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_lfsck.h>
72 mdl_mode_t mdt_mdl_lock_modes[] = {
73 [LCK_MINMODE] = MDL_MINMODE,
80 [LCK_GROUP] = MDL_GROUP
83 ldlm_mode_t mdt_dlm_lock_modes[] = {
84 [MDL_MINMODE] = LCK_MINMODE,
91 [MDL_GROUP] = LCK_GROUP
94 static struct mdt_device *mdt_dev(struct lu_device *d);
95 static int mdt_unpack_req_pack_rep(struct mdt_thread_info *info, __u32 flags);
97 static const struct lu_object_operations mdt_obj_ops;
99 /* Slab for MDT object allocation */
100 static struct kmem_cache *mdt_object_kmem;
102 /* For HSM restore handles */
103 struct kmem_cache *mdt_hsm_cdt_kmem;
105 /* For HSM request handles */
106 struct kmem_cache *mdt_hsm_car_kmem;
108 static struct lu_kmem_descr mdt_caches[] = {
110 .ckd_cache = &mdt_object_kmem,
111 .ckd_name = "mdt_obj",
112 .ckd_size = sizeof(struct mdt_object)
115 .ckd_cache = &mdt_hsm_cdt_kmem,
116 .ckd_name = "mdt_cdt_restore_handle",
117 .ckd_size = sizeof(struct cdt_restore_handle)
120 .ckd_cache = &mdt_hsm_car_kmem,
121 .ckd_name = "mdt_cdt_agent_req",
122 .ckd_size = sizeof(struct cdt_agent_req)
129 int mdt_get_disposition(struct ldlm_reply *rep, int flag)
133 return (rep->lock_policy_res1 & flag);
136 void mdt_clear_disposition(struct mdt_thread_info *info,
137 struct ldlm_reply *rep, int flag)
140 info->mti_opdata &= ~flag;
141 tgt_opdata_clear(info->mti_env, flag);
144 rep->lock_policy_res1 &= ~flag;
147 void mdt_set_disposition(struct mdt_thread_info *info,
148 struct ldlm_reply *rep, int flag)
151 info->mti_opdata |= flag;
152 tgt_opdata_set(info->mti_env, flag);
155 rep->lock_policy_res1 |= flag;
158 void mdt_lock_reg_init(struct mdt_lock_handle *lh, ldlm_mode_t lm)
160 lh->mlh_pdo_hash = 0;
161 lh->mlh_reg_mode = lm;
162 lh->mlh_rreg_mode = lm;
163 lh->mlh_type = MDT_REG_LOCK;
166 void mdt_lock_pdo_init(struct mdt_lock_handle *lh, ldlm_mode_t lock_mode,
167 const struct lu_name *lname)
169 lh->mlh_reg_mode = lock_mode;
170 lh->mlh_rreg_mode = lock_mode;
171 lh->mlh_type = MDT_PDO_LOCK;
173 if (lu_name_is_valid(lname)) {
174 lh->mlh_pdo_hash = full_name_hash(lname->ln_name,
176 /* XXX Workaround for LU-2856
178 * Zero is a valid return value of full_name_hash, but
179 * several users of mlh_pdo_hash assume a non-zero
180 * hash value. We therefore map zero onto an
181 * arbitrary, but consistent value (1) to avoid
182 * problems further down the road. */
183 if (unlikely(lh->mlh_pdo_hash == 0))
184 lh->mlh_pdo_hash = 1;
186 lh->mlh_pdo_hash = 0;
190 static void mdt_lock_pdo_mode(struct mdt_thread_info *info, struct mdt_object *o,
191 struct mdt_lock_handle *lh)
197 * Any dir access needs couple of locks:
199 * 1) on part of dir we gonna take lookup/modify;
201 * 2) on whole dir to protect it from concurrent splitting and/or to
202 * flush client's cache for readdir().
204 * so, for a given mode and object this routine decides what lock mode
205 * to use for lock #2:
207 * 1) if caller's gonna lookup in dir then we need to protect dir from
208 * being splitted only - LCK_CR
210 * 2) if caller's gonna modify dir then we need to protect dir from
211 * being splitted and to flush cache - LCK_CW
213 * 3) if caller's gonna modify dir and that dir seems ready for
214 * splitting then we need to protect it from any type of access
215 * (lookup/modify/split) - LCK_EX --bzzz
218 LASSERT(lh->mlh_reg_mode != LCK_MINMODE);
219 LASSERT(lh->mlh_pdo_mode == LCK_MINMODE);
222 * Ask underlaying level its opinion about preferable PDO lock mode
223 * having access type passed as regular lock mode:
225 * - MDL_MINMODE means that lower layer does not want to specify lock
228 * - MDL_NL means that no PDO lock should be taken. This is used in some
229 * cases. Say, for non-splittable directories no need to use PDO locks
232 mode = mdo_lock_mode(info->mti_env, mdt_object_child(o),
233 mdt_dlm_mode2mdl_mode(lh->mlh_reg_mode));
235 if (mode != MDL_MINMODE) {
236 lh->mlh_pdo_mode = mdt_mdl_mode2dlm_mode(mode);
239 * Lower layer does not want to specify locking mode. We do it
240 * our selves. No special protection is needed, just flush
241 * client's cache on modification and allow concurrent
244 switch (lh->mlh_reg_mode) {
246 lh->mlh_pdo_mode = LCK_EX;
249 lh->mlh_pdo_mode = LCK_CR;
252 lh->mlh_pdo_mode = LCK_CW;
255 CERROR("Not expected lock type (0x%x)\n",
256 (int)lh->mlh_reg_mode);
261 LASSERT(lh->mlh_pdo_mode != LCK_MINMODE);
265 int mdt_getstatus(struct tgt_session_info *tsi)
267 struct mdt_thread_info *info = tsi2mdt_info(tsi);
268 struct mdt_device *mdt = info->mti_mdt;
269 struct mdt_body *repbody;
274 rc = mdt_check_ucred(info);
276 GOTO(out, rc = err_serious(rc));
278 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK))
279 GOTO(out, rc = err_serious(-ENOMEM));
281 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
282 repbody->fid1 = mdt->mdt_md_root_fid;
283 repbody->valid |= OBD_MD_FLID;
285 if (tsi->tsi_tgt->lut_mds_capa &&
286 exp_connect_flags(info->mti_exp) & OBD_CONNECT_MDS_CAPA) {
287 struct mdt_object *root;
288 struct lustre_capa *capa;
290 root = mdt_object_find(info->mti_env, mdt, &repbody->fid1);
292 GOTO(out, rc = PTR_ERR(root));
294 capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA1);
296 capa->lc_opc = CAPA_OPC_MDS_DEFAULT;
297 rc = mo_capa_get(info->mti_env, mdt_object_child(root), capa,
299 mdt_object_put(info->mti_env, root);
301 repbody->valid |= OBD_MD_FLMDSCAPA;
305 mdt_thread_info_fini(info);
309 int mdt_statfs(struct tgt_session_info *tsi)
311 struct ptlrpc_request *req = tgt_ses_req(tsi);
312 struct mdt_thread_info *info = tsi2mdt_info(tsi);
313 struct md_device *next = info->mti_mdt->mdt_child;
314 struct ptlrpc_service_part *svcpt;
315 struct obd_statfs *osfs;
320 svcpt = req->rq_rqbd->rqbd_svcpt;
322 /* This will trigger a watchdog timeout */
323 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_STATFS_LCW_SLEEP,
324 (MDT_SERVICE_WATCHDOG_FACTOR *
325 at_get(&svcpt->scp_at_estimate)) + 1);
327 rc = mdt_check_ucred(info);
329 GOTO(out, rc = err_serious(rc));
331 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK))
332 GOTO(out, rc = err_serious(-ENOMEM));
334 osfs = req_capsule_server_get(info->mti_pill, &RMF_OBD_STATFS);
336 GOTO(out, rc = -EPROTO);
338 /** statfs information are cached in the mdt_device */
339 if (cfs_time_before_64(info->mti_mdt->mdt_osfs_age,
340 cfs_time_shift_64(-OBD_STATFS_CACHE_SECONDS))) {
341 /** statfs data is too old, get up-to-date one */
342 rc = next->md_ops->mdo_statfs(info->mti_env, next, osfs);
345 spin_lock(&info->mti_mdt->mdt_osfs_lock);
346 info->mti_mdt->mdt_osfs = *osfs;
347 info->mti_mdt->mdt_osfs_age = cfs_time_current_64();
348 spin_unlock(&info->mti_mdt->mdt_osfs_lock);
350 /** use cached statfs data */
351 spin_lock(&info->mti_mdt->mdt_osfs_lock);
352 *osfs = info->mti_mdt->mdt_osfs;
353 spin_unlock(&info->mti_mdt->mdt_osfs_lock);
357 mdt_counter_incr(req, LPROC_MDT_STATFS);
359 mdt_thread_info_fini(info);
364 * Pack SOM attributes into the reply.
365 * Call under a DLM UPDATE lock.
367 static void mdt_pack_size2body(struct mdt_thread_info *info,
368 struct mdt_object *mo)
371 struct md_attr *ma = &info->mti_attr;
373 LASSERT(ma->ma_attr.la_valid & LA_MODE);
374 b = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
376 /* Check if Size-on-MDS is supported, if this is a regular file,
377 * if SOM is enabled on the object and if SOM cache exists and valid.
378 * Otherwise do not pack Size-on-MDS attributes to the reply. */
379 if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) ||
380 !S_ISREG(ma->ma_attr.la_mode) ||
381 !mdt_object_is_som_enabled(mo) ||
382 !(ma->ma_valid & MA_SOM))
385 b->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
386 b->size = ma->ma_som->msd_size;
387 b->blocks = ma->ma_som->msd_blocks;
390 void mdt_pack_attr2body(struct mdt_thread_info *info, struct mdt_body *b,
391 const struct lu_attr *attr, const struct lu_fid *fid)
393 struct md_attr *ma = &info->mti_attr;
395 LASSERT(ma->ma_valid & MA_INODE);
397 b->atime = attr->la_atime;
398 b->mtime = attr->la_mtime;
399 b->ctime = attr->la_ctime;
400 b->mode = attr->la_mode;
401 b->size = attr->la_size;
402 b->blocks = attr->la_blocks;
403 b->uid = attr->la_uid;
404 b->gid = attr->la_gid;
405 b->flags = attr->la_flags;
406 b->nlink = attr->la_nlink;
407 b->rdev = attr->la_rdev;
409 /*XXX should pack the reply body according to lu_valid*/
410 b->valid |= OBD_MD_FLCTIME | OBD_MD_FLUID |
411 OBD_MD_FLGID | OBD_MD_FLTYPE |
412 OBD_MD_FLMODE | OBD_MD_FLNLINK | OBD_MD_FLFLAGS |
413 OBD_MD_FLATIME | OBD_MD_FLMTIME ;
415 if (!S_ISREG(attr->la_mode)) {
416 b->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS | OBD_MD_FLRDEV;
417 } else if (ma->ma_need & MA_LOV && !(ma->ma_valid & MA_LOV)) {
418 /* means no objects are allocated on osts. */
419 LASSERT(!(ma->ma_valid & MA_LOV));
420 /* just ignore blocks occupied by extend attributes on MDS */
422 /* if no object is allocated on osts, the size on mds is valid. b=22272 */
423 b->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
424 } else if ((ma->ma_valid & MA_LOV) && ma->ma_lmm != NULL &&
425 ma->ma_lmm->lmm_pattern & LOV_PATTERN_F_RELEASED) {
426 /* A released file stores its size on MDS. */
427 /* But return 1 block for released file, unless tools like tar
428 * will consider it fully sparse. (LU-3864)
430 if (unlikely(b->size == 0))
434 b->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
439 b->valid |= OBD_MD_FLID;
440 CDEBUG(D_INODE, DFID": nlink=%d, mode=%o, size="LPU64"\n",
441 PFID(fid), b->nlink, b->mode, b->size);
445 mdt_body_reverse_idmap(info, b);
447 if (fid != NULL && (b->valid & OBD_MD_FLSIZE))
448 CDEBUG(D_VFSTRACE, DFID": returning size %llu\n",
449 PFID(fid), (unsigned long long)b->size);
452 static inline int mdt_body_has_lov(const struct lu_attr *la,
453 const struct mdt_body *body)
455 return ((S_ISREG(la->la_mode) && (body->valid & OBD_MD_FLEASIZE)) ||
456 (S_ISDIR(la->la_mode) && (body->valid & OBD_MD_FLDIREA )) );
459 void mdt_client_compatibility(struct mdt_thread_info *info)
461 struct mdt_body *body;
462 struct ptlrpc_request *req = mdt_info_req(info);
463 struct obd_export *exp = req->rq_export;
464 struct md_attr *ma = &info->mti_attr;
465 struct lu_attr *la = &ma->ma_attr;
468 if (exp_connect_layout(exp))
469 /* the client can deal with 16-bit lmm_stripe_count */
472 body = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
474 if (!mdt_body_has_lov(la, body))
477 /* now we have a reply with a lov for a client not compatible with the
478 * layout lock so we have to clean the layout generation number */
479 if (S_ISREG(la->la_mode))
480 ma->ma_lmm->lmm_layout_gen = 0;
484 static int mdt_big_xattr_get(struct mdt_thread_info *info, struct mdt_object *o,
487 const struct lu_env *env = info->mti_env;
491 LASSERT(info->mti_big_lmm_used == 0);
492 rc = mo_xattr_get(env, mdt_object_child(o), &LU_BUF_NULL, name);
496 /* big_lmm may need to be grown */
497 if (info->mti_big_lmmsize < rc) {
498 int size = size_roundup_power2(rc);
500 if (info->mti_big_lmmsize > 0) {
501 /* free old buffer */
502 LASSERT(info->mti_big_lmm);
503 OBD_FREE_LARGE(info->mti_big_lmm,
504 info->mti_big_lmmsize);
505 info->mti_big_lmm = NULL;
506 info->mti_big_lmmsize = 0;
509 OBD_ALLOC_LARGE(info->mti_big_lmm, size);
510 if (info->mti_big_lmm == NULL)
512 info->mti_big_lmmsize = size;
514 LASSERT(info->mti_big_lmmsize >= rc);
516 info->mti_buf.lb_buf = info->mti_big_lmm;
517 info->mti_buf.lb_len = info->mti_big_lmmsize;
518 rc = mo_xattr_get(env, mdt_object_child(o), &info->mti_buf, name);
523 int mdt_stripe_get(struct mdt_thread_info *info, struct mdt_object *o,
524 struct md_attr *ma, const char *name)
526 struct md_object *next = mdt_object_child(o);
527 struct lu_buf *buf = &info->mti_buf;
530 if (strcmp(name, XATTR_NAME_LOV) == 0) {
531 buf->lb_buf = ma->ma_lmm;
532 buf->lb_len = ma->ma_lmm_size;
533 LASSERT(!(ma->ma_valid & MA_LOV));
534 } else if (strcmp(name, XATTR_NAME_LMV) == 0) {
535 buf->lb_buf = ma->ma_lmv;
536 buf->lb_len = ma->ma_lmv_size;
537 LASSERT(!(ma->ma_valid & MA_LMV));
538 } else if (strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0) {
539 buf->lb_buf = ma->ma_lmv;
540 buf->lb_len = ma->ma_lmv_size;
541 LASSERT(!(ma->ma_valid & MA_LMV_DEF));
546 rc = mo_xattr_get(info->mti_env, next, buf, name);
548 if (strcmp(name, XATTR_NAME_LOV) == 0) {
549 ma->ma_lmm_size = rc;
550 ma->ma_valid |= MA_LOV;
551 } else if (strcmp(name, XATTR_NAME_LMV) == 0) {
552 ma->ma_lmv_size = rc;
553 ma->ma_valid |= MA_LMV;
554 } else if (strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0) {
555 ma->ma_lmv_size = rc;
556 ma->ma_valid |= MA_LMV_DEF;
560 } else if (rc == -ENODATA) {
563 } else if (rc == -ERANGE) {
564 /* Default LMV has fixed size, so it must be able to fit
565 * in the original buffer */
566 if (strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0)
568 rc = mdt_big_xattr_get(info, o, name);
570 info->mti_big_lmm_used = 1;
571 if (!strcmp(name, XATTR_NAME_LOV)) {
572 ma->ma_valid |= MA_LOV;
573 ma->ma_lmm = info->mti_big_lmm;
574 ma->ma_lmm_size = rc;
575 } else if (!strcmp(name, XATTR_NAME_LMV)) {
576 ma->ma_valid |= MA_LMV;
577 ma->ma_lmv = info->mti_big_lmm;
578 ma->ma_lmv_size = rc;
583 /* update mdt_max_mdsize so all clients
584 * will be aware about that */
585 if (info->mti_mdt->mdt_max_mdsize < rc)
586 info->mti_mdt->mdt_max_mdsize = rc;
594 int mdt_attr_get_pfid(struct mdt_thread_info *info,
595 struct mdt_object *o, struct lu_fid *pfid)
597 struct lu_buf *buf = &info->mti_buf;
598 struct link_ea_header *leh;
599 struct link_ea_entry *lee;
603 buf->lb_buf = info->mti_big_lmm;
604 buf->lb_len = info->mti_big_lmmsize;
605 rc = mo_xattr_get(info->mti_env, mdt_object_child(o),
606 buf, XATTR_NAME_LINK);
607 /* ignore errors, MA_PFID won't be set and it is
608 * up to the caller to treat this as an error */
609 if (rc == -ERANGE || buf->lb_len == 0) {
610 rc = mdt_big_xattr_get(info, o, XATTR_NAME_LINK);
611 buf->lb_buf = info->mti_big_lmm;
612 buf->lb_len = info->mti_big_lmmsize;
617 if (rc < sizeof(*leh)) {
618 CERROR("short LinkEA on "DFID": rc = %d\n",
619 PFID(mdt_object_fid(o)), rc);
623 leh = (struct link_ea_header *) buf->lb_buf;
624 lee = (struct link_ea_entry *)(leh + 1);
625 if (leh->leh_magic == __swab32(LINK_EA_MAGIC)) {
626 leh->leh_magic = LINK_EA_MAGIC;
627 leh->leh_reccount = __swab32(leh->leh_reccount);
628 leh->leh_len = __swab64(leh->leh_len);
630 if (leh->leh_magic != LINK_EA_MAGIC)
632 if (leh->leh_reccount == 0)
635 memcpy(pfid, &lee->lee_parent_fid, sizeof(*pfid));
636 fid_be_to_cpu(pfid, pfid);
641 int mdt_attr_get_complex(struct mdt_thread_info *info,
642 struct mdt_object *o, struct md_attr *ma)
644 const struct lu_env *env = info->mti_env;
645 struct md_object *next = mdt_object_child(o);
646 struct lu_buf *buf = &info->mti_buf;
647 u32 mode = lu_object_attr(&next->mo_lu);
648 int need = ma->ma_need;
654 if (need & MA_INODE) {
655 ma->ma_need = MA_INODE;
656 rc = mo_attr_get(env, next, ma);
659 ma->ma_valid |= MA_INODE;
662 if (need & MA_PFID) {
663 rc = mdt_attr_get_pfid(info, o, &ma->ma_pfid);
665 ma->ma_valid |= MA_PFID;
666 /* ignore this error, parent fid is not mandatory */
670 if (need & MA_LOV && (S_ISREG(mode) || S_ISDIR(mode))) {
671 rc = mdt_stripe_get(info, o, ma, XATTR_NAME_LOV);
676 if (need & MA_LMV && S_ISDIR(mode)) {
677 rc = mdt_stripe_get(info, o, ma, XATTR_NAME_LMV);
682 if (need & MA_LMV_DEF && S_ISDIR(mode)) {
683 rc = mdt_stripe_get(info, o, ma, XATTR_NAME_DEFAULT_LMV);
688 if (need & MA_SOM && S_ISREG(mode)) {
689 buf->lb_buf = info->mti_xattr_buf;
690 buf->lb_len = sizeof(info->mti_xattr_buf);
691 CLASSERT(sizeof(struct som_attrs) <=
692 sizeof(info->mti_xattr_buf));
693 rc2 = mo_xattr_get(info->mti_env, next, buf, XATTR_NAME_SOM);
694 rc2 = lustre_buf2som(info->mti_xattr_buf, rc2, ma->ma_som);
696 ma->ma_valid |= MA_SOM;
697 else if (rc2 < 0 && rc2 != -ENODATA)
701 if (need & MA_HSM && S_ISREG(mode)) {
702 buf->lb_buf = info->mti_xattr_buf;
703 buf->lb_len = sizeof(info->mti_xattr_buf);
704 CLASSERT(sizeof(struct hsm_attrs) <=
705 sizeof(info->mti_xattr_buf));
706 rc2 = mo_xattr_get(info->mti_env, next, buf, XATTR_NAME_HSM);
707 rc2 = lustre_buf2hsm(info->mti_xattr_buf, rc2, &ma->ma_hsm);
709 ma->ma_valid |= MA_HSM;
710 else if (rc2 < 0 && rc2 != -ENODATA)
714 #ifdef CONFIG_FS_POSIX_ACL
715 if (need & MA_ACL_DEF && S_ISDIR(mode)) {
716 buf->lb_buf = ma->ma_acl;
717 buf->lb_len = ma->ma_acl_size;
718 rc2 = mo_xattr_get(env, next, buf, XATTR_NAME_ACL_DEFAULT);
720 ma->ma_acl_size = rc2;
721 ma->ma_valid |= MA_ACL_DEF;
722 } else if (rc2 == -ENODATA) {
731 CDEBUG(D_INODE, "after getattr rc = %d, ma_valid = "LPX64" ma_lmm=%p\n",
732 rc, ma->ma_valid, ma->ma_lmm);
736 static int mdt_getattr_internal(struct mdt_thread_info *info,
737 struct mdt_object *o, int ma_need)
739 struct md_object *next = mdt_object_child(o);
740 const struct mdt_body *reqbody = info->mti_body;
741 struct ptlrpc_request *req = mdt_info_req(info);
742 struct md_attr *ma = &info->mti_attr;
743 struct lu_attr *la = &ma->ma_attr;
744 struct req_capsule *pill = info->mti_pill;
745 const struct lu_env *env = info->mti_env;
746 struct mdt_body *repbody;
747 struct lu_buf *buffer = &info->mti_buf;
752 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK))
753 RETURN(err_serious(-ENOMEM));
755 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
759 if (mdt_object_remote(o)) {
760 /* This object is located on remote node.*/
761 /* Return -EIO for old client */
762 if (!mdt_is_dne_client(req->rq_export))
763 GOTO(out, rc = -EIO);
765 repbody->fid1 = *mdt_object_fid(o);
766 repbody->valid = OBD_MD_FLID | OBD_MD_MDS;
770 buffer->lb_len = reqbody->eadatasize;
771 if (buffer->lb_len > 0) {
772 buffer->lb_buf = req_capsule_server_get(pill, &RMF_MDT_MD);
773 if (buffer->lb_buf == NULL)
