From: Mikko Perttunen <hidden> Date: 2021-06-09 13:47:17
Hi,
I'm observing nouveau not initializing recently on linux-next on my
Tegra186 Jetson TX2 board. Specifically it looks like BO allocation is
failing when initializing the sync subsystem:
[ 21.858149] nouveau 17000000.gpu: DRM: failed to initialise sync
subsystem, -28
I have been bisecting and I have found two patches that affect this.
Firstly, things first break on
d02117f8efaa drm/ttm: remove special handling for non GEM drivers
starting to return error code -12. Then, at
d79025c7f5e3 drm/ttm: always initialize the full ttm_resource v2
the error code changes to the above -28.
If I checkout one commit prior to d79025c7f5e3 and revert d02117f8efaa,
things work again. There are a bunch of other TTM commits between this
and HEAD, so reverting these on top of HEAD doesn't work. However, I
checked that both yesterday's and today's nexts are also broken.
Thank you,
Mikko
From: Christian König <christian.koenig@amd.com> Date: 2021-06-09 14:18:02
Hi Mikko,
strange sounds like Nouveau was somehow also using the GEM workaround
for VMWGFX as well.
But -12 means -ENOMEM which doesn't fits into the picture.
I will try with a G710, but if that doesn't yields anything I need some
more input from you.
Thanks for the report,
Christian.
Am 09.06.21 um 15:47 schrieb Mikko Perttunen:
Hi,
I'm observing nouveau not initializing recently on linux-next on my
Tegra186 Jetson TX2 board. Specifically it looks like BO allocation is
failing when initializing the sync subsystem:
[ 21.858149] nouveau 17000000.gpu: DRM: failed to initialise sync
subsystem, -28
I have been bisecting and I have found two patches that affect this.
Firstly, things first break on
d02117f8efaa drm/ttm: remove special handling for non GEM drivers
starting to return error code -12. Then, at
d79025c7f5e3 drm/ttm: always initialize the full ttm_resource v2
the error code changes to the above -28.
If I checkout one commit prior to d79025c7f5e3 and revert
d02117f8efaa, things work again. There are a bunch of other TTM
commits between this and HEAD, so reverting these on top of HEAD
doesn't work. However, I checked that both yesterday's and today's
nexts are also broken.
Thank you,
Mikko
From: Ilia Mirkin <hidden> Date: 2021-06-09 14:52:38
Christian - potentially relevant is that Tegra doesn't have VRAM at
all -- all GTT (or GART or whatever it's called nowadays). No
fake/stolen VRAM.
Cheers,
-ilia
On Wed, Jun 9, 2021 at 10:18 AM Christian König
[off-list ref] wrote:
Hi Mikko,
strange sounds like Nouveau was somehow also using the GEM workaround
for VMWGFX as well.
But -12 means -ENOMEM which doesn't fits into the picture.
I will try with a G710, but if that doesn't yields anything I need some
more input from you.
Thanks for the report,
Christian.
Am 09.06.21 um 15:47 schrieb Mikko Perttunen:
quoted
Hi,
I'm observing nouveau not initializing recently on linux-next on my
Tegra186 Jetson TX2 board. Specifically it looks like BO allocation is
failing when initializing the sync subsystem:
[ 21.858149] nouveau 17000000.gpu: DRM: failed to initialise sync
subsystem, -28
I have been bisecting and I have found two patches that affect this.
Firstly, things first break on
d02117f8efaa drm/ttm: remove special handling for non GEM drivers
starting to return error code -12. Then, at
d79025c7f5e3 drm/ttm: always initialize the full ttm_resource v2
the error code changes to the above -28.
If I checkout one commit prior to d79025c7f5e3 and revert
d02117f8efaa, things work again. There are a bunch of other TTM
commits between this and HEAD, so reverting these on top of HEAD
doesn't work. However, I checked that both yesterday's and today's
nexts are also broken.
Thank you,
Mikko
From: Christian König <christian.koenig@amd.com> Date: 2021-06-09 14:58:28
Good point, but I think that is unrelated.
My suspicion is rather that nouveau is not initializing the underlying
GEM object for internal allocations.
So what happens is the same as on VMWGFX that TTM doesn't know anything
about the size to of the BO resulting in a kmalloc() with a random value
and eventually -ENOMEM.
Good news is that I can reproduce it, so going to look into that later
today.
Regards,
Christian.
Am 09.06.21 um 16:52 schrieb Ilia Mirkin:
Christian - potentially relevant is that Tegra doesn't have VRAM at
all -- all GTT (or GART or whatever it's called nowadays). No
fake/stolen VRAM.
Cheers,
-ilia
On Wed, Jun 9, 2021 at 10:18 AM Christian König
[off-list ref] wrote:
quoted
Hi Mikko,
strange sounds like Nouveau was somehow also using the GEM workaround
for VMWGFX as well.
But -12 means -ENOMEM which doesn't fits into the picture.
I will try with a G710, but if that doesn't yields anything I need some
more input from you.
Thanks for the report,
Christian.
Am 09.06.21 um 15:47 schrieb Mikko Perttunen:
quoted
Hi,
I'm observing nouveau not initializing recently on linux-next on my
Tegra186 Jetson TX2 board. Specifically it looks like BO allocation is
failing when initializing the sync subsystem:
[ 21.858149] nouveau 17000000.gpu: DRM: failed to initialise sync
subsystem, -28
I have been bisecting and I have found two patches that affect this.
