Re: [PATCH][2.6] Allow x86_64 to reenable interrupts on contention

[William Lee Irwin III]

On Tue, 27 Jul 2004 17:14:15 -0700, 
>> There are also flash and similar absolute space footprints to consider.
>> Experiments seem to suggest that consolidating the lock sections and
>> other spinlock code can reduce kernel image size by as much as 220KB on
>> ia32 with no performance impact (rigorous benchmarks still in progress).

On Wed, Jul 28, 2004 at 10:35:08AM +1000, Keith Owens wrote:
> I consolidated the spinlock contention path to a single routine on
> ia64, with big space savings.  The problem with the ia64 consolidation
> was backtracing through a contended lock; the ia64 unwind API is not
> designed for code that is shared between multiple code paths but uses
> non-standard entry and exit conventions.  In the end, David Mosberger
> did a patch to gcc to do lightweight calls to the out of line
> contention code, just to get reliable backtraces.
> kdb has workarounds for backtracing through ia64 contended locks when
> the kernel is built with older versions of gcc.  gdb (and hence kgdb)
> has no idea about the special out of line code.  Mind you, the same is
> true right now with the out of line i386 code, you need special
> heuristics to backtrace the existing spinlock code reliably.  That will
> only get worse with Zwane's patch, interrupts can now occur in the out
> of line code.

It's good to know there is a precedent and that the backtrace issue has
been looked at on other architectures.


On Wed, Jul 28, 2004 at 10:35:08AM +1000, Keith Owens wrote:
> Are you are planning to consolidate the out of line code for i386?  Is
> there a patch (even work in progress) so I can start thinking about
> doing reliable backtraces?

The experiments were carried out using the standard calling convention.
We may investigate less standard calling conventions, but they should
actually already be there given __write_lock_failed/__read_lock_failed.
i.e. if reliable backtraces are going to be an issue they should
already be an issue for rwlocks.


-- wli
-
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/