[TibixDev]

What

Write tracked Linux direct memory through a retained backing view. Keep access checks, write versions, and GPU ordering. Use the same path for completion labels.

The writer validates the full range before it changes any bytes. Unsupported mappings use the existing copy path.

Related: #1152 reduces GPU import count; this change reduces CPU write faults. Both change PhysicalBacking and need integration. #637 changes Windows tracking.

Tested

Fedora 44, Release build:

• Shared writes, write tracking, and guest memory tests.
• Shared writes with write watching disabled.
• RX 6600 completion-label storage and ordering tests with Vulkan validation. The recorder fixture now enables command-buffer reset.
• Contribution checks.

Isolated comparison: upstream e9846679 versus the same revision plus this PR. Six pinned processes per variant in ABBA order. A 1 MiB write took 285.954 → 68.110 µs (76% less time); process medians ranged from 274.610–365.925 µs before and 64.895–79.676 µs after. Write faults fell from 589,824 to zero over 2,304 measured writes. Smaller writes and labels had overlapping timing ranges. This is a component result, not title frame time. Windows was not tested.

Checklist

• Current main; related PRs and integration needs are described above.
• One topic.
• No new code comments.
• Unsupported paths still throw; UI stubs are documented.
• No notes, logs, or images in the repository.
• No new third-party code.
• General behavior.
• Depends on: none.
• AI-assisted: yes.

[TibixDev]

Rebased on e9846679. The test-list conflict is resolved; the production change is unchanged.

Measured upstream e9846679 against that revision plus this PR (349781da). Fedora 44, GCC 16 Release, Ryzen 5 5600, pinned to CPU 3. Six processes per variant, ABBA × 3; each has one warmup and nine batches of 256 writes. No game or build ran during timing.

Write size	Upstream median (µs)	PR median (µs)	Upstream process range	PR process range
4,096 bytes	50.023	41.315	40.184–57.437	37.069–54.181
65,536 bytes	65.957	46.791	54.190–76.309	37.947–56.446
1,048,576 bytes	285.954	68.110	274.610–365.925	64.895–79.676

The 1 MiB case uses 76% less time. Faults per 2,304 writes fell from 2,304 / 36,864 / 589,824 to zero for the three sizes. Both builds check the output bytes.

For 4/8/16-byte completion stores, medians were 49.860/49.247/40.664 → 40.406/39.346/42.053 µs. All ranges overlap, so this does not establish a small-store speedup; the 16-byte median regressed. The label fixture invokes the same access-check and store sequence as each revision's completion callback. Faults fell from 2,304 to zero in each case.

Shared-write, write-tracking, and graphics tests pass. Shared writes also pass with watching disabled. Completion storage and ordering pass with core and sync validation on RX 6600 RADV and llvmpipe. Contribution checks pass. No title frame time or Windows result is claimed. AI-assisted.

[popcoder565]

Tested on Windows, since the description says it wasn't: Windows 10 (19045), RTX 3060 Laptop (driver 616.92), MinGW-w64 GCC 15.2.0, Release. I tested main 7a178e2 merged with 349781d.

• agc_shared_write passes, and agc_driver_recorder_tests (not in ctest) passes 3 of 3 runs.
• Full ctest: 415 of 417 pass. The two failures happen on main without this PR too. agc_driver_srgb8_color_target fails 7 of 10 runs there, and guest_raise_exception times out intermittently, which fix(libkernel): stop losing and deadlocking sceKernelRaiseException deliveries on Windows #1421 fixes.

[oneandonlydean]

PhysicalBacking::HostWriteData (DirectMemory.cpp) returns nullptr when the mmap of the write view fails, and StoreOwnBytes then writes with memcpy instead. Failures stay explicit here (#182); please throw when that mmap fails, and keep false only for ranges that aren't direct-memory backed.

[adriadam10]

Rebased onto current main (conflicts in the agcTest list in libSceAgcDriver/CMakeLists.txt and in main of tests/Recorder.cpp) and added one commit:

• PhysicalBacking::HostWriteData (DirectMemory.cpp) now throws std::system_error when the mmap of the write view fails, as requested in the review. It still returns nullptr only for ranges that aren't backed by direct memory.

Rebuilt the AgcDriver, libc and libkernel targets; agc_shared_write passes, and the new test no longer builds without the non-test changes. Windows results from @popcoder565 above still apply (production code for Windows is unchanged).

Note: #1152 also changes PhysicalBacking in DirectMemory.cpp and conflicts with this PR; whichever lands second needs a rebase.

[github-actions]

Other open pull requests touch the same code. Merging one will break the other, or both implement the same thing. If this one builds on them, list them in Depends on.

Pull request	Conflicting files (lines on main)	Same exports	Same new files
#1046	core/libs/prx/libSceAgcDriver/CMakeLists.txt L673
#1152	core/libs/prx/libkernel/DirectMemory/DirectMemory.cpp L226-227, L255, L261-262, L282
#1640	core/libs/prx/libSceAgcDriver/CMakeLists.txt L673
#1959	core/libs/prx/libSceAgcDriver/CMakeLists.txt L673
#2470	core/libs/prx/libSceAgcDriver/CMakeLists.txt L673
#2501	core/libs/prx/libSceAgcDriver/CMakeLists.txt L673