On Sat, Jan 19, 2019 at 10:00:25AM +0200, Leon Romanovsky wrote:
> On Fri, Jan 18, 2019 at 09:59:31PM +0000, Jason Gunthorpe wrote:
> > On Thu, Jan 17, 2019 at 08:04:45PM +0200, Leon Romanovsky wrote:
> > > From: Leon Romanovsky <leonro@xxxxxxxxxxxx>
> > >
> > > Adapt mlx5 to share restrack ID instead of local variable.
> > >
> > > Signed-off-by: Leon Romanovsky <leonro@xxxxxxxxxxxx>
> > > drivers/infiniband/hw/mlx5/main.c | 22 +++++++++++++++-------
> > > drivers/infiniband/hw/mlx5/mlx5_ib.h | 1 -
> > > drivers/infiniband/hw/mlx5/mr.c | 16 ++++++++--------
> > > drivers/infiniband/hw/mlx5/qp.c | 22 +++++++++++++---------
> > > drivers/infiniband/hw/mlx5/srq.c | 2 +-
> > > 5 files changed, 37 insertions(+), 26 deletions(-)
> > >
> > > diff --git a/drivers/infiniband/hw/mlx5/main.c b/drivers/infiniband/hw/mlx5/main.c
> > > index b32b5c7f1cfd..9db041465266 100644
> > > +++ b/drivers/infiniband/hw/mlx5/main.c
> > > @@ -2262,11 +2262,14 @@ static struct ib_pd *mlx5_ib_alloc_pd(struct ib_device *ibdev,
> > > u32 out[MLX5_ST_SZ_DW(alloc_pd_out)] = {};
> > > u32 in[MLX5_ST_SZ_DW(alloc_pd_in)] = {};
> > > u16 uid = 0;
> > > + u32 pdn;
> > >
> > > pd = kzalloc(sizeof(*pd), GFP_KERNEL);
> > > if (!pd)
> > > return ERR_PTR(-ENOMEM);
> > >
> > > + pd->ibpd.device = ibdev;
> >
> > This needs to be re-ordered now? Don't like..
>
> I need "device" to know where to store restrack objects.
> Because we don't have "priv" for those ib_* objects, we can't
> preallocate them before drivers (we don't know the allocation
> size). What about if we introduce new functions in IB/core for
> all objects to allocate structure and set device?
I would prefer if the core could still do the restrack_add just using
the device provided ID number.. Is that not possible? I didn't look
closely yet
> Something like this:
> void *alloc_object(struct ib_device *dev, size_t s, uint64_t offset_to_device_pointer_in_driver)
> {
> void *a;
>
> a = kzalloc(s, GFP_KERNEL);
> if (!a)
> return ERR_PTR(-ENOMEM);
> *(a + offset_to_device_pointer_in_driver) = dev;
> return a;
> }
Yuk, if we are messing with this we should include the allocation
sizes in the ops structure and have the core provide an allocated
pointer. Would remove lots of driver code, but is a huge amount of
work..
Jason
