Relaxed Binding 是 Spring Boot 中一个有趣的机制，它可以让开发人员更灵活地进行配置。但偶然的机会，我发现这个机制在 Spring Boot 1.5 与 Spring Boot 2.0 版本中的实现是不一样的。
Relaxed Binding机制
关于Relaxed Binding机制在Spring Boot的官方文档中有描述，链接如下：
24.7 Type-safe Configuration Properties
24.7.2 Relaxed Binding
这个Relaxed Binding机制的主要作用是，当配置文件（properties或yml）中的配置项值与带@ConfigurationProperties注解的配置类属性进行绑定时，可以进行相对宽松的绑定。
本文并不是针对@ConfigurationProperties与类型安全的属性配置进行解释，相关的说明请参考上面的链接。
Relax Binding示例
下面的例子有助于理解Relaxed Binding。
假设有一个带@ConfigurationProperties注解的属性类
@ConfigurationProperties(prefix="my")
public class MyProperties {
private String firstName;
public String getFirstName() {
return this.firstName;
}
public void setFirstName(String firstName) {
this.firstName = firstName;
}
}
在配置文件（properties文件）中，以下的写法都能将值正确注入到firstName这个属性中。
| property 写法 | 说明 | 推荐场景 |
|---|---|---|
| my.firstName | 标准驼峰 | |
| my.first-name | 减号分隔 | 推荐在properties或yml文件中使用 |
| my.first_name | 下划线分隔 | |
| MY_FIRST_NAME | 大写+下划线分隔 | 推荐用于环境变量或命令行参数 |
源码分析
出于好奇，我研究了一下SpringBoot的源码，发现在1.5版本与2.0版本中，Relaxed Binding 的具体实现是不一样的。
为了解释两个版本的不同之处，我首先在properties配置文件中做如下的配置。
my.first_Name=fonoisrev
Spring Boot 1.5版本的实现（以1.5.14.RELEASE版本为例）
属性与配置值的绑定逻辑始于ConfigurationPropertiesBindingPostProcessor类的postProcessBeforeInitialization函数。
public class ConfigurationPropertiesBindingPostProcessor implements BeanPostProcessor,
BeanFactoryAware, EnvironmentAware, ApplicationContextAware, InitializingBean,
DisposableBean, ApplicationListener<ContextRefreshedEvent>, PriorityOrdered {
...
@Override
public Object postProcessBeforeInitialization(Object bean, String beanName)
throws BeansException {
...
ConfigurationProperties annotation = AnnotationUtils
.findAnnotation(bean.getClass(), ConfigurationProperties.class);
if (annotation != null) {
postProcessBeforeInitialization(bean, beanName, annotation);
}
...
}
private void postProcessBeforeInitialization(Object bean, String beanName,
ConfigurationProperties annotation) {
Object target = bean;
PropertiesConfigurationFactory<Object> factory = new PropertiesConfigurationFactory<Object>(
target);
factory.setPropertySources(this.propertySources);
...
try {
factory.bindPropertiesToTarget();
}
...
}
...
}
从以上代码可以看出，该类实现了BeanPostProcessor接口（BeanPostProcessor官方文档请点我）。BeanPostProcessor接口主要作用是允许自行实现Bean装配的逻辑，而postProcessBeforeInitialization函数则在执行Bean的初始化调用（如afterPropertiesSet）前被调用。
postProcessBeforeInitialization函数执行时，属性值绑定的工作被委派给了PropertiesConfigurationFactory<T>类。
public class PropertiesConfigurationFactory<T> implements FactoryBean<T>,
ApplicationContextAware, MessageSourceAware, InitializingBean {
...
public void bindPropertiesToTarget() throws BindException {
...
doBindPropertiesToTarget();
...
}
private void doBindPropertiesToTarget() throws BindException {
RelaxedDataBinder dataBinder = (this.targetName != null
? new RelaxedDataBinder(this.target, this.targetName)
: new RelaxedDataBinder(this.target));
...
