Subject: [PATCH RFC v1 3/3] PCI: qcom: ep: Add wake up host op to dw_pcie_ep_ops

Add wakeup host op to dw_pcie_ep_ops to wake up host from D3cold
or D3hot.

Signed-off-by: Krishna chaitanya chundru <[email protected]>
---
drivers/pci/controller/dwc/pcie-qcom-ep.c | 34 +++++++++++++++++++++++++++++++
1 file changed, 34 insertions(+)

diff --git a/drivers/pci/controller/dwc/pcie-qcom-ep.c b/drivers/pci/controller/dwc/pcie-qcom-ep.c
index 5d146ec..916a138 100644
--- a/drivers/pci/controller/dwc/pcie-qcom-ep.c
+++ b/drivers/pci/controller/dwc/pcie-qcom-ep.c
@@ -91,6 +91,7 @@
/* PARF_PM_CTRL register fields */
#define PARF_PM_CTRL_REQ_EXIT_L1 BIT(1)
#define PARF_PM_CTRL_READY_ENTR_L23 BIT(2)
+#define PARF_PM_CTRL_XMT_PME BIT(4)
#define PARF_PM_CTRL_REQ_NOT_ENTR_L1 BIT(5)

/* PARF_MHI_CLOCK_RESET_CTRL fields */
@@ -794,10 +795,43 @@ static void qcom_pcie_ep_init(struct dw_pcie_ep *ep)
dw_pcie_ep_reset_bar(pci, bar);
}

+static int qcom_pcie_ep_wakeup_host(struct dw_pcie_ep *ep, u8 func_no)
+{
+ struct dw_pcie *pci = to_dw_pcie_from_ep(ep);
+ struct qcom_pcie_ep *pcie_ep = to_pcie_ep(pci);
+ struct device *dev = pci->dev;
+ u32 perst, dstate, val;
+
+ perst = gpiod_get_value(pcie_ep->reset);
+ /* Toggle wake GPIO when device is in D3 cold */
+ if (perst) {
+ dev_info(dev, "Device is in D3 cold toggling wake\n");
+ gpiod_set_value_cansleep(pcie_ep->wake, 1);
+ usleep_range(WAKE_DELAY_US, WAKE_DELAY_US + 500);
+ gpiod_set_value_cansleep(pcie_ep->wake, 0);
+ return 0;
+ }
+
+ dstate = dw_pcie_readl_dbi(pci, DBI_CON_STATUS) &
+ DBI_CON_STATUS_POWER_STATE_MASK;
+ if (dstate == 3) {
+ dev_info(dev, "Device is in D3 hot sending inband PME\n");
+ val = readl_relaxed(pcie_ep->parf + PARF_PM_CTRL);
+ val |= PARF_PM_CTRL_XMT_PME;
+ writel_relaxed(val, pcie_ep->parf + PARF_PM_CTRL);
+ } else {
+ dev_err(dev, "Device is not in D3 state wakeup is not supported\n");
+ return -EPERM;
+ }
+
+ return 0;
+}
+
static const struct dw_pcie_ep_ops pci_ep_ops = {
.ep_init = qcom_pcie_ep_init,
.raise_irq = qcom_pcie_ep_raise_irq,
.get_features = qcom_pcie_epc_get_features,
+ .wakeup_host = qcom_pcie_ep_wakeup_host,
};

static int qcom_pcie_ep_probe(struct platform_device *pdev)
--
2.7.4



2023-06-23 06:29:26

by Manivannan Sadhasivam

[permalink] [raw]
Subject: Re: [PATCH RFC v1 3/3] PCI: qcom: ep: Add wake up host op to dw_pcie_ep_ops

On Wed, Jun 14, 2023 at 08:30:49PM +0530, Krishna chaitanya chundru wrote:
> Add wakeup host op to dw_pcie_ep_ops to wake up host from D3cold
> or D3hot.
>

Commit message should describe how the wakeup is implemented in the driver.

> Signed-off-by: Krishna chaitanya chundru <[email protected]>
> ---
> drivers/pci/controller/dwc/pcie-qcom-ep.c | 34 +++++++++++++++++++++++++++++++
> 1 file changed, 34 insertions(+)
>
> diff --git a/drivers/pci/controller/dwc/pcie-qcom-ep.c b/drivers/pci/controller/dwc/pcie-qcom-ep.c
> index 5d146ec..916a138 100644
> --- a/drivers/pci/controller/dwc/pcie-qcom-ep.c
> +++ b/drivers/pci/controller/dwc/pcie-qcom-ep.c
> @@ -91,6 +91,7 @@
> /* PARF_PM_CTRL register fields */
> #define PARF_PM_CTRL_REQ_EXIT_L1 BIT(1)
> #define PARF_PM_CTRL_READY_ENTR_L23 BIT(2)
> +#define PARF_PM_CTRL_XMT_PME BIT(4)
> #define PARF_PM_CTRL_REQ_NOT_ENTR_L1 BIT(5)
>
> /* PARF_MHI_CLOCK_RESET_CTRL fields */
> @@ -794,10 +795,43 @@ static void qcom_pcie_ep_init(struct dw_pcie_ep *ep)
> dw_pcie_ep_reset_bar(pci, bar);
> }
>
> +static int qcom_pcie_ep_wakeup_host(struct dw_pcie_ep *ep, u8 func_no)
> +{
> + struct dw_pcie *pci = to_dw_pcie_from_ep(ep);
> + struct qcom_pcie_ep *pcie_ep = to_pcie_ep(pci);
> + struct device *dev = pci->dev;
> + u32 perst, dstate, val;
> +
> + perst = gpiod_get_value(pcie_ep->reset);
> + /* Toggle wake GPIO when device is in D3 cold */
> + if (perst) {
> + dev_info(dev, "Device is in D3 cold toggling wake\n");

dev_dbg(). "Waking up the host by toggling WAKE#"

> + gpiod_set_value_cansleep(pcie_ep->wake, 1);

Waking a device from D3cold requires power-on sequence by the host and in the
presence of Vaux, the EPF should be prepared for that. In that case, the mode of
wakeup should be decided by the EPF driver. So the wakeup API should have an
argument to decide whether the wakeup is through PME or sideband WAKE#.

