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, 2002 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>
12 * This file is part of Lustre, http://www.lustre.org.
14 * Lustre is free software; you can redistribute it and/or
15 * modify it under the terms of version 2 of the GNU General Public
16 * License as published by the Free Software Foundation.
18 * Lustre is distributed in the hope that it will be useful,
19 * but WITHOUT ANY WARRANTY; without even the implied warranty of
20 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
21 * GNU General Public License for more details.
23 * You should have received a copy of the GNU General Public License
24 * along with Lustre; if not, write to the Free Software
25 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
29 #define DEBUG_SUBSYSTEM S_MDS
31 #include <linux/module.h>
32 #include <linux/lustre_mds.h>
33 #include <linux/lustre_dlm.h>
34 #include <linux/init.h>
35 #include <linux/obd_class.h>
36 #include <linux/random.h>
38 static kmem_cache_t *mds_file_cache;
40 extern int mds_get_lovtgts(struct obd_device *obd, int tgt_count,
41 obd_uuid_t *uuidarray);
42 extern int mds_get_lovdesc(struct obd_device *obd, struct lov_desc *desc);
43 extern int mds_update_last_rcvd(struct mds_obd *mds, void *handle,
44 struct ptlrpc_request *req);
45 static int mds_cleanup(struct obd_device * obddev);
47 inline struct mds_obd *mds_req2mds(struct ptlrpc_request *req)
49 return &req->rq_export->exp_obd->u.mds;
52 static int mds_bulk_timeout(void *data)
54 struct ptlrpc_bulk_desc *desc = data;
57 CERROR("(not yet) starting recovery of client %p\n", desc->bd_client);
61 /* Assumes caller has already pushed into the kernel filesystem context */
62 static int mds_sendpage(struct ptlrpc_request *req, struct file *file,
66 struct mds_obd *mds = mds_req2mds(req);
67 struct ptlrpc_bulk_desc *desc;
68 struct ptlrpc_bulk_page *bulk;
69 struct l_wait_info lwi;
73 desc = ptlrpc_prep_bulk(req->rq_connection);
75 GOTO(out, rc = -ENOMEM);
77 bulk = ptlrpc_prep_bulk_page(desc);
79 GOTO(cleanup_bulk, rc = -ENOMEM);
81 OBD_ALLOC(buf, PAGE_SIZE);
83 GOTO(cleanup_bulk, rc = -ENOMEM);
85 rc = mds_fs_readpage(mds, file, buf, PAGE_SIZE, (loff_t *)&offset);
88 GOTO(cleanup_buf, rc = -EIO);
90 bulk->bp_xid = req->rq_xid;
92 bulk->bp_buflen = PAGE_SIZE;
93 desc->bd_portal = MDS_BULK_PORTAL;
95 rc = ptlrpc_send_bulk(desc);
97 GOTO(cleanup_buf, rc);
99 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
100 CERROR("obd_fail_loc=%x, fail operation rc=%d\n",
101 OBD_FAIL_MDS_SENDPAGE, rc);
102 ptlrpc_abort_bulk(desc);
103 GOTO(cleanup_buf, rc);
106 lwi = LWI_TIMEOUT(obd_timeout * HZ, mds_bulk_timeout, desc);
107 rc = l_wait_event(desc->bd_waitq, desc->bd_flags & PTL_BULK_FL_SENT, &lwi);
109 if (rc != -ETIMEDOUT)
111 GOTO(cleanup_buf, rc);
116 OBD_FREE(buf, PAGE_SIZE);
118 ptlrpc_free_bulk(desc);
124 * Look up a named entry in a directory, and get an LDLM lock on it.
125 * 'dir' is a inode for which an LDLM lock has already been taken.
127 * If we do not need an exclusive or write lock on this entry (e.g.
128 * a read lock for attribute lookup only) then we do not hold the
129 * directory on return. It is up to the caller to know what type
130 * of lock it is getting, and clean up appropriately.
