[PATCH rdma-next] RDMA/mlx5: Free software WCs left on wc_list at CQ destroy

[PATCH rdma-next] RDMA/mlx5: Free software WCs left on wc_list at CQ destroy

[Edward Srouji]

Software WCs generated by mlx5_ib_generate_wc() are allocated one at a
time and linked onto cq->wc_list. The only thing that ever frees them is
poll_soft_wc(), i.e. a consumer actually reaping them. If the CQ is
destroyed while entries are still linked, those allocations are leaked.

Drain the list in destroy_cq_kernel(). Nothing can be walking wc_list by
then: the poll context is the only walker, notify_work the only thing
that can re-arm it, and both are stopped before post_destroy_cq() runs.

destroy_cq_kernel() is also the error unwind for mlx5_ib_create_cq(), so
move the INIT_LIST_HEAD() of wc_list up next to the other list
initialisations, ahead of anything that can jump to err_cqb.  The driver
object is zero-allocated, and list_for_each_entry_safe() evaluates the
next linkage before testing the loop condition, so draining a wc_list
that is still all zeroes would oops.

Fixes: 25361e02c448 ("IB/mlx5: Generate completions in software")
Reviewed-by: Maher Sanalla <msanalla@nvidia.com>
Signed-off-by: Edward Srouji <edwards@nvidia.com>
---
 drivers/infiniband/hw/mlx5/cq.c | 13 ++++++++++++-
 1 file changed, 12 insertions(+), 1 deletion(-)

diff --git a/drivers/infiniband/hw/mlx5/cq.c b/drivers/infiniband/hw/mlx5/cq.c
index 55ffd7d4de517e6ce3f04c98b4edcd94b12e43b3..2f7ac350c3d8f179ad1d4d8c1181b93f3151c7ce 100644
--- a/drivers/infiniband/hw/mlx5/cq.c
+++ b/drivers/infiniband/hw/mlx5/cq.c
@@ -933,8 +933,19 @@ static int create_cq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq,
 	return err;
 }
 
+static void drain_soft_wc(struct mlx5_ib_cq *cq)
+{
+	struct mlx5_ib_wc *soft_wc, *next;
+
+	list_for_each_entry_safe(soft_wc, next, &cq->wc_list, list) {
+		list_del(&soft_wc->list);
+		kfree(soft_wc);
+	}
+}
+
 static void destroy_cq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq)
 {
+	drain_soft_wc(cq);
 	free_cq_buf(dev, &cq->buf);
 	mlx5_db_free(dev->mdev, &cq->db);
 }
@@ -1067,6 +1078,7 @@ int mlx5_ib_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
 	spin_lock_init(&cq->lock);
 	INIT_LIST_HEAD(&cq->list_send_qp);
 	INIT_LIST_HEAD(&cq->list_recv_qp);
+	INIT_LIST_HEAD(&cq->wc_list);
 
 	cqe_size = cache_line_size() == 128 ? 128 : 64;
 	err = create_cq_kernel(dev, cq, entries, cqe_size, &cqb, &index,
@@ -1102,7 +1114,6 @@ int mlx5_ib_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
 	mlx5_ib_dbg(dev, "cqn 0x%x\n", cq->mcq.cqn);
 	cq->mcq.event = mlx5_ib_cq_event;
 
-	INIT_LIST_HEAD(&cq->wc_list);
 	kvfree(cqb);
 	return 0;
 

---
base-commit: 6628e7d4c6b25c8d18f1345e3a20d21829aae7e4
change-id: 20261103-free-sw-wcs-8cf40aa66fd8

Best regards,
-- 
Edward Srouji <edwards@nvidia.com>

Re: [PATCH rdma-next] RDMA/mlx5: Free software WCs left on wc_list at CQ destroy

[Leon Romanovsky]

On Tue, Nov 03, 2026 at 06:55:19PM +0200, Edward Srouji wrote:
> Software WCs generated by mlx5_ib_generate_wc() are allocated one at a
> time and linked onto cq->wc_list. The only thing that ever frees them is
> poll_soft_wc(), i.e. a consumer actually reaping them. If the CQ is
> destroyed while entries are still linked, those allocations are leaked.

Is this real issue? Just put WARN_ON(!list_empty(...)) instead.

