Commit 04157469b7b8 ("net: Use static_key for XPS maps") introduced a
static key for XPS, but the increments/decrements don't match.
First, the static key's counter is incremented once for each queue, but
only decremented once for a whole batch of queues, leading to large
unbalances.
Second, the xps_rxqs_needed key is decremented whenever we reset a batch
of queues, whether they had any rxqs mapping or not, so that if we setup
cpu-XPS on em1 and RXQS-XPS on em2, resetting the queues on em1 would
decrement the xps_rxqs_needed key.
This reworks the accounting scheme so that the xps_needed key is
incremented only once for each type of XPS for all the queues on a
device, and the xps_rxqs_needed key is incremented only once for all
queues. This is sufficient to let us retrieve queues via
get_xps_queue().
This patch introduces a new reset_xps_maps(), which reinitializes and
frees the appropriate map (xps_rxqs_map or xps_cpus_map), and drops a
reference to the needed keys:
- both xps_needed and xps_rxqs_needed, in case of rxqs maps,
- only xps_needed, in case of CPU maps.
Now, we also need to call reset_xps_maps() at the end of
__netif_set_xps_queue() when there's no active map left, for example
when writing '00000000,00000000' to all queues' xps_rxqs setting.
Fixes: 04157469b7b8 ("net: Use static_key for XPS maps")
Signed-off-by: Sabrina Dubroca <s...@queasysnail.net>
---
net/core/dev.c | 45 ++++++++++++++++++++++++---------------------
1 file changed, 24 insertions(+), 21 deletions(-)
diff --git a/net/core/dev.c b/net/core/dev.c
index 32a63f4c3a92..3470e7fff1f4 100644
--- a/net/core/dev.c
+++ b/net/core/dev.c
@@ -2175,6 +2175,20 @@ static bool remove_xps_queue_cpu(struct net_device *dev,
return active;
}
+static void reset_xps_maps(struct net_device *dev,
+ struct xps_dev_maps *dev_maps,
+ bool is_rxqs_map)
+{
+ if (is_rxqs_map) {
+ static_key_slow_dec_cpuslocked(&xps_rxqs_needed);
+ RCU_INIT_POINTER(dev->xps_rxqs_map, NULL);
+ } else {
+ RCU_INIT_POINTER(dev->xps_cpus_map, NULL);
+ }
+ static_key_slow_dec_cpuslocked(&xps_needed);
+ kfree_rcu(dev_maps, rcu);
+}
+
static void clean_xps_maps(struct net_device *dev, const unsigned long *mask,
struct xps_dev_maps *dev_maps, unsigned int nr_ids,
u16 offset, u16 count, bool is_rxqs_map)
@@ -2186,13 +2200,8 @@ static void clean_xps_maps(struct net_device *dev, const
unsigned long *mask,
j < nr_ids;)
active |= remove_xps_queue_cpu(dev, dev_maps, j, offset,
count);
- if (!active) {
- if (is_rxqs_map)
- RCU_INIT_POINTER(dev->xps_rxqs_map, NULL);
- else
- RCU_INIT_POINTER(dev->xps_cpus_map, NULL);
- kfree_rcu(dev_maps, rcu);
- }
+ if (!active)
+ reset_xps_maps(dev, dev_maps, is_rxqs_map);
if (!is_rxqs_map) {
for (i = offset + (count - 1); count--; i--) {
@@ -2236,10 +2245,6 @@ static void netif_reset_xps_queues(struct net_device
*dev, u16 offset,
false);
out_no_maps:
- if (static_key_enabled(&xps_rxqs_needed))
- static_key_slow_dec_cpuslocked(&xps_rxqs_needed);
-
- static_key_slow_dec_cpuslocked(&xps_needed);
mutex_unlock(&xps_map_mutex);
cpus_read_unlock();
}
@@ -2357,9 +2362,12 @@ int __netif_set_xps_queue(struct net_device *dev, const
unsigned long *mask,
if (!new_dev_maps)
goto out_no_new_maps;
- static_key_slow_inc_cpuslocked(&xps_needed);
- if (is_rxqs_map)
- static_key_slow_inc_cpuslocked(&xps_rxqs_needed);
+ if (!dev_maps) {
+ /* Increment static keys at most once per type */
+ static_key_slow_inc_cpuslocked(&xps_needed);
+ if (is_rxqs_map)
+ static_key_slow_inc_cpuslocked(&xps_rxqs_needed);
+ }
for (j = -1; j = netif_attrmask_next(j, possible_mask, nr_ids),
j < nr_ids;) {
@@ -2457,13 +2465,8 @@ int __netif_set_xps_queue(struct net_device *dev, const
unsigned long *mask,
}
/* free map if not active */
- if (!active) {
- if (is_rxqs_map)
- RCU_INIT_POINTER(dev->xps_rxqs_map, NULL);
- else
- RCU_INIT_POINTER(dev->xps_cpus_map, NULL);
- kfree_rcu(dev_maps, rcu);
- }
+ if (!active)
+ reset_xps_maps(dev, dev_maps, is_rxqs_map);
out_no_maps:
mutex_unlock(&xps_map_mutex);
--
2.19.2
