| Message ID | 20250122100214.489749-6-quic_ziqichen@quicinc.com (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | Support Multi-frequency scale for UFS | expand |
On 1/22/25 2:02 AM, Ziqi Chen wrote: > + if (target_gear) { > + memcpy(&new_pwr_info, &hba->pwr_info, > + sizeof(struct ufs_pa_layer_attr)); Why memcpy() instead of an assignment? The advantage of an assignment is that the compiler can perform type checking. Thanks, Bart.
On 1/23/2025 2:32 AM, Bart Van Assche wrote: > On 1/22/25 2:02 AM, Ziqi Chen wrote: >> + if (target_gear) { >> + memcpy(&new_pwr_info, &hba->pwr_info, >> + sizeof(struct ufs_pa_layer_attr)); > > Why memcpy() instead of an assignment? The advantage of an assignment is > that the compiler can perform type checking. > > Thanks, > > Bart. > Hi Bart, We use memcpy() here is due to memcpy() can be faster than direct assignment. We don't worry about safety because they are same struct "ufs_pa_layer_attr" so that we can ensure the accuracy of number of bytes and member type. -Ziqi
On 1/22/25 11:41 PM, Ziqi Chen wrote: > We use memcpy() here is due to memcpy() can be faster than direct > assignment. We don't worry about safety because they are same struct > "ufs_pa_layer_attr" so that we can ensure the accuracy of number of > bytes and member type. The memcpy() call we are discussing is not in the hot path so it doesn't have to be hyper-optimized. Making the compiler perform type checking is more important in this code path than micro-optimizing the code. Additionally, please do not try to be smarter than the compiler. Compilers are able to convert struct assignments into a memcpy() call if there are good reasons to assume that the memcpy() call will be faster. Given the small size of struct ufs_pa_layer_attr (7 * 4 = 28 bytes), memberwise assignment probably is faster than a memcpy() call. The trunk version of gcc (ARM64) translates a memberwise assignment of struct ufs_pa_layer_attr into the following four assembler instructions (x0 and x1 point to struct ufs_pa_layer_attr instances, q30 and q31 are 128 bit registers): ldr q30, [x1] ldr q31, [x1, 12] str q30, [x0] str q31, [x0, 12] Thanks, Bart.
On 1/24/2025 2:02 AM, Bart Van Assche wrote: > On 1/22/25 11:41 PM, Ziqi Chen wrote: >> We use memcpy() here is due to memcpy() can be faster than direct >> assignment. We don't worry about safety because they are same struct >> "ufs_pa_layer_attr" so that we can ensure the accuracy of number of >> bytes and member type. > > The memcpy() call we are discussing is not in the hot path so it doesn't > have to be hyper-optimized. Making the compiler perform type checking is > more important in this code path than micro-optimizing the code. > > Additionally, please do not try to be smarter than the compiler. > Compilers are able to convert struct assignments into a memcpy() call if > there are good reasons to assume that the memcpy() call will be faster. > > Given the small size of struct ufs_pa_layer_attr (7 * 4 = 28 bytes), > memberwise assignment probably is faster than a memcpy() call. The trunk > version of gcc (ARM64) translates a memberwise assignment of struct > ufs_pa_layer_attr into the following four assembler instructions (x0 and > x1 point to struct ufs_pa_layer_attr instances, q30 and q31 are 128 bit > registers): > > ldr q30, [x1] > ldr q31, [x1, 12] > str q30, [x0] > str q31, [x0, 12] > > Thanks, > > Bart. > Sure , Let me try and test it. If works fine , I will update in next version. -Ziqi
