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 #if OBD_OCD_VERSION(3, 0, 53, 0) > LUSTRE_VERSION_CODE
71 static int mdt_max_mod_rpcs_per_client_set(const char *val,
72 cfs_kernel_param_arg_t *kp)
77 rc = kstrtouint(val, 0, &num);
81 if (num < 1 || num > OBD_MAX_RIF_MAX)
84 CWARN("max_mod_rpcs_per_client is deprecated, set mdt.*.max_mod_rpcs_in_flight parameter instead\n");
86 max_mod_rpcs_per_client = num;
89 static const struct kernel_param_ops
90 param_ops_max_mod_rpcs_per_client = {
91 .set = mdt_max_mod_rpcs_per_client_set,
92 .get = param_get_uint,
95 #define param_check_max_mod_rpcs_per_client(name, p) \
96 __param_check(name, p, unsigned int)
98 module_param_cb(max_mod_rpcs_per_client,
99 ¶m_ops_max_mod_rpcs_per_client,
100 &max_mod_rpcs_per_client, 0644);
102 MODULE_PARM_DESC(max_mod_rpcs_per_client,
103 "maximum number of modify RPCs in flight allowed per client (Deprecated)");
106 mdl_mode_t mdt_mdl_lock_modes[] = {
107 [LCK_MINMODE] = MDL_MINMODE,
114 [LCK_GROUP] = MDL_GROUP
117 enum ldlm_mode mdt_dlm_lock_modes[] = {
118 [MDL_MINMODE] = LCK_MINMODE,
125 [MDL_GROUP] = LCK_GROUP
128 static struct mdt_device *mdt_dev(struct lu_device *d);
130 static const struct lu_object_operations mdt_obj_ops;
132 /* Slab for MDT object allocation */
133 static struct kmem_cache *mdt_object_kmem;
135 /* For HSM restore handles */
136 struct kmem_cache *mdt_hsm_cdt_kmem;
138 /* For HSM request handles */
139 struct kmem_cache *mdt_hsm_car_kmem;
141 static struct lu_kmem_descr mdt_caches[] = {
143 .ckd_cache = &mdt_object_kmem,
144 .ckd_name = "mdt_obj",
145 .ckd_size = sizeof(struct mdt_object)
148 .ckd_cache = &mdt_hsm_cdt_kmem,
149 .ckd_name = "mdt_cdt_restore_handle",
150 .ckd_size = sizeof(struct cdt_restore_handle)
153 .ckd_cache = &mdt_hsm_car_kmem,
154 .ckd_name = "mdt_cdt_agent_req",
155 .ckd_size = sizeof(struct cdt_agent_req)
162 __u64 mdt_get_disposition(struct ldlm_reply *rep, __u64 op_flag)
166 return rep->lock_policy_res1 & op_flag;
169 void mdt_clear_disposition(struct mdt_thread_info *info,
170 struct ldlm_reply *rep, __u64 op_flag)
173 info->mti_opdata &= ~op_flag;
174 tgt_opdata_clear(info->mti_env, op_flag);
177 rep->lock_policy_res1 &= ~op_flag;
180 void mdt_set_disposition(struct mdt_thread_info *info,
181 struct ldlm_reply *rep, __u64 op_flag)
184 info->mti_opdata |= op_flag;
185 tgt_opdata_set(info->mti_env, op_flag);
188 rep->lock_policy_res1 |= op_flag;
191 /* assert lock is unlocked before reuse */
192 static inline void mdt_lock_handle_assert(struct mdt_lock_handle *lh)
194 LASSERT(!lustre_handle_is_used(&lh->mlh_reg_lh));
195 LASSERT(!lustre_handle_is_used(&lh->mlh_pdo_lh));
196 LASSERT(!lustre_handle_is_used(&lh->mlh_rreg_lh));
199 void mdt_lock_reg_init(struct mdt_lock_handle *lh, enum ldlm_mode lm)
201 mdt_lock_handle_assert(lh);
202 lh->mlh_pdo_hash = 0;
203 lh->mlh_reg_mode = lm;
204 lh->mlh_rreg_mode = lm;
205 lh->mlh_type = MDT_REG_LOCK;
208 void mdt_lh_reg_init(struct mdt_lock_handle *lh, struct ldlm_lock *lock)
210 mdt_lock_reg_init(lh, lock->l_req_mode);
211 if (lock->l_req_mode == LCK_GROUP)
212 lh->mlh_gid = lock->l_policy_data.l_inodebits.li_gid;
215 void mdt_lock_pdo_init(struct mdt_lock_handle *lh, enum ldlm_mode lock_mode,
216 const struct lu_name *lname)
218 mdt_lock_handle_assert(lh);
219 lh->mlh_reg_mode = lock_mode;
220 lh->mlh_pdo_mode = LCK_MINMODE;
221 lh->mlh_rreg_mode = lock_mode;
222 lh->mlh_type = MDT_PDO_LOCK;
224 if (lu_name_is_valid(lname)) {
225 lh->mlh_pdo_hash = ll_full_name_hash(NULL, lname->ln_name,
227 /* XXX Workaround for LU-2856
229 * Zero is a valid return value of full_name_hash, but
230 * several users of mlh_pdo_hash assume a non-zero
231 * hash value. We therefore map zero onto an
232 * arbitrary, but consistent value (1) to avoid
233 * problems further down the road. */
234 if (unlikely(lh->mlh_pdo_hash == 0))
235 lh->mlh_pdo_hash = 1;
237 lh->mlh_pdo_hash = 0;
241 static void mdt_lock_pdo_mode(struct mdt_thread_info *info, struct mdt_object *o,
242 struct mdt_lock_handle *lh)
248 * Any dir access needs couple of locks:
250 * 1) on part of dir we gonna take lookup/modify;
252 * 2) on whole dir to protect it from concurrent splitting and/or to
253 * flush client's cache for readdir().
255 * so, for a given mode and object this routine decides what lock mode
256 * to use for lock #2:
258 * 1) if caller's gonna lookup in dir then we need to protect dir from
259 * being splitted only - LCK_CR
261 * 2) if caller's gonna modify dir then we need to protect dir from
262 * being splitted and to flush cache - LCK_CW
264 * 3) if caller's gonna modify dir and that dir seems ready for
265 * splitting then we need to protect it from any type of access
266 * (lookup/modify/split) - LCK_EX --bzzz
269 LASSERT(lh->mlh_reg_mode != LCK_MINMODE);
270 LASSERT(lh->mlh_pdo_mode == LCK_MINMODE);
273 * Ask underlaying level its opinion about preferable PDO lock mode
274 * having access type passed as regular lock mode:
276 * - MDL_MINMODE means that lower layer does not want to specify lock
279 * - MDL_NL means that no PDO lock should be taken. This is used in some
280 * cases. Say, for non-splittable directories no need to use PDO locks
283 mode = mdo_lock_mode(info->mti_env, mdt_object_child(o),
284 mdt_dlm_mode2mdl_mode(lh->mlh_reg_mode));
286 if (mode != MDL_MINMODE) {
287 lh->mlh_pdo_mode = mdt_mdl_mode2dlm_mode(mode);
290 * Lower layer does not want to specify locking mode. We do it
291 * our selves. No special protection is needed, just flush
292 * client's cache on modification and allow concurrent
295 switch (lh->mlh_reg_mode) {
297 lh->mlh_pdo_mode = LCK_EX;
300 lh->mlh_pdo_mode = LCK_CR;
303 lh->mlh_pdo_mode = LCK_CW;
306 CERROR("Not expected lock type (0x%x)\n",
307 (int)lh->mlh_reg_mode);
312 LASSERT(lh->mlh_pdo_mode != LCK_MINMODE);
317 * Check whether \a o is directory stripe object.
319 * \param[in] info thread environment
320 * \param[in] o MDT object
322 * \retval 1 is directory stripe.
323 * \retval 0 isn't directory stripe.
324 * \retval < 1 error code
326 static int mdt_is_dir_stripe(struct mdt_thread_info *info,
327 struct mdt_object *o)
329 struct md_attr *ma = &info->mti_attr;
330 struct lmv_mds_md_v1 *lmv;
333 rc = mdt_stripe_get(info, o, ma, XATTR_NAME_LMV);
337 if (!(ma->ma_valid & MA_LMV))
340 lmv = &ma->ma_lmv->lmv_md_v1;
342 if (!lmv_is_sane2(lmv))
345 if (le32_to_cpu(lmv->lmv_magic) == LMV_MAGIC_STRIPE)
351 static int mdt_lookup_fileset(struct mdt_thread_info *info, const char *fileset,
354 struct mdt_device *mdt = info->mti_mdt;
355 struct lu_name *lname = &info->mti_name;
356 const char *start = fileset;
357 char *filename = info->mti_filename;
358 struct mdt_object *obj;
361 LASSERT(!info->mti_cross_ref);
364 * We may want to allow this to mount a completely separate
365 * fileset from the MDT in the future, but keeping it to
366 * ROOT/ only for now avoid potential security issues.
368 *fid = mdt->mdt_md_root_fid;
370 while (rc == 0 && start != NULL && *start != '\0') {
371 const char *s1 = start;
377 while (*s2 != '/' && *s2 != '\0')
385 lname->ln_namelen = s2 - s1;
386 if (lname->ln_namelen > NAME_MAX) {
391 /* reject .. as a path component */
392 if (lname->ln_namelen == 2 && strncmp(s1, "..", 2) == 0) {
397 strncpy(filename, s1, lname->ln_namelen);
398 filename[lname->ln_namelen] = '\0';
399 lname->ln_name = filename;
401 obj = mdt_object_find(info->mti_env, mdt, fid);
406 /* Only got the fid of this obj by name */
408 rc = mdo_lookup(info->mti_env, mdt_object_child(obj), lname,
409 fid, &info->mti_spec);
410 if (!rc && !S_ISDIR(lu_object_attr(&obj->mot_obj)))
412 mdt_object_put(info->mti_env, obj);
418 static int mdt_get_root(struct tgt_session_info *tsi)
420 struct mdt_thread_info *info = tsi2mdt_info(tsi);
421 struct mdt_device *mdt = info->mti_mdt;
422 struct mdt_body *repbody;
423 char *fileset = NULL, *buffer = NULL;
425 struct obd_export *exp = info->mti_exp;
426 char *nodemap_fileset;
430 rc = mdt_check_ucred(info);
432 GOTO(out, rc = err_serious(rc));
434 if (CFS_FAIL_CHECK(OBD_FAIL_MDS_GET_ROOT_PACK))
435 GOTO(out, rc = err_serious(-ENOMEM));
437 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
438 if (req_capsule_get_size(info->mti_pill, &RMF_NAME, RCL_CLIENT) > 0) {
439 fileset = req_capsule_client_get(info->mti_pill, &RMF_NAME);
441 GOTO(out, rc = err_serious(-EFAULT));
444 nodemap_fileset = nodemap_get_fileset(exp->exp_target_data.ted_nodemap);
445 if (nodemap_fileset && nodemap_fileset[0]) {
446 CDEBUG(D_INFO, "nodemap fileset is %s\n", nodemap_fileset);
448 /* consider fileset from client as a sub-fileset
449 * of the nodemap one */
450 OBD_ALLOC(buffer, PATH_MAX + 1);
452 GOTO(out, rc = err_serious(-ENOMEM));
453 if (snprintf(buffer, PATH_MAX + 1, "%s/%s",
454 nodemap_fileset, fileset) >= PATH_MAX + 1)
455 GOTO(out, rc = err_serious(-EINVAL));
458 /* enforce fileset as specified in the nodemap */
459 fileset = nodemap_fileset;
464 CDEBUG(D_INFO, "Getting fileset %s\n", fileset);
465 rc = mdt_lookup_fileset(info, fileset, &repbody->mbo_fid1);
467 GOTO(out, rc = err_serious(rc));
469 repbody->mbo_fid1 = mdt->mdt_md_root_fid;
471 exp->exp_root_fid = repbody->mbo_fid1;
472 repbody->mbo_valid |= OBD_MD_FLID;
476 mdt_thread_info_fini(info);
478 OBD_FREE(buffer, PATH_MAX+1);
482 static int mdt_statfs(struct tgt_session_info *tsi)
484 struct ptlrpc_request *req = tgt_ses_req(tsi);
485 struct mdt_thread_info *info = tsi2mdt_info(tsi);
486 struct mdt_device *mdt = info->mti_mdt;
487 struct tg_grants_data *tgd = &mdt->mdt_lut.lut_tgd;
488 struct md_device *next = mdt->mdt_child;
489 struct ptlrpc_service_part *svcpt;
490 struct obd_statfs *osfs;
491 struct mdt_body *reqbody = NULL;
492 struct mdt_statfs_cache *msf;
493 ktime_t kstart = ktime_get();
494 int current_blockbits;
499 svcpt = req->rq_rqbd->rqbd_svcpt;
501 /* This will trigger a watchdog timeout */
502 CFS_FAIL_TIMEOUT(OBD_FAIL_MDS_STATFS_LCW_SLEEP,
503 (MDT_SERVICE_WATCHDOG_FACTOR *
504 at_get(&svcpt->scp_at_estimate)) + 1);
506 rc = mdt_check_ucred(info);
508 GOTO(out, rc = err_serious(rc));
510 if (CFS_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK))
511 GOTO(out, rc = err_serious(-ENOMEM));
513 osfs = req_capsule_server_get(info->mti_pill, &RMF_OBD_STATFS);
515 GOTO(out, rc = -EPROTO);
517 if (mdt_is_sum_statfs_client(req->rq_export) &&
518 lustre_packed_msg_size(req->rq_reqmsg) ==
519 req_capsule_fmt_size(req->rq_reqmsg->lm_magic,
520 &RQF_MDS_STATFS_NEW, RCL_CLIENT)) {
521 req_capsule_extend(info->mti_pill, &RQF_MDS_STATFS_NEW);
522 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
525 if (reqbody && reqbody->mbo_valid & OBD_MD_FLAGSTATFS)
526 msf = &mdt->mdt_sum_osfs;
528 msf = &mdt->mdt_osfs;
530 if (msf->msf_age + OBD_STATFS_CACHE_SECONDS <= ktime_get_seconds()) {
531 /** statfs data is too old, get up-to-date one */
532 if (reqbody && reqbody->mbo_valid & OBD_MD_FLAGSTATFS)
533 rc = next->md_ops->mdo_statfs(info->mti_env,
536 rc = dt_statfs(info->mti_env, mdt->mdt_bottom,
540 spin_lock(&mdt->mdt_lock);
541 msf->msf_osfs = *osfs;
542 msf->msf_age = ktime_get_seconds();
543 spin_unlock(&mdt->mdt_lock);
545 /** use cached statfs data */
546 spin_lock(&mdt->mdt_lock);
547 *osfs = msf->msf_osfs;
548 spin_unlock(&mdt->mdt_lock);
551 /* tgd_blockbit is recordsize bits set during mkfs.
552 * This once set does not change. However, 'zfs set'
553 * can be used to change the MDT blocksize. Instead
554 * of using cached value of 'tgd_blockbit' always
555 * calculate the blocksize bits which may have
558 current_blockbits = fls64(osfs->os_bsize) - 1;
560 /* at least try to account for cached pages. its still racy and
561 * might be under-reporting if clients haven't announced their
562 * caches with brw recently */
563 CDEBUG(D_SUPER | D_CACHE, "blocks cached %llu granted %llu"
564 " pending %llu free %llu avail %llu\n",
565 tgd->tgd_tot_dirty, tgd->tgd_tot_granted,
566 tgd->tgd_tot_pending,
567 osfs->os_bfree << current_blockbits,
568 osfs->os_bavail << current_blockbits);
570 osfs->os_bavail -= min_t(u64, osfs->os_bavail,
571 ((tgd->tgd_tot_dirty + tgd->tgd_tot_pending +
572 osfs->os_bsize - 1) >> current_blockbits));
574 tgt_grant_sanity_check(mdt->mdt_lu_dev.ld_obd, __func__);
575 CDEBUG(D_CACHE, "%llu blocks: %llu free, %llu avail; "
576 "%llu objects: %llu free; state %x\n",
577 osfs->os_blocks, osfs->os_bfree, osfs->os_bavail,
578 osfs->os_files, osfs->os_ffree, osfs->os_state);
580 if (!exp_grant_param_supp(tsi->tsi_exp) &&
581 current_blockbits > COMPAT_BSIZE_SHIFT) {
582 /* clients which don't support OBD_CONNECT_GRANT_PARAM
583 * should not see a block size > page size, otherwise
584 * cl_lost_grant goes mad. Therefore, we emulate a 4KB (=2^12)
585 * block size which is the biggest block size known to work
586 * with all client's page size. */
587 osfs->os_blocks <<= current_blockbits - COMPAT_BSIZE_SHIFT;
588 osfs->os_bfree <<= current_blockbits - COMPAT_BSIZE_SHIFT;
589 osfs->os_bavail <<= current_blockbits - COMPAT_BSIZE_SHIFT;
590 osfs->os_bsize = 1 << COMPAT_BSIZE_SHIFT;
593 mdt_counter_incr(req, LPROC_MDT_STATFS,
594 ktime_us_delta(ktime_get(), kstart));
596 mdt_thread_info_fini(info);
600 __u32 mdt_lmm_dom_entry_check(struct lov_mds_md *lmm, int *is_dom_only)
602 struct lov_comp_md_v1 *comp_v1;
603 struct lov_mds_md *v1;
605 __u32 dom_stripesize = 0;
607 bool has_ost_stripes = false;
614 if (le32_to_cpu(lmm->lmm_magic) != LOV_MAGIC_COMP_V1)
617 comp_v1 = (struct lov_comp_md_v1 *)lmm;
618 off = le32_to_cpu(comp_v1->lcm_entries[0].lcme_offset);
619 v1 = (struct lov_mds_md *)((char *)comp_v1 + off);
621 /* Fast check for DoM entry with no mirroring, should be the first */
622 if (le16_to_cpu(comp_v1->lcm_mirror_count) == 0 &&
623 !(lov_pattern(le32_to_cpu(v1->lmm_pattern)) & LOV_PATTERN_MDT))
626 /* check all entries otherwise */
627 for (i = 0; i < le16_to_cpu(comp_v1->lcm_entry_count); i++) {
628 struct lov_comp_md_entry_v1 *lcme;
630 lcme = &comp_v1->lcm_entries[i];
631 if (!(le32_to_cpu(lcme->lcme_flags) & LCME_FL_INIT))
634 off = le32_to_cpu(lcme->lcme_offset);
635 v1 = (struct lov_mds_md *)((char *)comp_v1 + off);
637 if (lov_pattern(le32_to_cpu(v1->lmm_pattern)) &
639 dom_stripesize = le32_to_cpu(v1->lmm_stripe_size);
641 has_ost_stripes = true;
643 if (dom_stripesize && has_ost_stripes)
644 RETURN(dom_stripesize);
646 /* DoM-only case exits here */
647 if (is_dom_only && dom_stripesize)
649 RETURN(dom_stripesize);
653 * Pack size attributes into the reply.
655 int mdt_pack_size2body(struct mdt_thread_info *info,
656 const struct lu_fid *fid, struct lustre_handle *lh)
659 struct md_attr *ma = &info->mti_attr;
661 bool dom_lock = false;
665 LASSERT(ma->ma_attr.la_valid & LA_MODE);
667 if (!S_ISREG(ma->ma_attr.la_mode) ||
668 !(ma->ma_valid & MA_LOV && ma->ma_lmm != NULL))
671 dom_stripe = mdt_lmm_dom_stripesize(ma->ma_lmm);
672 /* no DoM stripe, no size in reply */
676 if (lustre_handle_is_used(lh)) {
677 struct ldlm_lock *lock;
679 lock = ldlm_handle2lock(lh);
681 dom_lock = ldlm_has_dom(lock);
686 /* no DoM lock, no size in reply */
690 /* Either DoM lock exists or LMM has only DoM stripe then
691 * return size on body. */
692 b = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
694 mdt_dom_object_size(info->mti_env, info->mti_mdt, fid, b, dom_lock);
698 #ifdef CONFIG_LUSTRE_FS_POSIX_ACL
700 * Pack ACL data into the reply. UIDs/GIDs are mapped and filtered by nodemap.
702 * \param info thread info object
703 * \param repbody reply to pack ACLs into
704 * \param o mdt object of file to examine
705 * \param nodemap nodemap of client to reply to
707 * \retval -errno error getting or parsing ACL from disk
709 int mdt_pack_acl2body(struct mdt_thread_info *info, struct mdt_body *repbody,
710 struct mdt_object *o, struct lu_nodemap *nodemap)
712 const struct lu_env *env = info->mti_env;
713 struct md_object *next = mdt_object_child(o);
714 struct lu_buf *buf = &info->mti_buf;
715 struct mdt_device *mdt = info->mti_mdt;
716 struct req_capsule *pill = info->mti_pill;
721 buf->lb_buf = req_capsule_server_get(pill, &RMF_ACL);
722 buf->lb_len = req_capsule_get_size(pill, &RMF_ACL, RCL_SERVER);
723 if (buf->lb_len == 0)
726 LASSERT(!info->mti_big_acl_used);
728 rc = mo_xattr_get(env, next, buf, XATTR_NAME_ACL_ACCESS);
730 if (rc == -ENODATA) {
731 repbody->mbo_aclsize = 0;
732 repbody->mbo_valid |= OBD_MD_FLACL;
734 } else if (rc == -EOPNOTSUPP) {
736 } else if (rc == -ERANGE) {
737 if (exp_connect_large_acl(info->mti_exp) &&
738 !info->mti_big_acl_used) {
739 if (info->mti_big_acl == NULL) {
740 info->mti_big_aclsize =
742 mdt->mdt_max_ea_size,
744 OBD_ALLOC_LARGE(info->mti_big_acl,
745 info->mti_big_aclsize);
746 if (info->mti_big_acl == NULL) {
747 info->mti_big_aclsize = 0;
748 CERROR("%s: unable to grow "
751 PFID(mdt_object_fid(o)));
756 CDEBUG(D_INODE, "%s: grow the "DFID
757 " ACL buffer to size %d\n",
759 PFID(mdt_object_fid(o)),
760 info->mti_big_aclsize);
762 buf->lb_buf = info->mti_big_acl;
763 buf->lb_len = info->mti_big_aclsize;
764 info->mti_big_acl_used = 1;
767 /* FS has ACL bigger that our limits */
768 CDEBUG(D_INODE, "%s: "DFID" ACL can't fit into %d\n",
769 mdt_obd_name(mdt), PFID(mdt_object_fid(o)),
770 info->mti_big_aclsize);
773 CERROR("%s: unable to read "DFID" ACL: rc = %d\n",
774 mdt_obd_name(mdt), PFID(mdt_object_fid(o)), rc);
777 rc = nodemap_map_acl(nodemap, buf->lb_buf,
778 rc, NODEMAP_FS_TO_CLIENT);
779 /* if all ACLs mapped out, rc is still >= 0 */
781 CERROR("%s: nodemap_map_acl unable to parse "DFID
782 " ACL: rc = %d\n", mdt_obd_name(mdt),
783 PFID(mdt_object_fid(o)), rc);
784 repbody->mbo_aclsize = 0;
785 repbody->mbo_valid &= ~OBD_MD_FLACL;
787 repbody->mbo_aclsize = rc;
788 repbody->mbo_valid |= OBD_MD_FLACL;
797 /* XXX Look into layout in MDT layer. */
798 static inline bool mdt_hsm_is_released(struct lov_mds_md *lmm)
800 struct lov_comp_md_v1 *comp_v1;
801 struct lov_mds_md *v1;
804 if (lmm->lmm_magic == LOV_MAGIC_COMP_V1) {
805 comp_v1 = (struct lov_comp_md_v1 *)lmm;
807 for (i = 0; i < comp_v1->lcm_entry_count; i++) {
808 v1 = (struct lov_mds_md *)((char *)comp_v1 +
809 comp_v1->lcm_entries[i].lcme_offset);
810 /* We don't support partial release for now */
811 if (!(v1->lmm_pattern & LOV_PATTERN_F_RELEASED))
816 return (lmm->lmm_pattern & LOV_PATTERN_F_RELEASED) ?
821 void mdt_pack_attr2body(struct mdt_thread_info *info, struct mdt_body *b,
822 const struct lu_attr *attr, const struct lu_fid *fid)
824 struct mdt_device *mdt = info->mti_mdt;
825 struct obd_export *exp = info->mti_exp;
826 struct md_attr *ma = &info->mti_attr;
827 struct lu_nodemap *nodemap = NULL;
829 LASSERT(ma->ma_valid & MA_INODE);
831 if (attr->la_valid & LA_ATIME) {
832 b->mbo_atime = attr->la_atime;
833 b->mbo_valid |= OBD_MD_FLATIME;
835 if (attr->la_valid & LA_MTIME) {
836 b->mbo_mtime = attr->la_mtime;
837 b->mbo_valid |= OBD_MD_FLMTIME;
839 if (attr->la_valid & LA_CTIME) {
840 b->mbo_ctime = attr->la_ctime;
841 b->mbo_valid |= OBD_MD_FLCTIME;
843 if (attr->la_valid & LA_BTIME) {
844 b->mbo_btime = attr->la_btime;
845 b->mbo_valid |= OBD_MD_FLBTIME;
847 if (attr->la_valid & LA_FLAGS) {
848 b->mbo_flags = attr->la_flags;
849 b->mbo_valid |= OBD_MD_FLFLAGS;
851 if (attr->la_valid & LA_NLINK) {
852 b->mbo_nlink = attr->la_nlink;
853 b->mbo_valid |= OBD_MD_FLNLINK;
855 if (attr->la_valid & (LA_UID|LA_GID|LA_PROJID)) {
856 nodemap = nodemap_get_from_exp(exp);
860 if (attr->la_valid & LA_UID) {
861 b->mbo_uid = nodemap_map_id(nodemap, NODEMAP_UID,
862 NODEMAP_FS_TO_CLIENT,
864 b->mbo_valid |= OBD_MD_FLUID;
866 if (attr->la_valid & LA_GID) {
867 b->mbo_gid = nodemap_map_id(nodemap, NODEMAP_GID,
868 NODEMAP_FS_TO_CLIENT,
870 b->mbo_valid |= OBD_MD_FLGID;
873 if (attr->la_valid & LA_PROJID) {
874 b->mbo_projid = nodemap_map_id(nodemap, NODEMAP_PROJID,
875 NODEMAP_FS_TO_CLIENT,
877 b->mbo_valid |= OBD_MD_FLPROJID;
880 b->mbo_mode = attr->la_mode;
881 if (attr->la_valid & LA_MODE)
882 b->mbo_valid |= OBD_MD_FLMODE;
883 if (attr->la_valid & LA_TYPE)
884 b->mbo_valid |= OBD_MD_FLTYPE;
888 b->mbo_valid |= OBD_MD_FLID;
889 CDEBUG(D_INODE, DFID": nlink=%d, mode=%o, valid=%#llx\n",
890 PFID(fid), b->mbo_nlink, b->mbo_mode, b->mbo_valid);
893 if (!(attr->la_valid & LA_TYPE))
896 b->mbo_rdev = attr->la_rdev;
897 b->mbo_size = attr->la_size;
898 b->mbo_blocks = attr->la_blocks;
900 if (!S_ISREG(attr->la_mode)) {
901 b->mbo_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS | OBD_MD_FLRDEV;
902 } else if (ma->ma_need & MA_LOV && !(ma->ma_valid & MA_LOV)) {
903 /* means no objects are allocated on osts. */
904 LASSERT(!(ma->ma_valid & MA_LOV));
905 /* just ignore blocks occupied by extend attributes on MDS */
907 /* if no object is allocated on osts, the size on mds is valid.
909 b->mbo_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
910 } else if ((ma->ma_valid & MA_LOV) && ma->ma_lmm != NULL) {
911 if (mdt_hsm_is_released(ma->ma_lmm)) {
912 /* A released file stores its size on MDS. */
913 /* But return 1 block for released file, unless tools
914 * like tar will consider it fully sparse. (LU-3864)
916 if (unlikely(b->mbo_size == 0))
920 b->mbo_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
921 } else if (info->mti_som_strict &&
922 mdt->mdt_opts.mo_enable_strict_som) {
923 /* use SOM for size*/
924 b->mbo_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
925 } else if (ma->ma_valid & MA_SOM) { /* lsom is valid */
926 b->mbo_valid |= OBD_MD_FLLAZYSIZE | OBD_MD_FLLAZYBLOCKS;
927 b->mbo_size = ma->ma_som.ms_size;
928 b->mbo_blocks = ma->ma_som.ms_blocks;
932 if (fid != NULL && (b->mbo_valid & OBD_MD_FLSIZE ||
933 b->mbo_valid & OBD_MD_FLLAZYSIZE))
934 CDEBUG(D_VFSTRACE, DFID": returning size %llu\n",
935 PFID(fid), (unsigned long long)b->mbo_size);
938 if (!IS_ERR_OR_NULL(nodemap))
939 nodemap_putref(nodemap);
942 static inline int mdt_body_has_lov(const struct lu_attr *la,
943 const struct mdt_body *body)
945 return (S_ISREG(la->la_mode) && (body->mbo_valid & OBD_MD_FLEASIZE)) ||
946 (S_ISDIR(la->la_mode) && (body->mbo_valid & OBD_MD_FLDIREA));
949 void mdt_client_compatibility(struct mdt_thread_info *info)
951 struct mdt_body *body;
952 struct ptlrpc_request *req = mdt_info_req(info);
953 struct obd_export *exp = req->rq_export;
954 struct md_attr *ma = &info->mti_attr;
955 struct lu_attr *la = &ma->ma_attr;
958 if (exp_connect_layout(exp))
959 /* the client can deal with 16-bit lmm_stripe_count */
962 body = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
964 if (!mdt_body_has_lov(la, body))
967 /* now we have a reply with a lov for a client not compatible with the
968 * layout lock so we have to clean the layout generation number */
969 if (S_ISREG(la->la_mode))
970 ma->ma_lmm->lmm_layout_gen = 0;
974 static int mdt_attr_get_eabuf_size(struct mdt_thread_info *info,
975 struct mdt_object *o)
977 const struct lu_env *env = info->mti_env;
980 rc = mo_xattr_get(env, mdt_object_child(o), &LU_BUF_NULL,
989 /* Is it a directory? Let's check for the LMV as well */
990 if (S_ISDIR(lu_object_attr(&mdt_object_child(o)->mo_lu))) {
991 rc2 = mo_xattr_get(env, mdt_object_child(o), &LU_BUF_NULL,
995 rc2 = mo_xattr_get(env, mdt_object_child(o),
997 XATTR_NAME_DEFAULT_LMV);
999 if ((rc2 < 0 && rc2 != -ENODATA) || (rc2 > rc))
1007 int mdt_big_xattr_get(struct mdt_thread_info *info, struct mdt_object *o,
1010 const struct lu_env *env = info->mti_env;
1014 LASSERT(info->mti_big_lmm_used == 0);
1015 rc = mo_xattr_get(env, mdt_object_child(o), &LU_BUF_NULL, name);
1019 /* big_lmm may need to be grown */
1020 if (info->mti_big_lmmsize < rc) {
1021 int size = size_roundup_power2(rc);
1023 if (info->mti_big_lmmsize > 0) {
1024 /* free old buffer */
1025 LASSERT(info->mti_big_lmm);
1026 OBD_FREE_LARGE(info->mti_big_lmm,
1027 info->mti_big_lmmsize);
1028 info->mti_big_lmm = NULL;
1029 info->mti_big_lmmsize = 0;
1032 OBD_ALLOC_LARGE(info->mti_big_lmm, size);
1033 if (info->mti_big_lmm == NULL)
1035 info->mti_big_lmmsize = size;
1037 LASSERT(info->mti_big_lmmsize >= rc);
1039 info->mti_buf.lb_buf = info->mti_big_lmm;
1040 info->mti_buf.lb_len = info->mti_big_lmmsize;
1041 rc = mo_xattr_get(env, mdt_object_child(o), &info->mti_buf, name);
1046 int __mdt_stripe_get(struct mdt_thread_info *info, struct mdt_object *o,
1047 struct md_attr *ma, const char *name)
1049 struct md_object *next = mdt_object_child(o);
1050 struct lu_buf *buf = &info->mti_buf;
1053 if (strcmp(name, XATTR_NAME_LOV) == 0) {
1054 buf->lb_buf = ma->ma_lmm;
1055 buf->lb_len = ma->ma_lmm_size;
1056 LASSERT(!(ma->ma_valid & MA_LOV));
1057 } else if (strcmp(name, XATTR_NAME_LMV) == 0) {
1058 buf->lb_buf = ma->ma_lmv;
1059 buf->lb_len = ma->ma_lmv_size;
1060 LASSERT(!(ma->ma_valid & MA_LMV));
1061 } else if (strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0) {
1062 buf->lb_buf = ma->ma_default_lmv;
1063 buf->lb_len = ma->ma_default_lmv_size;
1064 LASSERT(!(ma->ma_valid & MA_LMV_DEF));
1069 LASSERT(buf->lb_buf);
1071 if (!mdt_object_exists(o))
1074 if (mdt_object_remote(o) && S_ISDIR(lu_object_attr(&o->mot_obj)))
1075 /* force reload layout for remote dir in case layout changed */
1076 mo_invalidate(info->mti_env, mdt_object_child(o));
1078 rc = mo_xattr_get(info->mti_env, next, buf, name);
1082 if (strcmp(name, XATTR_NAME_LOV) == 0) {
1083 if (info->mti_big_lmm_used)
1084 ma->ma_lmm = info->mti_big_lmm;
1086 /* NOT return LOV EA with hole to old client. */
1087 if (unlikely(le32_to_cpu(ma->ma_lmm->lmm_pattern) &
1088 LOV_PATTERN_F_HOLE) &&
1089 !(exp_connect_flags(info->mti_exp) &
1090 OBD_CONNECT_LFSCK)) {
1093 ma->ma_lmm_size = rc;
1094 ma->ma_valid |= MA_LOV;
1096 } else if (strcmp(name, XATTR_NAME_LMV) == 0) {
1097 if (info->mti_big_lmm_used)
1098 ma->ma_lmv = info->mti_big_lmm;
1100 ma->ma_lmv_size = rc;
1101 ma->ma_valid |= MA_LMV;
1102 } else if (strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0) {
1103 ma->ma_default_lmv_size = rc;
1104 ma->ma_valid |= MA_LMV_DEF;
1107 /* Update mdt_max_mdsize so all clients will be aware that */
1108 if (info->mti_mdt->mdt_max_mdsize < rc)
1109 info->mti_mdt->mdt_max_mdsize = rc;
1112 } else if (rc == -ENODATA) {
1115 } else if (rc == -ERANGE) {
1116 /* Default LMV has fixed size, so it must be able to fit
1117 * in the original buffer */
1118 if (strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0)
1120 rc = mdt_big_xattr_get(info, o, name);
1122 info->mti_big_lmm_used = 1;
1130 int mdt_stripe_get(struct mdt_thread_info *info, struct mdt_object *o,
1131 struct md_attr *ma, const char *name)
1135 if (!info->mti_big_lmm) {
1136 OBD_ALLOC(info->mti_big_lmm, PAGE_SIZE);
1137 if (!info->mti_big_lmm)
1139 info->mti_big_lmmsize = PAGE_SIZE;
1142 if (strcmp(name, XATTR_NAME_LOV) == 0) {
1143 ma->ma_lmm = info->mti_big_lmm;
1144 ma->ma_lmm_size = info->mti_big_lmmsize;
1145 ma->ma_valid &= ~MA_LOV;
1146 } else if (strcmp(name, XATTR_NAME_LMV) == 0) {
1147 ma->ma_lmv = info->mti_big_lmm;
1148 ma->ma_lmv_size = info->mti_big_lmmsize;
1149 ma->ma_valid &= ~MA_LMV;
1154 LASSERT(!info->mti_big_lmm_used);
1155 rc = __mdt_stripe_get(info, o, ma, name);
1156 /* since big_lmm is always used here, clear 'used' flag to avoid
1157 * assertion in mdt_big_xattr_get().
