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>
66 #include <lustre_crypto.h>
68 #include "mdt_internal.h"
70 static unsigned int max_mod_rpcs_per_client = 8;
71 module_param(max_mod_rpcs_per_client, uint, 0644);
72 MODULE_PARM_DESC(max_mod_rpcs_per_client, "maximum number of modify RPCs in flight allowed per client");
74 mdl_mode_t mdt_mdl_lock_modes[] = {
75 [LCK_MINMODE] = MDL_MINMODE,
82 [LCK_GROUP] = MDL_GROUP
85 enum ldlm_mode mdt_dlm_lock_modes[] = {
86 [MDL_MINMODE] = LCK_MINMODE,
93 [MDL_GROUP] = LCK_GROUP
96 static struct mdt_device *mdt_dev(struct lu_device *d);
98 static const struct lu_object_operations mdt_obj_ops;
100 /* Slab for MDT object allocation */
101 static struct kmem_cache *mdt_object_kmem;
103 /* For HSM restore handles */
104 struct kmem_cache *mdt_hsm_cdt_kmem;
106 /* For HSM request handles */
107 struct kmem_cache *mdt_hsm_car_kmem;
109 static struct lu_kmem_descr mdt_caches[] = {
111 .ckd_cache = &mdt_object_kmem,
112 .ckd_name = "mdt_obj",
113 .ckd_size = sizeof(struct mdt_object)
116 .ckd_cache = &mdt_hsm_cdt_kmem,
117 .ckd_name = "mdt_cdt_restore_handle",
118 .ckd_size = sizeof(struct cdt_restore_handle)
121 .ckd_cache = &mdt_hsm_car_kmem,
122 .ckd_name = "mdt_cdt_agent_req",
123 .ckd_size = sizeof(struct cdt_agent_req)
130 __u64 mdt_get_disposition(struct ldlm_reply *rep, __u64 op_flag)
134 return rep->lock_policy_res1 & op_flag;
137 void mdt_clear_disposition(struct mdt_thread_info *info,
138 struct ldlm_reply *rep, __u64 op_flag)
141 info->mti_opdata &= ~op_flag;
142 tgt_opdata_clear(info->mti_env, op_flag);
145 rep->lock_policy_res1 &= ~op_flag;
148 void mdt_set_disposition(struct mdt_thread_info *info,
149 struct ldlm_reply *rep, __u64 op_flag)
152 info->mti_opdata |= op_flag;
153 tgt_opdata_set(info->mti_env, op_flag);
156 rep->lock_policy_res1 |= op_flag;
159 void mdt_lock_reg_init(struct mdt_lock_handle *lh, enum ldlm_mode lm)
161 lh->mlh_pdo_hash = 0;
162 lh->mlh_reg_mode = lm;
163 lh->mlh_rreg_mode = lm;
164 lh->mlh_type = MDT_REG_LOCK;
167 void mdt_lh_reg_init(struct mdt_lock_handle *lh, struct ldlm_lock *lock)
169 mdt_lock_reg_init(lh, lock->l_req_mode);
170 if (lock->l_req_mode == LCK_GROUP)
171 lh->mlh_gid = lock->l_policy_data.l_inodebits.li_gid;
174 void mdt_lock_pdo_init(struct mdt_lock_handle *lh, enum ldlm_mode lock_mode,
175 const struct lu_name *lname)
177 lh->mlh_reg_mode = lock_mode;
178 lh->mlh_pdo_mode = LCK_MINMODE;
179 lh->mlh_rreg_mode = lock_mode;
180 lh->mlh_type = MDT_PDO_LOCK;
182 if (lu_name_is_valid(lname)) {
183 lh->mlh_pdo_hash = ll_full_name_hash(NULL, lname->ln_name,
185 /* XXX Workaround for LU-2856
187 * Zero is a valid return value of full_name_hash, but
188 * several users of mlh_pdo_hash assume a non-zero
189 * hash value. We therefore map zero onto an
190 * arbitrary, but consistent value (1) to avoid
191 * problems further down the road. */
192 if (unlikely(lh->mlh_pdo_hash == 0))
193 lh->mlh_pdo_hash = 1;
195 lh->mlh_pdo_hash = 0;
199 static void mdt_lock_pdo_mode(struct mdt_thread_info *info, struct mdt_object *o,
200 struct mdt_lock_handle *lh)
206 * Any dir access needs couple of locks:
208 * 1) on part of dir we gonna take lookup/modify;
210 * 2) on whole dir to protect it from concurrent splitting and/or to
211 * flush client's cache for readdir().
213 * so, for a given mode and object this routine decides what lock mode
214 * to use for lock #2:
216 * 1) if caller's gonna lookup in dir then we need to protect dir from
217 * being splitted only - LCK_CR
219 * 2) if caller's gonna modify dir then we need to protect dir from
220 * being splitted and to flush cache - LCK_CW
222 * 3) if caller's gonna modify dir and that dir seems ready for
223 * splitting then we need to protect it from any type of access
224 * (lookup/modify/split) - LCK_EX --bzzz
227 LASSERT(lh->mlh_reg_mode != LCK_MINMODE);
228 LASSERT(lh->mlh_pdo_mode == LCK_MINMODE);
231 * Ask underlaying level its opinion about preferable PDO lock mode
232 * having access type passed as regular lock mode:
234 * - MDL_MINMODE means that lower layer does not want to specify lock
237 * - MDL_NL means that no PDO lock should be taken. This is used in some
238 * cases. Say, for non-splittable directories no need to use PDO locks
241 mode = mdo_lock_mode(info->mti_env, mdt_object_child(o),
242 mdt_dlm_mode2mdl_mode(lh->mlh_reg_mode));
244 if (mode != MDL_MINMODE) {
245 lh->mlh_pdo_mode = mdt_mdl_mode2dlm_mode(mode);
248 * Lower layer does not want to specify locking mode. We do it
249 * our selves. No special protection is needed, just flush
250 * client's cache on modification and allow concurrent
253 switch (lh->mlh_reg_mode) {
255 lh->mlh_pdo_mode = LCK_EX;
258 lh->mlh_pdo_mode = LCK_CR;
261 lh->mlh_pdo_mode = LCK_CW;
264 CERROR("Not expected lock type (0x%x)\n",
265 (int)lh->mlh_reg_mode);
270 LASSERT(lh->mlh_pdo_mode != LCK_MINMODE);
275 * Check whether \a o is directory stripe object.
277 * \param[in] info thread environment
278 * \param[in] o MDT object
280 * \retval 1 is directory stripe.
281 * \retval 0 isn't directory stripe.
282 * \retval < 1 error code
284 static int mdt_is_dir_stripe(struct mdt_thread_info *info,
285 struct mdt_object *o)
287 struct md_attr *ma = &info->mti_attr;
288 struct lmv_mds_md_v1 *lmv;
291 rc = mdt_stripe_get(info, o, ma, XATTR_NAME_LMV);
295 if (!(ma->ma_valid & MA_LMV))
298 lmv = &ma->ma_lmv->lmv_md_v1;
300 if (!lmv_is_sane2(lmv))
303 if (le32_to_cpu(lmv->lmv_magic) == LMV_MAGIC_STRIPE)
309 static int mdt_lookup_fileset(struct mdt_thread_info *info, const char *fileset,
312 struct mdt_device *mdt = info->mti_mdt;
313 struct lu_name *lname = &info->mti_name;
314 const char *start = fileset;
315 char *filename = info->mti_filename;
316 struct mdt_object *parent;
320 LASSERT(!info->mti_cross_ref);
323 * We may want to allow this to mount a completely separate
324 * fileset from the MDT in the future, but keeping it to
325 * ROOT/ only for now avoid potential security issues.
327 *fid = mdt->mdt_md_root_fid;
329 while (rc == 0 && start != NULL && *start != '\0') {
330 const char *s1 = start;
336 while (*s2 != '/' && *s2 != '\0')
344 lname->ln_namelen = s2 - s1;
345 if (lname->ln_namelen > NAME_MAX) {
350 /* reject .. as a path component */
351 if (lname->ln_namelen == 2 &&
352 strncmp(s1, "..", 2) == 0) {
357 strncpy(filename, s1, lname->ln_namelen);
358 filename[lname->ln_namelen] = '\0';
359 lname->ln_name = filename;
361 parent = mdt_object_find(info->mti_env, mdt, fid);
362 if (IS_ERR(parent)) {
363 rc = PTR_ERR(parent);
366 /* Only got the fid of this obj by name */
368 rc = mdo_lookup(info->mti_env, mdt_object_child(parent), lname,
369 fid, &info->mti_spec);
370 mdt_object_put(info->mti_env, parent);
373 parent = mdt_object_find(info->mti_env, mdt, fid);
375 rc = PTR_ERR(parent);
377 mode = lu_object_attr(&parent->mot_obj);
378 if (!S_ISDIR(mode)) {
380 } else if (mdt_is_remote_object(info, parent, parent)) {
381 if (!mdt->mdt_enable_remote_subdir_mount) {
383 LCONSOLE_WARN("%s: subdir mount '%s' refused because 'enable_remote_subdir_mount=0': rc = %d\n",
387 LCONSOLE_INFO("%s: subdir mount '%s' is remote and may be slow\n",
392 mdt_object_put(info->mti_env, parent);
399 static int mdt_get_root(struct tgt_session_info *tsi)
401 struct mdt_thread_info *info = tsi2mdt_info(tsi);
402 struct mdt_device *mdt = info->mti_mdt;
403 struct mdt_body *repbody;
404 char *fileset = NULL, *buffer = NULL;
406 struct obd_export *exp = info->mti_exp;
407 char *nodemap_fileset;
411 rc = mdt_check_ucred(info);
413 GOTO(out, rc = err_serious(rc));
415 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GET_ROOT_PACK))
416 GOTO(out, rc = err_serious(-ENOMEM));
418 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
419 if (req_capsule_get_size(info->mti_pill, &RMF_NAME, RCL_CLIENT) > 0) {
420 fileset = req_capsule_client_get(info->mti_pill, &RMF_NAME);
422 GOTO(out, rc = err_serious(-EFAULT));
425 nodemap_fileset = nodemap_get_fileset(exp->exp_target_data.ted_nodemap);
426 if (nodemap_fileset && nodemap_fileset[0]) {
427 CDEBUG(D_INFO, "nodemap fileset is %s\n", nodemap_fileset);
429 /* consider fileset from client as a sub-fileset
430 * of the nodemap one */
431 OBD_ALLOC(buffer, PATH_MAX + 1);
433 GOTO(out, rc = err_serious(-ENOMEM));
434 if (snprintf(buffer, PATH_MAX + 1, "%s/%s",
435 nodemap_fileset, fileset) >= PATH_MAX + 1)
436 GOTO(out, rc = err_serious(-EINVAL));
439 /* enforce fileset as specified in the nodemap */
440 fileset = nodemap_fileset;
445 CDEBUG(D_INFO, "Getting fileset %s\n", fileset);
446 rc = mdt_lookup_fileset(info, fileset, &repbody->mbo_fid1);
448 GOTO(out, rc = err_serious(rc));
450 repbody->mbo_fid1 = mdt->mdt_md_root_fid;
452 repbody->mbo_valid |= OBD_MD_FLID;
456 mdt_thread_info_fini(info);
458 OBD_FREE(buffer, PATH_MAX+1);
462 static int mdt_statfs(struct tgt_session_info *tsi)
464 struct ptlrpc_request *req = tgt_ses_req(tsi);
465 struct mdt_thread_info *info = tsi2mdt_info(tsi);
466 struct mdt_device *mdt = info->mti_mdt;
467 struct tg_grants_data *tgd = &mdt->mdt_lut.lut_tgd;
468 struct md_device *next = mdt->mdt_child;
469 struct ptlrpc_service_part *svcpt;
470 struct obd_statfs *osfs;
471 struct mdt_body *reqbody = NULL;
472 struct mdt_statfs_cache *msf;
473 ktime_t kstart = ktime_get();
474 int current_blockbits;
479 svcpt = req->rq_rqbd->rqbd_svcpt;
481 /* This will trigger a watchdog timeout */
482 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_STATFS_LCW_SLEEP,
483 (MDT_SERVICE_WATCHDOG_FACTOR *
484 at_get(&svcpt->scp_at_estimate)) + 1);
486 rc = mdt_check_ucred(info);
488 GOTO(out, rc = err_serious(rc));
490 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK))
491 GOTO(out, rc = err_serious(-ENOMEM));
493 osfs = req_capsule_server_get(info->mti_pill, &RMF_OBD_STATFS);
495 GOTO(out, rc = -EPROTO);
497 if (mdt_is_sum_statfs_client(req->rq_export) &&
498 lustre_packed_msg_size(req->rq_reqmsg) ==
499 req_capsule_fmt_size(req->rq_reqmsg->lm_magic,
500 &RQF_MDS_STATFS_NEW, RCL_CLIENT)) {
501 req_capsule_extend(info->mti_pill, &RQF_MDS_STATFS_NEW);
502 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
505 if (reqbody && reqbody->mbo_valid & OBD_MD_FLAGSTATFS)
506 msf = &mdt->mdt_sum_osfs;
508 msf = &mdt->mdt_osfs;
510 if (msf->msf_age + OBD_STATFS_CACHE_SECONDS <= ktime_get_seconds()) {
511 /** statfs data is too old, get up-to-date one */
512 if (reqbody && reqbody->mbo_valid & OBD_MD_FLAGSTATFS)
513 rc = next->md_ops->mdo_statfs(info->mti_env,
516 rc = dt_statfs(info->mti_env, mdt->mdt_bottom,
520 spin_lock(&mdt->mdt_lock);
521 msf->msf_osfs = *osfs;
522 msf->msf_age = ktime_get_seconds();
523 spin_unlock(&mdt->mdt_lock);
525 /** use cached statfs data */
526 spin_lock(&mdt->mdt_lock);
527 *osfs = msf->msf_osfs;
528 spin_unlock(&mdt->mdt_lock);
531 /* tgd_blockbit is recordsize bits set during mkfs.
532 * This once set does not change. However, 'zfs set'
533 * can be used to change the MDT blocksize. Instead
534 * of using cached value of 'tgd_blockbit' always
535 * calculate the blocksize bits which may have
538 current_blockbits = fls64(osfs->os_bsize) - 1;
540 /* at least try to account for cached pages. its still racy and
541 * might be under-reporting if clients haven't announced their
542 * caches with brw recently */
543 CDEBUG(D_SUPER | D_CACHE, "blocks cached %llu granted %llu"
544 " pending %llu free %llu avail %llu\n",
545 tgd->tgd_tot_dirty, tgd->tgd_tot_granted,
546 tgd->tgd_tot_pending,
547 osfs->os_bfree << current_blockbits,
548 osfs->os_bavail << current_blockbits);
550 osfs->os_bavail -= min_t(u64, osfs->os_bavail,
551 ((tgd->tgd_tot_dirty + tgd->tgd_tot_pending +
552 osfs->os_bsize - 1) >> current_blockbits));
554 tgt_grant_sanity_check(mdt->mdt_lu_dev.ld_obd, __func__);
555 CDEBUG(D_CACHE, "%llu blocks: %llu free, %llu avail; "
556 "%llu objects: %llu free; state %x\n",
557 osfs->os_blocks, osfs->os_bfree, osfs->os_bavail,
558 osfs->os_files, osfs->os_ffree, osfs->os_state);
560 if (!exp_grant_param_supp(tsi->tsi_exp) &&
561 current_blockbits > COMPAT_BSIZE_SHIFT) {
562 /* clients which don't support OBD_CONNECT_GRANT_PARAM
563 * should not see a block size > page size, otherwise
564 * cl_lost_grant goes mad. Therefore, we emulate a 4KB (=2^12)
565 * block size which is the biggest block size known to work
566 * with all client's page size. */
567 osfs->os_blocks <<= current_blockbits - COMPAT_BSIZE_SHIFT;
568 osfs->os_bfree <<= current_blockbits - COMPAT_BSIZE_SHIFT;
569 osfs->os_bavail <<= current_blockbits - COMPAT_BSIZE_SHIFT;
570 osfs->os_bsize = 1 << COMPAT_BSIZE_SHIFT;
573 mdt_counter_incr(req, LPROC_MDT_STATFS,
574 ktime_us_delta(ktime_get(), kstart));
576 mdt_thread_info_fini(info);
580 __u32 mdt_lmm_dom_entry_check(struct lov_mds_md *lmm, int *is_dom_only)
582 struct lov_comp_md_v1 *comp_v1;
583 struct lov_mds_md *v1;
585 __u32 dom_stripesize = 0;
587 bool has_ost_stripes = false;
594 if (le32_to_cpu(lmm->lmm_magic) != LOV_MAGIC_COMP_V1)
597 comp_v1 = (struct lov_comp_md_v1 *)lmm;
598 off = le32_to_cpu(comp_v1->lcm_entries[0].lcme_offset);
599 v1 = (struct lov_mds_md *)((char *)comp_v1 + off);
601 /* Fast check for DoM entry with no mirroring, should be the first */
602 if (le16_to_cpu(comp_v1->lcm_mirror_count) == 0 &&
603 lov_pattern(le32_to_cpu(v1->lmm_pattern)) != LOV_PATTERN_MDT)
606 /* check all entries otherwise */
607 for (i = 0; i < le16_to_cpu(comp_v1->lcm_entry_count); i++) {
608 struct lov_comp_md_entry_v1 *lcme;
610 lcme = &comp_v1->lcm_entries[i];
611 if (!(le32_to_cpu(lcme->lcme_flags) & LCME_FL_INIT))
614 off = le32_to_cpu(lcme->lcme_offset);
615 v1 = (struct lov_mds_md *)((char *)comp_v1 + off);
617 if (lov_pattern(le32_to_cpu(v1->lmm_pattern)) ==
619 dom_stripesize = le32_to_cpu(v1->lmm_stripe_size);
621 has_ost_stripes = true;
623 if (dom_stripesize && has_ost_stripes)
624 RETURN(dom_stripesize);
626 /* DoM-only case exits here */
627 if (is_dom_only && dom_stripesize)
629 RETURN(dom_stripesize);
633 * Pack size attributes into the reply.
635 int mdt_pack_size2body(struct mdt_thread_info *info,
636 const struct lu_fid *fid, struct lustre_handle *lh)
639 struct md_attr *ma = &info->mti_attr;
641 bool dom_lock = false;
645 LASSERT(ma->ma_attr.la_valid & LA_MODE);
647 if (!S_ISREG(ma->ma_attr.la_mode) ||
648 !(ma->ma_valid & MA_LOV && ma->ma_lmm != NULL))
651 dom_stripe = mdt_lmm_dom_stripesize(ma->ma_lmm);
652 /* no DoM stripe, no size in reply */
656 if (lustre_handle_is_used(lh)) {
657 struct ldlm_lock *lock;
659 lock = ldlm_handle2lock(lh);
661 dom_lock = ldlm_has_dom(lock);
666 /* no DoM lock, no size in reply */
670 /* Either DoM lock exists or LMM has only DoM stripe then
671 * return size on body. */
672 b = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
674 mdt_dom_object_size(info->mti_env, info->mti_mdt, fid, b, dom_lock);
678 #ifdef CONFIG_LUSTRE_FS_POSIX_ACL
680 * Pack ACL data into the reply. UIDs/GIDs are mapped and filtered by nodemap.
682 * \param info thread info object
683 * \param repbody reply to pack ACLs into
684 * \param o mdt object of file to examine
685 * \param nodemap nodemap of client to reply to
687 * \retval -errno error getting or parsing ACL from disk
689 int mdt_pack_acl2body(struct mdt_thread_info *info, struct mdt_body *repbody,
690 struct mdt_object *o, struct lu_nodemap *nodemap)
692 const struct lu_env *env = info->mti_env;
693 struct md_object *next = mdt_object_child(o);
694 struct lu_buf *buf = &info->mti_buf;
695 struct mdt_device *mdt = info->mti_mdt;
696 struct req_capsule *pill = info->mti_pill;
701 buf->lb_buf = req_capsule_server_get(pill, &RMF_ACL);
702 buf->lb_len = req_capsule_get_size(pill, &RMF_ACL, RCL_SERVER);
703 if (buf->lb_len == 0)
706 LASSERT(!info->mti_big_acl_used);
708 rc = mo_xattr_get(env, next, buf, XATTR_NAME_ACL_ACCESS);
710 if (rc == -ENODATA) {
711 repbody->mbo_aclsize = 0;
712 repbody->mbo_valid |= OBD_MD_FLACL;
714 } else if (rc == -EOPNOTSUPP) {
716 } else if (rc == -ERANGE) {
717 if (exp_connect_large_acl(info->mti_exp) &&
718 !info->mti_big_acl_used) {
719 if (info->mti_big_acl == NULL) {
720 info->mti_big_aclsize =
722 mdt->mdt_max_ea_size,
724 OBD_ALLOC_LARGE(info->mti_big_acl,
725 info->mti_big_aclsize);
726 if (info->mti_big_acl == NULL) {
727 info->mti_big_aclsize = 0;
728 CERROR("%s: unable to grow "
731 PFID(mdt_object_fid(o)));
736 CDEBUG(D_INODE, "%s: grow the "DFID
737 " ACL buffer to size %d\n",
739 PFID(mdt_object_fid(o)),
740 info->mti_big_aclsize);
742 buf->lb_buf = info->mti_big_acl;
743 buf->lb_len = info->mti_big_aclsize;
744 info->mti_big_acl_used = 1;
747 /* FS has ACL bigger that our limits */
748 CDEBUG(D_INODE, "%s: "DFID" ACL can't fit into %d\n",
749 mdt_obd_name(mdt), PFID(mdt_object_fid(o)),
750 info->mti_big_aclsize);
753 CERROR("%s: unable to read "DFID" ACL: rc = %d\n",
754 mdt_obd_name(mdt), PFID(mdt_object_fid(o)), rc);
757 rc = nodemap_map_acl(nodemap, buf->lb_buf,
758 rc, NODEMAP_FS_TO_CLIENT);
759 /* if all ACLs mapped out, rc is still >= 0 */
761 CERROR("%s: nodemap_map_acl unable to parse "DFID
762 " ACL: rc = %d\n", mdt_obd_name(mdt),
763 PFID(mdt_object_fid(o)), rc);
764 repbody->mbo_aclsize = 0;
765 repbody->mbo_valid &= ~OBD_MD_FLACL;
767 repbody->mbo_aclsize = rc;
768 repbody->mbo_valid |= OBD_MD_FLACL;
777 /* XXX Look into layout in MDT layer. */
778 static inline bool mdt_hsm_is_released(struct lov_mds_md *lmm)
780 struct lov_comp_md_v1 *comp_v1;
781 struct lov_mds_md *v1;
784 if (lmm->lmm_magic == LOV_MAGIC_COMP_V1) {
785 comp_v1 = (struct lov_comp_md_v1 *)lmm;
787 for (i = 0; i < comp_v1->lcm_entry_count; i++) {
788 v1 = (struct lov_mds_md *)((char *)comp_v1 +
789 comp_v1->lcm_entries[i].lcme_offset);
790 /* We don't support partial release for now */
791 if (!(v1->lmm_pattern & LOV_PATTERN_F_RELEASED))
796 return (lmm->lmm_pattern & LOV_PATTERN_F_RELEASED) ?
801 void mdt_pack_attr2body(struct mdt_thread_info *info, struct mdt_body *b,
802 const struct lu_attr *attr, const struct lu_fid *fid)
804 struct md_attr *ma = &info->mti_attr;
805 struct obd_export *exp = info->mti_exp;
806 struct lu_nodemap *nodemap = NULL;
808 LASSERT(ma->ma_valid & MA_INODE);
810 if (attr->la_valid & LA_ATIME) {
811 b->mbo_atime = attr->la_atime;
812 b->mbo_valid |= OBD_MD_FLATIME;
814 if (attr->la_valid & LA_MTIME) {
815 b->mbo_mtime = attr->la_mtime;
816 b->mbo_valid |= OBD_MD_FLMTIME;
818 if (attr->la_valid & LA_CTIME) {
819 b->mbo_ctime = attr->la_ctime;
820 b->mbo_valid |= OBD_MD_FLCTIME;
822 if (attr->la_valid & LA_BTIME) {
823 b->mbo_btime = attr->la_btime;
824 b->mbo_valid |= OBD_MD_FLBTIME;
826 if (attr->la_valid & LA_FLAGS) {
827 b->mbo_flags = attr->la_flags;
828 b->mbo_valid |= OBD_MD_FLFLAGS;
830 if (attr->la_valid & LA_NLINK) {
831 b->mbo_nlink = attr->la_nlink;
832 b->mbo_valid |= OBD_MD_FLNLINK;
834 if (attr->la_valid & (LA_UID|LA_GID)) {
835 nodemap = nodemap_get_from_exp(exp);
839 if (attr->la_valid & LA_UID) {
840 b->mbo_uid = nodemap_map_id(nodemap, NODEMAP_UID,
841 NODEMAP_FS_TO_CLIENT,
843 b->mbo_valid |= OBD_MD_FLUID;
845 if (attr->la_valid & LA_GID) {
846 b->mbo_gid = nodemap_map_id(nodemap, NODEMAP_GID,
847 NODEMAP_FS_TO_CLIENT,
849 b->mbo_valid |= OBD_MD_FLGID;
852 if (attr->la_valid & LA_PROJID) {
853 /* TODO, nodemap for project id */
854 b->mbo_projid = attr->la_projid;
855 b->mbo_valid |= OBD_MD_FLPROJID;
858 b->mbo_mode = attr->la_mode;
859 if (attr->la_valid & LA_MODE)
860 b->mbo_valid |= OBD_MD_FLMODE;
861 if (attr->la_valid & LA_TYPE)
862 b->mbo_valid |= OBD_MD_FLTYPE;
866 b->mbo_valid |= OBD_MD_FLID;
867 CDEBUG(D_INODE, DFID": nlink=%d, mode=%o, valid=%#llx\n",
868 PFID(fid), b->mbo_nlink, b->mbo_mode, b->mbo_valid);
871 if (!(attr->la_valid & LA_TYPE))
874 b->mbo_rdev = attr->la_rdev;
875 b->mbo_size = attr->la_size;
876 b->mbo_blocks = attr->la_blocks;
878 if (!S_ISREG(attr->la_mode)) {
879 b->mbo_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS | OBD_MD_FLRDEV;
880 } else if (ma->ma_need & MA_LOV && !(ma->ma_valid & MA_LOV)) {
881 /* means no objects are allocated on osts. */
882 LASSERT(!(ma->ma_valid & MA_LOV));
883 /* just ignore blocks occupied by extend attributes on MDS */
885 /* if no object is allocated on osts, the size on mds is valid.
887 b->mbo_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
888 } else if ((ma->ma_valid & MA_LOV) && ma->ma_lmm != NULL) {
889 if (mdt_hsm_is_released(ma->ma_lmm)) {
890 /* A released file stores its size on MDS. */
891 /* But return 1 block for released file, unless tools
892 * like tar will consider it fully sparse. (LU-3864)
894 if (unlikely(b->mbo_size == 0))
898 b->mbo_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
899 } else if (info->mti_som_valid) { /* som is valid */
900 b->mbo_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
901 } else if (ma->ma_valid & MA_SOM) { /* lsom is valid */
902 b->mbo_valid |= OBD_MD_FLLAZYSIZE | OBD_MD_FLLAZYBLOCKS;
903 b->mbo_size = ma->ma_som.ms_size;
904 b->mbo_blocks = ma->ma_som.ms_blocks;
908 if (fid != NULL && (b->mbo_valid & OBD_MD_FLSIZE ||
909 b->mbo_valid & OBD_MD_FLLAZYSIZE))
910 CDEBUG(D_VFSTRACE, DFID": returning size %llu\n",
911 PFID(fid), (unsigned long long)b->mbo_size);
914 if (!IS_ERR_OR_NULL(nodemap))
915 nodemap_putref(nodemap);
918 static inline int mdt_body_has_lov(const struct lu_attr *la,
919 const struct mdt_body *body)
921 return (S_ISREG(la->la_mode) && (body->mbo_valid & OBD_MD_FLEASIZE)) ||
922 (S_ISDIR(la->la_mode) && (body->mbo_valid & OBD_MD_FLDIREA));
925 void mdt_client_compatibility(struct mdt_thread_info *info)
927 struct mdt_body *body;
928 struct ptlrpc_request *req = mdt_info_req(info);
929 struct obd_export *exp = req->rq_export;
930 struct md_attr *ma = &info->mti_attr;
931 struct lu_attr *la = &ma->ma_attr;
934 if (exp_connect_layout(exp))
935 /* the client can deal with 16-bit lmm_stripe_count */
938 body = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
940 if (!mdt_body_has_lov(la, body))
943 /* now we have a reply with a lov for a client not compatible with the
944 * layout lock so we have to clean the layout generation number */
945 if (S_ISREG(la->la_mode))
946 ma->ma_lmm->lmm_layout_gen = 0;
950 static int mdt_attr_get_eabuf_size(struct mdt_thread_info *info,
951 struct mdt_object *o)
953 const struct lu_env *env = info->mti_env;
956 rc = mo_xattr_get(env, mdt_object_child(o), &LU_BUF_NULL,
965 /* Is it a directory? Let's check for the LMV as well */
966 if (S_ISDIR(lu_object_attr(&mdt_object_child(o)->mo_lu))) {
967 rc2 = mo_xattr_get(env, mdt_object_child(o), &LU_BUF_NULL,
971 rc2 = mo_xattr_get(env, mdt_object_child(o),
973 XATTR_NAME_DEFAULT_LMV);
975 if ((rc2 < 0 && rc2 != -ENODATA) || (rc2 > rc))
983 int mdt_big_xattr_get(struct mdt_thread_info *info, struct mdt_object *o,
986 const struct lu_env *env = info->mti_env;
990 LASSERT(info->mti_big_lmm_used == 0);
991 rc = mo_xattr_get(env, mdt_object_child(o), &LU_BUF_NULL, name);
995 /* big_lmm may need to be grown */
996 if (info->mti_big_lmmsize < rc) {
997 int size = size_roundup_power2(rc);
999 if (info->mti_big_lmmsize > 0) {
1000 /* free old buffer */
1001 LASSERT(info->mti_big_lmm);
1002 OBD_FREE_LARGE(info->mti_big_lmm,
1003 info->mti_big_lmmsize);
1004 info->mti_big_lmm = NULL;
1005 info->mti_big_lmmsize = 0;
1008 OBD_ALLOC_LARGE(info->mti_big_lmm, size);
1009 if (info->mti_big_lmm == NULL)
1011 info->mti_big_lmmsize = size;
1013 LASSERT(info->mti_big_lmmsize >= rc);
1015 info->mti_buf.lb_buf = info->mti_big_lmm;
1016 info->mti_buf.lb_len = info->mti_big_lmmsize;
1017 rc = mo_xattr_get(env, mdt_object_child(o), &info->mti_buf, name);
1022 int __mdt_stripe_get(struct mdt_thread_info *info, struct mdt_object *o,
1023 struct md_attr *ma, const char *name)
1025 struct md_object *next = mdt_object_child(o);
1026 struct lu_buf *buf = &info->mti_buf;
1029 if (strcmp(name, XATTR_NAME_LOV) == 0) {
1030 buf->lb_buf = ma->ma_lmm;
1031 buf->lb_len = ma->ma_lmm_size;
1032 LASSERT(!(ma->ma_valid & MA_LOV));
1033 } else if (strcmp(name, XATTR_NAME_LMV) == 0) {
1034 buf->lb_buf = ma->ma_lmv;
1035 buf->lb_len = ma->ma_lmv_size;
1036 LASSERT(!(ma->ma_valid & MA_LMV));
1037 } else if (strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0) {
1038 buf->lb_buf = ma->ma_default_lmv;
1039 buf->lb_len = ma->ma_default_lmv_size;
1040 LASSERT(!(ma->ma_valid & MA_LMV_DEF));
1045 LASSERT(buf->lb_buf);
1047 rc = mo_xattr_get(info->mti_env, next, buf, name);
1051 if (strcmp(name, XATTR_NAME_LOV) == 0) {
1052 if (info->mti_big_lmm_used)
1053 ma->ma_lmm = info->mti_big_lmm;
1055 /* NOT return LOV EA with hole to old client. */
1056 if (unlikely(le32_to_cpu(ma->ma_lmm->lmm_pattern) &
1057 LOV_PATTERN_F_HOLE) &&
1058 !(exp_connect_flags(info->mti_exp) &
1059 OBD_CONNECT_LFSCK)) {
1062 ma->ma_lmm_size = rc;
1063 ma->ma_valid |= MA_LOV;
1065 } else if (strcmp(name, XATTR_NAME_LMV) == 0) {
1066 if (info->mti_big_lmm_used)
1067 ma->ma_lmv = info->mti_big_lmm;
1069 ma->ma_lmv_size = rc;
1070 ma->ma_valid |= MA_LMV;
1071 } else if (strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0) {
1072 ma->ma_default_lmv_size = rc;
1073 ma->ma_valid |= MA_LMV_DEF;
1076 /* Update mdt_max_mdsize so all clients will be aware that */
1077 if (info->mti_mdt->mdt_max_mdsize < rc)
1078 info->mti_mdt->mdt_max_mdsize = rc;
1081 } else if (rc == -ENODATA) {
1084 } else if (rc == -ERANGE) {
1085 /* Default LMV has fixed size, so it must be able to fit
1086 * in the original buffer */
1087 if (strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0)
1089 rc = mdt_big_xattr_get(info, o, name);
1091 info->mti_big_lmm_used = 1;
1099 int mdt_stripe_get(struct mdt_thread_info *info, struct mdt_object *o,
1100 struct md_attr *ma, const char *name)
1104 if (!info->mti_big_lmm) {
1105 OBD_ALLOC(info->mti_big_lmm, PAGE_SIZE);
1106 if (!info->mti_big_lmm)
1108 info->mti_big_lmmsize = PAGE_SIZE;
1111 if (strcmp(name, XATTR_NAME_LOV) == 0) {
1112 ma->ma_lmm = info->mti_big_lmm;
1113 ma->ma_lmm_size = info->mti_big_lmmsize;
1114 ma->ma_valid &= ~MA_LOV;
1115 } else if (strcmp(name, XATTR_NAME_LMV) == 0) {
1116 ma->ma_lmv = info->mti_big_lmm;
1117 ma->ma_lmv_size = info->mti_big_lmmsize;
1118 ma->ma_valid &= ~MA_LMV;
1123 LASSERT(!info->mti_big_lmm_used);
1124 rc = __mdt_stripe_get(info, o, ma, name);
1125 /* since big_lmm is always used here, clear 'used' flag to avoid
1126 * assertion in mdt_big_xattr_get().
