1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
5 * Lustre Metadata Server (mds) request handler
7 * Copyright (c) 2001, 2002 Cluster File Systems, Inc.
8 * Author: Peter Braam <braam@clusterfs.com>
9 * Author: Andreas Dilger <adilger@clusterfs.com>
10 * Author: Phil Schwan <phil@clusterfs.com>
12 * This file is part of Lustre, http://www.lustre.org.
14 * Lustre is free software; you can redistribute it and/or
15 * modify it under the terms of version 2 of the GNU General Public
16 * License as published by the Free Software Foundation.
18 * Lustre is distributed in the hope that it will be useful,
19 * but WITHOUT ANY WARRANTY; without even the implied warranty of
20 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
21 * GNU General Public License for more details.
23 * You should have received a copy of the GNU General Public License
24 * along with Lustre; if not, write to the Free Software
25 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
29 #define DEBUG_SUBSYSTEM S_MDS
31 #include <linux/module.h>
32 #include <linux/lustre_mds.h>
33 #include <linux/lustre_dlm.h>
34 #include <linux/init.h>
35 #include <linux/obd_class.h>
36 #include <linux/random.h>
37 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
38 #include <linux/buffer_head.h>
41 static kmem_cache_t *mds_file_cache;
43 extern int mds_get_lovtgts(struct obd_device *obd, int tgt_count,
44 obd_uuid_t *uuidarray);
45 extern int mds_get_lovdesc(struct obd_device *obd, struct lov_desc *desc);
46 extern int mds_update_last_rcvd(struct mds_obd *mds, void *handle,
47 struct ptlrpc_request *req);
48 static int mds_cleanup(struct obd_device * obddev);
50 inline struct mds_obd *mds_req2mds(struct ptlrpc_request *req)
52 return &req->rq_export->exp_obd->u.mds;
55 static int mds_bulk_timeout(void *data)
57 struct ptlrpc_bulk_desc *desc = data;
60 CERROR("(not yet) starting recovery of client %p\n", desc->bd_client);
64 /* Assumes caller has already pushed into the kernel filesystem context */
65 static int mds_sendpage(struct ptlrpc_request *req, struct file *file,
69 struct mds_obd *mds = mds_req2mds(req);
70 struct ptlrpc_bulk_desc *desc;
71 struct ptlrpc_bulk_page *bulk;
72 struct l_wait_info lwi;
76 desc = ptlrpc_prep_bulk(req->rq_connection);
78 GOTO(out, rc = -ENOMEM);
80 bulk = ptlrpc_prep_bulk_page(desc);
82 GOTO(cleanup_bulk, rc = -ENOMEM);
84 OBD_ALLOC(buf, PAGE_SIZE);
86 GOTO(cleanup_bulk, rc = -ENOMEM);
88 rc = mds_fs_readpage(mds, file, buf, PAGE_SIZE, (loff_t *)&offset);
91 GOTO(cleanup_buf, rc = -EIO);
93 bulk->bp_xid = req->rq_xid;
95 bulk->bp_buflen = PAGE_SIZE;
96 desc->bd_portal = MDS_BULK_PORTAL;
98 rc = ptlrpc_send_bulk(desc);
100 GOTO(cleanup_buf, rc);
102 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
103 CERROR("obd_fail_loc=%x, fail operation rc=%d\n",
104 OBD_FAIL_MDS_SENDPAGE, rc);
105 ptlrpc_abort_bulk(desc);
106 GOTO(cleanup_buf, rc);
109 lwi = LWI_TIMEOUT(obd_timeout * HZ, mds_bulk_timeout, desc);
110 rc = l_wait_event(desc->bd_waitq, desc->bd_flags & PTL_BULK_FL_SENT, &lwi);
112 if (rc != -ETIMEDOUT)
114 GOTO(cleanup_buf, rc);
119 OBD_FREE(buf, PAGE_SIZE);
121 ptlrpc_free_bulk(desc);
127 * Look up a named entry in a directory, and get an LDLM lock on it.
128 * 'dir' is a inode for which an LDLM lock has already been taken.
130 * If we do not need an exclusive or write lock on this entry (e.g.
131 * a read lock for attribute lookup only) then we do not hold the
132 * directory on return. It is up to the caller to know what type
133 * of lock it is getting, and clean up appropriately.
