| Message ID | 20241207173803.90744-8-kerneljasonxing@gmail.com (mailing list archive) |
|---|---|
| State | Changes Requested |
| Delegated to: | Netdev Maintainers |
| Headers | show |
| Series | net-timestamp: bpf extension to equip applications transparently | expand |
On 12/7/24 9:37 AM, Jason Xing wrote: > From: Jason Xing <kernelxing@tencent.com> > > Passing the hwtstamp to bpf prog which can print. > > Signed-off-by: Jason Xing <kernelxing@tencent.com> > --- > include/net/sock.h | 6 ++++-- > net/core/skbuff.c | 17 +++++++++++++---- > net/core/sock.c | 4 +++- > 3 files changed, 20 insertions(+), 7 deletions(-) > > diff --git a/include/net/sock.h b/include/net/sock.h > index f88a00108a2f..9bc883573208 100644 > --- a/include/net/sock.h > +++ b/include/net/sock.h > @@ -2921,9 +2921,11 @@ int sock_set_timestamping(struct sock *sk, int optname, > > void sock_enable_timestamps(struct sock *sk); > #if defined(CONFIG_CGROUP_BPF) && defined(CONFIG_BPF_SYSCALL) > -void bpf_skops_tx_timestamping(struct sock *sk, struct sk_buff *skb, int op); > +void bpf_skops_tx_timestamping(struct sock *sk, struct sk_buff *skb, int op, > + u32 nargs, u32 *args); > #else > -static inline void bpf_skops_tx_timestamping(struct sock *sk, struct sk_buff *skb, int op) > +static inline void bpf_skops_tx_timestamping(struct sock *sk, struct sk_buff *skb, int op, > + u32 nargs, u32 *args) > { > } > #endif > diff --git a/net/core/skbuff.c b/net/core/skbuff.c > index 48b0c71e9522..182a44815630 100644 > --- a/net/core/skbuff.c > +++ b/net/core/skbuff.c > @@ -5539,8 +5539,12 @@ void skb_complete_tx_timestamp(struct sk_buff *skb, > } > EXPORT_SYMBOL_GPL(skb_complete_tx_timestamp); > > -static void __skb_tstamp_tx_bpf(struct sock *sk, struct sk_buff *skb, int tstype) > +static void __skb_tstamp_tx_bpf(struct sock *sk, struct sk_buff *skb, > + struct skb_shared_hwtstamps *hwtstamps, > + int tstype) > { > + struct timespec64 tstamp; > + u32 args[2] = {0, 0}; > int op; > > if (!sk) > @@ -5552,6 +5556,11 @@ static void __skb_tstamp_tx_bpf(struct sock *sk, struct sk_buff *skb, int tstype > break; > case SCM_TSTAMP_SND: > op = BPF_SOCK_OPS_TS_SW_OPT_CB; > + if (hwtstamps) { > + tstamp = ktime_to_timespec64(hwtstamps->hwtstamp); Avoid this conversion which is likely not useful to the bpf prog. Directly pass hwtstamps->hwtstamp (in ns?) to the bpf prog. Put lower 32bits in args[0] and higher 32bits in args[1]. Also, how about adding a BPF_SOCK_OPS_TS_"HW"_OPT_CB for the "hwtstamps != NULL" case instead of reusing the BPF_SOCK_OPS_TS_"SW"_OPT_CB? A more subtle thing for the hwtstamps case is, afaik the bpf prog will not be called. All drivers are still only testing SKBTX_HW_TSTAMP instead of testing (SKBTX_HW_TSTAMP | SKBTX_BPF). There are a lot of drivers to change though. A quick thought is to rename the existing SKBTX_HW_TSTAMP (e.g. __SKBTX_HW_TSTAMP = 1 << 0) and define SKBTX_HW_TSTAMP like: #define SKBTX_HW_TSTAMP (__SKBTX_HW_TSTAMP | SKBTX_BPF) Then change some of the existing skb_shinfo(skb)->tx_flags "setting" site to use __SKBTX_HW_TSTAMP instead. e.g. in __sock_tx_timestamp(). Not very pretty but may be still better than changing many drivers. May be there is a better way... While talking