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).
19 #include <linux/version.h>
20 #include <linux/module.h>
22 #include <linux/stat.h>
23 #include <linux/locks.h>
24 #include <linux/ext2_fs.h>
25 #include <linux/quotaops.h>
26 #include <asm/unistd.h>
27 #include <asm/uaccess.h>
29 #define DEBUG_SUBSYSTEM S_MDS
31 #include <linux/obd_support.h>
32 #include <linux/obd.h>
33 #include <linux/lustre_lib.h>
34 #include <linux/lustre_idl.h>
35 #include <linux/lustre_mds.h>
36 #include <linux/lustre_net.h>
37 #include <linux/obd_class.h>
40 static struct mds_obd *MDS;
42 // XXX make this networked!
43 static int mds_queue_req(struct ptlrpc_request *req)
45 struct ptlrpc_request *srv_req;
52 OBD_ALLOC(srv_req, sizeof(*srv_req));
58 CDEBUG(0, "---> MDS at %d %p, incoming req %p, srv_req %p\n",
59 __LINE__, MDS, req, srv_req);
61 memset(srv_req, 0, sizeof(*req));
63 /* move the request buffer */
64 srv_req->rq_reqbuf = req->rq_reqbuf;
65 srv_req->rq_reqlen = req->rq_reqlen;
66 srv_req->rq_obd = MDS;
68 /* remember where it came from */
69 srv_req->rq_reply_handle = req;
71 list_add(&srv_req->rq_list, &MDS->mds_reqs);
72 wake_up(&MDS->mds_waitq);
76 int mds_sendpage(struct ptlrpc_request *req, struct file *file,
77 __u64 offset, struct niobuf *dst)
80 mm_segment_t oldfs = get_fs();
82 if (req->rq_peer.peer_nid == 0) {
83 /* dst->addr is a user address, but in a different task! */
85 rc = generic_file_read(file, (char *)(long)dst->addr,
94 OBD_ALLOC(buf, PAGE_SIZE);
99 rc = generic_file_read(file, buf, PAGE_SIZE, &offset);
102 if (rc != PAGE_SIZE) {
103 OBD_FREE(buf, PAGE_SIZE);
107 req->rq_bulkbuf = buf;
108 req->rq_bulklen = PAGE_SIZE;
109 rc = ptl_send_buf(req, &req->rq_peer, MDS_BULK_PORTAL, 0);
110 init_waitqueue_head(&req->rq_wait_for_bulk);
111 sleep_on(&req->rq_wait_for_bulk);
112 OBD_FREE(buf, PAGE_SIZE);
113 req->rq_bulklen = 0; /* FIXME: eek. */
119 int mds_reply(struct ptlrpc_request *req)
121 struct ptlrpc_request *clnt_req = req->rq_reply_handle;
125 if (req->rq_obd->mds_service != NULL) {
126 /* This is a request that came from the network via portals. */
128 /* FIXME: we need to increment the count of handled events */
129 ptl_send_buf(req, &req->rq_peer, MDS_REPLY_PORTAL, 0);
131 /* This is a local request that came from another thread. */
133 /* move the reply to the client */
134 clnt_req->rq_replen = req->rq_replen;
135 clnt_req->rq_repbuf = req->rq_repbuf;
136 req->rq_repbuf = NULL;
139 /* free the request buffer */
140 OBD_FREE(req->rq_reqbuf, req->rq_reqlen);
141 req->rq_reqbuf = NULL;
143 /* wake up the client */
144 wake_up_interruptible(&clnt_req->rq_wait_for_rep);
151 int mds_error(struct ptlrpc_request *req)
153 struct ptlrep_hdr *hdr;
157 OBD_ALLOC(hdr, sizeof(*hdr));
163 memset(hdr, 0, sizeof(*hdr));
165 hdr->seqno = req->rq_reqhdr->seqno;
166 hdr->status = req->rq_status;
167 hdr->type = MDS_TYPE_ERR;