diff --git a/drivers/ufs/core/ufshcd.c b/drivers/ufs/core/ufshcd.c index 8d295cc827cc..e0fc198328a5 100644 --- a/drivers/ufs/core/ufshcd.c +++ b/drivers/ufs/core/ufshcd.c @@ -1308,16 +1308,28 @@ static int ufshcd_wait_for_doorbell_clr(struct ufs_hba *hba, /** * ufshcd_scale_gear - scale up/down UFS gear * @hba: per adapter instance + * @target_gear: target gear to scale to * @scale_up: True for scaling up gear and false for scaling down * * Return: 0 for success; -EBUSY if scaling can't happen at this time; * non-zero for any other errors. */ -static int ufshcd_scale_gear(struct ufs_hba *hba, bool scale_up) +static int ufshcd_scale_gear(struct ufs_hba *hba, u32 target_gear, bool scale_up) { int ret = 0; struct ufs_pa_layer_attr new_pwr_info; + if (target_gear) { + memcpy(&new_pwr_info, &hba->pwr_info, + sizeof(struct ufs_pa_layer_attr)); + + new_pwr_info.gear_tx = target_gear; + new_pwr_info.gear_rx = target_gear; + + goto config_pwr_mode; + } + + /* Legacy gear scaling, in case vops_freq_to_gear_speed() is not implemented */ if (scale_up) { memcpy(&new_pwr_info, &hba->clk_scaling.saved_pwr_info, sizeof(struct ufs_pa_layer_attr)); @@ -1338,6 +1350,7 @@ static int ufshcd_scale_gear(struct ufs_hba *hba, bool scale_up) } } +config_pwr_mode: /* check if the power mode needs to be changed or not? */ ret = ufshcd_config_pwr_mode(hba, &new_pwr_info); if (ret) @@ -1408,15 +1421,19 @@ static void ufshcd_clock_scaling_unprepare(struct ufs_hba *hba, int err, bool sc static int ufshcd_devfreq_scale(struct ufs_hba *hba, unsigned long freq, bool scale_up) { + u32 old_gear = hba->pwr_info.gear_rx; + u32 new_gear = 0; int ret = 0; + ufshcd_vops_freq_to_gear_speed(hba, freq, &new_gear); + ret = ufshcd_clock_scaling_prepare(hba, 1 * USEC_PER_SEC); if (ret) return ret; /* scale down the gear before scaling down clocks */ if (!scale_up) { - ret = ufshcd_scale_gear(hba, false); + ret = ufshcd_scale_gear(hba, new_gear, false); if (ret) goto out_unprepare; } @@ -1424,13 +1441,13 @@ static int ufshcd_devfreq_scale(struct ufs_hba *hba, unsigned long freq, ret = ufshcd_scale_clks(hba, freq, scale_up); if (ret) { if (!scale_up) - ufshcd_scale_gear(hba, true); + ufshcd_scale_gear(hba, old_gear, true); goto out_unprepare; } /* scale up the gear after scaling up clocks */ if (scale_up) { - ret = ufshcd_scale_gear(hba, true); + ret = ufshcd_scale_gear(hba, new_gear, true); if (ret) { ufshcd_scale_clks(hba, hba->devfreq->previous_freq, false); @@ -1723,6 +1740,8 @@ static ssize_t ufshcd_clkscale_enable_store(struct device *dev, struct device_attribute *attr, const char *buf, size_t count) { struct ufs_hba *hba = dev_get_drvdata(dev); + struct ufs_clk_info *clki; + unsigned long freq; u32 value; int err = 0; @@ -1746,14 +1765,21 @@ static ssize_t ufshcd_clkscale_enable_store(struct device *dev, if (value) { ufshcd_resume_clkscaling(hba); - } else { - ufshcd_suspend_clkscaling(hba); - err = ufshcd_devfreq_scale(hba, ULONG_MAX, true); - if (err) - dev_err(hba->dev, "%s: failed to scale clocks up %d\n", - __func__, err); + goto out_rel; } + clki = list_first_entry(&hba->clk_list_head, struct ufs_clk_info, list); + freq = clki->max_freq; + + ufshcd_suspend_clkscaling(hba); + err = ufshcd_devfreq_scale(hba, freq, true); + if (err) + dev_err(hba->dev, "%s: failed to scale clocks up %d\n", + __func__, err); + else + hba->clk_scaling.target_freq = freq; + +out_rel: ufshcd_release(hba); ufshcd_rpm_put_sync(hba); out: