2020-06-12 19:56:52

by Aditya Pakki

Subject: [PATCH] RDMA/rvt: Fix potential memory leak caused by rvt_alloc_rq

In case of failure of alloc_ud_wq_attr, the memory allocated by
rvt_alloc_rq() is not freed. The patch fixes this issue by
calling rvt_free_rq().

Signed-off-by: Aditya Pakki <[email protected]>
---
drivers/infiniband/sw/rdmavt/qp.c | 1 +
1 file changed, 1 insertion(+)

diff --git a/drivers/infiniband/sw/rdmavt/qp.c b/drivers/infiniband/sw/rdmavt/qp.c
index 511b72809e14..17ea7da73bf9 100644
--- a/drivers/infiniband/sw/rdmavt/qp.c
+++ b/drivers/infiniband/sw/rdmavt/qp.c
@@ -1203,6 +1203,7 @@ struct ib_qp *rvt_create_qp(struct ib_pd *ibpd,
qp->s_flags = RVT_S_SIGNAL_REQ_WR;
err = alloc_ud_wq_attr(qp, rdi->dparms.node);
if (err) {
+ rvt_free_rq(&qp->r_rq);
ret = (ERR_PTR(err));
goto bail_driver_priv;
}
--
2.25.1

2020-06-13 17:29:32

by Dennis Dalessandro

Subject: Re: [PATCH] RDMA/rvt: Fix potential memory leak caused by rvt_alloc_rq

On 6/12/2020 3:54 PM, Aditya Pakki wrote:
> In case of failure of alloc_ud_wq_attr, the memory allocated by
> rvt_alloc_rq() is not freed. The patch fixes this issue by
> calling rvt_free_rq().
>
> Signed-off-by: Aditya Pakki <[email protected]>
> ---
> drivers/infiniband/sw/rdmavt/qp.c | 1 +
> 1 file changed, 1 insertion(+)
>
> diff --git a/drivers/infiniband/sw/rdmavt/qp.c b/drivers/infiniband/sw/rdmavt/qp.c
> index 511b72809e14..17ea7da73bf9 100644
> --- a/drivers/infiniband/sw/rdmavt/qp.c
> +++ b/drivers/infiniband/sw/rdmavt/qp.c
> @@ -1203,6 +1203,7 @@ struct ib_qp *rvt_create_qp(struct ib_pd *ibpd,
> qp->s_flags = RVT_S_SIGNAL_REQ_WR;
> err = alloc_ud_wq_attr(qp, rdi->dparms.node);
> if (err) {
> + rvt_free_rq(&qp->r_rq);
> ret = (ERR_PTR(err));
> goto bail_driver_priv;
> }
>

This should probably use the unwind code at the end to be consistent.

Looks like the rvt_free_rq and free_ud_wq_attr have gotten out of order
and shouldn't be tied together, so that needs fixed up too.

I'd need to study the git log a little more to see what happened but I
think d310c4bf8aea ("IB/{rdmavt, hfi1, qib}: Remove AH refcount for UD
QPs") just missed this.

-Denny