v4:
-Added reviewed-by tag
-Return appropriate error code from mhi_poll_reg_field()
-Fixed bug where mhi_poll_reg_field() returns success if polling times out
-Added an interval_us variable in mhi_ready_state_transition()
v3:
-Removed config changes that crept in in the first patch
v2:
-Addressed review comments
-Introduce new patch for to use controller defined read_reg() for polling
-Add usage in RDDM download panic path as well
Use polling instead of interrupt driven approach to wait for MHI ready state.
In certain devices, it is likely that there is no incoming MHI
interrupt for a transition to MHI READY state. One such example
is the move from Pass Through to an SBL or AMSS execution
environment. In order to facilitate faster bootup times as there
is no need to wait until timeout_ms completes, MHI host can poll
every 25 milliseconds to check if device has entered MHI READY
until a maximum timeout of twice the timeout_ms is reached.
This patch series has been tested on an arm64 device.
Bhaumik Bhatt (3):
bus: mhi: core: Introduce internal register poll helper function
bus: mhi: core: Move to polling method to wait for MHI ready
bus: mhi: core: Use poll register read API for RDDM download
drivers/bus/mhi/core/boot.c | 20 ++++++--------------
drivers/bus/mhi/core/internal.h | 3 +++
drivers/bus/mhi/core/main.c | 23 +++++++++++++++++++++++
drivers/bus/mhi/core/pm.c | 32 +++++++++++++++-----------------
4 files changed, 47 insertions(+), 31 deletions(-)
--
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum,
a Linux Foundation Collaborative Project
Introduce helper function to allow MHI core driver to poll for
a value in a register field. This helps reach a common path to
read and poll register values along with a retry time interval.
Signed-off-by: Bhaumik Bhatt <[email protected]>
---
drivers/bus/mhi/core/internal.h | 3 +++
drivers/bus/mhi/core/main.c | 23 +++++++++++++++++++++++
2 files changed, 26 insertions(+)
diff --git a/drivers/bus/mhi/core/internal.h b/drivers/bus/mhi/core/internal.h
index 6f80ec3..005286b 100644
--- a/drivers/bus/mhi/core/internal.h
+++ b/drivers/bus/mhi/core/internal.h
@@ -643,6 +643,9 @@ int __must_check mhi_read_reg(struct mhi_controller *mhi_cntrl,
int __must_check mhi_read_reg_field(struct mhi_controller *mhi_cntrl,
void __iomem *base, u32 offset, u32 mask,
u32 shift, u32 *out);
+int __must_check mhi_poll_reg_field(struct mhi_controller *mhi_cntrl,
+ void __iomem *base, u32 offset, u32 mask,
+ u32 shift, u32 val, u32 delayus);
void mhi_write_reg(struct mhi_controller *mhi_cntrl, void __iomem *base,
u32 offset, u32 val);
void mhi_write_reg_field(struct mhi_controller *mhi_cntrl, void __iomem *base,
diff --git a/drivers/bus/mhi/core/main.c b/drivers/bus/mhi/core/main.c
index 4e0131b..7c7f41a 100644
--- a/drivers/bus/mhi/core/main.c
+++ b/drivers/bus/mhi/core/main.c
@@ -4,6 +4,7 @@
*
*/
+#include <linux/delay.h>
#include <linux/device.h>
#include <linux/dma-direction.h>
#include <linux/dma-mapping.h>
@@ -37,6 +38,28 @@ int __must_check mhi_read_reg_field(struct mhi_controller *mhi_cntrl,
return 0;
}
+int __must_check mhi_poll_reg_field(struct mhi_controller *mhi_cntrl,
+ void __iomem *base, u32 offset,
+ u32 mask, u32 shift, u32 val, u32 delayus)
+{
+ int ret;
+ u32 out, retry = (mhi_cntrl->timeout_ms * 1000) / delayus;
+
+ while (retry--) {
+ ret = mhi_read_reg_field(mhi_cntrl, base, offset, mask, shift,
+ &out);
+ if (ret)
+ return ret;
+
+ if (out == val)
+ return 0;
+
+ udelay(delayus);
+ }
+
+ return -ENOENT;
+}
+
void mhi_write_reg(struct mhi_controller *mhi_cntrl, void __iomem *base,
u32 offset, u32 val)
{
--
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum,
a Linux Foundation Collaborative Project
Hi Bhaumik,
On Thu, 11 Mar 2021 at 00:31, Bhaumik Bhatt <[email protected]> wrote:
>
> Introduce helper function to allow MHI core driver to poll for
> a value in a register field. This helps reach a common path to
> read and poll register values along with a retry time interval.
