On 13/03/2015 02:38, Fam Zheng wrote:
> The caller's workflow is like
>
> if (!address_space_map()) {
> ...
> cpu_register_map_client();
> }
>
> If bounce buffer became available after address_space_map() but before
> cpu_register_map_client(), the caller could miss it and has to wait for the
> next bounce buffer uesr to release, which may never happen in the worse case.
>
> Just notify the caller with the passed callback in cpu_register_map_client().
>
> Signed-off-by: Fam Zheng <f...@redhat.com>
> ---
> exec.c | 3 +++
> 1 file changed, 3 insertions(+)
>
> diff --git a/exec.c b/exec.c
> index 93ccd5a..82781e4 100644
> --- a/exec.c
> +++ b/exec.c
> @@ -2502,6 +2502,9 @@ void *cpu_register_map_client(void *opaque, void
> (*callback)(void *opaque))
> client->opaque = opaque;
> client->callback = callback;
> QSLIST_INSERT_HEAD_ATOMIC(&map_client_list, client, link);
> + if (!atomic_read(&bounce.in_use)) {
> + callback(opaque);
This should call cpu_notify_map_clients instead, otherwise you get two
calls.
Paolo
> + }
> return client;
> }
>
>
