| Message ID | 155144955045.28287.6187020206969282688.stgit@alrua-x1 |
|---|---|
| State | Changes Requested |
| Delegated to: | BPF Maintainers |
| Headers | show |
| Series | xdp: Use a default map for xdp_redirect helper | expand |
On Fri, 01 Mar 2019 15:12:30 +0100, Toke Høiland-Jørgensen wrote: > An XDP program can redirect packets between interfaces using either the > xdp_redirect() helper or the xdp_redirect_map() helper. Apart from the > flexibility of updating maps from userspace, the redirect_map() helper also > uses the map structure to batch packets, which results in a significant > (around 50%) performance boost. However, the xdp_redirect() API is simpler > if one just wants to redirect to another interface, which means people tend > to use this interface and then wonder why they getter worse performance > than expected. > > This patch seeks to close this performance difference between the two APIs. > It achieves this by changing xdp_redirect() to use a hidden devmap for > looking up destination interfaces, thus gaining the batching benefit with > no visible difference from the user API point of view. > > A hidden per-namespace map is allocated when an XDP program that uses the > non-map xdp_redirect() helper is first loaded. This map is populated with > all available interfaces in its namespace, and kept up to date as > interfaces come and go. Once allocated, the map is kept around until the > namespace is removed. > > The hidden map uses the ifindex as map key, which means they are limited to > ifindexes smaller than the map size of 64. A later patch introduces a new > map type to lift this restriction. > > Performance numbers: > > Before patch: > xdp_redirect: 5426035 pkt/s > xdp_redirect_map: 8412754 pkt/s > > After patch: > xdp_redirect: 8314702 pkt/s > xdp_redirect_map: 8411854 pkt/s > > This corresponds to a 53% increase in xdp_redirect performance, or a > reduction in per-packet processing time by 64 nanoseconds. > > Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com> > diff --git a/include/linux/bpf.h b/include/linux/bpf.h > index de18227b3d95..060c5850af3b 100644 > --- a/include/linux/bpf.h > +++ b/include/linux/bpf.h > @@ -25,6 +25,7 @@ struct sock; > struct seq_file; > struct btf; > struct btf_type; > +struct net; > > /* map is generic key/value storage optionally accesible by eBPF programs */ > struct bpf_map_ops { > @@ -533,6 +534,7 @@ extern const struct bpf_verifier_ops tc_cls_act_analyzer_ops; > extern const struct bpf_verifier_ops xdp_analyzer_ops; > > struct bpf_prog *bpf_prog_get(u32 ufd); > +struct bpf_prog *bpf_prog_get_by_id(u32 id); > struct bpf_prog *bpf_prog_get_type_dev(u32 ufd, enum bpf_prog_type type, > bool attach_drv); > struct bpf_prog * __must_check bpf_prog_add(struct bpf_prog *prog, int i); > @@ -612,6 +614,11 @@ struct xdp_buff; > struct sk_buff; > > struct bpf_dtab_netdev *__dev_map_lookup_elem(struct bpf_map *map, u32 key); > +struct bpf_map *__dev_map_get_default_map(struct net_device *dev); > +int dev_map_ensure_default_map(struct net *net); > +void dev_map_put_default_map(struct net *net); > +int dev_map_inc_redirect_count(void); > +void dev_map_dec_redirect_count(void); > void __dev_map_insert_ctx(struct bpf_map *map, u32 index); > void __dev_map_flush(struct bpf_map *map); > int dev_map_enqueue(struct bpf_dtab_netdev *dst, struct xdp_buff *xdp, > @@ -641,6 +648,11 @@ static inline struct bpf_prog *bpf_prog_get(u32 ufd) > return ERR_PTR(-EOPNOTSUPP); > } > > +static inline struct bpf_prog *bpf_prog_get_by_id(u32 id) > +{ > + return ERR_PTR(-EOPNOTSUPP); > +} > + > static inline struct bpf_prog *bpf_prog_get_type_dev(u32 ufd, > enum bpf_prog_type type, > bool attach_drv) > @@ -693,6 +705,29 @@ static inline struct net_device *__dev_map_lookup_elem(struct bpf_map *map, > return NULL; > } > > +static inline struct bpf_map *__dev_map_get_default_map(struct net_device *dev) > +{ > + return NULL; > +} > + > +static inline int dev_map_ensure_default_map(struct net *net) > +{ > + return 0; > +} > + > +static inline void dev_map_put_default_map(struct net *net) > +{ > +} > + > +static inline int dev_map_inc_redirect_count(void) > +{ > + return 0; > +} > + > +static inline void dev_map_dec_redirect_count(void) > +{ > +} > + > static inline void __dev_map_insert_ctx(struct bpf_map *map, u32 index) > { > } > diff --git a/include/linux/filter.h b/include/linux/filter.h > index 95e2d7ebdf21..dd6bbbab32f7 100644 > --- a/include/linux/filter.h > +++ b/include/linux/filter.h > @@ -507,6 +507,8 @@ struct bpf_prog { > gpl_compatible:1, /* Is filter GPL compatible? */ > cb_access:1, /* Is control block accessed? */ > dst_needed:1, /* Do we need dst entry? */ > + redirect_needed:1, /* Does program need access to xdp_redirect? */ > + redirect_used:1, /* Does program use xdp_redirect? */ > blinded:1, /* Was blinded */ > is_func:1, /* program is a bpf function */ > kprobe_override:1, /* Do we override a kprobe? */ > diff --git a/include/net/net_namespace.h b/include/net/net_namespace.h > index a68ced28d8f4..6706ecc25d8f 100644 > --- a/include/net/net_namespace.h > +++ b/include/net/net_namespace.h > @@ -162,7 +162,7 @@ struct net { > #if IS_ENABLED(CONFIG_CAN) > struct netns_can can; > #endif > -#ifdef CONFIG_XDP_SOCKETS > +#ifdef CONFIG_BPF_SYSCALL > struct netns_xdp xdp; > #endif > struct sock *diag_nlsk; > diff --git a/include/net/netns/xdp.h b/include/net/netns/xdp.h > index e5734261ba0a..4935dfe1cf43 100644 > --- a/include/net/netns/xdp.h > +++ b/include/net/netns/xdp.h > @@ -4,10 +4,21 @@ > > #include <linux/rculist.h> > #include <linux/mutex.h> > +#include <linux/atomic.h> > + > +struct bpf_dtab; > + > +struct bpf_dtab_container { > + struct bpf_dtab __rcu *dtab; > + atomic_t refcnt; refcount_t ? I'm not sure what the rules are for non-performance critical stuff are.. > +}; > > struct netns_xdp { > +#ifdef CONFIG_XDP_SOCKETS > struct mutex lock; > struct hlist_head list; > +#endif > + struct bpf_dtab_container default_map; > }; > > #endif /* __NETNS_XDP_H__ */ > diff --git a/kernel/bpf/devmap.c b/kernel/bpf/devmap.c > index 1037fc08c504..e55707e62b60 100644 > --- a/kernel/bpf/devmap.c > +++ b/kernel/bpf/devmap.c > @@ -56,6 +56,9 @@ > (BPF_F_NUMA_NODE | BPF_F_RDONLY | BPF_F_WRONLY) > > #define DEV_MAP_BULK_SIZE 16 > +#define DEV_MAP_DEFAULT_SIZE 8 > +#define BPF_MAX_REFCNT 32768 > + > struct xdp_bulk_queue { > struct xdp_frame *q[DEV_MAP_BULK_SIZE]; > struct net_device *dev_rx; > @@ -80,6 +83,7 @@ struct bpf_dtab { > > static DEFINE_SPINLOCK(dev_map_lock); > static LIST_HEAD(dev_map_list); > +static atomic_t global_redirect_use = {}; REFCOUNT_INIT(0) or ATOMIC_INIT(0) > static u64 dev_map_bitmap_size(const union bpf_attr *attr) > { > @@ -332,6 +336,18 @@ struct bpf_dtab_netdev *__dev_map_lookup_elem(struct bpf_map *map, u32 key) > return obj; > } > > +/* This is only being called from xdp_do_redirect() if the xdp_redirect helper > + * is used; the default map is allocated on XDP program load if the helper is > + * used, so will always be available at this point. > + */ Yet it does check for NULL..? > +struct bpf_map *__dev_map_get_default_map(struct net_device *dev) > +{ > + struct net *net = dev_net(dev); > + struct bpf_dtab *dtab = rcu_dereference(net->xdp.default_map.dtab); This init is long and break rxmas tree rules, separate line perhaps? > + return dtab ? &dtab->map : NULL; > +} > + > /* Runs under RCU-read-side, plus in softirq under NAPI