On Wed, Oct 16, 2013 at 06:14:50PM +0100, Jiang Liu wrote:

On 10/16/2013 11:36 PM, Jiang Liu wrote:

quoted

On 10/16/2013 06:51 PM, Will Deacon wrote:

quoted

Take care here:

	- This doesn't guarantee that CPUs exclusively see new_insn (they
	  may see old_insn, even after patching)

Good point. Still need some synchronization mechanism lighter than
stop_machine(). How about kick_all_cpus_sync(), which should be lighter
than stop_machine()?

For jump label, x86 uses three smp_call_function() to replace a
stop_machine(). Seems it's valuable for us to use just one
kick_all_cpus_sync() to avoid a stop_machine().

Yes, the exception return after handling the IPI from kick_all_cpus_sync
should be enough to synchronise the pipeline, providing that all cache
maintenance by the CPU writing the new instructions has completed (if you're
using flush_icache_range then you're fine).

Will