On Thu, Sep 05, 2019 at 10:13:10AM +0200, Ingo Molnar wrote: > > * Alexei Starovoitov <alexei.starovoi...@gmail.com> wrote: > > > On Wed, Sep 4, 2019 at 10:47 AM Peter Zijlstra <pet...@infradead.org> wrote: > > > > > > On Wed, Sep 04, 2019 at 08:51:21AM -0700, Alexei Starovoitov wrote: > > > > Anything in tracing can be deleted. > > > > Tracing is about debugging and introspection. > > > > When underlying kernel code changes the introspection points change as > > > > well. > > > > > > Right; except when it breaks widely used tools; like say powertop. Been > > > there, done that. > > > > powertop was a lesson learned, but it's not a relevant example anymore. > > There are more widely used tools today. Like bcc tools. > > And bpftrace is quickly gaining momentum and large user base. > > bcc tools did break already several times and people fixed them. > > Are these tools using libtraceevents?
bcc tools and bpftrace are using libbcc. Which in turn is using libbpf. libtraceevents is not used. Interesting example is https://github.com/iovisor/bcc/blob/master/tools/tcplife.py It's using "inet_sock_set_state" tracepoint when available on newer kernels and kprobe in tcp_set_state() function on older kernels. That tracepoint changed significantly over time. It had different name 'tcp_set_state' and slightly different semantics. Hence the tool was fixed when that change in tracepoint happened: https://github.com/iovisor/bcc/commit/fd93dc0409b626b749b90f115d3d550a870ed125 Note that tcp:tcp_set_state tracepoint existed for full kernel release. Yet people didn't make fuzz about the fact it disappeared in 4.16. Though tcplife.py tool is simple there are more complex tools based on this idea that are deployed in netflix and fb that went through the same tcp_set_state->inet_sock_set_state fixes.
