Re: [PATCH 2/2 v3] lib/vsprintf: Initialize vsprintf's pointer hash once the random core is ready.

[Jason A. Donenfeld]

Hi,

On Mon, Aug 1, 2022 at 11:34 AM Sebastian Andrzej Siewior
<bigeasy@xxxxxxxxxxxxx> wrote:
>
> The printk code invokes vnsprintf in order to compute the complete
> string before adding it into its buffer. This happens in an IRQ-off
> region which leads to a warning on PREEMPT_RT in the random code if the
> format strings contains a %p for pointer printing. This happens because
> the random core acquires locks which become sleeping locks on PREEMPT_RT
> which must not be acquired with disabled interrupts and or preemption
> disabled.
> By default the pointers are hashed which requires a random value on the
> first invocation (either by printk or another user which comes first.
>
> One could argue that there is no need for printk to disable interrupts
> during the vsprintf() invocation which would fix the just mentioned
> problem. However printk itself can be invoked in a context with
> disabled interrupts which would lead to the very same problem.
>
> Move the initialization of ptr_key into a worker and schedule it from
> subsys_initcall(). This happens early but after the workqueue subsystem
> is ready. Use get_random_bytes() to retrieve the random value if the RNG
> core is ready, otherwise schedule a worker in two seconds and try again.
>
> Reported-by: Mike Galbraith <efault@xxxxxx>
> Signed-off-by: Sebastian Andrzej Siewior <bigeasy@xxxxxxxxxxxxx>
> ---
> v2…v3:
>    - schedule a worker every two seconds if the RNG core is not ready.
>
>  lib/vsprintf.c |   46 +++++++++++++++++++++++++++-------------------
>  1 file changed, 27 insertions(+), 19 deletions(-)
>
> --- a/lib/vsprintf.c
> +++ b/lib/vsprintf.c
> @@ -751,31 +751,39 @@ static int __init debug_boot_weak_hash_e
>  early_param("debug_boot_weak_hash", debug_boot_weak_hash_enable);
>
>  static bool filled_random_ptr_key;
> +static siphash_key_t ptr_key __read_mostly;
> +static void fill_ptr_key_workfn(struct work_struct *work);
> +static DECLARE_DELAYED_WORK(fill_ptr_key_work, fill_ptr_key_workfn);
> +
> +static void fill_ptr_key_workfn(struct work_struct *work)
> +{
> +       if (!rng_is_initialized()) {
> +               queue_delayed_work(system_unbound_wq, &fill_ptr_key_work, HZ  * 2);
> +               return;
> +       }
> +
> +       get_random_bytes(&ptr_key, sizeof(ptr_key));
> +
> +       /* Pairs with smp_rmb() before reading ptr_key. */
> +       smp_wmb();
> +       WRITE_ONCE(filled_random_ptr_key, true);
> +}
> +
> +static int __init vsprintf_init_hashval(void)
> +{
> +       fill_ptr_key_workfn(NULL);
> +       return 0;
> +}
> +subsys_initcall(vsprintf_init_hashval)
>
>  /* Maps a pointer to a 32 bit unique identifier. */
>  static inline int __ptr_to_hashval(const void *ptr, unsigned long *hashval_out)
>  {
> -       static siphash_key_t ptr_key __read_mostly;
>         unsigned long hashval;
>
> -       if (!READ_ONCE(filled_random_ptr_key)) {
> -               static bool filled = false;
> -               static DEFINE_SPINLOCK(filling);
> -               unsigned long flags;
> -
> -               if (!rng_is_initialized() ||
> -                   !spin_trylock_irqsave(&filling, flags))
> -                       return -EAGAIN;
> -
> -               if (!filled) {
> -                       get_random_bytes(&ptr_key, sizeof(ptr_key));
> -                       /* Pairs with smp_rmb() before reading ptr_key. */
> -                       smp_wmb();
> -                       WRITE_ONCE(filled_random_ptr_key, true);
> -                       filled = true;
> -               }
> -               spin_unlock_irqrestore(&filling, flags);
> -       }
> +       if (!READ_ONCE(filled_random_ptr_key))
> +               return -EBUSY;
> +
>         /* Pairs with smp_wmb() after writing ptr_key. */
>         smp_rmb();
>

As we discussed at Plumbers, it seems like this is the least-awful way
forward. If we wind up with another case sufficiently similar to this,
I'll add back a notifier to random.c. But while there's only this one
special case, the ugly timer thing will have to do.

So Petr - feel free to queue this up this v3, with my objection now removed.

Jason