On 9/14/20 11:47 AM, Gregor Haas wrote:
Hi Eric, Thank you for the quick response. I'm still pretty new to playing with filesystems so thank you also for the information about clustered filesystems. I'll be looking into these for my use case. So essentially, any synchronization for parallel clients would have to happen at the filesystem layer -- NBD has no mechanisms in place for that, and simply exports a raw block device. Is my understanding of that correct?
Correct - NBD is _just_ block storage, and has no synchronization primitives beyond NBD_CMD_FLUSH (which flushes all prior writes) or NBD_CMD_FLAG_FUA (which flushes the current write), where even those primitives are optional. (There may be other forked implementations of the NBD protocol that have added further primitives as additional extensions, but no one has proposed adding such extensions into the upstream NBD protocol documentation yet for interoperability between implementations).
How you use the block storage (whether as a building block to larger storage arrays, as swap space, as a file system, etc) is up to higher layers of software.
-- Eric Blake, Principal Software Engineer Red Hat, Inc. +1-919-301-3226 Virtualization: qemu.org | libvirt.org