>
> Signed-off-by: Bhaumik Bhatt <[email protected]>
> ---
> drivers/bus/mhi/core/internal.h | 3 +++
> drivers/bus/mhi/core/main.c | 23 +++++++++++++++++++++++
> 2 files changed, 26 insertions(+)
>
> diff --git a/drivers/bus/mhi/core/internal.h b/drivers/bus/mhi/core/internal.h
> index 6f80ec3..005286b 100644
> --- a/drivers/bus/mhi/core/internal.h
> +++ b/drivers/bus/mhi/core/internal.h
> @@ -643,6 +643,9 @@ int __must_check mhi_read_reg(struct mhi_controller *mhi_cntrl,
> int __must_check mhi_read_reg_field(struct mhi_controller *mhi_cntrl,
> void __iomem *base, u32 offset, u32 mask,
> u32 shift, u32 *out);
> +int __must_check mhi_poll_reg_field(struct mhi_controller *mhi_cntrl,
> + void __iomem *base, u32 offset, u32 mask,
> + u32 shift, u32 val, u32 delayus);
> void mhi_write_reg(struct mhi_controller *mhi_cntrl, void __iomem *base,
> u32 offset, u32 val);
> void mhi_write_reg_field(struct mhi_controller *mhi_cntrl, void __iomem *base,
> diff --git a/drivers/bus/mhi/core/main.c b/drivers/bus/mhi/core/main.c
> index 4e0131b..7c7f41a 100644
> --- a/drivers/bus/mhi/core/main.c
> +++ b/drivers/bus/mhi/core/main.c
> @@ -4,6 +4,7 @@
> *
> */
>
> +#include <linux/delay.h>
> #include <linux/device.h>
> #include <linux/dma-direction.h>
> #include <linux/dma-mapping.h>
> @@ -37,6 +38,28 @@ int __must_check mhi_read_reg_field(struct mhi_controller *mhi_cntrl,
> return 0;
> }
>
> +int __must_check mhi_poll_reg_field(struct mhi_controller *mhi_cntrl,
> + void __iomem *base, u32 offset,
> + u32 mask, u32 shift, u32 val, u32 delayus)
> +{
> + int ret;
> + u32 out, retry = (mhi_cntrl->timeout_ms * 1000) / delayus;
> +
> + while (retry--) {
> + ret = mhi_read_reg_field(mhi_cntrl, base, offset, mask, shift,
> + &out);
> + if (ret)
> + return ret;
> +
> + if (out == val)
> + return 0;
> +
> + udelay(delayus);
Have you read my previous comment?
Do you really want to risk hogging the CPU for several seconds? we
know that some devices take several seconds to start/boot.
Why not using msleep variant here?
Regards,
Loic
On 3/11/2021 1:00 AM, Loic Poulain wrote:
> Hi Bhaumik,
>
> On Thu, 11 Mar 2021 at 00:31, Bhaumik Bhatt <[email protected]> wrote:
>>
>> Introduce helper function to allow MHI core driver to poll for
>> a value in a register field. This helps reach a common path to
>> read and poll register values along with a retry time interval.