Iterable<String> relaxedTargetNames = getRelaxedTargetNames();
Set<String> names = getNames(relaxedTargetNames);
PropertyValues propertyValues = getPropertySourcesPropertyValues(names,
relaxedTargetNames);
dataBinder.bind(propertyValues);
...
}
private PropertyValues getPropertySourcesPropertyValues(Set<String> names,
Iterable<String> relaxedTargetNames) {
PropertyNamePatternsMatcher includes = getPropertyNamePatternsMatcher(names,
relaxedTargetNames);
return new PropertySourcesPropertyValues(this.propertySources, names, includes,
this.resolvePlaceholders);
}
...
}
以上的代码其逻辑可以描述如下：
- 借助RelaxedNames类，将注解@ConfigurationProperties(prefix="my")中的前缀“my”与MyProperties类的属性firstName可能存在的情况进行穷举（共28种情况，如下表）
| 序号 | RelaxedNames |
|---|---|
| 0 | my.first-name |
| 1 | my_first-name |
| 2 | my.first_name |
| 3 | my_first_name |
| 4 | my.firstName |
| 5 | my_firstName |
| 6 | my.firstname |
| 7 | my_firstname |
| 8 | my.FIRST-NAME |
| 9 | my_FIRST-NAME |
| 10 | my.FIRST_NAME |
| 11 | my_FIRST_NAME |
| 12 | my.FIRSTNAME |
| 13 | my_FIRSTNAME |
| 14 | MY.first-name |
| 15 | MY_first-name |
| 16 | MY.first_name |
| 17 | MY_first_name |
| 18 | MY.firstName |
| 19 | MY_firstName |
| 20 | MY.firstname |
| 21 | MY_firstname |
| 22 | MY.FIRST-NAME |
| 23 | MY_FIRST-NAME |
| 24 | MY.FIRST_NAME |
| 25 | MY_FIRST_NAME |
| 26 | MY.FIRSTNAME |
| 27 | MY_FIRSTNAME |
- 构造PropertySourcesPropertyValues对象从配置文件（如properties文件）中查找匹配的值，
- 使用DataBinder操作PropertySourcesPropertyValues实现MyProperties类的setFirstName方法调用，完成绑定。
public class PropertySourcesPropertyValues implements PropertyValues {
...
PropertySourcesPropertyValues(PropertySources propertySources,
Collection<String> nonEnumerableFallbackNames,
PropertyNamePatternsMatcher includes, boolean resolvePlaceholders) {
...
PropertySourcesPropertyResolver resolver = new PropertySourcesPropertyResolver(
propertySources);
for (PropertySource<?> source : propertySources) {
processPropertySource(source, resolver);
}
}
private void processPropertySource(PropertySource<?> source,
PropertySourcesPropertyResolver resolver) {
...
processEnumerablePropertySource((EnumerablePropertySource<?>) source,
resolver, this.includes);
...
}
private void processEnumerablePropertySource(EnumerablePropertySource<?> source,
PropertySourcesPropertyResolver resolver,
PropertyNamePatternsMatcher includes) {
if (source.getPropertyNames().length > 0) {
for (String propertyName : source.getPropertyNames()) {
if (includes.matches(propertyName)) {
Object value = getEnumerableProperty(source, resolver, propertyName);
putIfAbsent(propertyName, value, source);
}
}
}
}
...
}
从以上代码来看，整个匹配的过程其实是在PropertySourcesPropertyValues对象构造的过程中完成的。
更具体地说，是在DefaultPropertyNamePatternsMatcher的matches函数中完成字符串匹配的，如下代码。
class DefaultPropertyNamePatternsMatcher implements PropertyNamePatternsMatcher {
...
