Re: [PATCH 11/12] hw-breakpoints: ftrace plugin for kernel symbol tracing using HW Breakpoint interfaces

[David Daney <ddaney@caviumnetworks.com>]

Frederic Weisbecker wrote:
> On Tue, Jun 02, 2009 at 04:12:08PM -0700, David Daney wrote:
>> Frederic Weisbecker wrote:
>>> From: K.Prasad <prasad@linux.vnet.ibm.com>
>>>
>>> This patch adds an ftrace plugin to detect and profile memory access over kernel
>>> variables. It uses HW Breakpoint interfaces to 'watch memory addresses.
>>>
>>> Signed-off-by: K.Prasad <prasad@linux.vnet.ibm.com>
>>> Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
>>> ---
>>>  kernel/trace/Kconfig          |   21 ++
>>>  kernel/trace/Makefile         |    1 +
>>>  kernel/trace/trace.h          |   23 ++
>>>  kernel/trace/trace_ksym.c     |  525 +++++++++++++++++++++++++++++++++++++++++
>>>  kernel/trace/trace_selftest.c |   53 ++++
>>>  5 files changed, 623 insertions(+), 0 deletions(-)
>>>  create mode 100644 kernel/trace/trace_ksym.c
>> [...]
>>> +	entry->ksym_hbp->info.name = ksymname;
>>> +	entry->ksym_hbp->info.type = op;
>>> +	entry->ksym_addr = entry->ksym_hbp->info.address = addr;
>>> +#ifdef CONFIG_X86
>>> +	entry->ksym_hbp->info.len = HW_BREAKPOINT_LEN_4;
>>> +#endif
>> What if the symbol referred to an object of size other than 4?  This  
>> would clearly be incorrect in that case.
>>
>>
>>> +	entry->ksym_hbp->triggered = (void *)ksym_hbp_handler;
>>> +
>>> +	ret = register_kernel_hw_breakpoint(entry->ksym_hbp);
>> I hate to sound like a broken record, but could some one explain to me  
>> again why it is a good idea to design a new API that requires processor  
>> specific #ifdefs to be sprinkled all around generic kernel code?
>>
>> Back in:
>> http://lkml.org/lkml/2008/12/4/329
>> and
>> http://lkml.org/lkml/2009/5/21/189
>>
>> I raised doubts about this hw-breakpoint thing being generic and the  
>> responses made think that the processor specific portions would be  
>> isolated in the processor specific parts of the kernel.  I now see that  
>> I was wrong.
>>
>> When we add sparc, MIPS, ppc...  Support it would be nice to not have to  
>> add all our own #ifdefs to this, but instead have a generic interface  
>> that will not need changes.
>>
>> David Daney
> 
> I was discussing about it with Prasad few hours ago :)
> 
> The fact is that archs support the hardware breakpoints in
> very different ways each.
> Some of them support read breakpoint, others not (x86).
> Some support addresses range, others (x86).
> 
> But still it would be nice to gather the most common
> breakpoints operations through a real generic wrapper
> that relies on arch specific implmentation in
> background.
> 

Thanks for taking the time to think about it.  I do think this patch set 
will be useful.

One thing I was thinking was that when I am using gdb, all I have to do 
is say 'watch symbol' and the differences in hardware breakpoint 
facilities are abstracted away.  Having similar abstraction in this 
kernel API should be possible as well.

David Daney