| Message ID | 1664093662-32069-2-git-send-email-paulb@nvidia.com (mailing list archive) |
|---|---|
| State | Superseded |
| Delegated to: | Netdev Maintainers |
| Headers | show |
| Series | net: Fix return value of qdisc ingress handling on success | expand |
On Sun, Sep 25, 2022 at 11:14:21AM +0300, Paul Blakey wrote: > Currently qdisc ingress handling (sch_handle_ingress()) doesn't > set a return value and it is left to the old return value of > the caller (__netif_receive_skb_core()) which is RX drop, so if > the packet is consumed, caller will stop and return this value > as if the packet was dropped. > > This causes a problem in the kernel tcp stack when having a > egress tc rule forwarding to a ingress tc rule. > The tcp stack sending packets on the device having the egress rule > will see the packets as not successfully transmitted (although they > actually were), will not advance it's internal state of sent data, > and packets returning on such tcp stream will be dropped by the tcp > stack with reason ack-of-unsent-data. See reproduction in [0] below. > Hm, but how is this return value propagated to egress? I checked tcf_mirred_act() code, but don't see how it is even used there. 318 err = tcf_mirred_forward(want_ingress, skb2); 319 if (err) { 320 out: 321 tcf_action_inc_overlimit_qstats(&m->common); 322 if (tcf_mirred_is_act_redirect(m_eaction)) 323 retval = TC_ACT_SHOT; 324 } 325 __this_cpu_dec(mirred_rec_level); 326 327 return retval; What am I missing? Also, the offending commit is very old and this configuration is not uncommon at all, how could we even not notice this for such a long time? Thanks.
On 25/09/2022 21:00, Cong Wang wrote: > On Sun, Sep 25, 2022 at 11:14:21AM +0300, Paul Blakey wrote: >> Currently qdisc ingress handling (sch_handle_ingress()) doesn't >> set a return value and it is left to the old return value of >> the caller (__netif_receive_skb_core()) which is RX drop, so if >> the packet is consumed, caller will stop and return this value >> as if the packet was dropped. >> >> This causes a problem in the kernel tcp stack when having a >> egress tc rule forwarding to a ingress tc rule. >> The tcp stack sending packets on the device having the egress rule >> will see the packets as not successfully transmitted (although they >> actually were), will not advance it's internal state of sent data, >> and packets returning on such tcp stream will be dropped by the tcp >> stack with reason ack-of-unsent-data. See reproduction in [0] below. >> > > Hm, but how is this return value propagated to egress? I checked > tcf_mirred_act() code, but don't see how it is even used there. > > 318 err = tcf_mirred_forward(want_ingress, skb2); > 319 if (err) { > 320 out: > 321 tcf_action_inc_overlimit_qstats(&m->common); > 322 if (tcf_mirred_is_act_redirect(m_eaction)) > 323 retval = TC_ACT_SHOT; > 324 } > 325 __this_cpu_dec(mirred_rec_level); > 326 > 327 return retval; > > > What am I missing? for the ingress acting act_mirred it will return TC_ACT_CONSUMED above the code you mentioned (since redirect=1, use_reinsert=1. Although TC_ACT_STOLEN which is the retval set for this action, will also act the same) It is propagated as such (TX stack starting from tcp): __tcp_transmit_skb ip_queue_xmit __ip_queue_xmit ip_local_out dst_output ip_output ip_finish_output ip_finish_output2 neigh_output neigh_hh_output dev_queue_xmit sch_handle_egress tcf_classify tcf_mirred_act tcf_mirred_forward netif_receive_skb # here we moved to ingress processing netif_receive_skb_internal __netif_receive_skb __netif_receive_skb_core sch_handle_ingress tcf_classify tcf_mirred_act tcf_mirred_forward tcf_dev_queue_xmit dev_queue_xmit # sends packet ... return TC_ACT_CONSUMED return NULL, and leaves *ret untouched return NET_RX_DROP ... > > Also, the offending commit is very old and this configuration is not > uncommon at all, how could we even not notice this for such a long time? I blamed the commit that left the ret value unset for the first time, but normally netif_receive_skb return value is ignored (as mentioned in the comment above it), this time it isn't ignored because it was chained to egress processing of the tcp transmit path. And this specific case where I chain tc ingress to tc egress came much later (with the commit I blamed in v1 of this patch). > > Thanks.