>>
>> Signed-off-by: Bhaumik Bhatt <[email protected]>
>> ---
>> drivers/bus/mhi/core/internal.h | 3 +++
>> drivers/bus/mhi/core/main.c | 23 +++++++++++++++++++++++
>> 2 files changed, 26 insertions(+)
>>
>> diff --git a/drivers/bus/mhi/core/internal.h b/drivers/bus/mhi/core/internal.h
>> index 6f80ec3..005286b 100644
>> --- a/drivers/bus/mhi/core/internal.h
>> +++ b/drivers/bus/mhi/core/internal.h
>> @@ -643,6 +643,9 @@ int __must_check mhi_read_reg(struct mhi_controller *mhi_cntrl,
>> int __must_check mhi_read_reg_field(struct mhi_controller *mhi_cntrl,
>> void __iomem *base, u32 offset, u32 mask,
>> u32 shift, u32 *out);
>> +int __must_check mhi_poll_reg_field(struct mhi_controller *mhi_cntrl,
>> + void __iomem *base, u32 offset, u32 mask,
>> + u32 shift, u32 val, u32 delayus);
>> void mhi_write_reg(struct mhi_controller *mhi_cntrl, void __iomem *base,
>> u32 offset, u32 val);
>> void mhi_write_reg_field(struct mhi_controller *mhi_cntrl, void __iomem *base,
>> diff --git a/drivers/bus/mhi/core/main.c b/drivers/bus/mhi/core/main.c
>> index 4e0131b..7c7f41a 100644
>> --- a/drivers/bus/mhi/core/main.c
>> +++ b/drivers/bus/mhi/core/main.c
>> @@ -4,6 +4,7 @@
>> *
>> */
>>
>> +#include <linux/delay.h>
>> #include <linux/device.h>
>> #include <linux/dma-direction.h>
>> #include <linux/dma-mapping.h>
>> @@ -37,6 +38,28 @@ int __must_check mhi_read_reg_field(struct mhi_controller *mhi_cntrl,
>> return 0;
>> }
>>
>> +int __must_check mhi_poll_reg_field(struct mhi_controller *mhi_cntrl,
>> + void __iomem *base, u32 offset,
>> + u32 mask, u32 shift, u32 val, u32 delayus)
>> +{
>> + int ret;
>> + u32 out, retry = (mhi_cntrl->timeout_ms * 1000) / delayus;
>> +
>> + while (retry--) {
>> + ret = mhi_read_reg_field(mhi_cntrl, base, offset, mask, shift,
>> + &out);
>> + if (ret)
>> + return ret;
>> +
>> + if (out == val)
>> + return 0;
>> +
>> + udelay(delayus);
>
> Have you read my previous comment?
> Do you really want to risk hogging the CPU for several seconds? we
> know that some devices take several seconds to start/boot.
> Why not using msleep variant here?
usleep_range() if there is a desire to stay in us units?
Given that the use of this function is for 25ms in one case, I wonder if
this warning is applicable:
https://elixir.bootlin.com/linux/latest/source/include/linux/delay.h#L28
Counter point, 1ms latency over PCIe is not unusual. I know we've
removed the PCIe dependencies from MHI, but PCIe is the real usecase at
this time. Seems like this function could behave a bit weird if the
parameter to udelay is something like "100", but the
mhi_read_reg_field() call takes significantly longer than that. Feels
like in some scenarios, we could actually exceed the timeout by a
non-trivial margin.
I guess I'm going back and forth in determining if us scale timing is a
benefit in any way.
--
Jeffrey Hugo
Qualcomm Technologies, Inc. is a member of the
Code Aurora Forum, a Linux Foundation Collaborative Project.
On 2021-03-11 11:59 AM, Jeffrey Hugo wrote:
> On 3/11/2021 1:00 AM, Loic Poulain wrote:
>> Hi Bhaumik,
>>
>> On Thu, 11 Mar 2021 at 00:31, Bhaumik Bhatt <[email protected]>
>> wrote:
>>>
>>> Introduce helper function to allow MHI core driver to poll for
>>> a value in a register field. This helps reach a common path to
>>> read and poll register values along with a retry time interval.
>>>
>>> Signed-off-by: Bhaumik Bhatt <[email protected]>
>>> ---
>>> drivers/bus/mhi/core/internal.h | 3 +++
>>> drivers/bus/mhi/core/main.c | 23 +++++++++++++++++++++++
>>> 2 files changed, 26 insertions(+)
>>>
>>> diff --git a/drivers/bus/mhi/core/internal.h
>>> b/drivers/bus/mhi/core/internal.h
>>> index 6f80ec3..005286b 100644