1128 info->mti_big_lmm_used = 0;
1133 int mdt_attr_get_pfid(struct mdt_thread_info *info, struct mdt_object *o,
1134 struct lu_fid *pfid)
1136 struct lu_buf *buf = &info->mti_buf;
1137 struct link_ea_header *leh;
1138 struct link_ea_entry *lee;
1142 buf->lb_buf = info->mti_big_lmm;
1143 buf->lb_len = info->mti_big_lmmsize;
1144 rc = mo_xattr_get(info->mti_env, mdt_object_child(o),
1145 buf, XATTR_NAME_LINK);
1146 /* ignore errors, MA_PFID won't be set and it is
1147 * up to the caller to treat this as an error */
1148 if (rc == -ERANGE || buf->lb_len == 0) {
1149 rc = mdt_big_xattr_get(info, o, XATTR_NAME_LINK);
1150 buf->lb_buf = info->mti_big_lmm;
1151 buf->lb_len = info->mti_big_lmmsize;
1156 if (rc < sizeof(*leh)) {
1157 CERROR("short LinkEA on "DFID": rc = %d\n",
1158 PFID(mdt_object_fid(o)), rc);
1162 leh = (struct link_ea_header *) buf->lb_buf;
1163 lee = (struct link_ea_entry *)(leh + 1);
1164 if (leh->leh_magic == __swab32(LINK_EA_MAGIC)) {
1165 leh->leh_magic = LINK_EA_MAGIC;
1166 leh->leh_reccount = __swab32(leh->leh_reccount);
1167 leh->leh_len = __swab64(leh->leh_len);
1169 if (leh->leh_magic != LINK_EA_MAGIC)
1171 if (leh->leh_reccount == 0)
1174 memcpy(pfid, &lee->lee_parent_fid, sizeof(*pfid));
1175 fid_be_to_cpu(pfid, pfid);
1180 int mdt_attr_get_pfid_name(struct mdt_thread_info *info, struct mdt_object *o,
1181 struct lu_fid *pfid, struct lu_name *lname)
1183 struct lu_buf *buf = &info->mti_buf;
1184 struct link_ea_header *leh;
1185 struct link_ea_entry *lee;
1189 buf->lb_buf = info->mti_xattr_buf;
1190 buf->lb_len = sizeof(info->mti_xattr_buf);
1191 rc = mo_xattr_get(info->mti_env, mdt_object_child(o), buf,
1193 if (rc == -ERANGE) {
1194 rc = mdt_big_xattr_get(info, o, XATTR_NAME_LINK);
1195 buf->lb_buf = info->mti_big_lmm;
1196 buf->lb_len = info->mti_big_lmmsize;
1201 if (rc < sizeof(*leh)) {
1202 CERROR("short LinkEA on "DFID": rc = %d\n",
1203 PFID(mdt_object_fid(o)), rc);
1207 leh = (struct link_ea_header *)buf->lb_buf;
1208 lee = (struct link_ea_entry *)(leh + 1);
1209 if (leh->leh_magic == __swab32(LINK_EA_MAGIC)) {
1210 leh->leh_magic = LINK_EA_MAGIC;
1211 leh->leh_reccount = __swab32(leh->leh_reccount);
1212 leh->leh_len = __swab64(leh->leh_len);
1214 if (leh->leh_magic != LINK_EA_MAGIC)
1217 if (leh->leh_reccount == 0)
1220 linkea_entry_unpack(lee, &reclen, lname, pfid);
1225 int mdt_attr_get_complex(struct mdt_thread_info *info,
1226 struct mdt_object *o, struct md_attr *ma)
1228 const struct lu_env *env = info->mti_env;
1229 struct md_object *next = mdt_object_child(o);
1230 struct lu_buf *buf = &info->mti_buf;
1231 int need = ma->ma_need;
1238 if (mdt_object_exists(o) == 0)
1239 GOTO(out, rc = -ENOENT);
1240 mode = lu_object_attr(&next->mo_lu);
1242 if (need & MA_INODE) {
1243 ma->ma_need = MA_INODE;
1244 rc = mo_attr_get(env, next, ma);
1249 (void) mdt_get_som(info, o, ma);
1250 ma->ma_valid |= MA_INODE;
1253 if (need & MA_PFID) {
1254 rc = mdt_attr_get_pfid(info, o, &ma->ma_pfid);
1256 ma->ma_valid |= MA_PFID;
1257 /* ignore this error, parent fid is not mandatory */
1261 if (need & MA_LOV && (S_ISREG(mode) || S_ISDIR(mode))) {
1262 rc = __mdt_stripe_get(info, o, ma, XATTR_NAME_LOV);
1267 if (need & MA_LMV && S_ISDIR(mode)) {
1268 rc = __mdt_stripe_get(info, o, ma, XATTR_NAME_LMV);
1273 if (need & MA_LMV_DEF && S_ISDIR(mode)) {
1274 rc = __mdt_stripe_get(info, o, ma, XATTR_NAME_DEFAULT_LMV);
1280 * In the handle of MA_INODE, we may already get the SOM attr.
1282 if (need & MA_SOM && S_ISREG(mode) && !(ma->ma_valid & MA_SOM)) {
1283 rc = mdt_get_som(info, o, ma);
1288 if (need & MA_HSM && S_ISREG(mode)) {
1289 buf->lb_buf = info->mti_xattr_buf;
1290 buf->lb_len = sizeof(info->mti_xattr_buf);
1291 BUILD_BUG_ON(sizeof(struct hsm_attrs) >
1292 sizeof(info->mti_xattr_buf));
1293 rc2 = mo_xattr_get(info->mti_env, next, buf, XATTR_NAME_HSM);
1294 rc2 = lustre_buf2hsm(info->mti_xattr_buf, rc2, &ma->ma_hsm);
1296 ma->ma_valid |= MA_HSM;
1297 else if (rc2 < 0 && rc2 != -ENODATA)
1298 GOTO(out, rc = rc2);
1301 #ifdef CONFIG_LUSTRE_FS_POSIX_ACL
1302 if (need & MA_ACL_DEF && S_ISDIR(mode)) {
1303 buf->lb_buf = ma->ma_acl;
1304 buf->lb_len = ma->ma_acl_size;
1305 rc2 = mo_xattr_get(env, next, buf, XATTR_NAME_ACL_DEFAULT);
1307 ma->ma_acl_size = rc2;
1308 ma->ma_valid |= MA_ACL_DEF;
1309 } else if (rc2 == -ENODATA) {
1311 ma->ma_acl_size = 0;
1313 GOTO(out, rc = rc2);
1318 CDEBUG(D_INODE, "after getattr rc = %d, ma_valid = %#llx ma_lmm=%p\n",
1319 rc, ma->ma_valid, ma->ma_lmm);
1323 static int mdt_getattr_internal(struct mdt_thread_info *info,
1324 struct mdt_object *o, int ma_need)
1326 struct mdt_device *mdt = info->mti_mdt;
1327 struct md_object *next = mdt_object_child(o);
1328 const struct mdt_body *reqbody = info->mti_body;
1329 struct ptlrpc_request *req = mdt_info_req(info);
1330 struct md_attr *ma = &info->mti_attr;
1331 struct lu_attr *la = &ma->ma_attr;
1332 struct req_capsule *pill = info->mti_pill;
1333 const struct lu_env *env = info->mti_env;
1334 struct mdt_body *repbody;
1335 struct lu_buf *buffer = &info->mti_buf;
1336 struct obd_export *exp = info->mti_exp;
1337 ktime_t kstart = ktime_get();
1342 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK))
1343 RETURN(err_serious(-ENOMEM));
1345 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
1349 if (mdt_object_remote(o)) {
1350 /* This object is located on remote node.*/
1351 /* Return -ENOTSUPP for old client */
1352 if (!mdt_is_dne_client(req->rq_export))
1353 GOTO(out, rc = -ENOTSUPP);
1355 repbody->mbo_fid1 = *mdt_object_fid(o);
1356 repbody->mbo_valid = OBD_MD_FLID | OBD_MD_MDS;
1360 if (reqbody->mbo_eadatasize > 0) {
1361 buffer->lb_buf = req_capsule_server_get(pill, &RMF_MDT_MD);
1362 if (buffer->lb_buf == NULL)
1363 GOTO(out, rc = -EPROTO);
1364 buffer->lb_len = req_capsule_get_size(pill, &RMF_MDT_MD,
1367 buffer->lb_buf = NULL;
1369 ma_need &= ~(MA_LOV | MA_LMV);
1370 CDEBUG(D_INFO, "%s: RPC from %s: does not need LOVEA.\n",
1371 mdt_obd_name(info->mti_mdt),
1372 req->rq_export->exp_client_uuid.uuid);
1375 /* from 2.12.58 intent_getattr pack default LMV in reply */
1376 if (S_ISDIR(lu_object_attr(&next->mo_lu)) &&
1377 ((reqbody->mbo_valid & (OBD_MD_MEA | OBD_MD_DEFAULT_MEA)) ==
1378 (OBD_MD_MEA | OBD_MD_DEFAULT_MEA)) &&
1379 req_capsule_has_field(&req->rq_pill, &RMF_DEFAULT_MDT_MD,
1381 ma->ma_lmv = buffer->lb_buf;
1382 ma->ma_lmv_size = buffer->lb_len;
1383 ma->ma_default_lmv = req_capsule_server_get(pill,
1384 &RMF_DEFAULT_MDT_MD);
1385 ma->ma_default_lmv_size = req_capsule_get_size(pill,
1386 &RMF_DEFAULT_MDT_MD,
1388 ma->ma_need = MA_INODE;
1389 if (ma->ma_lmv_size > 0)
1390 ma->ma_need |= MA_LMV;
1391 if (ma->ma_default_lmv_size > 0)
1392 ma->ma_need |= MA_LMV_DEF;
1393 } else if (S_ISDIR(lu_object_attr(&next->mo_lu)) &&
1394 (reqbody->mbo_valid & (OBD_MD_MEA | OBD_MD_DEFAULT_MEA))) {
1395 /* If it is dir and client require MEA, then we got MEA */
1396 /* Assumption: MDT_MD size is enough for lmv size. */
1397 ma->ma_lmv = buffer->lb_buf;
1398 ma->ma_lmv_size = buffer->lb_len;
1399 ma->ma_need = MA_INODE;
1400 if (ma->ma_lmv_size > 0) {
1401 if (reqbody->mbo_valid & OBD_MD_MEA) {
1402 ma->ma_need |= MA_LMV;
1403 } else if (reqbody->mbo_valid & OBD_MD_DEFAULT_MEA) {
1404 ma->ma_need |= MA_LMV_DEF;
1405 ma->ma_default_lmv = buffer->lb_buf;
1407 ma->ma_default_lmv_size = buffer->lb_len;
1408 ma->ma_lmv_size = 0;
1412 ma->ma_lmm = buffer->lb_buf;
1413 ma->ma_lmm_size = buffer->lb_len;
1414 ma->ma_need = MA_INODE | MA_HSM;
1415 if (ma->ma_lmm_size > 0) {
1416 ma->ma_need |= MA_LOV;
1417 /* Older clients may crash if they getattr overstriped
1420 if (!exp_connect_overstriping(exp) &&
1421 mdt_lmm_is_overstriping(ma->ma_lmm))
1422 RETURN(-EOPNOTSUPP);
1426 if (S_ISDIR(lu_object_attr(&next->mo_lu)) &&
1427 reqbody->mbo_valid & OBD_MD_FLDIREA &&
1428 lustre_msg_get_opc(req->rq_reqmsg) == MDS_GETATTR) {
1429 /* get default stripe info for this dir. */
1430 ma->ma_need |= MA_LOV_DEF;
1432 ma->ma_need |= ma_need;
1434 rc = mdt_attr_get_complex(info, o, ma);
1436 CDEBUG_LIMIT(rc == -ENOENT ? D_OTHER : D_ERROR,
1437 "%s: getattr error for "DFID": rc = %d\n",
1438 mdt_obd_name(info->mti_mdt),
1439 PFID(mdt_object_fid(o)), rc);
1443 /* if file is released, check if a restore is running */
1444 if (ma->ma_valid & MA_HSM) {
1445 repbody->mbo_valid |= OBD_MD_TSTATE;
1446 if ((ma->ma_hsm.mh_flags & HS_RELEASED) &&
1447 mdt_hsm_restore_is_running(info, mdt_object_fid(o)))
1448 repbody->mbo_t_state = MS_RESTORE;
1451 if (unlikely(!(ma->ma_valid & MA_INODE)))
1454 mdt_pack_attr2body(info, repbody, la, mdt_object_fid(o));
1456 if (mdt_body_has_lov(la, reqbody)) {
1457 u32 stripe_count = 1;
1458 bool fixed_layout = false;
1460 if (ma->ma_valid & MA_LOV) {
1461 LASSERT(ma->ma_lmm_size);
1462 repbody->mbo_eadatasize = ma->ma_lmm_size;
1463 if (S_ISDIR(la->la_mode))
1464 repbody->mbo_valid |= OBD_MD_FLDIREA;
1466 repbody->mbo_valid |= OBD_MD_FLEASIZE;
1467 mdt_dump_lmm(D_INFO, ma->ma_lmm, repbody->mbo_valid);
1469 if (ma->ma_valid & MA_LMV) {
1470 struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1471 u32 magic = le32_to_cpu(lmv->lmv_magic);
1473 /* Return -ENOTSUPP for old client */
1474 if (!mdt_is_striped_client(req->rq_export))
1477 LASSERT(S_ISDIR(la->la_mode));
1478 mdt_dump_lmv(D_INFO, ma->ma_lmv);
1479 repbody->mbo_eadatasize = ma->ma_lmv_size;
1480 repbody->mbo_valid |= (OBD_MD_FLDIREA|OBD_MD_MEA);
1482 stripe_count = le32_to_cpu(lmv->lmv_stripe_count);
1483 fixed_layout = lmv_is_fixed(lmv);
1484 if (magic == LMV_MAGIC_STRIPE && lmv_is_restriping(lmv))
1485 mdt_restripe_migrate_add(info, o);
1486 else if (magic == LMV_MAGIC_V1 &&
1487 lmv_is_restriping(lmv))
1488 mdt_restripe_update_add(info, o);
1490 if (ma->ma_valid & MA_LMV_DEF) {
1491 /* Return -ENOTSUPP for old client */
1492 if (!mdt_is_striped_client(req->rq_export))
1494 LASSERT(S_ISDIR(la->la_mode));
1496 * when ll_dir_getstripe() gets default LMV, it
1497 * checks mbo_eadatasize.
1499 if (!(ma->ma_valid & MA_LMV))
1500 repbody->mbo_eadatasize =
1501 ma->ma_default_lmv_size;
1502 repbody->mbo_valid |= (OBD_MD_FLDIREA |
1503 OBD_MD_DEFAULT_MEA);
1506 "dirent count %llu stripe count %u MDT count %d\n",
1507 ma->ma_attr.la_dirent_count, stripe_count,
1508 atomic_read(&mdt->mdt_mds_mds_conns) + 1);
1509 if (ma->ma_attr.la_dirent_count != LU_DIRENT_COUNT_UNSET &&
1510 ma->ma_attr.la_dirent_count >
1511 mdt->mdt_restriper.mdr_dir_split_count &&
1512 !fid_is_root(mdt_object_fid(o)) &&
1513 mdt->mdt_enable_dir_auto_split &&
1514 !o->mot_restriping &&
1515 stripe_count < atomic_read(&mdt->mdt_mds_mds_conns) + 1 &&
1517 mdt_auto_split_add(info, o);
1518 } else if (S_ISLNK(la->la_mode) &&
1519 reqbody->mbo_valid & OBD_MD_LINKNAME) {
1520 buffer->lb_buf = ma->ma_lmm;
1521 /* eadatasize from client includes NULL-terminator, so
1522 * there is no need to read it */
1523 buffer->lb_len = reqbody->mbo_eadatasize - 1;
1524 rc = mo_readlink(env, next, buffer);
1525 if (unlikely(rc <= 0)) {
1526 CERROR("%s: readlink failed for "DFID": rc = %d\n",
1527 mdt_obd_name(info->mti_mdt),
1528 PFID(mdt_object_fid(o)), rc);
1531 int print_limit = min_t(int, PAGE_SIZE - 128, rc);
1533 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_READLINK_EPROTO))
1535 repbody->mbo_valid |= OBD_MD_LINKNAME;
1536 /* we need to report back size with NULL-terminator
1537 * because client expects that */
1538 repbody->mbo_eadatasize = rc + 1;
1539 if (repbody->mbo_eadatasize != reqbody->mbo_eadatasize)
1540 CDEBUG(D_INODE, "%s: Read shorter symlink %d "
1541 "on "DFID ", expected %d\n",
1542 mdt_obd_name(info->mti_mdt),
1543 rc, PFID(mdt_object_fid(o)),
1544 reqbody->mbo_eadatasize - 1);
1545 /* NULL terminate */
1546 ((char *)ma->ma_lmm)[rc] = 0;
1548 /* If the total CDEBUG() size is larger than a page, it
1549 * will print a warning to the console, avoid this by
1550 * printing just the last part of the symlink. */
1551 CDEBUG(D_INODE, "symlink dest %s%.*s, len = %d\n",
1552 print_limit < rc ? "..." : "", print_limit,
1553 (char *)ma->ma_lmm + rc - print_limit, rc);
1558 if (reqbody->mbo_valid & OBD_MD_FLMODEASIZE) {
1559 repbody->mbo_max_mdsize = info->mti_mdt->mdt_max_mdsize;
1560 repbody->mbo_valid |= OBD_MD_FLMODEASIZE;
1561 CDEBUG(D_INODE, "changing the max MD size to %u\n",
1562 repbody->mbo_max_mdsize);
1565 #ifdef CONFIG_LUSTRE_FS_POSIX_ACL
1566 if ((exp_connect_flags(req->rq_export) & OBD_CONNECT_ACL) &&
1567 (reqbody->mbo_valid & OBD_MD_FLACL)) {
1568 struct lu_nodemap *nodemap = nodemap_get_from_exp(exp);
1569 if (IS_ERR(nodemap))
1570 RETURN(PTR_ERR(nodemap));
1572 rc = mdt_pack_acl2body(info, repbody, o, nodemap);
1573 nodemap_putref(nodemap);
1579 mdt_counter_incr(req, LPROC_MDT_GETATTR,
1580 ktime_us_delta(ktime_get(), kstart));
1585 static int mdt_getattr(struct tgt_session_info *tsi)
1587 struct mdt_thread_info *info = tsi2mdt_info(tsi);
1588 struct mdt_object *obj = info->mti_object;
1589 struct req_capsule *pill = info->mti_pill;
1590 struct mdt_body *reqbody;
1591 struct mdt_body *repbody;
1595 if (unlikely(info->mti_object == NULL))
1598 reqbody = req_capsule_client_get(pill, &RMF_MDT_BODY);
1600 LASSERT(lu_object_assert_exists(&obj->mot_obj));
1602 /* Special case for Data-on-MDT files to get data version */
1603 if (unlikely(reqbody->mbo_valid & OBD_MD_FLDATAVERSION)) {
1604 rc = mdt_data_version_get(tsi);
1608 /* Unlike intent case where we need to pre-fill out buffers early on
1609 * in intent policy for ldlm reasons, here we can have a much better
1610 * guess at EA size by just reading it from disk.
1611 * Exceptions are readdir and (missing) directory striping */
1613 if (reqbody->mbo_valid & OBD_MD_LINKNAME) {
1614 /* No easy way to know how long is the symlink, but it cannot
1615 * be more than PATH_MAX, so we allocate +1 */
1617 /* A special case for fs ROOT: getattr there might fetch
1618 * default EA for entire fs, not just for this dir!
1620 } else if (lu_fid_eq(mdt_object_fid(obj),
1621 &info->mti_mdt->mdt_md_root_fid) &&
1622 (reqbody->mbo_valid & OBD_MD_FLDIREA) &&
1623 (lustre_msg_get_opc(mdt_info_req(info)->rq_reqmsg) ==
1625 /* Should the default strping be bigger, mdt_fix_reply
1626 * will reallocate */
1627 rc = DEF_REP_MD_SIZE;
1629 /* Read the actual EA size from disk */
1630 rc = mdt_attr_get_eabuf_size(info, obj);
1634 GOTO(out, rc = err_serious(rc));
1636 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER, rc);
1638 /* Set ACL reply buffer size as LUSTRE_POSIX_ACL_MAX_SIZE_OLD
1639 * by default. If the target object has more ACL entries, then
1640 * enlarge the buffer when necessary. */
1641 req_capsule_set_size(pill, &RMF_ACL, RCL_SERVER,
1642 LUSTRE_POSIX_ACL_MAX_SIZE_OLD);
1644 rc = req_capsule_server_pack(pill);
1645 if (unlikely(rc != 0))
1646 GOTO(out, rc = err_serious(rc));
1648 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
1649 LASSERT(repbody != NULL);
1650 repbody->mbo_eadatasize = 0;
1651 repbody->mbo_aclsize = 0;
1653 rc = mdt_check_ucred(info);
1655 GOTO(out_shrink, rc);
1657 info->mti_cross_ref = !!(reqbody->mbo_valid & OBD_MD_FLCROSSREF);
1659 rc = mdt_getattr_internal(info, obj, 0);
1662 mdt_client_compatibility(info);
1663 rc2 = mdt_fix_reply(info);
1667 mdt_thread_info_fini(info);
1672 * Handler of layout intent RPC requiring the layout modification
1674 * \param[in] info thread environment
1675 * \param[in] obj object
1676 * \param[out] lhc object ldlm lock handle
1677 * \param[in] layout layout change descriptor
1679 * \retval 0 on success
1680 * \retval < 0 error code
1682 int mdt_layout_change(struct mdt_thread_info *info, struct mdt_object *obj,
1683 struct mdt_lock_handle *lhc,
1684 struct md_layout_change *layout)
1690 if (!mdt_object_exists(obj))
1693 if (!S_ISREG(lu_object_attr(&obj->mot_obj)))
1696 rc = mo_permission(info->mti_env, NULL, mdt_object_child(obj), NULL,
1701 rc = mdt_check_resent_lock(info, obj, lhc);
1707 __u64 lockpart = MDS_INODELOCK_LAYOUT;
1709 /* take layout lock to prepare layout change */
1710 if (layout->mlc_opc == MD_LAYOUT_WRITE)
1711 lockpart |= MDS_INODELOCK_UPDATE;
1713 mdt_lock_handle_init(lhc);
1714 mdt_lock_reg_init(lhc, LCK_EX);
1715 rc = mdt_reint_object_lock(info, obj, lhc, lockpart, false);
1720 mutex_lock(&obj->mot_som_mutex);
1721 rc = mo_layout_change(info->mti_env, mdt_object_child(obj), layout);
1722 mutex_unlock(&obj->mot_som_mutex);
1725 mdt_object_unlock(info, obj, lhc, 1);
1731 * Exchange MOF_LOV_CREATED flags between two objects after a
1732 * layout swap. No assumption is made on whether o1 or o2 have
1733 * created objects or not.
1735 * \param[in,out] o1 First swap layout object
1736 * \param[in,out] o2 Second swap layout object
1738 static void mdt_swap_lov_flag(struct mdt_object *o1, struct mdt_object *o2)
1740 unsigned int o1_lov_created = o1->mot_lov_created;
1742 mutex_lock(&o1->mot_lov_mutex);
1743 mutex_lock(&o2->mot_lov_mutex);
1745 o1->mot_lov_created = o2->mot_lov_created;
1746 o2->mot_lov_created = o1_lov_created;
1748 mutex_unlock(&o2->mot_lov_mutex);
1749 mutex_unlock(&o1->mot_lov_mutex);
1752 static int mdt_swap_layouts(struct tgt_session_info *tsi)
1754 struct mdt_thread_info *info;
1755 struct ptlrpc_request *req = tgt_ses_req(tsi);
1756 struct obd_export *exp = req->rq_export;
1757 struct mdt_object *o1, *o2, *o;
1758 struct mdt_lock_handle *lh1, *lh2;
1759 struct mdc_swap_layouts *msl;
1763 /* client does not support layout lock, so layout swaping
1765 * FIXME: there is a problem for old clients which don't support
1766 * layout lock yet. If those clients have already opened the file
1767 * they won't be notified at all so that old layout may still be
1768 * used to do IO. This can be fixed after file release is landed by
1769 * doing exclusive open and taking full EX ibits lock. - Jinshan */
1770 if (!exp_connect_layout(exp))
1771 RETURN(-EOPNOTSUPP);
1773 info = tsi2mdt_info(tsi);
1774 if (unlikely(info->mti_object == NULL))
1777 if (info->mti_dlm_req != NULL)
1778 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1780 o1 = info->mti_object;
1781 o = o2 = mdt_object_find(info->mti_env, info->mti_mdt,
1782 &info->mti_body->mbo_fid2);
1784 GOTO(out, rc = PTR_ERR(o));
1786 if (mdt_object_remote(o) || !mdt_object_exists(o)) /* remote object */
1787 GOTO(put, rc = -ENOENT);
1789 rc = lu_fid_cmp(&info->mti_body->mbo_fid1, &info->mti_body->mbo_fid2);
1790 if (unlikely(rc == 0)) /* same file, you kidding me? no-op. */
1796 /* permission check. Make sure the calling process having permission
1797 * to write both files. */
1798 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o1), NULL,
1803 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o2), NULL,
1808 msl = req_capsule_client_get(info->mti_pill, &RMF_SWAP_LAYOUTS);
1810 GOTO(put, rc = -EPROTO);
1812 lh1 = &info->mti_lh[MDT_LH_NEW];
1813 mdt_lock_reg_init(lh1, LCK_EX);
1814 lh2 = &info->mti_lh[MDT_LH_OLD];
1815 mdt_lock_reg_init(lh2, LCK_EX);
1817 rc = mdt_object_lock(info, o1, lh1, MDS_INODELOCK_LAYOUT |
1818 MDS_INODELOCK_XATTR);
1822 rc = mdt_object_lock(info, o2, lh2, MDS_INODELOCK_LAYOUT |
1823 MDS_INODELOCK_XATTR);
1827 rc = mo_swap_layouts(info->mti_env, mdt_object_child(o1),
1828 mdt_object_child(o2), msl->msl_flags);
1832 mdt_swap_lov_flag(o1, o2);
1835 mdt_object_unlock(info, o2, lh2, rc);
1837 mdt_object_unlock(info, o1, lh1, rc);
1839 mdt_object_put(info->mti_env, o);
1841 mdt_thread_info_fini(info);
1845 static int mdt_raw_lookup(struct mdt_thread_info *info,
1846 struct mdt_object *parent,
1847 const struct lu_name *lname)
1849 struct lu_fid *fid = &info->mti_tmp_fid1;
1850 struct mdt_body *repbody;
1851 bool is_dotdot = false;
1852 bool is_old_parent_stripe = false;
1853 bool is_new_parent_checked = false;
1858 LASSERT(!info->mti_cross_ref);
1859 /* Always allow to lookup ".." */
1860 if (lname->ln_namelen == 2 &&
1861 lname->ln_name[0] == '.' && lname->ln_name[1] == '.') {
1862 info->mti_spec.sp_permitted = 1;
1864 if (mdt_is_dir_stripe(info, parent) == 1)
1865 is_old_parent_stripe = true;
1868 mdt_object_get(info->mti_env, parent);
1870 /* Only got the fid of this obj by name */
1872 rc = mdo_lookup(info->mti_env, mdt_object_child(parent), lname, fid,
1874 mdt_object_put(info->mti_env, parent);
1878 /* getattr_name("..") should return master object FID for striped dir */
1879 if (is_dotdot && (is_old_parent_stripe || !is_new_parent_checked)) {
1880 parent = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1882 RETURN(PTR_ERR(parent));
1884 /* old client getattr_name("..") with stripe FID */
1885 if (unlikely(is_old_parent_stripe)) {
1886 is_old_parent_stripe = false;
1890 /* ".." may be a stripe */
1891 if (unlikely(mdt_is_dir_stripe(info, parent) == 1)) {
1892 is_new_parent_checked = true;
1896 mdt_object_put(info->mti_env, parent);
1899 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1900 repbody->mbo_fid1 = *fid;
1901 repbody->mbo_valid = OBD_MD_FLID;
1907 * Find name matching hash
1909 * We search \a child LinkEA for a name whose hash matches \a lname
1910 * (it contains an encoded hash).
1912 * \param info mdt thread info
1913 * \param lname encoded hash to find
1914 * \param parent parent object
1915 * \param child object to search with LinkEA
1916 * \param force_check true to check hash even if LinkEA has only one entry
1918 * \retval 1 match found
1919 * \retval 0 no match found
1920 * \retval -ev negative errno upon error
1922 int find_name_matching_hash(struct mdt_thread_info *info, struct lu_name *lname,
1923 struct mdt_object *parent, struct mdt_object *child,
1926 /* Here, lname is an encoded hash of on-disk name, and
1927 * client is doing access without encryption key.
