net/9p: avoid -ERESTARTSYS leak to userspace
authorGreg Kurz <[email protected]>
Thu, 5 Apr 2018 23:19:44 +0000 (16:19 -0700)
committerLinus Torvalds <[email protected]>
Fri, 6 Apr 2018 04:36:22 +0000 (21:36 -0700)
commita85222435bd055b2d2cedc515d810a5ea6c05432
tree5cb9623e7cdb4229ab9bd90d559d5a6d895aa1fc
parentde37428638ec7d22b46c4c20b8536f3681559e2c
net/9p: avoid -ERESTARTSYS leak to userspace

If it was interrupted by a signal, the 9p client may need to send some
more requests to the server for cleanup before returning to userspace.

To avoid such a last minute request to be interrupted right away, the
client memorizes if a signal is pending, clears TIF_SIGPENDING, handles
the request and calls recalc_sigpending() before returning.

Unfortunately, if the transmission of this cleanup request fails for any
reason, the transport returns an error and the client propagates it
right away, without calling recalc_sigpending().

This ends up with -ERESTARTSYS from the initially interrupted request
crawling up to syscall exit, with TIF_SIGPENDING cleared by the cleanup
request.  The specific signal handling code, which is responsible for
converting -ERESTARTSYS to -EINTR is not called, and userspace receives
the confusing errno value:

  open: Unknown error 512 (512)

This is really hard to hit in real life.  I discovered the issue while
working on hot-unplug of a virtio-9p-pci device with an instrumented
QEMU allowing to control request completion.

Both p9_client_zc_rpc() and p9_client_rpc() functions have this buggy
error path actually.  Their code flow is a bit obscure and the best
thing to do would probably be a full rewrite: to really ensure this
situation of clearing TIF_SIGPENDING and returning -ERESTARTSYS can
never happen.

But given the general lack of interest for the 9p code, I won't risk
breaking more things.  So this patch simply fixes the buggy paths in
both functions with a trivial label+goto.

Thanks to Laurent Dufour for his help and suggestions on how to find the
root cause and how to fix it.

Link: http://lkml.kernel.org/r/[email protected]
Signed-off-by: Greg Kurz <[email protected]>
Reviewed-by: Andrew Morton <[email protected]>
Reviewed-by: Yiwen Jiang <[email protected]>
Cc: Eric Van Hensbergen <[email protected]>
Cc: Ron Minnich <[email protected]>
Cc: Latchesar Ionkov <[email protected]>
Cc: David Miller <[email protected]>
Cc: Laurent Dufour <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
net/9p/client.c