On Fri, Jan 29, 2021 at 10:08 AM Mark Thomas <ma...@apache.org> wrote:
> On 29/01/2021 09:01, Rémy Maucherat wrote: > > On Fri, Jan 29, 2021 at 9:12 AM Mark Thomas <ma...@apache.org> wrote: > > > >> On 29/01/2021 08:06, Mark Thomas wrote: > >>> Heads up. When I ran my pre-tag tests for 8.5.x I started to see a > bunch > >>> of JVM crashes on Windows with both Java 7 and Java 8. The crashes were > >>> always in a compilation thread which was unusual. > >>> > >>> I am currently doing the binary search to figure out which commit > >>> triggered the issue. > >> > >> That was quicker than I expected. Running the HTTP/2 tests on their own > >> is enough to trigger the issue. > >> > >> It was the fix for bug 65111 - freeing the direct buffers - that > >> triggered this. > >> > > > > Almost certainly the swap with SocketBufferHandler.EMPTY saves 9 and 10 > > from crashing. It's not certain this is 100% safe actually, but possibly > > good enough. I did reproduce a (rare) crash with 8.5, so no good option > > other than take it out for now. > > I was thinking about moving the cleaning of the direct buffers to > destroySocket() > > I'll do some tests. > Yes, ideally, waiting until after the socket is removed from the poller is a good plan. Rémy
