Re: [PATCH] PCI: Apply mandatory recovery delay on return from D3cold

Re: [PATCH] PCI: Apply mandatory recovery delay on return from D3cold

[Bjorn Helgaas]

[+cc Rafael, linux-pm]

On Wed, Jul 08, 2026 at 10:26:50AM -0500, Mario Limonciello wrote:
> Per the "PCI Bus Power Management Interface Specification", rev. 1.2,
> sec. 5.4, there is a minimum recovery time between programming a function
> from D3 to D0 and accessing it.  The spec does not limit this to D3hot; it
> applies to the D3cold to D0 transition as well.
> 
> pci_power_up() only honors this delay on the D3hot branch.  When a device
> returns from D3cold, platform_pci_set_power_state() has already restored
> main power before PCI_PM_CTRL is read, so the state read from the register
> is D0 and the transition delay block is skipped by the
> 
> 	if (state == PCI_D0)
> 		goto end;
> 
> early return.  The register value is masked with PCI_PM_CTRL_STATE_MASK
> and cannot represent D3cold, so only dev->current_state still reflects the
> D3cold origin at this point.
> 
> Apply the delay based on dev->current_state, ahead of the early return, so
> it takes effect on the D3cold to D0 path before the device is accessed.
> Use the device's d3cold_delay, which the platform may tune via _DSM and
> quirks may raise, rather than the D3hot delay.
> 
> To keep the existing D3hot callers unchanged, pci_dev_d3_sleep() now takes
> the delay in milliseconds and a pci_dev_d3hot_sleep() wrapper supplies the
> D3hot delay as before.
> 
> Reported-by: mrh@frame.work
> Closes: https://bugzilla.kernel.org/show_bug.cgi?id=221073
> Signed-off-by: Mario Limonciello <mario.limonciello@amd.com>
> ---
> Cc: mrh@frame.work
> Cc: stern@rowland.harvard.edu
> Cc: hannes@vonhaugwitz.com
> Cc: jase_harley@protonmail.com
> Cc: superveridical@gmail.com
> Cc: david.c.hubbard@gmail.com
> Cc: bugzilla@logical.ink
> Cc: michal.pecio@gmail.com
> ---
>  drivers/pci/pci.c | 26 ++++++++++++++++++++------
>  1 file changed, 20 insertions(+), 6 deletions(-)
> 
> diff --git a/drivers/pci/pci.c b/drivers/pci/pci.c
> index 77b17b13ee615..e09cfb28fe61c 100644
> --- a/drivers/pci/pci.c
> +++ b/drivers/pci/pci.c
> @@ -81,9 +81,8 @@ struct pci_pme_device {
>   */
>  #define PCIE_RESET_READY_POLL_MS 60000 /* msec */
>  
> -static void pci_dev_d3_sleep(struct pci_dev *dev)
> +static void pci_dev_d3_sleep(unsigned int delay_ms)
>  {
> -	unsigned int delay_ms = max(dev->d3hot_delay, pci_pm_d3hot_delay);
>  	unsigned int upper;
>  
>  	if (delay_ms) {
> @@ -94,6 +93,11 @@ static void pci_dev_d3_sleep(struct pci_dev *dev)
>  	}
>  }
>  
> +static void pci_dev_d3hot_sleep(struct pci_dev *dev)
> +{
> +	pci_dev_d3_sleep(max(dev->d3hot_delay, pci_pm_d3hot_delay));
> +}
> +
>  bool pci_reset_supported(struct pci_dev *dev)
>  {
>  	return dev->reset_methods[0] != 0;
> @@ -1333,6 +1337,16 @@ int pci_power_up(struct pci_dev *dev)
>  	need_restore = (state == PCI_D3hot || dev->current_state >= PCI_D3hot) &&
>  			!(pmcsr & PCI_PM_CTRL_NO_SOFT_RESET);
>  
> +	/*
> +	 * A device returning from D3cold has already been powered back on by
> +	 * platform_pci_set_power_state() above, so PCI_PM_CTRL now reads back
> +	 * as D0 and the transition delays below are skipped.  PCI PM 1.2 still
> +	 * requires a minimum recovery time on the D3 to D0 transition, so apply
> +	 * the device's D3cold recovery delay here before it is accessed.
> +	 */
> +	if (dev->current_state == PCI_D3cold)
> +		pci_dev_d3_sleep(dev->d3cold_delay);

My understanding is that the "mandatory transition delays" below only
cover a transition caused by the write to PCI_PM_CTRL, which would be
from D1, D2, or D3hot to D0.

