| Message ID | 20230703175048.151683-1-jthinz@mailbox.tu-berlin.de (mailing list archive) |
|---|---|
| Headers | show |
| Series | bpf, net: Allow setting SO_TIMESTAMPING* from BPF | expand |
Jörn-Thorben Hinz wrote: > BPF applications, e.g., a TCP congestion control, might benefit from > precise packet timestamps. These timestamps are already available in > __sk_buff and bpf_sock_ops, but could not be requested: A BPF program > was not allowed to set SO_TIMESTAMPING* on a socket. This change enables > BPF programs to actively request the generation of timestamps from a > stream socket. > > To reuse the setget_sockopt BPF prog test for > bpf_{get,set}sockopt(SO_TIMESTAMPING_NEW), also implement the missing > getsockopt(SO_TIMESTAMPING_NEW) in the network stack. > > I reckon the way I added getsockopt(SO_TIMESTAMPING_NEW) causes an API > change: For existing users that set SO_TIMESTAMPING_NEW but queried > SO_TIMESTAMPING_OLD afterwards, it would now look as if no timestamping > flags have been set. Is this an acceptable change? If not, I’m happy to > change getsockopt() to only be strict about the newly-implemented > getsockopt(SO_TIMESTAMPING_NEW), or not distinguish between > SO_TIMESTAMPING_NEW and SO_TIMESTAMPING_OLD at all. Yeah, I think it would be best if we keep the old behavior and let SO_TIMESTAMPING_OLD return timestamps for both new/old. It looks like it should be relatively easy to implement? Otherwise the series lgtm. > > Jörn-Thorben Hinz (2): > net: Implement missing getsockopt(SO_TIMESTAMPING_NEW) > bpf: Allow setting SO_TIMESTAMPING* with bpf_setsockopt() > > include/uapi/linux/bpf.h | 3 ++- > net/core/filter.c | 2 ++ > net/core/sock.c | 9 +++++++-- > tools/include/uapi/linux/bpf.h | 3 ++- > tools/testing/selftests/bpf/progs/bpf_tracing_net.h | 2 ++ > tools/testing/selftests/bpf/progs/setget_sockopt.c | 4 ++++ > 6 files changed, 19 insertions(+), 4 deletions(-) > > -- > 2.39.2 > >
Thank you for the feedback. Just noticed I missed the “bpf-next” designation in the subject. Will add that in v2. On Mon, 2023-07-03 at 14:25 -0700, John Fastabend wrote: > Jörn-Thorben Hinz wrote: > > BPF applications, e.g., a TCP congestion control, might benefit > > from > > precise packet timestamps. These timestamps are already available > > in > > __sk_buff and bpf_sock_ops, but could not be requested: A BPF > > program > > was not allowed to set SO_TIMESTAMPING* on a socket. This change > > enables > > BPF programs to actively request the generation of timestamps from > > a > > stream socket. > > > > To reuse the setget_sockopt BPF prog test for > > bpf_{get,set}sockopt(SO_TIMESTAMPING_NEW), also implement the > > missing > > getsockopt(SO_TIMESTAMPING_NEW) in the network stack. > > > > I reckon the way I added getsockopt(SO_TIMESTAMPING_NEW) causes an > > API > > change: For existing users that set SO_TIMESTAMPING_NEW but queried > > SO_TIMESTAMPING_OLD afterwards, it would now look as if no > > timestamping > > flags have been set. Is this an acceptable change? If not, I’m > > happy to > > change getsockopt() to only be strict about the newly-implemented > > getsockopt(SO_TIMESTAMPING_NEW), or not distinguish between > > SO_TIMESTAMPING_NEW and SO_TIMESTAMPING_OLD at all. > > Yeah, I think it would be best if we keep the old behavior and let > SO_TIMESTAMPING_OLD return timestamps for both new/old. It looks > like it should be relatively easy to implement? Alright, I guessed that would be preferred. Yes, if there is no objection to making the added getsockopt(SO_TIMESTAMPING_NEW) this tiny bit more “strict”, it’s just a matter of modifying the if inserted in sk_getsockopt(). (And, well, in the other case I would even remove this if.) > > Otherwise the series lgtm. Great, thanks. > > > > > Jörn-Thorben Hinz (2): > > net: Implement missing getsockopt(SO_TIMESTAMPING_NEW) > > bpf: Allow setting SO_TIMESTAMPING* with bpf_setsockopt() > > > > include/uapi/linux/bpf.h | 3 ++- > > net/core/filter.c | 2 ++ > > net/core/sock.c | 9 +++++++-- > > tools/include/uapi/linux/bpf.h | 3 ++- > > tools/testing/selftests/bpf/progs/bpf_tracing_net.h | 2 ++ > > tools/testing/selftests/bpf/progs/setget_sockopt.c | 4 ++++ > > 6 files changed, 19 insertions(+), 4 deletions(-) > > > > -- > > 2.39.2 > >
