[PATCH v2 3/3] uprobes: add speculative lockless system-wide uprobe filter check

[Andrii Nakryiko]

It's very common with BPF-based uprobe/uretprobe use cases to have
a system-wide (not PID specific) probes used. In this case uprobe's
trace_uprobe_filter->nr_systemwide counter is bumped at registration
time, and actual filtering is short circuited at the time when
uprobe/uretprobe is triggered.

This is a great optimization, and the only issue with it is that to even
get to checking this counter uprobe subsystem is taking
read-side trace_uprobe_filter->rwlock. This is actually noticeable in
profiles and is just another point of contention when uprobe is
triggered on multiple CPUs simultaneously.

This patch moves this nr_systemwide check outside of filter list's
rwlock scope, as rwlock is meant to protect list modification, while
nr_systemwide-based check is speculative and racy already, despite the
lock (as discussed in [0]). trace_uprobe_filter_remove() and
trace_uprobe_filter_add() already check for filter->nr_systewide
explicitly outside of __uprobe_perf_filter, so no modifications are
required there.

Confirming with BPF selftests's based benchmarks.

BEFORE (based on changes in previous patch)
===========================================
uprobe-nop     :    2.732 ± 0.022M/s
uprobe-push    :    2.621 ± 0.016M/s
uprobe-ret     :    1.105 ± 0.007M/s
uretprobe-nop  :    1.396 ± 0.007M/s
uretprobe-push :    1.347 ± 0.008M/s
uretprobe-ret  :    0.800 ± 0.006M/s

AFTER
=====
uprobe-nop     :    2.878 ± 0.017M/s (+5.5%, total +8.3%)
uprobe-push    :    2.753 ± 0.013M/s (+5.3%, total +10.2%)
uprobe-ret     :    1.142 ± 0.010M/s (+3.8%, total +3.8%)
uretprobe-nop  :    1.444 ± 0.008M/s (+3.5%, total +6.5%)
uretprobe-push :    1.410 ± 0.010M/s (+4.8%, total +7.1%)
uretprobe-ret  :    0.816 ± 0.002M/s (+2.0%, total +3.9%)

In the above, first percentage value is based on top of previous patch
(lazy uprobe buffer optimization), while the "total" percentage is
based on kernel without any of the changes in this patch set.

As can be seen, we get about 4% - 10% speed up, in total, with both lazy
uprobe buffer and speculative filter check optimizations.

  [0] https://lore.kernel.org/bpf/20240313131926.GA19986@xxxxxxxxxx/

Reviewed-by: Jiri Olsa <jolsa@xxxxxxxxxx>
Signed-off-by: Andrii Nakryiko <andrii@xxxxxxxxxx>
---
 kernel/trace/trace_uprobe.c | 10 +++++++---
 1 file changed, 7 insertions(+), 3 deletions(-)

diff --git a/kernel/trace/trace_uprobe.c b/kernel/trace/trace_uprobe.c
index b5da95240a31..ac05885a6ce6 100644
--- a/kernel/trace/trace_uprobe.c
+++ b/kernel/trace/trace_uprobe.c
@@ -1226,9 +1226,6 @@ __uprobe_perf_filter(struct trace_uprobe_filter *filter, struct mm_struct *mm)
 {
 	struct perf_event *event;
 
-	if (filter->nr_systemwide)
-		return true;
-
 	list_for_each_entry(event, &filter->perf_events, hw.tp_list) {
 		if (event->hw.target->mm == mm)
 			return true;
@@ -1353,6 +1350,13 @@ static bool uprobe_perf_filter(struct uprobe_consumer *uc,
 	tu = container_of(uc, struct trace_uprobe, consumer);
 	filter = tu->tp.event->filter;
 
+	/*
+	 * speculative short-circuiting check to avoid unnecessarily taking
+	 * filter->rwlock below, if the uprobe has system-wide consumer
+	 */
+	if (READ_ONCE(filter->nr_systemwide))
+		return true;
+
 	read_lock(&filter->rwlock);
 	ret = __uprobe_perf_filter(filter, mm);
 	read_unlock(&filter->rwlock);
-- 
2.43.0