>>> --- a/drivers/bus/mhi/core/internal.h
>>> +++ b/drivers/bus/mhi/core/internal.h
>>> @@ -643,6 +643,9 @@ int __must_check mhi_read_reg(struct
>>> mhi_controller *mhi_cntrl,
>>> int __must_check mhi_read_reg_field(struct mhi_controller
>>> *mhi_cntrl,
>>> void __iomem *base, u32 offset,
>>> u32 mask,
>>> u32 shift, u32 *out);
>>> +int __must_check mhi_poll_reg_field(struct mhi_controller
>>> *mhi_cntrl,
>>> + void __iomem *base, u32 offset,
>>> u32 mask,
>>> + u32 shift, u32 val, u32 delayus);
>>> void mhi_write_reg(struct mhi_controller *mhi_cntrl, void __iomem
>>> *base,
>>> u32 offset, u32 val);
>>> void mhi_write_reg_field(struct mhi_controller *mhi_cntrl, void
>>> __iomem *base,
>>> diff --git a/drivers/bus/mhi/core/main.c
>>> b/drivers/bus/mhi/core/main.c
>>> index 4e0131b..7c7f41a 100644
>>> --- a/drivers/bus/mhi/core/main.c
>>> +++ b/drivers/bus/mhi/core/main.c
>>> @@ -4,6 +4,7 @@
>>> *
>>> */
>>>
>>> +#include <linux/delay.h>
>>> #include <linux/device.h>
>>> #include <linux/dma-direction.h>
>>> #include <linux/dma-mapping.h>
>>> @@ -37,6 +38,28 @@ int __must_check mhi_read_reg_field(struct
>>> mhi_controller *mhi_cntrl,
>>> return 0;
>>> }
>>>
>>> +int __must_check mhi_poll_reg_field(struct mhi_controller
>>> *mhi_cntrl,
>>> + void __iomem *base, u32 offset,
>>> + u32 mask, u32 shift, u32 val, u32
>>> delayus)
>>> +{
>>> + int ret;
>>> + u32 out, retry = (mhi_cntrl->timeout_ms * 1000) / delayus;
>>> +
>>> + while (retry--) {
>>> + ret = mhi_read_reg_field(mhi_cntrl, base, offset,
>>> mask, shift,
>>> + &out);
>>> + if (ret)
>>> + return ret;
>>> +
>>> + if (out == val)
>>> + return 0;
>>> +
>>> + udelay(delayus);
>>
>> Have you read my previous comment?
>> Do you really want to risk hogging the CPU for several seconds? we
>> know that some devices take several seconds to start/boot.
>> Why not using msleep variant here?
>
> usleep_range() if there is a desire to stay in us units?
>
> Given that the use of this function is for 25ms in one case, I wonder
> if this warning is applicable:
> https://elixir.bootlin.com/linux/latest/source/include/linux/delay.h#L28
>
> Counter point, 1ms latency over PCIe is not unusual. I know we've
> removed the PCIe dependencies from MHI, but PCIe is the real usecase
> at this time. Seems like this function could behave a bit weird if
> the parameter to udelay is something like "100", but the
> mhi_read_reg_field() call takes significantly longer than that. Feels
> like in some scenarios, we could actually exceed the timeout by a
> non-trivial margin.
>
> I guess I'm going back and forth in determining if us scale timing is
> a benefit in any way.
Thanks for all the inputs. I think a good idea here would be to use
fsleep()
API as we need to allow any timeout the caller specifies. Also, plan is
to
drop the patch #3 in this series since that will require a busywait due
to
the code being in panic path.
I don't wish to accommodate another variable here for busywait but that
would be an option to pick sleep or delay depending on the caller's
path.
Please respond if there are any concerns.
Thanks,
Bhaumik
---
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora
Forum,
a Linux Foundation Collaborative Project
On 3/17/2021 3:26 PM, Bhaumik Bhatt wrote:
> On 2021-03-11 11:59 AM, Jeffrey Hugo wrote:
>> On 3/11/2021 1:00 AM, Loic Poulain wrote:
>>> Hi Bhaumik,
>>>
>>> On Thu, 11 Mar 2021 at 00:31, Bhaumik Bhatt <[email protected]>
>>> wrote:
>>>>
>>>> Introduce helper function to allow MHI core driver to poll for
>>>> a value in a register field. This helps reach a common path to
>>>> read and poll register values along with a retry time interval.
