TC: Introduce qevents
Subsystems (8)
cover TC: Introduce qevents
The Spectrum hardware allows execution of one of several actions as a result of queue management decisions: tail-dropping, early-dropping, marking a packet, or passing a configured latency threshold or buffer size. Such packets can be mirrored, trapped, or sampled.
Modeling the action to be taken as simply a TC action is very attractive, but it is not obvious where to put these actions. At least with ECN marking one could imagine a tree of qdiscs and classifiers that effectively accomplishes this task, albeit in an impractically complex manner. But there is just no way to match on dropped-ness of a packet, let alone dropped-ness due to a particular reason.
To allow configuring user-defined actions as a result of inner workings of a qdisc, this patch set introduces a concept of qevents. Those are attach points for TC blocks, where filters can be put that are executed as the packet hits well-defined points in the qdisc algorithms. The attached blocks can be shared, in a manner similar to clsact ingress and egress blocks, arbitrary classifiers with arbitrary actions can be put on them, etc.
For example:
# tc qdisc add dev eth0 root handle 1: \ red limit 500K avpkt 1K qevent early_drop block 10 # tc filter add block 10 \ matchall action mirred egress mirror dev eth1
The central patch #2 introduces several helpers to allow easy and uniform addition of qevents to qdiscs: initialization, destruction, qevent block number change validation, and qevent handling, i.e. dispatch of the filters attached to the block bound to a qevent.
Patch #1 adds root_lock argument to qdisc enqueue op. The problem this is tackling is that if a qevent filter pushes packets to the same qdisc tree that holds the qevent in the first place, attempt to take qdisc root lock for the second time will lead to a deadlock. To solve the issue, qevent handler needs to unlock and relock the root lock around the filter processing. Passing root_lock around makes it possible to get the lock where it is needed, and visibly so, such that it is obvious the lock will be used when invoking a qevent.
The following two patches, #3 and #4, then add two qevents to the RED
qdisc: "early_drop" qevent fires when a packet is early-dropped; "mark" qevent, when it is ECN-marked.
Patch #5 contains a selftest. I have mentioned this test when pushing the RED ECN nodrop mode and said that "I have no confidence in its portability to [...] different configurations". That still holds. The backlog and packet size are tuned to make the test deterministic. But it is better than nothing, and on the boxes that I ran it on it does work and shows that qevents work the way they are supposed to, and that their addition has not broken the other tested features.
This patch set does not deal with offloading. The idea there is that a driver will be able to figure out that a given block is used in qevent context by looking at binder type. A future patch-set will add a qdisc pointer to struct flow_block_offload, which a driver will be able to consult to glean the TC or other relevant attributes.
Changes from RFC to v1:
- Move a "q = qdisc_priv(sch)" from patch #3 to patch #4
- Fix deadlock caused by mirroring packet back to the same qdisc tree.
- Rename "tail" qevent to "tail_drop".
- Adapt to the new 100-column standard.
- Add a selftest
Petr Machata (5):
net: sched: Pass root lock to Qdisc_ops.enqueue
net: sched: Introduce helpers for qevent blocks
net: sched: sch_red: Split init and change callbacks
net: sched: sch_red: Add qevents "early_drop" and "mark"
selftests: forwarding: Add a RED test for SW datapath
include/net/flow_offload.h | 2 + include/net/pkt_cls.h | 49 ++ include/net/sch_generic.h | 6 +- include/uapi/linux/pkt_sched.h | 2 + net/core/dev.c | 4 +- net/sched/cls_api.c | 119 +++++ net/sched/sch_atm.c | 4 +- net/sched/sch_blackhole.c | 2 +- net/sched/sch_cake.c | 2 +- net/sched/sch_cbq.c | 4 +- net/sched/sch_cbs.c | 18 +- net/sched/sch_choke.c | 2 +- net/sched/sch_codel.c | 2 +- net/sched/sch_drr.c | 4 +- net/sched/sch_dsmark.c | 4 +- net/sched/sch_etf.c | 2 +- net/sched/sch_ets.c | 4 +- net/sched/sch_fifo.c | 6 +- net/sched/sch_fq.c | 2 +- net/sched/sch_fq_codel.c | 2 +- net/sched/sch_fq_pie.c | 2 +- net/sched/sch_generic.c | 4 +- net/sched/sch_gred.c | 2 +- net/sched/sch_hfsc.c | 6 +- net/sched/sch_hhf.c | 2 +- net/sched/sch_htb.c | 4 +- net/sched/sch_multiq.c | 4 +- net/sched/sch_netem.c | 8 +- net/sched/sch_pie.c | 2 +- net/sched/sch_plug.c | 2 +- net/sched/sch_prio.c | 6 +- net/sched/sch_qfq.c | 4 +- net/sched/sch_red.c | 102 +++- net/sched/sch_sfb.c | 4 +- net/sched/sch_sfq.c | 2 +- net/sched/sch_skbprio.c | 2 +- net/sched/sch_taprio.c | 4 +- net/sched/sch_tbf.c | 10 +- net/sched/sch_teql.c | 4 +- .../selftests/net/forwarding/sch_red.sh | 492 ++++++++++++++++++ 40 files changed, 822 insertions(+), 84 deletions(-) create mode 100755 tools/testing/selftests/net/forwarding/sch_red.sh
Signature · 1 line
2.20.1
PATCH 1/5 net: sched: Pass root lock to Qdisc_ops.enqueue
A following patch introduces qevents, points in qdisc algorithm where packet can be processed by user-defined filters. Should this processing lead to a situation where a new packet is to be enqueued on the same port, holding the root lock would lead to deadlocks. To solve the issue, qevent handler needs to unlock and relock the root lock when necessary.
To that end, add the root lock argument to the qdisc op enqueue, and propagate throughout.
--- include/net/sch_generic.h | 6 ++++-- net/core/dev.c | 4 ++-- net/sched/sch_atm.c | 4 ++-- net/sched/sch_blackhole.c | 2 +- net/sched/sch_cake.c | 2 +- net/sched/sch_cbq.c | 4 ++-- net/sched/sch_cbs.c | 18 +++++++++--------- net/sched/sch_choke.c | 2 +- net/sched/sch_codel.c | 2 +- net/sched/sch_drr.c | 4 ++-- net/sched/sch_dsmark.c | 4 ++-- net/sched/sch_etf.c | 2 +- net/sched/sch_ets.c | 4 ++-- net/sched/sch_fifo.c | 6 +++--- net/sched/sch_fq.c | 2 +- net/sched/sch_fq_codel.c | 2 +- net/sched/sch_fq_pie.c | 2 +- net/sched/sch_generic.c | 4 ++-- net/sched/sch_gred.c | 2 +- net/sched/sch_hfsc.c | 6 +++--- net/sched/sch_hhf.c | 2 +- net/sched/sch_htb.c | 4 ++-- net/sched/sch_multiq.c | 4 ++-- net/sched/sch_netem.c | 8 ++++---- net/sched/sch_pie.c | 2 +- net/sched/sch_plug.c | 2 +- net/sched/sch_prio.c | 6 +++--- net/sched/sch_qfq.c | 4 ++-- net/sched/sch_red.c | 4 ++-- net/sched/sch_sfb.c | 4 ++-- net/sched/sch_sfq.c | 2 +- net/sched/sch_skbprio.c | 2 +- net/sched/sch_taprio.c | 4 ++-- net/sched/sch_tbf.c | 10 +++++----- net/sched/sch_teql.c | 4 ++-- 35 files changed, 73 insertions(+), 71 deletions(-)
include/net/sch_generic.h +4 -2
@@ -57,6 +57,7 @@ struct qdisc_skb_head {struct Qdisc {int (*enqueue)(struct sk_buff *skb,struct Qdisc *sch, +spinlock_t *root_lock,struct sk_buff **to_free);struct sk_buff * (*dequeue)(struct Qdisc *sch);unsigned int flags;
@@ -241,6 +242,7 @@ struct Qdisc_ops {int (*enqueue)(struct sk_buff *skb,struct Qdisc *sch, +spinlock_t *root_lock,struct sk_buff **to_free);struct sk_buff * (*dequeue)(struct Qdisc *);struct sk_buff * (*peek)(struct Qdisc *);
@@ -788,11 +790,11 @@ static inline void qdisc_calculate_pkt_len(struct sk_buff *skb,#endif } -static inline int qdisc_enqueue (struct sk_buff *skb,struct Qdisc *sch, +static inline int qdisc_enqueue (struct sk_buff *skb,struct Qdisc *sch,spinlock_t *root_lock,struct sk_buff **to_free) {qdisc_calculate_pkt_len (skb, sch); -return sch->enqueue (skb, sch, to_free); +return sch->enqueue (skb, sch, root_lock, to_free); } static inline void _bstats_update(struct gnet_stats_basic_packed *bstats,
net/core/dev.c +2 -2
@@ -3743,7 +3743,7 @@ static inline int __dev_xmit_skb(struct sk_buff *skb, struct Qdisc *q,qdisc_calculate_pkt_len (skb, q);if (q->flags &TCQ_F_NOLOCK ) { - rc = q->enqueue (skb, q, &to_free) &NET_XMIT_MASK ; + rc = q->enqueue (skb, q,NULL , &to_free) &NET_XMIT_MASK ;qdisc_run (q);if (unlikely (to_free))
@@ -3786,7 +3786,7 @@ static inline int __dev_xmit_skb(struct sk_buff *skb, struct Qdisc *q,qdisc_run_end (q); rc =NET_XMIT_SUCCESS ; }else { - rc = q->enqueue (skb, q, &to_free) &NET_XMIT_MASK ; + rc = q->enqueue (skb, q, root_lock, &to_free) &NET_XMIT_MASK ;if (qdisc_run_begin (q)) {if (unlikely (contended)) {spin_unlock (&q->busylock);
net/sched/sch_atm.c +2 -2
@@ -374,7 +374,7 @@ static struct tcf_block *atm_tc_tcf_block(struct Qdisc *sch, unsigned long cl,/* --------------------------- Qdisc operations ---------------------------- */ -static int atm_tc_enqueue (struct sk_buff *skb,struct Qdisc *sch, +static int atm_tc_enqueue (struct sk_buff *skb,struct Qdisc *sch,spinlock_t *root_lock,struct sk_buff **to_free) {struct atm_qdisc_data *p =qdisc_priv (sch);
@@ -432,7 +432,7 @@ static int atm_tc_enqueue(struct sk_buff *skb, struct Qdisc *sch,#endif } - ret =qdisc_enqueue (skb, flow->q, to_free); + ret =qdisc_enqueue (skb, flow->q, root_lock, to_free);if (ret !=NET_XMIT_SUCCESS ) { drop:__maybe_unused if (net_xmit_drop_count (ret)) {
net/sched/sch_blackhole.c +1 -1
@@ -13,7 +13,7 @@#include <linux/skbuff.h> #include <net/pkt_sched.h> -static int blackhole_enqueue (struct sk_buff *skb,struct Qdisc *sch, +static int blackhole_enqueue (struct sk_buff *skb,struct Qdisc *sch,spinlock_t *root_lock,struct sk_buff **to_free) {qdisc_drop (skb, sch, to_free);
net/sched/sch_cake.c +1 -1
@@ -1663,7 +1663,7 @@ static u32 cake_classify(struct Qdisc *sch, struct cake_tin_data **t,static void cake_reconfigure (struct Qdisc *sch); -static s32 cake_enqueue (struct sk_buff *skb,struct Qdisc *sch, +static s32 cake_enqueue (struct sk_buff *skb,struct Qdisc *sch,spinlock_t *root_lock,struct sk_buff **to_free) {struct cake_sched_data *q =qdisc_priv (sch);
net/sched/sch_cbq.c +2 -2
@@ -356,7 +356,7 @@ cbq_mark_toplevel(struct cbq_sched_data *q, struct cbq_class *cl) }static int -cbq_enqueue (struct sk_buff *skb,struct Qdisc *sch, +cbq_enqueue (struct sk_buff *skb,struct Qdisc *sch,spinlock_t *root_lock,struct sk_buff **to_free) {struct cbq_sched_data *q =qdisc_priv (sch);
