ftrace introduces instability into kernel 2.6.27(-rc2,-rc3)

[Mathieu Desnoyers]

* Steven Rostedt (rostedt at goodmis.org) wrote:
> 
> 
> On Mon, 18 Aug 2008, Mathieu Desnoyers wrote:
> 
> > * Steven Rostedt (rostedt at goodmis.org) wrote:
> > > 
> > > On Tue, 19 Aug 2008, Benjamin Herrenschmidt wrote:
> > > 
> > > > 
> > > > > Hmm, this was originally copied from x86, where we did a cmpxchg, but that 
> > > > > is probably not needed since all of this is done in kstop_machine. Also, 
> > > > > only the "get" is needed. If we don't fault there, we wont fault on the 
> > > > > put (unless we have permissions wrong, and that would be a bug).
> > > > 
> > > > Would it ? How do we make sure the kernel text is mapped writeable ?
> > > 
> > > We map it writeable if FTRACE is enabled.
> > > 
> > 
> > Argh. See text_poke(). It's there exactly for this purpose on x86.
> > 
> 
> OK, I just tried text_poke and it unfortunately fails. The problem is that 
> it requires that the text you are changing is aligned and fits on one 
> page. We have no control over that.
> 
> -- Steve
> 

Ok, there are two cases where it's ok :

1 - in stop_machine, considering we are not touching code executed in
NMI handlers.
2 - when using my replace_instruction_safe() which uses a temporary
breakpoint when doing the instruction replacement.

In those cases you could use text_poke_early().

See
http://git.kernel.org/?p=linux/kernel/git/compudj/linux-2.6-lttng.git;a=blob;f=arch/x86/kernel/immediate.c;h=7789e2c75bf03e645f15759d5dff0c1698493f92;hb=HEAD

For a use example. Basically it looks like :


360                 pages[0] = virt_to_page((void *)bypass_eip);
361                 vaddr = vmap(pages, 1, VM_MAP, PAGE_KERNEL);
362                 BUG_ON(!vaddr);
363                 text_poke_early(&vaddr[bypass_eip & ~PAGE_MASK],
364                         (void *)addr, size);
365                 /*
366                  * Fill the rest with nops.
367                  */
368                 len = NR_NOPS - size;
369                 add_nops((void *)
370                         &vaddr[(bypass_eip & ~PAGE_MASK) + size],
371                         len);
372                 print_dbg_bytes("inserted nops",
373                         &vaddr[(bypass_eip & ~PAGE_MASK) + size], len);
374                 vunmap(vaddr);

Mathieu


-- 
Mathieu Desnoyers
OpenPGP key fingerprint: 8CD5 52C3 8E3C 4140 715F  BA06 3F25 A8FE 3BAE 9A68