>>>>
>>>> Signed-off-by: Bhaumik Bhatt <[email protected]>
>>>> ---
>>>> drivers/bus/mhi/core/internal.h | 3 +++
>>>> drivers/bus/mhi/core/main.c | 23 +++++++++++++++++++++++
>>>> 2 files changed, 26 insertions(+)
>>>>
>>>> diff --git a/drivers/bus/mhi/core/internal.h
>>>> b/drivers/bus/mhi/core/internal.h
>>>> index 6f80ec3..005286b 100644
>>>> --- a/drivers/bus/mhi/core/internal.h
>>>> +++ b/drivers/bus/mhi/core/internal.h
>>>> @@ -643,6 +643,9 @@ int __must_check mhi_read_reg(struct
>>>> mhi_controller *mhi_cntrl,
>>>> int __must_check mhi_read_reg_field(struct mhi_controller *mhi_cntrl,
>>>> void __iomem *base, u32 offset,
>>>> u32 mask,
>>>> u32 shift, u32 *out);
>>>> +int __must_check mhi_poll_reg_field(struct mhi_controller *mhi_cntrl,
>>>> + void __iomem *base, u32 offset,
>>>> u32 mask,
>>>> + u32 shift, u32 val, u32 delayus);
>>>> void mhi_write_reg(struct mhi_controller *mhi_cntrl, void __iomem
>>>> *base,
>>>> u32 offset, u32 val);
>>>> void mhi_write_reg_field(struct mhi_controller *mhi_cntrl, void
>>>> __iomem *base,
>>>> diff --git a/drivers/bus/mhi/core/main.c b/drivers/bus/mhi/core/main.c
>>>> index 4e0131b..7c7f41a 100644
>>>> --- a/drivers/bus/mhi/core/main.c
>>>> +++ b/drivers/bus/mhi/core/main.c
>>>> @@ -4,6 +4,7 @@
>>>> *
>>>> */
>>>>
>>>> +#include <linux/delay.h>
>>>> #include <linux/device.h>
>>>> #include <linux/dma-direction.h>
>>>> #include <linux/dma-mapping.h>
>>>> @@ -37,6 +38,28 @@ int __must_check mhi_read_reg_field(struct
>>>> mhi_controller *mhi_cntrl,
>>>> return 0;
>>>> }
>>>>
>>>> +int __must_check mhi_poll_reg_field(struct mhi_controller *mhi_cntrl,
>>>> + void __iomem *base, u32 offset,
>>>> + u32 mask, u32 shift, u32 val,
>>>> u32 delayus)
>>>> +{
>>>> + int ret;
>>>> + u32 out, retry = (mhi_cntrl->timeout_ms * 1000) / delayus;
>>>> +
>>>> + while (retry--) {
>>>> + ret = mhi_read_reg_field(mhi_cntrl, base, offset,
>>>> mask, shift,
>>>> + &out);
>>>> + if (ret)
>>>> + return ret;
>>>> +
>>>> + if (out == val)
>>>> + return 0;
>>>> +
>>>> + udelay(delayus);
>>>
>>> Have you read my previous comment?
>>> Do you really want to risk hogging the CPU for several seconds? we
>>> know that some devices take several seconds to start/boot.
>>> Why not using msleep variant here?
>>
>> usleep_range() if there is a desire to stay in us units?
>>
>> Given that the use of this function is for 25ms in one case, I wonder
>> if this warning is applicable:
>> https://elixir.bootlin.com/linux/latest/source/include/linux/delay.h#L28
>>
>> Counter point, 1ms latency over PCIe is not unusual. I know we've
>> removed the PCIe dependencies from MHI, but PCIe is the real usecase
>> at this time. Seems like this function could behave a bit weird if
>> the parameter to udelay is something like "100", but the
>> mhi_read_reg_field() call takes significantly longer than that. Feels
>> like in some scenarios, we could actually exceed the timeout by a
>> non-trivial margin.
>>
>> I guess I'm going back and forth in determining if us scale timing is
>> a benefit in any way.
> Thanks for all the inputs. I think a good idea here would be to use
> fsleep()
> API as we need to allow any timeout the caller specifies. Also, plan is to
> drop the patch #3 in this series since that will require a busywait due to
> the code being in panic path.
>
> I don't wish to accommodate another variable here for busywait but that
> would be an option to pick sleep or delay depending on the caller's path.
>
> Please respond if there are any concerns.
fsleep() would be some improvement, but I think there is still the issue
Loic points out where if delayus is small, but timeout_ms is large (say
50us and 25s), this function will end up burning a lot of cpu cycles.
However that is likely an edge case, and if your poll cycle is that
small, I think it should be assumed that the event is expected to happen
quickly, so the full timeout should not be hit.
fsleep() does nothing to address this function possibly taking quite a
bit longer than the timeout in overall wall time. Perhaps that is not a
significant concern though.
--
Jeffrey Hugo
Qualcomm Technologies, Inc. is a member of the
Code Aurora Forum, a Linux Foundation Collaborative Project.
On Thu, 18 Mar 2021 at 17:13, Jeffrey Hugo <[email protected]> wrote:
>
> On 3/17/2021 3:26 PM, Bhaumik Bhatt wrote:
> > On 2021-03-11 11:59 AM, Jeffrey Hugo wrote:
> >> On 3/11/2021 1:00 AM, Loic Poulain wrote:
> >>> Hi Bhaumik,
> >>>
> >>> On Thu, 11 Mar 2021 at 00:31, Bhaumik Bhatt <[email protected]>
> >>> wrote:
> >>>>
> >>>> Introduce helper function to allow MHI core driver to poll for
> >>>> a value in a register field. This helps reach a common path to
> >>>> read and poll register values along with a retry time interval.