1928 * So we need to get LinkEA, check parent fid is correct and
1929 * compare name hash with the one in the request.
1931 struct lu_buf *buf = &info->mti_big_buf;
1932 struct lu_name name;
1934 struct linkea_data ldata = { NULL };
1935 struct link_ea_header *leh;
1936 struct link_ea_entry *lee;
1937 struct lu_buf link = { 0 };
1939 int reclen, count, rc;
1943 if (lname->ln_namelen < LLCRYPT_FNAME_DIGEST_SIZE)
1946 buf = lu_buf_check_and_alloc(buf, PATH_MAX);
1951 rc = mdt_links_read(info, child, &ldata);
1956 if (force_check || leh->leh_reccount > 1) {
1957 hash = kmalloc(lname->ln_namelen, GFP_NOFS);
1960 rc = critical_decode(lname->ln_name, lname->ln_namelen, hash);
1962 lee = (struct link_ea_entry *)(leh + 1);
1963 for (count = 0; count < leh->leh_reccount; count++) {
1964 linkea_entry_unpack(lee, &reclen, &name, &pfid);
1965 if (!force_check && leh->leh_reccount == 1) {
1966 /* if there is only one rec, it has to be it */
1970 if (!parent || lu_fid_eq(&pfid, mdt_object_fid(parent))) {
1971 lu_buf_check_and_alloc(&link, name.ln_namelen);
1973 GOTO(out_match, rc = -ENOMEM);
1974 rc = critical_decode(name.ln_name, name.ln_namelen,
1977 if (memcmp(LLCRYPT_FNAME_DIGEST(link.lb_buf, rc),
1978 hash, LLCRYPT_FNAME_DIGEST_SIZE) == 0) {
1983 lee = (struct link_ea_entry *) ((char *)lee + reclen);
1985 if (count == leh->leh_reccount)
1998 * UPDATE lock should be taken against parent, and be released before exit;
1999 * child_bits lock should be taken against child, and be returned back:
2000 * (1)normal request should release the child lock;
2001 * (2)intent request will grant the lock to client.
2003 static int mdt_getattr_name_lock(struct mdt_thread_info *info,
2004 struct mdt_lock_handle *lhc,
2006 struct ldlm_reply *ldlm_rep)
2008 struct ptlrpc_request *req = mdt_info_req(info);
2009 struct mdt_body *reqbody = NULL;
2010 struct mdt_object *parent = info->mti_object;
2011 struct mdt_object *child = NULL;
2012 struct lu_fid *child_fid = &info->mti_tmp_fid1;
2013 struct lu_name *lname = NULL;
2014 struct mdt_lock_handle *lhp = NULL;
2015 struct ldlm_lock *lock;
2016 struct req_capsule *pill = info->mti_pill;
2024 is_resent = lustre_handle_is_used(&lhc->mlh_reg_lh);
2025 LASSERT(ergo(is_resent,
2026 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT));
2031 if (info->mti_cross_ref) {
2032 /* Only getattr on the child. Parent is on another node. */
2033 mdt_set_disposition(info, ldlm_rep,
2034 DISP_LOOKUP_EXECD | DISP_LOOKUP_POS);
2036 CDEBUG(D_INODE, "partial getattr_name child_fid = "DFID", "
2038 PFID(mdt_object_fid(child)), ldlm_rep);
2040 rc = mdt_check_resent_lock(info, child, lhc);
2043 } else if (rc > 0) {
2044 mdt_lock_handle_init(lhc);
2045 mdt_lock_reg_init(lhc, LCK_PR);
2048 * Object's name entry is on another MDS, it will
2049 * request PERM lock only because LOOKUP lock is owned
2050 * by the MDS where name entry resides.
2052 * TODO: it should try layout lock too. - Jinshan
2054 child_bits &= ~(MDS_INODELOCK_LOOKUP |
2055 MDS_INODELOCK_LAYOUT);
2056 child_bits |= MDS_INODELOCK_PERM;
2058 rc = mdt_object_lock(info, child, lhc, child_bits);
2063 /* Finally, we can get attr for child. */
2064 if (!mdt_object_exists(child)) {
2065 LU_OBJECT_DEBUG(D_INFO, info->mti_env,
2067 "remote object doesn't exist.");
2068 mdt_object_unlock(info, child, lhc, 1);
2072 rc = mdt_getattr_internal(info, child, 0);
2073 if (unlikely(rc != 0)) {
2074 mdt_object_unlock(info, child, lhc, 1);
2078 rc = mdt_pack_secctx_in_reply(info, child);
2080 mdt_object_unlock(info, child, lhc, 1);
2084 rc = mdt_pack_encctx_in_reply(info, child);
2086 mdt_object_unlock(info, child, lhc, 1);
2090 lname = &info->mti_name;
2091 mdt_name_unpack(pill, &RMF_NAME, lname, MNF_FIX_ANON);
2093 if (info->mti_body->mbo_valid & OBD_MD_NAMEHASH) {
2094 reqbody = req_capsule_client_get(pill, &RMF_MDT_BODY);
2095 if (unlikely(reqbody == NULL))
2096 RETURN(err_serious(-EPROTO));
2098 *child_fid = reqbody->mbo_fid2;
2099 if (unlikely(!fid_is_sane(child_fid)))
2100 RETURN(err_serious(-EINVAL));
2102 if (lu_fid_eq(mdt_object_fid(parent), child_fid)) {
2103 mdt_object_get(info->mti_env, parent);
2106 child = mdt_object_find(info->mti_env, info->mti_mdt,
2109 RETURN(PTR_ERR(child));
2112 CDEBUG(D_INODE, "getattr with lock for "DFID"/"DFID", "
2114 PFID(mdt_object_fid(parent)),
2115 PFID(&reqbody->mbo_fid2), ldlm_rep);
2116 } else if (lu_name_is_valid(lname)) {
2117 if (mdt_object_remote(parent)) {
2118 CERROR("%s: parent "DFID" is on remote target\n",
2119 mdt_obd_name(info->mti_mdt),
2120 PFID(mdt_object_fid(parent)));
2124 CDEBUG(D_INODE, "getattr with lock for "DFID"/"DNAME", "
2125 "ldlm_rep = %p\n", PFID(mdt_object_fid(parent)),
2126 PNAME(lname), ldlm_rep);
2128 reqbody = req_capsule_client_get(pill, &RMF_MDT_BODY);
2129 if (unlikely(reqbody == NULL))
2130 RETURN(err_serious(-EPROTO));
2132 *child_fid = reqbody->mbo_fid2;
2133 if (unlikely(!fid_is_sane(child_fid)))
2134 RETURN(err_serious(-EINVAL));
2136 if (lu_fid_eq(mdt_object_fid(parent), child_fid)) {
2137 mdt_object_get(info->mti_env, parent);
2140 child = mdt_object_find(info->mti_env, info->mti_mdt,
2143 RETURN(PTR_ERR(child));
2146 if (mdt_object_remote(child)) {
2147 CERROR("%s: child "DFID" is on remote target\n",
2148 mdt_obd_name(info->mti_mdt),
2149 PFID(mdt_object_fid(child)));
2150 GOTO(out_child, rc = -EPROTO);
2153 /* don't fetch LOOKUP lock if it's remote object */
2154 rc = mdt_is_remote_object(info, parent, child);
2156 GOTO(out_child, rc);
2158 child_bits &= ~MDS_INODELOCK_LOOKUP;
2160 CDEBUG(D_INODE, "getattr with lock for "DFID"/"DFID", "
2162 PFID(mdt_object_fid(parent)),
2163 PFID(&reqbody->mbo_fid2), ldlm_rep);
2166 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_EXECD);
2168 if (unlikely(!mdt_object_exists(parent)) &&
2169 !(info->mti_body->mbo_valid & OBD_MD_NAMEHASH) &&
2170 lu_name_is_valid(lname)) {
2171 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2173 "Parent doesn't exist!");
2174 GOTO(out_child, rc = -ESTALE);
2177 if (!(info->mti_body->mbo_valid & OBD_MD_NAMEHASH) &&
2178 lu_name_is_valid(lname)) {
2179 if (info->mti_body->mbo_valid == OBD_MD_FLID) {
2180 rc = mdt_raw_lookup(info, parent, lname);
2185 /* step 1: lock parent only if parent is a directory */
2186 if (S_ISDIR(lu_object_attr(&parent->mot_obj))) {
2187 lhp = &info->mti_lh[MDT_LH_PARENT];
2188 mdt_lock_pdo_init(lhp, LCK_PR, lname);
2189 rc = mdt_object_lock(info, parent, lhp,
2190 MDS_INODELOCK_UPDATE);
2191 if (unlikely(rc != 0))
2195 /* step 2: lookup child's fid by name */
2196 fid_zero(child_fid);
2197 rc = mdo_lookup(info->mti_env, mdt_object_child(parent), lname,
2198 child_fid, &info->mti_spec);
2200 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
2203 GOTO(unlock_parent, rc);
2205 child = mdt_object_find(info->mti_env, info->mti_mdt,
2207 if (unlikely(IS_ERR(child)))
2208 GOTO(unlock_parent, rc = PTR_ERR(child));
2211 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
2213 /* step 3: lock child regardless if it is local or remote. */
2216 if (info->mti_body->mbo_valid & OBD_MD_NAMEHASH) {
2217 /* Here, lname is an encoded hash of on-disk name, and
2218 * client is doing access without encryption key.
2219 * So we need to compare name hash with the one in the request.
2221 if (!find_name_matching_hash(info, lname, parent,
2223 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
2224 mdt_clear_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
2225 GOTO(out_child, rc = -ENOENT);
2229 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RESEND, obd_timeout * 2);
2230 if (!mdt_object_exists(child)) {
2231 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2233 "Object doesn't exist!");
2234 GOTO(out_child, rc = -ENOENT);
2237 rc = mdt_check_resent_lock(info, child, lhc);
2239 GOTO(out_child, rc);
2240 } else if (rc > 0) {
2241 mdt_lock_handle_init(lhc);
2242 mdt_lock_reg_init(lhc, LCK_PR);
2244 if (!(child_bits & MDS_INODELOCK_UPDATE) &&
2245 !mdt_object_remote(child)) {
2246 struct md_attr *ma = &info->mti_attr;
2249 ma->ma_need = MA_INODE;
2250 rc = mdt_attr_get_complex(info, child, ma);
2251 if (unlikely(rc != 0))
2252 GOTO(out_child, rc);
2254 /* If the file has not been changed for some time, we
2255 * return not only a LOOKUP lock, but also an UPDATE
2256 * lock and this might save us RPC on later STAT. For
2257 * directories, it also let negative dentry cache start
2258 * working for this dir. */
2259 if (ma->ma_valid & MA_INODE &&
2260 ma->ma_attr.la_valid & LA_CTIME &&
2261 info->mti_mdt->mdt_namespace->ns_ctime_age_limit +
2262 ma->ma_attr.la_ctime < ktime_get_real_seconds())
2263 child_bits |= MDS_INODELOCK_UPDATE;
2266 /* layout lock must be granted in a best-effort way
2267 * for IT operations */
2268 LASSERT(!(child_bits & MDS_INODELOCK_LAYOUT));
2269 if (S_ISREG(lu_object_attr(&child->mot_obj)) &&
2270 !mdt_object_remote(child) && ldlm_rep != NULL) {
2271 if (!OBD_FAIL_CHECK(OBD_FAIL_MDS_NO_LL_GETATTR) &&
2272 exp_connect_layout(info->mti_exp)) {
2273 /* try to grant layout lock for regular file. */
2274 try_bits = MDS_INODELOCK_LAYOUT;
2276 /* Acquire DOM lock in advance for data-on-mdt file */
2277 if (child != parent)
2278 try_bits |= MDS_INODELOCK_DOM;
2281 if (try_bits != 0) {
2282 /* try layout lock, it may fail to be granted due to
2283 * contention at LOOKUP or UPDATE */
2284 rc = mdt_object_lock_try(info, child, lhc, &child_bits,
2286 if (child_bits & MDS_INODELOCK_LAYOUT)
2289 /* Do not enqueue the UPDATE lock from MDT(cross-MDT),
2290 * client will enqueue the lock to the remote MDT */
2291 if (mdt_object_remote(child))
2292 child_bits &= ~MDS_INODELOCK_UPDATE;
2293 rc = mdt_object_lock(info, child, lhc, child_bits);
2295 if (unlikely(rc != 0))
2296 GOTO(out_child, rc);
2299 /* finally, we can get attr for child. */
2300 rc = mdt_getattr_internal(info, child, ma_need);
2301 if (unlikely(rc != 0)) {
2302 mdt_object_unlock(info, child, lhc, 1);
2303 GOTO(out_child, rc);
2306 rc = mdt_pack_secctx_in_reply(info, child);
2308 mdt_object_unlock(info, child, lhc, 1);
2309 GOTO(out_child, rc);
2312 rc = mdt_pack_encctx_in_reply(info, child);
2314 mdt_object_unlock(info, child, lhc, 1);
2315 GOTO(out_child, rc);
2318 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
2320 /* Debugging code. */
2321 LDLM_DEBUG(lock, "Returning lock to client");
2322 LASSERTF(fid_res_name_eq(mdt_object_fid(child),
2323 &lock->l_resource->lr_name),
2324 "Lock res_id: "DLDLMRES", fid: "DFID"\n",
2325 PLDLMRES(lock->l_resource),
2326 PFID(mdt_object_fid(child)));
2328 if (unlikely(OBD_FAIL_PRECHECK(OBD_FAIL_PTLRPC_ENQ_RESEND))) {
2329 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
2330 OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_ENQ_RESEND,
2332 req->rq_arrival_time.tv_sec +
2334 /* Put the lock to the waiting list and force the cancel */
2335 ldlm_set_ast_sent(lock);
2338 if (S_ISREG(lu_object_attr(&child->mot_obj)) &&
2339 !mdt_object_remote(child) && child != parent) {
2340 mdt_object_put(info->mti_env, child);
2341 rc = mdt_pack_size2body(info, child_fid,
2343 if (rc != 0 && child_bits & MDS_INODELOCK_DOM) {
2344 /* DOM lock was taken in advance but this is
2345 * not DoM file. Drop the lock.
2347 lock_res_and_lock(lock);
2348 ldlm_inodebits_drop(lock, MDS_INODELOCK_DOM);
2349 unlock_res_and_lock(lock);
2351 LDLM_LOCK_PUT(lock);
2352 GOTO(unlock_parent, rc = 0);
2354 LDLM_LOCK_PUT(lock);
2360 mdt_object_put(info->mti_env, child);
2363 mdt_object_unlock(info, parent, lhp, 1);
2367 /* normal handler: should release the child lock */
2368 static int mdt_getattr_name(struct tgt_session_info *tsi)
2370 struct mdt_thread_info *info = tsi2mdt_info(tsi);
2371 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_CHILD];
2372 struct mdt_body *reqbody;
2373 struct mdt_body *repbody;
2378 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
2379 LASSERT(reqbody != NULL);
2380 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2381 LASSERT(repbody != NULL);
2383 info->mti_cross_ref = !!(reqbody->mbo_valid & OBD_MD_FLCROSSREF);
2384 repbody->mbo_eadatasize = 0;
2385 repbody->mbo_aclsize = 0;
2387 rc = mdt_init_ucred(info, reqbody);
2389 GOTO(out_shrink, rc);
2391 rc = mdt_getattr_name_lock(info, lhc, MDS_INODELOCK_UPDATE, NULL);
2392 if (lustre_handle_is_used(&lhc->mlh_reg_lh)) {
2393 ldlm_lock_decref(&lhc->mlh_reg_lh, lhc->mlh_reg_mode);
2394 lhc->mlh_reg_lh.cookie = 0;
2396 mdt_exit_ucred(info);
2399 mdt_client_compatibility(info);
2400 rc2 = mdt_fix_reply(info);
2403 mdt_thread_info_fini(info);
2407 static int mdt_rmfid_unlink(struct mdt_thread_info *info,
2408 const struct lu_fid *pfid,
2409 const struct lu_name *name,
2410 struct mdt_object *obj, s64 ctime)
2412 struct lu_fid *child_fid = &info->mti_tmp_fid1;
2413 struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
2414 struct mdt_device *mdt = info->mti_mdt;
2415 struct md_attr *ma = &info->mti_attr;
2416 struct mdt_lock_handle *parent_lh;
2417 struct mdt_lock_handle *child_lh;
2418 struct mdt_object *pobj;
2419 bool cos_incompat = false;
2423 pobj = mdt_object_find(info->mti_env, mdt, pfid);
2425 GOTO(out, rc = PTR_ERR(pobj));
2427 parent_lh = &info->mti_lh[MDT_LH_PARENT];
2428 mdt_lock_pdo_init(parent_lh, LCK_PW, name);
2429 rc = mdt_object_lock(info, pobj, parent_lh, MDS_INODELOCK_UPDATE);
2431 GOTO(put_parent, rc);
2433 if (mdt_object_remote(pobj))
2434 cos_incompat = true;
2436 rc = mdo_lookup(info->mti_env, mdt_object_child(pobj),
2437 name, child_fid, &info->mti_spec);
2439 GOTO(unlock_parent, rc);
2441 if (!lu_fid_eq(child_fid, mdt_object_fid(obj)))
2442 GOTO(unlock_parent, rc = -EREMCHG);
2444 child_lh = &info->mti_lh[MDT_LH_CHILD];
2445 mdt_lock_reg_init(child_lh, LCK_EX);
2446 rc = mdt_reint_striped_lock(info, obj, child_lh,
2447 MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE,
2448 einfo, cos_incompat);
2450 GOTO(unlock_parent, rc);
2452 if (atomic_read(&obj->mot_open_count)) {
2453 CDEBUG(D_OTHER, "object "DFID" open, skip\n",
2454 PFID(mdt_object_fid(obj)));
2455 GOTO(unlock_child, rc = -EBUSY);
2459 ma->ma_valid = MA_INODE;
2460 ma->ma_attr.la_valid = LA_CTIME;
2461 ma->ma_attr.la_ctime = ctime;
2463 mutex_lock(&obj->mot_lov_mutex);
2465 rc = mdo_unlink(info->mti_env, mdt_object_child(pobj),
2466 mdt_object_child(obj), name, ma, 0);
2468 mutex_unlock(&obj->mot_lov_mutex);
2471 mdt_reint_striped_unlock(info, obj, child_lh, einfo, 1);
2473 mdt_object_unlock(info, pobj, parent_lh, 1);
2475 mdt_object_put(info->mti_env, pobj);
2480 static int mdt_rmfid_check_permission(struct mdt_thread_info *info,
2481 struct mdt_object *obj)
2483 struct lu_ucred *uc = lu_ucred(info->mti_env);
2484 struct md_attr *ma = &info->mti_attr;
2485 struct lu_attr *la = &ma->ma_attr;
2489 ma->ma_need = MA_INODE;
2490 rc = mo_attr_get(info->mti_env, mdt_object_child(obj), ma);
2494 if (la->la_flags & LUSTRE_IMMUTABLE_FL)
2497 if (cap_raised(uc->uc_cap, CAP_DAC_OVERRIDE))
2499 if (uc->uc_fsuid == la->la_uid) {
2500 if ((la->la_mode & S_IWUSR) == 0)
2502 } else if (uc->uc_fsgid == la->la_gid) {
2503 if ((la->la_mode & S_IWGRP) == 0)
2505 } else if ((la->la_mode & S_IWOTH) == 0) {
2513 static int mdt_rmfid_one(struct mdt_thread_info *info, struct lu_fid *fid,
2516 struct mdt_device *mdt = info->mti_mdt;
2517 struct mdt_object *obj = NULL;
2518 struct linkea_data ldata = { NULL };
2519 struct lu_buf *buf = &info->mti_big_buf;
2520 struct lu_name *name = &info->mti_name;
2521 struct lu_fid *pfid = &info->mti_tmp_fid1;
2522 struct link_ea_header *leh;
2523 struct link_ea_entry *lee;
2524 int reclen, count, rc = 0;
2527 if (!fid_is_sane(fid))
2528 GOTO(out, rc = -EINVAL);
2530 if (!fid_is_namespace_visible(fid))
2531 GOTO(out, rc = -EINVAL);
2533 obj = mdt_object_find(info->mti_env, mdt, fid);
2535 GOTO(out, rc = PTR_ERR(obj));
2537 if (mdt_object_remote(obj))
2538 GOTO(out, rc = -EREMOTE);
2539 if (!mdt_object_exists(obj) || lu_object_is_dying(&obj->mot_header))
2540 GOTO(out, rc = -ENOENT);
2542 rc = mdt_rmfid_check_permission(info, obj);
2547 buf = lu_buf_check_and_alloc(buf, PATH_MAX);
2549 GOTO(out, rc = -ENOMEM);
2552 rc = mdt_links_read(info, obj, &ldata);
2557 lee = (struct link_ea_entry *)(leh + 1);
2558 for (count = 0; count < leh->leh_reccount; count++) {
2559 /* remove every hardlink */
2560 linkea_entry_unpack(lee, &reclen, name, pfid);
2561 lee = (struct link_ea_entry *) ((char *)lee + reclen);
2562 rc = mdt_rmfid_unlink(info, pfid, name, obj, ctime);
2568 if (obj && !IS_ERR(obj))
2569 mdt_object_put(info->mti_env, obj);
2570 if (info->mti_big_buf.lb_buf)
2571 lu_buf_free(&info->mti_big_buf);
2576 static int mdt_rmfid(struct tgt_session_info *tsi)
2578 struct mdt_thread_info *mti = tsi2mdt_info(tsi);
2579 struct mdt_body *reqbody;
2580 struct lu_fid *fids, *rfids;
2586 reqbody = req_capsule_client_get(tsi->tsi_pill, &RMF_MDT_BODY);
2587 if (reqbody == NULL)
2589 bufsize = req_capsule_get_size(tsi->tsi_pill, &RMF_FID_ARRAY,
2591 nr = bufsize / sizeof(struct lu_fid);
2592 if (nr * sizeof(struct lu_fid) != bufsize)
2594 req_capsule_set_size(tsi->tsi_pill, &RMF_RCS,
2595 RCL_SERVER, nr * sizeof(__u32));
2596 req_capsule_set_size(tsi->tsi_pill, &RMF_FID_ARRAY,
2597 RCL_SERVER, nr * sizeof(struct lu_fid));
2598 rc = req_capsule_server_pack(tsi->tsi_pill);
2600 GOTO(out, rc = err_serious(rc));
2601 fids = req_capsule_client_get(tsi->tsi_pill, &RMF_FID_ARRAY);
2604 rcs = req_capsule_server_get(tsi->tsi_pill, &RMF_RCS);
2606 rfids = req_capsule_server_get(tsi->tsi_pill, &RMF_FID_ARRAY);
2609 mdt_init_ucred(mti, reqbody);
2610 for (i = 0; i < nr; i++) {
2612 rcs[i] = mdt_rmfid_one(mti, fids + i, reqbody->mbo_ctime);
2614 mdt_exit_ucred(mti);
2620 static int mdt_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2621 void *karg, void __user *uarg);
2623 static int mdt_set_info(struct tgt_session_info *tsi)
2625 struct ptlrpc_request *req = tgt_ses_req(tsi);
2628 int keylen, vallen, rc = 0;
2632 key = req_capsule_client_get(tsi->tsi_pill, &RMF_SETINFO_KEY);
2634 DEBUG_REQ(D_HA, req, "no set_info key");
2635 RETURN(err_serious(-EFAULT));
2638 keylen = req_capsule_get_size(tsi->tsi_pill, &RMF_SETINFO_KEY,
2641 val = req_capsule_client_get(tsi->tsi_pill, &RMF_SETINFO_VAL);
2643 DEBUG_REQ(D_HA, req, "no set_info val");
2644 RETURN(err_serious(-EFAULT));
2647 vallen = req_capsule_get_size(tsi->tsi_pill, &RMF_SETINFO_VAL,
2650 /* Swab any part of val you need to here */
2651 if (KEY_IS(KEY_READ_ONLY)) {
2652 spin_lock(&req->rq_export->exp_lock);
2654 *exp_connect_flags_ptr(req->rq_export) |=
2657 *exp_connect_flags_ptr(req->rq_export) &=
2658 ~OBD_CONNECT_RDONLY;
2659 spin_unlock(&req->rq_export->exp_lock);
2660 } else if (KEY_IS(KEY_CHANGELOG_CLEAR)) {
2661 struct changelog_setinfo *cs = val;
2663 if (vallen != sizeof(*cs)) {
2664 CERROR("%s: bad changelog_clear setinfo size %d\n",
2665 tgt_name(tsi->tsi_tgt), vallen);
2668 if (req_capsule_req_need_swab(&req->rq_pill)) {
2669 __swab64s(&cs->cs_recno);
2670 __swab32s(&cs->cs_id);
2673 if (!mdt_is_rootadmin(tsi2mdt_info(tsi)))
2675 rc = mdt_iocontrol(OBD_IOC_CHANGELOG_CLEAR, req->rq_export,
2677 } else if (KEY_IS(KEY_EVICT_BY_NID)) {
2679 obd_export_evict_by_nid(req->rq_export->exp_obd, val);
2686 static int mdt_readpage(struct tgt_session_info *tsi)
2688 struct mdt_thread_info *info = mdt_th_info(tsi->tsi_env);
2689 struct mdt_object *object = mdt_obj(tsi->tsi_corpus);
2690 struct lu_rdpg *rdpg = &info->mti_u.rdpg.mti_rdpg;
2691 const struct mdt_body *reqbody = tsi->tsi_mdt_body;
2692 struct mdt_body *repbody;
2698 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK))
2699 RETURN(err_serious(-ENOMEM));
2701 repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_MDT_BODY);
2702 if (repbody == NULL || reqbody == NULL)
2703 RETURN(err_serious(-EFAULT));
2706 * prepare @rdpg before calling lower layers and transfer itself. Here
2707 * reqbody->size contains offset of where to start to read and
2708 * reqbody->nlink contains number bytes to read.
2710 rdpg->rp_hash = reqbody->mbo_size;
2711 if (rdpg->rp_hash != reqbody->mbo_size) {
2712 CERROR("Invalid hash: %#llx != %#llx\n",
2713 rdpg->rp_hash, reqbody->mbo_size);
2717 rdpg->rp_attrs = reqbody->mbo_mode;
2718 if (exp_connect_flags(tsi->tsi_exp) & OBD_CONNECT_64BITHASH)
2719 rdpg->rp_attrs |= LUDA_64BITHASH;
2720 rdpg->rp_count = min_t(unsigned int, reqbody->mbo_nlink,
2721 exp_max_brw_size(tsi->tsi_exp));
2722 rdpg->rp_npages = (rdpg->rp_count + PAGE_SIZE - 1) >>
2724 OBD_ALLOC_PTR_ARRAY_LARGE(rdpg->rp_pages, rdpg->rp_npages);
2725 if (rdpg->rp_pages == NULL)
2728 for (i = 0; i < rdpg->rp_npages; ++i) {
2729 rdpg->rp_pages[i] = alloc_page(GFP_NOFS);
2730 if (rdpg->rp_pages[i] == NULL)
2731 GOTO(free_rdpg, rc = -ENOMEM);
2734 /* call lower layers to fill allocated pages with directory data */
2735 rc = mo_readpage(tsi->tsi_env, mdt_object_child(object), rdpg);
2737 GOTO(free_rdpg, rc);
2739 /* send pages to client */
2740 rc = tgt_sendpage(tsi, rdpg, rc);
2745 for (i = 0; i < rdpg->rp_npages; i++)
2746 if (rdpg->rp_pages[i] != NULL)
2747 __free_page(rdpg->rp_pages[i]);
2748 OBD_FREE_PTR_ARRAY_LARGE(rdpg->rp_pages, rdpg->rp_npages);
2750 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE))
2756 static int mdt_fix_attr_ucred(struct mdt_thread_info *info, __u32 op)
2758 struct lu_ucred *uc = mdt_ucred_check(info);
2759 struct lu_attr *attr = &info->mti_attr.ma_attr;
2764 if (op != REINT_SETATTR) {
2765 if ((attr->la_valid & LA_UID) && (attr->la_uid != -1))
2766 attr->la_uid = uc->uc_fsuid;
2767 /* for S_ISGID, inherit gid from his parent, such work will be
2768 * done in cmm/mdd layer, here set all cases as uc->uc_fsgid. */
2769 if ((attr->la_valid & LA_GID) && (attr->la_gid != -1))
2770 attr->la_gid = uc->uc_fsgid;
2776 static inline bool mdt_is_readonly_open(struct mdt_thread_info *info, __u32 op)
2778 return op == REINT_OPEN &&
2779 !(info->mti_spec.sp_cr_flags & (MDS_FMODE_WRITE | MDS_OPEN_CREAT));
2782 static void mdt_preset_secctx_size(struct mdt_thread_info *info)
2784 struct req_capsule *pill = info->mti_pill;
2786 if (req_capsule_has_field(pill, &RMF_FILE_SECCTX,
2788 req_capsule_has_field(pill, &RMF_FILE_SECCTX_NAME,
2790 if (req_capsule_get_size(pill, &RMF_FILE_SECCTX_NAME,
2792 /* pre-set size in server part with max size */
2793 req_capsule_set_size(pill, &RMF_FILE_SECCTX,
2795 OBD_MAX_DEFAULT_EA_SIZE);
2797 req_capsule_set_size(pill, &RMF_FILE_SECCTX,
2802 static void mdt_preset_encctx_size(struct mdt_thread_info *info)
2804 struct req_capsule *pill = info->mti_pill;
2806 if (req_capsule_has_field(pill, &RMF_FILE_ENCCTX,
2808 /* pre-set size in server part with max size */
2809 req_capsule_set_size(pill, &RMF_FILE_ENCCTX,
2811 info->mti_mdt->mdt_max_mdsize);
2814 static int mdt_reint_internal(struct mdt_thread_info *info,
2815 struct mdt_lock_handle *lhc,
2818 struct req_capsule *pill = info->mti_pill;
2819 struct mdt_body *repbody;
2824 rc = mdt_reint_unpack(info, op);
2826 CERROR("Can't unpack reint, rc %d\n", rc);
2827 RETURN(err_serious(rc));
2831 /* check if the file system is set to readonly. O_RDONLY open
2832 * is still allowed even the file system is set to readonly mode */
2833 if (mdt_rdonly(info->mti_exp) && !mdt_is_readonly_open(info, op))
2834 RETURN(err_serious(-EROFS));
2836 /* for replay (no_create) lmm is not needed, client has it already */
2837 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
2838 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER,
2841 /* llog cookies are always 0, the field is kept for compatibility */
2842 if (req_capsule_has_field(pill, &RMF_LOGCOOKIES, RCL_SERVER))
2843 req_capsule_set_size(pill, &RMF_LOGCOOKIES, RCL_SERVER, 0);
2845 /* Set ACL reply buffer size as LUSTRE_POSIX_ACL_MAX_SIZE_OLD
2846 * by default. If the target object has more ACL entries, then
2847 * enlarge the buffer when necessary. */
2848 if (req_capsule_has_field(pill, &RMF_ACL, RCL_SERVER))
2849 req_capsule_set_size(pill, &RMF_ACL, RCL_SERVER,
2850 LUSTRE_POSIX_ACL_MAX_SIZE_OLD);
2852 mdt_preset_secctx_size(info);
2853 mdt_preset_encctx_size(info);
2855 rc = req_capsule_server_pack(pill);
2857 CERROR("Can't pack response, rc %d\n", rc);
2858 RETURN(err_serious(rc));
2861 if (req_capsule_has_field(pill, &RMF_MDT_BODY, RCL_SERVER)) {
2862 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
2864 repbody->mbo_eadatasize = 0;
2865 repbody->mbo_aclsize = 0;
2868 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_REINT_DELAY, 10);
2870 /* for replay no cookkie / lmm need, because client have this already */
2871 if (info->mti_spec.no_create)
2872 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
2873 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER, 0);
2875 rc = mdt_init_ucred_reint(info);
2877 GOTO(out_shrink, rc);
2879 rc = mdt_fix_attr_ucred(info, op);
2881 GOTO(out_ucred, rc = err_serious(rc));
2883 rc = mdt_check_resent(info, mdt_reconstruct, lhc);
2885 GOTO(out_ucred, rc);
2886 } else if (rc == 1) {
2887 DEBUG_REQ(D_INODE, mdt_info_req(info), "resent opt");
2888 rc = lustre_msg_get_status(mdt_info_req(info)->rq_repmsg);
2889 GOTO(out_ucred, rc);
2891 rc = mdt_reint_rec(info, lhc);
2894 mdt_exit_ucred(info);
2896 mdt_client_compatibility(info);
2898 rc2 = mdt_fix_reply(info);
2903 * Data-on-MDT optimization - read data along with OPEN and return it
2904 * in reply when possible.