about where to test the SKBTX_BPF bit, I wonder if the new skb_tstamp_is_set() is needed. For the non SKBTX_HW_TSTAMP case, the number of tx_flags testing sites should be limited, so should be easy to add the SKBTX_BPF bit test. e.g. at the __dev_queue_xmit, test "if (unlikely(skb_shinfo(skb)->tx_flags & (SKBTX_SCHED_TSTAMP | SKBTX_BPF)))". Patch 6 has also tested the bpf specific bit at tcp_ack_tstamp() before calling the __skb_tstamp_tx(). At the beginning of __skb_tstamp_tx(), do something like this: void __skb_tstamp_tx(struct sk_buff *orig_skb, const struct sk_buff *ack_skb, struct skb_shared_hwtstamps *hwtstamps, struct sock *sk, int tstype) { if (cgroup_bpf_enabled(CGROUP_SOCK_OPS) && unlikely(skb_shinfo(skb)->tx_flags & SKBTX_BPF)) __skb_tstamp_tx_bpf(sk, orig_skb, hwtstamps, tstype); if (unlikely(!(skb_shinfo(skb)->tx_flags & ~SKBTX_BPF))) return; Then the new skb_tstamp_tx_output() shuffle is probably not be needed also. > + args[0] = tstamp.tv_sec; > + args[1] = tstamp.tv_nsec; > + } > break; > case SCM_TSTAMP_ACK: > op = BPF_SOCK_OPS_TS_ACK_OPT_CB; > @@ -5560,7 +5569,7 @@ static void __skb_tstamp_tx_bpf(struct sock *sk, struct sk_buff *skb, int tstype > return; > } > > - bpf_skops_tx_timestamping(sk, skb, op); > + bpf_skops_tx_timestamping(sk, skb, op, 2, args); > } > > static void skb_tstamp_tx_output(struct sk_buff *orig_skb, > @@ -5651,7 +5660,7 @@ void __skb_tstamp_tx(struct sk_buff *orig_skb, > if (unlikely(skb_tstamp_is_set(orig_skb, tstype, false))) > skb_tstamp_tx_output(orig_skb, ack_skb, hwtstamps, sk, tstype); > if (unlikely(skb_tstamp_is_set(orig_skb, tstype, true))) > - __skb_tstamp_tx_bpf(sk, orig_skb, tstype); > + __skb_tstamp_tx_bpf(sk, orig_skb, hwtstamps, tstype); > } > EXPORT_SYMBOL_GPL(__skb_tstamp_tx); > > @@ -5662,7 +5671,7 @@ void skb_tstamp_tx(struct sk_buff *orig_skb, > > skb_tstamp_tx_output(orig_skb, NULL, hwtstamps, orig_skb->sk, tstype); > if (unlikely(skb_tstamp_is_set(orig_skb, tstype, true))) > - __skb_tstamp_tx_bpf(orig_skb->sk, orig_skb, tstype); > + __skb_tstamp_tx_bpf(orig_skb->sk, orig_skb, hwtstamps, tstype); > } > EXPORT_SYMBOL_GPL(skb_tstamp_tx); > > diff --git a/net/core/sock.c b/net/core/sock.c > index 79cb5c74c76c..504939bafe0c 100644 > --- a/net/core/sock.c > +++ b/net/core/sock.c > @@ -942,7 +942,8 @@ int sock_set_timestamping(struct sock *sk, int optname, > } > > #if defined(CONFIG_CGROUP_BPF) && defined(CONFIG_BPF_SYSCALL) > -void bpf_skops_tx_timestamping(struct sock *sk, struct sk_buff *skb, int op) > +void bpf_skops_tx_timestamping(struct sock *sk, struct sk_buff *skb, int op, > + u32 nargs, u32 *args) Directly pass hwtstamps->hwtstamp instead of args and nargs. Save a memcpy below. > { > struct bpf_sock_ops_kern sock_ops; > > @@ -952,6 +953,7 @@ void bpf_skops_tx_timestamping(struct sock *sk, struct sk_buff *skb, int op) > sock_ops.op = op; > sock_ops.is_fullsock = 1; > sock_ops.sk = sk; > + memcpy(sock_ops.args, args, nargs * sizeof(*args)); > __cgroup_bpf_run_filter_sock_ops(sk, &sock_ops, CGROUP_SOCK_OPS); > } > #endif
