma->ma_valid & MA_LMV) {
1433 struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1434 u32 magic = le32_to_cpu(lmv->lmv_magic);
1436 /* Return -ENOTSUPP for old client */
1437 if (!mdt_is_striped_client(req->rq_export))
1440 LASSERT(S_ISDIR(la->la_mode));
1441 mdt_dump_lmv(D_INFO, ma->ma_lmv);
1442 repbody->mbo_eadatasize = ma->ma_lmv_size;
1443 repbody->mbo_valid |= (OBD_MD_FLDIREA|OBD_MD_MEA);
1445 stripe_count = le32_to_cpu(lmv->lmv_stripe_count);
1446 if (magic == LMV_MAGIC_STRIPE && lmv_is_restriping(lmv))
1447 mdt_restripe_migrate_add(info, o);
1448 else if (magic == LMV_MAGIC_V1 &&
1449 lmv_is_restriping(lmv))
1450 mdt_restripe_update_add(info, o);
1452 if (ma->ma_valid & MA_LMV_DEF) {
1453 /* Return -ENOTSUPP for old client */
1454 if (!mdt_is_striped_client(req->rq_export))
1456 LASSERT(S_ISDIR(la->la_mode));
1458 * when ll_dir_getstripe() gets default LMV, it
1459 * checks mbo_eadatasize.
1461 if (!(ma->ma_valid & MA_LMV))
1462 repbody->mbo_eadatasize =
1463 ma->ma_default_lmv_size;
1464 repbody->mbo_valid |= (OBD_MD_FLDIREA |
1465 OBD_MD_DEFAULT_MEA);
1468 "dirent count %llu stripe count %u MDT count %d\n",
1469 ma->ma_attr.la_dirent_count, stripe_count,
1470 atomic_read(&mdt->mdt_mds_mds_conns) + 1);
1471 if (ma->ma_attr.la_dirent_count != LU_DIRENT_COUNT_UNSET &&
1472 ma->ma_attr.la_dirent_count >
1473 mdt->mdt_restriper.mdr_dir_split_count &&
1474 !fid_is_root(mdt_object_fid(o)) &&
1475 mdt->mdt_enable_dir_auto_split &&
1476 !o->mot_restriping &&
1477 stripe_count < atomic_read(&mdt->mdt_mds_mds_conns) + 1)
1478 mdt_auto_split_add(info, o);
1479 } else if (S_ISLNK(la->la_mode) &&
1480 reqbody->mbo_valid & OBD_MD_LINKNAME) {
1481 buffer->lb_buf = ma->ma_lmm;
1482 /* eadatasize from client includes NULL-terminator, so
1483 * there is no need to read it */
1484 buffer->lb_len = reqbody->mbo_eadatasize - 1;
1485 rc = mo_readlink(env, next, buffer);
1486 if (unlikely(rc <= 0)) {
1487 CERROR("%s: readlink failed for "DFID": rc = %d\n",
1488 mdt_obd_name(info->mti_mdt),
1489 PFID(mdt_object_fid(o)), rc);
1492 int print_limit = min_t(int, PAGE_SIZE - 128, rc);
1494 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_READLINK_EPROTO))
1496 repbody->mbo_valid |= OBD_MD_LINKNAME;
1497 /* we need to report back size with NULL-terminator
1498 * because client expects that */
1499 repbody->mbo_eadatasize = rc + 1;
1500 if (repbody->mbo_eadatasize != reqbody->mbo_eadatasize)
1501 CDEBUG(D_INODE, "%s: Read shorter symlink %d "
1502 "on "DFID ", expected %d\n",
1503 mdt_obd_name(info->mti_mdt),
1504 rc, PFID(mdt_object_fid(o)),
1505 reqbody->mbo_eadatasize - 1);
1506 /* NULL terminate */
1507 ((char *)ma->ma_lmm)[rc] = 0;
1509 /* If the total CDEBUG() size is larger than a page, it
1510 * will print a warning to the console, avoid this by
1511 * printing just the last part of the symlink. */
1512 CDEBUG(D_INODE, "symlink dest %s%.*s, len = %d\n",
1513 print_limit < rc ? "..." : "", print_limit,
1514 (char *)ma->ma_lmm + rc - print_limit, rc);
1519 if (reqbody->mbo_valid & OBD_MD_FLMODEASIZE) {
1520 repbody->mbo_max_mdsize = info->mti_mdt->mdt_max_mdsize;
1521 repbody->mbo_valid |= OBD_MD_FLMODEASIZE;
1522 CDEBUG(D_INODE, "changing the max MD size to %u\n",
1523 repbody->mbo_max_mdsize);
1526 #ifdef CONFIG_LUSTRE_FS_POSIX_ACL
1527 if ((exp_connect_flags(req->rq_export) & OBD_CONNECT_ACL) &&
1528 (reqbody->mbo_valid & OBD_MD_FLACL)) {
1529 struct lu_nodemap *nodemap = nodemap_get_from_exp(exp);
1530 if (IS_ERR(nodemap))
1531 RETURN(PTR_ERR(nodemap));
1533 rc = mdt_pack_acl2body(info, repbody, o, nodemap);
1534 nodemap_putref(nodemap);
1540 mdt_counter_incr(req, LPROC_MDT_GETATTR,
1541 ktime_us_delta(ktime_get(), kstart));
1546 static int mdt_getattr(struct tgt_session_info *tsi)
1548 struct mdt_thread_info *info = tsi2mdt_info(tsi);
1549 struct mdt_object *obj = info->mti_object;
1550 struct req_capsule *pill = info->mti_pill;
1551 struct mdt_body *reqbody;
1552 struct mdt_body *repbody;
1556 if (unlikely(info->mti_object == NULL))
1559 reqbody = req_capsule_client_get(pill, &RMF_MDT_BODY);
1561 LASSERT(lu_object_assert_exists(&obj->mot_obj));
1563 /* Special case for Data-on-MDT files to get data version */
1564 if (unlikely(reqbody->mbo_valid & OBD_MD_FLDATAVERSION)) {
1565 rc = mdt_data_version_get(tsi);
1569 /* Unlike intent case where we need to pre-fill out buffers early on
1570 * in intent policy for ldlm reasons, here we can have a much better
1571 * guess at EA size by just reading it from disk.
1572 * Exceptions are readdir and (missing) directory striping */
1574 if (reqbody->mbo_valid & OBD_MD_LINKNAME) {
1575 /* No easy way to know how long is the symlink, but it cannot
1576 * be more than PATH_MAX, so we allocate +1 */
1578 /* A special case for fs ROOT: getattr there might fetch
1579 * default EA for entire fs, not just for this dir!
1581 } else if (lu_fid_eq(mdt_object_fid(obj),
1582 &info->mti_mdt->mdt_md_root_fid) &&
1583 (reqbody->mbo_valid & OBD_MD_FLDIREA) &&
1584 (lustre_msg_get_opc(mdt_info_req(info)->rq_reqmsg) ==
1586 /* Should the default strping be bigger, mdt_fix_reply
1587 * will reallocate */
1588 rc = DEF_REP_MD_SIZE;
1590 /* Read the actual EA size from disk */
1591 rc = mdt_attr_get_eabuf_size(info, obj);
1595 GOTO(out, rc = err_serious(rc));
1597 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER, rc);
1599 /* Set ACL reply buffer size as LUSTRE_POSIX_ACL_MAX_SIZE_OLD
1600 * by default. If the target object has more ACL entries, then
1601 * enlarge the buffer when necessary. */
1602 req_capsule_set_size(pill, &RMF_ACL, RCL_SERVER,
1603 LUSTRE_POSIX_ACL_MAX_SIZE_OLD);
1605 rc = req_capsule_server_pack(pill);
1606 if (unlikely(rc != 0))
1607 GOTO(out, rc = err_serious(rc));
1609 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
1610 LASSERT(repbody != NULL);
1611 repbody->mbo_eadatasize = 0;
1612 repbody->mbo_aclsize = 0;
1614 rc = mdt_check_ucred(info);
1616 GOTO(out_shrink, rc);
1618 info->mti_cross_ref = !!(reqbody->mbo_valid & OBD_MD_FLCROSSREF);
1620 rc = mdt_getattr_internal(info, obj, 0);
1623 mdt_client_compatibility(info);
1624 rc2 = mdt_fix_reply(info);
1628 mdt_thread_info_fini(info);
1633 * Handler of layout intent RPC requiring the layout modification
1635 * \param[in] info thread environment
1636 * \param[in] obj object
1637 * \param[out] lhc object ldlm lock handle
1638 * \param[in] layout layout change descriptor
1640 * \retval 0 on success
1641 * \retval < 0 error code
1643 int mdt_layout_change(struct mdt_thread_info *info, struct mdt_object *obj,
1644 struct mdt_lock_handle *lhc,
1645 struct md_layout_change *layout)
1651 if (!mdt_object_exists(obj))
1654 if (!S_ISREG(lu_object_attr(&obj->mot_obj)))
1657 rc = mo_permission(info->mti_env, NULL, mdt_object_child(obj), NULL,
1662 rc = mdt_check_resent_lock(info, obj, lhc);
1668 __u64 lockpart = MDS_INODELOCK_LAYOUT;
1670 /* take layout lock to prepare layout change */
1671 if (layout->mlc_opc == MD_LAYOUT_WRITE)
1672 lockpart |= MDS_INODELOCK_UPDATE;
1674 mdt_lock_handle_init(lhc);
1675 mdt_lock_reg_init(lhc, LCK_EX);
1676 rc = mdt_reint_object_lock(info, obj, lhc, lockpart, false);
1681 mutex_lock(&obj->mot_som_mutex);
1682 rc = mo_layout_change(info->mti_env, mdt_object_child(obj), layout);
1683 mutex_unlock(&obj->mot_som_mutex);
1686 mdt_object_unlock(info, obj, lhc, 1);
1692 * Exchange MOF_LOV_CREATED flags between two objects after a
1693 * layout swap. No assumption is made on whether o1 or o2 have
1694 * created objects or not.
1696 * \param[in,out] o1 First swap layout object
1697 * \param[in,out] o2 Second swap layout object
1699 static void mdt_swap_lov_flag(struct mdt_object *o1, struct mdt_object *o2)
1701 unsigned int o1_lov_created = o1->mot_lov_created;
1703 mutex_lock(&o1->mot_lov_mutex);
1704 mutex_lock(&o2->mot_lov_mutex);
1706 o1->mot_lov_created = o2->mot_lov_created;
1707 o2->mot_lov_created = o1_lov_created;
1709 mutex_unlock(&o2->mot_lov_mutex);
1710 mutex_unlock(&o1->mot_lov_mutex);
1713 static int mdt_swap_layouts(struct tgt_session_info *tsi)
1715 struct mdt_thread_info *info;
1716 struct ptlrpc_request *req = tgt_ses_req(tsi);
1717 struct obd_export *exp = req->rq_export;
1718 struct mdt_object *o1, *o2, *o;
1719 struct mdt_lock_handle *lh1, *lh2;
1720 struct mdc_swap_layouts *msl;
1724 /* client does not support layout lock, so layout swaping
1726 * FIXME: there is a problem for old clients which don't support
1727 * layout lock yet. If those clients have already opened the file
1728 * they won't be notified at all so that old layout may still be
1729 * used to do IO. This can be fixed after file release is landed by
1730 * doing exclusive open and taking full EX ibits lock. - Jinshan */
1731 if (!exp_connect_layout(exp))
1732 RETURN(-EOPNOTSUPP);
1734 info = tsi2mdt_info(tsi);
1735 if (unlikely(info->mti_object == NULL))
1738 if (info->mti_dlm_req != NULL)
1739 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1741 o1 = info->mti_object;
1742 o = o2 = mdt_object_find(info->mti_env, info->mti_mdt,
1743 &info->mti_body->mbo_fid2);
1745 GOTO(out, rc = PTR_ERR(o));
1747 if (mdt_object_remote(o) || !mdt_object_exists(o)) /* remote object */
1748 GOTO(put, rc = -ENOENT);
1750 rc = lu_fid_cmp(&info->mti_body->mbo_fid1, &info->mti_body->mbo_fid2);
1751 if (unlikely(rc == 0)) /* same file, you kidding me? no-op. */
1757 /* permission check. Make sure the calling process having permission
1758 * to write both files. */
1759 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o1), NULL,
1764 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o2), NULL,
1769 msl = req_capsule_client_get(info->mti_pill, &RMF_SWAP_LAYOUTS);
1771 GOTO(put, rc = -EPROTO);
1773 lh1 = &info->mti_lh[MDT_LH_NEW];
1774 mdt_lock_reg_init(lh1, LCK_EX);
1775 lh2 = &info->mti_lh[MDT_LH_OLD];
1776 mdt_lock_reg_init(lh2, LCK_EX);
1778 rc = mdt_object_lock(info, o1, lh1, MDS_INODELOCK_LAYOUT |
1779 MDS_INODELOCK_XATTR);
1783 rc = mdt_object_lock(info, o2, lh2, MDS_INODELOCK_LAYOUT |
1784 MDS_INODELOCK_XATTR);
1788 rc = mo_swap_layouts(info->mti_env, mdt_object_child(o1),
1789 mdt_object_child(o2), msl->msl_flags);
1793 mdt_swap_lov_flag(o1, o2);
1796 mdt_object_unlock(info, o2, lh2, rc);
1798 mdt_object_unlock(info, o1, lh1, rc);
1800 mdt_object_put(info->mti_env, o);
1802 mdt_thread_info_fini(info);
1806 static int mdt_raw_lookup(struct mdt_thread_info *info,
1807 struct mdt_object *parent,
1808 const struct lu_name *lname,
1809 struct ldlm_reply *ldlm_rep)
1811 struct lu_fid *child_fid = &info->mti_tmp_fid1;
1815 LASSERT(!info->mti_cross_ref);
1817 /* Only got the fid of this obj by name */
1818 fid_zero(child_fid);
1819 rc = mdo_lookup(info->mti_env, mdt_object_child(info->mti_object),
1820 lname, child_fid, &info->mti_spec);
1822 struct mdt_body *repbody;
1824 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1825 repbody->mbo_fid1 = *child_fid;
1826 repbody->mbo_valid = OBD_MD_FLID;
1827 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1828 } else if (rc == -ENOENT) {
1829 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1836 * UPDATE lock should be taken against parent, and be released before exit;
1837 * child_bits lock should be taken against child, and be returned back:
1838 * (1)normal request should release the child lock;
1839 * (2)intent request will grant the lock to client.
1841 static int mdt_getattr_name_lock(struct mdt_thread_info *info,
1842 struct mdt_lock_handle *lhc,
1844 struct ldlm_reply *ldlm_rep)
1846 struct ptlrpc_request *req = mdt_info_req(info);
1847 struct mdt_body *reqbody = NULL;
1848 struct mdt_object *parent = info->mti_object;
1849 struct mdt_object *child = NULL;
1850 struct lu_fid *child_fid = &info->mti_tmp_fid1;
1851 struct lu_name *lname = NULL;
1852 struct mdt_lock_handle *lhp = NULL;
1853 struct ldlm_lock *lock;
1854 struct req_capsule *pill = info->mti_pill;
1862 is_resent = lustre_handle_is_used(&lhc->mlh_reg_lh);
1863 LASSERT(ergo(is_resent,
1864 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT));
1869 if (info->mti_cross_ref) {
1870 /* Only getattr on the child. Parent is on another node. */
1871 mdt_set_disposition(info, ldlm_rep,
1872 DISP_LOOKUP_EXECD | DISP_LOOKUP_POS);
1874 CDEBUG(D_INODE, "partial getattr_name child_fid = "DFID", "
1876 PFID(mdt_object_fid(child)), ldlm_rep);
1878 rc = mdt_check_resent_lock(info, child, lhc);
1881 } else if (rc > 0) {
1882 mdt_lock_handle_init(lhc);
1883 mdt_lock_reg_init(lhc, LCK_PR);
1886 * Object's name entry is on another MDS, it will
1887 * request PERM lock only because LOOKUP lock is owned
1888 * by the MDS where name entry resides.
1890 * TODO: it should try layout lock too. - Jinshan
1892 child_bits &= ~(MDS_INODELOCK_LOOKUP |
1893 MDS_INODELOCK_LAYOUT);
1894 child_bits |= MDS_INODELOCK_PERM;
1896 rc = mdt_object_lock(info, child, lhc, child_bits);
1901 /* Finally, we can get attr for child. */
1902 if (!mdt_object_exists(child)) {
1903 LU_OBJECT_DEBUG(D_INFO, info->mti_env,
1905 "remote object doesn't exist.");
1906 mdt_object_unlock(info, child, lhc, 1);
1910 rc = mdt_getattr_internal(info, child, 0);
1911 if (unlikely(rc != 0)) {
1912 mdt_object_unlock(info, child, lhc, 1);
1916 rc = mdt_pack_secctx_in_reply(info, child);
1918 mdt_object_unlock(info, child, lhc, 1);
1922 rc = mdt_pack_encctx_in_reply(info, child);
1924 mdt_object_unlock(info, child, lhc, 1);
1928 lname = &info->mti_name;
1929 mdt_name_unpack(pill, &RMF_NAME, lname, MNF_FIX_ANON);
1931 if (lu_name_is_valid(lname)) {
1932 if (mdt_object_remote(parent)) {
1933 CERROR("%s: parent "DFID" is on remote target\n",
1934 mdt_obd_name(info->mti_mdt),
1935 PFID(mdt_object_fid(parent)));
1939 CDEBUG(D_INODE, "getattr with lock for "DFID"/"DNAME", "
1940 "ldlm_rep = %p\n", PFID(mdt_object_fid(parent)),
1941 PNAME(lname), ldlm_rep);
1943 reqbody = req_capsule_client_get(pill, &RMF_MDT_BODY);
1944 if (unlikely(reqbody == NULL))
1945 RETURN(err_serious(-EPROTO));
1947 *child_fid = reqbody->mbo_fid2;
1948 if (unlikely(!fid_is_sane(child_fid)))
1949 RETURN(err_serious(-EINVAL));
1951 if (lu_fid_eq(mdt_object_fid(parent), child_fid)) {
1952 mdt_object_get(info->mti_env, parent);
1955 child = mdt_object_find(info->mti_env, info->mti_mdt,
1958 RETURN(PTR_ERR(child));
1961 if (mdt_object_remote(child)) {
1962 CERROR("%s: child "DFID" is on remote target\n",
1963 mdt_obd_name(info->mti_mdt),
1964 PFID(mdt_object_fid(child)));
1965 GOTO(out_child, rc = -EPROTO);
1968 /* don't fetch LOOKUP lock if it's remote object */
1969 rc = mdt_is_remote_object(info, parent, child);
1971 GOTO(out_child, rc);
1973 child_bits &= ~MDS_INODELOCK_LOOKUP;
1975 CDEBUG(D_INODE, "getattr with lock for "DFID"/"DFID", "
1977 PFID(mdt_object_fid(parent)),
1978 PFID(&reqbody->mbo_fid2), ldlm_rep);
1981 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_EXECD);
1983 if (unlikely(!mdt_object_exists(parent)) && lu_name_is_valid(lname)) {
1984 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
1986 "Parent doesn't exist!");
1987 GOTO(out_child, rc = -ESTALE);
1990 if (lu_name_is_valid(lname)) {
1991 /* Always allow to lookup ".." */
1992 if (unlikely(lname->ln_namelen == 2 &&
1993 lname->ln_name[0] == '.' &&
1994 lname->ln_name[1] == '.'))
1995 info->mti_spec.sp_permitted = 1;
1997 if (info->mti_body->mbo_valid == OBD_MD_FLID) {
1998 rc = mdt_raw_lookup(info, parent, lname, ldlm_rep);
2003 /* step 1: lock parent only if parent is a directory */
2004 if (S_ISDIR(lu_object_attr(&parent->mot_obj))) {
2005 lhp = &info->mti_lh[MDT_LH_PARENT];
2006 mdt_lock_pdo_init(lhp, LCK_PR, lname);
2007 rc = mdt_object_lock(info, parent, lhp,
2008 MDS_INODELOCK_UPDATE);
2009 if (unlikely(rc != 0))
2013 /* step 2: lookup child's fid by name */
2014 fid_zero(child_fid);
2015 rc = mdo_lookup(info->mti_env, mdt_object_child(parent), lname,
2016 child_fid, &info->mti_spec);
2018 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
2021 GOTO(unlock_parent, rc);
2023 child = mdt_object_find(info->mti_env, info->mti_mdt,
2025 if (unlikely(IS_ERR(child)))
2026 GOTO(unlock_parent, rc = PTR_ERR(child));
2029 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
2031 /* step 3: lock child regardless if it is local or remote. */
2034 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RESEND, obd_timeout * 2);
2035 if (!mdt_object_exists(child)) {
2036 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2038 "Object doesn't exist!");
2039 GOTO(out_child, rc = -ENOENT);
2042 rc = mdt_check_resent_lock(info, child, lhc);
2044 GOTO(out_child, rc);
2045 } else if (rc > 0) {
2046 mdt_lock_handle_init(lhc);
2047 mdt_lock_reg_init(lhc, LCK_PR);
2049 if (!(child_bits & MDS_INODELOCK_UPDATE) &&
2050 !mdt_object_remote(child)) {
2051 struct md_attr *ma = &info->mti_attr;
2054 ma->ma_need = MA_INODE;
2055 rc = mdt_attr_get_complex(info, child, ma);
2056 if (unlikely(rc != 0))
2057 GOTO(out_child, rc);
2059 /* If the file has not been changed for some time, we
2060 * return not only a LOOKUP lock, but also an UPDATE
2061 * lock and this might save us RPC on later STAT. For
2062 * directories, it also let negative dentry cache start
2063 * working for this dir. */
2064 if (ma->ma_valid & MA_INODE &&
2065 ma->ma_attr.la_valid & LA_CTIME &&
2066 info->mti_mdt->mdt_namespace->ns_ctime_age_limit +
2067 ma->ma_attr.la_ctime < ktime_get_real_seconds())
2068 child_bits |= MDS_INODELOCK_UPDATE;
2071 /* layout lock must be granted in a best-effort way
2072 * for IT operations */
2073 LASSERT(!(child_bits & MDS_INODELOCK_LAYOUT));
2074 if (S_ISREG(lu_object_attr(&child->mot_obj)) &&
2075 !mdt_object_remote(child) && ldlm_rep != NULL) {
2076 if (!OBD_FAIL_CHECK(OBD_FAIL_MDS_NO_LL_GETATTR) &&
2077 exp_connect_layout(info->mti_exp)) {
2078 /* try to grant layout lock for regular file. */
2079 try_bits = MDS_INODELOCK_LAYOUT;
2081 /* Acquire DOM lock in advance for data-on-mdt file */
2082 if (child != parent)
2083 try_bits |= MDS_INODELOCK_DOM;
2086 if (try_bits != 0) {
2087 /* try layout lock, it may fail to be granted due to
2088 * contention at LOOKUP or UPDATE */
2089 rc = mdt_object_lock_try(info, child, lhc, &child_bits,
2091 if (child_bits & MDS_INODELOCK_LAYOUT)
2094 /* Do not enqueue the UPDATE lock from MDT(cross-MDT),
2095 * client will enqueue the lock to the remote MDT */
2096 if (mdt_object_remote(child))
2097 child_bits &= ~MDS_INODELOCK_UPDATE;
2098 rc = mdt_object_lock(info, child, lhc, child_bits);
2100 if (unlikely(rc != 0))
2101 GOTO(out_child, rc);
2104 /* finally, we can get attr for child. */
2105 rc = mdt_getattr_internal(info, child, ma_need);
2106 if (unlikely(rc != 0)) {
2107 mdt_object_unlock(info, child, lhc, 1);
2108 GOTO(out_child, rc);
2111 rc = mdt_pack_secctx_in_reply(info, child);
2113 mdt_object_unlock(info, child, lhc, 1);
2114 GOTO(out_child, rc);
2117 rc = mdt_pack_encctx_in_reply(info, child);
2119 mdt_object_unlock(info, child, lhc, 1);
2120 GOTO(out_child, rc);
2123 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
2125 /* Debugging code. */
2126 LDLM_DEBUG(lock, "Returning lock to client");
2127 LASSERTF(fid_res_name_eq(mdt_object_fid(child),
2128 &lock->l_resource->lr_name),
2129 "Lock res_id: "DLDLMRES", fid: "DFID"\n",
2130 PLDLMRES(lock->l_resource),
2131 PFID(mdt_object_fid(child)));
2133 if (S_ISREG(lu_object_attr(&child->mot_obj)) &&
2134 !mdt_object_remote(child) && child != parent) {
2135 mdt_object_put(info->mti_env, child);
2136 rc = mdt_pack_size2body(info, child_fid,
2138 if (rc != 0 && child_bits & MDS_INODELOCK_DOM) {
2139 /* DOM lock was taken in advance but this is
2140 * not DoM file. Drop the lock.
