I have also noticed this type of problem. It seems as though the RAID5
driver generates a growing write backlog and keeps allocating new
buffers when new asynchronous write requests get in. Eventually it
reserves all the available physical memory. Trying to swap data to
virtual memory storage would only make the situation worse.
I'm not sure where the responsibility lies for this problem. The md
driver can limit how much it allocates, but the memory manager should be
able to handle this situation better.
Markus Linnala wrote:
>
> v2.2.0 heavy writing at raid5 array kills processes randomly, including init.
>
> Normal user can force random processes to out of memory
> situation when writing stuff at raid5 array. This makes the raid
>
> I get 'Out of memory for init. ' etc. with following simple command:
>
> dd if=/dev/zero of=file
>
> Repeatable, file is between 100-200M after dd gets killed.
> I guess this killing action seems to be triggered by swapping.
>
