Re: [PATCH -v3 01/10] mm, swap: Make swap cluster size same of THP size on x86_64

[Johannes Weiner]

Hi Ying,

On Tue, Sep 20, 2016 at 10:01:30AM +0800, Huang, Ying wrote:
> It appears all patches other than [10/10] in the series is used by the
> last patch [10/10], directly or indirectly.  And Without [10/10], they
> don't make much sense.  So you suggest me to use one large patch?
> Something like below?  Does that help you to review?

I find this version a lot easier to review, thank you.

> As the first step, in this patch, the splitting huge page is
> delayed from almost the first step of swapping out to after allocating
> the swap space for the THP and adding the THP into the swap cache.
> This will reduce lock acquiring/releasing for the locks used for the
> swap cache management.

I agree that that's a fine goal for this patch series. We can worry
about 2MB IO submissions later on.

> @@ -503,6 +503,19 @@ config FRONTSWAP
>  
>  	  If unsure, say Y to enable frontswap.
>  
> +config ARCH_USES_THP_SWAP_CLUSTER
> +	bool
> +	default n
> +
> +config THP_SWAP_CLUSTER
> +	bool
> +	depends on SWAP && TRANSPARENT_HUGEPAGE && ARCH_USES_THP_SWAP_CLUSTER
> +	default y
> +	help
> +	  Use one swap cluster to hold the contents of the THP
> +	  (Transparent Huge Page) swapped out.  The size of the swap
> +	  cluster will be same as that of THP.

Making swap space allocation and swapcache handling THP-native is not
dependent on the architecture, it's generic VM code. Can you please
just define the cluster size depending on CONFIG_TRANSPARENT_HUGEPAGE?

> @@ -196,7 +196,11 @@ static void discard_swap_cluster(struct
>  	}
>  }
>  
> +#ifdef CONFIG_THP_SWAP_CLUSTER
> +#define SWAPFILE_CLUSTER	(HPAGE_SIZE / PAGE_SIZE)
> +#else
>  #define SWAPFILE_CLUSTER	256
> +#endif
>  #define LATENCY_LIMIT		256

I.e. this?

#ifdef CONFIG_TRANSPARENT_HUGEPAGE
#define SWAPFILE_CLUSTER	HPAGE_PMD_NR
#else
#define SWAPFILE_CLUSTER	256
#endif

