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;
500 svcpt = req->rq_rqbd->rqbd_svcpt;
502 /* This will trigger a watchdog timeout */
503 at_est = obd_at_get(mdt->mdt_lu_dev.ld_obd, &svcpt->scp_at_estimate);
504 CFS_FAIL_TIMEOUT(OBD_FAIL_MDS_STATFS_LCW_SLEEP,
505 (MDT_SERVICE_WATCHDOG_FACTOR * at_est) + 1);
507 rc = mdt_check_ucred(info);
509 GOTO(out, rc = err_serious(rc));
511 if (CFS_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK))
512 GOTO(out, rc = err_serious(-ENOMEM));
514 osfs = req_capsule_server_get(info->mti_pill, &RMF_OBD_STATFS);
516 GOTO(out, rc = -EPROTO);
518 if (mdt_is_sum_statfs_client(req->rq_export) &&
519 lustre_packed_msg_size(req->rq_reqmsg) ==
520 req_capsule_fmt_size(req->rq_reqmsg->lm_magic,
521 &RQF_MDS_STATFS_NEW, RCL_CLIENT)) {
522 req_capsule_extend(info->mti_pill, &RQF_MDS_STATFS_NEW);
523 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
526 if (reqbody && reqbody->mbo_valid & OBD_MD_FLAGSTATFS)
527 msf = &mdt->mdt_sum_osfs;
529 msf = &mdt->mdt_osfs;
531 if (msf->msf_age + OBD_STATFS_CACHE_SECONDS <= ktime_get_seconds()) {
532 /** statfs data is too old, get up-to-date one */
533 if (reqbody && reqbody->mbo_valid & OBD_MD_FLAGSTATFS)
534 rc = next->md_ops->mdo_statfs(info->mti_env,
537 rc = dt_statfs(info->mti_env, mdt->mdt_bottom,
541 spin_lock(&mdt->mdt_lock);
542 msf->msf_osfs = *osfs;
543 msf->msf_age = ktime_get_seconds();
544 spin_unlock(&mdt->mdt_lock);
546 /** use cached statfs data */
547 spin_lock(&mdt->mdt_lock);
548 *osfs = msf->msf_osfs;
549 spin_unlock(&mdt->mdt_lock);
552 /* tgd_blockbit is recordsize bits set during mkfs.
553 * This once set does not change. However, 'zfs set'
554 * can be used to change the MDT blocksize. Instead
555 * of using cached value of 'tgd_blockbit' always
556 * calculate the blocksize bits which may have
559 current_blockbits = fls64(osfs->os_bsize) - 1;
561 /* at least try to account for cached pages. its still racy and
562 * might be under-reporting if clients haven't announced their
563 * caches with brw recently */
564 CDEBUG(D_SUPER | D_CACHE, "blocks cached %llu granted %llu"
565 " pending %llu free %llu avail %llu\n",
566 tgd->tgd_tot_dirty, tgd->tgd_tot_granted,
567 tgd->tgd_tot_pending,
568 osfs->os_bfree << current_blockbits,
569 osfs->os_bavail << current_blockbits);
571 osfs->os_bavail -= min_t(u64, osfs->os_bavail,
572 ((tgd->tgd_tot_dirty + tgd->tgd_tot_pending +
573 osfs->os_bsize - 1) >> current_blockbits));
575 tgt_grant_sanity_check(mdt->mdt_lu_dev.ld_obd, __func__);
576 if (mdt->mdt_lut.lut_no_create)
577 osfs->os_state |= OS_STATFS_NOCREATE;
578 CDEBUG(D_CACHE, "%llu blocks: %llu free, %llu avail; "
579 "%llu objects: %llu free; state %x\n",
580 osfs->os_blocks, osfs->os_bfree, osfs->os_bavail,
581 osfs->os_files, osfs->os_ffree, osfs->os_state);
583 if (!exp_grant_param_supp(tsi->tsi_exp) &&
584 current_blockbits > COMPAT_BSIZE_SHIFT) {
585 /* clients which don't support OBD_CONNECT_GRANT_PARAM
586 * should not see a block size > page size, otherwise
587 * cl_lost_grant goes mad. Therefore, we emulate a 4KB (=2^12)
588 * block size which is the biggest block size known to work
589 * with all client's page size. */
590 osfs->os_blocks <<= current_blockbits - COMPAT_BSIZE_SHIFT;
591 osfs->os_bfree <<= current_blockbits - COMPAT_BSIZE_SHIFT;
592 osfs->os_bavail <<= current_blockbits - COMPAT_BSIZE_SHIFT;
593 osfs->os_bsize = 1 << COMPAT_BSIZE_SHIFT;
596 mdt_counter_incr(req, LPROC_MDT_STATFS,
597 ktime_us_delta(ktime_get(), kstart));
599 mdt_thread_info_fini(info);
603 __u32 mdt_lmm_dom_entry_check(struct lov_mds_md *lmm, int *is_dom_only)
605 struct lov_comp_md_v1 *comp_v1;
606 struct lov_mds_md *v1;
608 __u32 dom_stripesize = 0;
610 bool has_ost_stripes = false;
617 if (le32_to_cpu(lmm->lmm_magic) != LOV_MAGIC_COMP_V1)
620 comp_v1 = (struct lov_comp_md_v1 *)lmm;
621 off = le32_to_cpu(comp_v1->lcm_entries[0].lcme_offset);
622 v1 = (struct lov_mds_md *)((char *)comp_v1 + off);
624 /* Fast check for DoM entry with no mirroring, should be the first */
625 if (le16_to_cpu(comp_v1->lcm_mirror_count) == 0 &&
626 !(lov_pattern(le32_to_cpu(v1->lmm_pattern)) & LOV_PATTERN_MDT))
629 /* check all entries otherwise */
630 for (i = 0; i < le16_to_cpu(comp_v1->lcm_entry_count); i++) {
631 struct lov_comp_md_entry_v1 *lcme;
633 lcme = &comp_v1->lcm_entries[i];
634 if (!(le32_to_cpu(lcme->lcme_flags) & LCME_FL_INIT))
637 off = le32_to_cpu(lcme->lcme_offset);
638 v1 = (struct lov_mds_md *)((char *)comp_v1 + off);
640 if (lov_pattern(le32_to_cpu(v1->lmm_pattern)) &
642 dom_stripesize = le32_to_cpu(v1->lmm_stripe_size);
644 has_ost_stripes = true;
646 if (dom_stripesize && has_ost_stripes)
647 RETURN(dom_stripesize);
649 /* DoM-only case exits here */
650 if (is_dom_only && dom_stripesize)
652 RETURN(dom_stripesize);
656 * Pack size attributes into the reply.
658 int mdt_pack_size2body(struct mdt_thread_info *info,
659 const struct lu_fid *fid, struct lustre_handle *lh)
662 struct md_attr *ma = &info->mti_attr;
664 bool dom_lock = false;
668 LASSERT(ma->ma_attr.la_valid & LA_MODE);
670 if (!S_ISREG(ma->ma_attr.la_mode) ||
671 !(ma->ma_valid & MA_LOV && ma->ma_lmm != NULL))
674 dom_stripe = mdt_lmm_dom_stripesize(ma->ma_lmm);
675 /* no DoM stripe, no size in reply */
679 if (lustre_handle_is_used(lh)) {
680 struct ldlm_lock *lock;
682 lock = ldlm_handle2lock(lh);
684 dom_lock = ldlm_has_dom(lock);
689 /* no DoM lock, no size in reply */
693 /* Either DoM lock exists or LMM has only DoM stripe then
694 * return size on body. */
695 b = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
697 mdt_dom_object_size(info->mti_env, info->mti_mdt, fid, b, dom_lock);
701 #ifdef CONFIG_LUSTRE_FS_POSIX_ACL
703 * Pack ACL data into the reply. UIDs/GIDs are mapped and filtered by nodemap.
705 * \param info thread info object
706 * \param repbody reply to pack ACLs into
707 * \param o mdt object of file to examine
708 * \param nodemap nodemap of client to reply to
710 * \retval -errno error getting or parsing ACL from disk
712 int mdt_pack_acl2body(struct mdt_thread_info *info, struct mdt_body *repbody,
713 struct mdt_object *o, struct lu_nodemap *nodemap)
715 const struct lu_env *env = info->mti_env;
716 struct md_object *next = mdt_object_child(o);
717 struct lu_buf *buf = &info->mti_buf;
718 struct mdt_device *mdt = info->mti_mdt;
719 struct req_capsule *pill = info->mti_pill;
724 buf->lb_buf = req_capsule_server_get(pill, &RMF_ACL);
725 buf->lb_len = req_capsule_get_size(pill, &RMF_ACL, RCL_SERVER);
726 if (buf->lb_len == 0)
729 LASSERT(!info->mti_big_acl_used);
731 rc = mo_xattr_get(env, next, buf, XATTR_NAME_ACL_ACCESS);
733 if (rc == -ENODATA) {
734 repbody->mbo_aclsize = 0;
735 repbody->mbo_valid |= OBD_MD_FLACL;
737 } else if (rc == -EOPNOTSUPP) {
739 } else if (rc == -ERANGE) {
740 if (exp_connect_large_acl(info->mti_exp) &&
741 !info->mti_big_acl_used) {
742 if (info->mti_big_acl == NULL) {
743 info->mti_big_aclsize =
745 mdt->mdt_max_ea_size,
747 OBD_ALLOC_LARGE(info->mti_big_acl,
748 info->mti_big_aclsize);
749 if (info->mti_big_acl == NULL) {
750 info->mti_big_aclsize = 0;
751 CERROR("%s: unable to grow "
754 PFID(mdt_object_fid(o)));
759 CDEBUG(D_INODE, "%s: grow the "DFID
760 " ACL buffer to size %d\n",
762 PFID(mdt_object_fid(o)),
763 info->mti_big_aclsize);
765 buf->lb_buf = info->mti_big_acl;
766 buf->lb_len = info->mti_big_aclsize;
767 info->mti_big_acl_used = 1;
770 /* FS has ACL bigger that our limits */
771 CDEBUG(D_INODE, "%s: "DFID" ACL can't fit into %d\n",
772 mdt_obd_name(mdt), PFID(mdt_object_fid(o)),
773 info->mti_big_aclsize);
776 CERROR("%s: unable to read "DFID" ACL: rc = %d\n",
777 mdt_obd_name(mdt), PFID(mdt_object_fid(o)), rc);
780 rc = nodemap_map_acl(nodemap, buf->lb_buf,
781 rc, NODEMAP_FS_TO_CLIENT);
782 /* if all ACLs mapped out, rc is still >= 0 */
784 CERROR("%s: nodemap_map_acl unable to parse "DFID
785 " ACL: rc = %d\n", mdt_obd_name(mdt),
786 PFID(mdt_object_fid(o)), rc);
787 repbody->mbo_aclsize = 0;
788 repbody->mbo_valid &= ~OBD_MD_FLACL;
790 repbody->mbo_aclsize = rc;
791 repbody->mbo_valid |= OBD_MD_FLACL;
800 /* XXX Look into layout in MDT layer. */
801 static inline bool mdt_hsm_is_released(struct lov_mds_md *lmm)
803 struct lov_comp_md_v1 *comp_v1;
804 struct lov_mds_md *v1;
807 if (lmm->lmm_magic == LOV_MAGIC_COMP_V1) {
808 comp_v1 = (struct lov_comp_md_v1 *)lmm;
810 for (i = 0; i < comp_v1->lcm_entry_count; i++) {
811 v1 = (struct lov_mds_md *)((char *)comp_v1 +
812 comp_v1->lcm_entries[i].lcme_offset);
813 /* We don't support partial release for now */
814 if (!(v1->lmm_pattern & LOV_PATTERN_F_RELEASED))
819 return (lmm->lmm_pattern & LOV_PATTERN_F_RELEASED) ?
824 void mdt_pack_attr2body(struct mdt_thread_info *info, struct mdt_body *b,
825 const struct lu_attr *attr, const struct lu_fid *fid)
827 struct mdt_device *mdt = info->mti_mdt;
828 struct obd_export *exp = info->mti_exp;
829 struct md_attr *ma = &info->mti_attr;
830 struct lu_nodemap *nodemap = NULL;
832 LASSERT(ma->ma_valid & MA_INODE);
834 if (attr->la_valid & LA_ATIME) {
835 b->mbo_atime = attr->la_atime;
836 b->mbo_valid |= OBD_MD_FLATIME;
838 if (attr->la_valid & LA_MTIME) {
839 b->mbo_mtime = attr->la_mtime;
840 b->mbo_valid |= OBD_MD_FLMTIME;
842 if (attr->la_valid & LA_CTIME) {
843 b->mbo_ctime = attr->la_ctime;
844 b->mbo_valid |= OBD_MD_FLCTIME;
846 if (attr->la_valid & LA_BTIME) {
847 b->mbo_btime = attr->la_btime;
848 b->mbo_valid |= OBD_MD_FLBTIME;
850 if (attr->la_valid & LA_FLAGS) {
851 b->mbo_flags = attr->la_flags;
852 b->mbo_valid |= OBD_MD_FLFLAGS;
854 if (attr->la_valid & LA_NLINK) {
855 b->mbo_nlink = attr->la_nlink;
856 b->mbo_valid |= OBD_MD_FLNLINK;
858 if (attr->la_valid & (LA_UID|LA_GID|LA_PROJID)) {
859 nodemap = nodemap_get_from_exp(exp);
863 if (attr->la_valid & LA_UID) {
864 b->mbo_uid = nodemap_map_id(nodemap, NODEMAP_UID,
865 NODEMAP_FS_TO_CLIENT,
867 b->mbo_valid |= OBD_MD_FLUID;
869 if (attr->la_valid & LA_GID) {
870 b->mbo_gid = nodemap_map_id(nodemap, NODEMAP_GID,
871 NODEMAP_FS_TO_CLIENT,
873 b->mbo_valid |= OBD_MD_FLGID;
876 if (attr->la_valid & LA_PROJID) {
877 b->mbo_projid = nodemap_map_id(nodemap, NODEMAP_PROJID,
878 NODEMAP_FS_TO_CLIENT,
880 b->mbo_valid |= OBD_MD_FLPROJID;
883 b->mbo_mode = attr->la_mode;
884 if (attr->la_valid & LA_MODE)
885 b->mbo_valid |= OBD_MD_FLMODE;
886 if (attr->la_valid & LA_TYPE)
887 b->mbo_valid |= OBD_MD_FLTYPE;
891 b->mbo_valid |= OBD_MD_FLID;
892 CDEBUG(D_INODE, DFID": nlink=%d, mode=%o, valid=%#llx\n",
893 PFID(fid), b->mbo_nlink, b->mbo_mode, b->mbo_valid);
896 if (!(attr->la_valid & LA_TYPE))
899 b->mbo_rdev = attr->la_rdev;
900 b->mbo_size = attr->la_size;
901 b->mbo_blocks = attr->la_blocks;
903 if (!S_ISREG(attr->la_mode)) {
904 b->mbo_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS | OBD_MD_FLRDEV;
905 } else if (ma->ma_need & MA_LOV && !(ma->ma_valid & MA_LOV)) {
906 /* means no objects are allocated on osts. */
907 LASSERT(!(ma->ma_valid & MA_LOV));
908 /* just ignore blocks occupied by extend attributes on MDS */
910 /* if no object is allocated on osts, the size on mds is valid.
912 b->mbo_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
913 } else if ((ma->ma_valid & MA_LOV) && ma->ma_lmm != NULL) {
914 if (mdt_hsm_is_released(ma->ma_lmm)) {
915 /* A released file stores its size on MDS. */
916 /* But return 1 block for released file, unless tools
917 * like tar will consider it fully sparse. (LU-3864)
919 if (unlikely(b->mbo_size == 0))
923 b->mbo_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
924 } else if (info->mti_som_strict && mdt->mdt_enable_strict_som) {
925 /* use SOM for size*/
926 b->mbo_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
927 } else if (ma->ma_valid & MA_SOM) { /* lsom is valid */
928 b->mbo_valid |= OBD_MD_FLLAZYSIZE | OBD_MD_FLLAZYBLOCKS;
929 b->mbo_size = ma->ma_som.ms_size;
930 b->mbo_blocks = ma->ma_som.ms_blocks;
934 if (fid != NULL && (b->mbo_valid & OBD_MD_FLSIZE ||
935 b->mbo_valid & OBD_MD_FLLAZYSIZE))
936 CDEBUG(D_VFSTRACE, DFID": returning size %llu\n",
937 PFID(fid), (unsigned long long)b->mbo_size);
940 if (!IS_ERR_OR_NULL(nodemap))
941 nodemap_putref(nodemap);
944 static inline int mdt_body_has_lov(const struct lu_attr *la,
945 const struct mdt_body *body)
947 return (S_ISREG(la->la_mode) && (body->mbo_valid & OBD_MD_FLEASIZE)) ||
948 (S_ISDIR(la->la_mode) && (body->mbo_valid & OBD_MD_FLDIREA));
951 void mdt_client_compatibility(struct mdt_thread_info *info)
953 struct mdt_body *body;
954 struct ptlrpc_request *req = mdt_info_req(info);
955 struct obd_export *exp = req->rq_export;
956 struct md_attr *ma = &info->mti_attr;
957 struct lu_attr *la = &ma->ma_attr;
960 if (exp_connect_layout(exp))
961 /* the client can deal with 16-bit lmm_stripe_count */
964 body = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
966 if (!mdt_body_has_lov(la, body))
969 /* now we have a reply with a lov for a client not compatible with the
970 * layout lock so we have to clean the layout generation number */
971 if (S_ISREG(la->la_mode))
972 ma->ma_lmm->lmm_layout_gen = 0;
976 static int mdt_attr_get_eabuf_size(struct mdt_thread_info *info,
977 struct mdt_object *o)
979 const struct lu_env *env = info->mti_env;
982 rc = mo_xattr_get(env, mdt_object_child(o), &LU_BUF_NULL,
991 /* Is it a directory? Let's check for the LMV as well */
992 if (S_ISDIR(lu_object_attr(&mdt_object_child(o)->mo_lu))) {
993 rc2 = mo_xattr_get(env, mdt_object_child(o), &LU_BUF_NULL,
997 rc2 = mo_xattr_get(env, mdt_object_child(o),
999 XATTR_NAME_DEFAULT_LMV);
1001 if ((rc2 < 0 && rc2 != -ENODATA) || (rc2 > rc))
1009 int mdt_big_xattr_get(struct mdt_thread_info *info, struct mdt_object *o,
1012 const struct lu_env *env = info->mti_env;
1016 LASSERT(info->mti_big_lmm_used == 0);
1017 rc = mo_xattr_get(env, mdt_object_child(o), &LU_BUF_NULL, name);
1021 /* big_lmm may need to be grown */
1022 if (info->mti_big_lmmsize < rc) {
1023 int size = size_roundup_power2(rc);
1025 if (info->mti_big_lmmsize > 0) {
1026 /* free old buffer */
1027 LASSERT(info->mti_big_lmm);
1028 OBD_FREE_LARGE(info->mti_big_lmm,
1029 info->mti_big_lmmsize);
1030 info->mti_big_lmm = NULL;
1031 info->mti_big_lmmsize = 0;
1034 OBD_ALLOC_LARGE(info->mti_big_lmm, size);
1035 if (info->mti_big_lmm == NULL)
1037 info->mti_big_lmmsize = size;
1039 LASSERT(info->mti_big_lmmsize >= rc);
1041 info->mti_buf.lb_buf = info->mti_big_lmm;
1042 info->mti_buf.lb_len = info->mti_big_lmmsize;
1043 rc = mo_xattr_get(env, mdt_object_child(o), &info->mti_buf, name);
1048 int __mdt_stripe_get(struct mdt_thread_info *info, struct mdt_object *o,
1049 struct md_attr *ma, const char *name)
1051 struct md_object *next = mdt_object_child(o);
1052 struct lu_buf *buf = &info->mti_buf;
1055 if (strcmp(name, XATTR_NAME_LOV) == 0) {
1056 buf->lb_buf = ma->ma_lmm;
1057 buf->lb_len = ma->ma_lmm_size;
1058 LASSERT(!(ma->ma_valid & MA_LOV));
1059 } else if (strcmp(name, XATTR_NAME_LMV) == 0) {
1060 buf->lb_buf = ma->ma_lmv;
1061 buf->lb_len = ma->ma_lmv_size;
1062 LASSERT(!(ma->ma_valid & MA_LMV));
1063 } else if (strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0) {
1064 buf->lb_buf = ma->ma_default_lmv;
1065 buf->lb_len = ma->ma_default_lmv_size;
1066 LASSERT(!(ma->ma_valid & MA_LMV_DEF));
1071 LASSERT(buf->lb_buf);
1073 if (!mdt_object_exists(o))
1076 if (mdt_object_remote(o) && S_ISDIR(lu_object_attr(&o->mot_obj)))
1077 /* force reload layout for remote dir in case layout changed */
1078 mo_invalidate(info->mti_env, mdt_object_child(o));
1080 rc = mo_xattr_get(info->mti_env, next, buf, name);
1084 if (strcmp(name, XATTR_NAME_LOV) == 0) {
1085 if (info->mti_big_lmm_used)
1086 ma->ma_lmm = info->mti_big_lmm;
1088 /* NOT return LOV EA with hole to old client. */
1089 if (unlikely(le32_to_cpu(ma->ma_lmm->lmm_pattern) &
1090 LOV_PATTERN_F_HOLE) &&
1091 !(exp_connect_flags(info->mti_exp) &
1092 OBD_CONNECT_LFSCK)) {
1095 ma->ma_lmm_size = rc;
1096 ma->ma_valid |= MA_LOV;
1098 } else if (strcmp(name, XATTR_NAME_LMV) == 0) {
1099 if (info->mti_big_lmm_used)
1100 ma->ma_lmv = info->mti_big_lmm;
1102 ma->ma_lmv_size = rc;
1103 ma->ma_valid |= MA_LMV;
1104 } else if (strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0) {
1105 ma->ma_default_lmv_size = rc;
1106 ma->ma_valid |= MA_LMV_DEF;
1109 /* Update mdt_max_mdsize so all clients will be aware that */
1110 if (info->mti_mdt->mdt_max_mdsize < rc)
1111 info->mti_mdt->mdt_max_mdsize = rc;
1114 } else if (rc == -ENODATA) {
1117 } else if (rc == -ERANGE) {
1118 /* Default LMV has fixed size, so it must be able to fit
1119 * in the original buffer */
1120 if (strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0)
1122 rc = mdt_big_xattr_get(info, o, name);
1124 info->mti_big_lmm_used = 1;
1132 int mdt_stripe_get(struct mdt_thread_info *info, struct mdt_object *o,
1133 struct md_attr *ma, const char *name)
1137 if (!info->mti_big_lmm) {
1138 OBD_ALLOC(info->mti_big_lmm, PAGE_SIZE);
1139 if (!info->mti_big_lmm)
1141 info->mti_big_lmmsize = PAGE_SIZE;
1144 if (strcmp(name, XATTR_NAME_LOV) == 0) {
1145 ma->ma_lmm = info->mti_big_lmm;
1146 ma->ma_lmm_size = info->mti_big_lmmsize;
1147 ma->ma_valid &= ~MA_LOV;
1148 } else if (strcmp(name, XATTR_NAME_LMV) == 0) {
1149 ma->ma_lmv = info->mti_big_lmm;
1150 ma->ma_lmv_size = info->mti_big_lmmsize;
1151 ma->ma_valid &= ~MA_LMV;
1156 LASSERT(!info->mti_big_lmm_used);
1157 rc = __mdt_stripe_get(info, o, ma, name);
1158 /* since big_lmm is always used here, clear 'used' flag to avoid
1159 * assertion in mdt_big_xattr_get().
1161 info->mti_big_lmm_used = 0;
1166 int mdt_attr_get_pfid(struct mdt_thread_info *info, struct mdt_object *o,
1167 struct lu_fid *pfid)
1169 struct lu_buf *buf = &info->mti_buf;
1170 struct link_ea_header *leh;
1171 struct link_ea_entry *lee;
1175 buf->lb_buf = info->mti_big_lmm;
1176 buf->lb_len = info->mti_big_lmmsize;
1177 rc = mo_xattr_get(info->mti_env, mdt_object_child(o),
1178 buf, XATTR_NAME_LINK);
1179 /* ignore errors, MA_PFID won't be set and it is
1180 * up to the caller to treat this as an error */
1181 if (rc == -ERANGE || buf->lb_len == 0) {
1182 rc = mdt_big_xattr_get(info, o, XATTR_NAME_LINK);
1183 buf->lb_buf = info->mti_big_lmm;
1184 buf->lb_len = info->mti_big_lmmsize;
1189 if (rc < sizeof(*leh)) {
1190 CERROR("short LinkEA on "DFID": rc = %d\n",
1191 PFID(mdt_object_fid(o)), rc);
1195 leh = (struct link_ea_header *) buf->lb_buf;
1196 lee = (struct link_ea_entry *)(leh + 1);
1197 if (leh->leh_magic == __swab32(LINK_EA_MAGIC)) {
1198 leh->leh_magic = LINK_EA_MAGIC;
1199 leh->leh_reccount = __swab32(leh->leh_reccount);
1200 leh->leh_len = __swab64(leh->leh_len);
1202 if (leh->leh_magic != LINK_EA_MAGIC)
1204 if (leh->leh_reccount == 0)
1207 memcpy(pfid, &lee->lee_parent_fid, sizeof(*pfid));
1208 fid_be_to_cpu(pfid, pfid);
1213 int mdt_attr_get_pfid_name(struct mdt_thread_info *info, struct mdt_object *o,
1214 struct lu_fid *pfid, struct lu_name *lname)
1216 struct lu_buf *buf = &info->mti_buf;
1217 struct link_ea_header *leh;
1218 struct link_ea_entry *lee;
1222 buf->lb_buf = info->mti_xattr_buf;
1223 buf->lb_len = sizeof(info->mti_xattr_buf);
1224 rc = mo_xattr_get(info->mti_env, mdt_object_child(o), buf,
1226 if (rc == -ERANGE) {
1227 rc = mdt_big_xattr_get(info, o, XATTR_NAME_LINK);
1228 buf->lb_buf = info->mti_big_lmm;
1229 buf->lb_len = info->mti_big_lmmsize;
1234 if (rc < sizeof(*leh)) {
1235 CERROR("short LinkEA on "DFID": rc = %d\n",
1236 PFID(mdt_object_fid(o)), rc);
1240 leh = (struct link_ea_header *)buf->lb_buf;
1241 lee = (struct link_ea_entry *)(leh + 1);
1242 if (leh->leh_magic == __swab32(LINK_EA_MAGIC)) {
1243 leh->leh_magic = LINK_EA_MAGIC;
1244 leh->leh_reccount = __swab32(leh->leh_reccount);
1245 leh->leh_len = __swab64(leh->leh_len);
1247 if (leh->leh_magic != LINK_EA_MAGIC)
1250 if (leh->leh_reccount == 0)
1253 linkea_entry_unpack(lee, &reclen, lname, pfid);
1258 int mdt_attr_get_complex(struct mdt_thread_info *info,
1259 struct mdt_object *o, struct md_attr *ma)
1261 const struct lu_env *env = info->mti_env;
1262 struct md_object *next = mdt_object_child(o);
1263 struct lu_buf *buf = &info->mti_buf;
1264 int need = ma->ma_need;
1271 if (mdt_object_exists(o) == 0)
1272 GOTO(out, rc = -ENOENT);
1273 mode = lu_object_attr(&next->mo_lu);
1275 if (need & MA_INODE) {
1276 ma->ma_need = MA_INODE;
1277 if (need & MA_DIRENT_CNT)
1278 ma->ma_attr.la_valid |= LA_DIRENT_CNT;
1280 ma->ma_attr.la_valid &= ~LA_DIRENT_CNT;
1281 rc = mo_attr_get(env, next, ma);
1286 (void) mdt_get_som(info, o, ma);
1287 ma->ma_valid |= MA_INODE;
1290 if (need & MA_PFID) {
1291 rc = mdt_attr_get_pfid(info, o, &ma->ma_pfid);
1293 ma->ma_valid |= MA_PFID;
1294 /* ignore this error, parent fid is not mandatory */
1298 if (need & MA_LOV && (S_ISREG(mode) || S_ISDIR(mode))) {
1299 rc = __mdt_stripe_get(info, o, ma, XATTR_NAME_LOV);
1304 if (need & MA_LMV && S_ISDIR(mode)) {
1305 rc = __mdt_stripe_get(info, o, ma, XATTR_NAME_LMV);
1310 if (need & MA_LMV_DEF && S_ISDIR(mode)) {
1311 rc = __mdt_stripe_get(info, o, ma, XATTR_NAME_DEFAULT_LMV);
1317 * In the handle of MA_INODE, we may already get the SOM attr.
