Re: [Qemu-devel] [PATCH] pc: acpi: revert back to 1 SRAT entry for hotpluggable area

[Igor Mammedov]

On Thu, 23 Aug 2018 17:01:33 +0800
Yu Zhang <address@hidden> wrote:

> On 8/23/2018 2:01 AM, Eduardo Habkost wrote:
> > On Wed, Aug 22, 2018 at 03:05:36PM +0200, Igor Mammedov wrote:  
> >> On Wed, 22 Aug 2018 12:06:26 +0200
> >> Laszlo Ersek <address@hidden> wrote:
> >>  
> >>> On 08/22/18 11:46, Igor Mammedov wrote:  
> >>>> Commit
> >>>>    10efd7e108 "pc: acpi: fix memory hotplug regression by reducing stub 
> >>>> SRAT entry size"
> >>>> attemped to fix hotplug regression introduced by
> >>>>    848a1cc1e "hw/acpi-build: build SRAT memory affinity structures for 
> >>>> DIMM devices"
> >>>>
> >>>> fixed issue for Windows/3.0+ linux kernels, however it regressed 2.6 
> >>>> based
> >>>> kernels (RHEL6) to the point where guest might crash at boot.
> >>>> Reason is that 2.6 kernel discards SRAT table due too small last entry
> >>>> which down the road leads to crashes. Hack I've tried in 10efd7e108 is 
> >>>> also
> >>>> not ACPI spec compliant according to which whole possible RAM should be
> >>>> described in SRAT. Revert 10efd7e108 to fix regression for 2.6 based 
> >>>> kernels.
> >>>>
> >>>> With 10efd7e108 reverted, I've also tried splitting SRAT table statically
> >>>> in different ways %/node and %/slot but Windows still fails to online
> >>>> 2nd pc-dimm hot-plugged into node 0 (as described in 10efd7e108) and
> >>>> sometimes even coldplugged pc-dimms where affected with static SRAT
> >>>> partitioning.
> >>>> The only known so far way where Windows stays happy is when we have 1
> >>>> SRAT entry in the last node covering all hotplug area.
> >>>>
> >>>> Revert 848a1cc1e until we come up with a way to avoid regression
> >>>> on Windows with hotplug area split in several entries.
> >>>> Tested this with 2.6/3.0 based kernels (RHEL6/7) and 
> >>>> WS20[08/12/12R2/16]).
> >>>>
> >>>> Signed-off-by: Igor Mammedov <address@hidden>
> >>>> ---
> >>>> CC: address@hidden
> >>>> CC: address@hidden
> >>>> CC: address@hidden
> >>>> CC: address@hidden
> >>>> CC: address@hidden
> >>>> ---
> >>>>   hw/i386/acpi-build.c | 73 
> >>>> +++++++++-------------------------------------------
> >>>>   1 file changed, 12 insertions(+), 61 deletions(-)
> >>>>
> >>>> diff --git a/hw/i386/acpi-build.c b/hw/i386/acpi-build.c
> >>>> index e1ee8ae..1599caa 100644
> >>>> --- a/hw/i386/acpi-build.c
> >>>> +++ b/hw/i386/acpi-build.c
> >>>> @@ -2251,64 +2251,6 @@ build_tpm2(GArray *table_data, BIOSLinker 
> >>>> *linker, GArray *tcpalog)
> >>>>   #define HOLE_640K_START  (640 * KiB)
> >>>>   #define HOLE_640K_END   (1 * MiB)
> >>>>   
> >>>> -static void build_srat_hotpluggable_memory(GArray *table_data, uint64_t 
> >>>> base,
> >>>> -                                           uint64_t len, int 
> >>>> default_node)
> >>>> -{
> >>>> -    MemoryDeviceInfoList *info_list = qmp_memory_device_list();
> >>>> -    MemoryDeviceInfoList *info;
> >>>> -    MemoryDeviceInfo *mi;
> >>>> -    PCDIMMDeviceInfo *di;
> >>>> -    uint64_t end = base + len, cur, size;
> >>>> -    bool is_nvdimm;
> >>>> -    AcpiSratMemoryAffinity *numamem;
> >>>> -    MemoryAffinityFlags flags;
> >>>> -
> >>>> -    for (cur = base, info = info_list;
> >>>> -         cur < end;
> >>>> -         cur += size, info = info->next) {
> >>>> -        numamem = acpi_data_push(table_data, sizeof *numamem);
> >>>> -
> >>>> -        if (!info) {
> >>>> -            /*
> >>>> -             * 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;
> >>>> -        }
> >>>> -
> >>>> -        mi = info->value;
> >>>> -        is_nvdimm = (mi->type == MEMORY_DEVICE_INFO_KIND_NVDIMM);
> >>>> -        di = !is_nvdimm ? mi->u.dimm.data : mi->u.nvdimm.data;
> >>>> -
> >>>> -        if (cur < di->addr) {
> >>>> -            build_srat_memory(numamem, cur, di->addr - cur, 
> >>>> default_node,
> >>>> -                              MEM_AFFINITY_HOTPLUGGABLE | 
> >>>> MEM_AFFINITY_ENABLED);
> >>>> -            numamem = acpi_data_push(table_data, sizeof *numamem);
> >>>> -        }
> >>>> -
> >>>> -        size = di->size;
> >>>> -
> >>>> -        flags = MEM_AFFINITY_ENABLED;
> >>>> -        if (di->hotpluggable) {
> >>>> -            flags |= MEM_AFFINITY_HOTPLUGGABLE;
> >>>> -        }
> >>>> -        if (is_nvdimm) {
> >>>> -            flags |= MEM_AFFINITY_NON_VOLATILE;
> >>>> -        }
> >>>> -
> >>>> -        build_srat_memory(numamem, di->addr, size, di->node, flags);
> >>>> -    }
> >>>> -
> >>>> -    qapi_free_MemoryDeviceInfoList(info_list);
> >>>> -}
> >>>> -
> >>>>   static void
> >>>>   build_srat(GArray *table_data, BIOSLinker *linker, MachineState 
> >>>> *machine)
> >>>>   {
> >>>> @@ -2414,10 +2356,19 @@ build_srat(GArray *table_data, BIOSLinker 
> >>>> *linker, MachineState *machine)
