On Mon, Sep 03, 2018 at 10:15:16AM +0300, Leon Romanovsky wrote:
> On Sun, Sep 02, 2018 at 02:42:38PM -0600, Jason Gunthorpe wrote:
> > On Tue, Aug 28, 2018 at 02:12:44PM +0300, Leon Romanovsky wrote:
> > > From: Leon Romanovsky <leonro@xxxxxxxxxxxx>
> > >
> > > Change on stack allocation to be dynamically allocated structure to fix
> > > the following compilation warning.
> > >
> > > drivers/infiniband/hw/mlx5/cq.c: In function _mlx5_ib_create_cq_:
> > > drivers/infiniband/hw/mlx5/cq.c:1091:1: warning: the frame size of 1064
> > > bytes is larger than 1024 bytes [-Wframe-larger-than=]
> > > }
> > > ^
> > >
> > > Signed-off-by: Leon Romanovsky <leonro@xxxxxxxxxxxx>
> > > drivers/infiniband/hw/mlx5/cq.c | 49 ++++++++++++++++++++++++-----------------
> > > 1 file changed, 29 insertions(+), 20 deletions(-)
> > >
> > > diff --git a/drivers/infiniband/hw/mlx5/cq.c b/drivers/infiniband/hw/mlx5/cq.c
> > > index 088205d7f1a1..ada77c9e4b6b 100644
> > > +++ b/drivers/infiniband/hw/mlx5/cq.c
> > > @@ -778,7 +778,7 @@ static int create_cq_user(struct mlx5_ib_dev *dev, struct ib_udata *udata,
> > > int entries, u32 **cqb,
> > > int *cqe_size, int *index, int *inlen)
> > > {
> > > - struct mlx5_ib_create_cq ucmd = {};
> > > + struct mlx5_ib_create_cq *ucmd;
> >
> > This is a 24 byte structure, it seems silly to allocate that alone on
> > the heap.
> >
> > How did the stack frame get to 1064 bytes big? I can't get anywhere
> > near that by eyeballing the stack usage in create_cq_user and
> > mlx5_ib_create_cq..
>
> 1.
> Before this patch
> _ kernel git:(rdma-next) objdump -d drivers/infiniband/hw/mlx5/mlx5_ib.ko | scripts/checkstack.pl | grep mlx5_ib_create_cq
> 0x00014da2 mlx5_ib_create_cq [mlx5_ib]: 1064
>
> After this patch
> _ kernel git:(rdma-next) objdump -d drivers/infiniband/hw/mlx5/mlx5_ib.ko | scripts/checkstack.pl | grep mlx5_ib_create_cq
> 0x00014da2 mlx5_ib_create_cq [mlx5_ib]: 992
>
> 2.
> Main consumer of stack in create_cq_user() is mlx5_ib_cont_pages() with
> reverted patch in question and empty mlx5_ib_cont_pages():
> 0x00014da2 mlx5_ib_create_cq [mlx5_ib]: 848
Wow, that is still another 800 bytes scattered all around.. That is
alot of inlining for this function..
> However I prefer do not touch mlx5_ib_cont_pages() without real bug.
Well, none of this is a real bug, other places in the kernel require >
1024 in this warning today, so let's not make anything more complicated?
Jason
