4 * Lustre Metadata Server (mds) request handler
6 * Copyright (C) 2001, 2002 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 <asm/uaccess.h>
29 #include <linux/obd_support.h>
30 #include <linux/obd.h>
31 #include <linux/lustre_lib.h>
32 #include <linux/lustre_idl.h>
33 #include <linux/lustre_mds.h>
34 #include <linux/lustre_net.h>
35 #include <linux/obd_class.h>
38 static struct mds_obd *MDS;
40 // XXX make this networked!
41 static int mds_queue_req(struct ptlrpc_request *req)
43 struct ptlrpc_request *srv_req;
50 srv_req = kmalloc(sizeof(*srv_req), GFP_KERNEL);
56 printk("---> MDS at %d %p, incoming req %p, srv_req %p\n",
57 __LINE__, MDS, req, srv_req);
59 memset(srv_req, 0, sizeof(*req));
61 /* move the request buffer */
62 srv_req->rq_reqbuf = req->rq_reqbuf;
63 srv_req->rq_reqlen = req->rq_reqlen;
64 srv_req->rq_obd = MDS;
66 /* remember where it came from */
67 srv_req->rq_reply_handle = req;
69 list_add(&srv_req->rq_list, &MDS->mds_reqs);
70 wake_up(&MDS->mds_waitq);
74 int mds_sendpage(struct ptlrpc_request *req, struct file *file,
75 __u64 offset, struct niobuf *dst)
78 mm_segment_t oldfs = get_fs();
80 if (req->rq_peer.peer_nid == 0) {
81 /* dst->addr is a user address, but in a different task! */
83 rc = generic_file_read(file, (char *)(long)dst->addr,
92 buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
98 rc = generic_file_read(file, buf, PAGE_SIZE, &offset);
104 req->rq_bulkbuf = buf;
105 req->rq_bulklen = PAGE_SIZE;
106 rc = ptl_send_buf(req, &req->rq_peer, MDS_BULK_PORTAL, 0);
107 init_waitqueue_head(&req->rq_wait_for_bulk);
108 sleep_on(&req->rq_wait_for_bulk);
110 req->rq_bulklen = 0; /* FIXME: eek. */
116 int mds_reply(struct ptlrpc_request *req)
118 struct ptlrpc_request *clnt_req = req->rq_reply_handle;
122 if (req->rq_obd->mds_service != NULL) {
123 /* This is a request that came from the network via portals. */
125 /* FIXME: we need to increment the count of handled events */
126 ptl_send_buf(req, &req->rq_peer, MDS_REPLY_PORTAL, 0);
128 /* This is a local request that came from another thread. */
130 /* move the reply to the client */
131 clnt_req->rq_replen = req->rq_replen;
132 clnt_req->rq_repbuf = req->rq_repbuf;
133 req->rq_repbuf = NULL;
136 /* free the request buffer */
137 kfree(req->rq_reqbuf);
138 req->rq_reqbuf = NULL;
140 /* wake up the client */
141 wake_up_interruptible(&clnt_req->rq_wait_for_rep);
148 int mds_error(struct ptlrpc_request *req)
150 struct ptlrep_hdr *hdr;
154 hdr = kmalloc(sizeof(*hdr), GFP_KERNEL);
160 memset(hdr, 0, sizeof(*hdr));
162 hdr->seqno = req->rq_reqhdr->seqno;
163 hdr->status = req->rq_status;
164 hdr->type = MDS_TYPE_ERR;
166 req->rq_repbuf = (char *)hdr;
167 req->rq_replen = sizeof(*hdr);
170 return mds_reply(req);
173 struct dentry *mds_fid2dentry(struct mds_obd *mds, struct ll_fid *fid, struct vfsmount **mnt)
175 /* stolen from NFS */
176 struct super_block *sb = mds->mds_sb;
177 unsigned long ino = fid->id;
178 //__u32 generation = fid->generation;
179 __u32 generation = 0;
181 struct list_head *lp;
182 struct dentry *result;
185 *mnt = mntget(mds->mds_vfsmnt);
189 return ERR_PTR(-ESTALE);
191 inode = iget(sb, ino);
193 return ERR_PTR(-ENOMEM);
195 printk("--> mds_fid2dentry: sb %p\n", inode->i_sb);
197 if (is_bad_inode(inode)
198 || (generation && inode->i_generation != generation)
200 /* we didn't find the right inode.. */
202 "bad inode %lu, link: %d ct: %d or version %u/%u\n",
204 inode->i_nlink, atomic_read(&inode->i_count),
208 return ERR_PTR(-ESTALE);
211 /* now to find a dentry.
