Re: [Adeos-main] [PATCH] I-pipe tracer for ARM

[Jan Kiszka <jan.kiszka@domain.hid>]

[-- Attachment #1: Type: text/plain, Size: 1137 bytes --]

Sebastian Smolorz wrote:
> Jan Kiszka wrote:
>> Sebastian Smolorz wrote:
>>>>> <    +freeze  0xffffffff     0   97.444N __ipipe_frozen_ctrl+0x108
>>>>> (vfs_write+0xc4)
>>>> Hmm, takes quite a long time to complete the freeze though you only have
>>>> 10 post-trace points (which defines its complexity). Unusual, also when
>>>> comparing with other traced functions.
>>> It seems that it doesn't depend on post-trace-points. When I reduce them
>>> to 2 the value for the freeze does not get smaller. Maybe it has
>>> something to do with the following lines?
>> Are you tracing into vmalloc'ed memory?
> 
> Yes, without CONFIG_IPIPE_TRACE_VMALLOC the system does not boot.

Hmm, makes me wonder of some laziness of the page mapping or a missing
lock-against-swapping causes this. Could you have a closer look at the
tracer code if we are lacking some magic for the vmalloc trace buffer
(compare to the xnheap code, e.g.)?

> 
>> And what results do you get with 
>> latency -f?
> 
> These seem to be better (see attachment).

This might shadow earlier page faults, i.e. during the first triggerings.

Jan


[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 250 bytes --]