[PATCH V2 1/3] ufs: core: Add CPU latency QoS support for ufs driver
Naresh Maramaina
quic_mnaresh at quicinc.com
Thu Dec 7 03:26:44 PST 2023
On 12/6/2023 8:56 PM, Manivannan Sadhasivam wrote:
> On Mon, Dec 04, 2023 at 08:00:59PM +0530, Maramaina Naresh wrote:
>> Register ufs driver to CPU latency PM QoS framework can improves
>> ufs device random io performance.
>>
>> PM QoS initialization will insert new QoS request into the CPU
>> latency QoS list with the maximum latency PM_QOS_DEFAULT_VALUE
>> value.
>>
>> UFS driver will vote for performance mode on scale up and power
>> save mode for scale down.
>>
>> If clock scaling feature is not enabled then voting will be based
>> on clock on or off condition.
>>
>> tiotest benchmark tool io performance results on sm8550 platform:
>>
>> 1. Without PM QoS support
>> Type (Speed in) | Average of 18 iterations
>> Random Write(IPOS) | 41065.13
>> Random Read(IPOS) | 37101.3
>>
>> 2. With PM QoS support
>> Type (Speed in) | Average of 18 iterations
>> Random Write(IPOS) | 46784.9
>> Random Read(IPOS) | 42943.4
>> (Improvement % with PM QoS = ~15%).
>>
>> Co-developed-by: Nitin Rawat <quic_nitirawa at quicinc.com>
>> Signed-off-by: Nitin Rawat <quic_nitirawa at quicinc.com>
>> Signed-off-by: Naveen Kumar Goud Arepalli <quic_narepall at quicinc.com>
>> Signed-off-by: Maramaina Naresh <quic_mnaresh at quicinc.com>
>> ---
>> drivers/ufs/core/ufshcd-priv.h | 8 +++++
>> drivers/ufs/core/ufshcd.c | 62 ++++++++++++++++++++++++++++++++++
>> include/ufs/ufshcd.h | 16 +++++++++
>> 3 files changed, 86 insertions(+)
>>
>> diff --git a/drivers/ufs/core/ufshcd-priv.h b/drivers/ufs/core/ufshcd-priv.h
>> index f42d99ce5bf1..536805f6c4e1 100644
>> --- a/drivers/ufs/core/ufshcd-priv.h
>> +++ b/drivers/ufs/core/ufshcd-priv.h
>> @@ -241,6 +241,14 @@ static inline void ufshcd_vops_config_scaling_param(struct ufs_hba *hba,
>> hba->vops->config_scaling_param(hba, p, data);
>> }
>>
>> +static inline u32 ufshcd_vops_config_qos_vote(struct ufs_hba *hba)
>> +{
>> + if (hba->vops && hba->vops->config_qos_vote)
>> + return hba->vops->config_qos_vote(hba);
>
> Please remove this callback as Bart noted.
>
Sure Mani, will takecare of this comment.
>> +
>> + return UFSHCD_QOS_DEFAULT_VOTE;
>> +}
>> +
>> static inline void ufshcd_vops_reinit_notify(struct ufs_hba *hba)
>> {
>> if (hba->vops && hba->vops->reinit_notify)
>> diff --git a/drivers/ufs/core/ufshcd.c b/drivers/ufs/core/ufshcd.c
>> index ae9936fc6ffb..13370febd2b5 100644
>> --- a/drivers/ufs/core/ufshcd.c
>> +++ b/drivers/ufs/core/ufshcd.c
>> @@ -1001,6 +1001,20 @@ static bool ufshcd_is_unipro_pa_params_tuning_req(struct ufs_hba *hba)
>> return ufshcd_get_local_unipro_ver(hba) < UFS_UNIPRO_VER_1_6;
>> }
>>
>> +/**
>> + * ufshcd_pm_qos_perf - vote for PM QoS performance or power save mode
>
> ufshcd_pm_qos_update() - Update PM QoS request
>
Sure Mani, will takecare of this comment.
