1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright 2008 Sun Microsystems, Inc. All rights reserved
30 * Use is subject to license terms.
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>
50 # define EXPORT_SYMTAB
52 #define DEBUG_SUBSYSTEM S_MDS
54 #include <linux/module.h>
56 * struct OBD_{ALLOC,FREE}*()
58 #include <obd_support.h>
59 /* struct ptlrpc_request */
60 #include <lustre_net.h>
61 /* struct obd_export */
62 #include <lustre_export.h>
63 /* struct obd_device */
66 #include <dt_object.h>
67 #include <lustre_mds.h>
68 #include <lustre_mdt.h>
69 #include "mdt_internal.h"
70 #ifdef HAVE_QUOTA_SUPPORT
71 # include <lustre_quota.h>
73 #include <lustre_acl.h>
74 #include <lustre_param.h>
75 #include <lustre_fsfilt.h>
77 mdl_mode_t mdt_mdl_lock_modes[] = {
78 [LCK_MINMODE] = MDL_MINMODE,
85 [LCK_GROUP] = MDL_GROUP
88 ldlm_mode_t mdt_dlm_lock_modes[] = {
89 [MDL_MINMODE] = LCK_MINMODE,
96 [MDL_GROUP] = LCK_GROUP
100 * Initialized in mdt_mod_init().
102 unsigned long mdt_num_threads;
104 /* ptlrpc request handler for MDT. All handlers are
105 * grouped into several slices - struct mdt_opc_slice,
106 * and stored in an array - mdt_handlers[].
109 /* The name of this handler. */
111 /* Fail id for this handler, checked at the beginning of this handler*/
113 /* Operation code for this handler */
115 /* flags are listed in enum mdt_handler_flags below. */
117 /* The actual handler function to execute. */
118 int (*mh_act)(struct mdt_thread_info *info);
119 /* Request format for this request. */
120 const struct req_format *mh_fmt;
123 enum mdt_handler_flags {
125 * struct mdt_body is passed in the incoming message, and object
126 * identified by this fid exists on disk.
128 * "habeo corpus" == "I have a body"
130 HABEO_CORPUS = (1 << 0),
132 * struct ldlm_request is passed in the incoming message.
134 * "habeo clavis" == "I have a key"
136 HABEO_CLAVIS = (1 << 1),
138 * this request has fixed reply format, so that reply message can be
139 * packed by generic code.
141 * "habeo refero" == "I have a reply"
143 HABEO_REFERO = (1 << 2),
145 * this request will modify something, so check whether the filesystem
146 * is readonly or not, then return -EROFS to client asap if necessary.
148 * "mutabor" == "I shall modify"
153 struct mdt_opc_slice {
156 struct mdt_handler *mos_hs;
159 static struct mdt_opc_slice mdt_regular_handlers[];
160 static struct mdt_opc_slice mdt_readpage_handlers[];
161 static struct mdt_opc_slice mdt_xmds_handlers[];
162 static struct mdt_opc_slice mdt_seq_handlers[];
163 static struct mdt_opc_slice mdt_fld_handlers[];
165 static struct mdt_device *mdt_dev(struct lu_device *d);
166 static int mdt_regular_handle(struct ptlrpc_request *req);
167 static int mdt_unpack_req_pack_rep(struct mdt_thread_info *info, __u32 flags);
168 static int mdt_fid2path(struct lu_env *env, struct mdt_device *mdt,
169 struct getinfo_fid2path *fp);
171 static const struct lu_object_operations mdt_obj_ops;
173 int mdt_get_disposition(struct ldlm_reply *rep, int flag)
177 return (rep->lock_policy_res1 & flag);
180 void mdt_clear_disposition(struct mdt_thread_info *info,
181 struct ldlm_reply *rep, int flag)
184 info->mti_opdata &= ~flag;
186 rep->lock_policy_res1 &= ~flag;
189 void mdt_set_disposition(struct mdt_thread_info *info,
190 struct ldlm_reply *rep, int flag)
193 info->mti_opdata |= flag;
195 rep->lock_policy_res1 |= flag;
198 void mdt_lock_reg_init(struct mdt_lock_handle *lh, ldlm_mode_t lm)
200 lh->mlh_pdo_hash = 0;
201 lh->mlh_reg_mode = lm;
202 lh->mlh_type = MDT_REG_LOCK;
205 void mdt_lock_pdo_init(struct mdt_lock_handle *lh, ldlm_mode_t lm,
206 const char *name, int namelen)
208 lh->mlh_reg_mode = lm;
209 lh->mlh_type = MDT_PDO_LOCK;
212 LASSERT(namelen > 0);
213 lh->mlh_pdo_hash = full_name_hash(name, namelen);
215 LASSERT(namelen == 0);
216 lh->mlh_pdo_hash = 0ull;
220 static void mdt_lock_pdo_mode(struct mdt_thread_info *info, struct mdt_object *o,
221 struct mdt_lock_handle *lh)
227 * Any dir access needs couple of locks:
229 * 1) on part of dir we gonna take lookup/modify;
231 * 2) on whole dir to protect it from concurrent splitting and/or to
232 * flush client's cache for readdir().
234 * so, for a given mode and object this routine decides what lock mode
235 * to use for lock #2:
237 * 1) if caller's gonna lookup in dir then we need to protect dir from
238 * being splitted only - LCK_CR
240 * 2) if caller's gonna modify dir then we need to protect dir from
241 * being splitted and to flush cache - LCK_CW
243 * 3) if caller's gonna modify dir and that dir seems ready for
244 * splitting then we need to protect it from any type of access
245 * (lookup/modify/split) - LCK_EX --bzzz
248 LASSERT(lh->mlh_reg_mode != LCK_MINMODE);
249 LASSERT(lh->mlh_pdo_mode == LCK_MINMODE);
252 * Ask underlaying level its opinion about preferable PDO lock mode
253 * having access type passed as regular lock mode:
255 * - MDL_MINMODE means that lower layer does not want to specify lock
258 * - MDL_NL means that no PDO lock should be taken. This is used in some
259 * cases. Say, for non-splittable directories no need to use PDO locks
262 mode = mdo_lock_mode(info->mti_env, mdt_object_child(o),
263 mdt_dlm_mode2mdl_mode(lh->mlh_reg_mode));
265 if (mode != MDL_MINMODE) {
266 lh->mlh_pdo_mode = mdt_mdl_mode2dlm_mode(mode);
269 * Lower layer does not want to specify locking mode. We do it
270 * our selves. No special protection is needed, just flush
271 * client's cache on modification and allow concurrent
274 switch (lh->mlh_reg_mode) {
276 lh->mlh_pdo_mode = LCK_EX;
279 lh->mlh_pdo_mode = LCK_CR;
282 lh->mlh_pdo_mode = LCK_CW;
285 CERROR("Not expected lock type (0x%x)\n",
286 (int)lh->mlh_reg_mode);
291 LASSERT(lh->mlh_pdo_mode != LCK_MINMODE);
295 static int mdt_getstatus(struct mdt_thread_info *info)
297 struct mdt_device *mdt = info->mti_mdt;
298 struct md_device *next = mdt->mdt_child;
299 struct mdt_body *repbody;
304 rc = mdt_check_ucred(info);
306 RETURN(err_serious(rc));
308 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK))
309 RETURN(err_serious(-ENOMEM));
311 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
312 rc = next->md_ops->mdo_root_get(info->mti_env, next, &repbody->fid1);
316 repbody->valid |= OBD_MD_FLID;
318 if (mdt->mdt_opts.mo_mds_capa &&
319 info->mti_exp->exp_connect_flags & OBD_CONNECT_MDS_CAPA) {
320 struct mdt_object *root;
321 struct lustre_capa *capa;
323 root = mdt_object_find(info->mti_env, mdt, &repbody->fid1);
325 RETURN(PTR_ERR(root));
327 capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA1);
329 capa->lc_opc = CAPA_OPC_MDS_DEFAULT;
330 rc = mo_capa_get(info->mti_env, mdt_object_child(root), capa,
332 mdt_object_put(info->mti_env, root);
334 repbody->valid |= OBD_MD_FLMDSCAPA;
340 static int mdt_statfs(struct mdt_thread_info *info)
342 struct md_device *next = info->mti_mdt->mdt_child;
343 struct ptlrpc_service *svc;
344 struct obd_statfs *osfs;
349 svc = info->mti_pill->rc_req->rq_rqbd->rqbd_service;
351 /* This will trigger a watchdog timeout */
352 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_STATFS_LCW_SLEEP,
353 (MDT_SERVICE_WATCHDOG_FACTOR *
354 at_get(&svc->srv_at_estimate)) + 1);
356 rc = mdt_check_ucred(info);
358 RETURN(err_serious(rc));
360 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
361 rc = err_serious(-ENOMEM);
363 osfs = req_capsule_server_get(info->mti_pill, &RMF_OBD_STATFS);
364 rc = next->md_ops->mdo_statfs(info->mti_env, next,
366 statfs_pack(osfs, &info->mti_u.ksfs);
371 void mdt_pack_size2body(struct mdt_thread_info *info, struct mdt_object *o)
374 struct lu_attr *attr = &info->mti_attr.ma_attr;
376 b = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
378 /* Check if Size-on-MDS is enabled. */
379 if ((mdt_conn_flags(info) & OBD_CONNECT_SOM) &&
380 S_ISREG(attr->la_mode) && mdt_sizeonmds_enabled(o)) {
381 b->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS);
382 b->size = attr->la_size;
383 b->blocks = attr->la_blocks;
387 void mdt_pack_attr2body(struct mdt_thread_info *info, struct mdt_body *b,
388 const struct lu_attr *attr, const struct lu_fid *fid)
390 /*XXX should pack the reply body according to lu_valid*/
391 b->valid |= OBD_MD_FLCTIME | OBD_MD_FLUID |
392 OBD_MD_FLGID | OBD_MD_FLTYPE |
393 OBD_MD_FLMODE | OBD_MD_FLNLINK | OBD_MD_FLFLAGS |
394 OBD_MD_FLATIME | OBD_MD_FLMTIME ;
396 if (!S_ISREG(attr->la_mode))
397 b->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS | OBD_MD_FLRDEV;
399 b->atime = attr->la_atime;
400 b->mtime = attr->la_mtime;
401 b->ctime = attr->la_ctime;
402 b->mode = attr->la_mode;
403 b->size = attr->la_size;
404 b->blocks = attr->la_blocks;
405 b->uid = attr->la_uid;
406 b->gid = attr->la_gid;
407 b->flags = attr->la_flags;
408 b->nlink = attr->la_nlink;
409 b->rdev = attr->la_rdev;
413 b->valid |= OBD_MD_FLID;
415 /* FIXME: these should be fixed when new igif ready.*/
416 b->ino = fid_oid(fid); /* 1.6 compatibility */
417 b->generation = fid_ver(fid); /* 1.6 compatibility */
418 b->valid |= OBD_MD_FLGENER; /* 1.6 compatibility */
420 CDEBUG(D_INODE, DFID": nlink=%d, mode=%o, size="LPU64"\n",
421 PFID(fid), b->nlink, b->mode, b->size);
425 mdt_body_reverse_idmap(info, b);
428 static inline int mdt_body_has_lov(const struct lu_attr *la,
429 const struct mdt_body *body)
431 return ((S_ISREG(la->la_mode) && (body->valid & OBD_MD_FLEASIZE)) ||
432 (S_ISDIR(la->la_mode) && (body->valid & OBD_MD_FLDIREA )) );
435 static int mdt_getattr_internal(struct mdt_thread_info *info,
436 struct mdt_object *o)
438 struct md_object *next = mdt_object_child(o);
439 const struct mdt_body *reqbody = info->mti_body;
440 struct ptlrpc_request *req = mdt_info_req(info);
441 struct md_attr *ma = &info->mti_attr;
442 struct lu_attr *la = &ma->ma_attr;
443 struct req_capsule *pill = info->mti_pill;
444 const struct lu_env *env = info->mti_env;
445 struct mdt_body *repbody;
446 struct lu_buf *buffer = &info->mti_buf;
450 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK))
451 RETURN(err_serious(-ENOMEM));
453 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
457 rc = mdt_object_exists(o);
459 /* This object is located on remote node.*/
460 repbody->fid1 = *mdt_object_fid(o);
461 repbody->valid = OBD_MD_FLID | OBD_MD_MDS;
465 buffer->lb_buf = req_capsule_server_get(pill, &RMF_MDT_MD);
466 buffer->lb_len = req_capsule_get_size(pill, &RMF_MDT_MD, RCL_SERVER);
468 /* If it is dir object and client require MEA, then we got MEA */
469 if (S_ISDIR(lu_object_attr(&next->mo_lu)) &&
470 reqbody->valid & OBD_MD_MEA) {
471 /* Assumption: MDT_MD size is enough for lmv size. */
472 ma->ma_lmv = buffer->lb_buf;
473 ma->ma_lmv_size = buffer->lb_len;
474 ma->ma_need = MA_LMV | MA_INODE;
476 ma->ma_lmm = buffer->lb_buf;
477 ma->ma_lmm_size = buffer->lb_len;
478 ma->ma_need = MA_LOV | MA_INODE;
481 if (S_ISDIR(lu_object_attr(&next->mo_lu)) &&
482 reqbody->valid & OBD_MD_FLDIREA &&
483 lustre_msg_get_opc(req->rq_reqmsg) == MDS_GETATTR) {
484 /* get default stripe info for this dir. */
485 ma->ma_need |= MA_LOV_DEF;
487 rc = mo_attr_get(env, next, ma);
489 CERROR("getattr error for "DFID": %d\n",
490 PFID(mdt_object_fid(o)), rc);
494 if (likely(ma->ma_valid & MA_INODE))
495 mdt_pack_attr2body(info, repbody, la, mdt_object_fid(o));
499 if (mdt_body_has_lov(la, reqbody)) {
500 if (ma->ma_valid & MA_LOV) {
501 LASSERT(ma->ma_lmm_size);
502 mdt_dump_lmm(D_INFO, ma->ma_lmm);
503 repbody->eadatasize = ma->ma_lmm_size;
504 if (S_ISDIR(la->la_mode))
505 repbody->valid |= OBD_MD_FLDIREA;
507 repbody->valid |= OBD_MD_FLEASIZE;
509 if (ma->ma_valid & MA_LMV) {
510 LASSERT(S_ISDIR(la->la_mode));
511 repbody->eadatasize = ma->ma_lmv_size;
512 repbody->valid |= (OBD_MD_FLDIREA|OBD_MD_MEA);
514 if (!(ma->ma_valid & MA_LOV) && !(ma->ma_valid & MA_LMV)) {
515 repbody->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
517 } else if (S_ISLNK(la->la_mode) &&
518 reqbody->valid & OBD_MD_LINKNAME) {
519 buffer->lb_buf = ma->ma_lmm;
520 buffer->lb_len = reqbody->eadatasize;
521 rc = mo_readlink(env, next, buffer);
522 if (unlikely(rc <= 0)) {
523 CERROR("readlink failed: %d\n", rc);
526 repbody->valid |= OBD_MD_LINKNAME;
527 repbody->eadatasize = rc;
529 ((char*)ma->ma_lmm)[rc - 1] = 0;
530 CDEBUG(D_INODE, "symlink dest %s, len = %d\n",
531 (char*)ma->ma_lmm, rc);
536 if (reqbody->valid & OBD_MD_FLMODEASIZE) {
537 repbody->max_cookiesize = info->mti_mdt->mdt_max_cookiesize;
538 repbody->max_mdsize = info->mti_mdt->mdt_max_mdsize;
539 repbody->valid |= OBD_MD_FLMODEASIZE;
540 CDEBUG(D_INODE, "I am going to change the MAX_MD_SIZE & "
541 "MAX_COOKIE to : %d:%d\n", repbody->max_mdsize,
542 repbody->max_cookiesize);
545 if (exp_connect_rmtclient(info->mti_exp) &&
546 reqbody->valid & OBD_MD_FLRMTPERM) {
547 void *buf = req_capsule_server_get(pill, &RMF_ACL);
549 /* mdt_getattr_lock only */
550 rc = mdt_pack_remote_perm(info, o, buf);
552 repbody->valid &= ~OBD_MD_FLRMTPERM;
553 repbody->aclsize = 0;
556 repbody->valid |= OBD_MD_FLRMTPERM;
557 repbody->aclsize = sizeof(struct mdt_remote_perm);
560 #ifdef CONFIG_FS_POSIX_ACL
561 else if ((req->rq_export->exp_connect_flags & OBD_CONNECT_ACL) &&
562 (reqbody->valid & OBD_MD_FLACL)) {
563 buffer->lb_buf = req_capsule_server_get(pill, &RMF_ACL);
564 buffer->lb_len = req_capsule_get_size(pill,
565 &RMF_ACL, RCL_SERVER);
566 if (buffer->lb_len > 0) {
567 rc = mo_xattr_get(env, next, buffer,
568 XATTR_NAME_ACL_ACCESS);
570 if (rc == -ENODATA) {
571 repbody->aclsize = 0;
572 repbody->valid |= OBD_MD_FLACL;
574 } else if (rc == -EOPNOTSUPP) {
577 CERROR("got acl size: %d\n", rc);
580 repbody->aclsize = rc;
581 repbody->valid |= OBD_MD_FLACL;
588 if (reqbody->valid & OBD_MD_FLMDSCAPA &&
589 info->mti_mdt->mdt_opts.mo_mds_capa &&
590 info->mti_exp->exp_connect_flags & OBD_CONNECT_MDS_CAPA) {
591 struct lustre_capa *capa;
593 capa = req_capsule_server_get(pill, &RMF_CAPA1);
595 capa->lc_opc = CAPA_OPC_MDS_DEFAULT;
596 rc = mo_capa_get(env, next, capa, 0);
599 repbody->valid |= OBD_MD_FLMDSCAPA;
604 static int mdt_renew_capa(struct mdt_thread_info *info)
606 struct mdt_object *obj = info->mti_object;
607 struct mdt_body *body;
608 struct lustre_capa *capa, *c;
612 /* if object doesn't exist, or server has disabled capability,
613 * return directly, client will find body->valid OBD_MD_FLOSSCAPA
616 if (!obj || !info->mti_mdt->mdt_opts.mo_oss_capa ||
617 !(info->mti_exp->exp_connect_flags & OBD_CONNECT_OSS_CAPA))
620 body = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
621 LASSERT(body != NULL);
623 c = req_capsule_client_get(info->mti_pill, &RMF_CAPA1);
626 capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA2);
630 rc = mo_capa_get(info->mti_env, mdt_object_child(obj), capa, 1);
632 body->valid |= OBD_MD_FLOSSCAPA;
636 static int mdt_getattr(struct mdt_thread_info *info)
638 struct mdt_object *obj = info->mti_object;
639 struct req_capsule *pill = info->mti_pill;
640 struct mdt_body *reqbody;
641 struct mdt_body *repbody;
647 reqbody = req_capsule_client_get(pill, &RMF_MDT_BODY);
650 if (reqbody->valid & OBD_MD_FLOSSCAPA) {
651 rc = req_capsule_server_pack(pill);
653 RETURN(err_serious(rc));
654 rc = mdt_renew_capa(info);
655 GOTO(out_shrink, rc);
658 LASSERT(obj != NULL);
659 LASSERT(lu_object_assert_exists(&obj->mot_obj.mo_lu));
661 mode = lu_object_attr(&obj->mot_obj.mo_lu);
662 if (S_ISLNK(mode) && (reqbody->valid & OBD_MD_LINKNAME) &&
663 (reqbody->eadatasize > info->mti_mdt->mdt_max_mdsize))
664 md_size = reqbody->eadatasize;
666 md_size = info->mti_mdt->mdt_max_mdsize;
668 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER, md_size);
670 rc = req_capsule_server_pack(pill);
671 if (unlikely(rc != 0))
672 RETURN(err_serious(rc));
674 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
675 LASSERT(repbody != NULL);
676 repbody->eadatasize = 0;
677 repbody->aclsize = 0;
679 if (reqbody->valid & OBD_MD_FLRMTPERM)
680 rc = mdt_init_ucred(info, reqbody);
682 rc = mdt_check_ucred(info);
684 GOTO(out_shrink, rc);
686 info->mti_spec.sp_ck_split = !!(reqbody->valid & OBD_MD_FLCKSPLIT);
687 info->mti_cross_ref = !!(reqbody->valid & OBD_MD_FLCROSSREF);
690 * Don't check capability at all, because rename might getattr for
691 * remote obj, and at that time no capability is available.
693 mdt_set_capainfo(info, 1, &reqbody->fid1, BYPASS_CAPA);
694 rc = mdt_getattr_internal(info, obj);
695 if (reqbody->valid & OBD_MD_FLRMTPERM)
696 mdt_exit_ucred(info);
699 mdt_shrink_reply(info);
703 static int mdt_is_subdir(struct mdt_thread_info *info)
705 struct mdt_object *o = info->mti_object;
706 struct req_capsule *pill = info->mti_pill;
707 const struct mdt_body *body = info->mti_body;
708 struct mdt_body *repbody;
714 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
717 * We save last checked parent fid to @repbody->fid1 for remote
720 LASSERT(fid_is_sane(&body->fid2));
721 LASSERT(mdt_object_exists(o) > 0);
722 rc = mdo_is_subdir(info->mti_env, mdt_object_child(o),
723 &body->fid2, &repbody->fid1);
724 if (rc == 0 || rc == -EREMOTE)
725 repbody->valid |= OBD_MD_FLID;
730 static int mdt_raw_lookup(struct mdt_thread_info *info,
731 struct mdt_object *parent,
732 const struct lu_name *lname,
733 struct ldlm_reply *ldlm_rep)
735 struct md_object *next = mdt_object_child(info->mti_object);
736 const struct mdt_body *reqbody = info->mti_body;
737 struct lu_fid *child_fid = &info->mti_tmp_fid1;
738 struct mdt_body *repbody;
742 if (reqbody->valid != OBD_MD_FLID)
745 LASSERT(!info->mti_cross_ref);
747 /* Only got the fid of this obj by name */
748 rc = mdo_lookup(info->mti_env, next, lname, child_fid,
751 /* XXX is raw_lookup possible as intent operation? */
754 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
757 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
759 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
762 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
763 repbody->fid1 = *child_fid;
764 repbody->valid = OBD_MD_FLID;
770 * UPDATE lock should be taken against parent, and be release before exit;
771 * child_bits lock should be taken against child, and be returned back:
772 * (1)normal request should release the child lock;
773 * (2)intent request will grant the lock to client.