On 12/12/24 3:25 PM, Martin KaFai Lau wrote: > A more subtle thing for the hwtstamps case is, afaik the bpf prog will not be > called. All drivers are still only testing SKBTX_HW_TSTAMP instead of testing > (SKBTX_HW_TSTAMP | SKBTX_BPF). > > There are a lot of drivers to change though. A quick thought is to rename the > existing SKBTX_HW_TSTAMP (e.g. __SKBTX_HW_TSTAMP = 1 << 0) and define > SKBTX_HW_TSTAMP like: > > #define SKBTX_HW_TSTAMP (__SKBTX_HW_TSTAMP | SKBTX_BPF) > > Then change some of the existing skb_shinfo(skb)->tx_flags "setting" site to use > __SKBTX_HW_TSTAMP instead. e.g. in __sock_tx_timestamp(). Not very pretty but > may be still better than changing many drivers. May be there is a better way... > > While talking about where to test the SKBTX_BPF bit, I wonder if the new > skb_tstamp_is_set() is needed. For the non SKBTX_HW_TSTAMP case, the number of > tx_flags testing sites should be limited, so should be easy to add the SKBTX_BPF > bit test. e.g. at the __dev_queue_xmit, test "if (unlikely(skb_shinfo(skb)- > >tx_flags & (SKBTX_SCHED_TSTAMP | SKBTX_BPF)))". Patch 6 has also tested the > bpf specific bit at tcp_ack_tstamp() before calling the __skb_tstamp_tx(). > > At the beginning of __skb_tstamp_tx(), do something like this: > > void __skb_tstamp_tx(struct sk_buff *orig_skb, > const struct sk_buff *ack_skb, > struct skb_shared_hwtstamps *hwtstamps, > struct sock *sk, int tstype) > { > if (cgroup_bpf_enabled(CGROUP_SOCK_OPS) && > unlikely(skb_shinfo(skb)->tx_flags & SKBTX_BPF)) > __skb_tstamp_tx_bpf(sk, orig_skb, hwtstamps, tstype); > > if (unlikely(!(skb_shinfo(skb)->tx_flags & ~SKBTX_BPF))) This is not enough. I was wrong here. The test in skb_tstamp_is_set() is needed when SKBTX_BPF is not set. > return;
On Fri, Dec 13, 2024 at 7:26 AM Martin KaFai Lau <martin.lau@linux.dev> wrote: > > On 12/7/24 9:37 AM, Jason Xing wrote: > > From: Jason Xing <kernelxing@tencent.com> > > > > Passing the hwtstamp to bpf prog which can print. > > > > Signed-off-by: Jason Xing <kernelxing@tencent.com> > > --- > > include/net/sock.h | 6 ++++-- > > net/core/skbuff.c | 17 +++++++++++++---- > > net/core/sock.c | 4 +++- > > 3 files changed, 20 insertions(+), 7 deletions(-) > > > > diff --git a/include/net/sock.h b/include/net/sock.h > > index f88a00108a2f..9bc883573208 100644 > > --- a/include/net/sock.h > > +++ b/include/net/sock.h > > @@ -2921,9 +2921,11 @@ int sock_set_timestamping(struct sock *sk, int optname, > > > > void sock_enable_timestamps(struct sock *sk); > > #if defined(CONFIG_CGROUP_BPF) && defined(CONFIG_BPF_SYSCALL) > > -void bpf_skops_tx_timestamping(struct sock *sk, struct sk_buff *skb, int op); > > +void bpf_skops_tx_timestamping(struct sock *sk, struct sk_buff *skb, int op, > > + u32 nargs, u32 *args); > > #else > > -static inline void bpf_skops_tx_timestamping(struct sock *sk, struct sk_buff *skb, int op) > > +static inline void bpf_skops_tx_timestamping(struct sock *sk, struct sk_buff *skb, int op, > > + u32 nargs, u32 *args) > > { > > } > > #endif > > diff --git a/net/core/skbuff.c b/net/core/skbuff.c > > index 48b0c71e9522..182a44815630 100644 > > --- a/net/core/skbuff.c > > +++ b/net/core/skbuff.c > > @@ -5539,8 +5539,12 @@ void skb_complete_tx_timestamp(struct sk_buff *skb, > > } > > EXPORT_SYMBOL_GPL(skb_complete_tx_timestamp); > > > > -static void __skb_tstamp_tx_bpf(struct sock *sk, struct sk_buff *skb, int tstype) > > +static void __skb_tstamp_tx_bpf(struct sock *sk, struct sk_buff *skb, > > + struct skb_shared_hwtstamps *hwtstamps, > > + int tstype) > > { > > + struct timespec64 tstamp; > > + u32 args[2] = {0, 0}; > > int op; > > > > if (!sk) > > @@ -5552,6 +5556,11 @@ static void __skb_tstamp_tx_bpf(struct sock *sk, struct sk_buff *skb, int tstype > > break; > > case SCM_TSTAMP_SND: > > op = BPF_SOCK_OPS_TS_SW_OPT_CB; > > > + if (hwtstamps) { > > + tstamp = ktime_to_timespec64(hwtstamps->hwtstamp); > > Avoid this conversion which is likely not useful to the bpf prog. Directly pass > hwtstamps->hwtstamp (in ns?) to the bpf prog. Put lower 32bits in args[0] and > higher 32bits in args[1]. It makes sense. > > Also, how about adding a BPF_SOCK_OPS_TS_"HW"_OPT_CB for the "hwtstamps != NULL" > case instead of reusing the BPF_SOCK_OPS_TS_"SW"_OPT_CB? Good suggestion. Will do that. > > A more subtle thing for the hwtstamps case is, afaik the bpf prog will not be > called. All drivers are still only testing SKBTX_HW_TSTAMP instead of testing > (SKBTX_HW_TSTAMP | SKBTX_BPF). Ah, I didn't realize that!! > > There are a lot of drivers to change though. A quick thought is to rename the > existing SKBTX_HW_TSTAMP (e.g. __SKBTX_HW_TSTAMP = 1 << 0) and define > SKBTX_HW_TSTAMP like: > > #define SKBTX_HW_TSTAMP (__SKBTX_HW_TSTAMP | SKBTX_BPF) I will take it, thanks! > > Then change some of the existing skb_shinfo(skb)->tx_flags "setting" site to use > __SKBTX_HW_TSTAMP instead. e.g. in __sock_tx_timestamp(). Not very pretty but > may be still better than changing many drivers. May be there is a better way... I agree with your approach. Changing so many drivers would be a head-ache thing. > > While talking about where to test the SKBTX_BPF bit, I wonder if the new > skb_tstamp_is_set() is needed. For the non SKBTX_HW_TSTAMP case, the number of > tx_flags testing sites should be limited, so should be easy to add the SKBTX_BPF > bit test. e.g. at the __dev_queue_xmit, test "if > (unlikely(skb_shinfo(skb)->tx_flags & (SKBTX_SCHED_TSTAMP | SKBTX_BPF)))". Patch > 6 has also tested the bpf specific bit at tcp_ack_tstamp() before calling the > __skb_tstamp_tx(). > > At the beginning of __skb_tstamp_tx(), do something like this: > > void __skb_tstamp_tx(struct sk_buff *orig_skb, > const struct sk_buff *ack_skb, > struct skb_shared_hwtstamps *hwtstamps, > struct sock *sk, int tstype) > { > if (cgroup_bpf_enabled(CGROUP_SOCK_OPS) && > unlikely(skb_shinfo(skb)->tx_flags & SKBTX_BPF)) > __skb_tstamp_tx_bpf(sk, orig_skb, hwtstamps, tstype); > > if (unlikely(!(skb_shinfo(skb)->tx_flags & ~SKBTX_BPF))) > return; > > Then the new skb_tstamp_tx_output() shuffle is probably not be needed also. > > > + args[0] = tstamp.tv_sec; > > + args[1] = tstamp.tv_nsec; > > + } > > break; > > case SCM_TSTAMP_ACK: > > op = BPF_SOCK_OPS_TS_ACK_OPT_CB; > > @@ -5560,7 +5569,7 @@ static void __skb_tstamp_tx_bpf(struct sock *sk, struct sk_buff *skb, int tstype > > return; > > } > > > > - bpf_skops_tx_timestamping(sk, skb, op); > > + bpf_skops_tx_timestamping(sk, skb, op, 2, args); > > } > > > > static void skb_tstamp_tx_output(struct sk_buff *orig_skb, > > @@ -5651,7 +5660,7 @@ void __skb_tstamp_tx(struct