On Sun, Sep 9, 2018 at 1:04 AM, Alan Bateman <alan.bate...@oracle.com> wrote:
> On 09/09/2018 01:34, Martin Buchholz wrote: > > : > > See discussion in src/java.base/unix/native/libjava/ProcessImpl_md.c > > >> Later Oracle introduced something very similar with the jspawnhelper. >> But I see that jspawnhelper is not used at all for the vfork() case, >> just for posix_spawn(), yes? >> > > Implemented by other folks. I was mostly happy with vfork. I forget the > motivation to introduce jspawnhelper, but I'm sure it's in the fossil > record somewhere! > > jpsawnhelper is to deal with chdir and threading issues that arise when > using posix_spawn. The move to posix_spawn was motivated by eager > allocation of swap when using fork. It was originally done for Solaris, > then macOS, and I see the AIX implementation is declared to use posix_spawn > too. Linux is using vfork and seems to be stable. Martin will probably > remember a brief foray into clone(CLONE_VM) at one point but that didn't > work out. > Thanks. I followed my own advice and re-read the comments in ProcessImpl_md.c. I now recall not paying full attention to the posix_spawn implementation because I use Linux exclusively, where "my" vfork implementation works well. It seems too late to add HP-UX support to OpenJDK, but if SAP are willing to do the porting work I would support it (why didn't HP do it decades ago?). One decision to make for any Unix platform is which process implementation strategy to use.
