fix: prevent unsigned integer underflow in NVML memory info hook

[miaobyte]

Summary

Fix unsigned integer underflow in _nvmlDeviceGetMemoryInfo: when GPU memory usage exceeds the configured limit, (limit - usage) with unsigned size_t wraps to ~18 EB, breaking monitoring tools.

Changes

src/nvml/hook.c: compute clamped = (usage > limit) ? limit : usage, then derive both used = clamped and free = limit - clamped. This guarantees free + used = total always holds and free never underflows.

Fixes #200

🤖 Generated with Claude Code

Summary by CodeRabbit

• Bug Fixes
  • Corrected reported GPU and CUDA memory values when actual usage exceeds the configured memory limit.
  • Prevented invalid or underflowed free-memory results.
  • Added warnings when memory usage exceeds the configured limit.

[hami-robot]

[APPROVALNOTIFIER] This PR is NOT APPROVED

This pull-request has been approved by: miaobyte
Once this PR has been reviewed and has the lgtm label, please assign archlitchi for approval. For more information see the Kubernetes Code Review Process.

The full list of commands accepted by this bot can be found here.

Details

Needs approval from an approver in each of these files:

• OWNERS

Approvers can indicate their approval by writing /approve in a comment
Approvers can cancel approval by writing /approve cancel in a comment

[mesutoezdil]

This silently clamps the value without logging, so operators can't tell usage actually exceeded the limit.

[maverick123123]

Agreed, a LOG_WARN here would help operators detect when usage exceeded the limit. Something like:

size_t clamped = (usage > limit) ? limit : usage;
if (usage > limit) {
    LOG_WARN("NVML meminfo: usage %lu exceeds limit %lu, clamping", usage, limit);
}

[mesutoezdil]

This only fixes the NVML hook, not the CUDA hook underflow from the same issue, so closing #200 here would be premature.

[maverick123123]

已收到，正在 review 你的 PR。

[maverick123123]

Thanks for the review @mesutoezdil. Agreed on both points:

1. ✅ Need LOG_WARN — silently clamping hides the fact that usage drifted above limit, which operators need visibility into
2. ✅ CUDA hook also needs fixing — the return CUDA_ERROR_INVALID_VALUE path in src/cuda/memory.c should be similarly clamped rather than returning an error, otherwise applications calling cuMemGetInfo may crash or behave unpredictably

@miaobyte please update when you have a chance.

[coderabbitai]

Warning

Review limit reached

@miaobyte, you've reached your PR review limit, so we couldn't start this review.

Next review available in: 53 minutes

Enable usage-based reviews in Billing to review now. Otherwise, wait until the next included review is available.
You're only billed for reviews past your plan's rate limits ($0.25/file).

How can I continue?

After more reviews become available, a review can be triggered using the @coderabbitai review command as a PR comment. Alternatively, push new commits to this PR.

To avoid repeated limits, reduce automatic review volume by pausing incremental auto-reviews earlier, using label-based review opt-in, excluding WIP or generated PR titles, or requesting reviews manually when the PR is ready. If your team needs uninterrupted high-volume reviews, an organization admin can enable usage-based reviews.

How do review limits work?

CodeRabbit enforces per-developer PR review limits for each organization. Most developers receive the normal plan review availability.

For paid Pro and Pro+ PR reviews, CodeRabbit uses adaptive limits for sustained high-volume activity. When a developer's recent PR review activity reaches the 95th percentile or higher among CodeRabbit users, additional reviews become available more gradually as earlier reviews age out of the rolling window.

Please refer docs for additional details.

Review details

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 63c3def8-9c1a-44bd-a94f-5bfd11ab34e2

📥 Commits

Reviewing files that changed from the base of the PR and between b72a46f and c56e601.

📒 Files selected for processing (2)

• src/cuda/memory.c
• src/nvml/hook.c

📝 Walkthrough

Walkthrough

CUDA and NVML memory hooks now handle usage exceeding configured limits by logging warnings, clamping usage, and reporting bounded memory values instead of errors or underflowed free memory.

Changes

Memory Usage Clamping

Layer / File(s)	Summary
Clamp CUDA and NVML memory reports src/cuda/memory.c, src/nvml/hook.c	CUDA clamps over-limit usage while preserving the underlying query and limit calculations. NVML clamps usage and recalculates free, total, and used for both memory API versions.

Estimated code review effort: 2 (Simple) | ~10 minutes

Suggested reviewers: chaunceyjiang, archlitchi, mesutoezdil

Poem

I’m a rabbit with numbers neat,
No giant free-memory leaps.
CUDA clamps with warning bright,
NVML keeps its values right.
Bounded bytes now softly sleep.

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title is concise and matches the main bug fix, though it mentions only the NVML side of the change.
Linked Issues check	✅ Passed	The PR implements the requested NVML clamping and updates the CUDA hook to clamp instead of erroring, matching issue #200.
Out of Scope Changes check	✅ Passed	The changes stay within the memory-hook underflow fix in the two referenced source files and add no unrelated scope.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[coderabbitai]

Actionable comments posted: 1

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@src/nvml/hook.c`:
- Around line 356-369: Update the NVML memory handling around the version switch
to derive actual_limit as the minimum of the configured limit and the original
NVML-reported total. Use actual_limit for usage clamping and for the free and
total fields in both nvmlMemory_t and nvmlMemory_v2_t, and warn when usage
exceeds actual_limit.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 44929024-f7e6-470f-8905-80b61a875ce9

📥 Commits

Reviewing files that changed from the base of the PR and between 52f33fc and b72a46f.

