Skip to contentSkip to conversation

drm/panthor: intermittent freezes in Xorg xserver

BUG Marius Dinu 3 participants 6 messages Jun 26 – Jun 29 Arm Rockchip

lore
Share & tools
lore
#0 lore raw

Hi.

The system is Radxa Rock 5 ITX, RK3588, 32GB RAM,
mainline kernel v7.1.0 with panthor driver, mesa 26.0.7 with panfrost driver, xorg-server v21.1.22 accelerated by modesetting driver, x11perf v1.7.0.

The most clean way to reproduce this bug:
Form a console or ssh session, start Xorg in background, without a window manager, and then run x11perf on that display.

Xorg :0 &
x11perf -all -display :0

The bug: weird intermittent performance issues in some of the tests. Examples:

    1000 reps @  18.0799 msec (    55.3/sec): 10x10 rectangle outline
    1000 reps @  18.0791 msec (    55.3/sec): 10x10 rectangle outline
    1000 reps @  18.0790 msec (    55.3/sec): 10x10 rectangle outline
    1000 reps @  18.0789 msec (    55.3/sec): 10x10 rectangle outline
    1000 reps @   0.0287 msec ( 34800.0/sec): 10x10 rectangle outline
    5000 trep @  14.4691 msec (    69.1/sec): 10x10 rectangle outline
 5000000 reps @   0.0029 msec (345000.0/sec): 1-pixel solid circle
 5000000 reps @   0.0009 msec (1100000.0/sec): 1-pixel solid circle
 5000000 reps @   0.0012 msec (818000.0/sec): 1-pixel solid circle
 5000000 reps @   0.0000 msec (20900000.0/sec): 1-pixel solid circle
 5000000 reps @   0.0000 msec (32300000.0/sec): 1-pixel solid circle
25000000 trep @   0.0010 msec (979000.0/sec): 1-pixel solid circle
    1000 reps @  31.6607 msec (    31.6/sec): Fill 10x10 trapezoid
    1000 reps @  16.9012 msec (    59.2/sec): Fill 10x10 trapezoid
    1000 reps @   0.0234 msec ( 42700.0/sec): Fill 10x10 trapezoid
    1000 reps @   0.0272 msec ( 36700.0/sec): Fill 10x10 trapezoid
    1000 reps @  32.1268 msec (    31.1/sec): Fill 10x10 trapezoid
    5000 trep @  16.1479 msec (    61.9/sec): Fill 10x10 trapezoid

There are no messages in kernel log or Xorg log.

Visually, most of the tests runs fast as expected, but in some of the tests it gets "stuck" randomly. The test suddenly freezes while drawing the pattern. Then, after a while (a few seconds up to a couple of minutes), it gets unstuck by itself, draws a few more patterns extremely fast for less than a second, and then it gets stuck again.
The small test patterns (10px) are a lot more often affected than large patterns (100px or more).

The bug can also be seen by trying to move or resize windows. Setting governors to "performance" for the CPU and/or GPU has no effect.

I'm available for tests.
Thanks.

Marius Dinu

_______________________________________________
Linux-rockchip mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-rockchip

#1 lore raw
Quoted 5 lines · On Fri, 26 Jun 2026 14:50:17 +0300 Marius Dinu <[email protected]> wrote:

Hi.

The system is Radxa Rock 5 ITX, RK3588, 32GB RAM,
mainline kernel v7.1.0 with panthor driver, mesa 26.0.7 with panfrost driver, xorg-server v21.1.22 accelerated by modesetting driver, x11perf v1.7.0.

Are we sure that's a panthor issue? Also, is this a regression introduced by the latest kernel version, or has it been like that from the start?

Also, would you mind opening an issue on [1]?

Quoted 32 lines

The most clean way to reproduce this bug:
Form a console or ssh session, start Xorg in background, without a window manager, and then run x11perf on that display.

