read-after-write and read-after-read consistency

[dgiantsidi]

Hello,

I am wondering if ZFS serializes concurrent reads and writes from different threads such that a read is guaranteed to see the latest write even between multithreaded executions. I have the intuition that ZFS takes range locks on the files it touches based on the znode* (zfs_rangelock_enter_impl) and there is no per-thread caching. As such, concurrent operations on the same file are somehow serialized in a total order across threads.

Is my understanding correct?

• I also understand that the znode is the ZFS equivalent to an inode. So basically all threads/processes are accessing the ZFS files through a znode (even though they do not share the same file descriptor)

[amotin]

You are right. ZFS uses range locks to serialize conflicting VFS operations.