how to increase heap size?

LB4229 2012-04-02 07:32:55

win7使用Netbean
处理数列若一千个数据，跑着没问题，2000个数据速度比较慢。五千个数据就抱错，说heap size不够。
谁知道如何能提高这个heap size?看了下网上都是linux操作。本人不会linux，谁能给个windows不改系统的操作法？

见有用Regedit的，没敢尝试。sign。。

...全文

421 4 打赏收藏转发到动态举报

写回复

4 条回复

切换为时间正序

请发表友善的回复…

发表回复

xiazdong 2012-04-03

打赏
举报

回复

JVM PARAMETER里面设置 -Xmx1000M 表示分配1000M

dreamhunter_lan 2012-04-03

打赏
举报

回复

Netbean没用过，直接在命令行下运行的时候加上 -Xms512M和-Xmx1024M分别设置初始堆大小和堆的最大size

比如 java -Xms512M -Xmx1024M Test

eclipse下可以在run configuration中配置VM arguments设置相应参数

LB4229 2012-04-03

打赏
举报

回复

谢谢众位大牛回复哈。

TGITCIC 2012-04-03

打赏
举报

回复

就是你的java后面的参数， java -Xmx1024m -Xms1024m

不要直接拍脑袋说我把Xmx后的值给的越大越好，在给这个值时先用java -Xmx1024m -version这条命令测一下，如果能正常看到java的版本号，代表可以给到这个值。

32位机下jvm最大可给到2g，再大即使内存再多也没法给jvm了

另外，学JAVA的人不要用WINDOWS，用LINUX，JAVA本身就是UNIX下开发出来的，要依赖WINDOWS，学。NET吧。

举个例子来说，同样在WINODWS下，1000根线程在LINUX下可以达到2000，处理能力和速度都不一样的。

搞JAVA的人天生就和UNIX打交道，因为一些大型金融保险银行项目都不可能用WINDOWS去做他们的核心服务器的，太烂了，一定是UNIX，但UNNIX很贵，如：solary8都要好几百万，所以为了确保项目在移植时无误，就用LINUX来做环境。