Firstly, things first break on
d02117f8efaa drm/ttm: remove special handling for non GEM drivers
starting to return error code -12. Then, at
d79025c7f5e3 drm/ttm: always initialize the full ttm_resource v2
the error code changes to the above -28.
If I checkout one commit prior to d79025c7f5e3 and revert
d02117f8efaa, things work again. There are a bunch of other TTM
commits between this and HEAD, so reverting these on top of HEAD
doesn't work. However, I checked that both yesterday's and today's
nexts are also broken.
Thank you,
Mikko
From: Ilia Mirkin <hidden> Date: 2021-06-09 15:13:18
GEM init happens here:
https://cgit.freedesktop.org/drm/drm/tree/drivers/gpu/drm/nouveau/nouveau_gem.c#n221
Note the bo alloc / gem init / bo init dance.
I don't think there is a GEM object for internal allocations at all --
we just allocate bo's directly and that's it. Perhaps you meant
something else? I thought GEM was meant for externally-available
objects.
Cheers,
-ilia
On Wed, Jun 9, 2021 at 10:58 AM Christian König
[off-list ref] wrote:
Good point, but I think that is unrelated.
My suspicion is rather that nouveau is not initializing the underlying
GEM object for internal allocations.
So what happens is the same as on VMWGFX that TTM doesn't know anything
about the size to of the BO resulting in a kmalloc() with a random value
and eventually -ENOMEM.
Good news is that I can reproduce it, so going to look into that later
today.
Regards,
Christian.
Am 09.06.21 um 16:52 schrieb Ilia Mirkin:
quoted
Christian - potentially relevant is that Tegra doesn't have VRAM at
all -- all GTT (or GART or whatever it's called nowadays). No
fake/stolen VRAM.
Cheers,
-ilia
On Wed, Jun 9, 2021 at 10:18 AM Christian König
[off-list ref] wrote:
quoted
Hi Mikko,
strange sounds like Nouveau was somehow also using the GEM workaround
for VMWGFX as well.
But -12 means -ENOMEM which doesn't fits into the picture.
I will try with a G710, but if that doesn't yields anything I need some
more input from you.
Thanks for the report,
Christian.
Am 09.06.21 um 15:47 schrieb Mikko Perttunen:
quoted
Hi,
I'm observing nouveau not initializing recently on linux-next on my
Tegra186 Jetson TX2 board. Specifically it looks like BO allocation is
failing when initializing the sync subsystem:
[ 21.858149] nouveau 17000000.gpu: DRM: failed to initialise sync
subsystem, -28
I have been bisecting and I have found two patches that affect this.
Firstly, things first break on
d02117f8efaa drm/ttm: remove special handling for non GEM drivers
starting to return error code -12. Then, at
d79025c7f5e3 drm/ttm: always initialize the full ttm_resource v2
the error code changes to the above -28.
If I checkout one commit prior to d79025c7f5e3 and revert
d02117f8efaa, things work again. There are a bunch of other TTM
commits between this and HEAD, so reverting these on top of HEAD
doesn't work. However, I checked that both yesterday's and today's
nexts are also broken.
Thank you,
Mikko
From: Christian König <christian.koenig@amd.com> Date: 2021-06-09 15:22:07
Yeah, exactly that's the root cause of the problem.
GEM objects are the base class for TTM BOs for quite a while now.
So we at least need to initialize them or otherwise at least the size of
the object is not available.
Going to send a fix in a few minutes.
Thanks,
Christian.
Am 09.06.21 um 17:13 schrieb Ilia Mirkin:
Good point, but I think that is unrelated.
My suspicion is rather that nouveau is not initializing the underlying
GEM object for internal allocations.
So what happens is the same as on VMWGFX that TTM doesn't know anything
about the size to of the BO resulting in a kmalloc() with a random value
and eventually -ENOMEM.
Good news is that I can reproduce it, so going to look into that later
today.
Regards,
Christian.
Am 09.06.21 um 16:52 schrieb Ilia Mirkin:
quoted
Christian - potentially relevant is that Tegra doesn't have VRAM at
all -- all GTT (or GART or whatever it's called nowadays). No
fake/stolen VRAM.
Cheers,
-ilia
On Wed, Jun 9, 2021 at 10:18 AM Christian König
[off-list ref] wrote:
quoted
Hi Mikko,
strange sounds like Nouveau was somehow also using the GEM workaround
for VMWGFX as well.
But -12 means -ENOMEM which doesn't fits into the picture.
I will try with a G710, but if that doesn't yields anything I need some
more input from you.
Thanks for the report,
Christian.
Am 09.06.21 um 15:47 schrieb Mikko Perttunen:
quoted
Hi,
I'm observing nouveau not initializing recently on linux-next on my
Tegra186 Jetson TX2 board. Specifically it looks like BO allocation is
failing when initializing the sync subsystem:
[ 21.858149] nouveau 17000000.gpu: DRM: failed to initialise sync
subsystem, -28
I have been bisecting and I have found two patches that affect this.
Firstly, things first break on
d02117f8efaa drm/ttm: remove special handling for non GEM drivers
starting to return error code -12. Then, at
d79025c7f5e3 drm/ttm: always initialize the full ttm_resource v2
the error code changes to the above -28.
If I checkout one commit prior to d79025c7f5e3 and revert
d02117f8efaa, things work again. There are a bunch of other TTM
commits between this and HEAD, so reverting these on top of HEAD
doesn't work. However, I checked that both yesterday's and today's
nexts are also broken.
Thank you,
Mikko