Messages in this thread | | | Date | Sun, 29 Jul 2018 19:05:56 +0200 | From | Jiri Olsa <> | Subject | Re: [PATCH] perf c2c report: Fix crash for empty browser |
| |
On Fri, Jul 27, 2018 at 11:13:40AM -0300, Arnaldo Carvalho de Melo wrote: > Adding Joe to the CC list. > > Em Fri, Jul 27, 2018 at 09:06:23AM +0200, Jiri Olsa escreveu: > > On Thu, Jul 26, 2018 at 11:31:34PM +0000, rodia@autistici.org wrote: > > > On 2018-07-26 19:30, Arnaldo Carvalho de Melo wrote: > > > > Em Tue, Jul 24, 2018 at 08:20:08AM +0200, Jiri Olsa escreveu: > > > > > Do not try to display entry details if there's > > > > > not any. Currently this ends up in crash: > > > > > $ perf c2c report > > > > > perf: Segmentation fault > > > > > How to replicate this? > > > > > I tried: > > > > > $ perf record sleep 1 > > > > $ perf c2c report > > > > > But it didn't segfault > > > > Similarly I have tried : > > > $ perf record sleep 1 > > > $ perf c2c report > > > Then Press `d` to show the cache-line contents. > > > yep, sry I forgot to mention you need to press the 'd' to show details > > > > This replies the segfault on my machine (4.17.8-1). > > > The patch mentioned above should solve it, even tough I am not sure as I > > > haven't been able to recompile the kernel. > > > no need to recompile kernel > > > > The segfault by itself seems to be due to the report logic, as it did not > > > expect to report on an empty browser. > > > What has stepped me back is that application which I have been testing with > > > rely on multiple threads instantiated through pthread, which should be > > > counted in user-level threads right? But they still seem to return an empty > > > browser. > > > right, c2c scans read/write accesses and tries to find false sharing > > cases maybe there was nothing to be found > > > > When instead c2c is runned system-wide, with an application running on > > > multiple threads like firefox or julia, cache hits are measured and also > > > they are traced back in the source code. > > > I got a cache line (attached) for 'perf bench sched messaging' > > NOT being traced system wide and just for user (you'll get plenty > > of detected cachelines in kernel space): > > With that info in mind, we get: > > [root@seventh ~]# perf record sleep 1 > [ perf record: Woken up 1 times to write data ] > [ perf record: Captured and wrote 0.001 MB perf.data (6 samples) ] > [root@seventh ~]# > [root@seventh ~]# > [root@seventh ~]# perf c2c report # and press 'd' > perf: Segmentation fault > -------- backtrace -------- > perf[0x5b1d2a] > /lib64/libc.so.6(+0x346df)[0x7fcb566e36df] > perf[0x46fcae] > perf[0x4a9f1e] > perf[0x4aa220] > perf(main+0x301)[0x42c561] > /lib64/libc.so.6(__libc_start_main+0xe9)[0x7fcb566cff29] > perf(_start+0x29)[0x42c999] > [root@seventh ~]# > > With your patches the segfault is gone, but I'd do a follow up patch to > show some message telling the user why 'd' showed nothing and > instructing him about what is missing, i.e. is this done on a perf.data > file that has no events of interest? Suggest using 'perf c2c record' or > 'perf record -e events,of,interest,to,perf,c2c', was this done on some > workload where no false sharing was detected? Say so, etc.
ok, will try to come up with something
> > I applied your patch with a more detailed commit log to state how this > can reproduced, etc, as usual: > > https://git.kernel.org/acme/c/983eb6aa7098
I checked, looks good, thanks
jirka
| |