On Thu, Oct 8, 2020 at 12:40 AM Michel Lespinasse <walken@xxxxxxxxxx> wrote:
>
> On Wed, Oct 7, 2020 at 11:44 AM Axel Rasmussen <axelrasmussen@xxxxxxxxxx> wrote:
> > The goal of these tracepoints is to be able to debug lock contention
> > issues. This lock is acquired on most (all?) mmap / munmap / page fault
> > operations, so a multi-threaded process which does a lot of these can
> > experience significant contention.
> >
> > We trace just before we start acquisition, when the acquisition returns
> > (whether it succeeded or not), and when the lock is released (or
> > downgraded). The events are broken out by lock type (read / write).
> >
> > The events are also broken out by memcg path. For container-based
> > workloads, users often think of several processes in a memcg as a single
> > logical "task", so collecting statistics at this level is useful.
> >
> > The end goal is to get latency information. This isn't directly included
> > in the trace events. Instead, users are expected to compute the time
> > between "start locking" and "acquire returned", using e.g. synthetic
> > events or BPF. The benefit we get from this is simpler code.
> >
> > Because we use tracepoint_enabled() to decide whether or not to trace,
> > this patch has effectively no overhead unless tracepoints are enabled at
> > runtime. If tracepoints are enabled, there is a performance impact, but
> > how much depends on exactly what e.g. the BPF program does.
> >
> > Signed-off-by: Axel Rasmussen <axelrasmussen@xxxxxxxxxx>
>
> Thanks for working on this.
>
> I like that there is no overhead unless CONFIG_TRACING is set.
Actually, it's even a bit better. Even if CONFIG_TRACING is set, the
only overhead we add is a single static_key branch (much cheaper than
a normal branch). The overhead is only introduced when the tracepoints
are enabled at runtime, e.g. by writing to
/sys/kernel/trace/events/(event)/enabled.
> However, I think the __mmap_lock_traced_lock and similar functions are
> the wrong level of abstraction, especially considering that we are
> considering to switch away from using rwsem as the underlying lock
> implementation. Would you consider something along the following lines
> instead for include/linux/mmap_lock.h ?
Sure, there's no reason that can't work. It's just a tradeoff:
Breaking up these helpers the way you describe makes the locking
functions more verbose (more repeated code; overall this file would be
longer), but there are a few less lines to change when we switch from
rwsem to something else (because we'd only need to change the lock
functions vs. as-is we'd have to change a couple of lines in each of
the helpers too).
I prefer the way it is, but I don't feel that strongly about it. I'll
refactor and send a v3 that looks as you describe.
>
> #ifdef CONFIG_TRACING
>
> DECLARE_TRACEPOINT(...);
>
> void __mmap_lock_do_trace_start_locking(struct mm_struct *mm, bool write);
>
> static inline void mmap_lock_trace_start_locking(struct mm_struct *mm,
> bool write)
> {
> if (tracepoint_enabled(mmap_lock_start_locking))
> __mmap_lock_do_trace_start_locking(mm, write);
> }
>
> #else
>
> static inline void mmap_lock_trace_start_locking(struct mm_struct *mm,
> bool write) {}
>
> #endif
>
> static inline void mmap_write_lock(struct mm_struct *mm)
> {
> mmap_lock_trace_start_locking(mm, true);
> down_write(&mm->mmap_lock);
> mmap_lock_trace_acquire_returned(mm, true, true);
> }
>
> I think this is more straightforward, and also the
> mmap_lock_trace_start_locking and similar functions don't depend on
> the underlying lock implementation.
>
> The changes to the other files look fine to me.
>
> --
> Michel "Walken" Lespinasse
> A program is never fully debugged until the last user dies.