📒 Files selected for processing (2)

• src/cuda/memory.c
• src/nvml/hook.c

[coderabbitai]

🗄️ Data Integrity & Integration | 🟠 Major | ⚡ Quick win

Cap NVML’s effective limit by the physical total.

Unlike src/cuda/memory.c Lines 522-526, this path sets total = limit unconditionally. If the configured limit exceeds the underlying NVML-reported total, NVML can return physically impossible total and free values. Compute actual_limit = min(limit, original_total), clamp usage to that value, and warn when usage > actual_limit so both hooks report the same bounded memory contract.

Proposed fix

+    size_t original_total = version == 1
+        ? ((nvmlMemory_t*)memory)->total
+        : ((nvmlMemory_v2_t*)memory)->total;
+    size_t actual_limit = (limit < original_total) ? limit : original_total;
-    if (usage > limit) {
-        LOG_WARN("usage(%ld) exceeds limit(%ld), clamping", usage, limit);
+    if (usage > actual_limit) {
+        LOG_WARN("usage(%zu) exceeds effective limit(%zu), clamping",
+            usage, actual_limit);
     }
-    size_t clamped = (usage > limit) ? limit : usage;
+    size_t clamped = (usage > actual_limit) ? actual_limit : usage;
...
-        ((nvmlMemory_t*)memory)->free = limit - clamped;
-        ((nvmlMemory_t*)memory)->total = limit;
+        ((nvmlMemory_t*)memory)->free = actual_limit - clamped;
+        ((nvmlMemory_t*)memory)->total = actual_limit;
...
-        ((nvmlMemory_v2_t *)memory)->free = limit - clamped;
-        ((nvmlMemory_v2_t *)memory)->total = limit;
+        ((nvmlMemory_v2_t *)memory)->free = actual_limit - clamped;
+        ((nvmlMemory_v2_t *)memory)->total = actual_limit;

📝 Committable suggestion

‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.

Suggested change

	if (usage > limit) {
	LOG_WARN("usage(%ld) exceeds limit(%ld), clamping", usage, limit);
	}
	size_t clamped = (usage > limit) ? limit : usage;
	switch (version) {
	case 1:
	((nvmlMemory_t*)memory)->free = (limit-usage);
	((nvmlMemory_t*)memory)->free = limit - clamped;
	((nvmlMemory_t*)memory)->total = limit;
	((nvmlMemory_t*)memory)->used = usage;
	((nvmlMemory_t*)memory)->used = clamped;
	return NVML_SUCCESS;
	case 2:
	((nvmlMemory_v2_t *)memory)->free = (limit-usage);
	((nvmlMemory_v2_t *)memory)->free = limit - clamped;
	((nvmlMemory_v2_t *)memory)->total = limit;
	((nvmlMemory_v2_t *)memory)->used = usage;
	((nvmlMemory_v2_t *)memory)->used = clamped;
	size_t original_total = version == 1
	? ((nvmlMemory_t*)memory)->total
	: ((nvmlMemory_v2_t*)memory)->total;
	size_t actual_limit = (limit < original_total) ? limit : original_total;
	if (usage > actual_limit) {
	LOG_WARN("usage(%zu) exceeds effective limit(%zu), clamping",
	usage, actual_limit);
	}
	size_t clamped = (usage > actual_limit) ? actual_limit : usage;
	switch (version) {
	case 1:
	((nvmlMemory_t*)memory)->free = actual_limit - clamped;
	((nvmlMemory_t*)memory)->total = actual_limit;
	((nvmlMemory_t*)memory)->used = clamped;
	return NVML_SUCCESS;
	case 2:
	((nvmlMemory_v2_t *)memory)->free = actual_limit - clamped;
	((nvmlMemory_v2_t *)memory)->total = actual_limit;
	((nvmlMemory_v2_t *)memory)->used = clamped;

🧰 Tools

🪛 Cppcheck (2.21.0)

[warning] 357-357: If resource allocation fails, then there is a possible null pointer dereference

(nullPointerOutOfResources)

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@src/nvml/hook.c` around lines 356 - 369, Update the NVML memory handling
around the version switch to derive actual_limit as the minimum of the
configured limit and the original NVML-reported total. Use actual_limit for
usage clamping and for the free and total fields in both nvmlMemory_t and
nvmlMemory_v2_t, and warn when usage exceeds actual_limit.

[miaobyte]

@maverick123123 @mesutoezdil

done,now updated:

NVML hook (src/nvml/hook.c)

• Added LOG_WARN("NVML meminfo: usage %lu exceeds limit %lu, clamping", ...) when usage > limit, so operators have visibility into the over-limit condition.

CUDA hook (src/cuda/memory.c)

• Removed the return CUDA_ERROR_INVALID_VALUE path when usage > limit — it now clamps usage to limit and reports free=0, total=actual_limit, matching the NVML hook behavior. This prevents cuMemGetInfo callers from crashing or seeing unpredictable values.
• Added LOG_WARN("CUDA meminfo: usage %lu exceeds limit %lu, clamping", ...) with the same pattern for consistency.

Both hooks now handle usage > limit identically: clamp, warn, and report free=0 / total=limit.

[maverick123123]

收到，我们先测试一下。

[maverick123123]

测试通过，8x RTX 3090 上编译和运行均正常。@archlitchi 可以 merge 了。

[mesutoezdil]

@miaobyte Thanks for the updates, LOG_WARN and the CUDA fix look good.

One item from CodeRabbit's review is still open though: on the NVML side, total is set to limit unconditionally, without capping it against the GPU's actual physical total like the CUDA path does (actual_limit = min(limit, *total)). If the configured limit is set higher than the real GPU memory, NVML would report an impossible total.

Can you apply the same min() clamp on the NVML side before we merge?