powerpc64/bpf: Limit 'ldbrx' to processors compliant with ISA v2.06

Johan reported the below crash with test_bpf on ppc64 e5500:

  test_bpf: #296 ALU_END_FROM_LE 64: 0x0123456789abcdef -> 0x67452301 jited:1
  Oops: Exception in kernel mode, sig: 4 [#1]
  BE PAGE_SIZE=4K SMP NR_CPUS=24 QEMU e500
  Modules linked in: test_bpf(+)
  CPU: 0 PID: 76 Comm: insmod Not tainted 5.14.0-03771-g98c2059e008a-dirty #1
  NIP:  8000000000061c3c LR: 80000000006dea64 CTR: 8000000000061c18
  REGS: c0000000032d3420 TRAP: 0700   Not tainted (5.14.0-03771-g98c2059e008a-dirty)
  MSR:  0000000080089000 <EE,ME>  CR: 88002822  XER: 20000000 IRQMASK: 0
  <...>
  NIP [8000000000061c3c] 0x8000000000061c3c
  LR [80000000006dea64] .__run_one+0x104/0x17c [test_bpf]
  Call Trace:
   .__run_one+0x60/0x17c [test_bpf] (unreliable)
   .test_bpf_init+0x6a8/0xdc8 [test_bpf]
   .do_one_initcall+0x6c/0x28c
   .do_init_module+0x68/0x28c
   .load_module+0x2460/0x2abc
   .__do_sys_init_module+0x120/0x18c
   .system_call_exception+0x110/0x1b8
   system_call_common+0xf0/0x210
  --- interrupt: c00 at 0x101d0acc
  <...>
  ---[ end trace 47b2bf19090bb3d0 ]---

  Illegal instruction

The illegal instruction turned out to be 'ldbrx' emitted for
BPF_FROM_[L|B]E, which was only introduced in ISA v2.06. Guard use of
the same and implement an alternative approach for older processors.

Fixes: 156d0e2 ("powerpc/ebpf/jit: Implement JIT compiler for extended BPF")
Reported-by: Johan Almbladh <[email protected]>
Signed-off-by: Naveen N. Rao <[email protected]>
Tested-by: Johan Almbladh <[email protected]>
Acked-by: Johan Almbladh <[email protected]>
Signed-off-by: Michael Ellerman <[email protected]>
Link: https://lore.kernel.org/r/d1e51c6fdf572062cf3009a751c3406bda01b832.1641468127.git.naveen.n.rao@linux.vnet.ibm.com

Author: rnav

arch/powerpc/include/asm/ppc-opcode.h

@@ -500,6 +500,7 @@
 #define PPC_RAW_LDX(r, base, b)		(0x7c00002a | ___PPC_RT(r) | ___PPC_RA(base) | ___PPC_RB(b))
 #define PPC_RAW_LHZ(r, base, i)		(0xa0000000 | ___PPC_RT(r) | ___PPC_RA(base) | IMM_L(i))
 #define PPC_RAW_LHBRX(r, base, b)	(0x7c00062c | ___PPC_RT(r) | ___PPC_RA(base) | ___PPC_RB(b))
+#define PPC_RAW_LWBRX(r, base, b)	(0x7c00042c | ___PPC_RT(r) | ___PPC_RA(base) | ___PPC_RB(b))
 #define PPC_RAW_LDBRX(r, base, b)	(0x7c000428 | ___PPC_RT(r) | ___PPC_RA(base) | ___PPC_RB(b))
 #define PPC_RAW_STWCX(s, a, b)		(0x7c00012d | ___PPC_RS(s) | ___PPC_RA(a) | ___PPC_RB(b))
 #define PPC_RAW_CMPWI(a, i)		(0x2c000000 | ___PPC_RA(a) | IMM_L(i))

arch/powerpc/net/bpf_jit_comp64.c

@@ -634,17 +634,21 @@ int bpf_jit_build_body(struct bpf_prog *fp, u32 *image, struct codegen_context *
 				EMIT(PPC_RAW_MR(dst_reg, b2p[TMP_REG_1]));
 				break;
 			case 64:
-				/*
-				 * Way easier and faster(?) to store the value
-				 * into stack and then use ldbrx
-				 *
-				 * ctx->seen will be reliable in pass2, but
-				 * the instructions generated will remain the
-				 * same across all passes
-				 */
+				/* Store the value to stack and then use byte-reverse loads */
 				PPC_BPF_STL(dst_reg, 1, bpf_jit_stack_local(ctx));
 				EMIT(PPC_RAW_ADDI(b2p[TMP_REG_1], 1, bpf_jit_stack_local(ctx)));
-				EMIT(PPC_RAW_LDBRX(dst_reg, 0, b2p[TMP_REG_1]));
+				if (cpu_has_feature(CPU_FTR_ARCH_206)) {
+					EMIT(PPC_RAW_LDBRX(dst_reg, 0, b2p[TMP_REG_1]));
+				} else {
+					EMIT(PPC_RAW_LWBRX(dst_reg, 0, b2p[TMP_REG_1]));
+					if (IS_ENABLED(CONFIG_CPU_LITTLE_ENDIAN))
+						EMIT(PPC_RAW_SLDI(dst_reg, dst_reg, 32));
+					EMIT(PPC_RAW_LI(b2p[TMP_REG_2], 4));
+					EMIT(PPC_RAW_LWBRX(b2p[TMP_REG_2], b2p[TMP_REG_2], b2p[TMP_REG_1]));
+					if (IS_ENABLED(CONFIG_CPU_BIG_ENDIAN))
+						EMIT(PPC_RAW_SLDI(b2p[TMP_REG_2], b2p[TMP_REG_2], 32));
+					EMIT(PPC_RAW_OR(dst_reg, dst_reg, b2p[TMP_REG_2]));
+				}
 				break;
 			}
 			break;