Re: [PATCH v5] mm: softdirty: enable write notifications on VMAs after VM_SOFTDIRTY cleared

[Peter Feiner <pfeiner@google.com>]

On Thu, Sep 04, 2014 at 09:43:11AM -0700, Peter Feiner wrote:
> On Mon, Aug 25, 2014 at 09:45:34PM -0700, Hugh Dickins wrote:
> > That sets me wondering: have you placed the VM_SOFTDIRTY check in the
> > right place in this series of tests?
> > 
> > I think, once pgprot_modify() is correct on all architectures,
> > it should be possible to drop that pgprot_val() check from
> > vma_wants_writenotify() - which would be a welcome simplification.
> > 
> > But what about the VM_PFNMAP test below it?  If that test was necessary,
> > then having your VM_SOFTDIRTY check before it seems dangerous.  But I'm
> > hoping we can persuade ourselves that the VM_PFNMAP test was unnecessary,
> > and simply delete it.
> 
> If VM_PFNMAP is necessary, then I definitely put the VM_SOFTDIRTY check in the
> wrong spot :-) I don't know much (i.e., anything) about VM_PFNMAP, so I'll
> have to bone up on a lot of code before I have an informed opinion about the
> necessity of the check.

AFAICT, the VM_PFNMAP check is unnecessary since I can't find any drivers that
set VM_PFNMAP and enable dirty accounting on their mappings. If anything,
VM_PFNMAP precludes mapping dirty tracking since set_page_dirty takes a
struct_page argument! Perhaps the VM_PFNMAP check was originally put in
vma_wants_writenotify as a safeguard against bogus calls to set_page_dirty?
In any case, it seems harmless to me to put the VM_SOFTDIRTY check before the
VM_PFNMAP check since none of the fault handling code in mm/memory.c calls
set_page_dirty on a VM_PFNMAP fault because either vm_normal_page() returns
NULL or ->fault() / ->page_mkwrite() return VM_FAULT_NOPAGE. Moreover, for
the purpose of softdirty tracking, enabling write notifications on VM_PFNMAP
VMAs is OK since do_wp_page does the right thing when vm_normal_page() returns
NULL.