Re: [PATCH v2] mm: Introduce new function vm_insert_kmem_page

[Souptick Joarder]

On Thu, Oct 4, 2018 at 11:47 PM Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote:
>
> On Thu, Oct 04, 2018 at 11:42:18PM +0530, Souptick Joarder wrote:
> > On Thu, Oct 4, 2018 at 6:04 PM Russell King - ARM Linux
> > <linux@xxxxxxxxxxxxxxx> wrote:
> > > I'm confused, what are you trying to do?
> > >
> > > It seems that we already have:
> > >
> > > vm_insert_page() - returns an errno
> > > vmf_insert_page() - returns a VM_FAULT_* code
> > >
> > > From what I _think_ you're saying, you're trying to provide
> > > vm_insert_kmem_page() as a direct replacement for the existing
> > > vm_insert_page(), and then make vm_insert_page() behave as per
> > > vmf_insert_page(), so we end up with:
> >
> > yes, vm_insert_kmem_page() can be a direct replacement of vm_insert_page
> > or might be a wrapper function written using vm_insert_page whichever
> > suites better
> > based on feedback.
> >
> > >
> > > vm_insert_kmem_page() - returns an errno
> > > vm_insert_page() - returns a VM_FAULT_* code
> > > vmf_insert_page() - returns a VM_FAULT_* code and is identical to
> > >       vm_insert_page()
> > >
> >
> > After completion of conversion we end up with
> >
> >  vm_insert_kmem_page() - returns an errno
> >  vmf_insert_page() - returns a VM_FAULT_* code
> >
> >
> > > Given that the documentation for vm_insert_page() says:
> > >
> > >  * Usually this function is called from f_op->mmap() handler
> > >  * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
> > >  * Caller must set VM_MIXEDMAP on vma if it wants to call this
> > >  * function from other places, for example from page-fault handler.
> > >
> > > this says that the "usual" use method for vm_insert_page() is
> > > _outside_ of page fault handling - if it is used _inside_ page fault
> > > handling, then it states that additional fixups are required on the
> > > VMA.  So I don't get why your patch commentry seems to be saying that
> > > users of vm_insert_page() outside of page fault handling all need to
> > > be patched - isn't this the use case that this function is defined
> > > to be handling?
> >
> > The answer is yes best of my knowledge.
> >
> > But as mentioned in change log ->
> >
> > Going forward, the plan is to restrict future drivers not
> > to use vm_insert_page ( *it will generate new errno to
> > VM_FAULT_CODE mapping code for new drivers which were already
> > cleaned up for existing drivers*) in #PF (page fault handler)
> > context but to make use of vmf_insert_page which returns
> > VMF_FAULT_CODE and that is not possible until both vm_insert_page
> > and vmf_insert_page API exists.
> >
> > But there are some consumers of vm_insert_page which use it
> > outside #PF context. straight forward conversion of vm_insert_page
> > to vmf_insert_page won't work there as those function calls expects
> > errno not vm_fault_t in return.
> >
> > If both {vm, vmf}_insert_page exists, vm_insert_page might be used for
> > #PF context which we want to protect by removing/ replacing vm_insert_page
> > with another similar/ wrapper API.
> >
> > Is that the right answer of your question ? no ?
>
> I think this is a bad plan.  What we should rather do is examine the current
> users of vm_insert_page() and ask "What interface would better replace
> vm_insert_page()?"
>
> As I've said to you before, I believe the right answer is to have a
> vm_insert_range() which takes an array of struct page pointers.  That
> fits the majority of remaining users.

Ok, but it will take some time.
Is it a good idea to introduce the final vm_fault_t patch and then
start working on vm_insert_range as it will be bit time consuming ?

>
> ----
>
> If we do want to rename vm_insert_page() to vm_insert_kmem_page(), then
> the right answer is to _just do that_.  Not duplicate vm_insert_page()
> in its entirety.  I don't see the point to doing that.