In article <v03130306b41858a56262@[204.189.11.34]> you write:
>It runs fine for several days and then suddenly the squid process goes
>CPU-bound and stops doing anything useful. The last entry in the cache.log
>file is always:
>
>comm_accept: FD 0: (14) Bad address
>
>There seems to be plenty of RAM available; the Squid process is using
>somewhere around 310 MB and there is approximately 100 MB Free (as reported
>by "top").
>
>Shutting down and restarting the server process brings the cache back to
>life for another several days until the same thing happens again.
>
>Does anyone have any ideas about diagnosing this problem? I didn't see
>this problem reported in the Hypermail archives. TIA...
I have exactly the same problem and it seems to be a bug in the 4.0
implementation of accept in BSD/OS. accept sometimes returns the Bad
address error under load and then also leaks a file descriptor in the
kernel; after a while you get 'too many open files' errors and sometime
the comm_accept failure makes Squid use 100% CPU.
I have patched the kernel (source license) to get rid of the leak but
the real accept problem hasn't been solved yet by BSDI. I already sent a
problem report to BSD/OS. If you have a support contract, please contact
support@bsdi.com too.
Arjan
Received on Thu Sep 30 1999 - 13:10:23 MDT
This archive was generated by hypermail pre-2.1.9 : Tue Dec 09 2003 - 16:48:36 MST