#asahi · 2025-10-30 12:26
monthebiffseparate hardware /cache?#asahi · 2025-10-30 12:26
monthebiff*cries*#asahi · 2025-10-30 12:26
chaos_princessno#asahi · 2025-10-30 12:26
monthebiffiommu has their own page tables... they all sit in memory as part of 'the page tables'?#asahi · 2025-10-30 12:24
j`eymonthebiff: iommus have their own page tables#asahi · 2025-10-30 12:24
chaos_princess"sometimes devices can accessed via memory addresse" - s/sometimes/always/ for !x86, s/sometimes/almost always/ for x86#asahi · 2025-10-30 12:23
monthebiffyeh im reading about different types of page tables and its all very abstract. im sure its simplified to the point where im not getting the whole picture anyway.#asahi · 2025-10-30 12:22
chaos_princessbut hey, i said it was a bad diagram#asahi · 2025-10-30 12:22
chaos_princessand i guess there is also a mistake in that diagram in that each core can have a different set of page tables#asahi · 2025-10-30 12:22
chaos_princessyes#asahi · 2025-10-30 12:21
monthebiffso page tables are separate from iommu? (the stuff im currently reading talks about address translation happening for page tables, but im also aware that sometimes devices can accessed via memory addresses. thought the iommu unit(s) might be a place where all that coexists and is managed)#asahi · 2025-10-30 12:19
monthebiffthank you x #asahi · 2025-10-30 12:19
chaos_princessvery bad diagram https://pastebin.com/H7vJYfFc#asahi · 2025-10-30 12:13
chaos_princess(mostly because it is a network in some sense)#asahi · 2025-10-30 12:13
monthebiffand multiple iommu systems can access the same memory by being part of the same iommu groups?#asahi · 2025-10-30 12:13
chaos_princessthink of it as a network with memory/cpu/gpu sending requests between each other#asahi · 2025-10-30 12:12
monthebiffso its a single memory pool, being accessed by different hardware via their respective iommu systems all inside the SoC?#asahi · 2025-10-30 12:10
chaos_princesshere it is shared, on x86 - there are two pools, but gpu and cpu can access their "non-local" pool too#asahi · 2025-10-30 12:10
monthebiffi think#asahi · 2025-10-30 12:10
monthebiffcool#asahi · 2025-10-30 12:10
svenon apple silicon: what j`ey said#asahi · 2025-10-30 12:10
j`ey / unified#asahi · 2025-10-30 12:10
svenon x86 they usually have their dedicated memory, yes#asahi · 2025-10-30 12:10
j`eyit's a shared memory system#asahi · 2025-10-30 12:10
monthebiffand if this is a particularly dense question... i always thought GPU's shipped with their own dedicated memory? I'm guessing from the above that actually the GPU has direct access to system memory?#asahi · 2025-10-30 12:10
svenon apple silicon there's usually a single device in each group, on x86 you can have multiple devices (especially on the pci bus) in a single group#asahi · 2025-10-30 12:09
svenwhat you want to look for are iommu group assignments. each iommu group shares the same virtual address space#asahi · 2025-10-30 12:08
chaos_princessthere is a whole thing called "acs override" where a platform does not have enough iommus and people do some hacky thing to do what they want to do#asahi · 2025-10-30 12:08
monthebiff*makes a note to do that*#asahi · 2025-10-30 12:07
chaos_princessin general - run lspci and see what happens :P#asahi · 2025-10-30 12:07
monthebiffand more generally, in a none apple system with a GPU available somewhere else on the board via the bus, same question... if theres a general answer?#asahi · 2025-10-30 12:07
chaos_princessnot always the case on x86, as they can have a single iommu for a bunch of devices at the same time#asahi · 2025-10-30 12:06
chaos_princesson AS, there are dedicated iommus in front of every dma-capable hw block#asahi · 2025-10-30 12:06
monthebiffforgive my ignorance: this is not stuff familiar to me. Im picturing the CPU on a die with an mmu/iommu controller as part of it. the GPU is also present, yay SoC's? But does that GPU use the same iommu or does it have its own thing inside the die space or internal to the GPU?#asahi · 2025-10-30 12:04
sveneach hardware usually has a separate iommu with a separate virtual address space for DMA#asahi · 2025-10-30 12:04
svenwhat do you mean with "same controller"?#asahi · 2025-10-30 12:03
monthebiffsame controller? or 1 controller per hardware with its own address mapping stuff going on?#asahi · 2025-10-30 12:03
svenlooks like a quick google search confirms that m1 actually only supports 4k and 16k pages#asahi · 2025-10-30 12:02
svenmost hardware is behind some iommu which also has pagetables for DMA#asahi · 2025-10-30 12:02
monthebiffgpu uses page tables as well? this isn't just a CPU to ram lookup thing?#asahi · 2025-10-30 12:02
svenhm.. maybe i misremember, i thought m1+ allowed 64k pages#asahi · 2025-10-30 12:01
chaos_princessarm64 does, idk if as specifically does#asahi · 2025-10-30 12:01
sveni thought it did#asahi · 2025-10-30 12:01
svenhuh, do i remember that wrong?#asahi · 2025-10-30 12:01
chaos_princessAnd AS supporting 64k is news to me#asahi · 2025-10-30 12:01
chaos_princessit is set by hardware, the page table format is very dependent on exact page size.#asahi · 2025-10-30 12:00
svenwhich effectively means that on linux only 16k and 64k will work#asahi · 2025-10-30 11:59
svenDART (iommu) and the gpu can only do 16k#asahi · 2025-10-30 11:59
sventhe cpu can do 4k, 16k and 64k#asahi · 2025-10-30 11:58
monthebiffpage table question (because learning): is their an apple silicon constraint on page table size? The material im looking at talks about page table implementations within memory as a convention... but I'm not sure where the line is between hardware implementation and something set by the OS/ in an application running as a process. cheers in advance x