On 10/26/24 2:30 AM, Martin KaFai Lau wrote:
On 10/25/24 4:05 AM, D. Wythe wrote:
Our main concern is to avoid introducing kfuncs as much as possible. For our subsystem, we might
need to maintain it in a way that maintains a uapi, as we certainly have user applications
depending on it.
The smc_bpf_ops can read/write the tp and ireq. In patch 4, there is 'tp->syn_smc = 1'. I assume the
real bpf prog will read something from the tp to make the decision also. Note that tp/ireq is also
not in the uapi but the CO-RE can help in case the tp->syn_smc bool is moved around.
From looking at the selftest in patch 4 again, I think all it needs is for the bpf prog (i.e. the
ops) to return a bool instead of allowing the bpf prog to write or call a kfunc to change the tp/ireq.
Hi Martin,
At the beginning, I did modify it by returning values, but later I wanted to make this ops more
universal, so I considered influencing the behavior by modifying the tp without returning any value.
But considering we currently do not have any other needs, perhaps modifying it by returning a value
would be more appropriate.
And If that's the case, we won't need to add new prog parameters to the struct_access anymore. I'll
try this in the next series.
Thanks,
D. Wythe
