Skip to contentSkip to conversation

pi_stress: Bug fixes and improvements

PATCH New John Kacur 1 participant 6 messages Jun 30

lore
Share & tools
lore

John Kacur

author
#0 lore raw

cover pi_stress: Bug fixes and improvements

This series addresses several issues in pi_stress:

- Fixes a critical deadlock bug in error handling where inconsistent barrier wait logic could cause the program to hang
- Improves input validation to reject malformed numeric arguments
- Enhances error messages for better user experience
- Fixes cosmetic issues in debug/error output
- Ensures proper string null-termination

The first patch is the most important as it fixes a real deadlock scenario during error conditions.

John Kacur (5):
pi_stress: Fix inconsistent barrier wait logic to prevent deadlock
pi_stress: Validate numeric command-line arguments
pi_stress: Improve error message for permission failures
pi_stress: Fix error and debug message formatting
pi_stress: Ensure null-termination of JSON filename

 src/pi_tests/pi_stress.c | 44 ++++++++++++++++++++++++++--------------
 1 file changed, 29 insertions(+), 15 deletions(-)
Signature · 1 line
2.54.0
author
#1 lore raw

PATCH 1/5 pi_stress: Fix inconsistent barrier wait logic to prevent deadlock

The low_priority thread function unconditionally waited on the all_threads_done barrier, while med_priority and high_priority only waited when have_errors == 0. This inconsistency could cause deadlock during error conditions where med_priority and high_priority threads exit without waiting, leaving low_priority threads waiting forever.

Make low_priority consistent with the other thread functions by adding the same have_errors check before waiting on all_threads_done.

Assisted-by Claude:claude-sonnet-4-5
---
 src/pi_tests/pi_stress.c | 17 ++++++++++-------
 1 file changed, 10 insertions(+), 7 deletions(-)
src/pi_tests/pi_stress.c +10 -7
index 74e7bde0c21f..b7f7515fa9b1 100644
@@ -605,13 +605,16 @@ void *low_priority(void *arg)
 	}
 	set_shutdown_flag();
 	pi_debug("low_priority[%d]: entering done barrier\n", p->id);
-	/* wait for all threads to finish */
-	status = pthread_barrier_wait(&all_threads_done);
-	if (status && status != PTHREAD_BARRIER_SERIAL_THREAD) {
-		pi_error
-		    ("low_priority[%d]: pthread_barrier_wait(all_threads_done): %x",
-		     p->id, status);
-		return NULL;
+
+	if (have_errors == 0) {
+		/* wait for all threads to finish */
+		status = pthread_barrier_wait(&all_threads_done);
+		if (status && status != PTHREAD_BARRIER_SERIAL_THREAD) {
+			pi_error
+			    ("low_priority[%d]: pthread_barrier_wait(all_threads_done): %x",
+			     p->id, status);
+			return NULL;
+		}
 	}
 	pi_debug("low_priority[%d]: exiting\n", p->id);
 	return NULL;
Signature · 1 line
2.54.0
author
#2 lore raw

PATCH 2/5 pi_stress: Validate numeric command-line arguments

In reply to John Kacur

The parse_unsigned() and parse_signed() functions were using strtoul() and strtol() but not checking if the entire string was consumed. This allowed invalid input like "123abc" to be silently accepted as "123".

Additionally, the -g (groups) and -i (inversions) options were calling strtol() directly instead of using the parse_signed() helper function, bypassing any validation entirely.

Add validation to parse_unsigned() and parse_signed() to ensure the entire input string is a valid number by checking that the end pointer points to a null terminator. Update -g and -i options to use parse_signed() for consistent error handling.

Assisted-by Claude:claude-sonnet-4-5
---
 src/pi_tests/pi_stress.c | 14 ++++++++++++--
 1 file changed, 12 insertions(+), 2 deletions(-)
src/pi_tests/pi_stress.c +12 -2
index b7f7515fa9b1..c8c895a71e44 100644
@@ -1092,6 +1092,11 @@ unsigned long parse_unsigned(const char *str)
 		exit(EXIT_FAILURE);
 	}
 
+	if (*p != '\0') {
+		pi_error("invalid number format: %s\n", str);
+		exit(EXIT_FAILURE);
+	}
+
 	return n;
 }
 
@@ -1109,6 +1114,11 @@ long parse_signed(const char *str)
 		exit(EXIT_FAILURE);
 	}
 
+	if (*p != '\0') {
+		pi_error("invalid number format: %s\n", str);
+		exit(EXIT_FAILURE);
+	}
+
 	return n;
 }
 
@@ -1211,7 +1221,7 @@ void process_command_line(int argc, char **argv)
 			break;
 		case OPT_GROUPS:
 		case 'g':
-			ngroups = strtol(optarg, NULL, 10);
+			ngroups = parse_signed(optarg);
 			if (ngroups > num_processors) {
 				pi_error("the number of groups cannot exceed the number of online processors (%ld)\n", num_processors);
 				exit(-1);
@@ -1224,7 +1234,7 @@ void process_command_line(int argc, char **argv)
 			break;
 		case OPT_INVERSIONS:
 		case 'i':
-			inversions = strtol(optarg, NULL, 10);
+			inversions = parse_signed(optarg);
 			pi_info("doing %d inversion per group\n", inversions);
 			break;
 		case OPT_JSON:
Signature · 1 line
2.54.0
author
#3 lore raw

PATCH 3/5 pi_stress: Improve error message for permission failures

In reply to John Kacur

When pi_stress fails to set real-time scheduling priority, it printed a cryptic hex error code (0x1) that didn't explain what went wrong or how to fix it.

