Message ID | 20241023123212.15908-1-laoar.shao@gmail.com (mailing list archive) |
---|---|
State | Handled Elsewhere |
Headers | show |
Series | net: Add tcp_drop_reason tracepoint | expand |
On Wed, Oct 23, 2024 at 8:33 PM Yafang Shao <laoar.shao@gmail.com> wrote: > > We previously hooked the tcp_drop_reason() function using BPF to monitor > TCP drop reasons. However, after upgrading our compiler from GCC 9 to GCC > 11, tcp_drop_reason() is now inlined, preventing us from hooking into it. > To address this, it would be beneficial to introduce a dedicated tracepoint > for monitoring. Hello, Can the existing tracepoint kfree_skb do this work? AFAIK, you can attach you BPF to the kfree_skb tracepoint and do some filter according to the "protocol" field, or the information "sk" field. And this works fine in my tool. I hope I'm not missing something :/ BTW, I do such filter in probe_parse_skb_sk() in https://github.com/OpenCloudOS/nettrace/blob/master/shared/bpf/skb_parse.h Thanks! Menglong Dong > > Signed-off-by: Yafang Shao <laoar.shao@gmail.com> > Cc: Menglong Dong <menglong8.dong@gmail.com> > --- > include/trace/events/tcp.h | 53 ++++++++++++++++++++++++++++++++++++++ > net/ipv4/tcp_input.c | 1 + > 2 files changed, 54 insertions(+) > > diff --git a/include/trace/events/tcp.h b/include/trace/events/tcp.h > index a27c4b619dff..056f7026224c 100644 > --- a/include/trace/events/tcp.h > +++ b/include/trace/events/tcp.h > @@ -12,6 +12,7 @@ > #include <net/tcp.h> > #include <linux/sock_diag.h> > #include <net/rstreason.h> > +#include <net/dropreason-core.h> > > /* > * tcp event with arguments sk and skb > @@ -728,6 +729,58 @@ DEFINE_EVENT(tcp_ao_event_sne, tcp_ao_rcv_sne_update, > TP_ARGS(sk, new_sne) > ); > > +#undef FN > +#undef FNe > +#define FN(reason) { SKB_DROP_REASON_##reason, #reason }, > +#define FNe(reason) { SKB_DROP_REASON_##reason, #reason } > + > +TRACE_EVENT(tcp_drop_reason, > + > + TP_PROTO(const struct sock *sk, const struct sk_buff *skb, > + const enum skb_drop_reason reason), > + > + TP_ARGS(sk, skb, reason), > + > + TP_STRUCT__entry( > + __field(const void *, skbaddr) > + __field(const void *, skaddr) > + __field(int, state) > + __field(enum skb_drop_reason, reason) > + __array(__u8, saddr, sizeof(struct sockaddr_in6)) > + __array(__u8, daddr, sizeof(struct sockaddr_in6)) > + ), > + > + TP_fast_assign( > + __entry->skbaddr = skb; > + __entry->skaddr = sk; > + /* Zero means unknown state. */ > + __entry->state = sk ? sk->sk_state : 0; > + > + memset(__entry->saddr, 0, sizeof(struct sockaddr_in6)); > + memset(__entry->daddr, 0, sizeof(struct sockaddr_in6)); > + > + if (sk && sk_fullsock(sk)) { > + const struct inet_sock *inet = inet_sk(sk); > + > + TP_STORE_ADDR_PORTS(__entry, inet, sk); > + } else { > + const struct tcphdr *th = (const struct tcphdr *)skb->data; > + > + TP_STORE_ADDR_PORTS_SKB(skb, th, entry->saddr, entry->daddr); > + } > + __entry->reason = reason; > + ), > + > + TP_printk("skbaddr=%p skaddr=%p src=%pISpc dest=%pISpc state=%s reason=%s", > + __entry->skbaddr, __entry->skaddr, > + __entry->saddr, __entry->daddr, > + __entry->state ? show_tcp_state_name(__entry->state) : "UNKNOWN", > + __print_symbolic(__entry->reason, DEFINE_DROP_REASON(FN, FNe))) > +); > + > +#undef FN > +#undef FNe > + > #endif /* _TRACE_TCP_H */ > > /* This part must be outside protection */ > diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c > index cc05ec1faac8..44795555596a 100644 > --- a/net/ipv4/tcp_input.c > +++ b/net/ipv4/tcp_input.c > @@ -4897,6 +4897,7 @@ static bool tcp_ooo_try_coalesce(struct sock *sk, > static void tcp_drop_reason(struct sock *sk, struct sk_buff *skb, > enum skb_drop_reason reason) > { > + trace_tcp_drop_reason(sk, skb, reason); > sk_drops_add(sk, skb); > sk_skb_reason_drop(sk, skb, reason); > } > -- > 2.43.5 >
On Wed, Oct 23, 2024 at 2:33 PM Yafang Shao <laoar.shao@gmail.com> wrote: > > We previously hooked the tcp_drop_reason() function using BPF to monitor > TCP drop reasons. However, after upgrading our compiler from GCC 9 to GCC > 11, tcp_drop_reason() is now inlined, preventing us from hooking into it. > To address this, it would be beneficial to introduce a dedicated tracepoint > for monitoring. This patch would require changes in user space tracers. I am surprised no one came up with a noinline variant. __bpf_kfunc is using #define __bpf_kfunc __used __retain noinline I would rather not have include/trace/events/tcp.h becoming the biggest file in TCP stack...
