accf_http.ko breaks without COMPAT_IA32 (module linker problems?)

[John Baldwin]

On Tuesday 03 July 2007 04:55:39 pm Jeremy Chadwick wrote:
> On Tue, Jul 03, 2007 at 04:15:09PM -0400, John Baldwin wrote:
> > On Tuesday 03 July 2007 01:53:13 pm Jeremy Chadwick wrote:
> > > Came across this today (and is reproducable):
> > > 
> > > With COMPAT_IA32 removed from the kernel configuration, accf_http.ko
> > > breaks in a bad way.  Only shown on the console is the message "kldload:
> > > Unsupported file type" when kldload is used to load the module.  Worse,
> > > the module linker appears to load a portion of the module anyways:
> > > 
> > > # kldstat
> > > Id Refs Address            Size     Name
> > >  1    2 0xffffffff80100000 566cd0   kernel
> > >  2    1 0xffffffffb468a000 795      accf_http.ko
> > > 
> > > # kldstat -v -i 2
> > > Id Refs Address            Size     Name
> > >  2    1 0xffffffffb468a000 795      accf_http.ko
> > >         Contains modules:
> > >                 Id Name
> > >                 173 accf_http
> > > 
> > > And attempts to unload the module fail (which is where I question the
> > > reliability of the module linker (no offence intended)):
> > > 
> > > # kldunload accf_http.ko
> > > kldunload: can't unload file: Operation not supported
> > > icarus# kldunload -v -f accf_http.ko
> > > Unloading accf_http.ko, id=2
> > > kldunload: can't unload file: Operation not supported
> > > 
> > > Minor details:
> > > 
> > > -r-xr-xr-x  1 root  wheel  5544 Jul  3 10:35 /boot/kernel/accf_http.ko
> > > 
> > > /boot/kernel/accf_http.ko: ELF 64-bit LSB relocatable, AMD x86-64, 
version 1 
> > (FreeBSD), not stripped
> > > 
> > > Putting COMPAT_IA32 back in results in proper behaviour all around.
> > 
> > What is in dmseg after the kldload?  Usually if there is a missing symbol 
the 
> > kernel will print out its name in dmesg.
> 
> This is during bootup, so it happens right as rc.d/apache22 starts.
> rc.d/apache22 loads the module via apache22_http_accept_enable="yes".
> The script also redirects kldload's stderr output to /dev/null, so if
> kldload spits it out to stderr, that might be why nothing gets shown.

dmesg != stderr.  The _kernel_ prints out missing symbols to the _dmesg_, not 
to userland.

-- 
John Baldwin