2020-12-08 08:29:41

by Kai-Heng Feng

[permalink] [raw]
Subject: [PATCH 1/2] PCI/ASPM: Store disabled ASPM states

If we use sysfs to disable L1 ASPM, then enable one L1 ASPM substate
again, all other substates will also be enabled too:

link# grep . *
clkpm:1
l0s_aspm:1
l1_1_aspm:1
l1_1_pcipm:1
l1_2_aspm:1
l1_2_pcipm:1
l1_aspm:1

link# echo 0 > l1_aspm

link# grep . *
clkpm:1
l0s_aspm:1
l1_1_aspm:0
l1_1_pcipm:0
l1_2_aspm:0
l1_2_pcipm:0
l1_aspm:0

link# echo 1 > l1_2_aspm

link# grep . *
clkpm:1
l0s_aspm:1
l1_1_aspm:1
l1_1_pcipm:1
l1_2_aspm:1
l1_2_pcipm:1
l1_aspm:1

This is because disabled ASPM states weren't saved, so enable any of the
substate will also enable others.

So store the disabled ASPM states for consistency.

Signed-off-by: Kai-Heng Feng <[email protected]>
---
drivers/pci/pcie/aspm.c | 10 ++++++++--
1 file changed, 8 insertions(+), 2 deletions(-)

diff --git a/drivers/pci/pcie/aspm.c b/drivers/pci/pcie/aspm.c
index ac0557a305af..2ea9fddadfad 100644
--- a/drivers/pci/pcie/aspm.c
+++ b/drivers/pci/pcie/aspm.c
@@ -658,6 +658,8 @@ static void pcie_aspm_cap_init(struct pcie_link_state *link, int blacklist)
/* Setup initial capable state. Will be updated later */
link->aspm_capable = link->aspm_support;

+ link->aspm_disable = link->aspm_capable & ~link->aspm_default;
+
/* Get and check endpoint acceptable latencies */
list_for_each_entry(child, &linkbus->devices, bus_list) {
u32 reg32, encoding;
@@ -1226,11 +1228,15 @@ static ssize_t aspm_attr_store_common(struct device *dev,
mutex_lock(&aspm_lock);

if (state_enable) {
- link->aspm_disable &= ~state;
/* need to enable L1 for substates */
if (state & ASPM_STATE_L1SS)
- link->aspm_disable &= ~ASPM_STATE_L1;
+ state |= ASPM_STATE_L1;
+
+ link->aspm_disable &= ~state;
} else {
+ if (state == ASPM_STATE_L1)
+ state |= ASPM_STATE_L1SS;
+
link->aspm_disable |= state;
}

--
2.29.2


2020-12-08 17:21:47

by Heiner Kallweit

[permalink] [raw]
Subject: Re: [PATCH 1/2] PCI/ASPM: Store disabled ASPM states

Am 08.12.2020 um 09:25 schrieb Kai-Heng Feng:
> If we use sysfs to disable L1 ASPM, then enable one L1 ASPM substate
> again, all other substates will also be enabled too:
>
> link# grep . *
> clkpm:1
> l0s_aspm:1
> l1_1_aspm:1
> l1_1_pcipm:1
> l1_2_aspm:1
> l1_2_pcipm:1
> l1_aspm:1
>
> link# echo 0 > l1_aspm
>
> link# grep . *
> clkpm:1
> l0s_aspm:1
> l1_1_aspm:0
> l1_1_pcipm:0
> l1_2_aspm:0
> l1_2_pcipm:0
> l1_aspm:0
>
> link# echo 1 > l1_2_aspm
>
> link# grep . *
> clkpm:1
> l0s_aspm:1
> l1_1_aspm:1
> l1_1_pcipm:1
> l1_2_aspm:1
> l1_2_pcipm:1
> l1_aspm:1
>
> This is because disabled ASPM states weren't saved, so enable any of the
> substate will also enable others.
>
> So store the disabled ASPM states for consistency.
>
> Signed-off-by: Kai-Heng Feng <[email protected]>
> ---
> drivers/pci/pcie/aspm.c | 10 ++++++++--
> 1 file changed, 8 insertions(+), 2 deletions(-)
>
> diff --git a/drivers/pci/pcie/aspm.c b/drivers/pci/pcie/aspm.c
> index ac0557a305af..2ea9fddadfad 100644
> --- a/drivers/pci/pcie/aspm.c
> +++ b/drivers/pci/pcie/aspm.c
> @@ -658,6 +658,8 @@ static void pcie_aspm_cap_init(struct pcie_link_state *link, int blacklist)
> /* Setup initial capable state. Will be updated later */
> link->aspm_capable = link->aspm_support;
>
> + link->aspm_disable = link->aspm_capable & ~link->aspm_default;
> +

This makes sense only in combination with patch 2. However I think patch 1
should be independent of patch 2. Especially if we consider patch 1 a fix
that is applied to stable whilst patch 2 is an improvement for next.