169 req->rq_repbuf = (char *)hdr;
170 req->rq_replen = sizeof(*hdr);
173 return mds_reply(req);
176 struct dentry *mds_fid2dentry(struct mds_obd *mds, struct ll_fid *fid,
177 struct vfsmount **mnt)
179 /* stolen from NFS */
180 struct super_block *sb = mds->mds_sb;
181 unsigned long ino = fid->id;
182 //__u32 generation = fid->generation;
183 __u32 generation = 0;
185 struct list_head *lp;
186 struct dentry *result;
189 return ERR_PTR(-ESTALE);
191 inode = iget(sb, ino);
193 return ERR_PTR(-ENOMEM);
195 CDEBUG(D_DENTRY, "--> 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
215 *mnt = mds->mds_vfsmnt;
216 spin_lock(&dcache_lock);
217 for (lp = inode->i_dentry.next; lp != &inode->i_dentry ; lp=lp->next) {
218 result = list_entry(lp,struct dentry, d_alias);
219 if (! (result->d_flags & DCACHE_NFSD_DISCONNECTED)) {
221 result->d_vfs_flags |= DCACHE_REFERENCED;
222 spin_unlock(&dcache_lock);
229 spin_unlock(&dcache_lock);
230 result = d_alloc_root(inode);
231 if (result == NULL) {
233 return ERR_PTR(-ENOMEM);
237 result->d_flags |= DCACHE_NFSD_DISCONNECTED;
241 static inline void mds_get_objid(struct inode *inode, __u64 *id)
243 memcpy(id, &inode->u.ext2_i.i_data, sizeof(*id));
246 int mds_getattr(struct ptlrpc_request *req)
253 rc = mds_pack_rep(NULL, 0, NULL, 0, &req->rq_rephdr, &req->rq_rep.mds,
254 &req->rq_replen, &req->rq_repbuf);
257 printk("mds: out of memory\n");
258 req->rq_status = -ENOMEM;
262 req->rq_rephdr->seqno = req->rq_reqhdr->seqno;
263 rep = req->rq_rep.mds;
265 de = mds_fid2dentry(req->rq_obd, &req->rq_req.mds->fid1, NULL);
268 req->rq_rephdr->status = -ENOENT;
273 rep->ino = inode->i_ino;
274 rep->atime = inode->i_atime;
275 rep->ctime = inode->i_ctime;
276 rep->mtime = inode->i_mtime;
277 rep->uid = inode->i_uid;
278 rep->gid = inode->i_gid;
279 rep->size = inode->i_size;
280 rep->mode = inode->i_mode;
281 rep->nlink = inode->i_nlink;
283 mds_get_objid(inode, &rep->objid);
288 int mds_readpage(struct ptlrpc_request *req)
290 struct vfsmount *mnt;
293 struct niobuf *niobuf;
297 rc = mds_pack_rep(NULL, 0, NULL, 0, &req->rq_rephdr, &req->rq_rep.mds,
298 &req->rq_replen, &req->rq_repbuf);
301 printk("mds: out of memory\n");
302 req->rq_status = -ENOMEM;
306 req->rq_rephdr->seqno = req->rq_reqhdr->seqno;
307 rep = req->rq_rep.mds;
309 de = mds_fid2dentry(req->rq_obd, &req->rq_req.mds->fid1, &mnt);
312 req->rq_rephdr->status = PTR_ERR(de);
316 CDEBUG(D_INODE, "ino %ld\n", de->d_inode->i_ino);
318 file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
319 /* note: in case of an error, dentry_open puts dentry */
322 req->rq_rephdr->status = PTR_ERR(file);
326 niobuf = mds_req_tgt(req->rq_req.mds);
328 /* to make this asynchronous make sure that the handling function
329 doesn't send a reply when this function completes. Instead a
330 callback function would send the reply */
331 rc = mds_sendpage(req, file, req->rq_req.mds->size, niobuf);