Replace the error message with a clear explanation that indicates the failure reason and suggests running with root privileges.

Before:

    main: boosting to max priority: 0x1

After:
Failed to set real-time priority: Operation not permitted Please run with root privileges to set real-time priorities.

Assisted-by Claude:claude-sonnet-4-5
---
 src/pi_tests/pi_stress.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
src/pi_tests/pi_stress.c +2 -2
index c8c895a71e44..675953923c3a 100644
@@ -134,8 +134,8 @@ int main(int argc, char **argv)
 	status = pthread_setschedparam(pthread_self(), admin_sa.sched_policy,
 				       &thread_param);
 	if (status) {
-		pi_error("%s: boosting to max priority: 0x%x\n", __func__,
-			 status);
+		pi_error("Failed to set real-time priority: %s\n", strerror(status));
+		pi_error("Please run with root privileges to set real-time priorities.\n");
 		return FAILURE;
 	}
 	/* block unwanted signals */
Signature · 1 line
2.54.0
author
#4 lore raw

PATCH 4/5 pi_stress: Fix error and debug message formatting

In reply to John Kacur

Fix three message formatting issues:

1. Incorrect barrier name in error message (line 212): The error referenced "all_threads_ready" but was actually checking "all_threads_done"

2. Typo in SCHED_DEADLINE error message (line 746): Fixed malformed parenthesis in "affinity*dl)" to "affinity (dl)"

3. Wrong format specifier (line 808): Variable 'count' is unsigned long but was using %d instead of %lu

These are cosmetic fixes that improve debugging clarity but don't affect program behavior.

Assisted-by Claude:claude-sonnet-4-5
---
 src/pi_tests/pi_stress.c | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)
src/pi_tests/pi_stress.c +3 -3
index 675953923c3a..45ccac739646 100644
@@ -209,7 +209,7 @@ int main(int argc, char **argv)
 		pi_info("waiting for all threads to complete\n");
 		status = pthread_barrier_wait(&all_threads_done);
 		if (status && status != PTHREAD_BARRIER_SERIAL_THREAD) {
-			pi_error("%s: pthread_barrier_wait(all_threads_ready): 0x%x\n", __func__, status);
+			pi_error("%s: pthread_barrier_wait(all_threads_done): 0x%x\n", __func__, status);
 			return FAILURE;
 		}
 		pi_info("All threads terminated!\n");
@@ -743,7 +743,7 @@ void *high_priority(void *arg)
 		status = sched_setaffinity(0, sizeof(cpu_set_t), &cpu_mask);
 		if (status < 0) {
 			pi_error
-			    ("high_priority[%d]: set cpu affinity*dl): %x\n",
+			    ("high_priority[%d]: set cpu affinity (dl): %x\n",
 			    p->id, status);
 			return NULL;
 		}
@@ -805,7 +805,7 @@ void *high_priority(void *arg)
 			}
 			pthread_mutex_unlock(&shutdown_mtx);
 		}
-		pi_debug("high_priority[%d]: entering start state (%d)\n", p->id,
+		pi_debug("high_priority[%d]: entering start state (%lu)\n", p->id,
 		      count++);
 		status = pthread_barrier_wait(&p->start_barrier);
 		if (status && status != PTHREAD_BARRIER_SERIAL_THREAD) {
Signature · 1 line
2.54.0
author
#5 lore raw

PATCH 5/5 pi_stress: Ensure null-termination of JSON filename

In reply to John Kacur

The --json option used strncpy() with strnlen() to copy the filename, but did not guarantee null-termination if the input string was exactly MAX_PATH-1 characters long. When strncpy() copies exactly n characters, it does not add a null terminator, potentially causing undefined behavior when the string is later used.

Replace the strnlen() call with MAX_PATH-1 as the copy limit and explicitly null-terminate the buffer to ensure the string is always properly terminated regardless of input length.

Assisted-by Claude:claude-sonnet-4-5
---
 src/pi_tests/pi_stress.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
src/pi_tests/pi_stress.c +2 -1
index 45ccac739646..500b10d60737 100644
@@ -1238,7 +1238,8 @@ void process_command_line(int argc, char **argv)
 			pi_info("doing %d inversion per group\n", inversions);
 			break;
 		case OPT_JSON:
-			strncpy(jsonfile, optarg, strnlen(optarg, MAX_PATH-1));
+			strncpy(jsonfile, optarg, MAX_PATH-1);
+		jsonfile[MAX_PATH-1] = '\0';
 			break;
 		case OPT_MLOCKALL:
 		case 'm':
Signature · 1 line
2.54.0