1319 if (need & MA_SOM && S_ISREG(mode) && !(ma->ma_valid & MA_SOM)) {
1320 rc = mdt_get_som(info, o, ma);
1325 if (need & MA_HSM && S_ISREG(mode)) {
1326 buf->lb_buf = info->mti_xattr_buf;
1327 buf->lb_len = sizeof(info->mti_xattr_buf);
1328 BUILD_BUG_ON(sizeof(struct hsm_attrs) >
1329 sizeof(info->mti_xattr_buf));
1330 rc2 = mo_xattr_get(info->mti_env, next, buf, XATTR_NAME_HSM);
1331 rc2 = lustre_buf2hsm(info->mti_xattr_buf, rc2, &ma->ma_hsm);
1333 ma->ma_valid |= MA_HSM;
1334 else if (rc2 < 0 && rc2 != -ENODATA)
1335 GOTO(out, rc = rc2);
1338 #ifdef CONFIG_LUSTRE_FS_POSIX_ACL
1339 if (need & MA_ACL_DEF && S_ISDIR(mode)) {
1340 buf->lb_buf = ma->ma_acl;
1341 buf->lb_len = ma->ma_acl_size;
1342 rc2 = mo_xattr_get(env, next, buf, XATTR_NAME_ACL_DEFAULT);
1344 ma->ma_acl_size = rc2;
1345 ma->ma_valid |= MA_ACL_DEF;
1346 } else if (rc2 == -ENODATA) {
1348 ma->ma_acl_size = 0;
1350 GOTO(out, rc = rc2);
1355 CDEBUG(D_INODE, "after getattr rc = %d, ma_valid = %#llx ma_lmm=%p\n",
1356 rc, ma->ma_valid, ma->ma_lmm);
1360 static void mdt_preset_encctx_size(struct mdt_thread_info *info)
1362 struct req_capsule *pill = info->mti_pill;
1365 if (req_capsule_has_field(pill, &RMF_FILE_ENCCTX,
1367 /* pre-set size in server part with max size */
1368 req_capsule_set_size(pill, &RMF_FILE_ENCCTX,
1370 info->mti_mdt->mdt_max_mdsize);
1374 static int mdt_getattr_internal(struct mdt_thread_info *info,
1375 struct mdt_object *o, int ma_need)
1377 struct mdt_device *mdt = info->mti_mdt;
1378 struct md_object *next = mdt_object_child(o);
1379 const struct mdt_body *reqbody = info->mti_body;
1380 struct ptlrpc_request *req = mdt_info_req(info);
1381 struct md_attr *ma = &info->mti_attr;
1382 struct lu_attr *la = &ma->ma_attr;
1383 struct req_capsule *pill = info->mti_pill;
1384 const struct lu_env *env = info->mti_env;
1385 struct mdt_body *repbody;
1386 struct lu_buf *buffer = &info->mti_buf;
1387 struct obd_export *exp = info->mti_exp;
1388 ktime_t kstart = ktime_get();
1393 if (CFS_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK))
1394 RETURN(err_serious(-ENOMEM));
1396 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
1400 if (mdt_object_remote(o)) {
1401 /* This object is located on remote node.*/
1402 /* Return -ENOTSUPP for old client */
1403 if (!mdt_is_dne_client(req->rq_export))
1404 GOTO(out, rc = -ENOTSUPP);
1406 repbody->mbo_fid1 = *mdt_object_fid(o);
1407 repbody->mbo_valid = OBD_MD_FLID | OBD_MD_MDS;
1411 if (reqbody->mbo_eadatasize > 0) {
1412 buffer->lb_buf = req_capsule_server_get(pill, &RMF_MDT_MD);
1413 if (buffer->lb_buf == NULL)
1414 GOTO(out, rc = -EPROTO);
1415 buffer->lb_len = req_capsule_get_size(pill, &RMF_MDT_MD,
1418 buffer->lb_buf = NULL;
1420 ma_need &= ~(MA_LOV | MA_LMV);
1421 CDEBUG(D_INFO, "%s: RPC from %s: does not need LOVEA.\n",
1422 mdt_obd_name(info->mti_mdt),
1423 req->rq_export->exp_client_uuid.uuid);
1426 /* from 2.12.58 intent_getattr pack default LMV in reply */
1427 if (S_ISDIR(lu_object_attr(&next->mo_lu)) &&
1428 ((reqbody->mbo_valid & (OBD_MD_MEA | OBD_MD_DEFAULT_MEA)) ==
1429 (OBD_MD_MEA | OBD_MD_DEFAULT_MEA)) &&
1430 req_capsule_has_field(&req->rq_pill, &RMF_DEFAULT_MDT_MD,
1432 ma->ma_lmv = buffer->lb_buf;
1433 ma->ma_lmv_size = buffer->lb_len;
1434 ma->ma_default_lmv = req_capsule_server_get(pill,
1435 &RMF_DEFAULT_MDT_MD);
1436 ma->ma_default_lmv_size = req_capsule_get_size(pill,
1437 &RMF_DEFAULT_MDT_MD,
1439 ma->ma_need = MA_INODE;
1440 if (ma->ma_lmv_size > 0)
1441 ma->ma_need |= MA_LMV;
1442 if (ma->ma_default_lmv_size > 0)
1443 ma->ma_need |= MA_LMV_DEF;
1444 } else if (S_ISDIR(lu_object_attr(&next->mo_lu)) &&
1445 (reqbody->mbo_valid & (OBD_MD_MEA | OBD_MD_DEFAULT_MEA))) {
1446 /* If it is dir and client require MEA, then we got MEA */
1447 /* Assumption: MDT_MD size is enough for lmv size. */
1448 ma->ma_lmv = buffer->lb_buf;
1449 ma->ma_lmv_size = buffer->lb_len;
1450 ma->ma_need = MA_INODE;
1451 if (ma->ma_lmv_size > 0) {
1452 if (reqbody->mbo_valid & OBD_MD_MEA) {
1453 ma->ma_need |= MA_LMV;
1454 } else if (reqbody->mbo_valid & OBD_MD_DEFAULT_MEA) {
1455 ma->ma_need |= MA_LMV_DEF;
1456 ma->ma_default_lmv = buffer->lb_buf;
1458 ma->ma_default_lmv_size = buffer->lb_len;
1459 ma->ma_lmv_size = 0;
1463 ma->ma_lmm = buffer->lb_buf;
1464 ma->ma_lmm_size = buffer->lb_len;
1465 ma->ma_need = MA_INODE | MA_HSM;
1466 if (ma->ma_lmm_size > 0) {
1467 ma->ma_need |= MA_LOV;
1468 /* Older clients may crash if they getattr overstriped
1471 if (!exp_connect_overstriping(exp) &&
1472 mdt_lmm_is_overstriping(ma->ma_lmm))
1473 RETURN(-EOPNOTSUPP);
1477 if (S_ISDIR(lu_object_attr(&next->mo_lu)) &&
1478 reqbody->mbo_valid & OBD_MD_FLDIREA &&
1479 lustre_msg_get_opc(req->rq_reqmsg) == MDS_GETATTR) {
1480 /* get default stripe info for this dir. */
1481 ma->ma_need |= MA_LOV_DEF;
1483 ma->ma_need |= ma_need;
1485 rc = mdt_attr_get_complex(info, o, ma);
1487 CDEBUG_LIMIT(rc == -ENOENT ? D_OTHER : D_ERROR,
1488 "%s: getattr error for "DFID": rc = %d\n",
1489 mdt_obd_name(info->mti_mdt),
1490 PFID(mdt_object_fid(o)), rc);
1494 /* return immutable attr on fscrypt metadata files
1495 * if fscrypt admin is not permitted
1497 if (o->mot_obj.lo_header->loh_attr & LOHA_FSCRYPT_MD &&
1498 !mdt_ucred(info)->uc_rbac_fscrypt_admin)
1499 la->la_flags |= LUSTRE_IMMUTABLE_FL;
1501 /* if file is released, check if a restore is running */
1502 if (ma->ma_valid & MA_HSM) {
1503 repbody->mbo_valid |= OBD_MD_TSTATE;
1504 if ((ma->ma_hsm.mh_flags & HS_RELEASED) &&
1505 mdt_hsm_restore_is_running(info, mdt_object_fid(o)))
1506 repbody->mbo_t_state = MS_RESTORE;
1509 if (unlikely(!(ma->ma_valid & MA_INODE)))
1512 mdt_pack_attr2body(info, repbody, la, mdt_object_fid(o));
1514 if (mdt_body_has_lov(la, reqbody)) {
1515 u32 stripe_count = 1;
1516 bool fixed_layout = false;
1518 if (ma->ma_valid & MA_LOV) {
1519 LASSERT(ma->ma_lmm_size);
1520 repbody->mbo_eadatasize = ma->ma_lmm_size;
1521 if (S_ISDIR(la->la_mode))
1522 repbody->mbo_valid |= OBD_MD_FLDIREA;
1524 repbody->mbo_valid |= OBD_MD_FLEASIZE;
1525 mdt_dump_lmm(D_INFO, ma->ma_lmm, repbody->mbo_valid);
1527 if (ma->ma_valid & MA_LMV) {
1528 struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1529 u32 magic = le32_to_cpu(lmv->lmv_magic);
1531 /* Return -ENOTSUPP for old client */
1532 if (!mdt_is_striped_client(req->rq_export))
1535 LASSERT(S_ISDIR(la->la_mode));
1536 mdt_dump_lmv(D_INFO, ma->ma_lmv);
1537 repbody->mbo_eadatasize = ma->ma_lmv_size;
1538 repbody->mbo_valid |= (OBD_MD_FLDIREA|OBD_MD_MEA);
1540 stripe_count = le32_to_cpu(lmv->lmv_stripe_count);
1541 fixed_layout = lmv_is_fixed(lmv);
1542 if (magic == LMV_MAGIC_STRIPE && lmv_is_restriping(lmv))
1543 mdt_restripe_migrate_add(info, o);
1544 else if (magic == LMV_MAGIC_V1 &&
1545 lmv_is_restriping(lmv))
1546 mdt_restripe_update_add(info, o);
1548 if (ma->ma_valid & MA_LMV_DEF) {
1549 /* Return -ENOTSUPP for old client */
1550 if (!mdt_is_striped_client(req->rq_export))
1552 LASSERT(S_ISDIR(la->la_mode));
1554 * when ll_dir_getstripe() gets default LMV, it
1555 * checks mbo_eadatasize.
1557 if (!(ma->ma_valid & MA_LMV))
1558 repbody->mbo_eadatasize =
1559 ma->ma_default_lmv_size;
1560 repbody->mbo_valid |= (OBD_MD_FLDIREA |
1561 OBD_MD_DEFAULT_MEA);
1564 "dirent count %llu stripe count %u MDT count %d\n",
1565 ma->ma_attr.la_dirent_count, stripe_count,
1566 atomic_read(&mdt->mdt_mds_mds_conns) + 1);
1567 if (ma->ma_attr.la_dirent_count != LU_DIRENT_COUNT_UNSET &&
1568 ma->ma_attr.la_dirent_count >
1569 mdt->mdt_restriper.mdr_dir_split_count &&
1570 !fid_is_root(mdt_object_fid(o)) &&
1571 mdt->mdt_enable_dir_auto_split &&
1572 !o->mot_restriping &&
1573 stripe_count < atomic_read(&mdt->mdt_mds_mds_conns) + 1 &&
1575 mdt_auto_split_add(info, o);
1576 } else if (S_ISLNK(la->la_mode) &&
1577 reqbody->mbo_valid & OBD_MD_LINKNAME) {
1578 buffer->lb_buf = ma->ma_lmm;
1579 /* eadatasize from client includes NULL-terminator, so
1580 * there is no need to read it */
1581 buffer->lb_len = reqbody->mbo_eadatasize - 1;
1582 rc = mo_readlink(env, next, buffer);
1583 if (unlikely(rc <= 0)) {
1584 CERROR("%s: readlink failed for "DFID": rc = %d\n",
1585 mdt_obd_name(info->mti_mdt),
1586 PFID(mdt_object_fid(o)), rc);
1589 int print_limit = min_t(int, PAGE_SIZE - 128, rc);
1591 if (CFS_FAIL_CHECK(OBD_FAIL_MDS_READLINK_EPROTO))
1593 repbody->mbo_valid |= OBD_MD_LINKNAME;
1594 /* we need to report back size with NULL-terminator
1595 * because client expects that */
1596 repbody->mbo_eadatasize = rc + 1;
1597 if (repbody->mbo_eadatasize != reqbody->mbo_eadatasize)
1598 CDEBUG(D_INODE, "%s: Read shorter symlink %d "
1599 "on "DFID ", expected %d\n",
1600 mdt_obd_name(info->mti_mdt),
1601 rc, PFID(mdt_object_fid(o)),
1602 reqbody->mbo_eadatasize - 1);
1603 /* NULL terminate */
1604 ((char *)ma->ma_lmm)[rc] = 0;
1606 /* If the total CDEBUG() size is larger than a page, it
1607 * will print a warning to the console, avoid this by
1608 * printing just the last part of the symlink. */
1609 CDEBUG(D_INODE, "symlink dest %s%.*s, len = %d\n",
1610 print_limit < rc ? "..." : "", print_limit,
1611 (char *)ma->ma_lmm + rc - print_limit, rc);
1616 if (reqbody->mbo_valid & OBD_MD_FLMODEASIZE) {
1617 repbody->mbo_max_mdsize = info->mti_mdt->mdt_max_mdsize;
1618 repbody->mbo_valid |= OBD_MD_FLMODEASIZE;
1619 CDEBUG(D_INODE, "changing the max MD size to %u\n",
1620 repbody->mbo_max_mdsize);
1623 #ifdef CONFIG_LUSTRE_FS_POSIX_ACL
1624 if ((exp_connect_flags(req->rq_export) & OBD_CONNECT_ACL) &&
1625 (reqbody->mbo_valid & OBD_MD_FLACL)) {
1626 struct lu_nodemap *nodemap = nodemap_get_from_exp(exp);
1627 if (IS_ERR(nodemap))
1628 RETURN(PTR_ERR(nodemap));
1630 rc = mdt_pack_acl2body(info, repbody, o, nodemap);
1631 nodemap_putref(nodemap);
1637 mdt_counter_incr(req, LPROC_MDT_GETATTR,
1638 ktime_us_delta(ktime_get(), kstart));
1643 static int mdt_getattr(struct tgt_session_info *tsi)
1645 struct mdt_thread_info *info = tsi2mdt_info(tsi);
1646 struct mdt_object *obj = info->mti_object;
1647 struct req_capsule *pill = info->mti_pill;
1648 struct mdt_body *reqbody;
1649 struct mdt_body *repbody;
1653 if (unlikely(info->mti_object == NULL))
1656 reqbody = req_capsule_client_get(pill, &RMF_MDT_BODY);
1658 LASSERT(lu_object_assert_exists(&obj->mot_obj));
1660 /* Special case for Data-on-MDT files to get data version */
1661 if (unlikely(reqbody->mbo_valid & OBD_MD_FLDATAVERSION)) {
1662 rc = mdt_data_version_get(tsi);
1666 /* Unlike intent case where we need to pre-fill out buffers early on
1667 * in intent policy for ldlm reasons, here we can have a much better
1668 * guess at EA size by just reading it from disk.
1669 * Exceptions are readdir and (missing) directory striping */
1671 if (reqbody->mbo_valid & OBD_MD_LINKNAME) {
1672 /* No easy way to know how long is the symlink, but it cannot
1673 * be more than PATH_MAX, so we allocate +1 */
1675 /* A special case for fs ROOT: getattr there might fetch
1676 * default EA for entire fs, not just for this dir!
1678 } else if (lu_fid_eq(mdt_object_fid(obj),
1679 &info->mti_mdt->mdt_md_root_fid) &&
1680 (reqbody->mbo_valid & OBD_MD_FLDIREA) &&
1681 (lustre_msg_get_opc(mdt_info_req(info)->rq_reqmsg) ==
1683 /* Should the default strping be bigger, mdt_fix_reply
1684 * will reallocate */
1685 rc = DEF_REP_MD_SIZE;
1687 /* Read the actual EA size from disk */
1688 rc = mdt_attr_get_eabuf_size(info, obj);
1692 GOTO(out, rc = err_serious(rc));
1694 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER, rc);
1696 /* Set ACL reply buffer size as LUSTRE_POSIX_ACL_MAX_SIZE_OLD
1697 * by default. If the target object has more ACL entries, then
1698 * enlarge the buffer when necessary. */
1699 req_capsule_set_size(pill, &RMF_ACL, RCL_SERVER,
1700 LUSTRE_POSIX_ACL_MAX_SIZE_OLD);
1701 mdt_preset_encctx_size(info);
1703 rc = req_capsule_server_pack(pill);
1704 if (unlikely(rc != 0))
1705 GOTO(out, rc = err_serious(rc));
1707 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
1708 LASSERT(repbody != NULL);
1709 repbody->mbo_eadatasize = 0;
1710 repbody->mbo_aclsize = 0;
1712 rc = mdt_check_ucred(info);
1714 GOTO(out_shrink, rc);
1716 info->mti_cross_ref = !!(reqbody->mbo_valid & OBD_MD_FLCROSSREF);
1718 rc = mdt_init_ucred(info, reqbody);
1720 GOTO(out_shrink, rc);
1722 rc = mdt_getattr_internal(info, obj, 0);
1724 GOTO(out_ucred, rc);
1726 rc = mdt_pack_encctx_in_reply(info, obj);
1729 mdt_exit_ucred(info);
1731 mdt_client_compatibility(info);
1732 rc2 = mdt_fix_reply(info);
1736 mdt_thread_info_fini(info);
1741 * Handler of layout intent RPC requiring the layout modification
1743 * \param[in] info thread environment
1744 * \param[in] obj object
1745 * \param[out] lhc object ldlm lock handle
1746 * \param[in] layout layout change descriptor
1748 * \retval 0 on success
1749 * \retval < 0 error code
1751 int mdt_layout_change(struct mdt_thread_info *info, struct mdt_object *obj,
1752 struct mdt_lock_handle *lhc,
1753 struct md_layout_change *layout)
1759 if (!mdt_object_exists(obj))
1762 if (!S_ISREG(lu_object_attr(&obj->mot_obj)))
1765 rc = mo_permission(info->mti_env, NULL, mdt_object_child(obj), NULL,
1770 rc = mdt_check_resent_lock(info, obj, lhc);
1776 __u64 lockpart = MDS_INODELOCK_LAYOUT;
1778 /* take layout lock to prepare layout change */
1779 if (layout->mlc_opc == MD_LAYOUT_WRITE)
1780 lockpart |= MDS_INODELOCK_UPDATE;
1782 rc = mdt_object_lock(info, obj, lhc, lockpart, LCK_EX);
1787 mutex_lock(&obj->mot_som_mutex);
1788 rc = mo_layout_change(info->mti_env, mdt_object_child(obj), layout);
1789 mutex_unlock(&obj->mot_som_mutex);
1792 mdt_object_unlock(info, obj, lhc, 1);
1798 * Exchange MOF_LOV_CREATED flags between two objects after a
1799 * layout swap. No assumption is made on whether o1 or o2 have
1800 * created objects or not.
1802 * \param[in,out] o1 First swap layout object
1803 * \param[in,out] o2 Second swap layout object
1805 static void mdt_swap_lov_flag(struct mdt_object *o1, struct mdt_object *o2)
1807 unsigned int o1_lov_created = o1->mot_lov_created;
1809 mutex_lock(&o1->mot_lov_mutex);
1810 mutex_lock(&o2->mot_lov_mutex);
1812 o1->mot_lov_created = o2->mot_lov_created;
1813 o2->mot_lov_created = o1_lov_created;
1815 mutex_unlock(&o2->mot_lov_mutex);
1816 mutex_unlock(&o1->mot_lov_mutex);
1819 static int mdt_swap_layouts(struct tgt_session_info *tsi)
1821 struct mdt_thread_info *info;
1822 struct ptlrpc_request *req = tgt_ses_req(tsi);
1823 struct obd_export *exp = req->rq_export;
1824 struct mdt_object *o1, *o2, *o;
1825 struct mdt_lock_handle *lh1, *lh2;
1826 struct mdc_swap_layouts *msl;
1830 /* client does not support layout lock, so layout swaping
1832 * FIXME: there is a problem for old clients which don't support
1833 * layout lock yet. If those clients have already opened the file
1834 * they won't be notified at all so that old layout may still be
1835 * used to do IO. This can be fixed after file release is landed by
1836 * doing exclusive open and taking full EX ibits lock. - Jinshan */
1837 if (!exp_connect_layout(exp))
1838 RETURN(-EOPNOTSUPP);
1840 info = tsi2mdt_info(tsi);
1841 if (unlikely(info->mti_object == NULL))
1844 if (info->mti_dlm_req != NULL)
1845 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1847 o1 = info->mti_object;
1848 o = o2 = mdt_object_find(info->mti_env, info->mti_mdt,
1849 &info->mti_body->mbo_fid2);
1851 GOTO(out, rc = PTR_ERR(o));
1853 if (mdt_object_remote(o) || !mdt_object_exists(o)) /* remote object */
1854 GOTO(put, rc = -ENOENT);
1856 rc = lu_fid_cmp(&info->mti_body->mbo_fid1, &info->mti_body->mbo_fid2);
1857 if (unlikely(rc == 0)) /* same file, you kidding me? no-op. */
1863 /* permission check. Make sure the calling process having permission
1864 * to write both files. */
1865 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o1), NULL,
1870 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o2), NULL,
1875 msl = req_capsule_client_get(info->mti_pill, &RMF_SWAP_LAYOUTS);
1877 GOTO(put, rc = -EPROTO);
1879 lh1 = &info->mti_lh[MDT_LH_NEW];
1880 lh2 = &info->mti_lh[MDT_LH_OLD];
1881 rc = mdt_object_lock(info, o1, lh1, MDS_INODELOCK_LAYOUT |
1882 MDS_INODELOCK_XATTR, LCK_EX);
1886 rc = mdt_object_lock(info, o2, lh2, MDS_INODELOCK_LAYOUT |
1887 MDS_INODELOCK_XATTR, LCK_EX);
1891 rc = mo_swap_layouts(info->mti_env, mdt_object_child(o1),
1892 mdt_object_child(o2), msl->msl_flags);
1896 mdt_swap_lov_flag(o1, o2);
1899 mdt_object_unlock(info, o2, lh2, rc);
1901 mdt_object_unlock(info, o1, lh1, rc);
1903 mdt_object_put(info->mti_env, o);
1905 mdt_thread_info_fini(info);
1909 static int mdt_raw_lookup(struct mdt_thread_info *info,
1910 struct mdt_object *parent,
1911 const struct lu_name *lname)
1913 struct lu_fid *fid = &info->mti_tmp_fid1;
1914 struct mdt_body *repbody;
1915 bool is_dotdot = false;
1916 bool is_old_parent_stripe = false;
1917 bool is_new_parent_checked = false;
1922 LASSERT(!info->mti_cross_ref);
1923 /* Always allow to lookup ".." */
1924 if (lname->ln_namelen == 2 &&
1925 lname->ln_name[0] == '.' && lname->ln_name[1] == '.') {
1926 info->mti_spec.sp_permitted = 1;
1928 if (mdt_is_dir_stripe(info, parent) == 1)
1929 is_old_parent_stripe = true;
1932 mdt_object_get(info->mti_env, parent);
1934 /* Only got the fid of this obj by name */
1936 rc = mdo_lookup(info->mti_env, mdt_object_child(parent), lname, fid,
1938 mdt_object_put(info->mti_env, parent);
1942 /* getattr_name("..") should return master object FID for striped dir */
1943 if (is_dotdot && (is_old_parent_stripe || !is_new_parent_checked)) {
1944 parent = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1946 RETURN(PTR_ERR(parent));
1948 /* old client getattr_name("..") with stripe FID */
1949 if (unlikely(is_old_parent_stripe)) {
1950 is_old_parent_stripe = false;
1954 /* ".." may be a stripe */
1955 if (unlikely(mdt_is_dir_stripe(info, parent) == 1)) {
1956 is_new_parent_checked = true;
1960 mdt_object_put(info->mti_env, parent);
1963 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1964 repbody->mbo_fid1 = *fid;
1965 repbody->mbo_valid = OBD_MD_FLID;
1971 * Find name matching hash
1973 * We search \a child LinkEA for a name whose hash matches \a lname
1974 * (it contains an encoded hash).
1976 * \param info mdt thread info
1977 * \param lname encoded hash to find
1978 * \param parent parent object
1979 * \param child object to search with LinkEA
1981 * \retval 1 match found
1982 * \retval 0 no match found
1983 * \retval -ev negative errno upon error
1985 int find_name_matching_hash(struct mdt_thread_info *info, struct lu_name *lname,
1986 struct mdt_object *parent, struct mdt_object *child)
1988 /* Here, lname is an encoded hash of on-disk name, and
1989 * client is doing access without encryption key.
1990 * So we need to get LinkEA, check parent fid is correct and
1991 * compare name hash with the one in the request.
1993 struct lu_buf *buf = &info->mti_big_buf;
1994 struct lu_name name;
1996 struct linkea_data ldata = { NULL };
1997 struct link_ea_header *leh;
1998 struct link_ea_entry *lee;
1999 struct lu_buf link = { 0 };
2001 int reclen, count, rc;
2004 if (lname->ln_namelen < LL_CRYPTO_BLOCK_SIZE)
2007 buf = lu_buf_check_and_alloc(buf, PATH_MAX);
2012 rc = mdt_links_read(info, child, &ldata);
2016 hash = kmalloc(lname->ln_namelen, GFP_NOFS);
2019 rc = critical_decode(lname->ln_name, lname->ln_namelen, hash);
2022 lee = (struct link_ea_entry *)(leh + 1);
2023 for (count = 0; count < leh->leh_reccount; count++) {
2024 linkea_entry_unpack(lee, &reclen, &name, &pfid);
2025 if (!parent || lu_fid_eq(&pfid, mdt_object_fid(parent))) {
2026 lu_buf_check_and_alloc(&link, name.ln_namelen);
2028 GOTO(out_match, rc = -ENOMEM);
2029 rc = critical_decode(name.ln_name, name.ln_namelen,
2032 if (memcmp(LLCRYPT_EXTRACT_DIGEST(link.lb_buf, rc),
2033 hash, LL_CRYPTO_BLOCK_SIZE) == 0) {
2038 lee = (struct link_ea_entry *) ((char *)lee + reclen);
2040 if (count == leh->leh_reccount)
2053 * UPDATE lock should be taken against parent, and be released before exit;
2054 * child_bits lock should be taken against child, and be returned back:
2055 * (1)normal request should release the child lock;
2056 * (2)intent request will grant the lock to client.
2058 static int mdt_getattr_name_lock(struct mdt_thread_info *info,
2059 struct mdt_lock_handle *lhc,
2061 struct ldlm_reply *ldlm_rep)
2063 struct ptlrpc_request *req = mdt_info_req(info);
2064 struct mdt_body *reqbody = NULL;
2065 struct mdt_object *parent = info->mti_object;
2066 struct mdt_object *child = NULL;
2067 struct lu_fid *child_fid = &info->mti_tmp_fid1;
2068 struct lu_name *lname = NULL;
2069 struct mdt_lock_handle *lhp = NULL;
2070 struct ldlm_lock *lock;
2071 struct req_capsule *pill = info->mti_pill;
2072 bool fscrypt_md = false;
2080 is_resent = lustre_handle_is_used(&lhc->mlh_reg_lh);
2081 LASSERT(ergo(is_resent,
2082 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT));
2087 if (info->mti_mdt->mdt_enable_dir_auto_split)
2088 ma_need |= MA_DIRENT_CNT;
2090 if (info->mti_cross_ref) {
2091 /* Only getattr on the child. Parent is on another node. */
2092 mdt_set_disposition(info, ldlm_rep,
2093 DISP_LOOKUP_EXECD | DISP_LOOKUP_POS);
2095 CDEBUG(D_INODE, "partial getattr_name child_fid = "DFID", "
2097 PFID(mdt_object_fid(child)), ldlm_rep);
2099 rc = mdt_check_resent_lock(info, child, lhc);
2102 } else if (rc > 0) {
2104 * Object's name entry is on another MDS, it will
2105 * request PERM lock only because LOOKUP lock is owned
2106 * by the MDS where name entry resides.
2108 * TODO: it should try layout lock too. - Jinshan
2110 child_bits &= ~(MDS_INODELOCK_LOOKUP |
2111 MDS_INODELOCK_LAYOUT);
2112 child_bits |= MDS_INODELOCK_PERM;
2113 rc = mdt_object_lock(info, child, lhc, child_bits,
2119 /* Finally, we can get attr for child. */
2120 if (!mdt_object_exists(child)) {
2121 LU_OBJECT_DEBUG(D_INFO, info->mti_env,
2123 "remote object doesn't exist.");
2124 mdt_object_unlock(info, child, lhc, 1);
2128 rc = mdt_getattr_internal(info, child, ma_need);
2129 if (unlikely(rc != 0)) {
2130 mdt_object_unlock(info, child, lhc, 1);
2134 rc = mdt_pack_secctx_in_reply(info, child);
2136 mdt_object_unlock(info, child, lhc, 1);
2140 rc = mdt_pack_encctx_in_reply(info, child);
2142 mdt_object_unlock(info, child, lhc, 1);
2146 lname = &info->mti_name;
2147 mdt_name_unpack(pill, &RMF_NAME, lname, MNF_FIX_ANON);
2149 if (info->mti_body->mbo_valid & OBD_MD_NAMEHASH) {
2150 reqbody = req_capsule_client_get(pill, &RMF_MDT_BODY);
2151 if (unlikely(reqbody == NULL))
2152 RETURN(err_serious(-EPROTO));
2154 *child_fid = reqbody->mbo_fid2;
2155 if (unlikely(!fid_is_sane(child_fid)))
2156 RETURN(err_serious(-EINVAL));
2158 if (lu_fid_eq(mdt_object_fid(parent), child_fid)) {
2159 mdt_object_get(info->mti_env, parent);
2162 child = mdt_object_find(info->mti_env, info->mti_mdt,
2165 RETURN(PTR_ERR(child));
2168 CDEBUG(D_INODE, "getattr with lock for "DFID"/"DFID", "
2170 PFID(mdt_object_fid(parent)),
2171 PFID(&reqbody->mbo_fid2), ldlm_rep);
2172 } else if (lu_name_is_valid(lname)) {
2173 if (mdt_object_remote(parent)) {
2174 CERROR("%s: parent "DFID" is on remote target\n",
2175 mdt_obd_name(info->mti_mdt),
2176 PFID(mdt_object_fid(parent)));
2180 CDEBUG(D_INODE, "getattr with lock for "DFID"/"DNAME", "
2181 "ldlm_rep = %p\n", PFID(mdt_object_fid(parent)),
2182 PNAME(lname), ldlm_rep);
2184 if (parent->mot_obj.lo_header->loh_attr & LOHA_FSCRYPT_MD ||
2185 (fid_is_root(mdt_object_fid(parent)) &&
2186 lname->ln_namelen == strlen(dot_fscrypt_name) &&
2187 strncmp(lname->ln_name, dot_fscrypt_name,
2188 lname->ln_namelen) == 0))
2191 reqbody = req_capsule_client_get(pill, &RMF_MDT_BODY);
2192 if (unlikely(reqbody == NULL))
2193 RETURN(err_serious(-EPROTO));
2195 *child_fid = reqbody->mbo_fid2;
2196 if (unlikely(!fid_is_sane(child_fid)))
2197 RETURN(err_serious(-EINVAL));
2199 if (lu_fid_eq(mdt_object_fid(parent), child_fid)) {
2200 mdt_object_get(info->mti_env, parent);
2203 child = mdt_object_find(info->mti_env, info->mti_mdt,
2206 RETURN(PTR_ERR(child));
2209 if (mdt_object_remote(child)) {
2210 CERROR("%s: child "DFID" is on remote target\n",
2211 mdt_obd_name(info->mti_mdt),
2212 PFID(mdt_object_fid(child)));
2213 GOTO(out_child, rc = -EPROTO);
2216 /* don't fetch LOOKUP lock if it's remote object */
2217 rc = mdt_is_remote_object(info, parent, child);
2219 GOTO(out_child, rc);
2221 child_bits &= ~MDS_INODELOCK_LOOKUP;
2223 CDEBUG(D_INODE, "getattr with lock for "DFID"/"DFID", "
2225 PFID(mdt_object_fid(parent)),
2226 PFID(&reqbody->mbo_fid2), ldlm_rep);
2229 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_EXECD);
2231 if (unlikely(!mdt_object_exists(parent)) &&
2232 !(info->mti_body->mbo_valid & OBD_MD_NAMEHASH) &&
2233 lu_name_is_valid(lname)) {
2234 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2236 "Parent doesn't exist!");
2237 GOTO(out_child, rc = -ESTALE);
2240 if (!child && is_resent) {
2241 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
2243 /* Lock is pinned by ldlm_handle_enqueue0() as it is
2244 * a resend case, however, it could be already destroyed
2245 * due to client eviction or a raced cancel RPC.
2247 LDLM_DEBUG_NOLOCK("Invalid lock handle %#llx",
2248 lhc->mlh_reg_lh.cookie);
2251 fid_extract_from_res_name(child_fid,
2252 &lock->l_resource->lr_name);
2253 LDLM_LOCK_PUT(lock);
2254 child = mdt_object_find(info->mti_env, info->mti_mdt,
2257 RETURN(PTR_ERR(child));
2258 } else if (!(info->mti_body->mbo_valid & OBD_MD_NAMEHASH) &&
2259 lu_name_is_valid(lname)) {
2260 if (info->mti_body->mbo_valid == OBD_MD_FLID) {
2261 rc = mdt_raw_lookup(info, parent, lname);
2266 /* step 1: lock parent only if parent is a directory */
2267 if (S_ISDIR(lu_object_attr(&parent->mot_obj))) {
2268 lhp = &info->mti_lh[MDT_LH_PARENT];
2269 rc = mdt_parent_lock(info, parent, lhp, lname, LCK_PR);
2270 if (unlikely(rc != 0))
2274 /* step 2: lookup child's fid by name */
2275 fid_zero(child_fid);
2276 rc = mdo_lookup(info->mti_env, mdt_object_child(parent), lname,
2277 child_fid, &info->mti_spec);
2279 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
2282 GOTO(unlock_parent, rc);
2284 child = mdt_object_find(info->mti_env, info->mti_mdt,
2286 if (unlikely(IS_ERR(child)))
2287 GOTO(unlock_parent, rc = PTR_ERR(child));
2290 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
2292 /* step 3: lock child regardless if it is local or remote. */
2295 if (info->mti_body->mbo_valid & OBD_MD_NAMEHASH) {
2296 /* Here, lname is an encoded hash of on-disk name, and
2297 * client is doing access without encryption key.
2298 * So we need to compare name hash with the one in the request.