> >>>>           build_srat_memory(numamem, 0, 0, 0, MEM_AFFINITY_NOFLAGS);
> >>>>       }
> >>>>   
> >>>> +    /*
> >>>> +     * 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.
> >>>> +     * Memory devices may override proximity set by this entry,
> >>>> +     * providing _PXM method if necessary.
> >>>> +     */
> >>>>       if (hotplugabble_address_space_size) {
> >>>> -        build_srat_hotpluggable_memory(table_data, 
> >>>> machine->device_memory->base,
> >>>> -                                       hotplugabble_address_space_size,
> >>>> -                                       pcms->numa_nodes - 1);
> >>>> +        numamem = acpi_data_push(table_data, sizeof *numamem);
> >>>> +        build_srat_memory(numamem, machine->device_memory->base,
> >>>> +                          hotplugabble_address_space_size, 
> >>>> pcms->numa_nodes - 1,
> >>>> +                          MEM_AFFINITY_HOTPLUGGABLE | 
> >>>> MEM_AFFINITY_ENABLED);
> >>>>       }
> >>>>   
> >>>>       build_header(linker, table_data,
> >>>>      
> >>> So this reverts both 10efd7e108 (which only moved the regression around)
> >>> and the earlier 848a1cc1e (which had introduced the regression in the
> >>> first place).
> >>>
> >>> If I understand correctly, the issue that 848a1cc1e was meant to solve
> >>> was that "-device nvdimm,node=..." could be passed by the user such that
> >>> it lead to "proximity domain mismatch". (What was the higher-level goal
> >>> with that BTW?)
> >>>
> >>> However, that mismatch could as well be avoided by simply not passing
> >>> such "node=..." properties. Is that right?
> >>>
> >>> If so, should we perhaps add another patch (temporarily), beyond this
> >>> revert, that identifies the misconfig in question, and prints a warning
> >>> about it?  
> >> not exactly, before the patch node=... influenced only _PXM which is 
> >> supposed
> >> to be evaluated after SRAT table and override whatever was in static table.
> >>
> >> The patch, as I understood it, was about ACPI spec compliance where nvdimm 
> >> SPA
> >> in NFIT table should have a corresponding entry in SRAT table with
> >> MEM_AFFINITY_NON_VOLATILE flag set.
> >> Whether it solves any real issues aren't known to me and typically for
> >> Intel contributed patches, author's email doesn't exists anymore so ...
> >> And even if it fixes some new nvdimm issue, I'd rather have it broken
> >> than regress memory hotplug that worked fine so far and wait for another
> >> nvdimm patch that won't break existing features.
> >>  
> >>> Anyway the revert seems justified to me.  
> >> I've killed quite enough time trying to find out a way to keep
> >> everyone happy, but alas it's time to give up and let interested
> >> party to deal with regressions while introducing new stuff for
> >> nvdimm, hence this revert.  
> > If the original author isn't available, I agree the best option
> > is to revert it to avoid the regression by now.  
> 
> Thanks Edurado, for informing us, and sorry for the delayed reply.
> I'm not the author of the original patch(848a1cc1e), and is just trying 
> to understand
> this part now...
> 
> And my understanding is that the patch was cooked to make sure data in SRAT
> align with the ones in NFIT.
> 
> > Reviewed-by: Eduardo Habkost <address@hidden>
> >
> > However, have you considered keeping adding separate entries for
> > NVDIMM devices only (so we follow the spec), but add a single
> > (numa_nodes-1, MEM_AFFINITY_HOTPLUGGABLE|MEM_AFFINITY_ENABLED)
> > entry to the rest?  
> 
> And I think that is what Igor did in patch 10efd7e108.
> 
> But I feel puzzled why windows has such requirement. And I am also 
> curious, is
> this a windows specific issue?
It's is a Windows issue, but QEMU doesn't differentiate between guests
OSes so in general it should work for all of them or at least do not break
existing features.


> Reverting 848a1cc1e will lead the VM not able to figure out the correct 
> NUMA info,
> e.g. when we are trying to assign different regions from NVDIMMs 
> residing on different
> NUMA node.
that's probably non issue as QEMU's impl. supports only 1:1 mapping,
i.e. SPA for whole nvdimm goes to 1 node and SPA record already
contains node ID it belongs to.

The reason why I've acked 848a1cc1e is that it's correct per ACPI spec,
but unfortunately QE found a test case where it regresses memory hotplug,
hence a revert. We should find out a way to be spec compliant and
not to break existing features.

Perhaps you'd be able to figure out how to accomplish both goals,
I'd be happy to help you out if you'd need some assistance.


> 
> B.R.
> Yu
> 
> > This way both use cases should still work as expected.  If
> > Windows break if using NVDIMM + Memory Hotplug at the same time,
> > maybe that's just a Windows bug we can't work around.
> >