Date:   Tue, 26 Sep 2017 12:19:21 -0700
From:   Dan Williams <dan.j.williams@...el.com>
To:     Ross Zwisler <ross.zwisler@...ux.intel.com>,
        Dan Williams <dan.j.williams@...el.com>,
        Andrew Morton <akpm@...ux-foundation.org>,
        "linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>,
        "Darrick J. Wong" <darrick.wong@...cle.com>,
        "J. Bruce Fields" <bfields@...ldses.org>,
        Christoph Hellwig <hch@....de>,
        Dave Chinner <david@...morbit.com>, Jan Kara <jack@...e.cz>,
        Jeff Layton <jlayton@...chiereds.net>,
        linux-fsdevel <linux-fsdevel@...r.kernel.org>,
        Linux MM <linux-mm@...ck.org>,
        "linux-nvdimm@...ts.01.org" <linux-nvdimm@...ts.01.org>,
        linux-xfs@...r.kernel.org
Subject: Re: [PATCH 6/7] mm, fs: introduce file_operations->post_mmap()

On Tue, Sep 26, 2017 at 11:57 AM, Ross Zwisler
<ross.zwisler@...ux.intel.com> wrote:
> On Mon, Sep 25, 2017 at 04:38:45PM -0700, Dan Williams wrote:
>> On Mon, Sep 25, 2017 at 4:14 PM, Ross Zwisler
>> <ross.zwisler@...ux.intel.com> wrote:
>> > When mappings are created the vma->vm_flags that they use vary based on
>> > whether the inode being mapped is using DAX or not.  This setup happens in
>> > XFS via mmap_region()=>call_mmap()=>xfs_file_mmap().
>> >
>> > For us to be able to safely use the DAX per-inode flag we need to prevent
>> > S_DAX transitions when any mappings are present, and we will do that by
>> > looking at the address_space->i_mmap tree and returning -EBUSY if any
>> > mappings are present.
>> >
>> > Unfortunately at the time that the filesystem's file_operations->mmap()
>> > entry point is called the mapping has not yet been added to the
>> > address_space->i_mmap tree.  This means that at that point in time we
>> > cannot determine whether or not the mapping will be set up to support DAX.
>> >
>> > Fix this by adding a new file_operations entry called post_mmap() which is
>> > called after the mapping has been added to the address_space->i_mmap tree.
>> > This post_mmap() op now happens at a time when we can be sure whether the
>> > mapping will use DAX or not, and we can set up the vma->vm_flags
>> > appropriately.
>> >
>> > Signed-off-by: Ross Zwisler <ross.zwisler@...ux.intel.com>
>> > ---
>> >  fs/xfs/xfs_file.c  | 15 ++++++++++++++-
>> >  include/linux/fs.h |  1 +
>> >  mm/mmap.c          |  2 ++
>> >  3 files changed, 17 insertions(+), 1 deletion(-)
>> >
>> > diff --git a/fs/xfs/xfs_file.c b/fs/xfs/xfs_file.c
>> > index 2816858..9d66aaa 100644
>> > --- a/fs/xfs/xfs_file.c
>> > +++ b/fs/xfs/xfs_file.c
>> > @@ -1087,9 +1087,21 @@ xfs_file_mmap(
>> >  {
>> >         file_accessed(filp);
>> >         vma->vm_ops = &xfs_file_vm_ops;
>> > +       return 0;
>> > +}
>> > +
>> > +/* This call happens during mmap(), after the vma has been inserted into the
>> > + * inode->i_mapping->i_mmap tree.  At this point the decision on whether or
>> > + * not to use DAX for this mapping has been set and will not change for the
>> > + * duration of the mapping.
>> > + */
>> > +STATIC void
>> > +xfs_file_post_mmap(
>> > +       struct file     *filp,
>> > +       struct vm_area_struct *vma)
>> > +{
>> >         if (IS_DAX(file_inode(filp)))
>> >                 vma->vm_flags |= VM_MIXEDMAP | VM_HUGEPAGE;
>>
>> It's not clear to me what this is actually protecting? vma_is_dax()
>> returns true regardless of the vm_flags state , so what is the benefit
>> to delaying the vm_flags setting to ->post_mmap()?
>
> Right, but the point is that until the vma has been inserted into the
> inode->i_mapping->i_mmap tree, the results of IS_DAX() don't matter because it
> can still change.  Until this insertion happens we cannot know whether or not
> we should set up the vma->vm_flags to support DAX mappings (i.e. have
> VM_MIXEDMAP and VM_HUGEPAGE set).

Those flags are not DAX flags. The side effect of these being set on
non-DAX mappings is that we effectively auto madvise(MADV_HUGEPAGE)
and enable some page-less insertion paths. Both of those are
effectively no-ops for normal mappings since normal mappings always
have an associated struct page and the THP policy these days is
usually "always".

> This decision can only be made (in this
> proposed scheme) *after* the inode->i_mapping->i_mmap  tree has been
> populated, which means we need another call into the filesystem after this
> insertion has happened.

I get that, but it seems over-engineered and something that can also
be safely cleaned up after the fact by the code path that is disabling
DAX.

> We don't want to mess with the existing file_operations->mmap() call because
> in many filesystems that does sanity checking and setup that you really want
> to have happen *before* the mapping is completed and inserted into the
> inode->i_mapping->i_mmap tree.
>
>> Also, why is this a file_operation and not a vm_operation?
>
> Because ->mmap() is also a file_operation, and this is an analogous call from
> the mmap code that needs to happen at a different time.  Or are you suggesting
> that file_operations->mmap() should be moved to be a vm_operation?  If not,
> why would one be in one operations table and one in another?

Growing something as widely used as file_operations for this one-off
fixup feels like overkill. vm_operations is not much better, but it at
least constrains the data structure growth to something closer to the
problem space.

Powered by blists - more mailing lists