[PATCH 1/2] mm/slub: Fix cmp_loc_by_count() to return 0 when counts are equal

[PATCH 1/2] mm/slub: Fix cmp_loc_by_count() to return 0 when counts are equal

[Kuan-Wei Chiu]

The comparison function cmp_loc_by_count() used for sorting stack trace
locations in debugfs currently returns -1 if a->count > b->count and 1
otherwise. This breaks the antisymmetry property required by sort(),
because when two counts are equal, both cmp(a, b) and cmp(b, a) return
1.

This can lead to undefined or incorrect ordering results. Fix it by
explicitly returning 0 when the counts are equal, ensuring that the
comparison function follows the expected mathematical properties.

Fixes: 553c0369b3e1 ("mm/slub: sort debugfs output by frequency of stack traces")
Cc: stable@vger.kernel.org
Signed-off-by: Kuan-Wei Chiu <visitorckw@gmail.com>
---
 mm/slub.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/mm/slub.c b/mm/slub.c
index 30003763d224..c91b3744adbc 100644
--- a/mm/slub.c
+++ b/mm/slub.c
@@ -7718,8 +7718,9 @@ static int cmp_loc_by_count(const void *a, const void *b, const void *data)
 
 	if (loc1->count > loc2->count)
 		return -1;
-	else
+	if (loc1->count < loc2->count)
 		return 1;
+	return 0;
 }
 
 static void *slab_debugfs_start(struct seq_file *seq, loff_t *ppos)
-- 
2.34.1

Re: [PATCH 1/2] mm/slub: Fix cmp_loc_by_count() to return 0 when counts are equal

[Joshua Hahn]

On Mon, 25 Aug 2025 09:34:18 +0800 Kuan-Wei Chiu <visitorckw@gmail.com> wrote:

> The comparison function cmp_loc_by_count() used for sorting stack trace
> locations in debugfs currently returns -1 if a->count > b->count and 1
> otherwise. This breaks the antisymmetry property required by sort(),
> because when two counts are equal, both cmp(a, b) and cmp(b, a) return
> 1.
> 
> This can lead to undefined or incorrect ordering results. Fix it by
> explicitly returning 0 when the counts are equal, ensuring that the
> comparison function follows the expected mathematical properties.
> 
> Fixes: 553c0369b3e1 ("mm/slub: sort debugfs output by frequency of stack traces")
> Cc: stable@vger.kernel.org
> Signed-off-by: Kuan-Wei Chiu <visitorckw@gmail.com>
> ---
>  mm/slub.c | 3 ++-
>  1 file changed, 2 insertions(+), 1 deletion(-)
> 
> diff --git a/mm/slub.c b/mm/slub.c
> index 30003763d224..c91b3744adbc 100644
> --- a/mm/slub.c
> +++ b/mm/slub.c
> @@ -7718,8 +7718,9 @@ static int cmp_loc_by_count(const void *a, const void *b, const void *data)
>  
>  	if (loc1->count > loc2->count)
>  		return -1;
> -	else
> +	if (loc1->count < loc2->count)
>  		return 1;
> +	return 0;
>  }

Hello Kuan-Wei,

This is a great catch! I was thinking that in addition to separating out the
== case, we can also simplify the behavior by just opting to use the
cmp_int macro, which is defined in the <linux/sort.h> header, which is
already included in mm/slub.c. For the description, we have:

 * Return: 1 if the left argument is greater than the right one; 0 if the
 * arguments are equal; -1 if the left argument is less than the right one.

So in this case, we can replace the entire code block above with:

return cmp_int(loc2->count, loc1->count);

or

return -1 * cmp_int(loc1->count, loc2->count);

if you prefer to keep the position of loc1 and loc2. I guess we do lose
some interpretability of what -1 and 1 would refer to here, but I think
a comment should be able to take care of that.

