Skip to contentSkip to conversation

mem/min_free_kbytes: Fix incorrect pass/fail accounting

PATCH In review Sachin Sant 3 participants 4 messages Jul 9 – Jul 10

lore
Share & tools
lore
#0 lore raw

PATCH v2 1/1 mem/min_free_kbytes: Fix incorrect pass/fail accounting

When check_monitor() detected a violation (MemFree < min_free_kbytes) it called tst_res(TFAIL, ...) from the child process, which atomically incremented the shared results->failed counter, and then the child exited with status 0. Back in the parent, min_free_kbytes_test() checked the child exit status, found it was 0, and fell through to the unconditional tst_res(TPASS, ...) at the end of the function.

This produced a misleading summary of 'passed 1 / failed 1' on violation:
the TFAIL from the monitor child was correctly counted, but the unconditional TPASS that followed also added to the pass count regardless of the violation.

Fix this by moving TPASS/TFAIL reporting into check_monitor() in the child process. Emit tst_res(TFAIL, ...) immediately on each individual violation so that the TINFO diagnostic and its corresponding TFAIL are always paired; the 'violated' flag is kept to suppress the final TPASS when any breach was seen. Emit tst_res(TPASS, ...) only when no violation was observed across the entire monitoring run.
The parent's wait-result block is changed to guard against unexpected termination (signal death or non-zero exit) only; it emits no result for the monitor outcome since that is already reported by the child. The original condition WIFEXITED && WEXITSTATUS != 0 missed signal death; the corrected condition is !WIFEXITED || WEXITSTATUS != 0.

After the fix the summary correctly reflects the outcome:
- No violation: passed 1 / failed 0
- Violation: passed 0 / failed N (one TFAIL per breached sample)

---
v1 -> v2:
- Addressed review comments by moving TPASS/TFAIL reporting into check_monitor() in the child process.

---
 testcases/kernel/mem/tunable/min_free_kbytes.c | 13 ++++++++-----
 1 file changed, 8 insertions(+), 5 deletions(-)
testcases/kernel/mem/tunable/min_free_kbytes.c +8 -5
index bdc9126c2..132c8fe70 100644
@@ -68,11 +68,9 @@ static void min_free_kbytes_test(void)
 	SAFE_KILL(pid, SIGUSR1);
 	SAFE_WAITPID(pid, &status, WUNTRACED | WCONTINUED);
 
-	if (WIFEXITED(status) && WEXITSTATUS(status) != 0)
-		tst_res(TFAIL, "check_monitor child exit with status: %s",
+	if (!WIFEXITED(status) || WEXITSTATUS(status) != 0)
+		tst_res(TFAIL, "check_monitor child failed: %s",
 			tst_strstatus(status));
-
-	tst_res(TPASS, "min_free_kbytes test pass");
 }
 
 static void test_tune(unsigned long overcommit_policy)
@@ -173,6 +171,7 @@ static int eatup_mem(unsigned long overcommit_policy)
 
 static void check_monitor(void)
 {
+	int violated = 0;
 	unsigned long tune;
 	unsigned long memfree;
 
@@ -182,12 +181,16 @@ static void check_monitor(void)
 
 		if (memfree < tune) {
 			tst_res(TINFO, "MemFree is %lu kB, "
-				 "min_free_kbytes is %lu kB", memfree, tune);
+				"min_free_kbytes is %lu kB", memfree, tune);
 			tst_res(TFAIL, "MemFree < min_free_kbytes");
+			violated = 1;
 		}
 
 		sleep(2);
 	}
+
+	if (!violated)
+		tst_res(TPASS, "min_free_kbytes test pass");
 }
 
 static void sighandler(int signo LTP_ATTRIBUTE_UNUSED)
Signature · 5 lines
2.39.1


-- 
Mailing list info: https://lists.linux.it/listinfo/ltp
#1 lore raw

Hi Sachin,

Quoted 1 line · Sachin Sant

mem/min_free_kbytes: Fix incorrect pass/fail accounting

Verdict - Reviewed

---
Pre-existing issues:

In check_monitor(), the flag variable used for signal-handler communication is declared as:

volatile int end;

Strict C (C99 and later) requires volatile sig_atomic_t for variables read/written by a signal handler. This is pre-existing and harmless on common architectures, but worth fixing in a follow-up.

---
Note:

The agent can sometimes produce false positives although often its findings are genuine. If you find issues with the review, please comment this email or ignore the suggestions.

Regards,
LTP AI Reviewer

Signature · 1 line
Mailing list info: https://lists.linux.it/listinfo/ltp
#2 lore raw
In reply to Sachin Sant

Hi Sachin,

I think with this patch we are working around the problem. Which is.. the test is badly written and it needs a complete refactoring, since it seems like a mix of old and new API.

For instance, we could heavily simplify the test_tune(), adding a helper for setting the min free kb, then using SAFE_FORK():

static void set_min_free_kbytes(int i)

{
	unsigned long memfree, memtotal, tune;

	switch (i) {
	case 0:
		tune = default_tune;
		break;
	case 1:
		tune = 2 * default_tune;
		break;
	default:
		memfree = SAFE_READ_MEMINFO("MemFree:");
		memtotal = SAFE_READ_MEMINFO("MemTotal:");
		tune = MIN(memfree / 20, memtotal / 50);
		break;
	}

	TST_SYS_CONF_LONG_SET(PATH_VM_MIN_FREE_KBYTES, tune, 1);
}

and then handling SAFE_WAITPID() only once for all.

Anyway..below there are some considerations about the patch:

Quoted 12 lines

When check_monitor() detected a violation (MemFree < min_free_kbytes) it called tst_res(TFAIL, ...) from the child process, which atomically incremented the shared results->failed counter, and then the child exited with status 0. Back in the parent, min_free_kbytes_test() checked the child exit status, found it was 0, and fell through to the unconditional tst_res(TPASS, ...) at the end of the function.

