On Fri, Aug 27, 2021 at 12:41 AM Jesse Brandeburg
[off-list ref] wrote:

On 8/26/2021 9:18 AM, Eric Dumazet wrote:

quoted

quoted

+static inline int ixgbe_determine_xdp_q_idx(int cpu)
+{
+    if (static_key_enabled(&ixgbe_xdp_locking_key))
+            return cpu % IXGBE_MAX_XDP_QS;
+    else
+            return cpu;

Even if num_online_cpus() is 8, the returned cpu here could be

0, 32, 64, 96, 128, 161, 197, 224

Are we sure this will still be ok ?

I'm not sure about that one myself. Jason?

quoted

quoted

+}
+
 static inline u8 ixgbe_max_rss_indices(struct ixgbe_adapter *adapter)
 {
     switch (adapter->hw.mac.type) {

diff --git a/drivers/net/ethernet/intel/ixgbe/ixgbe_lib.c b/drivers/net/ethernet/intel/ixgbe/ixgbe_lib.c
index 0218f6c..884bf99 100644
--- a/drivers/net/ethernet/intel/ixgbe/ixgbe_lib.c
+++ b/drivers/net/ethernet/intel/ixgbe/ixgbe_lib.c

@@ -299,7 +299,10 @@ static void ixgbe_cache_ring_register(struct ixgbe_adapter *adapter)

 static int ixgbe_xdp_queues(struct ixgbe_adapter *adapter)
 {
-    return adapter->xdp_prog ? nr_cpu_ids : 0;
+    int queues;
+
+    queues = min_t(int, IXGBE_MAX_XDP_QS, num_online_cpus());

num_online_cpus() might change later...

I saw that too, but I wonder if it doesn't matter to the driver. If a
CPU goes offline or comes online after the driver loads, we will use
this logic to try to pick an available TX queue. But this is a
complicated thing that is easy to get wrong, is there a common example
of how to get it right?

Honestly, I'm a little confused right now. @nr_cpu_ids is the fixed
number which means the total number of cpus the machine has.
I think, using @nr_cpu_ids is safe one way or the other regardless of
whether the cpu goes offline or not. What do you think?

A possible problem I guess is that if the "static_key_enabled" check
returned false in the past, we would need to update that if the number
of CPUs changes, do we need a notifier?

Things get complicated. If the number decreases down to
@IXGBE_MAX_XDP_QS (which is 64), the notifier could be useful because
we wouldn't need to use the @tx_lock. I'm wondering if we really need
to implement one notifier for this kind of change?

Also, now that I'm asking it, I dislike the global as it would apply to
all ixgbe ports and each PF would increment and decrement it
independently. Showing my ignorance here, but I haven't seen this
utility in the kernel before in detail. Not sure if this is "OK" from
multiple device (with the same driver / global namespace) perspective.