Hi Matt,
Happy new year!
We also noticed this problem and proposed a change in
https://gerrit.fd.io/r/c/vpp/+/30386.
The change only uses single queue per thread to avoid this problem. We tried
with 1-4 crypto IDs per core and achieved
Can you have a look?
Regards,
Fan
From: vpp-dev@lists.fd.io <vpp-dev@lists.fd.io> On Behalf Of Matthew Smith via
lists.fd.io
Sent: Tuesday, December 22, 2020 7:31 PM
To: vpp-dev <vpp-dev@lists.fd.io>
Subject: [vpp-dev] question about crypto_sw_scheduler plugin
Hello,
I am looking into enabling VPP's asynchronous crypto for IPsec in Netgate's
control plane. I was looking at the crypto_sw_scheduler plugin and noticed that
crypto_sw_scheduler_dequeue_aead() and crypto_sw_scheduler_dequeue_link() each
have a loop like this:
vec_foreach_index (i, cm->per_thread_data)
{
ptd = cm->per_thread_data + i;
q = ptd->queues[async_op_id];
f = crypto_sw_scheduler_get_pending_frame (q);
if (f)
break;
}
It looks like this loop will iterate the per-thread data of each thread
sequentially starting at thread 0 and try to retrieve a pending frame from that
thread's queue and break the loop if one is retrieved. Will this end up causing
the operations which were queued by the threads with lower indices to be
handled more quickly than operations which were queued by the threads with
higher indices?
E.g. - assume 4 worker threads. The main thread has ID 0 and the workers have
IDs 1 - 4. If thread 1 has 5 frames queued for some crypto operation and thread
4 has 1 frame queued for the same operation, will all 5 frames from thread 1 be
handled before the 1 frame from thread 4 is? Or am I misunderstanding something?
Thanks!
-Matt
-=-=-=-=-=-=-=-=-=-=-=-
Links: You receive all messages sent to this group.
View/Reply Online (#18459): https://lists.fd.io/g/vpp-dev/message/18459
Mute This Topic: https://lists.fd.io/mt/79156990/21656
Group Owner: vpp-dev+ow...@lists.fd.io
Unsubscribe: https://lists.fd.io/g/vpp-dev/unsub [arch...@mail-archive.com]
-=-=-=-=-=-=-=-=-=-=-=-
