fuse: fix pipe buffer lifetime for direct_io
authorMiklos Szeredi <mszeredi@redhat.com>
Mon, 7 Mar 2022 15:30:44 +0000 (16:30 +0100)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Wed, 16 Mar 2022 13:23:42 +0000 (14:23 +0100)
commitca62747b38f59d4e75967ebf63c992de8852ca1b
treecae13d63fae78a7d12733eebffe46014508bc702
parentd60d34b4d6d1227923bcbe6fdba24c2f03ce69cf
fuse: fix pipe buffer lifetime for direct_io

commit 0c4bcfdecb1ac0967619ee7ff44871d93c08c909 upstream.

In FOPEN_DIRECT_IO mode, fuse_file_write_iter() calls
fuse_direct_write_iter(), which normally calls fuse_direct_io(), which then
imports the write buffer with fuse_get_user_pages(), which uses
iov_iter_get_pages() to grab references to userspace pages instead of
actually copying memory.

On the filesystem device side, these pages can then either be read to
userspace (via fuse_dev_read()), or splice()d over into a pipe using
fuse_dev_splice_read() as pipe buffers with &nosteal_pipe_buf_ops.

This is wrong because after fuse_dev_do_read() unlocks the FUSE request,
the userspace filesystem can mark the request as completed, causing write()
to return. At that point, the userspace filesystem should no longer have
access to the pipe buffer.

Fix by copying pages coming from the user address space to new pipe
buffers.

Reported-by: Jann Horn <jannh@google.com>
Fixes: c3021629a0d8 ("fuse: support splice() reading from fuse device")
Cc: <stable@vger.kernel.org>
Signed-off-by: Miklos Szeredi <mszeredi@redhat.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
fs/fuse/dev.c
fs/fuse/file.c
fs/fuse/fuse_i.h