Re: [Qemu-devel] [PATCH for-3.0 v2] pc: acpi: fix memory hotplug regression by reducing stub SRAT entry size

[Eduardo Habkost]

On Thu, Aug 02, 2018 at 12:09:37PM +0200, Igor Mammedov wrote:
> On Tue, 31 Jul 2018 12:03:22 -0300
> Eduardo Habkost <address@hidden> wrote:
> 
> > On Tue, Jul 31, 2018 at 11:53:40AM +0200, Igor Mammedov wrote:
> > > On Mon, 30 Jul 2018 17:26:24 -0300
> > > Eduardo Habkost <address@hidden> wrote:
> > >   
> > > > On Mon, Jul 30, 2018 at 11:41:41AM +0200, Igor Mammedov wrote:  
> > > > > Commit 848a1cc1e (hw/acpi-build: build SRAT memory affinity 
> > > > > structures for DIMM devices)
> > > > > broke the first dimm hotplug in following cases:
> > > > > 
> > > > >  1: there is no coldplugged dimm in the last numa node
> > > > >     but there is a coldplugged dimm in another node
> > > > > 
> > > > >   -m 4096,slots=4,maxmem=32G               \
> > > > >   -object memory-backend-ram,id=m0,size=2G \
> > > > >   -device pc-dimm,memdev=m0,node=0         \
> > > > >   -numa node,nodeid=0                      \
> > > > >   -numa node,nodeid=1
> > > > > 
> > > > >  2: if order of dimms on CLI is:
> > > > >        1st plugged dimm in node1
> > > > >        2nd plugged dimm in node0
> > > > > 
> > > > >   -m 4096,slots=4,maxmem=32G               \
> > > > >   -object memory-backend-ram,size=2G,id=m0 \
> > > > >   -device pc-dimm,memdev=m0,node=1         \
> > > > >   -object memory-backend-ram,id=m1,size=2G \
> > > > >   -device pc-dimm,memdev=m1,node=0         \
> > > > >   -numa node,nodeid=0                      \
> > > > >   -numa node,nodeid=1
> > > > > 
> > > > > (qemu) object_add memory-backend-ram,id=m2,size=1G
> > > > > (qemu) device_add pc-dimm,memdev=m2,node=0
> > > > > 
> > > > > the first DIMM hotplug to any node except the last one
> > > > > fails (Windows is unable to online it).
> > > > > 
> > > > > Length reduction of stub hotplug memory SRAT entry,
> > > > > fixes issue for some reason.
> > > > >     
> > > > 
> > > > I'm really bothered by the lack of automated testing for all
> > > > these NUMA/ACPI corner cases.
> > > > 
> > > > This looks like a good candidate for an avocado_qemu test case.
> > > > Can you show pseudo-code of how exactly the bug fix could be
> > > > verified, so we can try to write a test case?  
> > > Sadly I do it manually every time I'm suspect a patch would
> > > affect the feature. On just has to check if a new memory device
> > > appeared in device manager and it is in working state (started 
> > > successfully).
> > > One also need to run it against to test it against windows version
> > > that supports memory hot-add (DC ed.).
> > > 
> > > It's typically what RHEL QE does, and they just found
> > > a new case which wasn't on test list so proactive measures
> > > wouldn't work here in any case as we didn't know about
> > > this particular combination.
> > > 
> > > I'm not sure how it will work with upstream avocado though,
> > > windows testing implies tester would need access to MSDN
> > > subscription or multiple retail versions to test against.
> > > So with windows it becomes expensive and complicated
> > > hence I'd leave this job to QE which has resources and
> > > upstream would benefit from downstream when a bug is found
> > > (albeit it's a catch up game).  
> > 
> > I don't mean functional testing of Windows guests.  I'm just
> > looking for a way we can ensure we won't reintroduce this
> > particular bug later.  We should be able to encode known
> > requirements of existing guest OSes in test code (especially the
> > undocumented requirements).
> >
> > In other words, we need test code that will check if the entry
> > you are adding below is still present and contains the right
> > flags, so people won't remove it by mistake.
> known requirements are described in acpi code comment and commit
> messages and maintainer are supposed to check if a change showed
> by bios test is valid and doesn't regress existing state.

I really believe computers are better at that than humans.  If we
can't even describe to a computer how to look for mistakes, I
don't think we can expect humans to spot them.


> Parsing SRAT in test and ensuring that the last entry hasn't changed
> won't help, we already have this by doing comparison with reference
> SRAT.

Comparison against reference SRAT is useless when a patch is
expected to change the ACPI tables, which happens all the time.
I think we can do better than that.

> 
> And if there is a change, the only thing that can somewhat verify
> it is a functional test with windows (known combinations at
> least). Some new sequence/combination might regress it again
> (like one described in commit). An Avocado functional test running
> windows(es) might help if it will test random startup/hotplug combinations,
> run by someone with rights to use windows.
> 
> I think that once I've contributed cpu hotplug testcases to autotest
> but then there appears a new test suite and then another.
> I don't really feel nor have capacity to deal with it, if someone
> contributes testcase to Avocado and tells me how to easily use it,
> I'd gladly run it with windows guests I have access to
> whenever I review/test a patch that might affect windows.

Running tests with real guests is useful, but costly.  I would
like to be able to detect when we break known requirements
without running a guest.

I'm not asking you to write that test code right now, but I'm
trying to see if it's possible to do that and use it as reference
for testing future ACPI patches.  Let's see if I can enumerate
the known requirements for this stub entry that are not
documented in the ACPI spec:

1) Windows expect a stub entry with
   MEM_AFFINITY_HOTPLUGGABLE|MEM_AFFINITY_ENABLED at the end of
   the hotpluggable address space.
2) Linux won't enable SWIOTLB when booted with less 4GB unless
   we add a stub entry to cover the whole hotpluggable address
   space.
3) Windows expect that entry to be bound to the last NUMA node.
4) Windows won't behave well if the stub entry covers the whole
   hotpluggable address space.  We just need to be at the end of
   the address space, leaving a hole between the last DIMM and
   this stub entry.  We add a 1-byte stub entry for that.

The last one is a new requirement we didn't know about, right?
Is my description accurate?  What else we know about Windows
expectations?


> [...]
> > > > @@ -2269,7 +2269,16 @@ static void 
> > > > build_srat_hotpluggable_memory(GArray *table_data, uint64_t base,
> > > >          numamem = acpi_data_push(table_data, sizeof *numamem);
> > > >  
> > > >          if (!info) {
> > > > -            build_srat_memory(numamem, cur, end - cur, default_node,
> > > > +            /*
> > > > +             * Entry is required for Windows to enable memory hotplug 
> > > > in OS
> > > > +             * and for Linux to enable SWIOTLB when booted with less 
> > > > than
> > > > +             * 4G of RAM. Windows works better if the entry sets 
> > > > proximity
> > > > +             * to the highest NUMA node in the machine at the end of 
> > > > the
> > > > +             * reserved space.
> > > > +             * Memory devices may override proximity set by this entry,
> > > > +             * providing _PXM method if necessary.
> > > > +             */
> > > > +            build_srat_memory(numamem, end - 1, 1, default_node,
> > > >                                MEM_AFFINITY_HOTPLUGGABLE | 
> > > > MEM_AFFINITY_ENABLED);
> > > >              break;
> > > >          }

-- 
Eduardo