1 From: Eric Dumazet <eric.dumazet@gmail.com>
2 Date: Sat, 12 May 2012 21:23:23 +0000
3 Subject: [PATCH 6/7] codel: use u16 field instead of 31bits for rec_inv_sqrt
5 commit 6ff272c9ad65eda219cd975b9da2dbc31cc812ee upstream.
7 David pointed out gcc might generate poor code with 31bit fields.
9 Using u16 is more than enough and permits a better code output.
11 Also make the code intent more readable using constants, fixed point arithmetic
12 not being trivial for everybody.
14 Suggested-by: David Miller <davem@davemloft.net>
15 Signed-off-by: Eric Dumazet <edumazet@google.com>
16 Signed-off-by: David S. Miller <davem@davemloft.net>
18 include/net/codel.h | 25 +++++++++++++++----------
19 1 file changed, 15 insertions(+), 10 deletions(-)
21 diff --git a/include/net/codel.h b/include/net/codel.h
22 index bd8747c..7546517 100644
23 --- a/include/net/codel.h
24 +++ b/include/net/codel.h
25 @@ -133,13 +133,17 @@ struct codel_params {
30 - u32 rec_inv_sqrt:31;
33 codel_time_t first_above_time;
34 codel_time_t drop_next;
38 +#define REC_INV_SQRT_BITS (8 * sizeof(u16)) /* or sizeof_in_bits(rec_inv_sqrt) */
39 +/* needed shift to get a Q0.32 number from rec_inv_sqrt */
40 +#define REC_INV_SQRT_SHIFT (32 - REC_INV_SQRT_BITS)
43 * struct codel_stats - contains codel shared variables and stats
44 * @maxpacket: largest packet we've seen so far
45 @@ -173,17 +177,18 @@ static void codel_stats_init(struct codel_stats *stats)
46 * http://en.wikipedia.org/wiki/Methods_of_computing_square_roots#Iterative_methods_for_reciprocal_square_roots
47 * new_invsqrt = (invsqrt / 2) * (3 - count * invsqrt^2)
49 - * Here, invsqrt is a fixed point number (< 1.0), 31bit mantissa)
50 + * Here, invsqrt is a fixed point number (< 1.0), 32bit mantissa, aka Q0.32
52 static void codel_Newton_step(struct codel_vars *vars)
54 - u32 invsqrt = vars->rec_inv_sqrt;
55 - u32 invsqrt2 = ((u64)invsqrt * invsqrt) >> 31;
56 - u64 val = (3LL << 31) - ((u64)vars->count * invsqrt2);
57 + u32 invsqrt = ((u32)vars->rec_inv_sqrt) << REC_INV_SQRT_SHIFT;
58 + u32 invsqrt2 = ((u64)invsqrt * invsqrt) >> 32;
59 + u64 val = (3LL << 32) - ((u64)vars->count * invsqrt2);
61 - val = (val * invsqrt) >> 32;
62 + val >>= 2; /* avoid overflow in following multiply */
63 + val = (val * invsqrt) >> (32 - 2 + 1);
65 - vars->rec_inv_sqrt = val;
66 + vars->rec_inv_sqrt = val >> REC_INV_SQRT_SHIFT;
70 @@ -195,7 +200,7 @@ static codel_time_t codel_control_law(codel_time_t t,
71 codel_time_t interval,
74 - return t + reciprocal_divide(interval, rec_inv_sqrt << 1);
75 + return t + reciprocal_divide(interval, rec_inv_sqrt << REC_INV_SQRT_SHIFT);
79 @@ -326,7 +331,7 @@ static struct sk_buff *codel_dequeue(struct Qdisc *sch,
80 codel_Newton_step(vars);
83 - vars->rec_inv_sqrt = 0x7fffffff;
84 + vars->rec_inv_sqrt = ~0U >> REC_INV_SQRT_SHIFT;
86 vars->lastcount = vars->count;
87 vars->drop_next = codel_control_law(now, params->interval,