@@ -373,7 +373,7 @@ cbq_enqueue(struct sk_buff *skb, struct Qdisc *sch,return ret; } - ret =qdisc_enqueue (skb, cl->q, to_free); + ret =qdisc_enqueue (skb, cl->q, root_lock, to_free);if (ret ==NET_XMIT_SUCCESS ) { sch->q.qlen++;cbq_mark_toplevel (q, cl);
net/sched/sch_cbs.c +9 -9
@@ -77,7 +77,7 @@ struct cbs_sched_data {s64 sendslope;/* in bytes/s */ s64 idleslope;/* in bytes/s */ struct qdisc_watchdog watchdog; -int (*enqueue)(struct sk_buff *skb,struct Qdisc *sch, +int (*enqueue)(struct sk_buff *skb,struct Qdisc *sch,spinlock_t *root_lock,struct sk_buff **to_free);struct sk_buff *(*dequeue)(struct Qdisc *sch);struct Qdisc *qdisc;
@@ -85,13 +85,13 @@ struct cbs_sched_data { };static int cbs_child_enqueue (struct sk_buff *skb,struct Qdisc *sch, -struct Qdisc *child, +struct Qdisc *child,spinlock_t *root_lock,struct sk_buff **to_free) {unsigned int len =qdisc_pkt_len (skb);int err; - err = child->ops->enqueue (skb, child, to_free); + err = child->ops->enqueue (skb, child, root_lock, to_free);if (err !=NET_XMIT_SUCCESS )return err;
@@ -101,16 +101,16 @@ static int cbs_child_enqueue(struct sk_buff *skb, struct Qdisc *sch,return NET_XMIT_SUCCESS ; } -static int cbs_enqueue_offload (struct sk_buff *skb,struct Qdisc *sch, +static int cbs_enqueue_offload (struct sk_buff *skb,struct Qdisc *sch,spinlock_t *root_lock,struct sk_buff **to_free) {struct cbs_sched_data *q =qdisc_priv (sch);struct Qdisc *qdisc = q->qdisc; -return cbs_child_enqueue (skb, sch, qdisc, to_free); +return cbs_child_enqueue (skb, sch, qdisc, root_lock, to_free); } -static int cbs_enqueue_soft (struct sk_buff *skb,struct Qdisc *sch, +static int cbs_enqueue_soft (struct sk_buff *skb,struct Qdisc *sch,spinlock_t *root_lock,struct sk_buff **to_free) {struct cbs_sched_data *q =qdisc_priv (sch);
@@ -124,15 +124,15 @@ static int cbs_enqueue_soft(struct sk_buff *skb, struct Qdisc *sch, q->last =ktime_get_ns (); } -return cbs_child_enqueue (skb, sch, qdisc, to_free); +return cbs_child_enqueue (skb, sch, qdisc, root_lock, to_free); } -static int cbs_enqueue (struct sk_buff *skb,struct Qdisc *sch, +static int cbs_enqueue (struct sk_buff *skb,struct Qdisc *sch,spinlock_t *root_lock,struct sk_buff **to_free) {struct cbs_sched_data *q =qdisc_priv (sch); -return q->enqueue (skb, sch, to_free); +return q->enqueue (skb, sch, root_lock, to_free); }/* timediff is in ns, slope is in bytes/s */
net/sched/sch_choke.c +1 -1
@@ -210,7 +210,7 @@ static bool choke_match_random(const struct choke_sched_data *q,return choke_match_flow (oskb, nskb); } -static int choke_enqueue (struct sk_buff *skb,struct Qdisc *sch, +static int choke_enqueue (struct sk_buff *skb,struct Qdisc *sch,spinlock_t *root_lock,struct sk_buff **to_free) {struct choke_sched_data *q =qdisc_priv (sch);
net/sched/sch_codel.c +1 -1
@@ -108,7 +108,7 @@ static struct sk_buff *codel_qdisc_dequeue(struct Qdisc *sch)return skb; } -static int codel_qdisc_enqueue (struct sk_buff *skb,struct Qdisc *sch, +static int codel_qdisc_enqueue (struct sk_buff *skb,struct Qdisc *sch,spinlock_t *root_lock,struct sk_buff **to_free) {struct codel_sched_data *q;
net/sched/sch_drr.c +2 -2
@@ -337,7 +337,7 @@ static struct drr_class *drr_classify(struct sk_buff *skb, struct Qdisc *sch,return NULL ; } -static int drr_enqueue (struct sk_buff *skb,struct Qdisc *sch, +static int drr_enqueue (struct sk_buff *skb,struct Qdisc *sch,spinlock_t *root_lock,struct sk_buff **to_free) {unsigned int len =qdisc_pkt_len (skb);
@@ -355,7 +355,7 @@ static int drr_enqueue(struct sk_buff *skb, struct Qdisc *sch, } first = !cl->qdisc->q.qlen; - err =qdisc_enqueue (skb, cl->qdisc, to_free); + err =qdisc_enqueue (skb, cl->qdisc, root_lock, to_free);if (unlikely (err !=NET_XMIT_SUCCESS )) {if (net_xmit_drop_count (err)) { cl->qstats.drops++;
net/sched/sch_dsmark.c +2 -2
@@ -198,7 +198,7 @@ static struct tcf_block *dsmark_tcf_block(struct Qdisc *sch, unsigned long cl,/* --------------------------- Qdisc operations ---------------------------- */ -static int dsmark_enqueue (struct sk_buff *skb,struct Qdisc *sch, +static int dsmark_enqueue (struct sk_buff *skb,struct Qdisc *sch,spinlock_t *root_lock,struct sk_buff **to_free) {unsigned int len =qdisc_pkt_len (skb);
@@ -267,7 +267,7 @@ static int dsmark_enqueue(struct sk_buff *skb, struct Qdisc *sch, } } - err =qdisc_enqueue (skb, p->q, to_free); + err =qdisc_enqueue (skb, p->q, root_lock, to_free);if (err !=NET_XMIT_SUCCESS ) {if (net_xmit_drop_count (err))qdisc_qstats_drop (sch);
net/sched/sch_etf.c +1 -1
@@ -160,7 +160,7 @@ static void report_sock_error(struct sk_buff *skb, u32 err, u8 code) }static int etf_enqueue_timesortedlist (struct sk_buff *nskb,struct Qdisc *sch, -struct sk_buff **to_free) +spinlock_t *root_lock,struct sk_buff **to_free) {struct etf_sched_data *q =qdisc_priv (sch);struct rb_node **p = &q->head.rb_root.rb_node, *parent =NULL ;
net/sched/sch_ets.c +2 -2
@@ -415,7 +415,7 @@ static struct ets_class *ets_classify(struct sk_buff *skb, struct Qdisc *sch,return &q->classes[band]; } -static int ets_qdisc_enqueue (struct sk_buff *skb,struct Qdisc *sch, +static int ets_qdisc_enqueue (struct sk_buff *skb,struct Qdisc *sch,spinlock_t *root_lock,struct sk_buff **to_free) {unsigned int len =qdisc_pkt_len (skb);
@@ -433,7 +433,7 @@ static int ets_qdisc_enqueue(struct sk_buff *skb, struct Qdisc *sch, } first = !cl->qdisc->q.qlen; - err =qdisc_enqueue (skb, cl->qdisc, to_free); + err =qdisc_enqueue (skb, cl->qdisc, root_lock, to_free);if (unlikely (err !=NET_XMIT_SUCCESS )) {if (net_xmit_drop_count (err)) { cl->qstats.drops++;
net/sched/sch_fifo.c +3 -3
@@ -16,7 +16,7 @@/* 1 band FIFO pseudo-"scheduler" */ -static int bfifo_enqueue (struct sk_buff *skb,struct Qdisc *sch, +static int bfifo_enqueue (struct sk_buff *skb,struct Qdisc *sch,spinlock_t *root_lock,struct sk_buff **to_free) {if (likely (sch->qstats.backlog +qdisc_pkt_len (skb) <= sch->limit))
@@ -25,7 +25,7 @@ static int bfifo_enqueue(struct sk_buff *skb, struct Qdisc *sch,return qdisc_drop (skb, sch, to_free); } -static int pfifo_enqueue (struct sk_buff *skb,struct Qdisc *sch, +static int pfifo_enqueue (struct sk_buff *skb,struct Qdisc *sch,spinlock_t *root_lock,struct sk_buff **to_free) {if (likely (sch->q.qlen < sch->limit))
@@ -34,7 +34,7 @@ static int pfifo_enqueue(struct sk_buff *skb, struct Qdisc *sch,return qdisc_drop (skb, sch, to_free); } -static int pfifo_tail_enqueue (struct sk_buff *skb,struct Qdisc *sch, +static int pfifo_tail_enqueue (struct sk_buff *skb,struct Qdisc *sch,spinlock_t *root_lock,struct sk_buff **to_free) {unsigned int prev_backlog;
net/sched/sch_fq.c +1 -1
@@ -439,7 +439,7 @@ static bool fq_packet_beyond_horizon(const struct sk_buff *skb,return unlikely ((s64 )skb->tstamp > (s64)(q->ktime_cache + q->horizon)); } -static int fq_enqueue (struct sk_buff *skb,struct Qdisc *sch, +static int fq_enqueue (struct sk_buff *skb,struct Qdisc *sch,spinlock_t *root_lock,struct sk_buff **to_free) {struct fq_sched_data *q =qdisc_priv (sch);
net/sched/sch_fq_codel.c +1 -1
@@ -181,7 +181,7 @@ static unsigned int fq_codel_drop(struct Qdisc *sch, unsigned int max_packets,return idx; } -static int fq_codel_enqueue (struct sk_buff *skb,struct Qdisc *sch, +static int fq_codel_enqueue (struct sk_buff *skb,struct Qdisc *sch,spinlock_t *root_lock,struct sk_buff **to_free) {struct fq_codel_sched_data *q =qdisc_priv (sch);
net/sched/sch_fq_pie.c +1 -1
@@ -125,7 +125,7 @@ static inline void flow_queue_add(struct fq_pie_flow *flow, skb->next =NULL ; } -static int fq_pie_qdisc_enqueue (struct sk_buff *skb,struct Qdisc *sch, +static int fq_pie_qdisc_enqueue (struct sk_buff *skb,struct Qdisc *sch,spinlock_t *root_lock,struct sk_buff **to_free) {struct fq_pie_sched_data *q =qdisc_priv (sch);
net/sched/sch_generic.c +2 -2
@@ -520,7 +520,7 @@ EXPORT_SYMBOL(netif_carrier_off);cheaper. */ -static int noop_enqueue (struct sk_buff *skb,struct Qdisc *qdisc, +static int noop_enqueue (struct sk_buff *skb,struct Qdisc *qdisc,spinlock_t *root_lock,struct sk_buff **to_free) {__qdisc_drop (skb, to_free);
@@ -614,7 +614,7 @@ static inline struct skb_array *band2list(struct pfifo_fast_priv *priv,return &priv->q[band]; } -static int pfifo_fast_enqueue (struct sk_buff *skb,struct Qdisc *qdisc, +static int pfifo_fast_enqueue (struct sk_buff *skb,struct Qdisc *qdisc,spinlock_t *root_lock,struct sk_buff **to_free) {int band = prio2band[skb->priority &TC_PRIO_MAX ];
net/sched/sch_gred.c +1 -1
@@ -161,7 +161,7 @@ static bool gred_per_vq_red_flags_used(struct gred_sched *table)return false; } -static int gred_enqueue (struct sk_buff *skb,struct Qdisc *sch, +static int gred_enqueue (struct sk_buff *skb,struct Qdisc *sch,spinlock_t *root_lock,struct sk_buff **to_free) {struct gred_sched_data *q =NULL ;
net/sched/sch_hfsc.c +3 -3
@@ -1528,8 +1528,8 @@ hfsc_dump_qdisc(struct Qdisc *sch, struct sk_buff *skb)return -1 ; } -static int -hfsc_enqueue (struct sk_buff *skb,struct Qdisc *sch,struct sk_buff **to_free) +static int hfsc_enqueue (struct sk_buff *skb,struct Qdisc *sch,spinlock_t *root_lock, +struct sk_buff **to_free) {unsigned int len =qdisc_pkt_len (skb);struct hfsc_class *cl;
@@ -1545,7 +1545,7 @@ hfsc_enqueue(struct sk_buff *skb, struct Qdisc *sch, struct sk_buff **to_free) } first = !cl->qdisc->q.qlen; - err =qdisc_enqueue (skb, cl->qdisc, to_free); + err =qdisc_enqueue (skb, cl->qdisc, root_lock, to_free);if (unlikely (err !=NET_XMIT_SUCCESS )) {if (net_xmit_drop_count (err)) { cl->qstats.drops++;
net/sched/sch_hhf.c +1 -1
@@ -368,7 +368,7 @@ static unsigned int hhf_drop(struct Qdisc *sch, struct sk_buff **to_free)return bucket - q->buckets; } -static int hhf_enqueue (struct sk_buff *skb,struct Qdisc *sch, +static int hhf_enqueue (struct sk_buff *skb,struct Qdisc *sch,spinlock_t *root_lock,struct sk_buff **to_free) {struct hhf_sched_data *q =qdisc_priv (sch);
