[PATCH] test/debug: fix crash with mlx5 devices

[PATCH] test/debug: fix crash with mlx5 devices

[David Marchand]

Running rte_exit() in a forked process means that shared memory will be
released by the child process before the parent process does the same.
This issue has been seen recently when some GHA virtual machine (with
some mlx5 devices) runs the debug_autotest unit test.

Instead, run rte_panic() and rte_exit() from a new DPDK process spawned
like for other recursive unit tests.

Bugzilla ID: 1796
Fixes: af75078fece3 ("first public release")
Cc: stable@dpdk.org

Signed-off-by: David Marchand <david.marchand@redhat.com>
---
 app/test/process.h    |  2 +-
 app/test/test.c       |  2 +
 app/test/test.h       |  2 +
 app/test/test_debug.c | 94 ++++++++++++++++++++++++++++++-------------
 4 files changed, 71 insertions(+), 29 deletions(-)

diff --git a/app/test/process.h b/app/test/process.h
index 9fb2bf481c..8e11d0b059 100644
--- a/app/test/process.h
+++ b/app/test/process.h
@@ -203,7 +203,7 @@ process_dup(const char *const argv[], int numargs, const char *env_value)
  * tests attempting to use this function on FreeBSD.
  */
 #ifdef RTE_EXEC_ENV_LINUX
-static char *
+static inline char *
 get_current_prefix(char *prefix, int size)
 {
 	char path[PATH_MAX] = {0};
diff --git a/app/test/test.c b/app/test/test.c
index fd653cbbfd..8a4598baee 100644
--- a/app/test/test.c
+++ b/app/test/test.c
@@ -80,6 +80,8 @@ do_recursive_call(void)
 			{ "test_memory_flags", no_action },
 			{ "test_file_prefix", no_action },
 			{ "test_no_huge_flag", no_action },
+			{ "test_panic", test_panic },
+			{ "test_exit", test_exit },
 #ifdef RTE_LIB_TIMER
 #ifndef RTE_EXEC_ENV_WINDOWS
 			{ "timer_secondary_spawn_wait", test_timer_secondary },
diff --git a/app/test/test.h b/app/test/test.h
index ebc4864bf8..c6d7d23313 100644
--- a/app/test/test.h
+++ b/app/test/test.h
@@ -174,7 +174,9 @@ extern const char *prgname;
 int commands_init(void);
 int command_valid(const char *cmd);
 
+int test_exit(void);
 int test_mp_secondary(void);
+int test_panic(void);
 int test_timer_secondary(void);
 
 int test_set_rxtx_conf(cmdline_fixed_string_t mode);
diff --git a/app/test/test_debug.c b/app/test/test_debug.c
index 8ad6d40fcb..623b80a753 100644
--- a/app/test/test_debug.c
+++ b/app/test/test_debug.c
@@ -8,6 +8,18 @@
 #include <stdint.h>
 
 #ifdef RTE_EXEC_ENV_WINDOWS
+int
+test_panic(void)
+{
+	printf("debug not supported on Windows, skipping test\n");
+	return TEST_SKIPPED;
+}
+int
+test_exit(void)
+{
+	printf("debug not supported on Windows, skipping test\n");
+	return TEST_SKIPPED;
+}
 static int
 test_debug(void)
 {
@@ -25,34 +37,33 @@ test_debug(void)
 #include <rte_debug.h>
 #include <rte_common.h>
 #include <rte_eal.h>
-#include <rte_service_component.h>
+#include <rte_lcore.h>
+
+#include "process.h"
+
+#define launch_proc(ARGV) process_dup(ARGV, RTE_DIM(ARGV), __func__)
 
 /*
  * Debug test
  * ==========
  */
 
-/* use fork() to test rte_panic() */
-static int
+static const char *test_args[7];
+
+int
 test_panic(void)
 {
-	int pid;
 	int status;
 
-	pid = fork();
-
-	if (pid == 0) {
+	if (getenv(RECURSIVE_ENV_VAR) != NULL) {
 		struct rlimit rl;
 
