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

[Igor Mammedov]

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.
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.

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.

> 
> 
> [...]
> > > > @@ -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;
> > > >          }  
>