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.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2010, 2013, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/mdt/mdt_handler.c
38 * Lustre Metadata Target (mdt) request handler
40 * Author: Peter Braam <braam@clusterfs.com>
41 * Author: Andreas Dilger <adilger@clusterfs.com>
42 * Author: Phil Schwan <phil@clusterfs.com>
43 * Author: Mike Shaver <shaver@clusterfs.com>
44 * Author: Nikita Danilov <nikita@clusterfs.com>
45 * Author: Huang Hua <huanghua@clusterfs.com>
46 * Author: Yury Umanets <umka@clusterfs.com>
49 #define DEBUG_SUBSYSTEM S_MDS
51 #include <linux/module.h>
53 * struct OBD_{ALLOC,FREE}*()
55 #include <obd_support.h>
56 /* struct ptlrpc_request */
57 #include <lustre_net.h>
58 /* struct obd_export */
59 #include <lustre_export.h>
60 /* struct obd_device */
63 #include <dt_object.h>
64 #include <lustre_mds.h>
65 #include <lustre_log.h>
66 #include "mdt_internal.h"
67 #include <lustre_acl.h>
68 #include <lustre_param.h>
69 #include <lustre_quota.h>
70 #include <lustre_linkea.h>
71 #include <lustre_lfsck.h>
73 mdl_mode_t mdt_mdl_lock_modes[] = {
74 [LCK_MINMODE] = MDL_MINMODE,
81 [LCK_GROUP] = MDL_GROUP
84 ldlm_mode_t mdt_dlm_lock_modes[] = {
85 [MDL_MINMODE] = LCK_MINMODE,
92 [MDL_GROUP] = LCK_GROUP
95 static struct mdt_device *mdt_dev(struct lu_device *d);
96 static int mdt_unpack_req_pack_rep(struct mdt_thread_info *info, __u32 flags);
98 static const struct lu_object_operations mdt_obj_ops;
100 /* Slab for MDT object allocation */
101 static struct kmem_cache *mdt_object_kmem;
103 /* For HSM restore handles */
104 struct kmem_cache *mdt_hsm_cdt_kmem;
106 /* For HSM request handles */
107 struct kmem_cache *mdt_hsm_car_kmem;
109 static struct lu_kmem_descr mdt_caches[] = {
111 .ckd_cache = &mdt_object_kmem,
112 .ckd_name = "mdt_obj",
113 .ckd_size = sizeof(struct mdt_object)
116 .ckd_cache = &mdt_hsm_cdt_kmem,
117 .ckd_name = "mdt_cdt_restore_handle",
118 .ckd_size = sizeof(struct cdt_restore_handle)
121 .ckd_cache = &mdt_hsm_car_kmem,
122 .ckd_name = "mdt_cdt_agent_req",
123 .ckd_size = sizeof(struct cdt_agent_req)
130 int mdt_get_disposition(struct ldlm_reply *rep, int flag)
134 return (rep->lock_policy_res1 & flag);
137 void mdt_clear_disposition(struct mdt_thread_info *info,
138 struct ldlm_reply *rep, int flag)
141 info->mti_opdata &= ~flag;
142 tgt_opdata_clear(info->mti_env, flag);
145 rep->lock_policy_res1 &= ~flag;
148 void mdt_set_disposition(struct mdt_thread_info *info,
149 struct ldlm_reply *rep, int flag)
152 info->mti_opdata |= flag;
153 tgt_opdata_set(info->mti_env, flag);
156 rep->lock_policy_res1 |= flag;
159 void mdt_lock_reg_init(struct mdt_lock_handle *lh, ldlm_mode_t lm)
161 lh->mlh_pdo_hash = 0;
162 lh->mlh_reg_mode = lm;
163 lh->mlh_rreg_mode = lm;
164 lh->mlh_type = MDT_REG_LOCK;
167 void mdt_lock_pdo_init(struct mdt_lock_handle *lh, ldlm_mode_t lock_mode,
168 const struct lu_name *lname)
170 lh->mlh_reg_mode = lock_mode;
171 lh->mlh_rreg_mode = lock_mode;
172 lh->mlh_type = MDT_PDO_LOCK;
174 if (lu_name_is_valid(lname)) {
175 lh->mlh_pdo_hash = full_name_hash(lname->ln_name,
177 /* XXX Workaround for LU-2856
179 * Zero is a valid return value of full_name_hash, but
180 * several users of mlh_pdo_hash assume a non-zero
181 * hash value. We therefore map zero onto an
182 * arbitrary, but consistent value (1) to avoid
183 * problems further down the road. */
184 if (unlikely(lh->mlh_pdo_hash == 0))
185 lh->mlh_pdo_hash = 1;
187 lh->mlh_pdo_hash = 0;
191 static void mdt_lock_pdo_mode(struct mdt_thread_info *info, struct mdt_object *o,
192 struct mdt_lock_handle *lh)
198 * Any dir access needs couple of locks:
200 * 1) on part of dir we gonna take lookup/modify;
202 * 2) on whole dir to protect it from concurrent splitting and/or to
203 * flush client's cache for readdir().
205 * so, for a given mode and object this routine decides what lock mode
206 * to use for lock #2:
208 * 1) if caller's gonna lookup in dir then we need to protect dir from
209 * being splitted only - LCK_CR
211 * 2) if caller's gonna modify dir then we need to protect dir from
212 * being splitted and to flush cache - LCK_CW
214 * 3) if caller's gonna modify dir and that dir seems ready for
215 * splitting then we need to protect it from any type of access
216 * (lookup/modify/split) - LCK_EX --bzzz
219 LASSERT(lh->mlh_reg_mode != LCK_MINMODE);
220 LASSERT(lh->mlh_pdo_mode == LCK_MINMODE);
223 * Ask underlaying level its opinion about preferable PDO lock mode
224 * having access type passed as regular lock mode:
226 * - MDL_MINMODE means that lower layer does not want to specify lock
229 * - MDL_NL means that no PDO lock should be taken. This is used in some
230 * cases. Say, for non-splittable directories no need to use PDO locks
233 mode = mdo_lock_mode(info->mti_env, mdt_object_child(o),
234 mdt_dlm_mode2mdl_mode(lh->mlh_reg_mode));
236 if (mode != MDL_MINMODE) {
237 lh->mlh_pdo_mode = mdt_mdl_mode2dlm_mode(mode);
240 * Lower layer does not want to specify locking mode. We do it
241 * our selves. No special protection is needed, just flush
242 * client's cache on modification and allow concurrent
245 switch (lh->mlh_reg_mode) {
247 lh->mlh_pdo_mode = LCK_EX;
250 lh->mlh_pdo_mode = LCK_CR;
253 lh->mlh_pdo_mode = LCK_CW;
256 CERROR("Not expected lock type (0x%x)\n",
257 (int)lh->mlh_reg_mode);
262 LASSERT(lh->mlh_pdo_mode != LCK_MINMODE);
266 int mdt_getstatus(struct tgt_session_info *tsi)
268 struct mdt_thread_info *info = tsi2mdt_info(tsi);
269 struct mdt_device *mdt = info->mti_mdt;
270 struct mdt_body *repbody;
275 rc = mdt_check_ucred(info);
277 GOTO(out, rc = err_serious(rc));
279 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK))
280 GOTO(out, rc = err_serious(-ENOMEM));
282 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
283 repbody->fid1 = mdt->mdt_md_root_fid;
284 repbody->valid |= OBD_MD_FLID;
286 if (tsi->tsi_tgt->lut_mds_capa &&
287 exp_connect_flags(info->mti_exp) & OBD_CONNECT_MDS_CAPA) {
288 struct mdt_object *root;
289 struct lustre_capa *capa;
291 root = mdt_object_find(info->mti_env, mdt, &repbody->fid1);
293 GOTO(out, rc = PTR_ERR(root));
295 capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA1);
297 capa->lc_opc = CAPA_OPC_MDS_DEFAULT;
298 rc = mo_capa_get(info->mti_env, mdt_object_child(root), capa,
300 mdt_object_put(info->mti_env, root);
302 repbody->valid |= OBD_MD_FLMDSCAPA;
306 mdt_thread_info_fini(info);
310 int mdt_statfs(struct tgt_session_info *tsi)
312 struct ptlrpc_request *req = tgt_ses_req(tsi);
313 struct mdt_thread_info *info = tsi2mdt_info(tsi);
314 struct md_device *next = info->mti_mdt->mdt_child;
315 struct ptlrpc_service_part *svcpt;
316 struct obd_statfs *osfs;
321 svcpt = req->rq_rqbd->rqbd_svcpt;
323 /* This will trigger a watchdog timeout */
324 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_STATFS_LCW_SLEEP,
325 (MDT_SERVICE_WATCHDOG_FACTOR *
326 at_get(&svcpt->scp_at_estimate)) + 1);
328 rc = mdt_check_ucred(info);
330 GOTO(out, rc = err_serious(rc));
332 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK))
333 GOTO(out, rc = err_serious(-ENOMEM));
335 osfs = req_capsule_server_get(info->mti_pill, &RMF_OBD_STATFS);
337 GOTO(out, rc = -EPROTO);
339 /** statfs information are cached in the mdt_device */
340 if (cfs_time_before_64(info->mti_mdt->mdt_osfs_age,
341 cfs_time_shift_64(-OBD_STATFS_CACHE_SECONDS))) {
342 /** statfs data is too old, get up-to-date one */
343 rc = next->md_ops->mdo_statfs(info->mti_env, next, osfs);
346 spin_lock(&info->mti_mdt->mdt_osfs_lock);
347 info->mti_mdt->mdt_osfs = *osfs;
348 info->mti_mdt->mdt_osfs_age = cfs_time_current_64();
349 spin_unlock(&info->mti_mdt->mdt_osfs_lock);
351 /** use cached statfs data */
352 spin_lock(&info->mti_mdt->mdt_osfs_lock);
353 *osfs = info->mti_mdt->mdt_osfs;
354 spin_unlock(&info->mti_mdt->mdt_osfs_lock);
358 mdt_counter_incr(req, LPROC_MDT_STATFS);
360 mdt_thread_info_fini(info);
365 * Pack SOM attributes into the reply.
366 * Call under a DLM UPDATE lock.
368 static void mdt_pack_size2body(struct mdt_thread_info *info,
369 struct mdt_object *mo)
372 struct md_attr *ma = &info->mti_attr;
374 LASSERT(ma->ma_attr.la_valid & LA_MODE);
375 b = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
377 /* Check if Size-on-MDS is supported, if this is a regular file,
378 * if SOM is enabled on the object and if SOM cache exists and valid.
379 * Otherwise do not pack Size-on-MDS attributes to the reply. */
380 if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) ||
381 !S_ISREG(ma->ma_attr.la_mode) ||
382 !mdt_object_is_som_enabled(mo) ||
383 !(ma->ma_valid & MA_SOM))
386 b->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
387 b->size = ma->ma_som->msd_size;
388 b->blocks = ma->ma_som->msd_blocks;
391 void mdt_pack_attr2body(struct mdt_thread_info *info, struct mdt_body *b,
392 const struct lu_attr *attr, const struct lu_fid *fid)
394 struct md_attr *ma = &info->mti_attr;
396 LASSERT(ma->ma_valid & MA_INODE);
398 b->atime = attr->la_atime;
399 b->mtime = attr->la_mtime;
400 b->ctime = attr->la_ctime;
401 b->mode = attr->la_mode;
402 b->size = attr->la_size;
403 b->blocks = attr->la_blocks;
404 b->uid = attr->la_uid;
405 b->gid = attr->la_gid;
406 b->flags = attr->la_flags;
407 b->nlink = attr->la_nlink;
408 b->rdev = attr->la_rdev;
410 /*XXX should pack the reply body according to lu_valid*/
411 b->valid |= OBD_MD_FLCTIME | OBD_MD_FLUID |
412 OBD_MD_FLGID | OBD_MD_FLTYPE |
413 OBD_MD_FLMODE | OBD_MD_FLNLINK | OBD_MD_FLFLAGS |
414 OBD_MD_FLATIME | OBD_MD_FLMTIME ;
416 if (!S_ISREG(attr->la_mode)) {
417 b->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS | OBD_MD_FLRDEV;
418 } else if (ma->ma_need & MA_LOV && !(ma->ma_valid & MA_LOV)) {
419 /* means no objects are allocated on osts. */
420 LASSERT(!(ma->ma_valid & MA_LOV));
421 /* just ignore blocks occupied by extend attributes on MDS */
423 /* if no object is allocated on osts, the size on mds is valid. b=22272 */
424 b->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
425 } else if ((ma->ma_valid & MA_LOV) && ma->ma_lmm != NULL &&
426 ma->ma_lmm->lmm_pattern & LOV_PATTERN_F_RELEASED) {
427 /* A released file stores its size on MDS. */
428 /* But return 1 block for released file, unless tools like tar
429 * will consider it fully sparse. (LU-3864)
431 if (unlikely(b->size == 0))
435 b->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
440 b->valid |= OBD_MD_FLID;
441 CDEBUG(D_INODE, DFID": nlink=%d, mode=%o, size="LPU64"\n",
442 PFID(fid), b->nlink, b->mode, b->size);
446 mdt_body_reverse_idmap(info, b);
448 if (fid != NULL && (b->valid & OBD_MD_FLSIZE))
449 CDEBUG(D_VFSTRACE, DFID": returning size %llu\n",
450 PFID(fid), (unsigned long long)b->size);
453 static inline int mdt_body_has_lov(const struct lu_attr *la,
454 const struct mdt_body *body)
456 return ((S_ISREG(la->la_mode) && (body->valid & OBD_MD_FLEASIZE)) ||
457 (S_ISDIR(la->la_mode) && (body->valid & OBD_MD_FLDIREA )) );
460 void mdt_client_compatibility(struct mdt_thread_info *info)
462 struct mdt_body *body;
463 struct ptlrpc_request *req = mdt_info_req(info);
464 struct obd_export *exp = req->rq_export;
465 struct md_attr *ma = &info->mti_attr;
466 struct lu_attr *la = &ma->ma_attr;
469 if (exp_connect_layout(exp))
470 /* the client can deal with 16-bit lmm_stripe_count */
473 body = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
475 if (!mdt_body_has_lov(la, body))
478 /* now we have a reply with a lov for a client not compatible with the
479 * layout lock so we have to clean the layout generation number */
480 if (S_ISREG(la->la_mode))
481 ma->ma_lmm->lmm_layout_gen = 0;
485 static int mdt_big_xattr_get(struct mdt_thread_info *info, struct mdt_object *o,
488 const struct lu_env *env = info->mti_env;
492 LASSERT(info->mti_big_lmm_used == 0);
493 rc = mo_xattr_get(env, mdt_object_child(o), &LU_BUF_NULL, name);
497 /* big_lmm may need to be grown */
498 if (info->mti_big_lmmsize < rc) {
499 int size = size_roundup_power2(rc);
501 if (info->mti_big_lmmsize > 0) {
502 /* free old buffer */
503 LASSERT(info->mti_big_lmm);
504 OBD_FREE_LARGE(info->mti_big_lmm,
505 info->mti_big_lmmsize);
506 info->mti_big_lmm = NULL;
507 info->mti_big_lmmsize = 0;
510 OBD_ALLOC_LARGE(info->mti_big_lmm, size);
511 if (info->mti_big_lmm == NULL)
513 info->mti_big_lmmsize = size;
515 LASSERT(info->mti_big_lmmsize >= rc);
517 info->mti_buf.lb_buf = info->mti_big_lmm;
518 info->mti_buf.lb_len = info->mti_big_lmmsize;
519 rc = mo_xattr_get(env, mdt_object_child(o), &info->mti_buf, name);
524 int mdt_attr_get_lov(struct mdt_thread_info *info,
525 struct mdt_object *o, struct md_attr *ma)
527 struct md_object *next = mdt_object_child(o);
528 struct lu_buf *buf = &info->mti_buf;
531 buf->lb_buf = ma->ma_lmm;
532 buf->lb_len = ma->ma_lmm_size;
533 rc = mo_xattr_get(info->mti_env, next, buf, XATTR_NAME_LOV);
535 ma->ma_lmm_size = rc;
536 ma->ma_valid |= MA_LOV;
538 } else if (rc == -ENODATA) {
541 } else if (rc == -ERANGE) {
542 rc = mdt_big_xattr_get(info, o, XATTR_NAME_LOV);
544 info->mti_big_lmm_used = 1;
545 ma->ma_valid |= MA_LOV;
546 ma->ma_lmm = info->mti_big_lmm;
547 ma->ma_lmm_size = rc;
548 /* update mdt_max_mdsize so all clients
549 * will be aware about that */
550 if (info->mti_mdt->mdt_max_mdsize < rc)
551 info->mti_mdt->mdt_max_mdsize = rc;
559 int mdt_attr_get_pfid(struct mdt_thread_info *info,
560 struct mdt_object *o, struct lu_fid *pfid)
562 struct lu_buf *buf = &info->mti_buf;
563 struct link_ea_header *leh;
564 struct link_ea_entry *lee;
568 buf->lb_buf = info->mti_big_lmm;
569 buf->lb_len = info->mti_big_lmmsize;
570 rc = mo_xattr_get(info->mti_env, mdt_object_child(o),
571 buf, XATTR_NAME_LINK);
572 /* ignore errors, MA_PFID won't be set and it is
573 * up to the caller to treat this as an error */
574 if (rc == -ERANGE || buf->lb_len == 0) {
575 rc = mdt_big_xattr_get(info, o, XATTR_NAME_LINK);
576 buf->lb_buf = info->mti_big_lmm;
577 buf->lb_len = info->mti_big_lmmsize;
582 if (rc < sizeof(*leh)) {
583 CERROR("short LinkEA on "DFID": rc = %d\n",
584 PFID(mdt_object_fid(o)), rc);
588 leh = (struct link_ea_header *) buf->lb_buf;
589 lee = (struct link_ea_entry *)(leh + 1);
590 if (leh->leh_magic == __swab32(LINK_EA_MAGIC)) {
591 leh->leh_magic = LINK_EA_MAGIC;
592 leh->leh_reccount = __swab32(leh->leh_reccount);
593 leh->leh_len = __swab64(leh->leh_len);
595 if (leh->leh_magic != LINK_EA_MAGIC)
597 if (leh->leh_reccount == 0)
600 memcpy(pfid, &lee->lee_parent_fid, sizeof(*pfid));
601 fid_be_to_cpu(pfid, pfid);
606 int mdt_attr_get_complex(struct mdt_thread_info *info,
607 struct mdt_object *o, struct md_attr *ma)
609 const struct lu_env *env = info->mti_env;
610 struct md_object *next = mdt_object_child(o);
611 struct lu_buf *buf = &info->mti_buf;
612 u32 mode = lu_object_attr(&next->mo_lu);
613 int need = ma->ma_need;
619 if (need & MA_INODE) {
620 ma->ma_need = MA_INODE;
621 rc = mo_attr_get(env, next, ma);
624 ma->ma_valid |= MA_INODE;
627 if (need & MA_PFID) {
628 rc = mdt_attr_get_pfid(info, o, &ma->ma_pfid);
630 ma->ma_valid |= MA_PFID;
631 /* ignore this error, parent fid is not mandatory */
635 if (need & MA_LOV && (S_ISREG(mode) || S_ISDIR(mode))) {
636 rc = mdt_attr_get_lov(info, o, ma);
641 if (need & MA_LMV && S_ISDIR(mode)) {
642 buf->lb_buf = ma->ma_lmv;
643 buf->lb_len = ma->ma_lmv_size;
644 rc2 = mo_xattr_get(env, next, buf, XATTR_NAME_LMV);
646 ma->ma_lmv_size = rc2;
647 ma->ma_valid |= MA_LMV;
648 } else if (rc2 == -ENODATA) {
655 if (need & MA_SOM && S_ISREG(mode)) {
656 buf->lb_buf = info->mti_xattr_buf;
657 buf->lb_len = sizeof(info->mti_xattr_buf);
658 CLASSERT(sizeof(struct som_attrs) <=
659 sizeof(info->mti_xattr_buf));
660 rc2 = mo_xattr_get(info->mti_env, next, buf, XATTR_NAME_SOM);
661 rc2 = lustre_buf2som(info->mti_xattr_buf, rc2, ma->ma_som);
663 ma->ma_valid |= MA_SOM;
664 else if (rc2 < 0 && rc2 != -ENODATA)
668 if (need & MA_HSM && S_ISREG(mode)) {
669 buf->lb_buf = info->mti_xattr_buf;
670 buf->lb_len = sizeof(info->mti_xattr_buf);
671 CLASSERT(sizeof(struct hsm_attrs) <=
672 sizeof(info->mti_xattr_buf));
673 rc2 = mo_xattr_get(info->mti_env, next, buf, XATTR_NAME_HSM);
674 rc2 = lustre_buf2hsm(info->mti_xattr_buf, rc2, &ma->ma_hsm);
676 ma->ma_valid |= MA_HSM;
677 else if (rc2 < 0 && rc2 != -ENODATA)
681 #ifdef CONFIG_FS_POSIX_ACL
682 if (need & MA_ACL_DEF && S_ISDIR(mode)) {
683 buf->lb_buf = ma->ma_acl;
684 buf->lb_len = ma->ma_acl_size;
685 rc2 = mo_xattr_get(env, next, buf, XATTR_NAME_ACL_DEFAULT);
687 ma->ma_acl_size = rc2;
688 ma->ma_valid |= MA_ACL_DEF;
689 } else if (rc2 == -ENODATA) {
698 CDEBUG(D_INODE, "after getattr rc = %d, ma_valid = "LPX64" ma_lmm=%p\n",
699 rc, ma->ma_valid, ma->ma_lmm);
703 static int mdt_getattr_internal(struct mdt_thread_info *info,
704 struct mdt_object *o, int ma_need)
706 struct md_object *next = mdt_object_child(o);
707 const struct mdt_body *reqbody = info->mti_body;
708 struct ptlrpc_request *req = mdt_info_req(info);
709 struct md_attr *ma = &info->mti_attr;
710 struct lu_attr *la = &ma->ma_attr;
711 struct req_capsule *pill = info->mti_pill;
712 const struct lu_env *env = info->mti_env;
713 struct mdt_body *repbody;
714 struct lu_buf *buffer = &info->mti_buf;
719 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK))
720 RETURN(err_serious(-ENOMEM));
722 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
726 if (mdt_object_remote(o)) {
727 /* This object is located on remote node.*/
728 /* Return -EIO for old client */
729 if (!mdt_is_dne_client(req->rq_export))
730 GOTO(out, rc = -EIO);
732 repbody->fid1 = *mdt_object_fid(o);
