Re: [PATCH] RDMA/mlx5: Use memset_after() to zero struct mlx5_ib_mr

From: Kees Cook
Date: Wed Jan 12 2022 - 15:50:12 EST


On Tue, Dec 07, 2021 at 03:45:25PM -0400, Jason Gunthorpe wrote:
> On Tue, Dec 07, 2021 at 11:41:07AM -0800, Kees Cook wrote:
> > On Tue, Dec 07, 2021 at 02:47:29PM -0400, Jason Gunthorpe wrote:
> > > On Sun, Nov 21, 2021 at 03:54:55PM +0200, Leon Romanovsky wrote:
> > > > On Thu, Nov 18, 2021 at 12:31:38PM -0800, Kees Cook wrote:
> > > > > In preparation for FORTIFY_SOURCE performing compile-time and run-time
> > > > > field bounds checking for memset(), avoid intentionally writing across
> > > > > neighboring fields.
> > > > >
> > > > > Use memset_after() to zero the end of struct mlx5_ib_mr that should
> > > > > be initialized.
> > > > >
> > > > > Signed-off-by: Kees Cook <keescook@xxxxxxxxxxxx>
> > > > > drivers/infiniband/hw/mlx5/mlx5_ib.h | 5 ++---
> > > > > 1 file changed, 2 insertions(+), 3 deletions(-)
> > > > >
> > > > > diff --git a/drivers/infiniband/hw/mlx5/mlx5_ib.h b/drivers/infiniband/hw/mlx5/mlx5_ib.h
> > > > > index e636e954f6bf..af94c9fe8753 100644
> > > > > +++ b/drivers/infiniband/hw/mlx5/mlx5_ib.h
> > > > > @@ -665,8 +665,7 @@ struct mlx5_ib_mr {
> > > > > /* User MR data */
> > > > > struct mlx5_cache_ent *cache_ent;
> > > > > struct ib_umem *umem;
> > > > > -
> > > > > - /* This is zero'd when the MR is allocated */
> > > > > + /* Everything after umem is zero'd when the MR is allocated */
> > > > > union {
> > > > > /* Used only while the MR is in the cache */
> > > > > struct {
> > > > > @@ -718,7 +717,7 @@ struct mlx5_ib_mr {
> > > > > /* Zero the fields in the mr that are variant depending on usage */
> > > > > static inline void mlx5_clear_mr(struct mlx5_ib_mr *mr)
> > > > > {
> > > > > - memset(mr->out, 0, sizeof(*mr) - offsetof(struct mlx5_ib_mr, out));
> > > > > + memset_after(mr, 0, umem);
> > > >
> > > > I think that it is not equivalent change and you need "memset_after(mr, 0, cache_ent);"
> > > > to clear umem pointer too.
> > >
> > > Kees?
> >
> > Oops, sorry, I missed the ealrier reply!
> >
> > I don't think that matches -- the original code wipes from the start of
> > "out" to the end of the struct. "out" is the first thing in the union
> > after "umem", so "umem" was not wiped before. I retained that behavior
> > ("wipe everything after umem").
> >
> > Am I misunderstanding the desired behavior here?
>
> Ah, it is this patch:
>
> commit f0ae4afe3d35e67db042c58a52909e06262b740f
> Author: Alaa Hleihel <alaa@xxxxxxxxxx>
> Date: Mon Nov 22 13:41:51 2021 +0200
>
> RDMA/mlx5: Fix releasing unallocated memory in dereg MR flow
>
> Which moved umem into the union that is causing the confusion
>
> It hasn't quite made it to a rc release yet, so I suppose the answer
> is to rebase this on that then it is as Leon says about cache_ent

The umem patch appears to have been reverted. Should I send an updated
patch for memset_after()? I think it would simply be:

diff --git a/drivers/infiniband/hw/mlx5/mlx5_ib.h b/drivers/infiniband/hw/mlx5/mlx5_ib.h
index e3c33be9c5a0..a58f69b19705 100644
--- a/drivers/infiniband/hw/mlx5/mlx5_ib.h
+++ b/drivers/infiniband/hw/mlx5/mlx5_ib.h
@@ -667,7 +667,7 @@ struct mlx5_ib_mr {
struct mlx5_cache_ent *cache_ent;
struct ib_umem *umem;

- /* This is zero'd when the MR is allocated */
+ /* Everything after umem is zero'd when MR allocated */
union {
/* Used only while the MR is in the cache */
struct {
@@ -719,7 +719,7 @@ struct mlx5_ib_mr {
/* Zero the fields in the mr that are variant depending on usage */
static inline void mlx5_clear_mr(struct mlx5_ib_mr *mr)
{
- memset(mr->out, 0, sizeof(*mr) - offsetof(struct mlx5_ib_mr, out));
+ memset_after(mr, 0, umem);
}

static inline bool is_odp_mr(struct mlx5_ib_mr *mr)

--
Kees Cook