Add 32-bit unconditional jump instruction

This instruction is part of the cpu=v4 set of instructions recently added, and specified in https://www.ietf.org/archive/id/draft-ietf-bpf-isa-00.html#jump-instructions Fixes #532 Signed-off-by: Dave Thaler <[email protected]>
vbpf · Dec 27, 2023 · f269a44 · f269a44
1 parent 38ebf36
commit f269a44
Show file tree

Hide file tree

Showing 5 changed files with 61 additions and 11 deletions.
diff --git a/src/asm_marshal.cpp b/src/asm_marshal.cpp
@@ -94,7 +94,8 @@ struct MarshalVisitor {
     }
 
   public:
-    std::function<auto(label_t)->int16_t> label_to_offset;
+    std::function<auto(label_t)->int16_t> label_to_offset16;
+    std::function<auto(label_t)->int32_t> label_to_offset32;
 
     vector<ebpf_inst> operator()(Undefined const& a) {
         assert(false);
@@ -179,7 +180,7 @@ struct MarshalVisitor {
                 .opcode = static_cast<uint8_t>(INST_CLS_JMP | (op(b.cond->op) << 4)),
                 .dst = b.cond->left.v,
                 .src = 0,
-                .offset = label_to_offset(b.target),
+                .offset = label_to_offset16(b.target),
             };
             visit(overloaded{[&](Reg right) {
                                  res.opcode |= INST_SRC_REG;
@@ -189,7 +190,11 @@ struct MarshalVisitor {
                   b.cond->right);
             return {res};
         } else {
-            return {ebpf_inst{.opcode = INST_OP_JA, .dst = 0, .src = 0, .offset = label_to_offset(b.target), .imm = 0}};
+            int32_t imm = label_to_offset32(b.target);
+            if (imm != 0)
+                return {ebpf_inst{.opcode = INST_OP_JA32, .imm = imm}};
+            else
+                return {ebpf_inst{.opcode = INST_OP_JA16, .offset = label_to_offset16(b.target)}};
         }
     }
 
@@ -252,7 +257,9 @@ struct MarshalVisitor {
     }
 };
 
-vector<ebpf_inst> marshal(const Instruction& ins, pc_t pc) { return std::visit(MarshalVisitor{label_to_offset(pc)}, ins); }
+vector<ebpf_inst> marshal(const Instruction& ins, pc_t pc) {
+    return std::visit(MarshalVisitor{label_to_offset16(pc), label_to_offset32(pc)}, ins);
+}
 
 vector<ebpf_inst> marshal(const vector<Instruction>& insts) {
     vector<ebpf_inst> res;

diff --git a/src/asm_ostream.hpp b/src/asm_ostream.hpp
@@ -11,8 +11,21 @@
 
 #include "asm_syntax.hpp"
 
-inline std::function<int16_t(label_t)> label_to_offset(pc_t pc) {
-    return [=](const label_t& label) { return label.from - pc - 1; };
+// We use a 16-bit offset whenever it fits in 16 bits.
+inline std::function<int16_t(label_t)> label_to_offset16(pc_t pc) {
+    return [=](const label_t& label) {
+        int64_t offset = label.from - (int64_t)pc - 1;
+        return (offset >= INT16_MIN && offset <= INT16_MAX) ? (int16_t)offset : 0;
+    };
+}
+
+// We use the JA32 opcode with the offset in 'imm' when the offset
+// of an unconditional jump doesn't fit in a int16_t.
+inline std::function<int32_t(label_t)> label_to_offset32(pc_t pc) {
+    return [=](const label_t& label) {
+        int64_t offset = label.from - (int64_t)pc - 1;
+        return (offset >= INT16_MIN && offset <= INT16_MAX) ? 0 : (int32_t)offset;
+    };
 }
 
 std::ostream& operator<<(std::ostream& os, const btf_line_info_t& line_info);

diff --git a/src/asm_syntax.hpp b/src/asm_syntax.hpp
@@ -336,7 +336,8 @@ using InstructionSeq = std::vector<LabeledInstruction>;
 #define DECLARE_EQ1(T, f1) \
     inline bool operator==(T const& a, T const& b) { return a.f1 == b.f1; }
 
-using pc_t = uint16_t;
+// cpu=v4 supports 32-bit PC offsets so we need a large enough type.
+using pc_t = size_t;
 
