On 3/13/24 1:52 AM, Willem de Bruijn wrote:
Martin KaFai Lau wrote:
On 3/1/24 12:13 PM, Abhishek Chauhan wrote:
Bridge driver today has no support to forward the userspace timestamp
packets and ends up resetting the timestamp. ETF qdisc checks the
packet coming from userspace and encounters to be 0 thereby dropping
time sensitive packets. These changes will allow userspace timestamps
packets to be forwarded from the bridge to NIC drivers.
Setting the same bit (mono_delivery_time) to avoid dropping of
userspace tstamp packets in the forwarding path.
Existing functionality of mono_delivery_time remains unaltered here,
instead just extended with userspace tstamp support for bridge
forwarding path.
The patch currently broke the bpf selftest test_tc_dtime:
https://github.com/kernel-patches/bpf/actions/runs/8242487344/job/22541746675
In particular, there is a uapi field __sk_buff->tstamp_type which currently has
BPF_SKB_TSTAMP_DELIVERY_MONO to mean skb->tstamp has the MONO "delivery" time.
BPF_SKB_TSTAMP_UNSPEC means everything else (this could be a rx timestamp at
ingress or a delivery time set by user space).
__sk_buff->tstamp_type depends on skb->mono_delivery_time which does not
necessarily mean mono after this patch. I thought about fixing it on the bpf
side such that reading __sk_buff->tstamp_type only returns
BPF_SKB_TSTAMP_DELIVERY_MONO when the skb->mono_delivery_time is set and skb->sk
is IPPROTO_TCP. However, it won't work because of bpf_skb_set_tstamp().
There is a bpf helper, bpf_skb_set_tstamp(skb, tstamp,
BPF_SKB_TSTAMP_DELIVERY_MONO). This helper changes both the skb->tstamp and the
skb->mono_delivery_time. The expectation is this could change skb->tstamp in the
ingress skb and redirect to egress sch_fq. It could also set a mono time to
skb->tstamp where the udp sk->sk_clockid may not be necessary in mono and then
bpf_redirect to egress sch_fq. When bpf_skb_set_tstamp(skb, tstamp,
BPF_SKB_TSTAMP_DELIVERY_MONO) succeeds, reading __sk_buff->tstamp_type expects
BPF_SKB_TSTAMP_DELIVERY_MONO also.
I ran out of idea to solve this uapi breakage.
I am afraid it may need to go back to v1 idea and use another bit
(user_delivery_time) in the skb.
Is the only conflict when bpf_skb_set_tstamp is called for an skb from
a socket with sk_clockid set (and thus SO_TXTIME called)?
Right, because skb->mono_delivery_time does not mean skb->tstamp is mono now and
its interpretation depends on skb->sk->sk_clockid.
Interpreting skb->tstamp as mono if skb->mono_delivery_time is set and
skb->sk is NULL is fine. This is the ingress to egress redirect case.
skb->sk == NULL is fine. I tried something like this in
bpf_convert_tstamp_type_read() for reading __sk_buff->tstamp_type:
__sk_buff->tstamp_type is BPF_SKB_TSTAMP_DELIVERY_MONO when:
skb->mono_delivery_time == 1 &&
(!skb->sk ||
!sk_fullsock(skb->sk) /* tcp tw or req sk */ ||
skb->sk->sk_protocol == IPPROTO_TCP)
Not a small bpf instruction addition to bpf_convert_tstamp_type_read() but doable.
I don't see an immediate use for this BPF function on egress where it
would overwrite an SO_TXTIME timestamp and now skb->tstamp is mono,
but skb->sk != NULL and skb->sk->sk_clockid != CLOCK_MONOTONIC.
The bpf prog may act as a traffic shaper that limits the bandwidth usage of all
outgoing packets (tcp/udp/...) by setting the mono EDT in skb->tstamp before
sending to the sch_fq.
I currently also don't have a use case for skb->sk->sk_clockid !=
CLOCK_MONOTONIC. However, it is something that bpf_skb_set_tstamp() can do now
before queuing to sch_fq.
The container (in netns + veth) may use other sk_clockid/qdisc (e.g. sch_etf)
setup and the non mono skb->tstamp is not cleared now during dev_forward_skb()
between the veth pair.
Perhaps bpf_skb_set_tstamp() can just fail if another delivery time is
already explicitly programmed?
This will change the existing bpf_skb_set_tstamp() behavior, so probably not
acceptable.
skb->sk &&
sock_flag(sk, SOCK_TXTIME) &&
skb->sk->sk_clockid != CLOCK_MONOTONIC
Either that, or unset SOCK_TXTIME to make sk_clockid undefined and
fall back on interpreting as monotonic.
Change sk->sk_flags in tc bpf prog? hmm... I am not sure this will work well also.
sock_valbool_flag(SOCK_TXTIME) should require a lock_sock() to make changes. The
tc bpf prog cannot take the lock_sock, so bpf_skb_set_tstamp() currently only
changes skb and does not change skb->sk.
I think changing sock_valbool_flag(SOCK_TXTIME) will also have a new user space
visible side effect. The sendmsg for cmsg with SCM_TXTIME will start failing
from looking at __sock_cmsg_send().
There may be a short period of disconnect between what is in sk->sk_flags and
what is set in skb->tstamp. e.g. what if user space does setsockopt(SO_TXTIME)
again after skb->tstamp is set by bpf. This could be considered a small glitch
for some amount of skb(s) until the user space settled on setsockopt(SO_TXTIME).
I think all this is crying for another bit in skb to mean user_delivery_time
(skb->tstamp depends on skb->sk->sk_clockid) while mono_delivery_time is the
mono time either set by kernel-tcp or bpf. If we need to revert the
mono_delivery_time reuse patch later, we will need to revert the netdev patch
and the (to-be-made) bpf patch.