> >>>>
> >>>> Signed-off-by: Bhaumik Bhatt <[email protected]>
> >>>> ---
> >>>> drivers/bus/mhi/core/internal.h | 3 +++
> >>>> drivers/bus/mhi/core/main.c | 23 +++++++++++++++++++++++
> >>>> 2 files changed, 26 insertions(+)
> >>>>
> >>>> diff --git a/drivers/bus/mhi/core/internal.h
> >>>> b/drivers/bus/mhi/core/internal.h
> >>>> index 6f80ec3..005286b 100644
> >>>> --- a/drivers/bus/mhi/core/internal.h
> >>>> +++ b/drivers/bus/mhi/core/internal.h
> >>>> @@ -643,6 +643,9 @@ int __must_check mhi_read_reg(struct
> >>>> mhi_controller *mhi_cntrl,
> >>>> int __must_check mhi_read_reg_field(struct mhi_controller *mhi_cntrl,
> >>>> void __iomem *base, u32 offset,
> >>>> u32 mask,
> >>>> u32 shift, u32 *out);
> >>>> +int __must_check mhi_poll_reg_field(struct mhi_controller *mhi_cntrl,
> >>>> + void __iomem *base, u32 offset,
> >>>> u32 mask,
> >>>> + u32 shift, u32 val, u32 delayus);
> >>>> void mhi_write_reg(struct mhi_controller *mhi_cntrl, void __iomem
> >>>> *base,
> >>>> u32 offset, u32 val);
> >>>> void mhi_write_reg_field(struct mhi_controller *mhi_cntrl, void
> >>>> __iomem *base,
> >>>> diff --git a/drivers/bus/mhi/core/main.c b/drivers/bus/mhi/core/main.c
> >>>> index 4e0131b..7c7f41a 100644
> >>>> --- a/drivers/bus/mhi/core/main.c
> >>>> +++ b/drivers/bus/mhi/core/main.c
> >>>> @@ -4,6 +4,7 @@
> >>>> *
> >>>> */
> >>>>
> >>>> +#include <linux/delay.h>
> >>>> #include <linux/device.h>
> >>>> #include <linux/dma-direction.h>
> >>>> #include <linux/dma-mapping.h>
> >>>> @@ -37,6 +38,28 @@ int __must_check mhi_read_reg_field(struct
> >>>> mhi_controller *mhi_cntrl,
> >>>> return 0;
> >>>> }
> >>>>
> >>>> +int __must_check mhi_poll_reg_field(struct mhi_controller *mhi_cntrl,
> >>>> + void __iomem *base, u32 offset,
> >>>> + u32 mask, u32 shift, u32 val,
> >>>> u32 delayus)
> >>>> +{
> >>>> + int ret;
> >>>> + u32 out, retry = (mhi_cntrl->timeout_ms * 1000) / delayus;
> >>>> +
> >>>> + while (retry--) {
> >>>> + ret = mhi_read_reg_field(mhi_cntrl, base, offset,
> >>>> mask, shift,
> >>>> + &out);
> >>>> + if (ret)
> >>>> + return ret;
> >>>> +
> >>>> + if (out == val)
> >>>> + return 0;
> >>>> +
> >>>> + udelay(delayus);
> >>>
> >>> Have you read my previous comment?
> >>> Do you really want to risk hogging the CPU for several seconds? we
> >>> know that some devices take several seconds to start/boot.
> >>> Why not using msleep variant here?
> >>
> >> usleep_range() if there is a desire to stay in us units?
> >>
> >> Given that the use of this function is for 25ms in one case, I wonder
> >> if this warning is applicable:
> >> https://elixir.bootlin.com/linux/latest/source/include/linux/delay.h#L28
> >>
> >> Counter point, 1ms latency over PCIe is not unusual. I know we've
> >> removed the PCIe dependencies from MHI, but PCIe is the real usecase
> >> at this time. Seems like this function could behave a bit weird if
> >> the parameter to udelay is something like "100", but the
> >> mhi_read_reg_field() call takes significantly longer than that. Feels
> >> like in some scenarios, we could actually exceed the timeout by a
> >> non-trivial margin.
> >>
> >> I guess I'm going back and forth in determining if us scale timing is
> >> a benefit in any way.
