Hi Jirka Sorry for late!
On Tue, Nov 06, 2018 at 12:54:36PM +0100, Jiri Olsa wrote: > On Mon, Nov 05, 2018 at 08:53:42PM -0800, David Miller wrote: > > > > Jiri, > > > > Because you now run queued_events__queue() lockless with that condvar > > trick, it is possible for top->qe.in to be seen as one past the data[] > > array, this is because the rotate_queues() code goes: > > > > if (++top->qe.in > &top->qe.data[1]) > > top->qe.in = &top->qe.data[0]; > > > > So for a brief moment top->qe.in is out of range and thus > > perf_top__mmap_read_idx() can try to enqueue to top->qe.data[2] > > > > We can just do: > > > > if (top->qe.in == &top->qe.data[1]) > > top->qe.in = &top->qe.data[0]; > > else > > top->qe.in = &top->qe.data[1]; > > > > Or, make top->qe.in an index, and simply go: > > > > top->qe.in ^= 1; > > > > Either way will fix the bug. > > ah right.. I had originaly full mutex around that, > then I switched it off in the last patch and did > not realize this implication.. nice ;-) I like the rotate_queues() using cond-variable. Have you tried to use the same for hists->lock in hists__get_rotate_entries_in() too? Eventually it'd be nice to avoid locks when a single thread processes all the events. Thanks, Namhyung