> /* Get and check endpoint acceptable latencies */
> list_for_each_entry(child, &linkbus->devices, bus_list) {
> u32 reg32, encoding;
> @@ -1226,11 +1228,15 @@ static ssize_t aspm_attr_store_common(struct device *dev,
> mutex_lock(&aspm_lock);
>
> if (state_enable) {
> - link->aspm_disable &= ~state;
> /* need to enable L1 for substates */
> if (state & ASPM_STATE_L1SS)
> - link->aspm_disable &= ~ASPM_STATE_L1;
> + state |= ASPM_STATE_L1;
> +
> + link->aspm_disable &= ~state;

I don't see what this part of the patch changes. Can you elaborate on why
this is needed?

> } else {
> + if (state == ASPM_STATE_L1)
> + state |= ASPM_STATE_L1SS;
> +

I think this part should be sufficient to fix the behavior. because what
I think currently happens:

1. original status: policy powersupersave, nothing disabled -> L1 + L1SS active
2. disable L1: L1 disabled, pcie_config_aspm_link() disabled L1 and L1SS
w/o adding L1SS to link-> aspm_disabled
3. enable one L1SS state: aspm_attr_store_common() removes L1 from
link->aspm_disabled -> link->aspm_disabled is empty, resulting in
L1 + L1SS being active

> link->aspm_disable |= state;
> }
>
>

2020-12-09 17:38:24

by Kai-Heng Feng

[permalink] [raw]
Subject: Re: [PATCH 1/2] PCI/ASPM: Store disabled ASPM states



> On Dec 9, 2020, at 01:11, Heiner Kallweit <[email protected]> wrote:
>
> Am 08.12.2020 um 09:25 schrieb Kai-Heng Feng:
>> If we use sysfs to disable L1 ASPM, then enable one L1 ASPM substate
>> again, all other substates will also be enabled too:
>>
>> link# grep . *
>> clkpm:1
>> l0s_aspm:1
>> l1_1_aspm:1
>> l1_1_pcipm:1
>> l1_2_aspm:1
>> l1_2_pcipm:1
>> l1_aspm:1
>>
>> link# echo 0 > l1_aspm
>>
>> link# grep . *
>> clkpm:1
>> l0s_aspm:1
>> l1_1_aspm:0
>> l1_1_pcipm:0
>> l1_2_aspm:0
>> l1_2_pcipm:0
>> l1_aspm:0
>>
>> link# echo 1 > l1_2_aspm
>>
>> link# grep . *
>> clkpm:1
>> l0s_aspm:1
>> l1_1_aspm:1
>> l1_1_pcipm:1
>> l1_2_aspm:1
>> l1_2_pcipm:1
>> l1_aspm:1
>>
>> This is because disabled ASPM states weren't saved, so enable any of the
>> substate will also enable others.
>>
>> So store the disabled ASPM states for consistency.
>>
>> Signed-off-by: Kai-Heng Feng <[email protected]>
>> ---
>> drivers/pci/pcie/aspm.c | 10 ++++++++--
>> 1 file changed, 8 insertions(+), 2 deletions(-)
>>
>> diff --git a/drivers/pci/pcie/aspm.c b/drivers/pci/pcie/aspm.c
>> index ac0557a305af..2ea9fddadfad 100644
>> --- a/drivers/pci/pcie/aspm.c
>> +++ b/drivers/pci/pcie/aspm.c
>> @@ -658,6 +658,8 @@ static void pcie_aspm_cap_init(struct pcie_link_state *link, int blacklist)
>> /* Setup initial capable state. Will be updated later */
>> link->aspm_capable = link->aspm_support;
>>
>> + link->aspm_disable = link->aspm_capable & ~link->aspm_default;
>> +
>
> This makes sense only in combination with patch 2. However I think patch 1
> should be independent of patch 2. Especially if we consider patch 1 a fix
> that is applied to stable whilst patch 2 is an improvement for next.
>
>> /* Get and check endpoint acceptable latencies */
>> list_for_each_entry(child, &linkbus->devices, bus_list) {
>> u32 reg32, encoding;
>> @@ -1226,11 +1228,15 @@ static ssize_t aspm_attr_store_common(struct device *dev,
>> mutex_lock(&aspm_lock);
>>
>> if (state_enable) {
>> - link->aspm_disable &= ~state;
>> /* need to enable L1 for substates */
>> if (state & ASPM_STATE_L1SS)
>> - link->aspm_disable &= ~ASPM_STATE_L1;
>> + state |= ASPM_STATE_L1;
>> +
>> + link->aspm_disable &= ~state;
>
> I don't see what this part of the patch changes. Can you elaborate on why
> this is needed?

No this is just a cosmetic change. Of course "cosmetic" is really subjective.
I'll drop this part in v2.

>
>> } else {
>> + if (state == ASPM_STATE_L1)
>> + state |= ASPM_STATE_L1SS;
>> +
>
> I think this part should be sufficient to fix the behavior. because what
> I think currently happens:
>
> 1. original status: policy powersupersave, nothing disabled -> L1 + L1SS active
> 2. disable L1: L1 disabled, pcie_config_aspm_link() disabled L1 and L1SS
> w/o adding L1SS to link-> aspm_disabled
> 3. enable one L1SS state: aspm_attr_store_common() removes L1 from
> link->aspm_disabled -> link->aspm_disabled is empty, resulting in
> L1 + L1SS being active

Yes. This is the case the patch solves.

Kai-Heng

>
>> link->aspm_disable |= state;
>> }