Please let me know what you think. I hope you have a great day!
Joshua

Re: [PATCH 1/2] mm/slub: Fix cmp_loc_by_count() to return 0 when counts are equal

[Kuan-Wei Chiu]

Hi Joshua,

On Mon, Aug 25, 2025 at 07:48:36AM -0700, Joshua Hahn wrote:
> On Mon, 25 Aug 2025 09:34:18 +0800 Kuan-Wei Chiu <visitorckw@gmail.com> wrote:
> 
> > The comparison function cmp_loc_by_count() used for sorting stack trace
> > locations in debugfs currently returns -1 if a->count > b->count and 1
> > otherwise. This breaks the antisymmetry property required by sort(),
> > because when two counts are equal, both cmp(a, b) and cmp(b, a) return
> > 1.
> > 
> > This can lead to undefined or incorrect ordering results. Fix it by
> > explicitly returning 0 when the counts are equal, ensuring that the
> > comparison function follows the expected mathematical properties.
> > 
> > Fixes: 553c0369b3e1 ("mm/slub: sort debugfs output by frequency of stack traces")
> > Cc: stable@vger.kernel.org
> > Signed-off-by: Kuan-Wei Chiu <visitorckw@gmail.com>
> > ---
> >  mm/slub.c | 3 ++-
> >  1 file changed, 2 insertions(+), 1 deletion(-)
> > 
> > diff --git a/mm/slub.c b/mm/slub.c
> > index 30003763d224..c91b3744adbc 100644
> > --- a/mm/slub.c
> > +++ b/mm/slub.c
> > @@ -7718,8 +7718,9 @@ static int cmp_loc_by_count(const void *a, const void *b, const void *data)
> >  
> >  	if (loc1->count > loc2->count)
> >  		return -1;
> > -	else
> > +	if (loc1->count < loc2->count)
> >  		return 1;
> > +	return 0;
> >  }
> 
> Hello Kuan-Wei,
> 
> This is a great catch! I was thinking that in addition to separating out the
> == case, we can also simplify the behavior by just opting to use the
> cmp_int macro, which is defined in the <linux/sort.h> header, which is
> already included in mm/slub.c. For the description, we have:
> 
>  * Return: 1 if the left argument is greater than the right one; 0 if the
>  * arguments are equal; -1 if the left argument is less than the right one.
> 
> So in this case, we can replace the entire code block above with:
> 
> return cmp_int(loc2->count, loc1->count);
> 
> or
> 
> return -1 * cmp_int(loc1->count, loc2->count);
> 
> if you prefer to keep the position of loc1 and loc2. I guess we do lose
> some interpretability of what -1 and 1 would refer to here, but I think
> a comment should be able to take care of that.
> 
> Please let me know what you think. I hope you have a great day!
> Joshua

Thanks for the suggestion!
If we're going with the cmp_int() macro, I personally prefer
return cmp_int(loc2->count, loc1->count);
this avoids the need to explain the extra * (-1), and I think cmp_int()
is simple enough to be easily understood by readers.
That said, both options work fine for me.

Regards,
Kuan-Wei

Re: [PATCH 1/2] mm/slub: Fix cmp_loc_by_count() to return 0 when counts are equal

[Joshua Hahn]

On Tue, 26 Aug 2025 00:18:42 +0800 Kuan-Wei Chiu <visitorckw@gmail.com> wrote:

> Hi Joshua,
> 
> On Mon, Aug 25, 2025 at 07:48:36AM -0700, Joshua Hahn wrote:
> > On Mon, 25 Aug 2025 09:34:18 +0800 Kuan-Wei Chiu <visitorckw@gmail.com> wrote:
> > 
> > > The comparison function cmp_loc_by_count() used for sorting stack trace
> > > locations in debugfs currently returns -1 if a->count > b->count and 1
> > > otherwise. This breaks the antisymmetry property required by sort(),
> > > because when two counts are equal, both cmp(a, b) and cmp(b, a) return
> > > 1.
> > > 
> > > This can lead to undefined or incorrect ordering results. Fix it by
> > > explicitly returning 0 when the counts are equal, ensuring that the
> > > comparison function follows the expected mathematical properties.
> > > 
> > > Fixes: 553c0369b3e1 ("mm/slub: sort debugfs output by frequency of stack traces")
> > > Cc: stable@vger.kernel.org
> > > Signed-off-by: Kuan-Wei Chiu <visitorckw@gmail.com>
> > > ---
> > >  mm/slub.c | 3 ++-
> > >  1 file changed, 2 insertions(+), 1 deletion(-)
> > > 
> > > diff --git a/mm/slub.c b/mm/slub.c
> > > index 30003763d224..c91b3744adbc 100644
> > > --- a/mm/slub.c
> > > +++ b/mm/slub.c
> > > @@ -7718,8 +7718,9 @@ static int cmp_loc_by_count(const void *a, const void *b, const void *data)
> > >  
> > >  	if (loc1->count > loc2->count)
> > >  		return -1;
> > > -	else
> > > +	if (loc1->count < loc2->count)
> > >  		return 1;
> > > +	return 0;
> > >  }
> > 
> > Hello Kuan-Wei,
> > 
> > This is a great catch! I was thinking that in addition to separating out the
> > == case, we can also simplify the behavior by just opting to use the
> > cmp_int macro, which is defined in the <linux/sort.h> header, which is
> > already included in mm/slub.c. For the description, we have:
> > 
> >  * Return: 1 if the left argument is greater than the right one; 0 if the
> >  * arguments are equal; -1 if the left argument is less than the right one.
> > 
> > So in this case, we can replace the entire code block above with:
> > 
> > return cmp_int(loc2->count, loc1->count);
> > 
> > or
> > 
> > return -1 * cmp_int(loc1->count, loc2->count);
> > 
> > if you prefer to keep the position of loc1 and loc2. I guess we do lose
> > some interpretability of what -1 and 1 would refer to here, but I think
> > a comment should be able to take care of that.
> > 
> > Please let me know what you think. I hope you have a great day!
> > Joshua
> 
> Thanks for the suggestion!
> If we're going with the cmp_int() macro, I personally prefer
> return cmp_int(loc2->count, loc1->count);