212 * If possible, get a well-connected one
214 spin_lock(&dcache_lock);
215 for (lp = inode->i_dentry.next; lp != &inode->i_dentry ; lp=lp->next) {
216 result = list_entry(lp,struct dentry, d_alias);
217 if (! (result->d_flags & DCACHE_NFSD_DISCONNECTED)) {
219 result->d_vfs_flags |= DCACHE_REFERENCED;
220 spin_unlock(&dcache_lock);
225 spin_unlock(&dcache_lock);
226 result = d_alloc_root(inode);
227 if (result == NULL) {
229 return ERR_PTR(-ENOMEM);
231 result->d_flags |= DCACHE_NFSD_DISCONNECTED;
235 static inline void mds_get_objid(struct inode *inode, __u64 *id)
237 memcpy(id, &inode->u.ext2_i.i_data, sizeof(*id));
240 int mds_getattr(struct ptlrpc_request *req)
242 struct dentry *de = mds_fid2dentry(req->rq_obd, &req->rq_req.mds->fid1,
248 rc = mds_pack_rep(NULL, 0, NULL, 0, &req->rq_rephdr, &req->rq_rep.mds,
249 &req->rq_replen, &req->rq_repbuf);
252 printk("mds: out of memory\n");
253 req->rq_status = -ENOMEM;
257 req->rq_rephdr->seqno = req->rq_reqhdr->seqno;
258 rep = req->rq_rep.mds;
262 req->rq_rephdr->status = -ENOENT;
267 rep->ino = inode->i_ino;
268 rep->atime = inode->i_atime;
269 rep->ctime = inode->i_ctime;
270 rep->mtime = inode->i_mtime;
271 rep->uid = inode->i_uid;
272 rep->gid = inode->i_gid;
273 rep->size = inode->i_size;
274 rep->mode = inode->i_mode;
275 rep->nlink = inode->i_nlink;
277 mds_get_objid(inode, &rep->objid);
282 int mds_readpage(struct ptlrpc_request *req)
284 struct vfsmount *mnt;
285 struct dentry *de = mds_fid2dentry(req->rq_obd, &req->rq_req.mds->fid1,
288 struct niobuf *niobuf;
292 printk("mds_readpage: ino %ld\n", de->d_inode->i_ino);
293 rc = mds_pack_rep(NULL, 0, NULL, 0, &req->rq_rephdr, &req->rq_rep.mds,
294 &req->rq_replen, &req->rq_repbuf);
297 printk("mds: out of memory\n");
298 req->rq_status = -ENOMEM;
302 req->rq_rephdr->seqno = req->rq_reqhdr->seqno;
303 rep = req->rq_rep.mds;
307 req->rq_rephdr->status = PTR_ERR(de);
311 file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
312 /* note: in case of an error, dentry_open puts dentry */
315 req->rq_rephdr->status = PTR_ERR(file);
319 niobuf = mds_req_tgt(req->rq_req.mds);
321 /* to make this asynchronous make sure that the handling function
322 doesn't send a reply when this function completes. Instead a
323 callback function would send the reply */
324 rc = mds_sendpage(req, file, req->rq_req.mds->size, niobuf);
327 req->rq_rephdr->status = rc;
332 int mds_reint(struct ptlrpc_request *req)
335 char *buf = mds_req_tgt(req->rq_req.mds);
336 int len = req->rq_req.mds->tgtlen;
337 struct mds_update_record rec;
339 rc = mds_update_unpack(buf, len, &rec);
341 printk(__FUNCTION__ ": invalid record\n");
