[Nouveau] Nouveau: kernel hang on Optimus+Intel+NVidia GeForce 1060m

[Peter Wu]

Hi Nicholas,

Increasing the timeout won't help as the dGPU is not powered/available. If you have an external monitor connected, the dGPU will stay powered on and not suspend (with a (temporary) effect similar to booting with nouveau.runpm=0).

The reason for the failure to restore power is still unknown, it seems related to the PCI core.

Kind regards,
Peter
https://lekensteyn.nl
(pardon my brevity, top-posting and formatting, sent from my phone)


On 13 September 2017 23:13:59 BST, Nicolas Mercier <mercier.nicolas at gmail.com> wrote:
>Thanks Tobias, I tried this but unfortunately the only effect was thta
>the
>boot was delayed by an additional 4 seconds :(
>The original timeout is at
>drivers/gpu/drm/nouveau/nvkm/subdev/secboot/ls_ucode_msgqueue.c
>I tried to increase that timeout, but it did not seem to make a
>difference
>either.
>
>I think I get this error less often when I have a cable plugged in the
>output of that card at boot, whereas I always get this error when I
>boot
>without a monitor attached to the card.
>
>On Wed, Sep 13, 2017 at 2:28 PM, Tobias Klausmann <
>tobias.johannes.klausmann at mni.thm.de> wrote:
>
>> Hi,
>>
>> the system fails to initialize your vbios using secureboot (i had a
>rare
>> chance to on my system to witness it again), for now i traced it to
>> acr_boot_falcon() in
>> "linux/drivers/gpu/drm/nouveau/nvkm/falcon/msgqueue_0148cdec.c" where
>it
>> throws -110 which is -ETIMEDOUT. You could try to increase the
>timeout
>> and see if it helps something, similar to the following:
>>
>>
>> diff --git a/drivers/gpu/drm/nouveau/nvkm/falcon/msgqueue.c
>> b/drivers/gpu/drm/nouveau/nvkm/falcon/msgqueue.c
>> index 77273b53672c..fc0cb187d80d 100644
>> --- a/drivers/gpu/drm/nouveau/nvkm/falcon/msgqueue.c
>> +++ b/drivers/gpu/drm/nouveau/nvkm/falcon/msgqueue.c
>> @@ -326,7 +326,7 @@ nvkm_msgqueue_post(struct nvkm_msgqueue *priv,
>enum
>> msgqueue_msg_priority prio,
>>         int ret;
>>
>>         if (wait_init &&
>!wait_for_completion_timeout(&priv->init_done,
>> -                                        msecs_to_jiffies(1000)))
>> +                                        msecs_to_jiffies(5000)))
>>                 return -ETIMEDOUT;
>>
>>         queue = priv->func->cmd_queue(priv, prio);
>>
>> diff --git a/drivers/gpu/drm/nouveau/nvkm/falcon/msgqueue_0137c63d.c
>> b/drivers/gpu/drm/nouveau/nvkm/falcon/msgqueue_0137c63d.c
>> index fec0273158f6..c2ae525a0780 100644
>> --- a/drivers/gpu/drm/nouveau/nvkm/falcon/msgqueue_0137c63d.c
>> +++ b/drivers/gpu/drm/nouveau/nvkm/falcon/msgqueue_0137c63d.c
>> @@ -279,6 +279,7 @@ acr_boot_falcon(struct nvkm_msgqueue *priv, enum
>> nvkm_secboot_falcon falcon)
>>                 u32 flags;
>>                 u32 falcon_id;
>>         } cmd;
>> +       const struct nvkm_subdev *subdev = priv->falcon->owner;
>>
>>         memset(&cmd, 0, sizeof(cmd));
>>
>> @@ -290,7 +291,8 @@ acr_boot_falcon(struct nvkm_msgqueue *priv, enum
>> nvkm_secboot_falcon falcon)
>>         nvkm_msgqueue_post(priv, MSGQUEUE_MSG_PRIORITY_HIGH,
>&cmd.hdr,
>>                         acr_boot_falcon_callback, &completed, true);
>>
>> -       if (!wait_for_completion_timeout(&completed,
>> msecs_to_jiffies(1000)))
>> +       nvkm_error(subdev, "waiting for timeout in acr_boot_falcon
>> (msgqueue_0137bca5)\n");
>> +       if (!wait_for_completion_timeout(&completed,
>> msecs_to_jiffies(5000)))
>>                 return -ETIMEDOUT;
>>
>>         return 0;
>>
>>
>>
>> On 9/13/17 11:37 AM, Nicolas Mercier wrote:
>> > I am still looking for a solution. I have hacked around in the code
>> > and found out the following:
>> > - Nouveau prefers using PCIE power managemet over ACPI Optimus
>calls.
>> > I tried to force it to use Optimus ACPI calls, but there was an
>error
>> > calling the ACPI method so it bails out and uses PCIE PM anyway.
>> > - I tried to debug the PCIE pm states which internally uses ACPI to
>> > turn power on/off. I could print different statuses here and there.
>> > When the power is switched off, ACPI calls turn the power off then
>the
>> > kernel successfully puts the device in state D3Cold (also turning
>off
>> > power to the PCI Express port). When waking up, ACPI turns the
>power
>> > on, apparently successfully (Device [PEGP] transitioned to D0). But
>a
>> > read from the PCI bus to get the power state & other flags return
>> > 65535 (~0) and the kernel fails to set the device in D0 (although
>ACPI
>> > claims it is in D0)
>> > The call to pci_raw_set_power_state (in drivers/pci/pci.c) seems to
>> > fail because pci_read_config_word returns "~0" (and does not return
>> > any error code)
>> >
>> > I have tried different things; if I use pcie_port_pm=off, the
>NVidia
>> > card goes to state D3Hot (if I am not mistaken, its PCIE port is
>still
>> > powered) but that did not fix it. I tried to turn on or off
>different
>> > PCI/PCIexpress features such as hotplug, PM and so on. The only
>thing
>> > that works is that PM is fully disabled, which equals to the device
>> > not being powered off, so that would be equivalent to
>nouveau.runpm=0,
>> > which is not helping a lot. I have tried to force pcie aspm by
>> > recompiling the ACPI table, still no luck.
>> >
>> > I am still taking a look, but it seems like the problem comes from
>the
>> > PCIExpress PM functions and ACPI, not directly from Nouveau
>> >
>> > /n
>>