135 struct dentry *mds_name2locked_dentry(struct obd_device *obd,
136 struct dentry *dir, struct vfsmount **mnt,
137 char *name, int namelen, int lock_mode,
138 struct lustre_handle *lockh,
141 struct dentry *dchild;
143 __u64 res_id[3] = {0};
146 down(&dir->d_inode->i_sem);
147 dchild = lookup_one_len(name, dir, namelen);
148 if (IS_ERR(dchild)) {
149 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
150 up(&dir->d_inode->i_sem);
154 if (dir_lock_mode != LCK_EX && dir_lock_mode != LCK_PW) {
155 up(&dir->d_inode->i_sem);
156 ldlm_lock_decref(lockh, dir_lock_mode);
159 if (lock_mode == 0 || !dchild->d_inode)
162 res_id[0] = dchild->d_inode->i_ino;
163 rc = ldlm_match_or_enqueue(NULL, NULL, obd->obd_namespace, NULL,
164 res_id, LDLM_PLAIN, NULL, 0, lock_mode,
165 &flags, ldlm_completion_ast,
166 mds_blocking_ast, NULL, 0, lockh);
167 if (rc != ELDLM_OK) {
169 up(&dir->d_inode->i_sem);
170 RETURN(ERR_PTR(-ENOLCK)); /* XXX translate ldlm code */
176 struct dentry *mds_fid2locked_dentry(struct obd_device *obd, struct ll_fid *fid,
177 struct vfsmount **mnt, int lock_mode,
178 struct lustre_handle *lockh)
180 struct mds_obd *mds = &obd->u.mds;
181 struct dentry *de = mds_fid2dentry(mds, fid, mnt), *retval = de;
183 __u64 res_id[3] = {0};
189 res_id[0] = de->d_inode->i_ino;
190 rc = ldlm_match_or_enqueue(NULL, NULL, obd->obd_namespace, NULL,
191 res_id, LDLM_PLAIN, NULL, 0, lock_mode,
192 &flags, ldlm_completion_ast,
193 mds_blocking_ast, NULL, 0, lockh);
194 if (rc != ELDLM_OK) {
196 retval = ERR_PTR(-ENOLCK); /* XXX translate ldlm code */
202 #ifndef DCACHE_DISCONNECTED
203 #define DCACHE_DISCONNECTED DCACHE_NFSD_DISCONNECTED
206 /* Look up an entry by inode number. */
207 struct dentry *mds_fid2dentry(struct mds_obd *mds, struct ll_fid *fid,
208 struct vfsmount **mnt)
210 /* stolen from NFS */
211 struct super_block *sb = mds->mds_sb;
212 unsigned long ino = fid->id;
213 __u32 generation = fid->generation;
215 struct list_head *lp;
216 struct dentry *result;
219 RETURN(ERR_PTR(-ESTALE));
221 inode = iget(sb, ino);
223 RETURN(ERR_PTR(-ENOMEM));
225 CDEBUG(D_DENTRY, "--> mds_fid2dentry: sb %p\n", inode->i_sb);
227 if (is_bad_inode(inode) ||
228 (generation && inode->i_generation != generation)) {
229 /* we didn't find the right inode.. */
230 CERROR("bad inode %lu, link: %d ct: %d or version %u/%u\n",
231 inode->i_ino, inode->i_nlink,
232 atomic_read(&inode->i_count), inode->i_generation,
236 RETURN(ERR_PTR(-ESTALE));
239 /* now to find a dentry. If possible, get a well-connected one */
241 *mnt = mds->mds_vfsmnt;
242 spin_lock(&dcache_lock);
243 list_for_each(lp, &inode->i_dentry) {
244 result = list_entry(lp, struct dentry, d_alias);
245 if (!(result->d_flags & DCACHE_DISCONNECTED)) {
247 result->d_vfs_flags |= DCACHE_REFERENCED;
248 spin_unlock(&dcache_lock);
255 spin_unlock(&dcache_lock);
256 result = d_alloc_root(inode);
257 if (result == NULL) {
259 return ERR_PTR(-ENOMEM);
263 result->d_flags |= DCACHE_DISCONNECTED;
267 /* Establish a connection to the MDS.
269 * This will set up an export structure for the client to hold state data
270 * about that client, like open files, the last operation number it did
271 * on the server, etc.
273 static int mds_connect(struct lustre_handle *conn, struct obd_device *obd,
274 obd_uuid_t cluuid, struct recovd_obd *recovd,
275 ptlrpc_recovery_cb_t recover)
277 struct obd_export *exp;
278 struct mds_export_data *med;
279 struct mds_client_data *mcd;
284 if (!conn || !obd || !cluuid)
289 spin_lock(&obd->obd_dev_lock);
290 list_for_each(p, &obd->obd_exports) {
291 exp = list_entry(p, struct obd_export, exp_obd_chain);
292 mcd = exp->exp_mds_data.med_mcd;
293 if (!memcmp(cluuid, mcd->mcd_uuid, sizeof(mcd->mcd_uuid))) {
294 LASSERT(exp->exp_obd == obd);
296 if (!list_empty(&exp->exp_conn_chain)) {
297 CERROR("existing uuid/export, list not empty!\n");
298 spin_unlock(&obd->obd_dev_lock);
299 /* XXX should we MOD_DEC_USE_COUNT; here? */
302 conn->addr = (__u64) (unsigned long)exp;
303 conn->cookie = exp->exp_cookie;
304 spin_unlock(&obd->obd_dev_lock);
305 CDEBUG(D_INFO, "existing export for UUID '%s' at %p\n",
307 CDEBUG(D_IOCTL,"connect: addr %Lx cookie %Lx\n",
308 (long long)conn->addr, (long long)conn->cookie);
313 spin_unlock(&obd->obd_dev_lock);
314 /* XXX There is a small race between checking the list and adding a
315 * new connection for the same UUID, but the real threat (list
316 * corruption when multiple different clients connect) is solved.
