The 64-bit gw_factor is divided by BATADV_TQ_LOCAL_WINDOW_SIZE ** 2 * 64. But the rest of the calculation has nothing to do with the tq window size end therefore the calculation is just (tmp_gw_factor / (64 ** 3)).
The problem with 64 bit div is that it doesn't work on systems without native 64 bit div support. It has to be emulated using do_div or div_u64. The change in f63c54bba31d ("batman-adv: Avoid u32 overflow during gateway select") only compiled on such systems because the compiler converted the div to a (tmp_gw_factor >> 18). Making this explicit avoids having build problems in the future when BATADV_TQ_LOCAL_WINDOW_SIZE is changed in such a way that (BATADV_TQ_LOCAL_WINDOW_SIZE ** 2 * 64) is not a power of two.
Signed-off-by: Sven Eckelmann sven@narfation.org --- gateway_client.c | 6 +----- 1 file changed, 1 insertion(+), 5 deletions(-)
diff --git a/gateway_client.c b/gateway_client.c index 3f32357..0f8d06b 100644 --- a/gateway_client.c +++ b/gateway_client.c @@ -134,14 +134,10 @@ batadv_gw_get_best_gw_node(struct batadv_priv *bat_priv) struct batadv_neigh_ifinfo *router_ifinfo; struct batadv_gw_node *gw_node, *curr_gw = NULL; uint64_t max_gw_factor = 0, tmp_gw_factor = 0; - uint32_t gw_divisor; uint8_t max_tq = 0; uint8_t tq_avg; struct batadv_orig_node *orig_node;
- gw_divisor = BATADV_TQ_LOCAL_WINDOW_SIZE * BATADV_TQ_LOCAL_WINDOW_SIZE; - gw_divisor *= 64; - rcu_read_lock(); hlist_for_each_entry_rcu(gw_node, &bat_priv->gw.list, list) { if (gw_node->deleted) @@ -167,7 +163,7 @@ batadv_gw_get_best_gw_node(struct batadv_priv *bat_priv) tmp_gw_factor = tq_avg * tq_avg; tmp_gw_factor *= gw_node->bandwidth_down; tmp_gw_factor *= 100 * 100; - tmp_gw_factor /= gw_divisor; + tmp_gw_factor >>= 18;
if ((tmp_gw_factor > max_gw_factor) || ((tmp_gw_factor == max_gw_factor) &&
On 22/06/15 09:13, Sven Eckelmann wrote:
- gw_divisor = BATADV_TQ_LOCAL_WINDOW_SIZE * BATADV_TQ_LOCAL_WINDOW_SIZE;
- gw_divisor *= 64;
Marek can you comment on why this constant was used in this computation? To me it looks like we only needed a "good enough number" and not a real reference to the local window size.
The remaining change looks good to me.
If you merge this patch I'll then squash it with the uint64_t change that we already have and I will re-send the pull request to David.
Thanks,
- rcu_read_lock(); hlist_for_each_entry_rcu(gw_node, &bat_priv->gw.list, list) { if (gw_node->deleted)
@@ -167,7 +163,7 @@ batadv_gw_get_best_gw_node(struct batadv_priv *bat_priv) tmp_gw_factor = tq_avg * tq_avg; tmp_gw_factor *= gw_node->bandwidth_down; tmp_gw_factor *= 100 * 100;
tmp_gw_factor /= gw_divisor;
tmp_gw_factor >>= 18; if ((tmp_gw_factor > max_gw_factor) || ((tmp_gw_factor == max_gw_factor) &&
On Monday, June 22, 2015 09:13:23 Sven Eckelmann wrote:
The 64-bit gw_factor is divided by BATADV_TQ_LOCAL_WINDOW_SIZE ** 2 * 64. But the rest of the calculation has nothing to do with the tq window size end therefore the calculation is just (tmp_gw_factor / (64 ** 3)).
The problem with 64 bit div is that it doesn't work on systems without native 64 bit div support. It has to be emulated using do_div or div_u64. The change in f63c54bba31d ("batman-adv: Avoid u32 overflow during gateway select") only compiled on such systems because the compiler converted the div to a (tmp_gw_factor >> 18). Making this explicit avoids having build problems in the future when BATADV_TQ_LOCAL_WINDOW_SIZE is changed in such a way that (BATADV_TQ_LOCAL_WINDOW_SIZE ** 2 * 64) is not a power of two.
Signed-off-by: Sven Eckelmann sven@narfation.org
gateway_client.c | 6 +----- 1 file changed, 1 insertion(+), 5 deletions(-)
Applied in revision 013aab3.
Thanks, Marek
b.a.t.m.a.n@lists.open-mesh.org