> There are a couple of signals for which this is the appropriate behavior.
> The right fix is to install handlers with the SA_RESTART flag set. The
> next version of bash will install its SIGWINCH handler with SA_RESTART.
Here is a trivial patch doing just that, though IMHO it would be much better
and more consistent to check for EINTR and restart the write() system call,
as it's already done for open() in redir.c:redir_open(). The EINTR "error"
should never be user visible, no matter what signal had caused it.
The problem could also be reproduced with
printf '%065535d' 1 | sleep 10000
followed by a terminal resize, but not with
mkfifo /tmp/fifo; exec 7</tmp/fifo
because the EINTR is handled correctly in redir_open().
diff --git a/sig.c b/sig.c
index e5bb7399..cc2456c5 100644
--- a/sig.c
+++ b/sig.c
@@ -736,12 +736,17 @@ set_signal_handler (sig, handler)
/* We don't want a child death to interrupt interruptible system calls, even
if we take the time to reap children */
#if defined (SIGCHLD)
if (sig == SIGCHLD)
act.sa_flags |= SA_RESTART; /* XXX */
#endif
+ /* And neither should a terminal resize */
+#if defined (SIGWINCH)
+ if (sig == SIGWINCH)
+ act.sa_flags |= SA_RESTART; /* XXX */
+#endif
/* If we're installing a SIGTERM handler for interactive shells, we want
it to be as close to SIG_IGN as possible. */
if (sig == SIGTERM && handler == sigterm_sighandler)
act.sa_flags |= SA_RESTART; /* XXX */
sigemptyset (&act.sa_mask);
