Support madvise, msync, and mremap on high-VA mmap regions

[Max042004]

Problem

Three memory syscalls -- sys_madvise, sys_msync, and sys_mremap -- plus
the file-backed MAP_SHARED path in sys_mmap_high_va were all
primary-window-only. They computed off = addr - ipa_base and reached into
host_base + off, which only resolves for identity regions (gpa_base == start). High-VA mmap regions -- Rosetta's own slab/JIT and guest mmap(NULL)
placements -- back their pages at gpa_base (a named mapping or overflow
segment), so these paths rejected the range with -ENOMEM/-EFAULT or wrote to
the wrong host address.

Two concrete failures motivate this:

• x86_64 Node.js / V8. V8's page allocator decommits guard/code pages with
  mprotect(PROT_NONE) + madvise(MADV_DONTNEED) and CHECK_EQ(0, ret)s the
  madvise return, so the spurious -ENOMEM aborted Node the moment its JIT
  initialized:

  # Fatal error in , line 0
  # Check failed: 0 == ret.
  

  --verbose pins it to madvise(advice=4 MADV_DONTNEED) -> -12 (ENOMEM) on a
  page in the high-VA window (mmap(NULL) under Rosetta lands at e.g.
  0x7fffff7fd000).

• apt (issue msync() returns ENOMEM (errno 12) for high-VA guest mappings created by Rosetta #108). apt memory-maps its package-list cache with MAP_SHARED
  and msyncs it; under an x86_64 guest that mapping is high-VA, so sys_msync
  returned -ENOMEM and apt failed with
  E: Unable to synchronize mmap - msync (12: Cannot allocate memory). The
  mapping could not even be created, because sys_mmap_high_va refused
  file-backed MAP_SHARED with -ENODEV.

Fix

Route every admission and data-movement path through the region tracker and
host_ptr_for_gpa(gpa_base + ...) so primary and high-VA regions act on their
real backing. Identity regions have gpa_base == start, so this collapses to
host_base + off and primary behaviour is byte-for-byte unchanged.

• madvise MADV_DONTNEED accepts high-VA ranges the tracker records as
  mapped and zeroes / restores them on their gpa_base backing (drops the
  previous zero-fill-only scope-out for high-VA file-backed pages).
• msync admission drops the guest_size bound (the existing coverage loop
  validates both windows), and sync_shared_aliases_range /
  refresh_shared_region_range resolve the guest bytes through gpa_base.
• sys_mmap_high_va accepts file-backed MAP_SHARED as a snapshot-style
  shared region (pread on map, msync writes dirty bytes back), so the high-VA
  shared mappings msync operates on can be created.
• mremap admits high-VA sources and resolves the source read/zero through
  gpa_base in the shrink, MREMAP_FIXED, and MREMAP_MAYMOVE paths. The
  destination stays in the primary window (find_free_gap /
  mremap_extend_range), so an mremap(MAYMOVE) of a high-VA region relocates
  it there with its contents intact -- no high-VA destination backing or new
  Stage-2 machinery is needed.

Tests

Vendored x86_64 static ELFs run through elfuse + Rosetta (rebuild recipe in
tests/fixtures/rosetta/README.md):

• x86_64-rosetta-madvise -- MADV_DONTNEED on writable / PROT_NONE guard /
  multi-page high-VA ranges (the V8 decommit pattern).
• x86_64-rosetta-msync -- high-VA MAP_SHARED write-back
  (single/multi-page/MS_ASYNC), verified against the backing file.
• x86_64-rosetta-mremap -- MREMAP_MAYMOVE grow (contents preserved) and
  in-place shrink.

Wired as make test-rosetta-{madvise,msync,mremap} (and test-rosetta-all)
plus the matching suites in tests/test-matrix.sh.

The aarch64 unit tests test-madvise / test-msync / test-mremap /
test-mremap-infra cover the unchanged primary-window path. An end-to-end
MAP_SHARED store scenario (write + msync + cross-mapping refresh + mremap
grow + madvise) passes with this change and fails (-ENODEV/-ENOMEM) on the
pre-fix baseline.

Notes

• macOS CI runners cannot run guest tests (no HVF entitlement), so the Rosetta
  suites were verified locally on Apple silicon; they self-skip (exit 77) where
  the translator or timeout(1) is unavailable.
• The vendored x86_64 fixtures are not built in-tree (the Makefile targets
  aarch64 hosts); rebuild them out of tree per the README when the sources
  change.
• For high-VA file-backed MAP_SHARED served through the FUSE sysroot layer,
  the backing is materialized read-only, so msync write-back to the original
  sysroot file is best-effort; real host-fd backings (e.g. /dev/shm) write
  back exactly. This matches how the primary window already treats such
  mappings.

Fixes #108.

[cubic-dev-ai]

4 issues found across 7 files

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="tests/test-rosetta-madvise.sh">

<violation number="1" location="tests/test-rosetta-madvise.sh:66">
P2: Script exits 0 even when tests fail — missing `if [ "$fail" -gt 0 ]; then exit 1; fi` after `report_summary`. This masks test failures in standalone runs and weakens the matrix runner's belt-and-suspenders `|| rc=1` check.</violation>
</file>

<sub>Reply with feedback, questions, or to request a fix.

Re-trigger cubic</sub>

[cubic-dev-ai]

P2: Script exits 0 even when tests fail — missing if [ "$fail" -gt 0 ]; then exit 1; fi after report_summary. This masks test failures in standalone runs and weakens the matrix runner's belt-and-suspenders || rc=1 check.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At tests/test-rosetta-madvise.sh, line 66:

<comment>Script exits 0 even when tests fail — missing `if [ "$fail" -gt 0 ]; then exit 1; fi` after `report_summary`. This masks test failures in standalone runs and weakens the matrix runner's belt-and-suspenders `|| rc=1` check.</comment>

<file context>
@@ -0,0 +1,66 @@
+    printf '%s\n' "$madv_out" >&2
+fi
+
+report_summary "$total"
</file context>

[jserv]

Rebase latest main branch and resolve conflicts and refine per review.

[jserv]

Enforce rules described in https://cbea.ms/git-commit/ carefully.

[jserv]

Thank @Max042004 for contributing!