If the device started in D3cold, platform_pci_set_power_state(PCI_D0)
should transition it to D0uninitialized *and* take care of any
required delays [1].  The device should be Configuration-Ready upon
return (and we've already read PCI_PM_CTRL above, and the read
returned something other than PCI_ERROR_RESPONSE).

But evidently adding more delay does make a difference, even though
the config read of PCI_PM_CTRL seemed successful.  I guess it's
conceivable that platform AML doesn't wait quite long enough, although
it does seem to affect more than one platform (AMD Strix Halo,
Framework Desktop/AMD Ryzen AI Max 300, Lenovo ThinkPad T14 Gen 6 AMD
Ryzen AI 7 Pro 350), and I *assume* the issue doesn't happen under
Windows?

[1] https://lore.kernel.org/linux-pci/CAJZ5v0iZN5NtUztqe=MxCRcXdBaaqzZ749OqSUkadwwBy0ugUQ@mail.gmail.com/

>  	if (state == PCI_D0)
>  		goto end;
>  
> @@ -1344,7 +1358,7 @@ int pci_power_up(struct pci_dev *dev)
>  
>  	/* Mandatory transition delays; see PCI PM 1.2. */
>  	if (state == PCI_D3hot) {
> -		pci_dev_d3_sleep(dev);
> +		pci_dev_d3hot_sleep(dev);
>  		if (!(pmcsr & PCI_PM_CTRL_NO_SOFT_RESET)) {
>  			ret = pci_dev_wait(dev, "power up D3hot->D0uninitialized",
>  					   PCIE_RESET_READY_POLL_MS);
> @@ -1514,7 +1528,7 @@ static int pci_set_low_power_state(struct pci_dev *dev, pci_power_t state, bool
>  
>  	/* Mandatory power management transition delays; see PCI PM 1.2. */
>  	if (state == PCI_D3hot)
> -		pci_dev_d3_sleep(dev);
> +		pci_dev_d3hot_sleep(dev);
>  	else if (state == PCI_D2)
>  		udelay(PCI_PM_D2_DELAY);
>  
> @@ -4511,12 +4525,12 @@ static int pci_pm_reset(struct pci_dev *dev, bool probe)
>  	csr &= ~PCI_PM_CTRL_STATE_MASK;
>  	csr |= PCI_D3hot;
>  	pci_write_config_word(dev, dev->pm_cap + PCI_PM_CTRL, csr);
> -	pci_dev_d3_sleep(dev);
> +	pci_dev_d3hot_sleep(dev);
>  
>  	csr &= ~PCI_PM_CTRL_STATE_MASK;
>  	csr |= PCI_D0;
>  	pci_write_config_word(dev, dev->pm_cap + PCI_PM_CTRL, csr);
> -	pci_dev_d3_sleep(dev);
> +	pci_dev_d3hot_sleep(dev);
>  
>  	ret = pci_dev_wait(dev, "PM D3hot->D0", PCIE_RESET_READY_POLL_MS);
>  	pci_dev_reset_iommu_done(dev);
> -- 
> 2.43.0
> 

Re: [PATCH] PCI: Apply mandatory recovery delay on return from D3cold

[Mario Limonciello]