775 static int mdt_getattr_name_lock(struct mdt_thread_info *info,
776 struct mdt_lock_handle *lhc,
778 struct ldlm_reply *ldlm_rep)
780 struct ptlrpc_request *req = mdt_info_req(info);
781 struct mdt_body *reqbody = NULL;
782 struct mdt_object *parent = info->mti_object;
783 struct mdt_object *child;
784 struct md_object *next = mdt_object_child(parent);
785 struct lu_fid *child_fid = &info->mti_tmp_fid1;
786 struct lu_name *lname = NULL;
787 const char *name = NULL;
789 struct mdt_lock_handle *lhp;
790 struct ldlm_lock *lock;
791 struct ldlm_res_id *res_id;
797 is_resent = lustre_handle_is_used(&lhc->mlh_reg_lh);
798 LASSERT(ergo(is_resent,
799 lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT));
801 LASSERT(parent != NULL);
802 name = req_capsule_client_get(info->mti_pill, &RMF_NAME);
804 RETURN(err_serious(-EFAULT));
806 namelen = req_capsule_get_size(info->mti_pill, &RMF_NAME,
808 if (!info->mti_cross_ref) {
810 * XXX: Check for "namelen == 0" is for getattr by fid
811 * (OBD_CONNECT_ATTRFID), otherwise do not allow empty name,
812 * that is the name must contain at least one character and
813 * the terminating '\0'
816 reqbody = req_capsule_client_get(info->mti_pill,
818 LASSERT(fid_is_sane(&reqbody->fid2));
821 CDEBUG(D_INODE, "getattr with lock for "DFID"/"DFID", "
823 PFID(mdt_object_fid(parent)), PFID(&reqbody->fid2),
826 lname = mdt_name(info->mti_env, (char *)name, namelen);
827 CDEBUG(D_INODE, "getattr with lock for "DFID"/%s, "
828 "ldlm_rep = %p\n", PFID(mdt_object_fid(parent)),
832 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_EXECD);
834 rc = mdt_object_exists(parent);
835 if (unlikely(rc == 0)) {
836 LU_OBJECT_DEBUG(D_WARNING, info->mti_env,
837 &parent->mot_obj.mo_lu,
838 "Parent doesn't exist!\n");
840 } else if (!info->mti_cross_ref) {
841 LASSERTF(rc > 0, "Parent "DFID" is on remote server\n",
842 PFID(mdt_object_fid(parent)));
845 rc = mdt_raw_lookup(info, parent, lname, ldlm_rep);
853 if (info->mti_cross_ref) {
854 /* Only getattr on the child. Parent is on another node. */
855 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
857 CDEBUG(D_INODE, "partial getattr_name child_fid = "DFID", "
858 "ldlm_rep=%p\n", PFID(mdt_object_fid(child)), ldlm_rep);
861 /* Do not take lock for resent case. */
862 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
863 LASSERTF(lock != NULL, "Invalid lock handle "LPX64"\n",
864 lhc->mlh_reg_lh.cookie);
865 LASSERT(fid_res_name_eq(mdt_object_fid(child),
866 &lock->l_resource->lr_name));
870 mdt_lock_handle_init(lhc);
871 mdt_lock_reg_init(lhc, LCK_PR);
874 * Object's name is on another MDS, no lookup lock is
875 * needed here but update is.
877 child_bits &= ~MDS_INODELOCK_LOOKUP;
878 child_bits |= MDS_INODELOCK_UPDATE;
880 rc = mdt_object_lock(info, child, lhc, child_bits,
884 /* Finally, we can get attr for child. */
885 mdt_set_capainfo(info, 0, mdt_object_fid(child),
887 rc = mdt_getattr_internal(info, child);
888 if (unlikely(rc != 0))
889 mdt_object_unlock(info, child, lhc, 1);
894 /* step 1: lock parent */
895 lhp = &info->mti_lh[MDT_LH_PARENT];
896 mdt_lock_pdo_init(lhp, LCK_PR, name, namelen);
897 rc = mdt_object_lock(info, parent, lhp, MDS_INODELOCK_UPDATE,
900 if (unlikely(rc != 0))
904 /* step 2: lookup child's fid by name */
905 rc = mdo_lookup(info->mti_env, next, lname, child_fid,
910 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
911 GOTO(out_parent, rc);
913 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
915 *child_fid = reqbody->fid2;
916 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
920 *step 3: find the child object by fid & lock it.
921 * regardless if it is local or remote.
923 child = mdt_object_find(info->mti_env, info->mti_mdt, child_fid);
925 if (unlikely(IS_ERR(child)))
926 GOTO(out_parent, rc = PTR_ERR(child));
928 /* Do not take lock for resent case. */
929 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
930 LASSERTF(lock != NULL, "Invalid lock handle "LPX64"\n",
931 lhc->mlh_reg_lh.cookie);
933 res_id = &lock->l_resource->lr_name;
934 if (!fid_res_name_eq(mdt_object_fid(child),
935 &lock->l_resource->lr_name)) {
936 LASSERTF(fid_res_name_eq(mdt_object_fid(parent),
937 &lock->l_resource->lr_name),
938 "Lock res_id: %lu/%lu/%lu, Fid: "DFID".\n",
939 (unsigned long)res_id->name[0],
940 (unsigned long)res_id->name[1],
941 (unsigned long)res_id->name[2],
942 PFID(mdt_object_fid(parent)));
943 CWARN("Although resent, but still not get child lock"
944 "parent:"DFID" child:"DFID"\n",
945 PFID(mdt_object_fid(parent)),
946 PFID(mdt_object_fid(child)));
947 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
956 ma = &info->mti_attr;
958 mdt_lock_handle_init(lhc);
959 mdt_lock_reg_init(lhc, LCK_PR);
961 if (mdt_object_exists(child) == 0) {
962 LU_OBJECT_DEBUG(D_WARNING, info->mti_env,
963 &child->mot_obj.mo_lu,
964 "Object doesn't exist!\n");
965 GOTO(out_child, rc = -ESTALE);
969 ma->ma_need = MA_INODE;
970 rc = mo_attr_get(info->mti_env, next, ma);
971 if (unlikely(rc != 0))
974 /* If the file has not been changed for some time, we return
975 * not only a LOOKUP lock, but also an UPDATE lock and this
976 * might save us RPC on later STAT. For directories, it also
977 * let negative dentry starts working for this dir. */
978 if (ma->ma_valid & MA_INODE &&
979 ma->ma_attr.la_valid & LA_CTIME &&
980 info->mti_mdt->mdt_namespace->ns_ctime_age_limit +
981 ma->ma_attr.la_ctime < cfs_time_current_sec())
982 child_bits |= MDS_INODELOCK_UPDATE;
984 rc = mdt_object_lock(info, child, lhc, child_bits,
987 if (unlikely(rc != 0))
991 /* finally, we can get attr for child. */
992 mdt_set_capainfo(info, 1, child_fid, BYPASS_CAPA);
993 rc = mdt_getattr_internal(info, child);
994 if (unlikely(rc != 0)) {
995 mdt_object_unlock(info, child, lhc, 1);
997 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
999 struct mdt_body *repbody;
1001 /* Debugging code. */
1002 res_id = &lock->l_resource->lr_name;
1003 LDLM_DEBUG(lock, "Returning lock to client\n");
1004 LASSERTF(fid_res_name_eq(mdt_object_fid(child),
1005 &lock->l_resource->lr_name),
1006 "Lock res_id: %lu/%lu/%lu, Fid: "DFID".\n",
1007 (unsigned long)res_id->name[0],
1008 (unsigned long)res_id->name[1],
1009 (unsigned long)res_id->name[2],
1010 PFID(mdt_object_fid(child)));
1012 * Pack Size-on-MDS inode attributes to the body if
1013 * update lock is given.
1015 repbody = req_capsule_server_get(info->mti_pill,
1017 if (lock->l_policy_data.l_inodebits.bits &
1018 MDS_INODELOCK_UPDATE)
1019 mdt_pack_size2body(info, child);
1020 LDLM_LOCK_PUT(lock);
1025 mdt_object_put(info->mti_env, child);
1027 mdt_object_unlock(info, parent, lhp, 1);
1031 /* normal handler: should release the child lock */
1032 static int mdt_getattr_name(struct mdt_thread_info *info)
1034 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_CHILD];
1035 struct mdt_body *reqbody;
1036 struct mdt_body *repbody;
1040 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
1041 LASSERT(reqbody != NULL);
1042 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1043 LASSERT(repbody != NULL);
1045 info->mti_spec.sp_ck_split = !!(reqbody->valid & OBD_MD_FLCKSPLIT);
1046 info->mti_cross_ref = !!(reqbody->valid & OBD_MD_FLCROSSREF);
1047 repbody->eadatasize = 0;
1048 repbody->aclsize = 0;
1050 rc = mdt_init_ucred(info, reqbody);
1052 GOTO(out_shrink, rc);
1054 rc = mdt_getattr_name_lock(info, lhc, MDS_INODELOCK_UPDATE, NULL);
1055 if (lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1056 ldlm_lock_decref(&lhc->mlh_reg_lh, lhc->mlh_reg_mode);
1057 lhc->mlh_reg_lh.cookie = 0;
1059 mdt_exit_ucred(info);
1062 mdt_shrink_reply(info);
1066 static const struct lu_device_operations mdt_lu_ops;
1068 static int lu_device_is_mdt(struct lu_device *d)
1070 return ergo(d != NULL && d->ld_ops != NULL, d->ld_ops == &mdt_lu_ops);
1073 static int mdt_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
1074 void *karg, void *uarg);
1076 static int mdt_set_info(struct mdt_thread_info *info)
1078 struct ptlrpc_request *req = mdt_info_req(info);
1081 int keylen, vallen, rc = 0;
1084 rc = req_capsule_server_pack(info->mti_pill);
1088 key = req_capsule_client_get(info->mti_pill, &RMF_SETINFO_KEY);
1090 DEBUG_REQ(D_HA, req, "no set_info key");
1094 keylen = req_capsule_get_size(info->mti_pill, &RMF_SETINFO_KEY,
1097 val = req_capsule_client_get(info->mti_pill, &RMF_SETINFO_VAL);
1099 DEBUG_REQ(D_HA, req, "no set_info val");
1103 vallen = req_capsule_get_size(info->mti_pill, &RMF_SETINFO_VAL,
1106 if (KEY_IS(KEY_READ_ONLY)) {
1108 lustre_msg_set_status(req->rq_repmsg, 0);
1110 spin_lock(&req->rq_export->exp_lock);
1112 req->rq_export->exp_connect_flags |= OBD_CONNECT_RDONLY;
1114 req->rq_export->exp_connect_flags &=~OBD_CONNECT_RDONLY;
1115 spin_unlock(&req->rq_export->exp_lock);
1117 } else if (KEY_IS(KEY_CHANGELOG_CLEAR)) {
1118 struct changelog_setinfo *cs =
1119 (struct changelog_setinfo *)val;
1120 if (vallen != sizeof(*cs)) {
1121 CERROR("Bad changelog_clear setinfo size %d\n", vallen);
1124 if (lustre_msg_swabbed(req->rq_reqmsg)) {
1125 __swab64s(&cs->cs_recno);
1126 __swab32s(&cs->cs_id);
1129 rc = mdt_iocontrol(OBD_IOC_CHANGELOG_CLEAR, info->mti_exp,
1131 lustre_msg_set_status(req->rq_repmsg, rc);
1139 static int mdt_connect(struct mdt_thread_info *info)
1142 struct ptlrpc_request *req;
1144 req = mdt_info_req(info);
1145 rc = target_handle_connect(req);
1147 LASSERT(req->rq_export != NULL);
1148 info->mti_mdt = mdt_dev(req->rq_export->exp_obd->obd_lu_dev);
1149 rc = mdt_init_sec_level(info);
1151 rc = mdt_init_idmap(info);
1153 obd_disconnect(class_export_get(req->rq_export));
1155 rc = err_serious(rc);
1160 static int mdt_disconnect(struct mdt_thread_info *info)
1165 rc = target_handle_disconnect(mdt_info_req(info));
1167 rc = err_serious(rc);
1171 static int mdt_sendpage(struct mdt_thread_info *info,
1172 struct lu_rdpg *rdpg)
1174 struct ptlrpc_request *req = mdt_info_req(info);
1175 struct ptlrpc_bulk_desc *desc;
1176 struct l_wait_info *lwi = &info->mti_u.rdpg.mti_wait_info;
1184 desc = ptlrpc_prep_bulk_exp(req, rdpg->rp_npages, BULK_PUT_SOURCE,
1189 for (i = 0, tmpcount = rdpg->rp_count;
1190 i < rdpg->rp_npages; i++, tmpcount -= tmpsize) {
1191 tmpsize = min_t(int, tmpcount, CFS_PAGE_SIZE);
1192 ptlrpc_prep_bulk_page(desc, rdpg->rp_pages[i], 0, tmpsize);
1195 LASSERT(desc->bd_nob == rdpg->rp_count);
1196 rc = sptlrpc_svc_wrap_bulk(req, desc);
1198 GOTO(free_desc, rc);
1200 rc = ptlrpc_start_bulk_transfer(desc);
1202 GOTO(free_desc, rc);
1204 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE))
1205 GOTO(abort_bulk, rc = 0);
1207 timeout = (int) req->rq_deadline - cfs_time_current_sec();
1209 CERROR("Req deadline already passed %lu (now: %lu)\n",
1210 req->rq_deadline, cfs_time_current_sec());
1211 *lwi = LWI_TIMEOUT(cfs_time_seconds(max(timeout, 1)), NULL, NULL);
1212 rc = l_wait_event(desc->bd_waitq, !ptlrpc_server_bulk_active(desc), lwi);
1213 LASSERT (rc == 0 || rc == -ETIMEDOUT);
1216 if (desc->bd_success &&
1217 desc->bd_nob_transferred == rdpg->rp_count)
1218 GOTO(free_desc, rc);
1220 rc = -ETIMEDOUT; /* XXX should this be a different errno? */
1223 DEBUG_REQ(D_ERROR, req, "bulk failed: %s %d(%d), evicting %s@%s",
1224 (rc == -ETIMEDOUT) ? "timeout" : "network error",
1225 desc->bd_nob_transferred, rdpg->rp_count,
1226 req->rq_export->exp_client_uuid.uuid,
1227 req->rq_export->exp_connection->c_remote_uuid.uuid);
1229 class_fail_export(req->rq_export);
1233 ptlrpc_abort_bulk(desc);
1235 ptlrpc_free_bulk(desc);
1239 #ifdef HAVE_SPLIT_SUPPORT
1241 * Retrieve dir entry from the page and insert it to the slave object, actually,
1242 * this should be in osd layer, but since it will not in the final product, so
1243 * just do it here and do not define more moo api anymore for this.
1245 static int mdt_write_dir_page(struct mdt_thread_info *info, struct page *page,
1248 struct mdt_object *object = info->mti_object;
1249 struct lu_fid *lf = &info->mti_tmp_fid2;
1250 struct md_attr *ma = &info->mti_attr;
1251 struct lu_dirpage *dp;
1252 struct lu_dirent *ent;
1253 int rc = 0, offset = 0;
1256 /* Make sure we have at least one entry. */
1261 * Disable trans for this name insert, since it will include many trans
1264 info->mti_no_need_trans = 1;
1266 * When write_dir_page, no need update parent's ctime,
1267 * and no permission check for name_insert.
1269 ma->ma_attr.la_ctime = 0;
1270 ma->ma_attr.la_valid = LA_MODE;
1271 ma->ma_valid = MA_INODE;
1274 dp = page_address(page);
1275 offset = (int)((__u32)lu_dirent_start(dp) - (__u32)dp);
1277 for (ent = lu_dirent_start(dp); ent != NULL;
1278 ent = lu_dirent_next(ent)) {
1279 struct lu_name *lname;
1282 if (le16_to_cpu(ent->lde_namelen) == 0)
1285 fid_le_to_cpu(lf, &ent->lde_fid);
1286 if (le64_to_cpu(ent->lde_hash) & MAX_HASH_HIGHEST_BIT)
1287 ma->ma_attr.la_mode = S_IFDIR;
1289 ma->ma_attr.la_mode = 0;
1290 OBD_ALLOC(name, le16_to_cpu(ent->lde_namelen) + 1);
1292 GOTO(out, rc = -ENOMEM);
1294 memcpy(name, ent->lde_name, le16_to_cpu(ent->lde_namelen));
1295 lname = mdt_name(info->mti_env, name,
1296 le16_to_cpu(ent->lde_namelen));
1297 ma->ma_attr_flags |= (MDS_PERM_BYPASS | MDS_QUOTA_IGNORE);
1298 rc = mdo_name_insert(info->mti_env,
1299 md_object_next(&object->mot_obj),
1301 OBD_FREE(name, le16_to_cpu(ent->lde_namelen) + 1);
1303 CERROR("Can't insert %*.*s, rc %d\n",
1304 le16_to_cpu(ent->lde_namelen),
1305 le16_to_cpu(ent->lde_namelen),
1310 offset += lu_dirent_size(ent);
1320 static int mdt_bulk_timeout(void *data)
1324 CERROR("mdt bulk transfer timeout \n");
1329 static int mdt_writepage(struct mdt_thread_info *info)
1331 struct ptlrpc_request *req = mdt_info_req(info);
1332 struct mdt_body *reqbody;
1333 struct l_wait_info *lwi;
1334 struct ptlrpc_bulk_desc *desc;
1340 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
1341 if (reqbody == NULL)
1342 RETURN(err_serious(-EFAULT));
1344 desc = ptlrpc_prep_bulk_exp(req, 1, BULK_GET_SINK, MDS_BULK_PORTAL);
1346 RETURN(err_serious(-ENOMEM));
1348 /* allocate the page for the desc */
1349 page = cfs_alloc_page(CFS_ALLOC_STD);
1351 GOTO(desc_cleanup, rc = -ENOMEM);
1353 CDEBUG(D_INFO, "Received page offset %d size %d \n",
1354 (int)reqbody->size, (int)reqbody->nlink);
1356 ptlrpc_prep_bulk_page(desc, page, (int)reqbody->size,
1357 (int)reqbody->nlink);
1359 rc = sptlrpc_svc_prep_bulk(req, desc);
1361 GOTO(cleanup_page, rc);
1363 * Check if client was evicted while we were doing i/o before touching
1368 GOTO(cleanup_page, rc = -ENOMEM);
1370 if (desc->bd_export->exp_failed)
1373 rc = ptlrpc_start_bulk_transfer (desc);
1375 *lwi = LWI_TIMEOUT_INTERVAL(obd_timeout * HZ / 4, HZ,
1376 mdt_bulk_timeout, desc);
1377 rc = l_wait_event(desc->bd_waitq, !ptlrpc_bulk_active(desc) ||
1378 desc->bd_export->exp_failed, lwi);
1379 LASSERT(rc == 0 || rc == -ETIMEDOUT);
1380 if (rc == -ETIMEDOUT) {
1381 DEBUG_REQ(D_ERROR, req, "timeout on bulk GET");
1382 ptlrpc_abort_bulk(desc);
1383 } else if (desc->bd_export->exp_failed) {
1384 DEBUG_REQ(D_ERROR, req, "Eviction on bulk GET");
1386 ptlrpc_abort_bulk(desc);
1387 } else if (!desc->bd_success ||
1388 desc->bd_nob_transferred != desc->bd_nob) {
1389 DEBUG_REQ(D_ERROR, req, "%s bulk GET %d(%d)",
1391 "truncated" : "network error on",
1392 desc->bd_nob_transferred, desc->bd_nob);
1393 /* XXX should this be a different errno? */
1397 DEBUG_REQ(D_ERROR, req, "ptlrpc_bulk_get failed: rc %d", rc);
1400 GOTO(cleanup_lwi, rc);
1401 rc = mdt_write_dir_page(info, page, reqbody->nlink);
1406 cfs_free_page(page);
1408 ptlrpc_free_bulk(desc);
1413 static int mdt_readpage(struct mdt_thread_info *info)
1415 struct mdt_object *object = info->mti_object;
1416 struct lu_rdpg *rdpg = &info->mti_u.rdpg.mti_rdpg;
1417 struct mdt_body *reqbody;
1418 struct mdt_body *repbody;
1423 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK))
1424 RETURN(err_serious(-ENOMEM));
1426 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
1427 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1428 if (reqbody == NULL || repbody == NULL)
1429 RETURN(err_serious(-EFAULT));
1432 * prepare @rdpg before calling lower layers and transfer itself. Here
1433 * reqbody->size contains offset of where to start to read and
1434 * reqbody->nlink contains number bytes to read.
