mm: page_alloc: avoid marking zones full prematurely after zone_reclaim()
Hello Mel Gorman,
The patch 290d1a3ce0ec: "mm: page_alloc: avoid marking zones full prematurely after zone_reclaim()" from Mar 23, 2013, leads to the following warning:
"mm/page_alloc.c:1957 get_page_from_freelist()
warn: bitwise AND condition is false here"
mm/page_alloc.c
1948 /*
1949 * Failed to reclaim enough to meet watermark.
1950 * Only mark the zone full if checking the min
1951 * watermark or if we failed to reclaim just
1952 * 1<<order pages or else the page allocator
1953 * fastpath will prematurely mark zones full
1954 * when the watermark is between the low and
1955 * min watermarks.
1956 */
1957 if ((alloc_flags & ALLOC_WMARK_MIN) ||
^^^^^^^^^^^^^^^This is zero.
1958 ret == ZONE_RECLAIM_SOME)
1959 goto this_zone_full;
[snip]
2333 static inline int
2334 gfp_to_alloc_flags(gfp_t gfp_mask)
2335 {
2336 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
^^^^^^^^^^^^^^^ 2337 const gfp_t wait = gfp_mask & __GFP_WAIT;
regards,
dan carpenter
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>
[Adding Andrew into CC]
Hi Dan,
Quoted 7 lines · Dan Carpenter
Hello Mel Gorman,
The patch
290d1a3ce0ec: "mm:page_alloc: avoid marking zones full prematurely afterzone_reclaim()" from Mar 23, 2013, leads to the following warning:
"mm/page_alloc.c:1957get_page_from_freelist()
warn: bitwise AND condition is false here"
Dohh, I have totally missed this during review and I managed to burn myself on the similar issue in the past (gfp & GFP_NOWAIT). The follow up fix is bellow
Quoted 17 lines
mm/page_alloc.c
1948 /*1949 * Failed to reclaim enough to meet watermark. 1950 * Only mark the zone full if checking the min 1951 * watermark or if we failed to reclaim just 1952 * 1<<order pages or else the page allocator 1953 * fastpath will prematurely mark zones full 1954 * when the watermark is between the low and 1955 * min watermarks. 1956 */ 1957 if ((alloc_flags & ALLOC_WMARK_MIN) || ^^^^^^^^^^^^^^^This is zero.
1958 ret ==
ZONE_RECLAIM_SOME)
1959 gotothis_zone_full;
Quoted 8 lines
[snip]
2333 static inline int 2334 gfp_to_alloc_flags(gfp_t gfp_mask) 2335 { 2336 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET; ^^^^^^^^^^^^^^^2337 const
gfp_twait =gfp_mask&__GFP_WAIT;
---
And just for record. I wanted to give credit to Dan for reporting this but I have no idea how when this gets merged into the original patch which is still sitting in the -mm queue. I will leave that to Andrew ;)
Quoted 40 lines · Michal Hocko
From
b60e75c65b855a0df827a28a509e6761b4cf45ddMon Sep 17 00:00:00 2001
From: Michal Hocko <[email protected]>
Date: Wed, 27 Mar 2013 17:53:09 +0100
Subject: [PATCH] mm:
mm-page_alloc-avoid-marking-zones-full-prematurely-after-zone_reclaim-fixDan Carpenter has reported that (
alloc_flags&ALLOC_WMARK_MIN) test doesn't make much sense as the flag is 0 and it is in fact intended for wmark indexing rather than being used as a flag.Signed-off-by Michal Hocko <[email protected]>--- mm/page_alloc.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)mm/page_alloc.c +1
-1@@ -1953,7 +1953,7 @@ zonelist_scan:* when the watermark is between the low and * min watermarks. */ -if ((alloc_flags &ALLOC_WMARK_MIN ) || +if (((alloc_flags &ALLOC_WMARK_MASK ) ==ALLOC_WMARK_MIN ) || ret ==ZONE_RECLAIM_SOME ) goto this_zone_full;Signature · 11 lines
1.7.10.4 -- Michal Hocko SUSE Labs -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"[email protected]"> [email protected] </a>
Signature · 8 lines
Michal Hocko SUSE Labs -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"[email protected]"> [email protected] </a>
I still don't understand the code in gfp_to_alloc_flags().
int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
ORing with zero is odd.
regards,
dan carpenter
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>
Quoted 5 lines · Dan Carpenter
I still don't understand the code in
gfp_to_alloc_flags().int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;ORing with zero is odd.
It is a named zero ;) Something similar like GFP_NOWAIT AFAIU.
Signature · 8 lines
Michal Hocko SUSE Labs -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"[email protected]"> [email protected] </a>
Quoted 6 lines · Dan Carpenter
I still don't understand the code in
gfp_to_alloc_flags().int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;ORing with zero is odd.
Thanks Dan for the report and thanks Michal for fixing it. I was offline for last week which lead to my tardy response.
The odditiy is that ALLOC_WMARK_MIN is not treated as a flag but as an offset within the zone->wmark so it starts as 0. It could have been written as
int alloc_flags = ALLOC_CPUSET;
but then it would be easy to forget that in this path we are using the MIN watermark.
The "flag" is used as an offset because it eliminated a number of branches in the page allocator and was a micro-optimisation at the time.
Signature · 8 lines
Mel Gorman SUSE Labs -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"[email protected]"> [email protected] </a>