Re: [PATCH v2 04/20] qemu-file: We only call qemu_file_transferred_* on the sending side

[Peter Xu <peterx@redhat.com>]

On Tue, May 30, 2023 at 08:39:25PM +0200, Juan Quintela wrote:
> Remove the increase in qemu_file_fill_buffer() and add asserts to
> qemu_file_transferred* functions.
> 
> Signed-off-by: Juan Quintela <quintela@redhat.com>

The read side accounting does look a bit weird and never caught my notice..

Maybe worth also touching the document of QEMUFile::total_transferred to
clarify what it accounts?

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

Though when I'm looking at the counters (didn't follow every single recent
patch on this..), I found that now reading transferred value is actually
more expensive - qemu_file_transferred() needs flushing, even if for the
fast version, qemu_file_transferred_fast() loops over all possible iovs,
which can be as large as MAX_IOV_SIZE==64.

To be explicit, I _think_ for each guest page we now need to flush...

  ram_save_iterate
    migration_rate_exceeded
      migration_transferred_bytes
        qemu_file_transferred

I hope I'm wrong..

Does it mean that perhaps we simply need "sent and put into send buffer"
more than "what really got transferred"?  So I start to wonder what's the
origianl purpose of this change, and which one is better..

-- 
Peter Xu