>> + * @hba: per adapter instance
>> + * @on: If True, vote for perf PM QoS mode otherwise power save mode
>> + */
>> +static void ufshcd_pm_qos_perf(struct ufs_hba *hba, bool on)
>> +{
>> + if (!hba->pm_qos_init)
>> + return;
>> +
>> + cpu_latency_qos_update_request(&hba->pm_qos_req, on ? hba->qos_vote
>> + : PM_QOS_DEFAULT_VALUE);
>> +}
>> +
>> /**
>> * ufshcd_set_clk_freq - set UFS controller clock frequencies
>> * @hba: per adapter instance
>> @@ -1153,6 +1167,10 @@ static int ufshcd_scale_clks(struct ufs_hba *hba, unsigned long freq,
>> trace_ufshcd_profile_clk_scaling(dev_name(hba->dev),
>> (scale_up ? "up" : "down"),
>> ktime_to_us(ktime_sub(ktime_get(), start)), ret);
>> +
>> + if (!ret)
>> + ufshcd_pm_qos_perf(hba, scale_up);
>
> Can't you just move this before trace_ufshcd_profile_clk_scaling()? This also
> avoids checking for !ret.
>
In this case, we need to use goto out; inside if(ret) of
ufshcd_vops_clk_scale_notify.
will do the above change, to enable ufshcd_pm_qos_perf before the out flag.
>> +
>> return ret;
>> }
>>
>> @@ -9204,6 +9222,8 @@ static int ufshcd_setup_clocks(struct ufs_hba *hba, bool on)
>> if (ret)
>> return ret;
>>
>> + if (!ufshcd_is_clkscaling_supported(hba))
>> + ufshcd_pm_qos_perf(hba, on);
>> out:
>> if (ret) {
>> list_for_each_entry(clki, head, list) {
>> @@ -9296,6 +9316,45 @@ static int ufshcd_init_clocks(struct ufs_hba *hba)
>> return ret;
>> }
>>
>> +/**
>> + * ufshcd_pm_qos_init - initialize PM QoS instance
>
> "Initialize PM QoS request"
>
Sure Mani, will takecare of this comment.
>> + * @hba: per adapter instance
>> + */
>> +static void ufshcd_pm_qos_init(struct ufs_hba *hba)
>> +{
>> + if (!(hba->caps & UFSHCD_CAP_PM_QOS))
>> + return;
>> +
>> + /*
>> + * called to configure PM QoS vote value for UFS host,
>> + * expecting qos vote return value from caller else
>> + * default vote value will be return.
>> + */
>> + hba->qos_vote = ufshcd_vops_config_qos_vote(hba);
>
> No need of this variable too if you get rid of the callback.
>
>> + cpu_latency_qos_add_request(&hba->pm_qos_req,
>> + PM_QOS_DEFAULT_VALUE);
>> +
>> + if (cpu_latency_qos_request_active(&hba->pm_qos_req))
>> + hba->pm_qos_init = true;
>
> Why do you need this flag?
this flag ensure UFS qos request got added into the Global PM QoS list.
>
>> +
>> + dev_dbg(hba->dev, "%s: QoS %s, qos_vote: %u\n", __func__,
>> + hba->pm_qos_init ? "initialized" : "uninitialized",
>> + hba->qos_vote);
>> +}
>> +
>> +/**
>> + * ufshcd_pm_qos_exit - remove instance from PM QoS
>> + * @hba: per adapter instance
>> + */
>> +static void ufshcd_pm_qos_exit(struct ufs_hba *hba)
>> +{
>> + if (!hba->pm_qos_init)
>> + return;
>> +
>> + cpu_latency_qos_remove_request(&hba->pm_qos_req);
>> + hba->pm_qos_init = false;
>> +}
>> +
>
> [...]
>
>> /**
>> * struct ufs_hba - per adapter private structure
>> * @mmio_base: UFSHCI base register address
>> @@ -912,6 +923,8 @@ enum ufshcd_mcq_opr {
>> * @mcq_base: Multi circular queue registers base address
>> * @uhq: array of supported hardware queues
>> * @dev_cmd_queue: Queue for issuing device management commands
>> + * @pm_qos_req: PM QoS request handle
>> + * @pm_qos_init: flag to check if pm qos init completed
>> */
>> struct ufs_hba {
>> void __iomem *mmio_base;
>> @@ -1076,6 +1089,9 @@ struct ufs_hba {
>> struct ufs_hw_queue *uhq;
>> struct ufs_hw_queue *dev_cmd_queue;
>> struct ufshcd_mcq_opr_info_t mcq_opr[OPR_MAX];
>> + struct pm_qos_request pm_qos_req;
>> + bool pm_qos_init;
>> + u32 qos_vote;
>
> Order doesn't match Kdoc.
>
we are removing qos_vote variable in next patch series.
> - Mani
>
Thanks,
Naresh.
More information about the linux-arm-kernel
mailing list