Xorg :0 &
x11perf -all -display :0

The bug: weird intermittent performance issues in some of the tests. Examples:

    1000 reps @  18.0799 msec (    55.3/sec): 10x10 rectangle outline
    1000 reps @  18.0791 msec (    55.3/sec): 10x10 rectangle outline
    1000 reps @  18.0790 msec (    55.3/sec): 10x10 rectangle outline
    1000 reps @  18.0789 msec (    55.3/sec): 10x10 rectangle outline
    1000 reps @   0.0287 msec ( 34800.0/sec): 10x10 rectangle outline
    5000 trep @  14.4691 msec (    69.1/sec): 10x10 rectangle outline
 5000000 reps @   0.0029 msec (345000.0/sec): 1-pixel solid circle
 5000000 reps @   0.0009 msec (1100000.0/sec): 1-pixel solid circle
 5000000 reps @   0.0012 msec (818000.0/sec): 1-pixel solid circle
 5000000 reps @   0.0000 msec (20900000.0/sec): 1-pixel solid circle
 5000000 reps @   0.0000 msec (32300000.0/sec): 1-pixel solid circle
25000000 trep @   0.0010 msec (979000.0/sec): 1-pixel solid circle
    1000 reps @  31.6607 msec (    31.6/sec): Fill 10x10 trapezoid
    1000 reps @  16.9012 msec (    59.2/sec): Fill 10x10 trapezoid
    1000 reps @   0.0234 msec ( 42700.0/sec): Fill 10x10 trapezoid
    1000 reps @   0.0272 msec ( 36700.0/sec): Fill 10x10 trapezoid
    1000 reps @  32.1268 msec (    31.1/sec): Fill 10x10 trapezoid
    5000 trep @  16.1479 msec (    61.9/sec): Fill 10x10 trapezoid
Quoted 7 lines

There are no messages in kernel log or Xorg log.

Visually, most of the tests runs fast as expected, but in some of the tests it gets "stuck" randomly. The test suddenly freezes while drawing the pattern. Then, after a while (a few seconds up to a couple of minutes), it gets unstuck by itself, draws a few more patterns extremely fast for less than a second, and then it gets stuck again.

It smells like a signalling issue. Can you check the CPU activity and report any outstanding process? Can you maybe connect a gdb to the x11perf to see where it's blocked, and post the backtrace here (or in the gitlab issue)?

Regards,

Boris

[1]https://gitlab.freedesktop.org/mesa/mesa

#2 lore raw

I wonder if you're experiencing the same issue as this?

https://gitlab.freedesktop.org/mesa/mesa/-/work_items/15116

Quoted 70 lines · Boris Brezillon
Quoted 4 lines · On Fri, 26 Jun 2026 14:50:17 +0300 Marius Dinu <[email protected]> wrote:

Hi.

The system is Radxa Rock 5 ITX, RK3588, 32GB RAM,
mainline kernel v7.1.0 with panthor driver, mesa 26.0.7 with panfrost

driver,

Quoted 1 line

xorg-server v21.1.22 accelerated by modesetting driver, x11perf v1.7.0.

Are we sure that's a panthor issue? Also, is this a regression introduced by the latest kernel version, or has it been like that from the start?

Also, would you mind opening an issue on [1]?

Quoted 32 lines

The most clean way to reproduce this bug:
Form a console or ssh session, start Xorg in background, without a window manager, and then run x11perf on that display.

Xorg :0 &
x11perf -all -display :0

The bug: weird intermittent performance issues in some of the tests. Examples:

    1000 reps @  18.0799 msec (    55.3/sec): 10x10 rectangle outline
    1000 reps @  18.0791 msec (    55.3/sec): 10x10 rectangle outline
    1000 reps @  18.0790 msec (    55.3/sec): 10x10 rectangle outline
    1000 reps @  18.0789 msec (    55.3/sec): 10x10 rectangle outline
    1000 reps @   0.0287 msec ( 34800.0/sec): 10x10 rectangle outline
    5000 trep @  14.4691 msec (    69.1/sec): 10x10 rectangle outline
 5000000 reps @   0.0029 msec (345000.0/sec): 1-pixel solid circle
 5000000 reps @   0.0009 msec (1100000.0/sec): 1-pixel solid circle
 5000000 reps @   0.0012 msec (818000.0/sec): 1-pixel solid circle
 5000000 reps @   0.0000 msec (20900000.0/sec): 1-pixel solid circle
 5000000 reps @   0.0000 msec (32300000.0/sec): 1-pixel solid circle
25000000 trep @   0.0010 msec (979000.0/sec): 1-pixel solid circle
    1000 reps @  31.6607 msec (    31.6/sec): Fill 10x10 trapezoid
    1000 reps @  16.9012 msec (    59.2/sec): Fill 10x10 trapezoid
    1000 reps @   0.0234 msec ( 42700.0/sec): Fill 10x10 trapezoid
    1000 reps @   0.0272 msec ( 36700.0/sec): Fill 10x10 trapezoid
    1000 reps @  32.1268 msec (    31.1/sec): Fill 10x10 trapezoid
    5000 trep @  16.1479 msec (    61.9/sec): Fill 10x10 trapezoid
Quoted 3 lines

