Re: [bug report] iommu_dma_unmap_sg() is very slow then running IO from remote numa node

From: Robin Murphy <robin.murphy@arm.com>
To: Ming Lei <ming.lei@redhat.com>, John Garry <john.garry@huawei.com>
Cc: linux-kernel@vger.kernel.org, linux-nvme@lists.infradead.org,
	iommu@lists.linux-foundation.org, Will Deacon <will@kernel.org>,
	linux-arm-kernel@lists.infradead.org
Subject: Re: [bug report] iommu_dma_unmap_sg() is very slow then running IO from remote numa node
Date: Wed, 28 Jul 2021 16:39:23 +0100	[thread overview]
Message-ID: <e1ec45e5-8e8b-7295-4a95-af6fe92573ee@arm.com> (raw)
In-Reply-To: <YQF1AKS6Y14dLU/A@T590>

On 2021-07-28 16:17, Ming Lei wrote:
> On Wed, Jul 28, 2021 at 11:38:18AM +0100, John Garry wrote:
>> On 28/07/2021 02:32, Ming Lei wrote:
>>> On Mon, Jul 26, 2021 at 3:51 PM John Garry<john.garry@huawei.com>  wrote:
>>>> On 23/07/2021 11:21, Ming Lei wrote:
>>>>>> Thanks, I was also going to suggest the latter, since it's what
>>>>>> arm_smmu_cmdq_issue_cmdlist() does with IRQs masked that should be most
>>>>>> indicative of where the slowness most likely stems from.
>>>>> The improvement from 'iommu.strict=0' is very small:
>>>>>
>>>> Have you tried turning off the IOMMU to ensure that this is really just
>>>> an IOMMU problem?
>>>>
>>>> You can try setting CONFIG_ARM_SMMU_V3=n in the defconfig or passing
>>>> cmdline param iommu.passthrough=1 to bypass the the SMMU (equivalent to
>>>> disabling for kernel drivers).
>>> Bypassing SMMU via iommu.passthrough=1 basically doesn't make a difference
>>> on this issue.
>>
>> A ~90% throughput drop still seems to me to be too high to be a software
>> issue. More so since I don't see similar on my system. And that throughput
>> drop does not lead to a total CPU usage drop, from the fio log.

Indeed, it now sounds like $SUBJECT has been a complete red herring, and 
although the SMMU may be reflecting the underlying slowness it is not in 
fact a significant contributor to it. Presumably perf shows any 
difference in CPU time moving elsewhere once iommu_dma_unmap_sg() is out 
of the picture?

>> Do you know if anyone has run memory benchmark tests on this board to find
>> out NUMA effect? I think lmbench or stream could be used for this.
> 
> https://lore.kernel.org/lkml/YOhbc5C47IzC893B@T590/

Hmm, a ~4x discrepancy in CPU<->memory bandwidth is pretty significant, 
but it's still not the ~10x discrepancy in NVMe throughput. Possibly 
CPU<->PCIe and/or PCIe<->memory bandwidth is even further impacted 
between sockets, or perhaps all the individual latencies just add up - 
that level of detailed performance analysis is beyond my expertise. 
Either way I guess it's probably time to take it up with the system 
vendor to see if there's anything which can be tuned in hardware/firmware.

Robin.