733 repbody->valid = OBD_MD_FLID | OBD_MD_MDS;
737 buffer->lb_len = reqbody->eadatasize;
738 if (buffer->lb_len > 0) {
739 buffer->lb_buf = req_capsule_server_get(pill, &RMF_MDT_MD);
740 if (buffer->lb_buf == NULL)
741 GOTO(out, rc = -EPROTO);
743 buffer->lb_buf = NULL;
744 ma_need &= ~(MA_LOV | MA_LMV);
745 CDEBUG(D_INFO, "%s: RPC from %s: does not need LOVEA.\n",
746 mdt_obd_name(info->mti_mdt),
747 req->rq_export->exp_client_uuid.uuid);
750 /* If it is dir object and client require MEA, then we got MEA */
751 if (S_ISDIR(lu_object_attr(&next->mo_lu)) &&
752 reqbody->valid & OBD_MD_MEA) {
753 /* Assumption: MDT_MD size is enough for lmv size. */
754 ma->ma_lmv = buffer->lb_buf;
755 ma->ma_lmv_size = buffer->lb_len;
756 ma->ma_need = MA_INODE;
757 if (ma->ma_lmm_size > 0)
758 ma->ma_need |= MA_LMV;
760 ma->ma_lmm = buffer->lb_buf;
761 ma->ma_lmm_size = buffer->lb_len;
762 ma->ma_need = MA_INODE | MA_HSM;
763 if (ma->ma_lmm_size > 0)
764 ma->ma_need |= MA_LOV;
767 if (S_ISDIR(lu_object_attr(&next->mo_lu)) &&
768 reqbody->valid & OBD_MD_FLDIREA &&
769 lustre_msg_get_opc(req->rq_reqmsg) == MDS_GETATTR) {
770 /* get default stripe info for this dir. */
771 ma->ma_need |= MA_LOV_DEF;
773 ma->ma_need |= ma_need;
774 if (ma->ma_need & MA_SOM)
775 ma->ma_som = &info->mti_u.som.data;
777 rc = mdt_attr_get_complex(info, o, ma);
779 CERROR("%s: getattr error for "DFID": rc = %d\n",
780 mdt_obd_name(info->mti_mdt),
781 PFID(mdt_object_fid(o)), rc);
785 /* if file is released, check if a restore is running */
786 if ((ma->ma_valid & MA_HSM) && (ma->ma_hsm.mh_flags & HS_RELEASED) &&
787 mdt_hsm_restore_is_running(info, mdt_object_fid(o))) {
788 repbody->t_state = MS_RESTORE;
789 repbody->valid |= OBD_MD_TSTATE;
792 is_root = lu_fid_eq(mdt_object_fid(o), &info->mti_mdt->mdt_md_root_fid);
794 /* the Lustre protocol supposes to return default striping
795 * on the user-visible root if explicitly requested */
796 if ((ma->ma_valid & MA_LOV) == 0 && S_ISDIR(la->la_mode) &&
797 (ma->ma_need & MA_LOV_DEF && is_root) && ma->ma_need & MA_LOV) {
798 struct lu_fid rootfid;
799 struct mdt_object *root;
800 struct mdt_device *mdt = info->mti_mdt;
802 rc = dt_root_get(env, mdt->mdt_bottom, &rootfid);
805 root = mdt_object_find(env, mdt, &rootfid);
807 RETURN(PTR_ERR(root));
808 rc = mdt_attr_get_lov(info, root, ma);
809 mdt_object_put(info->mti_env, root);
811 CERROR("%s: getattr error for "DFID": rc = %d\n",
812 mdt_obd_name(info->mti_mdt),
813 PFID(mdt_object_fid(o)), rc);
818 if (likely(ma->ma_valid & MA_INODE))
819 mdt_pack_attr2body(info, repbody, la, mdt_object_fid(o));
823 if (mdt_body_has_lov(la, reqbody)) {
824 if (ma->ma_valid & MA_LOV) {
825 LASSERT(ma->ma_lmm_size);
826 mdt_dump_lmm(D_INFO, ma->ma_lmm);
827 repbody->eadatasize = ma->ma_lmm_size;
828 if (S_ISDIR(la->la_mode))
829 repbody->valid |= OBD_MD_FLDIREA;
831 repbody->valid |= OBD_MD_FLEASIZE;
833 if (ma->ma_valid & MA_LMV) {
834 LASSERT(S_ISDIR(la->la_mode));
835 repbody->eadatasize = ma->ma_lmv_size;
836 repbody->valid |= (OBD_MD_FLDIREA|OBD_MD_MEA);
838 } else if (S_ISLNK(la->la_mode) &&
839 reqbody->valid & OBD_MD_LINKNAME) {
840 buffer->lb_buf = ma->ma_lmm;
841 /* eadatasize from client includes NULL-terminator, so
842 * there is no need to read it */
843 buffer->lb_len = reqbody->eadatasize - 1;
844 rc = mo_readlink(env, next, buffer);
845 if (unlikely(rc <= 0)) {
846 CERROR("%s: readlink failed for "DFID": rc = %d\n",
847 mdt_obd_name(info->mti_mdt),
848 PFID(mdt_object_fid(o)), rc);
851 int print_limit = min_t(int, PAGE_CACHE_SIZE - 128, rc);
853 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_READLINK_EPROTO))
855 repbody->valid |= OBD_MD_LINKNAME;
856 /* we need to report back size with NULL-terminator
857 * because client expects that */
858 repbody->eadatasize = rc + 1;
859 if (repbody->eadatasize != reqbody->eadatasize)
860 CDEBUG(D_INODE, "%s: Read shorter symlink %d "
861 "on "DFID ", expected %d\n",
862 mdt_obd_name(info->mti_mdt),
863 rc, PFID(mdt_object_fid(o)),
864 reqbody->eadatasize - 1);
866 ((char *)ma->ma_lmm)[rc] = 0;
868 /* If the total CDEBUG() size is larger than a page, it
869 * will print a warning to the console, avoid this by
870 * printing just the last part of the symlink. */
871 CDEBUG(D_INODE, "symlink dest %s%.*s, len = %d\n",
872 print_limit < rc ? "..." : "", print_limit,
873 (char *)ma->ma_lmm + rc - print_limit, rc);
878 if (reqbody->valid & OBD_MD_FLMODEASIZE) {
879 repbody->max_cookiesize = 0;
880 repbody->max_mdsize = info->mti_mdt->mdt_max_mdsize;
881 repbody->valid |= OBD_MD_FLMODEASIZE;
882 CDEBUG(D_INODE, "I am going to change the MAX_MD_SIZE & "
883 "MAX_COOKIE to : %d:%d\n", repbody->max_mdsize,
884 repbody->max_cookiesize);
887 if (exp_connect_rmtclient(info->mti_exp) &&
888 reqbody->valid & OBD_MD_FLRMTPERM) {
889 void *buf = req_capsule_server_get(pill, &RMF_ACL);
891 /* mdt_getattr_lock only */
892 rc = mdt_pack_remote_perm(info, o, buf);
894 repbody->valid &= ~OBD_MD_FLRMTPERM;
895 repbody->aclsize = 0;
898 repbody->valid |= OBD_MD_FLRMTPERM;
899 repbody->aclsize = sizeof(struct mdt_remote_perm);
902 #ifdef CONFIG_FS_POSIX_ACL
903 else if ((exp_connect_flags(req->rq_export) & OBD_CONNECT_ACL) &&
904 (reqbody->valid & OBD_MD_FLACL)) {
905 buffer->lb_buf = req_capsule_server_get(pill, &RMF_ACL);
906 buffer->lb_len = req_capsule_get_size(pill,
907 &RMF_ACL, RCL_SERVER);
908 if (buffer->lb_len > 0) {
909 rc = mo_xattr_get(env, next, buffer,
910 XATTR_NAME_ACL_ACCESS);
912 if (rc == -ENODATA) {
913 repbody->aclsize = 0;
914 repbody->valid |= OBD_MD_FLACL;
916 } else if (rc == -EOPNOTSUPP) {
919 CERROR("%s: unable to read "DFID
921 mdt_obd_name(info->mti_mdt),
922 PFID(mdt_object_fid(o)), rc);
925 repbody->aclsize = rc;
926 repbody->valid |= OBD_MD_FLACL;
933 if (reqbody->valid & OBD_MD_FLMDSCAPA &&
934 info->mti_mdt->mdt_lut.lut_mds_capa &&
935 exp_connect_flags(info->mti_exp) & OBD_CONNECT_MDS_CAPA) {
936 struct lustre_capa *capa;
938 capa = req_capsule_server_get(pill, &RMF_CAPA1);
940 capa->lc_opc = CAPA_OPC_MDS_DEFAULT;
941 rc = mo_capa_get(env, next, capa, 0);
944 repbody->valid |= OBD_MD_FLMDSCAPA;
949 mdt_counter_incr(req, LPROC_MDT_GETATTR);
954 static int mdt_renew_capa(struct mdt_thread_info *info)
956 struct mdt_object *obj = info->mti_object;
957 struct mdt_body *body;
958 struct lustre_capa *capa, *c;
962 /* if object doesn't exist, or server has disabled capability,
963 * return directly, client will find body->valid OBD_MD_FLOSSCAPA
966 if (!obj || !info->mti_mdt->mdt_lut.lut_oss_capa ||
967 !(exp_connect_flags(info->mti_exp) & OBD_CONNECT_OSS_CAPA))
970 body = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
971 LASSERT(body != NULL);
973 c = req_capsule_client_get(info->mti_pill, &RMF_CAPA1);
976 capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA2);
980 rc = mo_capa_get(info->mti_env, mdt_object_child(obj), capa, 1);
982 body->valid |= OBD_MD_FLOSSCAPA;
986 int mdt_getattr(struct tgt_session_info *tsi)
988 struct mdt_thread_info *info = tsi2mdt_info(tsi);
989 struct mdt_object *obj = info->mti_object;
990 struct req_capsule *pill = info->mti_pill;
991 struct mdt_body *reqbody;
992 struct mdt_body *repbody;
997 reqbody = req_capsule_client_get(pill, &RMF_MDT_BODY);
1000 if (reqbody->valid & OBD_MD_FLOSSCAPA) {
1001 rc = req_capsule_server_pack(pill);
1003 RETURN(err_serious(rc));
1004 rc = mdt_renew_capa(info);
1005 GOTO(out_shrink, rc);
1008 LASSERT(obj != NULL);
1009 LASSERT(lu_object_assert_exists(&obj->mot_obj));
1011 mode = lu_object_attr(&obj->mot_obj);
1013 /* old clients may not report needed easize, use max value then */
1014 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER,
1015 reqbody->eadatasize == 0 ?
1016 info->mti_mdt->mdt_max_mdsize :
1017 reqbody->eadatasize);
1019 rc = req_capsule_server_pack(pill);
1020 if (unlikely(rc != 0))
1021 GOTO(out, rc = err_serious(rc));
1023 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
1024 LASSERT(repbody != NULL);
1025 repbody->eadatasize = 0;
1026 repbody->aclsize = 0;
1028 if (reqbody->valid & OBD_MD_FLRMTPERM)
1029 rc = mdt_init_ucred(info, reqbody);
1031 rc = mdt_check_ucred(info);
1033 GOTO(out_shrink, rc);
1035 info->mti_cross_ref = !!(reqbody->valid & OBD_MD_FLCROSSREF);
1038 * Don't check capability at all, because rename might getattr for
1039 * remote obj, and at that time no capability is available.
1041 mdt_set_capainfo(info, 1, &reqbody->fid1, BYPASS_CAPA);
1042 rc = mdt_getattr_internal(info, obj, 0);
1043 if (reqbody->valid & OBD_MD_FLRMTPERM)
1044 mdt_exit_ucred(info);
1047 mdt_client_compatibility(info);
1048 rc2 = mdt_fix_reply(info);
1052 mdt_thread_info_fini(info);
1056 int mdt_is_subdir(struct tgt_session_info *tsi)
1058 struct mdt_thread_info *info = tsi2mdt_info(tsi);
1059 struct mdt_object *o = info->mti_object;
1060 struct req_capsule *pill = info->mti_pill;
1061 const struct mdt_body *body = info->mti_body;
1062 struct mdt_body *repbody;
1068 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
1071 * We save last checked parent fid to @repbody->fid1 for remote
1074 LASSERT(fid_is_sane(&body->fid2));
1075 LASSERT(mdt_object_exists(o) && !mdt_object_remote(o));
1076 rc = mdo_is_subdir(info->mti_env, mdt_object_child(o),
1077 &body->fid2, &repbody->fid1);
1078 if (rc == 0 || rc == -EREMOTE)
1079 repbody->valid |= OBD_MD_FLID;
1081 mdt_thread_info_fini(info);
1085 int mdt_swap_layouts(struct tgt_session_info *tsi)
1087 struct mdt_thread_info *info;
1088 struct ptlrpc_request *req = tgt_ses_req(tsi);
1089 struct obd_export *exp = req->rq_export;
1090 struct mdt_object *o1, *o2, *o;
1091 struct mdt_lock_handle *lh1, *lh2;
1092 struct mdc_swap_layouts *msl;
1096 /* client does not support layout lock, so layout swaping
1098 * FIXME: there is a problem for old clients which don't support
1099 * layout lock yet. If those clients have already opened the file
1100 * they won't be notified at all so that old layout may still be
1101 * used to do IO. This can be fixed after file release is landed by
1102 * doing exclusive open and taking full EX ibits lock. - Jinshan */
1103 if (!exp_connect_layout(exp))
1104 RETURN(-EOPNOTSUPP);
1106 info = tsi2mdt_info(tsi);
1107 if (req_capsule_get_size(info->mti_pill, &RMF_CAPA1, RCL_CLIENT))
1108 mdt_set_capainfo(info, 0, &info->mti_body->fid1,
1109 req_capsule_client_get(info->mti_pill,
1112 if (req_capsule_get_size(info->mti_pill, &RMF_CAPA2, RCL_CLIENT))
1113 mdt_set_capainfo(info, 1, &info->mti_body->fid2,
1114 req_capsule_client_get(info->mti_pill,
1117 o1 = info->mti_object;
1118 o = o2 = mdt_object_find(info->mti_env, info->mti_mdt,
1119 &info->mti_body->fid2);
1121 GOTO(out, rc = PTR_ERR(o));
1123 if (mdt_object_remote(o) || !mdt_object_exists(o)) /* remote object */
1124 GOTO(put, rc = -ENOENT);
1126 rc = lu_fid_cmp(&info->mti_body->fid1, &info->mti_body->fid2);
1127 if (unlikely(rc == 0)) /* same file, you kidding me? no-op. */
1133 /* permission check. Make sure the calling process having permission
1134 * to write both files. */
1135 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o1), NULL,
1140 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o2), NULL,
1145 msl = req_capsule_client_get(info->mti_pill, &RMF_SWAP_LAYOUTS);
1147 GOTO(put, rc = -EPROTO);
1149 lh1 = &info->mti_lh[MDT_LH_NEW];
1150 mdt_lock_reg_init(lh1, LCK_EX);
1151 lh2 = &info->mti_lh[MDT_LH_OLD];
1152 mdt_lock_reg_init(lh2, LCK_EX);
1154 rc = mdt_object_lock(info, o1, lh1, MDS_INODELOCK_LAYOUT |
1155 MDS_INODELOCK_XATTR, MDT_LOCAL_LOCK);
1159 rc = mdt_object_lock(info, o2, lh2, MDS_INODELOCK_LAYOUT |
1160 MDS_INODELOCK_XATTR, MDT_LOCAL_LOCK);
1164 rc = mo_swap_layouts(info->mti_env, mdt_object_child(o1),
1165 mdt_object_child(o2), msl->msl_flags);
1168 mdt_object_unlock(info, o2, lh2, rc);
1170 mdt_object_unlock(info, o1, lh1, rc);
1172 mdt_object_put(info->mti_env, o);
1174 mdt_thread_info_fini(info);
1178 static int mdt_raw_lookup(struct mdt_thread_info *info,
1179 struct mdt_object *parent,
1180 const struct lu_name *lname,
1181 struct ldlm_reply *ldlm_rep)
1183 struct md_object *next = mdt_object_child(info->mti_object);
1184 const struct mdt_body *reqbody = info->mti_body;
1185 struct lu_fid *child_fid = &info->mti_tmp_fid1;
1186 struct mdt_body *repbody;
1190 if (reqbody->valid != OBD_MD_FLID)
1193 LASSERT(!info->mti_cross_ref);
1195 /* Only got the fid of this obj by name */
1196 fid_zero(child_fid);
1197 rc = mdo_lookup(info->mti_env, next, lname, child_fid,
1200 /* XXX is raw_lookup possible as intent operation? */
1203 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1206 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1208 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1211 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1212 repbody->fid1 = *child_fid;
1213 repbody->valid = OBD_MD_FLID;
1219 * UPDATE lock should be taken against parent, and be release before exit;
1220 * child_bits lock should be taken against child, and be returned back:
1221 * (1)normal request should release the child lock;
1222 * (2)intent request will grant the lock to client.
1224 static int mdt_getattr_name_lock(struct mdt_thread_info *info,
1225 struct mdt_lock_handle *lhc,
1227 struct ldlm_reply *ldlm_rep)
1229 struct ptlrpc_request *req = mdt_info_req(info);
1230 struct mdt_body *reqbody = NULL;
1231 struct mdt_object *parent = info->mti_object;
1232 struct mdt_object *child;
1233 struct md_object *next = mdt_object_child(parent);
1234 struct lu_fid *child_fid = &info->mti_tmp_fid1;
1235 struct lu_name *lname = NULL;
1236 struct mdt_lock_handle *lhp = NULL;
1237 struct ldlm_lock *lock;
1238 struct ldlm_res_id *res_id;
1245 is_resent = lustre_handle_is_used(&lhc->mlh_reg_lh);
1246 LASSERT(ergo(is_resent,
1247 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT));
1249 LASSERT(parent != NULL);
1251 lname = &info->mti_name;
1252 mdt_name_unpack(info->mti_pill, &RMF_NAME, lname, MNF_FIX_ANON);
1254 if (!info->mti_cross_ref) {
1256 * XXX: Check for anonymous name is for getattr by fid
1257 * (OBD_CONNECT_ATTRFID), otherwise do not allow empty name,
1258 * that is the name must contain at least one character and
1259 * the terminating '\0'.
1261 if (!lu_name_is_valid(lname)) {
1262 reqbody = req_capsule_client_get(info->mti_pill,
1264 if (unlikely(reqbody == NULL))
1265 RETURN(err_serious(-EFAULT));
1267 if (unlikely(!fid_is_sane(&reqbody->fid2)))
1268 RETURN(err_serious(-EINVAL));
1270 CDEBUG(D_INODE, "getattr with lock for "DFID"/"DFID", "
1272 PFID(mdt_object_fid(parent)),
1273 PFID(&reqbody->fid2), ldlm_rep);
1275 CDEBUG(D_INODE, "getattr with lock for "DFID"/"DNAME", "
1276 "ldlm_rep = %p\n", PFID(mdt_object_fid(parent)),
1277 PNAME(lname), ldlm_rep);
1281 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_EXECD);
1283 if (unlikely(!mdt_object_exists(parent)) && lu_name_is_valid(lname)) {
1284 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
1286 "Parent doesn't exist!\n");
1288 } else if (!info->mti_cross_ref) {
1289 LASSERTF(!mdt_object_remote(parent),
1290 "Parent "DFID" is on remote server\n",
1291 PFID(mdt_object_fid(parent)));
1294 if (lu_name_is_valid(lname)) {
1295 rc = mdt_raw_lookup(info, parent, lname, ldlm_rep);
1303 if (info->mti_cross_ref) {
1304 /* Only getattr on the child. Parent is on another node. */
1305 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1307 CDEBUG(D_INODE, "partial getattr_name child_fid = "DFID", "
1308 "ldlm_rep=%p\n", PFID(mdt_object_fid(child)), ldlm_rep);
1311 /* Do not take lock for resent case. */
1312 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
1313 LASSERTF(lock != NULL, "Invalid lock handle "LPX64"\n",
1314 lhc->mlh_reg_lh.cookie);
1315 LASSERT(fid_res_name_eq(mdt_object_fid(child),
1316 &lock->l_resource->lr_name));
1317 LDLM_LOCK_PUT(lock);
1320 mdt_lock_handle_init(lhc);
1321 mdt_lock_reg_init(lhc, LCK_PR);
1324 * Object's name is on another MDS, no lookup or layout
1325 * lock is needed here but update lock is.
1327 child_bits &= ~(MDS_INODELOCK_LOOKUP |
1328 MDS_INODELOCK_LAYOUT);
1329 child_bits |= MDS_INODELOCK_PERM | MDS_INODELOCK_UPDATE;
1331 rc = mdt_object_lock(info, child, lhc, child_bits,
1335 /* Finally, we can get attr for child. */
1336 if (!mdt_object_exists(child)) {
1337 LU_OBJECT_DEBUG(D_INFO, info->mti_env,
1339 "remote object doesn't exist.\n");
1340 mdt_object_unlock(info, child, lhc, 1);
1344 mdt_set_capainfo(info, 0, mdt_object_fid(child),
1346 rc = mdt_getattr_internal(info, child, 0);
1347 if (unlikely(rc != 0))
1348 mdt_object_unlock(info, child, lhc, 1);
1353 if (lu_name_is_valid(lname)) {
1354 /* step 1: lock parent only if parent is a directory */
1355 if (S_ISDIR(lu_object_attr(&parent->mot_obj))) {
1356 lhp = &info->mti_lh[MDT_LH_PARENT];
1357 mdt_lock_pdo_init(lhp, LCK_PR, lname);
1358 rc = mdt_object_lock(info, parent, lhp,
1359 MDS_INODELOCK_UPDATE,
1361 if (unlikely(rc != 0))
1365 /* step 2: lookup child's fid by name */
1366 fid_zero(child_fid);
1367 rc = mdo_lookup(info->mti_env, next, lname, child_fid,
1372 mdt_set_disposition(info, ldlm_rep,
1374 GOTO(out_parent, rc);
1376 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1378 *child_fid = reqbody->fid2;
1379 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1383 *step 3: find the child object by fid & lock it.
1384 * regardless if it is local or remote.
1386 *Note: LU-3240 (commit 762f2114d282a98ebfa4dbbeea9298a8088ad24e)
1387 * set parent dir fid the same as child fid in getattr by fid case
1388 * we should not lu_object_find() the object again, could lead
1389 * to hung if there is a concurrent unlink destroyed the object.