On Wed, Oct 23, 2024 at 8:59 PM Menglong Dong <menglong8.dong@gmail.com> wrote: > > On Wed, Oct 23, 2024 at 8:33 PM Yafang Shao <laoar.shao@gmail.com> wrote: > > > > We previously hooked the tcp_drop_reason() function using BPF to monitor > > TCP drop reasons. However, after upgrading our compiler from GCC 9 to GCC > > 11, tcp_drop_reason() is now inlined, preventing us from hooking into it. > > To address this, it would be beneficial to introduce a dedicated tracepoint > > for monitoring. > > Hello, > > Can the existing tracepoint kfree_skb do this work? AFAIK, you > can attach you BPF to the kfree_skb tracepoint and do some filter > according to the "protocol" field, or the information "sk" field. And > this works fine in my tool. > > I hope I'm not missing something :/ > > BTW, I do such filter in probe_parse_skb_sk() in > https://github.com/OpenCloudOS/nettrace/blob/master/shared/bpf/skb_parse.h We prefer not to hook the kfree_skb tracepoint, as we want to avoid the overhead of parsing extensive information from @skb. Since we now have a function that can be easily hooked, why not hook it directly? -- Regards Yafang
On Wed, Oct 23, 2024 at 9:01 PM Eric Dumazet <edumazet@google.com> wrote: > > On Wed, Oct 23, 2024 at 2:33 PM Yafang Shao <laoar.shao@gmail.com> wrote: > > > > We previously hooked the tcp_drop_reason() function using BPF to monitor > > TCP drop reasons. However, after upgrading our compiler from GCC 9 to GCC > > 11, tcp_drop_reason() is now inlined, preventing us from hooking into it. > > To address this, it would be beneficial to introduce a dedicated tracepoint > > for monitoring. > > This patch would require changes in user space tracers. > I am surprised no one came up with a noinline variant. > > __bpf_kfunc is using > > #define __bpf_kfunc __used __retain noinline > > I would rather not have include/trace/events/tcp.h becoming the > biggest file in TCP stack... I’d prefer not to introduce a new tracepoint if we can easily hook it with BPF. Does the following change look good to you? diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c index 092456b8f8af..ebea844cc974 100644 --- a/net/ipv4/tcp_input.c +++ b/net/ipv4/tcp_input.c @@ -4720,7 +4720,7 @@ static bool tcp_ooo_try_coalesce(struct sock *sk, return res; } -static void tcp_drop_reason(struct sock *sk, struct sk_buff *skb, +noinline static void tcp_drop_reason(struct sock *sk, struct sk_buff *skb, enum skb_drop_reason reason) { sk_drops_add(sk, skb); -- Regards Yafang
On Wed, Oct 23, 2024 at 4:35 PM Yafang Shao <laoar.shao@gmail.com> wrote: > > On Wed, Oct 23, 2024 at 9:01 PM Eric Dumazet <edumazet@google.com> wrote: > > > > On Wed, Oct 23, 2024 at 2:33 PM Yafang Shao <laoar.shao@gmail.com> wrote: > > > > > > We previously hooked the tcp_drop_reason() function using BPF to monitor > > > TCP drop reasons. However, after upgrading our compiler from GCC 9 to GCC > > > 11, tcp_drop_reason() is now inlined, preventing us from hooking into it. > > > To address this, it would be beneficial to introduce a dedicated tracepoint > > > for monitoring. > > > > This patch would require changes in user space tracers. > > I am surprised no one came up with a noinline variant. > > > > __bpf_kfunc is using > > > > #define __bpf_kfunc __used __retain noinline > > > > I would rather not have include/trace/events/tcp.h becoming the > > biggest file in TCP stack... > > I’d prefer not to introduce