In D61037#1399489, @kp wrote:

Should we try to account for a VLAN header here too?

Abandoned in favor of 9e7d55570f69.

• Report UDP packets hashed on the addresses as M_HASHTYPE_RSS_IPV4 / M_HASHTYPE_RSS_IPV6 instead of the UDP types, so that rss_mbuf_software_hash_v4() / _v6() trust the hardware hash rather than recomputing a 2-tuple hash for every packet.
• Add address-only SCTP profiles; ice never hashed SCTP.
• Treat the IPv6 _EX hash types as served by the IPv6 profiles. This also removes the "could not configure every requested hash type" warning printed on every attach.
• Summary updated to match.

• The hold is consulted only where the doorbell policy would ring for every packet, as its lowest band rather than a size test of its own.
• The end of a receive pass on the simple path uses the completion task's drain with the thread quota; the separate ring function is gone.
• The deregister epoch wait is dropped from this review; it fixes a pre-existing detach race and will come in a separate review.
• Man page entry shortened.

Two single-use helpers folded away: the hold looks up the batch itself, and batch_end rings the queues it releases.

This version is even slimmer. The transmit task now rings a held queue instead of being told not to, so the flush flag, the timer hook, the unheld check and its fences are gone. The cost is about 8 % more doorbells on mp_ring (0.063 to 0.069 per packet), throughput unchanged. The admin-task kick now says why it exists.

• The admin-task kick and the release fence now say which case they cover.
• iflib_txdb_unheld uses a seq_cst fence instead of acquire.
• The batch lookup is renamed iflib_txdb_current() so it no longer shares a name with the struct (better greppability).

In D60416#1397076, @kbowling wrote:

This looks cleaner but it still feels grafted on, and my earlier wonder remains: perhaps unnecessarily; doing what @gallatin said with the call or disproving that would be convincing to me that you've reasoned through the existing mechanism and found the minimal implementation.

In D60416#1393857, @gallatin wrote:

Did you try measure doing the call to iflib_txdb_batch() in iflib_simple_encap(), rather than adding an argument all the way down? I hate the perceived complexity adding an arg adds.. Pushing it down adds 2 loads from hot cache lines per packet in the case where we have contention and the deferral ring is used, at the human savings of some complexity, and possibly 8 bytes of stack space. I think it won't matter at all for your workload, since I'm guessing your workload is mostly uncontended, so the deferral ring is probably not used.

Also, did you try @kbowling 's suggestion of fixing the db deferral in simple_tx ? I think its just as follows, and I'd be interested to hear how much it helps compared to this complexity:

Rebased on main. Changes since the last diff:

D60756 was accepted by @kib.

In D60756#1393605, @kib wrote:

Since this is driver-specific, did you considered using osd(9) osd_register(OSD_THREAD)? [I do not object against adding a thread-local pointer, just want to check other possibilities before]

I have reworked this significantly.