bpf: Fix a data-race around bpf_jit_limit.

While reading bpf_jit_limit, it can be changed concurrently via sysctl,
WRITE_ONCE() in __do_proc_doulongvec_minmax(). The size of bpf_jit_limit
is long, so we need to add a paired READ_ONCE() to avoid load-tearing.

Fixes: ede95a63b5e8 ("bpf: add bpf_jit_limit knob to restrict unpriv allocations")
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Link: https://lore.kernel.org/bpf/20220823215804.2177-1-kuniyu@amazon.com

diff --git a/kernel/bpf/core.c b/kernel/bpf/core.c
index c1e10d088dbb7388c44b17a01711c0b5d505ddb4..3d9eb3ae334ce2738213e7c89ef63cf01a42a81e 100644 (file)
--- a/kernel/bpf/core.c
+++ b/kernel/bpf/core.c
@@ -971,7 +971,7 @@ pure_initcall(bpf_jit_charge_init);
 
 int bpf_jit_charge_modmem(u32 size)
 {
-       if (atomic_long_add_return(size, &bpf_jit_current) > bpf_jit_limit) {
+       if (atomic_long_add_return(size, &bpf_jit_current) > READ_ONCE(bpf_jit_limit)) {
                if (!bpf_capable()) {
                        atomic_long_sub(size, &bpf_jit_current);
                        return -EPERM;