si_copy_dma(9)

📦 version 7.3.1, in nixpkgs unstable

💅 si_copy_dma(9) — or, how to copy pages with the DMA engine

okay bestie, welcome to the kernel API manual, this one's si_copy_dma(9) and honestly it's giving "small function, big responsibility" ✨

🎀 SYNOPSIS

struct radeon_fence * si_copy_dma (struct radeon_device *rdev , uint64_t
src_offset , uint64_t dst_offset , unsigned num_gpu_pages , struct
dma_resv *resv );

that's the whole vibe, copy it exactly, no cap 🫶

🌸 ARGUMENTS

so here's who's showing up to the function party:

  • rdev — a radeon_device pointer. it's the main character of this whole thing, literally the device we're working with 💖
  • src_offset — the source GPU address. like, where the pages are coming from 🦋
  • dst_offset — the destination GPU address. where they're going to 🎀
  • num_gpu_pages — number of GPU pages to xfer. yes, "xfer", that's the manual's word, keeping it 🙈 this is literally how many pages get moved
  • resv — reservation object to sync to. the sync bestie, don't forget it 🌸

✨ DESCRIPTION

okay so the actual tea: this copies GPU paging using the DMA engine (SI). that's it, that's the job 🔥

and who uses it? the radeon ttm implementation, to move pages — but only if it's registered as the asic copy callback. conditional queen 👑 so like, no registration, no pages getting moved, iykyk.

💜 SEE ALSO

go check out the kernel file ./drivers/gpu/drm/radeon/si_dma.c and these iconic siblings:

family reunion fr fr 🌈

ngl this one's short but it eats. every argument matters, so pass them right or the vibe check fails 💀

mwah, that's a wrap 🫶 this came from ./drivers/gpu/drm/radeon, dated October 2026, and it goes by si_copy_dma(9) — periodt 💅✨

📦 where this page comes from