1436 rdpg->rp_hash = reqbody->size;
1437 if (rdpg->rp_hash != reqbody->size) {
1438 CERROR("Invalid hash: "LPX64" != "LPX64"\n",
1439 rdpg->rp_hash, reqbody->size);
1443 rdpg->rp_attrs = reqbody->mode;
1444 rdpg->rp_count = reqbody->nlink;
1445 rdpg->rp_npages = (rdpg->rp_count + CFS_PAGE_SIZE - 1)>>CFS_PAGE_SHIFT;
1446 OBD_ALLOC(rdpg->rp_pages, rdpg->rp_npages * sizeof rdpg->rp_pages[0]);
1447 if (rdpg->rp_pages == NULL)
1450 for (i = 0; i < rdpg->rp_npages; ++i) {
1451 rdpg->rp_pages[i] = cfs_alloc_page(CFS_ALLOC_STD);
1452 if (rdpg->rp_pages[i] == NULL)
1453 GOTO(free_rdpg, rc = -ENOMEM);
1456 /* call lower layers to fill allocated pages with directory data */
1457 rc = mo_readpage(info->mti_env, mdt_object_child(object), rdpg);
1459 GOTO(free_rdpg, rc);
1461 /* send pages to client */
1462 rc = mdt_sendpage(info, rdpg);
1467 for (i = 0; i < rdpg->rp_npages; i++)
1468 if (rdpg->rp_pages[i] != NULL)
1469 cfs_free_page(rdpg->rp_pages[i]);
1470 OBD_FREE(rdpg->rp_pages, rdpg->rp_npages * sizeof rdpg->rp_pages[0]);
1472 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE))
1478 static int mdt_reint_internal(struct mdt_thread_info *info,
1479 struct mdt_lock_handle *lhc,
1482 struct req_capsule *pill = info->mti_pill;
1483 struct mdt_device *mdt = info->mti_mdt;
1484 struct mdt_body *repbody;
1489 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
1490 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER,
1491 mdt->mdt_max_mdsize);
1492 if (req_capsule_has_field(pill, &RMF_LOGCOOKIES, RCL_SERVER))
1493 req_capsule_set_size(pill, &RMF_LOGCOOKIES, RCL_SERVER,
1494 mdt->mdt_max_cookiesize);
1496 rc = req_capsule_server_pack(pill);
1498 CERROR("Can't pack response, rc %d\n", rc);
1499 RETURN(err_serious(rc));
1502 if (req_capsule_has_field(pill, &RMF_MDT_BODY, RCL_SERVER)) {
1503 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
1505 repbody->eadatasize = 0;
1506 repbody->aclsize = 0;
1509 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK))
1510 GOTO(out_shrink, rc = err_serious(-EFAULT));
1512 rc = mdt_reint_unpack(info, op);
1514 CERROR("Can't unpack reint, rc %d\n", rc);
1515 GOTO(out_shrink, rc = err_serious(rc));
1518 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_REINT_DELAY, 10);
1520 /* for replay no cookkie / lmm need, because client have this already */
1521 if (info->mti_spec.no_create == 1) {
1522 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
1523 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER, 0);
1525 if (req_capsule_has_field(pill, &RMF_LOGCOOKIES, RCL_SERVER))
1526 req_capsule_set_size(pill, &RMF_LOGCOOKIES, RCL_SERVER,
1530 rc = mdt_init_ucred_reint(info);
1532 GOTO(out_shrink, rc);
1534 rc = mdt_fix_attr_ucred(info, op);
1536 GOTO(out_ucred, rc = err_serious(rc));
1538 if (mdt_check_resent(info, mdt_reconstruct, lhc)) {
1539 rc = lustre_msg_get_status(mdt_info_req(info)->rq_repmsg);
1540 GOTO(out_ucred, rc);
1542 rc = mdt_reint_rec(info, lhc);
1545 mdt_exit_ucred(info);
1547 mdt_shrink_reply(info);
1551 static long mdt_reint_opcode(struct mdt_thread_info *info,
1552 const struct req_format **fmt)
1554 struct mdt_rec_reint *rec;
1557 opc = err_serious(-EFAULT);
1558 rec = req_capsule_client_get(info->mti_pill, &RMF_REC_REINT);
1560 opc = rec->rr_opcode;
1561 DEBUG_REQ(D_INODE, mdt_info_req(info), "reint opt = %ld", opc);
1562 if (opc < REINT_MAX && fmt[opc] != NULL)
1563 req_capsule_extend(info->mti_pill, fmt[opc]);
1565 CERROR("Unsupported opc: %ld\n", opc);
1566 opc = err_serious(opc);
1572 static int mdt_reint(struct mdt_thread_info *info)
1577 static const struct req_format *reint_fmts[REINT_MAX] = {
1578 [REINT_SETATTR] = &RQF_MDS_REINT_SETATTR,
1579 [REINT_CREATE] = &RQF_MDS_REINT_CREATE,
1580 [REINT_LINK] = &RQF_MDS_REINT_LINK,
1581 [REINT_UNLINK] = &RQF_MDS_REINT_UNLINK,
1582 [REINT_RENAME] = &RQF_MDS_REINT_RENAME,
1583 [REINT_OPEN] = &RQF_MDS_REINT_OPEN,
1584 [REINT_SETXATTR] = &RQF_MDS_REINT_SETXATTR
1589 opc = mdt_reint_opcode(info, reint_fmts);
1592 * No lock possible here from client to pass it to reint code
1595 rc = mdt_reint_internal(info, NULL, opc);
1600 info->mti_fail_id = OBD_FAIL_MDS_REINT_NET_REP;
1604 /* this should sync the whole device */
1605 static int mdt_device_sync(const struct lu_env *env, struct mdt_device *mdt)
1607 struct dt_device *dt = mdt->mdt_bottom;
1611 rc = dt->dd_ops->dt_sync(env, dt);
1615 /* this should sync this object */
1616 static int mdt_object_sync(struct mdt_thread_info *info)
1618 struct md_object *next;
1622 if (!mdt_object_exists(info->mti_object)) {
1623 CWARN("Non existing object "DFID"!\n",
1624 PFID(mdt_object_fid(info->mti_object)));
1627 next = mdt_object_child(info->mti_object);
1628 rc = mo_object_sync(info->mti_env, next);
1633 static int mdt_sync(struct mdt_thread_info *info)
1635 struct req_capsule *pill = info->mti_pill;
1636 struct mdt_body *body;
1640 /* The fid may be zero, so we req_capsule_set manually */
1641 req_capsule_set(pill, &RQF_MDS_SYNC);
1643 body = req_capsule_client_get(pill, &RMF_MDT_BODY);
1645 RETURN(err_serious(-EINVAL));
1647 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SYNC_PACK))
1648 RETURN(err_serious(-ENOMEM));
1650 if (fid_seq(&body->fid1) == 0) {
1651 /* sync the whole device */
1652 rc = req_capsule_server_pack(pill);
1654 rc = mdt_device_sync(info->mti_env, info->mti_mdt);
1656 rc = err_serious(rc);
1658 /* sync an object */
1659 rc = mdt_unpack_req_pack_rep(info, HABEO_CORPUS|HABEO_REFERO);
1661 rc = mdt_object_sync(info);
1663 struct md_object *next;
1664 const struct lu_fid *fid;
1665 struct lu_attr *la = &info->mti_attr.ma_attr;
1667 next = mdt_object_child(info->mti_object);
1668 info->mti_attr.ma_need = MA_INODE;
1669 info->mti_attr.ma_valid = 0;
1670 rc = mo_attr_get(info->mti_env, next,
1673 body = req_capsule_server_get(pill,
1675 fid = mdt_object_fid(info->mti_object);
1676 mdt_pack_attr2body(info, body, la, fid);
1680 rc = err_serious(rc);
1685 #ifdef HAVE_QUOTA_SUPPORT
1686 static int mdt_quotacheck_handle(struct mdt_thread_info *info)
1688 struct obd_quotactl *oqctl;
1689 struct req_capsule *pill = info->mti_pill;
1690 struct obd_export *exp = info->mti_exp;
1691 struct md_device *next = info->mti_mdt->mdt_child;
1695 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_QUOTACHECK_NET))
1698 oqctl = req_capsule_client_get(pill, &RMF_OBD_QUOTACTL);
1702 /* remote client has no permission for quotacheck */
1703 if (unlikely(exp_connect_rmtclient(exp)))
1706 rc = req_capsule_server_pack(pill);
1710 rc = next->md_ops->mdo_quota.mqo_check(info->mti_env, next, exp,
1715 static int mdt_quotactl_handle(struct mdt_thread_info *info)
1717 struct obd_quotactl *oqctl, *repoqc;
1718 struct req_capsule *pill = info->mti_pill;
1719 struct obd_export *exp = info->mti_exp;
1720 struct md_device *next = info->mti_mdt->mdt_child;
1721 const struct md_quota_operations *mqo = &next->md_ops->mdo_quota;
1725 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_QUOTACTL_NET))
1728 oqctl = req_capsule_client_get(pill, &RMF_OBD_QUOTACTL);
1733 if (exp_connect_rmtclient(exp)) {
1734 struct ptlrpc_request *req = mdt_info_req(info);
1735 struct mdt_export_data *med = mdt_req2med(req);
1736 struct lustre_idmap_table *idmap = med->med_idmap;
1738 if (unlikely(oqctl->qc_cmd != Q_GETQUOTA &&
1739 oqctl->qc_cmd != Q_GETINFO))
1743 if (oqctl->qc_type == USRQUOTA)
1744 id = lustre_idmap_lookup_uid(NULL, idmap, 0,
1746 else if (oqctl->qc_type == GRPQUOTA)
1747 id = lustre_idmap_lookup_gid(NULL, idmap, 0,
1752 if (id == CFS_IDMAP_NOTFOUND) {
1753 CDEBUG(D_QUOTA, "no mapping for id %u\n",
1759 rc = req_capsule_server_pack(pill);
1763 repoqc = req_capsule_server_get(pill, &RMF_OBD_QUOTACTL);
1764 LASSERT(repoqc != NULL);
1766 switch (oqctl->qc_cmd) {
1768 rc = mqo->mqo_on(info->mti_env, next, oqctl->qc_type);
1771 rc = mqo->mqo_off(info->mti_env, next, oqctl->qc_type);
1774 rc = mqo->mqo_setinfo(info->mti_env, next, oqctl->qc_type, id,
1778 rc = mqo->mqo_getinfo(info->mti_env, next, oqctl->qc_type, id,
1782 rc = mqo->mqo_setquota(info->mti_env, next, oqctl->qc_type, id,
1786 rc = mqo->mqo_getquota(info->mti_env, next, oqctl->qc_type, id,
1790 rc = mqo->mqo_getoinfo(info->mti_env, next, oqctl->qc_type, id,
1794 rc = mqo->mqo_getoquota(info->mti_env, next, oqctl->qc_type, id,
1797 case LUSTRE_Q_INVALIDATE:
1798 rc = mqo->mqo_invalidate(info->mti_env, next, oqctl->qc_type);
1800 case LUSTRE_Q_FINVALIDATE:
1801 rc = mqo->mqo_finvalidate(info->mti_env, next, oqctl->qc_type);
1804 CERROR("unsupported mdt_quotactl command: %d\n",
1816 * OBD PING and other handlers.
1818 static int mdt_obd_ping(struct mdt_thread_info *info)
1823 req_capsule_set(info->mti_pill, &RQF_OBD_PING);
1825 rc = target_handle_ping(mdt_info_req(info));
1827 rc = err_serious(rc);
1831 static int mdt_obd_log_cancel(struct mdt_thread_info *info)
1833 return err_serious(-EOPNOTSUPP);
1836 static int mdt_obd_qc_callback(struct mdt_thread_info *info)
1838 return err_serious(-EOPNOTSUPP);
1846 /** clone llog ctxt from child (mdd)
1847 * This allows remote llog (replicator) access.
1848 * We can either pass all llog RPCs (eg mdt_llog_create) on to child where the
1849 * context was originally set up, or we can handle them directly.
1850 * I choose the latter, but that means I need any llog
1851 * contexts set up by child to be accessable by the mdt. So we clone the
1852 * context into our context list here.
1854 static int mdt_llog_ctxt_clone(const struct lu_env *env, struct mdt_device *mdt,
1857 struct md_device *next = mdt->mdt_child;
1858 struct llog_ctxt *ctxt;
1861 if (!llog_ctxt_null(mdt2obd_dev(mdt), idx))
1864 rc = next->md_ops->mdo_llog_ctxt_get(env, next, idx, (void **)&ctxt);
1865 if (rc || ctxt == NULL) {
1866 CERROR("Can't get mdd ctxt %d\n", rc);
1870 rc = llog_group_set_ctxt(&mdt2obd_dev(mdt)->obd_olg, ctxt, idx);
1872 CERROR("Can't set mdt ctxt %d\n", rc);
1877 static int mdt_llog_ctxt_unclone(const struct lu_env *env,
1878 struct mdt_device *mdt, int idx)
1880 struct llog_ctxt *ctxt;
1882 ctxt = llog_get_context(mdt2obd_dev(mdt), idx);
1885 /* Put once for the get we just did, and once for the clone */
1886 llog_ctxt_put(ctxt);
1887 llog_ctxt_put(ctxt);
1891 static int mdt_llog_create(struct mdt_thread_info *info)
1895 req_capsule_set(info->mti_pill, &RQF_LLOG_ORIGIN_HANDLE_CREATE);
1896 rc = llog_origin_handle_create(mdt_info_req(info));
1897 return (rc < 0 ? err_serious(rc) : rc);
1900 static int mdt_llog_destroy(struct mdt_thread_info *info)
1904 req_capsule_set(info->mti_pill, &RQF_LLOG_ORIGIN_HANDLE_DESTROY);
1905 rc = llog_origin_handle_destroy(mdt_info_req(info));
1906 return (rc < 0 ? err_serious(rc) : rc);
1909 static int mdt_llog_read_header(struct mdt_thread_info *info)
1913 req_capsule_set(info->mti_pill, &RQF_LLOG_ORIGIN_HANDLE_READ_HEADER);
1914 rc = llog_origin_handle_read_header(mdt_info_req(info));
1915 return (rc < 0 ? err_serious(rc) : rc);
1918 static int mdt_llog_next_block(struct mdt_thread_info *info)
1922 req_capsule_set(info->mti_pill, &RQF_LLOG_ORIGIN_HANDLE_NEXT_BLOCK);
1923 rc = llog_origin_handle_next_block(mdt_info_req(info));
1924 return (rc < 0 ? err_serious(rc) : rc);
1927 static int mdt_llog_prev_block(struct mdt_thread_info *info)
1931 req_capsule_set(info->mti_pill, &RQF_LLOG_ORIGIN_HANDLE_PREV_BLOCK);
1932 rc = llog_origin_handle_prev_block(mdt_info_req(info));
1933 return (rc < 0 ? err_serious(rc) : rc);
1940 static struct ldlm_callback_suite cbs = {
1941 .lcs_completion = ldlm_server_completion_ast,
1942 .lcs_blocking = ldlm_server_blocking_ast,
1946 static int mdt_enqueue(struct mdt_thread_info *info)
1948 struct ptlrpc_request *req;
1953 * info->mti_dlm_req already contains swapped and (if necessary)
1954 * converted dlm request.
1956 LASSERT(info->mti_dlm_req != NULL);
1958 req = mdt_info_req(info);
1961 * Lock without inodebits makes no sense and will oops later in
1962 * ldlm. Let's check it now to see if we have wrong lock from client or
1963 * bits get corrupted somewhere in mdt_intent_policy().
1965 req_bits = info->mti_dlm_req->lock_desc.l_policy_data.l_inodebits.bits;
1966 /* This is disabled because we need to support liblustre flock.
1967 * LASSERT(req_bits != 0);
1970 rc = ldlm_handle_enqueue0(info->mti_mdt->mdt_namespace,
1971 req, info->mti_dlm_req, &cbs);
1972 info->mti_fail_id = OBD_FAIL_LDLM_REPLY;
1973 return rc ? err_serious(rc) : req->rq_status;
1976 static int mdt_convert(struct mdt_thread_info *info)
1979 struct ptlrpc_request *req;
1981 LASSERT(info->mti_dlm_req);
1982 req = mdt_info_req(info);
1983 rc = ldlm_handle_convert0(req, info->mti_dlm_req);
1984 return rc ? err_serious(rc) : req->rq_status;
1987 static int mdt_bl_callback(struct mdt_thread_info *info)
1989 CERROR("bl callbacks should not happen on MDS\n");
1991 return err_serious(-EOPNOTSUPP);
1994 static int mdt_cp_callback(struct mdt_thread_info *info)
1996 CERROR("cp callbacks should not happen on MDS\n");
1998 return err_serious(-EOPNOTSUPP);
2002 * sec context handlers
2004 static int mdt_sec_ctx_handle(struct mdt_thread_info *info)
2008 rc = mdt_handle_idmap(info);
2011 struct ptlrpc_request *req = mdt_info_req(info);
2014 opc = lustre_msg_get_opc(req->rq_reqmsg);
2015 if (opc == SEC_CTX_INIT || opc == SEC_CTX_INIT_CONT)
2016 sptlrpc_svc_ctx_invalidate(req);
2019 OBD_FAIL_TIMEOUT(OBD_FAIL_SEC_CTX_HDL_PAUSE, obd_fail_val);
2024 static struct mdt_object *mdt_obj(struct lu_object *o)
2026 LASSERT(lu_device_is_mdt(o->lo_dev));
2027 return container_of0(o, struct mdt_object, mot_obj.mo_lu);
2030 struct mdt_object *mdt_object_find(const struct lu_env *env,
2031 struct mdt_device *d,
2032 const struct lu_fid *f)
2034 struct lu_object *o;
2035 struct mdt_object *m;
2038 CDEBUG(D_INFO, "Find object for "DFID"\n", PFID(f));
2039 o = lu_object_find(env, &d->mdt_md_dev.md_lu_dev, f, NULL);
2040 if (unlikely(IS_ERR(o)))
2041 m = (struct mdt_object *)o;
2048 * Asyncronous commit for mdt device.
2050 * Pass asynchonous commit call down the MDS stack.
2052 * \param env environment
2053 * \param mdt the mdt device
2055 static void mdt_device_commit_async(const struct lu_env *env,
2056 struct mdt_device *mdt)
2058 struct dt_device *dt = mdt->mdt_bottom;
2061 rc = dt->dd_ops->dt_commit_async(env, dt);
2062 if (unlikely(rc != 0))
2063 CWARN("async commit start failed with rc = %d", rc);
2067 * Mark the lock as "synchonous".
2069 * Mark the lock to deffer transaction commit to the unlock time.
2071 * \param lock the lock to mark as "synchonous"
2073 * \see mdt_is_lock_sync
2074 * \see mdt_save_lock
2076 static inline void mdt_set_lock_sync(struct ldlm_lock *lock)
2078 lock->l_ast_data = (void*)1;
2082 * Check whehter the lock "synchonous" or not.
2084 * \param lock the lock to check
2085 * \retval 1 the lock is "synchonous"
2086 * \retval 0 the lock isn't "synchronous"
2088 * \see mdt_set_lock_sync
2089 * \see mdt_save_lock
2091 static inline int mdt_is_lock_sync(struct ldlm_lock *lock)
2093 return lock->l_ast_data != NULL;
2097 * Blocking AST for mdt locks.
2099 * Starts transaction commit if in case of COS lock conflict or
2100 * deffers such a commit to the mdt_save_lock.
2102 * \param lock the lock which blocks a request or cancelling lock
2103 * \param desc unused
2104 * \param data unused
2105 * \param flag indicates whether this cancelling or blocking callback
2107 * \see ldlm_blocking_ast_nocheck
2109 int mdt_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
2110 void *data, int flag)
2112 struct obd_device *obd = lock->l_resource->lr_namespace->ns_obd;
2113 struct mdt_device *mdt = mdt_dev(obd->obd_lu_dev);
2117 if (flag == LDLM_CB_CANCELING)
2119 lock_res_and_lock(lock);
2120 if (lock->l_blocking_ast != mdt_blocking_ast) {
2121 unlock_res_and_lock(lock);
2124 if (mdt_cos_is_enabled(mdt) &&
2125 lock->l_req_mode & (LCK_PW | LCK_EX) &&
2126 lock->l_blocking_lock != NULL &&
2127 lock->l_client_cookie != lock->l_blocking_lock->l_client_cookie) {
2128 mdt_set_lock_sync(lock);
2130 rc = ldlm_blocking_ast_nocheck(lock);
2132 /* There is no lock conflict if l_blocking_lock == NULL,
2133 * it indicates a blocking ast sent from ldlm_lock_decref_internal
2134 * when the last reference to a local lock was released */
2135 if (lock->l_req_mode == LCK_COS && lock->l_blocking_lock != NULL) {
2138 rc = lu_env_init(&env, LCT_MD_THREAD);
2139 if (unlikely(rc != 0))
2140 CWARN("lu_env initialization failed with rc = %d,"
2141 "cannot start asynchronous commit\n", rc);
2143 mdt_device_commit_async(&env, mdt);
2149 int mdt_object_lock(struct mdt_thread_info *info, struct mdt_object *o,
2150 struct mdt_lock_handle *lh, __u64 ibits, int locality)
2152 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
2153 ldlm_policy_data_t *policy = &info->mti_policy;
2154 struct ldlm_res_id *res_id = &info->mti_res_id;
2158 LASSERT(!lustre_handle_is_used(&lh->mlh_reg_lh));
2159 LASSERT(!lustre_handle_is_used(&lh->mlh_pdo_lh));
2160 LASSERT(lh->mlh_reg_mode != LCK_MINMODE);
2161 LASSERT(lh->mlh_type != MDT_NUL_LOCK);
2163 if (mdt_object_exists(o) < 0) {
2164 if (locality == MDT_CROSS_LOCK) {
2165 /* cross-ref object fix */
2166 ibits &= ~MDS_INODELOCK_UPDATE;
2167 ibits |= MDS_INODELOCK_LOOKUP;
2169 LASSERT(!(ibits & MDS_INODELOCK_UPDATE));
2170 LASSERT(ibits & MDS_INODELOCK_LOOKUP);
2172 /* No PDO lock on remote object */
2173 LASSERT(lh->mlh_type != MDT_PDO_LOCK);
2176 if (lh->mlh_type == MDT_PDO_LOCK) {
2177 /* check for exists after object is locked */
2178 if (mdt_object_exists(o) == 0) {
2179 /* Non-existent object shouldn't have PDO lock */
2182 /* Non-dir object shouldn't have PDO lock */
2183 LASSERT(S_ISDIR(lu_object_attr(&o->mot_obj.mo_lu)));
2187 memset(policy, 0, sizeof(*policy));
2188 fid_build_reg_res_name(mdt_object_fid(o), res_id);
2191 * Take PDO lock on whole directory and build correct @res_id for lock
2192 * on part of directory.
2194 if (lh->mlh_pdo_hash != 0) {
2195 LASSERT(lh->mlh_type == MDT_PDO_LOCK);
2196 mdt_lock_pdo_mode(info, o, lh);
2197 if (lh->mlh_pdo_mode != LCK_NL) {
2199 * Do not use LDLM_FL_LOCAL_ONLY for parallel lock, it
2200 * is never going to be sent to client and we do not
2201 * want it slowed down due to possible cancels.
2203 policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
2204 rc = mdt_fid_lock(ns, &lh->mlh_pdo_lh, lh->mlh_pdo_mode,
2205 policy, res_id, LDLM_FL_ATOMIC_CB,
2206 &info->mti_exp->exp_handle.h_cookie);
2212 * Finish res_id initializing by name hash marking part of
2213 * directory which is taking modification.
2215 res_id->name[LUSTRE_RES_ID_HSH_OFF] = lh->mlh_pdo_hash;
2218 policy->l_inodebits.bits = ibits;
2221 * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
2222 * going to be sent to client. If it is - mdt_intent_policy() path will
2223 * fix it up and turn FL_LOCAL flag off.
2225 rc = mdt_fid_lock(ns, &lh->mlh_reg_lh, lh->mlh_reg_mode, policy,
2226 res_id, LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB,
2227 &info->mti_exp->exp_handle.h_cookie);
2233 mdt_object_unlock(info, o, lh, 1);
2240 * Save a lock within request object.
2242 * Keep the lock referenced until whether client ACK or transaction
2243 * commit happens or release the lock immediately depending on input
2244 * parameters. If COS is ON, a write lock is converted to COS lock
2247 * \param info thead info object
2248 * \param h lock handle
2249 * \param mode lock mode
2250 * \param decref force immediate lock releasing
2253 void mdt_save_lock(struct mdt_thread_info *info, struct lustre_handle *h,
2254 ldlm_mode_t mode, int decref)
2258 if (lustre_handle_is_used(h)) {
2259 if (decref || !info->mti_has_trans ||
2260 !(mode & (LCK_PW | LCK_EX))){
2261 mdt_fid_unlock(h, mode);
2263 struct mdt_device *mdt = info->mti_mdt;
2264 struct ldlm_lock *lock = ldlm_handle2lock(h);
2265 struct ptlrpc_request *req = mdt_info_req(info);
2268 LASSERTF(lock != NULL, "no lock for cookie "LPX64"\n",
2270 CDEBUG(D_HA, "request = %p reply state = %p"
2271 " transno = "LPD64"\n",
2272 req, req->rq_reply_state, req->rq_transno);
2273 if (mdt_cos_is_enabled(mdt)) {
2275 ldlm_lock_downgrade(lock, LCK_COS);
2278 ptlrpc_save_lock(req, h, mode, no_ack);
2279 if (mdt_is_lock_sync(lock)) {
2280 CDEBUG(D_HA, "found sync-lock,"
2281 " async commit started\n");
2282 mdt_device_commit_async(info->mti_env,
2285 LDLM_LOCK_PUT(lock);
2294 * Unlock mdt object.
2296 * Immeditely release the regular lock and the PDO lock or save the
2297 * lock in reqeuest and keep them referenced until client ACK or
2298 * transaction commit.
2300 * \param info thread info object
2301 * \param o mdt object
2302 * \param h mdt lock handle referencing regular and PDO locks
2303 * \param decref force immediate lock releasing
2305 void mdt_object_unlock(struct mdt_thread_info *info, struct mdt_object *o,
2306 struct mdt_lock_handle *lh, int decref)
2310 mdt_save_lock(info, &lh->mlh_pdo_lh, lh->mlh_pdo_mode, decref);
2311 mdt_save_lock(info, &lh->mlh_reg_lh, lh->mlh_reg_mode, decref);
2316 struct mdt_object *mdt_object_find_lock(struct mdt_thread_info *info,
2317 const struct lu_fid *f,
2318 struct mdt_lock_handle *lh,
2321 struct mdt_object *o;
2323 o = mdt_object_find(info->mti_env, info->mti_mdt, f);
2327 rc = mdt_object_lock(info, o, lh, ibits,
2330 mdt_object_put(info->mti_env, o);
2337 void mdt_object_unlock_put(struct mdt_thread_info * info,
2338 struct mdt_object * o,
2339 struct mdt_lock_handle *lh,
2342 mdt_object_unlock(info, o, lh, decref);
2343 mdt_object_put(info->mti_env, o);
2346 static struct mdt_handler *mdt_handler_find(__u32 opc,
2347 struct mdt_opc_slice *supported)
2349 struct mdt_opc_slice *s;
2350 struct mdt_handler *h;
2353 for (s = supported; s->mos_hs != NULL; s++) {
2354 if (s->mos_opc_start <= opc && opc < s->mos_opc_end) {
2355 h = s->mos_hs + (opc - s->mos_opc_start);
2356 if (likely(h->mh_opc != 0))
2357 LASSERTF(h->mh_opc == opc,
2358 "opcode mismatch %d != %d\n",
2361 h = NULL; /* unsupported opc */
2368 static int mdt_lock_resname_compat(struct mdt_device *m,
2369 struct ldlm_request *req)
2371 /* XXX something... later. */
2375 static int mdt_lock_reply_compat(struct mdt_device *m, struct ldlm_reply *rep)
2377 /* XXX something... later. */
2382 * Generic code handling requests that have struct mdt_body passed in:
2384 * - extract mdt_body from request and save it in @info, if present;
2386 * - create lu_object, corresponding to the fid in mdt_body, and save it in
2389 * - if HABEO_CORPUS flag is set for this request type check whether object
2390 * actually exists on storage (lu_object_exists()).
