On Tue, Oct 1, 2019 at 1:53 PM Eric Dumazet [off-list ref] wrote:

On 10/1/19 1:45 PM, John Ousterhout wrote:

quoted

But this isn't really about socket resource limits (though that is
conflated in the implementation); it's about limiting the time spent
in a single call to __release_sock, no?

The proxy used is memory usage, not time usage.

I apologize for being pedantic, but the proxy isn't memory usage; it's
actually "number of bytes added to the backlog since the last time it
was emptied". At the time the limit is hit, actual memory usage is
probably a lot less than the limit. This was the source of my
confusion, since I assumed you really *wanted* memory usage to be the
limit.

cond_resched() or a preemptible kernel makes anything based on time flaky,
you probably do not want to play with a time limit...