[3.11] gh-104372: Drop the GIL around the vfork() call. (#104782) (#104958)

gh-104372: Drop the GIL around the vfork() call. (#104782)

On Linux where the `subprocess` module can use the `vfork` syscall for
faster spawning, prevent the parent process from blocking other threads
by dropping the GIL while it waits for the vfork'ed child process `exec`
outcome.  This prevents spawning a binary from a slow filesystem from
blocking the rest of the application.

Fixes #104372.

(cherry picked from commit d08679212d9af52dd074cd4a6abb440edb944c9c)

1 files changed, 19 insertions, 2 deletions

diff --git a/Modules/_posixsubprocess.c b/Modules/_posixsubprocess.c
index ad9daae..072519c 100644
--- a/Modules/_posixsubprocess.c
+++ b/Modules/_posixsubprocess.c
@@ -537,7 +537,7 @@ reset_signal_handlers(const sigset_t *child_sigmask)
  * required by POSIX but not supported natively on Linux. Another reason to
  * avoid this family of functions is that sharing an address space between
  * processes running with different privileges is inherently insecure.
- * See bpo-35823 for further discussion and references.
+ * See https://bugs.python.org/issue35823 for discussion and references.
  *
  * In some C libraries, setrlimit() has the same thread list/signalling
  * behavior since resource limits were per-thread attributes before
@@ -774,6 +774,7 @@ do_fork_exec(char *const exec_array[],
     pid_t pid;
 
 #ifdef VFORK_USABLE
+    PyThreadState *vfork_tstate_save;
     if (child_sigmask) {
         /* These are checked by our caller; verify them in debug builds. */
         assert(!call_setuid);
@@ -781,8 +782,23 @@ do_fork_exec(char *const exec_array[],
         assert(!call_setgroups);
         assert(preexec_fn == Py_None);
 
+        /* Drop the GIL so that other threads can continue execution while this
+         * thread in the parent remains blocked per vfork-semantics on the
+         * child's exec syscall outcome. Exec does filesystem access which
+         * can take an arbitrarily long time. This addresses GH-104372.
+         *
+         * The vfork'ed child still runs in our address space. Per POSIX it
+         * must be limited to nothing but exec, but the Linux implementation
+         * is a little more usable. See the child_exec() comment - The child
+         * MUST NOT re-acquire the GIL.
+         */
+        vfork_tstate_save = PyEval_SaveThread();
         pid = vfork();
-        if (pid == -1) {
+        if (pid != 0) {
+            // Not in the child process, reacquire the GIL.
+            PyEval_RestoreThread(vfork_tstate_save);
+        }
+        if (pid == (pid_t)-1) {
             /* If vfork() fails, fall back to using fork(). When it isn't
              * allowed in a process by the kernel, vfork can return -1
              * with errno EINVAL. https://bugs.python.org/issue47151. */
@@ -795,6 +811,7 @@ do_fork_exec(char *const exec_array[],
     }
 
     if (pid != 0) {
+        // Parent process.
         return pid;
     }