Thanks

> 
> Drain the list in destroy_cq_kernel(). Nothing can be walking wc_list by
> then: the poll context is the only walker, notify_work the only thing
> that can re-arm it, and both are stopped before post_destroy_cq() runs.
> 
> destroy_cq_kernel() is also the error unwind for mlx5_ib_create_cq(), so
> move the INIT_LIST_HEAD() of wc_list up next to the other list
> initialisations, ahead of anything that can jump to err_cqb.  The driver
> object is zero-allocated, and list_for_each_entry_safe() evaluates the
> next linkage before testing the loop condition, so draining a wc_list
> that is still all zeroes would oops.
> 
> Fixes: 25361e02c448 ("IB/mlx5: Generate completions in software")
> Reviewed-by: Maher Sanalla <msanalla@nvidia.com>
> Signed-off-by: Edward Srouji <edwards@nvidia.com>
> ---
>  drivers/infiniband/hw/mlx5/cq.c | 13 ++++++++++++-
>  1 file changed, 12 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/infiniband/hw/mlx5/cq.c b/drivers/infiniband/hw/mlx5/cq.c
> index 55ffd7d4de517e6ce3f04c98b4edcd94b12e43b3..2f7ac350c3d8f179ad1d4d8c1181b93f3151c7ce 100644
> --- a/drivers/infiniband/hw/mlx5/cq.c
> +++ b/drivers/infiniband/hw/mlx5/cq.c
> @@ -933,8 +933,19 @@ static int create_cq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq,
>  	return err;
>  }
>  
> +static void drain_soft_wc(struct mlx5_ib_cq *cq)
> +{
> +	struct mlx5_ib_wc *soft_wc, *next;
> +
> +	list_for_each_entry_safe(soft_wc, next, &cq->wc_list, list) {
> +		list_del(&soft_wc->list);
> +		kfree(soft_wc);
> +	}
> +}
> +
>  static void destroy_cq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq)
>  {
> +	drain_soft_wc(cq);
>  	free_cq_buf(dev, &cq->buf);
>  	mlx5_db_free(dev->mdev, &cq->db);
>  }
> @@ -1067,6 +1078,7 @@ int mlx5_ib_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
>  	spin_lock_init(&cq->lock);
>  	INIT_LIST_HEAD(&cq->list_send_qp);
>  	INIT_LIST_HEAD(&cq->list_recv_qp);
> +	INIT_LIST_HEAD(&cq->wc_list);
>  
>  	cqe_size = cache_line_size() == 128 ? 128 : 64;
>  	err = create_cq_kernel(dev, cq, entries, cqe_size, &cqb, &index,
> @@ -1102,7 +1114,6 @@ int mlx5_ib_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
>  	mlx5_ib_dbg(dev, "cqn 0x%x\n", cq->mcq.cqn);
>  	cq->mcq.event = mlx5_ib_cq_event;
>  
> -	INIT_LIST_HEAD(&cq->wc_list);
>  	kvfree(cqb);
>  	return 0;
>  
> 
> ---
> base-commit: 6628e7d4c6b25c8d18f1345e3a20d21829aae7e4
> change-id: 20261103-free-sw-wcs-8cf40aa66fd8
> 
> Best regards,
> -- 
> Edward Srouji <edwards@nvidia.com>
> 

Re: [PATCH rdma-next] RDMA/mlx5: Free software WCs left on wc_list at CQ destroy

[Edward Srouji]

On 11/4/2026 4:01 PM, Leon Romanovsky wrote:
> External email: Use caution opening links or attachments
> 
> 
> On Tue, Nov 03, 2026 at 06:55:19PM +0200, Edward Srouji wrote:
>> Software WCs generated by mlx5_ib_generate_wc() are allocated one at a
>> time and linked onto cq->wc_list. The only thing that ever frees them is
>> poll_soft_wc(), i.e. a consumer actually reaping them. If the CQ is
>> destroyed while entries are still linked, those allocations are leaked.
> 
> Is this real issue? Just put WARN_ON(!list_empty(...)) instead.
> 
> Thanks

Yes, it's a real issue reachable on a normal device removal.
E.g. ib_mad destroys the GSI QP and frees the CQ without polling it 
(while it may have unpolled completions).
I'd say that this is a complementary fix to 
https://patch.msgid.link/20261008141209.6923-1-prathameshdeshpande7@gmail.com
Before it, a pending notify_work still ran and polled from freed memory, 
which was the UAF it fixed.