1391 if (lu_fid_eq(mdt_object_fid(parent), child_fid)) {
1392 mdt_object_get(info->mti_env, parent);
1395 child = mdt_object_find(info->mti_env, info->mti_mdt,
1399 if (unlikely(IS_ERR(child)))
1400 GOTO(out_parent, rc = PTR_ERR(child));
1402 /* Do not take lock for resent case. */
1403 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
1404 LASSERTF(lock != NULL, "Invalid lock handle "LPX64"\n",
1405 lhc->mlh_reg_lh.cookie);
1407 res_id = &lock->l_resource->lr_name;
1408 if (!fid_res_name_eq(mdt_object_fid(child),
1409 &lock->l_resource->lr_name)) {
1410 LASSERTF(fid_res_name_eq(mdt_object_fid(parent),
1411 &lock->l_resource->lr_name),
1412 "Lock res_id: "DLDLMRES", fid: "DFID"\n",
1413 PLDLMRES(lock->l_resource),
1414 PFID(mdt_object_fid(parent)));
1415 CWARN("Although resent, but still not get child lock"
1416 "parent:"DFID" child:"DFID"\n",
1417 PFID(mdt_object_fid(parent)),
1418 PFID(mdt_object_fid(child)));
1419 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
1420 LDLM_LOCK_PUT(lock);
1423 LDLM_LOCK_PUT(lock);
1426 bool try_layout = false;
1429 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RESEND, obd_timeout*2);
1430 mdt_lock_handle_init(lhc);
1431 mdt_lock_reg_init(lhc, LCK_PR);
1433 if (!mdt_object_exists(child)) {
1434 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
1436 "Object doesn't exist!\n");
1437 GOTO(out_child, rc = -ENOENT);
1440 if (!(child_bits & MDS_INODELOCK_UPDATE) &&
1441 mdt_object_exists(child) && !mdt_object_remote(child)) {
1442 struct md_attr *ma = &info->mti_attr;
1445 ma->ma_need = MA_INODE;
1446 rc = mdt_attr_get_complex(info, child, ma);
1447 if (unlikely(rc != 0))
1448 GOTO(out_child, rc);
1450 /* If the file has not been changed for some time, we
1451 * return not only a LOOKUP lock, but also an UPDATE
1452 * lock and this might save us RPC on later STAT. For
1453 * directories, it also let negative dentry cache start
1454 * working for this dir. */
1455 if (ma->ma_valid & MA_INODE &&
1456 ma->ma_attr.la_valid & LA_CTIME &&
1457 info->mti_mdt->mdt_namespace->ns_ctime_age_limit +
1458 ma->ma_attr.la_ctime < cfs_time_current_sec())
1459 child_bits |= MDS_INODELOCK_UPDATE;
1462 /* layout lock must be granted in a best-effort way
1463 * for IT operations */
1464 LASSERT(!(child_bits & MDS_INODELOCK_LAYOUT));
1465 if (!OBD_FAIL_CHECK(OBD_FAIL_MDS_NO_LL_GETATTR) &&
1466 exp_connect_layout(info->mti_exp) &&
1467 S_ISREG(lu_object_attr(&child->mot_obj)) &&
1468 !mdt_object_remote(child) && ldlm_rep != NULL) {
1469 /* try to grant layout lock for regular file. */
1475 child_bits |= MDS_INODELOCK_LAYOUT;
1476 /* try layout lock, it may fail to be granted due to
1477 * contention at LOOKUP or UPDATE */
1478 if (!mdt_object_lock_try(info, child, lhc, child_bits,
1480 child_bits &= ~MDS_INODELOCK_LAYOUT;
1481 LASSERT(child_bits != 0);
1482 rc = mdt_object_lock(info, child, lhc,
1483 child_bits, MDT_CROSS_LOCK);
1488 rc = mdt_object_lock(info, child, lhc, child_bits,
1491 if (unlikely(rc != 0))
1492 GOTO(out_child, rc);
1495 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
1496 /* Get MA_SOM attributes if update lock is given. */
1498 lock->l_policy_data.l_inodebits.bits & MDS_INODELOCK_UPDATE &&
1499 S_ISREG(lu_object_attr(&mdt_object_child(child)->mo_lu)))
1502 /* finally, we can get attr for child. */
1503 mdt_set_capainfo(info, 1, child_fid, BYPASS_CAPA);
1504 rc = mdt_getattr_internal(info, child, ma_need);
1505 if (unlikely(rc != 0)) {
1506 mdt_object_unlock(info, child, lhc, 1);
1508 /* Debugging code. */
1509 res_id = &lock->l_resource->lr_name;
1510 LDLM_DEBUG(lock, "Returning lock to client");
1511 LASSERTF(fid_res_name_eq(mdt_object_fid(child),
1512 &lock->l_resource->lr_name),
1513 "Lock res_id: "DLDLMRES", fid: "DFID"\n",
1514 PLDLMRES(lock->l_resource),
1515 PFID(mdt_object_fid(child)));
1516 if (mdt_object_exists(child) && !mdt_object_remote(child))
1517 mdt_pack_size2body(info, child);
1520 LDLM_LOCK_PUT(lock);
1524 mdt_object_put(info->mti_env, child);
1527 mdt_object_unlock(info, parent, lhp, 1);
1531 /* normal handler: should release the child lock */
1532 int mdt_getattr_name(struct tgt_session_info *tsi)
1534 struct mdt_thread_info *info = tsi2mdt_info(tsi);
1535 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_CHILD];
1536 struct mdt_body *reqbody;
1537 struct mdt_body *repbody;
1541 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
1542 LASSERT(reqbody != NULL);
1543 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1544 LASSERT(repbody != NULL);
1546 info->mti_cross_ref = !!(reqbody->valid & OBD_MD_FLCROSSREF);
1547 repbody->eadatasize = 0;
1548 repbody->aclsize = 0;
1550 rc = mdt_init_ucred(info, reqbody);
1552 GOTO(out_shrink, rc);
1554 rc = mdt_getattr_name_lock(info, lhc, MDS_INODELOCK_UPDATE, NULL);
1555 if (lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1556 ldlm_lock_decref(&lhc->mlh_reg_lh, lhc->mlh_reg_mode);
1557 lhc->mlh_reg_lh.cookie = 0;
1559 mdt_exit_ucred(info);
1562 mdt_client_compatibility(info);
1563 rc2 = mdt_fix_reply(info);
1566 mdt_thread_info_fini(info);
1570 static int mdt_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
1571 void *karg, void *uarg);
1573 int mdt_set_info(struct tgt_session_info *tsi)
1575 struct ptlrpc_request *req = tgt_ses_req(tsi);
1578 int keylen, vallen, rc = 0;
1582 key = req_capsule_client_get(tsi->tsi_pill, &RMF_SETINFO_KEY);
1584 DEBUG_REQ(D_HA, req, "no set_info key");
1585 RETURN(err_serious(-EFAULT));
1588 keylen = req_capsule_get_size(tsi->tsi_pill, &RMF_SETINFO_KEY,
1591 val = req_capsule_client_get(tsi->tsi_pill, &RMF_SETINFO_VAL);
1593 DEBUG_REQ(D_HA, req, "no set_info val");
1594 RETURN(err_serious(-EFAULT));
1597 vallen = req_capsule_get_size(tsi->tsi_pill, &RMF_SETINFO_VAL,
1600 /* Swab any part of val you need to here */
1601 if (KEY_IS(KEY_READ_ONLY)) {
1602 spin_lock(&req->rq_export->exp_lock);
1604 *exp_connect_flags_ptr(req->rq_export) |=
1607 *exp_connect_flags_ptr(req->rq_export) &=
1608 ~OBD_CONNECT_RDONLY;
1609 spin_unlock(&req->rq_export->exp_lock);
1610 } else if (KEY_IS(KEY_CHANGELOG_CLEAR)) {
1611 struct changelog_setinfo *cs = val;
1613 if (vallen != sizeof(*cs)) {
1614 CERROR("%s: bad changelog_clear setinfo size %d\n",
1615 tgt_name(tsi->tsi_tgt), vallen);
1618 if (ptlrpc_req_need_swab(req)) {
1619 __swab64s(&cs->cs_recno);
1620 __swab32s(&cs->cs_id);
1623 rc = mdt_iocontrol(OBD_IOC_CHANGELOG_CLEAR, req->rq_export,
1631 int mdt_readpage(struct tgt_session_info *tsi)
1633 struct mdt_thread_info *info = mdt_th_info(tsi->tsi_env);
1634 struct mdt_object *object = mdt_obj(tsi->tsi_corpus);
1635 struct lu_rdpg *rdpg = &info->mti_u.rdpg.mti_rdpg;
1636 const struct mdt_body *reqbody = tsi->tsi_mdt_body;
1637 struct mdt_body *repbody;
1643 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK))
1644 RETURN(err_serious(-ENOMEM));
1646 repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_MDT_BODY);
1647 if (repbody == NULL || reqbody == NULL)
1648 RETURN(err_serious(-EFAULT));
1651 * prepare @rdpg before calling lower layers and transfer itself. Here
1652 * reqbody->size contains offset of where to start to read and
1653 * reqbody->nlink contains number bytes to read.
1655 rdpg->rp_hash = reqbody->size;
1656 if (rdpg->rp_hash != reqbody->size) {
1657 CERROR("Invalid hash: "LPX64" != "LPX64"\n",
1658 rdpg->rp_hash, reqbody->size);
1662 rdpg->rp_attrs = reqbody->mode;
1663 if (exp_connect_flags(tsi->tsi_exp) & OBD_CONNECT_64BITHASH)
1664 rdpg->rp_attrs |= LUDA_64BITHASH;
1665 rdpg->rp_count = min_t(unsigned int, reqbody->nlink,
1666 exp_max_brw_size(tsi->tsi_exp));
1667 rdpg->rp_npages = (rdpg->rp_count + PAGE_CACHE_SIZE - 1) >>
1669 OBD_ALLOC(rdpg->rp_pages, rdpg->rp_npages * sizeof rdpg->rp_pages[0]);
1670 if (rdpg->rp_pages == NULL)
1673 for (i = 0; i < rdpg->rp_npages; ++i) {
1674 rdpg->rp_pages[i] = alloc_page(GFP_IOFS);
1675 if (rdpg->rp_pages[i] == NULL)
1676 GOTO(free_rdpg, rc = -ENOMEM);
1679 /* call lower layers to fill allocated pages with directory data */
1680 rc = mo_readpage(tsi->tsi_env, mdt_object_child(object), rdpg);
1682 GOTO(free_rdpg, rc);
1684 /* send pages to client */
1685 rc = tgt_sendpage(tsi, rdpg, rc);
1690 for (i = 0; i < rdpg->rp_npages; i++)
1691 if (rdpg->rp_pages[i] != NULL)
1692 __free_page(rdpg->rp_pages[i]);
1693 OBD_FREE(rdpg->rp_pages, rdpg->rp_npages * sizeof rdpg->rp_pages[0]);
1695 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE))
1701 static int mdt_reint_internal(struct mdt_thread_info *info,
1702 struct mdt_lock_handle *lhc,
1705 struct req_capsule *pill = info->mti_pill;
1706 struct mdt_body *repbody;
1711 rc = mdt_reint_unpack(info, op);
1713 CERROR("Can't unpack reint, rc %d\n", rc);
1714 RETURN(err_serious(rc));
1717 /* for replay (no_create) lmm is not needed, client has it already */
1718 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
1719 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER,
1720 info->mti_rr.rr_eadatalen);
1722 /* llog cookies are always 0, the field is kept for compatibility */
1723 if (req_capsule_has_field(pill, &RMF_LOGCOOKIES, RCL_SERVER))
1724 req_capsule_set_size(pill, &RMF_LOGCOOKIES, RCL_SERVER, 0);
1726 rc = req_capsule_server_pack(pill);
1728 CERROR("Can't pack response, rc %d\n", rc);
1729 RETURN(err_serious(rc));
1732 if (req_capsule_has_field(pill, &RMF_MDT_BODY, RCL_SERVER)) {
1733 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
1735 repbody->eadatasize = 0;
1736 repbody->aclsize = 0;
1739 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_REINT_DELAY, 10);
1741 /* for replay no cookkie / lmm need, because client have this already */
1742 if (info->mti_spec.no_create)
1743 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
1744 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER, 0);
1746 rc = mdt_init_ucred_reint(info);
1748 GOTO(out_shrink, rc);
1750 rc = mdt_fix_attr_ucred(info, op);
1752 GOTO(out_ucred, rc = err_serious(rc));
1754 if (mdt_check_resent(info, mdt_reconstruct, lhc)) {
1755 rc = lustre_msg_get_status(mdt_info_req(info)->rq_repmsg);
1756 GOTO(out_ucred, rc);
1758 rc = mdt_reint_rec(info, lhc);
1761 mdt_exit_ucred(info);
1763 mdt_client_compatibility(info);
1764 rc2 = mdt_fix_reply(info);
1770 static long mdt_reint_opcode(struct ptlrpc_request *req,
1771 const struct req_format **fmt)
1773 struct mdt_device *mdt;
1774 struct mdt_rec_reint *rec;
1777 rec = req_capsule_client_get(&req->rq_pill, &RMF_REC_REINT);
1779 opc = rec->rr_opcode;
1780 DEBUG_REQ(D_INODE, req, "reint opt = %ld", opc);
1781 if (opc < REINT_MAX && fmt[opc] != NULL)
1782 req_capsule_extend(&req->rq_pill, fmt[opc]);
1784 mdt = mdt_exp2dev(req->rq_export);
1785 CERROR("%s: Unsupported opcode '%ld' from client '%s':"
1786 " rc = %d\n", req->rq_export->exp_obd->obd_name,
1787 opc, mdt->mdt_ldlm_client->cli_name, -EFAULT);
1788 opc = err_serious(-EFAULT);
1791 opc = err_serious(-EFAULT);
1796 int mdt_reint(struct tgt_session_info *tsi)
1800 static const struct req_format *reint_fmts[REINT_MAX] = {
1801 [REINT_SETATTR] = &RQF_MDS_REINT_SETATTR,
1802 [REINT_CREATE] = &RQF_MDS_REINT_CREATE,
1803 [REINT_LINK] = &RQF_MDS_REINT_LINK,
1804 [REINT_UNLINK] = &RQF_MDS_REINT_UNLINK,
1805 [REINT_RENAME] = &RQF_MDS_REINT_RENAME,
1806 [REINT_OPEN] = &RQF_MDS_REINT_OPEN,
1807 [REINT_SETXATTR] = &RQF_MDS_REINT_SETXATTR,
1808 [REINT_RMENTRY] = &RQF_MDS_REINT_UNLINK
1813 opc = mdt_reint_opcode(tgt_ses_req(tsi), reint_fmts);
1815 struct mdt_thread_info *info = tsi2mdt_info(tsi);
1817 * No lock possible here from client to pass it to reint code
1820 rc = mdt_reint_internal(info, NULL, opc);
1821 mdt_thread_info_fini(info);
1826 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_REINT_NET_REP;
1830 /* this should sync the whole device */
1831 static int mdt_device_sync(const struct lu_env *env, struct mdt_device *mdt)
1833 struct dt_device *dt = mdt->mdt_bottom;
1837 rc = dt->dd_ops->dt_sync(env, dt);
1841 /* this should sync this object */
1842 static int mdt_object_sync(struct mdt_thread_info *info)
1844 struct md_object *next;
1848 if (!mdt_object_exists(info->mti_object)) {
1849 CWARN("Non existing object "DFID"!\n",
1850 PFID(mdt_object_fid(info->mti_object)));
1853 next = mdt_object_child(info->mti_object);
1854 rc = mo_object_sync(info->mti_env, next);
1859 int mdt_sync(struct tgt_session_info *tsi)
1861 struct ptlrpc_request *req = tgt_ses_req(tsi);
1862 struct req_capsule *pill = tsi->tsi_pill;
1863 struct mdt_body *body;
1868 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SYNC_PACK))
1869 RETURN(err_serious(-ENOMEM));
1871 if (fid_seq(&tsi->tsi_mdt_body->fid1) == 0) {
1872 rc = mdt_device_sync(tsi->tsi_env, mdt_exp2dev(tsi->tsi_exp));
1874 struct mdt_thread_info *info = tsi2mdt_info(tsi);
1876 /* sync an object */
1877 rc = mdt_object_sync(info);
1879 const struct lu_fid *fid;
1880 struct lu_attr *la = &info->mti_attr.ma_attr;
1882 info->mti_attr.ma_need = MA_INODE;
1883 info->mti_attr.ma_valid = 0;
1884 rc = mdt_attr_get_complex(info, info->mti_object,
1887 body = req_capsule_server_get(pill,
1889 fid = mdt_object_fid(info->mti_object);
1890 mdt_pack_attr2body(info, body, la, fid);
1893 mdt_thread_info_fini(info);
1896 mdt_counter_incr(req, LPROC_MDT_SYNC);
1902 * Handle quota control requests to consult current usage/limit, but also
1903 * to configure quota enforcement
1905 int mdt_quotactl(struct tgt_session_info *tsi)
1907 struct obd_export *exp = tsi->tsi_exp;
1908 struct req_capsule *pill = tsi->tsi_pill;
1909 struct obd_quotactl *oqctl, *repoqc;
1911 struct mdt_device *mdt = mdt_exp2dev(exp);
1912 struct lu_device *qmt = mdt->mdt_qmt_dev;
1915 oqctl = req_capsule_client_get(pill, &RMF_OBD_QUOTACTL);
1917 RETURN(err_serious(-EPROTO));
1919 rc = req_capsule_server_pack(pill);
1921 RETURN(err_serious(rc));
1923 switch (oqctl->qc_cmd) {
1925 case LUSTRE_Q_INVALIDATE:
1926 case LUSTRE_Q_FINVALIDATE:
1930 /* deprecated, not used any more */
1931 RETURN(-EOPNOTSUPP);
1932 /* master quotactl */
1938 RETURN(-EOPNOTSUPP);
1939 /* slave quotactl */
1944 CERROR("Unsupported quotactl command: %d\n", oqctl->qc_cmd);
1948 /* map uid/gid for remote client */
1950 if (exp_connect_rmtclient(exp)) {
1951 struct lustre_idmap_table *idmap;
1953 idmap = exp->exp_mdt_data.med_idmap;
1955 if (unlikely(oqctl->qc_cmd != Q_GETQUOTA &&
1956 oqctl->qc_cmd != Q_GETINFO))
1959 if (oqctl->qc_type == USRQUOTA)
1960 id = lustre_idmap_lookup_uid(NULL, idmap, 0,
1962 else if (oqctl->qc_type == GRPQUOTA)
1963 id = lustre_idmap_lookup_gid(NULL, idmap, 0,
1968 if (id == CFS_IDMAP_NOTFOUND) {
1969 CDEBUG(D_QUOTA, "no mapping for id %u\n", oqctl->qc_id);
1974 repoqc = req_capsule_server_get(pill, &RMF_OBD_QUOTACTL);
1976 RETURN(err_serious(-EFAULT));
1978 if (oqctl->qc_id != id)
1979 swap(oqctl->qc_id, id);
1981 switch (oqctl->qc_cmd) {
1987 /* forward quotactl request to QMT */
1988 rc = qmt_hdls.qmth_quotactl(tsi->tsi_env, qmt, oqctl);
1993 /* slave quotactl */
1994 rc = lquotactl_slv(tsi->tsi_env, tsi->tsi_tgt->lut_bottom,
1999 CERROR("Unsupported quotactl command: %d\n", oqctl->qc_cmd);
2003 if (oqctl->qc_id != id)
2004 swap(oqctl->qc_id, id);
2010 /** clone llog ctxt from child (mdd)
2011 * This allows remote llog (replicator) access.
2012 * We can either pass all llog RPCs (eg mdt_llog_create) on to child where the
2013 * context was originally set up, or we can handle them directly.
2014 * I choose the latter, but that means I need any llog
2015 * contexts set up by child to be accessable by the mdt. So we clone the
2016 * context into our context list here.
2018 static int mdt_llog_ctxt_clone(const struct lu_env *env, struct mdt_device *mdt,
2021 struct md_device *next = mdt->mdt_child;
2022 struct llog_ctxt *ctxt;
2025 if (!llog_ctxt_null(mdt2obd_dev(mdt), idx))
2028 rc = next->md_ops->mdo_llog_ctxt_get(env, next, idx, (void **)&ctxt);
2029 if (rc || ctxt == NULL) {
2033 rc = llog_group_set_ctxt(&mdt2obd_dev(mdt)->obd_olg, ctxt, idx);
2035 CERROR("Can't set mdt ctxt %d\n", rc);
2040 static int mdt_llog_ctxt_unclone(const struct lu_env *env,
2041 struct mdt_device *mdt, int idx)
2043 struct llog_ctxt *ctxt;
2045 ctxt = llog_get_context(mdt2obd_dev(mdt), idx);
2048 /* Put once for the get we just did, and once for the clone */
2049 llog_ctxt_put(ctxt);
2050 llog_ctxt_put(ctxt);
2055 * sec context handlers
2057 int mdt_sec_ctx_handle(struct tgt_session_info *tsi)
2061 rc = mdt_handle_idmap(tsi);
2063 struct ptlrpc_request *req = tgt_ses_req(tsi);
2066 opc = lustre_msg_get_opc(req->rq_reqmsg);
2067 if (opc == SEC_CTX_INIT || opc == SEC_CTX_INIT_CONT)
2068 sptlrpc_svc_ctx_invalidate(req);
2071 CFS_FAIL_TIMEOUT(OBD_FAIL_SEC_CTX_HDL_PAUSE, cfs_fail_val);
2077 * quota request handlers
2079 int mdt_quota_dqacq(struct tgt_session_info *tsi)
2081 struct mdt_device *mdt = mdt_exp2dev(tsi->tsi_exp);
2082 struct lu_device *qmt = mdt->mdt_qmt_dev;
2087 RETURN(err_serious(-EOPNOTSUPP));
2089 rc = qmt_hdls.qmth_dqacq(tsi->tsi_env, qmt, tgt_ses_req(tsi));
2093 struct mdt_object *mdt_object_new(const struct lu_env *env,
2094 struct mdt_device *d,
2095 const struct lu_fid *f)
2097 struct lu_object_conf conf = { .loc_flags = LOC_F_NEW };
2098 struct lu_object *o;
2099 struct mdt_object *m;
2102 CDEBUG(D_INFO, "Allocate object for "DFID"\n", PFID(f));
2103 o = lu_object_find(env, &d->mdt_lu_dev, f, &conf);
2104 if (unlikely(IS_ERR(o)))
2105 m = (struct mdt_object *)o;
2111 struct mdt_object *mdt_object_find(const struct lu_env *env,
2112 struct mdt_device *d,
2113 const struct lu_fid *f)
2115 struct lu_object *o;
2116 struct mdt_object *m;
2119 CDEBUG(D_INFO, "Find object for "DFID"\n", PFID(f));
2120 o = lu_object_find(env, &d->mdt_lu_dev, f, NULL);
2121 if (unlikely(IS_ERR(o)))
2122 m = (struct mdt_object *)o;
2130 * Asyncronous commit for mdt device.
2132 * Pass asynchonous commit call down the MDS stack.
2134 * \param env environment
2135 * \param mdt the mdt device
2137 static void mdt_device_commit_async(const struct lu_env *env,
2138 struct mdt_device *mdt)
2140 struct dt_device *dt = mdt->mdt_bottom;
2143 rc = dt->dd_ops->dt_commit_async(env, dt);
2144 if (unlikely(rc != 0))
2145 CWARN("async commit start failed with rc = %d", rc);
2149 * Mark the lock as "synchonous".
2151 * Mark the lock to deffer transaction commit to the unlock time.
2153 * \param lock the lock to mark as "synchonous"
2155 * \see mdt_is_lock_sync
2156 * \see mdt_save_lock
2158 static inline void mdt_set_lock_sync(struct ldlm_lock *lock)
2160 lock->l_ast_data = (void*)1;
2164 * Check whehter the lock "synchonous" or not.
