SMB/CIFS compounding support [LWN.net]

SMB/CIFS compounding support

Posted May 24, 2018 6:27 UTC (Thu) by mjthayer (guest, #39183)
Parent article: SMB/CIFS compounding support

The original kernel SMB code was developed before FUSE was in the kernel. Does it still have an advantage over a FUSE SMB filesystem? I still can't quite get used to the idea of an SMB client running in kernel space.

to post comments

SMB/CIFS compounding support

Posted Jun 2, 2018 21:25 UTC (Sat) by smfrench (subscriber, #124116) [Link]

Performance of the kernel SMB3 code in my testing was quite good, although in some cases Volker's userspace (Samba tool) smbclient improvements can beat it. In most of my large i/o tests it was faster (Linux to Linux, on localhost) than NFSv3 or NFSv4.2.

But ... SMB3 kernel client without compounding (and handle caching and directory leases) can be much slower for large directories and metadata searches. This is very exciting work that Ronnie at Redhat is doing to introduce compounding into the kernel client (cifs.ko) for SMB3.

In addition, Long Li's work (various "smbdirect" - ie network file system i/o over RDMA) is impressive. Windows, NetApp etc. servers support RDMA and it is nice to be able to take advantage of this with very easy to use SMB3 mounts using the work Long Li has done. He has a pending patchset as well for improving SMB3 direct i/o in the kernel. Hard to imagine a userspace FUSE client able to keep up with that.