Messages in this thread |  | | Date | Wed, 11 Dec 2013 14:47:24 +0100 | From | Ingo Molnar <> | Subject | Re: [PATCH v0 00/71] perf: Add support for Intel Processor Trace |
| |
* Alexander Shishkin <alexander.shishkin@linux.intel.com> wrote:
> Ingo Molnar <mingo@kernel.org> writes: > > > * Alexander Shishkin <alexander.shishkin@linux.intel.com> wrote: > > > >> Hi, > >> > >> This patchset adds support for Intel Processor Trace (PT) extension > >> [1] of Intel Architecture that allows the capture of information > >> about software execution flow, to the perf kernel and userspace > >> infrastructure. We provide an abstraction for it called "itrace" for > >> "instruction trace" ([2]). > > > > Ok, this feature looks rather interesting. > > > > On the hardware side this is essentially BTS (Branch Trace Store) > > on steroids (with many extensions), right? > > Yes, you get timestamps and all sorts of other useful data in the > trace and the performance intrusion is much less than that of BTS.
So the problem I see here right now that BTS is rarely used and AFAICS close to unmaintained. It has some very minimal support in 'perf script' but that's all I can see.
So one necessary precondition to merging PT support would be to have a convincing case that this kind of stuff is generally useful.
One good approach to do that would be to unify the BTS and PT tooling (the kernel side can be unified as well, to the extent it makes sense), and to prove it via actual functionality that this stuff matters. BTS is available widely, so the tooling can be tested by anyone who's interested.
Allow people to record crashes in core dumps, allow them to look at histograms/spectrograms of BTS/PT traces, zoom in on actual traces, etc. - make it easier to handle this huge amount of data and visualize traces in other ways you find useful, etc.
None of that is done right now via BTS so nobody uses it.
Thanks,
Ingo
|  |