On Thu, 24 Nov 2022 14:44:29 +0100
Nico Boehr <nrb@xxxxxxxxxxxxx> wrote:
> Add a test which modifies CMM page states while migration is in
> progress.
>
> Signed-off-by: Nico Boehr <nrb@xxxxxxxxxxxxx>
> ---
> s390x/Makefile | 1 +
> s390x/migration-during-cmm.c | 121 +++++++++++++++++++++++++++++++++++
> s390x/unittests.cfg | 5 ++
> 3 files changed, 127 insertions(+)
> create mode 100644 s390x/migration-during-cmm.c
>
> diff --git a/s390x/Makefile b/s390x/Makefile
> index 401cb6371cee..64c7c04409ae 100644
> --- a/s390x/Makefile
> +++ b/s390x/Makefile
> @@ -39,6 +39,7 @@ tests += $(TEST_DIR)/panic-loop-extint.elf
> tests += $(TEST_DIR)/panic-loop-pgm.elf
> tests += $(TEST_DIR)/migration-sck.elf
> tests += $(TEST_DIR)/exittime.elf
> +tests += $(TEST_DIR)/migration-during-cmm.elf
>
> pv-tests += $(TEST_DIR)/pv-diags.elf
>
> diff --git a/s390x/migration-during-cmm.c b/s390x/migration-during-cmm.c
> new file mode 100644
> index 000000000000..afe1f73605ba
> --- /dev/null
> +++ b/s390x/migration-during-cmm.c
> @@ -0,0 +1,121 @@
> +/* SPDX-License-Identifier: GPL-2.0-only */
> +/*
> + * Perform CMMA actions while migrating.
> + *
> + * Copyright IBM Corp. 2022
> + *
> + * Authors:
> + * Nico Boehr <nrb@xxxxxxxxxxxxx>
> + */
> +
> +#include <libcflat.h>
> +#include <smp.h>
> +#include <asm-generic/barrier.h>
> +
> +#include "cmm.h"
> +
> +#define NUM_PAGES 128
> +
> +/*
> + * Allocate 3 pages more than we need so we can start at different offsets. This ensures page states
> + * change on every loop iteration.
> + */
> +static uint8_t pagebuf[(NUM_PAGES + 3) * PAGE_SIZE] __attribute__((aligned(PAGE_SIZE)));
> +
> +static unsigned int thread_iters;
> +static int thread_should_exit;
> +static int thread_exited;
> +struct cmm_verify_result result;
> +
> +static void test_cmm_during_migration(void)
> +{
> + uint8_t *pagebuf_start;
> + /*
> + * The second CPU must not print on the console, otherwise it will race with
> + * the primary CPU on the SCLP buffer.
> + */
> + while (!thread_should_exit) {
I don't see any mb()s here, maybe use READ_ONCE
> + /*
> + * Start on a offset different from the last iteration so page states change with
> + * every iteration. This is why pagebuf has 3 extra pages.
> + */
> + pagebuf_start = pagebuf + (thread_iters % 4) * PAGE_SIZE;
> + cmm_set_page_states(pagebuf_start, NUM_PAGES);
> +
> + /*
> + * Always increment even if the verify fails. This ensures primary CPU knows where
> + * we left off and can do an additional verify round after migration finished.
> + */
> + thread_iters++;
> +
> + result = cmm_verify_page_states(pagebuf_start, NUM_PAGES);
> + if (result.verify_failed)
> + goto out;
> + }
> +
> +out:
> + WRITE_ONCE(thread_exited, 1);
> +}
> +
> +int main(void)
> +{
> + bool has_essa = check_essa_available();
> + struct psw psw;
> +
> + report_prefix_push("migration-during-cmm");
> + if (!has_essa) {
> + report_skip("ESSA is not available");
> + goto error;
> + }
> +
> + if (smp_query_num_cpus() == 1) {
> + report_skip("need at least 2 cpus for this test");
> + goto error;
> + }
> +
> + psw.mask = extract_psw_mask();
> + psw.addr = (unsigned long)test_cmm_during_migration;
> + smp_cpu_setup(1, psw);
> +
> + puts("Please migrate me, then press return\n");
> + (void)getchar();
> +
> + WRITE_ONCE(thread_should_exit, 1);
> +
> + while (!thread_exited)
> + mb();
> +
> + report_info("thread completed %u iterations", thread_iters);
> +
> + report_prefix_push("during migration");
> + cmm_report_verify(&result);
> + report_prefix_pop();
> +
> + /*
> + * Verification of page states occurs on the thread. We don't know if we
> + * were still migrating during the verification.
> + * To be sure, make another verification round after the migration
> + * finished to catch page states which might not have been migrated
> + * correctly.
> + */
> + report_prefix_push("after migration");
> + assert(thread_iters > 0);
> + result = cmm_verify_page_states(pagebuf + ((thread_iters - 1) % 4) * PAGE_SIZE, NUM_PAGES);
> + cmm_report_verify(&result);
> + report_prefix_pop();
> +
> + goto done;
> +
> +error:
> + /*
> + * If we just exit and don't ask migrate_cmd to migrate us, it
> + * will just hang forever. Hence, also ask for migration when we
> + * skip this test alltogether.
> + */
> + puts("Please migrate me, then press return\n");
> + (void)getchar();
> +
I wonder if this can be made easier to review and less error prone...
maybe something like:
static void migrate_once(void)
{
static bool migrated = false;
if (migrated)
return;
puts("Please migrate me, then press return\n");
migrated = true;
(void)getchar();
}
then you can put it where needed in the code _and_ unconditionally at the end
> +done:
> + report_prefix_pop();
> + return report_summary();
> +}
> diff --git a/s390x/unittests.cfg b/s390x/unittests.cfg
> index 3caf81eda396..f6889bd4da01 100644
> --- a/s390x/unittests.cfg
> +++ b/s390x/unittests.cfg
> @@ -208,3 +208,8 @@ groups = migration
> [exittime]
> file = exittime.elf
> smp = 2
> +
> +[migration-during-cmm]
> +file = migration-during-cmm.elf
> +groups = migration
> +smp = 2