2393 static int mdt_body_unpack(struct mdt_thread_info *info, __u32 flags)
2395 const struct mdt_body *body;
2396 struct mdt_object *obj;
2397 const struct lu_env *env;
2398 struct req_capsule *pill;
2402 env = info->mti_env;
2403 pill = info->mti_pill;
2405 body = info->mti_body = req_capsule_client_get(pill, &RMF_MDT_BODY);
2409 if (!(body->valid & OBD_MD_FLID))
2412 if (!fid_is_sane(&body->fid1)) {
2413 CERROR("Invalid fid: "DFID"\n", PFID(&body->fid1));
2418 * Do not get size or any capa fields before we check that request
2419 * contains capa actually. There are some requests which do not, for
2420 * instance MDS_IS_SUBDIR.
2422 if (req_capsule_has_field(pill, &RMF_CAPA1, RCL_CLIENT) &&
2423 req_capsule_get_size(pill, &RMF_CAPA1, RCL_CLIENT))
2424 mdt_set_capainfo(info, 0, &body->fid1,
2425 req_capsule_client_get(pill, &RMF_CAPA1));
2427 obj = mdt_object_find(env, info->mti_mdt, &body->fid1);
2429 if ((flags & HABEO_CORPUS) &&
2430 !mdt_object_exists(obj)) {
2431 mdt_object_put(env, obj);
2432 /* for capability renew ENOENT will be handled in
2434 if (body->valid & OBD_MD_FLOSSCAPA)
2439 info->mti_object = obj;
2448 static int mdt_unpack_req_pack_rep(struct mdt_thread_info *info, __u32 flags)
2450 struct req_capsule *pill = info->mti_pill;
2454 if (req_capsule_has_field(pill, &RMF_MDT_BODY, RCL_CLIENT))
2455 rc = mdt_body_unpack(info, flags);
2459 if (rc == 0 && (flags & HABEO_REFERO)) {
2460 struct mdt_device *mdt = info->mti_mdt;
2464 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
2465 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER,
2466 mdt->mdt_max_mdsize);
2467 if (req_capsule_has_field(pill, &RMF_LOGCOOKIES, RCL_SERVER))
2468 req_capsule_set_size(pill, &RMF_LOGCOOKIES, RCL_SERVER,
2469 mdt->mdt_max_cookiesize);
2471 rc = req_capsule_server_pack(pill);
2476 static int mdt_init_capa_ctxt(const struct lu_env *env, struct mdt_device *m)
2478 struct md_device *next = m->mdt_child;
2480 return next->md_ops->mdo_init_capa_ctxt(env, next,
2481 m->mdt_opts.mo_mds_capa,
2482 m->mdt_capa_timeout,
2488 * Invoke handler for this request opc. Also do necessary preprocessing
2489 * (according to handler ->mh_flags), and post-processing (setting of
2490 * ->last_{xid,committed}).
2492 static int mdt_req_handle(struct mdt_thread_info *info,
2493 struct mdt_handler *h, struct ptlrpc_request *req)
2495 int rc, serious = 0;
2500 LASSERT(h->mh_act != NULL);
2501 LASSERT(h->mh_opc == lustre_msg_get_opc(req->rq_reqmsg));
2502 LASSERT(current->journal_info == NULL);
2505 * Checking for various OBD_FAIL_$PREF_$OPC_NET codes. _Do_ not try
2506 * to put same checks into handlers like mdt_close(), mdt_reint(),
2507 * etc., without talking to mdt authors first. Checking same thing
2508 * there again is useless and returning 0 error wihtout packing reply
2509 * is buggy! Handlers either pack reply or return error.
2511 * We return 0 here and do not send any reply in order to emulate
2512 * network failure. Do not send any reply in case any of NET related
2513 * fail_id has occured.
2515 if (OBD_FAIL_CHECK_ORSET(h->mh_fail_id, OBD_FAIL_ONCE))
2519 flags = h->mh_flags;
2520 LASSERT(ergo(flags & (HABEO_CORPUS|HABEO_REFERO), h->mh_fmt != NULL));
2522 if (h->mh_fmt != NULL) {
2523 req_capsule_set(info->mti_pill, h->mh_fmt);
2524 rc = mdt_unpack_req_pack_rep(info, flags);
2527 if (rc == 0 && flags & MUTABOR &&
2528 req->rq_export->exp_connect_flags & OBD_CONNECT_RDONLY)
2529 /* should it be rq_status? */
2532 if (rc == 0 && flags & HABEO_CLAVIS) {
2533 struct ldlm_request *dlm_req;
2535 LASSERT(h->mh_fmt != NULL);
2537 dlm_req = req_capsule_client_get(info->mti_pill, &RMF_DLM_REQ);
2538 if (dlm_req != NULL) {
2539 if (info->mti_mdt->mdt_opts.mo_compat_resname)
2540 rc = mdt_lock_resname_compat(info->mti_mdt,
2542 info->mti_dlm_req = dlm_req;
2548 /* capability setting changed via /proc, needs reinitialize ctxt */
2549 if (info->mti_mdt && info->mti_mdt->mdt_capa_conf) {
2550 mdt_init_capa_ctxt(info->mti_env, info->mti_mdt);
2551 info->mti_mdt->mdt_capa_conf = 0;
2554 if (likely(rc == 0)) {
2556 * Process request, there can be two types of rc:
2557 * 1) errors with msg unpack/pack, other failures outside the
2558 * operation itself. This is counted as serious errors;
2559 * 2) errors during fs operation, should be placed in rq_status
2562 rc = h->mh_act(info);
2564 !req->rq_no_reply && req->rq_reply_state == NULL) {
2565 DEBUG_REQ(D_ERROR, req, "MDT \"handler\" %s did not "
2566 "pack reply and returned 0 error\n",
2570 serious = is_serious(rc);
2571 rc = clear_serious(rc);
2575 req->rq_status = rc;
2578 * ELDLM_* codes which > 0 should be in rq_status only as well as
2579 * all non-serious errors.
2581 if (rc > 0 || !serious)
2584 LASSERT(current->journal_info == NULL);
2586 if (rc == 0 && (flags & HABEO_CLAVIS) &&
2587 info->mti_mdt->mdt_opts.mo_compat_resname) {
2588 struct ldlm_reply *dlmrep;
2590 dlmrep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
2592 rc = mdt_lock_reply_compat(info->mti_mdt, dlmrep);
2595 /* If we're DISCONNECTing, the mdt_export_data is already freed */
2596 if (likely(rc == 0 && req->rq_export && h->mh_opc != MDS_DISCONNECT))
2597 target_committed_to_req(req);
2599 if (unlikely((lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) &&
2600 lustre_msg_get_transno(req->rq_reqmsg) == 0)) {
2601 DEBUG_REQ(D_ERROR, req, "transno is 0 during REPLAY");
2605 target_send_reply(req, rc, info->mti_fail_id);
2609 void mdt_lock_handle_init(struct mdt_lock_handle *lh)
2611 lh->mlh_type = MDT_NUL_LOCK;
2612 lh->mlh_reg_lh.cookie = 0ull;
2613 lh->mlh_reg_mode = LCK_MINMODE;
2614 lh->mlh_pdo_lh.cookie = 0ull;
2615 lh->mlh_pdo_mode = LCK_MINMODE;
2618 void mdt_lock_handle_fini(struct mdt_lock_handle *lh)
2620 LASSERT(!lustre_handle_is_used(&lh->mlh_reg_lh));
2621 LASSERT(!lustre_handle_is_used(&lh->mlh_pdo_lh));
2625 * Initialize fields of struct mdt_thread_info. Other fields are left in
2626 * uninitialized state, because it's too expensive to zero out whole
2627 * mdt_thread_info (> 1K) on each request arrival.
2629 static void mdt_thread_info_init(struct ptlrpc_request *req,
2630 struct mdt_thread_info *info)
2633 struct md_capainfo *ci;
2635 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2636 info->mti_pill = &req->rq_pill;
2639 for (i = 0; i < ARRAY_SIZE(info->mti_lh); i++)
2640 mdt_lock_handle_init(&info->mti_lh[i]);
2642 /* mdt device: it can be NULL while CONNECT */
2643 if (req->rq_export) {
2644 info->mti_mdt = mdt_dev(req->rq_export->exp_obd->obd_lu_dev);
2645 info->mti_exp = req->rq_export;
2647 info->mti_mdt = NULL;
2648 info->mti_env = req->rq_svc_thread->t_env;
2649 ci = md_capainfo(info->mti_env);
2650 memset(ci, 0, sizeof *ci);
2651 if (req->rq_export) {
2652 if (exp_connect_rmtclient(req->rq_export))
2653 ci->mc_auth = LC_ID_CONVERT;
2654 else if (req->rq_export->exp_connect_flags &
2655 OBD_CONNECT_MDS_CAPA)
2656 ci->mc_auth = LC_ID_PLAIN;
2658 ci->mc_auth = LC_ID_NONE;
2661 info->mti_fail_id = OBD_FAIL_MDS_ALL_REPLY_NET;
2662 info->mti_transno = lustre_msg_get_transno(req->rq_reqmsg);
2663 info->mti_mos[0] = NULL;
2664 info->mti_mos[1] = NULL;
2665 info->mti_mos[2] = NULL;
2666 info->mti_mos[3] = NULL;
2668 memset(&info->mti_attr, 0, sizeof(info->mti_attr));
2669 info->mti_body = NULL;
2670 info->mti_object = NULL;
2671 info->mti_dlm_req = NULL;
2672 info->mti_has_trans = 0;
2673 info->mti_no_need_trans = 0;
2674 info->mti_cross_ref = 0;
2675 info->mti_opdata = 0;
2677 /* To not check for split by default. */
2678 info->mti_spec.sp_ck_split = 0;
2679 info->mti_spec.no_create = 0;
2682 static void mdt_thread_info_fini(struct mdt_thread_info *info)
2686 req_capsule_fini(info->mti_pill);
2687 if (info->mti_object != NULL) {
2688 mdt_object_put(info->mti_env, info->mti_object);
2689 info->mti_object = NULL;
2691 for (i = 0; i < ARRAY_SIZE(info->mti_lh); i++)
2692 mdt_lock_handle_fini(&info->mti_lh[i]);
2693 info->mti_env = NULL;
2696 static int mdt_filter_recovery_request(struct ptlrpc_request *req,
2697 struct obd_device *obd, int *process)
2699 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2700 case MDS_CONNECT: /* This will never get here, but for completeness. */
2701 case OST_CONNECT: /* This will never get here, but for completeness. */
2702 case MDS_DISCONNECT:
2703 case OST_DISCONNECT:
2708 case MDS_DONE_WRITING:
2709 case MDS_SYNC: /* used in unmounting */
2715 *process = target_queue_recovery_request(req, obd);
2719 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
2726 * Handle recovery. Return:
2727 * +1: continue request processing;
2728 * -ve: abort immediately with the given error code;
2729 * 0: send reply with error code in req->rq_status;
2731 static int mdt_recovery(struct mdt_thread_info *info)
2733 struct ptlrpc_request *req = mdt_info_req(info);
2735 struct obd_device *obd;
2739 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2742 case SEC_CTX_INIT_CONT:
2748 rc = mdt_handle_idmap(info);
2757 if (unlikely(req->rq_export == NULL)) {
2758 CERROR("operation %d on unconnected MDS from %s\n",
2759 lustre_msg_get_opc(req->rq_reqmsg),
2760 libcfs_id2str(req->rq_peer));
2761 /* FIXME: For CMD cleanup, when mds_B stop, the req from
2762 * mds_A will get -ENOTCONN(especially for ping req),
2763 * which will cause that mds_A deactive timeout, then when
2764 * mds_A cleanup, the cleanup process will be suspended since
2765 * deactive timeout is not zero.
2767 req->rq_status = -ENOTCONN;
2768 target_send_reply(req, -ENOTCONN, info->mti_fail_id);
2772 /* sanity check: if the xid matches, the request must be marked as a
2773 * resent or replayed */
2774 if (req_xid_is_last(req)) {
2775 if (!(lustre_msg_get_flags(req->rq_reqmsg) &
2776 (MSG_RESENT | MSG_REPLAY))) {
2777 DEBUG_REQ(D_WARNING, req, "rq_xid "LPU64" matches last_xid, "
2778 "expected REPLAY or RESENT flag (%x)", req->rq_xid,
2779 lustre_msg_get_flags(req->rq_reqmsg));
2781 req->rq_status = -ENOTCONN;
2786 /* else: note the opposite is not always true; a RESENT req after a
2787 * failover will usually not match the last_xid, since it was likely
2788 * never committed. A REPLAYed request will almost never match the
2789 * last xid, however it could for a committed, but still retained,
2792 obd = req->rq_export->exp_obd;
2794 /* Check for aborted recovery... */
2795 spin_lock_bh(&obd->obd_processing_task_lock);
2796 recovering = obd->obd_recovering;
2797 spin_unlock_bh(&obd->obd_processing_task_lock);
2798 if (unlikely(recovering)) {
2801 DEBUG_REQ(D_INFO, req, "Got new replay");
2802 rc = mdt_filter_recovery_request(req, obd, &should_process);
2803 if (rc != 0 || !should_process)
2805 else if (should_process < 0) {
2806 req->rq_status = should_process;
2807 rc = ptlrpc_error(req);
2814 static int mdt_msg_check_version(struct lustre_msg *msg)
2818 switch (lustre_msg_get_opc(msg)) {
2820 case MDS_DISCONNECT:
2823 case SEC_CTX_INIT_CONT:
2825 rc = lustre_msg_check_version(msg, LUSTRE_OBD_VERSION);
2827 CERROR("bad opc %u version %08x, expecting %08x\n",
2828 lustre_msg_get_opc(msg),
2829 lustre_msg_get_version(msg),
2830 LUSTRE_OBD_VERSION);
2834 case MDS_GETATTR_NAME:
2841 case MDS_DONE_WRITING:
2848 case MDS_QUOTACHECK:
2854 rc = lustre_msg_check_version(msg, LUSTRE_MDS_VERSION);
2856 CERROR("bad opc %u version %08x, expecting %08x\n",
2857 lustre_msg_get_opc(msg),
2858 lustre_msg_get_version(msg),
2859 LUSTRE_MDS_VERSION);
2863 case LDLM_BL_CALLBACK:
2864 case LDLM_CP_CALLBACK:
2865 rc = lustre_msg_check_version(msg, LUSTRE_DLM_VERSION);
2867 CERROR("bad opc %u version %08x, expecting %08x\n",
2868 lustre_msg_get_opc(msg),
2869 lustre_msg_get_version(msg),
2870 LUSTRE_DLM_VERSION);
2872 case OBD_LOG_CANCEL:
2873 case LLOG_ORIGIN_HANDLE_CREATE:
2874 case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
2875 case LLOG_ORIGIN_HANDLE_READ_HEADER:
2876 case LLOG_ORIGIN_HANDLE_CLOSE:
2877 case LLOG_ORIGIN_HANDLE_DESTROY:
2878 case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
2880 rc = lustre_msg_check_version(msg, LUSTRE_LOG_VERSION);
2882 CERROR("bad opc %u version %08x, expecting %08x\n",
2883 lustre_msg_get_opc(msg),
2884 lustre_msg_get_version(msg),
2885 LUSTRE_LOG_VERSION);
2888 CERROR("MDS unknown opcode %d\n", lustre_msg_get_opc(msg));
2894 static int mdt_handle0(struct ptlrpc_request *req,
2895 struct mdt_thread_info *info,
2896 struct mdt_opc_slice *supported)
2898 struct mdt_handler *h;
2899 struct lustre_msg *msg;
2904 if (OBD_FAIL_CHECK_ORSET(OBD_FAIL_MDS_ALL_REQUEST_NET, OBD_FAIL_ONCE))
2907 LASSERT(current->journal_info == NULL);
2909 msg = req->rq_reqmsg;
2910 rc = mdt_msg_check_version(msg);
2911 if (likely(rc == 0)) {
2912 rc = mdt_recovery(info);
2913 if (likely(rc == +1)) {
2914 switch (lustre_msg_get_opc(msg)) {
2916 req->rq_bulk_read = 1;
2919 req->rq_bulk_write = 1;
2923 h = mdt_handler_find(lustre_msg_get_opc(msg),
2925 if (likely(h != NULL)) {
2926 rc = mdt_req_handle(info, h, req);
2928 CERROR("The unsupported opc: 0x%x\n",
2929 lustre_msg_get_opc(msg) );
2930 req->rq_status = -ENOTSUPP;
2931 rc = ptlrpc_error(req);
2936 CERROR(LUSTRE_MDT_NAME" drops mal-formed request\n");
2941 * MDT handler function called by ptlrpc service thread when request comes.
2943 * XXX common "target" functionality should be factored into separate module
2944 * shared by mdt, ost and stand-alone services like fld.
2946 static int mdt_handle_common(struct ptlrpc_request *req,
2947 struct mdt_opc_slice *supported)
2950 struct mdt_thread_info *info;
2954 env = req->rq_svc_thread->t_env;
2955 LASSERT(env != NULL);
2956 LASSERT(env->le_ses != NULL);
2957 LASSERT(env->le_ctx.lc_thread == req->rq_svc_thread);
2958 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
2959 LASSERT(info != NULL);
2961 mdt_thread_info_init(req, info);
2963 rc = mdt_handle0(req, info, supported);
2965 mdt_thread_info_fini(info);
2970 * This is called from recovery code as handler of _all_ RPC types, FLD and SEQ
2973 int mdt_recovery_handle(struct ptlrpc_request *req)
2978 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2980 rc = mdt_handle_common(req, mdt_fld_handlers);
2983 rc = mdt_handle_common(req, mdt_seq_handlers);
2986 rc = mdt_handle_common(req, mdt_regular_handlers);
2993 static int mdt_regular_handle(struct ptlrpc_request *req)
2995 return mdt_handle_common(req, mdt_regular_handlers);
2998 static int mdt_readpage_handle(struct ptlrpc_request *req)
3000 return mdt_handle_common(req, mdt_readpage_handlers);
3003 static int mdt_xmds_handle(struct ptlrpc_request *req)
3005 return mdt_handle_common(req, mdt_xmds_handlers);
3008 static int mdt_mdsc_handle(struct ptlrpc_request *req)
3010 return mdt_handle_common(req, mdt_seq_handlers);
3013 static int mdt_mdss_handle(struct ptlrpc_request *req)
3015 return mdt_handle_common(req, mdt_seq_handlers);
3018 static int mdt_dtss_handle(struct ptlrpc_request *req)
3020 return mdt_handle_common(req, mdt_seq_handlers);
3023 static int mdt_fld_handle(struct ptlrpc_request *req)
3025 return mdt_handle_common(req, mdt_fld_handlers);
3041 static int mdt_intent_getattr(enum mdt_it_code opcode,
3042 struct mdt_thread_info *info,
3043 struct ldlm_lock **,
3045 static int mdt_intent_reint(enum mdt_it_code opcode,
3046 struct mdt_thread_info *info,
3047 struct ldlm_lock **,
3050 static struct mdt_it_flavor {
3051 const struct req_format *it_fmt;
3053 int (*it_act)(enum mdt_it_code ,
3054 struct mdt_thread_info *,
3055 struct ldlm_lock **,
3058 } mdt_it_flavor[] = {
3060 .it_fmt = &RQF_LDLM_INTENT,
3061 /*.it_flags = HABEO_REFERO,*/
3063 .it_act = mdt_intent_reint,
3064 .it_reint = REINT_OPEN
3067 .it_fmt = &RQF_LDLM_INTENT,
3068 .it_flags = MUTABOR,
3069 .it_act = mdt_intent_reint,
3070 .it_reint = REINT_OPEN
3073 .it_fmt = &RQF_LDLM_INTENT,
3074 .it_flags = MUTABOR,
3075 .it_act = mdt_intent_reint,
3076 .it_reint = REINT_CREATE
3078 [MDT_IT_GETATTR] = {
3079 .it_fmt = &RQF_LDLM_INTENT_GETATTR,
3080 .it_flags = HABEO_REFERO,
3081 .it_act = mdt_intent_getattr
3083 [MDT_IT_READDIR] = {
3089 .it_fmt = &RQF_LDLM_INTENT_GETATTR,
3090 .it_flags = HABEO_REFERO,
3091 .it_act = mdt_intent_getattr
3094 .it_fmt = &RQF_LDLM_INTENT_UNLINK,
3095 .it_flags = MUTABOR,
3097 .it_reint = REINT_UNLINK
3101 .it_flags = MUTABOR,
3104 [MDT_IT_GETXATTR] = {
3111 int mdt_intent_lock_replace(struct mdt_thread_info *info,
3112 struct ldlm_lock **lockp,
3113 struct ldlm_lock *new_lock,
3114 struct mdt_lock_handle *lh,
3117 struct ptlrpc_request *req = mdt_info_req(info);
3118 struct ldlm_lock *lock = *lockp;
3121 * Get new lock only for cases when possible resent did not find any
3124 if (new_lock == NULL)
3125 new_lock = ldlm_handle2lock_long(&lh->mlh_reg_lh, 0);
3127 if (new_lock == NULL && (flags & LDLM_FL_INTENT_ONLY)) {
3128 lh->mlh_reg_lh.cookie = 0;
3132 LASSERTF(new_lock != NULL,
3133 "lockh "LPX64"\n", lh->mlh_reg_lh.cookie);
3136 * If we've already given this lock to a client once, then we should
3137 * have no readers or writers. Otherwise, we should have one reader
3138 * _or_ writer ref (which will be zeroed below) before returning the
3141 if (new_lock->l_export == req->rq_export) {
3142 LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
3144 LASSERT(new_lock->l_export == NULL);
3145 LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
3150 if (new_lock->l_export == req->rq_export) {
3152 * Already gave this to the client, which means that we
3153 * reconstructed a reply.
3155 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
3157 lh->mlh_reg_lh.cookie = 0;
3158 RETURN(ELDLM_LOCK_REPLACED);
3162 * Fixup the lock to be given to the client.
