Re: [PATCH 4/5] fuse: writeback_cache consistency enhancement (writeback_cache_v2)

[Bernd Schubert]

On 8/23/23 11:07, Miklos Szeredi wrote:
> On Tue, 11 Jul 2023 at 06:36, Jiachen Zhang
> <zhangjiachen.jaycee@xxxxxxxxxxxxx> wrote:
> > Some users may want both the high performance of the writeback_cahe mode
> > and a little bit more consistency among FUSE mounts. Current
> > writeback_cache mode never updates attributes from server, so can never
> > see the file attributes changed by other FUSE mounts, which means
> > 'zero-consisteny'.
> >
> > This commit introduces writeback_cache_v2 mode, which allows the attributes
> > to be updated from server to kernel when the inode is clean and no
> > writeback is in-progressing. FUSE daemons can select this mode by the
> > FUSE_WRITEBACK_CACHE_V2 init flag.
> >
> > In writeback_cache_v2 mode, the server generates official attributes.
> > Therefore,
> >
> >      1. For the cmtime, the cmtime generated by kernel are just temporary
> >      values that are never flushed to server by fuse_write_inode(), and they
> >      could be eventually updated by the official server cmtime. The
> >      mtime-based revalidation of the fc->auto_inval_data mode is also
> >      skipped, as the kernel-generated temporary cmtime are likely not equal
> >      to the offical server cmtime.
> >
> >      2. For the file size, we expect server updates its file size on
> >      FUSE_WRITEs. So we increase fi->attr_version in fuse_writepage_end() to
> >      check the staleness of the returning file size.
> >
> > Together with FOPEN_INVAL_ATTR, a FUSE daemon is able to implement
> > close-to-open (CTO) consistency like NFS client implementations.
>
> What I'd prefer is mode similar to NFS: getattr flushes pending writes
> so that server ctime/mtime are always in sync with client.  FUSE
> probably should have done that from the beginning, but at that time I
> wasn't aware of the NFS solution.


I think it would be good to have flush-on-getattr configurable - systems 
with a distributed lock manager (DLM) and notifications from 
server/daemon to kernel should not need it.


Thanks,
Bernd