On Tue, 2013-07-09 at 11:33 +0100, David Howells wrote: > Milosz Tanski <[email protected]> wrote: > > > It looks like both the cifs and NFS code do not bother with any > > locking around cifs_fscache_set_inode_cookie. Is there no concern over > > multiple open() calls racing to create the cookie in those > > filesystems? > > Yeah... That's probably wrong. AFS obviates the need for special locking by > doing it in afs_iget(). > > Hmmm... I think I've just spotted what might be the cause of pages getting > marked PG_fscache whilst belonging to the allocator. > > void nfs_fscache_set_inode_cookie(struct inode *inode, struct file > *filp) > { > if (NFS_FSCACHE(inode)) { > nfs_fscache_inode_lock(inode); > if ((filp->f_flags & O_ACCMODE) != O_RDONLY) > nfs_fscache_disable_inode_cookie(inode); > else > nfs_fscache_enable_inode_cookie(inode); > nfs_fscache_inode_unlock(inode); > } > } > > can release the cookie whilst reads are in progress on it when an inode being > read suddenly changes to an inode being written. We need some sort of > synchronisation on that there.
Change fscache_uncache_all_inode_pages to always take the page lock, check the value of page->mapping, and wait for outstanding writes to finish before trying to release PG_fscache? -- Trond Myklebust Linux NFS client maintainer NetApp [email protected] www.netapp.com N�����r��y����b�X��ǧv�^�){.n�+���z�]z���{ay� ʇڙ�,j��f���h���z� �w��� ���j:+v���w�j�m��������zZ+�����ݢj"��!�i
