On Fri, Oct 07, 2022 at 02:17:22PM -0700, Darrick J. Wong wrote:
> On Fri, Oct 07, 2022 at 12:01:03PM -0600, Jason A. Donenfeld wrote:
> > Rather than incurring a division or requesting too many random bytes for
> > the given range, use the prandom_u32_max() function, which only takes
> > the minimum required bytes from the RNG and avoids divisions.
> >
> > Reviewed-by: Kees Cook <keescook@xxxxxxxxxxxx>
> > Reviewed-by: KP Singh <kpsingh@xxxxxxxxxx>
> > Reviewed-by: Christoph Böhmwalder <christoph.boehmwalder@xxxxxxxxxx> # for drbd
> > Reviewed-by: Jan Kara <jack@xxxxxxx> # for ext2, ext4, and sbitmap
> > Signed-off-by: Jason A. Donenfeld <Jason@xxxxxxxxx>
> > ---
>
> <snip, skip to the xfs part>
>
> > diff --git a/fs/xfs/libxfs/xfs_alloc.c b/fs/xfs/libxfs/xfs_alloc.c
> > index e2bdf089c0a3..6261599bb389 100644
> > --- a/fs/xfs/libxfs/xfs_alloc.c
> > +++ b/fs/xfs/libxfs/xfs_alloc.c
> > @@ -1520,7 +1520,7 @@ xfs_alloc_ag_vextent_lastblock(
> >
> > #ifdef DEBUG
> > /* Randomly don't execute the first algorithm. */
> > - if (prandom_u32() & 1)
> > + if (prandom_u32_max(2))
>
> I wonder if these usecases (picking 0 or 1 randomly) ought to have a
> trivial wrapper to make it more obvious that we want boolean semantics:
>
> static inline bool prandom_bool(void)
> {
> return prandom_u32_max(2);
> }
>
> if (prandom_bool())
> use_crazy_algorithm(...);
>
Yea, I've had a lot of similar ideas there. Part of doing this (initial)
patchset is to get an intuitive sense of what's actually used and how
often. On my list for investigation are a get_random_u32_max() to return
uniform numbers by rejection sampling (prandom_u32_max() doesn't do
that uniformly) and adding a function for booleans or bits < 8. Possible
ideas for the latter include:
bool get_random_bool(void);
bool get_random_bool(unsigned int probability);
bool get_random_bits(u8 bits_less_than_eight);
With the core of all of those involving the same batching as the current
get_random_u{8,16,32,64}() functions, but also buffering the latest byte
and managing how many bits are left in it that haven't been shifted out
yet.
So API-wise, there are a few ways to go, so hopefully this series will
start to give a good picture of what's needed.
One thing I've noticed is that most of the prandom_u32_max(2)
invocations are in debug or test code, so that doesn't need to be
optimized. But kfence does that too in its hot path, so a
get_random_bool() function there would in theory lead to an 8x speed-up.
But I guess I just have to try some things and see.
Anyway, that is a long way to say, I share you curiosity on the matter
and I'm looking into it.
Jason
