Re: [PATCH] brcmfmac: Set SDIO workqueue as WQ_HIGHPRI
From: Sean Anderson <hidden>
Date: 2021-08-17 18:22:05
On 8/17/21 1:17 PM, Arend van Spriel wrote:
On August 17, 2021 6:50:50 PM Sean Anderson [off-list ref] wrote:quoted
ping?Good idea to ping with a top-level post :-p
Sorry, I'm not subscribed to this list; did I mess up the list/CC for the original?
quoted
On 8/2/21 1:09 PM, Sean Anderson wrote:quoted
This puts tasks submitted to the SDIO workqueue at the head of the queue and runs them immediately. This gets higher RX throughput with the SDIO bus. This was originally submitted as [1]. The original author Wright Feng reportsquoted
throughput result with 43455(11ac) on 1 core 1.6 Ghz platform is Without WQ_HIGGPRI TX/RX: 293/301 (mbps) With WQ_HIGHPRI TX/RX: 293/321 (mbps)I tested this with a 43364(11bgn) on a 1 core 800 MHz platform and got Without WQ_HIGHPRI TX/RX: 16/19 (Mbits/sec) With WQ_HIGHPRI TX/RX: 24/20 (MBits/sec) [1] https://lore.kernel.org/linux-wireless/1584604406-15452-4-git-send-email-wright.feng@cypress.com/ (local)While I understand the obvious gain it seems like a wrong move to me. What if all workqueues in the kernel would start using this flag? I bet the gain above would be negated and all are equal in the eyes of .. the kernel
Is there an official policy on what counts as high-priority? Using some
very-scientific methodology [1], it seems like most high-priority
workqueues are in drivers/net and fs. Making these queues high-priority
seems to be commonplace. For example, in fe101716c7c9 ("rtw88: replace
tx tasklet with work queue"), Po-Hao Huang remarks:
Since throughput is delay-sensitive in most cases, we allocate a dedicated, high priority wq for our needs.
which is effectively the same rationale as this patch. At least for my application, network transfer speed is one of the most important performance metrics. The original patch got the following feedback [2] from Kalle Valo:
Why would someone want to disable this? Like in patch 2, please avoid adding new module parameters as much as possible.
so for this patch I just made the workqueue high-priority unconditionally. --Sean [1] git grep -l WQ_HIGHPRI | cut -f 1,2 -d / | uniq -c | sort -n [2] https://lore.kernel.org/linux-wireless/87tv2gbgv1.fsf@kamboji.qca.qualcomm.com/ (local)