public boolean matches(String propertyName) {
char[] propertyNameChars = propertyName.toCharArray();
boolean[] match = new boolean[this.names.length];
boolean noneMatched = true;
for (int i = 0; i < this.names.length; i++) {
if (this.names[i].length() <= propertyNameChars.length) {
match[i] = true;
noneMatched = false;
}
}
if (noneMatched) {
return false;
}
for (int charIndex = 0; charIndex < propertyNameChars.length; charIndex++) {
for (int nameIndex = 0; nameIndex < this.names.length; nameIndex++) {
if (match[nameIndex]) {
match[nameIndex] = false;
if (charIndex < this.names[nameIndex].length()) {
if (isCharMatch(this.names[nameIndex].charAt(charIndex),
propertyNameChars[charIndex])) {
match[nameIndex] = true;
noneMatched = false;
}
}
else {
char charAfter = propertyNameChars[this.names[nameIndex]
.length()];
if (isDelimiter(charAfter)) {
match[nameIndex] = true;
noneMatched = false;
}
}
}
}
if (noneMatched) {
return false;
}
}
for (int i = 0; i < match.length; i++) {
if (match[i]) {
return true;
}
}
return false;
}
private boolean isCharMatch(char c1, char c2) {
if (this.ignoreCase) {
return Character.toLowerCase(c1) == Character.toLowerCase(c2);
}
return c1 == c2;
}
private boolean isDelimiter(char c) {
for (char delimiter : this.delimiters) {
if (c == delimiter) {
return true;
}
}
return false;
}
}
如上代码，将字符串拆为单个字符进行比较，使用了双层循环匹配。
Spring Boot 2.0版本的实现（以2.0.4.RELEASE版本为例）
与1.5.14.RELEASE有很大区别。
属性与配置值的绑定逻辑依旧始于
ConfigurationPropertiesBindingPostProcessor 类的
postProcessBeforeInitialization 函数。
但 ConfigurationPropertiesBindingPostProcessor 类的定义与实现均发生了变化。先看代码。
public class ConfigurationPropertiesBindingPostProcessor implements BeanPostProcessor,
PriorityOrdered, ApplicationContextAware, InitializingBean {
...
private ConfigurationPropertiesBinder configurationPropertiesBinder;
...
@Override
public Object postProcessBeforeInitialization(Object bean, String beanName)
throws BeansException {
ConfigurationProperties annotation = getAnnotation(bean, beanName,
ConfigurationProperties.class);
if (annotation != null) {
bind(bean, beanName, annotation);
}
return bean;
}
private void bind(Object bean, String beanName, ConfigurationProperties annotation) {
ResolvableType type = getBeanType(bean, beanName);
Validated validated = getAnnotation(bean, beanName, Validated.class);
Annotation[] annotations = (validated != null)
? new Annotation[] { annotation, validated }
: new Annotation[] { annotation };
Bindable<?> target = Bindable.of(type).withExistingValue(bean)
.withAnnotations(annotations);
try {
this.configurationPropertiesBinder.bind(target);
}
...
}
...
}
从以上代码可以看出，该类依旧实现了BeanPostProcessor接口，但调用postProcessBeforeInitialization函数，属性值绑定的工作则被委派给了ConfigurationPropertiesBinder类，调用了bind函数。
class ConfigurationPropertiesBinder {
...
public void bind(Bindable<?> target) {
ConfigurationProperties annotation = target
.getAnnotation(ConfigurationProperties.class);
Assert.state(annotation != null,
() -> "Missing @ConfigurationProperties on " + target);
List<Validator> validators = getValidators(target);
BindHandler bindHandler = getBindHandler(annotation, validators);
getBinder().bind(annotation.prefix(), target, bindHandler);
}
private Binder getBinder() {
if (this.binder == null) {
this.binder = new Binder(getConfigurationPropertySources(),
getPropertySourcesPlaceholdersResolver(), getConversionService(),
getPropertyEditorInitializer());
}
return this.binder;
}
...