774 GOTO(out, rc = -EPROTO);
776 buffer->lb_buf = NULL;
777 ma_need &= ~(MA_LOV | MA_LMV);
778 CDEBUG(D_INFO, "%s: RPC from %s: does not need LOVEA.\n",
779 mdt_obd_name(info->mti_mdt),
780 req->rq_export->exp_client_uuid.uuid);
783 /* If it is dir object and client require MEA, then we got MEA */
784 if (S_ISDIR(lu_object_attr(&next->mo_lu)) &&
785 (reqbody->valid & (OBD_MD_MEA | OBD_MD_DEFAULT_MEA))) {
786 /* Assumption: MDT_MD size is enough for lmv size. */
787 ma->ma_lmv = buffer->lb_buf;
788 ma->ma_lmv_size = buffer->lb_len;
789 ma->ma_need = MA_INODE;
790 if (ma->ma_lmv_size > 0) {
791 if (reqbody->valid & OBD_MD_MEA)
792 ma->ma_need |= MA_LMV;
793 else if (reqbody->valid & OBD_MD_DEFAULT_MEA)
794 ma->ma_need |= MA_LMV_DEF;
797 ma->ma_lmm = buffer->lb_buf;
798 ma->ma_lmm_size = buffer->lb_len;
799 ma->ma_need = MA_INODE | MA_HSM;
800 if (ma->ma_lmm_size > 0)
801 ma->ma_need |= MA_LOV;
804 if (S_ISDIR(lu_object_attr(&next->mo_lu)) &&
805 reqbody->valid & OBD_MD_FLDIREA &&
806 lustre_msg_get_opc(req->rq_reqmsg) == MDS_GETATTR) {
807 /* get default stripe info for this dir. */
808 ma->ma_need |= MA_LOV_DEF;
810 ma->ma_need |= ma_need;
811 if (ma->ma_need & MA_SOM)
812 ma->ma_som = &info->mti_u.som.data;
814 rc = mdt_attr_get_complex(info, o, ma);
816 CERROR("%s: getattr error for "DFID": rc = %d\n",
817 mdt_obd_name(info->mti_mdt),
818 PFID(mdt_object_fid(o)), rc);
822 /* if file is released, check if a restore is running */
823 if ((ma->ma_valid & MA_HSM) && (ma->ma_hsm.mh_flags & HS_RELEASED) &&
824 mdt_hsm_restore_is_running(info, mdt_object_fid(o))) {
825 repbody->t_state = MS_RESTORE;
826 repbody->valid |= OBD_MD_TSTATE;
829 is_root = lu_fid_eq(mdt_object_fid(o), &info->mti_mdt->mdt_md_root_fid);
831 /* the Lustre protocol supposes to return default striping
832 * on the user-visible root if explicitly requested */
833 if ((ma->ma_valid & MA_LOV) == 0 && S_ISDIR(la->la_mode) &&
834 (ma->ma_need & MA_LOV_DEF && is_root) && ma->ma_need & MA_LOV) {
835 struct lu_fid rootfid;
836 struct mdt_object *root;
837 struct mdt_device *mdt = info->mti_mdt;
839 rc = dt_root_get(env, mdt->mdt_bottom, &rootfid);
842 root = mdt_object_find(env, mdt, &rootfid);
844 RETURN(PTR_ERR(root));
845 rc = mdt_stripe_get(info, root, ma, XATTR_NAME_LOV);
846 mdt_object_put(info->mti_env, root);
848 CERROR("%s: getattr error for "DFID": rc = %d\n",
849 mdt_obd_name(info->mti_mdt),
850 PFID(mdt_object_fid(o)), rc);
855 if (likely(ma->ma_valid & MA_INODE))
856 mdt_pack_attr2body(info, repbody, la, mdt_object_fid(o));
860 if (mdt_body_has_lov(la, reqbody)) {
861 if (ma->ma_valid & MA_LOV) {
862 LASSERT(ma->ma_lmm_size);
863 mdt_dump_lmm(D_INFO, ma->ma_lmm);
864 repbody->eadatasize = ma->ma_lmm_size;
865 if (S_ISDIR(la->la_mode))
866 repbody->valid |= OBD_MD_FLDIREA;
868 repbody->valid |= OBD_MD_FLEASIZE;
870 if (ma->ma_valid & MA_LMV) {
871 LASSERT(S_ISDIR(la->la_mode));
872 mdt_dump_lmv(D_INFO, ma->ma_lmv);
873 repbody->eadatasize = ma->ma_lmv_size;
874 repbody->valid |= (OBD_MD_FLDIREA|OBD_MD_MEA);
876 if (ma->ma_valid & MA_LMV_DEF) {
877 LASSERT(S_ISDIR(la->la_mode));
878 repbody->eadatasize = ma->ma_lmv_size;
879 repbody->valid |= (OBD_MD_FLDIREA|OBD_MD_DEFAULT_MEA);
881 } else if (S_ISLNK(la->la_mode) &&
882 reqbody->valid & OBD_MD_LINKNAME) {
883 buffer->lb_buf = ma->ma_lmm;
884 /* eadatasize from client includes NULL-terminator, so
885 * there is no need to read it */
886 buffer->lb_len = reqbody->eadatasize - 1;
887 rc = mo_readlink(env, next, buffer);
888 if (unlikely(rc <= 0)) {
889 CERROR("%s: readlink failed for "DFID": rc = %d\n",
890 mdt_obd_name(info->mti_mdt),
891 PFID(mdt_object_fid(o)), rc);
894 int print_limit = min_t(int, PAGE_CACHE_SIZE - 128, rc);
896 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_READLINK_EPROTO))
898 repbody->valid |= OBD_MD_LINKNAME;
899 /* we need to report back size with NULL-terminator
900 * because client expects that */
901 repbody->eadatasize = rc + 1;
902 if (repbody->eadatasize != reqbody->eadatasize)
903 CDEBUG(D_INODE, "%s: Read shorter symlink %d "
904 "on "DFID ", expected %d\n",
905 mdt_obd_name(info->mti_mdt),
906 rc, PFID(mdt_object_fid(o)),
907 reqbody->eadatasize - 1);
909 ((char *)ma->ma_lmm)[rc] = 0;
911 /* If the total CDEBUG() size is larger than a page, it
912 * will print a warning to the console, avoid this by
913 * printing just the last part of the symlink. */
914 CDEBUG(D_INODE, "symlink dest %s%.*s, len = %d\n",
915 print_limit < rc ? "..." : "", print_limit,
916 (char *)ma->ma_lmm + rc - print_limit, rc);
921 if (reqbody->valid & OBD_MD_FLMODEASIZE) {
922 repbody->max_cookiesize = 0;
923 repbody->max_mdsize = info->mti_mdt->mdt_max_mdsize;
924 repbody->valid |= OBD_MD_FLMODEASIZE;
925 CDEBUG(D_INODE, "I am going to change the MAX_MD_SIZE & "
926 "MAX_COOKIE to : %d:%d\n", repbody->max_mdsize,
927 repbody->max_cookiesize);
930 if (exp_connect_rmtclient(info->mti_exp) &&
931 reqbody->valid & OBD_MD_FLRMTPERM) {
932 void *buf = req_capsule_server_get(pill, &RMF_ACL);
934 /* mdt_getattr_lock only */
935 rc = mdt_pack_remote_perm(info, o, buf);
937 repbody->valid &= ~OBD_MD_FLRMTPERM;
938 repbody->aclsize = 0;
941 repbody->valid |= OBD_MD_FLRMTPERM;
942 repbody->aclsize = sizeof(struct mdt_remote_perm);
945 #ifdef CONFIG_FS_POSIX_ACL
946 else if ((exp_connect_flags(req->rq_export) & OBD_CONNECT_ACL) &&
947 (reqbody->valid & OBD_MD_FLACL)) {
948 buffer->lb_buf = req_capsule_server_get(pill, &RMF_ACL);
949 buffer->lb_len = req_capsule_get_size(pill,
950 &RMF_ACL, RCL_SERVER);
951 if (buffer->lb_len > 0) {
952 rc = mo_xattr_get(env, next, buffer,
953 XATTR_NAME_ACL_ACCESS);
955 if (rc == -ENODATA) {
956 repbody->aclsize = 0;
957 repbody->valid |= OBD_MD_FLACL;
959 } else if (rc == -EOPNOTSUPP) {
962 CERROR("%s: unable to read "DFID
964 mdt_obd_name(info->mti_mdt),
965 PFID(mdt_object_fid(o)), rc);
968 repbody->aclsize = rc;
969 repbody->valid |= OBD_MD_FLACL;
976 if (reqbody->valid & OBD_MD_FLMDSCAPA &&
977 info->mti_mdt->mdt_lut.lut_mds_capa &&
978 exp_connect_flags(info->mti_exp) & OBD_CONNECT_MDS_CAPA) {
979 struct lustre_capa *capa;
981 capa = req_capsule_server_get(pill, &RMF_CAPA1);
983 capa->lc_opc = CAPA_OPC_MDS_DEFAULT;
984 rc = mo_capa_get(env, next, capa, 0);
987 repbody->valid |= OBD_MD_FLMDSCAPA;
992 mdt_counter_incr(req, LPROC_MDT_GETATTR);
997 static int mdt_renew_capa(struct mdt_thread_info *info)
999 struct mdt_object *obj = info->mti_object;
1000 struct mdt_body *body;
1001 struct lustre_capa *capa, *c;
1005 /* if object doesn't exist, or server has disabled capability,
1006 * return directly, client will find body->valid OBD_MD_FLOSSCAPA
1009 if (!obj || !info->mti_mdt->mdt_lut.lut_oss_capa ||
1010 !(exp_connect_flags(info->mti_exp) & OBD_CONNECT_OSS_CAPA))
1013 body = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1014 LASSERT(body != NULL);
1016 c = req_capsule_client_get(info->mti_pill, &RMF_CAPA1);
1019 capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA2);
1023 rc = mo_capa_get(info->mti_env, mdt_object_child(obj), capa, 1);
1025 body->valid |= OBD_MD_FLOSSCAPA;
1029 int mdt_getattr(struct tgt_session_info *tsi)
1031 struct mdt_thread_info *info = tsi2mdt_info(tsi);
1032 struct mdt_object *obj = info->mti_object;
1033 struct req_capsule *pill = info->mti_pill;
1034 struct mdt_body *reqbody;
1035 struct mdt_body *repbody;
1040 reqbody = req_capsule_client_get(pill, &RMF_MDT_BODY);
1043 if (reqbody->valid & OBD_MD_FLOSSCAPA) {
1044 rc = req_capsule_server_pack(pill);
1046 RETURN(err_serious(rc));
1047 rc = mdt_renew_capa(info);
1048 GOTO(out_shrink, rc);
1051 LASSERT(obj != NULL);
1052 LASSERT(lu_object_assert_exists(&obj->mot_obj));
1054 mode = lu_object_attr(&obj->mot_obj);
1056 /* old clients may not report needed easize, use max value then */
1057 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER,
1058 reqbody->eadatasize == 0 ?
1059 info->mti_mdt->mdt_max_mdsize :
1060 reqbody->eadatasize);
1062 rc = req_capsule_server_pack(pill);
1063 if (unlikely(rc != 0))
1064 GOTO(out, rc = err_serious(rc));
1066 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
1067 LASSERT(repbody != NULL);
1068 repbody->eadatasize = 0;
1069 repbody->aclsize = 0;
1071 if (reqbody->valid & OBD_MD_FLRMTPERM)
1072 rc = mdt_init_ucred(info, reqbody);
1074 rc = mdt_check_ucred(info);
1076 GOTO(out_shrink, rc);
1078 info->mti_cross_ref = !!(reqbody->valid & OBD_MD_FLCROSSREF);
1081 * Don't check capability at all, because rename might getattr for
1082 * remote obj, and at that time no capability is available.
1084 mdt_set_capainfo(info, 1, &reqbody->fid1, BYPASS_CAPA);
1085 rc = mdt_getattr_internal(info, obj, 0);
1086 if (reqbody->valid & OBD_MD_FLRMTPERM)
1087 mdt_exit_ucred(info);
1090 mdt_client_compatibility(info);
1091 rc2 = mdt_fix_reply(info);
1095 mdt_thread_info_fini(info);
1099 int mdt_is_subdir(struct tgt_session_info *tsi)
1101 struct mdt_thread_info *info = tsi2mdt_info(tsi);
1102 struct mdt_object *o = info->mti_object;
1103 struct req_capsule *pill = info->mti_pill;
1104 const struct mdt_body *body = info->mti_body;
1105 struct mdt_body *repbody;
1111 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
1114 * We save last checked parent fid to @repbody->fid1 for remote
1117 LASSERT(fid_is_sane(&body->fid2));
1118 LASSERT(mdt_object_exists(o) && !mdt_object_remote(o));
1119 rc = mdo_is_subdir(info->mti_env, mdt_object_child(o),
1120 &body->fid2, &repbody->fid1);
1121 if (rc == 0 || rc == -EREMOTE)
1122 repbody->valid |= OBD_MD_FLID;
1124 mdt_thread_info_fini(info);
1128 int mdt_swap_layouts(struct tgt_session_info *tsi)
1130 struct mdt_thread_info *info;
1131 struct ptlrpc_request *req = tgt_ses_req(tsi);
1132 struct obd_export *exp = req->rq_export;
1133 struct mdt_object *o1, *o2, *o;
1134 struct mdt_lock_handle *lh1, *lh2;
1135 struct mdc_swap_layouts *msl;
1139 /* client does not support layout lock, so layout swaping
1141 * FIXME: there is a problem for old clients which don't support
1142 * layout lock yet. If those clients have already opened the file
1143 * they won't be notified at all so that old layout may still be
1144 * used to do IO. This can be fixed after file release is landed by
1145 * doing exclusive open and taking full EX ibits lock. - Jinshan */
1146 if (!exp_connect_layout(exp))
1147 RETURN(-EOPNOTSUPP);
1149 info = tsi2mdt_info(tsi);
1150 if (req_capsule_get_size(info->mti_pill, &RMF_CAPA1, RCL_CLIENT))
1151 mdt_set_capainfo(info, 0, &info->mti_body->fid1,
1152 req_capsule_client_get(info->mti_pill,
1155 if (req_capsule_get_size(info->mti_pill, &RMF_CAPA2, RCL_CLIENT))
1156 mdt_set_capainfo(info, 1, &info->mti_body->fid2,
1157 req_capsule_client_get(info->mti_pill,
1160 o1 = info->mti_object;
1161 o = o2 = mdt_object_find(info->mti_env, info->mti_mdt,
1162 &info->mti_body->fid2);
1164 GOTO(out, rc = PTR_ERR(o));
1166 if (mdt_object_remote(o) || !mdt_object_exists(o)) /* remote object */
1167 GOTO(put, rc = -ENOENT);
1169 rc = lu_fid_cmp(&info->mti_body->fid1, &info->mti_body->fid2);
1170 if (unlikely(rc == 0)) /* same file, you kidding me? no-op. */
1176 /* permission check. Make sure the calling process having permission
1177 * to write both files. */
1178 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o1), NULL,
1183 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o2), NULL,
1188 msl = req_capsule_client_get(info->mti_pill, &RMF_SWAP_LAYOUTS);
1190 GOTO(put, rc = -EPROTO);
1192 lh1 = &info->mti_lh[MDT_LH_NEW];
1193 mdt_lock_reg_init(lh1, LCK_EX);
1194 lh2 = &info->mti_lh[MDT_LH_OLD];
1195 mdt_lock_reg_init(lh2, LCK_EX);
1197 rc = mdt_object_lock(info, o1, lh1, MDS_INODELOCK_LAYOUT |
1198 MDS_INODELOCK_XATTR, MDT_LOCAL_LOCK);
1202 rc = mdt_object_lock(info, o2, lh2, MDS_INODELOCK_LAYOUT |
1203 MDS_INODELOCK_XATTR, MDT_LOCAL_LOCK);
1207 rc = mo_swap_layouts(info->mti_env, mdt_object_child(o1),
1208 mdt_object_child(o2), msl->msl_flags);
1211 mdt_object_unlock(info, o2, lh2, rc);
1213 mdt_object_unlock(info, o1, lh1, rc);
1215 mdt_object_put(info->mti_env, o);
1217 mdt_thread_info_fini(info);
1221 static int mdt_raw_lookup(struct mdt_thread_info *info,
1222 struct mdt_object *parent,
1223 const struct lu_name *lname,
1224 struct ldlm_reply *ldlm_rep)
1226 struct md_object *next = mdt_object_child(info->mti_object);
1227 const struct mdt_body *reqbody = info->mti_body;
1228 struct lu_fid *child_fid = &info->mti_tmp_fid1;
1229 struct mdt_body *repbody;
1233 if (reqbody->valid != OBD_MD_FLID)
1236 LASSERT(!info->mti_cross_ref);
1238 /* Only got the fid of this obj by name */
1239 fid_zero(child_fid);
1240 rc = mdo_lookup(info->mti_env, next, lname, child_fid,
1243 /* XXX is raw_lookup possible as intent operation? */
1246 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1249 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1251 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1254 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1255 repbody->fid1 = *child_fid;
1256 repbody->valid = OBD_MD_FLID;
1262 * UPDATE lock should be taken against parent, and be release before exit;
1263 * child_bits lock should be taken against child, and be returned back:
1264 * (1)normal request should release the child lock;
1265 * (2)intent request will grant the lock to client.
1267 static int mdt_getattr_name_lock(struct mdt_thread_info *info,
1268 struct mdt_lock_handle *lhc,
1270 struct ldlm_reply *ldlm_rep)
1272 struct ptlrpc_request *req = mdt_info_req(info);
1273 struct mdt_body *reqbody = NULL;
1274 struct mdt_object *parent = info->mti_object;
1275 struct mdt_object *child;
1276 struct md_object *next = mdt_object_child(parent);
1277 struct lu_fid *child_fid = &info->mti_tmp_fid1;
1278 struct lu_name *lname = NULL;
1279 struct mdt_lock_handle *lhp = NULL;
1280 struct ldlm_lock *lock;
1281 struct ldlm_res_id *res_id;
1288 is_resent = lustre_handle_is_used(&lhc->mlh_reg_lh);
1289 LASSERT(ergo(is_resent,
1290 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT));
1292 LASSERT(parent != NULL);
1294 lname = &info->mti_name;
1295 mdt_name_unpack(info->mti_pill, &RMF_NAME, lname, MNF_FIX_ANON);
1297 if (!info->mti_cross_ref) {
1299 * XXX: Check for anonymous name is for getattr by fid
1300 * (OBD_CONNECT_ATTRFID), otherwise do not allow empty name,
1301 * that is the name must contain at least one character and
1302 * the terminating '\0'.
1304 if (!lu_name_is_valid(lname)) {
1305 reqbody = req_capsule_client_get(info->mti_pill,
1307 if (unlikely(reqbody == NULL))
1308 RETURN(err_serious(-EFAULT));
1310 if (unlikely(!fid_is_sane(&reqbody->fid2)))
1311 RETURN(err_serious(-EINVAL));
1313 CDEBUG(D_INODE, "getattr with lock for "DFID"/"DFID", "
1315 PFID(mdt_object_fid(parent)),
1316 PFID(&reqbody->fid2), ldlm_rep);
1318 CDEBUG(D_INODE, "getattr with lock for "DFID"/"DNAME", "
1319 "ldlm_rep = %p\n", PFID(mdt_object_fid(parent)),
1320 PNAME(lname), ldlm_rep);
1324 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_EXECD);
1326 if (unlikely(!mdt_object_exists(parent)) && lu_name_is_valid(lname)) {
1327 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
1329 "Parent doesn't exist!\n");
1331 } else if (!info->mti_cross_ref) {
1332 LASSERTF(!mdt_object_remote(parent),
1333 "Parent "DFID" is on remote server\n",
1334 PFID(mdt_object_fid(parent)));
1337 if (lu_name_is_valid(lname)) {
1338 rc = mdt_raw_lookup(info, parent, lname, ldlm_rep);
1346 if (info->mti_cross_ref) {
1347 /* Only getattr on the child. Parent is on another node. */
1348 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1350 CDEBUG(D_INODE, "partial getattr_name child_fid = "DFID", "
1351 "ldlm_rep=%p\n", PFID(mdt_object_fid(child)), ldlm_rep);
1354 /* Do not take lock for resent case. */
1355 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
1356 LASSERTF(lock != NULL, "Invalid lock handle "LPX64"\n",
1357 lhc->mlh_reg_lh.cookie);
1358 LASSERT(fid_res_name_eq(mdt_object_fid(child),
1359 &lock->l_resource->lr_name));
1360 LDLM_LOCK_PUT(lock);
1363 mdt_lock_handle_init(lhc);
1364 mdt_lock_reg_init(lhc, LCK_PR);
1367 * Object's name is on another MDS, no lookup or layout
1368 * lock is needed here but update lock is.
1370 child_bits &= ~(MDS_INODELOCK_LOOKUP |
1371 MDS_INODELOCK_LAYOUT);
1372 child_bits |= MDS_INODELOCK_PERM | MDS_INODELOCK_UPDATE;
1374 rc = mdt_object_lock(info, child, lhc, child_bits,
1378 /* Finally, we can get attr for child. */
1379 if (!mdt_object_exists(child)) {
1380 LU_OBJECT_DEBUG(D_INFO, info->mti_env,
1382 "remote object doesn't exist.\n");
1383 mdt_object_unlock(info, child, lhc, 1);
1387 mdt_set_capainfo(info, 0, mdt_object_fid(child),
1389 rc = mdt_getattr_internal(info, child, 0);
1390 if (unlikely(rc != 0))
1391 mdt_object_unlock(info, child, lhc, 1);
1396 if (lu_name_is_valid(lname)) {
1397 /* step 1: lock parent only if parent is a directory */
1398 if (S_ISDIR(lu_object_attr(&parent->mot_obj))) {
1399 lhp = &info->mti_lh[MDT_LH_PARENT];
1400 mdt_lock_pdo_init(lhp, LCK_PR, lname);
1401 rc = mdt_object_lock(info, parent, lhp,
1402 MDS_INODELOCK_UPDATE,
1404 if (unlikely(rc != 0))
1408 /* step 2: lookup child's fid by name */
1409 fid_zero(child_fid);
1410 rc = mdo_lookup(info->mti_env, next, lname, child_fid,
1415 mdt_set_disposition(info, ldlm_rep,
1417 GOTO(out_parent, rc);
1419 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1421 *child_fid = reqbody->fid2;
1422 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1426 *step 3: find the child object by fid & lock it.
1427 * regardless if it is local or remote.
1429 *Note: LU-3240 (commit 762f2114d282a98ebfa4dbbeea9298a8088ad24e)
1430 * set parent dir fid the same as child fid in getattr by fid case
1431 * we should not lu_object_find() the object again, could lead
1432 * to hung if there is a concurrent unlink destroyed the object.
