#include "QueryFuzzer.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include namespace DB { namespace ErrorCodes { extern const int TOO_DEEP_RECURSION; } Field QueryFuzzer::getRandomField(int type) { switch (type) { case 0: { static constexpr Int64 values[] = {-2, -1, 0, 1, 2, 3, 7, 10, 100, 255, 256, 257, 1023, 1024, 1025, 65535, 65536, 65537, 1024 * 1024 - 1, 1024 * 1024, 1024 * 1024 + 1, INT64_MIN, INT64_MAX}; return values[fuzz_rand() % (sizeof(values) / sizeof(*values))]; } case 1: { static constexpr float values[] = {NAN, INFINITY, -INFINITY, 0., 0.0001, 0.5, 0.9999, 1., 1.0001, 2., 10.0001, 100.0001, 1000.0001}; return values[fuzz_rand() % (sizeof(values) / sizeof(*values))]; } case 2: { static constexpr Int64 values[] = {-2, -1, 0, 1, 2, 3, 7, 10, 100, 255, 256, 257, 1023, 1024, 1025, 65535, 65536, 65537, 1024 * 1024 - 1, 1024 * 1024, 1024 * 1024 + 1, INT64_MIN, INT64_MAX}; static constexpr UInt64 scales[] = {0, 1, 2, 10}; return DecimalField( values[fuzz_rand() % (sizeof(values) / sizeof(*values))], scales[fuzz_rand() % (sizeof(scales) / sizeof(*scales))] ); } default: assert(false); return Null{}; } } Field QueryFuzzer::fuzzField(Field field) { const auto type = field.getType(); int type_index = -1; if (type == Field::Types::Int64 || type == Field::Types::UInt64) { type_index = 0; } else if (type == Field::Types::Float64) { type_index = 1; } else if (type == Field::Types::Decimal32 || type == Field::Types::Decimal64 || type == Field::Types::Decimal128) { type_index = 2; } if (fuzz_rand() % 20 == 0) { return Null{}; } if (type_index >= 0) { if (fuzz_rand() % 20 == 0) { // Change type sometimes, but not often, because it mostly leads to // boring errors. type_index = fuzz_rand() % 3; } return getRandomField(type_index); } if (type == Field::Types::String) { auto & str = field.get(); UInt64 action = fuzz_rand() % 10; switch (action) { case 0: str = ""; break; case 1: str = str + str; break; case 2: str = str + str + str + str; break; case 4: if (!str.empty()) { str[fuzz_rand() % str.size()] = '\0'; } break; default: // Do nothing break; } } else if (type == Field::Types::Array || type == Field::Types::Tuple) { auto & arr = field.reinterpret(); if (fuzz_rand() % 5 == 0 && !arr.empty()) { size_t pos = fuzz_rand() % arr.size(); arr.erase(arr.begin() + pos); fprintf(stderr, "erased\n"); } if (fuzz_rand() % 5 == 0) { if (!arr.empty()) { size_t pos = fuzz_rand() % arr.size(); arr.insert(arr.begin() + pos, fuzzField(arr[pos])); fprintf(stderr, "inserted (pos %zd)\n", pos); } else { arr.insert(arr.begin(), getRandomField(0)); fprintf(stderr, "inserted (0)\n"); } } for (auto & element : arr) { element = fuzzField(element); } } return field; } ASTPtr QueryFuzzer::getRandomColumnLike() { if (column_like.empty()) { return nullptr; } ASTPtr new_ast = column_like[fuzz_rand() % column_like.size()]->clone(); new_ast->setAlias(""); return new_ast; } void QueryFuzzer::replaceWithColumnLike(ASTPtr & ast) { if (column_like.empty()) { return; } std::string old_alias = ast->tryGetAlias(); ast = getRandomColumnLike(); ast->setAlias(old_alias); } void QueryFuzzer::replaceWithTableLike(ASTPtr & ast) { if (table_like.empty()) { return; } ASTPtr new_ast = table_like[fuzz_rand() % table_like.size()]->clone(); std::string old_alias = ast->tryGetAlias(); new_ast->setAlias(old_alias); ast = new_ast; } void QueryFuzzer::fuzzOrderByElement(ASTOrderByElement * elem) { switch (fuzz_rand() % 10) { case 0: elem->direction = -1; break; case 1: elem->direction = 1; break; case 2: elem->nulls_direction = -1; elem->nulls_direction_was_explicitly_specified = true; break; case 3: elem->nulls_direction = 1; elem->nulls_direction_was_explicitly_specified = true; break; case 4: elem->nulls_direction = elem->direction; elem->nulls_direction_was_explicitly_specified = false; break; default: // do nothing break; } } void QueryFuzzer::fuzzOrderByList(IAST * ast) { if (!ast) { return; } auto * list = assert_cast(ast); // Remove element if (fuzz_rand() % 50 == 0 && list->children.size() > 1) { // Don't remove last element -- this leads to questionable // constructs such as empty select. list->children.erase(list->children.begin() + fuzz_rand() % list->children.size()); } // Add element if (fuzz_rand() % 50 == 0) { auto pos = list->children.empty() ? list->children.begin() : list->children.begin() + fuzz_rand() % list->children.size(); auto col = getRandomColumnLike(); if (col) { auto elem = std::make_shared(); elem->children.push_back(col); elem->direction = 1; elem->nulls_direction = 1; elem->nulls_direction_was_explicitly_specified = false; elem->with_fill = false; list->children.insert(pos, elem); } else { fprintf(stderr, "no random col!