net/sched/sch_htb.c +2 -2
@@ -576,7 +576,7 @@ static inline void htb_deactivate(struct htb_sched *q, struct htb_class *cl) cl->prio_activity =0 ; } -static int htb_enqueue (struct sk_buff *skb,struct Qdisc *sch, +static int htb_enqueue (struct sk_buff *skb,struct Qdisc *sch,spinlock_t *root_lock,struct sk_buff **to_free) {int uninitialized_var (ret );
@@ -599,7 +599,7 @@ static int htb_enqueue(struct sk_buff *skb, struct Qdisc *sch,__qdisc_drop (skb, to_free);return ret;#endif - }else if ((ret =qdisc_enqueue (skb, cl->leaf.q, + }else if ((ret =qdisc_enqueue (skb, cl->leaf.q, root_lock, to_free)) !=NET_XMIT_SUCCESS ) {if (net_xmit_drop_count (ret)) {qdisc_qstats_drop (sch);
net/sched/sch_multiq.c +2 -2
@@ -57,7 +57,7 @@ multiq_classify(struct sk_buff *skb, struct Qdisc *sch, int *qerr) }static int -multiq_enqueue (struct sk_buff *skb,struct Qdisc *sch, +multiq_enqueue (struct sk_buff *skb,struct Qdisc *sch,spinlock_t *root_lock,struct sk_buff **to_free) {struct Qdisc *qdisc;
@@ -74,7 +74,7 @@ multiq_enqueue(struct sk_buff *skb, struct Qdisc *sch, }#endif - ret =qdisc_enqueue (skb, qdisc, to_free); + ret =qdisc_enqueue (skb, qdisc, root_lock, to_free);if (ret ==NET_XMIT_SUCCESS ) { sch->q.qlen++;return NET_XMIT_SUCCESS ;
net/sched/sch_netem.c +4 -4
@@ -431,7 +431,7 @@ static struct sk_buff *netem_segment(struct sk_buff *skb, struct Qdisc *sch,* NET_XMIT_DROP: queue length didn't change. * NET_XMIT_SUCCESS: one skb was queued. */ -static int netem_enqueue (struct sk_buff *skb,struct Qdisc *sch, +static int netem_enqueue (struct sk_buff *skb,struct Qdisc *sch,spinlock_t *root_lock,struct sk_buff **to_free) {struct netem_sched_data *q =qdisc_priv (sch);
@@ -480,7 +480,7 @@ static int netem_enqueue(struct sk_buff *skb, struct Qdisc *sch,u32 dupsave = q->duplicate;/* prevent duplicating a dup... */ q->duplicate =0 ; - rootq->enqueue (skb2, rootq, to_free); + rootq->enqueue (skb2, rootq, root_lock, to_free); q->duplicate = dupsave; rc_drop =NET_XMIT_SUCCESS ; }
@@ -604,7 +604,7 @@ static int netem_enqueue(struct sk_buff *skb, struct Qdisc *sch,skb_mark_not_on_list (segs);qdisc_skb_cb (segs)->pkt_len = segs->len; last_len = segs->len; - rc =qdisc_enqueue (segs, sch, to_free); + rc =qdisc_enqueue (segs, sch, root_lock, to_free);if (rc !=NET_XMIT_SUCCESS ) {if (net_xmit_drop_count (rc))qdisc_qstats_drop (sch);
@@ -720,7 +720,7 @@ static struct sk_buff *netem_dequeue(struct Qdisc *sch)struct sk_buff *to_free =NULL ;int err; - err =qdisc_enqueue (skb, q->qdisc, &to_free); + err =qdisc_enqueue (skb, q->qdisc,NULL , &to_free);kfree_skb_list (to_free);if (err !=NET_XMIT_SUCCESS &&net_xmit_drop_count (err)) {
net/sched/sch_pie.c +1 -1
@@ -82,7 +82,7 @@ bool pie_drop_early(struct Qdisc *sch, struct pie_params *params, }EXPORT_SYMBOL_GPL (pie_drop_early); -static int pie_qdisc_enqueue (struct sk_buff *skb,struct Qdisc *sch, +static int pie_qdisc_enqueue (struct sk_buff *skb,struct Qdisc *sch,spinlock_t *root_lock,struct sk_buff **to_free) {struct pie_sched_data *q =qdisc_priv (sch);
net/sched/sch_plug.c +1 -1
@@ -84,7 +84,7 @@ struct plug_sched_data {u32 pkts_to_release; }; -static int plug_enqueue (struct sk_buff *skb,struct Qdisc *sch, +static int plug_enqueue (struct sk_buff *skb,struct Qdisc *sch,spinlock_t *root_lock,struct sk_buff **to_free) {struct plug_sched_data *q =qdisc_priv (sch);
net/sched/sch_prio.c +3 -3
@@ -65,8 +65,8 @@ prio_classify(struct sk_buff *skb, struct Qdisc *sch, int *qerr)return q->queues[band]; } -static int -prio_enqueue (struct sk_buff *skb,struct Qdisc *sch,struct sk_buff **to_free) +static int prio_enqueue (struct sk_buff *skb,struct Qdisc *sch,spinlock_t *root_lock, +struct sk_buff **to_free) {unsigned int len =qdisc_pkt_len (skb);struct Qdisc *qdisc;
@@ -83,7 +83,7 @@ prio_enqueue(struct sk_buff *skb, struct Qdisc *sch, struct sk_buff **to_free) }#endif - ret =qdisc_enqueue (skb, qdisc, to_free); + ret =qdisc_enqueue (skb, qdisc, root_lock, to_free);if (ret ==NET_XMIT_SUCCESS ) { sch->qstats.backlog += len; sch->q.qlen++;
net/sched/sch_qfq.c +2 -2
@@ -1194,7 +1194,7 @@ static struct qfq_aggregate *qfq_choose_next_agg(struct qfq_sched *q)return agg; } -static int qfq_enqueue (struct sk_buff *skb,struct Qdisc *sch, +static int qfq_enqueue (struct sk_buff *skb,struct Qdisc *sch,spinlock_t *root_lock,struct sk_buff **to_free) {unsigned int len =qdisc_pkt_len (skb), gso_segs;
@@ -1225,7 +1225,7 @@ static int qfq_enqueue(struct sk_buff *skb, struct Qdisc *sch, gso_segs =skb_is_gso (skb) ?skb_shinfo (skb)->gso_segs :1 ; first = !cl->qdisc->q.qlen; - err =qdisc_enqueue (skb, cl->qdisc, to_free); + err =qdisc_enqueue (skb, cl->qdisc, root_lock, to_free);if (unlikely (err !=NET_XMIT_SUCCESS )) {pr_debug ("qfq_enqueue: enqueue failed %d\n" , err);if (net_xmit_drop_count (err)) {
net/sched/sch_red.c +2 -2
@@ -65,7 +65,7 @@ static int red_use_nodrop(struct red_sched_data *q)return q->flags &TC_RED_NODROP ; } -static int red_enqueue (struct sk_buff *skb,struct Qdisc *sch, +static int red_enqueue (struct sk_buff *skb,struct Qdisc *sch,spinlock_t *root_lock,struct sk_buff **to_free) {struct red_sched_data *q =qdisc_priv (sch);
@@ -118,7 +118,7 @@ static int red_enqueue(struct sk_buff *skb, struct Qdisc *sch,break ; } - ret =qdisc_enqueue (skb, child, to_free); + ret =qdisc_enqueue (skb, child, root_lock, to_free);if (likely (ret ==NET_XMIT_SUCCESS )) {qdisc_qstats_backlog_inc (sch, skb); sch->q.qlen++;
net/sched/sch_sfb.c +2 -2
@@ -276,7 +276,7 @@ static bool sfb_classify(struct sk_buff *skb, struct tcf_proto *fl,return false; } -static int sfb_enqueue (struct sk_buff *skb,struct Qdisc *sch, +static int sfb_enqueue (struct sk_buff *skb,struct Qdisc *sch,spinlock_t *root_lock,struct sk_buff **to_free) {
@@ -399,7 +399,7 @@ static int sfb_enqueue(struct sk_buff *skb, struct Qdisc *sch, } enqueue: - ret =qdisc_enqueue (skb, child, to_free); + ret =qdisc_enqueue (skb, child, root_lock, to_free);if (likely (ret ==NET_XMIT_SUCCESS )) {qdisc_qstats_backlog_inc (sch, skb); sch->q.qlen++;
net/sched/sch_sfq.c +1 -1
@@ -343,7 +343,7 @@ static int sfq_headdrop(const struct sfq_sched_data *q) }static int -sfq_enqueue (struct sk_buff *skb,struct Qdisc *sch,struct sk_buff **to_free) +sfq_enqueue (struct sk_buff *skb,struct Qdisc *sch,spinlock_t *root_lock,struct sk_buff **to_free) {struct sfq_sched_data *q =qdisc_priv (sch);unsigned int hash, dropped;
net/sched/sch_skbprio.c +1 -1
@@ -65,7 +65,7 @@ static u16 calc_new_low_prio(const struct skbprio_sched_data *q)return SKBPRIO_MAX_PRIORITY -1 ; } -static int skbprio_enqueue (struct sk_buff *skb,struct Qdisc *sch, +static int skbprio_enqueue (struct sk_buff *skb,struct Qdisc *sch,spinlock_t *root_lock,struct sk_buff **to_free) {const unsigned int max_priority =SKBPRIO_MAX_PRIORITY -1 ;
net/sched/sch_taprio.c +2 -2
@@ -410,7 +410,7 @@ static long get_packet_txtime(struct sk_buff *skb, struct Qdisc *sch)return txtime; } -static int taprio_enqueue (struct sk_buff *skb,struct Qdisc *sch, +static int taprio_enqueue (struct sk_buff *skb,struct Qdisc *sch,spinlock_t *root_lock,struct sk_buff **to_free) {struct taprio_sched *q =qdisc_priv (sch);
@@ -435,7 +435,7 @@ static int taprio_enqueue(struct sk_buff *skb, struct Qdisc *sch,qdisc_qstats_backlog_inc (sch, skb); sch->q.qlen++; -return qdisc_enqueue (skb, child, to_free); +return qdisc_enqueue (skb, child, root_lock, to_free); }static struct sk_buff *taprio_peek_soft (struct Qdisc *sch)
net/sched/sch_tbf.c +5 -5
@@ -187,7 +187,7 @@ static int tbf_offload_dump(struct Qdisc *sch)/* GSO packet is too big, segment it so that tbf can transmit * each segment in time */ -static int tbf_segment (struct sk_buff *skb,struct Qdisc *sch, +static int tbf_segment (struct sk_buff *skb,struct Qdisc *sch,spinlock_t *root_lock,struct sk_buff **to_free) {struct tbf_sched_data *q =qdisc_priv (sch);
@@ -206,7 +206,7 @@ static int tbf_segment(struct sk_buff *skb, struct Qdisc *sch,skb_mark_not_on_list (segs);qdisc_skb_cb (segs)->pkt_len = segs->len; len += segs->len; - ret =qdisc_enqueue (segs, q->qdisc, to_free); + ret =qdisc_enqueue (segs, q->qdisc, root_lock, to_free);if (ret !=NET_XMIT_SUCCESS ) {if (net_xmit_drop_count (ret))qdisc_qstats_drop (sch);
@@ -221,7 +221,7 @@ static int tbf_segment(struct sk_buff *skb, struct Qdisc *sch,return nb >0 ?NET_XMIT_SUCCESS :NET_XMIT_DROP ; } -static int tbf_enqueue (struct sk_buff *skb,struct Qdisc *sch, +static int tbf_enqueue (struct sk_buff *skb,struct Qdisc *sch,spinlock_t *root_lock,struct sk_buff **to_free) {struct tbf_sched_data *q =qdisc_priv (sch);
@@ -231,10 +231,10 @@ static int tbf_enqueue(struct sk_buff *skb, struct Qdisc *sch,if (qdisc_pkt_len (skb) > q->max_size) {if (skb_is_gso (skb) &&skb_gso_validate_mac_len (skb, q->max_size)) -return tbf_segment (skb, sch, to_free); +return tbf_segment (skb, sch, root_lock, to_free);return qdisc_drop (skb, sch, to_free); } - ret =qdisc_enqueue (skb, q->qdisc, to_free); + ret =qdisc_enqueue (skb, q->qdisc, root_lock, to_free);if (ret !=NET_XMIT_SUCCESS ) {if (net_xmit_drop_count (ret))qdisc_qstats_drop (sch);
net/sched/sch_teql.c +2 -2
@@ -72,8 +72,8 @@ struct teql_sched_data {/* "teql*" qdisc routines */ -static int -teql_enqueue (struct sk_buff *skb,struct Qdisc *sch,struct sk_buff **to_free) +static int teql_enqueue (struct sk_buff *skb,struct Qdisc *sch,spinlock_t *root_lock, +struct sk_buff **to_free) {struct net_device *dev =qdisc_dev (sch);struct teql_sched_data *q =qdisc_priv (sch);
Signature · 1 line
2.20.1
Re: [PATCH net-next v1 1/5] net: sched: Pass root lock to Qdisc_ops.enqueue
Quoted 9 lines · Petr Machata
A following patch introduces qevents, points in qdisc algorithm where packet can be processed by user-defined filters. Should this processing lead to a situation where a new packet is to be enqueued on the same port, holding the root lock would lead to deadlocks. To solve the issue, qevent handler needs to unlock and relock the root lock when necessary.
