Hi Jeffrey,
There are several platforms that uses e1000e can’t enter Opportunistic S0ix
(PC10) when the ethernet has a link partner.
This behavior also exits in out-of-tree e1000e driver 3.4.2.1, but seems
like 3.4.2.3 fixes the issue.
A quick diff between the two versions shows that this code section may be
our solution:
/* Read from EXTCNF_CTRL in e1000_acquire_swflag_ich8lan function
* may occur during global reset and cause system hang.
* Configuration space access creates the needed delay.
* Write to E1000_STRAP RO register E1000_PCI_VENDOR_ID_REGISTER value
* insures configuration space read is done before global reset.
*/
pci_read_config_word(hw->adapter->pdev, E1000_PCI_VENDOR_ID_REGISTER,
&pci_cfg);
ew32(STRAP, pci_cfg);
e_dbg("Issuing a global reset to ich8lan\n");
ew32(CTRL, (ctrl | E1000_CTRL_RST));
/* cannot issue a flush here because it hangs the hardware */
msleep(20);
/* Configuration space access improve HW level time sync mechanism.
* Write to E1000_STRAP RO register E1000_PCI_VENDOR_ID_REGISTER
* value to insure configuration space read is done
* before any access to mac register.
*/
pci_read_config_word(hw->adapter->pdev, E1000_PCI_VENDOR_ID_REGISTER,
&pci_cfg);
ew32(STRAP, pci_cfg);
Is there any plan to support this in the upstream kernel?
Kai-Heng
Hi Jeffrey,
at 19:08, Kai-Heng Feng <[email protected]> wrote:
> Hi Jeffrey,
>
> There are several platforms that uses e1000e can’t enter Opportunistic
> S0ix (PC10) when the ethernet has a link partner.
>
> This behavior also exits in out-of-tree e1000e driver 3.4.2.1, but seems
> like 3.4.2.3 fixes the issue.
>
> A quick diff between the two versions shows that this code section may be
> our solution:
>
> /* Read from EXTCNF_CTRL in e1000_acquire_swflag_ich8lan function
> * may occur during global reset and cause system hang.
> * Configuration space access creates the needed delay.
> * Write to E1000_STRAP RO register E1000_PCI_VENDOR_ID_REGISTER value
> * insures configuration space read is done before global reset.
> */
> pci_read_config_word(hw->adapter->pdev, E1000_PCI_VENDOR_ID_REGISTER,
> &pci_cfg);
> ew32(STRAP, pci_cfg);
> e_dbg("Issuing a global reset to ich8lan\n");
> ew32(CTRL, (ctrl | E1000_CTRL_RST));
> /* cannot issue a flush here because it hangs the hardware */
> msleep(20);
>
> /* Configuration space access improve HW level time sync mechanism.
> * Write to E1000_STRAP RO register E1000_PCI_VENDOR_ID_REGISTER
> * value to insure configuration space read is done
> * before any access to mac register.
> */
> pci_read_config_word(hw->adapter->pdev, E1000_PCI_VENDOR_ID_REGISTER,
> &pci_cfg);
> ew32(STRAP, pci_cfg);
Turns out the "extra sauce” is not this part, it’s called “Dynamic LTR
support”.
>
> Is there any plan to support this in the upstream kernel?
Is there any plan to support Dynamic LTR in upstream e1000e?
Kai-Heng
>
> Kai-Heng
On 6/24/2019 10:03, Kai-Heng Feng wrote:
> Hi Jeffrey,
>
> at 19:08, Kai-Heng Feng <[email protected]> wrote:
>
>> Hi Jeffrey,
>>
>> There are several platforms that uses e1000e can’t enter Opportunistic
>> S0ix (PC10) when the ethernet has a link partner.
>>
>> This behavior also exits in out-of-tree e1000e driver 3.4.2.1, but
>> seems like 3.4.2.3 fixes the issue.
>>
>> A quick diff between the two versions shows that this code section may
>> be our solution:
>>
>> /* Read from EXTCNF_CTRL in e1000_acquire_swflag_ich8lan function
>> * may occur during global reset and cause system hang.
>> * Configuration space access creates the needed delay.
>> * Write to E1000_STRAP RO register
>> E1000_PCI_VENDOR_ID_REGISTER value
>> * insures configuration space read is done before global reset.
