kswapd (& clock?) problems in 2.6.10

kswapd (& clock?) problems in 2.6.10

[aeriksson]

Since installing 2.6.10 kswapd has decided to loop wildly on two
occations. Both occations happened after starting a big compiles.
Checking vmstat, I noticed a steady stream of io/bi figures ranging
between 2-5K (which is about what I can get out of this box during
normal operation), but no swap io(?). Manwhile the load (xload)climbs
to about 10, and the system looses interactive responsiveness.

Now after the last time, after reboot, I noticed that the clock was
missing 3 hours, which is about the time I left the box unattended. It
might be that kswapd stresses the system to the point that the clock
cannot move forward, (or the clock does something bad and kswapd gets
tricked by it???).

Thoughts anyone?

/Anders

Re: kswapd (& clock?) problems in 2.6.10

[Nick Piggin]

aeriksson@fastmail.fm wrote:
> Since installing 2.6.10 kswapd has decided to loop wildly on two
> occations. Both occations happened after starting a big compiles.
> Checking vmstat, I noticed a steady stream of io/bi figures ranging
> between 2-5K (which is about what I can get out of this box during
> normal operation), but no swap io(?). Manwhile the load (xload)climbs
> to about 10, and the system looses interactive responsiveness.
> 
> Now after the last time, after reboot, I noticed that the clock was
> missing 3 hours, which is about the time I left the box unattended. It
> might be that kswapd stresses the system to the point that the clock
> cannot move forward, (or the clock does something bad and kswapd gets
> tricked by it???).
> 
> Thoughts anyone?
> 

Are you sure it is a bug? kswapd has been recently fixed so it properly
starts freeing memory before the critical shortage is reached where
processes have to synchronously free memory for themselves.

This will make kswapd take more CPU time, but it should be doing work
that would otherwise need to be done by someone else. The lack of any
swapping activity indicates that kswapd is not actually having too much
trouble freeing memory (or you have swap turned off!)

If you still think it is a bug, if kswapd keeps running after the compile
stops, or if the system stops doing useful work, then let's see what the
problem is:

Can you make sure you kernel is compiled with "magic sysrq key" turned
on (under kernel hacking menu). While the weird kswapd behaviour happens
next, press alt+sysrq+m 3 or 4 times, then dump dmesg to a file
`dmesg > dmesg.out`, and send that file in.

Thanks,
Nick