> @@ -18,6 +18,13 @@ struct swap_cgroup {
>  };
>  #define SC_PER_PAGE	(PAGE_SIZE/sizeof(struct swap_cgroup))
>  
> +struct swap_cgroup_iter {
> +	struct swap_cgroup_ctrl *ctrl;
> +	struct swap_cgroup *sc;
> +	swp_entry_t entry;
> +	unsigned long flags;
> +};
> +
>  /*
>   * SwapCgroup implements "lookup" and "exchange" operations.
>   * In typical usage, this swap_cgroup is accessed via memcg's charge/uncharge
> @@ -75,6 +82,35 @@ static struct swap_cgroup *lookup_swap_c
>  	return sc + offset % SC_PER_PAGE;
>  }
>  
> +static void swap_cgroup_iter_init(struct swap_cgroup_iter *iter,
> +				  swp_entry_t ent)
> +{
> +	iter->entry = ent;
> +	iter->sc = lookup_swap_cgroup(ent, &iter->ctrl);
> +	spin_lock_irqsave(&iter->ctrl->lock, iter->flags);
> +}
> +
> +static void swap_cgroup_iter_exit(struct swap_cgroup_iter *iter)
> +{
> +	spin_unlock_irqrestore(&iter->ctrl->lock, iter->flags);
> +}
> +
> +/*
> + * swap_cgroup is stored in a kind of discontinuous array.  That is,
> + * they are continuous in one page, but not across page boundary.  And
> + * there is one lock for each page.
> + */
> +static void swap_cgroup_iter_advance(struct swap_cgroup_iter *iter)
> +{
> +	iter->sc++;
> +	iter->entry.val++;
> +	if (!(((unsigned long)iter->sc) & PAGE_MASK)) {
> +		spin_unlock_irqrestore(&iter->ctrl->lock, iter->flags);
> +		iter->sc = lookup_swap_cgroup(iter->entry, &iter->ctrl);
> +		spin_lock_irqsave(&iter->ctrl->lock, iter->flags);
> +	}
> +}
> +
>  /**
>   * swap_cgroup_cmpxchg - cmpxchg mem_cgroup's id for this swp_entry.
>   * @ent: swap entry to be cmpxchged
> @@ -87,45 +123,49 @@ static struct swap_cgroup *lookup_swap_c
>  unsigned short swap_cgroup_cmpxchg(swp_entry_t ent,
>  					unsigned short old, unsigned short new)
>  {
> -	struct swap_cgroup_ctrl *ctrl;
> -	struct swap_cgroup *sc;
> -	unsigned long flags;
> +	struct swap_cgroup_iter iter;
>  	unsigned short retval;
>  
> -	sc = lookup_swap_cgroup(ent, &ctrl);
> +	swap_cgroup_iter_init(&iter, ent);
>  
> -	spin_lock_irqsave(&ctrl->lock, flags);
> -	retval = sc->id;
> +	retval = iter.sc->id;
>  	if (retval == old)
> -		sc->id = new;
> +		iter.sc->id = new;
>  	else
>  		retval = 0;
> -	spin_unlock_irqrestore(&ctrl->lock, flags);
> +
> +	swap_cgroup_iter_exit(&iter);
>  	return retval;
>  }
>  
>  /**
> - * swap_cgroup_record - record mem_cgroup for this swp_entry.
> - * @ent: swap entry to be recorded into
> + * swap_cgroup_record - record mem_cgroup for a set of swap entries
> + * @ent: the first swap entry to be recorded into
>   * @id: mem_cgroup to be recorded
> + * @nr_ents: number of swap entries to be recorded
>   *
>   * Returns old value at success, 0 at failure.
>   * (Of course, old value can be 0.)
>   */
> -unsigned short swap_cgroup_record(swp_entry_t ent, unsigned short id)
> +unsigned short swap_cgroup_record(swp_entry_t ent, unsigned short id,
> +				  unsigned int nr_ents)
>  {
> -	struct swap_cgroup_ctrl *ctrl;
> -	struct swap_cgroup *sc;
> +	struct swap_cgroup_iter iter;
>  	unsigned short old;
> -	unsigned long flags;
>  
> -	sc = lookup_swap_cgroup(ent, &ctrl);
> +	swap_cgroup_iter_init(&iter, ent);
>  
> -	spin_lock_irqsave(&ctrl->lock, flags);
> -	old = sc->id;
> -	sc->id = id;
> -	spin_unlock_irqrestore(&ctrl->lock, flags);
> +	old = iter.sc->id;
> +	for (;;) {
> +		VM_BUG_ON(iter.sc->id != old);
> +		iter.sc->id = id;
> +		nr_ents--;
> +		if (!nr_ents)
> +			break;
> +		swap_cgroup_iter_advance(&iter);
> +	}
>  
> +	swap_cgroup_iter_exit(&iter);
>  	return old;
>  }

The iterator seems overkill for one real user, and it's undesirable in
the single-slot access from swap_cgroup_cmpxchg(). How about something
like the following?

static struct swap_cgroup *lookup_swap_cgroup(struct swap_cgroup_ctrl *ctrl,
					      pgoff_t offset)
{
	struct page *page;

	page = page_address(ctrl->map[offset / SC_PER_PAGE]);
	return page + (offset % SC_PER_PAGE);
}

unsigned short swap_cgroup_cmpxchg(swp_entry_t ent,
					unsigned short old, unsigned short new)
{
	struct swap_cgroup_ctrl *ctrl;
	struct swap_cgroup *sc;
	unsigned long flags;
	unsigned short retval;
	pgoff_t off = swp_offset(ent);

	ctrl = &swap_cgroup_ctrl[swp_type(ent)];
	sc = lookup_swap_cgroup(ctrl, swp_offset(ent));

