Received-SPF: pass (google.com: domain of linux-wireless+bounces-2495-linux.lists.archive=gmail.com@vger.kernel.org designates 139.178.88.99 as permitted sender) client-ip=139.178.88.99;
Content-Type: text/plain; charset="utf-8"
Precedence: bulk
MIME-Version: 1.0
Content-Transfer-Encoding: 7bit
Subject: Re: [PATCH] wifi: ath11k: enable 36 bit mask for stream DMA
From: Kalle Valo <kvalo@kernel.org>
In-Reply-To: <20240123015201.28939-1-quic_bqiang@quicinc.com>
References: <20240123015201.28939-1-quic_bqiang@quicinc.com>
To: Baochen Qiang <quic_bqiang@quicinc.com>
Cc: <ath11k@lists.infradead.org>, <linux-wireless@vger.kernel.org>,
 <quic_bqiang@quicinc.com>
User-Agent: pwcli/0.1.1-git (https://github.com/kvalo/pwcli/) Python/3.11.2
Message-ID: <170620068905.500647.8187957964143425904.kvalo@kernel.org>
Date: Thu, 25 Jan 2024 16:38:11 +0000 (UTC)

Baochen Qiang <quic_bqiang@quicinc.com> wrote:

> Currently 32 bit DMA mask is used, telling kernel to get us an DMA
> address under 4GB when mapping a buffer. This results in a very high
> CPU overhead in the case where IOMMU is disabled and more than 4GB
> system memory is installed. The reason is, with more than 4GB memory
> installed, kernel is likely to allocate a buffer whose physical
> address is above 4GB. While with IOMMU disabled, kernel has to involve
> SWIOTLB to map/unmap that buffer, which consumes lots of CPU cycles.
> 
> We did hit an issue caused by the reason mentioned above: in a system
> that disables IOMMU and gets 8GB memory installed, a total of 40.5%
> CPU usage is observed in throughput test. CPU profiling shows nearly
> 60% of CPU cycles are consumed by SWIOTLB.
> 
> By enabling 36 bit DMA mask, we can bypass SWIOTLB for any buffer
> whose physical address is below 64GB. There are two types of DMA mask
> within struct device, named dma_mask and coherent_dma_mask. Here we
> only enable 36 bit for dma_mask, because firmware crashes if
> coherent_dma_mask is also enabled, due to some unknown hardware
> limitations. This is acceptable because coherent_dma_mask is used for
> mapping a consistent DMA buffer, which generally does not happen in
> a hot path.
> 
> With this change, the total CPU usage mentioned in above issue drops
> to 18.9%.
> 
> Tested-on: WCN6855 hw2.1 PCI WLAN.HSP.1.1-03125-QCAHSPSWPL_V1_V2_SILICONZ_LITE-3.6510.23
> Tested-on: QCN9074 hw1.0 PCI WLAN.HK.2.7.0.1-01744-QCAHKSWPL_SILICONZ-1
> Tested-on: IPQ8074 hw2.0 AHB WLAN.HK.2.7.0.1-01744-QCAHKSWPL_SILICONZ-1
> 
> Signed-off-by: Baochen Qiang <quic_bqiang@quicinc.com>
> Acked-by: Jeff Johnson <quic_jjohnson@quicinc.com>
> Signed-off-by: Kalle Valo <quic_kvalo@quicinc.com>

Patch applied to ath-next branch of ath.git, thanks.

dbd73acb22d8 wifi: ath11k: enable 36 bit mask for stream DMA

-- 
https://patchwork.kernel.org/project/linux-wireless/patch/20240123015201.28939-1-quic_bqiang@quicinc.com/

https://wireless.wiki.kernel.org/en/developers/documentation/submittingpatches