Re: [Devel] [PATCH VZ10 v9 8/8] selftests/ve: Add mount accounting selftest

[Konstantin Khorenko <khorenko@virtuozzo.com>]

On 9/15/26 19:22, Vladimir Riabchun wrote:
> There are 6 test cases, covered in the new test:
> 1. Simple mount accouting correctness, just mount/umount.
> 2. Verification of correct limit hits and changes, including
>    negative values.
> 3. Partial mounts test, when mount limit is hit in the middle
>    of creation.
> 4. Test that enabled pseudosuper allows overuse.
> 5. Test that pseudosuper doesn't affect mount accounting.
> 6. Failcount feature verification.
> 
> https://virtuozzo.atlassian.net/browse/VSTOR-135520
> 
> Feature: per-ve failcounters
> Signed-off-by: Vladimir Riabchun <vladimir.riabchun@virtuozzo.com>
> ---
>  tools/testing/selftests/ve/.gitignore         |   1 +
>  tools/testing/selftests/ve/Makefile           |   1 +
>  .../selftests/ve/ve_mount_accounting_test.c   | 550 ++++++++++++++++++
>  tools/testing/selftests/ve/ve_selftest.h      |  21 +
>  4 files changed, 573 insertions(+)
>  create mode 100644 tools/testing/selftests/ve/ve_mount_accounting_test.c
> 

<skipped>

> + * We can't enforce workqueue flush from userspace, so instead
> + * read several times until the value is the one we expect.
> + */
> +static int wait_on_mnt_avail(FIXTURE_DATA(ve_mnt_acc) * self, int expected)
> +{
> +#define WAIT_TIMEOUT_NS		(4096UL * 1000 * 1000)
> +	char path_avail[64];
> +	int mnt_avail_nr = 0;
> +	struct timespec delay = { .tv_nsec = 1000000 };
> +
> +	snprintf(path_avail, sizeof(path_avail), "%d/ve.mnt_avail_nr", self->ctid);
> +
> +	while (delay.tv_nsec <= WAIT_TIMEOUT_NS) {
> +		if (read_s32_at(self->cgv2_fd, path_avail, &mnt_avail_nr)) {
> +			fprintf(stderr, "Failed to read from %s\n", path_avail);
> +			return 1;
> +		}
> +		if (mnt_avail_nr == expected)
> +			return 0;
> +
> +		do_rcu_barrier();
> +		if (nanosleep(&delay, NULL)) {
> +			fprintf(stderr, "nanosleep failed: %s\n", strerror(errno));
> +			return 1;
> +		}
> +		delay.tv_nsec *= 2;

  nanosleep() requires tv_nsec < 1e9 and fails with EINVAL otherwise.
  The doubling goes 1ms, 2ms, ..., 512ms, and the next step is 1024ms,
  which no longer fits tv_nsec. So the loop never gets anywhere near
  WAIT_TIMEOUT_NS: after ~1s of total sleeping the next iteration bails
  out with "nanosleep failed: Invalid argument" and returns 1, instead
  of the intended ~8s of waiting and the timeout report with the
  expected/read values.

  This defeats the purpose of the helper: on a loaded machine, where
  the counter takes a while to settle, the test fails early with a
  misleading error - exactly the case this function was reworked for
  in v9.

  Please keep the delay in a plain nanoseconds variable and normalize
  it before sleeping:

        delay.tv_sec = ns / 1000000000;
        delay.tv_nsec = ns % 1000000000;

  As a side note, tv_nsec is long, so 2048ms would overflow it on a
  32-bit build; the normalization fixes that too.

<skipped>
_______________________________________________
Devel mailing list
Devel@openvz.org
https://lists.openvz.org/mailman/listinfo/devel