On Mon Mar 4, 2024 at 4:22 PM AEST, Thomas Huth wrote:

On 26/02/2024 10.38, Nicholas Piggin wrote:

quoted

This test stores to a bunch of pages and verifies previous stores,
while being continually migrated. This can fail due to a QEMU TCG
physical memory dirty bitmap bug.

Good idea, but could we then please drop "continuous" test from 
selftest-migration.c again? ... having two common tests to exercise the 
continuous migration that take quite a bunch of seconds to finish sounds 
like a waste of time in the long run to me.

Yeah if you like. I could shorten them up a bit. I did want to have
the selftests for just purely testing the harness with as little
"test" code as possible.

quoted

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
---
  common/memory-verify.c  | 48 +++++++++++++++++++++++++++++++++++++++++
  powerpc/Makefile.common |  1 +
  powerpc/memory-verify.c |  1 +
  powerpc/unittests.cfg   |  7 ++++++
  s390x/Makefile          |  1 +
  s390x/memory-verify.c   |  1 +
  s390x/unittests.cfg     |  6 ++++++
  7 files changed, 65 insertions(+)
  create mode 100644 common/memory-verify.c
  create mode 120000 powerpc/memory-verify.c
  create mode 120000 s390x/memory-verify.c

diff --git a/common/memory-verify.c b/common/memory-verify.c
new file mode 100644
index 000000000..7c4ec087b
--- /dev/null
+++ b/common/memory-verify.c

@@ -0,0 +1,48 @@
+// SPDX-License-Identifier: GPL-2.0-only
+/*
+ * Simple memory verification test, used to exercise dirty memory migration.
+ *
+ */
+#include <libcflat.h>
+#include <migrate.h>
+#include <alloc.h>
+#include <asm/page.h>
+#include <asm/time.h>
+
+#define NR_PAGES 32
+
+int main(int argc, char **argv)
+{
+	void *mem = malloc(NR_PAGES*PAGE_SIZE);
+	bool success = true;
+	uint64_t ms;
+	long i;
+
+	report_prefix_push("memory");
+
+	memset(mem, 0, NR_PAGES*PAGE_SIZE);
+
+	migrate_begin_continuous();
+	ms = get_clock_ms();
+	i = 0;
+	do {
+		int j;
+
+		for (j = 0; j < NR_PAGES*PAGE_SIZE; j += PAGE_SIZE) {
+			if (*(volatile long *)(mem + j) != i) {
+				success = false;
+				goto out;
+			}
+			*(volatile long *)(mem + j) = i + 1;
+		}
+		i++;
+	} while (get_clock_ms() - ms < 5000);

Maybe add a parameter so that the user can use different values for the 
runtime than always doing 5 seconds?

Sure.

Thanks,
Nick