On 7/10/26 15:16, Bjorn Helgaas wrote:
> [+cc Rafael, linux-pm]
> 
> On Wed, Jul 08, 2026 at 10:26:50AM -0500, Mario Limonciello wrote:
>> Per the "PCI Bus Power Management Interface Specification", rev. 1.2,
>> sec. 5.4, there is a minimum recovery time between programming a function
>> from D3 to D0 and accessing it.  The spec does not limit this to D3hot; it
>> applies to the D3cold to D0 transition as well.
>>
>> pci_power_up() only honors this delay on the D3hot branch.  When a device
>> returns from D3cold, platform_pci_set_power_state() has already restored
>> main power before PCI_PM_CTRL is read, so the state read from the register
>> is D0 and the transition delay block is skipped by the
>>
>> 	if (state == PCI_D0)
>> 		goto end;
>>
>> early return.  The register value is masked with PCI_PM_CTRL_STATE_MASK
>> and cannot represent D3cold, so only dev->current_state still reflects the
>> D3cold origin at this point.
>>
>> Apply the delay based on dev->current_state, ahead of the early return, so
>> it takes effect on the D3cold to D0 path before the device is accessed.
>> Use the device's d3cold_delay, which the platform may tune via _DSM and
>> quirks may raise, rather than the D3hot delay.
>>
>> To keep the existing D3hot callers unchanged, pci_dev_d3_sleep() now takes
>> the delay in milliseconds and a pci_dev_d3hot_sleep() wrapper supplies the
>> D3hot delay as before.
>>
>> Reported-by: mrh@frame.work
>> Closes: https://bugzilla.kernel.org/show_bug.cgi?id=221073
>> Signed-off-by: Mario Limonciello <mario.limonciello@amd.com>
>> ---
>> Cc: mrh@frame.work
>> Cc: stern@rowland.harvard.edu
>> Cc: hannes@vonhaugwitz.com
>> Cc: jase_harley@protonmail.com
>> Cc: superveridical@gmail.com
>> Cc: david.c.hubbard@gmail.com
>> Cc: bugzilla@logical.ink
>> Cc: michal.pecio@gmail.com
>> ---
>>   drivers/pci/pci.c | 26 ++++++++++++++++++++------
>>   1 file changed, 20 insertions(+), 6 deletions(-)
>>
>> diff --git a/drivers/pci/pci.c b/drivers/pci/pci.c
>> index 77b17b13ee615..e09cfb28fe61c 100644
>> --- a/drivers/pci/pci.c
>> +++ b/drivers/pci/pci.c
>> @@ -81,9 +81,8 @@ struct pci_pme_device {
>>    */
>>   #define PCIE_RESET_READY_POLL_MS 60000 /* msec */
>>   
>> -static void pci_dev_d3_sleep(struct pci_dev *dev)
>> +static void pci_dev_d3_sleep(unsigned int delay_ms)
>>   {
>> -	unsigned int delay_ms = max(dev->d3hot_delay, pci_pm_d3hot_delay);
>>   	unsigned int upper;
>>   
>>   	if (delay_ms) {
>> @@ -94,6 +93,11 @@ static void pci_dev_d3_sleep(struct pci_dev *dev)
>>   	}
>>   }
>>   
>> +static void pci_dev_d3hot_sleep(struct pci_dev *dev)
>> +{
>> +	pci_dev_d3_sleep(max(dev->d3hot_delay, pci_pm_d3hot_delay));
>> +}
>> +
>>   bool pci_reset_supported(struct pci_dev *dev)
>>   {
>>   	return dev->reset_methods[0] != 0;
>> @@ -1333,6 +1337,16 @@ int pci_power_up(struct pci_dev *dev)
>>   	need_restore = (state == PCI_D3hot || dev->current_state >= PCI_D3hot) &&
>>   			!(pmcsr & PCI_PM_CTRL_NO_SOFT_RESET);
>>   
>> +	/*
>> +	 * A device returning from D3cold has already been powered back on by
>> +	 * platform_pci_set_power_state() above, so PCI_PM_CTRL now reads back
>> +	 * as D0 and the transition delays below are skipped.  PCI PM 1.2 still
>> +	 * requires a minimum recovery time on the D3 to D0 transition, so apply
>> +	 * the device's D3cold recovery delay here before it is accessed.
>> +	 */
>> +	if (dev->current_state == PCI_D3cold)
>> +		pci_dev_d3_sleep(dev->d3cold_delay);
> 
> My understanding is that the "mandatory transition delays" below only
> cover a transition caused by the write to PCI_PM_CTRL, which would be
> from D1, D2, or D3hot to D0.
> 
> If the device started in D3cold, platform_pci_set_power_state(PCI_D0)
> should transition it to D0uninitialized *and* take care of any
> required delays [1].  The device should be Configuration-Ready upon
> return (and we've already read PCI_PM_CTRL above, and the read
> returned something other than PCI_ERROR_RESPONSE).

