On Wed, Nov 27, 2019 at 5:35 PM Mark Thomas <ma...@apache.org> wrote:
> On 27/11/2019 15:59, Rémy Maucherat wrote: > > On Wed, Nov 27, 2019 at 4:51 PM Mark Thomas <ma...@apache.org > > <mailto:ma...@apache.org>> wrote: > > > > I've been looking at the 9.0.x changes. > > > > What is clearer now (I think the differences were always there - just > > harder to see with the more complex code) is that there are some > minor > > differences in behaviour when all of the following are true: > > - the source buffer is emptied when it is copied to the socket buffer > > - the socket buffer is full after the source buffer has been copied > > > > In some cases there will be an immediate write to the network. In > other > > cases the network write will wait until the next write to the socket > > buffer or the next explicit flush. > > > > I think I prefer an immediate write in this scenario. WDYT to > updating > > (very carefully) all of the write methods to behave that way? > > > > > > Ah personally I did prefer not writing that full buffer (unless the > > source still has data, of course) on purpose since I considered not > > doing IO reduced the likelihood to block on IO (or write 0 for non > > blocking and having to use the poller). That's possible though if you > > like it better. > > My preference for writing earlier is only a slight one. I'm more > concerned that different methods take a different approach. Taking a > closer look at that is on my TODO list but it isn't a priority for me at > the moment. > I'll double check the inconsistency :) writeBlocking writes after. writeNonBlockingInternal has that "&& !socketBufferHandler.isWriteBufferWritable()" in the while loop that makes the thing inconsistent, it is a leftover from the old code. For now I'll fix that. Rémy