LINUX就是运行在PC上的UNIX系统，学JAVA的人不会LINUX/UNIX实在无法在将来取得突破，要用WINDOWS请选。NET！

Table of Contents Header Files The #define Guard Header File Dependencies Inline Functions The -inl.h Files Function Parameter Ordering Names and Order of Includes Scoping Namespaces Nested Classes Nonmember, Static Member, and Global Functions Local Variables Static and Global Variables Classes Doing Work in Constructors Default Constructors Explicit Constructors Copy Constructors Structs vs. Classes Inheritance Multiple Inheritance Interfaces Operator Overloading Access Control Declaration Order Write Short Functions Google-Specific Magic Smart Pointers cpplint Other C++ Features Reference Arguments Function Overloading Default Arguments Variable-Length Arrays and alloca() Friends Exceptions Run-Time Type Information (RTTI) Casting Streams Preincrement and Predecrement Use of const Integer Types 64-bit Portability Preprocessor Macros 0 and NULL sizeof Boost C++0x Naming General Naming Rules File Names Type Names Variable Names Constant Names Function Names Namespace Names Enumerator Names Macro Names Exceptions to Naming Rules Comments Comment Style File Comments Class Comments Function Comments Variable Comments Implementation Comments Punctuation, Spelling and Grammar TODO Comments Deprecation Comments Formatting Line Length Non-ASCII Characters Spaces vs. Tabs Function Declarations and Definitions Function Calls Conditionals Loops and Switch Statements Pointer and Reference Expressions Boolean Expressions Return Values Variable and Array Initialization Preprocessor Directives Class Format Constructor Initializer Lists Namespace Formatting Horizontal Whitespace Vertical Whitespace Exceptions to the Rules Existing Non-conformant Code Windows Code Important Note Displaying Hidden Details in this Guide link ▶This style guide contains many details that are initially hidden from view. They are marked by the triangle icon, which you see here on your left. Click it now. You should see "Hooray" appear below. Hooray! Now you know you can expand points to get more details. Alternatively, there's an "expand all" at the top of this document. Background C++ is the main development language used by many of Google's open-source projects. As every C++ programmer knows, the language has many powerful features, but this power brings with it complexity, which in turn can make code more bug-prone and harder to read and maintain. The goal of this guide is to manage this complexity by describing in detail the dos and don'ts of writing C++ code. These rules exist to keep the code base manageable while still allowing coders to use C++ language features productively. Style, also known as readability, is what we call the conventions that govern our C++ code. The term Style is a bit of a misnomer, since these conventions cover far more than just source file formatting. One way in which we keep the code base manageable is by enforcing consistency. It is very important that any programmer be able to look at another's code and quickly understand it. Maintaining a uniform style and following conventions means that we can more easily use "pattern-matching" to infer what various symbols are and what invariants are true about them. Creating common, required idioms and patterns makes code much easier to understand. In some cases there might be good arguments for changing certain style rules, but we nonetheless keep things as they are in order to preserve consistency. Another issue this guide addresses is that of C++ feature bloat. C++ is a huge language with many advanced features. In some cases we constrain, or even ban, use of certain features. We do this to keep code simple and to avoid the various common errors and problems that these features can cause. This guide lists these features and explains why their use is restricted. Open-source projects developed by Google conform to the requirements in this guide. Note that this guide is not a C++ tutorial: we assume that the reader is familiar with the language. Header Files In general, every .cc file should have an associated .h file. There are some common exceptions, such as unittests and small .cc files containing just a main() function. Correct use of header files can make a huge difference to the readability, size and performance of your code. The following rules will guide you through the various pitfalls of using header files. The #define Guard link ▶All header files should have #define guards to prevent multiple inclusion. The format of the symbol name should be ___H_. To guarantee uniqueness, they should be based on the full path in a project's source tree. For example, the file foo/src/bar/baz.h in project foo should have the following guard: #ifndef FOO_BAR_BAZ_H_ #define FOO_BAR_BAZ_H_ ... #endif // FOO_BAR_BAZ_H_ Header File Dependencies link ▶Don't use an #include when a forward declaration would suffice. When you include a header file you introduce a dependency that will cause your code to be recompiled whenever the header file changes. If your header file includes other header files, any change to those files will cause any code that includes your header to be recompiled. Therefore, we prefer to minimize includes, particularly includes of header files in other header files. You can significantly minimize the number of header files you need to include in your own header files by using forward declarations. For example, if your header file uses the File class in ways that do not require access to the declaration of the File class, your header file can just forward declare class File; instead of having to #include "file/base/file.h". How can we use a class Foo in a header file without access to its definition? We can declare data members of type Foo* or Foo&. We can declare (but not define) functions with arguments, and/or return values, of type Foo. (One exception is if an argument Foo or const Foo& has a non-explicit, one-argument constructor, in which case we need the full definition to support automatic type conversion.) We can declare static data members of type Foo. This is because static data members are defined outside the class definition. On the other hand, you must include the header file for Foo if your class subclasses Foo or has a data member of type Foo. Sometimes it makes sense to have pointer (or better, scoped_ptr) members instead of object members. However, this complicates code readability and imposes a performance penalty, so avoid doing this transformation if the only purpose is to minimize includes in header files. Of course, .cc files typically do require the definitions of the classes they use, and usually have to include several header files. Note: If you use a symbol Foo in your source file, you should bring in a definition for Foo yourself, either via an #include or via a forward declaration. Do not depend on the symbol being brought in transitively via headers not directly included. One exception is if Foo is used in myfile.cc, it's ok to #include (or forward-declare) Foo in myfile.h, instead of myfile.cc. Inline Functions link ▶Define functions inline only when they are small, say, 10 lines or less. Definition: You can declare functions in a way that allows the compiler to expand them inline rather than calling them through the usual function call mechanism. Pros: Inlining a function can generate more efficient object code, as long as the inlined function is small. Feel free to inline accessors and mutators, and other short, performance-critical functions. Cons: Overuse of inlining can actually make programs slower. Depending on a function's size, inlining it can cause the code size to increase or decrease. Inlining a very small accessor function will usually decrease code size while inlining a very large function can dramatically increase code size. On modern processors smaller code usually runs faster due to better use of the instruction cache. Decision: A decent rule of thumb is to not inline a function if it is more than 10 lines long. Beware of destructors, which are often longer than they appear because of implicit member- and base-destructor calls! Another useful rule of thumb: it's typically not cost effective to inline functions with loops or switch statements (unless, in the common case, the loop or switch statement is never executed). It is important to know that functions are not always inlined even if they are declared as such; for example, virtual and recursive functions are not normally inlined. Usually recursive functions should not be inline. The main reason for making a virtual function inline is to place its definition in the class, either for convenience or to document its behavior, e.g., for accessors and mutators. The -inl.h Files link ▶You may use file names with a -inl.h suffix to define complex inline functions when needed. The definition of an inline function needs to be in a header file, so that the compiler has the definition available for inlining at the call sites. However, implementation code properly belongs in .cc files, and we do not like to have much actual code in .h files unless there is a readability or performance advantage. If an inline function definition is short, with very little, if any, logic in it, you should put the code in your .h file. For example, accessors and mutators should certainly be inside a class definition. More complex inline functions may also be put in a .h file for the convenience of the implementer and callers, though if this makes the .h file too unwieldy you can instead put that code in a separate -inl.h file. This separates the implementation from the class definition, while still allowing the implementation to be included where necessary. Another use of -inl.h files is for definitions of function templates. This can be used to keep your template definitions easy to read. Do not forget that a -inl.h file requires a #define guard just like any other header file. Function Parameter Ordering link ▶When defining a function, parameter order is: inputs, then outputs. Parameters to C/C++ functions are either input to the function, output from the function, or both. Input parameters are usually values or const references, while output and input/output parameters will be non-const pointers. When ordering function parameters, put all input-only parameters before any output parameters. In particular, do not add new parameters to the end of the function just because they are new; place new input-only parameters before the output parameters. This is not a hard-and-fast rule. Parameters that are both input and output (often classes/structs) muddy the waters, and, as always, consistency with related functions may require you to bend the rule. Names and Order of Includes link ▶Use standard order for readability and to avoid hidden dependencies: C library, C++ library, other libraries' .h, your project's .h. All of a project's header files should be listed as descentants of the project's source directory without use of UNIX directory shortcuts . (the current directory) or .. (the parent directory). For example, google-awesome-project/src/base/logging.h should be included as #include "base/logging.h" In dir/foo.cc, whose main purpose is to implement or test the stuff in dir2/foo2.h, order your includes as follows: dir2/foo2.h (preferred location — see details below). C system files. C++ system files. Other libraries' .h files. Your project's .h files. The preferred ordering reduces hidden dependencies. We want every header file to be compilable on its own. The easiest way to achieve this is to make sure that every one of them is the first .h file #included in some .cc. dir/foo.cc and dir2/foo2.h are often in the same directory (e.g. base/basictypes_test.cc and base/basictypes.h), but can be in different directories too. Within each section it is nice to order the includes alphabetically. For example, the includes in google-awesome-project/src/foo/internal/fooserver.cc might look like this: #include "foo/public/fooserver.h" // Preferred location. #include #include #include #include #include "base/basictypes.h" #include "base/commandlineflags.h" #include "foo/public/bar.h" Scoping Namespaces link ▶Unnamed namespaces in .cc files are encouraged. With named namespaces, choose the name based on the project, and possibly its path. Do not use a using-directive. Definition: Namespaces subdivide the global scope into distinct, named scopes, and so are useful for preventing name collisions in the global scope. Pros: Namespaces provide a (hierarchical) axis of naming, in addition to the (also hierarchical) name axis provided by classes. For example, if two different projects have a class Foo in the global scope, these symbols may collide at compile time or at runtime. If each project places their code in a namespace, project1::Foo and project2::Foo are now distinct symbols that do not collide. Cons: Namespaces can be confusing, because they provide an additional (hierarchical) axis of naming, in addition to the (also hierarchical) name axis provided by classes. Use of unnamed spaces in header files can easily cause violations of the C++ One Definition Rule (ODR). Decision: Use namespaces according to the policy described below. Unnamed Namespaces Unnamed namespaces are allowed and even encouraged in .cc files, to avoid runtime naming conflicts: namespace { // This is in a .cc file. // The content of a namespace is not indented enum { kUnused, kEOF, kError }; // Commonly used tokens. bool AtEof() { return pos_ == kEOF; } // Uses our namespace's EOF. } // namespace However, file-scope declarations that are associated with a particular class may be declared in that class as types, static data members or static member functions rather than as members of an unnamed namespace. Terminate the unnamed namespace as shown, with a comment // namespace. Do not use unnamed namespaces in .h files. Named Namespaces Named namespaces should be used as follows: Namespaces wrap the entire source file after includes, gflags definitions/declarations, and forward declarations of classes from other namespaces: // In the .h file namespace mynamespace { // All declarations are within the namespace scope. // Notice the lack of indentation. class MyClass { public: ... void Foo(); }; } // namespace mynamespace // In the .cc file namespace mynamespace { // Definition of functions is within scope of the namespace. void MyClass::Foo() { ... } } // namespace mynamespace The typical .cc file might have more complex detail, including the need to reference classes in other namespaces. #include "a.h" DEFINE_bool(someflag, false, "dummy flag"); class C; // Forward declaration of class C in the global namespace. namespace a { class A; } // Forward declaration of a::A. namespace b { ...code for b... // Code goes against the left margin. } // namespace b Do not declare anything in namespace std, not even forward declarations of standard library classes. Declaring entities in namespace std is undefined behavior, i.e., not portable. To declare entities from the standard library, include the appropriate header file. You may not use a using-directive to make all names from a namespace available. // Forbidden -- This pollutes the namespace. using namespace foo; You may use a using-declaration anywhere in a .cc file, and in functions, methods or classes in .h files. // OK in .cc files. // Must be in a function, method or class in .h files. using ::foo::bar; Namespace aliases are allowed anywhere in a .cc file, anywhere inside the named namespace that wraps an entire .h file, and in functions and methods. // Shorten access to some commonly used names in .cc files. namespace fbz = ::foo::bar::baz; // Shorten access to some commonly used names (in a .h file). namespace librarian { // The following alias is available to all files including // this header (in namespace librarian): // alias names should therefore be chosen consistently // within a project. namespace pd_s = ::pipeline_diagnostics::sidetable; inline void my_inline_function() { // namespace alias local to a function (or method). namespace fbz = ::foo::bar::baz; ... } } // namespace librarian Note that an alias in a .h file is visible to everyone #including that file, so public headers (those available outside a project) and headers transitively #included by them, should avoid defining aliases, as part of the general goal of keeping public APIs as small as possible. Nested Classes link ▶Although you may use public nested classes when they are part of an interface, consider a namespace to keep declarations out of the global scope. Definition: A class can define another class within it; this is also called a member class. class Foo { private: // Bar is a member class, nested within Foo. class Bar { ... }; }; Pros: This is useful when the nested (or member) class is only used by the enclosing class; making it a member puts it in the enclosing class scope rather than polluting the outer scope with the class name. Nested classes can be forward declared within the enclosing class and then defined in the .cc file to avoid including the nested class definition in the enclosing class declaration, since the nested class definition is usually only relevant to the implementation. Cons: Nested classes can be forward-declared only within the definition of the enclosing class. Thus, any header file manipulating a Foo::Bar* pointer will have to include the full class declaration for Foo. Decision: Do not make nested classes public unless they are actually part of the interface, e.g., a class that holds a set of options for some method. Nonmember, Static Member, and Global Functions link ▶Prefer nonmember functions within a namespace or static member functions to global functions; use completely global functions rarely. Pros: Nonmember and static member functions can be useful in some situations. Putting nonmember functions in a namespace avoids polluting the global namespace. Cons: Nonmember and static member functions may make more sense as members of a new class, especially if they access external resources or have significant dependencies. Decision: Sometimes it is useful, or even necessary, to define a function not bound to a class instance. Such a function can be either a static member or a nonmember function. Nonmember functions should not depend on external variables, and should nearly always exist in a namespace. Rather than creating classes only to group static member functions which do not share static data, use namespaces instead. Functions defined in the same compilation unit as production classes may introduce unnecessary coupling and link-time dependencies when directly called from other compilation units; static member functions are particularly susceptible to this. Consider extracting a new class, or placing the functions in a namespace possibly in a separate library. If you must define a nonmember function and it is only needed in its .cc file, use an unnamed namespace or static linkage (eg static int Foo() {...