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) 2001-2003 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>
11 * Author: Mike Shaver <shaver@clusterfs.com>
13 * This file is part of Lustre, http://www.lustre.org.
15 * Lustre is free software; you can redistribute it and/or
16 * modify it under the terms of version 2 of the GNU General Public
17 * License as published by the Free Software Foundation.
19 * Lustre is distributed in the hope that it will be useful,
20 * but WITHOUT ANY WARRANTY; without even the implied warranty of
21 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
22 * GNU General Public License for more details.
24 * You should have received a copy of the GNU General Public License
25 * along with Lustre; if not, write to the Free Software
26 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
30 #define DEBUG_SUBSYSTEM S_MDS
32 #include <linux/module.h>
33 #include <linux/lustre_mds.h>
34 #include <linux/lustre_dlm.h>
35 #include <linux/init.h>
36 #include <linux/obd_class.h>
37 #include <linux/random.h>
38 #include <linux/locks.h>
39 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
40 #include <linux/buffer_head.h>
41 #include <linux/workqueue.h>
43 #include <linux/obd_lov.h>
44 #include <linux/lustre_mds.h>
45 #include <linux/lustre_fsfilt.h>
46 #include <linux/lprocfs_status.h>
48 kmem_cache_t *mds_file_cache;
50 extern int mds_get_lovtgts(struct mds_obd *obd, int tgt_count,
51 struct obd_uuid *uuidarray);
52 extern int mds_get_lovdesc(struct mds_obd *obd, struct lov_desc *desc);
53 extern void mds_start_transno(struct mds_obd *mds);
54 extern int mds_finish_transno(struct mds_obd *mds, void *handle,
55 struct ptlrpc_request *req, int rc);
56 static int mds_cleanup(struct obd_device * obddev);
58 inline struct mds_obd *mds_req2mds(struct ptlrpc_request *req)
60 return &req->rq_export->exp_obd->u.mds;
63 static int mds_bulk_timeout(void *data)
65 struct ptlrpc_bulk_desc *desc = data;
68 recovd_conn_fail(desc->bd_connection);
72 /* Assumes caller has already pushed into the kernel filesystem context */
73 static int mds_sendpage(struct ptlrpc_request *req, struct file *file,
74 __u64 offset, __u64 xid)
76 struct ptlrpc_bulk_desc *desc;
77 struct ptlrpc_bulk_page *bulk;
78 struct l_wait_info lwi;
83 desc = ptlrpc_prep_bulk(req->rq_connection);
85 GOTO(out, rc = -ENOMEM);
87 bulk = ptlrpc_prep_bulk_page(desc);
89 GOTO(cleanup_bulk, rc = -ENOMEM);
91 OBD_ALLOC(buf, PAGE_CACHE_SIZE);
93 GOTO(cleanup_bulk, rc = -ENOMEM);
95 CDEBUG(D_EXT2, "reading %lu@"LPU64" from dir %lu (size %llu)\n",
96 PAGE_CACHE_SIZE, offset, file->f_dentry->d_inode->i_ino,
97 file->f_dentry->d_inode->i_size);
98 rc = fsfilt_readpage(req->rq_export->exp_obd, file, buf,
99 PAGE_CACHE_SIZE, (loff_t *)&offset);
101 if (rc != PAGE_CACHE_SIZE)
102 GOTO(cleanup_buf, rc = -EIO);
106 bulk->bp_buflen = PAGE_CACHE_SIZE;
107 desc->bd_ptl_ev_hdlr = NULL;
108 desc->bd_portal = MDS_BULK_PORTAL;
110 rc = ptlrpc_bulk_put(desc);
112 GOTO(cleanup_buf, rc);
114 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
115 CERROR("obd_fail_loc=%x, fail operation rc=%d\n",
116 OBD_FAIL_MDS_SENDPAGE, rc);
117 ptlrpc_abort_bulk(desc);
118 GOTO(cleanup_buf, rc);
121 lwi = LWI_TIMEOUT(obd_timeout * HZ, mds_bulk_timeout, desc);
122 rc = l_wait_event(desc->bd_waitq, desc->bd_flags & PTL_BULK_FL_SENT,
125 if (rc != -ETIMEDOUT)
127 GOTO(cleanup_buf, rc);
132 OBD_FREE(buf, PAGE_SIZE);
134 ptlrpc_bulk_decref(desc);
139 /* only valid locked dentries or errors should be returned */
140 struct dentry *mds_fid2locked_dentry(struct obd_device *obd, struct ll_fid *fid,
141 struct vfsmount **mnt, int lock_mode,
142 struct lustre_handle *lockh)
144 struct mds_obd *mds = &obd->u.mds;
145 struct dentry *de = mds_fid2dentry(mds, fid, mnt), *retval = de;
146 struct ldlm_res_id res_id = { .name = {0} };
153 res_id.name[0] = de->d_inode->i_ino;
154 res_id.name[1] = de->d_inode->i_generation;
155 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, NULL,
156 res_id, LDLM_PLAIN, NULL, 0, lock_mode,
157 &flags, ldlm_completion_ast,
158 mds_blocking_ast, NULL, NULL, lockh);
159 if (rc != ELDLM_OK) {
161 retval = ERR_PTR(-ENOLCK); /* XXX translate ldlm code */
167 #ifndef DCACHE_DISCONNECTED
168 #define DCACHE_DISCONNECTED DCACHE_NFSD_DISCONNECTED
173 /* Look up an entry by inode number. */
174 /* this function ONLY returns valid dget'd dentries with an initialized inode
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;
184 struct list_head *lp;
185 struct dentry *result;
188 RETURN(ERR_PTR(-ESTALE));
190 inode = iget(sb, ino);
192 RETURN(ERR_PTR(-ENOMEM));
194 CDEBUG(D_DENTRY, "--> mds_fid2dentry: sb %p\n", inode->i_sb);
196 if (is_bad_inode(inode) ||
197 (generation && inode->i_generation != generation)) {
198 /* we didn't find the right inode.. */
199 CERROR("bad inode %lu, link: %d ct: %d or version %u/%u\n",
200 inode->i_ino, inode->i_nlink,
201 atomic_read(&inode->i_count), inode->i_generation,
204 RETURN(ERR_PTR(-ENOENT));
207 /* now to find a dentry. If possible, get a well-connected one */
209 *mnt = mds->mds_vfsmnt;
210 spin_lock(&dcache_lock);
211 list_for_each(lp, &inode->i_dentry) {
212 result = list_entry(lp, struct dentry, d_alias);
213 if (!(result->d_flags & DCACHE_DISCONNECTED)) {
215 result->d_vfs_flags |= DCACHE_REFERENCED;
216 spin_unlock(&dcache_lock);
223 spin_unlock(&dcache_lock);
224 result = d_alloc_root(inode);
225 if (result == NULL) {
227 return ERR_PTR(-ENOMEM);
231 result->d_flags |= DCACHE_DISCONNECTED;
235 static void mds_abort_recovery(void *data);
237 /* Establish a connection to the MDS.
239 * This will set up an export structure for the client to hold state data
240 * about that client, like open files, the last operation number it did
241 * on the server, etc.
243 static int mds_connect(struct lustre_handle *conn, struct obd_device *obd,
244 struct obd_uuid *cluuid, struct recovd_obd *recovd,
245 ptlrpc_recovery_cb_t recover)
247 struct obd_export *exp;
248 struct mds_export_data *med;
249 struct mds_client_data *mcd;
250 struct mds_obd *mds = &obd->u.mds;
254 if (!conn || !obd || !cluuid)
257 /* Check for aborted recovery. */
258 spin_lock_bh(&mds->mds_processing_task_lock);
259 if (obd->obd_flags & OBD_ABORT_RECOVERY)
260 mds_abort_recovery(mds);
261 spin_unlock_bh(&mds->mds_processing_task_lock);
263 /* XXX There is a small race between checking the list and adding a
264 * new connection for the same UUID, but the real threat (list
265 * corruption when multiple different clients connect) is solved.