2142 lock_res_and_lock(lock);
2143 ldlm_inodebits_drop(lock, MDS_INODELOCK_DOM);
2144 unlock_res_and_lock(lock);
2146 LDLM_LOCK_PUT(lock);
2147 GOTO(unlock_parent, rc = 0);
2149 LDLM_LOCK_PUT(lock);
2155 mdt_object_put(info->mti_env, child);
2158 mdt_object_unlock(info, parent, lhp, 1);
2162 /* normal handler: should release the child lock */
2163 static int mdt_getattr_name(struct tgt_session_info *tsi)
2165 struct mdt_thread_info *info = tsi2mdt_info(tsi);
2166 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_CHILD];
2167 struct mdt_body *reqbody;
2168 struct mdt_body *repbody;
2173 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
2174 LASSERT(reqbody != NULL);
2175 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2176 LASSERT(repbody != NULL);
2178 info->mti_cross_ref = !!(reqbody->mbo_valid & OBD_MD_FLCROSSREF);
2179 repbody->mbo_eadatasize = 0;
2180 repbody->mbo_aclsize = 0;
2182 rc = mdt_init_ucred(info, reqbody);
2184 GOTO(out_shrink, rc);
2186 rc = mdt_getattr_name_lock(info, lhc, MDS_INODELOCK_UPDATE, NULL);
2187 if (lustre_handle_is_used(&lhc->mlh_reg_lh)) {
2188 ldlm_lock_decref(&lhc->mlh_reg_lh, lhc->mlh_reg_mode);
2189 lhc->mlh_reg_lh.cookie = 0;
2191 mdt_exit_ucred(info);
2194 mdt_client_compatibility(info);
2195 rc2 = mdt_fix_reply(info);
2198 mdt_thread_info_fini(info);
2202 static int mdt_rmfid_unlink(struct mdt_thread_info *info,
2203 const struct lu_fid *pfid,
2204 const struct lu_name *name,
2205 struct mdt_object *obj, s64 ctime)
2207 struct lu_fid *child_fid = &info->mti_tmp_fid1;
2208 struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
2209 struct mdt_device *mdt = info->mti_mdt;
2210 struct md_attr *ma = &info->mti_attr;
2211 struct mdt_lock_handle *parent_lh;
2212 struct mdt_lock_handle *child_lh;
2213 struct mdt_object *pobj;
2214 bool cos_incompat = false;
2218 pobj = mdt_object_find(info->mti_env, mdt, pfid);
2220 GOTO(out, rc = PTR_ERR(pobj));
2222 parent_lh = &info->mti_lh[MDT_LH_PARENT];
2223 mdt_lock_pdo_init(parent_lh, LCK_PW, name);
2224 rc = mdt_object_lock(info, pobj, parent_lh, MDS_INODELOCK_UPDATE);
2226 GOTO(put_parent, rc);
2228 if (mdt_object_remote(pobj))
2229 cos_incompat = true;
2231 rc = mdo_lookup(info->mti_env, mdt_object_child(pobj),
2232 name, child_fid, &info->mti_spec);
2234 GOTO(unlock_parent, rc);
2236 if (!lu_fid_eq(child_fid, mdt_object_fid(obj)))
2237 GOTO(unlock_parent, rc = -EREMCHG);
2239 child_lh = &info->mti_lh[MDT_LH_CHILD];
2240 mdt_lock_reg_init(child_lh, LCK_EX);
2241 rc = mdt_reint_striped_lock(info, obj, child_lh,
2242 MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE,
2243 einfo, cos_incompat);
2245 GOTO(unlock_parent, rc);
2247 if (atomic_read(&obj->mot_open_count)) {
2248 CDEBUG(D_OTHER, "object "DFID" open, skip\n",
2249 PFID(mdt_object_fid(obj)));
2250 GOTO(unlock_child, rc = -EBUSY);
2254 ma->ma_valid = MA_INODE;
2255 ma->ma_attr.la_valid = LA_CTIME;
2256 ma->ma_attr.la_ctime = ctime;
2258 mutex_lock(&obj->mot_lov_mutex);
2260 rc = mdo_unlink(info->mti_env, mdt_object_child(pobj),
2261 mdt_object_child(obj), name, ma, 0);
2263 mutex_unlock(&obj->mot_lov_mutex);
2266 mdt_reint_striped_unlock(info, obj, child_lh, einfo, 1);
2268 mdt_object_unlock(info, pobj, parent_lh, 1);
2270 mdt_object_put(info->mti_env, pobj);
2275 static int mdt_rmfid_check_permission(struct mdt_thread_info *info,
2276 struct mdt_object *obj)
2278 struct lu_ucred *uc = lu_ucred(info->mti_env);
2279 struct md_attr *ma = &info->mti_attr;
2280 struct lu_attr *la = &ma->ma_attr;
2284 ma->ma_need = MA_INODE;
2285 rc = mo_attr_get(info->mti_env, mdt_object_child(obj), ma);
2289 if (la->la_flags & LUSTRE_IMMUTABLE_FL)
2292 if (md_capable(uc, CAP_DAC_OVERRIDE))
2294 if (uc->uc_fsuid == la->la_uid) {
2295 if ((la->la_mode & S_IWUSR) == 0)
2297 } else if (uc->uc_fsgid == la->la_gid) {
2298 if ((la->la_mode & S_IWGRP) == 0)
2300 } else if ((la->la_mode & S_IWOTH) == 0) {
2308 static int mdt_rmfid_one(struct mdt_thread_info *info, struct lu_fid *fid,
2311 struct mdt_device *mdt = info->mti_mdt;
2312 struct mdt_object *obj = NULL;
2313 struct linkea_data ldata = { NULL };
2314 struct lu_buf *buf = &info->mti_big_buf;
2315 struct lu_name *name = &info->mti_name;
2316 struct lu_fid *pfid = &info->mti_tmp_fid1;
2317 struct link_ea_header *leh;
2318 struct link_ea_entry *lee;
2319 int reclen, count, rc = 0;
2322 if (!fid_is_sane(fid))
2323 GOTO(out, rc = -EINVAL);
2325 if (!fid_is_namespace_visible(fid))
2326 GOTO(out, rc = -EINVAL);
2328 obj = mdt_object_find(info->mti_env, mdt, fid);
2330 GOTO(out, rc = PTR_ERR(obj));
2332 if (mdt_object_remote(obj))
2333 GOTO(out, rc = -EREMOTE);
2334 if (!mdt_object_exists(obj) || lu_object_is_dying(&obj->mot_header))
2335 GOTO(out, rc = -ENOENT);
2337 rc = mdt_rmfid_check_permission(info, obj);
2342 buf = lu_buf_check_and_alloc(buf, PATH_MAX);
2344 GOTO(out, rc = -ENOMEM);
2347 rc = mdt_links_read(info, obj, &ldata);
2352 lee = (struct link_ea_entry *)(leh + 1);
2353 for (count = 0; count < leh->leh_reccount; count++) {
2354 /* remove every hardlink */
2355 linkea_entry_unpack(lee, &reclen, name, pfid);
2356 lee = (struct link_ea_entry *) ((char *)lee + reclen);
2357 rc = mdt_rmfid_unlink(info, pfid, name, obj, ctime);
2363 if (obj && !IS_ERR(obj))
2364 mdt_object_put(info->mti_env, obj);
2365 if (info->mti_big_buf.lb_buf)
2366 lu_buf_free(&info->mti_big_buf);
2371 static int mdt_rmfid(struct tgt_session_info *tsi)
2373 struct mdt_thread_info *mti = tsi2mdt_info(tsi);
2374 struct mdt_body *reqbody;
2375 struct lu_fid *fids, *rfids;
2381 reqbody = req_capsule_client_get(tsi->tsi_pill, &RMF_MDT_BODY);
2382 if (reqbody == NULL)
2384 bufsize = req_capsule_get_size(tsi->tsi_pill, &RMF_FID_ARRAY,
2386 nr = bufsize / sizeof(struct lu_fid);
2387 if (nr * sizeof(struct lu_fid) != bufsize)
2389 req_capsule_set_size(tsi->tsi_pill, &RMF_RCS,
2390 RCL_SERVER, nr * sizeof(__u32));
2391 req_capsule_set_size(tsi->tsi_pill, &RMF_FID_ARRAY,
2392 RCL_SERVER, nr * sizeof(struct lu_fid));
2393 rc = req_capsule_server_pack(tsi->tsi_pill);
2395 GOTO(out, rc = err_serious(rc));
2396 fids = req_capsule_client_get(tsi->tsi_pill, &RMF_FID_ARRAY);
2399 rcs = req_capsule_server_get(tsi->tsi_pill, &RMF_RCS);
2401 rfids = req_capsule_server_get(tsi->tsi_pill, &RMF_FID_ARRAY);
2404 mdt_init_ucred(mti, reqbody);
2405 for (i = 0; i < nr; i++) {
2407 rcs[i] = mdt_rmfid_one(mti, fids + i, reqbody->mbo_ctime);
2409 mdt_exit_ucred(mti);
2415 static int mdt_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2416 void *karg, void __user *uarg);
2418 static int mdt_set_info(struct tgt_session_info *tsi)
2420 struct ptlrpc_request *req = tgt_ses_req(tsi);
2423 int keylen, vallen, rc = 0;
2427 key = req_capsule_client_get(tsi->tsi_pill, &RMF_SETINFO_KEY);
2429 DEBUG_REQ(D_HA, req, "no set_info key");
2430 RETURN(err_serious(-EFAULT));
2433 keylen = req_capsule_get_size(tsi->tsi_pill, &RMF_SETINFO_KEY,
2436 val = req_capsule_client_get(tsi->tsi_pill, &RMF_SETINFO_VAL);
2438 DEBUG_REQ(D_HA, req, "no set_info val");
2439 RETURN(err_serious(-EFAULT));
2442 vallen = req_capsule_get_size(tsi->tsi_pill, &RMF_SETINFO_VAL,
2445 /* Swab any part of val you need to here */
2446 if (KEY_IS(KEY_READ_ONLY)) {
2447 spin_lock(&req->rq_export->exp_lock);
2449 *exp_connect_flags_ptr(req->rq_export) |=
2452 *exp_connect_flags_ptr(req->rq_export) &=
2453 ~OBD_CONNECT_RDONLY;
2454 spin_unlock(&req->rq_export->exp_lock);
2455 } else if (KEY_IS(KEY_CHANGELOG_CLEAR)) {
2456 struct changelog_setinfo *cs = val;
2458 if (vallen != sizeof(*cs)) {
2459 CERROR("%s: bad changelog_clear setinfo size %d\n",
2460 tgt_name(tsi->tsi_tgt), vallen);
2463 if (ptlrpc_req_need_swab(req)) {
2464 __swab64s(&cs->cs_recno);
2465 __swab32s(&cs->cs_id);
2468 if (!mdt_is_rootadmin(tsi2mdt_info(tsi)))
2470 rc = mdt_iocontrol(OBD_IOC_CHANGELOG_CLEAR, req->rq_export,
2472 } else if (KEY_IS(KEY_EVICT_BY_NID)) {
2474 obd_export_evict_by_nid(req->rq_export->exp_obd, val);
2481 static int mdt_readpage(struct tgt_session_info *tsi)
2483 struct mdt_thread_info *info = mdt_th_info(tsi->tsi_env);
2484 struct mdt_object *object = mdt_obj(tsi->tsi_corpus);
2485 struct lu_rdpg *rdpg = &info->mti_u.rdpg.mti_rdpg;
2486 const struct mdt_body *reqbody = tsi->tsi_mdt_body;
2487 struct mdt_body *repbody;
2493 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK))
2494 RETURN(err_serious(-ENOMEM));
2496 repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_MDT_BODY);
2497 if (repbody == NULL || reqbody == NULL)
2498 RETURN(err_serious(-EFAULT));
2501 * prepare @rdpg before calling lower layers and transfer itself. Here
2502 * reqbody->size contains offset of where to start to read and
2503 * reqbody->nlink contains number bytes to read.
2505 rdpg->rp_hash = reqbody->mbo_size;
2506 if (rdpg->rp_hash != reqbody->mbo_size) {
2507 CERROR("Invalid hash: %#llx != %#llx\n",
2508 rdpg->rp_hash, reqbody->mbo_size);
2512 rdpg->rp_attrs = reqbody->mbo_mode;
2513 if (exp_connect_flags(tsi->tsi_exp) & OBD_CONNECT_64BITHASH)
2514 rdpg->rp_attrs |= LUDA_64BITHASH;
2515 rdpg->rp_count = min_t(unsigned int, reqbody->mbo_nlink,
2516 exp_max_brw_size(tsi->tsi_exp));
2517 rdpg->rp_npages = (rdpg->rp_count + PAGE_SIZE - 1) >>
2519 OBD_ALLOC_PTR_ARRAY_LARGE(rdpg->rp_pages, rdpg->rp_npages);
2520 if (rdpg->rp_pages == NULL)
2523 for (i = 0; i < rdpg->rp_npages; ++i) {
2524 rdpg->rp_pages[i] = alloc_page(GFP_NOFS);
2525 if (rdpg->rp_pages[i] == NULL)
2526 GOTO(free_rdpg, rc = -ENOMEM);
2529 /* call lower layers to fill allocated pages with directory data */
2530 rc = mo_readpage(tsi->tsi_env, mdt_object_child(object), rdpg);
2532 GOTO(free_rdpg, rc);
2534 /* send pages to client */
2535 rc = tgt_sendpage(tsi, rdpg, rc);
2540 for (i = 0; i < rdpg->rp_npages; i++)
2541 if (rdpg->rp_pages[i] != NULL)
2542 __free_page(rdpg->rp_pages[i]);
2543 OBD_FREE_PTR_ARRAY_LARGE(rdpg->rp_pages, rdpg->rp_npages);
2545 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE))
2551 static int mdt_fix_attr_ucred(struct mdt_thread_info *info, __u32 op)
2553 struct lu_ucred *uc = mdt_ucred_check(info);
2554 struct lu_attr *attr = &info->mti_attr.ma_attr;
2559 if (op != REINT_SETATTR) {
2560 if ((attr->la_valid & LA_UID) && (attr->la_uid != -1))
2561 attr->la_uid = uc->uc_fsuid;
2562 /* for S_ISGID, inherit gid from his parent, such work will be
2563 * done in cmm/mdd layer, here set all cases as uc->uc_fsgid. */
2564 if ((attr->la_valid & LA_GID) && (attr->la_gid != -1))
2565 attr->la_gid = uc->uc_fsgid;
2571 static inline bool mdt_is_readonly_open(struct mdt_thread_info *info, __u32 op)
2573 return op == REINT_OPEN &&
2574 !(info->mti_spec.sp_cr_flags & (MDS_FMODE_WRITE | MDS_OPEN_CREAT));
2577 static void mdt_preset_secctx_size(struct mdt_thread_info *info)
2579 struct req_capsule *pill = info->mti_pill;
2581 if (req_capsule_has_field(pill, &RMF_FILE_SECCTX,
2583 req_capsule_has_field(pill, &RMF_FILE_SECCTX_NAME,
2585 if (req_capsule_get_size(pill, &RMF_FILE_SECCTX_NAME,
2587 /* pre-set size in server part with max size */
2588 req_capsule_set_size(pill, &RMF_FILE_SECCTX,
2590 OBD_MAX_DEFAULT_EA_SIZE);
2592 req_capsule_set_size(pill, &RMF_FILE_SECCTX,
2597 static void mdt_preset_encctx_size(struct mdt_thread_info *info)
2599 struct req_capsule *pill = info->mti_pill;
2601 if (req_capsule_has_field(pill, &RMF_FILE_ENCCTX,
2603 /* pre-set size in server part with max size */
2604 req_capsule_set_size(pill, &RMF_FILE_ENCCTX,
2606 info->mti_mdt->mdt_max_mdsize);
2609 static int mdt_reint_internal(struct mdt_thread_info *info,
2610 struct mdt_lock_handle *lhc,
2613 struct req_capsule *pill = info->mti_pill;
2614 struct mdt_body *repbody;
2619 rc = mdt_reint_unpack(info, op);
2621 CERROR("Can't unpack reint, rc %d\n", rc);
2622 RETURN(err_serious(rc));
2626 /* check if the file system is set to readonly. O_RDONLY open
2627 * is still allowed even the file system is set to readonly mode */
2628 if (mdt_rdonly(info->mti_exp) && !mdt_is_readonly_open(info, op))
2629 RETURN(err_serious(-EROFS));
2631 /* for replay (no_create) lmm is not needed, client has it already */
2632 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
2633 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER,
2636 /* llog cookies are always 0, the field is kept for compatibility */
2637 if (req_capsule_has_field(pill, &RMF_LOGCOOKIES, RCL_SERVER))
2638 req_capsule_set_size(pill, &RMF_LOGCOOKIES, RCL_SERVER, 0);
2640 /* Set ACL reply buffer size as LUSTRE_POSIX_ACL_MAX_SIZE_OLD
2641 * by default. If the target object has more ACL entries, then
2642 * enlarge the buffer when necessary. */
2643 if (req_capsule_has_field(pill, &RMF_ACL, RCL_SERVER))
2644 req_capsule_set_size(pill, &RMF_ACL, RCL_SERVER,
2645 LUSTRE_POSIX_ACL_MAX_SIZE_OLD);
2647 mdt_preset_secctx_size(info);
2648 mdt_preset_encctx_size(info);
2650 rc = req_capsule_server_pack(pill);
2652 CERROR("Can't pack response, rc %d\n", rc);
2653 RETURN(err_serious(rc));
2656 if (req_capsule_has_field(pill, &RMF_MDT_BODY, RCL_SERVER)) {
2657 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
2659 repbody->mbo_eadatasize = 0;
2660 repbody->mbo_aclsize = 0;
2663 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_REINT_DELAY, 10);
2665 /* for replay no cookkie / lmm need, because client have this already */
2666 if (info->mti_spec.no_create)
2667 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
2668 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER, 0);
2670 rc = mdt_init_ucred_reint(info);
2672 GOTO(out_shrink, rc);
2674 rc = mdt_fix_attr_ucred(info, op);
2676 GOTO(out_ucred, rc = err_serious(rc));
2678 rc = mdt_check_resent(info, mdt_reconstruct, lhc);
2680 GOTO(out_ucred, rc);
2681 } else if (rc == 1) {
2682 DEBUG_REQ(D_INODE, mdt_info_req(info), "resent opt");
2683 rc = lustre_msg_get_status(mdt_info_req(info)->rq_repmsg);
2684 GOTO(out_ucred, rc);
2686 rc = mdt_reint_rec(info, lhc);
2689 mdt_exit_ucred(info);
2691 mdt_client_compatibility(info);
2693 rc2 = mdt_fix_reply(info);
2698 * Data-on-MDT optimization - read data along with OPEN and return it
2699 * in reply when possible.
2701 if (rc == 0 && op == REINT_OPEN && !req_is_replay(pill->rc_req))
2702 rc = mdt_dom_read_on_open(info, info->mti_mdt,
2708 static long mdt_reint_opcode(struct ptlrpc_request *req,
2709 const struct req_format **fmt)
2711 struct mdt_device *mdt;
2712 struct mdt_rec_reint *rec;
2715 rec = req_capsule_client_get(&req->rq_pill, &RMF_REC_REINT);
2717 opc = rec->rr_opcode;
2718 DEBUG_REQ(D_INODE, req, "reint opt = %ld", opc);
2719 if (opc < REINT_MAX && fmt[opc] != NULL)
2720 req_capsule_extend(&req->rq_pill, fmt[opc]);
2722 mdt = mdt_exp2dev(req->rq_export);
2723 CERROR("%s: Unsupported opcode '%ld' from client '%s':"
2724 " rc = %d\n", req->rq_export->exp_obd->obd_name,
2725 opc, mdt->mdt_ldlm_client->cli_name, -EFAULT);
2726 opc = err_serious(-EFAULT);
2729 opc = err_serious(-EFAULT);
2734 static int mdt_reint(struct tgt_session_info *tsi)
2738 static const struct req_format *reint_fmts[REINT_MAX] = {
2739 [REINT_SETATTR] = &RQF_MDS_REINT_SETATTR,
2740 [REINT_CREATE] = &RQF_MDS_REINT_CREATE,
2741 [REINT_LINK] = &RQF_MDS_REINT_LINK,
2742 [REINT_UNLINK] = &RQF_MDS_REINT_UNLINK,
2743 [REINT_RENAME] = &RQF_MDS_REINT_RENAME,
2744 [REINT_OPEN] = &RQF_MDS_REINT_OPEN,
2745 [REINT_SETXATTR] = &RQF_MDS_REINT_SETXATTR,
2746 [REINT_RMENTRY] = &RQF_MDS_REINT_UNLINK,
2747 [REINT_MIGRATE] = &RQF_MDS_REINT_MIGRATE,
2748 [REINT_RESYNC] = &RQF_MDS_REINT_RESYNC,
2753 opc = mdt_reint_opcode(tgt_ses_req(tsi), reint_fmts);
2755 struct mdt_thread_info *info = tsi2mdt_info(tsi);
2757 * No lock possible here from client to pass it to reint code
2760 rc = mdt_reint_internal(info, NULL, opc);
2761 mdt_thread_info_fini(info);
2766 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_REINT_NET_REP;
2770 /* this should sync the whole device */
2771 int mdt_device_sync(const struct lu_env *env, struct mdt_device *mdt)
2773 struct dt_device *dt = mdt->mdt_bottom;
2777 rc = dt->dd_ops->dt_sync(env, dt);
2781 /* this should sync this object */
2782 static int mdt_object_sync(const struct lu_env *env, struct obd_export *exp,
2783 struct mdt_object *mo)
2789 if (!mdt_object_exists(mo)) {
2790 CWARN("%s: non existing object "DFID": rc = %d\n",
2791 exp->exp_obd->obd_name, PFID(mdt_object_fid(mo)),
2796 if (S_ISREG(lu_object_attr(&mo->mot_obj))) {
2797 struct lu_target *tgt = tgt_ses_info(env)->tsi_tgt;
2798 dt_obj_version_t version;
2800 version = dt_version_get(env, mdt_obj2dt(mo));
2801 if (version > tgt->lut_obd->obd_last_committed)
2802 rc = mo_object_sync(env, mdt_object_child(mo));
2804 rc = mo_object_sync(env, mdt_object_child(mo));
2810 static int mdt_sync(struct tgt_session_info *tsi)
2812 struct ptlrpc_request *req = tgt_ses_req(tsi);
2813 struct req_capsule *pill = tsi->tsi_pill;
2814 struct mdt_body *body;
2815 ktime_t kstart = ktime_get();
2820 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SYNC_PACK))
2821 RETURN(err_serious(-ENOMEM));
2823 if (fid_seq(&tsi->tsi_mdt_body->mbo_fid1) == 0) {
2824 rc = mdt_device_sync(tsi->tsi_env, mdt_exp2dev(tsi->tsi_exp));
2826 struct mdt_thread_info *info = tsi2mdt_info(tsi);
2828 if (unlikely(info->mti_object == NULL))
2831 /* sync an object */
2832 rc = mdt_object_sync(tsi->tsi_env, tsi->tsi_exp,
2835 const struct lu_fid *fid;
2836 struct lu_attr *la = &info->mti_attr.ma_attr;
2838 info->mti_attr.ma_need = MA_INODE;
2839 info->mti_attr.ma_valid = 0;
2840 rc = mdt_attr_get_complex(info, info->mti_object,
2843 body = req_capsule_server_get(pill,
2845 fid = mdt_object_fid(info->mti_object);
2846 mdt_pack_attr2body(info, body, la, fid);
2849 mdt_thread_info_fini(info);
2852 mdt_counter_incr(req, LPROC_MDT_SYNC,
2853 ktime_us_delta(ktime_get(), kstart));
2858 static int mdt_data_sync(struct tgt_session_info *tsi)
2860 struct mdt_thread_info *info;
2861 struct mdt_device *mdt = mdt_exp2dev(tsi->tsi_exp);
2862 struct ost_body *body = tsi->tsi_ost_body;
2863 struct ost_body *repbody;
2864 struct mdt_object *mo = NULL;
2870 repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_OST_BODY);
2872 /* if no fid is specified then do nothing,
2873 * device sync is done via MDS_SYNC */
2874 if (fid_is_zero(&tsi->tsi_fid))
2877 mo = mdt_object_find(tsi->tsi_env, mdt, &tsi->tsi_fid);
2879 RETURN(PTR_ERR(mo));
2881 rc = mdt_object_sync(tsi->tsi_env, tsi->tsi_exp, mo);
2885 repbody->oa.o_oi = body->oa.o_oi;
2886 repbody->oa.o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
2888 info = tsi2mdt_info(tsi);
2889 ma = &info->mti_attr;
2890 ma->ma_need = MA_INODE;
2892 rc = mdt_attr_get_complex(info, mo, ma);
2894 obdo_from_la(&repbody->oa, &ma->ma_attr, VALID_FLAGS);
2897 mdt_thread_info_fini(info);
2902 mdt_object_put(tsi->tsi_env, mo);
2907 * Handle quota control requests to consult current usage/limit, but also
2908 * to configure quota enforcement
2910 static int mdt_quotactl(struct tgt_session_info *tsi)
2912 struct obd_export *exp = tsi->tsi_exp;
2913 struct req_capsule *pill = tsi->tsi_pill;
2914 struct obd_quotactl *oqctl, *repoqc;
2916 struct mdt_device *mdt = mdt_exp2dev(exp);
2917 struct lu_device *qmt = mdt->mdt_qmt_dev;
2918 struct lu_nodemap *nodemap;
2921 oqctl = req_capsule_client_get(pill, &RMF_OBD_QUOTACTL);
2923 RETURN(err_serious(-EPROTO));
2925 rc = req_capsule_server_pack(pill);
2927 RETURN(err_serious(rc));
2929 nodemap = nodemap_get_from_exp(exp);
2930 if (IS_ERR(nodemap))
2931 RETURN(PTR_ERR(nodemap));
2933 switch (oqctl->qc_cmd) {
2934 /* master quotactl */
2937 case LUSTRE_Q_SETDEFAULT:
2938 case LUSTRE_Q_SETQUOTAPOOL:
2939 case LUSTRE_Q_SETINFOPOOL:
2940 case LUSTRE_Q_SETDEFAULT_POOL:
2941 if (!nodemap_can_setquota(nodemap))
2942 GOTO(out_nodemap, rc = -EPERM);
2946 case LUSTRE_Q_GETDEFAULT:
2947 case LUSTRE_Q_GETQUOTAPOOL:
2948 case LUSTRE_Q_GETINFOPOOL:
2949 case LUSTRE_Q_GETDEFAULT_POOL:
2951 GOTO(out_nodemap, rc = -EOPNOTSUPP);
2952 /* slave quotactl */
2959 CERROR("%s: unsupported quotactl command %d: rc = %d\n",
2960 mdt_obd_name(mdt), oqctl->qc_cmd, rc);
2961 GOTO(out_nodemap, rc);
2965 switch (oqctl->qc_type) {
2967 id = nodemap_map_id(nodemap, NODEMAP_UID,
2968 NODEMAP_CLIENT_TO_FS, id);
2971 id = nodemap_map_id(nodemap, NODEMAP_GID,
2972 NODEMAP_CLIENT_TO_FS, id);
2975 /* todo: check/map project id */
2979 GOTO(out_nodemap, rc = -EOPNOTSUPP);
2981 repoqc = req_capsule_server_get(pill, &RMF_OBD_QUOTACTL);
2983 GOTO(out_nodemap, rc = err_serious(-EFAULT));
2985 if (oqctl->qc_cmd == Q_SETINFO || oqctl->qc_cmd == Q_SETQUOTA)
2986 barrier_exit(tsi->tsi_tgt->lut_bottom);
2988 if (oqctl->qc_id != id)
2989 swap(oqctl->qc_id, id);
2991 if (oqctl->qc_cmd == Q_SETINFO || oqctl->qc_cmd == Q_SETQUOTA) {
2992 if (unlikely(!barrier_entry(tsi->tsi_tgt->lut_bottom)))
2993 RETURN(-EINPROGRESS);
2996 switch (oqctl->qc_cmd) {
3002 case LUSTRE_Q_SETDEFAULT:
3003 case LUSTRE_Q_GETDEFAULT:
3004 case LUSTRE_Q_SETQUOTAPOOL:
3005 case LUSTRE_Q_GETQUOTAPOOL:
3006 case LUSTRE_Q_SETINFOPOOL:
3007 case LUSTRE_Q_GETINFOPOOL:
3008 case LUSTRE_Q_SETDEFAULT_POOL:
3009 case LUSTRE_Q_GETDEFAULT_POOL:
3010 /* forward quotactl request to QMT */
3011 rc = qmt_hdls.qmth_quotactl(tsi->tsi_env, qmt, oqctl);
3016 /* slave quotactl */
3017 rc = lquotactl_slv(tsi->tsi_env, tsi->tsi_tgt->lut_bottom,
3022 CERROR("Unsupported quotactl command: %d\n", oqctl->qc_cmd);
3023 GOTO(out_nodemap, rc = -EFAULT);
3026 if (oqctl->qc_id != id)
3027 swap(oqctl->qc_id, id);
3029 QCTL_COPY(repoqc, oqctl);
3033 nodemap_putref(nodemap);
3038 /** clone llog ctxt from child (mdd)
3039 * This allows remote llog (replicator) access.
