Re: [PATCH v7] printk: hash addresses printed with %p

[Tobin C. Harding]

On Thu, Oct 26, 2017 at 12:59:08AM +0200, Jason A. Donenfeld wrote:
> On Thu, Oct 26, 2017 at 12:27 AM, Tobin C. Harding <me@xxxxxxxx> wrote:
> > How good is unlikely()?
> 
> It places that branch way at the bottom of the function so that it's
> less likely to pollute the icache.
> 
> > It doesn't _feel_ right adding a check on every call to printk just to
> > check for a condition that was only true for the briefest time when the
> > kernel booted. But if unlikely() is good then I guess it doesn't hurt.
> >
> > I'm leaning towards the option 1, but then all those text books I read
> > are telling me to implement the simplest solution first then if we need
> > to go faster implement the more complex solution.
> >
> > This is a pretty airy fairy discussion now, but if you have an opinion
> > I'd love to hear it.
> 
> I don't think adding a single branch there really matters that much,
> considering how many random other branches there are all over the
> printk code. However, if you really want to optimize on the little
> bits, and sensibly don't want to go with the overcomplex
> workqueue-to-statickey thing, you could consider using a plain vanilla
> `bool has_gotten_random_ptr_secret` instead of using the atomic_t. The
> reason is that there's only ever one single writer, changing from a 0
> to a 1. Basically the only thing doing the atomic_t got you was a
> cache flush surrounding the read (and the write) so that assigning
> has_gotten_random_ptr_secret=true would take effect _immediately_.
> However, since you might not necessarily about that, going with a bool
> instead will save you an expensive cache flush, while potentially
> being a microsecond out of date the first time it's used. Seems like
> an okay trade off to me. (That kind of cache latency, also, is a few
> orders of magnitude better than using a work queue for the statickey
> stuff.)

Awesome. Patch to follow.

thanks,
Tobin.