sk_buff *orig_skb, > > if (unlikely(skb_tstamp_is_set(orig_skb, tstype, false))) > > skb_tstamp_tx_output(orig_skb, ack_skb, hwtstamps, sk, tstype); > > if (unlikely(skb_tstamp_is_set(orig_skb, tstype, true))) > > - __skb_tstamp_tx_bpf(sk, orig_skb, tstype); > > + __skb_tstamp_tx_bpf(sk, orig_skb, hwtstamps, tstype); > > } > > EXPORT_SYMBOL_GPL(__skb_tstamp_tx); > > > > @@ -5662,7 +5671,7 @@ void skb_tstamp_tx(struct sk_buff *orig_skb, > > > > skb_tstamp_tx_output(orig_skb, NULL, hwtstamps, orig_skb->sk, tstype); > > if (unlikely(skb_tstamp_is_set(orig_skb, tstype, true))) > > - __skb_tstamp_tx_bpf(orig_skb->sk, orig_skb, tstype); > > + __skb_tstamp_tx_bpf(orig_skb->sk, orig_skb, hwtstamps, tstype); > > } > > EXPORT_SYMBOL_GPL(skb_tstamp_tx); > > > > diff --git a/net/core/sock.c b/net/core/sock.c > > index 79cb5c74c76c..504939bafe0c 100644 > > --- a/net/core/sock.c > > +++ b/net/core/sock.c > > @@ -942,7 +942,8 @@ int sock_set_timestamping(struct sock *sk, int optname, > > } > > > > #if defined(CONFIG_CGROUP_BPF) && defined(CONFIG_BPF_SYSCALL) > > -void bpf_skops_tx_timestamping(struct sock *sk, struct sk_buff *skb, int op) > > +void bpf_skops_tx_timestamping(struct sock *sk, struct sk_buff *skb, int op, > > + u32 nargs, u32 *args) > > Directly pass hwtstamps->hwtstamp instead of args and nargs. Save a memcpy below. > > > { > > struct bpf_sock_ops_kern sock_ops; > > > > @@ -952,6 +953,7 @@ void bpf_skops_tx_timestamping(struct sock *sk, struct sk_buff *skb, int op) > > sock_ops.op = op; > > sock_ops.is_fullsock = 1; > > sock_ops.sk = sk; > > + memcpy(sock_ops.args, args, nargs * sizeof(*args)); > > __cgroup_bpf_run_filter_sock_ops(sk, &sock_ops, CGROUP_SOCK_OPS); > > } > > #endif >
On 12/13/24 7:13 AM, Jason Xing wrote: >>> -static void __skb_tstamp_tx_bpf(struct sock *sk, struct sk_buff *skb, int tstype) >>> +static void __skb_tstamp_tx_bpf(struct sock *sk, struct sk_buff *skb, >>> + struct skb_shared_hwtstamps *hwtstamps, >>> + int tstype) >>> { >>> + struct timespec64 tstamp; >>> + u32 args[2] = {0, 0}; >>> int op; >>> >>> if (!sk) >>> @@ -5552,6 +5556,11 @@ static void __skb_tstamp_tx_bpf(struct sock *sk, struct sk_buff *skb, int tstype >>> break; >>> case SCM_TSTAMP_SND: >>> op = BPF_SOCK_OPS_TS_SW_OPT_CB; >>> + if (hwtstamps) { >>> + tstamp = ktime_to_timespec64(hwtstamps->hwtstamp); >> Avoid this conversion which is likely not useful to the bpf prog. Directly pass >> hwtstamps->hwtstamp (in ns?) to the bpf prog. Put lower 32bits in args[0] and >> higher 32bits in args[1]. > It makes sense. When replying the patch 2 thread, I noticed it may not even have to pass the hwtstamps in args here. Can "*skb_hwtstamps(skb) = *hwtstamps;" be done before calling the bpf prog? Then the bpf prog can directly get it from skb_shinfo(skb)->hwtstamps. It is like reading other fields in skb_shinfo(skb), e.g. the skb_shinfo(skb)->tskey discussed in patch 10. The bpf prog will have a more consistent experience in reading different fields of the skb_shinfo(skb). skb_shinfo(skb)->hwtstamps is a more intuitive place to obtain the hwtstamp than the broken up args[0] and args[1]. On top of that, there is also an older "skb_hwtstamp" field in "struct bpf_sock_ops".
