lkml.org 
[lkml]   [2015]   [Apr]   [14]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH v3 07/28] IB/Verbs: Reform IB-ulp ipoib
On Tue, Apr 14, 2015 at 11:25:15AM -0600, Jason Gunthorpe wrote:
> On Tue, Apr 14, 2015 at 10:18:07AM -0400, ira.weiny wrote:
>
> > After more thought and reading other opinions, I must agree we should not
> > have cap_foo_dev.
>
> I looked at it a bit, and I think Sean has also basically said, CM
> does not support certain mixed port combinations. iWarp and IB simply
> cannot be mixed with the current CM and it doesn't look easy to fix
> that. We can fix a few point areas simply, but not all of it.
>
> So we have to have the _dev tests, only to fill the CM's need and they
> must have the all true/all false/BUG semantics CM demands.
>
> Verify on register.
>
> > While the ports in ib_sa and ib_umad probably can be orthogonal the current
> > implementation does not support that and this patch series obscures that a bit.
>
> Really? Do you see any bugs/missed things? Both were made port
> orthogonal when RoCEE was added, because RoCEE needs that.

They are not completely orthogonal:

A failure to init port 2 ends up ends up "killing" port 1 and releasing the
device associated resources.

static void ib_umad_add_one(struct ib_device *device)
{
...
if (ib_umad_init_port(device, i, umad_dev,
&umad_dev->port[i - s]))
goto err;
...

err:
while (--i >= s) {
if (!cap_ib_mad(device, i))
continue;

ib_umad_kill_port(&umad_dev->port[i - s]);
}

kobject_put(&umad_dev->kobj);
}

>
> CM wasn't because RoCEE and IB seem to use almost the same code,
> though I wonder if mixing really works 100%..

The support can (and should) be orthogonal but the implementation is
incomplete.

Ira

>
> Jason


\
 
 \ /
  Last update: 2015-04-14 20:01    [W:0.109 / U:3.968 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site