On 06/05/2012 07:07 PM, Christoph Hellwig wrote: > I didn't manage to finish my implementation in time. I'd suggest to > put your change 3 in for now, but I plan to get back to this. Okay, I think we can spin it for next version, current recovery works quite well now from the observation of our tests, so we'd better keep it for the release scheduled at the end of the June. There seems to be a big problem in forward_read/write_obj_req() which get a random ret value because, possibly, high concurrent requests to the same node are interleaved. Also block/sheepdog.c of QEMU have a fatal racy problem, which lead requests to be discarded by QEMU or segfault, in a high rate of requests bursting. Thanks, Yuan |