bpf: avoid excessive stack usage for perf_sample_data

perf_sample_data consumes 386 bytes on stack, reduce excessive stack
usage and move it to per cpu buffer. It's allowed due to preemption
being disabled for tracing, xdp and tc programs, thus at all times
only one program can run on a specific CPU and programs cannot run
from interrupt. We similarly also handle bpf_pt_regs.

Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

1 files changed, 6 insertions, 4 deletions

diff --git a/kernel/trace/bpf_trace.c b/kernel/trace/bpf_trace.c
index 08eb072430b9..051d7fca0c09 100644
--- a/kernel/trace/bpf_trace.c
+++ b/kernel/trace/bpf_trace.c
@@ -266,14 +266,16 @@ static const struct bpf_func_proto bpf_perf_event_read_proto = {
 	.arg2_type	= ARG_ANYTHING,
 };
 
+static DEFINE_PER_CPU(struct perf_sample_data, bpf_sd);
+
 static __always_inline u64
 __bpf_perf_event_output(struct pt_regs *regs, struct bpf_map *map,
 			u64 flags, struct perf_raw_record *raw)
 {
 	struct bpf_array *array = container_of(map, struct bpf_array, map);
+	struct perf_sample_data *sd = this_cpu_ptr(&bpf_sd);
 	unsigned int cpu = smp_processor_id();
 	u64 index = flags & BPF_F_INDEX_MASK;
-	struct perf_sample_data sample_data;
 	struct bpf_event_entry *ee;
 	struct perf_event *event;
 
@@ -294,9 +296,9 @@ __bpf_perf_event_output(struct pt_regs *regs, struct bpf_map *map,
 	if (unlikely(event->oncpu != cpu))
 		return -EOPNOTSUPP;
 
-	perf_sample_data_init(&sample_data, 0, 0);
-	sample_data.raw = raw;
-	perf_event_output(event, &sample_data, regs);
+	perf_sample_data_init(sd, 0, 0);
+	sd->raw = raw;
+	perf_event_output(event, sd, regs);
 	return 0;
 }