 // Helpers:
 

diff --git a/src/asm_unmarshal.cpp b/src/asm_unmarshal.cpp
@@ -381,16 +381,19 @@ struct Unmarshaller {
                 throw InvalidInstruction(pc, "Bad instruction");
             return Exit{};
         case 0x0:
-            if ((inst.opcode & INST_CLS_MASK) != INST_CLS_JMP)
+            if ((inst.opcode & INST_CLS_MASK) != INST_CLS_JMP &&
+                (inst.opcode & INST_CLS_MASK) != INST_CLS_JMP32)
                 throw InvalidInstruction(pc, "Bad instruction");
         default: {
-            pc_t new_pc = pc + 1 + inst.offset;
+            int32_t offset = (inst.opcode == INST_OP_JA32) ? inst.imm : inst.offset;
+            pc_t new_pc = pc + 1 + offset;
             if (new_pc >= insts.size())
                 throw InvalidInstruction(pc, "jump out of bounds");
             else if (insts[new_pc].opcode == 0)
                 throw InvalidInstruction(pc, "jump to middle of lddw");
 
-            auto cond = inst.opcode == INST_OP_JA ? std::optional<Condition>{}
+            auto cond = (inst.opcode == INST_OP_JA16 || inst.opcode == INST_OP_JA32)
+                                                  ? std::optional<Condition>{}
                                                   : Condition{
                                                         .op = getJmpOp(pc, inst.opcode),
                                                         .left = Reg{inst.dst},
@@ -400,7 +403,7 @@ struct Unmarshaller {
                                                     };
             return Jmp{
                 .cond = cond,
-                .target = label_t{new_pc},
+                .target = label_t{(int)new_pc},
             };
         }
         }

diff --git a/src/test/test_marshal.cpp b/src/test/test_marshal.cpp
@@ -6,6 +6,25 @@
 #include "asm_marshal.hpp"
 #include "asm_unmarshal.hpp"
 
+// Verify that if we unmarshal an instruction and then re-marshal it,
+// we get what we expect.
+static void compare_unmarshal_marshal(const ebpf_inst& ins, const ebpf_inst& expected_result) {
+    program_info info{.platform = &g_ebpf_platform_linux,
+                      .type = g_ebpf_platform_linux.get_program_type("unspec", "unspec")};
+    const ebpf_inst exit{.opcode = INST_OP_EXIT};
+    InstructionSeq parsed = std::get<InstructionSeq>(unmarshal(raw_program{"", "", {ins, exit, exit}, info}));
+    REQUIRE(parsed.size() == 3);
+    auto [_, single, _2] = parsed.front();
+    (void)_;  // unused
+    (void)_2; // unused
+    std::vector<ebpf_inst> marshaled = marshal(single, 0);
+    REQUIRE(marshaled.size() == 1);
+    ebpf_inst result = marshaled.back();
+    REQUIRE(memcmp(&expected_result, &result, sizeof(result)) == 0);
+}
+
+// Verify that if we marshal an instruction and then unmarshal it,
+// we get the original.
 static void compare_marshal_unmarshal(const Instruction& ins, bool double_cmd = false) {
     program_info info{.platform = &g_ebpf_platform_linux,
                       .type = g_ebpf_platform_linux.get_program_type("unspec", "unspec")};
@@ -81,6 +100,13 @@ TEST_CASE("disasm_marshal", "[disasm][marshal]") {
             // Condition::Op::NSET, does not exist in ebpf
                     Condition::Op::NE, Condition::Op::SGT, Condition::Op::SGE, Condition::Op::LT, Condition::Op::LE,
                     Condition::Op::SLT, Condition::Op::SLE};
+        SECTION("goto offset") {
+            ebpf_inst jmp_offset{.opcode = INST_OP_JA16, .offset = 1};
+            compare_unmarshal_marshal(jmp_offset, jmp_offset);
+
+            // JA32 +1 is equivalent to JA16 +1 since the offset fits in 16 bits.
+            compare_unmarshal_marshal(ebpf_inst{.opcode = INST_OP_JA32, .imm = 1}, jmp_offset);
+        }
         SECTION("Reg right") {
             for (auto op : ops) {
                 Condition cond{.op = op, .left = Reg{1}, .right = Reg{2}};