}) to limit its scope. Local Variables link ▶Place a function's variables in the narrowest scope possible, and initialize variables in the declaration. C++ allows you to declare variables anywhere in a function. We encourage you to declare them in as local a scope as possible, and as close to the first use as possible. This makes it easier for the reader to find the declaration and see what type the variable is and what it was initialized to. In particular, initialization should be used instead of declaration and assignment, e.g. int i; i = f(); // Bad -- initialization separate from declaration. int j = g(); // Good -- declaration has initialization. Note that gcc implements for (int i = 0; i < 10; ++i) correctly (the scope of i is only the scope of the for loop), so you can then reuse i in another for loop in the same scope. It also correctly scopes declarations in if and while statements, e.g. while (const char* p = strchr(str, '/')) str = p + 1; There is one caveat: if the variable is an object, its constructor is invoked every time it enters scope and is created, and its destructor is invoked every time it goes out of scope. // Inefficient implementation: for (int i = 0; i < 1000000; ++i) { Foo f; // My ctor and dtor get called 1000000 times each. f.DoSomething(i); } It may be more efficient to declare such a variable used in a loop outside that loop: Foo f; // My ctor and dtor get called once each. for (int i = 0; i < 1000000; ++i) { f.DoSomething(i); } Static and Global Variables link ▶Static or global variables of class type are forbidden: they cause hard-to-find bugs due to indeterminate order of construction and destruction. Objects with static storage duration, including global variables, static variables, static class member variables, and function static variables, must be Plain Old Data (POD): only ints, chars, floats, or pointers, or arrays/structs of POD. The order in which class constructors and initializers for static variables are called is only partially specified in C++ and can even change from build to build, which can cause bugs that are difficult to find. Therefore in addition to banning globals of class type, we do not allow static POD variables to be initialized with the result of a function, unless that function (such as getenv(), or getpid()) does not itself depend on any other globals. Likewise, the order in which destructors are called is defined to be the reverse of the order in which the constructors were called. Since constructor order is indeterminate, so is destructor order. For example, at program-end time a static variable might have been destroyed, but code still running -- perhaps in another thread -- tries to access it and fails. Or the destructor for a static 'string' variable might be run prior to the destructor for another variable that contains a reference to that string. As a result we only allow static variables to contain POD data. This rule completely disallows vector (use C arrays instead), or string (use const char []). If you need a static or global variable of a class type, consider initializing a pointer (which will never be freed), from either your main() function or from pthread_once(). Note that this must be a raw pointer, not a "smart" pointer, since the smart pointer's destructor will have the order-of-destructor issue that we are trying to avoid. Classes Classes are the fundamental unit of code in C++. Naturally, we use them extensively. This section lists the main dos and don'ts you should follow when writing a class. Doing Work in Constructors link ▶In general, constructors should merely set member variables to their initial values. Any complex initialization should go in an explicit Init() method. Definition: It is possible to perform initialization in the body of the constructor. Pros: Convenience in typing. No need to worry about whether the class has been initialized or not. Cons: The problems with doing work in constructors are: There is no easy way for constructors to signal errors, short of using exceptions (which are forbidden). If the work fails, we now have an object whose initialization code failed, so it may be an indeterminate state. If the work calls virtual functions, these calls will not get dispatched to the subclass implementations. Future modification to your class can quietly introduce this problem even if your class is not currently subclassed, causing much confusion. If someone creates a global variable of this type (which is against the rules, but still), the constructor code will be called before main(), possibly breaking some implicit assumptions in the constructor code. For instance, gflags will not yet have been initialized. Decision: If your object requires non-trivial initialization, consider having an explicit Init() method. In particular, constructors should not call virtual functions, attempt to raise errors, access potentially uninitialized global variables, etc. Default Constructors link ▶You must define a default constructor if your class defines member variables and has no other constructors. Otherwise the compiler will do it for you, badly. Definition: The default constructor is called when we new a class object with no arguments. It is always called when calling new[] (for arrays). Pros: Initializing structures by default, to hold "impossible" values, makes debugging much easier. Cons: Extra work for you, the code writer. Decision: If your class defines member variables and has no other constructors you must define a default constructor (one that takes no arguments). It should preferably initialize the object in such a way that its internal state is consistent and valid. The reason for this is that if you have no other constructors and do not define a default constructor, the compiler will generate one for you. This compiler generated constructor may not initialize your object sensibly. If your class inherits from an existing class but you add no new member variables, you are not required to have a default constructor. Explicit Constructors link ▶Use the C++ keyword explicit for constructors with one argument. Definition: Normally, if a constructor takes one argument, it can be used as a conversion. For instance, if you define Foo::Foo(string name) and then pass a string to a function that expects a Foo, the constructor will be called to convert the string into a Foo and will pass the Foo to your function for you. This can be convenient but is also a source of trouble when things get converted and new objects created without you meaning them to. Declaring a constructor explicit prevents it from being invoked implicitly as a conversion. Pros: Avoids undesirable conversions. Cons: None. Decision: We require all single argument constructors to be explicit. Always put explicit in front of one-argument constructors in the class definition: explicit Foo(string name); The exception is copy constructors, which, in the rare cases when we allow them, should probably not be explicit. Classes that are intended to be transparent wrappers around other classes are also exceptions. Such exceptions should be clearly marked with comments. Copy Constructors link ▶Provide a copy constructor and assignment operator only when necessary. Otherwise, disable them with DISALLOW_COPY_AND_ASSIGN. Definition: The copy constructor and assignment operator are used to create copies of objects. The copy constructor is implicitly invoked by the compiler in some situations, e.g. passing objects by value. Pros: Copy constructors make it easy to copy objects. STL containers require that all contents be copyable and assignable. Copy constructors can be more efficient than CopyFrom()-style workarounds because they combine construction with copying, the compiler can elide them in some contexts, and they make it easier to avoid heap allocation. Cons: Implicit copying of objects in C++ is a rich source of bugs and of performance problems. It also reduces readability, as it becomes hard to track which objects are being passed around by value as opposed to by reference, and therefore where changes to an object are reflected. Decision: Few classes need to be copyable. Most should have neither a copy constructor nor an assignment operator. In many situations, a pointer or reference will work just as well as a copied value, with better performance. For example, you can pass function parameters by reference or pointer instead of by value, and you can store pointers rather than objects in an STL container. If your class needs to be copyable, prefer providing a copy method, such as CopyFrom() or Clone(), rather than a copy constructor, because such methods cannot be invoked implicitly. If a copy method is insufficient in your situation (e.g. for performance reasons, or because your class needs to be stored by value in an STL container), provide both a copy constructor and assignment operator. If your class does not need a copy constructor or assignment operator, you must explicitly disable them. To do so, add dummy declarations for the copy constructor and assignment operator in the private: section of your class, but do not provide any corresponding definition (so that any attempt to use them results in a link error). For convenience, a DISALLOW_COPY_AND_ASSIGN macro can be used: // A macro to disallow the copy constructor and operator= functions // This should be used in the private: declarations for a class #define DISALLOW_COPY_AND_ASSIGN(TypeName) \ TypeName(const TypeName&); \ void operator=(const TypeName&) Then, in class Foo: class Foo { public: Foo(int f); ~Foo(); private: DISALLOW_COPY_AND_ASSIGN(Foo); }; Structs vs. Classes link ▶Use a struct only for passive objects that carry data; everything else is a class. The struct and class keywords behave almost identically in C++. We add our own semantic meanings to each keyword, so you should use the appropriate keyword for the data-type you're defining. structs should be used for passive objects that carry data, and may have associated constants, but lack any functionality other than access/setting the data members. The accessing/setting of fields is done by directly accessing the fields rather than through method invocations. Methods should not provide behavior but should only be used to set up the data members, e.g., constructor, destructor, Initialize(), Reset(), Validate(). If more functionality is required, a class is more appropriate. If in doubt, make it a class. For consistency with STL, you can use struct instead of class for functors and traits. Note that member variables in structs and classes have different naming rules. Inheritance link ▶Composition is often more appropriate than inheritance. When using inheritance, make it public. Definition: When a sub-class inherits from a base class, it includes the definitions of all the data and operations that the parent base class defines. In practice, inheritance is used in two major ways in C++: implementation inheritance, in which actual code is inherited by the child, and interface inheritance, in which only method names are inherited. Pros: Implementation inheritance reduces code size by re-using the base class code as it specializes an existing type. Because inheritance is a compile-time declaration, you and the compiler can understand the operation and detect errors. Interface inheritance can be used to programmatically enforce that a class expose a particular API. Again, the compiler can detect errors, in this case, when a class does not define a necessary method of the API. Cons: For implementation inheritance, because the code implementing a sub-class is spread between the base and the sub-class, it can be more difficult to understand an implementation. The sub-class cannot override functions that are not virtual, so the sub-class cannot change implementation. The base class may also define some data members, so that specifies physical layout of the base class. Decision: All inheritance should be public. If you want to do private inheritance, you should be including an instance of the base class as a member instead. Do not overuse implementation inheritance. Composition is often more appropriate. Try to restrict use of inheritance to the "is-a" case: Bar subclasses Foo if it can reasonably be said that Bar "is a kind of" Foo. Make your destructor virtual if necessary. If your class has virtual methods, its destructor should be virtual. Limit the use of protected to those member functions that might need to be accessed from subclasses. Note that data members should be private. When redefining an inherited virtual function, explicitly declare it virtual in the declaration of the derived class. Rationale: If virtual is omitted, the reader has to check all ancestors of the class in question to determine if the function is virtual or not. Multiple Inheritance link ▶Only very rarely is multiple implementation inheritance actually useful. We allow multiple inheritance only when at most one of the base classes has an implementation; all other base classes must be pure interface classes tagged with the Interface suffix. Definition: Multiple inheritance allows a sub-class to have more than one base class. We distinguish between base classes that are pure interfaces and those that have an implementation. Pros: Multiple implementation inheritance may let you re-use even more code than single inheritance (see Inheritance). Cons: Only very rarely is multiple implementation inheritance actually useful. When multiple implementation inheritance seems like the solution, you can usually find a different, more explicit, and cleaner solution. Decision: Multiple inheritance is allowed only when all superclasses, with the possible exception of the first one, are pure interfaces. In order to ensure that they remain pure interfaces, they must end with the Interface suffix. Note: There is an exception to this rule on Windows. Interfaces link ▶Classes that satisfy certain conditions are allowed, but not required, to end with an Interface suffix. Definition: A class is a pure interface if it meets the following requirements: It has only public pure virtual ("= 0") methods and static methods (but see below for destructor). It may not have non-static data members. It need not have any constructors defined. If a constructor is provided, it must take no arguments and it must be protected. If it is a subclass, it may only be derived from classes that satisfy these conditions and are tagged with the Interface suffix. An interface class can never be directly instantiated because of the pure virtual method(s) it declares. To make sure all implementations of the interface can be destroyed correctly, they must also declare a virtual destructor (in an exception to the first rule, this should not be pure). See Stroustrup, The C++ Programming Language, 3rd edition, section 12.4 for details. Pros: Tagging a class with the Interface suffix lets others know that they must not add implemented methods or non static data members. This is particularly important in the case of multiple inheritance. Additionally, the interface concept is already well-understood by Java programmers. Cons: The Interface suffix lengthens the class name, which can make it harder to read and understand. Also, the interface property may be considered an implementation detail that shouldn't be exposed to clients. Decision: A class may end with Interface only if it meets the above requirements. We do not require the converse, however: classes that meet the above requirements are not required to end with Interface. Operator Overloading link ▶Do not overload operators except in rare, special circumstances. Definition: A class can define that operators such as + and / operate on the class as if it were a built-in type. Pros: Can make code appear more intuitive because a class will behave in the same way as built-in types (such as int). Overloaded operators are more playful names for functions that are less-colorfully named, such as Equals() or Add(). For some template functions to work correctly, you may need to define operators. Cons: While operator overloading can make code more intuitive, it has several drawbacks: It can fool our intuition into thinking that expensive operations are cheap, built-in operations. It is much harder to find the call sites for overloaded operators. Searching for Equals() is much easier than searching for relevant invocations of ==. Some operators work on pointers too, making it easy to introduce bugs. Foo + 4 may do one thing, while &Foo + 4 does something totally different. The compiler does not complain for either of these, making this very hard to debug. Overloading also has surprising ramifications. For instance, if a class overloads unary operator&, it cannot safely be forward-declared. Decision: In general, do not overload operators. The assignment operator (operator=), in particular, is insidious and should be avoided. You can define functions like Equals() and CopyFrom() if you need them. Likewise, avoid the dangerous unary operator& at all costs, if there's any possibility the class might be forward-declared. However, there may be rare cases where you need to overload an operator to interoperate with templates or "standard" C++ classes (such as operator<<(ostream&, const T&) for logging). These are acceptable if fully justified, but you should try to avoid these whenever possible. In particular, do not overload operator== or operator< just so that your class can be used as a key in an STL container; instead, you should create equality and comparison functor types when declaring the container. Some of the STL algorithms do require you to overload operator==, and you may do so in these cases, provided you document why. See also Copy Constructors and Function Overloading. Access Control link ▶Make data members private, and provide access to them through accessor functions as needed (for technical reasons, we allow data members of a test fixture class to be protected when using Google Test). Typically a variable would be called foo_ and the accessor function foo(). You may also want a mutator function set_foo(). Exception: static const data members (typically called kFoo) need not be private. The definitions of accessors are usually inlined in the header file. See also Inheritance and Function Names. Declaration Order link ▶Use the specified order of declarations within a class: public: before private:, methods before data members (variables), etc. Your class definition should start with its public: section, followed by its protected: section and then its private: section. If any of these sections are empty, omit them. Within each section, the declarations generally should be in the following order: Typedefs and Enums Constants (static const data members) Constructors Destructor Methods, including static methods Data Members (except static const data members) Friend declarations should always be in the private section, and the DISALLOW_COPY_AND_ASSIGN macro invocation should be at the end of the private: section. It should be the last thing in the class. See Copy Constructors. Method definitions in the corresponding .cc file should be the same as the declaration order, as much as possible. Do not put large method definitions inline in the class definition. Usually, only trivial or performance-critical, and very short, methods may be defined inline. See Inline Functions for more details. Write Short Functions link ▶Prefer small and focused functions. We recognize that long functions are sometimes appropriate, so no hard limit is placed on functions length. If a function exceeds about 40 lines, think about whether it can be broken up without harming the structure of the program. Even if your long function works perfectly now, someone modifying it in a few months may add new behavior. This could result in bugs that are hard to find. Keeping your functions short and simple makes it easier for other people to read and modify your code. You could find long and complicated functions when working with some code. Do not be intimidated by modifying existing code: if working with such a function proves to be difficult, you find that errors are hard to debug, or you want to use a piece of it in several different contexts, consider breaking up the function into smaller and more manageable pieces. Google-Specific Magic There are various tricks and utilities that we use to make C++ code more robust, and various ways we use C++ that may differ from what you see elsewhere. Smart Pointers link ▶If you actually need pointer semantics, scoped_ptr is great. You should only use std::tr1::shared_ptr under very specific conditions, such as when objects need to be held by STL containers. You should never use auto_ptr. "Smart" pointers are objects that act like pointers but have added semantics. When a scoped_ptr is destroyed, for instance, it deletes the object it's pointing to. shared_ptr is the same way, but implements reference-counting so only the last pointer to an object deletes it. Generally speaking, we prefer that we design code with clear object ownership. The clearest object ownership is obtained by using an object directly as a field or local variable, without using pointers at all. On the other extreme, by their very definition, reference counted pointers are owned by nobody. The problem with this design is that it is easy to create circular references or other strange conditions that cause an object to never be deleted. It is also slow to perform atomic operations every time a value is copied or assigned. Although they are not recommended, reference counted pointers are sometimes the simplest and most elegant way to solve a problem. cpplint link ▶Use cpplint.py to detect style errors. cpplint.py is a tool that reads a source file and identifies many style errors. It is not perfect, and has both false positives and false negatives, but it is still a valuable tool. False positives can be ignored by putting // NOLINT at the end of the line. Some projects have instructions on how to run cpplint.py from their project tools. If the project you are contributing to does not, you can download cpplint.py separately. Other C++ Features Reference Arguments link ▶All parameters passed by reference must be labeled const. Definition: In C, if a function needs to modify a variable, the parameter must use a pointer, eg int foo(int *pval). In C++, the function can alternatively declare a reference parameter: int foo(int &val). Pros: Defining a parameter as reference avoids ugly code like (*pval)++. Necessary for some applications like copy constructors. Makes it clear, unlike with pointers, that NULL is not a possible value. Cons: References can be confusing, as they have value syntax but pointer semantics. Decision: Within function parameter lists all references must be const: void Foo(const string &in, string *out); In fact it is a very strong convention in Google code that input arguments are values or const references while output arguments are pointers. Input parameters may be const pointers, but we never allow non-const reference parameters. One case when you might want an input parameter to be a const pointer is if you want to emphasize that the argument is not copied, so it must exist for the lifetime of the object; it is usually best to document this in comments as well. STL adapters such as bind2nd and mem_fun do not permit reference parameters, so you must declare functions with pointer parameters in these cases, too. Function Overloading link ▶Use overloaded functions (including constructors) only if a reader looking at a call site can get a good idea of what is happening without having to first figure out exactly which overload is being called. Definition: You may write a function that takes a const string& and overload it with