From: Tom Talpey > Sent: 09 April 2021 18:49 > On 4/9/2021 12:27 PM, Haakon Bugge wrote: > > > > > >> On 9 Apr 2021, at 17:32, Tom Talpey <tom@xxxxxxxxxx> wrote: > >> > >> On 4/9/2021 10:45 AM, Chuck Lever III wrote: > >>>> On Apr 9, 2021, at 10:26 AM, Tom Talpey <tom@xxxxxxxxxx> wrote: > >>>> > >>>> On 4/6/2021 7:49 AM, Jason Gunthorpe wrote: > >>>>> On Mon, Apr 05, 2021 at 11:42:31PM +0000, Chuck Lever III wrote: > >>>>> > >>>>>> We need to get a better idea what correctness testing has been done, > >>>>>> and whether positive correctness testing results can be replicated > >>>>>> on a variety of platforms. > >>>>> RO has been rolling out slowly on mlx5 over a few years and storage > >>>>> ULPs are the last to change. eg the mlx5 ethernet driver has had RO > >>>>> turned on for a long time, userspace HPC applications have been using > >>>>> it for a while now too. > >>>> > >>>> I'd love to see RO be used more, it was always something the RDMA > >>>> specs supported and carefully architected for. My only concern is > >>>> that it's difficult to get right, especially when the platforms > >>>> have been running strictly-ordered for so long. The ULPs need > >>>> testing, and a lot of it. > >>>> > >>>>> We know there are platforms with broken RO implementations (like > >>>>> Haswell) but the kernel is supposed to globally turn off RO on all > >>>>> those cases. I'd be a bit surprised if we discover any more from this > >>>>> series. > >>>>> On the other hand there are platforms that get huge speed ups from > >>>>> turning this on, AMD is one example, there are a bunch in the ARM > >>>>> world too. > >>>> > >>>> My belief is that the biggest risk is from situations where completions > >>>> are batched, and therefore polling is used to detect them without > >>>> interrupts (which explicitly). The RO pipeline will completely reorder > >>>> DMA writes, and consumers which infer ordering from memory contents may > >>>> break. This can even apply within the provider code, which may attempt > >>>> to poll WR and CQ structures, and be tripped up. > >>> You are referring specifically to RPC/RDMA depending on Receive > >>> completions to guarantee that previous RDMA Writes have been > >>> retired? Or is there a particular implementation practice in > >>> the Linux RPC/RDMA code that worries you? > >> > >> Nothing in the RPC/RDMA code, which is IMO correct. The worry, which > >> is hopefully unfounded, is that the RO pipeline might not have flushed > >> when a completion is posted *after* posting an interrupt. > >> > >> Something like this... > >> > >> RDMA Write arrives > >> PCIe RO Write for data > >> PCIe RO Write for data > >> ... > >> RDMA Write arrives > >> PCIe RO Write for data > >> ... > >> RDMA Send arrives > >> PCIe RO Write for receive data > >> PCIe RO Write for receive descriptor > > > > Do you mean the Write of the CQE? It has to be Strongly Ordered for a correct implementation. Then > it will shure prior written RO date has global visibility when the CQE can be observed. > > I wasn't aware that a strongly-ordered PCIe Write will ensure that > prior relaxed-ordered writes went first. If that's the case, I'm > fine with it - as long as the providers are correctly coded!! I remember trying to read the relevant section of the PCIe spec. (Possibly in a book that was trying to make it easier to understand!) It is about as clear as mud. I presume this is all about allowing PCIe targets (eg ethernet cards) to use relaxed ordering on write requests to host memory. And that such writes can be completed out of order? It isn't entirely clear that you aren't talking of letting the cpu do 'relaxed order' writes to PCIe targets! For a typical ethernet driver the receive interrupt just means 'go and look at the receive descriptor ring'. So there is an absolute requirement that the writes for data buffer complete before the write to the receive descriptor. There is no requirement for the interrupt (requested after the descriptor write) to have been seen by the cpu. Quite often the driver will find the 'receive complete' descriptor when processing frames from an earlier interrupt (and nothing to do in response to the interrupt itself). So the write to the receive descriptor would have to have RO clear to ensure that all the buffer writes complete first. (The furthest I've got into PCIe internals was fixing the bug in some vendor-supplied FPGA logic that failed to correctly handle multiple data TLP responses to a single read TLP. Fortunately it wasn't in the hard-IP bit.) David - Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK Registration No: 1397386 (Wales)
