4 * Lustre Metadata Server (mds) request handler
6 * Copyright (C) 2001 Cluster File Systems, Inc.
8 * This code is issued under the GNU General Public License.
9 * See the file COPYING in this distribution
11 * by Peter Braam <braam@clusterfs.com>
13 * This server is single threaded at present (but can easily be multi threaded).
20 #include <linux/version.h>
21 #include <linux/module.h>
23 #include <linux/stat.h>
24 #include <linux/locks.h>
25 #include <linux/ext2_fs.h>
26 #include <linux/quotaops.h>
27 #include <asm/unistd.h>
28 #include <linux/obd_support.h>
29 #include <linux/obd.h>
30 #include <linux/lustre_lib.h>
31 #include <linux/lustre_idl.h>
32 #include <linux/lustre_mds.h>
33 #include <linux/obd_class.h>
36 static struct mds_obd *MDS;
39 static int mds_queue_req(struct mds_request *req)
47 list_add(&req->rq_list, &MDS->mds_reqs);
48 init_waitqueue_head(&req->rq_wait_for_mds_rep);
50 wake_up(&MDS->mds_waitq);
51 printk("-- sleeping\n");
52 interruptible_sleep_on(&req->rq_wait_for_mds_rep);
57 static struct dentry *mds_fid2dentry(struct mds_obd *mds, struct lustre_fid *fid)
62 inode = iget(mds->mds_sb, fid->id);
67 de = d_alloc_root(inode);
78 int mds_getattr(struct mds_request *req)
80 struct dentry *de = mds_fid2dentry(req->rq_obd, &req->rq_req->fid1);
85 rc = mds_pack_rep(NULL, 0, NULL, 0, &req->rq_rephdr, &req->rq_rep,
86 &req->rq_replen, &req->rq_repbuf);
89 printk("mds: out of memory\n");
90 req->rq_status = -ENOMEM;
94 req->rq_rephdr->seqno = req->rq_reqhdr->seqno;
99 req->rq_rephdr->status = -ENOENT;
104 rep->atime = inode->i_atime;
105 rep->ctime = inode->i_ctime;
106 rep->mtime = inode->i_mtime;
107 rep->uid = inode->i_uid;
108 rep->gid = inode->i_gid;
109 rep->size = inode->i_size;
110 rep->mode = inode->i_mode;
116 int mds_reply(struct mds_request *req)
119 kfree(req->rq_reqbuf);
120 req->rq_reqbuf = NULL;
121 wake_up_interruptible(&req->rq_wait_for_mds_rep);
126 int mds_error(struct mds_request *req)
128 struct mds_rep_hdr *hdr;
131 hdr = kmalloc(sizeof(*hdr), GFP_KERNEL);
137 memset(hdr, 0, sizeof(*hdr));
139 hdr->seqno = req->rq_reqhdr->seqno;
140 hdr->status = req->rq_status;
141 hdr->type = MDS_TYPE_ERR;
142 req->rq_repbuf = (char *)hdr;
145 return mds_reply(req);
148 //int mds_handle(struct mds_conn *conn, int len, char *buf)
149 int mds_handle(struct mds_request *req)
152 struct mds_req_hdr *hdr;
156 hdr = (struct mds_req_hdr *)req->rq_reqbuf;
158 if (NTOH__u32(hdr->type) != MDS_TYPE_REQ) {
159 printk("lustre_mds: wrong packet type sent %d\n",
160 NTOH__u32(hdr->type));
165 rc = mds_unpack_req(req->rq_reqbuf, req->rq_reqlen,
166 &req->rq_reqhdr, &req->rq_req);
168 printk("lustre_mds: Invalid request\n");
173 switch (req->rq_reqhdr->opc) {
176 CDEBUG(D_INODE, "getattr\n");
177 rc = mds_getattr(req);
181 return mds_getattr(req);
184 return mds_getattr(req);
187 return mds_getattr(req);
190 return mds_getattr(req);
193 return mds_getattr(req);
196 return mds_getattr(req);
199 return mds_getattr(req);
202 return mds_getattr(req);
205 return mds_getattr(req);
208 return mds_getattr(req);
211 return mds_error(req);
216 printk("mds: processing error %d\n", rc);
219 CDEBUG(D_INODE, "sending reply\n");
227 static void mds_timer_run(unsigned long __data)
229 struct task_struct * p = (struct task_struct *) __data;
234 int mds_main(void *arg)
236 struct mds_obd *mds = (struct mds_obd *) arg;
237 struct timer_list timer;
241 spin_lock_irq(¤t->sigmask_lock);
242 sigfillset(¤t->blocked);
243 recalc_sigpending(current);
244 spin_unlock_irq(¤t->sigmask_lock);