1159 info->mti_big_lmm_used = 0;
1164 int mdt_attr_get_pfid(struct mdt_thread_info *info, struct mdt_object *o,
1165 struct lu_fid *pfid)
1167 struct lu_buf *buf = &info->mti_buf;
1168 struct link_ea_header *leh;
1169 struct link_ea_entry *lee;
1173 buf->lb_buf = info->mti_big_lmm;
1174 buf->lb_len = info->mti_big_lmmsize;
1175 rc = mo_xattr_get(info->mti_env, mdt_object_child(o),
1176 buf, XATTR_NAME_LINK);
1177 /* ignore errors, MA_PFID won't be set and it is
1178 * up to the caller to treat this as an error */
1179 if (rc == -ERANGE || buf->lb_len == 0) {
1180 rc = mdt_big_xattr_get(info, o, XATTR_NAME_LINK);
1181 buf->lb_buf = info->mti_big_lmm;
1182 buf->lb_len = info->mti_big_lmmsize;
1187 if (rc < sizeof(*leh)) {
1188 CERROR("short LinkEA on "DFID": rc = %d\n",
1189 PFID(mdt_object_fid(o)), rc);
1193 leh = (struct link_ea_header *) buf->lb_buf;
1194 lee = (struct link_ea_entry *)(leh + 1);
1195 if (leh->leh_magic == __swab32(LINK_EA_MAGIC)) {
1196 leh->leh_magic = LINK_EA_MAGIC;
1197 leh->leh_reccount = __swab32(leh->leh_reccount);
1198 leh->leh_len = __swab64(leh->leh_len);
1200 if (leh->leh_magic != LINK_EA_MAGIC)
1202 if (leh->leh_reccount == 0)
1205 memcpy(pfid, &lee->lee_parent_fid, sizeof(*pfid));
1206 fid_be_to_cpu(pfid, pfid);
1211 int mdt_attr_get_pfid_name(struct mdt_thread_info *info, struct mdt_object *o,
1212 struct lu_fid *pfid, struct lu_name *lname)
1214 struct lu_buf *buf = &info->mti_buf;
1215 struct link_ea_header *leh;
1216 struct link_ea_entry *lee;
1220 buf->lb_buf = info->mti_xattr_buf;
1221 buf->lb_len = sizeof(info->mti_xattr_buf);
1222 rc = mo_xattr_get(info->mti_env, mdt_object_child(o), buf,
1224 if (rc == -ERANGE) {
1225 rc = mdt_big_xattr_get(info, o, XATTR_NAME_LINK);
1226 buf->lb_buf = info->mti_big_lmm;
1227 buf->lb_len = info->mti_big_lmmsize;
1232 if (rc < sizeof(*leh)) {
1233 CERROR("short LinkEA on "DFID": rc = %d\n",
1234 PFID(mdt_object_fid(o)), rc);
1238 leh = (struct link_ea_header *)buf->lb_buf;
1239 lee = (struct link_ea_entry *)(leh + 1);
1240 if (leh->leh_magic == __swab32(LINK_EA_MAGIC)) {
1241 leh->leh_magic = LINK_EA_MAGIC;
1242 leh->leh_reccount = __swab32(leh->leh_reccount);
1243 leh->leh_len = __swab64(leh->leh_len);
1245 if (leh->leh_magic != LINK_EA_MAGIC)
1248 if (leh->leh_reccount == 0)
1251 linkea_entry_unpack(lee, &reclen, lname, pfid);
1256 int mdt_attr_get_complex(struct mdt_thread_info *info,
1257 struct mdt_object *o, struct md_attr *ma)
1259 const struct lu_env *env = info->mti_env;
1260 struct md_object *next = mdt_object_child(o);
1261 struct lu_buf *buf = &info->mti_buf;
1262 int need = ma->ma_need;
1269 if (mdt_object_exists(o) == 0)
1270 GOTO(out, rc = -ENOENT);
1271 mode = lu_object_attr(&next->mo_lu);
1273 if (need & MA_INODE) {
1274 ma->ma_need = MA_INODE;
1275 rc = mo_attr_get(env, next, ma);
1280 (void) mdt_get_som(info, o, ma);
1281 ma->ma_valid |= MA_INODE;
1284 if (need & MA_PFID) {
1285 rc = mdt_attr_get_pfid(info, o, &ma->ma_pfid);
1287 ma->ma_valid |= MA_PFID;
1288 /* ignore this error, parent fid is not mandatory */
1292 if (need & MA_LOV && (S_ISREG(mode) || S_ISDIR(mode))) {
1293 rc = __mdt_stripe_get(info, o, ma, XATTR_NAME_LOV);
1298 if (need & MA_LMV && S_ISDIR(mode)) {
1299 rc = __mdt_stripe_get(info, o, ma, XATTR_NAME_LMV);
1304 if (need & MA_LMV_DEF && S_ISDIR(mode)) {
1305 rc = __mdt_stripe_get(info, o, ma, XATTR_NAME_DEFAULT_LMV);
1311 * In the handle of MA_INODE, we may already get the SOM attr.
1313 if (need & MA_SOM && S_ISREG(mode) && !(ma->ma_valid & MA_SOM)) {
1314 rc = mdt_get_som(info, o, ma);
1319 if (need & MA_HSM && S_ISREG(mode)) {
1320 buf->lb_buf = info->mti_xattr_buf;
1321 buf->lb_len = sizeof(info->mti_xattr_buf);
1322 BUILD_BUG_ON(sizeof(struct hsm_attrs) >
1323 sizeof(info->mti_xattr_buf));
1324 rc2 = mo_xattr_get(info->mti_env, next, buf, XATTR_NAME_HSM);
1325 rc2 = lustre_buf2hsm(info->mti_xattr_buf, rc2, &ma->ma_hsm);
1327 ma->ma_valid |= MA_HSM;
1328 else if (rc2 < 0 && rc2 != -ENODATA)
1329 GOTO(out, rc = rc2);
1332 #ifdef CONFIG_LUSTRE_FS_POSIX_ACL
1333 if (need & MA_ACL_DEF && S_ISDIR(mode)) {
1334 buf->lb_buf = ma->ma_acl;
1335 buf->lb_len = ma->ma_acl_size;
1336 rc2 = mo_xattr_get(env, next, buf, XATTR_NAME_ACL_DEFAULT);
1338 ma->ma_acl_size = rc2;
1339 ma->ma_valid |= MA_ACL_DEF;
1340 } else if (rc2 == -ENODATA) {
1342 ma->ma_acl_size = 0;
1344 GOTO(out, rc = rc2);
1349 CDEBUG(D_INODE, "after getattr rc = %d, ma_valid = %#llx ma_lmm=%p\n",
1350 rc, ma->ma_valid, ma->ma_lmm);
1354 static void mdt_preset_encctx_size(struct mdt_thread_info *info)
1356 struct req_capsule *pill = info->mti_pill;
1359 if (req_capsule_has_field(pill, &RMF_FILE_ENCCTX,
1361 /* pre-set size in server part with max size */
1362 req_capsule_set_size(pill, &RMF_FILE_ENCCTX,
1364 info->mti_mdt->mdt_max_mdsize);
1368 static int mdt_getattr_internal(struct mdt_thread_info *info,
1369 struct mdt_object *o, int ma_need)
1371 struct mdt_device *mdt = info->mti_mdt;
1372 struct md_object *next = mdt_object_child(o);
1373 const struct mdt_body *reqbody = info->mti_body;
1374 struct ptlrpc_request *req = mdt_info_req(info);
1375 struct md_attr *ma = &info->mti_attr;
1376 struct lu_attr *la = &ma->ma_attr;
1377 struct req_capsule *pill = info->mti_pill;
1378 const struct lu_env *env = info->mti_env;
1379 struct mdt_body *repbody;
1380 struct lu_buf *buffer = &info->mti_buf;
1381 struct obd_export *exp = info->mti_exp;
1382 ktime_t kstart = ktime_get();
1387 if (CFS_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK))
1388 RETURN(err_serious(-ENOMEM));
1390 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
1394 if (mdt_object_remote(o)) {
1395 /* This object is located on remote node.*/
1396 /* Return -ENOTSUPP for old client */
1397 if (!mdt_is_dne_client(req->rq_export))
1398 GOTO(out, rc = -ENOTSUPP);
1400 repbody->mbo_fid1 = *mdt_object_fid(o);
1401 repbody->mbo_valid = OBD_MD_FLID | OBD_MD_MDS;
1405 if (reqbody->mbo_eadatasize > 0) {
1406 buffer->lb_buf = req_capsule_server_get(pill, &RMF_MDT_MD);
1407 if (buffer->lb_buf == NULL)
1408 GOTO(out, rc = -EPROTO);
1409 buffer->lb_len = req_capsule_get_size(pill, &RMF_MDT_MD,
1412 buffer->lb_buf = NULL;
1414 ma_need &= ~(MA_LOV | MA_LMV);
1415 CDEBUG(D_INFO, "%s: RPC from %s: does not need LOVEA.\n",
1416 mdt_obd_name(info->mti_mdt),
1417 req->rq_export->exp_client_uuid.uuid);
1420 /* from 2.12.58 intent_getattr pack default LMV in reply */
1421 if (S_ISDIR(lu_object_attr(&next->mo_lu)) &&
1422 ((reqbody->mbo_valid & (OBD_MD_MEA | OBD_MD_DEFAULT_MEA)) ==
1423 (OBD_MD_MEA | OBD_MD_DEFAULT_MEA)) &&
1424 req_capsule_has_field(&req->rq_pill, &RMF_DEFAULT_MDT_MD,
1426 ma->ma_lmv = buffer->lb_buf;
1427 ma->ma_lmv_size = buffer->lb_len;
1428 ma->ma_default_lmv = req_capsule_server_get(pill,
1429 &RMF_DEFAULT_MDT_MD);
1430 ma->ma_default_lmv_size = req_capsule_get_size(pill,
1431 &RMF_DEFAULT_MDT_MD,
1433 ma->ma_need = MA_INODE;
1434 if (ma->ma_lmv_size > 0)
1435 ma->ma_need |= MA_LMV;
1436 if (ma->ma_default_lmv_size > 0)
1437 ma->ma_need |= MA_LMV_DEF;
1438 } else if (S_ISDIR(lu_object_attr(&next->mo_lu)) &&
1439 (reqbody->mbo_valid & (OBD_MD_MEA | OBD_MD_DEFAULT_MEA))) {
1440 /* If it is dir and client require MEA, then we got MEA */
1441 /* Assumption: MDT_MD size is enough for lmv size. */
1442 ma->ma_lmv = buffer->lb_buf;
1443 ma->ma_lmv_size = buffer->lb_len;
1444 ma->ma_need = MA_INODE;
1445 if (ma->ma_lmv_size > 0) {
1446 if (reqbody->mbo_valid & OBD_MD_MEA) {
1447 ma->ma_need |= MA_LMV;
1448 } else if (reqbody->mbo_valid & OBD_MD_DEFAULT_MEA) {
1449 ma->ma_need |= MA_LMV_DEF;
1450 ma->ma_default_lmv = buffer->lb_buf;
1452 ma->ma_default_lmv_size = buffer->lb_len;
1453 ma->ma_lmv_size = 0;
1457 ma->ma_lmm = buffer->lb_buf;
1458 ma->ma_lmm_size = buffer->lb_len;
1459 ma->ma_need = MA_INODE | MA_HSM;
1460 if (ma->ma_lmm_size > 0) {
1461 ma->ma_need |= MA_LOV;
1462 /* Older clients may crash if they getattr overstriped
1465 if (!exp_connect_overstriping(exp) &&
1466 mdt_lmm_is_overstriping(ma->ma_lmm))
1467 RETURN(-EOPNOTSUPP);
1471 if (S_ISDIR(lu_object_attr(&next->mo_lu)) &&
1472 reqbody->mbo_valid & OBD_MD_FLDIREA &&
1473 lustre_msg_get_opc(req->rq_reqmsg) == MDS_GETATTR) {
1474 /* get default stripe info for this dir. */
1475 ma->ma_need |= MA_LOV_DEF;
1477 ma->ma_need |= ma_need;
1479 rc = mdt_attr_get_complex(info, o, ma);
1481 CDEBUG_LIMIT(rc == -ENOENT ? D_OTHER : D_ERROR,
1482 "%s: getattr error for "DFID": rc = %d\n",
1483 mdt_obd_name(info->mti_mdt),
1484 PFID(mdt_object_fid(o)), rc);
1488 /* return immutable attr on fscrypt metadata files
1489 * if fscrypt admin is not permitted
1491 if (o->mot_obj.lo_header->loh_attr & LOHA_FSCRYPT_MD &&
1492 !mdt_ucred(info)->uc_rbac_fscrypt_admin)
1493 la->la_flags |= LUSTRE_IMMUTABLE_FL;
1495 /* if file is released, check if a restore is running */
1496 if (ma->ma_valid & MA_HSM) {
1497 repbody->mbo_valid |= OBD_MD_TSTATE;
1498 if ((ma->ma_hsm.mh_flags & HS_RELEASED) &&
1499 mdt_hsm_restore_is_running(info, mdt_object_fid(o)))
1500 repbody->mbo_t_state = MS_RESTORE;
1503 if (unlikely(!(ma->ma_valid & MA_INODE)))
1506 mdt_pack_attr2body(info, repbody, la, mdt_object_fid(o));
1508 if (mdt_body_has_lov(la, reqbody)) {
1509 u32 stripe_count = 1;
1510 bool fixed_layout = false;
1512 if (ma->ma_valid & MA_LOV) {
1513 LASSERT(ma->ma_lmm_size);
1514 repbody->mbo_eadatasize = ma->ma_lmm_size;
1515 if (S_ISDIR(la->la_mode))
1516 repbody->mbo_valid |= OBD_MD_FLDIREA;
1518 repbody->mbo_valid |= OBD_MD_FLEASIZE;
1519 mdt_dump_lmm(D_INFO, ma->ma_lmm, repbody->mbo_valid);
1521 if (ma->ma_valid & MA_LMV) {
1522 struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1523 u32 magic = le32_to_cpu(lmv->lmv_magic);
1525 /* Return -ENOTSUPP for old client */
1526 if (!mdt_is_striped_client(req->rq_export))
1529 LASSERT(S_ISDIR(la->la_mode));
1530 mdt_dump_lmv(D_INFO, ma->ma_lmv);
1531 repbody->mbo_eadatasize = ma->ma_lmv_size;
1532 repbody->mbo_valid |= (OBD_MD_FLDIREA|OBD_MD_MEA);
1534 stripe_count = le32_to_cpu(lmv->lmv_stripe_count);
1535 fixed_layout = lmv_is_fixed(lmv);
1536 if (magic == LMV_MAGIC_STRIPE && lmv_is_restriping(lmv))
1537 mdt_restripe_migrate_add(info, o);
1538 else if (magic == LMV_MAGIC_V1 &&
1539 lmv_is_restriping(lmv))
1540 mdt_restripe_update_add(info, o);
1542 if (ma->ma_valid & MA_LMV_DEF) {
1543 /* Return -ENOTSUPP for old client */
1544 if (!mdt_is_striped_client(req->rq_export))
1546 LASSERT(S_ISDIR(la->la_mode));
1548 * when ll_dir_getstripe() gets default LMV, it
1549 * checks mbo_eadatasize.
1551 if (!(ma->ma_valid & MA_LMV))
1552 repbody->mbo_eadatasize =
1553 ma->ma_default_lmv_size;
1554 repbody->mbo_valid |= (OBD_MD_FLDIREA |
1555 OBD_MD_DEFAULT_MEA);
1558 "dirent count %llu stripe count %u MDT count %d\n",
1559 ma->ma_attr.la_dirent_count, stripe_count,
1560 atomic_read(&mdt->mdt_mds_mds_conns) + 1);
1561 if (ma->ma_attr.la_dirent_count != LU_DIRENT_COUNT_UNSET &&
1562 ma->ma_attr.la_dirent_count >
1563 mdt->mdt_restriper.mdr_dir_split_count &&
1564 !fid_is_root(mdt_object_fid(o)) &&
1565 mdt->mdt_enable_dir_auto_split &&
1566 !o->mot_restriping &&
1567 stripe_count < atomic_read(&mdt->mdt_mds_mds_conns) + 1 &&
1569 mdt_auto_split_add(info, o);
1570 } else if (S_ISLNK(la->la_mode) &&
1571 reqbody->mbo_valid & OBD_MD_LINKNAME) {
1572 buffer->lb_buf = ma->ma_lmm;
1573 /* eadatasize from client includes NULL-terminator, so
1574 * there is no need to read it */
1575 buffer->lb_len = reqbody->mbo_eadatasize - 1;
1576 rc = mo_readlink(env, next, buffer);
1577 if (unlikely(rc <= 0)) {
1578 CERROR("%s: readlink failed for "DFID": rc = %d\n",
1579 mdt_obd_name(info->mti_mdt),
1580 PFID(mdt_object_fid(o)), rc);
1583 int print_limit = min_t(int, PAGE_SIZE - 128, rc);
1585 if (CFS_FAIL_CHECK(OBD_FAIL_MDS_READLINK_EPROTO))
1587 repbody->mbo_valid |= OBD_MD_LINKNAME;
1588 /* we need to report back size with NULL-terminator
1589 * because client expects that */
1590 repbody->mbo_eadatasize = rc + 1;
1591 if (repbody->mbo_eadatasize != reqbody->mbo_eadatasize)
1592 CDEBUG(D_INODE, "%s: Read shorter symlink %d "
1593 "on "DFID ", expected %d\n",
1594 mdt_obd_name(info->mti_mdt),
1595 rc, PFID(mdt_object_fid(o)),
1596 reqbody->mbo_eadatasize - 1);
1597 /* NULL terminate */
1598 ((char *)ma->ma_lmm)[rc] = 0;
1600 /* If the total CDEBUG() size is larger than a page, it
1601 * will print a warning to the console, avoid this by
1602 * printing just the last part of the symlink. */
1603 CDEBUG(D_INODE, "symlink dest %s%.*s, len = %d\n",
1604 print_limit < rc ? "..." : "", print_limit,
1605 (char *)ma->ma_lmm + rc - print_limit, rc);
1610 if (reqbody->mbo_valid & OBD_MD_FLMODEASIZE) {
1611 repbody->mbo_max_mdsize = info->mti_mdt->mdt_max_mdsize;
1612 repbody->mbo_valid |= OBD_MD_FLMODEASIZE;
1613 CDEBUG(D_INODE, "changing the max MD size to %u\n",
1614 repbody->mbo_max_mdsize);
1617 #ifdef CONFIG_LUSTRE_FS_POSIX_ACL
1618 if ((exp_connect_flags(req->rq_export) & OBD_CONNECT_ACL) &&
1619 (reqbody->mbo_valid & OBD_MD_FLACL)) {
1620 struct lu_nodemap *nodemap = nodemap_get_from_exp(exp);
1621 if (IS_ERR(nodemap))
1622 RETURN(PTR_ERR(nodemap));
1624 rc = mdt_pack_acl2body(info, repbody, o, nodemap);
1625 nodemap_putref(nodemap);
1631 mdt_counter_incr(req, LPROC_MDT_GETATTR,
1632 ktime_us_delta(ktime_get(), kstart));
1637 static int mdt_getattr(struct tgt_session_info *tsi)
1639 struct mdt_thread_info *info = tsi2mdt_info(tsi);
1640 struct mdt_object *obj = info->mti_object;
1641 struct req_capsule *pill = info->mti_pill;
1642 struct mdt_body *reqbody;
1643 struct mdt_body *repbody;
1647 if (unlikely(info->mti_object == NULL))
1650 reqbody = req_capsule_client_get(pill, &RMF_MDT_BODY);
1652 LASSERT(lu_object_assert_exists(&obj->mot_obj));
1654 /* Special case for Data-on-MDT files to get data version */
1655 if (unlikely(reqbody->mbo_valid & OBD_MD_FLDATAVERSION)) {
1656 rc = mdt_data_version_get(tsi);
1660 /* Unlike intent case where we need to pre-fill out buffers early on
1661 * in intent policy for ldlm reasons, here we can have a much better
1662 * guess at EA size by just reading it from disk.
1663 * Exceptions are readdir and (missing) directory striping */
1665 if (reqbody->mbo_valid & OBD_MD_LINKNAME) {
1666 /* No easy way to know how long is the symlink, but it cannot
1667 * be more than PATH_MAX, so we allocate +1 */
1669 /* A special case for fs ROOT: getattr there might fetch
1670 * default EA for entire fs, not just for this dir!
1672 } else if (lu_fid_eq(mdt_object_fid(obj),
1673 &info->mti_mdt->mdt_md_root_fid) &&
1674 (reqbody->mbo_valid & OBD_MD_FLDIREA) &&
1675 (lustre_msg_get_opc(mdt_info_req(info)->rq_reqmsg) ==
1677 /* Should the default strping be bigger, mdt_fix_reply
1678 * will reallocate */
1679 rc = DEF_REP_MD_SIZE;
1681 /* Read the actual EA size from disk */
1682 rc = mdt_attr_get_eabuf_size(info, obj);
1686 GOTO(out, rc = err_serious(rc));
1688 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER, rc);
1690 /* Set ACL reply buffer size as LUSTRE_POSIX_ACL_MAX_SIZE_OLD
1691 * by default. If the target object has more ACL entries, then
1692 * enlarge the buffer when necessary. */
1693 req_capsule_set_size(pill, &RMF_ACL, RCL_SERVER,
1694 LUSTRE_POSIX_ACL_MAX_SIZE_OLD);
1695 mdt_preset_encctx_size(info);
1697 rc = req_capsule_server_pack(pill);
1698 if (unlikely(rc != 0))
1699 GOTO(out, rc = err_serious(rc));
1701 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
1702 LASSERT(repbody != NULL);
1703 repbody->mbo_eadatasize = 0;
1704 repbody->mbo_aclsize = 0;
1706 rc = mdt_check_ucred(info);
1708 GOTO(out_shrink, rc);
1710 info->mti_cross_ref = !!(reqbody->mbo_valid & OBD_MD_FLCROSSREF);
1712 rc = mdt_init_ucred(info, reqbody);
1714 GOTO(out_shrink, rc);
1716 rc = mdt_getattr_internal(info, obj, 0);
1718 GOTO(out_ucred, rc);
1720 rc = mdt_pack_encctx_in_reply(info, obj);
1723 mdt_exit_ucred(info);
1725 mdt_client_compatibility(info);
1726 rc2 = mdt_fix_reply(info);
1730 mdt_thread_info_fini(info);
1735 * Handler of layout intent RPC requiring the layout modification
1737 * \param[in] info thread environment
1738 * \param[in] obj object
1739 * \param[out] lhc object ldlm lock handle
1740 * \param[in] layout layout change descriptor
1742 * \retval 0 on success
1743 * \retval < 0 error code
1745 int mdt_layout_change(struct mdt_thread_info *info, struct mdt_object *obj,
1746 struct mdt_lock_handle *lhc,
1747 struct md_layout_change *layout)
1753 if (!mdt_object_exists(obj))
1756 if (!S_ISREG(lu_object_attr(&obj->mot_obj)))
1759 rc = mo_permission(info->mti_env, NULL, mdt_object_child(obj), NULL,
1764 rc = mdt_check_resent_lock(info, obj, lhc);
1770 __u64 lockpart = MDS_INODELOCK_LAYOUT;
1772 /* take layout lock to prepare layout change */
1773 if (layout->mlc_opc == MD_LAYOUT_WRITE)
1774 lockpart |= MDS_INODELOCK_UPDATE;
1776 rc = mdt_object_lock(info, obj, lhc, lockpart, LCK_EX);
1781 mutex_lock(&obj->mot_som_mutex);
1782 rc = mo_layout_change(info->mti_env, mdt_object_child(obj), layout);
1783 mutex_unlock(&obj->mot_som_mutex);
1786 mdt_object_unlock(info, obj, lhc, 1);
1792 * Exchange MOF_LOV_CREATED flags between two objects after a
1793 * layout swap. No assumption is made on whether o1 or o2 have
1794 * created objects or not.
1796 * \param[in,out] o1 First swap layout object
1797 * \param[in,out] o2 Second swap layout object
1799 static void mdt_swap_lov_flag(struct mdt_object *o1, struct mdt_object *o2)
1801 unsigned int o1_lov_created = o1->mot_lov_created;
1803 mutex_lock(&o1->mot_lov_mutex);
1804 mutex_lock(&o2->mot_lov_mutex);
1806 o1->mot_lov_created = o2->mot_lov_created;
1807 o2->mot_lov_created = o1_lov_created;
1809 mutex_unlock(&o2->mot_lov_mutex);
1810 mutex_unlock(&o1->mot_lov_mutex);
1813 static int mdt_swap_layouts(struct tgt_session_info *tsi)
1815 struct mdt_thread_info *info;
1816 struct ptlrpc_request *req = tgt_ses_req(tsi);
1817 struct obd_export *exp = req->rq_export;
1818 struct mdt_object *o1, *o2, *o;
1819 struct mdt_lock_handle *lh1, *lh2;
1820 struct mdc_swap_layouts *msl;
1824 /* client does not support layout lock, so layout swaping
1826 * FIXME: there is a problem for old clients which don't support
1827 * layout lock yet. If those clients have already opened the file
1828 * they won't be notified at all so that old layout may still be
1829 * used to do IO. This can be fixed after file release is landed by
1830 * doing exclusive open and taking full EX ibits lock. - Jinshan */
1831 if (!exp_connect_layout(exp))
1832 RETURN(-EOPNOTSUPP);
1834 info = tsi2mdt_info(tsi);
1835 if (unlikely(info->mti_object == NULL))
1838 if (info->mti_dlm_req != NULL)
1839 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1841 o1 = info->mti_object;
1842 o = o2 = mdt_object_find(info->mti_env, info->mti_mdt,
1843 &info->mti_body->mbo_fid2);
1845 GOTO(out, rc = PTR_ERR(o));
1847 if (mdt_object_remote(o) || !mdt_object_exists(o)) /* remote object */
1848 GOTO(put, rc = -ENOENT);
1850 rc = lu_fid_cmp(&info->mti_body->mbo_fid1, &info->mti_body->mbo_fid2);
1851 if (unlikely(rc == 0)) /* same file, you kidding me? no-op. */
1857 /* permission check. Make sure the calling process having permission
1858 * to write both files. */
1859 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o1), NULL,
1864 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o2), NULL,
1869 msl = req_capsule_client_get(info->mti_pill, &RMF_SWAP_LAYOUTS);
1871 GOTO(put, rc = -EPROTO);
1873 lh1 = &info->mti_lh[MDT_LH_NEW];
1874 lh2 = &info->mti_lh[MDT_LH_OLD];
1875 rc = mdt_object_lock(info, o1, lh1, MDS_INODELOCK_LAYOUT |
1876 MDS_INODELOCK_XATTR, LCK_EX);
1880 rc = mdt_object_lock(info, o2, lh2, MDS_INODELOCK_LAYOUT |
1881 MDS_INODELOCK_XATTR, LCK_EX);
1885 rc = mo_swap_layouts(info->mti_env, mdt_object_child(o1),
1886 mdt_object_child(o2), msl->msl_flags);
1890 mdt_swap_lov_flag(o1, o2);
1893 mdt_object_unlock(info, o2, lh2, rc);
1895 mdt_object_unlock(info, o1, lh1, rc);
1897 mdt_object_put(info->mti_env, o);
1899 mdt_thread_info_fini(info);
1903 static int mdt_raw_lookup(struct mdt_thread_info *info,
1904 struct mdt_object *parent,
1905 const struct lu_name *lname)
1907 struct lu_fid *fid = &info->mti_tmp_fid1;
1908 struct mdt_body *repbody;
1909 bool is_dotdot = false;
1910 bool is_old_parent_stripe = false;
1911 bool is_new_parent_checked = false;
1916 LASSERT(!info->mti_cross_ref);
1917 /* Always allow to lookup ".." */
1918 if (lname->ln_namelen == 2 &&
1919 lname->ln_name[0] == '.' && lname->ln_name[1] == '.') {
1920 info->mti_spec.sp_permitted = 1;
1922 if (mdt_is_dir_stripe(info, parent) == 1)
1923 is_old_parent_stripe = true;
1926 mdt_object_get(info->mti_env, parent);
1928 /* Only got the fid of this obj by name */
1930 rc = mdo_lookup(info->mti_env, mdt_object_child(parent), lname, fid,
1932 mdt_object_put(info->mti_env, parent);
1936 /* getattr_name("..") should return master object FID for striped dir */
1937 if (is_dotdot && (is_old_parent_stripe || !is_new_parent_checked)) {
1938 parent = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1940 RETURN(PTR_ERR(parent));
1942 /* old client getattr_name("..") with stripe FID */
1943 if (unlikely(is_old_parent_stripe)) {
1944 is_old_parent_stripe = false;
1948 /* ".." may be a stripe */
1949 if (unlikely(mdt_is_dir_stripe(info, parent) == 1)) {
1950 is_new_parent_checked = true;
1954 mdt_object_put(info->mti_env, parent);
1957 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1958 repbody->mbo_fid1 = *fid;
1959 repbody->mbo_valid = OBD_MD_FLID;
1965 * Find name matching hash
1967 * We search \a child LinkEA for a name whose hash matches \a lname
1968 * (it contains an encoded hash).
1970 * \param info mdt thread info
1971 * \param lname encoded hash to find
1972 * \param parent parent object
1973 * \param child object to search with LinkEA
1975 * \retval 1 match found
1976 * \retval 0 no match found
1977 * \retval -ev negative errno upon error
1979 int find_name_matching_hash(struct mdt_thread_info *info, struct lu_name *lname,
1980 struct mdt_object *parent, struct mdt_object *child)
1982 /* Here, lname is an encoded hash of on-disk name, and
1983 * client is doing access without encryption key.
1984 * So we need to get LinkEA, check parent fid is correct and
1985 * compare name hash with the one in the request.
1987 struct lu_buf *buf = &info->mti_big_buf;
1988 struct lu_name name;
1990 struct linkea_data ldata = { NULL };
1991 struct link_ea_header *leh;
1992 struct link_ea_entry *lee;
1993 struct lu_buf link = { 0 };
1995 int reclen, count, rc;
1998 if (lname->ln_namelen < LL_CRYPTO_BLOCK_SIZE)
2001 buf = lu_buf_check_and_alloc(buf, PATH_MAX);
2006 rc = mdt_links_read(info, child, &ldata);
2010 hash = kmalloc(lname->ln_namelen, GFP_NOFS);
2013 rc = critical_decode(lname->ln_name, lname->ln_namelen, hash);
2016 lee = (struct link_ea_entry *)(leh + 1);
2017 for (count = 0; count < leh->leh_reccount; count++) {
2018 linkea_entry_unpack(lee, &reclen, &name, &pfid);
2019 if (!parent || lu_fid_eq(&pfid, mdt_object_fid(parent))) {
2020 lu_buf_check_and_alloc(&link, name.ln_namelen);
2022 GOTO(out_match, rc = -ENOMEM);
2023 rc = critical_decode(name.ln_name, name.ln_namelen,
2026 if (memcmp(LLCRYPT_EXTRACT_DIGEST(link.lb_buf, rc),
2027 hash, LL_CRYPTO_BLOCK_SIZE) == 0) {
2032 lee = (struct link_ea_entry *) ((char *)lee + reclen);
2034 if (count == leh->leh_reccount)
2047 * UPDATE lock should be taken against parent, and be released before exit;
2048 * child_bits lock should be taken against child, and be returned back:
2049 * (1)normal request should release the child lock;
2050 * (2)intent request will grant the lock to client.
