On Thu, Mar 3, 2022 at 12:02 PM Alexei Starovoitov <alexei.starovoitov@xxxxxxxxx> wrote: > > On Thu, Mar 3, 2022 at 11:43 AM Hao Luo <haoluo@xxxxxxxxxx> wrote: > > > > On Wed, Mar 2, 2022 at 6:29 PM Alexei Starovoitov > > <alexei.starovoitov@xxxxxxxxx> wrote: > > > > > > On Wed, Mar 2, 2022 at 5:09 PM Yonghong Song <yhs@xxxxxx> wrote: > > > > > > > > > > > > > > > > On 3/2/22 1:30 PM, Alexei Starovoitov wrote: > > > > > On Wed, Mar 2, 2022 at 1:23 PM Yonghong Song <yhs@xxxxxx> wrote: > > > > >> > > > > >> > > > > >> > > > > >> On 2/25/22 3:43 PM, Hao Luo wrote: > > > > >>> Add a new type of bpf tracepoints: sleepable tracepoints, which allows > > > > >>> the handler to make calls that may sleep. With sleepable tracepoints, a > > > > >>> set of syscall helpers (which may sleep) may also be called from > > > > >>> sleepable tracepoints. > > > > >> > > > > >> There are some old discussions on sleepable tracepoints, maybe > > > > >> worthwhile to take a look. > > > > >> > > > > >> https://lore.kernel.org/bpf/20210218222125.46565-5-mjeanson@xxxxxxxxxxxx/T/ > > > > > > > > > > Right. It's very much related, but obsolete too. > > > > > We don't need any of that for sleeptable _raw_ tps. > > > > > I prefer to stay with "sleepable" name as well to > > > > > match the rest of the bpf sleepable code. > > > > > In all cases it's faultable. > > > > > > > > sounds good to me. Agree that for the bpf user case, Hao's > > > > implementation should be enough. > > > > > > Just remembered that we can also do trivial noinline __weak > > > nop function and mark it sleepable on the verifier side. > > > That's what we were planning to do to trace map update/delete ops > > > in Joe Burton's series. > > > Then we don't need to extend tp infra. > > > I'm fine whichever way. I see pros and cons in both options. > > > > Joe is also cc'ed in this patchset, I will sync up with him on the > > status of trace map work. > > > > Alexei, do we have potentially other variants of tp? We can make the > > current u16 sleepable a flag, so we can reuse this flag later when we > > have another type of tracepoints. > > When we added the ability to attach to kernel functions and mark them > as allow_error_inject the usefulness of tracepoints and even > writeable tracepoints was deminissed. > If we do sleepable tracepoint, I suspect, it may be the last extension > in that area. > I guess I'm convincing myself that noinline weak nop func > is better here. Just like it's better for Joe's map tracing. To add to the above... The only downside of sleepable nop func comparing to tp is the lack of static_branch. So this nop call will always be there. For map tracing and for cgroup mkdir/rmdir the few nanosecond overhead of calling an empty function isn't even measurable.