3040 * We can either pass all llog RPCs (eg mdt_llog_create) on to child where the
3041 * context was originally set up, or we can handle them directly.
3042 * I choose the latter, but that means I need any llog
3043 * contexts set up by child to be accessable by the mdt. So we clone the
3044 * context into our context list here.
3046 static int mdt_llog_ctxt_clone(const struct lu_env *env, struct mdt_device *mdt,
3049 struct md_device *next = mdt->mdt_child;
3050 struct llog_ctxt *ctxt;
3053 if (!llog_ctxt_null(mdt2obd_dev(mdt), idx))
3056 rc = next->md_ops->mdo_llog_ctxt_get(env, next, idx, (void **)&ctxt);
3057 if (rc || ctxt == NULL) {
3061 rc = llog_group_set_ctxt(&mdt2obd_dev(mdt)->obd_olg, ctxt, idx);
3063 CERROR("Can't set mdt ctxt %d\n", rc);
3068 static int mdt_llog_ctxt_unclone(const struct lu_env *env,
3069 struct mdt_device *mdt, int idx)
3071 struct llog_ctxt *ctxt;
3073 ctxt = llog_get_context(mdt2obd_dev(mdt), idx);
3076 /* Put once for the get we just did, and once for the clone */
3077 llog_ctxt_put(ctxt);
3078 llog_ctxt_put(ctxt);
3083 * sec context handlers
3085 static int mdt_sec_ctx_handle(struct tgt_session_info *tsi)
3087 CFS_FAIL_TIMEOUT(OBD_FAIL_SEC_CTX_HDL_PAUSE, cfs_fail_val);
3093 * quota request handlers
3095 static int mdt_quota_dqacq(struct tgt_session_info *tsi)
3097 struct mdt_device *mdt = mdt_exp2dev(tsi->tsi_exp);
3098 struct lu_device *qmt = mdt->mdt_qmt_dev;
3103 RETURN(err_serious(-EOPNOTSUPP));
3105 rc = qmt_hdls.qmth_dqacq(tsi->tsi_env, qmt, tgt_ses_req(tsi));
3109 struct mdt_object *mdt_object_new(const struct lu_env *env,
3110 struct mdt_device *d,
3111 const struct lu_fid *f)
3113 struct lu_object_conf conf = { .loc_flags = LOC_F_NEW };
3114 struct lu_object *o;
3115 struct mdt_object *m;
3118 CDEBUG(D_INFO, "Allocate object for "DFID"\n", PFID(f));
3119 o = lu_object_find(env, &d->mdt_lu_dev, f, &conf);
3120 if (unlikely(IS_ERR(o)))
3121 m = (struct mdt_object *)o;
3127 struct mdt_object *mdt_object_find(const struct lu_env *env,
3128 struct mdt_device *d,
3129 const struct lu_fid *f)
3131 struct lu_object *o;
3132 struct mdt_object *m;
3135 CDEBUG(D_INFO, "Find object for "DFID"\n", PFID(f));
3136 o = lu_object_find(env, &d->mdt_lu_dev, f, NULL);
3137 if (unlikely(IS_ERR(o)))
3138 m = (struct mdt_object *)o;
3146 * Asyncronous commit for mdt device.
3148 * Pass asynchonous commit call down the MDS stack.
3150 * \param env environment
3151 * \param mdt the mdt device
3153 static void mdt_device_commit_async(const struct lu_env *env,
3154 struct mdt_device *mdt)
3156 struct dt_device *dt = mdt->mdt_bottom;
3160 rc = dt->dd_ops->dt_commit_async(env, dt);
3161 if (unlikely(rc != 0))
3162 CWARN("%s: async commit start failed: rc = %d\n",
3163 mdt_obd_name(mdt), rc);
3164 atomic_inc(&mdt->mdt_async_commit_count);
3169 * Mark the lock as "synchonous".
3171 * Mark the lock to deffer transaction commit to the unlock time.
3173 * \param lock the lock to mark as "synchonous"
3175 * \see mdt_is_lock_sync
3176 * \see mdt_save_lock
3178 static inline void mdt_set_lock_sync(struct ldlm_lock *lock)
3180 lock->l_ast_data = (void*)1;
3184 * Check whehter the lock "synchonous" or not.
3186 * \param lock the lock to check
3187 * \retval 1 the lock is "synchonous"
3188 * \retval 0 the lock isn't "synchronous"
3190 * \see mdt_set_lock_sync
3191 * \see mdt_save_lock
3193 static inline int mdt_is_lock_sync(struct ldlm_lock *lock)
3195 return lock->l_ast_data != NULL;
3199 * Blocking AST for mdt locks.
3201 * Starts transaction commit if in case of COS lock conflict or
3202 * deffers such a commit to the mdt_save_lock.
3204 * \param lock the lock which blocks a request or cancelling lock
3205 * \param desc unused
3206 * \param data unused
3207 * \param flag indicates whether this cancelling or blocking callback
3209 * \see ldlm_blocking_ast_nocheck
3211 int mdt_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
3212 void *data, int flag)
3214 struct obd_device *obd = ldlm_lock_to_ns(lock)->ns_obd;
3215 struct mdt_device *mdt = mdt_dev(obd->obd_lu_dev);
3216 struct ldlm_cb_set_arg *arg = data;
3217 bool commit_async = false;
3221 if (flag == LDLM_CB_CANCELING)
3224 lock_res_and_lock(lock);
3225 if (lock->l_blocking_ast != mdt_blocking_ast) {
3226 unlock_res_and_lock(lock);
3230 /* A blocking ast may be sent from ldlm_lock_decref_internal
3231 * when the last reference to a local lock was released and
3232 * during blocking event from ldlm_work_bl_ast_lock().
3233 * The 'data' parameter is l_ast_data in the first case and
3234 * callback arguments in the second one. Distinguish them by that.
3236 if (!data || data == lock->l_ast_data || !arg->bl_desc)
3237 goto skip_cos_checks;
3239 if (lock->l_req_mode & (LCK_PW | LCK_EX)) {
3240 if (mdt_cos_is_enabled(mdt)) {
3241 if (!arg->bl_desc->bl_same_client)
3242 mdt_set_lock_sync(lock);
3243 } else if (mdt_slc_is_enabled(mdt) &&
3244 arg->bl_desc->bl_cos_incompat) {
3245 mdt_set_lock_sync(lock);
3247 * we may do extra commit here, but there is a small
3248 * window to miss a commit: lock was unlocked (saved),
3249 * then a conflict lock queued and we come here, but
3250 * REP-ACK not received, so lock was not converted to
3252 * Fortunately this window is quite small, so the
3253 * extra commit should be rare (not to say distributed
3254 * operation is rare too).
3256 commit_async = true;
3258 } else if (lock->l_req_mode == LCK_COS) {
3259 commit_async = true;
3263 rc = ldlm_blocking_ast_nocheck(lock);
3268 rc = lu_env_init(&env, LCT_LOCAL);
3269 if (unlikely(rc != 0))
3270 CWARN("%s: lu_env initialization failed, cannot "
3271 "start asynchronous commit: rc = %d\n",
3274 mdt_device_commit_async(&env, mdt);
3281 * Blocking AST for cross-MDT lock
3283 * Discard lock from uncommitted_slc_locks and cancel it.
3285 * \param lock the lock which blocks a request or cancelling lock
3286 * \param desc unused
3287 * \param data unused
3288 * \param flag indicates whether this cancelling or blocking callback
3289 * \retval 0 on success
3290 * \retval negative number on error
3292 int mdt_remote_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
3293 void *data, int flag)
3299 case LDLM_CB_BLOCKING: {
3300 struct lustre_handle lockh;
3302 ldlm_lock2handle(lock, &lockh);
3303 rc = ldlm_cli_cancel(&lockh,
3304 ldlm_is_atomic_cb(lock) ? 0 : LCF_ASYNC);
3306 CDEBUG(D_INODE, "ldlm_cli_cancel: %d\n", rc);
3311 case LDLM_CB_CANCELING: {
3312 struct obd_device *obd = ldlm_lock_to_ns(lock)->ns_obd;
3313 struct mdt_device *mdt =
3314 mdt_dev(obd->obd_lu_dev->ld_site->ls_top_dev);
3316 LDLM_DEBUG(lock, "Revoke remote lock\n");
3318 /* discard slc lock here so that it can be cleaned anytime,
3319 * especially for cleanup_resource() */
3320 tgt_discard_slc_lock(&mdt->mdt_lut, lock);
3322 /* once we cache lock, l_ast_data is set to mdt_object */
3323 if (lock->l_ast_data != NULL) {
3324 struct mdt_object *mo = lock->l_ast_data;
3327 rc = lu_env_init(&env, LCT_MD_THREAD);
3328 if (unlikely(rc != 0)) {
3329 CWARN("%s: lu_env initialization failed, object %p "DFID" is leaked!: rc = %d\n",
3331 PFID(mdt_object_fid(mo)), rc);
3335 if (lock->l_policy_data.l_inodebits.bits &
3336 (MDS_INODELOCK_XATTR | MDS_INODELOCK_UPDATE)) {
3337 rc = mo_invalidate(&env, mdt_object_child(mo));
3338 mo->mot_cache_attr = 0;
3340 mdt_object_put(&env, mo);
3352 int mdt_check_resent_lock(struct mdt_thread_info *info,
3353 struct mdt_object *mo,
3354 struct mdt_lock_handle *lhc)
3356 /* the lock might already be gotten in ldlm_handle_enqueue() */
3357 if (unlikely(lustre_handle_is_used(&lhc->mlh_reg_lh))) {
3358 struct ptlrpc_request *req = mdt_info_req(info);
3359 struct ldlm_lock *lock;
3361 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
3362 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
3364 /* Lock is pinned by ldlm_handle_enqueue0() as it is
3365 * a resend case, however, it could be already destroyed
3366 * due to client eviction or a raced cancel RPC. */
3367 LDLM_DEBUG_NOLOCK("Invalid lock handle %#llx",
3368 lhc->mlh_reg_lh.cookie);
3372 if (!fid_res_name_eq(mdt_object_fid(mo),
3373 &lock->l_resource->lr_name)) {
3374 CWARN("%s: Although resent, but still not "
3375 "get child lock:"DFID"\n",
3376 info->mti_exp->exp_obd->obd_name,
3377 PFID(mdt_object_fid(mo)));
3378 LDLM_LOCK_PUT(lock);
3381 LDLM_LOCK_PUT(lock);
3387 static void mdt_remote_object_lock_created_cb(struct ldlm_lock *lock)
3389 mdt_object_get(NULL, lock->l_ast_data);
3392 int mdt_remote_object_lock_try(struct mdt_thread_info *mti,
3393 struct mdt_object *o, const struct lu_fid *fid,
3394 struct lustre_handle *lh, enum ldlm_mode mode,
3395 __u64 *ibits, __u64 trybits, bool cache)
3397 struct ldlm_enqueue_info *einfo = &mti->mti_remote_einfo;
3398 union ldlm_policy_data *policy = &mti->mti_policy;
3399 struct ldlm_res_id *res_id = &mti->mti_res_id;
3403 LASSERT(mdt_object_remote(o));
3405 fid_build_reg_res_name(fid, res_id);
3407 memset(einfo, 0, sizeof(*einfo));
3408 einfo->ei_type = LDLM_IBITS;
3409 einfo->ei_mode = mode;
3410 einfo->ei_cb_bl = mdt_remote_blocking_ast;
3411 einfo->ei_cb_cp = ldlm_completion_ast;
3412 einfo->ei_enq_slave = 0;
3413 einfo->ei_res_id = res_id;
3417 * if we cache lock, couple lock with mdt_object, so that object
3418 * can be easily found in lock ASTs.
3420 einfo->ei_cbdata = o;
3421 einfo->ei_cb_created = mdt_remote_object_lock_created_cb;
3424 memset(policy, 0, sizeof(*policy));
3425 policy->l_inodebits.bits = *ibits;
3426 policy->l_inodebits.try_bits = trybits;
3428 rc = mo_object_lock(mti->mti_env, mdt_object_child(o), lh, einfo,
3431 /* Return successfully acquired bits to a caller */
3433 struct ldlm_lock *lock = ldlm_handle2lock(lh);
3436 *ibits = lock->l_policy_data.l_inodebits.bits;
3437 LDLM_LOCK_PUT(lock);
3442 int mdt_remote_object_lock(struct mdt_thread_info *mti, struct mdt_object *o,
3443 const struct lu_fid *fid, struct lustre_handle *lh,
3444 enum ldlm_mode mode, __u64 ibits, bool cache)
3446 return mdt_remote_object_lock_try(mti, o, fid, lh, mode, &ibits, 0,
3450 int mdt_object_local_lock(struct mdt_thread_info *info, struct mdt_object *o,
3451 struct mdt_lock_handle *lh, __u64 *ibits,
3452 __u64 trybits, bool cos_incompat)
3454 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
3455 union ldlm_policy_data *policy = &info->mti_policy;
3456 struct ldlm_res_id *res_id = &info->mti_res_id;
3457 __u64 dlmflags = 0, *cookie = NULL;
3461 LASSERT(!lustre_handle_is_used(&lh->mlh_reg_lh));
3462 LASSERT(!lustre_handle_is_used(&lh->mlh_pdo_lh));
3463 LASSERT(lh->mlh_reg_mode != LCK_MINMODE);
3464 LASSERT(lh->mlh_type != MDT_NUL_LOCK);
3467 LASSERT(lh->mlh_reg_mode == LCK_PW ||
3468 lh->mlh_reg_mode == LCK_EX);
3469 dlmflags |= LDLM_FL_COS_INCOMPAT;
3470 } else if (mdt_cos_is_enabled(info->mti_mdt)) {
3471 dlmflags |= LDLM_FL_COS_ENABLED;
3474 /* Only enqueue LOOKUP lock for remote object */
3475 LASSERT(ergo(mdt_object_remote(o), *ibits == MDS_INODELOCK_LOOKUP));
3477 /* Lease lock are granted with LDLM_FL_CANCEL_ON_BLOCK */
3478 if (lh->mlh_type == MDT_REG_LOCK && lh->mlh_reg_mode == LCK_EX &&
3479 *ibits == MDS_INODELOCK_OPEN)
3480 dlmflags |= LDLM_FL_CANCEL_ON_BLOCK;
3482 if (lh->mlh_type == MDT_PDO_LOCK) {
3483 /* check for exists after object is locked */
3484 if (mdt_object_exists(o) == 0) {
3485 /* Non-existent object shouldn't have PDO lock */
3488 /* Non-dir object shouldn't have PDO lock */
3489 if (!S_ISDIR(lu_object_attr(&o->mot_obj)))
3494 fid_build_reg_res_name(mdt_object_fid(o), res_id);
3495 dlmflags |= LDLM_FL_ATOMIC_CB;
3498 cookie = &info->mti_exp->exp_handle.h_cookie;
3501 * Take PDO lock on whole directory and build correct @res_id for lock
3502 * on part of directory.
3504 if (lh->mlh_pdo_hash != 0) {
3505 LASSERT(lh->mlh_type == MDT_PDO_LOCK);
3506 mdt_lock_pdo_mode(info, o, lh);
3507 if (lh->mlh_pdo_mode != LCK_NL) {
3509 * Do not use LDLM_FL_LOCAL_ONLY for parallel lock, it
3510 * is never going to be sent to client and we do not
3511 * want it slowed down due to possible cancels.
3513 policy->l_inodebits.bits =
3514 *ibits & MDS_INODELOCK_UPDATE;
3515 policy->l_inodebits.try_bits =
3516 trybits & MDS_INODELOCK_UPDATE;
3517 /* at least one of them should be set */
3518 LASSERT(policy->l_inodebits.bits |
3519 policy->l_inodebits.try_bits);
3520 rc = mdt_fid_lock(info->mti_env, ns, &lh->mlh_pdo_lh,
3521 lh->mlh_pdo_mode, policy, res_id,
3523 if (unlikely(rc != 0))
3524 GOTO(out_unlock, rc);
3528 * Finish res_id initializing by name hash marking part of
3529 * directory which is taking modification.
3531 res_id->name[LUSTRE_RES_ID_HSH_OFF] = lh->mlh_pdo_hash;
3534 policy->l_inodebits.bits = *ibits;
3535 policy->l_inodebits.try_bits = trybits;
3536 policy->l_inodebits.li_gid = lh->mlh_gid;
3539 * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
3540 * going to be sent to client. If it is - mdt_intent_policy() path will
3541 * fix it up and turn FL_LOCAL flag off.
3543 rc = mdt_fid_lock(info->mti_env, ns, &lh->mlh_reg_lh, lh->mlh_reg_mode,
3544 policy, res_id, LDLM_FL_LOCAL_ONLY | dlmflags,
3548 mdt_object_unlock(info, o, lh, 1);
3549 else if (unlikely(OBD_FAIL_PRECHECK(OBD_FAIL_MDS_PDO_LOCK)) &&
3550 lh->mlh_pdo_hash != 0 &&
3551 (lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX))
3552 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK, 15);
3554 /* Return successfully acquired bits to a caller */
3556 struct ldlm_lock *lock = ldlm_handle2lock(&lh->mlh_reg_lh);
3559 *ibits = lock->l_policy_data.l_inodebits.bits;
3560 LDLM_LOCK_PUT(lock);
3566 mdt_object_lock_internal(struct mdt_thread_info *info, struct mdt_object *o,
3567 struct mdt_lock_handle *lh, __u64 *ibits,
3568 __u64 trybits, bool cos_incompat)
3570 struct mdt_lock_handle *local_lh = NULL;
3574 if (!mdt_object_remote(o)) {
3575 rc = mdt_object_local_lock(info, o, lh, ibits, trybits,
3580 /* XXX do not support PERM/LAYOUT/XATTR lock for remote object yet */
3581 *ibits &= ~(MDS_INODELOCK_PERM | MDS_INODELOCK_LAYOUT |
3582 MDS_INODELOCK_XATTR);
3584 /* Only enqueue LOOKUP lock for remote object */
3585 if (*ibits & MDS_INODELOCK_LOOKUP) {
3586 __u64 local = MDS_INODELOCK_LOOKUP;
3588 rc = mdt_object_local_lock(info, o, lh, &local, 0,
3596 if ((*ibits | trybits) & MDS_INODELOCK_UPDATE) {
3597 /* Sigh, PDO needs to enqueue 2 locks right now, but
3598 * enqueue RPC can only request 1 lock, to avoid extra
3599 * RPC, so it will instead enqueue EX lock for remote
3600 * object anyway XXX*/
3601 if (lh->mlh_type == MDT_PDO_LOCK &&
3602 lh->mlh_pdo_hash != 0) {
3604 "%s: "DFID" convert PDO lock to EX lock.\n",
3605 mdt_obd_name(info->mti_mdt),
3606 PFID(mdt_object_fid(o)));
3607 lh->mlh_pdo_hash = 0;
3608 lh->mlh_rreg_mode = LCK_EX;
3609 lh->mlh_type = MDT_REG_LOCK;
3612 rc = mdt_remote_object_lock_try(info, o, mdt_object_fid(o),
3615 ibits, trybits, false);
3616 if (rc != ELDLM_OK) {
3617 if (local_lh != NULL)
3618 mdt_object_unlock(info, o, local_lh, rc);
3623 /* other components like LFSCK can use lockless access
3624 * and populate cache, so we better invalidate it */
3625 mo_invalidate(info->mti_env, mdt_object_child(o));
3630 int mdt_object_lock(struct mdt_thread_info *info, struct mdt_object *o,
3631 struct mdt_lock_handle *lh, __u64 ibits)
3633 return mdt_object_lock_internal(info, o, lh, &ibits, 0, false);
3636 int mdt_reint_object_lock(struct mdt_thread_info *info, struct mdt_object *o,
3637 struct mdt_lock_handle *lh, __u64 ibits,
3640 LASSERT(lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX);
3641 return mdt_object_lock_internal(info, o, lh, &ibits, 0,
3645 int mdt_object_lock_try(struct mdt_thread_info *info, struct mdt_object *o,
3646 struct mdt_lock_handle *lh, __u64 *ibits,
3647 __u64 trybits, bool cos_incompat)
3649 bool trylock_only = *ibits == 0;
3652 LASSERT(!(*ibits & trybits));
3653 rc = mdt_object_lock_internal(info, o, lh, ibits, trybits,
3655 if (rc && trylock_only) { /* clear error for try ibits lock only */
3656 LASSERT(*ibits == 0);
3663 * Save a lock within request object.
