Synchronization problem with posix_spawn

[Corinna Vinschen]

Hi Ken,

On Jul 30 13:59, Corinna Vinschen wrote:
> On Jul 29 19:12, Ken Brown via Cygwin wrote:
> > On 7/29/2020 4:17 PM, Ken Brown via Cygwin wrote:
> > > posix_spawn(p) returns before the spawned process is fully up and
> > > running.  [...]
> > I just took a look at the source, and I see that posix_spawn was taken from
> > FreeBSD.  Does FreeBSD have the same problem?  Should applications just be
> > aware of this issue and insert a sleep after posix_spawn before sending
> > signals?
> 
> Actually, this is a Cygwin problem.  I just had a look into the
> newlib implementation myself, and it turns out that the code,
> in particular the do_posix_spawn() function, is BSD specific.  It
> relies on the BSD implementation of vfork(2).  Cygwin's vfork(2)
> on the other hand, is NOT following the historic idea of the
> BSD vfork(2), rather it's equivalent to fork(2).  This is
> POSIX compliant, but certainly the reliance of the BSD vfork
> behaviour makes do_posix_spawn() as it's implemented right now,
> not overly functional for Cygwin.
> 
> IOW, we need a Cygwin-specific do_posix_spawn() using fork(2)
> in conjunction with some synchronization the BSD function
> gets "for free" by using its specific vfork(2).

Below is a POC implementation for a Cygwin-specific do_posix_spawn().
If this does the trick (at least your testcase works in my testing),
then I'm planning to move the function over to the winsup/cygwin dir
so it can be streamlined further.

Can you give it a try?


Thanks,
Corinna


diff --git a/newlib/libc/posix/posix_spawn.c b/newlib/libc/posix/posix_spawn.c
index 19c5cd0fe986..3adbac29d7fd 100644
--- a/newlib/libc/posix/posix_spawn.c
+++ b/newlib/libc/posix/posix_spawn.c
@@ -254,6 +254,82 @@ process_file_actions(const posix_spawn_file_actions_t fa)
 	return (0);
 }
 
+#ifdef __CYGWIN__
+#include <windows.h>
+#include <sys/mman.h>
+#include <process.h>
+
+typedef struct
+{
+  pid_t pid;
+  int error;
+} child_info;
+
+static int
+do_posix_spawn(pid_t *pid, const char *path,
+	const posix_spawn_file_actions_t *fa,
+	const posix_spawnattr_t *sa,
+	char * const argv[], char * const envp[], int use_env_path)
+{
+	pid_t p;
+	int ret = 0;
+	child_info *ci;
+
+	ci = (child_info *) mmap(NULL, sizeof(child_info),
+				 PROT_READ | PROT_WRITE,
+				 MAP_SHARED | MAP_ANONYMOUS, -1, 0);
+	if (ci == MAP_FAILED)
+		return (errno);
+
+	/* Cygwin's vfork does not follow BSD vfork semantics.  Rather it's
+	   equivalent to fork.  While that's POSIX compliant, it means the
+	   below FreeBSD implementation relying on BSD vfork semantics
+	   doesn't work as expected on Cygwin. */
+	p = fork();
+	switch (p) {
+	case -1:
+		ret = errno;
+		break;
+	case 0:
+		if (sa != NULL) {
+			ci->error = process_spawnattr(*sa);
+			if (ci->error)
+				_exit(127);
+		}
+		if (fa != NULL) {
+			ci->error = process_file_actions(*fa);
+			if (ci->error)
+				_exit(127);
+		}
+		if (use_env_path)
+			p = spawnvpe(_P_NOWAITO, path,
+				     (const char * const *) argv,
+				     (const char * const *)
+				     (envp != NULL ? envp : *p_environ));
+		else
+			p = spawnve(_P_NOWAITO, path,
+				    (const char * const *) argv,
+				    (const char * const *)
+				    (envp != NULL ? envp : *p_environ));
+		if (p < 0) {
+			ci->error = errno;
+			_exit(127);
+		}
+		ci->pid = p;
+		_exit (0);
+	default:
+		if (waitpid(p, NULL, 0) < 0)
+			ret = errno;
+		else if (ci->error)
+			ret = ci->error;
+		else if (pid != NULL)
+			*pid = ci->pid;
+		break;
+	}
+	munmap(ci, sizeof(child_info));
+	return ret;
+}
+#else
 static int
 do_posix_spawn(pid_t *pid, const char *path,
 	const posix_spawn_file_actions_t *fa,
@@ -292,6 +368,7 @@ do_posix_spawn(pid_t *pid, const char *path,
 		return (error);
 	}
 }
+#endif
 
 int
 posix_spawn (pid_t *pid,

-- 
Corinna Vinschen
Cygwin Maintainer