On Fri, Jan 10, 2020 at 11:29:39AM -0800, [email protected] wrote:
> From: Ira Weiny <[email protected]>
>
> Page faults need to ensure the inode mode is correct and consistent with
> the vmf information at the time of the fault. There is no easy way to
> ensure the vmf information is correct if a mode change is in progress.
> Furthermore, there is no good use case to require a mode change while
> the file is mmap'ed.
>
> Track mmap's of the file and fail the mode change if the file is
> mmap'ed.
>
> Signed-off-by: Ira Weiny <[email protected]>
> ---
> fs/inode.c | 2 ++
> fs/xfs/xfs_ioctl.c | 8 ++++++++
> include/linux/fs.h | 1 +
> mm/mmap.c | 19 +++++++++++++++++--
> 4 files changed, 28 insertions(+), 2 deletions(-)
>
> diff --git a/fs/inode.c b/fs/inode.c
> index 2b0f51161918..944711aed6f8 100644
> --- a/fs/inode.c
> +++ b/fs/inode.c
> @@ -245,6 +245,8 @@ static struct inode *alloc_inode(struct super_block *sb)
> return NULL;
> }
>
> + atomic64_set(&inode->i_mapped, 0);
> +
> return inode;
> }
>
> diff --git a/fs/xfs/xfs_ioctl.c b/fs/xfs/xfs_ioctl.c
> index bc3654fe3b5d..1ab0906c6c7f 100644
> --- a/fs/xfs/xfs_ioctl.c
> +++ b/fs/xfs/xfs_ioctl.c
> @@ -1200,6 +1200,14 @@ xfs_ioctl_setattr_dax_invalidate(
> goto out_unlock;
> }
>
> + /*
> + * If there is a mapping in place we must remain in our current mode.
> + */
> + if (atomic64_read(&inode->i_mapped)) {
Urk, should we really be messing around with the address space
internals?
> + error = -EBUSY;
> + goto out_unlock;
> + }
> +
> error = filemap_write_and_wait(inode->i_mapping);
> if (error)
> goto out_unlock;
> diff --git a/include/linux/fs.h b/include/linux/fs.h
> index 631f11d6246e..6e7dc626b657 100644
> --- a/include/linux/fs.h
> +++ b/include/linux/fs.h
> @@ -740,6 +740,7 @@ struct inode {
> #endif
>
> void *i_private; /* fs or device private pointer */
> + atomic64_t i_mapped;
I would have expected to find this in struct address_space since the
mapping count is a function of the address space, right?
--D
> } __randomize_layout;
>
> struct timespec64 timestamp_truncate(struct timespec64 t, struct inode *inode);
> diff --git a/mm/mmap.c b/mm/mmap.c
> index dfaf1130e706..e6b68924b7ca 100644
> --- a/mm/mmap.c
> +++ b/mm/mmap.c
> @@ -171,12 +171,17 @@ void unlink_file_vma(struct vm_area_struct *vma)
> static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
> {
> struct vm_area_struct *next = vma->vm_next;
> + struct file *f = vma->vm_file;
>
> might_sleep();
> if (vma->vm_ops && vma->vm_ops->close)
> vma->vm_ops->close(vma);
> - if (vma->vm_file)
> - fput(vma->vm_file);
> + if (f) {
> + struct inode *inode = file_inode(f);
> + if (inode)
> + atomic64_dec(&inode->i_mapped);
> + fput(f);
> + }
> mpol_put(vma_policy(vma));
> vm_area_free(vma);
> return next;
> @@ -1837,6 +1842,16 @@ unsigned long mmap_region(struct file *file, unsigned long addr,
>
> vma_set_page_prot(vma);
>
> + /*
> + * Track if there is mapping in place such that a mode change
> + * does not occur on a file which is mapped
> + */
> + if (file) {
> + struct inode *inode = file_inode(file);
> +
> + atomic64_inc(&inode->i_mapped);
> + }
> +
> return addr;
>
> unmap_and_free_vma:
> --
> 2.21.0
>
On Mon, Jan 13, 2020 at 02:22:12PM -0800, Darrick J. Wong wrote:
> On Fri, Jan 10, 2020 at 11:29:39AM -0800, [email protected] wrote:
> > From: Ira Weiny <[email protected]>
> >
[snip]
> >
> > diff --git a/fs/xfs/xfs_ioctl.c b/fs/xfs/xfs_ioctl.c
> > index bc3654fe3b5d..1ab0906c6c7f 100644
> > --- a/fs/xfs/xfs_ioctl.c
> > +++ b/fs/xfs/xfs_ioctl.c
> > @@ -1200,6 +1200,14 @@ xfs_ioctl_setattr_dax_invalidate(
> > goto out_unlock;
> > }
> >
> > + /*
> > + * If there is a mapping in place we must remain in our current mode.
