Re: [PATCH v1 2/3] scsi: ufs: Optimize host lock on transfer requests send/compl paths

[Can Guo <cang@codeaurora.org>]

On 2021-05-25 04:10, Bart Van Assche wrote:
> On 5/24/21 1:36 AM, Can Guo wrote:
>> Current UFS IRQ handler is completely wrapped by host lock, and 
>> because
>> ufshcd_send_command() is also protected by host lock, when IRQ handler
>> fires, not only the CPU running the IRQ handler cannot send new 
>> requests,
>> the rest CPUs can neither. Move the host lock wrapping the IRQ handler 
>> into
>> specific branches, i.e., ufshcd_uic_cmd_compl(), 
>> ufshcd_check_errors(),
>> ufshcd_tmc_handler() and ufshcd_transfer_req_compl(). Meanwhile, to 
>> further
>> reduce occpuation of host lock in ufshcd_transfer_req_compl(), host 
>> lock is
>> no longer required to call __ufshcd_transfer_req_compl(). As per test, 
>> the
>> optimization can bring considerable gain to random read/write 
>> performance.
> 
> Hi Can,
> 
> Using the host lock to serialize the completion path against the
> submission path was a common practice 11 years ago, before the host 
> lock
> push-down (see also
> https://linux-scsi.vger.kernel.narkive.com/UEmGgwAc/rfc-patch-scsi-host-lock-push-down).
> Modern SCSI LLDs should not use the SCSI host lock. Please consider
> introducing one or more new synchronization objects in struct ufs_hba
> and to use these instead of the SCSI host lock. That will save multiple
> pointer dereferences in the hot path since hba->host->host_lock will
> become hba->new_spin_lock.
> 
> An additional question is whether it is necessary for v3.0 UFS devices
> to serialize the submission path against the completion path? Multiple
> high-performance SCSI LLDs support hardware with separate submission 
> and
> completion queues and hence do not need any serialization between the
> submission and the completion path. I'm asking this because it is 
> likely
> that sooner or later multiqueue support will be added in the UFS
> specification. Benefiting from multiqueue support will require to 
> rework
> locking in the UFS driver anyway.
> 

Hi Bart,

Agree with all above, and what you ask is right what we are doing in the
3rd change - get rid of host lock on dispatch and completion paths.

I agree with using dedicated spin locks for dedicated purposes in UFS 
driver,
e.g., clk gating has its own gating_lock and clk scaling has its own 
scaling_lock.
But this specific series is only for improving performance. We will take 
your
comments into consideration and address it in future.

Thanks,

Can Guo.

> Thanks,
> 
> Bart.