Possible read of uninitialized array values in reuseport_select_sock?

Hello,

While reading the code of net/core/sock_reuseport.c, I think I found a
possible race in reuseport_select_sock. The current code has the
following pattern:

   socks = READ_ONCE(reuse->num_socks);
   smp_rmb(); // paired with reuseport_add_sock to make sure
reuse->socks[i < num_socks] are initialized
   while (reuse->socks[i]->sk_state == TCP_ESTABLISHED) {
        if (i >= reuse->num_socks) // should be "socks" and not
"reuse->num_socks"?

The barrier seems to be here to make sure that all items of
reuse->socks are initialized before being read, but the barrier only
protects indexes < socks, not indexes < reuse->num_socks.

I have a patch ready to fix this issue, but I wanted to confirm that
the current code is indeed incorrect (if the code is correct for a
non-obvious reason, I'd be happy to add some comments to document the
behavior).


Here is the diff of the patch I was planning to submit:

diff --git a/net/core/sock_reuseport.c b/net/core/sock_reuseport.c
index bbdd3c7b6cb5..b065f0a103ed 100644
--- a/net/core/sock_reuseport.c
+++ b/net/core/sock_reuseport.c

@@ -293,7 +293,7 @@ struct sock *reuseport_select_sock(struct sock *sk,
             i = j = reciprocal_scale(hash, socks);
             while (reuse->socks[i]->sk_state == TCP_ESTABLISHED) {
                 i++;
-                if (i >= reuse->num_socks)
+                if (i >= socks)
                     i = 0;
                 if (i == j)
                     goto out;

Thanks,
Baptiste.

On Wed, Jan 6, 2021 at 10:54 PM Baptiste Lepers
[off-list ref] wrote:

Thanks for the clear description. Yes, I believe you're correct.

Thanks. I will submit a patch.

Baptiste.

On Thu, Jan 7, 2021 at 3:57 PM Willem de Bruijn
[off-list ref] wrote:

On Wed, Jan 6, 2021 at 10:54 PM Baptiste Lepers
[off-list ref] wrote:

quoted

Hello,

While reading the code of net/core/sock_reuseport.c, I think I found a
possible race in reuseport_select_sock. The current code has the
following pattern:

   socks = READ_ONCE(reuse->num_socks);
   smp_rmb(); // paired with reuseport_add_sock to make sure
reuse->socks[i < num_socks] are initialized
   while (reuse->socks[i]->sk_state == TCP_ESTABLISHED) {
        if (i >= reuse->num_socks) // should be "socks" and not
"reuse->num_socks"?

The barrier seems to be here to make sure that all items of
reuse->socks are initialized before being read, but the barrier only
protects indexes < socks, not indexes < reuse->num_socks.

I have a patch ready to fix this issue, but I wanted to confirm that
the current code is indeed incorrect (if the code is correct for a
non-obvious reason, I'd be happy to add some comments to document the
behavior).


Here is the diff of the patch I was planning to submit:

diff --git a/net/core/sock_reuseport.c b/net/core/sock_reuseport.c
index bbdd3c7b6cb5..b065f0a103ed 100644
--- a/net/core/sock_reuseport.c
+++ b/net/core/sock_reuseport.c

@@ -293,7 +293,7 @@ struct sock *reuseport_select_sock(struct sock *sk,
             i = j = reciprocal_scale(hash, socks);
             while (reuse->socks[i]->sk_state == TCP_ESTABLISHED) {
                 i++;
-                if (i >= reuse->num_socks)
+                if (i >= socks)
                     i = 0;
                 if (i == j)
                     goto out;

Thanks,
Baptiste.

Thanks for the clear description. Yes, I believe you're correct.