2166 * \param lock the lock to check
2167 * \retval 1 the lock is "synchonous"
2168 * \retval 0 the lock isn't "synchronous"
2170 * \see mdt_set_lock_sync
2171 * \see mdt_save_lock
2173 static inline int mdt_is_lock_sync(struct ldlm_lock *lock)
2175 return lock->l_ast_data != NULL;
2179 * Blocking AST for mdt locks.
2181 * Starts transaction commit if in case of COS lock conflict or
2182 * deffers such a commit to the mdt_save_lock.
2184 * \param lock the lock which blocks a request or cancelling lock
2185 * \param desc unused
2186 * \param data unused
2187 * \param flag indicates whether this cancelling or blocking callback
2189 * \see ldlm_blocking_ast_nocheck
2191 int mdt_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
2192 void *data, int flag)
2194 struct obd_device *obd = ldlm_lock_to_ns(lock)->ns_obd;
2195 struct mdt_device *mdt = mdt_dev(obd->obd_lu_dev);
2199 if (flag == LDLM_CB_CANCELING)
2201 lock_res_and_lock(lock);
2202 if (lock->l_blocking_ast != mdt_blocking_ast) {
2203 unlock_res_and_lock(lock);
2206 if (mdt_cos_is_enabled(mdt) &&
2207 lock->l_req_mode & (LCK_PW | LCK_EX) &&
2208 lock->l_blocking_lock != NULL &&
2209 lock->l_client_cookie != lock->l_blocking_lock->l_client_cookie) {
2210 mdt_set_lock_sync(lock);
2212 rc = ldlm_blocking_ast_nocheck(lock);
2214 /* There is no lock conflict if l_blocking_lock == NULL,
2215 * it indicates a blocking ast sent from ldlm_lock_decref_internal
2216 * when the last reference to a local lock was released */
2217 if (lock->l_req_mode == LCK_COS && lock->l_blocking_lock != NULL) {
2220 rc = lu_env_init(&env, LCT_LOCAL);
2221 if (unlikely(rc != 0))
2222 CWARN("lu_env initialization failed with rc = %d,"
2223 "cannot start asynchronous commit\n", rc);
2225 mdt_device_commit_async(&env, mdt);
2231 int mdt_md_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
2232 void *data, int flag)
2234 struct lustre_handle lockh;
2238 case LDLM_CB_BLOCKING:
2239 ldlm_lock2handle(lock, &lockh);
2240 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
2242 CDEBUG(D_INODE, "ldlm_cli_cancel: %d\n", rc);
2246 case LDLM_CB_CANCELING:
2247 LDLM_DEBUG(lock, "Revoke remote lock\n");
2255 int mdt_remote_object_lock(struct mdt_thread_info *mti,
2256 struct mdt_object *o, struct lustre_handle *lh,
2257 ldlm_mode_t mode, __u64 ibits)
2259 struct ldlm_enqueue_info *einfo = &mti->mti_einfo;
2260 ldlm_policy_data_t *policy = &mti->mti_policy;
2264 LASSERT(mdt_object_remote(o));
2266 LASSERT(ibits & MDS_INODELOCK_UPDATE);
2268 memset(einfo, 0, sizeof(*einfo));
2269 einfo->ei_type = LDLM_IBITS;
2270 einfo->ei_mode = mode;
2271 einfo->ei_cb_bl = mdt_md_blocking_ast;
2272 einfo->ei_cb_cp = ldlm_completion_ast;
2274 memset(policy, 0, sizeof(*policy));
2275 policy->l_inodebits.bits = ibits;
2277 rc = mo_object_lock(mti->mti_env, mdt_object_child(o), lh, einfo,
2282 static int mdt_object_lock0(struct mdt_thread_info *info, struct mdt_object *o,
2283 struct mdt_lock_handle *lh, __u64 ibits,
2284 bool nonblock, int locality)
2286 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
2287 ldlm_policy_data_t *policy = &info->mti_policy;
2288 struct ldlm_res_id *res_id = &info->mti_res_id;
2293 LASSERT(!lustre_handle_is_used(&lh->mlh_reg_lh));
2294 LASSERT(!lustre_handle_is_used(&lh->mlh_pdo_lh));
2295 LASSERT(lh->mlh_reg_mode != LCK_MINMODE);
2296 LASSERT(lh->mlh_type != MDT_NUL_LOCK);
2298 if (mdt_object_remote(o)) {
2299 if (locality == MDT_CROSS_LOCK) {
2300 ibits &= ~(MDS_INODELOCK_UPDATE | MDS_INODELOCK_PERM |
2301 MDS_INODELOCK_LAYOUT);
2302 ibits |= MDS_INODELOCK_LOOKUP;
2305 (MDS_INODELOCK_UPDATE | MDS_INODELOCK_PERM |
2306 MDS_INODELOCK_LAYOUT)),
2307 "%s: wrong bit "LPX64" for remote obj "DFID"\n",
2308 mdt_obd_name(info->mti_mdt), ibits,
2309 PFID(mdt_object_fid(o)));
2310 LASSERT(ibits & MDS_INODELOCK_LOOKUP);
2312 /* No PDO lock on remote object */
2313 LASSERT(lh->mlh_type != MDT_PDO_LOCK);
2316 if (lh->mlh_type == MDT_PDO_LOCK) {
2317 /* check for exists after object is locked */
2318 if (mdt_object_exists(o) == 0) {
2319 /* Non-existent object shouldn't have PDO lock */
2322 /* Non-dir object shouldn't have PDO lock */
2323 if (!S_ISDIR(lu_object_attr(&o->mot_obj)))
2328 memset(policy, 0, sizeof(*policy));
2329 fid_build_reg_res_name(mdt_object_fid(o), res_id);
2331 dlmflags = LDLM_FL_ATOMIC_CB;
2333 dlmflags |= LDLM_FL_BLOCK_NOWAIT;
2336 * Take PDO lock on whole directory and build correct @res_id for lock
2337 * on part of directory.
2339 if (lh->mlh_pdo_hash != 0) {
2340 LASSERT(lh->mlh_type == MDT_PDO_LOCK);
2341 mdt_lock_pdo_mode(info, o, lh);
2342 if (lh->mlh_pdo_mode != LCK_NL) {
2344 * Do not use LDLM_FL_LOCAL_ONLY for parallel lock, it
2345 * is never going to be sent to client and we do not
2346 * want it slowed down due to possible cancels.
2348 policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
2349 rc = mdt_fid_lock(ns, &lh->mlh_pdo_lh, lh->mlh_pdo_mode,
2350 policy, res_id, dlmflags,
2351 &info->mti_exp->exp_handle.h_cookie);
2357 * Finish res_id initializing by name hash marking part of
2358 * directory which is taking modification.
2360 res_id->name[LUSTRE_RES_ID_HSH_OFF] = lh->mlh_pdo_hash;
2363 policy->l_inodebits.bits = ibits;
2366 * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
2367 * going to be sent to client. If it is - mdt_intent_policy() path will
2368 * fix it up and turn FL_LOCAL flag off.
2370 rc = mdt_fid_lock(ns, &lh->mlh_reg_lh, lh->mlh_reg_mode, policy,
2371 res_id, LDLM_FL_LOCAL_ONLY | dlmflags,
2372 &info->mti_exp->exp_handle.h_cookie);
2374 mdt_object_unlock(info, o, lh, 1);
2375 else if (unlikely(OBD_FAIL_PRECHECK(OBD_FAIL_MDS_PDO_LOCK)) &&
2376 lh->mlh_pdo_hash != 0 &&
2377 (lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX)) {
2378 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK, 15);
2384 int mdt_object_lock(struct mdt_thread_info *info, struct mdt_object *o,
2385 struct mdt_lock_handle *lh, __u64 ibits, int locality)
2387 return mdt_object_lock0(info, o, lh, ibits, false, locality);
2390 int mdt_object_lock_try(struct mdt_thread_info *info, struct mdt_object *o,
2391 struct mdt_lock_handle *lh, __u64 ibits, int locality)
2393 struct mdt_lock_handle tmp = *lh;
2396 rc = mdt_object_lock0(info, o, &tmp, ibits, true, locality);
2404 * Save a lock within request object.
2406 * Keep the lock referenced until whether client ACK or transaction
2407 * commit happens or release the lock immediately depending on input
2408 * parameters. If COS is ON, a write lock is converted to COS lock
2411 * \param info thead info object
2412 * \param h lock handle
2413 * \param mode lock mode
2414 * \param decref force immediate lock releasing
2417 void mdt_save_lock(struct mdt_thread_info *info, struct lustre_handle *h,
2418 ldlm_mode_t mode, int decref)
2422 if (lustre_handle_is_used(h)) {
2423 if (decref || !info->mti_has_trans ||
2424 !(mode & (LCK_PW | LCK_EX))){
2425 mdt_fid_unlock(h, mode);
2427 struct mdt_device *mdt = info->mti_mdt;
2428 struct ldlm_lock *lock = ldlm_handle2lock(h);
2429 struct ptlrpc_request *req = mdt_info_req(info);
2432 LASSERTF(lock != NULL, "no lock for cookie "LPX64"\n",
2434 /* there is no request if mdt_object_unlock() is called
2435 * from mdt_export_cleanup()->mdt_add_dirty_flag() */
2436 if (likely(req != NULL)) {
2437 CDEBUG(D_HA, "request = %p reply state = %p"
2438 " transno = "LPD64"\n", req,
2439 req->rq_reply_state, req->rq_transno);
2440 if (mdt_cos_is_enabled(mdt)) {
2442 ldlm_lock_downgrade(lock, LCK_COS);
2445 ptlrpc_save_lock(req, h, mode, no_ack);
2447 ldlm_lock_decref(h, mode);
2449 if (mdt_is_lock_sync(lock)) {
2450 CDEBUG(D_HA, "found sync-lock,"
2451 " async commit started\n");
2452 mdt_device_commit_async(info->mti_env,
2455 LDLM_LOCK_PUT(lock);
2464 * Unlock mdt object.
2466 * Immeditely release the regular lock and the PDO lock or save the
2467 * lock in reqeuest and keep them referenced until client ACK or
2468 * transaction commit.
2470 * \param info thread info object
2471 * \param o mdt object
2472 * \param lh mdt lock handle referencing regular and PDO locks
2473 * \param decref force immediate lock releasing
2475 void mdt_object_unlock(struct mdt_thread_info *info, struct mdt_object *o,
2476 struct mdt_lock_handle *lh, int decref)
2480 mdt_save_lock(info, &lh->mlh_pdo_lh, lh->mlh_pdo_mode, decref);
2481 mdt_save_lock(info, &lh->mlh_reg_lh, lh->mlh_reg_mode, decref);
2483 if (lustre_handle_is_used(&lh->mlh_rreg_lh))
2484 ldlm_lock_decref(&lh->mlh_rreg_lh, lh->mlh_rreg_mode);
2489 struct mdt_object *mdt_object_find_lock(struct mdt_thread_info *info,
2490 const struct lu_fid *f,
2491 struct mdt_lock_handle *lh,
2494 struct mdt_object *o;
2496 o = mdt_object_find(info->mti_env, info->mti_mdt, f);
2500 rc = mdt_object_lock(info, o, lh, ibits,
2503 mdt_object_put(info->mti_env, o);
2510 void mdt_object_unlock_put(struct mdt_thread_info * info,
2511 struct mdt_object * o,
2512 struct mdt_lock_handle *lh,
2515 mdt_object_unlock(info, o, lh, decref);
2516 mdt_object_put(info->mti_env, o);
2520 * Generic code handling requests that have struct mdt_body passed in:
2522 * - extract mdt_body from request and save it in @info, if present;
2524 * - create lu_object, corresponding to the fid in mdt_body, and save it in
2527 * - if HABEO_CORPUS flag is set for this request type check whether object
2528 * actually exists on storage (lu_object_exists()).
2531 static int mdt_body_unpack(struct mdt_thread_info *info, __u32 flags)
2533 const struct mdt_body *body;
2534 struct mdt_object *obj;
2535 const struct lu_env *env;
2536 struct req_capsule *pill;
2540 env = info->mti_env;
2541 pill = info->mti_pill;
2543 body = info->mti_body = req_capsule_client_get(pill, &RMF_MDT_BODY);
2547 if (!(body->valid & OBD_MD_FLID))
2550 if (!fid_is_sane(&body->fid1)) {
2551 CERROR("Invalid fid: "DFID"\n", PFID(&body->fid1));
2556 * Do not get size or any capa fields before we check that request
2557 * contains capa actually. There are some requests which do not, for
2558 * instance MDS_IS_SUBDIR.
2560 if (req_capsule_has_field(pill, &RMF_CAPA1, RCL_CLIENT) &&
2561 req_capsule_get_size(pill, &RMF_CAPA1, RCL_CLIENT))
2562 mdt_set_capainfo(info, 0, &body->fid1,
2563 req_capsule_client_get(pill, &RMF_CAPA1));
2565 obj = mdt_object_find(env, info->mti_mdt, &body->fid1);
2567 if ((flags & HABEO_CORPUS) &&
2568 !mdt_object_exists(obj)) {
2569 mdt_object_put(env, obj);
2570 /* for capability renew ENOENT will be handled in
2572 if (body->valid & OBD_MD_FLOSSCAPA)
2577 info->mti_object = obj;
2586 static int mdt_unpack_req_pack_rep(struct mdt_thread_info *info, __u32 flags)
2588 struct req_capsule *pill = info->mti_pill;
2592 if (req_capsule_has_field(pill, &RMF_MDT_BODY, RCL_CLIENT))
2593 rc = mdt_body_unpack(info, flags);
2597 if (rc == 0 && (flags & HABEO_REFERO)) {
2599 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
2600 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER,
2601 info->mti_body->eadatasize);
2602 if (req_capsule_has_field(pill, &RMF_LOGCOOKIES, RCL_SERVER))
2603 req_capsule_set_size(pill, &RMF_LOGCOOKIES,
2606 rc = req_capsule_server_pack(pill);
2611 static int mdt_init_capa_ctxt(const struct lu_env *env, struct mdt_device *m)
2613 struct md_device *next = m->mdt_child;
2615 return next->md_ops->mdo_init_capa_ctxt(env, next,
2616 m->mdt_lut.lut_mds_capa,
2617 m->mdt_capa_timeout,
2622 void mdt_lock_handle_init(struct mdt_lock_handle *lh)
2624 lh->mlh_type = MDT_NUL_LOCK;
2625 lh->mlh_reg_lh.cookie = 0ull;
2626 lh->mlh_reg_mode = LCK_MINMODE;
2627 lh->mlh_pdo_lh.cookie = 0ull;
2628 lh->mlh_pdo_mode = LCK_MINMODE;
2629 lh->mlh_rreg_lh.cookie = 0ull;
2630 lh->mlh_rreg_mode = LCK_MINMODE;
2633 void mdt_lock_handle_fini(struct mdt_lock_handle *lh)
2635 LASSERT(!lustre_handle_is_used(&lh->mlh_reg_lh));
2636 LASSERT(!lustre_handle_is_used(&lh->mlh_pdo_lh));
2640 * Initialize fields of struct mdt_thread_info. Other fields are left in
2641 * uninitialized state, because it's too expensive to zero out whole
2642 * mdt_thread_info (> 1K) on each request arrival.
2644 void mdt_thread_info_init(struct ptlrpc_request *req,
2645 struct mdt_thread_info *info)
2649 info->mti_pill = &req->rq_pill;
2652 for (i = 0; i < ARRAY_SIZE(info->mti_lh); i++)
2653 mdt_lock_handle_init(&info->mti_lh[i]);
2655 /* mdt device: it can be NULL while CONNECT */
2656 if (req->rq_export) {
2657 info->mti_mdt = mdt_dev(req->rq_export->exp_obd->obd_lu_dev);
2658 info->mti_exp = req->rq_export;
2660 info->mti_mdt = NULL;
2661 info->mti_env = req->rq_svc_thread->t_env;
2662 info->mti_transno = lustre_msg_get_transno(req->rq_reqmsg);
2664 memset(&info->mti_attr, 0, sizeof(info->mti_attr));
2665 info->mti_big_buf = LU_BUF_NULL;
2666 info->mti_body = NULL;
2667 info->mti_object = NULL;
2668 info->mti_dlm_req = NULL;
2669 info->mti_has_trans = 0;
2670 info->mti_cross_ref = 0;
2671 info->mti_opdata = 0;
2672 info->mti_big_lmm_used = 0;
2674 /* To not check for split by default. */
2675 info->mti_spec.no_create = 0;
2676 info->mti_spec.sp_rm_entry = 0;
2679 void mdt_thread_info_fini(struct mdt_thread_info *info)
2683 if (info->mti_object != NULL) {
2684 mdt_object_put(info->mti_env, info->mti_object);
2685 info->mti_object = NULL;
2688 for (i = 0; i < ARRAY_SIZE(info->mti_lh); i++)
2689 mdt_lock_handle_fini(&info->mti_lh[i]);
2690 info->mti_env = NULL;
2691 info->mti_pill = NULL;
2692 info->mti_exp = NULL;
2694 if (unlikely(info->mti_big_buf.lb_buf != NULL))
2695 lu_buf_free(&info->mti_big_buf);
2698 struct mdt_thread_info *tsi2mdt_info(struct tgt_session_info *tsi)
2700 struct mdt_thread_info *mti;
2701 struct lustre_capa *lc;
2703 mti = mdt_th_info(tsi->tsi_env);
2704 LASSERT(mti != NULL);
2706 mdt_thread_info_init(tgt_ses_req(tsi), mti);
2707 if (tsi->tsi_corpus != NULL) {
2708 struct req_capsule *pill = tsi->tsi_pill;
2710 mti->mti_object = mdt_obj(tsi->tsi_corpus);
2711 lu_object_get(tsi->tsi_corpus);
2714 * XXX: must be part of tgt_mdt_body_unpack but moved here
2715 * due to mdt_set_capainfo().
2717 if (req_capsule_has_field(pill, &RMF_CAPA1, RCL_CLIENT) &&
2718 req_capsule_get_size(pill, &RMF_CAPA1, RCL_CLIENT) > 0) {
2719 lc = req_capsule_client_get(pill, &RMF_CAPA1);
2720 mdt_set_capainfo(mti, 0, &tsi->tsi_mdt_body->fid1, lc);
2723 mti->mti_body = tsi->tsi_mdt_body;
2724 mti->mti_dlm_req = tsi->tsi_dlm_req;
2729 int mdt_tgt_connect(struct tgt_session_info *tsi)
2731 struct ptlrpc_request *req = tgt_ses_req(tsi);
2736 rc = tgt_connect(tsi);
2740 rc = mdt_init_idmap(tsi);
2745 obd_disconnect(class_export_get(req->rq_export));
2764 static int mdt_intent_getattr(enum mdt_it_code opcode,
2765 struct mdt_thread_info *info,
2766 struct ldlm_lock **,
2769 static int mdt_intent_getxattr(enum mdt_it_code opcode,
2770 struct mdt_thread_info *info,
2771 struct ldlm_lock **lockp,
2774 static int mdt_intent_layout(enum mdt_it_code opcode,
2775 struct mdt_thread_info *info,
2776 struct ldlm_lock **,
2778 static int mdt_intent_reint(enum mdt_it_code opcode,
2779 struct mdt_thread_info *info,
2780 struct ldlm_lock **,
2783 static struct mdt_it_flavor {
2784 const struct req_format *it_fmt;
2786 int (*it_act)(enum mdt_it_code ,
2787 struct mdt_thread_info *,
2788 struct ldlm_lock **,
2791 } mdt_it_flavor[] = {
2793 .it_fmt = &RQF_LDLM_INTENT,
2794 /*.it_flags = HABEO_REFERO,*/
2796 .it_act = mdt_intent_reint,
2797 .it_reint = REINT_OPEN
2800 .it_fmt = &RQF_LDLM_INTENT,
2802 * OCREAT is not a MUTABOR request as if the file
2804 * We do the extra check of OBD_CONNECT_RDONLY in
2805 * mdt_reint_open() when we really need to create
2809 .it_act = mdt_intent_reint,
2810 .it_reint = REINT_OPEN
2813 .it_fmt = &RQF_LDLM_INTENT,
2814 .it_flags = MUTABOR,
2815 .it_act = mdt_intent_reint,
2816 .it_reint = REINT_CREATE
2818 [MDT_IT_GETATTR] = {
2819 .it_fmt = &RQF_LDLM_INTENT_GETATTR,
2820 .it_flags = HABEO_REFERO,
2821 .it_act = mdt_intent_getattr
2823 [MDT_IT_READDIR] = {
2829 .it_fmt = &RQF_LDLM_INTENT_GETATTR,
2830 .it_flags = HABEO_REFERO,
2831 .it_act = mdt_intent_getattr
2834 .it_fmt = &RQF_LDLM_INTENT_UNLINK,
2835 .it_flags = MUTABOR,
2837 .it_reint = REINT_UNLINK
2841 .it_flags = MUTABOR,
2844 [MDT_IT_GETXATTR] = {
2845 .it_fmt = &RQF_LDLM_INTENT_GETXATTR,
2846 .it_flags = HABEO_CORPUS,
2847 .it_act = mdt_intent_getxattr
2850 .it_fmt = &RQF_LDLM_INTENT_LAYOUT,
2852 .it_act = mdt_intent_layout
2856 int mdt_intent_lock_replace(struct mdt_thread_info *info,
2857 struct ldlm_lock **lockp,
2858 struct ldlm_lock *new_lock,
2859 struct mdt_lock_handle *lh,
2862 struct ptlrpc_request *req = mdt_info_req(info);
2863 struct ldlm_lock *lock = *lockp;
2866 * Get new lock only for cases when possible resent did not find any
2869 if (new_lock == NULL)
2870 new_lock = ldlm_handle2lock_long(&lh->mlh_reg_lh, 0);
2872 if (new_lock == NULL && (flags & LDLM_FL_INTENT_ONLY)) {
2873 lh->mlh_reg_lh.cookie = 0;
2877 LASSERTF(new_lock != NULL,
2878 "lockh "LPX64"\n", lh->mlh_reg_lh.cookie);
2881 * If we've already given this lock to a client once, then we should
2882 * have no readers or writers. Otherwise, we should have one reader
2883 * _or_ writer ref (which will be zeroed below) before returning the
2886 if (new_lock->l_export == req->rq_export) {
2887 LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
2889 LASSERT(new_lock->l_export == NULL);
2890 LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
2895 if (new_lock->l_export == req->rq_export) {
2897 * Already gave this to the client, which means that we
2898 * reconstructed a reply.
2900 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
2902 lh->mlh_reg_lh.cookie = 0;
2903 RETURN(ELDLM_LOCK_REPLACED);
2907 * Fixup the lock to be given to the client.