267 * There is a second race between adding the export to the list,
268 * and filling in the client data below. Hence skipping the case
269 * of NULL mcd above. We should already be controlling multiple
270 * connects at the client, and we can't hold the spinlock over
271 * memory allocations without risk of deadlocking.
273 rc = class_connect(conn, obd, cluuid);
276 exp = class_conn2export(conn);
278 med = &exp->exp_mds_data;
280 OBD_ALLOC(mcd, sizeof(*mcd));
282 CERROR("mds: out of memory for client data\n");
283 GOTO(out_export, rc = -ENOMEM);
286 memcpy(mcd->mcd_uuid, cluuid, sizeof(mcd->mcd_uuid));
289 INIT_LIST_HEAD(&med->med_open_head);
290 spin_lock_init(&med->med_open_lock);
292 rc = mds_client_add(&obd->u.mds, med, -1);
299 OBD_FREE(mcd, sizeof(*mcd));
301 class_disconnect(conn);
306 /* Call with med->med_open_lock held, please. */
307 inline int mds_close_mfd(struct mds_file_data *mfd, struct mds_export_data *med)
309 struct file *file = mfd->mfd_file;
311 struct dentry *de = NULL;
312 LASSERT(file->private_data == mfd);
314 LASSERT(mfd->mfd_servercookie != DEAD_HANDLE_MAGIC);
316 list_del(&mfd->mfd_list);
317 mfd->mfd_servercookie = DEAD_HANDLE_MAGIC;
318 kmem_cache_free(mds_file_cache, mfd);
320 if (file->f_dentry->d_parent)
321 de = dget(file->f_dentry->d_parent);
322 rc = filp_close(file, 0);
328 static int mds_disconnect(struct lustre_handle *conn)
330 struct obd_export *export = class_conn2export(conn);
331 struct list_head *tmp, *n;
332 struct mds_export_data *med = &export->exp_mds_data;
337 * Close any open files.
339 spin_lock(&med->med_open_lock);
340 list_for_each_safe(tmp, n, &med->med_open_head) {
341 struct mds_file_data *mfd =
342 list_entry(tmp, struct mds_file_data, mfd_list);
343 CERROR("force closing client file handle for %*s\n",
344 mfd->mfd_file->f_dentry->d_name.len,
345 mfd->mfd_file->f_dentry->d_name.name);
346 rc = mds_close_mfd(mfd, med);
348 CDEBUG(D_INODE, "Error closing file: %d\n", rc);
350 spin_unlock(&med->med_open_lock);
352 ldlm_cancel_locks_for_export(export);
353 mds_client_free(export);
355 rc = class_disconnect(conn);
361 * XXX This is NOT guaranteed to flush all transactions to disk (even though
362 * it is equivalent to calling sync()) because it only _starts_ the flush
363 * and does not wait for completion. It's better than nothing though.
364 * What we really want is a mild form of fsync_dev_lockfs(), but it is
365 * non-standard, or enabling do_sync_supers in ext3, just for this call.
367 static void mds_fsync_super(struct super_block *sb)
371 if (sb->s_dirt && sb->s_op && sb->s_op->write_super)
372 sb->s_op->write_super(sb);
377 static int mds_getstatus(struct ptlrpc_request *req)
379 struct mds_obd *mds = mds_req2mds(req);
380 struct mds_body *body;
381 int rc, size = sizeof(*body);
384 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
385 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK)) {
386 CERROR("mds: out of memory for message: size=%d\n", size);
387 req->rq_status = -ENOMEM;
391 /* Flush any outstanding transactions to disk so the client will
392 * get the latest last_committed value and can drop their local
393 * requests if they have any. This would be fsync_super() if it
396 mds_fsync_super(mds->mds_sb);
398 body = lustre_msg_buf(req->rq_repmsg, 0);
399 memcpy(&body->fid1, &mds->mds_rootfid, sizeof(body->fid1));
401 /* the last_committed and last_xid fields are filled in for all
402 * replies already - no need to do so here also.
407 static int mds_getlovinfo(struct ptlrpc_request *req)
409 struct mds_obd *mds = mds_req2mds(req);
410 struct mds_status_req *streq;
411 struct lov_desc *desc;
413 int rc, size[2] = {sizeof(*desc)};
416 streq = lustre_msg_buf(req->rq_reqmsg, 0);
417 streq->flags = NTOH__u32(streq->flags);
418 streq->repbuf = NTOH__u32(streq->repbuf);
419 size[1] = streq->repbuf;
421 rc = lustre_pack_msg(2, size, NULL, &req->rq_replen, &req->rq_repmsg);
423 CERROR("mds: out of memory for message: size=%d\n", size[1]);
424 req->rq_status = -ENOMEM;
428 if (!mds->mds_has_lov_desc) {
429 req->rq_status = -ENOENT;
433 desc = lustre_msg_buf(req->rq_repmsg, 0);
434 memcpy(desc, &mds->mds_lov_desc, sizeof *desc);
436 tgt_count = le32_to_cpu(desc->ld_tgt_count);
437 if (tgt_count * sizeof(struct obd_uuid) > streq->repbuf) {
438 CERROR("too many targets, enlarge client buffers\n");
439 req->rq_status = -ENOSPC;
443 rc = mds_get_lovtgts(mds, tgt_count,
444 lustre_msg_buf(req->rq_repmsg, 1));
446 CERROR("get_lovtgts error %d\n", rc);
453 int mds_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
454 void *data, int flag)
459 if (flag == LDLM_CB_CANCELING) {
460 /* Don't need to do anything here. */
464 /* XXX layering violation! -phil */
465 l_lock(&lock->l_resource->lr_namespace->ns_lock);
466 /* Get this: if mds_blocking_ast is racing with ldlm_intent_policy,
467 * such that mds_blocking_ast is called just before l_i_p takes the
468 * ns_lock, then by the time we get the lock, we might not be the
469 * correct blocking function anymore. So check, and return early, if
471 if (lock->l_blocking_ast != mds_blocking_ast) {
472 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
476 lock->l_flags |= LDLM_FL_CBPENDING;
477 do_ast = (!lock->l_readers && !lock->l_writers);
478 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
481 struct lustre_handle lockh;
484 LDLM_DEBUG0(lock, "already unused, calling ldlm_cli_cancel");
485 ldlm_lock2handle(lock, &lockh);
486 rc = ldlm_cli_cancel(&lockh);
488 CERROR("ldlm_cli_cancel: %d\n", rc);
490 LDLM_DEBUG0(lock, "Lock still has references, will be "
496 int mds_pack_md(struct obd_device *obd, struct lustre_msg *msg,
497 int offset, struct mds_body *body, struct inode *inode)
499 struct mds_obd *mds = &obd->u.mds;
500 struct lov_mds_md *lmm;
501 int lmm_size = msg->buflens[offset];
506 CDEBUG(D_INFO, "no space reserved for inode %lu MD\n",
511 lmm = lustre_msg_buf(msg, offset);
513 /* I don't really like this, but it is a sanity check on the client
514 * MD request. However, if the client doesn't know how much space
515 * to reserve for the MD, this shouldn't be fatal either...
