On 12 Oct 2020, at 10:21, Sebastian Andrzej Siewior wrote:
On 2020-10-12 10:14:42 [+0200], Eelco Chaudron wrote:
On 9 Oct 2020, at 17:41, Sebastian Andrzej Siewior wrote:
On 2020-10-09 14:47:59 [+0200], Juri Lelli wrote:
This happens because openvswitch/flow_table::flow_lookup() accesses
per-cpu data while being preemptible (and migratable).
Fix it by adding get/put_cpu_light(), so that, even if preempted,
the
task executing this code is not migrated (operation is also guarded
by
ovs_mutex mutex).
This warning is not limited to PREEMPT_RT it also present upstream
since
commit
eac87c413bf97 ("net: openvswitch: reorder masks array based on
usage")
You should be able to reproduce it there, too.
The path ovs_flow_tbl_lookup() -> flow_lookup() is guarded by
ovs_lock()
I can't say that this true for
ovs_vport_receive() -> ovs_dp_process_packet() ->
ovs_flow_tbl_lookup_stats() -> flow_lookup()
(means I don't know but it looks like coming from NAPI).
Which means u64_stats_update_begin() could have two writers. This
must
not happen.
There are two reader which do u64_stats_fetch_begin_irq(). Disabling
interrupts makes no sense since they perform cross-CPU access.
-> You need to ensure that there is only one writer at a time.
If mask_array gains a spinlock_t for writer protection then you can
acquire the lock prio grabbing ->masks_usage_cntr. But as of now
there
is one `ma->syncp'.
I’m not too familiar with the RT kernel, but in the none RT kernel,
this
function is called in run to completion parts only, hence does not
need a
lock. Actually, this was designed in such a way that it does not need
a lock
at all.
_no_ As explained above, this is not RT specific.
What guaranties that you don't have two flow_lookup() invocations on
the
same CPU? You are using u64_stats_update_begin() which must not be
preempted. This means even if preemption is disabled you must not have
another invocation in BH context. This is due to the
write_seqcount_begin() in u64_stats_update_begin().
If preemption / CPU migration is not a problem in the above part, you
can use annotation to disable the warning that led to the warning. But
the u64_stats invocation looks still problematic.
So maybe this needs a get_cpu() instead of the light variant in the
RT case?
Hi Sebastian,
I was not reading the splat correctly and thought it was from the NAPI
path but it looks like it's from the netlink part. I think this could be
fixed with the following patch, so both paths, NAPI, and netlink become
non-preemptive:
--- a/net/openvswitch/flow_table.c
+++ b/net/openvswitch/flow_table.c
@@ -850,9 +850,14 @@ struct sw_flow *ovs_flow_tbl_lookup(struct
flow_table *tbl,
struct mask_array *ma = rcu_dereference_ovsl(tbl->mask_array);
u32 __always_unused n_mask_hit;
u32 __always_unused n_cache_hit;
+ struct sw_flow *flow;
u32 index = 0;
- return flow_lookup(tbl, ti, ma, key, &n_mask_hit, &n_cache_hit,
&index);
+ preempt_disable();
+ flow = flow_lookup(tbl, ti, ma, key, &n_mask_hit, &n_cache_hit,
&index);
+ preempt_enable();
+
+ return flow;
}
struct sw_flow *ovs_flow_tbl_lookup_exact(struct flow_table *tbl,
Note that choosing to do this in ovs_flow_tbl_lookup() and not
flow_lookup() allows the fast-path, through NAPI, not having to call
preempt_disable().
Any comments? If not I can sent a proper patch trough netdev.
//Eelco