132 struct dentry *mds_name2locked_dentry(struct obd_device *obd,
133 struct dentry *dir, struct vfsmount **mnt,
134 char *name, int namelen, int lock_mode,
135 struct lustre_handle *lockh,
138 struct dentry *dchild;
140 __u64 res_id[3] = {0};
143 down(&dir->d_inode->i_sem);
144 dchild = lookup_one_len(name, dir, namelen);
145 if (IS_ERR(dchild)) {
146 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
147 up(&dir->d_inode->i_sem);
151 if (dir_lock_mode != LCK_EX && dir_lock_mode != LCK_PW) {
152 up(&dir->d_inode->i_sem);
153 ldlm_lock_decref(lockh, dir_lock_mode);
156 if (lock_mode == 0 || !dchild->d_inode)
159 res_id[0] = dchild->d_inode->i_ino;
160 rc = ldlm_match_or_enqueue(NULL, NULL, obd->obd_namespace, NULL,
161 res_id, LDLM_PLAIN, NULL, 0, lock_mode,
162 &flags, ldlm_completion_ast,
163 mds_blocking_ast, NULL, 0, lockh);
164 if (rc != ELDLM_OK) {
166 up(&dir->d_inode->i_sem);
167 RETURN(ERR_PTR(-ENOLCK)); /* XXX translate ldlm code */
173 struct dentry *mds_fid2locked_dentry(struct obd_device *obd, struct ll_fid *fid,
174 struct vfsmount **mnt, int lock_mode,
175 struct lustre_handle *lockh)
177 struct mds_obd *mds = &obd->u.mds;
178 struct dentry *de = mds_fid2dentry(mds, fid, mnt), *retval = de;
180 __u64 res_id[3] = {0};
186 res_id[0] = de->d_inode->i_ino;
187 rc = ldlm_match_or_enqueue(NULL, NULL, obd->obd_namespace, NULL,
188 res_id, LDLM_PLAIN, NULL, 0, lock_mode,
189 &flags, ldlm_completion_ast,
190 mds_blocking_ast, NULL, 0, lockh);
191 if (rc != ELDLM_OK) {
193 retval = ERR_PTR(-ENOLCK); /* XXX translate ldlm code */
199 /* Look up an entry by inode number. */
200 struct dentry *mds_fid2dentry(struct mds_obd *mds, struct ll_fid *fid,
201 struct vfsmount **mnt)
203 /* stolen from NFS */
204 struct super_block *sb = mds->mds_sb;
205 unsigned long ino = fid->id;
206 __u32 generation = fid->generation;
208 struct list_head *lp;
209 struct dentry *result;
212 RETURN(ERR_PTR(-ESTALE));
214 inode = iget(sb, ino);
216 RETURN(ERR_PTR(-ENOMEM));
218 CDEBUG(D_DENTRY, "--> mds_fid2dentry: sb %p\n", inode->i_sb);
220 if (is_bad_inode(inode) ||
221 (generation && inode->i_generation != generation)) {
222 /* we didn't find the right inode.. */
223 CERROR("bad inode %lu, link: %d ct: %d or version %u/%u\n",
224 inode->i_ino, inode->i_nlink,
225 atomic_read(&inode->i_count), inode->i_generation,
229 RETURN(ERR_PTR(-ESTALE));
232 /* now to find a dentry. If possible, get a well-connected one */
234 *mnt = mds->mds_vfsmnt;
235 spin_lock(&dcache_lock);
236 list_for_each(lp, &inode->i_dentry) {
237 result = list_entry(lp, struct dentry, d_alias);
238 if (!(result->d_flags & DCACHE_NFSD_DISCONNECTED)) {
240 result->d_vfs_flags |= DCACHE_REFERENCED;
241 spin_unlock(&dcache_lock);
248 spin_unlock(&dcache_lock);
249 result = d_alloc_root(inode);
250 if (result == NULL) {
252 return ERR_PTR(-ENOMEM);
256 result->d_flags |= DCACHE_NFSD_DISCONNECTED;
260 /* Establish a connection to the MDS.
262 * This will set up an export structure for the client to hold state data
263 * about that client, like open files, the last operation number it did
264 * on the server, etc.
266 static int mds_connect(struct lustre_handle *conn, struct obd_device *obd,
267 obd_uuid_t cluuid, struct recovd_obd *recovd,
268 ptlrpc_recovery_cb_t recover)
270 struct obd_export *exp;
271 struct mds_export_data *med;
272 struct mds_client_data *mcd;
277 if (!conn || !obd || !cluuid)
282 spin_lock(&obd->obd_dev_lock);
283 list_for_each(p, &obd->obd_exports) {
284 exp = list_entry(p, struct obd_export, exp_obd_chain);
285 mcd = exp->exp_mds_data.med_mcd;
286 if (!memcmp(cluuid, mcd->mcd_uuid, sizeof(mcd->mcd_uuid))) {
287 LASSERT(exp->exp_obd == obd);
289 if (!list_empty(&exp->exp_conn_chain)) {
290 CERROR("existing uuid/export, list not empty!\n");
291 spin_unlock(&obd->obd_dev_lock);
292 /* XXX should we MOD_DEC_USE_COUNT; here? */
295 conn->addr = (__u64) (unsigned long)exp;
296 conn->cookie = exp->exp_cookie;
297 spin_unlock(&obd->obd_dev_lock);
298 CDEBUG(D_INFO, "existing export for UUID '%s' at %p\n",
300 CDEBUG(D_IOCTL,"connect: addr %Lx cookie %Lx\n",
301 (long long)conn->addr, (long long)conn->cookie);
306 spin_unlock(&obd->obd_dev_lock);
307 /* XXX There is a small race between checking the list and adding a
308 * new connection for the same UUID, but the real threat (list
309 * corruption when multiple different clients connect) is solved.
311 rc = class_connect(conn, obd, cluuid);
314 exp = class_conn2export(conn);
316 med = &exp->exp_mds_data;
318 OBD_ALLOC(mcd, sizeof(*mcd));
320 CERROR("mds: out of memory for client data\n");
321 GOTO(out_export, rc = -ENOMEM);
324 memcpy(mcd->mcd_uuid, cluuid, sizeof(mcd->mcd_uuid));
327 INIT_LIST_HEAD(&med->med_open_head);
328 spin_lock_init(&med->med_open_lock);
330 rc = mds_client_add(med, -1);
337 OBD_FREE(mcd, sizeof(*mcd));
339 class_disconnect(conn);
346 static int mds_disconnect(struct lustre_handle *conn)
348 struct obd_export *export = class_conn2export(conn);
352 #warning "Mike: we need to close all files opened on med_open_head"
353 ldlm_cancel_locks_for_export(export);
354 mds_client_free(export);
356 rc = class_disconnect(conn);
363 static int mds_getstatus(struct ptlrpc_request *req)
365 struct mds_obd *mds = mds_req2mds(req);
366 struct mds_body *body;
367 int rc, size = sizeof(*body);
370 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
371 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK)) {
372 CERROR("mds: out of memory for message: size=%d\n", size);
373 req->rq_status = -ENOMEM;
377 /* Anything we need to do here with the client's trans no or so? */
378 body = lustre_msg_buf(req->rq_repmsg, 0);
379 memcpy(&body->fid1, &mds->mds_rootfid, sizeof(body->fid1));
381 /* the last_committed and last_xid fields are filled in for all
382 * replies already - no need to do so here also.