2300 if (!find_name_matching_hash(info, lname, parent,
2302 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
2303 mdt_clear_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
2304 GOTO(out_child, rc = -ENOENT);
2308 CFS_FAIL_TIMEOUT(OBD_FAIL_MDS_RESEND, obd_timeout * 2);
2309 if (!mdt_object_exists(child)) {
2310 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2312 "Object doesn't exist!");
2313 GOTO(out_child, rc = -ENOENT);
2316 rc = mdt_check_resent_lock(info, child, lhc);
2318 GOTO(out_child, rc);
2319 } else if (rc > 0) {
2320 if (!(child_bits & MDS_INODELOCK_UPDATE) &&
2321 !mdt_object_remote(child)) {
2322 struct md_attr *ma = &info->mti_attr;
2325 ma->ma_need = MA_INODE;
2326 rc = mdt_attr_get_complex(info, child, ma);
2327 if (unlikely(rc != 0))
2328 GOTO(out_child, rc);
2330 /* If the file has not been changed for some time, we
2331 * return not only a LOOKUP lock, but also an UPDATE
2332 * lock and this might save us RPC on later STAT. For
2333 * directories, it also let negative dentry cache start
2334 * working for this dir. */
2335 if (ma->ma_valid & MA_INODE &&
2336 ma->ma_attr.la_valid & LA_CTIME &&
2337 info->mti_mdt->mdt_namespace->ns_ctime_age_limit +
2338 ma->ma_attr.la_ctime < ktime_get_real_seconds())
2339 child_bits |= MDS_INODELOCK_UPDATE;
2342 /* layout lock must be granted in a best-effort way
2343 * for IT operations */
2344 LASSERT(!(child_bits & MDS_INODELOCK_LAYOUT));
2345 if (S_ISREG(lu_object_attr(&child->mot_obj)) &&
2346 !mdt_object_remote(child) && ldlm_rep != NULL) {
2347 if (!CFS_FAIL_CHECK(OBD_FAIL_MDS_NO_LL_GETATTR) &&
2348 exp_connect_layout(info->mti_exp)) {
2349 /* try to grant layout lock for regular file. */
2350 try_bits = MDS_INODELOCK_LAYOUT;
2352 /* Acquire DOM lock in advance for data-on-mdt file */
2353 if (child != parent)
2354 try_bits |= MDS_INODELOCK_DOM;
2358 * To avoid possible deadlock between batched statahead RPC
2359 * and rename()/migrate() operation, it should use trylock to
2360 * obtain the DLM PR ibits lock for file attributes in a
2361 * batched statahead RPC. A failed trylock means that other
2362 * users maybe modify the directory simultaneously as in current
2363 * Lustre design the server only grants read lock to a client.
2365 * When a trylock failed, the MDT reports the conflict with
2366 * error code -EBUSY, and stops statahead immediately.
2368 if (info->mti_batch_env) {
2370 * This is a sub stat-ahead request in a batched RPC.
2371 * However, the @child is a remote object, we just
2372 * return -EREMOTE here to forbid stat-ahead on it.
2374 if (mdt_object_remote(child))
2375 GOTO(out_child, rc = -EREMOTE);
2377 try_bits |= child_bits;
2381 if (try_bits != 0) {
2382 /* try layout lock, it may fail to be granted due to
2383 * contention at LOOKUP or UPDATE */
2384 rc = mdt_object_lock_try(info, child, lhc, &child_bits,
2386 if (child_bits & MDS_INODELOCK_LAYOUT)
2389 /* Do not enqueue the UPDATE lock from MDT(cross-MDT),
2390 * client will enqueue the lock to the remote MDT */
2391 if (mdt_object_remote(child))
2392 rc = mdt_object_lookup_lock(info, NULL, child,
2395 rc = mdt_object_lock(info, child, lhc,
2396 child_bits, LCK_PR);
2398 if (unlikely(rc != 0))
2399 GOTO(out_child, rc);
2400 if (info->mti_batch_env && child_bits == 0) {
2402 mdt_object_unlock(info, child, lhc, 1);
2403 GOTO(out_child, rc = -EBUSY);
2408 child->mot_obj.lo_header->loh_attr |= LOHA_FSCRYPT_MD;
2410 /* finally, we can get attr for child. */
2411 rc = mdt_getattr_internal(info, child, ma_need);
2412 if (unlikely(rc != 0)) {
2414 mdt_object_unlock(info, child, lhc, 1);
2415 GOTO(out_child, rc);
2418 rc = mdt_pack_secctx_in_reply(info, child);
2421 mdt_object_unlock(info, child, lhc, 1);
2422 GOTO(out_child, rc);
2425 rc = mdt_pack_encctx_in_reply(info, child);
2428 mdt_object_unlock(info, child, lhc, 1);
2429 GOTO(out_child, rc);
2432 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
2434 /* Debugging code. */
2435 LDLM_DEBUG(lock, "Returning lock to client");
2436 LASSERTF(fid_res_name_eq(mdt_object_fid(child),
2437 &lock->l_resource->lr_name),
2438 "Lock res_id: "DLDLMRES", fid: "DFID"\n",
2439 PLDLMRES(lock->l_resource),
2440 PFID(mdt_object_fid(child)));
2442 if (unlikely(CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_ENQ_RESEND))) {
2443 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
2444 CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_ENQ_RESEND,
2446 req->rq_arrival_time.tv_sec +
2448 /* Put the lock to the waiting list and force the cancel */
2449 ldlm_set_ast_sent(lock);
2452 if (S_ISREG(lu_object_attr(&child->mot_obj)) &&
2453 !mdt_object_remote(child) && child != parent) {
2454 mdt_object_put(info->mti_env, child);
2455 rc = mdt_pack_size2body(info, child_fid,
2457 if (rc != 0 && child_bits & MDS_INODELOCK_DOM) {
2458 /* DOM lock was taken in advance but this is
2459 * not DoM file. Drop the lock.
2461 lock_res_and_lock(lock);
2462 ldlm_inodebits_drop(lock, MDS_INODELOCK_DOM);
2463 unlock_res_and_lock(lock);
2465 LDLM_LOCK_PUT(lock);
2466 GOTO(unlock_parent, rc = 0);
2468 LDLM_LOCK_PUT(lock);
2474 mdt_object_put(info->mti_env, child);
2477 mdt_object_unlock(info, parent, lhp, 1);
2478 if (rc == -ENOENT) {
2479 /* return -ENOKEY instead of -ENOENT to encryption-unaware
2480 * client if trying to access an encrypted file
2482 int rc2 = mdt_check_enc(info, parent);
2490 /* normal handler: should release the child lock */
2491 static int mdt_getattr_name(struct tgt_session_info *tsi)
2493 struct mdt_thread_info *info = tsi2mdt_info(tsi);
2494 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_CHILD];
2495 struct mdt_body *reqbody;
2496 struct mdt_body *repbody;
2501 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
2502 LASSERT(reqbody != NULL);
2503 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2504 LASSERT(repbody != NULL);
2506 info->mti_cross_ref = !!(reqbody->mbo_valid & OBD_MD_FLCROSSREF);
2507 repbody->mbo_eadatasize = 0;
2508 repbody->mbo_aclsize = 0;
2510 rc = mdt_init_ucred(info, reqbody);
2512 GOTO(out_shrink, rc);
2514 rc = mdt_getattr_name_lock(info, lhc, MDS_INODELOCK_UPDATE, NULL);
2515 if (lustre_handle_is_used(&lhc->mlh_reg_lh)) {
2516 ldlm_lock_decref(&lhc->mlh_reg_lh, lhc->mlh_reg_mode);
2517 lhc->mlh_reg_lh.cookie = 0;
2519 mdt_exit_ucred(info);
2522 mdt_client_compatibility(info);
2523 rc2 = mdt_fix_reply(info);
2526 mdt_thread_info_fini(info);
2530 static int mdt_rmfid_unlink(struct mdt_thread_info *info,
2531 const struct lu_fid *pfid,
2532 const struct lu_name *name,
2533 struct mdt_object *obj, s64 ctime)
2535 struct lu_fid *child_fid = &info->mti_tmp_fid1;
2536 struct ldlm_enqueue_info *einfo = &info->mti_einfo;
2537 struct mdt_device *mdt = info->mti_mdt;
2538 struct md_attr *ma = &info->mti_attr;
2539 struct mdt_lock_handle *parent_lh;
2540 struct mdt_lock_handle *child_lh;
2541 struct mdt_object *pobj;
2545 pobj = mdt_object_find(info->mti_env, mdt, pfid);
2547 GOTO(out, rc = PTR_ERR(pobj));
2549 parent_lh = &info->mti_lh[MDT_LH_PARENT];
2550 rc = mdt_parent_lock(info, pobj, parent_lh, name, LCK_PW);
2552 GOTO(put_parent, rc);
2554 rc = mdo_lookup(info->mti_env, mdt_object_child(pobj),
2555 name, child_fid, &info->mti_spec);
2557 GOTO(unlock_parent, rc);
2559 if (!lu_fid_eq(child_fid, mdt_object_fid(obj)))
2560 GOTO(unlock_parent, rc = -EREMCHG);
2562 child_lh = &info->mti_lh[MDT_LH_CHILD];
2563 rc = mdt_object_stripes_lock(info, pobj, obj, child_lh, einfo,
2564 MDS_INODELOCK_LOOKUP |
2565 MDS_INODELOCK_UPDATE, LCK_EX);
2567 GOTO(unlock_parent, rc);
2569 if (atomic_read(&obj->mot_open_count)) {
2570 CDEBUG(D_OTHER, "object "DFID" open, skip\n",
2571 PFID(mdt_object_fid(obj)));
2572 GOTO(unlock_child, rc = -EBUSY);
2576 ma->ma_valid = MA_INODE;
2577 ma->ma_attr.la_valid = LA_CTIME;
2578 ma->ma_attr.la_ctime = ctime;
2580 mutex_lock(&obj->mot_lov_mutex);
2582 rc = mdo_unlink(info->mti_env, mdt_object_child(pobj),
2583 mdt_object_child(obj), name, ma, 0);
2585 mutex_unlock(&obj->mot_lov_mutex);
2588 mdt_object_stripes_unlock(info, obj, child_lh, einfo, 1);
2590 mdt_object_unlock(info, pobj, parent_lh, 1);
2592 mdt_object_put(info->mti_env, pobj);
2597 static int mdt_rmfid_check_permission(struct mdt_thread_info *info,
2598 struct mdt_object *obj)
2600 struct lu_ucred *uc = lu_ucred(info->mti_env);
2601 struct md_attr *ma = &info->mti_attr;
2602 struct lu_attr *la = &ma->ma_attr;
2606 ma->ma_need = MA_INODE;
2607 rc = mo_attr_get(info->mti_env, mdt_object_child(obj), ma);
2611 if (la->la_flags & LUSTRE_IMMUTABLE_FL)
2614 /* we want rbac roles to have precedence over any other
2615 * permission or capability checks
2617 if (!uc->uc_rbac_byfid_ops)
2619 if (cap_raised(uc->uc_cap, CAP_DAC_OVERRIDE))
2621 if (uc->uc_fsuid == la->la_uid) {
2622 if ((la->la_mode & S_IWUSR) == 0)
2624 } else if (uc->uc_fsgid == la->la_gid) {
2625 if ((la->la_mode & S_IWGRP) == 0)
2627 } else if ((la->la_mode & S_IWOTH) == 0) {
2635 static int mdt_rmfid_one(struct mdt_thread_info *info, struct lu_fid *fid,
2638 struct mdt_device *mdt = info->mti_mdt;
2639 struct mdt_object *obj = NULL;
2640 struct linkea_data ldata = { NULL };
2641 struct lu_buf *buf = &info->mti_big_buf;
2642 struct lu_name *name = &info->mti_name;
2643 struct lu_fid *pfid = &info->mti_tmp_fid1;
2644 struct link_ea_header *leh;
2645 struct link_ea_entry *lee;
2646 int reclen, count, rc = 0;
2649 if (!fid_is_sane(fid))
2650 GOTO(out, rc = -EINVAL);
2652 if (!fid_is_namespace_visible(fid))
2653 GOTO(out, rc = -EINVAL);
2655 obj = mdt_object_find(info->mti_env, mdt, fid);
2657 GOTO(out, rc = PTR_ERR(obj));
2659 if (mdt_object_remote(obj))
2660 GOTO(out, rc = -EREMOTE);
2661 if (!mdt_object_exists(obj) || lu_object_is_dying(&obj->mot_header))
2662 GOTO(out, rc = -ENOENT);
2664 rc = mdt_rmfid_check_permission(info, obj);
2669 buf = lu_buf_check_and_alloc(buf, PATH_MAX);
2671 GOTO(out, rc = -ENOMEM);
2674 rc = mdt_links_read(info, obj, &ldata);
2679 lee = (struct link_ea_entry *)(leh + 1);
2680 for (count = 0; count < leh->leh_reccount; count++) {
2681 /* remove every hardlink */
2682 linkea_entry_unpack(lee, &reclen, name, pfid);
2683 lee = (struct link_ea_entry *) ((char *)lee + reclen);
2684 rc = mdt_rmfid_unlink(info, pfid, name, obj, ctime);
2690 if (obj && !IS_ERR(obj))
2691 mdt_object_put(info->mti_env, obj);
2692 if (info->mti_big_buf.lb_buf)
2693 lu_buf_free(&info->mti_big_buf);
2698 static int mdt_rmfid(struct tgt_session_info *tsi)
2700 struct mdt_thread_info *mti = tsi2mdt_info(tsi);
2701 struct mdt_body *reqbody;
2702 struct lu_fid *fids, *rfids;
2708 reqbody = req_capsule_client_get(tsi->tsi_pill, &RMF_MDT_BODY);
2709 if (reqbody == NULL)
2711 bufsize = req_capsule_get_size(tsi->tsi_pill, &RMF_FID_ARRAY,
2713 nr = bufsize / sizeof(struct lu_fid);
2714 if (nr * sizeof(struct lu_fid) != bufsize)
2716 req_capsule_set_size(tsi->tsi_pill, &RMF_RCS,
2717 RCL_SERVER, nr * sizeof(__u32));
2718 req_capsule_set_size(tsi->tsi_pill, &RMF_FID_ARRAY,
2719 RCL_SERVER, nr * sizeof(struct lu_fid));
2720 rc = req_capsule_server_pack(tsi->tsi_pill);
2722 GOTO(out, rc = err_serious(rc));
2723 fids = req_capsule_client_get(tsi->tsi_pill, &RMF_FID_ARRAY);
2726 rcs = req_capsule_server_get(tsi->tsi_pill, &RMF_RCS);
2728 rfids = req_capsule_server_get(tsi->tsi_pill, &RMF_FID_ARRAY);
2731 mdt_init_ucred(mti, reqbody);
2732 for (i = 0; i < nr; i++) {
2734 rcs[i] = mdt_rmfid_one(mti, fids + i, reqbody->mbo_ctime);
2736 mdt_exit_ucred(mti);
2742 static int mdt_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2743 void *karg, void __user *uarg);
2745 int mdt_io_set_info(struct tgt_session_info *tsi)
2747 struct ptlrpc_request *req = tgt_ses_req(tsi);
2748 struct ost_body *body = NULL, *repbody;
2749 void *key, *val = NULL;
2750 int keylen, vallen, rc = 0;
2751 bool is_grant_shrink;
2755 key = req_capsule_client_get(tsi->tsi_pill, &RMF_SETINFO_KEY);
2757 DEBUG_REQ(D_HA, req, "no set_info key");
2758 RETURN(err_serious(-EFAULT));
2760 keylen = req_capsule_get_size(tsi->tsi_pill, &RMF_SETINFO_KEY,
2763 val = req_capsule_client_get(tsi->tsi_pill, &RMF_SETINFO_VAL);
2765 DEBUG_REQ(D_HA, req, "no set_info val");
2766 RETURN(err_serious(-EFAULT));
2768 vallen = req_capsule_get_size(tsi->tsi_pill, &RMF_SETINFO_VAL,
2771 is_grant_shrink = KEY_IS(KEY_GRANT_SHRINK);
2772 if (is_grant_shrink)
2773 /* In this case the value is actually an RMF_OST_BODY, so we
2774 * transmutate the type of this PTLRPC */
2775 req_capsule_extend(tsi->tsi_pill, &RQF_OST_SET_GRANT_INFO);
2777 rc = req_capsule_server_pack(tsi->tsi_pill);
2781 if (is_grant_shrink) {
2782 body = req_capsule_client_get(tsi->tsi_pill, &RMF_OST_BODY);
2784 repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_OST_BODY);
2787 /** handle grant shrink, similar to a read request */
2788 tgt_grant_prepare_read(tsi->tsi_env, tsi->tsi_exp,
2791 CERROR("%s: Unsupported key %s\n",
2792 tgt_name(tsi->tsi_tgt), (char *)key);
2800 static int mdt_set_info(struct tgt_session_info *tsi)
2802 struct ptlrpc_request *req = tgt_ses_req(tsi);
2805 int keylen, vallen, rc = 0;
2809 key = req_capsule_client_get(tsi->tsi_pill, &RMF_SETINFO_KEY);
2811 DEBUG_REQ(D_HA, req, "no set_info key");
2812 RETURN(err_serious(-EFAULT));
2815 keylen = req_capsule_get_size(tsi->tsi_pill, &RMF_SETINFO_KEY,
2818 val = req_capsule_client_get(tsi->tsi_pill, &RMF_SETINFO_VAL);
2820 DEBUG_REQ(D_HA, req, "no set_info val");
2821 RETURN(err_serious(-EFAULT));
2824 vallen = req_capsule_get_size(tsi->tsi_pill, &RMF_SETINFO_VAL,
2827 /* Swab any part of val you need to here */
2828 if (KEY_IS(KEY_READ_ONLY)) {
2829 spin_lock(&req->rq_export->exp_lock);
2831 *exp_connect_flags_ptr(req->rq_export) |=
2834 *exp_connect_flags_ptr(req->rq_export) &=
2835 ~OBD_CONNECT_RDONLY;
2836 spin_unlock(&req->rq_export->exp_lock);
2837 } else if (KEY_IS(KEY_CHANGELOG_CLEAR)) {
2838 struct changelog_setinfo *cs = val;
2840 if (vallen != sizeof(*cs)) {
2841 CERROR("%s: bad changelog_clear setinfo size %d\n",
2842 tgt_name(tsi->tsi_tgt), vallen);
2845 if (req_capsule_req_need_swab(&req->rq_pill)) {
2846 __swab64s(&cs->cs_recno);
2847 __swab32s(&cs->cs_id);
2850 if (!mdt_changelog_allow(tsi2mdt_info(tsi)))
2852 rc = mdt_iocontrol(OBD_IOC_CHANGELOG_CLEAR, req->rq_export,
2854 } else if (KEY_IS(KEY_EVICT_BY_NID)) {
2856 obd_export_evict_by_nid(req->rq_export->exp_obd, val);
2863 static int mdt_readpage(struct tgt_session_info *tsi)
2865 struct mdt_thread_info *info = mdt_th_info(tsi->tsi_env);
2866 struct mdt_object *object = mdt_obj(tsi->tsi_corpus);
2867 struct lu_rdpg *rdpg = &info->mti_u.rdpg.mti_rdpg;
2868 const struct mdt_body *reqbody = tsi->tsi_mdt_body;
2869 struct mdt_body *repbody;
2875 if (CFS_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK))
2876 RETURN(err_serious(-ENOMEM));
2878 repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_MDT_BODY);
2879 if (repbody == NULL || reqbody == NULL)
2880 RETURN(err_serious(-EFAULT));
2883 * prepare @rdpg before calling lower layers and transfer itself. Here
2884 * reqbody->size contains offset of where to start to read and
2885 * reqbody->nlink contains number bytes to read.
2887 rdpg->rp_hash = reqbody->mbo_size;
2888 if (rdpg->rp_hash != reqbody->mbo_size) {
2889 CERROR("Invalid hash: %#llx != %#llx\n",
2890 rdpg->rp_hash, reqbody->mbo_size);
2894 rdpg->rp_attrs = reqbody->mbo_mode;
2895 if (exp_connect_flags(tsi->tsi_exp) & OBD_CONNECT_64BITHASH)
2896 rdpg->rp_attrs |= LUDA_64BITHASH;
2897 rdpg->rp_count = min_t(unsigned int, reqbody->mbo_nlink,
2898 exp_max_brw_size(tsi->tsi_exp));
2899 rdpg->rp_npages = (rdpg->rp_count + PAGE_SIZE - 1) >>
2901 OBD_ALLOC_PTR_ARRAY_LARGE(rdpg->rp_pages, rdpg->rp_npages);
2902 if (rdpg->rp_pages == NULL)
2905 for (i = 0; i < rdpg->rp_npages; ++i) {
2906 rdpg->rp_pages[i] = alloc_page(GFP_NOFS);
2907 if (rdpg->rp_pages[i] == NULL)
2908 GOTO(free_rdpg, rc = -ENOMEM);
2911 /* call lower layers to fill allocated pages with directory data */
2912 rc = mo_readpage(tsi->tsi_env, mdt_object_child(object), rdpg);
2914 GOTO(free_rdpg, rc);
2916 /* send pages to client */
2917 rc = tgt_sendpage(tsi, rdpg, rc);
2922 for (i = 0; i < rdpg->rp_npages; i++)
2923 if (rdpg->rp_pages[i] != NULL)
2924 __free_page(rdpg->rp_pages[i]);
2925 OBD_FREE_PTR_ARRAY_LARGE(rdpg->rp_pages, rdpg->rp_npages);
2927 if (CFS_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE))
2933 static int mdt_fix_attr_ucred(struct mdt_thread_info *info, __u32 op)
2935 struct lu_ucred *uc = mdt_ucred_check(info);
2936 struct lu_attr *attr = &info->mti_attr.ma_attr;
2941 if (op != REINT_SETATTR) {
2942 if ((attr->la_valid & LA_UID) && (attr->la_uid != -1))
2943 attr->la_uid = uc->uc_fsuid;
2944 /* for S_ISGID, inherit gid from his parent, such work will be
2945 * done in cmm/mdd layer, here set all cases as uc->uc_fsgid. */
2946 if ((attr->la_valid & LA_GID) && (attr->la_gid != -1))
2947 attr->la_gid = uc->uc_fsgid;
2953 static inline bool mdt_is_readonly_open(struct mdt_thread_info *info, __u32 op)
2955 return op == REINT_OPEN &&
2956 !(info->mti_spec.sp_cr_flags & (MDS_FMODE_WRITE | MDS_OPEN_CREAT));
2959 static void mdt_preset_secctx_size(struct mdt_thread_info *info)
2961 struct req_capsule *pill = info->mti_pill;
2963 if (req_capsule_has_field(pill, &RMF_FILE_SECCTX,
2965 req_capsule_has_field(pill, &RMF_FILE_SECCTX_NAME,
2967 if (req_capsule_get_size(pill, &RMF_FILE_SECCTX_NAME,
2969 /* pre-set size in server part with max size */
2970 req_capsule_set_size(pill, &RMF_FILE_SECCTX,
2972 req_capsule_ptlreq(pill) ?
2973 OBD_MAX_DEFAULT_EA_SIZE :
2976 req_capsule_set_size(pill, &RMF_FILE_SECCTX,
2981 static int mdt_reint_internal(struct mdt_thread_info *info,
2982 struct mdt_lock_handle *lhc,
2985 struct req_capsule *pill = info->mti_pill;
2986 struct mdt_body *repbody;
2991 rc = mdt_reint_unpack(info, op);
2993 CERROR("Can't unpack reint, rc %d\n", rc);
2994 RETURN(err_serious(rc));
2998 /* check if the file system is set to readonly. O_RDONLY open
2999 * is still allowed even the file system is set to readonly mode */
3000 if (mdt_rdonly(info->mti_exp) && !mdt_is_readonly_open(info, op))
3001 RETURN(err_serious(-EROFS));
3003 /* for replay (no_create) lmm is not needed, client has it already */
3004 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
3005 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER,
3008 /* llog cookies are always 0, the field is kept for compatibility */
3009 if (req_capsule_has_field(pill, &RMF_LOGCOOKIES, RCL_SERVER))
3010 req_capsule_set_size(pill, &RMF_LOGCOOKIES, RCL_SERVER, 0);
3012 /* Set ACL reply buffer size as LUSTRE_POSIX_ACL_MAX_SIZE_OLD
3013 * by default. If the target object has more ACL entries, then
3014 * enlarge the buffer when necessary. */
3015 if (req_capsule_has_field(pill, &RMF_ACL, RCL_SERVER))
3016 req_capsule_set_size(pill, &RMF_ACL, RCL_SERVER,
3017 LUSTRE_POSIX_ACL_MAX_SIZE_OLD);
3019 mdt_preset_secctx_size(info);
3020 mdt_preset_encctx_size(info);
3022 rc = req_capsule_server_pack(pill);
3024 CERROR("Can't pack response, rc %d\n", rc);
3025 RETURN(err_serious(rc));
3028 if (req_capsule_has_field(pill, &RMF_MDT_BODY, RCL_SERVER)) {
3029 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
3031 repbody->mbo_eadatasize = 0;
3032 repbody->mbo_aclsize = 0;
3035 CFS_FAIL_TIMEOUT(OBD_FAIL_MDS_REINT_DELAY, 10);
3037 /* for replay no cookkie / lmm need, because client have this already */
3038 if (info->mti_spec.no_create)
3039 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
3040 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER, 0);
3042 rc = mdt_init_ucred_reint(info);
3044 GOTO(out_shrink, rc);
3046 rc = mdt_fix_attr_ucred(info, op);
3048 GOTO(out_ucred, rc = err_serious(rc));
3050 rc = mdt_check_resent(info, mdt_reconstruct, lhc);
3052 GOTO(out_ucred, rc);
3053 } else if (rc == 1) {
3054 DEBUG_REQ(D_INODE, mdt_info_req(info), "resent opt");
3055 rc = lustre_msg_get_status(mdt_info_req(info)->rq_repmsg);
3056 GOTO(out_ucred, rc);
3058 rc = mdt_reint_rec(info, lhc);
3061 mdt_exit_ucred(info);
3063 mdt_client_compatibility(info);
3065 rc2 = mdt_fix_reply(info);
3070 * Data-on-MDT optimization - read data along with OPEN and return it
3071 * in reply when possible.
3073 if (rc == 0 && op == REINT_OPEN && !req_is_replay(pill->rc_req))
3074 rc = mdt_dom_read_on_open(info, info->mti_mdt,
3080 static long mdt_reint_opcode(struct ptlrpc_request *req,
3081 const struct req_format **fmt)
3083 struct mdt_device *mdt;
3084 struct mdt_rec_reint *rec;
3087 rec = req_capsule_client_get(&req->rq_pill, &RMF_REC_REINT);
3089 opc = rec->rr_opcode;
3090 DEBUG_REQ(D_INODE, req, "reint opt = %ld", opc);
3091 if (opc < REINT_MAX && fmt[opc] != NULL)
3092 req_capsule_extend(&req->rq_pill, fmt[opc]);
3094 mdt = mdt_exp2dev(req->rq_export);
3095 CERROR("%s: Unsupported opcode '%ld' from client '%s':"
3096 " rc = %d\n", req->rq_export->exp_obd->obd_name,
3097 opc, mdt->mdt_ldlm_client->cli_name, -EFAULT);
3098 opc = err_serious(-EFAULT);
3101 opc = err_serious(-EFAULT);
3106 static int mdt_reint(struct tgt_session_info *tsi)
3110 static const struct req_format *reint_fmts[REINT_MAX] = {
3111 [REINT_SETATTR] = &RQF_MDS_REINT_SETATTR,
3112 [REINT_CREATE] = &RQF_MDS_REINT_CREATE,
3113 [REINT_LINK] = &RQF_MDS_REINT_LINK,
3114 [REINT_UNLINK] = &RQF_MDS_REINT_UNLINK,
3115 [REINT_RENAME] = &RQF_MDS_REINT_RENAME,
3116 [REINT_OPEN] = &RQF_MDS_REINT_OPEN,
3117 [REINT_SETXATTR] = &RQF_MDS_REINT_SETXATTR,
3118 [REINT_RMENTRY] = &RQF_MDS_REINT_UNLINK,
3119 [REINT_MIGRATE] = &RQF_MDS_REINT_MIGRATE,
3120 [REINT_RESYNC] = &RQF_MDS_REINT_RESYNC,
3125 opc = mdt_reint_opcode(tgt_ses_req(tsi), reint_fmts);
3127 struct mdt_thread_info *info = tsi2mdt_info(tsi);
3129 * No lock possible here from client to pass it to reint code
3132 rc = mdt_reint_internal(info, NULL, opc);
3133 mdt_thread_info_fini(info);
3138 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_REINT_NET_REP;
3142 /* this should sync the whole device */
3143 int mdt_device_sync(const struct lu_env *env, struct mdt_device *mdt)
3145 struct dt_device *dt = mdt->mdt_bottom;
3149 rc = dt->dd_ops->dt_sync(env, dt);
3153 /* this should sync this object */
3154 static int mdt_object_sync(const struct lu_env *env, struct obd_export *exp,
3155 struct mdt_object *mo)
3161 if (!mdt_object_exists(mo)) {
3162 CWARN("%s: non existing object "DFID": rc = %d\n",
3163 exp->exp_obd->obd_name, PFID(mdt_object_fid(mo)),
3168 if (S_ISREG(lu_object_attr(&mo->mot_obj))) {
3169 struct lu_target *tgt = tgt_ses_info(env)->tsi_tgt;
3170 dt_obj_version_t version;
3172 version = dt_version_get(env, mdt_obj2dt(mo));
3173 if (version > tgt->lut_obd->obd_last_committed)
3174 rc = mo_object_sync(env, mdt_object_child(mo));
3176 rc = mo_object_sync(env, mdt_object_child(mo));
3182 static int mdt_sync(struct tgt_session_info *tsi)
3184 struct ptlrpc_request *req = tgt_ses_req(tsi);
3185 struct req_capsule *pill = tsi->tsi_pill;
3186 struct mdt_body *body;
3187 ktime_t kstart = ktime_get();
3192 if (CFS_FAIL_CHECK(OBD_FAIL_MDS_SYNC_PACK))
3193 RETURN(err_serious(-ENOMEM));
3195 if (fid_seq(&tsi->tsi_mdt_body->mbo_fid1) == 0) {
3196 rc = mdt_device_sync(tsi->tsi_env, mdt_exp2dev(tsi->tsi_exp));
3198 struct mdt_thread_info *info = tsi2mdt_info(tsi);
3200 if (unlikely(info->mti_object == NULL))
3203 /* sync an object */
3204 rc = mdt_object_sync(tsi->tsi_env, tsi->tsi_exp,
3207 const struct lu_fid *fid;
3208 struct lu_attr *la = &info->mti_attr.ma_attr;
3210 info->mti_attr.ma_need = MA_INODE;
3211 info->mti_attr.ma_valid = 0;
3212 rc = mdt_attr_get_complex(info, info->mti_object,
3215 body = req_capsule_server_get(pill,
3217 fid = mdt_object_fid(info->mti_object);
3218 mdt_pack_attr2body(info, body, la, fid);
3221 mdt_thread_info_fini(info);
3224 mdt_counter_incr(req, LPROC_MDT_SYNC,
3225 ktime_us_delta(ktime_get(), kstart));
3230 static int mdt_data_sync(struct tgt_session_info *tsi)
3232 struct mdt_thread_info *info;
3233 struct mdt_device *mdt = mdt_exp2dev(tsi->tsi_exp);
3234 struct ost_body *body = tsi->tsi_ost_body;
3235 struct ost_body *repbody;
3236 struct mdt_object *mo = NULL;
3242 repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_OST_BODY);
3244 /* if no fid is specified then do nothing,
3245 * device sync is done via MDS_SYNC */
3246 if (fid_is_zero(&tsi->tsi_fid))
3249 mo = mdt_object_find(tsi->tsi_env, mdt, &tsi->tsi_fid);
3251 RETURN(PTR_ERR(mo));
3253 rc = mdt_object_sync(tsi->tsi_env, tsi->tsi_exp, mo);
3257 repbody->oa.o_oi = body->oa.o_oi;
3258 repbody->oa.o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
3260 info = tsi2mdt_info(tsi);
3261 ma = &info->mti_attr;
3262 ma->ma_need = MA_INODE;
3264 rc = mdt_attr_get_complex(info, mo, ma);
3266 obdo_from_la(&repbody->oa, &ma->ma_attr, VALID_FLAGS);
3269 mdt_thread_info_fini(info);
3274 mdt_object_put(tsi->tsi_env, mo);
3279 * Handle quota control requests to consult current usage/limit, but also
3280 * to configure quota enforcement
3282 static int mdt_quotactl(struct tgt_session_info *tsi)
3284 struct obd_export *exp = tsi->tsi_exp;
3285 struct req_capsule *pill = tsi->tsi_pill;
3286 struct obd_quotactl *oqctl, *repoqc;
3288 struct mdt_device *mdt = mdt_exp2dev(exp);
3289 struct lu_device *qmt = mdt->mdt_qmt_dev;
3290 struct lu_nodemap *nodemap;
3293 oqctl = req_capsule_client_get(pill, &RMF_OBD_QUOTACTL);
3295 RETURN(err_serious(-EPROTO));
3297 rc = req_capsule_server_pack(pill);
3299 RETURN(err_serious(rc));
3301 nodemap = nodemap_get_from_exp(exp);
3302 if (IS_ERR(nodemap))
3303 RETURN(PTR_ERR(nodemap));
3305 switch (oqctl->qc_cmd) {
3306 /* master quotactl */
3309 case LUSTRE_Q_SETDEFAULT:
3310 case LUSTRE_Q_SETQUOTAPOOL:
3311 case LUSTRE_Q_SETINFOPOOL:
3312 case LUSTRE_Q_SETDEFAULT_POOL:
3313 case LUSTRE_Q_DELETEQID:
3314 case LUSTRE_Q_RESETQID:
3315 if (!nodemap_can_setquota(nodemap, oqctl->qc_type,
3317 GOTO(out_nodemap, rc = -EPERM);
3321 case LUSTRE_Q_GETDEFAULT:
3322 case LUSTRE_Q_GETQUOTAPOOL:
3323 case LUSTRE_Q_GETINFOPOOL:
3324 case LUSTRE_Q_GETDEFAULT_POOL:
3326 GOTO(out_nodemap, rc = -EOPNOTSUPP);
3327 /* slave quotactl */
3334 CERROR("%s: unsupported quotactl command %d: rc = %d\n",
3335 mdt_obd_name(mdt), oqctl->qc_cmd, rc);
3336 GOTO(out_nodemap, rc);
3340 switch (oqctl->qc_type) {
3342 id = nodemap_map_id(nodemap, NODEMAP_UID,
3343 NODEMAP_CLIENT_TO_FS, id);
3346 id = nodemap_map_id(nodemap, NODEMAP_GID,
3347 NODEMAP_CLIENT_TO_FS, id);
3350 id = nodemap_map_id(nodemap, NODEMAP_PROJID,
3351 NODEMAP_CLIENT_TO_FS, id);
3354 GOTO(out_nodemap, rc = -EOPNOTSUPP);
3356 repoqc = req_capsule_server_get(pill, &RMF_OBD_QUOTACTL);
3358 GOTO(out_nodemap, rc = err_serious(-EFAULT));
3360 if (oqctl->qc_cmd == Q_SETINFO || oqctl->qc_cmd == Q_SETQUOTA)
3361 barrier_exit(tsi->tsi_tgt->lut_bottom);
3363 if (oqctl->qc_id != id)
3364 swap(oqctl->qc_id, id);
3366 if (oqctl->qc_cmd == Q_SETINFO || oqctl->qc_cmd == Q_SETQUOTA) {
3367 if (unlikely(!barrier_entry(tsi->tsi_tgt->lut_bottom)))
3368 RETURN(-EINPROGRESS);
3371 switch (oqctl->qc_cmd) {
3377 case LUSTRE_Q_SETDEFAULT:
3378 case LUSTRE_Q_GETDEFAULT:
3379 case LUSTRE_Q_SETQUOTAPOOL:
3380 case LUSTRE_Q_GETQUOTAPOOL:
3381 case LUSTRE_Q_SETINFOPOOL:
3382 case LUSTRE_Q_GETINFOPOOL:
3383 case LUSTRE_Q_SETDEFAULT_POOL:
3384 case LUSTRE_Q_GETDEFAULT_POOL:
3385 case LUSTRE_Q_DELETEQID:
3386 case LUSTRE_Q_RESETQID:
3387 /* forward quotactl request to QMT */
3388 rc = qmt_hdls.qmth_quotactl(tsi->tsi_env, qmt, oqctl);
3393 /* slave quotactl */
3394 rc = lquotactl_slv(tsi->tsi_env, tsi->tsi_tgt->lut_bottom,
3399 CERROR("Unsupported quotactl command: %d\n", oqctl->qc_cmd);
3400 GOTO(out_nodemap, rc = -EFAULT);
3403 if (oqctl->qc_id != id)
3404 swap(oqctl->qc_id, id);
3406 QCTL_COPY_NO_PNAME(repoqc, oqctl);
3410 nodemap_putref(nodemap);
3415 /** clone llog ctxt from child (mdd)
3416 * This allows remote llog (replicator) access.