2052 static int mdt_getattr_name_lock(struct mdt_thread_info *info,
2053 struct mdt_lock_handle *lhc,
2055 struct ldlm_reply *ldlm_rep)
2057 struct ptlrpc_request *req = mdt_info_req(info);
2058 struct mdt_body *reqbody = NULL;
2059 struct mdt_object *parent = info->mti_object;
2060 struct mdt_object *child = NULL;
2061 struct lu_fid *child_fid = &info->mti_tmp_fid1;
2062 struct lu_name *lname = NULL;
2063 struct mdt_lock_handle *lhp = NULL;
2064 struct ldlm_lock *lock;
2065 struct req_capsule *pill = info->mti_pill;
2066 bool fscrypt_md = false;
2074 is_resent = lustre_handle_is_used(&lhc->mlh_reg_lh);
2075 LASSERT(ergo(is_resent,
2076 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT));
2081 if (info->mti_cross_ref) {
2082 /* Only getattr on the child. Parent is on another node. */
2083 mdt_set_disposition(info, ldlm_rep,
2084 DISP_LOOKUP_EXECD | DISP_LOOKUP_POS);
2086 CDEBUG(D_INODE, "partial getattr_name child_fid = "DFID", "
2088 PFID(mdt_object_fid(child)), ldlm_rep);
2090 rc = mdt_check_resent_lock(info, child, lhc);
2093 } else if (rc > 0) {
2095 * Object's name entry is on another MDS, it will
2096 * request PERM lock only because LOOKUP lock is owned
2097 * by the MDS where name entry resides.
2099 * TODO: it should try layout lock too. - Jinshan
2101 child_bits &= ~(MDS_INODELOCK_LOOKUP |
2102 MDS_INODELOCK_LAYOUT);
2103 child_bits |= MDS_INODELOCK_PERM;
2104 rc = mdt_object_lock(info, child, lhc, child_bits,
2110 /* Finally, we can get attr for child. */
2111 if (!mdt_object_exists(child)) {
2112 LU_OBJECT_DEBUG(D_INFO, info->mti_env,
2114 "remote object doesn't exist.");
2115 mdt_object_unlock(info, child, lhc, 1);
2119 rc = mdt_getattr_internal(info, child, 0);
2120 if (unlikely(rc != 0)) {
2121 mdt_object_unlock(info, child, lhc, 1);
2125 rc = mdt_pack_secctx_in_reply(info, child);
2127 mdt_object_unlock(info, child, lhc, 1);
2131 rc = mdt_pack_encctx_in_reply(info, child);
2133 mdt_object_unlock(info, child, lhc, 1);
2137 lname = &info->mti_name;
2138 mdt_name_unpack(pill, &RMF_NAME, lname, MNF_FIX_ANON);
2140 if (info->mti_body->mbo_valid & OBD_MD_NAMEHASH) {
2141 reqbody = req_capsule_client_get(pill, &RMF_MDT_BODY);
2142 if (unlikely(reqbody == NULL))
2143 RETURN(err_serious(-EPROTO));
2145 *child_fid = reqbody->mbo_fid2;
2146 if (unlikely(!fid_is_sane(child_fid)))
2147 RETURN(err_serious(-EINVAL));
2149 if (lu_fid_eq(mdt_object_fid(parent), child_fid)) {
2150 mdt_object_get(info->mti_env, parent);
2153 child = mdt_object_find(info->mti_env, info->mti_mdt,
2156 RETURN(PTR_ERR(child));
2159 CDEBUG(D_INODE, "getattr with lock for "DFID"/"DFID", "
2161 PFID(mdt_object_fid(parent)),
2162 PFID(&reqbody->mbo_fid2), ldlm_rep);
2163 } else if (lu_name_is_valid(lname)) {
2164 if (mdt_object_remote(parent)) {
2165 CERROR("%s: parent "DFID" is on remote target\n",
2166 mdt_obd_name(info->mti_mdt),
2167 PFID(mdt_object_fid(parent)));
2171 CDEBUG(D_INODE, "getattr with lock for "DFID"/"DNAME", "
2172 "ldlm_rep = %p\n", PFID(mdt_object_fid(parent)),
2173 PNAME(lname), ldlm_rep);
2175 if (parent->mot_obj.lo_header->loh_attr & LOHA_FSCRYPT_MD ||
2176 (fid_is_root(mdt_object_fid(parent)) &&
2177 lname->ln_namelen == strlen(dot_fscrypt_name) &&
2178 strncmp(lname->ln_name, dot_fscrypt_name,
2179 lname->ln_namelen) == 0))
2182 reqbody = req_capsule_client_get(pill, &RMF_MDT_BODY);
2183 if (unlikely(reqbody == NULL))
2184 RETURN(err_serious(-EPROTO));
2186 *child_fid = reqbody->mbo_fid2;
2187 if (unlikely(!fid_is_sane(child_fid)))
2188 RETURN(err_serious(-EINVAL));
2190 if (lu_fid_eq(mdt_object_fid(parent), child_fid)) {
2191 mdt_object_get(info->mti_env, parent);
2194 child = mdt_object_find(info->mti_env, info->mti_mdt,
2197 RETURN(PTR_ERR(child));
2200 if (mdt_object_remote(child)) {
2201 CERROR("%s: child "DFID" is on remote target\n",
2202 mdt_obd_name(info->mti_mdt),
2203 PFID(mdt_object_fid(child)));
2204 GOTO(out_child, rc = -EPROTO);
2207 /* don't fetch LOOKUP lock if it's remote object */
2208 rc = mdt_is_remote_object(info, parent, child);
2210 GOTO(out_child, rc);
2212 child_bits &= ~MDS_INODELOCK_LOOKUP;
2214 CDEBUG(D_INODE, "getattr with lock for "DFID"/"DFID", "
2216 PFID(mdt_object_fid(parent)),
2217 PFID(&reqbody->mbo_fid2), ldlm_rep);
2220 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_EXECD);
2222 if (unlikely(!mdt_object_exists(parent)) &&
2223 !(info->mti_body->mbo_valid & OBD_MD_NAMEHASH) &&
2224 lu_name_is_valid(lname)) {
2225 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2227 "Parent doesn't exist!");
2228 GOTO(out_child, rc = -ESTALE);
2231 if (!child && is_resent) {
2232 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
2234 /* Lock is pinned by ldlm_handle_enqueue0() as it is
2235 * a resend case, however, it could be already destroyed
2236 * due to client eviction or a raced cancel RPC.
2238 LDLM_DEBUG_NOLOCK("Invalid lock handle %#llx",
2239 lhc->mlh_reg_lh.cookie);
2242 fid_extract_from_res_name(child_fid,
2243 &lock->l_resource->lr_name);
2244 LDLM_LOCK_PUT(lock);
2245 child = mdt_object_find(info->mti_env, info->mti_mdt,
2248 RETURN(PTR_ERR(child));
2249 } else if (!(info->mti_body->mbo_valid & OBD_MD_NAMEHASH) &&
2250 lu_name_is_valid(lname)) {
2251 if (info->mti_body->mbo_valid == OBD_MD_FLID) {
2252 rc = mdt_raw_lookup(info, parent, lname);
2257 /* step 1: lock parent only if parent is a directory */
2258 if (S_ISDIR(lu_object_attr(&parent->mot_obj))) {
2259 lhp = &info->mti_lh[MDT_LH_PARENT];
2260 rc = mdt_parent_lock(info, parent, lhp, lname, LCK_PR);
2261 if (unlikely(rc != 0))
2265 /* step 2: lookup child's fid by name */
2266 fid_zero(child_fid);
2267 rc = mdo_lookup(info->mti_env, mdt_object_child(parent), lname,
2268 child_fid, &info->mti_spec);
2270 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
2273 GOTO(unlock_parent, rc);
2275 child = mdt_object_find(info->mti_env, info->mti_mdt,
2277 if (unlikely(IS_ERR(child)))
2278 GOTO(unlock_parent, rc = PTR_ERR(child));
2281 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
2283 /* step 3: lock child regardless if it is local or remote. */
2286 if (info->mti_body->mbo_valid & OBD_MD_NAMEHASH) {
2287 /* Here, lname is an encoded hash of on-disk name, and
2288 * client is doing access without encryption key.
2289 * So we need to compare name hash with the one in the request.
2291 if (!find_name_matching_hash(info, lname, parent,
2293 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
2294 mdt_clear_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
2295 GOTO(out_child, rc = -ENOENT);
2299 CFS_FAIL_TIMEOUT(OBD_FAIL_MDS_RESEND, obd_timeout * 2);
2300 if (!mdt_object_exists(child)) {
2301 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2303 "Object doesn't exist!");
2304 GOTO(out_child, rc = -ENOENT);
2307 rc = mdt_check_resent_lock(info, child, lhc);
2309 GOTO(out_child, rc);
2310 } else if (rc > 0) {
2311 if (!(child_bits & MDS_INODELOCK_UPDATE) &&
2312 !mdt_object_remote(child)) {
2313 struct md_attr *ma = &info->mti_attr;
2316 ma->ma_need = MA_INODE;
2317 rc = mdt_attr_get_complex(info, child, ma);
2318 if (unlikely(rc != 0))
2319 GOTO(out_child, rc);
2321 /* If the file has not been changed for some time, we
2322 * return not only a LOOKUP lock, but also an UPDATE
2323 * lock and this might save us RPC on later STAT. For
2324 * directories, it also let negative dentry cache start
2325 * working for this dir. */
2326 if (ma->ma_valid & MA_INODE &&
2327 ma->ma_attr.la_valid & LA_CTIME &&
2328 info->mti_mdt->mdt_namespace->ns_ctime_age_limit +
2329 ma->ma_attr.la_ctime < ktime_get_real_seconds())
2330 child_bits |= MDS_INODELOCK_UPDATE;
2333 /* layout lock must be granted in a best-effort way
2334 * for IT operations */
2335 LASSERT(!(child_bits & MDS_INODELOCK_LAYOUT));
2336 if (S_ISREG(lu_object_attr(&child->mot_obj)) &&
2337 !mdt_object_remote(child) && ldlm_rep != NULL) {
2338 if (!CFS_FAIL_CHECK(OBD_FAIL_MDS_NO_LL_GETATTR) &&
2339 exp_connect_layout(info->mti_exp)) {
2340 /* try to grant layout lock for regular file. */
2341 try_bits = MDS_INODELOCK_LAYOUT;
2343 /* Acquire DOM lock in advance for data-on-mdt file */
2344 if (child != parent)
2345 try_bits |= MDS_INODELOCK_DOM;
2349 * To avoid possible deadlock between batched statahead RPC
2350 * and rename()/migrate() operation, it should use trylock to
2351 * obtain the DLM PR ibits lock for file attributes in a
2352 * batched statahead RPC. A failed trylock means that other
2353 * users maybe modify the directory simultaneously as in current
2354 * Lustre design the server only grants read lock to a client.
2356 * When a trylock failed, the MDT reports the conflict with
2357 * error code -EBUSY, and stops statahead immediately.
2359 if (info->mti_batch_env) {
2361 * This is a sub stat-ahead request in a batched RPC.
2362 * However, the @child is a remote object, we just
2363 * return -EREMOTE here to forbid stat-ahead on it.
2365 if (mdt_object_remote(child))
2366 GOTO(out_child, rc = -EREMOTE);
2368 try_bits |= child_bits;
2372 if (try_bits != 0) {
2373 /* try layout lock, it may fail to be granted due to
2374 * contention at LOOKUP or UPDATE */
2375 rc = mdt_object_lock_try(info, child, lhc, &child_bits,
2377 if (child_bits & MDS_INODELOCK_LAYOUT)
2380 /* Do not enqueue the UPDATE lock from MDT(cross-MDT),
2381 * client will enqueue the lock to the remote MDT */
2382 if (mdt_object_remote(child))
2383 rc = mdt_object_lookup_lock(info, NULL, child,
2386 rc = mdt_object_lock(info, child, lhc,
2387 child_bits, LCK_PR);
2389 if (unlikely(rc != 0))
2390 GOTO(out_child, rc);
2391 if (info->mti_batch_env && child_bits == 0) {
2393 mdt_object_unlock(info, child, lhc, 1);
2394 GOTO(out_child, rc = -EBUSY);
2399 child->mot_obj.lo_header->loh_attr |= LOHA_FSCRYPT_MD;
2401 /* finally, we can get attr for child. */
2402 rc = mdt_getattr_internal(info, child, ma_need);
2403 if (unlikely(rc != 0)) {
2405 mdt_object_unlock(info, child, lhc, 1);
2406 GOTO(out_child, rc);
2409 rc = mdt_pack_secctx_in_reply(info, child);
2412 mdt_object_unlock(info, child, lhc, 1);
2413 GOTO(out_child, rc);
2416 rc = mdt_pack_encctx_in_reply(info, child);
2419 mdt_object_unlock(info, child, lhc, 1);
2420 GOTO(out_child, rc);
2423 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
2425 /* Debugging code. */
2426 LDLM_DEBUG(lock, "Returning lock to client");
2427 LASSERTF(fid_res_name_eq(mdt_object_fid(child),
2428 &lock->l_resource->lr_name),
2429 "Lock res_id: "DLDLMRES", fid: "DFID"\n",
2430 PLDLMRES(lock->l_resource),
2431 PFID(mdt_object_fid(child)));
2433 if (unlikely(CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_ENQ_RESEND))) {
2434 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
2435 CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_ENQ_RESEND,
2437 req->rq_arrival_time.tv_sec +
2439 /* Put the lock to the waiting list and force the cancel */
2440 ldlm_set_ast_sent(lock);
2443 if (S_ISREG(lu_object_attr(&child->mot_obj)) &&
2444 !mdt_object_remote(child) && child != parent) {
2445 mdt_object_put(info->mti_env, child);
2446 rc = mdt_pack_size2body(info, child_fid,
2448 if (rc != 0 && child_bits & MDS_INODELOCK_DOM) {
2449 /* DOM lock was taken in advance but this is
2450 * not DoM file. Drop the lock.
2452 lock_res_and_lock(lock);
2453 ldlm_inodebits_drop(lock, MDS_INODELOCK_DOM);
2454 unlock_res_and_lock(lock);
2456 LDLM_LOCK_PUT(lock);
2457 GOTO(unlock_parent, rc = 0);
2459 LDLM_LOCK_PUT(lock);
2465 mdt_object_put(info->mti_env, child);
2468 mdt_object_unlock(info, parent, lhp, 1);
2469 if (rc == -ENOENT) {
2470 /* return -ENOKEY instead of -ENOENT to encryption-unaware
2471 * client if trying to access an encrypted file
2473 int rc2 = mdt_check_enc(info, parent);
2481 /* normal handler: should release the child lock */
2482 static int mdt_getattr_name(struct tgt_session_info *tsi)
2484 struct mdt_thread_info *info = tsi2mdt_info(tsi);
2485 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_CHILD];
2486 struct mdt_body *reqbody;
2487 struct mdt_body *repbody;
2492 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
2493 LASSERT(reqbody != NULL);
2494 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2495 LASSERT(repbody != NULL);
2497 info->mti_cross_ref = !!(reqbody->mbo_valid & OBD_MD_FLCROSSREF);
2498 repbody->mbo_eadatasize = 0;
2499 repbody->mbo_aclsize = 0;
2501 rc = mdt_init_ucred(info, reqbody);
2503 GOTO(out_shrink, rc);
2505 rc = mdt_getattr_name_lock(info, lhc, MDS_INODELOCK_UPDATE, NULL);
2506 if (lustre_handle_is_used(&lhc->mlh_reg_lh)) {
2507 ldlm_lock_decref(&lhc->mlh_reg_lh, lhc->mlh_reg_mode);
2508 lhc->mlh_reg_lh.cookie = 0;
2510 mdt_exit_ucred(info);
2513 mdt_client_compatibility(info);
2514 rc2 = mdt_fix_reply(info);
2517 mdt_thread_info_fini(info);
2521 static int mdt_rmfid_unlink(struct mdt_thread_info *info,
2522 const struct lu_fid *pfid,
2523 const struct lu_name *name,
2524 struct mdt_object *obj, s64 ctime)
2526 struct lu_fid *child_fid = &info->mti_tmp_fid1;
2527 struct ldlm_enqueue_info *einfo = &info->mti_einfo;
2528 struct mdt_device *mdt = info->mti_mdt;
2529 struct md_attr *ma = &info->mti_attr;
2530 struct mdt_lock_handle *parent_lh;
2531 struct mdt_lock_handle *child_lh;
2532 struct mdt_object *pobj;
2536 pobj = mdt_object_find(info->mti_env, mdt, pfid);
2538 GOTO(out, rc = PTR_ERR(pobj));
2540 parent_lh = &info->mti_lh[MDT_LH_PARENT];
2541 rc = mdt_parent_lock(info, pobj, parent_lh, name, LCK_PW);
2543 GOTO(put_parent, rc);
2545 rc = mdo_lookup(info->mti_env, mdt_object_child(pobj),
2546 name, child_fid, &info->mti_spec);
2548 GOTO(unlock_parent, rc);
2550 if (!lu_fid_eq(child_fid, mdt_object_fid(obj)))
2551 GOTO(unlock_parent, rc = -EREMCHG);
2553 child_lh = &info->mti_lh[MDT_LH_CHILD];
2554 rc = mdt_object_stripes_lock(info, pobj, obj, child_lh, einfo,
2555 MDS_INODELOCK_LOOKUP |
2556 MDS_INODELOCK_UPDATE, LCK_EX);
2558 GOTO(unlock_parent, rc);
2560 if (atomic_read(&obj->mot_open_count)) {
2561 CDEBUG(D_OTHER, "object "DFID" open, skip\n",
2562 PFID(mdt_object_fid(obj)));
2563 GOTO(unlock_child, rc = -EBUSY);
2567 ma->ma_valid = MA_INODE;
2568 ma->ma_attr.la_valid = LA_CTIME;
2569 ma->ma_attr.la_ctime = ctime;
2571 mutex_lock(&obj->mot_lov_mutex);
2573 rc = mdo_unlink(info->mti_env, mdt_object_child(pobj),
2574 mdt_object_child(obj), name, ma, 0);
2576 mutex_unlock(&obj->mot_lov_mutex);
2579 mdt_object_stripes_unlock(info, obj, child_lh, einfo, 1);
2581 mdt_object_unlock(info, pobj, parent_lh, 1);
2583 mdt_object_put(info->mti_env, pobj);
2588 static int mdt_rmfid_check_permission(struct mdt_thread_info *info,
2589 struct mdt_object *obj)
2591 struct lu_ucred *uc = lu_ucred(info->mti_env);
2592 struct md_attr *ma = &info->mti_attr;
2593 struct lu_attr *la = &ma->ma_attr;
2597 ma->ma_need = MA_INODE;
2598 rc = mo_attr_get(info->mti_env, mdt_object_child(obj), ma);
2602 if (la->la_flags & LUSTRE_IMMUTABLE_FL)
2605 /* we want rbac roles to have precedence over any other
2606 * permission or capability checks
2608 if (!uc->uc_rbac_byfid_ops)
2610 if (cap_raised(uc->uc_cap, CAP_DAC_OVERRIDE))
2612 if (uc->uc_fsuid == la->la_uid) {
2613 if ((la->la_mode & S_IWUSR) == 0)
2615 } else if (uc->uc_fsgid == la->la_gid) {
2616 if ((la->la_mode & S_IWGRP) == 0)
2618 } else if ((la->la_mode & S_IWOTH) == 0) {
2626 static int mdt_rmfid_one(struct mdt_thread_info *info, struct lu_fid *fid,
2629 struct mdt_device *mdt = info->mti_mdt;
2630 struct mdt_object *obj = NULL;
2631 struct linkea_data ldata = { NULL };
2632 struct lu_buf *buf = &info->mti_big_buf;
2633 struct lu_name *name = &info->mti_name;
2634 struct lu_fid *pfid = &info->mti_tmp_fid1;
2635 struct link_ea_header *leh;
2636 struct link_ea_entry *lee;
2637 int reclen, count, rc = 0;
2640 if (!fid_is_sane(fid))
2641 GOTO(out, rc = -EINVAL);
2643 if (!fid_is_namespace_visible(fid))
2644 GOTO(out, rc = -EINVAL);
2646 obj = mdt_object_find(info->mti_env, mdt, fid);
2648 GOTO(out, rc = PTR_ERR(obj));
2650 if (mdt_object_remote(obj))
2651 GOTO(out, rc = -EREMOTE);
2652 if (!mdt_object_exists(obj) || lu_object_is_dying(&obj->mot_header))
2653 GOTO(out, rc = -ENOENT);
2655 rc = mdt_rmfid_check_permission(info, obj);
2660 buf = lu_buf_check_and_alloc(buf, PATH_MAX);
2662 GOTO(out, rc = -ENOMEM);
2665 rc = mdt_links_read(info, obj, &ldata);
2670 lee = (struct link_ea_entry *)(leh + 1);
2671 for (count = 0; count < leh->leh_reccount; count++) {
2672 /* remove every hardlink */
2673 linkea_entry_unpack(lee, &reclen, name, pfid);
2674 lee = (struct link_ea_entry *) ((char *)lee + reclen);
2675 rc = mdt_rmfid_unlink(info, pfid, name, obj, ctime);
2681 if (obj && !IS_ERR(obj))
2682 mdt_object_put(info->mti_env, obj);
2683 if (info->mti_big_buf.lb_buf)
2684 lu_buf_free(&info->mti_big_buf);
2689 static int mdt_rmfid(struct tgt_session_info *tsi)
2691 struct mdt_thread_info *mti = tsi2mdt_info(tsi);
2692 struct mdt_body *reqbody;
2693 struct lu_fid *fids, *rfids;
2699 reqbody = req_capsule_client_get(tsi->tsi_pill, &RMF_MDT_BODY);
2700 if (reqbody == NULL)
2702 bufsize = req_capsule_get_size(tsi->tsi_pill, &RMF_FID_ARRAY,
2704 nr = bufsize / sizeof(struct lu_fid);
2705 if (nr * sizeof(struct lu_fid) != bufsize)
2707 req_capsule_set_size(tsi->tsi_pill, &RMF_RCS,
2708 RCL_SERVER, nr * sizeof(__u32));
2709 req_capsule_set_size(tsi->tsi_pill, &RMF_FID_ARRAY,
2710 RCL_SERVER, nr * sizeof(struct lu_fid));
2711 rc = req_capsule_server_pack(tsi->tsi_pill);
2713 GOTO(out, rc = err_serious(rc));
2714 fids = req_capsule_client_get(tsi->tsi_pill, &RMF_FID_ARRAY);
2717 rcs = req_capsule_server_get(tsi->tsi_pill, &RMF_RCS);
2719 rfids = req_capsule_server_get(tsi->tsi_pill, &RMF_FID_ARRAY);
2722 mdt_init_ucred(mti, reqbody);
2723 for (i = 0; i < nr; i++) {
2725 rcs[i] = mdt_rmfid_one(mti, fids + i, reqbody->mbo_ctime);
2727 mdt_exit_ucred(mti);
2733 static int mdt_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2734 void *karg, void __user *uarg);
2736 int mdt_io_set_info(struct tgt_session_info *tsi)
2738 struct ptlrpc_request *req = tgt_ses_req(tsi);
2739 struct ost_body *body = NULL, *repbody;
2740 void *key, *val = NULL;
2741 int keylen, vallen, rc = 0;
2742 bool is_grant_shrink;
2746 key = req_capsule_client_get(tsi->tsi_pill, &RMF_SETINFO_KEY);
2748 DEBUG_REQ(D_HA, req, "no set_info key");
2749 RETURN(err_serious(-EFAULT));
2751 keylen = req_capsule_get_size(tsi->tsi_pill, &RMF_SETINFO_KEY,
2754 val = req_capsule_client_get(tsi->tsi_pill, &RMF_SETINFO_VAL);
2756 DEBUG_REQ(D_HA, req, "no set_info val");
2757 RETURN(err_serious(-EFAULT));
2759 vallen = req_capsule_get_size(tsi->tsi_pill, &RMF_SETINFO_VAL,
2762 is_grant_shrink = KEY_IS(KEY_GRANT_SHRINK);
2763 if (is_grant_shrink)
2764 /* In this case the value is actually an RMF_OST_BODY, so we
2765 * transmutate the type of this PTLRPC */
2766 req_capsule_extend(tsi->tsi_pill, &RQF_OST_SET_GRANT_INFO);
2768 rc = req_capsule_server_pack(tsi->tsi_pill);
2772 if (is_grant_shrink) {
2773 body = req_capsule_client_get(tsi->tsi_pill, &RMF_OST_BODY);
2775 repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_OST_BODY);
2778 /** handle grant shrink, similar to a read request */
2779 tgt_grant_prepare_read(tsi->tsi_env, tsi->tsi_exp,
2782 CERROR("%s: Unsupported key %s\n",
2783 tgt_name(tsi->tsi_tgt), (char *)key);
2791 static int mdt_set_info(struct tgt_session_info *tsi)
2793 struct ptlrpc_request *req = tgt_ses_req(tsi);
2796 int keylen, vallen, rc = 0;
2800 key = req_capsule_client_get(tsi->tsi_pill, &RMF_SETINFO_KEY);
2802 DEBUG_REQ(D_HA, req, "no set_info key");
2803 RETURN(err_serious(-EFAULT));
2806 keylen = req_capsule_get_size(tsi->tsi_pill, &RMF_SETINFO_KEY,
2809 val = req_capsule_client_get(tsi->tsi_pill, &RMF_SETINFO_VAL);
2811 DEBUG_REQ(D_HA, req, "no set_info val");
2812 RETURN(err_serious(-EFAULT));
2815 vallen = req_capsule_get_size(tsi->tsi_pill, &RMF_SETINFO_VAL,
2818 /* Swab any part of val you need to here */
2819 if (KEY_IS(KEY_READ_ONLY)) {
2820 spin_lock(&req->rq_export->exp_lock);
2822 *exp_connect_flags_ptr(req->rq_export) |=
2825 *exp_connect_flags_ptr(req->rq_export) &=
2826 ~OBD_CONNECT_RDONLY;
2827 spin_unlock(&req->rq_export->exp_lock);
2828 } else if (KEY_IS(KEY_CHANGELOG_CLEAR)) {
2829 struct changelog_setinfo *cs = val;
2831 if (vallen != sizeof(*cs)) {
2832 CERROR("%s: bad changelog_clear setinfo size %d\n",
2833 tgt_name(tsi->tsi_tgt), vallen);
2836 if (req_capsule_req_need_swab(&req->rq_pill)) {
2837 __swab64s(&cs->cs_recno);
2838 __swab32s(&cs->cs_id);
2841 if (!mdt_changelog_allow(tsi2mdt_info(tsi)))
2843 rc = mdt_iocontrol(OBD_IOC_CHANGELOG_CLEAR, req->rq_export,
2845 } else if (KEY_IS(KEY_EVICT_BY_NID)) {
2847 obd_export_evict_by_nid(req->rq_export->exp_obd, val);
2854 static int mdt_readpage(struct tgt_session_info *tsi)
2856 struct mdt_thread_info *info = mdt_th_info(tsi->tsi_env);
2857 struct mdt_object *object = mdt_obj(tsi->tsi_corpus);
2858 struct lu_rdpg *rdpg = &info->mti_u.rdpg.mti_rdpg;
2859 const struct mdt_body *reqbody = tsi->tsi_mdt_body;
2860 struct mdt_body *repbody;
2866 if (CFS_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK))
2867 RETURN(err_serious(-ENOMEM));
2869 repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_MDT_BODY);
2870 if (repbody == NULL || reqbody == NULL)
2871 RETURN(err_serious(-EFAULT));
2874 * prepare @rdpg before calling lower layers and transfer itself. Here
2875 * reqbody->size contains offset of where to start to read and
2876 * reqbody->nlink contains number bytes to read.
2878 rdpg->rp_hash = reqbody->mbo_size;
2879 if (rdpg->rp_hash != reqbody->mbo_size) {
2880 CERROR("Invalid hash: %#llx != %#llx\n",
2881 rdpg->rp_hash, reqbody->mbo_size);
2885 rdpg->rp_attrs = reqbody->mbo_mode;
2886 if (exp_connect_flags(tsi->tsi_exp) & OBD_CONNECT_64BITHASH)
2887 rdpg->rp_attrs |= LUDA_64BITHASH;
2888 rdpg->rp_count = min_t(unsigned int, reqbody->mbo_nlink,
2889 exp_max_brw_size(tsi->tsi_exp));
2890 rdpg->rp_npages = (rdpg->rp_count + PAGE_SIZE - 1) >>
2892 OBD_ALLOC_PTR_ARRAY_LARGE(rdpg->rp_pages, rdpg->rp_npages);
2893 if (rdpg->rp_pages == NULL)
2896 for (i = 0; i < rdpg->rp_npages; ++i) {
2897 rdpg->rp_pages[i] = alloc_page(GFP_NOFS);
2898 if (rdpg->rp_pages[i] == NULL)
2899 GOTO(free_rdpg, rc = -ENOMEM);
2902 /* call lower layers to fill allocated pages with directory data */
2903 rc = mo_readpage(tsi->tsi_env, mdt_object_child(object), rdpg);
2905 GOTO(free_rdpg, rc);
2907 /* send pages to client */
2908 rc = tgt_sendpage(tsi, rdpg, rc);
2913 for (i = 0; i < rdpg->rp_npages; i++)
2914 if (rdpg->rp_pages[i] != NULL)
2915 __free_page(rdpg->rp_pages[i]);
2916 OBD_FREE_PTR_ARRAY_LARGE(rdpg->rp_pages, rdpg->rp_npages);
2918 if (CFS_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE))
2924 static int mdt_fix_attr_ucred(struct mdt_thread_info *info, __u32 op)
2926 struct lu_ucred *uc = mdt_ucred_check(info);
2927 struct lu_attr *attr = &info->mti_attr.ma_attr;
2932 if (op != REINT_SETATTR) {
2933 if ((attr->la_valid & LA_UID) && (attr->la_uid != -1))
2934 attr->la_uid = uc->uc_fsuid;
2935 /* for S_ISGID, inherit gid from his parent, such work will be
2936 * done in cmm/mdd layer, here set all cases as uc->uc_fsgid. */
2937 if ((attr->la_valid & LA_GID) && (attr->la_gid != -1))
2938 attr->la_gid = uc->uc_fsgid;
2944 static inline bool mdt_is_readonly_open(struct mdt_thread_info *info, __u32 op)
2946 return op == REINT_OPEN &&
2947 !(info->mti_spec.sp_cr_flags & (MDS_FMODE_WRITE | MDS_OPEN_CREAT));
2950 static void mdt_preset_secctx_size(struct mdt_thread_info *info)
2952 struct req_capsule *pill = info->mti_pill;
2954 if (req_capsule_has_field(pill, &RMF_FILE_SECCTX,
2956 req_capsule_has_field(pill, &RMF_FILE_SECCTX_NAME,
2958 if (req_capsule_get_size(pill, &RMF_FILE_SECCTX_NAME,
2960 /* pre-set size in server part with max size */
2961 req_capsule_set_size(pill, &RMF_FILE_SECCTX,
2963 req_capsule_ptlreq(pill) ?
2964 OBD_MAX_DEFAULT_EA_SIZE :
2967 req_capsule_set_size(pill, &RMF_FILE_SECCTX,
2972 static int mdt_reint_internal(struct mdt_thread_info *info,
2973 struct mdt_lock_handle *lhc,
2976 struct req_capsule *pill = info->mti_pill;
2977 struct mdt_body *repbody;
2982 rc = mdt_reint_unpack(info, op);
2984 CERROR("Can't unpack reint, rc %d\n", rc);
2985 RETURN(err_serious(rc));
2989 /* check if the file system is set to readonly. O_RDONLY open
2990 * is still allowed even the file system is set to readonly mode */
2991 if (mdt_rdonly(info->mti_exp) && !mdt_is_readonly_open(info, op))
2992 RETURN(err_serious(-EROFS));
2994 /* for replay (no_create) lmm is not needed, client has it already */
2995 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
2996 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER,
2999 /* llog cookies are always 0, the field is kept for compatibility */
3000 if (req_capsule_has_field(pill, &RMF_LOGCOOKIES, RCL_SERVER))
3001 req_capsule_set_size(pill, &RMF_LOGCOOKIES, RCL_SERVER, 0);
3003 /* Set ACL reply buffer size as LUSTRE_POSIX_ACL_MAX_SIZE_OLD
3004 * by default. If the target object has more ACL entries, then
3005 * enlarge the buffer when necessary. */
3006 if (req_capsule_has_field(pill, &RMF_ACL, RCL_SERVER))
3007 req_capsule_set_size(pill, &RMF_ACL, RCL_SERVER,
3008 LUSTRE_POSIX_ACL_MAX_SIZE_OLD);
3010 mdt_preset_secctx_size(info);
3011 mdt_preset_encctx_size(info);
3013 rc = req_capsule_server_pack(pill);
3015 CERROR("Can't pack response, rc %d\n", rc);
3016 RETURN(err_serious(rc));
3019 if (req_capsule_has_field(pill, &RMF_MDT_BODY, RCL_SERVER)) {
3020 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
3022 repbody->mbo_eadatasize = 0;
3023 repbody->mbo_aclsize = 0;
3026 CFS_FAIL_TIMEOUT(OBD_FAIL_MDS_REINT_DELAY, 10);
3028 /* for replay no cookkie / lmm need, because client have this already */
3029 if (info->mti_spec.no_create)
3030 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
3031 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER, 0);
3033 rc = mdt_init_ucred_reint(info);
3035 GOTO(out_shrink, rc);
3037 rc = mdt_fix_attr_ucred(info, op);
3039 GOTO(out_ucred, rc = err_serious(rc));
3041 rc = mdt_check_resent(info, mdt_reconstruct, lhc);
3043 GOTO(out_ucred, rc);
3044 } else if (rc == 1) {
3045 DEBUG_REQ(D_INODE, mdt_info_req(info), "resent opt");
3046 rc = lustre_msg_get_status(mdt_info_req(info)->rq_repmsg);
3047 GOTO(out_ucred, rc);
3049 rc = mdt_reint_rec(info, lhc);
3052 mdt_exit_ucred(info);
3054 mdt_client_compatibility(info);
3056 rc2 = mdt_fix_reply(info);
3061 * Data-on-MDT optimization - read data along with OPEN and return it
3062 * in reply when possible.
