Re: [Qemu-block] [PATCH v2 11/17] block-backend: Decrease in_flight only after callback

[Paolo Bonzini]

On 13/09/2018 18:59, Kevin Wolf wrote:
> Am 13.09.2018 um 17:10 hat Paolo Bonzini geschrieben:
>> On 13/09/2018 14:52, Kevin Wolf wrote:
>>> + if (qemu_get_current_aio_context() == qemu_get_aio_context()) {
>>> + /* If we are in the main thread, the callback is allowed to unref
>>> + * the BlockBackend, so we have to hold an additional reference */
>>> + blk_ref(acb->rwco.blk);
>>> + }
>>> acb->common.cb(acb->common.opaque, acb->rwco.ret);
>>> + blk_dec_in_flight(acb->rwco.blk);
>>> + if (qemu_get_current_aio_context() == qemu_get_aio_context()) {
>>> + blk_unref(acb->rwco.blk);
>>> + }
>>
>> Is this something that happens only for some specific callers?  That is,
>> which callers are sure that the callback is invoked from the main thread?
> 
> I can't seem to reproduce the problem I saw any more even when reverting
> the bdrv_ref/unref pair. If I remember correctly it was actually a
> nested aio_poll() that was running a block job completion or something
> like that - which would obviously only happen on the main thread because
> the job intentionally defers to the main thread.
> 
> The only reason I made this conditional is that I think bdrv_unref()
> still isn't safe outside the main thread, is it?

Yes, making it conditional is correct, but it is quite fishy even with
the conditional.

As you mention, you could have a nested aio_poll() in the main thread,
for example invoked from a bottom half, but in that case I'd rather
track the caller that is creating the bottom half and see if it lacks a
bdrv_ref/bdrv_unref (or perhaps it's even higher in the tree that is
missing).

Paolo