Re: [PATCH 1/5] ftrace: use module notifier for function tracer

[Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>]

* Anders Kaseorg (andersk@MIT.EDU) wrote:
> On Wed, 22 Apr 2009, Masami Hiramatsu wrote:
> > Could you also add "use text_poke() on x86" to your plan? :-)
> 
> That should be possible now that it is usable inside stop_machine().  It 
> would be nicer to have an API that isn’t x86-specific, though.
> 
> Another issue is that Ksplice supports patching rodata as well as text, 
> and text_poke() does not support changes bigger than PAGE_SIZE.  Though 
> perhaps text_poke() is not the right function for rodata patches anyway.
> 
> > Hmm, I can't agree that we allow module to modify kernel text...
> > Is that possible to separate kernel-text swapping routine from Ksplice
> > module? In that case, we don't need to export text_mutex.
> 
> No, it’s not enough for Ksplice to lock the kernel text only while 
> actively swapping it.  We also need to prevent changes to the kernel text 
> while Ksplice is doing run-pre matching to ensure safety.  This means that 
> Ksplice wants to hold text_mutex for essentially the entire time it’s 
> running.
> 

how about a kernel Ksplice API which lets your patch modules get
their handlers executed under the right execution context ? E.g. :

int do_ksplice(void (*pre_hook)(struct blah *context),
	void (*code_update_hook)(struct blah *context),
	void (*post_hook)(struct blah *context));

Which returns either 0 or -ESOMETHING. It would take care of locking and
everything within core kernel code. The "patch" modules would simply
provide the functions that does the code updates, assuming that the
do_ksplice function makes sure to provide correct locking semantics.
This way, the kernel code-base would contain the tricky locking bits.
It seems much better than exporting the text_mutex to modules.

Mathieu



> Anders
> 

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