On Wed, Sep 28, 2022 at 10:55:49AM +0300, Paul Blakey wrote: > > > On 25/09/2022 21:00, Cong Wang wrote: > > On Sun, Sep 25, 2022 at 11:14:21AM +0300, Paul Blakey wrote: > > > Currently qdisc ingress handling (sch_handle_ingress()) doesn't > > > set a return value and it is left to the old return value of > > > the caller (__netif_receive_skb_core()) which is RX drop, so if > > > the packet is consumed, caller will stop and return this value > > > as if the packet was dropped. > > > > > > This causes a problem in the kernel tcp stack when having a > > > egress tc rule forwarding to a ingress tc rule. > > > The tcp stack sending packets on the device having the egress rule > > > will see the packets as not successfully transmitted (although they > > > actually were), will not advance it's internal state of sent data, > > > and packets returning on such tcp stream will be dropped by the tcp > > > stack with reason ack-of-unsent-data. See reproduction in [0] below. > > > > > > > Hm, but how is this return value propagated to egress? I checked > > tcf_mirred_act() code, but don't see how it is even used there. > > > > 318 err = tcf_mirred_forward(want_ingress, skb2); > > 319 if (err) { > > 320 out: > > 321 tcf_action_inc_overlimit_qstats(&m->common); > > 322 if (tcf_mirred_is_act_redirect(m_eaction)) > > 323 retval = TC_ACT_SHOT; > > 324 } > > 325 __this_cpu_dec(mirred_rec_level); > > 326 > > 327 return retval; > > > > > > What am I missing? > > for the ingress acting act_mirred it will return TC_ACT_CONSUMED above > the code you mentioned (since redirect=1, use_reinsert=1. Although > TC_ACT_STOLEN which is the retval set for this action, will also act the > same) > > > It is propagated as such (TX stack starting from tcp): > Sorry for my misunderstanding. I meant to say those TC_ACT_* return value, not NET_RX_*, but I worried too much here, as mirred lets user specify the return value. BTW, it seems you at least miss the drop case, which is NET_RX_DROP for TC_ACT_SHOT at least? Possibly other code paths in sch_handle_ingress() too. Thanks.
On Sat, Oct 01, 2022 at 01:19:50PM -0700, Cong Wang wrote: > On Wed, Sep 28, 2022 at 10:55:49AM +0300, Paul Blakey wrote: > > > > > > On 25/09/2022 21:00, Cong Wang wrote: > > > On Sun, Sep 25, 2022 at 11:14:21AM +0300, Paul Blakey wrote: > > > > Currently qdisc ingress handling (sch_handle_ingress()) doesn't > > > > set a return value and it is left to the old return value of > > > > the caller (__netif_receive_skb_core()) which is RX drop, so if > > > > the packet is consumed, caller will stop and return this value > > > > as if the packet was dropped. > > > > > > > > This causes a problem in the kernel tcp stack when having a > > > > egress tc rule forwarding to a ingress tc rule. > > > > The tcp stack sending packets on the device having the egress rule > > > > will see the packets as not successfully transmitted (although they > > > > actually were), will not advance it's internal state of sent data, > > > > and packets returning on such tcp stream will be dropped by the tcp > > > > stack with reason ack-of-unsent-data. See reproduction in [0] below. > > > > > > > > > > Hm, but how is this return value propagated to egress? I checked > > > tcf_mirred_act() code, but don't see how it is even used there. > > > > > > 318 err = tcf_mirred_forward(want_ingress, skb2); > > > 319 if (err) { > > > 320 out: > > > 321 tcf_action_inc_overlimit_qstats(&m->common); > > > 322 if (tcf_mirred_is_act_redirect(m_eaction)) > > > 323 retval = TC_ACT_SHOT; > > > 324 } > > > 325 __this_cpu_dec(mirred_rec_level); > > > 326 > > > 327 return retval; > > > > > > > > > What am I missing? > > > > for the ingress acting act_mirred it will return TC_ACT_CONSUMED above > > the code you mentioned (since redirect=1, use_reinsert=1. Although > > TC_ACT_STOLEN which is the retval set for this action, will also act the > > same) > > > > > > It is propagated as such (TX stack starting from tcp): > > > > Sorry for my misunderstanding. > > I meant to say those