[PATCH] mm, compaction: avoid uninitialized variable use

[PATCH] mm, compaction: avoid uninitialized variable use

[Arnd Bergmann]

A recent rework of the compaction code introduced a warning about
an uninitialized variable when CONFIG_COMPACTION is disabled and
__alloc_pages_direct_compact() does not set its 'compact_result'
output argument:

mm/page_alloc.c: In function '__alloc_pages_nodemask':
mm/page_alloc.c:3651:6: error: 'compact_result' may be used uninitialized in this function [-Werror=maybe-uninitialized]

This adds another check for CONFIG_COMPACTION to ensure we never
evaluate the uninitialized variable in this configuration, which
is probably the simplest way to avoid the warning.

A more elaborate rework might make this more readable.

Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Fixes: 13cff7b81275 ("mm, compaction: simplify __alloc_pages_direct_compact feedback interface")
---
 mm/page_alloc.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index 477d9382f70d..bedadc686a22 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c
@@ -3640,7 +3640,8 @@ retry:
 		goto got_pg;
 
 	/* Checks for THP-specific high-order allocations */
-	if (is_thp_gfp_mask(gfp_mask)) {
+	if (IS_ENABLED(CONFIG_COMPACTION) &&
+		is_thp_gfp_mask(gfp_mask)) {
 		/*
 		 * If compaction is deferred for high-order allocations, it is
 		 * because sync compaction recently failed. If this is the case
-- 
2.7.0

Re: [PATCH] mm, compaction: avoid uninitialized variable use

[Michal Hocko]

On Wed 11-05-16 15:24:44, Arnd Bergmann wrote:
> A recent rework of the compaction code introduced a warning about
> an uninitialized variable when CONFIG_COMPACTION is disabled and
> __alloc_pages_direct_compact() does not set its 'compact_result'
> output argument:
> 
> mm/page_alloc.c: In function '__alloc_pages_nodemask':
> mm/page_alloc.c:3651:6: error: 'compact_result' may be used uninitialized in this function [-Werror=maybe-uninitialized]
> 
> This adds another check for CONFIG_COMPACTION to ensure we never
> evaluate the uninitialized variable in this configuration, which
> is probably the simplest way to avoid the warning.

I think that hiding this into __alloc_pages_direct_compact is a better
idea. See the diff below
> 
> A more elaborate rework might make this more readable.
> 
> Signed-off-by: Arnd Bergmann <arnd@arndb.de>
> Fixes: 13cff7b81275 ("mm, compaction: simplify __alloc_pages_direct_compact feedback interface")

Please do not use SHA for mmotm commits because they are unstable and
change each linux-next release.

--- 
diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index 4950d01ff935..14e3b4d93adc 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c
@@ -3300,6 +3300,7 @@ __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
 		unsigned int alloc_flags, const struct alloc_context *ac,
 		enum migrate_mode mode, enum compact_result *compact_result)
 {
+	*compact_result = COMPACT_DEFERRED;
 	return NULL;
 }
 
-- 
Michal Hocko
SUSE Labs

Re: [PATCH] mm, compaction: avoid uninitialized variable use

[Arnd Bergmann]

On Wednesday 11 May 2016 16:44:07 Michal Hocko wrote:
> On Wed 11-05-16 15:24:44, Arnd Bergmann wrote:
> > A recent rework of the compaction code introduced a warning about
> > an uninitialized variable when CONFIG_COMPACTION is disabled and
> > __alloc_pages_direct_compact() does not set its 'compact_result'
> > output argument:
> > 
> > mm/page_alloc.c: In function '__alloc_pages_nodemask':
> > mm/page_alloc.c:3651:6: error: 'compact_result' may be used uninitialized in this function [-Werror=maybe-uninitialized]
> > 
> > This adds another check for CONFIG_COMPACTION to ensure we never
> > evaluate the uninitialized variable in this configuration, which
> > is probably the simplest way to avoid the warning.
> 
> I think that hiding this into __alloc_pages_direct_compact is a better
> idea. See the diff below

Ok, sounds good.

> --- 
> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> index 4950d01ff935..14e3b4d93adc 100644
> --- a/mm/page_alloc.c
> +++ b/mm/page_alloc.c
> @@ -3300,6 +3300,7 @@ __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
>                 unsigned int alloc_flags, const struct alloc_context *ac,
>                 enum migrate_mode mode, enum compact_result *compact_result)
>  {
> +       *compact_result = COMPACT_DEFERRED;
>         return NULL;
>  }
> 

I thought about this but didn't know which COMPACT_* value was appropriate here.

The behavior then changes a bit with your approach compared to mine,
because 

                if (compact_result == COMPACT_DEFERRED)
                        goto nopage;

is true now. I assume this is what we want though.

