On Tue, Jun 30, 2020 at 05:13:54PM +0200, Vitaly Kuznetsov wrote: > Vivek Goyal <vgoyal@xxxxxxxxxx> writes: > > > On Tue, Jun 30, 2020 at 03:24:43PM +0200, Vitaly Kuznetsov wrote: > > >> > >> It's probably me who's missing something important here :-) but I think > >> you describe how it *should* work as I'm not seeing how we can leave the > >> loop in kvm_async_pf_task_wait_schedule() other than by > >> "if (hlist_unhashed(&n.link)) break;" and this only happens when APF > >> completes. > > > > We don't leave loop in kvm_async_pf_task_wait_schedule(). It will happen > > before you return to user space. > > > > I have not looked too closely but I think following code path might be taken > > after aync PF has completed. > > > > __kvm_handle_async_pf() > > idtentry_exit_cond_rcu() > > prepare_exit_to_usermode() > > __prepare_exit_to_usermode() > > exit_to_usermode_loop() > > do_signal() > > > > So once you have been woken up (because APF completed), > > Ah, OK so we still need to complete APF and we can't kill the process > before this happens, that's what I was missing. > > > you will > > return to user space and before that you will check if there are > > pending signals and handle that signal first before user space > > gets a chance to run again and retry faulting instruction. > > ... > > > > >> > >> When guest receives the 'page ready' event with an error it (like for > >> every other 'page ready' event) tries to wake up the corresponding > >> process but if the process is dead already it can do in-kernel probing > >> of the GFN, this way we guarantee that the error is always injected. I'm > >> not sure if it is needed though but in case it is, this can be a > >> solution. We can add a new feature bit and only deliver errors when the > >> guest indicates that it knows what to do with them. > > > > - Process will be delivered singal after async PF completion and during > > returning to user space. You have lost control by then. > > > > So actually there's no way for kernel to know if the userspace process > managed to re-try the instruction and get the error injected or if it > was killed prior to that. Yes. > > > - If you retry in kernel, we will change the context completely that > > who was trying to access the gfn in question. We want to retain > > the real context and retain information who was trying to access > > gfn in question. > > (Just so I understand the idea better) does the guest context matter to > the host? Or, more specifically, are we going to do anything besides > get_user_pages() which will actually analyze who triggered the access > *in the guest*? When we exit to user space, qemu prints bunch of register state. I am wondering what does that state represent. Does some of that traces back to the process which was trying to access that hva? I don't know. I think keeping a cache of error gfns might not be too bad from implemetation point of view. I will give it a try and see how bad does it look. Vivek