3064 if (rc == 0 && op == REINT_OPEN && !req_is_replay(pill->rc_req))
3065 rc = mdt_dom_read_on_open(info, info->mti_mdt,
3071 static long mdt_reint_opcode(struct ptlrpc_request *req,
3072 const struct req_format **fmt)
3074 struct mdt_device *mdt;
3075 struct mdt_rec_reint *rec;
3078 rec = req_capsule_client_get(&req->rq_pill, &RMF_REC_REINT);
3080 opc = rec->rr_opcode;
3081 DEBUG_REQ(D_INODE, req, "reint opt = %ld", opc);
3082 if (opc < REINT_MAX && fmt[opc] != NULL)
3083 req_capsule_extend(&req->rq_pill, fmt[opc]);
3085 mdt = mdt_exp2dev(req->rq_export);
3086 CERROR("%s: Unsupported opcode '%ld' from client '%s':"
3087 " rc = %d\n", req->rq_export->exp_obd->obd_name,
3088 opc, mdt->mdt_ldlm_client->cli_name, -EFAULT);
3089 opc = err_serious(-EFAULT);
3092 opc = err_serious(-EFAULT);
3097 static int mdt_reint(struct tgt_session_info *tsi)
3101 static const struct req_format *reint_fmts[REINT_MAX] = {
3102 [REINT_SETATTR] = &RQF_MDS_REINT_SETATTR,
3103 [REINT_CREATE] = &RQF_MDS_REINT_CREATE,
3104 [REINT_LINK] = &RQF_MDS_REINT_LINK,
3105 [REINT_UNLINK] = &RQF_MDS_REINT_UNLINK,
3106 [REINT_RENAME] = &RQF_MDS_REINT_RENAME,
3107 [REINT_OPEN] = &RQF_MDS_REINT_OPEN,
3108 [REINT_SETXATTR] = &RQF_MDS_REINT_SETXATTR,
3109 [REINT_RMENTRY] = &RQF_MDS_REINT_UNLINK,
3110 [REINT_MIGRATE] = &RQF_MDS_REINT_MIGRATE,
3111 [REINT_RESYNC] = &RQF_MDS_REINT_RESYNC,
3116 opc = mdt_reint_opcode(tgt_ses_req(tsi), reint_fmts);
3118 struct mdt_thread_info *info = tsi2mdt_info(tsi);
3120 * No lock possible here from client to pass it to reint code
3123 rc = mdt_reint_internal(info, NULL, opc);
3124 mdt_thread_info_fini(info);
3129 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_REINT_NET_REP;
3133 /* this should sync the whole device */
3134 int mdt_device_sync(const struct lu_env *env, struct mdt_device *mdt)
3136 struct dt_device *dt = mdt->mdt_bottom;
3140 rc = dt->dd_ops->dt_sync(env, dt);
3144 /* this should sync this object */
3145 static int mdt_object_sync(const struct lu_env *env, struct obd_export *exp,
3146 struct mdt_object *mo)
3152 if (!mdt_object_exists(mo)) {
3153 CWARN("%s: non existing object "DFID": rc = %d\n",
3154 exp->exp_obd->obd_name, PFID(mdt_object_fid(mo)),
3159 if (S_ISREG(lu_object_attr(&mo->mot_obj))) {
3160 struct lu_target *tgt = tgt_ses_info(env)->tsi_tgt;
3161 dt_obj_version_t version;
3163 version = dt_version_get(env, mdt_obj2dt(mo));
3164 if (version > tgt->lut_obd->obd_last_committed)
3165 rc = mo_object_sync(env, mdt_object_child(mo));
3167 rc = mo_object_sync(env, mdt_object_child(mo));
3173 static int mdt_sync(struct tgt_session_info *tsi)
3175 struct ptlrpc_request *req = tgt_ses_req(tsi);
3176 struct req_capsule *pill = tsi->tsi_pill;
3177 struct mdt_body *body;
3178 ktime_t kstart = ktime_get();
3183 if (CFS_FAIL_CHECK(OBD_FAIL_MDS_SYNC_PACK))
3184 RETURN(err_serious(-ENOMEM));
3186 if (fid_seq(&tsi->tsi_mdt_body->mbo_fid1) == 0) {
3187 rc = mdt_device_sync(tsi->tsi_env, mdt_exp2dev(tsi->tsi_exp));
3189 struct mdt_thread_info *info = tsi2mdt_info(tsi);
3191 if (unlikely(info->mti_object == NULL))
3194 /* sync an object */
3195 rc = mdt_object_sync(tsi->tsi_env, tsi->tsi_exp,
3198 const struct lu_fid *fid;
3199 struct lu_attr *la = &info->mti_attr.ma_attr;
3201 info->mti_attr.ma_need = MA_INODE;
3202 info->mti_attr.ma_valid = 0;
3203 rc = mdt_attr_get_complex(info, info->mti_object,
3206 body = req_capsule_server_get(pill,
3208 fid = mdt_object_fid(info->mti_object);
3209 mdt_pack_attr2body(info, body, la, fid);
3212 mdt_thread_info_fini(info);
3215 mdt_counter_incr(req, LPROC_MDT_SYNC,
3216 ktime_us_delta(ktime_get(), kstart));
3221 static int mdt_data_sync(struct tgt_session_info *tsi)
3223 struct mdt_thread_info *info;
3224 struct mdt_device *mdt = mdt_exp2dev(tsi->tsi_exp);
3225 struct ost_body *body = tsi->tsi_ost_body;
3226 struct ost_body *repbody;
3227 struct mdt_object *mo = NULL;
3233 repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_OST_BODY);
3235 /* if no fid is specified then do nothing,
3236 * device sync is done via MDS_SYNC */
3237 if (fid_is_zero(&tsi->tsi_fid))
3240 mo = mdt_object_find(tsi->tsi_env, mdt, &tsi->tsi_fid);
3242 RETURN(PTR_ERR(mo));
3244 rc = mdt_object_sync(tsi->tsi_env, tsi->tsi_exp, mo);
3248 repbody->oa.o_oi = body->oa.o_oi;
3249 repbody->oa.o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
3251 info = tsi2mdt_info(tsi);
3252 ma = &info->mti_attr;
3253 ma->ma_need = MA_INODE;
3255 rc = mdt_attr_get_complex(info, mo, ma);
3257 obdo_from_la(&repbody->oa, &ma->ma_attr, VALID_FLAGS);
3260 mdt_thread_info_fini(info);
3265 mdt_object_put(tsi->tsi_env, mo);
3270 * Handle quota control requests to consult current usage/limit, but also
3271 * to configure quota enforcement
3273 static int mdt_quotactl(struct tgt_session_info *tsi)
3275 struct obd_export *exp = tsi->tsi_exp;
3276 struct req_capsule *pill = tsi->tsi_pill;
3277 struct obd_quotactl *oqctl, *repoqc;
3279 struct mdt_device *mdt = mdt_exp2dev(exp);
3280 struct lu_device *qmt = mdt->mdt_qmt_dev;
3281 struct lu_nodemap *nodemap;
3284 oqctl = req_capsule_client_get(pill, &RMF_OBD_QUOTACTL);
3286 RETURN(err_serious(-EPROTO));
3288 rc = req_capsule_server_pack(pill);
3290 RETURN(err_serious(rc));
3292 nodemap = nodemap_get_from_exp(exp);
3293 if (IS_ERR(nodemap))
3294 RETURN(PTR_ERR(nodemap));
3296 switch (oqctl->qc_cmd) {
3297 /* master quotactl */
3300 case LUSTRE_Q_SETDEFAULT:
3301 case LUSTRE_Q_SETQUOTAPOOL:
3302 case LUSTRE_Q_SETINFOPOOL:
3303 case LUSTRE_Q_SETDEFAULT_POOL:
3304 case LUSTRE_Q_DELETEQID:
3305 case LUSTRE_Q_RESETQID:
3306 if (!nodemap_can_setquota(nodemap, oqctl->qc_type,
3308 GOTO(out_nodemap, rc = -EPERM);
3312 case LUSTRE_Q_GETDEFAULT:
3313 case LUSTRE_Q_GETQUOTAPOOL:
3314 case LUSTRE_Q_GETINFOPOOL:
3315 case LUSTRE_Q_GETDEFAULT_POOL:
3317 GOTO(out_nodemap, rc = -EOPNOTSUPP);
3318 /* slave quotactl */
3325 CERROR("%s: unsupported quotactl command %d: rc = %d\n",
3326 mdt_obd_name(mdt), oqctl->qc_cmd, rc);
3327 GOTO(out_nodemap, rc);
3331 switch (oqctl->qc_type) {
3333 id = nodemap_map_id(nodemap, NODEMAP_UID,
3334 NODEMAP_CLIENT_TO_FS, id);
3337 id = nodemap_map_id(nodemap, NODEMAP_GID,
3338 NODEMAP_CLIENT_TO_FS, id);
3341 id = nodemap_map_id(nodemap, NODEMAP_PROJID,
3342 NODEMAP_CLIENT_TO_FS, id);
3345 GOTO(out_nodemap, rc = -EOPNOTSUPP);
3347 repoqc = req_capsule_server_get(pill, &RMF_OBD_QUOTACTL);
3349 GOTO(out_nodemap, rc = err_serious(-EFAULT));
3351 if (oqctl->qc_cmd == Q_SETINFO || oqctl->qc_cmd == Q_SETQUOTA)
3352 barrier_exit(tsi->tsi_tgt->lut_bottom);
3354 if (oqctl->qc_id != id)
3355 swap(oqctl->qc_id, id);
3357 if (oqctl->qc_cmd == Q_SETINFO || oqctl->qc_cmd == Q_SETQUOTA) {
3358 if (unlikely(!barrier_entry(tsi->tsi_tgt->lut_bottom)))
3359 RETURN(-EINPROGRESS);
3362 switch (oqctl->qc_cmd) {
3368 case LUSTRE_Q_SETDEFAULT:
3369 case LUSTRE_Q_GETDEFAULT:
3370 case LUSTRE_Q_SETQUOTAPOOL:
3371 case LUSTRE_Q_GETQUOTAPOOL:
3372 case LUSTRE_Q_SETINFOPOOL:
3373 case LUSTRE_Q_GETINFOPOOL:
3374 case LUSTRE_Q_SETDEFAULT_POOL:
3375 case LUSTRE_Q_GETDEFAULT_POOL:
3376 case LUSTRE_Q_DELETEQID:
3377 case LUSTRE_Q_RESETQID:
3378 /* forward quotactl request to QMT */
3379 rc = qmt_hdls.qmth_quotactl(tsi->tsi_env, qmt, oqctl);
3384 /* slave quotactl */
3385 rc = lquotactl_slv(tsi->tsi_env, tsi->tsi_tgt->lut_bottom,
3390 CERROR("Unsupported quotactl command: %d\n", oqctl->qc_cmd);
3391 GOTO(out_nodemap, rc = -EFAULT);
3394 if (oqctl->qc_id != id)
3395 swap(oqctl->qc_id, id);
3397 QCTL_COPY_NO_PNAME(repoqc, oqctl);
3401 nodemap_putref(nodemap);
3406 /** clone llog ctxt from child (mdd)
3407 * This allows remote llog (replicator) access.
3408 * We can either pass all llog RPCs (eg mdt_llog_create) on to child where the
3409 * context was originally set up, or we can handle them directly.
3410 * I choose the latter, but that means I need any llog
3411 * contexts set up by child to be accessable by the mdt. So we clone the
3412 * context into our context list here.
3414 static int mdt_llog_ctxt_clone(const struct lu_env *env, struct mdt_device *mdt,
3417 struct md_device *next = mdt->mdt_child;
3418 struct llog_ctxt *ctxt;
3421 if (!llog_ctxt_null(mdt2obd_dev(mdt), idx))
3424 rc = next->md_ops->mdo_llog_ctxt_get(env, next, idx, (void **)&ctxt);
3425 if (rc || ctxt == NULL) {
3429 rc = llog_group_set_ctxt(&mdt2obd_dev(mdt)->obd_olg, ctxt, idx);
3431 CERROR("Can't set mdt ctxt %d\n", rc);
3436 static int mdt_llog_ctxt_unclone(const struct lu_env *env,
3437 struct mdt_device *mdt, int idx)
3439 struct llog_ctxt *ctxt;
3441 ctxt = llog_get_context(mdt2obd_dev(mdt), idx);
3444 /* Put once for the get we just did, and once for the clone */
3445 llog_ctxt_put(ctxt);
3446 llog_ctxt_put(ctxt);
3451 * sec context handlers
3453 static int mdt_sec_ctx_handle(struct tgt_session_info *tsi)
3455 CFS_FAIL_TIMEOUT(OBD_FAIL_SEC_CTX_HDL_PAUSE, cfs_fail_val);
3461 * quota request handlers
3463 static int mdt_quota_dqacq(struct tgt_session_info *tsi)
3465 struct mdt_device *mdt = mdt_exp2dev(tsi->tsi_exp);
3466 struct lu_device *qmt = mdt->mdt_qmt_dev;
3471 RETURN(err_serious(-EOPNOTSUPP));
3473 rc = qmt_hdls.qmth_dqacq(tsi->tsi_env, qmt, tgt_ses_req(tsi));
3477 struct mdt_object *mdt_object_new(const struct lu_env *env,
3478 struct mdt_device *d,
3479 const struct lu_fid *f)
3481 struct lu_object_conf conf = { .loc_flags = LOC_F_NEW };
3482 struct lu_object *o;
3483 struct mdt_object *m;
3486 CDEBUG(D_INFO, "Allocate object for "DFID"\n", PFID(f));
3487 o = lu_object_find(env, &d->mdt_lu_dev, f, &conf);
3488 if (unlikely(IS_ERR(o)))
3489 m = (struct mdt_object *)o;
3495 struct mdt_object *mdt_object_find(const struct lu_env *env,
3496 struct mdt_device *d,
3497 const struct lu_fid *f)
3499 struct lu_object *o;
3500 struct mdt_object *m;
3503 CDEBUG(D_INFO, "Find object for "DFID"\n", PFID(f));
3504 o = lu_object_find(env, &d->mdt_lu_dev, f, NULL);
3505 if (unlikely(IS_ERR(o)))
3506 m = (struct mdt_object *)o;
3514 * Asyncronous commit for mdt device.
3516 * Pass asynchonous commit call down the MDS stack.
3518 * \param env environment
3519 * \param mdt the mdt device
3521 static void mdt_device_commit_async(const struct lu_env *env,
3522 struct mdt_device *mdt)
3524 struct dt_device *dt = mdt->mdt_bottom;
3528 rc = dt->dd_ops->dt_commit_async(env, dt);
3529 if (unlikely(rc != 0))
3530 CWARN("%s: async commit start failed: rc = %d\n",
3531 mdt_obd_name(mdt), rc);
3532 atomic_inc(&mdt->mdt_async_commit_count);
3537 * Mark the lock as "synchonous".
3539 * Mark the lock to deffer transaction commit to the unlock time.
3541 * \param lock the lock to mark as "synchonous"
3543 * \see mdt_is_lock_sync
3544 * \see mdt_save_lock
3546 static inline void mdt_set_lock_sync(struct ldlm_lock *lock)
3548 lock->l_ast_data = (void*)1;
3552 * Check whehter the lock "synchonous" or not.
3554 * \param lock the lock to check
3555 * \retval 1 the lock is "synchonous"
3556 * \retval 0 the lock isn't "synchronous"
3558 * \see mdt_set_lock_sync
3559 * \see mdt_save_lock
3561 static inline int mdt_is_lock_sync(struct ldlm_lock *lock)
3563 return lock->l_ast_data != NULL;
3567 * Blocking AST for mdt locks.
3569 * Starts transaction commit if in case of COS lock conflict or
3570 * deffers such a commit to the mdt_save_lock.
3572 * \param lock the lock which blocks a request or cancelling lock
3573 * \param desc unused
3574 * \param data unused
3575 * \param flag indicates whether this cancelling or blocking callback
3577 * \see ldlm_blocking_ast_nocheck
3579 int mdt_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
3580 void *data, int flag)
3582 struct obd_device *obd = ldlm_lock_to_ns(lock)->ns_obd;
3583 struct mdt_device *mdt = mdt_dev(obd->obd_lu_dev);
3584 struct ldlm_cb_set_arg *arg = data;
3585 bool commit_async = false;
3589 if (flag == LDLM_CB_CANCELING)
3592 lock_res_and_lock(lock);
3593 if (lock->l_blocking_ast != mdt_blocking_ast) {
3594 unlock_res_and_lock(lock);
3598 /* A blocking ast may be sent from ldlm_lock_decref_internal
3599 * when the last reference to a local lock was released and
3600 * during blocking event from ldlm_work_bl_ast_lock().
3601 * The 'data' parameter is l_ast_data in the first case and
3602 * callback arguments in the second one. Distinguish them by that.
3604 if (data && data != lock->l_ast_data && arg->bl_desc) {
3605 if (lock->l_req_mode & (LCK_COS | LCK_TXN))
3606 commit_async = true;
3607 else if ((lock->l_req_mode & (LCK_PW | LCK_EX)) &&
3608 ((mdt_cos_is_enabled(mdt) &&
3609 !arg->bl_desc->bl_same_client) ||
3610 (mdt_slc_is_enabled(mdt) &&
3611 arg->bl_desc->bl_txn_dependent)))
3612 mdt_set_lock_sync(lock);
3615 rc = ldlm_blocking_ast_nocheck(lock);
3620 rc = lu_env_init(&env, LCT_LOCAL);
3621 if (unlikely(rc != 0))
3622 CWARN("%s: lu_env initialization failed, cannot "
3623 "start asynchronous commit: rc = %d\n",
3626 mdt_device_commit_async(&env, mdt);
3633 * Blocking AST for cross-MDT lock
3635 * Discard lock from uncommitted_slc_locks and cancel it.
3637 * \param lock the lock which blocks a request or cancelling lock
3638 * \param desc unused
3639 * \param data unused
3640 * \param flag indicates whether this cancelling or blocking callback
3641 * \retval 0 on success
3642 * \retval negative number on error
3644 int mdt_remote_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
3645 void *data, int flag)
3651 case LDLM_CB_BLOCKING: {
3652 struct lustre_handle lockh;
3654 ldlm_lock2handle(lock, &lockh);
3655 rc = ldlm_cli_cancel(&lockh,
3656 ldlm_is_atomic_cb(lock) ? 0 : LCF_ASYNC);
3658 CDEBUG(D_INODE, "ldlm_cli_cancel: %d\n", rc);
3663 case LDLM_CB_CANCELING: {
3664 struct obd_device *obd = ldlm_lock_to_ns(lock)->ns_obd;
3665 struct mdt_device *mdt =
3666 mdt_dev(obd->obd_lu_dev->ld_site->ls_top_dev);
3668 LDLM_DEBUG(lock, "Revoke remote lock");
3670 /* discard slc lock here so that it can be cleaned anytime,
3671 * especially for cleanup_resource() */
3672 tgt_discard_slc_lock(&mdt->mdt_lut, lock);
3674 /* once we cache lock, l_ast_data is set to mdt_object */
3675 if (lock->l_ast_data != NULL) {
3676 struct mdt_object *mo = lock->l_ast_data;
3679 rc = lu_env_init(&env, LCT_MD_THREAD);
3680 if (unlikely(rc != 0)) {
3681 CWARN("%s: lu_env initialization failed, object %p "DFID" is leaked!: rc = %d\n",
3683 PFID(mdt_object_fid(mo)), rc);
3687 if (lock->l_policy_data.l_inodebits.bits &
3688 (MDS_INODELOCK_XATTR | MDS_INODELOCK_UPDATE)) {
3689 rc = mo_invalidate(&env, mdt_object_child(mo));
3690 mo->mot_cache_attr = 0;
3692 mdt_object_put(&env, mo);
3704 int mdt_check_resent_lock(struct mdt_thread_info *info,
3705 struct mdt_object *mo,
3706 struct mdt_lock_handle *lhc)
3708 /* the lock might already be gotten in ldlm_handle_enqueue() */
3709 if (unlikely(lustre_handle_is_used(&lhc->mlh_reg_lh))) {
3710 struct ptlrpc_request *req = mdt_info_req(info);
3711 struct ldlm_lock *lock;
3713 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
3714 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
3716 /* Lock is pinned by ldlm_handle_enqueue0() as it is
3717 * a resend case, however, it could be already destroyed
3718 * due to client eviction or a raced cancel RPC. */
3719 LDLM_DEBUG_NOLOCK("Invalid lock handle %#llx",
3720 lhc->mlh_reg_lh.cookie);
3724 if (!fid_res_name_eq(mdt_object_fid(mo),
3725 &lock->l_resource->lr_name)) {
3726 CWARN("%s: Although resent, but still not "
3727 "get child lock:"DFID"\n",
3728 info->mti_exp->exp_obd->obd_name,
3729 PFID(mdt_object_fid(mo)));
3730 LDLM_LOCK_PUT(lock);
3733 LDLM_LOCK_PUT(lock);
3739 static void mdt_remote_object_lock_created_cb(struct ldlm_lock *lock)
3741 mdt_object_get(NULL, lock->l_ast_data);
3744 static int mdt_remote_object_lock_try(struct mdt_thread_info *mti,
3745 struct mdt_object *obj,
3746 struct lustre_handle *lh,
3747 enum ldlm_mode mode,
3748 union ldlm_policy_data *policy,
3749 struct ldlm_res_id *res_id,
3752 struct ldlm_enqueue_info *einfo = &mti->mti_remote_einfo;
3755 LASSERT(mdt_object_remote(obj));
3757 memset(einfo, 0, sizeof(*einfo));
3758 einfo->ei_type = LDLM_IBITS;
3759 einfo->ei_mode = mode;
3760 einfo->ei_cb_bl = mdt_remote_blocking_ast;
3761 einfo->ei_cb_cp = ldlm_completion_ast;
3762 einfo->ei_enq_slave = 0;
3763 einfo->ei_res_id = res_id;
3764 einfo->ei_req_slot = 1;
3767 * if we cache lock, couple lock with mdt_object, so that object
3768 * can be easily found in lock ASTs.
3770 einfo->ei_cbdata = obj;
3771 einfo->ei_cb_created = mdt_remote_object_lock_created_cb;
3774 rc = mo_object_lock(mti->mti_env, mdt_object_child(obj), lh, einfo,
3781 /* other components like LFSCK can use lockless access
3782 * and populate cache, so we better invalidate it
3784 if (policy->l_inodebits.bits &
3785 (MDS_INODELOCK_UPDATE | MDS_INODELOCK_XATTR))
3786 mo_invalidate(mti->mti_env, mdt_object_child(obj));
3792 * Helper function to take PDO and hash lock.
3794 * if \a pdo_lock is false, don't take PDO lock, this is case in rename.
3796 int mdt_object_pdo_lock(struct mdt_thread_info *info, struct mdt_object *obj,
3797 struct mdt_lock_handle *lh, const struct lu_name *name,
3798 enum ldlm_mode mode, bool pdo_lock)
3800 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
3801 union ldlm_policy_data *policy = &info->mti_policy;
3802 struct ldlm_res_id *res_id = &info->mti_res_id;
3804 * Do not use LDLM_FL_LOCAL_ONLY for parallel lock, it is never going to
3805 * be sent to client and we do not want it slowed down due to possible
3808 __u64 dlmflags = LDLM_FL_ATOMIC_CB;
3809 __u64 *cookie = NULL;
3814 /* check for exists after object is locked */
3815 if (!mdt_object_exists(obj))
3816 /* Non-existent object shouldn't have PDO lock */
3819 /* Non-dir object shouldn't have PDO lock */
3820 if (!S_ISDIR(lu_object_attr(&obj->mot_obj)))
3823 policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
3824 policy->l_inodebits.try_bits = 0;
3825 policy->l_inodebits.li_gid = 0;
3826 policy->l_inodebits.li_initiator_id = mdt_node_id(info->mti_mdt);
3827 fid_build_reg_res_name(mdt_object_fid(obj), res_id);
3829 cookie = &info->mti_exp->exp_handle.h_cookie;
3831 mdt_lock_pdo_init(lh, mode, name);
3832 mdt_lock_pdo_mode(info, obj, lh);
3833 if (lh->mlh_pdo_mode != LCK_NL) {
3835 if (mdt_object_remote(obj)) {
3836 rc = mdt_remote_object_lock_try(info, obj,
3837 &lh->mlh_pdo_lh, lh->mlh_pdo_mode,
3838 policy, res_id, false);
3839 lh->mlh_pdo_remote = 1;
3841 rc = mdt_fid_lock(info->mti_env, ns,
3842 &lh->mlh_pdo_lh, lh->mlh_pdo_mode,
3843 policy, res_id, dlmflags, cookie);
3846 mdt_object_unlock(info, obj, lh, 1);
3850 res_id->name[LUSTRE_RES_ID_HSH_OFF] = lh->mlh_pdo_hash;
3853 if (mdt_object_remote(obj))
3854 rc = mdt_remote_object_lock_try(info, obj, &lh->mlh_rreg_lh,
3855 lh->mlh_rreg_mode, policy, res_id, false);
3857 rc = mdt_fid_lock(info->mti_env, ns, &lh->mlh_reg_lh,
3858 lh->mlh_reg_mode, policy, res_id, dlmflags, cookie);
3860 mdt_object_unlock(info, obj, lh, 1);
3861 else if (CFS_FAIL_PRECHECK(OBD_FAIL_MDS_PDO_LOCK) &&
3862 lh->mlh_pdo_hash != 0 &&
3863 (lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX))
3864 CFS_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK, 15);
3869 int mdt_object_lock_internal(struct mdt_thread_info *info,
3870 struct mdt_object *obj, const struct lu_fid *fid,
3871 struct mdt_lock_handle *lh, __u64 *ibits,
3872 __u64 trybits, bool cache)
3874 union ldlm_policy_data *policy = &info->mti_policy;
3875 struct ldlm_res_id *res_id = &info->mti_res_id;
3876 struct lustre_handle *handle;
3879 policy->l_inodebits.bits = *ibits;
3880 policy->l_inodebits.try_bits = trybits;
3881 policy->l_inodebits.li_gid = lh->mlh_gid;
3882 policy->l_inodebits.li_initiator_id = mdt_node_id(info->mti_mdt);
3883 fid_build_reg_res_name(fid, res_id);
3885 if (obj && mdt_object_remote(obj)) {
3886 handle = &lh->mlh_rreg_lh;
3887 LASSERT(!lustre_handle_is_used(handle));
3888 LASSERT(lh->mlh_rreg_mode != LCK_MINMODE);
3889 LASSERT(lh->mlh_type != MDT_NUL_LOCK);
3890 rc = mdt_remote_object_lock_try(info, obj, handle,
3891 lh->mlh_rreg_mode, policy,
3894 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
3896 * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if
3897 * it is going to be sent to client. If it is -
3898 * mdt_intent_policy() path will fix it up and turn FL_LOCAL
3901 __u64 dlmflags = LDLM_FL_ATOMIC_CB | LDLM_FL_LOCAL_ONLY;
3902 __u64 *cookie = NULL;
3904 handle = &lh->mlh_reg_lh;
3905 LASSERT(!lustre_handle_is_used(handle));
3906 LASSERT(lh->mlh_reg_mode != LCK_MINMODE);
3907 LASSERT(lh->mlh_type != MDT_NUL_LOCK);
3909 /* Lease lock are granted with LDLM_FL_CANCEL_ON_BLOCK */
3910 if (lh->mlh_type == MDT_REG_LOCK &&
3911 lh->mlh_reg_mode == LCK_EX && *ibits == MDS_INODELOCK_OPEN)
3912 dlmflags |= LDLM_FL_CANCEL_ON_BLOCK;
3916 cookie = &info->mti_exp->exp_handle.h_cookie;
3918 rc = mdt_fid_lock(info->mti_env, ns, handle, lh->mlh_reg_mode,
3919 policy, res_id, dlmflags, cookie);
3921 mdt_object_unlock(info, obj, lh, 1);
3925 struct ldlm_lock *lock;
3927 /* Return successfully acquired bits to a caller */
3928 lock = ldlm_handle2lock(handle);
3930 *ibits = lock->l_policy_data.l_inodebits.bits;
3931 LDLM_LOCK_PUT(lock);
3938 * MDT object locking functions:
3939 * mdt_object_lock(): lock object, this is used in most places, and normally
3940 * lock ibits doesn't contain LOOKUP, unless the caller knows it's not
3942 * mdt_object_check_lock(): lock object with LOOKUP and other ibits, it needs
3943 * to check whether parent is on remote MDT, if so, take LOOKUP on parent
3944 * MDT separately, and then lock other ibits on child object.
3945 * mdt_parent_lock(): take parent UPDATE lock with specific mode, if parent is
3946 * local, take PDO lock by name hash, otherwise take regular lock.
3947 * mdt_object_stripes_lock(): lock object which should be local, and if it's a
3948 * striped directory, lock its stripes, this is called in operations which
3949 * modify both object and stripes.
3950 * mdt_object_lock_try(): lock object with trybits, the trybits contains
3951 * optional inode lock bits that can be granted. This is called by
3952 * getattr/open to fetch more inode lock bits to client, and is also called
3953 * by dir migration to lock link parent in non-block mode to avoid
3960 * this is used to lock object in most places, and normally lock ibits doesn't
3961 * contain LOOKUP, unless the caller knows it's not remote object.
3963 * \param info struct mdt_thread_info
3965 * \param lh lock handle
3966 * \param ibits MDS inode lock bits
3967 * \param mode lock mode
3969 * \retval 0 on success, -ev on error.
3971 int mdt_object_lock(struct mdt_thread_info *info, struct mdt_object *obj,
3972 struct mdt_lock_handle *lh, __u64 ibits,
3973 enum ldlm_mode mode)
3978 mdt_lock_reg_init(lh, mode);
3979 rc = mdt_object_lock_internal(info, obj, mdt_object_fid(obj), lh,
3985 * lock object with LOOKUP and other ibits
3987 * it will check whether parent and child are on different MDTs, if so, take
3988 * LOOKUP lock on parent MDT, and lock other ibits on child MDT, otherwise lock
3989 * all ibits on child MDT. Note, parent and child shouldn't be both on remote
3990 * MDTs, in which case specific lock function should be used, and it's in
3991 * rename and migrate only.
3993 * \param info struct mdt_thread_info
3994 * \param parent parent object
3995 * \param child child object
3996 * \param lh lock handle
3997 * \param ibits MDS inode lock bits
3998 * \param mode lock mode
4000 * \retval 0 on success, -ev on error.
4002 int mdt_object_check_lock(struct mdt_thread_info *info,
4003 struct mdt_object *parent, struct mdt_object *child,
4004 struct mdt_lock_handle *lh, __u64 ibits,
4005 enum ldlm_mode mode)
4010 /* if LOOKUP ibit is not set, use mdt_object_lock() */
4011 LASSERT(ibits & MDS_INODELOCK_LOOKUP);
4012 /* if only LOOKUP ibit is needed, use mdt_object_lookup_lock() */
4013 LASSERT(ibits != MDS_INODELOCK_LOOKUP);
4015 /* @parent and @child shouldn't both be on remote MDTs */
4016 LASSERT(!(mdt_object_remote(parent) && mdt_object_remote(child)));
4018 mdt_lock_reg_init(lh, mode);
4019 if (mdt_object_remote(parent) ^ mdt_object_remote(child)) {
4020 __u64 lookup_ibits = MDS_INODELOCK_LOOKUP;
4022 rc = mdt_object_lock_internal(info, parent,
4023 mdt_object_fid(child), lh,
4024 &lookup_ibits, 0, false);
4028 ibits &= ~MDS_INODELOCK_LOOKUP;
4031 rc = mdt_object_lock_internal(info, child, mdt_object_fid(child), lh,
4033 if (rc && !(ibits & MDS_INODELOCK_LOOKUP))
4034 mdt_object_unlock(info, NULL, lh, 1);
4040 * take parent UPDATE lock
4042 * if parent is local or mode is LCK_PW, take PDO lock, otherwise take regular
4045 * \param info struct mdt_thread_info
4046 * \param obj parent object
4047 * \param lh lock handle
4048 * \param lname child name
4049 * \param mode lock mode
4051 * \retval 0 on success, -ev on error.