318 rc = class_connect(conn, obd, cluuid);
321 exp = class_conn2export(conn);
323 med = &exp->exp_mds_data;
325 OBD_ALLOC(mcd, sizeof(*mcd));
327 CERROR("mds: out of memory for client data\n");
328 GOTO(out_export, rc = -ENOMEM);
331 memcpy(mcd->mcd_uuid, cluuid, sizeof(mcd->mcd_uuid));
334 INIT_LIST_HEAD(&med->med_open_head);
335 spin_lock_init(&med->med_open_lock);
337 rc = mds_client_add(med, -1);
344 OBD_FREE(mcd, sizeof(*mcd));
346 class_disconnect(conn);
353 static int mds_disconnect(struct lustre_handle *conn)
355 struct obd_export *export = class_conn2export(conn);
359 #warning "Mike: we need to close all files opened on med_open_head"
360 ldlm_cancel_locks_for_export(export);
361 mds_client_free(export);
363 rc = class_disconnect(conn);
370 static int mds_getstatus(struct ptlrpc_request *req)
372 struct mds_obd *mds = mds_req2mds(req);
373 struct mds_body *body;
374 int rc, size = sizeof(*body);
377 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
378 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK)) {
379 CERROR("mds: out of memory for message: size=%d\n", size);
380 req->rq_status = -ENOMEM;
384 /* Anything we need to do here with the client's trans no or so? */
385 body = lustre_msg_buf(req->rq_repmsg, 0);
386 memcpy(&body->fid1, &mds->mds_rootfid, sizeof(body->fid1));
388 /* the last_committed and last_xid fields are filled in for all
389 * replies already - no need to do so here also.
394 static int mds_getlovinfo(struct ptlrpc_request *req)
396 struct mds_obd *mds = mds_req2mds(req);
397 struct mds_status_req *streq;
398 struct lov_desc *desc;
400 int rc, size[2] = {sizeof(*desc)};
403 streq = lustre_msg_buf(req->rq_reqmsg, 0);
404 streq->flags = NTOH__u32(streq->flags);
405 streq->repbuf = NTOH__u32(streq->repbuf);
406 size[1] = streq->repbuf;
408 rc = lustre_pack_msg(2, size, NULL, &req->rq_replen, &req->rq_repmsg);
410 CERROR("mds: out of memory for message: size=%d\n", size[1]);
411 req->rq_status = -ENOMEM;
415 desc = lustre_msg_buf(req->rq_repmsg, 0);
416 rc = mds_get_lovdesc(req->rq_obd, desc);
418 CERROR("mds_get_lovdesc error %d", rc);
423 tgt_count = le32_to_cpu(desc->ld_tgt_count);
424 if (tgt_count * sizeof(obd_uuid_t) > streq->repbuf) {
425 CERROR("too many targets, enlarge client buffers\n");
426 req->rq_status = -ENOSPC;
430 mds->mds_max_mdsize = sizeof(struct lov_mds_md) +
431 tgt_count * sizeof(struct lov_object_id);
432 rc = mds_get_lovtgts(req->rq_obd, tgt_count,
433 lustre_msg_buf(req->rq_repmsg, 1));
435 CERROR("get_lovtgts error %d\n", rc);
442 int mds_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
443 void *data, __u32 data_len, int flag)
448 if (flag == LDLM_CB_CANCELING) {
449 /* Don't need to do anything here. */
453 /* XXX layering violation! -phil */
454 l_lock(&lock->l_resource->lr_namespace->ns_lock);
455 lock->l_flags |= LDLM_FL_CBPENDING;
456 do_ast = (!lock->l_readers && !lock->l_writers);
457 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
460 struct lustre_handle lockh;
463 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
464 ldlm_lock2handle(lock, &lockh);
465 rc = ldlm_cli_cancel(&lockh);
467 CERROR("ldlm_cli_cancel: %d\n", rc);
469 LDLM_DEBUG(lock, "Lock still has references, will be"
474 static int mds_getattr_internal(struct mds_obd *mds, struct dentry *dentry,
475 struct ptlrpc_request *req,
476 struct mds_body *reqbody, int reply_off)
478 struct mds_body *body;
479 struct inode *inode = dentry->d_inode;
486 body = lustre_msg_buf(req->rq_repmsg, reply_off);
488 mds_pack_inode2fid(&body->fid1, inode);
489 mds_pack_inode2body(body, inode);
491 if (S_ISREG(inode->i_mode)) {
492 struct lov_mds_md *lmm;
494 lmm = lustre_msg_buf(req->rq_repmsg, reply_off + 1);
495 lmm->lmm_easize = mds->mds_max_mdsize;
496 rc = mds_fs_get_md(mds, inode, lmm);
501 CERROR("mds_fs_get_md failed: %d\n", rc);
504 body->valid |= OBD_MD_FLEASIZE;
505 } else if (S_ISLNK(inode->i_mode) && reqbody->valid & OBD_MD_LINKNAME) {
506 char *symname = lustre_msg_buf(req->rq_repmsg, reply_off + 1);
507 int len = req->rq_repmsg->buflens[reply_off + 1];
509 rc = inode->i_op->readlink(dentry, symname, len);
511 CERROR("readlink failed: %d\n", rc);
514 CDEBUG(D_INODE, "read symlink dest %s\n", symname);
516 body->valid |= OBD_MD_LINKNAME;
521 static int mds_getattr_name(int offset, struct ptlrpc_request *req)
523 struct mds_obd *mds = mds_req2mds(req);
