Re: [PATCH v4 2/2] mm: hugetlb: optionally allocate gigantic hugepages using cma

[Roman Gushchin]

On Tue, Apr 07, 2020 at 05:40:05PM +0200, Michal Hocko wrote:
> On Tue 07-04-20 08:25:44, Roman Gushchin wrote:
> > On Tue, Apr 07, 2020 at 09:03:31AM +0200, Michal Hocko wrote:
> > > On Mon 06-04-20 18:04:31, Roman Gushchin wrote:
> > > [...]
> > > My ack still applies but I have only noticed two minor things now.
> > 
> > Hello, Michal!
> > 
> > > 
> > > [...]
> > > > @@ -1281,8 +1308,14 @@ static void update_and_free_page(struct hstate *h, struct page *page)
> > > >  	set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
> > > >  	set_page_refcounted(page);
> > > >  	if (hstate_is_gigantic(h)) {
> > > > +		/*
> > > > +		 * Temporarily drop the hugetlb_lock, because
> > > > +		 * we might block in free_gigantic_page().
> > > > +		 */
> > > > +		spin_unlock(&hugetlb_lock);
> > > >  		destroy_compound_gigantic_page(page, huge_page_order(h));
> > > >  		free_gigantic_page(page, huge_page_order(h));
> > > > +		spin_lock(&hugetlb_lock);
> > > 
> > > This is OK with the current code because existing paths do not have to
> > > revalidate the state AFAICS but it is a bit subtle. I have checked the
> > > cma_free path and it can only sleep on the cma->lock unless I am missing
> > > something. This lock is only used for cma bitmap manipulation and the
> > > mutex sounds like an overkill there and it can be replaced by a
> > > spinlock.
> > > 
> > > Sounds like a follow up patch material to me.
> > 
> > I had the same idea and even posted a patch:
> > https://lore.kernel.org/linux-mm/20200403174559.GC220160@xxxxxxxxxx/T/#m87be98bdacda02cea3dd6759b48a28bd23f29ff0
> > 
> > However, Joonsoo pointed out that in some cases the bitmap operation might
> > be too long for a spinlock.
> 
> I was not aware of this email thread. I will have a look. Thanks!
>  
> > Alternatively, we can implement an asynchronous delayed release on the cma side,
> > I just don't know if it's worth it (I mean adding code/complexity).
> > 
> > > 
> > > [...]
> > > > +	for_each_node_state(nid, N_ONLINE) {
> > > > +		int res;
> > > > +
> > > > +		size = min(per_node, hugetlb_cma_size - reserved);
> > > > +		size = round_up(size, PAGE_SIZE << order);
> > > > +
> > > > +		res = cma_declare_contiguous_nid(0, size, 0, PAGE_SIZE << order,
> > > > +						 0, false, "hugetlb",
> > > > +						 &hugetlb_cma[nid], nid);
> > > > +		if (res) {
> > > > +			pr_warn("hugetlb_cma: reservation failed: err %d, node %d",
> > > > +				res, nid);
> > > > +			break;
> > > 
> > > Do we really have to break out after a single node failure? There might
> > > be other nodes that can satisfy the allocation. You are not cleaning up
> > > previous allocations so there is a partial state and then it would make
> > > more sense to me to simply s@break@continue@ here.
> > 
> > But then we should iterate over all nodes in alloc_gigantic_page()?
> 
> OK, I've managed to miss the early break on hugetlb_cma[node] == NULL
> there as well. I do not think this makes much sense. Just consider a
> setup with one node much smaller than others (not unseen on LPAR
> configurations) and then you are potentially using CMA areas on some
> nodes without a good reason.
> 
> > Currently if hugetlb_cma[0] is NULL it will immediately switch back
> > to the fallback approach.
> > 
> > Actually, Idk how realistic are use cases with complex node configuration,
> > so that we can hugetlb_cma areas can be allocated only on some of them.
> > I'd leave it up to the moment when we'll have a real world example.
> > Then we probably want something more sophisticated anyway...
> 
> I do not follow. Isn't the s@break@continue@ in this and
> alloc_gigantic_page path enough to make it work?

Well, of course it will. But for a highly asymmetrical configuration
there is probably not much sense to try allocate cma areas of a similar
size on each node and rely on allocation failures on some of them.

But, again, if you strictly prefer s/break/continue, I can send a v5.
Just let me know.

Thanks!