This patch set is heavily inspired by Peter Zijlstra's uprobe optimization
patches ([0]) and continues that work, albeit trying to keep complexity to the
minimum, and attepting to reuse existing primitives as much as possible. The
goal here is to optimize common uprobe triggering hot path, while keeping the
rest of locking mostly intact.
I've added uprobe_unregister_sync() into the error handling code path inside
uprobe_unregister(). This is due to recent refactorings from Oleg Nesterov
([1]), which necessitates this addition.
Except for refcounting change patch (which I stongly believe is a good
improvement we should do and forget about quasi-refcounting schema of
uprobe->consumers list), the rest of the changes are similar to Peter's
initial changes in [0].
Main differences would be:
- no special RCU protection for mmap and fork handling, we just stick to
refcounts there, as those are infrequent and not a performance-sensitive
code path, while being complex and thus benefiting from proper locking;
- the above means we don't need to do any custom SRCU additions to handle
forking code path;
- I handled UPROBE_HANDLER_REMOVE problem in handler_chain() differently,
again, leveraging existing locking scheam;
- I kept refcount usage for uretprobe and single-stepping uprobes, I plan to
address that in a separate follow up patches. The plan is to avoid
task_work with a lockless atomic xchg()/cmpxchg()-based solution;
- finally, I dutifully was using SRCU throughout all the changes, and only
last patch switches SRCU to RCU Tasks Trace and demonstrates significant
performance and scalability gains from this.
The changes in this patch set were tested using BPF selftests and using
uprobe-stress ([2]) tool.
Now, for the benchmarking results. I've used the following script (which
utilizes BPF selftests-based bench tool). The CPU used was 80-core Intel Xeon
Gold 6138 CPU @ 2.00GHz running kernel with production-like config. I minimized
background noise by stopping any service I could identify and stop, so results
are pretty stable and variability is pretty small, overall.
Benchmark script:
#!/bin/bash
set -eufo pipefail
for i in uprobe-nop uretprobe-nop; do
for p in 1 2 4 8 16 32 64; do
summary=$(sudo ./bench -w3 -d5 -p$p -a trig-$i | tail -n1)
total=$(echo "$summary" | cut -d'(' -f1 | cut -d' ' -f3-)
percpu=$(echo "$summary" | cut -d'(' -f2 | cut -d')' -f1 | cut -d'/' -f1)
printf "%-15s (%2d cpus): %s (%s/s/cpu)\n" $i $p "$total" "$percpu"
done
echo
done
With all the lock-avoiding changes done in this patch set, we get a pretty
decent improvement in performance and scalability of uprobes with number of
CPUs, even though we are still nowhere near linear scalability. This is due to
the remaining mmap_lock, which is currently taken to resolve interrupt address
to inode+offset and then uprobe instance. And, of course, uretprobes still need
similar RCU to avoid refcount in the hot path, which will be addressed in the
follow up patches.
BASELINE (on top of Oleg's clean up patches)
============================================
uprobe-nop ( 1 cpus): 3.032 ± 0.023M/s ( 3.032M/s/cpu)
uprobe-nop ( 2 cpus): 3.452 ± 0.005M/s ( 1.726M/s/cpu)
uprobe-nop ( 4 cpus): 3.663 ± 0.005M/s ( 0.916M/s/cpu)
uprobe-nop ( 8 cpus): 3.718 ± 0.038M/s ( 0.465M/s/cpu)
uprobe-nop (16 cpus): 3.344 ± 0.008M/s ( 0.209M/s/cpu)
uprobe-nop (32 cpus): 2.288 ± 0.021M/s ( 0.071M/s/cpu)
uprobe-nop (64 cpus): 3.205 ± 0.004M/s ( 0.050M/s/cpu)
uretprobe-nop ( 1 cpus): 1.979 ± 0.005M/s ( 1.979M/s/cpu)
uretprobe-nop ( 2 cpus): 2.361 ± 0.005M/s ( 1.180M/s/cpu)
uretprobe-nop ( 4 cpus): 2.309 ± 0.002M/s ( 0.577M/s/cpu)
uretprobe-nop ( 8 cpus): 2.253 ± 0.001M/s ( 0.282M/s/cpu)
uretprobe-nop (16 cpus): 2.007 ± 0.000M/s ( 0.125M/s/cpu)
uretprobe-nop (32 cpus): 1.624 ± 0.003M/s ( 0.051M/s/cpu)
uretprobe-nop (64 cpus): 2.149 ± 0.001M/s ( 0.034M/s/cpu)
Up to second-to-last patch (i.e., SRCU-based optimizations)
===========================================================
