Re: [PATCH] perf data: export to JSON

[Jiri Olsa]

On Wed, Mar 31, 2021 at 06:38:16AM -0400, Nicholas Fraser wrote:
> Hi Jiri,
> 
> Thanks for the review. I've addressed your suggestions; some notes are
> below. I'll send a new patch.
> 
> 
> On 2021-03-24 2:20 p.m., Jiri Olsa wrote:
> > On Wed, Mar 24, 2021 at 09:06:50AM -0400, Nicholas Fraser wrote:
> >> [...] We use this to import the data into a tool on Windows
> >> where integrating perf or libbabeltrace is impractical.
> > 
> > hi,
> > exciting ;-) and curious, which tool is that?
> > 
> 
> The tool is called gpuvis. The perf JSON parsing support is here:
> 
> https://github.com/ludocode/gpuvis
> 
> The idea is to be able to line up samples from perf with GPU trace events, so
> you can do things like timebox all perf samples in a particular frame of
> rendering.

I recall you did not add support for walltime clock,
don't you need it to sync with other events?

> 
> 
> > we already have zstd support compiled in for compressing samples,
> > should be easy to use it for compressing the output of this right
> > away
> 
> This would require that apps that consume this integrate zstd as well. It's
> simpler (both conceptually and from an integration standpoint) to just compress
> on command-line if you need with whatever compressor you want. You can even do
> this inline by writing to /dev/stdout, e.g.:
> 
>     perf data convert --to-json /dev/stdout --force | zstd > out.json.zstd
> 
> Since we're transferring to Windows, more likely we'd output the JSON and then
> put it in a .zip container.
> 
> 

ok

> > I understand not supporting opts.all or opts.tod, but 'force'
> > support means just assigning 'force' to struct perf_data
> 
> It's not clear to me what 'force' does on 'struct perf_data' since we're only
> reading it. I assumed for data export it meant the output file should be
> overwritten. I've made it do both in the replacement patch.
> 

it tells perf to skip ownership validation, perf will not
open other user data file if it's not forced

jirka