On Mon, Mar 12, 2018 at 04:19:01PM -0500, Steve Wise wrote:
> > On Mon, Mar 12, 2018 at 11:14:28AM -0600, Jason Gunthorpe wrote:
> > > On Mon, Mar 12, 2018 at 12:03:55PM -0500, Steve Wise wrote:
> > > > >
> > > > > On Mon, Mar 12, 2018 at 06:05:31PM +0200, Leon Romanovsky wrote:
> > > > > > On Mon, Mar 12, 2018 at 09:22:32AM -0600, Jason Gunthorpe wrote:
> > > > > > > On Mon, Mar 12, 2018 at 04:16:16PM +0200, Leon Romanovsky wrote:
> > > > > > > > From: Leon Romanovsky <leonro@xxxxxxxxxxxx>
> > > > > > > >
> > > > > > > > Hi,
> > > > > > > >
> > > > > > > > The reason to this patchset is described in "RDMA/mlx4: Clear
> all
> > > > > > > > allocated memory by default" accommodated with relevant crash
> > > > report.
> > > > > > >
> > > > > > > Well, but you never explained what the actual problem is, and
> I've
> > > > failed
> > > > > > > to guess...
> > > > > >
> > > > > > 109 void rdma_restrack_add(struct rdma_restrack_entry *res)
> > > > > > 110 {
> > > > > > 111 struct ib_device *dev = res_to_dev(res);
> > > > > > 112
> > > > > > 113 if (!dev)
> > > > > > 114 return;
> > > > > > 115
> > > > > > 116 if (res_is_user(res)) {
> > > > > > 117 if (!res->task)
> > > > > >
> > > > > > ^^^^^^^^^^^ this part wasn't zero in new
> objects,
> > > > > > sometimes it failed in QP (mlx4), sometimes it failed in CQ (mlx5)
> and
> > > > my
> > > > > > grep showed that other drivers in worst situation.
> > > > >
> > > > > So this is only broken recently by 00313983 ?
> > > > >
> > > > > That is a pretty big change to require all users to provide zero'd
> > > > > memory.. There are certainly simpler ways to fix this than changing
> > > > > every driver..
> > > > >
> > > > > What do you think Steve?
> > > > >
> > > >
> > > > Perhaps an rdma_restrack_entry_init()? It would be called everywhere
> the
> > > > restrack_entry is embedded in an object struct.
> > >
> > > I was thinking simply a flag to rdma_restrack_add that lets the caller
> > > specify the kernel_name and task are inited...
> >
> > It will look awful.
> >
>
> Hiding the initialization of restrack->task via kzalloc() of the parent
> struct is pretty awful too, I think.
Yes, I agree. Very hhard to maintain. I don't like to trust drivers to
do the right thing. They usually don't.
Given all the pushback in the last few hours, this looks NAK'd to me..
Jason
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
