target/linux/generic/patches-3.3/045-codel-use-u16-field-instead-of-31bits-for-rec_inv_sq.patch

   1 From 6ff272c9ad65eda219cd975b9da2dbc31cc812ee Mon Sep 17 00:00:00 2001
   2 From: Eric Dumazet <eric.dumazet@gmail.com>
   3 Date: Sat, 12 May 2012 21:23:23 +0000
   4 Subject: [PATCH 46/50] codel: use u16 field instead of 31bits for
   5  rec_inv_sqrt
   6
   7 David pointed out gcc might generate poor code with 31bit fields.
   8
   9 Using u16 is more than enough and permits a better code output.
  10
  11 Also make the code intent more readable using constants, fixed point arithmetic
  12 not being trivial for everybody.
  13
  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>
  17 ---
  18  include/net/codel.h |   25 +++++++++++++++----------
  19  1 file changed, 15 insertions(+), 10 deletions(-)
  20
  21 --- a/include/net/codel.h
  22 +++ b/include/net/codel.h
  23 @@ -133,13 +133,17 @@ struct codel_params {
  24  struct codel_vars {
  25         u32             count;
  26         u32             lastcount;
  27 -       bool            dropping:1;
  28 -       u32             rec_inv_sqrt:31;
  29 +       bool            dropping;
  30 +       u16             rec_inv_sqrt;
  31         codel_time_t    first_above_time;
  32         codel_time_t    drop_next;
  33         codel_time_t    ldelay;
  34  };
  35
  36 +#define REC_INV_SQRT_BITS (8 * sizeof(u16)) /* or sizeof_in_bits(rec_inv_sqrt) */
  37 +/* needed shift to get a Q0.32 number from rec_inv_sqrt */
  38 +#define REC_INV_SQRT_SHIFT (32 - REC_INV_SQRT_BITS)
  39 +
  40  /**
  41   * struct codel_stats - contains codel shared variables and stats
  42   * @maxpacket: largest packet we've seen so far
  43 @@ -173,17 +177,18 @@ static void codel_stats_init(struct code
  44   * http://en.wikipedia.org/wiki/Methods_of_computing_square_roots#Iterative_methods_for_reciprocal_square_roots
  45   * new_invsqrt = (invsqrt / 2) * (3 - count * invsqrt^2)
  46   *
  47 - * Here, invsqrt is a fixed point number (< 1.0), 31bit mantissa)
  48 + * Here, invsqrt is a fixed point number (< 1.0), 32bit mantissa, aka Q0.32
  49   */
  50  static void codel_Newton_step(struct codel_vars *vars)
  51  {
  52 -       u32 invsqrt = vars->rec_inv_sqrt;
  53 -       u32 invsqrt2 = ((u64)invsqrt * invsqrt) >> 31;
  54 -       u64 val = (3LL << 31) - ((u64)vars->count * invsqrt2);
  55 +       u32 invsqrt = ((u32)vars->rec_inv_sqrt) << REC_INV_SQRT_SHIFT;
  56 +       u32 invsqrt2 = ((u64)invsqrt * invsqrt) >> 32;
  57 +       u64 val = (3LL << 32) - ((u64)vars->count * invsqrt2);
  58
  59 -       val = (val * invsqrt) >> 32;
  60 +       val >>= 2; /* avoid overflow in following multiply */
  61 +       val = (val * invsqrt) >> (32 - 2 + 1);
  62
  63 -       vars->rec_inv_sqrt = val;
  64 +       vars->rec_inv_sqrt = val >> REC_INV_SQRT_SHIFT;
  65  }
  66
  67  /*
  68 @@ -195,7 +200,7 @@ static codel_time_t codel_control_law(co
  69                                       codel_time_t interval,
  70                                       u32 rec_inv_sqrt)
  71  {
  72 -       return t + reciprocal_divide(interval, rec_inv_sqrt << 1);
  73 +       return t + reciprocal_divide(interval, rec_inv_sqrt << REC_INV_SQRT_SHIFT);
  74  }
  75
  76
  77 @@ -326,7 +331,7 @@ static struct sk_buff *codel_dequeue(str
  78                         codel_Newton_step(vars);
  79                 } else {
  80                         vars->count = 1;
  81 -                       vars->rec_inv_sqrt = 0x7fffffff;
  82 +                       vars->rec_inv_sqrt = ~0U >> REC_INV_SQRT_SHIFT;
  83                 }
  84                 vars->lastcount = vars->count;
  85                 vars->drop_next = codel_control_law(now, params->interval,