bpf: Fix combination of jit blinding and pointers to bpf subprogs.

commit 4b6313c upstream.

The combination of jit blinding and pointers to bpf subprogs causes:
[   36.989548] BUG: unable to handle page fault for address: 0000000100000001
[   36.990342] #PF: supervisor instruction fetch in kernel mode
[   36.990968] #PF: error_code(0x0010) - not-present page
[   36.994859] RIP: 0010:0x100000001
[   36.995209] Code: Unable to access opcode bytes at RIP 0xffffffd7.
[   37.004091] Call Trace:
[   37.004351]  <TASK>
[   37.004576]  ? bpf_loop+0x4d/0x70
[   37.004932]  ? bpf_prog_3899083f75e4c5de_F+0xe3/0x13b

The jit blinding logic didn't recognize that ld_imm64 with an address
of bpf subprogram is a special instruction and proceeded to randomize it.
By itself it wouldn't have been an issue, but jit_subprogs() logic
relies on two step process to JIT all subprogs and then JIT them
again when addresses of all subprogs are known.
Blinding process in the first JIT phase caused second JIT to miss
adjustment of special ld_imm64.

Fix this issue by ignoring special ld_imm64 instructions that don't have
user controlled constants and shouldn't be blinded.

Fixes: 69c087b ("bpf: Add bpf_for_each_map_elem() helper")
Reported-by: Andrii Nakryiko <andrii@kernel.org>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Andrii Nakryiko <andrii@kernel.org>
Acked-by: Martin KaFai Lau <kafai@fb.com>
Link: https://lore.kernel.org/bpf/20220513011025.13344-1-alexei.starovoitov@gmail.com
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

kernel/bpf/core.c

@@ -1436,6 +1436,16 @@ struct bpf_prog *bpf_jit_blind_constants(struct bpf_prog *prog)
 	insn = clone->insnsi;
 
 	for (i = 0; i < insn_cnt; i++, insn++) {
+		if (bpf_pseudo_func(insn)) {
+			/* ld_imm64 with an address of bpf subprog is not
+			 * a user controlled constant. Don't randomize it,
+			 * since it will conflict with jit_subprogs() logic.
+			 */
+			insn++;
+			i++;
+			continue;
+		}
+
 		/* We temporarily need to hold the original ld64 insn
 		 * so that we can still access the first part in the
 		 * second blinding run.