[PATCH 2 of 3] Tests: handling of EAGAIN from sysread() with IO::Socket::SSL

[Sergey Kandaurov]

> On 11 Mar 2023, at 13:30, Maxim Dounin <mdounin at mdounin.ru> wrote:
> 
> Hello!
> 
> On Fri, Mar 10, 2023 at 08:00:05AM +0300, Maxim Dounin wrote:
> 
>> # HG changeset patch
>> # User Maxim Dounin <mdounin at mdounin.ru>
>> # Date 1678424071 -10800
>> #      Fri Mar 10 07:54:31 2023 +0300
>> # Node ID 49d12f8c4cf69e1cbe7feccae3b0ea1ac2ca8c2f
>> # Parent  fdebeebd07b160f1d30e18d56e64dfb08570f8b1
>> Tests: handling of EAGAIN from sysread() with IO::Socket::SSL.
>> 
>> With IO::Socket::SSL, when select() reports that the socket is readable,
>> reading from it might still fail with EAGAIN, since no application data is
>> available in the socket.  In particular, this might happen with TLSv1.3
>> when a session ticket is received after the handshake.  Fix is to explicitly
>> check for EAGAIN errors.
> 
> Err, IO::Socket::SSL actually generates EWOULDBLOCK rather than 
> EAGAIN, and this is important on some systems (notably Windows).
> 
> s/EAGAIN/EWOULDBLOCK/g;
> 
> # HG changeset patch
> # User Maxim Dounin <mdounin at mdounin.ru>
> # Date 1678522238 -10800
> #      Sat Mar 11 11:10:38 2023 +0300
> # Node ID 0fefa04c5be1e8095072d176cdf847c7c3766fbf
> # Parent  fdebeebd07b160f1d30e18d56e64dfb08570f8b1
> Tests: handling of EWOULDBLOCK from sysread() with IO::Socket::SSL.
> 
> With IO::Socket::SSL, when select() reports that the socket is readable,
> reading from it might still fail with EWOULDBLOCK, since no application
> data is available in the socket.  In particular, this might happen with
> TLSv1.3 when a session ticket is received after the handshake.  Fix is
> to explicitly check for EWOULDBLOCK errors.
> 
> diff --git a/lib/Test/Nginx/IMAP.pm b/lib/Test/Nginx/IMAP.pm
> --- a/lib/Test/Nginx/IMAP.pm
> +++ b/lib/Test/Nginx/IMAP.pm
> @@ -68,7 +68,9 @@ sub getline {
> 	while (IO::Select->new($socket)->can_read(8)) {
> 		$socket->blocking(0);
> 		my $n = $socket->sysread(my $buf, 1024);
> +		my $again = !defined $n && $!{EWOULDBLOCK};
> 		$socket->blocking(1);
> +		next if $again;
> 		last unless $n;
> 
> 		$self->{_read_buffer} .= $buf;
> diff --git a/lib/Test/Nginx/POP3.pm b/lib/Test/Nginx/POP3.pm
> --- a/lib/Test/Nginx/POP3.pm
> +++ b/lib/Test/Nginx/POP3.pm
> @@ -68,7 +68,9 @@ sub getline {
> 	while (IO::Select->new($socket)->can_read(8)) {
> 		$socket->blocking(0);
> 		my $n = $socket->sysread(my $buf, 1024);
> +		my $again = !defined $n && $!{EWOULDBLOCK};
> 		$socket->blocking(1);
> +		next if $again;
> 		last unless $n;
> 
> 		$self->{_read_buffer} .= $buf;
> diff --git a/lib/Test/Nginx/SMTP.pm b/lib/Test/Nginx/SMTP.pm
> --- a/lib/Test/Nginx/SMTP.pm
> +++ b/lib/Test/Nginx/SMTP.pm
> @@ -68,7 +68,9 @@ sub getline {
> 	while (IO::Select->new($socket)->can_read(8)) {
> 		$socket->blocking(0);
> 		my $n = $socket->sysread(my $buf, 1024);
> +		my $again = !defined $n && $!{EWOULDBLOCK};
> 		$socket->blocking(1);
> +		next if $again;
> 		last unless $n;
> 
> 		$self->{_read_buffer} .= $buf;
> diff --git a/lib/Test/Nginx/Stream.pm b/lib/Test/Nginx/Stream.pm
> --- a/lib/Test/Nginx/Stream.pm
> +++ b/lib/Test/Nginx/Stream.pm
> @@ -84,8 +84,10 @@ sub read {
> 	$s = $self->{_socket};
> 
> 	$s->blocking(0);
> -	if (IO::Select->new($s)->can_read($extra{read_timeout} || 8)) {
> -		$s->sysread($buf, 1024);
> +	while (IO::Select->new($s)->can_read($extra{read_timeout} || 8)) {
> +		my $n = $s->sysread($buf, 1024);
> +		next if !defined $n && $!{EWOULDBLOCK};
> +		last;
> 	}
> 
> 	log_in($buf);
> 

Looks good.

Note that this occurs with non-blocking sockets, as seen with diff.

From POD:
Using Non-Blocking Sockets
   If you have a non-blocking socket, the expected behavior on read,
   write, accept or connect is to set $! to EWOULDBLOCK if the operation
   cannot be completed immediately. Note that EWOULDBLOCK is the same as
   EAGAIN on UNIX systems, but is different on Windows.

This reminds me how OpenSSL has changed the SSL_MODE_AUTO_RETRY default,
then IO::Socket::SSL disabled it again on non-blocking sockets.

-- 
Sergey Kandaurov