Hi, On Fri, Mar 3, 2023 at 10:52 AM Andreas Gruenbacher <agrue...@redhat.com> wrote: > > Now, let me get to the core of the matter. We've been talking about > using user-space (SDT) trace points for collecting the data, and I still > think that that's what we should do instead of introducing a new > dlm_controld log file. In the dlm_controld code, this would look like > the below patch. > > Note that <sys/sdt.h> is part of the systemtap-sdt-devel package, so a > "BuildRequires: systemtap-sdt-devel" dependency will be needed in > dlm.spec. >
ah, ok. This answers my other question I had. > With that, we can use standard tools like perf, bpftrace, etc. for > collecting all the relevant information without any further > modifications to dlm_controld. We can also collect additional kernel > and user-space trace point data at the same time with very little > additional effort. > > For example, here is how to register the four plock dlm_controld trace > points in perf: > > for ev in \ > sdt_dlm_controld:plock_lock_begin \ > sdt_dlm_controld:plock_lock_end \ > sdt_dlm_controld:plock_wait_begin \ > sdt_dlm_controld:plock_wait_end; do \ > perf probe -x /usr/sbin/dlm_controld $ev; \ > done > > The events can then be recorded with "perf record": > > perf record \ > -e sdt_dlm_controld:plock_lock_begin \ > -e sdt_dlm_controld:plock_lock_end \ > -e sdt_dlm_controld:plock_wait_begin \ > -e sdt_dlm_controld:plock_wait_end \ > [...] > > We've already gone through how the resulting log can be processed with > "perf script". One possible result would be the log file format that > lockdb_plot expects, but there are countless other possibilities. > > Other useful "tricks": > > $ bpftrace -l 'usdt:/usr/sbin/dlm_controld:*' > > $ readelf -n /usr/sbin/dlm_controld | sed -ne '/\.note\.stapsdt/,/^$/p' > look easy enough. - Alex
