mtx_init/lock_init and uninitialized struct mtx

Matthew Fleming mdf356 at gmail.com
Thu Feb 24 16:56:50 UTC 2011


On Thu, Feb 24, 2011 at 7:47 AM, Dmitry Krivenok
<krivenok.dmitry at gmail.com> wrote:
> Hello Hackers,
>
> Is it allowed to call mtx_init on a mutex defined as an auto variable
> and not initialized explicitly, i.e.:

We recently ran into this problem at $WORK because we turned on the
deadc0de checking in uma zones for any zone without an explicit
init/fini function, in order to detect more use-after-free scenarios.
It happens that one of the bits in 0xDEADC0DE is the LO_INITIALIZED
bit.

It's a bit of a tough call, since one would like mtx_init(9) and
family to just work on any blob of memory, but one would also like to
catch a programmer error where a lock is re-initialized.

I suppose that for INVARIANTS the kernel can remember the address of
all initialized locks and forget it on lock destroy, and in this way
have a useful assert and also allow lock_init on random storage.  This
would also allow for detecting if the memory for a lock was released
but the lock wasn't destroyed.

Sadly, I have just enough time to propose this and not enough to write
a patch at the moment.

Thanks,
matthew

>
> static int foo()
> {
>   struct mtx m;  // Uninitialized auto variable, so it's value is undefined.
>   mtx_init(&m, "my_mutex", NULL, MTX_DEF);
>   …
>   // Do something
>   ...
>   mtx_destroy(&m);
>   return 0;
> }
>
> I encountered a problem with such code on a kernel compiled with
> INVARIANTS option.
> The problem is that mtx_init calls lock_init(&m->lock_object) and
> lock_init, in turn, calls:
>
>  79         /* Check for double-init and zero object. */
>  80         KASSERT(!lock_initalized(lock), ("lock \"%s\" %p already
> initialized",
>  81             name, lock));
>
> lock_initialized() just checks that a bit is set in lo_flags field of
> struct lock_object:
>
> 178 #define lock_initalized(lo)     ((lo)->lo_flags & LO_INITIALIZED)
>
> However, the structure containing this field is never initialized
> (neither in mtx_init nor in lock_init).
> So, assuming that the mutex was defined as auto variable, the content
> of lock_object field of struct mtx
> is also undefined:
>
>  37 struct mtx {
>  38         struct lock_object      lock_object;    /* Common lock
> properties. */
>  39         volatile uintptr_t      mtx_lock;       /* Owner and flags. */
>  40 };
>
> In some cases, the initial value of lo_flags _may_ have the
> "initialized" bit set and KASSERT will call panic.
>
> Is it user's responsibility to properly (how exactly?) initialize
> struct mtx, e.g.
> memset(&m, '\0', sizeof(struct mtx));
>
> Or should mtx_init() explicitly initialize all fields of struct mtx?
>
> Thanks in advance!
>
> --
> Sincerely yours, Dmitry V. Krivenok
> e-mail: krivenok.dmitry at gmail.com
> skype: krivenok_dmitry
> jabber: krivenok_dmitry at jabber.ru
> icq: 242-526-443


More information about the freebsd-hackers mailing list