Re: crash in entry.S restore_all, 2.6.12-rc2, x86, PAGEALLOC

[Stas Sergeev <stsp@aknet.ru>]

Hi,

Linus Torvalds wrote:
> Yes. But how do you have _such_ an empty stack when the interrupt comes
> in? See what I mean?
Yes, I hope so.

> IOW, that requires that the kernel stack would have
> only two words on it when the interrupt happens. How?
Well, you can simply do something like this:

--- entry.S.old1	2005-04-05 22:54:43.000000000 +0400
+++ entry.S	2005-04-06 19:35:14.000000000 +0400
@@ -179,9 +179,9 @@
 ENTRY(sysenter_entry)
 	movl TSS_sysenter_esp0(%esp),%esp
 sysenter_past_esp:
-	sti
 	pushl $(__USER_DS)
 	pushl %ebp
+	sti
 	pushfl
 	pushl $(__USER_CS)
 	pushl $SYSENTER_RETURN

And this will "elimenate" the problem
(modulo NMI and there could be other places
too, but for me it elimenates it completely).
So I don't think this is something strange.

> So I definitely think the "bug" is in your optimization,
Yes, and I think the patch I posted, can
just work, or are there the problems with
the taken forward jump on a fast path?

> I just think it 
> should be a valid optimization
But it is totally bogus, why not should it
crash? It is probably even very good that
it does:)

> and we should just make sure our kernel 
> stack is never _so_ empty that "struct pt_regs" is not safe to 
> dereference.
I guess you'll just need to adjust the tss.esp0
then, but do you really want this? Accesing
the registers that are simply not there, doesn't
sound too good I think.
Or am I still missing your point?