1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * Lustre Metadata Server (mds) request handler
8 * Copyright (C) 2001, 2002 Cluster File Systems, Inc.
10 * This code is issued under the GNU General Public License.
11 * See the file COPYING in this distribution
13 * by Peter Braam <braam@clusterfs.com>
15 * This server is single threaded at present (but can easily be multi threaded).
21 #include <linux/version.h>
22 #include <linux/module.h>
24 #include <linux/stat.h>
25 #include <linux/locks.h>
26 #include <linux/ext2_fs.h>
27 #include <linux/quotaops.h>
28 #include <asm/unistd.h>
29 #include <asm/uaccess.h>
31 #define DEBUG_SUBSYSTEM S_MDS
33 #include <linux/lustre_mds.h>
34 #include <linux/lustre_lib.h>
36 int mds_sendpage(struct ptlrpc_request *req, struct file *file,
37 __u64 offset, struct niobuf *dst)
40 mm_segment_t oldfs = get_fs();
42 if (req->rq_peer.peer_nid == 0) {
43 /* dst->addr is a user address, but in a different task! */
45 /* FIXME: Can't use file->f_op->read() on a directory */
46 rc = generic_file_read(file, (char *)(long)dst->addr,
50 if (rc != PAGE_SIZE) {
57 struct ptlrpc_bulk_desc *bulk;
60 bulk = ptlrpc_prep_bulk(&req->rq_peer);
67 bulk->b_xid = req->rq_xid;
69 OBD_ALLOC(buf, PAGE_SIZE);
77 /* FIXME: Can't use file->f_op->read() on a directory */
78 //rc = file->f_op->read(file, buf, PAGE_SIZE, &offset);
79 rc = generic_file_read(file, buf, PAGE_SIZE, &offset);
82 if (rc != PAGE_SIZE) {
89 bulk->b_buflen = PAGE_SIZE;
91 rc = ptlrpc_send_bulk(bulk, MDS_BULK_PORTAL);
92 wait_event_interruptible(bulk->b_waitq,
93 ptlrpc_check_bulk_sent(bulk));
95 if (bulk->b_flags == PTL_RPC_INTR) {
103 OBD_FREE(buf, PAGE_SIZE);
105 OBD_FREE(bulk, sizeof(*bulk));
111 struct dentry *mds_fid2dentry(struct mds_obd *mds, struct ll_fid *fid,
112 struct vfsmount **mnt)
114 /* stolen from NFS */
115 struct super_block *sb = mds->mds_sb;
116 unsigned long ino = fid->id;
117 //__u32 generation = fid->generation;
118 __u32 generation = 0;
120 struct list_head *lp;
121 struct dentry *result;
124 return ERR_PTR(-ESTALE);
126 inode = iget(sb, ino);
128 return ERR_PTR(-ENOMEM);
130 CDEBUG(D_DENTRY, "--> mds_fid2dentry: sb %p\n", inode->i_sb);
132 if (is_bad_inode(inode)
133 || (generation && inode->i_generation != generation)
135 /* we didn't find the right inode.. */
136 CERROR("bad inode %lu, link: %d ct: %d or version %u/%u\n",
138 inode->i_nlink, atomic_read(&inode->i_count),
143 return ERR_PTR(-ESTALE);
146 /* now to find a dentry.
