Calling system() in multi-threads.

[Takashi Yano]

On Fri, 18 Jul 2025 00:44:46 +0900
Takashi Yano wrote:
> On Thu, 17 Jul 2025 17:19:49 +0200
> Corinna Vinschen wrote:
> > On Jul 17 23:14, Takashi Yano via Cygwin wrote:
> > > Hi Corinna,
> > > 
> > > On Wed, 16 Jul 2025 17:36:42 +0200
> > > Corinna Vinschen wrote:
> > > > On Jul 16 23:52, Takashi Yano via Cygwin wrote:
> > > > > Do you have any idea?
> > > > 
> > > > Locking would be super-simple.
> > > > 
> > > > But theoretically it should be possible to use a local child_info_spawn
> > > > variable at this point.  The ch_spawn child_info_spawn instance is not
> > > > copied to the child anyway, so that should be safe.  The same goes for
> > > > posix_spawn() then, btw.
> > > > 
> > > > I checked the sources and I don't see any dependency to ch_spawn
> > > > from a spawning process, in contrast to an exec'ing process.  That
> > > > doesn't mean there is none, just that I didn't find any.
> > > 
> > > Thanks!
> > > As a starting point, I tried tntroducing locking. It almost works
> > > as expected, however, sometimes my STC in my first report is hangs
> > > if N is large e.g. 100. The patch is as attached.
> > > 
> > > What am I missing?
> > 
> > I don't know.  You're perhaps not releasing the lock in all cases.
> > But I would have to debug this just like you ¯\_(ツ)_/¯
> > 
> > Out of curiosity, did you try using a locale child_info_spawn instance
> > instead?  That would be a rather nice solution, but I'm pretty sure
> > there's some other problem lurking in the dark...
> 
> I'm not sure what to do with local child_info_spawn.
> Some of other modules refer to ch_spawn, such as exception.cc and
> pinfo.cc. Also, has_execed* uses ch_spawn. What should we do for that?
> 
> I've just tried simply the following patch, however, this also hangs
> with my STC.
> 
> diff --git a/winsup/cygwin/spawn.cc b/winsup/cygwin/spawn.cc
> index cb58b6eed..56fca6e45 100644
> --- a/winsup/cygwin/spawn.cc
> +++ b/winsup/cygwin/spawn.cc
> @@ -944,6 +944,7 @@ spawnve (int mode, const char *path, const char *const *argv,
>    int ret;
>  
>    syscall_printf ("spawnve (%s, %s, %p)", path, argv[0], envp);
> +  child_info_spawn ch_spawn_local;
>  
>    if (!envp)
>      envp = empty_env;
> @@ -951,7 +952,7 @@ spawnve (int mode, const char *path, const char *const *argv,
>    switch (_P_MODE (mode))
>      {
>      case _P_OVERLAY:
> -      ch_spawn.worker (path, argv, envp, mode);
> +      ch_spawn_local.worker (path, argv, envp, mode);
>        /* Errno should be set by worker.  */
>        ret = -1;
>        break;
> @@ -961,7 +962,7 @@ spawnve (int mode, const char *path, const char *const *argv,
>      case _P_WAIT:
>      case _P_DETACH:
>      case _P_SYSTEM:
> -      ret = ch_spawn.worker (path, argv, envp, mode);
> +      ret = ch_spawn_local.worker (path, argv, envp, mode);
>        break;
>      default:
>        set_errno (EINVAL);

The hang seems to be at acquiring the cygheap_protect lock in child sh.exe.
This lock is aquired only in _cfree() and _cmalloc(), so I am not sure why
cygheap_protect cannot be acquired at this point at all...

-- 
Takashi Yano <takashi.yano@nifty.ne.jp>