Re: 9 TiB vm memory creation

[David Hildenbrand]

On 15.02.22 10:40, Ani Sinha wrote:
> On Tue, Feb 15, 2022 at 2:08 PM David Hildenbrand <david@redhat.com> wrote:
>>
>> On 15.02.22 09:12, Ani Sinha wrote:
>>> On Tue, Feb 15, 2022 at 1:25 PM David Hildenbrand <david@redhat.com> wrote:
>>>>
>>>> On 15.02.22 08:00, Ani Sinha wrote:
>>>>>
>>>>>
>>>>> On Mon, 14 Feb 2022, David Hildenbrand wrote:
>>>>>
>>>>>> On 14.02.22 13:36, Igor Mammedov wrote:
>>>>>>> On Mon, 14 Feb 2022 10:54:22 +0530 (IST)
>>>>>>> Ani Sinha <ani@anisinha.ca> wrote:
>>>>>>>
>>>>>>>> Hi Igor:
>>>>>>>>
>>>>>>>> I failed to spawn a 9 Tib VM. The max I could do was a 2 TiB vm on my
>>>>>>>> system with the following commandline before either the system
>>>>>>>> destabilized or the OOM killed killed qemu
>>>>>>>>
>>>>>>>> -m 2T,maxmem=9T,slots=1 \
>>>>>>>> -object 
>>>>>>>> memory-backend-file,id=mem0,size=2T,mem-path=/data/temp/memfile,prealloc=off
>>>>>>>>  \
>>>>>>>> -machine memory-backend=mem0 \
>>>>>>>> -chardev file,path=/tmp/debugcon2.txt,id=debugcon \
>>>>>>>> -device isa-debugcon,iobase=0x402,chardev=debugcon \
>>>>>>>>
>>>>>>>> I have attached the debugcon output from 2 TiB vm.
>>>>>>>> Is there any other commandline parameters or options I should try?
>>>>>>>>
>>>>>>>> thanks
>>>>>>>> ani
>>>>>>>
>>>>>>> $ truncate -s 9T 9tb_sparse_disk.img
>>>>>>> $ qemu-system-x86_64 -m 9T \
>>>>>>>   -object 
>>>>>>> memory-backend-file,id=mem0,size=9T,mem-path=9tb_sparse_disk.img,prealloc=off,share=on
>>>>>>>  \
>>>>>>>   -machine memory-backend=mem0
>>>>>>>
>>>>>>> works for me till GRUB menu, with sufficient guest kernel
>>>>>>> persuasion (i.e. CLI limit ram size to something reasonable) you can 
>>>>>>> boot linux
>>>>>>> guest on it and inspect SMBIOS tables comfortably.
>>>>>>>
>>>>>>>
>>>>>>> With KVM enabled it bails out with:
>>>>>>>    qemu-system-x86_64: kvm_set_user_memory_region: 
>>>>>>> KVM_SET_USER_MEMORY_REGION failed, slot=1, start=0x100000000, 
>>>>>>> size=0x8ff40000000: Invalid argument
>>>>>>>
>>>>>
>>>>> I have seen this in my system but not always. Maybe I should have dug
>>>>> deeper as to why i do see this all the time.
>>>>>
>>>>>>> all of that on a host with 32G of RAM/no swap.
>>>>>>>
>>>>>
>>>>> My system in 16 Gib of main memory, no swap.
>>>>>
>>>>>>
>>>>>> #define KVM_MEM_MAX_NR_PAGES ((1UL << 31) - 1)
>>>>>>
>>>>>> ~8 TiB (7,999999)
>>>>>
>>>>> That's not 8 Tib, thats 2 GiB. But yes, 0x8ff40000000 is certainly greater
>>>>> than 2 Gib * 4K (assuming 4K size pages).
>>>>
>>>> "pages" don't carry the unit "GiB/TiB", so I was talking about the
>>>> actual size with 4k pages (your setup, I assume)
>>>
>>> yes I got that after reading your email again.
>>> The interesting question now is how is redhat QE running 9 TiB vm with kvm?
>>
>> As already indicated by me regarding s390x only having single large NUMA
>> nodes, x86 is usually using multiple NUMA nodes with such large memory.
>> And QE seems to be using virtual numa nodes:
>>
>> Each of the 32 virtual numa nodes receive a:
>>
>>   -object memory-backend-ram,id=ram-node20,size=309237645312,host-
>>    nodes=0-31,policy=bind
>>
>> which results in a dedicated KVM memslot (just like each DIMM would)
>>
>>
>> 32 * 309237645312 == 9 TiB :)
> 
> ah, I should have looked closely at the other commandlines before
> shooting off the email. Yes the limitation is per mem-slot and they
> have 32 slots one per node.
> ok so should we do
> kvm_set_max_memslot_size(KVM_SLOT_MAX_BYTES);
> from i386 kvm_arch_init()?


As I said, I'm not a friend of these workarounds in user space.

Assume you have one KVM memslot left and you hotplug a huge DIMM that
will consume more than one KVM memslot -- you're in trouble, because
hotplug will succeed but creating the second memslot will fail. So you
need additional logic in memory device code to special-case on these
corner cases.

We should try increasing the limit in KVM and handle it gracefully in
QEMU. But that's just my 2 cents.

-- 
Thanks,

David / dhildenb