> -----Original Message-----
> From: Moni Shoua <monis@xxxxxxxxxxxx>
> Sent: Wednesday, February 6, 2019 1:21 AM
> To: Parav Pandit <parav@xxxxxxxxxxxx>
> Cc: Leon Romanovsky <leon@xxxxxxxxxx>; Doug Ledford
> <dledford@xxxxxxxxxx>; Jason Gunthorpe <jgg@xxxxxxxxxxxx>; Leon
> Romanovsky <leonro@xxxxxxxxxxxx>; RDMA mailing list <linux-
> rdma@xxxxxxxxxxxxxxx>
> Subject: Re: [PATCH rdma-next 1/2] IB/mlx5: Protect against prefetch of
> invalid MR
>
> > > struct pf_frame *head = NULL, *frame;
> > > struct mlx5_core_mkey *mmkey;
> > > struct mlx5_ib_mr *mr;
> > > @@ -772,6 +777,10 @@ static int pagefault_single_data_segment(struct
> > > mlx5_ib_dev *dev, u32 key,
> > > switch (mmkey->type) {
> > > case MLX5_MKEY_MR:
> > > mr = container_of(mmkey, struct mlx5_ib_mr, mmkey);
> > > +
> > > + if (deferred)
> > > + atomic_dec(&mr->num_pending_prefetch);
> > > +
>
> > Work item handler is still running on MR at this point and
> flush_workqueue would skip to flush.
> > atomic_dec() should be after completing pagefault_mr(() where we are
> done with the MR.
> Here, we are in srcu critical section and dereg_mr() waits for it after flushing
> the workqueue with synchronize_srcu()
okey
