Discard dirty guest memory at exit

[henrybear327]

Currently, the disk IO is very heavy when gvisor / LTP conformance test is run. On local development machine, it's easy to see around 1TB of accumulated write to SSD.

The main issue lies in the flush mechanism of our file-backed memory file - when a process terminates, the file is flushed, though no one would be reading it anymore.

Experiments

For workloads that isn't fork-heavy, we can see improvements.

Workload	main	branch	Wall time
gVisor readv_test	2056.3 MiB	12.2 MiB	1.15 s -> 0.41 s
gVisor close_range_test	832.9 MiB	0.4 MiB	2.01 s -> 0.10 s

---

Summary by cubic

Discards a guest's dirty RAM at process exit instead of writing it back to disk, eliminating a major source of disk I/O (roughly 1 TB of writes on the worst local workloads; e.g. gVisor readv_test dropped from 2056 MiB to 12 MiB written).

guest_destroy now truncates the unlinked slab file before its last close. The truncate is skipped when fclonefileat failed and a fork child received the live slab fd (fallback path, marked via shm_exported), because truncating under that child's private mapping would stall the host.

• Adds test-slab-exit-writes, which reads the process's block-write count from its zombie to confirm dirtying 256 MiB writes nothing at exit, with a 32 MiB fsync control.
• Adds test-slab-exit-fork, which runs both the APFS clone path and the forced live-fd fallback, verifying a fork child still reads its memory after the parent exits and that the truncate is skipped only in the fallback; probe-disk-writes records the counter.
• Documents the truncate and fallback in docs/internals.md, including ELFUSE_DISABLE_FORK_CLONEFILE=1 to force the fallback on APFS.

^{Written for commit 5a3ab6f. Summary will update on new commits.}

[cubic-dev-ai]

1 issue found across 3 files (changes from recent commits).

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. When an issue isn't valid or won't be fixed in this PR, reply in its thread with the reason and then resolve the thread. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="src/runtime/forkipc.c">

<violation number="1" location="src/runtime/forkipc.c:1842">
P3: This new branch routes a deliberately disabled clone through the error path, so on Rosetta the pre-existing handler logs "clone: rosetta CoW snapshot via fclonefileat failed (Operation not supported)" even though fclonefileat was never attempted, and native guests log a fabricated ENOTSUP. Emit the diagnostic before faking the failure, or use a distinct message for the disabled case, so the logs report the real cause.</violation>
</file>

<sub>Tip: Review your code locally with the cubic CLI to iterate faster.

Turn on auto-fix | Re-trigger cubic</sub>

[cubic-dev-ai]

P3: This new branch routes a deliberately disabled clone through the error path, so on Rosetta the pre-existing handler logs "clone: rosetta CoW snapshot via fclonefileat failed (Operation not supported)" even though fclonefileat was never attempted, and native guests log a fabricated ENOTSUP. Emit the diagnostic before faking the failure, or use a distinct message for the disabled case, so the logs report the real cause.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. When an issue isn't valid or won't be fixed in this PR, reply in its thread with the reason and then resolve the thread. At src/runtime/forkipc.c, line 1842:

<comment>This new branch routes a deliberately disabled clone through the error path, so on Rosetta the pre-existing handler logs "clone: rosetta CoW snapshot via fclonefileat failed (Operation not supported)" even though fclonefileat was never attempted, and native guests log a fabricated ENOTSUP. Emit the diagnostic before faking the failure, or use a distinct message for the disabled case, so the logs report the real cause.</comment>

<file context>
@@ -1834,8 +1834,15 @@ int64_t sys_clone(hv_vcpu_t vcpu,
+        const char *disable_clone = getenv("ELFUSE_DISABLE_FORK_CLONEFILE");
+        if (disable_clone && strcmp(disable_clone, "1") == 0) {
+            snapshot_shm_fd = -1;
+            errno = ENOTSUP;
+        } else {
+            snapshot_shm_fd = fork_snapshot_shm_via_clonefile(g->shm_fd);
</file context>