* [PATCH bpf] net, xsk: Avoid taking multiple skbuff references
@ 2020-11-23 13:12 Björn Töpel
2020-11-23 13:53 ` Daniel Borkmann
0 siblings, 1 reply; 3+ messages in thread
From: Björn Töpel @ 2020-11-23 13:12 UTC (permalink / raw)
To: ast, daniel, netdev, bpf
Cc: Björn Töpel, jonathan.lemon, yhs, weqaar.janjua,
magnus.karlsson, weqaar.a.janjua
From: Björn Töpel <bjorn.topel@intel.com>
Commit 642e450b6b59 ("xsk: Do not discard packet when NETDEV_TX_BUSY")
addressed the problem that packets were discarded from the Tx AF_XDP
ring, when the driver returned NETDEV_TX_BUSY. Part of the fix was
bumping the skbuff reference count, so that the buffer would not be
freed by dev_direct_xmit(). A reference count larger than one means
that the skbuff is "shared", which is not the case.
If the "shared" skbuff is sent to the generic XDP receive path,
netif_receive_generic_xdp(), and pskb_expand_head() is entered the
BUG_ON(skb_shared(skb)) will trigger.
This patch adds a variant to dev_direct_xmit(), __dev_direct_xmit(),
where a user can select the skbuff free policy. This allows AF_XDP to
avoid bumping the reference count, but still keep the NETDEV_TX_BUSY
behavior.
Reported-by: Yonghong Song <yhs@fb.com>
Fixes: 642e450b6b59 ("xsk: Do not discard packet when NETDEV_TX_BUSY")
Signed-off-by: Björn Töpel <bjorn.topel@intel.com>
---
include/linux/netdevice.h | 1 +
net/core/dev.c | 9 +++++++--
net/xdp/xsk.c | 8 +-------
3 files changed, 9 insertions(+), 9 deletions(-)
diff --git a/include/linux/netdevice.h b/include/linux/netdevice.h
index 964b494b0e8d..e7402fca7752 100644
--- a/include/linux/netdevice.h
+++ b/include/linux/netdevice.h
@@ -2815,6 +2815,7 @@ u16 dev_pick_tx_cpu_id(struct net_device *dev, struct sk_buff *skb,
struct net_device *sb_dev);
int dev_queue_xmit(struct sk_buff *skb);
int dev_queue_xmit_accel(struct sk_buff *skb, struct net_device *sb_dev);
+int __dev_direct_xmit(struct sk_buff *skb, u16 queue_id, bool free_on_busy);
int dev_direct_xmit(struct sk_buff *skb, u16 queue_id);
int register_netdevice(struct net_device *dev);
void unregister_netdevice_queue(struct net_device *dev, struct list_head *head);
diff --git a/net/core/dev.c b/net/core/dev.c
index 82dc6b48e45f..2af79a4253bb 100644
--- a/net/core/dev.c
+++ b/net/core/dev.c
@@ -4180,7 +4180,7 @@ int dev_queue_xmit_accel(struct sk_buff *skb, struct net_device *sb_dev)
}
EXPORT_SYMBOL(dev_queue_xmit_accel);
-int dev_direct_xmit(struct sk_buff *skb, u16 queue_id)
+int __dev_direct_xmit(struct sk_buff *skb, u16 queue_id, bool free_on_busy)
{
struct net_device *dev = skb->dev;
struct sk_buff *orig_skb = skb;
@@ -4211,7 +4211,7 @@ int dev_direct_xmit(struct sk_buff *skb, u16 queue_id)
local_bh_enable();
- if (!dev_xmit_complete(ret))
+ if (free_on_busy && !dev_xmit_complete(ret))
kfree_skb(skb);
return ret;
@@ -4220,6 +4220,11 @@ int dev_direct_xmit(struct sk_buff *skb, u16 queue_id)
kfree_skb_list(skb);
return NET_XMIT_DROP;
}
+
+int dev_direct_xmit(struct sk_buff *skb, u16 queue_id)
+{
+ return __dev_direct_xmit(skb, queue_id, true);
+}
EXPORT_SYMBOL(dev_direct_xmit);
/*************************************************************************
diff --git a/net/xdp/xsk.c b/net/xdp/xsk.c
index 5a6cdf7b320d..c6ad31b374b7 100644
--- a/net/xdp/xsk.c
+++ b/net/xdp/xsk.c
@@ -411,11 +411,7 @@ static int xsk_generic_xmit(struct sock *sk)
skb_shinfo(skb)->destructor_arg = (void *)(long)desc.addr;
skb->destructor = xsk_destruct_skb;
- /* Hinder dev_direct_xmit from freeing the packet and
- * therefore completing it in the destructor
- */
- refcount_inc(&skb->users);
- err = dev_direct_xmit(skb, xs->queue_id);
+ err = __dev_direct_xmit(skb, xs->queue_id, false);
