Re: [Qemu-devel] [PATCH v2 01/11] block: Add op blocker type "device IO"

[Fam Zheng <famz@redhat.com>]

On Wed, 05/13 14:04, Paolo Bonzini wrote:
> 
> 
> On 13/05/2015 19:28, Fam Zheng wrote:
> > @@ -478,6 +478,10 @@ static int blk_check_request(BlockBackend *blk, int64_t sector_num,
> >          return -EIO;
> >      }
> >  
> > +    if (bdrv_op_is_blocked(blk->bs, BLOCK_OP_TYPE_DEVICE_IO, NULL)) {
> > +        return -EBUSY;
> > +    }
> 
> I think this is incorrect.  It's fine for backends to generate more I/O
> after a blocker is submitted, as long as it's bounded.
> 
> For example, SCSI requests can result in many consecutive I/Os:
> 
> (1) FUA requests are split in write+flush
> 
> (2) adapters that do not use QEMUSGList-based I/O only read 128K at a time
> 
> (3) WRITE SAME operations are also split in chunks
> 
> (4) UNMAP operations process one descriptor at a time

I don't understand the point of these examples. If we don't return -EBUSY here,
the request will sneak into block/io.c and perhaps break qmp transaction
semantics, if it lands between two backups.

Fam

> 
> Paolo
> 
> >      return blk_check_byte_request(blk, sector_num * BDRV_SECTOR_SIZE,
> >                                    nb_sectors * BDRV_SECTOR_SIZE);
> >  }