3164 lock_res_and_lock(new_lock);
3165 /* Zero new_lock->l_readers and new_lock->l_writers without triggering
3166 * possible blocking AST. */
3167 while (new_lock->l_readers > 0) {
3168 lu_ref_del(&new_lock->l_reference, "reader", new_lock);
3169 lu_ref_del(&new_lock->l_reference, "user", new_lock);
3170 new_lock->l_readers--;
3172 while (new_lock->l_writers > 0) {
3173 lu_ref_del(&new_lock->l_reference, "writer", new_lock);
3174 lu_ref_del(&new_lock->l_reference, "user", new_lock);
3175 new_lock->l_writers--;
3178 new_lock->l_export = class_export_get(req->rq_export);
3179 new_lock->l_blocking_ast = lock->l_blocking_ast;
3180 new_lock->l_completion_ast = lock->l_completion_ast;
3181 new_lock->l_remote_handle = lock->l_remote_handle;
3182 new_lock->l_flags &= ~LDLM_FL_LOCAL;
3184 unlock_res_and_lock(new_lock);
3186 lustre_hash_add(new_lock->l_export->exp_lock_hash,
3187 &new_lock->l_remote_handle,
3188 &new_lock->l_exp_hash);
3190 LDLM_LOCK_RELEASE(new_lock);
3191 lh->mlh_reg_lh.cookie = 0;
3193 RETURN(ELDLM_LOCK_REPLACED);
3196 static void mdt_intent_fixup_resent(struct mdt_thread_info *info,
3197 struct ldlm_lock *new_lock,
3198 struct ldlm_lock **old_lock,
3199 struct mdt_lock_handle *lh)
3201 struct ptlrpc_request *req = mdt_info_req(info);
3202 struct obd_export *exp = req->rq_export;
3203 struct lustre_handle remote_hdl;
3204 struct ldlm_request *dlmreq;
3205 struct ldlm_lock *lock;
3207 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
3210 dlmreq = req_capsule_client_get(info->mti_pill, &RMF_DLM_REQ);
3211 remote_hdl = dlmreq->lock_handle[0];
3213 lock = lustre_hash_lookup(exp->exp_lock_hash, &remote_hdl);
3215 if (lock != new_lock) {
3216 lh->mlh_reg_lh.cookie = lock->l_handle.h_cookie;
3217 lh->mlh_reg_mode = lock->l_granted_mode;
3219 LDLM_DEBUG(lock, "Restoring lock cookie");
3220 DEBUG_REQ(D_DLMTRACE, req,
3221 "restoring lock cookie "LPX64,
3222 lh->mlh_reg_lh.cookie);
3224 *old_lock = LDLM_LOCK_GET(lock);
3225 lh_put(exp->exp_lock_hash, &lock->l_exp_hash);
3229 lh_put(exp->exp_lock_hash, &lock->l_exp_hash);
3233 * If the xid matches, then we know this is a resent request, and allow
3234 * it. (It's probably an OPEN, for which we don't send a lock.
3236 if (req_xid_is_last(req))
3240 * This remote handle isn't enqueued, so we never received or processed
3241 * this request. Clear MSG_RESENT, because it can be handled like any
3242 * normal request now.
3244 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
3246 DEBUG_REQ(D_DLMTRACE, req, "no existing lock with rhandle "LPX64,
3250 static int mdt_intent_getattr(enum mdt_it_code opcode,
3251 struct mdt_thread_info *info,
3252 struct ldlm_lock **lockp,
3255 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
3256 struct ldlm_lock *new_lock = NULL;
3258 struct ldlm_reply *ldlm_rep;
3259 struct ptlrpc_request *req;
3260 struct mdt_body *reqbody;
3261 struct mdt_body *repbody;
3265 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
3268 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
3271 info->mti_spec.sp_ck_split = !!(reqbody->valid & OBD_MD_FLCKSPLIT);
3272 info->mti_cross_ref = !!(reqbody->valid & OBD_MD_FLCROSSREF);
3273 repbody->eadatasize = 0;
3274 repbody->aclsize = 0;
3278 child_bits = MDS_INODELOCK_LOOKUP;
3280 case MDT_IT_GETATTR:
3281 child_bits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
3284 CERROR("Unhandled till now");
3285 GOTO(out_shrink, rc = -EINVAL);
3288 rc = mdt_init_ucred(info, reqbody);
3290 GOTO(out_shrink, rc);
3292 req = info->mti_pill->rc_req;
3293 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
3294 mdt_set_disposition(info, ldlm_rep, DISP_IT_EXECD);
3296 /* Get lock from request for possible resent case. */
3297 mdt_intent_fixup_resent(info, *lockp, &new_lock, lhc);
3299 ldlm_rep->lock_policy_res2 =
3300 mdt_getattr_name_lock(info, lhc, child_bits, ldlm_rep);
3302 if (mdt_get_disposition(ldlm_rep, DISP_LOOKUP_NEG))
3303 ldlm_rep->lock_policy_res2 = 0;
3304 if (!mdt_get_disposition(ldlm_rep, DISP_LOOKUP_POS) ||
3305 ldlm_rep->lock_policy_res2) {
3306 lhc->mlh_reg_lh.cookie = 0ull;
3307 GOTO(out_ucred, rc = ELDLM_LOCK_ABORTED);
3310 rc = mdt_intent_lock_replace(info, lockp, new_lock, lhc, flags);
3313 mdt_exit_ucred(info);
3315 mdt_shrink_reply(info);
3319 static int mdt_intent_reint(enum mdt_it_code opcode,
3320 struct mdt_thread_info *info,
3321 struct ldlm_lock **lockp,
3324 struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
3325 struct ldlm_reply *rep = NULL;
3329 static const struct req_format *intent_fmts[REINT_MAX] = {
3330 [REINT_CREATE] = &RQF_LDLM_INTENT_CREATE,
3331 [REINT_OPEN] = &RQF_LDLM_INTENT_OPEN
3336 opc = mdt_reint_opcode(info, intent_fmts);
3340 if (mdt_it_flavor[opcode].it_reint != opc) {
3341 CERROR("Reint code %ld doesn't match intent: %d\n",
3343 RETURN(err_serious(-EPROTO));
3346 /* Get lock from request for possible resent case. */
3347 mdt_intent_fixup_resent(info, *lockp, NULL, lhc);
3349 rc = mdt_reint_internal(info, lhc, opc);
3351 /* Check whether the reply has been packed successfully. */
3352 if (mdt_info_req(info)->rq_repmsg != NULL)
3353 rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
3355 RETURN(err_serious(-EFAULT));
3357 /* MDC expects this in any case */
3359 mdt_set_disposition(info, rep, DISP_LOOKUP_EXECD);
3361 /* Cross-ref case, the lock should be returned to the client */
3362 if (rc == -EREMOTE) {
3363 LASSERT(lustre_handle_is_used(&lhc->mlh_reg_lh));
3364 rep->lock_policy_res2 = 0;
3365 rc = mdt_intent_lock_replace(info, lockp, NULL, lhc, flags);
3368 rep->lock_policy_res2 = clear_serious(rc);
3370 if (rc == -ENOTCONN || rc == -ENODEV ||
3371 rc == -EOVERFLOW) { /**< if VBR failure then return error */
3373 * If it is the disconnect error (ENODEV & ENOCONN), the error
3374 * will be returned by rq_status, and client at ptlrpc layer
3375 * will detect this, then disconnect, reconnect the import
3376 * immediately, instead of impacting the following the rpc.
3378 lhc->mlh_reg_lh.cookie = 0ull;
3382 * For other cases, the error will be returned by intent.
3383 * and client will retrieve the result from intent.
3386 * FIXME: when open lock is finished, that should be
3389 if (lustre_handle_is_used(&lhc->mlh_reg_lh)) {
3390 rep->lock_policy_res2 = 0;
3391 rc = mdt_intent_lock_replace(info, lockp, NULL, lhc, flags);
3394 lhc->mlh_reg_lh.cookie = 0ull;
3395 RETURN(ELDLM_LOCK_ABORTED);
3400 static int mdt_intent_code(long itcode)
3408 case IT_OPEN|IT_CREAT:
3415 rc = MDT_IT_READDIR;
3418 rc = MDT_IT_GETATTR;
3430 rc = MDT_IT_GETXATTR;
3433 CERROR("Unknown intent opcode: %ld\n", itcode);
3440 static int mdt_intent_opc(long itopc, struct mdt_thread_info *info,
3441 struct ldlm_lock **lockp, int flags)
3443 struct req_capsule *pill;
3444 struct mdt_it_flavor *flv;
3449 opc = mdt_intent_code(itopc);
3453 pill = info->mti_pill;
3454 flv = &mdt_it_flavor[opc];
3456 if (flv->it_fmt != NULL)
3457 req_capsule_extend(pill, flv->it_fmt);
3459 rc = mdt_unpack_req_pack_rep(info, flv->it_flags);
3461 struct ptlrpc_request *req = mdt_info_req(info);
3462 if (flv->it_flags & MUTABOR &&
3463 req->rq_export->exp_connect_flags & OBD_CONNECT_RDONLY)
3466 if (rc == 0 && flv->it_act != NULL) {
3467 /* execute policy */
3468 rc = flv->it_act(opc, info, lockp, flags);
3475 static int mdt_intent_policy(struct ldlm_namespace *ns,
3476 struct ldlm_lock **lockp, void *req_cookie,
3477 ldlm_mode_t mode, int flags, void *data)
3479 struct mdt_thread_info *info;
3480 struct ptlrpc_request *req = req_cookie;
3481 struct ldlm_intent *it;
3482 struct req_capsule *pill;
3487 LASSERT(req != NULL);
3489 info = lu_context_key_get(&req->rq_svc_thread->t_env->le_ctx,
3491 LASSERT(info != NULL);
3492 pill = info->mti_pill;
3493 LASSERT(pill->rc_req == req);
3495 if (req->rq_reqmsg->lm_bufcount > DLM_INTENT_IT_OFF) {
3496 req_capsule_extend(pill, &RQF_LDLM_INTENT);
3497 it = req_capsule_client_get(pill, &RMF_LDLM_INTENT);
3499 const struct ldlm_request *dlmreq;
3502 rc = mdt_intent_opc(it->opc, info, lockp, flags);
3507 * Lock without inodebits makes no sense and will oops
3508 * later in ldlm. Let's check it now to see if we have
3509 * wrong lock from client or bits get corrupted
3510 * somewhere in mdt_intent_opc().
3512 dlmreq = info->mti_dlm_req;
3513 req_bits = dlmreq->lock_desc.l_policy_data.l_inodebits.bits;
3514 LASSERT(req_bits != 0);
3517 rc = err_serious(-EFAULT);
3519 /* No intent was provided */
3520 LASSERT(pill->rc_fmt == &RQF_LDLM_ENQUEUE);
3521 rc = req_capsule_server_pack(pill);
3523 rc = err_serious(rc);
3531 static void mdt_seq_adjust(const struct lu_env *env,
3532 struct mdt_device *m, int lost)
3534 struct md_site *ms = mdt_md_site(m);
3535 struct lu_seq_range out;
3538 LASSERT(ms && ms->ms_server_seq);
3540 /* get extra seq from seq_server, moving it's range up */
3541 while (lost-- > 0) {
3542 seq_server_alloc_meta(ms->ms_server_seq, NULL, &out, env);
3547 static int mdt_seq_fini(const struct lu_env *env,
3548 struct mdt_device *m)
3550 struct md_site *ms = mdt_md_site(m);
3554 if (ms->ms_server_seq) {
3555 seq_server_fini(ms->ms_server_seq, env);
3556 OBD_FREE_PTR(ms->ms_server_seq);
3557 ms->ms_server_seq = NULL;
3560 if (ms->ms_control_seq) {
3561 seq_server_fini(ms->ms_control_seq, env);
3562 OBD_FREE_PTR(ms->ms_control_seq);
3563 ms->ms_control_seq = NULL;
3566 if (ms->ms_client_seq) {
3567 seq_client_fini(ms->ms_client_seq);
3568 OBD_FREE_PTR(ms->ms_client_seq);
3569 ms->ms_client_seq = NULL;
3576 static int mdt_seq_init(const struct lu_env *env,
3578 struct mdt_device *m)
3585 ms = mdt_md_site(m);
3588 * This is sequence-controller node. Init seq-controller server on local
3591 if (ms->ms_node_id == 0) {
3592 LASSERT(ms->ms_control_seq == NULL);
3594 OBD_ALLOC_PTR(ms->ms_control_seq);
3595 if (ms->ms_control_seq == NULL)
3598 rc = seq_server_init(ms->ms_control_seq,
3599 m->mdt_bottom, uuid,
3600 LUSTRE_SEQ_CONTROLLER,
3605 GOTO(out_seq_fini, rc);
3607 OBD_ALLOC_PTR(ms->ms_client_seq);
3608 if (ms->ms_client_seq == NULL)
3609 GOTO(out_seq_fini, rc = -ENOMEM);
3611 OBD_ALLOC(prefix, MAX_OBD_NAME + 5);
3612 if (prefix == NULL) {
3613 OBD_FREE_PTR(ms->ms_client_seq);
3614 GOTO(out_seq_fini, rc = -ENOMEM);
3617 snprintf(prefix, MAX_OBD_NAME + 5, "ctl-%s",
3621 * Init seq-controller client after seq-controller server is
3622 * ready. Pass ms->ms_control_seq to it for direct talking.
3624 rc = seq_client_init(ms->ms_client_seq, NULL,
3625 LUSTRE_SEQ_METADATA, prefix,
3626 ms->ms_control_seq);
3627 OBD_FREE(prefix, MAX_OBD_NAME + 5);
3630 GOTO(out_seq_fini, rc);
3633 /* Init seq-server on local MDT */
3634 LASSERT(ms->ms_server_seq == NULL);
3636 OBD_ALLOC_PTR(ms->ms_server_seq);
3637 if (ms->ms_server_seq == NULL)
3638 GOTO(out_seq_fini, rc = -ENOMEM);
3640 rc = seq_server_init(ms->ms_server_seq,
3641 m->mdt_bottom, uuid,
3646 GOTO(out_seq_fini, rc = -ENOMEM);
3648 /* Assign seq-controller client to local seq-server. */
3649 if (ms->ms_node_id == 0) {
3650 LASSERT(ms->ms_client_seq != NULL);
3652 rc = seq_server_set_cli(ms->ms_server_seq,
3660 mdt_seq_fini(env, m);
3665 * Init client sequence manager which is used by local MDS to talk to sequence
3666 * controller on remote node.
3668 static int mdt_seq_init_cli(const struct lu_env *env,
3669 struct mdt_device *m,
3670 struct lustre_cfg *cfg)
3672 struct md_site *ms = mdt_md_site(m);
3673 struct obd_device *mdc;
3674 struct obd_uuid *uuidp, *mdcuuidp;
3675 char *uuid_str, *mdc_uuid_str;
3678 struct mdt_thread_info *info;
3679 char *p, *index_string = lustre_cfg_string(cfg, 2);
3682 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
3683 uuidp = &info->mti_u.uuid[0];
3684 mdcuuidp = &info->mti_u.uuid[1];
3686 LASSERT(index_string);
3688 index = simple_strtol(index_string, &p, 10);
3690 CERROR("Invalid index in lustre_cgf, offset 2\n");
3694 /* check if this is adding the first MDC and controller is not yet
3696 if (index != 0 || ms->ms_client_seq)
3699 uuid_str = lustre_cfg_string(cfg, 1);
3700 mdc_uuid_str = lustre_cfg_string(cfg, 4);
3701 obd_str2uuid(uuidp, uuid_str);
3702 obd_str2uuid(mdcuuidp, mdc_uuid_str);
3704 mdc = class_find_client_obd(uuidp, LUSTRE_MDC_NAME, mdcuuidp);
3706 CERROR("can't find controller MDC by uuid %s\n",
3709 } else if (!mdc->obd_set_up) {
3710 CERROR("target %s not set up\n", mdc->obd_name);
3713 LASSERT(ms->ms_control_exp);
3714 OBD_ALLOC_PTR(ms->ms_client_seq);
3715 if (ms->ms_client_seq != NULL) {
3718 OBD_ALLOC(prefix, MAX_OBD_NAME + 5);
3722 snprintf(prefix, MAX_OBD_NAME + 5, "ctl-%s",
3725 rc = seq_client_init(ms->ms_client_seq,
3727 LUSTRE_SEQ_METADATA,
3729 OBD_FREE(prefix, MAX_OBD_NAME + 5);
3736 LASSERT(ms->ms_server_seq != NULL);
3737 rc = seq_server_set_cli(ms->ms_server_seq, ms->ms_client_seq,
3744 static void mdt_seq_fini_cli(struct mdt_device *m)
3750 ms = mdt_md_site(m);
3753 if (ms->ms_server_seq)
3754 seq_server_set_cli(ms->ms_server_seq,
3757 if (ms->ms_control_exp) {
3758 class_export_put(ms->ms_control_exp);
3759 ms->ms_control_exp = NULL;
3768 static int mdt_fld_fini(const struct lu_env *env,
3769 struct mdt_device *m)
3771 struct md_site *ms = mdt_md_site(m);
3774 if (ms && ms->ms_server_fld) {
3775 fld_server_fini(ms->ms_server_fld, env);
3776 OBD_FREE_PTR(ms->ms_server_fld);
3777 ms->ms_server_fld = NULL;
3783 static int mdt_fld_init(const struct lu_env *env,
3785 struct mdt_device *m)
3791 ms = mdt_md_site(m);
3793 OBD_ALLOC_PTR(ms->ms_server_fld);
3794 if (ms->ms_server_fld == NULL)
3795 RETURN(rc = -ENOMEM);
3797 rc = fld_server_init(ms->ms_server_fld,
3798 m->mdt_bottom, uuid,
3799 env, ms->ms_node_id);
3801 OBD_FREE_PTR(ms->ms_server_fld);
3802 ms->ms_server_fld = NULL;
3809 /* device init/fini methods */
3810 static void mdt_stop_ptlrpc_service(struct mdt_device *m)
3813 if (m->mdt_regular_service != NULL) {
3814 ptlrpc_unregister_service(m->mdt_regular_service);
3815 m->mdt_regular_service = NULL;
3817 if (m->mdt_readpage_service != NULL) {
3818 ptlrpc_unregister_service(m->mdt_readpage_service);
3819 m->mdt_readpage_service = NULL;
3821 if (m->mdt_xmds_service != NULL) {
3822 ptlrpc_unregister_service(m->mdt_xmds_service);
3823 m->mdt_xmds_service = NULL;
3825 if (m->mdt_setattr_service != NULL) {
3826 ptlrpc_unregister_service(m->mdt_setattr_service);
3827 m->mdt_setattr_service = NULL;
3829 if (m->mdt_mdsc_service != NULL) {
3830 ptlrpc_unregister_service(m->mdt_mdsc_service);
3831 m->mdt_mdsc_service = NULL;
3833 if (m->mdt_mdss_service != NULL) {
3834 ptlrpc_unregister_service(m->mdt_mdss_service);
3835 m->mdt_mdss_service = NULL;
3837 if (m->mdt_dtss_service != NULL) {
3838 ptlrpc_unregister_service(m->mdt_dtss_service);
3839 m->mdt_dtss_service = NULL;
3841 if (m->mdt_fld_service != NULL) {
3842 ptlrpc_unregister_service(m->mdt_fld_service);
3843 m->mdt_fld_service = NULL;
3848 static int mdt_start_ptlrpc_service(struct mdt_device *m)
3851 static struct ptlrpc_service_conf conf;
3852 cfs_proc_dir_entry_t *procfs_entry;
3855 procfs_entry = m->mdt_md_dev.md_lu_dev.ld_obd->obd_proc_entry;
3857 conf = (typeof(conf)) {
3858 .psc_nbufs = MDS_NBUFS,
3859 .psc_bufsize = MDS_BUFSIZE,
3860 .psc_max_req_size = MDS_MAXREQSIZE,
3861 .psc_max_reply_size = MDS_MAXREPSIZE,
3862 .psc_req_portal = MDS_REQUEST_PORTAL,
3863 .psc_rep_portal = MDC_REPLY_PORTAL,
3864 .psc_watchdog_factor = MDT_SERVICE_WATCHDOG_FACTOR,
3866 * We'd like to have a mechanism to set this on a per-device
3867 * basis, but alas...
3869 .psc_min_threads = min(max(mdt_num_threads, MDT_MIN_THREADS),
3871 .psc_max_threads = MDT_MAX_THREADS,
3872 .psc_ctx_tags = LCT_MD_THREAD
3875 m->mdt_ldlm_client = &m->mdt_md_dev.md_lu_dev.ld_obd->obd_ldlm_client;
3876 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
3877 "mdt_ldlm_client", m->mdt_ldlm_client);
3879 m->mdt_regular_service =
3880 ptlrpc_init_svc_conf(&conf, mdt_regular_handle, LUSTRE_MDT_NAME,
3881 procfs_entry, target_print_req,
3883 if (m->mdt_regular_service == NULL)
3886 rc = ptlrpc_start_threads(NULL, m->mdt_regular_service);
3888 GOTO(err_mdt_svc, rc);
3891 * readpage service configuration. Parameters have to be adjusted,
3894 conf = (typeof(conf)) {
3895 .psc_nbufs = MDS_NBUFS,
3896 .psc_bufsize = MDS_BUFSIZE,
3897 .psc_max_req_size = MDS_MAXREQSIZE,
3898 .psc_max_reply_size = MDS_MAXREPSIZE,
3899 .psc_req_portal = MDS_READPAGE_PORTAL,
3900 .psc_rep_portal = MDC_REPLY_PORTAL,
3901 .psc_watchdog_factor = MDT_SERVICE_WATCHDOG_FACTOR,
3902 .psc_min_threads = min(max(mdt_num_threads, MDT_MIN_THREADS),
3904 .psc_max_threads = MDT_MAX_THREADS,
3905 .psc_ctx_tags = LCT_MD_THREAD
3907 m->mdt_readpage_service =
3908 ptlrpc_init_svc_conf(&conf, mdt_readpage_handle,
3909 LUSTRE_MDT_NAME "_readpage",
3910 procfs_entry, target_print_req,"mdt_rdpg");
3912 if (m->mdt_readpage_service == NULL) {
3913 CERROR("failed to start readpage service\n");
3914 GOTO(err_mdt_svc, rc = -ENOMEM);
3917 rc = ptlrpc_start_threads(NULL, m->mdt_readpage_service);
3920 * setattr service configuration.
