On Sat, Jun 18, 2022 at 12:41:11PM +0100, Ralph Corderoy wrote:

Hi Nate,

quoted

One manifestation of this is a race conditions in system(), which
(depending on the implementation) is non-atomic in that it first calls
a fork() and then an exec().

The need for O_CLOFORK might be made more clear by looking at a
long-standing Go issue, i.e. unrelated to system(3), which was started
in 2017 by Russ Cox when he summed up the current race-condition
behaviour of trying to execve(2) a newly created file:
https://github.com/golang/go/issues/22315.  I raised it on linux-kernel
in 2017, https://marc.info/?l=linux-kernel&m=150834137201488, and linked
to a proposed patch from 2011, ‘[PATCH] fs: add FD_CLOFORK and
O_CLOFORK’ by Changli Gao.  As I said, long-standing.

The problem is that people advocating for O_CLOFORK understand its
value, but not its cost.  Other google employees have a system which has
literally millions of file descriptors in a single process.  Having to
maintain this extra state per-fd is a cost they don't want to pay
(and have been quite vocal about earlier in this thread).

Fundamentally, fork()+exec() is a terrible model.  Mind you, so is
spawn().  I haven't seen a good model yet.