1434 if (lu_fid_eq(mdt_object_fid(parent), child_fid)) {
1435 mdt_object_get(info->mti_env, parent);
1438 child = mdt_object_find(info->mti_env, info->mti_mdt,
1442 if (unlikely(IS_ERR(child)))
1443 GOTO(out_parent, rc = PTR_ERR(child));
1445 /* Do not take lock for resent case. */
1446 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
1447 LASSERTF(lock != NULL, "Invalid lock handle "LPX64"\n",
1448 lhc->mlh_reg_lh.cookie);
1450 res_id = &lock->l_resource->lr_name;
1451 if (!fid_res_name_eq(mdt_object_fid(child),
1452 &lock->l_resource->lr_name)) {
1453 LASSERTF(fid_res_name_eq(mdt_object_fid(parent),
1454 &lock->l_resource->lr_name),
1455 "Lock res_id: "DLDLMRES", fid: "DFID"\n",
1456 PLDLMRES(lock->l_resource),
1457 PFID(mdt_object_fid(parent)));
1458 CWARN("Although resent, but still not get child lock"
1459 "parent:"DFID" child:"DFID"\n",
1460 PFID(mdt_object_fid(parent)),
1461 PFID(mdt_object_fid(child)));
1462 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
1463 LDLM_LOCK_PUT(lock);
1466 LDLM_LOCK_PUT(lock);
1469 bool try_layout = false;
1472 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RESEND, obd_timeout*2);
1473 mdt_lock_handle_init(lhc);
1474 mdt_lock_reg_init(lhc, LCK_PR);
1476 if (!mdt_object_exists(child)) {
1477 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
1479 "Object doesn't exist!\n");
1480 GOTO(out_child, rc = -ENOENT);
1483 if (!(child_bits & MDS_INODELOCK_UPDATE) &&
1484 mdt_object_exists(child) && !mdt_object_remote(child)) {
1485 struct md_attr *ma = &info->mti_attr;
1488 ma->ma_need = MA_INODE;
1489 rc = mdt_attr_get_complex(info, child, ma);
1490 if (unlikely(rc != 0))
1491 GOTO(out_child, rc);
1493 /* If the file has not been changed for some time, we
1494 * return not only a LOOKUP lock, but also an UPDATE
1495 * lock and this might save us RPC on later STAT. For
1496 * directories, it also let negative dentry cache start
1497 * working for this dir. */
1498 if (ma->ma_valid & MA_INODE &&
1499 ma->ma_attr.la_valid & LA_CTIME &&
1500 info->mti_mdt->mdt_namespace->ns_ctime_age_limit +
1501 ma->ma_attr.la_ctime < cfs_time_current_sec())
1502 child_bits |= MDS_INODELOCK_UPDATE;
1505 /* layout lock must be granted in a best-effort way
1506 * for IT operations */
1507 LASSERT(!(child_bits & MDS_INODELOCK_LAYOUT));
1508 if (!OBD_FAIL_CHECK(OBD_FAIL_MDS_NO_LL_GETATTR) &&
1509 exp_connect_layout(info->mti_exp) &&
1510 S_ISREG(lu_object_attr(&child->mot_obj)) &&
1511 !mdt_object_remote(child) && ldlm_rep != NULL) {
1512 /* try to grant layout lock for regular file. */
1518 child_bits |= MDS_INODELOCK_LAYOUT;
1519 /* try layout lock, it may fail to be granted due to
1520 * contention at LOOKUP or UPDATE */
1521 if (!mdt_object_lock_try(info, child, lhc, child_bits,
1523 child_bits &= ~MDS_INODELOCK_LAYOUT;
1524 LASSERT(child_bits != 0);
1525 rc = mdt_object_lock(info, child, lhc,
1526 child_bits, MDT_CROSS_LOCK);
1531 rc = mdt_object_lock(info, child, lhc, child_bits,
1534 if (unlikely(rc != 0))
1535 GOTO(out_child, rc);
1538 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
1539 /* Get MA_SOM attributes if update lock is given. */
1541 lock->l_policy_data.l_inodebits.bits & MDS_INODELOCK_UPDATE &&
1542 S_ISREG(lu_object_attr(&mdt_object_child(child)->mo_lu)))
1545 /* finally, we can get attr for child. */
1546 mdt_set_capainfo(info, 1, child_fid, BYPASS_CAPA);
1547 rc = mdt_getattr_internal(info, child, ma_need);
1548 if (unlikely(rc != 0)) {
1549 mdt_object_unlock(info, child, lhc, 1);
1551 /* Debugging code. */
1552 res_id = &lock->l_resource->lr_name;
1553 LDLM_DEBUG(lock, "Returning lock to client");
1554 LASSERTF(fid_res_name_eq(mdt_object_fid(child),
1555 &lock->l_resource->lr_name),
1556 "Lock res_id: "DLDLMRES", fid: "DFID"\n",
1557 PLDLMRES(lock->l_resource),
1558 PFID(mdt_object_fid(child)));
1559 if (mdt_object_exists(child) && !mdt_object_remote(child))
1560 mdt_pack_size2body(info, child);
1563 LDLM_LOCK_PUT(lock);
1567 mdt_object_put(info->mti_env, child);
1570 mdt_object_unlock(info, parent, lhp, 1);
1574 /* normal handler: should release the child lock */
1575 int mdt_getattr_name(struct tgt_session_info *tsi)
1577 struct mdt_thread_info *info = tsi2mdt_info(tsi);
1578 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_CHILD];
1579 struct mdt_body *reqbody;
1580 struct mdt_body *repbody;
1584 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
1585 LASSERT(reqbody != NULL);
1586 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1587 LASSERT(repbody != NULL);
1589 info->mti_cross_ref = !!(reqbody->valid & OBD_MD_FLCROSSREF);
1590 repbody->eadatasize = 0;
1591 repbody->aclsize = 0;
1593 rc = mdt_init_ucred(info, reqbody);
1595 GOTO(out_shrink, rc);
1597 rc = mdt_getattr_name_lock(info, lhc, MDS_INODELOCK_UPDATE, NULL);
1598 if (lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1599 ldlm_lock_decref(&lhc->mlh_reg_lh, lhc->mlh_reg_mode);
1600 lhc->mlh_reg_lh.cookie = 0;
1602 mdt_exit_ucred(info);
1605 mdt_client_compatibility(info);
1606 rc2 = mdt_fix_reply(info);
1609 mdt_thread_info_fini(info);
1613 static int mdt_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
1614 void *karg, void *uarg);
1616 int mdt_set_info(struct tgt_session_info *tsi)
1618 struct ptlrpc_request *req = tgt_ses_req(tsi);
1621 int keylen, vallen, rc = 0;
1625 key = req_capsule_client_get(tsi->tsi_pill, &RMF_SETINFO_KEY);
1627 DEBUG_REQ(D_HA, req, "no set_info key");
1628 RETURN(err_serious(-EFAULT));
1631 keylen = req_capsule_get_size(tsi->tsi_pill, &RMF_SETINFO_KEY,
1634 val = req_capsule_client_get(tsi->tsi_pill, &RMF_SETINFO_VAL);
1636 DEBUG_REQ(D_HA, req, "no set_info val");
1637 RETURN(err_serious(-EFAULT));
1640 vallen = req_capsule_get_size(tsi->tsi_pill, &RMF_SETINFO_VAL,
1643 /* Swab any part of val you need to here */
1644 if (KEY_IS(KEY_READ_ONLY)) {
1645 spin_lock(&req->rq_export->exp_lock);
1647 *exp_connect_flags_ptr(req->rq_export) |=
1650 *exp_connect_flags_ptr(req->rq_export) &=
1651 ~OBD_CONNECT_RDONLY;
1652 spin_unlock(&req->rq_export->exp_lock);
1653 } else if (KEY_IS(KEY_CHANGELOG_CLEAR)) {
1654 struct changelog_setinfo *cs = val;
1656 if (vallen != sizeof(*cs)) {
1657 CERROR("%s: bad changelog_clear setinfo size %d\n",
1658 tgt_name(tsi->tsi_tgt), vallen);
1661 if (ptlrpc_req_need_swab(req)) {
1662 __swab64s(&cs->cs_recno);
1663 __swab32s(&cs->cs_id);
1666 rc = mdt_iocontrol(OBD_IOC_CHANGELOG_CLEAR, req->rq_export,
1674 int mdt_readpage(struct tgt_session_info *tsi)
1676 struct mdt_thread_info *info = mdt_th_info(tsi->tsi_env);
1677 struct mdt_object *object = mdt_obj(tsi->tsi_corpus);
1678 struct lu_rdpg *rdpg = &info->mti_u.rdpg.mti_rdpg;
1679 const struct mdt_body *reqbody = tsi->tsi_mdt_body;
1680 struct mdt_body *repbody;
1686 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK))
1687 RETURN(err_serious(-ENOMEM));
1689 repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_MDT_BODY);
1690 if (repbody == NULL || reqbody == NULL)
1691 RETURN(err_serious(-EFAULT));
1694 * prepare @rdpg before calling lower layers and transfer itself. Here
1695 * reqbody->size contains offset of where to start to read and
1696 * reqbody->nlink contains number bytes to read.
1698 rdpg->rp_hash = reqbody->size;
1699 if (rdpg->rp_hash != reqbody->size) {
1700 CERROR("Invalid hash: "LPX64" != "LPX64"\n",
1701 rdpg->rp_hash, reqbody->size);
1705 rdpg->rp_attrs = reqbody->mode;
1706 if (exp_connect_flags(tsi->tsi_exp) & OBD_CONNECT_64BITHASH)
1707 rdpg->rp_attrs |= LUDA_64BITHASH;
1708 rdpg->rp_count = min_t(unsigned int, reqbody->nlink,
1709 exp_max_brw_size(tsi->tsi_exp));
1710 rdpg->rp_npages = (rdpg->rp_count + PAGE_CACHE_SIZE - 1) >>
1712 OBD_ALLOC(rdpg->rp_pages, rdpg->rp_npages * sizeof rdpg->rp_pages[0]);
1713 if (rdpg->rp_pages == NULL)
1716 for (i = 0; i < rdpg->rp_npages; ++i) {
1717 rdpg->rp_pages[i] = alloc_page(GFP_IOFS);
1718 if (rdpg->rp_pages[i] == NULL)
1719 GOTO(free_rdpg, rc = -ENOMEM);
1722 /* call lower layers to fill allocated pages with directory data */
1723 rc = mo_readpage(tsi->tsi_env, mdt_object_child(object), rdpg);
1725 GOTO(free_rdpg, rc);
1727 /* send pages to client */
1728 rc = tgt_sendpage(tsi, rdpg, rc);
1733 for (i = 0; i < rdpg->rp_npages; i++)
1734 if (rdpg->rp_pages[i] != NULL)
1735 __free_page(rdpg->rp_pages[i]);
1736 OBD_FREE(rdpg->rp_pages, rdpg->rp_npages * sizeof rdpg->rp_pages[0]);
1738 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE))
1744 static int mdt_reint_internal(struct mdt_thread_info *info,
1745 struct mdt_lock_handle *lhc,
1748 struct req_capsule *pill = info->mti_pill;
1749 struct mdt_body *repbody;
1754 rc = mdt_reint_unpack(info, op);
1756 CERROR("Can't unpack reint, rc %d\n", rc);
1757 RETURN(err_serious(rc));
1760 /* for replay (no_create) lmm is not needed, client has it already */
1761 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
1762 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER,
1763 info->mti_rr.rr_eadatalen);
1765 /* llog cookies are always 0, the field is kept for compatibility */
1766 if (req_capsule_has_field(pill, &RMF_LOGCOOKIES, RCL_SERVER))
1767 req_capsule_set_size(pill, &RMF_LOGCOOKIES, RCL_SERVER, 0);
1769 rc = req_capsule_server_pack(pill);
1771 CERROR("Can't pack response, rc %d\n", rc);
1772 RETURN(err_serious(rc));
1775 if (req_capsule_has_field(pill, &RMF_MDT_BODY, RCL_SERVER)) {
1776 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
1778 repbody->eadatasize = 0;
1779 repbody->aclsize = 0;
1782 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_REINT_DELAY, 10);
1784 /* for replay no cookkie / lmm need, because client have this already */
1785 if (info->mti_spec.no_create)
1786 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
1787 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER, 0);
1789 rc = mdt_init_ucred_reint(info);
1791 GOTO(out_shrink, rc);
1793 rc = mdt_fix_attr_ucred(info, op);
1795 GOTO(out_ucred, rc = err_serious(rc));
1797 if (mdt_check_resent(info, mdt_reconstruct, lhc)) {
1798 DEBUG_REQ(D_INODE, mdt_info_req(info), "resent opt.");
1799 rc = lustre_msg_get_status(mdt_info_req(info)->rq_repmsg);
1800 GOTO(out_ucred, rc);
1802 rc = mdt_reint_rec(info, lhc);
1805 mdt_exit_ucred(info);
1807 mdt_client_compatibility(info);
1808 rc2 = mdt_fix_reply(info);
1814 static long mdt_reint_opcode(struct ptlrpc_request *req,
1815 const struct req_format **fmt)
1817 struct mdt_device *mdt;
1818 struct mdt_rec_reint *rec;
1821 rec = req_capsule_client_get(&req->rq_pill, &RMF_REC_REINT);
1823 opc = rec->rr_opcode;
1824 DEBUG_REQ(D_INODE, req, "reint opt = %ld", opc);
1825 if (opc < REINT_MAX && fmt[opc] != NULL)
1826 req_capsule_extend(&req->rq_pill, fmt[opc]);
1828 mdt = mdt_exp2dev(req->rq_export);
1829 CERROR("%s: Unsupported opcode '%ld' from client '%s':"
1830 " rc = %d\n", req->rq_export->exp_obd->obd_name,
1831 opc, mdt->mdt_ldlm_client->cli_name, -EFAULT);
1832 opc = err_serious(-EFAULT);
1835 opc = err_serious(-EFAULT);
1840 int mdt_reint(struct tgt_session_info *tsi)
1844 static const struct req_format *reint_fmts[REINT_MAX] = {
1845 [REINT_SETATTR] = &RQF_MDS_REINT_SETATTR,
1846 [REINT_CREATE] = &RQF_MDS_REINT_CREATE,
1847 [REINT_LINK] = &RQF_MDS_REINT_LINK,
1848 [REINT_UNLINK] = &RQF_MDS_REINT_UNLINK,
1849 [REINT_RENAME] = &RQF_MDS_REINT_RENAME,
1850 [REINT_OPEN] = &RQF_MDS_REINT_OPEN,
1851 [REINT_SETXATTR] = &RQF_MDS_REINT_SETXATTR,
1852 [REINT_RMENTRY] = &RQF_MDS_REINT_UNLINK,
1853 [REINT_MIGRATE] = &RQF_MDS_REINT_RENAME
1858 opc = mdt_reint_opcode(tgt_ses_req(tsi), reint_fmts);
1860 struct mdt_thread_info *info = tsi2mdt_info(tsi);
1862 * No lock possible here from client to pass it to reint code
1865 rc = mdt_reint_internal(info, NULL, opc);
1866 mdt_thread_info_fini(info);
1871 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_REINT_NET_REP;
1875 /* this should sync the whole device */
1876 static int mdt_device_sync(const struct lu_env *env, struct mdt_device *mdt)
1878 struct dt_device *dt = mdt->mdt_bottom;
1882 rc = dt->dd_ops->dt_sync(env, dt);
1886 /* this should sync this object */
1887 static int mdt_object_sync(struct mdt_thread_info *info)
1889 struct md_object *next;
1893 if (!mdt_object_exists(info->mti_object)) {
1894 CWARN("Non existing object "DFID"!\n",
1895 PFID(mdt_object_fid(info->mti_object)));
1898 next = mdt_object_child(info->mti_object);
1899 rc = mo_object_sync(info->mti_env, next);
1904 int mdt_sync(struct tgt_session_info *tsi)
1906 struct ptlrpc_request *req = tgt_ses_req(tsi);
1907 struct req_capsule *pill = tsi->tsi_pill;
1908 struct mdt_body *body;
1913 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SYNC_PACK))
1914 RETURN(err_serious(-ENOMEM));
1916 if (fid_seq(&tsi->tsi_mdt_body->fid1) == 0) {
1917 rc = mdt_device_sync(tsi->tsi_env, mdt_exp2dev(tsi->tsi_exp));
1919 struct mdt_thread_info *info = tsi2mdt_info(tsi);
1921 /* sync an object */
1922 rc = mdt_object_sync(info);
1924 const struct lu_fid *fid;
1925 struct lu_attr *la = &info->mti_attr.ma_attr;
1927 info->mti_attr.ma_need = MA_INODE;
1928 info->mti_attr.ma_valid = 0;
1929 rc = mdt_attr_get_complex(info, info->mti_object,
1932 body = req_capsule_server_get(pill,
1934 fid = mdt_object_fid(info->mti_object);
1935 mdt_pack_attr2body(info, body, la, fid);
1938 mdt_thread_info_fini(info);
1941 mdt_counter_incr(req, LPROC_MDT_SYNC);
1947 * Handle quota control requests to consult current usage/limit, but also
1948 * to configure quota enforcement
1950 int mdt_quotactl(struct tgt_session_info *tsi)
1952 struct obd_export *exp = tsi->tsi_exp;
1953 struct req_capsule *pill = tsi->tsi_pill;
1954 struct obd_quotactl *oqctl, *repoqc;
1956 struct mdt_device *mdt = mdt_exp2dev(exp);
1957 struct lu_device *qmt = mdt->mdt_qmt_dev;
1960 oqctl = req_capsule_client_get(pill, &RMF_OBD_QUOTACTL);
1962 RETURN(err_serious(-EPROTO));
1964 rc = req_capsule_server_pack(pill);
1966 RETURN(err_serious(rc));
1968 switch (oqctl->qc_cmd) {
1970 case LUSTRE_Q_INVALIDATE:
1971 case LUSTRE_Q_FINVALIDATE:
1975 /* deprecated, not used any more */
1976 RETURN(-EOPNOTSUPP);
1977 /* master quotactl */
1983 RETURN(-EOPNOTSUPP);
1984 /* slave quotactl */
1989 CERROR("Unsupported quotactl command: %d\n", oqctl->qc_cmd);
1993 /* map uid/gid for remote client */
1995 if (exp_connect_rmtclient(exp)) {
1996 struct lustre_idmap_table *idmap;
1998 idmap = exp->exp_mdt_data.med_idmap;
2000 if (unlikely(oqctl->qc_cmd != Q_GETQUOTA &&
2001 oqctl->qc_cmd != Q_GETINFO))
2004 if (oqctl->qc_type == USRQUOTA)
2005 id = lustre_idmap_lookup_uid(NULL, idmap, 0,
2007 else if (oqctl->qc_type == GRPQUOTA)
2008 id = lustre_idmap_lookup_gid(NULL, idmap, 0,
2013 if (id == CFS_IDMAP_NOTFOUND) {
2014 CDEBUG(D_QUOTA, "no mapping for id %u\n", oqctl->qc_id);
2019 repoqc = req_capsule_server_get(pill, &RMF_OBD_QUOTACTL);
2021 RETURN(err_serious(-EFAULT));
2023 if (oqctl->qc_id != id)
2024 swap(oqctl->qc_id, id);
2026 switch (oqctl->qc_cmd) {
2032 /* forward quotactl request to QMT */
2033 rc = qmt_hdls.qmth_quotactl(tsi->tsi_env, qmt, oqctl);
2038 /* slave quotactl */
2039 rc = lquotactl_slv(tsi->tsi_env, tsi->tsi_tgt->lut_bottom,
2044 CERROR("Unsupported quotactl command: %d\n", oqctl->qc_cmd);
2048 if (oqctl->qc_id != id)
2049 swap(oqctl->qc_id, id);
2055 /** clone llog ctxt from child (mdd)
2056 * This allows remote llog (replicator) access.
2057 * We can either pass all llog RPCs (eg mdt_llog_create) on to child where the
2058 * context was originally set up, or we can handle them directly.
2059 * I choose the latter, but that means I need any llog
2060 * contexts set up by child to be accessable by the mdt. So we clone the
2061 * context into our context list here.
2063 static int mdt_llog_ctxt_clone(const struct lu_env *env, struct mdt_device *mdt,
2066 struct md_device *next = mdt->mdt_child;
2067 struct llog_ctxt *ctxt;
2070 if (!llog_ctxt_null(mdt2obd_dev(mdt), idx))
2073 rc = next->md_ops->mdo_llog_ctxt_get(env, next, idx, (void **)&ctxt);
2074 if (rc || ctxt == NULL) {
2078 rc = llog_group_set_ctxt(&mdt2obd_dev(mdt)->obd_olg, ctxt, idx);
2080 CERROR("Can't set mdt ctxt %d\n", rc);
2085 static int mdt_llog_ctxt_unclone(const struct lu_env *env,
2086 struct mdt_device *mdt, int idx)
2088 struct llog_ctxt *ctxt;
2090 ctxt = llog_get_context(mdt2obd_dev(mdt), idx);
2093 /* Put once for the get we just did, and once for the clone */
2094 llog_ctxt_put(ctxt);
2095 llog_ctxt_put(ctxt);
2100 * sec context handlers
2102 int mdt_sec_ctx_handle(struct tgt_session_info *tsi)
2106 rc = mdt_handle_idmap(tsi);
2108 struct ptlrpc_request *req = tgt_ses_req(tsi);
2111 opc = lustre_msg_get_opc(req->rq_reqmsg);
2112 if (opc == SEC_CTX_INIT || opc == SEC_CTX_INIT_CONT)
2113 sptlrpc_svc_ctx_invalidate(req);
2116 CFS_FAIL_TIMEOUT(OBD_FAIL_SEC_CTX_HDL_PAUSE, cfs_fail_val);
2122 * quota request handlers
2124 int mdt_quota_dqacq(struct tgt_session_info *tsi)
2126 struct mdt_device *mdt = mdt_exp2dev(tsi->tsi_exp);
2127 struct lu_device *qmt = mdt->mdt_qmt_dev;
2132 RETURN(err_serious(-EOPNOTSUPP));
2134 rc = qmt_hdls.qmth_dqacq(tsi->tsi_env, qmt, tgt_ses_req(tsi));
2138 struct mdt_object *mdt_object_new(const struct lu_env *env,
2139 struct mdt_device *d,
2140 const struct lu_fid *f)
2142 struct lu_object_conf conf = { .loc_flags = LOC_F_NEW };
2143 struct lu_object *o;
2144 struct mdt_object *m;
2147 CDEBUG(D_INFO, "Allocate object for "DFID"\n", PFID(f));
2148 o = lu_object_find(env, &d->mdt_lu_dev, f, &conf);
2149 if (unlikely(IS_ERR(o)))
2150 m = (struct mdt_object *)o;
2156 struct mdt_object *mdt_object_find(const struct lu_env *env,
2157 struct mdt_device *d,
2158 const struct lu_fid *f)
2160 struct lu_object *o;
2161 struct mdt_object *m;
2164 CDEBUG(D_INFO, "Find object for "DFID"\n", PFID(f));
2165 o = lu_object_find(env, &d->mdt_lu_dev, f, NULL);
2166 if (unlikely(IS_ERR(o)))
2167 m = (struct mdt_object *)o;
2175 * Asyncronous commit for mdt device.
2177 * Pass asynchonous commit call down the MDS stack.
2179 * \param env environment
2180 * \param mdt the mdt device
2182 static void mdt_device_commit_async(const struct lu_env *env,
2183 struct mdt_device *mdt)
2185 struct dt_device *dt = mdt->mdt_bottom;
2188 rc = dt->dd_ops->dt_commit_async(env, dt);
2189 if (unlikely(rc != 0))
2190 CWARN("async commit start failed with rc = %d", rc);
2194 * Mark the lock as "synchonous".
2196 * Mark the lock to deffer transaction commit to the unlock time.
2198 * \param lock the lock to mark as "synchonous"
2200 * \see mdt_is_lock_sync
2201 * \see mdt_save_lock
2203 static inline void mdt_set_lock_sync(struct ldlm_lock *lock)
2205 lock->l_ast_data = (void*)1;
2209 * Check whehter the lock "synchonous" or not.
2211 * \param lock the lock to check
2212 * \retval 1 the lock is "synchonous"
2213 * \retval 0 the lock isn't "synchronous"
2215 * \see mdt_set_lock_sync
2216 * \see mdt_save_lock
2218 static inline int mdt_is_lock_sync(struct ldlm_lock *lock)
2220 return lock->l_ast_data != NULL;
2224 * Blocking AST for mdt locks.
