On Wed, Jan 25, 2023 at 12:34:26PM +0100, Petr Mladek wrote: > On Tue 2023-01-24 11:21:39, Seth Forshee wrote: > > On Tue, Jan 24, 2023 at 03:17:43PM +0100, Petr Mladek wrote: > > > On Fri 2023-01-20 16:12:22, Seth Forshee (DigitalOcean) wrote: > > > > Livepatch relies on stack checking of sleeping tasks to switch kthreads, > > > > so a busy kthread can block a livepatch transition indefinitely. We've > > > > seen this happen fairly often with busy vhost kthreads. > > > > > > To be precise, it would be "indefinitely" only when the kthread never > > > sleeps. > > > > > > But yes. I believe that the problem is real. It might be almost > > > impossible to livepatch some busy kthreads, especially when they > > > have a dedicated CPU. > > > > > > > > > > Add a check to call klp_switch_current() from vhost_worker() when a > > > > livepatch is pending. In testing this allowed vhost kthreads to switch > > > > immediately when they had previously blocked livepatch transitions for > > > > long periods of time. > > > > > > > > Signed-off-by: Seth Forshee (DigitalOcean) <sforshee@xxxxxxxxxx> > > > > --- > > > > drivers/vhost/vhost.c | 4 ++++ > > > > 1 file changed, 4 insertions(+) > > > > > > > > diff --git a/drivers/vhost/vhost.c b/drivers/vhost/vhost.c > > > > index cbe72bfd2f1f..d8624f1f2d64 100644 > > > > --- a/drivers/vhost/vhost.c > > > > +++ b/drivers/vhost/vhost.c > > > > @@ -366,6 +367,9 @@ static int vhost_worker(void *data) > > > > if (need_resched()) > > > > schedule(); > > > > } > > > > + > > > > + if (unlikely(klp_patch_pending(current))) > > > > + klp_switch_current(); > > > > > > I suggest to use the following intead: > > > > > > if (unlikely(klp_patch_pending(current))) > > > klp_update_patch_state(current); > > > > > > We already use this in do_idle(). The reason is basically the same. > > > It is almost impossible to livepatch the idle task when a CPU is > > > very idle. > > > > > > klp_update_patch_state(current) does not check the stack. > > > It switches the task immediately. > > > > > > It should be safe because the kthread never leaves vhost_worker(). > > > It means that the same kthread could never re-enter this function > > > and use the new code. > > > > My knowledge of livepatching internals is fairly limited, so I'll accept > > it if you say that it's safe to do it this way. But let me ask about one > > scenario. > > > > Let's say that a livepatch is loaded which replaces vhost_worker(). New > > vhost worker threads are started which use the replacement function. Now > > if the patch is disabled, these new worker threads would be switched > > despite still running the code from the patch module, correct? Could the > > module then be unloaded, freeing the memory containing the code these > > kthreads are executing? > > Great catch! Yes, this might theoretically happen. > > The above scenario would require calling klp_update_patch_state() from > the code in the livepatch module. It is not possible at the moment because > this function is not exported for modules. vhost can be built as a module, so in order to call klp_update_patch_state() from vhost_worker() it would have to be exported to modules. > Hmm, the same problem might be when we livepatch a function that calls > another function that calls klp_update_patch_state(). But in this case > it would be kthread() from kernel/kthread.c. It would affect any > running kthread. I doubt that anyone would seriously think about > livepatching this function. Yes, there are clearly certain functions that are not safe/practical to patch, and authors need to know what they are doing. Most kthread main() functions probably qualify as impractical at best, at least without a strategy to restart relevant kthreads. But a livepatch transition will normally stall if patching these functions when a relevant kthread is running (unless the patch is forced), so a patch author who made a mistake should quickly notice. vhost_worker() would behave differently. > A good enough solution might be to document this. Livepatches could > not be created blindly. There are more situations where the > livepatch is tricky or not possible at all. I can add this if you like. Is Documentation/livepatch/livepatch.rst the right place for this? > Crazy idea. We could prevent this problem even technically. A solution > would be to increment a per-process counter in klp_ftrace_handler() when a > function is redirected(). And klp_update_patch_state() might refuse > the migration when this counter is not zero. But it would require > to use a trampoline on return that would decrement the counter. > I am not sure if this is worth the complexity. > > One the other hand, this counter might actually remove the need > of the reliable backtrace. It is possible that I miss something > or that it is not easy/possible to implement the return trampoline. I agree this should work for unpatching, and even for patching a function which is already patched. Maybe I'm misunderstanding, but this would only work for unpatching or patching an already-patched function, wouldn't it? Because the original functions would not increment the counter so you would not know if tasks still had those on their call stacks. > Back to the original problem. I still consider calling > klp_update_patch_state(current) in vhost_worker() safe. Okay, I can send a v2 which does this, so long as it's okay to export klp_update_patch_state() to modules. Thanks, Seth