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 EXPORT_SYMTAB
32 #define DEBUG_SUBSYSTEM S_MDS
34 #include <linux/module.h>
35 #include <linux/lustre_mds.h>
36 #include <linux/lustre_dlm.h>
37 #include <linux/init.h>
38 #include <linux/obd_class.h>
39 #include <linux/random.h>
41 #include <linux/jbd.h>
42 #include <linux/ext3_fs.h>
43 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
44 # include <linux/smp_lock.h>
45 # include <linux/buffer_head.h>
46 # include <linux/workqueue.h>
47 # include <linux/mount.h>
49 # include <linux/locks.h>
51 #include <linux/obd_lov.h>
52 #include <linux/lustre_mds.h>
53 #include <linux/lustre_fsfilt.h>
54 #include <linux/lprocfs_status.h>
55 #include <linux/lustre_commit_confd.h>
57 #include "mds_internal.h"
59 static int mds_postsetup(struct obd_device *obd);
60 static int mds_cleanup(struct obd_device *obd, int flags);
62 static int mds_bulk_timeout(void *data)
64 struct ptlrpc_bulk_desc *desc = data;
65 struct obd_export *exp = desc->bd_export;
67 DEBUG_REQ(D_ERROR, desc->bd_req,"bulk send timed out: evicting %s@%s\n",
68 exp->exp_client_uuid.uuid,
69 exp->exp_connection->c_remote_uuid.uuid);
70 ptlrpc_fail_export(exp);
71 ptlrpc_abort_bulk (desc);
75 /* Assumes caller has already pushed into the kernel filesystem context */
76 static int mds_sendpage(struct ptlrpc_request *req, struct file *file,
77 loff_t offset, int count)
79 struct ptlrpc_bulk_desc *desc;
80 struct l_wait_info lwi;
82 int rc = 0, npages, i, tmpcount, tmpsize = 0;
85 LASSERT((offset & (PAGE_SIZE - 1)) == 0); /* I'm dubious about this */
87 npages = (count + PAGE_SIZE - 1) >> PAGE_SHIFT;
88 OBD_ALLOC(pages, sizeof(*pages) * npages);
90 GOTO(out, rc = -ENOMEM);
92 desc = ptlrpc_prep_bulk_exp (req, BULK_PUT_SOURCE, MDS_BULK_PORTAL);
94 GOTO(out_free, rc = -ENOMEM);
96 for (i = 0, tmpcount = count; i < npages; i++, tmpcount -= tmpsize) {
97 tmpsize = tmpcount > PAGE_SIZE ? PAGE_SIZE : tmpcount;
99 pages[i] = alloc_pages(GFP_KERNEL, 0);
100 if (pages[i] == NULL)
101 GOTO(cleanup_buf, rc = -ENOMEM);
103 rc = ptlrpc_prep_bulk_page(desc, pages[i], 0, tmpsize);
105 GOTO(cleanup_buf, rc);
108 for (i = 0, tmpcount = count; i < npages; i++, tmpcount -= tmpsize) {
109 tmpsize = tmpcount > PAGE_SIZE ? PAGE_SIZE : tmpcount;
110 CDEBUG(D_EXT2, "reading %u@%llu from dir %lu (size %llu)\n",
111 tmpsize, offset, file->f_dentry->d_inode->i_ino,
112 file->f_dentry->d_inode->i_size);
114 rc = fsfilt_readpage(req->rq_export->exp_obd, file,
115 page_address(pages[i]), tmpsize, &offset);
118 GOTO(cleanup_buf, rc = -EIO);
121 rc = ptlrpc_bulk_put(desc);
123 GOTO(cleanup_buf, rc);
125 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
126 CERROR("obd_fail_loc=%x, fail operation rc=%d\n",
127 OBD_FAIL_MDS_SENDPAGE, rc);
128 ptlrpc_abort_bulk(desc);
129 GOTO(cleanup_buf, rc);
132 lwi = LWI_TIMEOUT(obd_timeout * HZ / 4, mds_bulk_timeout, desc);
133 rc = l_wait_event(desc->bd_waitq, ptlrpc_bulk_complete (desc), &lwi);
135 LASSERT (rc == -ETIMEDOUT);
136 GOTO(cleanup_buf, rc);
141 for (i = 0; i < npages; i++)
143 __free_pages(pages[i], 0);
145 ptlrpc_free_bulk(desc);
147 OBD_FREE(pages, sizeof(*pages) * npages);
152 /* only valid locked dentries or errors should be returned */
153 struct dentry *mds_fid2locked_dentry(struct obd_device *obd, struct ll_fid *fid,
154 struct vfsmount **mnt, int lock_mode,
155 struct lustre_handle *lockh,
156 char *name, int namelen)
158 struct mds_obd *mds = &obd->u.mds;
159 struct dentry *de = mds_fid2dentry(mds, fid, mnt), *retval = de;
160 struct ldlm_res_id res_id = { .name = {0} };
167 res_id.name[0] = de->d_inode->i_ino;
168 res_id.name[1] = de->d_inode->i_generation;
169 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, NULL,
170 res_id, LDLM_PLAIN, NULL, 0, lock_mode,
171 &flags, ldlm_completion_ast,
172 mds_blocking_ast, NULL, lockh);
173 if (rc != ELDLM_OK) {
175 retval = ERR_PTR(-EIO); /* XXX translate ldlm code */
181 #ifndef DCACHE_DISCONNECTED
182 #define DCACHE_DISCONNECTED DCACHE_NFSD_DISCONNECTED
186 /* Look up an entry by inode number. */
187 /* this function ONLY returns valid dget'd dentries with an initialized inode
189 struct dentry *mds_fid2dentry(struct mds_obd *mds, struct ll_fid *fid,
190 struct vfsmount **mnt)
193 unsigned long ino = fid->id;
194 __u32 generation = fid->generation;
196 struct dentry *result;
199 RETURN(ERR_PTR(-ESTALE));
201 snprintf(fid_name, sizeof(fid_name), "0x%lx", ino);
203 CDEBUG(D_DENTRY, "--> mds_fid2dentry: ino/gen %lu/%u, sb %p\n",
204 ino, generation, mds->mds_sb);
206 /* under ext3 this is neither supposed to return bad inodes
208 result = ll_lookup_one_len(fid_name, mds->mds_fid_de, strlen(fid_name));
212 inode = result->d_inode;
214 RETURN(ERR_PTR(-ENOENT));
216 if (generation && inode->i_generation != generation) {
217 /* we didn't find the right inode.. */
218 CERROR("bad inode %lu, link: %lu ct: %d or generation %u/%u\n",
219 inode->i_ino, (unsigned long)inode->i_nlink,
220 atomic_read(&inode->i_count), inode->i_generation,
223 RETURN(ERR_PTR(-ENOENT));
227 *mnt = mds->mds_vfsmnt;
235 /* Establish a connection to the MDS.
237 * This will set up an export structure for the client to hold state data
238 * about that client, like open files, the last operation number it did
239 * on the server, etc.
241 static int mds_connect(struct lustre_handle *conn, struct obd_device *obd,
242 struct obd_uuid *cluuid)
244 struct obd_export *exp;
245 struct mds_export_data *med; /* */
246 struct mds_client_data *mcd;
247 int rc, abort_recovery;
250 if (!conn || !obd || !cluuid)
253 /* Check for aborted recovery. */
254 spin_lock_bh(&obd->obd_processing_task_lock);
255 abort_recovery = obd->obd_abort_recovery;
256 spin_unlock_bh(&obd->obd_processing_task_lock);
258 target_abort_recovery(obd);
260 /* XXX There is a small race between checking the list and adding a
261 * new connection for the same UUID, but the real threat (list
262 * corruption when multiple different clients connect) is solved.