another that takes const char*. class MyClass { public: void Analyze(const string &text); void Analyze(const char *text, size_t textlen); }; Pros: Overloading can make code more intuitive by allowing an identically-named function to take different arguments. It may be necessary for templatized code, and it can be convenient for Visitors. Cons: If a function is overloaded by the argument types alone, a reader may have to understand C++'s complex matching rules in order to tell what's going on. Also many people are confused by the semantics of inheritance if a derived class overrides only some of the variants of a function. Decision: If you want to overload a function, consider qualifying the name with some information about the arguments, e.g., AppendString(), AppendInt() rather than just Append(). Default Arguments link ▶We do not allow default function parameters, except in a few uncommon situations explained below. Pros: Often you have a function that uses lots of default values, but occasionally you want to override the defaults. Default parameters allow an easy way to do this without having to define many functions for the rare exceptions. Cons: People often figure out how to use an API by looking at existing code that uses it. Default parameters are more difficult to maintain because copy-and-paste from previous code may not reveal all the parameters. Copy-and-pasting of code segments can cause major problems when the default arguments are not appropriate for the new code. Decision: Except as described below, we require all arguments to be explicitly specified, to force programmers to consider the API and the values they are passing for each argument rather than silently accepting defaults they may not be aware of. One specific exception is when default arguments are used to simulate variable-length argument lists. // Support up to 4 params by using a default empty AlphaNum. string StrCat(const AlphaNum &a, const AlphaNum &b = gEmptyAlphaNum, const AlphaNum &c = gEmptyAlphaNum, const AlphaNum &d = gEmptyAlphaNum); Variable-Length Arrays and alloca() link ▶We do not allow variable-length arrays or alloca(). Pros: Variable-length arrays have natural-looking syntax. Both variable-length arrays and alloca() are very efficient. Cons: Variable-length arrays and alloca are not part of Standard C++. More importantly, they allocate a data-dependent amount of stack space that can trigger difficult-to-find memory overwriting bugs: "It ran fine on my machine, but dies mysteriously in production". Decision: Use a safe allocator instead, such as scoped_ptr/scoped_array. Friends link ▶We allow use of friend classes and functions, within reason. Friends should usually be defined in the same file so that the reader does not have to look in another file to find uses of the private members of a class. A common use of friend is to have a FooBuilder class be a friend of Foo so that it can construct the inner state of Foo correctly, without exposing this state to the world. In some cases it may be useful to make a unittest class a friend of the class it tests. Friends extend, but do not break, the encapsulation boundary of a class. In some cases this is better than making a member public when you want to give only one other class access to it. However, most classes should interact with other classes solely through their public members. Exceptions link ▶We do not use C++ exceptions. Pros: Exceptions allow higher levels of an application to decide how to handle "can't happen" failures in deeply nested functions, without the obscuring and error-prone bookkeeping of error codes. Exceptions are used by most other modern languages. Using them in C++ would make it more consistent with Python, Java, and the C++ that others are familiar with. Some third-party C++ libraries use exceptions, and turning them off internally makes it harder to integrate with those libraries. Exceptions are the only way for a constructor to fail. We can simulate this with a factory function or an Init() method, but these require heap allocation or a new "invalid" state, respectively. Exceptions are really handy in testing frameworks. Cons: When you add a throw statement to an existing function, you must examine all of its transitive callers. Either they must make at least the basic exception safety guarantee, or they must never catch the exception and be happy with the program terminating as a result. For instance, if f() calls g() calls h(), and h throws an exception that f catches, g has to be careful or it may not clean up properly. More generally, exceptions make the control flow of programs difficult to evaluate by looking at code: functions may return in places you don't expect. This causes maintainability and debugging difficulties. You can minimize this cost via some rules on how and where exceptions can be used, but at the cost of more that a developer needs to know and understand. Exception safety requires both RAII and different coding practices. Lots of supporting machinery is needed to make writing correct exception-safe code easy. Further, to avoid requiring readers to understand the entire call graph, exception-safe code must isolate logic that writes to persistent state into a "commit" phase. This will have both benefits and costs (perhaps where you're forced to obfuscate code to isolate the commit). Allowing exceptions would force us to always pay those costs even when they're not worth it. Turning on exceptions adds data to each binary produced, increasing compile time (probably slightly) and possibly increasing address space pressure. The availability of exceptions may encourage developers to throw them when they are not appropriate or recover from them when it's not safe to do so. For example, invalid user input should not cause exceptions to be thrown. We would need to make the style guide even longer to document these restrictions! Decision: On their face, the benefits of using exceptions outweigh the costs, especially in new projects. However, for existing code, the introduction of exceptions has implications on all dependent code. If exceptions can be propagated beyond a new project, it also becomes problematic to integrate the new project into existing exception-free code. Because most existing C++ code at Google is not prepared to deal with exceptions, it is comparatively difficult to adopt new code that generates exceptions. Given that Google's existing code is not exception-tolerant, the costs of using exceptions are somewhat greater than the costs in a new project. The conversion process would be slow and error-prone. We don't believe that the available alternatives to exceptions, such as error codes and assertions, introduce a significant burden. Our advice against using exceptions is not predicated on philosophical or moral grounds, but practical ones. Because we'd like to use our open-source projects at Google and it's difficult to do so if those projects use exceptions, we need to advise against exceptions in Google open-source projects as well. Things would probably be different if we had to do it all over again from scratch. There is an exception to this rule (no pun intended) for Windows code. Run-Time Type Information (RTTI) link ▶We do not use Run Time Type Information (RTTI). Definition: RTTI allows a programmer to query the C++ class of an object at run time. Pros: It is useful in some unittests. For example, it is useful in tests of factory classes where the test has to verify that a newly created object has the expected dynamic type. In rare circumstances, it is useful even outside of tests. Cons: A query of type during run-time typically means a design problem. If you need to know the type of an object at runtime, that is often an indication that you should reconsider the design of your class. Decision: Do not use RTTI, except in unittests. If you find yourself in need of writing code that behaves differently based on the class of an object, consider one of the alternatives to querying the type. Virtual methods are the preferred way of executing different code paths depending on a specific subclass type. This puts the work within the object itself. If the work belongs outside the object and instead in some processing code, consider a double-dispatch solution, such as the Visitor design pattern. This allows a facility outside the object itself to determine the type of class using the built-in type system. If you think you truly cannot use those ideas, you may use RTTI. But think twice about it. :-) Then think twice again. Do not hand-implement an RTTI-like workaround. The arguments against RTTI apply just as much to workarounds like class hierarchies with type tags. Casting link ▶Use C++ casts like static_cast(). Do not use other cast formats like int y = (int)x; or int y = int(x);. Definition: C++ introduced a different cast system from C that distinguishes the types of cast operations. Pros: The problem with C casts is the ambiguity of the operation; sometimes you are doing a conversion (e.g., (int)3.5) and sometimes you are doing a cast (e.g., (int)"hello"); C++ casts avoid this. Additionally C++ casts are more visible when searching for them. Cons: The syntax is nasty. Decision: Do not use C-style casts. Instead, use these C++-style casts. Use static_cast as the equivalent of a C-style cast that does value conversion, or when you need to explicitly up-cast a pointer from a class to its superclass. Use const_cast to remove the const qualifier (see const). Use reinterpret_cast to do unsafe conversions of pointer types to and from integer and other pointer types. Use this only if you know what you are doing and you understand the aliasing issues. Do not use dynamic_cast except in test code. If you need to know type information at runtime in this way outside of a unittest, you probably have a design flaw. Streams link ▶Use streams only for logging. Definition: Streams are a replacement for printf() and scanf(). Pros: With streams, you do not need to know the type of the object you are printing. You do not have problems with format strings not matching the argument list. (Though with gcc, you do not have that problem with printf either.) Streams have automatic constructors and destructors that open and close the relevant files. Cons: Streams make it difficult to do functionality like pread(). Some formatting (particularly the common format string idiom %.*s) is difficult if not impossible to do efficiently using streams without using printf-like hacks. Streams do not support operator reordering (the %1s directive), which is helpful for internationalization. Decision: Do not use streams, except where required by a logging interface. Use printf-like routines instead. There are various pros and cons to using streams, but in this case, as in many other cases, consistency trumps the debate. Do not use streams in your code. Extended Discussion There has been debate on this issue, so this explains the reasoning in greater depth. Recall the Only One Way guiding principle: we want to make sure that whenever we do a certain type of I/O, the code looks the same in all those places. Because of this, we do not want to allow users to decide between using streams or using printf plus Read/Write/etc. Instead, we should settle on one or the other. We made an exception for logging because it is a pretty specialized application, and for historical reasons. Proponents of streams have argued that streams are the obvious choice of the two, but the issue is not actually so clear. For every advantage of streams they point out, there is an equivalent disadvantage. The biggest advantage is that you do not need to know the type of the object to be printing. This is a fair point. But, there is a downside: you can easily use the wrong type, and the compiler will not warn you. It is easy to make this kind of mistake without knowing when using streams. cout << this; // Prints the address cout << *this; // Prints the contents The compiler does not generate an error because << has been overloaded. We discourage overloading for just this reason. Some say printf formatting is ugly and hard to read, but streams are often no better. Consider the following two fragments, both with the same typo. Which is easier to discover? cerr << "Error connecting to '" hostname.first << ":" hostname.second << ": " hostname.first, foo->bar()->hostname.second, strerror(errno)); And so on and so forth for any issue you might bring up. (You could argue, "Things would be better with the right wrappers," but if it is true for one scheme, is it not also true for the other? Also, remember the goal is to make the language smaller, not add yet more machinery that someone has to learn.) Either path would yield different advantages and disadvantages, and there is not a clearly superior solution. The simplicity doctrine mandates we settle on one of them though, and the majority decision was on printf + read/write. Preincrement and Predecrement link ▶Use prefix form (++i) of the increment and decrement operators with iterators and other template objects. Definition: When a variable is incremented (++i or i++) or decremented (--i or i--) and the value of the expression is not used, one must decide whether to preincrement (decrement) or postincrement (decrement). Pros: When the return value is ignored, the "pre" form (++i) is never less efficient than the "post" form (i++), and is often more efficient. This is because post-increment (or decrement) requires a copy of i to be made, which is the value of the expression. If i is an iterator or other non-scalar type, copying i could be expensive. Since the two types of increment behave the same when the value is ignored, why not just always pre-increment? Cons: The tradition developed, in C, of using post-increment when the expression value is not used, especially in for loops. Some find post-increment easier to read, since the "subject" (i) precedes the "verb" (++), just like in English. Decision: For simple scalar (non-object) values there is no reason to prefer one form and we allow either. For iterators and other template types, use pre-increment. Use of const link ▶We strongly recommend that you use const whenever it makes sense to do so. Definition: Declared variables and parameters can be preceded by the keyword const to indicate the variables are not changed (e.g., const int foo). Class functions can have the const qualifier to indicate the function does not change the state of the class member variables (e.g., class Foo { int Bar(char c) const; };). Pros: Easier for people to understand how variables are being used. Allows the compiler to do better type checking, and, conceivably, generate better code. Helps people convince themselves of program correctness because they know the functions they call are limited in how they can modify your variables. Helps people know what functions are safe to use without locks in multi-threaded programs. Cons: const is viral: if you pass a const variable to a function, that function must have const in its prototype (or the variable will need a const_cast). This can be a particular problem when calling library functions. Decision: const variables, data members, methods and arguments add a level of compile-time type checking; it is better to detect errors as soon as possible. Therefore we strongly recommend that you use const whenever it makes sense to do so: If a function does not modify an argument passed by reference or by pointer, that argument should be const. Declare methods to be const whenever possible. Accessors should almost always be const. Other methods should be const if they do not modify any data members, do not call any non-const methods, and do not return a non-const pointer or non-const reference to a data member. Consider making data members const whenever they do not need to be modified after construction. However, do not go crazy with const. Something like const int * const * const x; is likely overkill, even if it accurately describes how const x is. Focus on what's really useful to know: in this case, const int** x is probably sufficient. The mutable keyword is allowed but is unsafe when used with threads, so thread safety should be carefully considered first. Where to put the const Some people favor the form int const *foo to const int* foo. They argue that this is more readable because it's more consistent: it keeps the rule that const always follows the object it's describing. However, this consistency argument doesn't apply in this case, because the "don't go crazy" dictum eliminates most of the uses you'd have to be consistent with. Putting the const first is arguably more readable, since it follows English in putting the "adjective" (const) before the "noun" (int). That said, while we encourage putting const first, we do not require it. But be consistent with the code around you! Integer Types link ▶Of the built-in C++ integer types, the only one used is int. If a program needs a variable of a different size, use a precise-width integer type from , such as int16_t. Definition: C++ does not specify the sizes of its integer types. Typically people assume that short is 16 bits, int is 32 bits, long is 32 bits and long long is 64 bits. Pros: Uniformity of declaration. Cons: The sizes of integral types in C++ can vary based on compiler and architecture. Decision: defines types like int16_t, uint32_t, int64_t, etc. You should always use those in preference to short, unsigned long long and the like, when you need a guarantee on the size of an integer. Of the C integer types, only int should be used. When appropriate, you are welcome to use standard types like size_t and ptrdiff_t. We use int very often, for integers we know are not going to be too big, e.g., loop counters. Use plain old int for such things. You should assume that an int is at least 32 bits, but don't assume that it has more than 32 bits. If you need a 64-bit integer type, use int64_t or uint64_t. For integers we know can be "big", use int64_t. You should not use the unsigned integer types such as uint32_t, unless the quantity you are representing is really a bit pattern rather than a number, or unless you need defined twos-complement overflow. In particular, do not use unsigned types to say a number will never be negative. Instead, use assertions for this. On Unsigned Integers Some people, including some textbook authors, recommend using unsigned types to represent numbers that are never negative. This is intended as a form of self-documentation. However, in C, the advantages of such documentation are outweighed by the real bugs it can introduce. Consider: for (unsigned int i = foo.Length()-1; i >= 0; --i) ... This code will never terminate! Sometimes gcc will notice this bug and warn you, but often it will not. Equally bad bugs can occur when comparing signed and unsigned variables. Basically, C's type-promotion scheme causes unsigned types to behave differently than one might expect. So, document that a variable is non-negative using assertions. Don't use an unsigned type. 64-bit Portability link ▶Code should be 64-bit and 32-bit friendly. Bear in mind problems of printing, comparisons, and structure alignment. printf() specifiers for some types are not cleanly portable between 32-bit and 64-bit systems. C99 defines some portable format specifiers. Unfortunately, MSVC 7.1 does not understand some of these specifiers and the standard is missing a few, so we have to define our own ugly versions in some cases (in the style of the standard include file inttypes.h): // printf macros for size_t, in the style of inttypes.h #ifdef _LP64 #define __PRIS_PREFIX "z" #else #define __PRIS_PREFIX #endif // Use these macros after a % in a printf format string // to get correct 32/64 bit behavior, like this: // size_t size = records.size(); // printf("%"PRIuS"\n", size); #define PRIdS __PRIS_PREFIX "d" #define PRIxS __PRIS_PREFIX "x" #define PRIuS __PRIS_PREFIX "u" #define PRIXS __PRIS_PREFIX "X" #define PRIoS __PRIS_PREFIX "o" Type DO NOT use DO use Notes void * (or any pointer) %lx %p int64_t %qd, %lld %"PRId64" uint64_t %qu, %llu, %llx %"PRIu64", %"PRIx64" size_t %u %"PRIuS", %"PRIxS" C99 specifies %zu ptrdiff_t %d %"PRIdS" C99 specifies %zd Note that the PRI* macros expand to independent strings which are concatenated by the compiler. Hence if you are using a non-constant format string, you need to insert the value of the macro into the format, rather than the name. It is still possible, as usual, to include length specifiers, etc., after the % when using the PRI* macros. So, e.g. printf("x = %30"PRIuS"\n", x) would expand on 32-bit Linux to printf("x = %30" "u" "\n", x), which the compiler will treat as printf("x = %30u\n", x). Remember that sizeof(void *) != sizeof(int). Use intptr_t if you want a pointer-sized integer. You may need to be careful with structure alignments, particularly for structures being stored on disk. Any class/structure with a int64_t/uint64_t member will by default end up being 8-byte aligned on a 64-bit system. If you have such structures being shared on disk between 32-bit and 64-bit code, you will need to ensure that they are packed the same on both architectures. Most compilers offer a way to alter structure alignment. For gcc, you can use __attribute__((packed)). MSVC offers #pragma pack() and __declspec(align()). Use the LL or ULL suffixes a

