fix: Use system search path when locate driver library by reddevillg · Pull Request #1820 · NVIDIA/nvidia-container-toolkit

reddevillg · 2026-05-12T05:25:23Z

on Debian driver library may installed on both /usr/lib/<archtriplet> and /usr/lib/<archtriplet>/nvidia/current

copy-pr-bot · 2026-05-12T05:25:26Z

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

elezar · 2026-05-12T08:03:44Z

+		"/lib/aarch64-linux-gnu/nvidia/current",
+		libcudasoParentDirPath,
+    }
 	for _, dir := range additionalDirs {


We already search these paths for libcuda.so.RM_VERSION and now we're specifying them again. I agree that we can make this more robust, but I would rather use a specific set of directories here than broadening the set again.

As a follow-up question: Could we plumb through the other directories through additionalDirs instead of hardcoding them here?

elezar · 2026-05-12T08:23:46Z

@reddevillg I have added a commit on top that should allow multiple driver library paths to be supported. Could you confirm that this works in your environment?

reddevillg · 2026-05-12T09:17:55Z

@reddevillg I have added a commit on top that should allow multiple driver library paths to be supported. Could you confirm that this works in your environment?

It works perfectly.

elezar · 2026-05-12T11:47:15Z

@reddevillg I have added a commit on top that should allow multiple driver library paths to be supported. Could you confirm that this works in your environment?

It works perfectly.

That's great! Did you want to squash the two commits and then we could get @cdesiniotis or @tariq1890 to also review this?

reddevillg · 2026-05-12T12:47:26Z

@reddevillg I have added a commit on top that should allow multiple driver library paths to be supported. Could you confirm that this works in your environment?

It works perfectly.

That's great! Did you want to squash the two commits and then we could get @cdesiniotis or @tariq1890 to also review this?

it's ok, you can edit it directly, or tell me what I need to do

elezar · 2026-05-13T08:56:05Z

@reddevillg I have added a commit on top that should allow multiple driver library paths to be supported. Could you confirm that this works in your environment?

It works perfectly.

That's great! Did you want to squash the two commits and then we could get @cdesiniotis or @tariq1890 to also review this?

it's ok, you can edit it directly, or tell me what I need to do

I have squashed the commits.

tariq1890 · 2026-05-15T19:50:29Z

+
 // buildXOrgSearchPaths returns the ordered list of search paths for XOrg files.
-func buildXOrgSearchPaths(libRoot string) []string {
+func buildXOrgSearchPathsForSingle(libRoot string) []string {


Can we update the comment so it matches the new method name?

it has now been updated

tariq1890 · 2026-05-15T19:51:35Z

 	return match
 }

+func buildXOrgSearchPaths(roots ...string) []string {


This comment needs to be moved above this method?

// buildXOrgSearchPaths returns the ordered list of search paths for XOrg files.

Add comment "// buildXOrgSearchPaths returns search paths from all roots"

tariq1890 · 2026-05-15T19:52:42Z

 		driverLibPaths, err := r.Libraries().Locate(driverLib + versionSuffix)
 		if err != nil {
-			errs = errors.Join(errs, fmt.Errorf("failed to locate libcuda.so: %w", err))
+			errs = errors.Join(errs, fmt.Errorf("failed to locate %v: %w", driverLib, err))


Suggested change

errs = errors.Join(errs, fmt.Errorf("failed to locate %v: %w", driverLib, err))

errs = errors.Join(errs, fmt.Errorf("failed to locate %q: %w", driverLib, err))

It has already been fixed

On some systems (e.g. Debian), the driver libraries may be located in multiple paths on the system. This change allows these driver libraries to be found by locating all libraries with the expected version suffix. Signed-off-by: Evan Lezar <elezar@nvidia.com> Co-authored--by: Guo Lv <reddevillg@gmail.com> Signed-off-by: Guo Lv <reddevillg@gmail.com>

cdesiniotis · 2026-05-26T23:13:26Z

 		Name:  "NVIDIA_CTK_LIBCUDA_DIR",
-		Value: driverLibDirectory,
+		Value: strings.Join(driverLibDirectories, string(filepath.ListSeparator)),


@elezar can you remind me what tooling depends on NVIDIA_CTK_LIBCUDA_DIR? Since this now can return a :-separated list of directories, do any downstream projects need to be updated? #1030 seems to indicate NixOS uses this.

cdesiniotis · 2026-05-26T23:22:33Z

Can we add a test case to this file that locates two driver library directories?

cdesiniotis · 2026-05-26T23:25:15Z

+}
+
+// buildXOrgSearchPathsForSingle returns the ordered list of search paths for XOrg files.
+func buildXOrgSearchPathsForSingle(libRoot string) []string {


nit: can we rename this method to buildXOrgSearchPathsAtRoot?

reddevillg force-pushed the fix_lib_locate branch from d06b791 to 589fb1b Compare May 12, 2026 05:32

elezar reviewed May 12, 2026

View reviewed changes

elezar force-pushed the fix_lib_locate branch from 978c826 to 37ea7cc Compare May 13, 2026 08:55

elezar requested review from cdesiniotis and tariq1890 May 13, 2026 08:56

elezar added this to the next-minor milestone May 13, 2026

tariq1890 reviewed May 15, 2026

View reviewed changes

github-actions Bot mentioned this pull request May 19, 2026

Remove Vulkan ICD symlinks if CUDA packages update install path claytono/infra#1120

Open

reddevillg force-pushed the fix_lib_locate branch from 37ea7cc to a787f48 Compare May 26, 2026 05:16

reddevillg force-pushed the fix_lib_locate branch from a787f48 to 8e83f9f Compare May 26, 2026 05:21

reddevillg requested review from elezar and tariq1890 May 26, 2026 05:24

cdesiniotis reviewed May 26, 2026

View reviewed changes

cdesiniotis self-assigned this May 26, 2026

	errs = errors.Join(errs, fmt.Errorf("failed to locate %v: %w", driverLib, err))
	errs = errors.Join(errs, fmt.Errorf("failed to locate %q: %w", driverLib, err))

Conversation

reddevillg commented May 12, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

copy-pr-bot Bot commented May 12, 2026

Uh oh!

Choose a reason for hiding this comment

Uh oh!

elezar commented May 12, 2026

Uh oh!

reddevillg commented May 12, 2026

Uh oh!

elezar commented May 12, 2026

Uh oh!

reddevillg commented May 12, 2026

Uh oh!

elezar commented May 13, 2026

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

reddevillg commented May 12, 2026 •

edited

Loading