[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index][Old Index]
Re: Sane support for SMP kernel profiling
In article <20101211041353.GA20866%coyotepoint.com@localhost>,
Thor Simon <tls%coyotepoint.com@localhost> wrote:
>We've fixed SMP kernel profiling, which worked poorly at best (particularly
>on systems with high HZ) since a lock was taken and released around every
>single entry to mcount. Thanks to Andy for the suggestion as to how.
>The following patch includes:
> * A convenience change to build.sh to make it simpler to
> cross-build profiling kernels (-p)
> * Per-CPU profiling space
> * Changes to kgmon to fetch profiling data per-CPU rather than
> globally on MP systems. gprof already knows how to merge
> multiple profiles so there seems little reason to include this
> functionaity in the kernel code or kgmon, both of which should
> be as small and simple as possible.
> * The MCOUNT macro is cleaned up to an inline function on
> architectures which define PROFILE_TASTEFUL_FUNCTIONS. The idea is
> to eventually eliminate the macro.
> * The kernel and userspace mcount implementations are separated
> into two source files to clean up (some of) a really ugly #ifdef
> * A correct, non-locking mcount implementation is included for
> i386. Other ports still build and should be no worse than they
> were before, but also no better.
>Patch is against netbsd-5. Comments much appreciated.
Perhaps move it from common to sys and libc since nothing is shared anymore.
Main Index |
Thread Index |