3665 * Keep the lock referenced until whether client ACK or transaction
3666 * commit happens or release the lock immediately depending on input
3667 * parameters. If COS is ON, a write lock is converted to COS lock
3670 * \param info thead info object
3671 * \param h lock handle
3672 * \param mode lock mode
3673 * \param decref force immediate lock releasing
3675 void mdt_save_lock(struct mdt_thread_info *info, struct lustre_handle *h,
3676 enum ldlm_mode mode, int decref)
3680 if (lustre_handle_is_used(h)) {
3681 if (decref || !info->mti_has_trans ||
3682 !(mode & (LCK_PW | LCK_EX))) {
3683 mdt_fid_unlock(h, mode);
3685 struct mdt_device *mdt = info->mti_mdt;
3686 struct ldlm_lock *lock = ldlm_handle2lock(h);
3687 struct ptlrpc_request *req = mdt_info_req(info);
3688 bool cos = mdt_cos_is_enabled(mdt);
3689 bool convert_lock = !cos && mdt_slc_is_enabled(mdt);
3691 LASSERTF(lock != NULL, "no lock for cookie %#llx\n",
3694 /* there is no request if mdt_object_unlock() is called
3695 * from mdt_export_cleanup()->mdt_add_dirty_flag() */
3696 if (likely(req != NULL)) {
3697 LDLM_DEBUG(lock, "save lock request %p reply "
3698 "state %p transno %lld\n", req,
3699 req->rq_reply_state, req->rq_transno);
3701 ldlm_lock_mode_downgrade(lock, LCK_COS);
3704 if (req->rq_export->exp_disconnected)
3705 mdt_fid_unlock(h, mode);
3707 ptlrpc_save_lock(req, h, mode, cos,
3710 mdt_fid_unlock(h, mode);
3712 if (mdt_is_lock_sync(lock)) {
3713 CDEBUG(D_HA, "found sync-lock,"
3714 " async commit started\n");
3715 mdt_device_commit_async(info->mti_env,
3718 LDLM_LOCK_PUT(lock);
3727 * Save cross-MDT lock in uncommitted_slc_locks
3729 * Keep the lock referenced until transaction commit happens or release the lock
3730 * immediately depending on input parameters.
3732 * \param info thead info object
3733 * \param h lock handle
3734 * \param mode lock mode
3735 * \param decref force immediate lock releasing
3737 static void mdt_save_remote_lock(struct mdt_thread_info *info,
3738 struct mdt_object *o, struct lustre_handle *h,
3739 enum ldlm_mode mode, int decref)
3743 if (lustre_handle_is_used(h)) {
3744 struct ldlm_lock *lock = ldlm_handle2lock(h);
3745 struct ptlrpc_request *req = mdt_info_req(info);
3748 (lock->l_policy_data.l_inodebits.bits &
3749 (MDS_INODELOCK_XATTR | MDS_INODELOCK_UPDATE)))
3750 mo_invalidate(info->mti_env, mdt_object_child(o));
3752 if (decref || !info->mti_has_trans || !req ||
3753 !(mode & (LCK_PW | LCK_EX))) {
3754 ldlm_lock_decref_and_cancel(h, mode);
3755 LDLM_LOCK_PUT(lock);
3757 tgt_save_slc_lock(&info->mti_mdt->mdt_lut, lock,
3759 ldlm_lock_decref(h, mode);
3768 * Unlock mdt object.
3770 * Immeditely release the regular lock and the PDO lock or save the
3771 * lock in request and keep them referenced until client ACK or
3772 * transaction commit.
3774 * \param info thread info object
3775 * \param o mdt object
3776 * \param lh mdt lock handle referencing regular and PDO locks
3777 * \param decref force immediate lock releasing
3779 * XXX o is not used and may be NULL, see hsm_cdt_request_completed().
3781 void mdt_object_unlock(struct mdt_thread_info *info, struct mdt_object *o,
3782 struct mdt_lock_handle *lh, int decref)
3786 mdt_save_lock(info, &lh->mlh_pdo_lh, lh->mlh_pdo_mode, decref);
3787 mdt_save_lock(info, &lh->mlh_reg_lh, lh->mlh_reg_mode, decref);
3788 mdt_save_remote_lock(info, o, &lh->mlh_rreg_lh, lh->mlh_rreg_mode,
3794 struct mdt_object *mdt_object_find_lock(struct mdt_thread_info *info,
3795 const struct lu_fid *f,
3796 struct mdt_lock_handle *lh,
3799 struct mdt_object *o;
3801 o = mdt_object_find(info->mti_env, info->mti_mdt, f);
3805 rc = mdt_object_lock(info, o, lh, ibits);
3807 mdt_object_put(info->mti_env, o);
3814 void mdt_object_unlock_put(struct mdt_thread_info * info,
3815 struct mdt_object * o,
3816 struct mdt_lock_handle *lh,
3819 mdt_object_unlock(info, o, lh, decref);
3820 mdt_object_put(info->mti_env, o);
3824 * Generic code handling requests that have struct mdt_body passed in:
3826 * - extract mdt_body from request and save it in @info, if present;
3828 * - create lu_object, corresponding to the fid in mdt_body, and save it in
3831 * - if HAS_BODY flag is set for this request type check whether object
3832 * actually exists on storage (lu_object_exists()).
3835 static int mdt_body_unpack(struct mdt_thread_info *info,
3836 enum tgt_handler_flags flags)
3838 const struct mdt_body *body;
3839 struct mdt_object *obj;
3840 const struct lu_env *env;
3841 struct req_capsule *pill;
3845 env = info->mti_env;
3846 pill = info->mti_pill;
3848 body = info->mti_body = req_capsule_client_get(pill, &RMF_MDT_BODY);
3852 if (!(body->mbo_valid & OBD_MD_FLID))
3855 if (!fid_is_sane(&body->mbo_fid1)) {
3856 CERROR("Invalid fid: "DFID"\n", PFID(&body->mbo_fid1));
3860 obj = mdt_object_find(env, info->mti_mdt, &body->mbo_fid1);
3862 if ((flags & HAS_BODY) && !mdt_object_exists(obj)) {
3863 mdt_object_put(env, obj);
3866 info->mti_object = obj;
3875 static int mdt_unpack_req_pack_rep(struct mdt_thread_info *info,
3876 enum tgt_handler_flags flags)
3878 struct req_capsule *pill = info->mti_pill;
3883 if (req_capsule_has_field(pill, &RMF_MDT_BODY, RCL_CLIENT))
3884 rc = mdt_body_unpack(info, flags);
3888 if (rc == 0 && (flags & HAS_REPLY)) {
3890 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
3891 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER,
3893 if (req_capsule_has_field(pill, &RMF_LOGCOOKIES, RCL_SERVER))
3894 req_capsule_set_size(pill, &RMF_LOGCOOKIES,
3897 /* Set ACL reply buffer size as LUSTRE_POSIX_ACL_MAX_SIZE_OLD
3898 * by default. If the target object has more ACL entries, then
3899 * enlarge the buffer when necessary. */
3900 if (req_capsule_has_field(pill, &RMF_ACL, RCL_SERVER))
3901 req_capsule_set_size(pill, &RMF_ACL, RCL_SERVER,
3902 LUSTRE_POSIX_ACL_MAX_SIZE_OLD);
3904 mdt_preset_secctx_size(info);
3905 mdt_preset_encctx_size(info);
3907 rc = req_capsule_server_pack(pill);
3909 CWARN("%s: cannot pack response: rc = %d\n",
3910 mdt_obd_name(info->mti_mdt), rc);
3915 void mdt_lock_handle_init(struct mdt_lock_handle *lh)
3917 lh->mlh_type = MDT_NUL_LOCK;
3918 lh->mlh_reg_lh.cookie = 0ull;
3919 lh->mlh_reg_mode = LCK_MINMODE;
3920 lh->mlh_pdo_lh.cookie = 0ull;
3921 lh->mlh_pdo_mode = LCK_MINMODE;
3922 lh->mlh_rreg_lh.cookie = 0ull;
3923 lh->mlh_rreg_mode = LCK_MINMODE;
3926 void mdt_lock_handle_fini(struct mdt_lock_handle *lh)
3928 LASSERT(!lustre_handle_is_used(&lh->mlh_reg_lh));
3929 LASSERT(!lustre_handle_is_used(&lh->mlh_pdo_lh));
3933 * Initialize fields of struct mdt_thread_info. Other fields are left in
3934 * uninitialized state, because it's too expensive to zero out whole
3935 * mdt_thread_info (> 1K) on each request arrival.
3937 void mdt_thread_info_init(struct ptlrpc_request *req,
3938 struct mdt_thread_info *info)
3942 info->mti_pill = &req->rq_pill;
3945 for (i = 0; i < ARRAY_SIZE(info->mti_lh); i++)
3946 mdt_lock_handle_init(&info->mti_lh[i]);
3948 /* mdt device: it can be NULL while CONNECT */
3949 if (req->rq_export) {
3950 info->mti_mdt = mdt_dev(req->rq_export->exp_obd->obd_lu_dev);
3951 info->mti_exp = req->rq_export;
3953 info->mti_mdt = NULL;
3954 info->mti_env = req->rq_svc_thread->t_env;
3955 info->mti_transno = lustre_msg_get_transno(req->rq_reqmsg);
3957 memset(&info->mti_attr, 0, sizeof(info->mti_attr));
3958 info->mti_big_buf = LU_BUF_NULL;
3959 info->mti_body = NULL;
3960 info->mti_object = NULL;
3961 info->mti_dlm_req = NULL;
3962 info->mti_has_trans = 0;
3963 info->mti_cross_ref = 0;
3964 info->mti_opdata = 0;
3965 info->mti_big_lmm_used = 0;
3966 info->mti_big_acl_used = 0;
3967 info->mti_som_valid = 0;
3969 info->mti_spec.no_create = 0;
3970 info->mti_spec.sp_rm_entry = 0;
3971 info->mti_spec.sp_permitted = 0;
3973 info->mti_spec.u.sp_ea.eadata = NULL;
3974 info->mti_spec.u.sp_ea.eadatalen = 0;
3977 void mdt_thread_info_fini(struct mdt_thread_info *info)
3981 if (info->mti_object != NULL) {
3982 mdt_object_put(info->mti_env, info->mti_object);
3983 info->mti_object = NULL;
3986 for (i = 0; i < ARRAY_SIZE(info->mti_lh); i++)
3987 mdt_lock_handle_fini(&info->mti_lh[i]);
3988 info->mti_env = NULL;
3989 info->mti_pill = NULL;
3990 info->mti_exp = NULL;
3991 info->mti_mdt = NULL;
3993 if (unlikely(info->mti_big_buf.lb_buf != NULL))
3994 lu_buf_free(&info->mti_big_buf);
3997 struct mdt_thread_info *tsi2mdt_info(struct tgt_session_info *tsi)
3999 struct mdt_thread_info *mti;
4001 mti = mdt_th_info(tsi->tsi_env);
4002 LASSERT(mti != NULL);
4004 mdt_thread_info_init(tgt_ses_req(tsi), mti);
4005 if (tsi->tsi_corpus != NULL) {
4006 mti->mti_object = mdt_obj(tsi->tsi_corpus);
4007 lu_object_get(tsi->tsi_corpus);
4009 mti->mti_body = tsi->tsi_mdt_body;
4010 mti->mti_dlm_req = tsi->tsi_dlm_req;
4015 static int mdt_tgt_connect(struct tgt_session_info *tsi)
4017 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_DELAY_CONDITIONAL) &&
4019 tsi2mdt_info(tsi)->mti_mdt->mdt_seq_site.ss_node_id)
4020 schedule_timeout_uninterruptible(cfs_time_seconds(3));
4022 return tgt_connect(tsi);
4025 static int mdt_intent_glimpse(enum ldlm_intent_flags it_opc,
4026 struct mdt_thread_info *info,
4027 struct ldlm_lock **lockp, __u64 flags)
4029 return mdt_glimpse_enqueue(info, info->mti_mdt->mdt_namespace,
4032 static int mdt_intent_brw(enum ldlm_intent_flags it_opc,
4033 struct mdt_thread_info *info,
4034 struct ldlm_lock **lockp, __u64 flags)
4036 return mdt_brw_enqueue(info, info->mti_mdt->mdt_namespace,
4040 int mdt_intent_lock_replace(struct mdt_thread_info *info,
4041 struct ldlm_lock **lockp,
4042 struct mdt_lock_handle *lh,
4043 __u64 flags, int result)
4045 struct ptlrpc_request *req = mdt_info_req(info);
4046 struct ldlm_lock *lock = *lockp;
4047 struct ldlm_lock *new_lock;
4049 /* If possible resent found a lock, @lh is set to its handle */
4050 new_lock = ldlm_handle2lock_long(&lh->mlh_reg_lh, 0);
4052 if (new_lock == NULL && (flags & LDLM_FL_INTENT_ONLY)) {
4053 lh->mlh_reg_lh.cookie = 0;
4057 if (new_lock == NULL && (flags & LDLM_FL_RESENT)) {
4058 /* Lock is pinned by ldlm_handle_enqueue0() as it is
4059 * a resend case, however, it could be already destroyed
4060 * due to client eviction or a raced cancel RPC. */
4061 LDLM_DEBUG_NOLOCK("Invalid lock handle %#llx\n",
4062 lh->mlh_reg_lh.cookie);
4063 lh->mlh_reg_lh.cookie = 0;
4067 LASSERTF(new_lock != NULL,
4068 "lockh %#llx flags %#llx : rc = %d\n",
4069 lh->mlh_reg_lh.cookie, flags, result);
4072 * If we've already given this lock to a client once, then we should
4073 * have no readers or writers. Otherwise, we should have one reader
4074 * _or_ writer ref (which will be zeroed below) before returning the
4077 if (new_lock->l_export == req->rq_export) {
4078 LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
4080 LASSERT(new_lock->l_export == NULL);
4081 LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
4086 if (new_lock->l_export == req->rq_export) {
4088 * Already gave this to the client, which means that we
4089 * reconstructed a reply.
4091 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
4094 LDLM_LOCK_RELEASE(new_lock);
4095 lh->mlh_reg_lh.cookie = 0;
4096 RETURN(ELDLM_LOCK_REPLACED);
4100 * Fixup the lock to be given to the client.
4102 lock_res_and_lock(new_lock);
4103 /* Zero new_lock->l_readers and new_lock->l_writers without triggering
4104 * possible blocking AST. */
4105 while (new_lock->l_readers > 0) {
4106 lu_ref_del(&new_lock->l_reference, "reader", new_lock);
4107 lu_ref_del(&new_lock->l_reference, "user", new_lock);
4108 new_lock->l_readers--;
4110 while (new_lock->l_writers > 0) {
4111 lu_ref_del(&new_lock->l_reference, "writer", new_lock);
4112 lu_ref_del(&new_lock->l_reference, "user", new_lock);
4113 new_lock->l_writers--;
4116 new_lock->l_export = class_export_lock_get(req->rq_export, new_lock);
4117 new_lock->l_blocking_ast = lock->l_blocking_ast;
4118 new_lock->l_completion_ast = lock->l_completion_ast;
4119 if (ldlm_has_dom(new_lock))
4120 new_lock->l_glimpse_ast = ldlm_server_glimpse_ast;
4121 new_lock->l_remote_handle = lock->l_remote_handle;
4122 new_lock->l_flags &= ~LDLM_FL_LOCAL;
4124 unlock_res_and_lock(new_lock);
4126 cfs_hash_add(new_lock->l_export->exp_lock_hash,
4127 &new_lock->l_remote_handle,
4128 &new_lock->l_exp_hash);
4130 LDLM_LOCK_RELEASE(new_lock);
4131 lh->mlh_reg_lh.cookie = 0;
4133 RETURN(ELDLM_LOCK_REPLACED);
4136 void mdt_intent_fixup_resent(struct mdt_thread_info *info,
4137 struct ldlm_lock *new_lock,
4138 struct mdt_lock_handle *lh, __u64 flags)
4140 struct ptlrpc_request *req = mdt_info_req(info);
4141 struct ldlm_request *dlmreq;
4143 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
4146 dlmreq = req_capsule_client_get(info->mti_pill, &RMF_DLM_REQ);
4148 /* Check if this is a resend case (MSG_RESENT is set on RPC) and a
4149 * lock was found by ldlm_handle_enqueue(); if so @lh must be
4151 if (flags & LDLM_FL_RESENT) {
4152 lh->mlh_reg_lh.cookie = new_lock->l_handle.h_cookie;
4153 lh->mlh_reg_mode = new_lock->l_granted_mode;
4155 LDLM_DEBUG(new_lock, "Restoring lock cookie");
4156 DEBUG_REQ(D_DLMTRACE, req, "restoring lock cookie %#llx",
4157 lh->mlh_reg_lh.cookie);
4162 * If the xid matches, then we know this is a resent request, and allow
4163 * it. (It's probably an OPEN, for which we don't send a lock.
4165 if (req_can_reconstruct(req, NULL) != 0)
4169 * This remote handle isn't enqueued, so we never received or processed
4170 * this request. Clear MSG_RESENT, because it can be handled like any
4171 * normal request now.
4173 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
4175 DEBUG_REQ(D_DLMTRACE, req, "no existing lock with rhandle %#llx",
4176 dlmreq->lock_handle[0].cookie);
4179 static int mdt_intent_getxattr(enum ldlm_intent_flags it_opc,
4180 struct mdt_thread_info *info,
4181 struct ldlm_lock **lockp,
4184 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
4185 struct ldlm_reply *ldlm_rep = NULL;
4190 * Initialize lhc->mlh_reg_lh either from a previously granted lock
4191 * (for the resend case) or a new lock. Below we will use it to
4192 * replace the original lock.
4194 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
4195 if (!lustre_handle_is_used(&lhc->mlh_reg_lh)) {
4196 mdt_lock_reg_init(lhc, (*lockp)->l_req_mode);
4197 rc = mdt_object_lock(info, info->mti_object, lhc,
4198 MDS_INODELOCK_XATTR);
4203 rc = mdt_getxattr(info);
4205 if (mdt_info_req(info)->rq_repmsg != NULL)
4206 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
4208 if (ldlm_rep == NULL ||
4209 OBD_FAIL_CHECK(OBD_FAIL_MDS_XATTR_REP)) {
4210 mdt_object_unlock(info, info->mti_object, lhc, 1);
4214 RETURN(err_serious(-EFAULT));
4217 ldlm_rep->lock_policy_res2 = clear_serious(rc);
4219 /* This is left for interop instead of adding a new interop flag.
4221 #if LUSTRE_VERSION_CODE > OBD_OCD_VERSION(3, 0, 0, 0)
4222 if (ldlm_rep->lock_policy_res2) {
4223 mdt_object_unlock(info, info->mti_object, lhc, 1);
4224 RETURN(ELDLM_LOCK_ABORTED);
4228 rc = mdt_intent_lock_replace(info, lockp, lhc, flags, rc);
4232 static int mdt_intent_getattr(enum ldlm_intent_flags it_opc,
4233 struct mdt_thread_info *info,
4234 struct ldlm_lock **lockp,
4237 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
4239 struct ldlm_reply *ldlm_rep;
4240 struct mdt_body *reqbody;
4241 struct mdt_body *repbody;
4245 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
4248 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
4251 info->mti_cross_ref = !!(reqbody->mbo_valid & OBD_MD_FLCROSSREF);
4252 repbody->mbo_eadatasize = 0;
4253 repbody->mbo_aclsize = 0;
4257 child_bits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
4260 child_bits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE |
4264 CERROR("%s: unsupported intent %#x\n",
4265 mdt_obd_name(info->mti_mdt), (unsigned int)it_opc);
4266 GOTO(out_shrink, rc = -EINVAL);
4269 rc = mdt_init_ucred(info, reqbody);
4271 GOTO(out_shrink, rc);
4273 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
4274 mdt_set_disposition(info, ldlm_rep, DISP_IT_EXECD);
4276 /* Get lock from request for possible resent case. */
4277 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
4279 rc = mdt_getattr_name_lock(info, lhc, child_bits, ldlm_rep);
4280 ldlm_rep->lock_policy_res2 = clear_serious(rc);
4282 if (mdt_get_disposition(ldlm_rep, DISP_LOOKUP_NEG))
4283 ldlm_rep->lock_policy_res2 = 0;
4284 if (!mdt_get_disposition(ldlm_rep, DISP_LOOKUP_POS) ||
4285 ldlm_rep->lock_policy_res2) {
4286 lhc->mlh_reg_lh.cookie = 0ull;
4287 GOTO(out_ucred, rc = ELDLM_LOCK_ABORTED);
4290 rc = mdt_intent_lock_replace(info, lockp, lhc, flags, rc);
4293 mdt_exit_ucred(info);
4295 mdt_client_compatibility(info);
4296 rc2 = mdt_fix_reply(info);
4302 static int mdt_intent_layout(enum ldlm_intent_flags it_opc,
4303 struct mdt_thread_info *info,
4304 struct ldlm_lock **lockp,
4307 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
4308 struct md_layout_change layout = { .mlc_opc = MD_LAYOUT_NOP };
4309 struct layout_intent *intent;
4310 struct ldlm_reply *ldlm_rep;
4311 struct lu_fid *fid = &info->mti_tmp_fid2;
4312 struct mdt_object *obj = NULL;
4313 int layout_size = 0;
4314 struct lu_buf *buf = &layout.mlc_buf;
4319 fid_extract_from_res_name(fid, &(*lockp)->l_resource->lr_name);
4321 intent = req_capsule_client_get(info->mti_pill, &RMF_LAYOUT_INTENT);
4325 CDEBUG(D_INFO, DFID "got layout change request from client: "
4326 "opc:%u flags:%#x extent "DEXT"\n",
4327 PFID(fid), intent->li_opc, intent->li_flags,
4328 PEXT(&intent->li_extent));
4330 switch (intent->li_opc) {
4331 case LAYOUT_INTENT_TRUNC:
4332 case LAYOUT_INTENT_WRITE:
4333 layout.mlc_opc = MD_LAYOUT_WRITE;
4334 layout.mlc_intent = intent;
4336 case LAYOUT_INTENT_ACCESS:
4338 case LAYOUT_INTENT_READ:
4339 case LAYOUT_INTENT_GLIMPSE:
4340 case LAYOUT_INTENT_RELEASE:
4341 case LAYOUT_INTENT_RESTORE:
4342 CERROR("%s: Unsupported layout intent opc %d\n",
4343 mdt_obd_name(info->mti_mdt), intent->li_opc);
4346 CERROR("%s: Unknown layout intent opc %d\n",
4347 mdt_obd_name(info->mti_mdt), intent->li_opc);
4351 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
4353 RETURN(PTR_ERR(obj));
4355 if (mdt_object_exists(obj) && !mdt_object_remote(obj)) {
4356 /* if layout is going to be changed don't use the current EA
4357 * size but the maximum one. That buffer will be shrinked
4358 * to the actual size in req_capsule_shrink() before reply.
4360 if (layout.mlc_opc == MD_LAYOUT_WRITE) {
4361 layout_size = info->mti_mdt->mdt_max_mdsize;
4363 layout_size = mdt_attr_get_eabuf_size(info, obj);
4364 if (layout_size < 0)
4365 GOTO(out, rc = layout_size);
4367 if (layout_size > info->mti_mdt->mdt_max_mdsize)
4368 info->mti_mdt->mdt_max_mdsize = layout_size;
4370 CDEBUG(D_INFO, "%s: layout_size %d\n",
4371 mdt_obd_name(info->mti_mdt), layout_size);
4375 * set reply buffer size, so that ldlm_handle_enqueue0()->
4376 * ldlm_lvbo_fill() will fill the reply buffer with lovea.
4378 req_capsule_set_size(info->mti_pill, &RMF_DLM_LVB, RCL_SERVER,
4380 rc = req_capsule_server_pack(info->mti_pill);
4384 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
4386 GOTO(out, rc = -EPROTO);
4388 mdt_set_disposition(info, ldlm_rep, DISP_IT_EXECD);
4390 /* take lock in ldlm_lock_enqueue() for LAYOUT_INTENT_ACCESS */
4391 if (layout.mlc_opc == MD_LAYOUT_NOP)
4394 rc = mdt_check_resent(info, mdt_reconstruct_generic, lhc);
4398 DEBUG_REQ(D_INODE, mdt_info_req(info), "resent opt.");
4399 rc = lustre_msg_get_status(mdt_info_req(info)->rq_repmsg);
4405 if (unlikely(req_is_replay(mdt_info_req(info)))) {
4406 buf->lb_buf = req_capsule_client_get(info->mti_pill,
4408 buf->lb_len = req_capsule_get_size(info->mti_pill,
4409 &RMF_EADATA, RCL_CLIENT);
4411 * If it's a replay of layout write intent RPC, the client has
4412 * saved the extended lovea when it get reply then.
4414 if (buf->lb_len > 0)
4415 mdt_fix_lov_magic(info, buf->lb_buf);
4418 /* Get lock from request for possible resent case. */
4419 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
4420 (*lockp)->l_lvb_type = LVB_T_LAYOUT;
4423 * Instantiate some layout components, if @buf contains lovea, then it's
4424 * a replay of the layout intent write RPC.