3417 * We can either pass all llog RPCs (eg mdt_llog_create) on to child where the
3418 * context was originally set up, or we can handle them directly.
3419 * I choose the latter, but that means I need any llog
3420 * contexts set up by child to be accessable by the mdt. So we clone the
3421 * context into our context list here.
3423 static int mdt_llog_ctxt_clone(const struct lu_env *env, struct mdt_device *mdt,
3426 struct md_device *next = mdt->mdt_child;
3427 struct llog_ctxt *ctxt;
3430 if (!llog_ctxt_null(mdt2obd_dev(mdt), idx))
3433 rc = next->md_ops->mdo_llog_ctxt_get(env, next, idx, (void **)&ctxt);
3434 if (rc || ctxt == NULL) {
3438 rc = llog_group_set_ctxt(&mdt2obd_dev(mdt)->obd_olg, ctxt, idx);
3440 CERROR("Can't set mdt ctxt %d\n", rc);
3445 static int mdt_llog_ctxt_unclone(const struct lu_env *env,
3446 struct mdt_device *mdt, int idx)
3448 struct llog_ctxt *ctxt;
3450 ctxt = llog_get_context(mdt2obd_dev(mdt), idx);
3453 /* Put once for the get we just did, and once for the clone */
3454 llog_ctxt_put(ctxt);
3455 llog_ctxt_put(ctxt);
3460 * sec context handlers
3462 static int mdt_sec_ctx_handle(struct tgt_session_info *tsi)
3464 CFS_FAIL_TIMEOUT(OBD_FAIL_SEC_CTX_HDL_PAUSE, cfs_fail_val);
3470 * quota request handlers
3472 static int mdt_quota_dqacq(struct tgt_session_info *tsi)
3474 struct mdt_device *mdt = mdt_exp2dev(tsi->tsi_exp);
3475 struct lu_device *qmt = mdt->mdt_qmt_dev;
3480 RETURN(err_serious(-EOPNOTSUPP));
3482 rc = qmt_hdls.qmth_dqacq(tsi->tsi_env, qmt, tgt_ses_req(tsi));
3486 struct mdt_object *mdt_object_new(const struct lu_env *env,
3487 struct mdt_device *d,
3488 const struct lu_fid *f)
3490 struct lu_object_conf conf = { .loc_flags = LOC_F_NEW };
3491 struct lu_object *o;
3492 struct mdt_object *m;
3495 CDEBUG(D_INFO, "Allocate object for "DFID"\n", PFID(f));
3496 o = lu_object_find(env, &d->mdt_lu_dev, f, &conf);
3497 if (unlikely(IS_ERR(o)))
3498 m = (struct mdt_object *)o;
3504 struct mdt_object *mdt_object_find(const struct lu_env *env,
3505 struct mdt_device *d,
3506 const struct lu_fid *f)
3508 struct lu_object *o;
3509 struct mdt_object *m;
3512 CDEBUG(D_INFO, "Find object for "DFID"\n", PFID(f));
3513 o = lu_object_find(env, &d->mdt_lu_dev, f, NULL);
3514 if (unlikely(IS_ERR(o)))
3515 m = (struct mdt_object *)o;
3523 * Asyncronous commit for mdt device.
3525 * Pass asynchonous commit call down the MDS stack.
3527 * \param env environment
3528 * \param mdt the mdt device
3530 static void mdt_device_commit_async(const struct lu_env *env,
3531 struct mdt_device *mdt)
3533 struct dt_device *dt = mdt->mdt_bottom;
3537 rc = dt->dd_ops->dt_commit_async(env, dt);
3538 if (unlikely(rc != 0))
3539 CWARN("%s: async commit start failed: rc = %d\n",
3540 mdt_obd_name(mdt), rc);
3541 atomic_inc(&mdt->mdt_async_commit_count);
3546 * Mark the lock as "synchonous".
3548 * Mark the lock to deffer transaction commit to the unlock time.
3550 * \param lock the lock to mark as "synchonous"
3552 * \see mdt_is_lock_sync
3553 * \see mdt_save_lock
3555 static inline void mdt_set_lock_sync(struct ldlm_lock *lock)
3557 lock->l_ast_data = (void*)1;
3561 * Check whehter the lock "synchonous" or not.
3563 * \param lock the lock to check
3564 * \retval 1 the lock is "synchonous"
3565 * \retval 0 the lock isn't "synchronous"
3567 * \see mdt_set_lock_sync
3568 * \see mdt_save_lock
3570 static inline int mdt_is_lock_sync(struct ldlm_lock *lock)
3572 return lock->l_ast_data != NULL;
3576 * Blocking AST for mdt locks.
3578 * Starts transaction commit if in case of COS lock conflict or
3579 * deffers such a commit to the mdt_save_lock.
3581 * \param lock the lock which blocks a request or cancelling lock
3582 * \param desc unused
3583 * \param data unused
3584 * \param flag indicates whether this cancelling or blocking callback
3586 * \see ldlm_blocking_ast_nocheck
3588 int mdt_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
3589 void *data, int flag)
3591 struct obd_device *obd = ldlm_lock_to_ns(lock)->ns_obd;
3592 struct mdt_device *mdt = mdt_dev(obd->obd_lu_dev);
3593 struct ldlm_cb_set_arg *arg = data;
3594 bool commit_async = false;
3598 if (flag == LDLM_CB_CANCELING)
3601 lock_res_and_lock(lock);
3602 if (lock->l_blocking_ast != mdt_blocking_ast) {
3603 unlock_res_and_lock(lock);
3607 /* A blocking ast may be sent from ldlm_lock_decref_internal
3608 * when the last reference to a local lock was released and
3609 * during blocking event from ldlm_work_bl_ast_lock().
3610 * The 'data' parameter is l_ast_data in the first case and
3611 * callback arguments in the second one. Distinguish them by that.
3613 if (data && data != lock->l_ast_data && arg->bl_desc) {
3614 if (lock->l_req_mode & (LCK_COS | LCK_TXN))
3615 commit_async = true;
3616 else if ((lock->l_req_mode & (LCK_PW | LCK_EX)) &&
3617 ((mdt_cos_is_enabled(mdt) &&
3618 !arg->bl_desc->bl_same_client) ||
3619 (mdt_slc_is_enabled(mdt) &&
3620 arg->bl_desc->bl_txn_dependent)))
3621 mdt_set_lock_sync(lock);
3624 rc = ldlm_blocking_ast_nocheck(lock);
3629 rc = lu_env_init(&env, LCT_LOCAL);
3630 if (unlikely(rc != 0))
3631 CWARN("%s: lu_env initialization failed, cannot "
3632 "start asynchronous commit: rc = %d\n",
3635 mdt_device_commit_async(&env, mdt);
3642 * Blocking AST for cross-MDT lock
3644 * Discard lock from uncommitted_slc_locks and cancel it.
3646 * \param lock the lock which blocks a request or cancelling lock
3647 * \param desc unused
3648 * \param data unused
3649 * \param flag indicates whether this cancelling or blocking callback
3650 * \retval 0 on success
3651 * \retval negative number on error
3653 int mdt_remote_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
3654 void *data, int flag)
3660 case LDLM_CB_BLOCKING: {
3661 struct lustre_handle lockh;
3663 ldlm_lock2handle(lock, &lockh);
3664 rc = ldlm_cli_cancel(&lockh,
3665 ldlm_is_atomic_cb(lock) ? 0 : LCF_ASYNC);
3667 CDEBUG(D_INODE, "ldlm_cli_cancel: %d\n", rc);
3672 case LDLM_CB_CANCELING: {
3673 struct obd_device *obd = ldlm_lock_to_ns(lock)->ns_obd;
3674 struct mdt_device *mdt =
3675 mdt_dev(obd->obd_lu_dev->ld_site->ls_top_dev);
3677 LDLM_DEBUG(lock, "Revoke remote lock");
3679 /* discard slc lock here so that it can be cleaned anytime,
3680 * especially for cleanup_resource() */
3681 tgt_discard_slc_lock(&mdt->mdt_lut, lock);
3683 /* once we cache lock, l_ast_data is set to mdt_object */
3684 if (lock->l_ast_data != NULL) {
3685 struct mdt_object *mo = lock->l_ast_data;
3688 rc = lu_env_init(&env, LCT_MD_THREAD);
3689 if (unlikely(rc != 0)) {
3690 CWARN("%s: lu_env initialization failed, object %p "DFID" is leaked!: rc = %d\n",
3692 PFID(mdt_object_fid(mo)), rc);
3696 if (lock->l_policy_data.l_inodebits.bits &
3697 (MDS_INODELOCK_XATTR | MDS_INODELOCK_UPDATE)) {
3698 rc = mo_invalidate(&env, mdt_object_child(mo));
3699 mo->mot_cache_attr = 0;
3701 mdt_object_put(&env, mo);
3713 int mdt_check_resent_lock(struct mdt_thread_info *info,
3714 struct mdt_object *mo,
3715 struct mdt_lock_handle *lhc)
3717 /* the lock might already be gotten in ldlm_handle_enqueue() */
3718 if (unlikely(lustre_handle_is_used(&lhc->mlh_reg_lh))) {
3719 struct ptlrpc_request *req = mdt_info_req(info);
3720 struct ldlm_lock *lock;
3722 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
3723 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
3725 /* Lock is pinned by ldlm_handle_enqueue0() as it is
3726 * a resend case, however, it could be already destroyed
3727 * due to client eviction or a raced cancel RPC. */
3728 LDLM_DEBUG_NOLOCK("Invalid lock handle %#llx",
3729 lhc->mlh_reg_lh.cookie);
3733 if (!fid_res_name_eq(mdt_object_fid(mo),
3734 &lock->l_resource->lr_name)) {
3735 CWARN("%s: Although resent, but still not "
3736 "get child lock:"DFID"\n",
3737 info->mti_exp->exp_obd->obd_name,
3738 PFID(mdt_object_fid(mo)));
3739 LDLM_LOCK_PUT(lock);
3742 LDLM_LOCK_PUT(lock);
3748 static void mdt_remote_object_lock_created_cb(struct ldlm_lock *lock)
3750 mdt_object_get(NULL, lock->l_ast_data);
3753 static int mdt_remote_object_lock_try(struct mdt_thread_info *mti,
3754 struct mdt_object *obj,
3755 struct lustre_handle *lh,
3756 enum ldlm_mode mode,
3757 union ldlm_policy_data *policy,
3758 struct ldlm_res_id *res_id,
3761 struct ldlm_enqueue_info *einfo = &mti->mti_remote_einfo;
3764 LASSERT(mdt_object_remote(obj));
3766 memset(einfo, 0, sizeof(*einfo));
3767 einfo->ei_type = LDLM_IBITS;
3768 einfo->ei_mode = mode;
3769 einfo->ei_cb_bl = mdt_remote_blocking_ast;
3770 einfo->ei_cb_cp = ldlm_completion_ast;
3771 einfo->ei_enq_slave = 0;
3772 einfo->ei_res_id = res_id;
3773 einfo->ei_req_slot = 1;
3776 * if we cache lock, couple lock with mdt_object, so that object
3777 * can be easily found in lock ASTs.
3779 einfo->ei_cbdata = obj;
3780 einfo->ei_cb_created = mdt_remote_object_lock_created_cb;
3783 rc = mo_object_lock(mti->mti_env, mdt_object_child(obj), lh, einfo,
3790 /* other components like LFSCK can use lockless access
3791 * and populate cache, so we better invalidate it
3793 if (policy->l_inodebits.bits &
3794 (MDS_INODELOCK_UPDATE | MDS_INODELOCK_XATTR))
3795 mo_invalidate(mti->mti_env, mdt_object_child(obj));
3801 * Helper function to take PDO and hash lock.
3803 * if \a pdo_lock is false, don't take PDO lock, this is case in rename.
3805 int mdt_object_pdo_lock(struct mdt_thread_info *info, struct mdt_object *obj,
3806 struct mdt_lock_handle *lh, const struct lu_name *name,
3807 enum ldlm_mode mode, bool pdo_lock)
3809 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
3810 union ldlm_policy_data *policy = &info->mti_policy;
3811 struct ldlm_res_id *res_id = &info->mti_res_id;
3813 * Do not use LDLM_FL_LOCAL_ONLY for parallel lock, it is never going to
3814 * be sent to client and we do not want it slowed down due to possible
3817 __u64 dlmflags = LDLM_FL_ATOMIC_CB;
3818 __u64 *cookie = NULL;
3823 /* check for exists after object is locked */
3824 if (!mdt_object_exists(obj))
3825 /* Non-existent object shouldn't have PDO lock */
3828 /* Non-dir object shouldn't have PDO lock */
3829 if (!S_ISDIR(lu_object_attr(&obj->mot_obj)))
3832 policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
3833 policy->l_inodebits.try_bits = 0;
3834 policy->l_inodebits.li_gid = 0;
3835 policy->l_inodebits.li_initiator_id = mdt_node_id(info->mti_mdt);
3836 fid_build_reg_res_name(mdt_object_fid(obj), res_id);
3838 cookie = &info->mti_exp->exp_handle.h_cookie;
3840 mdt_lock_pdo_init(lh, mode, name);
3841 mdt_lock_pdo_mode(info, obj, lh);
3842 if (lh->mlh_pdo_mode != LCK_NL) {
3844 if (mdt_object_remote(obj)) {
3845 rc = mdt_remote_object_lock_try(info, obj,
3846 &lh->mlh_pdo_lh, lh->mlh_pdo_mode,
3847 policy, res_id, false);
3848 lh->mlh_pdo_remote = 1;
3850 rc = mdt_fid_lock(info->mti_env, ns,
3851 &lh->mlh_pdo_lh, lh->mlh_pdo_mode,
3852 policy, res_id, dlmflags, cookie);
3855 mdt_object_unlock(info, obj, lh, 1);
3859 res_id->name[LUSTRE_RES_ID_HSH_OFF] = lh->mlh_pdo_hash;
3862 if (mdt_object_remote(obj))
3863 rc = mdt_remote_object_lock_try(info, obj, &lh->mlh_rreg_lh,
3864 lh->mlh_rreg_mode, policy, res_id, false);
3866 rc = mdt_fid_lock(info->mti_env, ns, &lh->mlh_reg_lh,
3867 lh->mlh_reg_mode, policy, res_id, dlmflags, cookie);
3869 mdt_object_unlock(info, obj, lh, 1);
3870 else if (CFS_FAIL_PRECHECK(OBD_FAIL_MDS_PDO_LOCK) &&
3871 lh->mlh_pdo_hash != 0 &&
3872 (lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX))
3873 CFS_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK, 15);
3878 int mdt_object_lock_internal(struct mdt_thread_info *info,
3879 struct mdt_object *obj, const struct lu_fid *fid,
3880 struct mdt_lock_handle *lh, __u64 *ibits,
3881 __u64 trybits, bool cache)
3883 union ldlm_policy_data *policy = &info->mti_policy;
3884 struct ldlm_res_id *res_id = &info->mti_res_id;
3885 struct lustre_handle *handle;
3888 policy->l_inodebits.bits = *ibits;
3889 policy->l_inodebits.try_bits = trybits;
3890 policy->l_inodebits.li_gid = lh->mlh_gid;
3891 policy->l_inodebits.li_initiator_id = mdt_node_id(info->mti_mdt);
3892 fid_build_reg_res_name(fid, res_id);
3894 if (obj && mdt_object_remote(obj)) {
3895 handle = &lh->mlh_rreg_lh;
3896 LASSERT(!lustre_handle_is_used(handle));
3897 LASSERT(lh->mlh_rreg_mode != LCK_MINMODE);
3898 LASSERT(lh->mlh_type != MDT_NUL_LOCK);
3899 rc = mdt_remote_object_lock_try(info, obj, handle,
3900 lh->mlh_rreg_mode, policy,
3903 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
3905 * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if
3906 * it is going to be sent to client. If it is -
3907 * mdt_intent_policy() path will fix it up and turn FL_LOCAL
3910 __u64 dlmflags = LDLM_FL_ATOMIC_CB | LDLM_FL_LOCAL_ONLY;
3911 __u64 *cookie = NULL;
3913 handle = &lh->mlh_reg_lh;
3914 LASSERT(!lustre_handle_is_used(handle));
3915 LASSERT(lh->mlh_reg_mode != LCK_MINMODE);
3916 LASSERT(lh->mlh_type != MDT_NUL_LOCK);
3918 /* Lease lock are granted with LDLM_FL_CANCEL_ON_BLOCK */
3919 if (lh->mlh_type == MDT_REG_LOCK &&
3920 lh->mlh_reg_mode == LCK_EX && *ibits == MDS_INODELOCK_OPEN)
3921 dlmflags |= LDLM_FL_CANCEL_ON_BLOCK;
3925 cookie = &info->mti_exp->exp_handle.h_cookie;
3927 rc = mdt_fid_lock(info->mti_env, ns, handle, lh->mlh_reg_mode,
3928 policy, res_id, dlmflags, cookie);
3930 mdt_object_unlock(info, obj, lh, 1);
3934 struct ldlm_lock *lock;
3936 /* Return successfully acquired bits to a caller */
3937 lock = ldlm_handle2lock(handle);
3939 *ibits = lock->l_policy_data.l_inodebits.bits;
3940 LDLM_LOCK_PUT(lock);
3947 * MDT object locking functions:
3948 * mdt_object_lock(): lock object, this is used in most places, and normally
3949 * lock ibits doesn't contain LOOKUP, unless the caller knows it's not
3951 * mdt_object_check_lock(): lock object with LOOKUP and other ibits, it needs
3952 * to check whether parent is on remote MDT, if so, take LOOKUP on parent
3953 * MDT separately, and then lock other ibits on child object.
3954 * mdt_parent_lock(): take parent UPDATE lock with specific mode, if parent is
3955 * local, take PDO lock by name hash, otherwise take regular lock.
3956 * mdt_object_stripes_lock(): lock object which should be local, and if it's a
3957 * striped directory, lock its stripes, this is called in operations which
3958 * modify both object and stripes.
3959 * mdt_object_lock_try(): lock object with trybits, the trybits contains
3960 * optional inode lock bits that can be granted. This is called by
3961 * getattr/open to fetch more inode lock bits to client, and is also called
3962 * by dir migration to lock link parent in non-block mode to avoid
3969 * this is used to lock object in most places, and normally lock ibits doesn't
3970 * contain LOOKUP, unless the caller knows it's not remote object.
3972 * \param info struct mdt_thread_info
3974 * \param lh lock handle
3975 * \param ibits MDS inode lock bits
3976 * \param mode lock mode
3978 * \retval 0 on success, -ev on error.
3980 int mdt_object_lock(struct mdt_thread_info *info, struct mdt_object *obj,
3981 struct mdt_lock_handle *lh, __u64 ibits,
3982 enum ldlm_mode mode)
3987 mdt_lock_reg_init(lh, mode);
3988 rc = mdt_object_lock_internal(info, obj, mdt_object_fid(obj), lh,
3994 * lock object with LOOKUP and other ibits
3996 * it will check whether parent and child are on different MDTs, if so, take
3997 * LOOKUP lock on parent MDT, and lock other ibits on child MDT, otherwise lock
3998 * all ibits on child MDT. Note, parent and child shouldn't be both on remote
3999 * MDTs, in which case specific lock function should be used, and it's in
4000 * rename and migrate only.
4002 * \param info struct mdt_thread_info
4003 * \param parent parent object
4004 * \param child child object
4005 * \param lh lock handle
4006 * \param ibits MDS inode lock bits
4007 * \param mode lock mode
4009 * \retval 0 on success, -ev on error.
4011 int mdt_object_check_lock(struct mdt_thread_info *info,
4012 struct mdt_object *parent, struct mdt_object *child,
4013 struct mdt_lock_handle *lh, __u64 ibits,
4014 enum ldlm_mode mode)
4019 /* if LOOKUP ibit is not set, use mdt_object_lock() */
4020 LASSERT(ibits & MDS_INODELOCK_LOOKUP);
4021 /* if only LOOKUP ibit is needed, use mdt_object_lookup_lock() */
4022 LASSERT(ibits != MDS_INODELOCK_LOOKUP);
4024 /* @parent and @child shouldn't both be on remote MDTs */
4025 LASSERT(!(mdt_object_remote(parent) && mdt_object_remote(child)));
4027 mdt_lock_reg_init(lh, mode);
4028 if (mdt_object_remote(parent) ^ mdt_object_remote(child)) {
4029 __u64 lookup_ibits = MDS_INODELOCK_LOOKUP;
4031 rc = mdt_object_lock_internal(info, parent,
4032 mdt_object_fid(child), lh,
4033 &lookup_ibits, 0, false);
4037 ibits &= ~MDS_INODELOCK_LOOKUP;
4040 rc = mdt_object_lock_internal(info, child, mdt_object_fid(child), lh,
4042 if (rc && !(ibits & MDS_INODELOCK_LOOKUP))
4043 mdt_object_unlock(info, NULL, lh, 1);
4049 * take parent UPDATE lock
4051 * if parent is local or mode is LCK_PW, take PDO lock, otherwise take regular
4054 * \param info struct mdt_thread_info
4055 * \param obj parent object
4056 * \param lh lock handle
4057 * \param lname child name
4058 * \param mode lock mode
4060 * \retval 0 on success, -ev on error.
4062 int mdt_parent_lock(struct mdt_thread_info *info, struct mdt_object *obj,
4063 struct mdt_lock_handle *lh, const struct lu_name *lname,
4064 enum ldlm_mode mode)
4069 LASSERT(obj && lname);
4070 LASSERT(mode == LCK_PW || mode == LCK_PR);
4071 if (mdt_object_remote(obj) && mode == LCK_PR) {
4072 __u64 ibits = MDS_INODELOCK_UPDATE;
4074 mdt_lock_reg_init(lh, mode);
4075 rc = mdt_object_lock_internal(info, obj, mdt_object_fid(obj),
4076 lh, &ibits, 0, false);
4078 rc = mdt_object_pdo_lock(info, obj, lh, lname, mode, true);
4084 * lock object with trybits
4086 * the trybits contains optional inode lock bits that can be granted. This is
4087 * called by getattr/open to fetch more inode lock bits to client, and is also
4088 * called by dir migration to lock link parent in non-block mode to avoid
4091 * \param info struct mdt_thread_info
4093 * \param lh lock handle
4094 * \param ibits MDS inode lock bits
4095 * \param trybits optional inode lock bits
4096 * \param mode lock mode
4098 * \retval 0 on success, -ev on error.
4100 int mdt_object_lock_try(struct mdt_thread_info *info, struct mdt_object *obj,
4101 struct mdt_lock_handle *lh, __u64 *ibits,
4102 __u64 trybits, enum ldlm_mode mode)
4104 bool trylock_only = *ibits == 0;
4108 LASSERT(!(*ibits & trybits));
4109 mdt_lock_reg_init(lh, mode);
4110 rc = mdt_object_lock_internal(info, obj, mdt_object_fid(obj), lh, ibits,
4112 if (rc && trylock_only) { /* clear error for try ibits lock only */
4113 LASSERT(*ibits == 0);
4120 * Helper function to take \a obj LOOKUP lock.
4122 * Both \a pobj and \a obj may be located on remote MDTs.
4124 int mdt_object_lookup_lock(struct mdt_thread_info *info,
4125 struct mdt_object *pobj, struct mdt_object *obj,
4126 struct mdt_lock_handle *lh, enum ldlm_mode mode)
4128 __u64 ibits = MDS_INODELOCK_LOOKUP;
4132 /* if @parent is NULL, it's on local MDT, and @child is remote,
4133 * this is case in getattr/unlink/open by name.
4135 LASSERT(ergo(!pobj, mdt_object_remote(obj)));
4136 mdt_lock_reg_init(lh, mode);
4137 rc = mdt_object_lock_internal(info, pobj, mdt_object_fid(obj), lh,
4143 * Save a lock within request object.
4145 * Keep the lock referenced until whether client ACK or transaction
4146 * commit happens or release the lock immediately depending on input
4147 * parameters. If COS is ON, a write lock is converted to COS lock
4150 * \param info thead info object
4151 * \param h lock handle
4152 * \param mode lock mode
4153 * \param decref force immediate lock releasing
4155 static void mdt_save_lock(struct mdt_thread_info *info, struct lustre_handle *h,
4156 enum ldlm_mode mode, int decref)
4158 struct tgt_session_info *tsi = info->mti_env->le_ses ?
4159 tgt_ses_info(info->mti_env) : NULL;
4163 if (lustre_handle_is_used(h)) {
4164 bool has_trans = tsi && tsi->tsi_has_trans;
4166 if (decref || !has_trans || !(mode & (LCK_PW | LCK_EX))) {
4167 mdt_fid_unlock(h, mode);
4169 struct mdt_device *mdt = info->mti_mdt;
4170 struct ldlm_lock *lock = ldlm_handle2lock(h);
4171 struct ptlrpc_request *req = mdt_info_req(info);
4172 bool no_ack = false;
4174 LASSERTF(lock != NULL, "no lock for cookie %#llx\n",
4177 /* there is no request if mdt_object_unlock() is called
4178 * from mdt_export_cleanup()->mdt_add_dirty_flag() */
4179 if (likely(req != NULL)) {
4180 LDLM_DEBUG(lock, "save lock request %p reply "
4181 "state %p transno %lld", req,
4182 req->rq_reply_state, req->rq_transno);
4183 if (mdt_cos_is_enabled(mdt)) {
4186 ldlm_lock_mode_downgrade(lock, mode);
4187 } else if (mdt_slc_is_enabled(mdt)) {
4189 if (mode != LCK_TXN) {
4191 ldlm_lock_mode_downgrade(lock,
4195 if (req->rq_export->exp_disconnected)
4196 mdt_fid_unlock(h, mode);
4198 ptlrpc_save_lock(req, h, mode, no_ack);
4200 mdt_fid_unlock(h, mode);
4203 if (mdt_is_lock_sync(lock)) {
4204 CDEBUG(D_HA, "sync_lock, do async commit\n");
4205 mdt_device_commit_async(info->mti_env, mdt);
4207 LDLM_LOCK_PUT(lock);
4216 * Save cross-MDT lock in uncommitted_slc_locks
4218 * Keep the lock referenced until transaction commit happens or release the lock
4219 * immediately depending on input parameters.
4221 * \param info thead info object
4222 * \param h lock handle
4223 * \param mode lock mode
4224 * \param decref force immediate lock releasing
4226 static void mdt_save_remote_lock(struct mdt_thread_info *info,
4227 struct mdt_object *o, struct lustre_handle *h,
4228 enum ldlm_mode mode, int decref)
4232 if (lustre_handle_is_used(h)) {
4233 struct ldlm_lock *lock = ldlm_handle2lock(h);
4234 struct ptlrpc_request *req = mdt_info_req(info);
4237 (lock->l_policy_data.l_inodebits.bits &
4238 (MDS_INODELOCK_XATTR | MDS_INODELOCK_UPDATE)))
4239 mo_invalidate(info->mti_env, mdt_object_child(o));
4241 if (decref || !req || !(mode & (LCK_PW | LCK_EX)) ||
4242 !tgt_ses_info(info->mti_env)->tsi_has_trans) {
4243 ldlm_lock_decref_and_cancel(h, mode);
4244 LDLM_LOCK_PUT(lock);
4246 tgt_save_slc_lock(&info->mti_mdt->mdt_lut, lock,
4248 ldlm_lock_decref(h, mode);
4257 * Unlock mdt object.
