problems with UNIX domain sockets

Hi all,

I've implemented the webserver side of the FastCGI protocol.  I've built 
some test apps in Perl using the Perl FCGI library version 0.67, which is 
based on the FastCGI development kit version 2.4.0.  I'm having two 
problems.

#1
The library closes the connection if it's a UNIX domain socket and it does 
not receive any data within 2 seconds.  There is a big long comment in 
os_unix.c about how this is supposed to work around an obscure bug in Linux 
2.0.x and SCO UnixWare wherein you could accept a connection that was 
already closed on the client side, blah blah blah.  I see in the mailing 
list archives that this has been mentioned before, but I don't think 
anything has been done about it yet.  The problem I'm having is that my 
webserver creates a connection to the FastCGI app at init time and keeps it 
open (persistent); it only sends data to the FastCGI app when there is an 
actual request, which probably won't be in the first 2 seconds.  I've chosen 
this model for performance reasons.

Shouldn't this kludge be off by default, and only on for those systems that 
are affected?  By "those systems that are affected", I mean you could either 
use #ifdefs based on architecture, or perhaps even a simple simulation could 
be performed at compile time to determine if the system is affected.  I know 
that I can make a special versin of the FCGI library myself to disable the 
kludge, but given that my aim is to incorporate FastCGI support into a 
*webserver*, I want that webserver to be able to work with a wide variety of 
FastCGI apps out there.  This kludge is very annoying.

#2
To try to work around the kludge without modifying the Perl FCGI library, I 
decided that I could have my webserver send an FCGI_GET_VALUES message to 
the FastCGI app immediately after connecting to it.  Here is a hexdump of 
the exact message I send, in case somebody can spot something I'm doing 
wrong:

01 09 00 00  00 30 00 00  0E 00 46 43  47 49 5F 4D  .....0....FCGI_M
41 58 5F 43  4F 4E 4E 53  0D 00 46 43  47 49 5F 4D  AX_CONNS..FCGI_M
41 58 5F 52  45 51 53 0F  00 46 43 47  49 5F 4D 50  AX_REQS..FCGI_MP
58 53 5F 43  4F 4E 4E 53                            XS_CONNS


The problem is I never get a response back; the FastCGI app reads in the 
three name-value pairs but then enters a perma-blocked state trying to read 
more name-value pairs.  In fcgiapp.c, line 1464, ProcessManagementRecord() 
calls ReadParams() which starts to read the name-value pairs.  But it 
doesn't know when to stop.  Once it has exhausted the three pairs, it keeps 
on going and calls FCGX_GetChar() again, which ends up doing a read() 
because there's no more data.  It sits in the read(), so ReadParams() never 
returns, and consequently we never get back into ProcessManagementRecord() 
so that it can send back the FCGI_GET_VALUES_RESULT message.  BTW this 
problem is not specific to UNIX domain sockets.

So in order to get the FCGI_GET_VALUES / FCGI_GET_VALUES_RESULT exchange to 
work, I would have to hack up the FCGI library which is what I wanted to 
avoid.  Basically I have no way of communicating reliably with an 
off-the-shelf FCGI library over a UNIX domain socket in my model where I 
don't necessarily send data immediately after connecting.

Any help is greatly appreciated!

___________________________________
fastcgi-developers mailing list
http://fastcgi.com/fastcgi-developers/