517 if (lmm_size > mds->mds_max_mdsize) {
518 CERROR("Reading MD for inode %lu of %d bytes > max %d\n",
519 inode->i_ino, lmm_size, mds->mds_max_mdsize);
523 /* We don't need to store the reply size, because this buffer is
524 * discarded right after unpacking, and the LOV can figure out the
525 * size itself from the ost count.
527 if ((rc = fsfilt_get_md(obd, inode, lmm, lmm_size)) < 0) {
528 CDEBUG(D_INFO, "No md for ino %lu: rc = %d\n",
531 body->valid |= OBD_MD_FLEASIZE;
538 static int mds_getattr_internal(struct obd_device *obd, struct dentry *dentry,
539 struct ptlrpc_request *req,
540 struct mds_body *reqbody, int reply_off)
542 struct mds_body *body;
543 struct inode *inode = dentry->d_inode;
550 body = lustre_msg_buf(req->rq_repmsg, reply_off);
552 mds_pack_inode2fid(&body->fid1, inode);
553 mds_pack_inode2body(body, inode);
555 if (S_ISREG(inode->i_mode) && reqbody->valid & OBD_MD_FLEASIZE) {
556 rc = mds_pack_md(obd, req->rq_repmsg, reply_off + 1,
558 } else if (S_ISLNK(inode->i_mode) && reqbody->valid & OBD_MD_LINKNAME) {
559 char *symname = lustre_msg_buf(req->rq_repmsg, reply_off + 1);
560 int len = req->rq_repmsg->buflens[reply_off + 1];
562 rc = inode->i_op->readlink(dentry, symname, len);
564 CERROR("readlink failed: %d\n", rc);
566 CDEBUG(D_INODE, "read symlink dest %s\n", symname);
567 body->valid |= OBD_MD_LINKNAME;
574 static int mds_getattr_pack_msg(struct ptlrpc_request *req, struct inode *inode,
577 struct mds_obd *mds = mds_req2mds(req);
578 struct mds_body *body;
579 int rc = 0, size[2] = {sizeof(*body)}, bufcount = 1;
582 body = lustre_msg_buf(req->rq_reqmsg, offset);
584 if (S_ISREG(inode->i_mode) && body->valid & OBD_MD_FLEASIZE) {
585 int rc = fsfilt_get_md(req->rq_export->exp_obd, inode, NULL, 0);
586 CDEBUG(D_INODE, "got %d bytes MD data for inode %lu\n",
590 CERROR("error getting inode %lu MD: rc = %d\n",
593 } else if (rc > mds->mds_max_mdsize) {
595 CERROR("MD size %d larger than maximum possible %u\n",
596 rc, mds->mds_max_mdsize);
600 } else if (body->valid & OBD_MD_LINKNAME) {
601 size[bufcount] = MIN(inode->i_size + 1, body->size);
603 CDEBUG(D_INODE, "symlink size: %Lu, reply space: "LPU64"\n",
604 inode->i_size + 1, body->size);
607 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
608 CERROR("failed MDS_GETATTR_PACK test\n");
609 req->rq_status = -ENOMEM;
610 GOTO(out, rc = -ENOMEM);
613 rc = lustre_pack_msg(bufcount, size, NULL, &req->rq_replen,
616 CERROR("out of memoryK\n");
626 static int mds_getattr_name(int offset, struct ptlrpc_request *req,
627 struct lustre_handle *child_lockh)
629 struct ldlm_intent *it = lustre_msg_buf(req->rq_reqmsg, 1);
631 struct mds_obd *mds = mds_req2mds(req);
632 struct obd_device *obd = req->rq_export->exp_obd;
633 struct obd_run_ctxt saved;
634 struct mds_body *body;
635 struct dentry *de = NULL, *dchild = NULL;
638 struct ldlm_res_id child_res_id = { .name = {0} };
639 struct lustre_handle parent_lockh;
640 int namelen, flags = 0, rc = 0;
644 LASSERT(!strcmp(obd->obd_type->typ_name, "mds"));
646 if (req->rq_reqmsg->bufcount <= offset + 1) {
648 GOTO(out_pre_de, rc = -EINVAL);
651 body = lustre_msg_buf(req->rq_reqmsg, offset);
652 name = lustre_msg_buf(req->rq_reqmsg, offset + 1);
653 namelen = req->rq_reqmsg->buflens[offset + 1];
654 /* requests were at offset 2, replies go back at 1 */
658 uc.ouc_fsuid = body->fsuid;
659 uc.ouc_fsgid = body->fsgid;
660 uc.ouc_cap = body->capability;
661 uc.ouc_suppgid = body->suppgid;
662 push_ctxt(&saved, &mds->mds_ctxt, &uc);
663 /* Step 1: Lookup/lock parent */
664 de = mds_fid2locked_dentry(obd, &body->fid1, NULL, LCK_PR,
667 GOTO(out_pre_de, rc = PTR_ERR(de));
671 CDEBUG(D_INODE, "parent ino %lu, name %*s\n", dir->i_ino,namelen,name);
673 /* Step 2: Lookup child */
674 dchild = lookup_one_len(name, de, namelen - 1);
675 if (IS_ERR(dchild)) {
676 CDEBUG(D_INODE, "child lookup error %ld\n", PTR_ERR(dchild));
677 GOTO(out_step_1, rc = PTR_ERR(dchild));
678 } else if (dchild->d_inode == NULL) {
679 GOTO(out_step_2, rc = -ENOENT);
682 /* Step 3: Lock child */
683 if (it->opc == IT_SETATTR)
687 child_res_id.name[0] = dchild->d_inode->i_ino;
688 child_res_id.name[1] = dchild->d_inode->i_generation;
689 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, NULL,
690 child_res_id, LDLM_PLAIN, NULL, 0, lock_mode,
691 &flags, ldlm_completion_ast, mds_blocking_ast,
692 NULL, NULL, child_lockh);
693 if (rc != ELDLM_OK) {
694 CERROR("ldlm_cli_enqueue: %d\n", rc);
695 GOTO(out_step_2, rc = -EIO);
698 if (req->rq_repmsg == NULL)
699 mds_getattr_pack_msg(req, dchild->d_inode, offset);
701 rc = mds_getattr_internal(obd, dchild, req, body, offset);
703 GOTO(out_step_3, rc);
704 GOTO(out_step_2, rc); /* returns the lock to the client */
706 ldlm_lock_decref(child_lockh, LCK_PR);
710 ldlm_lock_decref(&parent_lockh, LCK_PR);
714 pop_ctxt(&saved, &mds->mds_ctxt, &uc);
718 static int mds_getattr(int offset, struct ptlrpc_request *req)
720 struct mds_obd *mds = mds_req2mds(req);
721 struct obd_device *obd = req->rq_export->exp_obd;
722 struct obd_run_ctxt saved;
724 struct mds_body *body;
729 body = lustre_msg_buf(req->rq_reqmsg, offset);
730 uc.ouc_fsuid = body->fsuid;
731 uc.ouc_fsgid = body->fsgid;
732 uc.ouc_cap = body->capability;
733 push_ctxt(&saved, &mds->mds_ctxt, &uc);
734 de = mds_fid2dentry(mds, &body->fid1, NULL);
736 rc = req->rq_status = -ENOENT;
737 GOTO(out_pop, PTR_ERR(de));
740 rc = mds_getattr_pack_msg(req, de->d_inode, offset);
742 req->rq_status = mds_getattr_internal(obd, de, req, body, 0);
747 pop_ctxt(&saved, &mds->mds_ctxt, &uc);
751 static int mds_statfs(struct ptlrpc_request *req)
753 struct obd_device *obd = req->rq_export->exp_obd;
754 struct obd_statfs *osfs;
755 int rc, size = sizeof(*osfs);
758 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
759 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
760 CERROR("mds: statfs lustre_pack_msg failed: rc = %d\n", rc);
764 osfs = lustre_msg_buf(req->rq_repmsg, 0);
765 rc = fsfilt_statfs(obd, obd->u.mds.mds_sb, osfs);
767 CERROR("mds: statfs failed: rc %d\n", rc);
