Re: [PATCH v2 1/3] Revert "perf session: Fix decompression of PERF_RECORD_COMPRESSED records"

From: Jiri Olsa <jolsa@redhat.com>
To: Petr Malat <oss@malat.biz>
Cc: linux-kernel@vger.kernel.org,
	Peter Zijlstra <peterz@infradead.org>,
	Ingo Molnar <mingo@redhat.com>,
	Arnaldo Carvalho de Melo <acme@kernel.org>,
	Mark Rutland <mark.rutland@arm.com>,
	Alexander Shishkin <alexander.shishkin@linux.intel.com>,
	Namhyung Kim <namhyung@kernel.org>,
	Adrian Hunter <adrian.hunter@intel.com>,
	Kan Liang <kan.liang@linux.intel.com>,
	Alexey Budankov <alexey.budankov@linux.intel.com>
Subject: Re: [PATCH v2 1/3] Revert "perf session: Fix decompression of PERF_RECORD_COMPRESSED records"
Date: Tue, 1 Dec 2020 20:09:28 +0100	[thread overview]
Message-ID: <20201201190928.GB3169083@krava> (raw)
In-Reply-To: <20201130114020.GA29476@ntb.petris.klfree.czf>

On Mon, Nov 30, 2020 at 12:40:20PM +0100, Petr Malat wrote:
> Hi Jiří,
> were you able to reproduce the issue? I may also upload perf-archive
> if that would help.

oh yea ;-) seems like those 2 commits you reverted broke 32 bits
perf for data files > 32MB

but the fix you did does not work for Alexey's test he mentioned
in the commit:

      $ perf record -z -- some_long_running_workload
      $ perf report --stdio -vv

it's failing for me with:

	# ./perf report
	Couldn't allocate memory for decompression
	0xfe6f3a [0x60]: failed to process type: 81 [Operation not permitted]
	Error:
	failed to process sample
	# To display the perf.data header info, please use --header/--header-only options.
	#

I think that's why here's special handling for compressed
events, but I'll need to check on that in more detail,
I was hoping for Alexey to answer ;-)

jirka

>   Petr
> 
> On Tue, Nov 24, 2020 at 07:15:19PM +0100, Petr Malat wrote:
> > Hi!
> > On Tue, Nov 24, 2020 at 03:36:45PM +0100, Jiri Olsa wrote:
> > > On Tue, Nov 24, 2020 at 11:29:15AM +0100, Petr Malat wrote:
> > > > Both mmapped and compressed events can be split by the buffer boundary,
> > > > it doesn't make sense to handle them differently.
> > > I'm going to need more than this, if there's a problem
> > > with current code please share more details, what's
> > > broken and how it shows
> > It's easy to trigger the problem - make a perf recording larger than
> > MMAP_SIZE (32MB on 32-bit platform) and run perf report on it. There
> > is a small chance recorded events will be aligned on the 32 MB
> > boundary and in that case just repeat the test.
> > 
> > The problem was introduced by "perf session: Avoid infinite loop when
> > seeing invalid header.size", which instead of aborting the execution
> > when there is a truncated event at the end of the file just terminated
> > execution whenever there is a split event. Later then the problem has
> > been noticed for compressed events and fixed by "perf session: Fix
> > decompression of PERF_RECORD_COMPRESSED records" by effectively
> > reverting "perf session: Avoid infinite loop when seeing invalid
> > header.size" for compressed events, which left uncompressed events
> > broken.
> > 
> > I think the best is to revert these 2 changes and fix the original
> > problem by aborting when there is no actual shift during remapping - as
> > long as we shift, it's clear we must approach the end of the file so
> > such an algorithm can't loop forever.
> > BR,
> >   Petr
>