On Mon, Sep 30, 2019 at 5:14 PM Eric Dumazet <eric.duma...@gmail.com> wrote: > > On 9/30/19 4:58 PM, John Ousterhout wrote: > > As of 4.16.10, it appears to me that sk->sk_backlog_len does not > > provide an accurate estimate of backlog length; this reduces the > > usefulness of the "limit" argument to sk_add_backlog. > > > > The problem is that, under heavy load, sk->sk_backlog_len can grow > > arbitrarily large, even though the actual amount of data in the > > backlog is small. This happens because __release_sock doesn't reset > > the backlog length until it gets completely caught up. Under heavy > > load, new packets can be arriving continuously into the backlog > > (which increases sk_backlog.len) while other packets are being > > serviced. This can go on forever, so sk_backlog.len never gets reset > > and it can become arbitrarily large. > > Certainly not. > > It can not grow arbitrarily large, unless a backport gone wrong maybe.
Can you help me understand what would limit the growth of this value? Suppose that new packets are arriving as quickly as they are processed. Every time __release_sock calls sk_backlog_rcv, a new packet arrives during the call, which is added to the backlog, incrementing sk_backlog.len. However, sk_backlog_len doesn't get decreased when sk_backlog_rcv completes, since the backlog hasn't emptied (as you said, it's not "safe"). As a result, sk_backlog.len has increased, but the actual backlog length is unchanged (one packet was added, one was removed). Why can't this process repeat indefinitely, until eventually sk_backlog.len reaches whatever limit the transport specifies when it invokes sk_add_backlog? At this point packets will be dropped by the transport even though the backlog isn't actually very large. > > Because of this, the "limit" argument to sk_add_backlog may not be > > useful, since it could result in packets being discarded even though > > the backlog is not very large. > > > > > You will have to study git log/history for the details, the limit _is_ useful, > and we reset the limit in __release_sock() only when _safe_. > > Assuming you talk about TCP, then I suggest you use a more recent kernel. > > linux-5.0 got coalescing in the backlog queue, which helped quite a bit.
