The following reply was made to PR general/1680; it has been noted by GNATS.
From: Marc Slemko <[EMAIL PROTECTED]> To: Morgan Davis <[EMAIL PROTECTED]> Cc: Apache bugs database <[EMAIL PROTECTED]> Subject: Re: general/1680: Critically anomalous behavior when upgrading from 1.3a1 to 1.3b3 Date: Thu, 15 Jan 1998 18:45:58 -0700 (MST) On Thu, 15 Jan 1998, Morgan Davis wrote: > At 04:05 PM 1/15/98 -0700, you wrote: > >On Thu, 15 Jan 1998, Morgan Davis wrote: > > > >> At 02:38 PM 1/15/98 -0700, you wrote: > >> >On 15 Jan 1998, Morgan Davis wrote: > >> > > >> >> Compiled 1.3b3 from official sources with the no2slash patch, > >> >> and compiler options and modules identical to the 1.3a1 previously > >> >> built on the same server (same OS version). Runs fine while load > >> >> is light. But under load (approx. 100 virtual hosts taking about > >> >> 700 cumulative hits a minute), the parent and most children will > >> >> suddenly die off, with a single child remaining as a zombie. > >> >> Subsequent attempts to restart httpd fail since Port 80 is already > >> >> bound by the zombie child. Prior to the collapse, odd behavior is > >> >> noticed, such as references to a particular virtual host end up > >> >> bringing up a totally different virtual host site, or there is a > >> >> failure to connect at all. A reboot is necessary to get the server > >> >> back up, but will eventually collapse again with exact same state. > >> >> No core file to be found. > >> > > >> >Are you using NFS? > >> > >> Yes. > >> > >Where is your ServerRoot? If it is NFS mounted, you should move it to a > >local directory using the LockFile directive (eg. LockFile > >/var/run/accept.lock). > > Done that already. Here are the relevant settings: > > ServerRoot /usr/local/apache > LockFile /var/spool/lock/accept.lock > > Both of these are on a local filesystem. Also, PID and Log files all point > to local filesystems. Only DocumentRoot points to the NFS filesystem. > > Keep in mind that 1.3a1 is using the exact same settings as was 1.3b3, and > it doesn't exhibit this behavior. Yes, but USE_FLOCK_SERIALIZED_ACCEPT was broken with 1.3a1. It was fixed in b2. > > >If this isn't the case or don't help, there are a few other things to try. > > Ready and willing. Comment out the USE_FLOCK_SERIALIZED_ACCEPT bit from the section appropriate to your OS in src/main/conf.h. You could also try using 1.3b2 to help narrow down where the difference comes in... Are you _sure_ the child is a zombie? If the parent is dead, it shouldn't be. What releveant processes and info does a ps -axlw show? Nothing in the error log? Are you sure you aren't running low on file descriptors? Unlikely anything should change between a1 and b3, but...
