Re: [PATCH v2 3/4] perf tools: Support reading from backward ring buffer

[Wangnan (F)]

On 2016/5/6 4:07, Arnaldo Carvalho de Melo wrote:
> Em Wed, Apr 27, 2016 at 02:19:22AM +0000, Wang Nan escreveu:
> > perf_evlist__mmap_read_backward() is introduced for reading backward
> > ring buffer. Different from reading forward, before reading, caller
> > needs to call perf_evlist__mmap_read_catchup() first.
> >
> > Backward ring buffer should be read from 'head' pointer, not '0'.
> > perf_evlist__mmap_read_catchup() saves 'head' to 'md->prev', then
> > make it remember the start position after each reading.
> >
> > Signed-off-by: Wang Nan <wangnan0@xxxxxxxxxx>
> > Cc: Arnaldo Carvalho de Melo <acme@xxxxxxxxxx>
> > Cc: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
> > Cc: Zefan Li <lizefan@xxxxxxxxxx>
> > Cc: pi3orama@xxxxxxx
> > ---
> >   tools/perf/util/evlist.c | 39 +++++++++++++++++++++++++++++++++++++++
> >   tools/perf/util/evlist.h |  4 ++++
> >   2 files changed, 43 insertions(+)
> >
> > diff --git a/tools/perf/util/evlist.c b/tools/perf/util/evlist.c
> > index 17cd014..2e0b7b0 100644
> > --- a/tools/perf/util/evlist.c
> > +++ b/tools/perf/util/evlist.c
> > @@ -766,6 +766,45 @@ union perf_event *perf_evlist__mmap_read(struct perf_evlist *evlist, int idx)
> >   	return perf_mmap__read(md, evlist->overwrite, old, head, &md->prev);
> >   }
> >   
> > +union perf_event *
> > +perf_evlist__mmap_read_backward(struct perf_evlist *evlist, int idx)
> > +{
> > +	struct perf_mmap *md = &evlist->mmap[idx];
> > +	u64 head, end;
> > +	u64 start = md->prev;
> > +
> > +	/*
> > +	 * Check if event was unmapped due to a POLLHUP/POLLERR.
> > +	 */
> > +	if (!atomic_read(&md->refcnt))
> > +		return NULL;
> > +
> > +	/* NOTE: head is negative in this case */
> What is this comment for, can you ellaborate? Are you double sure this
> arithmetic with u64, negative values, that -head below are all ok?

Yes. In backward ring buffer, kernel write data from '0'. Each
time it write a record, it minus sizeof(record) from 'head'
pointer. So '-head' means number of bytes already written.

> I've applied the first two patches in this series.
>
> I also need to check why we now need that catchup thing :-\

I think catchup is necessary. I tried to get rid of it by setting
md->prev to -1 and initialize it to 0 when reading forwardly and
to head when reading backwardly, but it require too much code
adjustments.

> > +	head = perf_mmap__read_head(md);
> > +
> > +	if (!head)
> > +		return NULL;
> > +
> > +	end = head + md->mask + 1;
> > +
> > +	if ((end - head) > -head)
> > +		end = 0;
> > +

This '-head' is used to detect wrapping. Never use positive pointer.

> > +	return perf_mmap__read(md, false, start, end, &md->prev);
> > +}
> > +
> > +void perf_evlist__mmap_read_catchup(struct perf_evlist *evlist, int idx)
> > +{
> > +	struct perf_mmap *md = &evlist->mmap[idx];
> > +	u64 head;
> > +
> > +	if (!atomic_read(&md->refcnt))
> > +		return;
> > +
> > +	head = perf_mmap__read_head(md);
> > +	md->prev = head;
> > +}
> > +
> >   static bool perf_mmap__empty(struct perf_mmap *md)
> >   {
> >   	return perf_mmap__read_head(md) == md->prev && !md->auxtrace_mmap.base;
> > diff --git a/tools/perf/util/evlist.h b/tools/perf/util/evlist.h
> > index 208897a..85d1b59 100644
> > --- a/tools/perf/util/evlist.h
> > +++ b/tools/perf/util/evlist.h
> > @@ -129,6 +129,10 @@ struct perf_sample_id *perf_evlist__id2sid(struct perf_evlist *evlist, u64 id);
> >   
> >   union perf_event *perf_evlist__mmap_read(struct perf_evlist *evlist, int idx);
> >   
> > +union perf_event *perf_evlist__mmap_read_backward(struct perf_evlist *evlist,
> > +						  int idx);
> > +void perf_evlist__mmap_read_catchup(struct perf_evlist *evlist, int idx);
> > +
> >   void perf_evlist__mmap_consume(struct perf_evlist *evlist, int idx);
> >   
> >   int perf_evlist__open(struct perf_evlist *evlist);
> > --
> > 1.8.3.4