To that end, add the root lock argument to the qdisc op enqueue, and propagate throughout.
Hmm, but why do you pass root lock down to each ->enqueue()?
You can find root lock with sch_tree_lock() (or qdisc_lock() if you don't care about hierarchy), and you already have qdisc as a parameter of tcf_qevent_handle().
Thanks.
Re: [PATCH net-next v1 1/5] net: sched: Pass root lock to Qdisc_ops.enqueue
Quoted 8 lines · Cong Wang
Quoted 7 lines · Petr Machata
A following patch introduces qevents, points in qdisc algorithm where packet can be processed by user-defined filters. Should this processing lead to a situation where a new packet is to be enqueued on the same port, holding the root lock would lead to deadlocks. To solve the issue, qevent handler needs to unlock and relock the root lock when necessary.
Quoted 8 lines
Quoted 2 lines
To that end, add the root lock argument to the qdisc op enqueue, and propagate throughout.
Hmm, but why do you pass root lock down to each ->enqueue()?
You can find root lock with
sch_tree_lock()(orqdisc_lock()if you don't care about hierarchy), and you already have qdisc as a parameter oftcf_qevent_handle().
I know, I wanted to make it clear that the lock may end up being used, instead of doing it "stealthily". If you find this inelegant I can push a follow-up that converts tcf_qevent_handle() to sch_tree_unlock().
Re: [PATCH net-next v1 1/5] net: sched: Pass root lock to Qdisc_ops.enqueue
Quoted 17 lines · Petr Machata
Quoted 13 lines · Cong Wang
Quoted 9 lines · Petr Machata
A following patch introduces qevents, points in qdisc algorithm where packet can be processed by user-defined filters. Should this processing lead to a situation where a new packet is to be enqueued on the same port, holding the root lock would lead to deadlocks. To solve the issue, qevent handler needs to unlock and relock the root lock when necessary.
To that end, add the root lock argument to the qdisc op enqueue, and propagate throughout.
Hmm, but why do you pass root lock down to each ->enqueue()?
Quoted 7 lines
Quoted 3 lines
You can find root lock with
sch_tree_lock()(orqdisc_lock()if you don't care about hierarchy), and you already have qdisc as a parameter oftcf_qevent_handle().I know, I wanted to make it clear that the lock may end up being used, instead of doing it "stealthily". If you find this inelegant I can push a follow-up that converts
tcf_qevent_handle()tosch_tree_unlock().
So far only sch_red uses tcf_qevent_handle(), changing the rest qdisc's just for sch_red seems overkill here.
Of course, if we could eliminate the root lock release, all the pains would go away.
Thanks.
PATCH 1/4 uapi: pkt_sched: Add two new RED attributes
In reply to Petr Machata--- include/uapi/linux/pkt_sched.h | 2 ++ 1 file changed, 2 insertions(+)
include/uapi/linux/pkt_sched.h +2 -0
@@ -257,6 +257,8 @@ enum {TCA_RED_STAB ,TCA_RED_MAX_P ,TCA_RED_FLAGS ,/* bitfield32 */ +TCA_RED_EARLY_DROP_BLOCK ,/* u32 */ +TCA_RED_MARK_BLOCK ,/* u32 */ __TCA_RED_MAX, };
Signature · 1 line
2.20.1
PATCH 2/4 tc: Add helpers to support qevent handling
Introduce a set of helpers to make it easy to add support for qevents into qdisc.
The idea behind this is that qevent types will be generally reused between qdiscs, rather than each having a completely idiosyncratic set of qevents. The qevent module holds functions for parsing, dumping and formatting of these common qevent types, and for dispatch to the appropriate set of handlers based on the qevent name.
--- tc/Makefile | 1 + tc/tc_qevent.c | 202 +++++++++++++++++++++++++++++++++++++++++++++++++ tc/tc_qevent.h | 49 ++++++++++++ 3 files changed, 252 insertions(+) create mode 100644 tc/tc_qevent.c create mode 100644 tc/tc_qevent.h
tc/Makefile +1 -0
@@ -122,6 +122,7 @@ TCLIB += tc_red.o TCLIB += tc_cbq.o TCLIB += tc_estimator.o TCLIB += tc_stab.o +TCLIB += tc_qevent.o CFLAGS += -DCONFIG_GACT -DCONFIG_GACT_PROB ifneq ($(IPT_LIB_DIR),)
tc/tc_qevent.c new file +202 -0
202 lines not on this page · raw message
tc/tc_qevent.h new file +49 -0
49 lines not on this page · raw message
Signature · 1 line
2.20.1
PATCH 3/4 man: tc: Describe qevents
In reply to Petr MachataAdd some general remarks about qevents.
--- man/man8/tc.8 | 19 +++++++++++++++++++ 1 file changed, 19 insertions(+)
man/man8/tc.8 +19 -0
@@ -254,6 +254,25 @@ Traffic control filter that matches every packet. See .BR tc-matchall (8) for details. +.SH QEVENTS +Qdiscs may invoke user-configured actions when certain interesting events +take place in the qdisc. Each qevent can either be unused, or can have a +block attached to it. To this block are then attached filters using the "tc +block BLOCK_IDX" syntax. The block is executed when the qevent associated +with the attachment point takes place. For example, packet could be +dropped, or delayed, etc., depending on the qdisc and the qevent in +question. + +For example: +.PP +.RS +tc qdisc add dev eth0 root handle 1: red limit 500K avpkt 1K \\ + qevent early block 10 +.RE +.RS +tc filter add block 10 matchall action mirred egress mirror dev eth1 +.RE + .SH CLASSLESS QDISCS The classless qdiscs are: .TP
Signature · 1 line
2.20.1
PATCH 4/4 tc: q_red: Add support for qevents "mark" and "early_drop"
In reply to Petr MachataThe "early_drop" qevent matches packets that have been early-dropped. The "mark" qevent matches packets that have been ECN-marked.
--- man/man8/tc-red.8 | 18 +++++++++++++++++- tc/q_red.c | 30 +++++++++++++++++++++++++++--- 2 files changed, 44 insertions(+), 4 deletions(-)
man/man8/tc-red.8 +17 -1
@@ -17,7 +17,11 @@ packets rate .B ] [ probability chance -.B ] [ adaptive ] +.B ] [ adaptive ] [ qevent early_drop block +index +.B ] [ qevent mark block +index +.B ] .SH DESCRIPTION Random Early Detection is a classless qdisc which manages its queue size
@@ -134,6 +138,18 @@ Goal of Adaptive RED is to make 'probability' dynamic value between 1% and 50% t .B (max - min) / 2 .fi +.SH QEVENTS +See tc (8) for some general notes about qevents. The RED qdisc supports the +following qevents: + +.TP +early_drop +The associated block is executed when packets are early-dropped. This includes +non-ECT packets in ECN mode. +.TP +mark +The associated block is executed when packets are marked in ECN mode. + .SH EXAMPLE .P
tc/q_red.c +27 -3
@@ -22,6 +22,7 @@#include "utils.h" #include "tc_util.h" +#include "tc_qevent.h" #include "tc_red.h"
@@ -30,11 +31,20 @@ static void explain(void)fprintf (stderr,"Usage: ... red limit BYTES [min BYTES] [max BYTES] avpkt BYTES [burst PACKETS]\n" " [adaptive] [probability PROBABILITY] [bandwidth KBPS]\n" -" [ecn] [harddrop] [nodrop]\n" ); +" [ecn] [harddrop] [nodrop]\n" +" [qevent early_drop block IDX] [qevent mark block IDX]\n" ); }#define RED_SUPPORTED_FLAGS (TC_RED_HISTORIC_FLAGS | TC_RED_NODROP) +static struct qevent_plain qe_early_drop = {}; +static struct qevent_plain qe_mark = {}; +static struct qevent_util qevents[] = { +QEVENT ("early_drop" , plain, &qe_early_drop,TCA_RED_EARLY_DROP_BLOCK ), +QEVENT ("mark" , plain, &qe_mark,TCA_RED_MARK_BLOCK ), + {}, +}; +static int red_parse_opt (struct qdisc_util *qu,int argc,char **argv,struct nlmsghdr *n,const char *dev) {
@@ -51,6 +61,8 @@ static int red_parse_opt(struct qdisc_util *qu, int argc, char **argv,__u32 max_P;struct rtattr *tail; +qevents_init (qevents); +while (argc >0 ) {if (strcmp (*argv,"limit" ) ==0 ) {NEXT_ARG ();
@@ -109,6 +121,11 @@ static int red_parse_opt(struct qdisc_util *qu, int argc, char **argv, flags_bf.value |=TC_RED_ADAPTATIVE ; }else if (strcmp (*argv,"adaptive" ) ==0 ) { flags_bf.value |=TC_RED_ADAPTATIVE ; + }else if (matches (*argv,"qevent" ) ==0 ) { +NEXT_ARG (); +if (qevent_parse (qevents, &argc, &argv)) +return -1 ; +continue ; }else if (strcmp (*argv,"help" ) ==0 ) {explain ();return -1 ;
@@ -162,6 +179,8 @@ static int red_parse_opt(struct qdisc_util *qu, int argc, char **argv, max_P = probability *pow (2 ,32 );addattr_l (n,1024 ,TCA_RED_MAX_P , &max_P,sizeof (max_P));addattr_l (n,1024 ,TCA_RED_FLAGS , &flags_bf,sizeof (flags_bf)); +if (qevents_dump (qevents, n)) +return -1 ;addattr_nest_end (n, tail);return 0 ; }
@@ -203,12 +222,12 @@ static int red_print_opt(struct qdisc_util *qu, FILE *f, struct rtattr *opt)print_uint (PRINT_JSON ,"min" ,NULL , qopt->qth_min);print_string (PRINT_FP ,NULL ,"min %s " ,sprint_size (qopt->qth_min, b2));print_uint (PRINT_JSON ,"max" ,NULL , qopt->qth_max); -print_string (PRINT_FP ,NULL ,"max %s " ,sprint_size (qopt->qth_max, b3)); +print_string (PRINT_FP ,NULL ,"max %s" ,sprint_size (qopt->qth_max, b3));tc_red_print_flags (qopt->flags);if (show_details) { -print_uint (PRINT_ANY ,"ewma" ,"ewma %u " , qopt->Wlog); +print_uint (PRINT_ANY ,"ewma" ," ewma %u " , qopt->Wlog);if (max_P)print_float (PRINT_ANY ,"probability" ,"probability %lg " , max_P /pow (2 ,32 ));
@@ -217,6 +236,11 @@ static int red_print_opt(struct qdisc_util *qu, FILE *f, struct rtattr *opt)print_uint (PRINT_ANY ,"Scell_log" ,"Scell_log %u" , qopt->Scell_log); } + +qevents_init (qevents); +if (qevents_read (qevents, tb)) +return -1 ; +qevents_print (qevents, f);return 0 ; }
Signature · 1 line
2.20.1
PATCH 2/5 net: sched: Introduce helpers for qevent blocks
In reply to Petr MachataQevents are attach points for TC blocks, where filters can be put that are executed when "interesting events" take place in a qdisc. The data to keep and the functions to invoke to maintain a qevent will be largely the same between qevents. Therefore introduce sched-wide helpers for qevent management.