if (err == NETDEV_TX_BUSY) {
/* Tell user-space to retry the send */
skb->destructor = sock_wfree;
@@ -429,12 +425,10 @@ static int xsk_generic_xmit(struct sock *sk)
/* Ignore NET_XMIT_CN as packet might have been sent */
if (err == NET_XMIT_DROP) {
/* SKB completed but not sent */
- kfree_skb(skb);
err = -EBUSY;
goto out;
}
- consume_skb(skb);
sent_frame = true;
}
base-commit: 178648916e73e00de83150eb0c90c0d3a977a46a
--
2.27.0
^ permalink raw reply related [flat|nested] 3+ messages in thread* Re: [PATCH bpf] net, xsk: Avoid taking multiple skbuff references
2020-11-23 13:12 [PATCH bpf] net, xsk: Avoid taking multiple skbuff references Björn Töpel
@ 2020-11-23 13:53 ` Daniel Borkmann
2020-11-23 14:05 ` Björn Töpel
0 siblings, 1 reply; 3+ messages in thread
From: Daniel Borkmann @ 2020-11-23 13:53 UTC (permalink / raw)
To: Björn Töpel, ast, netdev, bpf
Cc: Björn Töpel, jonathan.lemon, yhs, weqaar.janjua,
magnus.karlsson, weqaar.a.janjua
On 11/23/20 2:12 PM, Björn Töpel wrote:
> From: Björn Töpel <bjorn.topel@intel.com>
>
> Commit 642e450b6b59 ("xsk: Do not discard packet when NETDEV_TX_BUSY")
> addressed the problem that packets were discarded from the Tx AF_XDP
> ring, when the driver returned NETDEV_TX_BUSY. Part of the fix was
> bumping the skbuff reference count, so that the buffer would not be
> freed by dev_direct_xmit(). A reference count larger than one means
> that the skbuff is "shared", which is not the case.
>
> If the "shared" skbuff is sent to the generic XDP receive path,
> netif_receive_generic_xdp(), and pskb_expand_head() is entered the
> BUG_ON(skb_shared(skb)) will trigger.
>
> This patch adds a variant to dev_direct_xmit(), __dev_direct_xmit(),
> where a user can select the skbuff free policy. This allows AF_XDP to
> avoid bumping the reference count, but still keep the NETDEV_TX_BUSY
> behavior.
>
> Reported-by: Yonghong Song <yhs@fb.com>
> Fixes: 642e450b6b59 ("xsk: Do not discard packet when NETDEV_TX_BUSY")
> Signed-off-by: Björn Töpel <bjorn.topel@intel.com>
> ---
> include/linux/netdevice.h | 1 +
> net/core/dev.c | 9 +++++++--
> net/xdp/xsk.c | 8 +-------
> 3 files changed, 9 insertions(+), 9 deletions(-)
>
> diff --git a/include/linux/netdevice.h b/include/linux/netdevice.h
> index 964b494b0e8d..e7402fca7752 100644
> --- a/include/linux/netdevice.h
> +++ b/include/linux/netdevice.h
> @@ -2815,6 +2815,7 @@ u16 dev_pick_tx_cpu_id(struct net_device *dev, struct sk_buff *skb,
> struct net_device *sb_dev);
> int dev_queue_xmit(struct sk_buff *skb);
> int dev_queue_xmit_accel(struct sk_buff *skb, struct net_device *sb_dev);
> +int __dev_direct_xmit(struct sk_buff *skb, u16 queue_id, bool free_on_busy);
> int dev_direct_xmit(struct sk_buff *skb, u16 queue_id);
> int register_netdevice(struct net_device *dev);
> void unregister_netdevice_queue(struct net_device *dev, struct list_head *head);
> diff --git a/net/core/dev.c b/net/core/dev.c
> index 82dc6b48e45f..2af79a4253bb 100644
> --- a/net/core/dev.c
> +++ b/net/core/dev.c
> @@ -4180,7 +4180,7 @@ int dev_queue_xmit_accel(struct sk_buff *skb, struct net_device *sb_dev)
> }
> EXPORT_SYMBOL(dev_queue_xmit_accel);
>
> -int dev_direct_xmit(struct sk_buff *skb, u16 queue_id)
> +int __dev_direct_xmit(struct sk_buff *skb, u16 queue_id, bool free_on_busy)
> {
> struct net_device *dev = skb->dev;
> struct sk_buff *orig_skb = skb;
> @@ -4211,7 +4211,7 @@ int dev_direct_xmit(struct sk_buff *skb, u16 queue_id)
>
> local_bh_enable();
>
> - if (!dev_xmit_complete(ret))
> + if (free_on_busy && !dev_xmit_complete(ret))
> kfree_skb(skb);
>
> return ret;
Hm, but this way free_on_busy, even though constant, cannot be optimized away?