264 * There is a second race between adding the export to the list,
265 * and filling in the client data below. Hence skipping the case
266 * of NULL mcd above. We should already be controlling multiple
267 * connects at the client, and we can't hold the spinlock over
268 * memory allocations without risk of deadlocking.
270 rc = class_connect(conn, obd, cluuid);
273 exp = class_conn2export(conn);
275 med = &exp->exp_mds_data;
277 OBD_ALLOC(mcd, sizeof(*mcd));
279 CERROR("mds: out of memory for client data\n");
280 GOTO(out, rc = -ENOMEM);
283 memcpy(mcd->mcd_uuid, cluuid, sizeof(mcd->mcd_uuid));
286 rc = mds_client_add(obd, &obd->u.mds, med, -1);
291 OBD_FREE(mcd, sizeof(*mcd));
292 class_disconnect(exp, 0);
294 class_export_put(exp);
299 static int mds_init_export(struct obd_export *exp)
301 struct mds_export_data *med = &exp->exp_mds_data;
303 INIT_LIST_HEAD(&med->med_open_head);
304 spin_lock_init(&med->med_open_lock);
308 static int mds_destroy_export(struct obd_export *export)
310 struct mds_export_data *med;
311 struct obd_device *obd = export->exp_obd;
312 struct obd_run_ctxt saved;
316 med = &export->exp_mds_data;
317 target_destroy_export(export);
319 push_ctxt(&saved, &obd->obd_ctxt, NULL);
320 /* Close any open files (which may also cause orphan unlinking). */
321 spin_lock(&med->med_open_lock);
322 while (!list_empty(&med->med_open_head)) {
323 struct list_head *tmp = med->med_open_head.next;
324 struct mds_file_data *mfd =
325 list_entry(tmp, struct mds_file_data, mfd_list);
326 BDEVNAME_DECLARE_STORAGE(btmp);
328 /* bug 1579: fix force-closing for 2.5 */
329 struct dentry *dentry = mfd->mfd_dentry;
331 list_del(&mfd->mfd_list);
332 spin_unlock(&med->med_open_lock);
334 CERROR("force closing client file handle for %*s (%s:%lu)\n",
335 dentry->d_name.len, dentry->d_name.name,
336 ll_bdevname(dentry->d_inode->i_sb, btmp),
337 dentry->d_inode->i_ino);
338 rc = mds_mfd_close(NULL, obd, mfd,
339 !(export->exp_flags & OBD_OPT_FAILOVER));
342 CDEBUG(D_INODE, "Error closing file: %d\n", rc);
343 spin_lock(&med->med_open_lock);
345 spin_unlock(&med->med_open_lock);
346 pop_ctxt(&saved, &obd->obd_ctxt, NULL);
348 mds_client_free(export, !(export->exp_flags & OBD_OPT_FAILOVER));
353 static int mds_disconnect(struct obd_export *export, int flags)
355 unsigned long irqflags;
359 ldlm_cancel_locks_for_export(export);
361 spin_lock_irqsave(&export->exp_lock, irqflags);
362 export->exp_flags = flags;
363 spin_unlock_irqrestore(&export->exp_lock, irqflags);
365 rc = class_disconnect(export, flags);
369 static int mds_getstatus(struct ptlrpc_request *req)
371 struct mds_obd *mds = mds_req2mds(req);
372 struct mds_body *body;
373 int rc, size = sizeof(*body);
376 rc = lustre_pack_reply(req, 1, &size, NULL);
377 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK)) {
378 CERROR("mds: out of memory for message: size=%d\n", size);
379 req->rq_status = -ENOMEM; /* superfluous? */
383 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
384 memcpy(&body->fid1, &mds->mds_rootfid, sizeof(body->fid1));
386 /* the last_committed and last_xid fields are filled in for all
387 * replies already - no need to do so here also.
392 int mds_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
393 void *data, int flag)
398 if (flag == LDLM_CB_CANCELING) {
399 /* Don't need to do anything here. */
403 /* XXX layering violation! -phil */
404 l_lock(&lock->l_resource->lr_namespace->ns_lock);
405 /* Get this: if mds_blocking_ast is racing with ldlm_intent_policy,
406 * such that mds_blocking_ast is called just before l_i_p takes the
407 * ns_lock, then by the time we get the lock, we might not be the
408 * correct blocking function anymore. So check, and return early, if
410 if (lock->l_blocking_ast != mds_blocking_ast) {
411 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
415 lock->l_flags |= LDLM_FL_CBPENDING;
416 do_ast = (!lock->l_readers && !lock->l_writers);
417 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
420 struct lustre_handle lockh;
423 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
424 ldlm_lock2handle(lock, &lockh);
425 rc = ldlm_cli_cancel(&lockh);
427 CERROR("ldlm_cli_cancel: %d\n", rc);
429 LDLM_DEBUG(lock, "Lock still has references, will be "
435 /* Call with lock=1 if you want mds_pack_md to take the i_sem.
436 * Call with lock=0 if the caller has already taken the i_sem. */
437 int mds_pack_md(struct obd_device *obd, struct lustre_msg *msg, int offset,
438 struct mds_body *body, struct inode *inode, int lock)
440 struct mds_obd *mds = &obd->u.mds;
446 lmm = lustre_msg_buf(msg, offset, 0);
448 /* Some problem with getting eadata when I sized the reply
450 CDEBUG(D_INFO, "no space reserved for inode %lu MD\n",
454 lmm_size = msg->buflens[offset];
456 /* I don't really like this, but it is a sanity check on the client
457 * MD request. However, if the client doesn't know how much space
458 * to reserve for the MD, it shouldn't be bad to have too much space.
