right, libuv[1] is a good example of an IO completion API[2] that works across both epoll/kqueue and IOCP. There have been a couple offhand discussions about libuv on httpd-dev, but no one is clamoring for adoption.
APR currently has no such abstraction. pollcb was added to provide a polling API that was lock free, allowed cross-thread addition and didn't keep extra data structures around for N sockets under management. (this contrasts to APR's pollset which has locks, uses a couple RINGs of structures to keep track of things, etc) [1] - https://github.com/libuv/libuv [2] - http://nikhilm.github.io/uvbook/filesystem.html#buffers-and-streams On Thu, Mar 12, 2015 at 3:40 PM, Ivan Zhakov <i...@visualsvn.com> wrote: > On 9 March 2015 at 20:04, Jim Jagielski <j...@jagunet.com> wrote: > > As far as I can tell, there are no real consumers of apr_pollcb_*() > > other than Motorz (and Simple)... Not sure, but it looks like the > > pollcb stuff may be lagging, since I'm getting some weird behavior > > when using that impl. So I'll likely be moving Motorz off of > > pollcb and use the "traditional" impl which has been proven in > > Event. > > > > I'd like to share some thoughts about pollcb as Windows developer. > Most efficient way to handle a lot of connections on Windows is I/O > completion ports (IOCP). They are like epoll, but different. Here is > good writeup about difference [1]. In short: IOCP provides > notification when I/O completes, while epoll provides notifications > when socket is ready to perform I/O operation. It would require > additional buffer to implement pollcb using IOCP on Windows. Or > APR/HTTPD should have different abstraction that is responsible for > I/O and waiting for completion. > > [1] > http://www.ulduzsoft.com/2014/01/practical-difference-between-epoll-and-windows-io-completion-ports-iocp/ > > -- > Ivan Zhakov >