Can't you just move the dev_xmit_complete() check out into dev_direct_xmit()
instead? That way you can just drop the bool, and the below dev_direct_xmit()
should probably just become an __always_line function in netdevice.h so you
avoid the double call.
> @@ -4220,6 +4220,11 @@ int dev_direct_xmit(struct sk_buff *skb, u16 queue_id)
> kfree_skb_list(skb);
> return NET_XMIT_DROP;
> }
> +
> +int dev_direct_xmit(struct sk_buff *skb, u16 queue_id)
> +{
> + return __dev_direct_xmit(skb, queue_id, true);
> +}
> EXPORT_SYMBOL(dev_direct_xmit);
>
> /*************************************************************************
> diff --git a/net/xdp/xsk.c b/net/xdp/xsk.c
> index 5a6cdf7b320d..c6ad31b374b7 100644
> --- a/net/xdp/xsk.c
> +++ b/net/xdp/xsk.c
> @@ -411,11 +411,7 @@ static int xsk_generic_xmit(struct sock *sk)
> skb_shinfo(skb)->destructor_arg = (void *)(long)desc.addr;
> skb->destructor = xsk_destruct_skb;
>
> - /* Hinder dev_direct_xmit from freeing the packet and
> - * therefore completing it in the destructor
> - */
> - refcount_inc(&skb->users);
> - err = dev_direct_xmit(skb, xs->queue_id);
> + err = __dev_direct_xmit(skb, xs->queue_id, false);
> if (err == NETDEV_TX_BUSY) {
> /* Tell user-space to retry the send */
> skb->destructor = sock_wfree;
> @@ -429,12 +425,10 @@ static int xsk_generic_xmit(struct sock *sk)
> /* Ignore NET_XMIT_CN as packet might have been sent */
> if (err == NET_XMIT_DROP) {
> /* SKB completed but not sent */
> - kfree_skb(skb);
> err = -EBUSY;
> goto out;
> }
>
> - consume_skb(skb);
> sent_frame = true;
> }
>
>
> base-commit: 178648916e73e00de83150eb0c90c0d3a977a46a
>
^ permalink raw reply [flat|nested] 3+ messages in thread* Re: [PATCH bpf] net, xsk: Avoid taking multiple skbuff references
2020-11-23 13:53 ` Daniel Borkmann
@ 2020-11-23 14:05 ` Björn Töpel
0 siblings, 0 replies; 3+ messages in thread
From: Björn Töpel @ 2020-11-23 14:05 UTC (permalink / raw)
To: Daniel Borkmann, Björn Töpel, ast, netdev, bpf
Cc: jonathan.lemon, yhs, weqaar.janjua, magnus.karlsson, weqaar.a.janjua
On 2020-11-23 14:53, Daniel Borkmann wrote:
[...]
>
> Hm, but this way free_on_busy, even though constant, cannot be optimized
> away?
> Can't you just move the dev_xmit_complete() check out into
> dev_direct_xmit()
> instead? That way you can just drop the bool, and the below
> dev_direct_xmit()
> should probably just become an __always_line function in netdevice.h so you
> avoid the double call.
>
Good suggestion! I'll spin a v2.
Björn
^ permalink raw reply [flat|nested] 3+ messages in thread
end of thread, other threads:[~2020-11-23 14:07 UTC | newest]
Thread overview: 3+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2020-11-23 13:12 [PATCH bpf] net, xsk: Avoid taking multiple skbuff references Björn Töpel
2020-11-23 13:53 ` Daniel Borkmann
2020-11-23 14:05 ` Björn Töpel
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).