2226 * Starts transaction commit if in case of COS lock conflict or
2227 * deffers such a commit to the mdt_save_lock.
2229 * \param lock the lock which blocks a request or cancelling lock
2230 * \param desc unused
2231 * \param data unused
2232 * \param flag indicates whether this cancelling or blocking callback
2234 * \see ldlm_blocking_ast_nocheck
2236 int mdt_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
2237 void *data, int flag)
2239 struct obd_device *obd = ldlm_lock_to_ns(lock)->ns_obd;
2240 struct mdt_device *mdt = mdt_dev(obd->obd_lu_dev);
2244 if (flag == LDLM_CB_CANCELING)
2246 lock_res_and_lock(lock);
2247 if (lock->l_blocking_ast != mdt_blocking_ast) {
2248 unlock_res_and_lock(lock);
2251 if (mdt_cos_is_enabled(mdt) &&
2252 lock->l_req_mode & (LCK_PW | LCK_EX) &&
2253 lock->l_blocking_lock != NULL &&
2254 lock->l_client_cookie != lock->l_blocking_lock->l_client_cookie) {
2255 mdt_set_lock_sync(lock);
2257 rc = ldlm_blocking_ast_nocheck(lock);
2259 /* There is no lock conflict if l_blocking_lock == NULL,
2260 * it indicates a blocking ast sent from ldlm_lock_decref_internal
2261 * when the last reference to a local lock was released */
2262 if (lock->l_req_mode == LCK_COS && lock->l_blocking_lock != NULL) {
2265 rc = lu_env_init(&env, LCT_LOCAL);
2266 if (unlikely(rc != 0))
2267 CWARN("lu_env initialization failed with rc = %d,"
2268 "cannot start asynchronous commit\n", rc);
2270 mdt_device_commit_async(&env, mdt);
2276 /* Used for cross-MDT lock */
2277 int mdt_remote_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
2278 void *data, int flag)
2280 struct lustre_handle lockh;
2284 case LDLM_CB_BLOCKING:
2285 ldlm_lock2handle(lock, &lockh);
2286 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
2288 CDEBUG(D_INODE, "ldlm_cli_cancel: %d\n", rc);
2292 case LDLM_CB_CANCELING:
2293 LDLM_DEBUG(lock, "Revoke remote lock\n");
2301 int mdt_remote_object_lock(struct mdt_thread_info *mti,
2302 struct mdt_object *o, struct lustre_handle *lh,
2303 ldlm_mode_t mode, __u64 ibits)
2305 struct ldlm_enqueue_info *einfo = &mti->mti_einfo;
2306 ldlm_policy_data_t *policy = &mti->mti_policy;
2310 LASSERT(mdt_object_remote(o));
2312 LASSERT(ibits & MDS_INODELOCK_UPDATE);
2314 memset(einfo, 0, sizeof(*einfo));
2315 einfo->ei_type = LDLM_IBITS;
2316 einfo->ei_mode = mode;
2317 einfo->ei_cb_bl = mdt_remote_blocking_ast;
2318 einfo->ei_cb_cp = ldlm_completion_ast;
2319 einfo->ei_enq_slave = 0;
2321 memset(policy, 0, sizeof(*policy));
2322 policy->l_inodebits.bits = ibits;
2324 rc = mo_object_lock(mti->mti_env, mdt_object_child(o), lh, einfo,
2329 static int mdt_object_lock0(struct mdt_thread_info *info, struct mdt_object *o,
2330 struct mdt_lock_handle *lh, __u64 ibits,
2331 bool nonblock, int locality)
2333 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
2334 ldlm_policy_data_t *policy = &info->mti_policy;
2335 struct ldlm_res_id *res_id = &info->mti_res_id;
2340 LASSERT(!lustre_handle_is_used(&lh->mlh_reg_lh));
2341 LASSERT(!lustre_handle_is_used(&lh->mlh_pdo_lh));
2342 LASSERT(lh->mlh_reg_mode != LCK_MINMODE);
2343 LASSERT(lh->mlh_type != MDT_NUL_LOCK);
2345 if (mdt_object_remote(o)) {
2346 if (locality == MDT_CROSS_LOCK) {
2347 ibits &= ~(MDS_INODELOCK_UPDATE | MDS_INODELOCK_PERM |
2348 MDS_INODELOCK_LAYOUT);
2349 ibits |= MDS_INODELOCK_LOOKUP;
2352 (MDS_INODELOCK_UPDATE | MDS_INODELOCK_PERM |
2353 MDS_INODELOCK_LAYOUT)),
2354 "%s: wrong bit "LPX64" for remote obj "DFID"\n",
2355 mdt_obd_name(info->mti_mdt), ibits,
2356 PFID(mdt_object_fid(o)));
2357 LASSERT(ibits & MDS_INODELOCK_LOOKUP);
2359 /* No PDO lock on remote object */
2360 LASSERT(lh->mlh_type != MDT_PDO_LOCK);
2363 if (lh->mlh_type == MDT_PDO_LOCK) {
2364 /* check for exists after object is locked */
2365 if (mdt_object_exists(o) == 0) {
2366 /* Non-existent object shouldn't have PDO lock */
2369 /* Non-dir object shouldn't have PDO lock */
2370 if (!S_ISDIR(lu_object_attr(&o->mot_obj)))
2375 memset(policy, 0, sizeof(*policy));
2376 fid_build_reg_res_name(mdt_object_fid(o), res_id);
2378 dlmflags = LDLM_FL_ATOMIC_CB;
2380 dlmflags |= LDLM_FL_BLOCK_NOWAIT;
2383 * Take PDO lock on whole directory and build correct @res_id for lock
2384 * on part of directory.
2386 if (lh->mlh_pdo_hash != 0) {
2387 LASSERT(lh->mlh_type == MDT_PDO_LOCK);
2388 mdt_lock_pdo_mode(info, o, lh);
2389 if (lh->mlh_pdo_mode != LCK_NL) {
2391 * Do not use LDLM_FL_LOCAL_ONLY for parallel lock, it
2392 * is never going to be sent to client and we do not
2393 * want it slowed down due to possible cancels.
2395 policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
2396 rc = mdt_fid_lock(ns, &lh->mlh_pdo_lh, lh->mlh_pdo_mode,
2397 policy, res_id, dlmflags,
2398 &info->mti_exp->exp_handle.h_cookie);
2404 * Finish res_id initializing by name hash marking part of
2405 * directory which is taking modification.
2407 res_id->name[LUSTRE_RES_ID_HSH_OFF] = lh->mlh_pdo_hash;
2410 policy->l_inodebits.bits = ibits;
2413 * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
2414 * going to be sent to client. If it is - mdt_intent_policy() path will
2415 * fix it up and turn FL_LOCAL flag off.
2417 rc = mdt_fid_lock(ns, &lh->mlh_reg_lh, lh->mlh_reg_mode, policy,
2418 res_id, LDLM_FL_LOCAL_ONLY | dlmflags,
2419 &info->mti_exp->exp_handle.h_cookie);
2421 mdt_object_unlock(info, o, lh, 1);
2422 else if (unlikely(OBD_FAIL_PRECHECK(OBD_FAIL_MDS_PDO_LOCK)) &&
2423 lh->mlh_pdo_hash != 0 &&
2424 (lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX)) {
2425 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK, 15);
2431 int mdt_object_lock(struct mdt_thread_info *info, struct mdt_object *o,
2432 struct mdt_lock_handle *lh, __u64 ibits, int locality)
2434 return mdt_object_lock0(info, o, lh, ibits, false, locality);
2437 int mdt_object_lock_try(struct mdt_thread_info *info, struct mdt_object *o,
2438 struct mdt_lock_handle *lh, __u64 ibits, int locality)
2440 struct mdt_lock_handle tmp = *lh;
2443 rc = mdt_object_lock0(info, o, &tmp, ibits, true, locality);
2451 * Save a lock within request object.
2453 * Keep the lock referenced until whether client ACK or transaction
2454 * commit happens or release the lock immediately depending on input
2455 * parameters. If COS is ON, a write lock is converted to COS lock
2458 * \param info thead info object
2459 * \param h lock handle
2460 * \param mode lock mode
2461 * \param decref force immediate lock releasing
2464 void mdt_save_lock(struct mdt_thread_info *info, struct lustre_handle *h,
2465 ldlm_mode_t mode, int decref)
2469 if (lustre_handle_is_used(h)) {
2470 if (decref || !info->mti_has_trans ||
2471 !(mode & (LCK_PW | LCK_EX))){
2472 mdt_fid_unlock(h, mode);
2474 struct mdt_device *mdt = info->mti_mdt;
2475 struct ldlm_lock *lock = ldlm_handle2lock(h);
2476 struct ptlrpc_request *req = mdt_info_req(info);
2479 LASSERTF(lock != NULL, "no lock for cookie "LPX64"\n",
2481 /* there is no request if mdt_object_unlock() is called
2482 * from mdt_export_cleanup()->mdt_add_dirty_flag() */
2483 if (likely(req != NULL)) {
2484 CDEBUG(D_HA, "request = %p reply state = %p"
2485 " transno = "LPD64"\n", req,
2486 req->rq_reply_state, req->rq_transno);
2487 if (mdt_cos_is_enabled(mdt)) {
2489 ldlm_lock_downgrade(lock, LCK_COS);
2492 ptlrpc_save_lock(req, h, mode, no_ack);
2494 ldlm_lock_decref(h, mode);
2496 if (mdt_is_lock_sync(lock)) {
2497 CDEBUG(D_HA, "found sync-lock,"
2498 " async commit started\n");
2499 mdt_device_commit_async(info->mti_env,
2502 LDLM_LOCK_PUT(lock);
2511 * Unlock mdt object.
2513 * Immeditely release the regular lock and the PDO lock or save the
2514 * lock in reqeuest and keep them referenced until client ACK or
2515 * transaction commit.
2517 * \param info thread info object
2518 * \param o mdt object
2519 * \param lh mdt lock handle referencing regular and PDO locks
2520 * \param decref force immediate lock releasing
2522 void mdt_object_unlock(struct mdt_thread_info *info, struct mdt_object *o,
2523 struct mdt_lock_handle *lh, int decref)
2527 mdt_save_lock(info, &lh->mlh_pdo_lh, lh->mlh_pdo_mode, decref);
2528 mdt_save_lock(info, &lh->mlh_reg_lh, lh->mlh_reg_mode, decref);
2530 if (lustre_handle_is_used(&lh->mlh_rreg_lh))
2531 ldlm_lock_decref(&lh->mlh_rreg_lh, lh->mlh_rreg_mode);
2536 struct mdt_object *mdt_object_find_lock(struct mdt_thread_info *info,
2537 const struct lu_fid *f,
2538 struct mdt_lock_handle *lh,
2541 struct mdt_object *o;
2543 o = mdt_object_find(info->mti_env, info->mti_mdt, f);
2547 rc = mdt_object_lock(info, o, lh, ibits,
2550 mdt_object_put(info->mti_env, o);
2557 void mdt_object_unlock_put(struct mdt_thread_info * info,
2558 struct mdt_object * o,
2559 struct mdt_lock_handle *lh,
2562 mdt_object_unlock(info, o, lh, decref);
2563 mdt_object_put(info->mti_env, o);
2567 * Generic code handling requests that have struct mdt_body passed in:
2569 * - extract mdt_body from request and save it in @info, if present;
2571 * - create lu_object, corresponding to the fid in mdt_body, and save it in
2574 * - if HABEO_CORPUS flag is set for this request type check whether object
2575 * actually exists on storage (lu_object_exists()).
2578 static int mdt_body_unpack(struct mdt_thread_info *info, __u32 flags)
2580 const struct mdt_body *body;
2581 struct mdt_object *obj;
2582 const struct lu_env *env;
2583 struct req_capsule *pill;
2587 env = info->mti_env;
2588 pill = info->mti_pill;
2590 body = info->mti_body = req_capsule_client_get(pill, &RMF_MDT_BODY);
2594 if (!(body->valid & OBD_MD_FLID))
2597 if (!fid_is_sane(&body->fid1)) {
2598 CERROR("Invalid fid: "DFID"\n", PFID(&body->fid1));
2603 * Do not get size or any capa fields before we check that request
2604 * contains capa actually. There are some requests which do not, for
2605 * instance MDS_IS_SUBDIR.
2607 if (req_capsule_has_field(pill, &RMF_CAPA1, RCL_CLIENT) &&
2608 req_capsule_get_size(pill, &RMF_CAPA1, RCL_CLIENT))
2609 mdt_set_capainfo(info, 0, &body->fid1,
2610 req_capsule_client_get(pill, &RMF_CAPA1));
2612 obj = mdt_object_find(env, info->mti_mdt, &body->fid1);
2614 if ((flags & HABEO_CORPUS) &&
2615 !mdt_object_exists(obj)) {
2616 mdt_object_put(env, obj);
2617 /* for capability renew ENOENT will be handled in
2619 if (body->valid & OBD_MD_FLOSSCAPA)
2624 info->mti_object = obj;
2633 static int mdt_unpack_req_pack_rep(struct mdt_thread_info *info, __u32 flags)
2635 struct req_capsule *pill = info->mti_pill;
2639 if (req_capsule_has_field(pill, &RMF_MDT_BODY, RCL_CLIENT))
2640 rc = mdt_body_unpack(info, flags);
2644 if (rc == 0 && (flags & HABEO_REFERO)) {
2646 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
2647 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER,
2648 info->mti_body->eadatasize);
2649 if (req_capsule_has_field(pill, &RMF_LOGCOOKIES, RCL_SERVER))
2650 req_capsule_set_size(pill, &RMF_LOGCOOKIES,
2653 rc = req_capsule_server_pack(pill);
2658 static int mdt_init_capa_ctxt(const struct lu_env *env, struct mdt_device *m)
2660 struct md_device *next = m->mdt_child;
2662 return next->md_ops->mdo_init_capa_ctxt(env, next,
2663 m->mdt_lut.lut_mds_capa,
2664 m->mdt_capa_timeout,
2669 void mdt_lock_handle_init(struct mdt_lock_handle *lh)
2671 lh->mlh_type = MDT_NUL_LOCK;
2672 lh->mlh_reg_lh.cookie = 0ull;
2673 lh->mlh_reg_mode = LCK_MINMODE;
2674 lh->mlh_pdo_lh.cookie = 0ull;
2675 lh->mlh_pdo_mode = LCK_MINMODE;
2676 lh->mlh_rreg_lh.cookie = 0ull;
2677 lh->mlh_rreg_mode = LCK_MINMODE;
2680 void mdt_lock_handle_fini(struct mdt_lock_handle *lh)
2682 LASSERT(!lustre_handle_is_used(&lh->mlh_reg_lh));
2683 LASSERT(!lustre_handle_is_used(&lh->mlh_pdo_lh));
2687 * Initialize fields of struct mdt_thread_info. Other fields are left in
2688 * uninitialized state, because it's too expensive to zero out whole
2689 * mdt_thread_info (> 1K) on each request arrival.
2691 void mdt_thread_info_init(struct ptlrpc_request *req,
2692 struct mdt_thread_info *info)
2696 info->mti_pill = &req->rq_pill;
2699 for (i = 0; i < ARRAY_SIZE(info->mti_lh); i++)
2700 mdt_lock_handle_init(&info->mti_lh[i]);
2702 /* mdt device: it can be NULL while CONNECT */
2703 if (req->rq_export) {
2704 info->mti_mdt = mdt_dev(req->rq_export->exp_obd->obd_lu_dev);
2705 info->mti_exp = req->rq_export;
2707 info->mti_mdt = NULL;
2708 info->mti_env = req->rq_svc_thread->t_env;
2709 info->mti_transno = lustre_msg_get_transno(req->rq_reqmsg);
2711 memset(&info->mti_attr, 0, sizeof(info->mti_attr));
2712 info->mti_big_buf = LU_BUF_NULL;
2713 info->mti_body = NULL;
2714 info->mti_object = NULL;
2715 info->mti_dlm_req = NULL;
2716 info->mti_has_trans = 0;
2717 info->mti_cross_ref = 0;
2718 info->mti_opdata = 0;
2719 info->mti_big_lmm_used = 0;
2721 info->mti_spec.no_create = 0;
2722 info->mti_spec.sp_rm_entry = 0;
2724 info->mti_spec.u.sp_ea.eadata = NULL;
2725 info->mti_spec.u.sp_ea.eadatalen = 0;
2728 void mdt_thread_info_fini(struct mdt_thread_info *info)
2732 if (info->mti_object != NULL) {
2733 mdt_object_put(info->mti_env, info->mti_object);
2734 info->mti_object = NULL;
2737 for (i = 0; i < ARRAY_SIZE(info->mti_lh); i++)
2738 mdt_lock_handle_fini(&info->mti_lh[i]);
2739 info->mti_env = NULL;
2740 info->mti_pill = NULL;
2741 info->mti_exp = NULL;
2743 if (unlikely(info->mti_big_buf.lb_buf != NULL))
2744 lu_buf_free(&info->mti_big_buf);
2747 struct mdt_thread_info *tsi2mdt_info(struct tgt_session_info *tsi)
2749 struct mdt_thread_info *mti;
2750 struct lustre_capa *lc;
2752 mti = mdt_th_info(tsi->tsi_env);
2753 LASSERT(mti != NULL);
2755 mdt_thread_info_init(tgt_ses_req(tsi), mti);
2756 if (tsi->tsi_corpus != NULL) {
2757 struct req_capsule *pill = tsi->tsi_pill;
2759 mti->mti_object = mdt_obj(tsi->tsi_corpus);
2760 lu_object_get(tsi->tsi_corpus);
2763 * XXX: must be part of tgt_mdt_body_unpack but moved here
2764 * due to mdt_set_capainfo().
2766 if (req_capsule_has_field(pill, &RMF_CAPA1, RCL_CLIENT) &&
2767 req_capsule_get_size(pill, &RMF_CAPA1, RCL_CLIENT) > 0) {
2768 lc = req_capsule_client_get(pill, &RMF_CAPA1);
2769 mdt_set_capainfo(mti, 0, &tsi->tsi_mdt_body->fid1, lc);
2772 mti->mti_body = tsi->tsi_mdt_body;
2773 mti->mti_dlm_req = tsi->tsi_dlm_req;
2778 int mdt_tgt_connect(struct tgt_session_info *tsi)
2780 struct ptlrpc_request *req = tgt_ses_req(tsi);
2785 rc = tgt_connect(tsi);
2789 rc = mdt_init_idmap(tsi);
2794 obd_disconnect(class_export_get(req->rq_export));
2813 static int mdt_intent_getattr(enum mdt_it_code opcode,
2814 struct mdt_thread_info *info,
2815 struct ldlm_lock **,
2818 static int mdt_intent_getxattr(enum mdt_it_code opcode,
2819 struct mdt_thread_info *info,
2820 struct ldlm_lock **lockp,
2823 static int mdt_intent_layout(enum mdt_it_code opcode,
2824 struct mdt_thread_info *info,
2825 struct ldlm_lock **,
2827 static int mdt_intent_reint(enum mdt_it_code opcode,
2828 struct mdt_thread_info *info,
2829 struct ldlm_lock **,
2832 static struct mdt_it_flavor {
2833 const struct req_format *it_fmt;
2835 int (*it_act)(enum mdt_it_code ,
2836 struct mdt_thread_info *,
2837 struct ldlm_lock **,
2840 } mdt_it_flavor[] = {
2842 .it_fmt = &RQF_LDLM_INTENT,
2843 /*.it_flags = HABEO_REFERO,*/
2845 .it_act = mdt_intent_reint,
2846 .it_reint = REINT_OPEN
2849 .it_fmt = &RQF_LDLM_INTENT,
2851 * OCREAT is not a MUTABOR request as if the file
2853 * We do the extra check of OBD_CONNECT_RDONLY in
2854 * mdt_reint_open() when we really need to create
2858 .it_act = mdt_intent_reint,
2859 .it_reint = REINT_OPEN
2862 .it_fmt = &RQF_LDLM_INTENT,
2863 .it_flags = MUTABOR,
2864 .it_act = mdt_intent_reint,
2865 .it_reint = REINT_CREATE
2867 [MDT_IT_GETATTR] = {
2868 .it_fmt = &RQF_LDLM_INTENT_GETATTR,
2869 .it_flags = HABEO_REFERO,
2870 .it_act = mdt_intent_getattr
2872 [MDT_IT_READDIR] = {
2878 .it_fmt = &RQF_LDLM_INTENT_GETATTR,
2879 .it_flags = HABEO_REFERO,
2880 .it_act = mdt_intent_getattr
2883 .it_fmt = &RQF_LDLM_INTENT_UNLINK,
2884 .it_flags = MUTABOR,
2886 .it_reint = REINT_UNLINK
2890 .it_flags = MUTABOR,
2893 [MDT_IT_GETXATTR] = {
2894 .it_fmt = &RQF_LDLM_INTENT_GETXATTR,
2895 .it_flags = HABEO_CORPUS,
2896 .it_act = mdt_intent_getxattr
2899 .it_fmt = &RQF_LDLM_INTENT_LAYOUT,
2901 .it_act = mdt_intent_layout
2905 int mdt_intent_lock_replace(struct mdt_thread_info *info,
2906 struct ldlm_lock **lockp,
2907 struct ldlm_lock *new_lock,
2908 struct mdt_lock_handle *lh,
2911 struct ptlrpc_request *req = mdt_info_req(info);
2912 struct ldlm_lock *lock = *lockp;
2915 * Get new lock only for cases when possible resent did not find any
2918 if (new_lock == NULL)
2919 new_lock = ldlm_handle2lock_long(&lh->mlh_reg_lh, 0);
2921 if (new_lock == NULL && (flags & LDLM_FL_INTENT_ONLY)) {
2922 lh->mlh_reg_lh.cookie = 0;
2926 LASSERTF(new_lock != NULL,
2927 "lockh "LPX64"\n", lh->mlh_reg_lh.cookie);
2930 * If we've already given this lock to a client once, then we should
2931 * have no readers or writers. Otherwise, we should have one reader
2932 * _or_ writer ref (which will be zeroed below) before returning the
2935 if (new_lock->l_export == req->rq_export) {
2936 LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
2938 LASSERT(new_lock->l_export == NULL);
2939 LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
2944 if (new_lock->l_export == req->rq_export) {
2946 * Already gave this to the client, which means that we
2947 * reconstructed a reply.
2949 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
2951 lh->mlh_reg_lh.cookie = 0;
2952 RETURN(ELDLM_LOCK_REPLACED);
2956 * Fixup the lock to be given to the client.