}
ConfigurationPropertiesBinder类并不是一个public的类。实际上这相当于ConfigurationPropertiesBindingPostProcessor的一个内部静态类，表面上负责处理@ConfigurationProperties注解的绑定任务。
但从代码中可以看出，具体的工作委派给另一个Binder类的对象。
Binder类是SpringBoot 2.0版本后加入的类，其负责处理对象与多个ConfigurationPropertySource之间的绑定。
public class Binder {
...
private static final List<BeanBinder> BEAN_BINDERS;
static {
List<BeanBinder> binders = new ArrayList<>();
binders.add(new JavaBeanBinder());
BEAN_BINDERS = Collections.unmodifiableList(binders);
}
...
public <T> BindResult<T> bind(String name, Bindable<T> target, BindHandler handler) {
return bind(ConfigurationPropertyName.of(name), target, handler);
}
public <T> BindResult<T> bind(ConfigurationPropertyName name, Bindable<T> target,
BindHandler handler) {
...
Context context = new Context();
T bound = bind(name, target, handler, context, false);
return BindResult.of(bound);
}
protected final <T> T bind(ConfigurationPropertyName name, Bindable<T> target,
BindHandler handler, Context context, boolean allowRecursiveBinding) {
...
try {
...
Object bound = bindObject(name, target, handler, context,
allowRecursiveBinding);
return handleBindResult(name, target, handler, context, bound);
}
...
}
private <T> Object bindObject(ConfigurationPropertyName name, Bindable<T> target,
BindHandler handler, Context context, boolean allowRecursiveBinding) {
ConfigurationProperty property = findProperty(name, context);
if (property == null && containsNoDescendantOf(context.streamSources(), name)) {
return null;
}
AggregateBinder<?> aggregateBinder = getAggregateBinder(target, context);
if (aggregateBinder != null) {
return bindAggregate(name, target, handler, context, aggregateBinder);
}
if (property != null) {
try {
return bindProperty(target, context, property);
}
catch (ConverterNotFoundException ex) {
// We might still be able to bind it as a bean
Object bean = bindBean(name, target, handler, context,
allowRecursiveBinding);
if (bean != null) {
return bean;
}
throw ex;
}
}
return bindBean(name, target, handler, context, allowRecursiveBinding);
}
private Object bindBean(ConfigurationPropertyName name, Bindable<?> target,
BindHandler handler, Context context, boolean allowRecursiveBinding) {
...
BeanPropertyBinder propertyBinder = (propertyName, propertyTarget) -> bind(
name.append(propertyName), propertyTarget, handler, context, false);
Class<?> type = target.getType().resolve(Object.class);
...
return context.withBean(type, () -> {
Stream<?> boundBeans = BEAN_BINDERS.stream()
.map((b) -> b.bind(name, target, context, propertyBinder));
return boundBeans.filter(Objects::nonNull).findFirst().orElse(null);
});
}
...
private ConfigurationProperty findProperty(ConfigurationPropertyName name,
Context context) {
...
return context.streamSources()
.map((source) -> source.getConfigurationProperty(name))
.filter(Objects::nonNull).findFirst().orElse(null);
}
}
如下代码，Binder类中实现了一个比较复杂的递归调用。
- ConfigurationPropertiesBinder 调用Binder类的bind函数时，参数通过层层转换，来到bindObject函数中。
- bindObject函数中，通过bindAggregate，bindProperty与bindBean等私有方法逐步推导绑定，bindBean是最后一步。
- bindBean函数通过定义BeanPropertyBinder的lambda表达式，允许bean绑定过程递归调用bindObject函数。
实际上，bindObject函数中findProperty函数调用是从properties文件中查找匹配项的关键，根据lambda表达式的执行结果，properties配置文件的配置项对应的是SpringIterableConfigurationPropertySource类型，因此调用的是其getConfigurationProperty函数，关键代码如下。
class SpringIterableConfigurationPropertySource extends SpringConfigurationPropertySource
implements IterableConfigurationPropertySource {
...
@Override
public ConfigurationProperty getConfigurationProperty(
ConfigurationPropertyName name) {
ConfigurationProperty configurationProperty = super.getConfigurationProperty(
name);
if (configurationProperty == null) {
configurationProperty = find(getPropertyMappings(getCache()), name);
}
return configurationProperty;
}
protected final ConfigurationProperty find(PropertyMapping[] mappings,
ConfigurationPropertyName name) {
for (PropertyMapping candidate : mappings) {
if (candidate.isApplicable(name)) {
ConfigurationProperty result = find(candidate);
if (result != null) {
return result;
}
}
}
return null;
}
...
