[koushikbillakanti-amd]

Motivation

Cover existing FCLK error reporting on mixed GPUs without hiding malformed driver data. The empty Raphael FCLK table still needs a driver fix.

Technical Details

• Use shared sysfs/device fixtures to test valid and newline-only FCLK tables through the real public APIs.
• Route clock-range reads through the existing private sysfs-root seam; the production path is unchanged.
• Drop the parser return-value changes and unrelated reset/sweep changes. Newline-only data remains IO; zero-byte and missing-node behavior are unchanged.
• Keep memory-partition coverage, --gpu all, and saved clock-limit restoration. No user-facing change remains, so the earlier changelog entries are removed.

Issue Tracking

JIRA ID: ROCM-33079

Separate driver ticket: pending required Targeted ASIC selection; the November 2 logs are attached to the original ticket.

Test Plan

BUILD and ASAN_BUILD below point to configured Debug builds (BUILD_TESTS=ON; the latter also has ADDRESS_SANITIZER=ON).

• cmake --build "$BUILD" --target amdsmitst amd_smi amdsmi_cli amdsmi_cli_subcommands -j4 — passed.
• "$BUILD/tests/amd_smi_test/amdsmitst" --gtest_filter='*Unit*' — passed; also passed with sudo -n.
• ASAN_OPTIONS=detect_leaks=1:strict_string_checks=1:detect_stack_use_after_return=1:check_initialization_order=1 LSAN_OPTIONS="suppressions=$PWD/tests/amd_smi_test/lsan.supp:print_suppressions=1" "$ASAN_BUILD/tests/amd_smi_test/amdsmitst" --gtest_filter='*Unit*' — passed; privileged unit cases also passed.
• sudo -n env PYTHONDONTWRITEBYTECODE=1 LD_LIBRARY_PATH="$BUILD/src" AMDSMI_PATH="$BUILD/py-interface/python_package" python3 tests/python/unit_tests.py -v -k cli — passed.
• Pinned project pre-commit hooks on changed files and git diff --check — passed.

Test Result

C++: 313 passed / 25 skipped without root; 328 passed / 10 skipped with root, including under ASAN with leak checks. Python: 360 passed. Clock/RAS fixtures also passed 10 shuffled repetitions.

The hardware set sweep and driver reload were not run; pending memory-partition cancellation is not claimed. Previous-head OS CI stopped at DRIVER_NOT_LOADED; GPU ASAN produced no test results. New-head CI must still complete. The ASAN build retains one existing E-SMI discarded-const warning.

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?

[gabrpham]

Please see comments below.

[charis-poag-amd]

ROCM-33079 - states driver has empty pp_dpm_fclk file. Which is a driver issue.

Lets close this PR and re-assign to them (after providing evidence). Otherwise, if there are other issues we are fixing - please outline in the Jira ticket.

[gabrpham]

ROCM-33079 - states driver has empty pp_dpm_fclk file. Which is a driver issue.

Lets close this PR and re-assign to them (after providing evidence). Otherwise, if there are other issues we are fixing - please outline in the Jira ticket.

The PR still addresses other issues in ROCM-33079 which were the main issues the ticket wanted addressed in the first place. I would think we want this PR to go through but a new ticket raised for the pp_dpm_fclk empty file issue.

[charis-poag-amd]

Thanks for digging into this, Koushik. I left inline comments; the summary is below.

What ROCM-33079 shows

• The rc=12 IO on --clk-limit FCLK MAX comes from the Raphael iGPU's pp_dpm_fclk containing only a newline. The parser already returns AMDSMI_STATUS_IO for that on develop, so nothing in AMD SMI needs to change for it. It is a driver bug, as Maisam noted on the ticket, and the fix belongs there.
• MIXED_DEVICE_ERRORS (204) on the other FCLK commands is the CLI combining the MI300X's and the iGPU's different error codes.

Requests

1. Please remove the changes that aren't needed for this ticket and aren't related to pp_dpm:
   • the reset --clocks FCLK cap restore, with its docs and tests
   • the readDevInfoMultiLineStr read-failure check
   • the assert(false) removal in rocm_smi.cc
   • the sweep rework
   • the --memory-partition exclusion
2. Please don't make AMD SMI succeed on bad driver data. Blank lines in pp_dpm_* shouldn't be skipped, so a newline-only table stays IO.
3. Please keep the sweep coverage: --memory-partition, --gpu all, and the clk-limit restore. The ticket asks to confirm the memory-partition restore works, not to remove it.
4. If you keep any library change, say exactly which inputs change behaviour (for example, a zero-byte file going from SUCCESS to IO). Update the CHANGELOG to match.
5. For tests, use the shared fixtures in test_fixture_utils.h (FakeSysfsTree, ScopedRocmSmiDevice) instead of FakeDpmDevice.
6. Please file and link a driver ticket for the empty pp_dpm_fclk. Until it's fixed, mixed hosts will keep failing the FCLK commands, and the sweep shouldn't be changed to hide that.

[gabrpham]

Comments have been addressed. Approved.