[root-over]

What

guest_mmap_executable fails on main in containers. The NVIDIA and AMD GPU CI runs on #10069 list it under "Failing on main too" (Subprocess aborted).

CheckExecutableSharedMappings maps named shared memory with PROT_EXEC. On Linux, named objects are files under /dev/shm, and Docker mounts /dev/shm noexec by default. Reproduced with docker run --tmpfs /dev/shm:rw,noexec,nosuid,nodev (ubuntu:24.04, host binaries through chroot):

mmap(0x200000000, 32768, PROT_READ|PROT_EXEC, MAP_SHARED|MAP_FIXED, 5, 0) = -1 EPERM (Operation not permitted)
Guest memory check failed at core/libs/tests/GuestMemory.cpp:3576

The same container with /dev/shm mounted exec passes, so the mount option is the only cause.

• The test reads the flags of /dev/shm with statvfs. When it is noexec, the test:
  • still runs every regular-file, anonymous (memfd), JIT and non-executable named mapping;
  • skips only the executable named ones;
  • prints skipped, the host mounts /dev/shm noexec, so named shared memory is not mapped executable and returns 77.
• guest_mmap_executable gets SKIP_RETURN_CODE 77. Windows is unchanged.
• New debt entry: on such a host a PROT_EXEC mapping of a named object fails, where FreeBSD 11.4 maps POSIX shared memory executable. The library change this would need (named objects outside /dev/shm) is left out of this PR.

Tested

Ubuntu, kernel 7.0, GCC 15.2.0, Release.

/dev/shm	main	this PR
host (rw,nosuid,nodev)	passes	passes
Docker, noexec	aborts at GuestMemory.cpp:3576	skipped (77), the other mappings checked

• guest_memory passes.
• tools/check_conventions.py passes.
• tools/pr_overlap.py: feat(aot): consolidate native ARM64 compiler, runtime and consumers #5021, fix(libkernel): return EACCES for unknown ranges on direct-memory release paths #7105, feat(libkernel): direct memory mappings on macOS hosts #7631 and fix(relinker): match AppContent utility import to provider #13486 touch GuestMemory.cpp, the libkernel debt file or CMakeLists.txt, but none touches CheckExecutableSharedMappings or the guest_mmap_executable registration.

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
• Fixes: none
• Unblocks: none
• GPUs tested: n/a
• AI-assisted: yes

[adam25136]

Windows run, for head bbf62b9350 merged into main 9ab6a281ea (git merge --no-ff, clean merge). On Windows the change sets namedExecutable to true, so this checks that the named mapping is still checked as executable there.

Host	Windows 10 Pro 22H2 (19045), i9-12900KF, WinLibs GCC 15.2.0 r7, Release, BUILD_TESTING=ON
GPU	NVIDIA GeForce RTX 5080 (10de:2c02), driver 621.14
Selection	3,940 tests, chosen by tools/select_tests.py select for the 3 changed files
Run	select_tests.py ctest build --parallel 4 --timeout 120 with ANYPS5_REQUIRE_VULKAN=1 ANYPS5_NO_SHADER_CACHE=1; main 9ab6a281ea ran its full suite (4,030 tests) the same way on this machine
Result	3,926 passed, 12 skipped, 2 failed, 0 new

• guest_mmap_executable passes in the selection and 3 of 3 more times with ctest -V, without the new skipped, the host mounts /dev/shm noexec ... line. guest_memory passes 3 of 3 too.
• Both failures fail on main 9ab6a281ea here too. agc_driver_direct_write_completion is AGC: agc_driver_direct_write_completion fails on Windows since #13159 #15564. agc_driver_quad_pixel_mask is AGC: agc_driver_quad_pixel_mask samples the wrong mip level under a partial quad mask on Windows (RTX 5060 Ti and Radeon 860M) #14237: it fails on main with driver 621.14 and passed with 617.42. The 12 skips also skip on main here.
• The slowest selected test, agc_driver_pipeline_journal, took 28.7 s; on main here it takes 29.8 s.

No title was run.

AI-assisted: yes. Claude Code ran the builds and tests and drafted this comment; the numbers are from the runs on this machine.