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

[Edward Srouji <edwards@nvidia.com>]

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>
>>