3922 conf = (typeof(conf)) {
3923 .psc_nbufs = MDS_NBUFS,
3924 .psc_bufsize = MDS_BUFSIZE,
3925 .psc_max_req_size = MDS_MAXREQSIZE,
3926 .psc_max_reply_size = MDS_MAXREPSIZE,
3927 .psc_req_portal = MDS_SETATTR_PORTAL,
3928 .psc_rep_portal = MDC_REPLY_PORTAL,
3929 .psc_watchdog_factor = MDT_SERVICE_WATCHDOG_FACTOR,
3930 .psc_min_threads = min(max(mdt_num_threads, MDT_MIN_THREADS),
3932 .psc_max_threads = MDT_MAX_THREADS,
3933 .psc_ctx_tags = LCT_MD_THREAD
3936 m->mdt_setattr_service =
3937 ptlrpc_init_svc_conf(&conf, mdt_regular_handle,
3938 LUSTRE_MDT_NAME "_setattr",
3939 procfs_entry, target_print_req,"mdt_attr");
3941 if (!m->mdt_setattr_service) {
3942 CERROR("failed to start setattr service\n");
3943 GOTO(err_mdt_svc, rc = -ENOMEM);
3946 rc = ptlrpc_start_threads(NULL, m->mdt_setattr_service);
3948 GOTO(err_mdt_svc, rc);
3951 * sequence controller service configuration
3953 conf = (typeof(conf)) {
3954 .psc_nbufs = MDS_NBUFS,
3955 .psc_bufsize = MDS_BUFSIZE,
3956 .psc_max_req_size = SEQ_MAXREQSIZE,
3957 .psc_max_reply_size = SEQ_MAXREPSIZE,
3958 .psc_req_portal = SEQ_CONTROLLER_PORTAL,
3959 .psc_rep_portal = MDC_REPLY_PORTAL,
3960 .psc_watchdog_factor = MDT_SERVICE_WATCHDOG_FACTOR,
3961 .psc_min_threads = SEQ_NUM_THREADS,
3962 .psc_max_threads = SEQ_NUM_THREADS,
3963 .psc_ctx_tags = LCT_MD_THREAD|LCT_DT_THREAD
3966 m->mdt_mdsc_service =
3967 ptlrpc_init_svc_conf(&conf, mdt_mdsc_handle,
3968 LUSTRE_MDT_NAME"_mdsc",
3969 procfs_entry, target_print_req,"mdt_mdsc");
3970 if (!m->mdt_mdsc_service) {
3971 CERROR("failed to start seq controller service\n");
3972 GOTO(err_mdt_svc, rc = -ENOMEM);
3975 rc = ptlrpc_start_threads(NULL, m->mdt_mdsc_service);
3977 GOTO(err_mdt_svc, rc);
3980 * metadata sequence server service configuration
3982 conf = (typeof(conf)) {
3983 .psc_nbufs = MDS_NBUFS,
3984 .psc_bufsize = MDS_BUFSIZE,
3985 .psc_max_req_size = SEQ_MAXREQSIZE,
3986 .psc_max_reply_size = SEQ_MAXREPSIZE,
3987 .psc_req_portal = SEQ_METADATA_PORTAL,
3988 .psc_rep_portal = MDC_REPLY_PORTAL,
3989 .psc_watchdog_factor = MDT_SERVICE_WATCHDOG_FACTOR,
3990 .psc_min_threads = SEQ_NUM_THREADS,
3991 .psc_max_threads = SEQ_NUM_THREADS,
3992 .psc_ctx_tags = LCT_MD_THREAD|LCT_DT_THREAD
3995 m->mdt_mdss_service =
3996 ptlrpc_init_svc_conf(&conf, mdt_mdss_handle,
3997 LUSTRE_MDT_NAME"_mdss",
3998 procfs_entry, target_print_req,"mdt_mdss");
3999 if (!m->mdt_mdss_service) {
4000 CERROR("failed to start metadata seq server service\n");
4001 GOTO(err_mdt_svc, rc = -ENOMEM);
4004 rc = ptlrpc_start_threads(NULL, m->mdt_mdss_service);
4006 GOTO(err_mdt_svc, rc);
4010 * Data sequence server service configuration. We want to have really
4011 * cluster-wide sequences space. This is why we start only one sequence
4012 * controller which manages space.
4014 conf = (typeof(conf)) {
4015 .psc_nbufs = MDS_NBUFS,
4016 .psc_bufsize = MDS_BUFSIZE,
4017 .psc_max_req_size = SEQ_MAXREQSIZE,
4018 .psc_max_reply_size = SEQ_MAXREPSIZE,
4019 .psc_req_portal = SEQ_DATA_PORTAL,
4020 .psc_rep_portal = OSC_REPLY_PORTAL,
4021 .psc_watchdog_factor = MDT_SERVICE_WATCHDOG_FACTOR,
4022 .psc_min_threads = SEQ_NUM_THREADS,
4023 .psc_max_threads = SEQ_NUM_THREADS,
4024 .psc_ctx_tags = LCT_MD_THREAD|LCT_DT_THREAD
4027 m->mdt_dtss_service =
4028 ptlrpc_init_svc_conf(&conf, mdt_dtss_handle,
4029 LUSTRE_MDT_NAME"_dtss",
4030 procfs_entry, target_print_req,"mdt_dtss");
4031 if (!m->mdt_dtss_service) {
4032 CERROR("failed to start data seq server service\n");
4033 GOTO(err_mdt_svc, rc = -ENOMEM);
4036 rc = ptlrpc_start_threads(NULL, m->mdt_dtss_service);
4038 GOTO(err_mdt_svc, rc);
4040 /* FLD service start */
4041 conf = (typeof(conf)) {
4042 .psc_nbufs = MDS_NBUFS,
4043 .psc_bufsize = MDS_BUFSIZE,
4044 .psc_max_req_size = FLD_MAXREQSIZE,
4045 .psc_max_reply_size = FLD_MAXREPSIZE,
4046 .psc_req_portal = FLD_REQUEST_PORTAL,
4047 .psc_rep_portal = MDC_REPLY_PORTAL,
4048 .psc_watchdog_factor = MDT_SERVICE_WATCHDOG_FACTOR,
4049 .psc_min_threads = FLD_NUM_THREADS,
4050 .psc_max_threads = FLD_NUM_THREADS,
4051 .psc_ctx_tags = LCT_DT_THREAD|LCT_MD_THREAD
4054 m->mdt_fld_service =
4055 ptlrpc_init_svc_conf(&conf, mdt_fld_handle,
4056 LUSTRE_MDT_NAME"_fld",
4057 procfs_entry, target_print_req, "mdt_fld");
4058 if (!m->mdt_fld_service) {
4059 CERROR("failed to start fld service\n");
4060 GOTO(err_mdt_svc, rc = -ENOMEM);
4063 rc = ptlrpc_start_threads(NULL, m->mdt_fld_service);
4065 GOTO(err_mdt_svc, rc);
4068 * mds-mds service configuration. Separate portal is used to allow
4069 * mds-mds requests be not blocked during recovery.
4071 conf = (typeof(conf)) {
4072 .psc_nbufs = MDS_NBUFS,
4073 .psc_bufsize = MDS_BUFSIZE,
4074 .psc_max_req_size = MDS_MAXREQSIZE,
4075 .psc_max_reply_size = MDS_MAXREPSIZE,
4076 .psc_req_portal = MDS_MDS_PORTAL,
4077 .psc_rep_portal = MDC_REPLY_PORTAL,
4078 .psc_watchdog_factor = MDT_SERVICE_WATCHDOG_FACTOR,
4079 .psc_min_threads = min(max(mdt_num_threads, MDT_MIN_THREADS),
4081 .psc_max_threads = MDT_MAX_THREADS,
4082 .psc_ctx_tags = LCT_MD_THREAD
4084 m->mdt_xmds_service =
4085 ptlrpc_init_svc_conf(&conf, mdt_xmds_handle,
4086 LUSTRE_MDT_NAME "_mds",
4087 procfs_entry, target_print_req,"mdt_xmds");
4089 if (m->mdt_xmds_service == NULL) {
4090 CERROR("failed to start readpage service\n");
4091 GOTO(err_mdt_svc, rc = -ENOMEM);
4094 rc = ptlrpc_start_threads(NULL, m->mdt_xmds_service);
4096 GOTO(err_mdt_svc, rc);
4101 mdt_stop_ptlrpc_service(m);
4106 static void mdt_stack_fini(const struct lu_env *env,
4107 struct mdt_device *m, struct lu_device *top)
4109 struct obd_device *obd = mdt2obd_dev(m);
4110 struct lustre_cfg_bufs *bufs;
4111 struct lustre_cfg *lcfg;
4112 struct mdt_thread_info *info;
4116 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
4117 LASSERT(info != NULL);
4119 bufs = &info->mti_u.bufs;
4120 /* process cleanup, pass mdt obd name to get obd umount flags */
4121 lustre_cfg_bufs_reset(bufs, obd->obd_name);
4126 lustre_cfg_bufs_set_string(bufs, 1, flags);
4127 lcfg = lustre_cfg_new(LCFG_CLEANUP, bufs);
4129 CERROR("Cannot alloc lcfg!\n");
4134 top->ld_ops->ldo_process_config(env, top, lcfg);
4135 lustre_cfg_free(lcfg);
4137 lu_stack_fini(env, top);
4138 m->mdt_child = NULL;
4139 m->mdt_bottom = NULL;
4142 static struct lu_device *mdt_layer_setup(struct lu_env *env,
4143 const char *typename,
4144 struct lu_device *child,
4145 struct lustre_cfg *cfg)
4147 const char *dev = lustre_cfg_string(cfg, 0);
4148 struct obd_type *type;
4149 struct lu_device_type *ldt;
4150 struct lu_device *d;
4155 type = class_get_type(typename);
4157 CERROR("Unknown type: '%s'\n", typename);
4158 GOTO(out, rc = -ENODEV);
4161 rc = lu_env_refill((struct lu_env *)env);
4163 CERROR("Failure to refill session: '%d'\n", rc);
4169 CERROR("type: '%s'\n", typename);
4170 GOTO(out_type, rc = -EINVAL);
4173 ldt->ldt_obd_type = type;
4174 d = ldt->ldt_ops->ldto_device_alloc(env, ldt, cfg);
4176 CERROR("Cannot allocate device: '%s'\n", typename);
4177 GOTO(out_type, rc = -ENODEV);
4180 LASSERT(child->ld_site);
4181 d->ld_site = child->ld_site;
4184 rc = ldt->ldt_ops->ldto_device_init(env, d, dev, child);
4186 CERROR("can't init device '%s', rc %d\n", typename, rc);
4187 GOTO(out_alloc, rc);
4190 lu_ref_add(&d->ld_reference, "lu-stack", &lu_site_init);
4195 ldt->ldt_ops->ldto_device_free(env, d);
4198 class_put_type(type);
4203 static int mdt_stack_init(struct lu_env *env,
4204 struct mdt_device *m,
4205 struct lustre_cfg *cfg,
4206 struct lustre_mount_info *lmi)
4208 struct lu_device *d = &m->mdt_md_dev.md_lu_dev;
4209 struct lu_device *tmp;
4210 struct md_device *md;
4211 struct lu_device *child_lu_dev;
4215 /* init the stack */
4216 tmp = mdt_layer_setup(env, LUSTRE_OSD_NAME, d, cfg);
4218 RETURN(PTR_ERR(tmp));
4220 m->mdt_bottom = lu2dt_dev(tmp);
4222 tmp = mdt_layer_setup(env, LUSTRE_MDD_NAME, d, cfg);
4224 GOTO(out, rc = PTR_ERR(tmp));
4229 tmp = mdt_layer_setup(env, LUSTRE_CMM_NAME, d, cfg);
4231 GOTO(out, rc = PTR_ERR(tmp));
4234 /*set mdd upcall device*/
4235 md_upcall_dev_set(md, lu2md_dev(d));
4238 /*set cmm upcall device*/
4239 md_upcall_dev_set(md, &m->mdt_md_dev);
4241 m->mdt_child = lu2md_dev(d);
4243 /* process setup config */
4244 tmp = &m->mdt_md_dev.md_lu_dev;
4245 rc = tmp->ld_ops->ldo_process_config(env, tmp, cfg);
4249 /* initialize local objects */
4250 child_lu_dev = &m->mdt_child->md_lu_dev;
4252 rc = child_lu_dev->ld_ops->ldo_prepare(env,
4253 &m->mdt_md_dev.md_lu_dev,
4256 /* fini from last known good lu_device */
4258 mdt_stack_fini(env, m, d);
4264 * setup CONFIG_ORIG context, used to access local config log.
4265 * this may need to be rewrite as part of llog rewrite for lu-api.
4267 static int mdt_obd_llog_setup(struct obd_device *obd,
4268 struct lustre_sb_info *lsi)
4272 LASSERT(obd->obd_fsops == NULL);
4274 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
4275 if (IS_ERR(obd->obd_fsops))
4276 return PTR_ERR(obd->obd_fsops);
4278 rc = fsfilt_setup(obd, lsi->lsi_srv_mnt->mnt_sb);
4280 fsfilt_put_ops(obd->obd_fsops);
4284 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
4285 obd->obd_lvfs_ctxt.pwdmnt = lsi->lsi_srv_mnt;
4286 obd->obd_lvfs_ctxt.pwd = lsi->lsi_srv_mnt->mnt_root;
4287 obd->obd_lvfs_ctxt.fs = get_ds();
4289 rc = llog_setup(obd, &obd->obd_olg, LLOG_CONFIG_ORIG_CTXT, obd,
4290 0, NULL, &llog_lvfs_ops);
4292 CERROR("llog_setup() failed: %d\n", rc);
4293 fsfilt_put_ops(obd->obd_fsops);
4299 static void mdt_obd_llog_cleanup(struct obd_device *obd)
4301 struct llog_ctxt *ctxt;
4303 ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
4307 if (obd->obd_fsops) {
4308 fsfilt_put_ops(obd->obd_fsops);
4309 obd->obd_fsops = NULL;
4313 static void mdt_fini(const struct lu_env *env, struct mdt_device *m)
4315 struct md_device *next = m->mdt_child;
4316 struct lu_device *d = &m->mdt_md_dev.md_lu_dev;
4317 struct lu_site *ls = d->ld_site;
4318 struct obd_device *obd = mdt2obd_dev(m);
4322 target_recovery_fini(obd);
4323 /* At this point, obd exports might still be on the "obd_zombie_exports"
4324 * list, and obd_zombie_impexp_thread() is trying to destroy them.
4325 * We wait a little bit until all exports (except the self-export)
4326 * have been destroyed, because the whole mdt stack might be accessed
4327 * in mdt_destroy_export(). This will not be a long time, maybe one or
4328 * two seconds are enough. This is not a problem while umounting.
4330 * The three references that should be remaining are the
4331 * obd_self_export and the attach and setup references.
4333 while (atomic_read(&obd->obd_refcount) > 3) {
4334 cfs_schedule_timeout(CFS_TASK_UNINT, cfs_time_seconds(1));
4336 if (waited > 5 && IS_PO2(waited))
4337 LCONSOLE_WARN("Waiting for obd_zombie_impexp_thread "
4338 "more than %d seconds to destroy all "
4339 "the exports. The current obd refcount ="
4340 " %d. Is it stuck there?\n",
4341 waited, atomic_read(&obd->obd_refcount));
4344 ping_evictor_stop();
4346 mdt_stop_ptlrpc_service(m);
4347 mdt_llog_ctxt_unclone(env, m, LLOG_CHANGELOG_ORIG_CTXT);
4348 mdt_obd_llog_cleanup(obd);
4349 obd_zombie_barrier();
4350 #ifdef HAVE_QUOTA_SUPPORT
4351 next->md_ops->mdo_quota.mqo_cleanup(env, next);
4353 lut_fini(env, &m->mdt_lut);
4354 mdt_fs_cleanup(env, m);
4355 upcall_cache_cleanup(m->mdt_identity_cache);
4356 m->mdt_identity_cache = NULL;
4358 if (m->mdt_namespace != NULL) {
4359 ldlm_namespace_free(m->mdt_namespace, NULL, d->ld_obd->obd_force);
4360 d->ld_obd->obd_namespace = m->mdt_namespace = NULL;
4363 cfs_free_nidlist(&m->mdt_nosquash_nids);
4364 if (m->mdt_nosquash_str) {
4365 OBD_FREE(m->mdt_nosquash_str, m->mdt_nosquash_strlen);
4366 m->mdt_nosquash_str = NULL;
4367 m->mdt_nosquash_strlen = 0;
4370 mdt_seq_fini(env, m);
4371 mdt_seq_fini_cli(m);
4372 mdt_fld_fini(env, m);
4373 sptlrpc_rule_set_free(&m->mdt_sptlrpc_rset);
4375 next->md_ops->mdo_init_capa_ctxt(env, next, 0, 0, 0, NULL);
4376 cfs_timer_disarm(&m->mdt_ck_timer);
4377 mdt_ck_thread_stop(m);
4382 mdt_stack_fini(env, m, md2lu_dev(m->mdt_child));
4385 if (obd->obd_proc_exports_entry) {
4386 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
4387 obd->obd_proc_exports_entry = NULL;
4389 lprocfs_free_per_client_stats(obd);
4390 lprocfs_free_obd_stats(obd);
4391 ptlrpc_lprocfs_unregister_obd(obd);
4392 lprocfs_obd_cleanup(obd);
4395 struct md_site *mite;
4398 mite = lu_site2md(ls);
4402 LASSERT(atomic_read(&d->ld_ref) == 0);
4407 static int mdt_adapt_sptlrpc_conf(struct obd_device *obd, int initial)
4409 struct mdt_device *m = mdt_dev(obd->obd_lu_dev);
4410 struct sptlrpc_rule_set tmp_rset;
4413 sptlrpc_rule_set_init(&tmp_rset);
4414 rc = sptlrpc_conf_target_get_rules(obd, &tmp_rset, initial);
4416 CERROR("mdt %s: failed get sptlrpc rules: %d\n",
4421 sptlrpc_target_update_exp_flavor(obd, &tmp_rset);
4423 write_lock(&m->mdt_sptlrpc_lock);
4424 sptlrpc_rule_set_free(&m->mdt_sptlrpc_rset);
4425 m->mdt_sptlrpc_rset = tmp_rset;
4426 write_unlock(&m->mdt_sptlrpc_lock);
4431 static void fsoptions_to_mdt_flags(struct mdt_device *m, char *options)
4435 m->mdt_opts.mo_mds_capa = 1;
4436 m->mdt_opts.mo_oss_capa = 1;
4437 #ifdef CONFIG_FS_POSIX_ACL
4438 /* ACLs should be enabled by default (b=13829) */
4439 m->mdt_opts.mo_acl = 1;
4440 LCONSOLE_INFO("Enabling ACL\n");
4442 m->mdt_opts.mo_acl = 0;
4443 LCONSOLE_INFO("Disabling ACL\n");
4452 while (*p && *p != ',')
4456 if ((len == sizeof("user_xattr") - 1) &&
4457 (memcmp(options, "user_xattr", len) == 0)) {
4458 m->mdt_opts.mo_user_xattr = 1;
4459 LCONSOLE_INFO("Enabling user_xattr\n");
4460 } else if ((len == sizeof("nouser_xattr") - 1) &&
4461 (memcmp(options, "nouser_xattr", len) == 0)) {
4462 m->mdt_opts.mo_user_xattr = 0;
4463 LCONSOLE_INFO("Disabling user_xattr\n");
4464 } else if ((len == sizeof("noacl") - 1) &&
4465 (memcmp(options, "noacl", len) == 0)) {
4466 m->mdt_opts.mo_acl = 0;
4467 LCONSOLE_INFO("Disabling ACL\n");
4474 int mdt_postrecov(const struct lu_env *, struct mdt_device *);
4476 static int mdt_init0(const struct lu_env *env, struct mdt_device *m,
4477 struct lu_device_type *ldt, struct lustre_cfg *cfg)
4479 struct lprocfs_static_vars lvars;
4480 struct mdt_thread_info *info;
4481 struct obd_device *obd;
4482 const char *dev = lustre_cfg_string(cfg, 0);
4483 const char *num = lustre_cfg_string(cfg, 2);
4484 struct lustre_mount_info *lmi = NULL;
4485 struct lustre_sb_info *lsi;
4486 struct lustre_disk_data *ldd;
4488 struct md_site *mite;
4489 const char *identity_upcall = "NONE";
4490 #ifdef HAVE_QUOTA_SUPPORT
4491 struct md_device *next;
4497 md_device_init(&m->mdt_md_dev, ldt);
4499 * Environment (env) might be missing mdt_thread_key values at that
4500 * point, if device is allocated when mdt_thread_key is in QUIESCENT
4503 * Usually device allocation path doesn't use module key values, but
4504 * mdt has to do a lot of work here, so allocate key value.
4506 rc = lu_env_refill((struct lu_env *)env);
4510 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
4511 LASSERT(info != NULL);
4513 obd = class_name2obd(dev);
4514 LASSERT(obd != NULL);
4516 spin_lock_init(&m->mdt_transno_lock);
4518 m->mdt_max_mdsize = MAX_MD_SIZE;
4519 m->mdt_max_cookiesize = sizeof(struct llog_cookie);
4521 m->mdt_opts.mo_user_xattr = 0;
4522 m->mdt_opts.mo_acl = 0;
4523 m->mdt_opts.mo_cos = MDT_COS_DEFAULT;
4524 lmi = server_get_mount_2(dev);
4526 CERROR("Cannot get mount info for %s!\n", dev);
4529 lsi = s2lsi(lmi->lmi_sb);
4530 fsoptions_to_mdt_flags(m, lsi->lsi_lmd->lmd_opts);
4531 /* CMD is supported only in IAM mode */
4534 node_id = simple_strtol(num, NULL, 10);
4535 if (!(ldd->ldd_flags & LDD_F_IAM_DIR) && node_id) {
4536 CERROR("CMD Operation not allowed in IOP mode\n");
4537 GOTO(err_lmi, rc = -EINVAL);
4541 rwlock_init(&m->mdt_sptlrpc_lock);
4542 sptlrpc_rule_set_init(&m->mdt_sptlrpc_rset);
4544 spin_lock_init(&m->mdt_ioepoch_lock);
4545 m->mdt_opts.mo_compat_resname = 0;
4546 m->mdt_capa_timeout = CAPA_TIMEOUT;
4547 m->mdt_capa_alg = CAPA_HMAC_ALG_SHA1;
4548 m->mdt_ck_timeout = CAPA_KEY_TIMEOUT;
4549 m->mdt_squash_uid = 0;
4550 m->mdt_squash_gid = 0;
4551 CFS_INIT_LIST_HEAD(&m->mdt_nosquash_nids);
4552 m->mdt_nosquash_str = NULL;
4553 m->mdt_nosquash_strlen = 0;
4554 init_rwsem(&m->mdt_squash_sem);
4556 spin_lock_init(&m->mdt_client_bitmap_lock);
4558 OBD_ALLOC_PTR(mite);
4560 GOTO(err_lmi, rc = -ENOMEM);
4564 m->mdt_md_dev.md_lu_dev.ld_ops = &mdt_lu_ops;
4565 m->mdt_md_dev.md_lu_dev.ld_obd = obd;
4566 /* set this lu_device to obd, because error handling need it */
4567 obd->obd_lu_dev = &m->mdt_md_dev.md_lu_dev;
4569 rc = lu_site_init(s, &m->mdt_md_dev.md_lu_dev);
4571 CERROR("Can't init lu_site, rc %d\n", rc);
4572 GOTO(err_free_site, rc);
4575 lprocfs_mdt_init_vars(&lvars);
4576 rc = lprocfs_obd_setup(obd, lvars.obd_vars);
4578 CERROR("Can't init lprocfs, rc %d\n", rc);
4579 GOTO(err_fini_site, rc);
4581 ptlrpc_lprocfs_register_obd(obd);
4583 rc = mdt_procfs_init(m, dev);
4585 CERROR("Can't init MDT lprocfs, rc %d\n", rc);
4586 GOTO(err_fini_proc, rc);
4589 obd->obd_proc_exports_entry = proc_mkdir("exports",
4590 obd->obd_proc_entry);
4591 if (obd->obd_proc_exports_entry)
4592 lprocfs_add_simple(obd->obd_proc_exports_entry,
4593 "clear", lprocfs_nid_stats_clear_read,
4594 lprocfs_nid_stats_clear_write, obd, NULL);
4596 /* set server index */
4597 lu_site2md(s)->ms_node_id = node_id;
4599 /* failover is the default
4600 * FIXME: we do not failout mds0/mgs, which may cause some problems.
