bpf, arm64: Jit BPF_CALL to direct call when possible

Currently, BPF_CALL is always jited to indirect call. When target is
within the range of direct call, BPF_CALL can be jited to direct call.

For example, the following BPF_CALL

    call __htab_map_lookup_elem

is always jited to indirect call:

    mov     x10, #0xffffffffffff18f4
    movk    x10, #0x821, lsl #16
    movk    x10, #0x8000, lsl #32
    blr     x10

When the address of target __htab_map_lookup_elem is within the range of
direct call, the BPF_CALL can be jited to:

    bl      0xfffffffffd33bc98

This patch does such jit optimization by emitting arm64 direct calls for
BPF_CALL when possible, indirect calls otherwise.

Without this patch, the jit works as follows.

1. First pass
   A. Determine jited position and size for each bpf instruction.
   B. Computed the jited image size.

2. Allocate jited image with size computed in step 1.

3. Second pass
   A. Adjust jump offset for jump instructions
   B. Write the final image.

This works because, for a given bpf prog, regardless of where the jited
image is allocated, the jited result for each instruction is fixed. The
second pass differs from the first only in adjusting the jump offsets,
like changing "jmp imm1" to "jmp imm2", while the position and size of
the "jmp" instruction remain unchanged.

Now considering whether to jit BPF_CALL to arm64 direct or indirect call
instruction. The choice depends solely on the jump offset: direct call
if the jump offset is within 128MB, indirect call otherwise.

For a given BPF_CALL, the target address is known, so the jump offset is
decided by the jited address of the BPF_CALL instruction. In other words,
for a given bpf prog, the jited result for each BPF_CALL is determined
by its jited address.

The jited address for a BPF_CALL is the jited image address plus the
total jited size of all preceding instructions. For a given bpf prog,
there are clearly no BPF_CALL instructions before the first BPF_CALL
instruction. Since the jited result for all other instructions other
than BPF_CALL are fixed, the total jited size preceding the first
BPF_CALL is also fixed. Therefore, once the jited image is allocated,
the jited address for the first BPF_CALL is fixed.

Now that the jited result for the first BPF_CALL is fixed, the jited
results for all instructions preceding the second BPF_CALL are fixed.
So the jited address and result for the second BPF_CALL are also fixed.

Similarly, we can conclude that the jited addresses and results for all
subsequent BPF_CALL instructions are fixed.

This means that, for a given bpf prog, once the jited image is allocated,
the jited address and result for all instructions, including all BPF_CALL
instructions, are fixed.

Based on the observation, with this patch, the jit works as follows.

1. First pass
   Estimate the maximum jited image size. In this pass, all BPF_CALLs
   are jited to arm64 indirect calls since the jump offsets are unknown
   because the jited image is not allocated.

2. Allocate jited image with size estimated in step 1.

3. Second pass
   A. Determine the jited result for each BPF_CALL.
   B. Determine jited address and size for each bpf instruction.

4. Third pass
   A. Adjust jump offset for jump instructions.
   B. Write the final image.

Signed-off-by: Xu Kuohai <xukuohai@huawei.com>
Reviewed-by: Puranjay Mohan <puranjay@kernel.org>
Link: https://lore.kernel.org/r/20240903094407.601107-1-xukuohai@huaweicloud.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>

arch/arm64/net/bpf_jit_comp.c

@@ -84,6 +84,7 @@ struct jit_ctx {
 	u64 user_vm_start;
 	u64 arena_vm_start;
 	bool fp_used;
+	bool write;
 };
 
 struct bpf_plt {
@@ -97,7 +98,7 @@ struct bpf_plt {
 
 static inline void emit(const u32 insn, struct jit_ctx *ctx)
 {
-	if (ctx->image != NULL)
+	if (ctx->image != NULL && ctx->write)
 		ctx->image[ctx->idx] = cpu_to_le32(insn);
 
 	ctx->idx++;
@@ -182,14 +183,47 @@ static inline void emit_addr_mov_i64(const int reg, const u64 val,
 	}
 }
 
-static inline void emit_call(u64 target, struct jit_ctx *ctx)
+static bool should_emit_indirect_call(long target, const struct jit_ctx *ctx)
 {
-	u8 tmp = bpf2a64[TMP_REG_1];
+	long offset;
 
+	/* when ctx->ro_image is not allocated or the target is unknown,
+	 * emit indirect call
+	 */
+	if (!ctx->ro_image || !target)
+		return true;
+
+	offset = target - (long)&ctx->ro_image[ctx->idx];
+	return offset < -SZ_128M || offset >= SZ_128M;
+}
+
+static void emit_direct_call(u64 target, struct jit_ctx *ctx)
+{
+	u32 insn;
+	unsigned long pc;
+
+	pc = (unsigned long)&ctx->ro_image[ctx->idx];
+	insn = aarch64_insn_gen_branch_imm(pc, target, AARCH64_INSN_BRANCH_LINK);
+	emit(insn, ctx);
+}
+
+static void emit_indirect_call(u64 target, struct jit_ctx *ctx)
+{
+	u8 tmp;
+
+	tmp = bpf2a64[TMP_REG_1];
 	emit_addr_mov_i64(tmp, target, ctx);
 	emit(A64_BLR(tmp), ctx);
 }
 
