Re: [PATCH] uprobes: reduce contention on uprobes_tree access
From: Jonthan Haslam
Date: Wed Apr 03 2024 - 07:05:30 EST
> > > > Given the discussion around per-cpu rw semaphore and need for
> > > > (internal) batched attachment API for uprobes, do you think you can
> > > > apply this patch as is for now? We can then gain initial improvements
> > > > in scalability that are also easy to backport, and Jonathan will work
> > > > on a more complete solution based on per-cpu RW semaphore, as
> > > > suggested by Ingo.
> > >
> > > Yeah, it is interesting to use per-cpu rw semaphore on uprobe.
> > > I would like to wait for the next version.
> >
> > My initial tests show a nice improvement on the over RW spinlocks but
> > significant regression in acquiring a write lock. I've got a few days
> > vacation over Easter but I'll aim to get some more formalised results out
> > to the thread toward the end of next week.
>
> As far as the write lock is only on the cold path, I think you can choose
> per-cpu RW semaphore. Since it does not do busy wait, the total system
> performance impact will be small.
> I look forward to your formalized results :)
Sorry for the delay in getting back to you on this Masami.
I have used one of the bpf selftest benchmarks to provide some form of
comparison of the 3 different approaches (spinlock, RW spinlock and
per-cpu RW semaphore). The benchmark used here is the 'trig-uprobe-nop'
benchmark which just executes a single uprobe with a minimal bpf program
attached. The tests were done on a 32 core qemu/kvm instance.
Things to note about the results:
- The results are slightly variable so don't get too caught up on
individual thread count - it's the trend that is important.
- In terms of throughput with this specific benchmark a *very* macro view
is that the RW spinlock provides 40-60% more throughput than the
spinlock. The per-CPU RW semaphore provides in the order of 50-100%
more throughput then the spinlock.
- This doesn't fully reflect the large reduction in latency that we have
seen in application based measurements. However, it does demonstrate
that even the trivial change of going to a RW spinlock provides
significant benefits.
I haven't included the measurements on per-CPU RW semaphore write
performance as they are completely in line with those that Paul McKenney
posted on his journal [0]. On a 32 core system I see semaphore writes to
take in the order of 25-28 millisecs - the cost of the synchronize_rcu().
Each block of results below show 1 line per execution of the benchmark (the
"Summary" line) and each line is a run with one more thread added - a
thread is a "producer". The lines are edited to remove extraneous output
that adds no value here.
The tests were executed with this driver script:
for num_threads in {1..20}
do
sudo ./bench -p $num_threads trig-uprobe-nop | grep Summary
done
spinlock
Summary: hits 1.453 ± 0.005M/s ( 1.453M/prod)
Summary: hits 2.087 ± 0.005M/s ( 1.043M/prod)
