Re: [RFC PATCH v5 01/16] RDMA/irdma: Add driver framework definitions

[Jason Gunthorpe <jgg@ziepe.ca>]

On Tue, Apr 21, 2020 at 06:19:07PM +0000, Saleem, Shiraz wrote:
> > Subject: Re: [RFC PATCH v5 01/16] RDMA/irdma: Add driver framework
> > definitions
> > 
> > On Tue, Apr 21, 2020 at 12:23:45AM +0000, Saleem, Shiraz wrote:
> > > > Subject: Re: [RFC PATCH v5 01/16] RDMA/irdma: Add driver framework
> > > > definitions
> > > >
> > > > On Fri, Apr 17, 2020 at 10:12:36AM -0700, Jeff Kirsher wrote:
> > > > > From: Mustafa Ismail <mustafa.ismail@intel.com>
> > > > >
> > > > > Register irdma as a virtbus driver capable of supporting virtbus
> > > > > devices from multi-generation RDMA capable Intel HW. Establish the
> > > > > interface with all supported netdev peer drivers and initialize HW.
> > > > >
> > > > > Signed-off-by: Mustafa Ismail <mustafa.ismail@intel.com>
> > > > > Signed-off-by: Shiraz Saleem <shiraz.saleem@intel.com>
> > > > > drivers/infiniband/hw/irdma/i40iw_if.c | 228 ++++++++++
> > > > > drivers/infiniband/hw/irdma/irdma_if.c | 449 ++++++++++++++++++
> > > > >  drivers/infiniband/hw/irdma/main.c     | 573 +++++++++++++++++++++++
> > > > >  drivers/infiniband/hw/irdma/main.h     | 599
> > +++++++++++++++++++++++++
> > > > >  4 files changed, 1849 insertions(+)  create mode 100644
> > > > > drivers/infiniband/hw/irdma/i40iw_if.c
> > > > >  create mode 100644 drivers/infiniband/hw/irdma/irdma_if.c
> > > > >  create mode 100644 drivers/infiniband/hw/irdma/main.c
> > > > >  create mode 100644 drivers/infiniband/hw/irdma/main.h
> > > > >
> > > >
> > > > I didn't look in too much details, but three things caught my attention
> > immediately:
> > > > 1. Existence of ARP cache management logic in RDMA driver.
> > >
> > > Our HW has an independent ARP table for the rdma block.
> > > driver needs to add an ARP table entry via an rdma admin queue command
> > > before QP transitions to RTS.
> > >
> > > > 2. Extensive use of dev_*() prints while we have ibdev_*() prints
> > > The ib device object is not available till the end of the device init
> > > similarly its unavailable early on in device deinit flows. So dev_* is
> > > all we can use in those places.
> > 
> > hns guys were thinking about changing this. It looks fine to just move the name
> > assignment to the device allocation, then we don't have this weirdness
> 
> Did you mean moving name setting from ib_register_device to ib_device_alloc?
> Will that work ok for how rvt is handling the names in rvt_set_ibdev_name
> and its register?

I don't see why not? rvt_set_ibdev_name is always directly after
rvt_alloc_device, which is the thing that calls ib_alloc_device

> This could migrate a lot of the dev_* to ibdev_* but there is still
> going to be a handful of dev_* usages from our HW initialization in
> irdma_prob_dev since ib device allocation is done in irdma_open.

Don't do that?

> > Alternatively, you could do as netdev does and have a special name string when
> > the name is NULL
> 
> Not sure I found what your referring to. 
> Did you mean similar to use of netdev_name in __netdev_printk?

Yes

Jason