FWIW - I was wondering if we shouldn't be reading PCI_PM_CTRL until the 
mandatory delay too.

> 
> But evidently adding more delay does make a difference, even though
> the config read of PCI_PM_CTRL seemed successful.  I guess it's
> conceivable that platform AML doesn't wait quite long enough, although
> it does seem to affect more than one platform (AMD Strix Halo,
> Framework Desktop/AMD Ryzen AI Max 300, Lenovo ThinkPad T14 Gen 6 AMD
> Ryzen AI 7 Pro 350), and I *assume* the issue doesn't happen under
> Windows?

I don't want to make any rash decisions until we have more testing 
evidence than one person testing 4 cycles.  I would like more of the 
people who reported this to confirm the patches really help and it 
wasn't just luck in those 4 cycles for that 1 person.

The part that still doesn't sit well with me is this behavior being tied 
to the VRAM size.  I don't feel these should be connected.

> 
> [1] https://lore.kernel.org/linux-pci/CAJZ5v0iZN5NtUztqe=MxCRcXdBaaqzZ749OqSUkadwwBy0ugUQ@mail.gmail.com/
> 
>>   	if (state == PCI_D0)
>>   		goto end;
>>   
>> @@ -1344,7 +1358,7 @@ int pci_power_up(struct pci_dev *dev)
>>   
>>   	/* Mandatory transition delays; see PCI PM 1.2. */
>>   	if (state == PCI_D3hot) {
>> -		pci_dev_d3_sleep(dev);
>> +		pci_dev_d3hot_sleep(dev);
>>   		if (!(pmcsr & PCI_PM_CTRL_NO_SOFT_RESET)) {
>>   			ret = pci_dev_wait(dev, "power up D3hot->D0uninitialized",
>>   					   PCIE_RESET_READY_POLL_MS);
>> @@ -1514,7 +1528,7 @@ static int pci_set_low_power_state(struct pci_dev *dev, pci_power_t state, bool
>>   
>>   	/* Mandatory power management transition delays; see PCI PM 1.2. */
>>   	if (state == PCI_D3hot)
>> -		pci_dev_d3_sleep(dev);
>> +		pci_dev_d3hot_sleep(dev);
>>   	else if (state == PCI_D2)
>>   		udelay(PCI_PM_D2_DELAY);
>>   
>> @@ -4511,12 +4525,12 @@ static int pci_pm_reset(struct pci_dev *dev, bool probe)
>>   	csr &= ~PCI_PM_CTRL_STATE_MASK;
>>   	csr |= PCI_D3hot;
>>   	pci_write_config_word(dev, dev->pm_cap + PCI_PM_CTRL, csr);
>> -	pci_dev_d3_sleep(dev);
>> +	pci_dev_d3hot_sleep(dev);
>>   
>>   	csr &= ~PCI_PM_CTRL_STATE_MASK;
>>   	csr |= PCI_D0;
>>   	pci_write_config_word(dev, dev->pm_cap + PCI_PM_CTRL, csr);
>> -	pci_dev_d3_sleep(dev);
>> +	pci_dev_d3hot_sleep(dev);
>>   
>>   	ret = pci_dev_wait(dev, "PM D3hot->D0", PCIE_RESET_READY_POLL_MS);
>>   	pci_dev_reset_iommu_done(dev);
>> -- 
>> 2.43.0
>>

