Re: __GFP_REPEAT usage in fq_alloc_node

[Eric Dumazet]

On Mon, Jan 9, 2017 at 2:22 AM, Michal Hocko <mhocko@xxxxxxxxxx> wrote:
>
> the changelog doesn't mention it but this, unlike other kvmalloc
> conversions is not without functional changes. The kmalloc part
> will be weaker than it is with the original code for !costly (<64kB)
> requests, because we are enforcing __GFP_NORETRY to break out from the
> page allocator which doesn't really fail such a small requests.
>
> Now the question is what those code paths really prefer. Do they really
> want to potentially loop in the page allocator and invoke the OOM killer
> when the memory is short/fragmeted? I mean we can get into a situation
> when no order-3 pages can be compacted and shooting the system down just
> for that reason sounds quite dangerous to me.
>
> So the main question is how hard should we try before falling back to
> vmalloc here?

This patch is fine :

1) Default hash size is 1024 slots, 8192 bytes on 64bit arches.
2) Most of the times, qdisc are setup at boot time.