Also note that as per PCIe Spec 3.0, the devices can support PME generation from
D3cold provided that the Vaux is supplied to the device. I do not know if that
is supported by Qcom devices but API should honor the spec. So the wakeup
control should come from EPF driver as I suggested above.

> + usleep_range(WAKE_DELAY_US, WAKE_DELAY_US + 500);
> + gpiod_set_value_cansleep(pcie_ep->wake, 0);
> + return 0;
> + }
> +
> + dstate = dw_pcie_readl_dbi(pci, DBI_CON_STATUS) &
> + DBI_CON_STATUS_POWER_STATE_MASK;
> + if (dstate == 3) {
> + dev_info(dev, "Device is in D3 hot sending inband PME\n");

dev_dbg(). As I said, the device can sent PME from D3cold also. So the log could
be "Waking up the host using PME".

> + val = readl_relaxed(pcie_ep->parf + PARF_PM_CTRL);
> + val |= PARF_PM_CTRL_XMT_PME;
> + writel_relaxed(val, pcie_ep->parf + PARF_PM_CTRL);
> + } else {
> + dev_err(dev, "Device is not in D3 state wakeup is not supported\n");
> + return -EPERM;

-ENOTSUPP

- Mani

> + }
> +
> + return 0;
> +}
> +
> static const struct dw_pcie_ep_ops pci_ep_ops = {
> .ep_init = qcom_pcie_ep_init,
> .raise_irq = qcom_pcie_ep_raise_irq,
> .get_features = qcom_pcie_epc_get_features,
> + .wakeup_host = qcom_pcie_ep_wakeup_host,
> };
>
> static int qcom_pcie_ep_probe(struct platform_device *pdev)
> --
> 2.7.4
>

--
மணிவண்ணன் சதாசிவம்

Subject: Re: [PATCH RFC v1 3/3] PCI: qcom: ep: Add wake up host op to dw_pcie_ep_ops


On 6/23/2023 11:48 AM, Manivannan Sadhasivam wrote:
> On Wed, Jun 14, 2023 at 08:30:49PM +0530, Krishna chaitanya chundru wrote:
>> Add wakeup host op to dw_pcie_ep_ops to wake up host from D3cold
>> or D3hot.
>>
> Commit message should describe how the wakeup is implemented in the driver.
I will correct this in next series.
>
>> Signed-off-by: Krishna chaitanya chundru <[email protected]>
>> ---
>> drivers/pci/controller/dwc/pcie-qcom-ep.c | 34 +++++++++++++++++++++++++++++++
>> 1 file changed, 34 insertions(+)
>>
>> diff --git a/drivers/pci/controller/dwc/pcie-qcom-ep.c b/drivers/pci/controller/dwc/pcie-qcom-ep.c
>> index 5d146ec..916a138 100644
>> --- a/drivers/pci/controller/dwc/pcie-qcom-ep.c
>> +++ b/drivers/pci/controller/dwc/pcie-qcom-ep.c
>> @@ -91,6 +91,7 @@
>> /* PARF_PM_CTRL register fields */
>> #define PARF_PM_CTRL_REQ_EXIT_L1 BIT(1)
>> #define PARF_PM_CTRL_READY_ENTR_L23 BIT(2)
>> +#define PARF_PM_CTRL_XMT_PME BIT(4)
>> #define PARF_PM_CTRL_REQ_NOT_ENTR_L1 BIT(5)
>>
>> /* PARF_MHI_CLOCK_RESET_CTRL fields */
>> @@ -794,10 +795,43 @@ static void qcom_pcie_ep_init(struct dw_pcie_ep *ep)
>> dw_pcie_ep_reset_bar(pci, bar);
>> }
>>
>> +static int qcom_pcie_ep_wakeup_host(struct dw_pcie_ep *ep, u8 func_no)
>> +{
>> + struct dw_pcie *pci = to_dw_pcie_from_ep(ep);
>> + struct qcom_pcie_ep *pcie_ep = to_pcie_ep(pci);
>> + struct device *dev = pci->dev;
>> + u32 perst, dstate, val;
>> +
>> + perst = gpiod_get_value(pcie_ep->reset);
>> + /* Toggle wake GPIO when device is in D3 cold */
>> + if (perst) {
>> + dev_info(dev, "Device is in D3 cold toggling wake\n");
> dev_dbg(). "Waking up the host by toggling WAKE#"
>
>> + gpiod_set_value_cansleep(pcie_ep->wake, 1);
> Waking a device from D3cold requires power-on sequence by the host and in the
> presence of Vaux, the EPF should be prepared for that. In that case, the mode of
> wakeup should be decided by the EPF driver. So the wakeup API should have an
> argument to decide whether the wakeup is through PME or sideband WAKE#.
>
> Also note that as per PCIe Spec 3.0, the devices can support PME generation from
> D3cold provided that the Vaux is supplied to the device. I do not know if that
> is supported by Qcom devices but API should honor the spec. So the wakeup
> control should come from EPF driver as I suggested above.

I aggre with you, but how will EPF know the PCI device state is in
D3cold or D3hot.

And how the EPF knows whether Vaux is supported or not in D3cold?

If there is any existing mechanism can you please point me that.

FYI Qcom does not support vaux power in D3 cold.

