Public bug reported:

On an Arduino VENTUNO Q, leaving the GNOME login screen idle causes
shared and unevictable memory to grow continuously. The msm DRM GEM
buffer totals grow with it. I reproduced this with App Lab closed, no
running Docker containers, and no model server or inference process
running.

Expected: memory use settles while the login screen is idle.
Actual: pinned graphics buffers keep accumulating. During an earlier unattended 
boot, available RAM fell from roughly 12.8 GiB to 476 MiB over three hours, and 
later reached 174 MiB. SSH eventually stopped responding, and I manually 
rebooted. The logs do not prove that the GPU allocation growth alone caused the 
SSH failure.

System:
- Arduino VENTUNO Q, 16 GB RAM, Arduino Ubuntu 24.04 image updated with apt 
(currently reports Ubuntu 24.04.5 LTS).
- linux-image-6.8.0-1084-qcom 6.8.0-1084.89.
- libgl1-mesa-dri and libegl-mesa0 25.2.8-0ubuntu0.24.04.4.
- libmutter-14-0 46.2-1ubuntu0.24.04.18.
- gnome-shell 46.0-0ubuntu6~24.04.16; gdm3 46.2-1ubuntu1~24.04.9.
- linux-firmware-dragonwing 20260612.
- Monitor attached, GDM using Wayland, no swap configured.

Reproduction on 29 October, after a fresh boot:
1. Leave the board at the login screen, with no desktop user logged in.
2. Read /proc/meminfo and /sys/kernel/debug/dri/0/gem at 15-second intervals.

The following samples are from the same boot. Memory fields are in KiB,
GEM totals in bytes:

UTC time  Shmem   Unevictable  GEM objects  GEM total
14:32:43  463032  474836       299          462893056
14:32:58  483608  495412       302          483962880
14:33:13  514424  526228       306          515518464
14:33:28  534964  546768       309          536588288

The growing shmem/unevictable allocation was charged to the gdm user's
cgroup (user-123.slice), while its anonymous memory stayed around 216
MiB.

Stopping GDM released the buffers immediately:
- Before stop: 309 GEM objects, 536588288 bytes; Shmem 534996 KiB; Unevictable 
546804 KiB.
- After stop: 10 GEM objects, 25169920 bytes; Shmem 27944 KiB; Unevictable 
51468 KiB.
- Available RAM increased from 13826392 to 14587720 KiB.

Restarting GDM brought the growth back.

Workaround tested:
Create /etc/dconf/db/gdm.d/99-ventuno-memory-workaround with:

[org/gnome/desktop/interface]
enable-animations=false

Then run sudo dconf update and restart GDM. Restarting GDM ends
graphical sessions, so I did this only while the board was at the login
screen.

With animations disabled, the GEM total stayed at 259 objects /
219639808 bytes, Shmem around 225492 KiB, and Unevictable at 237284 KiB
across repeated samples over approximately 2.5 minutes. This is a short
controlled test, not an overnight stability result.

The previous boot logged 4,117 instances of:
MESA: error: flush_submit_list:135: submit failed: -28 (No space left on device)

It also logged msm/a6xx GMU HFI_H2F_MSG_GX_BW_PERF_VOTE response
timeouts. There was no recorded OOM kill, kernel panic, or filesystem
I/O error, and pstore was empty. The ENOSPC above came from GPU
submission, not a full root filesystem.

The board has a local REFGEN device-tree backport and
pcie_aspm.policy=performance from a separate NVMe investigation. That
change adds the regulator at 0x0891c000 and refgen-supply links to both
PCIe PHYs. The M.2 slot was empty during this memory test. I have not
repeated the memory test with the stock device tree, so I cannot rule
out that modification. NVMe report:
https://bugs.launchpad.net/ubuntu/+source/linux-qcom/+bug/2172593

I am filing against linux-qcom because the measured growth is in msm DRM
buffers, but I have not isolated whether the defect is in the kernel
driver, Mesa, or Mutter/GDM. Please retarget if another package is
appropriate.

** Affects: linux-qcom (Ubuntu)
     Importance: Undecided
         Status: New

-- 
You received this bug notification because you are a member of Ubuntu
Bugs, which is subscribed to Ubuntu.
https://bugs.launchpad.net/bugs/2172594

Title:
  VENTUNO Q: idle GDM login screen accumulates pinned msm GPU buffers
  until memory is exhausted

To manage notifications about this bug go to:
https://bugs.launchpad.net/ubuntu/+source/linux-qcom/+bug/2172594/+subscriptions


-- 
ubuntu-bugs mailing list
[email protected]
https://lists.ubuntu.com/mailman/listinfo/ubuntu-bugs

Reply via email to