On Mon, Aug 23, 2010 at 12:32 AM, Erik Faye-Lund<address@hidden> wrote:
I've recently looked into replacing the poll-emulation in Git for
Windows with the code from gnulib, due to the improved pipe-support in
gnulib.
However, I've noticed that win32_compute_revents() never returns
SIGHUP for FILE_TYPE_PIPE, which is something the code I'm trying to
port currently depends on. What seems to happen is the following: one
of the pipe-ends is closed, and thus PeekNamedPipe() fails.
win32_compute_revents() mistakenly believes that the failure means
that the pipe can be written because NtQueryInformationFile succeeds
and returns 8192 for both fpli.WriteQuotaAvailable and
fpli.OutboundQuota.
In my case, a simple "happened |= POLLHUP;"-hack in the error-case
solves the issue in this particular case, but this is obviously not
correct in the general case.
My guess is that a proper solution would be to find device a robust
way of checking if a pipe is closed or not. I haven't looked much into
this yet, but at least GetNamedPipeInfo() doesn't seem to error out on
my pipe.
The following seems to work:
diff --git a/lib/poll.c b/lib/poll.c
index 49c4159..83532c4 100644
--- a/lib/poll.c
+++ b/lib/poll.c
@@ -150,6 +150,8 @@ win32_compute_revents (HANDLE h, int *p_sought)
if (avail)
happened |= *p_sought& (POLLIN | POLLRDNORM);
}
+ else if (GetLastError() == ERROR_BROKEN_PIPE)
+ happened |= POLLHUP;
else
{