net: flow: Prevent bringing up new CPUs during per-CPU initialization
In a rare race-condition, it's possible that a CPU will be brought
online between when the for_each_online_cpu() loop executes to
call flow_cache_cpu_prepare(), and when the hotplug notifier for
calling this same function is registered. If this happens,
flow_cache_cpu_prepare() will never be called for the new CPU,
resulting in crashes due to uninitialized per-cpu data.
Fix this by preventing CPUs from being added or removed during
this small but sensitive window.
Change-Id: Iafbbaa8a50e5c527392d130561874313720849d0
Signed-off-by: Matt Wagantall <mattw@codeaurora.org>
diff --git a/net/core/flow.c b/net/core/flow.c
index e318c7e..9a517c6 100644
--- a/net/core/flow.c
+++ b/net/core/flow.c
@@ -423,6 +423,7 @@
if (!fc->percpu)
return -ENOMEM;
+ get_online_cpus();
for_each_online_cpu(i) {
if (flow_cache_cpu_prepare(fc, i))
goto err;
@@ -431,6 +432,7 @@
.notifier_call = flow_cache_cpu,
};
register_hotcpu_notifier(&fc->hotcpu_notifier);
+ put_online_cpus();
setup_timer(&fc->rnd_timer, flow_cache_new_hashrnd,
(unsigned long) fc);
@@ -440,6 +442,7 @@
return 0;
err:
+ put_online_cpus();
for_each_possible_cpu(i) {
struct flow_cache_percpu *fcp = per_cpu_ptr(fc->percpu, i);
kfree(fcp->hash_table);