On Wed, Oct 13, 2021 at 06:55:46PM +0100, Matthew Wilcox wrote:
> On Wed, Oct 13, 2021 at 09:49:31AM -0400, Johannes Weiner wrote:
> > On Wed, Oct 13, 2021 at 04:19:18AM +0100, Matthew Wilcox wrote:
> > > For today, testing PageSlab on the tail page helps the test proceed
> > > in parallel with the action. Looking at slub's kfree() for an example:
> > >
> > > page = virt_to_head_page(x);
> > > if (unlikely(!PageSlab(page))) {
> > > free_nonslab_page(page, object);
> > > return;
> > > }
> > > slab_free(page->slab_cache, page, object, NULL, 1, _RET_IP_);
> > >
> > > Your proposal is certainly an improvement (since gcc doesn't know
> > > that compound_head(compound_head(x)) == compound_head(x)), but I
> > > think checking on the tail page is even better:
> > >
> > > page = virt_to_page(x);
> > > if (unlikely(!PageSlab(page))) {
> > > free_nonslab_page(compound_head(page), object);
> > > return;
> > > }
> > > slab = page_slab(page);
> > > slab_free(slab->slab_cache, slab, object, NULL, 1, _RET_IP_);
> > >
> > > The compound_head() parts can proceed in parallel with the check of
> > > PageSlab().
> > >
> > > As far as the cost of setting PageSlab, those cachelines are already
> > > dirty because we set compound_head on each of those pages already
> > > (or in the case of freeing, we're about to clear compound_head on
> > > each of those pages).
> >
> > ... but this is not. I think the performance gains from this would
> > have to be significant to justify complicating page flags further.
>
> My argument isn't really "this is more efficient", because I think
> the performance gains are pretty minimal. More that I would like to
> be able to write code in the style which we'll want to use when we're
> using dynamically allocated memory descriptors. It's all just code,
> and we can change it at any time, but better to change it to something
> that continues to work well in the future.
>
> I don't think we end up with "virt_to_head_page()" in a dynamically
> allocated memory descriptor world. The head page contains no different
> information from the tail pages, and indeed the tail pages don't know
> that they're tail pages, or where the head page is. Or maybe they're
> all tail pages.
I agree with that, but future-provisioning is a tradeoff.
It'll be trivial to replace virt_to_head_page() with virt_to_page()
and remove compound_head() calls when whatever is left of struct page
will unconditionally point to a memory descriptor. And that can be
part of the changes that make it so.
But in today's codebase, maintaining type flags in the tail pages
while having to go through the headpage to find the type descriptor
makes things unnecessarily complicated in an area that already has
accrued too much tech debt.
I don't think that's a sensible thing to do as of today.
> I could see a world where we had a virt_to_memdesc() which returned
> a generic memory descriptor that could be cast to a struct slab if
> the flags within that memdesc said it was a slab. But I think it works
> out better to tag the memory descriptor pointer with a discriminator
> that defines what the pointer is. Plus it saves a page flag.
>
> Maybe that's the best way to approach it -- how would you want to write
> this part of kfree() when memory descriptors are dynamically allocated?
There are still many question marks on how the split out memory
descriptors actually will look like, and which state is tracked
where. 'struct slab' is an excellent trial balloon.
It's good to have common north stars to set the direction of where to
place efforts ("small struct page, dynamically allocated descriptors
etc.") but I don't think it makes sense to take on yet more tech debt
in this area for a future that may not pan out the way we think now.