Currently, similarly to ingress and egress blocks of clsact pseudo-qdisc, blocks attachment cannot be changed after the qdisc is created. To that end, add a helper tcf_qevent_validate_change(), which verifies whether block index attribute is not attached, or if it is, whether its value matches the current one (i.e. there is no material change).
The function tcf_qevent_handle() should be invoked when qdisc hits the "interesting event" corresponding to a block. This function releases root lock for the duration of executing the attached filters, to allow packets generated through user actions (notably mirred) to be reinserted to the same qdisc tree.
--- include/net/pkt_cls.h | 49 +++++++++++++++++ net/sched/cls_api.c | 119 ++++++++++++++++++++++++++++++++++++++++++ 2 files changed, 168 insertions(+)
include/net/pkt_cls.h +49 -0
@@ -32,6 +32,12 @@ struct tcf_block_ext_info {u32 block_index; }; +struct tcf_qevent { +struct tcf_block *block; +struct tcf_block_ext_info info; +struct tcf_proto __rcu *filter_chain; +}; +struct tcf_block_cb ;bool tcf_queue_work (struct rcu_work *rwork,work_func_t func);
@@ -552,6 +558,49 @@ int tc_setup_cb_reoffload(struct tcf_block *block, struct tcf_proto *tp, void *cb_priv, u32 *flags, unsigned int *in_hw_count);unsigned int tcf_exts_num_actions (struct tcf_exts *exts); +#ifdef CONFIG_NET_CLS_ACT +int tcf_qevent_init (struct tcf_qevent *qe,struct Qdisc *sch, +enum flow_block_binder_type binder_type, +struct nlattr *block_index_attr, +struct netlink_ext_ack *extack); +void tcf_qevent_destroy (struct tcf_qevent *qe,struct Qdisc *sch); +int tcf_qevent_validate_change (struct tcf_qevent *qe,struct nlattr *block_index_attr, +struct netlink_ext_ack *extack); +struct sk_buff *tcf_qevent_handle (struct tcf_qevent *qe,struct Qdisc *sch,struct sk_buff *skb, +spinlock_t *root_lock,struct sk_buff **to_free,int *ret); +int tcf_qevent_dump (struct sk_buff *skb,int attr_name,struct tcf_qevent *qe); +#else +static inline int tcf_qevent_init (struct tcf_qevent *qe,struct Qdisc *sch, +enum flow_block_binder_type binder_type, +struct nlattr *block_index_attr, +struct netlink_ext_ack *extack) +{ +return 0 ; +} + +static inline void tcf_qevent_destroy (struct tcf_qevent *qe,struct Qdisc *sch) +{ +} + +static inline int tcf_qevent_validate_change (struct tcf_qevent *qe,struct nlattr *block_index_attr, +struct netlink_ext_ack *extack) +{ +return 0 ; +} + +static inline struct sk_buff * +tcf_qevent_handle (struct tcf_qevent *qe,struct Qdisc *sch,struct sk_buff *skb, +spinlock_t *root_lock,struct sk_buff **to_free,int *ret) +{ +return skb; +} + +static inline int tcf_qevent_dump (struct sk_buff *skb,int attr_name,struct tcf_qevent *qe) +{ +return 0 ; +} +#endif +struct tc_cls_u32_knode {struct tcf_exts *exts;struct tcf_result *res;
net/sched/cls_api.c +119 -0
125 lines not on this page · raw message
Signature · 1 line
2.20.1
Re: [PATCH net-next v1 2/5] net: sched: Introduce helpers for qevent blocks
Quoted 5 lines · Petr Machata
The function
tcf_qevent_handle()should be invoked when qdisc hits the "interesting event" corresponding to a block. This function releases root lock for the duration of executing the attached filters, to allow packets generated through user actions (notably mirred) to be reinserted to the same qdisc tree.
Are you sure releasing the root lock in the middle of an enqueue operation is a good idea? I mean, it seems racy with qdisc change or reset path, for example, __red_change() could update some RED parameters immediately after you release the root lock.
Thanks.
Re: [PATCH net-next v1 2/5] net: sched: Introduce helpers for qevent blocks
Quoted 11 lines · Cong Wang
Quoted 5 lines · Petr Machata
The function
tcf_qevent_handle()should be invoked when qdisc hits the "interesting event" corresponding to a block. This function releases root lock for the duration of executing the attached filters, to allow packets generated through user actions (notably mirred) to be reinserted to the same qdisc tree.Are you sure releasing the root lock in the middle of an enqueue operation is a good idea? I mean, it seems racy with qdisc change or reset path, for example,
__red_change()could update some RED parameters immediately after you release the root lock.
So I had mulled over this for a while. If packets are enqueued or dequeued while the lock is released, maybe the packet under consideration should have been hard_marked instead of prob_marked, or vice versa. (I can't really go to not marked at all, because the fact that we ran the qevent is a very observable commitment to put the packet in the queue with marking.) I figured that is not such a big deal.
Regarding a configuration change, for a brief period after the change, a few not-yet-pushed packets could have been enqueued with ECN marking even as I e.g. disabled ECN. This does not seem like a big deal either, these are transient effects.
Re: [PATCH net-next v1 2/5] net: sched: Introduce helpers for qevent blocks
Quoted 23 lines · Petr Machata
Quoted 11 lines · Cong Wang
Quoted 5 lines · Petr Machata
The function
tcf_qevent_handle()should be invoked when qdisc hits the "interesting event" corresponding to a block. This function releases root lock for the duration of executing the attached filters, to allow packets generated through user actions (notably mirred) to be reinserted to the same qdisc tree.Are you sure releasing the root lock in the middle of an enqueue operation is a good idea? I mean, it seems racy with qdisc change or reset path, for example,
__red_change()could update some RED parameters immediately after you release the root lock.So I had mulled over this for a while. If packets are enqueued or dequeued while the lock is released, maybe the packet under consideration should have been
hard_markedinstead ofprob_marked, or vice versa. (I can't really go to not marked at all, because the fact that we ran the qevent is a very observable commitment to put the packet in the queue with marking.) I figured that is not such a big deal.
Quoted 4 lines
Regarding a configuration change, for a brief period after the change, a few not-yet-pushed packets could have been enqueued with ECN marking even as I e.g. disabled ECN. This does not seem like a big deal either, these are transient effects.
Hmm, let's see:
1. red_enqueue() caches a pointer to child qdisc, child = q->qdisc
2. root lock is released by tcf_qevent_handle().
3. __red_change() acquires the root lock and then changes child qdisc with a new one
4. The old child qdisc is put by qdisc_put()
5. tcf_qevent_handle() acquires the root lock again, and still uses the cached but now freed old child qdisc.
Isn't this a problem?
Even if it really is not, why do we make tcf_qevent_handle() callers' life so hard? They have to be very careful with race conditions like this.
Thanks.
Re: [PATCH net-next v1 2/5] net: sched: Introduce helpers for qevent blocks
Quoted 39 lines · Cong Wang
Quoted 23 lines · Petr Machata
Quoted 11 lines · Cong Wang
Quoted 5 lines · Petr Machata
The function
tcf_qevent_handle()should be invoked when qdisc hits the "interesting event" corresponding to a block. This function releases root lock for the duration of executing the attached filters, to allow packets generated through user actions (notably mirred) to be reinserted to the same qdisc tree.Are you sure releasing the root lock in the middle of an enqueue operation is a good idea? I mean, it seems racy with qdisc change or reset path, for example,
__red_change()could update some RED parameters immediately after you release the root lock.So I had mulled over this for a while. If packets are enqueued or dequeued while the lock is released, maybe the packet under consideration should have been
hard_markedinstead ofprob_marked, or vice versa. (I can't really go to not marked at all, because the fact that we ran the qevent is a very observable commitment to put the packet in the queue with marking.) I figured that is not such a big deal.Quoted 4 lines
Regarding a configuration change, for a brief period after the change, a few not-yet-pushed packets could have been enqueued with ECN marking even as I e.g. disabled ECN. This does not seem like a big deal either, these are transient effects.
Hmm, let's see:
1.
red_enqueue()caches a pointer to child qdisc, child =q->qdisc
2. root lock is released bytcf_qevent_handle().
3.__red_change()acquires the root lock and then changes child qdisc with a new one
4. The old child qdisc is put byqdisc_put()
5.tcf_qevent_handle()acquires the root lock again, and still uses the cached but now freed old child qdisc.
Quoted 1 line
Isn't this a problem?
I missed this. It is a problem, destroy gets called right away and then enqueue would use invalid data.
Also qdisc_graft() calls cops->graft, which locks the tree and swaps the qdisc pointes, and then qdisc_put()s the original one. So dropping the root lock can lead to destruction of the qdisc whose enqueue is currently executed.
So that's no good. The lock has to be held throughout.
Quoted 2 lines
Even if it really is not, why do we make
tcf_qevent_handle()callers' life so hard? They have to be very careful with race conditions like this.
Do you have another solution in mind here? I think the deadlock (in both classification and qevents) is an issue, but really don't know how to avoid it except by dropping the lock.
Re: [PATCH net-next v1 2/5] net: sched: Introduce helpers for qevent blocks
Quoted 51 lines · Petr Machata
Quoted 39 lines · Cong Wang
Quoted 23 lines · Petr Machata
Quoted 11 lines · Cong Wang
Quoted 5 lines · Petr Machata
The function
tcf_qevent_handle()should be invoked when qdisc hits the "interesting event" corresponding to a block. This function releases root lock for the duration of executing the attached filters, to allow packets generated through user actions (notably mirred) to be reinserted to the same qdisc tree.Are you sure releasing the root lock in the middle of an enqueue operation is a good idea? I mean, it seems racy with qdisc change or reset path, for example,
__red_change()could update some RED parameters immediately after you release the root lock.So I had mulled over this for a while. If packets are enqueued or dequeued while the lock is released, maybe the packet under consideration should have been
hard_markedinstead ofprob_marked, or vice versa. (I can't really go to not marked at all, because the fact that we ran the qevent is a very observable commitment to put the packet in the queue with marking.) I figured that is not such a big deal.Quoted 4 lines
Regarding a configuration change, for a brief period after the change, a few not-yet-pushed packets could have been enqueued with ECN marking even as I e.g. disabled ECN. This does not seem like a big deal either, these are transient effects.
Hmm, let's see:
1.
red_enqueue()caches a pointer to child qdisc, child =q->qdisc
2. root lock is released bytcf_qevent_handle().
3.__red_change()acquires the root lock and then changes child qdisc with a new one
4. The old child qdisc is put byqdisc_put()
5.tcf_qevent_handle()acquires the root lock again, and still uses the cached but now freed old child qdisc.Quoted 1 line
Isn't this a problem?
I missed this. It is a problem, destroy gets called right away and then enqueue would use invalid data.
Also
qdisc_graft()callscops->graft, which locks the tree and swaps the qdisc pointes, and then qdisc_put()s the original one. So dropping the root lock can lead to destruction of the qdisc whose enqueue is currently executed.
Quoted 8 lines
So that's no good. The lock has to be held throughout.
Quoted 2 lines
Even if it really is not, why do we make
tcf_qevent_handle()callers' life so hard? They have to be very careful with race conditions like this.Do you have another solution in mind here? I think the deadlock (in both classification and qevents) is an issue, but really don't know how to avoid it except by dropping the lock.
Actually I guess I could qdisc_refcount_inc() the current qdisc so that it doesn't go away. Then when enqueing I could access the child directly, not relying on the now-obsolete cache from the beginning of the enqueue function. I suppose that a similar approach could be used in other users of tcf_classify() as well. What do you think?
