[PATCH v4 36/39] netfs: Implement a write-through caching option

[Jeff Layton]

On Tue, 2023-12-19 at 16:51 +0000, David Howells wrote:
> Jeff Layton <jlayton at kernel.org> wrote:
> 
> > > This can't be used with content encryption as that may require expansion of
> > > the write RPC beyond the write being made.
> > > 
> > > This doesn't affect writes via mmap - those are written back in the normal
> > > way; similarly failed writethrough writes are marked dirty and left to
> > > writeback to retry.  Another option would be to simply invalidate them, but
> > > the contents can be simultaneously accessed by read() and through mmap.
> > > 
> > 
> > I do wish Linux were less of a mess in this regard. Different
> > filesystems behave differently when writeback fails.
> 
> Cifs is particularly, um, entertaining in this regard as it allows the write
> to fail on the server due to a checksum failure if the source data changes
> during the write and then just retries it later.
> 

Should they be using bounce pages here? Maybe that's more efficient in
the common case though and worth the extra hit if it happens seldom
enough.

> > That said, the modern consensus with local filesystems is to just leave
> > the pages clean when buffered writeback fails, but set a writeback error
> > on the inode. That at least keeps dirty pages from stacking up in the
> > cache. In the case of something like a netfs, we usually invalidate the
> > inode and the pages -- netfs's usually have to spontaneously deal with
> > that anyway, so we might as well.
> > 
> > Marking the pages dirty here should mean that they'll effectively get a
> > second try at writeback, which is a change in behavior from most
> > filesystems. I'm not sure it's a bad one, but writeback can take a long
> > time if you have a laggy network.
> 
> I'm not sure what the best thing to do is.  If everything is doing
> O_DSYNC/writethrough I/O on an inode and there is no mmap, then invalidating
> the pages is probably not a bad way to deal with failure here.
> 

That's a big if ;)

> > When a write has already failed once, why do you think it'll succeed on
> > a second attempt (and probably with page-aligned I/O, I guess)?
> 
> See above with cifs.  I wonder if the pages being written to should be made RO
> and page_mkwrite() forced to lock against DSYNC writethrough.
> 

That sounds pretty heavy handed, particularly if the server goes offline
for a bit. Now you're stuck in some locking call in page_mkwrite...

> > Another question: when the writeback is (re)attempted, will it end up
> > just doing page-aligned I/O, or is the byte range still going to be
> > limited to the written range?
> 
> At the moment, it then happens exactly as it would if it wasn't doing
> writethrough - so it will write partial folios if it's doing a streaming write
> and will do full folios otherwise.
> 
>
> > The more I consider it, I think it might be a lot simpler to just "fail
> > fast" here rather than remarking the write dirty.
> 
> You may be right - but, again, mmap:-/
> 

There's nothing we can do about mmap -- we're stuck page-sized I/Os
there.

With normal buffered I/O I still think just leaving the pages clean is
probably the least bad option. I think it's also sort of the Linux
"standard" behavior (for better or worse).

Willy, do you have any thoughts here?
-- 
Jeff Layton <jlayton at kernel.org>