Re: [PATCH] migration/mm: Add WARN_ON to try_offline_node

On Tue 02-10-18 12:45:50, Tyrel Datwyler wrote:

On 10/02/2018 11:13 AM, Michael Bringmann wrote:

quoted

On 10/02/2018 11:04 AM, Michal Hocko wrote:

quoted

On Tue 02-10-18 10:14:49, Michael Bringmann wrote:

quoted

On 10/02/2018 09:59 AM, Michal Hocko wrote:

quoted

On Tue 02-10-18 09:51:40, Michael Bringmann wrote:
[...]

quoted

When the device-tree affinity attributes have changed for memory,
the 'nid' affinity calculated points to a different node for the
memory block than the one used to install it, previously on the
source system.  The newly calculated 'nid' affinity may not yet
be initialized on the target system.  The current memory tracking
mechanisms do not record the node to which a memory block was
associated when it was added.  Nathan is looking at adding this
feature to the new implementation of LMBs, but it is not there
yet, and won't be present in earlier kernels without backporting a
significant number of changes.

Then the patch you have proposed here just papers over a real issue, no?
IIUC then you simply do not remove the memory if you lose the race.

The problem occurs when removing memory after an affinity change
references a node that was previously unreferenced.  Other code
in 'kernel/mm/memory_hotplug.c' deals with initializing an empty
node when adding memory to a system.  The 'removing memory' case is
specific to systems that perform LPM and allow device-tree changes.
The powerpc kernel does not have the option of accepting some PRRN
requests and accepting others.  It must perform them all.

I am sorry, but you are still too cryptic for me. Either there is a
correctness issue and the the patch doesn't really fix anything or the
final race doesn't make any difference and then the ppc code should be
explicit about that. Checking the node inside the hotplug core code just
looks as a wrong layer to mitigate an arch specific problem. I am not
saying the patch is a no-go but if anything we want a big fat comment
explaining how this is possible because right now it just points to an
incorrect API usage.

That being said, this sounds pretty much ppc specific problem and I
would _prefer_ it to be handled there (along with a big fat comment of
course).

Let me try again.  Regardless of the path to which we get to this condition,
we currently crash the kernel.  This patch changes that to a WARN_ON notice
and continues executing the kernel without shutting down the system.  I saw
the problem during powerpc testing, because that is the focus of my work.
There are other paths to this function besides powerpc.  I feel that the
kernel should keep running instead of halting.

This is still basically a hack to get around a known race. In itself
this patch is still worth while in that we shouldn't crash the kernel
on a null pointer dereference. However, I think the actual problem
still needs to be addressed. We shouldn't run any PRRN events for the
source system on the target after a migration. The device tree update
should have taken care of telling us about new affinities and what
not. Can we just throw out any queued PRRN events when we wake up on
the target?

And until a proper fix is developed can we have NODE_DATA test in the
affected code rather than pollute the generic code with something that
is essentially a wrong usage of the API? With a big fat warning
explaining what is going on here?
-- 
Michal Hocko
SUSE Labs