Re: [PATCH v5 00/17] Reference count checker and related fixes

[Arnaldo Carvalho de Melo]

Em Tue, Apr 04, 2023 at 08:25:41PM +0300, Adrian Hunter escreveu:
> On 4/04/23 18:58, Ian Rogers wrote:
> > Ping. It would be nice to have this landed or at least the first 10
> > patches that refactor the map API and are the bulk of the
> > lines-of-code changed. Having those landed would make it easier to
> > rebase in the future, but I also think the whole series is ready to
> > go.
> 
> I was wondering if the handling of dynamic data like struct map makes
> any sense at present.  Perhaps someone can reassure me.
> 
> A struct map can be updated when an MMAP event is processed.  So it

Yes, it can, and the update is made via a new PERF_RECORD_MMAP, right?

So:

	perf_event__process_mmap()
	  machine__process_mmap2_event()
	    map__new() + thread__insert_map(thread, map)
	    	maps__fixup_overlappings()
			maps__insert(thread->maps, map);

Ok, from this point on new samples on ] map->start .. map->end ] will
grab a refcount to this new map in its hist_entry, right?

When we want to sort by dso we will look at hist_entry->map->dso, etc.

> seems like anything racing with event processing is already broken, and
> reference counting / locking cannot help - unless there is also
> copy-on-write (which there isn't at present)?

> For struct maps, referencing it while simultaneously processing
> events seems to make even less sense?

Can you elaborate some more?

- Arnaldo