> > Thanks for all the inputs. I think a good idea here would be to use
> > fsleep()
> > API as we need to allow any timeout the caller specifies. Also, plan is to
> > drop the patch #3 in this series since that will require a busywait due to
> > the code being in panic path.
> >
> > I don't wish to accommodate another variable here for busywait but that
> > would be an option to pick sleep or delay depending on the caller's path.
> >
> > Please respond if there are any concerns.
>
> fsleep() would be some improvement, but I think there is still the issue
> Loic points out where if delayus is small, but timeout_ms is large (say
> 50us and 25s), this function will end up burning a lot of cpu cycles
> However that is likely an edge case, and if your poll cycle is that
> small, I think it should be assumed that the event is expected to happen
> quickly, so the full timeout should not be hit.
Well, my point is that during initial power_up, with a device
cold-booting, it can take several seconds for it to reach ready state
(not a corner case). That why timeout_ms can be as large as 20 seconds
for mhi_pci_modem. If polling is based on busy-wait, that means the
while loop will not let the CPU running anything else for several
seconds. Not sure what is the expected meaning of this timeout_ms in
first place... maybe I just use it badly.
Moreover, do we need microsecond latency on detecting ready
transition, this is not a critical path, right?
Regards,
Loic
On 2021-03-18 09:43 AM, Loic Poulain wrote:
> On Thu, 18 Mar 2021 at 17:13, Jeffrey Hugo <[email protected]>
> wrote:
>>
>> On 3/17/2021 3:26 PM, Bhaumik Bhatt wrote:
>> > On 2021-03-11 11:59 AM, Jeffrey Hugo wrote:
>> >> On 3/11/2021 1:00 AM, Loic Poulain wrote:
>> >>> Hi Bhaumik,
>> >>>
>> >>> On Thu, 11 Mar 2021 at 00:31, Bhaumik Bhatt <[email protected]>
>> >>> wrote:
>> >>>>
>> >>>> Introduce helper function to allow MHI core driver to poll for
>> >>>> a value in a register field. This helps reach a common path to
>> >>>> read and poll register values along with a retry time interval.
>> >>>>
>> >>>> Signed-off-by: Bhaumik Bhatt <[email protected]>
>> >>>> ---
>> >>>> drivers/bus/mhi/core/internal.h | 3 +++
>> >>>> drivers/bus/mhi/core/main.c | 23 +++++++++++++++++++++++
>> >>>> 2 files changed, 26 insertions(+)
>> >>>>
>> >>>> diff --git a/drivers/bus/mhi/core/internal.h
>> >>>> b/drivers/bus/mhi/core/internal.h
>> >>>> index 6f80ec3..005286b 100644
>> >>>> --- a/drivers/bus/mhi/core/internal.h
>> >>>> +++ b/drivers/bus/mhi/core/internal.h
>> >>>> @@ -643,6 +643,9 @@ int __must_check mhi_read_reg(struct
>> >>>> mhi_controller *mhi_cntrl,
>> >>>> int __must_check mhi_read_reg_field(struct mhi_controller *mhi_cntrl,
>> >>>> void __iomem *base, u32 offset,
>> >>>> u32 mask,
>> >>>> u32 shift, u32 *out);
>> >>>> +int __must_check mhi_poll_reg_field(struct mhi_controller *mhi_cntrl,
>> >>>> + void __iomem *base, u32 offset,
>> >>>> u32 mask,
>> >>>> + u32 shift, u32 val, u32 delayus);
>> >>>> void mhi_write_reg(struct mhi_controller *mhi_cntrl, void __iomem
>> >>>> *base,
>> >>>> u32 offset, u32 val);
>> >>>> void mhi_write_reg_field(struct mhi_controller *mhi_cntrl, void
>> >>>> __iomem *base,
>> >>>> diff --git a/drivers/bus/mhi/core/main.c b/drivers/bus/mhi/core/main.c
>> >>>> index 4e0131b..7c7f41a 100644
>> >>>> --- a/drivers/bus/mhi/core/main.c
>> >>>> +++ b/drivers/bus/mhi/core/main.c
>> >>>> @@ -4,6 +4,7 @@
>> >>>> *
>> >>>> */
>> >>>>
>> >>>> +#include <linux/delay.h>
>> >>>> #include <linux/device.h>
>> >>>> #include <linux/dma-direction.h>
>> >>>> #include <linux/dma-mapping.h>
>> >>>> @@ -37,6 +38,28 @@ int __must_check mhi_read_reg_field(struct
>> >>>> mhi_controller *mhi_cntrl,
>> >>>> return 0;
>> >>>> }
>> >>>>
>> >>>> +int __must_check mhi_poll_reg_field(struct mhi_controller *mhi_cntrl,
>> >>>> + void __iomem *base, u32 offset,
>> >>>> + u32 mask, u32 shift, u32 val,
>> >>>> u32 delayus)
>> >>>> +{
>> >>>> + int ret;
>> >>>> + u32 out, retry = (mhi_cntrl->timeout_ms * 1000) / delayus;
>> >>>> +
>> >>>> + while (retry--) {
>> >>>> + ret = mhi_read_reg_field(mhi_cntrl, base, offset,
>> >>>> mask, shift,
>> >>>> + &out);
>> >>>> + if (ret)
>> >>>> + return ret;
>> >>>> +
>> >>>> + if (out == val)
>> >>>> + return 0;
>> >>>> +
>> >>>> + udelay(delayus);
>> >>>
>> >>> Have you read my previous comment?