246 sprintf(current->comm, "lustre_mds");
248 /* Set up an interval timer which can be used to trigger a
249 wakeup after the interval expires */
251 timer.data = (unsigned long) current;
252 timer.function = mds_timer_run;
253 mds->mds_timer = &timer;
255 /* Record that the thread is running */
256 mds->mds_thread = current;
257 wake_up(&mds->mds_done_waitq);
259 printk(KERN_INFO "lustre_mds starting. Commit interval %d seconds\n",
260 mds->mds_interval / HZ);
262 /* XXX maintain a list of all managed devices: insert here */
264 /* And now, wait forever for commit wakeup events. */
266 struct mds_request *request;
269 if (mds->mds_flags & MDS_UNMOUNT)
273 wake_up(&mds->mds_done_waitq);
274 interruptible_sleep_on(&mds->mds_waitq);
276 CDEBUG(D_INODE, "lustre_mds wakes\n");
277 CDEBUG(D_INODE, "pick up req here and continue\n");
279 if (list_empty(&mds->mds_reqs)) {
280 CDEBUG(D_INODE, "woke because of timer\n");
282 request = list_entry(mds->mds_reqs.next,
283 struct mds_request, rq_list);
284 list_del(&request->rq_list);
285 rc = mds_handle(request);
289 del_timer_sync(mds->mds_timer);
291 /* XXX maintain a list of all managed devices: cleanup here */
293 mds->mds_thread = NULL;
294 wake_up(&mds->mds_done_waitq);
295 printk("lustre_mds: exiting\n");
299 static void mds_stop_srv_thread(struct mds_obd *mds)
301 mds->mds_flags |= MDS_UNMOUNT;
303 while (mds->mds_thread) {
304 wake_up(&mds->mds_waitq);
305 sleep_on(&mds->mds_done_waitq);
309 static void mds_start_srv_thread(struct mds_obd *mds)
311 init_waitqueue_head(&mds->mds_waitq);
312 init_waitqueue_head(&mds->mds_done_waitq);
313 kernel_thread(mds_main, (void *)mds,
314 CLONE_VM | CLONE_FS | CLONE_FILES);
315 while (!mds->mds_thread)
316 sleep_on(&mds->mds_done_waitq);
319 /* mount the file system (secretly) */
320 static int mds_setup(struct obd_device *obddev, obd_count len,
324 struct obd_ioctl_data* data = buf;
325 struct mds_obd *mds = &obddev->u.mds;
326 struct vfsmount *mnt;
330 mnt = do_kern_mount(data->ioc_inlbuf2, 0,
331 data->ioc_inlbuf1, NULL);
338 mds->mds_sb = mnt->mnt_root->d_inode->i_sb;
339 if (!obddev->u.mds.mds_sb) {
344 INIT_LIST_HEAD(&mds->mds_reqs);
345 mds->mds_thread = NULL;
347 mds->mds_interval = 3 * HZ;
348 mds->mds_vfsmnt = mnt;
349 obddev->u.mds.mds_fstype = strdup(data->ioc_inlbuf2);
351 mds->mds_ctxt.pwdmnt = mnt;
352 mds->mds_ctxt.pwd = mnt->mnt_root;
353 mds->mds_ctxt.fs = KERNEL_DS;
356 spin_lock_init(&obddev->u.mds.fo_lock);
358 mds_start_srv_thread(mds);
365 static int mds_cleanup(struct obd_device * obddev)
367 struct super_block *sb;
368 struct mds_obd *mds = &obddev->u.mds;
372 if ( !(obddev->obd_flags & OBD_SET_UP) ) {
377 if ( !list_empty(&obddev->obd_gen_clients) ) {
378 printk(KERN_WARNING __FUNCTION__ ": still has clients!\n");
384 mds_stop_srv_thread(mds);
391 if (!list_empty(&mds->mds_reqs)) {
392 // XXX reply with errors and clean up
393 CDEBUG(D_INODE, "Request list not empty!\n");
397 mntput(mds->mds_vfsmnt);
399 kfree(mds->mds_fstype);
408 /* use obd ops to offer management infrastructure */
409 static struct obd_ops mds_obd_ops = {
411 o_cleanup: mds_cleanup,
414 static int __init mds_init(void)
416 obd_register_type(&mds_obd_ops, LUSTRE_MDS_NAME);
420 static void __exit mds_exit(void)
422 obd_unregister_type(LUSTRE_MDS_NAME);
425 MODULE_AUTHOR("Peter J. Braam <braam@clusterfs.com>");
426 MODULE_DESCRIPTION("Lustre Metadata Server (MDS) v0.01");
427 MODULE_LICENSE("GPL");
430 // for testing (maybe this stays)
431 EXPORT_SYMBOL(mds_queue_req);
433 module_init(mds_init);
434 module_exit(mds_exit);