4601 * assumed whose ms_node_id == 0 XXX
4603 obd->obd_replayable = 1;
4604 /* No connection accepted until configurations will finish */
4605 obd->obd_no_conn = 1;
4607 if (cfg->lcfg_bufcount > 4 && LUSTRE_CFG_BUFLEN(cfg, 4) > 0) {
4608 char *str = lustre_cfg_string(cfg, 4);
4609 if (strchr(str, 'n')) {
4610 CWARN("%s: recovery disabled\n", obd->obd_name);
4611 obd->obd_replayable = 0;
4615 /* init the stack */
4616 rc = mdt_stack_init((struct lu_env *)env, m, cfg, lmi);
4618 CERROR("Can't init device stack, rc %d\n", rc);
4619 GOTO(err_fini_proc, rc);
4622 rc = lut_init(env, &m->mdt_lut, obd, m->mdt_bottom);
4624 GOTO(err_fini_stack, rc);
4626 rc = mdt_fld_init(env, obd->obd_name, m);
4630 rc = mdt_seq_init(env, obd->obd_name, m);
4632 GOTO(err_fini_fld, rc);
4634 snprintf(info->mti_u.ns_name, sizeof info->mti_u.ns_name,
4635 LUSTRE_MDT_NAME"-%p", m);
4636 m->mdt_namespace = ldlm_namespace_new(obd, info->mti_u.ns_name,
4637 LDLM_NAMESPACE_SERVER,
4638 LDLM_NAMESPACE_GREEDY);
4639 if (m->mdt_namespace == NULL)
4640 GOTO(err_fini_seq, rc = -ENOMEM);
4642 ldlm_register_intent(m->mdt_namespace, mdt_intent_policy);
4643 /* set obd_namespace for compatibility with old code */
4644 obd->obd_namespace = m->mdt_namespace;
4646 /* XXX: to support suppgid for ACL, we enable identity_upcall
4647 * by default, otherwise, maybe got unexpected -EACCESS. */
4648 if (m->mdt_opts.mo_acl)
4649 identity_upcall = MDT_IDENTITY_UPCALL_PATH;
4651 m->mdt_identity_cache = upcall_cache_init(obd->obd_name, identity_upcall,
4652 &mdt_identity_upcall_cache_ops);
4653 if (IS_ERR(m->mdt_identity_cache)) {
4654 rc = PTR_ERR(m->mdt_identity_cache);
4655 m->mdt_identity_cache = NULL;
4656 GOTO(err_free_ns, rc);
4659 cfs_timer_init(&m->mdt_ck_timer, mdt_ck_timer_callback, m);
4661 rc = mdt_ck_thread_start(m);
4663 GOTO(err_free_ns, rc);
4665 rc = mdt_fs_setup(env, m, obd, lsi);
4669 rc = mdt_obd_llog_setup(obd, lsi);
4671 GOTO(err_fs_cleanup, rc);
4673 rc = mdt_llog_ctxt_clone(env, m, LLOG_CHANGELOG_ORIG_CTXT);
4675 GOTO(err_llog_cleanup, rc);
4677 mdt_adapt_sptlrpc_conf(obd, 1);
4679 #ifdef HAVE_QUOTA_SUPPORT
4680 next = m->mdt_child;
4681 rc = next->md_ops->mdo_quota.mqo_setup(env, next, lmi->lmi_mnt);
4683 GOTO(err_llog_cleanup, rc);
4686 server_put_mount_2(dev, lmi->lmi_mnt);
4689 target_recovery_init(&m->mdt_lut, mdt_recovery_handle);
4691 rc = mdt_start_ptlrpc_service(m);
4693 GOTO(err_recovery, rc);
4695 ping_evictor_start();
4697 rc = lu_site_init_finish(s);
4699 GOTO(err_stop_service, rc);
4701 if (obd->obd_recovering == 0)
4702 mdt_postrecov(env, m);
4704 mdt_init_capa_ctxt(env, m);
4706 /* Reduce the initial timeout on an MDS because it doesn't need such
4707 * a long timeout as an OST does. Adaptive timeouts will adjust this
4708 * value appropriately. */
4709 if (ldlm_timeout == LDLM_TIMEOUT_DEFAULT)
4710 ldlm_timeout = MDS_LDLM_TIMEOUT_DEFAULT;
4715 ping_evictor_stop();
4716 mdt_stop_ptlrpc_service(m);
4718 target_recovery_fini(obd);
4719 #ifdef HAVE_QUOTA_SUPPORT
4720 next->md_ops->mdo_quota.mqo_cleanup(env, next);
4723 mdt_llog_ctxt_unclone(env, m, LLOG_CHANGELOG_ORIG_CTXT);
4724 mdt_obd_llog_cleanup(obd);
4726 mdt_fs_cleanup(env, m);
4728 cfs_timer_disarm(&m->mdt_ck_timer);
4729 mdt_ck_thread_stop(m);
4731 upcall_cache_cleanup(m->mdt_identity_cache);
4732 m->mdt_identity_cache = NULL;
4733 ldlm_namespace_free(m->mdt_namespace, NULL, 0);
4734 obd->obd_namespace = m->mdt_namespace = NULL;
4736 mdt_seq_fini(env, m);
4738 mdt_fld_fini(env, m);
4740 lut_fini(env, &m->mdt_lut);
4742 mdt_stack_fini(env, m, md2lu_dev(m->mdt_child));
4745 if (obd->obd_proc_exports_entry) {
4746 lprocfs_remove_proc_entry("clear", obd->obd_proc_exports_entry);
4747 obd->obd_proc_exports_entry = NULL;
4749 ptlrpc_lprocfs_unregister_obd(obd);
4750 lprocfs_obd_cleanup(obd);
4757 server_put_mount_2(dev, lmi->lmi_mnt);
4761 /* used by MGS to process specific configurations */
4762 static int mdt_process_config(const struct lu_env *env,
4763 struct lu_device *d, struct lustre_cfg *cfg)
4765 struct mdt_device *m = mdt_dev(d);
4766 struct md_device *md_next = m->mdt_child;
4767 struct lu_device *next = md2lu_dev(md_next);
4771 switch (cfg->lcfg_command) {
4773 struct lprocfs_static_vars lvars;
4774 struct obd_device *obd = d->ld_obd;
4777 * For interoperability between 1.8 and 2.0,
4778 * skip old "mdt.group_upcall" param.
4781 char *param = lustre_cfg_string(cfg, 1);
4782 if (param && !strncmp("mdt.group_upcall", param, 16)) {
4783 CWARN("For 1.8 interoperability, skip this"
4784 " mdt.group_upcall. It is obsolete\n");
4789 lprocfs_mdt_init_vars(&lvars);
4790 rc = class_process_proc_param(PARAM_MDT, lvars.obd_vars,
4792 if (rc > 0 || rc == -ENOSYS)
4793 /* we don't understand; pass it on */
4794 rc = next->ld_ops->ldo_process_config(env, next, cfg);
4799 * Add mdc hook to get first MDT uuid and connect it to
4800 * ls->controller to use for seq manager.
4802 rc = next->ld_ops->ldo_process_config(env, next, cfg);
4804 CERROR("Can't add mdc, rc %d\n", rc);
4806 rc = mdt_seq_init_cli(env, mdt_dev(d), cfg);
4809 /* others are passed further */
4810 rc = next->ld_ops->ldo_process_config(env, next, cfg);
4816 static struct lu_object *mdt_object_alloc(const struct lu_env *env,
4817 const struct lu_object_header *hdr,
4818 struct lu_device *d)
4820 struct mdt_object *mo;
4826 struct lu_object *o;
4827 struct lu_object_header *h;
4829 o = &mo->mot_obj.mo_lu;
4830 h = &mo->mot_header;
4831 lu_object_header_init(h);
4832 lu_object_init(o, h, d);
4833 lu_object_add_top(h, o);
4834 o->lo_ops = &mdt_obj_ops;
4840 static int mdt_object_init(const struct lu_env *env, struct lu_object *o,
4841 const struct lu_object_conf *_)
4843 struct mdt_device *d = mdt_dev(o->lo_dev);
4844 struct lu_device *under;
4845 struct lu_object *below;
4849 CDEBUG(D_INFO, "object init, fid = "DFID"\n",
4850 PFID(lu_object_fid(o)));
4852 under = &d->mdt_child->md_lu_dev;
4853 below = under->ld_ops->ldo_object_alloc(env, o->lo_header, under);
4854 if (below != NULL) {
4855 lu_object_add(o, below);
4862 static void mdt_object_free(const struct lu_env *env, struct lu_object *o)
4864 struct mdt_object *mo = mdt_obj(o);
4865 struct lu_object_header *h;
4869 CDEBUG(D_INFO, "object free, fid = "DFID"\n",
4870 PFID(lu_object_fid(o)));
4873 lu_object_header_fini(h);
4878 static int mdt_object_print(const struct lu_env *env, void *cookie,
4879 lu_printer_t p, const struct lu_object *o)
4881 struct mdt_object *mdto = mdt_obj((struct lu_object *)o);
4882 return (*p)(env, cookie, LUSTRE_MDT_NAME"-object@%p(ioepoch=%llu "
4883 "flags=%llx, epochcount=%d, writecount=%d)",
4884 mdto, mdto->mot_ioepoch, mdto->mot_flags,
4885 mdto->mot_epochcount, mdto->mot_writecount);
4888 static const struct lu_device_operations mdt_lu_ops = {
4889 .ldo_object_alloc = mdt_object_alloc,
4890 .ldo_process_config = mdt_process_config,
4893 static const struct lu_object_operations mdt_obj_ops = {
4894 .loo_object_init = mdt_object_init,
4895 .loo_object_free = mdt_object_free,
4896 .loo_object_print = mdt_object_print
4899 static int mdt_obd_set_info_async(struct obd_export *exp,
4900 __u32 keylen, void *key,
4901 __u32 vallen, void *val,
4902 struct ptlrpc_request_set *set)
4904 struct obd_device *obd = exp->exp_obd;
4910 if (KEY_IS(KEY_SPTLRPC_CONF)) {
4911 rc = mdt_adapt_sptlrpc_conf(obd, 0);
4918 /* mds_connect_internal */
4919 static int mdt_connect_internal(struct obd_export *exp,
4920 struct mdt_device *mdt,
4921 struct obd_connect_data *data)
4924 data->ocd_connect_flags &= MDT_CONNECT_SUPPORTED;
4925 data->ocd_ibits_known &= MDS_INODELOCK_FULL;
4927 /* If no known bits (which should not happen, probably,
4928 as everybody should support LOOKUP and UPDATE bits at least)
4929 revert to compat mode with plain locks. */
4930 if (!data->ocd_ibits_known &&
4931 data->ocd_connect_flags & OBD_CONNECT_IBITS)
4932 data->ocd_connect_flags &= ~OBD_CONNECT_IBITS;
4934 if (!mdt->mdt_opts.mo_acl)
4935 data->ocd_connect_flags &= ~OBD_CONNECT_ACL;
4937 if (!mdt->mdt_opts.mo_user_xattr)
4938 data->ocd_connect_flags &= ~OBD_CONNECT_XATTR;
4940 spin_lock(&exp->exp_lock);
4941 exp->exp_connect_flags = data->ocd_connect_flags;
4942 spin_unlock(&exp->exp_lock);
4943 data->ocd_version = LUSTRE_VERSION_CODE;
4944 exp->exp_mdt_data.med_ibits_known = data->ocd_ibits_known;
4948 if (mdt->mdt_opts.mo_acl &&
4949 ((exp->exp_connect_flags & OBD_CONNECT_ACL) == 0)) {
4950 CWARN("%s: MDS requires ACL support but client does not\n",
4951 mdt->mdt_md_dev.md_lu_dev.ld_obd->obd_name);
4956 if ((exp->exp_connect_flags & OBD_CONNECT_FID) == 0) {
4957 CWARN("%s: MDS requires FID support, but client not\n",
4958 mdt->mdt_md_dev.md_lu_dev.ld_obd->obd_name);
4965 static int mdt_connect_check_sptlrpc(struct mdt_device *mdt,
4966 struct obd_export *exp,
4967 struct ptlrpc_request *req)
4969 struct sptlrpc_flavor flvr;
4972 if (exp->exp_flvr.sf_rpc == SPTLRPC_FLVR_INVALID) {
4973 read_lock(&mdt->mdt_sptlrpc_lock);
4974 sptlrpc_target_choose_flavor(&mdt->mdt_sptlrpc_rset,
4978 read_unlock(&mdt->mdt_sptlrpc_lock);
4980 spin_lock(&exp->exp_lock);
4982 exp->exp_sp_peer = req->rq_sp_from;
4983 exp->exp_flvr = flvr;
4985 if (exp->exp_flvr.sf_rpc != SPTLRPC_FLVR_ANY &&
4986 exp->exp_flvr.sf_rpc != req->rq_flvr.sf_rpc) {
4987 CERROR("unauthorized rpc flavor %x from %s, "
4988 "expect %x\n", req->rq_flvr.sf_rpc,
4989 libcfs_nid2str(req->rq_peer.nid),
4990 exp->exp_flvr.sf_rpc);
4994 spin_unlock(&exp->exp_lock);
4996 if (exp->exp_sp_peer != req->rq_sp_from) {
4997 CERROR("RPC source %s doesn't match %s\n",
4998 sptlrpc_part2name(req->rq_sp_from),
4999 sptlrpc_part2name(exp->exp_sp_peer));
5002 rc = sptlrpc_target_export_check(exp, req);
5009 /* mds_connect copy */
5010 static int mdt_obd_connect(const struct lu_env *env,
5011 struct obd_export **exp, struct obd_device *obd,
5012 struct obd_uuid *cluuid,
5013 struct obd_connect_data *data,
5016 struct mdt_thread_info *info;
5017 struct lsd_client_data *lcd;
5018 struct obd_export *lexp;
5019 struct lustre_handle conn = { 0 };
5020 struct mdt_device *mdt;
5021 struct ptlrpc_request *req;
5025 LASSERT(env != NULL);
5026 if (!exp || !obd || !cluuid)
5029 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
5030 req = info->mti_pill->rc_req;
5031 mdt = mdt_dev(obd->obd_lu_dev);
5033 rc = class_connect(&conn, obd, cluuid);
5037 lexp = class_conn2export(&conn);
5038 LASSERT(lexp != NULL);
5040 rc = mdt_connect_check_sptlrpc(mdt, lexp, req);
5044 rc = mdt_connect_internal(lexp, mdt, data);
5048 struct mdt_thread_info *mti;
5049 mti = lu_context_key_get(&env->le_ctx,
5051 LASSERT(mti != NULL);
5052 mti->mti_exp = lexp;
5053 memcpy(lcd->lcd_uuid, cluuid, sizeof lcd->lcd_uuid);
5054 lexp->exp_mdt_data.med_lcd = lcd;
5055 rc = mdt_client_new(env, mdt);
5058 lexp->exp_mdt_data.med_lcd = NULL;
5060 mdt_export_stats_init(obd, lexp, localdata);
5068 class_disconnect(lexp);
5075 static int mdt_obd_reconnect(const struct lu_env *env,
5076 struct obd_export *exp, struct obd_device *obd,
5077 struct obd_uuid *cluuid,
5078 struct obd_connect_data *data,
5081 struct mdt_thread_info *info;
5082 struct mdt_device *mdt;
5083 struct ptlrpc_request *req;
5087 if (exp == NULL || obd == NULL || cluuid == NULL)
5090 info = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
5091 req = info->mti_pill->rc_req;
5092 mdt = mdt_dev(obd->obd_lu_dev);
5094 rc = mdt_connect_check_sptlrpc(mdt, exp, req);
5098 rc = mdt_connect_internal(exp, mdt_dev(obd->obd_lu_dev), data);
5100 mdt_export_stats_init(obd, exp, localdata);
5105 static int mdt_obd_disconnect(struct obd_export *exp)
5107 struct mdt_device *mdt = mdt_dev(exp->exp_obd->obd_lu_dev);
5112 class_export_get(exp);
5114 /* Disconnect early so that clients can't keep using export */
5115 rc = class_disconnect(exp);
5116 if (mdt->mdt_namespace != NULL || exp->exp_obd->obd_namespace != NULL)
5117 ldlm_cancel_locks_for_export(exp);
5119 /* release nid stat refererence */
5120 lprocfs_exp_cleanup(exp);
5122 /* complete all outstanding replies */
5123 spin_lock(&exp->exp_lock);
5124 while (!list_empty(&exp->exp_outstanding_replies)) {
5125 struct ptlrpc_reply_state *rs =
5126 list_entry(exp->exp_outstanding_replies.next,
5127 struct ptlrpc_reply_state, rs_exp_list);
5128 struct ptlrpc_service *svc = rs->rs_service;
5130 spin_lock(&svc->srv_lock);
5131 list_del_init(&rs->rs_exp_list);
5132 spin_lock(&rs->rs_lock);
5133 ptlrpc_schedule_difficult_reply(rs);
5134 spin_unlock(&rs->rs_lock);
5135 spin_unlock(&svc->srv_lock);
5137 spin_unlock(&exp->exp_lock);
5139 class_export_put(exp);
5143 /* FIXME: Can we avoid using these two interfaces? */
5144 static int mdt_init_export(struct obd_export *exp)
5146 struct mdt_export_data *med = &exp->exp_mdt_data;
5150 CFS_INIT_LIST_HEAD(&med->med_open_head);
5151 spin_lock_init(&med->med_open_lock);
5152 sema_init(&med->med_idmap_sem, 1);
5153 med->med_idmap = NULL;
5154 spin_lock(&exp->exp_lock);
5155 exp->exp_connecting = 1;
5156 spin_unlock(&exp->exp_lock);
5157 rc = ldlm_init_export(exp);
5159 CERROR("Error %d while initializing export\n", rc);
5163 static int mdt_destroy_export(struct obd_export *export)
5165 struct mdt_export_data *med;
5166 struct obd_device *obd = export->exp_obd;
5167 struct mdt_device *mdt;
5168 struct mdt_thread_info *info;
5173 CFS_LIST_HEAD(closing_list);
5174 struct mdt_file_data *mfd, *n;
5178 med = &export->exp_mdt_data;
5179 if (exp_connect_rmtclient(export))
5180 mdt_cleanup_idmap(med);
5182 target_destroy_export(export);
5183 ldlm_destroy_export(export);
5185 if (obd_uuid_equals(&export->exp_client_uuid, &obd->obd_uuid))
5188 mdt = mdt_dev(obd->obd_lu_dev);
5189 LASSERT(mdt != NULL);
5191 rc = lu_env_init(&env, LCT_MD_THREAD);
5195 info = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
5196 LASSERT(info != NULL);
5197 memset(info, 0, sizeof *info);
5198 info->mti_env = &env;
5199 info->mti_mdt = mdt;
5200 info->mti_exp = export;
5202 ma = &info->mti_attr;
5203 lmm_size = ma->ma_lmm_size = mdt->mdt_max_mdsize;
5204 cookie_size = ma->ma_cookie_size = mdt->mdt_max_cookiesize;
5205 OBD_ALLOC(ma->ma_lmm, lmm_size);
5206 OBD_ALLOC(ma->ma_cookie, cookie_size);
5208 if (ma->ma_lmm == NULL || ma->ma_cookie == NULL)
5209 GOTO(out, rc = -ENOMEM);
5210 ma->ma_need = MA_LOV | MA_COOKIE;
5212 /* Close any open files (which may also cause orphan unlinking). */
5213 spin_lock(&med->med_open_lock);
5214 while (!list_empty(&med->med_open_head)) {
5215 struct list_head *tmp = med->med_open_head.next;
5216 mfd = list_entry(tmp, struct mdt_file_data, mfd_list);
5218 /* Remove mfd handle so it can't be found again.