2906 if (rc == 0 && op == REINT_OPEN && !req_is_replay(pill->rc_req))
2907 rc = mdt_dom_read_on_open(info, info->mti_mdt,
2913 static long mdt_reint_opcode(struct ptlrpc_request *req,
2914 const struct req_format **fmt)
2916 struct mdt_device *mdt;
2917 struct mdt_rec_reint *rec;
2920 rec = req_capsule_client_get(&req->rq_pill, &RMF_REC_REINT);
2922 opc = rec->rr_opcode;
2923 DEBUG_REQ(D_INODE, req, "reint opt = %ld", opc);
2924 if (opc < REINT_MAX && fmt[opc] != NULL)
2925 req_capsule_extend(&req->rq_pill, fmt[opc]);
2927 mdt = mdt_exp2dev(req->rq_export);
2928 CERROR("%s: Unsupported opcode '%ld' from client '%s':"
2929 " rc = %d\n", req->rq_export->exp_obd->obd_name,
2930 opc, mdt->mdt_ldlm_client->cli_name, -EFAULT);
2931 opc = err_serious(-EFAULT);
2934 opc = err_serious(-EFAULT);
2939 static int mdt_reint(struct tgt_session_info *tsi)
2943 static const struct req_format *reint_fmts[REINT_MAX] = {
2944 [REINT_SETATTR] = &RQF_MDS_REINT_SETATTR,
2945 [REINT_CREATE] = &RQF_MDS_REINT_CREATE,
2946 [REINT_LINK] = &RQF_MDS_REINT_LINK,
2947 [REINT_UNLINK] = &RQF_MDS_REINT_UNLINK,
2948 [REINT_RENAME] = &RQF_MDS_REINT_RENAME,
2949 [REINT_OPEN] = &RQF_MDS_REINT_OPEN,
2950 [REINT_SETXATTR] = &RQF_MDS_REINT_SETXATTR,
2951 [REINT_RMENTRY] = &RQF_MDS_REINT_UNLINK,
2952 [REINT_MIGRATE] = &RQF_MDS_REINT_MIGRATE,
2953 [REINT_RESYNC] = &RQF_MDS_REINT_RESYNC,
2958 opc = mdt_reint_opcode(tgt_ses_req(tsi), reint_fmts);
2960 struct mdt_thread_info *info = tsi2mdt_info(tsi);
2962 * No lock possible here from client to pass it to reint code
2965 rc = mdt_reint_internal(info, NULL, opc);
2966 mdt_thread_info_fini(info);
2971 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_REINT_NET_REP;
2975 /* this should sync the whole device */
2976 int mdt_device_sync(const struct lu_env *env, struct mdt_device *mdt)
2978 struct dt_device *dt = mdt->mdt_bottom;
2982 rc = dt->dd_ops->dt_sync(env, dt);
2986 /* this should sync this object */
2987 static int mdt_object_sync(const struct lu_env *env, struct obd_export *exp,
2988 struct mdt_object *mo)
2994 if (!mdt_object_exists(mo)) {
2995 CWARN("%s: non existing object "DFID": rc = %d\n",
2996 exp->exp_obd->obd_name, PFID(mdt_object_fid(mo)),
3001 if (S_ISREG(lu_object_attr(&mo->mot_obj))) {
3002 struct lu_target *tgt = tgt_ses_info(env)->tsi_tgt;
3003 dt_obj_version_t version;
3005 version = dt_version_get(env, mdt_obj2dt(mo));
3006 if (version > tgt->lut_obd->obd_last_committed)
3007 rc = mo_object_sync(env, mdt_object_child(mo));
3009 rc = mo_object_sync(env, mdt_object_child(mo));
3015 static int mdt_sync(struct tgt_session_info *tsi)
3017 struct ptlrpc_request *req = tgt_ses_req(tsi);
3018 struct req_capsule *pill = tsi->tsi_pill;
3019 struct mdt_body *body;
3020 ktime_t kstart = ktime_get();
3025 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SYNC_PACK))
3026 RETURN(err_serious(-ENOMEM));
3028 if (fid_seq(&tsi->tsi_mdt_body->mbo_fid1) == 0) {
3029 rc = mdt_device_sync(tsi->tsi_env, mdt_exp2dev(tsi->tsi_exp));
3031 struct mdt_thread_info *info = tsi2mdt_info(tsi);
3033 if (unlikely(info->mti_object == NULL))
3036 /* sync an object */
3037 rc = mdt_object_sync(tsi->tsi_env, tsi->tsi_exp,
3040 const struct lu_fid *fid;
3041 struct lu_attr *la = &info->mti_attr.ma_attr;
3043 info->mti_attr.ma_need = MA_INODE;
3044 info->mti_attr.ma_valid = 0;
3045 rc = mdt_attr_get_complex(info, info->mti_object,
3048 body = req_capsule_server_get(pill,
3050 fid = mdt_object_fid(info->mti_object);
3051 mdt_pack_attr2body(info, body, la, fid);
3054 mdt_thread_info_fini(info);
3057 mdt_counter_incr(req, LPROC_MDT_SYNC,
3058 ktime_us_delta(ktime_get(), kstart));
3063 static int mdt_data_sync(struct tgt_session_info *tsi)
3065 struct mdt_thread_info *info;
3066 struct mdt_device *mdt = mdt_exp2dev(tsi->tsi_exp);
3067 struct ost_body *body = tsi->tsi_ost_body;
3068 struct ost_body *repbody;
3069 struct mdt_object *mo = NULL;
3075 repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_OST_BODY);
3077 /* if no fid is specified then do nothing,
3078 * device sync is done via MDS_SYNC */
3079 if (fid_is_zero(&tsi->tsi_fid))
3082 mo = mdt_object_find(tsi->tsi_env, mdt, &tsi->tsi_fid);
3084 RETURN(PTR_ERR(mo));
3086 rc = mdt_object_sync(tsi->tsi_env, tsi->tsi_exp, mo);
3090 repbody->oa.o_oi = body->oa.o_oi;
3091 repbody->oa.o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
3093 info = tsi2mdt_info(tsi);
3094 ma = &info->mti_attr;
3095 ma->ma_need = MA_INODE;
3097 rc = mdt_attr_get_complex(info, mo, ma);
3099 obdo_from_la(&repbody->oa, &ma->ma_attr, VALID_FLAGS);
3102 mdt_thread_info_fini(info);
3107 mdt_object_put(tsi->tsi_env, mo);
3112 * Handle quota control requests to consult current usage/limit, but also
3113 * to configure quota enforcement
3115 static int mdt_quotactl(struct tgt_session_info *tsi)
3117 struct obd_export *exp = tsi->tsi_exp;
3118 struct req_capsule *pill = tsi->tsi_pill;
3119 struct obd_quotactl *oqctl, *repoqc;
3121 struct mdt_device *mdt = mdt_exp2dev(exp);
3122 struct lu_device *qmt = mdt->mdt_qmt_dev;
3123 struct lu_nodemap *nodemap;
3126 oqctl = req_capsule_client_get(pill, &RMF_OBD_QUOTACTL);
3128 RETURN(err_serious(-EPROTO));
3130 rc = req_capsule_server_pack(pill);
3132 RETURN(err_serious(rc));
3134 nodemap = nodemap_get_from_exp(exp);
3135 if (IS_ERR(nodemap))
3136 RETURN(PTR_ERR(nodemap));
3138 switch (oqctl->qc_cmd) {
3139 /* master quotactl */
3142 case LUSTRE_Q_SETDEFAULT:
3143 case LUSTRE_Q_SETQUOTAPOOL:
3144 case LUSTRE_Q_SETINFOPOOL:
3145 case LUSTRE_Q_SETDEFAULT_POOL:
3146 if (!nodemap_can_setquota(nodemap))
3147 GOTO(out_nodemap, rc = -EPERM);
3151 case LUSTRE_Q_GETDEFAULT:
3152 case LUSTRE_Q_GETQUOTAPOOL:
3153 case LUSTRE_Q_GETINFOPOOL:
3154 case LUSTRE_Q_GETDEFAULT_POOL:
3156 GOTO(out_nodemap, rc = -EOPNOTSUPP);
3157 /* slave quotactl */
3164 CERROR("%s: unsupported quotactl command %d: rc = %d\n",
3165 mdt_obd_name(mdt), oqctl->qc_cmd, rc);
3166 GOTO(out_nodemap, rc);
3170 switch (oqctl->qc_type) {
3172 id = nodemap_map_id(nodemap, NODEMAP_UID,
3173 NODEMAP_CLIENT_TO_FS, id);
3176 id = nodemap_map_id(nodemap, NODEMAP_GID,
3177 NODEMAP_CLIENT_TO_FS, id);
3180 /* todo: check/map project id */
3184 GOTO(out_nodemap, rc = -EOPNOTSUPP);
3186 repoqc = req_capsule_server_get(pill, &RMF_OBD_QUOTACTL);
3188 GOTO(out_nodemap, rc = err_serious(-EFAULT));
3190 if (oqctl->qc_cmd == Q_SETINFO || oqctl->qc_cmd == Q_SETQUOTA)
3191 barrier_exit(tsi->tsi_tgt->lut_bottom);
3193 if (oqctl->qc_id != id)
3194 swap(oqctl->qc_id, id);
3196 if (oqctl->qc_cmd == Q_SETINFO || oqctl->qc_cmd == Q_SETQUOTA) {
3197 if (unlikely(!barrier_entry(tsi->tsi_tgt->lut_bottom)))
3198 RETURN(-EINPROGRESS);
3201 switch (oqctl->qc_cmd) {
3207 case LUSTRE_Q_SETDEFAULT:
3208 case LUSTRE_Q_GETDEFAULT:
3209 case LUSTRE_Q_SETQUOTAPOOL:
3210 case LUSTRE_Q_GETQUOTAPOOL:
3211 case LUSTRE_Q_SETINFOPOOL:
3212 case LUSTRE_Q_GETINFOPOOL:
3213 case LUSTRE_Q_SETDEFAULT_POOL:
3214 case LUSTRE_Q_GETDEFAULT_POOL:
3215 /* forward quotactl request to QMT */
3216 rc = qmt_hdls.qmth_quotactl(tsi->tsi_env, qmt, oqctl);
3221 /* slave quotactl */
3222 rc = lquotactl_slv(tsi->tsi_env, tsi->tsi_tgt->lut_bottom,
3227 CERROR("Unsupported quotactl command: %d\n", oqctl->qc_cmd);
3228 GOTO(out_nodemap, rc = -EFAULT);
3231 if (oqctl->qc_id != id)
3232 swap(oqctl->qc_id, id);
3234 QCTL_COPY(repoqc, oqctl);
3238 nodemap_putref(nodemap);
3243 /** clone llog ctxt from child (mdd)
3244 * This allows remote llog (replicator) access.
3245 * We can either pass all llog RPCs (eg mdt_llog_create) on to child where the
3246 * context was originally set up, or we can handle them directly.
3247 * I choose the latter, but that means I need any llog
3248 * contexts set up by child to be accessable by the mdt. So we clone the
3249 * context into our context list here.
3251 static int mdt_llog_ctxt_clone(const struct lu_env *env, struct mdt_device *mdt,
3254 struct md_device *next = mdt->mdt_child;
3255 struct llog_ctxt *ctxt;
3258 if (!llog_ctxt_null(mdt2obd_dev(mdt), idx))
3261 rc = next->md_ops->mdo_llog_ctxt_get(env, next, idx, (void **)&ctxt);
3262 if (rc || ctxt == NULL) {
3266 rc = llog_group_set_ctxt(&mdt2obd_dev(mdt)->obd_olg, ctxt, idx);
3268 CERROR("Can't set mdt ctxt %d\n", rc);
3273 static int mdt_llog_ctxt_unclone(const struct lu_env *env,
3274 struct mdt_device *mdt, int idx)
3276 struct llog_ctxt *ctxt;
3278 ctxt = llog_get_context(mdt2obd_dev(mdt), idx);
3281 /* Put once for the get we just did, and once for the clone */
3282 llog_ctxt_put(ctxt);
3283 llog_ctxt_put(ctxt);
3288 * sec context handlers
3290 static int mdt_sec_ctx_handle(struct tgt_session_info *tsi)
3292 CFS_FAIL_TIMEOUT(OBD_FAIL_SEC_CTX_HDL_PAUSE, cfs_fail_val);
3298 * quota request handlers
3300 static int mdt_quota_dqacq(struct tgt_session_info *tsi)
3302 struct mdt_device *mdt = mdt_exp2dev(tsi->tsi_exp);
3303 struct lu_device *qmt = mdt->mdt_qmt_dev;
3308 RETURN(err_serious(-EOPNOTSUPP));
3310 rc = qmt_hdls.qmth_dqacq(tsi->tsi_env, qmt, tgt_ses_req(tsi));
3314 struct mdt_object *mdt_object_new(const struct lu_env *env,
3315 struct mdt_device *d,
3316 const struct lu_fid *f)
3318 struct lu_object_conf conf = { .loc_flags = LOC_F_NEW };
3319 struct lu_object *o;
3320 struct mdt_object *m;
3323 CDEBUG(D_INFO, "Allocate object for "DFID"\n", PFID(f));
3324 o = lu_object_find(env, &d->mdt_lu_dev, f, &conf);
3325 if (unlikely(IS_ERR(o)))
3326 m = (struct mdt_object *)o;
3332 struct mdt_object *mdt_object_find(const struct lu_env *env,
3333 struct mdt_device *d,
3334 const struct lu_fid *f)
3336 struct lu_object *o;
3337 struct mdt_object *m;
3340 CDEBUG(D_INFO, "Find object for "DFID"\n", PFID(f));
3341 o = lu_object_find(env, &d->mdt_lu_dev, f, NULL);
3342 if (unlikely(IS_ERR(o)))
3343 m = (struct mdt_object *)o;
3351 * Asyncronous commit for mdt device.
3353 * Pass asynchonous commit call down the MDS stack.
3355 * \param env environment
3356 * \param mdt the mdt device
3358 static void mdt_device_commit_async(const struct lu_env *env,
3359 struct mdt_device *mdt)
3361 struct dt_device *dt = mdt->mdt_bottom;
3365 rc = dt->dd_ops->dt_commit_async(env, dt);
3366 if (unlikely(rc != 0))
3367 CWARN("%s: async commit start failed: rc = %d\n",
3368 mdt_obd_name(mdt), rc);
3369 atomic_inc(&mdt->mdt_async_commit_count);
3374 * Mark the lock as "synchonous".
3376 * Mark the lock to deffer transaction commit to the unlock time.
3378 * \param lock the lock to mark as "synchonous"
3380 * \see mdt_is_lock_sync
3381 * \see mdt_save_lock
3383 static inline void mdt_set_lock_sync(struct ldlm_lock *lock)
3385 lock->l_ast_data = (void*)1;
3389 * Check whehter the lock "synchonous" or not.
3391 * \param lock the lock to check
3392 * \retval 1 the lock is "synchonous"
3393 * \retval 0 the lock isn't "synchronous"
3395 * \see mdt_set_lock_sync
3396 * \see mdt_save_lock
3398 static inline int mdt_is_lock_sync(struct ldlm_lock *lock)
3400 return lock->l_ast_data != NULL;
3404 * Blocking AST for mdt locks.
3406 * Starts transaction commit if in case of COS lock conflict or
3407 * deffers such a commit to the mdt_save_lock.
3409 * \param lock the lock which blocks a request or cancelling lock
3410 * \param desc unused
3411 * \param data unused
3412 * \param flag indicates whether this cancelling or blocking callback
3414 * \see ldlm_blocking_ast_nocheck
3416 int mdt_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
3417 void *data, int flag)
3419 struct obd_device *obd = ldlm_lock_to_ns(lock)->ns_obd;
3420 struct mdt_device *mdt = mdt_dev(obd->obd_lu_dev);
3421 struct ldlm_cb_set_arg *arg = data;
3422 bool commit_async = false;
3426 if (flag == LDLM_CB_CANCELING)
3429 lock_res_and_lock(lock);
3430 if (lock->l_blocking_ast != mdt_blocking_ast) {
3431 unlock_res_and_lock(lock);
3435 /* A blocking ast may be sent from ldlm_lock_decref_internal
3436 * when the last reference to a local lock was released and
3437 * during blocking event from ldlm_work_bl_ast_lock().
3438 * The 'data' parameter is l_ast_data in the first case and
3439 * callback arguments in the second one. Distinguish them by that.
3441 if (!data || data == lock->l_ast_data || !arg->bl_desc)
3442 goto skip_cos_checks;
3444 if (lock->l_req_mode & (LCK_PW | LCK_EX)) {
3445 if (mdt_cos_is_enabled(mdt)) {
3446 if (!arg->bl_desc->bl_same_client)
3447 mdt_set_lock_sync(lock);
3448 } else if (mdt_slc_is_enabled(mdt) &&
3449 arg->bl_desc->bl_cos_incompat) {
3450 mdt_set_lock_sync(lock);
3452 * we may do extra commit here, but there is a small
3453 * window to miss a commit: lock was unlocked (saved),
3454 * then a conflict lock queued and we come here, but
3455 * REP-ACK not received, so lock was not converted to
3457 * Fortunately this window is quite small, so the
3458 * extra commit should be rare (not to say distributed
3459 * operation is rare too).
3461 commit_async = true;
3463 } else if (lock->l_req_mode == LCK_COS) {
3464 commit_async = true;
3468 rc = ldlm_blocking_ast_nocheck(lock);
3473 rc = lu_env_init(&env, LCT_LOCAL);
3474 if (unlikely(rc != 0))
3475 CWARN("%s: lu_env initialization failed, cannot "
3476 "start asynchronous commit: rc = %d\n",
3479 mdt_device_commit_async(&env, mdt);
3486 * Blocking AST for cross-MDT lock
3488 * Discard lock from uncommitted_slc_locks and cancel it.
3490 * \param lock the lock which blocks a request or cancelling lock
3491 * \param desc unused
3492 * \param data unused
3493 * \param flag indicates whether this cancelling or blocking callback
3494 * \retval 0 on success
3495 * \retval negative number on error
3497 int mdt_remote_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
3498 void *data, int flag)
3504 case LDLM_CB_BLOCKING: {
3505 struct lustre_handle lockh;
3507 ldlm_lock2handle(lock, &lockh);
3508 rc = ldlm_cli_cancel(&lockh,
3509 ldlm_is_atomic_cb(lock) ? 0 : LCF_ASYNC);
3511 CDEBUG(D_INODE, "ldlm_cli_cancel: %d\n", rc);
3516 case LDLM_CB_CANCELING: {
3517 struct obd_device *obd = ldlm_lock_to_ns(lock)->ns_obd;
3518 struct mdt_device *mdt =
3519 mdt_dev(obd->obd_lu_dev->ld_site->ls_top_dev);
3521 LDLM_DEBUG(lock, "Revoke remote lock\n");
3523 /* discard slc lock here so that it can be cleaned anytime,
3524 * especially for cleanup_resource() */
3525 tgt_discard_slc_lock(&mdt->mdt_lut, lock);
3527 /* once we cache lock, l_ast_data is set to mdt_object */
3528 if (lock->l_ast_data != NULL) {
3529 struct mdt_object *mo = lock->l_ast_data;
3532 rc = lu_env_init(&env, LCT_MD_THREAD);
3533 if (unlikely(rc != 0)) {
3534 CWARN("%s: lu_env initialization failed, object %p "DFID" is leaked!: rc = %d\n",
3536 PFID(mdt_object_fid(mo)), rc);
3540 if (lock->l_policy_data.l_inodebits.bits &
3541 (MDS_INODELOCK_XATTR | MDS_INODELOCK_UPDATE)) {
3542 rc = mo_invalidate(&env, mdt_object_child(mo));
3543 mo->mot_cache_attr = 0;
3545 mdt_object_put(&env, mo);
3557 int mdt_check_resent_lock(struct mdt_thread_info *info,
3558 struct mdt_object *mo,
3559 struct mdt_lock_handle *lhc)
3561 /* the lock might already be gotten in ldlm_handle_enqueue() */
3562 if (unlikely(lustre_handle_is_used(&lhc->mlh_reg_lh))) {
3563 struct ptlrpc_request *req = mdt_info_req(info);
3564 struct ldlm_lock *lock;
3566 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
3567 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
3569 /* Lock is pinned by ldlm_handle_enqueue0() as it is
3570 * a resend case, however, it could be already destroyed
3571 * due to client eviction or a raced cancel RPC. */
3572 LDLM_DEBUG_NOLOCK("Invalid lock handle %#llx",
3573 lhc->mlh_reg_lh.cookie);
3577 if (!fid_res_name_eq(mdt_object_fid(mo),
3578 &lock->l_resource->lr_name)) {
3579 CWARN("%s: Although resent, but still not "
3580 "get child lock:"DFID"\n",
3581 info->mti_exp->exp_obd->obd_name,
3582 PFID(mdt_object_fid(mo)));
3583 LDLM_LOCK_PUT(lock);
3586 LDLM_LOCK_PUT(lock);
3592 static void mdt_remote_object_lock_created_cb(struct ldlm_lock *lock)
3594 mdt_object_get(NULL, lock->l_ast_data);
3597 int mdt_remote_object_lock_try(struct mdt_thread_info *mti,
3598 struct mdt_object *o, const struct lu_fid *fid,
3599 struct lustre_handle *lh, enum ldlm_mode mode,
3600 __u64 *ibits, __u64 trybits, bool cache)
3602 struct ldlm_enqueue_info *einfo = &mti->mti_remote_einfo;
3603 union ldlm_policy_data *policy = &mti->mti_policy;
3604 struct ldlm_res_id *res_id = &mti->mti_res_id;
3608 LASSERT(mdt_object_remote(o));
3610 fid_build_reg_res_name(fid, res_id);
3612 memset(einfo, 0, sizeof(*einfo));
3613 einfo->ei_type = LDLM_IBITS;
3614 einfo->ei_mode = mode;
3615 einfo->ei_cb_bl = mdt_remote_blocking_ast;
3616 einfo->ei_cb_cp = ldlm_completion_ast;
3617 einfo->ei_enq_slave = 0;
3618 einfo->ei_res_id = res_id;
3622 * if we cache lock, couple lock with mdt_object, so that object
3623 * can be easily found in lock ASTs.
3625 einfo->ei_cbdata = o;
3626 einfo->ei_cb_created = mdt_remote_object_lock_created_cb;
3629 memset(policy, 0, sizeof(*policy));
3630 policy->l_inodebits.bits = *ibits;
3631 policy->l_inodebits.try_bits = trybits;
3633 rc = mo_object_lock(mti->mti_env, mdt_object_child(o), lh, einfo,
3636 /* Return successfully acquired bits to a caller */
3638 struct ldlm_lock *lock = ldlm_handle2lock(lh);
3641 *ibits = lock->l_policy_data.l_inodebits.bits;
3642 LDLM_LOCK_PUT(lock);
3647 int mdt_remote_object_lock(struct mdt_thread_info *mti, struct mdt_object *o,
3648 const struct lu_fid *fid, struct lustre_handle *lh,
3649 enum ldlm_mode mode, __u64 ibits, bool cache)
3651 return mdt_remote_object_lock_try(mti, o, fid, lh, mode, &ibits, 0,
3655 int mdt_object_local_lock(struct mdt_thread_info *info, struct mdt_object *o,
3656 struct mdt_lock_handle *lh, __u64 *ibits,
3657 __u64 trybits, bool cos_incompat)
3659 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
3660 union ldlm_policy_data *policy = &info->mti_policy;
3661 struct ldlm_res_id *res_id = &info->mti_res_id;
3662 __u64 dlmflags = 0, *cookie = NULL;
3666 LASSERT(!lustre_handle_is_used(&lh->mlh_reg_lh));
3667 LASSERT(!lustre_handle_is_used(&lh->mlh_pdo_lh));
3668 LASSERT(lh->mlh_reg_mode != LCK_MINMODE);
3669 LASSERT(lh->mlh_type != MDT_NUL_LOCK);
3672 LASSERT(lh->mlh_reg_mode == LCK_PW ||
3673 lh->mlh_reg_mode == LCK_EX);
3674 dlmflags |= LDLM_FL_COS_INCOMPAT;
3675 } else if (mdt_cos_is_enabled(info->mti_mdt)) {
3676 dlmflags |= LDLM_FL_COS_ENABLED;
3679 /* Only enqueue LOOKUP lock for remote object */
3680 LASSERT(ergo(mdt_object_remote(o), *ibits == MDS_INODELOCK_LOOKUP));
3682 /* Lease lock are granted with LDLM_FL_CANCEL_ON_BLOCK */
3683 if (lh->mlh_type == MDT_REG_LOCK && lh->mlh_reg_mode == LCK_EX &&
3684 *ibits == MDS_INODELOCK_OPEN)
3685 dlmflags |= LDLM_FL_CANCEL_ON_BLOCK;
3687 if (lh->mlh_type == MDT_PDO_LOCK) {
3688 /* check for exists after object is locked */
3689 if (mdt_object_exists(o) == 0) {
3690 /* Non-existent object shouldn't have PDO lock */
3693 /* Non-dir object shouldn't have PDO lock */
3694 if (!S_ISDIR(lu_object_attr(&o->mot_obj)))
3699 fid_build_reg_res_name(mdt_object_fid(o), res_id);
3700 dlmflags |= LDLM_FL_ATOMIC_CB;
3703 cookie = &info->mti_exp->exp_handle.h_cookie;
3706 * Take PDO lock on whole directory and build correct @res_id for lock
3707 * on part of directory.
3709 if (lh->mlh_pdo_hash != 0) {
3710 LASSERT(lh->mlh_type == MDT_PDO_LOCK);
3711 mdt_lock_pdo_mode(info, o, lh);
3712 if (lh->mlh_pdo_mode != LCK_NL) {
3714 * Do not use LDLM_FL_LOCAL_ONLY for parallel lock, it
3715 * is never going to be sent to client and we do not
3716 * want it slowed down due to possible cancels.
3718 policy->l_inodebits.bits =
3719 *ibits & MDS_INODELOCK_UPDATE;
3720 policy->l_inodebits.try_bits =
3721 trybits & MDS_INODELOCK_UPDATE;
3722 /* at least one of them should be set */
3723 LASSERT(policy->l_inodebits.bits |
3724 policy->l_inodebits.try_bits);
3725 rc = mdt_fid_lock(info->mti_env, ns, &lh->mlh_pdo_lh,
3726 lh->mlh_pdo_mode, policy, res_id,
3728 if (unlikely(rc != 0))
3729 GOTO(out_unlock, rc);
3733 * Finish res_id initializing by name hash marking part of
3734 * directory which is taking modification.
3736 res_id->name[LUSTRE_RES_ID_HSH_OFF] = lh->mlh_pdo_hash;
3739 policy->l_inodebits.bits = *ibits;
3740 policy->l_inodebits.try_bits = trybits;
3741 policy->l_inodebits.li_gid = lh->mlh_gid;
3744 * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
3745 * going to be sent to client. If it is - mdt_intent_policy() path will
3746 * fix it up and turn FL_LOCAL flag off.
3748 rc = mdt_fid_lock(info->mti_env, ns, &lh->mlh_reg_lh, lh->mlh_reg_mode,
3749 policy, res_id, LDLM_FL_LOCAL_ONLY | dlmflags,
3753 mdt_object_unlock(info, o, lh, 1);
3754 else if (unlikely(OBD_FAIL_PRECHECK(OBD_FAIL_MDS_PDO_LOCK)) &&
3755 lh->mlh_pdo_hash != 0 &&
3756 (lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX))
3757 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK, 15);
3759 /* Return successfully acquired bits to a caller */
3761 struct ldlm_lock *lock = ldlm_handle2lock(&lh->mlh_reg_lh);
3764 *ibits = lock->l_policy_data.l_inodebits.bits;
3765 LDLM_LOCK_PUT(lock);
3771 mdt_object_lock_internal(struct mdt_thread_info *info, struct mdt_object *o,
3772 struct mdt_lock_handle *lh, __u64 *ibits,
3773 __u64 trybits, bool cos_incompat)
3775 struct mdt_lock_handle *local_lh = NULL;
3779 if (!mdt_object_remote(o)) {
3780 rc = mdt_object_local_lock(info, o, lh, ibits, trybits,
3785 /* XXX do not support PERM/LAYOUT/XATTR lock for remote object yet */
3786 *ibits &= ~(MDS_INODELOCK_PERM | MDS_INODELOCK_LAYOUT |
3787 MDS_INODELOCK_XATTR);
3789 /* Only enqueue LOOKUP lock for remote object */
3790 if (*ibits & MDS_INODELOCK_LOOKUP) {
3791 __u64 local = MDS_INODELOCK_LOOKUP;
3793 rc = mdt_object_local_lock(info, o, lh, &local, 0,
3801 if ((*ibits | trybits) & MDS_INODELOCK_UPDATE) {
3802 /* Sigh, PDO needs to enqueue 2 locks right now, but
3803 * enqueue RPC can only request 1 lock, to avoid extra
3804 * RPC, so it will instead enqueue EX lock for remote
3805 * object anyway XXX*/
3806 if (lh->mlh_type == MDT_PDO_LOCK &&
3807 lh->mlh_pdo_hash != 0) {
3809 "%s: "DFID" convert PDO lock to EX lock.\n",
3810 mdt_obd_name(info->mti_mdt),
3811 PFID(mdt_object_fid(o)));
3812 lh->mlh_pdo_hash = 0;
3813 lh->mlh_rreg_mode = LCK_EX;
3814 lh->mlh_type = MDT_REG_LOCK;
3817 rc = mdt_remote_object_lock_try(info, o, mdt_object_fid(o),
3820 ibits, trybits, false);
3821 if (rc != ELDLM_OK) {
3822 if (local_lh != NULL)
3823 mdt_object_unlock(info, o, local_lh, rc);
3828 /* other components like LFSCK can use lockless access
3829 * and populate cache, so we better invalidate it */
3830 mo_invalidate(info->mti_env, mdt_object_child(o));
3835 int mdt_object_lock(struct mdt_thread_info *info, struct mdt_object *o,
3836 struct mdt_lock_handle *lh, __u64 ibits)
3838 return mdt_object_lock_internal(info, o, lh, &ibits, 0, false);
3841 int mdt_reint_object_lock(struct mdt_thread_info *info, struct mdt_object *o,
3842 struct mdt_lock_handle *lh, __u64 ibits,
3845 LASSERT(lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX);
3846 return mdt_object_lock_internal(info, o, lh, &ibits, 0,
3850 int mdt_object_lock_try(struct mdt_thread_info *info, struct mdt_object *o,
3851 struct mdt_lock_handle *lh, __u64 *ibits,
3852 __u64 trybits, bool cos_incompat)
3854 bool trylock_only = *ibits == 0;
3857 LASSERT(!(*ibits & trybits));
3858 rc = mdt_object_lock_internal(info, o, lh, ibits, trybits,
3860 if (rc && trylock_only) { /* clear error for try ibits lock only */
3861 LASSERT(*ibits == 0);
3868 * Save a lock within request object.
