[PATCH v2 1/3] scsi: ufs: core: Introduce mcq ops to config cqid

Bart Van Assche bvanassche at acm.org
Thu Jun 1 13:54:38 PDT 2023


On 6/1/23 07:41, Peter Wang (王信友) wrote:
> On Thu, 2023-06-01 at 07:23 -0700, Bart Van Assche wrote:
>> Thanks, I had overlooked this. Do you agree that the above shows that the
>> flag I proposed in my previous email (UFSHCD_CAP_SINGLE_CQ) is sufficient
>> to support the MediaTek use case? I want to keep the SQ-CQ association code
>> in the UFS driver core because the next step will probably to switch from
>> one CQ per SQ to one CQ per CPU core for UFS controllers that support
>> multiple completion interrupts.
>
> If the UFS device speed is geting higher and higher, one CQ may not sufficient.
> 
> So, UFSHCD_CAP_SINGLE_CQ is not flexible for us beacuse we may want to map to two CQs.

Hi Peter,

Let's take a step back. The MediaTek UFSHCI 4.0 host controller only 
supports a single completion interrupt. A significant disadvantage of 
the single completion interrupt approach is that all completion 
interrupts are processed by the same CPU core. This is known to cause 
problems on Android. If sufficient time is spent in an interrupt 
handler, threads that run on the same CPU core as the interrupt handler 
may get scheduled too late. This can result in e.g. audio glitches 
noticeable by humans. Hardware designers told me that the area occupied 
by a single interrupt line is small. So I think it is fair to say that 
the (nonstandard!) approach of only supporting a single completion 
interrupt in an UFSHCI 4.0 controller is not a good choice.

The UFS driver already supports multiple hardware queue types 
(HCTX_TYPE_DEFAULT, READ and POLL). An interesting optimization would be 
to combine the completion queues for at least the DEFAULT and READ queue 
types. Introducing a vop to configure the completion queue ID would make 
it almost impossible to implement this optimization in a generic way.

Asking to add a vop that improves performance by only a few percent for 
a *nonstandard* controller and at the same time that makes it very hard 
to optimize the driver for standards compliant controllers is something 
that I consider unreasonable.

Bart.



More information about the Linux-mediatek mailing list