524 struct obd_device *obd = req->rq_export->exp_obd;
525 struct obd_run_ctxt saved;
526 struct mds_body *body;
527 struct dentry *de = NULL, *dchild = NULL;
529 struct lustre_handle lockh;
531 int namelen, flags, lock_mode, rc = 0;
533 __u64 res_id[3] = {0, 0, 0};
536 LASSERT(!strcmp(req->rq_export->exp_obd->obd_type->typ_name, "mds"));
538 if (req->rq_reqmsg->bufcount <= offset + 1) {
540 GOTO(out_pre_de, rc = -EINVAL);
543 body = lustre_msg_buf(req->rq_reqmsg, offset);
544 name = lustre_msg_buf(req->rq_reqmsg, offset + 1);
545 namelen = req->rq_reqmsg->buflens[offset + 1];
546 /* requests were at offset 2, replies go back at 1 */
550 uc.ouc_fsuid = body->fsuid;
551 uc.ouc_fsgid = body->fsgid;
552 push_ctxt(&saved, &mds->mds_ctxt, &uc);
553 de = mds_fid2dentry(mds, &body->fid1, NULL);
556 GOTO(out_pre_de, rc = -ESTALE);
560 CDEBUG(D_INODE, "parent ino %ld, name %*s\n", dir->i_ino,namelen,name);
562 lock_mode = (req->rq_reqmsg->opc == MDS_REINT) ? LCK_CW : LCK_PW;
563 res_id[0] = dir->i_ino;
565 rc = ldlm_lock_match(obd->obd_namespace, res_id, LDLM_PLAIN,
566 NULL, 0, lock_mode, &lockh);
568 LDLM_DEBUG_NOLOCK("enqueue res "LPU64, res_id[0]);
569 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, NULL,
570 res_id, LDLM_PLAIN, NULL, 0, lock_mode,
571 &flags, ldlm_completion_ast,
572 mds_blocking_ast, NULL, 0, &lockh);
573 if (rc != ELDLM_OK) {
574 CERROR("lock enqueue: err: %d\n", rc);
575 GOTO(out_create_de, rc = -EIO);
578 ldlm_lock_dump((void *)(unsigned long)lockh.addr);
581 dchild = lookup_one_len(name, de, namelen - 1);
582 if (IS_ERR(dchild)) {
583 CDEBUG(D_INODE, "child lookup error %ld\n", PTR_ERR(dchild));
586 GOTO(out_create_dchild, rc = -ESTALE);
589 rc = mds_getattr_internal(mds, dchild, req, body, offset);
595 ldlm_lock_decref(&lockh, lock_mode);
604 static int mds_getattr(int offset, struct ptlrpc_request *req)
606 struct mds_obd *mds = mds_req2mds(req);
607 struct obd_run_ctxt saved;
610 struct mds_body *body;
612 int rc = 0, size[2] = {sizeof(*body)}, bufcount = 1;
615 body = lustre_msg_buf(req->rq_reqmsg, offset);
616 uc.ouc_fsuid = body->fsuid;
617 uc.ouc_fsgid = body->fsgid;
618 push_ctxt(&saved, &mds->mds_ctxt, &uc);
619 de = mds_fid2dentry(mds, &body->fid1, NULL);
621 rc = req->rq_status = -ENOENT;
622 GOTO(out_pop, PTR_ERR(de));
626 if (S_ISREG(body->fid1.f_type)) {
628 size[1] = mds->mds_max_mdsize;
629 } else if (body->valid & OBD_MD_LINKNAME) {
631 size[1] = MIN(inode->i_size + 1, body->size);
632 CDEBUG(D_INODE, "symlink size: %d, reply space: %d\n",
633 inode->i_size + 1, body->size);
636 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
637 CERROR("failed GETATTR_PACK test\n");
638 req->rq_status = -ENOMEM;
639 GOTO(out, rc = -ENOMEM);
642 rc = lustre_pack_msg(bufcount, size, NULL, &req->rq_replen,
645 CERROR("out of memory or FAIL_MDS_GETATTR_PACK\n");
650 req->rq_status = mds_getattr_internal(mds, de, req, body, 0);
659 static int mds_statfs(struct ptlrpc_request *req)
661 struct mds_obd *mds = mds_req2mds(req);
662 struct obd_statfs *osfs;
664 int rc, size = sizeof(*osfs);
667 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
669 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
670 CERROR("mds: statfs lustre_pack_msg failed: rc = %d\n", rc);
674 rc = vfs_statfs(mds->mds_sb, &sfs);
676 CERROR("mds: statfs failed: rc %d\n", rc);
679 osfs = lustre_msg_buf(req->rq_repmsg, 0);
680 memset(osfs, 0, size);
681 statfs_pack(osfs, &sfs);
682 obd_statfs_pack(osfs, osfs);
689 static struct mds_file_data *mds_handle2mfd(struct lustre_handle *handle)
691 struct mds_file_data *mfd = NULL;
693 if (!handle || !handle->addr)
696 mfd = (struct mds_file_data *)(unsigned long)(handle->addr);
697 if (!kmem_cache_validate(mds_file_cache, mfd))
700 if (mfd->mfd_servercookie != handle->cookie)
706 static int mds_store_ea(struct mds_obd *mds, struct ptlrpc_request *req,
707 struct mds_body *body, struct dentry *de,
708 struct lov_mds_md *lmm)
710 struct obd_run_ctxt saved;
715 uc.ouc_fsuid = body->fsuid;
716 uc.ouc_fsgid = body->fsgid;
717 push_ctxt(&saved, &mds->mds_ctxt, &uc);
718 handle = mds_fs_start(mds, de->d_inode, MDS_FSOP_SETATTR);
720 GOTO(out_ea, rc = -ENOMEM);
722 rc = mds_fs_set_md(mds, de->d_inode, handle, lmm);
724 rc = mds_update_last_rcvd(mds, handle, req);
726 rc2 = mds_fs_commit(mds, de->d_inode, handle);
735 static int mds_open(struct ptlrpc_request *req)
