Philipp Lutz

Results 76 comments of Philipp Lutz

> Do you see any error at dmesg from amdgpu? Ups, yes, I forgot to add that to the issue description, sorry. I just added it at the end.

> NVM, I see this is darktable. > > Wonder if related to [#5590](https://github.com/ROCm/ROCm/issues/5590) I'm not sure, the symptom (page fault / illegal mem access) might be the same, but...

I think the culprit is the wrong reporting of available GPU memory within the OpenCL platform, I added the `clinfo` output in my first post and here are the reported...

> [@da-phil](https://github.com/da-phil) what do you see with `$ sudo dmesg | grep amdgpu | grep memory` ? .. it might be easier to let amdgpu handle the allocation completely. Since...

> [@da-phil](https://github.com/da-phil) what do you see with `$ sudo dmesg | grep amdgpu | grep memory` ? .. it might be easier to let amdgpu handle the allocation completely. Since...

After the reboot I'm getting the values I set before: ``` > amd-ttm 💻 Current TTM pages limit: 1048576 pages (4.00 GB) 💻 Total system memory: 27.19 GB ``` Checking...

> Hi [@da-phil](https://github.com/da-phil), yes I agree with [@ianbmacdonald](https://github.com/ianbmacdonald) that you should let TTM to manage instead of configuring carveout in BIOS. Thanks for your reply, let me follow-up with some...

Only today I encountered another page fault which rendered my Ubuntu desktop completely unusable and required a reboot. This one however was not the typical "Memory access fault by GPU"...

Here is another kernel log file after a page fault with the classical "Memory access fault by GPU node-1" error (filtered only for amdgpu msgs): [another-page-fault.log](https://github.com/user-attachments/files/23697145/another-page-fault.log)

@amd-nicknick thanks a lot for your elaborate answer, this actually helps a lot to understand the merits of the different memory mapping mechanisms going on! Will try out the module...