On 04/25/2012 03:03 PM, Christoph Hellwig wrote: > Instead of obscuring the callbacks into sheepdog from the cluster drivers > by various means of callbacks just call them directly like you would do in > normal C code. The block_cb callback to the notify routine is left for now > as that area needs broader attention later. I basically agree to remove callbacks from upper layer(sheep) for simplicity, thought I guess originally those callbacks were designed to decouple sheep and cluster driver completely. I am more interested in how do you plan to deal with block_cb()? We already meet some subtle problem that cluster gets hung at block state for ever running a 1000 sheep daemon on dozen of machines, but not yet come to any conclusion useful. We can only say that the block mechanism would leave some holes to hang the whole cluster by only several minor failed nodes (be it whether EIO-exiting or down). Thanks, Yuan |