Skip to contentSkip to conversation

old radeon latency problem still unfixed?

DISCUSSION Lee Revell 6 participants 17 messages Feb 26, 2006 – Feb 27, 2006

lore
Share & tools
lore
author
#2 lore raw
Quoted 11 lines · Fernando Lopez-Lezcano
Quoted 7 lines · Lee Revell

Users report that this patch:

https://www.redhat.com/archives/fedora-devel-list/2004-June/msg00072.html

is still needed to eliminate audio underruns for Radeon users.

Any news on this?

You mean on the plain vanilla stable kernel tree? Users running what?

Presumably.

Quoted 2 lines

I'm using 2.6.15-rt18 currently on radeon machines (9250 chipset) with no problems that I can see.

We don't want to require the -rt kernel to use JACK at modest latencies, that's like killing a fly with a gun.

Lee

#11 lore raw
In reply to Lee Revell
Quoted 6 lines · Lee Revell

Users report that this patch:

https://www.redhat.com/archives/fedora-devel-list/2004-June/msg00072.html

is still needed to eliminate audio underruns for Radeon users.

That's a 2.6.4 patch which generates 100% rejects.

But still, if that patch helped and didn't throw a billion might_sleep() and people were using preemptible kernels then we have a lock_kernel() problem. A suitable fix would be to make sure all the locking's tight and to convert DRM to use unlocked_ioctl.

#12 lore raw
Quoted 6 lines · Andrew Morton
Quoted 5 lines · Lee Revell

Users report that this patch:

https://www.redhat.com/archives/fedora-devel-list/2004-June/msg00072.html

Quoted 7 lines
Quoted 1 line

is still needed to eliminate audio underruns for Radeon users.

That's a 2.6.4 patch which generates 100% rejects.

But still, if that patch helped and didn't throw a billion might_sleep() and people were using preemptible kernels then we have a lock_kernel() problem.

lock_kernel() is a semaphore nowadays.... unless those people just turned that off, at which point.. their problem ;)

author
#4 lore raw
In reply to Arjan van de Ven
Quoted 12 lines · Arjan van de Ven
Quoted 7 lines · Lee Revell

Users report that this patch:

https://www.redhat.com/archives/fedora-devel-list/2004-June/msg00072.html

is still needed to eliminate audio underruns for Radeon users.

Any news on this?

well that patch is not working (it's already in the mail, it schedules with spinlocks ;)

Understood, your response indicated that you would investigate a proper solution.

Quoted 3 lines

the other angle is that you're trading 3D performance vs audio performance....

AFAICT it's more like trading 3D performance for having audio work at all. Other video drivers that were too aggressive and caused audio dropouts (VIA) were fixed, even though there was a slight performance cost.

Lee

author
#13 lore raw
In reply to Arjan van de Ven
Quoted 3 lines · Arjan van de Ven

lock_kernel() is a semaphore nowadays.... unless those people just turned that off, at which point.. their problem ;)

Thanks I will check on this.

Lee

#5 lore raw
In reply to Lee Revell
Quoted 5 lines

AFAICT it's more like trading 3D performance for having audio work at all. Other video drivers that were too aggressive and caused audio dropouts (VIA) were fixed, even though there was a slight performance cost.

In addition, the radeon DRI shouldn't do active spinning like that in "normal" circumstances ... it should instead block on interrupts. if it does, I suppose that could safely be considered as a bug in the radeon DRM/DRI driver. It will do such loops on engine reset and such, which happen on X launch, VT switches or in case of lockups... I have to double check what happens in the code path used for 2d/3d transitions though, those might be a problem.

Ben.

#14 lore raw
In reply to Lee Revell
Quoted 3 lines

Thanks I will check on this.

I don't suppose they are running with X renice -10?

or some such.. that was done by a few vendors previously.. if X is using hw accel, then it will be in the DRM driver a bit...

Dave.

author
#6 lore raw
In reply to Benjamin Herrenschmidt
Quoted 5 lines · Benjamin Herrenschmidt
Quoted 4 lines

AFAICT it's more like trading 3D performance for having audio work at all. Other video drivers that were too aggressive and caused audio dropouts (VIA) were fixed, even though there was a slight performance cost.

Quoted 7 lines

In addition, the radeon DRI shouldn't do active spinning like that in "normal" circumstances ... it should instead block on interrupts. if it does, I suppose that could safely be considered as a bug in the radeon DRM/DRI driver. It will do such loops on engine reset and such, which happen on X launch, VT switches or in case of lockups... I have to double check what happens in the code path used for 2d/3d transitions though, those might be a problem.