}
最终，isApplicable函数中判断properties文件中的配置项（my.first_Name）与MyProperties类中的属性（firstName）是否匹配，其字符串比较过程使用的是ConfigurationPropertyName类重写的equals方法。
public final class ConfigurationPropertyName
implements Comparable<ConfigurationPropertyName> {
@Override
public boolean equals(Object obj) {
...
ConfigurationPropertyName other = (ConfigurationPropertyName) obj;
...
for (int i = 0; i < this.elements.length; i++) {
if (!elementEquals(this.elements[i], other.elements[i])) {
return false;
}
}
return true;
}
private boolean elementEquals(CharSequence e1, CharSequence e2) {
int l1 = e1.length();
int l2 = e2.length();
boolean indexed1 = isIndexed(e1);
int offset1 = indexed1 ? 1 : 0;
boolean indexed2 = isIndexed(e2);
int offset2 = indexed2 ? 1 : 0;
int i1 = offset1;
int i2 = offset2;
while (i1 < l1 - offset1) {
if (i2 >= l2 - offset2) {
return false;
}
char ch1 = indexed1 ? e1.charAt(i1) : Character.toLowerCase(e1.charAt(i1));
char ch2 = indexed2 ? e2.charAt(i2) : Character.toLowerCase(e2.charAt(i2));
if (ch1 == '-' || ch1 == '_') {
i1++;
}
else if (ch2 == '-' || ch2 == '_') {
i2++;
}
else if (ch1 != ch2) {
return false;
}
else {
i1++;
i2++;
}
}
while (i2 < l2 - offset2) {
char ch = e2.charAt(i2++);
if (ch != '-' && ch != '_') {
return false;
}
}
return true;
}
}
两者的异同
总结一下两个版本实现的不同。
| 对比科目 | Spring Boot-1.5.14.RELEASE版本 | Spring Boot-2.0.4.RELEASE版本 |
|---|---|---|
| java版本 | jdk1.7以上 | jdk1.8以上 |
| 关键类 | ConfigurationPropertiesBindingPostProcessor，PropertiesConfigurationFactory<T>，PropertySourcesPropertyValues，DefaultPropertyNamePatternsMatcher | ConfigurationPropertiesBindingPostProcessor，ConfigurationPropertiesBinder，SpringIterableConfigurationPropertySource，ConfigurationPropertyName |
| 查找方式 | 将属性拼接为字符串，穷举字符串的可能性 | 递归查找，分级匹配 |
| 字符串匹配方式 | 双层循环匹配 | 单层循环匹配 |
| 代码风格 | 传统java代码 | 大量使用lambda表达式 |
| 优点 | 由于使用传统代码风格，代码层次相对简单，可读性较强 | 不采用穷举的方式，匹配更精准，查找更有效 |
| 缺点 | 由于使用拼接穷举，当属性数量大且有“-”或“_”分隔单词时，组合会变多，影响查找效率 | 使用lambda表达式，产生各种延迟绑定，代码可读性差，不易调试 |
题外话：为什么我发现了这两个版本实现的不一致？
前一两个月，我写了一个自动识别消息的框架，在消息中模糊查找可能存在的关键信息，用了1.5版本的RelaxedNames类，虽然这个类在官方文档上没有提到，但是挺好用的。
然而最近，我想把框架迁移到2.0版本上，结果遇到了编译错误。因此特别研究了一下，发现这些类都不见了。
幸好只是小问题，我把RelaxedNames源码拷贝过来一份，就解决了问题，但这也给我提了个醒，大版本升级还是要仔细阅读下代码，避免留下一些坑。
另外也从一个侧面反映出，Spring Boot的2.0版本相比于1.5确实做了不小的更改，有空再好好琢磨下。