Re: [PATCH/RFC G-U-P experts] IB/umem: Modernize our get_user_pages() parameters

[Hugh Dickins <hughd@google.com>]

[-- Attachment #1: Type: TEXT/PLAIN, Size: 2193 bytes --]

On Fri, 27 Jan 2012, Roland Dreier wrote:

> > Sigh, what a mess ... it seems what we really want to do is know
> > if userspace might trigger a COW because or not, and only do a
> > preemptive COW in that case.  (We're not really concerned with
> > userspace fork()ing and setting up a COW in the future, since that's
> > what we have MADV_DONTFORK for)
> >
> > The status quo works for userspace anonymous mappings but
> > it doesn't work for my case of mapping a kernel buffer read-only
> > into userspace.  And fixing my case breaks the anonymous case.
> > Do you see a way out of this dilemma?  Do we need to add yet
> > another flag to get_user_pages()?
> 
> So thinking about this a bit more... it seems what we want is at least
> to first order that we do the equivalent of write==1 exactly when the vma
> for a mapping has VM_WRITE set

My first impression is that that's not what you want at all: that will
not do a preliminary COW of anonymous page to be written into by the
driver when the user only wants VM_READ access.  But perhaps I'm
worrying about the second order while you're sticking to first order.

Or perhaps I'm misunderstanding the context in which you want to do
this: are you now accepting to do a different get_user_pages in the
anonymous and driver-memory cases, and this suggestion was for the
driver-memory case only?

> (or is it VMA_MAYWRITE / force==1?
> I don't quite understand the distinction between WRITE and MAYWRITE).

I may have told you more than you wanted to know in the other mail.

> 
> Right now, one call to get_user_pages() might involve more than one vma,
> but we could simulate the above by doing find_vma() and making sure our
> call to get_user_pages() goes one vma at a time.  Of course that would be
> inefficient since get_user_pages() will redo the find_vma() internally, so it
> would I guess make sense to add another FOLL_ flag to tell
> get_user_pages() to do this?

I cannot go further, without explanation for why you need get_user_pages
in the driver-memory case at all.

> 
> Am I all wet, or am I becoming an MM hacker?

Certainly more than I'll ever be an RDMA hacker,

Hugh