diff mbox series

[v2,1/2] mm/page_alloc: Do not calculate node's total pages and memmap pages when empty

Message ID 20220621041717.6355-2-osalvador@suse.de (mailing list archive)
State New
Headers show
Series Minor memoryhotplug refactoring | expand

Commit Message

Oscar Salvador June 21, 2022, 4:17 a.m. UTC
free_area_init_node() calls calculate_node_totalpages() and
free_area_init_core(). The former to get node's {spanned,present}_pages,
and the latter to calculate, among other things, how many pages per zone
we spent on memmap_pages, which is used to substract zone's free pages.

On memoryless-nodes, it is pointless to perform such a bunch of work, so
make sure we skip the calculations when having a node or empty zone.

Signed-off-by: Oscar Salvador <osalvador@suse.de>
---
 mm/page_alloc.c | 16 ++++++++++++----
 1 file changed, 12 insertions(+), 4 deletions(-)

Comments

David Hildenbrand June 21, 2022, 7:44 a.m. UTC | #1
On 21.06.22 06:17, Oscar Salvador wrote:
> free_area_init_node() calls calculate_node_totalpages() and
> free_area_init_core(). The former to get node's {spanned,present}_pages,
> and the latter to calculate, among other things, how many pages per zone
> we spent on memmap_pages, which is used to substract zone's free pages.
> 
> On memoryless-nodes, it is pointless to perform such a bunch of work, so
> make sure we skip the calculations when having a node or empty zone.
> 
> Signed-off-by: Oscar Salvador <osalvador@suse.de>
> ---
>  mm/page_alloc.c | 16 ++++++++++++----
>  1 file changed, 12 insertions(+), 4 deletions(-)
> 
> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> index e008a3df0485..2b9b2422ba32 100644
> --- a/mm/page_alloc.c
> +++ b/mm/page_alloc.c
> @@ -7361,6 +7361,10 @@ static void __init calculate_node_totalpages(struct pglist_data *pgdat,
>  	unsigned long realtotalpages = 0, totalpages = 0;
>  	enum zone_type i;
>  
> +	/* Skip calculation for memoryless nodes */
> +	if (pgdat_is_empty(pgdat))
> +		goto no_pages;
> +
>  	for (i = 0; i < MAX_NR_ZONES; i++) {
>  		struct zone *zone = pgdat->node_zones + i;
>  		unsigned long zone_start_pfn, zone_end_pfn;
> @@ -7393,6 +7397,7 @@ static void __init calculate_node_totalpages(struct pglist_data *pgdat,
>  		realtotalpages += real_size;
>  	}
>  
> +no_pages:
>  	pgdat->node_spanned_pages = totalpages;
>  	pgdat->node_present_pages = realtotalpages;
>  	pr_debug("On node %d totalpages: %lu\n", pgdat->node_id, realtotalpages);
> @@ -7610,6 +7615,12 @@ static void __init free_area_init_core(struct pglist_data *pgdat)
>  		size = zone->spanned_pages;
>  		freesize = zone->present_pages;
>  
> +		/* No pages? Nothing to calculate then. */
> +		if (!size) {
> +			zone_init_internals(zone, j, nid, 0);
> +			continue;
> +		}
> +
>  		/*
>  		 * Adjust freesize so that it accounts for how much memory
>  		 * is used by this zone for memmap. This affects the watermark
> @@ -7647,9 +7658,6 @@ static void __init free_area_init_core(struct pglist_data *pgdat)
>  		 */
>  		zone_init_internals(zone, j, nid, freesize);
>  
> -		if (!size)
> -			continue;
> -
>  		set_pageblock_order();
>  		setup_usemap(zone);
>  		init_currently_empty_zone(zone, zone->zone_start_pfn, size);
> @@ -7730,7 +7738,7 @@ static void __init free_area_init_node(int nid)
>  	pgdat->node_start_pfn = start_pfn;
>  	pgdat->per_cpu_nodestats = NULL;
>  
> -	if (start_pfn != end_pfn) {
> +	if (!pgdat_is_empty(pgdat)) {
>  		pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
>  			(u64)start_pfn << PAGE_SHIFT,
>  			end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);


It's worth noting that the check in pgdat_is_empty() is slightly
different. I *think* it doesn't matter in practice, yet I wonder if we
should simply fixup (currently unused) pgdat_is_empty().