4053 int mdt_parent_lock(struct mdt_thread_info *info, struct mdt_object *obj,
4054 struct mdt_lock_handle *lh, const struct lu_name *lname,
4055 enum ldlm_mode mode)
4060 LASSERT(obj && lname);
4061 LASSERT(mode == LCK_PW || mode == LCK_PR);
4062 if (mdt_object_remote(obj) && mode == LCK_PR) {
4063 __u64 ibits = MDS_INODELOCK_UPDATE;
4065 mdt_lock_reg_init(lh, mode);
4066 rc = mdt_object_lock_internal(info, obj, mdt_object_fid(obj),
4067 lh, &ibits, 0, false);
4069 rc = mdt_object_pdo_lock(info, obj, lh, lname, mode, true);
4075 * lock object with trybits
4077 * the trybits contains optional inode lock bits that can be granted. This is
4078 * called by getattr/open to fetch more inode lock bits to client, and is also
4079 * called by dir migration to lock link parent in non-block mode to avoid
4082 * \param info struct mdt_thread_info
4084 * \param lh lock handle
4085 * \param ibits MDS inode lock bits
4086 * \param trybits optional inode lock bits
4087 * \param mode lock mode
4089 * \retval 0 on success, -ev on error.
4091 int mdt_object_lock_try(struct mdt_thread_info *info, struct mdt_object *obj,
4092 struct mdt_lock_handle *lh, __u64 *ibits,
4093 __u64 trybits, enum ldlm_mode mode)
4095 bool trylock_only = *ibits == 0;
4099 LASSERT(!(*ibits & trybits));
4100 mdt_lock_reg_init(lh, mode);
4101 rc = mdt_object_lock_internal(info, obj, mdt_object_fid(obj), lh, ibits,
4103 if (rc && trylock_only) { /* clear error for try ibits lock only */
4104 LASSERT(*ibits == 0);
4111 * Helper function to take \a obj LOOKUP lock.
4113 * Both \a pobj and \a obj may be located on remote MDTs.
4115 int mdt_object_lookup_lock(struct mdt_thread_info *info,
4116 struct mdt_object *pobj, struct mdt_object *obj,
4117 struct mdt_lock_handle *lh, enum ldlm_mode mode)
4119 __u64 ibits = MDS_INODELOCK_LOOKUP;
4123 /* if @parent is NULL, it's on local MDT, and @child is remote,
4124 * this is case in getattr/unlink/open by name.
4126 LASSERT(ergo(!pobj, mdt_object_remote(obj)));
4127 mdt_lock_reg_init(lh, mode);
4128 rc = mdt_object_lock_internal(info, pobj, mdt_object_fid(obj), lh,
4134 * Save a lock within request object.
4136 * Keep the lock referenced until whether client ACK or transaction
4137 * commit happens or release the lock immediately depending on input
4138 * parameters. If COS is ON, a write lock is converted to COS lock
4141 * \param info thead info object
4142 * \param h lock handle
4143 * \param mode lock mode
4144 * \param decref force immediate lock releasing
4146 static void mdt_save_lock(struct mdt_thread_info *info, struct lustre_handle *h,
4147 enum ldlm_mode mode, int decref)
4149 struct tgt_session_info *tsi = info->mti_env->le_ses ?
4150 tgt_ses_info(info->mti_env) : NULL;
4154 if (lustre_handle_is_used(h)) {
4155 bool has_trans = tsi && tsi->tsi_has_trans;
4157 if (decref || !has_trans || !(mode & (LCK_PW | LCK_EX))) {
4158 mdt_fid_unlock(h, mode);
4160 struct mdt_device *mdt = info->mti_mdt;
4161 struct ldlm_lock *lock = ldlm_handle2lock(h);
4162 struct ptlrpc_request *req = mdt_info_req(info);
4163 bool no_ack = false;
4165 LASSERTF(lock != NULL, "no lock for cookie %#llx\n",
4168 /* there is no request if mdt_object_unlock() is called
4169 * from mdt_export_cleanup()->mdt_add_dirty_flag() */
4170 if (likely(req != NULL)) {
4171 LDLM_DEBUG(lock, "save lock request %p reply "
4172 "state %p transno %lld\n", req,
4173 req->rq_reply_state, req->rq_transno);
4174 if (mdt_cos_is_enabled(mdt)) {
4177 ldlm_lock_mode_downgrade(lock, mode);
4178 } else if (mdt_slc_is_enabled(mdt)) {
4180 if (mode != LCK_TXN) {
4182 ldlm_lock_mode_downgrade(lock,
4186 if (req->rq_export->exp_disconnected)
4187 mdt_fid_unlock(h, mode);
4189 ptlrpc_save_lock(req, h, mode, no_ack);
4191 mdt_fid_unlock(h, mode);
4194 if (mdt_is_lock_sync(lock)) {
4195 CDEBUG(D_HA, "sync_lock, do async commit\n");
4196 mdt_device_commit_async(info->mti_env, mdt);
4198 LDLM_LOCK_PUT(lock);
4207 * Save cross-MDT lock in uncommitted_slc_locks
4209 * Keep the lock referenced until transaction commit happens or release the lock
4210 * immediately depending on input parameters.
4212 * \param info thead info object
4213 * \param h lock handle
4214 * \param mode lock mode
4215 * \param decref force immediate lock releasing
4217 static void mdt_save_remote_lock(struct mdt_thread_info *info,
4218 struct mdt_object *o, struct lustre_handle *h,
4219 enum ldlm_mode mode, int decref)
4223 if (lustre_handle_is_used(h)) {
4224 struct ldlm_lock *lock = ldlm_handle2lock(h);
4225 struct ptlrpc_request *req = mdt_info_req(info);
4228 (lock->l_policy_data.l_inodebits.bits &
4229 (MDS_INODELOCK_XATTR | MDS_INODELOCK_UPDATE)))
4230 mo_invalidate(info->mti_env, mdt_object_child(o));
4232 if (decref || !req || !(mode & (LCK_PW | LCK_EX)) ||
4233 !tgt_ses_info(info->mti_env)->tsi_has_trans) {
4234 ldlm_lock_decref_and_cancel(h, mode);
4235 LDLM_LOCK_PUT(lock);
4237 tgt_save_slc_lock(&info->mti_mdt->mdt_lut, lock,
4239 ldlm_lock_decref(h, mode);
4248 * Unlock mdt object.
4250 * Immeditely release the regular lock and the PDO lock or save the
4251 * lock in request and keep them referenced until client ACK or
4252 * transaction commit.
4254 * \param info thread info object
4255 * \param o mdt object
4256 * \param lh mdt lock handle referencing regular and PDO locks
4257 * \param decref force immediate lock releasing
4259 * XXX o is not used and may be NULL, see hsm_cdt_request_completed().
4261 void mdt_object_unlock(struct mdt_thread_info *info, struct mdt_object *o,
4262 struct mdt_lock_handle *lh, int decref)
4266 if (lh->mlh_pdo_remote)
4267 mdt_save_remote_lock(info, o, &lh->mlh_pdo_lh,
4268 lh->mlh_pdo_mode, decref);
4270 mdt_save_lock(info, &lh->mlh_pdo_lh, lh->mlh_pdo_mode, decref);
4271 mdt_save_lock(info, &lh->mlh_reg_lh, lh->mlh_reg_mode, decref);
4272 mdt_save_remote_lock(info, o, &lh->mlh_rreg_lh, lh->mlh_rreg_mode,
4278 struct mdt_object *mdt_object_find_lock(struct mdt_thread_info *info,
4279 const struct lu_fid *f,
4280 struct mdt_lock_handle *lh,
4281 __u64 ibits, enum ldlm_mode mode)
4283 struct mdt_object *o;
4285 o = mdt_object_find(info->mti_env, info->mti_mdt, f);
4289 rc = mdt_object_lock(info, o, lh, ibits, mode);
4291 mdt_object_put(info->mti_env, o);
4298 void mdt_object_unlock_put(struct mdt_thread_info * info,
4299 struct mdt_object *o,
4300 struct mdt_lock_handle *lh,
4303 mdt_object_unlock(info, o, lh, decref);
4304 mdt_object_put(info->mti_env, o);
4308 * Generic code handling requests that have struct mdt_body passed in:
4310 * - extract mdt_body from request and save it in @info, if present;
4312 * - create lu_object, corresponding to the fid in mdt_body, and save it in
4315 * - if HAS_BODY flag is set for this request type check whether object
4316 * actually exists on storage (lu_object_exists()).
4319 static int mdt_body_unpack(struct mdt_thread_info *info,
4320 enum tgt_handler_flags flags)
4322 const struct mdt_body *body;
4323 struct mdt_object *obj;
4324 const struct lu_env *env;
4325 struct req_capsule *pill;
4329 env = info->mti_env;
4330 pill = info->mti_pill;
4332 body = info->mti_body = req_capsule_client_get(pill, &RMF_MDT_BODY);
4336 if (!(body->mbo_valid & OBD_MD_FLID))
4339 if (!fid_is_sane(&body->mbo_fid1)) {
4340 CERROR("Invalid fid: "DFID"\n", PFID(&body->mbo_fid1));
4344 obj = mdt_object_find(env, info->mti_mdt, &body->mbo_fid1);
4346 if ((flags & HAS_BODY) && !mdt_object_exists(obj)) {
4347 mdt_object_put(env, obj);
4350 info->mti_object = obj;
4359 static int mdt_unpack_req_pack_rep(struct mdt_thread_info *info,
4360 enum tgt_handler_flags flags)
4362 struct req_capsule *pill = info->mti_pill;
4367 if (req_capsule_has_field(pill, &RMF_MDT_BODY, RCL_CLIENT))
4368 rc = mdt_body_unpack(info, flags);
4372 if (rc == 0 && (flags & HAS_REPLY)) {
4374 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
4375 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER,
4376 req_capsule_ptlreq(pill) ?
4377 DEF_REP_MD_SIZE : MAX_MD_SIZE_OLD);
4379 if (req_capsule_has_field(pill, &RMF_LOGCOOKIES, RCL_SERVER))
4380 req_capsule_set_size(pill, &RMF_LOGCOOKIES,
4383 /* Set ACL reply buffer size as LUSTRE_POSIX_ACL_MAX_SIZE_OLD
4384 * by default. If the target object has more ACL entries, then
4385 * enlarge the buffer when necessary. */
4386 if (req_capsule_has_field(pill, &RMF_ACL, RCL_SERVER))
4387 req_capsule_set_size(pill, &RMF_ACL, RCL_SERVER,
4388 LUSTRE_POSIX_ACL_MAX_SIZE_OLD);
4390 mdt_preset_secctx_size(info);
4391 mdt_preset_encctx_size(info);
4393 rc = req_capsule_server_pack(pill);
4395 CWARN("%s: cannot pack response: rc = %d\n",
4396 mdt_obd_name(info->mti_mdt), rc);
4401 void mdt_thread_info_reset(struct mdt_thread_info *info)
4403 memset(&info->mti_attr, 0, sizeof(info->mti_attr));
4404 info->mti_body = NULL;
4405 info->mti_dlm_req = NULL;
4406 info->mti_cross_ref = 0;
4407 info->mti_opdata = 0;
4408 info->mti_big_lmm_used = 0;
4409 info->mti_big_acl_used = 0;
4410 info->mti_som_strict = 0;
4412 info->mti_spec.no_create = 0;
4413 info->mti_spec.sp_rm_entry = 0;
4414 info->mti_spec.sp_permitted = 0;
4416 info->mti_spec.u.sp_ea.eadata = NULL;
4417 info->mti_spec.u.sp_ea.eadatalen = 0;
4419 if (info->mti_batch_env && info->mti_object != NULL) {
4420 mdt_object_put(info->mti_env, info->mti_object);
4421 info->mti_object = NULL;
4426 * Initialize fields of struct mdt_thread_info. Other fields are left in
4427 * uninitialized state, because it's too expensive to zero out whole
4428 * mdt_thread_info (> 1K) on each request arrival.
4430 void mdt_thread_info_init(struct ptlrpc_request *req,
4431 struct mdt_thread_info *info)
4433 info->mti_pill = &req->rq_pill;
4435 /* mdt device: it can be NULL while CONNECT */
4436 if (req->rq_export) {
4437 info->mti_mdt = mdt_dev(req->rq_export->exp_obd->obd_lu_dev);
4438 info->mti_exp = req->rq_export;
4440 info->mti_mdt = NULL;
4441 info->mti_env = req->rq_svc_thread->t_env;
4442 info->mti_transno = lustre_msg_get_transno(req->rq_reqmsg);
4443 info->mti_big_buf = LU_BUF_NULL;
4444 info->mti_batch_env = 0;
4445 info->mti_object = NULL;
4447 mdt_thread_info_reset(info);
4450 void mdt_thread_info_fini(struct mdt_thread_info *info)
4454 if (info->mti_object != NULL) {
4455 mdt_object_put(info->mti_env, info->mti_object);
4456 info->mti_object = NULL;
4459 for (i = 0; i < ARRAY_SIZE(info->mti_lh); i++)
4460 mdt_lock_handle_assert(&info->mti_lh[i]);
4461 info->mti_env = NULL;
4462 info->mti_pill = NULL;
4463 info->mti_exp = NULL;
4464 info->mti_mdt = NULL;
4466 if (unlikely(info->mti_big_buf.lb_buf != NULL))
4467 lu_buf_free(&info->mti_big_buf);
4470 struct mdt_thread_info *tsi2mdt_info(struct tgt_session_info *tsi)
4472 struct mdt_thread_info *mti;
4474 mti = mdt_th_info(tsi->tsi_env);
4475 LASSERT(mti != NULL);
4477 mdt_thread_info_init(tgt_ses_req(tsi), mti);
4478 if (tsi->tsi_corpus != NULL) {
4479 mti->mti_object = mdt_obj(tsi->tsi_corpus);
4480 lu_object_get(tsi->tsi_corpus);
4482 mti->mti_body = tsi->tsi_mdt_body;
4483 mti->mti_dlm_req = tsi->tsi_dlm_req;
4488 static int mdt_tgt_connect(struct tgt_session_info *tsi)
4490 if (CFS_FAIL_CHECK(OBD_FAIL_TGT_DELAY_CONDITIONAL) &&
4492 tsi2mdt_info(tsi)->mti_mdt->mdt_seq_site.ss_node_id)
4493 schedule_timeout_uninterruptible(cfs_time_seconds(3));
4495 return tgt_connect(tsi);
4498 static int mdt_intent_glimpse(enum ldlm_intent_flags it_opc,
4499 struct mdt_thread_info *info,
4500 struct ldlm_lock **lockp, __u64 flags)
4502 return mdt_glimpse_enqueue(info, info->mti_mdt->mdt_namespace,
4505 static int mdt_intent_brw(enum ldlm_intent_flags it_opc,
4506 struct mdt_thread_info *info,
4507 struct ldlm_lock **lockp, __u64 flags)
4509 return mdt_brw_enqueue(info, info->mti_mdt->mdt_namespace,
4513 int mdt_intent_lock_replace(struct mdt_thread_info *info,
4514 struct ldlm_lock **lockp,
4515 struct mdt_lock_handle *lh,
4516 __u64 flags, int result)
4518 struct ptlrpc_request *req = mdt_info_req(info);
4519 struct ldlm_lock *lock = *lockp;
4520 struct ldlm_lock *new_lock;
4522 /* If possible resent found a lock, @lh is set to its handle */
4523 new_lock = ldlm_handle2lock_long(&lh->mlh_reg_lh, 0);
4525 if (new_lock == NULL) {
4526 if (flags & LDLM_FL_INTENT_ONLY) {
4528 } else if (flags & LDLM_FL_RESENT) {
4529 /* Lock is pinned by ldlm_handle_enqueue0() as it is a
4530 * resend case, however, it could be already destroyed
4531 * due to client eviction or a raced cancel RPC.
4533 LDLM_DEBUG_NOLOCK("Invalid lock handle %#llx\n",
4534 lh->mlh_reg_lh.cookie);
4537 CERROR("%s: Invalid lockh=%#llx flags=%#llx fid1="DFID" fid2="DFID": rc = %d\n",
4538 mdt_obd_name(info->mti_mdt),
4539 lh->mlh_reg_lh.cookie, flags,
4540 PFID(&info->mti_tmp_fid1),
4541 PFID(&info->mti_tmp_fid2), result);
4544 lh->mlh_reg_lh.cookie = 0;
4549 * If we've already given this lock to a client once, then we should
4550 * have no readers or writers. Otherwise, we should have one reader
4551 * _or_ writer ref (which will be zeroed below) before returning the
4554 if (new_lock->l_export == req->rq_export) {
4555 LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
4557 LASSERT(new_lock->l_export == NULL);
4558 LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
4563 if (new_lock->l_export == req->rq_export) {
4565 * Already gave this to the client, which means that we
4566 * reconstructed a reply.
4568 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
4571 LDLM_LOCK_RELEASE(new_lock);
4572 lh->mlh_reg_lh.cookie = 0;
4573 RETURN(ELDLM_LOCK_REPLACED);
4577 * Fixup the lock to be given to the client.
4579 lock_res_and_lock(new_lock);
4580 /* Zero new_lock->l_readers and new_lock->l_writers without triggering
4581 * possible blocking AST. */
4582 while (new_lock->l_readers > 0) {
4583 lu_ref_del(&new_lock->l_reference, "reader", new_lock);
4584 lu_ref_del(&new_lock->l_reference, "user", new_lock);
4585 new_lock->l_readers--;
4587 while (new_lock->l_writers > 0) {
4588 lu_ref_del(&new_lock->l_reference, "writer", new_lock);
4589 lu_ref_del(&new_lock->l_reference, "user", new_lock);
4590 new_lock->l_writers--;
4593 new_lock->l_export = class_export_lock_get(req->rq_export, new_lock);
4594 new_lock->l_blocking_ast = lock->l_blocking_ast;
4595 new_lock->l_completion_ast = lock->l_completion_ast;
4596 if (ldlm_has_dom(new_lock))
4597 new_lock->l_glimpse_ast = ldlm_server_glimpse_ast;
4598 new_lock->l_remote_handle = lock->l_remote_handle;
4599 new_lock->l_flags &= ~LDLM_FL_LOCAL;
4601 unlock_res_and_lock(new_lock);
4603 cfs_hash_add(new_lock->l_export->exp_lock_hash,
4604 &new_lock->l_remote_handle,
4605 &new_lock->l_exp_hash);
4607 LDLM_LOCK_RELEASE(new_lock);
4608 lh->mlh_reg_lh.cookie = 0;
4610 RETURN(ELDLM_LOCK_REPLACED);
4613 void mdt_intent_fixup_resent(struct mdt_thread_info *info,
4614 struct ldlm_lock *new_lock,
4615 struct mdt_lock_handle *lh, __u64 flags)
4617 struct ptlrpc_request *req = mdt_info_req(info);
4618 struct ldlm_request *dlmreq;
4620 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
4623 dlmreq = req_capsule_client_get(info->mti_pill, &RMF_DLM_REQ);
4625 /* Check if this is a resend case (MSG_RESENT is set on RPC) and a
4626 * lock was found by ldlm_handle_enqueue(); if so @lh must be
4628 if (flags & LDLM_FL_RESENT) {
4629 lh->mlh_reg_lh.cookie = new_lock->l_handle.h_cookie;
4630 lh->mlh_reg_mode = new_lock->l_granted_mode;
4632 LDLM_DEBUG(new_lock, "Restoring lock cookie");
4633 DEBUG_REQ(D_DLMTRACE, req, "restoring lock cookie %#llx",
4634 lh->mlh_reg_lh.cookie);
4639 * If the xid matches, then we know this is a resent request, and allow
4640 * it. (It's probably an OPEN, for which we don't send a lock.
4642 if (req_can_reconstruct(req, NULL) != 0)
4646 * This remote handle isn't enqueued, so we never received or processed
4647 * this request. Clear MSG_RESENT, because it can be handled like any
4648 * normal request now.
4650 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
4652 DEBUG_REQ(D_DLMTRACE, req, "no existing lock with rhandle %#llx",
4653 dlmreq->lock_handle[0].cookie);
4656 static int mdt_intent_getxattr(enum ldlm_intent_flags it_opc,
4657 struct mdt_thread_info *info,
4658 struct ldlm_lock **lockp,
4661 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
4662 struct ldlm_reply *ldlm_rep = NULL;
4667 * Initialize lhc->mlh_reg_lh either from a previously granted lock
4668 * (for the resend case) or a new lock. Below we will use it to
4669 * replace the original lock.
4671 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
4672 if (!lustre_handle_is_used(&lhc->mlh_reg_lh)) {
4673 rc = mdt_object_lock(info, info->mti_object, lhc,
4674 MDS_INODELOCK_XATTR, (*lockp)->l_req_mode);
4679 rc = mdt_getxattr(info);
4681 if (mdt_info_req(info)->rq_repmsg != NULL)
4682 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
4684 if (ldlm_rep == NULL ||
4685 CFS_FAIL_CHECK(OBD_FAIL_MDS_XATTR_REP)) {
4686 mdt_object_unlock(info, info->mti_object, lhc, 1);
4690 RETURN(err_serious(-EFAULT));
4693 ldlm_rep->lock_policy_res2 = clear_serious(rc);
4695 /* This is left for interop instead of adding a new interop flag.
4697 #if LUSTRE_VERSION_CODE > OBD_OCD_VERSION(3, 0, 0, 0)
4698 if (ldlm_rep->lock_policy_res2) {
4699 mdt_object_unlock(info, info->mti_object, lhc, 1);
4700 RETURN(ELDLM_LOCK_ABORTED);
4704 rc = mdt_intent_lock_replace(info, lockp, lhc, flags, rc);
4708 static int mdt_intent_getattr(enum ldlm_intent_flags it_opc,
4709 struct mdt_thread_info *info,
4710 struct ldlm_lock **lockp,
4713 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
4715 struct ldlm_reply *ldlm_rep;
4716 struct mdt_body *reqbody;
4717 struct mdt_body *repbody;
4721 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
4724 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
4727 info->mti_cross_ref = !!(reqbody->mbo_valid & OBD_MD_FLCROSSREF);
4728 repbody->mbo_eadatasize = 0;
4729 repbody->mbo_aclsize = 0;
4733 child_bits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
4736 child_bits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE |
4740 CERROR("%s: unsupported intent %#x\n",
4741 mdt_obd_name(info->mti_mdt), (unsigned int)it_opc);
4742 GOTO(out_shrink, rc = -EINVAL);
4745 rc = mdt_init_ucred(info, reqbody);
4747 GOTO(out_shrink, rc);
4749 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
4750 mdt_set_disposition(info, ldlm_rep, DISP_IT_EXECD);
4752 /* Get lock from request for possible resent case. */
4753 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
4755 rc = mdt_getattr_name_lock(info, lhc, child_bits, ldlm_rep);
4756 ldlm_rep->lock_policy_res2 = clear_serious(rc);
4758 if (mdt_get_disposition(ldlm_rep, DISP_LOOKUP_NEG) &&
4759 ldlm_rep->lock_policy_res2 != -ENOKEY)
4760 ldlm_rep->lock_policy_res2 = 0;
4761 if (!mdt_get_disposition(ldlm_rep, DISP_LOOKUP_POS) ||
4762 ldlm_rep->lock_policy_res2) {
4763 lhc->mlh_reg_lh.cookie = 0ull;
4764 GOTO(out_ucred, rc = ELDLM_LOCK_ABORTED);
4767 rc = mdt_intent_lock_replace(info, lockp, lhc, flags, rc);
4770 mdt_exit_ucred(info);
4772 mdt_client_compatibility(info);
4773 rc2 = mdt_fix_reply(info);
4779 static int mdt_intent_layout(enum ldlm_intent_flags it_opc,
4780 struct mdt_thread_info *info,
4781 struct ldlm_lock **lockp,
4784 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
4785 struct md_layout_change layout = { .mlc_opc = MD_LAYOUT_NOP };
4786 struct layout_intent *intent;
4787 struct ldlm_reply *ldlm_rep;
4788 struct lu_fid *fid = &info->mti_tmp_fid2;
4789 struct mdt_object *obj = NULL;
4790 int layout_size = 0;
4791 struct lu_buf *buf = &layout.mlc_buf;
4796 fid_extract_from_res_name(fid, &(*lockp)->l_resource->lr_name);
4798 intent = req_capsule_client_get(info->mti_pill, &RMF_LAYOUT_INTENT);
4802 CDEBUG(D_INFO, DFID "got layout change request from client: "
4803 "opc:%u flags:%#x extent "DEXT"\n",
4804 PFID(fid), intent->li_opc, intent->li_flags,
4805 PEXT(&intent->li_extent));
4807 switch (intent->li_opc) {
4808 case LAYOUT_INTENT_TRUNC:
4809 case LAYOUT_INTENT_WRITE:
4810 layout.mlc_opc = MD_LAYOUT_WRITE;
4811 layout.mlc_intent = intent;
4813 case LAYOUT_INTENT_ACCESS:
4815 case LAYOUT_INTENT_READ:
4816 case LAYOUT_INTENT_GLIMPSE:
4817 case LAYOUT_INTENT_RELEASE:
4818 case LAYOUT_INTENT_RESTORE:
4819 CERROR("%s: Unsupported layout intent opc %d\n",
4820 mdt_obd_name(info->mti_mdt), intent->li_opc);
4823 CERROR("%s: Unknown layout intent opc %d\n",
4824 mdt_obd_name(info->mti_mdt), intent->li_opc);
4828 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
4830 RETURN(PTR_ERR(obj));
4832 if (mdt_object_exists(obj) && !mdt_object_remote(obj)) {
4833 /* if layout is going to be changed don't use the current EA
4834 * size but the maximum one. That buffer will be shrinked
4835 * to the actual size in req_capsule_shrink() before reply.
4837 if (layout.mlc_opc == MD_LAYOUT_WRITE) {
4838 layout_size = info->mti_mdt->mdt_max_mdsize;
4840 layout_size = mdt_attr_get_eabuf_size(info, obj);
4841 if (layout_size < 0)
4842 GOTO(out, rc = layout_size);
4844 if (layout_size > info->mti_mdt->mdt_max_mdsize)
4845 info->mti_mdt->mdt_max_mdsize = layout_size;
4847 CDEBUG(D_INFO, "%s: layout_size %d\n",
4848 mdt_obd_name(info->mti_mdt), layout_size);
4852 * set reply buffer size, so that ldlm_handle_enqueue0()->
4853 * ldlm_lvbo_fill() will fill the reply buffer with lovea.
4855 req_capsule_set_size(info->mti_pill, &RMF_DLM_LVB, RCL_SERVER,
4857 rc = req_capsule_server_pack(info->mti_pill);
4861 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
4863 GOTO(out, rc = -EPROTO);
4865 mdt_set_disposition(info, ldlm_rep, DISP_IT_EXECD);
4867 /* take lock in ldlm_lock_enqueue() for LAYOUT_INTENT_ACCESS */
4868 if (layout.mlc_opc == MD_LAYOUT_NOP)
4871 rc = mdt_check_resent(info, mdt_reconstruct_generic, lhc);
4875 DEBUG_REQ(D_INODE, mdt_info_req(info), "resent opt.");
4876 rc = lustre_msg_get_status(mdt_info_req(info)->rq_repmsg);
4882 if (unlikely(req_is_replay(mdt_info_req(info)))) {
4883 buf->lb_buf = req_capsule_client_get(info->mti_pill,
4885 buf->lb_len = req_capsule_get_size(info->mti_pill,
4886 &RMF_EADATA, RCL_CLIENT);
4888 * If it's a replay of layout write intent RPC, the client has
4889 * saved the extended lovea when it get reply then.
4891 if (buf->lb_len > 0)
4892 mdt_fix_lov_magic(info, buf->lb_buf);
4895 /* Get lock from request for possible resent case. */
4896 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
4897 (*lockp)->l_lvb_type = LVB_T_LAYOUT;
4900 * Instantiate some layout components, if @buf contains lovea, then it's
4901 * a replay of the layout intent write RPC.
4903 rc = mdt_layout_change(info, obj, lhc, &layout);
4904 ldlm_rep->lock_policy_res2 = clear_serious(rc);
4906 if (lustre_handle_is_used(&lhc->mlh_reg_lh)) {
4907 rc = mdt_intent_lock_replace(info, lockp, lhc, flags, rc);
4908 if (rc == ELDLM_LOCK_REPLACED &&
4909 (*lockp)->l_granted_mode == LCK_EX)
4910 ldlm_lock_mode_downgrade(*lockp, LCK_CR);
4915 mdt_object_put(info->mti_env, obj);
4919 static int mdt_intent_open(enum ldlm_intent_flags it_opc,
4920 struct mdt_thread_info *info,
4921 struct ldlm_lock **lockp,
4924 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
4925 struct ldlm_reply *rep = NULL;
4928 struct ptlrpc_request *req = mdt_info_req(info);
4930 static const struct req_format *intent_fmts[REINT_MAX] = {
4931 [REINT_CREATE] = &RQF_LDLM_INTENT_CREATE,
4932 [REINT_OPEN] = &RQF_LDLM_INTENT_OPEN
4937 opc = mdt_reint_opcode(mdt_info_req(info), intent_fmts);
4941 /* Get lock from request for possible resent case. */
4942 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
4944 rc = mdt_reint_internal(info, lhc, opc);
4946 if (rc < 0 && lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
4947 DEBUG_REQ(D_ERROR, req, "Replay open failed with %d", rc);
4949 /* Check whether the reply has been packed successfully. */
4950 if (mdt_info_req(info)->rq_repmsg != NULL)
4951 rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
4956 RETURN(err_serious(-EFAULT));
4959 /* MDC expects this in any case */
4961 mdt_set_disposition(info, rep, DISP_LOOKUP_EXECD);
4963 /* the open lock or the lock for cross-ref object should be
4964 * returned to the client */
4965 if (lustre_handle_is_used(&lhc->mlh_reg_lh) &&
4966 (rc == 0 || rc == -MDT_EREMOTE_OPEN)) {
4967 rep->lock_policy_res2 = 0;
4968 rc = mdt_intent_lock_replace(info, lockp, lhc, flags, rc);
4972 rep->lock_policy_res2 = clear_serious(rc);
4974 if (rep->lock_policy_res2 == -ENOENT &&
4975 mdt_get_disposition(rep, DISP_LOOKUP_NEG) &&
4976 !mdt_get_disposition(rep, DISP_OPEN_CREATE))
4977 rep->lock_policy_res2 = 0;
4979 lhc->mlh_reg_lh.cookie = 0ull;
4980 if (rc == -ENOTCONN || rc == -ENODEV ||
4981 rc == -EOVERFLOW) { /**< if VBR failure then return error */
4983 * If it is the disconnect error (ENODEV & ENOCONN), the error
4984 * will be returned by rq_status, and client at ptlrpc layer
4985 * will detect this, then disconnect, reconnect the import
4986 * immediately, instead of impacting the following the rpc.
4991 * For other cases, the error will be returned by intent, and client
4992 * will retrieve the result from intent.
4994 RETURN(ELDLM_LOCK_ABORTED);
4997 static int mdt_intent_opc(enum ldlm_intent_flags it_opc,
4998 struct mdt_thread_info *info,
4999 struct ldlm_lock **lockp,
5000 u64 flags /* LDLM_FL_* */)
5002 struct req_capsule *pill = info->mti_pill;
5003 struct ptlrpc_request *req = mdt_info_req(info);
5004 const struct req_format *it_format;
5005 int (*it_handler)(enum ldlm_intent_flags,
5006 struct mdt_thread_info *,
5007 struct ldlm_lock **,
5009 enum tgt_handler_flags it_handler_flags = 0;
5010 struct ldlm_reply *rep;
5011 bool check_mdt_object = false;
5017 case IT_OPEN|IT_CREAT:
5019 * OCREAT is not a IS_MUTABLE request since the file may
5020 * already exist. We do the extra check of
5021 * OBD_CONNECT_RDONLY in mdt_reint_open() when we
5022 * really need to create the object.
