On 10/11/2015 15:14, Stefan Hajnoczi wrote:
> From: Alberto Garcia <be...@igalia.com>
>
> This patch switches to QEMU_CLOCK_VIRTUAL for the accounting code in
> qtest mode, and makes the latency of the operation constant. This way we
> can perform tests on the accounting code with reproducible results.
>
> Signed-off-by: Alberto Garcia <be...@igalia.com>
> Message-id:
> 35ed0501450fa572684e9b5e92c361ab6cce565b.1446044838.git.be...@igalia.com
> Signed-off-by: Stefan Hajnoczi <stefa...@redhat.com>
> ---
> block/accounting.c | 14 ++++++++++++++
> 1 file changed, 14 insertions(+)
>
> diff --git a/block/accounting.c b/block/accounting.c
> index a941931..05a5c5f 100644
> --- a/block/accounting.c
> +++ b/block/accounting.c
> @@ -25,14 +25,20 @@
> #include "block/accounting.h"
> #include "block/block_int.h"
> #include "qemu/timer.h"
> +#include "sysemu/qtest.h"
>
> static QEMUClockType clock_type = QEMU_CLOCK_REALTIME;
> +static const int qtest_latency_ns = NANOSECONDS_PER_SECOND / 1000;
>
> void block_acct_init(BlockAcctStats *stats, bool account_invalid,
> bool account_failed)
> {
> stats->account_invalid = account_invalid;
> stats->account_failed = account_failed;
> +
> + if (qtest_enabled()) {
> + clock_type = QEMU_CLOCK_VIRTUAL;
> + }
> }
>
> void block_acct_cleanup(BlockAcctStats *stats)
> @@ -84,6 +90,10 @@ void block_acct_done(BlockAcctStats *stats,
> BlockAcctCookie *cookie)
> int64_t time_ns = qemu_clock_get_ns(clock_type);
> int64_t latency_ns = time_ns - cookie->start_time_ns;
>
> + if (qtest_enabled()) {
> + latency_ns = qtest_latency_ns;
> + }
> +
> assert(cookie->type < BLOCK_MAX_IOTYPE);
>
> stats->nr_bytes[cookie->type] += cookie->bytes;
> @@ -107,6 +117,10 @@ void block_acct_failed(BlockAcctStats *stats,
> BlockAcctCookie *cookie)
> int64_t time_ns = qemu_clock_get_ns(clock_type);
> int64_t latency_ns = time_ns - cookie->start_time_ns;
>
> + if (qtest_enabled()) {
> + latency_ns = qtest_latency_ns;
> + }
> +
> stats->total_time_ns[cookie->type] += latency_ns;
> stats->last_access_time_ns = time_ns;
>
>
Using QEMU_CLOCK_VIRTUAL makes sense but for the other part, could the
test instead use the null backend?
Paolo
