MADVISE LOGIC - SPLITTING AND TRAVERSAL(9)

📦 version 7.3.1, in nixpkgs unstable

💅 MADVISE LOGIC — SPLITTING AND TRAVERSAL 💅

okay bestie, this is the Madvise Logic - Splitting and Traversal page from the Kernel API Manual, and it's literally the glow-up your GPU VA ranges have been waiting for ✨

This logic handles GPU VA range updates by generating remap and map operations without performing unmaps or merging existing mappings. No deleting, no smooshing together — just clean remaps and maps, periodt 🎀

🌸 THE FIVE SITUATIONS 🌸

There are exactly five (5) flavor moments this logic can hit, and honestly? Each one ate 💅

1️⃣ everything lives inside ONE drm_gpuva

The requested range lies entirely within a single drm_gpuva. The logic splits the existing mapping at the start and end boundaries and inserts a new map.

                    a      start    end     b
               pre: |-----------------------|
                           drm_gpuva1

                    a      start    end     b
               new: |-----|=========|-------|
                     remap   map      remap

one REMAP and one MAP — Same behaviour as SPLIT and MERGE 🫶

2️⃣ the range spans MULTIPLE drm_gpuva regions

The requested range spans multiple drm_gpuva regions. The logic traverses across boundaries, remapping the start and end segments, and inserting two map operations to cover the full range.

       ::           a       start      b              c        end       d pre:
       |------------------|--------------|------------------| drm_gpuva1
       drm_gpuva2         drm_gpuva3

       a       start      b              c        end       d new:
       |-------|==========|--------------|========|---------| remap1   map1
       drm_gpuva2    map2     remap2

two REMAPS and two MAPS — she's doing the most and we love her for it 😭🔥

3️⃣ just ONE side lands inside

Either start or end lies within a drm_gpuva. A single remap and map operation are generated to update the affected portion.

First scenario — start is the one inside:

       ::           a/start            b              c        end       d pre:
       |------------------|--------------|------------------| drm_gpuva1
       drm_gpuva2         drm_gpuva3

       a/start            b              c        end       d new:
       |------------------|--------------|========|---------| drm_gpuva1
       drm_gpuva2     map1     remap1

Second scenario — end is the one inside:

       ::           a       start      b              c/end              d pre:
       |------------------|--------------|------------------| drm_gpuva1
       drm_gpuva2         drm_gpuva3

       a       start      b              c/end              d new:
       |-------|==========|--------------|------------------| remap1   map1
       drm_gpuva2        drm_gpuva3

one REMAP and one MAP — minimal effort, maximum slay 💅✨

4️⃣ both sides already line up

Both start and end align with existing drm_gpuva boundaries. No operations are needed as the range is already covered.

She showed up already dressed, nothing to do here, iconic 👑

5️⃣ no existing drm_gpuvas at all

No existing drm_gpuvas. No operations. That's it, that's the whole mood 💀

🦋 THE WHOLE POINT 🦋

Unlike drm_gpuvm_sm_map_ops_create, this logic avoids unmaps and merging, focusing solely on remap and map operations for efficient traversal and update.

That's the flex, bestie: remap + map only, traversal stays efficient, no unmap-and-rebuild drama 🙈💖

✨ SEE ALSO ✨

Kernel file ./drivers/gpu/drm/drm_gpuvm.c

and that's the page, bestie 🫶 this one shipped from ./drivers/gpu/drm back in October 2026, so go forth and remap responsibly 🌈⭐💜

📦 where this page comes from