387 static int mds_getlovinfo(struct ptlrpc_request *req)
389 struct mds_obd *mds = mds_req2mds(req);
390 struct mds_status_req *streq;
391 struct lov_desc *desc;
393 int rc, size[2] = {sizeof(*desc)};
396 streq = lustre_msg_buf(req->rq_reqmsg, 0);
397 streq->flags = NTOH__u32(streq->flags);
398 streq->repbuf = NTOH__u32(streq->repbuf);
399 size[1] = streq->repbuf;
401 rc = lustre_pack_msg(2, size, NULL, &req->rq_replen, &req->rq_repmsg);
403 CERROR("mds: out of memory for message: size=%d\n", size[1]);
404 req->rq_status = -ENOMEM;
408 desc = lustre_msg_buf(req->rq_repmsg, 0);
409 rc = mds_get_lovdesc(req->rq_obd, desc);
411 CERROR("mds_get_lovdesc error %d", rc);
416 tgt_count = le32_to_cpu(desc->ld_tgt_count);
417 if (tgt_count * sizeof(obd_uuid_t) > streq->repbuf) {
418 CERROR("too many targets, enlarge client buffers\n");
419 req->rq_status = -ENOSPC;
423 mds->mds_max_mdsize = sizeof(struct lov_mds_md) +
424 tgt_count * sizeof(struct lov_object_id);
425 rc = mds_get_lovtgts(req->rq_obd, tgt_count,
426 lustre_msg_buf(req->rq_repmsg, 1));
428 CERROR("get_lovtgts error %d\n", rc);
435 int mds_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
436 void *data, __u32 data_len, int flag)
441 if (flag == LDLM_CB_CANCELING) {
442 /* Don't need to do anything here. */
446 l_lock(&lock->l_resource->lr_namespace->ns_lock);
447 lock->l_flags |= LDLM_FL_CBPENDING;
448 do_ast = (!lock->l_readers && !lock->l_writers);
449 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
452 struct lustre_handle lockh;
455 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
456 ldlm_lock2handle(lock, &lockh);
457 rc = ldlm_cli_cancel(&lockh);
459 CERROR("ldlm_cli_cancel: %d\n", rc);
461 LDLM_DEBUG(lock, "Lock still has references, will be"
466 static int mds_getattr_internal(struct mds_obd *mds, struct dentry *dentry,
467 struct ptlrpc_request *req,
468 struct mds_body *reqbody, int reply_off)
470 struct mds_body *body;
471 struct inode *inode = dentry->d_inode;
478 body = lustre_msg_buf(req->rq_repmsg, reply_off);
480 mds_pack_inode2fid(&body->fid1, inode);
481 mds_pack_inode2body(body, inode);
483 if (S_ISREG(inode->i_mode)) {
484 struct lov_mds_md *lmm;
486 lmm = lustre_msg_buf(req->rq_repmsg, reply_off + 1);
487 lmm->lmm_easize = mds->mds_max_mdsize;
488 rc = mds_fs_get_md(mds, inode, lmm);
493 CERROR("mds_fs_get_md failed: %d\n", rc);
496 body->valid |= OBD_MD_FLEASIZE;
497 } else if (S_ISLNK(inode->i_mode) && reqbody->valid & OBD_MD_LINKNAME) {
498 char *symname = lustre_msg_buf(req->rq_repmsg, reply_off + 1);
499 int len = req->rq_repmsg->buflens[reply_off + 1];
501 rc = inode->i_op->readlink(dentry, symname, len);
503 CERROR("readlink failed: %d\n", rc);
506 CDEBUG(D_INODE, "read symlink dest %s\n", symname);
508 body->valid |= OBD_MD_LINKNAME;
513 static int mds_getattr_name(int offset, struct ptlrpc_request *req)
515 struct mds_obd *mds = mds_req2mds(req);
516 struct obd_device *obd = req->rq_export->exp_obd;
517 struct obd_run_ctxt saved;
518 struct mds_body *body;
519 struct dentry *de = NULL, *dchild = NULL;
521 struct lustre_handle lockh;
523 int namelen, flags, lock_mode, rc = 0;
525 __u64 res_id[3] = {0, 0, 0};
528 LASSERT(!strcmp(req->rq_export->exp_obd->obd_type->typ_name, "mds"));
530 if (req->rq_reqmsg->bufcount <= offset + 1) {
532 GOTO(out_pre_de, rc = -EINVAL);
535 body = lustre_msg_buf(req->rq_reqmsg, offset);
536 name = lustre_msg_buf(req->rq_reqmsg, offset + 1);
537 namelen = req->rq_reqmsg->buflens[offset + 1];
538 /* requests were at offset 2, replies go back at 1 */
542 uc.ouc_fsuid = body->fsuid;
543 uc.ouc_fsgid = body->fsgid;
544 push_ctxt(&saved, &mds->mds_ctxt, &uc);
545 de = mds_fid2dentry(mds, &body->fid1, NULL);
548 GOTO(out_pre_de, rc = -ESTALE);
552 CDEBUG(D_INODE, "parent ino %ld, name %*s\n", dir->i_ino,namelen,name);
554 lock_mode = (req->rq_reqmsg->opc == MDS_REINT) ? LCK_CW : LCK_PW;
555 res_id[0] = dir->i_ino;
557 rc = ldlm_lock_match(obd->obd_namespace, res_id, LDLM_PLAIN,
558 NULL, 0, lock_mode, &lockh);
