Re: [PATCH 2/3] find: micro-optimize for_each_{set,clear}_bit()

[Yury Norov <yury.norov@gmail.com>]

On Sat, Jun 19, 2021 at 10:28:07AM -0700, Yury Norov wrote:
> On Sat, Jun 19, 2021 at 05:24:15PM +0100, Marc Zyngier wrote:
> > On Fri, 18 Jun 2021 20:57:34 +0100,
> > Yury Norov <yury.norov@gmail.com> wrote:
> > > 
> > > The macros iterate thru all set/clear bits in a bitmap. They search a
> > > first bit using find_first_bit(), and the rest bits using find_next_bit().
> > > 
> > > Since find_next_bit() is called shortly after find_first_bit(), we can
> > > save few lines of I-cache by not using find_first_bit().
> > 
> > Really?
> > 
> > > 
> > > Signed-off-by: Yury Norov <yury.norov@gmail.com>
> > > ---
> > >  include/linux/find.h | 4 ++--
> > >  1 file changed, 2 insertions(+), 2 deletions(-)
> > > 
> > > diff --git a/include/linux/find.h b/include/linux/find.h
> > > index 4500e8ab93e2..ae9ed52b52b8 100644
> > > --- a/include/linux/find.h
> > > +++ b/include/linux/find.h
> > > @@ -280,7 +280,7 @@ unsigned long find_next_bit_le(const void *addr, unsigned
> > >  #endif
> > >  
> > >  #define for_each_set_bit(bit, addr, size) \
> > > -	for ((bit) = find_first_bit((addr), (size));		\
> > > +	for ((bit) = find_next_bit((addr), (size), 0);		\
> > 
> > On which architecture do you observe a gain? Only 32bit ARM and m68k
> > implement their own version of find_first_bit(), and everyone else
> > uses the canonical implementation:
> 
> And those who enable GENERIC_FIND_FIRST_BIT - x86, arm64, arc, mips
> and s390.
> 
> > #ifndef find_first_bit
> > #define find_first_bit(addr, size) find_next_bit((addr), (size), 0)
> > #endif
> > 
> > These architectures explicitly have different implementations for
> > find_first_bit() and find_next_bit() because they can do better
> > (whether that is true or not is another debate). I don't think you
> > should remove this optimisation until it has been measured on these
> > two architectures.
> 
> This patch is based on a series that enables separate implementation
> of find_first_bit() for all architectures; according to my tests,
> find_first* is ~ twice faster than find_next* on arm64 and x86.
> 
> https://lore.kernel.org/lkml/20210612123639.329047-1-yury.norov@gmail.com/T/#t
> 
> After applying the series, I noticed that my small kernel module that
> calls for_each_set_bit() is now using find_first_bit() to just find
> one bit, and find_next_bit() for all others. I think it's better to
> always use find_next_bit() in this case to minimize the chance of
> cache miss. But if it's not that obvious, I'll try to write some test.

This test measures the difference between for_each_set_bit() and
for_each_set_bit_from().

diff --git a/lib/find_bit_benchmark.c b/lib/find_bit_benchmark.c
index 5637c5711db9..1f37e99090b0 100644
--- a/lib/find_bit_benchmark.c
+++ b/lib/find_bit_benchmark.c
@@ -111,6 +111,59 @@ static int __init test_find_next_and_bit(const void *bitmap,
 	return 0;
 }
 
+#ifdef CONFIG_X86_64
+#define flush_cache_all() wbinvd()
+#endif
+
+static int __init test_for_each_set_bit(int flags)
+{
+#ifdef flush_cache_all
+	DECLARE_BITMAP(bm, BITS_PER_LONG * 2);
+	unsigned long i, cnt = 0;
+	ktime_t time;
+
+	bm[0] = 1; bm[1] = 0;
+
+	time = ktime_get();
+	while (cnt < 1000) {
+		if (flags)
+			flush_cache_all();
+		for_each_set_bit(i, bm, BITS_PER_LONG * 2)
+			cnt++;
+	}
+
+	time = ktime_get() - time;
+
+	pr_err("for_each_set_bit:   %18llu ns, %6ld iterations\n",  time, cnt);
+#endif
+	return 0;
+}
+
+static int __init test_for_each_set_bit_from(int flags)
+{
+#ifdef flush_cache_all
+	DECLARE_BITMAP(bm, BITS_PER_LONG * 2);
+	unsigned long i, cnt = 0;
+	ktime_t time;
+
+	bm[0] = 1; bm[1] = 0;
+
+	time = ktime_get();
+	while (cnt < 1000) {
+		if (flags)
+			flush_cache_all();
+		i = 0;
+		for_each_set_bit_from(i, bm, BITS_PER_LONG * 2)
+			cnt++;
+	}
+
+	time = ktime_get() - time;
+
+	pr_err("for_each_set_bit_from:%16llu ns, %6ld iterations\n", time, cnt);
+#endif
+	return 0;
+}
+
 static int __init find_bit_test(void)
 {
 	unsigned long nbits = BITMAP_LEN / SPARSE;
@@ -147,6 +200,16 @@ static int __init find_bit_test(void)
 	test_find_first_bit(bitmap, BITMAP_LEN);
 	test_find_next_and_bit(bitmap, bitmap2, BITMAP_LEN);
 
+	pr_err("\nStart testing for_each_bit()\n");
+
+	test_for_each_set_bit(0);
+	test_for_each_set_bit_from(0);
+
+	pr_err("\nStart testing for_each_bit() with cash flushing\n");
+
+	test_for_each_set_bit(1);
+	test_for_each_set_bit_from(1);
+
 	/*
 	 * Everything is OK. Return error just to let user run benchmark
 	 * again without annoying rmmod.

Here on each iteration: 
 - for_each_set_bit() calls find_first_bit() once, and find_next_bit() once.
 - for_each_set_bit_from() calls  find_next_bit() twice.

On my AMD Ryzen 7 4700U, the result is like this:

Start testing for_each_bit()
for_each_set_bit:                15296 ns,   1000 iterations
for_each_set_bit_from:           15225 ns,   1000 iterations

Start testing for_each_bit() with cash flushing
for_each_set_bit:               547626 ns,   1000 iterations
for_each_set_bit_from:          497899 ns,   1000 iterations

for_each_set_bit_from() is ~10% faster than for_each_set_bit() in
case of cold caches, and no significant difference was observed if
flush_cache_all() is not called.

So, it looks reasonable to switch for_each_set_bit() to use
find_next_bit() only.

Thanks,
Yury