Makes sense with me, please feel free to add my reviewed-by tag as well!
Have a great day!
Joshua

Reviewed-by: Joshua Hahn <joshua.hahnjy@gmail.com>

Re: [PATCH 1/2] mm/slub: Fix cmp_loc_by_count() to return 0 when counts are equal

[Vlastimil Babka]

On 8/25/25 03:34, Kuan-Wei Chiu wrote:
> The comparison function cmp_loc_by_count() used for sorting stack trace
> locations in debugfs currently returns -1 if a->count > b->count and 1
> otherwise. This breaks the antisymmetry property required by sort(),
> because when two counts are equal, both cmp(a, b) and cmp(b, a) return
> 1.

Good catch.

> This can lead to undefined or incorrect ordering results. Fix it by

Wonder if it can really affect anything in practice other than swapping
needlessly some records with an equal count?

> explicitly returning 0 when the counts are equal, ensuring that the
> comparison function follows the expected mathematical properties.

Agreed with the cmp_int() suggestion for a v2.

> Fixes: 553c0369b3e1 ("mm/slub: sort debugfs output by frequency of stack traces")
> Cc: stable@vger.kernel.org

I don't think it can cause any serious bugs so Cc: stable is unnecessary.

> Signed-off-by: Kuan-Wei Chiu <visitorckw@gmail.com>

Thanks!

> ---
>  mm/slub.c | 3 ++-
>  1 file changed, 2 insertions(+), 1 deletion(-)
> 
> diff --git a/mm/slub.c b/mm/slub.c
> index 30003763d224..c91b3744adbc 100644
> --- a/mm/slub.c
> +++ b/mm/slub.c
> @@ -7718,8 +7718,9 @@ static int cmp_loc_by_count(const void *a, const void *b, const void *data)
>  
>  	if (loc1->count > loc2->count)
>  		return -1;
> -	else
> +	if (loc1->count < loc2->count)
>  		return 1;
> +	return 0;
>  }
>  
>  static void *slab_debugfs_start(struct seq_file *seq, loff_t *ppos)

Re: [PATCH 1/2] mm/slub: Fix cmp_loc_by_count() to return 0 when counts are equal

[Kuan-Wei Chiu]

Hi Vlastimil,

On Mon, Aug 25, 2025 at 07:28:17PM +0200, Vlastimil Babka wrote:
> On 8/25/25 03:34, Kuan-Wei Chiu wrote:
> > The comparison function cmp_loc_by_count() used for sorting stack trace
> > locations in debugfs currently returns -1 if a->count > b->count and 1
> > otherwise. This breaks the antisymmetry property required by sort(),
> > because when two counts are equal, both cmp(a, b) and cmp(b, a) return
> > 1.
> 
> Good catch.
> 
> > This can lead to undefined or incorrect ordering results. Fix it by
> 
> Wonder if it can really affect anything in practice other than swapping
> needlessly some records with an equal count?
> 
It could result in some elements being incorrectly ordered, similar to
what happened before in ACPI causing issues with s2idle [1][2]. But in
this case, the worst impact is just the display order not matching the
count, so it's not too critical.

[1]: https://lore.kernel.org/lkml/70674dc7-5586-4183-8953-8095567e73df@gmail.com/
[2]: https://lore.kernel.org/lkml/20240701205639.117194-1-visitorckw@gmail.com/

> > explicitly returning 0 when the counts are equal, ensuring that the
> > comparison function follows the expected mathematical properties.
> 
> Agreed with the cmp_int() suggestion for a v2.
> 
I'll make that change in v2.

