08:10 MrCooper: mupuf: it was opt-in via AMDGPU_GEM_CREATE_VRAM_CLEARED until recently, a "please don't leak information to me" flag never really made sense though :)
08:11 mupuf: Right! that's what I.remembered
08:11 mupuf: thanks for taking the time to educate me 😅
08:13 MrCooper: no worries
14:17 Venemo: DemiMarie: it's interesting that you're also talking about gpu resets. there has been a lot of talk about it recently, as well as some work in progress patches sent to the mailing lists.
14:18 Venemo: DemiMarie: consider looking into the enforce_isolation option and Christian's recent series to improve it (note that it's still a work in progress though).
14:21 Venemo: ideally, with that option (the fixed version), a GPU hang shouldn't bring down the whole system anymore.
14:21 Venemo: though, as said above, you'll lose some performance. it's impossible to say how much because there are a lot of factors it depends on
21:13 DemiMarie: Venemo: will hard hangs no longer require PCI resets?
21:26 Venemo: DemiMarie: yes, that is how I understand it. but it only applies to cases that can be resolved by performing a queue reset.
21:28 Venemo: DemiMarie: (as a side note, they already didn't require PCI resets, as the PCIe interface remains active while the chip is being reset. the real issue why that destroys the system is that dGPUs lose the contents of VRAM after that kind of reset.)