Another Shambhala snapshot... [this one's more stable]

["Robert S. Thau" <rs...@ai.mit.edu>]

I've put another snapshot of my "server-by-components", Shambhala, up
for anonymous ftp, in ftp://ftp.ai.mit.edu/pub/users/rst/shambhala3.tar.Z.

The main difference between this snapshot and the last is stability
--- this code has been on my port 80 since Sunday noon without
apparent problems, although that was admittedly a low-traffic day
here.  (It only served a bit over 20K hits during prime time, before
I restarted it to make sure that the 3:00 AM restart would work; I'd
like to try it with real traffic, but we are hosting the International
Conference on Computer Vision this week, with the conference pages on
the lab main server, and so if Shambhala causes any problems, no
matter how minor, I'm booting it instantly to port 8000).

So far, I haven't noticed any particular descriptor leak problems
(from noon to 8:00 yesterday, when I was around watching it, the
number of open file descriptors system-wide, as reported by pstat -T,
stayed stable at around 200 for pretty much the whole time, with the
original pool of 20 processes still running).

As to features, it's the same story as the last time around (no
virtual hosts, no XBITHACK, most of the other stuff works), although
some of it, particularly basic authentication, has had more of a
workout, and some bugfixes; also, ErrorDocument passes a smoke test,
and it's finally setting the DOCUMENT_ROOT CGI variable.  I still
consider it alpha-test code; it works for me, at least when I'm there
watching it, but your mileage may vary.

As a side note, what made the difference between this version and last
week's was apparently guarding the critical sections in alloc.c.
(There are wrapper routines around fopen, etc., which remember the
whatever-it-is that is being allocated, to be sure it gets closed
later; they now take care that if any timeouts or client aborts
trigger while they are running, they're put off until after the
allocation routine is done.  This sort of protection may well be
necessary if you want to handle timeouts and keep on running...).

rst