Re: [PATCH 2/2] kgdb,debug_core,kgdbts: End DEBUG_RODATA limitation using kprobe breakpoints

[Jason Wessel <jason.wessel@windriver.com>]

On 03/21/2012 09:53 PM, Masami Hiramatsu wrote:
> (2012/03/22 2:55), Jason Wessel wrote:
>> There has long been a limitation using software breakpoints with a
>> kernel compiled with CONFIG_DEBUG_RODATA.  The kprobe breakpoint code
>> has its own text_poke() function which accommodates writing a
>> breakpoint into a read-only page.  The debug_core can make use of the
>> text_poke() capabilities by using the kprobes API, specifically
>> arch_arm_kprobe() and arch_disarm_kprobe().  For now it is safe to use
>> a single statically allocated kprobe structure to call the kprobes API
>> because the debug_core breakpoint API is only used when the kernel is
>> in the debug state.
> 
> You might misunderstand it. arch_*_kprobe() are not open APIs.
> Those are kprobes internal APIs (which means that those functions
> should be used only by kprobes).
> 


I was looking for an interface that solved the problem, without having
to use text_poke directly which is arch specific.  Eventually I would
like to use the kprobes high level API, but it cannot not be used
without taking a mutex presently.  This is a separate problem to deal
with at a later time, because the generic use of kprobes would be
aimed at having robust single stepping.

> 
>> The debug_core will first attempt to use the traditional
>> probe_kernel_write(), and next try using a kprobe breakpoint.  The
>> kgdb test suite was updated to run all the software breakpoint tests
>> when using a kernel with built with CONFIG_DEBUG_RODATA.
>>
>> Signed-off-by: Jason Wessel <jason.wessel@windriver.com>
> 
> Nak.
> 
> No, please don't use kprobes internal function this way, because
> you can't ensure that the arch_arm_kprobe() has no side-effect.
> 
> Why don't you use text_poke()? I see that the text_poke()
> is only for x86, but you already have arch/x86/kernel/kgdb.c for
> making your own wrapper function.

I will use the arch specific provision to override the
kgdb_arch_set_breakpoint() and use the text_poke() directly.

Eventually I would like to use the same software breakpoint
reservation system as kprobes, and that would happen if kgdb ever
starts using kprobes for single stepping.  A few years back we solved
the problem for hardware breakpoints reservations where the kernel
debugger, perf, ptrace, and kprobes all play nice.


Thanks,
Jason.