\n"); } } // We don't have to recurse here to fuzz the children, this is handled by // the generic recursion into IAST.children. } void QueryFuzzer::fuzzColumnLikeExpressionList(IAST * ast) { if (!ast) { return; } auto * impl = assert_cast(ast); // Remove element if (fuzz_rand() % 50 == 0 && impl->children.size() > 1) { // Don't remove last element -- this leads to questionable // constructs such as empty select. impl->children.erase(impl->children.begin() + fuzz_rand() % impl->children.size()); } // Add element if (fuzz_rand() % 50 == 0) { auto pos = impl->children.empty() ? impl->children.begin() : impl->children.begin() + fuzz_rand() % impl->children.size(); auto col = getRandomColumnLike(); if (col) { impl->children.insert(pos, col); } else { fprintf(stderr, "no random col!\n"); } } // We don't have to recurse here to fuzz the children, this is handled by // the generic recursion into IAST.children. } void QueryFuzzer::fuzzWindowFrame(WindowFrame & frame) { switch (fuzz_rand() % 40) { case 0: { const auto r = fuzz_rand() % 3; frame.type = r == 0 ? WindowFrame::FrameType::Rows : r == 1 ? WindowFrame::FrameType::Range : WindowFrame::FrameType::Groups; break; } case 1: { const auto r = fuzz_rand() % 3; frame.begin_type = r == 0 ? WindowFrame::BoundaryType::Unbounded : r == 1 ? WindowFrame::BoundaryType::Current : WindowFrame::BoundaryType::Offset; break; } case 2: { const auto r = fuzz_rand() % 3; frame.end_type = r == 0 ? WindowFrame::BoundaryType::Unbounded : r == 1 ? WindowFrame::BoundaryType::Current : WindowFrame::BoundaryType::Offset; break; } case 3: { frame.begin_offset = getRandomField(0).get(); break; } case 4: { frame.end_offset = getRandomField(0).get(); break; } default: break; } frame.is_default = (frame == WindowFrame{}); } void QueryFuzzer::fuzz(ASTs & asts) { for (auto & ast : asts) { fuzz(ast); } } struct ScopedIncrement { size_t & counter; explicit ScopedIncrement(size_t & counter_) : counter(counter_) { ++counter; } ~ScopedIncrement() { --counter; } }; void QueryFuzzer::fuzz(ASTPtr & ast) { if (!ast) return; // Check for exceeding max depth. ScopedIncrement depth_increment(current_ast_depth); if (current_ast_depth > 500) { // The AST is too deep (see the comment for current_ast_depth). Throw // an exception to fail fast and not use this query as an etalon, or we'll // end up in a very slow and useless loop. It also makes sense to set it // lower than the default max parse depth on the server (1000), so that // we don't get the useless error about parse depth from the server either. throw Exception(ErrorCodes::TOO_DEEP_RECURSION, "AST depth exceeded while fuzzing ({})", current_ast_depth); } // Check for loops. auto [_, inserted] = debug_visited_nodes.insert(ast.get()); if (!inserted) { fmt::print(stderr, "The AST node '{}' was already visited before." " Depth {}, {} visited nodes, current top AST:\n{}\n", static_cast(ast.get()), current_ast_depth, debug_visited_nodes.size(), (*debug_top_ast)->dumpTree()); assert(false); } // The fuzzing. if (auto * with_union = typeid_cast(ast.get())) { fuzz(with_union->list_of_selects); } else if (auto * tables = typeid_cast(ast.get())) { fuzz(tables->children); } else if (auto * tables_element = typeid_cast(ast.get())) { fuzz(tables_element->table_join); fuzz(tables_element->table_expression); fuzz(tables_element->array_join); } else if (auto * table_expr = typeid_cast(ast.get())) { fuzz(table_expr->database_and_table_name); fuzz(table_expr->subquery); fuzz(table_expr->table_function); } else