Date:   Tue, 7 Jul 2020 18:35:56 +0100
From:   Robin Murphy <robin.murphy@....com>
To:     Jonathan Lemon <jonathan.lemon@...il.com>
Cc:     netdev@...r.kernel.org, iommu@...ts.linux-foundation.org,
        Björn Töpel <bjorn.topel@...el.com>,
        Christoph Hellwig <hch@....de>
Subject: Re: the XSK buffer pool needs be to reverted

On 2020-07-06 20:59, Jonathan Lemon wrote:
> On Wed, Jul 01, 2020 at 10:46:40AM +0100, Robin Murphy wrote:
>> On 2020-06-30 20:08, Jonathan Lemon wrote:
>>> On Mon, Jun 29, 2020 at 02:15:16PM +0100, Robin Murphy wrote:
>>>> On 2020-06-27 08:02, Christoph Hellwig wrote:
>>>>> On Fri, Jun 26, 2020 at 01:54:12PM -0700, Jonathan Lemon wrote:
>>>>>> On Fri, Jun 26, 2020 at 09:47:25AM +0200, Christoph Hellwig wrote:
>>>>>>>
>>>>>>> Note that this is somewhat urgent, as various of the APIs that the code
>>>>>>> is abusing are slated to go away for Linux 5.9, so this addition comes
>>>>>>> at a really bad time.
>>>>>>
>>>>>> Could you elaborate on what is upcoming here?
>>>>>
>>>>> Moving all these calls out of line, and adding a bypass flag to avoid
>>>>> the indirect function call for IOMMUs in direct mapped mode.
>>>>>
>>>>>> Also, on a semi-related note, are there limitations on how many pages
>>>>>> can be left mapped by the iommu?  Some of the page pool work involves
>>>>>> leaving the pages mapped instead of constantly mapping/unmapping them.
>>>>>
>>>>> There are, but I think for all modern IOMMUs they are so big that they
>>>>> don't matter.  Maintaines of the individual IOMMU drivers might know
>>>>> more.
>>>>
>>>> Right - I don't know too much about older and more esoteric stuff like POWER
>>>> TCE, but for modern pagetable-based stuff like Intel VT-d, AMD-Vi, and Arm
>>>> SMMU, the only "limits" are such that legitimate DMA API use should never
>>>> get anywhere near them (you'd run out of RAM for actual buffers long
>>>> beforehand). The most vaguely-realistic concern might be a pathological
>>>> system topology where some old 32-bit PCI device doesn't have ACS isolation
>>>> from your high-performance NIC such that they have to share an address
>>>> space, where the NIC might happen to steal all the low addresses and prevent
>>>> the soundcard or whatever from being able to map a usable buffer.
>>>>
>>>> With an IOMMU, you typically really *want* to keep a full working set's
>>>> worth of pages mapped, since dma_map/unmap are expensive while dma_sync is
>>>> somewhere between relatively cheap and free. With no IOMMU it makes no real
>>>> difference from the DMA API perspective since map/unmap are effectively no
>>>> more than the equivalent sync operations anyway (I'm assuming we're not
>>>> talking about the kind of constrained hardware that might need SWIOTLB).
>>>>
>>>>>> On a heavily loaded box with iommu enabled, it seems that quite often
>>>>>> there is contention on the iova_lock.  Are there known issues in this
>>>>>> area?
>>>>>
>>>>> I'll have to defer to the IOMMU maintainers, and for that you'll need
>>>>> to say what code you are using.  Current mainlaine doesn't even have
>>>>> an iova_lock anywhere.
>>>>
>>>> Again I can't speak for non-mainstream stuff outside drivers/iommu, but it's
>>>> been over 4 years now since merging the initial scalability work for the
>>>> generic IOVA allocator there that focused on minimising lock contention, and
>>>> it's had considerable evaluation and tweaking since. But if we can achieve
>>>> the goal of efficiently recycling mapped buffers then we shouldn't need to
>>>> go anywhere near IOVA allocation either way except when expanding the pool.
>>>
>>>
>>> I'm running a set of patches which uses the page pool to try and keep
>>> all the RX buffers mapped as the skb goes up the stack, returning the
>>> pages to the pool when the skb is freed.
>>>
>>> On a dual-socket 12-core Intel machine (48 processors), and 256G of
>>> memory, when iommu is enabled, I see the following from 'perf top -U',
>>> as the hottest function being run:
>>>
>>> -   43.42%  worker      [k] queued_spin_lock_slowpath
>>>      - 43.42% queued_spin_lock_slowpath
>>>         - 41.69% _raw_spin_lock_irqsave
>>>            + 41.39% alloc_iova
>>>            + 0.28% iova_magazine_free_pfns
>>>         + 1.07% lock_sock_nested
>>>
>>> Which likely is heavy contention on the iovad->iova_rbtree_lock.
>>> (This is on a 5.6 based system, BTW).  More scripts and data are below.
>>> Is there a way to reduce the contention here?
>>
>> Hmm, how big are your DMA mappings? If you're still hitting the rbtree a
>> lot, that most likely implies that either you're making giant IOVA
>> allocations that are too big to be cached, or you're allocating/freeing
>> IOVAs in a pathological pattern that defeats the whole magazine cache
>> mechanism (It's optimised for relatively-balanced allocation and freeing of
>> sizes up order 6). On a further hunch, does the "intel_iommu=forcedac"
>> option make any difference at all?
> 
> The allocations are only 4K in size (packet memory) but there are a lot
> of them.  I tried running with "forcedac" over the weekend, and that
> seems to have made a huge difference.  Why the name of 'forcedac'?  It
> would seem this should be the default for a system with a sane 64-bit PCI
> device.

