On Sun, Oct 21, 2018 at 12:53:17AM +0900, Akira Yokosawa wrote: > Hi Paul, > > I just noticed occasional error of count_lim_atomic.c on POWER8 at current master. > As I've recently touched the code under Codesamples/count/, I also tested on > the tag "v2017.11.22a", and saw the same behavior. > > The POWER8 virtual machine is Ubuntu 16.04. > > Example output: > > $ ./count_lim_atomic 6 uperf 1 > !!! Count mismatch: 0 counted vs. 8 final value > n_reads: 0 n_updates: 26038000 nreaders: 0 nupdaters: 6 duration: 240 > ns/read: nan ns/update: 55.3038 > > $ ./count_lim_atomic 6 perf 1 > !!! Count mismatch: 0 counted vs. 11 final value > n_reads: 287000 n_updates: 1702000 nreaders: 6 nupdaters: 1 duration: 240 > ns/read: 5017.42 ns/update: 141.011 > > As you see, the final count check of zero fails even when nupdaters == 1. Yow!!! Thank you for checking this! That said, it probably wasn't really single threaded, at least assuming that you had at least one reader. > I have no idea what's wrong in count_lim_atomic.c. > > Can you look into this? There might be something wrong in the header file > under CodeSamples/arch-ppc64.h. There isn't much in that file anymore because we now rely on the gcc intrinsics for the most part. Which might well be the problem, depending on compiler versions and so on. Could you please send me the output of "objdump -d" on count_lim_atomic.o? And on the full count_lim_atomic binary, just in case gcc decides to be tricky in its code generation? In the meantime, there might well be a generic bug in count_lim_atomic.c that just happens not to be exercised on x86, and I will look into that. I am on travel next week, so will be in odd timezones, but should have at least a little useful airplane time to look into this. > On x86_64, I've never seen the count mismatch. Well, David Goldblatt's first C++11 signal-based litmus test wouldn't fail on PowerPC but did on x86, so I guess that they are now even. ;-) Thanx, Paul