2909 lock_res_and_lock(new_lock);
2910 /* Zero new_lock->l_readers and new_lock->l_writers without triggering
2911 * possible blocking AST. */
2912 while (new_lock->l_readers > 0) {
2913 lu_ref_del(&new_lock->l_reference, "reader", new_lock);
2914 lu_ref_del(&new_lock->l_reference, "user", new_lock);
2915 new_lock->l_readers--;
2917 while (new_lock->l_writers > 0) {
2918 lu_ref_del(&new_lock->l_reference, "writer", new_lock);
2919 lu_ref_del(&new_lock->l_reference, "user", new_lock);
2920 new_lock->l_writers--;
2923 new_lock->l_export = class_export_lock_get(req->rq_export, new_lock);
2924 new_lock->l_blocking_ast = lock->l_blocking_ast;
2925 new_lock->l_completion_ast = lock->l_completion_ast;
2926 new_lock->l_remote_handle = lock->l_remote_handle;
2927 new_lock->l_flags &= ~LDLM_FL_LOCAL;
2929 unlock_res_and_lock(new_lock);
2931 cfs_hash_add(new_lock->l_export->exp_lock_hash,
2932 &new_lock->l_remote_handle,
2933 &new_lock->l_exp_hash);
2935 LDLM_LOCK_RELEASE(new_lock);
2936 lh->mlh_reg_lh.cookie = 0;
2938 RETURN(ELDLM_LOCK_REPLACED);
2941 static void mdt_intent_fixup_resent(struct mdt_thread_info *info,
2942 struct ldlm_lock *new_lock,
2943 struct ldlm_lock **old_lock,
2944 struct mdt_lock_handle *lh)
2946 struct ptlrpc_request *req = mdt_info_req(info);
2947 struct obd_export *exp = req->rq_export;
2948 struct lustre_handle remote_hdl;
2949 struct ldlm_request *dlmreq;
2950 struct ldlm_lock *lock;
2952 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
2955 dlmreq = req_capsule_client_get(info->mti_pill, &RMF_DLM_REQ);
2956 remote_hdl = dlmreq->lock_handle[0];
2958 /* In the function below, .hs_keycmp resolves to
2959 * ldlm_export_lock_keycmp() */
2960 /* coverity[overrun-buffer-val] */
2961 lock = cfs_hash_lookup(exp->exp_lock_hash, &remote_hdl);
2963 if (lock != new_lock) {
2964 lh->mlh_reg_lh.cookie = lock->l_handle.h_cookie;
2965 lh->mlh_reg_mode = lock->l_granted_mode;
2967 LDLM_DEBUG(lock, "Restoring lock cookie");
2968 DEBUG_REQ(D_DLMTRACE, req,
2969 "restoring lock cookie "LPX64,
2970 lh->mlh_reg_lh.cookie);
2972 *old_lock = LDLM_LOCK_GET(lock);
2973 cfs_hash_put(exp->exp_lock_hash, &lock->l_exp_hash);
2977 cfs_hash_put(exp->exp_lock_hash, &lock->l_exp_hash);
2981 * If the xid matches, then we know this is a resent request, and allow
2982 * it. (It's probably an OPEN, for which we don't send a lock.
2984 if (req_xid_is_last(req))
2988 * This remote handle isn't enqueued, so we never received or processed
2989 * this request. Clear MSG_RESENT, because it can be handled like any
2990 * normal request now.
2992 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2994 DEBUG_REQ(D_DLMTRACE, req, "no existing lock with rhandle "LPX64,
2998 static int mdt_intent_getxattr(enum mdt_it_code opcode,
2999 struct mdt_thread_info *info,
3000 struct ldlm_lock **lockp,
3003 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
3004 struct ldlm_reply *ldlm_rep = NULL;
3008 * Initialize lhc->mlh_reg_lh either from a previously granted lock
3009 * (for the resend case) or a new lock. Below we will use it to
3010 * replace the original lock.
3012 mdt_intent_fixup_resent(info, *lockp, NULL, lhc);
3013 if (!lustre_handle_is_used(&lhc->mlh_reg_lh)) {
3014 mdt_lock_reg_init(lhc, (*lockp)->l_req_mode);
3015 rc = mdt_object_lock(info, info->mti_object, lhc,
3016 MDS_INODELOCK_XATTR,
3022 grc = mdt_getxattr(info);
3024 rc = mdt_intent_lock_replace(info, lockp, NULL, lhc, flags);
3026 if (mdt_info_req(info)->rq_repmsg != NULL)
3027 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
3028 if (ldlm_rep == NULL)
3029 RETURN(err_serious(-EFAULT));
3031 ldlm_rep->lock_policy_res2 = grc;
3036 static int mdt_intent_getattr(enum mdt_it_code opcode,
3037 struct mdt_thread_info *info,
3038 struct ldlm_lock **lockp,
3041 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
3042 struct ldlm_lock *new_lock = NULL;
3044 struct ldlm_reply *ldlm_rep;
3045 struct ptlrpc_request *req;
3046 struct mdt_body *reqbody;
3047 struct mdt_body *repbody;
3051 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
3054 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
3057 info->mti_cross_ref = !!(reqbody->valid & OBD_MD_FLCROSSREF);
3058 repbody->eadatasize = 0;
3059 repbody->aclsize = 0;
3063 child_bits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
3065 case MDT_IT_GETATTR:
3066 child_bits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE |
3070 CERROR("Unsupported intent (%d)\n", opcode);
3071 GOTO(out_shrink, rc = -EINVAL);
3074 rc = mdt_init_ucred(info, reqbody);
3076 GOTO(out_shrink, rc);
3078 req = info->mti_pill->rc_req;
3079 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
3080 mdt_set_disposition(info, ldlm_rep, DISP_IT_EXECD);
3082 /* Get lock from request for possible resent case. */
3083 mdt_intent_fixup_resent(info, *lockp, &new_lock, lhc);
3085 rc = mdt_getattr_name_lock(info, lhc, child_bits, ldlm_rep);
3086 ldlm_rep->lock_policy_res2 = clear_serious(rc);
3088 if (mdt_get_disposition(ldlm_rep, DISP_LOOKUP_NEG))
3089 ldlm_rep->lock_policy_res2 = 0;
3090 if (!mdt_get_disposition(ldlm_rep, DISP_LOOKUP_POS) ||
3091 ldlm_rep->lock_policy_res2) {
3092 lhc->mlh_reg_lh.cookie = 0ull;
3093 GOTO(out_ucred, rc = ELDLM_LOCK_ABORTED);
3096 rc = mdt_intent_lock_replace(info, lockp, new_lock, lhc, flags);
3099 mdt_exit_ucred(info);
3101 mdt_client_compatibility(info);
3102 rc2 = mdt_fix_reply(info);
3108 static int mdt_intent_layout(enum mdt_it_code opcode,
3109 struct mdt_thread_info *info,
3110 struct ldlm_lock **lockp,
3113 struct layout_intent *layout;
3115 struct mdt_object *obj = NULL;
3116 struct md_object *child = NULL;
3120 if (opcode != MDT_IT_LAYOUT) {
3121 CERROR("%s: Unknown intent (%d)\n", mdt_obd_name(info->mti_mdt),
3126 fid = &info->mti_tmp_fid2;
3127 fid_extract_from_res_name(fid, &(*lockp)->l_resource->lr_name);
3129 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
3131 RETURN(PTR_ERR(obj));
3133 if (mdt_object_exists(obj) && !mdt_object_remote(obj)) {
3134 child = mdt_object_child(obj);
3136 /* get the length of lsm */
3137 rc = mo_xattr_get(info->mti_env, child, &LU_BUF_NULL,
3140 if (rc > info->mti_mdt->mdt_max_mdsize)
3141 info->mti_mdt->mdt_max_mdsize = rc;
3144 mdt_object_put(info->mti_env, obj);
3146 (*lockp)->l_lvb_type = LVB_T_LAYOUT;
3147 req_capsule_set_size(info->mti_pill, &RMF_DLM_LVB, RCL_SERVER,
3148 ldlm_lvbo_size(*lockp));
3149 rc = req_capsule_server_pack(info->mti_pill);
3153 layout = req_capsule_client_get(info->mti_pill, &RMF_LAYOUT_INTENT);
3154 LASSERT(layout != NULL);
3155 if (layout->li_opc == LAYOUT_INTENT_ACCESS)
3156 /* return to normal ldlm handling */
3159 CERROR("%s: Unsupported layout intent (%d)\n",
3160 mdt_obd_name(info->mti_mdt), layout->li_opc);
3164 static int mdt_intent_reint(enum mdt_it_code opcode,
3165 struct mdt_thread_info *info,
3166 struct ldlm_lock **lockp,
3169 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
3170 struct ldlm_reply *rep = NULL;
3174 static const struct req_format *intent_fmts[REINT_MAX] = {
3175 [REINT_CREATE] = &RQF_LDLM_INTENT_CREATE,
3176 [REINT_OPEN] = &RQF_LDLM_INTENT_OPEN
3181 opc = mdt_reint_opcode(mdt_info_req(info), intent_fmts);
3185 if (mdt_it_flavor[opcode].it_reint != opc) {
3186 CERROR("Reint code %ld doesn't match intent: %d\n",
3188 RETURN(err_serious(-EPROTO));
3191 /* Get lock from request for possible resent case. */
3192 mdt_intent_fixup_resent(info, *lockp, NULL, lhc);
3194 rc = mdt_reint_internal(info, lhc, opc);
3196 /* Check whether the reply has been packed successfully. */
3197 if (mdt_info_req(info)->rq_repmsg != NULL)
3198 rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
3200 RETURN(err_serious(-EFAULT));
3202 /* MDC expects this in any case */
3204 mdt_set_disposition(info, rep, DISP_LOOKUP_EXECD);
3206 /* the open lock or the lock for cross-ref object should be
3207 * returned to the client */
3208 if (rc == -EREMOTE || mdt_get_disposition(rep, DISP_OPEN_LOCK)) {
3209 LASSERT(lustre_handle_is_used(&lhc->mlh_reg_lh));
3210 rep->lock_policy_res2 = 0;
3211 rc = mdt_intent_lock_replace(info, lockp, NULL, lhc, flags);
3215 rep->lock_policy_res2 = clear_serious(rc);
3217 if (rep->lock_policy_res2 == -ENOENT &&
3218 mdt_get_disposition(rep, DISP_LOOKUP_NEG))
3219 rep->lock_policy_res2 = 0;
3221 if (rc == -ENOTCONN || rc == -ENODEV ||
3222 rc == -EOVERFLOW) { /**< if VBR failure then return error */
3224 * If it is the disconnect error (ENODEV & ENOCONN), the error
3225 * will be returned by rq_status, and client at ptlrpc layer
3226 * will detect this, then disconnect, reconnect the import
3227 * immediately, instead of impacting the following the rpc.
3229 lhc->mlh_reg_lh.cookie = 0ull;
3233 * For other cases, the error will be returned by intent.
3234 * and client will retrieve the result from intent.
3237 * FIXME: when open lock is finished, that should be
3240 if (lustre_handle_is_used(&lhc->mlh_reg_lh)) {
3241 LASSERTF(rc == 0, "Error occurred but lock handle "
3242 "is still in use, rc = %d\n", rc);
3243 rep->lock_policy_res2 = 0;
3244 rc = mdt_intent_lock_replace(info, lockp, NULL, lhc, flags);
3247 lhc->mlh_reg_lh.cookie = 0ull;
3248 RETURN(ELDLM_LOCK_ABORTED);
3253 static int mdt_intent_code(long itcode)
3261 case IT_OPEN|IT_CREAT:
3268 rc = MDT_IT_READDIR;
3271 rc = MDT_IT_GETATTR;
3283 rc = MDT_IT_GETXATTR;
3288 case IT_QUOTA_DQACQ:
3293 CERROR("Unknown intent opcode: %ld\n", itcode);
3300 static int mdt_intent_opc(long itopc, struct mdt_thread_info *info,
3301 struct ldlm_lock **lockp, __u64 flags)
3303 struct req_capsule *pill;
3304 struct mdt_it_flavor *flv;
3309 opc = mdt_intent_code(itopc);
3313 pill = info->mti_pill;
3315 if (opc == MDT_IT_QUOTA) {
3316 struct lu_device *qmt = info->mti_mdt->mdt_qmt_dev;
3319 RETURN(-EOPNOTSUPP);
3321 (*lockp)->l_lvb_type = LVB_T_LQUOTA;
3322 /* pass the request to quota master */
3323 rc = qmt_hdls.qmth_intent_policy(info->mti_env, qmt,
3324 mdt_info_req(info), lockp,
3329 flv = &mdt_it_flavor[opc];
3330 if (flv->it_fmt != NULL)
3331 req_capsule_extend(pill, flv->it_fmt);
3333 rc = mdt_unpack_req_pack_rep(info, flv->it_flags);
3335 struct ptlrpc_request *req = mdt_info_req(info);
3336 if (flv->it_flags & MUTABOR &&
3337 exp_connect_flags(req->rq_export) & OBD_CONNECT_RDONLY)
3340 if (rc == 0 && flv->it_act != NULL) {
3341 struct ldlm_reply *rep;
3343 /* execute policy */
3344 rc = flv->it_act(opc, info, lockp, flags);
3346 /* Check whether the reply has been packed successfully. */
3347 if (mdt_info_req(info)->rq_repmsg != NULL) {
3348 rep = req_capsule_server_get(info->mti_pill,
3350 rep->lock_policy_res2 =
3351 ptlrpc_status_hton(rep->lock_policy_res2);
3359 static int mdt_intent_policy(struct ldlm_namespace *ns,
3360 struct ldlm_lock **lockp, void *req_cookie,
3361 ldlm_mode_t mode, __u64 flags, void *data)
3363 struct tgt_session_info *tsi;
3364 struct mdt_thread_info *info;
3365 struct ptlrpc_request *req = req_cookie;
3366 struct ldlm_intent *it;
3367 struct req_capsule *pill;
3372 LASSERT(req != NULL);
3374 tsi = tgt_ses_info(req->rq_svc_thread->t_env);
3376 info = tsi2mdt_info(tsi);
3377 LASSERT(info != NULL);
3378 pill = info->mti_pill;
3379 LASSERT(pill->rc_req == req);
3381 if (req->rq_reqmsg->lm_bufcount > DLM_INTENT_IT_OFF) {
3382 req_capsule_extend(pill, &RQF_LDLM_INTENT_BASIC);
3383 it = req_capsule_client_get(pill, &RMF_LDLM_INTENT);
3385 rc = mdt_intent_opc(it->opc, info, lockp, flags);
3389 /* Lock without inodebits makes no sense and will oops
3390 * later in ldlm. Let's check it now to see if we have
3391 * ibits corrupted somewhere in mdt_intent_opc().
3392 * The case for client miss to set ibits has been
3393 * processed by others. */
3394 LASSERT(ergo(info->mti_dlm_req->lock_desc.l_resource.\
3395 lr_type == LDLM_IBITS,
3396 info->mti_dlm_req->lock_desc.\
3397 l_policy_data.l_inodebits.bits != 0));
3399 rc = err_serious(-EFAULT);
3401 /* No intent was provided */
3402 LASSERT(pill->rc_fmt == &RQF_LDLM_ENQUEUE);
3403 req_capsule_set_size(pill, &RMF_DLM_LVB, RCL_SERVER, 0);
3404 rc = req_capsule_server_pack(pill);
3406 rc = err_serious(rc);
3408 mdt_thread_info_fini(info);
3412 static void mdt_deregister_seq_exp(struct mdt_device *mdt)
3414 struct seq_server_site *ss = mdt_seq_site(mdt);
3416 if (ss->ss_node_id == 0)
3419 if (ss->ss_client_seq != NULL) {
3420 lustre_deregister_lwp_item(&ss->ss_client_seq->lcs_exp);
3421 ss->ss_client_seq->lcs_exp = NULL;
3424 if (ss->ss_server_fld != NULL) {
3425 lustre_deregister_lwp_item(&ss->ss_server_fld->lsf_control_exp);
3426 ss->ss_server_fld->lsf_control_exp = NULL;
3430 static void mdt_seq_fini_cli(struct mdt_device *mdt)
3432 struct seq_server_site *ss = mdt_seq_site(mdt);
3437 if (ss->ss_server_seq == NULL)
3438 seq_server_set_cli(NULL, ss->ss_server_seq, NULL);
3443 static int mdt_seq_fini(const struct lu_env *env, struct mdt_device *mdt)
3445 mdt_seq_fini_cli(mdt);
3446 mdt_deregister_seq_exp(mdt);
3448 return seq_site_fini(env, mdt_seq_site(mdt));
3452 * It will retrieve its FLDB entries from MDT0, and it only happens
3453 * when upgrading existent FS to 2.6 or when local FLDB is corrupted,
3454 * and it needs to refresh FLDB from the MDT0.
3456 static int mdt_register_lwp_callback(void *data)
3459 struct mdt_device *mdt = data;
3460 struct lu_server_fld *fld = mdt_seq_site(mdt)->ss_server_fld;
3464 LASSERT(mdt_seq_site(mdt)->ss_node_id != 0);
3466 if (!likely(fld->lsf_new))
3469 rc = lu_env_init(&env, LCT_MD_THREAD);
3471 CERROR("%s: cannot init env: rc = %d\n", mdt_obd_name(mdt), rc);
3475 rc = fld_update_from_controller(&env, fld);
3477 CERROR("%s: cannot update controller: rc = %d\n",
3478 mdt_obd_name(mdt), rc);
3486 static int mdt_register_seq_exp(struct mdt_device *mdt)
3488 struct seq_server_site *ss = mdt_seq_site(mdt);
3489 char *lwp_name = NULL;
3492 if (ss->ss_node_id == 0)
3495 OBD_ALLOC(lwp_name, MAX_OBD_NAME);
3496 if (lwp_name == NULL)
3497 GOTO(out_free, rc = -ENOMEM);
3499 rc = tgt_name2lwpname(mdt_obd_name(mdt), lwp_name);
3503 rc = lustre_register_lwp_item(lwp_name, &ss->ss_client_seq->lcs_exp,
3508 rc = lustre_register_lwp_item(lwp_name,
3509 &ss->ss_server_fld->lsf_control_exp,
3510 mdt_register_lwp_callback, mdt);
3512 lustre_deregister_lwp_item(&ss->ss_client_seq->lcs_exp);
3513 ss->ss_client_seq->lcs_exp = NULL;
3517 if (lwp_name != NULL)
3518 OBD_FREE(lwp_name, MAX_OBD_NAME);
3524 * Init client sequence manager which is used by local MDS to talk to sequence
3525 * controller on remote node.
3527 static int mdt_seq_init_cli(const struct lu_env *env, struct mdt_device *mdt)
3529 struct seq_server_site *ss = mdt_seq_site(mdt);
3534 /* check if this is adding the first MDC and controller is not yet
3536 OBD_ALLOC_PTR(ss->ss_client_seq);
3537 if (ss->ss_client_seq == NULL)
3540 OBD_ALLOC(prefix, MAX_OBD_NAME + 5);
3541 if (prefix == NULL) {
3542 OBD_FREE_PTR(ss->ss_client_seq);
3543 ss->ss_client_seq = NULL;
3547 /* Note: seq_client_fini will be called in seq_site_fini */
3548 snprintf(prefix, MAX_OBD_NAME + 5, "ctl-%s", mdt_obd_name(mdt));
3549 rc = seq_client_init(ss->ss_client_seq, NULL, LUSTRE_SEQ_METADATA,
3550 prefix, ss->ss_node_id == 0 ? ss->ss_control_seq :
3552 OBD_FREE(prefix, MAX_OBD_NAME + 5);
3554 OBD_FREE_PTR(ss->ss_client_seq);
3555 ss->ss_client_seq = NULL;
3559 rc = seq_server_set_cli(env, ss->ss_server_seq, ss->ss_client_seq);
3564 static int mdt_seq_init(const struct lu_env *env, struct mdt_device *mdt)
3566 struct seq_server_site *ss;
3570 ss = mdt_seq_site(mdt);
3571 /* init sequence controller server(MDT0) */
3572 if (ss->ss_node_id == 0) {
3573 OBD_ALLOC_PTR(ss->ss_control_seq);
3574 if (ss->ss_control_seq == NULL)
3577 rc = seq_server_init(env, ss->ss_control_seq, mdt->mdt_bottom,
3578 mdt_obd_name(mdt), LUSTRE_SEQ_CONTROLLER,
3581 GOTO(out_seq_fini, rc);
3584 /* Init normal sequence server */
3585 OBD_ALLOC_PTR(ss->ss_server_seq);
3586 if (ss->ss_server_seq == NULL)
3587 GOTO(out_seq_fini, rc = -ENOMEM);
3589 rc = seq_server_init(env, ss->ss_server_seq, mdt->mdt_bottom,
3590 mdt_obd_name(mdt), LUSTRE_SEQ_SERVER, ss);
3592 GOTO(out_seq_fini, rc);
3594 /* init seq client for seq server to talk to seq controller(MDT0) */
3595 rc = mdt_seq_init_cli(env, mdt);
3597 GOTO(out_seq_fini, rc);
3599 if (ss->ss_node_id != 0)
3600 /* register controler export through lwp */
3601 rc = mdt_register_seq_exp(mdt);
3606 mdt_seq_fini(env, mdt);
3614 static int mdt_fld_fini(const struct lu_env *env,
3615 struct mdt_device *m)
3617 struct seq_server_site *ss = mdt_seq_site(m);
3620 if (ss && ss->ss_server_fld) {
3621 fld_server_fini(env, ss->ss_server_fld);
3622 OBD_FREE_PTR(ss->ss_server_fld);
3623 ss->ss_server_fld = NULL;
3629 static int mdt_fld_init(const struct lu_env *env,
3631 struct mdt_device *m)
3633 struct seq_server_site *ss;
3637 ss = mdt_seq_site(m);
3639 OBD_ALLOC_PTR(ss->ss_server_fld);
3640 if (ss->ss_server_fld == NULL)
3641 RETURN(rc = -ENOMEM);
3643 rc = fld_server_init(env, ss->ss_server_fld, m->mdt_bottom, uuid,
3646 OBD_FREE_PTR(ss->ss_server_fld);
3647 ss->ss_server_fld = NULL;
3654 static void mdt_stack_pre_fini(const struct lu_env *env,
3655 struct mdt_device *m, struct lu_device *top)
3657 struct obd_device *obd;
3658 struct lustre_cfg_bufs *bufs;
3659 struct lustre_cfg *lcfg;
3660 struct mdt_thread_info *info;
3665 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
3666 LASSERT(info != NULL);
3668 bufs = &info->mti_u.bufs;
3670 LASSERT(m->mdt_child_exp);
3671 LASSERT(m->mdt_child_exp->exp_obd);
3672 obd = m->mdt_child_exp->exp_obd;
3674 /* process cleanup, pass mdt obd name to get obd umount flags */
3675 /* XXX: this is needed because all layers are referenced by
3676 * objects (some of them are pinned by osd, for example *
3677 * the proper solution should be a model where object used
3678 * by osd only doesn't have mdt/mdd slices -bzzz */
3679 lustre_cfg_bufs_reset(bufs, mdt_obd_name(m));
3680 lustre_cfg_bufs_set_string(bufs, 1, NULL);
3681 lcfg = lustre_cfg_new(LCFG_PRE_CLEANUP, bufs);
3683 CERROR("%s:Cannot alloc lcfg!\n", mdt_obd_name(m));
3686 top->ld_ops->ldo_process_config(env, top, lcfg);
3687 lustre_cfg_free(lcfg);
3691 static void mdt_stack_fini(const struct lu_env *env,
3692 struct mdt_device *m, struct lu_device *top)
3694 struct obd_device *obd = mdt2obd_dev(m);
3695 struct lustre_cfg_bufs *bufs;
3696 struct lustre_cfg *lcfg;
3697 struct mdt_thread_info *info;
3701 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
3702 LASSERT(info != NULL);
3704 lu_dev_del_linkage(top->ld_site, top);
3706 lu_site_purge(env, top->ld_site, -1);
3708 bufs = &info->mti_u.bufs;
3709 /* process cleanup, pass mdt obd name to get obd umount flags */
3710 /* another purpose is to let all layers to release their objects */
3711 lustre_cfg_bufs_reset(bufs, mdt_obd_name(m));
3716 lustre_cfg_bufs_set_string(bufs, 1, flags);
3717 lcfg = lustre_cfg_new(LCFG_CLEANUP, bufs);
3719 CERROR("Cannot alloc lcfg!\n");
3723 top->ld_ops->ldo_process_config(env, top, lcfg);
3724 lustre_cfg_free(lcfg);
3726 lu_site_purge(env, top->ld_site, -1);
3728 m->mdt_child = NULL;
3729 m->mdt_bottom = NULL;
3731 obd_disconnect(m->mdt_child_exp);
3732 m->mdt_child_exp = NULL;
3734 obd_disconnect(m->mdt_bottom_exp);
3735 m->mdt_child_exp = NULL;
3738 static int mdt_connect_to_next(const struct lu_env *env, struct mdt_device *m,
3739 const char *next, struct obd_export **exp)
3741 struct obd_connect_data *data = NULL;
3742 struct obd_device *obd;
3746 OBD_ALLOC_PTR(data);
3748 GOTO(out, rc = -ENOMEM);
3750 obd = class_name2obd(next);
3752 CERROR("%s: can't locate next device: %s\n",
3753 mdt_obd_name(m), next);
3754 GOTO(out, rc = -ENOTCONN);
3757 data->ocd_connect_flags = OBD_CONNECT_VERSION;
3758 data->ocd_version = LUSTRE_VERSION_CODE;
3760 rc = obd_connect(NULL, exp, obd, &obd->obd_uuid, data, NULL);
3762 CERROR("%s: cannot connect to next dev %s (%d)\n",
3763 mdt_obd_name(m), next, rc);
3773 static int mdt_stack_init(const struct lu_env *env, struct mdt_device *mdt,
3774 struct lustre_cfg *cfg)
3776 char *dev = lustre_cfg_string(cfg, 0);
3777 int rc, name_size, uuid_size;
3778 char *name, *uuid, *p;
3779 struct lustre_cfg_bufs *bufs;
3780 struct lustre_cfg *lcfg;
3781 struct obd_device *obd;
3782 struct lustre_profile *lprof;
3783 struct lu_site *site;
3786 /* in 1.8 we had the only device in the stack - MDS.
