[exim] Load average not being looked at in long connections?

Startseite
Nachricht löschen
Nachricht beantworten
Autor: Mark Moseley
Datum:  
To: exim-users
Betreff: [exim] Load average not being looked at in long connections?
Hi all. We've been working with a new commercial spam filter. It's sitting
on the edge in front of our Exim (4.66 - haven't gotten around to upgrading
-- running on Debian Sarge on Dell 750/850's) boxes. I'm seeing an odd
thing: Exim is queueing incoming messages from the spam filter box due to
high load even when the load is low. One interesting thing is that the spam
filter box keeps connections open for a long time to save on session start
up. All I can guess is that the initial load average gets reused over and
over, despite the fact that /proc/loadavg is rechecked for each message. Or
perhaps once it hit the 'queue_only_load' threshold, some flag got set that
hasn't gotten unset. Here's some bits extracted from strace:


19:21:22.044329 open("/proc/loadavg", O_RDONLY|O_LARGEFILE) = 7
19:21:22.044466 read(7, "29.97 24.32 17.25 17/149 30502\n", 40) = 31
...
19:21:22.103730 write(3, "2007-06-14 19:21:22 1Hyycz-0002rS-VA no immediate
delivery: load average 45.82\n", 79) = 79
...
19:21:45.552132 open("/proc/loadavg", O_RDONLY|O_LARGEFILE) = 7
19:21:45.552305 read(7, "22.91 23.15 17.05 1/127 30926\n", 40) = 30
...
19:21:45.598339 write(3, "2007-06-14 19:21:45 1HyydN-0002rS-FB no immediate
delivery: load average 45.82\n", 79) = 79
...
19:22:05.841367 open("/proc/loadavg", O_RDONLY|O_LARGEFILE) = 7
19:22:05.841599 read(7, "17.00 21.78 16.73 3/131 31091\n", 40) = 30
...
19:22:05.896607 write(3, "2007-06-14 19:22:05 1Hyydh-0002rS-Oj no immediate
delivery: load average 45.82\n", 79) = 79
...
19:22:46.319279 open("/proc/loadavg", O_RDONLY|O_LARGEFILE) = 7
19:22:46.319495 read(7, "9.28 19.19 16.07 1/128 31557\n", 40) = 29
...
19:22:46.373883 write(3, "2007-06-14 19:22:46 1HyyeM-0002rS-7q no immediate
delivery: load average 45.82\n", 79) = 79


... and so on. The connection in question has been connected for about 6
hours! Though I've observed it on a number of other processes on other boxes
also behind the spam filter box.

Obviously not a critical bug, but I figured worth mentioning. I'm still
using 4.66 but the changelog for 4.67 doesn't mention anything remotely
resembling this (barring blindness on my part).