[Intel-gfx] 5.3-rc3: Frozen graphics with kcompactd migrating i915 pages

[Martin Wilck]

Hi Chris,

On Tue, 2019-09-10 at 17:20 +0300, Leho Kraav wrote:
> On Fri, Aug 09, 2019 at 01:53:43PM +0100, Chris Wilson wrote:
> > Quoting Martin Wilck (2019-08-09 13:41:42)
> > > This happened to me today, running kernel 5.3.0-rc3-1.g571863b-
> > > default
> > > (5.3-rc3 with just a few patches on top), after starting a KVM
> > > virtual
> > > machine. The X screen was frozen. Remote login via ssh was still
> > > possible, thus I was able to retrieve basic logs.
> > > 
> > > sysrq-w showed two blocked processes (kcompactd0 and KVM). After
> > > a
> > > minute, the same two processes were still blocked. KVM seems to
> > > try to
> > > acquire a lock that kcompactd is holding. kcompactd is waiting
> > > for IO
> > > to complete on pages owned by the i915 driver.
> > 
> > My bad, it's known. We haven't decided on whether to revert the
> > unfortunate recursive locking (and so hit another warn elsewhere)
> > or to
> > ignore the dirty pages (and so risk losing data across swap).
> > 
> > cb6d7c7dc7ff ("drm/i915/userptr: Acquire the page lock around
> > set_page_dirty()")
> > -Chris
> 
> Hi Chris. Is this exactly what I'm hitting at
> https://bugs.freedesktop.org/show_bug.cgi?id=111500 perhaps?
> 
> It reliably breaks the graphics userland, as the machine consistently
> freezes at any random moment.
> 
> Any workaround options, even if with a performance penalty? Revert
> cb6d7c7dc7ff but side effects?
> 
> 5.3 has useful NVMe power mgmt updates for laptops, I'd like to stick
> with the newest if possible.

There's a considerable risk that many users will start seeing this
regression when 5.3 is released. I am not aware of a workaround.

Is there an alternative to reverting aa56a292ce62 ("drm/i915/userptr:
Acquire the page lock around set_page_dirty()")? And if we do, what
would be the consequences? Would other patches need to be reverted,
too?

Thanks,
Martin