3870 * Keep the lock referenced until whether client ACK or transaction
3871 * commit happens or release the lock immediately depending on input
3872 * parameters. If COS is ON, a write lock is converted to COS lock
3875 * \param info thead info object
3876 * \param h lock handle
3877 * \param mode lock mode
3878 * \param decref force immediate lock releasing
3880 void mdt_save_lock(struct mdt_thread_info *info, struct lustre_handle *h,
3881 enum ldlm_mode mode, int decref)
3885 if (lustre_handle_is_used(h)) {
3886 if (decref || !info->mti_has_trans ||
3887 !(mode & (LCK_PW | LCK_EX))) {
3888 mdt_fid_unlock(h, mode);
3890 struct mdt_device *mdt = info->mti_mdt;
3891 struct ldlm_lock *lock = ldlm_handle2lock(h);
3892 struct ptlrpc_request *req = mdt_info_req(info);
3893 bool cos = mdt_cos_is_enabled(mdt);
3894 bool convert_lock = !cos && mdt_slc_is_enabled(mdt);
3896 LASSERTF(lock != NULL, "no lock for cookie %#llx\n",
3899 /* there is no request if mdt_object_unlock() is called
3900 * from mdt_export_cleanup()->mdt_add_dirty_flag() */
3901 if (likely(req != NULL)) {
3902 LDLM_DEBUG(lock, "save lock request %p reply "
3903 "state %p transno %lld\n", req,
3904 req->rq_reply_state, req->rq_transno);
3906 ldlm_lock_mode_downgrade(lock, LCK_COS);
3909 if (req->rq_export->exp_disconnected)
3910 mdt_fid_unlock(h, mode);
3912 ptlrpc_save_lock(req, h, mode, cos,
3915 mdt_fid_unlock(h, mode);
3917 if (mdt_is_lock_sync(lock)) {
3918 CDEBUG(D_HA, "found sync-lock,"
3919 " async commit started\n");
3920 mdt_device_commit_async(info->mti_env,
3923 LDLM_LOCK_PUT(lock);
3932 * Save cross-MDT lock in uncommitted_slc_locks
3934 * Keep the lock referenced until transaction commit happens or release the lock
3935 * immediately depending on input parameters.
3937 * \param info thead info object
3938 * \param h lock handle
3939 * \param mode lock mode
3940 * \param decref force immediate lock releasing
3942 static void mdt_save_remote_lock(struct mdt_thread_info *info,
3943 struct mdt_object *o, struct lustre_handle *h,
3944 enum ldlm_mode mode, int decref)
3948 if (lustre_handle_is_used(h)) {
3949 struct ldlm_lock *lock = ldlm_handle2lock(h);
3950 struct ptlrpc_request *req = mdt_info_req(info);
3953 (lock->l_policy_data.l_inodebits.bits &
3954 (MDS_INODELOCK_XATTR | MDS_INODELOCK_UPDATE)))
3955 mo_invalidate(info->mti_env, mdt_object_child(o));
3957 if (decref || !info->mti_has_trans || !req ||
3958 !(mode & (LCK_PW | LCK_EX))) {
3959 ldlm_lock_decref_and_cancel(h, mode);
3960 LDLM_LOCK_PUT(lock);
3962 tgt_save_slc_lock(&info->mti_mdt->mdt_lut, lock,
3964 ldlm_lock_decref(h, mode);
3973 * Unlock mdt object.
3975 * Immeditely release the regular lock and the PDO lock or save the
3976 * lock in request and keep them referenced until client ACK or
3977 * transaction commit.
3979 * \param info thread info object
3980 * \param o mdt object
3981 * \param lh mdt lock handle referencing regular and PDO locks
3982 * \param decref force immediate lock releasing
3984 * XXX o is not used and may be NULL, see hsm_cdt_request_completed().
3986 void mdt_object_unlock(struct mdt_thread_info *info, struct mdt_object *o,
3987 struct mdt_lock_handle *lh, int decref)
3991 mdt_save_lock(info, &lh->mlh_pdo_lh, lh->mlh_pdo_mode, decref);
3992 mdt_save_lock(info, &lh->mlh_reg_lh, lh->mlh_reg_mode, decref);
3993 mdt_save_remote_lock(info, o, &lh->mlh_rreg_lh, lh->mlh_rreg_mode,
3999 struct mdt_object *mdt_object_find_lock(struct mdt_thread_info *info,
4000 const struct lu_fid *f,
4001 struct mdt_lock_handle *lh,
4004 struct mdt_object *o;
4006 o = mdt_object_find(info->mti_env, info->mti_mdt, f);
4010 rc = mdt_object_lock(info, o, lh, ibits);
4012 mdt_object_put(info->mti_env, o);
4019 void mdt_object_unlock_put(struct mdt_thread_info * info,
4020 struct mdt_object * o,
4021 struct mdt_lock_handle *lh,
4024 mdt_object_unlock(info, o, lh, decref);
4025 mdt_object_put(info->mti_env, o);
4029 * Generic code handling requests that have struct mdt_body passed in:
4031 * - extract mdt_body from request and save it in @info, if present;
4033 * - create lu_object, corresponding to the fid in mdt_body, and save it in
4036 * - if HAS_BODY flag is set for this request type check whether object
4037 * actually exists on storage (lu_object_exists()).
4040 static int mdt_body_unpack(struct mdt_thread_info *info,
4041 enum tgt_handler_flags flags)
4043 const struct mdt_body *body;
4044 struct mdt_object *obj;
4045 const struct lu_env *env;
4046 struct req_capsule *pill;
4050 env = info->mti_env;
4051 pill = info->mti_pill;
4053 body = info->mti_body = req_capsule_client_get(pill, &RMF_MDT_BODY);
4057 if (!(body->mbo_valid & OBD_MD_FLID))
4060 if (!fid_is_sane(&body->mbo_fid1)) {
4061 CERROR("Invalid fid: "DFID"\n", PFID(&body->mbo_fid1));
4065 obj = mdt_object_find(env, info->mti_mdt, &body->mbo_fid1);
4067 if ((flags & HAS_BODY) && !mdt_object_exists(obj)) {
4068 mdt_object_put(env, obj);
4071 info->mti_object = obj;
4080 static int mdt_unpack_req_pack_rep(struct mdt_thread_info *info,
4081 enum tgt_handler_flags flags)
4083 struct req_capsule *pill = info->mti_pill;
4088 if (req_capsule_has_field(pill, &RMF_MDT_BODY, RCL_CLIENT))
4089 rc = mdt_body_unpack(info, flags);
4093 if (rc == 0 && (flags & HAS_REPLY)) {
4095 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
4096 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER,
4098 if (req_capsule_has_field(pill, &RMF_LOGCOOKIES, RCL_SERVER))
4099 req_capsule_set_size(pill, &RMF_LOGCOOKIES,
4102 /* Set ACL reply buffer size as LUSTRE_POSIX_ACL_MAX_SIZE_OLD
4103 * by default. If the target object has more ACL entries, then
4104 * enlarge the buffer when necessary. */
4105 if (req_capsule_has_field(pill, &RMF_ACL, RCL_SERVER))
4106 req_capsule_set_size(pill, &RMF_ACL, RCL_SERVER,
4107 LUSTRE_POSIX_ACL_MAX_SIZE_OLD);
4109 mdt_preset_secctx_size(info);
4110 mdt_preset_encctx_size(info);
4112 rc = req_capsule_server_pack(pill);
4114 CWARN("%s: cannot pack response: rc = %d\n",
4115 mdt_obd_name(info->mti_mdt), rc);
4120 void mdt_lock_handle_init(struct mdt_lock_handle *lh)
4122 lh->mlh_type = MDT_NUL_LOCK;
4123 lh->mlh_reg_lh.cookie = 0ull;
4124 lh->mlh_reg_mode = LCK_MINMODE;
4125 lh->mlh_pdo_lh.cookie = 0ull;
4126 lh->mlh_pdo_mode = LCK_MINMODE;
4127 lh->mlh_rreg_lh.cookie = 0ull;
4128 lh->mlh_rreg_mode = LCK_MINMODE;
4131 void mdt_lock_handle_fini(struct mdt_lock_handle *lh)
4133 LASSERT(!lustre_handle_is_used(&lh->mlh_reg_lh));
4134 LASSERT(!lustre_handle_is_used(&lh->mlh_pdo_lh));
4138 * Initialize fields of struct mdt_thread_info. Other fields are left in
4139 * uninitialized state, because it's too expensive to zero out whole
4140 * mdt_thread_info (> 1K) on each request arrival.
4142 void mdt_thread_info_init(struct ptlrpc_request *req,
4143 struct mdt_thread_info *info)
4147 info->mti_pill = &req->rq_pill;
4150 for (i = 0; i < ARRAY_SIZE(info->mti_lh); i++)
4151 mdt_lock_handle_init(&info->mti_lh[i]);
4153 /* mdt device: it can be NULL while CONNECT */
4154 if (req->rq_export) {
4155 info->mti_mdt = mdt_dev(req->rq_export->exp_obd->obd_lu_dev);
4156 info->mti_exp = req->rq_export;
4158 info->mti_mdt = NULL;
4159 info->mti_env = req->rq_svc_thread->t_env;
4160 info->mti_transno = lustre_msg_get_transno(req->rq_reqmsg);
4162 memset(&info->mti_attr, 0, sizeof(info->mti_attr));
4163 info->mti_big_buf = LU_BUF_NULL;
4164 info->mti_body = NULL;
4165 info->mti_object = NULL;
4166 info->mti_dlm_req = NULL;
4167 info->mti_has_trans = 0;
4168 info->mti_cross_ref = 0;
4169 info->mti_opdata = 0;
4170 info->mti_big_lmm_used = 0;
4171 info->mti_big_acl_used = 0;
4172 info->mti_som_valid = 0;
4174 info->mti_spec.no_create = 0;
4175 info->mti_spec.sp_rm_entry = 0;
4176 info->mti_spec.sp_permitted = 0;
4178 info->mti_spec.u.sp_ea.eadata = NULL;
4179 info->mti_spec.u.sp_ea.eadatalen = 0;
4182 void mdt_thread_info_fini(struct mdt_thread_info *info)
4186 if (info->mti_object != NULL) {
4187 mdt_object_put(info->mti_env, info->mti_object);
4188 info->mti_object = NULL;
4191 for (i = 0; i < ARRAY_SIZE(info->mti_lh); i++)
4192 mdt_lock_handle_fini(&info->mti_lh[i]);
4193 info->mti_env = NULL;
4194 info->mti_pill = NULL;
4195 info->mti_exp = NULL;
4196 info->mti_mdt = NULL;
4198 if (unlikely(info->mti_big_buf.lb_buf != NULL))
4199 lu_buf_free(&info->mti_big_buf);
4202 struct mdt_thread_info *tsi2mdt_info(struct tgt_session_info *tsi)
4204 struct mdt_thread_info *mti;
4206 mti = mdt_th_info(tsi->tsi_env);
4207 LASSERT(mti != NULL);
4209 mdt_thread_info_init(tgt_ses_req(tsi), mti);
4210 if (tsi->tsi_corpus != NULL) {
4211 mti->mti_object = mdt_obj(tsi->tsi_corpus);
4212 lu_object_get(tsi->tsi_corpus);
4214 mti->mti_body = tsi->tsi_mdt_body;
4215 mti->mti_dlm_req = tsi->tsi_dlm_req;
4220 static int mdt_tgt_connect(struct tgt_session_info *tsi)
4222 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_DELAY_CONDITIONAL) &&
4224 tsi2mdt_info(tsi)->mti_mdt->mdt_seq_site.ss_node_id)
4225 schedule_timeout_uninterruptible(cfs_time_seconds(3));
4227 return tgt_connect(tsi);
4230 static int mdt_intent_glimpse(enum ldlm_intent_flags it_opc,
4231 struct mdt_thread_info *info,
4232 struct ldlm_lock **lockp, __u64 flags)
4234 return mdt_glimpse_enqueue(info, info->mti_mdt->mdt_namespace,
4237 static int mdt_intent_brw(enum ldlm_intent_flags it_opc,
4238 struct mdt_thread_info *info,
4239 struct ldlm_lock **lockp, __u64 flags)
4241 return mdt_brw_enqueue(info, info->mti_mdt->mdt_namespace,
4245 int mdt_intent_lock_replace(struct mdt_thread_info *info,
4246 struct ldlm_lock **lockp,
4247 struct mdt_lock_handle *lh,
4248 __u64 flags, int result)
4250 struct ptlrpc_request *req = mdt_info_req(info);
4251 struct ldlm_lock *lock = *lockp;
4252 struct ldlm_lock *new_lock;
4254 /* If possible resent found a lock, @lh is set to its handle */
4255 new_lock = ldlm_handle2lock_long(&lh->mlh_reg_lh, 0);
4257 if (new_lock == NULL) {
4258 if (flags & LDLM_FL_INTENT_ONLY) {
4260 } else if (flags & LDLM_FL_RESENT) {
4261 /* Lock is pinned by ldlm_handle_enqueue0() as it is a
4262 * resend case, however, it could be already destroyed
4263 * due to client eviction or a raced cancel RPC.
4265 LDLM_DEBUG_NOLOCK("Invalid lock handle %#llx\n",
4266 lh->mlh_reg_lh.cookie);
4269 CERROR("%s: Invalid lockh=%#llx flags=%#llx fid1="DFID" fid2="DFID": rc = %d\n",
4270 mdt_obd_name(info->mti_mdt),
4271 lh->mlh_reg_lh.cookie, flags,
4272 PFID(&info->mti_tmp_fid1),
4273 PFID(&info->mti_tmp_fid2), result);
4276 lh->mlh_reg_lh.cookie = 0;
4281 * If we've already given this lock to a client once, then we should
4282 * have no readers or writers. Otherwise, we should have one reader
4283 * _or_ writer ref (which will be zeroed below) before returning the
4286 if (new_lock->l_export == req->rq_export) {
4287 LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
4289 LASSERT(new_lock->l_export == NULL);
4290 LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
4295 if (new_lock->l_export == req->rq_export) {
4297 * Already gave this to the client, which means that we
4298 * reconstructed a reply.
4300 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
4303 LDLM_LOCK_RELEASE(new_lock);
4304 lh->mlh_reg_lh.cookie = 0;
4305 RETURN(ELDLM_LOCK_REPLACED);
4309 * Fixup the lock to be given to the client.
4311 lock_res_and_lock(new_lock);
4312 /* Zero new_lock->l_readers and new_lock->l_writers without triggering
4313 * possible blocking AST. */
4314 while (new_lock->l_readers > 0) {
4315 lu_ref_del(&new_lock->l_reference, "reader", new_lock);
4316 lu_ref_del(&new_lock->l_reference, "user", new_lock);
4317 new_lock->l_readers--;
4319 while (new_lock->l_writers > 0) {
4320 lu_ref_del(&new_lock->l_reference, "writer", new_lock);
4321 lu_ref_del(&new_lock->l_reference, "user", new_lock);
4322 new_lock->l_writers--;
4325 new_lock->l_export = class_export_lock_get(req->rq_export, new_lock);
4326 new_lock->l_blocking_ast = lock->l_blocking_ast;
4327 new_lock->l_completion_ast = lock->l_completion_ast;
4328 if (ldlm_has_dom(new_lock))
4329 new_lock->l_glimpse_ast = ldlm_server_glimpse_ast;
4330 new_lock->l_remote_handle = lock->l_remote_handle;
4331 new_lock->l_flags &= ~LDLM_FL_LOCAL;
4333 unlock_res_and_lock(new_lock);
4335 cfs_hash_add(new_lock->l_export->exp_lock_hash,
4336 &new_lock->l_remote_handle,
4337 &new_lock->l_exp_hash);
4339 LDLM_LOCK_RELEASE(new_lock);
4340 lh->mlh_reg_lh.cookie = 0;
4342 RETURN(ELDLM_LOCK_REPLACED);
4345 void mdt_intent_fixup_resent(struct mdt_thread_info *info,
4346 struct ldlm_lock *new_lock,
4347 struct mdt_lock_handle *lh, __u64 flags)
4349 struct ptlrpc_request *req = mdt_info_req(info);
4350 struct ldlm_request *dlmreq;
4352 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
4355 dlmreq = req_capsule_client_get(info->mti_pill, &RMF_DLM_REQ);
4357 /* Check if this is a resend case (MSG_RESENT is set on RPC) and a
4358 * lock was found by ldlm_handle_enqueue(); if so @lh must be
4360 if (flags & LDLM_FL_RESENT) {
4361 lh->mlh_reg_lh.cookie = new_lock->l_handle.h_cookie;
4362 lh->mlh_reg_mode = new_lock->l_granted_mode;
4364 LDLM_DEBUG(new_lock, "Restoring lock cookie");
4365 DEBUG_REQ(D_DLMTRACE, req, "restoring lock cookie %#llx",
4366 lh->mlh_reg_lh.cookie);
4371 * If the xid matches, then we know this is a resent request, and allow
4372 * it. (It's probably an OPEN, for which we don't send a lock.
4374 if (req_can_reconstruct(req, NULL) != 0)
4378 * This remote handle isn't enqueued, so we never received or processed
4379 * this request. Clear MSG_RESENT, because it can be handled like any
4380 * normal request now.
4382 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
4384 DEBUG_REQ(D_DLMTRACE, req, "no existing lock with rhandle %#llx",
4385 dlmreq->lock_handle[0].cookie);
4388 static int mdt_intent_getxattr(enum ldlm_intent_flags it_opc,
4389 struct mdt_thread_info *info,
4390 struct ldlm_lock **lockp,
4393 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
4394 struct ldlm_reply *ldlm_rep = NULL;
4399 * Initialize lhc->mlh_reg_lh either from a previously granted lock
4400 * (for the resend case) or a new lock. Below we will use it to
4401 * replace the original lock.
4403 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
4404 if (!lustre_handle_is_used(&lhc->mlh_reg_lh)) {
4405 mdt_lock_reg_init(lhc, (*lockp)->l_req_mode);
4406 rc = mdt_object_lock(info, info->mti_object, lhc,
4407 MDS_INODELOCK_XATTR);
4412 rc = mdt_getxattr(info);
4414 if (mdt_info_req(info)->rq_repmsg != NULL)
4415 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
4417 if (ldlm_rep == NULL ||
4418 OBD_FAIL_CHECK(OBD_FAIL_MDS_XATTR_REP)) {
4419 mdt_object_unlock(info, info->mti_object, lhc, 1);
4423 RETURN(err_serious(-EFAULT));
4426 ldlm_rep->lock_policy_res2 = clear_serious(rc);
4428 /* This is left for interop instead of adding a new interop flag.
4430 #if LUSTRE_VERSION_CODE > OBD_OCD_VERSION(3, 0, 0, 0)
4431 if (ldlm_rep->lock_policy_res2) {
4432 mdt_object_unlock(info, info->mti_object, lhc, 1);
4433 RETURN(ELDLM_LOCK_ABORTED);
4437 rc = mdt_intent_lock_replace(info, lockp, lhc, flags, rc);
4441 static int mdt_intent_getattr(enum ldlm_intent_flags it_opc,
4442 struct mdt_thread_info *info,
4443 struct ldlm_lock **lockp,
4446 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
4448 struct ldlm_reply *ldlm_rep;
4449 struct mdt_body *reqbody;
4450 struct mdt_body *repbody;
4454 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
4457 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
4460 info->mti_cross_ref = !!(reqbody->mbo_valid & OBD_MD_FLCROSSREF);
4461 repbody->mbo_eadatasize = 0;
4462 repbody->mbo_aclsize = 0;
4466 child_bits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
4469 child_bits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE |
4473 CERROR("%s: unsupported intent %#x\n",
4474 mdt_obd_name(info->mti_mdt), (unsigned int)it_opc);
4475 GOTO(out_shrink, rc = -EINVAL);
4478 rc = mdt_init_ucred(info, reqbody);
4480 GOTO(out_shrink, rc);
4482 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
4483 mdt_set_disposition(info, ldlm_rep, DISP_IT_EXECD);
4485 /* Get lock from request for possible resent case. */
4486 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
4488 rc = mdt_getattr_name_lock(info, lhc, child_bits, ldlm_rep);
4489 ldlm_rep->lock_policy_res2 = clear_serious(rc);
4491 if (mdt_get_disposition(ldlm_rep, DISP_LOOKUP_NEG))
4492 ldlm_rep->lock_policy_res2 = 0;
4493 if (!mdt_get_disposition(ldlm_rep, DISP_LOOKUP_POS) ||
4494 ldlm_rep->lock_policy_res2) {
4495 lhc->mlh_reg_lh.cookie = 0ull;
4496 GOTO(out_ucred, rc = ELDLM_LOCK_ABORTED);
4499 rc = mdt_intent_lock_replace(info, lockp, lhc, flags, rc);
4502 mdt_exit_ucred(info);
4504 mdt_client_compatibility(info);
4505 rc2 = mdt_fix_reply(info);
4511 static int mdt_intent_layout(enum ldlm_intent_flags it_opc,
4512 struct mdt_thread_info *info,
4513 struct ldlm_lock **lockp,
4516 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
4517 struct md_layout_change layout = { .mlc_opc = MD_LAYOUT_NOP };
4518 struct layout_intent *intent;
4519 struct ldlm_reply *ldlm_rep;
4520 struct lu_fid *fid = &info->mti_tmp_fid2;
4521 struct mdt_object *obj = NULL;
4522 int layout_size = 0;
4523 struct lu_buf *buf = &layout.mlc_buf;
4528 fid_extract_from_res_name(fid, &(*lockp)->l_resource->lr_name);
4530 intent = req_capsule_client_get(info->mti_pill, &RMF_LAYOUT_INTENT);
4534 CDEBUG(D_INFO, DFID "got layout change request from client: "
4535 "opc:%u flags:%#x extent "DEXT"\n",
4536 PFID(fid), intent->li_opc, intent->li_flags,
4537 PEXT(&intent->li_extent));
4539 switch (intent->li_opc) {
4540 case LAYOUT_INTENT_TRUNC:
4541 case LAYOUT_INTENT_WRITE:
4542 layout.mlc_opc = MD_LAYOUT_WRITE;
4543 layout.mlc_intent = intent;
4545 case LAYOUT_INTENT_ACCESS:
4547 case LAYOUT_INTENT_READ:
4548 case LAYOUT_INTENT_GLIMPSE:
4549 case LAYOUT_INTENT_RELEASE:
4550 case LAYOUT_INTENT_RESTORE:
4551 CERROR("%s: Unsupported layout intent opc %d\n",
4552 mdt_obd_name(info->mti_mdt), intent->li_opc);
4555 CERROR("%s: Unknown layout intent opc %d\n",
4556 mdt_obd_name(info->mti_mdt), intent->li_opc);
4560 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
4562 RETURN(PTR_ERR(obj));
4564 if (mdt_object_exists(obj) && !mdt_object_remote(obj)) {
4565 /* if layout is going to be changed don't use the current EA
4566 * size but the maximum one. That buffer will be shrinked
4567 * to the actual size in req_capsule_shrink() before reply.
4569 if (layout.mlc_opc == MD_LAYOUT_WRITE) {
4570 layout_size = info->mti_mdt->mdt_max_mdsize;
4572 layout_size = mdt_attr_get_eabuf_size(info, obj);
4573 if (layout_size < 0)
4574 GOTO(out, rc = layout_size);
4576 if (layout_size > info->mti_mdt->mdt_max_mdsize)
4577 info->mti_mdt->mdt_max_mdsize = layout_size;
4579 CDEBUG(D_INFO, "%s: layout_size %d\n",
4580 mdt_obd_name(info->mti_mdt), layout_size);
4584 * set reply buffer size, so that ldlm_handle_enqueue0()->
4585 * ldlm_lvbo_fill() will fill the reply buffer with lovea.
4587 req_capsule_set_size(info->mti_pill, &RMF_DLM_LVB, RCL_SERVER,
4589 rc = req_capsule_server_pack(info->mti_pill);
4593 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
4595 GOTO(out, rc = -EPROTO);
4597 mdt_set_disposition(info, ldlm_rep, DISP_IT_EXECD);
4599 /* take lock in ldlm_lock_enqueue() for LAYOUT_INTENT_ACCESS */
4600 if (layout.mlc_opc == MD_LAYOUT_NOP)
4603 rc = mdt_check_resent(info, mdt_reconstruct_generic, lhc);
4607 DEBUG_REQ(D_INODE, mdt_info_req(info), "resent opt.");
4608 rc = lustre_msg_get_status(mdt_info_req(info)->rq_repmsg);
4614 if (unlikely(req_is_replay(mdt_info_req(info)))) {
4615 buf->lb_buf = req_capsule_client_get(info->mti_pill,
4617 buf->lb_len = req_capsule_get_size(info->mti_pill,
4618 &RMF_EADATA, RCL_CLIENT);
4620 * If it's a replay of layout write intent RPC, the client has
4621 * saved the extended lovea when it get reply then.
4623 if (buf->lb_len > 0)
4624 mdt_fix_lov_magic(info, buf->lb_buf);
4627 /* Get lock from request for possible resent case. */
4628 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
4629 (*lockp)->l_lvb_type = LVB_T_LAYOUT;
4632 * Instantiate some layout components, if @buf contains lovea, then it's
4633 * a replay of the layout intent write RPC.
4635 rc = mdt_layout_change(info, obj, lhc, &layout);
4636 ldlm_rep->lock_policy_res2 = clear_serious(rc);
4638 if (lustre_handle_is_used(&lhc->mlh_reg_lh)) {
4639 rc = mdt_intent_lock_replace(info, lockp, lhc, flags, rc);
4640 if (rc == ELDLM_LOCK_REPLACED &&
4641 (*lockp)->l_granted_mode == LCK_EX)
4642 ldlm_lock_mode_downgrade(*lockp, LCK_CR);
4647 mdt_object_put(info->mti_env, obj);
4651 static int mdt_intent_open(enum ldlm_intent_flags it_opc,
4652 struct mdt_thread_info *info,
4653 struct ldlm_lock **lockp,
4656 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
4657 struct ldlm_reply *rep = NULL;
4660 struct ptlrpc_request *req = mdt_info_req(info);
4662 static const struct req_format *intent_fmts[REINT_MAX] = {
4663 [REINT_CREATE] = &RQF_LDLM_INTENT_CREATE,
4664 [REINT_OPEN] = &RQF_LDLM_INTENT_OPEN
4669 opc = mdt_reint_opcode(mdt_info_req(info), intent_fmts);
4673 /* Get lock from request for possible resent case. */
4674 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
4676 rc = mdt_reint_internal(info, lhc, opc);
4678 if (rc < 0 && lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
4679 DEBUG_REQ(D_ERROR, req, "Replay open failed with %d", rc);
4681 /* Check whether the reply has been packed successfully. */
4682 if (mdt_info_req(info)->rq_repmsg != NULL)
4683 rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
4688 RETURN(err_serious(-EFAULT));
4691 /* MDC expects this in any case */
4693 mdt_set_disposition(info, rep, DISP_LOOKUP_EXECD);
4695 /* the open lock or the lock for cross-ref object should be
4696 * returned to the client */
4697 if (lustre_handle_is_used(&lhc->mlh_reg_lh) &&
4698 (rc == 0 || rc == -MDT_EREMOTE_OPEN)) {
4699 rep->lock_policy_res2 = 0;
4700 rc = mdt_intent_lock_replace(info, lockp, lhc, flags, rc);
4704 rep->lock_policy_res2 = clear_serious(rc);
4706 if (rep->lock_policy_res2 == -ENOENT &&
4707 mdt_get_disposition(rep, DISP_LOOKUP_NEG) &&
4708 !mdt_get_disposition(rep, DISP_OPEN_CREATE))
4709 rep->lock_policy_res2 = 0;
4711 lhc->mlh_reg_lh.cookie = 0ull;
4712 if (rc == -ENOTCONN || rc == -ENODEV ||
4713 rc == -EOVERFLOW) { /**< if VBR failure then return error */
4715 * If it is the disconnect error (ENODEV & ENOCONN), the error
4716 * will be returned by rq_status, and client at ptlrpc layer
4717 * will detect this, then disconnect, reconnect the import
4718 * immediately, instead of impacting the following the rpc.
4723 * For other cases, the error will be returned by intent, and client
4724 * will retrieve the result from intent.
4726 RETURN(ELDLM_LOCK_ABORTED);
4729 static int mdt_intent_opc(enum ldlm_intent_flags it_opc,
4730 struct mdt_thread_info *info,
4731 struct ldlm_lock **lockp,
4732 u64 flags /* LDLM_FL_* */)
4734 struct req_capsule *pill = info->mti_pill;
4735 struct ptlrpc_request *req = mdt_info_req(info);
4736 const struct req_format *it_format;
4737 int (*it_handler)(enum ldlm_intent_flags,
4738 struct mdt_thread_info *,
4739 struct ldlm_lock **,
4741 enum tgt_handler_flags it_handler_flags = 0;
4742 struct ldlm_reply *rep;
4743 bool check_mdt_object = false;
4749 case IT_OPEN|IT_CREAT:
4751 * OCREAT is not a IS_MUTABLE request since the file may
4752 * already exist. We do the extra check of
4753 * OBD_CONNECT_RDONLY in mdt_reint_open() when we
4754 * really need to create the object.