5024 it_format = &RQF_LDLM_INTENT;
5025 it_handler = &mdt_intent_open;
5028 check_mdt_object = true;
5031 it_format = &RQF_LDLM_INTENT_GETATTR;
5032 it_handler = &mdt_intent_getattr;
5033 it_handler_flags = HAS_REPLY;
5036 check_mdt_object = true;
5037 it_format = &RQF_LDLM_INTENT_GETXATTR;
5038 it_handler = &mdt_intent_getxattr;
5039 it_handler_flags = HAS_BODY;
5042 it_format = &RQF_LDLM_INTENT_LAYOUT;
5043 it_handler = &mdt_intent_layout;
5046 it_format = &RQF_LDLM_INTENT;
5047 it_handler = &mdt_intent_glimpse;
5050 it_format = &RQF_LDLM_INTENT;
5051 it_handler = &mdt_intent_brw;
5053 case IT_QUOTA_DQACQ:
5054 case IT_QUOTA_CONN: {
5055 struct lu_device *qmt = info->mti_mdt->mdt_qmt_dev;
5058 RETURN(-EOPNOTSUPP);
5060 if (mdt_rdonly(req->rq_export))
5063 (*lockp)->l_lvb_type = LVB_T_LQUOTA;
5064 /* pass the request to quota master */
5065 rc = qmt_hdls.qmth_intent_policy(info->mti_env, qmt,
5066 mdt_info_req(info), lockp,
5071 CERROR("%s: unknown intent code %#x\n",
5072 mdt_obd_name(info->mti_mdt), it_opc);
5076 if (!info->mti_batch_env)
5077 req_capsule_extend(pill, it_format);
5079 rc = mdt_unpack_req_pack_rep(info, it_handler_flags);
5083 if (unlikely(info->mti_object == NULL && check_mdt_object))
5086 if (it_handler_flags & IS_MUTABLE && mdt_rdonly(req->rq_export))
5089 CFS_FAIL_TIMEOUT(OBD_FAIL_MDS_INTENT_DELAY, 10);
5091 /* execute policy */
5092 rc = (*it_handler)(it_opc, info, lockp, flags);
5094 /* Check whether the reply has been packed successfully. */
5095 if (info->mti_batch_env || req->rq_repmsg != NULL) {
5096 rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
5097 rep->lock_policy_res2 =
5098 ptlrpc_status_hton(rep->lock_policy_res2);
5104 static void mdt_ptlrpc_stats_update(struct ptlrpc_request *req,
5105 enum ldlm_intent_flags it_opc)
5107 struct lprocfs_stats *srv_stats = ptlrpc_req2svc(req)->srv_stats;
5109 /* update stats when IT code is known */
5110 if (srv_stats != NULL)
5111 lprocfs_counter_incr(srv_stats,
5112 PTLRPC_LAST_CNTR + (it_opc == IT_GLIMPSE ?
5113 LDLM_GLIMPSE_ENQUEUE : LDLM_IBITS_ENQUEUE));
5116 static int mdt_intent_policy(const struct lu_env *env,
5117 struct ldlm_namespace *ns,
5118 struct ldlm_lock **lockp,
5120 enum ldlm_mode mode,
5121 __u64 flags, void *data)
5123 struct tgt_session_info *tsi;
5124 struct mdt_thread_info *info;
5125 struct ptlrpc_request *req = req_cookie;
5126 struct ldlm_intent *it;
5127 struct req_capsule *pill;
5128 const struct ldlm_lock_desc *ldesc;
5133 LASSERT(req != NULL);
5135 tsi = tgt_ses_info(env);
5137 info = mdt_th_info(env);
5138 LASSERT(info != NULL);
5140 /* Check whether it is a sub request processing in a batch request */
5141 if (info->mti_batch_env) {
5142 pill = info->mti_pill;
5143 LASSERT(pill == &info->mti_sub_pill);
5145 info = tsi2mdt_info(tsi);
5146 pill = info->mti_pill;
5149 LASSERT(pill->rc_req == req);
5150 ldesc = &info->mti_dlm_req->lock_desc;
5152 if (info->mti_batch_env ||
5153 req->rq_reqmsg->lm_bufcount > DLM_INTENT_IT_OFF) {
5155 * For batch processing environment, the request format has
5158 if (!info->mti_batch_env)
5159 req_capsule_extend(pill, &RQF_LDLM_INTENT_BASIC);
5161 it = req_capsule_client_get(pill, &RMF_LDLM_INTENT);
5163 mdt_ptlrpc_stats_update(req, it->opc);
5164 rc = mdt_intent_opc(it->opc, info, lockp, flags);
5168 /* Lock without inodebits makes no sense and will oops
5169 * later in ldlm. Let's check it now to see if we have
5170 * ibits corrupted somewhere in mdt_intent_opc().
5171 * The case for client miss to set ibits has been
5172 * processed by others. */
5173 LASSERT(ergo(ldesc->l_resource.lr_type == LDLM_IBITS,
5174 ldesc->l_policy_data.l_inodebits.bits != 0));
5176 rc = err_serious(-EFAULT);
5178 } else if (ldesc->l_resource.lr_type == LDLM_IBITS &&
5179 ldesc->l_policy_data.l_inodebits.bits == MDS_INODELOCK_DOM) {
5180 struct ldlm_reply *rep;
5182 /* No intent was provided but INTENT flag is set along with
5183 * DOM bit, this is considered as GLIMPSE request.
5184 * This logic is common for MDT and OST glimpse
5186 mdt_ptlrpc_stats_update(req, IT_GLIMPSE);
5187 rc = mdt_glimpse_enqueue(info, ns, lockp, flags);
5188 /* Check whether the reply has been packed successfully. */
5189 if (req->rq_repmsg != NULL) {
5190 rep = req_capsule_server_get(info->mti_pill,
5192 rep->lock_policy_res2 =
5193 ptlrpc_status_hton(rep->lock_policy_res2);
5196 /* No intent was provided */
5197 req_capsule_set_size(pill, &RMF_DLM_LVB, RCL_SERVER, 0);
5198 rc = req_capsule_server_pack(pill);
5200 rc = err_serious(rc);
5203 if (!info->mti_batch_env)
5204 mdt_thread_info_fini(info);
5208 static void mdt_deregister_seq_exp(struct mdt_device *mdt)
5210 struct seq_server_site *ss = mdt_seq_site(mdt);
5212 if (ss->ss_node_id == 0)
5215 if (ss->ss_client_seq != NULL) {
5216 lustre_deregister_lwp_item(&ss->ss_client_seq->lcs_exp);
5217 ss->ss_client_seq->lcs_exp = NULL;
5220 if (ss->ss_server_fld != NULL) {
5221 lustre_deregister_lwp_item(&ss->ss_server_fld->lsf_control_exp);
5222 ss->ss_server_fld->lsf_control_exp = NULL;
5226 static void mdt_seq_fini_cli(struct mdt_device *mdt)
5228 struct seq_server_site *ss = mdt_seq_site(mdt);
5233 if (ss->ss_server_seq != NULL)
5234 seq_server_set_cli(NULL, ss->ss_server_seq, NULL);
5237 static int mdt_seq_fini(const struct lu_env *env, struct mdt_device *mdt)
5239 mdt_seq_fini_cli(mdt);
5240 mdt_deregister_seq_exp(mdt);
5242 return seq_site_fini(env, mdt_seq_site(mdt));
5246 * It will retrieve its FLDB entries from MDT0, and it only happens
5247 * when upgrading existent FS to 2.6 or when local FLDB is corrupted,
5248 * and it needs to refresh FLDB from the MDT0.
5250 static int mdt_register_lwp_callback(void *data)
5253 struct mdt_device *mdt = data;
5254 struct lu_server_fld *fld = mdt_seq_site(mdt)->ss_server_fld;
5258 LASSERT(mdt_seq_site(mdt)->ss_node_id != 0);
5260 rc = lu_env_init(&env, LCT_MD_THREAD);
5262 CERROR("%s: cannot init env: rc = %d\n", mdt_obd_name(mdt), rc);
5266 /* Allocate new sequence now to avoid creating local transaction
5267 * in the normal transaction process */
5268 rc = seq_server_check_and_alloc_super(&env,
5269 mdt_seq_site(mdt)->ss_server_seq);
5274 rc = fld_update_from_controller(&env, fld);
5276 CERROR("%s: cannot update controller: rc = %d\n",
5277 mdt_obd_name(mdt), rc);
5286 static int mdt_register_seq_exp(struct mdt_device *mdt)
5288 struct seq_server_site *ss = mdt_seq_site(mdt);
5289 char *lwp_name = NULL;
5292 if (ss->ss_node_id == 0)
5295 OBD_ALLOC(lwp_name, MAX_OBD_NAME);
5296 if (lwp_name == NULL)
5297 GOTO(out_free, rc = -ENOMEM);
5299 rc = tgt_name2lwp_name(mdt_obd_name(mdt), lwp_name, MAX_OBD_NAME, 0);
5303 rc = lustre_register_lwp_item(lwp_name, &ss->ss_client_seq->lcs_exp,
5308 rc = lustre_register_lwp_item(lwp_name,
5309 &ss->ss_server_fld->lsf_control_exp,
5310 mdt_register_lwp_callback, mdt);
5312 lustre_deregister_lwp_item(&ss->ss_client_seq->lcs_exp);
5313 ss->ss_client_seq->lcs_exp = NULL;
5317 if (lwp_name != NULL)
5318 OBD_FREE(lwp_name, MAX_OBD_NAME);
5324 * Init client sequence manager which is used by local MDS to talk to sequence
5325 * controller on remote node.
5327 static int mdt_seq_init_cli(const struct lu_env *env, struct mdt_device *mdt)
5329 struct seq_server_site *ss = mdt_seq_site(mdt);
5333 /* check if this is adding the first MDC and controller is not yet
5335 OBD_ALLOC_PTR(ss->ss_client_seq);
5336 if (ss->ss_client_seq == NULL)
5339 OBD_ALLOC(prefix, MAX_OBD_NAME + 5);
5340 if (prefix == NULL) {
5341 OBD_FREE_PTR(ss->ss_client_seq);
5342 ss->ss_client_seq = NULL;
5346 /* Note: seq_client_fini will be called in seq_site_fini */
5347 snprintf(prefix, MAX_OBD_NAME + 5, "ctl-%s", mdt_obd_name(mdt));
5348 seq_client_init(ss->ss_client_seq, NULL, LUSTRE_SEQ_METADATA,
5349 prefix, ss->ss_node_id == 0 ? ss->ss_control_seq :
5351 OBD_FREE(prefix, MAX_OBD_NAME + 5);
5353 RETURN(seq_server_set_cli(env, ss->ss_server_seq, ss->ss_client_seq));
5356 static int mdt_seq_init(const struct lu_env *env, struct mdt_device *mdt)
5358 struct seq_server_site *ss;
5362 ss = mdt_seq_site(mdt);
5363 /* init sequence controller server(MDT0) */
5364 if (ss->ss_node_id == 0) {
5365 OBD_ALLOC_PTR(ss->ss_control_seq);
5366 if (ss->ss_control_seq == NULL)
5369 rc = seq_server_init(env, ss->ss_control_seq, mdt->mdt_bottom,
5370 mdt_obd_name(mdt), LUSTRE_SEQ_CONTROLLER,
5373 GOTO(out_seq_fini, rc);
5376 /* Init normal sequence server */
5377 OBD_ALLOC_PTR(ss->ss_server_seq);
5378 if (ss->ss_server_seq == NULL)
5379 GOTO(out_seq_fini, rc = -ENOMEM);
5381 rc = seq_server_init(env, ss->ss_server_seq, mdt->mdt_bottom,
5382 mdt_obd_name(mdt), LUSTRE_SEQ_SERVER, ss);
5384 GOTO(out_seq_fini, rc);
5386 /* init seq client for seq server to talk to seq controller(MDT0) */
5387 rc = mdt_seq_init_cli(env, mdt);
5389 GOTO(out_seq_fini, rc);
5391 if (ss->ss_node_id != 0)
5392 /* register controller export through lwp */
5393 rc = mdt_register_seq_exp(mdt);
5398 mdt_seq_fini(env, mdt);
5406 static int mdt_fld_fini(const struct lu_env *env,
5407 struct mdt_device *m)
5409 struct seq_server_site *ss = mdt_seq_site(m);
5412 if (ss && ss->ss_server_fld) {
5413 fld_server_fini(env, ss->ss_server_fld);
5414 OBD_FREE_PTR(ss->ss_server_fld);
5415 ss->ss_server_fld = NULL;
5421 static int mdt_fld_init(const struct lu_env *env,
5423 struct mdt_device *m)
5425 struct seq_server_site *ss;
5429 ss = mdt_seq_site(m);
5431 OBD_ALLOC_PTR(ss->ss_server_fld);
5432 if (ss->ss_server_fld == NULL)
5433 RETURN(rc = -ENOMEM);
5435 rc = fld_server_init(env, ss->ss_server_fld, m->mdt_bottom, uuid,
5438 OBD_FREE_PTR(ss->ss_server_fld);
5439 ss->ss_server_fld = NULL;
5446 static void mdt_stack_pre_fini(const struct lu_env *env,
5447 struct mdt_device *m, struct lu_device *top)
5449 struct lustre_cfg_bufs *bufs;
5450 struct lustre_cfg *lcfg;
5451 struct mdt_thread_info *info;
5456 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
5457 LASSERT(info != NULL);
5459 bufs = &info->mti_u.bufs;
5461 LASSERT(m->mdt_child_exp);
5462 LASSERT(m->mdt_child_exp->exp_obd);
5464 /* process cleanup, pass mdt obd name to get obd umount flags */
5465 /* XXX: this is needed because all layers are referenced by
5466 * objects (some of them are pinned by osd, for example *
5467 * the proper solution should be a model where object used
5468 * by osd only doesn't have mdt/mdd slices -bzzz */
5469 lustre_cfg_bufs_reset(bufs, mdt_obd_name(m));
5470 lustre_cfg_bufs_set_string(bufs, 1, NULL);
5471 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
5474 lustre_cfg_init(lcfg, LCFG_PRE_CLEANUP, bufs);
5476 top->ld_ops->ldo_process_config(env, top, lcfg);
5477 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens));
5481 static void mdt_stack_fini(const struct lu_env *env,
5482 struct mdt_device *m, struct lu_device *top)
5484 struct obd_device *obd = mdt2obd_dev(m);
5485 struct lustre_cfg_bufs *bufs;
5486 struct lustre_cfg *lcfg;
5487 struct mdt_thread_info *info;
5491 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
5492 LASSERT(info != NULL);
5494 lu_dev_del_linkage(top->ld_site, top);
5496 lu_site_purge(env, top->ld_site, -1);
5498 bufs = &info->mti_u.bufs;
5499 /* process cleanup, pass mdt obd name to get obd umount flags */
5500 /* another purpose is to let all layers to release their objects */
5501 lustre_cfg_bufs_reset(bufs, mdt_obd_name(m));
5506 lustre_cfg_bufs_set_string(bufs, 1, flags);
5507 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
5510 lustre_cfg_init(lcfg, LCFG_CLEANUP, bufs);
5513 top->ld_ops->ldo_process_config(env, top, lcfg);
5514 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens));
5516 lu_site_purge(env, top->ld_site, -1);
5518 m->mdt_child = NULL;
5519 m->mdt_bottom = NULL;
5521 obd_disconnect(m->mdt_child_exp);
5522 m->mdt_child_exp = NULL;
5524 obd_disconnect(m->mdt_bottom_exp);
5525 m->mdt_child_exp = NULL;
5528 static int mdt_connect_to_next(const struct lu_env *env, struct mdt_device *m,
5529 const char *next, struct obd_export **exp)
5531 struct obd_connect_data *data = NULL;
5532 struct obd_device *obd;
5536 OBD_ALLOC_PTR(data);
5538 GOTO(out, rc = -ENOMEM);
5540 obd = class_name2obd(next);
5542 CERROR("%s: can't locate next device: %s\n",
5543 mdt_obd_name(m), next);
5544 GOTO(out, rc = -ENOTCONN);
5547 data->ocd_connect_flags = OBD_CONNECT_VERSION;
5548 data->ocd_version = LUSTRE_VERSION_CODE;
5550 rc = obd_connect(NULL, exp, obd, &obd->obd_uuid, data, NULL);
5552 CERROR("%s: cannot connect to next dev %s (%d)\n",
5553 mdt_obd_name(m), next, rc);
5563 static int mdt_stack_init(const struct lu_env *env, struct mdt_device *mdt,
5564 struct lustre_cfg *cfg)
5566 char *dev = lustre_cfg_string(cfg, 0);
5567 int rc, name_size, uuid_size;
5568 char *name, *uuid, *p;
5569 struct lustre_cfg_bufs *bufs;
5570 struct lustre_cfg *lcfg;
5571 struct obd_device *obd;
5572 struct lustre_profile *lprof;
5573 struct lu_site *site;
5576 /* in 1.8 we had the only device in the stack - MDS.
5577 * 2.0 introduces MDT, MDD, OSD; MDT starts others internally.
5578 * in 2.3 OSD is instantiated by obd_mount.c, so we need
5579 * to generate names and setup MDT, MDD. MDT will be using
5580 * generated name to connect to MDD. for MDD the next device
5581 * will be LOD with name taken from so called "profile" which
5582 * is generated by mount_option line
5584 * 1.8 MGS generates config. commands like this:
5585 * #06 (104)mount_option 0: 1:lustre-MDT0000 2:lustre-mdtlov
5586 * #08 (120)setup 0:lustre-MDT0000 1:dev 2:type 3:lustre-MDT0000
5587 * 2.0 MGS generates config. commands like this:
5588 * #07 (112)mount_option 0: 1:lustre-MDT0000 2:lustre-MDT0000-mdtlov
5589 * #08 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
5590 * 3:lustre-MDT0000-mdtlov 4:f
5592 * we generate MDD name from MDT one, just replacing T with D
5594 * after all the preparations, the logical equivalent will be
5595 * #01 (160)setup 0:lustre-MDD0000 1:lustre-MDD0000_UUID 2:0
5596 * 3:lustre-MDT0000-mdtlov 4:f
5597 * #02 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
5598 * 3:lustre-MDD0000 4:f
5600 * notice we build the stack from down to top: MDD first, then MDT */
5602 name_size = MAX_OBD_NAME;
5603 uuid_size = MAX_OBD_NAME;
5605 OBD_ALLOC(name, name_size);
5606 OBD_ALLOC(uuid, uuid_size);
5607 if (name == NULL || uuid == NULL)
5608 GOTO(cleanup_mem, rc = -ENOMEM);
5610 OBD_ALLOC_PTR(bufs);
5612 GOTO(cleanup_mem, rc = -ENOMEM);
5615 p = strstr(name, "-MDT");
5617 GOTO(free_bufs, rc = -ENOMEM);
5620 snprintf(uuid, MAX_OBD_NAME, "%s_UUID", name);
5622 lprof = class_get_profile(lustre_cfg_string(cfg, 0));
5623 if (lprof == NULL || lprof->lp_dt == NULL) {
5624 CERROR("can't find the profile: %s\n",
5625 lustre_cfg_string(cfg, 0));
5626 GOTO(free_bufs, rc = -EINVAL);
5629 lustre_cfg_bufs_reset(bufs, name);
5630 lustre_cfg_bufs_set_string(bufs, 1, LUSTRE_MDD_NAME);
5631 lustre_cfg_bufs_set_string(bufs, 2, uuid);
5632 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
5634 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
5636 GOTO(put_profile, rc = -ENOMEM);
5637 lustre_cfg_init(lcfg, LCFG_ATTACH, bufs);
5639 rc = class_attach(lcfg);
5641 GOTO(lcfg_cleanup, rc);
5643 obd = class_name2obd(name);
5645 CERROR("Can not find obd %s (%s in config)\n",
5646 MDD_OBD_NAME, lustre_cfg_string(cfg, 0));
5647 GOTO(lcfg_cleanup, rc = -EINVAL);
5650 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens));
5652 lustre_cfg_bufs_reset(bufs, name);
5653 lustre_cfg_bufs_set_string(bufs, 1, uuid);
5654 lustre_cfg_bufs_set_string(bufs, 2, dev);
5655 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
5657 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
5659 GOTO(class_detach, rc = -ENOMEM);
5660 lustre_cfg_init(lcfg, LCFG_SETUP, bufs);
5662 rc = class_setup(obd, lcfg);
5664 GOTO(class_detach, rc);
5666 /* connect to MDD we just setup */
5667 rc = mdt_connect_to_next(env, mdt, name, &mdt->mdt_child_exp);
5669 GOTO(class_detach, rc);
5671 site = mdt->mdt_child_exp->exp_obd->obd_lu_dev->ld_site;
5673 LASSERT(mdt_lu_site(mdt) == NULL);
5674 mdt->mdt_lu_dev.ld_site = site;
5675 site->ls_top_dev = &mdt->mdt_lu_dev;
5676 mdt->mdt_child = lu2md_dev(mdt->mdt_child_exp->exp_obd->obd_lu_dev);
5678 /* now connect to bottom OSD */
5679 snprintf(name, MAX_OBD_NAME, "%s-osd", dev);
5680 rc = mdt_connect_to_next(env, mdt, name, &mdt->mdt_bottom_exp);
5682 GOTO(class_detach, rc);
5684 lu2dt_dev(mdt->mdt_bottom_exp->exp_obd->obd_lu_dev);
5686 rc = lu_env_refill((struct lu_env *)env);
5688 CERROR("Failure to refill session: '%d'\n", rc);
5690 lu_dev_add_linkage(site, &mdt->mdt_lu_dev);
5695 class_detach(obd, lcfg);
5697 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens));
5699 class_put_profile(lprof);
5704 OBD_FREE(name, name_size);
5706 OBD_FREE(uuid, uuid_size);
5710 /* setup quota master target on MDT0 */
5711 static int mdt_quota_init(const struct lu_env *env, struct mdt_device *mdt,
5712 struct lustre_cfg *cfg)
5714 struct obd_device *obd;
5715 char *dev = lustre_cfg_string(cfg, 0);
5716 char *qmtname, *uuid, *p;
5717 struct lustre_cfg_bufs *bufs;
5718 struct lustre_cfg *lcfg;
5719 struct lustre_profile *lprof;
5720 struct obd_connect_data *data;
5724 LASSERT(mdt->mdt_qmt_exp == NULL);
5725 LASSERT(mdt->mdt_qmt_dev == NULL);
5727 /* quota master is on MDT0 only for now */
5728 if (mdt->mdt_seq_site.ss_node_id != 0)
5731 /* MGS generates config commands which look as follows:
5732 * #01 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
5733 * 3:lustre-MDT0000-mdtlov 4:f
5735 * We generate the QMT name from the MDT one, just replacing MD with QM
5736 * after all the preparations, the logical equivalent will be:
5737 * #01 (160)setup 0:lustre-QMT0000 1:lustre-QMT0000_UUID 2:0
5738 * 3:lustre-MDT0000-osd 4:f */
5739 OBD_ALLOC(qmtname, MAX_OBD_NAME);
5740 OBD_ALLOC(uuid, UUID_MAX);
5741 OBD_ALLOC_PTR(bufs);
5742 OBD_ALLOC_PTR(data);
5743 if (qmtname == NULL || uuid == NULL || bufs == NULL || data == NULL)
5744 GOTO(cleanup_mem, rc = -ENOMEM);
5746 strcpy(qmtname, dev);
5747 p = strstr(qmtname, "-MDT");
5749 GOTO(cleanup_mem, rc = -ENOMEM);
5750 /* replace MD with QM */
5754 snprintf(uuid, UUID_MAX, "%s_UUID", qmtname);
5756 lprof = class_get_profile(lustre_cfg_string(cfg, 0));
5757 if (lprof == NULL || lprof->lp_dt == NULL) {
5758 CERROR("can't find profile for %s\n",
5759 lustre_cfg_string(cfg, 0));
5760 GOTO(cleanup_mem, rc = -EINVAL);
5763 lustre_cfg_bufs_reset(bufs, qmtname);
5764 lustre_cfg_bufs_set_string(bufs, 1, LUSTRE_QMT_NAME);
5765 lustre_cfg_bufs_set_string(bufs, 2, uuid);
5766 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
5768 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
5770 GOTO(put_profile, rc = -ENOMEM);
5771 lustre_cfg_init(lcfg, LCFG_ATTACH, bufs);
5773 rc = class_attach(lcfg);
5775 GOTO(lcfg_cleanup, rc);
5777 obd = class_name2obd(qmtname);
5779 CERROR("Can not find obd %s (%s in config)\n", qmtname,
5780 lustre_cfg_string(cfg, 0));
5781 GOTO(lcfg_cleanup, rc = -EINVAL);
5784 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens));
5786 lustre_cfg_bufs_reset(bufs, qmtname);
5787 lustre_cfg_bufs_set_string(bufs, 1, uuid);
5788 lustre_cfg_bufs_set_string(bufs, 2, dev);
5790 /* for quota, the next device should be the OSD device */
5791 lustre_cfg_bufs_set_string(bufs, 3,
5792 mdt->mdt_bottom->dd_lu_dev.ld_obd->obd_name);
5794 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
5796 GOTO(class_detach, rc = -ENOMEM);
5797 lustre_cfg_init(lcfg, LCFG_SETUP, bufs);
5799 rc = class_setup(obd, lcfg);
5801 GOTO(class_detach, rc);
5803 mdt->mdt_qmt_dev = obd->obd_lu_dev;
5805 /* configure local quota objects */
5806 if (CFS_FAIL_CHECK(OBD_FAIL_QUOTA_INIT))
5809 rc = mdt->mdt_qmt_dev->ld_ops->ldo_prepare(env,
5813 GOTO(class_cleanup, rc);
5815 /* connect to quota master target */
5816 data->ocd_connect_flags = OBD_CONNECT_VERSION;
5817 data->ocd_version = LUSTRE_VERSION_CODE;
5818 rc = obd_connect(NULL, &mdt->mdt_qmt_exp, obd, &obd->obd_uuid,
5821 CERROR("cannot connect to quota master device %s (%d)\n",
5823 GOTO(class_cleanup, rc);
5829 class_manual_cleanup(obd);
5830 mdt->mdt_qmt_dev = NULL;
5831 GOTO(lcfg_cleanup, rc);
5835 class_detach(obd, lcfg);
5837 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens));
5839 class_put_profile(lprof);
5844 OBD_FREE(qmtname, MAX_OBD_NAME);
5846 OBD_FREE(uuid, UUID_MAX);
5852 /* Shutdown quota master target associated with mdt */
5853 static void mdt_quota_fini(const struct lu_env *env, struct mdt_device *mdt)
5857 if (mdt->mdt_qmt_exp == NULL)
5859 LASSERT(mdt->mdt_qmt_dev != NULL);
5861 /* the qmt automatically shuts down when the mdt disconnects */
5862 obd_disconnect(mdt->mdt_qmt_exp);
5863 mdt->mdt_qmt_exp = NULL;
5864 mdt->mdt_qmt_dev = NULL;
5868 /* mdt_getxattr() is used from mdt_intent_getxattr(), use this wrapper
5869 * for now. This will be removed along with converting rest of MDT code
5870 * to use tgt_session_info */
5871 static int mdt_tgt_getxattr(struct tgt_session_info *tsi)
5873 struct mdt_thread_info *info = tsi2mdt_info(tsi);
5876 if (unlikely(info->mti_object == NULL))
5879 rc = mdt_getxattr(info);
5881 mdt_thread_info_fini(info);
5885 static int mdt_llog_open(struct tgt_session_info *tsi)
5889 if (!mdt_changelog_allow(tsi2mdt_info(tsi)))
5890 RETURN(err_serious(-EACCES));
5892 RETURN(tgt_llog_open(tsi));
5895 #define OBD_FAIL_OST_READ_NET OBD_FAIL_OST_BRW_NET
5896 #define OBD_FAIL_OST_WRITE_NET OBD_FAIL_OST_BRW_NET
5897 #define OST_BRW_READ OST_READ
5898 #define OST_BRW_WRITE OST_WRITE
5900 static struct tgt_handler mdt_tgt_handlers[] = {
5901 TGT_RPC_HANDLER(MDS_FIRST_OPC,
5902 0, MDS_CONNECT, mdt_tgt_connect,
5903 &RQF_CONNECT, LUSTRE_OBD_VERSION),
5904 TGT_RPC_HANDLER(MDS_FIRST_OPC,
5905 0, MDS_DISCONNECT, tgt_disconnect,
5906 &RQF_MDS_DISCONNECT, LUSTRE_OBD_VERSION),
5907 TGT_RPC_HANDLER(MDS_FIRST_OPC,
5908 HAS_REPLY, MDS_SET_INFO, mdt_set_info,
5909 &RQF_MDT_SET_INFO, LUSTRE_MDS_VERSION),
5910 TGT_MDT_HDL(0, MDS_GET_INFO, mdt_get_info),
5911 TGT_MDT_HDL(HAS_REPLY, MDS_GET_ROOT, mdt_get_root),
5912 TGT_MDT_HDL(HAS_BODY, MDS_GETATTR, mdt_getattr),
5913 TGT_MDT_HDL(HAS_BODY | HAS_REPLY, MDS_GETATTR_NAME,
5915 TGT_MDT_HDL(HAS_BODY, MDS_GETXATTR, mdt_tgt_getxattr),
5916 TGT_MDT_HDL(HAS_REPLY, MDS_STATFS, mdt_statfs),
5917 TGT_MDT_HDL(IS_MUTABLE, MDS_REINT, mdt_reint),
5918 TGT_MDT_HDL(HAS_BODY, MDS_CLOSE, mdt_close),
5919 TGT_MDT_HDL(HAS_BODY | HAS_REPLY, MDS_READPAGE, mdt_readpage),
5920 TGT_MDT_HDL(HAS_BODY | HAS_REPLY, MDS_SYNC, mdt_sync),
5921 TGT_MDT_HDL(0, MDS_QUOTACTL, mdt_quotactl),
5922 TGT_MDT_HDL(HAS_BODY | HAS_REPLY | IS_MUTABLE, MDS_HSM_PROGRESS,
5924 TGT_MDT_HDL(HAS_BODY | HAS_REPLY | IS_MUTABLE, MDS_HSM_CT_REGISTER,
5925 mdt_hsm_ct_register),
5926 TGT_MDT_HDL(HAS_BODY | HAS_REPLY | IS_MUTABLE, MDS_HSM_CT_UNREGISTER,
5927 mdt_hsm_ct_unregister),
5928 TGT_MDT_HDL(HAS_BODY | HAS_REPLY, MDS_HSM_STATE_GET,
5930 TGT_MDT_HDL(HAS_BODY | HAS_REPLY | IS_MUTABLE, MDS_HSM_STATE_SET,
5932 TGT_MDT_HDL(HAS_BODY | HAS_REPLY, MDS_HSM_ACTION, mdt_hsm_action),
5933 TGT_MDT_HDL(HAS_BODY | HAS_REPLY, MDS_HSM_REQUEST,
5935 TGT_MDT_HDL(HAS_KEY | HAS_BODY | HAS_REPLY | IS_MUTABLE,
5938 TGT_MDT_HDL(IS_MUTABLE, MDS_RMFID, mdt_rmfid),
5939 TGT_MDT_HDL(IS_MUTABLE, MDS_BATCH, mdt_batch),
5942 static struct tgt_handler mdt_io_ops[] = {
5943 TGT_OST_HDL_HP(HAS_BODY | HAS_REPLY, OST_BRW_READ, tgt_brw_read,
5945 TGT_OST_HDL_HP(HAS_BODY | IS_MUTABLE, OST_BRW_WRITE, tgt_brw_write,
5947 TGT_OST_HDL_HP(HAS_BODY | HAS_REPLY | IS_MUTABLE,
5948 OST_PUNCH, mdt_punch_hdl,
5950 TGT_OST_HDL(HAS_BODY | HAS_REPLY, OST_SYNC, mdt_data_sync),
5951 TGT_OST_HDL(HAS_BODY | HAS_REPLY | IS_MUTABLE, OST_FALLOCATE,
5953 TGT_OST_HDL(HAS_BODY | HAS_REPLY, OST_SEEK, tgt_lseek),
5954 TGT_RPC_HANDLER(OST_FIRST_OPC,
5955 0, OST_SET_INFO, mdt_io_set_info,
5956 &RQF_OBD_SET_INFO, LUSTRE_OST_VERSION),
5959 static struct tgt_handler mdt_sec_ctx_ops[] = {
5960 TGT_SEC_HDL_VAR(0, SEC_CTX_INIT, mdt_sec_ctx_handle),
5961 TGT_SEC_HDL_VAR(0, SEC_CTX_INIT_CONT,mdt_sec_ctx_handle),
5962 TGT_SEC_HDL_VAR(0, SEC_CTX_FINI, mdt_sec_ctx_handle)
5965 static struct tgt_handler mdt_quota_ops[] = {
5966 TGT_QUOTA_HDL(HAS_REPLY, QUOTA_DQACQ, mdt_quota_dqacq),
5969 static struct tgt_handler mdt_llog_handlers[] = {
5970 TGT_LLOG_HDL(0, LLOG_ORIGIN_HANDLE_CREATE, mdt_llog_open),
5971 TGT_LLOG_HDL(0, LLOG_ORIGIN_HANDLE_NEXT_BLOCK, tgt_llog_next_block),
5972 TGT_LLOG_HDL(0, LLOG_ORIGIN_HANDLE_READ_HEADER, tgt_llog_read_header),
5973 TGT_LLOG_HDL(0, LLOG_ORIGIN_HANDLE_PREV_BLOCK, tgt_llog_prev_block),
5976 static struct tgt_opc_slice mdt_common_slice[] = {
5978 .tos_opc_start = MDS_FIRST_OPC,
5979 .tos_opc_end = MDS_LAST_OPC,
5980 .tos_hs = mdt_tgt_handlers
5983 .tos_opc_start = OBD_FIRST_OPC,
5984 .tos_opc_end = OBD_LAST_OPC,
5985 .tos_hs = tgt_obd_handlers
5988 .tos_opc_start = LDLM_FIRST_OPC,
5989 .tos_opc_end = LDLM_LAST_OPC,
5990 .tos_hs = tgt_dlm_handlers
5993 .tos_opc_start = SEC_FIRST_OPC,
5994 .tos_opc_end = SEC_LAST_OPC,
5995 .tos_hs = mdt_sec_ctx_ops
5998 .tos_opc_start = OUT_UPDATE_FIRST_OPC,
5999 .tos_opc_end = OUT_UPDATE_LAST_OPC,
6000 .tos_hs = tgt_out_handlers
6003 .tos_opc_start = FLD_FIRST_OPC,
6004 .tos_opc_end = FLD_LAST_OPC,
6005 .tos_hs = fld_handlers
6008 .tos_opc_start = SEQ_FIRST_OPC,
6009 .tos_opc_end = SEQ_LAST_OPC,
6010 .tos_hs = seq_handlers
6013 .tos_opc_start = QUOTA_DQACQ,
6014 .tos_opc_end = QUOTA_LAST_OPC,
6015 .tos_hs = mdt_quota_ops
6018 .tos_opc_start = LLOG_FIRST_OPC,
6019 .tos_opc_end = LLOG_LAST_OPC,
6020 .tos_hs = mdt_llog_handlers
6023 .tos_opc_start = LFSCK_FIRST_OPC,
6024 .tos_opc_end = LFSCK_LAST_OPC,
6025 .tos_hs = tgt_lfsck_handlers
6028 .tos_opc_start = OST_FIRST_OPC,
6029 .tos_opc_end = OST_LAST_OPC,
6030 .tos_hs = mdt_io_ops
6037 static void mdt_fini(const struct lu_env *env, struct mdt_device *m)
6039 struct md_device *next = m->mdt_child;
6040 struct lu_device *d = &m->mdt_lu_dev;
6041 struct obd_device *obd = mdt2obd_dev(m);
6042 struct lfsck_stop stop;
6045 stop.ls_status = LS_PAUSED;
6047 next->md_ops->mdo_iocontrol(env, next, OBD_IOC_STOP_LFSCK, 0, &stop);
6049 mdt_stack_pre_fini(env, m, md2lu_dev(m->mdt_child));
6051 mdt_restriper_stop(m);
6052 ping_evictor_stop();
6054 /* Remove the HSM /proc entry so the coordinator cannot be
6055 * restarted by a user while it's shutting down.