Re: [PATCH] PCI: Apply mandatory recovery delay on return from D3cold

[Mario Limonciello]

On 7/10/26 15:16, Bjorn Helgaas wrote:
> [+cc Rafael, linux-pm]
> 
> On Wed, Jul 08, 2026 at 10:26:50AM -0500, Mario Limonciello wrote:
>> Per the "PCI Bus Power Management Interface Specification", rev. 1.2,
>> sec. 5.4, there is a minimum recovery time between programming a function
>> from D3 to D0 and accessing it.  The spec does not limit this to D3hot; it
>> applies to the D3cold to D0 transition as well.
>>
>> pci_power_up() only honors this delay on the D3hot branch.  When a device
>> returns from D3cold, platform_pci_set_power_state() has already restored
>> main power before PCI_PM_CTRL is read, so the state read from the register
>> is D0 and the transition delay block is skipped by the
>>
>> 	if (state == PCI_D0)
>> 		goto end;
>>
>> early return.  The register value is masked with PCI_PM_CTRL_STATE_MASK
>> and cannot represent D3cold, so only dev->current_state still reflects the
>> D3cold origin at this point.
>>
>> Apply the delay based on dev->current_state, ahead of the early return, so
>> it takes effect on the D3cold to D0 path before the device is accessed.
>> Use the device's d3cold_delay, which the platform may tune via _DSM and
>> quirks may raise, rather than the D3hot delay.
>>
>> To keep the existing D3hot callers unchanged, pci_dev_d3_sleep() now takes
>> the delay in milliseconds and a pci_dev_d3hot_sleep() wrapper supplies the
>> D3hot delay as before.
>>
>> Reported-by: mrh@frame.work
>> Closes: https://bugzilla.kernel.org/show_bug.cgi?id=221073
>> Signed-off-by: Mario Limonciello <mario.limonciello@amd.com>
>> ---
>> Cc: mrh@frame.work
>> Cc: stern@rowland.harvard.edu
>> Cc: hannes@vonhaugwitz.com
>> Cc: jase_harley@protonmail.com
>> Cc: superveridical@gmail.com
>> Cc: david.c.hubbard@gmail.com
>> Cc: bugzilla@logical.ink
>> Cc: michal.pecio@gmail.com
>> ---
>>   drivers/pci/pci.c | 26 ++++++++++++++++++++------
>>   1 file changed, 20 insertions(+), 6 deletions(-)
>>
>> diff --git a/drivers/pci/pci.c b/drivers/pci/pci.c
>> index 77b17b13ee615..e09cfb28fe61c 100644
>> --- a/drivers/pci/pci.c
>> +++ b/drivers/pci/pci.c
>> @@ -81,9 +81,8 @@ struct pci_pme_device {
>>    */
>>   #define PCIE_RESET_READY_POLL_MS 60000 /* msec */
>>   
>> -static void pci_dev_d3_sleep(struct pci_dev *dev)
>> +static void pci_dev_d3_sleep(unsigned int delay_ms)
>>   {
>> -	unsigned int delay_ms = max(dev->d3hot_delay, pci_pm_d3hot_delay);
>>   	unsigned int upper;
>>   
>>   	if (delay_ms) {
>> @@ -94,6 +93,11 @@ static void pci_dev_d3_sleep(struct pci_dev *dev)
>>   	}
>>   }
>>   
>> +static void pci_dev_d3hot_sleep(struct pci_dev *dev)
>> +{
>> +	pci_dev_d3_sleep(max(dev->d3hot_delay, pci_pm_d3hot_delay));
>> +}
>> +
>>   bool pci_reset_supported(struct pci_dev *dev)
>>   {
>>   	return dev->reset_methods[0] != 0;
>> @@ -1333,6 +1337,16 @@ int pci_power_up(struct pci_dev *dev)
>>   	need_restore = (state == PCI_D3hot || dev->current_state >= PCI_D3hot) &&
>>   			!(pmcsr & PCI_PM_CTRL_NO_SOFT_RESET);
>>   
>> +	/*
>> +	 * A device returning from D3cold has already been powered back on by
>> +	 * platform_pci_set_power_state() above, so PCI_PM_CTRL now reads back
>> +	 * as D0 and the transition delays below are skipped.  PCI PM 1.2 still
>> +	 * requires a minimum recovery time on the D3 to D0 transition, so apply
>> +	 * the device's D3cold recovery delay here before it is accessed.
>> +	 */
>> +	if (dev->current_state == PCI_D3cold)
>> +		pci_dev_d3_sleep(dev->d3cold_delay);
> 
> My understanding is that the "mandatory transition delays" below only
> cover a transition caused by the write to PCI_PM_CTRL, which would be
> from D1, D2, or D3hot to D0.
> 
> If the device started in D3cold, platform_pci_set_power_state(PCI_D0)
> should transition it to D0uninitialized *and* take care of any
> required delays [1].  The device should be Configuration-Ready upon
> return (and we've already read PCI_PM_CTRL above, and the read
> returned something other than PCI_ERROR_RESPONSE).
> 
> But evidently adding more delay does make a difference, even though
> the config read of PCI_PM_CTRL seemed successful.  I guess it's
> conceivable that platform AML doesn't wait quite long enough, although
> it does seem to affect more than one platform (AMD Strix Halo,
> Framework Desktop/AMD Ryzen AI Max 300, Lenovo ThinkPad T14 Gen 6 AMD
> Ryzen AI 7 Pro 350), and I *assume* the issue doesn't happen under
> Windows?

