On Wed, Aug 11, 2021 at 03:48:36PM -0700, Cong Wang wrote: > On Wed, Aug 11, 2021 at 2:23 PM Martin KaFai Lau <kafai@xxxxxx> wrote: > > Instead of adding tracepoints, the bpf fexit prog can be used here and > > the bpf prog will have the sk, skb, and ret available (example in fexit_test.c). > > Some tracepoints in this set can also be done with bpf fentry/fexit. > > Does bpf fentry/fexit work for your use case? > > Well, kprobe works too in this perspective. The problem with kprobe > or fexit is that there is no guarantee the function still exists in kernel > during iteration. Kernel is free to delete or rename it. With tracepoint, > even if ip_queue_xmit() were renamed, the same tracepoint must > remain in the kernel. Some of the function names are hardly changed. Considering it is not always cost free based on another thread, this is not a strong enough reason to add so many tracepoints while other options are available.