> On Mon, 2023-07-03 at 14:25 -0700, John Fastabend wrote: > > Jörn-Thorben Hinz wrote: > > > BPF applications, e.g., a TCP congestion control, might benefit > > > from > > > precise packet timestamps. These timestamps are already available > > > in > > > __sk_buff and bpf_sock_ops, but could not be requested: A BPF > > > program > > > was not allowed to set SO_TIMESTAMPING* on a socket. This change > > > enables > > > BPF programs to actively request the generation of timestamps from > > > a > > > stream socket. > > > > > > To reuse the setget_sockopt BPF prog test for > > > bpf_{get,set}sockopt(SO_TIMESTAMPING_NEW), also implement the > > > missing > > > getsockopt(SO_TIMESTAMPING_NEW) in the network stack. > > > > > > I reckon the way I added getsockopt(SO_TIMESTAMPING_NEW) causes an > > > API > > > change: For existing users that set SO_TIMESTAMPING_NEW but queried > > > SO_TIMESTAMPING_OLD afterwards, it would now look as if no > > > timestamping > > > flags have been set. Is this an acceptable change? If not, I’m > > > happy to > > > change getsockopt() to only be strict about the newly-implemented > > > getsockopt(SO_TIMESTAMPING_NEW), or not distinguish between > > > SO_TIMESTAMPING_NEW and SO_TIMESTAMPING_OLD at all. > > > > Yeah, I think it would be best if we keep the old behavior and let > > SO_TIMESTAMPING_OLD return timestamps for both new/old. It looks > > like it should be relatively easy to implement? > Alright, I guessed that would be preferred. > > Yes, if there is no objection to making the added > getsockopt(SO_TIMESTAMPING_NEW) this tiny bit more “strict”, it’s just > a matter of modifying the if inserted in sk_getsockopt(). (And, well, > in the other case I would even remove this if.) The difference is in the struct that is returned, on 32-bit platforms. Both calls should always be allowed? See also put_cmsg_scm_timestamping64 vs put_cmsg_scm_timestamping. For the second patch: the _OLD/_NEW was introduced to work around limitations on 32-bit platforms. This is intended to be transparent to users, by defining SO_TIMESTAMPING accordingly. Can the new BPF code always enforce the 64-bit version, that is, only implement the _NEW variants? And perhaps just call it SO_TIMESTAMPING directly.
On Tue, Jul 4, 2023, at 21:36, Willem de Bruijn wrote: >> On Mon, 2023-07-03 at 14:25 -0700, John Fastabend wrote: >> > Jörn-Thorben Hinz wrote: >> Yes, if there is no objection to making the added >> getsockopt(SO_TIMESTAMPING_NEW) this tiny bit more “strict”, it’s just >> a matter of modifying the if inserted in sk_getsockopt(). (And, well, >> in the other case I would even remove this if.) > > The difference is in the struct that is returned, on 32-bit platforms. > Both calls should always be allowed? See also > put_cmsg_scm_timestamping64 vs put_cmsg_scm_timestamping. > > For the second patch: the _OLD/_NEW was introduced to work around > limitations on 32-bit platforms. This is intended to be transparent to > users, by defining SO_TIMESTAMPING accordingly. > > Can the new BPF code always enforce the 64-bit version, that is, only > implement the _NEW variants? And perhaps just call it SO_TIMESTAMPING > directly. I guess that depends on how the returned timestamps are interpreted. In normal userspace code, the 'struct scm_timestamping' is defined in terms of the libc-provided 'struct timespec'. If this is a normal glibc based distro binary, then it probably expects the old format. OTOH, if the code reading the timestamp data is in BPF code itself, it's probably safe to mandate that to use the time64 format and define the timespec type as __kernel_timespec with 64-bit members. Arnd
BPF applications, e.g., a TCP congestion control, might benefit from precise packet timestamps. These timestamps are already available in __sk_buff and bpf_sock_ops, but could not be requested: A BPF program was not allowed to set SO_TIMESTAMPING* on a socket. This change enables BPF programs to actively request the generation of timestamps from a stream socket. To reuse the setget_sockopt BPF prog test for bpf_{get,set}sockopt(SO_TIMESTAMPING_NEW), also implement the missing getsockopt(SO_TIMESTAMPING_NEW) in the network stack. I reckon the way I added getsockopt(SO_TIMESTAMPING_NEW) causes an API change: For existing users that set SO_TIMESTAMPING_NEW but queried SO_TIMESTAMPING_OLD afterwards, it would now look as if no timestamping flags have been set. Is this an acceptable change? If not, I’m happy to change getsockopt() to only be strict about the newly-implemented getsockopt(SO_TIMESTAMPING_NEW), or not distinguish between SO_TIMESTAMPING_NEW and SO_TIMESTAMPING_OLD at all. Jörn-Thorben Hinz (2): net: Implement missing getsockopt(SO_TIMESTAMPING_NEW) bpf: Allow setting SO_TIMESTAMPING* with bpf_setsockopt() include/uapi/linux/bpf.h | 3 ++- net/core/filter.c | 2 ++ net/core/sock.c | 9 +++++++-- tools/include/uapi/linux/bpf.h | 3 ++- tools/testing/selftests/bpf/progs/bpf_tracing_net.h | 2 ++ tools/testing/selftests/bpf/progs/setget_sockopt.c | 4 ++++ 6 files changed, 19 insertions(+), 4 deletions(-)