4259 * Immeditely release the regular lock and the PDO lock or save the
4260 * lock in request and keep them referenced until client ACK or
4261 * transaction commit.
4263 * \param info thread info object
4264 * \param o mdt object
4265 * \param lh mdt lock handle referencing regular and PDO locks
4266 * \param decref force immediate lock releasing
4268 * XXX o is not used and may be NULL, see hsm_cdt_request_completed().
4270 void mdt_object_unlock(struct mdt_thread_info *info, struct mdt_object *o,
4271 struct mdt_lock_handle *lh, int decref)
4275 if (lh->mlh_pdo_remote)
4276 mdt_save_remote_lock(info, o, &lh->mlh_pdo_lh,
4277 lh->mlh_pdo_mode, decref);
4279 mdt_save_lock(info, &lh->mlh_pdo_lh, lh->mlh_pdo_mode, decref);
4280 mdt_save_lock(info, &lh->mlh_reg_lh, lh->mlh_reg_mode, decref);
4281 mdt_save_remote_lock(info, o, &lh->mlh_rreg_lh, lh->mlh_rreg_mode,
4287 struct mdt_object *mdt_object_find_lock(struct mdt_thread_info *info,
4288 const struct lu_fid *f,
4289 struct mdt_lock_handle *lh,
4290 __u64 ibits, enum ldlm_mode mode)
4292 struct mdt_object *o;
4294 o = mdt_object_find(info->mti_env, info->mti_mdt, f);
4298 rc = mdt_object_lock(info, o, lh, ibits, mode);
4300 mdt_object_put(info->mti_env, o);
4307 void mdt_object_unlock_put(struct mdt_thread_info * info,
4308 struct mdt_object *o,
4309 struct mdt_lock_handle *lh,
4312 mdt_object_unlock(info, o, lh, decref);
4313 mdt_object_put(info->mti_env, o);
4317 * Generic code handling requests that have struct mdt_body passed in:
4319 * - extract mdt_body from request and save it in @info, if present;
4321 * - create lu_object, corresponding to the fid in mdt_body, and save it in
4324 * - if HAS_BODY flag is set for this request type check whether object
4325 * actually exists on storage (lu_object_exists()).
4328 static int mdt_body_unpack(struct mdt_thread_info *info,
4329 enum tgt_handler_flags flags)
4331 const struct mdt_body *body;
4332 struct mdt_object *obj;
4333 const struct lu_env *env;
4334 struct req_capsule *pill;
4338 env = info->mti_env;
4339 pill = info->mti_pill;
4341 body = info->mti_body = req_capsule_client_get(pill, &RMF_MDT_BODY);
4345 if (!(body->mbo_valid & OBD_MD_FLID))
4348 if (!fid_is_sane(&body->mbo_fid1)) {
4349 CERROR("Invalid fid: "DFID"\n", PFID(&body->mbo_fid1));
4353 obj = mdt_object_find(env, info->mti_mdt, &body->mbo_fid1);
4355 if ((flags & HAS_BODY) && !mdt_object_exists(obj)) {
4356 mdt_object_put(env, obj);
4359 info->mti_object = obj;
4368 static int mdt_unpack_req_pack_rep(struct mdt_thread_info *info,
4369 enum tgt_handler_flags flags)
4371 struct req_capsule *pill = info->mti_pill;
4376 if (req_capsule_has_field(pill, &RMF_MDT_BODY, RCL_CLIENT))
4377 rc = mdt_body_unpack(info, flags);
4381 if (rc == 0 && (flags & HAS_REPLY)) {
4383 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
4384 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER,
4385 req_capsule_ptlreq(pill) ?
4386 DEF_REP_MD_SIZE : MAX_MD_SIZE_OLD);
4388 if (req_capsule_has_field(pill, &RMF_LOGCOOKIES, RCL_SERVER))
4389 req_capsule_set_size(pill, &RMF_LOGCOOKIES,
4392 /* Set ACL reply buffer size as LUSTRE_POSIX_ACL_MAX_SIZE_OLD
4393 * by default. If the target object has more ACL entries, then
4394 * enlarge the buffer when necessary. */
4395 if (req_capsule_has_field(pill, &RMF_ACL, RCL_SERVER))
4396 req_capsule_set_size(pill, &RMF_ACL, RCL_SERVER,
4397 LUSTRE_POSIX_ACL_MAX_SIZE_OLD);
4399 mdt_preset_secctx_size(info);
4400 mdt_preset_encctx_size(info);
4402 rc = req_capsule_server_pack(pill);
4404 CWARN("%s: cannot pack response: rc = %d\n",
4405 mdt_obd_name(info->mti_mdt), rc);
4410 void mdt_thread_info_reset(struct mdt_thread_info *info)
4412 memset(&info->mti_attr, 0, sizeof(info->mti_attr));
4413 info->mti_body = NULL;
4414 info->mti_dlm_req = NULL;
4415 info->mti_cross_ref = 0;
4416 info->mti_opdata = 0;
4417 info->mti_big_lmm_used = 0;
4418 info->mti_big_acl_used = 0;
4419 info->mti_som_strict = 0;
4421 info->mti_spec.no_create = 0;
4422 info->mti_spec.sp_rm_entry = 0;
4423 info->mti_spec.sp_permitted = 0;
4425 info->mti_spec.u.sp_ea.eadata = NULL;
4426 info->mti_spec.u.sp_ea.eadatalen = 0;
4428 if (info->mti_batch_env && info->mti_object != NULL) {
4429 mdt_object_put(info->mti_env, info->mti_object);
4430 info->mti_object = NULL;
4435 * Initialize fields of struct mdt_thread_info. Other fields are left in
4436 * uninitialized state, because it's too expensive to zero out whole
4437 * mdt_thread_info (> 1K) on each request arrival.
4439 void mdt_thread_info_init(struct ptlrpc_request *req,
4440 struct mdt_thread_info *info)
4442 info->mti_pill = &req->rq_pill;
4444 /* mdt device: it can be NULL while CONNECT */
4445 if (req->rq_export) {
4446 info->mti_mdt = mdt_dev(req->rq_export->exp_obd->obd_lu_dev);
4447 info->mti_exp = req->rq_export;
4449 info->mti_mdt = NULL;
4450 info->mti_env = req->rq_svc_thread->t_env;
4451 info->mti_transno = lustre_msg_get_transno(req->rq_reqmsg);
4452 info->mti_big_buf = LU_BUF_NULL;
4453 info->mti_batch_env = 0;
4454 info->mti_object = NULL;
4456 mdt_thread_info_reset(info);
4459 void mdt_thread_info_fini(struct mdt_thread_info *info)
4463 if (info->mti_object != NULL) {
4464 mdt_object_put(info->mti_env, info->mti_object);
4465 info->mti_object = NULL;
4468 for (i = 0; i < ARRAY_SIZE(info->mti_lh); i++)
4469 mdt_lock_handle_assert(&info->mti_lh[i]);
4470 info->mti_env = NULL;
4471 info->mti_pill = NULL;
4472 info->mti_exp = NULL;
4473 info->mti_mdt = NULL;
4475 if (unlikely(info->mti_big_buf.lb_buf != NULL))
4476 lu_buf_free(&info->mti_big_buf);
4479 struct mdt_thread_info *tsi2mdt_info(struct tgt_session_info *tsi)
4481 struct mdt_thread_info *mti;
4483 mti = mdt_th_info(tsi->tsi_env);
4484 LASSERT(mti != NULL);
4486 mdt_thread_info_init(tgt_ses_req(tsi), mti);
4487 if (tsi->tsi_corpus != NULL) {
4488 mti->mti_object = mdt_obj(tsi->tsi_corpus);
4489 lu_object_get(tsi->tsi_corpus);
4491 mti->mti_body = tsi->tsi_mdt_body;
4492 mti->mti_dlm_req = tsi->tsi_dlm_req;
4497 static int mdt_tgt_connect(struct tgt_session_info *tsi)
4499 if (CFS_FAIL_CHECK(OBD_FAIL_TGT_DELAY_CONDITIONAL) &&
4501 tsi2mdt_info(tsi)->mti_mdt->mdt_seq_site.ss_node_id)
4502 schedule_timeout_uninterruptible(cfs_time_seconds(3));
4504 return tgt_connect(tsi);
4507 static int mdt_intent_glimpse(enum ldlm_intent_flags it_opc,
4508 struct mdt_thread_info *info,
4509 struct ldlm_lock **lockp, __u64 flags)
4511 return mdt_glimpse_enqueue(info, info->mti_mdt->mdt_namespace,
4514 static int mdt_intent_brw(enum ldlm_intent_flags it_opc,
4515 struct mdt_thread_info *info,
4516 struct ldlm_lock **lockp, __u64 flags)
4518 return mdt_brw_enqueue(info, info->mti_mdt->mdt_namespace,
4522 int mdt_intent_lock_replace(struct mdt_thread_info *info,
4523 struct ldlm_lock **lockp,
4524 struct mdt_lock_handle *lh,
4525 __u64 flags, int result)
4527 struct ptlrpc_request *req = mdt_info_req(info);
4528 struct ldlm_lock *lock = *lockp;
4529 struct ldlm_lock *new_lock;
4531 /* If possible resent found a lock, @lh is set to its handle */
4532 new_lock = ldlm_handle2lock_long(&lh->mlh_reg_lh, 0);
4534 if (new_lock == NULL) {
4535 if (flags & LDLM_FL_INTENT_ONLY) {
4537 } else if (flags & LDLM_FL_RESENT) {
4538 /* Lock is pinned by ldlm_handle_enqueue0() as it is a
4539 * resend case, however, it could be already destroyed
4540 * due to client eviction or a raced cancel RPC.
4542 LDLM_DEBUG_NOLOCK("Invalid lock handle %#llx\n",
4543 lh->mlh_reg_lh.cookie);
4546 CERROR("%s: Invalid lockh=%#llx flags=%#llx fid1="DFID" fid2="DFID": rc = %d\n",
4547 mdt_obd_name(info->mti_mdt),
4548 lh->mlh_reg_lh.cookie, flags,
4549 PFID(&info->mti_tmp_fid1),
4550 PFID(&info->mti_tmp_fid2), result);
4553 lh->mlh_reg_lh.cookie = 0;
4558 * If we've already given this lock to a client once, then we should
4559 * have no readers or writers. Otherwise, we should have one reader
4560 * _or_ writer ref (which will be zeroed below) before returning the
4563 if (new_lock->l_export == req->rq_export) {
4564 LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
4566 LASSERT(new_lock->l_export == NULL);
4567 LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
4572 if (new_lock->l_export == req->rq_export) {
4574 * Already gave this to the client, which means that we
4575 * reconstructed a reply.
4577 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
4580 LDLM_LOCK_RELEASE(new_lock);
4581 lh->mlh_reg_lh.cookie = 0;
4582 RETURN(ELDLM_LOCK_REPLACED);
4586 * Fixup the lock to be given to the client.
4588 lock_res_and_lock(new_lock);
4589 /* Zero new_lock->l_readers and new_lock->l_writers without triggering
4590 * possible blocking AST. */
4591 while (new_lock->l_readers > 0) {
4592 lu_ref_del(&new_lock->l_reference, "reader", new_lock);
4593 lu_ref_del(&new_lock->l_reference, "user", new_lock);
4594 new_lock->l_readers--;
4596 while (new_lock->l_writers > 0) {
4597 lu_ref_del(&new_lock->l_reference, "writer", new_lock);
4598 lu_ref_del(&new_lock->l_reference, "user", new_lock);
4599 new_lock->l_writers--;
4602 new_lock->l_export = class_export_lock_get(req->rq_export, new_lock);
4603 new_lock->l_blocking_ast = lock->l_blocking_ast;
4604 new_lock->l_completion_ast = lock->l_completion_ast;
4605 if (ldlm_has_dom(new_lock))
4606 new_lock->l_glimpse_ast = ldlm_server_glimpse_ast;
4607 new_lock->l_remote_handle = lock->l_remote_handle;
4608 new_lock->l_flags &= ~LDLM_FL_LOCAL;
4610 unlock_res_and_lock(new_lock);
4612 cfs_hash_add(new_lock->l_export->exp_lock_hash,
4613 &new_lock->l_remote_handle,
4614 &new_lock->l_exp_hash);
4616 LDLM_LOCK_RELEASE(new_lock);
4617 lh->mlh_reg_lh.cookie = 0;
4619 RETURN(ELDLM_LOCK_REPLACED);
4622 void mdt_intent_fixup_resent(struct mdt_thread_info *info,
4623 struct ldlm_lock *new_lock,
4624 struct mdt_lock_handle *lh, __u64 flags)
4626 struct ptlrpc_request *req = mdt_info_req(info);
4627 struct ldlm_request *dlmreq;
4629 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
4632 dlmreq = req_capsule_client_get(info->mti_pill, &RMF_DLM_REQ);
4634 /* Check if this is a resend case (MSG_RESENT is set on RPC) and a
4635 * lock was found by ldlm_handle_enqueue(); if so @lh must be
4637 if (flags & LDLM_FL_RESENT) {
4638 lh->mlh_reg_lh.cookie = new_lock->l_handle.h_cookie;
4639 lh->mlh_reg_mode = new_lock->l_granted_mode;
4641 LDLM_DEBUG(new_lock, "Restoring lock cookie");
4642 DEBUG_REQ(D_DLMTRACE, req, "restoring lock cookie %#llx",
4643 lh->mlh_reg_lh.cookie);
4648 * If the xid matches, then we know this is a resent request, and allow
4649 * it. (It's probably an OPEN, for which we don't send a lock.
4651 if (req_can_reconstruct(req, NULL) != 0)
4655 * This remote handle isn't enqueued, so we never received or processed
4656 * this request. Clear MSG_RESENT, because it can be handled like any
4657 * normal request now.
4659 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
4661 DEBUG_REQ(D_DLMTRACE, req, "no existing lock with rhandle %#llx",
4662 dlmreq->lock_handle[0].cookie);
4665 static int mdt_intent_getxattr(enum ldlm_intent_flags it_opc,
4666 struct mdt_thread_info *info,
4667 struct ldlm_lock **lockp,
4670 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
4671 struct ldlm_reply *ldlm_rep = NULL;
4676 * Initialize lhc->mlh_reg_lh either from a previously granted lock
4677 * (for the resend case) or a new lock. Below we will use it to
4678 * replace the original lock.
4680 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
4681 if (!lustre_handle_is_used(&lhc->mlh_reg_lh)) {
4682 rc = mdt_object_lock(info, info->mti_object, lhc,
4683 MDS_INODELOCK_XATTR, (*lockp)->l_req_mode);
4688 rc = mdt_getxattr(info);
4690 if (mdt_info_req(info)->rq_repmsg != NULL)
4691 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
4693 if (ldlm_rep == NULL ||
4694 CFS_FAIL_CHECK(OBD_FAIL_MDS_XATTR_REP)) {
4695 mdt_object_unlock(info, info->mti_object, lhc, 1);
4699 RETURN(err_serious(-EFAULT));
4702 ldlm_rep->lock_policy_res2 = clear_serious(rc);
4704 /* This is left for interop instead of adding a new interop flag.
4706 #if LUSTRE_VERSION_CODE > OBD_OCD_VERSION(3, 0, 0, 0)
4707 if (ldlm_rep->lock_policy_res2) {
4708 mdt_object_unlock(info, info->mti_object, lhc, 1);
4709 RETURN(ELDLM_LOCK_ABORTED);
4713 rc = mdt_intent_lock_replace(info, lockp, lhc, flags, rc);
4717 static int mdt_intent_getattr(enum ldlm_intent_flags it_opc,
4718 struct mdt_thread_info *info,
4719 struct ldlm_lock **lockp,
4722 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
4724 struct ldlm_reply *ldlm_rep;
4725 struct mdt_body *reqbody;
4726 struct mdt_body *repbody;
4730 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
4733 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
4736 info->mti_cross_ref = !!(reqbody->mbo_valid & OBD_MD_FLCROSSREF);
4737 repbody->mbo_eadatasize = 0;
4738 repbody->mbo_aclsize = 0;
4742 child_bits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
4745 child_bits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE |
4749 CERROR("%s: unsupported intent %#x\n",
4750 mdt_obd_name(info->mti_mdt), (unsigned int)it_opc);
4751 GOTO(out_shrink, rc = -EINVAL);
4754 rc = mdt_init_ucred(info, reqbody);
4756 GOTO(out_shrink, rc);
4758 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
4759 mdt_set_disposition(info, ldlm_rep, DISP_IT_EXECD);
4761 /* Get lock from request for possible resent case. */
4762 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
4764 rc = mdt_getattr_name_lock(info, lhc, child_bits, ldlm_rep);
4765 ldlm_rep->lock_policy_res2 = clear_serious(rc);
4767 if (mdt_get_disposition(ldlm_rep, DISP_LOOKUP_NEG) &&
4768 ldlm_rep->lock_policy_res2 != -ENOKEY)
4769 ldlm_rep->lock_policy_res2 = 0;
4770 if (!mdt_get_disposition(ldlm_rep, DISP_LOOKUP_POS) ||
4771 ldlm_rep->lock_policy_res2) {
4772 lhc->mlh_reg_lh.cookie = 0ull;
4773 GOTO(out_ucred, rc = ELDLM_LOCK_ABORTED);
4776 rc = mdt_intent_lock_replace(info, lockp, lhc, flags, rc);
4779 mdt_exit_ucred(info);
4781 mdt_client_compatibility(info);
4782 rc2 = mdt_fix_reply(info);
4788 static int mdt_intent_layout(enum ldlm_intent_flags it_opc,
4789 struct mdt_thread_info *info,
4790 struct ldlm_lock **lockp,
4793 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
4794 struct md_layout_change layout = { .mlc_opc = MD_LAYOUT_NOP };
4795 struct layout_intent *intent;
4796 struct ldlm_reply *ldlm_rep;
4797 struct lu_fid *fid = &info->mti_tmp_fid2;
4798 struct mdt_object *obj = NULL;
4799 int layout_size = 0;
4800 struct lu_buf *buf = &layout.mlc_buf;
4805 fid_extract_from_res_name(fid, &(*lockp)->l_resource->lr_name);
4807 intent = req_capsule_client_get(info->mti_pill, &RMF_LAYOUT_INTENT);
4811 CDEBUG(D_INFO, DFID "got layout change request from client: "
4812 "opc:%u flags:%#x extent "DEXT"\n",
4813 PFID(fid), intent->lai_opc, intent->lai_flags,
4814 PEXT(&intent->lai_extent));
4816 switch (intent->lai_opc) {
4817 case LAYOUT_INTENT_TRUNC:
4818 case LAYOUT_INTENT_WRITE:
4819 case LAYOUT_INTENT_PCCRO_SET:
4820 case LAYOUT_INTENT_PCCRO_CLEAR:
4821 layout.mlc_opc = MD_LAYOUT_WRITE;
4822 layout.mlc_intent = intent;
4824 case LAYOUT_INTENT_ACCESS:
4826 case LAYOUT_INTENT_READ:
4827 case LAYOUT_INTENT_GLIMPSE:
4828 case LAYOUT_INTENT_RELEASE:
4829 case LAYOUT_INTENT_RESTORE:
4830 CERROR("%s: Unsupported layout intent opc %d\n",
4831 mdt_obd_name(info->mti_mdt), intent->lai_opc);
4834 CERROR("%s: Unknown layout intent opc %d\n",
4835 mdt_obd_name(info->mti_mdt), intent->lai_opc);
4839 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
4841 RETURN(PTR_ERR(obj));
4843 if (mdt_object_exists(obj) && !mdt_object_remote(obj)) {
4844 /* if layout is going to be changed don't use the current EA
4845 * size but the maximum one. That buffer will be shrinked
4846 * to the actual size in req_capsule_shrink() before reply.
4848 if (layout.mlc_opc == MD_LAYOUT_WRITE) {
4849 layout_size = info->mti_mdt->mdt_max_mdsize;
4851 layout_size = mdt_attr_get_eabuf_size(info, obj);
4852 if (layout_size < 0)
4853 GOTO(out, rc = layout_size);
4855 if (layout_size > info->mti_mdt->mdt_max_mdsize)
4856 info->mti_mdt->mdt_max_mdsize = layout_size;
4858 CDEBUG(D_INFO, "%s: layout_size %d\n",
4859 mdt_obd_name(info->mti_mdt), layout_size);
4863 * set reply buffer size, so that ldlm_handle_enqueue0()->
4864 * ldlm_lvbo_fill() will fill the reply buffer with lovea.
4866 req_capsule_set_size(info->mti_pill, &RMF_DLM_LVB, RCL_SERVER,
4868 rc = req_capsule_server_pack(info->mti_pill);
4872 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
4874 GOTO(out, rc = -EPROTO);
4876 mdt_set_disposition(info, ldlm_rep, DISP_IT_EXECD);
4878 /* take lock in ldlm_lock_enqueue() for LAYOUT_INTENT_ACCESS */
4879 if (layout.mlc_opc == MD_LAYOUT_NOP)
4882 rc = mdt_check_resent(info, mdt_reconstruct_generic, lhc);
4886 DEBUG_REQ(D_INODE, mdt_info_req(info), "resent opt.");
4887 rc = lustre_msg_get_status(mdt_info_req(info)->rq_repmsg);
4893 if (unlikely(req_is_replay(mdt_info_req(info)))) {
4894 buf->lb_buf = req_capsule_client_get(info->mti_pill,
4896 buf->lb_len = req_capsule_get_size(info->mti_pill,
4897 &RMF_EADATA, RCL_CLIENT);
4899 * If it's a replay of layout write intent RPC, the client has
4900 * saved the extended lovea when it get reply then.
4902 if (buf->lb_len > 0)
4903 mdt_fix_lov_magic(info, buf->lb_buf);
4906 /* Get lock from request for possible resent case. */
4907 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
4908 (*lockp)->l_lvb_type = LVB_T_LAYOUT;
4911 * Instantiate some layout components, if @buf contains lovea, then it's
4912 * a replay of the layout intent write RPC.
4914 rc = mdt_layout_change(info, obj, lhc, &layout);
4915 ldlm_rep->lock_policy_res2 = clear_serious(rc);
4917 if (lustre_handle_is_used(&lhc->mlh_reg_lh)) {
4918 rc = mdt_intent_lock_replace(info, lockp, lhc, flags, rc);
4919 if (rc == ELDLM_LOCK_REPLACED &&
4920 (*lockp)->l_granted_mode == LCK_EX)
4921 ldlm_lock_mode_downgrade(*lockp, LCK_CR);
4926 mdt_object_put(info->mti_env, obj);
4930 static int mdt_intent_open(enum ldlm_intent_flags it_opc,
4931 struct mdt_thread_info *info,
4932 struct ldlm_lock **lockp,
4935 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
4936 struct ldlm_reply *rep = NULL;
4939 struct ptlrpc_request *req = mdt_info_req(info);
4941 static const struct req_format *intent_fmts[REINT_MAX] = {
4942 [REINT_CREATE] = &RQF_LDLM_INTENT_CREATE,
4943 [REINT_OPEN] = &RQF_LDLM_INTENT_OPEN
4948 opc = mdt_reint_opcode(mdt_info_req(info), intent_fmts);
4952 /* Get lock from request for possible resent case. */
4953 mdt_intent_fixup_resent(info, *lockp, lhc, flags);
4955 rc = mdt_reint_internal(info, lhc, opc);
4957 if (rc < 0 && lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
4958 DEBUG_REQ(D_ERROR, req, "Replay open failed with %d", rc);
4960 /* Check whether the reply has been packed successfully. */
4961 if (mdt_info_req(info)->rq_repmsg != NULL)
4962 rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
4967 RETURN(err_serious(-EFAULT));
4970 /* MDC expects this in any case */
4972 mdt_set_disposition(info, rep, DISP_LOOKUP_EXECD);
4974 /* the open lock or the lock for cross-ref object should be
4975 * returned to the client */
4976 if (lustre_handle_is_used(&lhc->mlh_reg_lh) &&
4977 (rc == 0 || rc == -MDT_EREMOTE_OPEN)) {
4978 rep->lock_policy_res2 = 0;
4979 rc = mdt_intent_lock_replace(info, lockp, lhc, flags, rc);
4983 rep->lock_policy_res2 = clear_serious(rc);
4985 if (rep->lock_policy_res2 == -ENOENT &&
4986 mdt_get_disposition(rep, DISP_LOOKUP_NEG) &&
4987 !mdt_get_disposition(rep, DISP_OPEN_CREATE))
4988 rep->lock_policy_res2 = 0;
4990 lhc->mlh_reg_lh.cookie = 0ull;
4991 if (rc == -ENOTCONN || rc == -ENODEV ||
4992 rc == -EOVERFLOW) { /**< if VBR failure then return error */
4994 * If it is the disconnect error (ENODEV & ENOCONN), the error
4995 * will be returned by rq_status, and client at ptlrpc layer
4996 * will detect this, then disconnect, reconnect the import
4997 * immediately, instead of impacting the following the rpc.
5002 * For other cases, the error will be returned by intent, and client
5003 * will retrieve the result from intent.
5005 RETURN(ELDLM_LOCK_ABORTED);
5008 static int mdt_intent_opc(enum ldlm_intent_flags it_opc,
5009 struct mdt_thread_info *info,
5010 struct ldlm_lock **lockp,
5011 u64 flags /* LDLM_FL_* */)
5013 struct req_capsule *pill = info->mti_pill;
5014 struct ptlrpc_request *req = mdt_info_req(info);
5015 const struct req_format *it_format;
5016 int (*it_handler)(enum ldlm_intent_flags,
5017 struct mdt_thread_info *,
5018 struct ldlm_lock **,
5020 enum tgt_handler_flags it_handler_flags = 0;
5021 struct ldlm_reply *rep;
5022 bool check_mdt_object = false;
5028 case IT_OPEN|IT_CREAT:
5030 * OCREAT is not a IS_MUTABLE request since the file may
5031 * already exist. We do the extra check of
5032 * OBD_CONNECT_RDONLY in mdt_reint_open() when we
5033 * really need to create the object.
5035 it_format = &RQF_LDLM_INTENT;
5036 it_handler = &mdt_intent_open;
5039 check_mdt_object = true;
5042 it_format = &RQF_LDLM_INTENT_GETATTR;
5043 it_handler = &mdt_intent_getattr;
5044 it_handler_flags = HAS_REPLY;
5047 check_mdt_object = true;
5048 it_format = &RQF_LDLM_INTENT_GETXATTR;
5049 it_handler = &mdt_intent_getxattr;
5050 it_handler_flags = HAS_BODY;
5053 it_format = &RQF_LDLM_INTENT_LAYOUT;
5054 it_handler = &mdt_intent_layout;
5057 it_format = &RQF_LDLM_INTENT;
5058 it_handler = &mdt_intent_glimpse;
5061 it_format = &RQF_LDLM_INTENT;
5062 it_handler = &mdt_intent_brw;
5064 case IT_QUOTA_DQACQ:
5065 case IT_QUOTA_CONN: {
5066 struct lu_device *qmt = info->mti_mdt->mdt_qmt_dev;
5069 RETURN(-EOPNOTSUPP);
5071 if (mdt_rdonly(req->rq_export))
5074 (*lockp)->l_lvb_type = LVB_T_LQUOTA;
5075 /* pass the request to quota master */
5076 rc = qmt_hdls.qmth_intent_policy(info->mti_env, qmt,
5077 mdt_info_req(info), lockp,
5082 CERROR("%s: unknown intent code %#x\n",
5083 mdt_obd_name(info->mti_mdt), it_opc);
5087 if (!info->mti_batch_env)
5088 req_capsule_extend(pill, it_format);
5090 rc = mdt_unpack_req_pack_rep(info, it_handler_flags);
5094 if (unlikely(info->mti_object == NULL && check_mdt_object))
5097 if (it_handler_flags & IS_MUTABLE && mdt_rdonly(req->rq_export))
5100 CFS_FAIL_TIMEOUT(OBD_FAIL_MDS_INTENT_DELAY, 10);
5102 /* execute policy */
5103 rc = (*it_handler)(it_opc, info, lockp, flags);
5105 /* Check whether the reply has been packed successfully. */
5106 if (info->mti_batch_env || req->rq_repmsg != NULL) {
5107 rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
5108 rep->lock_policy_res2 =
5109 ptlrpc_status_hton(rep->lock_policy_res2);
5115 static void mdt_ptlrpc_stats_update(struct ptlrpc_request *req,
5116 enum ldlm_intent_flags it_opc)
5118 struct lprocfs_stats *srv_stats = ptlrpc_req2svc(req)->srv_stats;
5120 /* update stats when IT code is known */
5121 if (srv_stats != NULL)
5122 lprocfs_counter_incr(srv_stats,
5123 PTLRPC_LAST_CNTR + (it_opc == IT_GLIMPSE ?
5124 LDLM_GLIMPSE_ENQUEUE : LDLM_IBITS_ENQUEUE));
5127 static int mdt_intent_policy(const struct lu_env *env,
5128 struct ldlm_namespace *ns,
5129 struct ldlm_lock **lockp,
5131 enum ldlm_mode mode,
5132 __u64 flags, void *data)
5134 struct tgt_session_info *tsi;
5135 struct mdt_thread_info *info;
5136 struct ptlrpc_request *req = req_cookie;
5137 struct ldlm_intent *it;
5138 struct req_capsule *pill;
5139 const struct ldlm_lock_desc *ldesc;
5144 LASSERT(req != NULL);
5146 tsi = tgt_ses_info(env);
5148 info = mdt_th_info(env);
5149 LASSERT(info != NULL);
5151 /* Check whether it is a sub request processing in a batch request */
5152 if (info->mti_batch_env) {
5153 pill = info->mti_pill;
5154 LASSERT(pill == &info->mti_sub_pill);
5156 info = tsi2mdt_info(tsi);
5157 pill = info->mti_pill;
5160 LASSERT(pill->rc_req == req);
5161 ldesc = &info->mti_dlm_req->lock_desc;
5163 if (info->mti_batch_env ||
5164 req->rq_reqmsg->lm_bufcount > DLM_INTENT_IT_OFF) {
5166 * For batch processing environment, the request format has
5169 if (!info->mti_batch_env)
5170 req_capsule_extend(pill, &RQF_LDLM_INTENT_BASIC);
5172 it = req_capsule_client_get(pill, &RMF_LDLM_INTENT);
5174 mdt_ptlrpc_stats_update(req, it->opc);
5175 rc = mdt_intent_opc(it->opc, info, lockp, flags);
5179 /* Lock without inodebits makes no sense and will oops
5180 * later in ldlm. Let's check it now to see if we have
5181 * ibits corrupted somewhere in mdt_intent_opc().
5182 * The case for client miss to set ibits has been
5183 * processed by others. */
5184 LASSERT(ergo(ldesc->l_resource.lr_type == LDLM_IBITS,
5185 ldesc->l_policy_data.l_inodebits.bits != 0));
5187 rc = err_serious(-EFAULT);
5189 } else if (ldesc->l_resource.lr_type == LDLM_IBITS &&
5190 ldesc->l_policy_data.l_inodebits.bits == MDS_INODELOCK_DOM) {
5191 struct ldlm_reply *rep;
5193 /* No intent was provided but INTENT flag is set along with
5194 * DOM bit, this is considered as GLIMPSE request.
