Re: [PATCH v4] lib/vsprintf: defer filling siphash key on RT

["Jason A. Donenfeld" <Jason@zx2c4.com>]

Hi Sebastian,

On Mon, Aug 01, 2022 at 02:46:35PM +0200, Sebastian Andrzej Siewior wrote:
> On 2022-08-01 14:39:46 [+0200], Jason A. Donenfeld wrote:
> > On RT, we can't call get_random_bytes() from inside of the raw locks
> > that callers of vsprintf might take, because get_random_bytes() takes
> > normal spinlocks. So on those RT systems, defer the siphash key
> > generation to a worker.
> > 
> > Also, avoid using a static_branch, as this isn't the fast path.
> > Using static_branch_likely() to signal that ptr_key has been filled is a
> > bit much given that it is not a fast path.
> > 
> > Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
> > Reported-by: Mike Galbraith <efault@gmx.de>
> > Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
> > ---
> > Sebastian - feel free to take this and tweak it as needed. Sending this
> > mostly as something illustrative of what the "simpler" thing would be
> > that I had in mind. -Jason
> 
> Can have the same behaviour regardless of CONFIG_PREEMPT_RT? Here
> lockdep _may_ yell with !RT because it is broken for RT.
> If we agree that we drop the first %p print here, can we do this on
> both (regardless of CONFIG_PREEMPT_RT)?

"Lockdep may yell" -- but this would be when lockdep is turned on to
catch RT bugs, not to catch non-RT bugs. The actual bug only exists on
RT. This is an RT problem. Stop pretending that this is a real issue
outside of RT. It isn't. This is *only* an RT issue. So why would we
make things worse for an issue that doesn't actually exist on non-RT?

I too generally prefer having only one code path and not two. But the
way this patch is written, the worker function just gets reused with a
straight call on the non-RT case, so it doesn't actually require
duplicating code.

Jason