2958 lock_res_and_lock(new_lock);
2959 /* Zero new_lock->l_readers and new_lock->l_writers without triggering
2960 * possible blocking AST. */
2961 while (new_lock->l_readers > 0) {
2962 lu_ref_del(&new_lock->l_reference, "reader", new_lock);
2963 lu_ref_del(&new_lock->l_reference, "user", new_lock);
2964 new_lock->l_readers--;
2966 while (new_lock->l_writers > 0) {
2967 lu_ref_del(&new_lock->l_reference, "writer", new_lock);
2968 lu_ref_del(&new_lock->l_reference, "user", new_lock);
2969 new_lock->l_writers--;
2972 new_lock->l_export = class_export_lock_get(req->rq_export, new_lock);
2973 new_lock->l_blocking_ast = lock->l_blocking_ast;
2974 new_lock->l_completion_ast = lock->l_completion_ast;
2975 new_lock->l_remote_handle = lock->l_remote_handle;
2976 new_lock->l_flags &= ~LDLM_FL_LOCAL;
2978 unlock_res_and_lock(new_lock);
2980 cfs_hash_add(new_lock->l_export->exp_lock_hash,
2981 &new_lock->l_remote_handle,
2982 &new_lock->l_exp_hash);
2984 LDLM_LOCK_RELEASE(new_lock);
2985 lh->mlh_reg_lh.cookie = 0;
2987 RETURN(ELDLM_LOCK_REPLACED);
2990 static void mdt_intent_fixup_resent(struct mdt_thread_info *info,
2991 struct ldlm_lock *new_lock,
2992 struct ldlm_lock **old_lock,
2993 struct mdt_lock_handle *lh,
2994 enum mdt_it_code opcode)
2996 struct ptlrpc_request *req = mdt_info_req(info);
2997 struct obd_export *exp = req->rq_export;
2998 struct lustre_handle remote_hdl;
2999 struct ldlm_request *dlmreq;
3000 struct ldlm_lock *lock;
3002 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
3005 dlmreq = req_capsule_client_get(info->mti_pill, &RMF_DLM_REQ);
3006 remote_hdl = dlmreq->lock_handle[0];
3007 /* If the client does not require open lock, it does not need to
3008 * search lock in exp_lock_hash, since the server thread will
3009 * make sure the lock will be released, and the resend request
3010 * can always re-enqueue the lock */
3011 if ((opcode != MDT_IT_OPEN) || (opcode == MDT_IT_OPEN &&
3012 info->mti_spec.sp_cr_flags & MDS_OPEN_LOCK)) {
3013 /* In the function below, .hs_keycmp resolves to
3014 * ldlm_export_lock_keycmp() */
3015 /* coverity[overrun-buffer-val] */
3016 lock = cfs_hash_lookup(exp->exp_lock_hash, &remote_hdl);
3018 lock_res_and_lock(lock);
3019 if (lock != new_lock) {
3020 lh->mlh_reg_lh.cookie = lock->l_handle.h_cookie;
3021 lh->mlh_reg_mode = lock->l_granted_mode;
3023 LDLM_DEBUG(lock, "Restoring lock cookie");
3024 DEBUG_REQ(D_DLMTRACE, req,
3025 "restoring lock cookie "LPX64,
3026 lh->mlh_reg_lh.cookie);
3028 *old_lock = LDLM_LOCK_GET(lock);
3029 cfs_hash_put(exp->exp_lock_hash,
3031 unlock_res_and_lock(lock);
3034 cfs_hash_put(exp->exp_lock_hash, &lock->l_exp_hash);
3035 unlock_res_and_lock(lock);
3039 * If the xid matches, then we know this is a resent request, and allow
3040 * it. (It's probably an OPEN, for which we don't send a lock.
3042 if (req_xid_is_last(req))
3046 * This remote handle isn't enqueued, so we never received or processed
3047 * this request. Clear MSG_RESENT, because it can be handled like any
3048 * normal request now.
3050 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
3052 DEBUG_REQ(D_DLMTRACE, req, "no existing lock with rhandle "LPX64,
3056 static int mdt_intent_getxattr(enum mdt_it_code opcode,
3057 struct mdt_thread_info *info,
3058 struct ldlm_lock **lockp,
3061 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
3062 struct ldlm_reply *ldlm_rep = NULL;
3066 * Initialize lhc->mlh_reg_lh either from a previously granted lock
3067 * (for the resend case) or a new lock. Below we will use it to
3068 * replace the original lock.
3070 mdt_intent_fixup_resent(info, *lockp, NULL, lhc, opcode);
3071 if (!lustre_handle_is_used(&lhc->mlh_reg_lh)) {
3072 mdt_lock_reg_init(lhc, (*lockp)->l_req_mode);
3073 rc = mdt_object_lock(info, info->mti_object, lhc,
3074 MDS_INODELOCK_XATTR,
3080 grc = mdt_getxattr(info);
3082 rc = mdt_intent_lock_replace(info, lockp, NULL, lhc, flags);
3084 if (mdt_info_req(info)->rq_repmsg != NULL)
3085 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
3086 if (ldlm_rep == NULL)
3087 RETURN(err_serious(-EFAULT));
3089 ldlm_rep->lock_policy_res2 = grc;
3094 static int mdt_intent_getattr(enum mdt_it_code opcode,
3095 struct mdt_thread_info *info,
3096 struct ldlm_lock **lockp,
3099 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
3100 struct ldlm_lock *new_lock = NULL;
3102 struct ldlm_reply *ldlm_rep;
3103 struct ptlrpc_request *req;
3104 struct mdt_body *reqbody;
3105 struct mdt_body *repbody;
3109 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
3112 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
3115 info->mti_cross_ref = !!(reqbody->valid & OBD_MD_FLCROSSREF);
3116 repbody->eadatasize = 0;
3117 repbody->aclsize = 0;
3121 child_bits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
3123 case MDT_IT_GETATTR:
3124 child_bits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE |
3128 CERROR("Unsupported intent (%d)\n", opcode);
3129 GOTO(out_shrink, rc = -EINVAL);
3132 rc = mdt_init_ucred(info, reqbody);
3134 GOTO(out_shrink, rc);
3136 req = info->mti_pill->rc_req;
3137 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
3138 mdt_set_disposition(info, ldlm_rep, DISP_IT_EXECD);
3140 /* Get lock from request for possible resent case. */
3141 mdt_intent_fixup_resent(info, *lockp, &new_lock, lhc, opcode);
3143 rc = mdt_getattr_name_lock(info, lhc, child_bits, ldlm_rep);
3144 ldlm_rep->lock_policy_res2 = clear_serious(rc);
3146 if (mdt_get_disposition(ldlm_rep, DISP_LOOKUP_NEG))
3147 ldlm_rep->lock_policy_res2 = 0;
3148 if (!mdt_get_disposition(ldlm_rep, DISP_LOOKUP_POS) ||
3149 ldlm_rep->lock_policy_res2) {
3150 lhc->mlh_reg_lh.cookie = 0ull;
3151 GOTO(out_ucred, rc = ELDLM_LOCK_ABORTED);
3154 rc = mdt_intent_lock_replace(info, lockp, new_lock, lhc, flags);
3157 mdt_exit_ucred(info);
3159 mdt_client_compatibility(info);
3160 rc2 = mdt_fix_reply(info);
3166 static int mdt_intent_layout(enum mdt_it_code opcode,
3167 struct mdt_thread_info *info,
3168 struct ldlm_lock **lockp,
3171 struct layout_intent *layout;
3173 struct mdt_object *obj = NULL;
3174 struct md_object *child = NULL;
3178 if (opcode != MDT_IT_LAYOUT) {
3179 CERROR("%s: Unknown intent (%d)\n", mdt_obd_name(info->mti_mdt),
3184 fid = &info->mti_tmp_fid2;
3185 fid_extract_from_res_name(fid, &(*lockp)->l_resource->lr_name);
3187 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
3189 RETURN(PTR_ERR(obj));
3191 if (mdt_object_exists(obj) && !mdt_object_remote(obj)) {
3192 child = mdt_object_child(obj);
3194 /* get the length of lsm */
3195 rc = mo_xattr_get(info->mti_env, child, &LU_BUF_NULL,
3198 if (rc > info->mti_mdt->mdt_max_mdsize)
3199 info->mti_mdt->mdt_max_mdsize = rc;
3202 mdt_object_put(info->mti_env, obj);
3204 (*lockp)->l_lvb_type = LVB_T_LAYOUT;
3205 req_capsule_set_size(info->mti_pill, &RMF_DLM_LVB, RCL_SERVER,
3206 ldlm_lvbo_size(*lockp));
3207 rc = req_capsule_server_pack(info->mti_pill);
3211 layout = req_capsule_client_get(info->mti_pill, &RMF_LAYOUT_INTENT);
3212 LASSERT(layout != NULL);
3213 if (layout->li_opc == LAYOUT_INTENT_ACCESS)
3214 /* return to normal ldlm handling */
3217 CERROR("%s: Unsupported layout intent (%d)\n",
3218 mdt_obd_name(info->mti_mdt), layout->li_opc);
3222 static int mdt_intent_reint(enum mdt_it_code opcode,
3223 struct mdt_thread_info *info,
3224 struct ldlm_lock **lockp,
3227 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
3228 struct ldlm_reply *rep = NULL;
3232 static const struct req_format *intent_fmts[REINT_MAX] = {
3233 [REINT_CREATE] = &RQF_LDLM_INTENT_CREATE,
3234 [REINT_OPEN] = &RQF_LDLM_INTENT_OPEN
3239 opc = mdt_reint_opcode(mdt_info_req(info), intent_fmts);
3243 if (mdt_it_flavor[opcode].it_reint != opc) {
3244 CERROR("Reint code %ld doesn't match intent: %d\n",
3246 RETURN(err_serious(-EPROTO));
3249 /* Get lock from request for possible resent case. */
3250 mdt_intent_fixup_resent(info, *lockp, NULL, lhc, opcode);
3252 rc = mdt_reint_internal(info, lhc, opc);
3254 /* Check whether the reply has been packed successfully. */
3255 if (mdt_info_req(info)->rq_repmsg != NULL)
3256 rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
3258 RETURN(err_serious(-EFAULT));
3260 /* MDC expects this in any case */
3262 mdt_set_disposition(info, rep, DISP_LOOKUP_EXECD);
3264 /* the open lock or the lock for cross-ref object should be
3265 * returned to the client */
3266 if (rc == -EREMOTE || mdt_get_disposition(rep, DISP_OPEN_LOCK)) {
3267 LASSERT(lustre_handle_is_used(&lhc->mlh_reg_lh));
3268 rep->lock_policy_res2 = 0;
3269 rc = mdt_intent_lock_replace(info, lockp, NULL, lhc, flags);
3273 rep->lock_policy_res2 = clear_serious(rc);
3275 if (rep->lock_policy_res2 == -ENOENT &&
3276 mdt_get_disposition(rep, DISP_LOOKUP_NEG))
3277 rep->lock_policy_res2 = 0;
3279 if (rc == -ENOTCONN || rc == -ENODEV ||
3280 rc == -EOVERFLOW) { /**< if VBR failure then return error */
3282 * If it is the disconnect error (ENODEV & ENOCONN), the error
3283 * will be returned by rq_status, and client at ptlrpc layer
3284 * will detect this, then disconnect, reconnect the import
3285 * immediately, instead of impacting the following the rpc.
3287 lhc->mlh_reg_lh.cookie = 0ull;
3291 * For other cases, the error will be returned by intent.
3292 * and client will retrieve the result from intent.
3295 * FIXME: when open lock is finished, that should be
3298 if (lustre_handle_is_used(&lhc->mlh_reg_lh)) {
3299 LASSERTF(rc == 0, "Error occurred but lock handle "
3300 "is still in use, rc = %d\n", rc);
3301 rep->lock_policy_res2 = 0;
3302 rc = mdt_intent_lock_replace(info, lockp, NULL, lhc, flags);
3305 lhc->mlh_reg_lh.cookie = 0ull;
3306 RETURN(ELDLM_LOCK_ABORTED);
3311 static int mdt_intent_code(long itcode)
3319 case IT_OPEN|IT_CREAT:
3326 rc = MDT_IT_READDIR;
3329 rc = MDT_IT_GETATTR;
3341 rc = MDT_IT_GETXATTR;
3346 case IT_QUOTA_DQACQ:
3351 CERROR("Unknown intent opcode: %ld\n", itcode);
3358 static int mdt_intent_opc(long itopc, struct mdt_thread_info *info,
3359 struct ldlm_lock **lockp, __u64 flags)
3361 struct req_capsule *pill;
3362 struct mdt_it_flavor *flv;
3367 opc = mdt_intent_code(itopc);
3371 pill = info->mti_pill;
3373 if (opc == MDT_IT_QUOTA) {
3374 struct lu_device *qmt = info->mti_mdt->mdt_qmt_dev;
3377 RETURN(-EOPNOTSUPP);
3379 (*lockp)->l_lvb_type = LVB_T_LQUOTA;
3380 /* pass the request to quota master */
3381 rc = qmt_hdls.qmth_intent_policy(info->mti_env, qmt,
3382 mdt_info_req(info), lockp,
3387 flv = &mdt_it_flavor[opc];
3388 if (flv->it_fmt != NULL)
3389 req_capsule_extend(pill, flv->it_fmt);
3391 rc = mdt_unpack_req_pack_rep(info, flv->it_flags);
3393 struct ptlrpc_request *req = mdt_info_req(info);
3394 if (flv->it_flags & MUTABOR &&
3395 exp_connect_flags(req->rq_export) & OBD_CONNECT_RDONLY)
3398 if (rc == 0 && flv->it_act != NULL) {
3399 struct ldlm_reply *rep;
3401 /* execute policy */
3402 rc = flv->it_act(opc, info, lockp, flags);
3404 /* Check whether the reply has been packed successfully. */
3405 if (mdt_info_req(info)->rq_repmsg != NULL) {
3406 rep = req_capsule_server_get(info->mti_pill,
3408 rep->lock_policy_res2 =
3409 ptlrpc_status_hton(rep->lock_policy_res2);
3417 static int mdt_intent_policy(struct ldlm_namespace *ns,
3418 struct ldlm_lock **lockp, void *req_cookie,
3419 ldlm_mode_t mode, __u64 flags, void *data)
3421 struct tgt_session_info *tsi;
3422 struct mdt_thread_info *info;
3423 struct ptlrpc_request *req = req_cookie;
3424 struct ldlm_intent *it;
3425 struct req_capsule *pill;
3430 LASSERT(req != NULL);
3432 tsi = tgt_ses_info(req->rq_svc_thread->t_env);
3434 info = tsi2mdt_info(tsi);
3435 LASSERT(info != NULL);
3436 pill = info->mti_pill;
3437 LASSERT(pill->rc_req == req);
3439 if (req->rq_reqmsg->lm_bufcount > DLM_INTENT_IT_OFF) {
3440 req_capsule_extend(pill, &RQF_LDLM_INTENT_BASIC);
3441 it = req_capsule_client_get(pill, &RMF_LDLM_INTENT);
3443 rc = mdt_intent_opc(it->opc, info, lockp, flags);
3447 /* Lock without inodebits makes no sense and will oops
3448 * later in ldlm. Let's check it now to see if we have
3449 * ibits corrupted somewhere in mdt_intent_opc().
3450 * The case for client miss to set ibits has been
3451 * processed by others. */
3452 LASSERT(ergo(info->mti_dlm_req->lock_desc.l_resource.\
3453 lr_type == LDLM_IBITS,
3454 info->mti_dlm_req->lock_desc.\
3455 l_policy_data.l_inodebits.bits != 0));
3457 rc = err_serious(-EFAULT);
3459 /* No intent was provided */
3460 LASSERT(pill->rc_fmt == &RQF_LDLM_ENQUEUE);
3461 req_capsule_set_size(pill, &RMF_DLM_LVB, RCL_SERVER, 0);
3462 rc = req_capsule_server_pack(pill);
3464 rc = err_serious(rc);
3466 mdt_thread_info_fini(info);
3470 static void mdt_deregister_seq_exp(struct mdt_device *mdt)
3472 struct seq_server_site *ss = mdt_seq_site(mdt);
3474 if (ss->ss_node_id == 0)
3477 if (ss->ss_client_seq != NULL) {
3478 lustre_deregister_lwp_item(&ss->ss_client_seq->lcs_exp);
3479 ss->ss_client_seq->lcs_exp = NULL;
3482 if (ss->ss_server_fld != NULL) {
3483 lustre_deregister_lwp_item(&ss->ss_server_fld->lsf_control_exp);
3484 ss->ss_server_fld->lsf_control_exp = NULL;
3488 static void mdt_seq_fini_cli(struct mdt_device *mdt)
3490 struct seq_server_site *ss = mdt_seq_site(mdt);
3495 if (ss->ss_server_seq == NULL)
3496 seq_server_set_cli(NULL, ss->ss_server_seq, NULL);
3501 static int mdt_seq_fini(const struct lu_env *env, struct mdt_device *mdt)
3503 mdt_seq_fini_cli(mdt);
3504 mdt_deregister_seq_exp(mdt);
3506 return seq_site_fini(env, mdt_seq_site(mdt));
3510 * It will retrieve its FLDB entries from MDT0, and it only happens
3511 * when upgrading existent FS to 2.6 or when local FLDB is corrupted,
3512 * and it needs to refresh FLDB from the MDT0.
3514 static int mdt_register_lwp_callback(void *data)
3517 struct mdt_device *mdt = data;
3518 struct lu_server_fld *fld = mdt_seq_site(mdt)->ss_server_fld;
3522 LASSERT(mdt_seq_site(mdt)->ss_node_id != 0);
3524 if (!likely(fld->lsf_new))
3527 rc = lu_env_init(&env, LCT_MD_THREAD);
3529 CERROR("%s: cannot init env: rc = %d\n", mdt_obd_name(mdt), rc);
3533 rc = fld_update_from_controller(&env, fld);
3535 CERROR("%s: cannot update controller: rc = %d\n",
3536 mdt_obd_name(mdt), rc);
3544 static int mdt_register_seq_exp(struct mdt_device *mdt)
3546 struct seq_server_site *ss = mdt_seq_site(mdt);
3547 char *lwp_name = NULL;
3550 if (ss->ss_node_id == 0)
3553 OBD_ALLOC(lwp_name, MAX_OBD_NAME);
3554 if (lwp_name == NULL)
3555 GOTO(out_free, rc = -ENOMEM);
3557 rc = tgt_name2lwp_name(mdt_obd_name(mdt), lwp_name, MAX_OBD_NAME, 0);
3561 rc = lustre_register_lwp_item(lwp_name, &ss->ss_client_seq->lcs_exp,
3566 rc = lustre_register_lwp_item(lwp_name,
3567 &ss->ss_server_fld->lsf_control_exp,
3568 mdt_register_lwp_callback, mdt);
3570 lustre_deregister_lwp_item(&ss->ss_client_seq->lcs_exp);
3571 ss->ss_client_seq->lcs_exp = NULL;
3575 if (lwp_name != NULL)
3576 OBD_FREE(lwp_name, MAX_OBD_NAME);
3582 * Init client sequence manager which is used by local MDS to talk to sequence
3583 * controller on remote node.
3585 static int mdt_seq_init_cli(const struct lu_env *env, struct mdt_device *mdt)
3587 struct seq_server_site *ss = mdt_seq_site(mdt);
3592 /* check if this is adding the first MDC and controller is not yet
3594 OBD_ALLOC_PTR(ss->ss_client_seq);
3595 if (ss->ss_client_seq == NULL)
3598 OBD_ALLOC(prefix, MAX_OBD_NAME + 5);
3599 if (prefix == NULL) {
3600 OBD_FREE_PTR(ss->ss_client_seq);
3601 ss->ss_client_seq = NULL;
3605 /* Note: seq_client_fini will be called in seq_site_fini */
3606 snprintf(prefix, MAX_OBD_NAME + 5, "ctl-%s", mdt_obd_name(mdt));
3607 rc = seq_client_init(ss->ss_client_seq, NULL, LUSTRE_SEQ_METADATA,
3608 prefix, ss->ss_node_id == 0 ? ss->ss_control_seq :
3610 OBD_FREE(prefix, MAX_OBD_NAME + 5);
3612 OBD_FREE_PTR(ss->ss_client_seq);
3613 ss->ss_client_seq = NULL;
3617 rc = seq_server_set_cli(env, ss->ss_server_seq, ss->ss_client_seq);
3622 static int mdt_seq_init(const struct lu_env *env, struct mdt_device *mdt)
3624 struct seq_server_site *ss;
3628 ss = mdt_seq_site(mdt);
3629 /* init sequence controller server(MDT0) */
3630 if (ss->ss_node_id == 0) {
3631 OBD_ALLOC_PTR(ss->ss_control_seq);
3632 if (ss->ss_control_seq == NULL)
3635 rc = seq_server_init(env, ss->ss_control_seq, mdt->mdt_bottom,
3636 mdt_obd_name(mdt), LUSTRE_SEQ_CONTROLLER,
3639 GOTO(out_seq_fini, rc);
3642 /* Init normal sequence server */
3643 OBD_ALLOC_PTR(ss->ss_server_seq);
3644 if (ss->ss_server_seq == NULL)
3645 GOTO(out_seq_fini, rc = -ENOMEM);
3647 rc = seq_server_init(env, ss->ss_server_seq, mdt->mdt_bottom,
3648 mdt_obd_name(mdt), LUSTRE_SEQ_SERVER, ss);
3650 GOTO(out_seq_fini, rc);
3652 /* init seq client for seq server to talk to seq controller(MDT0) */
3653 rc = mdt_seq_init_cli(env, mdt);
3655 GOTO(out_seq_fini, rc);
3657 if (ss->ss_node_id != 0)
3658 /* register controler export through lwp */
3659 rc = mdt_register_seq_exp(mdt);
3664 mdt_seq_fini(env, mdt);
3672 static int mdt_fld_fini(const struct lu_env *env,
3673 struct mdt_device *m)
3675 struct seq_server_site *ss = mdt_seq_site(m);
3678 if (ss && ss->ss_server_fld) {
3679 fld_server_fini(env, ss->ss_server_fld);
3680 OBD_FREE_PTR(ss->ss_server_fld);
3681 ss->ss_server_fld = NULL;
3687 static int mdt_fld_init(const struct lu_env *env,
3689 struct mdt_device *m)
3691 struct seq_server_site *ss;
3695 ss = mdt_seq_site(m);
3697 OBD_ALLOC_PTR(ss->ss_server_fld);
3698 if (ss->ss_server_fld == NULL)
3699 RETURN(rc = -ENOMEM);
3701 rc = fld_server_init(env, ss->ss_server_fld, m->mdt_bottom, uuid,
3704 OBD_FREE_PTR(ss->ss_server_fld);
3705 ss->ss_server_fld = NULL;
3712 static void mdt_stack_pre_fini(const struct lu_env *env,
3713 struct mdt_device *m, struct lu_device *top)
3715 struct obd_device *obd;
3716 struct lustre_cfg_bufs *bufs;
3717 struct lustre_cfg *lcfg;
3718 struct mdt_thread_info *info;
3723 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
3724 LASSERT(info != NULL);
3726 bufs = &info->mti_u.bufs;
3728 LASSERT(m->mdt_child_exp);
3729 LASSERT(m->mdt_child_exp->exp_obd);
3730 obd = m->mdt_child_exp->exp_obd;
3732 /* process cleanup, pass mdt obd name to get obd umount flags */
3733 /* XXX: this is needed because all layers are referenced by
3734 * objects (some of them are pinned by osd, for example *
3735 * the proper solution should be a model where object used
3736 * by osd only doesn't have mdt/mdd slices -bzzz */
3737 lustre_cfg_bufs_reset(bufs, mdt_obd_name(m));
3738 lustre_cfg_bufs_set_string(bufs, 1, NULL);
3739 lcfg = lustre_cfg_new(LCFG_PRE_CLEANUP, bufs);
3741 CERROR("%s:Cannot alloc lcfg!\n", mdt_obd_name(m));
3744 top->ld_ops->ldo_process_config(env, top, lcfg);
3745 lustre_cfg_free(lcfg);
3749 static void mdt_stack_fini(const struct lu_env *env,
3750 struct mdt_device *m, struct lu_device *top)
3752 struct obd_device *obd = mdt2obd_dev(m);
3753 struct lustre_cfg_bufs *bufs;
3754 struct lustre_cfg *lcfg;
3755 struct mdt_thread_info *info;
3759 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
3760 LASSERT(info != NULL);
3762 lu_dev_del_linkage(top->ld_site, top);
3764 lu_site_purge(env, top->ld_site, -1);
3766 bufs = &info->mti_u.bufs;
3767 /* process cleanup, pass mdt obd name to get obd umount flags */
3768 /* another purpose is to let all layers to release their objects */
3769 lustre_cfg_bufs_reset(bufs, mdt_obd_name(m));
3774 lustre_cfg_bufs_set_string(bufs, 1, flags);
3775 lcfg = lustre_cfg_new(LCFG_CLEANUP, bufs);
3777 CERROR("Cannot alloc lcfg!\n");
3781 top->ld_ops->ldo_process_config(env, top, lcfg);
3782 lustre_cfg_free(lcfg);
3784 lu_site_purge(env, top->ld_site, -1);
3786 m->mdt_child = NULL;
3787 m->mdt_bottom = NULL;
3789 obd_disconnect(m->mdt_child_exp);
3790 m->mdt_child_exp = NULL;
3792 obd_disconnect(m->mdt_bottom_exp);
3793 m->mdt_child_exp = NULL;
3796 static int mdt_connect_to_next(const struct lu_env *env, struct mdt_device *m,
3797 const char *next, struct obd_export **exp)
3799 struct obd_connect_data *data = NULL;
3800 struct obd_device *obd;
3804 OBD_ALLOC_PTR(data);
3806 GOTO(out, rc = -ENOMEM);
3808 obd = class_name2obd(next);
3810 CERROR("%s: can't locate next device: %s\n",
3811 mdt_obd_name(m), next);
3812 GOTO(out, rc = -ENOTCONN);
3815 data->ocd_connect_flags = OBD_CONNECT_VERSION;
3816 data->ocd_version = LUSTRE_VERSION_CODE;
3818 rc = obd_connect(NULL, exp, obd, &obd->obd_uuid, data, NULL);
3820 CERROR("%s: cannot connect to next dev %s (%d)\n",
3821 mdt_obd_name(m), next, rc);
3831 static int mdt_stack_init(const struct lu_env *env, struct mdt_device *mdt,
3832 struct lustre_cfg *cfg)
3834 char *dev = lustre_cfg_string(cfg, 0);
3835 int rc, name_size, uuid_size;
3836 char *name, *uuid, *p;
3837 struct lustre_cfg_bufs *bufs;
3838 struct lustre_cfg *lcfg;
3839 struct obd_device *obd;
3840 struct lustre_profile *lprof;
3841 struct lu_site *site;
3844 /* in 1.8 we had the only device in the stack - MDS.