> 
>>
>> Drain the list in destroy_cq_kernel(). Nothing can be walking wc_list by
>> then: the poll context is the only walker, notify_work the only thing
>> that can re-arm it, and both are stopped before post_destroy_cq() runs.
>>
>> destroy_cq_kernel() is also the error unwind for mlx5_ib_create_cq(), so
>> move the INIT_LIST_HEAD() of wc_list up next to the other list
>> initialisations, ahead of anything that can jump to err_cqb.  The driver
>> object is zero-allocated, and list_for_each_entry_safe() evaluates the
>> next linkage before testing the loop condition, so draining a wc_list
>> that is still all zeroes would oops.
>>
>> Fixes: 25361e02c448 ("IB/mlx5: Generate completions in software")
>> Reviewed-by: Maher Sanalla <msanalla@nvidia.com>
>> Signed-off-by: Edward Srouji <edwards@nvidia.com>
>> ---
>>   drivers/infiniband/hw/mlx5/cq.c | 13 ++++++++++++-
>>   1 file changed, 12 insertions(+), 1 deletion(-)
>>
>> diff --git a/drivers/infiniband/hw/mlx5/cq.c b/drivers/infiniband/hw/mlx5/cq.c
>> index 55ffd7d4de517e6ce3f04c98b4edcd94b12e43b3..2f7ac350c3d8f179ad1d4d8c1181b93f3151c7ce 100644
>> --- a/drivers/infiniband/hw/mlx5/cq.c
>> +++ b/drivers/infiniband/hw/mlx5/cq.c
>> @@ -933,8 +933,19 @@ static int create_cq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq,
>>        return err;
>>   }
>>
>> +static void drain_soft_wc(struct mlx5_ib_cq *cq)
>> +{
>> +     struct mlx5_ib_wc *soft_wc, *next;
>> +
>> +     list_for_each_entry_safe(soft_wc, next, &cq->wc_list, list) {
>> +             list_del(&soft_wc->list);
>> +             kfree(soft_wc);
>> +     }
>> +}
>> +
>>   static void destroy_cq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq)
>>   {
>> +     drain_soft_wc(cq);
>>        free_cq_buf(dev, &cq->buf);
>>        mlx5_db_free(dev->mdev, &cq->db);
>>   }
>> @@ -1067,6 +1078,7 @@ int mlx5_ib_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
>>        spin_lock_init(&cq->lock);
>>        INIT_LIST_HEAD(&cq->list_send_qp);
>>        INIT_LIST_HEAD(&cq->list_recv_qp);
>> +     INIT_LIST_HEAD(&cq->wc_list);
>>
>>        cqe_size = cache_line_size() == 128 ? 128 : 64;
>>        err = create_cq_kernel(dev, cq, entries, cqe_size, &cqb, &index,
>> @@ -1102,7 +1114,6 @@ int mlx5_ib_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
>>        mlx5_ib_dbg(dev, "cqn 0x%x\n", cq->mcq.cqn);
>>        cq->mcq.event = mlx5_ib_cq_event;
>>
>> -     INIT_LIST_HEAD(&cq->wc_list);
>>        kvfree(cqb);
>>        return 0;
>>
>>
>> ---
>> base-commit: 6628e7d4c6b25c8d18f1345e3a20d21829aae7e4
>> change-id: 20261103-free-sw-wcs-8cf40aa66fd8
>>
>> Best regards,
>> --
>> Edward Srouji <edwards@nvidia.com>
>>

Re: [PATCH rdma-next] RDMA/mlx5: Free software WCs left on wc_list at CQ destroy

[Leon Romanovsky]

On Thu, Nov 05, 2026 at 12:05:47PM +0200, Edward Srouji wrote:
> 
> 
> On 11/4/2026 4:01 PM, Leon Romanovsky wrote:
> > External email: Use caution opening links or attachments
> > 
> > 
> > On Tue, Nov 03, 2026 at 06:55:19PM +0200, Edward Srouji wrote:
> > > Software WCs generated by mlx5_ib_generate_wc() are allocated one at a
> > > time and linked onto cq->wc_list. The only thing that ever frees them is
> > > poll_soft_wc(), i.e. a consumer actually reaping them. If the CQ is
> > > destroyed while entries are still linked, those allocations are leaked.
> > 
> > Is this real issue? Just put WARN_ON(!list_empty(...)) instead.
> > 
> > Thanks
> 
> Yes, it's a real issue reachable on a normal device removal.
> E.g. ib_mad destroys the GSI QP and frees the CQ without polling it (while
> it may have unpolled completions).

