Re^19: Net::OpenSSH loosing lines ins reply

write(1, "torm-control broadcast level 10."..., 4096) = 4096
write(1, "rol broadcast level 10.00\n1782: "..., 4096) = -1 EAGAIN (Re
+source temporarily unavailable)
[download]

The only plausible explanation I can think of is OpenSSH ssh setting STDOUT into non-blocking mode. BTW, how are you capturing STDOUT?

Add a dump of /proc/$$/fdinfo:

open OUT, ">", "./trace.out";

for (0, 1, 2) {
    local $/;
    open my $fdinfo, '<', "/proc/$$/fdinfo/$_";
    my $info = <$fdinfo>;
    print OUT "fdinfo $_:\n$info\n\n";
}

my $line =0;
foreach (@cmdout) {
    $line++;
    my $bytes = print  $line.": ".$_;
    print OUT "$line: bytes: $bytes, err: $! \n";
}
close(OUT);
[download]

The non-blocking flag is 0x4000. You may also like to generate dumps of fdinfo before the constructor call and before and after calling capture.

Comment on Re^19: Net::OpenSSH loosing lines ins reply Select or Download Code

Replies are listed 'Best First'.
Re^20: Net::OpenSSH loosing lines ins reply by Andy16 (Acolyte) on Jun 05, 2014 at 09:28 UTC
Hi Salva, on a normal run: `============================== start fdinfo 0: pos: 0 flags: 0100002 fdinfo 1: pos: 0 flags: 01 fdinfo 2: pos: 0 flags: 01 ============================== after new fdinfo 0: pos: 0 flags: 0100002 fdinfo 1: pos: 0 flags: 01 fdinfo 2: pos: 0 flags: 01 ============================== after capture fdinfo 0: pos: 0 flags: 0100002 fdinfo 1: pos: 0 flags: 04001 fdinfo 2: pos: 0 flags: 04001` [download] for a failed run: `============================== start fdinfo 0: pos: 0 flags: 0100002 fdinfo 1: pos: 0 flags: 01 fdinfo 2: pos: 0 flags: 01 ============================== after new fdinfo 0: pos: 0 flags: 0100002 fdinfo 1: pos: 0 flags: 01 fdinfo 2: pos: 0 flags: 01 ============================== after capture fdinfo 0: pos: 0 flags: 0100002 fdinfo 1: pos: 0 flags: 04001 fdinfo 2: pos: 0 flags: 04001` [download] so the same.... doesn't mean any thing to me... :-( I capture output using "tee" or even simpler only using "wc -l". For really using the script it would be called by another script and stdout would be read by caller.	[reply] [d/l] [select]
Re^21: Net::OpenSSH loosing lines ins reply by salva (Canon) on Jun 05, 2014 at 09:57 UTC
As I had supposed, after the `capture` call, the `O_NONBLOCK` flag is set. The `ssh` command that is being run under the hood by `capture` is leaving `STDOUT` in non-blocking mode. The issue seems fixed in newer versions of OpenSSH, or at least I am unable to reproduce it with the latest one (6.6.1p1). In any case it is easy to workaround. Just using `capture2` instead of `capture` or setting `stderr_discard => 1` should make it go. BTW, there isn't any difference on the `fdinfo` dumps between the successfully and failed invocations because the problem is hidden behind a race condition. The perl process needs to be faster writing than `tee` reading for the intermediate pipe buffer to fill.	[reply] [d/l] [select]
Re^22: Net::OpenSSH loosing lines ins reply by Andy16 (Acolyte) on Jun 05, 2014 at 10:45 UTC
Hi Salva, GREAT changed to capture2 and problem did not reoccur up to now... Follwing your explanations - it shall be solved now! hero of my day! no chance for me digging that out....	[reply]

GREAT