>> + usleep_range(WAKE_DELAY_US, WAKE_DELAY_US + 500);
>> + gpiod_set_value_cansleep(pcie_ep->wake, 0);
>> + return 0;
>> + }
>> +
>> + dstate = dw_pcie_readl_dbi(pci, DBI_CON_STATUS) &
>> + DBI_CON_STATUS_POWER_STATE_MASK;
>> + if (dstate == 3) {
>> + dev_info(dev, "Device is in D3 hot sending inband PME\n");
> dev_dbg(). As I said, the device can sent PME from D3cold also. So the log could
> be "Waking up the host using PME".
>
>> + val = readl_relaxed(pcie_ep->parf + PARF_PM_CTRL);
>> + val |= PARF_PM_CTRL_XMT_PME;
>> + writel_relaxed(val, pcie_ep->parf + PARF_PM_CTRL);
>> + } else {
>> + dev_err(dev, "Device is not in D3 state wakeup is not supported\n");
>> + return -EPERM;
> -ENOTSUPP
>
> - Mani
>
>> + }
>> +
>> + return 0;
>> +}
>> +
>> static const struct dw_pcie_ep_ops pci_ep_ops = {
>> .ep_init = qcom_pcie_ep_init,
>> .raise_irq = qcom_pcie_ep_raise_irq,
>> .get_features = qcom_pcie_epc_get_features,
>> + .wakeup_host = qcom_pcie_ep_wakeup_host,
>> };
>>
>> static int qcom_pcie_ep_probe(struct platform_device *pdev)
>> --
>> 2.7.4
>>

2023-06-27 14:02:15

by Manivannan Sadhasivam

[permalink] [raw]
Subject: Re: [PATCH RFC v1 3/3] PCI: qcom: ep: Add wake up host op to dw_pcie_ep_ops

On Mon, Jun 26, 2023 at 07:18:49PM +0530, Krishna Chaitanya Chundru wrote:
>
> On 6/23/2023 11:48 AM, Manivannan Sadhasivam wrote:
> > On Wed, Jun 14, 2023 at 08:30:49PM +0530, Krishna chaitanya chundru wrote:
> > > Add wakeup host op to dw_pcie_ep_ops to wake up host from D3cold
> > > or D3hot.
> > >
> > Commit message should describe how the wakeup is implemented in the driver.
> I will correct this in next series.
> >
> > > Signed-off-by: Krishna chaitanya chundru <[email protected]>
> > > ---
> > > drivers/pci/controller/dwc/pcie-qcom-ep.c | 34 +++++++++++++++++++++++++++++++
> > > 1 file changed, 34 insertions(+)
> > >
> > > diff --git a/drivers/pci/controller/dwc/pcie-qcom-ep.c b/drivers/pci/controller/dwc/pcie-qcom-ep.c
> > > index 5d146ec..916a138 100644
> > > --- a/drivers/pci/controller/dwc/pcie-qcom-ep.c
> > > +++ b/drivers/pci/controller/dwc/pcie-qcom-ep.c
> > > @@ -91,6 +91,7 @@
> > > /* PARF_PM_CTRL register fields */
> > > #define PARF_PM_CTRL_REQ_EXIT_L1 BIT(1)
> > > #define PARF_PM_CTRL_READY_ENTR_L23 BIT(2)
> > > +#define PARF_PM_CTRL_XMT_PME BIT(4)
> > > #define PARF_PM_CTRL_REQ_NOT_ENTR_L1 BIT(5)
> > > /* PARF_MHI_CLOCK_RESET_CTRL fields */
> > > @@ -794,10 +795,43 @@ static void qcom_pcie_ep_init(struct dw_pcie_ep *ep)
> > > dw_pcie_ep_reset_bar(pci, bar);
> > > }
> > > +static int qcom_pcie_ep_wakeup_host(struct dw_pcie_ep *ep, u8 func_no)
> > > +{
> > > + struct dw_pcie *pci = to_dw_pcie_from_ep(ep);
> > > + struct qcom_pcie_ep *pcie_ep = to_pcie_ep(pci);
> > > + struct device *dev = pci->dev;
> > > + u32 perst, dstate, val;
> > > +
> > > + perst = gpiod_get_value(pcie_ep->reset);
> > > + /* Toggle wake GPIO when device is in D3 cold */
> > > + if (perst) {
> > > + dev_info(dev, "Device is in D3 cold toggling wake\n");
> > dev_dbg(). "Waking up the host by toggling WAKE#"
> >
> > > + gpiod_set_value_cansleep(pcie_ep->wake, 1);
> > Waking a device from D3cold requires power-on sequence by the host and in the
> > presence of Vaux, the EPF should be prepared for that. In that case, the mode of
> > wakeup should be decided by the EPF driver. So the wakeup API should have an
> > argument to decide whether the wakeup is through PME or sideband WAKE#.
> >
> > Also note that as per PCIe Spec 3.0, the devices can support PME generation from
> > D3cold provided that the Vaux is supplied to the device. I do not know if that
> > is supported by Qcom devices but API should honor the spec. So the wakeup
> > control should come from EPF driver as I suggested above.
>
> I aggre with you, but how will EPF know the PCI device state is in D3cold or
> D3hot.
>

We should add a notifier in the controller driver which signals EPF when it
receives the DState events.. Take a look at pci_epc_linkdown().

> And how the EPF knows whether Vaux is supported or not in D3cold?
>
> If there is any existing mechanism can you please point me that.
>
> FYI Qcom does not support vaux power in D3 cold.
>

If the EPF can trigger wakeup event during D3Cold then it means it is powered
by Vaux, isn't it?

- Mani

> > > + usleep_range(WAKE_DELAY_US, WAKE_DELAY_US + 500);
> > > + gpiod_set_value_cansleep(pcie_ep->wake, 0);
> > > + return 0;
> > > + }
> > > +
> > > + dstate = dw_pcie_readl_dbi(pci, DBI_CON_STATUS) &
> > > + DBI_CON_STATUS_POWER_STATE_MASK;
> > > + if (dstate == 3) {
> > > + dev_info(dev, "Device is in D3 hot sending inband PME\n");
> > dev_dbg(). As I said, the device can sent PME from D3cold also. So the log could
> > be "Waking up the host using PME".
> >
> > > + val = readl_relaxed(pcie_ep->parf + PARF_PM_CTRL);
> > > + val |= PARF_PM_CTRL_XMT_PME;
> > > + writel_relaxed(val, pcie_ep->parf + PARF_PM_CTRL);
> > > + } else {
> > > + dev_err(dev, "Device is not in D3 state wakeup is not supported\n");
> > > + return -EPERM;
> > -ENOTSUPP
> >
> > - Mani
> >
> > > + }
> > > +
> > > + return 0;
> > > +}
> > > +
> > > static const struct dw_pcie_ep_ops pci_ep_ops = {
> > > .ep_init = qcom_pcie_ep_init,
> > > .raise_irq = qcom_pcie_ep_raise_irq,
> > > .get_features = qcom_pcie_epc_get_features,
> > > + .wakeup_host = qcom_pcie_ep_wakeup_host,
> > > };
> > > static int qcom_pcie_ep_probe(struct platform_device *pdev)
> > > --
> > > 2.7.4
> > >