4426 rc = mdt_layout_change(info, obj, lhc, &layout);
4427 ldlm_rep->lock_policy_res2 = clear_serious(rc);
4429 if (lustre_handle_is_used(&lhc->mlh_reg_lh)) {
4430 rc = mdt_intent_lock_replace(info, lockp, lhc, flags, rc);
4431 if (rc == ELDLM_LOCK_REPLACED &&
4432 (*lockp)->l_granted_mode == LCK_EX)
4433 ldlm_lock_mode_downgrade(*lockp, LCK_CR);
4438 mdt_object_put(info->mti_env, obj);
4442 static int mdt_intent_open(enum ldlm_intent_flags it_opc,
4443 struct mdt_thread_info *info,
4444 struct ldlm_lock **lockp,
4447 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
4448 struct ldlm_reply *rep = NULL;
4451 struct ptlrpc_request *req = mdt_info_req(info);
4453 static const struct req_format *intent_fmts[REINT_MAX] = {
4454 [REINT_CREATE] = &RQF_LDLM_INTENT_CREATE,
4455 [REINT_OPEN] = &RQF_LDLM_INTENT_OPEN
4460 opc = mdt_reint_opcode(mdt_info_req(info), intent_fmts);
4464 /* Get lock from request for possible resent case. */
4465 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
4467 rc = mdt_reint_internal(info, lhc, opc);
4469 if (rc < 0 && lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
4470 DEBUG_REQ(D_ERROR, req, "Replay open failed with %d", rc);
4472 /* Check whether the reply has been packed successfully. */
4473 if (mdt_info_req(info)->rq_repmsg != NULL)
4474 rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
4479 RETURN(err_serious(-EFAULT));
4482 /* MDC expects this in any case */
4484 mdt_set_disposition(info, rep, DISP_LOOKUP_EXECD);
4486 /* the open lock or the lock for cross-ref object should be
4487 * returned to the client */
4488 if (lustre_handle_is_used(&lhc->mlh_reg_lh) &&
4489 (rc == 0 || rc == -MDT_EREMOTE_OPEN)) {
4490 rep->lock_policy_res2 = 0;
4491 rc = mdt_intent_lock_replace(info, lockp, lhc, flags, rc);
4495 rep->lock_policy_res2 = clear_serious(rc);
4497 if (rep->lock_policy_res2 == -ENOENT &&
4498 mdt_get_disposition(rep, DISP_LOOKUP_NEG) &&
4499 !mdt_get_disposition(rep, DISP_OPEN_CREATE))
4500 rep->lock_policy_res2 = 0;
4502 lhc->mlh_reg_lh.cookie = 0ull;
4503 if (rc == -ENOTCONN || rc == -ENODEV ||
4504 rc == -EOVERFLOW) { /**< if VBR failure then return error */
4506 * If it is the disconnect error (ENODEV & ENOCONN), the error
4507 * will be returned by rq_status, and client at ptlrpc layer
4508 * will detect this, then disconnect, reconnect the import
4509 * immediately, instead of impacting the following the rpc.
4514 * For other cases, the error will be returned by intent, and client
4515 * will retrieve the result from intent.
4517 RETURN(ELDLM_LOCK_ABORTED);
4520 static int mdt_intent_opc(enum ldlm_intent_flags it_opc,
4521 struct mdt_thread_info *info,
4522 struct ldlm_lock **lockp,
4523 u64 flags /* LDLM_FL_* */)
4525 struct req_capsule *pill = info->mti_pill;
4526 struct ptlrpc_request *req = mdt_info_req(info);
4527 const struct req_format *it_format;
4528 int (*it_handler)(enum ldlm_intent_flags,
4529 struct mdt_thread_info *,
4530 struct ldlm_lock **,
4532 enum tgt_handler_flags it_handler_flags = 0;
4533 struct ldlm_reply *rep;
4534 bool check_mdt_object = false;
4540 case IT_OPEN|IT_CREAT:
4542 * OCREAT is not a IS_MUTABLE request since the file may
4543 * already exist. We do the extra check of
4544 * OBD_CONNECT_RDONLY in mdt_reint_open() when we
4545 * really need to create the object.
4547 it_format = &RQF_LDLM_INTENT;
4548 it_handler = &mdt_intent_open;
4551 check_mdt_object = true;
4554 it_format = &RQF_LDLM_INTENT_GETATTR;
4555 it_handler = &mdt_intent_getattr;
4556 it_handler_flags = HAS_REPLY;
4559 check_mdt_object = true;
4560 it_format = &RQF_LDLM_INTENT_GETXATTR;
4561 it_handler = &mdt_intent_getxattr;
4562 it_handler_flags = HAS_BODY;
4565 it_format = &RQF_LDLM_INTENT_LAYOUT;
4566 it_handler = &mdt_intent_layout;
4569 it_format = &RQF_LDLM_INTENT;
4570 it_handler = &mdt_intent_glimpse;
4573 it_format = &RQF_LDLM_INTENT;
4574 it_handler = &mdt_intent_brw;
4576 case IT_QUOTA_DQACQ:
4577 case IT_QUOTA_CONN: {
4578 struct lu_device *qmt = info->mti_mdt->mdt_qmt_dev;
4581 RETURN(-EOPNOTSUPP);
4583 if (mdt_rdonly(req->rq_export))
4586 (*lockp)->l_lvb_type = LVB_T_LQUOTA;
4587 /* pass the request to quota master */
4588 rc = qmt_hdls.qmth_intent_policy(info->mti_env, qmt,
4589 mdt_info_req(info), lockp,
4594 CERROR("%s: unknown intent code %#x\n",
4595 mdt_obd_name(info->mti_mdt), it_opc);
4599 req_capsule_extend(pill, it_format);
4601 rc = mdt_unpack_req_pack_rep(info, it_handler_flags);
4605 if (unlikely(info->mti_object == NULL && check_mdt_object))
4608 if (it_handler_flags & IS_MUTABLE && mdt_rdonly(req->rq_export))
4611 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_INTENT_DELAY, 10);
4613 /* execute policy */
4614 rc = (*it_handler)(it_opc, info, lockp, flags);
4616 /* Check whether the reply has been packed successfully. */
4617 if (req->rq_repmsg != NULL) {
4618 rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
4619 rep->lock_policy_res2 =
4620 ptlrpc_status_hton(rep->lock_policy_res2);
4626 static void mdt_ptlrpc_stats_update(struct ptlrpc_request *req,
4627 enum ldlm_intent_flags it_opc)
4629 struct lprocfs_stats *srv_stats = ptlrpc_req2svc(req)->srv_stats;
4631 /* update stats when IT code is known */
4632 if (srv_stats != NULL)
4633 lprocfs_counter_incr(srv_stats,
4634 PTLRPC_LAST_CNTR + (it_opc == IT_GLIMPSE ?
4635 LDLM_GLIMPSE_ENQUEUE : LDLM_IBITS_ENQUEUE));
4638 static int mdt_intent_policy(const struct lu_env *env,
4639 struct ldlm_namespace *ns,
4640 struct ldlm_lock **lockp,
4642 enum ldlm_mode mode,
4643 __u64 flags, void *data)
4645 struct tgt_session_info *tsi;
4646 struct mdt_thread_info *info;
4647 struct ptlrpc_request *req = req_cookie;
4648 struct ldlm_intent *it;
4649 struct req_capsule *pill;
4650 const struct ldlm_lock_desc *ldesc;
4655 LASSERT(req != NULL);
4657 tsi = tgt_ses_info(env);
4659 info = tsi2mdt_info(tsi);
4660 LASSERT(info != NULL);
4661 pill = info->mti_pill;
4662 LASSERT(pill->rc_req == req);
4663 ldesc = &info->mti_dlm_req->lock_desc;
4665 if (req->rq_reqmsg->lm_bufcount > DLM_INTENT_IT_OFF) {
4666 req_capsule_extend(pill, &RQF_LDLM_INTENT_BASIC);
4667 it = req_capsule_client_get(pill, &RMF_LDLM_INTENT);
4669 mdt_ptlrpc_stats_update(req, it->opc);
4670 rc = mdt_intent_opc(it->opc, info, lockp, flags);
4674 /* Lock without inodebits makes no sense and will oops
4675 * later in ldlm. Let's check it now to see if we have
4676 * ibits corrupted somewhere in mdt_intent_opc().
4677 * The case for client miss to set ibits has been
4678 * processed by others. */
4679 LASSERT(ergo(ldesc->l_resource.lr_type == LDLM_IBITS,
4680 ldesc->l_policy_data.l_inodebits.bits != 0));
4682 rc = err_serious(-EFAULT);
4684 } else if (ldesc->l_resource.lr_type == LDLM_IBITS &&
4685 ldesc->l_policy_data.l_inodebits.bits == MDS_INODELOCK_DOM) {
4686 struct ldlm_reply *rep;
4688 /* No intent was provided but INTENT flag is set along with
4689 * DOM bit, this is considered as GLIMPSE request.
4690 * This logic is common for MDT and OST glimpse
4692 mdt_ptlrpc_stats_update(req, IT_GLIMPSE);
4693 rc = mdt_glimpse_enqueue(info, ns, lockp, flags);
4694 /* Check whether the reply has been packed successfully. */
4695 if (req->rq_repmsg != NULL) {
4696 rep = req_capsule_server_get(info->mti_pill,
4698 rep->lock_policy_res2 =
4699 ptlrpc_status_hton(rep->lock_policy_res2);
4702 /* No intent was provided */
4703 req_capsule_set_size(pill, &RMF_DLM_LVB, RCL_SERVER, 0);
4704 rc = req_capsule_server_pack(pill);
4706 rc = err_serious(rc);
4708 mdt_thread_info_fini(info);
4712 static void mdt_deregister_seq_exp(struct mdt_device *mdt)
4714 struct seq_server_site *ss = mdt_seq_site(mdt);
4716 if (ss->ss_node_id == 0)
4719 if (ss->ss_client_seq != NULL) {
4720 lustre_deregister_lwp_item(&ss->ss_client_seq->lcs_exp);
4721 ss->ss_client_seq->lcs_exp = NULL;
4724 if (ss->ss_server_fld != NULL) {
4725 lustre_deregister_lwp_item(&ss->ss_server_fld->lsf_control_exp);
4726 ss->ss_server_fld->lsf_control_exp = NULL;
4730 static void mdt_seq_fini_cli(struct mdt_device *mdt)
4732 struct seq_server_site *ss = mdt_seq_site(mdt);
4737 if (ss->ss_server_seq != NULL)
4738 seq_server_set_cli(NULL, ss->ss_server_seq, NULL);
4741 static int mdt_seq_fini(const struct lu_env *env, struct mdt_device *mdt)
4743 mdt_seq_fini_cli(mdt);
4744 mdt_deregister_seq_exp(mdt);
4746 return seq_site_fini(env, mdt_seq_site(mdt));
4750 * It will retrieve its FLDB entries from MDT0, and it only happens
4751 * when upgrading existent FS to 2.6 or when local FLDB is corrupted,
4752 * and it needs to refresh FLDB from the MDT0.
4754 static int mdt_register_lwp_callback(void *data)
4757 struct mdt_device *mdt = data;
4758 struct lu_server_fld *fld = mdt_seq_site(mdt)->ss_server_fld;
4762 LASSERT(mdt_seq_site(mdt)->ss_node_id != 0);
4764 rc = lu_env_init(&env, LCT_MD_THREAD);
4766 CERROR("%s: cannot init env: rc = %d\n", mdt_obd_name(mdt), rc);
4770 /* Allocate new sequence now to avoid creating local transaction
4771 * in the normal transaction process */
4772 rc = seq_server_check_and_alloc_super(&env,
4773 mdt_seq_site(mdt)->ss_server_seq);
4778 rc = fld_update_from_controller(&env, fld);
4780 CERROR("%s: cannot update controller: rc = %d\n",
4781 mdt_obd_name(mdt), rc);
4790 static int mdt_register_seq_exp(struct mdt_device *mdt)
4792 struct seq_server_site *ss = mdt_seq_site(mdt);
4793 char *lwp_name = NULL;
4796 if (ss->ss_node_id == 0)
4799 OBD_ALLOC(lwp_name, MAX_OBD_NAME);
4800 if (lwp_name == NULL)
4801 GOTO(out_free, rc = -ENOMEM);
4803 rc = tgt_name2lwp_name(mdt_obd_name(mdt), lwp_name, MAX_OBD_NAME, 0);
4807 rc = lustre_register_lwp_item(lwp_name, &ss->ss_client_seq->lcs_exp,
4812 rc = lustre_register_lwp_item(lwp_name,
4813 &ss->ss_server_fld->lsf_control_exp,
4814 mdt_register_lwp_callback, mdt);
4816 lustre_deregister_lwp_item(&ss->ss_client_seq->lcs_exp);
4817 ss->ss_client_seq->lcs_exp = NULL;
4821 if (lwp_name != NULL)
4822 OBD_FREE(lwp_name, MAX_OBD_NAME);
4828 * Init client sequence manager which is used by local MDS to talk to sequence
4829 * controller on remote node.
4831 static int mdt_seq_init_cli(const struct lu_env *env, struct mdt_device *mdt)
4833 struct seq_server_site *ss = mdt_seq_site(mdt);
4837 /* check if this is adding the first MDC and controller is not yet
4839 OBD_ALLOC_PTR(ss->ss_client_seq);
4840 if (ss->ss_client_seq == NULL)
4843 OBD_ALLOC(prefix, MAX_OBD_NAME + 5);
4844 if (prefix == NULL) {
4845 OBD_FREE_PTR(ss->ss_client_seq);
4846 ss->ss_client_seq = NULL;
4850 /* Note: seq_client_fini will be called in seq_site_fini */
4851 snprintf(prefix, MAX_OBD_NAME + 5, "ctl-%s", mdt_obd_name(mdt));
4852 seq_client_init(ss->ss_client_seq, NULL, LUSTRE_SEQ_METADATA,
4853 prefix, ss->ss_node_id == 0 ? ss->ss_control_seq :
4855 OBD_FREE(prefix, MAX_OBD_NAME + 5);
4857 RETURN(seq_server_set_cli(env, ss->ss_server_seq, ss->ss_client_seq));
4860 static int mdt_seq_init(const struct lu_env *env, struct mdt_device *mdt)
4862 struct seq_server_site *ss;
4866 ss = mdt_seq_site(mdt);
4867 /* init sequence controller server(MDT0) */
4868 if (ss->ss_node_id == 0) {
4869 OBD_ALLOC_PTR(ss->ss_control_seq);
4870 if (ss->ss_control_seq == NULL)
4873 rc = seq_server_init(env, ss->ss_control_seq, mdt->mdt_bottom,
4874 mdt_obd_name(mdt), LUSTRE_SEQ_CONTROLLER,
4877 GOTO(out_seq_fini, rc);
4880 /* Init normal sequence server */
4881 OBD_ALLOC_PTR(ss->ss_server_seq);
4882 if (ss->ss_server_seq == NULL)
4883 GOTO(out_seq_fini, rc = -ENOMEM);
4885 rc = seq_server_init(env, ss->ss_server_seq, mdt->mdt_bottom,
4886 mdt_obd_name(mdt), LUSTRE_SEQ_SERVER, ss);
4888 GOTO(out_seq_fini, rc);
4890 /* init seq client for seq server to talk to seq controller(MDT0) */
4891 rc = mdt_seq_init_cli(env, mdt);
4893 GOTO(out_seq_fini, rc);
4895 if (ss->ss_node_id != 0)
4896 /* register controller export through lwp */
4897 rc = mdt_register_seq_exp(mdt);
4902 mdt_seq_fini(env, mdt);
4910 static int mdt_fld_fini(const struct lu_env *env,
4911 struct mdt_device *m)
4913 struct seq_server_site *ss = mdt_seq_site(m);
4916 if (ss && ss->ss_server_fld) {
4917 fld_server_fini(env, ss->ss_server_fld);
4918 OBD_FREE_PTR(ss->ss_server_fld);
4919 ss->ss_server_fld = NULL;
4925 static int mdt_fld_init(const struct lu_env *env,
4927 struct mdt_device *m)
4929 struct seq_server_site *ss;
4933 ss = mdt_seq_site(m);
4935 OBD_ALLOC_PTR(ss->ss_server_fld);
4936 if (ss->ss_server_fld == NULL)
4937 RETURN(rc = -ENOMEM);
4939 rc = fld_server_init(env, ss->ss_server_fld, m->mdt_bottom, uuid,
4942 OBD_FREE_PTR(ss->ss_server_fld);
4943 ss->ss_server_fld = NULL;
4950 static void mdt_stack_pre_fini(const struct lu_env *env,
4951 struct mdt_device *m, struct lu_device *top)
4953 struct lustre_cfg_bufs *bufs;
4954 struct lustre_cfg *lcfg;
4955 struct mdt_thread_info *info;
4960 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
4961 LASSERT(info != NULL);
4963 bufs = &info->mti_u.bufs;
4965 LASSERT(m->mdt_child_exp);
4966 LASSERT(m->mdt_child_exp->exp_obd);
4968 /* process cleanup, pass mdt obd name to get obd umount flags */
4969 /* XXX: this is needed because all layers are referenced by
4970 * objects (some of them are pinned by osd, for example *
4971 * the proper solution should be a model where object used
4972 * by osd only doesn't have mdt/mdd slices -bzzz */
4973 lustre_cfg_bufs_reset(bufs, mdt_obd_name(m));
4974 lustre_cfg_bufs_set_string(bufs, 1, NULL);
4975 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
4978 lustre_cfg_init(lcfg, LCFG_PRE_CLEANUP, bufs);
4980 top->ld_ops->ldo_process_config(env, top, lcfg);
4981 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens));
4985 static void mdt_stack_fini(const struct lu_env *env,
4986 struct mdt_device *m, struct lu_device *top)
4988 struct obd_device *obd = mdt2obd_dev(m);
4989 struct lustre_cfg_bufs *bufs;
4990 struct lustre_cfg *lcfg;
4991 struct mdt_thread_info *info;
4995 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
4996 LASSERT(info != NULL);
4998 lu_dev_del_linkage(top->ld_site, top);
5000 lu_site_purge(env, top->ld_site, -1);
5002 bufs = &info->mti_u.bufs;
5003 /* process cleanup, pass mdt obd name to get obd umount flags */
5004 /* another purpose is to let all layers to release their objects */
5005 lustre_cfg_bufs_reset(bufs, mdt_obd_name(m));
5010 lustre_cfg_bufs_set_string(bufs, 1, flags);
5011 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
5014 lustre_cfg_init(lcfg, LCFG_CLEANUP, bufs);
5017 top->ld_ops->ldo_process_config(env, top, lcfg);
5018 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens));
5020 lu_site_purge(env, top->ld_site, -1);
5022 m->mdt_child = NULL;
5023 m->mdt_bottom = NULL;
5025 obd_disconnect(m->mdt_child_exp);
5026 m->mdt_child_exp = NULL;
5028 obd_disconnect(m->mdt_bottom_exp);
5029 m->mdt_child_exp = NULL;
5032 static int mdt_connect_to_next(const struct lu_env *env, struct mdt_device *m,
5033 const char *next, struct obd_export **exp)
5035 struct obd_connect_data *data = NULL;
5036 struct obd_device *obd;
5040 OBD_ALLOC_PTR(data);
5042 GOTO(out, rc = -ENOMEM);
5044 obd = class_name2obd(next);
5046 CERROR("%s: can't locate next device: %s\n",
5047 mdt_obd_name(m), next);
5048 GOTO(out, rc = -ENOTCONN);
5051 data->ocd_connect_flags = OBD_CONNECT_VERSION;
5052 data->ocd_version = LUSTRE_VERSION_CODE;
5054 rc = obd_connect(NULL, exp, obd, &obd->obd_uuid, data, NULL);
5056 CERROR("%s: cannot connect to next dev %s (%d)\n",
5057 mdt_obd_name(m), next, rc);
5067 static int mdt_stack_init(const struct lu_env *env, struct mdt_device *mdt,
5068 struct lustre_cfg *cfg)
5070 char *dev = lustre_cfg_string(cfg, 0);
5071 int rc, name_size, uuid_size;
5072 char *name, *uuid, *p;
5073 struct lustre_cfg_bufs *bufs;
5074 struct lustre_cfg *lcfg;
5075 struct obd_device *obd;
5076 struct lustre_profile *lprof;
5077 struct lu_site *site;
5080 /* in 1.8 we had the only device in the stack - MDS.
5081 * 2.0 introduces MDT, MDD, OSD; MDT starts others internally.
