[REGRESSION] drm/etnaviv: command buffer outside valid memory window

[Lucas Stach]

Am Donnerstag, den 27.06.2019, 15:32 +0100 schrieb Russell King - ARM Linux admin:
> On Thu, Jun 27, 2019 at 11:04:17AM +0100, Russell King - ARM Linux admin wrote:
> > On Thu, Jun 27, 2019 at 11:20:15AM +0200, Lucas Stach wrote:
> > > Am Samstag, den 22.06.2019, 17:16 +0100 schrieb Russell King - ARM Linux admin:
> > > > While updating my various systems for the TCP SACK issue, I notice
> > > > that while most platforms are happy, the Cubox-i4 is not.  During
> > > > boot, we get:
> > > > 
> > > > [    0.000000] cma: Reserved 256 MiB at 0x30000000
> > > > ...
> > > > [    0.000000] Kernel command line: console=ttymxc0,115200n8 console=tty1 video=mxcfb0:dev=hdmi root=/dev/nfs rw cma=256M ahci_imx.hotplug=1 splash resume=/dev/sda1
> > > > [    0.000000] Dentry cache hash table entries: 131072 (order: 7, 524288 bytes)
> > > > [    0.000000] Inode-cache hash table entries: 65536 (order: 6, 262144 bytes)
> > > > [    0.000000] Memory: 1790972K/2097152K available (8471K kernel code, 693K rwdata, 2844K rodata, 500K init, 8062K bss, 44036K reserved, 262144K cma-reserved, 1310720K highmem)
> > > > ...
> > > > [   13.101098] etnaviv-gpu 130000.gpu: command buffer outside valid memory window
> > > > [   13.171963] etnaviv-gpu 134000.gpu: command buffer outside valid memory window
> > > 
> > > Yes, that's a regression due to different default CMA area placement
> > > and etnaviv not being smart enough to move the linear window to the
> > > right offset.
> > 
> > As it's a user visible regression, it needs fixing, either by reverting
> > the changes that caused it or by some other issue.  In the kernel, the
> > policy is "if a bug fix causes a regression, the bug fix was itself
> > wrong".  We don't fix one person's bug if it causes a regression for
> > someone else.
> > 
> > Please resolve the acknowledged regression.

The regression is caused due to a different CMA placement, which is
outside of the control of etnaviv. If you can point to the commit
causing this change in placement we could work with the
authors/maintainers of this code to get rid of the regression.
Currently I don't have the bandwidth to pinpoint the offending code
change.

> > > > and shortly after the login prompt appears, the entire SoC appears to
> > > > lock up - it becomes unresponsive on the network, or via serial console
> > > > to sysrq requests.
> > > > 
> > > > I suspect the GPU ends up scribbling over the CPU's vector page/kernel
> > > > as a result of the above two etnaviv errors when Xorg attempts to start
> > > > using the GPU.
> > > 
> > > This should not be possible. The driver notices that the command buffer
> > > isn't accessible to the GPU, which aborts the GPU init. While the
> > > etnaviv DRM device is still accessible, it will not expose any
> > > enumerable GPU cores to userspace. So there is no way for userspace to
> > > actually submit GPU commands.
> > 
> > Yep, I came to that conclusion.  Nevertheless, if I allow Xorg to start
> > with 5.1, the system totally hangs shortly thereafter.  I need to try
> > without etnaviv loaded at all.
> 
> Well, it seems to get worse.  I just tried to unload etnaviv, and was
> greeted by this oops.  It's another regression; etnaviv used to unload
> perfectly fine.  Please can you add module unload testing to your
> workflow?

As you can see from the patch I've just sent, this is a missing error
cleanup. So it's really the same regression. A module unload after
successful init of all GPU cores doesn't show this crash. The issue is
only unmasked due to the CMA placement regression.

Regards,
Lucas