[patch 0/3] Re: 2.6.12-rc2-mm3

[Stas Sergeev <stsp@aknet.ru>]

Hello.

Andrew Morton wrote:
> OK, the `int $3' is part of the CONFIG_TRAP_BAD_SYSCALL_EXITS thing which I
> never use.
> I'm not sure what problem is actually being reported here, now you mention it.
The problem being reported here is that
CONFIG_TRAP_BAD_SYSCALL_EXITS does nothing
but locking up your machine. Actually the
bug was so obvious, that I had real troubles
finding it (the obvious things are difficult
to spot), so I found some more bugs in a
mean time.
What was the bug? GET_THREAD_INFO(%ebp) was
missing before TI_preempt_count(%ebp), hence
the accesses beyond the stack. I'll have
troubles beleiving this code worked without
a lock-ups for someone sometimes.
I fixed it differently though. The subsequent
patches are addressing the issue.

> Yup.  But are you sure that the "+ 8" is correct, given these offsets are<>
> larger than that?
I don't think they are indeed larger. The %esp
points to "struct pt_regs", which is 60 bytes
in size, and without the xss/esp = 52. Adding
8 to this gives 60, so 56(+3) looks safe to me.

> Probably it decided that some syscall got a "bad exit".  Disable
> CONFIG_TRAP_BAD_SYSCALL_EXITS.
Yes, that's the fix too.

>> > -	p->thread.esp0 = (unsigned long) (childregs+1) - 8;
>> > +	p->thread.esp0 = (unsigned long) (childregs+1) - 15;
>> 15 is somewhat nasty - it will make the
>> stack unaligned, should better be 16 I
>> think.
> ?  It's still 4-byte-aligned.
I don't see your point. Why do you think that
I substract the stack pointer by 32 bytes, for
example? I literally substract it by 8 bytes,
you propose to substract it by 15 *bytes* (not
dwords), so why would it still be aligned?
But anyway, fortunately this bug is not about
the esp0 stuff at all.

> I'm suspecting this is all due to CONFIG_TRAP_BAD_SYSCALL_EXITS taking the
> debug trap..
Sure. And that looks silly. I removed "int $3".
Patches follow. Seems to work reliable now,
but I don't know how to test it since there
seem to be no such an offending syscalls here
to test.