3787 * 2.0 introduces MDT, MDD, OSD; MDT starts others internally.
3788 * in 2.3 OSD is instantiated by obd_mount.c, so we need
3789 * to generate names and setup MDT, MDD. MDT will be using
3790 * generated name to connect to MDD. for MDD the next device
3791 * will be LOD with name taken from so called "profile" which
3792 * is generated by mount_option line
3794 * 1.8 MGS generates config. commands like this:
3795 * #06 (104)mount_option 0: 1:lustre-MDT0000 2:lustre-mdtlov
3796 * #08 (120)setup 0:lustre-MDT0000 1:dev 2:type 3:lustre-MDT0000
3797 * 2.0 MGS generates config. commands like this:
3798 * #07 (112)mount_option 0: 1:lustre-MDT0000 2:lustre-MDT0000-mdtlov
3799 * #08 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
3800 * 3:lustre-MDT0000-mdtlov 4:f
3802 * we generate MDD name from MDT one, just replacing T with D
3804 * after all the preparations, the logical equivalent will be
3805 * #01 (160)setup 0:lustre-MDD0000 1:lustre-MDD0000_UUID 2:0
3806 * 3:lustre-MDT0000-mdtlov 4:f
3807 * #02 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
3808 * 3:lustre-MDD0000 4:f
3810 * notice we build the stack from down to top: MDD first, then MDT */
3812 name_size = MAX_OBD_NAME;
3813 uuid_size = MAX_OBD_NAME;
3815 OBD_ALLOC(name, name_size);
3816 OBD_ALLOC(uuid, uuid_size);
3817 if (name == NULL || uuid == NULL)
3818 GOTO(cleanup_mem, rc = -ENOMEM);
3820 OBD_ALLOC_PTR(bufs);
3822 GOTO(cleanup_mem, rc = -ENOMEM);
3825 p = strstr(name, "-MDT");
3827 GOTO(free_bufs, rc = -ENOMEM);
3830 snprintf(uuid, MAX_OBD_NAME, "%s_UUID", name);
3832 lprof = class_get_profile(lustre_cfg_string(cfg, 0));
3833 if (lprof == NULL || lprof->lp_dt == NULL) {
3834 CERROR("can't find the profile: %s\n",
3835 lustre_cfg_string(cfg, 0));
3836 GOTO(free_bufs, rc = -EINVAL);
3839 lustre_cfg_bufs_reset(bufs, name);
3840 lustre_cfg_bufs_set_string(bufs, 1, LUSTRE_MDD_NAME);
3841 lustre_cfg_bufs_set_string(bufs, 2, uuid);
3842 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
3844 lcfg = lustre_cfg_new(LCFG_ATTACH, bufs);
3846 GOTO(free_bufs, rc = -ENOMEM);
3848 rc = class_attach(lcfg);
3850 GOTO(lcfg_cleanup, rc);
3852 obd = class_name2obd(name);
3854 CERROR("Can not find obd %s (%s in config)\n",
3855 MDD_OBD_NAME, lustre_cfg_string(cfg, 0));
3856 GOTO(class_detach, rc = -EINVAL);
3859 lustre_cfg_free(lcfg);
3861 lustre_cfg_bufs_reset(bufs, name);
3862 lustre_cfg_bufs_set_string(bufs, 1, uuid);
3863 lustre_cfg_bufs_set_string(bufs, 2, dev);
3864 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
3866 lcfg = lustre_cfg_new(LCFG_SETUP, bufs);
3868 rc = class_setup(obd, lcfg);
3870 GOTO(class_detach, rc);
3872 /* connect to MDD we just setup */
3873 rc = mdt_connect_to_next(env, mdt, name, &mdt->mdt_child_exp);
3875 GOTO(class_detach, rc);
3877 site = mdt->mdt_child_exp->exp_obd->obd_lu_dev->ld_site;
3879 LASSERT(mdt_lu_site(mdt) == NULL);
3880 mdt->mdt_lu_dev.ld_site = site;
3881 site->ls_top_dev = &mdt->mdt_lu_dev;
3882 mdt->mdt_child = lu2md_dev(mdt->mdt_child_exp->exp_obd->obd_lu_dev);
3885 /* now connect to bottom OSD */
3886 snprintf(name, MAX_OBD_NAME, "%s-osd", dev);
3887 rc = mdt_connect_to_next(env, mdt, name, &mdt->mdt_bottom_exp);
3889 GOTO(class_detach, rc);
3891 lu2dt_dev(mdt->mdt_bottom_exp->exp_obd->obd_lu_dev);
3894 rc = lu_env_refill((struct lu_env *)env);
3896 CERROR("Failure to refill session: '%d'\n", rc);
3898 lu_dev_add_linkage(site, &mdt->mdt_lu_dev);
3903 class_detach(obd, lcfg);
3905 lustre_cfg_free(lcfg);
3910 OBD_FREE(name, name_size);
3912 OBD_FREE(uuid, uuid_size);
3916 /* setup quota master target on MDT0 */
3917 static int mdt_quota_init(const struct lu_env *env, struct mdt_device *mdt,
3918 struct lustre_cfg *cfg)
3920 struct obd_device *obd;
3921 char *dev = lustre_cfg_string(cfg, 0);
3922 char *qmtname, *uuid, *p;
3923 struct lustre_cfg_bufs *bufs;
3924 struct lustre_cfg *lcfg;
3925 struct lustre_profile *lprof;
3926 struct obd_connect_data *data;
3930 LASSERT(mdt->mdt_qmt_exp == NULL);
3931 LASSERT(mdt->mdt_qmt_dev == NULL);
3933 /* quota master is on MDT0 only for now */
3934 if (mdt->mdt_seq_site.ss_node_id != 0)
3937 /* MGS generates config commands which look as follows:
3938 * #01 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
3939 * 3:lustre-MDT0000-mdtlov 4:f
3941 * We generate the QMT name from the MDT one, just replacing MD with QM
3942 * after all the preparations, the logical equivalent will be:
3943 * #01 (160)setup 0:lustre-QMT0000 1:lustre-QMT0000_UUID 2:0
3944 * 3:lustre-MDT0000-osd 4:f */
3945 OBD_ALLOC(qmtname, MAX_OBD_NAME);
3946 OBD_ALLOC(uuid, UUID_MAX);
3947 OBD_ALLOC_PTR(bufs);
3948 OBD_ALLOC_PTR(data);
3949 if (qmtname == NULL || uuid == NULL || bufs == NULL || data == NULL)
3950 GOTO(cleanup_mem, rc = -ENOMEM);
3952 strcpy(qmtname, dev);
3953 p = strstr(qmtname, "-MDT");
3955 GOTO(cleanup_mem, rc = -ENOMEM);
3956 /* replace MD with QM */
3960 snprintf(uuid, UUID_MAX, "%s_UUID", qmtname);
3962 lprof = class_get_profile(lustre_cfg_string(cfg, 0));
3963 if (lprof == NULL || lprof->lp_dt == NULL) {
3964 CERROR("can't find profile for %s\n",
3965 lustre_cfg_string(cfg, 0));
3966 GOTO(cleanup_mem, rc = -EINVAL);
3969 lustre_cfg_bufs_reset(bufs, qmtname);
3970 lustre_cfg_bufs_set_string(bufs, 1, LUSTRE_QMT_NAME);
3971 lustre_cfg_bufs_set_string(bufs, 2, uuid);
3972 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
3974 lcfg = lustre_cfg_new(LCFG_ATTACH, bufs);
3976 GOTO(cleanup_mem, rc = -ENOMEM);
3978 rc = class_attach(lcfg);
3980 GOTO(lcfg_cleanup, rc);
3982 obd = class_name2obd(qmtname);
3984 CERROR("Can not find obd %s (%s in config)\n", qmtname,
3985 lustre_cfg_string(cfg, 0));
3986 GOTO(class_detach, rc = -EINVAL);
3989 lustre_cfg_free(lcfg);
3991 lustre_cfg_bufs_reset(bufs, qmtname);
3992 lustre_cfg_bufs_set_string(bufs, 1, uuid);
3993 lustre_cfg_bufs_set_string(bufs, 2, dev);
3995 /* for quota, the next device should be the OSD device */
3996 lustre_cfg_bufs_set_string(bufs, 3,
3997 mdt->mdt_bottom->dd_lu_dev.ld_obd->obd_name);
3999 lcfg = lustre_cfg_new(LCFG_SETUP, bufs);
4001 rc = class_setup(obd, lcfg);
4003 GOTO(class_detach, rc);
4005 mdt->mdt_qmt_dev = obd->obd_lu_dev;
4007 /* configure local quota objects */
4008 rc = mdt->mdt_qmt_dev->ld_ops->ldo_prepare(env,
4012 GOTO(class_cleanup, rc);
4014 /* connect to quota master target */
4015 data->ocd_connect_flags = OBD_CONNECT_VERSION;
4016 data->ocd_version = LUSTRE_VERSION_CODE;
4017 rc = obd_connect(NULL, &mdt->mdt_qmt_exp, obd, &obd->obd_uuid,
4020 CERROR("cannot connect to quota master device %s (%d)\n",
4022 GOTO(class_cleanup, rc);
4028 class_manual_cleanup(obd);
4029 mdt->mdt_qmt_dev = NULL;
4033 class_detach(obd, lcfg);
4035 lustre_cfg_free(lcfg);
4040 OBD_FREE(qmtname, MAX_OBD_NAME);
4042 OBD_FREE(uuid, UUID_MAX);
4048 /* Shutdown quota master target associated with mdt */
4049 static void mdt_quota_fini(const struct lu_env *env, struct mdt_device *mdt)
4053 if (mdt->mdt_qmt_exp == NULL)
4055 LASSERT(mdt->mdt_qmt_dev != NULL);
4057 /* the qmt automatically shuts down when the mdt disconnects */
4058 obd_disconnect(mdt->mdt_qmt_exp);
4059 mdt->mdt_qmt_exp = NULL;
4060 mdt->mdt_qmt_dev = NULL;
4064 /* mdt_getxattr() is used from mdt_intent_getxattr(), use this wrapper
4065 * for now. This will be removed along with converting rest of MDT code
4066 * to use tgt_session_info */
4067 int mdt_tgt_getxattr(struct tgt_session_info *tsi)
4069 struct mdt_thread_info *info = tsi2mdt_info(tsi);
4072 rc = mdt_getxattr(info);
4074 mdt_thread_info_fini(info);
4078 static struct tgt_handler mdt_tgt_handlers[] = {
4079 TGT_RPC_HANDLER(MDS_FIRST_OPC,
4080 0, MDS_CONNECT, mdt_tgt_connect,
4081 &RQF_CONNECT, LUSTRE_OBD_VERSION),
4082 TGT_RPC_HANDLER(MDS_FIRST_OPC,
4083 0, MDS_DISCONNECT, tgt_disconnect,
4084 &RQF_MDS_DISCONNECT, LUSTRE_OBD_VERSION),
4085 TGT_RPC_HANDLER(MDS_FIRST_OPC,
4086 HABEO_REFERO, MDS_SET_INFO, mdt_set_info,
4087 &RQF_OBD_SET_INFO, LUSTRE_MDS_VERSION),
4088 TGT_MDT_HDL(0, MDS_GET_INFO, mdt_get_info),
4089 TGT_MDT_HDL(0 | HABEO_REFERO, MDS_GETSTATUS, mdt_getstatus),
4090 TGT_MDT_HDL(HABEO_CORPUS, MDS_GETATTR, mdt_getattr),
4091 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_GETATTR_NAME,
4093 TGT_MDT_HDL(HABEO_CORPUS, MDS_GETXATTR, mdt_tgt_getxattr),
4094 TGT_MDT_HDL(0 | HABEO_REFERO, MDS_STATFS, mdt_statfs),
4095 TGT_MDT_HDL(0 | MUTABOR, MDS_REINT, mdt_reint),
4096 TGT_MDT_HDL(HABEO_CORPUS, MDS_CLOSE, mdt_close),
4097 TGT_MDT_HDL(HABEO_CORPUS, MDS_DONE_WRITING,
4099 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_READPAGE, mdt_readpage),
4100 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_SYNC, mdt_sync),
4101 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_IS_SUBDIR, mdt_is_subdir),
4102 TGT_MDT_HDL(0, MDS_QUOTACTL, mdt_quotactl),
4103 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO | MUTABOR, MDS_HSM_PROGRESS,
4105 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO | MUTABOR, MDS_HSM_CT_REGISTER,
4106 mdt_hsm_ct_register),
4107 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO | MUTABOR, MDS_HSM_CT_UNREGISTER,
4108 mdt_hsm_ct_unregister),
4109 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_HSM_STATE_GET,
4111 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO | MUTABOR, MDS_HSM_STATE_SET,
4113 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_HSM_ACTION, mdt_hsm_action),
4114 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_HSM_REQUEST,
4116 TGT_MDT_HDL(HABEO_CORPUS|HABEO_REFERO | MUTABOR, MDS_SWAP_LAYOUTS,
4120 static struct tgt_handler mdt_sec_ctx_ops[] = {
4121 TGT_SEC_HDL_VAR(0, SEC_CTX_INIT, mdt_sec_ctx_handle),
4122 TGT_SEC_HDL_VAR(0, SEC_CTX_INIT_CONT,mdt_sec_ctx_handle),
4123 TGT_SEC_HDL_VAR(0, SEC_CTX_FINI, mdt_sec_ctx_handle)
4126 static struct tgt_handler mdt_quota_ops[] = {
4127 TGT_QUOTA_HDL(HABEO_REFERO, QUOTA_DQACQ, mdt_quota_dqacq),
4130 static struct tgt_opc_slice mdt_common_slice[] = {
4132 .tos_opc_start = MDS_FIRST_OPC,
4133 .tos_opc_end = MDS_LAST_OPC,
4134 .tos_hs = mdt_tgt_handlers
4137 .tos_opc_start = OBD_FIRST_OPC,
4138 .tos_opc_end = OBD_LAST_OPC,
4139 .tos_hs = tgt_obd_handlers
4142 .tos_opc_start = LDLM_FIRST_OPC,
4143 .tos_opc_end = LDLM_LAST_OPC,
4144 .tos_hs = tgt_dlm_handlers
4147 .tos_opc_start = SEC_FIRST_OPC,
4148 .tos_opc_end = SEC_LAST_OPC,
4149 .tos_hs = mdt_sec_ctx_ops
4152 .tos_opc_start = UPDATE_OBJ,
4153 .tos_opc_end = UPDATE_LAST_OPC,
4154 .tos_hs = tgt_out_handlers
4157 .tos_opc_start = FLD_FIRST_OPC,
4158 .tos_opc_end = FLD_LAST_OPC,
4159 .tos_hs = fld_handlers
4162 .tos_opc_start = SEQ_FIRST_OPC,
4163 .tos_opc_end = SEQ_LAST_OPC,
4164 .tos_hs = seq_handlers
4167 .tos_opc_start = QUOTA_DQACQ,
4168 .tos_opc_end = QUOTA_LAST_OPC,
4169 .tos_hs = mdt_quota_ops
4172 .tos_opc_start = LLOG_FIRST_OPC,
4173 .tos_opc_end = LLOG_LAST_OPC,
4174 .tos_hs = tgt_llog_handlers
4182 static void mdt_fini(const struct lu_env *env, struct mdt_device *m)
4184 struct md_device *next = m->mdt_child;
4185 struct lu_device *d = &m->mdt_lu_dev;
4186 struct obd_device *obd = mdt2obd_dev(m);
4189 target_recovery_fini(obd);
4191 ping_evictor_stop();
4193 mdt_stack_pre_fini(env, m, md2lu_dev(m->mdt_child));
4195 if (m->mdt_opts.mo_coordinator)
4196 mdt_hsm_cdt_stop(m);
4198 mdt_hsm_cdt_fini(m);
4200 mdt_llog_ctxt_unclone(env, m, LLOG_AGENT_ORIG_CTXT);
4201 mdt_llog_ctxt_unclone(env, m, LLOG_CHANGELOG_ORIG_CTXT);
4202 obd_exports_barrier(obd);
4203 obd_zombie_barrier();
4207 tgt_fini(env, &m->mdt_lut);
4208 mdt_fs_cleanup(env, m);
4209 upcall_cache_cleanup(m->mdt_identity_cache);
4210 m->mdt_identity_cache = NULL;
4212 if (m->mdt_namespace != NULL) {
4213 ldlm_namespace_free(m->mdt_namespace, NULL,
4214 d->ld_obd->obd_force);
4215 d->ld_obd->obd_namespace = m->mdt_namespace = NULL;
4218 mdt_quota_fini(env, m);
4220 cfs_free_nidlist(&m->mdt_nosquash_nids);
4221 if (m->mdt_nosquash_str) {
4222 OBD_FREE(m->mdt_nosquash_str, m->mdt_nosquash_strlen);
4223 m->mdt_nosquash_str = NULL;
4224 m->mdt_nosquash_strlen = 0;
4227 next->md_ops->mdo_iocontrol(env, next, OBD_IOC_PAUSE_LFSCK, 0, NULL);
4229 mdt_seq_fini(env, m);
4230 mdt_fld_fini(env, m);
4232 next->md_ops->mdo_init_capa_ctxt(env, next, 0, 0, 0, NULL);
4233 cfs_timer_disarm(&m->mdt_ck_timer);
4234 mdt_ck_thread_stop(m);
4239 mdt_stack_fini(env, m, md2lu_dev(m->mdt_child));
4241 LASSERT(cfs_atomic_read(&d->ld_ref) == 0);
4243 server_put_mount(mdt_obd_name(m));
4248 int mdt_postrecov(const struct lu_env *, struct mdt_device *);
4250 static int mdt_init0(const struct lu_env *env, struct mdt_device *m,
4251 struct lu_device_type *ldt, struct lustre_cfg *cfg)
4253 struct mdt_thread_info *info;
4254 struct obd_device *obd;
4255 const char *dev = lustre_cfg_string(cfg, 0);
4256 const char *num = lustre_cfg_string(cfg, 2);
4257 struct lustre_mount_info *lmi = NULL;
4258 struct lustre_sb_info *lsi;
4260 struct seq_server_site *ss_site;
4261 const char *identity_upcall = "NONE";
4262 struct md_device *next;
4268 lu_device_init(&m->mdt_lu_dev, ldt);
4270 * Environment (env) might be missing mdt_thread_key values at that
4271 * point, if device is allocated when mdt_thread_key is in QUIESCENT
4274 * Usually device allocation path doesn't use module key values, but
4275 * mdt has to do a lot of work here, so allocate key value.
4277 rc = lu_env_refill((struct lu_env *)env);
4281 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
4282 LASSERT(info != NULL);
4284 obd = class_name2obd(dev);
4285 LASSERT(obd != NULL);
4287 m->mdt_max_mdsize = MAX_MD_SIZE; /* 4 stripes */
4289 m->mdt_som_conf = 0;
4291 m->mdt_opts.mo_cos = MDT_COS_DEFAULT;
4293 /* default is coordinator off, it is started through conf_param
4295 m->mdt_opts.mo_coordinator = 0;
4297 lmi = server_get_mount(dev);
4299 CERROR("Cannot get mount info for %s!\n", dev);
4302 lsi = s2lsi(lmi->lmi_sb);
4303 /* CMD is supported only in IAM mode */
4305 node_id = simple_strtol(num, NULL, 10);
4306 obd->u.obt.obt_magic = OBT_MAGIC;
4309 spin_lock_init(&m->mdt_ioepoch_lock);
4310 m->mdt_capa_timeout = CAPA_TIMEOUT;
4311 m->mdt_capa_alg = CAPA_HMAC_ALG_SHA1;
4312 m->mdt_ck_timeout = CAPA_KEY_TIMEOUT;
4313 m->mdt_squash_uid = 0;
4314 m->mdt_squash_gid = 0;
4315 CFS_INIT_LIST_HEAD(&m->mdt_nosquash_nids);
4316 m->mdt_nosquash_str = NULL;
4317 m->mdt_nosquash_strlen = 0;
4318 init_rwsem(&m->mdt_squash_sem);
4319 spin_lock_init(&m->mdt_osfs_lock);
4320 m->mdt_osfs_age = cfs_time_shift_64(-1000);
4321 m->mdt_enable_remote_dir = 0;
4322 m->mdt_enable_remote_dir_gid = 0;
4324 m->mdt_lu_dev.ld_ops = &mdt_lu_ops;
4325 m->mdt_lu_dev.ld_obd = obd;
4326 /* Set this lu_device to obd for error handling purposes. */
4327 obd->obd_lu_dev = &m->mdt_lu_dev;
4329 /* init the stack */
4330 rc = mdt_stack_init((struct lu_env *)env, m, cfg);
4332 CERROR("%s: Can't init device stack, rc %d\n",
4333 mdt_obd_name(m), rc);
4338 ss_site = mdt_seq_site(m);
4339 s->ld_seq_site = ss_site;
4342 /* set server index */
4343 ss_site->ss_node_id = node_id;
4345 /* failover is the default
4346 * FIXME: we do not failout mds0/mgs, which may cause some problems.
