2014-02-13 02:41:09

by Weijie Yang

[permalink] [raw]
Subject: [PATCH 1/2] mm/vmscan: restore sc->gfp_mask after promoting it to __GFP_HIGHMEM

We promote sc->gfp_mask to __GFP_HIGHMEM to forcibly scan highmem if
there are too many buffer_heads pinning highmem. see: cc715d99e5

This patch restores sc->gfp_mask to its caller original value after
finishing the scan job, to avoid the impact on other invocations from
its upper caller, such as vmpressure_prio(), shrink_slab().

Signed-off-by: Weijie Yang <[email protected]>
---
mm/vmscan.c | 8 +++++++-
1 file changed, 7 insertions(+), 1 deletion(-)
mode change 100644 => 100755 mm/vmscan.c

diff --git a/mm/vmscan.c b/mm/vmscan.c
old mode 100644
new mode 100755
index a9c74b4..35879f0
--- a/mm/vmscan.c
+++ b/mm/vmscan.c
@@ -2298,14 +2298,17 @@ static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
unsigned long nr_soft_reclaimed;
unsigned long nr_soft_scanned;
bool aborted_reclaim = false;
+ bool promoted_mask = false;

/*
* If the number of buffer_heads in the machine exceeds the maximum
* allowed level, force direct reclaim to scan the highmem zone as
* highmem pages could be pinning lowmem pages storing buffer_heads
*/
- if (buffer_heads_over_limit)
+ if (buffer_heads_over_limit) {
+ promoted_mask = !(sc->gfp_mask & __GFP_HIGHMEM);
sc->gfp_mask |= __GFP_HIGHMEM;
+ }

for_each_zone_zonelist_nodemask(zone, z, zonelist,
gfp_zone(sc->gfp_mask), sc->nodemask) {
@@ -2354,6 +2357,9 @@ static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
shrink_zone(zone, sc);
}

+ if (promoted_mask)
+ sc->gfp_mask &= ~__GFP_HIGHMEM;
+
return aborted_reclaim;
}

--
1.7.10.4


2014-02-13 16:57:55

by Rik van Riel

[permalink] [raw]
Subject: Re: [PATCH 1/2] mm/vmscan: restore sc->gfp_mask after promoting it to __GFP_HIGHMEM

On 02/12/2014 09:39 PM, Weijie Yang wrote:

> --- a/mm/vmscan.c
> +++ b/mm/vmscan.c
> @@ -2298,14 +2298,17 @@ static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
> unsigned long nr_soft_reclaimed;
> unsigned long nr_soft_scanned;
> bool aborted_reclaim = false;
> + bool promoted_mask = false;
>
> /*
> * If the number of buffer_heads in the machine exceeds the maximum
> * allowed level, force direct reclaim to scan the highmem zone as
> * highmem pages could be pinning lowmem pages storing buffer_heads
> */
> - if (buffer_heads_over_limit)
> + if (buffer_heads_over_limit) {

It took me a minute to figure out why you are doing things this way,
so maybe this could use a comment, or maybe it could be done in a
simpler way, by simply saving and restoring the original mask?

orig_mask = sc->gfp_mask;

> + promoted_mask = !(sc->gfp_mask & __GFP_HIGHMEM);
> sc->gfp_mask |= __GFP_HIGHMEM;
> + }
>
> for_each_zone_zonelist_nodemask(zone, z, zonelist,
> gfp_zone(sc->gfp_mask), sc->nodemask) {
> @@ -2354,6 +2357,9 @@ static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
> shrink_zone(zone, sc);
> }
>
> + if (promoted_mask)
sc->gfp_mask = orig_mask;

> + sc->gfp_mask &= ~__GFP_HIGHMEM;
> +
> return aborted_reclaim;
> }
>
>


--
All rights reversed

2014-02-14 03:39:47

by Weijie Yang

[permalink] [raw]
Subject: Re: [PATCH 1/2] mm/vmscan: restore sc->gfp_mask after promoting it to __GFP_HIGHMEM

On Fri, Feb 14, 2014 at 12:10 AM, Rik van Riel <[email protected]> wrote:
> On 02/12/2014 09:39 PM, Weijie Yang wrote:
>
>> --- a/mm/vmscan.c
>> +++ b/mm/vmscan.c
>> @@ -2298,14 +2298,17 @@ static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
>> unsigned long nr_soft_reclaimed;
>> unsigned long nr_soft_scanned;
>> bool aborted_reclaim = false;
>> + bool promoted_mask = false;
>>
>> /*
>> * If the number of buffer_heads in the machine exceeds the maximum
>> * allowed level, force direct reclaim to scan the highmem zone as
>> * highmem pages could be pinning lowmem pages storing buffer_heads
>> */
>> - if (buffer_heads_over_limit)
>> + if (buffer_heads_over_limit) {
>
> It took me a minute to figure out why you are doing things this way,
> so maybe this could use a comment, or maybe it could be done in a
> simpler way, by simply saving and restoring the original mask?
> orig_mask = sc->gfp_mask;

Yes, you are right. This simpler way is better. I will turn to it in
V2 resend patch.

Thanks!

>> + promoted_mask = !(sc->gfp_mask & __GFP_HIGHMEM);
>> sc->gfp_mask |= __GFP_HIGHMEM;
>> + }
>>
>> for_each_zone_zonelist_nodemask(zone, z, zonelist,
>> gfp_zone(sc->gfp_mask), sc->nodemask) {
>> @@ -2354,6 +2357,9 @@ static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
>> shrink_zone(zone, sc);
>> }
>>
>> + if (promoted_mask)
> sc->gfp_mask = orig_mask;
>
>> + sc->gfp_mask &= ~__GFP_HIGHMEM;
>> +
>> return aborted_reclaim;
>> }
>>
>>
>
>
> --
> All rights reversed