Re: [PATCH 1/3] Balance Freeing of Huge Pages across Nodes

[Lee Schermerhorn <Lee.Schermerhorn@hp.com>]

On Tue, 2009-06-30 at 14:05 +0100, Mel Gorman wrote:
> On Mon, Jun 29, 2009 at 05:52:34PM -0400, Lee Schermerhorn wrote:
> > [PATCH] 1/3 Balance Freeing of Huge Pages across Nodes
> > 
> > Against:  25jun09 mmotm
> > 
> > Free huges pages from nodes in round robin fashion in an
> > attempt to keep [persistent a.k.a static] hugepages balanced
> > across nodes
> > 
> > New function free_pool_huge_page() is modeled on and
> > performs roughly the inverse of alloc_fresh_huge_page().
> > Replaces dequeue_huge_page() which now has no callers,
> > so this patch removes it.
> > 
> > Helper function hstate_next_node_to_free() uses new hstate
> > member next_to_free_nid to distribute "frees" across all
> > nodes with huge pages.
> > 
> > V2:
> > 
> > At Mel Gorman's suggestion:  renamed hstate_next_node() to
> > hstate_next_node_to_alloc() for symmetry.  Also, renamed
> > hstate member hugetlb_next_node to next_node_to_free.
> > ["hugetlb" is implicit in the hstate struct, I think].
> > 
> > New in this version:
> > 
> > Modified adjust_pool_surplus() to use hstate_next_node_to_alloc()
> > and hstate_next_node_to_free() to advance node id for adjusting
> > surplus huge page count, as this is equivalent to allocating and
> > freeing persistent huge pages.  [Can't blame Mel for this part.]
> > 
> > V3:
> > 
> > Minor cleanup: rename 'nid' to 'next_nid' in free_pool_huge_page() to
> > better match alloc_fresh_huge_page() conventions.
> > 
> > Acked-by: David Rientjes <rientjes@google.com>
> > Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
> > 
> >  include/linux/hugetlb.h |    3 -
> >  mm/hugetlb.c            |  132 +++++++++++++++++++++++++++++++-----------------
> >  2 files changed, 88 insertions(+), 47 deletions(-)
> > 
> > Index: linux-2.6.31-rc1-mmotm-090625-1549/include/linux/hugetlb.h
> > ===================================================================
> > --- linux-2.6.31-rc1-mmotm-090625-1549.orig/include/linux/hugetlb.h	2009-06-29 10:21:12.000000000 -0400
> > +++ linux-2.6.31-rc1-mmotm-090625-1549/include/linux/hugetlb.h	2009-06-29 10:27:18.000000000 -0400
> > @@ -183,7 +183,8 @@ unsigned long hugetlb_get_unmapped_area(
> >  #define HSTATE_NAME_LEN 32
> >  /* Defines one hugetlb page size */
> >  struct hstate {
> > -	int hugetlb_next_nid;
> > +	int next_nid_to_alloc;
> > +	int next_nid_to_free;
> >  	unsigned int order;
> >  	unsigned long mask;
> >  	unsigned long max_huge_pages;
> > Index: linux-2.6.31-rc1-mmotm-090625-1549/mm/hugetlb.c
> > ===================================================================
> > --- linux-2.6.31-rc1-mmotm-090625-1549.orig/mm/hugetlb.c	2009-06-29 10:21:12.000000000 -0400
> > +++ linux-2.6.31-rc1-mmotm-090625-1549/mm/hugetlb.c	2009-06-29 15:53:55.000000000 -0400
> > @@ -455,24 +455,6 @@ static void enqueue_huge_page(struct hst
> >  	h->free_huge_pages_node[nid]++;
> >  }
> >  
> > -static struct page *dequeue_huge_page(struct hstate *h)
> > -{
> > -	int nid;
> > -	struct page *page = NULL;
> > -
> > -	for (nid = 0; nid < MAX_NUMNODES; ++nid) {
> > -		if (!list_empty(&h->hugepage_freelists[nid])) {
> > -			page = list_entry(h->hugepage_freelists[nid].next,
> > -					  struct page, lru);
> > -			list_del(&page->lru);
> > -			h->free_huge_pages--;
> > -			h->free_huge_pages_node[nid]--;
> > -			break;
> > -		}
> > -	}
> > -	return page;
> > -}
> > -
> >  static struct page *dequeue_huge_page_vma(struct hstate *h,
> >  				struct vm_area_struct *vma,
> >  				unsigned long address, int avoid_reserve)
> > @@ -640,7 +622,7 @@ static struct page *alloc_fresh_huge_pag
> >  
> >  /*
> >   * Use a helper variable to find the next node and then
> > - * copy it back to hugetlb_next_nid afterwards:
> > + * copy it back to next_nid_to_alloc afterwards:
> >   * otherwise there's a window in which a racer might
> >   * pass invalid nid MAX_NUMNODES to alloc_pages_exact_node.
> >   * But we don't need to use a spin_lock here: it really
> > @@ -649,13 +631,13 @@ static struct page *alloc_fresh_huge_pag
> >   * if we just successfully allocated a hugepage so that
> >   * the next caller gets hugepages on the next node.
> >   */
> > -static int hstate_next_node(struct hstate *h)
> > +static int hstate_next_node_to_alloc(struct hstate *h)
> >  {
> >  	int next_nid;
> > -	next_nid = next_node(h->hugetlb_next_nid, node_online_map);
> > +	next_nid = next_node(h->next_nid_to_alloc, node_online_map);
> >  	if (next_nid == MAX_NUMNODES)
> >  		next_nid = first_node(node_online_map);
> > -	h->hugetlb_next_nid = next_nid;
> > +	h->next_nid_to_alloc = next_nid;
> >  	return next_nid;
> >  }
> >  
> 
> Strictly speaking, next_nid_to_alloc looks more like last_nid_alloced but I
> don't think it makes an important difference. Implementing it this way is
> shorter and automatically ensures next_nid is an online node. 
> 
> If you wanted to be pedantic, I think the following untested code would
> make it really next_nid_to_alloc but I don't think it's terribly
> important.
> 
> static int hstate_next_node_to_alloc(struct hstate *h)
> {
> 	int this_nid = h->next_nid_to_alloc;
> 
> 	/* Check the node didn't get off-lined since */
> 	if (unlikely(!node_online(next_nid))) {
> 		this_nid = next_node(h->next_nid_to_alloc, node_online_map);
> 		h->next_nid_to_alloc = this_nid;
> 	}
> 
> 	h->next_nid_to_alloc = next_node(h->next_nid_to_alloc, node_online_map);
> 	if (h->next_nid_to_alloc == MAX_NUMNODES)
> 		h->next_nid_to_alloc = first_node(node_online_map);
> 
> 	return this_nid;
> }

