Re: [RFC v3 0/5] Transparent on-demand struct page initialization embedded in the buddy allocator

[Ingo Molnar <mingo@kernel.org>]

* Linus Torvalds <torvalds@linux-foundation.org> wrote:

> [...]
> 
> Ok, so I don't know all the issues, and in many ways I don't even really 
> care. You could do it other ways, I don't think this is a big deal. The 
> part I hate is the runtime hook into the core MM page allocation code, 
> so I'm just throwing out any random thing that comes to my mind that 
> could be used to avoid that part.

So, my hope was that it's possible to have a single, simple, zero-cost 
runtime check [zero cost for already initialized pages], because it can be 
merged into already existing page flag mask checks present here and 
executed for every freshly allocated page:

static inline int check_new_page(struct page *page)
{
        if (unlikely(page_mapcount(page) |
                (page->mapping != NULL)  |
                (atomic_read(&page->_count) != 0)  |
                (page->flags & PAGE_FLAGS_CHECK_AT_PREP) |
                (mem_cgroup_bad_page_check(page)))) {
                bad_page(page);
                return 1;
        }
        return 0;
}

We already run this for every new page allocated and the initialization 
check could hide in PAGE_FLAGS_CHECK_AT_PREP in a zero-cost fashion.

I'd not do any of the ensure_page_is_initialized() or 
__expand_page_initialization() complications in this patch-set - each page 
head represents itself and gets iterated when check_new_page() is done.

During regular bootup we'd initialize like before, except we don't set up 
the page heads but memset() them to zero. With each page head 32 bytes 
this would mean 8 GB of page head memory to clear per 1 TB - with 16 TB 
that's 128 GB to clear - that ought to be possible to do rather quickly, 
perhaps with some smart SMP cross-call approach that makes sure that each 
memset is done in a node-local fashion. [*]

Such an approach should IMO be far smaller and less invasive than the 
patches presented so far: it should be below 100 lines or so.

I don't know why there's such a big difference between the theory I 
outlined and the invasive patch-set implemented so far in practice, 
perhaps I'm missing some complication. I was trying to probe that 
difference, before giving up on the idea and punting back to the async 
hotplug-ish approach which would obviously work well too.

All in one, I think async init just hides the real problem - there's no 
way memory init should take this long.

Thanks,

	Ingo

[*] alternatively maybe the main performance problem is that node-local 
    memory is set up on a remote (boot) node? In that case I'd try to
    optimize it by migrating the memory init code's current node by using
    set_cpus_allowed() to live migrate from node to node, tracking the 
    node whose struct page array is being initialized.