--
மணிவண்ணன் சதாசிவம்

Subject: Re: [PATCH RFC v1 3/3] PCI: qcom: ep: Add wake up host op to dw_pcie_ep_ops


On 6/27/2023 7:23 PM, Manivannan Sadhasivam wrote:
> On Mon, Jun 26, 2023 at 07:18:49PM +0530, Krishna Chaitanya Chundru wrote:
>> On 6/23/2023 11:48 AM, Manivannan Sadhasivam wrote:
>>> On Wed, Jun 14, 2023 at 08:30:49PM +0530, Krishna chaitanya chundru wrote:
>>>> Add wakeup host op to dw_pcie_ep_ops to wake up host from D3cold
>>>> or D3hot.
>>>>
>>> Commit message should describe how the wakeup is implemented in the driver.
>> I will correct this in next series.
>>>> Signed-off-by: Krishna chaitanya chundru <[email protected]>
>>>> ---
>>>> drivers/pci/controller/dwc/pcie-qcom-ep.c | 34 +++++++++++++++++++++++++++++++
>>>> 1 file changed, 34 insertions(+)
>>>>
>>>> diff --git a/drivers/pci/controller/dwc/pcie-qcom-ep.c b/drivers/pci/controller/dwc/pcie-qcom-ep.c
>>>> index 5d146ec..916a138 100644
>>>> --- a/drivers/pci/controller/dwc/pcie-qcom-ep.c
>>>> +++ b/drivers/pci/controller/dwc/pcie-qcom-ep.c
>>>> @@ -91,6 +91,7 @@
>>>> /* PARF_PM_CTRL register fields */
>>>> #define PARF_PM_CTRL_REQ_EXIT_L1 BIT(1)
>>>> #define PARF_PM_CTRL_READY_ENTR_L23 BIT(2)
>>>> +#define PARF_PM_CTRL_XMT_PME BIT(4)
>>>> #define PARF_PM_CTRL_REQ_NOT_ENTR_L1 BIT(5)
>>>> /* PARF_MHI_CLOCK_RESET_CTRL fields */
>>>> @@ -794,10 +795,43 @@ static void qcom_pcie_ep_init(struct dw_pcie_ep *ep)
>>>> dw_pcie_ep_reset_bar(pci, bar);
>>>> }
>>>> +static int qcom_pcie_ep_wakeup_host(struct dw_pcie_ep *ep, u8 func_no)
>>>> +{
>>>> + struct dw_pcie *pci = to_dw_pcie_from_ep(ep);
>>>> + struct qcom_pcie_ep *pcie_ep = to_pcie_ep(pci);
>>>> + struct device *dev = pci->dev;
>>>> + u32 perst, dstate, val;
>>>> +
>>>> + perst = gpiod_get_value(pcie_ep->reset);
>>>> + /* Toggle wake GPIO when device is in D3 cold */
>>>> + if (perst) {
>>>> + dev_info(dev, "Device is in D3 cold toggling wake\n");
>>> dev_dbg(). "Waking up the host by toggling WAKE#"
>>>
>>>> + gpiod_set_value_cansleep(pcie_ep->wake, 1);
>>> Waking a device from D3cold requires power-on sequence by the host and in the
>>> presence of Vaux, the EPF should be prepared for that. In that case, the mode of
>>> wakeup should be decided by the EPF driver. So the wakeup API should have an
>>> argument to decide whether the wakeup is through PME or sideband WAKE#.
>>>
>>> Also note that as per PCIe Spec 3.0, the devices can support PME generation from
>>> D3cold provided that the Vaux is supplied to the device. I do not know if that
>>> is supported by Qcom devices but API should honor the spec. So the wakeup
>>> control should come from EPF driver as I suggested above.
>> I aggre with you, but how will EPF know the PCI device state is in D3cold or
>> D3hot.
>>
> We should add a notifier in the controller driver which signals EPF when it
> receives the DState events.. Take a look at pci_epc_linkdown().
Ok I will add this kind of Dstate change event
>
>> And how the EPF knows whether Vaux is supported or not in D3cold?
>>
>> If there is any existing mechanism can you please point me that.
>>
>> FYI Qcom does not support vaux power in D3 cold.
>>
> If the EPF can trigger wakeup event during D3Cold then it means it is powered
> by Vaux, isn't it?
>
> - Mani

EPF needs to know that if the endpoint is getting vaux in D3cold or not
without this info

EPF doesn't know wheather to send toggle wake or send Inband PME right.

I mean EPF should have some  prior information wheather vaux is supplied
or not to decide

wheather to toggle wake or send in band PME.