if (auto * expr_list = typeid_cast(ast.get())) { fuzz(expr_list->children); } else if (auto * order_by_element = typeid_cast(ast.get())) { fuzzOrderByElement(order_by_element); } else if (auto * fn = typeid_cast(ast.get())) { fuzzColumnLikeExpressionList(fn->arguments.get()); fuzzColumnLikeExpressionList(fn->parameters.get()); if (fn->is_window_function && fn->window_definition) { auto & def = fn->window_definition->as(); fuzzColumnLikeExpressionList(def.partition_by.get()); fuzzOrderByList(def.order_by.get()); fuzzWindowFrame(def.frame); } fuzz(fn->children); } else if (auto * select = typeid_cast(ast.get())) { fuzzColumnLikeExpressionList(select->select().get()); fuzzColumnLikeExpressionList(select->groupBy().get()); fuzzOrderByList(select->orderBy().get()); fuzz(select->children); } /* * The time to fuzz the settings has not yet come. * Apparently we don't have any infractructure to validate the values of * the settings, and the first query with max_block_size = -1 breaks * because of overflows here and there. *//* * else if (auto * set = typeid_cast(ast.get())) * { * for (auto & c : set->changes) * { * if (fuzz_rand() % 50 == 0) * { * c.value = fuzzField(c.value); * } * } * } */ else if (auto * literal = typeid_cast(ast.get())) { // There is a caveat with fuzzing the children: many ASTs also keep the // links to particular children in own fields. This means that replacing // the child with another object might lead to error. Many of these fields // are ASTPtr -- this is redundant ownership, but hides the error if the // child field is replaced. Others can be ASTLiteral * or the like, which // leads to segfault if the pointed-to AST is replaced. // Replacing children is safe in case of ASTExpressionList. In a more // general case, we can change the value of ASTLiteral, which is what we // do here. if (fuzz_rand() % 11 == 0) { literal->value = fuzzField(literal->value); } } else { fuzz(ast->children); } } /* * This functions collects various parts of query that we can then substitute * to a query being fuzzed. * * TODO: we just stop remembering new parts after our corpus reaches certain size. * This is boring, should implement a random replacement of existing parst with * small probability. Do this after we add this fuzzer to CI and fix all the * problems it can routinely find even in this boring version. */ void QueryFuzzer::collectFuzzInfoMain(const ASTPtr ast) { collectFuzzInfoRecurse(ast); aliases.clear(); for (const auto & alias : aliases_set) { aliases.push_back(alias); } column_like.clear(); for (const auto & [name, value] : column_like_map) { column_like.push_back(value); } table_like.clear(); for (const auto & [name, value] : table_like_map) { table_like.push_back(value); } } void QueryFuzzer::addTableLike(const ASTPtr ast) { if (table_like_map.size() > 1000) { table_like_map.clear(); } const auto name = ast->formatForErrorMessage(); if (name.size() < 200) { table_like_map.insert({name, ast}); } } void QueryFuzzer::addColumnLike(const ASTPtr ast) { if (column_like_map.size() > 1000) { column_like_map.clear(); } const auto name = ast->formatForErrorMessage(); if (name.size() < 200) { column_like_map.insert({name, ast}); } } void QueryFuzzer::collectFuzzInfoRecurse(const ASTPtr ast) { if (auto * impl = dynamic_cast(ast.get())) { if (aliases_set.size() > 1000) { aliases_set.clear(); } aliases_set.insert(impl->alias); } if (typeid_cast(ast.get())) { addColumnLike(ast); } else if (typeid_cast(ast.get())) { addColumnLike(ast); } else if (typeid_cast(ast.get())) { addColumnLike(ast); } else if (typeid_cast(ast.get())) { addTableLike(ast); } else if (typeid_cast(ast.get())) { addTableLike(ast); } for (const auto & child : ast->children) { collectFuzzInfoRecurse(child); } } void QueryFuzzer::fuzzMain(ASTPtr & ast) { current_ast_depth = 0; debug_visited_nodes.clear(); debug_top_ast = * collectFuzzInfoMain(ast); fuzz(ast); std::cout << std::endl; WriteBufferFromOStream ast_buf(std::cout, 4096); formatAST(*ast, ast_buf, false /*highlight*/); ast_buf.next(); std::cout << std::endl << std::endl; } }