Hello,
On Friday, February 03, 2012 3:04 PM Mel Gorman wrote:
On Fri, Feb 03, 2012 at 01:18:54PM +0100, Marek Szyprowski wrote:
alloc_contig_range() performs memory allocation so it also should keep track on keeping the correct level of memory watermarks. This commit adds a call to *_slowpath style reclaim to grab enough pages to make sure that the final collection of contiguous pages from freelists will not starve the system.
Signed-off-by: Marek Szyprowski m.szyprowski@samsung.com Signed-off-by: Kyungmin Park kyungmin.park@samsung.com CC: Michal Nazarewicz mina86@mina86.com Tested-by: Rob Clark rob.clark@linaro.org Tested-by: Ohad Ben-Cohen ohad@wizery.com Tested-by: Benjamin Gaignard benjamin.gaignard@linaro.org
I still do not intend to ack this patch and any damage is confined to CMA but I have a few comments anyway.
mm/page_alloc.c | 47 +++++++++++++++++++++++++++++++++++++++++++++++ 1 files changed, 47 insertions(+), 0 deletions(-)
diff --git a/mm/page_alloc.c b/mm/page_alloc.c index 983ccba..371a79f 100644 --- a/mm/page_alloc.c +++ b/mm/page_alloc.c @@ -5632,6 +5632,46 @@ static int __alloc_contig_migrate_range(unsigned long start, unsigned
long end)
return ret > 0 ? 0 : ret; }
+/*
- Trigger memory pressure bump to reclaim some pages in order to be able to
- allocate 'count' pages in single page units. Does similar work as
- *__alloc_pages_slowpath() function.
- */
+static int __reclaim_pages(struct zone *zone, gfp_t gfp_mask, int count) +{
- enum zone_type high_zoneidx = gfp_zone(gfp_mask);
- struct zonelist *zonelist = node_zonelist(0, gfp_mask);
- int did_some_progress = 0;
- int order = 1;
- unsigned long watermark;
- /*
* Increase level of watermarks to force kswapd do his job* to stabilize at new watermark level.*/- min_free_kbytes += count * PAGE_SIZE / 1024;
There is a risk of overflow here although it is incredibly small. Still, a potentially nicer way of doing this was
count << (PAGE_SHIFT - 10)
- setup_per_zone_wmarks();
Nothing prevents two or more processes updating the wmarks at the same time which is racy and unpredictable. Today it is not much of a problem but CMA makes this path hotter than it was and you may see weirdness if two processes are updating zonelists at the same time. Swap-over-NFS actually starts with a patch that serialises setup_per_zone_wmarks()
You also potentially have a BIG problem here if this happens
min_free_kbytes = 32768 Process a: min_free_kbytes += 65536 Process a: start direct reclaim echo 16374 > /proc/sys/vm/min_free_kbytes Process a: exit direct_reclaim Process a: min_free_kbytes -= 65536
min_free_kbytes now wraps negative and the machine hangs.
The damage is confined to CMA though so I am not going to lose sleep over it but you might want to consider at least preventing parallel updates to min_free_kbytes from proc.
Right. This approach was definitely too hacky. What do you think about replacing it with the following code (I assume that setup_per_zone_wmarks() serialization patch will be merged anyway so I skipped it here):
diff --git a/include/linux/mmzone.h b/include/linux/mmzone.h index 82f4fa5..bb9ae41 100644 --- a/include/linux/mmzone.h +++ b/include/linux/mmzone.h @@ -371,6 +371,13 @@ struct zone { /* see spanned/present_pages for more description */ seqlock_t span_seqlock; #endif +#ifdef CONFIG_CMA + /* + * CMA needs to increase watermark levels during the allocation + * process to make sure that the system is not starved. + */ + unsigned long min_cma_pages; +#endif struct free_area free_area[MAX_ORDER];
#ifndef CONFIG_SPARSEMEM diff --git a/mm/page_alloc.c b/mm/page_alloc.c index 824fb37..1ca52f0 100644 --- a/mm/page_alloc.c +++ b/mm/page_alloc.c @@ -5044,6 +5044,11 @@ void setup_per_zone_wmarks(void)
zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2); zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1); +#ifdef CONFIG_CMA + zone->watermark[WMARK_MIN] += zone->min_cma_pages; + zone->watermark[WMARK_LOW] += zone->min_cma_pages; + zone->watermark[WMARK_HIGH] += zone->min_cma_pages; +#endif setup_zone_migrate_reserve(zone); spin_unlock_irqrestore(&zone->lock, flags); } @@ -5625,13 +5630,15 @@ static int __reclaim_pages(struct zone *zone, gfp_t gfp_mask, int count) struct zonelist *zonelist = node_zonelist(0, gfp_mask); int did_some_progress = 0; int order = 1; - unsigned long watermark; + unsigned long watermark, flags;
/* * Increase level of watermarks to force kswapd do his job * to stabilize at new watermark level. */ - min_free_kbytes += count * PAGE_SIZE / 1024; + spin_lock_irqsave(&zone->lock, flags); + zone->min_cma_pages += count; + spin_unlock_irqrestore(&zone->lock, flags); setup_per_zone_wmarks();
/* Obey watermarks as if the page was being allocated */ @@ -5648,7 +5655,9 @@ static int __reclaim_pages(struct zone *zone, gfp_t gfp_mask, int count) }
/* Restore original watermark levels. */ - min_free_kbytes -= count * PAGE_SIZE / 1024; + spin_lock_irqsave(&zone->lock, flags); + zone->min_cma_pages -= count; + spin_unlock_irqrestore(&zone->lock, flags); setup_per_zone_wmarks();
return count;
Best regards