147 * If possible, get a well-connected one
150 *mnt = mds->mds_vfsmnt;
151 spin_lock(&dcache_lock);
152 for (lp = inode->i_dentry.next; lp != &inode->i_dentry ; lp=lp->next) {
153 result = list_entry(lp,struct dentry, d_alias);
154 if (! (result->d_flags & DCACHE_NFSD_DISCONNECTED)) {
156 result->d_vfs_flags |= DCACHE_REFERENCED;
157 spin_unlock(&dcache_lock);
164 spin_unlock(&dcache_lock);
165 result = d_alloc_root(inode);
166 if (result == NULL) {
168 return ERR_PTR(-ENOMEM);
172 result->d_flags |= DCACHE_NFSD_DISCONNECTED;
176 static inline void mds_get_objid(struct inode *inode, __u64 *id)
178 /* FIXME: it is only by luck that this works on ext3 */
179 memcpy(id, &inode->u.ext2_i.i_data, sizeof(*id));
182 int mds_getattr(struct ptlrpc_request *req)
189 rc = mds_pack_rep(NULL, 0, NULL, 0, &req->rq_rephdr, &req->rq_rep,
190 &req->rq_replen, &req->rq_repbuf);
193 CERROR("mds: out of memory\n");
194 req->rq_status = -ENOMEM;
198 req->rq_rephdr->xid = req->rq_reqhdr->xid;
199 rep = req->rq_rep.mds;
201 de = mds_fid2dentry(&req->rq_obd->u.mds, &req->rq_req.mds->fid1, NULL);
204 req->rq_rephdr->status = -ENOENT;
209 rep->ino = inode->i_ino;
210 rep->atime = inode->i_atime;
211 rep->ctime = inode->i_ctime;
212 rep->mtime = inode->i_mtime;
213 rep->uid = inode->i_uid;
214 rep->gid = inode->i_gid;
215 rep->size = inode->i_size;
216 rep->mode = inode->i_mode;
217 rep->nlink = inode->i_nlink;
219 mds_get_objid(inode, &rep->objid);
224 int mds_open(struct ptlrpc_request *req)
229 struct vfsmount *mnt;
233 rc = mds_pack_rep(NULL, 0, NULL, 0, &req->rq_rephdr, &req->rq_rep,
234 &req->rq_replen, &req->rq_repbuf);
237 CERROR("mds: out of memory\n");
238 req->rq_status = -ENOMEM;
242 req->rq_rephdr->xid = req->rq_reqhdr->xid;
243 rep = req->rq_rep.mds;
245 de = mds_fid2dentry(&req->rq_obd->u.mds, &req->rq_req.mds->fid1, &mnt);
248 req->rq_rephdr->status = -ENOENT;
251 flags = req->rq_req.mds->flags;
252 file = dentry_open(de, mnt, flags);
253 if (!file || IS_ERR(file)) {
254 req->rq_rephdr->status = -EINVAL;
258 rep->objid = (__u64) (unsigned long)file;
262 int mds_close(struct ptlrpc_request *req)
267 struct vfsmount *mnt;
270 rc = mds_pack_rep(NULL, 0, NULL, 0, &req->rq_rephdr, &req->rq_rep,
271 &req->rq_replen, &req->rq_repbuf);
274 CERROR("mds: out of memory\n");
275 req->rq_status = -ENOMEM;
279 req->rq_rephdr->xid = req->rq_reqhdr->xid;
280 rep = req->rq_rep.mds;
282 de = mds_fid2dentry(&req->rq_obd->u.mds, &req->rq_req.mds->fid1, &mnt);
285 req->rq_rephdr->status = -ENOENT;
289 file = (struct file *)(unsigned long) req->rq_req.mds->objid;
291 req->rq_rephdr->status = filp_close(file, 0);
297 int mds_readpage(struct ptlrpc_request *req)
299 struct vfsmount *mnt;
302 struct niobuf *niobuf;
306 rc = mds_pack_rep(NULL, 0, NULL, 0, &req->rq_rephdr, &req->rq_rep,
307 &req->rq_replen, &req->rq_repbuf);
310 CERROR("mds: out of memory\n");
311 req->rq_status = -ENOMEM;
315 req->rq_rephdr->xid = req->rq_reqhdr->xid;
316 rep = req->rq_rep.mds;
318 de = mds_fid2dentry(&req->rq_obd->u.mds, &req->rq_req.mds->fid1, &mnt);
321 req->rq_rephdr->status = PTR_ERR(de);
325 CDEBUG(D_INODE, "ino %ld\n", de->d_inode->i_ino);
327 file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
328 /* note: in case of an error, dentry_open puts dentry */
331 req->rq_rephdr->status = PTR_ERR(file);
335 niobuf = mds_req_tgt(req->rq_req.mds);
337 /* to make this asynchronous make sure that the handling function