737 struct mds_obd *mds = mds_req2mds(req);
738 struct mds_body *body;
739 struct mds_export_data *med;
740 struct mds_file_data *mfd;
743 struct vfsmount *mnt;
745 struct list_head *tmp;
746 int rc, size = sizeof(*body);
749 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
750 CERROR("test case OBD_FAIL_MDS_OPEN_PACK\n");
751 req->rq_status = -ENOMEM;
755 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
757 CERROR("mds: pack error: rc = %d\n", rc);
762 body = lustre_msg_buf(req->rq_reqmsg, 0);
764 /* was this animal open already and the client lost the reply? */
765 /* XXX need some way to detect a reopen, to avoid locked list walks */
766 med = &req->rq_export->exp_mds_data;
767 spin_lock(&med->med_open_lock);
768 list_for_each(tmp, &med->med_open_head) {
769 mfd = list_entry(tmp, typeof(*mfd), mfd_list);
770 if (!memcmp(&mfd->mfd_clienthandle, &body->handle,
771 sizeof(mfd->mfd_clienthandle)) &&
772 body->fid1.id == mfd->mfd_file->f_dentry->d_inode->i_ino) {
773 de = mfd->mfd_file->f_dentry;
774 spin_unlock(&med->med_open_lock);
775 CERROR("Re opening "LPD64"\n", body->fid1.id);
776 GOTO(out_pack, rc = 0);
779 spin_unlock(&med->med_open_lock);
781 mfd = kmem_cache_alloc(mds_file_cache, GFP_KERNEL);
783 CERROR("mds: out of memory\n");
784 req->rq_status = -ENOMEM;
788 de = mds_fid2dentry(mds, &body->fid1, &mnt);
790 GOTO(out_free, rc = PTR_ERR(de));
792 /* check if this inode has seen a delayed object creation */
793 if (req->rq_reqmsg->bufcount > 1) {
794 struct lov_mds_md *lmm = lustre_msg_buf(req->rq_reqmsg, 1);
796 rc = mds_store_ea(mds, req, body, de, lmm);
805 /* dentry_open does a dput(de) and mntput(mnt) on error */
806 file = dentry_open(de, mnt, flags & ~O_DIRECT);
812 file->private_data = mfd;
813 mfd->mfd_file = file;
814 memcpy(&mfd->mfd_clienthandle, &body->handle, sizeof(body->handle));
815 get_random_bytes(&mfd->mfd_servercookie, sizeof(mfd->mfd_servercookie));
816 spin_lock(&med->med_open_lock);
817 list_add(&mfd->mfd_list, &med->med_open_head);
818 spin_unlock(&med->med_open_lock);
821 body = lustre_msg_buf(req->rq_repmsg, 0);
822 mds_pack_inode2fid(&body->fid1, de->d_inode);
823 mds_pack_inode2body(body, de->d_inode);
824 body->handle.addr = (__u64)(unsigned long)mfd;
825 body->handle.cookie = mfd->mfd_servercookie;
826 CDEBUG(D_INODE, "llite file "LPX64": addr %p, cookie "LPX64"\n",
827 mfd->mfd_clienthandle.addr, mfd, mfd->mfd_servercookie);
831 mfd->mfd_servercookie = DEAD_HANDLE_MAGIC;
832 kmem_cache_free(mds_file_cache, mfd);
837 static int mds_close(struct ptlrpc_request *req)
839 struct mds_export_data *med = &req->rq_export->exp_mds_data;
840 struct mds_body *body;
842 struct mds_file_data *mfd;
846 body = lustre_msg_buf(req->rq_reqmsg, 0);
848 mfd = mds_handle2mfd(&body->handle);
850 CERROR("no handle for file close "LPD64
851 ": addr "LPX64", cookie "LPX64"\n",
852 body->fid1.id, body->handle.addr, body->handle.cookie);
856 file = mfd->mfd_file;
857 LASSERT(file->private_data == mfd);
859 spin_lock(&med->med_open_lock);
860 list_del(&mfd->mfd_list);
861 spin_unlock(&med->med_open_lock);
862 mfd->mfd_servercookie = DEAD_HANDLE_MAGIC;
863 kmem_cache_free(mds_file_cache, mfd);
865 req->rq_status = filp_close(file, 0);
867 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
868 CERROR("test case OBD_FAIL_MDS_CLOSE_PACK\n");
869 req->rq_status = -ENOMEM;
873 rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen, &req->rq_repmsg);
875 CERROR("mds: lustre_pack_msg: rc = %d\n", rc);
882 static int mds_readpage(struct ptlrpc_request *req)
884 struct mds_obd *mds = mds_req2mds(req);
885 struct vfsmount *mnt;
888 struct mds_body *body, *repbody;
889 struct obd_run_ctxt saved;
890 int rc, size = sizeof(*body);
894 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
895 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK)) {
896 CERROR("mds: out of memory\n");
897 GOTO(out, rc = -ENOMEM);
900 body = lustre_msg_buf(req->rq_reqmsg, 0);
901 uc.ouc_fsuid = body->fsuid;
902 uc.ouc_fsgid = body->fsgid;
903 push_ctxt(&saved, &mds->mds_ctxt, &uc);
904 de = mds_fid2dentry(mds, &body->fid1, &mnt);
906 GOTO(out_pop, rc = PTR_ERR(de));
908 CDEBUG(D_INODE, "ino %ld\n", de->d_inode->i_ino);
910 file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
911 /* note: in case of an error, dentry_open puts dentry */
913 GOTO(out_pop, rc = PTR_ERR(file));