Summary: hits 2.701 ± 0.012M/s ( 0.900M/prod)
Summary: hits 1.917 ± 0.011M/s ( 0.479M/prod)
Summary: hits 2.105 ± 0.003M/s ( 0.421M/prod)
Summary: hits 1.615 ± 0.006M/s ( 0.269M/prod)
Summary: hits 2.046 ± 0.004M/s ( 0.292M/prod)
Summary: hits 1.818 ± 0.002M/s ( 0.227M/prod)
Summary: hits 1.867 ± 0.024M/s ( 0.207M/prod)
Summary: hits 1.692 ± 0.003M/s ( 0.169M/prod)
Summary: hits 1.778 ± 0.004M/s ( 0.162M/prod)
Summary: hits 1.710 ± 0.011M/s ( 0.142M/prod)
Summary: hits 1.687 ± 0.022M/s ( 0.130M/prod)
Summary: hits 1.697 ± 0.004M/s ( 0.121M/prod)
Summary: hits 1.645 ± 0.011M/s ( 0.110M/prod)
Summary: hits 1.671 ± 0.002M/s ( 0.104M/prod)
Summary: hits 1.692 ± 0.014M/s ( 0.100M/prod)
Summary: hits 1.700 ± 0.015M/s ( 0.094M/prod)
Summary: hits 1.668 ± 0.005M/s ( 0.088M/prod)
Summary: hits 1.644 ± 0.004M/s ( 0.082M/prod)
RW spinlock
Summary: hits 1.465 ± 0.008M/s ( 1.465M/prod)
Summary: hits 1.750 ± 0.035M/s ( 0.875M/prod)
Summary: hits 2.164 ± 0.008M/s ( 0.721M/prod)
Summary: hits 2.235 ± 0.014M/s ( 0.559M/prod)
Summary: hits 2.202 ± 0.005M/s ( 0.440M/prod)
Summary: hits 2.816 ± 0.003M/s ( 0.469M/prod)
Summary: hits 2.364 ± 0.002M/s ( 0.338M/prod)
Summary: hits 2.327 ± 0.008M/s ( 0.291M/prod)
Summary: hits 2.147 ± 0.005M/s ( 0.239M/prod)
Summary: hits 2.266 ± 0.011M/s ( 0.227M/prod)
Summary: hits 2.483 ± 0.003M/s ( 0.226M/prod)
Summary: hits 2.573 ± 0.008M/s ( 0.214M/prod)
Summary: hits 2.569 ± 0.004M/s ( 0.198M/prod)
Summary: hits 2.507 ± 0.013M/s ( 0.179M/prod)
Summary: hits 2.165 ± 0.008M/s ( 0.144M/prod)
Summary: hits 2.524 ± 0.004M/s ( 0.158M/prod)
Summary: hits 2.059 ± 0.020M/s ( 0.121M/prod)
Summary: hits 2.332 ± 0.007M/s ( 0.130M/prod)
Summary: hits 2.404 ± 0.017M/s ( 0.127M/prod)
Summary: hits 2.187 ± 0.002M/s ( 0.109M/prod)
per-CPU RW semaphore
Summary: hits 1.341 ± 0.017M/s ( 1.341M/prod)
Summary: hits 2.397 ± 0.011M/s ( 1.198M/prod)
Summary: hits 3.547 ± 0.041M/s ( 1.182M/prod)
Summary: hits 4.108 ± 0.016M/s ( 1.027M/prod)
Summary: hits 3.138 ± 0.055M/s ( 0.628M/prod)
Summary: hits 3.247 ± 0.017M/s ( 0.541M/prod)
Summary: hits 2.877 ± 0.005M/s ( 0.411M/prod)
Summary: hits 2.880 ± 0.002M/s ( 0.360M/prod)
Summary: hits 2.579 ± 0.001M/s ( 0.287M/prod)
Summary: hits 2.982 ± 0.011M/s ( 0.298M/prod)
Summary: hits 2.603 ± 0.002M/s ( 0.237M/prod)
Summary: hits 3.013 ± 0.004M/s ( 0.251M/prod)
Summary: hits 3.059 ± 0.001M/s ( 0.235M/prod)
Summary: hits 2.721 ± 0.014M/s ( 0.194M/prod)
Summary: hits 2.943 ± 0.005M/s ( 0.196M/prod)
Summary: hits 3.366 ± 0.011M/s ( 0.210M/prod)
Summary: hits 2.459 ± 0.001M/s ( 0.145M/prod)
Summary: hits 3.023 ± 0.024M/s ( 0.168M/prod)
Summary: hits 2.919 ± 0.002M/s ( 0.154M/prod)
Summary: hits 2.569 ± 0.002M/s ( 0.128M/prod)
[0] https://paulmck.livejournal.com/67547.html
Thanks.
Jon.
>
> Thank you,
>
> >
> > Jon.
> >
> > >
> > > Thank you,
> > >
> > > >
> > > > >
> > > > > BTW, how did you measure the overhead? I think spinlock overhead
> > > > > will depend on how much lock contention happens.
