#asahi-dev · 2026-08-18 20:22
sven+what#asahi-dev · 2026-08-18 20:21
svenand dma-ranges did you use?#asahi-dev · 2026-08-18 20:19
svenaren't you stuck with using the iommu api then?#asahi-dev · 2026-08-18 20:18
jannaumaybe if the driver creates a domain and doesn't use the default domain?#asahi-dev · 2026-08-18 20:18
svenreserved regions? though that would be a big hack if that worked :D#asahi-dev · 2026-08-18 20:17
svenhm, there's no way to limit the iova aperture somehow when dma-iommu is used outside of the iommu driver, is there?#asahi-dev · 2026-08-18 20:16
jannauI think it is not fully correct because 0x00x_xxxx_xxxx would be valid IOVA as well with that mask#asahi-dev · 2026-08-18 20:13
svenit's to encode that weird 0x1000... + 36bit iova space#asahi-dev · 2026-08-18 20:13
svennvm#asahi-dev · 2026-08-18 20:11
sventhat 1ULL<<40 | DMA_BIT_MASK(36) looks odd though#asahi-dev · 2026-08-18 20:11
sveni've long given up on the idea that we can support M(n+1) with just device tree updates anyway#asahi-dev · 2026-08-18 20:11
jannauso we would need to specify the dma_mask has hardware data in the device table#asahi-dev · 2026-08-18 20:09
jannauit works but requires `dma_set_mask_and_coherent(&pdev->dev, (1ULL << 40) | DMA_BIT_MASK(36));`#asahi-dev · 2026-08-18 20:06
svenpity :(#asahi-dev · 2026-08-18 20:05
jannauit does not work#asahi-dev · 2026-08-18 19:59
svenyes, but i would enjoy this very much because it strengthens the nhi as child of acio argument#asahi-dev · 2026-08-18 19:59
jannaulet me check. it would be a little annoying since it's a solution that only works for tbt#asahi-dev · 2026-08-18 19:57
svenlet me check if we can just use dma-ranges here since we already have that parent node anyway#asahi-dev · 2026-08-18 19:47
jannaunot sure how everyone feels about inscreasing #iommu-cells#asahi-dev · 2026-08-18 19:47
jannauagreed, it feels like it should belong to the node which uses the dart#asahi-dev · 2026-08-18 19:44
svenI really don’t want to upstream that property because it feels wrong on the dart node #asahi-dev · 2026-08-18 19:43
jannauso "apple,dma-range = <0x100 0x0 0x1 0x0>;" seems to work#asahi-dev · 2026-08-18 19:42
jannauiova are as expected but it looks like acio has only a 36-bit address space (which still starts at 0x100_0000_0000#asahi-dev · 2026-08-18 19:30
sveni dunno if the iova code respects that#asahi-dev · 2026-08-18 19:30
svenyou might be able to hack this with dma-ranges in the parent node as well :D#asahi-dev · 2026-08-18 19:30
svenit probably faults once the first entry is added to that ring#asahi-dev · 2026-08-18 19:30
svenring_iowrite64desc(ring, ring->descriptors_dma, 0); is probably the iova#asahi-dev · 2026-08-18 19:29
svenyou can add prints to tb_ring_start#asahi-dev · 2026-08-18 19:25
jannauno, I suspect linux uses 0xeffff8000 and acio somehow offsets it by vm-base (0x100_0000_0000) anyway#asahi-dev · 2026-08-18 19:22
svenbut that should still be in that range, shouldn't it?#asahi-dev · 2026-08-18 19:22
svenhm, true, it at least used to allocate top-down#asahi-dev · 2026-08-18 19:22
jannauwith "apple,dma-range = <0x100 0x0 0x2ff 0x0>;"#asahi-dev · 2026-08-18 19:22
jannauiirc it allocates top down so I would expect iovas of the form 0x3fe_xxxx_xxxx#asahi-dev · 2026-08-18 19:20
sveni suspected that the tbt code was never tested with iovas >32bit#asahi-dev · 2026-08-18 19:19
jannaunot sure linux' dma iommu code does the right thing#asahi-dev · 2026-08-18 19:18
jannau0x10effff8000 and 0x10effffc000#asahi-dev · 2026-08-18 19:17
svenwhat's the iova that fails?#asahi-dev · 2026-08-18 19:16
jannaustill fails with that (and set_dma_mask_and_coherent(DMA_BIT_MASK(42)))#asahi-dev · 2026-08-18 18:57
jannauno, it probably needs "iommu: apple-dart: Support specifying the DMA aperture in the DT"#asahi-dev · 2026-08-18 18:54
svendoes the iova it faults on look like something linux would've allocated?#asahi-dev · 2026-08-18 18:52
svenugh, some t602x thing then#asahi-dev · 2026-08-18 18:52
jannaut6001 still works#asahi-dev · 2026-08-18 18:45
svenhmmmm#asahi-dev · 2026-08-18 18:45
chaos_princessno, that is i think m4 pro/max and m3 ultra#asahi-dev · 2026-08-18 18:44
svendoes that already have tbt5?#asahi-dev · 2026-08-18 18:44
svencould also be a t602x thing, sure