[PATCH -mm] splice: Disable vmsplice on nommu.

[PATCH -mm] splice: Disable vmsplice on nommu.

[Paul Mundt]

The vmsplice rework in -mm breaks nommu:

fs/built-in.o: In function `splice_setup_vma':
fs/splice.c:1272: undefined reference to `may_expand_vm'
fs/splice.c:1272: undefined reference to `protection_map'
fs/splice.c:1272: undefined reference to `insert_vm_struct'
fs/splice.c:1272: undefined reference to `vm_stat_account'

This is going to be non-trivial to wire up properly on nommu, and
it's debateable whether there's even any point in trying. For now,
disable vmsplice if CONFIG_MMU=n and wire it up as a cond syscall
for the platforms that support both.

Signed-off-by: Paul Mundt <lethal@linux-sh.org>

--

 fs/splice.c     |    2 ++
 kernel/sys_ni.c |    1 +
 2 files changed, 3 insertions(+)

--- linux-2.6.23-rc4-mm1.orig/fs/splice.c	2007-09-11 15:15:54.000000000 +0900
+++ linux-2.6.23-rc4-mm1/fs/splice.c	2007-09-11 17:11:02.000000000 +0900
@@ -1180,6 +1180,7 @@
 	return -EINVAL;
 }
 
+#ifdef CONFIG_MMU
 /*
  * Undo vmsplice_to_user map. Basically just lookup the vmas and
  * detach/unmap/remove them.
@@ -1686,6 +1687,7 @@
 
 	return error;
 }
+#endif /* CONFIG_MMU */
 
 asmlinkage long sys_splice(int fd_in, loff_t __user *off_in,
 			   int fd_out, loff_t __user *off_out,
--- linux-2.6.23-rc4-mm1.orig/kernel/sys_ni.c	2007-09-11 15:15:56.000000000 +0900
+++ linux-2.6.23-rc4-mm1/kernel/sys_ni.c	2007-09-11 17:09:24.000000000 +0900
@@ -139,6 +139,7 @@
 cond_syscall(sys_madvise);
 cond_syscall(sys_mremap);
 cond_syscall(sys_remap_file_pages);
+cond_syscall(sys_vmsplice);
 cond_syscall(compat_sys_move_pages);
 cond_syscall(compat_sys_migrate_pages);
 

Re: [PATCH -mm] splice: Disable vmsplice on nommu.

[Jens Axboe]

On Tue, Sep 11 2007, Paul Mundt wrote:
> The vmsplice rework in -mm breaks nommu:
> 
> fs/built-in.o: In function `splice_setup_vma':
> fs/splice.c:1272: undefined reference to `may_expand_vm'
> fs/splice.c:1272: undefined reference to `protection_map'
> fs/splice.c:1272: undefined reference to `insert_vm_struct'
> fs/splice.c:1272: undefined reference to `vm_stat_account'
> 
> This is going to be non-trivial to wire up properly on nommu, and
> it's debateable whether there's even any point in trying. For now,
> disable vmsplice if CONFIG_MMU=n and wire it up as a cond syscall
> for the platforms that support both.

It's a bit of a sledge hammer... Since get_user_pages() appears to work
on nummu, we can just disable the vmsplice-to-userspace bits and leave
the (working) vmsplice-from-userspace in place.

I'm reworking this code anyway in the block branches, so I'll update and
take CONFIG_MMU into consideration.

-- 
Jens Axboe

Re: [PATCH -mm] splice: Disable vmsplice on nommu.

[David Howells]

Jens Axboe <jens.axboe@oracle.com> wrote:

> It's a bit of a sledge hammer... Since get_user_pages() appears to work
> on nummu, we can just disable the vmsplice-to-userspace bits and leave
> the (working) vmsplice-from-userspace in place.

That sounds about right.  I don't think you can do the latter without an MMU,
but I may be wrong about what it does.

David

Re: [PATCH -mm] splice: Disable vmsplice on nommu.

[Jens Axboe]

On Tue, Sep 11 2007, David Howells wrote:
> Jens Axboe <jens.axboe@oracle.com> wrote:
> 
> > It's a bit of a sledge hammer... Since get_user_pages() appears to work
> > on nummu, we can just disable the vmsplice-to-userspace bits and leave
> > the (working) vmsplice-from-userspace in place.
> 
> That sounds about right.  I don't think you can do the latter without
> an MMU, but I may be wrong about what it does.

You mean the former? The latter is vmsplice() filling a pipe with user
pages, using get_user_pages() to set them up. And that looks like it
should work on nommu.

The former is vmsplice() using mmap like functionality to map pages into
a user vma.

Just making sure we agree :-)

-- 
Jens Axboe

Re: [PATCH -mm] splice: Disable vmsplice on nommu.

[David Howells]

Jens Axboe <jens.axboe@oracle.com> wrote:

> > > It's a bit of a sledge hammer... Since get_user_pages() appears to work
> > > on nummu, we can just disable the vmsplice-to-userspace bits and leave
> > > the (working) vmsplice-from-userspace in place.
> > 
> > That sounds about right.  I don't think you can do the latter without
> > an MMU, but I may be wrong about what it does.
> 
> You mean the former? The latter is vmsplice() filling a pipe with user
> pages, using get_user_pages() to set them up. And that looks like it
> should work on nommu.
> 
> The former is vmsplice() using mmap like functionality to map pages into
> a user vma.

Sorry, yes.  The former.

David