4756 it_format = &RQF_LDLM_INTENT;
4757 it_handler = &mdt_intent_open;
4760 check_mdt_object = true;
4763 it_format = &RQF_LDLM_INTENT_GETATTR;
4764 it_handler = &mdt_intent_getattr;
4765 it_handler_flags = HAS_REPLY;
4768 check_mdt_object = true;
4769 it_format = &RQF_LDLM_INTENT_GETXATTR;
4770 it_handler = &mdt_intent_getxattr;
4771 it_handler_flags = HAS_BODY;
4774 it_format = &RQF_LDLM_INTENT_LAYOUT;
4775 it_handler = &mdt_intent_layout;
4778 it_format = &RQF_LDLM_INTENT;
4779 it_handler = &mdt_intent_glimpse;
4782 it_format = &RQF_LDLM_INTENT;
4783 it_handler = &mdt_intent_brw;
4785 case IT_QUOTA_DQACQ:
4786 case IT_QUOTA_CONN: {
4787 struct lu_device *qmt = info->mti_mdt->mdt_qmt_dev;
4790 RETURN(-EOPNOTSUPP);
4792 if (mdt_rdonly(req->rq_export))
4795 (*lockp)->l_lvb_type = LVB_T_LQUOTA;
4796 /* pass the request to quota master */
4797 rc = qmt_hdls.qmth_intent_policy(info->mti_env, qmt,
4798 mdt_info_req(info), lockp,
4803 CERROR("%s: unknown intent code %#x\n",
4804 mdt_obd_name(info->mti_mdt), it_opc);
4808 req_capsule_extend(pill, it_format);
4810 rc = mdt_unpack_req_pack_rep(info, it_handler_flags);
4814 if (unlikely(info->mti_object == NULL && check_mdt_object))
4817 if (it_handler_flags & IS_MUTABLE && mdt_rdonly(req->rq_export))
4820 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_INTENT_DELAY, 10);
4822 /* execute policy */
4823 rc = (*it_handler)(it_opc, info, lockp, flags);
4825 /* Check whether the reply has been packed successfully. */
4826 if (req->rq_repmsg != NULL) {
4827 rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
4828 rep->lock_policy_res2 =
4829 ptlrpc_status_hton(rep->lock_policy_res2);
4835 static void mdt_ptlrpc_stats_update(struct ptlrpc_request *req,
4836 enum ldlm_intent_flags it_opc)
4838 struct lprocfs_stats *srv_stats = ptlrpc_req2svc(req)->srv_stats;
4840 /* update stats when IT code is known */
4841 if (srv_stats != NULL)
4842 lprocfs_counter_incr(srv_stats,
4843 PTLRPC_LAST_CNTR + (it_opc == IT_GLIMPSE ?
4844 LDLM_GLIMPSE_ENQUEUE : LDLM_IBITS_ENQUEUE));
4847 static int mdt_intent_policy(const struct lu_env *env,
4848 struct ldlm_namespace *ns,
4849 struct ldlm_lock **lockp,
4851 enum ldlm_mode mode,
4852 __u64 flags, void *data)
4854 struct tgt_session_info *tsi;
4855 struct mdt_thread_info *info;
4856 struct ptlrpc_request *req = req_cookie;
4857 struct ldlm_intent *it;
4858 struct req_capsule *pill;
4859 const struct ldlm_lock_desc *ldesc;
4864 LASSERT(req != NULL);
4866 tsi = tgt_ses_info(env);
4868 info = tsi2mdt_info(tsi);
4869 LASSERT(info != NULL);
4870 pill = info->mti_pill;
4871 LASSERT(pill->rc_req == req);
4872 ldesc = &info->mti_dlm_req->lock_desc;
4874 if (req->rq_reqmsg->lm_bufcount > DLM_INTENT_IT_OFF) {
4875 req_capsule_extend(pill, &RQF_LDLM_INTENT_BASIC);
4876 it = req_capsule_client_get(pill, &RMF_LDLM_INTENT);
4878 mdt_ptlrpc_stats_update(req, it->opc);
4879 rc = mdt_intent_opc(it->opc, info, lockp, flags);
4883 /* Lock without inodebits makes no sense and will oops
4884 * later in ldlm. Let's check it now to see if we have
4885 * ibits corrupted somewhere in mdt_intent_opc().
4886 * The case for client miss to set ibits has been
4887 * processed by others. */
4888 LASSERT(ergo(ldesc->l_resource.lr_type == LDLM_IBITS,
4889 ldesc->l_policy_data.l_inodebits.bits != 0));
4891 rc = err_serious(-EFAULT);
4893 } else if (ldesc->l_resource.lr_type == LDLM_IBITS &&
4894 ldesc->l_policy_data.l_inodebits.bits == MDS_INODELOCK_DOM) {
4895 struct ldlm_reply *rep;
4897 /* No intent was provided but INTENT flag is set along with
4898 * DOM bit, this is considered as GLIMPSE request.
4899 * This logic is common for MDT and OST glimpse
4901 mdt_ptlrpc_stats_update(req, IT_GLIMPSE);
4902 rc = mdt_glimpse_enqueue(info, ns, lockp, flags);
4903 /* Check whether the reply has been packed successfully. */
4904 if (req->rq_repmsg != NULL) {
4905 rep = req_capsule_server_get(info->mti_pill,
4907 rep->lock_policy_res2 =
4908 ptlrpc_status_hton(rep->lock_policy_res2);
4911 /* No intent was provided */
4912 req_capsule_set_size(pill, &RMF_DLM_LVB, RCL_SERVER, 0);
4913 rc = req_capsule_server_pack(pill);
4915 rc = err_serious(rc);
4917 mdt_thread_info_fini(info);
4921 static void mdt_deregister_seq_exp(struct mdt_device *mdt)
4923 struct seq_server_site *ss = mdt_seq_site(mdt);
4925 if (ss->ss_node_id == 0)
4928 if (ss->ss_client_seq != NULL) {
4929 lustre_deregister_lwp_item(&ss->ss_client_seq->lcs_exp);
4930 ss->ss_client_seq->lcs_exp = NULL;
4933 if (ss->ss_server_fld != NULL) {
4934 lustre_deregister_lwp_item(&ss->ss_server_fld->lsf_control_exp);
4935 ss->ss_server_fld->lsf_control_exp = NULL;
4939 static void mdt_seq_fini_cli(struct mdt_device *mdt)
4941 struct seq_server_site *ss = mdt_seq_site(mdt);
4946 if (ss->ss_server_seq != NULL)
4947 seq_server_set_cli(NULL, ss->ss_server_seq, NULL);
4950 static int mdt_seq_fini(const struct lu_env *env, struct mdt_device *mdt)
4952 mdt_seq_fini_cli(mdt);
4953 mdt_deregister_seq_exp(mdt);
4955 return seq_site_fini(env, mdt_seq_site(mdt));
4959 * It will retrieve its FLDB entries from MDT0, and it only happens
4960 * when upgrading existent FS to 2.6 or when local FLDB is corrupted,
4961 * and it needs to refresh FLDB from the MDT0.
4963 static int mdt_register_lwp_callback(void *data)
4966 struct mdt_device *mdt = data;
4967 struct lu_server_fld *fld = mdt_seq_site(mdt)->ss_server_fld;
4971 LASSERT(mdt_seq_site(mdt)->ss_node_id != 0);
4973 rc = lu_env_init(&env, LCT_MD_THREAD);
4975 CERROR("%s: cannot init env: rc = %d\n", mdt_obd_name(mdt), rc);
4979 /* Allocate new sequence now to avoid creating local transaction
4980 * in the normal transaction process */
4981 rc = seq_server_check_and_alloc_super(&env,
4982 mdt_seq_site(mdt)->ss_server_seq);
4987 rc = fld_update_from_controller(&env, fld);
4989 CERROR("%s: cannot update controller: rc = %d\n",
4990 mdt_obd_name(mdt), rc);
4999 static int mdt_register_seq_exp(struct mdt_device *mdt)
5001 struct seq_server_site *ss = mdt_seq_site(mdt);
5002 char *lwp_name = NULL;
5005 if (ss->ss_node_id == 0)
5008 OBD_ALLOC(lwp_name, MAX_OBD_NAME);
5009 if (lwp_name == NULL)
5010 GOTO(out_free, rc = -ENOMEM);
5012 rc = tgt_name2lwp_name(mdt_obd_name(mdt), lwp_name, MAX_OBD_NAME, 0);
5016 rc = lustre_register_lwp_item(lwp_name, &ss->ss_client_seq->lcs_exp,
5021 rc = lustre_register_lwp_item(lwp_name,
5022 &ss->ss_server_fld->lsf_control_exp,
5023 mdt_register_lwp_callback, mdt);
5025 lustre_deregister_lwp_item(&ss->ss_client_seq->lcs_exp);
5026 ss->ss_client_seq->lcs_exp = NULL;
5030 if (lwp_name != NULL)
5031 OBD_FREE(lwp_name, MAX_OBD_NAME);
5037 * Init client sequence manager which is used by local MDS to talk to sequence
5038 * controller on remote node.
5040 static int mdt_seq_init_cli(const struct lu_env *env, struct mdt_device *mdt)
5042 struct seq_server_site *ss = mdt_seq_site(mdt);
5046 /* check if this is adding the first MDC and controller is not yet
5048 OBD_ALLOC_PTR(ss->ss_client_seq);
5049 if (ss->ss_client_seq == NULL)
5052 OBD_ALLOC(prefix, MAX_OBD_NAME + 5);
5053 if (prefix == NULL) {
5054 OBD_FREE_PTR(ss->ss_client_seq);
5055 ss->ss_client_seq = NULL;
5059 /* Note: seq_client_fini will be called in seq_site_fini */
5060 snprintf(prefix, MAX_OBD_NAME + 5, "ctl-%s", mdt_obd_name(mdt));
5061 seq_client_init(ss->ss_client_seq, NULL, LUSTRE_SEQ_METADATA,
5062 prefix, ss->ss_node_id == 0 ? ss->ss_control_seq :
5064 OBD_FREE(prefix, MAX_OBD_NAME + 5);
5066 RETURN(seq_server_set_cli(env, ss->ss_server_seq, ss->ss_client_seq));
5069 static int mdt_seq_init(const struct lu_env *env, struct mdt_device *mdt)
5071 struct seq_server_site *ss;
5075 ss = mdt_seq_site(mdt);
5076 /* init sequence controller server(MDT0) */
5077 if (ss->ss_node_id == 0) {
5078 OBD_ALLOC_PTR(ss->ss_control_seq);
5079 if (ss->ss_control_seq == NULL)
5082 rc = seq_server_init(env, ss->ss_control_seq, mdt->mdt_bottom,
5083 mdt_obd_name(mdt), LUSTRE_SEQ_CONTROLLER,
5086 GOTO(out_seq_fini, rc);
5089 /* Init normal sequence server */
5090 OBD_ALLOC_PTR(ss->ss_server_seq);
5091 if (ss->ss_server_seq == NULL)
5092 GOTO(out_seq_fini, rc = -ENOMEM);
5094 rc = seq_server_init(env, ss->ss_server_seq, mdt->mdt_bottom,
5095 mdt_obd_name(mdt), LUSTRE_SEQ_SERVER, ss);
5097 GOTO(out_seq_fini, rc);
5099 /* init seq client for seq server to talk to seq controller(MDT0) */
5100 rc = mdt_seq_init_cli(env, mdt);
5102 GOTO(out_seq_fini, rc);
5104 if (ss->ss_node_id != 0)
5105 /* register controller export through lwp */
5106 rc = mdt_register_seq_exp(mdt);
5111 mdt_seq_fini(env, mdt);
5119 static int mdt_fld_fini(const struct lu_env *env,
5120 struct mdt_device *m)
5122 struct seq_server_site *ss = mdt_seq_site(m);
5125 if (ss && ss->ss_server_fld) {
5126 fld_server_fini(env, ss->ss_server_fld);
5127 OBD_FREE_PTR(ss->ss_server_fld);
5128 ss->ss_server_fld = NULL;
5134 static int mdt_fld_init(const struct lu_env *env,
5136 struct mdt_device *m)
5138 struct seq_server_site *ss;
5142 ss = mdt_seq_site(m);
5144 OBD_ALLOC_PTR(ss->ss_server_fld);
5145 if (ss->ss_server_fld == NULL)
5146 RETURN(rc = -ENOMEM);
5148 rc = fld_server_init(env, ss->ss_server_fld, m->mdt_bottom, uuid,
5151 OBD_FREE_PTR(ss->ss_server_fld);
5152 ss->ss_server_fld = NULL;
5159 static void mdt_stack_pre_fini(const struct lu_env *env,
5160 struct mdt_device *m, struct lu_device *top)
5162 struct lustre_cfg_bufs *bufs;
5163 struct lustre_cfg *lcfg;
5164 struct mdt_thread_info *info;
5169 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
5170 LASSERT(info != NULL);
5172 bufs = &info->mti_u.bufs;
5174 LASSERT(m->mdt_child_exp);
5175 LASSERT(m->mdt_child_exp->exp_obd);
5177 /* process cleanup, pass mdt obd name to get obd umount flags */
5178 /* XXX: this is needed because all layers are referenced by
5179 * objects (some of them are pinned by osd, for example *
5180 * the proper solution should be a model where object used
5181 * by osd only doesn't have mdt/mdd slices -bzzz */
5182 lustre_cfg_bufs_reset(bufs, mdt_obd_name(m));
5183 lustre_cfg_bufs_set_string(bufs, 1, NULL);
5184 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
5187 lustre_cfg_init(lcfg, LCFG_PRE_CLEANUP, bufs);
5189 top->ld_ops->ldo_process_config(env, top, lcfg);
5190 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens));
5194 static void mdt_stack_fini(const struct lu_env *env,
5195 struct mdt_device *m, struct lu_device *top)
5197 struct obd_device *obd = mdt2obd_dev(m);
5198 struct lustre_cfg_bufs *bufs;
5199 struct lustre_cfg *lcfg;
5200 struct mdt_thread_info *info;
5204 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
5205 LASSERT(info != NULL);
5207 lu_dev_del_linkage(top->ld_site, top);
5209 lu_site_purge(env, top->ld_site, -1);
5211 bufs = &info->mti_u.bufs;
5212 /* process cleanup, pass mdt obd name to get obd umount flags */
5213 /* another purpose is to let all layers to release their objects */
5214 lustre_cfg_bufs_reset(bufs, mdt_obd_name(m));
5219 lustre_cfg_bufs_set_string(bufs, 1, flags);
5220 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
5223 lustre_cfg_init(lcfg, LCFG_CLEANUP, bufs);
5226 top->ld_ops->ldo_process_config(env, top, lcfg);
5227 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens));
5229 lu_site_purge(env, top->ld_site, -1);
5231 m->mdt_child = NULL;
5232 m->mdt_bottom = NULL;
5234 obd_disconnect(m->mdt_child_exp);
5235 m->mdt_child_exp = NULL;
5237 obd_disconnect(m->mdt_bottom_exp);
5238 m->mdt_child_exp = NULL;
5241 static int mdt_connect_to_next(const struct lu_env *env, struct mdt_device *m,
5242 const char *next, struct obd_export **exp)
5244 struct obd_connect_data *data = NULL;
5245 struct obd_device *obd;
5249 OBD_ALLOC_PTR(data);
5251 GOTO(out, rc = -ENOMEM);
5253 obd = class_name2obd(next);
5255 CERROR("%s: can't locate next device: %s\n",
5256 mdt_obd_name(m), next);
5257 GOTO(out, rc = -ENOTCONN);
5260 data->ocd_connect_flags = OBD_CONNECT_VERSION;
5261 data->ocd_version = LUSTRE_VERSION_CODE;
5263 rc = obd_connect(NULL, exp, obd, &obd->obd_uuid, data, NULL);
5265 CERROR("%s: cannot connect to next dev %s (%d)\n",
5266 mdt_obd_name(m), next, rc);
5276 static int mdt_stack_init(const struct lu_env *env, struct mdt_device *mdt,
5277 struct lustre_cfg *cfg)
5279 char *dev = lustre_cfg_string(cfg, 0);
5280 int rc, name_size, uuid_size;
5281 char *name, *uuid, *p;
5282 struct lustre_cfg_bufs *bufs;
5283 struct lustre_cfg *lcfg;
5284 struct obd_device *obd;
5285 struct lustre_profile *lprof;
5286 struct lu_site *site;
5289 /* in 1.8 we had the only device in the stack - MDS.
5290 * 2.0 introduces MDT, MDD, OSD; MDT starts others internally.
5291 * in 2.3 OSD is instantiated by obd_mount.c, so we need
5292 * to generate names and setup MDT, MDD. MDT will be using
5293 * generated name to connect to MDD. for MDD the next device
5294 * will be LOD with name taken from so called "profile" which
5295 * is generated by mount_option line
5297 * 1.8 MGS generates config. commands like this:
5298 * #06 (104)mount_option 0: 1:lustre-MDT0000 2:lustre-mdtlov
5299 * #08 (120)setup 0:lustre-MDT0000 1:dev 2:type 3:lustre-MDT0000
5300 * 2.0 MGS generates config. commands like this:
5301 * #07 (112)mount_option 0: 1:lustre-MDT0000 2:lustre-MDT0000-mdtlov
5302 * #08 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
5303 * 3:lustre-MDT0000-mdtlov 4:f
5305 * we generate MDD name from MDT one, just replacing T with D
5307 * after all the preparations, the logical equivalent will be
5308 * #01 (160)setup 0:lustre-MDD0000 1:lustre-MDD0000_UUID 2:0
5309 * 3:lustre-MDT0000-mdtlov 4:f
5310 * #02 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
5311 * 3:lustre-MDD0000 4:f
5313 * notice we build the stack from down to top: MDD first, then MDT */
5315 name_size = MAX_OBD_NAME;
5316 uuid_size = MAX_OBD_NAME;
5318 OBD_ALLOC(name, name_size);
5319 OBD_ALLOC(uuid, uuid_size);
5320 if (name == NULL || uuid == NULL)
5321 GOTO(cleanup_mem, rc = -ENOMEM);
5323 OBD_ALLOC_PTR(bufs);
5325 GOTO(cleanup_mem, rc = -ENOMEM);
5328 p = strstr(name, "-MDT");
5330 GOTO(free_bufs, rc = -ENOMEM);
5333 snprintf(uuid, MAX_OBD_NAME, "%s_UUID", name);
5335 lprof = class_get_profile(lustre_cfg_string(cfg, 0));
5336 if (lprof == NULL || lprof->lp_dt == NULL) {
5337 CERROR("can't find the profile: %s\n",
5338 lustre_cfg_string(cfg, 0));
5339 GOTO(free_bufs, rc = -EINVAL);
5342 lustre_cfg_bufs_reset(bufs, name);
5343 lustre_cfg_bufs_set_string(bufs, 1, LUSTRE_MDD_NAME);
5344 lustre_cfg_bufs_set_string(bufs, 2, uuid);
5345 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
5347 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
5349 GOTO(put_profile, rc = -ENOMEM);
5350 lustre_cfg_init(lcfg, LCFG_ATTACH, bufs);
5352 rc = class_attach(lcfg);
5354 GOTO(lcfg_cleanup, rc);
5356 obd = class_name2obd(name);
5358 CERROR("Can not find obd %s (%s in config)\n",
5359 MDD_OBD_NAME, lustre_cfg_string(cfg, 0));
5360 GOTO(lcfg_cleanup, rc = -EINVAL);
5363 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens));
5365 lustre_cfg_bufs_reset(bufs, name);
5366 lustre_cfg_bufs_set_string(bufs, 1, uuid);
5367 lustre_cfg_bufs_set_string(bufs, 2, dev);
5368 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
5370 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
5372 GOTO(class_detach, rc = -ENOMEM);
5373 lustre_cfg_init(lcfg, LCFG_SETUP, bufs);
5375 rc = class_setup(obd, lcfg);
5377 GOTO(class_detach, rc);
5379 /* connect to MDD we just setup */
5380 rc = mdt_connect_to_next(env, mdt, name, &mdt->mdt_child_exp);
5382 GOTO(class_detach, rc);
5384 site = mdt->mdt_child_exp->exp_obd->obd_lu_dev->ld_site;
5386 LASSERT(mdt_lu_site(mdt) == NULL);
5387 mdt->mdt_lu_dev.ld_site = site;
5388 site->ls_top_dev = &mdt->mdt_lu_dev;
5389 mdt->mdt_child = lu2md_dev(mdt->mdt_child_exp->exp_obd->obd_lu_dev);
5391 /* now connect to bottom OSD */
5392 snprintf(name, MAX_OBD_NAME, "%s-osd", dev);
5393 rc = mdt_connect_to_next(env, mdt, name, &mdt->mdt_bottom_exp);
5395 GOTO(class_detach, rc);
5397 lu2dt_dev(mdt->mdt_bottom_exp->exp_obd->obd_lu_dev);
5399 rc = lu_env_refill((struct lu_env *)env);
5401 CERROR("Failure to refill session: '%d'\n", rc);
5403 lu_dev_add_linkage(site, &mdt->mdt_lu_dev);
5408 class_detach(obd, lcfg);
5410 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens));
5412 class_put_profile(lprof);
5417 OBD_FREE(name, name_size);
5419 OBD_FREE(uuid, uuid_size);
5423 /* setup quota master target on MDT0 */
5424 static int mdt_quota_init(const struct lu_env *env, struct mdt_device *mdt,
5425 struct lustre_cfg *cfg)
5427 struct obd_device *obd;
5428 char *dev = lustre_cfg_string(cfg, 0);
5429 char *qmtname, *uuid, *p;
5430 struct lustre_cfg_bufs *bufs;
5431 struct lustre_cfg *lcfg;
5432 struct lustre_profile *lprof;
5433 struct obd_connect_data *data;
5437 LASSERT(mdt->mdt_qmt_exp == NULL);
5438 LASSERT(mdt->mdt_qmt_dev == NULL);
5440 /* quota master is on MDT0 only for now */
5441 if (mdt->mdt_seq_site.ss_node_id != 0)
5444 /* MGS generates config commands which look as follows:
5445 * #01 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
5446 * 3:lustre-MDT0000-mdtlov 4:f
5448 * We generate the QMT name from the MDT one, just replacing MD with QM
5449 * after all the preparations, the logical equivalent will be:
5450 * #01 (160)setup 0:lustre-QMT0000 1:lustre-QMT0000_UUID 2:0
5451 * 3:lustre-MDT0000-osd 4:f */
5452 OBD_ALLOC(qmtname, MAX_OBD_NAME);
5453 OBD_ALLOC(uuid, UUID_MAX);
5454 OBD_ALLOC_PTR(bufs);
5455 OBD_ALLOC_PTR(data);
5456 if (qmtname == NULL || uuid == NULL || bufs == NULL || data == NULL)
5457 GOTO(cleanup_mem, rc = -ENOMEM);
5459 strcpy(qmtname, dev);
5460 p = strstr(qmtname, "-MDT");
5462 GOTO(cleanup_mem, rc = -ENOMEM);
5463 /* replace MD with QM */
5467 snprintf(uuid, UUID_MAX, "%s_UUID", qmtname);
5469 lprof = class_get_profile(lustre_cfg_string(cfg, 0));
5470 if (lprof == NULL || lprof->lp_dt == NULL) {
5471 CERROR("can't find profile for %s\n",
5472 lustre_cfg_string(cfg, 0));
5473 GOTO(cleanup_mem, rc = -EINVAL);
5476 lustre_cfg_bufs_reset(bufs, qmtname);
5477 lustre_cfg_bufs_set_string(bufs, 1, LUSTRE_QMT_NAME);
5478 lustre_cfg_bufs_set_string(bufs, 2, uuid);
5479 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
5481 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
5483 GOTO(put_profile, rc = -ENOMEM);
5484 lustre_cfg_init(lcfg, LCFG_ATTACH, bufs);
5486 rc = class_attach(lcfg);
5488 GOTO(lcfg_cleanup, rc);
5490 obd = class_name2obd(qmtname);
5492 CERROR("Can not find obd %s (%s in config)\n", qmtname,
5493 lustre_cfg_string(cfg, 0));
5494 GOTO(lcfg_cleanup, rc = -EINVAL);
5497 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens));
5499 lustre_cfg_bufs_reset(bufs, qmtname);
5500 lustre_cfg_bufs_set_string(bufs, 1, uuid);
5501 lustre_cfg_bufs_set_string(bufs, 2, dev);
5503 /* for quota, the next device should be the OSD device */
5504 lustre_cfg_bufs_set_string(bufs, 3,
5505 mdt->mdt_bottom->dd_lu_dev.ld_obd->obd_name);
5507 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
5509 GOTO(class_detach, rc = -ENOMEM);
5510 lustre_cfg_init(lcfg, LCFG_SETUP, bufs);
5512 rc = class_setup(obd, lcfg);
5514 GOTO(class_detach, rc);
5516 mdt->mdt_qmt_dev = obd->obd_lu_dev;
5518 /* configure local quota objects */
5519 if (OBD_FAIL_CHECK(OBD_FAIL_QUOTA_INIT))
5522 rc = mdt->mdt_qmt_dev->ld_ops->ldo_prepare(env,
5526 GOTO(class_cleanup, rc);
5528 /* connect to quota master target */
5529 data->ocd_connect_flags = OBD_CONNECT_VERSION;
5530 data->ocd_version = LUSTRE_VERSION_CODE;
5531 rc = obd_connect(NULL, &mdt->mdt_qmt_exp, obd, &obd->obd_uuid,
5534 CERROR("cannot connect to quota master device %s (%d)\n",
5536 GOTO(class_cleanup, rc);
5542 class_manual_cleanup(obd);
5543 mdt->mdt_qmt_dev = NULL;
5544 GOTO(lcfg_cleanup, rc);
5548 class_detach(obd, lcfg);
5550 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens));
5552 class_put_profile(lprof);
5557 OBD_FREE(qmtname, MAX_OBD_NAME);
5559 OBD_FREE(uuid, UUID_MAX);
5565 /* Shutdown quota master target associated with mdt */
5566 static void mdt_quota_fini(const struct lu_env *env, struct mdt_device *mdt)
5570 if (mdt->mdt_qmt_exp == NULL)
5572 LASSERT(mdt->mdt_qmt_dev != NULL);
5574 /* the qmt automatically shuts down when the mdt disconnects */
5575 obd_disconnect(mdt->mdt_qmt_exp);
5576 mdt->mdt_qmt_exp = NULL;
5577 mdt->mdt_qmt_dev = NULL;
5581 /* mdt_getxattr() is used from mdt_intent_getxattr(), use this wrapper
5582 * for now. This will be removed along with converting rest of MDT code
5583 * to use tgt_session_info */
5584 static int mdt_tgt_getxattr(struct tgt_session_info *tsi)
5586 struct mdt_thread_info *info = tsi2mdt_info(tsi);
5589 if (unlikely(info->mti_object == NULL))
5592 rc = mdt_getxattr(info);
5594 mdt_thread_info_fini(info);
5598 static int mdt_llog_open(struct tgt_session_info *tsi)
5602 if (!mdt_is_rootadmin(tsi2mdt_info(tsi)))
5603 RETURN(err_serious(-EACCES));
5605 RETURN(tgt_llog_open(tsi));
5608 #define OBD_FAIL_OST_READ_NET OBD_FAIL_OST_BRW_NET
5609 #define OBD_FAIL_OST_WRITE_NET OBD_FAIL_OST_BRW_NET
5610 #define OST_BRW_READ OST_READ
5611 #define OST_BRW_WRITE OST_WRITE
5613 static struct tgt_handler mdt_tgt_handlers[] = {
5614 TGT_RPC_HANDLER(MDS_FIRST_OPC,
5615 0, MDS_CONNECT, mdt_tgt_connect,
5616 &RQF_CONNECT, LUSTRE_OBD_VERSION),
5617 TGT_RPC_HANDLER(MDS_FIRST_OPC,
5618 0, MDS_DISCONNECT, tgt_disconnect,
5619 &RQF_MDS_DISCONNECT, LUSTRE_OBD_VERSION),
5620 TGT_RPC_HANDLER(MDS_FIRST_OPC,
5621 HAS_REPLY, MDS_SET_INFO, mdt_set_info,
5622 &RQF_MDT_SET_INFO, LUSTRE_MDS_VERSION),
5623 TGT_MDT_HDL(0, MDS_GET_INFO, mdt_get_info),
5624 TGT_MDT_HDL(HAS_REPLY, MDS_GET_ROOT, mdt_get_root),
5625 TGT_MDT_HDL(HAS_BODY, MDS_GETATTR, mdt_getattr),
5626 TGT_MDT_HDL(HAS_BODY | HAS_REPLY, MDS_GETATTR_NAME,
5628 TGT_MDT_HDL(HAS_BODY, MDS_GETXATTR, mdt_tgt_getxattr),
5629 TGT_MDT_HDL(HAS_REPLY, MDS_STATFS, mdt_statfs),
5630 TGT_MDT_HDL(IS_MUTABLE, MDS_REINT, mdt_reint),
5631 TGT_MDT_HDL(HAS_BODY, MDS_CLOSE, mdt_close),
5632 TGT_MDT_HDL(HAS_BODY | HAS_REPLY, MDS_READPAGE, mdt_readpage),
5633 TGT_MDT_HDL(HAS_BODY | HAS_REPLY, MDS_SYNC, mdt_sync),
5634 TGT_MDT_HDL(0, MDS_QUOTACTL, mdt_quotactl),
5635 TGT_MDT_HDL(HAS_BODY | HAS_REPLY | IS_MUTABLE, MDS_HSM_PROGRESS,
5637 TGT_MDT_HDL(HAS_BODY | HAS_REPLY | IS_MUTABLE, MDS_HSM_CT_REGISTER,
5638 mdt_hsm_ct_register),
5639 TGT_MDT_HDL(HAS_BODY | HAS_REPLY | IS_MUTABLE, MDS_HSM_CT_UNREGISTER,
5640 mdt_hsm_ct_unregister),
5641 TGT_MDT_HDL(HAS_BODY | HAS_REPLY, MDS_HSM_STATE_GET,
5643 TGT_MDT_HDL(HAS_BODY | HAS_REPLY | IS_MUTABLE, MDS_HSM_STATE_SET,
5645 TGT_MDT_HDL(HAS_BODY | HAS_REPLY, MDS_HSM_ACTION, mdt_hsm_action),
5646 TGT_MDT_HDL(HAS_BODY | HAS_REPLY, MDS_HSM_REQUEST,
5648 TGT_MDT_HDL(HAS_KEY | HAS_BODY | HAS_REPLY | IS_MUTABLE,
5651 TGT_MDT_HDL(IS_MUTABLE, MDS_RMFID, mdt_rmfid),
5654 static struct tgt_handler mdt_io_ops[] = {
5655 TGT_OST_HDL_HP(HAS_BODY | HAS_REPLY, OST_BRW_READ, tgt_brw_read,
5657 TGT_OST_HDL_HP(HAS_BODY | IS_MUTABLE, OST_BRW_WRITE, tgt_brw_write,
5659 TGT_OST_HDL_HP(HAS_BODY | HAS_REPLY | IS_MUTABLE,
5660 OST_PUNCH, mdt_punch_hdl,
5662 TGT_OST_HDL(HAS_BODY | HAS_REPLY, OST_SYNC, mdt_data_sync),
5663 TGT_OST_HDL(HAS_BODY | HAS_REPLY | IS_MUTABLE, OST_FALLOCATE,
5665 TGT_OST_HDL(HAS_BODY | HAS_REPLY, OST_SEEK, tgt_lseek),
5668 static struct tgt_handler mdt_sec_ctx_ops[] = {
5669 TGT_SEC_HDL_VAR(0, SEC_CTX_INIT, mdt_sec_ctx_handle),
5670 TGT_SEC_HDL_VAR(0, SEC_CTX_INIT_CONT,mdt_sec_ctx_handle),
5671 TGT_SEC_HDL_VAR(0, SEC_CTX_FINI, mdt_sec_ctx_handle)
5674 static struct tgt_handler mdt_quota_ops[] = {
5675 TGT_QUOTA_HDL(HAS_REPLY, QUOTA_DQACQ, mdt_quota_dqacq),
5678 static struct tgt_handler mdt_llog_handlers[] = {
5679 TGT_LLOG_HDL(0, LLOG_ORIGIN_HANDLE_CREATE, mdt_llog_open),
5680 TGT_LLOG_HDL(0, LLOG_ORIGIN_HANDLE_NEXT_BLOCK, tgt_llog_next_block),
5681 TGT_LLOG_HDL(0, LLOG_ORIGIN_HANDLE_READ_HEADER, tgt_llog_read_header),
5682 TGT_LLOG_HDL(0, LLOG_ORIGIN_HANDLE_PREV_BLOCK, tgt_llog_prev_block),
5685 static struct tgt_opc_slice mdt_common_slice[] = {
5687 .tos_opc_start = MDS_FIRST_OPC,
5688 .tos_opc_end = MDS_LAST_OPC,
5689 .tos_hs = mdt_tgt_handlers
5692 .tos_opc_start = OBD_FIRST_OPC,
5693 .tos_opc_end = OBD_LAST_OPC,
5694 .tos_hs = tgt_obd_handlers
5697 .tos_opc_start = LDLM_FIRST_OPC,
5698 .tos_opc_end = LDLM_LAST_OPC,
5699 .tos_hs = tgt_dlm_handlers
5702 .tos_opc_start = SEC_FIRST_OPC,
5703 .tos_opc_end = SEC_LAST_OPC,
5704 .tos_hs = mdt_sec_ctx_ops
5707 .tos_opc_start = OUT_UPDATE_FIRST_OPC,
5708 .tos_opc_end = OUT_UPDATE_LAST_OPC,
5709 .tos_hs = tgt_out_handlers
5712 .tos_opc_start = FLD_FIRST_OPC,
5713 .tos_opc_end = FLD_LAST_OPC,
5714 .tos_hs = fld_handlers
5717 .tos_opc_start = SEQ_FIRST_OPC,
5718 .tos_opc_end = SEQ_LAST_OPC,
5719 .tos_hs = seq_handlers
5722 .tos_opc_start = QUOTA_DQACQ,
5723 .tos_opc_end = QUOTA_LAST_OPC,
5724 .tos_hs = mdt_quota_ops
5727 .tos_opc_start = LLOG_FIRST_OPC,
5728 .tos_opc_end = LLOG_LAST_OPC,
5729 .tos_hs = mdt_llog_handlers
5732 .tos_opc_start = LFSCK_FIRST_OPC,
5733 .tos_opc_end = LFSCK_LAST_OPC,
5734 .tos_hs = tgt_lfsck_handlers
5737 .tos_opc_start = OST_FIRST_OPC,
5738 .tos_opc_end = OST_LAST_OPC,
5739 .tos_hs = mdt_io_ops
5746 static void mdt_fini(const struct lu_env *env, struct mdt_device *m)
5748 struct md_device *next = m->mdt_child;
5749 struct lu_device *d = &m->mdt_lu_dev;
5750 struct obd_device *obd = mdt2obd_dev(m);
5751 struct lfsck_stop stop;
5754 stop.ls_status = LS_PAUSED;
5756 next->md_ops->mdo_iocontrol(env, next, OBD_IOC_STOP_LFSCK, 0, &stop);
5758 mdt_stack_pre_fini(env, m, md2lu_dev(m->mdt_child));
5760 mdt_restriper_stop(m);
5761 ping_evictor_stop();
5763 /* Remove the HSM /proc entry so the coordinator cannot be
5764 * restarted by a user while it's shutting down.
