nfs: remove extraneous and problematic calls to nfs_clear_request
authorTrond Myklebust <Trond.Myklebust@netapp.com>
Wed, 8 Dec 2010 03:39:17 +0000 (22:39 -0500)
committerTrond Myklebust <Trond.Myklebust@netapp.com>
Wed, 8 Dec 2010 04:02:44 +0000 (23:02 -0500)
When a nfs_page is freed, nfs_free_request is called which also calls
nfs_clear_request to clean out the lock and open contexts and free the
pagecache page.

However, a couple of places in the nfs code call nfs_clear_request
themselves. What happens here if the refcount on the request is still high?
We'll be releasing contexts and freeing pointers while the request is
possibly still in use.

Remove those bare calls to nfs_clear_context. That should only be done when
the request is being freed.

Note that when doing this, we need to watch out for tests of req->wb_page.
Previously, nfs_set_page_tag_locked() and nfs_clear_page_tag_locked()
would check the value of req->wb_page to figure out if the page is mapped
into the nfsi->nfs_page_tree. We now indicate the page is mapped using
the new bit PG_MAPPED in req->wb_flags .

Reported-by: Jeff Layton <jlayton@redhat.com>
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
fs/nfs/pagelist.c
fs/nfs/read.c
fs/nfs/write.c
include/linux/nfs_page.h

index 137b549..b68536c 100644 (file)
@@ -115,7 +115,7 @@ int nfs_set_page_tag_locked(struct nfs_page *req)
 {
        if (!nfs_lock_request_dontget(req))
                return 0;
-       if (req->wb_page != NULL)
+       if (test_bit(PG_MAPPED, &req->wb_flags))
                radix_tree_tag_set(&NFS_I(req->wb_context->path.dentry->d_inode)->nfs_page_tree, req->wb_index, NFS_PAGE_TAG_LOCKED);
        return 1;
 }
@@ -125,7 +125,7 @@ int nfs_set_page_tag_locked(struct nfs_page *req)
  */
 void nfs_clear_page_tag_locked(struct nfs_page *req)
 {
-       if (req->wb_page != NULL) {
+       if (test_bit(PG_MAPPED, &req->wb_flags)) {
                struct inode *inode = req->wb_context->path.dentry->d_inode;
                struct nfs_inode *nfsi = NFS_I(inode);
 
index e4b62c6..aedcaa7 100644 (file)
@@ -152,7 +152,6 @@ static void nfs_readpage_release(struct nfs_page *req)
                        (long long)NFS_FILEID(req->wb_context->path.dentry->d_inode),
                        req->wb_bytes,
                        (long long)req_offset(req));
-       nfs_clear_request(req);
        nfs_release_request(req);
 }
 
index 4c14c17..10d648e 100644 (file)
@@ -390,6 +390,7 @@ static int nfs_inode_add_request(struct inode *inode, struct nfs_page *req)
                if (nfs_have_delegation(inode, FMODE_WRITE))
                        nfsi->change_attr++;
        }
+       set_bit(PG_MAPPED, &req->wb_flags);
        SetPagePrivate(req->wb_page);
        set_page_private(req->wb_page, (unsigned long)req);
        nfsi->npages++;
@@ -415,6 +416,7 @@ static void nfs_inode_remove_request(struct nfs_page *req)
        spin_lock(&inode->i_lock);
        set_page_private(req->wb_page, 0);
        ClearPagePrivate(req->wb_page);
+       clear_bit(PG_MAPPED, &req->wb_flags);
        radix_tree_delete(&nfsi->nfs_page_tree, req->wb_index);
        nfsi->npages--;
        if (!nfsi->npages) {
@@ -422,7 +424,6 @@ static void nfs_inode_remove_request(struct nfs_page *req)
                iput(inode);
        } else
                spin_unlock(&inode->i_lock);
-       nfs_clear_request(req);
        nfs_release_request(req);
 }
 
index f8b60e7..d55cee7 100644 (file)
@@ -29,6 +29,7 @@
  */
 enum {
        PG_BUSY = 0,
+       PG_MAPPED,
        PG_CLEAN,
        PG_NEED_COMMIT,
        PG_NEED_RESCHED,