a new tracepoint if we can easily hook it > with BPF. Does the following change look good to you? > > diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c > index 092456b8f8af..ebea844cc974 100644 > --- a/net/ipv4/tcp_input.c > +++ b/net/ipv4/tcp_input.c > @@ -4720,7 +4720,7 @@ static bool tcp_ooo_try_coalesce(struct sock *sk, > return res; > } > > -static void tcp_drop_reason(struct sock *sk, struct sk_buff *skb, > +noinline static void tcp_drop_reason(struct sock *sk, struct sk_buff *skb, > enum skb_drop_reason reason) > { > sk_drops_add(sk, skb); > I would suggest adding an explicit keyword, like the one we have for noinline_for_stack, for documentation purposes. noinline_for_tracing perhaps ? diff --git a/include/linux/compiler_types.h b/include/linux/compiler_types.h index 1a957ea2f4fe78ed12d7f6a65e5759d07cea4449..9a687ca4bb4392583d150349ee11015bcb82ec74 100644 --- a/include/linux/compiler_types.h +++ b/include/linux/compiler_types.h @@ -265,6 +265,12 @@ struct ftrace_likely_data { */ #define noinline_for_stack noinline +/* + * Use noinline_for_tracing for functions that should not be inlined, + * for tracing reasons. + */ +#define noinline_for_tracing noinline + /* * Sanitizer helper attributes: Because using __always_inline and * __no_sanitize_* conflict, provide helper attributes that will either expand
On Wed, Oct 23, 2024 at 11:43 PM Eric Dumazet <edumazet@google.com> wrote: > > On Wed, Oct 23, 2024 at 4:35 PM Yafang Shao <laoar.shao@gmail.com> wrote: > > > > On Wed, Oct 23, 2024 at 9:01 PM Eric Dumazet <edumazet@google.com> wrote: > > > > > > On Wed, Oct 23, 2024 at 2:33 PM Yafang Shao <laoar.shao@gmail.com> wrote: > > > > > > > > We previously hooked the tcp_drop_reason() function using BPF to monitor > > > > TCP drop reasons. However, after upgrading our compiler from GCC 9 to GCC > > > > 11, tcp_drop_reason() is now inlined, preventing us from hooking into it. > > > > To address this, it would be beneficial to introduce a dedicated tracepoint > > > > for monitoring. > > > > > > This patch would require changes in user space tracers. > > > I am surprised no one came up with a noinline variant. > > > > > > __bpf_kfunc is using > > > > > > #define __bpf_kfunc __used __retain noinline > > > > > > I would rather not have include/trace/events/tcp.h becoming the > > > biggest file in TCP stack... > > > > I’d prefer not to introduce a new tracepoint if we can easily hook it > > with BPF. Does the following change look good to you? > > > > diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c > > index 092456b8f8af..ebea844cc974 100644 > > --- a/net/ipv4/tcp_input.c > > +++ b/net/ipv4/tcp_input.c > > @@ -4720,7 +4720,7 @@ static bool tcp_ooo_try_coalesce(struct sock *sk, > > return res; > > } > > > > -static void tcp_drop_reason(struct sock *sk, struct sk_buff *skb, > > +noinline static void tcp_drop_reason(struct sock *sk, struct sk_buff *skb, > > enum skb_drop_reason reason) > > { > > sk_drops_add(sk, skb); > > > > I would suggest adding an explicit keyword, like the one we have for > noinline_for_stack, for documentation purposes. > > noinline_for_tracing perhaps ? Good suggestion! This approach eliminates the need to add comments for each noinline. > > diff --git a/include/linux/compiler_types.h b/include/linux/compiler_types.h > index 1a957ea2f4fe78ed12d7f6a65e5759d07cea4449..9a687ca4bb4392583d150349ee11015bcb82ec74 > 100644 > --- a/include/linux/compiler_types.h > +++ b/include/linux/compiler_types.h > @@ -265,6 +265,12 @@ struct ftrace_likely_data { > */ > #define noinline_for_stack noinline > > +/* > + * Use noinline_for_tracing for functions that should not be inlined, > + * for tracing reasons. > + */ > +#define noinline_for_tracing noinline > + > /* > * Sanitizer helper attributes: Because using __always_inline and > * __no_sanitize_* conflict, provide helper attributes that will either expand -- Regards Yafang