Contents Overview 1 Lesson 1: Concepts – Locks and Lock Manager 3 Lesson 2: Concepts – Batch and Transaction 31 Lesson 3: Concepts – Locks and Applications 51 Lesson 4: Information Collection and Analysis 63 Lesson 5: Concepts – Formulating and Implementing Resolution 81 Module 4: Troubleshooting Locking and Blocking Overview At the end of this module, you will be able to:  Discuss how lock manager uses lock mode, lock resources, and lock compatibility to achieve transaction isolation.  Describe the various transaction types and how transactions differ from batches.  Describe how to troubleshoot blocking and locking issues.  Analyze the output of blocking scripts and Microsoft® SQL Server™ Profiler to troubleshoot locking and blocking issues.  Formulate hypothesis to resolve locking and blocking issues. Lesson 1: Concepts – Locks and Lock Manager This lesson outlines some of the common causes that contribute to the perception of a slow server. What You Will Learn After completing this lesson, you will be able to:  Describe locking architecture used by SQL Server.  Identify the various lock modes used by SQL Server.  Discuss lock compatibility and concurrent access.  Identify different types of lock resources.  Discuss dynamic locking and lock escalation.  Differentiate locks, latches, and other SQL Server internal “locking” mechanism such as spinlocks and other synchronization objects. Recommended Reading  Chapter 14 “Locking”, Inside SQL Server 2000 by Kalen Delaney  SOX000821700049 – SQL 7.0 How to interpret lock resource Ids  SOX000925700237 – TITLE: Lock escalation in SQL 7.0  SOX001109700040 – INF: Queries with PREFETCH in the plan hold lock until the end of transaction Locking Concepts Delivery Tip Prior to delivering this material, test the class to see if they fully understand the different isolation levels. If the class is not confident in their understanding, review appendix A04_Locking and its accompanying PowerPoint® file. Transactions in SQL Server provide the ACID properties: Atomicity A transaction either commits or aborts. If a transaction commits, all of its effects remain. If it aborts, all of its effects are undone. It is an “all or nothing” operation. Consistency An application should maintain the consistency of a database. For example, if you defer constraint checking, it is your responsibility to ensure that the database is consistent. Isolation Concurrent transactions are isolated from the updates of other incomplete transactions. These updates do not constitute a consistent state. This property is often called serializability. For example, a second transaction traversing the doubly linked list mentioned above would see the list before or after the insert, but it will see only complete changes. Durability After a transaction commits, its effects will persist even if there are system failures. Consistency and isolation are the most important in describing SQL Server’s locking model. It is up to the application to define what consistency means, and isolation in some form is needed to achieve consistent results. SQL Server uses locking to achieve isolation. Definition of Dependency: A set of transactions can run concurrently if their outputs are disjoint from the union of one another’s input and output sets. For example, if T1 writes some object that is in T2’s input or output set, there is a dependency between T1 and T2. Bad Dependencies These include lost updates, dirty reads, non-repeatable reads, and phantoms. ANSI SQL Isolation Levels An isolation level determines the degree to which data is isolated for use by one process and guarded against interference from other processes. Prior to SQL Server 7.0, REPEATABLE READ and SERIALIZABLE isolation levels were synonymous. There was no way to prevent non-repeatable reads while not preventing phantoms. By default, SQL Server 2000 operates at an isolation level of READ COMMITTED. To make use of either more or less strict isolation levels in applications, locking can be customized for an entire session by setting the isolation level of the session with the SET TRANSACTION ISOLATION LEVEL statement. To determine the transaction isolation level currently set, use the DBCC USEROPTIONS statement, for example: USE pubs GO SET TRANSACTION ISOLATION LEVEL REPEATABLE READ GO DBCC USEROPTIONS GO Multigranular Locking Multigranular Locking In our example, if one transaction (T1) holds an exclusive lock at the table level, and another transaction (T2) holds an exclusive lock at the row level, each of the transactions believe they have exclusive access to the resource. In this scenario, since T1 believes it locks the entire table, it might inadvertently make changes to the same row that T2 thought it has locked exclusively. In a multigranular locking environment, there must be a way to effectively overcome this scenario. Intent lock is the answer to this problem. Intent Lock Intent Lock is the term used to mean placing a marker in a higher-level lock queue. The type of intent lock can also be called the multigranular lock mode. An intent lock indicates that SQL Server wants to acquire a shared (S) lock or exclusive (X) lock on some of the resources lower down in the hierarchy. For example, a shared intent lock placed at the table level means that a transaction intends on placing shared (S) locks on pages or rows within that table. Setting an intent lock at the table level prevents another transaction from subsequently acquiring an exclusive (X) lock on the table containing that page. Intent locks improve performance because SQL Server examines intent locks only at the table level to determine whether a transaction can safely acquire a lock on that table. This removes the requirement to examine every row or page lock on the table to determine whether a transaction can lock the entire table. Lock Mode The code shown in the slide represents how the lock mode is stored internally. You can see these codes by querying the master.dbo.spt_values table: SELECT * FROM master.dbo.spt_values WHERE type = N'L' However, the req_mode column of master.dbo.syslockinfo has lock mode code that is one less than the code values shown here. For example, value of req_mode = 3 represents the Shared lock mode rather than the Schema Modification lock mode. Lock Compatibility These locks can apply at any coarser level of granularity. If a row is locked, SQL Server will apply intent locks at both the page and the table level. If a page is locked, SQL Server will apply an intent lock at the table level. SIX locks imply that we have shared access to a resource and we have also placed X locks at a lower level in the hierarchy. SQL Server never asks for SIX locks directly, they are always the result of a conversion. For example, suppose a transaction scanned a page using an S lock and then subsequently decided to perform a row level update. The row would obtain an X lock, but now the page would require an IX lock. The resultant mode on the page would be SIX. Another type of table lock is a schema stability lock (Sch-S) and is compatible with all table locks except the schema modification lock (Sch-M). The schema modification lock (Sch-M) is incompatible with all table locks. Locking Resources Delivery Tip Note the differences between Key and Key Range locks. Key Range locks will be covered in a couple of slides. SQL Server can lock these resources: Item Description DB A database. File A database file Index An entire index of a table. Table An entire table, including all data and indexes. Extent A contiguous group of data pages or index pages. Page An 8-KB data page or index page. Key Row lock within an index. Key-range A key-range. Used to lock ranges between records in a table to prevent phantom insertions or deletions into a set of records. Ensures serializable transactions. RID A Row Identifier. Used to individually lock a single row within a table. Application A lock resource defined by an application. The lock manager knows nothing about the resource format. It simply compares the 'strings' representing the lock resources to determine whether it has found a match. If a match is found, it knows that resource is already locked. Some of the resources have “sub-resources.” The followings are sub-resources displayed by the sp_lock output: Database Lock Sub-Resources: Full Database Lock (default) [BULK-OP-DB] – Bulk Operation Lock for Database [BULK-OP-LOG] – Bulk Operation Lock for Log Table Lock Sub-Resources: Full Table Lock (default) [UPD-STATS] – Update statistics Lock [COMPILE] – Compile Lock Index Lock sub-Resources: Full Index Lock (default) [INDEX_ID] – Index ID Lock [INDEX_NAME] – Index Name Lock [BULK_ALLOC] – Bulk Allocation Lock [DEFRAG] – Defragmentation Lock For more information, see also… SOX000821700049 SQL 7.0 How to interpret lock resource Ids Lock Resource Block The resource type has the following resource block format: Resource Type (Code) Content DB (2) Data 1: sub-resource; Data 2: 0; Data 3: 0 File (3) Data 1: File ID; Data 2: 0; Data 3: 0 Index (4) Data 1: Object ID; Data 2: sub-resource; Data 3: Index ID Table (5) Data 1: Object ID; Data 2: sub-resource; Data 3: 0. Page (6) Data 1: Page Number; Data 3: 0. Key (7) Data 1: Object ID; Data 2: Index ID; Data 3: Hashed Key Extent (8) Data 1: Extent ID; Data 3: 0. RID (9) Data 1: RID; Data 3: 0. Application (10) Data 1: Application resource name The rsc_bin column of master..syslockinfo contains the resource block in hexadecimal format. For an example of how to decode value from this column using the information above, let us assume we have the following value: 0x000705001F83D775010002014F0BEC4E With byte swapping within each field, this can be decoded as: Byte 0: Flag – 0x00 Byte 1: Resource Type – 0x07 (Key) Byte 2-3: DBID – 0x0005 Byte 4-7: ObjectID – 0x 75D7831F (1977058079) Byte 8-9: IndexID – 0x0001 Byte 10-16: Hash Key value – 0x 02014F0BEC4E For more information about how to decode this value, see also… Inside SQL Server 2000, pages 803 and 806. Key Range Locking Key Range Locking To support SERIALIZABLE transaction semantics, SQL Server needs to lock sets of rows specified by a predicate, such as WHERE salary BETWEEN 30000 AND 50000 SQL Server needs to lock data that does not exist! If no rows satisfy the WHERE condition the first time the range is scanned, no rows should be returned on any subsequent scans. Key range locks are similar to row locks on index keys (whether clustered or not). The locks are placed on individual keys rather than at the node level. The hash value consists of all the key components and the locator. So, for a nonclustered index over a heap, where columns c1 and c2 where indexed, the hash would contain contributions from c1, c2 and the RID. A key range lock applied to a particular key means that all keys between the value locked and the next value would be locked for all data modification. Key range locks can lock a slightly larger range than that implied by the WHERE clause. Suppose the following select was executed in a transaction with isolation level SERIALIZABLE: SELECT * FROM members WHERE first_name between ‘Al’ and ‘Carl’ If 'Al', 'Bob', and 'Dave' are index keys in the table, the first two of these would acquire key range locks. Although this would prevent anyone from inserting either 'Alex' or 'Ben', it would also prevent someone from inserting 'Dan', which is not within the range of the WHERE clause. Prior to SQL Server 7.0, page locking was used to prevent phantoms by locking the entire set of pages on which the phantom would exist. This can be too conservative. Key Range locking lets SQL Server lock only a much more restrictive area of the table. Impact Key-range locking ensures that these scenarios are SERIALIZABLE:  Range scan query  Singleton fetch of nonexistent row  Delete operation  Insert operation However, the following conditions must be satisfied before key-range locking can occur:  The transaction-isolation level must be set to SERIALIZABLE.  The operation performed on the data must use an index range access. Range locking is activated only when query processing (such as the optimizer) chooses an index path to access the data. Key Range Lock Mode Again, the req_mode column of master.dbo.syslockinfo has lock mode code that is one less than the code values shown here. Dynamic Locking When modifying individual rows, SQL Server typically would take row locks to maximize concurrency (for example, OLTP, order-entry application). When scanning larger volumes of data, it would be more appropriate to take page or table locks to minimize the cost of acquiring locks (for example, DSS, data warehouse, reporting). Locking Decision The decision about which unit to lock is made dynamically, taking many factors into account, including other activity on the system. For example, if there are multiple transactions currently accessing a table, SQL Server will tend to favor row locking more so than it otherwise would. It may mean the difference between scanning the table now and paying a bit more in locking cost, or having to wait to acquire a more coarse lock. A preliminary locking decision is made during query optimization, but that decision can be adjusted when the query is actually executed. Lock Escalation When the lock count for the transaction exceeds and is a multiple of ESCALATION_THRESHOLD (1250), the Lock Manager attempts to escalate. For example, when a transaction acquired 1250 locks, lock manager will try to escalate. The number of locks held may continue to increase after the escalation attempt (for example, because new tables are accessed, or the previous lock escalation attempts failed due to incompatible locks held by another spid). If the lock count for this transaction reaches 2500 (1250 * 2), Lock Manager will attempt escalation again. The Lock Manager looks at the lock memory it is using and if it is more than 40 percent of SQL Server’s allocated buffer pool memory, it tries to find a scan (SDES) where no escalation has already been performed. It then repeats the search operation until all scans have been escalated or until the memory used drops under the MEMORY_LOAD_ESCALATION_THRESHOLD (40%) value. If lock escalation is not possible or fails to significantly reduce lock memory footprint, SQL Server can continue to acquire locks until the total lock memory reaches 60 percent of the buffer pool (MAX_LOCK_RESOURCE_MEMORY_PERCENTAGE=60). Lock escalation may be also done when a single scan (SDES) holds more than LOCK_ESCALATION_THRESHOLD (765) locks. There is no lock escalation on temporary tables or system tables. Trace Flag 1211 disables lock escalation. Important Do not relay this to the customer without careful consideration. Lock escalation is a necessary feature, not something to be avoided completely. Trace flags are global and disabling lock escalation could lead to out of memory situations, extremely poor performing queries, or other problems. Lock escalation tracing can be seen using the Profiler or with the general locking trace flag, -T1200. However, Trace Flag 1200 shows all lock activity so it should not be usable on a production system. For more information, see also… SOX000925700237 “TITLE: SQL 7.0 Lock escalation in SQL 7.0” Lock Timeout Application Lock Timeout An application can set lock timeout for a session with the SET option: SET LOCK_TIMEOUT N where N is a number of milliseconds. A value of -1 means that there will be no timeout, which is equivalent to the version 6.5 behavior. A value of 0 means that there will be no waiting; if a process finds a resource locked, it will generate error message 1222 and continue with the next statement. The current value of LOCK_TIMEOUT is stored in the global variable @@lock_timeout. Note After a lock timeout any transaction containing the statement, is rolled back or canceled by SQL Server 2000 (bug#352640 was filed). This behavior is different from that of SQL Server 7.0. With SQL Server 7.0, the application must have an error handler that can trap error 1222 and if an application does not trap the error, it can proceed unaware that an individual statement within a transaction has been canceled, and errors can occur because statements later in the transaction may depend on the statement that was never executed. Bug#352640 is fixed in hotfix build 8.00.266 whereby a lock timeout will only Internal Lock Timeout At time, internal operations within SQL Server will attempt to acquire locks via lock manager. Typically, these lock requests are issued with “no waiting.” For example, the ghost record processing might try to clean up rows on a particular page, and before it can do that, it needs to lock the page. Thus, the ghost record manager will request a page lock with no wait so that if it cannot lock the page, it will just move on to other pages; it can always come back to this page later. If you look at SQL Profiler Lock: Timeout events, internal lock timeout typically have a duration value of zero. Lock Duration Lock Mode and Transaction Isolation Level For REPEATABLE READ transaction isolation level, update locks are held until data is read and processed, unless promoted to exclusive locks. "Data is processed" means that we have decided whether the row in question matched the search criteria; if not then the update lock is released, otherwise, we get an exclusive lock and make the modification. Consider the following query: use northwind go dbcc traceon(3604, 1200, 1211) -- turn on lock tracing -- and disable escalation go set transaction isolation level repeatable read begin tran update dbo.