OK, I think I can imagine what's happening there - I could elaborate if 
you're particularly interested, but suffice to say that despite all the 
tricks we've put in to avoid spending unnecessary time down that path 
wherever possible, there are still ways the allocator could behave 
pathologically once the low end below 32 bits gets (nearly) full. As far 
as that command-line option goes, I just happened to notice that it 
already existed and has the same effect as the hack I was going to 
suggest ;)

Either way, though, I'm inclined to agree that it might be time to stop 
applying that particular "optimisation" indiscriminately - I'll cook up 
some patches.

>> Either way if this persists after some initial warm-up period, it further
>> implies that the page pool is not doing its job properly (or at least in the
>> way I would have expected). The alloc_iova() call is part of the dma_map_*()
>> overhead, and if the aim is to keep pages mapped then that should only be
>> called relatively infrequently. The optimal behaviour would be to dma_map()
>> new clean pages as they are added to the pool, use dma_sync() when they are
>> claimed and returned by the driver, and only dma_unmap() if they're actually
>> freed back to the page allocator. And if you're still seeing a lot of
>> dma_map/unmap time after that, then the pool itself is churning pages and
>> clearly needs its size/thresholds tuning.
> 
> Most of the dma_map() overhead is coming from the TX path, which isn't
> using the page_pool.

Ah, OK, I guess I assumed the TX path might copy into preallocated DMA 
buffers such that it would use a pool too. Fair enough if it doesn't.
(Although perhaps it's still worth noting that IOMMUs are liable to skew 
the threshold for zero-copy benefits in general)

Robin.

> With the RX datapath I'm using, pages are returned to the pool in a
> non-napi context (so they enter the ring), then the driver refills the
> cache from the ring.  I'm seeing a 1% overflow (ring_full), so the
> default size seems appropriate.  The single ring lock is likely a
> problem, but can be addressed now that things aren't blowing up in
> alloc_iova().
> 
>  From a box with the mlx4 driver + persistent page pool:
> 
>       rx_alloc_pages: 35239524767
>       cache_hit: 34381377143
>       cache_full: 19935
>       cache_empty: 859266494
>       ring_consume: 31665993135
>       ring_produce: 31666119491
>       ring_full: 353805844
> 

Powered by blists - more mailing lists