batman-adv: bcast: queue per interface, if needed
PATCH v2 1/1 batman-adv: bcast: queue per interface, if needed
Currently we schedule a broadcast packet like:
3x: [ [(re-)queue] --> for(hard-if): maybe-transmit ]
The intention of queueing a broadcast packet multiple times is to increase robustness for wireless interfaces. However on interfaces which we only broadcast on once the queueing induces an unnecessary penalty. This patch restructures the queueing to be performed on a per interface basis:
for(hard-if):
- transmit
- if wireless: [queue] --> transmit --> [requeue] --> transmit
Next to the performance benefits on non-wireless interfaces this should also make it easier to apply alternative strategies for transmissions on wireless interfaces in the future (for instance sending via unicast transmissions on wireless interfaces, without queueing in batman-adv, if appropriate).
---
Changelog v2:
* removed the other two patches from the patchset for now, only the broadcast queueing cleanup to start with
* fixed spelling of "unnecessary" in commit message (thanks Sven)
* removed now superflous kerneldoc for hard_iface in batadv_forw_packet_bcasts_left() (thanks Sven)
* removed delay check for queued (re)broadcasts in batadv_forw_bcast_packet_if(): the only case where a delay is set for this function is for a delayed, DAT fallback ARP Request from this node, then however num_bcasts will be >=1, too, and the fallback ARP Request will be scheduled anyway
--- net/batman-adv/main.h | 1 - net/batman-adv/routing.c | 9 +- net/batman-adv/send.c | 392 ++++++++++++++++++++++---------- net/batman-adv/send.h | 12 +- net/batman-adv/soft-interface.c | 12 +- 5 files changed, 288 insertions(+), 138 deletions(-)
net/batman-adv/main.h +0 -1
@@ -88,7 +88,6 @@/* number of packets to send for broadcasts on different interface types */ #define BATADV_NUM_BCASTS_DEFAULT 1#define BATADV_NUM_BCASTS_WIRELESS 3 -#define BATADV_NUM_BCASTS_MAX 3/* length of the single packet used by the TP meter */ #define BATADV_TP_PACKET_LEN ETH_DATA_LEN
net/batman-adv/routing.c +6 -3
@@ -1182,9 +1182,9 @@ int batadv_recv_bcast_packet(struct sk_buff *skb,struct batadv_bcast_packet *bcast_packet;struct ethhdr *ethhdr;int hdr_size =sizeof (*bcast_packet); -int ret =NET_RX_DROP ;s32 seq_diff;u32 seqno; +int ret;/* drop packet if it has not necessary minimum size */ if (unlikely (!pskb_may_pull (skb, hdr_size)))
@@ -1210,7 +1210,7 @@ int batadv_recv_bcast_packet(struct sk_buff *skb,if (batadv_is_my_mac (bat_priv, bcast_packet->orig)) goto free_skb; -if (bcast_packet->ttl <2 ) +if (bcast_packet->ttl-- <2 ) goto free_skb; orig_node =batadv_orig_hash_find (bat_priv, bcast_packet->orig);
@@ -1249,7 +1249,9 @@ int batadv_recv_bcast_packet(struct sk_buff *skb,batadv_skb_set_priority (skb,sizeof (struct batadv_bcast_packet ));/* rebroadcast packet */ -batadv_add_bcast_packet_to_list (bat_priv, skb,1 , false); + ret =batadv_forw_bcast_packet (bat_priv, skb,0 , false); +if (ret ==NETDEV_TX_BUSY ) + goto free_skb;/* don't hand the broadcast up if it is from an originator * from the same backbone.