3845 * 2.0 introduces MDT, MDD, OSD; MDT starts others internally.
3846 * in 2.3 OSD is instantiated by obd_mount.c, so we need
3847 * to generate names and setup MDT, MDD. MDT will be using
3848 * generated name to connect to MDD. for MDD the next device
3849 * will be LOD with name taken from so called "profile" which
3850 * is generated by mount_option line
3852 * 1.8 MGS generates config. commands like this:
3853 * #06 (104)mount_option 0: 1:lustre-MDT0000 2:lustre-mdtlov
3854 * #08 (120)setup 0:lustre-MDT0000 1:dev 2:type 3:lustre-MDT0000
3855 * 2.0 MGS generates config. commands like this:
3856 * #07 (112)mount_option 0: 1:lustre-MDT0000 2:lustre-MDT0000-mdtlov
3857 * #08 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
3858 * 3:lustre-MDT0000-mdtlov 4:f
3860 * we generate MDD name from MDT one, just replacing T with D
3862 * after all the preparations, the logical equivalent will be
3863 * #01 (160)setup 0:lustre-MDD0000 1:lustre-MDD0000_UUID 2:0
3864 * 3:lustre-MDT0000-mdtlov 4:f
3865 * #02 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
3866 * 3:lustre-MDD0000 4:f
3868 * notice we build the stack from down to top: MDD first, then MDT */
3870 name_size = MAX_OBD_NAME;
3871 uuid_size = MAX_OBD_NAME;
3873 OBD_ALLOC(name, name_size);
3874 OBD_ALLOC(uuid, uuid_size);
3875 if (name == NULL || uuid == NULL)
3876 GOTO(cleanup_mem, rc = -ENOMEM);
3878 OBD_ALLOC_PTR(bufs);
3880 GOTO(cleanup_mem, rc = -ENOMEM);
3883 p = strstr(name, "-MDT");
3885 GOTO(free_bufs, rc = -ENOMEM);
3888 snprintf(uuid, MAX_OBD_NAME, "%s_UUID", name);
3890 lprof = class_get_profile(lustre_cfg_string(cfg, 0));
3891 if (lprof == NULL || lprof->lp_dt == NULL) {
3892 CERROR("can't find the profile: %s\n",
3893 lustre_cfg_string(cfg, 0));
3894 GOTO(free_bufs, rc = -EINVAL);
3897 lustre_cfg_bufs_reset(bufs, name);
3898 lustre_cfg_bufs_set_string(bufs, 1, LUSTRE_MDD_NAME);
3899 lustre_cfg_bufs_set_string(bufs, 2, uuid);
3900 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
3902 lcfg = lustre_cfg_new(LCFG_ATTACH, bufs);
3904 GOTO(free_bufs, rc = -ENOMEM);
3906 rc = class_attach(lcfg);
3908 GOTO(lcfg_cleanup, rc);
3910 obd = class_name2obd(name);
3912 CERROR("Can not find obd %s (%s in config)\n",
3913 MDD_OBD_NAME, lustre_cfg_string(cfg, 0));
3914 GOTO(class_detach, rc = -EINVAL);
3917 lustre_cfg_free(lcfg);
3919 lustre_cfg_bufs_reset(bufs, name);
3920 lustre_cfg_bufs_set_string(bufs, 1, uuid);
3921 lustre_cfg_bufs_set_string(bufs, 2, dev);
3922 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
3924 lcfg = lustre_cfg_new(LCFG_SETUP, bufs);
3926 rc = class_setup(obd, lcfg);
3928 GOTO(class_detach, rc);
3930 /* connect to MDD we just setup */
3931 rc = mdt_connect_to_next(env, mdt, name, &mdt->mdt_child_exp);
3933 GOTO(class_detach, rc);
3935 site = mdt->mdt_child_exp->exp_obd->obd_lu_dev->ld_site;
3937 LASSERT(mdt_lu_site(mdt) == NULL);
3938 mdt->mdt_lu_dev.ld_site = site;
3939 site->ls_top_dev = &mdt->mdt_lu_dev;
3940 mdt->mdt_child = lu2md_dev(mdt->mdt_child_exp->exp_obd->obd_lu_dev);
3942 /* now connect to bottom OSD */
3943 snprintf(name, MAX_OBD_NAME, "%s-osd", dev);
3944 rc = mdt_connect_to_next(env, mdt, name, &mdt->mdt_bottom_exp);
3946 GOTO(class_detach, rc);
3948 lu2dt_dev(mdt->mdt_bottom_exp->exp_obd->obd_lu_dev);
3950 rc = lu_env_refill((struct lu_env *)env);
3952 CERROR("Failure to refill session: '%d'\n", rc);
3954 lu_dev_add_linkage(site, &mdt->mdt_lu_dev);
3959 class_detach(obd, lcfg);
3961 lustre_cfg_free(lcfg);
3966 OBD_FREE(name, name_size);
3968 OBD_FREE(uuid, uuid_size);
3972 /* setup quota master target on MDT0 */
3973 static int mdt_quota_init(const struct lu_env *env, struct mdt_device *mdt,
3974 struct lustre_cfg *cfg)
3976 struct obd_device *obd;
3977 char *dev = lustre_cfg_string(cfg, 0);
3978 char *qmtname, *uuid, *p;
3979 struct lustre_cfg_bufs *bufs;
3980 struct lustre_cfg *lcfg;
3981 struct lustre_profile *lprof;
3982 struct obd_connect_data *data;
3986 LASSERT(mdt->mdt_qmt_exp == NULL);
3987 LASSERT(mdt->mdt_qmt_dev == NULL);
3989 /* quota master is on MDT0 only for now */
3990 if (mdt->mdt_seq_site.ss_node_id != 0)
3993 /* MGS generates config commands which look as follows:
3994 * #01 (160)setup 0:lustre-MDT0000 1:lustre-MDT0000_UUID 2:0
3995 * 3:lustre-MDT0000-mdtlov 4:f
3997 * We generate the QMT name from the MDT one, just replacing MD with QM
3998 * after all the preparations, the logical equivalent will be:
3999 * #01 (160)setup 0:lustre-QMT0000 1:lustre-QMT0000_UUID 2:0
4000 * 3:lustre-MDT0000-osd 4:f */
4001 OBD_ALLOC(qmtname, MAX_OBD_NAME);
4002 OBD_ALLOC(uuid, UUID_MAX);
4003 OBD_ALLOC_PTR(bufs);
4004 OBD_ALLOC_PTR(data);
4005 if (qmtname == NULL || uuid == NULL || bufs == NULL || data == NULL)
4006 GOTO(cleanup_mem, rc = -ENOMEM);
4008 strcpy(qmtname, dev);
4009 p = strstr(qmtname, "-MDT");
4011 GOTO(cleanup_mem, rc = -ENOMEM);
4012 /* replace MD with QM */
4016 snprintf(uuid, UUID_MAX, "%s_UUID", qmtname);
4018 lprof = class_get_profile(lustre_cfg_string(cfg, 0));
4019 if (lprof == NULL || lprof->lp_dt == NULL) {
4020 CERROR("can't find profile for %s\n",
4021 lustre_cfg_string(cfg, 0));
4022 GOTO(cleanup_mem, rc = -EINVAL);
4025 lustre_cfg_bufs_reset(bufs, qmtname);
4026 lustre_cfg_bufs_set_string(bufs, 1, LUSTRE_QMT_NAME);
4027 lustre_cfg_bufs_set_string(bufs, 2, uuid);
4028 lustre_cfg_bufs_set_string(bufs, 3, lprof->lp_dt);
4030 lcfg = lustre_cfg_new(LCFG_ATTACH, bufs);
4032 GOTO(cleanup_mem, rc = -ENOMEM);
4034 rc = class_attach(lcfg);
4036 GOTO(lcfg_cleanup, rc);
4038 obd = class_name2obd(qmtname);
4040 CERROR("Can not find obd %s (%s in config)\n", qmtname,
4041 lustre_cfg_string(cfg, 0));
4042 GOTO(class_detach, rc = -EINVAL);
4045 lustre_cfg_free(lcfg);
4047 lustre_cfg_bufs_reset(bufs, qmtname);
4048 lustre_cfg_bufs_set_string(bufs, 1, uuid);
4049 lustre_cfg_bufs_set_string(bufs, 2, dev);
4051 /* for quota, the next device should be the OSD device */
4052 lustre_cfg_bufs_set_string(bufs, 3,
4053 mdt->mdt_bottom->dd_lu_dev.ld_obd->obd_name);
4055 lcfg = lustre_cfg_new(LCFG_SETUP, bufs);
4057 rc = class_setup(obd, lcfg);
4059 GOTO(class_detach, rc);
4061 mdt->mdt_qmt_dev = obd->obd_lu_dev;
4063 /* configure local quota objects */
4064 rc = mdt->mdt_qmt_dev->ld_ops->ldo_prepare(env,
4068 GOTO(class_cleanup, rc);
4070 /* connect to quota master target */
4071 data->ocd_connect_flags = OBD_CONNECT_VERSION;
4072 data->ocd_version = LUSTRE_VERSION_CODE;
4073 rc = obd_connect(NULL, &mdt->mdt_qmt_exp, obd, &obd->obd_uuid,
4076 CERROR("cannot connect to quota master device %s (%d)\n",
4078 GOTO(class_cleanup, rc);
4084 class_manual_cleanup(obd);
4085 mdt->mdt_qmt_dev = NULL;
4089 class_detach(obd, lcfg);
4091 lustre_cfg_free(lcfg);
4096 OBD_FREE(qmtname, MAX_OBD_NAME);
4098 OBD_FREE(uuid, UUID_MAX);
4104 /* Shutdown quota master target associated with mdt */
4105 static void mdt_quota_fini(const struct lu_env *env, struct mdt_device *mdt)
4109 if (mdt->mdt_qmt_exp == NULL)
4111 LASSERT(mdt->mdt_qmt_dev != NULL);
4113 /* the qmt automatically shuts down when the mdt disconnects */
4114 obd_disconnect(mdt->mdt_qmt_exp);
4115 mdt->mdt_qmt_exp = NULL;
4116 mdt->mdt_qmt_dev = NULL;
4120 /* mdt_getxattr() is used from mdt_intent_getxattr(), use this wrapper
4121 * for now. This will be removed along with converting rest of MDT code
4122 * to use tgt_session_info */
4123 int mdt_tgt_getxattr(struct tgt_session_info *tsi)
4125 struct mdt_thread_info *info = tsi2mdt_info(tsi);
4128 rc = mdt_getxattr(info);
4130 mdt_thread_info_fini(info);
4134 static struct tgt_handler mdt_tgt_handlers[] = {
4135 TGT_RPC_HANDLER(MDS_FIRST_OPC,
4136 0, MDS_CONNECT, mdt_tgt_connect,
4137 &RQF_CONNECT, LUSTRE_OBD_VERSION),
4138 TGT_RPC_HANDLER(MDS_FIRST_OPC,
4139 0, MDS_DISCONNECT, tgt_disconnect,
4140 &RQF_MDS_DISCONNECT, LUSTRE_OBD_VERSION),
4141 TGT_RPC_HANDLER(MDS_FIRST_OPC,
4142 HABEO_REFERO, MDS_SET_INFO, mdt_set_info,
4143 &RQF_OBD_SET_INFO, LUSTRE_MDS_VERSION),
4144 TGT_MDT_HDL(0, MDS_GET_INFO, mdt_get_info),
4145 TGT_MDT_HDL(0 | HABEO_REFERO, MDS_GETSTATUS, mdt_getstatus),
4146 TGT_MDT_HDL(HABEO_CORPUS, MDS_GETATTR, mdt_getattr),
4147 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_GETATTR_NAME,
4149 TGT_MDT_HDL(HABEO_CORPUS, MDS_GETXATTR, mdt_tgt_getxattr),
4150 TGT_MDT_HDL(0 | HABEO_REFERO, MDS_STATFS, mdt_statfs),
4151 TGT_MDT_HDL(0 | MUTABOR, MDS_REINT, mdt_reint),
4152 TGT_MDT_HDL(HABEO_CORPUS, MDS_CLOSE, mdt_close),
4153 TGT_MDT_HDL(HABEO_CORPUS, MDS_DONE_WRITING,
4155 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_READPAGE, mdt_readpage),
4156 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_SYNC, mdt_sync),
4157 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_IS_SUBDIR, mdt_is_subdir),
4158 TGT_MDT_HDL(0, MDS_QUOTACTL, mdt_quotactl),
4159 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO | MUTABOR, MDS_HSM_PROGRESS,
4161 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO | MUTABOR, MDS_HSM_CT_REGISTER,
4162 mdt_hsm_ct_register),
4163 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO | MUTABOR, MDS_HSM_CT_UNREGISTER,
4164 mdt_hsm_ct_unregister),
4165 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_HSM_STATE_GET,
4167 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO | MUTABOR, MDS_HSM_STATE_SET,
4169 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_HSM_ACTION, mdt_hsm_action),
4170 TGT_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_HSM_REQUEST,
4172 TGT_MDT_HDL(HABEO_CORPUS|HABEO_REFERO | MUTABOR, MDS_SWAP_LAYOUTS,
4176 static struct tgt_handler mdt_sec_ctx_ops[] = {
4177 TGT_SEC_HDL_VAR(0, SEC_CTX_INIT, mdt_sec_ctx_handle),
4178 TGT_SEC_HDL_VAR(0, SEC_CTX_INIT_CONT,mdt_sec_ctx_handle),
4179 TGT_SEC_HDL_VAR(0, SEC_CTX_FINI, mdt_sec_ctx_handle)
4182 static struct tgt_handler mdt_quota_ops[] = {
4183 TGT_QUOTA_HDL(HABEO_REFERO, QUOTA_DQACQ, mdt_quota_dqacq),
4186 static struct tgt_opc_slice mdt_common_slice[] = {
4188 .tos_opc_start = MDS_FIRST_OPC,
4189 .tos_opc_end = MDS_LAST_OPC,
4190 .tos_hs = mdt_tgt_handlers
4193 .tos_opc_start = OBD_FIRST_OPC,
4194 .tos_opc_end = OBD_LAST_OPC,
4195 .tos_hs = tgt_obd_handlers
4198 .tos_opc_start = LDLM_FIRST_OPC,
4199 .tos_opc_end = LDLM_LAST_OPC,
4200 .tos_hs = tgt_dlm_handlers
4203 .tos_opc_start = SEC_FIRST_OPC,
4204 .tos_opc_end = SEC_LAST_OPC,
4205 .tos_hs = mdt_sec_ctx_ops
4208 .tos_opc_start = OUT_UPDATE_FIRST_OPC,
4209 .tos_opc_end = OUT_UPDATE_LAST_OPC,
4210 .tos_hs = tgt_out_handlers
4213 .tos_opc_start = FLD_FIRST_OPC,
4214 .tos_opc_end = FLD_LAST_OPC,
4215 .tos_hs = fld_handlers
4218 .tos_opc_start = SEQ_FIRST_OPC,
4219 .tos_opc_end = SEQ_LAST_OPC,
4220 .tos_hs = seq_handlers
4223 .tos_opc_start = QUOTA_DQACQ,
4224 .tos_opc_end = QUOTA_LAST_OPC,
4225 .tos_hs = mdt_quota_ops
4228 .tos_opc_start = LLOG_FIRST_OPC,
4229 .tos_opc_end = LLOG_LAST_OPC,
4230 .tos_hs = tgt_llog_handlers
4233 .tos_opc_start = LFSCK_FIRST_OPC,
4234 .tos_opc_end = LFSCK_LAST_OPC,
4235 .tos_hs = tgt_lfsck_handlers
4243 static void mdt_fini(const struct lu_env *env, struct mdt_device *m)
4245 struct md_device *next = m->mdt_child;
4246 struct lu_device *d = &m->mdt_lu_dev;
4247 struct obd_device *obd = mdt2obd_dev(m);
4248 struct lfsck_stop stop;
4251 target_recovery_fini(obd);
4252 ping_evictor_stop();
4253 mdt_stack_pre_fini(env, m, md2lu_dev(m->mdt_child));
4255 if (m->mdt_opts.mo_coordinator)
4256 mdt_hsm_cdt_stop(m);
4258 mdt_hsm_cdt_fini(m);
4260 mdt_llog_ctxt_unclone(env, m, LLOG_AGENT_ORIG_CTXT);
4261 mdt_llog_ctxt_unclone(env, m, LLOG_CHANGELOG_ORIG_CTXT);
4262 obd_exports_barrier(obd);
4263 obd_zombie_barrier();
4267 tgt_fini(env, &m->mdt_lut);
4268 mdt_fs_cleanup(env, m);
4269 upcall_cache_cleanup(m->mdt_identity_cache);
4270 m->mdt_identity_cache = NULL;
4272 if (m->mdt_namespace != NULL) {
4273 ldlm_namespace_free(m->mdt_namespace, NULL,
4274 d->ld_obd->obd_force);
4275 d->ld_obd->obd_namespace = m->mdt_namespace = NULL;
4278 mdt_quota_fini(env, m);
4280 cfs_free_nidlist(&m->mdt_nosquash_nids);
4281 if (m->mdt_nosquash_str) {
4282 OBD_FREE(m->mdt_nosquash_str, m->mdt_nosquash_strlen);
4283 m->mdt_nosquash_str = NULL;
4284 m->mdt_nosquash_strlen = 0;
4287 stop.ls_status = LS_PAUSED;
4289 next->md_ops->mdo_iocontrol(env, next, OBD_IOC_STOP_LFSCK, 0, &stop);
4291 mdt_seq_fini(env, m);
4292 mdt_fld_fini(env, m);
4294 next->md_ops->mdo_init_capa_ctxt(env, next, 0, 0, 0, NULL);
4295 cfs_timer_disarm(&m->mdt_ck_timer);
4296 mdt_ck_thread_stop(m);
4301 mdt_stack_fini(env, m, md2lu_dev(m->mdt_child));
4303 LASSERT(atomic_read(&d->ld_ref) == 0);
4305 server_put_mount(mdt_obd_name(m));
4310 int mdt_postrecov(const struct lu_env *, struct mdt_device *);
4312 static int mdt_init0(const struct lu_env *env, struct mdt_device *m,
4313 struct lu_device_type *ldt, struct lustre_cfg *cfg)
4315 struct mdt_thread_info *info;
4316 struct obd_device *obd;
4317 const char *dev = lustre_cfg_string(cfg, 0);
4318 const char *num = lustre_cfg_string(cfg, 2);
4319 struct lustre_mount_info *lmi = NULL;
4320 struct lustre_sb_info *lsi;
4322 struct seq_server_site *ss_site;
4323 const char *identity_upcall = "NONE";
4324 struct md_device *next;
4330 lu_device_init(&m->mdt_lu_dev, ldt);
4332 * Environment (env) might be missing mdt_thread_key values at that
4333 * point, if device is allocated when mdt_thread_key is in QUIESCENT
4336 * Usually device allocation path doesn't use module key values, but
4337 * mdt has to do a lot of work here, so allocate key value.
4339 rc = lu_env_refill((struct lu_env *)env);
4343 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
4344 LASSERT(info != NULL);
4346 obd = class_name2obd(dev);
4347 LASSERT(obd != NULL);
4349 m->mdt_max_mdsize = MAX_MD_SIZE; /* 4 stripes */
4351 m->mdt_som_conf = 0;
4353 m->mdt_opts.mo_cos = MDT_COS_DEFAULT;
4355 /* default is coordinator off, it is started through conf_param
4357 m->mdt_opts.mo_coordinator = 0;
4359 lmi = server_get_mount(dev);
4361 CERROR("Cannot get mount info for %s!\n", dev);
4364 lsi = s2lsi(lmi->lmi_sb);
4365 /* CMD is supported only in IAM mode */
4367 node_id = simple_strtol(num, NULL, 10);
4368 obd->u.obt.obt_magic = OBT_MAGIC;
4371 spin_lock_init(&m->mdt_ioepoch_lock);
4372 m->mdt_capa_timeout = CAPA_TIMEOUT;
4373 m->mdt_capa_alg = CAPA_HMAC_ALG_SHA1;
4374 m->mdt_ck_timeout = CAPA_KEY_TIMEOUT;
4375 m->mdt_squash_uid = 0;
4376 m->mdt_squash_gid = 0;
4377 CFS_INIT_LIST_HEAD(&m->mdt_nosquash_nids);
4378 m->mdt_nosquash_str = NULL;
4379 m->mdt_nosquash_strlen = 0;
4380 init_rwsem(&m->mdt_squash_sem);
4381 spin_lock_init(&m->mdt_osfs_lock);
4382 m->mdt_osfs_age = cfs_time_shift_64(-1000);
4383 m->mdt_enable_remote_dir = 0;
4384 m->mdt_enable_remote_dir_gid = 0;
4386 m->mdt_lu_dev.ld_ops = &mdt_lu_ops;
4387 m->mdt_lu_dev.ld_obd = obd;
4388 /* Set this lu_device to obd for error handling purposes. */
4389 obd->obd_lu_dev = &m->mdt_lu_dev;
4391 /* init the stack */
4392 rc = mdt_stack_init((struct lu_env *)env, m, cfg);
4394 CERROR("%s: Can't init device stack, rc %d\n",
4395 mdt_obd_name(m), rc);
4400 ss_site = mdt_seq_site(m);
4401 s->ld_seq_site = ss_site;
4404 /* set server index */
4405 ss_site->ss_node_id = node_id;
4407 /* failover is the default
4408 * FIXME: we do not failout mds0/mgs, which may cause some problems.
