fix: reload_nvidia_peermem uses 'sleep inf' and trap after sleep

[andrewwhitecdw]

This PR addresses the following issue in ubuntu24.04/nvidia-driver: reload_nvidia_peermem uses 'sleep inf' and trap after sleep.

Changes

• ubuntu24.04/nvidia-driver: reload_nvidia_peermem uses 'sleep inf' and trap after sleep.

Details

--- a/ubuntu24.04/nvidia-driver
+++ b/ubuntu24.04/nvidia-driver
@@ -1,33 +1,33 @@
-reload_nvidia_peermem() {
-    if [ "$USE_HOST_MOFED" = "true" ]; then
-        until  lsmod | grep mlx5_core > /dev/null 2>&1 && [ -f /run/nvidia/validations/.driver-ctr-ready ];
-        do
-            echo "waiting for mellanox ofed and nvidia drivers to be installed"
-            sleep 10
-        done
-    else
-        # use driver readiness flag created by MOFED container
-        until  [ -f /run/mellanox/drivers/.driver-ready ] && [ -f /run/nvidia/validations/.driver-ctr-ready ];
-        do
-            echo "waiting for mellanox ofed and nvidia drivers to be installed"
-            sleep 10
-        done
-    fi
-    # get any parameters provided for nvidia-peermem
-    _get_module_params && set +o nounset
-    # If nvidia-peermem is already loaded (e.g. fast-path restart), skip modprobe
-    if [ -f /sys/module/nvidia_peermem/refcnt ]; then
-        echo "nvidia-peermem module already loaded, skipping modprobe"
-        sleep inf
-        trap "echo 'Caught signal'; exit 1" HUP INT QUIT PIPE TERM
-    fi
-    if chroot /run/nvidia/driver modprobe nvidia-peermem "${NVIDIA_PEERMEM_MODULE_PARAMS[@]}"; then
-        if [ -f /sys/module/nvidia_peermem/refcnt ]; then
-            echo "successfully loaded nvidia-peermem module, now waiting for signal"
-            sleep inf
-            trap "echo 'Caught signal'; exit 1" HUP INT QUIT PIPE TERM
-        fi
-    fi
-    echo "failed to load nvidia-peermem module"
-    exit 1
-}
+reload_nvidia_peermem() {
+    if [ "$USE_HOST_MOFED" = "true" ]; then
+        until  lsmod | grep mlx5_core > /dev/null 2>&1 && [ -f /run/nvidia/validations/.driver-ctr-ready ];
+        do
+            echo "waiting for mellanox ofed and nvidia drivers to be installed"
+            sleep 10
+        done
+    else
+        # use driver readiness flag created by MOFED container
+        until  [ -f /run/mellanox/drivers/.driver-ready ] && [ -f /run/nvidia/validations/.driver-ctr-ready ];
+        do
+            echo "waiting for mellanox ofed and nvidia drivers to be installed"
+            sleep 10
+        done
+    fi
+    # get any parameters provided for nvidia-peermem
+    _get_module_params && set +o nounset
+    # If nvidia-peermem is already loaded (e.g. fast-path restart), skip modprobe
+    if [ -f /sys/module/nvidia_peermem/refcnt ]; then
+        echo "nvidia-peermem module already loaded, skipping modprobe"
+        trap "echo 'Caught signal'; exit 1" HUP INT QUIT PIPE TERM
+        sleep infinity
+    fi
+    if chroot /run/nvidia/driver modprobe nvidia-peermem "${NVIDIA_PEERMEM_MODULE_PARAMS[@]}"; then
+        if [ -f /sys/module/nvidia_peermem/refcnt ]; then
+            echo "successfully loaded nvidia-peermem module, now waiting for signal"
+            trap "echo 'Caught signal'; exit 1" HUP INT QUIT PIPE TERM
+            sleep infinity
+        fi
+    fi
+    echo "failed to load nvidia-peermem module"
+    exit 1
+}

Tests

• tests/test_nvidia-driver.sh

--- /dev/null
+++ tests/test_nvidia-driver.sh
@@ -0,0 +1,35 @@
+#!/bin/bash
+# Minimal regression tests for ubuntu24.04/nvidia-driver
+# Usage: bash tests/test_nvidia-driver.sh [path-to-nvidia-driver]
+set -eu
+
+script="${1:-ubuntu24.04/nvidia-driver}"
+fail=0
+
+test_reload_peermem_wait() {
+    if grep -q 'sleep inf' "$script"; then
+        echo "FAIL: 'sleep inf' is not portable; use 'sleep infinity'"
+        fail=1
+    fi
+
+    awk '/^reload_nvidia_peermem\(\)/,/^\}/' "$script" > /tmp/reload_peermem_func.txt
+    while grep -q 'sleep' /tmp/reload_peermem_func.txt; do
+        line=$(grep -n 'sleep' /tmp/reload_peermem_func.txt | head -1 | cut -d: -f1)
+        prev=$((line - 1))
+        if ! sed -n "${prev}p" /tmp/reload_peermem_func.txt | grep -q 'trap'; then
+            echo "FAIL: reload_nvidia_peermem sleeps without a preceding signal trap"
+            fail=1
+        fi
+        sed -i "${line}d" /tmp/reload_peermem_func.txt
+    done
+}
+
+test_reload_peermem_wait
+
+# --- next test insertion point ---
+
+exit $fail

Contributor guidelines

Per this repo's CONTRIBUTING.md:

• All commits are signed off (Signed-off-by trailer, DCO).

[copy-pr-bot]

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.