Re: [PATCH RFC] introduce ioctl to completely invalidate page cache

[Dave Chinner <david@fromorbit.com>]

On Mon, Oct 06, 2014 at 04:30:19PM +0200, Jan Kara wrote:
> On Mon 06-10-14 11:33:23, Thanos Makatos wrote:
> > > > Trond also had a comment that if we extended the ioctl to work for all
> > > > inodes (not just blkdev) and allowed some additional flags of what
> > > > needs to be invalidated, the new ioctl would be also useful to NFS
> > > > userspace - see Trond's email at
> > > >
> > > > http://www.spinics.net/lists/linux-fsdevel/msg78917.html
> > > >
> > > > and the following thread. I would prefer to cover that usecase when we
> > > > are introducing new invalidation ioctl. Have you considered that Thanos?
> > > 
> > > Sure, though I don't really know how to do it. I'll start by looking at the code
> > > flow when someone does " echo 3 > /proc/sys/vm/drop_caches", unless you
> > > already have a rough idea how to do that.
> > 
> > I realise I haven't clearly understood what the semantics of this new ioctl
> > should be.
> > 
> > My initial goal was to implement an ioctl that would _completely_ invalidate
> > the buffer cache of a block device when there is no file-system involved.
> > Unless I'm mistaken the patch I posted achieves this goal.
>   Yes.
> 
> > We now want to extend this patch to take care of cached metadata, which seems
> > to be of particular importance for NFS, and I suspect that this piece of
> > functionality will still be applicable to any kind of file-system, correct?
>   So most notably they want the ioctl to work not only for block devices
> but also for any regular file. That's easily doable - you just call
> filemap_write_and_wait() and invalidate_inode_pages2() in the ioctl handler
> for regular files.
> 
> Also they wanted to be able to specify a range of a mapping to invalidate -
> that's easily doable as well. Finally they wanted a 'flags' argument so you
> can additionally ask fs to invalidate also some metadata. How invalidation
> is done will be a fs specific thing and for now I guess we don't need to go
> into details. NFS guys can sort that out when they decide to implement it.
> So in the beginning we can just have u64 flags argument and in
> it a single 'INVAL_DATA' flag meaning that invalidation of data in a given
> range is requested. Later NFS guys can add further flags.

Why do we need a new ioctl to do this? fadvise64() seems like it's
the exact fit for "FADV_INVALIDATE_[META]DATA" flags...

And before anyone shouts "posix_fadvise sucks!" note that I'm
talking about adding flags to the syscall that the kernel defines,
not the glibc posix wrapper....

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com