> > > > >
> > > > > Thank you,
> > > > >
> > > > > >
> > > > > > [0] https://docs.kernel.org/locking/spinlocks.html
> > > > > >
> > > > > > Signed-off-by: Jonathan Haslam <jonathan.haslam@xxxxxxxxx>
> > > > > > ---
> > > > > > kernel/events/uprobes.c | 22 +++++++++++-----------
> > > > > > 1 file changed, 11 insertions(+), 11 deletions(-)
> > > > > >
> > > > > > diff --git a/kernel/events/uprobes.c b/kernel/events/uprobes.c
> > > > > > index 929e98c62965..42bf9b6e8bc0 100644
> > > > > > --- a/kernel/events/uprobes.c
> > > > > > +++ b/kernel/events/uprobes.c
> > > > > > @@ -39,7 +39,7 @@ static struct rb_root uprobes_tree = RB_ROOT;
> > > > > > */
> > > > > > #define no_uprobe_events() RB_EMPTY_ROOT(&uprobes_tree)
> > > > > >
> > > > > > -static DEFINE_SPINLOCK(uprobes_treelock); /* serialize rbtree access */
> > > > > > +static DEFINE_RWLOCK(uprobes_treelock); /* serialize rbtree access */
> > > > > >
> > > > > > #define UPROBES_HASH_SZ 13
> > > > > > /* serialize uprobe->pending_list */
> > > > > > @@ -669,9 +669,9 @@ static struct uprobe *find_uprobe(struct inode *inode, loff_t offset)
> > > > > > {
> > > > > > struct uprobe *uprobe;
> > > > > >
> > > > > > - spin_lock(&uprobes_treelock);
> > > > > > + read_lock(&uprobes_treelock);
> > > > > > uprobe = __find_uprobe(inode, offset);
> > > > > > - spin_unlock(&uprobes_treelock);
> > > > > > + read_unlock(&uprobes_treelock);
> > > > > >
> > > > > > return uprobe;
> > > > > > }
> > > > > > @@ -701,9 +701,9 @@ static struct uprobe *insert_uprobe(struct uprobe *uprobe)
> > > > > > {
> > > > > > struct uprobe *u;
> > > > > >
> > > > > > - spin_lock(&uprobes_treelock);
> > > > > > + write_lock(&uprobes_treelock);
> > > > > > u = __insert_uprobe(uprobe);
> > > > > > - spin_unlock(&uprobes_treelock);
> > > > > > + write_unlock(&uprobes_treelock);
> > > > > >
> > > > > > return u;
> > > > > > }
> > > > > > @@ -935,9 +935,9 @@ static void delete_uprobe(struct uprobe *uprobe)
> > > > > > if (WARN_ON(!uprobe_is_active(uprobe)))
> > > > > > return;
> > > > > >
> > > > > > - spin_lock(&uprobes_treelock);
> > > > > > + write_lock(&uprobes_treelock);
> > > > > > rb_erase(&uprobe->rb_node, &uprobes_tree);
> > > > > > - spin_unlock(&uprobes_treelock);
> > > > > > + write_unlock(&uprobes_treelock);
> > > > > > RB_CLEAR_NODE(&uprobe->rb_node); /* for uprobe_is_active() */
> > > > > > put_uprobe(uprobe);
> > > > > > }
> > > > > > @@ -1298,7 +1298,7 @@ static void build_probe_list(struct inode *inode,
> > > > > > min = vaddr_to_offset(vma, start);
> > > > > > max = min + (end - start) - 1;
> > > > > >
> > > > > > - spin_lock(&uprobes_treelock);
> > > > > > + read_lock(&uprobes_treelock);
> > > > > > n = find_node_in_range(inode, min, max);
> > > > > > if (n) {
> > > > > > for (t = n; t; t = rb_prev(t)) {
> > > > > > @@ -1316,7 +1316,7 @@ static void build_probe_list(struct inode *inode,
> > > > > > get_uprobe(u);
> > > > > > }
> > > > > > }
> > > > > > - spin_unlock(&uprobes_treelock);
> > > > > > + read_unlock(&uprobes_treelock);
> > > > > > }
> > > > > >
> > > > > > /* @vma contains reference counter, not the probed instruction. */
> > > > > > @@ -1407,9 +1407,9 @@ vma_has_uprobes(struct vm_area_struct *vma, unsigned long start, unsigned long e
> > > > > > min = vaddr_to_offset(vma, start);
> > > > > > max = min + (end - start) - 1;
> > > > > >
> > > > > > - spin_lock(&uprobes_treelock);
> > > > > > + read_lock(&uprobes_treelock);
> > > > > > n = find_node_in_range(inode, min, max);
> > > > > > - spin_unlock(&uprobes_treelock);
> > > > > > + read_unlock(&uprobes_treelock);
> > > > > >
> > > > > > return !!n;
> > > > > > }
> > > > > > --
> > > > > > 2.43.0
> > > > > >
> > > > >
> > > > >
> > > > > --
> > > > > Masami Hiramatsu (Google) <mhiramat@xxxxxxxxxx>
> > >
> > >
> > > --
> > > Masami Hiramatsu (Google) <mhiramat@xxxxxxxxxx>
>
>
> --
> Masami Hiramatsu (Google) <mhiramat@xxxxxxxxxx>