On Sat, Dec 14, 2024 at 7:15 AM Martin KaFai Lau <martin.lau@linux.dev> wrote: > > On 12/13/24 7:13 AM, Jason Xing wrote: > >>> -static void __skb_tstamp_tx_bpf(struct sock *sk, struct sk_buff *skb, int tstype) > >>> +static void __skb_tstamp_tx_bpf(struct sock *sk, struct sk_buff *skb, > >>> + struct skb_shared_hwtstamps *hwtstamps, > >>> + int tstype) > >>> { > >>> + struct timespec64 tstamp; > >>> + u32 args[2] = {0, 0}; > >>> int op; > >>> > >>> if (!sk) > >>> @@ -5552,6 +5556,11 @@ static void __skb_tstamp_tx_bpf(struct sock *sk, struct sk_buff *skb, int tstype > >>> break; > >>> case SCM_TSTAMP_SND: > >>> op = BPF_SOCK_OPS_TS_SW_OPT_CB; > >>> + if (hwtstamps) { > >>> + tstamp = ktime_to_timespec64(hwtstamps->hwtstamp); > >> Avoid this conversion which is likely not useful to the bpf prog. Directly pass > >> hwtstamps->hwtstamp (in ns?) to the bpf prog. Put lower 32bits in args[0] and > >> higher 32bits in args[1]. > > It makes sense. > > When replying the patch 2 thread, I noticed it may not even have to pass the > hwtstamps in args here. > > Can "*skb_hwtstamps(skb) = *hwtstamps;" be done before calling the bpf prog? > Then the bpf prog can directly get it from skb_shinfo(skb)->hwtstamps. > It is like reading other fields in skb_shinfo(skb), e.g. the > skb_shinfo(skb)->tskey discussed in patch 10. The bpf prog will have a more > consistent experience in reading different fields of the skb_shinfo(skb). > skb_shinfo(skb)->hwtstamps is a more intuitive place to obtain the hwtstamp than > the broken up args[0] and args[1]. On top of that, there is also an older > "skb_hwtstamp" field in "struct bpf_sock_ops". Right, right, last night, fortunately, I also spotted it. Let the bpf prog parse the shared info from skb then. A new callback for hwtstamp is needed, I suppose.
On 12/13/24 4:02 PM, Jason Xing wrote: > On Sat, Dec 14, 2024 at 7:15 AM Martin KaFai Lau <martin.lau@linux.dev> wrote: >> >> On 12/13/24 7:13 AM, Jason Xing wrote: >>>>> -static void __skb_tstamp_tx_bpf(struct sock *sk, struct sk_buff *skb, int tstype) >>>>> +static void __skb_tstamp_tx_bpf(struct sock *sk, struct sk_buff *skb, >>>>> + struct skb_shared_hwtstamps *hwtstamps, >>>>> + int tstype) >>>>> { >>>>> + struct timespec64 tstamp; >>>>> + u32 args[2] = {0, 0}; >>>>> int op; >>>>> >>>>> if (!sk) >>>>> @@ -5552,6 +5556,11 @@ static void __skb_tstamp_tx_bpf(struct sock *sk, struct sk_buff *skb, int tstype >>>>> break; >>>>> case SCM_TSTAMP_SND: >>>>> op = BPF_SOCK_OPS_TS_SW_OPT_CB; >>>>> + if (hwtstamps) { >>>>> + tstamp = ktime_to_timespec64(hwtstamps->hwtstamp); >>>> Avoid this conversion which is likely not useful to the bpf prog. Directly pass >>>> hwtstamps->hwtstamp (in ns?) to the bpf prog. Put lower 32bits in args[0] and >>>> higher 32bits in args[1]. >>> It makes sense. >> >> When replying the patch 2 thread, I noticed it may not even have to pass the >> hwtstamps in args here. >> >> Can "*skb_hwtstamps(skb) = *hwtstamps;" be done before calling the bpf prog? >> Then the bpf prog can directly get it from skb_shinfo(skb)->hwtstamps. >> It is like reading other fields in skb_shinfo(skb), e.g. the >> skb_shinfo(skb)->tskey discussed in patch 10. The bpf prog will have a more >> consistent experience in reading different fields of the skb_shinfo(skb). >> skb_shinfo(skb)->hwtstamps is a more intuitive place to obtain the hwtstamp than >> the broken up args[0] and args[1]. On top of that, there is also an older >> "skb_hwtstamp" field in "struct bpf_sock_ops". > > Right, right, last night, fortunately, I also spotted it. Let the bpf > prog parse the shared info from skb then. A new callback for hwtstamp > is needed, I suppose. Why a new callback is needed? "*skb_hwtstamps(skb) = *hwtstamps;" cannot be done in __skb_tstamp_tx_bpf?