	Arnd

Re: [PATCH] mm, compaction: avoid uninitialized variable use

[Michal Hocko]

On Wed 11-05-16 16:44:07, Michal Hocko wrote:
> On Wed 11-05-16 15:24:44, Arnd Bergmann wrote:
> > A recent rework of the compaction code introduced a warning about
> > an uninitialized variable when CONFIG_COMPACTION is disabled and
> > __alloc_pages_direct_compact() does not set its 'compact_result'
> > output argument:
> > 
> > mm/page_alloc.c: In function '__alloc_pages_nodemask':
> > mm/page_alloc.c:3651:6: error: 'compact_result' may be used uninitialized in this function [-Werror=maybe-uninitialized]
> > 
> > This adds another check for CONFIG_COMPACTION to ensure we never
> > evaluate the uninitialized variable in this configuration, which
> > is probably the simplest way to avoid the warning.
> 
> I think that hiding this into __alloc_pages_direct_compact is a better
> idea. See the diff below
> > 
> > A more elaborate rework might make this more readable.
> > 
> > Signed-off-by: Arnd Bergmann <arnd@arndb.de>
> > Fixes: 13cff7b81275 ("mm, compaction: simplify __alloc_pages_direct_compact feedback interface")
> 
> Please do not use SHA for mmotm commits because they are unstable and
> change each linux-next release.
> 
> --- 
> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> index 4950d01ff935..14e3b4d93adc 100644
> --- a/mm/page_alloc.c
> +++ b/mm/page_alloc.c
> @@ -3300,6 +3300,7 @@ __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
>  		unsigned int alloc_flags, const struct alloc_context *ac,
>  		enum migrate_mode mode, enum compact_result *compact_result)
>  {
> +	*compact_result = COMPACT_DEFERRED;

Sorry, this should have been COMPACT_SKIPPED.

-- 
Michal Hocko
SUSE Labs

Re: [PATCH] mm, compaction: avoid uninitialized variable use

[Michal Hocko]

On Wed 11-05-16 16:52:41, Arnd Bergmann wrote:
> On Wednesday 11 May 2016 16:44:07 Michal Hocko wrote:
> > On Wed 11-05-16 15:24:44, Arnd Bergmann wrote:
> > > A recent rework of the compaction code introduced a warning about
> > > an uninitialized variable when CONFIG_COMPACTION is disabled and
> > > __alloc_pages_direct_compact() does not set its 'compact_result'
> > > output argument:
> > > 
> > > mm/page_alloc.c: In function '__alloc_pages_nodemask':
> > > mm/page_alloc.c:3651:6: error: 'compact_result' may be used uninitialized in this function [-Werror=maybe-uninitialized]
> > > 
> > > This adds another check for CONFIG_COMPACTION to ensure we never
> > > evaluate the uninitialized variable in this configuration, which
> > > is probably the simplest way to avoid the warning.
> > 
> > I think that hiding this into __alloc_pages_direct_compact is a better
> > idea. See the diff below
> 
> Ok, sounds good.
> 
> > --- 
> > diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> > index 4950d01ff935..14e3b4d93adc 100644
> > --- a/mm/page_alloc.c
> > +++ b/mm/page_alloc.c
> > @@ -3300,6 +3300,7 @@ __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
> >                 unsigned int alloc_flags, const struct alloc_context *ac,
> >                 enum migrate_mode mode, enum compact_result *compact_result)
> >  {
> > +       *compact_result = COMPACT_DEFERRED;
> >         return NULL;
> >  }
> > 
> 
> I thought about this but didn't know which COMPACT_* value was appropriate here.
> 
> The behavior then changes a bit with your approach compared to mine,
> because 
> 
>                 if (compact_result == COMPACT_DEFERRED)
>                         goto nopage;
> 
> is true now. I assume this is what we want though.

No, we actually want to try to reclaim, that's why COMPACT_SKIPPED is
more appropriate. COMPACT_SKIPPED is even correct semantically because
the compaction wasn't invoked.

-- 
Michal Hocko
SUSE Labs

Re: [PATCH] mm, compaction: avoid uninitialized variable use

[Michal Hocko]

I think this would be slightly better than your proposal. Andrew, could
you fold it into the original
mm-compaction-simplify-__alloc_pages_direct_compact-feedback-interface.patch
patch?
---
>From 434bc8b6f3787724327499998c4fe651e8ce5d68 Mon Sep 17 00:00:00 2001
From: Michal Hocko <mhocko@suse.com>
Date: Thu, 12 May 2016 08:10:33 +0200
Subject: [PATCH] mmotm:
 mm-compaction-simplify-__alloc_pages_direct_compact-feedback-interface-fix

Arnd has reported the following compilation warning:
mm/page_alloc.c: In function '__alloc_pages_nodemask':
mm/page_alloc.c:3651:6: error: 'compact_result' may be used uninitialized in this function [-Werror=maybe-uninitialized]

This should be a false positive TRANSPARENT_HUGEPAGE depends on COMPACTION
so is_thp_gfp_mask shouldn't be true. GFP_TRANSHUGE is a bit tricky
and somebody might be using this accidently. Make sure that compact_result
is defined also for !CONFIG_COMPACT and set it to COMPACT_SKIPPED because
the compaction was really withdrawn.

