On 09/02/2018 18:23, Kevin Wolf wrote:
> Am 09.02.2018 um 17:28 hat Paolo Bonzini geschrieben:
>> On 08/02/2018 18:18, Stefan Hajnoczi wrote:
>>> +    BlockDriverState *bs = blk_bs(blk);
>>> +
>>> +    if (bs) {
>>> +        bdrv_drained_begin(bs);
>>> +    }
>>> +
>>> +    /* We may have aio requests like -ENOMEDIUM in flight */
>>> +    while (atomic_mb_read(&blk->in_flight) > 0) {
>>> +        aio_poll(blk_get_aio_context(blk), true);
>>> +    }
>>> +
>>> +    if (bs) {
>>> +        bdrv_drained_end(bs);
>>
>> This only works if you are in the default AioContext, otherwise you can
>> run handlers for one AioContexts in two threads.
> 
> Should aio_poll() assert that it's called in the right thread to make
> sure we obey the rules?
> 
>> bdrv_dec_in_flight uses bdrv_wakeup and BDRV_POLL_WHILE to ensure that
>> this doesn't happen, so there would be more code that you have to copy
>> into block-backend.c.
> 
> Instead of copying, can't we generalise it into a POLL_WHILE(ctx,
> wakeup, cond) and make BDRV_POLL_WHILE() a wrapper for that?

Yes, or even move bdrv_wakeup to AioContext would do.  We already have
block layer-specific fields such as the linux-aio state(*).

        (*) though now that linux-aio has been improved to do something
            like epoll, there may be a better reason to place linux-aio
            state in AioContext.

Paolo

Reply via email to