Fix #951: Fallback gracefully if a library can't be loaded by mdboom · Pull Request #1052 · NVIDIA/cuda-python

mdboom · 2025-09-30T17:10:48Z

Description

Allows falling back to another library more lazily.

closes #951

Checklist

New or existing tests cover these changes.
The documentation is up to date with these changes.

copy-pr-bot · 2025-09-30T17:10:52Z

Auto-sync is disabled for ready for review pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

mdboom · 2025-09-30T17:12:20Z

/ok to test

cpcloud · 2025-09-30T17:18:04Z

cuda_bindings/cuda/bindings/_internal/nvjitlink_linux.pyx

+    try:
+        handle = load_nvidia_dynamic_lib("nvJitLink")._handle_uint
+    except DynamicLibNotFoundError:
+        handle = 0


Is calling dlsym on a NULL pointer handle undefined behavior? It seems with this change that code path is now easily possible.

According to https://pubs.opengroup.org/onlinepubs/009604299/functions/dlsym.html, it says

If handle does not refer to a valid object opened by dlopen(), or if the named symbol cannot be found within any of the objects associated with handle, dlsym() shall return NULL.

But then there's the issue of the special values of handle, RTLD_DEFAULT and RTLD_NEXT. I suspect either of these could be defined as (void*)0, but in any case the specific value doesn't matter, you just can't use whatever their value is unless you want that behavior.

Seems like from the below code that RTLD_DEFAULT is the desired behavior, so maybe that should be returned to ensure that we're not triggering undefined behavior.

FWIW, this is reverting to behavior prior to the move to using pathfinder, so it's been testing for a long time in the wild with handle == NULL in those situations.

But to your question, it's kinda/sorta. On GNU, the special value RTLD_DEFAULT is equal to 0, so it's defined behavior to lookup the symbol in the global namespace -- for example if the library is statically linked. All that is fine -- we are already doing that before looking in the .so. Other POSIX platforms mileage may vary, of course.

Jinx. Yeah, I agree that maybe setting to RTLD_DEFAULT here would be better.

Windows is another kettle of fish -- something else might be better there.

Ok, yeah jinxing all over the place 🤣

~~For Windows it looks like its dlsym-equivalent has a similar behavior to RTLD_DEFAULT specifically when it is given NULL.~~

Not sure about this actually.

FWIW, this is reverting to behavior prior to the move to using pathfinder, so it's been testing for a long time in the wild with handle == NULL in those situations.

I was wondering about this too. Turns out we did not test this behavior at all. This is from 12.8.0 (whose behavior we should restore to):

cuda-python/cuda_bindings/cuda/bindings/_internal/nvjitlink_linux.pyx

Lines 55 to 65 in c04025d

cdef void* load_library(const int driver_ver) except* with gil:

cdef void* handle

for suffix in get_nvjitlink_dso_version_suffix(driver_ver):

so_name = "libnvJitLink.so" + (f".{suffix}" if suffix else suffix)

handle = dlopen(so_name.encode(), RTLD_NOW | RTLD_GLOBAL)

if handle != NULL:

break

else:

err_msg = dlerror()

raise RuntimeError(f'Failed to dlopen libnvJitLink ({err_msg.decode()})')

return handle

If a library is not loadable, we just raised an exception in load_library, so we never returned the null handle to complete the remaining symbol loading. I think to fix #951 we just need to swallow the exception raised by the pathfinder, and raise the one we used to raise?

Ah, ok. I missed that in all this. That certainly makes this less weird. Will change.

Hmm... That can't be the whole story. A DynamicLibNotFoundError is a subclass of RuntimeError. Puzzle to solve...

github-actions · 2025-09-30T17:28:10Z

Doc Preview CI
🚀 View preview at https://nvidia.github.io/cuda-python/pr-preview/pr-1052/
https://nvidia.github.io/cuda-python/pr-preview/pr-1052/cuda-core/
https://nvidia.github.io/cuda-python/pr-preview/pr-1052/cuda-bindings/
https://nvidia.github.io/cuda-python/pr-preview/pr-1052/cuda-pathfinder/
Preview will be ready when the GitHub Pages deployment is complete.

copy-pr-bot · 2025-09-30T19:03:54Z

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

mdboom · 2025-09-30T19:48:01Z

Closing -- I'm not sure it's fixing the right thing. See comment: #951 (comment)

Removed preview folders for the following PRs: - PR #1021 - PR #1034 - PR #1052 - PR #1059 - PR #1069 - PR #1086 - PR #1090 - PR #1096 - PR #1102 - PR #1103 - PR #1106 - PR #1107 - PR #1117 - PR #1133 - PR #1140 - PR #1166 - PR #1174 - PR #1185 - PR #1188 - PR #1191 ... and 41 more

mdboom added 2 commits September 30, 2025 13:07

Fix NVIDIA#951: Fallback gracefully if a library can't be loaded

f46d0f4

Add relnote

3ea9f69

mdboom self-assigned this Sep 30, 2025

mdboom requested review from leofang and rwgk September 30, 2025 17:12

cpcloud reviewed Sep 30, 2025

View reviewed changes

mdboom marked this pull request as draft September 30, 2025 19:03

mdboom closed this Sep 30, 2025

mdboom deleted the issue-951 branch December 9, 2025 16:13

	cdef void* load_library(const int driver_ver) except* with gil:
	cdef void* handle
	for suffix in get_nvjitlink_dso_version_suffix(driver_ver):
	so_name = "libnvJitLink.so" + (f".{suffix}" if suffix else suffix)
	handle = dlopen(so_name.encode(), RTLD_NOW \| RTLD_GLOBAL)
	if handle != NULL:
	break
	else:
	err_msg = dlerror()
	raise RuntimeError(f'Failed to dlopen libnvJitLink ({err_msg.decode()})')
	return handle

Conversation

mdboom commented Sep 30, 2025

Description

Checklist

Uh oh!

copy-pr-bot bot commented Sep 30, 2025

Uh oh!

mdboom commented Sep 30, 2025

Uh oh!

cpcloud Sep 30, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

cpcloud Sep 30, 2025

Choose a reason for hiding this comment

Uh oh!

cpcloud Sep 30, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

mdboom Sep 30, 2025

Choose a reason for hiding this comment

Uh oh!

mdboom Sep 30, 2025

Choose a reason for hiding this comment

Uh oh!

cpcloud Sep 30, 2025

Choose a reason for hiding this comment

Uh oh!

cpcloud Sep 30, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

leofang Sep 30, 2025

Choose a reason for hiding this comment

Uh oh!

mdboom Sep 30, 2025

Choose a reason for hiding this comment

Uh oh!

mdboom Sep 30, 2025

Choose a reason for hiding this comment

Uh oh!

github-actions bot commented Sep 30, 2025

Preview will be ready when the GitHub Pages deployment is complete.

Uh oh!

copy-pr-bot bot commented Sep 30, 2025

Uh oh!

mdboom commented Sep 30, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

cpcloud Sep 30, 2025 •

edited

Loading

cpcloud Sep 30, 2025 •

edited

Loading

cpcloud Sep 30, 2025 •

edited

Loading