770 obd_statfs_pack(osfs, osfs);
778 static struct mds_file_data *mds_handle2mfd(struct lustre_handle *handle)
780 struct mds_file_data *mfd = NULL;
783 if (!handle || !handle->addr)
786 mfd = (struct mds_file_data *)(unsigned long)(handle->addr);
787 if (!kmem_cache_validate(mds_file_cache, mfd))
790 if (mfd->mfd_servercookie != handle->cookie)
798 static int mds_store_md(struct mds_obd *mds, struct ptlrpc_request *req,
799 int offset, struct mds_body *body, struct inode *inode)
801 struct obd_device *obd = req->rq_export->exp_obd;
802 struct lov_mds_md *lmm = lustre_msg_buf(req->rq_reqmsg, offset);
803 int lmm_size = req->rq_reqmsg->buflens[offset];
804 struct obd_run_ctxt saved;
810 /* I don't really like this, but it is a sanity check on the client
813 if (lmm_size > mds->mds_max_mdsize) {
814 CERROR("Saving MD for inode %lu of %d bytes > max %d\n",
815 inode->i_ino, lmm_size, mds->mds_max_mdsize);
819 CDEBUG(D_INODE, "storing %d bytes MD for inode %lu\n",
820 lmm_size, inode->i_ino);
821 uc.ouc_fsuid = body->fsuid;
822 uc.ouc_fsgid = body->fsgid;
823 uc.ouc_cap = body->capability;
824 push_ctxt(&saved, &mds->mds_ctxt, &uc);
825 mds_start_transno(mds);
826 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR);
827 if (IS_ERR(handle)) {
828 rc = PTR_ERR(handle);
829 mds_finish_transno(mds, handle, req, rc);
833 rc = fsfilt_set_md(obd, inode,handle,lmm,lmm_size);
834 rc = mds_finish_transno(mds, handle, req, rc);
836 rc2 = fsfilt_commit(obd, inode, handle);
840 pop_ctxt(&saved, &mds->mds_ctxt, &uc);
847 static int mds_close(struct ptlrpc_request *req)
849 struct mds_export_data *med = &req->rq_export->exp_mds_data;
850 struct mds_body *body;
851 struct mds_file_data *mfd;
855 body = lustre_msg_buf(req->rq_reqmsg, 0);
857 mfd = mds_handle2mfd(&body->handle);
859 DEBUG_REQ(D_ERROR, req, "no handle for file close "LPD64
860 ": addr "LPX64", cookie "LPX64"\n",
861 body->fid1.id, body->handle.addr,
862 body->handle.cookie);
866 spin_lock(&med->med_open_lock);
867 req->rq_status = mds_close_mfd(mfd, med);
868 spin_unlock(&med->med_open_lock);
870 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
871 CERROR("test case OBD_FAIL_MDS_CLOSE_PACK\n");
872 req->rq_status = -ENOMEM;
876 rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen, &req->rq_repmsg);
878 CERROR("mds: lustre_pack_msg: rc = %d\n", rc);
885 static int mds_readpage(struct ptlrpc_request *req)
887 struct mds_obd *mds = mds_req2mds(req);
888 struct vfsmount *mnt;
891 struct mds_body *body, *repbody;
892 struct obd_run_ctxt saved;
893 int rc, size = sizeof(*body);
897 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
898 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK)) {
899 CERROR("mds: out of memory\n");
900 GOTO(out, rc = -ENOMEM);
903 body = lustre_msg_buf(req->rq_reqmsg, 0);
904 uc.ouc_fsuid = body->fsuid;
905 uc.ouc_fsgid = body->fsgid;
906 uc.ouc_cap = body->capability;
907 push_ctxt(&saved, &mds->mds_ctxt, &uc);
908 de = mds_fid2dentry(mds, &body->fid1, &mnt);
910 GOTO(out_pop, rc = PTR_ERR(de));
912 CDEBUG(D_INODE, "ino %lu\n", de->d_inode->i_ino);
914 file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
915 /* note: in case of an error, dentry_open puts dentry */
917 GOTO(out_pop, rc = PTR_ERR(file));
919 repbody = lustre_msg_buf(req->rq_repmsg, 0);
920 repbody->size = file->f_dentry->d_inode->i_size;
921 repbody->valid = OBD_MD_FLSIZE;
923 /* to make this asynchronous make sure that the handling function
924 doesn't send a reply when this function completes. Instead a
925 callback function would send the reply */
926 /* body->blocks is actually the xid -phil */
927 rc = mds_sendpage(req, file, body->size, body->blocks);
931 pop_ctxt(&saved, &mds->mds_ctxt, &uc);
937 int mds_reint(struct ptlrpc_request *req, int offset,
938 struct lustre_handle *lockh)
940 struct mds_update_record *rec; /* 116 bytes on the stack? no sir! */
943 OBD_ALLOC(rec, sizeof(*rec));
947 rc = mds_update_unpack(req, offset, rec);
948 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
949 CERROR("invalid record\n");
950 GOTO(out, req->rq_status = -EINVAL);
952 /* rc will be used to interrupt a for loop over multiple records */
953 rc = mds_reint_rec(rec, offset, req, lockh);
955 OBD_FREE(rec, sizeof(*rec));
959 /* forward declaration */
960 int mds_handle(struct ptlrpc_request *req);
962 static void abort_delayed_replies(struct mds_obd *mds)
964 struct ptlrpc_request *req;
965 struct list_head *tmp, *n;
966 list_for_each_safe(tmp, n, &mds->mds_delayed_reply_queue) {
967 req = list_entry(tmp, struct ptlrpc_request, rq_list);
968 DEBUG_REQ(D_ERROR, req, "aborted:");
969 req->rq_status = -ENOTCONN;
970 req->rq_type = PTL_RPC_MSG_ERR;
971 ptlrpc_reply(req->rq_svc, req);
972 list_del(&req->rq_list);
973 OBD_FREE(req, sizeof *req);
977 static void mds_abort_recovery(void *data)
979 struct mds_obd *mds = data;
980 struct obd_device *obd = list_entry(mds, struct obd_device, u.mds);
981 CERROR("disconnecting clients and aborting recovery\n");
982 mds->mds_recoverable_clients = 0;
983 obd->obd_flags &= ~(OBD_RECOVERING | OBD_ABORT_RECOVERY);
984 abort_delayed_replies(mds);
985 spin_unlock_bh(&mds->mds_processing_task_lock);
986 class_disconnect_all(obd);
987 spin_lock_bh(&mds->mds_processing_task_lock);
990 static void mds_recovery_expired(unsigned long castmeharder)
992 struct mds_obd *mds = (struct mds_obd *)castmeharder;
993 struct obd_device *obd = list_entry(mds, struct obd_device, u.mds);
994 CERROR("recovery timed out, aborting\n");
995 spin_lock_bh(&mds->mds_processing_task_lock);
996 obd->obd_flags |= OBD_ABORT_RECOVERY;
997 wake_up(&mds->mds_next_transno_waitq);
998 spin_unlock_bh(&mds->mds_processing_task_lock);
1001 static void reset_recovery_timer(struct mds_obd *mds)
1003 CDEBUG(D_ERROR, "timer will expire in %ld seconds\n",
1004 MDS_RECOVERY_TIMEOUT / HZ);
1005 mod_timer(&mds->mds_recovery_timer, jiffies + MDS_RECOVERY_TIMEOUT);
1008 static void start_recovery_timer(struct mds_obd *mds)
1010 mds->mds_recovery_timer.function = mds_recovery_expired;
