kernfs: invoke kernfs_unmap_bin_file() directly from __kernfs_remove()
authorTejun Heo <tj@kernel.org>
Fri, 10 Jan 2014 13:57:23 +0000 (08:57 -0500)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Fri, 10 Jan 2014 21:48:08 +0000 (13:48 -0800)
commitf601f9a2bf7dc1f7ee18feece4c4e2fc6845d6c4
treef8213b3d7542c85c42cb1a20668098609d65762c
parent45a140e587f3d32d8d424ed940dffb61e1739047
kernfs: invoke kernfs_unmap_bin_file() directly from __kernfs_remove()

kernfs_unmap_bin_file() is supposed to unmap all memory mappings of
the target file before kernfs_remove() finishes; however, it currently
is being called from kernfs_addrm_finish() and has the same race
problem as the original implementation of deactivation when there are
multiple removers - only the remover which snatches the node to its
addrm_cxt->removed list is guaranteed to wait for its completion
before returning.

It can be fixed by moving kernfs_unmap_bin_file() invocation from
kernfs_addrm_finish() to __kernfs_remove().  The function may be
called multiple times but that shouldn't do any harm.

We end up dropping kernfs_mutex in the removal loop and the node may
be removed inbetween by someone else.  kernfs_unlink_sibling() is
updated to test whether the node has already been removed and return
accordingly.  __kernfs_remove() in turn performs post-unlinking
cleanup only if it actually unlinked the node.

KERNFS_HAS_MMAP test is moved out of the unmap function into
__kernfs_remove() so that we don't unlock kernfs_mutex unnecessarily.
While at it, drop the now meaningless "bin" qualifier from the
function name.

v2: Rewritten to fit the v2 restructuring of removal path.  HAS_MMAP
    test relocated.

Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
fs/kernfs/dir.c
fs/kernfs/file.c
fs/kernfs/kernfs-internal.h