5219 * We are consuming the mfd_list reference here. */
5220 class_handle_unhash(&mfd->mfd_handle);
5221 list_move_tail(&mfd->mfd_list, &closing_list);
5223 spin_unlock(&med->med_open_lock);
5225 list_for_each_entry_safe(mfd, n, &closing_list, mfd_list) {
5226 list_del_init(&mfd->mfd_list);
5227 mdt_mfd_close(info, mfd);
5228 /* TODO: if we close the unlinked file,
5229 * we need to remove its objects from OST */
5230 memset(&ma->ma_attr, 0, sizeof(ma->ma_attr));
5231 spin_lock(&med->med_open_lock);
5232 ma->ma_lmm_size = lmm_size;
5233 ma->ma_cookie_size = cookie_size;
5234 ma->ma_need = MA_LOV | MA_COOKIE;
5236 spin_unlock(&med->med_open_lock);
5239 info->mti_mdt = NULL;
5240 mdt_client_del(&env, mdt);
5245 OBD_FREE(ma->ma_lmm, lmm_size);
5249 OBD_FREE(ma->ma_cookie, cookie_size);
5250 ma->ma_cookie = NULL;
5257 static void mdt_allow_cli(struct mdt_device *m, unsigned int flag)
5259 if (flag & CONFIG_LOG)
5260 m->mdt_fl_cfglog = 1;
5262 /* also notify active event */
5263 if (flag & CONFIG_SYNC)
5264 m->mdt_fl_synced = 1;
5266 if (m->mdt_fl_cfglog && m->mdt_fl_synced)
5267 /* Open for clients */
5268 m->mdt_md_dev.md_lu_dev.ld_obd->obd_no_conn = 0;
5271 static int mdt_upcall(const struct lu_env *env, struct md_device *md,
5272 enum md_upcall_event ev)
5274 struct mdt_device *m = mdt_dev(&md->md_lu_dev);
5275 struct md_device *next = m->mdt_child;
5276 struct mdt_thread_info *mti;
5282 rc = next->md_ops->mdo_maxsize_get(env, next,
5284 &m->mdt_max_cookiesize);
5285 CDEBUG(D_INFO, "get max mdsize %d max cookiesize %d\n",
5286 m->mdt_max_mdsize, m->mdt_max_cookiesize);
5287 mdt_allow_cli(m, CONFIG_SYNC);
5290 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
5291 mti->mti_no_need_trans = 1;
5292 CDEBUG(D_INFO, "disable mdt trans for this thread\n");
5295 /* Check that MDT is not yet configured */
5296 LASSERT(!m->mdt_fl_cfglog);
5298 #ifdef HAVE_QUOTA_SUPPORT
5300 if (md->md_lu_dev.ld_obd->obd_recovering == 0 &&
5301 likely(md->md_lu_dev.ld_obd->obd_stopping == 0))
5302 next->md_ops->mdo_quota.mqo_recovery(env, next);
5306 CERROR("invalid event\n");
5313 static int mdt_obd_notify(struct obd_device *host,
5314 struct obd_device *watched,
5315 enum obd_notify_event ev, void *data)
5317 struct mdt_device *mdt = mdt_dev(host->obd_lu_dev);
5318 #ifdef HAVE_QUOTA_SUPPORT
5319 struct md_device *next = mdt->mdt_child;
5324 case OBD_NOTIFY_CONFIG:
5325 mdt_allow_cli(mdt, (unsigned long)data);
5327 #ifdef HAVE_QUOTA_SUPPORT
5328 /* quota_type has been processed, we can now handle
5329 * incoming quota requests */
5330 next->md_ops->mdo_quota.mqo_notify(NULL, next);
5334 CDEBUG(D_INFO, "Unhandled notification %#x\n", ev);
5339 /* This routine is quite similar to mdc_ioc_fid2path() */
5340 static int mdt_ioc_fid2path(struct lu_env *env, struct mdt_device *mdt,
5341 struct obd_ioctl_data *data)
5343 struct getinfo_fid2path *fp = NULL;
5350 pathlen = data->ioc_plen1;
5351 if (pathlen > PATH_MAX)
5352 GOTO(out, rc = -EINVAL);
5354 GOTO(out, rc = -EINVAL);
5356 alloclen = sizeof(*fp) + pathlen;
5357 OBD_ALLOC(fp, alloclen);
5359 GOTO(out, rc = -ENOMEM);
5361 memcpy(&fp->gf_fid, data->ioc_inlbuf1, sizeof(struct lu_fid));
5362 memcpy(&fp->gf_recno, data->ioc_inlbuf2, sizeof(fp->gf_recno));
5363 memcpy(&fp->gf_linkno, data->ioc_inlbuf3,
5364 sizeof(fp->gf_linkno));
5365 fp->gf_pathlen = pathlen;
5367 if (!fid_is_sane(&fp->gf_fid))
5368 GOTO(out, rc = -EINVAL);
5370 rc = mdt_fid2path(env, mdt, fp);
5372 if (!rc && copy_to_user(data->ioc_pbuf1, fp->gf_path, pathlen))
5373 GOTO(out, rc = -EFAULT);
5374 memcpy(data->ioc_inlbuf2, &fp->gf_recno, sizeof(fp->gf_recno));
5375 memcpy(data->ioc_inlbuf3, &fp->gf_linkno, sizeof(fp->gf_linkno));
5378 OBD_FREE(fp, alloclen);
5382 static int mdt_rpc_fid2path(struct mdt_thread_info *info, void *key,
5383 int keylen, void *val, int vallen)
5386 struct mdt_device *mdt = mdt_dev(info->mti_exp->exp_obd->obd_lu_dev);
5387 struct ptlrpc_request *req = mdt_info_req(info);
5388 struct getinfo_fid2path *fpout, *fpin;
5391 fpin = key + size_round(sizeof(KEY_FID2PATH));
5394 if (lustre_msg_swabbed(req->rq_reqmsg))
5395 lustre_swab_fid2path(fpin);
5397 memcpy(fpout, fpin, sizeof(*fpin));
5398 if (fpout->gf_pathlen != vallen - sizeof(*fpin))
5401 rc = lu_env_init(&env, LCT_MD_THREAD);
5404 rc = mdt_fid2path(&env, mdt, fpout);
5410 static int mdt_fid2path(struct lu_env *env, struct mdt_device *mdt,
5411 struct getinfo_fid2path *fp)
5413 struct lu_context ioctl_session;
5414 struct mdt_object *obj;
5418 CDEBUG(D_IOCTL, "path get "DFID" from "LPU64" #%d\n",
5419 PFID(&fp->gf_fid), fp->gf_recno, fp->gf_linkno);
5421 if (!fid_is_sane(&fp->gf_fid))
5424 rc = lu_context_init(&ioctl_session, LCT_SESSION);
5427 ioctl_session.lc_thread = (struct ptlrpc_thread *)cfs_current();
5428 lu_context_enter(&ioctl_session);
5429 env->le_ses = &ioctl_session;
5431 obj = mdt_object_find(env, mdt, &fp->gf_fid);
5432 if (obj == NULL || IS_ERR(obj)) {
5433 CDEBUG(D_IOCTL, "no object "DFID": %ld\n",PFID(&fp->gf_fid),
5435 GOTO(out, rc = -EINVAL);
5438 rc = lu_object_exists(&obj->mot_obj.mo_lu);
5444 mdt_object_put(env, obj);
5445 CDEBUG(D_IOCTL, "nonlocal object "DFID": %d\n",
5446 PFID(&fp->gf_fid), rc);
5450 rc = mo_path(env, md_object_next(&obj->mot_obj), fp->gf_path,
5451 fp->gf_pathlen, &fp->gf_recno, &fp->gf_linkno);
5452 mdt_object_put(env, obj);
5457 lu_context_exit(&ioctl_session);
5458 lu_context_fini(&ioctl_session);
5462 static int mdt_get_info(struct mdt_thread_info *info)
5472 struct ptlrpc_request *req = mdt_info_req(info);
5476 key = req_capsule_client_get(info->mti_pill, &RMF_GETINFO_KEY);
5478 CDEBUG(D_IOCTL, "No GETINFO key");
5481 keysize = req_capsule_get_size(info->mti_pill, &RMF_GETINFO_KEY,
5483 keylen = strnlen(key, keysize);
5485 valin = req_capsule_client_get(info->mti_pill, &RMF_GETINFO_VALLEN);
5486 if (valin == NULL) {
5487 CDEBUG(D_IOCTL, "Unable to get RMF_GETINFO_VALLEN buffer");
5490 valsize = req_capsule_get_size(info->mti_pill, &RMF_GETINFO_VALLEN,
5492 if (valsize != sizeof(vallen)) {
5493 CDEBUG(D_IOCTL, "RMF_GETINFO_VALLEN has invalid size");
5498 req_capsule_set_size(info->mti_pill, &RMF_GETINFO_VAL, RCL_SERVER,
5500 rc = req_capsule_server_pack(info->mti_pill);
5501 valout = req_capsule_server_get(info->mti_pill, &RMF_GETINFO_VAL);
5502 if (valout == NULL) {
5503 CDEBUG(D_IOCTL, "Unable to get get-info RPC out buffer");
5507 if (KEY_IS(KEY_FID2PATH))
5508 rc = mdt_rpc_fid2path(info, key, keysize, valout, vallen);
5512 lustre_msg_set_status(req->rq_repmsg, rc);
5517 /* Pass the ioc down */
5518 static int mdt_ioc_child(struct lu_env *env, struct mdt_device *mdt,
5519 unsigned int cmd, int len, void *data)
5521 struct lu_context ioctl_session;
5522 struct md_device *next = mdt->mdt_child;
5526 rc = lu_context_init(&ioctl_session, LCT_SESSION);
5529 ioctl_session.lc_thread = (struct ptlrpc_thread *)cfs_current();
5530 lu_context_enter(&ioctl_session);
5531 env->le_ses = &ioctl_session;
5533 LASSERT(next->md_ops->mdo_iocontrol);
5534 rc = next->md_ops->mdo_iocontrol(env, next, cmd, len, data);
5536 lu_context_exit(&ioctl_session);
5537 lu_context_fini(&ioctl_session);
5541 /* ioctls on obd dev */
5542 static int mdt_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
5543 void *karg, void *uarg)
5546 struct obd_device *obd = exp->exp_obd;
5547 struct mdt_device *mdt = mdt_dev(obd->obd_lu_dev);
5548 struct dt_device *dt = mdt->mdt_bottom;
5552 CDEBUG(D_IOCTL, "handling ioctl cmd %#x\n", cmd);
5553 rc = lu_env_init(&env, LCT_MD_THREAD);
5559 rc = mdt_device_sync(&env, mdt);
5561 case OBD_IOC_SET_READONLY:
5562 dt->dd_ops->dt_ro(&env, dt);
5564 case OBD_IOC_ABORT_RECOVERY:
5565 CERROR("Aborting recovery for device %s\n", obd->obd_name);
5566 target_stop_recovery_thread(obd);
5569 case OBD_IOC_FID2PATH:
5570 rc = mdt_ioc_fid2path(&env, mdt, karg);
5572 case OBD_IOC_CHANGELOG_REG:
5573 case OBD_IOC_CHANGELOG_DEREG:
5574 case OBD_IOC_CHANGELOG_CLEAR:
5575 rc = mdt_ioc_child(&env, mdt, cmd, len, karg);
5578 CERROR("Not supported cmd = %d for device %s\n",
5579 cmd, obd->obd_name);
5587 int mdt_postrecov(const struct lu_env *env, struct mdt_device *mdt)
5589 struct lu_device *ld = md2lu_dev(mdt->mdt_child);
5590 struct obd_device *obd = mdt2obd_dev(mdt);
5591 #ifdef HAVE_QUOTA_SUPPORT
5592 struct md_device *next = mdt->mdt_child;
5596 /* if some clients didn't participate in recovery then we can possibly
5597 * lost sequence. Now we should increase sequence for safe value */
5598 lost = obd->obd_max_recoverable_clients - obd->obd_connected_clients;
5599 mdt_seq_adjust(env, mdt, lost);
5601 rc = ld->ld_ops->ldo_recovery_complete(env, ld);
5602 #ifdef HAVE_QUOTA_SUPPORT
5603 if (likely(obd->obd_stopping == 0))
5604 next->md_ops->mdo_quota.mqo_recovery(env, next);
5609 int mdt_obd_postrecov(struct obd_device *obd)
5614 rc = lu_env_init(&env, LCT_MD_THREAD);
5617 rc = mdt_postrecov(&env, mdt_dev(obd->obd_lu_dev));
5622 static struct obd_ops mdt_obd_device_ops = {
5623 .o_owner = THIS_MODULE,
5624 .o_set_info_async = mdt_obd_set_info_async,
5625 .o_connect = mdt_obd_connect,
5626 .o_reconnect = mdt_obd_reconnect,
5627 .o_disconnect = mdt_obd_disconnect,
5628 .o_init_export = mdt_init_export,
5629 .o_destroy_export = mdt_destroy_export,
5630 .o_iocontrol = mdt_iocontrol,
5631 .o_postrecov = mdt_obd_postrecov,
5632 .o_notify = mdt_obd_notify
5635 static struct lu_device* mdt_device_fini(const struct lu_env *env,
5636 struct lu_device *d)
5638 struct mdt_device *m = mdt_dev(d);
5645 static struct lu_device *mdt_device_free(const struct lu_env *env,
5646 struct lu_device *d)
5648 struct mdt_device *m = mdt_dev(d);
5651 md_device_fini(&m->mdt_md_dev);
5656 static struct lu_device *mdt_device_alloc(const struct lu_env *env,
5657 struct lu_device_type *t,
5658 struct lustre_cfg *cfg)
5660 struct lu_device *l;
5661 struct mdt_device *m;
5667 l = &m->mdt_md_dev.md_lu_dev;
5668 rc = mdt_init0(env, m, t, cfg);
5670 mdt_device_free(env, l);
5674 md_upcall_init(&m->mdt_md_dev, mdt_upcall);
5676 l = ERR_PTR(-ENOMEM);
5680 /* context key constructor/destructor: mdt_key_init, mdt_key_fini */
5681 LU_KEY_INIT_FINI(mdt, struct mdt_thread_info);
5683 /* context key: mdt_thread_key */
5684 LU_CONTEXT_KEY_DEFINE(mdt, LCT_MD_THREAD);
5686 /* context key constructor/destructor: mdt_txn_key_init, mdt_txn_key_fini */
5687 LU_KEY_INIT_FINI(mdt_txn, struct mdt_txn_info);
5689 struct lu_context_key mdt_txn_key = {
5690 .lct_tags = LCT_TX_HANDLE,
5691 .lct_init = mdt_txn_key_init,
5692 .lct_fini = mdt_txn_key_fini
5695 struct md_ucred *mdt_ucred(const struct mdt_thread_info *info)
5697 return md_ucred(info->mti_env);
5701 * Enable/disable COS.
5703 * Set/Clear the COS flag in mdt options.
5705 * \param mdt mdt device
5706 * \param val 0 disables COS, other values enable COS
5708 void mdt_enable_cos(struct mdt_device *mdt, int val)
5713 mdt->mdt_opts.mo_cos = !!val;
5714 rc = lu_env_init(&env, LCT_MD_THREAD);
5715 if (unlikely(rc != 0)) {
5716 CWARN("lu_env initialization failed with rc = %d,"
5717 "cannot sync\n", rc);
5720 mdt_device_sync(&env, mdt);
5727 * Return COS flag status/
5729 * \param mdt mdt device
5731 int mdt_cos_is_enabled(struct mdt_device *mdt)
5733 return mdt->mdt_opts.mo_cos != 0;
5736 /* type constructor/destructor: mdt_type_init, mdt_type_fini */
5737 LU_TYPE_INIT_FINI(mdt, &mdt_thread_key, &mdt_txn_key);
5739 static struct lu_device_type_operations mdt_device_type_ops = {
5740 .ldto_init = mdt_type_init,
5741 .ldto_fini = mdt_type_fini,
5743 .ldto_start = mdt_type_start,
5744 .ldto_stop = mdt_type_stop,
5746 .ldto_device_alloc = mdt_device_alloc,
5747 .ldto_device_free = mdt_device_free,
5748 .ldto_device_fini = mdt_device_fini
5751 static struct lu_device_type mdt_device_type = {
5752 .ldt_tags = LU_DEVICE_MD,
5753 .ldt_name = LUSTRE_MDT_NAME,
5754 .ldt_ops = &mdt_device_type_ops,
5755 .ldt_ctx_tags = LCT_MD_THREAD
5758 static struct lu_local_obj_desc mdt_last_recv = {
5759 .llod_name = LAST_RCVD,
5760 .llod_oid = MDT_LAST_RECV_OID,
5764 static int __init mdt_mod_init(void)
5766 struct lprocfs_static_vars lvars;
5769 llo_local_obj_register(&mdt_last_recv);
5771 mdt_num_threads = MDT_NUM_THREADS;
5772 lprocfs_mdt_init_vars(&lvars);
5773 rc = class_register_type(&mdt_obd_device_ops, NULL,
5774 lvars.module_vars, LUSTRE_MDT_NAME,
5780 static void __exit mdt_mod_exit(void)
5782 llo_local_obj_unregister(&mdt_last_recv);
5783 class_unregister_type(LUSTRE_MDT_NAME);
5787 #define DEF_HNDL(prefix, base, suffix, flags, opc, fn, fmt) \
5788 [prefix ## _ ## opc - prefix ## _ ## base] = { \
5790 .mh_fail_id = OBD_FAIL_ ## prefix ## _ ## opc ## suffix, \
5791 .mh_opc = prefix ## _ ## opc, \
5792 .mh_flags = flags, \
5797 #define DEF_MDT_HNDL(flags, name, fn, fmt) \
5798 DEF_HNDL(MDS, GETATTR, _NET, flags, name, fn, fmt)
5800 #define DEF_SEQ_HNDL(flags, name, fn, fmt) \
5801 DEF_HNDL(SEQ, QUERY, _NET, flags, name, fn, fmt)
5803 #define DEF_FLD_HNDL(flags, name, fn, fmt) \
5804 DEF_HNDL(FLD, QUERY, _NET, flags, name, fn, fmt)
5806 * Request with a format known in advance
5808 #define DEF_MDT_HNDL_F(flags, name, fn) \
5809 DEF_HNDL(MDS, GETATTR, _NET, flags, name, fn, &RQF_MDS_ ## name)
5811 #define DEF_SEQ_HNDL_F(flags, name, fn) \
5812 DEF_HNDL(SEQ, QUERY, _NET, flags, name, fn, &RQF_SEQ_ ## name)
5814 #define DEF_FLD_HNDL_F(flags, name, fn) \
5815 DEF_HNDL(FLD, QUERY, _NET, flags, name, fn, &RQF_FLD_ ## name)
5817 * Request with a format we do not yet know
5819 #define DEF_MDT_HNDL_0(flags, name, fn) \
5820 DEF_HNDL(MDS, GETATTR, _NET, flags, name, fn, NULL)
5822 static struct mdt_handler mdt_mds_ops[] = {
5823 DEF_MDT_HNDL_F(0, CONNECT, mdt_connect),
5824 DEF_MDT_HNDL_F(0, DISCONNECT, mdt_disconnect),
5825 DEF_MDT_HNDL_F(0, SET_INFO, mdt_set_info),
5826 DEF_MDT_HNDL_F(0, GET_INFO, mdt_get_info),
5827 DEF_MDT_HNDL_F(0 |HABEO_REFERO, GETSTATUS, mdt_getstatus),
5828 DEF_MDT_HNDL_F(HABEO_CORPUS, GETATTR, mdt_getattr),
5829 DEF_MDT_HNDL_F(HABEO_CORPUS|HABEO_REFERO, GETATTR_NAME, mdt_getattr_name),
5830 DEF_MDT_HNDL_F(HABEO_CORPUS, GETXATTR, mdt_getxattr),
5831 DEF_MDT_HNDL_F(0 |HABEO_REFERO, STATFS, mdt_statfs),
5832 DEF_MDT_HNDL_F(0 |MUTABOR, REINT, mdt_reint),
5833 DEF_MDT_HNDL_F(HABEO_CORPUS, CLOSE, mdt_close),
5834 DEF_MDT_HNDL_F(HABEO_CORPUS, DONE_WRITING, mdt_done_writing),
5835 DEF_MDT_HNDL_F(0 |HABEO_REFERO, PIN, mdt_pin),
5836 DEF_MDT_HNDL_0(0, SYNC, mdt_sync),
5837 DEF_MDT_HNDL_F(HABEO_CORPUS|HABEO_REFERO, IS_SUBDIR, mdt_is_subdir),
5838 #ifdef HAVE_QUOTA_SUPPORT
5839 DEF_MDT_HNDL_F(0, QUOTACHECK, mdt_quotacheck_handle),
5840 DEF_MDT_HNDL_F(0, QUOTACTL, mdt_quotactl_handle)
5844 #define DEF_OBD_HNDL(flags, name, fn) \
5845 DEF_HNDL(OBD, PING, _NET, flags, name, fn, NULL)
5848 static struct mdt_handler mdt_obd_ops[] = {
5849 DEF_OBD_HNDL(0, PING, mdt_obd_ping),
5850 DEF_OBD_HNDL(0, LOG_CANCEL, mdt_obd_log_cancel),
5851 DEF_OBD_HNDL(0, QC_CALLBACK, mdt_obd_qc_callback)
5854 #define DEF_DLM_HNDL_0(flags, name, fn) \
5855 DEF_HNDL(LDLM, ENQUEUE, , flags, name, fn, NULL)
5856 #define DEF_DLM_HNDL_F(flags, name, fn) \
5857 DEF_HNDL(LDLM, ENQUEUE, , flags, name, fn, &RQF_LDLM_ ## name)
5859 static struct mdt_handler mdt_dlm_ops[] = {
5860 DEF_DLM_HNDL_F(HABEO_CLAVIS, ENQUEUE, mdt_enqueue),
5861 DEF_DLM_HNDL_0(HABEO_CLAVIS, CONVERT, mdt_convert),
5862 DEF_DLM_HNDL_0(0, BL_CALLBACK, mdt_bl_callback),
5863 DEF_DLM_HNDL_0(0, CP_CALLBACK, mdt_cp_callback)
5866 #define DEF_LLOG_HNDL(flags, name, fn) \
5867 DEF_HNDL(LLOG, ORIGIN_HANDLE_CREATE, _NET, flags, name, fn, NULL)
5869 static struct mdt_handler mdt_llog_ops[] = {
5870 DEF_LLOG_HNDL(0, ORIGIN_HANDLE_CREATE, mdt_llog_create),
5871 DEF_LLOG_HNDL(0, ORIGIN_HANDLE_NEXT_BLOCK, mdt_llog_next_block),
5872 DEF_LLOG_HNDL(0, ORIGIN_HANDLE_READ_HEADER, mdt_llog_read_header),
5873 DEF_LLOG_HNDL(0, ORIGIN_HANDLE_WRITE_REC, NULL),
5874 DEF_LLOG_HNDL(0, ORIGIN_HANDLE_CLOSE, NULL),
5875 DEF_LLOG_HNDL(0, ORIGIN_CONNECT, NULL),
5876 DEF_LLOG_HNDL(0, CATINFO, NULL),
5877 DEF_LLOG_HNDL(0, ORIGIN_HANDLE_PREV_BLOCK, mdt_llog_prev_block),
5878 DEF_LLOG_HNDL(0, ORIGIN_HANDLE_DESTROY, mdt_llog_destroy),
5881 #define DEF_SEC_CTX_HNDL(name, fn) \
5882 DEF_HNDL(SEC_CTX, INIT, _NET, 0, name, fn, NULL)
5884 static struct mdt_handler mdt_sec_ctx_ops[] = {
5885 DEF_SEC_CTX_HNDL(INIT, mdt_sec_ctx_handle),
5886 DEF_SEC_CTX_HNDL(INIT_CONT, mdt_sec_ctx_handle),
5887 DEF_SEC_CTX_HNDL(FINI, mdt_sec_ctx_handle)
5890 static struct mdt_opc_slice mdt_regular_handlers[] = {
5892 .mos_opc_start = MDS_GETATTR,
5893 .mos_opc_end = MDS_LAST_OPC,
5894 .mos_hs = mdt_mds_ops
5897 .mos_opc_start = OBD_PING,
5898 .mos_opc_end = OBD_LAST_OPC,
5899 .mos_hs = mdt_obd_ops
5902 .mos_opc_start = LDLM_ENQUEUE,
5903 .mos_opc_end = LDLM_LAST_OPC,
5904 .mos_hs = mdt_dlm_ops
5907 .mos_opc_start = LLOG_ORIGIN_HANDLE_CREATE,
5908 .mos_opc_end = LLOG_LAST_OPC,
5909 .mos_hs = mdt_llog_ops
5912 .mos_opc_start = SEC_CTX_INIT,
5913 .mos_opc_end = SEC_LAST_OPC,
5914 .mos_hs = mdt_sec_ctx_ops
5921 static struct mdt_handler mdt_readpage_ops[] = {
5922 DEF_MDT_HNDL_F(0, CONNECT, mdt_connect),
5923 DEF_MDT_HNDL_F(HABEO_CORPUS|HABEO_REFERO, READPAGE, mdt_readpage),
5924 #ifdef HAVE_SPLIT_SUPPORT
5925 DEF_MDT_HNDL_F(HABEO_CORPUS|HABEO_REFERO, WRITEPAGE, mdt_writepage),
5929 * XXX: this is ugly and should be fixed one day, see mdc_close() for
5930 * detailed comments. --umka
5932 DEF_MDT_HNDL_F(HABEO_CORPUS, CLOSE, mdt_close),
5933 DEF_MDT_HNDL_F(HABEO_CORPUS, DONE_WRITING, mdt_done_writing),
5936 static struct mdt_opc_slice mdt_readpage_handlers[] = {
5938 .mos_opc_start = MDS_GETATTR,
5939 .mos_opc_end = MDS_LAST_OPC,
5940 .mos_hs = mdt_readpage_ops
5947 static struct mdt_handler mdt_xmds_ops[] = {
5948 DEF_MDT_HNDL_F(0, CONNECT, mdt_connect),
5949 DEF_MDT_HNDL_F(HABEO_CORPUS , GETATTR, mdt_getattr),
5950 DEF_MDT_HNDL_F(0 | MUTABOR , REINT, mdt_reint),
5951 DEF_MDT_HNDL_F(HABEO_CORPUS|HABEO_REFERO, IS_SUBDIR, mdt_is_subdir),
5954 static struct mdt_opc_slice mdt_xmds_handlers[] = {
5956 .mos_opc_start = MDS_GETATTR,
5957 .mos_opc_end = MDS_LAST_OPC,
5958 .mos_hs = mdt_xmds_ops
5961 .mos_opc_start = OBD_PING,
5962 .mos_opc_end = OBD_LAST_OPC,
5963 .mos_hs = mdt_obd_ops
5966 .mos_opc_start = SEC_CTX_INIT,
5967 .mos_opc_end = SEC_LAST_OPC,
5968 .mos_hs = mdt_sec_ctx_ops
5975 static struct mdt_handler mdt_seq_ops[] = {
5976 DEF_SEQ_HNDL_F(0, QUERY, (int (*)(struct mdt_thread_info *))seq_query)
5979 static struct mdt_opc_slice mdt_seq_handlers[] = {
5981 .mos_opc_start = SEQ_QUERY,
5982 .mos_opc_end = SEQ_LAST_OPC,
5983 .mos_hs = mdt_seq_ops
5990 static struct mdt_handler mdt_fld_ops[] = {
5991 DEF_FLD_HNDL_F(0, QUERY, (int (*)(struct mdt_thread_info *))fld_query)
5994 static struct mdt_opc_slice mdt_fld_handlers[] = {
5996 .mos_opc_start = FLD_QUERY,
5997 .mos_opc_end = FLD_LAST_OPC,
5998 .mos_hs = mdt_fld_ops
6005 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
6006 MODULE_DESCRIPTION("Lustre Meta-data Target ("LUSTRE_MDT_NAME")");
6007 MODULE_LICENSE("GPL");
6009 CFS_MODULE_PARM(mdt_num_threads, "ul", ulong, 0444,
6010 "number of mdt service threads to start");
6012 cfs_module(mdt, "0.2.0", mdt_mod_init, mdt_mod_exit);