Re: [RFC] perf session: Fix compression processing

[Jiri Olsa]

On Mon, Nov 11, 2019 at 06:53:35PM +0300, Alexey Budankov wrote:
> On 11.11.2019 18:46, Jiri Olsa wrote:
> > On Mon, Nov 11, 2019 at 06:41:47PM +0300, Alexey Budankov wrote:
> >> On 11.11.2019 17:56, Jiri Olsa wrote:
> >>> On Mon, Nov 11, 2019 at 05:38:49PM +0300, Alexey Budankov wrote:
> >>>>
> >>>> On 04.11.2019 1:24, Jiri Olsa wrote:
> >>>>> hi,
> >>>> <SNIP>
> >>>>> ---
> >>>>> The compressed data processing occasionally fails with:
> >>>>>   $ perf report --stdio -vv
> >>>>>   decomp (B): 44519 to 163000
> >>>>>   decomp (B): 48119 to 174800
> >>>>>   decomp (B): 65527 to 131072
> >>>>>   fetch_mmaped_event: head=0x1ffe0 event->header_size=0x28, mmap_size=0x20000: fuzzed perf.data?
> >>>>>   Error:
> >>>>>   failed to process sample
> >>>>>   ...
> >>>>>
> >>>>> It's caused by recent fuzzer fix that does not take into account
> >>>>> that compressed data do not need to by fully present in the buffer,
> >>>>> so it's ok to just return NULL and not to fail.
> >>>>>
> >>>>> Fixes: 57fc032ad643 ("perf session: Avoid infinite loop when seeing invalid header.size")
> >>>>> Link: http://lkml.kernel.org/n/tip-q1biqscs4stcmc9bs1iokfro@xxxxxxxxxxxxxx
> >>>>> Signed-off-by: Jiri Olsa <jolsa@xxxxxxxxxx>
> >>>>> ---
> >>>>>  tools/perf/util/session.c | 8 +++++---
> >>>>>  1 file changed, 5 insertions(+), 3 deletions(-)
> >>>>>
> >>>>> diff --git a/tools/perf/util/session.c b/tools/perf/util/session.c
> >>>>> index f07b8ecb91bc..3589ed14a629 100644
> >>>>> --- a/tools/perf/util/session.c
> >>>>> +++ b/tools/perf/util/session.c
> >>>>> @@ -1959,7 +1959,7 @@ static int __perf_session__process_pipe_events(struct perf_session *session)
> >>>>>  
> >>>>>  static union perf_event *
> >>>>>  fetch_mmaped_event(struct perf_session *session,
> >>>>> -		   u64 head, size_t mmap_size, char *buf)
> >>>>> +		   u64 head, size_t mmap_size, char *buf, bool decomp)
> >>>>
> >>>> bools in interface make code less transparent.
> >>>>
> >>>>>  {
> >>>>>  	union perf_event *event;
> >>>>>  
> >>>>> @@ -1979,6 +1979,8 @@ fetch_mmaped_event(struct perf_session *session,
> >>>>>  		/* We're not fetching the event so swap back again */
> >>>>>  		if (session->header.needs_swap)
> >>>>>  			perf_event_header__bswap(&event->header);
> >>>>> +		if (decomp)
> >>>>> +			return NULL;
> >>>>>  		pr_debug("%s: head=%#" PRIx64 " event->header_size=%#x, mmap_size=%#zx: fuzzed perf.data?\n",
> >>>>>  			 __func__, head, event->header.size, mmap_size);
> >>>>>  		return ERR_PTR(-EINVAL);
> >>>>> @@ -1997,7 +1999,7 @@ static int __perf_session__process_decomp_events(struct perf_session *session)
> >>>>>  		return 0;
> >>>>>  
> >>>>>  	while (decomp->head < decomp->size && !session_done()) {
> >>>>> -		union perf_event *event = fetch_mmaped_event(session, decomp->head, decomp->size, decomp->data);
> >>>>> +		union perf_event *event = fetch_mmaped_event(session, decomp->head, decomp->size, decomp->data, true);
> >>>>
> >>>> It looks like this call can be skipped, at all, in this case.
> >>>
> >>> not sure what you mean, we are in decomp code no?
> >>
> >> Ok, it is inside "not fetching" branch. 
> >> NULL return value means to proceed getting further over the trace.
> >> Checking record type == COMPRESSED at the higher level could 
> >> probably be cleaner fix and also work faster.
> > 
> > any chance you could post the fix? the patch I did was a
> > quick fix to get the feature working for presentation ;-)
> > you're probably thinking of the proper approach
> 
> Please share the exact reproducing steps 
> so I could come up with something.

'perf record -z' for longer workloads

jirka