915 repbody = lustre_msg_buf(req->rq_repmsg, 0);
916 repbody->size = file->f_dentry->d_inode->i_size;
917 repbody->valid = OBD_MD_FLSIZE;
919 /* to make this asynchronous make sure that the handling function
920 doesn't send a reply when this function completes. Instead a
921 callback function would send the reply */
922 /* note: in case of an error, dentry_open puts dentry */
923 rc = mds_sendpage(req, file, body->size);
933 int mds_reint(int offset, struct ptlrpc_request *req)
936 struct mds_update_record rec;
938 rc = mds_update_unpack(req, offset, &rec);
939 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
940 CERROR("invalid record\n");
941 req->rq_status = -EINVAL;
944 /* rc will be used to interrupt a for loop over multiple records */
945 rc = mds_reint_rec(&rec, offset, req);
949 int mds_handle(struct ptlrpc_request *req)
954 rc = lustre_unpack_msg(req->rq_reqmsg, req->rq_reqlen);
955 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_HANDLE_UNPACK)) {
956 CERROR("lustre_mds: Invalid request\n");
960 if (req->rq_reqmsg->opc != MDS_CONNECT && req->rq_export == NULL)
961 GOTO(out, rc = -ENOTCONN);
963 LASSERT(!strcmp(req->rq_obd->obd_type->typ_name, "mds"));
965 switch (req->rq_reqmsg->opc) {
967 CDEBUG(D_INODE, "connect\n");
968 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
969 rc = target_handle_connect(req);
973 CDEBUG(D_INODE, "disconnect\n");
974 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
975 rc = target_handle_disconnect(req);
979 CDEBUG(D_INODE, "getstatus\n");
980 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
981 rc = mds_getstatus(req);
985 CDEBUG(D_INODE, "getlovinfo\n");
986 rc = mds_getlovinfo(req);
990 CDEBUG(D_INODE, "getattr\n");
991 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
992 rc = mds_getattr(0, req);
996 CDEBUG(D_INODE, "statfs\n");
997 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
998 rc = mds_statfs(req);
1002 CDEBUG(D_INODE, "readpage\n");
1003 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
1004 rc = mds_readpage(req);
1006 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE))
1011 int size = sizeof(struct mds_body);
1012 CDEBUG(D_INODE, "reint\n");
1013 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
1015 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
1018 req->rq_status = rc;
1021 rc = mds_reint(0, req);
1022 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET_REP, 0);
1027 CDEBUG(D_INODE, "open\n");
1028 OBD_FAIL_RETURN(OBD_FAIL_MDS_OPEN_NET, 0);
1033 CDEBUG(D_INODE, "close\n");
1034 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
1035 rc = mds_close(req);
1039 CDEBUG(D_INODE, "enqueue\n");
1040 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
1041 rc = ldlm_handle_enqueue(req);
1047 CDEBUG(D_INODE, "convert\n");
1048 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
1049 rc = ldlm_handle_convert(req);
1055 CDEBUG(D_INODE, "cancel\n");
1056 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CANCEL, 0);
1057 rc = ldlm_handle_cancel(req);
1061 case LDLM_BL_CALLBACK:
1062 case LDLM_CP_CALLBACK:
1063 CDEBUG(D_INODE, "callback\n");
1064 CERROR("callbacks should not happen on MDS\n");
1066 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
1069 rc = ptlrpc_error(req->rq_svc, req);
1076 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1077 struct mds_obd *mds = mds_req2mds(req);
1079 req->rq_repmsg->last_xid =
1080 HTON__u64(le64_to_cpu(med->med_mcd->mcd_last_xid));
1081 req->rq_repmsg->last_committed =
1082 HTON__u64(mds->mds_last_committed);
1083 CDEBUG(D_INFO, "last_rcvd ~%Lu, last_committed %Lu, xid %d\n",
1084 (unsigned long long)mds->mds_last_rcvd,
1085 (unsigned long long)mds->mds_last_committed,
1086 cpu_to_le32(req->rq_xid));
1090 CERROR("mds: processing error (opcode %d): %d\n",
1091 req->rq_reqmsg->opc, rc);
1092 ptlrpc_error(req->rq_svc, req);
1094 CDEBUG(D_NET, "sending reply\n");
1095 ptlrpc_reply(req->rq_svc, req);
1100 /* Update the server data on disk. This stores the new mount_count and
1101 * also the last_rcvd value to disk. If we don't have a clean shutdown,
1102 * then the server last_rcvd value may be less than that of the clients.
1103 * This will alert us that we may need to do client recovery.
1105 * Assumes we are already in the server filesystem context.
1107 * Also assumes for mds_last_rcvd that we are not modifying it (no locking).