uprobe-nop ( 1 cpus): 3.276 ± 0.005M/s ( 3.276M/s/cpu)
uprobe-nop ( 2 cpus): 4.125 ± 0.002M/s ( 2.063M/s/cpu)
uprobe-nop ( 4 cpus): 7.713 ± 0.002M/s ( 1.928M/s/cpu)
uprobe-nop ( 8 cpus): 8.097 ± 0.006M/s ( 1.012M/s/cpu)
uprobe-nop (16 cpus): 6.501 ± 0.056M/s ( 0.406M/s/cpu)
uprobe-nop (32 cpus): 4.398 ± 0.084M/s ( 0.137M/s/cpu)
uprobe-nop (64 cpus): 6.452 ± 0.000M/s ( 0.101M/s/cpu)
uretprobe-nop ( 1 cpus): 2.055 ± 0.001M/s ( 2.055M/s/cpu)
uretprobe-nop ( 2 cpus): 2.677 ± 0.000M/s ( 1.339M/s/cpu)
uretprobe-nop ( 4 cpus): 4.561 ± 0.003M/s ( 1.140M/s/cpu)
uretprobe-nop ( 8 cpus): 5.291 ± 0.002M/s ( 0.661M/s/cpu)
uretprobe-nop (16 cpus): 5.065 ± 0.019M/s ( 0.317M/s/cpu)
uretprobe-nop (32 cpus): 3.622 ± 0.003M/s ( 0.113M/s/cpu)
uretprobe-nop (64 cpus): 3.723 ± 0.002M/s ( 0.058M/s/cpu)
RCU Tasks Trace
===============
uprobe-nop ( 1 cpus): 3.396 ± 0.002M/s ( 3.396M/s/cpu)
uprobe-nop ( 2 cpus): 4.271 ± 0.006M/s ( 2.135M/s/cpu)
uprobe-nop ( 4 cpus): 8.499 ± 0.015M/s ( 2.125M/s/cpu)
uprobe-nop ( 8 cpus): 10.355 ± 0.028M/s ( 1.294M/s/cpu)
uprobe-nop (16 cpus): 7.615 ± 0.099M/s ( 0.476M/s/cpu)
uprobe-nop (32 cpus): 4.430 ± 0.007M/s ( 0.138M/s/cpu)
uprobe-nop (64 cpus): 6.887 ± 0.020M/s ( 0.108M/s/cpu)
uretprobe-nop ( 1 cpus): 2.174 ± 0.001M/s ( 2.174M/s/cpu)
uretprobe-nop ( 2 cpus): 2.853 ± 0.001M/s ( 1.426M/s/cpu)
uretprobe-nop ( 4 cpus): 4.913 ± 0.002M/s ( 1.228M/s/cpu)
uretprobe-nop ( 8 cpus): 5.883 ± 0.002M/s ( 0.735M/s/cpu)
uretprobe-nop (16 cpus): 5.147 ± 0.001M/s ( 0.322M/s/cpu)
uretprobe-nop (32 cpus): 3.738 ± 0.008M/s ( 0.117M/s/cpu)
uretprobe-nop (64 cpus): 4.397 ± 0.002M/s ( 0.069M/s/cpu)
For baseline vs SRCU, peak througput increased from 3.7 M/s (million uprobe
triggerings per second) up to about 8 M/s. For uretprobes it's a bit more
modest with bump from 2.4 M/s to 5 M/s.
For SRCU vs RCU Tasks Trace, peak throughput for uprobes increases further from
8 M/s to 10.3 M/s (+28%!), and for uretprobes from 5.3 M/s to 5.8 M/s (+11%),
as we have more work to do on uretprobes side.
Even single-thread (no contention) performance is slightly better: 3.276 M/s to
3.396 M/s (+3.5%) for uprobes, and 2.055 M/s to 2.174 M/s (+5.8%)
for uretprobes.
[0] https://lore.kernel.org/linux-trace-kernel/20240711110235.098009979@xxxxxxxxxxxxx/
[1] https://lore.kernel.org/linux-trace-kernel/20240729134444.GA12293@xxxxxxxxxx/
[2] https://github.com/libbpf/libbpf-bootstrap/tree/uprobe-stress
v4->v5:
- added missing linux/rcupdate_trace.h include (Kernel Test Bot);
- added comment in uprobe_consumer about expected consistency between
handler() returning UPROBE_HANDLER_REMOVE and the need for filter() to
confirm the decision (Oleg);
- this time added Oleg's Reviewed-by;
v3->v4:
- added back consumer_rwsem into consumer_del(), which was accidentally
omitted earlier (Jiri);
- left out RFC patches, we can continue discussion on v3 patch set, if
necessary;
v2->v3:
- undid rcu and rb_node fields colocation which were causing crashes (Oleg);
- ensure synchronize_srcu() on registration failure in patch #4 (Oleg);
v1->v2:
- added back missed kfree() in patch #1 (Oleg);
- forgot the rest, but there were a few small things here and there.
Andrii Nakryiko (6):
uprobes: revamp uprobe refcounting and lifetime management
uprobes: protected uprobe lifetime with SRCU
uprobes: get rid of enum uprobe_filter_ctx in uprobe filter callbacks
uprobes: travers uprobe's consumer list locklessly under SRCU
protection
uprobes: perform lockless SRCU-protected uprobes_tree lookup
uprobes: switch to RCU Tasks Trace flavor for better performance
Peter Zijlstra (2):
perf/uprobe: split uprobe_unregister()
rbtree: provide rb_find_rcu() / rb_find_add_rcu()
include/linux/rbtree.h | 67 +++
include/linux/uprobes.h | 28 +-
kernel/events/uprobes.c | 392 +++++++++++-------
kernel/trace/bpf_trace.c | 8 +-
kernel/trace/trace_uprobe.c | 15 +-
.../selftests/bpf/bpf_testmod/bpf_testmod.c | 3 +-
6 files changed, 330 insertions(+), 183 deletions(-)
--
2.43.5
