Re: [PATCH v2 5/9] block/write-threshold: don't use aio context lock

[Max Reitz <mreitz@redhat.com>]

On 04.05.21 10:25, Vladimir Sementsov-Ogievskiy wrote:
> Instead of relying on aio context lock, let's make use of atomic
> operations.
> 
> The tricky place is bdrv_write_threshold_check_write(): we want
> atomically unset bs->write_threshold_offset iff
>    offset + bytes > bs->write_threshold_offset
> We don't have such atomic operation, so let's go in a loop:

This could also be solved by untangling the overloaded meaning of 
write_threshold_offset – if we had an additional boolean 
write_threshold_check, then this wouldn’t be a problem, and we could do 
this:

if (end > bdrv_write_threshold_get(bs)) {
     if (qatomic_xchg(&bs->write_threshold_check, false) == true) {
         qapi_event_send(...);
     }
}

However, the problem then becomes thinking about the memory access 
semantics, because if some other thread sets the threshold offset and 
enables the threshold check, we need to ensure that we see the updated 
offset here...

So I suppose your loop is simpler then.

> 1. fetch wtr atomically
> 2. if condition satisfied, try cmpxchg (if not satisfied, we are done,
>     don't send event)
> 3. if cmpxchg succeeded, we are done (send event), else go to [1]
> 
> Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
> ---
>   block/write-threshold.c | 32 +++++++++++++++++---------------
>   1 file changed, 17 insertions(+), 15 deletions(-)

Reviewed-by: Max Reitz <mreitz@redhat.com>