@@ -50,6 +50,23 @@ not cause harm to the host, their actual behavior is not guaranteed by
the API. See "General description" for details on the ioctl usage
model that is supported by KVM.
+It is important to note that althought VM ioctls may only be issued from
+the process that created the VM, a VM's lifecycle is associated with its
+file descriptor, not its creator (process). In other words, the VM and
+its resources, *including the associated address space*, are not freed
+until the last reference to the VM's file descriptor has been released.
+For example, if fork() is issued after ioctl(KVM_CREATE_VM), the VM will
+not be freed until both the parent (original) process and its child have
+put their references to the VM's file descriptor.
+
+Because a VM's resources are not freed until the last reference to its
+file descriptor is released, creating additional references to a VM via
+via fork(), dup(), etc... without careful consideration is strongly
+discouraged and may have unwanted side effects, e.g. memory allocated
+by and on behalf of the VM's process may not be freed/unaccounted when
+the VM is shut down.
+
+
3. Extensions
-------------
The series to add memcg accounting to KVM allocations[1] states: There are many KVM kernel memory allocations which are tied to the life of the VM process and should be charged to the VM process's cgroup. While it is correct to account KVM kernel allocations to the cgroup of the process that created the VM, it's technically incorrect to state that the KVM kernel memory allocations are tied to the life of the VM process. This is because the VM itself, i.e. struct kvm, is not tied to the life of the process which created it, rather it is tied to the life of its associated file descriptor. In other words, kvm_destroy_vm() is not invoked until fput() decrements its associated file's refcount to zero. A simple example is to fork() in Qemu and have the child sleep indefinitely; kvm_destroy_vm() isn't called until Qemu closes its file descriptor *and* the rogue child is killed. The allocations are guaranteed to be *accounted* to the process which created the VM, but only because KVM's per-{VM,vCPU} ioctls reject the ioctl() with -EIO if kvm->mm != current->mm. I.e. the child can keep the VM "alive" but can't do anything useful with its reference. Note that because 'struct kvm' also holds a reference to the mm_struct of its owner, the above behavior also applies to userspace allocations. Given that mucking with a VM's file descriptor can lead to subtle and undesirable behavior, e.g. memcg charges persisting after a VM is shut down, explicitly document a VM's lifecycle and its impact on the VM's resources. Alternatively, KVM could aggressively free resources when the creating process exits, e.g. via mmu_notifier->release(). However, mmu_notifier isn't guaranteed to be available, and freeing resources when the creator exits is likely to be error prone and fragile as KVM would need to ensure that it only freed resources that are truly out of reach. In practice, the existing behavior shouldn't be problematic as a properly configured system will prevent a child process from being moved out of the appropriate cgroup hierarchy, i.e. prevent hiding the process from the OOM killer, and will prevent an unprivileged user from being able to to hold a reference to struct kvm via another method, e.g. debugfs. [1]https://patchwork.kernel.org/patch/10806707/ Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com> --- Documentation/virtual/kvm/api.txt | 17 +++++++++++++++++ 1 file changed, 17 insertions(+)