Mel:  

I'm about to send out a series that constrains [persistent] huge page
alloc and free using task mempolicy, per your suggestion.  The functions
'next_node_to_{alloc|free} and how they're used get reworked in that
series quite a bit, and the name becomes more accurate, I think.  And, I
think it does handle the node going offline along with handling changing
to a new policy nodemask that doesn't include the value saved in the
hstate.  We can revisit this, then.

However, the way we currently use these functions, they do update the
'next_node_*' field in the hstate, and where the return value is tested
[against start_nid], it really is the "next" node.  If the alloc/free
succeeds, then the return value does turn out to be the [last] node we
just alloc'd/freed on.  But, again, we've advanced the next node to
alloc/free in the hstate.  A nit, I think :).

> 
> > @@ -666,14 +648,15 @@ static int alloc_fresh_huge_page(struct 
> >  	int next_nid;
> >  	int ret = 0;
> >  
> > -	start_nid = h->hugetlb_next_nid;
> > +	start_nid = h->next_nid_to_alloc;
> > +	next_nid = start_nid;
> >  
> >  	do {
> > -		page = alloc_fresh_huge_page_node(h, h->hugetlb_next_nid);
> > +		page = alloc_fresh_huge_page_node(h, next_nid);
> >  		if (page)
> >  			ret = 1;
> > -		next_nid = hstate_next_node(h);
> > -	} while (!page && h->hugetlb_next_nid != start_nid);
> > +		next_nid = hstate_next_node_to_alloc(h);
> > +	} while (!page && next_nid != start_nid);
> >  
> >  	if (ret)
> >  		count_vm_event(HTLB_BUDDY_PGALLOC);
> > @@ -683,6 +666,52 @@ static int alloc_fresh_huge_page(struct 
> >  	return ret;
> >  }
> >  
> > +/*
> > + * helper for free_pool_huge_page() - find next node
> > + * from which to free a huge page
> > + */
> > +static int hstate_next_node_to_free(struct hstate *h)
> > +{
> > +	int next_nid;
> > +	next_nid = next_node(h->next_nid_to_free, node_online_map);
> > +	if (next_nid == MAX_NUMNODES)
> > +		next_nid = first_node(node_online_map);
> > +	h->next_nid_to_free = next_nid;
> > +	return next_nid;
> > +}
> > +
> > +/*
> > + * Free huge page from pool from next node to free.
> > + * Attempt to keep persistent huge pages more or less
> > + * balanced over allowed nodes.
> > + * Called with hugetlb_lock locked.
> > + */
> > +static int free_pool_huge_page(struct hstate *h)
> > +{
> > +	int start_nid;
> > +	int next_nid;
> > +	int ret = 0;
> > +
> > +	start_nid = h->next_nid_to_free;
> > +	next_nid = start_nid;
> > +
> > +	do {
> > +		if (!list_empty(&h->hugepage_freelists[next_nid])) {
> > +			struct page *page =
> > +				list_entry(h->hugepage_freelists[next_nid].next,
> > +					  struct page, lru);
> > +			list_del(&page->lru);
> > +			h->free_huge_pages--;
> > +			h->free_huge_pages_node[next_nid]--;
> > +			update_and_free_page(h, page);
> > +			ret = 1;
> > +		}
> > +		next_nid = hstate_next_node_to_free(h);
> > +	} while (!ret && next_nid != start_nid);
> > +
> > +	return ret;
> > +}
> > +
> >  static struct page *alloc_buddy_huge_page(struct hstate *h,
> >  			struct vm_area_struct *vma, unsigned long address)
> >  {
> > @@ -1007,7 +1036,7 @@ int __weak alloc_bootmem_huge_page(struc
> >  		void *addr;
> >  
> >  		addr = __alloc_bootmem_node_nopanic(
> > -				NODE_DATA(h->hugetlb_next_nid),
> > +				NODE_DATA(h->next_nid_to_alloc),
> >  				huge_page_size(h), huge_page_size(h), 0);
> >  
> >  		if (addr) {
