loader: preserve resolved relocations across dynamic loads

src/loader/runtimeLinker.cpp

@@ -1007,17 +1007,12 @@ static RelocationInfo GetRelocationInfo(Elf64_Rela* r, Program* program) {
 	return ret;
 }
 
-static void RelocateRecord(uint32_t index, Elf64_Rela* r, Program* program, bool jmprela_table,
+static bool RelocateRecord(uint32_t index, Elf64_Rela* r, Program* program, bool jmprela_table,
-                           bool imports_only, std::vector<std::string>* unresolved) {
+                           std::vector<std::string>* unresolved) {
 	KYTY_PROFILER_FUNCTION();
 
 	auto ri = GetRelocationInfo(r, program);
 
-	if (imports_only &&
-	    (ri.bind_self || (ri.bind != BindType::Global && ri.bind != BindType::Weak))) {
-		return;
-	}
-
 	[[maybe_unused]] bool patched        = false;
 	bool                  stubbed_import = false;
 	bool                  stubbed_func   = false;
@@ -1098,17 +1093,23 @@ static void RelocateRecord(uint32_t index, Elf64_Rela* r, Program* program, bool
 			LOGF("Relocate: %s\n", dbg_str.c_str());
 		}
 	}
+	return ri.resolved;
 }
 
 static void RelocateRecords(Elf64_Rela* records, uint64_t size, Program* program,
-                            bool jmprela_table, bool imports_only,
+                            bool jmprela_table, uint32_t bit_base,
                             std::vector<std::string>* unresolved) {
 	KYTY_PROFILER_FUNCTION();
 
 	uint32_t index = 0;
 	for (auto* r = records;
 	     reinterpret_cast<uint8_t*>(r) < reinterpret_cast<uint8_t*>(records) + size; r++, index++) {
-		RelocateRecord(index, r, program, jmprela_table, imports_only, unresolved);
+		const auto bit_index = bit_base + index;
+		if ((program->rela_bits[bit_index >> 3u] & (1u << (bit_index & 7u))) == 0) {
+			if (RelocateRecord(index, r, program, jmprela_table, unresolved)) {
+				program->rela_bits[bit_index >> 3u] |= 1u << (bit_index & 7u);
+			}
+		}
 	}
 }
 
@@ -1314,8 +1315,6 @@ void RuntimeLinker::RelocateAll() {
 	for (auto* p: m_programs) {
 		Relocate(p);
 	}
-
-	m_relocated = true;
 }
 
 void RuntimeLinker::RelocateProgram(Program* program) {
@@ -1342,10 +1341,6 @@ void RuntimeLinker::UnloadProgram(Program* program) {
 	} else {
 		EXIT("program not found");
 	}
-
-	if (m_relocated) {
-		RelocateAll();
-	}
 }
 
 RuntimeLinker::RuntimeLinker(): m_symbols(std::make_unique<SymbolDatabase>()) {
@@ -1499,7 +1494,6 @@ void RuntimeLinker::Clear() {
 	g_tls_cached_main_tcb     = nullptr;
 	g_desired_base_addr       = SYSTEM_RESERVED + CODE_BASE_OFFSET;
 	m_symbols.reset();
-	m_relocated = false;
 }
 
 void RuntimeLinker::Resolve(const std::string& name, SymbolType type, Program* program,
@@ -2281,6 +2275,11 @@ void RuntimeLinker::ParseProgramDynamicInfo(Program* program) {
 	GetDynValue(elf, &program->dynamic_info->rela_table_total_size, DT_RELASZ);
 	GetDynValue(elf, &program->dynamic_info->rela_table_entry_size, DT_OS_RELAENT);
 	GetDynValue(elf, &program->dynamic_info->rela_table_entry_size, DT_RELAENT);
+	EXIT_NOT_IMPLEMENTED(program->dynamic_info->rela_table_total_size % sizeof(Elf64_Rela) != 0 ||
+	                     program->dynamic_info->jmprela_table_size % sizeof(Elf64_Rela) != 0);
+	const auto rela_count = program->dynamic_info->rela_table_total_size / sizeof(Elf64_Rela);
+	const auto jmprela_count = program->dynamic_info->jmprela_table_size / sizeof(Elf64_Rela);
+	program->rela_bits.resize((rela_count + jmprela_count + 7u) / 8u);
 
 	GetDynValue(elf, &program->dynamic_info->relative_count, DT_RELACOUNT);
 
@@ -2348,7 +2347,6 @@ static void InstallRelocateHandler(Program* program) {
 	}
 
 	// TODO(): check if this table already generated by compiler (sometimes it is missing)
-	if (program->custom_call_plt_vaddr == 0) {
 	program->custom_call_plt_num =
 	    program->dynamic_info->jmprela_table_size / sizeof(Elf64_Rela);
 	auto size                      = Jit::CallPlt::GetSize(program->custom_call_plt_num);
@@ -2362,7 +2360,6 @@ static void InstallRelocateHandler(Program* program) {
 	                                                     Common::VirtualMemory::Mode::Execute));
 	Common::VirtualMemory::FlushInstructionCache(program->custom_call_plt_vaddr, size);
 }
-}
 
 void RuntimeLinker::Relocate(Program* program) {
 	KYTY_PROFILER_FUNCTION();
@@ -2385,16 +2382,18 @@ void RuntimeLinker::Relocate(Program* program) {
 	EXIT_NOT_IMPLEMENTED(program->dynamic_info->symbol_table == nullptr);
 	EXIT_NOT_IMPLEMENTED(program->dynamic_info->pltgot_vaddr == 0);
 
+	if (program->custom_call_plt_vaddr == 0) {
 		InstallRelocateHandler(program);
+	}
 
 	std::vector<std::string> unresolved;
-	const bool               imports_only = program->relocated;
+	const auto rela_count = static_cast<uint32_t>(
+	    program->dynamic_info->rela_table_total_size / sizeof(Elf64_Rela));
 
 	RelocateRecords(program->dynamic_info->rela_table, program->dynamic_info->rela_table_total_size,
-	                program, false, imports_only, &unresolved);
+	                program, false, 0, &unresolved);
 	RelocateRecords(program->dynamic_info->jmprela_table, program->dynamic_info->jmprela_table_size,
-	                program, true, imports_only, &unresolved);
+	                program, true, rela_count, &unresolved);
-	program->relocated = true;
 
 	if (program->tls.image_vaddr != 0 && program->tls.init_size != 0 &&
 	    program->tls.init_image.empty()) {

src/loader/runtimeLinker.h

@@ -132,7 +132,7 @@ struct Program {
 	ThreadLocalStorage              tls;
 	bool                            fail_if_global_not_resolved = true;
 	bool                            dbg_print_reloc             = false;
-	bool                            relocated                   = false;
+	std::vector<uint8_t>            rela_bits;
 	uint64_t                        proc_param_vaddr            = 0;
 	uint64_t                        custom_call_plt_vaddr       = 0;
 	uint32_t                        custom_call_plt_num         = 0;
@@ -201,7 +201,6 @@ private:
 
 	std::vector<Program*>           m_programs;
 	std::unique_ptr<SymbolDatabase> m_symbols;
-	bool                            m_relocated = false;
 	Common::Mutex                   m_mutex;
 
 	application_heap_malloc_func_t         m_application_heap_malloc         = nullptr;