There are no messages in kernel log or Xorg log.

Visually, most of the tests runs fast as expected, but in some of the

tests

Quoted 1 line

it gets "stuck" randomly. The test suddenly freezes while drawing the

pattern.

Quoted 2 lines

Then, after a while (a few seconds up to a couple of minutes), it gets unstuck by itself, draws a few more patterns extremely fast for less

than a

Quoted 1 line

second, and then it gets stuck again.

It smells like a signalling issue. Can you check the CPU activity and report any outstanding process? Can you maybe connect a gdb to the x11perf to see where it's blocked, and post the backtrace here (or in the gitlab issue)?

Regards,

Boris

[1]https://gitlab.freedesktop.org/mesa/mesa

#3 lore raw

Silly me, I forgot to disable HTML email, so this one bounced for some, resending. Sorry about the duplicate.

I wonder if you're experiencing the same issue as this?

https://gitlab.freedesktop.org/mesa/mesa/-/work_items/15116

Quoted 72 lines · Erik Faye-Lund

I wonder if you're experiencing the same issue as this?

https://gitlab.freedesktop.org/mesa/mesa/-/work_items/15116

Quoted 66 lines · Boris Brezillon
Quoted 5 lines · On Fri, 26 Jun 2026 14:50:17 +0300 Marius Dinu <[email protected]> wrote:

Hi.

The system is Radxa Rock 5 ITX, RK3588, 32GB RAM,
mainline kernel v7.1.0 with panthor driver, mesa 26.0.7 with panfrost driver, xorg-server v21.1.22 accelerated by modesetting driver, x11perf v1.7.0.

Are we sure that's a panthor issue? Also, is this a regression introduced by the latest kernel version, or has it been like that from the start?

Also, would you mind opening an issue on [1]?

Quoted 32 lines

The most clean way to reproduce this bug:
Form a console or ssh session, start Xorg in background, without a window manager, and then run x11perf on that display.

Xorg :0 &
x11perf -all -display :0

The bug: weird intermittent performance issues in some of the tests. Examples:

    1000 reps @  18.0799 msec (    55.3/sec): 10x10 rectangle outline
    1000 reps @  18.0791 msec (    55.3/sec): 10x10 rectangle outline
    1000 reps @  18.0790 msec (    55.3/sec): 10x10 rectangle outline
    1000 reps @  18.0789 msec (    55.3/sec): 10x10 rectangle outline
    1000 reps @   0.0287 msec ( 34800.0/sec): 10x10 rectangle outline
    5000 trep @  14.4691 msec (    69.1/sec): 10x10 rectangle outline
 5000000 reps @   0.0029 msec (345000.0/sec): 1-pixel solid circle
 5000000 reps @   0.0009 msec (1100000.0/sec): 1-pixel solid circle
 5000000 reps @   0.0012 msec (818000.0/sec): 1-pixel solid circle
 5000000 reps @   0.0000 msec (20900000.0/sec): 1-pixel solid circle
 5000000 reps @   0.0000 msec (32300000.0/sec): 1-pixel solid circle
25000000 trep @   0.0010 msec (979000.0/sec): 1-pixel solid circle
    1000 reps @  31.6607 msec (    31.6/sec): Fill 10x10 trapezoid
    1000 reps @  16.9012 msec (    59.2/sec): Fill 10x10 trapezoid
    1000 reps @   0.0234 msec ( 42700.0/sec): Fill 10x10 trapezoid
    1000 reps @   0.0272 msec ( 36700.0/sec): Fill 10x10 trapezoid
    1000 reps @  32.1268 msec (    31.1/sec): Fill 10x10 trapezoid
    5000 trep @  16.1479 msec (    61.9/sec): Fill 10x10 trapezoid
Quoted 7 lines

There are no messages in kernel log or Xorg log.