>> >>> Do you really want to risk hogging the CPU for several seconds? we
>> >>> know that some devices take several seconds to start/boot.
>> >>> Why not using msleep variant here?
>> >>
>> >> usleep_range() if there is a desire to stay in us units?
>> >>
>> >> Given that the use of this function is for 25ms in one case, I wonder
>> >> if this warning is applicable:
>> >> https://elixir.bootlin.com/linux/latest/source/include/linux/delay.h#L28
>> >>
>> >> Counter point, 1ms latency over PCIe is not unusual. I know we've
>> >> removed the PCIe dependencies from MHI, but PCIe is the real usecase
>> >> at this time. Seems like this function could behave a bit weird if
>> >> the parameter to udelay is something like "100", but the
>> >> mhi_read_reg_field() call takes significantly longer than that. Feels
>> >> like in some scenarios, we could actually exceed the timeout by a
>> >> non-trivial margin.
>> >>
>> >> I guess I'm going back and forth in determining if us scale timing is
>> >> a benefit in any way.
>> > Thanks for all the inputs. I think a good idea here would be to use
>> > fsleep()
>> > API as we need to allow any timeout the caller specifies. Also, plan is to
>> > drop the patch #3 in this series since that will require a busywait due to
>> > the code being in panic path.
>> >
>> > I don't wish to accommodate another variable here for busywait but that
>> > would be an option to pick sleep or delay depending on the caller's path.
>> >
>> > Please respond if there are any concerns.
>>
>> fsleep() would be some improvement, but I think there is still the
>> issue
>> Loic points out where if delayus is small, but timeout_ms is large
>> (say
>> 50us and 25s), this function will end up burning a lot of cpu cycles
>> However that is likely an edge case, and if your poll cycle is that
>> small, I think it should be assumed that the event is expected to
>> happen
>> quickly, so the full timeout should not be hit.
>
> Well, my point is that during initial power_up, with a device
> cold-booting, it can take several seconds for it to reach ready state
> (not a corner case). That why timeout_ms can be as large as 20 seconds
> for mhi_pci_modem. If polling is based on busy-wait, that means the
> while loop will not let the CPU running anything else for several
> seconds. Not sure what is the expected meaning of this timeout_ms in
> first place... maybe I just use it badly.
>
> Moreover, do we need microsecond latency on detecting ready
> transition, this is not a critical path, right?
>
> Regards,
> Loic
At initial boot, yes, device could take longer to boot.
If we were to force caller to use an interval in the order of
milliseconds, I'd
still be using fsleep() internally anyway and just multiply the value by
1000
before passing it on as there's a need to check if the value is greater
than 20ms
or not.
I don't wish to reinvent the wheel and implement what we already have in
fsleep()
internally for msec.
It would be recommended that the caller specifies an interval of at
least 20+
msec but I don't think we can enforce that.
A point in favor of using microseconds is, if we were to expand usage of
this
API in the future for panic path to do a busywait, we wouldn't have to
change
the parameters.
3 options:
1. Use msec granularity and implement a partial fsleep for msec within
the new
API.
2. Use fsleep and leave it as usec granularity.
3. Leave it at usec, and add a busywait boolean allowing caller to
choose between
udelay() and fsleep() to also allow usage of this API in panic path (for
patch #3).
I like options 2 and 3. Hemant/Mani, your guidance is welcome.
Thanks,
Bhaumik
---
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora
Forum,
a Linux Foundation Collaborative Project