4347 * assumed whose ss_node_id == 0 XXX
4349 obd->obd_replayable = 1;
4350 /* No connection accepted until configurations will finish */
4351 obd->obd_no_conn = 1;
4353 if (cfg->lcfg_bufcount > 4 && LUSTRE_CFG_BUFLEN(cfg, 4) > 0) {
4354 char *str = lustre_cfg_string(cfg, 4);
4355 if (strchr(str, 'n')) {
4356 CWARN("%s: recovery disabled\n", mdt_obd_name(m));
4357 obd->obd_replayable = 0;
4361 rc = mdt_fld_init(env, mdt_obd_name(m), m);
4363 GOTO(err_fini_stack, rc);
4365 rc = mdt_seq_init(env, m);
4367 GOTO(err_fini_fld, rc);
4369 snprintf(info->mti_u.ns_name, sizeof(info->mti_u.ns_name), "%s-%s",
4370 LUSTRE_MDT_NAME, obd->obd_uuid.uuid);
4371 m->mdt_namespace = ldlm_namespace_new(obd, info->mti_u.ns_name,
4372 LDLM_NAMESPACE_SERVER,
4373 LDLM_NAMESPACE_GREEDY,
4375 if (m->mdt_namespace == NULL)
4376 GOTO(err_fini_seq, rc = -ENOMEM);
4378 m->mdt_namespace->ns_lvbp = m;
4379 m->mdt_namespace->ns_lvbo = &mdt_lvbo;
4381 ldlm_register_intent(m->mdt_namespace, mdt_intent_policy);
4382 /* set obd_namespace for compatibility with old code */
4383 obd->obd_namespace = m->mdt_namespace;
4385 cfs_timer_init(&m->mdt_ck_timer, mdt_ck_timer_callback, m);
4387 rc = mdt_hsm_cdt_init(m);
4389 CERROR("%s: error initializing coordinator, rc %d\n",
4390 mdt_obd_name(m), rc);
4391 GOTO(err_free_ns, rc);
4394 rc = mdt_ck_thread_start(m);
4396 GOTO(err_free_hsm, rc);
4398 rc = tgt_init(env, &m->mdt_lut, obd, m->mdt_bottom, mdt_common_slice,
4399 OBD_FAIL_MDS_ALL_REQUEST_NET,
4400 OBD_FAIL_MDS_ALL_REPLY_NET);
4404 rc = mdt_fs_setup(env, m, obd, lsi);
4408 tgt_adapt_sptlrpc_conf(&m->mdt_lut, 1);
4410 next = m->mdt_child;
4411 rc = next->md_ops->mdo_iocontrol(env, next, OBD_IOC_GET_MNTOPT, 0,
4414 GOTO(err_fs_cleanup, rc);
4416 if (mntopts & MNTOPT_USERXATTR)
4417 m->mdt_opts.mo_user_xattr = 1;
4419 m->mdt_opts.mo_user_xattr = 0;
4421 rc = next->md_ops->mdo_maxeasize_get(env, next, &m->mdt_max_ea_size);
4423 GOTO(err_fs_cleanup, rc);
4425 if (mntopts & MNTOPT_ACL)
4426 m->mdt_opts.mo_acl = 1;
4428 m->mdt_opts.mo_acl = 0;
4430 /* XXX: to support suppgid for ACL, we enable identity_upcall
4431 * by default, otherwise, maybe got unexpected -EACCESS. */
4432 if (m->mdt_opts.mo_acl)
4433 identity_upcall = MDT_IDENTITY_UPCALL_PATH;
4435 m->mdt_identity_cache = upcall_cache_init(mdt_obd_name(m),
4437 &mdt_identity_upcall_cache_ops);
4438 if (IS_ERR(m->mdt_identity_cache)) {
4439 rc = PTR_ERR(m->mdt_identity_cache);
4440 m->mdt_identity_cache = NULL;
4441 GOTO(err_fs_cleanup, rc);
4444 rc = mdt_procfs_init(m, dev);
4446 CERROR("Can't init MDT lprocfs, rc %d\n", rc);
4447 GOTO(err_recovery, rc);
4450 rc = mdt_quota_init(env, m, cfg);
4452 GOTO(err_procfs, rc);
4454 m->mdt_ldlm_client = &mdt2obd_dev(m)->obd_ldlm_client;
4455 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
4456 "mdt_ldlm_client", m->mdt_ldlm_client);
4458 ping_evictor_start();
4460 /* recovery will be started upon mdt_prepare()
4461 * when the whole stack is complete and ready
4462 * to serve the requests */
4464 mdt_init_capa_ctxt(env, m);
4466 /* Reduce the initial timeout on an MDS because it doesn't need such
4467 * a long timeout as an OST does. Adaptive timeouts will adjust this
4468 * value appropriately. */
4469 if (ldlm_timeout == LDLM_TIMEOUT_DEFAULT)
4470 ldlm_timeout = MDS_LDLM_TIMEOUT_DEFAULT;
4476 target_recovery_fini(obd);
4477 upcall_cache_cleanup(m->mdt_identity_cache);
4478 m->mdt_identity_cache = NULL;
4480 mdt_fs_cleanup(env, m);
4482 tgt_fini(env, &m->mdt_lut);
4484 cfs_timer_disarm(&m->mdt_ck_timer);
4485 mdt_ck_thread_stop(m);
4487 mdt_hsm_cdt_fini(m);
4489 ldlm_namespace_free(m->mdt_namespace, NULL, 0);
4490 obd->obd_namespace = m->mdt_namespace = NULL;
4492 mdt_seq_fini(env, m);
4494 mdt_fld_fini(env, m);
4496 mdt_stack_fini(env, m, md2lu_dev(m->mdt_child));
4499 server_put_mount(dev);
4503 /* For interoperability, the left element is old parameter, the right one
4504 * is the new version of the parameter, if some parameter is deprecated,
4505 * the new version should be set as NULL. */
4506 static struct cfg_interop_param mdt_interop_param[] = {
4507 { "mdt.group_upcall", NULL },
4508 { "mdt.quota_type", NULL },
4509 { "mdd.quota_type", NULL },
4510 { "mdt.rootsquash", "mdt.root_squash" },
4511 { "mdt.nosquash_nid", "mdt.nosquash_nids" },
4515 /* used by MGS to process specific configurations */
4516 static int mdt_process_config(const struct lu_env *env,
4517 struct lu_device *d, struct lustre_cfg *cfg)
4519 struct mdt_device *m = mdt_dev(d);
4520 struct md_device *md_next = m->mdt_child;
4521 struct lu_device *next = md2lu_dev(md_next);
4525 switch (cfg->lcfg_command) {
4527 struct lprocfs_static_vars lvars;
4528 struct obd_device *obd = d->ld_obd;
4530 /* For interoperability */
4531 struct cfg_interop_param *ptr = NULL;
4532 struct lustre_cfg *old_cfg = NULL;
4535 param = lustre_cfg_string(cfg, 1);
4536 if (param == NULL) {
4537 CERROR("param is empty\n");
4542 ptr = class_find_old_param(param, mdt_interop_param);
4544 if (ptr->new_param == NULL) {
4546 CWARN("For interoperability, skip this %s."
4547 " It is obsolete.\n", ptr->old_param);
4551 CWARN("Found old param %s, changed it to %s.\n",
4552 ptr->old_param, ptr->new_param);
4555 cfg = lustre_cfg_rename(old_cfg, ptr->new_param);
4562 lprocfs_mdt_init_vars(&lvars);
4563 rc = class_process_proc_param(PARAM_MDT, lvars.obd_vars,
4565 if (rc > 0 || rc == -ENOSYS) {
4566 /* is it an HSM var ? */
4567 rc = class_process_proc_param(PARAM_HSM,
4568 hsm_cdt_get_proc_vars(),
4570 if (rc > 0 || rc == -ENOSYS)
4571 /* we don't understand; pass it on */
4572 rc = next->ld_ops->ldo_process_config(env, next,
4576 if (old_cfg != NULL)
4577 lustre_cfg_free(cfg);
4582 /* others are passed further */
4583 rc = next->ld_ops->ldo_process_config(env, next, cfg);
4589 static struct lu_object *mdt_object_alloc(const struct lu_env *env,
4590 const struct lu_object_header *hdr,
4591 struct lu_device *d)
4593 struct mdt_object *mo;
4597 OBD_SLAB_ALLOC_PTR_GFP(mo, mdt_object_kmem, __GFP_IO);
4599 struct lu_object *o;
4600 struct lu_object_header *h;
4603 h = &mo->mot_header;
4604 lu_object_header_init(h);
4605 lu_object_init(o, h, d);
4606 lu_object_add_top(h, o);
4607 o->lo_ops = &mdt_obj_ops;
4608 mutex_init(&mo->mot_ioepoch_mutex);
4609 mutex_init(&mo->mot_lov_mutex);
4610 init_rwsem(&mo->mot_open_sem);
4616 static int mdt_object_init(const struct lu_env *env, struct lu_object *o,
4617 const struct lu_object_conf *unused)
4619 struct mdt_device *d = mdt_dev(o->lo_dev);
4620 struct lu_device *under;
4621 struct lu_object *below;
4625 CDEBUG(D_INFO, "object init, fid = "DFID"\n",
4626 PFID(lu_object_fid(o)));
4628 under = &d->mdt_child->md_lu_dev;
4629 below = under->ld_ops->ldo_object_alloc(env, o->lo_header, under);
4630 if (below != NULL) {
4631 lu_object_add(o, below);
4638 static void mdt_object_free(const struct lu_env *env, struct lu_object *o)
4640 struct mdt_object *mo = mdt_obj(o);
4641 struct lu_object_header *h;
4645 CDEBUG(D_INFO, "object free, fid = "DFID"\n",
4646 PFID(lu_object_fid(o)));
4648 LASSERT(atomic_read(&mo->mot_open_count) == 0);
4649 LASSERT(atomic_read(&mo->mot_lease_count) == 0);
4652 lu_object_header_fini(h);
4653 OBD_SLAB_FREE_PTR(mo, mdt_object_kmem);
4658 static int mdt_object_print(const struct lu_env *env, void *cookie,
4659 lu_printer_t p, const struct lu_object *o)
4661 struct mdt_object *mdto = mdt_obj((struct lu_object *)o);
4662 return (*p)(env, cookie, LUSTRE_MDT_NAME"-object@%p(ioepoch="LPU64" "
4663 "flags="LPX64", epochcount=%d, writecount=%d)",
4664 mdto, mdto->mot_ioepoch, mdto->mot_flags,
4665 mdto->mot_ioepoch_count, mdto->mot_writecount);
4668 static int mdt_prepare(const struct lu_env *env,
4669 struct lu_device *pdev,
4670 struct lu_device *cdev)
4672 struct mdt_device *mdt = mdt_dev(cdev);
4673 struct lu_device *next = &mdt->mdt_child->md_lu_dev;
4674 struct obd_device *obd = cdev->ld_obd;
4675 struct lfsck_start_param lsp;
4682 rc = next->ld_ops->ldo_prepare(env, cdev, next);
4686 rc = mdt_llog_ctxt_clone(env, mdt, LLOG_CHANGELOG_ORIG_CTXT);
4690 rc = mdt_llog_ctxt_clone(env, mdt, LLOG_AGENT_ORIG_CTXT);
4694 lsp.lsp_start = NULL;
4695 lsp.lsp_namespace = mdt->mdt_namespace;
4696 rc = mdt->mdt_child->md_ops->mdo_iocontrol(env, mdt->mdt_child,
4697 OBD_IOC_START_LFSCK,
4700 CWARN("%s: auto trigger paused LFSCK failed: rc = %d\n",
4701 mdt_obd_name(mdt), rc);
4705 if (mdt->mdt_seq_site.ss_node_id == 0) {
4706 rc = mdt->mdt_child->md_ops->mdo_root_get(env, mdt->mdt_child,
4707 &mdt->mdt_md_root_fid);
4712 LASSERT(!test_bit(MDT_FL_CFGLOG, &mdt->mdt_state));
4713 target_recovery_init(&mdt->mdt_lut, tgt_request_handle);
4714 set_bit(MDT_FL_CFGLOG, &mdt->mdt_state);
4715 LASSERT(obd->obd_no_conn);
4716 spin_lock(&obd->obd_dev_lock);
4717 obd->obd_no_conn = 0;
4718 spin_unlock(&obd->obd_dev_lock);
4720 if (obd->obd_recovering == 0)
4721 mdt_postrecov(env, mdt);
4726 const struct lu_device_operations mdt_lu_ops = {
4727 .ldo_object_alloc = mdt_object_alloc,
4728 .ldo_process_config = mdt_process_config,
4729 .ldo_prepare = mdt_prepare,
4732 static const struct lu_object_operations mdt_obj_ops = {
4733 .loo_object_init = mdt_object_init,
4734 .loo_object_free = mdt_object_free,
4735 .loo_object_print = mdt_object_print
4738 static int mdt_obd_set_info_async(const struct lu_env *env,
4739 struct obd_export *exp,
4740 __u32 keylen, void *key,
4741 __u32 vallen, void *val,
4742 struct ptlrpc_request_set *set)
4748 if (KEY_IS(KEY_SPTLRPC_CONF)) {
4749 rc = tgt_adapt_sptlrpc_conf(class_exp2tgt(exp), 0);
4757 * Match client and server connection feature flags.
4759 * Compute the compatibility flags for a connection request based on
4760 * features mutually supported by client and server.
4762 * The obd_export::exp_connect_data.ocd_connect_flags field in \a exp
4763 * must not be updated here, otherwise a partially initialized value may
4764 * be exposed. After the connection request is successfully processed,
4765 * the top-level MDT connect request handler atomically updates the export
4766 * connect flags from the obd_connect_data::ocd_connect_flags field of the
4767 * reply. \see mdt_connect().
4769 * \param exp the obd_export associated with this client/target pair
4770 * \param mdt the target device for the connection
4771 * \param data stores data for this connect request
4774 * \retval -EPROTO \a data unexpectedly has zero obd_connect_data::ocd_brw_size
4775 * \retval -EBADE client and server feature requirements are incompatible
4777 static int mdt_connect_internal(struct obd_export *exp,
4778 struct mdt_device *mdt,
4779 struct obd_connect_data *data)
4781 LASSERT(data != NULL);
4783 data->ocd_connect_flags &= MDT_CONNECT_SUPPORTED;
4784 data->ocd_ibits_known &= MDS_INODELOCK_FULL;
4786 /* If no known bits (which should not happen, probably,
4787 as everybody should support LOOKUP and UPDATE bits at least)
4788 revert to compat mode with plain locks. */
4789 if (!data->ocd_ibits_known &&
4790 data->ocd_connect_flags & OBD_CONNECT_IBITS)
4791 data->ocd_connect_flags &= ~OBD_CONNECT_IBITS;
4793 if (!mdt->mdt_opts.mo_acl)
4794 data->ocd_connect_flags &= ~OBD_CONNECT_ACL;
4796 if (!mdt->mdt_opts.mo_user_xattr)
4797 data->ocd_connect_flags &= ~OBD_CONNECT_XATTR;
4799 if (!mdt->mdt_som_conf)
4800 data->ocd_connect_flags &= ~OBD_CONNECT_SOM;
4802 if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
4803 data->ocd_brw_size = min(data->ocd_brw_size,
4804 (__u32)MD_MAX_BRW_SIZE);
4805 if (data->ocd_brw_size == 0) {
4806 CERROR("%s: cli %s/%p ocd_connect_flags: "LPX64
4807 " ocd_version: %x ocd_grant: %d "
4808 "ocd_index: %u ocd_brw_size is "
4809 "unexpectedly zero, network data "
4810 "corruption? Refusing connection of this"
4813 exp->exp_client_uuid.uuid,
4814 exp, data->ocd_connect_flags, data->ocd_version,
4815 data->ocd_grant, data->ocd_index);
4820 /* NB: Disregard the rule against updating
4821 * exp_connect_data.ocd_connect_flags in this case, since
4822 * tgt_client_new() needs to know if this is a lightweight
4823 * connection, and it is safe to expose this flag before
4824 * connection processing completes. */
4825 if (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) {
4826 spin_lock(&exp->exp_lock);
4827 *exp_connect_flags_ptr(exp) |= OBD_CONNECT_LIGHTWEIGHT;
4828 spin_unlock(&exp->exp_lock);
4831 data->ocd_version = LUSTRE_VERSION_CODE;
4833 if ((data->ocd_connect_flags & OBD_CONNECT_FID) == 0) {
4834 CWARN("%s: MDS requires FID support, but client not\n",
4839 if (mdt->mdt_som_conf &&
4840 !(data->ocd_connect_flags & (OBD_CONNECT_LIGHTWEIGHT |
4841 OBD_CONNECT_MDS_MDS |
4842 OBD_CONNECT_SOM))) {
4843 CWARN("%s: MDS has SOM enabled, but client does not support "
4844 "it\n", mdt_obd_name(mdt));
4848 if (OCD_HAS_FLAG(data, PINGLESS)) {
4849 if (ptlrpc_pinger_suppress_pings()) {
4850 spin_lock(&exp->exp_obd->obd_dev_lock);
4851 list_del_init(&exp->exp_obd_chain_timed);
4852 spin_unlock(&exp->exp_obd->obd_dev_lock);
4854 data->ocd_connect_flags &= ~OBD_CONNECT_PINGLESS;
4858 data->ocd_max_easize = mdt->mdt_max_ea_size;
4863 /* mds_connect copy */
4864 static int mdt_obd_connect(const struct lu_env *env,
4865 struct obd_export **exp, struct obd_device *obd,
4866 struct obd_uuid *cluuid,
4867 struct obd_connect_data *data,
4870 struct obd_export *lexp;
4871 struct lustre_handle conn = { 0 };
4872 struct mdt_device *mdt;
4876 LASSERT(env != NULL);
4877 if (!exp || !obd || !cluuid)
4880 mdt = mdt_dev(obd->obd_lu_dev);
4883 * first, check whether the stack is ready to handle requests
4884 * XXX: probably not very appropriate method is used now
4885 * at some point we should find a better one
4887 if (!test_bit(MDT_FL_SYNCED, &mdt->mdt_state) && data != NULL &&
4888 !(data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT)) {
4889 rc = obd_get_info(env, mdt->mdt_child_exp,
4890 sizeof(KEY_OSP_CONNECTED),
4891 KEY_OSP_CONNECTED, NULL, NULL, NULL);
4894 set_bit(MDT_FL_SYNCED, &mdt->mdt_state);
4897 rc = class_connect(&conn, obd, cluuid);
4901 lexp = class_conn2export(&conn);
4902 LASSERT(lexp != NULL);
4904 rc = mdt_connect_internal(lexp, mdt, data);
4906 struct lsd_client_data *lcd = lexp->exp_target_data.ted_lcd;
4909 memcpy(lcd->lcd_uuid, cluuid, sizeof lcd->lcd_uuid);
4910 rc = tgt_client_new(env, lexp);
4912 mdt_export_stats_init(obd, lexp, localdata);
4916 class_disconnect(lexp);
4925 static int mdt_obd_reconnect(const struct lu_env *env,
4926 struct obd_export *exp, struct obd_device *obd,
4927 struct obd_uuid *cluuid,
4928 struct obd_connect_data *data,
4934 if (exp == NULL || obd == NULL || cluuid == NULL)
4937 rc = mdt_connect_internal(exp, mdt_dev(obd->obd_lu_dev), data);
4939 mdt_export_stats_init(obd, exp, localdata);
4944 static int mdt_ctxt_add_dirty_flag(struct lu_env *env,
4945 struct mdt_thread_info *info,
4946 struct mdt_file_data *mfd)
4948 struct lu_context ses;
4952 rc = lu_context_init(&ses, LCT_SERVER_SESSION);
4957 lu_context_enter(&ses);
4959 mdt_ucred(info)->uc_valid = UCRED_OLD;
4960 rc = mdt_add_dirty_flag(info, mfd->mfd_object, &info->mti_attr);
4962 lu_context_exit(&ses);
4963 lu_context_fini(&ses);
4969 static int mdt_export_cleanup(struct obd_export *exp)
4971 struct mdt_export_data *med = &exp->exp_mdt_data;
4972 struct obd_device *obd = exp->exp_obd;
4973 struct mdt_device *mdt;
4974 struct mdt_thread_info *info;
4976 CFS_LIST_HEAD(closing_list);
4977 struct mdt_file_data *mfd, *n;
4981 spin_lock(&med->med_open_lock);
4982 while (!cfs_list_empty(&med->med_open_head)) {
4983 cfs_list_t *tmp = med->med_open_head.next;
4984 mfd = cfs_list_entry(tmp, struct mdt_file_data, mfd_list);
4986 /* Remove mfd handle so it can't be found again.
4987 * We are consuming the mfd_list reference here. */
4988 class_handle_unhash(&mfd->mfd_handle);
4989 cfs_list_move_tail(&mfd->mfd_list, &closing_list);
4991 spin_unlock(&med->med_open_lock);
4992 mdt = mdt_dev(obd->obd_lu_dev);
4993 LASSERT(mdt != NULL);
4995 rc = lu_env_init(&env, LCT_MD_THREAD);
4999 info = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
5000 LASSERT(info != NULL);
5001 memset(info, 0, sizeof *info);
5002 info->mti_env = &env;
5003 info->mti_mdt = mdt;
5004 info->mti_exp = exp;
5006 if (!cfs_list_empty(&closing_list)) {
5007 struct md_attr *ma = &info->mti_attr;
5009 /* Close any open files (which may also cause orphan unlinking). */
5010 cfs_list_for_each_entry_safe(mfd, n, &closing_list, mfd_list) {
5011 cfs_list_del_init(&mfd->mfd_list);
5012 ma->ma_need = ma->ma_valid = 0;
5014 /* This file is being closed due to an eviction, it
5015 * could have been modified and now dirty regarding to
5016 * HSM archive, check this!
