On Fri, May 03, 2024 at 12:41:52AM +0100, Al Viro wrote: > On Thu, May 02, 2024 at 04:21:13PM -0700, Kees Cook wrote: > > On Fri, May 03, 2024 at 12:12:28AM +0100, Al Viro wrote: > > > On Thu, May 02, 2024 at 03:52:21PM -0700, Kees Cook wrote: > > > > > > > As for semantics, what do you mean? Detecting dec-below-zero means we > > > > catch underflow, and detected inc-from-zero means we catch resurrection > > > > attempts. In both cases we avoid double-free, but we have already lost > > > > to a potential dangling reference to a freed struct file. But just > > > > letting f_count go bad seems dangerous. > > > > > > Detected inc-from-zero can also mean an RCU lookup detecting a descriptor > > > in the middle of getting closed. And it's more subtle than that, actually, > > > thanks to SLAB_TYPESAFE_BY_RCU for struct file. > > > > But isn't that already handled by __get_file_rcu()? i.e. shouldn't it be > > impossible for a simple get_file() to ever see a 0 f_count under normal > > conditions? > > For get_file() it is impossible. The comment about semantics had been > about the sane ways to recover if such crap gets detected. > > __get_file_rcu() is a separate story - consider the comment in there: > * atomic_long_inc_not_zero() above provided a full memory > * barrier when we acquired a reference. > * > * This is paired with the write barrier from assigning to the > * __rcu protected file pointer so that if that pointer still > * matches the current file, we know we have successfully > * acquired a reference to the right file. > > and IIRC, refcount_t is weaker wrt barriers. I think that was also fixed for refcount_t. I'll need to go dig out the commit... But anyway, there needs to be a general "oops I hit 0"-aware form of get_file(), and it seems like it should just be get_file() itself... -- Kees Cook