 		/* No need to generate a coredump when panicking. */
 		rl.rlim_cur = rl.rlim_max = 0;
 		setrlimit(RLIMIT_CORE, &rl);
 		rte_panic("Test Debug\n");
-	} else if (pid < 0) {
-		printf("Fork Failed\n");
-		return -1;
 	}
-	wait(&status);
+	status = launch_proc(test_args);
 	if(status == 0){
 		printf("Child process terminated normally!\n");
 		return -1;
@@ -62,27 +73,16 @@ test_panic(void)
 	return 0;
 }
 
-/* use fork() to test rte_exit() */
 static int
 test_exit_val(int exit_val)
 {
-	int pid;
+	char buf[5];
 	int status;
 
-	/* manually cleanup EAL memory, as the fork() below would otherwise
-	 * cause the same hugepages to be free()-ed multiple times.
-	 */
-	rte_service_finalize();
-
-	pid = fork();
-
-	if (pid == 0)
-		rte_exit(exit_val, __func__);
-	else if (pid < 0){
-		printf("Fork Failed\n");
-		return -1;
-	}
-	wait(&status);
+	sprintf(buf, "%d", exit_val);
+	if (setenv("TEST_DEBUG_EXIT_VAL", buf, 1) == -1)
+		rte_panic("Failed to set exit value in env\n");
+	status = launch_proc(test_args);
 	printf("Child process status: %d\n", status);
 	if(!WIFEXITED(status) || WEXITSTATUS(status) != (uint8_t)exit_val){
 		printf("Child process terminated with incorrect status (expected = %d)!\n",
@@ -92,11 +92,22 @@ test_exit_val(int exit_val)
 	return 0;
 }
 
-static int
+int
 test_exit(void)
 {
 	int test_vals[] = { 0, 1, 2, 255, -1 };
 	unsigned i;
+
+	if (getenv(RECURSIVE_ENV_VAR) != NULL) {
+		int exit_val;
+
+		if (!getenv("TEST_DEBUG_EXIT_VAL"))
+			rte_panic("No exit value set in env\n");
+
+		exit_val = strtol(getenv("TEST_DEBUG_EXIT_VAL"), NULL, 0);
+		rte_exit(exit_val, __func__);
+	}
+
 	for (i = 0; i < RTE_DIM(test_vals); i++) {
 		if (test_exit_val(test_vals[i]) < 0)
 			return -1;
@@ -128,6 +139,33 @@ test_usage(void)
 static int
 test_debug(void)
 {
+#ifdef RTE_EXEC_ENV_FREEBSD
+	/* BSD target doesn't support prefixes at this point, and we also need to
+	 * run another primary process here.
+	 */
+	const char * prefix = "--no-shconf";
+#else
+	const char * prefix = "--file-prefix=debug";
+#endif
+	char core[10];
+
+	sprintf(core, "%d", rte_get_main_lcore());
+
+	test_args[0] = prgname;
+	test_args[1] = prefix;
+	test_args[2] = "-l";
+	test_args[3] = core;
+
+	if (rte_eal_has_hugepages()) {
+		test_args[4] = "";
+		test_args[5] = "";
+		test_args[6] = "";
+	} else {
+		test_args[4] = "--no-huge";
+		test_args[5] = "-m";
+		test_args[6] = "2048";
+	}
+
 	rte_dump_stack();
 	if (test_panic() < 0)
 		return -1;
-- 
2.51.0

Re: [PATCH] test/debug: fix crash with mlx5 devices

[David Marchand]

On Thu, 2 Oct 2025 at 18:56, David Marchand <david.marchand@redhat.com> wrote:
>
> Running rte_exit() in a forked process means that shared memory will be
> released by the child process before the parent process does the same.

Another way of describing the issue is that the unit test code by
forking at rte_eal_init made it so two primary processes were
accessing the DPDK shared memory.
Nothing good can come from this...


-- 
David Marchand