5195 * This logic is common for MDT and OST glimpse
5197 mdt_ptlrpc_stats_update(req, IT_GLIMPSE);
5198 rc = mdt_glimpse_enqueue(info, ns, lockp, flags);
5199 /* Check whether the reply has been packed successfully. */
5200 if (req->rq_repmsg != NULL) {
5201 rep = req_capsule_server_get(info->mti_pill,
5203 rep->lock_policy_res2 =
5204 ptlrpc_status_hton(rep->lock_policy_res2);
5207 /* No intent was provided */
5208 req_capsule_set_size(pill, &RMF_DLM_LVB, RCL_SERVER, 0);
5209 rc = req_capsule_server_pack(pill);
5211 rc = err_serious(rc);
5214 if (!info->mti_batch_env)
5215 mdt_thread_info_fini(info);
5219 static void mdt_deregister_seq_exp(struct mdt_device *mdt)
5221 struct seq_server_site *ss = mdt_seq_site(mdt);
5223 if (ss->ss_node_id == 0)
5226 if (ss->ss_client_seq != NULL) {
5227 lustre_deregister_lwp_item(&ss->ss_client_seq->lcs_exp);
5228 ss->ss_client_seq->lcs_exp = NULL;
5231 if (ss->ss_server_fld != NULL) {
5232 lustre_deregister_lwp_item(&ss->ss_server_fld->lsf_control_exp);
5233 ss->ss_server_fld->lsf_control_exp = NULL;
5237 static void mdt_seq_fini_cli(struct mdt_device *mdt)
5239 struct seq_server_site *ss = mdt_seq_site(mdt);
5244 if (ss->ss_server_seq != NULL)
5245 seq_server_set_cli(NULL, ss->ss_server_seq, NULL);
5248 static int mdt_seq_fini(const struct lu_env *env, struct mdt_device *mdt)
5250 mdt_seq_fini_cli(mdt);
5251 mdt_deregister_seq_exp(mdt);
5253 return seq_site_fini(env, mdt_seq_site(mdt));
5257 * It will retrieve its FLDB entries from MDT0, and it only happens
5258 * when upgrading existent FS to 2.6 or when local FLDB is corrupted,
5259 * and it needs to refresh FLDB from the MDT0.
5261 static int mdt_register_lwp_callback(void *data)
5264 struct mdt_device *mdt = data;
5265 struct lu_server_fld *fld = mdt_seq_site(mdt)->ss_server_fld;
5269 LASSERT(mdt_seq_site(mdt)->ss_node_id != 0);
5271 rc = lu_env_init(&env, LCT_MD_THREAD);
5273 CERROR("%s: cannot init env: rc = %d\n", mdt_obd_name(mdt), rc);
5277 /* Allocate new sequence now to avoid creating local transaction
5278 * in the normal transaction process */
5279 rc = seq_server_check_and_alloc_super(&env,
5280 mdt_seq_site(mdt)->ss_server_seq);
5285 rc = fld_update_from_controller(&env, fld);
5287 CERROR("%s: cannot update controller: rc = %d\n",
5288 mdt_obd_name(mdt), rc);
5297 static int mdt_register_seq_exp(struct mdt_device *mdt)
5299 struct seq_server_site *ss = mdt_seq_site(mdt);
5300 char *lwp_name = NULL;
5303 if (ss->ss_node_id == 0)
5306 OBD_ALLOC(lwp_name, MAX_OBD_NAME);
5307 if (lwp_name == NULL)
5308 GOTO(out_free, rc = -ENOMEM);
5310 rc = tgt_name2lwp_name(mdt_obd_name(mdt), lwp_name, MAX_OBD_NAME, 0);
5314 rc = lustre_register_lwp_item(lwp_name, &ss->ss_client_seq->lcs_exp,
5319 rc = lustre_register_lwp_item(lwp_name,
5320 &ss->ss_server_fld->lsf_control_exp,
5321 mdt_register_lwp_callback, mdt);
5323 lustre_deregister_lwp_item(&ss->ss_client_seq->lcs_exp);
5324 ss->ss_client_seq->lcs_exp = NULL;
5328 if (lwp_name != NULL)
5329 OBD_FREE(lwp_name, MAX_OBD_NAME);
5335 * Init client sequence manager which is used by local MDS to talk to sequence
5336 * controller on remote node.
5338 static int mdt_seq_init_cli(const struct lu_env *env, struct mdt_device *mdt)
5340 struct seq_server_site *ss = mdt_seq_site(mdt);
5344 /* check if this is adding the first MDC and controller is not yet
5346 OBD_ALLOC_PTR(ss->ss_client_seq);
5347 if (ss->ss_client_seq == NULL)
5350 OBD_ALLOC(prefix, MAX_OBD_NAME + 5);
5351 if (prefix == NULL) {
5352 OBD_FREE_PTR(ss->ss_client_seq);
5353 ss->ss_client_seq = NULL;
5357 /* Note: seq_client_fini will be called in seq_site_fini */
5358 snprintf(prefix, MAX_OBD_NAME + 5, "ctl-%s", mdt_obd_name(mdt));
5359 seq_client_init(ss->ss_client_seq, NULL, LUSTRE_SEQ_METADATA,
5360 prefix, ss->ss_node_id == 0 ? ss->ss_control_seq :
5362 OBD_FREE(prefix, MAX_OBD_NAME + 5);
5364 RETURN(seq_server_set_cli(env, ss->ss_server_seq, ss->ss_client_seq));
5367 static int mdt_seq_init(const struct lu_env *env, struct mdt_device *mdt)
5369 struct seq_server_site *ss;
5373 ss = mdt_seq_site(mdt);
5374 /* init sequence controller server(MDT0) */
5375 if (ss->ss_node_id == 0) {
5376 OBD_ALLOC_PTR(ss->ss_control_seq);
5377 if (ss->ss_control_seq == NULL)
5380 rc = seq_server_init(env, ss->ss_control_seq, mdt->mdt_bottom,
5381 mdt_obd_name(mdt), LUSTRE_SEQ_CONTROLLER,
5384 GOTO(out_seq_fini, rc);
5387 /* Init normal sequence server */
5388 OBD_ALLOC_PTR(ss->ss_server_seq);
5389 if (ss->ss_server_seq == NULL)
5390 GOTO(out_seq_fini, rc = -ENOMEM);
5392 rc = seq_server_init(env, ss->ss_server_seq, mdt->mdt_bottom,
5393 mdt_obd_name(mdt), LUSTRE_SEQ_SERVER, ss);
5395 GOTO(out_seq_fini, rc);
5397 /* init seq client for seq server to talk to seq controller(MDT0) */
5398 rc = mdt_seq_init_cli(env, mdt);
5400 GOTO(out_seq_fini, rc);
5402 if (ss->ss_node_id != 0)
5403 /* register controller export through lwp */
5404 rc = mdt_register_seq_exp(mdt);
5409 mdt_seq_fini(env, mdt);
5417 static int mdt_fld_fini(const struct lu_env *env,
5418 struct mdt_device *m)
5420 struct seq_server_site *ss = mdt_seq_site(m);
5423 if (ss && ss->ss_server_fld) {
5424 fld_server_fini(env, ss->ss_server_fld);
5425 OBD_FREE_PTR(ss->ss_server_fld);
5426 ss->ss_server_fld = NULL;
5432 static int mdt_fld_init(const struct lu_env *env,
5434 struct mdt_device *m)
5436 struct seq_server_site *ss;
5440 ss = mdt_seq_site(m);
5442 OBD_ALLOC_PTR(ss->ss_server_fld);
5443 if (ss->ss_server_fld == NULL)
5444 RETURN(rc = -ENOMEM);
5446 rc = fld_server_init(env, ss->ss_server_fld, m->mdt_bottom, uuid,
5449 OBD_FREE_PTR(ss->ss_server_fld);
5450 ss->ss_server_fld = NULL;
5457 static void mdt_stack_pre_fini(const struct lu_env *env,
5458 struct mdt_device *m, struct lu_device *top)
5460 struct lustre_cfg_bufs *bufs;
5461 struct lustre_cfg *lcfg;
5462 struct mdt_thread_info *info;
5467 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
5468 LASSERT(info != NULL);
5470 bufs = &info->mti_u.bufs;
5472 LASSERT(m->mdt_child_exp);
5473 LASSERT(m->mdt_child_exp->exp_obd);
5475 /* process cleanup, pass mdt obd name to get obd umount flags */
5476 /* XXX: this is needed because all layers are referenced by
5477 * objects (some of them are pinned by osd, for example *
5478 * the proper solution should be a model where object used
5479 * by osd only doesn't have mdt/mdd slices -bzzz */
5480 lustre_cfg_bufs_reset(bufs, mdt_obd_name(m));
5481 lustre_cfg_bufs_set_string(bufs, 1, NULL);
5482 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
5485 lustre_cfg_init(lcfg, LCFG_PRE_CLEANUP, bufs);
5487 top->ld_ops->ldo_process_config(env, top, lcfg);
5488 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens));
5492 static void mdt_stack_fini(const struct lu_env *env,
5493 struct mdt_device *m, struct lu_device *top)
5495 struct obd_device *obd = mdt2obd_dev(m);
5496 struct lustre_cfg_bufs *bufs;
5497 struct lustre_cfg *lcfg;
5498 struct mdt_thread_info *info;
5502 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
5503 LASSERT(info != NULL);
5505 lu_dev_del_linkage(top->ld_site, top);
5507 lu_site_purge(env, top->ld_site, -1);
5509 bufs = &info->mti_u.bufs;
5510 /* process cleanup, pass mdt obd name to get obd umount flags */
5511 /* another purpose is to let all layers to release their objects */
5512 lustre_cfg_bufs_reset(bufs, mdt_obd_name(m));
5517 lustre_cfg_bufs_set_string(bufs, 1, flags);
5518 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
5521 lustre_cfg_init(lcfg, LCFG_CLEANUP, bufs);
5524 top->ld_ops->ldo_process_config(env, top, lcfg);
5525 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens));
5527 lu_site_purge(env, top->ld_site, -1);
5529 m->mdt_child = NULL;
5530 m->mdt_bottom = NULL;
5532 obd_disconnect(m->mdt_child_exp);
5533 m->mdt_child_exp = NULL;
5535 obd_disconnect(m->mdt_bottom_exp);
5536 m->mdt_child_exp = NULL;
5539 static int mdt_connect_to_next(const struct lu_env *env, struct mdt_device *m,
5540 const char *next, struct obd_export **exp)
5542 struct obd_connect_data *data = NULL;
5543 struct obd_device *obd;
5547 OBD_ALLOC_PTR(data);
5549 GOTO(out, rc = -ENOMEM);
5551 obd = class_name2obd(next);
5553 CERROR("%s: can't locate next device: %s\n",
5554 mdt_obd_name(m), next);
5555 GOTO(out, rc = -ENOTCONN);
5558 data->ocd_connect_flags = OBD_CONNECT_VERSION;
5559 data->ocd_version = LUSTRE_VERSION_CODE;
5561 rc = obd_connect(NULL, exp, obd, &obd->obd_uuid, data, NULL);
5563 CERROR("%s: cannot connect to next dev %s (%d)\n",
5564 mdt_obd_name(m), next, rc);
5574 static int mdt_stack_init(const struct lu_env *env, struct mdt_device *mdt,
5575 struct lustre_cfg *cfg)
5577 char *dev = lustre_cfg_string(cfg, 0);
5578 int rc, name_size, uuid_size;
5579 char *name, *uuid, *p;
5580 struct lustre_cfg_bufs *bufs;
5581 struct lustre_cfg *lcfg;
5582 struct obd_device *obd;
5583 struct lustre_profile *lprof;
5584 struct lu_site *site;
5587 /* in 1.8 we had the only device in the stack - MDS.
5588 * 2.0 introduces MDT, MDD, OSD; MDT starts others internally.
5589 * in 2.3 OSD is instantiated by obd_mount.c, so we need
5590 * to generate names and setup MDT, MDD. MDT will be using
5591 * generated name to connect to MDD. for MDD the next device
5592 * will be LOD with name taken from so called "profile" which
5593 * is generated by mount_option line
5595 * 1.8 MGS generates config. commands like this:
5596 * #06 (104)mount_option 0: 1:lustre-MDT0000 2:lustre-mdtlov
5597 * #08 (120)setup 0:lustre-MDT0000 1:dev 2:type 3:lustre-MDT0000
5598 * 2.0 MGS generates config. commands like this:
5599 * #07 (112)mount_option 0: 1:lustre-MDT0000 2:lustre-MDT0000-mdtlov
5600 * #08 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
5601 * 3:lustre-MDT0000-mdtlov 4:f
5603 * we generate MDD name from MDT one, just replacing T with D
5605 * after all the preparations, the logical equivalent will be
5606 * #01 (160)setup 0:lustre-MDD0000 1:lustre-MDD0000_UUID 2:0
5607 * 3:lustre-MDT0000-mdtlov 4:f
5608 * #02 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
5609 * 3:lustre-MDD0000 4:f
5611 * notice we build the stack from down to top: MDD first, then MDT */
5613 name_size = MAX_OBD_NAME;
5614 uuid_size = MAX_OBD_NAME;
5616 OBD_ALLOC(name, name_size);
5617 OBD_ALLOC(uuid, uuid_size);
5618 if (name == NULL || uuid == NULL)
5619 GOTO(cleanup_mem, rc = -ENOMEM);
5621 OBD_ALLOC_PTR(bufs);
5623 GOTO(cleanup_mem, rc = -ENOMEM);
5626 p = strstr(name, "-MDT");
5628 GOTO(free_bufs, rc = -ENOMEM);
5631 snprintf(uuid, MAX_OBD_NAME, "%s_UUID", name);
5633 lprof = class_get_profile(lustre_cfg_string(cfg, 0));
5634 if (lprof == NULL || lprof->lp_dt == NULL) {
5635 CERROR("can't find the profile: %s\n",
5636 lustre_cfg_string(cfg, 0));
5637 GOTO(free_bufs, rc = -EINVAL);
5640 lustre_cfg_bufs_reset(bufs, name);
5641 lustre_cfg_bufs_set_string(bufs, 1, LUSTRE_MDD_NAME);
5642 lustre_cfg_bufs_set_string(bufs, 2, uuid);
5643 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
5645 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
5647 GOTO(put_profile, rc = -ENOMEM);
5648 lustre_cfg_init(lcfg, LCFG_ATTACH, bufs);
5650 rc = class_attach(lcfg);
5652 GOTO(lcfg_cleanup, rc);
5654 obd = class_name2obd(name);
5656 CERROR("Can not find obd %s (%s in config)\n",
5657 MDD_OBD_NAME, lustre_cfg_string(cfg, 0));
5658 GOTO(lcfg_cleanup, rc = -EINVAL);
5661 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens));
5663 lustre_cfg_bufs_reset(bufs, name);
5664 lustre_cfg_bufs_set_string(bufs, 1, uuid);
5665 lustre_cfg_bufs_set_string(bufs, 2, dev);
5666 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
5668 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
5670 GOTO(class_detach, rc = -ENOMEM);
5671 lustre_cfg_init(lcfg, LCFG_SETUP, bufs);
5673 rc = class_setup(obd, lcfg);
5675 GOTO(class_detach, rc);
5677 /* connect to MDD we just setup */
5678 rc = mdt_connect_to_next(env, mdt, name, &mdt->mdt_child_exp);
5680 GOTO(class_detach, rc);
5682 site = mdt->mdt_child_exp->exp_obd->obd_lu_dev->ld_site;
5684 LASSERT(mdt_lu_site(mdt) == NULL);
5685 mdt->mdt_lu_dev.ld_site = site;
5686 site->ls_top_dev = &mdt->mdt_lu_dev;
5687 mdt->mdt_child = lu2md_dev(mdt->mdt_child_exp->exp_obd->obd_lu_dev);
5689 /* now connect to bottom OSD */
5690 snprintf(name, MAX_OBD_NAME, "%s-osd", dev);
5691 rc = mdt_connect_to_next(env, mdt, name, &mdt->mdt_bottom_exp);
5693 GOTO(class_detach, rc);
5695 lu2dt_dev(mdt->mdt_bottom_exp->exp_obd->obd_lu_dev);
5697 rc = lu_env_refill((struct lu_env *)env);
5699 CERROR("Failure to refill session: '%d'\n", rc);
5701 lu_dev_add_linkage(site, &mdt->mdt_lu_dev);
5706 class_detach(obd, lcfg);
5708 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens));
5710 class_put_profile(lprof);
5715 OBD_FREE(name, name_size);
5717 OBD_FREE(uuid, uuid_size);
5721 /* setup quota master target on MDT0 */
5722 static int mdt_quota_init(const struct lu_env *env, struct mdt_device *mdt,
5723 struct lustre_cfg *cfg)
5725 struct obd_device *obd;
5726 char *dev = lustre_cfg_string(cfg, 0);
5727 char *qmtname, *uuid, *p;
5728 struct lustre_cfg_bufs *bufs;
5729 struct lustre_cfg *lcfg;
5730 struct lustre_profile *lprof;
5731 struct obd_connect_data *data;
5735 LASSERT(mdt->mdt_qmt_exp == NULL);
5736 LASSERT(mdt->mdt_qmt_dev == NULL);
5738 /* quota master is on MDT0 only for now */
5739 if (mdt->mdt_seq_site.ss_node_id != 0)
5742 /* MGS generates config commands which look as follows:
5743 * #01 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
5744 * 3:lustre-MDT0000-mdtlov 4:f
5746 * We generate the QMT name from the MDT one, just replacing MD with QM
5747 * after all the preparations, the logical equivalent will be:
5748 * #01 (160)setup 0:lustre-QMT0000 1:lustre-QMT0000_UUID 2:0
5749 * 3:lustre-MDT0000-osd 4:f */
5750 OBD_ALLOC(qmtname, MAX_OBD_NAME);
5751 OBD_ALLOC(uuid, UUID_MAX);
5752 OBD_ALLOC_PTR(bufs);
5753 OBD_ALLOC_PTR(data);
5754 if (qmtname == NULL || uuid == NULL || bufs == NULL || data == NULL)
5755 GOTO(cleanup_mem, rc = -ENOMEM);
5757 strcpy(qmtname, dev);
5758 p = strstr(qmtname, "-MDT");
5760 GOTO(cleanup_mem, rc = -ENOMEM);
5761 /* replace MD with QM */
5765 snprintf(uuid, UUID_MAX, "%s_UUID", qmtname);
5767 lprof = class_get_profile(lustre_cfg_string(cfg, 0));
5768 if (lprof == NULL || lprof->lp_dt == NULL) {
5769 CERROR("can't find profile for %s\n",
5770 lustre_cfg_string(cfg, 0));
5771 GOTO(cleanup_mem, rc = -EINVAL);
5774 lustre_cfg_bufs_reset(bufs, qmtname);
5775 lustre_cfg_bufs_set_string(bufs, 1, LUSTRE_QMT_NAME);
5776 lustre_cfg_bufs_set_string(bufs, 2, uuid);
5777 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
5779 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
5781 GOTO(put_profile, rc = -ENOMEM);
5782 lustre_cfg_init(lcfg, LCFG_ATTACH, bufs);
5784 rc = class_attach(lcfg);
5786 GOTO(lcfg_cleanup, rc);
5788 obd = class_name2obd(qmtname);
5790 CERROR("Can not find obd %s (%s in config)\n", qmtname,
5791 lustre_cfg_string(cfg, 0));
5792 GOTO(lcfg_cleanup, rc = -EINVAL);
5795 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens));
5797 lustre_cfg_bufs_reset(bufs, qmtname);
5798 lustre_cfg_bufs_set_string(bufs, 1, uuid);
5799 lustre_cfg_bufs_set_string(bufs, 2, dev);
5801 /* for quota, the next device should be the OSD device */
5802 lustre_cfg_bufs_set_string(bufs, 3,
5803 mdt->mdt_bottom->dd_lu_dev.ld_obd->obd_name);
5805 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
5807 GOTO(class_detach, rc = -ENOMEM);
5808 lustre_cfg_init(lcfg, LCFG_SETUP, bufs);
5810 rc = class_setup(obd, lcfg);
5812 GOTO(class_detach, rc);
5814 mdt->mdt_qmt_dev = obd->obd_lu_dev;
5816 /* configure local quota objects */
5817 if (CFS_FAIL_CHECK(OBD_FAIL_QUOTA_INIT))
5820 rc = mdt->mdt_qmt_dev->ld_ops->ldo_prepare(env,
5824 GOTO(class_cleanup, rc);
5826 /* connect to quota master target */
5827 data->ocd_connect_flags = OBD_CONNECT_VERSION;
5828 data->ocd_version = LUSTRE_VERSION_CODE;
5829 rc = obd_connect(NULL, &mdt->mdt_qmt_exp, obd, &obd->obd_uuid,
5832 CERROR("cannot connect to quota master device %s (%d)\n",
5834 GOTO(class_cleanup, rc);
5840 class_manual_cleanup(obd);
5841 mdt->mdt_qmt_dev = NULL;
5842 GOTO(lcfg_cleanup, rc);
5846 class_detach(obd, lcfg);
5848 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens));
5850 class_put_profile(lprof);
5855 OBD_FREE(qmtname, MAX_OBD_NAME);
5857 OBD_FREE(uuid, UUID_MAX);
5863 /* Shutdown quota master target associated with mdt */
5864 static void mdt_quota_fini(const struct lu_env *env, struct mdt_device *mdt)
5868 if (mdt->mdt_qmt_exp == NULL)
5870 LASSERT(mdt->mdt_qmt_dev != NULL);
5872 /* the qmt automatically shuts down when the mdt disconnects */
5873 obd_disconnect(mdt->mdt_qmt_exp);
5874 mdt->mdt_qmt_exp = NULL;
5875 mdt->mdt_qmt_dev = NULL;
5879 /* mdt_getxattr() is used from mdt_intent_getxattr(), use this wrapper
5880 * for now. This will be removed along with converting rest of MDT code
5881 * to use tgt_session_info */
5882 static int mdt_tgt_getxattr(struct tgt_session_info *tsi)
5884 struct mdt_thread_info *info = tsi2mdt_info(tsi);
5887 if (unlikely(info->mti_object == NULL))
5890 rc = mdt_getxattr(info);
5892 mdt_thread_info_fini(info);
5896 static int mdt_llog_open(struct tgt_session_info *tsi)
5900 if (!mdt_changelog_allow(tsi2mdt_info(tsi)))
5901 RETURN(err_serious(-EACCES));
5903 RETURN(tgt_llog_open(tsi));
5906 #define OBD_FAIL_OST_READ_NET OBD_FAIL_OST_BRW_NET
5907 #define OBD_FAIL_OST_WRITE_NET OBD_FAIL_OST_BRW_NET
5908 #define OST_BRW_READ OST_READ
5909 #define OST_BRW_WRITE OST_WRITE
5911 static struct tgt_handler mdt_tgt_handlers[] = {
5912 TGT_RPC_HANDLER(MDS_FIRST_OPC,
5913 0, MDS_CONNECT, mdt_tgt_connect,
5914 &RQF_CONNECT, LUSTRE_OBD_VERSION),
5915 TGT_RPC_HANDLER(MDS_FIRST_OPC,
5916 0, MDS_DISCONNECT, tgt_disconnect,
5917 &RQF_MDS_DISCONNECT, LUSTRE_OBD_VERSION),
5918 TGT_RPC_HANDLER(MDS_FIRST_OPC,
5919 HAS_REPLY, MDS_SET_INFO, mdt_set_info,
5920 &RQF_MDT_SET_INFO, LUSTRE_MDS_VERSION),
5921 TGT_MDT_HDL(0, MDS_GET_INFO, mdt_get_info),
5922 TGT_MDT_HDL(HAS_REPLY, MDS_GET_ROOT, mdt_get_root),
5923 TGT_MDT_HDL(HAS_BODY, MDS_GETATTR, mdt_getattr),
5924 TGT_MDT_HDL(HAS_BODY | HAS_REPLY, MDS_GETATTR_NAME,
5926 TGT_MDT_HDL(HAS_BODY, MDS_GETXATTR, mdt_tgt_getxattr),
5927 TGT_MDT_HDL(HAS_REPLY, MDS_STATFS, mdt_statfs),
5928 TGT_MDT_HDL(IS_MUTABLE, MDS_REINT, mdt_reint),
5929 TGT_MDT_HDL(HAS_BODY, MDS_CLOSE, mdt_close),
5930 TGT_MDT_HDL(HAS_BODY | HAS_REPLY, MDS_READPAGE, mdt_readpage),
5931 TGT_MDT_HDL(HAS_BODY | HAS_REPLY, MDS_SYNC, mdt_sync),
5932 TGT_MDT_HDL(0, MDS_QUOTACTL, mdt_quotactl),
5933 TGT_MDT_HDL(HAS_BODY | HAS_REPLY | IS_MUTABLE, MDS_HSM_PROGRESS,
5935 TGT_MDT_HDL(HAS_BODY | HAS_REPLY | IS_MUTABLE, MDS_HSM_CT_REGISTER,
5936 mdt_hsm_ct_register),
5937 TGT_MDT_HDL(HAS_BODY | HAS_REPLY | IS_MUTABLE, MDS_HSM_CT_UNREGISTER,
5938 mdt_hsm_ct_unregister),
5939 TGT_MDT_HDL(HAS_BODY | HAS_REPLY, MDS_HSM_STATE_GET,
5941 TGT_MDT_HDL(HAS_BODY | HAS_REPLY | IS_MUTABLE, MDS_HSM_STATE_SET,
5943 TGT_MDT_HDL(HAS_BODY | HAS_REPLY, MDS_HSM_ACTION, mdt_hsm_action),
5944 TGT_MDT_HDL(HAS_BODY | HAS_REPLY, MDS_HSM_REQUEST,
5946 TGT_MDT_HDL(HAS_KEY | HAS_BODY | HAS_REPLY | IS_MUTABLE,
5949 TGT_MDT_HDL(IS_MUTABLE, MDS_RMFID, mdt_rmfid),
5950 TGT_MDT_HDL(IS_MUTABLE, MDS_BATCH, mdt_batch),
5953 static struct tgt_handler mdt_io_ops[] = {
5954 TGT_OST_HDL_HP(HAS_BODY | HAS_REPLY, OST_BRW_READ, tgt_brw_read,
5956 TGT_OST_HDL_HP(HAS_BODY | IS_MUTABLE, OST_BRW_WRITE, tgt_brw_write,
5958 TGT_OST_HDL_HP(HAS_BODY | HAS_REPLY | IS_MUTABLE,
5959 OST_PUNCH, mdt_punch_hdl,
5961 TGT_OST_HDL(HAS_BODY | HAS_REPLY, OST_SYNC, mdt_data_sync),
5962 TGT_OST_HDL(HAS_BODY | HAS_REPLY | IS_MUTABLE, OST_FALLOCATE,
5964 TGT_OST_HDL(HAS_BODY | HAS_REPLY, OST_SEEK, tgt_lseek),
5965 TGT_RPC_HANDLER(OST_FIRST_OPC,
5966 0, OST_SET_INFO, mdt_io_set_info,
5967 &RQF_OBD_SET_INFO, LUSTRE_OST_VERSION),
5970 static struct tgt_handler mdt_sec_ctx_ops[] = {
5971 TGT_SEC_HDL_VAR(0, SEC_CTX_INIT, mdt_sec_ctx_handle),
5972 TGT_SEC_HDL_VAR(0, SEC_CTX_INIT_CONT,mdt_sec_ctx_handle),
5973 TGT_SEC_HDL_VAR(0, SEC_CTX_FINI, mdt_sec_ctx_handle)
5976 static struct tgt_handler mdt_quota_ops[] = {
5977 TGT_QUOTA_HDL(HAS_REPLY, QUOTA_DQACQ, mdt_quota_dqacq),
5980 static struct tgt_handler mdt_llog_handlers[] = {
5981 TGT_LLOG_HDL(0, LLOG_ORIGIN_HANDLE_CREATE, mdt_llog_open),
5982 TGT_LLOG_HDL(0, LLOG_ORIGIN_HANDLE_NEXT_BLOCK, tgt_llog_next_block),
5983 TGT_LLOG_HDL(0, LLOG_ORIGIN_HANDLE_READ_HEADER, tgt_llog_read_header),
5984 TGT_LLOG_HDL(0, LLOG_ORIGIN_HANDLE_PREV_BLOCK, tgt_llog_prev_block),
5987 static struct tgt_opc_slice mdt_common_slice[] = {
5989 .tos_opc_start = MDS_FIRST_OPC,
5990 .tos_opc_end = MDS_LAST_OPC,
5991 .tos_hs = mdt_tgt_handlers
5994 .tos_opc_start = OBD_FIRST_OPC,
5995 .tos_opc_end = OBD_LAST_OPC,
5996 .tos_hs = tgt_obd_handlers
5999 .tos_opc_start = LDLM_FIRST_OPC,
6000 .tos_opc_end = LDLM_LAST_OPC,
6001 .tos_hs = tgt_dlm_handlers
6004 .tos_opc_start = SEC_FIRST_OPC,
6005 .tos_opc_end = SEC_LAST_OPC,
6006 .tos_hs = mdt_sec_ctx_ops
6009 .tos_opc_start = OUT_UPDATE_FIRST_OPC,
6010 .tos_opc_end = OUT_UPDATE_LAST_OPC,
6011 .tos_hs = tgt_out_handlers
6014 .tos_opc_start = FLD_FIRST_OPC,
6015 .tos_opc_end = FLD_LAST_OPC,
6016 .tos_hs = fld_handlers
6019 .tos_opc_start = SEQ_FIRST_OPC,
6020 .tos_opc_end = SEQ_LAST_OPC,
6021 .tos_hs = seq_handlers
6024 .tos_opc_start = QUOTA_DQACQ,
6025 .tos_opc_end = QUOTA_LAST_OPC,
6026 .tos_hs = mdt_quota_ops
6029 .tos_opc_start = LLOG_FIRST_OPC,
6030 .tos_opc_end = LLOG_LAST_OPC,
6031 .tos_hs = mdt_llog_handlers
6034 .tos_opc_start = LFSCK_FIRST_OPC,
6035 .tos_opc_end = LFSCK_LAST_OPC,
6036 .tos_hs = tgt_lfsck_handlers
6039 .tos_opc_start = OST_FIRST_OPC,
6040 .tos_opc_end = OST_LAST_OPC,
6041 .tos_hs = mdt_io_ops
6048 static void mdt_fini(const struct lu_env *env, struct mdt_device *m)
6050 struct md_device *next = m->mdt_child;
6051 struct lu_device *d = &m->mdt_lu_dev;
6052 struct obd_device *obd = mdt2obd_dev(m);
6053 struct lfsck_stop stop;
6056 stop.ls_status = LS_PAUSED;
6058 next->md_ops->mdo_iocontrol(env, next, OBD_IOC_STOP_LFSCK, 0, &stop);
6060 mdt_stack_pre_fini(env, m, md2lu_dev(m->mdt_child));
6062 mdt_restriper_stop(m);
6063 ping_evictor_stop();
6065 /* Remove the HSM /proc entry so the coordinator cannot be
6066 * restarted by a user while it's shutting down.
6068 mdt_hsm_cdt_stop(m);
6070 mdt_llog_ctxt_unclone(env, m, LLOG_AGENT_ORIG_CTXT);
6071 mdt_llog_ctxt_unclone(env, m, LLOG_CHANGELOG_ORIG_CTXT);
6073 if (m->mdt_namespace != NULL)
6074 ldlm_namespace_free_prior(m->mdt_namespace, NULL,
6075 d->ld_obd->obd_force);
6077 obd_exports_barrier(obd);
6078 obd_zombie_barrier();
6080 mdt_quota_fini(env, m);
6082 cfs_free_nidlist(&m->mdt_squash.rsi_nosquash_nids);
6084 /* Calling the cleanup functions in the same order as in the mdt_init0
6087 mdt_tunables_fini(m);
6089 target_recovery_fini(obd);
6090 upcall_cache_cleanup(m->mdt_identity_cache);
6091 m->mdt_identity_cache = NULL;
6093 tgt_fini(env, &m->mdt_lut);
6095 mdt_hsm_cdt_fini(m);
6097 if (m->mdt_los != NULL) {
6098 local_oid_storage_fini(env, m->mdt_los);
6102 if (m->mdt_namespace != NULL) {
6103 ldlm_namespace_free_post(m->mdt_namespace);
6104 d->ld_obd->obd_namespace = m->mdt_namespace = NULL;
6107 if (m->mdt_md_root != NULL) {
6108 mdt_object_put(env, m->mdt_md_root);
6109 m->mdt_md_root = NULL;
6112 mdt_seq_fini(env, m);
6114 mdt_fld_fini(env, m);
6119 mdt_stack_fini(env, m, md2lu_dev(m->mdt_child));
6121 LASSERT(atomic_read(&d->ld_ref) == 0);
6123 server_put_mount(mdt_obd_name(m), true);
6128 static int mdt_postrecov(const struct lu_env *, struct mdt_device *);
6130 static int mdt_init0(const struct lu_env *env, struct mdt_device *m,
6131 struct lu_device_type *ldt, struct lustre_cfg *cfg)
6133 const struct dt_device_param *dt_conf;
6134 struct mdt_thread_info *info;
6135 struct obd_device *obd;
6136 const char *dev = lustre_cfg_string(cfg, 0);
6137 const char *num = lustre_cfg_string(cfg, 2);
6138 struct tg_grants_data *tgd = &m->mdt_lut.lut_tgd;
6139 struct lustre_mount_info *lmi = NULL;
6140 struct lustre_sb_info *lsi;
6142 struct seq_server_site *ss_site;
6143 const char *identity_upcall = "NONE";
6144 struct md_device *next;
6151 lu_device_init(&m->mdt_lu_dev, ldt);
6153 * Environment (env) might be missing mdt_thread_key values at that
6154 * point, if device is allocated when mdt_thread_key is in QUIESCENT
6157 * Usually device allocation path doesn't use module key values, but
6158 * mdt has to do a lot of work here, so allocate key value.