560 LDLM_DEBUG_NOLOCK("enqueue res "LPU64, res_id[0]);
561 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, NULL,
562 res_id, LDLM_PLAIN, NULL, 0, lock_mode,
563 &flags, ldlm_completion_ast,
564 mds_blocking_ast, NULL, 0, &lockh);
565 if (rc != ELDLM_OK) {
566 CERROR("lock enqueue: err: %d\n", rc);
567 GOTO(out_create_de, rc = -EIO);
570 ldlm_lock_dump((void *)(unsigned long)lockh.addr);
573 dchild = lookup_one_len(name, de, namelen - 1);
574 if (IS_ERR(dchild)) {
575 CDEBUG(D_INODE, "child lookup error %ld\n", PTR_ERR(dchild));
578 GOTO(out_create_dchild, rc = -ESTALE);
581 rc = mds_getattr_internal(mds, dchild, req, body, offset);
587 ldlm_lock_decref(&lockh, lock_mode);
596 static int mds_getattr(int offset, struct ptlrpc_request *req)
598 struct mds_obd *mds = mds_req2mds(req);
599 struct obd_run_ctxt saved;
602 struct mds_body *body;
604 int rc = 0, size[2] = {sizeof(*body)}, bufcount = 1;
607 body = lustre_msg_buf(req->rq_reqmsg, offset);
608 uc.ouc_fsuid = body->fsuid;
609 uc.ouc_fsgid = body->fsgid;
610 push_ctxt(&saved, &mds->mds_ctxt, &uc);
611 de = mds_fid2dentry(mds, &body->fid1, NULL);
613 rc = req->rq_status = -ENOENT;
614 GOTO(out_pop, PTR_ERR(de));
618 if (S_ISREG(body->fid1.f_type)) {
620 size[1] = mds->mds_max_mdsize;
621 } else if (body->valid & OBD_MD_LINKNAME) {
623 size[1] = MIN(inode->i_size + 1, body->size);
624 CDEBUG(D_INODE, "symlink size: %d, reply space: %d\n",
625 inode->i_size + 1, body->size);
628 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
629 CERROR("failed GETATTR_PACK test\n");
630 req->rq_status = -ENOMEM;
631 GOTO(out, rc = -ENOMEM);
634 rc = lustre_pack_msg(bufcount, size, NULL, &req->rq_replen,
637 CERROR("out of memory or FAIL_MDS_GETATTR_PACK\n");
642 req->rq_status = mds_getattr_internal(mds, de, req, body, 0);
651 static int mds_statfs(struct ptlrpc_request *req)
653 struct mds_obd *mds = mds_req2mds(req);
654 struct obd_statfs *osfs;
656 int rc, size = sizeof(*osfs);
659 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
661 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
662 CERROR("mds: statfs lustre_pack_msg failed: rc = %d\n", rc);
666 rc = vfs_statfs(mds->mds_sb, &sfs);
668 CERROR("mds: statfs failed: rc %d\n", rc);
671 osfs = lustre_msg_buf(req->rq_repmsg, 0);
672 memset(osfs, 0, size);
673 statfs_pack(osfs, &sfs);
674 obd_statfs_pack(osfs, osfs);
681 static struct mds_file_data *mds_handle2mfd(struct lustre_handle *handle)
683 struct mds_file_data *mfd = NULL;
685 if (!handle || !handle->addr)
688 mfd = (struct mds_file_data *)(unsigned long)(handle->addr);
689 if (!kmem_cache_validate(mds_file_cache, mfd))
692 if (mfd->mfd_servercookie != handle->cookie)
698 static int mds_store_ea(struct mds_obd *mds, struct ptlrpc_request *req,
699 struct mds_body *body, struct dentry *de,
700 struct lov_mds_md *lmm)
702 struct obd_run_ctxt saved;
707 uc.ouc_fsuid = body->fsuid;
708 uc.ouc_fsgid = body->fsgid;
709 push_ctxt(&saved, &mds->mds_ctxt, &uc);
710 handle = mds_fs_start(mds, de->d_inode, MDS_FSOP_SETATTR);
712 GOTO(out_ea, rc = -ENOMEM);
714 rc = mds_fs_set_md(mds, de->d_inode, handle, lmm);
716 rc = mds_update_last_rcvd(mds, handle, req);
718 rc2 = mds_fs_commit(mds, de->d_inode, handle);
727 static int mds_open(struct ptlrpc_request *req)
729 struct mds_obd *mds = mds_req2mds(req);
730 struct mds_body *body;
731 struct mds_export_data *med;
732 struct mds_file_data *mfd;
735 struct vfsmount *mnt;
737 struct list_head *tmp;
738 int rc, size = sizeof(*body);
741 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
742 CERROR("test case OBD_FAIL_MDS_OPEN_PACK\n");
743 req->rq_status = -ENOMEM;
747 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
749 CERROR("mds: pack error: rc = %d\n", rc);
754 body = lustre_msg_buf(req->rq_reqmsg, 0);
756 /* was this animal open already and the client lost the reply? */
757 /* XXX need some way to detect a reopen, to avoid locked list walks */
758 med = &req->rq_export->exp_mds_data;
759 spin_lock(&med->med_open_lock);
