On 27.03.19 10:03, David Gibson wrote: > On Wed, Mar 27, 2019 at 09:10:01AM +0100, David Hildenbrand wrote: >> On 27.03.19 01:12, David Gibson wrote: >>> On Tue, Mar 26, 2019 at 06:02:51PM +0100, David Hildenbrand wrote: >>>> On 26.03.19 15:08, Igor Mammedov wrote: >>>>> On Tue, 26 Mar 2019 14:50:58 +1100 >>>>> David Gibson <da...@gibson.dropbear.id.au> wrote: >>>>> >>>>>> qemu_getrampagesize() works out the minimum host page size backing any of >>>>>> guest RAM. This is required in a few places, such as for POWER8 PAPR KVM >>>>>> guests, because limitations of the hardware virtualization mean the guest >>>>>> can't use pagesizes larger than the host pages backing its memory. >>>>>> >>>>>> However, it currently checks against *every* memory backend, whether or >>>>>> not >>>>>> it is actually mapped into guest memory at the moment. This is >>>>>> incorrect. >>>>>> >>>>>> This can cause a problem attempting to add memory to a POWER8 pseries KVM >>>>>> guest which is configured to allow hugepages in the guest (e.g. >>>>>> -machine cap-hpt-max-page-size=16m). If you attempt to add non-hugepage, >>>>>> you can (correctly) create a memory backend, however it (correctly) will >>>>>> throw an error when you attempt to map that memory into the guest by >>>>>> 'device_add'ing a pc-dimm. >>>>>> >>>>>> What's not correct is that if you then reset the guest a startup check >>>>>> against qemu_getrampagesize() will cause a fatal error because of the new >>>>>> memory object, even though it's not mapped into the guest. >>>>> I'd say that backend should be remove by mgmt app since device_add failed >>>>> instead of leaving it to hang around. (but fatal error either not a nice >>>>> behavior on QEMU part) >>>> >>>> Indeed, it should be removed. Depending on the options (huge pages with >>>> prealloc?) memory might be consumed for unused memory. Undesired. >>> >>> Right, but if the guest initiates a reboot before the management gets >>> to that, we'll have a crash. >>> >> >> Yes, I agree. >> >> At least on s390x (extending on what Igor said): >> >> mc->init() -> s390_memory_init() -> >> memory_region_allocate_system_memory() -> host_memory_backend_set_mapped() >> >> >> ac->init_machine() -> kvm_arch_init() -> >> kvm_s390_configure_mempath_backing() -> qemu_getrampagesize() >> >> >> And in vl.c >> >> configure_accelerator(current_machine, argv[0]); >> ... >> machine_run_board_init() >> >> So memory is indeed not mapped before calling qemu_getrampagesize(). >> >> >> We *could* move the call to kvm_s390_configure_mempath_backing() to >> s390_memory_init(). >> >> cap_hpage_1m is not needed before we create VCPUs, so this would work fine. >> >> We could than eventually make qemu_getrampagesize() asssert if no >> backends are mapped at all, to catch other user that rely on this being >> correct. > > So.. I had a look at the usage in kvm_s390_configure_mempath_backing() > and I'm pretty sure it's broken. It will work in the case where > there's only one backend. And if that's the default -mem-path rather > than an explicit memory backend then my patch won't break it any > further.
On the second look, I think I get your point. 1. Why on earth does "find_max_supported_pagesize" find the "minimum page size". What kind of nasty stuff is this. 2. qemu_mempath_getpagesize() is not affected by your patch and that seems to be the only thing used on s390x for now. I sent a patch to move the call on s390x. But we really have to detect the maximum page size (what find_max_supported_pagesize promises), not the minimum page size. > > qemu_getrampagesize() returns the smallest host page size for any > memory backend. That's what matters for pcc KVM (in several cases) > because we need certain things to be host-contiguous, not just > guest-contiguous. Bigger host page sizes are fine for that purpose, > clearly. > > AFAICT on s390 you're looking to determine if any backend is using > hugepages, because KVM may not support that. The minimum host page > size isn't adequate to determine that, so qemu_getrampagesize() won't > tell you what you need. > Indeed. -- Thanks, David / dhildenb