1110 int mds_update_server_data(struct mds_obd *mds)
1112 struct mds_server_data *msd = mds->mds_server_data;
1113 struct file *filp = mds->mds_rcvd_filp;
1117 msd->msd_last_rcvd = cpu_to_le64(mds->mds_last_rcvd);
1118 msd->msd_mount_count = cpu_to_le64(mds->mds_mount_count);
1120 CDEBUG(D_SUPER, "MDS mount_count is %Lu, last_rcvd is %Lu\n",
1121 (unsigned long long)mds->mds_mount_count,
1122 (unsigned long long)mds->mds_last_rcvd);
1123 rc = lustre_fwrite(filp, (char *)msd, sizeof(*msd), &off);
1124 if (rc != sizeof(*msd)) {
1125 CERROR("error writing MDS server data: rc = %d\n", rc);
1130 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
1131 rc = fsync_dev(filp->f_dentry->d_inode->i_rdev);
1133 rc = file_fsync(filp, filp->f_dentry, 1);
1136 CERROR("error flushing MDS server data: rc = %d\n", rc);
1141 /* Do recovery actions for the MDS */
1142 static int mds_recover(struct obd_device *obddev)
1144 struct mds_obd *mds = &obddev->u.mds;
1145 struct obd_run_ctxt saved;
1148 /* This happens at the end when recovery is complete */
1149 ++mds->mds_mount_count;
1150 push_ctxt(&saved, &mds->mds_ctxt, NULL);
1151 rc = mds_update_server_data(mds);
1157 #define MDS_NUM_THREADS 8
1158 /* mount the file system (secretly) */
1159 static int mds_setup(struct obd_device *obddev, obd_count len, void *buf)
1162 struct obd_ioctl_data* data = buf;
1163 struct mds_obd *mds = &obddev->u.mds;
1164 struct vfsmount *mnt;
1169 #ifdef CONFIG_DEV_RDONLY
1170 dev_clear_rdonly(2);
1172 if (!data->ioc_inlbuf1 || !data->ioc_inlbuf2)
1173 GOTO(err_dec, rc = -EINVAL);
1175 mds->mds_fstype = strdup(data->ioc_inlbuf2);
1177 mnt = do_kern_mount(mds->mds_fstype, 0, data->ioc_inlbuf1, NULL);
1180 CERROR("do_kern_mount failed: rc = %d\n", rc);
1181 GOTO(err_kfree, rc);
1184 mds->mds_sb = mnt->mnt_root->d_inode->i_sb;
1186 GOTO(err_put, rc = -ENODEV);
1188 spin_lock_init(&mds->mds_last_lock);
1189 mds->mds_max_mdsize = sizeof(struct lov_mds_md);
1190 rc = mds_fs_setup(obddev, mnt);
1192 CERROR("MDS filesystem method init failed: rc = %d\n", rc);
1196 mds->mds_service = ptlrpc_init_svc(MDS_NEVENTS, MDS_NBUFS,
1197 MDS_BUFSIZE, MDS_MAXREQSIZE,
1198 MDS_REQUEST_PORTAL, MDC_REPLY_PORTAL,
1199 "self", mds_handle, "mds");
1200 if (!mds->mds_service) {
1201 CERROR("failed to start service\n");
1202 GOTO(err_fs, rc = -EINVAL);
1205 obddev->obd_namespace =
1206 ldlm_namespace_new("mds_server", LDLM_NAMESPACE_SERVER);
1207 if (obddev->obd_namespace == NULL) {
1208 mds_cleanup(obddev);
1209 GOTO(err_svc, rc = -ENOMEM);
1212 for (i = 0; i < MDS_NUM_THREADS; i++) {
1214 sprintf(name, "lustre_MDS_%02d", i);
1215 rc = ptlrpc_start_thread(obddev, mds->mds_service, name);
1217 CERROR("cannot start MDS thread #%d: rc %d\n", i, rc);
1218 GOTO(err_thread, rc);
1222 rc = mds_recover(obddev);
1224 GOTO(err_thread, rc);
1226 ptlrpc_init_client(LDLM_REQUEST_PORTAL, LDLM_REPLY_PORTAL,
1227 "mds_ldlm_client", &obddev->obd_ldlm_client);
1232 ptlrpc_stop_all_threads(mds->mds_service);
1234 ptlrpc_unregister_service(mds->mds_service);
1236 mds_fs_cleanup(obddev);
1239 mntput(mds->mds_vfsmnt);
1243 kfree(mds->mds_fstype);
1249 static int mds_cleanup(struct obd_device *obddev)
1251 struct super_block *sb;
1252 struct mds_obd *mds = &obddev->u.mds;
1253 struct obd_run_ctxt saved;
1256 ptlrpc_stop_all_threads(mds->mds_service);
1257 ptlrpc_unregister_service(mds->mds_service);
1263 push_ctxt(&saved, &mds->mds_ctxt, NULL);
1264 mds_update_server_data(mds);
1266 if (mds->mds_rcvd_filp) {
1267 int rc = filp_close(mds->mds_rcvd_filp, 0);
1268 mds->mds_rcvd_filp = NULL;
1271 CERROR("last_rcvd file won't close, rc=%d\n", rc);
1276 mntput(mds->mds_vfsmnt);
1278 kfree(mds->mds_fstype);
1280 ldlm_namespace_free(obddev->obd_namespace);
1283 #ifdef CONFIG_DEV_RDONLY
1284 dev_clear_rdonly(2);
1286 mds_fs_cleanup(obddev);
1292 static int ldlm_intent_policy(struct ldlm_lock *lock, void *req_cookie,
1293 ldlm_mode_t mode, void *data)
1295 struct ptlrpc_request *req = req_cookie;