760 list_for_each(tmp, &med->med_open_head) {
761 mfd = list_entry(tmp, typeof(*mfd), mfd_list);
762 if (!memcmp(&mfd->mfd_clienthandle, &body->handle,
763 sizeof(mfd->mfd_clienthandle)) &&
764 body->fid1.id == mfd->mfd_file->f_dentry->d_inode->i_ino) {
765 de = mfd->mfd_file->f_dentry;
766 spin_unlock(&med->med_open_lock);
767 CERROR("Re opening "LPD64"\n", body->fid1.id);
768 GOTO(out_pack, rc = 0);
771 spin_unlock(&med->med_open_lock);
773 mfd = kmem_cache_alloc(mds_file_cache, GFP_KERNEL);
775 CERROR("mds: out of memory\n");
776 req->rq_status = -ENOMEM;
780 de = mds_fid2dentry(mds, &body->fid1, &mnt);
782 GOTO(out_free, rc = PTR_ERR(de));
784 /* check if this inode has seen a delayed object creation */
785 if (req->rq_reqmsg->bufcount > 1) {
786 struct lov_mds_md *lmm = lustre_msg_buf(req->rq_reqmsg, 1);
788 rc = mds_store_ea(mds, req, body, de, lmm);
797 /* dentry_open does a dput(de) and mntput(mnt) on error */
798 file = dentry_open(de, mnt, flags & ~O_DIRECT);
804 file->private_data = mfd;
805 mfd->mfd_file = file;
806 memcpy(&mfd->mfd_clienthandle, &body->handle, sizeof(body->handle));
807 get_random_bytes(&mfd->mfd_servercookie, sizeof(mfd->mfd_servercookie));
808 spin_lock(&med->med_open_lock);
809 list_add(&mfd->mfd_list, &med->med_open_head);
810 spin_unlock(&med->med_open_lock);
813 body = lustre_msg_buf(req->rq_repmsg, 0);
814 mds_pack_inode2fid(&body->fid1, de->d_inode);
815 mds_pack_inode2body(body, de->d_inode);
816 body->handle.addr = (__u64)(unsigned long)mfd;
817 body->handle.cookie = mfd->mfd_servercookie;
818 CDEBUG(D_INODE, "llite file "LPX64": addr %p, cookie "LPX64"\n",
819 mfd->mfd_clienthandle.addr, mfd, mfd->mfd_servercookie);
823 mfd->mfd_servercookie = DEAD_HANDLE_MAGIC;
824 kmem_cache_free(mds_file_cache, mfd);
829 static int mds_close(struct ptlrpc_request *req)
831 struct mds_export_data *med = &req->rq_export->exp_mds_data;
832 struct mds_body *body;
834 struct mds_file_data *mfd;
838 body = lustre_msg_buf(req->rq_reqmsg, 0);
840 mfd = mds_handle2mfd(&body->handle);
842 CERROR("no handle for file close "LPD64
843 ": addr "LPX64", cookie "LPX64"\n",
844 body->fid1.id, body->handle.addr, body->handle.cookie);
848 file = mfd->mfd_file;
849 LASSERT(file->private_data == mfd);
851 spin_lock(&med->med_open_lock);
852 list_del(&mfd->mfd_list);
853 spin_unlock(&med->med_open_lock);
854 mfd->mfd_servercookie = DEAD_HANDLE_MAGIC;
855 kmem_cache_free(mds_file_cache, mfd);
857 req->rq_status = filp_close(file, 0);
859 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
860 CERROR("test case OBD_FAIL_MDS_CLOSE_PACK\n");
861 req->rq_status = -ENOMEM;
865 rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen, &req->rq_repmsg);
867 CERROR("mds: lustre_pack_msg: rc = %d\n", rc);
874 static int mds_readpage(struct ptlrpc_request *req)
876 struct mds_obd *mds = mds_req2mds(req);
877 struct vfsmount *mnt;
880 struct mds_body *body, *repbody;
881 struct obd_run_ctxt saved;
882 int rc, size = sizeof(*body);
886 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
887 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK)) {
888 CERROR("mds: out of memory\n");
889 GOTO(out, rc = -ENOMEM);
892 body = lustre_msg_buf(req->rq_reqmsg, 0);
893 uc.ouc_fsuid = body->fsuid;
894 uc.ouc_fsgid = body->fsgid;
895 push_ctxt(&saved, &mds->mds_ctxt, &uc);
896 de = mds_fid2dentry(mds, &body->fid1, &mnt);
898 GOTO(out_pop, rc = PTR_ERR(de));
900 CDEBUG(D_INODE, "ino %ld\n", de->d_inode->i_ino);
902 file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
903 /* note: in case of an error, dentry_open puts dentry */
905 GOTO(out_pop, rc = PTR_ERR(file));
907 repbody = lustre_msg_buf(req->rq_repmsg, 0);
908 repbody->size = file->f_dentry->d_inode->i_size;
909 repbody->valid = OBD_MD_FLSIZE;
911 /* to make this asynchronous make sure that the handling function
912 doesn't send a reply when this function completes. Instead a
913 callback function would send the reply */
914 /* note: in case of an error, dentry_open puts dentry */