> > + */
> > + if (atomic64_read(&inode->i_mapped)) {
>
> Urk, should we really be messing around with the address space
> internals?
I contemplated a function call instead of checking i_mapped directly? Is that
what you mean?
>
> > + error = -EBUSY;
> > + goto out_unlock;
> > + }
> > +
> > error = filemap_write_and_wait(inode->i_mapping);
> > if (error)
> > goto out_unlock;
> > diff --git a/include/linux/fs.h b/include/linux/fs.h
> > index 631f11d6246e..6e7dc626b657 100644
> > --- a/include/linux/fs.h
> > +++ b/include/linux/fs.h
> > @@ -740,6 +740,7 @@ struct inode {
> > #endif
> >
> > void *i_private; /* fs or device private pointer */
> > + atomic64_t i_mapped;
>
> I would have expected to find this in struct address_space since the
> mapping count is a function of the address space, right?
I suppose but the only external call (above) would be passing an inode. So to
me it seemed better here.
Ira
>
> --D
>
On Mon, Jan 13, 2020 at 04:46:10PM -0800, Ira Weiny wrote:
> On Mon, Jan 13, 2020 at 02:22:12PM -0800, Darrick J. Wong wrote:
> > On Fri, Jan 10, 2020 at 11:29:39AM -0800, [email protected] wrote:
> > > From: Ira Weiny <[email protected]>
> > >
>
> [snip]
>
> > >
> > > diff --git a/fs/xfs/xfs_ioctl.c b/fs/xfs/xfs_ioctl.c
> > > index bc3654fe3b5d..1ab0906c6c7f 100644
> > > --- a/fs/xfs/xfs_ioctl.c
> > > +++ b/fs/xfs/xfs_ioctl.c
> > > @@ -1200,6 +1200,14 @@ xfs_ioctl_setattr_dax_invalidate(
> > > goto out_unlock;
> > > }
> > >
> > > + /*
> > > + * If there is a mapping in place we must remain in our current mode.
> > > + */
> > > + if (atomic64_read(&inode->i_mapped)) {
> >
> > Urk, should we really be messing around with the address space
> > internals?
>
> I contemplated a function call instead of checking i_mapped directly? Is that
> what you mean?
Yeah. Abstracting the details just enough that filesystems don't have
to know that i_mapped is atomic64 etc.
>
> >
> > > + error = -EBUSY;
> > > + goto out_unlock;
> > > + }
> > > +
> > > error = filemap_write_and_wait(inode->i_mapping);
> > > if (error)
> > > goto out_unlock;
> > > diff --git a/include/linux/fs.h b/include/linux/fs.h
> > > index 631f11d6246e..6e7dc626b657 100644
> > > --- a/include/linux/fs.h
> > > +++ b/include/linux/fs.h
> > > @@ -740,6 +740,7 @@ struct inode {
> > > #endif
> > >
> > > void *i_private; /* fs or device private pointer */
> > > + atomic64_t i_mapped;
> >
> > I would have expected to find this in struct address_space since the
> > mapping count is a function of the address space, right?
>
> I suppose but the only external call (above) would be passing an inode. So to
> me it seemed better here.
But the number of memory mappings reflects the state of the address
space, not the inode. Or maybe put another way, if I were an mm
developer I would not expect to look in struct inode for mm state.
static inline bool inode_has_mappings(struct inode *inode)
{
return atomic64_read(&inode->i_mapping->mapcount) > 0;
}
OTOH if there exist other mm developers who /do/ find that storing the
mmap count in struct inode is more logical, please let me know. :)
--D
> Ira
>
> >
> > --D
> >