Bug 214425 - [drm][amdgpu][TTM] Page pool memory never gets freed
Summary: [drm][amdgpu][TTM] Page pool memory never gets freed
Status: NEW
Alias: None
Product: Drivers
Classification: Unclassified
Component: Video(DRI - non Intel) (show other bugs)
Hardware: x86-64 Linux
: P1 normal
Assignee: drivers_video-dri
URL:
Keywords:
Depends on:
Blocks:
 
Reported: 2021-09-15 21:09 UTC by Martin Doucha
Modified: 2022-10-15 17:18 UTC (History)
3 users (show)

See Also:
Kernel Version: 5.14.3
Subsystem:
Regression: No
Bisected commit-id:


Attachments

Description Martin Doucha 2021-09-15 21:09:55 UTC
Hello,
while playing certain webGL games, I've noticed what appears to be a memory leak in the kernel. Further investigation revealed that after about an hour of gameplay, over 3GB of memory (half of all available RAM on my machine) will be taken by the TTM page pool.

While the excessive allocation may be caused by a resource leak in the game itself (I need to investigate that further), the larger problem is that TTM never releases the memory even after I quit the game. Closing the game only moves the allocated memory from active buffer objects to idle memory pool where it'll get stuck until I reboot the system. Shutting down X server doesn't release the memory either.

System specs:
HP Probook 455 G7
AMD Ryzen 5 4500U CPU
AMD Renoir GPU (Mesa 21.2.1, LLVM 12.0)
Gentoo Linux

TTM statistics before quitting the game:
/sys/kernel/debug/ttm/buffer_objects:
3116

/sys/kernel/debug/ttm/page_pool:
          --- 0--- --- 1--- --- 2--- --- 3--- --- 4--- --- 5--- --- 6--- --- 7--- --- 8--- --- 9--- ---10---
wc      :        2        2        1        1        8        2        0        1        2        1        2
uc      :        0        0        0        0        0        0        0        0        0        0        0
wc 32   :        0        0        0        0        0        0        0        0        0        0        0
uc 32   :        0        0        0        0        0        0        0        0        0        0        0

total   :     3410 of   939433

/sys/kernel/debug/ttm/page_pool_shrink:
2898/512


=======================================

TTM statistics after quitting the game (until reboot):
/sys/kernel/debug/ttm/buffer_objects:
403

/sys/kernel/debug/ttm/page_pool:
          --- 0--- --- 1--- --- 2--- --- 3--- --- 4--- --- 5--- --- 6--- --- 7--- --- 8--- --- 9--- ---10---
wc      :      151      134       20        5      255      241      790      193      416     1121       83
uc      :        0        0        0        0        0        0        0        0        0        0        0
wc 32   :        0        0        0        0        0        0        0        0        0        0        0
uc 32   :        0        0        0        0        0        0        0        0        0        0        0

total   :   853035 of   939433

/sys/kernel/debug/ttm/page_pool_shrink:
853034/1
Comment 1 Rafael Ristovski 2022-10-15 14:34:11 UTC
According to amdgpu devs, this is a feature where the allocated pages are kept around in case they are needed later on. TTM is able to release the memory in case the memory pressure increases.

See comment here: https://gitlab.freedesktop.org/drm/amd/-/issues/1942#note_1311016
Comment 2 Martin Doucha 2022-10-15 14:47:05 UTC
(In reply to Rafael Ristovski from comment #1)
> According to amdgpu devs, this is a feature where the allocated pages are
> kept around in case they are needed later on. TTM is able to release the
> memory in case the memory pressure increases.

I understand the logic behind keeping idle buffers allocated for a while. But it does not make sense to keep them for hours after last use and the release mechanism on increased memory pressure does not seem to be working.

When I run a large compilation overnight, starting from a fresh reboot and shutting down all graphics software including the X server, I'll often come back in the morning to find that 70% of all RAM is allocated in idle TTM buffers and GCC is stuck swapping for hours. The TTM buffers were likely allocated by some GPU-accelerated build computation halfway through the night. But this is harder to reproduce than the games I've mentioned in the initial bugreport.
Comment 3 Rafael Ristovski 2022-10-15 15:05:38 UTC
(In reply to Martin Doucha from comment #2)
> (In reply to Rafael Ristovski from comment #1)
> > According to amdgpu devs, this is a feature where the allocated pages are
> > kept around in case they are needed later on. TTM is able to release the
> > memory in case the memory pressure increases.
> 
> I understand the logic behind keeping idle buffers allocated for a while.
> But it does not make sense to keep them for hours after last use and the
> release mechanism on increased memory pressure does not seem to be working.
> 
> When I run a large compilation overnight, starting from a fresh reboot and
> shutting down all graphics software including the X server, I'll often come
> back in the morning to find that 70% of all RAM is allocated in idle TTM
> buffers and GCC is stuck swapping for hours. The TTM buffers were likely
> allocated by some GPU-accelerated build computation halfway through the
> night. But this is harder to reproduce than the games I've mentioned in the
> initial bugreport.

Indeed, I too run into situations where even if I purposefully trigger an OOM situation just to get the TTM "cache" to evict itself through memory pressure, _it still does not end up releasing all of the memory_.

There are also the following two sysfs files, simply reading them triggers an eviction of GTT/VRAM:
> cat /sys/kernel/debug/dri/0/amdgpu_evict_vram
> cat /sys/kernel/debug/dri/0/amdgpu_evict_gtt

this can be confirmed as working with tools like `radeontop`/`nvtop`.

However, this once again does not release the TTM buffers.

As you can see in the issue I linked, I never got a reply about a mechanism to manually release TTM memory. I will attempt coercing an answer on IRC, perhaps I will have better luck asking directly there.
Comment 4 Rafael Ristovski 2022-10-15 16:14:02 UTC
For what its worth, the following horrible incantation managed to release 2+GB of TTM buffers on one of my machines, after I purposefully ran a VRAM intensive game:
> for i in {1..1000}; do cat /sys/kernel/debug/ttm/page_pool_shrink; done

This seems to be the only sysfs mechanism to cause the memory to get released, and as of now I am not aware of a... better and mainly "cleaner" alternative.

Newer kernel versions seem to feature https://www.kernel.org/doc/html/next/admin-guide/mm/shrinker_debugfs.html, which might be a better alternative, but I have not tested it yet, and its usage is not exactly clear to me.

Note You need to log in before you can comment on or make changes to this bug.