On 7/10/19 11:02 AM, Eric Biggers wrote: > I already mentioned that io_uring triggers it too. > > Those are just 2 cases that syzbot happened to generate reproducers for. I > expect there are many others too, since many places in the kernel allocate > workqueues. AFAICS most are placed in static or global variables which avoids > this issue, but there are still many cases where a workqueue is owned by some > dynamic structure that can have a much shorter lifetime. > > You can also check the other syzbot reports that look similar > (https://lore.kernel.org/lkml/20190710055838.GC2152@sol.localdomain/). > Two of them have C reproducers too.
As you may know lockdep cannot use dynamic memory allocation because doing so would introduce a circular dependency between lockdep and the memory allocator. Hence the fixed size arrays in the lockdep code. Additionally, as far as I know lockdep works fine for human kernel developers and only syzbot runs code that triggers the lockdep limits. So I think it's up to the syzbot authors to come up with a solution. I mean another solution than finger pointing at kernel developers. Bart.