-KC

>
>>>> + usleep_range(WAKE_DELAY_US, WAKE_DELAY_US + 500);
>>>> + gpiod_set_value_cansleep(pcie_ep->wake, 0);
>>>> + return 0;
>>>> + }
>>>> +
>>>> + dstate = dw_pcie_readl_dbi(pci, DBI_CON_STATUS) &
>>>> + DBI_CON_STATUS_POWER_STATE_MASK;
>>>> + if (dstate == 3) {
>>>> + dev_info(dev, "Device is in D3 hot sending inband PME\n");
>>> dev_dbg(). As I said, the device can sent PME from D3cold also. So the log could
>>> be "Waking up the host using PME".
>>>
>>>> + val = readl_relaxed(pcie_ep->parf + PARF_PM_CTRL);
>>>> + val |= PARF_PM_CTRL_XMT_PME;
>>>> + writel_relaxed(val, pcie_ep->parf + PARF_PM_CTRL);
>>>> + } else {
>>>> + dev_err(dev, "Device is not in D3 state wakeup is not supported\n");
>>>> + return -EPERM;
>>> -ENOTSUPP
>>>
>>> - Mani
>>>
>>>> + }
>>>> +
>>>> + return 0;
>>>> +}
>>>> +
>>>> static const struct dw_pcie_ep_ops pci_ep_ops = {
>>>> .ep_init = qcom_pcie_ep_init,
>>>> .raise_irq = qcom_pcie_ep_raise_irq,
>>>> .get_features = qcom_pcie_epc_get_features,
>>>> + .wakeup_host = qcom_pcie_ep_wakeup_host,
>>>> };
>>>> static int qcom_pcie_ep_probe(struct platform_device *pdev)
>>>> --
>>>> 2.7.4
>>>>

2023-06-28 08:37:14

by Manivannan Sadhasivam

[permalink] [raw]
Subject: Re: [PATCH RFC v1 3/3] PCI: qcom: ep: Add wake up host op to dw_pcie_ep_ops

On Wed, Jun 28, 2023 at 11:49:04AM +0530, Krishna Chaitanya Chundru wrote:
>
> On 6/28/2023 10:27 AM, Manivannan Sadhasivam wrote:
> > On Wed, Jun 28, 2023 at 08:01:45AM +0530, Krishna Chaitanya Chundru wrote:
> > > On 6/27/2023 7:23 PM, Manivannan Sadhasivam wrote:
> > > > On Mon, Jun 26, 2023 at 07:18:49PM +0530, Krishna Chaitanya Chundru wrote:
> > > > > On 6/23/2023 11:48 AM, Manivannan Sadhasivam wrote:
> > > > > > On Wed, Jun 14, 2023 at 08:30:49PM +0530, Krishna chaitanya chundru wrote:
> > > > > > > Add wakeup host op to dw_pcie_ep_ops to wake up host from D3cold
> > > > > > > or D3hot.
> > > > > > >
> > > > > > Commit message should describe how the wakeup is implemented in the driver.
> > > > > I will correct this in next series.
> > > > > > > Signed-off-by: Krishna chaitanya chundru <[email protected]>
> > > > > > > ---
> > > > > > > drivers/pci/controller/dwc/pcie-qcom-ep.c | 34 +++++++++++++++++++++++++++++++
> > > > > > > 1 file changed, 34 insertions(+)
> > > > > > >
> > > > > > > diff --git a/drivers/pci/controller/dwc/pcie-qcom-ep.c b/drivers/pci/controller/dwc/pcie-qcom-ep.c
> > > > > > > index 5d146ec..916a138 100644
> > > > > > > --- a/drivers/pci/controller/dwc/pcie-qcom-ep.c
> > > > > > > +++ b/drivers/pci/controller/dwc/pcie-qcom-ep.c
> > > > > > > @@ -91,6 +91,7 @@
> > > > > > > /* PARF_PM_CTRL register fields */
> > > > > > > #define PARF_PM_CTRL_REQ_EXIT_L1 BIT(1)
> > > > > > > #define PARF_PM_CTRL_READY_ENTR_L23 BIT(2)
> > > > > > > +#define PARF_PM_CTRL_XMT_PME BIT(4)
> > > > > > > #define PARF_PM_CTRL_REQ_NOT_ENTR_L1 BIT(5)
> > > > > > > /* PARF_MHI_CLOCK_RESET_CTRL fields */
> > > > > > > @@ -794,10 +795,43 @@ static void qcom_pcie_ep_init(struct dw_pcie_ep *ep)
> > > > > > > dw_pcie_ep_reset_bar(pci, bar);
> > > > > > > }
> > > > > > > +static int qcom_pcie_ep_wakeup_host(struct dw_pcie_ep *ep, u8 func_no)
> > > > > > > +{
> > > > > > > + struct dw_pcie *pci = to_dw_pcie_from_ep(ep);
> > > > > > > + struct qcom_pcie_ep *pcie_ep = to_pcie_ep(pci);
> > > > > > > + struct device *dev = pci->dev;
> > > > > > > + u32 perst, dstate, val;
> > > > > > > +
> > > > > > > + perst = gpiod_get_value(pcie_ep->reset);
> > > > > > > + /* Toggle wake GPIO when device is in D3 cold */
> > > > > > > + if (perst) {
> > > > > > > + dev_info(dev, "Device is in D3 cold toggling wake\n");
> > > > > > dev_dbg(). "Waking up the host by toggling WAKE#"
> > > > > >
> > > > > > > + gpiod_set_value_cansleep(pcie_ep->wake, 1);
> > > > > > Waking a device from D3cold requires power-on sequence by the host and in the
> > > > > > presence of Vaux, the EPF should be prepared for that. In that case, the mode of
> > > > > > wakeup should be decided by the EPF driver. So the wakeup API should have an
> > > > > > argument to decide whether the wakeup is through PME or sideband WAKE#.
> > > > > >
> > > > > > Also note that as per PCIe Spec 3.0, the devices can support PME generation from
> > > > > > D3cold provided that the Vaux is supplied to the device. I do not know if that
> > > > > > is supported by Qcom devices but API should honor the spec. So the wakeup
> > > > > > control should come from EPF driver as I suggested above.
> > > > > I aggre with you, but how will EPF know the PCI device state is in D3cold or
> > > > > D3hot.
> > > > >
> > > > We should add a notifier in the controller driver which signals EPF when it
> > > > receives the DState events.. Take a look at pci_epc_linkdown().
> > > Ok I will add this kind of Dstate change event
> > > > > And how the EPF knows whether Vaux is supported or not in D3cold?
> > > > >
> > > > > If there is any existing mechanism can you please point me that.
> > > > >
> > > > > FYI Qcom does not support vaux power in D3 cold.
> > > > >
> > > > If the EPF can trigger wakeup event during D3Cold then it means it is powered
> > > > by Vaux, isn't it?
> > > >
> > > > - Mani
> > > EPF needs to know that if the endpoint is getting vaux in D3cold or not
> > > without this info
> > >
> > > EPF doesn't know wheather to send toggle wake or send Inband PME right.
> > >
> > > I mean EPF should have some  prior information wheather vaux is supplied or
> > > not to decide
> > >
> > > wheather to toggle wake or send in band PME.
> > >
> > Controller driver can use the #PERST level to detect D3Cold. Then it can pass
> > that info to EPF over notifiers. EPF may decide whether to toggle #WAKE or
> > send PME based on its configuration. For MHI EPF, it can toggle #WAKE as PME
> > during D3Cold is not supported.
> >
> > - Mani
>
> Can we add a new variable in ep features to know whether vaux is supported
> in D3 cold
>
> or is there any register in the config space of EP which indicates that vaux
> is supported or not.
>