5766 mdt_hsm_cdt_stop(m);
5768 mdt_llog_ctxt_unclone(env, m, LLOG_AGENT_ORIG_CTXT);
5769 mdt_llog_ctxt_unclone(env, m, LLOG_CHANGELOG_ORIG_CTXT);
5771 if (m->mdt_namespace != NULL)
5772 ldlm_namespace_free_prior(m->mdt_namespace, NULL,
5773 d->ld_obd->obd_force);
5775 obd_exports_barrier(obd);
5776 obd_zombie_barrier();
5778 mdt_quota_fini(env, m);
5780 cfs_free_nidlist(&m->mdt_squash.rsi_nosquash_nids);
5782 /* Calling the cleanup functions in the same order as in the mdt_init0
5785 mdt_tunables_fini(m);
5787 target_recovery_fini(obd);
5788 upcall_cache_cleanup(m->mdt_identity_cache);
5789 m->mdt_identity_cache = NULL;
5791 mdt_fs_cleanup(env, m);
5793 tgt_fini(env, &m->mdt_lut);
5795 mdt_hsm_cdt_fini(m);
5797 if (m->mdt_los != NULL) {
5798 local_oid_storage_fini(env, m->mdt_los);
5802 if (m->mdt_namespace != NULL) {
5803 ldlm_namespace_free_post(m->mdt_namespace);
5804 d->ld_obd->obd_namespace = m->mdt_namespace = NULL;
5807 if (m->mdt_md_root != NULL) {
5808 mdt_object_put(env, m->mdt_md_root);
5809 m->mdt_md_root = NULL;
5812 mdt_seq_fini(env, m);
5814 mdt_fld_fini(env, m);
5819 mdt_stack_fini(env, m, md2lu_dev(m->mdt_child));
5821 LASSERT(atomic_read(&d->ld_ref) == 0);
5823 server_put_mount(mdt_obd_name(m), true);
5828 static int mdt_postrecov(const struct lu_env *, struct mdt_device *);
5830 static int mdt_init0(const struct lu_env *env, struct mdt_device *m,
5831 struct lu_device_type *ldt, struct lustre_cfg *cfg)
5833 const struct dt_device_param *dt_conf;
5834 struct mdt_thread_info *info;
5835 struct obd_device *obd;
5836 const char *dev = lustre_cfg_string(cfg, 0);
5837 const char *num = lustre_cfg_string(cfg, 2);
5838 struct tg_grants_data *tgd = &m->mdt_lut.lut_tgd;
5839 struct lustre_mount_info *lmi = NULL;
5840 struct lustre_sb_info *lsi;
5842 struct seq_server_site *ss_site;
5843 const char *identity_upcall = "NONE";
5844 struct md_device *next;
5851 lu_device_init(&m->mdt_lu_dev, ldt);
5853 * Environment (env) might be missing mdt_thread_key values at that
5854 * point, if device is allocated when mdt_thread_key is in QUIESCENT
5857 * Usually device allocation path doesn't use module key values, but
5858 * mdt has to do a lot of work here, so allocate key value.
5860 rc = lu_env_refill((struct lu_env *)env);
5864 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
5865 LASSERT(info != NULL);
5867 obd = class_name2obd(dev);
5868 LASSERT(obd != NULL);
5870 m->mdt_max_mdsize = MAX_MD_SIZE_OLD;
5871 m->mdt_opts.mo_evict_tgt_nids = 1;
5872 m->mdt_opts.mo_cos = MDT_COS_DEFAULT;
5874 lmi = server_get_mount(dev);
5876 CERROR("Cannot get mount info for %s!\n", dev);
5879 lsi = s2lsi(lmi->lmi_sb);
5880 LASSERT(lsi->lsi_lmd);
5881 /* CMD is supported only in IAM mode */
5883 rc = kstrtol(num, 10, &node_id);
5887 obd->u.obt.obt_magic = OBT_MAGIC;
5888 if (lsi->lsi_lmd->lmd_flags & LMD_FLG_SKIP_LFSCK)
5889 m->mdt_skip_lfsck = 1;
5892 /* Just try to get a DoM lock by default. Otherwise, having a group
5893 * lock granted, it may get blocked for a long time. */
5894 m->mdt_opts.mo_dom_lock = TRYLOCK_DOM_ON_OPEN;
5895 /* DoM files are read at open and data is packed in the reply */
5896 m->mdt_opts.mo_dom_read_open = 1;
5898 m->mdt_squash.rsi_uid = 0;
5899 m->mdt_squash.rsi_gid = 0;
5900 INIT_LIST_HEAD(&m->mdt_squash.rsi_nosquash_nids);
5901 spin_lock_init(&m->mdt_squash.rsi_lock);
5902 spin_lock_init(&m->mdt_lock);
5903 m->mdt_enable_remote_dir = 1;
5904 m->mdt_enable_striped_dir = 1;
5905 m->mdt_enable_dir_migration = 1;
5906 m->mdt_enable_dir_restripe = 0;
5907 m->mdt_enable_dir_auto_split = 0;
5908 m->mdt_enable_remote_dir_gid = 0;
5909 m->mdt_enable_chprojid_gid = 0;
5910 m->mdt_enable_remote_rename = 1;
5911 m->mdt_dir_restripe_nsonly = 1;
5912 m->mdt_enable_remote_subdir_mount = 1;
5914 atomic_set(&m->mdt_mds_mds_conns, 0);
5915 atomic_set(&m->mdt_async_commit_count, 0);
5917 m->mdt_lu_dev.ld_ops = &mdt_lu_ops;
5918 m->mdt_lu_dev.ld_obd = obd;
5919 /* Set this lu_device to obd for error handling purposes. */
5920 obd->obd_lu_dev = &m->mdt_lu_dev;
5922 /* init the stack */
5923 rc = mdt_stack_init((struct lu_env *)env, m, cfg);
5925 CERROR("%s: Can't init device stack, rc %d\n",
5926 mdt_obd_name(m), rc);
5931 ss_site = mdt_seq_site(m);
5932 s->ld_seq_site = ss_site;
5935 /* set server index */
5936 ss_site->ss_node_id = node_id;
5938 /* failover is the default
5939 * FIXME: we do not failout mds0/mgs, which may cause some problems.
5940 * assumed whose ss_node_id == 0 XXX
5942 obd->obd_replayable = 1;
5943 /* No connection accepted until configurations will finish */
5944 obd->obd_no_conn = 1;
5946 if (cfg->lcfg_bufcount > 4 && LUSTRE_CFG_BUFLEN(cfg, 4) > 0) {
5947 char *str = lustre_cfg_string(cfg, 4);
5948 if (strchr(str, 'n')) {
5949 CWARN("%s: recovery disabled\n", mdt_obd_name(m));
5950 obd->obd_replayable = 0;
5954 rc = mdt_fld_init(env, mdt_obd_name(m), m);
5956 GOTO(err_fini_stack, rc);
5958 rc = mdt_seq_init(env, m);
5960 GOTO(err_fini_fld, rc);
5962 snprintf(info->mti_u.ns_name, sizeof(info->mti_u.ns_name), "%s-%s",
5963 LUSTRE_MDT_NAME, obd->obd_uuid.uuid);
5964 m->mdt_namespace = ldlm_namespace_new(obd, info->mti_u.ns_name,
5965 LDLM_NAMESPACE_SERVER,
5966 LDLM_NAMESPACE_GREEDY,
5968 if (IS_ERR(m->mdt_namespace)) {
5969 rc = PTR_ERR(m->mdt_namespace);
5970 CERROR("%s: unable to create server namespace: rc = %d\n",
5972 m->mdt_namespace = NULL;
5973 GOTO(err_fini_seq, rc);
5976 m->mdt_namespace->ns_lvbp = m;
5977 m->mdt_namespace->ns_lvbo = &mdt_lvbo;
5979 ldlm_register_intent(m->mdt_namespace, mdt_intent_policy);
5980 /* set obd_namespace for compatibility with old code */
5981 obd->obd_namespace = m->mdt_namespace;
5983 rc = tgt_init(env, &m->mdt_lut, obd, m->mdt_bottom, mdt_common_slice,
5984 OBD_FAIL_MDS_ALL_REQUEST_NET,
5985 OBD_FAIL_MDS_ALL_REPLY_NET);
5987 GOTO(err_free_ns, rc);
5989 /* Amount of available space excluded from granting and reserved
5990 * for metadata. It is a percentage of the total MDT size. */
5991 tgd->tgd_reserved_pcnt = 10;
5993 if (ONE_MB_BRW_SIZE < (1U << tgd->tgd_blockbits))
5994 m->mdt_brw_size = 1U << tgd->tgd_blockbits;
5996 m->mdt_brw_size = ONE_MB_BRW_SIZE;
5998 rc = mdt_fs_setup(env, m, obd, lsi);
6002 fid.f_seq = FID_SEQ_LOCAL_NAME;
6005 rc = local_oid_storage_init(env, m->mdt_bottom, &fid, &m->mdt_los);
6007 GOTO(err_fs_cleanup, rc);
6009 rc = mdt_hsm_cdt_init(m);
6011 CERROR("%s: error initializing coordinator, rc %d\n",
6012 mdt_obd_name(m), rc);
6013 GOTO(err_los_fini, rc);
6016 tgt_adapt_sptlrpc_conf(&m->mdt_lut);
6018 next = m->mdt_child;
6019 dt_conf = next->md_ops->mdo_dtconf_get(env, next);
6021 mntopts = dt_conf->ddp_mntopts;
6023 if (mntopts & MNTOPT_USERXATTR)
6024 m->mdt_opts.mo_user_xattr = 1;
6026 m->mdt_opts.mo_user_xattr = 0;
6028 m->mdt_max_ea_size = dt_conf->ddp_max_ea_size;
6030 if (mntopts & MNTOPT_ACL)
6031 m->mdt_opts.mo_acl = 1;
6033 m->mdt_opts.mo_acl = 0;
6035 /* XXX: to support suppgid for ACL, we enable identity_upcall
6036 * by default, otherwise, maybe got unexpected -EACCESS. */
6037 if (m->mdt_opts.mo_acl)
6038 identity_upcall = MDT_IDENTITY_UPCALL_PATH;
6040 m->mdt_identity_cache = upcall_cache_init(mdt_obd_name(m),
6042 &mdt_identity_upcall_cache_ops);
6043 if (IS_ERR(m->mdt_identity_cache)) {
6044 rc = PTR_ERR(m->mdt_identity_cache);
6045 m->mdt_identity_cache = NULL;
6046 GOTO(err_free_hsm, rc);
6049 rc = mdt_tunables_init(m, dev);
6051 CERROR("Can't init MDT lprocfs, rc %d\n", rc);
6052 GOTO(err_recovery, rc);
6055 rc = mdt_quota_init(env, m, cfg);
6057 GOTO(err_procfs, rc);
6059 m->mdt_ldlm_client = &mdt2obd_dev(m)->obd_ldlm_client;
6060 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
6061 "mdt_ldlm_client", m->mdt_ldlm_client);
6063 ping_evictor_start();
6065 /* recovery will be started upon mdt_prepare()
6066 * when the whole stack is complete and ready
6067 * to serve the requests */
6069 /* Reduce the initial timeout on an MDS because it doesn't need such
6070 * a long timeout as an OST does. Adaptive timeouts will adjust this
6071 * value appropriately. */
6072 if (ldlm_timeout == LDLM_TIMEOUT_DEFAULT)
6073 ldlm_timeout = MDS_LDLM_TIMEOUT_DEFAULT;
6075 if ((lsi->lsi_lmd->lmd_flags & LMD_FLG_LOCAL_RECOV))
6076 m->mdt_lut.lut_local_recovery = 1;
6078 rc = mdt_restriper_start(m);
6080 GOTO(err_ping_evictor, rc);
6085 ping_evictor_stop();
6087 mdt_tunables_fini(m);
6089 upcall_cache_cleanup(m->mdt_identity_cache);
6090 m->mdt_identity_cache = NULL;
6092 mdt_hsm_cdt_fini(m);
6094 local_oid_storage_fini(env, m->mdt_los);
6097 mdt_fs_cleanup(env, m);
6099 /* keep recoverable clients */
6101 target_recovery_fini(obd);
6102 obd_exports_barrier(obd);
6103 obd_zombie_barrier();
6104 tgt_fini(env, &m->mdt_lut);
6106 ldlm_namespace_free(m->mdt_namespace, NULL, 0);
6107 obd->obd_namespace = m->mdt_namespace = NULL;
6109 mdt_seq_fini(env, m);
6111 mdt_fld_fini(env, m);
6113 mdt_stack_fini(env, m, md2lu_dev(m->mdt_child));
6116 server_put_mount(dev, true);
6120 /* For interoperability, the left element is old parameter, the right one
6121 * is the new version of the parameter, if some parameter is deprecated,
6122 * the new version should be set as NULL. */
6123 static struct cfg_interop_param mdt_interop_param[] = {
6124 { "mdt.group_upcall", NULL },
6125 { "mdt.quota_type", NULL },
6126 { "mdd.quota_type", NULL },
6127 { "mdt.som", NULL },
6128 { "mdt.rootsquash", "mdt.root_squash" },
6129 { "mdt.nosquash_nid", "mdt.nosquash_nids" },
6133 /* used by MGS to process specific configurations */
6134 static int mdt_process_config(const struct lu_env *env,
6135 struct lu_device *d, struct lustre_cfg *cfg)
6137 struct mdt_device *m = mdt_dev(d);
6138 struct md_device *md_next = m->mdt_child;
6139 struct lu_device *next = md2lu_dev(md_next);
6143 switch (cfg->lcfg_command) {
6145 struct obd_device *obd = d->ld_obd;
6146 /* For interoperability */
6147 struct cfg_interop_param *ptr = NULL;
6148 struct lustre_cfg *old_cfg = NULL;
6152 param = lustre_cfg_string(cfg, 1);
6153 if (param == NULL) {
6154 CERROR("param is empty\n");
6159 ptr = class_find_old_param(param, mdt_interop_param);
6161 if (ptr->new_param == NULL) {
6163 CWARN("For interoperability, skip this %s."
6164 " It is obsolete.\n", ptr->old_param);
6168 CWARN("Found old param %s, changed it to %s.\n",
6169 ptr->old_param, ptr->new_param);
6172 cfg = lustre_cfg_rename(old_cfg, ptr->new_param);
6179 count = class_modify_config(cfg, PARAM_MDT,
6180 &obd->obd_kset.kobj);
6182 struct coordinator *cdt = &m->mdt_coordinator;
6184 /* is it an HSM var ? */
6185 count = class_modify_config(cfg, PARAM_HSM,
6186 &cdt->cdt_hsm_kobj);
6188 /* we don't understand; pass it on */
6189 rc = next->ld_ops->ldo_process_config(env, next,
6192 rc = count > 0 ? 0 : count;
6194 rc = count > 0 ? 0 : count;
6198 OBD_FREE(cfg, lustre_cfg_len(cfg->lcfg_bufcount,
6199 cfg->lcfg_buflens));
6203 /* others are passed further */
6204 rc = next->ld_ops->ldo_process_config(env, next, cfg);
6210 static struct lu_object *mdt_object_alloc(const struct lu_env *env,
6211 const struct lu_object_header *hdr,
6212 struct lu_device *d)
6214 struct mdt_object *mo;
6218 OBD_SLAB_ALLOC_PTR_GFP(mo, mdt_object_kmem, GFP_NOFS);
6220 struct lu_object *o;
6221 struct lu_object_header *h;
6224 h = &mo->mot_header;
6225 lu_object_header_init(h);
6226 lu_object_init(o, h, d);
6227 lu_object_add_top(h, o);
6228 o->lo_ops = &mdt_obj_ops;
6229 spin_lock_init(&mo->mot_write_lock);
6230 mutex_init(&mo->mot_som_mutex);
6231 mutex_init(&mo->mot_lov_mutex);
6232 init_rwsem(&mo->mot_dom_sem);
6233 init_rwsem(&mo->mot_open_sem);
6234 atomic_set(&mo->mot_open_count, 0);
6235 mo->mot_restripe_offset = 0;
6236 INIT_LIST_HEAD(&mo->mot_restripe_linkage);
6242 static int mdt_object_init(const struct lu_env *env, struct lu_object *o,
6243 const struct lu_object_conf *unused)
6245 struct mdt_device *d = mdt_dev(o->lo_dev);
6246 struct lu_device *under;
6247 struct lu_object *below;
6251 CDEBUG(D_INFO, "object init, fid = "DFID"\n",
6252 PFID(lu_object_fid(o)));
6254 under = &d->mdt_child->md_lu_dev;
6255 below = under->ld_ops->ldo_object_alloc(env, o->lo_header, under);
6256 if (below != NULL) {
6257 lu_object_add(o, below);
6264 static void mdt_object_free_rcu(struct rcu_head *head)
6266 struct mdt_object *mo = container_of(head, struct mdt_object,
6267 mot_header.loh_rcu);
6269 kmem_cache_free(mdt_object_kmem, mo);
6272 static void mdt_object_free(const struct lu_env *env, struct lu_object *o)
6274 struct mdt_object *mo = mdt_obj(o);
6275 struct lu_object_header *h;
6279 CDEBUG(D_INFO, "object free, fid = "DFID"\n",
6280 PFID(lu_object_fid(o)));
6282 LASSERT(atomic_read(&mo->mot_open_count) == 0);
6283 LASSERT(atomic_read(&mo->mot_lease_count) == 0);
6286 lu_object_header_fini(h);
6287 OBD_FREE_PRE(mo, sizeof(*mo), "slab-freed");
6288 call_rcu(&mo->mot_header.loh_rcu, mdt_object_free_rcu);
6293 static int mdt_object_print(const struct lu_env *env, void *cookie,
6294 lu_printer_t p, const struct lu_object *o)
6296 struct mdt_object *mdto = mdt_obj((struct lu_object *)o);
6298 return (*p)(env, cookie,
6299 LUSTRE_MDT_NAME"-object@%p(%s %s, writecount=%d)",
6300 mdto, mdto->mot_lov_created ? "lov_created" : "",
6301 mdto->mot_cache_attr ? "cache_attr" : "",
6302 mdto->mot_write_count);
6305 static int mdt_prepare(const struct lu_env *env,
6306 struct lu_device *pdev,
6307 struct lu_device *cdev)
6309 struct mdt_device *mdt = mdt_dev(cdev);
6310 struct lu_device *next = &mdt->mdt_child->md_lu_dev;
6311 struct obd_device *obd = cdev->ld_obd;
6318 rc = next->ld_ops->ldo_prepare(env, cdev, next);
6322 rc = mdt_llog_ctxt_clone(env, mdt, LLOG_CHANGELOG_ORIG_CTXT);
6326 rc = mdt_llog_ctxt_clone(env, mdt, LLOG_AGENT_ORIG_CTXT);
6330 rc = lfsck_register_namespace(env, mdt->mdt_bottom, mdt->mdt_namespace);
6331 /* The LFSCK instance is registered just now, so it must be there when
6332 * register the namespace to such instance. */
6333 LASSERTF(rc == 0, "register namespace failed: rc = %d\n", rc);
6335 if (mdt->mdt_seq_site.ss_node_id == 0) {
6336 rc = mdt->mdt_child->md_ops->mdo_root_get(env, mdt->mdt_child,
6337 &mdt->mdt_md_root_fid);
6342 LASSERT(!test_bit(MDT_FL_CFGLOG, &mdt->mdt_state));
6344 target_recovery_init(&mdt->mdt_lut, tgt_request_handle);
6345 set_bit(MDT_FL_CFGLOG, &mdt->mdt_state);
6346 LASSERT(obd->obd_no_conn);
6347 spin_lock(&obd->obd_dev_lock);
6348 obd->obd_no_conn = 0;
6349 spin_unlock(&obd->obd_dev_lock);
6351 if (obd->obd_recovering == 0)
6352 mdt_postrecov(env, mdt);
6357 const struct lu_device_operations mdt_lu_ops = {
6358 .ldo_object_alloc = mdt_object_alloc,
6359 .ldo_process_config = mdt_process_config,
6360 .ldo_prepare = mdt_prepare,
6363 static const struct lu_object_operations mdt_obj_ops = {
6364 .loo_object_init = mdt_object_init,
6365 .loo_object_free = mdt_object_free,
6366 .loo_object_print = mdt_object_print
6369 static int mdt_obd_set_info_async(const struct lu_env *env,
6370 struct obd_export *exp,
6371 __u32 keylen, void *key,
6372 __u32 vallen, void *val,
6373 struct ptlrpc_request_set *set)
6379 if (KEY_IS(KEY_SPTLRPC_CONF)) {
6380 rc = tgt_adapt_sptlrpc_conf(class_exp2tgt(exp));
6387 static inline void mdt_enable_slc(struct mdt_device *mdt)
6389 if (mdt->mdt_lut.lut_sync_lock_cancel == SYNC_LOCK_CANCEL_NEVER)
6390 mdt->mdt_lut.lut_sync_lock_cancel = SYNC_LOCK_CANCEL_BLOCKING;
6393 static inline void mdt_disable_slc(struct mdt_device *mdt)
6395 if (mdt->mdt_lut.lut_sync_lock_cancel == SYNC_LOCK_CANCEL_BLOCKING)
6396 mdt->mdt_lut.lut_sync_lock_cancel = SYNC_LOCK_CANCEL_NEVER;
6400 * Match client and server connection feature flags.
6402 * Compute the compatibility flags for a connection request based on
6403 * features mutually supported by client and server.
6405 * The obd_export::exp_connect_data.ocd_connect_flags field in \a exp
6406 * must not be updated here, otherwise a partially initialized value may
6407 * be exposed. After the connection request is successfully processed,
6408 * the top-level MDT connect request handler atomically updates the export
6409 * connect flags from the obd_connect_data::ocd_connect_flags field of the
6410 * reply. \see mdt_connect().
6412 * Before 2.7.50 clients will send a struct obd_connect_data_v1 rather than a
6413 * full struct obd_connect_data. So care must be taken when accessing fields
6414 * that are not present in struct obd_connect_data_v1. See LU-16.
6416 * \param exp the obd_export associated with this client/target pair
6417 * \param mdt the target device for the connection
6418 * \param data stores data for this connect request
6421 * \retval -EPROTO \a data unexpectedly has zero obd_connect_data::ocd_brw_size
6422 * \retval -EBADE client and server feature requirements are incompatible
6424 static int mdt_connect_internal(const struct lu_env *env,
6425 struct obd_export *exp,
6426 struct mdt_device *mdt,
6427 struct obd_connect_data *data, bool reconnect)
6429 const char *obd_name = mdt_obd_name(mdt);
6430 LASSERT(data != NULL);
6432 data->ocd_connect_flags &= MDT_CONNECT_SUPPORTED;
6434 if (mdt->mdt_bottom->dd_rdonly &&
6435 !(data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
6436 !(data->ocd_connect_flags & OBD_CONNECT_RDONLY))
6439 if (data->ocd_connect_flags & OBD_CONNECT_FLAGS2)
6440 data->ocd_connect_flags2 &= MDT_CONNECT_SUPPORTED2;
6442 data->ocd_ibits_known &= MDS_INODELOCK_FULL;
6444 if (!mdt->mdt_opts.mo_acl)
6445 data->ocd_connect_flags &= ~OBD_CONNECT_ACL;
6447 if (!mdt->mdt_opts.mo_user_xattr)
6448 data->ocd_connect_flags &= ~OBD_CONNECT_XATTR;
6450 if (OCD_HAS_FLAG(data, BRW_SIZE)) {
6451 data->ocd_brw_size = min(data->ocd_brw_size,
6453 if (data->ocd_brw_size == 0) {
6454 CERROR("%s: cli %s/%p ocd_connect_flags: %#llx "
6455 "ocd_version: %x ocd_grant: %d ocd_index: %u "
6456 "ocd_brw_size unexpectedly zero, network data "
6457 "corruption? Refusing to connect this client\n",
6458 obd_name, exp->exp_client_uuid.uuid,
6459 exp, data->ocd_connect_flags, data->ocd_version,
6460 data->ocd_grant, data->ocd_index);
6465 if (OCD_HAS_FLAG(data, GRANT_PARAM)) {
6466 struct dt_device_param *ddp = &mdt->mdt_lut.lut_dt_conf;
6468 /* client is reporting its page size, for future use */
6469 exp->exp_target_data.ted_pagebits = data->ocd_grant_blkbits;
6470 data->ocd_grant_blkbits = mdt->mdt_lut.lut_tgd.tgd_blockbits;
6471 /* ddp_inodespace may not be power-of-two value, eg. for ldiskfs
6472 * it's LDISKFS_DIR_REC_LEN(20) = 28. */
6473 data->ocd_grant_inobits = fls(ddp->ddp_inodespace - 1);
6474 /* ocd_grant_tax_kb is in 1K byte blocks */
6475 data->ocd_grant_tax_kb = ddp->ddp_extent_tax >> 10;
6476 data->ocd_grant_max_blks = ddp->ddp_max_extent_blks;
6479 /* Save connect_data we have so far because tgt_grant_connect()
6480 * uses it to calculate grant, and we want to save the client
6481 * version before it is overwritten by LUSTRE_VERSION_CODE. */
6482 exp->exp_connect_data = *data;
6483 if (OCD_HAS_FLAG(data, GRANT))
6484 tgt_grant_connect(env, exp, data, !reconnect);
6486 if (OCD_HAS_FLAG(data, MAXBYTES))
6487 data->ocd_maxbytes = mdt->mdt_lut.lut_dt_conf.ddp_maxbytes;
6489 /* NB: Disregard the rule against updating
6490 * exp_connect_data.ocd_connect_flags in this case, since
6491 * tgt_client_new() needs to know if this is a lightweight
6492 * connection, and it is safe to expose this flag before
6493 * connection processing completes. */
6494 if (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) {
6495 spin_lock(&exp->exp_lock);
6496 *exp_connect_flags_ptr(exp) |= OBD_CONNECT_LIGHTWEIGHT;
6497 spin_unlock(&exp->exp_lock);
6500 data->ocd_version = LUSTRE_VERSION_CODE;
6502 if ((data->ocd_connect_flags & OBD_CONNECT_FID) == 0) {
6503 CWARN("%s: MDS requires FID support, but client not\n",
6508 if (OCD_HAS_FLAG(data, PINGLESS)) {
6509 if (ptlrpc_pinger_suppress_pings()) {
6510 spin_lock(&exp->exp_obd->obd_dev_lock);
6511 list_del_init(&exp->exp_obd_chain_timed);
6512 spin_unlock(&exp->exp_obd->obd_dev_lock);
6514 data->ocd_connect_flags &= ~OBD_CONNECT_PINGLESS;
6518 data->ocd_max_easize = mdt->mdt_max_ea_size;
6520 /* NB: Disregard the rule against updating
6521 * exp_connect_data.ocd_connect_flags in this case, since
6522 * tgt_client_new() needs to know if this is client supports
6523 * multiple modify RPCs, and it is safe to expose this flag before
6524 * connection processing completes. */
6525 if (data->ocd_connect_flags & OBD_CONNECT_MULTIMODRPCS) {
6526 data->ocd_maxmodrpcs = max_mod_rpcs_per_client;
6527 spin_lock(&exp->exp_lock);
6528 *exp_connect_flags_ptr(exp) |= OBD_CONNECT_MULTIMODRPCS;
6529 spin_unlock(&exp->exp_lock);
6532 if (OCD_HAS_FLAG(data, CKSUM)) {
6533 __u32 cksum_types = data->ocd_cksum_types;
6535 tgt_mask_cksum_types(&mdt->mdt_lut, &data->ocd_cksum_types);
6537 if (unlikely(data->ocd_cksum_types == 0)) {
6538 CERROR("%s: Connect with checksum support but no "
6539 "ocd_cksum_types is set\n",
6540 exp->exp_obd->obd_name);
6544 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
6545 "%x\n", exp->exp_obd->obd_name, obd_export_nid2str(exp),
6546 cksum_types, data->ocd_cksum_types);
6548 /* This client does not support OBD_CONNECT_CKSUM
6549 * fall back to CRC32 */
6550 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
6551 "OBD_CONNECT_CKSUM, CRC32 will be used\n",
6552 exp->exp_obd->obd_name, obd_export_nid2str(exp));
6555 if ((data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
6556 !(data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT)) {
6557 atomic_inc(&mdt->mdt_mds_mds_conns);
6558 mdt_enable_slc(mdt);
6561 if (!mdt->mdt_lut.lut_dt_conf.ddp_has_lseek_data_hole)
6562 data->ocd_connect_flags2 &= ~OBD_CONNECT2_LSEEK;
6567 static int mdt_ctxt_add_dirty_flag(struct lu_env *env,
6568 struct mdt_thread_info *info,
6569 struct mdt_file_data *mfd)
6571 struct lu_context ses;
6575 rc = lu_context_init(&ses, LCT_SERVER_SESSION);
6580 lu_context_enter(&ses);
6582 mdt_ucred(info)->uc_valid = UCRED_OLD;
6583 rc = mdt_add_dirty_flag(info, mfd->mfd_object, &info->mti_attr);
6585 lu_context_exit(&ses);
6586 lu_context_fini(&ses);
6592 static int mdt_export_cleanup(struct obd_export *exp)
6594 LIST_HEAD(closing_list);
6595 struct mdt_export_data *med = &exp->exp_mdt_data;
6596 struct obd_device *obd = exp->exp_obd;
6597 struct mdt_device *mdt;
6598 struct mdt_thread_info *info;
6600 struct mdt_file_data *mfd, *n;
6604 spin_lock(&med->med_open_lock);
6605 while (!list_empty(&med->med_open_head)) {
6606 struct list_head *tmp = med->med_open_head.next;
6607 mfd = list_entry(tmp, struct mdt_file_data, mfd_list);
6609 /* Remove mfd handle so it can't be found again.
