fix hang caused by iopoll+eventfd

.github/workflows/build-and-test-latest.yml

@@ -87,7 +87,7 @@ jobs:
       - name: Run Tests on QEMU
         timeout-minutes: 15
         run: |
-          bash ./scripts/vm-run.sh -c "./tests" ./.cache/bzImage ./build/tests/tests 2>&1 | tee vm-run.log
+          bash ./scripts/vm-run.sh -c "CONDY_TEST_SSD_BASE_DIR=/mnt/ssd ./tests" ./.cache/bzImage ./build/tests/tests 2>&1 | tee vm-run.log
           if grep -q "SUCCESS!" vm-run.log; then
             echo "::notice:: All tests passed"
           else

include/condy/runtime.hpp

@@ -322,6 +322,11 @@ class Runtime {
             // No-op
             assert(cqe->res != -EINVAL); // If EINVAL, something is wrong
         } else if (type == WorkType::Notify) {
+            if (cqe->res == -EOPNOTSUPP) {
+                // Notification not supported, ignore. This may happen if we use
+                // eventfd for notification and iopoll is enabled.
+                return;
+            }
             std::lock_guard<std::mutex> lock(mutex_);
             flush_global_queue_();
         } else if (type == WorkType::SendFd) {

scripts/light-kernel.sh

@@ -56,6 +56,12 @@ cd "$SOURCE_DIR"
 make defconfig
 # Configure kvm guest support
 make kvm_guest.config
+# Configure NVMe support
+./scripts/config --enable CONFIG_NVME_CORE
+./scripts/config --enable CONFIG_BLK_DEV_NVME
+./scripts/config --enable CONFIG_NVME_MULTIPATH
+# Use olddefconfig to set new options to default values
+make olddefconfig
 
 # Build the kernel
 make -j"$(nproc)" bzImage

scripts/vm-run.sh

@@ -73,29 +73,43 @@ trap "rm -rf $TEMP_DIR" EXIT
 
 cat > "$TEMP_DIR/init.sh" <<EOF
 #!/bin/sh
+mkdir -p /mnt/ssd
+mount -t ext4 /dev/nvme0n1 /mnt/ssd
 $COMMAND
+umount /mnt/ssd
 EOF
 chmod +x "$TEMP_DIR/init.sh"
 
 echo "Building initrd image..."
 bash "$LIGHT_INITRD_SCRIPT" -o "$TEMP_DIR/$INITRD_OUTPUT" -s "/root/init.sh" $FILES "$TEMP_DIR/init.sh"
 echo "Initrd image created at $TEMP_DIR/$INITRD_OUTPUT"
 
+# Simulate NVMe SSD with a tmpfs-backed disk image
+SSD_IMG="/dev/shm/vm-ssd.img.$$"
+truncate -s 1G "$SSD_IMG"
+mkfs.ext4 -q "$SSD_IMG"
+trap "rm -f '$SSD_IMG'; rm -rf $TEMP_DIR" EXIT
+SSD_DRIVE="-drive file=$SSD_IMG,if=none,id=ssd0,format=raw,cache=none,aio=io_uring"
+SSD_DEVICE="-device nvme,drive=ssd0,serial=ssd0"
+
 KVM_FLAG=""
 if [ "$KVM_ENABLED" = true ]; then
     KVM_FLAG="-enable-kvm"
 fi
 
-KERNEL_ARGS="console=ttyS0"
+KERNEL_ARGS="console=ttyS0 nvme.poll_queues=2"
 if [ "$QUIET_MODE" = true ]; then
     KERNEL_ARGS="$KERNEL_ARGS quiet"
 fi
 
 # Run the VM,
 qemu-system-x86_64 \
     $KVM_FLAG \
+    -smp 4 \
     -m 512M \
     -kernel "$KERNEL_IMAGE" \
     -initrd "$TEMP_DIR/$INITRD_OUTPUT" \
     -append "$KERNEL_ARGS" \
-    -nographic
+    -nographic \
+    $SSD_DRIVE \
+    $SSD_DEVICE

tests/test_runtime_options.cpp

@@ -4,6 +4,7 @@
 #include "condy/sync_wait.hpp"
 #include "condy/task.hpp"
 #include <doctest/doctest.h>
+#include <limits>
 
 TEST_CASE("test runtime_options - event_interval") {
     condy::RuntimeOptions options;
@@ -28,11 +29,19 @@ TEST_CASE("test runtime_options - event_interval") {
         }
     };
 
-    condy::sync_wait(func());
+    condy::sync_wait(runtime, func());
 }
 
 TEST_CASE("test runtime_options - enable_iopoll") {
-    char name[32] = "XXXXXX";
+    const char *ssd_base_dir = std::getenv("CONDY_TEST_SSD_BASE_DIR");
+    if (ssd_base_dir == nullptr) {
+        MESSAGE("CONDY_TEST_SSD_BASE_DIR not set, skipping");
+        return;
+    }
+    std::string template_path = std::string(ssd_base_dir) + "/XXXXXX";
+    char name[32];
+    strncpy(name, template_path.c_str(), sizeof(name));
+    name[sizeof(name) - 1] = '\0';
     int fd = mkstemp(name);
     REQUIRE(fd >= 0);
 
@@ -56,6 +65,8 @@ TEST_CASE("test runtime_options - enable_iopoll") {
 #else
     options.enable_iopoll(/*hybrid=*/false);
 #endif
+    // Disable periodic event peeking to exposure hang caused by eventfd+iopoll
+    options.event_interval(std::numeric_limits<size_t>::max());
     condy::Runtime runtime(options);
 
     alignas(4096) char buffer[4096];
@@ -66,7 +77,7 @@ TEST_CASE("test runtime_options - enable_iopoll") {
         REQUIRE(std::string_view(buffer, msg.size()) == msg);
     };
 
-    condy::sync_wait(func());
+    condy::sync_wait(runtime, func());
 }
 
 TEST_CASE("test runtime_options - enable_sqpoll") {
@@ -92,7 +103,7 @@ TEST_CASE("test runtime_options - enable_sqpoll") {
         }
     };
 
-    condy::sync_wait(func());
+    condy::sync_wait(runtime, func());
 }
 
 TEST_CASE("test runtime_options - enable_defer_taskrun") {
@@ -118,7 +129,7 @@ TEST_CASE("test runtime_options - enable_defer_taskrun") {
         }
     };
 
-    condy::sync_wait(func());
+    condy::sync_wait(runtime, func());
 }
 
 TEST_CASE("test runtime_options - enable_attach_wq") {
@@ -178,7 +189,7 @@ TEST_CASE("test runtime_options - enable_coop_taskrun") {
         }
     };
 
-    condy::sync_wait(func());
+    condy::sync_wait(runtime, func());
 }
 
 TEST_CASE("test runtime_options - enable_sqe128 & enable_cqe32") {
@@ -204,7 +215,7 @@ TEST_CASE("test runtime_options - enable_sqe128 & enable_cqe32") {
         }
     };
 
-    condy::sync_wait(func());
+    condy::sync_wait(runtime, func());
 }
 
 TEST_CASE("test runtime_options - enable_no_mmap") {
@@ -235,5 +246,5 @@ TEST_CASE("test runtime_options - enable_no_mmap") {
         }
     };
 
-    condy::sync_wait(func());
+    condy::sync_wait(runtime, func());
 }