What about switching from 2D->3D mode, like when xscreensaver kicks in? IIRC people reported audio underruns when that happened but I could never narrow it down any further.

If as Arjan said the only lock this driver takes is the BKL then it's either a local config issue (ancient kernel or failure to enable preempt BKL) or something at the hardware level... I'm waiting for more info from the original reporter.

Lee

author
#15 lore raw
In reply to Dave Airlie
Quoted 8 lines · Dave Airlie
Quoted 3 lines

Thanks I will check on this.

I don't suppose they are running with X renice -10?

or some such.. that was done by a few vendors previously.. if X is using hw accel, then it will be in the DRM driver a bit...

Radeon uses DRI for regular 2D XAA acceleration? That's good to know.

This is not very common right?

Lee

#16 lore raw
Quoted 5 lines · Lee Revell
Quoted 4 lines · Dave Airlie
Quoted 3 lines

Thanks I will check on this.

Quoted 8 lines
Quoted 4 lines

I don't suppose they are running with X renice -10?

or some such.. that was done by a few vendors previously.. if X is using hw accel, then it will be in the DRM driver a bit...

Radeon uses DRI for regular 2D XAA acceleration? That's good to know.

This is not very common right?

Common in what sense? all radeons that have DRM enabled use the CP to do 2D XAA accel, the drm is the interface to the CP, granted it doesn't do a huge amount of work, X sets up all the accel in an indirect buffer, and just tells the DRM where the buffer starts.. so it isn't doing a lot in the kernel.

Dave.

Quoted 4 lines

Lee

#7 lore raw
In reply to Lee Revell
Quoted 14 lines · Lee Revell
Quoted 5 lines · Benjamin Herrenschmidt
Quoted 4 lines

AFAICT it's more like trading 3D performance for having audio work at all. Other video drivers that were too aggressive and caused audio dropouts (VIA) were fixed, even though there was a slight performance cost.

Quoted 7 lines

In addition, the radeon DRI shouldn't do active spinning like that in "normal" circumstances ... it should instead block on interrupts. if it does, I suppose that could safely be considered as a bug in the radeon DRM/DRI driver. It will do such loops on engine reset and such, which happen on X launch, VT switches or in case of lockups... I have to double check what happens in the code path used for 2d/3d transitions though, those might be a problem.

Quoted 3 lines

What about switching from 2D->3D mode, like when xscreensaver kicks in? IIRC people reported audio underruns when that happened but I could never narrow it down any further.

Well... as soon as a 3d window appears, the server starts switching all the time. there might be some spin loop in there remaining...

Quoted 4 lines

If as Arjan said the only lock this driver takes is the BKL then it's either a local config issue (ancient kernel or failure to enable preempt BKL) or something at the hardware level... I'm waiting for more info from the original reporter.

Ben.

author
#8 lore raw
Quoted 2 lines · Benjamin Herrenschmidt

Well... as soon as a 3d window appears, the server starts switching all the time. there might be some spin loop in there remaining...

But if the only lock taken in the radeon driver is the BKL, the SCHED_NORMAL X server should not be able to delay a SCHED_FIFO process right?

Lee

#9 lore raw
Quoted 7 lines · Lee Revell
Quoted 2 lines · Benjamin Herrenschmidt

Well... as soon as a 3d window appears, the server starts switching all the time. there might be some spin loop in there remaining...

But if the only lock taken in the radeon driver is the BKL, the SCHED_NORMAL X server should not be able to delay a SCHED_FIFO process right?

If you have preempt enabled, I suppose .... can we preempt with the BKL nowadays ?

author
#10 lore raw
Quoted 11 lines · Benjamin Herrenschmidt
Quoted 7 lines · Lee Revell
Quoted 2 lines · Benjamin Herrenschmidt

Well... as soon as a 3d window appears, the server starts switching all the time. there might be some spin loop in there remaining...

But if the only lock taken in the radeon driver is the BKL, the SCHED_NORMAL X server should not be able to delay a SCHED_FIFO process right?

If you have preempt enabled, I suppose .... can we preempt with the BKL nowadays ?

Yes.

This seems to have been user error - old kernel and/or broken config. Until I get more feedback from the user, this can probably be disregarded.

Lee