342 req->rq_status = -EINVAL;
345 /* rc will be used to interrupt a for loop over multiple records */
346 rc = mds_reint_rec(&rec, req);
350 //int mds_handle(struct mds_conn *conn, int len, char *buf)
351 int mds_handle(struct ptlrpc_request *req)
354 struct ptlreq_hdr *hdr;
358 hdr = (struct ptlreq_hdr *)req->rq_reqbuf;
360 if (NTOH__u32(hdr->type) != MDS_TYPE_REQ) {
361 printk("lustre_mds: wrong packet type sent %d\n",
362 NTOH__u32(hdr->type));
367 rc = mds_unpack_req(req->rq_reqbuf, req->rq_reqlen,
368 &req->rq_reqhdr, &req->rq_req.mds);
370 printk("lustre_mds: Invalid request\n");
375 switch (req->rq_reqhdr->opc) {
378 CDEBUG(D_INODE, "getattr\n");
379 rc = mds_getattr(req);
383 CDEBUG(D_INODE, "readpage\n");
384 rc = mds_readpage(req);
388 CDEBUG(D_INODE, "reint\n");
393 return mds_error(req);
398 printk(__FUNCTION__ ": no header\n");
402 if( req->rq_status) {
405 CDEBUG(D_INODE, "sending reply\n");
413 static void mds_timer_run(unsigned long __data)
415 struct task_struct * p = (struct task_struct *) __data;
420 int mds_main(void *arg)
422 struct mds_obd *mds = (struct mds_obd *) arg;
423 struct timer_list timer;
427 spin_lock_irq(¤t->sigmask_lock);
428 sigfillset(¤t->blocked);
429 recalc_sigpending(current);
430 spin_unlock_irq(¤t->sigmask_lock);
432 sprintf(current->comm, "lustre_mds");
434 /* Set up an interval timer which can be used to trigger a
435 wakeup after the interval expires */
437 timer.data = (unsigned long) current;
438 timer.function = mds_timer_run;
439 mds->mds_timer = &timer;
441 /* Record that the thread is running */
442 mds->mds_thread = current;
443 wake_up(&mds->mds_done_waitq);
445 printk(KERN_INFO "lustre_mds starting. Commit interval %d seconds\n",
446 mds->mds_interval / HZ);
448 /* XXX maintain a list of all managed devices: insert here */
450 /* And now, wait forever for commit wakeup events. */
454 if (mds->mds_flags & MDS_UNMOUNT)
457 wake_up(&mds->mds_done_waitq);
458 interruptible_sleep_on(&mds->mds_waitq);
460 CDEBUG(D_INODE, "lustre_mds wakes\n");
461 CDEBUG(D_INODE, "pick up req here and continue\n");
463 if (mds->mds_service != NULL) {
467 struct ptlrpc_request request;
469 rc = PtlEQGet(mds->mds_service->srv_eq_h, &ev);
470 if (rc != PTL_OK && rc != PTL_EQ_DROPPED)
472 /* FIXME: If we move to an event-driven model,
473 * we should put the request on the stack of
474 * mds_handle instead. */
475 memset(&request, 0, sizeof(request));
476 request.rq_reqbuf = ev.mem_desc.start +
478 request.rq_reqlen = ev.mem_desc.length;
479 request.rq_obd = MDS;
480 request.rq_xid = ev.match_bits;
482 request.rq_peer.peer_nid = ev.initiator.nid;
483 /* FIXME: this NI should be the incoming NI.