"may"? Does it have unpolled completions or not?

> I'd say that this is a complementary fix to
> https://patch.msgid.link/20261008141209.6923-1-prathameshdeshpande7@gmail.com
> Before it, a pending notify_work still ran and polled from freed memory,
> which was the UAF it fixed.

Do you have dmesg stack trace?

Thanks

> 
> > 
> > > 
> > > Drain the list in destroy_cq_kernel(). Nothing can be walking wc_list by
> > > then: the poll context is the only walker, notify_work the only thing
> > > that can re-arm it, and both are stopped before post_destroy_cq() runs.
> > > 
> > > destroy_cq_kernel() is also the error unwind for mlx5_ib_create_cq(), so
> > > move the INIT_LIST_HEAD() of wc_list up next to the other list
> > > initialisations, ahead of anything that can jump to err_cqb.  The driver
> > > object is zero-allocated, and list_for_each_entry_safe() evaluates the
> > > next linkage before testing the loop condition, so draining a wc_list
> > > that is still all zeroes would oops.
> > > 
> > > Fixes: 25361e02c448 ("IB/mlx5: Generate completions in software")
> > > Reviewed-by: Maher Sanalla <msanalla@nvidia.com>
> > > Signed-off-by: Edward Srouji <edwards@nvidia.com>
> > > ---
> > >   drivers/infiniband/hw/mlx5/cq.c | 13 ++++++++++++-
> > >   1 file changed, 12 insertions(+), 1 deletion(-)
> > > 
> > > diff --git a/drivers/infiniband/hw/mlx5/cq.c b/drivers/infiniband/hw/mlx5/cq.c
> > > index 55ffd7d4de517e6ce3f04c98b4edcd94b12e43b3..2f7ac350c3d8f179ad1d4d8c1181b93f3151c7ce 100644
> > > --- a/drivers/infiniband/hw/mlx5/cq.c
> > > +++ b/drivers/infiniband/hw/mlx5/cq.c
> > > @@ -933,8 +933,19 @@ static int create_cq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq,
> > >        return err;
> > >   }
> > > 
> > > +static void drain_soft_wc(struct mlx5_ib_cq *cq)
> > > +{
> > > +     struct mlx5_ib_wc *soft_wc, *next;
> > > +
> > > +     list_for_each_entry_safe(soft_wc, next, &cq->wc_list, list) {
> > > +             list_del(&soft_wc->list);
> > > +             kfree(soft_wc);
> > > +     }
> > > +}
> > > +
> > >   static void destroy_cq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq)
> > >   {
> > > +     drain_soft_wc(cq);
> > >        free_cq_buf(dev, &cq->buf);
> > >        mlx5_db_free(dev->mdev, &cq->db);
> > >   }
> > > @@ -1067,6 +1078,7 @@ int mlx5_ib_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
> > >        spin_lock_init(&cq->lock);
> > >        INIT_LIST_HEAD(&cq->list_send_qp);
> > >        INIT_LIST_HEAD(&cq->list_recv_qp);
> > > +     INIT_LIST_HEAD(&cq->wc_list);
> > > 
> > >        cqe_size = cache_line_size() == 128 ? 128 : 64;
> > >        err = create_cq_kernel(dev, cq, entries, cqe_size, &cqb, &index,
> > > @@ -1102,7 +1114,6 @@ int mlx5_ib_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
> > >        mlx5_ib_dbg(dev, "cqn 0x%x\n", cq->mcq.cqn);
> > >        cq->mcq.event = mlx5_ib_cq_event;
> > > 
> > > -     INIT_LIST_HEAD(&cq->wc_list);
> > >        kvfree(cqb);
> > >        return 0;
> > > 
> > > 
> > > ---
> > > base-commit: 6628e7d4c6b25c8d18f1345e3a20d21829aae7e4
> > > change-id: 20261103-free-sw-wcs-8cf40aa66fd8
> > > 
> > > Best regards,
> > > --
> > > Edward Srouji <edwards@nvidia.com>
> > > 
>