	spin_lock_irqsave(&ctrl->lock, flags);
	retval = sc->id;
	if (retval == old)
		sc->id = new;
	else
		retval = 0;
	spin_unlock_irqrestore(&ctrl->lock, flags);

	return retval;
}

unsigned short swap_cgroup_record(swp_entry_t ent, unsigned short id,
				  unsigned int nr_entries)
{
	struct swap_cgroup_ctrl *ctrl;
	struct swap_cgroup *sc;
	unsigned short old;
	unsigned long flags;

	ctrl = &swap_cgroup_ctrl[swp_type(ent)];
	sc = lookup_swap_cgroup(ctrl, offset);
	end = offset + nr_entries;

	spin_lock_irqsave(&ctrl->lock, flags);
	old = sc->id;
	while (offset != end) {
		sc->id = id;
		offset++;
		if (offset % SC_PER_PAGE)
			sc++;
		else
			sc = lookup_swap_cgroup(ctrl, offset);
	}
	spin_unlock_irqrestore(&ctrl->lock, flags);

	return old;
}

> @@ -145,20 +162,66 @@ void __delete_from_swap_cache(struct pag
>  
>  	entry.val = page_private(page);
>  	address_space = swap_address_space(entry);
> -	radix_tree_delete(&address_space->page_tree, page_private(page));
> -	set_page_private(page, 0);
>  	ClearPageSwapCache(page);
> -	address_space->nrpages--;
> -	__dec_node_page_state(page, NR_FILE_PAGES);
> -	INC_CACHE_INFO(del_total);
> +	for (i = 0; i < nr; i++) {
> +		struct page *cur_page = page + i;
> +
> +		radix_tree_delete(&address_space->page_tree,
> +				  page_private(cur_page));
> +		set_page_private(cur_page, 0);
> +	}
> +	address_space->nrpages -= nr;
> +	__mod_node_page_state(page_pgdat(page), NR_FILE_PAGES, -nr);
> +	ADD_CACHE_INFO(del_total, nr);
> +}
> +
> +#ifdef CONFIG_THP_SWAP_CLUSTER
> +int add_to_swap_trans_huge(struct page *page, struct list_head *list)
> +{
> +	swp_entry_t entry;
> +	int ret = 0;
> +
> +	/* cannot split, which may be needed during swap in, skip it */
> +	if (!can_split_huge_page(page))
> +		return -EBUSY;
> +	/* fallback to split huge page firstly if no PMD map */
> +	if (!compound_mapcount(page))
> +		return 0;

The can_split_huge_page() (and maybe also the compound_mapcount())
optimizations look like they could be split out into separate followup
patches. They're definitely nice to have, but don't seem necessary to
make this patch minimally complete.

> @@ -168,11 +231,23 @@ int add_to_swap(struct page *page, struc
>  	VM_BUG_ON_PAGE(!PageLocked(page), page);
>  	VM_BUG_ON_PAGE(!PageUptodate(page), page);
>  
> +	if (unlikely(PageTransHuge(page))) {
> +		err = add_to_swap_trans_huge(page, list);
> +		switch (err) {
> +		case 1:
> +			return 1;
> +		case 0:
> +			/* fallback to split firstly if return 0 */
> +			break;
> +		default:
> +			return 0;
> +		}
> +	}
>  	entry = get_swap_page();
>  	if (!entry.val)
>  		return 0;
>  
> -	if (mem_cgroup_try_charge_swap(page, entry)) {
> +	if (mem_cgroup_try_charge_swap(page, entry, 1)) {
>  		swapcache_free(entry);
>  		return 0;
>  	}

Instead of duplicating the control flow at such a high level -
add_to_swap() and add_to_swap_trans_huge() are basically identical -
it's better push down the THP handling as low as possible:

Pass the page to get_swap_page(), and then decide in there whether
it's THP and you need to allocate a single entry or a cluster.

And mem_cgroup_try_charge_swap() already gets the page. Again, check
in there how much swap to charge based on the passed page instead of
passing the same information twice.

Doing that will change the structure of the patch too much to review
the paths below in their current form. I'll have a closer look in the
next version.

Thanks!