915 rc = mds_sendpage(req, file, body->size);
925 int mds_reint(int offset, struct ptlrpc_request *req)
928 struct mds_update_record rec;
930 rc = mds_update_unpack(req, offset, &rec);
931 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
932 CERROR("invalid record\n");
933 req->rq_status = -EINVAL;
936 /* rc will be used to interrupt a for loop over multiple records */
937 rc = mds_reint_rec(&rec, offset, req);
941 int mds_handle(struct ptlrpc_request *req)
946 rc = lustre_unpack_msg(req->rq_reqmsg, req->rq_reqlen);
947 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_HANDLE_UNPACK)) {
948 CERROR("lustre_mds: Invalid request\n");
952 if (req->rq_reqmsg->opc != MDS_CONNECT && req->rq_export == NULL)
953 GOTO(out, rc = -ENOTCONN);
955 LASSERT(!strcmp(req->rq_obd->obd_type->typ_name, "mds"));
957 switch (req->rq_reqmsg->opc) {
959 CDEBUG(D_INODE, "connect\n");
960 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
961 rc = target_handle_connect(req);
965 CDEBUG(D_INODE, "disconnect\n");
966 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
967 rc = target_handle_disconnect(req);
971 CDEBUG(D_INODE, "getstatus\n");
972 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
973 rc = mds_getstatus(req);
977 CDEBUG(D_INODE, "getlovinfo\n");
978 rc = mds_getlovinfo(req);
982 CDEBUG(D_INODE, "getattr\n");
983 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
984 rc = mds_getattr(0, req);
988 CDEBUG(D_INODE, "statfs\n");
989 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
990 rc = mds_statfs(req);
994 CDEBUG(D_INODE, "readpage\n");
995 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
996 rc = mds_readpage(req);
998 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE))
1003 int size = sizeof(struct mds_body);
1004 CDEBUG(D_INODE, "reint\n");
1005 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
1007 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
1010 req->rq_status = rc;
1013 rc = mds_reint(0, req);
1014 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET_REP, 0);
1019 CDEBUG(D_INODE, "open\n");
1020 OBD_FAIL_RETURN(OBD_FAIL_MDS_OPEN_NET, 0);
1025 CDEBUG(D_INODE, "close\n");
1026 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
1027 rc = mds_close(req);
1031 CDEBUG(D_INODE, "enqueue\n");
1032 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
1033 rc = ldlm_handle_enqueue(req);
1039 CDEBUG(D_INODE, "convert\n");
1040 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
1041 rc = ldlm_handle_convert(req);
1047 CDEBUG(D_INODE, "cancel\n");
1048 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CANCEL, 0);
1049 rc = ldlm_handle_cancel(req);
1053 case LDLM_BL_CALLBACK:
1054 case LDLM_CP_CALLBACK:
1055 CDEBUG(D_INODE, "callback\n");
1056 CERROR("callbacks should not happen on MDS\n");
1058 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
1061 rc = ptlrpc_error(req->rq_svc, req);
1068 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1069 struct mds_obd *mds = mds_req2mds(req);
1071 req->rq_repmsg->last_xid =
1072 HTON__u64(le64_to_cpu(med->med_mcd->mcd_last_xid));
1073 req->rq_repmsg->last_committed =
1074 HTON__u64(mds->mds_last_committed);
1075 CDEBUG(D_INFO, "last_rcvd ~%Lu, last_committed %Lu, xid %d\n",
1076 (unsigned long long)mds->mds_last_rcvd,
1077 (unsigned long long)mds->mds_last_committed,
1078 cpu_to_le32(req->rq_xid));
1082 CERROR("mds: processing error (opcode %d): %d\n",
1083 req->rq_reqmsg->opc, rc);
1084 ptlrpc_error(req->rq_svc, req);
1086 CDEBUG(D_NET, "sending reply\n");
1087 ptlrpc_reply(req->rq_svc, req);
1092 /* Update the server data on disk. This stores the new mount_count and
1093 * also the last_rcvd value to disk. If we don't have a clean shutdown,
1094 * then the server last_rcvd value may be less than that of the clients.
1095 * This will alert us that we may need to do client recovery.
1097 * Assumes we are already in the server filesystem context.
1099 * Also assumes for mds_last_rcvd that we are not modifying it (no locking).
