On Fri, 4 Jan 2019 18:26:16 +0100 Dmitry Vyukov <dvyu...@google.com> wrote:
> On Fri, Jan 4, 2019 at 6:14 PM Stefano Brivio <sbri...@redhat.com> wrote: > > > > On Fri, 4 Jan 2019 12:05:04 +0100 > > Dmitry Vyukov <dvyu...@google.com> wrote: > > > > > I've added these as tests: > > > > > > https://github.com/google/syzkaller/blob/master/pkg/report/testdata/linux/report/341 > > > https://github.com/google/syzkaller/blob/master/pkg/report/testdata/linux/report/342 > > > https://github.com/google/syzkaller/blob/master/pkg/report/testdata/linux/report/343 > > > https://github.com/google/syzkaller/blob/master/pkg/report/testdata/linux/report/344 > > > > > > Will try to figure out how to distinguish them from true corrupted > > > reports. Usually when Call Trace does not have any frames, it's a sign > > > of a corrupted report, and in other crashes we see the same report but > > > with a stack trace. But some stack-corruption-related reliably don't > > > have stack traces (not corrupted). But then some other > > > stack-corruption-related crashes do have stack traces, and for these > > > no stack trace again means a corrupted kernel output. Amusingly this > > > is one of the most complex parts of syzkaller. > > > > I'm not sure how complicated that would be, but what about some metric > > based on valid symbol names being reported? > > Please elaborate. What do you mean by "valid symbol names"? I mean a symbol name listed in /proc/kallsyms on the running system. This is usually my minimum threshold for "I can do something with this report" -- which doesn't mean it's necessarily valid, but well, if you have that, it means that at least something worked in the reporting, and you can at least start having a look at a specific function. > Note that corrupted output detection solves 2 problems: > 1. Do we think the output is truncated to the point of being not useful? > E.g. sometimes kernel produces just 1 line: > > general protection fault: 0000 [#1] PREEMPT SMP KASAN > > This is sure a crash, but it's not too useful to report. Sure. In those tests above you have: - 341: udp6_lib_lookup2+0x622, handle_irq+0x2cb - 342: __sanitizer_cov_trace_pc+0x8, handle_irq+0x2cb - 343: __udp6_lib_err, etc. - 344: __udp6_lib_lookup+0x1d, etc. and this makes all those reports at least minimally useful. > 2. Do we have any reasons to think we extracted bogus crash identity? > E.g. crash intermixed with output from another thread so that we say > "something-bad in function foo", when in fact function foo come from > output of the second non-crashing thread. Okay, this looks way more complicated :) -- Stefano