Based on some more testing that occurred this patch doesn't help.  Back 
to the drawing board.

> 
> [1] https://lore.kernel.org/linux-pci/CAJZ5v0iZN5NtUztqe=MxCRcXdBaaqzZ749OqSUkadwwBy0ugUQ@mail.gmail.com/
> 
>>   	if (state == PCI_D0)
>>   		goto end;
>>   
>> @@ -1344,7 +1358,7 @@ int pci_power_up(struct pci_dev *dev)
>>   
>>   	/* Mandatory transition delays; see PCI PM 1.2. */
>>   	if (state == PCI_D3hot) {
>> -		pci_dev_d3_sleep(dev);
>> +		pci_dev_d3hot_sleep(dev);
>>   		if (!(pmcsr & PCI_PM_CTRL_NO_SOFT_RESET)) {
>>   			ret = pci_dev_wait(dev, "power up D3hot->D0uninitialized",
>>   					   PCIE_RESET_READY_POLL_MS);
>> @@ -1514,7 +1528,7 @@ static int pci_set_low_power_state(struct pci_dev *dev, pci_power_t state, bool
>>   
>>   	/* Mandatory power management transition delays; see PCI PM 1.2. */
>>   	if (state == PCI_D3hot)
>> -		pci_dev_d3_sleep(dev);
>> +		pci_dev_d3hot_sleep(dev);
>>   	else if (state == PCI_D2)
>>   		udelay(PCI_PM_D2_DELAY);
>>   
>> @@ -4511,12 +4525,12 @@ static int pci_pm_reset(struct pci_dev *dev, bool probe)
>>   	csr &= ~PCI_PM_CTRL_STATE_MASK;
>>   	csr |= PCI_D3hot;
>>   	pci_write_config_word(dev, dev->pm_cap + PCI_PM_CTRL, csr);
>> -	pci_dev_d3_sleep(dev);
>> +	pci_dev_d3hot_sleep(dev);
>>   
>>   	csr &= ~PCI_PM_CTRL_STATE_MASK;
>>   	csr |= PCI_D0;
>>   	pci_write_config_word(dev, dev->pm_cap + PCI_PM_CTRL, csr);
>> -	pci_dev_d3_sleep(dev);
>> +	pci_dev_d3hot_sleep(dev);
>>   
>>   	ret = pci_dev_wait(dev, "PM D3hot->D0", PCIE_RESET_READY_POLL_MS);
>>   	pci_dev_reset_iommu_done(dev);
>> -- 
>> 2.43.0
>>