[PATCH v2 02/28] vmscan: take at least one pass with shrinkers

Glauber Costa glommer at parallels.com
Mon Apr 8 08:47:14 UTC 2013


On 04/08/2013 12:42 PM, Joonsoo Kim wrote:
> Hello, Glauber.
> 
> On Fri, Mar 29, 2013 at 01:13:44PM +0400, Glauber Costa wrote:
>> In very low free kernel memory situations, it may be the case that we
>> have less objects to free than our initial batch size. If this is the
>> case, it is better to shrink those, and open space for the new workload
>> then to keep them and fail the new allocations.
>>
>> More specifically, this happens because we encode this in a loop with
>> the condition: "while (total_scan >= batch_size)". So if we are in such
>> a case, we'll not even enter the loop.
>>
>> This patch modifies turns it into a do () while {} loop, that will
>> guarantee that we scan it at least once, while keeping the behaviour
>> exactly the same for the cases in which total_scan > batch_size.
> 
> Current user of shrinker not only use their own condition, but also
> use batch_size and seeks to throttle their behavior. So IMHO,
> this behavior change is very dangerous to some users.
> 
> For example, think lowmemorykiller.
> With this patch, he always kill some process whenever shrink_slab() is
> called and their low memory condition is satisfied.
> Before this, total_scan also prevent us to go into lowmemorykiller, so
> killing innocent process is limited as much as possible.
> 
shrinking is part of the normal operation of the Linux kernel and
happens all the time. Not only the call to shrink_slab, but actual
shrinking of unused objects.

I don't know therefore about any code that would kill process only
because they have reached shrink_slab.

In normal systems, this loop will be executed many, many times. So we're
not shrinking *more*, we're just guaranteeing that at least one pass
will be made.

Also, anyone looking at this to see if we should kill processes, is a
lot more likely to kill something if we tried to shrink but didn't, than
if we successfully shrunk something.



More information about the Containers mailing list