334 req->rq_rephdr->status = rc;
339 int mds_reint(struct ptlrpc_request *req)
342 char *buf = mds_req_tgt(req->rq_req.mds);
343 int len = req->rq_req.mds->tgtlen;
344 struct mds_update_record rec;
346 rc = mds_update_unpack(buf, len, &rec);
348 printk(__FUNCTION__ ": invalid record\n");
349 req->rq_status = -EINVAL;
352 /* rc will be used to interrupt a for loop over multiple records */
353 rc = mds_reint_rec(&rec, req);
357 //int mds_handle(struct mds_conn *conn, int len, char *buf)
358 int mds_handle(struct ptlrpc_request *req)
361 struct ptlreq_hdr *hdr;
365 hdr = (struct ptlreq_hdr *)req->rq_reqbuf;
367 if (NTOH__u32(hdr->type) != MDS_TYPE_REQ) {
368 printk("lustre_mds: wrong packet type sent %d\n",
369 NTOH__u32(hdr->type));
374 rc = mds_unpack_req(req->rq_reqbuf, req->rq_reqlen,
375 &req->rq_reqhdr, &req->rq_req.mds);
377 printk("lustre_mds: Invalid request\n");
382 switch (req->rq_reqhdr->opc) {
385 CDEBUG(D_INODE, "getattr\n");
386 rc = mds_getattr(req);
390 CDEBUG(D_INODE, "readpage\n");
391 rc = mds_readpage(req);
395 CDEBUG(D_INODE, "reint\n");
400 return mds_error(req);
405 printk(__FUNCTION__ ": no header\n");
409 if( req->rq_status) {
412 CDEBUG(D_INODE, "sending reply\n");
420 static void mds_timer_run(unsigned long __data)
422 struct task_struct * p = (struct task_struct *) __data;
427 int mds_main(void *arg)
429 struct mds_obd *mds = (struct mds_obd *) arg;
430 struct timer_list timer;
434 spin_lock_irq(¤t->sigmask_lock);
435 sigfillset(¤t->blocked);
436 recalc_sigpending(current);
437 spin_unlock_irq(¤t->sigmask_lock);
439 sprintf(current->comm, "lustre_mds");
441 /* Set up an interval timer which can be used to trigger a
442 wakeup after the interval expires */
444 timer.data = (unsigned long) current;
445 timer.function = mds_timer_run;
446 mds->mds_timer = &timer;
448 /* Record that the thread is running */
449 mds->mds_thread = current;
450 wake_up(&mds->mds_done_waitq);
452 /* And now, wait forever for commit wakeup events. */
456 if (mds->mds_flags & MDS_UNMOUNT)
459 wake_up(&mds->mds_done_waitq);
460 interruptible_sleep_on(&mds->mds_waitq);
462 CDEBUG(D_INODE, "lustre_mds wakes\n");
463 CDEBUG(D_INODE, "pick up req here and continue\n");
465 if (mds->mds_service != NULL) {
469 struct ptlrpc_request request;
470 struct ptlrpc_service *service;
472 rc = PtlEQGet(mds->mds_service->srv_eq_h, &ev);
473 if (rc != PTL_OK && rc != PTL_EQ_DROPPED)
476 service = (struct ptlrpc_service *)ev.mem_desc.user_ptr;
478 /* FIXME: If we move to an event-driven model,
479 * we should put the request on the stack of
480 * mds_handle instead. */
481 memset(&request, 0, sizeof(request));
482 request.rq_reqbuf = ev.mem_desc.start +
484 request.rq_reqlen = ev.mem_desc.length;
485 request.rq_obd = MDS;
486 request.rq_xid = ev.match_bits;
488 request.rq_peer.peer_nid = ev.initiator.nid;
489 /* FIXME: this NI should be the incoming NI.