+static void emit_call(u64 target, struct jit_ctx *ctx)
+{
+	if (should_emit_indirect_call((long)target, ctx))
+		emit_indirect_call(target, ctx);
+	else
+		emit_direct_call(target, ctx);
+}
+
 static inline int bpf2a64_offset(int bpf_insn, int off,
 				 const struct jit_ctx *ctx)
 {
@@ -1649,13 +1683,11 @@ static int build_body(struct jit_ctx *ctx, bool extra_pass)
 		const struct bpf_insn *insn = &prog->insnsi[i];
 		int ret;
 
-		if (ctx->image == NULL)
-			ctx->offset[i] = ctx->idx;
+		ctx->offset[i] = ctx->idx;
 		ret = build_insn(insn, ctx, extra_pass);
 		if (ret > 0) {
 			i++;
-			if (ctx->image == NULL)
-				ctx->offset[i] = ctx->idx;
+			ctx->offset[i] = ctx->idx;
 			continue;
 		}
 		if (ret)
@@ -1666,8 +1698,7 @@ static int build_body(struct jit_ctx *ctx, bool extra_pass)
 	 * the last element with the offset after the last
 	 * instruction (end of program)
 	 */
-	if (ctx->image == NULL)
-		ctx->offset[i] = ctx->idx;
+	ctx->offset[i] = ctx->idx;
 
 	return 0;
 }
@@ -1721,6 +1752,8 @@ struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog)
 	struct jit_ctx ctx;
 	u8 *image_ptr;
 	u8 *ro_image_ptr;
+	int body_idx;
+	int exentry_idx;
 
 	if (!prog->jit_requested)
 		return orig_prog;
@@ -1768,8 +1801,7 @@ struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog)
 	ctx.user_vm_start = bpf_arena_get_user_vm_start(prog->aux->arena);
 	ctx.arena_vm_start = bpf_arena_get_kern_vm_start(prog->aux->arena);
 
-	/*
-	 * 1. Initial fake pass to compute ctx->idx and ctx->offset.
+	/* Pass 1: Estimate the maximum image size.
 	 *
 	 * BPF line info needs ctx->offset[i] to be the offset of
 	 * instruction[i] in jited image, so build prologue first.
@@ -1792,7 +1824,7 @@ struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog)
 	extable_size = prog->aux->num_exentries *
 		sizeof(struct exception_table_entry);
 
-	/* Now we know the actual image size. */
+	/* Now we know the maximum image size. */
 	prog_size = sizeof(u32) * ctx.idx;
 	/* also allocate space for plt target */
 	extable_offset = round_up(prog_size + PLT_TARGET_SIZE, extable_align);
@@ -1805,7 +1837,7 @@ struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog)
 		goto out_off;
 	}
 
-	/* 2. Now, the actual pass. */
+	/* Pass 2: Determine jited position and result for each instruction */
 
 	/*
 	 * Use the image(RW) for writing the JITed instructions. But also save
@@ -1821,30 +1853,56 @@ struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog)
 skip_init_ctx:
 	ctx.idx = 0;
 	ctx.exentry_idx = 0;
+	ctx.write = true;
 
 	build_prologue(&ctx, was_classic);
 
+	/* Record exentry_idx and body_idx before first build_body */
+	exentry_idx = ctx.exentry_idx;
+	body_idx = ctx.idx;
+	/* Dont write body instructions to memory for now */
+	ctx.write = false;
+
 	if (build_body(&ctx, extra_pass)) {
 		prog = orig_prog;
 		goto out_free_hdr;
 	}
 
+	ctx.epilogue_offset = ctx.idx;
+	ctx.exentry_idx = exentry_idx;
+	ctx.idx = body_idx;
+	ctx.write = true;
+
+	/* Pass 3: Adjust jump offset and write final image */
+	if (build_body(&ctx, extra_pass) ||
+		WARN_ON_ONCE(ctx.idx != ctx.epilogue_offset)) {
+		prog = orig_prog;
+		goto out_free_hdr;
+	}
+
 	build_epilogue(&ctx);
 	build_plt(&ctx);
 
-	/* 3. Extra pass to validate JITed code. */
+	/* Extra pass to validate JITed code. */
 	if (validate_ctx(&ctx)) {
 		prog = orig_prog;
 		goto out_free_hdr;
 	}
 
+	/* update the real prog size */
+	prog_size = sizeof(u32) * ctx.idx;
+
 	/* And we're done. */
 	if (bpf_jit_enable > 1)
 		bpf_jit_dump(prog->len, prog_size, 2, ctx.image);
 
 	if (!prog->is_func || extra_pass) {
-		if (extra_pass && ctx.idx != jit_data->ctx.idx) {
-			pr_err_once("multi-func JIT bug %d != %d\n",
+		/* The jited image may shrink since the jited result for
+		 * BPF_CALL to subprog may be changed from indirect call
+		 * to direct call.
+		 */
+		if (extra_pass && ctx.idx > jit_data->ctx.idx) {
+			pr_err_once("multi-func JIT bug %d > %d\n",
 				    ctx.idx, jit_data->ctx.idx);
 			prog->bpf_func = NULL;
 			prog->jited = 0;
@@ -2315,6 +2373,7 @@ int arch_prepare_bpf_trampoline(struct bpf_tramp_image *im, void *ro_image,
 		.image = image,
 		.ro_image = ro_image,
 		.idx = 0,
+		.write = true,
 	};
 
 	nregs = btf_func_model_nregs(m);