On Sat, Dec 14, 2024 at 8:17 AM Martin KaFai Lau <martin.lau@linux.dev> wrote: > > On 12/13/24 4:02 PM, Jason Xing wrote: > > On Sat, Dec 14, 2024 at 7:15 AM Martin KaFai Lau <martin.lau@linux.dev> wrote: > >> > >> On 12/13/24 7:13 AM, Jason Xing wrote: > >>>>> -static void __skb_tstamp_tx_bpf(struct sock *sk, struct sk_buff *skb, int tstype) > >>>>> +static void __skb_tstamp_tx_bpf(struct sock *sk, struct sk_buff *skb, > >>>>> + struct skb_shared_hwtstamps *hwtstamps, > >>>>> + int tstype) > >>>>> { > >>>>> + struct timespec64 tstamp; > >>>>> + u32 args[2] = {0, 0}; > >>>>> int op; > >>>>> > >>>>> if (!sk) > >>>>> @@ -5552,6 +5556,11 @@ static void __skb_tstamp_tx_bpf(struct sock *sk, struct sk_buff *skb, int tstype > >>>>> break; > >>>>> case SCM_TSTAMP_SND: > >>>>> op = BPF_SOCK_OPS_TS_SW_OPT_CB; > >>>>> + if (hwtstamps) { > >>>>> + tstamp = ktime_to_timespec64(hwtstamps->hwtstamp); > >>>> Avoid this conversion which is likely not useful to the bpf prog. Directly pass > >>>> hwtstamps->hwtstamp (in ns?) to the bpf prog. Put lower 32bits in args[0] and > >>>> higher 32bits in args[1]. > >>> It makes sense. > >> > >> When replying the patch 2 thread, I noticed it may not even have to pass the > >> hwtstamps in args here. > >> > >> Can "*skb_hwtstamps(skb) = *hwtstamps;" be done before calling the bpf prog? > >> Then the bpf prog can directly get it from skb_shinfo(skb)->hwtstamps. > >> It is like reading other fields in skb_shinfo(skb), e.g. the > >> skb_shinfo(skb)->tskey discussed in patch 10. The bpf prog will have a more > >> consistent experience in reading different fields of the skb_shinfo(skb). > >> skb_shinfo(skb)->hwtstamps is a more intuitive place to obtain the hwtstamp than > >> the broken up args[0] and args[1]. On top of that, there is also an older > >> "skb_hwtstamp" field in "struct bpf_sock_ops". > > > > Right, right, last night, fortunately, I also spotted it. Let the bpf > > prog parse the shared info from skb then. A new callback for hwtstamp > > is needed, I suppose. > > Why a new callback is needed? "*skb_hwtstamps(skb) = *hwtstamps;" cannot be done > in __skb_tstamp_tx_bpf? Oh, I have no preference on this point. I will abort adding a new callback then.
