On Fri, 27 Feb 2015 12:09:20 +0800 zhanghailiang <zhang.zhanghailiang@xxxxxxxxxx> wrote: > On 2015/2/27 3:26, Luiz Capitulino wrote: > > Hello, > > > > Reproducer: > > > > 1. Start QEMU with balloon and memory hotplug support: > > > > # qemu [...] -m 1G,slots=2,maxmem=2G -balloon virtio > > > > 2. Check balloon size: > > > > (qemu) info balloon > > balloon: actual=1024 > > (qemu) > > > > 3. Hotplug some memory: > > > > (qemu) object_add memory-backend-ram,id=mem1,size=1G > > (qemu) device_add pc-dimm,id=dimm1,memdev=mem1 > > > > 4. This is step is _not_ needed to reproduce the problem, > > but you may need to online memory manually on Linux so > > that it becomes available in the guest > > > > 5. Check balloon size again: > > > > (qemu) info balloon > > balloon: actual=1024 > > (qemu) > > > > BUG: The guest now has 2GB of memory, but the balloon thinks > > the guest has 1GB > > > > One may think that the problem is that the balloon driver is > > ignoring hotplugged memory. This is not what's happening. If > > you do balloon your guest, there's nothing stopping the > > balloon driver in the guest from ballooning hotplugged memory. > > > > The problem is that the balloon device in QEMU needs to know > > the current amount of memory available to the guest. > > > > Before memory hotplug this information was easy to obtain: the > > current amount of memory available to the guest is the memory the > > guest was booted with. This value is stored in the ram_size global > > variable in QEMU and this is what the balloon device emulation > > code uses today. However, when memory is hotplugged ram_size is > > _not_ updated and the balloon device breaks. > > > > I see two possible solutions for this problem: > > > > 1. In addition to reading ram_size, the balloon device in QEMU > > could scan pc-dimm devices to account for hotplugged memory. > > > > This solution was already implemented by zhanghailiang: > > > > http://lists.gnu.org/archive/html/qemu-devel/2014-11/msg02362.html > > > > It works, except that on Linux memory hotplug is a two-step > > procedure: first memory is inserted then it has to be onlined > > from user-space. So, if memory is inserted but not onlined > > this solution gives the opposite problem: the balloon device > > will report a larger memory amount than the guest actually has. > > > > Can we live with that? I guess not, but I'm open for discussion. > > > > If QEMU could be notified when Linux makes memory online, then > > the problem would be gone. But I guess this can't be done. > > > > Yes, it is really a problem, balloon can't work well with memory block online/offline now. > virtio-balloon can't be notified when memory block online/offline now, actually, we can > add this capability by using the exist kernel memory hotplug/unplug notifier mechanism. ( > just a simple register_memory_notifier().) I'm leaning towards applying your series now and do anything that involves the guest on top if needed. > > 2. Modify the balloon driver in the guest to inform the balloon > > device on the host about the current memory available to the > > guest. This way, whenever the balloon device in QEMU needs > > to know the current amount of memory in the guest, it asks > > the guest. This drops any usage of ram_size in the balloon > > device > > > > I'm not completely sure this is feasible though. For example, > > what happens if the guest reports a memory amount to QEMU and > > right after this more memory is plugged? > > > > Hmm, i wonder why we notify the number of pages which should be adjusted to virtio-balloon, > why not the memory 'target' size ? Is there any special reason ? I don't know either. I guess it's just how the balloon feature was designed. > For linux guest, it can always know exactly its current real memory size, but QEMU may not, because > guest can do online/offline memory block by themselves. > > If virtio-balloon in guest know the balloon's 'target' size, it can calculate the exact memory size > that should be adjuested. and also can do corresponding action (fill or leak balloon) > when there is online/offline memory block occurred. I'm not sure this would work as all sorts of races are possible with memory allocations that may occur during or after the calculation is done. Your series makes the best case work, which is memory is inserted and the guest uses all of it. Today not even the best case works. > > > Besides, this solution is more complex than solution 1 and > > won't address older guests. > > > > Another important detail is that, I *suspect* that a very similar > > bug already exists with 32-bit guests even without memory > > hotplug: what happens if you assign 6GB to a 32-bit without PAE > > support? I think the same problem we're seeing with memory > > hotplug will happen and solution 1 won't fix this, although > > no one seems to care about 32-bit guests... > > > > . > > > > -- To unsubscribe from this list: send the line "unsubscribe kvm" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