I'm not aware of any method as of now but you do not need to worry. I was
just pointing it out as a possibility as per the spec. For this series, just
expose the wakeup logic to EPF as a flag.

- Mani

> >
> > > -KC
> > >
> > > > > > > + usleep_range(WAKE_DELAY_US, WAKE_DELAY_US + 500);
> > > > > > > + gpiod_set_value_cansleep(pcie_ep->wake, 0);
> > > > > > > + return 0;
> > > > > > > + }
> > > > > > > +
> > > > > > > + dstate = dw_pcie_readl_dbi(pci, DBI_CON_STATUS) &
> > > > > > > + DBI_CON_STATUS_POWER_STATE_MASK;
> > > > > > > + if (dstate == 3) {
> > > > > > > + dev_info(dev, "Device is in D3 hot sending inband PME\n");
> > > > > > dev_dbg(). As I said, the device can sent PME from D3cold also. So the log could
> > > > > > be "Waking up the host using PME".
> > > > > >
> > > > > > > + val = readl_relaxed(pcie_ep->parf + PARF_PM_CTRL);
> > > > > > > + val |= PARF_PM_CTRL_XMT_PME;
> > > > > > > + writel_relaxed(val, pcie_ep->parf + PARF_PM_CTRL);
> > > > > > > + } else {
> > > > > > > + dev_err(dev, "Device is not in D3 state wakeup is not supported\n");
> > > > > > > + return -EPERM;
> > > > > > -ENOTSUPP
> > > > > >
> > > > > > - Mani
> > > > > >
> > > > > > > + }
> > > > > > > +
> > > > > > > + return 0;
> > > > > > > +}
> > > > > > > +
> > > > > > > static const struct dw_pcie_ep_ops pci_ep_ops = {
> > > > > > > .ep_init = qcom_pcie_ep_init,
> > > > > > > .raise_irq = qcom_pcie_ep_raise_irq,
> > > > > > > .get_features = qcom_pcie_epc_get_features,
> > > > > > > + .wakeup_host = qcom_pcie_ep_wakeup_host,
> > > > > > > };
> > > > > > > static int qcom_pcie_ep_probe(struct platform_device *pdev)
> > > > > > > --
> > > > > > > 2.7.4
> > > > > > >

--
மணிவண்ணன் சதாசிவம்

Subject: Re: [PATCH RFC v1 3/3] PCI: qcom: ep: Add wake up host op to dw_pcie_ep_ops


