Allow UFS suspend/resume callbacks to run in parallel with other
suspend/resume callbacks. This can recoup dozens of milliseconds on the
resume path if UFS hardware needs to be powered back on.
Suspending and resuming asynchronously is safe to do so long as the driver
callbacks only depend on resources made available by either a) parent
devices or b) devices explicitly marked as suppliers with device_link_add.
Cc: Bjorn Helgaas <[email protected]>
Cc: Jaegeuk Kim <[email protected]>
Cc: Bart Van Assche <[email protected]>
Cc: Adrian Hunter <[email protected]>
Cc: Stanley Chu <[email protected]>
Cc: Can Guo <[email protected]>
Cc: Asutosh Das <[email protected]>
Cc: Avri Altman <[email protected]>
Cc: Martin K. Petersen <[email protected]>
Signed-off-by: Vincent Palomares <[email protected]>
---
Are there any suspend/resume dependencies for UFS drivers not tracked by
the device parent relationship?
drivers/scsi/ufs/ufshcd.c | 1 +
1 file changed, 1 insertion(+)
diff --git a/drivers/scsi/ufs/ufshcd.c b/drivers/scsi/ufs/ufshcd.c
index b87ff68aa9aa..9ec5c308a0ea 100644
--- a/drivers/scsi/ufs/ufshcd.c
+++ b/drivers/scsi/ufs/ufshcd.c
@@ -9625,6 +9625,7 @@ int ufshcd_init(struct ufs_hba *hba, void __iomem *mmio_base, unsigned int irq)
async_schedule(ufshcd_async_scan, hba);
ufs_sysfs_add_nodes(hba->dev);
+ device_enable_async_suspend(dev);
return 0;
free_tmf_queue:
--
2.32.0.432.gabb21c7263-goog
On 7/27/21 6:27 PM, Vincent Palomares wrote:
> Are there any suspend/resume dependencies for UFS drivers not tracked by
> the device parent relationship?
Not that I know of. Since I like this patch:
Reviewed-by: Bart Van Assche <[email protected]>
> Allow UFS suspend/resume callbacks to run in parallel with other
> suspend/resume callbacks. This can recoup dozens of milliseconds on the
> resume path if UFS hardware needs to be powered back on.
>
> Suspending and resuming asynchronously is safe to do so long as the driver
> callbacks only depend on resources made available by either a) parent
> devices or b) devices explicitly marked as suppliers with device_link_add.
>
> Cc: Bjorn Helgaas <[email protected]>
> Cc: Jaegeuk Kim <[email protected]>
> Cc: Bart Van Assche <[email protected]>
> Cc: Adrian Hunter <[email protected]>
> Cc: Stanley Chu <[email protected]>
> Cc: Can Guo <[email protected]>
> Cc: Asutosh Das <[email protected]>
> Cc: Avri Altman <[email protected]>
> Cc: Martin K. Petersen <[email protected]>
> Signed-off-by: Vincent Palomares <[email protected]>
> ---
>
> Are there any suspend/resume dependencies for UFS drivers not tracked by
> the device parent relationship?
>
> drivers/scsi/ufs/ufshcd.c | 1 +
> 1 file changed, 1 insertion(+)
>
> diff --git a/drivers/scsi/ufs/ufshcd.c b/drivers/scsi/ufs/ufshcd.c
> index b87ff68aa9aa..9ec5c308a0ea 100644
> --- a/drivers/scsi/ufs/ufshcd.c
> +++ b/drivers/scsi/ufs/ufshcd.c
> @@ -9625,6 +9625,7 @@ int ufshcd_init(struct ufs_hba *hba, void __iomem
> *mmio_base, unsigned int irq)
> async_schedule(ufshcd_async_scan, hba);
> ufs_sysfs_add_nodes(hba->dev);
>
> + device_enable_async_suspend(dev);
> return 0;
Isn't device_enable_async_suspend is being called for each lun in scsi_sysfs_add_sdev Anyway?
Thanks,
Avri
>
> free_tmf_queue:
> --
> 2.32.0.432.gabb21c7263-goog
On 7/28/21 11:48 PM, Avri Altman wrote:
> Vincent wrote:
>> diff --git a/drivers/scsi/ufs/ufshcd.c b/drivers/scsi/ufs/ufshcd.c
>> index b87ff68aa9aa..9ec5c308a0ea 100644
>> --- a/drivers/scsi/ufs/ufshcd.c
>> +++ b/drivers/scsi/ufs/ufshcd.c
>> @@ -9625,6 +9625,7 @@ int ufshcd_init(struct ufs_hba *hba, void __iomem
>> *mmio_base, unsigned int irq)
>> async_schedule(ufshcd_async_scan, hba);
>> ufs_sysfs_add_nodes(hba->dev);
>>
>> + device_enable_async_suspend(dev);
>> return 0;
> Isn't device_enable_async_suspend is being called for each lun in scsi_sysfs_add_sdev Anyway?
Hi Avri,
Our measurements have shown that resume takes longer than it should with
encryption enabled. While suspending we change the power mode of the UFS
device to a mode in which it loses crypto keys. Restoring crypto keys
during resume (blk_ksm_reprogram_all_keys()) takes about 31 ms. This is
the long pole and takes much more time than resuming LUNs. This patch
makes UFS resume happen concurrently with resuming other devices in the
system instead of serializing it. Measurements have shown that this
patch significantly improves the time needed to resume an Android device.
Bart.
> On 7/28/21 11:48 PM, Avri Altman wrote:
> > Vincent wrote:
> >> diff --git a/drivers/scsi/ufs/ufshcd.c b/drivers/scsi/ufs/ufshcd.c
> >> index b87ff68aa9aa..9ec5c308a0ea 100644
> >> --- a/drivers/scsi/ufs/ufshcd.c
> >> +++ b/drivers/scsi/ufs/ufshcd.c
> >> @@ -9625,6 +9625,7 @@ int ufshcd_init(struct ufs_hba *hba, void __iomem
> >> *mmio_base, unsigned int irq)
> >> async_schedule(ufshcd_async_scan, hba);
> >> ufs_sysfs_add_nodes(hba->dev);
> >>
> >> + device_enable_async_suspend(dev);
> >> return 0;
> > Isn't device_enable_async_suspend is being called for each lun in
> scsi_sysfs_add_sdev Anyway?
>
> Hi Avri,
>
> Our measurements have shown that resume takes longer than it should with
> encryption enabled. While suspending we change the power mode of the UFS
> device to a mode in which it loses crypto keys. Restoring crypto keys
> during resume (blk_ksm_reprogram_all_keys()) takes about 31 ms. This is
> the long pole and takes much more time than resuming LUNs. This patch
> makes UFS resume happen concurrently with resuming other devices in the
> system instead of serializing it. Measurements have shown that this
> patch significantly improves the time needed to resume an Android device.
OK.
Thanks for the extra info.
Thanks,
Avri
>
> Bart.
Vincent,
> Allow UFS suspend/resume callbacks to run in parallel with other
> suspend/resume callbacks. This can recoup dozens of milliseconds on the
> resume path if UFS hardware needs to be powered back on.
Applied to 5.15/scsi-staging, thanks!
--
Martin K. Petersen Oracle Linux Engineering