460 if (lmm_size > mds->mds_max_mdsize) {
461 CWARN("Reading MD for inode %lu of %d bytes > max %d\n",
462 inode->i_ino, lmm_size, mds->mds_max_mdsize);
468 rc = fsfilt_get_md(obd, inode, lmm, lmm_size);
472 CERROR("Error %d reading eadata for ino %lu\n",
476 rc = mds_convert_lov_ea(obd, inode, lmm, lmm_size);
480 body->valid |= OBD_MD_FLEASIZE;
481 body->eadatasize = lmm_size;
488 static int mds_getattr_internal(struct obd_device *obd, struct dentry *dentry,
489 struct ptlrpc_request *req,
490 struct mds_body *reqbody, int reply_off)
492 struct mds_body *body;
493 struct inode *inode = dentry->d_inode;
500 body = lustre_msg_buf(req->rq_repmsg, reply_off, sizeof(*body));
501 LASSERT(body != NULL); /* caller prepped reply */
503 mds_pack_inode2fid(&body->fid1, inode);
504 mds_pack_inode2body(body, inode);
506 if (S_ISREG(inode->i_mode) && (reqbody->valid & OBD_MD_FLEASIZE) != 0) {
507 rc = mds_pack_md(obd, req->rq_repmsg, reply_off + 1, body,
510 /* If we have LOV EA data, the OST holds size, atime, mtime */
511 if (!(body->valid & OBD_MD_FLEASIZE))
512 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
513 OBD_MD_FLATIME | OBD_MD_FLMTIME);
514 } else if (S_ISLNK(inode->i_mode) &&
515 (reqbody->valid & OBD_MD_LINKNAME) != 0) {
516 char *symname = lustre_msg_buf(req->rq_repmsg, reply_off + 1,0);
519 LASSERT (symname != NULL); /* caller prepped reply */
520 len = req->rq_repmsg->buflens[reply_off + 1];
522 rc = inode->i_op->readlink(dentry, symname, len);
524 CERROR("readlink failed: %d\n", rc);
525 } else if (rc != len - 1) {
526 CERROR ("Unexpected readlink rc %d: expecting %d\n",
530 CDEBUG(D_INODE, "read symlink dest %s\n", symname);
531 body->valid |= OBD_MD_LINKNAME;
532 body->eadatasize = rc + 1;
533 symname[rc] = 0; /* NULL terminate */
541 static int mds_getattr_pack_msg(struct ptlrpc_request *req, struct inode *inode,
544 struct mds_obd *mds = mds_req2mds(req);
545 struct mds_body *body;
546 int rc = 0, size[2] = {sizeof(*body)}, bufcount = 1;
549 body = lustre_msg_buf(req->rq_reqmsg, offset, sizeof (*body));
550 LASSERT(body != NULL); /* checked by caller */
551 LASSERT_REQSWABBED(req, offset); /* swabbed by caller */
553 if (S_ISREG(inode->i_mode) && (body->valid & OBD_MD_FLEASIZE)) {
556 rc = fsfilt_get_md(req->rq_export->exp_obd, inode, NULL, 0);
558 CDEBUG(D_INODE, "got %d bytes MD data for inode %lu\n",
562 CERROR("error getting inode %lu MD: rc = %d\n",
565 } else if (rc > mds->mds_max_mdsize) {
567 CERROR("MD size %d larger than maximum possible %u\n",
568 rc, mds->mds_max_mdsize);
573 } else if (S_ISLNK(inode->i_mode) && (body->valid & OBD_MD_LINKNAME)) {
574 if (inode->i_size + 1 != body->eadatasize)
575 CERROR("symlink size: %Lu, reply space: %d\n",
576 inode->i_size + 1, body->eadatasize);
577 size[bufcount] = min_t(int, inode->i_size+1, body->eadatasize);
579 CDEBUG(D_INODE, "symlink size: %Lu, reply space: %d\n",
580 inode->i_size + 1, body->eadatasize);
583 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
584 CERROR("failed MDS_GETATTR_PACK test\n");
585 req->rq_status = -ENOMEM;
586 GOTO(out, rc = -ENOMEM);
589 rc = lustre_pack_reply(req, bufcount, size, NULL);
591 CERROR("out of memory\n");
592 GOTO(out, req->rq_status = rc);
600 static int mds_getattr_name(int offset, struct ptlrpc_request *req,
601 struct lustre_handle *child_lockh)
603 struct obd_device *obd = req->rq_export->exp_obd;
604 struct ldlm_reply *rep = NULL;
605 struct obd_run_ctxt saved;
606 struct mds_body *body;
607 struct dentry *dparent = NULL, *dchild = NULL;
609 struct lustre_handle parent_lockh;
611 int rc = 0, cleanup_phase = 0, resent_req = 0;
615 LASSERT(!strcmp(obd->obd_type->typ_name, "mds"));
617 /* Swab now, before anyone looks inside the request */
619 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
620 lustre_swab_mds_body);
622 CERROR("Can't swab mds_body\n");
623 GOTO(cleanup, rc = -EFAULT);
626 LASSERT_REQSWAB(req, offset + 1);
627 name = lustre_msg_string(req->rq_reqmsg, offset + 1, 0);
629 CERROR("Can't unpack name\n");
630 GOTO(cleanup, rc = -EFAULT);
632 namesize = req->rq_reqmsg->buflens[offset + 1];
634 LASSERT (offset == 0 || offset == 2);
635 /* if requests were at offset 2, the getattr reply goes back at 1 */
637 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
641 uc.ouc_fsuid = body->fsuid;
642 uc.ouc_fsgid = body->fsgid;
643 uc.ouc_cap = body->capability;
644 uc.ouc_suppgid1 = body->suppgid;
645 uc.ouc_suppgid2 = -1;
646 push_ctxt(&saved, &obd->obd_ctxt, &uc);
647 cleanup_phase = 1; /* kernel context */
648 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
650 /* FIXME: handle raw lookup */
652 if (body->valid == OBD_MD_FLID) {
653 struct mds_body *mds_reply;
654 int size = sizeof(*mds_reply);
656 // The user requested ONLY the inode number, so do a raw lookup
657 rc = lustre_pack_reply(req, 1, &size, NULL);
659 CERROR("out of memory\n");
663 rc = dir->i_op->lookup_raw(dir, name, namesize - 1, &inum);
665 mds_reply = lustre_msg_buf(req->rq_repmsg, offset,
667 mds_reply->fid1.id = inum;
668 mds_reply->valid = OBD_MD_FLID;
673 if (child_lockh->cookie != 0) {
674 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
678 if (resent_req == 0) {
679 rc = mds_get_parent_child_locked(obd, &obd->u.mds, &body->fid1,
680 &parent_lockh, &dparent,
681 LCK_PR, name, namesize,
682 child_lockh, &dchild, LCK_PR);
686 struct ldlm_lock *granted_lock;
687 struct ll_fid child_fid;
688 struct ldlm_resource *res;
689 DEBUG_REQ(D_DLMTRACE, req, "resent, not enqueuing new locks");
690 granted_lock = ldlm_handle2lock(child_lockh);
691 LASSERT(granted_lock);
693 res = granted_lock->l_resource;
694 child_fid.id = res->lr_name.name[0];
695 child_fid.generation = res->lr_name.name[1];
696 dchild = mds_fid2dentry(&obd->u.mds, &child_fid, NULL);
698 LDLM_LOCK_PUT(granted_lock);
701 cleanup_phase = 2; /* dchild, dparent, locks */
703 if (dchild->d_inode == NULL) {
704 intent_set_disposition(rep, DISP_LOOKUP_NEG);
705 /* in the intent case, the policy clears this error:
706 the disposition is enough */
707 GOTO(cleanup, rc = -ENOENT);
709 intent_set_disposition(rep, DISP_LOOKUP_POS);
712 if (req->rq_repmsg == NULL) {
713 rc = mds_getattr_pack_msg(req, dchild->d_inode, offset);
715 CERROR ("mds_getattr_pack_msg: %d\n", rc);
720 rc = mds_getattr_internal(obd, dchild, req, body, offset);
721 GOTO(cleanup, rc); /* returns the lock to the client */
724 switch (cleanup_phase) {
726 if (resent_req == 0) {
727 if (rc && dchild->d_inode)
728 ldlm_lock_decref(child_lockh, LCK_PR);
729 ldlm_lock_decref(&parent_lockh, LCK_PR);
734 pop_ctxt(&saved, &obd->obd_ctxt, &uc);
740 static int mds_getattr(int offset, struct ptlrpc_request *req)
742 struct mds_obd *mds = mds_req2mds(req);
743 struct obd_device *obd = req->rq_export->exp_obd;
744 struct obd_run_ctxt saved;
746 struct mds_body *body;
751 body = lustre_swab_reqbuf (req, offset, sizeof (*body),
752 lustre_swab_mds_body);
754 CERROR ("Can't unpack body\n");
758 uc.ouc_fsuid = body->fsuid;
759 uc.ouc_fsgid = body->fsgid;
760 uc.ouc_cap = body->capability;
761 push_ctxt(&saved, &obd->obd_ctxt, &uc);
762 de = mds_fid2dentry(mds, &body->fid1, NULL);
764 rc = req->rq_status = -ENOENT;
765 GOTO(out_pop, PTR_ERR(de));
768 rc = mds_getattr_pack_msg(req, de->d_inode, offset);
770 CERROR ("mds_getattr_pack_msg: %d\n", rc);
774 req->rq_status = mds_getattr_internal(obd, de, req, body, 0);
779 pop_ctxt(&saved, &obd->obd_ctxt, &uc);
784 static int mds_obd_statfs(struct obd_device *obd, struct obd_statfs *osfs,
785 unsigned long max_age)
789 spin_lock(&obd->obd_osfs_lock);
790 rc = fsfilt_statfs(obd, obd->u.mds.mds_sb, max_age);
792 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
793 spin_unlock(&obd->obd_osfs_lock);
798 static int mds_statfs(struct ptlrpc_request *req)
800 struct obd_device *obd = req->rq_export->exp_obd;
801 int rc, size = sizeof(struct obd_statfs);
804 rc = lustre_pack_reply(req, 1, &size, NULL);
805 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
806 CERROR("mds: statfs lustre_pack_reply failed: rc = %d\n", rc);
810 /* We call this so that we can cache a bit - 1 jiffie worth */
811 rc = mds_obd_statfs(obd, lustre_msg_buf(req->rq_repmsg, 0, size),
814 CERROR("mds_obd_statfs failed: rc %d\n", rc);
824 static int mds_sync(struct ptlrpc_request *req)
826 struct obd_device *obd = req->rq_export->exp_obd;
827 struct mds_obd *mds = &obd->u.mds;
828 struct mds_body *body;
829 int rc, size = sizeof(*body);
832 body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof(*body));
834 GOTO(out, rc = -EPROTO);
836 rc = lustre_pack_reply(req, 1, &size, NULL);
837 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_SYNC_PACK)) {
838 CERROR("fsync lustre_pack_reply failed: rc = %d\n", rc);
842 if (body->fid1.id == 0) {
843 /* a fid of zero is taken to mean "sync whole filesystem" */
844 rc = fsfilt_sync(obd, mds->mds_sb);
848 /* just any file to grab fsync method - "file" arg unused */
849 struct file *file = mds->mds_rcvd_filp;
852 de = mds_fid2dentry(mds, &body->fid1, NULL);
854 GOTO(out, rc = PTR_ERR(de));
856 rc = file->f_op->fsync(NULL, de, 1);
861 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
862 mds_pack_inode2fid(&body->fid1, de->d_inode);
863 mds_pack_inode2body(body, de->d_inode);
870 /* mds_readpage does not take a DLM lock on the inode, because the client must
871 * already have a PR lock.
873 * If we were to take another one here, a deadlock will result, if another
874 * thread is already waiting for a PW lock. */
875 static int mds_readpage(struct ptlrpc_request *req)
877 struct obd_device *obd = req->rq_export->exp_obd;
878 struct vfsmount *mnt;
881 struct mds_body *body, *repbody;
882 struct obd_run_ctxt saved;
883 int rc, size = sizeof(*repbody);
887 rc = lustre_pack_reply(req, 1, &size, NULL);
888 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK)) {
889 CERROR("mds: out of memory\n");
890 GOTO(out, rc = -ENOMEM);
893 body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_mds_body);
895 GOTO (out, rc = -EFAULT);
897 uc.ouc_fsuid = body->fsuid;
898 uc.ouc_fsgid = body->fsgid;
899 uc.ouc_cap = body->capability;
900 push_ctxt(&saved, &obd->obd_ctxt, &uc);
901 de = mds_fid2dentry(&obd->u.mds, &body->fid1, &mnt);
903 GOTO(out_pop, rc = PTR_ERR(de));
905 CDEBUG(D_INODE, "ino %lu\n", de->d_inode->i_ino);
907 file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
908 /* note: in case of an error, dentry_open puts dentry */
910 GOTO(out_pop, rc = PTR_ERR(file));
912 /* body->size is actually the offset -eeb */
913 if ((body->size & (de->d_inode->i_blksize - 1)) != 0) {
914 CERROR("offset "LPU64" not on a block boundary of %lu\n",
915 body->size, de->d_inode->i_blksize);
916 GOTO(out_file, rc = -EFAULT);
919 /* body->nlink is actually the #bytes to read -eeb */
920 if (body->nlink & (de->d_inode->i_blksize - 1)) {
921 CERROR("size %u is not multiple of blocksize %lu\n",
922 body->nlink, de->d_inode->i_blksize);
923 GOTO(out_file, rc = -EFAULT);
926 repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*repbody));
927 repbody->size = file->f_dentry->d_inode->i_size;
928 repbody->valid = OBD_MD_FLSIZE;
930 /* to make this asynchronous make sure that the handling function
931 doesn't send a reply when this function completes. Instead a
932 callback function would send the reply */
933 /* body->size is actually the offset -eeb */
934 rc = mds_sendpage(req, file, body->size, body->nlink);
939 pop_ctxt(&saved, &obd->obd_ctxt, &uc);
945 int mds_reint(struct ptlrpc_request *req, int offset,
946 struct lustre_handle *lockh)
948 struct mds_update_record *rec; /* 116 bytes on the stack? no sir! */
951 OBD_ALLOC(rec, sizeof(*rec));
955 rc = mds_update_unpack(req, offset, rec);
956 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
957 CERROR("invalid record\n");
958 GOTO(out, req->rq_status = -EINVAL);
960 /* rc will be used to interrupt a for loop over multiple records */
961 rc = mds_reint_rec(rec, offset, req, lockh);
963 OBD_FREE(rec, sizeof(*rec));
967 static int mds_filter_recovery_request(struct ptlrpc_request *req,
968 struct obd_device *obd, int *process)
970 switch (req->rq_reqmsg->opc) {
971 case MDS_CONNECT: /* This will never get here, but for completeness. */
972 case OST_CONNECT: /* This will never get here, but for completeness. */
979 case MDS_SYNC: /* used in unmounting */
983 *process = target_queue_recovery_request(req, obd);
987 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
989 /* XXX what should we set rq_status to here? */
990 req->rq_status = -EAGAIN;
991 RETURN(ptlrpc_error(req));
995 static char *reint_names[] = {
996 [REINT_SETATTR] "setattr",
997 [REINT_CREATE] "create",
999 [REINT_UNLINK] "unlink",
1000 [REINT_RENAME] "rename",
1001 [REINT_OPEN] "open",
1004 int mds_handle(struct ptlrpc_request *req)
1006 int should_process, fail = OBD_FAIL_MDS_ALL_REPLY_NET;
1008 struct mds_obd *mds = NULL; /* quell gcc overwarning */
1009 struct obd_device *obd = NULL;
1012 OBD_FAIL_RETURN(OBD_FAIL_MDS_ALL_REQUEST_NET | OBD_FAIL_ONCE, 0);
1014 LASSERT(current->journal_info == NULL);
1015 /* XXX identical to OST */