1302 if (req->rq_reqmsg->bufcount > 1) {
1303 /* an intent needs to be considered */
1304 struct ldlm_intent *it = lustre_msg_buf(req->rq_reqmsg, 1);
1305 struct mds_obd *mds= &req->rq_export->exp_obd->u.mds;
1306 struct mds_body *mds_rep;
1307 struct ldlm_reply *rep;
1308 __u64 new_resid[3] = {0, 0, 0}, old_res;
1309 int rc, size[3] = {sizeof(struct ldlm_reply),
1310 sizeof(struct mds_body),
1311 mds->mds_max_mdsize};
1313 it->opc = NTOH__u64(it->opc);
1315 LDLM_DEBUG(lock, "intent policy, opc: %s",
1316 ldlm_it2str(it->opc));
1318 rc = lustre_pack_msg(3, size, NULL, &req->rq_replen,
1321 rc = req->rq_status = -ENOMEM;
1325 rep = lustre_msg_buf(req->rq_repmsg, 0);
1326 rep->lock_policy_res1 = 1;
1328 /* execute policy */
1329 switch ((long)it->opc) {
1330 case IT_CREAT|IT_OPEN:
1331 rc = mds_reint(2, req);
1332 if (rc || (req->rq_status != 0 &&
1333 req->rq_status != -EEXIST)) {
1334 rep->lock_policy_res2 = req->rq_status;
1335 RETURN(ELDLM_LOCK_ABORTED);
1346 rc = mds_reint(2, req);
1347 if (rc || (req->rq_status != 0 &&
1348 req->rq_status != -EISDIR &&
1349 req->rq_status != -ENOTDIR)) {
1350 rep->lock_policy_res2 = req->rq_status;
1351 RETURN(ELDLM_LOCK_ABORTED);
1362 rc = mds_getattr_name(2, req);
1363 /* FIXME: we need to sit down and decide on who should
1364 * set req->rq_status, who should return negative and
1365 * positive return values, and what they all mean. */
1366 if (rc || req->rq_status != 0) {
1367 rep->lock_policy_res2 = req->rq_status;
1368 RETURN(ELDLM_LOCK_ABORTED);
1371 case IT_READDIR|IT_OPEN:
1375 CERROR("Unhandled intent\n");
1379 /* We don't bother returning a lock to the client for a file
1380 * or directory we are removing.
1382 * As for link and rename, there is no reason for the client
1383 * to get a lock on the target at this point. If they are
1384 * going to modify the file/directory later they will get a
1385 * lock at that time.
1387 if (it->opc & (IT_UNLINK | IT_RMDIR | IT_LINK | IT_LINK2 |
1388 IT_RENAME | IT_RENAME2))
1389 RETURN(ELDLM_LOCK_ABORTED);
1391 rep->lock_policy_res2 = req->rq_status;
1392 mds_rep = lustre_msg_buf(req->rq_repmsg, 1);
1394 /* If the client is about to open a file that doesn't have an MD
1395 * stripe record, it's going to need a write lock. */
1396 if (it->opc & IT_OPEN) {
1397 struct lov_mds_md *lmm =
1398 lustre_msg_buf(req->rq_repmsg, 2);
1399 if (lmm->lmm_easize == 0) {
1400 LDLM_DEBUG(lock, "open with no EA; returning PW"
1402 lock->l_req_mode = LCK_PW;
1406 /* Give the client a lock on the child object, instead of the
1407 * parent that it requested. */
1408 new_resid[0] = NTOH__u32(mds_rep->ino);
1409 if (new_resid[0] == 0)
1411 old_res = lock->l_resource->lr_name[0];
1413 ldlm_lock_change_resource(lock, new_resid);
1414 if (lock->l_resource == NULL) {
1418 LDLM_DEBUG(lock, "intent policy, old res %ld",
1420 RETURN(ELDLM_LOCK_CHANGED);
1422 int size = sizeof(struct ldlm_reply);
1423 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
1434 extern int mds_iocontrol(long cmd, struct lustre_handle *conn,
1435 int len, void *karg, void *uarg);
1437 /* use obd ops to offer management infrastructure */
1438 static struct obd_ops mds_obd_ops = {
1439 o_connect: mds_connect,
1440 o_disconnect: mds_disconnect,
1442 o_cleanup: mds_cleanup,
1443 o_iocontrol: mds_iocontrol
1446 static int __init mds_init(void)
1448 mds_file_cache = kmem_cache_create("ll_mds_file_data",
1449 sizeof(struct mds_file_data),
1451 if (mds_file_cache == NULL)
1454 class_register_type(&mds_obd_ops, LUSTRE_MDS_NAME);
1455 ldlm_register_intent(ldlm_intent_policy);
1459 static void __exit mds_exit(void)
1461 ldlm_unregister_intent();
1462 class_unregister_type(LUSTRE_MDS_NAME);
1463 if (kmem_cache_destroy(mds_file_cache))
1464 CERROR("couldn't free MDS file cache\n");
1467 MODULE_AUTHOR("Cluster File Systems <info@clusterfs.com>");
1468 MODULE_DESCRIPTION("Lustre Metadata Server (MDS) v0.01");
1469 MODULE_LICENSE("GPL");
1471 module_init(mds_init);
1472 module_exit(mds_exit);