protection. > * Thus, safe percpu variable access. > */ > @@ -533,14 +549,210 @@ const struct bpf_map_ops dev_map_ops = { > .map_check_btf = map_check_no_btf, > }; > > +static inline struct net *bpf_default_map_to_net(struct bpf_dtab_container *cont) > +{ > + struct netns_xdp *xdp = container_of(cont, struct netns_xdp, default_map); > + > + return container_of(xdp, struct net, xdp); > +} > + > +static void __dev_map_release_default_map(struct bpf_dtab_container *cont) > +{ > + struct bpf_dtab *dtab = NULL; > + > + lockdep_assert_held(&dev_map_lock); > + > + dtab = rcu_dereference(cont->dtab); > + if (dtab) { How can it be NULL if it's refcounted? hm.. > + list_del_rcu(&dtab->list); > + rcu_assign_pointer(cont->dtab, NULL); > + bpf_clear_redirect_map(&dtab->map); > + call_rcu(&dtab->rcu, __dev_map_free); > + } > +} > + > +void dev_map_put_default_map(struct net *net) > +{ > + if (atomic_dec_and_test(&net->xdp.default_map.refcnt)) { > + spin_lock(&dev_map_lock); > + __dev_map_release_default_map(&net->xdp.default_map); > + spin_unlock(&dev_map_lock); > + } > +} > + > +static int __init_default_map(struct bpf_dtab_container *cont) > +{ > + struct net *net = bpf_default_map_to_net(cont); > + struct bpf_dtab *dtab, *old_dtab; > + int size = DEV_MAP_DEFAULT_SIZE; > + struct net_device *netdev; > + union bpf_attr attr = {}; > + u32 idx; > + int err; > + > + lockdep_assert_held(&dev_map_lock); > + > + if (!atomic_read(&global_redirect_use)) > + return 0; > + > + for_each_netdev(net, netdev) > + if (netdev->ifindex >= size) > + size <<= 1; > + > + old_dtab = rcu_dereference(cont->dtab); > + if (old_dtab && old_dtab->map.max_entries == size) > + return 0; > + > + dtab = kzalloc(sizeof(*dtab), GFP_USER); You are calling this with a spin lock held :( > + if (!dtab) > + return -ENOMEM; > + > + attr.map_type = BPF_MAP_TYPE_DEVMAP; > + attr.max_entries = size; > + attr.value_size = 4; > + attr.key_size = 4; > + > + err = dev_map_init_map(dtab, &attr, false); > + if (err) { > + kfree(dtab); > + return err; > + } > + > + for_each_netdev(net, netdev) { > + idx = netdev->ifindex; > + err = __dev_map_update_elem(net, &dtab->map, &idx, &idx, 0); > + if (err) { > + __dev_map_free(&dtab->rcu); This map wasn't visible yet, so probably no need to go through RCU? Not that it matters much for an error path.. > + return err; > + } > + } > + > + rcu_assign_pointer(cont->dtab, dtab); > + list_add_tail_rcu(&dtab->list, &dev_map_list); > + > + if (old_dtab) { > + list_del_rcu(&old_dtab->list); > + bpf_clear_redirect_map(&old_dtab->map); > + call_rcu(&old_dtab->rcu, __dev_map_free); > + } > + > + return 0; > +} > + > diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c > index 1b9496c41383..f1b2f01e7ca1 100644 > --- a/kernel/bpf/verifier.c > +++ b/kernel/bpf/verifier.c > @@ -7609,6 +7609,17 @@ static int fixup_bpf_calls(struct bpf_verifier_env *env) > prog->dst_needed = 1; > if (insn->imm == BPF_FUNC_get_prandom_u32) > bpf_user_rnd_init_once(); > + if (insn->imm == BPF_FUNC_redirect) { > + prog->redirect_needed = true; nit: this field is a bitfield not boolean > + if (!prog->redirect_used) { > + int err = dev_map_inc_redirect_count(); Don't call complex functions as variable init, please > + if (err) > + return err; > + prog->redirect_used = true; > + } > + } > + > if (insn->imm == BPF_FUNC_override_return) > prog->kprobe_override = 1; > if (insn->imm == BPF_FUNC_tail_call) { > @@ -7618,6 +7629,7 @@ static int fixup_bpf_calls(struct bpf_verifier_env *env) > * the program array. > */ > prog->cb_access = 1; > + prog->redirect_needed = true; > env->prog->aux->stack_depth = MAX_BPF_STACK; > env->prog->aux->max_pkt_offset = MAX_PACKET_OFF; > > diff --git a/net/core/dev.c b/net/core/dev.c > index 2b67f2aa59dd..1df20d529026 100644 > --- a/net/core/dev.c > +++ b/net/core/dev.c > @@ -7990,6 +7990,21 @@ u32 __dev_xdp_query(struct net_device *dev, bpf_op_t bpf_op, > return xdp.prog_id; > } > > +static struct bpf_prog *dev_xdp_get_prog(struct net_device *dev) > +{ > + struct bpf_prog *prog; > + u32 prog_id; > + > + prog_id = __dev_xdp_query(dev, dev->netdev_ops->ndo_bpf, XDP_QUERY_PROG); if (!prog_id) return NULL; ... > + if (prog_id) { > + prog = bpf_prog_get_by_id(prog_id); > + if (!IS_ERR(prog)) > + return prog; > + } > + > + return NULL; > +} > + > static int dev_xdp_install(struct net_device *dev, bpf_op_t bpf_op, > struct netlink_ext_ack *extack, u32 flags, > struct bpf_prog *prog) > @@ -8024,9 +8039,18 @@ static void dev_xdp_uninstall(struct net_device *dev) > memset(&xdp, 0, sizeof(xdp)); > xdp.command = XDP_QUERY_PROG; > WARN_ON(ndo_bpf(dev, &xdp)); > - if (xdp.prog_id) > + if (xdp.prog_id) { > + struct bpf_prog *prog = bpf_prog_get_by_id(xdp.prog_id); > + > + if (!IS_ERR(prog)) { > + if (prog->redirect_needed) > + dev_map_put_default_map(dev_net(dev)); > + bpf_prog_put(prog); > + } > + > WARN_ON(dev_xdp_install(dev, ndo_bpf, NULL, xdp.prog_flags, > NULL)); > + } > > /* Remove HW offload */ > memset(&xdp, 0, sizeof(xdp)); > @@ -8091,6 +8115,23 @@ int dev_change_xdp_fd(struct net_device *dev, struct netlink_ext_ack *extack, > bpf_prog_put(prog); > return -EINVAL; > } > + > + if (!offload && bpf_op == ops->ndo_bpf && > + prog->redirect_needed) { > + err = dev_map_ensure_default_map(dev_net(dev)); > + if (err) { > + NL_SET_ERR_MSG(extack, "unable to allocate default map for xdp_redirect()"); > + return err; > + } > + } > + } else { > + struct bpf_prog *old_prog = dev_xdp_get_prog(dev); > + > + if (old_prog) { > + if (old_prog->redirect_needed) > + dev_map_put_default_map(dev_net(dev)); I think you should hold the ref to this program and only release the map _after_ new program was installed. You also have to do this for replacing a program with a different one. Could you please add tests for the replacement combinations? > + bpf_prog_put(old_prog); > + } > } > err = dev_xdp_install(dev, bpf_op, extack, flags, prog); > @@ -9333,6 +9374,7 @@ EXPORT_SYMBOL(unregister_netdev); > int dev_change_net_namespace(struct net_device *dev, struct net *net, const char *pat) > { > int err, new_nsid, new_ifindex; > + struct bpf_prog *prog = NULL; > > ASSERT_RTNL(); > > @@ -9350,6 +9392,15 @@ int dev_change_net_namespace(struct net_device *dev, struct net *net, const char > if (net_eq(dev_net(dev), net)) > goto out; > > + prog = dev_xdp_get_prog(dev); > + if (prog) { > + if (prog->redirect_needed) > + err = dev_map_ensure_default_map(net); > + > + if (err) > + goto out; > + } prog = dev_xdp_get_prog(dev); if (prog && prog->redirect_needed) { err = dev_map_ensure_default_map(net); if (err) goto out; } > /* Pick the destination device name, and ensure > * we can use it in the destination network namespace. > */ > @@ -9388,6 +9439,9 @@ int dev_change_net_namespace(struct net_device *dev, struct net *net, const char > call_netdevice_notifiers(NETDEV_UNREGISTER, dev); > rcu_barrier(); > > + if (prog && prog->redirect_needed) > + dev_map_put_default_map(dev_net(dev)); > + > new_nsid = peernet2id_alloc(dev_net(dev), net); > /* If there is an ifindex conflict assign a new one */ > if (__dev_get_by_index(net, dev->ifindex)) > @@ -9435,6 +9489,9 @@ int dev_change_net_namespace(struct net_device *dev, struct net *net, const char > synchronize_net(); > err = 0; > out: > + if (prog) > + bpf_prog_put(prog); > + > return err; > } > EXPORT_SYMBOL_GPL(dev_change_net_namespace);