484 * We don't know how to find that from here. */
485 request.rq_peer.peer_ni =
486 mds->mds_service->srv_self.peer_ni;
487 rc = mds_handle(&request);
490 struct ptlrpc_request *request;
492 if (list_empty(&mds->mds_reqs)) {
493 CDEBUG(D_INODE, "woke because of timer\n");
495 request = list_entry(mds->mds_reqs.next,
496 struct ptlrpc_request,
498 list_del(&request->rq_list);
499 rc = mds_handle(request);
504 del_timer_sync(mds->mds_timer);
506 /* XXX maintain a list of all managed devices: cleanup here */
508 mds->mds_thread = NULL;
509 wake_up(&mds->mds_done_waitq);
510 printk("lustre_mds: exiting\n");
514 static void mds_stop_srv_thread(struct mds_obd *mds)
516 mds->mds_flags |= MDS_UNMOUNT;
518 while (mds->mds_thread) {
519 wake_up(&mds->mds_waitq);
520 sleep_on(&mds->mds_done_waitq);
524 static void mds_start_srv_thread(struct mds_obd *mds)
526 init_waitqueue_head(&mds->mds_waitq);
527 init_waitqueue_head(&mds->mds_done_waitq);
528 kernel_thread(mds_main, (void *)mds,
529 CLONE_VM | CLONE_FS | CLONE_FILES);
530 while (!mds->mds_thread)
531 sleep_on(&mds->mds_done_waitq);
534 /* mount the file system (secretly) */
535 static int mds_setup(struct obd_device *obddev, obd_count len,
539 struct obd_ioctl_data* data = buf;
540 struct mds_obd *mds = &obddev->u.mds;
541 struct vfsmount *mnt;
542 struct lustre_peer peer;
547 mnt = do_kern_mount(data->ioc_inlbuf2, 0, data->ioc_inlbuf1, NULL);
554 mds->mds_sb = mnt->mnt_root->d_inode->i_sb;
555 if (!obddev->u.mds.mds_sb) {
560 mds->mds_vfsmnt = mnt;
561 obddev->u.mds.mds_fstype = strdup(data->ioc_inlbuf2);
563 mds->mds_ctxt.pwdmnt = mnt;
564 mds->mds_ctxt.pwd = mnt->mnt_root;
565 mds->mds_ctxt.fs = KERNEL_DS;
566 mds->mds_remote_nid = 0;
568 INIT_LIST_HEAD(&mds->mds_reqs);
569 mds->mds_thread = NULL;
571 mds->mds_interval = 3 * HZ;
574 spin_lock_init(&obddev->u.mds.mds_lock);
576 err = kportal_uuid_to_peer("self", &peer);
578 mds->mds_service = kmalloc(sizeof(*mds->mds_service),
580 if (mds->mds_service == NULL)
582 mds->mds_service->srv_buf_size = 64 * 1024;
583 mds->mds_service->srv_portal = MDS_REQUEST_PORTAL;
584 memcpy(&mds->mds_service->srv_self, &peer, sizeof(peer));
585 mds->mds_service->srv_wait_queue = &mds->mds_waitq;
587 rpc_register_service(mds->mds_service, "self");
590 mds_start_srv_thread(mds);
597 static int mds_cleanup(struct obd_device * obddev)
599 struct super_block *sb;
600 struct mds_obd *mds = &obddev->u.mds;
604 if ( !(obddev->obd_flags & OBD_SET_UP) ) {
609 if ( !list_empty(&obddev->obd_gen_clients) ) {
610 printk(KERN_WARNING __FUNCTION__ ": still has clients!\n");
616 mds_stop_srv_thread(mds);
623 if (!list_empty(&mds->mds_reqs)) {
624 // XXX reply with errors and clean up
625 CDEBUG(D_INODE, "Request list not empty!\n");
629 mntput(mds->mds_vfsmnt);
631 kfree(mds->mds_fstype);
639 /* use obd ops to offer management infrastructure */
640 static struct obd_ops mds_obd_ops = {
642 o_cleanup: mds_cleanup,
645 static int __init mds_init(void)
647 obd_register_type(&mds_obd_ops, LUSTRE_MDS_NAME);
651 static void __exit mds_exit(void)
653 obd_unregister_type(LUSTRE_MDS_NAME);
656 MODULE_AUTHOR("Peter J. Braam <braam@clusterfs.com>");
657 MODULE_DESCRIPTION("Lustre Metadata Server (MDS) v0.01");
658 MODULE_LICENSE("GPL");
661 // for testing (maybe this stays)
662 EXPORT_SYMBOL(mds_queue_req);
664 module_init(mds_init);
665 module_exit(mds_exit);