[1/4] migration: keep bytes_xfer_prev init'd to zero

The first time migration_bitmap_sync() is called, bytes_xfer_prev is set
to ram_state.bytes_transferred which is, at this point, zero. The next
time migration_bitmap_sync() is called, an iteration has happened and
bytes_xfer_prev is set to 'x' bytes. Most likely, more than one second
has passed, so the auto converge logic will be triggered and
bytes_xfer_now will also be set to 'x' bytes.

This condition is currently masked by dirty_rate_high_cnt, which will
wait for a few iterations before throttling. It would otherwise always
assume zero bytes have been copied and therefore throttle the guest
(possibly) prematurely.

Given bytes_xfer_prev is only used by the auto convergence logic, it
makes sense to only set its value after a check has been made against
bytes_xfer_now.

Signed-off-by: Felipe Franciosi <felipe@nutanix.com>
---
 migration/ram.c | 4 ----
 1 file changed, 4 deletions(-)

On Wed, May 24, 2017 at 05:10:00PM +0100, Felipe Franciosi wrote:
> The first time migration_bitmap_sync() is called, bytes_xfer_prev is set
> to ram_state.bytes_transferred which is, at this point, zero. The next
> time migration_bitmap_sync() is called, an iteration has happened and
> bytes_xfer_prev is set to 'x' bytes. Most likely, more than one second
> has passed, so the auto converge logic will be triggered and
> bytes_xfer_now will also be set to 'x' bytes.
> 
> This condition is currently masked by dirty_rate_high_cnt, which will
> wait for a few iterations before throttling. It would otherwise always
> assume zero bytes have been copied and therefore throttle the guest
> (possibly) prematurely.
> 
> Given bytes_xfer_prev is only used by the auto convergence logic, it
> makes sense to only set its value after a check has been made against
> bytes_xfer_now.
> 
> Signed-off-by: Felipe Franciosi <felipe@nutanix.com>

Yes, with patch 3, I think it should be accurate.

Reviewed-by: Peter Xu <peterx@redhat.com>

> ---
>  migration/ram.c | 4 ----
>  1 file changed, 4 deletions(-)
> 
> diff --git a/migration/ram.c b/migration/ram.c
> index c07a9c0..36bf720 100644
> --- a/migration/ram.c
> +++ b/migration/ram.c
> @@ -673,10 +673,6 @@ static void migration_bitmap_sync(RAMState *rs)
>  
>      rs->bitmap_sync_count++;
>  
> -    if (!rs->bytes_xfer_prev) {
> -        rs->bytes_xfer_prev = ram_bytes_transferred();
> -    }
> -
>      if (!rs->time_last_bitmap_sync) {
>          rs->time_last_bitmap_sync = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
>      }
> -- 
> 1.9.5
>

Felipe Franciosi <felipe@nutanix.com> wrote:
> The first time migration_bitmap_sync() is called, bytes_xfer_prev is set
> to ram_state.bytes_transferred which is, at this point, zero. The next
> time migration_bitmap_sync() is called, an iteration has happened and
> bytes_xfer_prev is set to 'x' bytes. Most likely, more than one second
> has passed, so the auto converge logic will be triggered and
> bytes_xfer_now will also be set to 'x' bytes.
>
> This condition is currently masked by dirty_rate_high_cnt, which will
> wait for a few iterations before throttling. It would otherwise always
> assume zero bytes have been copied and therefore throttle the guest
> (possibly) prematurely.
>
> Given bytes_xfer_prev is only used by the auto convergence logic, it
> makes sense to only set its value after a check has been made against
> bytes_xfer_now.
>
> Signed-off-by: Felipe Franciosi <felipe@nutanix.com>

Reviewed-by: Juan Quintela <quintela@redhat.com>