On 3/21/23 12:48?PM, Linus Torvalds wrote: > On Tue, Mar 21, 2023 at 9:43?AM David Howells <dhowells@xxxxxxxxxx> wrote: >> >> (1) Fix request_key() so that it doesn't cache a looked up key on the >> current thread if that thread is a kernel thread. The cache is >> cleared during notify_resume - but that doesn't happen in kernel >> threads. This is causing cifs DNS keys to be un-invalidateable. > > I've pulled this, but I'd like people to look a bit more at this. > > The issue with TIF_NOTIFY_RESUME is that it is only done on return to > user space. > > And these days, PF_KTHREAD isn't the only case that never returns to > user space. PF_IO_WORKER has the exact same behaviour. > > Now, to counteract this, as of this merge window (and marked for > stable) IO threads do a fake "return to user mode" handling in > io_run_task_work(), and so I think we're all good, but I'd like people > to at least think about this. I haven't seen the patch yet as it hasn't been pushed, but can imagine what it looks like. It may make sense to add some debug check for PF_KTHREAD having TIF_NOTIFY_RESUME set, or task_work pending for that matter, as that is generally not workable without doing something to handle it explicitly. For PF_IO_WORKER, with the commit you mentioned, those threads should deal with TIF_NOTIFY_RESUME just fine. Until something else gets added that is also run from exit_to_user_mode... -- Jens Axboe
