Re: [Qemu-devel] [PATCH v2] block/ssh: Avoid segfault if inet_connect doesn't set errno.

[Kevin Wolf]

Am 22.07.2015 um 15:10 hat Daniel P. Berrange geschrieben:
> On Wed, Jul 22, 2015 at 02:07:22PM +0100, Richard W.M. Jones wrote:
> > On some (but not all) systems:
> > 
> >   $ qemu-img create -f qcow2 overlay -b ssh://xen/
> >   Segmentation fault
> > 
> > It turns out this happens when inet_connect returns -1 in the
> > following code, but errno == 0.
> > 
> >   s->sock = inet_connect(s->hostport, errp);
> >   if (s->sock < 0) {
> >       ret = -errno;
> >       goto err;
> >   }
> > 
> > In the test case above, no host called "xen" exists, so getaddrinfo fails.
> > 
> > On Fedora 22, getaddrinfo happens to set errno = ENOENT (although it
> > is *not* documented to do that), so it doesn't segfault.
> > 
> > On RHEL 7, errno is not set by the failing getaddrinfo, so ret =
> > -errno = 0, so the caller doesn't know there was an error and
> > continues with a half-initialized BDRVSSHState struct, and everything
> > goes south from there, eventually resulting in a segfault.
> > 
> > Fix this by setting ret to -EINVAL.  The real error is saved in the
> > Error** errp struct, so it is printed correctly:
> > 
> >   $ ./qemu-img create -f qcow2 overlay -b ssh://xen/
> >   qemu-img: overlay: address resolution failed for xen:22: No address 
> > associated with hostname
> > 
> > Signed-off-by: Richard W.M. Jones <address@hidden>
> > Reported-by: Jun Li
> > BZ: https://bugzilla.redhat.com/show_bug.cgi?id=1147343
> > ---
> >  block/ssh.c | 2 +-
> >  1 file changed, 1 insertion(+), 1 deletion(-)
> > 
> > diff --git a/block/ssh.c b/block/ssh.c
> > index aebb18c..8d4dc2a 100644
> > --- a/block/ssh.c
> > +++ b/block/ssh.c
> > @@ -563,7 +563,7 @@ static int connect_to_ssh(BDRVSSHState *s, QDict 
> > *options,
> >      /* Open the socket and connect. */
> >      s->sock = inet_connect(s->hostport, errp);
> >      if (s->sock < 0) {
> > -        ret = -errno;
> > +        ret = -EINVAL;
> >          goto err;
> 
> There are a reasonable number of other uses of inet_connect() in QEMU,
> so can't we fix inet_connect() itself to set EINVAL in the error case
> instead of just fixing one caller.

None of the other callers try to use errno, which isn't even documented
as part of the inet_connect() interface. So I think setting errno inside
inet_connect() would be wrong.

If we were to change anything, we might consider returning -EINVAL
instead of -1, though I'm not sure how useful that change would be. We
would still have to fix this call in the ssh block driver.

Kevin