locking.c: Explain why the code has an apparently pointless `stat' call.

diff --git a/locking.c b/locking.c
index bfc5c771927e13d04703108224992ce92aef6933..6317b2e619888bd7c276f1c4d28aea4dfd7ee827 100644 (file)
--- a/locking.c
+++ b/locking.c
@@ -192,6 +192,13 @@ again:
   if (fstat(fd, &st))
     die(111, "error from fstat on `%s': %s", file, strerror(errno));
   err = fcntl(fd, f & f_wait ? F_SETLKW : F_SETLK, &l) >= 0 ? 0 : errno;
   if (fstat(fd, &st))
     die(111, "error from fstat on `%s': %s", file, strerror(errno));
   err = fcntl(fd, f & f_wait ? F_SETLKW : F_SETLK, &l) >= 0 ? 0 : errno;

+  /* It's tempting to `optimize' this code by opening a new file descriptor
+   * here so as to elide the additional call to fstat(2) above.  But this
+   * doesn't work: if we successfully acquire the lock, we then have two file
+   * descriptors open on the lock file, so we have to close one -- but, under
+   * the daft fcntl(2) rules, even closing `nfd' will release the lock
+   * immediately.
+   */

   if (stat(file, &nst)) {
     if (errno == ENOENT) { close(fd); goto again; }
     else die(111, "error from stat on `%s': %s", file, strerror(errno));
   if (stat(file, &nst)) {
     if (errno == ENOENT) { close(fd); goto again; }
     else die(111, "error from stat on `%s': %s", file, strerror(errno));