5082 * in 2.3 OSD is instantiated by obd_mount.c, so we need
5083 * to generate names and setup MDT, MDD. MDT will be using
5084 * generated name to connect to MDD. for MDD the next device
5085 * will be LOD with name taken from so called "profile" which
5086 * is generated by mount_option line
5088 * 1.8 MGS generates config. commands like this:
5089 * #06 (104)mount_option 0: 1:lustre-MDT0000 2:lustre-mdtlov
5090 * #08 (120)setup 0:lustre-MDT0000 1:dev 2:type 3:lustre-MDT0000
5091 * 2.0 MGS generates config. commands like this:
5092 * #07 (112)mount_option 0: 1:lustre-MDT0000 2:lustre-MDT0000-mdtlov
5093 * #08 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
5094 * 3:lustre-MDT0000-mdtlov 4:f
5096 * we generate MDD name from MDT one, just replacing T with D
5098 * after all the preparations, the logical equivalent will be
5099 * #01 (160)setup 0:lustre-MDD0000 1:lustre-MDD0000_UUID 2:0
5100 * 3:lustre-MDT0000-mdtlov 4:f
5101 * #02 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
5102 * 3:lustre-MDD0000 4:f
5104 * notice we build the stack from down to top: MDD first, then MDT */
5106 name_size = MAX_OBD_NAME;
5107 uuid_size = MAX_OBD_NAME;
5109 OBD_ALLOC(name, name_size);
5110 OBD_ALLOC(uuid, uuid_size);
5111 if (name == NULL || uuid == NULL)
5112 GOTO(cleanup_mem, rc = -ENOMEM);
5114 OBD_ALLOC_PTR(bufs);
5116 GOTO(cleanup_mem, rc = -ENOMEM);
5119 p = strstr(name, "-MDT");
5121 GOTO(free_bufs, rc = -ENOMEM);
5124 snprintf(uuid, MAX_OBD_NAME, "%s_UUID", name);
5126 lprof = class_get_profile(lustre_cfg_string(cfg, 0));
5127 if (lprof == NULL || lprof->lp_dt == NULL) {
5128 CERROR("can't find the profile: %s\n",
5129 lustre_cfg_string(cfg, 0));
5130 GOTO(free_bufs, rc = -EINVAL);
5133 lustre_cfg_bufs_reset(bufs, name);
5134 lustre_cfg_bufs_set_string(bufs, 1, LUSTRE_MDD_NAME);
5135 lustre_cfg_bufs_set_string(bufs, 2, uuid);
5136 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
5138 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
5140 GOTO(put_profile, rc = -ENOMEM);
5141 lustre_cfg_init(lcfg, LCFG_ATTACH, bufs);
5143 rc = class_attach(lcfg);
5145 GOTO(lcfg_cleanup, rc);
5147 obd = class_name2obd(name);
5149 CERROR("Can not find obd %s (%s in config)\n",
5150 MDD_OBD_NAME, lustre_cfg_string(cfg, 0));
5151 GOTO(lcfg_cleanup, rc = -EINVAL);
5154 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens));
5156 lustre_cfg_bufs_reset(bufs, name);
5157 lustre_cfg_bufs_set_string(bufs, 1, uuid);
5158 lustre_cfg_bufs_set_string(bufs, 2, dev);
5159 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
5161 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
5163 GOTO(class_detach, rc = -ENOMEM);
5164 lustre_cfg_init(lcfg, LCFG_SETUP, bufs);
5166 rc = class_setup(obd, lcfg);
5168 GOTO(class_detach, rc);
5170 /* connect to MDD we just setup */
5171 rc = mdt_connect_to_next(env, mdt, name, &mdt->mdt_child_exp);
5173 GOTO(class_detach, rc);
5175 site = mdt->mdt_child_exp->exp_obd->obd_lu_dev->ld_site;
5177 LASSERT(mdt_lu_site(mdt) == NULL);
5178 mdt->mdt_lu_dev.ld_site = site;
5179 site->ls_top_dev = &mdt->mdt_lu_dev;
5180 mdt->mdt_child = lu2md_dev(mdt->mdt_child_exp->exp_obd->obd_lu_dev);
5182 /* now connect to bottom OSD */
5183 snprintf(name, MAX_OBD_NAME, "%s-osd", dev);
5184 rc = mdt_connect_to_next(env, mdt, name, &mdt->mdt_bottom_exp);
5186 GOTO(class_detach, rc);
5188 lu2dt_dev(mdt->mdt_bottom_exp->exp_obd->obd_lu_dev);
5190 rc = lu_env_refill((struct lu_env *)env);
5192 CERROR("Failure to refill session: '%d'\n", rc);
5194 lu_dev_add_linkage(site, &mdt->mdt_lu_dev);
5199 class_detach(obd, lcfg);
5201 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens));
5203 class_put_profile(lprof);
5208 OBD_FREE(name, name_size);
5210 OBD_FREE(uuid, uuid_size);
5214 /* setup quota master target on MDT0 */
5215 static int mdt_quota_init(const struct lu_env *env, struct mdt_device *mdt,
5216 struct lustre_cfg *cfg)
5218 struct obd_device *obd;
5219 char *dev = lustre_cfg_string(cfg, 0);
5220 char *qmtname, *uuid, *p;
5221 struct lustre_cfg_bufs *bufs;
5222 struct lustre_cfg *lcfg;
5223 struct lustre_profile *lprof;
5224 struct obd_connect_data *data;
5228 LASSERT(mdt->mdt_qmt_exp == NULL);
5229 LASSERT(mdt->mdt_qmt_dev == NULL);
5231 /* quota master is on MDT0 only for now */
5232 if (mdt->mdt_seq_site.ss_node_id != 0)
5235 /* MGS generates config commands which look as follows:
5236 * #01 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
5237 * 3:lustre-MDT0000-mdtlov 4:f
5239 * We generate the QMT name from the MDT one, just replacing MD with QM
5240 * after all the preparations, the logical equivalent will be:
5241 * #01 (160)setup 0:lustre-QMT0000 1:lustre-QMT0000_UUID 2:0
5242 * 3:lustre-MDT0000-osd 4:f */
5243 OBD_ALLOC(qmtname, MAX_OBD_NAME);
5244 OBD_ALLOC(uuid, UUID_MAX);
5245 OBD_ALLOC_PTR(bufs);
5246 OBD_ALLOC_PTR(data);
5247 if (qmtname == NULL || uuid == NULL || bufs == NULL || data == NULL)
5248 GOTO(cleanup_mem, rc = -ENOMEM);
5250 strcpy(qmtname, dev);
5251 p = strstr(qmtname, "-MDT");
5253 GOTO(cleanup_mem, rc = -ENOMEM);
5254 /* replace MD with QM */
5258 snprintf(uuid, UUID_MAX, "%s_UUID", qmtname);
5260 lprof = class_get_profile(lustre_cfg_string(cfg, 0));
5261 if (lprof == NULL || lprof->lp_dt == NULL) {
5262 CERROR("can't find profile for %s\n",
5263 lustre_cfg_string(cfg, 0));
5264 GOTO(cleanup_mem, rc = -EINVAL);
5267 lustre_cfg_bufs_reset(bufs, qmtname);
5268 lustre_cfg_bufs_set_string(bufs, 1, LUSTRE_QMT_NAME);
5269 lustre_cfg_bufs_set_string(bufs, 2, uuid);
5270 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
5272 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
5274 GOTO(put_profile, rc = -ENOMEM);
5275 lustre_cfg_init(lcfg, LCFG_ATTACH, bufs);
5277 rc = class_attach(lcfg);
5279 GOTO(lcfg_cleanup, rc);
5281 obd = class_name2obd(qmtname);
5283 CERROR("Can not find obd %s (%s in config)\n", qmtname,
5284 lustre_cfg_string(cfg, 0));
5285 GOTO(lcfg_cleanup, rc = -EINVAL);
5288 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens));
5290 lustre_cfg_bufs_reset(bufs, qmtname);
5291 lustre_cfg_bufs_set_string(bufs, 1, uuid);
5292 lustre_cfg_bufs_set_string(bufs, 2, dev);
5294 /* for quota, the next device should be the OSD device */
5295 lustre_cfg_bufs_set_string(bufs, 3,
5296 mdt->mdt_bottom->dd_lu_dev.ld_obd->obd_name);
5298 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
5300 GOTO(class_detach, rc = -ENOMEM);
5301 lustre_cfg_init(lcfg, LCFG_SETUP, bufs);
5303 rc = class_setup(obd, lcfg);
5305 GOTO(class_detach, rc);
5307 mdt->mdt_qmt_dev = obd->obd_lu_dev;
5309 /* configure local quota objects */
5310 if (OBD_FAIL_CHECK(OBD_FAIL_QUOTA_INIT))
5313 rc = mdt->mdt_qmt_dev->ld_ops->ldo_prepare(env,
5317 GOTO(class_cleanup, rc);
5319 /* connect to quota master target */
5320 data->ocd_connect_flags = OBD_CONNECT_VERSION;
5321 data->ocd_version = LUSTRE_VERSION_CODE;
5322 rc = obd_connect(NULL, &mdt->mdt_qmt_exp, obd, &obd->obd_uuid,
5325 CERROR("cannot connect to quota master device %s (%d)\n",
5327 GOTO(class_cleanup, rc);
5333 class_manual_cleanup(obd);
5334 mdt->mdt_qmt_dev = NULL;
5335 GOTO(lcfg_cleanup, rc);
5339 class_detach(obd, lcfg);
5341 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens));
5343 class_put_profile(lprof);
5348 OBD_FREE(qmtname, MAX_OBD_NAME);
5350 OBD_FREE(uuid, UUID_MAX);
5356 /* Shutdown quota master target associated with mdt */
5357 static void mdt_quota_fini(const struct lu_env *env, struct mdt_device *mdt)
5361 if (mdt->mdt_qmt_exp == NULL)
5363 LASSERT(mdt->mdt_qmt_dev != NULL);
5365 /* the qmt automatically shuts down when the mdt disconnects */
5366 obd_disconnect(mdt->mdt_qmt_exp);
5367 mdt->mdt_qmt_exp = NULL;
5368 mdt->mdt_qmt_dev = NULL;
5372 /* mdt_getxattr() is used from mdt_intent_getxattr(), use this wrapper
5373 * for now. This will be removed along with converting rest of MDT code
5374 * to use tgt_session_info */
5375 static int mdt_tgt_getxattr(struct tgt_session_info *tsi)
5377 struct mdt_thread_info *info = tsi2mdt_info(tsi);
5380 if (unlikely(info->mti_object == NULL))
5383 rc = mdt_getxattr(info);
5385 mdt_thread_info_fini(info);
5389 static int mdt_llog_open(struct tgt_session_info *tsi)
5393 if (!mdt_is_rootadmin(tsi2mdt_info(tsi)))
5394 RETURN(err_serious(-EACCES));
5396 RETURN(tgt_llog_open(tsi));
5399 #define OBD_FAIL_OST_READ_NET OBD_FAIL_OST_BRW_NET
5400 #define OBD_FAIL_OST_WRITE_NET OBD_FAIL_OST_BRW_NET
5401 #define OST_BRW_READ OST_READ
5402 #define OST_BRW_WRITE OST_WRITE
5404 static struct tgt_handler mdt_tgt_handlers[] = {
5405 TGT_RPC_HANDLER(MDS_FIRST_OPC,
5406 0, MDS_CONNECT, mdt_tgt_connect,
5407 &RQF_CONNECT, LUSTRE_OBD_VERSION),
5408 TGT_RPC_HANDLER(MDS_FIRST_OPC,
5409 0, MDS_DISCONNECT, tgt_disconnect,
5410 &RQF_MDS_DISCONNECT, LUSTRE_OBD_VERSION),
5411 TGT_RPC_HANDLER(MDS_FIRST_OPC,
5412 HAS_REPLY, MDS_SET_INFO, mdt_set_info,
5413 &RQF_MDT_SET_INFO, LUSTRE_MDS_VERSION),
5414 TGT_MDT_HDL(0, MDS_GET_INFO, mdt_get_info),
5415 TGT_MDT_HDL(HAS_REPLY, MDS_GET_ROOT, mdt_get_root),
5416 TGT_MDT_HDL(HAS_BODY, MDS_GETATTR, mdt_getattr),
5417 TGT_MDT_HDL(HAS_BODY | HAS_REPLY, MDS_GETATTR_NAME,
5419 TGT_MDT_HDL(HAS_BODY, MDS_GETXATTR, mdt_tgt_getxattr),
5420 TGT_MDT_HDL(HAS_REPLY, MDS_STATFS, mdt_statfs),
5421 TGT_MDT_HDL(IS_MUTABLE, MDS_REINT, mdt_reint),
5422 TGT_MDT_HDL(HAS_BODY, MDS_CLOSE, mdt_close),
5423 TGT_MDT_HDL(HAS_BODY | HAS_REPLY, MDS_READPAGE, mdt_readpage),
5424 TGT_MDT_HDL(HAS_BODY | HAS_REPLY, MDS_SYNC, mdt_sync),
5425 TGT_MDT_HDL(0, MDS_QUOTACTL, mdt_quotactl),
5426 TGT_MDT_HDL(HAS_BODY | HAS_REPLY | IS_MUTABLE, MDS_HSM_PROGRESS,
5428 TGT_MDT_HDL(HAS_BODY | HAS_REPLY | IS_MUTABLE, MDS_HSM_CT_REGISTER,
5429 mdt_hsm_ct_register),
5430 TGT_MDT_HDL(HAS_BODY | HAS_REPLY | IS_MUTABLE, MDS_HSM_CT_UNREGISTER,
5431 mdt_hsm_ct_unregister),
5432 TGT_MDT_HDL(HAS_BODY | HAS_REPLY, MDS_HSM_STATE_GET,
5434 TGT_MDT_HDL(HAS_BODY | HAS_REPLY | IS_MUTABLE, MDS_HSM_STATE_SET,
5436 TGT_MDT_HDL(HAS_BODY | HAS_REPLY, MDS_HSM_ACTION, mdt_hsm_action),
5437 TGT_MDT_HDL(HAS_BODY | HAS_REPLY, MDS_HSM_REQUEST,
5439 TGT_MDT_HDL(HAS_KEY | HAS_BODY | HAS_REPLY | IS_MUTABLE,
5442 TGT_MDT_HDL(IS_MUTABLE, MDS_RMFID, mdt_rmfid),
5445 static struct tgt_handler mdt_io_ops[] = {
5446 TGT_OST_HDL_HP(HAS_BODY | HAS_REPLY, OST_BRW_READ, tgt_brw_read,
5448 TGT_OST_HDL_HP(HAS_BODY | IS_MUTABLE, OST_BRW_WRITE, tgt_brw_write,
5450 TGT_OST_HDL_HP(HAS_BODY | HAS_REPLY | IS_MUTABLE,
5451 OST_PUNCH, mdt_punch_hdl,
5453 TGT_OST_HDL(HAS_BODY | HAS_REPLY, OST_SYNC, mdt_data_sync),
5454 TGT_OST_HDL(HAS_BODY | HAS_REPLY, OST_SEEK, tgt_lseek),
5457 static struct tgt_handler mdt_sec_ctx_ops[] = {
5458 TGT_SEC_HDL_VAR(0, SEC_CTX_INIT, mdt_sec_ctx_handle),
5459 TGT_SEC_HDL_VAR(0, SEC_CTX_INIT_CONT,mdt_sec_ctx_handle),
5460 TGT_SEC_HDL_VAR(0, SEC_CTX_FINI, mdt_sec_ctx_handle)
5463 static struct tgt_handler mdt_quota_ops[] = {
5464 TGT_QUOTA_HDL(HAS_REPLY, QUOTA_DQACQ, mdt_quota_dqacq),
5467 static struct tgt_handler mdt_llog_handlers[] = {
5468 TGT_LLOG_HDL(0, LLOG_ORIGIN_HANDLE_CREATE, mdt_llog_open),
5469 TGT_LLOG_HDL(0, LLOG_ORIGIN_HANDLE_NEXT_BLOCK, tgt_llog_next_block),
5470 TGT_LLOG_HDL(0, LLOG_ORIGIN_HANDLE_READ_HEADER, tgt_llog_read_header),
5471 TGT_LLOG_HDL(0, LLOG_ORIGIN_HANDLE_PREV_BLOCK, tgt_llog_prev_block),
5474 static struct tgt_opc_slice mdt_common_slice[] = {
5476 .tos_opc_start = MDS_FIRST_OPC,
5477 .tos_opc_end = MDS_LAST_OPC,
5478 .tos_hs = mdt_tgt_handlers
5481 .tos_opc_start = OBD_FIRST_OPC,
5482 .tos_opc_end = OBD_LAST_OPC,
5483 .tos_hs = tgt_obd_handlers
5486 .tos_opc_start = LDLM_FIRST_OPC,
5487 .tos_opc_end = LDLM_LAST_OPC,
5488 .tos_hs = tgt_dlm_handlers
5491 .tos_opc_start = SEC_FIRST_OPC,
5492 .tos_opc_end = SEC_LAST_OPC,
5493 .tos_hs = mdt_sec_ctx_ops
5496 .tos_opc_start = OUT_UPDATE_FIRST_OPC,
5497 .tos_opc_end = OUT_UPDATE_LAST_OPC,
5498 .tos_hs = tgt_out_handlers
5501 .tos_opc_start = FLD_FIRST_OPC,
5502 .tos_opc_end = FLD_LAST_OPC,
5503 .tos_hs = fld_handlers
5506 .tos_opc_start = SEQ_FIRST_OPC,
5507 .tos_opc_end = SEQ_LAST_OPC,
5508 .tos_hs = seq_handlers
5511 .tos_opc_start = QUOTA_DQACQ,
5512 .tos_opc_end = QUOTA_LAST_OPC,
5513 .tos_hs = mdt_quota_ops
5516 .tos_opc_start = LLOG_FIRST_OPC,
5517 .tos_opc_end = LLOG_LAST_OPC,
5518 .tos_hs = mdt_llog_handlers
5521 .tos_opc_start = LFSCK_FIRST_OPC,
5522 .tos_opc_end = LFSCK_LAST_OPC,
5523 .tos_hs = tgt_lfsck_handlers
5526 .tos_opc_start = OST_FIRST_OPC,
5527 .tos_opc_end = OST_LAST_OPC,
5528 .tos_hs = mdt_io_ops
5535 static void mdt_fini(const struct lu_env *env, struct mdt_device *m)
5537 struct md_device *next = m->mdt_child;
5538 struct lu_device *d = &m->mdt_lu_dev;
5539 struct obd_device *obd = mdt2obd_dev(m);
5540 struct lfsck_stop stop;
5543 stop.ls_status = LS_PAUSED;
5545 next->md_ops->mdo_iocontrol(env, next, OBD_IOC_STOP_LFSCK, 0, &stop);
5547 mdt_stack_pre_fini(env, m, md2lu_dev(m->mdt_child));
5549 mdt_restriper_stop(m);
5550 ping_evictor_stop();
5552 /* Remove the HSM /proc entry so the coordinator cannot be
5553 * restarted by a user while it's shutting down.
5555 mdt_hsm_cdt_stop(m);
5557 mdt_llog_ctxt_unclone(env, m, LLOG_AGENT_ORIG_CTXT);
5558 mdt_llog_ctxt_unclone(env, m, LLOG_CHANGELOG_ORIG_CTXT);
5560 if (m->mdt_namespace != NULL)
5561 ldlm_namespace_free_prior(m->mdt_namespace, NULL,
5562 d->ld_obd->obd_force);
5564 obd_exports_barrier(obd);
5565 obd_zombie_barrier();
5567 mdt_quota_fini(env, m);
5569 cfs_free_nidlist(&m->mdt_squash.rsi_nosquash_nids);
5571 /* Calling the cleanup functions in the same order as in the mdt_init0
5574 mdt_tunables_fini(m);
5576 target_recovery_fini(obd);
5577 upcall_cache_cleanup(m->mdt_identity_cache);
5578 m->mdt_identity_cache = NULL;
5580 mdt_fs_cleanup(env, m);
5582 tgt_fini(env, &m->mdt_lut);
5584 mdt_hsm_cdt_fini(m);
5586 if (m->mdt_los != NULL) {
5587 local_oid_storage_fini(env, m->mdt_los);
5591 if (m->mdt_namespace != NULL) {
5592 ldlm_namespace_free_post(m->mdt_namespace);
5593 d->ld_obd->obd_namespace = m->mdt_namespace = NULL;
5596 if (m->mdt_md_root != NULL) {
5597 mdt_object_put(env, m->mdt_md_root);
5598 m->mdt_md_root = NULL;
5601 mdt_seq_fini(env, m);
5603 mdt_fld_fini(env, m);
5608 mdt_stack_fini(env, m, md2lu_dev(m->mdt_child));
5610 LASSERT(atomic_read(&d->ld_ref) == 0);
5612 server_put_mount(mdt_obd_name(m), true);
5617 static int mdt_postrecov(const struct lu_env *, struct mdt_device *);
5619 static int mdt_init0(const struct lu_env *env, struct mdt_device *m,
5620 struct lu_device_type *ldt, struct lustre_cfg *cfg)
5622 const struct dt_device_param *dt_conf;
5623 struct mdt_thread_info *info;
5624 struct obd_device *obd;
5625 const char *dev = lustre_cfg_string(cfg, 0);
5626 const char *num = lustre_cfg_string(cfg, 2);
5627 struct tg_grants_data *tgd = &m->mdt_lut.lut_tgd;
5628 struct lustre_mount_info *lmi = NULL;
5629 struct lustre_sb_info *lsi;
5631 struct seq_server_site *ss_site;
5632 const char *identity_upcall = "NONE";
5633 struct md_device *next;
5640 lu_device_init(&m->mdt_lu_dev, ldt);
5642 * Environment (env) might be missing mdt_thread_key values at that
5643 * point, if device is allocated when mdt_thread_key is in QUIESCENT
5646 * Usually device allocation path doesn't use module key values, but
5647 * mdt has to do a lot of work here, so allocate key value.
5649 rc = lu_env_refill((struct lu_env *)env);
5653 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
5654 LASSERT(info != NULL);
5656 obd = class_name2obd(dev);
5657 LASSERT(obd != NULL);
5659 m->mdt_max_mdsize = MAX_MD_SIZE_OLD;
5660 m->mdt_opts.mo_evict_tgt_nids = 1;
5661 m->mdt_opts.mo_cos = MDT_COS_DEFAULT;
5663 lmi = server_get_mount(dev);
5665 CERROR("Cannot get mount info for %s!\n", dev);
5668 lsi = s2lsi(lmi->lmi_sb);
5669 LASSERT(lsi->lsi_lmd);
5670 /* CMD is supported only in IAM mode */
5672 rc = kstrtol(num, 10, &node_id);
5676 obd->u.obt.obt_magic = OBT_MAGIC;
5677 if (lsi->lsi_lmd->lmd_flags & LMD_FLG_SKIP_LFSCK)
5678 m->mdt_skip_lfsck = 1;
5681 /* Just try to get a DoM lock by default. Otherwise, having a group
5682 * lock granted, it may get blocked for a long time. */
5683 m->mdt_opts.mo_dom_lock = TRYLOCK_DOM_ON_OPEN;
5684 /* DoM files are read at open and data is packed in the reply */
5685 m->mdt_opts.mo_dom_read_open = 1;
5687 m->mdt_squash.rsi_uid = 0;
5688 m->mdt_squash.rsi_gid = 0;
5689 INIT_LIST_HEAD(&m->mdt_squash.rsi_nosquash_nids);
5690 spin_lock_init(&m->mdt_squash.rsi_lock);
5691 spin_lock_init(&m->mdt_lock);
5692 m->mdt_enable_remote_dir = 1;
5693 m->mdt_enable_striped_dir = 1;
5694 m->mdt_enable_dir_migration = 1;
5695 m->mdt_enable_dir_restripe = 0;
5696 m->mdt_enable_dir_auto_split = 0;
5697 m->mdt_enable_remote_dir_gid = 0;
5698 m->mdt_enable_chprojid_gid = 0;
5699 m->mdt_enable_remote_rename = 1;
5700 m->mdt_dir_restripe_nsonly = 1;
5701 m->mdt_enable_remote_subdir_mount = 1;
5703 atomic_set(&m->mdt_mds_mds_conns, 0);
5704 atomic_set(&m->mdt_async_commit_count, 0);
5706 m->mdt_lu_dev.ld_ops = &mdt_lu_ops;
5707 m->mdt_lu_dev.ld_obd = obd;
5708 /* Set this lu_device to obd for error handling purposes. */
5709 obd->obd_lu_dev = &m->mdt_lu_dev;
5711 /* init the stack */
5712 rc = mdt_stack_init((struct lu_env *)env, m, cfg);
5714 CERROR("%s: Can't init device stack, rc %d\n",
5715 mdt_obd_name(m), rc);
5720 ss_site = mdt_seq_site(m);
5721 s->ld_seq_site = ss_site;
5724 /* set server index */
5725 ss_site->ss_node_id = node_id;
5727 /* failover is the default
5728 * FIXME: we do not failout mds0/mgs, which may cause some problems.
5729 * assumed whose ss_node_id == 0 XXX
5731 obd->obd_replayable = 1;
5732 /* No connection accepted until configurations will finish */
5733 obd->obd_no_conn = 1;
5735 if (cfg->lcfg_bufcount > 4 && LUSTRE_CFG_BUFLEN(cfg, 4) > 0) {
5736 char *str = lustre_cfg_string(cfg, 4);
5737 if (strchr(str, 'n')) {
5738 CWARN("%s: recovery disabled\n", mdt_obd_name(m));
5739 obd->obd_replayable = 0;
5743 rc = mdt_fld_init(env, mdt_obd_name(m), m);
5745 GOTO(err_fini_stack, rc);
5747 rc = mdt_seq_init(env, m);
5749 GOTO(err_fini_fld, rc);
5751 snprintf(info->mti_u.ns_name, sizeof(info->mti_u.ns_name), "%s-%s",
5752 LUSTRE_MDT_NAME, obd->obd_uuid.uuid);
5753 m->mdt_namespace = ldlm_namespace_new(obd, info->mti_u.ns_name,
5754 LDLM_NAMESPACE_SERVER,
5755 LDLM_NAMESPACE_GREEDY,
5757 if (IS_ERR(m->mdt_namespace)) {
5758 rc = PTR_ERR(m->mdt_namespace);
5759 CERROR("%s: unable to create server namespace: rc = %d\n",
5761 m->mdt_namespace = NULL;
5762 GOTO(err_fini_seq, rc);
5765 m->mdt_namespace->ns_lvbp = m;
5766 m->mdt_namespace->ns_lvbo = &mdt_lvbo;
5768 ldlm_register_intent(m->mdt_namespace, mdt_intent_policy);
5769 /* set obd_namespace for compatibility with old code */
5770 obd->obd_namespace = m->mdt_namespace;
5772 rc = tgt_init(env, &m->mdt_lut, obd, m->mdt_bottom, mdt_common_slice,
5773 OBD_FAIL_MDS_ALL_REQUEST_NET,
5774 OBD_FAIL_MDS_ALL_REPLY_NET);
5776 GOTO(err_free_ns, rc);
5778 /* Amount of available space excluded from granting and reserved
5779 * for metadata. It is a percentage of the total MDT size. */
5780 tgd->tgd_reserved_pcnt = 10;
5782 if (ONE_MB_BRW_SIZE < (1U << tgd->tgd_blockbits))
5783 m->mdt_brw_size = 1U << tgd->tgd_blockbits;
5785 m->mdt_brw_size = ONE_MB_BRW_SIZE;
5787 rc = mdt_fs_setup(env, m, obd, lsi);
5791 fid.f_seq = FID_SEQ_LOCAL_NAME;
5794 rc = local_oid_storage_init(env, m->mdt_bottom, &fid, &m->mdt_los);
5796 GOTO(err_fs_cleanup, rc);
5798 rc = mdt_hsm_cdt_init(m);
5800 CERROR("%s: error initializing coordinator, rc %d\n",
5801 mdt_obd_name(m), rc);
5802 GOTO(err_los_fini, rc);
5805 tgt_adapt_sptlrpc_conf(&m->mdt_lut);
5807 next = m->mdt_child;
5808 dt_conf = next->md_ops->mdo_dtconf_get(env, next);
5810 mntopts = dt_conf->ddp_mntopts;
5812 if (mntopts & MNTOPT_USERXATTR)
5813 m->mdt_opts.mo_user_xattr = 1;
5815 m->mdt_opts.mo_user_xattr = 0;
5817 m->mdt_max_ea_size = dt_conf->ddp_max_ea_size;
5819 if (mntopts & MNTOPT_ACL)
5820 m->mdt_opts.mo_acl = 1;
5822 m->mdt_opts.mo_acl = 0;
5824 /* XXX: to support suppgid for ACL, we enable identity_upcall
5825 * by default, otherwise, maybe got unexpected -EACCESS. */
5826 if (m->mdt_opts.mo_acl)
5827 identity_upcall = MDT_IDENTITY_UPCALL_PATH;
5829 m->mdt_identity_cache = upcall_cache_init(mdt_obd_name(m),
5831 &mdt_identity_upcall_cache_ops);
5832 if (IS_ERR(m->mdt_identity_cache)) {
5833 rc = PTR_ERR(m->mdt_identity_cache);
5834 m->mdt_identity_cache = NULL;
5835 GOTO(err_free_hsm, rc);
5838 rc = mdt_tunables_init(m, dev);
5840 CERROR("Can't init MDT lprocfs, rc %d\n", rc);
5841 GOTO(err_recovery, rc);
5844 rc = mdt_quota_init(env, m, cfg);
5846 GOTO(err_procfs, rc);
5848 m->mdt_ldlm_client = &mdt2obd_dev(m)->obd_ldlm_client;
5849 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
5850 "mdt_ldlm_client", m->mdt_ldlm_client);
5852 ping_evictor_start();
5854 /* recovery will be started upon mdt_prepare()
5855 * when the whole stack is complete and ready
5856 * to serve the requests */
5858 /* Reduce the initial timeout on an MDS because it doesn't need such
5859 * a long timeout as an OST does. Adaptive timeouts will adjust this
5860 * value appropriately. */
5861 if (ldlm_timeout == LDLM_TIMEOUT_DEFAULT)
5862 ldlm_timeout = MDS_LDLM_TIMEOUT_DEFAULT;
5864 if ((lsi->lsi_lmd->lmd_flags & LMD_FLG_LOCAL_RECOV))
5865 m->mdt_lut.lut_local_recovery = 1;
5867 rc = mdt_restriper_start(m);
5869 GOTO(err_ping_evictor, rc);
5874 ping_evictor_stop();
5876 mdt_tunables_fini(m);
5878 upcall_cache_cleanup(m->mdt_identity_cache);
5879 m->mdt_identity_cache = NULL;
5881 mdt_hsm_cdt_fini(m);
5883 local_oid_storage_fini(env, m->mdt_los);
5886 mdt_fs_cleanup(env, m);
5888 /* keep recoverable clients */
5890 target_recovery_fini(obd);
5891 obd_exports_barrier(obd);
5892 obd_zombie_barrier();
5893 tgt_fini(env, &m->mdt_lut);
5895 ldlm_namespace_free(m->mdt_namespace, NULL, 0);
5896 obd->obd_namespace = m->mdt_namespace = NULL;
5898 mdt_seq_fini(env, m);
5900 mdt_fld_fini(env, m);
5902 mdt_stack_fini(env, m, md2lu_dev(m->mdt_child));
5905 server_put_mount(dev, true);
5909 /* For interoperability, the left element is old parameter, the right one
5910 * is the new version of the parameter, if some parameter is deprecated,
5911 * the new version should be set as NULL. */
5912 static struct cfg_interop_param mdt_interop_param[] = {
5913 { "mdt.group_upcall", NULL },
5914 { "mdt.quota_type", NULL },
5915 { "mdd.quota_type", NULL },
5916 { "mdt.som", NULL },
5917 { "mdt.rootsquash", "mdt.root_squash" },
5918 { "mdt.nosquash_nid", "mdt.nosquash_nids" },
5922 /* used by MGS to process specific configurations */
5923 static int mdt_process_config(const struct lu_env *env,
5924 struct lu_device *d, struct lustre_cfg *cfg)
5926 struct mdt_device *m = mdt_dev(d);
5927 struct md_device *md_next = m->mdt_child;
5928 struct lu_device *next = md2lu_dev(md_next);
5932 switch (cfg->lcfg_command) {
5934 struct obd_device *obd = d->ld_obd;
5935 /* For interoperability */
5936 struct cfg_interop_param *ptr = NULL;
5937 struct lustre_cfg *old_cfg = NULL;
5941 param = lustre_cfg_string(cfg, 1);
5942 if (param == NULL) {
5943 CERROR("param is empty\n");
5948 ptr = class_find_old_param(param, mdt_interop_param);