Anyhow

Reviewed-by: David Hildenbrand <david@redhat.com>
Oscar Salvador June 22, 2022, 3:47 a.m. UTC | #2
On Tue, Jun 21, 2022 at 09:44:47AM +0200, David Hildenbrand wrote:
> 
> 
> It's worth noting that the check in pgdat_is_empty() is slightly
> different. I *think* it doesn't matter in practice, yet I wonder if we
> should simply fixup (currently unused) pgdat_is_empty().

I guess we could change it to

 static inline bool pgdat_is_empty(pg_data_t *pgdat)
 {
	 return node_start_pfn(pgdat->node_id) == node_end_pfn(pgdat->node_id)
 }

? And maybe even rename it to to node_is_empty (not sure why but I tend to like
that more than pgdat) 

I could squeeze a "fixup" patch for that before this one. 

> 
> Anyhow
> 
> Reviewed-by: David Hildenbrand <david@redhat.com>

Thanks!
Muchun Song June 22, 2022, 3:56 a.m. UTC | #3
On Wed, Jun 22, 2022 at 05:47:22AM +0200, Oscar Salvador wrote:
> On Tue, Jun 21, 2022 at 09:44:47AM +0200, David Hildenbrand wrote:
> > 
> > 
> > It's worth noting that the check in pgdat_is_empty() is slightly
> > different. I *think* it doesn't matter in practice, yet I wonder if we
> > should simply fixup (currently unused) pgdat_is_empty().
> 
> I guess we could change it to
> 
>  static inline bool pgdat_is_empty(pg_data_t *pgdat)
>  {
> 	 return node_start_pfn(pgdat->node_id) == node_end_pfn(pgdat->node_id)
>  }
> 
> ? And maybe even rename it to to node_is_empty (not sure why but I tend to like

At least I like this name (node_is_empty) as well.

Thanks.

> that more than pgdat) 
> 
> I could squeeze a "fixup" patch for that before this one. 
> 
> > 
> > Anyhow
> > 
> > Reviewed-by: David Hildenbrand <david@redhat.com>
> 
> Thanks!
> 
> 
> -- 
> Oscar Salvador
> SUSE Labs
>
David Hildenbrand June 22, 2022, 8:31 a.m. UTC | #4
On 22.06.22 05:56, Muchun Song wrote:
> On Wed, Jun 22, 2022 at 05:47:22AM +0200, Oscar Salvador wrote:
>> On Tue, Jun 21, 2022 at 09:44:47AM +0200, David Hildenbrand wrote:
>>>
>>>
>>> It's worth noting that the check in pgdat_is_empty() is slightly
>>> different. I *think* it doesn't matter in practice, yet I wonder if we
>>> should simply fixup (currently unused) pgdat_is_empty().
>>
>> I guess we could change it to
>>
>>  static inline bool pgdat_is_empty(pg_data_t *pgdat)
>>  {
>> 	 return node_start_pfn(pgdat->node_id) == node_end_pfn(pgdat->node_id)
>>  }
>>
>> ? And maybe even rename it to to node_is_empty (not sure why but I tend to like
> 
> At least I like this name (node_is_empty) as well.
> 

Let's try keeping it consistent. I think node_is_empty() might indicate
that we're punching in a node id instead of a pgdat.
Muchun Song June 22, 2022, 8:54 a.m. UTC | #5
On Wed, Jun 22, 2022 at 10:31:12AM +0200, David Hildenbrand wrote:
> On 22.06.22 05:56, Muchun Song wrote:
> > On Wed, Jun 22, 2022 at 05:47:22AM +0200, Oscar Salvador wrote:
> >> On Tue, Jun 21, 2022 at 09:44:47AM +0200, David Hildenbrand wrote:
> >>>
> >>>
> >>> It's worth noting that the check in pgdat_is_empty() is slightly
> >>> different. I *think* it doesn't matter in practice, yet I wonder if we
> >>> should simply fixup (currently unused) pgdat_is_empty().
> >>
> >> I guess we could change it to
> >>
> >>  static inline bool pgdat_is_empty(pg_data_t *pgdat)
> >>  {
> >> 	 return node_start_pfn(pgdat->node_id) == node_end_pfn(pgdat->node_id)
> >>  }
> >>
> >> ? And maybe even rename it to to node_is_empty (not sure why but I tend to like
> > 
> > At least I like this name (node_is_empty) as well.
> > 
> 
> Let's try keeping it consistent. I think node_is_empty() might indicate
> that we're punching in a node id instead of a pgdat.
>

I suspect Oscar will change the argument to "nid" as well, like:

static inline bool node_is_empty(int nid)
{
	return node_start_pfn(nid) == node_end_pfn(nid);
}

Does this look good?

