10:16 tzimmermann: i have a switcheroo patch that affects nouveau. please review: https://lore.kernel.org/dri-devel/20251105161549.98836-1-tzimmermann@suse.de/
10:32 esdrastarsis[d]: karolherbst: Is there any plans to upstream the kernel patch for nvdec/nvenc engjne?
10:33 x512[m]: karolherbst: How ready is NVK part?
19:59 _lyude[d]: Hm. Is the drm_buddy API supposed to be used for a reusable pool of DMA memory that can be used for things like noncoherent contiguous allocations that almost never fail even in high memory pressure?
20:02 _lyude[d]: Discovered another bug in nouveau where it seems like it's actually possible for us to fail allocating memory (order 5) to use for storing the firmware image we use for shutdown (fwsec-sb). And I'm trying to figure out if we should maybe just keep firmware memory allocated for firmware images we know we'll need later for GPU shutdown so that we don't need to worry about allocations failing later,
20:02 _lyude[d]: or maybe hook up drm_buddy for allocating firmware if it would help?
21:03 airlied[d]: _lyude[d]: no, buddy is just for vram mgmt, does that fw need to be in a single allocation or is there a page table? we could just load it early alright to avoid it
21:07 _lyude[d]: airlied[d]: the way it's currently setup is using the DMA firmware target . so I would assume the answer is it needs to be in a single table unless we just never bothered trying to hook up sgt to `nvkm_gsp_fwsec`
21:08 _lyude[d]: and yeah, considering that (assuming I was looking at the right firmware file) it seems like the uncompressed firmware in question here is only like 36K it probably couldn't hurt to just allocate early
21:13 airlied[d]: I'm no fully awake, but it does appear as it we could use a non-contig allocation there, but also I'd fix it first by just allocating it early and keeping it
21:14 airlied[d]: using non-contig would probably involve rewriting the pio/dma uploader code
21:14 airlied[d]: and maybe it isn't possible
21:16 mhenning[d]: silly question but if we're turning the gpu off then why do we have a ton of memory allocated anyway?
21:17 airlied[d]: you need firmware to turn it off
21:17 mhenning[d]: right, I got that part
21:17 mhenning[d]: but we can free all of the user allocations right
21:17 airlied[d]: for laptops, user might not be using the GPU at all, but it keeps waking up and shutting down
21:18 airlied[d]: eventually one of the shut downs fails and takes out the system
21:18 mhenning[d]: oh, is vram preserved for this type of "shut down"?
21:18 airlied[d]: yes
21:18 mhenning[d]: okay makes sense