Hi Jakub Thanks for the review! A few comments/questions below. > On Fri, 01 Mar 2019 15:12:30 +0100, Toke Høiland-Jørgensen wrote: >> An XDP program can redirect packets between interfaces using either the >> xdp_redirect() helper or the xdp_redirect_map() helper. Apart from the >> flexibility of updating maps from userspace, the redirect_map() helper also >> uses the map structure to batch packets, which results in a significant >> (around 50%) performance boost. However, the xdp_redirect() API is simpler >> if one just wants to redirect to another interface, which means people tend >> to use this interface and then wonder why they getter worse performance >> than expected. >> >> This patch seeks to close this performance difference between the two APIs. >> It achieves this by changing xdp_redirect() to use a hidden devmap for >> looking up destination interfaces, thus gaining the batching benefit with >> no visible difference from the user API point of view. >> >> A hidden per-namespace map is allocated when an XDP program that uses the >> non-map xdp_redirect() helper is first loaded. This map is populated with >> all available interfaces in its namespace, and kept up to date as >> interfaces come and go. Once allocated, the map is kept around until the >> namespace is removed. >> >> The hidden map uses the ifindex as map key, which means they are limited to >> ifindexes smaller than the map size of 64. A later patch introduces a new >> map type to lift this restriction. >> >> Performance numbers: >> >> Before patch: >> xdp_redirect: 5426035 pkt/s >> xdp_redirect_map: 8412754 pkt/s >> >> After patch: >> xdp_redirect: 8314702 pkt/s >> xdp_redirect_map: 8411854 pkt/s >> >> This corresponds to a 53% increase in xdp_redirect performance, or a >> reduction in per-packet processing time by 64 nanoseconds. >> >> Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com> > >> diff --git a/include/linux/bpf.h b/include/linux/bpf.h >> index de18227b3d95..060c5850af3b 100644 >> --- a/include/linux/bpf.h >> +++ b/include/linux/bpf.h >> @@ -25,6 +25,7 @@ struct sock; >> struct seq_file; >> struct btf; >> struct btf_type; >> +struct net; >> >> /* map is generic key/value storage optionally accesible by eBPF programs */ >> struct bpf_map_ops { >> @@ -533,6 +534,7 @@ extern const struct bpf_verifier_ops tc_cls_act_analyzer_ops; >> extern const struct bpf_verifier_ops xdp_analyzer_ops; >> >> struct bpf_prog *bpf_prog_get(u32 ufd); >> +struct bpf_prog *bpf_prog_get_by_id(u32 id); >> struct bpf_prog *bpf_prog_get_type_dev(u32 ufd, enum bpf_prog_type type, >> bool attach_drv); >> struct bpf_prog * __must_check bpf_prog_add(struct bpf_prog *prog, int i); >> @@ -612,6 +614,11 @@ struct xdp_buff; >> struct sk_buff; >> >> struct bpf_dtab_netdev *__dev_map_lookup_elem(struct bpf_map *map, u32 key); >> +struct bpf_map *__dev_map_get_default_map(struct net_device *dev); >> +int dev_map_ensure_default_map(struct net *net); >> +void dev_map_put_default_map(struct net *net); >> +int dev_map_inc_redirect_count(void); >> +void dev_map_dec_redirect_count(void); >> void __dev_map_insert_ctx(struct bpf_map *map, u32 index); >> void __dev_map_flush(struct bpf_map *map); >> int dev_map_enqueue(struct bpf_dtab_netdev *dst, struct xdp_buff *xdp, >> @@ -641,6 +648,11 @@ static inline struct bpf_prog *bpf_prog_get(u32 ufd) >> return ERR_PTR(-EOPNOTSUPP); >> } >> >> +static inline struct bpf_prog *bpf_prog_get_by_id(u32 id) >> +{ >> + return ERR_PTR(-EOPNOTSUPP); >> +} >> + >> static inline struct bpf_prog *bpf_prog_get_type_dev(u32 ufd, >> enum bpf_prog_type type, >> bool attach_drv) >> @@ -693,6 +705,29 @@ static inline struct net_device *__dev_map_lookup_elem(struct bpf_map *map, >> return NULL; >> } >> >> +static inline struct bpf_map *__dev_map_get_default_map(struct net_device *dev) >> +{ >> + return NULL; >> +} >> + >> +static inline int dev_map_ensure_default_map(struct net *net) >> +{ >> + return 0; >> +} >> + >> +static inline void dev_map_put_default_map(struct net *net) >> +{ >> +} >> + >> +static inline int dev_map_inc_redirect_count(void) >> +{ >> + return 0; >> +} >> + >> +static inline void dev_map_dec_redirect_count(void) >> +{ >> +} >> + >> static inline void __dev_map_insert_ctx(struct bpf_map *map, u32 index) >> { >> } >> diff --git a/include/linux/filter.h b/include/linux/filter.h >> index 95e2d7ebdf21..dd6bbbab32f7 100644 >> --- a/include/linux/filter.h >> +++ b/include/linux/filter.h >> @@ -507,6 +507,8 @@ struct bpf_prog { >> gpl_compatible:1, /* Is filter GPL compatible? */ >> cb_access:1, /* Is control block accessed? */ >> dst_needed:1, /* Do we need dst entry? */ >> + redirect_needed:1, /* Does program need access to xdp_redirect? */ >> + redirect_used:1, /* Does program use xdp_redirect? */ >> blinded:1, /* Was blinded */ >> is_func:1, /* program is a bpf function */ >> kprobe_override:1, /* Do we override a kprobe? */ >> diff --git a/include/net/net_namespace.h b/include/net/net_namespace.h >> index a68ced28d8f4..6706ecc25d8f 100644 >> --- a/include/net/net_namespace.h >> +++ b/include/net/net_namespace.h >> @@ -162,7 +162,7 @@ struct net { >> #if IS_ENABLED(CONFIG_CAN) >> struct netns_can can; >> #endif >> -#ifdef CONFIG_XDP_SOCKETS >> +#ifdef CONFIG_BPF_SYSCALL >> struct netns_xdp xdp; >> #endif >> struct sock *diag_nlsk; >> diff --git a/include/net/netns/xdp.h b/include/net/netns/xdp.h >> index e5734261ba0a..4935dfe1cf43 100644 >> --- a/include/net/netns/xdp.h >> +++ b/include/net/netns/xdp.h >> @@ -4,10 +4,21 @@ >> >> #include <linux/rculist.h> >> #include <linux/mutex.h> >> +#include <linux/atomic.h> >> + >> +struct bpf_dtab; >> + >> +struct bpf_dtab_container { >> + struct bpf_dtab __rcu *dtab; >> + atomic_t refcnt; > > refcount_t ? I'm not sure what the rules are for non-performance > critical stuff are.. I based the use of atomic_t on what bpf/syscall.c is doing. The problem with using refcount_t is that it's a bit of a weird refcount because it's not strictly tied to the object lifetime (as the objects are only allocated if *both* the globals and the per-namespace refcnt are >0). [... snip things that I'll just fix ...] > >> +static void __dev_map_release_default_map(struct bpf_dtab_container *cont) >> +{ >> + struct bpf_dtab *dtab = NULL; >> + >> + lockdep_assert_held(&dev_map_lock); >> + >> + dtab = rcu_dereference(cont->dtab); >> + if (dtab) { > > How can it be NULL if it's refcounted? hm.. See above; it's not actually the object that is refcounted, it's a count of the number of active XDP programs in the namespace, which still needs to be kept so we can allocate the devmaps when a REDIRECT program is loaded. >> + list_del_rcu(&dtab->list); >> + rcu_assign_pointer(cont->dtab, NULL); >> + bpf_clear_redirect_map(&dtab->map); >> + call_rcu(&dtab->rcu, __dev_map_free); >> + } >> +} >> + >> +void dev_map_put_default_map(struct net *net) >> +{ >> + if (atomic_dec_and_test(&net->xdp.default_map.refcnt)) { >> + spin_lock(&dev_map_lock); >> + __dev_map_release_default_map(&net->xdp.default_map); >> + spin_unlock(&dev_map_lock); >> + } >> +} >> + >> +static int __init_default_map(struct bpf_dtab_container *cont) >> +{ >> + struct net *net = bpf_default_map_to_net(cont); >> + struct bpf_dtab *dtab, *old_dtab; >> + int size = DEV_MAP_DEFAULT_SIZE; >> + struct net_device *netdev; >> + union bpf_attr attr = {}; >> + u32 idx; >> + int err; >> + >> + lockdep_assert_held(&dev_map_lock); >> + >> + if (!atomic_read(&global_redirect_use)) >> + return 0; >> + >> + for_each_netdev(net, netdev) >> + if (netdev->ifindex >= size) >> + size <<= 1; >> + >> + old_dtab = rcu_dereference(cont->dtab); >> + if (old_dtab && old_dtab->map.max_entries == size) >> + return 0; >> + >> + dtab = kzalloc(sizeof(*dtab), GFP_USER); > > You are calling this with a spin lock held :( Yeah, not just this, there are more allocations in dev_map_init_map(). Hmm, I guess I need to think about the concurrency bits some more; it is made a bit tricky by the refcounts being outside the objects. Basically what I'm trying to protect against is the case where the global refcount goes to zero and then immediately back to one (or vice versa), which results in a deallocation immediately followed by an allocation. I was worried about the allocation and deallocation racing with each other (because there's a window after the refcnt is checked before the (de)allocation begins). It's not quite clear to me whether RCU is enough to protect against this... Thoughts? >> + if (!dtab) >> + return -ENOMEM; >> + >> + attr.map_type = BPF_MAP_TYPE_DEVMAP; >> + attr.max_entries = size; >> + attr.value_size = 4; >> + attr.key_size = 4; >> + >> + err = dev_map_init_map(dtab, &attr, false); >> + if (err) { >> + kfree(dtab); >> + return err; >> + } >> + >> + for_each_netdev(net, netdev) { >> + idx = netdev->ifindex; >> + err = __dev_map_update_elem(net, &dtab->map, &idx, &idx, 0); >> + if (err) { >> + __dev_map_free(&dtab->rcu); > > This map wasn't visible yet, so probably no need to go through RCU? > Not that it matters much for an error path.. It's not actually using call_rcu(), it's just calling a function that expects a struct_rcu as a parameter (to avoid another wrapper function) :) >> + return err; >> + } >> + } >> + >> + rcu_assign_pointer(cont->dtab, dtab); >> + list_add_tail_rcu(&dtab->list, &dev_map_list); >> + >> + if (old_dtab) { >> + list_del_rcu(&old_dtab->list); >> + bpf_clear_redirect_map(&old_dtab->map); >> + call_rcu(&old_dtab->rcu, __dev_map_free); >> + } >> + >> + return 0; >> +} >> + > >> diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c >> index 1b9496c41383..f1b2f01e7ca1 100644 >> --- a/kernel/bpf/verifier.c >> +++ b/kernel/bpf/verifier.c >> @@ -7609,6 +7609,17 @@ static int fixup_bpf_calls(struct bpf_verifier_env *env) >> prog->dst_needed = 1; >> if (insn->imm == BPF_FUNC_get_prandom_u32) >> bpf_user_rnd_init_once(); >> + if (insn->imm == BPF_FUNC_redirect) { >> + prog->redirect_needed = true; > > nit: this field is a bitfield not boolean > >> + if (!prog->redirect_used) { >> + int err = dev_map_inc_redirect_count(); > > Don't call complex functions as variable init, please > >> + if (err) >> + return err; >> + prog->redirect_used = true; >> + } >> + } >> + >> if (insn->imm == BPF_FUNC_override_return) >> prog->kprobe_override = 1; >> if (insn->imm == BPF_FUNC_tail_call) { >> @@ -7618,6 +7629,7 @@ static int fixup_bpf_calls(struct bpf_verifier_env *env) >> * the program array. >> */ >> prog->cb_access = 1; >> + prog->redirect_needed = true; >> env->prog->aux->stack_depth = MAX_BPF_STACK; >> env->prog->aux->max_pkt_offset = MAX_PACKET_OFF; >> >> diff --git a/net/core/dev.c b/net/core/dev.c >> index 2b67f2aa59dd..1df20d529026 100644 >> --- a/net/core/dev.c >> +++ b/net/core/dev.c >> @@ -7990,6 +7990,21 @@ u32 __dev_xdp_query(struct net_device *dev, bpf_op_t bpf_op, >> return xdp.prog_id; >> } >> >> +static struct bpf_prog *dev_xdp_get_prog(struct net_device *dev) >> +{ >> + struct bpf_prog *prog; >> + u32 prog_id; >> + >> + prog_id = __dev_xdp_query(dev, dev->netdev_ops->ndo_bpf, XDP_QUERY_PROG); > > if (!prog_id) > return NULL; > > ... > >> + if (prog_id) { >> + prog = bpf_prog_get_by_id(prog_id); >> + if (!IS_ERR(prog)) >> + return prog; >> + } >> + >> + return NULL; >> +} >> + >> static int dev_xdp_install(struct net_device *dev, bpf_op_t bpf_op, >> struct netlink_ext_ack *extack, u32 flags, >> struct bpf_prog *prog) >> @@ -8024,9 +8039,18 @@ static void dev_xdp_uninstall(struct net_device *dev) >> memset(&xdp, 0, sizeof(xdp)); >> xdp.command = XDP_QUERY_PROG; >> WARN_ON(ndo_bpf(dev, &xdp)); >> - if (xdp.prog_id) >> + if (xdp.prog_id) { >> + struct bpf_prog *prog = bpf_prog_get_by_id(xdp.prog_id); >> + >> + if (!IS_ERR(prog)) { >> + if (prog->redirect_needed) >> + dev_map_put_default_map(dev_net(dev)); >> + bpf_prog_put(prog); >> + } >> + >> WARN_ON(dev_xdp_install(dev, ndo_bpf, NULL, xdp.prog_flags, >> NULL)); >> + } >> >> /* Remove HW offload */ >> memset(&xdp, 0, sizeof(xdp)); >> @@ -8091,6 +8115,23 @@ int dev_change_xdp_fd(struct net_device *dev, struct netlink_ext_ack *extack, >> bpf_prog_put(prog); >> return -EINVAL; >> } >> + >> + if (!offload && bpf_op == ops->ndo_bpf && >> + prog->redirect_needed) { >> + err = dev_map_ensure_default_map(dev_net(dev)); >> + if (err) { >> + NL_SET_ERR_MSG(extack, "unable to allocate default map for xdp_redirect()"); >> + return err; >> + } >> + } >> + } else { >> + struct bpf_prog *old_prog = dev_xdp_get_prog(dev); >> + >> + if (old_prog) { >> + if (old_prog->redirect_needed) >> + dev_map_put_default_map(dev_net(dev)); > > I think you should hold the ref to this program and only release the > map _after_ new program was installed. > > You also have to do this for replacing a program with a different one. Ah, right, good points. > Could you please add tests for the replacement combinations? Sure. But, erm, where? selftests? :) [... snip a few more things that I'll just fix ...] -Toke
On Mon, 04 Mar 2019 12:58:51 +0100, Toke Høiland-Jørgensen wrote: > >> diff --git a/include/net/netns/xdp.h b/include/net/netns/xdp.h > >> index e5734261ba0a..4935dfe1cf43 100644 > >> --- a/include/net/netns/xdp.h > >> +++ b/include/net/netns/xdp.h > >> @@ -4,10 +4,21 @@ > >> > >> #include <linux/rculist.h> > >> #include <linux/mutex.h> > >> +#include <linux/atomic.h> > >> + > >> +struct bpf_dtab; > >> + > >> +struct bpf_dtab_container { > >> + struct bpf_dtab __rcu *dtab; > >> + atomic_t refcnt; > > > > refcount_t ? I'm not sure what the rules are for non-performance > > critical stuff are.. > > I based the use of atomic_t on what bpf/syscall.c is doing. The problem > with using refcount_t is that it's a bit of a weird refcount because > it's not strictly tied to the object lifetime (as the objects are only > allocated if *both* the globals and the per-namespace refcnt are >0). > > >> +static void __dev_map_release_default_map(struct bpf_dtab_container *cont) > >> +{ > >> + struct bpf_dtab *dtab = NULL; > >> + > >> + lockdep_assert_held(&dev_map_lock); > >> + > >> + dtab = rcu_dereference(cont->dtab); > >> + if (dtab) { > > > > How can it be NULL if it's refcounted? hm.. > > See above; it's not actually the object that is refcounted, it's a count > of the number of active XDP programs in the namespace, which still needs > to be kept so we can allocate the devmaps when a REDIRECT program is > loaded. I think I got it now. I wonder if anything could be done to improve the readability of the dual-refcnting :S Maybe use the "needed"/"used" terms throughout? > >> + list_del_rcu(&dtab->list); > >> + rcu_assign_pointer(cont->dtab, NULL); > >> + bpf_clear_redirect_map(&dtab->map); > >> + call_rcu(&dtab->rcu, __dev_map_free); > >> + } > >> +} > >> + > >> +void dev_map_put_default_map(struct net *net) > >> +{ > >> + if (atomic_dec_and_test(&net->xdp.default_map.refcnt)) { > >> + spin_lock(&dev_map_lock); > >> + __dev_map_release_default_map(&net->xdp.default_map); > >> + spin_unlock(&dev_map_lock); > >> + } > >> +} > >> + > >> +static