[order details] set discount = convert (real, discount) where discount = 0.0 exec sp_lock Update locks are promoted to exclusive locks when there is a match; otherwise, the update lock is released. The sp_lock output verifies that the SPID does not hold any update locks or shared locks at the end of the query. Lock escalation is turned off so that exclusive table lock is not held at the end. Warning Do not use trace flag 1200 in a production environment because it produces a lot of output and slows down the server. Trace flag 1211 should not be used unless you have done extensive study to make sure it helps with performance. These trace flags are used here for illustration and learning purposes only. Lock Ownership Most of the locking discussion in this lesson relates to locks owned by “transactions.” In addition to transaction, cursor and session can be owners of locks and they both affect how long locks are held. For every row that is fetched, when SCROLL_LOCKS option is used, regardless of the state of a transaction, a cursor lock is held until the next row is fetched or when the cursor is closed. Locks owned by session are outside the scope of a transaction. The duration of these locks are bounded by the connection and the process will continue to hold these locks until the process disconnects. A typical lock owned by session is the database (DB) lock. Locking – Read Committed Scan Under read committed isolation level, when database pages are scanned, shared locks are held when the page is read and processed. The shared locks are released “behind” the scan and allow other transactions to update rows. It is important to note that the shared lock currently acquired will not be released until shared lock for the next page is successfully acquired (this is commonly know as “crabbing”). If the same pages are scanned again, rows may be modified or deleted by other transactions. Locking – Repeatable Read Scan Under repeatable read isolation level, when database pages are scanned, shared locks are held when the page is read and processed. SQL Server continues to hold these shared locks, thus preventing other transactions to update rows. If the same pages are scanned again, previously scanned rows will not change but new rows may be added by other transactions. Locking – Serializable Read Scan Under serializable read isolation level, when database pages are scanned, shared locks are held not only on rows but also on scanned key range. SQL Server continues to hold these shared locks until the end of transaction. Because key range locks are held, not only will this prevent other transactions from modifying the rows, no new rows can be inserted. Prefetch and Isolation Level Prefetch and Locking Behavior The prefetch feature is available for use with SQL Server 7.0 and SQL Server 2000. When searching for data using a nonclustered index, the index is searched for a particular value. When that value is found, the index points to the disk address. The traditional approach would be to immediately issue an I/O for that row, given the disk address. The result is one synchronous I/O per row and, at most, one disk at a time working to evaluate the query. This does not take advantage of striped disk sets. The prefetch feature takes a different approach. It continues looking for more record pointers in the nonclustered index. When it has collected a number of them, it provides the storage engine with prefetch hints. These hints tell the storage engine that the query processor will need these particular records soon. The storage engine can now issue several I/Os simultaneously, taking advantage of striped disk sets to execute multiple operations simultaneously. For example, if the engine is scanning a nonclustered index to determine which rows qualify but will eventually need to visit the data page as well to access columns that are not in the index, it may decide to submit asynchronous page read requests for a group of qualifying rows. The prefetch data pages are then revisited later to avoid waiting for each individual page read to complete in a serial fashion. This data access path requires that a lock be held between the prefetch request and the row lookup to stabilize the row on the page so it is not to be moved by a page split or clustered key update. For our example, the isolation level of the query is escalated to REPEATABLE READ, overriding the transaction isolation level. With SQL Server 7.0 and SQL Server 2000, portions of a transaction can execute at a different transaction isolation level than the entire transaction itself. This is implemented as lock classes. Lock classes are used to control lock lifetime when portions of a transaction need to execute at a stricter isolation level than the underlying transaction. Unfortunately, in SQL Server 7.0 and SQL Server 2000, the lock class is created at the topmost operator of the query and hence released only at the end of the query. Currently there is no support to release the lock (lock class) after the row has been discarded or fetched by the filter or join operator. This is because isolation level can be set at the query level via a lock class, but no lower. Because of this, locks acquired during the query will not be released until the query completes. If prefetch is occurring you may see a single SPID that holds hundreds of Shared KEY or PAG locks even though the connection’s isolation level is READ COMMITTED. Isolation level can be determined from DBCC PSS output. For details about this behavior see “SOX001109700040 INF: Queries with PREFETCH in the plan hold lock until the end of transaction”. Other Locking Mechanism Lock manager does not manage latches and spinlocks. Latches Latches are internal mechanisms used to protect pages while doing operations such as placing a row physically on a page, compressing space on a page, or retrieving rows from a page. Latches can roughly be divided into I/O latches and non-I/O latches. If you see a high number of non-I/O related latches, SQL Server is usually doing a large number of hash or sort operations in tempdb. You can monitor latch activities via DBCC SQLPERF(‘WAITSTATS’) command. Spinlock A spinlock is an internal data structure that is used to protect vital information that is shared within SQL Server. On a multi-processor machine, when SQL Server tries to access a particular resource protected by a spinlock, it must first acquire the spinlock. If it fails, it executes a loop that will check to see if the lock is available and if not, decrements a counter. If the counter reaches zero, it yields the processor to another thread and goes into a “sleep” (wait) state for a pre-determined amount of time. When it wakes, hopefully, the lock is free and available. If not, the loop starts again and it is terminated only when the lock is acquired. The reason for implementing a spinlock is that it is probably less costly to “spin” for a short time rather than yielding the processor. Yielding the processor will force an expensive context switch where:  The old thread’s state must be saved  The new thread’s state must be reloaded  The data stored in the L1 and L2 cache are useless to the processor On a single-processor computer, the loop is not useful because no other thread can be running and thus, no one can release the spinlock for the currently executing thread to acquire. In this situation, the thread yields the processor immediately. Lesson 2: Concepts – Batch and Transaction This lesson outlines some of the common causes that contribute to the perception of a slow server. What You Will Learn After completing this lesson, you will be able to:  Review batch processing and error checking.  Review explicit, implicit and autocommit transactions and transaction nesting level.  Discuss how commit and rollback transaction done in stored procedure and trigger affects transaction nesting level.  Discuss various transaction isolation level and their impact on locking.  Discuss the difference between aborting a statement, a transaction, and a batch.  Describe how @@error, @@transcount, and @@rowcount can be used for error checking and handling. Recommended Reading  Charter 12 “Transactions and Triggers”, Inside SQL Server 2000 by Kalen Delaney Batch Definition SQL Profiler Statements and Batches To help further your understanding of what is a batch and what is a statement, you can use SQL Profiler to study the definition of batch and statement.  Try This: Using SQL Profiler to Analyze Batch 1. Log on to a server with Query Analyzer 2. Startup the SQL Profiler against the same server 3. Start a trace using the “StandardSQLProfiler” template 4. Execute the following using Query Analyzer: SELECT @@VERSION SELECT @@SPID The ‘SQL:BatchCompleted’ event is captured by the trace. It shows both the statements as a single batch. 5. Now execute the following using Query Analyzer {call sp_who()} What shows up? The ‘RPC:Completed’ with the sp_who information. RPC is simply another entry point to the SQL Server to call stored procedures with native data types. This allows one to avoid parsing. The ‘RPC:Completed’ event should be considered the same as a batch for the purposes of this discussion. Stop the current trace and start a new trace using the “SQLProfilerTSQL_SPs” template. Issue the same command as outlines in step 5 above. Looking at the output, not only can you see the batch markers but each statement as executed within the batch. Autocommit, Explicit, and Implicit Transaction Autocommit Transaction Mode (Default) Autocommit mode is the default transaction management mode of SQL Server. Every Transact-SQL statement, whether it is a standalone statement or part of a batch, is committed or rolled back when it completes. If a statement completes successfully, it is committed; if it encounters any error, it is rolled back. A SQL Server connection operates in autocommit mode whenever this default mode has not been overridden by either explicit or implicit transactions. Autocommit mode is also the default mode for ADO, OLE DB, ODBC, and DB-Library. A SQL Server connection operates in autocommit mode until a BEGIN TRANSACTION statement starts an explicit transaction, or implicit transaction mode is set on. When the explicit transaction is committed or rolled back, or when implicit transaction mode is turned off, SQL Server returns to autocommit mode. Explicit Transaction Mode An explicit transaction is a transaction that starts with a BEGIN TRANSACTION statement. An explicit transaction can contain one or more statements and must be terminated by either a COMMIT TRANSACTION or a ROLLBACK TRANSACTION statement. Implicit Transaction Mode SQL Server can automatically or, more precisely, implicitly start a transaction for you if a SET IMPLICIT_TRANSACTIONS ON statement is run or if the implicit transaction option is turned on globally by running sp_configure ‘user options’ 2. (Actually, the bit mask 0x2 must be turned on for the user option so you might have to perform an ‘OR’ operation with the existing user option value.) See SQL Server 2000 Books Online on how to turn on implicit transaction under ODBC and OLE DB (acdata.chm::/ac_8_md_06_2g6r.htm). Transaction Nesting Explicit transactions can be nested. Committing inner transactions is ignored by SQL Server other than to decrements @@TRANCOUNT. The transaction is either committed or rolled back based on the action taken at the end of the outermost transaction. If the outer transaction is committed, the inner nested transactions are also committed. If the outer transaction is rolled back, then all inner transactions are also rolled back, regardless of whether the inner transactions were individually committed. Each call to COMMIT TRANSACTION applies to the last executed BEGIN TRANSACTION. If the BEGIN TRANSACTION statements are nested, then a COMMIT statement applies only to the last nested transaction, which is the innermost transaction. Even if a COMMIT TRANSACTION transaction_name statement within a nested transaction refers to the transaction name of the outer transaction, the commit applies only to the innermost transaction. If a ROLLBACK TRANSACTION statement without a transaction_name parameter is executed at any level of a set of nested transaction, it rolls back all the nested transactions, including the outermost transaction. The @@TRANCOUNT function records the current transaction nesting level. Each BEGIN TRANSACTION statement increments @@TRANCOUNT by one. Each COMMIT TRANSACTION statement decrements @@TRANCOUNT by one. A ROLLBACK TRANSACTION statement that does not have a transaction name rolls back all nested transactions and decrements @@TRANCOUNT to 0. A ROLLBACK TRANSACTION that uses the transaction name of the outermost transaction in a set of nested transactions rolls back all the nested transactions and decrements @@TRANCOUNT to 0. When you are unsure if you are already in a transaction, SELECT @@TRANCOUNT to determine whether it is 1 or more. If @@TRANCOUNT is 0 you are not in a transaction. You can also find the transaction nesting level by checking the sysprocess.open_tran column. See SQL Server 2000 Books Online topic “Nesting Transactions” (acdata.chm::/ac_8_md_06_66nq.htm) for more information. Statement, Transaction, and Batch Abort One batch can have many statements and one transaction can have multiple statements, also. One transaction can span multiple batches and one batch can have multiple transactions. Statement Abort Currently executing statement is aborted. This can be a bit confusing when you start talking about statements in a trigger or stored procedure. Let us look closely at the following trigger: CREATE TRIGGER TRG8134 ON TBL8134 AFTER INSERT AS BEGIN SELECT 1/0 SELECT 'Next command in trigger' END To fire the INSERT trigger, the batch could be as simple as ‘INSERT INTO TBL8134 VALUES(1)’. However, the trigger contains two statements that must be executed as part of the batch to satisfy the clients insert request. When the ‘SELECT 1/0’ causes the divide by zero error, a statement abort is issued for the ‘SELECT 1/0’ statement. Batch and Transaction Abort On SQL Server 2000 (and SQL Server 7.0) whenever a non-informational error is encountered in a trigger, the statement abort is promoted to a batch and transactional abort. Thus, in the example the statement abort for ‘select 1/0’ promotion results in an entire batch abort. No further statements in the trigger or batch will be executed and a rollback is issued. On SQL Server 6.5, the statement aborts immediately and results in a transaction abort. However, the rest of the statements within the trigger are executed. This trigger could return ‘Next command in trigger’ as a result set. Once the trigger completes the batch abort promotion takes effect. Conversely, submitting a similar set of statements in a standalone batch can result in different behavior. SELECT 1/0 SELECT 'Next command in batch' Not considering the set option possibilities, a divide by zero error generally results in a statement abort. Since it is not in a trigger, the promotion to a batch abort is avoided and subsequent SELECT statement can execute. The programmer should add an “if @@ERROR” check immediately after the ‘select 1/0’ to T-SQL execution to control the flow correctly. Aborting and Set Options ARITHABORT If SET ARITHABORT is ON, these error conditions cause the query or batch to terminate. If the errors occur in a transaction, the transaction is rolled back. If SET ARITHABORT is OFF and one of these errors occurs, a warning message is displayed, and NULL is assigned to the result of the arithmetic operation. When an INSERT, DELETE, or UPDATE statement encounters an arithmetic error (overflow, divide-by-zero, or a domain error) during expression evaluation when SET ARITHABORT is OFF, SQL Server inserts or updates a NULL value. If the target column is not nullable, the insert or update action fails and the user receives an error. XACT_ABORT When SET XACT_ABORT is ON, if a Transact-SQL statement raises a run-time error, the entire transaction is terminated and rolled back. When OFF, only the Transact-SQL statement that raised the error is rolled back and the transaction continues processing. Compile errors, such as syntax errors, are not affected by SET XACT_ABORT. For example: CREATE TABLE t1 (a int PRIMARY KEY) CREATE TABLE t2 (a int REFERENCES t1(a)) GO INSERT INTO t1 VALUES (1) INSERT INTO t1 VALUES (3) INSERT INTO t1 VALUES (4) INSERT INTO t1 VALUES (6) GO SET XACT_ABORT OFF GO BEGIN TRAN INSERT INTO t2 VALUES (1) INSERT INTO t2 VALUES (2) /* Foreign key error */ INSERT INTO t2 VALUES (3) COMMIT TRAN SELECT 'Continue running batch 1...' GO SET XACT_ABORT ON GO BEGIN TRAN INSERT INTO t2 VALUES (4) INSERT INTO t2 VALUES (5) /* Foreign key error */ INSERT INTO t2 VALUES (6) COMMIT TRAN SELECT 'Continue running batch 2...' GO /* Select shows only keys 1 and 3 added. Key 2 insert failed and was rolled back, but XACT_ABORT was OFF and rest of transaction succeeded. Key 5 insert error with XACT_ABORT ON caused all of the second transaction to roll back. Also note that 'Continue running batch 2...' is not Returned to indicate that the batch is aborted. */ SELECT * FROM t2 GO DROP TABLE t2 DROP TABLE t1 GO Compile and Run-time Errors Compile Errors Compile errors are encountered during syntax checks, security checks, and other general operations to prepare the batch for execution. These errors can prevent the optimization of the query and thus lead to immediate abort. The statement is not run and the batch is aborted. The transaction state is generally left untouched. For example, assume there are four statements in a particular batch. If the third statement has a syntax error, none of the statements in the batch is executed. Optimization Errors Optimization errors would include rare situations where the statement encounters a problem when attempting to build an optimal execution plan. Example: “too many tables referenced in the query” error is reported because a “work table” was added to the plan. Runtime Errors Runtime errors are those that are encountered during the execution of the query. Consider the following batch: SELECT * FROM pubs.dbo.titles UPDATE pubs.dbo.authors SET au_lname = au_lname SELECT * FROM foo UPDATE pubs.dbo.authors SET au_lname = au_lname If you run the above statements in a batch, the first two statements will be executed, the third statement will fail because table foo does not exist, and the batch will terminate. Deferred Name Resolution is the feature that allows this batch to start executing before resolving the object foo. This feature allows SQL Server to delay object resolution and place a “placeholder” in the query’s execution. The object referenced by the placeholder is resolved until the query is executed. In our example, the execution of the statement “SELECT * FROM foo” will trigger another compile process to resolve the name again. This time, error message 208 is returned. Error: 208, Level 16, State 1, Line 1 Invalid object name 'foo'. Message 208 can be encountered as a runtime or compile error depending on whether the Deferred Name Resolution feature is available. In SQL Server 6.5 this would be considered a compile error and on SQL Server 2000 (and SQL Server7.0) as a runtime error due to Deferred Name Resolution. In the following example, if a trigger referenced authors2, the error is detected as SQL Server attempts to execute the trigger. However, under SQL Server 6.5 the create trigger statement fails because authors2 does not exist at compile time. When errors are encountered in a trigger, generally, the statement, batch, and transaction are aborted. You should be able to observe this by running the following script in pubs database: Create table tblTest(iID int) go create trigger trgInsert on tblTest for INSERT as begin select * from authors select * from authors2 select * from titles end go begin tran select 'Before' insert into tblTest values(1) select 'After' go select @@TRANCOUNT go When run in a batch, the statement and the batch are aborted but the transaction remains active. The follow script illustrates this: begin tran select 'Before' select * from authors2 select 'After' go select @@TRANCOUNT go One other factor in a compile versus runtime error is implicit data type conversions. If you were to run the following statements on SQL Server 6.5 and SQL Server 2000 (and SQL Server 7.0): create table tblData(dtData datetime) go select 1 insert into tblData values(12/13/99) go On SQL Server 6.5, you get an error before execution of the batch begins so no statements are executed and the batch is aborted. Error: 206, Level 16, State 2, Line 2 Operand type clash: int is incompatible with datetime On SQL Server 2000, you get the default value (1900-01-01 00:00:00.000) inserted into the table. SQL Server 2000 implicit data type conversion treats this as integer division. The integer division of 12/13/99 is 0, so the default date and time value is inserted, no error returned. To correct the problem on either version is to wrap the date string with quotes. See Bug #56118 (sqlbug_70) for more details about this situation. Another example of a runtime error is a 605 message. Error: 605 Attempt to fetch logical page %S_PGID in database '%.*ls' belongs to object '%.*ls', not to object '%.