On Mon 2015-09-21 17:33:21, Sheng Yang wrote: > Thank you Paul! That's exactly the issue I met. I've read the whole > thread and got a general idea of the issue. > > I try to summarize it and please correct me if I'm wrong: > > 1. The issue is the result of kill_bdev() when connection has been cut > when IO is still flying. > 2. Other block devices driver didn't have this issue because they > normally keep the buffer and device, deny all the requests and only > kill the device when all request has been settled down and device has > been umounted. > 3. Why NBD cannot handle it in the same way because NBD has dependency > on userspace nfs-client, which would handle the reconnection/retry. If > DO_IT ioctl didn't return, then there is no way for userspace to > reconnect. How to coordinate between kernel and userspace on > reconnecting while leaving the device open until it's safe to close > leads to several choices and would leads to quite amount of change. > > Sounds like Goswin's suggestion really make sense( > http://sourceforge.net/p/nbd/mailman/message/31678340/ ). This issue > has been there for years and impact the stability of NBD a lot. I > think we should get it fixed.
You can get fix this one, but getting memory management right so that client & server on same machine is safe is not going to be easy. Think "nbd-server swapped out, because memory was needed by dirty buffers". Think "nbd-server can't allocate memory because it is all in the dirty buffers". Pavel -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/