Investigate and enable epoll under manylinux for Python binary distributions

[mehrdada]

Per @jonparrot, manylinux1 should support epoll. The possibility of compiling epoll to Python binary packages should be re-evaluated:

epoll was introduced in glibc 2.3.2, manylinux1 supports glibc <= 2.5 so it should be possible to enable epoll without breaking manylinux1 compatibility.

[mehrdada]

As tried in #13951, epoll_create1 which is the syscall used by gRPC is unavailable under manylinux1

[kpayson64]

Can we use epoll_create instead?

[mehrdada]

@kpayson64 I believe that would be a fundamental change: epoll_create has different semantics and does not support the flag we pass to epoll_create1, so even if possible to make epoll_create work, it would require substantial changes to the polling engine, and we might as well fork it into another polling engine specific to Python.

[kpayson64]

grpc/src/core/lib/iomgr/ev_epollex_linux.cc

Line 439 in 3ea0e2d

int epfd = epoll_create1(EPOLL_CLOEXEC);

It looks like the only flag we pass is EPOLL_CLOEXEC? We should be able to replicate that exact behavior with an additional call to fcntl(FD_CLOEXEC)

[mehrdada]

Yes as far as I know that is the only flag we pass to it. If there is a way to duplicate that exactly it should work, but I am not well-versed on the semantics of the flag.

[mehrdada]

@nicolasnoble Is there a catch to replacing epoll_create1 with epoll_create immediately followed by fcntl? Can their non-atomicity be a problem?

[mehrdada]

Summary of the offline discussion with @nicolasnoble: we should ifdef on libc version and use epoll_create followed by setting the flag.