6160 rc = lu_env_refill((struct lu_env *)env);
6164 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
6165 LASSERT(info != NULL);
6167 obd = class_name2obd(dev);
6168 LASSERT(obd != NULL);
6170 m->mdt_max_mdsize = MAX_MD_SIZE_OLD;
6171 m->mdt_evict_tgt_nids = 1;
6172 m->mdt_opts.mo_cos = MDT_COS_DEFAULT;
6174 lmi = server_get_mount(dev);
6176 CERROR("Cannot get mount info for %s!\n", dev);
6179 lsi = s2lsi(lmi->lmi_sb);
6180 LASSERT(lsi->lsi_lmd);
6181 /* CMD is supported only in IAM mode */
6183 rc = kstrtol(num, 10, &node_id);
6188 if (test_bit(LMD_FLG_SKIP_LFSCK, lsi->lsi_lmd->lmd_flags))
6189 m->mdt_skip_lfsck = 1;
6190 if (test_bit(LMD_FLG_NO_CREATE, lsi->lsi_lmd->lmd_flags))
6191 m->mdt_lut.lut_no_create = 1;
6194 /* Just try to get a DoM lock by default. Otherwise, having a group
6195 * lock granted, it may get blocked for a long time. */
6196 m->mdt_opts.mo_dom_lock = TRYLOCK_DOM_ON_OPEN;
6197 /* DoM files are read at open and data is packed in the reply */
6198 m->mdt_dom_read_open = 1;
6200 m->mdt_squash.rsi_uid = 0;
6201 m->mdt_squash.rsi_gid = 0;
6202 INIT_LIST_HEAD(&m->mdt_squash.rsi_nosquash_nids);
6203 spin_lock_init(&m->mdt_squash.rsi_lock);
6204 spin_lock_init(&m->mdt_lock);
6205 m->mdt_enable_chprojid_gid = 0;
6206 m->mdt_enable_dir_migration = 1;
6207 m->mdt_enable_dir_restripe = 0;
6208 m->mdt_enable_dir_auto_split = 0;
6209 m->mdt_enable_parallel_rename_dir = 1;
6210 m->mdt_enable_parallel_rename_file = 1;
6211 m->mdt_enable_remote_dir = 1;
6212 m->mdt_enable_remote_dir_gid = 0;
6213 m->mdt_enable_remote_rename = 1;
6214 m->mdt_enable_striped_dir = 1;
6215 m->mdt_enable_dmv_implicit_inherit = 1;
6216 m->mdt_dir_restripe_nsonly = 1;
6217 m->mdt_max_mod_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
6219 atomic_set(&m->mdt_mds_mds_conns, 0);
6220 atomic_set(&m->mdt_async_commit_count, 0);
6221 atomic_set(&m->mdt_dmv_old_client_count, 0);
6223 m->mdt_lu_dev.ld_ops = &mdt_lu_ops;
6224 m->mdt_lu_dev.ld_obd = obd;
6225 /* Set this lu_device to obd for error handling purposes. */
6226 obd->obd_lu_dev = &m->mdt_lu_dev;
6228 strncpy(m->mdt_job_xattr, XATTR_NAME_JOB_DEFAULT, XATTR_JOB_MAX_LEN);
6230 /* init the stack */
6231 rc = mdt_stack_init((struct lu_env *)env, m, cfg);
6233 CERROR("%s: Can't init device stack, rc %d\n",
6234 mdt_obd_name(m), rc);
6239 ss_site = mdt_seq_site(m);
6240 s->ld_seq_site = ss_site;
6243 /* set server index */
6244 ss_site->ss_node_id = node_id;
6246 /* failover is the default
6247 * FIXME: we do not failout mds0/mgs, which may cause some problems.
6248 * assumed whose ss_node_id == 0 XXX
6250 obd->obd_replayable = 1;
6251 /* No connection accepted until configurations will finish */
6252 obd->obd_no_conn = 1;
6254 if (cfg->lcfg_bufcount > 4 && LUSTRE_CFG_BUFLEN(cfg, 4) > 0) {
6255 char *str = lustre_cfg_string(cfg, 4);
6256 if (strchr(str, 'n')) {
6257 CWARN("%s: recovery disabled\n", mdt_obd_name(m));
6258 obd->obd_replayable = 0;
6262 rc = mdt_fld_init(env, mdt_obd_name(m), m);
6264 GOTO(err_fini_stack, rc);
6266 rc = mdt_seq_init(env, m);
6268 GOTO(err_fini_fld, rc);
6270 snprintf(info->mti_u.ns_name, sizeof(info->mti_u.ns_name), "%s-%s",
6271 LUSTRE_MDT_NAME, obd->obd_uuid.uuid);
6272 m->mdt_namespace = ldlm_namespace_new(obd, info->mti_u.ns_name,
6273 LDLM_NAMESPACE_SERVER,
6274 LDLM_NAMESPACE_GREEDY,
6276 if (IS_ERR(m->mdt_namespace)) {
6277 rc = PTR_ERR(m->mdt_namespace);
6278 CERROR("%s: unable to create server namespace: rc = %d\n",
6280 m->mdt_namespace = NULL;
6281 GOTO(err_fini_seq, rc);
6284 m->mdt_namespace->ns_lvbp = m;
6285 m->mdt_namespace->ns_lvbo = &mdt_lvbo;
6287 ldlm_register_intent(m->mdt_namespace, mdt_intent_policy);
6288 /* set obd_namespace for compatibility with old code */
6289 obd->obd_namespace = m->mdt_namespace;
6291 rc = tgt_init(env, &m->mdt_lut, obd, m->mdt_bottom, mdt_common_slice,
6292 OBD_FAIL_MDS_ALL_REQUEST_NET,
6293 OBD_FAIL_MDS_ALL_REPLY_NET);
6295 GOTO(err_free_ns, rc);
6297 /* Amount of available space excluded from granting and reserved
6298 * for metadata. It is a percentage of the total MDT size. */
6299 tgd->tgd_reserved_pcnt = 10;
6301 if (ONE_MB_BRW_SIZE < (1U << tgd->tgd_blockbits))
6302 m->mdt_brw_size = 1U << tgd->tgd_blockbits;
6304 m->mdt_brw_size = ONE_MB_BRW_SIZE;
6306 if (CFS_FAIL_CHECK(OBD_FAIL_MDS_FS_SETUP))
6307 GOTO(err_tgt, rc = -ENOENT);
6309 fid.f_seq = FID_SEQ_LOCAL_NAME;
6312 rc = local_oid_storage_init(env, m->mdt_bottom, &fid, &m->mdt_los);
6316 rc = mdt_hsm_cdt_init(m);
6318 CERROR("%s: error initializing coordinator, rc %d\n",
6319 mdt_obd_name(m), rc);
6320 GOTO(err_los_fini, rc);
6323 tgt_adapt_sptlrpc_conf(&m->mdt_lut);
6325 next = m->mdt_child;
6326 dt_conf = next->md_ops->mdo_dtconf_get(env, next);
6328 mntopts = dt_conf->ddp_mntopts;
6330 if (mntopts & MNTOPT_USERXATTR)
6331 m->mdt_opts.mo_user_xattr = 1;
6333 m->mdt_opts.mo_user_xattr = 0;
6335 m->mdt_max_ea_size = dt_conf->ddp_max_ea_size;
6337 if (mntopts & MNTOPT_ACL)
6338 m->mdt_opts.mo_acl = 1;
6340 m->mdt_opts.mo_acl = 0;
6342 m->mdt_enable_strict_som = 1;
6344 /* XXX: to support suppgid for ACL, we enable identity_upcall
6345 * by default, otherwise, maybe got unexpected -EACCESS. */
6346 if (m->mdt_opts.mo_acl)
6347 identity_upcall = MDT_IDENTITY_UPCALL_PATH;
6349 m->mdt_identity_cache = upcall_cache_init(mdt_obd_name(m),
6351 UC_IDCACHE_HASH_SIZE,
6352 1200, /* entry expire: 20 mn */
6353 30, /* acquire expire: 30 s */
6354 true, /* acquire can replay */
6355 &mdt_identity_upcall_cache_ops);
6356 if (IS_ERR(m->mdt_identity_cache)) {
6357 rc = PTR_ERR(m->mdt_identity_cache);
6358 m->mdt_identity_cache = NULL;
6359 GOTO(err_free_hsm, rc);
6362 rc = mdt_tunables_init(m, dev);
6364 CERROR("Can't init MDT lprocfs, rc %d\n", rc);
6365 GOTO(err_recovery, rc);
6368 rc = mdt_quota_init(env, m, cfg);
6370 GOTO(err_procfs, rc);
6372 m->mdt_ldlm_client = &mdt2obd_dev(m)->obd_ldlm_client;
6373 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
6374 "mdt_ldlm_client", m->mdt_ldlm_client);
6376 ping_evictor_start();
6378 /* recovery will be started upon mdt_prepare()
6379 * when the whole stack is complete and ready
6380 * to serve the requests */
6382 /* Reduce the initial timeout on an MDS because it doesn't need such
6383 * a long timeout as an OST does. Adaptive timeouts will adjust this
6384 * value appropriately. */
6385 if (ldlm_timeout == LDLM_TIMEOUT_DEFAULT)
6386 ldlm_timeout = MDS_LDLM_TIMEOUT_DEFAULT;
6388 if (test_bit(LMD_FLG_LOCAL_RECOV, lsi->lsi_lmd->lmd_flags))
6389 m->mdt_lut.lut_local_recovery = 1;
6391 rc = mdt_restriper_start(m);
6393 GOTO(err_ping_evictor, rc);
6398 ping_evictor_stop();
6400 mdt_tunables_fini(m);
6402 upcall_cache_cleanup(m->mdt_identity_cache);
6403 m->mdt_identity_cache = NULL;
6405 mdt_hsm_cdt_fini(m);
6407 local_oid_storage_fini(env, m->mdt_los);
6410 /* keep recoverable clients */
6412 target_recovery_fini(obd);
6413 obd_exports_barrier(obd);
6414 obd_zombie_barrier();
6415 tgt_fini(env, &m->mdt_lut);
6417 ldlm_namespace_free(m->mdt_namespace, NULL, 0);
6418 obd->obd_namespace = m->mdt_namespace = NULL;
6420 mdt_seq_fini(env, m);
6422 mdt_fld_fini(env, m);
6424 mdt_stack_fini(env, m, md2lu_dev(m->mdt_child));
6427 server_put_mount(dev, true);
6431 /* For interoperability, the left element is old parameter, the right one
6432 * is the new version of the parameter, if some parameter is deprecated,
6433 * the new version should be set as NULL. */
6434 static struct cfg_interop_param mdt_interop_param[] = {
6435 { "mdt.group_upcall", NULL },
6436 { "mdt.quota_type", NULL },
6437 { "mdd.quota_type", NULL },
6438 { "mdt.som", NULL },
6439 { "mdt.rootsquash", "mdt.root_squash" },
6440 { "mdt.nosquash_nid", "mdt.nosquash_nids" },
6444 /* used by MGS to process specific configurations */
6445 static int mdt_process_config(const struct lu_env *env,
6446 struct lu_device *d, struct lustre_cfg *cfg)
6448 struct mdt_device *m = mdt_dev(d);
6449 struct md_device *md_next = m->mdt_child;
6450 struct lu_device *next = md2lu_dev(md_next);
6454 switch (cfg->lcfg_command) {
6456 struct obd_device *obd = d->ld_obd;
6457 /* For interoperability */
6458 struct cfg_interop_param *ptr = NULL;
6459 struct lustre_cfg *old_cfg = NULL;
6463 param = lustre_cfg_string(cfg, 1);
6464 if (param == NULL) {
6465 CERROR("param is empty\n");
6470 ptr = class_find_old_param(param, mdt_interop_param);
6472 if (ptr->new_param == NULL) {
6474 CWARN("For interoperability, skip this %s."
6475 " It is obsolete.\n", ptr->old_param);
6479 CWARN("Found old param %s, changed it to %s.\n",
6480 ptr->old_param, ptr->new_param);
6483 cfg = lustre_cfg_rename(old_cfg, ptr->new_param);
6490 count = class_modify_config(cfg, PARAM_MDT,
6491 &obd->obd_kset.kobj);
6493 struct coordinator *cdt = &m->mdt_coordinator;
6495 /* is it an HSM var ? */
6496 count = class_modify_config(cfg, PARAM_HSM,
6497 &cdt->cdt_hsm_kobj);
6499 /* we don't understand; pass it on */
6500 rc = next->ld_ops->ldo_process_config(env, next,
6503 rc = count > 0 ? 0 : count;
6505 rc = count > 0 ? 0 : count;
6509 OBD_FREE(cfg, lustre_cfg_len(cfg->lcfg_bufcount,
6510 cfg->lcfg_buflens));
6514 /* others are passed further */
6515 rc = next->ld_ops->ldo_process_config(env, next, cfg);
6521 static struct lu_object *mdt_object_alloc(const struct lu_env *env,
6522 const struct lu_object_header *hdr,
6523 struct lu_device *d)
6525 struct mdt_object *mo;
6529 OBD_SLAB_ALLOC_PTR_GFP(mo, mdt_object_kmem, GFP_NOFS);
6531 struct lu_object *o;
6532 struct lu_object_header *h;
6535 h = &mo->mot_header;
6536 lu_object_header_init(h);
6537 lu_object_init(o, h, d);
6538 lu_object_add_top(h, o);
6539 o->lo_ops = &mdt_obj_ops;
6540 spin_lock_init(&mo->mot_write_lock);
6541 mutex_init(&mo->mot_som_mutex);
6542 mutex_init(&mo->mot_lov_mutex);
6543 init_rwsem(&mo->mot_dom_sem);
6544 init_rwsem(&mo->mot_open_sem);
6545 atomic_set(&mo->mot_open_count, 0);
6546 mo->mot_restripe_offset = 0;
6547 INIT_LIST_HEAD(&mo->mot_restripe_linkage);
6548 mo->mot_lsom_size = 0;
6549 mo->mot_lsom_blocks = 0;
6550 mo->mot_lsom_inited = false;
6556 static int mdt_object_init(const struct lu_env *env, struct lu_object *o,
6557 const struct lu_object_conf *unused)
6559 struct mdt_device *d = mdt_dev(o->lo_dev);
6560 struct lu_device *under;
6561 struct lu_object *below;
6565 CDEBUG(D_INFO, "object init, fid = "DFID"\n",
6566 PFID(lu_object_fid(o)));
6568 under = &d->mdt_child->md_lu_dev;
6569 below = under->ld_ops->ldo_object_alloc(env, o->lo_header, under);
6570 if (below != NULL) {
6571 lu_object_add(o, below);
6578 static void mdt_object_free_rcu(struct rcu_head *head)
6580 struct mdt_object *mo = container_of(head, struct mdt_object,
6581 mot_header.loh_rcu);
6583 kmem_cache_free(mdt_object_kmem, mo);
6586 static void mdt_object_free(const struct lu_env *env, struct lu_object *o)
6588 struct mdt_object *mo = mdt_obj(o);
6589 struct lu_object_header *h;
6593 CDEBUG(D_INFO, "object free, fid = "DFID"\n",
6594 PFID(lu_object_fid(o)));
6596 LASSERT(atomic_read(&mo->mot_open_count) == 0);
6597 LASSERT(atomic_read(&mo->mot_lease_count) == 0);
6600 lu_object_header_fini(h);
6601 OBD_FREE_PRE(mo, sizeof(*mo), "slab-freed");
6602 call_rcu(&mo->mot_header.loh_rcu, mdt_object_free_rcu);
6607 static int mdt_object_print(const struct lu_env *env, void *cookie,
6608 lu_printer_t p, const struct lu_object *o)
6610 struct mdt_object *mdto = mdt_obj((struct lu_object *)o);
6612 return (*p)(env, cookie,
6613 LUSTRE_MDT_NAME"-object@%p(%s %s, writecount=%d)",
6614 mdto, mdto->mot_lov_created ? "lov_created" : "",
6615 mdto->mot_cache_attr ? "cache_attr" : "",
6616 mdto->mot_write_count);
6619 static int mdt_prepare(const struct lu_env *env,
6620 struct lu_device *pdev,
6621 struct lu_device *cdev)
6623 struct mdt_device *mdt = mdt_dev(cdev);
6624 struct lu_device *next = &mdt->mdt_child->md_lu_dev;
6625 struct obd_device *obd = cdev->ld_obd;
6632 rc = next->ld_ops->ldo_prepare(env, cdev, next);
6636 rc = mdt_llog_ctxt_clone(env, mdt, LLOG_CHANGELOG_ORIG_CTXT);
6640 rc = mdt_llog_ctxt_clone(env, mdt, LLOG_AGENT_ORIG_CTXT);
6644 rc = lfsck_register_namespace(env, mdt->mdt_bottom, mdt->mdt_namespace);
6645 /* The LFSCK instance is registered just now, so it must be there when
6646 * register the namespace to such instance. */
6647 LASSERTF(rc == 0, "register namespace failed: rc = %d\n", rc);
6649 if (mdt->mdt_seq_site.ss_node_id == 0) {
6650 rc = mdt->mdt_child->md_ops->mdo_root_get(env, mdt->mdt_child,
6651 &mdt->mdt_md_root_fid);
6656 LASSERT(!test_bit(MDT_FL_CFGLOG, &mdt->mdt_state));
6658 target_recovery_init(&mdt->mdt_lut, tgt_request_handle);
6659 set_bit(MDT_FL_CFGLOG, &mdt->mdt_state);
6660 LASSERT(obd->obd_no_conn);
6661 spin_lock(&obd->obd_dev_lock);
6662 obd->obd_no_conn = 0;
6663 spin_unlock(&obd->obd_dev_lock);
6665 if (obd->obd_recovering == 0)
6666 mdt_postrecov(env, mdt);
6671 const struct lu_device_operations mdt_lu_ops = {
6672 .ldo_object_alloc = mdt_object_alloc,
6673 .ldo_process_config = mdt_process_config,
6674 .ldo_prepare = mdt_prepare,
6677 static const struct lu_object_operations mdt_obj_ops = {
6678 .loo_object_init = mdt_object_init,
6679 .loo_object_free = mdt_object_free,
6680 .loo_object_print = mdt_object_print
6683 static int mdt_obd_set_info_async(const struct lu_env *env,
6684 struct obd_export *exp,
6685 __u32 keylen, void *key,
6686 __u32 vallen, void *val,
6687 struct ptlrpc_request_set *set)
6693 if (KEY_IS(KEY_SPTLRPC_CONF)) {
6694 rc = tgt_adapt_sptlrpc_conf(class_exp2tgt(exp));
6701 static inline void mdt_enable_slc(struct mdt_device *mdt)
6703 if (mdt->mdt_lut.lut_sync_lock_cancel == SYNC_LOCK_CANCEL_NEVER)
6704 mdt->mdt_lut.lut_sync_lock_cancel = SYNC_LOCK_CANCEL_BLOCKING;
6707 static inline void mdt_disable_slc(struct mdt_device *mdt)
6709 if (mdt->mdt_lut.lut_sync_lock_cancel == SYNC_LOCK_CANCEL_BLOCKING)
6710 mdt->mdt_lut.lut_sync_lock_cancel = SYNC_LOCK_CANCEL_NEVER;
6714 * Match client and server connection feature flags.
6716 * Compute the compatibility flags for a connection request based on
6717 * features mutually supported by client and server.
6719 * The obd_export::exp_connect_data.ocd_connect_flags field in \a exp
6720 * must not be updated here, otherwise a partially initialized value may
6721 * be exposed. After the connection request is successfully processed,
6722 * the top-level MDT connect request handler atomically updates the export
6723 * connect flags from the obd_connect_data::ocd_connect_flags field of the
6724 * reply. \see mdt_connect().
6726 * Before 2.7.50 clients will send a struct obd_connect_data_v1 rather than a
6727 * full struct obd_connect_data. So care must be taken when accessing fields
6728 * that are not present in struct obd_connect_data_v1. See LU-16.
6730 * \param exp the obd_export associated with this client/target pair
6731 * \param mdt the target device for the connection
6732 * \param data stores data for this connect request
6735 * \retval -EPROTO \a data unexpectedly has zero obd_connect_data::ocd_brw_size
6736 * \retval -EBADE client and server feature requirements are incompatible
6738 static int mdt_connect_internal(const struct lu_env *env,
6739 struct obd_export *exp,
6740 struct mdt_device *mdt,
6741 struct obd_connect_data *data, bool reconnect)
6743 const char *obd_name = mdt_obd_name(mdt);
6744 LASSERT(data != NULL);
6746 data->ocd_connect_flags &= MDT_CONNECT_SUPPORTED;
6748 if (mdt->mdt_bottom->dd_rdonly &&
6749 !(data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
6750 !(data->ocd_connect_flags & OBD_CONNECT_RDONLY))
6753 if (data->ocd_connect_flags & OBD_CONNECT_FLAGS2)
6754 data->ocd_connect_flags2 &= MDT_CONNECT_SUPPORTED2;
6756 data->ocd_ibits_known &= MDS_INODELOCK_FULL;
6758 if (!mdt->mdt_opts.mo_acl)
6759 data->ocd_connect_flags &= ~OBD_CONNECT_ACL;
6761 if (!mdt->mdt_opts.mo_user_xattr)
6762 data->ocd_connect_flags &= ~OBD_CONNECT_XATTR;
6764 if (OCD_HAS_FLAG(data, BRW_SIZE)) {
6765 data->ocd_brw_size = min(data->ocd_brw_size,
6767 if (data->ocd_brw_size == 0) {
6768 CERROR("%s: cli %s/%p ocd_connect_flags: %#llx "
6769 "ocd_version: %x ocd_grant: %d ocd_index: %u "
6770 "ocd_brw_size unexpectedly zero, network data "
6771 "corruption? Refusing to connect this client\n",
6772 obd_name, exp->exp_client_uuid.uuid,
6773 exp, data->ocd_connect_flags, data->ocd_version,
6774 data->ocd_grant, data->ocd_index);
6779 if (OCD_HAS_FLAG(data, GRANT_PARAM)) {
6780 struct dt_device_param *ddp = &mdt->mdt_lut.lut_dt_conf;
6782 /* client is reporting its page size, for future use */
6783 exp->exp_target_data.ted_pagebits = data->ocd_grant_blkbits;
6784 data->ocd_grant_blkbits = mdt->mdt_lut.lut_tgd.tgd_blockbits;
6785 /* ddp_inodespace may not be power-of-two value, eg. for ldiskfs
6786 * it's LDISKFS_DIR_REC_LEN(20) = 28. */
6787 data->ocd_grant_inobits = fls(ddp->ddp_inodespace - 1);
6788 /* ocd_grant_tax_kb is in 1K byte blocks */
6789 data->ocd_grant_tax_kb = ddp->ddp_extent_tax >> 10;
6790 data->ocd_grant_max_blks = ddp->ddp_max_extent_blks;
6793 /* Save connect_data we have so far because tgt_grant_connect()
6794 * uses it to calculate grant, and we want to save the client
6795 * version before it is overwritten by LUSTRE_VERSION_CODE. */
6796 exp->exp_connect_data = *data;
6797 if (OCD_HAS_FLAG(data, GRANT))
6798 tgt_grant_connect(env, exp, data, !reconnect);
6800 if (OCD_HAS_FLAG(data, MAXBYTES))
6801 data->ocd_maxbytes = mdt->mdt_lut.lut_dt_conf.ddp_maxbytes;
6803 /* NB: Disregard the rule against updating
6804 * exp_connect_data.ocd_connect_flags in this case, since
6805 * tgt_client_new() needs to know if this is a lightweight
6806 * connection, and it is safe to expose this flag before
6807 * connection processing completes. */
6808 if (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) {
6809 spin_lock(&exp->exp_lock);
6810 *exp_connect_flags_ptr(exp) |= OBD_CONNECT_LIGHTWEIGHT;
6811 spin_unlock(&exp->exp_lock);
6814 data->ocd_version = LUSTRE_VERSION_CODE;
6816 if ((data->ocd_connect_flags & OBD_CONNECT_FID) == 0) {
6817 CWARN("%s: MDS requires FID support, but client not\n",
6822 if (OCD_HAS_FLAG(data, PINGLESS)) {
6823 if (ptlrpc_pinger_suppress_pings()) {
6824 spin_lock(&exp->exp_obd->obd_dev_lock);
6825 list_del_init(&exp->exp_obd_chain_timed);
6826 spin_unlock(&exp->exp_obd->obd_dev_lock);
6828 data->ocd_connect_flags &= ~OBD_CONNECT_PINGLESS;
6832 data->ocd_max_easize = mdt->mdt_max_ea_size;
6834 /* NB: Disregard the rule against updating
6835 * exp_connect_data.ocd_connect_flags in this case, since
6836 * tgt_client_new() needs to know if this is client supports
6837 * multiple modify RPCs, and it is safe to expose this flag before
6838 * connection processing completes. */
6839 if (data->ocd_connect_flags & OBD_CONNECT_MULTIMODRPCS) {
6840 if (mdt_max_mod_rpcs_changed(mdt))
6841 /* The new mdt.*.max_mod_rpcs_in_flight parameter
6842 * has not changed since initialization, but the
6843 * deprecated module parameter was changed,
6844 * so use that instead.
6846 data->ocd_maxmodrpcs = max_mod_rpcs_per_client;
6848 data->ocd_maxmodrpcs = mdt->mdt_max_mod_rpcs_in_flight;
6849 spin_lock(&exp->exp_lock);
6850 *exp_connect_flags_ptr(exp) |= OBD_CONNECT_MULTIMODRPCS;
6851 spin_unlock(&exp->exp_lock);
6854 if (OCD_HAS_FLAG(data, CKSUM)) {
6855 __u32 cksum_types = data->ocd_cksum_types;
6857 tgt_mask_cksum_types(&mdt->mdt_lut, &data->ocd_cksum_types);
6859 if (unlikely(data->ocd_cksum_types == 0)) {
6860 CERROR("%s: Connect with checksum support but no "
6861 "ocd_cksum_types is set\n",
6862 exp->exp_obd->obd_name);
6866 CDEBUG(D_RPCTRACE, "%s: cli %s supports cksum type %x, return "
6867 "%x\n", exp->exp_obd->obd_name, obd_export_nid2str(exp),
6868 cksum_types, data->ocd_cksum_types);
6870 /* This client does not support OBD_CONNECT_CKSUM
6871 * fall back to CRC32 */
6872 CDEBUG(D_RPCTRACE, "%s: cli %s does not support "
6873 "OBD_CONNECT_CKSUM, CRC32 will be used\n",
6874 exp->exp_obd->obd_name, obd_export_nid2str(exp));
6877 if ((data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
6878 !(data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT)) {
6879 atomic_inc(&mdt->mdt_mds_mds_conns);
6880 mdt_enable_slc(mdt);
6883 if (!mdt->mdt_lut.lut_dt_conf.ddp_has_lseek_data_hole)
6884 data->ocd_connect_flags2 &= ~OBD_CONNECT2_LSEEK;
6886 if (!OCD_HAS_FLAG(data, MDS_MDS) && !OCD_HAS_FLAG(data, LIGHTWEIGHT) &&
6887 !OCD_HAS_FLAG2(data, DMV_IMP_INHERIT)) {
6888 atomic_inc(&mdt->mdt_dmv_old_client_count);
6889 mdt->mdt_enable_dmv_implicit_inherit = 0;
6895 static int mdt_ctxt_add_dirty_flag(struct lu_env *env,
6896 struct mdt_thread_info *info,
6897 struct mdt_file_data *mfd)
6899 struct lu_context ses;
6903 rc = lu_context_init(&ses, LCT_SERVER_SESSION);
6908 lu_context_enter(&ses);
6910 mdt_ucred(info)->uc_valid = UCRED_OLD;
6911 /* do not let rbac interfere with dirty flag internal system event */
6912 mdt_ucred(info)->uc_rbac_file_perms = 1;
6913 mdt_ucred(info)->uc_rbac_dne_ops = 1;
6914 mdt_ucred(info)->uc_rbac_quota_ops = 1;
6915 mdt_ucred(info)->uc_rbac_byfid_ops = 1;
6916 mdt_ucred(info)->uc_rbac_chlg_ops = 1;
6917 mdt_ucred(info)->uc_rbac_fscrypt_admin = 1;
6918 rc = mdt_add_dirty_flag(info, mfd->mfd_object, &info->mti_attr);
6920 lu_context_exit(&ses);
6921 lu_context_fini(&ses);
6927 static int mdt_export_cleanup(struct obd_export *exp)
6929 LIST_HEAD(closing_list);
6930 struct mdt_export_data *med = &exp->exp_mdt_data;
6931 struct obd_device *obd = exp->exp_obd;
6932 struct mdt_device *mdt;
6933 struct mdt_thread_info *info;
6935 struct mdt_file_data *mfd, *n;
6939 spin_lock(&med->med_open_lock);
6940 while (!list_empty(&med->med_open_head)) {
6941 struct list_head *tmp = med->med_open_head.next;
6942 mfd = list_entry(tmp, struct mdt_file_data, mfd_list);
6944 /* Remove mfd handle so it can't be found again.
6945 * We are consuming the mfd_list reference here. */
6946 class_handle_unhash(&mfd->mfd_open_handle);
6947 list_move_tail(&mfd->mfd_list, &closing_list);
6949 spin_unlock(&med->med_open_lock);
6950 mdt = mdt_dev(obd->obd_lu_dev);
6951 LASSERT(mdt != NULL);
6953 rc = lu_env_init(&env, LCT_MD_THREAD);
6957 info = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
6958 LASSERT(info != NULL);
6959 memset(info, 0, sizeof *info);
6960 info->mti_env = &env;
6961 info->mti_mdt = mdt;
6962 info->mti_exp = exp;
6964 if (!list_empty(&closing_list)) {
6965 struct md_attr *ma = &info->mti_attr;
6967 /* Close any open files (which may also cause orphan
6969 list_for_each_entry_safe(mfd, n, &closing_list, mfd_list) {
6970 list_del_init(&mfd->mfd_list);
6971 ma->ma_need = ma->ma_valid = 0;
6973 /* This file is being closed due to an eviction, it
6974 * could have been modified and now dirty regarding to
6975 * HSM archive, check this!
