replace EGAIN with EWOULDBLOCK - these are same on UNIX but on Windows y... #11
Conversation
…s you need the latter one
|
For what it's worth we've run into this and this patch solves the problem |
|
I think that the change will need to be applied to https://github.com/libwww-perl/libwww-perl/blob/master/lib/LWP/Protocol/http.pm too. |
Yes, this is done in libwww-perl/libwww-perl#66 |
|
To be properly correct, you'd have to check for both |
You are right. Adapted the fix accordingly. |
|
Wow. This issue was driving me crazy! I had to patch both Net/HTTP/Methods.pm and LWP/Protocol/HTTP.pm. The only workaround without patching them is to use IO::Socket::SSL version 2.005, which is the last version still not supporting non-blocking sockets on Windows. |
|
I'm uncomfortable merging this without more voices in agreement - e.g. there was just a big discussion on the perl5porters list about EAGAIN/EWOULDBLOCK on various platforms (and e.g. see https://api.metacpan.org/source/MLEHMANN/AnyEvent-7.09/Changes). |
|
As far as I can see from the Changes and the discussion the problems described affect EWOULDBLOCK only if |
|
@noxxi Can you rebase this please? |
|
That's not a rebase. |
… but not on Windows)
…se_size % 1024 == 0
my_readline() is broken for response sizes (headers + raw body) that are multiples of 1024 (sysread block size).
When reading from date from the socket, we keep reading until we get fewer bytes than we asked for. The code implicitly assumes that if we're getting *exactly* what we asked for, then there is more data available. That assumption does not hold true for response sizes of 1024, 2048, etc...
In most cases, this bug is masked by either not using keep-alive or by a lot of servers having short keep-alive timeouts. It is easily reproducible though by using a simple Plack one-liner and tweaking the response size until hitting e.g. 1024 bytes in the first sysread(). The following works with starman:
$ cat foo.psgi
sub { [ 200, [], [ "x" x 904 ] ] } # Tweak response size as needed, depending on headers sent by starman
$
$ starman --keepalive-timeout 10 foo.psgi
Elsewhere, run:
$ time perl -MLWP::UserAgent -e 'print LWP::UserAgent->new(keep_alive => 1)->post(q(http://127.0.0.1:5000))->status_line'
200 OK
real 0m10.067s
user 0m0.052s
sys 0m0.004s
$
Net::HTTP hangs, waiting for more data until the server disconnects.
The simplest fix seems to be just checking can_read() before another read if we get exactly what we asked for in sysread().
…CE at apache.org. Apache recently changed their server config and no longer allow TRACE requests.
Test against http://httpbin.org/headers and iterate until responses are exactly 1024, 2048 or 3072 bytes in size.
|
Something got messed up while trying to rebase. I'll restart with a new repo because I cannot sort this out. Continued in #24. |
Since 2.006 IO::Socket::SSL supports non-blocking on Windows. But on Windows the relevant $ERRNO is EWOULDBLOCK, not EAGAIN. On Unix these both are the same. Thus the code needs be changed to use EWOULDBLOCK instead.
This patch needs to go together with libwww-perl/libwww-perl#66
See also http://stackoverflow.com/a/27863737/3081018