Re: [PATCH] sched/topology: Do not override sd->shared allocation

Hello Peter,

On 7/13/2026 12:26 PM, K Prateek Nayak wrote:
> Breno reported a kmemleak for sd->shared assignment on an asymmetric
> system with a single LLC.
> 
> Post the unconditional assignment of sd->shared for sd_llc in commit
> 9e005ed21152d ("sched/topology: Allow multiple domains to claim
> sched_domain_shared"), init_sched_domain_shared() can overrwrite the
> sd->shared assignment for as_asym_cpucapacity when sd_llc and
> as_asym_cpucapacity point to the same domain.
> 
> Return early from init_sched_domain_shared() if a valid sd->shared
> assignment is detected. Since the degeneration path only decrements the
> refcount by 1, it is correct to only account the sd->shared assignment
> once per sched-domain level, even if the flags overlap.
> 
> While at it, correct a typo in init_sched_domain_shared().
> 
> Reported-by: Breno Leitao <[email protected]>
> Closes: https://lore.kernel.org/lkml/[email protected]/
> Fixes: 9e005ed21152d ("sched/topology: Allow multiple domains to claim sched_domain_shared")
> Tested-by: Breno Leitao <[email protected]>
> Tested-by: Dietmar Eggemann <[email protected]>
> Signed-off-by: K Prateek Nayak <[email protected]>

Sorry for the late notice! Was stuck with some internal fire but in case
sched/urgent is still open, could you stick this fix for a memory leak
on top.

sched/core is good too otherwise - AUTOSEL should pick this up for
stable, and if it doesn't, I'll send out a patch to Greg for v7.2.
Sorry again for the late ping.

> ---
>  kernel/sched/topology.c | 10 +++++++++-
>  1 file changed, 9 insertions(+), 1 deletion(-)
> 
> diff --git a/kernel/sched/topology.c b/kernel/sched/topology.c
> index 622e2e01974c4..b761ecda790a8 100644
> --- a/kernel/sched/topology.c
> +++ b/kernel/sched/topology.c
> @@ -2942,6 +2942,14 @@ init_sched_domain_shared(struct s_data *d, struct sched_domain *sd, int flags)
>  	struct sched_domain_shared *sds = NULL;
>  	int cpu;
>  
> +	/*
> +	 * sd->shared is already assigned for this domain level for a
> +	 * different flag. Nothing to do since this allocation is
> +	 * already accounted for.
> +	 */
> +	if (sd->shared)
> +		return;
> +
>  	/*
>  	 * Multiple domains can try to claim a shared object like
>  	 * SD_ASYM_CPUCAPACITY and SD_SHARE_LLC which can alias to
> @@ -2970,7 +2978,7 @@ init_sched_domain_shared(struct s_data *d, struct sched_domain *sd, int flags)
>  
>  	/*
>  	 * Use the sd_shared corresponding to the last
> -	 * CPU in the span if none are avaialable.
> +	 * CPU in the span if none are available.
>  	 */
>  	if (WARN_ON_ONCE(!sd->shared))
>  		sd->shared = sds;
> 
> base-commit: 04998aa54848f15332202d0bea008d2ca1ed1713

-- 
Thanks and Regards,
Prateek