> > Fixes: 553c0369b3e1 ("mm/slub: sort debugfs output by frequency of stack traces")
> > Cc: stable@vger.kernel.org
> 
> I don't think it can cause any serious bugs so Cc: stable is unnecessary.
> 
I'll drop it in v2.

Regards,
Kuan-Wei

> > Signed-off-by: Kuan-Wei Chiu <visitorckw@gmail.com>
> 
> Thanks!
> 
> > ---
> >  mm/slub.c | 3 ++-
> >  1 file changed, 2 insertions(+), 1 deletion(-)
> > 
> > diff --git a/mm/slub.c b/mm/slub.c
> > index 30003763d224..c91b3744adbc 100644
> > --- a/mm/slub.c
> > +++ b/mm/slub.c
> > @@ -7718,8 +7718,9 @@ static int cmp_loc_by_count(const void *a, const void *b, const void *data)
> >  
> >  	if (loc1->count > loc2->count)
> >  		return -1;
> > -	else
> > +	if (loc1->count < loc2->count)
> >  		return 1;
> > +	return 0;
> >  }
> >  
> >  static void *slab_debugfs_start(struct seq_file *seq, loff_t *ppos)
> 

Re: [PATCH 1/2] mm/slub: Fix cmp_loc_by_count() to return 0 when counts are equal

[Harry Yoo]

On Tue, Aug 26, 2025 at 01:54:49AM +0800, Kuan-Wei Chiu wrote:
> Hi Vlastimil,
> 
> On Mon, Aug 25, 2025 at 07:28:17PM +0200, Vlastimil Babka wrote:
> > On 8/25/25 03:34, Kuan-Wei Chiu wrote:
> > > The comparison function cmp_loc_by_count() used for sorting stack trace
> > > locations in debugfs currently returns -1 if a->count > b->count and 1
> > > otherwise. This breaks the antisymmetry property required by sort(),
> > > because when two counts are equal, both cmp(a, b) and cmp(b, a) return
> > > 1.
> > 
> > Good catch.
> > 
> > > This can lead to undefined or incorrect ordering results. Fix it by
> > 
> > Wonder if it can really affect anything in practice other than swapping
> > needlessly some records with an equal count?
> > 
> It could result in some elements being incorrectly ordered, similar to
> what happened before in ACPI causing issues with s2idle [1][2]. But in
> this case, the worst impact is just the display order not matching the
> count, so it's not too critical.

Could you give an example where the previous cmp_loc_by_count() code
produces an incorrectly sorted array?

> [1]: https://lore.kernel.org/lkml/70674dc7-5586-4183-8953-8095567e73df@gmail.com
> [2]: https://lore.kernel.org/lkml/20240701205639.117194-1-visitorckw@gmail.com
> 
> > > explicitly returning 0 when the counts are equal, ensuring that the
> > > comparison function follows the expected mathematical properties.
> > 
> > Agreed with the cmp_int() suggestion for a v2.
> > 
> I'll make that change in v2.

-- 
Cheers,
Harry / Hyeonggon

Re: [PATCH 1/2] mm/slub: Fix cmp_loc_by_count() to return 0 when counts are equal

[Kuan-Wei Chiu]

On Tue, Aug 26, 2025 at 04:53:34PM +0900, Harry Yoo wrote:
> On Tue, Aug 26, 2025 at 01:54:49AM +0800, Kuan-Wei Chiu wrote:
> > Hi Vlastimil,
> > 
> > On Mon, Aug 25, 2025 at 07:28:17PM +0200, Vlastimil Babka wrote:
> > > On 8/25/25 03:34, Kuan-Wei Chiu wrote:
> > > > The comparison function cmp_loc_by_count() used for sorting stack trace
> > > > locations in debugfs currently returns -1 if a->count > b->count and 1
> > > > otherwise. This breaks the antisymmetry property required by sort(),
> > > > because when two counts are equal, both cmp(a, b) and cmp(b, a) return
> > > > 1.
> > > 
> > > Good catch.
> > > 
> > > > This can lead to undefined or incorrect ordering results. Fix it by
> > > 
> > > Wonder if it can really affect anything in practice other than swapping
> > > needlessly some records with an equal count?
> > > 
> > It could result in some elements being incorrectly ordered, similar to
> > what happened before in ACPI causing issues with s2idle [1][2]. But in
> > this case, the worst impact is just the display order not matching the
> > count, so it's not too critical.
> 
> Could you give an example where the previous cmp_loc_by_count() code
> produces an incorrectly sorted array?
> 
Sorry for the late reply.