1011 mds->mds_recovery_timer.data = (unsigned long)mds;
1012 init_timer(&mds->mds_recovery_timer);
1013 reset_recovery_timer(mds);
1016 static void cancel_recovery_timer(struct mds_obd *mds)
1018 del_timer(&mds->mds_recovery_timer);
1021 static int check_for_next_transno(struct mds_obd *mds)
1023 struct ptlrpc_request *req;
1024 struct obd_device *obd = list_entry(mds, struct obd_device, u.mds);
1025 req = list_entry(mds->mds_recovery_queue.next,
1026 struct ptlrpc_request, rq_list);
1027 LASSERT(req->rq_reqmsg->transno >= mds->mds_next_recovery_transno);
1029 return req->rq_reqmsg->transno == mds->mds_next_recovery_transno ||
1030 (obd->obd_flags & OBD_RECOVERING) == 0;
1033 static void process_recovery_queue(struct mds_obd *mds)
1035 struct ptlrpc_request *req;
1036 struct obd_device *obd = list_entry(mds, struct obd_device, u.mds);
1041 spin_lock_bh(&mds->mds_processing_task_lock);
1042 LASSERT(mds->mds_processing_task == current->pid);
1043 req = list_entry(mds->mds_recovery_queue.next,
1044 struct ptlrpc_request, rq_list);
1046 if (req->rq_reqmsg->transno != mds->mds_next_recovery_transno) {
1047 spin_unlock_bh(&mds->mds_processing_task_lock);
1048 CDEBUG(D_HA, "Waiting for transno "LPD64" (1st is "
1050 mds->mds_next_recovery_transno,
1051 req->rq_reqmsg->transno);
1052 wait_event(mds->mds_next_transno_waitq,
1053 check_for_next_transno(mds));
1054 spin_lock_bh(&mds->mds_processing_task_lock);
1055 if (obd->obd_flags & OBD_ABORT_RECOVERY) {
1056 mds_abort_recovery(mds);
1059 spin_unlock_bh(&mds->mds_processing_task_lock);
1064 list_del_init(&req->rq_list);
1065 spin_unlock_bh(&mds->mds_processing_task_lock);
1067 DEBUG_REQ(D_ERROR, req, "processing: ");
1068 (void)mds_handle(req);
1069 reset_recovery_timer(mds);
1070 mds_fsync_super(mds->mds_sb);
1071 OBD_FREE(req, sizeof *req);
1072 spin_lock_bh(&mds->mds_processing_task_lock);
1073 mds->mds_next_recovery_transno++;
1074 if (list_empty(&mds->mds_recovery_queue)) {
1075 mds->mds_processing_task = 0;
1076 spin_unlock_bh(&mds->mds_processing_task_lock);
1079 spin_unlock_bh(&mds->mds_processing_task_lock);
1084 static int queue_recovery_request(struct ptlrpc_request *req,
1085 struct mds_obd *mds)
1087 struct list_head *tmp;
1089 __u64 transno = req->rq_reqmsg->transno;
1090 struct ptlrpc_request *saved_req;
1093 INIT_LIST_HEAD(&req->rq_list);
1094 DEBUG_REQ(D_HA, req, "not queueing");
1098 spin_lock_bh(&mds->mds_processing_task_lock);
1100 if (mds->mds_processing_task == current->pid) {
1101 /* Processing the queue right now, don't re-add. */
1102 LASSERT(list_empty(&req->rq_list));
1103 spin_unlock_bh(&mds->mds_processing_task_lock);
1107 OBD_ALLOC(saved_req, sizeof *saved_req);
1110 memcpy(saved_req, req, sizeof *req);
1112 INIT_LIST_HEAD(&req->rq_list);
1115 list_for_each(tmp, &mds->mds_recovery_queue) {
1116 struct ptlrpc_request *reqiter =
1117 list_entry(tmp, struct ptlrpc_request, rq_list);
1119 if (reqiter->rq_reqmsg->transno > transno) {
1120 list_add_tail(&req->rq_list, &reqiter->rq_list);
1127 list_add_tail(&req->rq_list, &mds->mds_recovery_queue);
1130 if (mds->mds_processing_task != 0) {
1131 /* Someone else is processing this queue, we'll leave it to
1134 if (transno == mds->mds_next_recovery_transno)
1135 wake_up(&mds->mds_next_transno_waitq);
1136 spin_unlock_bh(&mds->mds_processing_task_lock);
1140 /* Nobody is processing, and we know there's (at least) one to process
1141 * now, so we'll do the honours.
1143 mds->mds_processing_task = current->pid;
1144 spin_unlock_bh(&mds->mds_processing_task_lock);
1146 process_recovery_queue(mds);
1150 static int filter_recovery_request(struct ptlrpc_request *req,
1151 struct mds_obd *mds, int *process)
1153 switch (req->rq_reqmsg->opc) {
1154 case MDS_CONNECT: /* This will never get here, but for completeness. */
1155 case MDS_DISCONNECT:
1160 case MDS_GETSTATUS: /* used in unmounting */
1163 *process = queue_recovery_request(req, mds);
1167 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
1169 /* XXX what should we set rq_status to here? */
1170 RETURN(ptlrpc_error(req->rq_svc, req));
1174 static int mds_queue_final_reply(struct ptlrpc_request *req, int rc)
1176 struct mds_obd *mds = mds_req2mds(req);
1177 struct obd_device *mds_obd = list_entry(mds, struct obd_device, u.mds);
1178 struct ptlrpc_request *saved_req;
1180 spin_lock_bh(&mds->mds_processing_task_lock);
1182 /* Just like ptlrpc_error, but without the sending. */
1183 lustre_pack_msg(0, NULL, NULL, &req->rq_replen,
1185 req->rq_type = PTL_RPC_MSG_ERR;
1188 LASSERT(list_empty(&req->rq_list));
1189 OBD_ALLOC(saved_req, sizeof *saved_req);
1190 memcpy(saved_req, req, sizeof *saved_req);
1192 list_add(&req->rq_list, &mds->mds_delayed_reply_queue);
1193 if (--mds->mds_recoverable_clients == 0) {
1194 struct list_head *tmp, *n;
1195 ldlm_reprocess_all_ns(req->rq_export->exp_obd->obd_namespace);
1197 "all clients recovered, sending delayed replies\n");
1198 mds_obd->obd_flags &= ~OBD_RECOVERING;
1199 list_for_each_safe(tmp, n, &mds->mds_delayed_reply_queue) {
1200 req = list_entry(tmp, struct ptlrpc_request, rq_list);
1201 DEBUG_REQ(D_ERROR, req, "delayed:");
1202 ptlrpc_reply(req->rq_svc, req);
1203 list_del(&req->rq_list);
1204 OBD_FREE(req, sizeof *req);
1206 cancel_recovery_timer(mds);
1208 CERROR("%d recoverable clients remain\n",
1209 mds->mds_recoverable_clients);
1212 spin_unlock_bh(&mds->mds_processing_task_lock);
1216 static char *reint_names[] = {
1217 [REINT_SETATTR] "setattr",
1218 [REINT_CREATE] "create",
1219 [REINT_LINK] "link",
1220 [REINT_UNLINK] "unlink",
1221 [REINT_RENAME] "rename",
1222 [REINT_OPEN] "open",
1225 int mds_handle(struct ptlrpc_request *req)
1227 int should_process, rc;
1228 struct mds_obd *mds = NULL; /* quell gcc overwarning */
1229 struct obd_device *mds_obd = NULL;
1232 rc = lustre_unpack_msg(req->rq_reqmsg, req->rq_reqlen);
1233 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_HANDLE_UNPACK)) {
1234 DEBUG_REQ(D_ERROR, req, "invalid request (%d)", rc);
1238 OBD_FAIL_RETURN(OBD_FAIL_MDS_ALL_REQUEST_NET | OBD_FAIL_ONCE, 0);
