Thread (288 messages) 288 messages, 13 authors, 2014-04-01

Re: [PATCH 6/6] net: Free skbs from irqs when possible.

From: Eric Dumazet <hidden>
Date: 2014-03-18 13:22:39

On Mon, 2014-03-17 at 23:27 -0700, Eric W. Biederman wrote:
quoted hunk ↗ jump to hunk
Add a test skb_irq_freeable to report when it is safe to free a skb
from irq context.

It is not safe to free an skb from irq context when:
- The skb has a destructor as some skb destructors call local_bh_disable
  or spin_lock_bh.
- There is xfrm state as __xfrm_state_destroy calls spin_lock_bh.
- There is netfilter conntrack state as destroy_conntrack calls
  spin_lock_bh.
- If there is a refcounted dst entry on the skb, as __dst_free
  calls spin_lock_bh.
- If there is a frag_list, which could be a list of any skbs.
Otherwise it appears safe to free a skb from interrupt context.

- Update the warning in skb_releae_head_state to warn about freeing
  skb's in the wrong context.

- Update __dev_kfree_skb_irq to free all skbs that it can immediately

- Kill zap_completion_queue because there is no point going through
  a queue of packets that are not safe to free and looking for packets
  that are safe to free.

Signed-off-by: "Eric W. Biederman" <redacted>
---
 include/linux/skbuff.h |   13 +++++++++++++
 net/core/dev.c         |   14 +++++++++-----
 net/core/netpoll.c     |   32 --------------------------------
 net/core/skbuff.c      |   13 ++++++++++---
 4 files changed, 32 insertions(+), 40 deletions(-)
diff --git a/include/linux/skbuff.h b/include/linux/skbuff.h
index 03db95ab8a8c..53f72b53fd47 100644
--- a/include/linux/skbuff.h
+++ b/include/linux/skbuff.h
@@ -2833,6 +2833,19 @@ static inline void skb_init_secmark(struct sk_buff *skb)
 { }
 #endif
 
+static inline bool skb_irq_freeable(struct sk_buff *skb)
+{
+	return !skb->destructor &&
+#if IS_ENABLED(CONFIG_XFRM)
+		!skb->sp &&
+#endif
+#if IS_ENABLED(CONFIG_NF_CONNTRACK)
+		!skb->nfct &&
+#endif
+		(!skb->_skb_refdst || (skb->_skb_refdst & SKB_DST_NOREF)) &&
+		!skb_has_frag_list(skb);
+}
+
It would be a serious bug having (skb->_skb_refdst & SKB_DST_NOREF) at
this point. dst would be RCU protected, but this can not be true as the
packet was queued in TX ring buffer for a possibly long period. And even
before reaching the driver, skb might have been queued in qdisc layer
and escape rcu protection section anyway.

Thats why we use skb_dst_force() from __dev_xmit_skb()
Keyboard shortcuts
hback out one level
jnext message in thread
kprevious message in thread
ldrill in
Escclose help / fold thread tree
?toggle this help