Re: [PATCH v2 18/20] scsi: ufs: Optimize the command queueing code

["Asutosh Das (asd)" <asutoshd@codeaurora.org>]

On 11/22/2021 3:48 PM, Bart Van Assche wrote:
> On 11/22/21 3:02 PM, Asutosh Das (asd) wrote:
>> Current code waits for the already issued requests to complete. It 
>> doesn't issue the yet-to-be issued requests. Wouldn't freezing the 
>> queue issue the requests in the context of scaling_{up/down}?
>> If yes, I don't think the current code is doing that.
> 
> Hi Asutosh,
> 
> How about the patch below that preserves most of the existing code for
> preparing for clock scaling?
> 
> Thanks,
> 
> Bart.
> 
Hi Bart,
This looks good to me. Please push a change and I can test it out.

-asd

> 
> Subject: [PATCH] scsi: ufs: Optimize the command queueing code
> 
> Remove the clock scaling lock from ufshcd_queuecommand() since it is a
> performance bottleneck. Instead, use synchronize_rcu_expedited() to wait
> for ongoing ufshcd_queuecommand() calls.
> 
> Signed-off-by: Bart Van Assche <bvanassche@acm.org>
> ---
>   drivers/scsi/ufs/ufshcd.c | 12 +++++++-----
>   drivers/scsi/ufs/ufshcd.h |  1 +
>   2 files changed, 8 insertions(+), 5 deletions(-)
> 
> diff --git a/drivers/scsi/ufs/ufshcd.c b/drivers/scsi/ufs/ufshcd.c
> index 5d214456bf82..1d929c28efaf 100644
> --- a/drivers/scsi/ufs/ufshcd.c
> +++ b/drivers/scsi/ufs/ufshcd.c
> @@ -1196,6 +1196,13 @@ static int ufshcd_clock_scaling_prepare(struct 
> ufs_hba *hba)
>       /* let's not get into low power until clock scaling is completed */
>       ufshcd_hold(hba, false);
> 
> +    /*
> +     * Wait for ongoing ufshcd_queuecommand() calls. Calling
> +     * synchronize_rcu_expedited() instead of synchronize_rcu() reduces 
> the
> +     * waiting time from milliseconds to microseconds.
> +     */
> +    synchronize_rcu_expedited();
> +
>   out:
>       return ret;
>   }
> @@ -2699,9 +2706,6 @@ static int ufshcd_queuecommand(struct Scsi_Host 
> *host, struct scsi_cmnd *cmd)
> 
>       WARN_ONCE(tag < 0, "Invalid tag %d\n", tag);
> 
> -    if (!down_read_trylock(&hba->clk_scaling_lock))
> -        return SCSI_MLQUEUE_HOST_BUSY;
> -
>       /*
>        * Allows the UFS error handler to wait for prior 
> ufshcd_queuecommand()
>        * calls.
> @@ -2790,8 +2794,6 @@ static int ufshcd_queuecommand(struct Scsi_Host 
> *host, struct scsi_cmnd *cmd)
>   out:
>       rcu_read_unlock();
> 
> -    up_read(&hba->clk_scaling_lock);
> -
>       if (ufs_trigger_eh()) {
>           unsigned long flags;
> 
> diff --git a/drivers/scsi/ufs/ufshcd.h b/drivers/scsi/ufs/ufshcd.h
> index c13ae56fbff8..695bede14dac 100644
> --- a/drivers/scsi/ufs/ufshcd.h
> +++ b/drivers/scsi/ufs/ufshcd.h
> @@ -777,6 +777,7 @@ struct ufs_hba_monitor {
>    * @clk_list_head: UFS host controller clocks list node head
>    * @pwr_info: holds current power mode
>    * @max_pwr_info: keeps the device max valid pwm
> + * @clk_scaling_lock: used to serialize device commands and clock scaling
>    * @desc_size: descriptor sizes reported by device
>    * @urgent_bkops_lvl: keeps track of urgent bkops level for device
>    * @is_urgent_bkops_lvl_checked: keeps track if the urgent bkops level 
> for


-- 
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum,
Linux Foundation Collaborative Project