Subject: Re: [virtio-dev] RFC: Doorbell suppression, packed-ring mode and hardware offload
This can result in a very high rate of doorbells with some drivers, which can become a severe bottleneck (because x86 CPUs can't emit MMIOs at very high rates).Interesting. Is there any data you could share to help guide the design? E.g. what's the highest rate of MMIO writes supported etc?On an E3-1270 v5 @ 3.60GHz, max rate across all cores is ~18M/s. On an E5-2637 v4 @ 3.50GHz, max rate on PCIe-local socket is ~14M/s.Â On PCIe-remote socket ~8M/s.Is that mega bytes? Or million writes? With 32 bit writes are we limited to 2M then?
Sorry, this is million-writes-per-second.Â The size of the write doesn't matter.
This doesn't just impose a limit on aggregate packet rate: If you hit this bottleneck then the CPU core is back-pressured by the MMIO writes, and so instr/cycle takes a huge hit.The proposed offset/wrap field allows devices to disable doorbells when appropriate, and determine the latest fill level via a PCIe read.This kind of looks like a split ring though, does it not?I don't agree, because the ring isn't split.Â Split-ring is very painful for hardware because of the potentially random accesses to the descriptor table (including pointer chasing) which inhibit batching of descriptor fetches, as well as causing complexity in implementation.That's different with IN_ORDER, right?
Yes, IN_ORDER will give some of the benefit of packed ring, and is also a win because you can merge 'used' writes.Â (But packed-ring still wins due to not having to fetch ring and descriptor table separately).
Packed ring is a huge improvement on both dimensions, but introduces a specific pain point for hardware offload.The issue is we will again need to bounce more cache lines to communicate.You'll only bounce cache lines if the device chooses to read the idx.Â A PV device need not offer this feature.Â A PCIe device will, but the cost to the driver of writing an in-memory idx is much smaller than the cost of an MMIO, which is always a strongly ordered barrier on x86/64. With vDPA you ideally would have this feature enabled, and the device would sometimes be PV and sometimes PCIe.Â The PV device would ignore the new idx field and so cache lines would not bounce then either. Ie. The only time cache lines are shared is when sharing with a PCIe device, which is the scenario when this is a win.True. OTOH on non-x86 you will need more write barriers :( It would be natural to say driver can reuse the barrier before flags update, but note that that would mean hardware still needs to support re-fetching if flags value is wrong. Such re-fetch is probably rare so fine by me, but it does add a bit more complexity.
I would prefer to have the write barrier before writing the idx.Note that drivers could take advantage of this feature to avoid read barriers when consuming descriptors: At the moment there is a virtio_rmb() per descriptor read.Â With the proposed feature the driver can do just one barrier after reading idx.Â I expect that on platforms where write barriers have a cost read barriers likely have a significant cost too, so this might be a win with PV devices too.
So I wonder: what if we made a change to spec that would allow prefetch of ring entries? E.g. you would be able to read at random and if you guessed right then you can just use what you have read, no need to re-fetch?Unless I've misunderstood I think this would imply that the driver would have to ensure strict ordering for each descriptor it wrote, which would impose a cost to the driver.Â At the moment a write barrier is only needed before writing the flags of the first descriptor in a batch.On non-x86 right? OTOH the extra data structure also adds more ordering requirements.
Yes on non-x86.Â The extra data structure only adds an ordering once per request (when enabled) whereas allowing prefetch would add an ordering per descriptor.Â The number of descriptors is always >= the number of requests, and can be much larger (esp. for a net rx virt-queue).
I suggest the best place to put this would be in the driver area, immediately after the event suppression structure.Could you comment on why is that a good place though?The new field is written by the driver, as are the other fields in the driver area.Â Also I expect that devices might be able to read this new idx together with the interrupt suppression fields in a single read, reducing PCIe overheads.OK then you need it in the same aligned 256 byte boundary.
The event suppression structs currently require 4byte align.Â Are you suggesting we increase the align required when VIRTIO_F_RING_PACKED_AVAIL_IDX is negotiated?Â Sounds good to me, but 8bytes would presumably suffice.
Thanks for the feedback. David -- David Riddoch <email@example.com> -- Chief Architect, Solarflare