6610 * We are consuming the mfd_list reference here. */
6611 class_handle_unhash(&mfd->mfd_open_handle);
6612 list_move_tail(&mfd->mfd_list, &closing_list);
6614 spin_unlock(&med->med_open_lock);
6615 mdt = mdt_dev(obd->obd_lu_dev);
6616 LASSERT(mdt != NULL);
6618 rc = lu_env_init(&env, LCT_MD_THREAD);
6622 info = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
6623 LASSERT(info != NULL);
6624 memset(info, 0, sizeof *info);
6625 info->mti_env = &env;
6626 info->mti_mdt = mdt;
6627 info->mti_exp = exp;
6629 if (!list_empty(&closing_list)) {
6630 struct md_attr *ma = &info->mti_attr;
6632 /* Close any open files (which may also cause orphan
6634 list_for_each_entry_safe(mfd, n, &closing_list, mfd_list) {
6635 list_del_init(&mfd->mfd_list);
6636 ma->ma_need = ma->ma_valid = 0;
6638 /* This file is being closed due to an eviction, it
6639 * could have been modified and now dirty regarding to
6640 * HSM archive, check this!
6641 * The logic here is to mark a file dirty if there's a
6642 * chance it was dirtied before the client was evicted,
6643 * so that we don't have to wait for a release attempt
6644 * before finding out the file was actually dirty and
6645 * fail the release. Aggressively marking it dirty here
6646 * will cause the policy engine to attempt to
6647 * re-archive it; when rearchiving, we can compare the
6648 * current version to the HSM data_version and make the
6649 * archive request into a noop if it's not actually
6652 if (mfd->mfd_open_flags & MDS_FMODE_WRITE)
6653 rc = mdt_ctxt_add_dirty_flag(&env, info, mfd);
6655 /* Don't unlink orphan on failover umount, LU-184 */
6656 if (exp->exp_flags & OBD_OPT_FAILOVER ||
6657 exp->exp_obd->obd_stopping) {
6658 ma->ma_valid = MA_FLAGS;
6659 ma->ma_attr_flags |= MDS_KEEP_ORPHAN;
6661 ma->ma_valid |= MA_FORCE_LOG;
6662 mdt_mfd_close(info, mfd);
6665 info->mti_mdt = NULL;
6666 /* cleanup client slot early */
6667 /* Do not erase record for recoverable client. */
6668 if (!(exp->exp_flags & OBD_OPT_FAILOVER) || exp->exp_failed)
6669 tgt_client_del(&env, exp);
6675 static int mdt_obd_disconnect(struct obd_export *exp)
6682 class_export_get(exp);
6684 if (!(exp->exp_flags & OBD_OPT_FORCE))
6685 tgt_grant_sanity_check(exp->exp_obd, __func__);
6687 if ((exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) &&
6688 !(exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)) {
6689 struct mdt_device *mdt = mdt_dev(exp->exp_obd->obd_lu_dev);
6691 if (atomic_dec_and_test(&mdt->mdt_mds_mds_conns))
6692 mdt_disable_slc(mdt);
6695 rc = server_disconnect_export(exp);
6697 CDEBUG(D_IOCTL, "server disconnect error: rc = %d\n", rc);
6699 tgt_grant_discard(exp);
6701 rc = mdt_export_cleanup(exp);
6702 nodemap_del_member(exp);
6703 class_export_put(exp);
6707 /* mds_connect copy */
6708 static int mdt_obd_connect(const struct lu_env *env,
6709 struct obd_export **exp, struct obd_device *obd,
6710 struct obd_uuid *cluuid,
6711 struct obd_connect_data *data,
6714 struct obd_export *lexp;
6715 struct lustre_handle conn = { 0 };
6716 struct mdt_device *mdt;
6718 lnet_nid_t *client_nid = localdata;
6721 LASSERT(env != NULL);
6722 LASSERT(data != NULL);
6724 if (!exp || !obd || !cluuid)
6727 mdt = mdt_dev(obd->obd_lu_dev);
6730 * first, check whether the stack is ready to handle requests
6731 * XXX: probably not very appropriate method is used now
6732 * at some point we should find a better one
6734 if (!test_bit(MDT_FL_SYNCED, &mdt->mdt_state) &&
6735 !(data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) &&
6736 !(data->ocd_connect_flags & OBD_CONNECT_MDS_MDS)) {
6737 rc = obd_get_info(env, mdt->mdt_child_exp,
6738 sizeof(KEY_OSP_CONNECTED),
6739 KEY_OSP_CONNECTED, NULL, NULL);
6742 set_bit(MDT_FL_SYNCED, &mdt->mdt_state);
6745 rc = class_connect(&conn, obd, cluuid);
6749 lexp = class_conn2export(&conn);
6750 LASSERT(lexp != NULL);
6752 rc = nodemap_add_member(*client_nid, lexp);
6753 if (rc != 0 && rc != -EEXIST)
6756 rc = mdt_connect_internal(env, lexp, mdt, data, false);
6758 struct lsd_client_data *lcd = lexp->exp_target_data.ted_lcd;
6761 memcpy(lcd->lcd_uuid, cluuid, sizeof lcd->lcd_uuid);
6762 rc = tgt_client_new(env, lexp);
6764 mdt_export_stats_init(obd, lexp, localdata);
6768 class_disconnect(lexp);
6769 nodemap_del_member(lexp);
6773 /* Because we do not want this export to be evicted by pinger,
6774 * let's not add this export to the timed chain list. */
6775 if (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) {
6776 spin_lock(&lexp->exp_obd->obd_dev_lock);
6777 list_del_init(&lexp->exp_obd_chain_timed);
6778 spin_unlock(&lexp->exp_obd->obd_dev_lock);
6785 static int mdt_obd_reconnect(const struct lu_env *env,
6786 struct obd_export *exp, struct obd_device *obd,
6787 struct obd_uuid *cluuid,
6788 struct obd_connect_data *data,
6791 lnet_nid_t *client_nid = localdata;
6795 if (exp == NULL || obd == NULL || cluuid == NULL)
6798 rc = nodemap_add_member(*client_nid, exp);
6799 if (rc != 0 && rc != -EEXIST)
6802 rc = mdt_connect_internal(env, exp, mdt_dev(obd->obd_lu_dev), data,
6805 mdt_export_stats_init(obd, exp, localdata);
6807 nodemap_del_member(exp);
6812 /* FIXME: Can we avoid using these two interfaces? */
6813 static int mdt_init_export(struct obd_export *exp)
6815 struct mdt_export_data *med = &exp->exp_mdt_data;
6819 INIT_LIST_HEAD(&med->med_open_head);
6820 spin_lock_init(&med->med_open_lock);
6821 spin_lock(&exp->exp_lock);
6822 exp->exp_connecting = 1;
6823 spin_unlock(&exp->exp_lock);
6825 OBD_ALLOC(exp->exp_used_slots,
6826 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
6827 if (exp->exp_used_slots == NULL)
6830 /* self-export doesn't need client data and ldlm initialization */
6831 if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
6832 &exp->exp_client_uuid)))
6835 rc = tgt_client_alloc(exp);
6839 rc = ldlm_init_export(exp);
6846 tgt_client_free(exp);
6848 OBD_FREE(exp->exp_used_slots,
6849 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
6850 exp->exp_used_slots = NULL;
6852 CERROR("%s: Failed to initialize export: rc = %d\n",
6853 exp->exp_obd->obd_name, rc);
6857 static int mdt_destroy_export(struct obd_export *exp)
6861 target_destroy_export(exp);
6862 if (exp->exp_used_slots)
6863 OBD_FREE(exp->exp_used_slots,
6864 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
6866 /* destroy can be called from failed obd_setup, so
6867 * checking uuid is safer than obd_self_export */
6868 if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
6869 &exp->exp_client_uuid)))
6872 ldlm_destroy_export(exp);
6873 tgt_client_free(exp);
6875 LASSERT(list_empty(&exp->exp_outstanding_replies));
6876 LASSERT(list_empty(&exp->exp_mdt_data.med_open_head));
6879 * discard grants once we're sure no more
6880 * interaction with the client is possible
6882 tgt_grant_discard(exp);
6883 if (exp_connect_flags(exp) & OBD_CONNECT_GRANT)
6884 exp->exp_obd->u.obt.obt_lut->lut_tgd.tgd_tot_granted_clients--;
6886 if (!(exp->exp_flags & OBD_OPT_FORCE))
6887 tgt_grant_sanity_check(exp->exp_obd, __func__);
6892 int mdt_links_read(struct mdt_thread_info *info, struct mdt_object *mdt_obj,
6893 struct linkea_data *ldata)
6897 LASSERT(ldata->ld_buf->lb_buf != NULL);
6899 if (!mdt_object_exists(mdt_obj))
6902 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
6903 ldata->ld_buf, XATTR_NAME_LINK);
6904 if (rc == -ERANGE) {
6905 /* Buf was too small, figure out what we need. */
6906 lu_buf_free(ldata->ld_buf);
6907 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
6908 ldata->ld_buf, XATTR_NAME_LINK);
6911 ldata->ld_buf = lu_buf_check_and_alloc(ldata->ld_buf, rc);
6912 if (ldata->ld_buf->lb_buf == NULL)
6914 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
6915 ldata->ld_buf, XATTR_NAME_LINK);
6920 return linkea_init_with_rec(ldata);
6924 * Given an MDT object, try to look up the full path to the object.
6925 * Part of the MDT layer implementation of lfs fid2path.
6927 * \param[in] info Per-thread common data shared by MDT level handlers.
6928 * \param[in] obj Object to do path lookup of
6929 * \param[in,out] fp User-provided struct to store path information
6930 * \param[in] root_fid Root FID of current path should reach
6932 * \retval 0 Lookup successful, path information stored in fp
6933 * \retval -EAGAIN Lookup failed, usually because object is being moved
6934 * \retval negative errno if there was a problem
6936 static int mdt_path_current(struct mdt_thread_info *info,
6937 struct mdt_object *obj,
6938 struct getinfo_fid2path *fp,
6939 struct lu_fid *root_fid)
6941 struct mdt_device *mdt = info->mti_mdt;
6942 struct lu_name *tmpname = &info->mti_name;
6943 struct lu_fid *tmpfid = &info->mti_tmp_fid1;
6944 struct lu_buf *buf = &info->mti_big_buf;
6945 struct linkea_data ldata = { NULL };
6947 struct mdt_object *mdt_obj;
6948 struct link_ea_header *leh;
6949 struct link_ea_entry *lee;
6956 /* temp buffer for path element, the buffer will be finally freed
6957 * in mdt_thread_info_fini */
6958 buf = lu_buf_check_and_alloc(buf, PATH_MAX);
6959 if (buf->lb_buf == NULL)
6963 ptr = fp->gf_u.gf_path + fp->gf_pathlen - 1;
6966 *tmpfid = fp->gf_fid = *mdt_object_fid(obj);
6968 while (!lu_fid_eq(root_fid, &fp->gf_fid)) {
6969 if (!lu_fid_eq(root_fid, &mdt->mdt_md_root_fid) &&
6970 lu_fid_eq(&mdt->mdt_md_root_fid, &fp->gf_fid))
6971 GOTO(out, rc = -ENOENT);
6973 if (lu_fid_eq(mdt_object_fid(obj), tmpfid)) {
6975 mdt_object_get(info->mti_env, mdt_obj);
6977 mdt_obj = mdt_object_find(info->mti_env, mdt, tmpfid);
6978 if (IS_ERR(mdt_obj))
6979 GOTO(out, rc = PTR_ERR(mdt_obj));
6982 if (!mdt_object_exists(mdt_obj)) {
6983 mdt_object_put(info->mti_env, mdt_obj);
6984 GOTO(out, rc = -ENOENT);
6987 if (mdt_object_remote(mdt_obj)) {
6988 mdt_object_put(info->mti_env, mdt_obj);
6989 GOTO(remote_out, rc = -EREMOTE);
6992 rc = mdt_links_read(info, mdt_obj, &ldata);
6994 mdt_object_put(info->mti_env, mdt_obj);
6999 lee = (struct link_ea_entry *)(leh + 1); /* link #0 */
7000 linkea_entry_unpack(lee, &reclen, tmpname, tmpfid);
7001 /* If set, use link #linkno for path lookup, otherwise use
7002 link #0. Only do this for the final path element. */
7003 if (first && fp->gf_linkno < leh->leh_reccount) {
7005 for (count = 0; count < fp->gf_linkno; count++) {
7006 lee = (struct link_ea_entry *)
7007 ((char *)lee + reclen);
7008 linkea_entry_unpack(lee, &reclen, tmpname,
7011 if (fp->gf_linkno < leh->leh_reccount - 1)
7012 /* indicate to user there are more links */
7016 /* Check if it is slave stripes */
7017 rc = mdt_is_dir_stripe(info, mdt_obj);
7018 mdt_object_put(info->mti_env, mdt_obj);
7022 fp->gf_fid = *tmpfid;
7026 /* Pack the name in the end of the buffer */
7027 ptr -= tmpname->ln_namelen;
7028 if (ptr - 1 <= fp->gf_u.gf_path)
7029 GOTO(out, rc = -EOVERFLOW);
7030 strncpy(ptr, tmpname->ln_name, tmpname->ln_namelen);
7033 /* keep the last resolved fid to the client, so the
7034 * client will build the left path on another MDT for
7036 fp->gf_fid = *tmpfid;
7041 /* non-zero will be treated as an error */
7045 ptr++; /* skip leading / */
7046 memmove(fp->gf_u.gf_path, ptr,
7047 fp->gf_u.gf_path + fp->gf_pathlen - ptr);
7054 * Given an MDT object, use mdt_path_current to get the path.
7055 * Essentially a wrapper to retry mdt_path_current a set number of times
7056 * if -EAGAIN is returned (usually because an object is being moved).
7058 * Part of the MDT layer implementation of lfs fid2path.
7060 * \param[in] info Per-thread common data shared by mdt level handlers.
7061 * \param[in] obj Object to do path lookup of
7062 * \param[in,out] fp User-provided struct for arguments and to store path
7065 * \retval 0 Lookup successful, path information stored in fp
7066 * \retval negative errno if there was a problem
7068 static int mdt_path(struct mdt_thread_info *info, struct mdt_object *obj,
7069 struct getinfo_fid2path *fp, struct lu_fid *root_fid)
7071 struct mdt_device *mdt = info->mti_mdt;
7076 if (fp->gf_pathlen < 3)
7079 if (root_fid == NULL)
7080 root_fid = &mdt->mdt_md_root_fid;
7082 if (lu_fid_eq(root_fid, mdt_object_fid(obj))) {
7083 fp->gf_u.gf_path[0] = '\0';
7087 /* Retry multiple times in case file is being moved */
7088 while (tries-- && rc == -EAGAIN)
7089 rc = mdt_path_current(info, obj, fp, root_fid);
7095 * Get the full path of the provided FID, as of changelog record recno.
7097 * This checks sanity and looks up object for user provided FID
7098 * before calling the actual path lookup code.
7100 * Part of the MDT layer implementation of lfs fid2path.
7102 * \param[in] info Per-thread common data shared by mdt level handlers.
7103 * \param[in,out] fp User-provided struct for arguments and to store path
7106 * \retval 0 Lookup successful, path information and recno stored in fp
7107 * \retval -ENOENT, object does not exist
7108 * \retval negative errno if there was a problem
7110 static int mdt_fid2path(struct mdt_thread_info *info,
7111 struct lu_fid *root_fid,
7112 struct getinfo_fid2path *fp)
7114 struct mdt_device *mdt = info->mti_mdt;
7115 struct mdt_object *obj;
7119 CDEBUG(D_IOCTL, "path get "DFID" from %llu #%d\n",
7120 PFID(&fp->gf_fid), fp->gf_recno, fp->gf_linkno);
7122 if (!fid_is_sane(&fp->gf_fid))
7125 if (!fid_is_namespace_visible(&fp->gf_fid)) {
7126 CDEBUG(D_INFO, "%s: "DFID" is invalid, f_seq should be >= %#llx"
7127 ", or f_oid != 0, or f_ver == 0\n", mdt_obd_name(mdt),
7128 PFID(&fp->gf_fid), (__u64)FID_SEQ_NORMAL);
7132 obj = mdt_object_find(info->mti_env, mdt, &fp->gf_fid);
7135 CDEBUG(D_IOCTL, "cannot find "DFID": rc = %d\n",
7136 PFID(&fp->gf_fid), rc);
7140 if (mdt_object_remote(obj)) {
7142 } else if (!mdt_object_exists(obj)) {
7145 struct lu_attr la = { 0 };
7146 struct dt_object *dt = mdt_obj2dt(obj);
7148 if (dt && dt->do_ops && dt->do_ops->do_attr_get)
7149 dt_attr_get(info->mti_env, mdt_obj2dt(obj), &la);
7150 if (la.la_valid & LA_FLAGS && la.la_flags & LUSTRE_ENCRYPT_FL)
7151 /* path resolution cannot be carried out on server
7152 * side for encrypted files
7160 mdt_object_put(info->mti_env, obj);
7161 CDEBUG(D_IOCTL, "nonlocal object "DFID": rc = %d\n",
7162 PFID(&fp->gf_fid), rc);
7166 rc = mdt_path(info, obj, fp, root_fid);
7168 CDEBUG(D_INFO, "fid "DFID", path %s recno %#llx linkno %u\n",
7169 PFID(&fp->gf_fid), fp->gf_u.gf_path,
7170 fp->gf_recno, fp->gf_linkno);
7172 mdt_object_put(info->mti_env, obj);
7177 static int mdt_rpc_fid2path(struct mdt_thread_info *info, void *key, int keylen,
7178 void *val, int vallen)
7180 struct getinfo_fid2path *fpout, *fpin;
7181 struct lu_fid *root_fid = NULL;
7184 fpin = key + cfs_size_round(sizeof(KEY_FID2PATH));
7187 if (req_capsule_req_need_swab(info->mti_pill))
7188 lustre_swab_fid2path(fpin);
7190 memcpy(fpout, fpin, sizeof(*fpin));
7191 if (fpout->gf_pathlen != vallen - sizeof(*fpin))
7194 if (keylen >= cfs_size_round(sizeof(KEY_FID2PATH)) + sizeof(*fpin) +
7195 sizeof(struct lu_fid)) {
7196 /* client sent its root FID, which is normally fileset FID */
7197 root_fid = fpin->gf_u.gf_root_fid;
7198 if (req_capsule_req_need_swab(info->mti_pill))
7199 lustre_swab_lu_fid(root_fid);
7201 if (root_fid != NULL && !fid_is_sane(root_fid))
7205 rc = mdt_fid2path(info, root_fid, fpout);
7209 int mdt_get_info(struct tgt_session_info *tsi)
7219 key = req_capsule_client_get(tsi->tsi_pill, &RMF_GETINFO_KEY);
7221 CDEBUG(D_IOCTL, "No GETINFO key\n");
7222 RETURN(err_serious(-EFAULT));
7224 keylen = req_capsule_get_size(tsi->tsi_pill, &RMF_GETINFO_KEY,
7227 vallen = req_capsule_client_get(tsi->tsi_pill, &RMF_GETINFO_VALLEN);
7228 if (vallen == NULL) {
7229 CDEBUG(D_IOCTL, "%s: cannot get RMF_GETINFO_VALLEN buffer\n",
7230 tgt_name(tsi->tsi_tgt));
7231 RETURN(err_serious(-EFAULT));
7234 req_capsule_set_size(tsi->tsi_pill, &RMF_GETINFO_VAL, RCL_SERVER,
7236 rc = req_capsule_server_pack(tsi->tsi_pill);
7238 RETURN(err_serious(rc));
7240 valout = req_capsule_server_get(tsi->tsi_pill, &RMF_GETINFO_VAL);
7241 if (valout == NULL) {
7242 CDEBUG(D_IOCTL, "%s: cannot get get-info RPC out buffer\n",
7243 tgt_name(tsi->tsi_tgt));
7244 RETURN(err_serious(-EFAULT));
7247 if (KEY_IS(KEY_FID2PATH)) {
7248 struct mdt_thread_info *info = tsi2mdt_info(tsi);
7250 rc = mdt_rpc_fid2path(info, key, keylen, valout, *vallen);
7251 mdt_thread_info_fini(info);
7258 static int mdt_ioc_version_get(struct mdt_thread_info *mti, void *karg)
7260 struct obd_ioctl_data *data = karg;
7263 struct mdt_object *obj;
7264 struct mdt_lock_handle *lh;
7268 if (data->ioc_inlbuf1 == NULL || data->ioc_inllen1 != sizeof(*fid) ||
7269 data->ioc_inlbuf2 == NULL || data->ioc_inllen2 != sizeof(version))
7272 fid = (struct lu_fid *)data->ioc_inlbuf1;
7274 if (!fid_is_sane(fid))
7277 CDEBUG(D_IOCTL, "getting version for "DFID"\n", PFID(fid));
7279 lh = &mti->mti_lh[MDT_LH_PARENT];
7280 mdt_lock_reg_init(lh, LCK_CR);
7282 obj = mdt_object_find_lock(mti, fid, lh, MDS_INODELOCK_UPDATE);
7284 RETURN(PTR_ERR(obj));
7286 if (mdt_object_remote(obj)) {
7289 * before calling version get the correct MDS should be
7290 * fid, this is error to find remote object here
7292 CERROR("nonlocal object "DFID"\n", PFID(fid));
7293 } else if (!mdt_object_exists(obj)) {
7294 *(__u64 *)data->ioc_inlbuf2 = ENOENT_VERSION;
7297 version = dt_version_get(mti->mti_env, mdt_obj2dt(obj));
7298 *(__u64 *)data->ioc_inlbuf2 = version;
7301 mdt_object_unlock_put(mti, obj, lh, 1);
7305 /* ioctls on obd dev */
7306 static int mdt_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
7307 void *karg, void __user *uarg)
7310 struct obd_device *obd = exp->exp_obd;
7311 struct mdt_device *mdt = mdt_dev(obd->obd_lu_dev);
7312 struct dt_device *dt = mdt->mdt_bottom;
7316 CDEBUG(D_IOCTL, "handling ioctl cmd %#x\n", cmd);
7317 rc = lu_env_init(&env, LCT_MD_THREAD);
7323 rc = mdt_device_sync(&env, mdt);
7325 case OBD_IOC_SET_READONLY:
7326 rc = dt_sync(&env, dt);
7328 rc = dt_ro(&env, dt);
7330 case OBD_IOC_ABORT_RECOVERY: {
7331 struct obd_ioctl_data *data = karg;
7333 CERROR("%s: Aborting recovery for device\n", mdt_obd_name(mdt));
7334 if (data->ioc_type & OBD_FLG_ABORT_RECOV_MDT) {
7335 obd->obd_abort_recov_mdt = 1;
7336 wake_up(&obd->obd_next_transno_waitq);
7337 } else { /* if (data->ioc_type & OBD_FLG_ABORT_RECOV_OST) */
7338 /* lctl didn't set OBD_FLG_ABORT_RECOV_OST < 2.13.57 */
7339 obd->obd_abort_recovery = 1;
7340 target_stop_recovery_thread(obd);
7345 case OBD_IOC_CHANGELOG_REG:
7346 case OBD_IOC_CHANGELOG_DEREG:
7347 case OBD_IOC_CHANGELOG_CLEAR:
7348 rc = mdt->mdt_child->md_ops->mdo_iocontrol(&env,
7352 case OBD_IOC_START_LFSCK: {
7353 struct md_device *next = mdt->mdt_child;
7354 struct obd_ioctl_data *data = karg;
7355 struct lfsck_start_param lsp;
7357 if (unlikely(data == NULL)) {
7362 lsp.lsp_start = (struct lfsck_start *)(data->ioc_inlbuf1);
7363 lsp.lsp_index_valid = 0;
7364 rc = next->md_ops->mdo_iocontrol(&env, next, cmd, 0, &lsp);
7367 case OBD_IOC_STOP_LFSCK: {
7368 struct md_device *next = mdt->mdt_child;
7369 struct obd_ioctl_data *data = karg;
7370 struct lfsck_stop stop;
7372 stop.ls_status = LS_STOPPED;
7373 /* Old lfsck utils may pass NULL @stop. */
7374 if (data->ioc_inlbuf1 == NULL)
7378 ((struct lfsck_stop *)(data->ioc_inlbuf1))->ls_flags;
7380 rc = next->md_ops->mdo_iocontrol(&env, next, cmd, 0, &stop);
7383 case OBD_IOC_QUERY_LFSCK: {
7384 struct md_device *next = mdt->mdt_child;
7385 struct obd_ioctl_data *data = karg;
7387 rc = next->md_ops->mdo_iocontrol(&env, next, cmd, 0,
7391 case OBD_IOC_GET_OBJ_VERSION: {
7392 struct mdt_thread_info *mti;
7393 mti = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
7394 memset(mti, 0, sizeof *mti);
7395 mti->mti_env = &env;
7399 rc = mdt_ioc_version_get(mti, karg);
7402 case OBD_IOC_CATLOGLIST: {
7403 struct mdt_thread_info *mti;
7405 mti = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
7406 lu_local_obj_fid(&mti->mti_tmp_fid1, LLOG_CATALOGS_OID);
7407 rc = llog_catalog_list(&env, mdt->mdt_bottom, 0, karg,
7408 &mti->mti_tmp_fid1);
7413 CERROR("%s: Not supported cmd = %d, rc = %d\n",
7414 mdt_obd_name(mdt), cmd, rc);
7421 static int mdt_postrecov(const struct lu_env *env, struct mdt_device *mdt)
7423 struct lu_device *ld = md2lu_dev(mdt->mdt_child);
7427 if (!mdt->mdt_skip_lfsck && !mdt->mdt_bottom->dd_rdonly) {
7428 struct lfsck_start_param lsp;
7430 lsp.lsp_start = NULL;
7431 lsp.lsp_index_valid = 0;
7432 rc = mdt->mdt_child->md_ops->mdo_iocontrol(env, mdt->mdt_child,
7433 OBD_IOC_START_LFSCK,
7435 if (rc != 0 && rc != -EALREADY)
7436 CWARN("%s: auto trigger paused LFSCK failed: rc = %d\n",
7437 mdt_obd_name(mdt), rc);
7440 rc = ld->ld_ops->ldo_recovery_complete(env, ld);
7444 static int mdt_obd_postrecov(struct obd_device *obd)
7449 rc = lu_env_init(&env, LCT_MD_THREAD);
7452 rc = mdt_postrecov(&env, mdt_dev(obd->obd_lu_dev));
7457 static const struct obd_ops mdt_obd_device_ops = {
7458 .o_owner = THIS_MODULE,
7459 .o_set_info_async = mdt_obd_set_info_async,
7460 .o_connect = mdt_obd_connect,
7461 .o_reconnect = mdt_obd_reconnect,
7462 .o_disconnect = mdt_obd_disconnect,
7463 .o_init_export = mdt_init_export,
7464 .o_destroy_export = mdt_destroy_export,
7465 .o_iocontrol = mdt_iocontrol,
7466 .o_postrecov = mdt_obd_postrecov,
7467 /* Data-on-MDT IO methods */
7468 .o_preprw = mdt_obd_preprw,
7469 .o_commitrw = mdt_obd_commitrw,
7472 static struct lu_device* mdt_device_fini(const struct lu_env *env,
7473 struct lu_device *d)
7475 struct mdt_device *m = mdt_dev(d);
7482 static struct lu_device *mdt_device_free(const struct lu_env *env,
7483 struct lu_device *d)
7485 struct mdt_device *m = mdt_dev(d);
7488 lu_device_fini(&m->mdt_lu_dev);
7494 static struct lu_device *mdt_device_alloc(const struct lu_env *env,
7495 struct lu_device_type *t,
7496 struct lustre_cfg *cfg)
7498 struct lu_device *l;
7499 struct mdt_device *m;
7506 rc = mdt_init0(env, m, t, cfg);
7508 mdt_device_free(env, l);
7513 l = ERR_PTR(-ENOMEM);
7517 /* context key constructor/destructor: mdt_key_init, mdt_key_fini */
7518 LU_KEY_INIT(mdt, struct mdt_thread_info);
7520 static void mdt_key_fini(const struct lu_context *ctx,
7521 struct lu_context_key *key, void* data)
7523 struct mdt_thread_info *info = data;
7525 if (info->mti_big_lmm) {
7526 OBD_FREE_LARGE(info->mti_big_lmm, info->mti_big_lmmsize);
7527 info->mti_big_lmm = NULL;
7528 info->mti_big_lmmsize = 0;
7531 if (info->mti_big_acl) {
7532 OBD_FREE_LARGE(info->mti_big_acl, info->mti_big_aclsize);
7533 info->mti_big_acl = NULL;
7534 info->mti_big_aclsize = 0;
7540 /* context key: mdt_thread_key */
7541 LU_CONTEXT_KEY_DEFINE(mdt, LCT_MD_THREAD);
7543 struct lu_ucred *mdt_ucred(const struct mdt_thread_info *info)
7545 return lu_ucred(info->mti_env);
7548 struct lu_ucred *mdt_ucred_check(const struct mdt_thread_info *info)
7550 return lu_ucred_check(info->mti_env);
7554 * Enable/disable COS (Commit On Sharing).
7556 * Set/Clear the COS flag in mdt options.
7558 * \param mdt mdt device
7559 * \param val 0 disables COS, other values enable COS
7561 void mdt_enable_cos(struct mdt_device *mdt, bool val)
7566 mdt->mdt_opts.mo_cos = val;
7567 rc = lu_env_init(&env, LCT_LOCAL);
7568 if (unlikely(rc != 0)) {
7569 CWARN("%s: lu_env initialization failed, cannot "
7570 "sync: rc = %d\n", mdt_obd_name(mdt), rc);
7573 mdt_device_sync(&env, mdt);
7578 * Check COS (Commit On Sharing) status.
7580 * Return COS flag status.
7582 * \param mdt mdt device
7584 int mdt_cos_is_enabled(struct mdt_device *mdt)
7586 return mdt->mdt_opts.mo_cos != 0;
7589 static const struct lu_device_type_operations mdt_device_type_ops = {
7590 .ldto_device_alloc = mdt_device_alloc,
7591 .ldto_device_free = mdt_device_free,
7592 .ldto_device_fini = mdt_device_fini
7595 static struct lu_device_type mdt_device_type = {
7596 .ldt_tags = LU_DEVICE_MD,
7597 .ldt_name = LUSTRE_MDT_NAME,
7598 .ldt_ops = &mdt_device_type_ops,
7599 .ldt_ctx_tags = LCT_MD_THREAD
7602 static int __init mdt_init(void)
7606 BUILD_BUG_ON(sizeof("0x0123456789ABCDEF:0x01234567:0x01234567") !=
7607 FID_NOBRACE_LEN + 1);
7608 BUILD_BUG_ON(sizeof("[0x0123456789ABCDEF:0x01234567:0x01234567]") !=
7610 rc = lu_kmem_init(mdt_caches);
7614 rc = mds_mod_init();
7618 rc = class_register_type(&mdt_obd_device_ops, NULL, true,
7619 LUSTRE_MDT_NAME, &mdt_device_type);
7624 lu_kmem_fini(mdt_caches);
7631 static void __exit mdt_exit(void)
7633 class_unregister_type(LUSTRE_MDT_NAME);
7635 lu_kmem_fini(mdt_caches);
7638 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
7639 MODULE_DESCRIPTION("Lustre Metadata Target ("LUSTRE_MDT_NAME")");
7640 MODULE_VERSION(LUSTRE_VERSION_STRING);
7641 MODULE_LICENSE("GPL");
7643 module_init(mdt_init);
7644 module_exit(mdt_exit);