5950 if (ptr->new_param == NULL) {
5952 CWARN("For interoperability, skip this %s."
5953 " It is obsolete.\n", ptr->old_param);
5957 CWARN("Found old param %s, changed it to %s.\n",
5958 ptr->old_param, ptr->new_param);
5961 cfg = lustre_cfg_rename(old_cfg, ptr->new_param);
5968 count = class_modify_config(cfg, PARAM_MDT,
5969 &obd->obd_kset.kobj);
5971 struct coordinator *cdt = &m->mdt_coordinator;
5973 /* is it an HSM var ? */
5974 count = class_modify_config(cfg, PARAM_HSM,
5975 &cdt->cdt_hsm_kobj);
5977 /* we don't understand; pass it on */
5978 rc = next->ld_ops->ldo_process_config(env, next,
5981 rc = count > 0 ? 0 : count;
5983 rc = count > 0 ? 0 : count;
5987 OBD_FREE(cfg, lustre_cfg_len(cfg->lcfg_bufcount,
5988 cfg->lcfg_buflens));
5992 /* others are passed further */
5993 rc = next->ld_ops->ldo_process_config(env, next, cfg);
5999 static struct lu_object *mdt_object_alloc(const struct lu_env *env,
6000 const struct lu_object_header *hdr,
6001 struct lu_device *d)
6003 struct mdt_object *mo;
6007 OBD_SLAB_ALLOC_PTR_GFP(mo, mdt_object_kmem, GFP_NOFS);
6009 struct lu_object *o;
6010 struct lu_object_header *h;
6013 h = &mo->mot_header;
6014 lu_object_header_init(h);
6015 lu_object_init(o, h, d);
6016 lu_object_add_top(h, o);
6017 o->lo_ops = &mdt_obj_ops;
6018 spin_lock_init(&mo->mot_write_lock);
6019 mutex_init(&mo->mot_som_mutex);
6020 mutex_init(&mo->mot_lov_mutex);
6021 init_rwsem(&mo->mot_dom_sem);
6022 init_rwsem(&mo->mot_open_sem);
6023 atomic_set(&mo->mot_open_count, 0);
6024 mo->mot_restripe_offset = 0;
6025 INIT_LIST_HEAD(&mo->mot_restripe_linkage);
6031 static int mdt_object_init(const struct lu_env *env, struct lu_object *o,
6032 const struct lu_object_conf *unused)
6034 struct mdt_device *d = mdt_dev(o->lo_dev);
6035 struct lu_device *under;
6036 struct lu_object *below;
6040 CDEBUG(D_INFO, "object init, fid = "DFID"\n",
6041 PFID(lu_object_fid(o)));
6043 under = &d->mdt_child->md_lu_dev;
6044 below = under->ld_ops->ldo_object_alloc(env, o->lo_header, under);
6045 if (below != NULL) {
6046 lu_object_add(o, below);
6053 static void mdt_object_free_rcu(struct rcu_head *head)
6055 struct mdt_object *mo = container_of(head, struct mdt_object,
6056 mot_header.loh_rcu);
6058 kmem_cache_free(mdt_object_kmem, mo);
6061 static void mdt_object_free(const struct lu_env *env, struct lu_object *o)
6063 struct mdt_object *mo = mdt_obj(o);
6064 struct lu_object_header *h;
6068 CDEBUG(D_INFO, "object free, fid = "DFID"\n",
6069 PFID(lu_object_fid(o)));
6071 LASSERT(atomic_read(&mo->mot_open_count) == 0);
6072 LASSERT(atomic_read(&mo->mot_lease_count) == 0);
6075 lu_object_header_fini(h);
6076 OBD_FREE_PRE(mo, sizeof(*mo), "slab-freed");
6077 call_rcu(&mo->mot_header.loh_rcu, mdt_object_free_rcu);
6082 static int mdt_object_print(const struct lu_env *env, void *cookie,
6083 lu_printer_t p, const struct lu_object *o)
6085 struct mdt_object *mdto = mdt_obj((struct lu_object *)o);
6087 return (*p)(env, cookie,
6088 LUSTRE_MDT_NAME"-object@%p(%s %s, writecount=%d)",
6089 mdto, mdto->mot_lov_created ? "lov_created" : "",
6090 mdto->mot_cache_attr ? "cache_attr" : "",
6091 mdto->mot_write_count);
6094 static int mdt_prepare(const struct lu_env *env,
6095 struct lu_device *pdev,
6096 struct lu_device *cdev)
6098 struct mdt_device *mdt = mdt_dev(cdev);
6099 struct lu_device *next = &mdt->mdt_child->md_lu_dev;
6100 struct obd_device *obd = cdev->ld_obd;
6107 rc = next->ld_ops->ldo_prepare(env, cdev, next);
6111 rc = mdt_llog_ctxt_clone(env, mdt, LLOG_CHANGELOG_ORIG_CTXT);
6115 rc = mdt_llog_ctxt_clone(env, mdt, LLOG_AGENT_ORIG_CTXT);
6119 rc = lfsck_register_namespace(env, mdt->mdt_bottom, mdt->mdt_namespace);
6120 /* The LFSCK instance is registered just now, so it must be there when
6121 * register the namespace to such instance. */
6122 LASSERTF(rc == 0, "register namespace failed: rc = %d\n", rc);
6124 if (mdt->mdt_seq_site.ss_node_id == 0) {
6125 rc = mdt->mdt_child->md_ops->mdo_root_get(env, mdt->mdt_child,
6126 &mdt->mdt_md_root_fid);
6131 LASSERT(!test_bit(MDT_FL_CFGLOG, &mdt->mdt_state));
6133 target_recovery_init(&mdt->mdt_lut, tgt_request_handle);
6134 set_bit(MDT_FL_CFGLOG, &mdt->mdt_state);
6135 LASSERT(obd->obd_no_conn);
6136 spin_lock(&obd->obd_dev_lock);
6137 obd->obd_no_conn = 0;
6138 spin_unlock(&obd->obd_dev_lock);
6140 if (obd->obd_recovering == 0)
6141 mdt_postrecov(env, mdt);
6146 const struct lu_device_operations mdt_lu_ops = {
6147 .ldo_object_alloc = mdt_object_alloc,
6148 .ldo_process_config = mdt_process_config,
6149 .ldo_prepare = mdt_prepare,
6152 static const struct lu_object_operations mdt_obj_ops = {
6153 .loo_object_init = mdt_object_init,
6154 .loo_object_free = mdt_object_free,
6155 .loo_object_print = mdt_object_print
6158 static int mdt_obd_set_info_async(const struct lu_env *env,
6159 struct obd_export *exp,
6160 __u32 keylen, void *key,
6161 __u32 vallen, void *val,
6162 struct ptlrpc_request_set *set)
6168 if (KEY_IS(KEY_SPTLRPC_CONF)) {
6169 rc = tgt_adapt_sptlrpc_conf(class_exp2tgt(exp));
6176 static inline void mdt_enable_slc(struct mdt_device *mdt)
6178 if (mdt->mdt_lut.lut_sync_lock_cancel == SYNC_LOCK_CANCEL_NEVER)
6179 mdt->mdt_lut.lut_sync_lock_cancel = SYNC_LOCK_CANCEL_BLOCKING;
6182 static inline void mdt_disable_slc(struct mdt_device *mdt)
6184 if (mdt->mdt_lut.lut_sync_lock_cancel == SYNC_LOCK_CANCEL_BLOCKING)
6185 mdt->mdt_lut.lut_sync_lock_cancel = SYNC_LOCK_CANCEL_NEVER;
6189 * Match client and server connection feature flags.
6191 * Compute the compatibility flags for a connection request based on
6192 * features mutually supported by client and server.
6194 * The obd_export::exp_connect_data.ocd_connect_flags field in \a exp
6195 * must not be updated here, otherwise a partially initialized value may
6196 * be exposed. After the connection request is successfully processed,
6197 * the top-level MDT connect request handler atomically updates the export
6198 * connect flags from the obd_connect_data::ocd_connect_flags field of the
6199 * reply. \see mdt_connect().
6201 * Before 2.7.50 clients will send a struct obd_connect_data_v1 rather than a
6202 * full struct obd_connect_data. So care must be taken when accessing fields
6203 * that are not present in struct obd_connect_data_v1. See LU-16.
6205 * \param exp the obd_export associated with this client/target pair
6206 * \param mdt the target device for the connection
6207 * \param data stores data for this connect request
6210 * \retval -EPROTO \a data unexpectedly has zero obd_connect_data::ocd_brw_size
6211 * \retval -EBADE client and server feature requirements are incompatible
6213 static int mdt_connect_internal(const struct lu_env *env,
6214 struct obd_export *exp,
6215 struct mdt_device *mdt,
6216 struct obd_connect_data *data, bool reconnect)
6218 const char *obd_name = mdt_obd_name(mdt);
6219 LASSERT(data != NULL);
6221 data->ocd_connect_flags &= MDT_CONNECT_SUPPORTED;
6223 if (mdt->mdt_bottom->dd_rdonly &&
6224 !(data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
6225 !(data->ocd_connect_flags & OBD_CONNECT_RDONLY))
6228 if (data->ocd_connect_flags & OBD_CONNECT_FLAGS2)
6229 data->ocd_connect_flags2 &= MDT_CONNECT_SUPPORTED2;
6231 data->ocd_ibits_known &= MDS_INODELOCK_FULL;
6233 if (!mdt->mdt_opts.mo_acl)
6234 data->ocd_connect_flags &= ~OBD_CONNECT_ACL;
6236 if (!mdt->mdt_opts.mo_user_xattr)
6237 data->ocd_connect_flags &= ~OBD_CONNECT_XATTR;
6239 if (OCD_HAS_FLAG(data, BRW_SIZE)) {
6240 data->ocd_brw_size = min(data->ocd_brw_size,
6242 if (data->ocd_brw_size == 0) {
6243 CERROR("%s: cli %s/%p ocd_connect_flags: %#llx "
6244 "ocd_version: %x ocd_grant: %d ocd_index: %u "
6245 "ocd_brw_size unexpectedly zero, network data "
6246 "corruption? Refusing to connect this client\n",
6247 obd_name, exp->exp_client_uuid.uuid,
6248 exp, data->ocd_connect_flags, data->ocd_version,
6249 data->ocd_grant, data->ocd_index);
6254 if (OCD_HAS_FLAG(data, GRANT_PARAM)) {
6255 struct dt_device_param *ddp = &mdt->mdt_lut.lut_dt_conf;
6257 /* client is reporting its page size, for future use */
6258 exp->exp_target_data.ted_pagebits = data->ocd_grant_blkbits;
6259 data->ocd_grant_blkbits = mdt->mdt_lut.lut_tgd.tgd_blockbits;
6260 /* ddp_inodespace may not be power-of-two value, eg. for ldiskfs
6261 * it's LDISKFS_DIR_REC_LEN(20) = 28. */
6262 data->ocd_grant_inobits = fls(ddp->ddp_inodespace - 1);
6263 /* ocd_grant_tax_kb is in 1K byte blocks */
6264 data->ocd_grant_tax_kb = ddp->ddp_extent_tax >> 10;
6265 data->ocd_grant_max_blks = ddp->ddp_max_extent_blks;
6268 /* Save connect_data we have so far because tgt_grant_connect()
6269 * uses it to calculate grant, and we want to save the client
6270 * version before it is overwritten by LUSTRE_VERSION_CODE. */
6271 exp->exp_connect_data = *data;
6272 if (OCD_HAS_FLAG(data, GRANT))
6273 tgt_grant_connect(env, exp, data, !reconnect);
6275 if (OCD_HAS_FLAG(data, MAXBYTES))
6276 data->ocd_maxbytes = mdt->mdt_lut.lut_dt_conf.ddp_maxbytes;
6278 /* NB: Disregard the rule against updating
6279 * exp_connect_data.ocd_connect_flags in this case, since
6280 * tgt_client_new() needs to know if this is a lightweight
6281 * connection, and it is safe to expose this flag before
6282 * connection processing completes. */
6283 if (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) {
6284 spin_lock(&exp->exp_lock);
6285 *exp_connect_flags_ptr(exp) |= OBD_CONNECT_LIGHTWEIGHT;
6286 spin_unlock(&exp->exp_lock);
6289 data->ocd_version = LUSTRE_VERSION_CODE;
6291 if ((data->ocd_connect_flags & OBD_CONNECT_FID) == 0) {
6292 CWARN("%s: MDS requires FID support, but client not\n",
6297 if (OCD_HAS_FLAG(data, PINGLESS)) {
6298 if (ptlrpc_pinger_suppress_pings()) {
6299 spin_lock(&exp->exp_obd->obd_dev_lock);
6300 list_del_init(&exp->exp_obd_chain_timed);
6301 spin_unlock(&exp->exp_obd->obd_dev_lock);
6303 data->ocd_connect_flags &= ~OBD_CONNECT_PINGLESS;
6307 data->ocd_max_easize = mdt->mdt_max_ea_size;
6309 /* NB: Disregard the rule against updating
6310 * exp_connect_data.ocd_connect_flags in this case, since
6311 * tgt_client_new() needs to know if this is client supports
6312 * multiple modify RPCs, and it is safe to expose this flag before
6313 * connection processing completes. */
6314 if (data->ocd_connect_flags & OBD_CONNECT_MULTIMODRPCS) {
6315 data->ocd_maxmodrpcs = max_mod_rpcs_per_client;
6316 spin_lock(&exp->exp_lock);
6317 *exp_connect_flags_ptr(exp) |= OBD_CONNECT_MULTIMODRPCS;
6318 spin_unlock(&exp->exp_lock);
6321 if (OCD_HAS_FLAG(data, CKSUM)) {
6322 __u32 cksum_types = data->ocd_cksum_types;
6324 tgt_mask_cksum_types(&mdt->mdt_lut, &data->ocd_cksum_types);
6326 if (unlikely(data->ocd_cksum_types == 0)) {
6327 CERROR("%s: Connect with checksum support but no "
6328 "ocd_cksum_types is set\n",
6329 exp->exp_obd->obd_name);
6333 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
6334 "%x\n", exp->exp_obd->obd_name, obd_export_nid2str(exp),
6335 cksum_types, data->ocd_cksum_types);
6337 /* This client does not support OBD_CONNECT_CKSUM
6338 * fall back to CRC32 */
6339 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
6340 "OBD_CONNECT_CKSUM, CRC32 will be used\n",
6341 exp->exp_obd->obd_name, obd_export_nid2str(exp));
6344 if ((data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
6345 !(data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT)) {
6346 atomic_inc(&mdt->mdt_mds_mds_conns);
6347 mdt_enable_slc(mdt);
6350 if (!mdt->mdt_lut.lut_dt_conf.ddp_has_lseek_data_hole)
6351 data->ocd_connect_flags2 &= ~OBD_CONNECT2_LSEEK;
6356 static int mdt_ctxt_add_dirty_flag(struct lu_env *env,
6357 struct mdt_thread_info *info,
6358 struct mdt_file_data *mfd)
6360 struct lu_context ses;
6364 rc = lu_context_init(&ses, LCT_SERVER_SESSION);
6369 lu_context_enter(&ses);
6371 mdt_ucred(info)->uc_valid = UCRED_OLD;
6372 rc = mdt_add_dirty_flag(info, mfd->mfd_object, &info->mti_attr);
6374 lu_context_exit(&ses);
6375 lu_context_fini(&ses);
6381 static int mdt_export_cleanup(struct obd_export *exp)
6383 LIST_HEAD(closing_list);
6384 struct mdt_export_data *med = &exp->exp_mdt_data;
6385 struct obd_device *obd = exp->exp_obd;
6386 struct mdt_device *mdt;
6387 struct mdt_thread_info *info;
6389 struct mdt_file_data *mfd, *n;
6393 spin_lock(&med->med_open_lock);
6394 while (!list_empty(&med->med_open_head)) {
6395 struct list_head *tmp = med->med_open_head.next;
6396 mfd = list_entry(tmp, struct mdt_file_data, mfd_list);
6398 /* Remove mfd handle so it can't be found again.
6399 * We are consuming the mfd_list reference here. */
6400 class_handle_unhash(&mfd->mfd_open_handle);
6401 list_move_tail(&mfd->mfd_list, &closing_list);
6403 spin_unlock(&med->med_open_lock);
6404 mdt = mdt_dev(obd->obd_lu_dev);
6405 LASSERT(mdt != NULL);
6407 rc = lu_env_init(&env, LCT_MD_THREAD);
6411 info = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
6412 LASSERT(info != NULL);
6413 memset(info, 0, sizeof *info);
6414 info->mti_env = &env;
6415 info->mti_mdt = mdt;
6416 info->mti_exp = exp;
6418 if (!list_empty(&closing_list)) {
6419 struct md_attr *ma = &info->mti_attr;
6421 /* Close any open files (which may also cause orphan
6423 list_for_each_entry_safe(mfd, n, &closing_list, mfd_list) {
6424 list_del_init(&mfd->mfd_list);
6425 ma->ma_need = ma->ma_valid = 0;
6427 /* This file is being closed due to an eviction, it
6428 * could have been modified and now dirty regarding to
6429 * HSM archive, check this!
6430 * The logic here is to mark a file dirty if there's a
6431 * chance it was dirtied before the client was evicted,
6432 * so that we don't have to wait for a release attempt
6433 * before finding out the file was actually dirty and
6434 * fail the release. Aggressively marking it dirty here
6435 * will cause the policy engine to attempt to
6436 * re-archive it; when rearchiving, we can compare the
6437 * current version to the HSM data_version and make the
6438 * archive request into a noop if it's not actually
6441 if (mfd->mfd_open_flags & MDS_FMODE_WRITE)
6442 rc = mdt_ctxt_add_dirty_flag(&env, info, mfd);
6444 /* Don't unlink orphan on failover umount, LU-184 */
6445 if (exp->exp_flags & OBD_OPT_FAILOVER) {
6446 ma->ma_valid = MA_FLAGS;
6447 ma->ma_attr_flags |= MDS_KEEP_ORPHAN;
6449 ma->ma_valid |= MA_FORCE_LOG;
6450 mdt_mfd_close(info, mfd);
6453 info->mti_mdt = NULL;
6454 /* cleanup client slot early */
6455 /* Do not erase record for recoverable client. */
6456 if (!(exp->exp_flags & OBD_OPT_FAILOVER) || exp->exp_failed)
6457 tgt_client_del(&env, exp);
6463 static int mdt_obd_disconnect(struct obd_export *exp)
6470 class_export_get(exp);
6472 if (!(exp->exp_flags & OBD_OPT_FORCE))
6473 tgt_grant_sanity_check(exp->exp_obd, __func__);
6475 if ((exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) &&
6476 !(exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)) {
6477 struct mdt_device *mdt = mdt_dev(exp->exp_obd->obd_lu_dev);
6479 if (atomic_dec_and_test(&mdt->mdt_mds_mds_conns))
6480 mdt_disable_slc(mdt);
6483 rc = server_disconnect_export(exp);
6485 CDEBUG(D_IOCTL, "server disconnect error: rc = %d\n", rc);
6487 tgt_grant_discard(exp);
6489 rc = mdt_export_cleanup(exp);
6490 nodemap_del_member(exp);
6491 class_export_put(exp);
6495 /* mds_connect copy */
6496 static int mdt_obd_connect(const struct lu_env *env,
6497 struct obd_export **exp, struct obd_device *obd,
6498 struct obd_uuid *cluuid,
6499 struct obd_connect_data *data,
6502 struct obd_export *lexp;
6503 struct lustre_handle conn = { 0 };
6504 struct mdt_device *mdt;
6506 lnet_nid_t *client_nid = localdata;
6509 LASSERT(env != NULL);
6510 LASSERT(data != NULL);
6512 if (!exp || !obd || !cluuid)
6515 mdt = mdt_dev(obd->obd_lu_dev);
6518 * first, check whether the stack is ready to handle requests
6519 * XXX: probably not very appropriate method is used now
6520 * at some point we should find a better one
6522 if (!test_bit(MDT_FL_SYNCED, &mdt->mdt_state) &&
6523 !(data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) &&
6524 !(data->ocd_connect_flags & OBD_CONNECT_MDS_MDS)) {
6525 rc = obd_get_info(env, mdt->mdt_child_exp,
6526 sizeof(KEY_OSP_CONNECTED),
6527 KEY_OSP_CONNECTED, NULL, NULL);
6530 set_bit(MDT_FL_SYNCED, &mdt->mdt_state);
6533 rc = class_connect(&conn, obd, cluuid);
6537 lexp = class_conn2export(&conn);
6538 LASSERT(lexp != NULL);
6540 rc = nodemap_add_member(*client_nid, lexp);
6541 if (rc != 0 && rc != -EEXIST)
6544 rc = mdt_connect_internal(env, lexp, mdt, data, false);
6546 struct lsd_client_data *lcd = lexp->exp_target_data.ted_lcd;
6549 memcpy(lcd->lcd_uuid, cluuid, sizeof lcd->lcd_uuid);
6550 rc = tgt_client_new(env, lexp);
6552 mdt_export_stats_init(obd, lexp, localdata);
6556 class_disconnect(lexp);
6557 nodemap_del_member(lexp);
6561 /* Because we do not want this export to be evicted by pinger,
6562 * let's not add this export to the timed chain list. */
6563 if (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) {
6564 spin_lock(&lexp->exp_obd->obd_dev_lock);
6565 list_del_init(&lexp->exp_obd_chain_timed);
6566 spin_unlock(&lexp->exp_obd->obd_dev_lock);
6573 static int mdt_obd_reconnect(const struct lu_env *env,
6574 struct obd_export *exp, struct obd_device *obd,
6575 struct obd_uuid *cluuid,
6576 struct obd_connect_data *data,
6579 lnet_nid_t *client_nid = localdata;
6583 if (exp == NULL || obd == NULL || cluuid == NULL)
6586 rc = nodemap_add_member(*client_nid, exp);
6587 if (rc != 0 && rc != -EEXIST)
6590 rc = mdt_connect_internal(env, exp, mdt_dev(obd->obd_lu_dev), data,
6593 mdt_export_stats_init(obd, exp, localdata);
6595 nodemap_del_member(exp);
6600 /* FIXME: Can we avoid using these two interfaces? */
6601 static int mdt_init_export(struct obd_export *exp)
6603 struct mdt_export_data *med = &exp->exp_mdt_data;
6607 INIT_LIST_HEAD(&med->med_open_head);
6608 spin_lock_init(&med->med_open_lock);
6609 spin_lock(&exp->exp_lock);
6610 exp->exp_connecting = 1;
6611 spin_unlock(&exp->exp_lock);
6613 OBD_ALLOC(exp->exp_used_slots,
6614 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
6615 if (exp->exp_used_slots == NULL)
6618 /* self-export doesn't need client data and ldlm initialization */
6619 if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
6620 &exp->exp_client_uuid)))
6623 rc = tgt_client_alloc(exp);
6627 rc = ldlm_init_export(exp);
6634 tgt_client_free(exp);
6636 OBD_FREE(exp->exp_used_slots,
6637 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
6638 exp->exp_used_slots = NULL;
6640 CERROR("%s: Failed to initialize export: rc = %d\n",
6641 exp->exp_obd->obd_name, rc);
6645 static int mdt_destroy_export(struct obd_export *exp)
6649 target_destroy_export(exp);
6650 if (exp->exp_used_slots)
6651 OBD_FREE(exp->exp_used_slots,
6652 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
6654 /* destroy can be called from failed obd_setup, so
6655 * checking uuid is safer than obd_self_export */
6656 if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
6657 &exp->exp_client_uuid)))
6660 ldlm_destroy_export(exp);
6661 tgt_client_free(exp);
6663 LASSERT(list_empty(&exp->exp_outstanding_replies));
6664 LASSERT(list_empty(&exp->exp_mdt_data.med_open_head));
6667 * discard grants once we're sure no more
6668 * interaction with the client is possible
6670 tgt_grant_discard(exp);
6671 if (exp_connect_flags(exp) & OBD_CONNECT_GRANT)
6672 exp->exp_obd->u.obt.obt_lut->lut_tgd.tgd_tot_granted_clients--;
6674 if (!(exp->exp_flags & OBD_OPT_FORCE))
6675 tgt_grant_sanity_check(exp->exp_obd, __func__);
6680 int mdt_links_read(struct mdt_thread_info *info, struct mdt_object *mdt_obj,
6681 struct linkea_data *ldata)
6685 LASSERT(ldata->ld_buf->lb_buf != NULL);
6687 if (!mdt_object_exists(mdt_obj))
6690 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
6691 ldata->ld_buf, XATTR_NAME_LINK);
6692 if (rc == -ERANGE) {
6693 /* Buf was too small, figure out what we need. */
6694 lu_buf_free(ldata->ld_buf);
6695 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
6696 ldata->ld_buf, XATTR_NAME_LINK);
6699 ldata->ld_buf = lu_buf_check_and_alloc(ldata->ld_buf, rc);
6700 if (ldata->ld_buf->lb_buf == NULL)
6702 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
6703 ldata->ld_buf, XATTR_NAME_LINK);
6708 return linkea_init_with_rec(ldata);
6712 * Given an MDT object, try to look up the full path to the object.