Visually, most of the tests runs fast as expected, but in some of the tests it gets "stuck" randomly. The test suddenly freezes while drawing the pattern. Then, after a while (a few seconds up to a couple of minutes), it gets unstuck by itself, draws a few more patterns extremely fast for less than a second, and then it gets stuck again.

It smells like a signalling issue. Can you check the CPU activity and report any outstanding process? Can you maybe connect a gdb to the x11perf to see where it's blocked, and post the backtrace here (or in the gitlab issue)?

Regards,

Boris

[1]https://gitlab.freedesktop.org/mesa/mesa

#5 lore raw
In reply to Boris Brezillon
Quoted 5 lines · Boris Brezillon
Quoted 2 lines · On Fri, 26 Jun 2026 14:50:17 +0300 Marius Dinu <[email protected]> wrote:

Hi.

Quoted 7 lines
Quoted 3 lines

The system is Radxa Rock 5 ITX, RK3588, 32GB RAM,
mainline kernel v7.1.0 with panthor driver, mesa 26.0.7 with panfrost driver, xorg-server v21.1.22 accelerated by modesetting driver, x11perf v1.7.0.

Are we sure that's a panthor issue? Also, is this a regression introduced by the latest kernel version, or has it been like that from the start?

It was from the start, not a regression. I think I started from v6.18 on this board IIRC. If you need me to test earlier versions, say so.

Quoted 3 lines

Also, would you mind opening an issue on [1]?

I already did.
https://gitlab.freedesktop.org/mesa/mesa/-/work_items/15721

Quoted 40 lines
Quoted 32 lines

The most clean way to reproduce this bug:
Form a console or ssh session, start Xorg in background, without a window manager, and then run x11perf on that display.

Xorg :0 &
x11perf -all -display :0

The bug: weird intermittent performance issues in some of the tests. Examples:

    1000 reps @  18.0799 msec (    55.3/sec): 10x10 rectangle outline
    1000 reps @  18.0791 msec (    55.3/sec): 10x10 rectangle outline
    1000 reps @  18.0790 msec (    55.3/sec): 10x10 rectangle outline
    1000 reps @  18.0789 msec (    55.3/sec): 10x10 rectangle outline
    1000 reps @   0.0287 msec ( 34800.0/sec): 10x10 rectangle outline
    5000 trep @  14.4691 msec (    69.1/sec): 10x10 rectangle outline
 5000000 reps @   0.0029 msec (345000.0/sec): 1-pixel solid circle
 5000000 reps @   0.0009 msec (1100000.0/sec): 1-pixel solid circle
 5000000 reps @   0.0012 msec (818000.0/sec): 1-pixel solid circle
 5000000 reps @   0.0000 msec (20900000.0/sec): 1-pixel solid circle
 5000000 reps @   0.0000 msec (32300000.0/sec): 1-pixel solid circle
25000000 trep @   0.0010 msec (979000.0/sec): 1-pixel solid circle
    1000 reps @  31.6607 msec (    31.6/sec): Fill 10x10 trapezoid
    1000 reps @  16.9012 msec (    59.2/sec): Fill 10x10 trapezoid
    1000 reps @   0.0234 msec ( 42700.0/sec): Fill 10x10 trapezoid
    1000 reps @   0.0272 msec ( 36700.0/sec): Fill 10x10 trapezoid
    1000 reps @  32.1268 msec (    31.1/sec): Fill 10x10 trapezoid
    5000 trep @  16.1479 msec (    61.9/sec): Fill 10x10 trapezoid
Quoted 7 lines

There are no messages in kernel log or Xorg log.

Visually, most of the tests runs fast as expected, but in some of the tests it gets "stuck" randomly. The test suddenly freezes while drawing the pattern. Then, after a while (a few seconds up to a couple of minutes), it gets unstuck by itself, draws a few more patterns extremely fast for less than a second, and then it gets stuck again.

