On Mar 18 21:46, Corinna Vinschen wrote: > On Mar 18 13:10, Jeremy Drake via Cygwin-developers wrote: > > On Tue, 18 Mar 2025, Corinna Vinschen wrote: > > > > > Fodder for future Cygwin development: > > > > > > If you have other ideas what could be helpful for Cygwin, don't be shy > > > to add it to this thread. Ideally with the intention to look into > > > implementing it by yourself, but at least a certain willingness to look > > > into the source code and discuss potential ways to implement it given > > > the current framework. > > > > I periodically think about posix_spawn and the potential to fast-path > > simple cases to one of the spawn functions instead of a fork/exec. This > > should give a relatively good effort to performance improvement ratio. > > One of the biggies here is the way descriptor passing works in Cygwin. > > We're coming from a historical API model which didn't know the > STARTUPINFOEX datatype, nor the UpdateProcThreadAttribute function > and PROC_THREAD_ATTRIBUTE_HANDLE_LIST. Thus, Cygwin is handling > handles free-form. Some handles required for Cygwin internally are > simply inherited, FD_CLOEXEC is implemented by closing the handles > in the forked child, etc. > > For posix_spawn without forking, this complicates matters. For > instance, we don't want having to close FD_CLOEXEC handles in the > spawned child because that's a security problem. > > An excellent preparation for a faster posix_spawn which spawns w/o > forking would be: > > - Full handle bookkeeping. Every single handle opened by Cygwin (not > only those in an fhandler) is added to a list with the information > if it has to be inherited by forked and execed/spawned children. > And maybe more generic info, but that's not the point here. > > - After having all handles booked in this central list, change > frok::parent and child_info_spawn::worker to collect all handles > to be inherited in a PROC_THREAD_ATTRIBUTE_HANDLE_LIST. > > Incidentally this would not only simplify posix_spawn. It would > catch the new FD_CLOFORK almost automatically.
...and even simplify FD_CLOEXEC if I didn't miss something. This might even speed up fork() slightly. > > > Corinna