>> */
>> pci_read_config_word(hw->adapter->pdev,
>> E1000_PCI_VENDOR_ID_REGISTER,
>> &pci_cfg);
>> ew32(STRAP, pci_cfg);
>> e_dbg("Issuing a global reset to ich8lan\n");
>> ew32(CTRL, (ctrl | E1000_CTRL_RST));
>> /* cannot issue a flush here because it hangs the hardware */
>> msleep(20);
>>
>> /* Configuration space access improve HW level time sync
>> mechanism.
>> * Write to E1000_STRAP RO register E1000_PCI_VENDOR_ID_REGISTER
>> * value to insure configuration space read is done
>> * before any access to mac register.
>> */
>> pci_read_config_word(hw->adapter->pdev,
>> E1000_PCI_VENDOR_ID_REGISTER,
>> &pci_cfg);
>> ew32(STRAP, pci_cfg);
>
> Turns out the "extra sauce” is not this part, it’s called “Dynamic LTR
> support”.
> >>
>> Is there any plan to support this in the upstream kernel?
>
> Is there any plan to support Dynamic LTR in upstream e1000e?
>
Dynamic LTR is not stable solution. So, we can not put this solution to
upstream. I hope we will be able to fix this in HW for a future projects.
S0ix support is under discussion with our architecture. We will try
enable S0ix in our e1000e OOT driver as first step.
> Kai-Heng
>
>>
>> Kai-Heng
>
>
> _______________________________________________
> Intel-wired-lan mailing list
> [email protected]
> https://lists.osuosl.org/mailman/listinfo/intel-wired-lan
Thanks
Sasha
at 19:56, Neftin, Sasha <[email protected]> wrote:
> On 6/24/2019 10:03, Kai-Heng Feng wrote:
>> Hi Jeffrey,
>> at 19:08, Kai-Heng Feng <[email protected]> wrote:
>>> Hi Jeffrey,
>>>
>>> There are several platforms that uses e1000e can’t enter Opportunistic
>>> S0ix (PC10) when the ethernet has a link partner.
>>>
>>> This behavior also exits in out-of-tree e1000e driver 3.4.2.1, but
>>> seems like 3.4.2.3 fixes the issue.
>>>
>>> A quick diff between the two versions shows that this code section may
>>> be our solution:
>>>
>>> /* Read from EXTCNF_CTRL in e1000_acquire_swflag_ich8lan function
>>> * may occur during global reset and cause system hang.
>>> * Configuration space access creates the needed delay.
>>> * Write to E1000_STRAP RO register E1000_PCI_VENDOR_ID_REGISTER value
>>> * insures configuration space read is done before global reset.
>>> */
>>> pci_read_config_word(hw->adapter->pdev, E1000_PCI_VENDOR_ID_REGISTER,
>>> &pci_cfg);
>>> ew32(STRAP, pci_cfg);
>>> e_dbg("Issuing a global reset to ich8lan\n");
>>> ew32(CTRL, (ctrl | E1000_CTRL_RST));
>>> /* cannot issue a flush here because it hangs the hardware */
>>> msleep(20);
>>>
>>> /* Configuration space access improve HW level time sync mechanism.
>>> * Write to E1000_STRAP RO register E1000_PCI_VENDOR_ID_REGISTER
>>> * value to insure configuration space read is done
>>> * before any access to mac register.
>>> */
>>> pci_read_config_word(hw->adapter->pdev, E1000_PCI_VENDOR_ID_REGISTER,
>>> &pci_cfg);
>>> ew32(STRAP, pci_cfg);
>> Turns out the "extra sauce” is not this part, it’s called “Dynamic LTR
>> support”.
>> >>
>>> Is there any plan to support this in the upstream kernel?
>> Is there any plan to support Dynamic LTR in upstream e1000e?
> Dynamic LTR is not stable solution. So, we can not put this solution to
> upstream. I hope we will be able to fix this in HW for a future projects.
Does this mean current generation hardware won’t get the fix?
> S0ix support is under discussion with our architecture. We will try
> enable S0ix in our e1000e OOT driver as first step.
Is it possible to add Dynamic LTR as an option so users and downstream
distros can still benefit from it?
Kai-Heng
>> Kai-Heng
>>> Kai-Heng
>> _______________________________________________
>> Intel-wired-lan mailing list
>> [email protected]
>> https://lists.osuosl.org/mailman/listinfo/intel-wired-lan
>
> Thanks
> Sasha
On 6/24/2019 18:06, Kai-Heng Feng wrote:
> at 19:56, Neftin, Sasha <[email protected]> wrote:
>
>> On 6/24/2019 10:03, Kai-Heng Feng wrote:
>>> Hi Jeffrey,
>>> at 19:08, Kai-Heng Feng <[email protected]> wrote:
>>>> Hi Jeffrey,
>>>>
>>>> There are several platforms that uses e1000e can’t enter
>>>> Opportunistic S0ix (PC10) when the ethernet has a link partner.