Quoted 5 lines

It smells like a signalling issue. Can you check the CPU activity and report any outstanding process? Can you maybe connect a gdb to the x11perf to see where it's blocked, and post the backtrace here (or in the gitlab issue)?

x11perf is waiting on I/O (red "D" in htop). CPU is completely idle, not even grey "io-wait" activity.

I have no experience using gdb. Please give me the exact command to run.

Quoted 5 lines

Regards,

Boris

[1]https://gitlab.freedesktop.org/mesa/mesa

Marius

#4 lore raw
In reply to Erik Faye-Lund
Quoted 7 lines · Erik Faye-Lund

Silly me, I forgot to disable HTML email, so this one bounced for some, resending. Sorry about the duplicate.

I wonder if you're experiencing the same issue as this?

https://gitlab.freedesktop.org/mesa/mesa/-/work_items/15116

I have no ideea. That bug report is way above my understanding. Sorry.

Quoted 63 lines
Quoted 63 lines · Boris Brezillon
Quoted 61 lines
Quoted 5 lines · On Fri, 26 Jun 2026 14:50:17 +0300 Marius Dinu <[email protected]> wrote:

Hi.

The system is Radxa Rock 5 ITX, RK3588, 32GB RAM,
mainline kernel v7.1.0 with panthor driver, mesa 26.0.7 with panfrost driver, xorg-server v21.1.22 accelerated by modesetting driver, x11perf v1.7.0.

Are we sure that's a panthor issue? Also, is this a regression introduced by the latest kernel version, or has it been like that from the start?

Also, would you mind opening an issue on [1]?

Quoted 32 lines

The most clean way to reproduce this bug:
Form a console or ssh session, start Xorg in background, without a window manager, and then run x11perf on that display.

Xorg :0 &
x11perf -all -display :0

The bug: weird intermittent performance issues in some of the tests. Examples:

    1000 reps @  18.0799 msec (    55.3/sec): 10x10 rectangle outline
    1000 reps @  18.0791 msec (    55.3/sec): 10x10 rectangle outline
    1000 reps @  18.0790 msec (    55.3/sec): 10x10 rectangle outline
    1000 reps @  18.0789 msec (    55.3/sec): 10x10 rectangle outline
    1000 reps @   0.0287 msec ( 34800.0/sec): 10x10 rectangle outline
    5000 trep @  14.4691 msec (    69.1/sec): 10x10 rectangle outline
 5000000 reps @   0.0029 msec (345000.0/sec): 1-pixel solid circle
 5000000 reps @   0.0009 msec (1100000.0/sec): 1-pixel solid circle
 5000000 reps @   0.0012 msec (818000.0/sec): 1-pixel solid circle
 5000000 reps @   0.0000 msec (20900000.0/sec): 1-pixel solid circle
 5000000 reps @   0.0000 msec (32300000.0/sec): 1-pixel solid circle
25000000 trep @   0.0010 msec (979000.0/sec): 1-pixel solid circle
    1000 reps @  31.6607 msec (    31.6/sec): Fill 10x10 trapezoid
    1000 reps @  16.9012 msec (    59.2/sec): Fill 10x10 trapezoid
    1000 reps @   0.0234 msec ( 42700.0/sec): Fill 10x10 trapezoid
    1000 reps @   0.0272 msec ( 36700.0/sec): Fill 10x10 trapezoid
    1000 reps @  32.1268 msec (    31.1/sec): Fill 10x10 trapezoid
    5000 trep @  16.1479 msec (    61.9/sec): Fill 10x10 trapezoid
Quoted 7 lines

There are no messages in kernel log or Xorg log.

Visually, most of the tests runs fast as expected, but in some of the tests it gets "stuck" randomly. The test suddenly freezes while drawing the pattern. Then, after a while (a few seconds up to a couple of minutes), it gets unstuck by itself, draws a few more patterns extremely fast for less than a second, and then it gets stuck again.

It smells like a signalling issue. Can you check the CPU activity and report any outstanding process? Can you maybe connect a gdb to the x11perf to see where it's blocked, and post the backtrace here (or in the gitlab issue)?

Quoted 5 lines
Quoted 5 lines
Quoted 5 lines

Regards,

Boris

[1]https://gitlab.freedesktop.org/mesa/mesa

Marius