1102 int mds_update_server_data(struct mds_obd *mds)
1104 struct mds_server_data *msd = mds->mds_server_data;
1105 struct file *filp = mds->mds_rcvd_filp;
1109 msd->msd_last_rcvd = cpu_to_le64(mds->mds_last_rcvd);
1110 msd->msd_mount_count = cpu_to_le64(mds->mds_mount_count);
1112 CDEBUG(D_SUPER, "MDS mount_count is %Lu, last_rcvd is %Lu\n",
1113 (unsigned long long)mds->mds_mount_count,
1114 (unsigned long long)mds->mds_last_rcvd);
1115 rc = lustre_fwrite(filp, (char *)msd, sizeof(*msd), &off);
1116 if (rc != sizeof(*msd)) {
1117 CERROR("error writing MDS server data: rc = %d\n", rc);
1122 rc = fsync_dev(filp->f_dentry->d_inode->i_rdev);
1124 CERROR("error flushing MDS server data: rc = %d\n", rc);
1129 /* Do recovery actions for the MDS */
1130 static int mds_recover(struct obd_device *obddev)
1132 struct mds_obd *mds = &obddev->u.mds;
1133 struct obd_run_ctxt saved;
1136 /* This happens at the end when recovery is complete */
1137 ++mds->mds_mount_count;
1138 push_ctxt(&saved, &mds->mds_ctxt, NULL);
1139 rc = mds_update_server_data(mds);
1145 #define MDS_NUM_THREADS 8
1146 /* mount the file system (secretly) */
1147 static int mds_setup(struct obd_device *obddev, obd_count len, void *buf)
1150 struct obd_ioctl_data* data = buf;
1151 struct mds_obd *mds = &obddev->u.mds;
1152 struct vfsmount *mnt;
1157 #ifdef CONFIG_DEV_RDONLY
1158 dev_clear_rdonly(2);
1160 if (!data->ioc_inlbuf1 || !data->ioc_inlbuf2)
1161 GOTO(err_dec, rc = -EINVAL);
1163 mds->mds_fstype = strdup(data->ioc_inlbuf2);
1165 mnt = do_kern_mount(mds->mds_fstype, 0, data->ioc_inlbuf1, NULL);
1168 CERROR("do_kern_mount failed: rc = %d\n", rc);
1169 GOTO(err_kfree, rc);
1172 mds->mds_sb = mnt->mnt_root->d_inode->i_sb;
1174 GOTO(err_put, rc = -ENODEV);
1176 spin_lock_init(&mds->mds_last_lock);
1177 mds->mds_max_mdsize = sizeof(struct lov_mds_md);
1178 rc = mds_fs_setup(obddev, mnt);
1180 CERROR("MDS filesystem method init failed: rc = %d\n", rc);
1184 mds->mds_service = ptlrpc_init_svc(MDS_NEVENTS, MDS_NBUFS,
1185 MDS_BUFSIZE, MDS_MAXREQSIZE,
1186 MDS_REQUEST_PORTAL, MDC_REPLY_PORTAL,
1187 "self", mds_handle, "mds");
1188 if (!mds->mds_service) {
1189 CERROR("failed to start service\n");
1190 GOTO(err_fs, rc = -EINVAL);
1193 obddev->obd_namespace =
1194 ldlm_namespace_new("mds_server", LDLM_NAMESPACE_SERVER);
1195 if (obddev->obd_namespace == NULL) {
1196 mds_cleanup(obddev);
1197 GOTO(err_svc, rc = -ENOMEM);
1200 for (i = 0; i < MDS_NUM_THREADS; i++) {
1202 sprintf(name, "lustre_MDS_%02d", i);
1203 rc = ptlrpc_start_thread(obddev, mds->mds_service, name);
1205 CERROR("cannot start MDS thread #%d: rc %d\n", i, rc);
1206 GOTO(err_thread, rc);
1210 rc = mds_recover(obddev);
1212 GOTO(err_thread, rc);
1214 ptlrpc_init_client(LDLM_REQUEST_PORTAL, LDLM_REPLY_PORTAL,
1215 "mds_ldlm_client", &obddev->obd_ldlm_client);
1220 ptlrpc_stop_all_threads(mds->mds_service);
1222 ptlrpc_unregister_service(mds->mds_service);
1224 mds_fs_cleanup(obddev);
1227 mntput(mds->mds_vfsmnt);
1231 kfree(mds->mds_fstype);
1237 static int mds_cleanup(struct obd_device *obddev)
1239 struct super_block *sb;
1240 struct mds_obd *mds = &obddev->u.mds;
1241 struct obd_run_ctxt saved;
1244 ptlrpc_stop_all_threads(mds->mds_service);
1245 ptlrpc_unregister_service(mds->mds_service);
1251 push_ctxt(&saved, &mds->mds_ctxt, NULL);
1252 mds_update_server_data(mds);
1254 if (mds->mds_rcvd_filp) {
1255 int rc = filp_close(mds->mds_rcvd_filp, 0);
1256 mds->mds_rcvd_filp = NULL;
1259 CERROR("last_rcvd file won't close, rc=%d\n", rc);
1264 mntput(mds->mds_vfsmnt);
1266 kfree(mds->mds_fstype);
1268 ldlm_namespace_free(obddev->obd_namespace);
1271 #ifdef CONFIG_DEV_RDONLY
1272 dev_clear_rdonly(2);
1274 mds_fs_cleanup(obddev);
1280 static int ldlm_intent_policy(struct ldlm_lock *lock, void *req_cookie,
1281 ldlm_mode_t mode, void *data)
1283 struct ptlrpc_request *req = req_cookie;
1290 if (req->rq_reqmsg->bufcount > 1) {