diff --git a/include/net/sock.h b/include/net/sock.h index f88a00108a2f..9bc883573208 100644 --- a/include/net/sock.h +++ b/include/net/sock.h @@ -2921,9 +2921,11 @@ int sock_set_timestamping(struct sock *sk, int optname, void sock_enable_timestamps(struct sock *sk); #if defined(CONFIG_CGROUP_BPF) && defined(CONFIG_BPF_SYSCALL) -void bpf_skops_tx_timestamping(struct sock *sk, struct sk_buff *skb, int op); +void bpf_skops_tx_timestamping(struct sock *sk, struct sk_buff *skb, int op, + u32 nargs, u32 *args); #else -static inline void bpf_skops_tx_timestamping(struct sock *sk, struct sk_buff *skb, int op) +static inline void bpf_skops_tx_timestamping(struct sock *sk, struct sk_buff *skb, int op, + u32 nargs, u32 *args) { } #endif diff --git a/net/core/skbuff.c b/net/core/skbuff.c index 48b0c71e9522..182a44815630 100644 --- a/net/core/skbuff.c +++ b/net/core/skbuff.c @@ -5539,8 +5539,12 @@ void skb_complete_tx_timestamp(struct sk_buff *skb, } EXPORT_SYMBOL_GPL(skb_complete_tx_timestamp); -static void __skb_tstamp_tx_bpf(struct sock *sk, struct sk_buff *skb, int tstype) +static void __skb_tstamp_tx_bpf(struct sock *sk, struct sk_buff *skb, + struct skb_shared_hwtstamps *hwtstamps, + int tstype) { + struct timespec64 tstamp; + u32 args[2] = {0, 0}; int op; if (!sk) @@ -5552,6 +5556,11 @@ static void __skb_tstamp_tx_bpf(struct sock *sk, struct sk_buff *skb, int tstype break; case SCM_TSTAMP_SND: op = BPF_SOCK_OPS_TS_SW_OPT_CB; + if (hwtstamps) { + tstamp = ktime_to_timespec64(hwtstamps->hwtstamp); + args[0] = tstamp.tv_sec; + args[1] = tstamp.tv_nsec; + } break; case SCM_TSTAMP_ACK: op = BPF_SOCK_OPS_TS_ACK_OPT_CB; @@ -5560,7 +5569,7 @@ static void __skb_tstamp_tx_bpf(struct sock *sk, struct sk_buff *skb, int tstype return; } - bpf_skops_tx_timestamping(sk, skb, op); + bpf_skops_tx_timestamping(sk, skb, op, 2, args); } static void skb_tstamp_tx_output(struct sk_buff *orig_skb, @@ -5651,7 +5660,7 @@ void __skb_tstamp_tx(struct sk_buff *orig_skb, if (unlikely(skb_tstamp_is_set(orig_skb, tstype, false))) skb_tstamp_tx_output(orig_skb, ack_skb, hwtstamps, sk, tstype); if (unlikely(skb_tstamp_is_set(orig_skb, tstype, true))) - __skb_tstamp_tx_bpf(sk, orig_skb, tstype); + __skb_tstamp_tx_bpf(sk, orig_skb, hwtstamps, tstype); } EXPORT_SYMBOL_GPL(__skb_tstamp_tx); @@ -5662,7 +5671,7 @@ void skb_tstamp_tx(struct sk_buff *orig_skb, skb_tstamp_tx_output(orig_skb, NULL, hwtstamps, orig_skb->sk, tstype); if (unlikely(skb_tstamp_is_set(orig_skb, tstype, true))) - __skb_tstamp_tx_bpf(orig_skb->sk, orig_skb, tstype); + __skb_tstamp_tx_bpf(orig_skb->sk, orig_skb, hwtstamps, tstype); } EXPORT_SYMBOL_GPL(skb_tstamp_tx); diff --git a/net/core/sock.c b/net/core/sock.c index 79cb5c74c76c..504939bafe0c 100644 --- a/net/core/sock.c +++ b/net/core/sock.c @@ -942,7 +942,8 @@ int sock_set_timestamping(struct sock *sk, int optname, } #if defined(CONFIG_CGROUP_BPF) && defined(CONFIG_BPF_SYSCALL) -void bpf_skops_tx_timestamping(struct sock *sk, struct sk_buff *skb, int op) +void bpf_skops_tx_timestamping(struct sock *sk, struct sk_buff *skb, int op, + u32 nargs, u32 *args) { struct bpf_sock_ops_kern sock_ops; @@ -952,6 +953,7 @@ void bpf_skops_tx_timestamping(struct sock *sk, struct sk_buff *skb, int op) sock_ops.op = op; sock_ops.is_fullsock = 1; sock_ops.sk = sk; + memcpy(sock_ops.args, args, nargs * sizeof(*args)); __cgroup_bpf_run_filter_sock_ops(sk, &sock_ops, CGROUP_SOCK_OPS); } #endif