Date:   Mon, 20 Feb 2017 15:40:25 -0800
From:   Alexander Duyck <alexander.duyck@...il.com>
To:     Saeed Mahameed <saeedm@...lanox.com>
Cc:     Jesper Dangaard Brouer <brouer@...hat.com>,
        Alexei Starovoitov <alexei.starovoitov@...il.com>,
        John Fastabend <john.fastabend@...il.com>,
        David Miller <davem@...emloft.net>,
        Tom Herbert <tom@...bertland.com>,
        "netdev@...r.kernel.org" <netdev@...r.kernel.org>,
        Brenden Blanco <bblanco@...il.com>
Subject: Re: Focusing the XDP project

On Mon, Feb 20, 2017 at 2:57 PM, Saeed Mahameed <saeedm@...lanox.com> wrote:
>
>
> On 02/20/2017 10:09 PM, Alexander Duyck wrote:
>> On Mon, Feb 20, 2017 at 2:13 AM, Jesper Dangaard Brouer
>> <brouer@...hat.com> wrote:
>>>
>>> First thing to bring in order for the XDP project:
>>>
>>>   RX batching is missing.
>>>
>>> I don't want to discuss packet page-sizes or multi-port forwarding,
>>> before we have established the most fundamental principal that all
>>> other solution use; RX batching.
>>
>> That is all well and good, but some of us would like to discuss other
>> items as it has a direct impact on our driver implementation and
>> future driver design.  Rx batching really seems tangential to the
>> whole XDP discussion anyway unless you are talking about rewriting the
>> core BPF code and kernel API itself to process multiple frames at a
>> time.
>>
>> That said, if something seems like it would break the concept you have
>> for Rx batching please bring it up.  What I would like to see is well
>> defined APIs and a usable interface so that I can present XDP to
>> management and they will see the use of it and be willing to let me
>> dedicate developer heads to enabling it on our drivers.
>>
>>> Without building in RX batching, from the beginning/now, the XDP
>>> architecture have lost.  As adding features and capabilities, will
>>> just lead us back to the exact same performance problems as before!
>>
>> I would argue you have much bigger issues to deal with.  Here is a short list:
>> 1.  The Tx code is mostly just a toy.  We need support for more
>> functional use cases.
>> 2.  1 page per packet is costly, and blocks use on the intel drivers,
>> mlx4 (after Eric's patches), and 64K page architectures.
>> 3.  Should we support scatter-gather to support 9K jumbo frames
>> instead of allocating order 2 pages?
>>
>> Focusing on Rx batching seems like bike shedding more than anything
>> else.  I would much rather be focused on what the API definitions
>> should be for the drivers and the BPF code rather than focus on the
>> inner workings of the drivers themselves.  Then at that point we can
>> start looking at expanding this out to other drivers and coming up
>> with good test cases to test the functionality.  We really need the
>> interfaces clearly defines so that we can then look at having those
>> pulled into the distros so we have some sort of ABI we can work with
>> in customer environments.
>>
>> Dropping frames is all well and good, but only so useful.  With the
>> addition of DMA_ATTR_SKIP_CPU_SYNC we should be able to do writable
>> pages so we could now do encap/decap type workloads.  If we can add
>> support for routing pages between interfaces that gets us close to
>> being able to OVS style demos.  At that point we can then start
>> comparing ourselves to DPDK and FD.io and seeing what we can do to
>> improve performance.
>>
>
> Well, although I think Jesper is a little bit exaggerating ;) I guess he has a point
> and i am on his side on this discussion. you see, if we define the APIs and ABIs now
> and they turn out to be a bottleneck for the whole XDP arch performance, at that
> point it will be too late to compare XDP to DPDK and other kernel bypass solutions.

Yes, but at the same time we cannot hold due to decision paralysis.
We should be moving forward, not holding waiting on things that may or
may not get done.

> What we need to do is to bring XDP to a state where it performs at least the same as other
> kernel bypass solutions. I know that the DPDK team here at mellanox spent years working
> on DPDK performance, squeezing every bit out of the code/dcache/icache/cpu you name it..
> We simply need to do the same for XDP to prove it worthy and can deliver the required
> rates. Only then, when we have the performance baseline numbers, we can start expanding XDP features
> and defining new use cases and a uniform API, while making sure the performance is kept at it max.

The problem is performance without features is useless.  I can make a
driver that received and drops all packets that goes really fast, but
it isn't too terribly useful and nobody will use it.  I don't want us
locking in on one use case and spending all of our time optimizing for
that when there is a good chance that nobody cares.  For example the
FIB argument Jesper was making is likely completely useless to most
people who will want to use XDP.  While there are some that may want a
router implemented in XDP it is much more likely that they will want
to do VM to VM switching via something more like OVS.

