On Thu, Jul 11, 2024 at 4:07 AM Peter Zijlstra [off-list ref] wrote:

Hi!

These patches implement the (S)RCU based proposal to optimize uprobes.

On my c^Htrusty old IVB-EP -- where each (of the 40) CPU calls 'func' in a
tight loop:

  perf probe -x ./uprobes test=func
  perf stat -ae probe_uprobe:test  -- sleep 1

  perf probe -x ./uprobes test=func%return
  perf stat -ae probe_uprobe:test__return -- sleep 1

PRE:

  4,038,804      probe_uprobe:test
  2,356,275      probe_uprobe:test__return

POST:

  7,216,579      probe_uprobe:test
  6,744,786      probe_uprobe:test__return

(copy-paste FTW, I didn't do new numbers because the fast paths didn't change --
 and quick test run shows similar numbers)

Patches also available here:

  git://git.kernel.org/pub/scm/linux/kernel/git/peterz/queue.git perf/uprobes


Changes since last time:
 - better split with intermediate inc_not_zero()
 - fix UPROBE_HANDLER_REMOVE
 - restored the lost rcu_assign_pointer()
 - avoid lockdep for uretprobe_srcu
 - add missing put_uprobe() -> srcu_read_unlock() conversion
 - actually initialize return_instance::has_ref
 - a few comments
 - things I don't remember

Hey Peter!

Thanks for the v2, I plan to look at it more thoroughly tomorrow. But
meanwhile I spent a good chunk of today to write an uprobes
stress-test, so we can validate that we are not regressing anything
(yes, I don't trust lockless code and people in general ;)

Anyways, if you'd like to use it, it's at [0]. All you should need to
build and run it is:

  $ cd examples/c
  $ make -j$(nproc) uprobe-stress
  $ sudo ./uprobe-stress -tN -aM -mP -fR


N, M, P, R are number of threads dedicated to one of four functions of
the stress test: triggering user space functions (N),
attaching/detaching various random subsets of uprobes (M), mmap()ing
parts of executable with uprobes (P), and forking the process and
triggering uprobes for a little bit (R). The idea is to test various
timings and interleavings of uprobe-related logic.

You should only need not-too-old Clang to build everything (Clang 12+
should work, I believe). But do let me know if you run into troubles.

I did run this stress test for a little while on current
bpf-next/master with no issues detected (yay!).

But then I also ran it on Linux built from perf/uprobes branch (these
patches), and after a few seconds I see that there is no more
attachment/detachment happening. Eventually I got splats, which you
can see in [1]. I used `sudo ./uprobe-stress -a10 -t5 -m5 -f3` command
to run it inside my QEMU image.

So there is still something off, hopefully this will help to debug and
hammer out any remaining kinks. Thanks!

  [0] https://github.com/libbpf/libbpf-bootstrap/commit/2f88cef90f9728ec8c7bee7bd48fdbcf197806c3
  [1] https://gist.github.com/anakryiko/f761690addf7aa5f08caec95fda9ef1a