338 doesn't send a reply when this function completes. Instead a
339 callback function would send the reply */
340 rc = mds_sendpage(req, file, req->rq_req.mds->size, niobuf);
343 req->rq_rephdr->status = rc;
348 int mds_reint(struct ptlrpc_request *req)
351 char *buf = mds_req_tgt(req->rq_req.mds);
352 int len = req->rq_req.mds->tgtlen;
353 struct mds_update_record rec;
355 rc = mds_update_unpack(buf, len, &rec);
357 CERROR("invalid record\n");
358 req->rq_status = -EINVAL;
361 /* rc will be used to interrupt a for loop over multiple records */
362 rc = mds_reint_rec(&rec, req);
366 int mds_handle(struct obd_device *dev, struct ptlrpc_service *svc,
367 struct ptlrpc_request *req)
370 struct ptlreq_hdr *hdr;
374 hdr = (struct ptlreq_hdr *)req->rq_reqbuf;
376 if (NTOH__u32(hdr->type) != MDS_TYPE_REQ) {
377 CERROR("lustre_mds: wrong packet type sent %d\n",
378 NTOH__u32(hdr->type));
383 rc = mds_unpack_req(req->rq_reqbuf, req->rq_reqlen,
384 &req->rq_reqhdr, &req->rq_req);
386 CERROR("lustre_mds: Invalid request\n");
391 switch (req->rq_reqhdr->opc) {
394 CDEBUG(D_INODE, "getattr\n");
395 rc = mds_getattr(req);
399 CDEBUG(D_INODE, "readpage\n");
400 rc = mds_readpage(req);
404 CDEBUG(D_INODE, "reint\n");
409 CDEBUG(D_INODE, "open\n");
414 CDEBUG(D_INODE, "close\n");
419 return ptlrpc_error(dev, svc, req);
424 CERROR("no header\n");
428 if( req->rq_status) {
429 ptlrpc_error(dev, svc, req);
431 CDEBUG(D_INODE, "sending reply\n");
432 ptlrpc_reply(dev, svc, req);
439 /* mount the file system (secretly) */
440 static int mds_setup(struct obd_device *obddev, obd_count len, void *buf)
442 struct obd_ioctl_data* data = buf;
443 struct mds_obd *mds = &obddev->u.mds;
444 struct vfsmount *mnt;
448 mnt = do_kern_mount(data->ioc_inlbuf2, 0, data->ioc_inlbuf1, NULL);
451 CERROR("do_kern_mount failed: %d\n", err);
456 mds->mds_sb = mnt->mnt_root->d_inode->i_sb;
462 mds->mds_vfsmnt = mnt;
463 mds->mds_fstype = strdup(data->ioc_inlbuf2);
465 mds->mds_ctxt.pwdmnt = mnt;
466 mds->mds_ctxt.pwd = mnt->mnt_root;
467 mds->mds_ctxt.fs = KERNEL_DS;
469 mds->mds_service = ptlrpc_init_svc( 64 * 1024,
477 rpc_register_service(mds->mds_service, "self");
479 err = ptlrpc_start_thread(obddev, mds->mds_service, "lustre_mds");
481 CERROR("cannot start thread\n");
489 static int mds_cleanup(struct obd_device * obddev)
491 struct super_block *sb;
492 struct mds_obd *mds = &obddev->u.mds;
496 if ( !(obddev->obd_flags & OBD_SET_UP) ) {
501 if ( !list_empty(&obddev->obd_gen_clients) ) {
502 CERROR("still has clients!\n");
507 ptlrpc_stop_thread(mds->mds_service);
508 rpc_unregister_service(mds->mds_service);
510 if (!list_empty(&mds->mds_service->srv_reqs)) {
511 // XXX reply with errors and clean up
512 CERROR("Request list not empty!\n");
515 rpc_unregister_service(mds->mds_service);
516 OBD_FREE(mds->mds_service, sizeof(*mds->mds_service));
525 mntput(mds->mds_vfsmnt);
527 kfree(mds->mds_fstype);
535 /* use obd ops to offer management infrastructure */
536 static struct obd_ops mds_obd_ops = {
538 o_cleanup: mds_cleanup,
541 static int __init mds_init(void)
543 obd_register_type(&mds_obd_ops, LUSTRE_MDS_NAME);
547 static void __exit mds_exit(void)
549 obd_unregister_type(LUSTRE_MDS_NAME);
552 MODULE_AUTHOR("Peter J. Braam <braam@clusterfs.com>");
553 MODULE_DESCRIPTION("Lustre Metadata Server (MDS) v0.01");
554 MODULE_LICENSE("GPL");
556 module_init(mds_init);
557 module_exit(mds_exit);