int __init_default_map(struct bpf_dtab_container *cont) > >> +{ > >> + struct net *net = bpf_default_map_to_net(cont); > >> + struct bpf_dtab *dtab, *old_dtab; > >> + int size = DEV_MAP_DEFAULT_SIZE; > >> + struct net_device *netdev; > >> + union bpf_attr attr = {}; > >> + u32 idx; > >> + int err; > >> + > >> + lockdep_assert_held(&dev_map_lock); > >> + > >> + if (!atomic_read(&global_redirect_use)) > >> + return 0; > >> + > >> + for_each_netdev(net, netdev) > >> + if (netdev->ifindex >= size) > >> + size <<= 1; > >> + > >> + old_dtab = rcu_dereference(cont->dtab); > >> + if (old_dtab && old_dtab->map.max_entries == size) > >> + return 0; > >> + > >> + dtab = kzalloc(sizeof(*dtab), GFP_USER); > > > > You are calling this with a spin lock held :( > > Yeah, not just this, there are more allocations in dev_map_init_map(). > Hmm, I guess I need to think about the concurrency bits some more; it is > made a bit tricky by the refcounts being outside the objects. > > Basically what I'm trying to protect against is the case where the > global refcount goes to zero and then immediately back to one (or vice > versa), which results in a deallocation immediately followed by an > allocation. I was worried about the allocation and deallocation racing > with each other (because there's a window after the refcnt is checked > before the (de)allocation begins). It's not quite clear to me whether > RCU is enough to protect against this... Thoughts? Hm. I think you'll still need a lock (mutex?) on the alloc path, but the free path should be fine as long as you load the map pointer before looking at the refcnt (atomic op ensuring the barrier there). > > Could you please add tests for the replacement combinations? > > Sure. But, erm, where? selftests? :) Yes :) selftests/bpf/ Look around at the scripts, AFAIK we don't have much in the way of standard infra for system interaction tests like this :S
Jakub Kicinski <jakub.kicinski@netronome.com> writes: > On Mon, 04 Mar 2019 12:58:51 +0100, Toke Høiland-Jørgensen wrote: >> >> diff --git a/include/net/netns/xdp.h b/include/net/netns/xdp.h >> >> index e5734261ba0a..4935dfe1cf43 100644 >> >> --- a/include/net/netns/xdp.h >> >> +++ b/include/net/netns/xdp.h >> >> @@ -4,10 +4,21 @@ >> >> >> >> #include <linux/rculist.h> >> >> #include <linux/mutex.h> >> >> +#include <linux/atomic.h> >> >> + >> >> +struct bpf_dtab; >> >> + >> >> +struct bpf_dtab_container { >> >> + struct bpf_dtab __rcu *dtab; >> >> + atomic_t refcnt; >> > >> > refcount_t ? I'm not sure what the rules are for non-performance >> > critical stuff are.. >> >> I based the use of atomic_t on what bpf/syscall.c is doing. The problem >> with using refcount_t is that it's a bit of a weird refcount because >> it's not strictly tied to the object lifetime (as the objects are only >> allocated if *both* the globals and the per-namespace refcnt are >0). >> >> >> +static void __dev_map_release_default_map(struct bpf_dtab_container *cont) >> >> +{ >> >> + struct bpf_dtab *dtab = NULL; >> >> + >> >> + lockdep_assert_held(&dev_map_lock); >> >> + >> >> + dtab = rcu_dereference(cont->dtab); >> >> + if (dtab) { >> > >> > How can it be NULL if it's refcounted? hm.. >> >> See above; it's not actually the object that is refcounted, it's a count >> of the number of active XDP programs in the namespace, which still needs >> to be kept so we can allocate the devmaps when a REDIRECT program is >> loaded. > > I think I got it now. I wonder if anything could be done to improve > the readability of the dual-refcnting :S Maybe use the "needed"/"used" > terms throughout? Yeah, good idea, I'll try something along those lines. >> >> + list_del_rcu(&dtab->list); >> >> + rcu_assign_pointer(cont->dtab, NULL); >> >> + bpf_clear_redirect_map(&dtab->map); >> >> + call_rcu(&dtab->rcu, __dev_map_free); >> >> + } >> >> +} >> >> + >> >> +void dev_map_put_default_map(struct net *net) >> >> +{ >> >> + if (atomic_dec_and_test(&net->xdp.default_map.refcnt)) { >> >> + spin_lock(&dev_map_lock); >> >> + __dev_map_release_default_map(&net->xdp.default_map); >> >> + spin_unlock(&dev_map_lock); >> >> + } >> >> +} >> >> + >> >> +static int __init_default_map(struct bpf_dtab_container *cont) >> >> +{ >> >> + struct net *net = bpf_default_map_to_net(cont); >> >> + struct bpf_dtab *dtab, *old_dtab; >> >> + int size = DEV_MAP_DEFAULT_SIZE; >> >> + struct net_device *netdev; >> >> + union bpf_attr attr = {}; >> >> + u32 idx; >> >> + int err; >> >> + >> >> + lockdep_assert_held(&dev_map_lock); >> >> + >> >> + if (!atomic_read(&global_redirect_use)) >> >> + return 0; >> >> + >> >> + for_each_netdev(net, netdev) >> >> + if (netdev->ifindex >= size) >> >> + size <<= 1; >> >> + >> >> + old_dtab = rcu_dereference(cont->dtab); >> >> + if (old_dtab && old_dtab->map.max_entries == size) >> >> + return 0; >> >> + >> >> + dtab = kzalloc(sizeof(*dtab), GFP_USER); >> > >> > You are calling this with a spin lock held :( >> >> Yeah, not just this, there are more allocations in dev_map_init_map(). >> Hmm, I guess I need to think about the concurrency bits some more; it is >> made a bit tricky by the refcounts being outside the objects. >> >> Basically what I'm trying to protect against is the case where the >> global refcount goes to zero and then immediately back to one (or vice >> versa), which results in a deallocation immediately followed by an >> allocation. I was worried about the allocation and deallocation racing >> with each other (because there's a window after the refcnt is checked >> before the (de)allocation begins). It's not quite clear to me whether >> RCU is enough to protect against this... Thoughts? > > Hm. I think you'll still need a lock (mutex?) on the alloc path, but > the free path should be fine as long as you load the map pointer before > looking at the refcnt (atomic op ensuring the barrier there). Yeah, for the per-namespace refcnt it's pretty straight forward, the trouble is the global count that needs to iterate over all namespaces; probably need to put that all behind a (non-spin)lock, right? >> > Could you please add tests for the replacement combinations? >> >> Sure. But, erm, where? selftests? :) > > Yes :) selftests/bpf/ Look around at the scripts, AFAIK we don't have > much in the way of standard infra for system interaction tests like > this :S Right. I'll shamelessly copy from the tests already in there and see if I can't get something decent working :) -Toke
On Mon, 04 Mar 2019 20:05:30 +0100, Toke Høiland-Jørgensen wrote: > > Hm. I think you'll still need a lock (mutex?) on the alloc path, but > > the free path should be fine as long as you load the map pointer before > > looking at the refcnt (atomic op ensuring the barrier there). > > Yeah, for the per-namespace refcnt it's pretty straight forward, the > trouble is the global count that needs to iterate over all namespaces; > probably need to put that all behind a (non-spin)lock, right? Because net iteration is under RCU? You can switch to taking net_rwsem for that one, no? I'm probably confused again ;)