TC_ACT_* return value, not NET_RX_*, but I worried > too much here, as mirred lets user specify the return value. > > BTW, it seems you at least miss the drop case, which is NET_RX_DROP for > TC_ACT_SHOT at least? Possibly other code paths in sch_handle_ingress() > too. > I mean: diff --git a/net/core/dev.c b/net/core/dev.c index fa53830d0683..d1db8210d671 100644 --- a/net/core/dev.c +++ b/net/core/dev.c @@ -5109,6 +5109,7 @@ sch_handle_ingress(struct sk_buff *skb, struct packet_type **pt_prev, int *ret, struct mini_Qdisc *miniq = rcu_dereference_bh(skb->dev->miniq_ingress); struct tcf_result cl_res; + *ret = NET_RX_SUCCESS; /* If there's at least one ingress present somewhere (so * we get here via enabled static key), remaining devices * that are not configured with an ingress qdisc will bail @@ -5136,6 +5137,7 @@ sch_handle_ingress(struct sk_buff *skb, struct packet_type **pt_prev, int *ret, case TC_ACT_SHOT: mini_qdisc_qstats_cpu_drop(miniq); kfree_skb_reason(skb, SKB_DROP_REASON_TC_INGRESS); + *ret = NET_RX_DROP; return NULL; case TC_ACT_STOLEN: case TC_ACT_QUEUED: @@ -5160,6 +5162,7 @@ sch_handle_ingress(struct sk_buff *skb, struct packet_type **pt_prev, int *ret, break; } #endif /* CONFIG_NET_CLS_ACT */ + *ret = NET_RX_SUCCESS; return skb; }
On 01/10/2022 23:19, Cong Wang wrote: > On Wed, Sep 28, 2022 at 10:55:49AM +0300, Paul Blakey wrote: >> >> >> On 25/09/2022 21:00, Cong Wang wrote: >>> On Sun, Sep 25, 2022 at 11:14:21AM +0300, Paul Blakey wrote: >>>> Currently qdisc ingress handling (sch_handle_ingress()) doesn't >>>> set a return value and it is left to the old return value of >>>> the caller (__netif_receive_skb_core()) which is RX drop, so if >>>> the packet is consumed, caller will stop and return this value >>>> as if the packet was dropped. >>>> >>>> This causes a problem in the kernel tcp stack when having a >>>> egress tc rule forwarding to a ingress tc rule. >>>> The tcp stack sending packets on the device having the egress rule >>>> will see the packets as not successfully transmitted (although they >>>> actually were), will not advance it's internal state of sent data, >>>> and packets returning on such tcp stream will be dropped by the tcp >>>> stack with reason ack-of-unsent-data. See reproduction in [0] below. >>>> >>> >>> Hm, but how is this return value propagated to egress? I checked >>> tcf_mirred_act() code, but don't see how it is even used there. >>> >>> 318 err = tcf_mirred_forward(want_ingress, skb2); >>> 319 if (err) { >>> 320 out: >>> 321 tcf_action_inc_overlimit_qstats(&m->common); >>> 322 if (tcf_mirred_is_act_redirect(m_eaction)) >>> 323 retval = TC_ACT_SHOT; >>> 324 } >>> 325 __this_cpu_dec(mirred_rec_level); >>> 326 >>> 327 return retval; >>> >>> >>> What am I missing? >> >> for the ingress acting act_mirred it will return TC_ACT_CONSUMED above >> the code you mentioned (since redirect=1, use_reinsert=1. Although >> TC_ACT_STOLEN which is the retval set for this action, will also act the >> same) >> >> >> It is propagated as such (TX stack starting from tcp): >> > > Sorry for my misunderstanding. > > I meant to say those TC_ACT_* return value, not NET_RX_*, but I worried > too much here, as mirred lets user specify the return value Yes TC_ACT_* start at the action mirred case, and end in tcf_handle_ingress/egresss switch cases, which then should be converted to NET_RX and NET_XMIT if done. > > BTW, it seems you at least miss the drop case, which is NET_RX_DROP for > TC_ACT_SHOT at least? Possibly other code paths in sch_handle_ingress() > too. I'll add the SHOT for v3 as the packet was handled in this case, but I should only update ret where the packet/skb was handled, which is where we also return NULL, as otherwise rx pipeline should continue and will update ret once handled (say in running the rx_handler). For example, if there are not tc filters (tcf_classify returns TC_ACT_UNSPEC) I should not update *ret, and it will continue to the rx handler, and if there isn't any, it would return the default ret RX_DROP value. > > Thanks.
