Re: [PATCH 0/2 v2] Syscalls tracing

[Mathieu Desnoyers]

* Ingo Molnar (mingo@xxxxxxx) wrote:
> 
> * Frederic Weisbecker <fweisbec@xxxxxxxxx> wrote:
> 
> > And actually I don't think two copy_from_user will really change a 
> > lot the tracing throughput.
> 
> Correct. It's already in the CPU cache so it is a performance 
> non-issue and essentially for free. Copy avoidance is only an issue 
> when touchig cache-cold data.
> 
> ( Yes, a few cycles could be shaven off but the beauty of 
>   all-encompassing non-source-code-invasive syscall tracing covering 
>   hundreds of syscalls straight away trumps those concerns. )
> 

I agree. I just wanted to make sure we agreed on the tradeoff here. I
also think hitting data already in cache-lines a second time with
copy_from_user should not be a big concern.

> > The idea would be now to join the syscalls metadata with such 
> > quick handlers. We will have to think about how to join these in a 
> > proper way.
> 
> We could allow per syscall tracepoints via the attribute table. The 
> call signature could be a standard:
> 
>    long sys_call(unsinged long arg1, unsigned long arg2,
> 		 unsigned long arg3, unsigned long arg4,
> 		 unsigned long arg5, unsigned long arg6);
> 
> This would allow interested plugins/tools to install a system call 
> specific callback. (we might allow two tracepoints - one before and 
> one after the syscall)
> 
> The registration API could be driven by the name or by the syscall 
> index - NR_sys_open or so.

Hrm, given the syscalls are defined with their number of arguments
with the SYSCALL_DEFINE* macros, then we could create, in syscalls.h
(example from open.c) :

SYSCALL_DECLARE2(statfs, const char __user *, pathname, struct statfs __user *, buf))

SYSCALL_DECLARE3(statfs64, const char __user *, pathname, size_t, sz, struct statfs64 __user *, buf)

creating SYSCALL_DECLARE0 to 6, which would map to a tracepoint
declaration _and_ a syscall prototype, e.g.

#define __SC_ARGS1(t1, a1)      a1
#define __SC_ARGS2(t2, a2, ...) a2, __SC_ARGS1(__VA_ARGS__)
#define __SC_ARGS3(t3, a3, ...) a3, __SC_ARGS2(__VA_ARGS__)
#define __SC_ARGS4(t4, a4, ...) a4, __SC_ARGS3(__VA_ARGS__)
#define __SC_ARGS5(t5, a5, ...) a5, __SC_ARGS4(__VA_ARGS__)
#define __SC_ARGS6(t6, a6, ...) a6, __SC_ARGS5(__VA_ARGS__)

#define SYSCALL_DECLARE2(name, ...) SYSCALL_DECLAREx(2, _##name, __VA_ARGS__) 

#define SYSCALL_DECLAREx(x, name, ...)				\
	long sys##name(__SC_DECL##x(__VA_ARGS__));		\
	DECLARE_TRACE(sys_##name,				\
			TP_PROTO(__SC_DECL##x(__VA_ARGS__)),	\
				TP_ARGS(__SC_ARGS##x(__VA_ARGS__)))

Those could be declared in a system-wide header (syscalls.h ?) which
would be included by each files using SYSCALL_DEFINE*. Those
declarations would declare the tracepoints and therefore make sure we
spot any SYSCALL_DEFINE* change at compile-time, and we could create a
tracing module which would contain the callbacks that would register on
those syscall tracepoint declarations. This would all be type-safe,
which is a very nice thing to have, even if we don't expect the system
calls to change often at all.

Mathieu

> 
> 	Ingo

-- 
Mathieu Desnoyers
OpenPGP key fingerprint: 8CD5 52C3 8E3C 4140 715F  BA06 3F25 A8FE 3BAE 9A68
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/