[adam360x]

Motivation

A process that calls amdsmi_init() and exits without amdsmi_shut_down() leaks every socket and processor object the library created, so under LeakSanitizer its exit fails with code 1 (for example, a tool that only prints --version). Same ownership fix as #12362, one level up.

Technical Details

Library (src/amd_smi/amd_smi_system.cc):

• AMDSmiSystem::getInstance() allocates the system once and never destroys it, so the GPU, CPU and NIC objects it owns stay reachable until the process ends. The function-local static it replaces was destroyed at exit, which freed only its containers and left those objects unreachable
• Never destroying it is deliberate, as in fix(amdsmi): stop leaking the per-GPU process-list cache #12362: freeing the objects at exit would pull them out from under any thread still inside the library while static destructors run
• amdsmi_shut_down() still frees everything; init and shut-down cycles are unchanged
• Trade-off: a host that unloads the library with dlclose() leaves the system object (about 0.5 KB) unreachable, as it already does fix(amdsmi): stop leaking the per-GPU process-list cache #12362's map

Headers (include/amd_smi/impl/amd_smi_system.h, include/amd_smi/impl/amd_smi_drm.h):

• Give the NIC context pointer and the DRM vendor ID default member initializers (nullptr, 0). The static object was zero-initialized before its constructor ran; the heap object is not, so both would otherwise start indeterminate

Tests (tests/amd_smi_test/functional/system/exit_without_shutdown_test.cc):

• Add SystemFunctionalReadOnly.ExitWithoutShutDownIsClean: a threadsafe death test re-runs itself in a fresh process that exits right after amdsmi_init() without amdsmi_shut_down(), and expects exit code 0. In the AMDSMI ASAN GPU job, LeakSanitizer turns a leak into a failure
• Skips when the amdgpu driver is not loaded

No public header, wrapper, CLI or docs change. No CHANGELOG entry: the leak only shows up at exit.

Issue Tracking

JIRA ID: TBD (ticket to be created)

Test Plan

• Debug ASAN build (-DADDRESS_SANITIZER=ON -DBUILD_TESTS=ON) with the CI's ASAN_OPTIONS and LSAN_OPTIONS, A/B against the same build with only the library change reverted:
  • 2 Navi GPUs (gfx1030 + gfx1100), non-root, AMDSMI_NON_PRIVILEGED=1: full amdsmitst
  • 8x MI308X, root: *Unit*:GpuFunctionalReadOnly.*:SystemFunctionalReadOnly.*
  • MI455X host with an AMD CPU (HSMP) and an AI NIC, root: the same filter, plus a small program that calls amdsmi_init() with the GPU, CPU, NIC and combined flags and exits without amdsmi_shut_down()
• gdb on the Navi box (ASAN build): the constructed NIC context pointer and vendor ID, and a retry of amdsmi_init() with GPU and NIC flags after forcing the first call to fail once rocm_smi is initialized
• Non-ASAN RelWithDebInfo build: SystemFunctionalReadOnly.*, 3 runs
• pre-commit with projects/amdsmi/.pre-commit-config.yaml

Test Result

• Without the fix, the new test fails with the LeakSanitizer report (exit status 1): 3,712 bytes in 24 allocations on 8x MI308X, 928 bytes on the 2 Navi GPUs, 464 bytes on MI455X. With the fix it passes on all three
• Init-and-exit program on MI455X, before → after: GPU 464 B → 0, CPU 7,408 B in 132 allocations → 0, AI NIC 649,320 B → 0, all three 657,192 B → 0
• Suites: the new test is the only difference; with the fix, 0 sanitizer reports (Navi 243 passed / 26 skipped, MI308X 244 / 3, MI455X 244 / 3)
• Constructed values with the initializers: null context and vendor ID 0, as on develop (without them, ASAN's fill pattern 0xbebe...). The forced retry crashed in smi_nic_destroy_context() without the initializers and succeeds with them
• Non-ASAN: 6/6 system tests pass in all 3 runs

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/rocm-systems/blob/develop/CONTRIBUTING.md.

[therock-pr-bot]

✅ All Policy Checks Passed

Check	Status	Details
📝 PR Description	✅ Pass	—
⛔ Forbidden Files	✅ Pass	—
🧪 Unit Test	✅ Pass	—
🚫 Draft PR	🔜 To Be Enabled	—
🚩 Feature Flag	🔜 To Be Enabled	—
📊 Code Coverage	🔜 To Be Enabled	—

🎉 All policy checks passed!

📖 Need help? See the Policy FAQ for details on every check and how to fix failures.

🙋 Wish to Override Policy?

[Copilot]

Copilot review overview

🟡 Changes recommended

Heap allocation removes implicit zero-initialization, leaving members indeterminate and allowing combined GPU/NIC initialization to crash.

Review effort: Balanced
Findings: 1

Open (1)

• Initialize AMD SMI context and DRM vendor ID fields · New

What changed in this PR

Extends #12362’s lifetime approach to AMD-SMI’s system singleton, preventing exit-time leak reports when shutdown is omitted.

Changes:

• Keeps the singleton alive until process termination while preserving explicit shutdown.
• Adds a subprocess regression test for initialization followed by exit without shutdown.

File	Description
projects/​amdsmi/​tests/​amd_smi_test/​functional/​system/​exit_without_shutdown_test.cc	Tests exit without shutdown under LeakSanitizer.
projects/​amdsmi/​src/​amd_smi/​amd_smi_system.cc	Replaces the destructed static singleton with permanent heap storage.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[systems-assistant]

TheRock Submodule Bump Activity

Newest first

• 2026-10-09 — Included in TheRock via ROCm/TheRock#8885.