diff --git a/net/core/dev.c b/net/core/dev.c index 56c8b0921c9f..c58ab657b164 100644 --- a/net/core/dev.c +++ b/net/core/dev.c @@ -5141,6 +5141,7 @@ sch_handle_ingress(struct sk_buff *skb, struct packet_type **pt_prev, int *ret, case TC_ACT_QUEUED: case TC_ACT_TRAP: consume_skb(skb); + *ret = NET_RX_SUCCESS; return NULL; case TC_ACT_REDIRECT: /* skb_mac_header check was done by cls/act_bpf, so @@ -5153,8 +5154,10 @@ sch_handle_ingress(struct sk_buff *skb, struct packet_type **pt_prev, int *ret, *another = true; break; } + *ret = NET_RX_SUCCESS; return NULL; case TC_ACT_CONSUMED: + *ret = NET_RX_SUCCESS; return NULL; default: break;
Currently qdisc ingress handling (sch_handle_ingress()) doesn't set a return value and it is left to the old return value of the caller (__netif_receive_skb_core()) which is RX drop, so if the packet is consumed, caller will stop and return this value as if the packet was dropped. This causes a problem in the kernel tcp stack when having a egress tc rule forwarding to a ingress tc rule. The tcp stack sending packets on the device having the egress rule will see the packets as not successfully transmitted (although they actually were), will not advance it's internal state of sent data, and packets returning on such tcp stream will be dropped by the tcp stack with reason ack-of-unsent-data. See reproduction in [0] below. Fix that by setting the return value to RX success if the packet was handled successfully. [0] Reproduction steps: $ ip link add veth1 type veth peer name peer1 $ ip link add veth2 type veth peer name peer2 $ ifconfig peer1 5.5.5.6/24 up $ ip netns add ns0 $ ip link set dev peer2 netns ns0 $ ip netns exec ns0 ifconfig peer2 5.5.5.5/24 up $ ifconfig veth2 0 up $ ifconfig veth1 0 up #ingress forwarding veth1 <-> veth2 $ tc qdisc add dev veth2 ingress $ tc qdisc add dev veth1 ingress $ tc filter add dev veth2 ingress prio 1 proto all flower \ action mirred egress redirect dev veth1 $ tc filter add dev veth1 ingress prio 1 proto all flower \ action mirred egress redirect dev veth2 #steal packet from peer1 egress to veth2 ingress, bypassing the veth pipe $ tc qdisc add dev peer1 clsact $ tc filter add dev peer1 egress prio 20 proto ip flower \ action mirred ingress redirect dev veth1 #run iperf and see connection not running $ iperf3 -s& $ ip netns exec ns0 iperf3 -c 5.5.5.6 -i 1 #delete egress rule, and run again, now should work $ tc filter del dev peer1 egress $ ip netns exec ns0 iperf3 -c 5.5.5.6 -i 1 Fixes: f697c3e8b35c ("[NET]: Avoid unnecessary cloning for ingress filtering") Signed-off-by: Paul Blakey <paulb@nvidia.com> --- net/core/dev.c | 3 +++ 1 file changed, 3 insertions(+)