Re: [RFC 0/7] Peer-direct memory

[Sagi Grimberg <sagig-LDSdmyG8hGV8YrgS2mwiifqBs+8SCbDb@public.gmane.org>]

>> Agreed on all points.  Nevermind that seems to be missing a user of all
>> the newly added infrastructure, which should come with the series.
>
> Yep!
>
> I've heard from the people working on NVMe cards that this approach
> isn't needed, desired, or correct. I'm deeply skeptical that there
> would be an in-kernel user.

Hey Jason,

So regarding NVMe, a possible user for this would be a user-space NVMe
target implementation which wants to use a CMB (controller memory
buffer on the device bar) as an intermediate memory for RDMA. In case
the NVMe unplugs we can't rely on user-space to handle it, and need a
way to tell the HCA to tear down the CMB mappings.

I do think that this problem is solvable if we create a mmu_notitifier
client for this mapping (assuming that device unplug will trigger mmu
invalidate callouts on the relevant pages). But I'm still not exactly
sure how ZONE_DEVICE works with I/O device bar though (I'll appreciate
if someone can educate me here).

CC'ing sbates who played with this stuff at some point...
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html