1016 if (req->rq_reqmsg->opc != MDS_CONNECT) {
1017 struct mds_export_data *med;
1018 int recovering, abort_recovery;
1020 if (req->rq_export == NULL) {
1021 CERROR("lustre_mds: operation %d on unconnected MDS\n",
1022 req->rq_reqmsg->opc);
1023 req->rq_status = -ENOTCONN;
1024 GOTO(out, rc = -ENOTCONN);
1027 med = &req->rq_export->exp_mds_data;
1028 obd = req->rq_export->exp_obd;
1031 /* Check for aborted recovery. */
1032 spin_lock_bh(&obd->obd_processing_task_lock);
1033 abort_recovery = obd->obd_abort_recovery;
1034 recovering = obd->obd_recovering;
1035 spin_unlock_bh(&obd->obd_processing_task_lock);
1036 if (abort_recovery) {
1037 target_abort_recovery(obd);
1038 } else if (recovering) {
1039 rc = mds_filter_recovery_request(req, obd,
1041 if (rc || !should_process)
1046 switch (req->rq_reqmsg->opc) {
1048 DEBUG_REQ(D_INODE, req, "connect");
1049 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
1050 rc = target_handle_connect(req, mds_handle);
1052 /* Now that we have an export, set mds. */
1053 mds = mds_req2mds(req);
1056 case MDS_DISCONNECT:
1057 DEBUG_REQ(D_INODE, req, "disconnect");
1058 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
1059 rc = target_handle_disconnect(req);
1060 req->rq_status = rc; /* superfluous? */
1064 DEBUG_REQ(D_INODE, req, "getstatus");
1065 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
1066 rc = mds_getstatus(req);
1070 DEBUG_REQ(D_INODE, req, "getattr");
1071 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
1072 rc = mds_getattr(0, req);
1075 case MDS_GETATTR_NAME: {
1076 struct lustre_handle lockh;
1077 DEBUG_REQ(D_INODE, req, "getattr_name");
1078 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NAME_NET, 0);
1080 /* If this request gets a reconstructed reply, we won't be
1081 * acquiring any new locks in mds_getattr_name, so we don't
1085 rc = mds_getattr_name(0, req, &lockh);
1086 /* this non-intent call (from an ioctl) is special */
1087 req->rq_status = rc;
1088 if (rc == 0 && lockh.cookie)
1089 ldlm_lock_decref(&lockh, LCK_PR);
1093 DEBUG_REQ(D_INODE, req, "statfs");
1094 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
1095 rc = mds_statfs(req);
1099 DEBUG_REQ(D_INODE, req, "readpage");
1100 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
1101 rc = mds_readpage(req);
1103 OBD_FAIL_RETURN(OBD_FAIL_MDS_SENDPAGE, 0);
1108 __u32 *opcp = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*opcp));
1110 int size[3] = {sizeof(struct mds_body), mds->mds_max_mdsize,
1111 mds->mds_max_cookiesize};
1114 /* NB only peek inside req now; mds_reint() will swab it */
1116 CERROR ("Can't inspect opcode\n");
1121 if (lustre_msg_swabbed (req->rq_reqmsg))
1124 DEBUG_REQ(D_INODE, req, "reint %d (%s)", opc,
1125 (opc < sizeof(reint_names) / sizeof(reint_names[0]) ||
1126 reint_names[opc] == NULL) ? reint_names[opc] :
1129 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
1131 if (opc == REINT_UNLINK)
1133 else if (opc == REINT_OPEN || opc == REINT_RENAME)
1138 rc = lustre_pack_reply(req, bufcount, size, NULL);
1142 rc = mds_reint(req, 0, NULL);
1143 fail = OBD_FAIL_MDS_REINT_NET_REP;
1148 DEBUG_REQ(D_INODE, req, "close");
1149 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
1150 rc = mds_close(req);
1153 case MDS_DONE_WRITING:
1154 DEBUG_REQ(D_INODE, req, "done_writing");
1155 OBD_FAIL_RETURN(OBD_FAIL_MDS_DONE_WRITING_NET, 0);
1156 rc = mds_done_writing(req);
1160 DEBUG_REQ(D_INODE, req, "pin");
1161 OBD_FAIL_RETURN(OBD_FAIL_MDS_PIN_NET, 0);
1166 DEBUG_REQ(D_INODE, req, "sync");
1167 OBD_FAIL_RETURN(OBD_FAIL_MDS_SYNC_NET, 0);
1172 DEBUG_REQ(D_INODE, req, "ping");
1173 rc = target_handle_ping(req);
1176 case OBD_LOG_CANCEL:
1177 CDEBUG(D_INODE, "log cancel\n");
1178 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
1179 rc = -ENOTSUPP; /* la la la */
1183 DEBUG_REQ(D_INODE, req, "enqueue");
1184 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
1185 rc = ldlm_handle_enqueue(req, ldlm_server_completion_ast,
1186 ldlm_server_blocking_ast);
1189 DEBUG_REQ(D_INODE, req, "convert");
1190 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
1191 rc = ldlm_handle_convert(req);
1193 case LDLM_BL_CALLBACK:
1194 case LDLM_CP_CALLBACK:
1195 DEBUG_REQ(D_INODE, req, "callback");
1196 CERROR("callbacks should not happen on MDS\n");
1198 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
1200 case LLOG_ORIGIN_HANDLE_CREATE:
1201 DEBUG_REQ(D_INODE, req, "llog_init");
1202 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1203 rc = llog_origin_handle_create(req);
1205 case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
1206 DEBUG_REQ(D_INODE, req, "llog next block");
1207 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1208 rc = llog_origin_handle_next_block(req);
1210 case LLOG_ORIGIN_HANDLE_READ_HEADER:
1211 DEBUG_REQ(D_INODE, req, "llog read header");
1212 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1213 rc = llog_origin_handle_read_header(req);
1215 case LLOG_ORIGIN_HANDLE_CLOSE:
1216 DEBUG_REQ(D_INODE, req, "llog close");
1217 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1218 rc = llog_origin_handle_close(req);
1221 DEBUG_REQ(D_INODE, req, "llog catinfo");
1222 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1223 rc = llog_catinfo(req);
1226 req->rq_status = -ENOTSUPP;
1227 rc = ptlrpc_error(req);
1231 LASSERT(current->journal_info == NULL);
1235 /* If we're DISCONNECTing, the mds_export_data is already freed */
1236 if (!rc && req->rq_reqmsg->opc != MDS_DISCONNECT) {
1237 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1238 struct obd_device *obd = list_entry(mds, struct obd_device,
1240 req->rq_repmsg->last_xid =
1241 le64_to_cpu(med->med_mcd->mcd_last_xid);
1243 if (!obd->obd_no_transno) {
1244 req->rq_repmsg->last_committed =
1245 obd->obd_last_committed;
1247 DEBUG_REQ(D_IOCTL, req,
1248 "not sending last_committed update");
1250 CDEBUG(D_INFO, "last_transno "LPU64", last_committed "LPU64
1252 mds->mds_last_transno, obd->obd_last_committed,
1257 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LAST_REPLAY) {
1258 if (obd && obd->obd_recovering) {
1259 DEBUG_REQ(D_HA, req, "LAST_REPLAY, queuing reply");
1260 return target_queue_final_reply(req, rc);
1262 /* Lost a race with recovery; let the error path DTRT. */
1263 rc = req->rq_status = -ENOTCONN;
1266 target_send_reply(req, rc, fail);
1270 /* Update the server data on disk. This stores the new mount_count and
1271 * also the last_rcvd value to disk. If we don't have a clean shutdown,
1272 * then the server last_rcvd value may be less than that of the clients.