> > @@ -1019,7 +1048,7 @@ int __weak alloc_bootmem_huge_page(struc
> >  			m = addr;
> >  			goto found;
> >  		}
> > -		hstate_next_node(h);
> > +		hstate_next_node_to_alloc(h);
> >  		nr_nodes--;
> >  	}
> >  	return 0;
> > @@ -1140,31 +1169,43 @@ static inline void try_to_free_low(struc
> >   */
> >  static int adjust_pool_surplus(struct hstate *h, int delta)
> >  {
> > -	static int prev_nid;
> > -	int nid = prev_nid;
> > +	int start_nid, next_nid;
> >  	int ret = 0;
> >  
> >  	VM_BUG_ON(delta != -1 && delta != 1);
> > -	do {
> > -		nid = next_node(nid, node_online_map);
> > -		if (nid == MAX_NUMNODES)
> > -			nid = first_node(node_online_map);
> >  
> > -		/* To shrink on this node, there must be a surplus page */
> > -		if (delta < 0 && !h->surplus_huge_pages_node[nid])
> > -			continue;
> > -		/* Surplus cannot exceed the total number of pages */
> > -		if (delta > 0 && h->surplus_huge_pages_node[nid] >=
> > +	if (delta < 0)
> > +		start_nid = h->next_nid_to_alloc;
> > +	else
> > +		start_nid = h->next_nid_to_free;
> > +	next_nid = start_nid;
> > +
> > +	do {
> > +		int nid = next_nid;
> > +		if (delta < 0)  {
> > +			next_nid = hstate_next_node_to_alloc(h);
> > +			/*
> > +			 * To shrink on this node, there must be a surplus page
> > +			 */
> > +			if (!h->surplus_huge_pages_node[nid])
> > +				continue;
> > +		}
> > +		if (delta > 0) {
> > +			next_nid = hstate_next_node_to_free(h);
> > +			/*
> > +			 * Surplus cannot exceed the total number of pages
> > +			 */
> > +			if (h->surplus_huge_pages_node[nid] >=
> >  						h->nr_huge_pages_node[nid])
> > -			continue;
> > +				continue;
> > +		}
> >  
> >  		h->surplus_huge_pages += delta;
> >  		h->surplus_huge_pages_node[nid] += delta;
> >  		ret = 1;
> >  		break;
> > -	} while (nid != prev_nid);
> > +	} while (next_nid != start_nid);
> >  
> > -	prev_nid = nid;
> >  	return ret;
> >  }
> >  
> > @@ -1226,10 +1267,8 @@ static unsigned long set_max_huge_pages(
> >  	min_count = max(count, min_count);
> >  	try_to_free_low(h, min_count);
> >  	while (min_count < persistent_huge_pages(h)) {
> > -		struct page *page = dequeue_huge_page(h);
> > -		if (!page)
> > +		if (!free_pool_huge_page(h))
> >  			break;
> > -		update_and_free_page(h, page);
> >  	}
> >  	while (count < persistent_huge_pages(h)) {
> >  		if (!adjust_pool_surplus(h, 1))
> > @@ -1441,7 +1480,8 @@ void __init hugetlb_add_hstate(unsigned 
> >  	h->free_huge_pages = 0;
> >  	for (i = 0; i < MAX_NUMNODES; ++i)
> >  		INIT_LIST_HEAD(&h->hugepage_freelists[i]);
> > -	h->hugetlb_next_nid = first_node(node_online_map);
> > +	h->next_nid_to_alloc = first_node(node_online_map);
> > +	h->next_nid_to_free = first_node(node_online_map);
> >  	snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
> >  					huge_page_size(h)/1024);
> >  
> 
> Nothing problematic jumps out at me. Even with hstate_next_node_to_alloc()
> as it is;
> 
> Acked-by: Mel Gorman <mel@csn.ul.ie>
> 

Thanks.  It did seem to test out OK on ia64 [12jun mmotm; 25jun mmotm
has a problem there--TBI] and x86_64.  Could use more testing, tho'.
Especially with various combinations of persistent and surplus huge
pages.  I saw you mention that you have a hugetlb regression test suite.
Is that available "out there, somewhere"?  I just grabbed a libhugetlbfs
source rpm, but haven't cracked it yet.  Maybe it's there?

Lee

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>