1240 LASSERT(!strcmp(req->rq_obd->obd_type->typ_name, LUSTRE_MDT_NAME));
1242 if (req->rq_reqmsg->opc != MDS_CONNECT) {
1243 struct mds_export_data *med;
1244 if (req->rq_export == NULL) {
1245 req->rq_status = -ENOTCONN;
1246 GOTO(out, rc = -ENOTCONN);
1249 med = &req->rq_export->exp_mds_data;
1250 mds_obd = req->rq_export->exp_obd;
1251 mds = &mds_obd->u.mds;
1252 spin_lock_bh(&mds->mds_processing_task_lock);
1253 if (mds_obd->obd_flags & OBD_ABORT_RECOVERY)
1254 mds_abort_recovery(mds);
1255 spin_unlock_bh(&mds->mds_processing_task_lock);
1257 if (mds_obd->obd_flags & OBD_RECOVERING) {
1258 rc = filter_recovery_request(req, mds, &should_process);
1259 if (rc || !should_process)
1261 } else if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
1262 if (req->rq_xid == med->med_last_xid) {
1263 DEBUG_REQ(D_HA, req, "resending reply");
1264 OBD_ALLOC(req->rq_repmsg, med->med_last_replen);
1265 req->rq_replen = med->med_last_replen;
1266 memcpy(req->rq_repmsg, med->med_last_reply,
1268 ptlrpc_reply(req->rq_svc, req);
1271 DEBUG_REQ(D_HA, req, "no reply for resend, continuing");
1276 switch (req->rq_reqmsg->opc) {
1278 DEBUG_REQ(D_INODE, req, "connect");
1279 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
1280 rc = target_handle_connect(req);
1281 /* Make sure that last_rcvd is correct. */
1283 /* Now that we have an export, set mds. */
1284 mds = mds_req2mds(req);
1285 mds_fsync_super(mds->mds_sb);
1289 case MDS_DISCONNECT:
1290 DEBUG_REQ(D_INODE, req, "disconnect");
1291 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
1292 rc = target_handle_disconnect(req);
1293 /* Make sure that last_rcvd is correct. */
1295 mds_fsync_super(mds->mds_sb);
1296 req->rq_status = rc;
1300 DEBUG_REQ(D_INODE, req, "getstatus");
1301 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
1302 rc = mds_getstatus(req);
1305 case MDS_GETLOVINFO:
1306 DEBUG_REQ(D_INODE, req, "getlovinfo");
1307 rc = mds_getlovinfo(req);
1311 DEBUG_REQ(D_INODE, req, "getattr");
1312 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
1313 rc = mds_getattr(0, req);
1316 case MDS_GETATTR_NAME: {
1317 struct lustre_handle lockh;
1318 DEBUG_REQ(D_INODE, req, "getattr_name");
1319 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NAME_NET, 0);
1320 rc = mds_getattr_name(0, req, &lockh);
1322 ldlm_lock_decref(&lockh, LCK_PR);
1326 DEBUG_REQ(D_INODE, req, "statfs");
1327 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
1328 rc = mds_statfs(req);
1332 DEBUG_REQ(D_INODE, req, "readpage\n");
1333 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
1334 rc = mds_readpage(req);
1336 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE))
1341 int opc = *(u32 *)lustre_msg_buf(req->rq_reqmsg, 0);
1342 int size[2] = {sizeof(struct mds_body), mds->mds_max_mdsize};
1345 DEBUG_REQ(D_INODE, req, "reint (%s%s)",
1346 reint_names[opc & REINT_OPCODE_MASK],
1347 opc & REINT_REPLAYING ? "|REPLAYING" : "");
1349 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
1351 if (opc == REINT_UNLINK)
1356 rc = lustre_pack_msg(bufcount, size, NULL,
1357 &req->rq_replen, &req->rq_repmsg);
1361 rc = mds_reint(req, 0, NULL);
1362 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET_REP, 0);
1367 DEBUG_REQ(D_INODE, req, "close");
1368 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
1369 rc = mds_close(req);
1373 DEBUG_REQ(D_INODE, req, "enqueue");
1374 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
1375 rc = ldlm_handle_enqueue(req, ldlm_server_completion_ast,
1376 ldlm_server_blocking_ast);
1379 DEBUG_REQ(D_INODE, req, "convert");
1380 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
1381 rc = ldlm_handle_convert(req);
1383 case LDLM_BL_CALLBACK:
1384 case LDLM_CP_CALLBACK:
1385 DEBUG_REQ(D_INODE, req, "callback");
1386 CERROR("callbacks should not happen on MDS\n");
1388 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
1391 rc = ptlrpc_error(req->rq_svc, req);
1397 /* If we're DISCONNECTing, the mds_export_data is already freed */
1398 if (!rc && req->rq_reqmsg->opc != MDS_DISCONNECT) {
1399 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1400 struct obd_device *obd = list_entry(mds, struct obd_device,
1402 req->rq_repmsg->last_xid =
1403 HTON__u64(le64_to_cpu(med->med_mcd->mcd_last_xid));
1404 if ((obd->obd_flags & OBD_NO_TRANSNO) == 0) {
1405 req->rq_repmsg->last_committed =
1406 HTON__u64(obd->obd_last_committed);
1408 DEBUG_REQ(D_IOCTL, req,
1409 "not sending last_committed update");
1411 CDEBUG(D_INFO, "last_transno %Lu, last_committed %Lu, xid %d\n",
1412 (unsigned long long)mds->mds_last_rcvd,
1413 (unsigned long long)obd->obd_last_committed,
1414 cpu_to_le32(req->rq_xid));
1418 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LAST_REPLAY) {
1419 if (mds_obd && (mds_obd->obd_flags & OBD_RECOVERING)) {
1420 DEBUG_REQ(D_HA, req, "LAST_REPLAY, queuing reply");
1421 return mds_queue_final_reply(req, rc);
1423 /* Lost a race with recovery; let the error path DTRT. */
1424 rc = req->rq_status = -ENOTCONN;
1427 if (req->rq_export && mds_obd &&
1428 (mds_obd->obd_flags & OBD_RECOVERING) == 0) {
1429 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1430 if (med->med_last_reply)
1431 OBD_FREE(med->med_last_reply, med->med_last_replen);
1432 OBD_ALLOC(med->med_last_reply, req->rq_replen);
1433 med->med_last_replen = req->rq_replen;
1434 med->med_last_xid = req->rq_xid;
1435 memcpy(med->med_last_reply, req->rq_repmsg, req->rq_replen);
1439 if (!OBD_FAIL_CHECK(OBD_FAIL_MDS_ALL_REPLY_NET | OBD_FAIL_ONCE)) {
1441 DEBUG_REQ(D_ERROR, req, "processing error (%d)", rc);
1442 ptlrpc_error(req->rq_svc, req);
1444 DEBUG_REQ(D_NET, req, "sending reply");
1445 ptlrpc_reply(req->rq_svc, req);
1448 obd_fail_loc |= OBD_FAIL_ONCE | OBD_FAILED;
1449 DEBUG_REQ(D_ERROR, req, "dropping reply");
1451 OBD_FREE(req->rq_repmsg, req->rq_replen);
1457 /* Update the server data on disk. This stores the new mount_count and
1458 * also the last_rcvd value to disk. If we don't have a clean shutdown,
1459 * then the server last_rcvd value may be less than that of the clients.