490 * We don't know how to find that from here. */
491 request.rq_peer.peer_ni =
492 mds->mds_service->srv_self.peer_ni;
493 rc = mds_handle(&request);
495 /* Inform the rpc layer the event has been handled */
496 ptl_received_rpc(service);
499 struct ptlrpc_request *request;
501 if (list_empty(&mds->mds_reqs)) {
502 CDEBUG(D_INODE, "woke because of timer\n");
504 request = list_entry(mds->mds_reqs.next,
505 struct ptlrpc_request,
507 list_del(&request->rq_list);
508 rc = mds_handle(request);
513 del_timer_sync(mds->mds_timer);
515 /* XXX maintain a list of all managed devices: cleanup here */
517 mds->mds_thread = NULL;
518 wake_up(&mds->mds_done_waitq);
519 printk("lustre_mds: exiting\n");
523 static void mds_stop_srv_thread(struct mds_obd *mds)
525 mds->mds_flags |= MDS_UNMOUNT;
527 while (mds->mds_thread) {
528 wake_up(&mds->mds_waitq);
529 sleep_on(&mds->mds_done_waitq);
533 static void mds_start_srv_thread(struct mds_obd *mds)
535 init_waitqueue_head(&mds->mds_waitq);
536 init_waitqueue_head(&mds->mds_done_waitq);
537 kernel_thread(mds_main, (void *)mds,
538 CLONE_VM | CLONE_FS | CLONE_FILES);
539 while (!mds->mds_thread)
540 sleep_on(&mds->mds_done_waitq);
543 /* mount the file system (secretly) */
544 static int mds_setup(struct obd_device *obddev, obd_count len,
548 struct obd_ioctl_data* data = buf;
549 struct mds_obd *mds = &obddev->u.mds;
550 struct vfsmount *mnt;
551 struct lustre_peer peer;
556 mnt = do_kern_mount(data->ioc_inlbuf2, 0, data->ioc_inlbuf1, NULL);
563 mds->mds_sb = mnt->mnt_root->d_inode->i_sb;
564 if (!obddev->u.mds.mds_sb) {
569 mds->mds_vfsmnt = mnt;
570 obddev->u.mds.mds_fstype = strdup(data->ioc_inlbuf2);
572 mds->mds_ctxt.pwdmnt = mnt;
573 mds->mds_ctxt.pwd = mnt->mnt_root;
574 mds->mds_ctxt.fs = KERNEL_DS;
575 mds->mds_remote_nid = 0;
577 INIT_LIST_HEAD(&mds->mds_reqs);
578 mds->mds_thread = NULL;
580 mds->mds_interval = 3 * HZ;
583 spin_lock_init(&obddev->u.mds.mds_lock);
585 err = kportal_uuid_to_peer("self", &peer);
587 OBD_ALLOC(mds->mds_service, sizeof(*mds->mds_service));
588 if (mds->mds_service == NULL)
590 mds->mds_service->srv_buf_size = 64 * 1024;
591 mds->mds_service->srv_portal = MDS_REQUEST_PORTAL;
592 memcpy(&mds->mds_service->srv_self, &peer, sizeof(peer));
593 mds->mds_service->srv_wait_queue = &mds->mds_waitq;
595 rpc_register_service(mds->mds_service, "self");
598 mds_start_srv_thread(mds);
605 static int mds_cleanup(struct obd_device * obddev)
607 struct super_block *sb;
608 struct mds_obd *mds = &obddev->u.mds;
612 if ( !(obddev->obd_flags & OBD_SET_UP) ) {
617 if ( !list_empty(&obddev->obd_gen_clients) ) {
618 printk(KERN_WARNING __FUNCTION__ ": still has clients!\n");
624 mds_stop_srv_thread(mds);
625 rpc_unregister_service(mds->mds_service);
626 OBD_FREE(mds->mds_service, sizeof(*mds->mds_service));
634 if (!list_empty(&mds->mds_reqs)) {
635 // XXX reply with errors and clean up
636 CDEBUG(D_INODE, "Request list not empty!\n");
640 mntput(mds->mds_vfsmnt);
642 kfree(mds->mds_fstype);
650 /* use obd ops to offer management infrastructure */
651 static struct obd_ops mds_obd_ops = {
653 o_cleanup: mds_cleanup,
656 static int __init mds_init(void)
658 obd_register_type(&mds_obd_ops, LUSTRE_MDS_NAME);
662 static void __exit mds_exit(void)
664 obd_unregister_type(LUSTRE_MDS_NAME);
667 MODULE_AUTHOR("Peter J. Braam <braam@clusterfs.com>");
668 MODULE_DESCRIPTION("Lustre Metadata Server (MDS) v0.01");
669 MODULE_LICENSE("GPL");
672 // for testing (maybe this stays)
673 EXPORT_SYMBOL(mds_queue_req);
675 module_init(mds_init);
676 module_exit(mds_exit);