Reported-by: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Michal Hocko <mhocko@suse.com>
---
 mm/page_alloc.c | 1 +
 1 file changed, 1 insertion(+)

diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index 4950d01ff935..0d9008042efa 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c
@@ -3300,6 +3300,7 @@ __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
 		unsigned int alloc_flags, const struct alloc_context *ac,
 		enum migrate_mode mode, enum compact_result *compact_result)
 {
+	*compact_result = COMPACT_SKIPPED;
 	return NULL;
 }
 
-- 
2.8.1

-- 
Michal Hocko
SUSE Labs

Re: [PATCH] mm, compaction: avoid uninitialized variable use

[Vlastimil Babka]

On 05/12/2016 08:16 AM, Michal Hocko wrote:
> I think this would be slightly better than your proposal. Andrew, could
> you fold it into the original
> mm-compaction-simplify-__alloc_pages_direct_compact-feedback-interface.patch
> patch?
> ---
>  From 434bc8b6f3787724327499998c4fe651e8ce5d68 Mon Sep 17 00:00:00 2001
> From: Michal Hocko <mhocko@suse.com>
> Date: Thu, 12 May 2016 08:10:33 +0200
> Subject: [PATCH] mmotm:
>   mm-compaction-simplify-__alloc_pages_direct_compact-feedback-interface-fix
>
> Arnd has reported the following compilation warning:
> mm/page_alloc.c: In function '__alloc_pages_nodemask':
> mm/page_alloc.c:3651:6: error: 'compact_result' may be used uninitialized in this function [-Werror=maybe-uninitialized]
>
> This should be a false positive TRANSPARENT_HUGEPAGE depends on COMPACTION
> so is_thp_gfp_mask shouldn't be true. GFP_TRANSHUGE is a bit tricky
> and somebody might be using this accidently. Make sure that compact_result
> is defined also for !CONFIG_COMPACT and set it to COMPACT_SKIPPED because
> the compaction was really withdrawn.
>
> Reported-by: Arnd Bergmann <arnd@arndb.de>
> Signed-off-by: Michal Hocko <mhocko@suse.com>

Ack.

> ---
>   mm/page_alloc.c | 1 +
>   1 file changed, 1 insertion(+)
>
> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> index 4950d01ff935..0d9008042efa 100644
> --- a/mm/page_alloc.c
> +++ b/mm/page_alloc.c
> @@ -3300,6 +3300,7 @@ __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
>   		unsigned int alloc_flags, const struct alloc_context *ac,
>   		enum migrate_mode mode, enum compact_result *compact_result)
>   {
> +	*compact_result = COMPACT_SKIPPED;
>   	return NULL;
>   }
>
>

Re: [PATCH] mm, compaction: avoid uninitialized variable use

[Arnd Bergmann]

On Thursday 12 May 2016 08:16:36 Michal Hocko wrote:
> I think this would be slightly better than your proposal. Andrew, could
> you fold it into the original
> mm-compaction-simplify-__alloc_pages_direct_compact-feedback-interface.patch
> patch?
> ---
> From 434bc8b6f3787724327499998c4fe651e8ce5d68 Mon Sep 17 00:00:00 2001
> From: Michal Hocko <mhocko@suse.com>
> Date: Thu, 12 May 2016 08:10:33 +0200
> Subject: [PATCH] mmotm:
>  mm-compaction-simplify-__alloc_pages_direct_compact-feedback-interface-fix
> 
> Arnd has reported the following compilation warning:
> mm/page_alloc.c: In function '__alloc_pages_nodemask':
> mm/page_alloc.c:3651:6: error: 'compact_result' may be used uninitialized in this function [-Werror=maybe-uninitialized]
> 
> This should be a false positive TRANSPARENT_HUGEPAGE depends on COMPACTION
> so is_thp_gfp_mask shouldn't be true. GFP_TRANSHUGE is a bit tricky
> and somebody might be using this accidently. Make sure that compact_result
> is defined also for !CONFIG_COMPACT and set it to COMPACT_SKIPPED because
> the compaction was really withdrawn.
> 
> Reported-by: Arnd Bergmann <arnd@arndb.de>
> Signed-off-by: Michal Hocko <mhocko@suse.com>
> 

Acked-by: Arnd Bergmann <arnd@arndb.de>

Looks much nicer than my version.

	Arnd