1460 * This will alert us that we may need to do client recovery.
1462 * Also assumes for mds_last_rcvd that we are not modifying it (no locking).
1464 int mds_update_server_data(struct mds_obd *mds)
1466 struct mds_server_data *msd = mds->mds_server_data;
1467 struct file *filp = mds->mds_rcvd_filp;
1468 struct obd_run_ctxt saved;
1472 push_ctxt(&saved, &mds->mds_ctxt, NULL);
1473 msd->msd_last_rcvd = cpu_to_le64(mds->mds_last_rcvd);
1474 msd->msd_mount_count = cpu_to_le64(mds->mds_mount_count);
1476 CDEBUG(D_SUPER, "MDS mount_count is %Lu, last_rcvd is %Lu\n",
1477 (unsigned long long)mds->mds_mount_count,
1478 (unsigned long long)mds->mds_last_rcvd);
1479 rc = lustre_fwrite(filp, (char *)msd, sizeof(*msd), &off);
1480 if (rc != sizeof(*msd)) {
1481 CERROR("error writing MDS server data: rc = %d\n", rc);
1486 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
1487 rc = fsync_dev(filp->f_dentry->d_inode->i_rdev);
1489 rc = file_fsync(filp, filp->f_dentry, 1);
1492 CERROR("error flushing MDS server data: rc = %d\n", rc);
1495 pop_ctxt(&saved, &mds->mds_ctxt, NULL);
1499 /* mount the file system (secretly) */
1500 static int mds_setup(struct obd_device *obddev, obd_count len, void *buf)
1502 struct obd_ioctl_data* data = buf;
1503 struct mds_obd *mds = &obddev->u.mds;
1504 struct vfsmount *mnt;
1508 #ifdef CONFIG_DEV_RDONLY
1509 dev_clear_rdonly(2);
1511 if (!data->ioc_inlbuf1 || !data->ioc_inlbuf2)
1512 RETURN(rc = -EINVAL);
1514 obddev->obd_fsops = fsfilt_get_ops(data->ioc_inlbuf2);
1515 if (IS_ERR(obddev->obd_fsops))
1516 RETURN(rc = PTR_ERR(obddev->obd_fsops));
1518 mnt = do_kern_mount(data->ioc_inlbuf2, 0, data->ioc_inlbuf1, NULL);
1521 CERROR("do_kern_mount failed: rc = %d\n", rc);
1525 CDEBUG(D_SUPER, "%s: mnt = %p\n", data->ioc_inlbuf1, mnt);
1526 mds->mds_sb = mnt->mnt_root->d_inode->i_sb;
1528 GOTO(err_put, rc = -ENODEV);
1530 init_MUTEX(&mds->mds_transno_sem);
1531 mds->mds_max_mdsize = sizeof(struct lov_mds_md);
1532 rc = mds_fs_setup(obddev, mnt);
1534 CERROR("MDS filesystem method init failed: rc = %d\n", rc);
1538 if (obddev->obd_flags & OBD_RECOVERING)
1539 start_recovery_timer(mds);
1541 obddev->obd_namespace =
1542 ldlm_namespace_new("mds_server", LDLM_NAMESPACE_SERVER);
1543 if (obddev->obd_namespace == NULL) {
1544 mds_cleanup(obddev);
1545 GOTO(err_fs, rc = -ENOMEM);
1548 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
1549 "mds_ldlm_client", &obddev->obd_ldlm_client);
1551 spin_lock_init(&mds->mds_processing_task_lock);
1552 mds->mds_processing_task = 0;
1553 mds->mds_has_lov_desc = 0;
1554 INIT_LIST_HEAD(&mds->mds_recovery_queue);
1555 INIT_LIST_HEAD(&mds->mds_delayed_reply_queue);
1556 init_waitqueue_head(&mds->mds_next_transno_waitq);
1561 mds_fs_cleanup(obddev);
1564 mntput(mds->mds_vfsmnt);
1568 fsfilt_put_ops(obddev->obd_fsops);
1572 static int mds_cleanup(struct obd_device *obddev)
1574 struct super_block *sb;
1575 struct mds_obd *mds = &obddev->u.mds;
1582 mds_update_server_data(mds);
1583 mds_fs_cleanup(obddev);
1586 mntput(mds->mds_vfsmnt);
1589 ldlm_namespace_free(obddev->obd_namespace);
1592 #ifdef CONFIG_DEV_RDONLY
1593 dev_clear_rdonly(2);
1595 fsfilt_put_ops(obddev->obd_fsops);
1600 static int ldlm_intent_policy(struct ldlm_namespace *ns,
1601 struct ldlm_lock **lockp, void *req_cookie,
1602 ldlm_mode_t mode, int flags, void *data)
1604 struct ptlrpc_request *req = req_cookie;
1605 struct ldlm_lock *lock = *lockp;
1612 if (req->rq_reqmsg->bufcount > 1) {
1613 /* an intent needs to be considered */
1614 struct ldlm_intent *it = lustre_msg_buf(req->rq_reqmsg, 1);
1615 struct mds_obd *mds = &req->rq_export->exp_obd->u.mds;
1616 struct mds_body *mds_body;
1617 struct ldlm_reply *rep;
1618 struct lustre_handle lockh;
1619 struct ldlm_lock *new_lock;
1620 int rc, offset = 2, repsize[3] = {sizeof(struct ldlm_reply),
1621 sizeof(struct mds_body),
1622 mds->mds_max_mdsize};
1624 it->opc = NTOH__u64(it->opc);
1626 LDLM_DEBUG(lock, "intent policy, opc: %s",
1627 ldlm_it2str(it->opc));
1629 rc = lustre_pack_msg(3, repsize, NULL, &req->rq_replen,
1632 rc = req->rq_status = -ENOMEM;
1636 rep = lustre_msg_buf(req->rq_repmsg, 0);
1637 rep->lock_policy_res1 = IT_INTENT_EXEC;
1639 /* execute policy */
1640 switch ((long)it->opc) {
1642 case IT_CREAT|IT_OPEN:
1643 rc = mds_reint(req, offset, &lockh);
1644 /* We return a dentry to the client if IT_OPEN_POS is
1645 * set, or if we make it to the OPEN portion of the
1646 * programme (which implies that we created) */
1647 if (!(rep->lock_policy_res1 & IT_OPEN_POS ||
1648 rep->lock_policy_res1 & IT_OPEN_OPEN)) {
1649 rep->lock_policy_res2 = rc;
1650 RETURN(ELDLM_LOCK_ABORTED);
1654 rc = mds_reint(req, offset, &lockh);
1655 /* Don't return a lock if the unlink failed, or if we're
1656 * not sending back an EA */
1658 rep->lock_policy_res2 = rc;
1659 RETURN(ELDLM_LOCK_ABORTED);
1661 if (req->rq_status != 0) {
1662 rep->lock_policy_res2 = req->rq_status;
1663 RETURN(ELDLM_LOCK_ABORTED);
1665 mds_body = lustre_msg_buf(req->rq_repmsg, 1);
1666 if (!(mds_body->valid & OBD_MD_FLEASIZE)) {
1667 rep->lock_policy_res2 = rc;
1668 RETURN(ELDLM_LOCK_ABORTED);
1675 rc = mds_getattr_name(offset, req, &lockh);
1676 /* FIXME: we need to sit down and decide on who should
1677 * set req->rq_status, who should return negative and
1678 * positive return values, and what they all mean. */
1680 rep->lock_policy_res2 = rc;
1681 RETURN(ELDLM_LOCK_ABORTED);
