bpf, arm64: Fixed a BTI error on returning to patched function

[ Upstream commit 738a96c ]

When BPF_TRAMP_F_CALL_ORIG is set, BPF trampoline uses BLR to jump
back to the instruction next to call site to call the patched function.
For BTI-enabled kernel, the instruction next to call site is usually
PACIASP, in this case, it's safe to jump back with BLR. But when
the call site is not followed by a PACIASP or bti, a BTI exception
is triggered.

Here is a fault log:

 Unhandled 64-bit el1h sync exception on CPU0, ESR 0x0000000034000002 -- BTI
 CPU: 0 PID: 263 Comm: test_progs Tainted: GF
 Hardware name: linux,dummy-virt (DT)
 pstate: 40400805 (nZcv daif +PAN -UAO -TCO -DIT -SSBS BTYPE=-c)
 pc : bpf_fentry_test1+0xc/0x30
 lr : bpf_trampoline_6442573892_0+0x48/0x1000
 sp : ffff80000c0c3a50
 x29: ffff80000c0c3a90 x28: ffff0000c2e6c080 x27: 0000000000000000
 x26: 0000000000000000 x25: 0000000000000000 x24: 0000000000000050
 x23: 0000000000000000 x22: 0000ffffcfd2a7f0 x21: 000000000000000a
 x20: 0000ffffcfd2a7f0 x19: 0000000000000000 x18: 0000000000000000
 x17: 0000000000000000 x16: 0000000000000000 x15: 0000ffffcfd2a7f0
 x14: 0000000000000000 x13: 0000000000000000 x12: 0000000000000000
 x11: 0000000000000000 x10: ffff80000914f5e4 x9 : ffff8000082a1528
 x8 : 0000000000000000 x7 : 0000000000000000 x6 : 0101010101010101
 x5 : 0000000000000000 x4 : 00000000fffffff2 x3 : 0000000000000001
 x2 : ffff8001f4b82000 x1 : 0000000000000000 x0 : 0000000000000001
 Kernel panic - not syncing: Unhandled exception
 CPU: 0 PID: 263 Comm: test_progs Tainted: GF
 Hardware name: linux,dummy-virt (DT)
 Call trace:
  dump_backtrace+0xec/0x144
  show_stack+0x24/0x7c
  dump_stack_lvl+0x8c/0xb8
  dump_stack+0x18/0x34
  panic+0x1cc/0x3ec
  __el0_error_handler_common+0x0/0x130
  el1h_64_sync_handler+0x60/0xd0
  el1h_64_sync+0x78/0x7c
  bpf_fentry_test1+0xc/0x30
  bpf_fentry_test1+0xc/0x30
  bpf_prog_test_run_tracing+0xdc/0x2a0
  __sys_bpf+0x438/0x22a0
  __arm64_sys_bpf+0x30/0x54
  invoke_syscall+0x78/0x110
  el0_svc_common.constprop.0+0x6c/0x1d0
  do_el0_svc+0x38/0xe0
  el0_svc+0x30/0xd0
  el0t_64_sync_handler+0x1ac/0x1b0
  el0t_64_sync+0x1a0/0x1a4
 Kernel Offset: disabled
 CPU features: 0x0000,00034c24,f994fdab
 Memory Limit: none

And the instruction next to call site of bpf_fentry_test1 is ADD,
not PACIASP:

<bpf_fentry_test1>:
	bti     c
	nop
	nop
	add     w0, w0, #0x1
	paciasp

For BPF prog, JIT always puts a PACIASP after call site for BTI-enabled
kernel, so there is no problem. To fix it, replace BLR with RET to bypass
the branch target check.

Fixes: efc9909 ("bpf, arm64: Add bpf trampoline for arm64")
Reported-by: Florent Revest <revest@chromium.org>
Signed-off-by: Xu Kuohai <xukuohai@huawei.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Tested-by: Florent Revest <revest@chromium.org>
Acked-by: Florent Revest <revest@chromium.org>
Link: https://lore.kernel.org/bpf/20230401234144.3719742-1-xukuohai@huaweicloud.com
Signed-off-by: Sasha Levin <sashal@kernel.org>

arch/arm64/net/bpf_jit.h

@@ -281,4 +281,8 @@
 /* DMB */
 #define A64_DMB_ISH aarch64_insn_gen_dmb(AARCH64_INSN_MB_ISH)
 
+/* ADR */
+#define A64_ADR(Rd, offset) \
+	aarch64_insn_gen_adr(0, offset, Rd, AARCH64_INSN_ADR_TYPE_ADR)
+
 #endif /* _BPF_JIT_H */

arch/arm64/net/bpf_jit_comp.c

@@ -1900,7 +1900,8 @@ static int prepare_trampoline(struct jit_ctx *ctx, struct bpf_tramp_image *im,
 		restore_args(ctx, args_off, nargs);
 		/* call original func */
 		emit(A64_LDR64I(A64_R(10), A64_SP, retaddr_off), ctx);
-		emit(A64_BLR(A64_R(10)), ctx);
+		emit(A64_ADR(A64_LR, AARCH64_INSN_SIZE * 2), ctx);
+		emit(A64_RET(A64_R(10)), ctx);
 		/* store return value */
 		emit(A64_STR64I(A64_R(0), A64_SP, retval_off), ctx);
 		/* reserve a nop for bpf_tramp_image_put */