This produced a misleading summary of 'passed 1 / failed 1' on violation:
the TFAIL from the monitor child was correctly counted, but the unconditional TPASS that followed also added to the pass count regardless of the violation.

Quoted 11 lines

Fix this by moving TPASS/TFAIL reporting into check_monitor() in the child process. Emit tst_res(TFAIL, ...) immediately on each individual violation so that the TINFO diagnostic and its corresponding TFAIL are always paired; the 'violated' flag is kept to suppress the final TPASS when any breach was seen. Emit tst_res(TPASS, ...) only when no violation was observed across the entire monitoring run.
The parent's wait-result block is changed to guard against unexpected termination (signal death or non-zero exit) only; it emits no result for the monitor outcome since that is already reported by the child. The original condition WIFEXITED && WEXITSTATUS != 0 missed signal death; the corrected condition is !WIFEXITED || WEXITSTATUS != 0.

Commit message is really verbose and a bit hard to follow without reading the code. It's good practice to describe why patch has been created in a short way so everyone will understand the patch by reading it at the very first look.

Quoted 20 lines

After the fix the summary correctly reflects the outcome:
- No violation: passed 1 / failed 0
- Violation: passed 0 / failed N (one TFAIL per breached sample)

---
v1 -> v2:
- Addressed review comments by moving TPASS/TFAIL reporting into check_monitor() in the child process.

---
 testcases/kernel/mem/tunable/min_free_kbytes.c | 13 ++++++++-----
 1 file changed, 8 insertions(+), 5 deletions(-)
testcases/kernel/mem/tunable/min_free_kbytes.c +0 -0
index bdc9126c2..132c8fe70 100644

testcases/kernel/mem/tunable/min_free_kbytes.c static void min_free_kbytes_test(void) · line 75

Quoted 10 lines
testcases/kernel/mem/tunable/min_free_kbytes.c +2 -4
@@ -68,11 +68,9 @@ static void min_free_kbytes_test(void)
 	SAFE_KILL(pid, SIGUSR1);
 	SAFE_WAITPID(pid, &status, WUNTRACED | WCONTINUED);
 
-	if (WIFEXITED(status) && WEXITSTATUS(status) != 0)
-		tst_res(TFAIL, "check_monitor child exit with status: %s",
+	if (!WIFEXITED(status) || WEXITSTATUS(status) != 0)
+		tst_res(TFAIL, "check_monitor child failed: %s",
 			tst_strstatus(status));
-
-	tst_res(TPASS, "min_free_kbytes test pass");

We can simply remove the whole check on the exit status, since LTP already calls tst_reap_children().

Quoted 16 lines
testcases/kernel/mem/tunable/min_free_kbytes.c +1 -1
 }
 
 static void test_tune(unsigned long overcommit_policy)
@@ -173,6 +171,7 @@ static int eatup_mem(unsigned long overcommit_policy)
 
 static void check_monitor(void)
 {
+	int violated = 0;
 	unsigned long tune;
 	unsigned long memfree;
 
@@ -182,12 +181,16 @@ static void check_monitor(void)
 
 		if (memfree < tune) {
 			tst_res(TINFO, "MemFree is %lu kB, "
-				 "min_free_kbytes is %lu kB", memfree, tune);

testcases/kernel/mem/tunable/min_free_kbytes.c static void check_monitor(void) · line 193

Quoted 10 lines
testcases/kernel/mem/tunable/min_free_kbytes.c +5 -0
+				"min_free_kbytes is %lu kB", memfree, tune);
 			tst_res(TFAIL, "MemFree < min_free_kbytes");
+			violated = 1;
 		}
 
 		sleep(2);
 	}
+
+	if (!violated)
+		tst_res(TPASS, "min_free_kbytes test pass");

This is correct and good practice. tst_reap_children() will take care of it.

Regards,
--
Andrea Cervesato
SUSE QE Automation Engineer Linux
[email protected]

Signature · 1 line
Mailing list info: https://lists.linux.it/listinfo/ltp
#3 lore raw
Quoted 5 lines · Andrea Cervesato

Hi Sachin,

I think with this patch we are working around the problem. Which is.. the test is badly written and it needs a complete refactoring, since it seems like a mix of old and new API.

Hi Andrea,

Thank you for the detailed review.

You are right — the original patch was a workaround rather than a proper fix.  The test was mixing old bare-fork patterns with the newer LTP API, and patching around the symptom without addressing the structural problem was the wrong approach.

I will rework the patch as a full refactor addressing all points raised:

1. Extract a set_min_free_kbytes() helper that consolidates the three tune-selection cases using a switch statement and MIN() as suggested, replacing the open-coded if/else if/else block in test_tune().

2. Replace all bare fork() calls with SAFE_FORK() throughout.  The eatup_mem children in test_tune() are still waited with SAFE_WAITPID and their exit status is still inspected per-policy, because overcommit_policy=1 expects SIGKILL — delegating those to tst_reap_children() would incorrectly produce TBROK on a valid outcome.

3. Drop the SAFE_WAITPID on the monitor child and its status check from min_free_kbytes_test() entirely.  The monitor child reports its own TPASS/TFAIL via tst_res() and always exits 0, so the framework's tst_reap_children() (called automatically after test_all returns) handles it correctly without any duplicate logic in the test.

4. Revise the commit message to be short and self-explanatory without re-narrating the diff.

I will send a v3 with the changes included.

Signature · 7 lines
Thanks
- Sachin



-- 
Mailing list info: https://lists.linux.it/listinfo/ltp