6976 * The logic here is to mark a file dirty if there's a
6977 * chance it was dirtied before the client was evicted,
6978 * so that we don't have to wait for a release attempt
6979 * before finding out the file was actually dirty and
6980 * fail the release. Aggressively marking it dirty here
6981 * will cause the policy engine to attempt to
6982 * re-archive it; when rearchiving, we can compare the
6983 * current version to the HSM data_version and make the
6984 * archive request into a noop if it's not actually
6987 if (mfd->mfd_open_flags & MDS_FMODE_WRITE)
6988 rc = mdt_ctxt_add_dirty_flag(&env, info, mfd);
6990 /* Don't unlink orphan on failover umount, LU-184 */
6991 if (exp->exp_flags & OBD_OPT_FAILOVER ||
6992 exp->exp_obd->obd_stopping) {
6993 ma->ma_valid = MA_FLAGS;
6994 ma->ma_attr_flags |= MDS_KEEP_ORPHAN;
6996 ma->ma_valid |= MA_FORCE_LOG;
6997 mdt_mfd_close(info, mfd);
7000 info->mti_mdt = NULL;
7001 /* cleanup client slot early */
7002 /* Do not erase record for recoverable client. */
7003 if (!(exp->exp_flags & OBD_OPT_FAILOVER) || exp->exp_failed)
7004 tgt_client_del(&env, exp);
7010 static int mdt_obd_disconnect(struct obd_export *exp)
7012 struct obd_connect_data *data = &exp->exp_connect_data;
7013 struct mdt_device *mdt = mdt_dev(exp->exp_obd->obd_lu_dev);
7019 class_export_get(exp);
7021 if (!(exp->exp_flags & OBD_OPT_FORCE))
7022 tgt_grant_sanity_check(exp->exp_obd, __func__);
7024 if (OCD_HAS_FLAG(data, MDS_MDS) && !OCD_HAS_FLAG(data, LIGHTWEIGHT) &&
7025 atomic_dec_and_test(&mdt->mdt_mds_mds_conns))
7026 mdt_disable_slc(mdt);
7028 if (!OCD_HAS_FLAG(data, MDS_MDS) && !OCD_HAS_FLAG(data, LIGHTWEIGHT) &&
7029 !OCD_HAS_FLAG2(data, DMV_IMP_INHERIT) &&
7030 atomic_dec_and_test(&mdt->mdt_dmv_old_client_count))
7031 mdt->mdt_enable_dmv_implicit_inherit = 1;
7033 rc = server_disconnect_export(exp);
7035 CDEBUG(D_IOCTL, "server disconnect error: rc = %d\n", rc);
7037 tgt_grant_discard(exp);
7039 rc = mdt_export_cleanup(exp);
7040 nodemap_del_member(exp);
7041 class_export_put(exp);
7045 /* mds_connect copy */
7046 static int mdt_obd_connect(const struct lu_env *env,
7047 struct obd_export **exp, struct obd_device *obd,
7048 struct obd_uuid *cluuid,
7049 struct obd_connect_data *data,
7052 struct obd_export *lexp;
7053 struct lustre_handle conn = { 0 };
7054 struct mdt_device *mdt;
7056 struct lnet_nid *client_nid = localdata;
7059 LASSERT(env != NULL);
7060 LASSERT(data != NULL);
7062 if (!exp || !obd || !cluuid)
7065 mdt = mdt_dev(obd->obd_lu_dev);
7068 * first, check whether the stack is ready to handle requests
7069 * XXX: probably not very appropriate method is used now
7070 * at some point we should find a better one
7072 if (!test_bit(MDT_FL_SYNCED, &mdt->mdt_state) &&
7073 !(data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) &&
7074 !(data->ocd_connect_flags & OBD_CONNECT_MDS_MDS)) {
7075 rc = obd_get_info(env, mdt->mdt_child_exp,
7076 sizeof(KEY_OSP_CONNECTED),
7077 KEY_OSP_CONNECTED, NULL, NULL);
7080 set_bit(MDT_FL_SYNCED, &mdt->mdt_state);
7083 rc = class_connect(&conn, obd, cluuid);
7087 lexp = class_conn2export(&conn);
7088 LASSERT(lexp != NULL);
7090 rc = nodemap_add_member(client_nid, lexp);
7091 if (rc != 0 && rc != -EEXIST)
7094 rc = mdt_connect_internal(env, lexp, mdt, data, false);
7096 struct lsd_client_data *lcd = lexp->exp_target_data.ted_lcd;
7099 memcpy(lcd->lcd_uuid, cluuid, sizeof lcd->lcd_uuid);
7100 rc = tgt_client_new(env, lexp);
7102 mdt_export_stats_init(obd, lexp, localdata);
7106 class_disconnect(lexp);
7107 nodemap_del_member(lexp);
7111 /* Because we do not want this export to be evicted by pinger,
7112 * let's not add this export to the timed chain list. */
7113 if (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) {
7114 spin_lock(&lexp->exp_obd->obd_dev_lock);
7115 list_del_init(&lexp->exp_obd_chain_timed);
7116 spin_unlock(&lexp->exp_obd->obd_dev_lock);
7123 static int mdt_obd_reconnect(const struct lu_env *env,
7124 struct obd_export *exp, struct obd_device *obd,
7125 struct obd_uuid *cluuid,
7126 struct obd_connect_data *data,
7129 struct lnet_nid *client_nid = localdata;
7133 if (exp == NULL || obd == NULL || cluuid == NULL)
7136 rc = nodemap_add_member(client_nid, exp);
7137 if (rc != 0 && rc != -EEXIST)
7140 rc = mdt_connect_internal(env, exp, mdt_dev(obd->obd_lu_dev), data,
7143 mdt_export_stats_init(obd, exp, localdata);
7145 nodemap_del_member(exp);
7150 /* FIXME: Can we avoid using these two interfaces? */
7151 static int mdt_init_export(struct obd_export *exp)
7153 struct mdt_export_data *med = &exp->exp_mdt_data;
7157 INIT_LIST_HEAD(&med->med_open_head);
7158 spin_lock_init(&med->med_open_lock);
7159 spin_lock(&exp->exp_lock);
7160 exp->exp_connecting = 1;
7161 spin_unlock(&exp->exp_lock);
7163 OBD_ALLOC(exp->exp_used_slots,
7164 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
7165 if (exp->exp_used_slots == NULL)
7168 /* self-export doesn't need client data and ldlm initialization */
7169 if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
7170 &exp->exp_client_uuid)))
7173 rc = tgt_client_alloc(exp);
7177 rc = ldlm_init_export(exp);
7184 tgt_client_free(exp);
7186 OBD_FREE(exp->exp_used_slots,
7187 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
7188 exp->exp_used_slots = NULL;
7190 CERROR("%s: Failed to initialize export: rc = %d\n",
7191 exp->exp_obd->obd_name, rc);
7195 static int mdt_destroy_export(struct obd_export *exp)
7199 target_destroy_export(exp);
7200 if (exp->exp_used_slots)
7201 OBD_FREE(exp->exp_used_slots,
7202 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
7204 /* destroy can be called from failed obd_setup, so
7205 * checking uuid is safer than obd_self_export */
7206 if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
7207 &exp->exp_client_uuid)))
7210 ldlm_destroy_export(exp);
7211 tgt_client_free(exp);
7213 LASSERT(list_empty(&exp->exp_outstanding_replies));
7214 LASSERT(list_empty(&exp->exp_mdt_data.med_open_head));
7217 * discard grants once we're sure no more
7218 * interaction with the client is possible
7220 tgt_grant_discard(exp);
7221 if (exp_connect_flags(exp) & OBD_CONNECT_GRANT)
7222 obd2obt(exp->exp_obd)->obt_lut->lut_tgd.tgd_tot_granted_clients--;
7224 if (!(exp->exp_flags & OBD_OPT_FORCE))
7225 tgt_grant_sanity_check(exp->exp_obd, __func__);
7230 int mdt_links_read(struct mdt_thread_info *info, struct mdt_object *mdt_obj,
7231 struct linkea_data *ldata)
7235 LASSERT(ldata->ld_buf->lb_buf != NULL);
7237 if (!mdt_object_exists(mdt_obj))
7240 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
7241 ldata->ld_buf, XATTR_NAME_LINK);
7242 if (rc == -ERANGE) {
7243 /* Buf was too small, figure out what we need. */
7244 lu_buf_free(ldata->ld_buf);
7245 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
7246 ldata->ld_buf, XATTR_NAME_LINK);
7249 ldata->ld_buf = lu_buf_check_and_alloc(ldata->ld_buf, rc);
7250 if (ldata->ld_buf->lb_buf == NULL)
7252 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
7253 ldata->ld_buf, XATTR_NAME_LINK);
7258 return linkea_init_with_rec(ldata);
7262 * Given an MDT object, try to look up the full path to the object.
7263 * Part of the MDT layer implementation of lfs fid2path.
7265 * \param[in] info Per-thread common data shared by MDT level handlers.
7266 * \param[in] obj Object to do path lookup of
7267 * \param[in,out] fp User-provided struct to store path information
7268 * \param[in] root_fid Root FID of current path should reach
7270 * \retval 0 Lookup successful, path information stored in fp
7271 * \retval -EAGAIN Lookup failed, usually because object is being moved
7272 * \retval negative errno if there was a problem
7274 static int mdt_path_current(struct mdt_thread_info *info,
7275 struct mdt_object *obj,
7276 struct getinfo_fid2path *fp,
7277 struct lu_fid *root_fid)
7279 struct mdt_device *mdt = info->mti_mdt;
7280 struct lu_name *tmpname = &info->mti_name;
7281 struct lu_fid *tmpfid = &info->mti_tmp_fid1;
7282 struct lu_buf *buf = &info->mti_big_buf;
7283 struct linkea_data ldata = { NULL };
7285 struct mdt_object *mdt_obj;
7286 struct link_ea_header *leh;
7287 struct link_ea_entry *lee;
7288 bool worthchecking = true;
7289 bool needsfid = false;
7290 bool supported = false;
7298 /* temp buffer for path element, the buffer will be finally freed
7299 * in mdt_thread_info_fini */
7300 buf = lu_buf_check_and_alloc(buf, PATH_MAX);
7301 if (buf->lb_buf == NULL)
7305 ptr = fp->gf_u.gf_path + fp->gf_pathlen - 1;
7308 *tmpfid = fp->gf_fid = *mdt_object_fid(obj);
7310 while (!lu_fid_eq(root_fid, &fp->gf_fid)) {
7311 if (!lu_fid_eq(root_fid, &mdt->mdt_md_root_fid) &&
7312 lu_fid_eq(&mdt->mdt_md_root_fid, &fp->gf_fid))
7313 GOTO(out, rc = -ENOENT);
7315 if (lu_fid_eq(mdt_object_fid(obj), tmpfid)) {
7317 mdt_object_get(info->mti_env, mdt_obj);
7319 mdt_obj = mdt_object_find(info->mti_env, mdt, tmpfid);
7320 if (IS_ERR(mdt_obj))
7321 GOTO(out, rc = PTR_ERR(mdt_obj));
7324 if (!mdt_object_exists(mdt_obj)) {
7325 mdt_object_put(info->mti_env, mdt_obj);
7326 GOTO(out, rc = -ENOENT);
7329 if (mdt_object_remote(mdt_obj)) {
7330 mdt_object_put(info->mti_env, mdt_obj);
7331 GOTO(remote_out, rc = -EREMOTE);
7334 if (worthchecking) {
7335 /* we need to know if the FID being
7336 * looked up is encrypted
7338 struct lu_attr la = { 0 };
7339 struct dt_object *dt = mdt_obj2dt(mdt_obj);
7341 if (dt && dt->do_ops && dt->do_ops->do_attr_get)
7342 dt_attr_get(info->mti_env, dt, &la);
7343 if (la.la_valid & LA_FLAGS &&
7344 la.la_flags & LUSTRE_ENCRYPT_FL) {
7345 if (!supported && mdt_info_req(info) &&
7346 !exp_connect_encrypt_fid2path(
7347 mdt_info_req(info)->rq_export)) {
7348 /* client does not support fid2path
7349 * for encrypted files
7351 mdt_object_put(info->mti_env, mdt_obj);
7352 GOTO(out, rc = -ENODATA);
7360 worthchecking = false;
7367 rc = mdt_links_read(info, mdt_obj, &ldata);
7369 mdt_object_put(info->mti_env, mdt_obj);
7374 lee = (struct link_ea_entry *)(leh + 1); /* link #0 */
7375 linkea_entry_unpack(lee, &reclen, tmpname, tmpfid);
7376 /* If set, use link #linkno for path lookup, otherwise use
7377 link #0. Only do this for the final path element. */
7378 if (first && fp->gf_linkno < leh->leh_reccount) {
7380 for (count = 0; count < fp->gf_linkno; count++) {
7381 lee = (struct link_ea_entry *)
7382 ((char *)lee + reclen);
7383 linkea_entry_unpack(lee, &reclen, tmpname,
7386 if (fp->gf_linkno < leh->leh_reccount - 1)
7387 /* indicate to user there are more links */
7391 /* Check if it is slave stripes */
7392 rc = mdt_is_dir_stripe(info, mdt_obj);
7393 mdt_object_put(info->mti_env, mdt_obj);
7397 fp->gf_fid = *tmpfid;
7401 /* Pack the name in the end of the buffer */
7402 ptr -= tmpname->ln_namelen;
7403 if (ptr - 1 <= fp->gf_u.gf_path)
7404 GOTO(out, rc = -ENAMETOOLONG);
7405 strncpy(ptr, tmpname->ln_name, tmpname->ln_namelen);
7407 /* Pack FID before file name, so that client can build
7408 * encoded/digested form.
7410 char fidstr[FID_LEN + 1];
7412 snprintf(fidstr, sizeof(fidstr), DFID,
7414 ptr -= strlen(fidstr);
7415 if (ptr - 1 <= fp->gf_u.gf_path)
7416 GOTO(out, rc = -ENAMETOOLONG);
7417 strncpy(ptr, fidstr, strlen(fidstr));
7421 /* keep the last resolved fid to the client, so the
7422 * client will build the left path on another MDT for
7424 fp->gf_fid = *tmpfid;
7429 /* non-zero will be treated as an error */
7434 ptr++; /* skip leading / unless this is an encrypted file */
7435 memmove(fp->gf_u.gf_path, ptr,
7436 fp->gf_u.gf_path + fp->gf_pathlen - ptr);
7443 * Given an MDT object, use mdt_path_current to get the path.
7444 * Essentially a wrapper to retry mdt_path_current a set number of times
7445 * if -EAGAIN is returned (usually because an object is being moved).
7447 * Part of the MDT layer implementation of lfs fid2path.
7449 * \param[in] info Per-thread common data shared by mdt level handlers.
7450 * \param[in] obj Object to do path lookup of
7451 * \param[in,out] fp User-provided struct for arguments and to store path
7454 * \retval 0 Lookup successful, path information stored in fp
7455 * \retval negative errno if there was a problem
7457 static int mdt_path(struct mdt_thread_info *info, struct mdt_object *obj,
7458 struct getinfo_fid2path *fp, struct lu_fid *root_fid)
7460 struct mdt_device *mdt = info->mti_mdt;
7465 if (fp->gf_pathlen < 3)
7468 if (root_fid == NULL)
7469 root_fid = &mdt->mdt_md_root_fid;
7471 if (lu_fid_eq(root_fid, mdt_object_fid(obj))) {
7472 fp->gf_u.gf_path[0] = '\0';
7476 /* Retry multiple times in case file is being moved */
7477 while (tries-- && rc == -EAGAIN)
7478 rc = mdt_path_current(info, obj, fp, root_fid);
7484 * Get the full path of the provided FID, as of changelog record recno.
7486 * This checks sanity and looks up object for user provided FID
7487 * before calling the actual path lookup code.
7489 * Part of the MDT layer implementation of lfs fid2path.
7491 * \param[in] info Per-thread common data shared by mdt level handlers.
7492 * \param[in,out] fp User-provided struct for arguments and to store path
7495 * \retval 0 Lookup successful, path information and recno stored in fp
7496 * \retval -ENOENT, object does not exist
7497 * \retval negative errno if there was a problem
7499 static int mdt_fid2path(struct mdt_thread_info *info,
7500 struct lu_fid *root_fid,
7501 struct getinfo_fid2path *fp)
7503 struct mdt_device *mdt = info->mti_mdt;
7504 struct mdt_object *obj;
7508 CDEBUG(D_IOCTL, "path get "DFID" from %llu #%d\n",
7509 PFID(&fp->gf_fid), fp->gf_recno, fp->gf_linkno);
7511 if (!fid_is_sane(&fp->gf_fid))
7514 if (!fid_is_namespace_visible(&fp->gf_fid)) {
7515 CDEBUG(D_INFO, "%s: "DFID" is invalid, f_seq should be >= %#llx"
7516 ", or f_oid != 0, or f_ver == 0\n", mdt_obd_name(mdt),
7517 PFID(&fp->gf_fid), (__u64)FID_SEQ_NORMAL);
7521 /* return error if client-provided root fid
7522 * is not the one stored in the export
7524 if (root_fid && !fid_is_zero(&info->mti_exp->exp_root_fid) &&
7525 !lu_fid_eq(root_fid, &info->mti_exp->exp_root_fid)) {
7527 "%s: root fid from client "DFID" but "DFID" stored in export\n",
7528 mdt_obd_name(mdt), PFID(root_fid),
7529 PFID(&info->mti_exp->exp_root_fid));
7533 obj = mdt_object_find(info->mti_env, mdt, &fp->gf_fid);
7536 CDEBUG(D_IOCTL, "cannot find "DFID": rc = %d\n",
7537 PFID(&fp->gf_fid), rc);
7541 if (mdt_object_remote(obj))
7543 else if (!mdt_object_exists(obj))
7549 mdt_object_put(info->mti_env, obj);
7550 CDEBUG(D_IOCTL, "nonlocal object "DFID": rc = %d\n",
7551 PFID(&fp->gf_fid), rc);
7555 rc = mdt_path(info, obj, fp, root_fid);
7557 CDEBUG(D_INFO, "fid "DFID", path %s recno %#llx linkno %u\n",
7558 PFID(&fp->gf_fid), fp->gf_u.gf_path,
7559 fp->gf_recno, fp->gf_linkno);
7561 mdt_object_put(info->mti_env, obj);
7566 static int mdt_rpc_fid2path(struct mdt_thread_info *info, void *key, int keylen,
7567 void *val, int vallen)
7569 struct getinfo_fid2path *fpout, *fpin;
7570 struct lu_fid *root_fid = NULL;
7573 fpin = key + round_up(sizeof(KEY_FID2PATH), 8);
7576 if (req_capsule_req_need_swab(info->mti_pill))
7577 lustre_swab_fid2path(fpin);
7579 memcpy(fpout, fpin, sizeof(*fpin));
7580 if (fpout->gf_pathlen != vallen - sizeof(*fpin))
7583 if (keylen >= round_up(sizeof(KEY_FID2PATH), 8) + sizeof(*fpin) +
7584 sizeof(struct lu_fid)) {
7585 /* client sent its root FID, which is normally fileset FID */
7586 root_fid = fpin->gf_u.gf_root_fid;
7587 if (req_capsule_req_need_swab(info->mti_pill))
7588 lustre_swab_lu_fid(root_fid);
7590 if (root_fid != NULL && !fid_is_sane(root_fid))
7594 rc = mdt_fid2path(info, root_fid, fpout);
7598 int mdt_get_info(struct tgt_session_info *tsi)
7608 key = req_capsule_client_get(tsi->tsi_pill, &RMF_GETINFO_KEY);
7610 CDEBUG(D_IOCTL, "No GETINFO key\n");
7611 RETURN(err_serious(-EFAULT));
7613 keylen = req_capsule_get_size(tsi->tsi_pill, &RMF_GETINFO_KEY,
7616 vallen = req_capsule_client_get(tsi->tsi_pill, &RMF_GETINFO_VALLEN);
7617 if (vallen == NULL) {
7618 CDEBUG(D_IOCTL, "%s: cannot get RMF_GETINFO_VALLEN buffer\n",
7619 tgt_name(tsi->tsi_tgt));
7620 RETURN(err_serious(-EFAULT));
7623 req_capsule_set_size(tsi->tsi_pill, &RMF_GETINFO_VAL, RCL_SERVER,
7625 rc = req_capsule_server_pack(tsi->tsi_pill);
7627 RETURN(err_serious(rc));
7629 valout = req_capsule_server_get(tsi->tsi_pill, &RMF_GETINFO_VAL);
7630 if (valout == NULL) {
7631 CDEBUG(D_IOCTL, "%s: cannot get get-info RPC out buffer\n",
7632 tgt_name(tsi->tsi_tgt));
7633 RETURN(err_serious(-EFAULT));
7636 if (KEY_IS(KEY_FID2PATH)) {
7637 struct mdt_thread_info *info = tsi2mdt_info(tsi);
7639 rc = mdt_rpc_fid2path(info, key, keylen, valout, *vallen);
7640 mdt_thread_info_fini(info);
7647 static int mdt_ioc_version_get(struct mdt_thread_info *mti, void *karg)
7649 struct obd_ioctl_data *data = karg;
7652 struct mdt_object *obj;
7653 struct mdt_lock_handle *lh;
7657 if (data->ioc_inlbuf1 == NULL || data->ioc_inllen1 != sizeof(*fid) ||
7658 data->ioc_inlbuf2 == NULL || data->ioc_inllen2 != sizeof(version))
7661 fid = (struct lu_fid *)data->ioc_inlbuf1;
7663 if (!fid_is_sane(fid))
7666 CDEBUG(D_IOCTL, "getting version for "DFID"\n", PFID(fid));
7668 lh = &mti->mti_lh[MDT_LH_PARENT];
7669 obj = mdt_object_find_lock(mti, fid, lh, MDS_INODELOCK_UPDATE, LCK_CR);
7671 RETURN(PTR_ERR(obj));
7673 if (mdt_object_remote(obj)) {
7676 * before calling version get the correct MDS should be
7677 * fid, this is error to find remote object here
7679 CERROR("nonlocal object "DFID"\n", PFID(fid));
7680 } else if (!mdt_object_exists(obj)) {
7681 *(__u64 *)data->ioc_inlbuf2 = ENOENT_VERSION;
7684 version = dt_version_get(mti->mti_env, mdt_obj2dt(obj));
7685 *(__u64 *)data->ioc_inlbuf2 = version;
7688 mdt_object_unlock_put(mti, obj, lh, 1);
7692 /* ioctls on obd dev */
7693 static int mdt_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
7694 void *karg, void __user *uarg)
7696 struct obd_device *obd = exp->exp_obd;
7697 struct mdt_device *mdt = mdt_dev(obd->obd_lu_dev);
7698 struct dt_device *dt = mdt->mdt_bottom;
7699 struct obd_ioctl_data *data;
7704 CDEBUG(D_IOCTL, "%s: cmd=%x len=%u karg=%pK uarg=%pK\n",
7705 obd->obd_name, cmd, len, karg, uarg);
7707 rc = lu_env_init(&env, LCT_MD_THREAD);
7711 /* handle commands that don't use @karg first */
7714 rc = mdt_device_sync(&env, mdt);
7716 case OBD_IOC_SET_READONLY:
7717 rc = dt_sync(&env, dt);
7719 rc = dt_ro(&env, dt);
7723 if (unlikely(karg == NULL)) {
7724 OBD_IOC_ERROR(obd->obd_name, cmd, "karg=NULL", rc = -EINVAL);
7730 case OBD_IOC_ABORT_RECOVERY: {
7731 if (data->ioc_type & OBD_FLG_ABORT_RECOV_MDT) {
7732 LCONSOLE_WARN("%s: Aborting MDT recovery\n",
7734 obd->obd_abort_mdt_recovery = 1;
7735 wake_up(&obd->obd_next_transno_waitq);
7736 } else { /* if (data->ioc_type & OBD_FLG_ABORT_RECOV_OST) */
7737 /* lctl didn't set OBD_FLG_ABORT_RECOV_OST < 2.13.57 */
7738 LCONSOLE_WARN("%s: Aborting client recovery\n",
7740 obd->obd_abort_recovery = 1;
7741 target_stop_recovery_thread(obd);
7746 case OBD_IOC_CHANGELOG_REG:
7747 case OBD_IOC_CHANGELOG_DEREG:
7748 case OBD_IOC_CHANGELOG_CLEAR:
7749 case OBD_IOC_LLOG_PRINT:
7750 case OBD_IOC_LLOG_CANCEL:
7751 rc = mdt->mdt_child->md_ops->mdo_iocontrol(&env, mdt->mdt_child,
7754 case OBD_IOC_START_LFSCK: {
7755 struct md_device *next = mdt->mdt_child;
7756 struct lfsck_start_param lsp;
7758 lsp.lsp_start = (struct lfsck_start *)(data->ioc_inlbuf1);
7759 lsp.lsp_index_valid = 0;
7760 rc = next->md_ops->mdo_iocontrol(&env, next, cmd, 0, &lsp);
7763 case OBD_IOC_STOP_LFSCK: {
7764 struct md_device *next = mdt->mdt_child;
7765 struct lfsck_stop stop;
7767 stop.ls_status = LS_STOPPED;
7768 /* Old lfsck utils may pass NULL @stop. */
7769 if (data->ioc_inlbuf1 == NULL)
7773 ((struct lfsck_stop *)(data->ioc_inlbuf1))->ls_flags;
7775 rc = next->md_ops->mdo_iocontrol(&env, next, cmd, 0, &stop);
7778 case OBD_IOC_QUERY_LFSCK: {
7779 struct md_device *next = mdt->mdt_child;
7781 rc = next->md_ops->mdo_iocontrol(&env, next, cmd, 0,
7785 case OBD_IOC_GET_OBJ_VERSION: {
7786 struct mdt_thread_info *mti;
7788 mti = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
7789 memset(mti, 0, sizeof(*mti));
7790 mti->mti_env = &env;
7794 rc = mdt_ioc_version_get(mti, karg);
7797 case OBD_IOC_CATLOGLIST: {
7798 struct mdt_thread_info *mti;
7800 mti = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
7801 lu_local_obj_fid(&mti->mti_tmp_fid1, LLOG_CATALOGS_OID);
7802 rc = llog_catalog_list(&env, mdt->mdt_bottom, 0, karg,
7803 &mti->mti_tmp_fid1);
7807 rc = OBD_IOC_ERROR(obd->obd_name, cmd, "unrecognized", -ENOTTY);
7815 static int mdt_postrecov(const struct lu_env *env, struct mdt_device *mdt)
7817 struct lu_device *ld = md2lu_dev(mdt->mdt_child);
7821 if (!mdt->mdt_skip_lfsck && !mdt->mdt_bottom->dd_rdonly) {
7822 struct lfsck_start_param lsp;
7824 lsp.lsp_start = NULL;
7825 lsp.lsp_index_valid = 0;
7826 rc = mdt->mdt_child->md_ops->mdo_iocontrol(env, mdt->mdt_child,
7827 OBD_IOC_START_LFSCK,
7829 if (rc != 0 && rc != -EALREADY)
7830 CWARN("%s: auto trigger paused LFSCK failed: rc = %d\n",
7831 mdt_obd_name(mdt), rc);
7834 rc = ld->ld_ops->ldo_recovery_complete(env, ld);
7838 static int mdt_obd_postrecov(struct obd_device *obd)
7843 rc = lu_env_init(&env, LCT_MD_THREAD);
7846 rc = mdt_postrecov(&env, mdt_dev(obd->obd_lu_dev));
7851 static const struct obd_ops mdt_obd_device_ops = {
7852 .o_owner = THIS_MODULE,
7853 .o_set_info_async = mdt_obd_set_info_async,
7854 .o_connect = mdt_obd_connect,
7855 .o_reconnect = mdt_obd_reconnect,
7856 .o_disconnect = mdt_obd_disconnect,
7857 .o_init_export = mdt_init_export,
7858 .o_destroy_export = mdt_destroy_export,
7859 .o_iocontrol = mdt_iocontrol,
7860 .o_postrecov = mdt_obd_postrecov,
7861 /* Data-on-MDT IO methods */
7862 .o_preprw = mdt_obd_preprw,
7863 .o_commitrw = mdt_obd_commitrw,
7866 static struct lu_device* mdt_device_fini(const struct lu_env *env,
7867 struct lu_device *d)
7869 struct mdt_device *m = mdt_dev(d);
7876 static struct lu_device *mdt_device_free(const struct lu_env *env,
7877 struct lu_device *d)
7879 struct mdt_device *m = mdt_dev(d);
7882 lu_device_fini(&m->mdt_lu_dev);
7888 static struct lu_device *mdt_device_alloc(const struct lu_env *env,
7889 struct lu_device_type *t,
7890 struct lustre_cfg *cfg)
7892 struct lu_device *l;
7893 struct mdt_device *m;
7900 rc = mdt_init0(env, m, t, cfg);
7902 mdt_device_free(env, l);
7907 l = ERR_PTR(-ENOMEM);
7911 /* context key constructor/destructor: mdt_key_init, mdt_key_fini */
7912 LU_KEY_INIT(mdt, struct mdt_thread_info);
7914 static void mdt_key_fini(const struct lu_context *ctx,
7915 struct lu_context_key *key, void* data)
7917 struct mdt_thread_info *info = data;
7919 if (info->mti_big_lmm) {
7920 OBD_FREE_LARGE(info->mti_big_lmm, info->mti_big_lmmsize);
7921 info->mti_big_lmm = NULL;
7922 info->mti_big_lmmsize = 0;
7925 if (info->mti_big_acl) {
7926 OBD_FREE_LARGE(info->mti_big_acl, info->mti_big_aclsize);
7927 info->mti_big_acl = NULL;
7928 info->mti_big_aclsize = 0;
7934 /* context key: mdt_thread_key */
7935 LU_CONTEXT_KEY_DEFINE(mdt, LCT_MD_THREAD);
7937 struct lu_ucred *mdt_ucred(const struct mdt_thread_info *info)
7939 return lu_ucred(info->mti_env);
7942 struct lu_ucred *mdt_ucred_check(const struct mdt_thread_info *info)
7944 return lu_ucred_check(info->mti_env);
7948 * Enable/disable COS (Commit On Sharing).
7950 * Set/Clear the COS flag in mdt options.
7952 * \param mdt mdt device
7953 * \param val 0 disables COS, other values enable COS
7955 void mdt_enable_cos(struct mdt_device *mdt, bool val)
7960 mdt->mdt_opts.mo_cos = val;
7961 rc = lu_env_init(&env, LCT_LOCAL);
7962 if (unlikely(rc != 0)) {
7963 CWARN("%s: lu_env initialization failed, cannot "
7964 "sync: rc = %d\n", mdt_obd_name(mdt), rc);
7967 mdt_device_sync(&env, mdt);
7972 * Check COS (Commit On Sharing) status.
7974 * Return COS flag status.
7976 * \param mdt mdt device
7978 int mdt_cos_is_enabled(struct mdt_device *mdt)
7980 return mdt->mdt_opts.mo_cos != 0;
7983 static const struct lu_device_type_operations mdt_device_type_ops = {
7984 .ldto_device_alloc = mdt_device_alloc,
7985 .ldto_device_free = mdt_device_free,
7986 .ldto_device_fini = mdt_device_fini
7989 static struct lu_device_type mdt_device_type = {
7990 .ldt_tags = LU_DEVICE_MD,
7991 .ldt_name = LUSTRE_MDT_NAME,
7992 .ldt_ops = &mdt_device_type_ops,
7993 .ldt_ctx_tags = LCT_MD_THREAD
7996 static int __init mdt_init(void)
8000 BUILD_BUG_ON(sizeof("0x0123456789ABCDEF:0x01234567:0x01234567") !=
8001 FID_NOBRACE_LEN + 1);
8002 BUILD_BUG_ON(sizeof("[0x0123456789ABCDEF:0x01234567:0x01234567]") !=
8005 rc = libcfs_setup();
8009 rc = lu_kmem_init(mdt_caches);
8013 rc = mds_mod_init();
8017 rc = class_register_type(&mdt_obd_device_ops, NULL, true,
8018 LUSTRE_MDT_NAME, &mdt_device_type);
8023 lu_kmem_fini(mdt_caches);
8030 static void __exit mdt_exit(void)
8032 class_unregister_type(LUSTRE_MDT_NAME);
8034 lu_kmem_fini(mdt_caches);
8037 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
8038 MODULE_DESCRIPTION("Lustre Metadata Target ("LUSTRE_MDT_NAME")");
8039 MODULE_VERSION(LUSTRE_VERSION_STRING);
8040 MODULE_LICENSE("GPL");
8042 module_init(mdt_init);
8043 module_exit(mdt_exit);