*ls'. A 605 error is always a runtime error. However, depending on the transaction isolation level, (e.g. using the NOLOCK lock hint), established by the SPID the handling of the error can vary. Specifically, a 605 error is considered an ACCESS error. Errors associated with buffer and page access are found in the 600 series of errors. When the error is encountered, the isolation level of the SPID is examined to determine proper handling based on information or fatal error level. Transaction Error Checking Not all errors cause transactions to automatically rollback. Although it is difficult to determine exactly which errors will rollback transactions and which errors will not, the main idea here is that programmers must perform error checking and handle errors appropriately. Error Handling Raiserror Details Raiserror seems to be a source of confusion but is really rather simple. Raiserror with severity levels of 20 or higher will terminate the connection. Of course, when the connection is terminated a full rollback of any open transaction will immediately be instantiated by the SQL Server (except distributed transaction with DTC involved). Severity levels lower than 20 will simply result in the error message being returned to the client. They do not affect the transaction scope of the connection. Consider the following batch: use pubs begin tran update authors set au_lname = 'smith' raiserror ('This is bad', 19, 1) with log select @@trancount With severity set at 19, the 'select @@trancount' will be executed after the raiserror statement and will return a value of 1. If severity is changed to 20, then the select statement will not run and the connection is broken. Important Error handling must occur not only in T-SQL batches and stored procedures, but also in application program code. Transactions and Triggers (1 of 2) Basic behavior assumes the implicit transactions setting is set to OFF. This behavior makes it possible to identify business logic errors in a trigger, raise an error, rollback the action, and add an audit table entry. Logically, the insert to the audit table cannot take place before the ROLLBACK action and you would not want to build in the audit table insert into every applications error handler that violated the business rule of the trigger. For more information, see also… SQL Server 2000 Books Online topic “Rollbacks in stored procedure and triggers“ (acdata.chm::/ac_8_md_06_4qcz.htm) IMPLICIT_TRANSACTIONS ON Behavior The behavior of firing other triggers on the same table can be tricky. Say you added a trigger that checks the CODE field. Read only versions of the rows contain the code ‘RO’ and read/write versions use ‘RW.’ Whenever someone tries to delete a row with a code ‘RO’ the trigger issues the rollback and logs an audit table entry. However, you also have a second trigger that is responsible for cascading delete operations. One client could issue the delete without implicit transactions on and only the current trigger would execute and then terminate the batch. However, a second client with implicit transactions on could issue the same delete and the secondary trigger would fire. You end up with a situation in which the cascading delete operations can take place (are committed) but the initial row remains in the table because of the rollback operation. None of the delete operations should be allowed but because the transaction scope was restarted because of the implicit transactions setting, they did. Transactions and Triggers (2 of 2) It is extremely difficult to determine the execution state of a trigger when using explicit rollback statements in combination with implicit transactions. The RETURN statement is not allowed to return a value. The only way I have found to set the @@ERROR is using a ‘raiserror’ as the last execution statement in the last trigger to execute. If you modify the example, this following RAISERROR statement will set @@ERROR to 50000: CREATE TRIGGER trgTest on tblTest for INSERT AS BEGIN ROLLBACK INSERT INTO tblAudit VALUES (1) RAISERROR('This is bad', 14,1) END However, this value does not carry over to a secondary trigger for the same table. If you raise an error at the end of the first trigger and then look at @@ERROR in the secondary trigger the @@ERROR remains 0. Carrying Forward an Active/Open Transaction It is possible to exit from a trigger and carry forward an open transaction by issuing a BEGIN TRAN or by setting implicit transaction on and doing INSERT, UPDATE, or DELETE. Warning It is never recommended that a trigger call BEGIN TRANSACTION. By doing this you increment the transaction count. Invalid code logic, not calling commit transaction, can lead to a situation where the transaction count remains elevated upon exit of the trigger. Transaction Count The behavior is better explained by understanding how the server works. It does not matter whether you are in a transaction, when a modification takes place the transaction count is incremented. So, in the simplest form, during the processing of an insert the transaction count is 1. On completion of the insert, the server will commit (and thus decrement the transaction count). If the commit identifies the transaction count has returned to 0, the actual commit processing is completed. Issuing a commit when the transaction count is greater than 1 simply decrements the nested transaction counter. Thus, when we enter a trigger, the transaction count is 1. At the completion of the trigger, the transaction count will be 0 due to the commit issued at the end of the modification statement (insert). In our example, if the connection was already in a transaction and called the second INSERT, since implicit transaction is ON, the transaction count in the trigger will be 2 as long as the ROLLBACK is not executed. At the end of the insert, the commit is again issued to decrement the transaction reference count to 1. However, the value does not return to 0 so the transaction remains open/active. Subsequent triggers are only fired if the transaction count at the end of the trigger remains greater than or equal to 1. The key to continuation of secondary triggers and the batch is the transaction count at the end of a trigger execution. If the trigger that performs a rollback has done an explicit begin transaction or uses implicit transactions, subsequent triggers and the batch will continue. If the transaction count is not 1 or greater, subsequent triggers and the batch will not execute. Warning Forcing the transaction count after issuing a rollback is dangerous because you can easily loose track of your transaction nesting level. When performing an explicit rollback in a trigger, you should immediately issue a return statement to maintain consistent behavior between a connection with and without implicit transaction settings. This will force the trigger(s) and batch to terminate immediately. One of the methods of dealing with this issue is to run ‘SET IMPLICIT_TRANSACTIONS OFF’ as the first statement of any trigger. Other methods may entails checking @@TRANCOUNT at the end of the trigger and continue to COMMIT the transaction as long as @@TRANCOUNT is greater than 1. Examples The following examples are based on this table: create table tbl50000Insert (iID int NOT NULL) go Note If more than one trigger is used, to guarantee the trigger firing sequence, the sp_settriggerorder command should be used. This command is omitted in these examples to simplify the complexity of the statements. First Example In the first example, the second trigger was never fired and the batch, starting with the insert statement, was aborted. Thus, the print statement was never issued. print('Trigger issues rollback - cancels batch') go create trigger trg50000Insert on tbl50000Insert for INSERT as begin select 'Inserted', * from inserted rollback tran select 'End of trigger', @@TRANCOUNT as 'TRANCOUNT' end go create trigger trg50000Insert2 on tbl50000Insert for INSERT as begin select 'In Trigger2' select 'Trigger 2 Inserted', * from inserted end go insert into tbl50000Insert values(1) print('---------------------- In same batch') select * from tbl50000Insert go -- Cleanup drop trigger trg50000Insert drop trigger trg50000Insert2 go delete from tbl50000Insert Second Example The next example shows that since a new transaction is started, the second trigger will be fired and the print statement in the batch will be executed. Note that the insert is rolled back. print('Trigger issues rollback - increases tran count to continue batch') go create trigger trg50000Insert on tbl50000Insert for INSERT as begin select 'Inserted', * from inserted rollback tran begin tran end go create trigger trg50000Insert2 on tbl50000Insert for INSERT as begin select 'In Trigger2' select 'Trigger 2 Inserted', * from inserted end go insert into tbl50000Insert values(2) print('---------------------- In same batch') select * from tbl50000Insert go -- Cleanup drop trigger trg50000Insert drop trigger trg50000Insert2 go delete from tbl50000Insert Third Example In the third example, the raiserror statement is used to set the @@ERROR value and the BEGIN TRAN statement is used in the trigger to allow the batch to continue to run. print('Trigger issues rollback - uses raiserror to set @@ERROR') go create trigger trg50000Insert on tbl50000Insert for INSERT as begin select 'Inserted', * from inserted rollback tran begin tran -- Increase @@trancount to allow -- batch to continue select @@trancount as ‘Trancount’ raiserror('This is from the trigger', 14,1) end go insert into tbl50000Insert values(3) select @@ERROR as 'ERROR', @@TRANCOUNT as 'Trancount' go -- Cleanup drop trigger trg50000Insert go delete from tbl50000Insert Fourth Example For the fourth example, a second trigger is added to illustrate the fact that @@ERROR value set in the first trigger will not be seen in the second trigger nor will it show up in the batch after the second trigger is fired. print('Trigger issues rollback - uses raiserror to set @@ERROR, not seen in second trigger and cleared in batch') go create trigger trg50000Insert on tbl50000Insert for INSERT as begin select 'Inserted', * from inserted rollback begin tran -- Increase @@trancount to -- allow batch to continue select @@TRANCOUNT as 'Trancount' raiserror('This is from the trigger', 14,1) end go create trigger trg50000Insert2 on tbl50000Insert for INSERT as begin select @@ERROR as 'ERROR', @@TRANCOUNT as 'Trancount' end go insert into tbl50000Insert values(4) select @@ERROR as 'ERROR', @@TRANCOUNT as 'Trancount' go -- Cleanup drop trigger trg50000Insert drop trigger trg50000Insert2 go delete from tbl50000Insert Lesson 3: Concepts – Locks and Applications This lesson outlines some of the common causes that contribute to the perception of a slow server. What You Will Learn After completing this lesson, you will be able to:  Explain how lock hints are used and their impact.  Discuss the effect on locking when an application uses Microsoft Transaction Server.  Identify the different kinds of deadlocks including distributed deadlock. Recommended Reading  Charter 14 “Locking”, Inside SQL Server 2000 by Kalen Delaney  Charter 16 “Query Tuning”, Inside SQL Server 2000 by Kalen Delaney Q239753 – Deadlock Situation Not Detected by SQL Server Q288752 – Blocked SPID Not Participating in Deadlock May Incorrectly be Chosen as victim Locking Hints UPDLOCK If update locks are used instead of shared locks while reading a table, the locks are held until the end of the statement or transaction. UPDLOCK has the advantage of allowing you to read data (without blocking other readers) and update it later with the assurance that the data has not changed since you last read it. READPAST READPAST is an optimizer hint for use with SELECT statements. When this hint is used, SQL Server will read past locked rows. For example, assume table T1 contains a single integer column with the values of 1, 2, 3, 4, and 5. If transaction A changes the value of 3 to 8 but has not yet committed, a SELECT * FROM T1 (READPAST) yields values 1, 2, 4, 5. Tip READPAST only applies to transactions operating at READ COMMITTED isolation and only reads past row-level locks. This lock hint can be used to implement a work queue on a SQL Server table. For example, assume there are many external work requests being thrown into a table and they should be serviced in approximate insertion order but they do not have to be completely FIFO. If you have 4 worker threads consuming work items from the queue they could each pick up a record using read past locking and then delete the entry from the queue and commit when they're done. If they fail, they could rollback, leaving the entry on the queue for the next worker thread to pick up. Caution The READPAST hint is not compatible with HOLDLOCK.  Try This: Using Locking Hints 1. Open a Query Window and connect to the pubs database. 2. Execute the following statements (--Conn 1 is optional to help you keep track of each connection): BEGIN TRANSACTION -- Conn 1 UPDATE titles SET price = price * 0.9 WHERE title_id = 'BU1032' 3. Open a second connection and execute the following statements: SELECT @@lock_timeout -- Conn 2 GO SELECT * FROM titles SELECT * FROM authors 4. Open a third connection and execute the following statements: SET LOCK_TIMEOUT 0 -- Conn 3 SELECT * FROM titles SELECT * FROM authors 5. Open a fourth connection and execute the following statement: SELECT * FROM titles (READPAST) -- Conn 4 WHERE title_ID < 'C' SELECT * FROM authors How many records were returned? 3 6. Open a fifth connection and execute the following statement: SELECT * FROM titles (NOLOCK) -- Conn 5 WHERE title_ID 0 the lock manager also checks for deadlocks every time a SPID gets blocked. So a single deadlock will trigger 20 seconds of more immediate deadlock detection, but if no additional deadlocks occur in that 20 seconds, the lock manager no longer checks for deadlocks at each block and detection again only happens every 5 seconds. Although normally not needed, you may use trace flag -T1205 to trace the deadlock detection process. Note Please note the distinction between application lock and other locks’ deadlock detection. For application lock, we do not rollback the transaction of the deadlock victim but simply return a -3 to sp_getapplock, which the application needs to handle itself. Deadlock Resolution How is a deadlock resolved? SQL Server picks one of the connections as a deadlock victim. The victim is chosen based on either which is the least expensive transaction (calculated using the number and size of the log records) to roll back or in which process “SET DEADLOCK_PRIORITY LOW” is specified. The victim’s transaction is rolled back, held locks are released, and SQL Server sends error 1205 to the victim’s client application to notify it that it was chosen as a victim. The other process can then obtain access to the resource it was waiting on and continue. Error 1205: Your transaction (process ID #%d) was deadlocked with another process and has been chosen as the deadlock victim. Rerun your transaction. Symptoms of deadlocking Error 1205 usually is not written to the SQL Server errorlog. Unfortunately, you cannot use sp_altermessage to cause 1205 to be written to the errorlog. If the client application does not capture and display error 1205, some of the symptoms of deadlock occurring are:  Clients complain of mysteriously canceled queries when using certain features of an application.  May be accompanied by excessive blocking. Lock contention increases the chances that a deadlock will occur. Triggers and Deadlock Triggers promote the deadlock priority of the SPID for the life of the trigger execution when the DEADLOCK PRIORITY is not set to low. When a statement in a trigger causes a deadlock to occur, the SPID executing the trigger is given preferential treatment and will not become the victim. Warning Bug 235794 is filed against SQL Server 2000 where a blocked SPID that is not a participant of a deadlock may incorrectly be chosen as a deadlock victim if the SPID is blocked by one of the deadlock participants and the SPID has the least amount of transaction logging. See KB article Q288752: “Blocked Spid Not Participating in Deadlock May Incorrectly be Chosen as victim” for more information. Distributed Deadlock – Scenario 1 Distributed Deadlocks The term distributed deadlock is ambiguous. There are many types of distributed deadlocks. Scenario 1 Client application opens connection A, begins a transaction, acquires some locks, opens connection B, connection B gets blocked by A but the application is designed to not commit A’s transaction until B completes. Note SQL Server has no way of knowing that connection A is somehow dependent on B – they are two distinct connections with two distinct transactions. This situation is discussed in scenario #4 in “Q224453 INF: Understanding and Resolving SQL Server 7.0 Blocking Problems”. Distributed Deadlock – Scenario 2 Scenario 2 Distributed deadlock involving bound connections. Two connections can be bound into a single transaction context with sp_getbindtoken/sp_bindsession or via DTC. Spid 60 enlists in a transaction with spid 61. A third spid 62 is blocked by spid 60, but spid 61 is blocked by spid 62. Because they are doing work in the same transaction, spid 60 cannot commit until spid 61 finishes his work, but spid 61 is blocked by 62 who is blocked by 60. This scenario is described in article “Q239753 - Deadlock Situation Not Detected by SQL Server.” Note SQL Server 6.5 and 7.0 do not detect this deadlock. The SQL Server 2000 deadlock detection algorithm has been enhanced to detect this type of distributed deadlock. The diagram in the slide illustrates this situation. Resources locked by a spid are below that spid (in a box). Arrows indicate blocking and are drawn from the blocked spid to the resource that the spid requires. A circle represents a transaction; spids in the same transaction are shown in the same circle. Distributed Deadlock – Scenario 3 Scenario 3 Distributed deadlock involving linked servers or server-to-server RPC. Spid 60 on Server 1 executes a stored procedure on Server 2 via linked server. This stored procedure does a loopback linked server query against a table on Server 1, and this connection is blocked by a lock held by Spid 60. Note No version of SQL Server is currently designed to detect this distributed deadlock. Lesson 4: Information Collection and Analysis This lesson outlines some of the common causes that contribute to the perception of a slow server. What You Will Learn After completing this lesson, you will be able to:  Identify specific information needed for troubleshooting issues.  Locate and collect information needed for troubleshooting issues.  Analyze output of DBCC Inputbuffer, DBCC PSS, and DBCC Page commands.  Review information collected from master.dbo.sysprocesses table.  Review information collected from master.dbo.syslockinfo table.  Review output of sp_who, sp_who2, sp_lock.  Analyze Profiler log for query usage pattern.  Review output of trace flags to help troubleshoot deadlocks. Recommended Reading Q244455 - INF: Definition of Sysprocesses Waittype and Lastwaittype Fields Q244456 - INF: Description of DBCC PSS Command for SQL Server 7.0 Q271509 - INF: How to Monitor SQL Server 2000 Blocking Q251004 - How to Monitor SQL Server 7.0 Blocking Q224453 - Understanding and Resolving SQL Server 7.0 Blocking Problem Q282749 – BUG: Deadlock information reported with SQL Server 2000 Profiler Locking and Blocking  Try This: Examine Blocked Processes 1. Open a Query Window and connect to the pubs database. Execute the following statements: BEGIN TRAN -- connection 1 UPDATE titles SET price = price + 1 2. Open another connection and execute the following statement: SELECT * FROM titles-- connection 2 3. Open a third connection and execute sp_who; note the process id (spid) of the blocked process. (Connection 3) 4. In the same connection, execute the following: SELECT spid, cmd, waittype FROM master..sysprocesses WHERE waittype 0 -- connection 3 5. Do not close any of the connections! What was the wait type of the blocked process?  Try This: Look at locks held Assumes all your connections are still open from the previous exercise. • Execute sp_lock -- Connection 3 What locks is the process from the previous example holding? Make sure you run ROLLBACK TRAN in Connection 1 to clean up your transaction. Collecting Information See Module 2 for more about how to gather this information using various tools. Recognizing Blocking Problems How to Recognize Blocking Problems  Users complain about poor performance at a certain time of day, or after a certain number of users connect.  SELECT * FROM sysprocesses or sp_who2 shows non-zero values in the blocked or BlkBy column.  More severe blocking incidents will have long blocking chains or large sysprocesses.waittime values for blocked spids.  Possibl