Re: [PATCH net-next v1 2/5] net: sched: Introduce helpers for qevent blocks
Quoted 6 lines · Petr Machata
Actually I guess I could
qdisc_refcount_inc()the current qdisc so that it doesn't go away. Then when enqueing I could access the child directly, not relying on the now-obsolete cache from the beginning of the enqueue function. I suppose that a similar approach could be used in other users oftcf_classify()as well. What do you think?
The above example is just a quick one I can think of, there could be more race conditions that lead to other kinds of bugs.
I am sure you can fix that one, but the point is that it is hard to audit and fix them all. The best solution here is of course not to release that lock, but again it requires some more work to avoid the deadlock.
Thanks.
Re: [PATCH net-next v1 2/5] net: sched: Introduce helpers for qevent blocks
In reply to Petr MachataQuoted 3 lines · Petr Machata
Do you have another solution in mind here? I think the deadlock (in both classification and qevents) is an issue, but really don't know how to avoid it except by dropping the lock.
Ideally we should only take the lock once, but it clearly requires some work to teach the dev_queue_xmit() in act_mirred not to acquire it again.
Thanks.
Re: [PATCH net-next v1 2/5] net: sched: Introduce helpers for qevent blocks
Quoted 7 lines · Cong Wang
Quoted 3 lines · Petr Machata
Do you have another solution in mind here? I think the deadlock (in both classification and qevents) is an issue, but really don't know how to avoid it except by dropping the lock.
Ideally we should only take the lock once, but it clearly requires some work to teach the
dev_queue_xmit()inact_mirrednot to acquire it again.
act_mirred does not acquire it though. The egress path does. And the packet can traverse several mirred instances on several devices before it gets to the deadlock. Currently I don't see a way to give the egress path a way to know that the lock is already taken.
I'll think about it some more. For now I will at least fix the lack of locking.
Re: [PATCH net-next v1 2/5] net: sched: Introduce helpers for qevent blocks
Quoted 4 lines · Petr Machata
Cong Wang <[email protected]> writes:
I'll think about it some more. For now I will at least fix the lack of locking.
I guess I could store smp_processor_id() that acquired the lock in struct qdisc_skb_head. Do a trylock instead of lock, and on fail check the stored value. I'll need to be careful about the race between unsuccessful trylock and the test, and about making sure CPU ID doesn't change after it is read. I'll probe this tomorrow.
Re: [PATCH net-next v1 2/5] net: sched: Introduce helpers for qevent blocks
Quoted 6 lines · Petr Machata
Quoted 2 lines · Petr Machata
Cong Wang <[email protected]> writes:
Quoted 8 lines
Quoted 2 lines
I'll think about it some more. For now I will at least fix the lack of locking.
I guess I could store
smp_processor_id()that acquired the lock in structqdisc_skb_head. Do a trylock instead of lock, and on fail check the stored value. I'll need to be careful about the race between unsuccessful trylock and the test, and about making sure CPU ID doesn't change after it is read. I'll probe this tomorrow.
Like __netif_tx_lock(), right? Seems doable.
Thanks.
Re: [PATCH net-next v1 2/5] net: sched: Introduce helpers for qevent blocks
Quoted 10 lines · Cong Wang
Quoted 9 lines · Petr Machata
Quoted 4 lines · Petr Machata
Cong Wang <[email protected]> writes:
I'll think about it some more. For now I will at least fix the lack of locking.
Quoted 7 lines
Quoted 5 lines
I guess I could store
smp_processor_id()that acquired the lock in structqdisc_skb_head. Do a trylock instead of lock, and on fail check the stored value. I'll need to be careful about the race between unsuccessful trylock and the test, and about making sure CPU ID doesn't change after it is read. I'll probe this tomorrow.Like
__netif_tx_lock(), right? Seems doable.
Good to see it actually used, I wasn't sure if the idea made sense :)
Unfortunately it is not enough.
Consider two threads (A, B) and two netdevices (eth0, eth1):
- "A" takes eth0's root lock and proceeds to classification
- "B" takes eth1's root lock and proceeds to classification
- "A" invokes mirror to eth1, waits on lock held by "B"
- "B" invakes mirror to eth0, waits on lock held by "A"
- Some say they are still waiting to this day.
So one option that I see is to just stash the mirrored packet in a queue instead of delivering it right away:
- s/netif_receive_skb/netif_rx/ in act_mirred
- Reuse the RX queue for TX packets as well, differentiating the two by a bit in SKB CB. Then process_backlog() would call either __netif_receive_skb() or dev_queue_transmit().
- Drop mirred_rec_level guard.
This seems to work, but I might be missing something non-obvious, such as CB actually being used for something already in that context. I would really rather not introduce a second backlog queue just for mirred though.
Since mirred_rec_level does not kick in anymore, the same packet can end up being forwarded from the backlog queue, to the qdisc, and back to the backlog queue, forever. But that seems OK, that's what the admin configured, so that's what's happening.
If this is not a good idea for some reason, this might work as well:
- Convert the current root lock to an rw lock. Convert all current lockers to write lock (which should be safe), except of enqueue, which will take read lock. That will allow many concurrent threads to enter enqueue, or one thread several times, but it will exclude all other users.
So this guards configuration access to the qdisc tree, makes sure qdiscs don't go away from under one's feet.
- Introduce another spin lock to guard the private data of the qdisc tree, counters etc., things that even two concurrent enqueue operations shouldn't trample on. Enqueue takes this spin lock after read-locking the root lock. act_mirred drops it before injecting the packet and takes it again afterwards.
Any opinions y'all?
Re: [PATCH net-next v1 2/5] net: sched: Introduce helpers for qevent blocks
Quoted 26 lines · Petr Machata
Quoted 10 lines · Cong Wang
Quoted 9 lines · Petr Machata
Quoted 4 lines · Petr Machata
Cong Wang <[email protected]> writes:
I'll think about it some more. For now I will at least fix the lack of locking.
Quoted 7 lines
Quoted 5 lines
I guess I could store
smp_processor_id()that acquired the lock in structqdisc_skb_head. Do a trylock instead of lock, and on fail check the stored value. I'll need to be careful about the race between unsuccessful trylock and the test, and about making sure CPU ID doesn't change after it is read. I'll probe this tomorrow.Like
__netif_tx_lock(), right? Seems doable.Good to see it actually used, I wasn't sure if the idea made sense :)
Unfortunately it is not enough.
Quoted 7 lines
Consider two threads (A, B) and two netdevices (eth0, eth1):
- "A" takes eth0's root lock and proceeds to classification
- "B" takes eth1's root lock and proceeds to classification
- "A" invokes mirror to eth1, waits on lock held by "B"
- "B" invakes mirror to eth0, waits on lock held by "A"
- Some say they are still waiting to this day.
Sure, AA or ABBA deadlock.
Quoted 11 lines
So one option that I see is to just stash the mirrored packet in a queue instead of delivering it right away:
- s/netif_receive_skb/netif_rx/ in
act_mirred- Reuse the RX queue for TX packets as well, differentiating the two by a bit in SKB CB. Then
process_backlog()would call either__netif_receive_skb()ordev_queue_transmit().- Drop
mirred_rec_levelguard.
I don't think I follow you, the root qdisc lock is on egress which has nothing to do with ingress, so I don't see how netif_rx() is even involved.
Quoted 11 lines
This seems to work, but I might be missing something non-obvious, such as CB actually being used for something already in that context. I would really rather not introduce a second backlog queue just for mirred though.
Since
mirred_rec_leveldoes not kick in anymore, the same packet can end up being forwarded from the backlog queue, to the qdisc, and back to the backlog queue, forever. But that seems OK, that's what the admin configured, so that's what's happening.
Quoted 7 lines
If this is not a good idea for some reason, this might work as well:
- Convert the current root lock to an rw lock. Convert all current lockers to write lock (which should be safe), except of enqueue, which will take read lock. That will allow many concurrent threads to enter enqueue, or one thread several times, but it will exclude all other users.
Are you sure we can parallelize enqueue()? They all need to move skb into some queue, which is not able to parallelize with just a read lock. Even the "lockless" qdisc takes a spinlock, r->producer_lock, for enqueue().
Quoted 11 lines
So this guards configuration access to the qdisc tree, makes sure qdiscs don't go away from under one's feet.
- Introduce another spin lock to guard the private data of the qdisc tree, counters etc., things that even two concurrent enqueue operations shouldn't trample on. Enqueue takes this spin lock after read-locking the root lock.
act_mirreddrops it before injecting the packet and takes it again afterwards.Any opinions y'all?
I thought about forbidding mirror/redirecting to the same device, but there might be some legitimate use cases of such. So, I don't have any other ideas yet, perhaps there is some way to refactor dev_queue_xmit() to avoid this deadlock.
Thanks.
Re: [PATCH net-next v1 2/5] net: sched: Introduce helpers for qevent blocks
Quoted 43 lines · Cong Wang
Quoted 26 lines · Petr Machata
Quoted 10 lines · Cong Wang
Quoted 9 lines · Petr Machata
Quoted 4 lines · Petr Machata
Cong Wang <[email protected]> writes:
I'll think about it some more. For now I will at least fix the lack of locking.
Quoted 7 lines
Quoted 5 lines
I guess I could store
smp_processor_id()that acquired the lock in structqdisc_skb_head. Do a trylock instead of lock, and on fail check the stored value. I'll need to be careful about the race between unsuccessful trylock and the test, and about making sure CPU ID doesn't change after it is read. I'll probe this tomorrow.Like
__netif_tx_lock(), right? Seems doable.Good to see it actually used, I wasn't sure if the idea made sense :)
Unfortunately it is not enough.
Quoted 7 lines
Consider two threads (A, B) and two netdevices (eth0, eth1):
- "A" takes eth0's root lock and proceeds to classification
- "B" takes eth1's root lock and proceeds to classification
- "A" invokes mirror to eth1, waits on lock held by "B"
- "B" invakes mirror to eth0, waits on lock held by "A"
- Some say they are still waiting to this day.Sure, AA or ABBA deadlock.
Quoted 6 lines
So one option that I see is to just stash the mirrored packet in a queue instead of delivering it right away:
- s/netif_receive_skb/netif_rx/ in
act_mirred
Quoted 8 lines
Quoted 5 lines
- Reuse the RX queue for TX packets as well, differentiating the two by a bit in SKB CB. Then
process_backlog()would call either__netif_receive_skb()ordev_queue_transmit().- Drop
mirred_rec_levelguard.I don't think I follow you, the root qdisc lock is on egress which has nothing to do with ingress, so I don't see how
netif_rx()is even involved.
netif_rx() isn't, but __netif_receive_skb() is, and that can lead to the deadlock as well when another mirred redirects it back to the locked egress queue.
So a way to solve "mirred ingress dev" action deadlock is to s/netif_receive_skb/netif_rx/. I.e. don't resolve the mirror right away, go through the per-CPU queue.
Then "mirred egress dev" could be fixed similarly by repurposing the queue for both ingress and egress, differentiating ingress packets from egress ones by a bit in SKB CB.
Quoted 19 lines
Quoted 11 lines
This seems to work, but I might be missing something non-obvious, such as CB actually being used for something already in that context. I would really rather not introduce a second backlog queue just for mirred though.
Since
mirred_rec_leveldoes not kick in anymore, the same packet can end up being forwarded from the backlog queue, to the qdisc, and back to the backlog queue, forever. But that seems OK, that's what the admin configured, so that's what's happening.Quoted 7 lines
If this is not a good idea for some reason, this might work as well:
- Convert the current root lock to an rw lock. Convert all current lockers to write lock (which should be safe), except of enqueue, which will take read lock. That will allow many concurrent threads to enter enqueue, or one thread several times, but it will exclude all other users.
Quoted 4 lines
Are you sure we can parallelize
enqueue()? They all need to move skb into some queue, which is not able to parallelize with just a read lock. Even the "lockless" qdisc takes a spinlock,r->producer_lock, forenqueue().
That's why the second spin lock is for. In guards private data, including the queues.
Quoted 10 lines
Quoted 10 lines
So this guards configuration access to the qdisc tree, makes sure qdiscs don't go away from under one's feet.
