uprobes: turn trace_uprobe's nhit counter to be per-CPU one

trace_uprobe->nhit counter is not incremented atomically, so its value
is questionable in when uprobe is hit on multiple CPUs simultaneously.

Also, doing this shared counter increment across many CPUs causes heavy
cache line bouncing, limiting uprobe/uretprobe performance scaling with
number of CPUs.

Solve both problems by making this a per-CPU counter.

Link: https://lore.kernel.org/all/[e-mail]/

Reviewed-by: Oleg Nesterov <[e-mail]>
Signed-off-by: Andrii Nakryiko <[e-mail]>
Signed-off-by: Masami Hiramatsu (Google) <[e-mail]>

kernel/trace/trace_uprobe.c

···
 #include <linux/string.h>
 #include <linux/rculist.h>
 #include <linux/filter.h>
+#include <linux/percpu.h>
 
 #include "trace_dynevent.h"
 #include "trace_probe.h"
···
 	char				*filename;
 	unsigned long			offset;
 	unsigned long			ref_ctr_offset;
-	unsigned long			nhit;
+	unsigned long __percpu		*nhits;
 	struct trace_probe		tp;
 };
 
···
 	if (!tu)
 		return ERR_PTR(-ENOMEM);
 
+	tu->nhits = alloc_percpu(unsigned long);
+	if (!tu->nhits) {
+		ret = -ENOMEM;
+		goto error;
+	}
+
 	ret = trace_probe_init(&tu->tp, event, group, true, nargs);
 	if (ret < 0)
 		goto error;
···
 	return tu;
 
 error:
+	free_percpu(tu->nhits);
 	kfree(tu);
 
 	return ERR_PTR(ret);
···
 	path_put(&tu->path);
 	trace_probe_cleanup(&tu->tp);
 	kfree(tu->filename);
+	free_percpu(tu->nhits);
 	kfree(tu);
 }
 
···
 {
 	struct dyn_event *ev = v;
 	struct trace_uprobe *tu;
+	unsigned long nhits;
+	int cpu;
 
 	if (!is_trace_uprobe(ev))
 		return 0;
 
 	tu = to_trace_uprobe(ev);
+
+	nhits = 0;
+	for_each_possible_cpu(cpu) {
+		nhits += per_cpu(*tu->nhits, cpu);
+	}
+
 	seq_printf(m, "  %s %-44s %15lu\n", tu->filename,
-			trace_probe_name(&tu->tp), tu->nhit);
+		   trace_probe_name(&tu->tp), nhits);
 	return 0;
 }
 
···
 	int ret = 0;
 
 	tu = container_of(con, struct trace_uprobe, consumer);
-	tu->nhit++;
+
+	this_cpu_inc(*tu->nhits);
 
 	udd.tu = tu;
 	udd.bp_addr = instruction_pointer(regs);