1273 * This will alert us that we may need to do client recovery.
1275 * Also assumes for mds_last_transno that we are not modifying it (no locking).
1277 int mds_update_server_data(struct obd_device *obd, int force_sync)
1279 struct mds_obd *mds = &obd->u.mds;
1280 struct mds_server_data *msd = mds->mds_server_data;
1281 struct file *filp = mds->mds_rcvd_filp;
1282 struct obd_run_ctxt saved;
1287 push_ctxt(&saved, &obd->obd_ctxt, NULL);
1288 msd->msd_last_transno = cpu_to_le64(mds->mds_last_transno);
1290 CDEBUG(D_SUPER, "MDS mount_count is "LPU64", last_transno is "LPU64"\n",
1291 mds->mds_mount_count, mds->mds_last_transno);
1292 rc = fsfilt_write_record(obd, filp, msd, sizeof(*msd), &off,force_sync);
1294 CERROR("error writing MDS server data: rc = %d\n", rc);
1295 pop_ctxt(&saved, &obd->obd_ctxt, NULL);
1301 /* mount the file system (secretly) */
1302 static int mds_setup(struct obd_device *obd, obd_count len, void *buf)
1304 struct lustre_cfg* lcfg = buf;
1305 struct mds_obd *mds = &obd->u.mds;
1306 struct vfsmount *mnt;
1311 dev_clear_rdonly(2);
1313 if (!lcfg->lcfg_inlbuf1 || !lcfg->lcfg_inlbuf2)
1314 RETURN(rc = -EINVAL);
1316 obd->obd_fsops = fsfilt_get_ops(lcfg->lcfg_inlbuf2);
1317 if (IS_ERR(obd->obd_fsops))
1318 RETURN(rc = PTR_ERR(obd->obd_fsops));
1320 if (!(page = __get_free_page(GFP_KERNEL)))
1323 memset((void *)page, 0, PAGE_SIZE);
1324 sprintf((char *)page, "iopen_nopriv");
1326 mnt = do_kern_mount(lcfg->lcfg_inlbuf2, 0,
1327 lcfg->lcfg_inlbuf1, (void *)page);
1331 CERROR("do_kern_mount failed: rc = %d\n", rc);
1335 CDEBUG(D_SUPER, "%s: mnt = %p\n", lcfg->lcfg_inlbuf1, mnt);
1337 sema_init(&mds->mds_orphan_recovery_sem, 1);
1338 sema_init(&mds->mds_epoch_sem, 1);
1339 spin_lock_init(&mds->mds_transno_lock);
1340 mds->mds_max_mdsize = sizeof(struct lov_mds_md);
1341 mds->mds_max_cookiesize = sizeof(struct llog_cookie);
1342 atomic_set(&mds->mds_open_count, 0);
1344 obd->obd_namespace = ldlm_namespace_new("mds_server",
1345 LDLM_NAMESPACE_SERVER);
1346 if (obd->obd_namespace == NULL) {
1347 mds_cleanup(obd, 0);
1348 GOTO(err_put, rc = -ENOMEM);
1351 rc = mds_fs_setup(obd, mnt);
1353 CERROR("MDS filesystem method init failed: rc = %d\n", rc);
1357 rc = llog_start_commit_thread();
1362 if (lcfg->lcfg_inllen3 > 0 && lcfg->lcfg_inlbuf3) {
1365 generate_random_uuid(uuid);
1366 class_uuid_unparse(uuid, &mds->mds_lov_uuid);
1368 OBD_ALLOC(mds->mds_profile, lcfg->lcfg_inllen3);
1369 if (mds->mds_profile == NULL)
1370 GOTO(err_fs, rc = -ENOMEM);
1372 memcpy(mds->mds_profile, lcfg->lcfg_inlbuf3,
1373 lcfg->lcfg_inllen3);
1377 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
1378 "mds_ldlm_client", &obd->obd_ldlm_client);
1379 obd->obd_replayable = 1;
1381 rc = mds_postsetup(obd);
1387 /* No extra cleanup needed for llog_init_commit_thread() */
1388 mds_fs_cleanup(obd, 0);
1390 ldlm_namespace_free(obd->obd_namespace, 0);
1391 obd->obd_namespace = NULL;
1394 mntput(mds->mds_vfsmnt);
1398 fsfilt_put_ops(obd->obd_fsops);
1402 static int mds_postsetup(struct obd_device *obd)
1404 struct mds_obd *mds = &obd->u.mds;
1409 rc = llog_setup(obd, LLOG_CONFIG_ORIG_CTXT, obd, 0, NULL,
1414 if (mds->mds_profile) {
1415 struct obd_run_ctxt saved;
1416 struct lustre_profile *lprof;
1417 struct config_llog_instance cfg;
1419 cfg.cfg_instance = NULL;
1420 cfg.cfg_uuid = mds->mds_lov_uuid;
1421 push_ctxt(&saved, &obd->obd_ctxt, NULL);
1422 rc = class_config_parse_llog(llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT),
1423 mds->mds_profile, &cfg);
1424 pop_ctxt(&saved, &obd->obd_ctxt, NULL);
1428 lprof = class_get_profile(mds->mds_profile);
1429 if (lprof == NULL) {
1430 CERROR("No profile found: %s\n", mds->mds_profile);
1431 GOTO(err_cleanup, rc = -ENOENT);
1433 rc = mds_lov_connect(obd, lprof->lp_osc);
1435 GOTO(err_cleanup, rc);
1443 llog_cleanup(llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT));
1447 static int mds_postrecov(struct obd_device *obd)
1452 LASSERT(!obd->obd_recovering);
1453 LASSERT(llog_get_context(obd, LLOG_UNLINK_ORIG_CTXT) != NULL);
1455 rc = llog_connect(llog_get_context(obd, LLOG_UNLINK_ORIG_CTXT),
1456 obd->u.mds.mds_lov_desc.ld_tgt_count, NULL, NULL);
1458 CERROR("faild at llog_origin_connect: %d\n", rc);
1461 rc = mds_cleanup_orphans(obd);
1463 rc2 = mds_lov_set_nextid(obd);
1469 int mds_lov_clean(struct obd_device *obd)
1471 struct mds_obd *mds = &obd->u.mds;
1473 if (mds->mds_profile) {
1475 struct config_llog_instance cfg;
1476 struct obd_run_ctxt saved;
1477 int len = strlen(mds->mds_profile) + sizeof("-clean") + 1;
1479 OBD_ALLOC(cln_prof, len);
1480 sprintf(cln_prof, "%s-clean", mds->mds_profile);
1482 cfg.cfg_instance = NULL;
1483 cfg.cfg_uuid = mds->mds_lov_uuid;
1485 push_ctxt(&saved, &obd->obd_ctxt, NULL);
1486 class_config_parse_llog(llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT),
1488 pop_ctxt(&saved, &obd->obd_ctxt, NULL);
1490 OBD_FREE(cln_prof, len);
1491 OBD_FREE(mds->mds_profile, strlen(mds->mds_profile) + 1);
1492 mds->mds_profile = NULL;
1497 static int mds_precleanup(struct obd_device *obd, int flags)
1502 mds_lov_disconnect(obd, flags);
1504 llog_cleanup(llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT));
1508 static int mds_cleanup(struct obd_device *obd, int flags)
1510 struct mds_obd *mds = &obd->u.mds;
1513 if (mds->mds_sb == NULL)
1516 mds_update_server_data(obd, 1);
1517 if (mds->mds_lov_objids != NULL) {
1518 OBD_FREE(mds->mds_lov_objids,