5017 * The logic here is to mark a file dirty if there's a
5018 * chance it was dirtied before the client was evicted,
5019 * so that we don't have to wait for a release attempt
5020 * before finding out the file was actually dirty and
5021 * fail the release. Aggressively marking it dirty here
5022 * will cause the policy engine to attempt to
5023 * re-archive it; when rearchiving, we can compare the
5024 * current version to the HSM data_version and make the
5025 * archive request into a noop if it's not actually
5028 if (mfd->mfd_mode & (FMODE_WRITE|MDS_FMODE_TRUNC))
5029 rc = mdt_ctxt_add_dirty_flag(&env, info, mfd);
5031 /* Don't unlink orphan on failover umount, LU-184 */
5032 if (exp->exp_flags & OBD_OPT_FAILOVER) {
5033 ma->ma_valid = MA_FLAGS;
5034 ma->ma_attr_flags |= MDS_KEEP_ORPHAN;
5036 mdt_mfd_close(info, mfd);
5039 info->mti_mdt = NULL;
5040 /* cleanup client slot early */
5041 /* Do not erase record for recoverable client. */
5042 if (!(exp->exp_flags & OBD_OPT_FAILOVER) || exp->exp_failed)
5043 tgt_client_del(&env, exp);
5049 static int mdt_obd_disconnect(struct obd_export *exp)
5055 class_export_get(exp);
5057 rc = server_disconnect_export(exp);
5059 CDEBUG(D_IOCTL, "server disconnect error: %d\n", rc);
5061 rc = mdt_export_cleanup(exp);
5062 class_export_put(exp);
5066 /* FIXME: Can we avoid using these two interfaces? */
5067 static int mdt_init_export(struct obd_export *exp)
5069 struct mdt_export_data *med = &exp->exp_mdt_data;
5073 CFS_INIT_LIST_HEAD(&med->med_open_head);
5074 spin_lock_init(&med->med_open_lock);
5075 mutex_init(&med->med_idmap_mutex);
5076 med->med_idmap = NULL;
5077 spin_lock(&exp->exp_lock);
5078 exp->exp_connecting = 1;
5079 spin_unlock(&exp->exp_lock);
5081 /* self-export doesn't need client data and ldlm initialization */
5082 if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
5083 &exp->exp_client_uuid)))
5086 rc = tgt_client_alloc(exp);
5090 rc = ldlm_init_export(exp);
5097 tgt_client_free(exp);
5099 CERROR("%s: Failed to initialize export: rc = %d\n",
5100 exp->exp_obd->obd_name, rc);
5104 static int mdt_destroy_export(struct obd_export *exp)
5108 if (exp_connect_rmtclient(exp))
5109 mdt_cleanup_idmap(&exp->exp_mdt_data);
5111 target_destroy_export(exp);
5112 /* destroy can be called from failed obd_setup, so
5113 * checking uuid is safer than obd_self_export */
5114 if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
5115 &exp->exp_client_uuid)))
5118 ldlm_destroy_export(exp);
5119 tgt_client_free(exp);
5121 LASSERT(cfs_list_empty(&exp->exp_outstanding_replies));
5122 LASSERT(cfs_list_empty(&exp->exp_mdt_data.med_open_head));
5127 /** The maximum depth that fid2path() will search.
5128 * This is limited only because we want to store the fids for
5129 * historical path lookup purposes.
5131 #define MAX_PATH_DEPTH 100
5133 /** mdt_path() lookup structure. */
5134 struct path_lookup_info {
5135 __u64 pli_recno; /**< history point */
5136 __u64 pli_currec; /**< current record */
5137 struct lu_fid pli_fid;
5138 struct lu_fid pli_fids[MAX_PATH_DEPTH]; /**< path, in fids */
5139 struct mdt_object *pli_mdt_obj;
5140 char *pli_path; /**< full path */
5142 int pli_linkno; /**< which hardlink to follow */
5143 int pli_fidcount; /**< number of \a pli_fids */
5146 static int mdt_links_read(struct mdt_thread_info *info,
5147 struct mdt_object *mdt_obj, struct linkea_data *ldata)
5151 LASSERT(ldata->ld_buf->lb_buf != NULL);
5153 if (!mdt_object_exists(mdt_obj))
5156 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
5157 ldata->ld_buf, XATTR_NAME_LINK);
5158 if (rc == -ERANGE) {
5159 /* Buf was too small, figure out what we need. */
5160 lu_buf_free(ldata->ld_buf);
5161 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
5162 ldata->ld_buf, XATTR_NAME_LINK);
5165 ldata->ld_buf = lu_buf_check_and_alloc(ldata->ld_buf, rc);
5166 if (ldata->ld_buf->lb_buf == NULL)
5168 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
5169 ldata->ld_buf, XATTR_NAME_LINK);
5174 return linkea_init(ldata);
5177 static int mdt_path_current(struct mdt_thread_info *info,
5178 struct path_lookup_info *pli)
5180 struct mdt_device *mdt = info->mti_mdt;
5181 struct mdt_object *mdt_obj;
5182 struct link_ea_header *leh;
5183 struct link_ea_entry *lee;
5184 struct lu_name *tmpname = &info->mti_name;
5185 struct lu_fid *tmpfid = &info->mti_tmp_fid1;
5186 struct lu_buf *buf = &info->mti_big_buf;
5189 struct linkea_data ldata = { 0 };
5193 /* temp buffer for path element, the buffer will be finally freed
5194 * in mdt_thread_info_fini */
5195 buf = lu_buf_check_and_alloc(buf, PATH_MAX);
5196 if (buf->lb_buf == NULL)
5200 ptr = pli->pli_path + pli->pli_pathlen - 1;
5203 pli->pli_fidcount = 0;
5204 pli->pli_fids[0] = *(struct lu_fid *)mdt_object_fid(pli->pli_mdt_obj);
5206 /* root FID only exists on MDT0, and fid2path should also ends at MDT0,
5207 * so checking root_fid can only happen on MDT0. */
5208 while (!lu_fid_eq(&mdt->mdt_md_root_fid,
5209 &pli->pli_fids[pli->pli_fidcount])) {
5210 mdt_obj = mdt_object_find(info->mti_env, mdt,
5211 &pli->pli_fids[pli->pli_fidcount]);
5212 if (IS_ERR(mdt_obj))
5213 GOTO(out, rc = PTR_ERR(mdt_obj));
5214 if (mdt_object_remote(mdt_obj)) {
5215 mdt_object_put(info->mti_env, mdt_obj);
5216 GOTO(remote_out, rc = -EREMOTE);
5218 if (!mdt_object_exists(mdt_obj)) {
5219 mdt_object_put(info->mti_env, mdt_obj);
5220 GOTO(out, rc = -ENOENT);
5223 rc = mdt_links_read(info, mdt_obj, &ldata);
5224 mdt_object_put(info->mti_env, mdt_obj);
5229 lee = (struct link_ea_entry *)(leh + 1); /* link #0 */
5230 linkea_entry_unpack(lee, &reclen, tmpname, tmpfid);
5231 /* If set, use link #linkno for path lookup, otherwise use
5232 link #0. Only do this for the final path element. */
5233 if (pli->pli_fidcount == 0 &&
5234 pli->pli_linkno < leh->leh_reccount) {
5236 for (count = 0; count < pli->pli_linkno; count++) {
5237 lee = (struct link_ea_entry *)
5238 ((char *)lee + reclen);
5239 linkea_entry_unpack(lee, &reclen, tmpname,
5242 if (pli->pli_linkno < leh->leh_reccount - 1)
5243 /* indicate to user there are more links */
5247 /* Pack the name in the end of the buffer */
5248 ptr -= tmpname->ln_namelen;
5249 if (ptr - 1 <= pli->pli_path)
5250 GOTO(out, rc = -EOVERFLOW);
5251 strncpy(ptr, tmpname->ln_name, tmpname->ln_namelen);
5254 /* Store the parent fid for historic lookup */
5255 if (++pli->pli_fidcount >= MAX_PATH_DEPTH)
5256 GOTO(out, rc = -EOVERFLOW);
5257 pli->pli_fids[pli->pli_fidcount] = *tmpfid;
5261 ptr++; /* skip leading / */
5262 memmove(pli->pli_path, ptr, pli->pli_path + pli->pli_pathlen - ptr);
5269 /* Returns the full path to this fid, as of changelog record recno. */
5270 static int mdt_path(struct mdt_thread_info *info, struct mdt_object *obj,
5271 char *path, int pathlen, __u64 *recno, int *linkno,
5274 struct mdt_device *mdt = info->mti_mdt;
5275 struct path_lookup_info *pli;
5283 if (lu_fid_eq(&mdt->mdt_md_root_fid, mdt_object_fid(obj))) {
5292 pli->pli_mdt_obj = obj;
5293 pli->pli_recno = *recno;
5294 pli->pli_path = path;
5295 pli->pli_pathlen = pathlen;
5296 pli->pli_linkno = *linkno;
5298 /* Retry multiple times in case file is being moved */
5299 while (tries-- && rc == -EAGAIN)
5300 rc = mdt_path_current(info, pli);
5302 /* return the last resolved fids to the client, so the client will
5303 * build the left path on another MDT for remote object */
5304 *fid = pli->pli_fids[pli->pli_fidcount];
5306 *recno = pli->pli_currec;
5307 /* Return next link index to caller */
5308 *linkno = pli->pli_linkno;
5315 static int mdt_fid2path(struct mdt_thread_info *info,
5316 struct getinfo_fid2path *fp)
5318 struct mdt_device *mdt = info->mti_mdt;
5319 struct mdt_object *obj;
5323 CDEBUG(D_IOCTL, "path get "DFID" from "LPU64" #%d\n",
5324 PFID(&fp->gf_fid), fp->gf_recno, fp->gf_linkno);
5326 if (!fid_is_sane(&fp->gf_fid))
5329 if (!fid_is_namespace_visible(&fp->gf_fid)) {
5330 CWARN("%s: "DFID" is invalid, sequence should be "
5331 ">= "LPX64"\n", mdt_obd_name(mdt),
5332 PFID(&fp->gf_fid), (__u64)FID_SEQ_NORMAL);
5336 obj = mdt_object_find(info->mti_env, mdt, &fp->gf_fid);
5337 if (obj == NULL || IS_ERR(obj)) {
5338 CDEBUG(D_IOCTL, "no object "DFID": %ld\n", PFID(&fp->gf_fid),
5343 if (mdt_object_remote(obj))
5345 else if (!mdt_object_exists(obj))
5351 mdt_object_put(info->mti_env, obj);
5352 CDEBUG(D_IOCTL, "nonlocal object "DFID": %d\n",
5353 PFID(&fp->gf_fid), rc);
5357 rc = mdt_path(info, obj, fp->gf_path, fp->gf_pathlen, &fp->gf_recno,
5358 &fp->gf_linkno, &fp->gf_fid);
5360 CDEBUG(D_INFO, "fid "DFID", path %s recno "LPX64" linkno %u\n",
5361 PFID(&fp->gf_fid), fp->gf_path, fp->gf_recno, fp->gf_linkno);
5363 mdt_object_put(info->mti_env, obj);
5368 static int mdt_rpc_fid2path(struct mdt_thread_info *info, void *key,
5369 void *val, int vallen)
5371 struct getinfo_fid2path *fpout, *fpin;
5374 fpin = key + cfs_size_round(sizeof(KEY_FID2PATH));
5377 if (ptlrpc_req_need_swab(info->mti_pill->rc_req))
5378 lustre_swab_fid2path(fpin);
5380 memcpy(fpout, fpin, sizeof(*fpin));
5381 if (fpout->gf_pathlen != vallen - sizeof(*fpin))
5384 rc = mdt_fid2path(info, fpout);
5388 int mdt_get_info(struct tgt_session_info *tsi)
5398 key = req_capsule_client_get(tsi->tsi_pill, &RMF_GETINFO_KEY);
5400 CDEBUG(D_IOCTL, "No GETINFO key");
5401 RETURN(err_serious(-EFAULT));
5403 keylen = req_capsule_get_size(tsi->tsi_pill, &RMF_GETINFO_KEY,
5406 vallen = req_capsule_client_get(tsi->tsi_pill, &RMF_GETINFO_VALLEN);
5407 if (vallen == NULL) {
5408 CDEBUG(D_IOCTL, "Unable to get RMF_GETINFO_VALLEN buffer");
5409 RETURN(err_serious(-EFAULT));
5412 req_capsule_set_size(tsi->tsi_pill, &RMF_GETINFO_VAL, RCL_SERVER,
5414 rc = req_capsule_server_pack(tsi->tsi_pill);
5416 RETURN(err_serious(rc));
5418 valout = req_capsule_server_get(tsi->tsi_pill, &RMF_GETINFO_VAL);
5419 if (valout == NULL) {
5420 CDEBUG(D_IOCTL, "Unable to get get-info RPC out buffer");
5421 RETURN(err_serious(-EFAULT));
5424 if (KEY_IS(KEY_FID2PATH)) {
5425 struct mdt_thread_info *info = tsi2mdt_info(tsi);
5427 rc = mdt_rpc_fid2path(info, key, valout, *vallen);
5428 mdt_thread_info_fini(info);
5435 /* Pass the ioc down */
5436 static int mdt_ioc_child(struct lu_env *env, struct mdt_device *mdt,
5437 unsigned int cmd, int len, void *data)
5439 struct lu_context ioctl_session;
5440 struct md_device *next = mdt->mdt_child;
5444 rc = lu_context_init(&ioctl_session, LCT_SERVER_SESSION);
5447 ioctl_session.lc_thread = (struct ptlrpc_thread *)current;
5448 lu_context_enter(&ioctl_session);
5449 env->le_ses = &ioctl_session;
5451 LASSERT(next->md_ops->mdo_iocontrol);
5452 rc = next->md_ops->mdo_iocontrol(env, next, cmd, len, data);
5454 lu_context_exit(&ioctl_session);
5455 lu_context_fini(&ioctl_session);
5459 static int mdt_ioc_version_get(struct mdt_thread_info *mti, void *karg)
5461 struct obd_ioctl_data *data = karg;
5464 struct mdt_object *obj;
5465 struct mdt_lock_handle *lh;
5469 if (data->ioc_inlbuf1 == NULL || data->ioc_inllen1 != sizeof(*fid) ||
5470 data->ioc_inlbuf2 == NULL || data->ioc_inllen2 != sizeof(version))
5473 fid = (struct lu_fid *)data->ioc_inlbuf1;
5475 if (!fid_is_sane(fid))
5478 CDEBUG(D_IOCTL, "getting version for "DFID"\n", PFID(fid));
5480 lh = &mti->mti_lh[MDT_LH_PARENT];
5481 mdt_lock_reg_init(lh, LCK_CR);
5483 obj = mdt_object_find_lock(mti, fid, lh, MDS_INODELOCK_UPDATE);
5485 RETURN(PTR_ERR(obj));
5487 if (mdt_object_remote(obj)) {
5490 * before calling version get the correct MDS should be
5491 * fid, this is error to find remote object here
5493 CERROR("nonlocal object "DFID"\n", PFID(fid));
5494 } else if (!mdt_object_exists(obj)) {
5495 *(__u64 *)data->ioc_inlbuf2 = ENOENT_VERSION;
5498 version = dt_version_get(mti->mti_env, mdt_obj2dt(obj));
5499 *(__u64 *)data->ioc_inlbuf2 = version;
5502 mdt_object_unlock_put(mti, obj, lh, 1);
5506 /* ioctls on obd dev */
5507 static int mdt_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
5508 void *karg, void *uarg)
5511 struct obd_device *obd = exp->exp_obd;
5512 struct mdt_device *mdt = mdt_dev(obd->obd_lu_dev);
5513 struct dt_device *dt = mdt->mdt_bottom;
5517 CDEBUG(D_IOCTL, "handling ioctl cmd %#x\n", cmd);
5518 rc = lu_env_init(&env, LCT_MD_THREAD);
5524 rc = mdt_device_sync(&env, mdt);
5526 case OBD_IOC_SET_READONLY:
5527 rc = dt->dd_ops->dt_ro(&env, dt);
5529 case OBD_IOC_ABORT_RECOVERY:
5530 CERROR("%s: Aborting recovery for device\n", mdt_obd_name(mdt));
5531 target_stop_recovery_thread(obd);
5534 case OBD_IOC_CHANGELOG_REG:
5535 case OBD_IOC_CHANGELOG_DEREG:
5536 case OBD_IOC_CHANGELOG_CLEAR:
5537 rc = mdt_ioc_child(&env, mdt, cmd, len, karg);
5539 case OBD_IOC_START_LFSCK: {
5540 struct md_device *next = mdt->mdt_child;
5541 struct obd_ioctl_data *data = karg;
5542 struct lfsck_start_param lsp;
5544 if (unlikely(data == NULL)) {
5549 lsp.lsp_start = (struct lfsck_start *)(data->ioc_inlbuf1);
5550 lsp.lsp_namespace = mdt->mdt_namespace;
5551 rc = next->md_ops->mdo_iocontrol(&env, next, cmd, 0, &lsp);
5554 case OBD_IOC_STOP_LFSCK: {
5555 struct md_device *next = mdt->mdt_child;
5557 rc = next->md_ops->mdo_iocontrol(&env, next, cmd, 0, NULL);
5560 case OBD_IOC_GET_OBJ_VERSION: {
5561 struct mdt_thread_info *mti;
5562 mti = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
5563 memset(mti, 0, sizeof *mti);
5564 mti->mti_env = &env;
5568 rc = mdt_ioc_version_get(mti, karg);
5571 case OBD_IOC_CATLOGLIST:
5572 rc = llog_catalog_list(&env, mdt->mdt_bottom, 0, karg);
5576 CERROR("%s: Not supported cmd = %d, rc = %d\n",
5577 mdt_obd_name(mdt), cmd, rc);
5584 int mdt_postrecov(const struct lu_env *env, struct mdt_device *mdt)
5586 struct lu_device *ld = md2lu_dev(mdt->mdt_child);
5590 rc = ld->ld_ops->ldo_recovery_complete(env, ld);
5594 int mdt_obd_postrecov(struct obd_device *obd)
5599 rc = lu_env_init(&env, LCT_MD_THREAD);
5602 rc = mdt_postrecov(&env, mdt_dev(obd->obd_lu_dev));
5607 static struct obd_ops mdt_obd_device_ops = {
5608 .o_owner = THIS_MODULE,
5609 .o_set_info_async = mdt_obd_set_info_async,
5610 .o_connect = mdt_obd_connect,
5611 .o_reconnect = mdt_obd_reconnect,
5612 .o_disconnect = mdt_obd_disconnect,
5613 .o_init_export = mdt_init_export,
5614 .o_destroy_export = mdt_destroy_export,
5615 .o_iocontrol = mdt_iocontrol,
5616 .o_postrecov = mdt_obd_postrecov,
5619 static struct lu_device* mdt_device_fini(const struct lu_env *env,
5620 struct lu_device *d)
5622 struct mdt_device *m = mdt_dev(d);
5629 static struct lu_device *mdt_device_free(const struct lu_env *env,
5630 struct lu_device *d)
5632 struct mdt_device *m = mdt_dev(d);
5635 lu_device_fini(&m->mdt_lu_dev);
5641 static struct lu_device *mdt_device_alloc(const struct lu_env *env,
5642 struct lu_device_type *t,
5643 struct lustre_cfg *cfg)
5645 struct lu_device *l;
5646 struct mdt_device *m;
5653 rc = mdt_init0(env, m, t, cfg);
5655 mdt_device_free(env, l);
5660 l = ERR_PTR(-ENOMEM);
5664 /* context key constructor/destructor: mdt_key_init, mdt_key_fini */
5665 LU_KEY_INIT(mdt, struct mdt_thread_info);
5667 static void mdt_key_fini(const struct lu_context *ctx,
5668 struct lu_context_key *key, void* data)
5670 struct mdt_thread_info *info = data;
5672 if (info->mti_big_lmm) {
5673 OBD_FREE_LARGE(info->mti_big_lmm, info->mti_big_lmmsize);
5674 info->mti_big_lmm = NULL;
5675 info->mti_big_lmmsize = 0;
5680 /* context key: mdt_thread_key */
5681 LU_CONTEXT_KEY_DEFINE(mdt, LCT_MD_THREAD);
5683 struct lu_ucred *mdt_ucred(const struct mdt_thread_info *info)
5685 return lu_ucred(info->mti_env);
5688 struct lu_ucred *mdt_ucred_check(const struct mdt_thread_info *info)
5690 return lu_ucred_check(info->mti_env);
5694 * Enable/disable COS (Commit On Sharing).
5696 * Set/Clear the COS flag in mdt options.
5698 * \param mdt mdt device
5699 * \param val 0 disables COS, other values enable COS
5701 void mdt_enable_cos(struct mdt_device *mdt, int val)
5706 mdt->mdt_opts.mo_cos = !!val;
5707 rc = lu_env_init(&env, LCT_LOCAL);
5708 if (unlikely(rc != 0)) {
5709 CWARN("lu_env initialization failed with rc = %d,"
5710 "cannot sync\n", rc);
5713 mdt_device_sync(&env, mdt);
5718 * Check COS (Commit On Sharing) status.
5720 * Return COS flag status.
5722 * \param mdt mdt device
5724 int mdt_cos_is_enabled(struct mdt_device *mdt)
5726 return mdt->mdt_opts.mo_cos != 0;
5729 static struct lu_device_type_operations mdt_device_type_ops = {
5730 .ldto_device_alloc = mdt_device_alloc,
5731 .ldto_device_free = mdt_device_free,
5732 .ldto_device_fini = mdt_device_fini
5735 static struct lu_device_type mdt_device_type = {
5736 .ldt_tags = LU_DEVICE_MD,
5737 .ldt_name = LUSTRE_MDT_NAME,
5738 .ldt_ops = &mdt_device_type_ops,
5739 .ldt_ctx_tags = LCT_MD_THREAD
5742 static int __init mdt_mod_init(void)
5744 struct lprocfs_static_vars lvars;
5747 CLASSERT(sizeof("0x0123456789ABCDEF:0x01234567:0x01234567") ==
5748 FID_NOBRACE_LEN + 1);
5749 CLASSERT(sizeof("[0x0123456789ABCDEF:0x01234567:0x01234567]") ==
5751 rc = lu_kmem_init(mdt_caches);
5755 rc = mds_mod_init();
5759 lprocfs_mdt_init_vars(&lvars);
5760 rc = class_register_type(&mdt_obd_device_ops, NULL, NULL,
5761 #ifndef HAVE_ONLY_PROCFS_SEQ
5764 LUSTRE_MDT_NAME, &mdt_device_type);
5769 lu_kmem_fini(mdt_caches);
5776 static void __exit mdt_mod_exit(void)
5778 class_unregister_type(LUSTRE_MDT_NAME);
5780 lu_kmem_fini(mdt_caches);
5783 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
5784 MODULE_DESCRIPTION("Lustre Metadata Target ("LUSTRE_MDT_NAME")");
5785 MODULE_LICENSE("GPL");
5787 cfs_module(mdt, LUSTRE_VERSION_STRING, mdt_mod_init, mdt_mod_exit);