handle_vm86_fault may fault? was Re: Kernel Oops in 2.6.0-test2-mm4

[Matt Mackall <mpm@selenic.com>]

On Tue, Aug 05, 2003 at 08:10:56PM -0400, s0be wrote:
> On Tue, 5 Aug 2003 19:05:24 -0500
> Matt Mackall <mpm@selenic.com> wrote:
> 
> > On Tue, Aug 05, 2003 at 05:05:58PM -0400, s0be wrote:
> > > here's the oops from dmesg and the surrounding messages.  I'm guessing it was caused by smb, but I can't confirm it.  trying to recreate it.
> ..snip..
> > > Debug: sleeping function called from invalid context at include/asm/uaccess.h:512Call Trace:
> > >  [<c011fd3c>] __might_sleep+0x5c/0x5e
> > >  [<c010da1a>] save_v86_state+0x6a/0x200
> > >  [<c010e565>] handle_vm86_fault+0xa5/0x8c0
> > >  [<c0170a23>] dput+0x23/0x200
> > >  [<c010c030>] do_general_protection+0x0/0xa0
> > >  [<c032519f>] error_code+0x2f/0x38
> > >  [<c0324733>] syscall_call+0x7/0xb
> ..snip..
> > 
> > This is not an oops, just a debug trace that says something tried to
> > do something unsafe (namely calling copy_from_user while in_atomic()
> > was true). 
> > 
> > Looks like we've got:
> > 
> >  do_general_protection
> >   handle_vm86_fault
> >    return_to_32bit
> >     save_v86_state
> >      copy_to_user   
> > 
> > and the destination of the copy is current->thread.vm86_info->regs,
> > which is labelled __user. Presuming this is actually in userspace,
> > this could be a problem.
> > 
> 
> I'd have to agree. I wrote a script that sits here mounting a samba
> mount, then, either writing to it, and unmount, write to it and
> no,unmount, just unmount it, or nothing, pseudo randomly, and it
> pissed off my windows machine it was connecting to, but could NOT
> reproduce this problem. I can't seem to get it to happen again
> though. I'm sort of at a loss. I can provide any extra info that
> might help.

This appears to be a video driver running a card's video BIOS in vm86
mode and hitting one of many possible faults. The fault handler saves
a subset of the state to userspace with copy_to_user. This could
potentially cause a second fault and I don't see any guarantees that
this page is locked or anything.

So a) is this safe, somehow? And b) if so, what check can might_sleep use
to avoid false positives?

-- 
Matt Mackall : http://www.selenic.com : of or relating to the moon