> -----Original Message----- > From: linux-rdma-owner@xxxxxxxxxxxxxxx <linux-rdma- > owner@xxxxxxxxxxxxxxx> On Behalf Of Gal Pressman > Sent: Wednesday, April 3, 2019 2:14 AM > To: Leon Romanovsky <leon@xxxxxxxxxx> > Cc: Doug Ledford <dledford@xxxxxxxxxx>; Jason Gunthorpe > <jgg@xxxxxxxxxxxx>; RDMA mailing list <linux-rdma@xxxxxxxxxxxxxxx>; > Majd Dibbiny <majd@xxxxxxxxxxxx>; Mark Zhang <markz@xxxxxxxxxxxx>; > Saeed Mahameed <saeedm@xxxxxxxxxxxx> > Subject: Re: [PATCH rdma-next 06/16] RDMA/counter: Add "auto" > configuration mode support > > On 02-Apr-19 21:29, Leon Romanovsky wrote: > > On Mon, Apr 01, 2019 at 05:19:30PM +0300, Gal Pressman wrote: > >> On 01-Apr-19 16:35, Leon Romanovsky wrote: > >>> On Mon, Apr 01, 2019 at 03:07:09PM +0300, Gal Pressman wrote: > >>>> On 01-Apr-19 11:47, Leon Romanovsky wrote: > >>>>> From: Mark Zhang <markz@xxxxxxxxxxxx> > >>>>> > >>>>> In auto mode all QPs belong to one category are bind automatically > >>>>> to a single counter set. Currently only "qp type" is supported. > >>>>> > >>>>> In this mode the qp counter is set in RST2INIT modification, and > >>>>> when a qp is destroyed the counter is unbound. > >>>> > >>>> This seems like a strange lifetime choice, why set it on RST2INIT > >>>> and not on creation? Isn't that very tied to mlx5 implementation? > >>> > >>> It is our attempt to delay depletion of HW objects (counters). They > >>> are limited in number and change RST2INIT means that this QP was > enabled. > >> > >> We can leave it to the individual driver to decide when to do the > >> actual hw allocation, it's better than the core forcing all drivers > >> to do it on QP modification. > > > > Other drivers cal allocate counter whenever they want, in RST2INIT > > they actually connected. > > The bind (on the software level) should happen when the object (QP) is > allocated, if mlx5 driver wants to do the hardware bind on RST2INIT that can > be solved inside the driver. It is not mlx5 driver. It is the IB spec. Rdma device can have multiple ports. stats are per port. RESET->INIT is when the Qp's port binding occurs. Table 96 of IB spec. ibv_create_qp doesn't have port number in it. ibv_modify_qp sets IBV_QP_PORT attribute. > Mark that the QP has a counter bind on create and delay the operation to > mlx5_ib_modify_qp. Don't make the core force it for everyone.