@@ -1275,6 +1277,7 @@ int batadv_recv_bcast_packet(struct sk_buff *skb,spin_unlock_bh (&orig_node->bcast_seqno_lock); free_skb:kfree_skb (skb); + ret =NET_RX_DROP ; out:if (orig_node)batadv_orig_node_put (orig_node);
net/batman-adv/send.c +271 -121
472 lines not on this page · raw message
net/batman-adv/send.h +8 -4
@@ -39,10 +39,14 @@ int batadv_send_broadcast_skb(struct sk_buff *skb, struct batadv_hard_iface *hard_iface);int batadv_send_unicast_skb (struct sk_buff *skb,struct batadv_neigh_node *neigh_node); -int batadv_add_bcast_packet_to_list (struct batadv_priv *bat_priv, -const struct sk_buff *skb, -unsigned long delay, -bool own_packet); +int batadv_forw_bcast_packet (struct batadv_priv *bat_priv, +struct sk_buff *skb, +unsigned long delay, +bool own_packet); +void batadv_send_bcast_packet (struct batadv_priv *bat_priv, +struct sk_buff *skb, +unsigned long delay, +bool own_packet);void batadv_purge_outstanding_packets (struct batadv_priv *bat_priv,const struct batadv_hard_iface *hard_iface);
net/batman-adv/soft-interface.c +3 -9
@@ -191,7 +191,7 @@ static netdev_tx_t batadv_interface_tx(struct sk_buff *skb,struct vlan_ethhdr *vhdr;unsigned int header_len =0 ;int data_len = skb->len, ret; -unsigned long brd_delay =1 ; +unsigned long brd_delay =0 ;bool do_bcast = false, client_added;unsigned short vid;u32 seqno;
@@ -330,7 +330,7 @@ static netdev_tx_t batadv_interface_tx(struct sk_buff *skb, bcast_packet = (struct batadv_bcast_packet *)skb->data; bcast_packet->version =BATADV_COMPAT_VERSION ; - bcast_packet->ttl =BATADV_TTL ; + bcast_packet->ttl =BATADV_TTL -1 ;/* batman packet type: broadcast */ bcast_packet->packet_type =BATADV_BCAST ;
@@ -346,13 +346,7 @@ static netdev_tx_t batadv_interface_tx(struct sk_buff *skb, seqno =atomic_inc_return (&bat_priv->bcast_seqno); bcast_packet->seqno =htonl (seqno); -batadv_add_bcast_packet_to_list (bat_priv, skb, brd_delay, true); - -/* a copy is stored in the bcast list, therefore removing -* the original skb. -*/ -consume_skb (skb); - +batadv_send_bcast_packet (bat_priv, skb, brd_delay, true);/* unicast packet */ }else {/* DHCP packets going to a server will use the GW feature */
Signature · 1 line
2.31.0
Sven Eckelmann
maintainerQuoted 5 lines · On Tuesday, 27 April 2021 20:45:27 CEST Linus Lüssing wrote:
- * The skb is not consumed, so the caller should make sure that the
- * skb is freed.
+ * This call clones the given skb, hence the caller needs to take into
+ * account that the data segment of the original skb might not be
+ * modifiable anymore.
But none of your callers is now taking care of it because you've removed all skb_copy's. All you do is to clone the control data and give it to the underlying layers. And they may write freely to the data. Thus breaking parallel (and under some circumstances sequential) running code which operates on the skbs.
The review was stopped after noticing this problem.
Kind regards,
Sven
Quoted 7 lines · Sven Eckelmann
Quoted 5 lines · On Tuesday, 27 April 2021 20:45:27 CEST Linus Lüssing wrote:
- * The skb is not consumed, so the caller should make sure that the
- * skb is freed.
+ * This call clones the given skb, hence the caller needs to take into
+ * account that the data segment of the original skb might not be
+ * modifiable anymore.
Quoted 5 lines
But none of your callers is now taking care of it because you've removed all skb_copy's. All you do is to clone the control data and give it to the underlying layers. And they may write freely to the data. Thus breaking parallel (and under some circumstances sequential) running code which operates on the skbs.
Hi Sven,
Thanks for looking at it so far. I'm not quite sure if the skb_copy() is needed though. Because there is a new skb_cow(). Let me explain my thoughts:
We have two cases: A) Packet originating from ourself, via batadv_interface_tx(). Or B) received+forwarded from a neighbor node via batadv_recv_bcast_packet().
For case A), self generated:
When we send the packet multiple times, for each rebroadcast or interface we will push the ethernet header and write the ether-src, ether-dest and ether protocol in batadv_send_skb_packet(). Before that batadv_send_skb_packet() calls batadv_skb_head_push() which calls skb_cow_head(). So the ethernet header should be modifiable safely then, even if it is an skb clone.
For case B), received/forwarded:
Rebroadcasts same as in A), but additionally after rebroadcasting with potential requeuing in batadv_recv_bcast_packet() via batadv_forw_bcast_packet() we will also call batadv_interface_rx() and strip the batman header. Betweeen these calls there is the following though:
batadv_forw_bcast_packet(skb, ...)
-> __batadv_forw_bcast_packet(skb, ...);
...
skb_cow(skb, 0)
So the original skb will have been made uncloned/writeable again via the skb_cow() before being handed to batadv_interface_rx().
Let me know if I'm missing something.
Regards, Linus
Sven Eckelmann
maintainerQuoted 14 lines · On Friday, 14 May 2021 13:35:32 CEST Linus Lüssing wrote:
Quoted 7 lines · Sven Eckelmann
Quoted 5 lines · On Tuesday, 27 April 2021 20:45:27 CEST Linus Lüssing wrote:
- * The skb is not consumed, so the caller should make sure that the
- * skb is freed.
+ * This call clones the given skb, hence the caller needs to take into
+ * account that the data segment of the original skb might not be
+ * modifiable anymore.Quoted 5 lines
But none of your callers is now taking care of it because you've removed all skb_copy's. All you do is to clone the control data and give it to the underlying layers. And they may write freely to the data. Thus breaking parallel (and under some circumstances sequential) running code which operates on the skbs.
Quoted 5 lines
Hi Sven,
Thanks for looking at it so far. I'm not quite sure if the
skb_copy()is needed though. Because there is a newskb_cow(). Let me explain my thoughts:
[...]
Haven't checked this in detail. But please split this part in a separate patch with the corresponding explanation. Should make it easier to understand why the skb_copy is no longer needed.
Kind regards,
Sven