Performance of l4_task_unmap call
adam at os.inf.tu-dresden.de
Thu Apr 26 11:20:02 CEST 2012
On Tue Apr 24, 2012 at 16:36:52 -0800, Chen Tian wrote:
> I did a small test: a parent task first maps and then unmaps a page to its
> child. Both parent and child task are running on the same core.
> I found the execution is very slow and 99% time are spent on l4_task_unmap(
> L4_FP_OTHER_SPACES flag is used).
> After some investigation, I found the flag "need_xcpu_tlb_flush" in
> map_util.cpp:766 is set to true. My guess is that this flag controls the
> tlb of the cpu at which the child task is running. However, in my case,
> both tasks are running on the same core, setting this flag can have huge
> negative impact on performance.
> I am thinking to put a check before this statement: if (current_cpu() !=
> space's CPU) set the flag to true. Is this correct? If so, how to get
> space's cpu?
> space->cpu() ?
This is a valid optimization, however, only if it is made sure that no
code of that task has ever run on any other core. A space has no CPU
because a space does not execute anything.
Adam adam at os.inf.tu-dresden.de
More information about the l4-hackers