On 02.11.18 11:10, Okash Khawaja wrote: > On Wed, Oct 31, 2018 at 1:17 AM Okash Khawaja <okash.khawaja@xxxxxxxxx> wrote: >> >> Hi, >> >> Apologies if this has already been discussed but hasn't crossed my >> eyes. It seems that KVM_CREATE_VM ioctl will fail with -ENOMEM if the >> call to kvm_create_vm_debugfs() returns negative value. Is debugfs >> essential for KVM operation? >> >> I came across this when testing with a bpf-next kernel I manually >> compiled, from around v4.17. Creating a VM from QEMU threw following >> error: >> >> ioctl(KVM_CREATE_VM) failed: 12 Cannot allocate memory >> >> Running 4.19.0 kernel fixed the problem. Please note that >> kvm_create_vm_debugfs() would not fail if debugfs hasn't been >> registered at all. >> >> Thanks, >> Okash > > `kvm_create_vm_debugfs()` currently fails only if it fails to allocate > memory so in a way it does make sense that KVM_CREATE_VM ioctl fails > with -ENOMEM. However, failing to create VM because debugfs failed > doesn't seem optimal to me. For example, let's say we change > `kvm_create_vm_debugfs()` to add more error exit conditions which > aren't memory allocation related, then still the ioctl will fail with > -ENOMEM, unless we change the calling code to check return value of > `kvm_create_vm_debugfs()`. Also, if the memory pressure is transient > and `kvm_create_vm_debugfs()` fails because of that, the whole VM > creation shouldn't fail. But most importantly, since VM debugfs stat > isn't necessary for VM to function, VM creation shouldn't fail. It's okay to ignore corner cases if it makes life easier (e.g. handling debugfs created or not when tearing down a VM). > > Please correct me if I'm wrong with my reasoning above. Ultimately, we > should check errors in `kvm_create_vm_debugfs()` and clean up files > and directories if needed and not fail KVM_CREATE_VM ioctl. If this > makes sense, I can put together a patch :) > Is this really worth investing time and effort? Why I am asking: Your kernel cannot even allocate memory for debugfs. How are you supposed to do anything with that VM? Maybe you cannot even create a VCPU. Maybe the OOM killer will zap your process. Is there any reasonable setup where this would be expected to work? "memory pressure is transient" is not a convincing argument. This sounds like a broken setup to me. -- Thanks, David / dhildenb
