Re: [PATCH net] net: let skb_orphan_partial wake-up waiters.

On Tue, Mar 30, 2021 at 5:18 PM Paolo Abeni [off-list ref] wrote:

On Tue, 2021-03-30 at 16:40 +0200, Eric Dumazet wrote:

quoted

On Tue, Mar 30, 2021 at 4:39 PM Eric Dumazet [off-list ref] wrote:

quoted

On Tue, Mar 30, 2021 at 4:25 PM Paolo Abeni [off-list ref] wrote:

quoted

Currently the mentioned helper can end-up freeing the socket wmem
without waking-up any processes waiting for more write memory.

If the partially orphaned skb is attached to an UDP (or raw) socket,
the lack of wake-up can hang the user-space.

Address the issue invoking the write_space callback after
releasing the memory, if the old skb destructor requires that.

Fixes: f6ba8d33cfbb ("netem: fix skb_orphan_partial()")
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
---
 net/core/sock.c | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/net/core/sock.c b/net/core/sock.c
index 0ed98f20448a2..7a38332d748e7 100644
--- a/net/core/sock.c
+++ b/net/core/sock.c

@@ -2137,6 +2137,8 @@ void skb_orphan_partial(struct sk_buff *skb)

                if (refcount_inc_not_zero(&sk->sk_refcnt)) {
                        WARN_ON(refcount_sub_and_test(skb->truesize, &sk->sk_wmem_alloc));
+                       if (skb->destructor == sock_wfree)
+                               sk->sk_write_space(sk);

Interesting.

Why TCP is not a problem here ?

AFAICS, tcp_wfree() does not call sk->sk_write_space(). Processes
waiting for wmem are woken by ack processing.

quoted

quoted

I would rather replace WARN_ON(refcount_sub_and_test(skb->truesize,
&sk->sk_wmem_alloc)) by :
                        skb_orphan(skb);

And of course re-add
                        skb->sk = sk;

Double checking to be sure. The patched slice of skb_orphan_partial()
will then look like:

        if (can_skb_orphan_partial(skb)) {
                struct sock *sk = skb->sk;

                if (refcount_inc_not_zero(&sk->sk_refcnt)) {
                        skb_orphan(skb);
                        skb->sk = sk;
                        skb->destructor = sock_efree;
                }
        } // ...

Am I correct?

Yes.

We also could add a helper for the whole construct, since many other
paths do almost the same.
(They might use sock_hold(), but it seems safe to use the
refcount_inc_not_zero())
Or they omit the skb_orphan() (see can_skb_set_owner), which seems also risky.

static inline void skb_set_owner_sk_safe(sk, skb)
{
    if (sk && refcount_inc_not_zero(&sk->sk_refcnt)) {
           skb_orphan(skb);
           skb->destructor = sock_efree;
          skb->sk = sk;
    }
}

Thanks!

Paolo