4409 * assumed whose ss_node_id == 0 XXX
4411 obd->obd_replayable = 1;
4412 /* No connection accepted until configurations will finish */
4413 obd->obd_no_conn = 1;
4415 if (cfg->lcfg_bufcount > 4 && LUSTRE_CFG_BUFLEN(cfg, 4) > 0) {
4416 char *str = lustre_cfg_string(cfg, 4);
4417 if (strchr(str, 'n')) {
4418 CWARN("%s: recovery disabled\n", mdt_obd_name(m));
4419 obd->obd_replayable = 0;
4423 rc = mdt_fld_init(env, mdt_obd_name(m), m);
4425 GOTO(err_fini_stack, rc);
4427 rc = mdt_seq_init(env, m);
4429 GOTO(err_fini_fld, rc);
4431 snprintf(info->mti_u.ns_name, sizeof(info->mti_u.ns_name), "%s-%s",
4432 LUSTRE_MDT_NAME, obd->obd_uuid.uuid);
4433 m->mdt_namespace = ldlm_namespace_new(obd, info->mti_u.ns_name,
4434 LDLM_NAMESPACE_SERVER,
4435 LDLM_NAMESPACE_GREEDY,
4437 if (m->mdt_namespace == NULL)
4438 GOTO(err_fini_seq, rc = -ENOMEM);
4440 m->mdt_namespace->ns_lvbp = m;
4441 m->mdt_namespace->ns_lvbo = &mdt_lvbo;
4443 ldlm_register_intent(m->mdt_namespace, mdt_intent_policy);
4444 /* set obd_namespace for compatibility with old code */
4445 obd->obd_namespace = m->mdt_namespace;
4447 cfs_timer_init(&m->mdt_ck_timer, mdt_ck_timer_callback, m);
4449 rc = mdt_hsm_cdt_init(m);
4451 CERROR("%s: error initializing coordinator, rc %d\n",
4452 mdt_obd_name(m), rc);
4453 GOTO(err_free_ns, rc);
4456 rc = mdt_ck_thread_start(m);
4458 GOTO(err_free_hsm, rc);
4460 rc = tgt_init(env, &m->mdt_lut, obd, m->mdt_bottom, mdt_common_slice,
4461 OBD_FAIL_MDS_ALL_REQUEST_NET,
4462 OBD_FAIL_MDS_ALL_REPLY_NET);
4466 rc = mdt_fs_setup(env, m, obd, lsi);
4470 tgt_adapt_sptlrpc_conf(&m->mdt_lut, 1);
4472 next = m->mdt_child;
4473 rc = next->md_ops->mdo_iocontrol(env, next, OBD_IOC_GET_MNTOPT, 0,
4476 GOTO(err_fs_cleanup, rc);
4478 if (mntopts & MNTOPT_USERXATTR)
4479 m->mdt_opts.mo_user_xattr = 1;
4481 m->mdt_opts.mo_user_xattr = 0;
4483 rc = next->md_ops->mdo_maxeasize_get(env, next, &m->mdt_max_ea_size);
4485 GOTO(err_fs_cleanup, rc);
4487 if (mntopts & MNTOPT_ACL)
4488 m->mdt_opts.mo_acl = 1;
4490 m->mdt_opts.mo_acl = 0;
4492 /* XXX: to support suppgid for ACL, we enable identity_upcall
4493 * by default, otherwise, maybe got unexpected -EACCESS. */
4494 if (m->mdt_opts.mo_acl)
4495 identity_upcall = MDT_IDENTITY_UPCALL_PATH;
4497 m->mdt_identity_cache = upcall_cache_init(mdt_obd_name(m),
4499 &mdt_identity_upcall_cache_ops);
4500 if (IS_ERR(m->mdt_identity_cache)) {
4501 rc = PTR_ERR(m->mdt_identity_cache);
4502 m->mdt_identity_cache = NULL;
4503 GOTO(err_fs_cleanup, rc);
4506 rc = mdt_procfs_init(m, dev);
4508 CERROR("Can't init MDT lprocfs, rc %d\n", rc);
4509 GOTO(err_recovery, rc);
4512 rc = mdt_quota_init(env, m, cfg);
4514 GOTO(err_procfs, rc);
4516 m->mdt_ldlm_client = &mdt2obd_dev(m)->obd_ldlm_client;
4517 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
4518 "mdt_ldlm_client", m->mdt_ldlm_client);
4520 ping_evictor_start();
4522 /* recovery will be started upon mdt_prepare()
4523 * when the whole stack is complete and ready
4524 * to serve the requests */
4526 mdt_init_capa_ctxt(env, m);
4528 /* Reduce the initial timeout on an MDS because it doesn't need such
4529 * a long timeout as an OST does. Adaptive timeouts will adjust this
4530 * value appropriately. */
4531 if (ldlm_timeout == LDLM_TIMEOUT_DEFAULT)
4532 ldlm_timeout = MDS_LDLM_TIMEOUT_DEFAULT;
4538 target_recovery_fini(obd);
4539 upcall_cache_cleanup(m->mdt_identity_cache);
4540 m->mdt_identity_cache = NULL;
4542 mdt_fs_cleanup(env, m);
4544 tgt_fini(env, &m->mdt_lut);
4546 cfs_timer_disarm(&m->mdt_ck_timer);
4547 mdt_ck_thread_stop(m);
4549 mdt_hsm_cdt_fini(m);
4551 ldlm_namespace_free(m->mdt_namespace, NULL, 0);
4552 obd->obd_namespace = m->mdt_namespace = NULL;
4554 mdt_seq_fini(env, m);
4556 mdt_fld_fini(env, m);
4558 mdt_stack_fini(env, m, md2lu_dev(m->mdt_child));
4561 server_put_mount(dev);
4565 /* For interoperability, the left element is old parameter, the right one
4566 * is the new version of the parameter, if some parameter is deprecated,
4567 * the new version should be set as NULL. */
4568 static struct cfg_interop_param mdt_interop_param[] = {
4569 { "mdt.group_upcall", NULL },
4570 { "mdt.quota_type", NULL },
4571 { "mdd.quota_type", NULL },
4572 { "mdt.rootsquash", "mdt.root_squash" },
4573 { "mdt.nosquash_nid", "mdt.nosquash_nids" },
4577 /* used by MGS to process specific configurations */
4578 static int mdt_process_config(const struct lu_env *env,
4579 struct lu_device *d, struct lustre_cfg *cfg)
4581 struct mdt_device *m = mdt_dev(d);
4582 struct md_device *md_next = m->mdt_child;
4583 struct lu_device *next = md2lu_dev(md_next);
4587 switch (cfg->lcfg_command) {
4589 struct lprocfs_static_vars lvars;
4590 struct obd_device *obd = d->ld_obd;
4592 /* For interoperability */
4593 struct cfg_interop_param *ptr = NULL;
4594 struct lustre_cfg *old_cfg = NULL;
4597 param = lustre_cfg_string(cfg, 1);
4598 if (param == NULL) {
4599 CERROR("param is empty\n");
4604 ptr = class_find_old_param(param, mdt_interop_param);
4606 if (ptr->new_param == NULL) {
4608 CWARN("For interoperability, skip this %s."
4609 " It is obsolete.\n", ptr->old_param);
4613 CWARN("Found old param %s, changed it to %s.\n",
4614 ptr->old_param, ptr->new_param);
4617 cfg = lustre_cfg_rename(old_cfg, ptr->new_param);
4624 lprocfs_mdt_init_vars(&lvars);
4625 rc = class_process_proc_param(PARAM_MDT, lvars.obd_vars,
4627 if (rc > 0 || rc == -ENOSYS) {
4628 /* is it an HSM var ? */
4629 rc = class_process_proc_param(PARAM_HSM,
4630 hsm_cdt_get_proc_vars(),
4632 if (rc > 0 || rc == -ENOSYS)
4633 /* we don't understand; pass it on */
4634 rc = next->ld_ops->ldo_process_config(env, next,
4638 if (old_cfg != NULL)
4639 lustre_cfg_free(cfg);
4644 /* others are passed further */
4645 rc = next->ld_ops->ldo_process_config(env, next, cfg);
4651 static struct lu_object *mdt_object_alloc(const struct lu_env *env,
4652 const struct lu_object_header *hdr,
4653 struct lu_device *d)
4655 struct mdt_object *mo;
4659 OBD_SLAB_ALLOC_PTR_GFP(mo, mdt_object_kmem, GFP_NOFS);
4661 struct lu_object *o;
4662 struct lu_object_header *h;
4665 h = &mo->mot_header;
4666 lu_object_header_init(h);
4667 lu_object_init(o, h, d);
4668 lu_object_add_top(h, o);
4669 o->lo_ops = &mdt_obj_ops;
4670 mutex_init(&mo->mot_ioepoch_mutex);
4671 mutex_init(&mo->mot_lov_mutex);
4672 init_rwsem(&mo->mot_open_sem);
4678 static int mdt_object_init(const struct lu_env *env, struct lu_object *o,
4679 const struct lu_object_conf *unused)
4681 struct mdt_device *d = mdt_dev(o->lo_dev);
4682 struct lu_device *under;
4683 struct lu_object *below;
4687 CDEBUG(D_INFO, "object init, fid = "DFID"\n",
4688 PFID(lu_object_fid(o)));
4690 under = &d->mdt_child->md_lu_dev;
4691 below = under->ld_ops->ldo_object_alloc(env, o->lo_header, under);
4692 if (below != NULL) {
4693 lu_object_add(o, below);
4700 static void mdt_object_free(const struct lu_env *env, struct lu_object *o)
4702 struct mdt_object *mo = mdt_obj(o);
4703 struct lu_object_header *h;
4707 CDEBUG(D_INFO, "object free, fid = "DFID"\n",
4708 PFID(lu_object_fid(o)));
4710 LASSERT(atomic_read(&mo->mot_open_count) == 0);
4711 LASSERT(atomic_read(&mo->mot_lease_count) == 0);
4714 lu_object_header_fini(h);
4715 OBD_SLAB_FREE_PTR(mo, mdt_object_kmem);
4720 static int mdt_object_print(const struct lu_env *env, void *cookie,
4721 lu_printer_t p, const struct lu_object *o)
4723 struct mdt_object *mdto = mdt_obj((struct lu_object *)o);
4724 return (*p)(env, cookie, LUSTRE_MDT_NAME"-object@%p(ioepoch="LPU64" "
4725 "flags="LPX64", epochcount=%d, writecount=%d)",
4726 mdto, mdto->mot_ioepoch, mdto->mot_flags,
4727 mdto->mot_ioepoch_count, mdto->mot_writecount);
4730 static int mdt_prepare(const struct lu_env *env,
4731 struct lu_device *pdev,
4732 struct lu_device *cdev)
4734 struct mdt_device *mdt = mdt_dev(cdev);
4735 struct lu_device *next = &mdt->mdt_child->md_lu_dev;
4736 struct obd_device *obd = cdev->ld_obd;
4737 struct lfsck_start_param lsp;
4744 rc = next->ld_ops->ldo_prepare(env, cdev, next);
4748 rc = mdt_llog_ctxt_clone(env, mdt, LLOG_CHANGELOG_ORIG_CTXT);
4752 rc = mdt_llog_ctxt_clone(env, mdt, LLOG_AGENT_ORIG_CTXT);
4756 rc = lfsck_register_namespace(env, mdt->mdt_bottom, mdt->mdt_namespace);
4757 /* The LFSCK instance is registered just now, so it must be there when
4758 * register the namespace to such instance. */
4759 LASSERTF(rc == 0, "register namespace failed: rc = %d\n", rc);
4761 lsp.lsp_start = NULL;
4762 lsp.lsp_index_valid = 0;
4763 rc = mdt->mdt_child->md_ops->mdo_iocontrol(env, mdt->mdt_child,
4764 OBD_IOC_START_LFSCK,
4767 CWARN("%s: auto trigger paused LFSCK failed: rc = %d\n",
4768 mdt_obd_name(mdt), rc);
4772 if (mdt->mdt_seq_site.ss_node_id == 0) {
4773 rc = mdt->mdt_child->md_ops->mdo_root_get(env, mdt->mdt_child,
4774 &mdt->mdt_md_root_fid);
4779 LASSERT(!test_bit(MDT_FL_CFGLOG, &mdt->mdt_state));
4780 target_recovery_init(&mdt->mdt_lut, tgt_request_handle);
4781 set_bit(MDT_FL_CFGLOG, &mdt->mdt_state);
4782 LASSERT(obd->obd_no_conn);
4783 spin_lock(&obd->obd_dev_lock);
4784 obd->obd_no_conn = 0;
4785 spin_unlock(&obd->obd_dev_lock);
4787 if (obd->obd_recovering == 0)
4788 mdt_postrecov(env, mdt);
4793 const struct lu_device_operations mdt_lu_ops = {
4794 .ldo_object_alloc = mdt_object_alloc,
4795 .ldo_process_config = mdt_process_config,
4796 .ldo_prepare = mdt_prepare,
4799 static const struct lu_object_operations mdt_obj_ops = {
4800 .loo_object_init = mdt_object_init,
4801 .loo_object_free = mdt_object_free,
4802 .loo_object_print = mdt_object_print
4805 static int mdt_obd_set_info_async(const struct lu_env *env,
4806 struct obd_export *exp,
4807 __u32 keylen, void *key,
4808 __u32 vallen, void *val,
4809 struct ptlrpc_request_set *set)
4815 if (KEY_IS(KEY_SPTLRPC_CONF)) {
4816 rc = tgt_adapt_sptlrpc_conf(class_exp2tgt(exp), 0);
4824 * Match client and server connection feature flags.
4826 * Compute the compatibility flags for a connection request based on
4827 * features mutually supported by client and server.
4829 * The obd_export::exp_connect_data.ocd_connect_flags field in \a exp
4830 * must not be updated here, otherwise a partially initialized value may
4831 * be exposed. After the connection request is successfully processed,
4832 * the top-level MDT connect request handler atomically updates the export
4833 * connect flags from the obd_connect_data::ocd_connect_flags field of the
4834 * reply. \see mdt_connect().
4836 * \param exp the obd_export associated with this client/target pair
4837 * \param mdt the target device for the connection
4838 * \param data stores data for this connect request
4841 * \retval -EPROTO \a data unexpectedly has zero obd_connect_data::ocd_brw_size
4842 * \retval -EBADE client and server feature requirements are incompatible
4844 static int mdt_connect_internal(struct obd_export *exp,
4845 struct mdt_device *mdt,
4846 struct obd_connect_data *data)
4848 LASSERT(data != NULL);
4850 data->ocd_connect_flags &= MDT_CONNECT_SUPPORTED;
4851 data->ocd_ibits_known &= MDS_INODELOCK_FULL;
4853 /* If no known bits (which should not happen, probably,
4854 as everybody should support LOOKUP and UPDATE bits at least)
4855 revert to compat mode with plain locks. */
4856 if (!data->ocd_ibits_known &&
4857 data->ocd_connect_flags & OBD_CONNECT_IBITS)
4858 data->ocd_connect_flags &= ~OBD_CONNECT_IBITS;
4860 if (!mdt->mdt_opts.mo_acl)
4861 data->ocd_connect_flags &= ~OBD_CONNECT_ACL;
4863 if (!mdt->mdt_opts.mo_user_xattr)
4864 data->ocd_connect_flags &= ~OBD_CONNECT_XATTR;
4866 if (!mdt->mdt_som_conf)
4867 data->ocd_connect_flags &= ~OBD_CONNECT_SOM;
4869 if (data->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
4870 data->ocd_brw_size = min(data->ocd_brw_size,
4871 (__u32)MD_MAX_BRW_SIZE);
4872 if (data->ocd_brw_size == 0) {
4873 CERROR("%s: cli %s/%p ocd_connect_flags: "LPX64
4874 " ocd_version: %x ocd_grant: %d "
4875 "ocd_index: %u ocd_brw_size is "
4876 "unexpectedly zero, network data "
4877 "corruption? Refusing connection of this"
4880 exp->exp_client_uuid.uuid,
4881 exp, data->ocd_connect_flags, data->ocd_version,
4882 data->ocd_grant, data->ocd_index);
4887 /* NB: Disregard the rule against updating
4888 * exp_connect_data.ocd_connect_flags in this case, since
4889 * tgt_client_new() needs to know if this is a lightweight
4890 * connection, and it is safe to expose this flag before
4891 * connection processing completes. */
4892 if (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) {
4893 spin_lock(&exp->exp_lock);
4894 *exp_connect_flags_ptr(exp) |= OBD_CONNECT_LIGHTWEIGHT;
4895 spin_unlock(&exp->exp_lock);
4898 data->ocd_version = LUSTRE_VERSION_CODE;
4900 if ((data->ocd_connect_flags & OBD_CONNECT_FID) == 0) {
4901 CWARN("%s: MDS requires FID support, but client not\n",
4906 if (mdt->mdt_som_conf &&
4907 !(data->ocd_connect_flags & (OBD_CONNECT_LIGHTWEIGHT |
4908 OBD_CONNECT_MDS_MDS |
4909 OBD_CONNECT_SOM))) {
4910 CWARN("%s: MDS has SOM enabled, but client does not support "
4911 "it\n", mdt_obd_name(mdt));
4915 if (OCD_HAS_FLAG(data, PINGLESS)) {
4916 if (ptlrpc_pinger_suppress_pings()) {
4917 spin_lock(&exp->exp_obd->obd_dev_lock);
4918 list_del_init(&exp->exp_obd_chain_timed);
4919 spin_unlock(&exp->exp_obd->obd_dev_lock);
4921 data->ocd_connect_flags &= ~OBD_CONNECT_PINGLESS;
4925 data->ocd_max_easize = mdt->mdt_max_ea_size;
4930 /* mds_connect copy */
4931 static int mdt_obd_connect(const struct lu_env *env,
4932 struct obd_export **exp, struct obd_device *obd,
4933 struct obd_uuid *cluuid,
4934 struct obd_connect_data *data,
4937 struct obd_export *lexp;
4938 struct lustre_handle conn = { 0 };
4939 struct mdt_device *mdt;
4943 LASSERT(env != NULL);
4944 if (!exp || !obd || !cluuid)
4947 mdt = mdt_dev(obd->obd_lu_dev);
4950 * first, check whether the stack is ready to handle requests
4951 * XXX: probably not very appropriate method is used now
4952 * at some point we should find a better one
4954 if (!test_bit(MDT_FL_SYNCED, &mdt->mdt_state) && data != NULL &&
4955 !(data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT)) {
4956 rc = obd_get_info(env, mdt->mdt_child_exp,
4957 sizeof(KEY_OSP_CONNECTED),
4958 KEY_OSP_CONNECTED, NULL, NULL, NULL);
4961 set_bit(MDT_FL_SYNCED, &mdt->mdt_state);
4964 rc = class_connect(&conn, obd, cluuid);
4968 lexp = class_conn2export(&conn);
4969 LASSERT(lexp != NULL);
4971 rc = mdt_connect_internal(lexp, mdt, data);
4973 struct lsd_client_data *lcd = lexp->exp_target_data.ted_lcd;
4976 memcpy(lcd->lcd_uuid, cluuid, sizeof lcd->lcd_uuid);
4977 rc = tgt_client_new(env, lexp);
4979 mdt_export_stats_init(obd, lexp, localdata);
4981 /* For phase I, sync for cross-ref operation. */
4982 lexp->exp_keep_sync = 1;
4986 class_disconnect(lexp);
4995 static int mdt_obd_reconnect(const struct lu_env *env,
4996 struct obd_export *exp, struct obd_device *obd,
4997 struct obd_uuid *cluuid,
4998 struct obd_connect_data *data,
5004 if (exp == NULL || obd == NULL || cluuid == NULL)
5007 rc = mdt_connect_internal(exp, mdt_dev(obd->obd_lu_dev), data);
5009 mdt_export_stats_init(obd, exp, localdata);
5014 static int mdt_ctxt_add_dirty_flag(struct lu_env *env,
5015 struct mdt_thread_info *info,
5016 struct mdt_file_data *mfd)
5018 struct lu_context ses;
5022 rc = lu_context_init(&ses, LCT_SERVER_SESSION);
5027 lu_context_enter(&ses);
5029 mdt_ucred(info)->uc_valid = UCRED_OLD;
5030 rc = mdt_add_dirty_flag(info, mfd->mfd_object, &info->mti_attr);
5032 lu_context_exit(&ses);
5033 lu_context_fini(&ses);
5039 static int mdt_export_cleanup(struct obd_export *exp)
5041 struct mdt_export_data *med = &exp->exp_mdt_data;
5042 struct obd_device *obd = exp->exp_obd;
5043 struct mdt_device *mdt;
5044 struct mdt_thread_info *info;
5046 CFS_LIST_HEAD(closing_list);
5047 struct mdt_file_data *mfd, *n;
5051 spin_lock(&med->med_open_lock);
5052 while (!cfs_list_empty(&med->med_open_head)) {
5053 cfs_list_t *tmp = med->med_open_head.next;
5054 mfd = cfs_list_entry(tmp, struct mdt_file_data, mfd_list);
5056 /* Remove mfd handle so it can't be found again.
5057 * We are consuming the mfd_list reference here. */
5058 class_handle_unhash(&mfd->mfd_handle);
5059 cfs_list_move_tail(&mfd->mfd_list, &closing_list);
5061 spin_unlock(&med->med_open_lock);
5062 mdt = mdt_dev(obd->obd_lu_dev);
5063 LASSERT(mdt != NULL);
5065 rc = lu_env_init(&env, LCT_MD_THREAD);
5069 info = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
5070 LASSERT(info != NULL);
5071 memset(info, 0, sizeof *info);
5072 info->mti_env = &env;
5073 info->mti_mdt = mdt;
5074 info->mti_exp = exp;
5076 if (!cfs_list_empty(&closing_list)) {
5077 struct md_attr *ma = &info->mti_attr;
5079 /* Close any open files (which may also cause orphan unlinking). */
5080 cfs_list_for_each_entry_safe(mfd, n, &closing_list, mfd_list) {
5081 cfs_list_del_init(&mfd->mfd_list);
5082 ma->ma_need = ma->ma_valid = 0;
5084 /* This file is being closed due to an eviction, it
5085 * could have been modified and now dirty regarding to
5086 * HSM archive, check this!
