On Tue, Aug 15, 2017 at 9:29 AM, Dan Williams [off-list ref] wrote:

On Tue, Aug 15, 2017 at 5:42 AM, Jan Kara [off-list ref] wrote:

quoted

On Mon 14-08-17 23:12:22, Dan Williams wrote:

quoted

diff --git a/include/linux/mm_types.h b/include/linux/mm_types.h
index ff151814a02d..73fdc0ada9ee 100644
--- a/include/linux/mm_types.h
+++ b/include/linux/mm_types.h

@@ -306,6 +306,7 @@ struct vm_area_struct {
      struct mm_struct *vm_mm;        /* The address space we belong to. */
      pgprot_t vm_page_prot;          /* Access permissions of this VMA. */
      unsigned long vm_flags;         /* Flags, see mm.h. */
+     unsigned long fs_flags;         /* fs flags, see MAP_DIRECT etc */

      /*
       * For areas with an address space and backing store,

Ah, OK, here are VMA flags I was missing in the previous patch :) But why
did you create separate fs_flags field for this? on 64-bit archs there's
still space in vm_flags and frankly I don't see why we should separate
MAP_DIRECT or MAP_SYNC from other flags?

Where would MAP_DIRECT go in the 32-bit case?

quoted

After all a difference in these
flags must also prevent VMA merging (which you forgot to handle I think)
and they need to be copied on split (which happens by chance even now).

Ah, yes I did miss blocking the merge of a vma with MAP_DIRECT and one
without. However, the vma split path looks ok.

The merge path already blocks merging vmas that have the ->close()
operation defined in is_mergeable_vma().