On 6/28/2023 10:27 AM, Manivannan Sadhasivam wrote:
> On Wed, Jun 28, 2023 at 08:01:45AM +0530, Krishna Chaitanya Chundru wrote:
>> On 6/27/2023 7:23 PM, Manivannan Sadhasivam wrote:
>>> On Mon, Jun 26, 2023 at 07:18:49PM +0530, Krishna Chaitanya Chundru wrote:
>>>> On 6/23/2023 11:48 AM, Manivannan Sadhasivam wrote:
>>>>> On Wed, Jun 14, 2023 at 08:30:49PM +0530, Krishna chaitanya chundru wrote:
>>>>>> Add wakeup host op to dw_pcie_ep_ops to wake up host from D3cold
>>>>>> or D3hot.
>>>>>>
>>>>> Commit message should describe how the wakeup is implemented in the driver.
>>>> I will correct this in next series.
>>>>>> Signed-off-by: Krishna chaitanya chundru <[email protected]>
>>>>>> ---
>>>>>> drivers/pci/controller/dwc/pcie-qcom-ep.c | 34 +++++++++++++++++++++++++++++++
>>>>>> 1 file changed, 34 insertions(+)
>>>>>>
>>>>>> diff --git a/drivers/pci/controller/dwc/pcie-qcom-ep.c b/drivers/pci/controller/dwc/pcie-qcom-ep.c
>>>>>> index 5d146ec..916a138 100644
>>>>>> --- a/drivers/pci/controller/dwc/pcie-qcom-ep.c
>>>>>> +++ b/drivers/pci/controller/dwc/pcie-qcom-ep.c
>>>>>> @@ -91,6 +91,7 @@
>>>>>> /* PARF_PM_CTRL register fields */
>>>>>> #define PARF_PM_CTRL_REQ_EXIT_L1 BIT(1)
>>>>>> #define PARF_PM_CTRL_READY_ENTR_L23 BIT(2)
>>>>>> +#define PARF_PM_CTRL_XMT_PME BIT(4)
>>>>>> #define PARF_PM_CTRL_REQ_NOT_ENTR_L1 BIT(5)
>>>>>> /* PARF_MHI_CLOCK_RESET_CTRL fields */
>>>>>> @@ -794,10 +795,43 @@ static void qcom_pcie_ep_init(struct dw_pcie_ep *ep)
>>>>>> dw_pcie_ep_reset_bar(pci, bar);
>>>>>> }
>>>>>> +static int qcom_pcie_ep_wakeup_host(struct dw_pcie_ep *ep, u8 func_no)
>>>>>> +{
>>>>>> + struct dw_pcie *pci = to_dw_pcie_from_ep(ep);
>>>>>> + struct qcom_pcie_ep *pcie_ep = to_pcie_ep(pci);
>>>>>> + struct device *dev = pci->dev;
>>>>>> + u32 perst, dstate, val;
>>>>>> +
>>>>>> + perst = gpiod_get_value(pcie_ep->reset);
>>>>>> + /* Toggle wake GPIO when device is in D3 cold */
>>>>>> + if (perst) {
>>>>>> + dev_info(dev, "Device is in D3 cold toggling wake\n");
>>>>> dev_dbg(). "Waking up the host by toggling WAKE#"
>>>>>
>>>>>> + gpiod_set_value_cansleep(pcie_ep->wake, 1);
>>>>> Waking a device from D3cold requires power-on sequence by the host and in the
>>>>> presence of Vaux, the EPF should be prepared for that. In that case, the mode of
>>>>> wakeup should be decided by the EPF driver. So the wakeup API should have an
>>>>> argument to decide whether the wakeup is through PME or sideband WAKE#.
>>>>>
>>>>> Also note that as per PCIe Spec 3.0, the devices can support PME generation from
>>>>> D3cold provided that the Vaux is supplied to the device. I do not know if that
>>>>> is supported by Qcom devices but API should honor the spec. So the wakeup
>>>>> control should come from EPF driver as I suggested above.
>>>> I aggre with you, but how will EPF know the PCI device state is in D3cold or
>>>> D3hot.
>>>>
>>> We should add a notifier in the controller driver which signals EPF when it
>>> receives the DState events.. Take a look at pci_epc_linkdown().
>> Ok I will add this kind of Dstate change event
>>>> And how the EPF knows whether Vaux is supported or not in D3cold?
>>>>
>>>> If there is any existing mechanism can you please point me that.
>>>>
>>>> FYI Qcom does not support vaux power in D3 cold.
>>>>
>>> If the EPF can trigger wakeup event during D3Cold then it means it is powered
>>> by Vaux, isn't it?
>>>
>>> - Mani
>> EPF needs to know that if the endpoint is getting vaux in D3cold or not
>> without this info
>>
>> EPF doesn't know wheather to send toggle wake or send Inband PME right.
>>
>> I mean EPF should have some  prior information wheather vaux is supplied or
>> not to decide
>>
>> wheather to toggle wake or send in band PME.
>>
> Controller driver can use the #PERST level to detect D3Cold. Then it can pass
> that info to EPF over notifiers. EPF may decide whether to toggle #WAKE or
> send PME based on its configuration. For MHI EPF, it can toggle #WAKE as PME
> during D3Cold is not supported.
>
> - Mani

Can we add a new variable in ep features to know whether vaux is
supported in D3 cold

or is there any register in the config space of EP which indicates that
vaux is supported or not.

>
>> -KC
>>
>>>>>> + usleep_range(WAKE_DELAY_US, WAKE_DELAY_US + 500);
>>>>>> + gpiod_set_value_cansleep(pcie_ep->wake, 0);
>>>>>> + return 0;
>>>>>> + }
>>>>>> +
>>>>>> + dstate = dw_pcie_readl_dbi(pci, DBI_CON_STATUS) &
>>>>>> + DBI_CON_STATUS_POWER_STATE_MASK;
>>>>>> + if (dstate == 3) {
>>>>>> + dev_info(dev, "Device is in D3 hot sending inband PME\n");
>>>>> dev_dbg(). As I said, the device can sent PME from D3cold also. So the log could
>>>>> be "Waking up the host using PME".
>>>>>
>>>>>> + val = readl_relaxed(pcie_ep->parf + PARF_PM_CTRL);
>>>>>> + val |= PARF_PM_CTRL_XMT_PME;
>>>>>> + writel_relaxed(val, pcie_ep->parf + PARF_PM_CTRL);
>>>>>> + } else {
>>>>>> + dev_err(dev, "Device is not in D3 state wakeup is not supported\n");
>>>>>> + return -EPERM;
>>>>> -ENOTSUPP
>>>>>
>>>>> - Mani
>>>>>
>>>>>> + }
>>>>>> +
>>>>>> + return 0;
>>>>>> +}
>>>>>> +
>>>>>> static const struct dw_pcie_ep_ops pci_ep_ops = {
>>>>>> .ep_init = qcom_pcie_ep_init,
>>>>>> .raise_irq = qcom_pcie_ep_raise_irq,
>>>>>> .get_features = qcom_pcie_epc_get_features,
>>>>>> + .wakeup_host = qcom_pcie_ep_wakeup_host,
>>>>>> };
>>>>>> static int qcom_pcie_ep_probe(struct platform_device *pdev)
>>>>>> --
>>>>>> 2.7.4
>>>>>>

2023-06-28 09:29:21

by Manivannan Sadhasivam

[permalink] [raw]
Subject: Re: [PATCH RFC v1 3/3] PCI: qcom: ep: Add wake up host op to dw_pcie_ep_ops