1291 /* an intent needs to be considered */
1292 struct ldlm_intent *it = lustre_msg_buf(req->rq_reqmsg, 1);
1293 struct mds_obd *mds= &req->rq_export->exp_obd->u.mds;
1294 struct mds_body *mds_rep;
1295 struct ldlm_reply *rep;
1296 __u64 new_resid[3] = {0, 0, 0}, old_res;
1297 int rc, size[3] = {sizeof(struct ldlm_reply),
1298 sizeof(struct mds_body),
1299 mds->mds_max_mdsize};
1301 it->opc = NTOH__u64(it->opc);
1303 LDLM_DEBUG(lock, "intent policy, opc: %s",
1304 ldlm_it2str(it->opc));
1306 rc = lustre_pack_msg(3, size, NULL, &req->rq_replen,
1309 rc = req->rq_status = -ENOMEM;
1313 rep = lustre_msg_buf(req->rq_repmsg, 0);
1314 rep->lock_policy_res1 = 1;
1316 /* execute policy */
1317 switch ((long)it->opc) {
1318 case IT_CREAT|IT_OPEN:
1319 rc = mds_reint(2, req);
1320 if (rc || (req->rq_status != 0 &&
1321 req->rq_status != -EEXIST)) {
1322 rep->lock_policy_res2 = req->rq_status;
1323 RETURN(ELDLM_LOCK_ABORTED);
1334 rc = mds_reint(2, req);
1335 if (rc || (req->rq_status != 0 &&
1336 req->rq_status != -EISDIR &&
1337 req->rq_status != -ENOTDIR)) {
1338 rep->lock_policy_res2 = req->rq_status;
1339 RETURN(ELDLM_LOCK_ABORTED);
1350 rc = mds_getattr_name(2, req);
1351 /* FIXME: we need to sit down and decide on who should
1352 * set req->rq_status, who should return negative and
1353 * positive return values, and what they all mean. */
1354 if (rc || req->rq_status != 0) {
1355 rep->lock_policy_res2 = req->rq_status;
1356 RETURN(ELDLM_LOCK_ABORTED);
1359 case IT_READDIR|IT_OPEN:
1363 CERROR("Unhandled intent\n");
1367 /* We don't bother returning a lock to the client for a file
1368 * or directory we are removing.
1370 * As for link and rename, there is no reason for the client
1371 * to get a lock on the target at this point. If they are
1372 * going to modify the file/directory later they will get a
1373 * lock at that time.
1375 if (it->opc & (IT_UNLINK | IT_RMDIR | IT_LINK | IT_LINK2 |
1376 IT_RENAME | IT_RENAME2))
1377 RETURN(ELDLM_LOCK_ABORTED);
1379 rep->lock_policy_res2 = req->rq_status;
1380 mds_rep = lustre_msg_buf(req->rq_repmsg, 1);
1382 /* If the client is about to open a file that doesn't have an MD
1383 * stripe record, it's going to need a write lock. */
1384 if (it->opc & IT_OPEN) {
1385 struct lov_mds_md *lmm =
1386 lustre_msg_buf(req->rq_repmsg, 2);
1387 if (lmm->lmm_easize == 0) {
1388 LDLM_DEBUG(lock, "open with no EA; returning PW"
1390 lock->l_req_mode = LCK_PW;
1394 /* Give the client a lock on the child object, instead of the
1395 * parent that it requested. */
1396 new_resid[0] = NTOH__u32(mds_rep->ino);
1397 if (new_resid[0] == 0)
1399 old_res = lock->l_resource->lr_name[0];
1401 ldlm_lock_change_resource(lock, new_resid);
1402 if (lock->l_resource == NULL) {
1406 LDLM_DEBUG(lock, "intent policy, old res %ld",
1408 RETURN(ELDLM_LOCK_CHANGED);
1410 int size = sizeof(struct ldlm_reply);
1411 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
1422 extern int mds_iocontrol(long cmd, struct lustre_handle *conn,
1423 int len, void *karg, void *uarg);
1425 /* use obd ops to offer management infrastructure */
1426 static struct obd_ops mds_obd_ops = {
1427 o_connect: mds_connect,
1428 o_disconnect: mds_disconnect,
1430 o_cleanup: mds_cleanup,
1431 o_iocontrol: mds_iocontrol
1434 static int __init mds_init(void)
1436 mds_file_cache = kmem_cache_create("ll_mds_file_data",
1437 sizeof(struct mds_file_data),
1439 if (mds_file_cache == NULL)
1442 class_register_type(&mds_obd_ops, LUSTRE_MDS_NAME);
1443 ldlm_register_intent(ldlm_intent_policy);
1447 static void __exit mds_exit(void)
1449 ldlm_unregister_intent();
1450 class_unregister_type(LUSTRE_MDS_NAME);
1451 if (kmem_cache_destroy(mds_file_cache))
1452 CERROR("couldn't free MDS file cache\n");
1455 MODULE_AUTHOR("Cluster File Systems <info@clusterfs.com>");
1456 MODULE_DESCRIPTION("Lustre Metadata Server (MDS) v0.01");
1457 MODULE_LICENSE("GPL");
1459 module_init(mds_init);
1460 module_exit(mds_exit);