1519 mds->mds_lov_desc.ld_tgt_count * sizeof(obd_id));
1521 mds_fs_cleanup(obd, flags);
1525 /* 2 seems normal on mds, (may_umount() also expects 2
1526 fwiw), but we only see 1 at this point in obdfilter. */
1527 if (atomic_read(&obd->u.mds.mds_vfsmnt->mnt_count) > 2)
1528 CERROR("%s: mount busy, mnt_count %d != 2\n", obd->obd_name,
1529 atomic_read(&obd->u.mds.mds_vfsmnt->mnt_count));
1531 mntput(mds->mds_vfsmnt);
1535 ldlm_namespace_free(obd->obd_namespace, flags & OBD_OPT_FORCE);
1537 spin_lock_bh(&obd->obd_processing_task_lock);
1538 if (obd->obd_recovering) {
1539 target_cancel_recovery_timer(obd);
1540 obd->obd_recovering = 0;
1542 spin_unlock_bh(&obd->obd_processing_task_lock);
1545 dev_clear_rdonly(2);
1546 fsfilt_put_ops(obd->obd_fsops);
1551 static void fixup_handle_for_resent_req(struct ptlrpc_request *req,
1552 struct ldlm_lock *new_lock,
1553 struct lustre_handle *lockh)
1555 struct obd_export *exp = req->rq_export;
1556 struct obd_device *obd = exp->exp_obd;
1557 struct ldlm_request *dlmreq =
1558 lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*dlmreq));
1559 struct lustre_handle remote_hdl = dlmreq->lock_handle1;
1560 struct list_head *iter;
1562 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
1565 l_lock(&obd->obd_namespace->ns_lock);
1566 list_for_each(iter, &exp->exp_ldlm_data.led_held_locks) {
1567 struct ldlm_lock *lock;
1568 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
1569 if (lock == new_lock)
1571 if (lock->l_remote_handle.cookie == remote_hdl.cookie) {
1572 lockh->cookie = lock->l_handle.h_cookie;
1573 DEBUG_REQ(D_HA, req, "restoring lock cookie "LPX64,
1575 l_unlock(&obd->obd_namespace->ns_lock);
1579 l_unlock(&obd->obd_namespace->ns_lock);
1581 /* This remote handle isn't enqueued, so we never received or
1582 * processed this request. Clear MSG_RESENT, because it can
1583 * be handled like any normal request now. */
1585 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
1587 DEBUG_REQ(D_HA, req, "no existing lock with rhandle "LPX64,
1591 int intent_disposition(struct ldlm_reply *rep, int flag)
1595 return (rep->lock_policy_res1 & flag);
1598 void intent_set_disposition(struct ldlm_reply *rep, int flag)
1602 rep->lock_policy_res1 |= flag;
1605 static int ldlm_intent_policy(struct ldlm_namespace *ns,
1606 struct ldlm_lock **lockp, void *req_cookie,
1607 ldlm_mode_t mode, int flags, void *data)
1609 struct ptlrpc_request *req = req_cookie;
1610 struct ldlm_lock *lock = *lockp;
1617 if (req->rq_reqmsg->bufcount > 1) {
1618 /* an intent needs to be considered */
1619 struct ldlm_intent *it;
1620 struct mds_obd *mds = &req->rq_export->exp_obd->u.mds;
1621 struct ldlm_reply *rep;
1622 struct lustre_handle lockh = { 0 };
1623 struct ldlm_lock *new_lock;
1624 int offset = 2, repsize[4] = {sizeof(struct ldlm_reply),
1625 sizeof(struct mds_body),
1626 mds->mds_max_mdsize,
1627 mds->mds_max_cookiesize};
1629 it = lustre_swab_reqbuf(req, 1, sizeof (*it),
1630 lustre_swab_ldlm_intent);
1632 CERROR ("Intent missing\n");
1633 req->rq_status = -EFAULT;
1634 RETURN(req->rq_status);
1637 LDLM_DEBUG(lock, "intent policy, opc: %s",
1638 ldlm_it2str(it->opc));
1640 rc = lustre_pack_reply(req, it->opc == IT_UNLINK ? 4 : 3,
1643 RETURN(req->rq_status = rc);
1645 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
1646 intent_set_disposition(rep, DISP_IT_EXECD);
1648 fixup_handle_for_resent_req(req, lock, &lockh);
1650 /* execute policy */
1651 switch ((long)it->opc) {
1653 case IT_CREAT|IT_OPEN:
1654 /* XXX swab here to assert that an mds_open reint
1655 * packet is following */
1656 rep->lock_policy_res2 = mds_reint(req, offset, &lockh);
1658 /* We abort the lock if the lookup was negative and
1659 * we did not make it to the OPEN portion */
1660 if (!intent_disposition(rep, DISP_LOOKUP_EXECD))
1661 RETURN(ELDLM_LOCK_ABORTED);
1662 if (intent_disposition(rep, DISP_LOOKUP_NEG) &&
1663 !intent_disposition(rep, DISP_OPEN_OPEN))
1665 RETURN(ELDLM_LOCK_ABORTED);
1670 rep->lock_policy_res2 = mds_getattr_name(offset, req,
1672 /* FIXME: LDLM can set req->rq_status. MDS sets
1673 policy_res{1,2} with disposition and status.
1674 - replay: returns 0 & req->status is old status
1675 - otherwise: returns req->status */
1676 if (intent_disposition(rep, DISP_LOOKUP_NEG))
1677 rep->lock_policy_res2 = 0;
1678 if (!intent_disposition(rep, DISP_LOOKUP_POS) ||
1679 rep->lock_policy_res2)
1680 RETURN(ELDLM_LOCK_ABORTED);
1681 if (req->rq_status != 0) {
1683 rep->lock_policy_res2 = req->rq_status;
1684 RETURN(ELDLM_LOCK_ABORTED);
1688 CERROR("Unhandled intent "LPD64"\n", it->opc);
1692 /* By this point, whatever function we called above must have
1693 * either filled in 'lockh', been an intent replay, or returned
1694 * an error. We want to allow replayed RPCs to not get a lock,
1695 * since we would just drop it below anyways because lock replay
1696 * is done separately by the client afterwards. For regular
1697 * RPCs we want to give the new lock to the client instead of
1698 * whatever lock it was about to get.
1700 new_lock = ldlm_handle2lock(&lockh);
1701 if (new_lock == NULL && (flags & LDLM_FL_INTENT_ONLY))
1704 LASSERT(new_lock != NULL);
1706 /* If we've already given this lock to a client once, then we
1707 * should have no readers or writers. Otherwise, we should
1708 * have one reader _or_ writer ref (which will be zeroed below)
1709 * before returning the lock to a client.
