p5pRT
commented
17 years ago From perlbug@plan9.de
Created by perlbug@plan9.de
It seems that $| and autoflush ability is lost on the first EAGAIN.
Here is a test program that might or might not work (it should be run under a terminal emulator\, and it likely works best under linux because I tested it with that kernel) that illustrates the problem:
!/usr/bin/perl
use Fcntl;
open my $fh\, ">&1" or die; fcntl $fh\, F_SETFL\, O_NONBLOCK or die;
$|=1;
print "-"; syswrite $fh\, "\x00" x 8192; print "+"; sleep 1; print "*"; sleep 60;
Here is an explanation of what the program does:
1. get another filehandle for stdout 2. make it nonblocking (this also makes fd 1 nonblocking) 3. print a "-"\, which gets printed immediately. 4. write 8192 bytes\, which is hopefully more than the tty supports 5. print a single "+"\, causing a write()\, but results in EAGAIN 6. wait a while\, print "*"\, which should cause a write()\, but doesn't 7. give us time to look at the output.
Here is what I expect:
-* \<---cursor here
(-+* is another expected possibility)
here is what I usually get:
- \<---cursor here
If I strace the program (carefully)\, I get (among others)\, this:
write(1\, "-"\, 1) = 1 write(3\, "\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0"...\, 8192) = 4095 write(1\, "+"\, 1) = -1 EAGAIN (Resource temporarily unavailable) nanosleep({1\, 0}\, {1\, 0}) = 0 nanosleep({60\, 0}\, ...
As you can see\, "-" gets printed\, then our 8k dump fills the tty buffer\, then "+" gets written\, but gets EAGAIN.
After a second\, we write another "*"\, but perl will not even attempt the write.
This is IMHO a bug\, as $| should cause a flush\, and thus a write().
Doing $|=0;$|=1 flushes the buffer\, and the "*" appears\, but subsequent writes will still be buffered\, and I found no way of resurrecting STDOUT with autoflush\, short of closing and reopening it.
Please note that I didn't investigate this further (it could quite likely be stdio that would change buffering modes on EAGAIN\, for example\, and it took me some days to locate the code that causes this behaviour and to write a testcase).
I still think perl most be involved somehow\, as an explicit flush (setting $|=1) actually prints the buffer out fully - if perl would do a flush after every print\, as documented\, I would expect it to still just work\, regardless of what stdio thinks about buffering.
Please also note that the "+" usually gets lost\, which is somehow expected or not\, but the "*" and all subsequent prints get buffered internally and will be written on excplit flush (setting $|) but will not be autoflushed.
Migrated from rt.perl.org#41043 (status was 'open')
Searchable as RT41043$