On Wed, Jun 28, 2023 at 08:01:45AM +0530, Krishna Chaitanya Chundru wrote:
>
> On 6/27/2023 7:23 PM, Manivannan Sadhasivam wrote:
> > On Mon, Jun 26, 2023 at 07:18:49PM +0530, Krishna Chaitanya Chundru wrote:
> > > On 6/23/2023 11:48 AM, Manivannan Sadhasivam wrote:
> > > > On Wed, Jun 14, 2023 at 08:30:49PM +0530, Krishna chaitanya chundru wrote:
> > > > > Add wakeup host op to dw_pcie_ep_ops to wake up host from D3cold
> > > > > or D3hot.
> > > > >
> > > > Commit message should describe how the wakeup is implemented in the driver.
> > > I will correct this in next series.
> > > > > Signed-off-by: Krishna chaitanya chundru <[email protected]>
> > > > > ---
> > > > > drivers/pci/controller/dwc/pcie-qcom-ep.c | 34 +++++++++++++++++++++++++++++++
> > > > > 1 file changed, 34 insertions(+)
> > > > >
> > > > > diff --git a/drivers/pci/controller/dwc/pcie-qcom-ep.c b/drivers/pci/controller/dwc/pcie-qcom-ep.c
> > > > > index 5d146ec..916a138 100644
> > > > > --- a/drivers/pci/controller/dwc/pcie-qcom-ep.c
> > > > > +++ b/drivers/pci/controller/dwc/pcie-qcom-ep.c
> > > > > @@ -91,6 +91,7 @@
> > > > > /* PARF_PM_CTRL register fields */
> > > > > #define PARF_PM_CTRL_REQ_EXIT_L1 BIT(1)
> > > > > #define PARF_PM_CTRL_READY_ENTR_L23 BIT(2)
> > > > > +#define PARF_PM_CTRL_XMT_PME BIT(4)
> > > > > #define PARF_PM_CTRL_REQ_NOT_ENTR_L1 BIT(5)
> > > > > /* PARF_MHI_CLOCK_RESET_CTRL fields */
> > > > > @@ -794,10 +795,43 @@ static void qcom_pcie_ep_init(struct dw_pcie_ep *ep)
> > > > > dw_pcie_ep_reset_bar(pci, bar);
> > > > > }
> > > > > +static int qcom_pcie_ep_wakeup_host(struct dw_pcie_ep *ep, u8 func_no)
> > > > > +{
> > > > > + struct dw_pcie *pci = to_dw_pcie_from_ep(ep);
> > > > > + struct qcom_pcie_ep *pcie_ep = to_pcie_ep(pci);
> > > > > + struct device *dev = pci->dev;
> > > > > + u32 perst, dstate, val;
> > > > > +
> > > > > + perst = gpiod_get_value(pcie_ep->reset);
> > > > > + /* Toggle wake GPIO when device is in D3 cold */
> > > > > + if (perst) {
> > > > > + dev_info(dev, "Device is in D3 cold toggling wake\n");
> > > > dev_dbg(). "Waking up the host by toggling WAKE#"
> > > >
> > > > > + gpiod_set_value_cansleep(pcie_ep->wake, 1);
> > > > Waking a device from D3cold requires power-on sequence by the host and in the
> > > > presence of Vaux, the EPF should be prepared for that. In that case, the mode of
> > > > wakeup should be decided by the EPF driver. So the wakeup API should have an
> > > > argument to decide whether the wakeup is through PME or sideband WAKE#.
> > > >
> > > > Also note that as per PCIe Spec 3.0, the devices can support PME generation from
> > > > D3cold provided that the Vaux is supplied to the device. I do not know if that
> > > > is supported by Qcom devices but API should honor the spec. So the wakeup
> > > > control should come from EPF driver as I suggested above.
> > > I aggre with you, but how will EPF know the PCI device state is in D3cold or
> > > D3hot.
> > >
> > We should add a notifier in the controller driver which signals EPF when it
> > receives the DState events.. Take a look at pci_epc_linkdown().
> Ok I will add this kind of Dstate change event
> >
> > > And how the EPF knows whether Vaux is supported or not in D3cold?
> > >
> > > If there is any existing mechanism can you please point me that.
> > >
> > > FYI Qcom does not support vaux power in D3 cold.
> > >
> > If the EPF can trigger wakeup event during D3Cold then it means it is powered
> > by Vaux, isn't it?
> >
> > - Mani
>
> EPF needs to know that if the endpoint is getting vaux in D3cold or not
> without this info
>
> EPF doesn't know wheather to send toggle wake or send Inband PME right.
>
> I mean EPF should have some  prior information wheather vaux is supplied or
> not to decide
>
> wheather to toggle wake or send in band PME.
>

Controller driver can use the #PERST level to detect D3Cold. Then it can pass
that info to EPF over notifiers. EPF may decide whether to toggle #WAKE or
send PME based on its configuration. For MHI EPF, it can toggle #WAKE as PME
during D3Cold is not supported.

- Mani

> -KC
>
> >
> > > > > + usleep_range(WAKE_DELAY_US, WAKE_DELAY_US + 500);
> > > > > + gpiod_set_value_cansleep(pcie_ep->wake, 0);
> > > > > + return 0;
> > > > > + }
> > > > > +
> > > > > + dstate = dw_pcie_readl_dbi(pci, DBI_CON_STATUS) &
> > > > > + DBI_CON_STATUS_POWER_STATE_MASK;
> > > > > + if (dstate == 3) {
> > > > > + dev_info(dev, "Device is in D3 hot sending inband PME\n");
> > > > dev_dbg(). As I said, the device can sent PME from D3cold also. So the log could
> > > > be "Waking up the host using PME".
> > > >
> > > > > + val = readl_relaxed(pcie_ep->parf + PARF_PM_CTRL);
> > > > > + val |= PARF_PM_CTRL_XMT_PME;
> > > > > + writel_relaxed(val, pcie_ep->parf + PARF_PM_CTRL);
> > > > > + } else {
> > > > > + dev_err(dev, "Device is not in D3 state wakeup is not supported\n");
> > > > > + return -EPERM;
> > > > -ENOTSUPP
> > > >
> > > > - Mani
> > > >
> > > > > + }
> > > > > +
> > > > > + return 0;
> > > > > +}
> > > > > +
> > > > > static const struct dw_pcie_ep_ops pci_ep_ops = {
> > > > > .ep_init = qcom_pcie_ep_init,
> > > > > .raise_irq = qcom_pcie_ep_raise_irq,
> > > > > .get_features = qcom_pcie_epc_get_features,
> > > > > + .wakeup_host = qcom_pcie_ep_wakeup_host,
> > > > > };
> > > > > static int qcom_pcie_ep_probe(struct platform_device *pdev)
> > > > > --
> > > > > 2.7.4
> > > > >

--
மணிவண்ணன் சதாசிவம்