- Introduce another spin lock to guard the private data of the qdisc tree, counters etc., things that even two concurrent enqueue operations shouldn't trample on. Enqueue takes this spin lock after read-locking the root lock.
act_mirreddrops it before injecting the packet and takes it again afterwards.
Quoted 4 lines
Quoted 1 line
Any opinions y'all?
I thought about forbidding mirror/redirecting to the same device, but there might be some legitimate use cases of such. So, I don't
Yes, and also that's not enough:
- A chain of mirreds can achieve the deadlocks as well (i.e. mirror to eth1, redirect back to eth0). Or the ABA case shown above, where it's two actions that don't even work with the same packets causing the deadlock.
- I suspect general forwarding could cause this deadlock as well. E.g. redirecting to ingress of a device, where bridge, router take over and bring the packet back to egress. I have not tried reproducing this though, maybe there's a queue or delayed work etc. somewhere in there that makes this not an issue.
Quoted 2 lines
have any other ideas yet, perhaps there is some way to refactor
dev_queue_xmit()to avoid this deadlock.
Re: [PATCH net-next v1 2/5] net: sched: Introduce helpers for qevent blocks
In reply to Petr MachataQuoted 5 lines · Petr Machata
The function
tcf_qevent_handle()should be invoked when qdisc hits the "interesting event" corresponding to a block. This function releases root lock for the duration of executing the attached filters, to allow packets generated through user actions (notably mirred) to be reinserted to the same qdisc tree.
I read this again, another question here is: why is tcf_qevent_handle() special here? We call tcf_classify() under root lock in many other places too, for example htb_enqueue(), which of course includes act_mirred execution, so why isn't it a problem there?
People added MIRRED_RECURSION_LIMIT for this kinda recursion, but never released that root lock.
Thanks.
Re: [PATCH net-next v1 2/5] net: sched: Introduce helpers for qevent blocks
Quoted 7 lines · Cong Wang
Quoted 5 lines · Petr Machata
The function
tcf_qevent_handle()should be invoked when qdisc hits the "interesting event" corresponding to a block. This function releases root lock for the duration of executing the attached filters, to allow packets generated through user actions (notably mirred) to be reinserted to the same qdisc tree.
Quoted 7 lines
I read this again, another question here is: why is
tcf_qevent_handle()special here? We calltcf_classify()under root lock in many other places too, for examplehtb_enqueue(), which of course includesact_mirredexecution, so why isn't it a problem there?People added
MIRRED_RECURSION_LIMITfor this kinda recursion, but never released that root lock.
Yes, I realized later that the qdiscs that use tcf_classify() for classification have this exact problem as well. My intention was to fix it by dropping the lock. Since the classification is the first step of enqueing it should not really lead to races, so hopefully this will be OK. I don't have any code as of yet.
The recursion limit makes sense for clsact, which is handled out of the root lock.
PATCH 3/5 net: sched: sch_red: Split init and change callbacks
In reply to Petr MachataIn the following patches, RED will get two qevents. The implementation will be clearer if the callback for change is not a pure subset of the callback for init. Split the two and promote attribute parsing to the callbacks themselves from the common code, because it will be handy there.
--- net/sched/sch_red.c | 42 ++++++++++++++++++++++++++++++------------ 1 file changed, 30 insertions(+), 12 deletions(-)
net/sched/sch_red.c +30 -12
@@ -215,12 +215,11 @@ static const struct nla_policy red_policy[TCA_RED_MAX + 1] = { [TCA_RED_FLAGS] = NLA_POLICY_BITFIELD32(TC_RED_SUPPORTED_FLAGS), }; -static int red_change (struct Qdisc *sch,struct nlattr *opt, -struct netlink_ext_ack *extack) +static int __red_change (struct Qdisc *sch,struct nlattr **tb, +struct netlink_ext_ack *extack) {struct Qdisc *old_child =NULL , *child =NULL ;struct red_sched_data *q =qdisc_priv (sch); -struct nlattr *tb[TCA_RED_MAX +1 ];struct nla_bitfield32 flags_bf;struct tc_red_qopt *ctl;unsigned char userbits;
@@ -228,14 +227,6 @@ static int red_change(struct Qdisc *sch, struct nlattr *opt,int err;u32 max_P; -if (opt ==NULL ) -return -EINVAL ; - - err =nla_parse_nested_deprecated (tb,TCA_RED_MAX , opt, red_policy, -NULL ); -if (err <0 ) -return err; -if (tb[TCA_RED_PARMS ] ==NULL || tb[TCA_RED_STAB ] ==NULL )return -EINVAL ;
@@ -323,11 +314,38 @@ static int red_init(struct Qdisc *sch, struct nlattr *opt, struct netlink_ext_ack *extack) {struct red_sched_data *q =qdisc_priv (sch); +struct nlattr *tb[TCA_RED_MAX +1 ]; +int err; + +if (!opt) +return -EINVAL ; + + err =nla_parse_nested_deprecated (tb,TCA_RED_MAX , opt, red_policy, + extack); +if (err <0 ) +return err; q->qdisc = &noop_qdisc; q->sch = sch;timer_setup (&q->adapt_timer, red_adaptative_timer,0 ); -return red_change (sch, opt, extack); +return __red_change (sch, tb, extack); +} + +static int red_change (struct Qdisc *sch,struct nlattr *opt, +struct netlink_ext_ack *extack) +{ +struct nlattr *tb[TCA_RED_MAX +1 ]; +int err; + +if (!opt) +return -EINVAL ; + + err =nla_parse_nested_deprecated (tb,TCA_RED_MAX , opt, red_policy, + extack); +if (err <0 ) +return err; + +return __red_change (sch, tb, extack); }static int red_dump_offload_stats (struct Qdisc *sch)
Signature · 1 line
2.20.1
PATCH 4/5 net: sched: sch_red: Add qevents "early_drop" and "mark"
In reply to Petr MachataIn order to allow acting on dropped and/or ECN-marked packets, add two new qevents to the RED qdisc: "early_drop" and "mark". Filters attached at "early_drop" block are executed as packets are early-dropped, those attached at the "mark" block are executed as packets are ECN-marked.
Two new attributes are introduced: TCA_RED_EARLY_DROP_BLOCK with the block index for the "early_drop" qevent, and TCA_RED_MARK_BLOCK for the "mark" qevent. Absence of these attributes signifies "don't care": no block is allocated in that case, or the existing blocks are left intact in case of the change callback.
For purposes of offloading, blocks attached to these qevents appear with newly-introduced binder types, FLOW_BLOCK_BINDER_TYPE_RED_EARLY_DROP and FLOW_BLOCK_BINDER_TYPE_RED_MARK.
--- include/net/flow_offload.h | 2 ++ include/uapi/linux/pkt_sched.h | 2 ++ net/sched/sch_red.c | 58 ++++++++++++++++++++++++++++++++-- 3 files changed, 60 insertions(+), 2 deletions(-)
include/net/flow_offload.h +2 -0
@@ -419,6 +419,8 @@ enum flow_block_binder_type {FLOW_BLOCK_BINDER_TYPE_UNSPEC ,FLOW_BLOCK_BINDER_TYPE_CLSACT_INGRESS ,FLOW_BLOCK_BINDER_TYPE_CLSACT_EGRESS , +FLOW_BLOCK_BINDER_TYPE_RED_EARLY_DROP , + FLOW_BLOCK_BINDER_TYPE_RED_MARK, }; struct flow_block {
include/uapi/linux/pkt_sched.h +2 -0
@@ -257,6 +257,8 @@ enum {TCA_RED_STAB ,TCA_RED_MAX_P ,TCA_RED_FLAGS ,/* bitfield32 */ +TCA_RED_EARLY_DROP_BLOCK ,/* u32 */ +TCA_RED_MARK_BLOCK ,/* u32 */ __TCA_RED_MAX, };
net/sched/sch_red.c +56 -2
117 lines not on this page · raw message
Signature · 1 line
2.20.1
PATCH 5/5 selftests: forwarding: Add a RED test for SW datapath
In reply to Petr MachataThis test is inspired by the mlxsw RED selftest. It is much simpler to set up (also because there is no point in testing PRIO / RED encapsulation). It tests bare RED, ECN and ECN+nodrop modes of operation. On top of that it tests RED early_drop and mark qevents.
--- .../selftests/net/forwarding/sch_red.sh | 492 ++++++++++++++++++ 1 file changed, 492 insertions(+) create mode 100755 tools/testing/selftests/net/forwarding/sch_red.sh
tools/testing/selftests/net/forwarding/sch_red.sh new file +492 -0
@@ -0,0 +1,492 @@ +# SPDX-License-Identifier: GPL-2.0 + +# This test sends one stream of traffic from H1 through a TBF shaper, to a RED +# within TBF shaper on $swp3. The two shapers have the same configuration, and +# thus the resulting stream should fill all available bandwidth on the latter +# shaper. A second stream is sent from H2 also via $swp3, and used to inject +# additional traffic. Since all available bandwidth is taken, this traffic has +# to go to backlog. +# +# +--------------------------+ +--------------------------+ +# | H1 | | H2 | +# | + $h1 | | + $h2 | +# | | 192.0.2.1/28 | | | 192.0.2.2/28 | +# | | TBF 10Mbps | | | | +# +-----|--------------------+ +-----|--------------------+ +# | | +# +-----|------------------------------------------------|--------------------+ +# | SW | | | +# | +--|------------------------------------------------|----------------+ | +# | | + $swp1 + $swp2 | | +# | | BR | | +# | | | | +# | | + $swp3 | | +# | | | TBF 10Mbps / RED | | +# | +--------------------------------|-----------------------------------+ | +# | | | +# +-----------------------------------|---------------------------------------+ +# | +# +-----|--------------------+ +# | H3 | | +# | + $h1 | +# | 192.0.2.3/28 | +# | | +# +--------------------------+ + +ALL_TESTS=" + ping_ipv4 + ecn_test + ecn_nodrop_test + red_test + red_qevent_test + ecn_qevent_test +" + +NUM_NETIFS=6 +CHECK_TC="yes" +source lib.sh + +BACKLOG=30000 +PKTSZ=1400 + +h1_create() +{ + simple_if_init $h1 192.0.2.1/28 + mtu_set $h1 10000 + tc qdisc replace dev $h1 root handle 1: tbf \ + rate 10Mbit burst 10K limit 1M +} + +h1_destroy() +{ + tc qdisc del dev $h1 root + mtu_restore $h1 + simple_if_fini $h1 192.0.2.1/28 +} + +h2_create() +{ + simple_if_init $h2 192.0.2.2/28 + mtu_set $h2 10000 +} + +h2_destroy() +{ + mtu_restore $h2 + simple_if_fini $h2 192.0.2.2/28 +} + +h3_create() +{ + simple_if_init $h3 192.0.2.3/28 + mtu_set $h3 10000 +} + +h3_destroy() +{ + mtu_restore $h3 + simple_if_fini $h3 192.0.2.3/28 +} + +switch_create() +{ + ip link add dev br up type bridge + ip link set dev $swp1 up master br + ip link set dev $swp2 up master br + ip link set dev $swp3 up master br + + mtu_set $swp1 10000 + mtu_set $swp2 10000 + mtu_set $swp3 10000 + + tc qdisc replace dev $swp3 root handle 1: tbf \ + rate 10Mbit burst 10K limit 1M + ip link add name _drop_test up type dummy +} + +switch_destroy() +{ + ip link del dev _drop_test + tc qdisc del dev $swp3 root + + mtu_restore $h3 + mtu_restore $h2 + mtu_restore $h1 + + ip link set dev $swp3 down nomaster + ip link set dev $swp2 down nomaster + ip link set dev $swp1 down nomaster + ip link del dev br +} + +setup_prepare() +{ + h1=${NETIFS[p1]} + swp1=${NETIFS[p2]} + + h2=${NETIFS[p3]} + swp2=${NETIFS[p4]} + + swp3=${NETIFS[p5]} + h3=${NETIFS[p6]} + + h3_mac=$(mac_get $h3) + + vrf_prepare + + h1_create + h2_create + h3_create + switch_create +} + +cleanup() +{ + pre_cleanup + + switch_destroy + h3_destroy + h2_destroy + h1_destroy + + vrf_cleanup +} + +ping_ipv4() +{ + ping_test $h1 192.0.2.3 " from host 1" + ping_test $h2 192.0.2.3 " from host 2" +} + +get_qdisc_backlog() +{ + qdisc_stats_get $swp3 11: .backlog +} + +get_nmarked() +{ + qdisc_stats_get $swp3 11: .marked +} + +get_qdisc_npackets() +{ + qdisc_stats_get $swp3 11: .packets +} + +get_nmirrored() +{ + link_stats_get _drop_test tx packets +} + +send_packets() +{ + local proto=$1; shift + local pkts=$1; shift + + $MZ $h2 -p $PKTSZ -a own -b $h3_mac -A 192.0.2.2 -B 192.0.2.3 -t $proto -q -c $pkts "$@" +} + +# This sends traffic in an attempt to build a backlog of $size. Returns 0 on +# success. After 10 failed attempts it bails out and returns 1. It dumps the +# backlog size to stdout. +build_backlog() +{ + local size=$1; shift + local proto=$1; shift + + local i=0 + + while :; do + local cur=$(get_qdisc_backlog) + local diff=$((size - cur)) + local pkts=$(((diff + PKTSZ - 1) / PKTSZ)) + + if ((cur >= size)); then + echo $cur + return 0 + elif ((i++ > 10)); then + echo $cur + return 1 + fi + + send_packets $proto $pkts "$@" + sleep 1 + done +} + +check_marking() +{ + local cond=$1; shift + + local npackets_0=$(get_qdisc_npackets) + local nmarked_0=$(get_nmarked) + sleep 5 + local npackets_1=$(get_qdisc_npackets) + local nmarked_1=$(get_nmarked) + + local nmarked_d=$((nmarked_1 - nmarked_0)) + local npackets_d=$((npackets_1 - npackets_0)) + local pct=$((100 * nmarked_d / npackets_d)) + + echo $pct + ((pct $cond)) +} + +check_mirroring() +{ + local cond=$1; shift + + local npackets_0=$(get_qdisc_npackets) + local nmirrored_0=$(get_nmirrored) + sleep 5 + local npackets_1=$(get_qdisc_npackets) + local nmirrored_1=$(get_nmirrored) + + local nmirrored_d=$((nmirrored_1 - nmirrored_0)) + local npackets_d=$((npackets_1 - npackets_0)) + local pct=$((100 * nmirrored_d / npackets_d)) + + echo $pct + ((pct $cond)) +} + +ecn_test_common() +{ + local name=$1; shift + local limit=$1; shift + local backlog + local pct + + # Build the below-the-limit backlog using UDP. We could use TCP just + # fine, but this way we get a proof that UDP is accepted when queue + # length is below the limit. The main stream is using TCP, and if the + # limit is misconfigured, we would see this traffic being ECN marked. + RET=0 + backlog=$(build_backlog $((2 * limit / 3)) udp) + check_err $? "Could not build the requested backlog" + pct=$(check_marking "== 0") + check_err $? "backlog $backlog / $limit Got $pct% marked packets, expected == 0." + log_test "$name backlog < limit" + + # Now push TCP, because non-TCP traffic would be early-dropped after the + # backlog crosses the limit, and we want to make sure that the backlog + # is above the limit. + RET=0 + backlog=$(build_backlog $((3 * limit / 2)) tcp tos=0x01) + check_err $? "Could not build the requested backlog" + pct=$(check_marking ">= 95") + check_err $? "backlog $backlog / $limit Got $pct% marked packets, expected >= 95." + log_test "$name backlog > limit" +} + +do_ecn_test() +{ + local limit=$1; shift + local name=ECN + + $MZ $h1 -p $PKTSZ -A 192.0.2.1 -B 192.0.2.3 -c 0 \ + -a own -b $h3_mac -t tcp -q tos=0x01 & + sleep 1 + + ecn_test_common "$name" $limit + + # Up there we saw that UDP gets accepted when backlog is below the + # limit. Now that it is above, it should all get dropped, and backlog + # building should fail. + RET=0 + build_backlog $((2 * limit)) udp >/dev/null + check_fail $? "UDP traffic went into backlog instead of being early-dropped" + log_test "$name backlog > limit: UDP early-dropped" + + stop_traffic + sleep 1 +} + +do_ecn_nodrop_test() +{ + local limit=$1; shift + local name="ECN nodrop" + + $MZ $h1 -p $PKTSZ -A 192.0.2.1 -B 192.0.2.3 -c 0 \ + -a own -b $h3_mac -t tcp -q tos=0x01 & + sleep 1 + + ecn_test_common "$name" $limit + + # Up there we saw that UDP gets accepted when backlog is below the + # limit. Now that it is above, in nodrop mode, make sure it goes to + # backlog as well. + RET=0 + build_backlog $((2 * limit)) udp >/dev/null + check_err $? "UDP traffic was early-dropped instead of getting into backlog" + log_test "$name backlog > limit: UDP not dropped" + + stop_traffic + sleep 1 +} + +do_red_test() +{ + local limit=$1; shift + local backlog + local pct + + # Use ECN-capable TCP to verify there's no marking even though the queue + # is above limit. + $MZ $h1 -p $PKTSZ -A 192.0.2.1 -B 192.0.2.3 -c 0 \ + -a own -b $h3_mac -t tcp -q tos=0x01 & + + # Pushing below the queue limit should work. + RET=0 + backlog=$(build_backlog $((2 * limit / 3)) tcp tos=0x01) + check_err $? "Could not build the requested backlog" + pct=$(check_marking "== 0") + check_err $? "backlog $backlog / $limit Got $pct% marked packets, expected == 0." + log_test "RED backlog < limit" + + # Pushing above should not. + RET=0 + backlog=$(build_backlog $((3 * limit / 2)) tcp tos=0x01) + check_fail $? "Traffic went into backlog instead of being early-dropped" + pct=$(check_marking "== 0") + check_err $? "backlog $backlog / $limit Got $pct% marked packets, expected == 0." + log_test "RED backlog > limit" + + stop_traffic + sleep 1 +} + +do_red_qevent_test() +{ + local limit=$1; shift + local backlog + local base + local now + local pct + + RET=0 + + $MZ $h1 -p $PKTSZ -A 192.0.2.1 -B 192.0.2.3 -c 0 \ + -a own -b $h3_mac -t udp -q & + sleep 1 + + tc filter add block 10 pref 1234 handle 102 matchall skip_hw \ + action mirred egress mirror dev _drop_test + + # Push to the queue until it's at the limit. The configured limit is + # rounded by the qdisc, so this is the best we can do to get to the real + # limit. + build_backlog $((3 * limit / 2)) udp >/dev/null + + base=$(get_nmirrored) + send_packets udp 100 + sleep 1 + now=$(get_nmirrored) + ((now >= base + 100)) + check_err $? "Dropped packets not observed: 100 expected, $((now - base)) seen" + + tc filter del block 10 pref 1234 handle 102 matchall + + base=$(get_nmirrored) + send_packets udp 100 + sleep 1 + now=$(get_nmirrored) + ((now == base)) + check_err $? "Dropped packets still observed: 0 expected, $((now - base)) seen" + + log_test "RED early_dropped packets mirrored" + + stop_traffic + sleep 1 +} + +do_ecn_qevent_test() +{ + local limit=$1; shift + local name=ECN + + RET=0 + + $MZ $h1 -p $PKTSZ -A 192.0.2.1 -B 192.0.2.3 -c 0 \ + -a own -b $h3_mac -t tcp -q tos=0x01 & + sleep 1 + + tc filter add block 10 pref 1234 handle 102 matchall skip_hw \ + action mirred egress mirror dev _drop_test + + backlog=$(build_backlog $((2 * limit / 3)) tcp tos=0x01) + check_err $? "Could not build the requested backlog" + pct=$(check_mirroring "== 0") + check_err $? "backlog $backlog / $limit Got $pct% mirrored packets, expected == 0." + + backlog=$(build_backlog $((3 * limit / 2)) tcp tos=0x01) + check_err $? "Could not build the requested backlog" + pct=$(check_mirroring ">= 95") + check_err $? "backlog $backlog / $limit Got $pct% mirrored packets, expected >= 95." + + tc filter del block 10 pref 1234 handle 102 matchall + + log_test "ECN marked packets mirrored" + + stop_traffic + sleep 1 +} + +install_qdisc() +{ + local -a args=("$@") + + tc qdisc replace dev $swp3 parent 1:1 handle 11: red \ + limit 1M avpkt $PKTSZ probability 1 \ + min $BACKLOG max $((BACKLOG + 1)) burst 38 "${args[@]}" + sleep 1 +} + +uninstall_qdisc() +{ + tc qdisc del dev $swp3 parent 1:1 +} + +ecn_test() +{ + install_qdisc ecn + do_ecn_test $BACKLOG + uninstall_qdisc +} + +ecn_nodrop_test() +{ + install_qdisc ecn nodrop + do_ecn_nodrop_test $BACKLOG + uninstall_qdisc +} + +red_test() +{ + install_qdisc + do_red_test $BACKLOG + uninstall_qdisc +} + +red_qevent_test() +{ + install_qdisc qevent early_drop block 10 + do_red_qevent_test $BACKLOG + uninstall_qdisc +} + +ecn_qevent_test() +{ + install_qdisc ecn qevent mark block 10 + do_ecn_qevent_test $BACKLOG + uninstall_qdisc +} + +trap cleanup EXIT + +setup_prepare +setup_wait + +tests_run + +exit $EXIT_STATUS
Signature · 1 line
2.20.1
Stephen Hemminger
maintainerQuoted 11 lines · On Sat, 27 Jun 2020 01:45:24 +0300 Petr Machata <[email protected]> wrote:
The Spectrum hardware allows execution of one of several actions as a result of queue management decisions: tail-dropping, early-dropping, marking a packet, or passing a configured latency threshold or buffer size. Such packets can be mirrored, trapped, or sampled.
Modeling the action to be taken as simply a TC action is very attractive, but it is not obvious where to put these actions. At least with ECN marking one could imagine a tree of qdiscs and classifiers that effectively accomplishes this task, albeit in an impractically complex manner. But there is just no way to match on dropped-ness of a packet, let alone dropped-ness due to a particular reason.
Would a BPF based hook be more flexible and reuse more existing infrastructure?
Quoted 15 lines · Stephen Hemminger
Quoted 11 lines · On Sat, 27 Jun 2020 01:45:24 +0300 Petr Machata <[email protected]> wrote:
The Spectrum hardware allows execution of one of several actions as a result of queue management decisions: tail-dropping, early-dropping, marking a packet, or passing a configured latency threshold or buffer size. Such packets can be mirrored, trapped, or sampled.
Modeling the action to be taken as simply a TC action is very attractive, but it is not obvious where to put these actions. At least with ECN marking one could imagine a tree of qdiscs and classifiers that effectively accomplishes this task, albeit in an impractically complex manner. But there is just no way to match on dropped-ness of a packet, let alone dropped-ness due to a particular reason.
Quoted 2 lines
Would a BPF based hook be more flexible and reuse more existing infrastructure?
This does reuse the existing infrastructure though: filters, actions, shared blocks, qdiscs invoking blocks, none of that is new.
And BPF can still be invoked though classifier and / or action bpf. It looks like you get the best of both worlds here: something symbolic for those of us that use the filter infrastructure, and a well-defined hook for those of us who like the BPF approach.
David Miller
maintainerFrom: Petr Machata <[email protected]>
Date: Sat, 27 Jun 2020 01:45:24 +0300
Quoted 12 lines
The Spectrum hardware allows execution of one of several actions as a result of queue management decisions: tail-dropping, early-dropping, marking a packet, or passing a configured latency threshold or buffer size. Such packets can be mirrored, trapped, or sampled.
Modeling the action to be taken as simply a TC action is very attractive, but it is not obvious where to put these actions. At least with ECN marking one could imagine a tree of qdiscs and classifiers that effectively accomplishes this task, albeit in an impractically complex manner. But there is just no way to match on dropped-ness of a packet, let alone dropped-ness due to a particular reason.
Quoted 7 lines
To allow configuring user-defined actions as a result of inner workings of a qdisc, this patch set introduces a concept of qevents. Those are attach points for TC blocks, where filters can be put that are executed as the packet hits well-defined points in the qdisc algorithms. The attached blocks can be shared, in a manner similar to clsact ingress and egress blocks, arbitrary classifiers with arbitrary actions can be put on them, etc.
...
Series applied, thank you.