[Intel-gfx] [PATCH RFC] mm: Add f_ops->populate()

[Matthew Wilcox]

On Sun, Mar 06, 2022 at 06:11:21AM +0200, Jarkko Sakkinen wrote:
> On Sun, Mar 06, 2022 at 03:52:12AM +0000, Matthew Wilcox wrote:
> > On Sun, Mar 06, 2022 at 05:21:11AM +0200, Jarkko Sakkinen wrote:
> > > On Sun, Mar 06, 2022 at 02:57:55AM +0000, Matthew Wilcox wrote:
> > > > On Sun, Mar 06, 2022 at 04:15:33AM +0200, Jarkko Sakkinen wrote:
> > > > > Sometimes you might want to use MAP_POPULATE to ask a device driver to
> > > > > initialize the device memory in some specific manner. SGX driver can use
> > > > > this to request more memory by issuing ENCLS[EAUG] x86 opcode for each
> > > > > page in the address range.
> > > > > 
> > > > > Add f_ops->populate() with the same parameters as f_ops->mmap() and make
> > > > > it conditionally called inside call_mmap(). Update call sites
> > > > > accodingly.
> > > > 
> > > > Your device driver has a ->mmap operation.  Why does it need another
> > > > one?  More explanation required here.
> > > 
> > > f_ops->mmap() would require an additional parameter, which results
> > > heavy refactoring.
> > > 
> > > struct file_operations has 1125 references in the kernel tree, so I
> > > decided to check this way around first. 
> > 
> > Are you saying that your device driver behaves differently if
> > MAP_POPULATE is set versus if it isn't?  That seems hideously broken.
> 
> MAP_POPULATE does not do anything (according to __mm_populate in mm/gup.c)
> with VMA's that have some sort of device/IO memory, i.e. vm_flags
> intersecting with VM_PFNMAP | VM_IO.
> 
> I can extend the guard obviously to:
> 
> if (!ret && do_populate && file->f_op->populate &&
>     !!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
>         file->f_op->populate(file, vma);

Are you deliberately avoiding the question?  I'm not asking about
implementation.  I'm asking about the semantics of MAP_POPULATE with
your driver.