Thanks.
 
> 
> -- 
> Thanks,
> 
> David / dhildenb
> 
>
David Hildenbrand June 22, 2022, 10:49 a.m. UTC | #6
On 22.06.22 10:54, Muchun Song wrote:
> On Wed, Jun 22, 2022 at 10:31:12AM +0200, David Hildenbrand wrote:
>> On 22.06.22 05:56, Muchun Song wrote:
>>> On Wed, Jun 22, 2022 at 05:47:22AM +0200, Oscar Salvador wrote:
>>>> On Tue, Jun 21, 2022 at 09:44:47AM +0200, David Hildenbrand wrote:
>>>>>
>>>>>
>>>>> It's worth noting that the check in pgdat_is_empty() is slightly
>>>>> different. I *think* it doesn't matter in practice, yet I wonder if we
>>>>> should simply fixup (currently unused) pgdat_is_empty().
>>>>
>>>> I guess we could change it to
>>>>
>>>>  static inline bool pgdat_is_empty(pg_data_t *pgdat)
>>>>  {
>>>> 	 return node_start_pfn(pgdat->node_id) == node_end_pfn(pgdat->node_id)
>>>>  }
>>>>
>>>> ? And maybe even rename it to to node_is_empty (not sure why but I tend to like
>>>
>>> At least I like this name (node_is_empty) as well.
>>>
>>
>> Let's try keeping it consistent. I think node_is_empty() might indicate
>> that we're punching in a node id instead of a pgdat.
>>
> 
> I suspect Oscar will change the argument to "nid" as well, like:
> 
> static inline bool node_is_empty(int nid)
> {
> 	return node_start_pfn(nid) == node_end_pfn(nid);
> }
> 
> Does this look good?

Then we have to lookup the pgdat multiple times for (IMHO) no real
compelling reason.
diff mbox series

Patch

diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index e008a3df0485..2b9b2422ba32 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c
@@ -7361,6 +7361,10 @@  static void __init calculate_node_totalpages(struct pglist_data *pgdat,
 	unsigned long realtotalpages = 0, totalpages = 0;
 	enum zone_type i;
 
+	/* Skip calculation for memoryless nodes */
+	if (pgdat_is_empty(pgdat))
+		goto no_pages;
+
 	for (i = 0; i < MAX_NR_ZONES; i++) {
 		struct zone *zone = pgdat->node_zones + i;
 		unsigned long zone_start_pfn, zone_end_pfn;
@@ -7393,6 +7397,7 @@  static void __init calculate_node_totalpages(struct pglist_data *pgdat,
 		realtotalpages += real_size;
 	}
 
+no_pages:
 	pgdat->node_spanned_pages = totalpages;
 	pgdat->node_present_pages = realtotalpages;
 	pr_debug("On node %d totalpages: %lu\n", pgdat->node_id, realtotalpages);
@@ -7610,6 +7615,12 @@  static void __init free_area_init_core(struct pglist_data *pgdat)
 		size = zone->spanned_pages;
 		freesize = zone->present_pages;
 
+		/* No pages? Nothing to calculate then. */
+		if (!size) {
+			zone_init_internals(zone, j, nid, 0);
+			continue;
+		}
+
 		/*
 		 * Adjust freesize so that it accounts for how much memory
 		 * is used by this zone for memmap. This affects the watermark
@@ -7647,9 +7658,6 @@  static void __init free_area_init_core(struct pglist_data *pgdat)
 		 */
 		zone_init_internals(zone, j, nid, freesize);
 
-		if (!size)
-			continue;
-
 		set_pageblock_order();
 		setup_usemap(zone);
 		init_currently_empty_zone(zone, zone->zone_start_pfn, size);
@@ -7730,7 +7738,7 @@  static void __init free_area_init_node(int nid)
 	pgdat->node_start_pfn = start_pfn;
 	pgdat->per_cpu_nodestats = NULL;
 
-	if (start_pfn != end_pfn) {
+	if (!pgdat_is_empty(pgdat)) {
 		pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
 			(u64)start_pfn << PAGE_SHIFT,
 			end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);