I tried generating random arrays to find a concrete example where the
old cmp_loc_by_count() causes a wrong ordering, but I couldn't
reproduce one. So I would like to withdraw my earlier claim that it
definitely leads to incorrect results, since I cannot demonstrate a
failing case.

The complexity of the sort() implementation also makes it hard to
reason precisely whether such inputs exist.

That said, I still believe the patch should be merged, because sort()
only guarantees correct behavior if the comparison function satisfies
antisymmetry and transitivity. When those are violated, correctness
depends on implementation details, and future changes (e.g., switching
to a different sorting algorithm) could potentially break the ordering.

Regards,
Kuan-Wei

> > [1]: https://lore.kernel.org/lkml/70674dc7-5586-4183-8953-8095567e73df@gmail.com
> > [2]: https://lore.kernel.org/lkml/20240701205639.117194-1-visitorckw@gmail.com
> > 
> > > > explicitly returning 0 when the counts are equal, ensuring that the
> > > > comparison function follows the expected mathematical properties.
> > > 
> > > Agreed with the cmp_int() suggestion for a v2.
> > > 
> > I'll make that change in v2.
> 
> -- 
> Cheers,
> Harry / Hyeonggon

Re: [PATCH 1/2] mm/slub: Fix cmp_loc_by_count() to return 0 when counts are equal

[Harry Yoo]

On Fri, Aug 29, 2025 at 01:13:58AM +0800, Kuan-Wei Chiu wrote:
> On Tue, Aug 26, 2025 at 04:53:34PM +0900, Harry Yoo wrote:
> > On Tue, Aug 26, 2025 at 01:54:49AM +0800, Kuan-Wei Chiu wrote:
> > > Hi Vlastimil,
> > > 
> > > On Mon, Aug 25, 2025 at 07:28:17PM +0200, Vlastimil Babka wrote:
> > > > On 8/25/25 03:34, Kuan-Wei Chiu wrote:
> > > > > The comparison function cmp_loc_by_count() used for sorting stack trace
> > > > > locations in debugfs currently returns -1 if a->count > b->count and 1
> > > > > otherwise. This breaks the antisymmetry property required by sort(),
> > > > > because when two counts are equal, both cmp(a, b) and cmp(b, a) return
> > > > > 1.
> > > > 
> > > > Good catch.
> > > > 
> > > > > This can lead to undefined or incorrect ordering results. Fix it by
> > > > 
> > > > Wonder if it can really affect anything in practice other than swapping
> > > > needlessly some records with an equal count?
> > > > 
> > > It could result in some elements being incorrectly ordered, similar to
> > > what happened before in ACPI causing issues with s2idle [1][2]. But in
> > > this case, the worst impact is just the display order not matching the
> > > count, so it's not too critical.
> > 
> > Could you give an example where the previous cmp_loc_by_count() code
> > produces an incorrectly sorted array?
> > 
> Sorry for the late reply.

No problem ;)

> I tried generating random arrays to find a concrete example where the
> old cmp_loc_by_count() causes a wrong ordering, but I couldn't
> reproduce one. So I would like to withdraw my earlier claim that it
> definitely leads to incorrect results, since I cannot demonstrate a
> failing case.

Yeah I couldn't either. Maybe mathematical proof would work, but I
didn't try.

> That said, I still believe the patch should be merged, because sort()
> only guarantees correct behavior if the comparison function satisfies
> antisymmetry and transitivity. When those are violated, correctness
> depends on implementation details, and future changes (e.g., switching
> to a different sorting algorithm) could potentially break the ordering.

Agreed. No doubt the series is worth merging, just wanted to clarify
that bit.

Thanks!

-- 
Cheers,
Harry / Hyeonggon

> Regards,
> Kuan-Wei
> 
> > > [1]: https://lore.kernel.org/lkml/70674dc7-5586-4183-8953-8095567e73df@gmail.com
> > > [2]: https://lore.kernel.org/lkml/20240701205639.117194-1-visitorckw@gmail.com
> > > 
> > > > > explicitly returning 0 when the counts are equal, ensuring that the
> > > > > comparison function follows the expected mathematical properties.
> > > > 
> > > > Agreed with the cmp_int() suggestion for a v2.
> > > > 
> > > I'll make that change in v2.