1711 if (new_lock->l_export == req->rq_export) {
1712 LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
1714 LASSERT(new_lock->l_export == NULL);
1715 LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
1720 if (new_lock->l_export == req->rq_export) {
1721 /* Already gave this to the client, which means that we
1722 * reconstructed a reply. */
1723 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
1725 RETURN(ELDLM_LOCK_REPLACED);
1728 /* Fixup the lock to be given to the client */
1729 l_lock(&new_lock->l_resource->lr_namespace->ns_lock);
1730 new_lock->l_readers = 0;
1731 new_lock->l_writers = 0;
1733 new_lock->l_export = class_export_get(req->rq_export);
1734 list_add(&new_lock->l_export_chain,
1735 &new_lock->l_export->exp_ldlm_data.led_held_locks);
1737 new_lock->l_blocking_ast = lock->l_blocking_ast;
1738 new_lock->l_completion_ast = lock->l_completion_ast;
1740 memcpy(&new_lock->l_remote_handle, &lock->l_remote_handle,
1741 sizeof(lock->l_remote_handle));
1743 new_lock->l_flags &= ~LDLM_FL_LOCAL;
1745 LDLM_LOCK_PUT(new_lock);
1746 l_unlock(&new_lock->l_resource->lr_namespace->ns_lock);
1748 RETURN(ELDLM_LOCK_REPLACED);
1750 int size = sizeof(struct ldlm_reply);
1751 rc = lustre_pack_reply(req, 1, &size, NULL);
1760 int mds_attach(struct obd_device *dev, obd_count len, void *data)
1762 struct lprocfs_static_vars lvars;
1764 lprocfs_init_multi_vars(0, &lvars);
1765 return lprocfs_obd_attach(dev, lvars.obd_vars);
1768 int mds_detach(struct obd_device *dev)
1770 return lprocfs_obd_detach(dev);
1773 int mdt_attach(struct obd_device *dev, obd_count len, void *data)
1775 struct lprocfs_static_vars lvars;
1777 lprocfs_init_multi_vars(1, &lvars);
1778 return lprocfs_obd_attach(dev, lvars.obd_vars);
1781 int mdt_detach(struct obd_device *dev)
1783 return lprocfs_obd_detach(dev);
1786 static int mdt_setup(struct obd_device *obddev, obd_count len, void *buf)
1788 struct mds_obd *mds = &obddev->u.mds;
1792 mds->mds_service = ptlrpc_init_svc(MDS_NEVENTS, MDS_NBUFS,
1793 MDS_BUFSIZE, MDS_MAXREQSIZE,
1794 MDS_REQUEST_PORTAL, MDC_REPLY_PORTAL,
1796 obddev->obd_proc_entry);
1798 if (!mds->mds_service) {
1799 CERROR("failed to start service\n");
1800 RETURN(rc = -ENOMEM);
1803 rc = ptlrpc_start_n_threads(obddev, mds->mds_service, MDT_NUM_THREADS,
1806 GOTO(err_thread, rc);
1808 mds->mds_setattr_service =
1809 ptlrpc_init_svc(MDS_NEVENTS, MDS_NBUFS,
1810 MDS_BUFSIZE, MDS_MAXREQSIZE,
1811 MDS_SETATTR_PORTAL, MDC_REPLY_PORTAL,
1812 mds_handle, "mds_setattr",
1813 obddev->obd_proc_entry);
1814 if (!mds->mds_setattr_service) {
1815 CERROR("failed to start getattr service\n");
1816 GOTO(err_thread, rc = -ENOMEM);
1819 rc = ptlrpc_start_n_threads(obddev, mds->mds_setattr_service,
1820 MDT_NUM_THREADS, "ll_mdt_attr");
1822 GOTO(err_thread2, rc);
1824 mds->mds_readpage_service =
1825 ptlrpc_init_svc(MDS_NEVENTS, MDS_NBUFS,
1826 MDS_BUFSIZE, MDS_MAXREQSIZE,
1827 MDS_READPAGE_PORTAL, MDC_REPLY_PORTAL,
1828 mds_handle, "mds_readpage",
1829 obddev->obd_proc_entry);
1830 if (!mds->mds_readpage_service) {
1831 CERROR("failed to start readpage service\n");
1832 GOTO(err_thread2, rc = -ENOMEM);
1835 rc = ptlrpc_start_n_threads(obddev, mds->mds_readpage_service,
1836 MDT_NUM_THREADS, "ll_mdt_rdpg");
1839 GOTO(err_thread3, rc);
1844 ptlrpc_unregister_service(mds->mds_readpage_service);
1846 ptlrpc_unregister_service(mds->mds_setattr_service);
1848 ptlrpc_unregister_service(mds->mds_service);
1853 static int mdt_cleanup(struct obd_device *obddev, int flags)
1855 struct mds_obd *mds = &obddev->u.mds;
1858 ptlrpc_stop_all_threads(mds->mds_readpage_service);
1859 ptlrpc_unregister_service(mds->mds_readpage_service);
1861 ptlrpc_stop_all_threads(mds->mds_setattr_service);
1862 ptlrpc_unregister_service(mds->mds_setattr_service);
1864 ptlrpc_stop_all_threads(mds->mds_service);
1865 ptlrpc_unregister_service(mds->mds_service);
1870 static struct dentry *mds_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr, void *data)
1872 struct obd_device *obd = data;
1875 fid.generation = gen;
1876 return mds_fid2dentry(&obd->u.mds, &fid, NULL);
1879 struct lvfs_callback_ops mds_lvfs_ops = {
1880 l_fid2dentry: mds_lvfs_fid2dentry,
1883 /* use obd ops to offer management infrastructure */
1884 static struct obd_ops mds_obd_ops = {
1885 o_owner: THIS_MODULE,
1886 o_attach: mds_attach,
1887 o_detach: mds_detach,
1888 o_connect: mds_connect,
1889 o_init_export: mds_init_export,
1890 o_destroy_export: mds_destroy_export,
1891 o_disconnect: mds_disconnect,
1893 o_precleanup: mds_precleanup,
1894 o_cleanup: mds_cleanup,
1895 o_postrecov: mds_postrecov,
1896 o_statfs: mds_obd_statfs,
1897 o_iocontrol: mds_iocontrol,
1898 o_create: mds_obd_create,
1899 o_destroy: mds_obd_destroy,
1900 o_llog_init: mds_llog_init,
1901 o_llog_finish: mds_llog_finish,
1902 o_notify: mds_notify,
1905 static struct obd_ops mdt_obd_ops = {
1906 o_owner: THIS_MODULE,
1907 o_attach: mdt_attach,
1908 o_detach: mdt_detach,
1910 o_cleanup: mdt_cleanup,
1913 static int __init mds_init(void)
1915 struct lprocfs_static_vars lvars;
1917 lprocfs_init_multi_vars(0, &lvars);
1918 class_register_type(&mds_obd_ops, lvars.module_vars, LUSTRE_MDS_NAME);
1919 lprocfs_init_multi_vars(1, &lvars);
1920 class_register_type(&mdt_obd_ops, lvars.module_vars, LUSTRE_MDT_NAME);
1921 ldlm_register_intent(ldlm_intent_policy);
1926 static void /*__exit*/ mds_exit(void)
1928 ldlm_unregister_intent();
1929 class_unregister_type(LUSTRE_MDS_NAME);
1930 class_unregister_type(LUSTRE_MDT_NAME);
1933 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
1934 MODULE_DESCRIPTION("Lustre Metadata Server (MDS)");
1935 MODULE_LICENSE("GPL");
1937 module_init(mds_init);
1938 module_exit(mds_exit);