diff --git a/include/trace/events/tcp.h b/include/trace/events/tcp.h index a27c4b619dff..056f7026224c 100644 --- a/include/trace/events/tcp.h +++ b/include/trace/events/tcp.h @@ -12,6 +12,7 @@ #include <net/tcp.h> #include <linux/sock_diag.h> #include <net/rstreason.h> +#include <net/dropreason-core.h> /* * tcp event with arguments sk and skb @@ -728,6 +729,58 @@ DEFINE_EVENT(tcp_ao_event_sne, tcp_ao_rcv_sne_update, TP_ARGS(sk, new_sne) ); +#undef FN +#undef FNe +#define FN(reason) { SKB_DROP_REASON_##reason, #reason }, +#define FNe(reason) { SKB_DROP_REASON_##reason, #reason } + +TRACE_EVENT(tcp_drop_reason, + + TP_PROTO(const struct sock *sk, const struct sk_buff *skb, + const enum skb_drop_reason reason), + + TP_ARGS(sk, skb, reason), + + TP_STRUCT__entry( + __field(const void *, skbaddr) + __field(const void *, skaddr) + __field(int, state) + __field(enum skb_drop_reason, reason) + __array(__u8, saddr, sizeof(struct sockaddr_in6)) + __array(__u8, daddr, sizeof(struct sockaddr_in6)) + ), + + TP_fast_assign( + __entry->skbaddr = skb; + __entry->skaddr = sk; + /* Zero means unknown state. */ + __entry->state = sk ? sk->sk_state : 0; + + memset(__entry->saddr, 0, sizeof(struct sockaddr_in6)); + memset(__entry->daddr, 0, sizeof(struct sockaddr_in6)); + + if (sk && sk_fullsock(sk)) { + const struct inet_sock *inet = inet_sk(sk); + + TP_STORE_ADDR_PORTS(__entry, inet, sk); + } else { + const struct tcphdr *th = (const struct tcphdr *)skb->data; + + TP_STORE_ADDR_PORTS_SKB(skb, th, entry->saddr, entry->daddr); + } + __entry->reason = reason; + ), + + TP_printk("skbaddr=%p skaddr=%p src=%pISpc dest=%pISpc state=%s reason=%s", + __entry->skbaddr, __entry->skaddr, + __entry->saddr, __entry->daddr, + __entry->state ? show_tcp_state_name(__entry->state) : "UNKNOWN", + __print_symbolic(__entry->reason, DEFINE_DROP_REASON(FN, FNe))) +); + +#undef FN +#undef FNe + #endif /* _TRACE_TCP_H */ /* This part must be outside protection */ diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c index cc05ec1faac8..44795555596a 100644 --- a/net/ipv4/tcp_input.c +++ b/net/ipv4/tcp_input.c @@ -4897,6 +4897,7 @@ static bool tcp_ooo_try_coalesce(struct sock *sk, static void tcp_drop_reason(struct sock *sk, struct sk_buff *skb, enum skb_drop_reason reason) { + trace_tcp_drop_reason(sk, skb, reason); sk_drops_add(sk, skb); sk_skb_reason_drop(sk, skb, reason); }
We previously hooked the tcp_drop_reason() function using BPF to monitor TCP drop reasons. However, after upgrading our compiler from GCC 9 to GCC 11, tcp_drop_reason() is now inlined, preventing us from hooking into it. To address this, it would be beneficial to introduce a dedicated tracepoint for monitoring. Signed-off-by: Yafang Shao <laoar.shao@gmail.com> Cc: Menglong Dong <menglong8.dong@gmail.com> --- include/trace/events/tcp.h | 53 ++++++++++++++++++++++++++++++++++++++ net/ipv4/tcp_input.c | 1 + 2 files changed, 54 insertions(+)