On Tue, 25 Sep 2018, Vladis Dronov wrote: > > > What about adding a WARN_ON()? It doesn't crash the kernel and it will > > > be detected and reported by syzbot. > > Yes, that would be a great solution. > > > Sure, we could do that. But would be the point? > > We know when usb_find_alt_setting() callers do smth weird and go fix them. > > > After c9a4cb204e9e, calling usb_find_alt_setting() with a NULL config is > > no more of a bug than calling kfree() with a NULL pointer. > > Yes, exactly. > > > You wouldn't want to put a WARN_ON in kfree(), would you? > > Honestly, in the ideal world I would, again, to be aware when some code does > something weird so we know about it. But this world is this world, it needs > more performance to the throne of performance. But is it really worthwhile? In terms of catching bugs, this would help in only one situation: when the programmer thinks the argument should always be non-NULL because a NULL argument indicates a bug. Such situations seem to be relatively rare, and we can handle them by inserting a WARN_ON() at the call site if need be. So it's a choice between: 1. Putting a single test for NULL in the function being called, together with WARN_ON() at a small number of call sites, or 2. Putting a WARN_ON() (or allowing a crash) in the function being called, together with tests for NULL at a potentially large number of call sites. 1 has two advantages over 2. First, it involves adding less code overall. Second, it doesn't require the programmer to remember to add special code (a test or a WARN_ON) in situation where it doesn't matter -- presumably the majority of them. Now consider the case at hand: the call to usb_find_alt_setting() from check_ctrlrecip(). In this case ps->dev->actconfig being NULL doesn't indicate an error or a bug; it merely indicates that the user is trying to send a control request to a device which happens to be unconfigured, which is a perfectly valid thing to do. Therefore it shouldn't require any special handling at the call site. Alan Stern