My argument is that we need to figure out what features we need, then
we can focus on performance.  I would much rather deliver a feature
and then improve the performance, than show the performance and not be
able to meet that after adding a feature.  It is all a matter of
setting expectations.

> Yes, there is a down side to this, that currently most of the optimizations and implementations we can do
> are inside the device driver and they are driver dependent, but once we have a clear image
> on how things should work, we can pause and think on how to generalize the approaches
> to all device drivers.

I'm fine with the optimizations being in the device driver, however
feature implementations are another matter.  Historically once
something is in a driver it takes a long time if ever for it to be
generalized out of the driver.  More often than not the driver vendors
prefer to leave their code as-is for a competitive advantage.
Historically the way we deal with this as a community is that if an
interface is likely to be used by more than one device it has to start
out generalized.

>>> Today we already have the 64 packets NAPI budget, but we are not
>>> taking advantage of this. For XDP as long as eBPF always return
>>> XDP_DROP or XDP_TX, then we (falsely) experience the effect of bulking
>>> (as code fits within the icache) and see huge perf boosts.
>>
>> This makes a lot of assumptions.  First, the budget is up to 64, it
>> isn't always 64.  Second, you say we are "falsely" seeing icache
>> improvements, and I would argue that it isn't false as we are
>> intentionally bypassing most of the stack to perform the drop early.
>> That was kind of the point of all this.  Finally, this completely
>> discounts GRO/LRO which would take care of aggregating the frames and
>> reducing much of this overhead for TCP flows being received over the
>> interface.
>>
>>> The initial principal of bulking/batching packets to amortize per
>>> packet costs.  The next step is just as important: Lookup table sizes
>>> (FIB) kills performance again. The solution is implementing a smart
>>> table lookup scheme that prefetch hash table key-cells and afterwards
>>> prefetch data-cells, based on the RX batch of packets.  Notice VPP
>>> revolves around similar tricks, and why it beats DPDK, and why it
>>> scales with 1Millon routes.
>>
>> This is where things go completely sideways in your argument.  If you
>> want to implement some sort of custom FIB lookup library for XDP be my
>> guest.  If you are talking about hacking on the kernel I would
>> question how this is even related to XDP?  The lookup that is in the
>> kernel is designed to provide the best possible lookup under a number
>> of different conditions.  It is a "jack of all trades, master of none"
>> type of implementation.
>>
>> Also, why should we be focused on FIB?  Seems like this is getting
>> back into routing territory and what I am looking for is uses well
>> beyond just routing.
>>
>>> I hope I've made it very clear where the focus for XDP should be.
>>> This involves implementing what I call RX-stages in the drivers. While
>>> doing that we can figure out the most optimal data structure for
>>> packet batching.
>>
>> Yes Jesper, your point of view is clear.  This is the same agenda you
>> have been pushing for the last several years.  I just don't see how
>> this can be made a priority now for a project where it isn't even
>> necessarily related.  In order for any of this to work the stack needs
>> support for bulk Rx, and we still seem pretty far from that happening.
>>
>>>  I know Saeed is already working on RX-stages for mlx5, and I've tested
>>> the initial version of his patch, and the results are excellent.
>>
>> That is great!  I look forward to seeing it when they push it to net-next.
>>
>> By the way, after looking over the mlx5 driver it seems like there is
>> a bug in the logic.  From what I can tell it is using build_skb to
>> build frames around the page, but it doesn't bother to take care of
>> handling the mappings correctly.  So mlx5 can end up with data
>> corruption when the pages are unmapped.  My advice would be to look at
>> updating the driver to do something like what I did in ixgbe to make
>> use of the DMA_ATTR_SKIP_CPU_SYNC DMA attribute so that it won't
>> invalidate any updates made when adding headers or shared info.
>>
>
> hmmm, are you talking about the mlx5 rx page cache ? will take a look at the ixgbe code for sure
> but we didn't experience any issue of the sort, can you shed more light on the issue ?
>
> Thanks,
> -Saeed.

Basically the issue is that there are some architectures where
dma_unmap_page will invalidate the page and cause any data written to
it from the CPU side to be invalidated.  On x86 the only way to
recreate this is to use the kernel parameter "swiotlb=force".
Basically when a page was mapped you couldn't unmap it without running
the risk of invalidating any data you had written to it.  I added a
DMA attribute called DMA_ATTR_SKIP_CPU_SYNC which is meant to prevent
that from taking place on unmap.  It also ends up being a performance
gain on architectures that do this since it avoids looping through
cache lines invalidating them on unmap.

Hope that helps.

- Alex

Powered by blists - more mailing lists