> From: linux-rdma-owner@xxxxxxxxxxxxxxx <linux-rdma-
> owner@xxxxxxxxxxxxxxx> On Behalf Of Jason Gunthorpe
>
> On Tue, May 28, 2019 at 06:41:28PM +0000, Michal Kalderon wrote:
> > > From: Jason Gunthorpe <jgg@xxxxxxxx>
> > > Sent: Tuesday, May 28, 2019 7:16 PM
> > >
> > > On Tue, May 28, 2019 at 02:24:00PM +0300, Michal Kalderon wrote:
> > >
> > > > +static int qedr_init_user_db_rec(struct ib_udata *udata,
> > > > + struct qedr_dev *dev, struct qedr_userq *q,
> > > > + u64 db_rec_addr, int access, int dmasync) {
> > > > + /* Aborting for non doorbell userqueue (SRQ) */
> > > > + if (db_rec_addr == 0)
> > > > + return 0;
> > > > +
> > > > + q->db_rec_addr = db_rec_addr;
> > > > + q->db_rec_umem = ib_umem_get(udata, q->db_rec_addr,
> > > PAGE_SIZE,
> > > > + access, dmasync);
> > > > +
> > > > + if (IS_ERR(q->db_rec_umem)) {
> > > > + DP_ERR(dev,
> > > > + "create user queue: failed db_rec ib_umem_get, error
> > > was %ld, db_rec_addr was %llx\n",
> > > > + PTR_ERR(q->db_rec_umem), db_rec_addr);
> > > > + return PTR_ERR(q->db_rec_umem);
> > > > + }
> > > > +
> > > > + q->db_rec_page = sg_page(q->db_rec_umem->sg_head.sgl);
> > > > + q->db_rec_virt = kmap(q->db_rec_page);
> > >
> > > Is this something new? You are much better to use user-triggered
> > > mmap to get a shared page than to use long term kmap.
> >
> > This was the fix for previously using sg_virt which as you stated won't
> always work.
> > Just to make sure I understand, by user-triggered mmap do you mean
> > allocating the memory in kernel and passing the physical pointer to
> > user to mmap it ?
>
> Yes, if the ABI allows for it, this is a better choice for this kind of long lived
> usage.
>
Ok thanks, will take a look at it.
> > > > cq->ibcq.cqe = chain_entries;
> > > > + cq->q.db_addr = (void __iomem *)(uintptr_t)ctx->dpi_addr +
> > > > + db_offset;
> > >
> > > Seems like something has gone wrong here if you have to type __iomem
> > > like this
> > The dpi_addr is an io address to the doorbell-bar received from the
> > qed module, the qed/qedr interface passes it as a u64 ( it is casted
> > from u8 __iomem * to u64) so I need to cast it back.
>
> Don't cast __iommem * to u8. Make a patch to fix it.
>
Ok
> Jason
