free pages in 2.1 and 2.0, was: 2.0.30 bends but doesn't break mke2fs

bofh@snoopy.virtual.net.au
Wed, 16 Apr 97 10:25:56 +1100

>> > I've seen similar messages for quite a while, but the changes to buffer.c in
>> > 2.0.30 do seem to make them more common. I'm fairly sure that some work is
>> > still needed in 2.0.30+ to be more agressive in making memory available to the
>> > kernel when necessary.
>> >
>> > Leonard
>>
>> It is relatively easy to cause slower machines (386's) with low
>> mem (4 -> 8MB) to get stuck in an endless loop spewing this message once
>> every 10s even when there are many MB of free swap still available.
>> Hitting reset is about the only way out (happens with 2.0.29 and 2.1.x).

>On my 386, 8MB the problem started with 2.0.30. It seems to be triggered
>by high NFS activity.

>> A cure for the symptom (but not a real fix for the problem) is to up the
>> /proc/sys/vm/freepages values. The defaults of 16/24/32 for all machines
>> with 8MB or less RAM seem simply too low. With 30/40/50 on a 5MB 386-40
>> the problem never rears its ugly head.

>Yes, with this little patch my problems went away. Maybe this should
>become standard because otherwise small machines may become unreliable.
>Without the patch, the system will hang. I'm testing without any swap
>activated.

I've found the default in 2.1.33 of 32/48/64 to be too low for a machine with 16megs of RAM. If we're going to increase the minimum then I think that we should use a number higher than 30. Of course the best solution would be to fix whatever problem is in the kernel that causes it to crash. Earlier kernels didn't crash on me on the default settings of 32/48/64...

Russell Coker