On Sat, Nov 23, 2019 at 8:54 PM Mark Thomas <ma...@apache.org> wrote:
> On 22/11/2019 15:25, Rémy Maucherat wrote: > > Hi, > > > > I'm looking at the endpoint code a lot in preparation for ports to 8.5. > > > > I noticed the write(Non)BlockingDirect methods in particular as they had > > a fixme which wondered about the buffering strategy. Given the use, the > > thinking seems very legacy optimization oriented, because: > > - If the socket buffer is configured as direct, then it is a good idea > > to always use it > > - The async API is a better API for "direct" writes (where the caller > > owns the buffer that is ultimately written to the socket or the SSL > > engine) with a clearer contract > > - HTTP/2 and websocket don't use them anymore by default, and sendfile > > also bypasses them > > > > I have not made a performance study but it's fairly certain the benefit > > is zero. As a result, I'm considering removing them (quick experiment: > > no particular problem with the testsuite). > > My only concern would be stability. With careful review we should be > able to handle that. > It looked ok and removed a big chunk of code so I went ahead with it. As far as porting goes, I have: - This one, also the easiest one: drop SocketWrapperBase.write(Non)BlockingDirect methods - Move the processor tracking to the wrapper and the connection map to the endpoint; this one can improve performance - The pile ups of close refactoring; big cleanup and move close to the wrapper, also improves robustness, I would do it last If everything is ported, the two branches should be similar and patch cherry pick should work better for future fixes. Rémy
