Am 24.06.2013 15:49, schrieb Paolo Bonzini:
> Il 24/06/2013 10:13, Stefan Hajnoczi ha scritto:
>> On Fri, Jun 21, 2013 at 11:18:42AM +0200, Kevin Wolf wrote:
>>> Am 20.06.2013 um 20:20 hat Peter Lieven geschrieben:
>>>> @@ -800,6 +801,60 @@ iscsi_getlength(BlockDriverState *bs)
>>>> return len;
>>>> }
>>>>
>>>> +static int coroutine_fn iscsi_co_is_allocated(BlockDriverState *bs,
>>>> + int64_t sector_num,
>>>> + int nb_sectors, int *pnum)
>>>> +{
>>>> + IscsiLun *iscsilun = bs->opaque;
>>>> + struct scsi_task *task = NULL;
>>>> + struct scsi_get_lba_status *lbas = NULL;
>>>> + struct scsi_lba_status_descriptor *lbasd = NULL;
>>>> + int ret;
>>>> +
>>>> + *pnum = nb_sectors;
>>>> +
>>>> + if (iscsilun->lbpme == 0) {
>>>> + return 1;
>>>> + }
>>>> +
>>>> + /* in-flight requests could invalidate the lba status result */
>>>> + while (iscsi_process_flush(iscsilun)) {
>>>> + qemu_aio_wait();
>>>> + }
>>> Note that you're blocking here. The preferred way would be something
>>> involving a yield from the coroutine and a reenter as soon as all
>>> requests are done. Maybe a CoRwLock does what you need?
>> The other option is to avoid synchronization here and instead process
>> bs->tracked_requests so that any in-flight writes count as allocated.
> I think it's a bug if the caller doesn't take into account in-flight
> requests. For example mirroring expects writes to mark sectors as
> dirty, which will pick up everything that is_allocated fails to pick up.
>
> If all else fails, you can always add a bdrv_drain_all before the query.
>
> Hence, this check is not needed. In fact, raw-posix does not perform it.
>
> Paolo
Please ignore this patch and look at the async version in the series
which also contained the write zero optimizations.
thanks,
Peter