6057 mdt_hsm_cdt_stop(m);
6059 mdt_llog_ctxt_unclone(env, m, LLOG_AGENT_ORIG_CTXT);
6060 mdt_llog_ctxt_unclone(env, m, LLOG_CHANGELOG_ORIG_CTXT);
6062 if (m->mdt_namespace != NULL)
6063 ldlm_namespace_free_prior(m->mdt_namespace, NULL,
6064 d->ld_obd->obd_force);
6066 obd_exports_barrier(obd);
6067 obd_zombie_barrier();
6069 mdt_quota_fini(env, m);
6071 cfs_free_nidlist(&m->mdt_squash.rsi_nosquash_nids);
6073 /* Calling the cleanup functions in the same order as in the mdt_init0
6076 mdt_tunables_fini(m);
6078 target_recovery_fini(obd);
6079 upcall_cache_cleanup(m->mdt_identity_cache);
6080 m->mdt_identity_cache = NULL;
6082 tgt_fini(env, &m->mdt_lut);
6084 mdt_hsm_cdt_fini(m);
6086 if (m->mdt_los != NULL) {
6087 local_oid_storage_fini(env, m->mdt_los);
6091 if (m->mdt_namespace != NULL) {
6092 ldlm_namespace_free_post(m->mdt_namespace);
6093 d->ld_obd->obd_namespace = m->mdt_namespace = NULL;
6096 if (m->mdt_md_root != NULL) {
6097 mdt_object_put(env, m->mdt_md_root);
6098 m->mdt_md_root = NULL;
6101 mdt_seq_fini(env, m);
6103 mdt_fld_fini(env, m);
6108 mdt_stack_fini(env, m, md2lu_dev(m->mdt_child));
6110 LASSERT(atomic_read(&d->ld_ref) == 0);
6112 server_put_mount(mdt_obd_name(m), true);
6117 static int mdt_postrecov(const struct lu_env *, struct mdt_device *);
6119 static int mdt_init0(const struct lu_env *env, struct mdt_device *m,
6120 struct lu_device_type *ldt, struct lustre_cfg *cfg)
6122 const struct dt_device_param *dt_conf;
6123 struct mdt_thread_info *info;
6124 struct obd_device *obd;
6125 const char *dev = lustre_cfg_string(cfg, 0);
6126 const char *num = lustre_cfg_string(cfg, 2);
6127 struct tg_grants_data *tgd = &m->mdt_lut.lut_tgd;
6128 struct lustre_mount_info *lmi = NULL;
6129 struct lustre_sb_info *lsi;
6131 struct seq_server_site *ss_site;
6132 const char *identity_upcall = "NONE";
6133 struct md_device *next;
6140 lu_device_init(&m->mdt_lu_dev, ldt);
6142 * Environment (env) might be missing mdt_thread_key values at that
6143 * point, if device is allocated when mdt_thread_key is in QUIESCENT
6146 * Usually device allocation path doesn't use module key values, but
6147 * mdt has to do a lot of work here, so allocate key value.
6149 rc = lu_env_refill((struct lu_env *)env);
6153 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
6154 LASSERT(info != NULL);
6156 obd = class_name2obd(dev);
6157 LASSERT(obd != NULL);
6159 m->mdt_max_mdsize = MAX_MD_SIZE_OLD;
6160 m->mdt_opts.mo_evict_tgt_nids = 1;
6161 m->mdt_opts.mo_cos = MDT_COS_DEFAULT;
6163 lmi = server_get_mount(dev);
6165 CERROR("Cannot get mount info for %s!\n", dev);
6168 lsi = s2lsi(lmi->lmi_sb);
6169 LASSERT(lsi->lsi_lmd);
6170 /* CMD is supported only in IAM mode */
6172 rc = kstrtol(num, 10, &node_id);
6177 if (test_bit(LMD_FLG_SKIP_LFSCK, lsi->lsi_lmd->lmd_flags))
6178 m->mdt_skip_lfsck = 1;
6181 /* Just try to get a DoM lock by default. Otherwise, having a group
6182 * lock granted, it may get blocked for a long time. */
6183 m->mdt_opts.mo_dom_lock = TRYLOCK_DOM_ON_OPEN;
6184 /* DoM files are read at open and data is packed in the reply */
6185 m->mdt_opts.mo_dom_read_open = 1;
6187 m->mdt_squash.rsi_uid = 0;
6188 m->mdt_squash.rsi_gid = 0;
6189 INIT_LIST_HEAD(&m->mdt_squash.rsi_nosquash_nids);
6190 spin_lock_init(&m->mdt_squash.rsi_lock);
6191 spin_lock_init(&m->mdt_lock);
6192 m->mdt_enable_chprojid_gid = 0;
6193 m->mdt_enable_dir_migration = 1;
6194 m->mdt_enable_dir_restripe = 0;
6195 m->mdt_enable_dir_auto_split = 0;
6196 m->mdt_enable_parallel_rename_dir = 1;
6197 m->mdt_enable_parallel_rename_file = 1;
6198 m->mdt_enable_remote_dir = 1;
6199 m->mdt_enable_remote_dir_gid = 0;
6200 m->mdt_enable_remote_rename = 1;
6201 m->mdt_enable_striped_dir = 1;
6202 m->mdt_enable_dmv_implicit_inherit = 1;
6203 m->mdt_dir_restripe_nsonly = 1;
6204 m->mdt_max_mod_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
6206 atomic_set(&m->mdt_mds_mds_conns, 0);
6207 atomic_set(&m->mdt_async_commit_count, 0);
6208 atomic_set(&m->mdt_dmv_old_client_count, 0);
6210 m->mdt_lu_dev.ld_ops = &mdt_lu_ops;
6211 m->mdt_lu_dev.ld_obd = obd;
6212 /* Set this lu_device to obd for error handling purposes. */
6213 obd->obd_lu_dev = &m->mdt_lu_dev;
6215 strncpy(m->mdt_job_xattr, XATTR_NAME_JOB_DEFAULT, XATTR_JOB_MAX_LEN);
6217 /* init the stack */
6218 rc = mdt_stack_init((struct lu_env *)env, m, cfg);
6220 CERROR("%s: Can't init device stack, rc %d\n",
6221 mdt_obd_name(m), rc);
6226 ss_site = mdt_seq_site(m);
6227 s->ld_seq_site = ss_site;
6230 /* set server index */
6231 ss_site->ss_node_id = node_id;
6233 /* failover is the default
6234 * FIXME: we do not failout mds0/mgs, which may cause some problems.
6235 * assumed whose ss_node_id == 0 XXX
6237 obd->obd_replayable = 1;
6238 /* No connection accepted until configurations will finish */
6239 obd->obd_no_conn = 1;
6241 if (cfg->lcfg_bufcount > 4 && LUSTRE_CFG_BUFLEN(cfg, 4) > 0) {
6242 char *str = lustre_cfg_string(cfg, 4);
6243 if (strchr(str, 'n')) {
6244 CWARN("%s: recovery disabled\n", mdt_obd_name(m));
6245 obd->obd_replayable = 0;
6249 rc = mdt_fld_init(env, mdt_obd_name(m), m);
6251 GOTO(err_fini_stack, rc);
6253 rc = mdt_seq_init(env, m);
6255 GOTO(err_fini_fld, rc);
6257 snprintf(info->mti_u.ns_name, sizeof(info->mti_u.ns_name), "%s-%s",
6258 LUSTRE_MDT_NAME, obd->obd_uuid.uuid);
6259 m->mdt_namespace = ldlm_namespace_new(obd, info->mti_u.ns_name,
6260 LDLM_NAMESPACE_SERVER,
6261 LDLM_NAMESPACE_GREEDY,
6263 if (IS_ERR(m->mdt_namespace)) {
6264 rc = PTR_ERR(m->mdt_namespace);
6265 CERROR("%s: unable to create server namespace: rc = %d\n",
6267 m->mdt_namespace = NULL;
6268 GOTO(err_fini_seq, rc);
6271 m->mdt_namespace->ns_lvbp = m;
6272 m->mdt_namespace->ns_lvbo = &mdt_lvbo;
6274 ldlm_register_intent(m->mdt_namespace, mdt_intent_policy);
6275 /* set obd_namespace for compatibility with old code */
6276 obd->obd_namespace = m->mdt_namespace;
6278 rc = tgt_init(env, &m->mdt_lut, obd, m->mdt_bottom, mdt_common_slice,
6279 OBD_FAIL_MDS_ALL_REQUEST_NET,
6280 OBD_FAIL_MDS_ALL_REPLY_NET);
6282 GOTO(err_free_ns, rc);
6284 /* Amount of available space excluded from granting and reserved
6285 * for metadata. It is a percentage of the total MDT size. */
6286 tgd->tgd_reserved_pcnt = 10;
6288 if (ONE_MB_BRW_SIZE < (1U << tgd->tgd_blockbits))
6289 m->mdt_brw_size = 1U << tgd->tgd_blockbits;
6291 m->mdt_brw_size = ONE_MB_BRW_SIZE;
6293 if (CFS_FAIL_CHECK(OBD_FAIL_MDS_FS_SETUP))
6294 GOTO(err_tgt, rc = -ENOENT);
6296 fid.f_seq = FID_SEQ_LOCAL_NAME;
6299 rc = local_oid_storage_init(env, m->mdt_bottom, &fid, &m->mdt_los);
6303 rc = mdt_hsm_cdt_init(m);
6305 CERROR("%s: error initializing coordinator, rc %d\n",
6306 mdt_obd_name(m), rc);
6307 GOTO(err_los_fini, rc);
6310 tgt_adapt_sptlrpc_conf(&m->mdt_lut);
6312 next = m->mdt_child;
6313 dt_conf = next->md_ops->mdo_dtconf_get(env, next);
6315 mntopts = dt_conf->ddp_mntopts;
6317 if (mntopts & MNTOPT_USERXATTR)
6318 m->mdt_opts.mo_user_xattr = 1;
6320 m->mdt_opts.mo_user_xattr = 0;
6322 m->mdt_max_ea_size = dt_conf->ddp_max_ea_size;
6324 if (mntopts & MNTOPT_ACL)
6325 m->mdt_opts.mo_acl = 1;
6327 m->mdt_opts.mo_acl = 0;
6329 m->mdt_opts.mo_enable_strict_som = 1;
6331 /* XXX: to support suppgid for ACL, we enable identity_upcall
6332 * by default, otherwise, maybe got unexpected -EACCESS. */
6333 if (m->mdt_opts.mo_acl)
6334 identity_upcall = MDT_IDENTITY_UPCALL_PATH;
6336 m->mdt_identity_cache = upcall_cache_init(mdt_obd_name(m),
6338 UC_IDCACHE_HASH_SIZE,
6339 &mdt_identity_upcall_cache_ops);
6340 if (IS_ERR(m->mdt_identity_cache)) {
6341 rc = PTR_ERR(m->mdt_identity_cache);
6342 m->mdt_identity_cache = NULL;
6343 GOTO(err_free_hsm, rc);
6346 rc = mdt_tunables_init(m, dev);
6348 CERROR("Can't init MDT lprocfs, rc %d\n", rc);
6349 GOTO(err_recovery, rc);
6352 rc = mdt_quota_init(env, m, cfg);
6354 GOTO(err_procfs, rc);
6356 m->mdt_ldlm_client = &mdt2obd_dev(m)->obd_ldlm_client;
6357 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
6358 "mdt_ldlm_client", m->mdt_ldlm_client);
6360 ping_evictor_start();
6362 /* recovery will be started upon mdt_prepare()
6363 * when the whole stack is complete and ready
6364 * to serve the requests */
6366 /* Reduce the initial timeout on an MDS because it doesn't need such
6367 * a long timeout as an OST does. Adaptive timeouts will adjust this
6368 * value appropriately. */
6369 if (ldlm_timeout == LDLM_TIMEOUT_DEFAULT)
6370 ldlm_timeout = MDS_LDLM_TIMEOUT_DEFAULT;
6372 if (test_bit(LMD_FLG_LOCAL_RECOV, lsi->lsi_lmd->lmd_flags))
6373 m->mdt_lut.lut_local_recovery = 1;
6375 rc = mdt_restriper_start(m);
6377 GOTO(err_ping_evictor, rc);
6382 ping_evictor_stop();
6384 mdt_tunables_fini(m);
6386 upcall_cache_cleanup(m->mdt_identity_cache);
6387 m->mdt_identity_cache = NULL;
6389 mdt_hsm_cdt_fini(m);
6391 local_oid_storage_fini(env, m->mdt_los);
6394 /* keep recoverable clients */
6396 target_recovery_fini(obd);
6397 obd_exports_barrier(obd);
6398 obd_zombie_barrier();
6399 tgt_fini(env, &m->mdt_lut);
6401 ldlm_namespace_free(m->mdt_namespace, NULL, 0);
6402 obd->obd_namespace = m->mdt_namespace = NULL;
6404 mdt_seq_fini(env, m);
6406 mdt_fld_fini(env, m);
6408 mdt_stack_fini(env, m, md2lu_dev(m->mdt_child));
6411 server_put_mount(dev, true);
6415 /* For interoperability, the left element is old parameter, the right one
6416 * is the new version of the parameter, if some parameter is deprecated,
6417 * the new version should be set as NULL. */
6418 static struct cfg_interop_param mdt_interop_param[] = {
6419 { "mdt.group_upcall", NULL },
6420 { "mdt.quota_type", NULL },
6421 { "mdd.quota_type", NULL },
6422 { "mdt.som", NULL },
6423 { "mdt.rootsquash", "mdt.root_squash" },
6424 { "mdt.nosquash_nid", "mdt.nosquash_nids" },
6428 /* used by MGS to process specific configurations */
6429 static int mdt_process_config(const struct lu_env *env,
6430 struct lu_device *d, struct lustre_cfg *cfg)
6432 struct mdt_device *m = mdt_dev(d);
6433 struct md_device *md_next = m->mdt_child;
6434 struct lu_device *next = md2lu_dev(md_next);
6438 switch (cfg->lcfg_command) {
6440 struct obd_device *obd = d->ld_obd;
6441 /* For interoperability */
6442 struct cfg_interop_param *ptr = NULL;
6443 struct lustre_cfg *old_cfg = NULL;
6447 param = lustre_cfg_string(cfg, 1);
6448 if (param == NULL) {
6449 CERROR("param is empty\n");
6454 ptr = class_find_old_param(param, mdt_interop_param);
6456 if (ptr->new_param == NULL) {
6458 CWARN("For interoperability, skip this %s."
6459 " It is obsolete.\n", ptr->old_param);
6463 CWARN("Found old param %s, changed it to %s.\n",
6464 ptr->old_param, ptr->new_param);
6467 cfg = lustre_cfg_rename(old_cfg, ptr->new_param);
6474 count = class_modify_config(cfg, PARAM_MDT,
6475 &obd->obd_kset.kobj);
6477 struct coordinator *cdt = &m->mdt_coordinator;
6479 /* is it an HSM var ? */
6480 count = class_modify_config(cfg, PARAM_HSM,
6481 &cdt->cdt_hsm_kobj);
6483 /* we don't understand; pass it on */
6484 rc = next->ld_ops->ldo_process_config(env, next,
6487 rc = count > 0 ? 0 : count;
6489 rc = count > 0 ? 0 : count;
6493 OBD_FREE(cfg, lustre_cfg_len(cfg->lcfg_bufcount,
6494 cfg->lcfg_buflens));
6498 /* others are passed further */
6499 rc = next->ld_ops->ldo_process_config(env, next, cfg);
6505 static struct lu_object *mdt_object_alloc(const struct lu_env *env,
6506 const struct lu_object_header *hdr,
6507 struct lu_device *d)
6509 struct mdt_object *mo;
6513 OBD_SLAB_ALLOC_PTR_GFP(mo, mdt_object_kmem, GFP_NOFS);
6515 struct lu_object *o;
6516 struct lu_object_header *h;
6519 h = &mo->mot_header;
6520 lu_object_header_init(h);
6521 lu_object_init(o, h, d);
6522 lu_object_add_top(h, o);
6523 o->lo_ops = &mdt_obj_ops;
6524 spin_lock_init(&mo->mot_write_lock);
6525 mutex_init(&mo->mot_som_mutex);
6526 mutex_init(&mo->mot_lov_mutex);
6527 init_rwsem(&mo->mot_dom_sem);
6528 init_rwsem(&mo->mot_open_sem);
6529 atomic_set(&mo->mot_open_count, 0);
6530 mo->mot_restripe_offset = 0;
6531 INIT_LIST_HEAD(&mo->mot_restripe_linkage);
6532 mo->mot_lsom_size = 0;
6533 mo->mot_lsom_blocks = 0;
6534 mo->mot_lsom_inited = false;
6540 static int mdt_object_init(const struct lu_env *env, struct lu_object *o,
6541 const struct lu_object_conf *unused)
6543 struct mdt_device *d = mdt_dev(o->lo_dev);
6544 struct lu_device *under;
6545 struct lu_object *below;
6549 CDEBUG(D_INFO, "object init, fid = "DFID"\n",
6550 PFID(lu_object_fid(o)));
6552 under = &d->mdt_child->md_lu_dev;
6553 below = under->ld_ops->ldo_object_alloc(env, o->lo_header, under);
6554 if (below != NULL) {
6555 lu_object_add(o, below);
6562 static void mdt_object_free_rcu(struct rcu_head *head)
6564 struct mdt_object *mo = container_of(head, struct mdt_object,
6565 mot_header.loh_rcu);
6567 kmem_cache_free(mdt_object_kmem, mo);
6570 static void mdt_object_free(const struct lu_env *env, struct lu_object *o)
6572 struct mdt_object *mo = mdt_obj(o);
6573 struct lu_object_header *h;
6577 CDEBUG(D_INFO, "object free, fid = "DFID"\n",
6578 PFID(lu_object_fid(o)));
6580 LASSERT(atomic_read(&mo->mot_open_count) == 0);
6581 LASSERT(atomic_read(&mo->mot_lease_count) == 0);
6584 lu_object_header_fini(h);
6585 OBD_FREE_PRE(mo, sizeof(*mo), "slab-freed");
6586 call_rcu(&mo->mot_header.loh_rcu, mdt_object_free_rcu);
6591 static int mdt_object_print(const struct lu_env *env, void *cookie,
6592 lu_printer_t p, const struct lu_object *o)
6594 struct mdt_object *mdto = mdt_obj((struct lu_object *)o);
6596 return (*p)(env, cookie,
6597 LUSTRE_MDT_NAME"-object@%p(%s %s, writecount=%d)",
6598 mdto, mdto->mot_lov_created ? "lov_created" : "",
6599 mdto->mot_cache_attr ? "cache_attr" : "",
6600 mdto->mot_write_count);
6603 static int mdt_prepare(const struct lu_env *env,
6604 struct lu_device *pdev,
6605 struct lu_device *cdev)
6607 struct mdt_device *mdt = mdt_dev(cdev);
6608 struct lu_device *next = &mdt->mdt_child->md_lu_dev;
6609 struct obd_device *obd = cdev->ld_obd;
6616 rc = next->ld_ops->ldo_prepare(env, cdev, next);
6620 rc = mdt_llog_ctxt_clone(env, mdt, LLOG_CHANGELOG_ORIG_CTXT);
6624 rc = mdt_llog_ctxt_clone(env, mdt, LLOG_AGENT_ORIG_CTXT);
6628 rc = lfsck_register_namespace(env, mdt->mdt_bottom, mdt->mdt_namespace);
6629 /* The LFSCK instance is registered just now, so it must be there when
6630 * register the namespace to such instance. */
6631 LASSERTF(rc == 0, "register namespace failed: rc = %d\n", rc);
6633 if (mdt->mdt_seq_site.ss_node_id == 0) {
6634 rc = mdt->mdt_child->md_ops->mdo_root_get(env, mdt->mdt_child,
6635 &mdt->mdt_md_root_fid);
6640 LASSERT(!test_bit(MDT_FL_CFGLOG, &mdt->mdt_state));
6642 target_recovery_init(&mdt->mdt_lut, tgt_request_handle);
6643 set_bit(MDT_FL_CFGLOG, &mdt->mdt_state);
6644 LASSERT(obd->obd_no_conn);
6645 spin_lock(&obd->obd_dev_lock);
6646 obd->obd_no_conn = 0;
6647 spin_unlock(&obd->obd_dev_lock);
6649 if (obd->obd_recovering == 0)
6650 mdt_postrecov(env, mdt);
6655 const struct lu_device_operations mdt_lu_ops = {
6656 .ldo_object_alloc = mdt_object_alloc,
6657 .ldo_process_config = mdt_process_config,
6658 .ldo_prepare = mdt_prepare,
6661 static const struct lu_object_operations mdt_obj_ops = {
6662 .loo_object_init = mdt_object_init,
6663 .loo_object_free = mdt_object_free,
6664 .loo_object_print = mdt_object_print
6667 static int mdt_obd_set_info_async(const struct lu_env *env,
6668 struct obd_export *exp,
6669 __u32 keylen, void *key,
6670 __u32 vallen, void *val,
6671 struct ptlrpc_request_set *set)
6677 if (KEY_IS(KEY_SPTLRPC_CONF)) {
6678 rc = tgt_adapt_sptlrpc_conf(class_exp2tgt(exp));
6685 static inline void mdt_enable_slc(struct mdt_device *mdt)
6687 if (mdt->mdt_lut.lut_sync_lock_cancel == SYNC_LOCK_CANCEL_NEVER)
6688 mdt->mdt_lut.lut_sync_lock_cancel = SYNC_LOCK_CANCEL_BLOCKING;
6691 static inline void mdt_disable_slc(struct mdt_device *mdt)
6693 if (mdt->mdt_lut.lut_sync_lock_cancel == SYNC_LOCK_CANCEL_BLOCKING)
6694 mdt->mdt_lut.lut_sync_lock_cancel = SYNC_LOCK_CANCEL_NEVER;
6698 * Match client and server connection feature flags.
6700 * Compute the compatibility flags for a connection request based on
6701 * features mutually supported by client and server.
6703 * The obd_export::exp_connect_data.ocd_connect_flags field in \a exp
6704 * must not be updated here, otherwise a partially initialized value may
6705 * be exposed. After the connection request is successfully processed,
6706 * the top-level MDT connect request handler atomically updates the export
6707 * connect flags from the obd_connect_data::ocd_connect_flags field of the
6708 * reply. \see mdt_connect().
6710 * Before 2.7.50 clients will send a struct obd_connect_data_v1 rather than a
6711 * full struct obd_connect_data. So care must be taken when accessing fields
6712 * that are not present in struct obd_connect_data_v1. See LU-16.
6714 * \param exp the obd_export associated with this client/target pair
6715 * \param mdt the target device for the connection
6716 * \param data stores data for this connect request
6719 * \retval -EPROTO \a data unexpectedly has zero obd_connect_data::ocd_brw_size
6720 * \retval -EBADE client and server feature requirements are incompatible
6722 static int mdt_connect_internal(const struct lu_env *env,
6723 struct obd_export *exp,
6724 struct mdt_device *mdt,
6725 struct obd_connect_data *data, bool reconnect)
6727 const char *obd_name = mdt_obd_name(mdt);
6728 LASSERT(data != NULL);
6730 data->ocd_connect_flags &= MDT_CONNECT_SUPPORTED;
6732 if (mdt->mdt_bottom->dd_rdonly &&
6733 !(data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
6734 !(data->ocd_connect_flags & OBD_CONNECT_RDONLY))
6737 if (data->ocd_connect_flags & OBD_CONNECT_FLAGS2)
6738 data->ocd_connect_flags2 &= MDT_CONNECT_SUPPORTED2;
6740 data->ocd_ibits_known &= MDS_INODELOCK_FULL;
6742 if (!mdt->mdt_opts.mo_acl)
6743 data->ocd_connect_flags &= ~OBD_CONNECT_ACL;
6745 if (!mdt->mdt_opts.mo_user_xattr)
6746 data->ocd_connect_flags &= ~OBD_CONNECT_XATTR;
6748 if (OCD_HAS_FLAG(data, BRW_SIZE)) {
6749 data->ocd_brw_size = min(data->ocd_brw_size,
6751 if (data->ocd_brw_size == 0) {
6752 CERROR("%s: cli %s/%p ocd_connect_flags: %#llx "
6753 "ocd_version: %x ocd_grant: %d ocd_index: %u "
6754 "ocd_brw_size unexpectedly zero, network data "
6755 "corruption? Refusing to connect this client\n",
6756 obd_name, exp->exp_client_uuid.uuid,
6757 exp, data->ocd_connect_flags, data->ocd_version,
6758 data->ocd_grant, data->ocd_index);
6763 if (OCD_HAS_FLAG(data, GRANT_PARAM)) {
6764 struct dt_device_param *ddp = &mdt->mdt_lut.lut_dt_conf;
6766 /* client is reporting its page size, for future use */
6767 exp->exp_target_data.ted_pagebits = data->ocd_grant_blkbits;
6768 data->ocd_grant_blkbits = mdt->mdt_lut.lut_tgd.tgd_blockbits;
6769 /* ddp_inodespace may not be power-of-two value, eg. for ldiskfs
6770 * it's LDISKFS_DIR_REC_LEN(20) = 28. */
6771 data->ocd_grant_inobits = fls(ddp->ddp_inodespace - 1);
6772 /* ocd_grant_tax_kb is in 1K byte blocks */
6773 data->ocd_grant_tax_kb = ddp->ddp_extent_tax >> 10;
6774 data->ocd_grant_max_blks = ddp->ddp_max_extent_blks;
6777 /* Save connect_data we have so far because tgt_grant_connect()
6778 * uses it to calculate grant, and we want to save the client
6779 * version before it is overwritten by LUSTRE_VERSION_CODE. */
6780 exp->exp_connect_data = *data;
6781 if (OCD_HAS_FLAG(data, GRANT))
6782 tgt_grant_connect(env, exp, data, !reconnect);
6784 if (OCD_HAS_FLAG(data, MAXBYTES))
6785 data->ocd_maxbytes = mdt->mdt_lut.lut_dt_conf.ddp_maxbytes;
6787 /* NB: Disregard the rule against updating
6788 * exp_connect_data.ocd_connect_flags in this case, since
6789 * tgt_client_new() needs to know if this is a lightweight
6790 * connection, and it is safe to expose this flag before
6791 * connection processing completes. */
6792 if (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) {
6793 spin_lock(&exp->exp_lock);
6794 *exp_connect_flags_ptr(exp) |= OBD_CONNECT_LIGHTWEIGHT;
6795 spin_unlock(&exp->exp_lock);
6798 data->ocd_version = LUSTRE_VERSION_CODE;
6800 if ((data->ocd_connect_flags & OBD_CONNECT_FID) == 0) {
6801 CWARN("%s: MDS requires FID support, but client not\n",
6806 if (OCD_HAS_FLAG(data, PINGLESS)) {
6807 if (ptlrpc_pinger_suppress_pings()) {
6808 spin_lock(&exp->exp_obd->obd_dev_lock);
6809 list_del_init(&exp->exp_obd_chain_timed);
6810 spin_unlock(&exp->exp_obd->obd_dev_lock);
6812 data->ocd_connect_flags &= ~OBD_CONNECT_PINGLESS;
6816 data->ocd_max_easize = mdt->mdt_max_ea_size;
6818 /* NB: Disregard the rule against updating
6819 * exp_connect_data.ocd_connect_flags in this case, since
6820 * tgt_client_new() needs to know if this is client supports
6821 * multiple modify RPCs, and it is safe to expose this flag before
6822 * connection processing completes. */
6823 if (data->ocd_connect_flags & OBD_CONNECT_MULTIMODRPCS) {
6824 if (mdt_max_mod_rpcs_changed(mdt))
6825 /* The new mdt.*.max_mod_rpcs_in_flight parameter
6826 * has not changed since initialization, but the
6827 * deprecated module parameter was changed,
6828 * so use that instead.
6830 data->ocd_maxmodrpcs = max_mod_rpcs_per_client;
6832 data->ocd_maxmodrpcs = mdt->mdt_max_mod_rpcs_in_flight;
6833 spin_lock(&exp->exp_lock);
6834 *exp_connect_flags_ptr(exp) |= OBD_CONNECT_MULTIMODRPCS;
6835 spin_unlock(&exp->exp_lock);
6838 if (OCD_HAS_FLAG(data, CKSUM)) {
6839 __u32 cksum_types = data->ocd_cksum_types;
6841 tgt_mask_cksum_types(&mdt->mdt_lut, &data->ocd_cksum_types);
6843 if (unlikely(data->ocd_cksum_types == 0)) {
6844 CERROR("%s: Connect with checksum support but no "
6845 "ocd_cksum_types is set\n",
6846 exp->exp_obd->obd_name);
6850 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
6851 "%x\n", exp->exp_obd->obd_name, obd_export_nid2str(exp),
6852 cksum_types, data->ocd_cksum_types);
6854 /* This client does not support OBD_CONNECT_CKSUM
6855 * fall back to CRC32 */
6856 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
6857 "OBD_CONNECT_CKSUM, CRC32 will be used\n",
6858 exp->exp_obd->obd_name, obd_export_nid2str(exp));
6861 if ((data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
6862 !(data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT)) {
6863 atomic_inc(&mdt->mdt_mds_mds_conns);
6864 mdt_enable_slc(mdt);
6867 if (!mdt->mdt_lut.lut_dt_conf.ddp_has_lseek_data_hole)
6868 data->ocd_connect_flags2 &= ~OBD_CONNECT2_LSEEK;
6870 if (!OCD_HAS_FLAG(data, MDS_MDS) && !OCD_HAS_FLAG(data, LIGHTWEIGHT) &&
6871 !OCD_HAS_FLAG2(data, DMV_IMP_INHERIT)) {
6872 atomic_inc(&mdt->mdt_dmv_old_client_count);
6873 mdt->mdt_enable_dmv_implicit_inherit = 0;
6879 static int mdt_ctxt_add_dirty_flag(struct lu_env *env,
6880 struct mdt_thread_info *info,
6881 struct mdt_file_data *mfd)
6883 struct lu_context ses;
6887 rc = lu_context_init(&ses, LCT_SERVER_SESSION);
6892 lu_context_enter(&ses);
6894 mdt_ucred(info)->uc_valid = UCRED_OLD;
6895 /* do not let rbac interfere with dirty flag internal system event */
6896 mdt_ucred(info)->uc_rbac_file_perms = 1;
6897 mdt_ucred(info)->uc_rbac_dne_ops = 1;
6898 mdt_ucred(info)->uc_rbac_quota_ops = 1;
6899 mdt_ucred(info)->uc_rbac_byfid_ops = 1;
6900 mdt_ucred(info)->uc_rbac_chlg_ops = 1;
6901 mdt_ucred(info)->uc_rbac_fscrypt_admin = 1;
6902 rc = mdt_add_dirty_flag(info, mfd->mfd_object, &info->mti_attr);
6904 lu_context_exit(&ses);
6905 lu_context_fini(&ses);
6911 static int mdt_export_cleanup(struct obd_export *exp)
6913 LIST_HEAD(closing_list);
6914 struct mdt_export_data *med = &exp->exp_mdt_data;
6915 struct obd_device *obd = exp->exp_obd;
6916 struct mdt_device *mdt;
6917 struct mdt_thread_info *info;
6919 struct mdt_file_data *mfd, *n;
6923 spin_lock(&med->med_open_lock);
6924 while (!list_empty(&med->med_open_head)) {
6925 struct list_head *tmp = med->med_open_head.next;
6926 mfd = list_entry(tmp, struct mdt_file_data, mfd_list);
6928 /* Remove mfd handle so it can't be found again.
