On Fri, Sep 16, 2016 at 10:31:41PM +0100, Al Viro wrote: > On Fri, Sep 16, 2016 at 01:59:38PM -0700, Guenter Roeck wrote: > > Yes, reverting 6e050503a150 fixes the problem. > > > > I added a BUG() into the "if (unlikely())" below, but it doesn't catch, > > and I still get the ip: OVERRUN errors. Which leaves me a bit puzzled. > > > > Guenter > > > > > The change in question is > > > if (__copy_size && __access_ok(__copy_from, __copy_size)) > > > - return __copy_user(to, from, __copy_size); > > > + __copy_size = __copy_user(to, from, __copy_size); > > > + > > > + if (unlikely(__copy_size)) > > > + memset(to + (n - __copy_size), 0, __copy_size); > > > > > > return __copy_size; > > So we don't even hit that memset()? What the hell? __copy_user() is > declared as > __kernel_size_t __copy_user(void *to, const void *from, __kernel_size_t n); > > and __copy_size copy_from_user() is > > __kernel_size_t __copy_size = (__kernel_size_t) n; > > So > return __copy_user(to, from, __copy_size); > and > __copy_size = __copy_user(to, from, __copy_size); > return __copy_size; > ought to be doing exactly the same thing. At that point it's starting to > smell like a compiler bug somewhere in there. > > Try to remove that (not triggered) if (unlikely(__copy_size)) memset(...) > and see if that's enough to recover. And it would be nice to see what > all three variants (as it is, with commit reverted and with just that if > removed) generate in e.g. sys_utimensat() (fs/utimes.s)
Adding pr_info() just before the "if (unlikely..." fixes the problem. Commenting out the "if (unlikely())" code fixes the problem. Using a new variable "unsigned long x" for the return code instead of re-using __copy_size fixes the problem. Replacing "return __copy_size;" with "return __copy_size & 0xffffffff;" fixes the problem. The problem only seems to be seen if the copy size length is odd (more specifically, the failing copy always has a length of 25 bytes). No idea what is going on. Bug in __copy_user() ? Compiler bug ? Guenter