>>>>
>>>> This behavior also exits in out-of-tree e1000e driver 3.4.2.1, but
>>>> seems like 3.4.2.3 fixes the issue.
>>>>
>>>> A quick diff between the two versions shows that this code section
>>>> may be our solution:
>>>>
>>>> /* Read from EXTCNF_CTRL in e1000_acquire_swflag_ich8lan
>>>> function
>>>> * may occur during global reset and cause system hang.
>>>> * Configuration space access creates the needed delay.
>>>> * Write to E1000_STRAP RO register
>>>> E1000_PCI_VENDOR_ID_REGISTER value
>>>> * insures configuration space read is done before global
>>>> reset.
>>>> */
>>>> pci_read_config_word(hw->adapter->pdev,
>>>> E1000_PCI_VENDOR_ID_REGISTER,
>>>> &pci_cfg);
>>>> ew32(STRAP, pci_cfg);
>>>> e_dbg("Issuing a global reset to ich8lan\n");
>>>> ew32(CTRL, (ctrl | E1000_CTRL_RST));
>>>> /* cannot issue a flush here because it hangs the hardware */
>>>> msleep(20);
>>>>
>>>> /* Configuration space access improve HW level time sync
>>>> mechanism.
>>>> * Write to E1000_STRAP RO register
>>>> E1000_PCI_VENDOR_ID_REGISTER
>>>> * value to insure configuration space read is done
>>>> * before any access to mac register.
>>>> */
>>>> pci_read_config_word(hw->adapter->pdev,
>>>> E1000_PCI_VENDOR_ID_REGISTER,
>>>> &pci_cfg);
>>>> ew32(STRAP, pci_cfg);
>>> Turns out the "extra sauce” is not this part, it’s called “Dynamic
>>> LTR support”.
>>> >>
>>>> Is there any plan to support this in the upstream kernel?
>>> Is there any plan to support Dynamic LTR in upstream e1000e?
>> Dynamic LTR is not stable solution. So, we can not put this solution
>> to upstream. I hope we will be able to fix this in HW for a future
>> projects.
>
> Does this mean current generation hardware won’t get the fix?
Current HW have a limitation. Please, try follow workaround on your
platform: echo 3 > /sys/kernel/debug/pmc_core/ltr_ignore
> >> S0ix support is under discussion with our architecture. We will try
>> enable S0ix in our e1000e OOT driver as first step.
>
> Is it possible to add Dynamic LTR as an option so users and downstream
> distros can still benefit from it?
As I said before, this is not a stable solution. No guarantee that HW
will work as properly.
>
> Kai-Heng
>
>>> Kai-Heng
>>>> Kai-Heng
>>> _______________________________________________
>>> Intel-wired-lan mailing list
>>> [email protected]
>>> https://lists.osuosl.org/mailman/listinfo/intel-wired-lan
>>
>> Thanks
>> Sasha
>
>
at 6:25 PM, Neftin, Sasha <[email protected]> wrote:
> On 6/24/2019 18:06, Kai-Heng Feng wrote:
>> at 19:56, Neftin, Sasha <[email protected]> wrote:
[Snipped]
> Current HW have a limitation. Please, try follow workaround on your
> platform: echo 3 > /sys/kernel/debug/pmc_core/ltr_ignore
Yes, this does the trick.
On 4.15 based kernel I can see the SoC enters PC10 but SLP_S0 is not
asserted.
On mainline kernel the SoC, PC10 is hit and SLP_S0 is asserted. Once SLP_S0
is asserted the SSH connection becomes really sluggish.
>> >> S0ix support is under discussion with our architecture. We will try
>>> enable S0ix in our e1000e OOT driver as first step.
>> Is it possible to add Dynamic LTR as an option so users and downstream
>> distros can still benefit from it?
> As I said before, this is not a stable solution. No guarantee that HW
> will work as properly.
Can you describe the symptom of "HW will work as properly”? Is this the
sluggish connection I observed?
Kai-Heng
>> Kai-Heng
>>>> Kai-Heng
>>>>> Kai-Heng
>>>> _______________________________________________
>>>> Intel-wired-lan mailing list
>>>> [email protected]
>>>> https://lists.osuosl.org/mailman/listinfo/intel-wired-lan
>>>
>>> Thanks
>>> Sasha