Interesting bug :-)
Just so I get this right - this could just as well be fixed with instead
changing msg_controllen to this:
mh.msg_controllen = CMSG_SPACE(sizeof(int) * nfd);
That said, your version works equally well and is slightly cleaner. I
don't recall we have any problems with VLAs.
Perhaps adding a comment to your finding and a registered kernel
bug/patch/etc would be helpful.
Anyhow, good work, so:
Acked-by: David Henningsson <diwic at ubuntu.com>
On 2016-11-15 17:48, Ahmed S. Darwish wrote:
> Users reported audio breakage for 32-bit pulse clients connected
> to a 64-bit server over memfds. Investigating the issue further,
> the problem is twofold:
>
> 1. iochannel's file-descriptor passing code is liberal in what it
> issues: produced ancillary data object's "data" section exceeds
> length field. How such an extra space is handled is a grey area
> in the POSIX.1g spec, the IETF RFC #2292 "Advanced Sockets API
> for IPv6" memo, and the cmsg(3) manpage.
>
> 2. A 64-bit kernel handling of such extra space differs by whether
> the app is 64-bit or 32-bit. For 64-bit apps, the kernel
> smartly ducks the issue. For 32-bit apps, an -EINVAL is
> directly returned; that's due to a kernel CMSG header traversal
> bug in the networking stack "32-bit sockets emulation layer".
>
> Compare Linux Kernel's socket.h cmsg_nxthdr() code and the
> 32-bit emulation layer version of it at net/compat.c
> cmsg_compat_nxthdr() for further info. Notice how the former
> graciously ignores incomplete CMSGs while the latter _directly_
> complains about them -- as of kernel version 4.9-rc5.
>
> (A kernel patch is to be submitted)
>
> Details:
>
> iochannel typically uses sendmsg() for passing FDs & credentials.
> From RFC 2292, sendmsg() control data is just a heterogeneous
> array of embedded ancillary objects that can differ in length.
> Linguistically, a "control message" is an ancillary data object.
>
> For example, below is a sendmsg() "msg_control" containing two
> ancillary objects:
>
> |<---------------------- msg_controllen---------------------->|
> | |
> |<--- ancillary data object -->|<----- ancillary data object->|
> |<------- CMSG_SPACE() ------->|<------- CMSG_SPACE() ------->|
> | | |
> |<-------- cmsg_len ------->| |<-------- cmsg_len ------->| |
> |<------- CMSG_LEN() ------>| |<------- CMSG_LEN() ------>| |
> | | | | |
> +-----+-----+-----+--+------+--+-----+-----+-----+--+------+--+
> |cmsg_|cmsg_|cmsg_|XX|cmsg_ |XX|cmsg_|cmsg_|cmsg_|XX|cmsg_ |XX|
> |len |level|type |XX|data[]|XX|len |level|type |XX|data[]|XX|
> +-----+-----+-----+--+------+--+-----+-----+-----+--+----+-+--+
> ^^^^^^^ Ancil Object #1 ^^^^^^^ Ancil Object #2
> (control message) (control message)
> ^
> |
> +--- sendmsg() "msg_control" points here
>
> Problem is, while passing FDs, iochannel's code try to avoid
> variable-length arrays by creating a single cmsg object that can
> fit as much FDs as possible:
>
> union {
> struct cmsghdr hdr;
> uint8_t data[CMSG_SPACE(sizeof(int) * MAX_ANCIL_DATA_FDS)];
> } cmsg; ^^^^^^^^^^^^^^^^^^
>
> Most of the time though the number of FDs to be passed is less
> than the maximum above, thus "cmsg_len" is set to the _actual_ FD
> array size:
>
> cmsg.hdr.cmsg_len = CMSG_LEN(sizeof(int) * nfd);
> ^^^
> This inconsistency tricks the kernel into thinking that we have 2
> ancillay data objects instead of one! First cmsg is valid as
> intended, but the second is instantly _corrupt_ since it has a
> cmsg_len size of 0 -- thus failing kernel's CMSG_OK() tests.
>
> For 32-bit apps on a 32-bit kernel, and 64-bit apps over a 64-bit
> one, the kernel's own CMSG header traversal macros just ignore the
> second "incomplete" cmsg. For 32-bit apps over a 64-bit kernel
> though, the kernel 32-bit socket emulation macros does not forgive
> such incompleteness and directly complains of invalid args (due to
> a subtle bug).
>
> Avoid this ugly problem, which can also bite us in a pure 64-bit
> environment if MAX_ANCIL_DATA_FDS got extended to 5 FDs, by
> setting "cmsg_data[]" array size to "cmsg_len".
>
> BugLink: https://bugs.freedesktop.org/show_bug.cgi?id=97769
>
> Signed-off-by: Ahmed S. Darwish <darwish.07 at gmail.com>
> ---
>
> Hopefully variable-length arrays won't be problematic in esoteric
> build environments?
>
> src/pulsecore/iochannel.c | 4 +++-
> 1 file changed, 3 insertions(+), 1 deletion(-)
>
> diff --git a/src/pulsecore/iochannel.c b/src/pulsecore/iochannel.c
> index e62750b..8ace297 100644
> --- a/src/pulsecore/iochannel.c
> +++ b/src/pulsecore/iochannel.c
> @@ -346,6 +346,8 @@ ssize_t pa_iochannel_write_with_creds(pa_iochannel*io, const void*data, size_t l
> return r;
> }
>
> +/* For more details on FD passing, check the cmsg(3) manpage
> + * and IETF RFC #2292: "Advanced Sockets API for IPv6" */
> ssize_t pa_iochannel_write_with_fds(pa_iochannel*io, const void*data, size_t l, int nfd, const int *fds) {
> ssize_t r;
> int *msgdata;
> @@ -353,7 +355,7 @@ ssize_t pa_iochannel_write_with_fds(pa_iochannel*io, const void*data, size_t l,
> struct iovec iov;
> union {
> struct cmsghdr hdr;
> - uint8_t data[CMSG_SPACE(sizeof(int) * MAX_ANCIL_DATA_FDS)];
> + uint8_t data[CMSG_SPACE(sizeof(int) * nfd)];
> } cmsg;
>
> pa_assert(io);
>
