Re: [ 05/48] x86_64, switch_to(): Load TLS descriptors before switching DS and ES

[Ben Hutchings]

On Fri, 2015-05-15 at 15:38 +0200, Willy Tarreau wrote:
> Hi Ben,
> 
> On Fri, May 15, 2015 at 01:32:20PM +0100, Ben Hutchings wrote:
> > On Fri, 2015-05-15 at 10:05 +0200, Willy Tarreau wrote:
> > > 2.6.32-longterm review patch.  If anyone has any objections, please let me know.
> > > 
> > > ------------------
> > > 
> > > From: Andy Lutomirski <luto@xxxxxxxxxxxxxx>
> > > 
> > > commit f647d7c155f069c1a068030255c300663516420e upstream.
> > > 
> > > Otherwise, if buggy user code points DS or ES into the TLS
> > > array, they would be corrupted after a context switch.
> > > 
> > > This also significantly improves the comments and documents some
> > > gotchas in the code.
> > > 
> > > Before this patch, the both tests below failed.  With this
> > > patch, the es test passes, although the gsbase test still fails.
> > [...]
> > 
> > This depends on the changes to FPU/MMX/SSE state management that you
> > didn't apply to 2.6.32.  Note this comment:
> > 
> > 	/* Must be after DS reload */
> > 	unlazy_fpu(prev_p);
> 
> Are you sure you're not confusing with another one ? When running
> estest without this patch, I get "FAIL: ES corrupted 1000/1000 times"
> while I get "OK: ES was preserved" once applied, so it does seem to
> do what it's intended for.
>
> Also I'm not seeing any reference to the comment above in the patch
> nor around it, which leaves me confused :-/

v2.6.32.65:arch/x86/kernel/process_64.c:425:    /* Must be after DS reload */

If this comment is correct then the patch will cause a regression for
FPU state management.  The comment was introduced by:

commit 0a5ace2ab08d45cd78d7ef0067cdcd5c812ac54f
Author: Andi Kleen <ak@xxxxxxx>
Date:   Thu Oct 5 18:47:22 2006 +0200

    [PATCH] x86-64: Fix FPU corruption

Ben.

-- 
Ben Hutchings
It is impossible to make anything foolproof because fools are so ingenious.

Attachment: signature.asc
Description: This is a digitally signed message part