[Intel-gfx] [Bug 104805] regression in 4.15 unable to reboot/poweroff, vgaswitcheroo doesn't work, RIP: 0010:i915_pci_probe+0x11/0x70 [i915]

[Chris Murphy]

On Sun, Jan 28, 2018 at 2:37 PM, Lukas Wunner <lukas at wunner.de> wrote:
> On Sat, Jan 27, 2018 at 10:42:45AM +0000, bugzilla-daemon at freedesktop.org wrote:
>> https://bugs.freedesktop.org/show_bug.cgi?id=104805
>>
>> --- Comment #6 from Chris Wilson <chris at chris-wilson.co.uk> ---
>> Sigh. Revert then solve the bloat another way. I think we can move it to a
>> seperate module and only load it during pci_probe. (That module can then play
>> some tricks to construct driver_info using private commands to allow feature
>> enabling.)
>
> The following might work:  Set i915_pci_driver->id_table = NULL.
> Then in i915_init(), after calling pci_register_driver(), iterate
> over pciidlist[] and call pci_add_dynid() for each entry.  Finally,
> call pci_free_dynids().
>
> The problem is that after manually unbinding the driver via sysfs,
> you can't bind it again.
>
> Also, as soon as Intel introduces a discrete GPU (e.g. in a dock,
> hotplugged via Thunderbolt), the whole idea falls apart.
>
> How many bytes are we talking about here anyway?  I imagine it's less
> than a page.  Honestly I think it's not worth the effort.
>
> I'm wondering why noone else is seeing this crash, myself included.
> Is the driver somehow unbound and rebound on Chris Murphy's machine?

I can't answer this question, my understanding this model defaults to
discrete GPU (AMD) graphics by the firmware. Fedora has enabled
/usr/lib/systemd/system/switcheroo-control.service by default to try
to disable discrete GPUs on dual GPU computers, favoring the more
power efficient integrated GPU. I see evidence the service is enabled
but no evidence whether it was run and in dmesg it's definitely not
enabled like what happens on 4.14. So maybe the switcheroo attempt is
what triggers the crash?

>
> @Chris Murphy: You may have noticed an 8 second delay on reboot with
> 4.15.  Before you waste time bisecting it:  A fix is already queued
> for 4.16 and marked for stable:
> https://git.kernel.org/pub/scm/linux/kernel/git/helgaas/pci.git/commit/?id=493fb50e958c

I haven't noticed such a delay.

-- 
Chris Murphy