Jakub Kicinski <jakub.kicinski@netronome.com> writes: > On Mon, 04 Mar 2019 20:05:30 +0100, Toke Høiland-Jørgensen wrote: >> > Hm. I think you'll still need a lock (mutex?) on the alloc path, but >> > the free path should be fine as long as you load the map pointer before >> > looking at the refcnt (atomic op ensuring the barrier there). >> >> Yeah, for the per-namespace refcnt it's pretty straight forward, the >> trouble is the global count that needs to iterate over all namespaces; >> probably need to put that all behind a (non-spin)lock, right? > > Because net iteration is under RCU? You can switch to taking net_rwsem > for that one, no? I'm probably confused again ;) Because there's a single refcount that needs to trigger creation/deletion of *all* the default maps. I.e. if (atomic_dec_return(&global_refcnt)) for_each_namespace(net) destroy_default_map(net); which needs to not step on the toes of a subsequent if (atomic_inc_return(&global_refcnt) == 1) for_each_namespace(net) create_default_map(net); (or vice versa, of course). Not sure there's a way to do that without wrapping both of those constructs (including the refcnt inc/dec) in a mutex? -Toke
On Mon, 04 Mar 2019 23:28:14 +0100, Toke Høiland-Jørgensen wrote: > Jakub Kicinski <jakub.kicinski@netronome.com> writes: > > On Mon, 04 Mar 2019 20:05:30 +0100, Toke Høiland-Jørgensen wrote: > >> > Hm. I think you'll still need a lock (mutex?) on the alloc path, but > >> > the free path should be fine as long as you load the map pointer before > >> > looking at the refcnt (atomic op ensuring the barrier there). > >> > >> Yeah, for the per-namespace refcnt it's pretty straight forward, the > >> trouble is the global count that needs to iterate over all namespaces; > >> probably need to put that all behind a (non-spin)lock, right? > > > > Because net iteration is under RCU? You can switch to taking net_rwsem > > for that one, no? I'm probably confused again ;) > > Because there's a single refcount that needs to trigger > creation/deletion of *all* the default maps. I.e. > > if (atomic_dec_return(&global_refcnt)) > for_each_namespace(net) > destroy_default_map(net); > > which needs to not step on the toes of a subsequent > > if (atomic_inc_return(&global_refcnt) == 1) > for_each_namespace(net) > create_default_map(net); > > (or vice versa, of course). > > Not sure there's a way to do that without wrapping both of those > constructs (including the refcnt inc/dec) in a mutex? Seems like it :(
Jakub Kicinski <jakub.kicinski@netronome.com> writes: > On Mon, 04 Mar 2019 23:28:14 +0100, Toke Høiland-Jørgensen wrote: >> Jakub Kicinski <jakub.kicinski@netronome.com> writes: >> > On Mon, 04 Mar 2019 20:05:30 +0100, Toke Høiland-Jørgensen wrote: >> >> > Hm. I think you'll still need a lock (mutex?) on the alloc path, but >> >> > the free path should be fine as long as you load the map pointer before >> >> > looking at the refcnt (atomic op ensuring the barrier there). >> >> >> >> Yeah, for the per-namespace refcnt it's pretty straight forward, the >> >> trouble is the global count that needs to iterate over all namespaces; >> >> probably need to put that all behind a (non-spin)lock, right? >> > >> > Because net iteration is under RCU? You can switch to taking net_rwsem >> > for that one, no? I'm probably confused again ;) >> >> Because there's a single refcount that needs to trigger >> creation/deletion of *all* the default maps. I.e. >> >> if (atomic_dec_return(&global_refcnt)) >> for_each_namespace(net) >> destroy_default_map(net); >> >> which needs to not step on the toes of a subsequent >> >> if (atomic_inc_return(&global_refcnt) == 1) >> for_each_namespace(net) >> create_default_map(net); >> >> (or vice versa, of course). >> >> Not sure there's a way to do that without wrapping both of those >> constructs (including the refcnt inc/dec) in a mutex? > > Seems like it :( Right. Well I'm glad that we at least agree, so I can stop racking my brain for other ways to do it ;) -Toke
diff --git a/include/linux/bpf.h b/include/linux/bpf.h index de18227b3d95..060c5850af3b 100644 --- a/include/linux/bpf.h +++ b/include/linux/bpf.h @@ -25,6 +25,7 @@ struct sock; struct seq_file; struct btf; struct btf_type; +struct net; /* map is generic key/value storage optionally accesible by eBPF programs */ struct bpf_map_ops { @@ -533,6 +534,7 @@ extern const struct bpf_verifier_ops tc_cls_act_analyzer_ops; extern const struct bpf_verifier_ops xdp_analyzer_ops; struct bpf_prog *bpf_prog_get(u32 ufd); +struct bpf_prog *bpf_prog_get_by_id(u32 id); struct bpf_prog *bpf_prog_get_type_dev(u32 ufd, enum bpf_prog_type type, bool attach_drv); struct bpf_prog * __must_check bpf_prog_add(struct bpf_prog *prog, int i); @@ -612,6 +614,11 @@ struct xdp_buff; struct sk_buff; struct bpf_dtab_netdev *__dev_map_lookup_elem(struct bpf_map *map, u32 key); +struct bpf_map *__dev_map_get_default_map(struct net_device *dev); +int dev_map_ensure_default_map(struct net *net); +void dev_map_put_default_map(struct net *net); +int dev_map_inc_redirect_count(void); +void dev_map_dec_redirect_count(void); void __dev_map_insert_ctx(struct bpf_map *map, u32 index); void __dev_map_flush(struct bpf_map *map); int dev_map_enqueue(struct bpf_dtab_netdev *dst, struct xdp_buff *xdp, @@ -641,6 +648,11 @@ static inline struct bpf_prog *bpf_prog_get(u32 ufd) return ERR_PTR(-EOPNOTSUPP); } +static inline struct bpf_prog *bpf_prog_get_by_id(u32 id) +{ + return ERR_PTR(-EOPNOTSUPP); +} + static inline struct bpf_prog *bpf_prog_get_type_dev(u32 ufd, enum bpf_prog_type type, bool attach_drv) @@ -693,6 +705,29 @@ static inline struct net_device *__dev_map_lookup_elem(struct bpf_map *map, return NULL; } +static inline struct bpf_map *__dev_map_get_default_map(struct net_device *dev) +{ + return NULL; +} + +static inline int dev_map_ensure_default_map(struct net *net) +{ + return 0; +} + +static inline void dev_map_put_default_map(struct net *net) +{ +} + +static inline int dev_map_inc_redirect_count(void) +{ + return 0; +} + +static inline void dev_map_dec_redirect_count(void) +{ +} + static inline void __dev_map_insert_ctx(struct bpf_map *map, u32 index) { } diff --git a/include/linux/filter.h b/include/linux/filter.h index 95e2d7ebdf21..dd6bbbab32f7 100644 --- a/include/linux/filter.h +++ b/include/linux/filter.h @@ -507,6 +507,8 @@ struct bpf_prog { gpl_compatible:1, /* Is filter GPL compatible? */ cb_access:1, /* Is control block accessed? */ dst_needed:1, /* Do we need dst entry? */ + redirect_needed:1, /* Does program need access to xdp_redirect? */ + redirect_used:1, /* Does program use xdp_redirect? */ blinded:1, /* Was blinded */ is_func:1, /* program is a bpf function */ kprobe_override:1, /* Do we override a kprobe? */ diff --git a/include/net/net_namespace.h b/include/net/net_namespace.h index a68ced28d8f4..6706ecc25d8f 100644 --- a/include/net/net_namespace.h +++ b/include/net/net_namespace.h @@ -162,7 +162,7 @@ struct net { #if IS_ENABLED(CONFIG_CAN) struct netns_can can; #endif -#ifdef CONFIG_XDP_SOCKETS +#ifdef CONFIG_BPF_SYSCALL struct netns_xdp xdp; #endif struct sock *diag_nlsk; diff --git a/include/net/netns/xdp.h b/include/net/netns/xdp.h index e5734261ba0a..4935dfe1cf43 100644 --- a/include/net/netns/xdp.h +++ b/include/net/netns/xdp.h @@ -4,10 +4,21 @@ #include <linux/rculist.h> #include <linux/mutex.h> +#include <linux/atomic.h> + +struct bpf_dtab; + +struct bpf_dtab_container { + struct bpf_dtab __rcu *dtab; + atomic_t refcnt; +}; struct netns_xdp { +#ifdef CONFIG_XDP_SOCKETS struct mutex lock; struct hlist_head list; +#endif + struct bpf_dtab_container default_map; }; #endif /* __NETNS_XDP_H__ */ diff --git a/kernel/bpf/devmap.c b/kernel/bpf/devmap.c index 1037fc08c504..e55707e62b60 100644 --- a/kernel/bpf/devmap.c +++ b/kernel/bpf/devmap.c @@ -56,6 +56,9 @@ (BPF_F_NUMA_NODE | BPF_F_RDONLY | BPF_F_WRONLY) #define DEV_MAP_BULK_SIZE 16 +#define DEV_MAP_DEFAULT_SIZE 8 +#define BPF_MAX_REFCNT 32768 + struct xdp_bulk_queue { struct xdp_frame *q[DEV_MAP_BULK_SIZE]; struct net_device *dev_rx; @@ -80,6 +83,7 @@ struct bpf_dtab { static DEFINE_SPINLOCK(dev_map_lock); static LIST_HEAD(dev_map_list); +static atomic_t global_redirect_use = {}; static u64 dev_map_bitmap_size(const union bpf_attr *attr) { @@ -332,6 +336,18 @@ struct bpf_dtab_netdev *__dev_map_lookup_elem(struct bpf_map *map, u32 key) return obj; } +/* This is only being called from xdp_do_redirect() if the xdp_redirect helper + * is used; the default map is allocated on XDP program load if the helper is + * used, so will always be available at this point. + */ +struct bpf_map *__dev_map_get_default_map(struct net_device *dev) +{ + struct net *net = dev_net(dev); + struct bpf_dtab *dtab = rcu_dereference(net->xdp.default_map.dtab); + + return dtab ? &dtab->map : NULL; +} + /* Runs under RCU-read-side, plus in softirq under NAPI protection. * Thus, safe percpu variable access. */ @@ -533,14 +549,210 @@ const struct bpf_map_ops dev_map_ops = { .map_check_btf = map_check_no_btf, }; +static inline struct net *bpf_default_map_to_net(struct bpf_dtab_container *cont) +{ + struct netns_xdp *xdp = container_of(cont, struct netns_xdp, default_map); + + return container_of(xdp, struct net, xdp); +} + +static void __dev_map_release_default_map(struct bpf_dtab_container *cont) +{ + struct bpf_dtab *dtab = NULL; + + lockdep_assert_held(&dev_map_lock); + + dtab = rcu_dereference(cont->dtab); + if (dtab) { + list_del_rcu(&dtab->list); + rcu_assign_pointer(cont->dtab, NULL); + bpf_clear_redirect_map(&dtab->map); + call_rcu(&dtab->rcu, __dev_map_free); + } +} + +void dev_map_put_default_map(struct net *net) +{ + if (atomic_dec_and_test(&net->xdp.default_map.refcnt)) { + spin_lock(&dev_map_lock); + __dev_map_release_default_map(&net->xdp.default_map); + spin_unlock(&dev_map_lock); + } +} + +static int __init_default_map(struct bpf_dtab_container *cont) +{ + struct net *net = bpf_default_map_to_net(cont); + struct bpf_dtab *dtab, *old_dtab; + int size = DEV_MAP_DEFAULT_SIZE; + struct net_device *netdev; + union bpf_attr attr = {}; + u32 idx; + int err; + + lockdep_assert_held(&dev_map_lock); + + if (!atomic_read(&global_redirect_use)) + return 0; + + for_each_netdev(net, netdev) + if (netdev->ifindex >= size) + size <<= 1; + + old_dtab = rcu_dereference(cont->dtab); + if (old_dtab && old_dtab->map.max_entries == size) + return 0; + + dtab = kzalloc(sizeof(*dtab), GFP_USER); + if (!dtab) + return -ENOMEM; + + attr.map_type = BPF_MAP_TYPE_DEVMAP; + attr.max_entries = size; + attr.value_size = 4; + attr.key_size = 4; + + err = dev_map_init_map(dtab, &attr, false); + if (err) { + kfree(dtab); + return err; + } + + for_each_netdev(net, netdev) { + idx = netdev->ifindex; + err = __dev_map_update_elem(net, &dtab->map, &idx, &idx, 0); + if (err) { + __dev_map_free(&dtab->rcu); + return err; + } + } + + rcu_assign_pointer(cont->dtab, dtab); + list_add_tail_rcu(&dtab->list, &dev_map_list); + + if (old_dtab) { + list_del_rcu(&old_dtab->list); + bpf_clear_redirect_map(&old_dtab->map); + call_rcu(&old_dtab->rcu, __dev_map_free); + } + + return 0; +} + +static int maybe_inc_refcnt(atomic_t *v) +{ + int refcnt; + + refcnt = atomic_inc_return(v); + if (refcnt > BPF_MAX_REFCNT) { + atomic_dec(v); + return -EBUSY; + } + + return refcnt; +} + +int dev_map_ensure_default_map(struct net *net) +{ + int refcnt, err = 0; + + refcnt = maybe_inc_refcnt(&net->xdp.default_map.refcnt); + if (refcnt < 0) + return refcnt; + + if (refcnt == 1) { + spin_lock(&dev_map_lock); + err = __init_default_map(&net->xdp.default_map); + spin_unlock(&dev_map_lock); + } + + return err; +} + +static void __dev_map_dec_redirect_count(void) +{ + struct net *net; + + lockdep_assert_held(&dev_map_lock); + + if (atomic_dec_and_test(&global_redirect_use)) + for_each_net_rcu(net) + __dev_map_release_default_map(&net->xdp.default_map); +} + +void dev_map_dec_redirect_count(void) +{ + spin_lock(&dev_map_lock); + __dev_map_dec_redirect_count(); + spin_unlock(&dev_map_lock); +} + +static int __dev_map_init_redirect_use(void) +{ + struct net *net; + int err; + + lockdep_assert_held(&dev_map_lock); + + for_each_net_rcu(net) { + if (atomic_read(&net->xdp.default_map.refcnt)) { + err = __init_default_map(&net->xdp.default_map); + if (err) + return err; + } + } + + return 0; +} + +int dev_map_inc_redirect_count(void) +{ + int refcnt, err = 0; + + spin_lock(&dev_map_lock); + refcnt = maybe_inc_refcnt(&global_redirect_use); + if (refcnt < 0) { + err = refcnt; + goto out; + } + + if (refcnt == 1) + err = __dev_map_init_redirect_use(); + + if (err) + __dev_map_dec_redirect_count(); + + out: + spin_unlock(&dev_map_lock); + return err; +} + static int dev_map_notification(struct notifier_block *notifier, ulong event, void *ptr) { struct net_device *netdev = netdev_notifier_info_to_dev(ptr); + struct net *net = dev_net(netdev); + u32 idx = netdev->ifindex; struct bpf_dtab *dtab; - int i; + int i, err; switch (event) { + case NETDEV_REGISTER: + rcu_read_lock(); + dtab = rcu_dereference(net->xdp.default_map.dtab); + if (dtab) { + err = __dev_map_update_elem(net, &dtab->map, + &idx, &idx, 0); + if (err == -E2BIG) { + spin_lock(&dev_map_lock); + err = __init_default_map(&net->xdp.default_map); + if (err) + net_warn_ratelimited("Unable to re-allocate default map, xdp_redirect() may fail on some ifindexes\n"); + spin_unlock(&dev_map_lock); + } + } + rcu_read_unlock(); + break; case NETDEV_UNREGISTER: /* This rcu_read_lock/unlock pair is needed because * dev_map_list is an RCU list AND to ensure a delete diff --git a/kernel/bpf/syscall.c b/kernel/bpf/syscall.c index ec7c552af76b..fd1b76f5da2d 100644 --- a/kernel/bpf/syscall.c +++ b/kernel/bpf/syscall.c @@ -1265,6 +1265,9 @@ static void __bpf_prog_put(struct bpf_prog *prog, bool do_idr_lock) kvfree(prog->aux->func_info); bpf_prog_free_linfo(prog); + if (prog->redirect_used) + dev_map_dec_redirect_count(); + call_rcu(&prog->aux->rcu, __bpf_prog_put_rcu); } } @@ -1962,6 +1965,21 @@ static int bpf_obj_get_next_id(const union bpf_attr *attr, #define BPF_PROG_GET_FD_BY_ID_LAST_FIELD prog_id +struct bpf_prog *bpf_prog_get_by_id(u32 id) +{ + struct bpf_prog *prog; + + spin_lock_bh(&prog_idr_lock); + prog = idr_find(&prog_idr, id); + if (prog) + prog = bpf_prog_inc_not_zero(prog); + else + prog = ERR_PTR(-ENOENT); + spin_unlock_bh(&prog_idr_lock); + + return prog; +} + static int bpf_prog_get_fd_by_id(const union bpf_attr *attr) { struct bpf_prog *prog; @@ -1974,14 +1992,7 @@ static int bpf_prog_get_fd_by_id(const union bpf_attr *attr) if (!capable(CAP_SYS_ADMIN)) return -EPERM; - spin_lock_bh(&prog_idr_lock); - prog = idr_find(&prog_idr, id); - if (prog) - prog = bpf_prog_inc_not_zero(prog); - else - prog = ERR_PTR(-ENOENT); - spin_unlock_bh(&prog_idr_lock); - + prog = bpf_prog_get_by_id(id); if (IS_ERR(prog)) return PTR_ERR(prog); diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c index 1b9496c41383..f1b2f01e7ca1 100644 --- a/kernel/bpf/verifier.c +++ b/kernel/bpf/verifier.c @@ -7609,6 +7609,17 @@ static int fixup_bpf_calls(struct bpf_verifier_env *env) prog->dst_needed = 1; if (insn->imm == BPF_FUNC_get_prandom_u32) bpf_user_rnd_init_once(); + if (insn->imm == BPF_FUNC_redirect) { + prog->redirect_needed = true; + if (!prog->redirect_used) { + int err = dev_map_inc_redirect_count(); + + if (err) + return err; + prog->redirect_used = true; + } + } + if (insn->imm == BPF_FUNC_override_return) prog->kprobe_override = 1; if (insn->imm == BPF_FUNC_tail_call) { @@ -7618,6 +7629,7 @@ static int fixup_bpf_calls(struct bpf_verifier_env *env) * the program array. */ prog->cb_access = 1; + prog->redirect_needed = true; env->prog->aux->stack_depth = MAX_BPF_STACK; env->prog->aux->max_pkt_offset = MAX_PACKET_OFF; diff --git a/net/core/dev.c b/net/core/dev.c index 2b67f2aa59dd..1df20d529026 100644 --- a/net/core/dev.c +++ b/net/core/dev.c @@ -7990,6 +7990,21 @@ u32 __dev_xdp_query(struct net_device *dev, bpf_op_t bpf_op, return xdp.prog_id; } +static struct bpf_prog *dev_xdp_get_prog(struct net_device *dev) +{ + struct bpf_prog *prog; + u32 prog_id; + + prog_id = __dev_xdp_query(dev, dev->netdev_ops->ndo_bpf, XDP_QUERY_PROG); + if (prog_id) { + prog = bpf_prog_get_by_id(prog_id); + if (!IS_ERR(prog)) + return prog; + } + + return NULL; +} + static int dev_xdp_install(struct net_device *dev, bpf_op_t bpf_op, struct netlink_ext_ack *extack, u32 flags, struct bpf_prog *prog) @@ -8024,9 +8039,18 @@ static void dev_xdp_uninstall(struct net_device *dev) memset(&xdp, 0, sizeof(xdp)); xdp.command = XDP_QUERY_PROG; WARN_ON(ndo_bpf(dev, &xdp)); - if (xdp.prog_id) + if (xdp.prog_id) { + struct bpf_prog *prog = bpf_prog_get_by_id(xdp.prog_id); + + if (!IS_ERR(prog)) { + if (prog->redirect_needed) + dev_map_put_default_map(dev_net(dev)); + bpf_prog_put(prog); + } + WARN_ON(dev_xdp_install(dev, ndo_bpf, NULL, xdp.prog_flags, NULL)); + } /* Remove HW offload */ memset(&xdp, 0, sizeof(xdp)); @@ -8091,6 +8115,23 @@ int dev_change_xdp_fd(struct net_device *dev, struct netlink_ext_ack *extack, bpf_prog_put(prog); return -EINVAL; } + + if (!offload && bpf_op == ops->ndo_bpf && + prog->redirect_needed) { + err = dev_map_ensure_default_map(dev_net(dev)); + if (err) { + NL_SET_ERR_MSG(extack, "unable to allocate default map for xdp_redirect()"); + return err; + } + } + } else { + struct bpf_prog *old_prog = dev_xdp_get_prog(dev); + + if (old_prog) { + if (old_prog->redirect_needed) + dev_map_put_default_map(dev_net(dev)); + bpf_prog_put(old_prog); + } } err = dev_xdp_install(dev, bpf_op, extack, flags, prog); @@ -9333,6 +9374,7 @@ EXPORT_SYMBOL(unregister_netdev); int dev_change_net_namespace(struct net_device *dev, struct net *net, const char *pat) { int err, new_nsid, new_ifindex; + struct bpf_prog *prog = NULL; ASSERT_RTNL(); @@ -9350,6 +9392,15 @@ int dev_change_net_namespace(struct net_device *dev, struct net *net, const char if (net_eq(dev_net(dev), net)) goto out; + prog = dev_xdp_get_prog(dev); + if (prog) { + if (prog->redirect_needed) + err = dev_map_ensure_default_map(net); + + if (err) + goto out; + } + /* Pick the destination device name, and ensure * we can use it in the destination network namespace. */ @@ -9388,6 +9439,9 @@ int dev_change_net_namespace(struct net_device *dev, struct net *net, const char call_netdevice_notifiers(NETDEV_UNREGISTER, dev); rcu_barrier(); + if (prog && prog->redirect_needed) + dev_map_put_default_map(dev_net(dev)); + new_nsid = peernet2id_alloc(dev_net(dev), net); /* If there is an ifindex conflict assign a new one */ if (__dev_get_by_index(net, dev->ifindex)) @@ -9435,6 +9489,9 @@ int dev_change_net_namespace(struct net_device *dev, struct net *net, const char synchronize_net(); err = 0; out: + if (prog) + bpf_prog_put(prog); + return err; } EXPORT_SYMBOL_GPL(dev_change_net_namespace); diff --git a/net/core/filter.c b/net/core/filter.c index 5132c054c981..be02ea103d05 100644 --- a/net/core/filter.c +++ b/net/core/filter.c @@ -3337,58 +3337,6 @@ static const struct bpf_func_proto bpf_xdp_adjust_meta_proto = { .arg2_type = ARG_ANYTHING, }; -static int __bpf_tx_xdp(struct net_device *dev, - struct bpf_map *map, - struct xdp_buff *xdp, - u32 index) -{ - struct xdp_frame *xdpf; - int err, sent; - - if (!dev->netdev_ops->ndo_xdp_xmit) { - return -EOPNOTSUPP; - } - - err = xdp_ok_fwd_dev(dev, xdp->data_end - xdp->data); - if (unlikely(err)) - return err; - - xdpf = convert_to_xdp_frame(xdp); - if (unlikely(!xdpf)) - return -EOVERFLOW; - - sent = dev->netdev_ops->ndo_xdp_xmit(dev, 1, &xdpf, XDP_XMIT_FLUSH); - if (sent <= 0) - return sent; - return 0; -} - -static noinline int -xdp_do_redirect_slow(struct net_device *dev, struct xdp_buff *xdp, - struct bpf_prog *xdp_prog, struct bpf_redirect_info *ri) -{ - struct net_device *fwd; - u32 index = ri->ifindex; - int err; - - fwd = dev_get_by_index_rcu(dev_net(dev), index); - ri->ifindex = 0; - if (unlikely(!fwd)) { - err = -EINVAL; - goto err; - } - - err = __bpf_tx_xdp(fwd, NULL, xdp, 0); - if (unlikely(err)) - goto err; - - _trace_xdp_redirect(dev, xdp_prog, index); - return 0; -err: - _trace_xdp_redirect_err(dev, xdp_prog, index, err); - return err; -} - static int __bpf_tx_xdp_map(struct net_device *dev_rx, void *fwd, struct bpf_map *map, struct xdp_buff *xdp, @@ -3519,10 +3467,10 @@ int xdp_do_redirect(struct net_device *dev, struct xdp_buff *xdp, struct bpf_redirect_info *ri = this_cpu_ptr(&bpf_redirect_info); struct bpf_map *map = READ_ONCE(ri->map); - if (likely(map)) - return xdp_do_redirect_map(dev, xdp, xdp_prog, map, ri); + if (unlikely(!map)) + map = __dev_map_get_default_map(dev); - return xdp_do_redirect_slow(dev, xdp, xdp_prog, ri); + return xdp_do_redirect_map(dev, xdp, xdp_prog, map, ri); } EXPORT_SYMBOL_GPL(xdp_do_redirect);
An XDP program can redirect packets between interfaces using either the xdp_redirect() helper or the xdp_redirect_map() helper. Apart from the flexibility of updating maps from userspace, the redirect_map() helper also uses the map structure to batch packets, which results in a significant (around 50%) performance boost. However, the xdp_redirect() API is simpler if one just wants to redirect to another interface, which means people tend to use this interface and then wonder why they getter worse performance than expected. This patch seeks to close this performance difference between the two APIs. It achieves this by changing xdp_redirect() to use a hidden devmap for looking up destination interfaces, thus gaining the batching benefit with no visible difference from the user API point of view. A hidden per-namespace map is allocated when an XDP program that uses the non-map xdp_redirect() helper is first loaded. This map is populated with all available interfaces in its namespace, and kept up to date as interfaces come and go. Once allocated, the map is kept around until the namespace is removed. The hidden map uses the ifindex as map key, which means they are limited to ifindexes smaller than the map size of 64. A later patch introduces a new map type to lift this restriction. Performance numbers: Before patch: xdp_redirect: 5426035 pkt/s xdp_redirect_map: 8412754 pkt/s After patch: xdp_redirect: 8314702 pkt/s xdp_redirect_map: 8411854 pkt/s This corresponds to a 53% increase in xdp_redirect performance, or a reduction in per-packet processing time by 64 nanoseconds. Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com> --- include/linux/bpf.h | 35 +++++++ include/linux/filter.h | 2 include/net/net_namespace.h | 2 include/net/netns/xdp.h | 11 ++ kernel/bpf/devmap.c | 214 +++++++++++++++++++++++++++++++++++++++++++ kernel/bpf/syscall.c | 27 ++++- kernel/bpf/verifier.c | 12 ++ net/core/dev.c | 59 ++++++++++++ net/core/filter.c | 58 +----------- 9 files changed, 354 insertions(+), 66 deletions(-)