5087 * The logic here is to mark a file dirty if there's a
5088 * chance it was dirtied before the client was evicted,
5089 * so that we don't have to wait for a release attempt
5090 * before finding out the file was actually dirty and
5091 * fail the release. Aggressively marking it dirty here
5092 * will cause the policy engine to attempt to
5093 * re-archive it; when rearchiving, we can compare the
5094 * current version to the HSM data_version and make the
5095 * archive request into a noop if it's not actually
5098 if (mfd->mfd_mode & (FMODE_WRITE|MDS_FMODE_TRUNC))
5099 rc = mdt_ctxt_add_dirty_flag(&env, info, mfd);
5101 /* Don't unlink orphan on failover umount, LU-184 */
5102 if (exp->exp_flags & OBD_OPT_FAILOVER) {
5103 ma->ma_valid = MA_FLAGS;
5104 ma->ma_attr_flags |= MDS_KEEP_ORPHAN;
5106 mdt_mfd_close(info, mfd);
5109 info->mti_mdt = NULL;
5110 /* cleanup client slot early */
5111 /* Do not erase record for recoverable client. */
5112 if (!(exp->exp_flags & OBD_OPT_FAILOVER) || exp->exp_failed)
5113 tgt_client_del(&env, exp);
5119 static int mdt_obd_disconnect(struct obd_export *exp)
5125 class_export_get(exp);
5127 rc = server_disconnect_export(exp);
5129 CDEBUG(D_IOCTL, "server disconnect error: %d\n", rc);
5131 rc = mdt_export_cleanup(exp);
5132 class_export_put(exp);
5136 /* FIXME: Can we avoid using these two interfaces? */
5137 static int mdt_init_export(struct obd_export *exp)
5139 struct mdt_export_data *med = &exp->exp_mdt_data;
5143 CFS_INIT_LIST_HEAD(&med->med_open_head);
5144 spin_lock_init(&med->med_open_lock);
5145 mutex_init(&med->med_idmap_mutex);
5146 med->med_idmap = NULL;
5147 spin_lock(&exp->exp_lock);
5148 exp->exp_connecting = 1;
5149 spin_unlock(&exp->exp_lock);
5151 /* self-export doesn't need client data and ldlm initialization */
5152 if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
5153 &exp->exp_client_uuid)))
5156 rc = tgt_client_alloc(exp);
5160 rc = ldlm_init_export(exp);
5167 tgt_client_free(exp);
5169 CERROR("%s: Failed to initialize export: rc = %d\n",
5170 exp->exp_obd->obd_name, rc);
5174 static int mdt_destroy_export(struct obd_export *exp)
5178 if (exp_connect_rmtclient(exp))
5179 mdt_cleanup_idmap(&exp->exp_mdt_data);
5181 target_destroy_export(exp);
5182 /* destroy can be called from failed obd_setup, so
5183 * checking uuid is safer than obd_self_export */
5184 if (unlikely(obd_uuid_equals(&exp->exp_obd->obd_uuid,
5185 &exp->exp_client_uuid)))
5188 ldlm_destroy_export(exp);
5189 tgt_client_free(exp);
5191 LASSERT(cfs_list_empty(&exp->exp_outstanding_replies));
5192 LASSERT(cfs_list_empty(&exp->exp_mdt_data.med_open_head));
5197 /** The maximum depth that fid2path() will search.
5198 * This is limited only because we want to store the fids for
5199 * historical path lookup purposes.
5201 #define MAX_PATH_DEPTH 100
5203 /** mdt_path() lookup structure. */
5204 struct path_lookup_info {
5205 __u64 pli_recno; /**< history point */
5206 __u64 pli_currec; /**< current record */
5207 struct lu_fid pli_fid;
5208 struct lu_fid pli_fids[MAX_PATH_DEPTH]; /**< path, in fids */
5209 struct mdt_object *pli_mdt_obj;
5210 char *pli_path; /**< full path */
5212 int pli_linkno; /**< which hardlink to follow */
5213 int pli_fidcount; /**< number of \a pli_fids */
5216 int mdt_links_read(struct mdt_thread_info *info, struct mdt_object *mdt_obj,
5217 struct linkea_data *ldata)
5221 LASSERT(ldata->ld_buf->lb_buf != NULL);
5223 if (!mdt_object_exists(mdt_obj))
5226 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
5227 ldata->ld_buf, XATTR_NAME_LINK);
5228 if (rc == -ERANGE) {
5229 /* Buf was too small, figure out what we need. */
5230 lu_buf_free(ldata->ld_buf);
5231 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
5232 ldata->ld_buf, XATTR_NAME_LINK);
5235 ldata->ld_buf = lu_buf_check_and_alloc(ldata->ld_buf, rc);
5236 if (ldata->ld_buf->lb_buf == NULL)
5238 rc = mo_xattr_get(info->mti_env, mdt_object_child(mdt_obj),
5239 ldata->ld_buf, XATTR_NAME_LINK);
5244 return linkea_init(ldata);
5247 static int mdt_path_current(struct mdt_thread_info *info,
5248 struct path_lookup_info *pli)
5250 struct mdt_device *mdt = info->mti_mdt;
5251 struct mdt_object *mdt_obj;
5252 struct link_ea_header *leh;
5253 struct link_ea_entry *lee;
5254 struct lu_name *tmpname = &info->mti_name;
5255 struct lu_fid *tmpfid = &info->mti_tmp_fid1;
5256 struct lu_buf *buf = &info->mti_big_buf;
5259 struct linkea_data ldata = { 0 };
5263 /* temp buffer for path element, the buffer will be finally freed
5264 * in mdt_thread_info_fini */
5265 buf = lu_buf_check_and_alloc(buf, PATH_MAX);
5266 if (buf->lb_buf == NULL)
5270 ptr = pli->pli_path + pli->pli_pathlen - 1;
5273 pli->pli_fidcount = 0;
5274 pli->pli_fids[0] = *(struct lu_fid *)mdt_object_fid(pli->pli_mdt_obj);
5276 /* root FID only exists on MDT0, and fid2path should also ends at MDT0,
5277 * so checking root_fid can only happen on MDT0. */
5278 while (!lu_fid_eq(&mdt->mdt_md_root_fid,
5279 &pli->pli_fids[pli->pli_fidcount])) {
5280 mdt_obj = mdt_object_find(info->mti_env, mdt,
5281 &pli->pli_fids[pli->pli_fidcount]);
5282 if (IS_ERR(mdt_obj))
5283 GOTO(out, rc = PTR_ERR(mdt_obj));
5284 if (mdt_object_remote(mdt_obj)) {
5285 mdt_object_put(info->mti_env, mdt_obj);
5286 GOTO(remote_out, rc = -EREMOTE);
5288 if (!mdt_object_exists(mdt_obj)) {
5289 mdt_object_put(info->mti_env, mdt_obj);
5290 GOTO(out, rc = -ENOENT);
5293 rc = mdt_links_read(info, mdt_obj, &ldata);
5294 mdt_object_put(info->mti_env, mdt_obj);
5299 lee = (struct link_ea_entry *)(leh + 1); /* link #0 */
5300 linkea_entry_unpack(lee, &reclen, tmpname, tmpfid);
5301 /* If set, use link #linkno for path lookup, otherwise use
5302 link #0. Only do this for the final path element. */
5303 if (pli->pli_fidcount == 0 &&
5304 pli->pli_linkno < leh->leh_reccount) {
5306 for (count = 0; count < pli->pli_linkno; count++) {
5307 lee = (struct link_ea_entry *)
5308 ((char *)lee + reclen);
5309 linkea_entry_unpack(lee, &reclen, tmpname,
5312 if (pli->pli_linkno < leh->leh_reccount - 1)
5313 /* indicate to user there are more links */
5317 /* Pack the name in the end of the buffer */
5318 ptr -= tmpname->ln_namelen;
5319 if (ptr - 1 <= pli->pli_path)
5320 GOTO(out, rc = -EOVERFLOW);
5321 strncpy(ptr, tmpname->ln_name, tmpname->ln_namelen);
5324 /* Store the parent fid for historic lookup */
5325 if (++pli->pli_fidcount >= MAX_PATH_DEPTH)
5326 GOTO(out, rc = -EOVERFLOW);
5327 pli->pli_fids[pli->pli_fidcount] = *tmpfid;
5331 ptr++; /* skip leading / */
5332 memmove(pli->pli_path, ptr, pli->pli_path + pli->pli_pathlen - ptr);
5339 /* Returns the full path to this fid, as of changelog record recno. */
5340 static int mdt_path(struct mdt_thread_info *info, struct mdt_object *obj,
5341 char *path, int pathlen, __u64 *recno, int *linkno,
5344 struct mdt_device *mdt = info->mti_mdt;
5345 struct path_lookup_info *pli;
5353 if (lu_fid_eq(&mdt->mdt_md_root_fid, mdt_object_fid(obj))) {
5362 pli->pli_mdt_obj = obj;
5363 pli->pli_recno = *recno;
5364 pli->pli_path = path;
5365 pli->pli_pathlen = pathlen;
5366 pli->pli_linkno = *linkno;
5368 /* Retry multiple times in case file is being moved */
5369 while (tries-- && rc == -EAGAIN)
5370 rc = mdt_path_current(info, pli);
5372 /* return the last resolved fids to the client, so the client will
5373 * build the left path on another MDT for remote object */
5374 *fid = pli->pli_fids[pli->pli_fidcount];
5376 *recno = pli->pli_currec;
5377 /* Return next link index to caller */
5378 *linkno = pli->pli_linkno;
5385 static int mdt_fid2path(struct mdt_thread_info *info,
5386 struct getinfo_fid2path *fp)
5388 struct mdt_device *mdt = info->mti_mdt;
5389 struct mdt_object *obj;
5393 CDEBUG(D_IOCTL, "path get "DFID" from "LPU64" #%d\n",
5394 PFID(&fp->gf_fid), fp->gf_recno, fp->gf_linkno);
5396 if (!fid_is_sane(&fp->gf_fid))
5399 if (!fid_is_namespace_visible(&fp->gf_fid)) {
5400 CWARN("%s: "DFID" is invalid, sequence should be "
5401 ">= "LPX64"\n", mdt_obd_name(mdt),
5402 PFID(&fp->gf_fid), (__u64)FID_SEQ_NORMAL);
5406 obj = mdt_object_find(info->mti_env, mdt, &fp->gf_fid);
5407 if (obj == NULL || IS_ERR(obj)) {
5408 CDEBUG(D_IOCTL, "no object "DFID": %ld\n", PFID(&fp->gf_fid),
5413 if (mdt_object_remote(obj))
5415 else if (!mdt_object_exists(obj))
5421 mdt_object_put(info->mti_env, obj);
5422 CDEBUG(D_IOCTL, "nonlocal object "DFID": %d\n",
5423 PFID(&fp->gf_fid), rc);
5427 rc = mdt_path(info, obj, fp->gf_path, fp->gf_pathlen, &fp->gf_recno,
5428 &fp->gf_linkno, &fp->gf_fid);
5430 CDEBUG(D_INFO, "fid "DFID", path %s recno "LPX64" linkno %u\n",
5431 PFID(&fp->gf_fid), fp->gf_path, fp->gf_recno, fp->gf_linkno);
5433 mdt_object_put(info->mti_env, obj);
5438 static int mdt_rpc_fid2path(struct mdt_thread_info *info, void *key,
5439 void *val, int vallen)
5441 struct getinfo_fid2path *fpout, *fpin;
5444 fpin = key + cfs_size_round(sizeof(KEY_FID2PATH));
5447 if (ptlrpc_req_need_swab(info->mti_pill->rc_req))
5448 lustre_swab_fid2path(fpin);
5450 memcpy(fpout, fpin, sizeof(*fpin));
5451 if (fpout->gf_pathlen != vallen - sizeof(*fpin))
5454 rc = mdt_fid2path(info, fpout);
5458 int mdt_get_info(struct tgt_session_info *tsi)
5468 key = req_capsule_client_get(tsi->tsi_pill, &RMF_GETINFO_KEY);
5470 CDEBUG(D_IOCTL, "No GETINFO key");
5471 RETURN(err_serious(-EFAULT));
5473 keylen = req_capsule_get_size(tsi->tsi_pill, &RMF_GETINFO_KEY,
5476 vallen = req_capsule_client_get(tsi->tsi_pill, &RMF_GETINFO_VALLEN);
5477 if (vallen == NULL) {
5478 CDEBUG(D_IOCTL, "Unable to get RMF_GETINFO_VALLEN buffer");
5479 RETURN(err_serious(-EFAULT));
5482 req_capsule_set_size(tsi->tsi_pill, &RMF_GETINFO_VAL, RCL_SERVER,
5484 rc = req_capsule_server_pack(tsi->tsi_pill);
5486 RETURN(err_serious(rc));
5488 valout = req_capsule_server_get(tsi->tsi_pill, &RMF_GETINFO_VAL);
5489 if (valout == NULL) {
5490 CDEBUG(D_IOCTL, "Unable to get get-info RPC out buffer");
5491 RETURN(err_serious(-EFAULT));
5494 if (KEY_IS(KEY_FID2PATH)) {
5495 struct mdt_thread_info *info = tsi2mdt_info(tsi);
5497 rc = mdt_rpc_fid2path(info, key, valout, *vallen);
5498 mdt_thread_info_fini(info);
5505 /* Pass the ioc down */
5506 static int mdt_ioc_child(struct lu_env *env, struct mdt_device *mdt,
5507 unsigned int cmd, int len, void *data)
5509 struct lu_context ioctl_session;
5510 struct md_device *next = mdt->mdt_child;
5514 rc = lu_context_init(&ioctl_session, LCT_SERVER_SESSION);
5517 ioctl_session.lc_thread = (struct ptlrpc_thread *)current;
5518 lu_context_enter(&ioctl_session);
5519 env->le_ses = &ioctl_session;
5521 LASSERT(next->md_ops->mdo_iocontrol);
5522 rc = next->md_ops->mdo_iocontrol(env, next, cmd, len, data);
5524 lu_context_exit(&ioctl_session);
5525 lu_context_fini(&ioctl_session);
5529 static int mdt_ioc_version_get(struct mdt_thread_info *mti, void *karg)
5531 struct obd_ioctl_data *data = karg;
5534 struct mdt_object *obj;
5535 struct mdt_lock_handle *lh;
5539 if (data->ioc_inlbuf1 == NULL || data->ioc_inllen1 != sizeof(*fid) ||
5540 data->ioc_inlbuf2 == NULL || data->ioc_inllen2 != sizeof(version))
5543 fid = (struct lu_fid *)data->ioc_inlbuf1;
5545 if (!fid_is_sane(fid))
5548 CDEBUG(D_IOCTL, "getting version for "DFID"\n", PFID(fid));
5550 lh = &mti->mti_lh[MDT_LH_PARENT];
5551 mdt_lock_reg_init(lh, LCK_CR);
5553 obj = mdt_object_find_lock(mti, fid, lh, MDS_INODELOCK_UPDATE);
5555 RETURN(PTR_ERR(obj));
5557 if (mdt_object_remote(obj)) {
5560 * before calling version get the correct MDS should be
5561 * fid, this is error to find remote object here
5563 CERROR("nonlocal object "DFID"\n", PFID(fid));
5564 } else if (!mdt_object_exists(obj)) {
5565 *(__u64 *)data->ioc_inlbuf2 = ENOENT_VERSION;
5568 version = dt_version_get(mti->mti_env, mdt_obj2dt(obj));
5569 *(__u64 *)data->ioc_inlbuf2 = version;
5572 mdt_object_unlock_put(mti, obj, lh, 1);
5576 /* ioctls on obd dev */
5577 static int mdt_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
5578 void *karg, void *uarg)
5581 struct obd_device *obd = exp->exp_obd;
5582 struct mdt_device *mdt = mdt_dev(obd->obd_lu_dev);
5583 struct dt_device *dt = mdt->mdt_bottom;
5587 CDEBUG(D_IOCTL, "handling ioctl cmd %#x\n", cmd);
5588 rc = lu_env_init(&env, LCT_MD_THREAD);
5594 rc = mdt_device_sync(&env, mdt);
5596 case OBD_IOC_SET_READONLY:
5597 rc = dt->dd_ops->dt_ro(&env, dt);
5599 case OBD_IOC_ABORT_RECOVERY:
5600 CERROR("%s: Aborting recovery for device\n", mdt_obd_name(mdt));
5601 target_stop_recovery_thread(obd);
5604 case OBD_IOC_CHANGELOG_REG:
5605 case OBD_IOC_CHANGELOG_DEREG:
5606 case OBD_IOC_CHANGELOG_CLEAR:
5607 rc = mdt_ioc_child(&env, mdt, cmd, len, karg);
5609 case OBD_IOC_START_LFSCK: {
5610 struct md_device *next = mdt->mdt_child;
5611 struct obd_ioctl_data *data = karg;
5612 struct lfsck_start_param lsp;
5614 if (unlikely(data == NULL)) {
5619 lsp.lsp_start = (struct lfsck_start *)(data->ioc_inlbuf1);
5620 lsp.lsp_index_valid = 0;
5621 rc = next->md_ops->mdo_iocontrol(&env, next, cmd, 0, &lsp);
5624 case OBD_IOC_STOP_LFSCK: {
5625 struct md_device *next = mdt->mdt_child;
5626 struct obd_ioctl_data *data = karg;
5627 struct lfsck_stop stop;
5629 stop.ls_status = LS_STOPPED;
5630 /* Old lfsck utils may pass NULL @stop. */
5631 if (data->ioc_inlbuf1 == NULL)
5635 ((struct lfsck_stop *)(data->ioc_inlbuf1))->ls_flags;
5637 rc = next->md_ops->mdo_iocontrol(&env, next, cmd, 0, &stop);
5640 case OBD_IOC_GET_OBJ_VERSION: {
5641 struct mdt_thread_info *mti;
5642 mti = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
5643 memset(mti, 0, sizeof *mti);
5644 mti->mti_env = &env;
5648 rc = mdt_ioc_version_get(mti, karg);
5651 case OBD_IOC_CATLOGLIST: {
5652 struct mdt_thread_info *mti;
5654 mti = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
5655 lu_local_obj_fid(&mti->mti_tmp_fid1, LLOG_CATALOGS_OID);
5656 rc = llog_catalog_list(&env, mdt->mdt_bottom, 0, karg,
5657 &mti->mti_tmp_fid1);
5662 CERROR("%s: Not supported cmd = %d, rc = %d\n",
5663 mdt_obd_name(mdt), cmd, rc);
5670 int mdt_postrecov(const struct lu_env *env, struct mdt_device *mdt)
5672 struct lu_device *ld = md2lu_dev(mdt->mdt_child);
5676 rc = ld->ld_ops->ldo_recovery_complete(env, ld);
5680 int mdt_obd_postrecov(struct obd_device *obd)
5685 rc = lu_env_init(&env, LCT_MD_THREAD);
5688 rc = mdt_postrecov(&env, mdt_dev(obd->obd_lu_dev));
5693 static struct obd_ops mdt_obd_device_ops = {
5694 .o_owner = THIS_MODULE,
5695 .o_set_info_async = mdt_obd_set_info_async,
5696 .o_connect = mdt_obd_connect,
5697 .o_reconnect = mdt_obd_reconnect,
5698 .o_disconnect = mdt_obd_disconnect,
5699 .o_init_export = mdt_init_export,
5700 .o_destroy_export = mdt_destroy_export,
5701 .o_iocontrol = mdt_iocontrol,
5702 .o_postrecov = mdt_obd_postrecov,
5705 static struct lu_device* mdt_device_fini(const struct lu_env *env,
5706 struct lu_device *d)
5708 struct mdt_device *m = mdt_dev(d);
5715 static struct lu_device *mdt_device_free(const struct lu_env *env,
5716 struct lu_device *d)
5718 struct mdt_device *m = mdt_dev(d);
5721 lu_device_fini(&m->mdt_lu_dev);
5727 static struct lu_device *mdt_device_alloc(const struct lu_env *env,
5728 struct lu_device_type *t,
5729 struct lustre_cfg *cfg)
5731 struct lu_device *l;
5732 struct mdt_device *m;
5739 rc = mdt_init0(env, m, t, cfg);
5741 mdt_device_free(env, l);
5746 l = ERR_PTR(-ENOMEM);
5750 /* context key constructor/destructor: mdt_key_init, mdt_key_fini */
5751 LU_KEY_INIT(mdt, struct mdt_thread_info);
5753 static void mdt_key_fini(const struct lu_context *ctx,
5754 struct lu_context_key *key, void* data)
5756 struct mdt_thread_info *info = data;
5758 if (info->mti_big_lmm) {
5759 OBD_FREE_LARGE(info->mti_big_lmm, info->mti_big_lmmsize);
5760 info->mti_big_lmm = NULL;
5761 info->mti_big_lmmsize = 0;
5766 /* context key: mdt_thread_key */
5767 LU_CONTEXT_KEY_DEFINE(mdt, LCT_MD_THREAD);
5769 struct lu_ucred *mdt_ucred(const struct mdt_thread_info *info)
5771 return lu_ucred(info->mti_env);
5774 struct lu_ucred *mdt_ucred_check(const struct mdt_thread_info *info)
5776 return lu_ucred_check(info->mti_env);
5780 * Enable/disable COS (Commit On Sharing).
5782 * Set/Clear the COS flag in mdt options.
5784 * \param mdt mdt device
5785 * \param val 0 disables COS, other values enable COS
5787 void mdt_enable_cos(struct mdt_device *mdt, int val)
5792 mdt->mdt_opts.mo_cos = !!val;
5793 rc = lu_env_init(&env, LCT_LOCAL);
5794 if (unlikely(rc != 0)) {
5795 CWARN("lu_env initialization failed with rc = %d,"
5796 "cannot sync\n", rc);
5799 mdt_device_sync(&env, mdt);
5804 * Check COS (Commit On Sharing) status.
5806 * Return COS flag status.
5808 * \param mdt mdt device
5810 int mdt_cos_is_enabled(struct mdt_device *mdt)
5812 return mdt->mdt_opts.mo_cos != 0;
5815 static struct lu_device_type_operations mdt_device_type_ops = {
5816 .ldto_device_alloc = mdt_device_alloc,
5817 .ldto_device_free = mdt_device_free,
5818 .ldto_device_fini = mdt_device_fini
5821 static struct lu_device_type mdt_device_type = {
5822 .ldt_tags = LU_DEVICE_MD,
5823 .ldt_name = LUSTRE_MDT_NAME,
5824 .ldt_ops = &mdt_device_type_ops,
5825 .ldt_ctx_tags = LCT_MD_THREAD
5828 static int __init mdt_mod_init(void)
5830 struct lprocfs_static_vars lvars;
5833 CLASSERT(sizeof("0x0123456789ABCDEF:0x01234567:0x01234567") ==
5834 FID_NOBRACE_LEN + 1);
5835 CLASSERT(sizeof("[0x0123456789ABCDEF:0x01234567:0x01234567]") ==
5837 rc = lu_kmem_init(mdt_caches);
5841 rc = mds_mod_init();
5845 lprocfs_mdt_init_vars(&lvars);
5846 rc = class_register_type(&mdt_obd_device_ops, NULL, NULL,
5847 #ifndef HAVE_ONLY_PROCFS_SEQ
5850 LUSTRE_MDT_NAME, &mdt_device_type);
5855 lu_kmem_fini(mdt_caches);
5862 static void __exit mdt_mod_exit(void)
5864 class_unregister_type(LUSTRE_MDT_NAME);
5866 lu_kmem_fini(mdt_caches);
5869 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
5870 MODULE_DESCRIPTION("Lustre Metadata Target ("LUSTRE_MDT_NAME")");
5871 MODULE_LICENSE("GPL");
5873 cfs_module(mdt, LUSTRE_VERSION_STRING, mdt_mod_init, mdt_mod_exit);