6713 * Part of the MDT layer implementation of lfs fid2path.
6715 * \param[in] info Per-thread common data shared by MDT level handlers.
6716 * \param[in] obj Object to do path lookup of
6717 * \param[in,out] fp User-provided struct to store path information
6718 * \param[in] root_fid Root FID of current path should reach
6720 * \retval 0 Lookup successful, path information stored in fp
6721 * \retval -EAGAIN Lookup failed, usually because object is being moved
6722 * \retval negative errno if there was a problem
6724 static int mdt_path_current(struct mdt_thread_info *info,
6725 struct mdt_object *obj,
6726 struct getinfo_fid2path *fp,
6727 struct lu_fid *root_fid)
6729 struct mdt_device *mdt = info->mti_mdt;
6730 struct lu_name *tmpname = &info->mti_name;
6731 struct lu_fid *tmpfid = &info->mti_tmp_fid1;
6732 struct lu_buf *buf = &info->mti_big_buf;
6733 struct md_attr *ma = &info->mti_attr;
6734 struct linkea_data ldata = { NULL };
6736 struct mdt_object *mdt_obj;
6737 struct link_ea_header *leh;
6738 struct link_ea_entry *lee;
6745 /* temp buffer for path element, the buffer will be finally freed
6746 * in mdt_thread_info_fini */
6747 buf = lu_buf_check_and_alloc(buf, PATH_MAX);
6748 if (buf->lb_buf == NULL)
6752 ptr = fp->gf_u.gf_path + fp->gf_pathlen - 1;
6755 *tmpfid = fp->gf_fid = *mdt_object_fid(obj);
6757 while (!lu_fid_eq(root_fid, &fp->gf_fid)) {
6758 if (!lu_fid_eq(root_fid, &mdt->mdt_md_root_fid) &&
6759 lu_fid_eq(&mdt->mdt_md_root_fid, &fp->gf_fid))
6760 GOTO(out, rc = -ENOENT);
6762 if (lu_fid_eq(mdt_object_fid(obj), tmpfid)) {
6764 mdt_object_get(info->mti_env, mdt_obj);
6766 mdt_obj = mdt_object_find(info->mti_env, mdt, tmpfid);
6767 if (IS_ERR(mdt_obj))
6768 GOTO(out, rc = PTR_ERR(mdt_obj));
6771 if (!mdt_object_exists(mdt_obj)) {
6772 mdt_object_put(info->mti_env, mdt_obj);
6773 GOTO(out, rc = -ENOENT);
6776 if (mdt_object_remote(mdt_obj)) {
6777 mdt_object_put(info->mti_env, mdt_obj);
6778 GOTO(remote_out, rc = -EREMOTE);
6781 rc = mdt_links_read(info, mdt_obj, &ldata);
6783 mdt_object_put(info->mti_env, mdt_obj);
6788 lee = (struct link_ea_entry *)(leh + 1); /* link #0 */
6789 linkea_entry_unpack(lee, &reclen, tmpname, tmpfid);
6790 /* If set, use link #linkno for path lookup, otherwise use
6791 link #0. Only do this for the final path element. */
6792 if (first && fp->gf_linkno < leh->leh_reccount) {
6794 for (count = 0; count < fp->gf_linkno; count++) {
6795 lee = (struct link_ea_entry *)
6796 ((char *)lee + reclen);
6797 linkea_entry_unpack(lee, &reclen, tmpname,
6800 if (fp->gf_linkno < leh->leh_reccount - 1)
6801 /* indicate to user there are more links */
6805 /* Check if it is slave stripes */
6806 rc = mdt_stripe_get(info, mdt_obj, ma, XATTR_NAME_LMV);
6807 mdt_object_put(info->mti_env, mdt_obj);
6811 if (ma->ma_valid & MA_LMV) {
6812 struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
6814 if (!lmv_is_sane2(lmv))
6815 GOTO(out, rc = -EBADF);
6817 /* For slave stripes, get its master */
6818 if (le32_to_cpu(lmv->lmv_magic) == LMV_MAGIC_STRIPE) {
6819 fp->gf_fid = *tmpfid;
6824 /* Pack the name in the end of the buffer */
6825 ptr -= tmpname->ln_namelen;
6826 if (ptr - 1 <= fp->gf_u.gf_path)
6827 GOTO(out, rc = -EOVERFLOW);
6828 strncpy(ptr, tmpname->ln_name, tmpname->ln_namelen);
6831 /* keep the last resolved fid to the client, so the
6832 * client will build the left path on another MDT for
6834 fp->gf_fid = *tmpfid;
6839 /* non-zero will be treated as an error */
6843 ptr++; /* skip leading / */
6844 memmove(fp->gf_u.gf_path, ptr,
6845 fp->gf_u.gf_path + fp->gf_pathlen - ptr);
6852 * Given an MDT object, use mdt_path_current to get the path.
6853 * Essentially a wrapper to retry mdt_path_current a set number of times
6854 * if -EAGAIN is returned (usually because an object is being moved).
6856 * Part of the MDT layer implementation of lfs fid2path.
6858 * \param[in] info Per-thread common data shared by mdt level handlers.
6859 * \param[in] obj Object to do path lookup of
6860 * \param[in,out] fp User-provided struct for arguments and to store path
6863 * \retval 0 Lookup successful, path information stored in fp
6864 * \retval negative errno if there was a problem
6866 static int mdt_path(struct mdt_thread_info *info, struct mdt_object *obj,
6867 struct getinfo_fid2path *fp, struct lu_fid *root_fid)
6869 struct mdt_device *mdt = info->mti_mdt;
6874 if (fp->gf_pathlen < 3)
6877 if (root_fid == NULL)
6878 root_fid = &mdt->mdt_md_root_fid;
6880 if (lu_fid_eq(root_fid, mdt_object_fid(obj))) {
6881 fp->gf_u.gf_path[0] = '\0';
6885 /* Retry multiple times in case file is being moved */
6886 while (tries-- && rc == -EAGAIN)
6887 rc = mdt_path_current(info, obj, fp, root_fid);
6893 * Get the full path of the provided FID, as of changelog record recno.
6895 * This checks sanity and looks up object for user provided FID
6896 * before calling the actual path lookup code.
6898 * Part of the MDT layer implementation of lfs fid2path.
6900 * \param[in] info Per-thread common data shared by mdt level handlers.
6901 * \param[in,out] fp User-provided struct for arguments and to store path
6904 * \retval 0 Lookup successful, path information and recno stored in fp
6905 * \retval -ENOENT, object does not exist
6906 * \retval negative errno if there was a problem
6908 static int mdt_fid2path(struct mdt_thread_info *info,
6909 struct lu_fid *root_fid,
6910 struct getinfo_fid2path *fp)
6912 struct mdt_device *mdt = info->mti_mdt;
6913 struct mdt_object *obj;
6917 CDEBUG(D_IOCTL, "path get "DFID" from %llu #%d\n",
6918 PFID(&fp->gf_fid), fp->gf_recno, fp->gf_linkno);
6920 if (!fid_is_sane(&fp->gf_fid))
6923 if (!fid_is_namespace_visible(&fp->gf_fid)) {
6924 CDEBUG(D_INFO, "%s: "DFID" is invalid, f_seq should be >= %#llx"
6925 ", or f_oid != 0, or f_ver == 0\n", mdt_obd_name(mdt),
6926 PFID(&fp->gf_fid), (__u64)FID_SEQ_NORMAL);
6930 obj = mdt_object_find(info->mti_env, mdt, &fp->gf_fid);
6933 CDEBUG(D_IOCTL, "cannot find "DFID": rc = %d\n",
6934 PFID(&fp->gf_fid), rc);
6938 if (mdt_object_remote(obj))
6940 else if (!mdt_object_exists(obj))
6946 mdt_object_put(info->mti_env, obj);
6947 CDEBUG(D_IOCTL, "nonlocal object "DFID": rc = %d\n",
6948 PFID(&fp->gf_fid), rc);
6952 rc = mdt_path(info, obj, fp, root_fid);
6954 CDEBUG(D_INFO, "fid "DFID", path %s recno %#llx linkno %u\n",
6955 PFID(&fp->gf_fid), fp->gf_u.gf_path,
6956 fp->gf_recno, fp->gf_linkno);
6958 mdt_object_put(info->mti_env, obj);
6963 static int mdt_rpc_fid2path(struct mdt_thread_info *info, void *key, int keylen,
6964 void *val, int vallen)
6966 struct getinfo_fid2path *fpout, *fpin;
6967 struct lu_fid *root_fid = NULL;
6970 fpin = key + cfs_size_round(sizeof(KEY_FID2PATH));
6973 if (ptlrpc_req_need_swab(info->mti_pill->rc_req))
6974 lustre_swab_fid2path(fpin);
6976 memcpy(fpout, fpin, sizeof(*fpin));
6977 if (fpout->gf_pathlen != vallen - sizeof(*fpin))
6980 if (keylen >= cfs_size_round(sizeof(KEY_FID2PATH)) + sizeof(*fpin) +
6981 sizeof(struct lu_fid)) {
6982 /* client sent its root FID, which is normally fileset FID */
6983 root_fid = fpin->gf_u.gf_root_fid;
6984 if (ptlrpc_req_need_swab(info->mti_pill->rc_req))
6985 lustre_swab_lu_fid(root_fid);
6987 if (root_fid != NULL && !fid_is_sane(root_fid))
6991 rc = mdt_fid2path(info, root_fid, fpout);
6995 int mdt_get_info(struct tgt_session_info *tsi)
7005 key = req_capsule_client_get(tsi->tsi_pill, &RMF_GETINFO_KEY);
7007 CDEBUG(D_IOCTL, "No GETINFO key\n");
7008 RETURN(err_serious(-EFAULT));
7010 keylen = req_capsule_get_size(tsi->tsi_pill, &RMF_GETINFO_KEY,
7013 vallen = req_capsule_client_get(tsi->tsi_pill, &RMF_GETINFO_VALLEN);
7014 if (vallen == NULL) {
7015 CDEBUG(D_IOCTL, "%s: cannot get RMF_GETINFO_VALLEN buffer\n",
7016 tgt_name(tsi->tsi_tgt));
7017 RETURN(err_serious(-EFAULT));
7020 req_capsule_set_size(tsi->tsi_pill, &RMF_GETINFO_VAL, RCL_SERVER,
7022 rc = req_capsule_server_pack(tsi->tsi_pill);
7024 RETURN(err_serious(rc));
7026 valout = req_capsule_server_get(tsi->tsi_pill, &RMF_GETINFO_VAL);
7027 if (valout == NULL) {
7028 CDEBUG(D_IOCTL, "%s: cannot get get-info RPC out buffer\n",
7029 tgt_name(tsi->tsi_tgt));
7030 RETURN(err_serious(-EFAULT));
7033 if (KEY_IS(KEY_FID2PATH)) {
7034 struct mdt_thread_info *info = tsi2mdt_info(tsi);
7036 rc = mdt_rpc_fid2path(info, key, keylen, valout, *vallen);
7037 mdt_thread_info_fini(info);
7044 static int mdt_ioc_version_get(struct mdt_thread_info *mti, void *karg)
7046 struct obd_ioctl_data *data = karg;
7049 struct mdt_object *obj;
7050 struct mdt_lock_handle *lh;
7054 if (data->ioc_inlbuf1 == NULL || data->ioc_inllen1 != sizeof(*fid) ||
7055 data->ioc_inlbuf2 == NULL || data->ioc_inllen2 != sizeof(version))
7058 fid = (struct lu_fid *)data->ioc_inlbuf1;
7060 if (!fid_is_sane(fid))
7063 CDEBUG(D_IOCTL, "getting version for "DFID"\n", PFID(fid));
7065 lh = &mti->mti_lh[MDT_LH_PARENT];
7066 mdt_lock_reg_init(lh, LCK_CR);
7068 obj = mdt_object_find_lock(mti, fid, lh, MDS_INODELOCK_UPDATE);
7070 RETURN(PTR_ERR(obj));
7072 if (mdt_object_remote(obj)) {
7075 * before calling version get the correct MDS should be
7076 * fid, this is error to find remote object here
7078 CERROR("nonlocal object "DFID"\n", PFID(fid));
7079 } else if (!mdt_object_exists(obj)) {
7080 *(__u64 *)data->ioc_inlbuf2 = ENOENT_VERSION;
7083 version = dt_version_get(mti->mti_env, mdt_obj2dt(obj));
7084 *(__u64 *)data->ioc_inlbuf2 = version;
7087 mdt_object_unlock_put(mti, obj, lh, 1);
7091 /* ioctls on obd dev */
7092 static int mdt_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
7093 void *karg, void __user *uarg)
7096 struct obd_device *obd = exp->exp_obd;
7097 struct mdt_device *mdt = mdt_dev(obd->obd_lu_dev);
7098 struct dt_device *dt = mdt->mdt_bottom;
7102 CDEBUG(D_IOCTL, "handling ioctl cmd %#x\n", cmd);
7103 rc = lu_env_init(&env, LCT_MD_THREAD);
7109 rc = mdt_device_sync(&env, mdt);
7111 case OBD_IOC_SET_READONLY:
7112 rc = dt_sync(&env, dt);
7114 rc = dt_ro(&env, dt);
7116 case OBD_IOC_ABORT_RECOVERY: {
7117 struct obd_ioctl_data *data = karg;
7119 CERROR("%s: Aborting recovery for device\n", mdt_obd_name(mdt));
7120 if (data->ioc_type & OBD_FLG_ABORT_RECOV_MDT)
7121 obd->obd_abort_recov_mdt = 1;
7122 else /* if (data->ioc_type & OBD_FLG_ABORT_RECOV_OST) */
7123 /* lctl didn't set OBD_FLG_ABORT_RECOV_OST < 2.13.57 */
7124 obd->obd_abort_recovery = 1;
7126 target_stop_recovery_thread(obd);
7130 case OBD_IOC_CHANGELOG_REG:
7131 case OBD_IOC_CHANGELOG_DEREG:
7132 case OBD_IOC_CHANGELOG_CLEAR:
7133 rc = mdt->mdt_child->md_ops->mdo_iocontrol(&env,
7137 case OBD_IOC_START_LFSCK: {
7138 struct md_device *next = mdt->mdt_child;
7139 struct obd_ioctl_data *data = karg;
7140 struct lfsck_start_param lsp;
7142 if (unlikely(data == NULL)) {
7147 lsp.lsp_start = (struct lfsck_start *)(data->ioc_inlbuf1);
7148 lsp.lsp_index_valid = 0;
7149 rc = next->md_ops->mdo_iocontrol(&env, next, cmd, 0, &lsp);
7152 case OBD_IOC_STOP_LFSCK: {
7153 struct md_device *next = mdt->mdt_child;
7154 struct obd_ioctl_data *data = karg;
7155 struct lfsck_stop stop;
7157 stop.ls_status = LS_STOPPED;
7158 /* Old lfsck utils may pass NULL @stop. */
7159 if (data->ioc_inlbuf1 == NULL)
7163 ((struct lfsck_stop *)(data->ioc_inlbuf1))->ls_flags;
7165 rc = next->md_ops->mdo_iocontrol(&env, next, cmd, 0, &stop);
7168 case OBD_IOC_QUERY_LFSCK: {
7169 struct md_device *next = mdt->mdt_child;
7170 struct obd_ioctl_data *data = karg;
7172 rc = next->md_ops->mdo_iocontrol(&env, next, cmd, 0,
7176 case OBD_IOC_GET_OBJ_VERSION: {
7177 struct mdt_thread_info *mti;
7178 mti = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
7179 memset(mti, 0, sizeof *mti);
7180 mti->mti_env = &env;
7184 rc = mdt_ioc_version_get(mti, karg);
7187 case OBD_IOC_CATLOGLIST: {
7188 struct mdt_thread_info *mti;
7190 mti = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
7191 lu_local_obj_fid(&mti->mti_tmp_fid1, LLOG_CATALOGS_OID);
7192 rc = llog_catalog_list(&env, mdt->mdt_bottom, 0, karg,
7193 &mti->mti_tmp_fid1);
7198 CERROR("%s: Not supported cmd = %d, rc = %d\n",
7199 mdt_obd_name(mdt), cmd, rc);
7206 static int mdt_postrecov(const struct lu_env *env, struct mdt_device *mdt)
7208 struct lu_device *ld = md2lu_dev(mdt->mdt_child);
7212 if (!mdt->mdt_skip_lfsck && !mdt->mdt_bottom->dd_rdonly) {
7213 struct lfsck_start_param lsp;
7215 lsp.lsp_start = NULL;
7216 lsp.lsp_index_valid = 0;
7217 rc = mdt->mdt_child->md_ops->mdo_iocontrol(env, mdt->mdt_child,
7218 OBD_IOC_START_LFSCK,
7220 if (rc != 0 && rc != -EALREADY)
7221 CWARN("%s: auto trigger paused LFSCK failed: rc = %d\n",
7222 mdt_obd_name(mdt), rc);
7225 rc = ld->ld_ops->ldo_recovery_complete(env, ld);
7229 static int mdt_obd_postrecov(struct obd_device *obd)
7234 rc = lu_env_init(&env, LCT_MD_THREAD);
7237 rc = mdt_postrecov(&env, mdt_dev(obd->obd_lu_dev));
7242 static const struct obd_ops mdt_obd_device_ops = {
7243 .o_owner = THIS_MODULE,
7244 .o_set_info_async = mdt_obd_set_info_async,
7245 .o_connect = mdt_obd_connect,
7246 .o_reconnect = mdt_obd_reconnect,
7247 .o_disconnect = mdt_obd_disconnect,
7248 .o_init_export = mdt_init_export,
7249 .o_destroy_export = mdt_destroy_export,
7250 .o_iocontrol = mdt_iocontrol,
7251 .o_postrecov = mdt_obd_postrecov,
7252 /* Data-on-MDT IO methods */
7253 .o_preprw = mdt_obd_preprw,
7254 .o_commitrw = mdt_obd_commitrw,
7257 static struct lu_device* mdt_device_fini(const struct lu_env *env,
7258 struct lu_device *d)
7260 struct mdt_device *m = mdt_dev(d);
7267 static struct lu_device *mdt_device_free(const struct lu_env *env,
7268 struct lu_device *d)
7270 struct mdt_device *m = mdt_dev(d);
7273 lu_device_fini(&m->mdt_lu_dev);
7279 static struct lu_device *mdt_device_alloc(const struct lu_env *env,
7280 struct lu_device_type *t,
7281 struct lustre_cfg *cfg)
7283 struct lu_device *l;
7284 struct mdt_device *m;
7291 rc = mdt_init0(env, m, t, cfg);
7293 mdt_device_free(env, l);
7298 l = ERR_PTR(-ENOMEM);
7302 /* context key constructor/destructor: mdt_key_init, mdt_key_fini */
7303 LU_KEY_INIT(mdt, struct mdt_thread_info);
7305 static void mdt_key_fini(const struct lu_context *ctx,
7306 struct lu_context_key *key, void* data)
7308 struct mdt_thread_info *info = data;
7310 if (info->mti_big_lmm) {
7311 OBD_FREE_LARGE(info->mti_big_lmm, info->mti_big_lmmsize);
7312 info->mti_big_lmm = NULL;
7313 info->mti_big_lmmsize = 0;
7316 if (info->mti_big_acl) {
7317 OBD_FREE_LARGE(info->mti_big_acl, info->mti_big_aclsize);
7318 info->mti_big_acl = NULL;
7319 info->mti_big_aclsize = 0;
7325 /* context key: mdt_thread_key */
7326 LU_CONTEXT_KEY_DEFINE(mdt, LCT_MD_THREAD);
7328 struct lu_ucred *mdt_ucred(const struct mdt_thread_info *info)
7330 return lu_ucred(info->mti_env);
7333 struct lu_ucred *mdt_ucred_check(const struct mdt_thread_info *info)
7335 return lu_ucred_check(info->mti_env);
7339 * Enable/disable COS (Commit On Sharing).
7341 * Set/Clear the COS flag in mdt options.
7343 * \param mdt mdt device
7344 * \param val 0 disables COS, other values enable COS
7346 void mdt_enable_cos(struct mdt_device *mdt, bool val)
7351 mdt->mdt_opts.mo_cos = val;
7352 rc = lu_env_init(&env, LCT_LOCAL);
7353 if (unlikely(rc != 0)) {
7354 CWARN("%s: lu_env initialization failed, cannot "
7355 "sync: rc = %d\n", mdt_obd_name(mdt), rc);
7358 mdt_device_sync(&env, mdt);
7363 * Check COS (Commit On Sharing) status.
7365 * Return COS flag status.
7367 * \param mdt mdt device
7369 int mdt_cos_is_enabled(struct mdt_device *mdt)
7371 return mdt->mdt_opts.mo_cos != 0;
7374 static const struct lu_device_type_operations mdt_device_type_ops = {
7375 .ldto_device_alloc = mdt_device_alloc,
7376 .ldto_device_free = mdt_device_free,
7377 .ldto_device_fini = mdt_device_fini
7380 static struct lu_device_type mdt_device_type = {
7381 .ldt_tags = LU_DEVICE_MD,
7382 .ldt_name = LUSTRE_MDT_NAME,
7383 .ldt_ops = &mdt_device_type_ops,
7384 .ldt_ctx_tags = LCT_MD_THREAD
7387 static int __init mdt_init(void)
7391 BUILD_BUG_ON(sizeof("0x0123456789ABCDEF:0x01234567:0x01234567") !=
7392 FID_NOBRACE_LEN + 1);
7393 BUILD_BUG_ON(sizeof("[0x0123456789ABCDEF:0x01234567:0x01234567]") !=
7395 rc = lu_kmem_init(mdt_caches);
7399 rc = mds_mod_init();
7403 rc = class_register_type(&mdt_obd_device_ops, NULL, true,
7404 LUSTRE_MDT_NAME, &mdt_device_type);
7409 lu_kmem_fini(mdt_caches);
7416 static void __exit mdt_exit(void)
7418 class_unregister_type(LUSTRE_MDT_NAME);
7420 lu_kmem_fini(mdt_caches);
7423 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
7424 MODULE_DESCRIPTION("Lustre Metadata Target ("LUSTRE_MDT_NAME")");
7425 MODULE_VERSION(LUSTRE_VERSION_STRING);
7426 MODULE_LICENSE("GPL");
7428 module_init(mdt_init);
7429 module_exit(mdt_exit);