6929 * We are consuming the mfd_list reference here. */
6930 class_handle_unhash(&mfd->mfd_open_handle);
6931 list_move_tail(&mfd->mfd_list, &closing_list);
6933 spin_unlock(&med->med_open_lock);
6934 mdt = mdt_dev(obd->obd_lu_dev);
6935 LASSERT(mdt != NULL);
6937 rc = lu_env_init(&env, LCT_MD_THREAD);
6941 info = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
6942 LASSERT(info != NULL);
6943 memset(info, 0, sizeof *info);
6944 info->mti_env = &env;
6945 info->mti_mdt = mdt;
6946 info->mti_exp = exp;
6948 if (!list_empty(&closing_list)) {
6949 struct md_attr *ma = &info->mti_attr;
6951 /* Close any open files (which may also cause orphan
6953 list_for_each_entry_safe(mfd, n, &closing_list, mfd_list) {
6954 list_del_init(&mfd->mfd_list);
6955 ma->ma_need = ma->ma_valid = 0;
6957 /* This file is being closed due to an eviction, it
6958 * could have been modified and now dirty regarding to
6959 * HSM archive, check this!
6960 * The logic here is to mark a file dirty if there's a
6961 * chance it was dirtied before the client was evicted,
6962 * so that we don't have to wait for a release attempt
6963 * before finding out the file was actually dirty and
6964 * fail the release. Aggressively marking it dirty here
6965 * will cause the policy engine to attempt to
6966 * re-archive it; when rearchiving, we can compare the
6967 * current version to the HSM data_version and make the
6968 * archive request into a noop if it's not actually
6971 if (mfd->mfd_open_flags & MDS_FMODE_WRITE)
6972 rc = mdt_ctxt_add_dirty_flag(&env, info, mfd);
6974 /* Don't unlink orphan on failover umount, LU-184 */
6975 if (exp->exp_flags & OBD_OPT_FAILOVER ||
6976 exp->exp_obd->obd_stopping) {
6977 ma->ma_valid = MA_FLAGS;
6978 ma->ma_attr_flags |= MDS_KEEP_ORPHAN;
6980 ma->ma_valid |= MA_FORCE_LOG;
6981 mdt_mfd_close(info, mfd);
6984 info->mti_mdt = NULL;
6985 /* cleanup client slot early */
6986 /* Do not erase record for recoverable client. */
6987 if (!(exp->exp_flags & OBD_OPT_FAILOVER) || exp->exp_failed)
6988 tgt_client_del(&env, exp);
6994 static int mdt_obd_disconnect(struct obd_export *exp)
6996 struct obd_connect_data *data = &exp->exp_connect_data;
6997 struct mdt_device *mdt = mdt_dev(exp->exp_obd->obd_lu_dev);
7003 class_export_get(exp);
7005 if (!(exp->exp_flags & OBD_OPT_FORCE))
7006 tgt_grant_sanity_check(exp->exp_obd, __func__);
7008 if (OCD_HAS_FLAG(data, MDS_MDS) && !OCD_HAS_FLAG(data, LIGHTWEIGHT) &&
7009 atomic_dec_and_test(&mdt->mdt_mds_mds_conns))
7010 mdt_disable_slc(mdt);
7012 if (!OCD_HAS_FLAG(data, MDS_MDS) && !OCD_HAS_FLAG(data, LIGHTWEIGHT) &&
7013 !OCD_HAS_FLAG2(data, DMV_IMP_INHERIT) &&
7014 atomic_dec_and_test(&mdt->mdt_dmv_old_client_count))
7015 mdt->mdt_enable_dmv_implicit_inherit = 1;
7017 rc = server_disconnect_export(exp);
7019 CDEBUG(D_IOCTL, "server disconnect error: rc = %d\n", rc);
7021 tgt_grant_discard(exp);
7023 rc = mdt_export_cleanup(exp);
7024 nodemap_del_member(exp);
7025 class_export_put(exp);
7029 /* mds_connect copy */
7030 static int mdt_obd_connect(const struct lu_env *env,
7031 struct obd_export **exp, struct obd_device *obd,
7032 struct obd_uuid *cluuid,
7033 struct obd_connect_data *data,
7036 struct obd_export *lexp;
7037 struct lustre_handle conn = { 0 };
7038 struct mdt_device *mdt;
7040 struct lnet_nid *client_nid = localdata;
7043 LASSERT(env != NULL);
7044 LASSERT(data != NULL);
7046 if (!exp || !obd || !cluuid)
7049 mdt = mdt_dev(obd->obd_lu_dev);
7052 * first, check whether the stack is ready to handle requests
7053 * XXX: probably not very appropriate method is used now
7054 * at some point we should find a better one
7056 if (!test_bit(MDT_FL_SYNCED, &mdt->mdt_state) &&
7057 !(data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) &&
7058 !(data->ocd_connect_flags & OBD_CONNECT_MDS_MDS)) {
7059 rc = obd_get_info(env, mdt->mdt_child_exp,
7060 sizeof(KEY_OSP_CONNECTED),
7061 KEY_OSP_CONNECTED, NULL, NULL);
7064 set_bit(MDT_FL_SYNCED, &mdt->mdt_state);
7067 rc = class_connect(&conn, obd, cluuid);
7071 lexp = class_conn2export(&conn);
7072 LASSERT(lexp != NULL);
7074 if (nid_is_nid4(client_nid)) {
7075 rc = nodemap_add_member(lnet_nid_to_nid4(client_nid), lexp);
7076 if (rc != 0 && rc != -EEXIST)
7080 rc = mdt_connect_internal(env, lexp, mdt, data, false);
7082 struct lsd_client_data *lcd = lexp->exp_target_data.ted_lcd;
7085 memcpy(lcd->lcd_uuid, cluuid, sizeof lcd->lcd_uuid);
7086 rc = tgt_client_new(env, lexp);
7088 mdt_export_stats_init(obd, lexp, localdata);
7092 class_disconnect(lexp);
7093 nodemap_del_member(lexp);
7097 /* Because we do not want this export to be evicted by pinger,
7098 * let's not add this export to the timed chain list. */
7099 if (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) {
7100 spin_lock(&lexp->exp_obd->obd_dev_lock);
7101 list_del_init(&lexp->exp_obd_chain_timed);
7102 spin_unlock(&lexp->exp_obd->obd_dev_lock);
7109 static int mdt_obd_reconnect(const struct lu_env *env,
7110 struct obd_export *exp, struct obd_device *obd,
7111 struct obd_uuid *cluuid,
7112 struct obd_connect_data *data,
7115 struct lnet_nid *client_nid = localdata;
7119 if (exp == NULL || obd == NULL || cluuid == NULL)
7122 if (nid_is_nid4(client_nid)) {
7123 rc = nodemap_add_member(lnet_nid_to_nid4(client_nid), exp);
7124 if (rc != 0 && rc != -EEXIST)
7128 rc = mdt_connect_internal(env, exp, mdt_dev(obd->obd_lu_dev), data,
7131 mdt_export_stats_init(obd, exp, localdata);
7133 nodemap_del_member(exp);
7138 /* FIXME: Can we avoid using these two interfaces? */
7139 static int mdt_init_export(struct obd_export *exp)
7141 struct mdt_export_data *med = &exp->exp_mdt_data;
7145 INIT_LIST_HEAD(&med->med_open_head);
7146 spin_lock_init(&med->med_open_lock);
7147 spin_lock(&exp->exp_lock);
7148 exp->exp_connecting = 1;
7149 spin_unlock(&exp->exp_lock);
7151 OBD_ALLOC(exp->exp_used_slots,
7152 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
7153 if (exp->exp_used_slots == NULL)
7156 /* self-export doesn't need client data and ldlm initialization */
7157 if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
7158 &exp->exp_client_uuid)))
7161 rc = tgt_client_alloc(exp);
7165 rc = ldlm_init_export(exp);
7172 tgt_client_free(exp);
7174 OBD_FREE(exp->exp_used_slots,
7175 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
7176 exp->exp_used_slots = NULL;
7178 CERROR("%s: Failed to initialize export: rc = %d\n",
7179 exp->exp_obd->obd_name, rc);
7183 static int mdt_destroy_export(struct obd_export *exp)
7187 target_destroy_export(exp);
7188 if (exp->exp_used_slots)
7189 OBD_FREE(exp->exp_used_slots,
7190 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
7192 /* destroy can be called from failed obd_setup, so
7193 * checking uuid is safer than obd_self_export */
7194 if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
7195 &exp->exp_client_uuid)))
7198 ldlm_destroy_export(exp);
7199 tgt_client_free(exp);
7201 LASSERT(list_empty(&exp->exp_outstanding_replies));
7202 LASSERT(list_empty(&exp->exp_mdt_data.med_open_head));
7205 * discard grants once we're sure no more
7206 * interaction with the client is possible
7208 tgt_grant_discard(exp);
7209 if (exp_connect_flags(exp) & OBD_CONNECT_GRANT)
7210 obd2obt(exp->exp_obd)->obt_lut->lut_tgd.tgd_tot_granted_clients--;
7212 if (!(exp->exp_flags & OBD_OPT_FORCE))
7213 tgt_grant_sanity_check(exp->exp_obd, __func__);
7218 int mdt_links_read(struct mdt_thread_info *info, struct mdt_object *mdt_obj,
7219 struct linkea_data *ldata)
7223 LASSERT(ldata->ld_buf->lb_buf != NULL);
7225 if (!mdt_object_exists(mdt_obj))
7228 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
7229 ldata->ld_buf, XATTR_NAME_LINK);
7230 if (rc == -ERANGE) {
7231 /* Buf was too small, figure out what we need. */
7232 lu_buf_free(ldata->ld_buf);
7233 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
7234 ldata->ld_buf, XATTR_NAME_LINK);
7237 ldata->ld_buf = lu_buf_check_and_alloc(ldata->ld_buf, rc);
7238 if (ldata->ld_buf->lb_buf == NULL)
7240 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
7241 ldata->ld_buf, XATTR_NAME_LINK);
7246 return linkea_init_with_rec(ldata);
7250 * Given an MDT object, try to look up the full path to the object.
7251 * Part of the MDT layer implementation of lfs fid2path.
7253 * \param[in] info Per-thread common data shared by MDT level handlers.
7254 * \param[in] obj Object to do path lookup of
7255 * \param[in,out] fp User-provided struct to store path information
7256 * \param[in] root_fid Root FID of current path should reach
7258 * \retval 0 Lookup successful, path information stored in fp
7259 * \retval -EAGAIN Lookup failed, usually because object is being moved
7260 * \retval negative errno if there was a problem
7262 static int mdt_path_current(struct mdt_thread_info *info,
7263 struct mdt_object *obj,
7264 struct getinfo_fid2path *fp,
7265 struct lu_fid *root_fid)
7267 struct mdt_device *mdt = info->mti_mdt;
7268 struct lu_name *tmpname = &info->mti_name;
7269 struct lu_fid *tmpfid = &info->mti_tmp_fid1;
7270 struct lu_buf *buf = &info->mti_big_buf;
7271 struct linkea_data ldata = { NULL };
7273 struct mdt_object *mdt_obj;
7274 struct link_ea_header *leh;
7275 struct link_ea_entry *lee;
7276 bool worthchecking = true;
7277 bool needsfid = false;
7278 bool supported = false;
7286 /* temp buffer for path element, the buffer will be finally freed
7287 * in mdt_thread_info_fini */
7288 buf = lu_buf_check_and_alloc(buf, PATH_MAX);
7289 if (buf->lb_buf == NULL)
7293 ptr = fp->gf_u.gf_path + fp->gf_pathlen - 1;
7296 *tmpfid = fp->gf_fid = *mdt_object_fid(obj);
7298 while (!lu_fid_eq(root_fid, &fp->gf_fid)) {
7299 if (!lu_fid_eq(root_fid, &mdt->mdt_md_root_fid) &&
7300 lu_fid_eq(&mdt->mdt_md_root_fid, &fp->gf_fid))
7301 GOTO(out, rc = -ENOENT);
7303 if (lu_fid_eq(mdt_object_fid(obj), tmpfid)) {
7305 mdt_object_get(info->mti_env, mdt_obj);
7307 mdt_obj = mdt_object_find(info->mti_env, mdt, tmpfid);
7308 if (IS_ERR(mdt_obj))
7309 GOTO(out, rc = PTR_ERR(mdt_obj));
7312 if (!mdt_object_exists(mdt_obj)) {
7313 mdt_object_put(info->mti_env, mdt_obj);
7314 GOTO(out, rc = -ENOENT);
7317 if (mdt_object_remote(mdt_obj)) {
7318 mdt_object_put(info->mti_env, mdt_obj);
7319 GOTO(remote_out, rc = -EREMOTE);
7322 if (worthchecking) {
7323 /* we need to know if the FID being
7324 * looked up is encrypted
7326 struct lu_attr la = { 0 };
7327 struct dt_object *dt = mdt_obj2dt(mdt_obj);
7329 if (dt && dt->do_ops && dt->do_ops->do_attr_get)
7330 dt_attr_get(info->mti_env, dt, &la);
7331 if (la.la_valid & LA_FLAGS &&
7332 la.la_flags & LUSTRE_ENCRYPT_FL) {
7333 if (!supported && mdt_info_req(info) &&
7334 !exp_connect_encrypt_fid2path(
7335 mdt_info_req(info)->rq_export)) {
7336 /* client does not support fid2path
7337 * for encrypted files
7339 mdt_object_put(info->mti_env, mdt_obj);
7340 GOTO(out, rc = -ENODATA);
7348 worthchecking = false;
7355 rc = mdt_links_read(info, mdt_obj, &ldata);
7357 mdt_object_put(info->mti_env, mdt_obj);
7362 lee = (struct link_ea_entry *)(leh + 1); /* link #0 */
7363 linkea_entry_unpack(lee, &reclen, tmpname, tmpfid);
7364 /* If set, use link #linkno for path lookup, otherwise use
7365 link #0. Only do this for the final path element. */
7366 if (first && fp->gf_linkno < leh->leh_reccount) {
7368 for (count = 0; count < fp->gf_linkno; count++) {
7369 lee = (struct link_ea_entry *)
7370 ((char *)lee + reclen);
7371 linkea_entry_unpack(lee, &reclen, tmpname,
7374 if (fp->gf_linkno < leh->leh_reccount - 1)
7375 /* indicate to user there are more links */
7379 /* Check if it is slave stripes */
7380 rc = mdt_is_dir_stripe(info, mdt_obj);
7381 mdt_object_put(info->mti_env, mdt_obj);
7385 fp->gf_fid = *tmpfid;
7389 /* Pack the name in the end of the buffer */
7390 ptr -= tmpname->ln_namelen;
7391 if (ptr - 1 <= fp->gf_u.gf_path)
7392 GOTO(out, rc = -ENAMETOOLONG);
7393 strncpy(ptr, tmpname->ln_name, tmpname->ln_namelen);
7395 /* Pack FID before file name, so that client can build
7396 * encoded/digested form.
7398 char fidstr[FID_LEN + 1];
7400 snprintf(fidstr, sizeof(fidstr), DFID,
7402 ptr -= strlen(fidstr);
7403 if (ptr - 1 <= fp->gf_u.gf_path)
7404 GOTO(out, rc = -ENAMETOOLONG);
7405 strncpy(ptr, fidstr, strlen(fidstr));
7409 /* keep the last resolved fid to the client, so the
7410 * client will build the left path on another MDT for
7412 fp->gf_fid = *tmpfid;
7417 /* non-zero will be treated as an error */
7422 ptr++; /* skip leading / unless this is an encrypted file */
7423 memmove(fp->gf_u.gf_path, ptr,
7424 fp->gf_u.gf_path + fp->gf_pathlen - ptr);
7431 * Given an MDT object, use mdt_path_current to get the path.
7432 * Essentially a wrapper to retry mdt_path_current a set number of times
7433 * if -EAGAIN is returned (usually because an object is being moved).
7435 * Part of the MDT layer implementation of lfs fid2path.
7437 * \param[in] info Per-thread common data shared by mdt level handlers.
7438 * \param[in] obj Object to do path lookup of
7439 * \param[in,out] fp User-provided struct for arguments and to store path
7442 * \retval 0 Lookup successful, path information stored in fp
7443 * \retval negative errno if there was a problem
7445 static int mdt_path(struct mdt_thread_info *info, struct mdt_object *obj,
7446 struct getinfo_fid2path *fp, struct lu_fid *root_fid)
7448 struct mdt_device *mdt = info->mti_mdt;
7453 if (fp->gf_pathlen < 3)
7456 if (root_fid == NULL)
7457 root_fid = &mdt->mdt_md_root_fid;
7459 if (lu_fid_eq(root_fid, mdt_object_fid(obj))) {
7460 fp->gf_u.gf_path[0] = '\0';
7464 /* Retry multiple times in case file is being moved */
7465 while (tries-- && rc == -EAGAIN)
7466 rc = mdt_path_current(info, obj, fp, root_fid);
7472 * Get the full path of the provided FID, as of changelog record recno.
7474 * This checks sanity and looks up object for user provided FID
7475 * before calling the actual path lookup code.
7477 * Part of the MDT layer implementation of lfs fid2path.
7479 * \param[in] info Per-thread common data shared by mdt level handlers.
7480 * \param[in,out] fp User-provided struct for arguments and to store path
7483 * \retval 0 Lookup successful, path information and recno stored in fp
7484 * \retval -ENOENT, object does not exist
7485 * \retval negative errno if there was a problem
7487 static int mdt_fid2path(struct mdt_thread_info *info,
7488 struct lu_fid *root_fid,
7489 struct getinfo_fid2path *fp)
7491 struct mdt_device *mdt = info->mti_mdt;
7492 struct mdt_object *obj;
7496 CDEBUG(D_IOCTL, "path get "DFID" from %llu #%d\n",
7497 PFID(&fp->gf_fid), fp->gf_recno, fp->gf_linkno);
7499 if (!fid_is_sane(&fp->gf_fid))
7502 if (!fid_is_namespace_visible(&fp->gf_fid)) {
7503 CDEBUG(D_INFO, "%s: "DFID" is invalid, f_seq should be >= %#llx"
7504 ", or f_oid != 0, or f_ver == 0\n", mdt_obd_name(mdt),
7505 PFID(&fp->gf_fid), (__u64)FID_SEQ_NORMAL);
7509 /* return error if client-provided root fid
7510 * is not the one stored in the export
7512 if (root_fid && !fid_is_zero(&info->mti_exp->exp_root_fid) &&
7513 !lu_fid_eq(root_fid, &info->mti_exp->exp_root_fid)) {
7515 "%s: root fid from client "DFID" but "DFID" stored in export\n",
7516 mdt_obd_name(mdt), PFID(root_fid),
7517 PFID(&info->mti_exp->exp_root_fid));
7521 obj = mdt_object_find(info->mti_env, mdt, &fp->gf_fid);
7524 CDEBUG(D_IOCTL, "cannot find "DFID": rc = %d\n",
7525 PFID(&fp->gf_fid), rc);
7529 if (mdt_object_remote(obj))
7531 else if (!mdt_object_exists(obj))
7537 mdt_object_put(info->mti_env, obj);
7538 CDEBUG(D_IOCTL, "nonlocal object "DFID": rc = %d\n",
7539 PFID(&fp->gf_fid), rc);
7543 rc = mdt_path(info, obj, fp, root_fid);
7545 CDEBUG(D_INFO, "fid "DFID", path %s recno %#llx linkno %u\n",
7546 PFID(&fp->gf_fid), fp->gf_u.gf_path,
7547 fp->gf_recno, fp->gf_linkno);
7549 mdt_object_put(info->mti_env, obj);
7554 static int mdt_rpc_fid2path(struct mdt_thread_info *info, void *key, int keylen,
7555 void *val, int vallen)
7557 struct getinfo_fid2path *fpout, *fpin;
7558 struct lu_fid *root_fid = NULL;
7561 fpin = key + round_up(sizeof(KEY_FID2PATH), 8);
7564 if (req_capsule_req_need_swab(info->mti_pill))
7565 lustre_swab_fid2path(fpin);
7567 memcpy(fpout, fpin, sizeof(*fpin));
7568 if (fpout->gf_pathlen != vallen - sizeof(*fpin))
7571 if (keylen >= round_up(sizeof(KEY_FID2PATH), 8) + sizeof(*fpin) +
7572 sizeof(struct lu_fid)) {
7573 /* client sent its root FID, which is normally fileset FID */
7574 root_fid = fpin->gf_u.gf_root_fid;
7575 if (req_capsule_req_need_swab(info->mti_pill))
7576 lustre_swab_lu_fid(root_fid);
7578 if (root_fid != NULL && !fid_is_sane(root_fid))
7582 rc = mdt_fid2path(info, root_fid, fpout);
7586 int mdt_get_info(struct tgt_session_info *tsi)
7596 key = req_capsule_client_get(tsi->tsi_pill, &RMF_GETINFO_KEY);
7598 CDEBUG(D_IOCTL, "No GETINFO key\n");
7599 RETURN(err_serious(-EFAULT));
7601 keylen = req_capsule_get_size(tsi->tsi_pill, &RMF_GETINFO_KEY,
7604 vallen = req_capsule_client_get(tsi->tsi_pill, &RMF_GETINFO_VALLEN);
7605 if (vallen == NULL) {
7606 CDEBUG(D_IOCTL, "%s: cannot get RMF_GETINFO_VALLEN buffer\n",
7607 tgt_name(tsi->tsi_tgt));
7608 RETURN(err_serious(-EFAULT));
7611 req_capsule_set_size(tsi->tsi_pill, &RMF_GETINFO_VAL, RCL_SERVER,
7613 rc = req_capsule_server_pack(tsi->tsi_pill);
7615 RETURN(err_serious(rc));
7617 valout = req_capsule_server_get(tsi->tsi_pill, &RMF_GETINFO_VAL);
7618 if (valout == NULL) {
7619 CDEBUG(D_IOCTL, "%s: cannot get get-info RPC out buffer\n",
7620 tgt_name(tsi->tsi_tgt));
7621 RETURN(err_serious(-EFAULT));
7624 if (KEY_IS(KEY_FID2PATH)) {
7625 struct mdt_thread_info *info = tsi2mdt_info(tsi);
7627 rc = mdt_rpc_fid2path(info, key, keylen, valout, *vallen);
7628 mdt_thread_info_fini(info);
7635 static int mdt_ioc_version_get(struct mdt_thread_info *mti, void *karg)
7637 struct obd_ioctl_data *data = karg;
7640 struct mdt_object *obj;
7641 struct mdt_lock_handle *lh;
7645 if (data->ioc_inlbuf1 == NULL || data->ioc_inllen1 != sizeof(*fid) ||
7646 data->ioc_inlbuf2 == NULL || data->ioc_inllen2 != sizeof(version))
7649 fid = (struct lu_fid *)data->ioc_inlbuf1;
7651 if (!fid_is_sane(fid))
7654 CDEBUG(D_IOCTL, "getting version for "DFID"\n", PFID(fid));
7656 lh = &mti->mti_lh[MDT_LH_PARENT];
7657 obj = mdt_object_find_lock(mti, fid, lh, MDS_INODELOCK_UPDATE, LCK_CR);
7659 RETURN(PTR_ERR(obj));
7661 if (mdt_object_remote(obj)) {
7664 * before calling version get the correct MDS should be
7665 * fid, this is error to find remote object here
7667 CERROR("nonlocal object "DFID"\n", PFID(fid));
7668 } else if (!mdt_object_exists(obj)) {
7669 *(__u64 *)data->ioc_inlbuf2 = ENOENT_VERSION;
7672 version = dt_version_get(mti->mti_env, mdt_obj2dt(obj));
7673 *(__u64 *)data->ioc_inlbuf2 = version;
7676 mdt_object_unlock_put(mti, obj, lh, 1);
7680 /* ioctls on obd dev */
7681 static int mdt_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
7682 void *karg, void __user *uarg)
7684 struct obd_device *obd = exp->exp_obd;
7685 struct mdt_device *mdt = mdt_dev(obd->obd_lu_dev);
7686 struct dt_device *dt = mdt->mdt_bottom;
7687 struct obd_ioctl_data *data;
7692 CDEBUG(D_IOCTL, "%s: cmd=%x len=%u karg=%pK uarg=%pK\n",
7693 obd->obd_name, cmd, len, karg, uarg);
7695 rc = lu_env_init(&env, LCT_MD_THREAD);
7699 /* handle commands that don't use @karg first */
7702 rc = mdt_device_sync(&env, mdt);
7704 case OBD_IOC_SET_READONLY:
7705 rc = dt_sync(&env, dt);
7707 rc = dt_ro(&env, dt);
7711 if (unlikely(karg == NULL)) {
7712 OBD_IOC_ERROR(obd->obd_name, cmd, "karg=NULL", rc = -EINVAL);
7718 case OBD_IOC_ABORT_RECOVERY: {
7719 if (data->ioc_type & OBD_FLG_ABORT_RECOV_MDT) {
7720 LCONSOLE_WARN("%s: Aborting MDT recovery\n",
7722 obd->obd_abort_mdt_recovery = 1;
7723 wake_up(&obd->obd_next_transno_waitq);
7724 } else { /* if (data->ioc_type & OBD_FLG_ABORT_RECOV_OST) */
7725 /* lctl didn't set OBD_FLG_ABORT_RECOV_OST < 2.13.57 */
7726 LCONSOLE_WARN("%s: Aborting client recovery\n",
7728 obd->obd_abort_recovery = 1;
7729 target_stop_recovery_thread(obd);
7734 case OBD_IOC_CHANGELOG_REG:
7735 case OBD_IOC_CHANGELOG_DEREG:
7736 case OBD_IOC_CHANGELOG_CLEAR:
7737 case OBD_IOC_LLOG_PRINT:
7738 case OBD_IOC_LLOG_CANCEL:
7739 rc = mdt->mdt_child->md_ops->mdo_iocontrol(&env, mdt->mdt_child,
7742 case OBD_IOC_START_LFSCK: {
7743 struct md_device *next = mdt->mdt_child;
7744 struct lfsck_start_param lsp;
7746 lsp.lsp_start = (struct lfsck_start *)(data->ioc_inlbuf1);
7747 lsp.lsp_index_valid = 0;
7748 rc = next->md_ops->mdo_iocontrol(&env, next, cmd, 0, &lsp);
7751 case OBD_IOC_STOP_LFSCK: {
7752 struct md_device *next = mdt->mdt_child;
7753 struct lfsck_stop stop;
7755 stop.ls_status = LS_STOPPED;
7756 /* Old lfsck utils may pass NULL @stop. */
7757 if (data->ioc_inlbuf1 == NULL)
7761 ((struct lfsck_stop *)(data->ioc_inlbuf1))->ls_flags;
7763 rc = next->md_ops->mdo_iocontrol(&env, next, cmd, 0, &stop);
7766 case OBD_IOC_QUERY_LFSCK: {
7767 struct md_device *next = mdt->mdt_child;
7769 rc = next->md_ops->mdo_iocontrol(&env, next, cmd, 0,
7773 case OBD_IOC_GET_OBJ_VERSION: {
7774 struct mdt_thread_info *mti;
7776 mti = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
7777 memset(mti, 0, sizeof(*mti));
7778 mti->mti_env = &env;
7782 rc = mdt_ioc_version_get(mti, karg);
7785 case OBD_IOC_CATLOGLIST: {
7786 struct mdt_thread_info *mti;
7788 mti = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
7789 lu_local_obj_fid(&mti->mti_tmp_fid1, LLOG_CATALOGS_OID);
7790 rc = llog_catalog_list(&env, mdt->mdt_bottom, 0, karg,
7791 &mti->mti_tmp_fid1);
7795 rc = OBD_IOC_ERROR(obd->obd_name, cmd, "unrecognized", -ENOTTY);
7803 static int mdt_postrecov(const struct lu_env *env, struct mdt_device *mdt)
7805 struct lu_device *ld = md2lu_dev(mdt->mdt_child);
7809 if (!mdt->mdt_skip_lfsck && !mdt->mdt_bottom->dd_rdonly) {
7810 struct lfsck_start_param lsp;
7812 lsp.lsp_start = NULL;
7813 lsp.lsp_index_valid = 0;
7814 rc = mdt->mdt_child->md_ops->mdo_iocontrol(env, mdt->mdt_child,
7815 OBD_IOC_START_LFSCK,
7817 if (rc != 0 && rc != -EALREADY)
7818 CWARN("%s: auto trigger paused LFSCK failed: rc = %d\n",
7819 mdt_obd_name(mdt), rc);
7822 rc = ld->ld_ops->ldo_recovery_complete(env, ld);
7826 static int mdt_obd_postrecov(struct obd_device *obd)
7831 rc = lu_env_init(&env, LCT_MD_THREAD);
7834 rc = mdt_postrecov(&env, mdt_dev(obd->obd_lu_dev));
7839 static const struct obd_ops mdt_obd_device_ops = {
7840 .o_owner = THIS_MODULE,
7841 .o_set_info_async = mdt_obd_set_info_async,
7842 .o_connect = mdt_obd_connect,
7843 .o_reconnect = mdt_obd_reconnect,
7844 .o_disconnect = mdt_obd_disconnect,
7845 .o_init_export = mdt_init_export,
7846 .o_destroy_export = mdt_destroy_export,
7847 .o_iocontrol = mdt_iocontrol,
7848 .o_postrecov = mdt_obd_postrecov,
7849 /* Data-on-MDT IO methods */
7850 .o_preprw = mdt_obd_preprw,
7851 .o_commitrw = mdt_obd_commitrw,
7854 static struct lu_device* mdt_device_fini(const struct lu_env *env,
7855 struct lu_device *d)
7857 struct mdt_device *m = mdt_dev(d);
7864 static struct lu_device *mdt_device_free(const struct lu_env *env,
7865 struct lu_device *d)
7867 struct mdt_device *m = mdt_dev(d);
7870 lu_device_fini(&m->mdt_lu_dev);
7876 static struct lu_device *mdt_device_alloc(const struct lu_env *env,
7877 struct lu_device_type *t,
7878 struct lustre_cfg *cfg)
7880 struct lu_device *l;
7881 struct mdt_device *m;
7888 rc = mdt_init0(env, m, t, cfg);
7890 mdt_device_free(env, l);
7895 l = ERR_PTR(-ENOMEM);
7899 /* context key constructor/destructor: mdt_key_init, mdt_key_fini */
7900 LU_KEY_INIT(mdt, struct mdt_thread_info);
7902 static void mdt_key_fini(const struct lu_context *ctx,
7903 struct lu_context_key *key, void* data)
7905 struct mdt_thread_info *info = data;
7907 if (info->mti_big_lmm) {
7908 OBD_FREE_LARGE(info->mti_big_lmm, info->mti_big_lmmsize);
7909 info->mti_big_lmm = NULL;
7910 info->mti_big_lmmsize = 0;
7913 if (info->mti_big_acl) {
7914 OBD_FREE_LARGE(info->mti_big_acl, info->mti_big_aclsize);
7915 info->mti_big_acl = NULL;
7916 info->mti_big_aclsize = 0;
7922 /* context key: mdt_thread_key */
7923 LU_CONTEXT_KEY_DEFINE(mdt, LCT_MD_THREAD);
7925 struct lu_ucred *mdt_ucred(const struct mdt_thread_info *info)
7927 return lu_ucred(info->mti_env);
7930 struct lu_ucred *mdt_ucred_check(const struct mdt_thread_info *info)
7932 return lu_ucred_check(info->mti_env);
7936 * Enable/disable COS (Commit On Sharing).
7938 * Set/Clear the COS flag in mdt options.
7940 * \param mdt mdt device
7941 * \param val 0 disables COS, other values enable COS
7943 void mdt_enable_cos(struct mdt_device *mdt, bool val)
7948 mdt->mdt_opts.mo_cos = val;
7949 rc = lu_env_init(&env, LCT_LOCAL);
7950 if (unlikely(rc != 0)) {
7951 CWARN("%s: lu_env initialization failed, cannot "
7952 "sync: rc = %d\n", mdt_obd_name(mdt), rc);
7955 mdt_device_sync(&env, mdt);
7960 * Check COS (Commit On Sharing) status.
7962 * Return COS flag status.
7964 * \param mdt mdt device
7966 int mdt_cos_is_enabled(struct mdt_device *mdt)
7968 return mdt->mdt_opts.mo_cos != 0;
7971 static const struct lu_device_type_operations mdt_device_type_ops = {
7972 .ldto_device_alloc = mdt_device_alloc,
7973 .ldto_device_free = mdt_device_free,
7974 .ldto_device_fini = mdt_device_fini
7977 static struct lu_device_type mdt_device_type = {
7978 .ldt_tags = LU_DEVICE_MD,
7979 .ldt_name = LUSTRE_MDT_NAME,
7980 .ldt_ops = &mdt_device_type_ops,
7981 .ldt_ctx_tags = LCT_MD_THREAD
7984 static int __init mdt_init(void)
7988 BUILD_BUG_ON(sizeof("0x0123456789ABCDEF:0x01234567:0x01234567") !=
7989 FID_NOBRACE_LEN + 1);
7990 BUILD_BUG_ON(sizeof("[0x0123456789ABCDEF:0x01234567:0x01234567]") !=
7992 rc = lu_kmem_init(mdt_caches);
7996 rc = mds_mod_init();
8000 rc = class_register_type(&mdt_obd_device_ops, NULL, true,
8001 LUSTRE_MDT_NAME, &mdt_device_type);
8006 lu_kmem_fini(mdt_caches);
8013 static void __exit mdt_exit(void)
8015 class_unregister_type(LUSTRE_MDT_NAME);
8017 lu_kmem_fini(mdt_caches);
8020 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
8021 MODULE_DESCRIPTION("Lustre Metadata Target ("LUSTRE_MDT_NAME")");
8022 MODULE_VERSION(LUSTRE_VERSION_STRING);
8023 MODULE_LICENSE("GPL");
8025 module_init(mdt_init);
8026 module_exit(mdt_exit);