[marifamd]

Motivation

PR #10179 (feat(amdsmi): expand amdsmi_gpu_block_t with new IP blocks) added 20
RAS IP blocks at bits 19-38 and moved AMDSMI_GPU_BLOCK_LAST from
AMDSMI_GPU_BLOCK_MPIO (1ULL << 18) to AMDSMI_GPU_BLOCK_UCIE_PCS
(1ULL << 38). It updated every AMD SMI consumer of that enum, but missed
projects/rdc, which had hard-coded the old tail. develop sat at a ~76%
TheRock CI failure rate until #10947 reverted it.

This PR fixes RDC first, relands #10179 unchanged on top, then repairs an
out-of-bounds read in an AMD SMI example that the wider enum makes worse.

Technical Details

Three commits, in dependency order. Each builds on its own.

1. fix(rdc): remove hard-coded AMD SMI GPU block enum bounds

Build break (tests/rdc_tests/test_common.cc, test_common.h) — the CI failure:

• kBlockNameMap was a copy of AMD SMI's block name table guarded by
  static_assert(AMDSMI_GPU_BLOCK_LAST == AMDSMI_GPU_BLOCK_MPIO). Its only
  consumer, GetBlockNameStr(), has zero callers anywhere in projects/rdc.
  The map and its guard did nothing except fail the build when AMD SMI added a
  block. Removed the map, the assert, the function, and its declaration.
  GetErrStateNameStr() / kErrStateNameMap are left untouched.

Latent hang (rdc_libs/rdc/src/RdcMetricFetcherImpl.cc) — not a test, production library code:

• get_ecc_total() and get_ecc_deferred_total() walked the enum with
  for (uint32_t b = AMDSMI_GPU_BLOCK_FIRST; b <= AMDSMI_GPU_BLOCK_LAST; b = b * 2).
  amdsmi_gpu_block_t is 64 bits wide (AMDSMI_GPU_BLOCK_RESERVED = 1ULL << 63).
  Once any block sits above bit 31, b * 2 wraps to 0, 0 <= LAST stays true,
  and the loop never terminates. Both counters widened to uint64_t.

PR #10179 had already applied this same widening inside AMD SMI, in
tests/amd_smi_test/functional/gpu/ras/err_cnt_read_test.cc. RDC's two copies
were the ones missed.

2. feat(amdsmi): expand amdsmi_gpu_block_t with new IP blocks

Relands be17474f7 verbatim (reverted in 2d4df4291), credited to the original
author. Header, generated Python wrapper, Python interface, Rust bindings, and
the AMD SMI test suite.

3. fix(amdsmi): bound the DRM example's RAS block name lookup

example/amd_smi_drm_example.cc walked every amdsmi_gpu_block_t value but
indexed char block_names[14][10], reading past the end of the array and
printing whatever followed on the stack:

Enum state	Loop iterations	Array entries	Overread
Before #10179 (1ULL << 18)	19	14	5 entries (already broken)
After commit 2 (1ULL << 38)	39	14	25 entries

The bug predates #10179 — the array was already missing MCA, VCN, JPEG,
IH, and MPIO — but commit 2 widens it fivefold, so it is fixed here rather
than left in tree.

• Filled in all 39 names, verified position-by-position against the enum.
• Switched to const char*; the old 10-byte cell would have truncated
  KPX_SERDES (10 chars + NUL).
• Bounded the lookup so a future block addition prints UNKNOWN instead of
  reading out of bounds.

Full-monorepo sweep

Checked every consumer of AMDSMI_GPU_BLOCK_LAST outside the header to confirm
nothing else breaks the way RDC did:

Consumer	Status
rdc/.../RdcMetricFetcherImpl.cc (x2)	Fixed in commit 1 (uint64_t)
rdc/tests/.../test_common.cc	Coupling removed in commit 1
amdsmi/src/amd_smi/amd_smi.cc (x2)	Safe — auto deduces the 64-bit enum type
amdsmi/tests/.../err_cnt_read_test.cc	Already uint64_t via #10179
amdsmi/tests/.../test_common.cc	Map + assert updated by #10179
amdsmi_cli/subcommands/metric.py	Safe — name lookup, not enumeration
amdsmi/example/amd_smi_drm_example.cc	Pre-existing out-of-bounds read, fixed in commit 3
rocm-smi-lib/.../err_cnt_read.cc	Unaffected — different enum (rsmi_gpu_block_t)

Issue Tracking

None.

Test Plan

• Reproduced the original CI failure locally by restoring the expanded enum and
  compiling the affected translation units.
• Compiled the RDC sources against both enum widths (1ULL << 18 and
  1ULL << 38) to confirm commit 1 stands alone and is not tied to commit 2.
• Compiled rdc/tests/rdc_tests/test_common.cc, rdc_libs/rdc/src/RdcMetricFetcherImpl.cc,
  amdsmi/tests/amd_smi_test/test_common.cc, and amdsmi/example/amd_smi_drm_example.cc
  against the final combined tree (-Wall -Wextra).
• Cross-checked the example's block name list against the enum programmatically,
  position by position, to confirm all 39 entries line up with their bit positions.
• Standalone reproduction of the iteration bug comparing uint32_t vs uint64_t
  counters at both enum widths, with an iteration cap to detect non-termination.
• Verified the enum expansion is consistent across all binding layers (header,
  ctypes wrapper, Python interface, Rust) and that generated values match the
  intended bit positions.
• clang-format plus the full projects/rdc and projects/amdsmi pre-commit
  hook sets on every changed file.

Test Result

• Before: test_common.cc:62:37: error: static assertion failed: kBlockNameMap needs to be updated
  (expression evaluates to '274877906944 == 262144') — identical to the CI log.
• After: all three translation units compile clean, no warnings, on the combined tree.
• Commit 1 alone compiles clean against the pre-expansion enum, so the two commits
  are independently valid.
• Iteration reproduction: LAST = 1<<18 with uint32_t terminates in 19 steps;
  LAST = 1<<38 with uint32_t does not terminate; LAST = 1<<38 with
  uint64_t terminates in 39 steps.
• Binding values spot-checked: MMSCH = 524288 (1<<19), OXRP = 2147483648
  (1<<31), PCIE_PL = 4294967296 (1<<32), UCIE_PCS = LAST = 274877906944
  (1<<38); 20 new blocks present in the header, ctypes wrapper, Python
  interface, and Rust bindings.
• Example block names: 39/39 match the enum in order (bit 0 resolves to UMC, which
  aliases AMDSMI_GPU_BLOCK_FIRST); longest name KPX_SERDES at 10 chars confirms the
  old [10] cell would have truncated it.
• pre-commit: codespell, trailing whitespace, end-of-file, large files,
  clang-format, ruff, mixed line ending, C++ test conventions, SPDX headers all pass.

The only removed symbol is an unused test-only helper in RDC. The enum expansion
is additive, and the example fix is confined to a sample program.

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?

[systems-assistant]

TheRock Submodule Bump Activity

Newest first

• 2026-09-02 — Included in TheRock via ROCm/TheRock#7825.