1683 if (req->rq_status != 0) {
1684 rep->lock_policy_res2 = req->rq_status;
1685 RETURN(ELDLM_LOCK_ABORTED);
1689 CERROR("Unhandled intent "LPD64"\n", it->opc);
1693 if (flags & LDLM_FL_INTENT_ONLY) {
1694 LDLM_DEBUG0(lock, "INTENT_ONLY, aborting lock");
1695 RETURN(ELDLM_LOCK_ABORTED);
1698 /* By this point, whatever function we called above must have
1699 * filled in 'lockh' or returned an error. We want to give the
1700 * new lock to the client instead of whatever lock it was about
1702 new_lock = ldlm_handle2lock(&lockh);
1703 LASSERT(new_lock != NULL);
1704 mds_body = lustre_msg_buf(req->rq_repmsg, 1);
1707 /* Fixup the lock to be given to the client */
1708 l_lock(&new_lock->l_resource->lr_namespace->ns_lock);
1709 LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
1710 new_lock->l_readers = 0;
1711 new_lock->l_writers = 0;
1713 new_lock->l_export = req->rq_export;
1714 list_add(&new_lock->l_export_chain,
1715 &new_lock->l_export->exp_ldlm_data.led_held_locks);
1717 /* We don't need to worry about completion_ast (which isn't set
1718 * in 'lock' yet anyways), because this lock is already
1720 new_lock->l_blocking_ast = lock->l_blocking_ast;
1722 memcpy(&new_lock->l_remote_handle, &lock->l_remote_handle,
1723 sizeof(lock->l_remote_handle));
1725 new_lock->l_flags &= ~(LDLM_FL_LOCAL | LDLM_FL_AST_SENT |
1728 LDLM_LOCK_PUT(new_lock);
1729 l_unlock(&new_lock->l_resource->lr_namespace->ns_lock);
1731 rep->lock_policy_res2 = req->rq_status;
1733 RETURN(ELDLM_LOCK_REPLACED);
1735 int size = sizeof(struct ldlm_reply);
1736 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
1746 int mds_attach(struct obd_device *dev, obd_count len, void *data)
1748 struct lprocfs_static_vars lvars;
1750 lprocfs_init_multi_vars(0, &lvars);
1751 return lprocfs_obd_attach(dev, lvars.obd_vars);
1754 int mds_detach(struct obd_device *dev)
1756 return lprocfs_obd_detach(dev);
1759 int mdt_attach(struct obd_device *dev, obd_count len, void *data)
1761 struct lprocfs_static_vars lvars;
1763 lprocfs_init_multi_vars(1, &lvars);
1764 return lprocfs_obd_attach(dev, lvars.obd_vars);
1767 int mdt_detach(struct obd_device *dev)
1769 return lprocfs_obd_detach(dev);
1772 static int mdt_setup(struct obd_device *obddev, obd_count len, void *buf)
1774 struct mds_obd *mds = &obddev->u.mds;
1775 struct obd_uuid uuid = { "self" };
1779 mds->mds_service = ptlrpc_init_svc(MDS_NEVENTS, MDS_NBUFS,
1780 MDS_BUFSIZE, MDS_MAXREQSIZE,
1781 MDS_REQUEST_PORTAL, MDC_REPLY_PORTAL,
1782 &uuid, mds_handle, "mds");
1783 if (!mds->mds_service) {
1784 CERROR("failed to start service\n");
1785 RETURN(rc = -ENOMEM);
1788 for (i = 0; i < MDT_NUM_THREADS; i++) {
1790 sprintf(name, "ll_mdt_%02d", i);
1791 rc = ptlrpc_start_thread(obddev, mds->mds_service, name);
1793 CERROR("cannot start MDT thread #%d: rc %d\n", i, rc);
1794 GOTO(err_thread, rc);
1798 mds->mds_getattr_service =
1799 ptlrpc_init_svc(MDS_NEVENTS, MDS_NBUFS,
1800 MDS_BUFSIZE, MDS_MAXREQSIZE,
1801 MDS_GETATTR_PORTAL, MDC_REPLY_PORTAL,
1802 &uuid, mds_handle, "mds");
1803 if (!mds->mds_getattr_service) {
1804 CERROR("failed to start getattr service\n");
1805 GOTO(err_thread, rc = -ENOMEM);
1808 for (i = 0; i < MDT_NUM_THREADS; i++) {
1810 sprintf(name, "ll_mdt_attr_%02d", i);
1811 rc = ptlrpc_start_thread(obddev, mds->mds_getattr_service,
1814 CERROR("cannot start MDT getattr thread #%d: rc %d\n",
1816 GOTO(err_thread2, rc);
1823 ptlrpc_stop_all_threads(mds->mds_getattr_service);
1824 ptlrpc_unregister_service(mds->mds_getattr_service);
1826 ptlrpc_stop_all_threads(mds->mds_service);
1827 ptlrpc_unregister_service(mds->mds_service);
1832 static int mdt_cleanup(struct obd_device *obddev)
1834 struct mds_obd *mds = &obddev->u.mds;
1837 ptlrpc_stop_all_threads(mds->mds_getattr_service);
1838 ptlrpc_unregister_service(mds->mds_getattr_service);
1840 ptlrpc_stop_all_threads(mds->mds_service);
1841 ptlrpc_unregister_service(mds->mds_service);
1846 extern int mds_iocontrol(unsigned int cmd, struct lustre_handle *conn,
1847 int len, void *karg, void *uarg);
1849 /* use obd ops to offer management infrastructure */
1850 static struct obd_ops mds_obd_ops = {
1851 o_owner: THIS_MODULE,
1852 o_attach: mds_attach,
1853 o_detach: mds_detach,
1854 o_connect: mds_connect,
1855 o_disconnect: mds_disconnect,
1857 o_cleanup: mds_cleanup,
1858 o_iocontrol: mds_iocontrol
1861 static struct obd_ops mdt_obd_ops = {
1862 o_owner: THIS_MODULE,
1863 o_attach: mdt_attach,
1864 o_detach: mdt_detach,
1866 o_cleanup: mdt_cleanup,
1870 static int __init mds_init(void)
1872 struct lprocfs_static_vars lvars;
1873 mds_file_cache = kmem_cache_create("ll_mds_file_data",
1874 sizeof(struct mds_file_data),
1876 if (mds_file_cache == NULL)
1879 lprocfs_init_multi_vars(0, &lvars);
1880 class_register_type(&mds_obd_ops, lvars.module_vars, LUSTRE_MDS_NAME);
1881 lprocfs_init_multi_vars(1, &lvars);
1882 class_register_type(&mdt_obd_ops, lvars.module_vars, LUSTRE_MDT_NAME);
1883 ldlm_register_intent(ldlm_intent_policy);
1888 static void __exit mds_exit(void)
1890 ldlm_unregister_intent();
1891 class_unregister_type(LUSTRE_MDS_NAME);
1892 class_unregister_type(LUSTRE_MDT_NAME);
1893 if (kmem_cache_destroy(mds_file_cache))
1894 CERROR("couldn't free MDS file cache\n");
1897 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
1898 MODULE_DESCRIPTION("Lustre Metadata Server (MDS)");
1899 MODULE_LICENSE("GPL");
1901 module_init(mds_init);
1902 module_exit(mds_exit);