[Raghav314]

What

#6160 maps 2 MiB of AGC driver memory at 0xfe0000000 (the FS table and its routine) and registers it as a guest allocation. On Windows it is committed inside the guest arena, which is write-watched. On Linux the address is in the arena's system-reserved hole, so it is a plain mmap outside the write watch.

Every address-based build maps the registered ranges, so it includes this range. When the host import budget refuses its import, a range normally gets a heap mirror refreshed by the blocks the CPU writes. This range is not watched, so it gets no mirror (unwatched heaps on the [buffers] line) and is copied on the CPU, all 2 MiB, on every build.

• On Linux the driver memory is now registered with the write watch after its table is written, so a refused import is served from a mirror like any other watched range.
• If the write watch is not available, nothing changes.

PPSA36082

Whether the import is refused depends on whether the first build that maps the range comes before the import budget fills, so it varies by run. With APS5_HOST_IMPORT_MIB=6784, APS5_PROFILE_DRAW=1, level UI standing still, about 84 s from launch, on main 1e774a3 with #5656 and #6946:

• before: 19964 imports (39908 MiB) refused so far, last 0xfe0000000+0x200000, and 96000 completed CPU-copy operations ... 96000 outside imports (192000 MiB): 0xfe0000000:96000/192000MiB
• after, in a run where the import was refused as well (1 imports (2 MiB) refused so far, ... last 0xfe0000000+0x200000 by build): no completed CPU-copy report in 60 s from launch, in the menu, on main a700809 with perf(agc): data-only draw hits compare only the user words a stage reads #5656 and perf(agc): draw finisher pipeline (APS5_DRAW_PIPELINE) #6946. Without the change, runs that refuse this import report CPU copies of it within 45 s, in the menu.

The frame rate did not change in that scene (72.2 against 71.6 fps), since the copies are not what limits it there. With the default import budget the range was imported in the runs I made, so nothing was copied before or after.

Tested

Linux (CachyOS), x86-64, GCC, 7800X3D / RTX 4090, NVIDIA 615.78.08, on main 0929be5, rebased without conflicts onto 0d8c51a, whose new commits do not touch these files.

• agc_driver_compute_fs_table_calls also checks that the driver memory is write-watched when the write watch is available. It fails without the change ("the AGC driver memory at 0xfe0000000 is not write-watched") and passes with it.
• The linux job of current build.yml run locally: configure with SPIRV tools on, test registration check, build, libs, the whole ctest and conventions, on main 0929be5: 1662/1662 pass, 16 expected skips, conventions OK.

Checklist

• Based on current main; no other open PR implements the same functions
• One topic per PR; follow-ups go in a new PR
• No comments in code except technical debt
• Unimplemented paths throw (NotImplemented_nid_no_patch); silent stubs are listed under Silent stubs in their library's technical debt file
• No notes, investigation .md files or images added to the repository (attach them to this PR)
• New third-party code is a submodule built from source
• Generic behaviour, not specific to one title
• Depends on: none
• Unblocks: none
• GPUs tested: NVIDIA GeForce RTX 4090, 615.78.08, Linux 7.2.10 (CachyOS)
• AI-assisted: yes

[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
PR 8658	core/libs/prx/libSceAgcDriver/tests/execution/ComputeFsTableCalls.cpp L22-23

Some comparisons could not be completed because a pull request conflicts with the base branch. Review these shared files manually before concluding that there is no overlap.

Pull request	Shared files not compared
PR 5021	core/libs/prx/libSceAgcDriver/Execution/src/Driver/Lifecycle.cpp
PR 5925	core/libs/prx/libSceAgcDriver/Execution/src/Driver/Lifecycle.cpp
PR 6188	core/libs/prx/libSceAgcDriver/Execution/src/Driver/Lifecycle.cpp
PR 7201	core/libs/prx/libSceAgcDriver/Execution/src/Driver/Lifecycle.cpp
PR 7642	core/libs/prx/libSceAgcDriver/Execution/src/Driver/Lifecycle.cpp

[DavidAmsellem]

First AMD-on-Linux (RADV) run of this PR. Head 7372189a7 merged onto main 2fee39f1b (clean merge), built in full and compared with the same main, on the same machine.

AMD Radeon RX 7900 XT (RADV NAVI31, gfx1100), Mesa 25.2.8, Linux Mint 22.3 (Ubuntu 24.04 base), kernel 7.0.0-38, GCC 13.3, Release.

Full suite (ANYPS5_REQUIRE_VULKAN=1 ctest -j8) at 8ec57596f: 1715/1716 with this PR, against 1715/1716 on main. The one failure is the same on both, so this PR adds none: agc_driver_guest_memory_lifetime (#7784, which #7849 addresses).

Comparing the JUnit output with main: this PR adds no ctest entry, and none is skipped or removed. The test it extends, agc_driver_compute_fs_table_calls, runs and passes.

GPUs tested: AMD Radeon RX 7900 XT (RADV NAVI31, gfx1100), Mesa 25.2.8, Linux Mint 22.3 / kernel 7.0.0-38

AI-assisted: yes