文将对 Linux™ 程序员可以使用的内存管理技术进行概述，虽然关注的重点是 C 语言，但同样也适用于其他语言。文中将为您提供如何管理内存的细节，然后将进一步展示如何手工管理内存，如何使用引用计数或者内存池来半手工地管理内存，以及如何使用垃圾收集自动管理内存。为什么必须管理内存内存管理是计算机编程最为基本的领域之一。在很多脚本语言中，您不必担心内存是如何管理的，这并不能使得内存管理的重要性有一点点降低。对实际编程来说，理解您的内存管理器的能力与局限性至关重要。在大部分系统语言中，比如 C 和 C++，您必须进行内存管理。本文将介绍手工的、半手工的以及自动的内存管理实践的基本概念。追溯到在 Apple II 上进行汇编语言编程的时代，那时内存管理还不是个大问题。您实际上在运行整个系统。系统有多少内存，您就有多少内存。您甚至不必费心思去弄明白它有多少内存，因为每一台机器的内存数量都相同。所以，如果内存需要非常固定，那么您只需要选择一个内存范围并使用它即可。不过，即使是在这样一个简单的计算机中，您也会有问题，尤其是当您不知道程序的每个部分将需要多少内存时。如果您的空间有限，而内存需求是变化的，那么您需要一些方法来满足这些需求：确定您是否有足够的内存来处理数据。从可用的内存中获取一部分内存。向可用内存池（pool）中返回部分内存，以使其可以由程序的其他部分或者其他程序使用。实现这些需求的程序库称为分配程序（allocators），因为它们负责分配和回收内存。程序的动态性越强，内存管理就越重要，您的内存分配程序的选择也就更重要。让我们来了解可用于内存管理的不同方法，它们的好处与不足，以及它们最适用的情形。回页首 C 风格的内存分配程序 C 编程语言提供了两个函数来满足我们的三个需求： malloc：该函数分配给定的字节数，并返回一个指向它们的指针。如果没有足够的可用内存，那么它返回一个空指针。 free：该函数获得指向由 malloc 分配的内存片段的指针，并将其释放，以便以后的程序或操作系统使用（实际上，一些 malloc 实现只能将内存归还给程序，而无法将内存归还给操作系统）。物理内存和虚拟内存要理解内存在程序中是如何分配的，首先需要理解如何将内存从操作系统分配给程序。计算机上的每一个进程都认为自己可以访问所有的物理内存。显然，由于同时在运行多个程序，所以每个进程不可能拥有全部内存。实际上，这些进程使用的是虚拟内存。只是作为一个例子，让我们假定您的程序正在访问地址为 629 的内存。不过，虚拟内存系统不需要将其存储在位置为 629 的 RAM 中。实际上，它甚至可以不在 RAM 中 —— 如果物理 RAM 已经满了，它甚至可能已经被转移到硬盘上！由于这类地址不必反映内存所在的物理位置，所以它们被称为虚拟内存。操作系统维持着一个虚拟地址到物理地址的转换的表，以便计算机硬件可以正确地响应地址请求。并且，如果地址在硬盘上而不是在 RAM 中，那么操作系统将暂时停止您的进程，将其他内存转存到硬盘中，从硬盘上加载被请求的内存，然后再重新启动您的进程。这样，每个进程都获得了自己可以使用的地址空间，可以访问比您物理上安装的内存更多的内存。在 32-位 x86 系统上，每一个进程可以访问 4 GB 内存。现在，大部分人的系统上并没有 4 GB 内存，即使您将 swap 也算上，每个进程所使用的内存也肯定少于 4 GB。因此，当加载一个进程时，它会得到一个取决于某个称为系统中断点（system break）的特定地址的初始内存分配。该地址之后是未被映射的内存 —— 用于在 RAM 或者硬盘中没有分配相应物理位置的内存。因此，如果一个进程运行超出了它初始分配的内存，那么它必须请求操作系统“映射进来（map in）”更多的内存。（映射是一个表示一一对应关系的数学术语 —— 当内存的虚拟地址有一个对应的物理地址来存储内存内容时，该内存将被映射。）基于 UNIX 的系统有两个可映射到附加内存中的基本系统调用： brk： brk() 是一个非常简单的系统调用。还记得系统中断点吗？该位置是进程映射的内存边界。 brk() 只是简单地将这个位置向前或者向后移动，就可以向进程添加内存或者从进程取走内存。 mmap： mmap()，或者说是“内存映像”，类似于 brk()，但是更为灵活。首先，它可以映射任何位置的内存，而不单单只局限于进程。其次，它不仅可以将虚拟地址映射到物理的 RAM 或者 swap，它还可以将它们映射到文件和文件位置，这样，读写内存将对文件中的数据进行读写。不过，在这里，我们只关心 mmap 向进程添加被映射的内存的能力。 munmap() 所做的事情与 mmap() 相反。如您所见， brk() 或者 mmap() 都可以用来向我们的进程添加额外的虚拟内存。在我们的例子中将使用 brk()，因为它更简单，更通用。实现一个简单的分配程序如果您曾经编写过很多 C 程序，那么您可能曾多次使用过 malloc() 和 free()。不过，您可能没有用一些时间去思考它们在您的操作系统中是如何实现的。本节将向您展示 malloc 和 free 的一个最简化实现的代码，来帮助说明管理内存时都涉及到了哪些事情。要试着运行这些示例，需要先复制本代码清单，并将其粘贴到一个名为 malloc.c 的文件中。接下来，我将一次一个部分地对该清单进行解释。在大部分操作系统中，内存分配由以下两个简单的函数来处理： void *malloc(long numbytes)：该函数负责分配 numbytes 大小的内存，并返回指向第一个字节的指针。 void free(void *firstbyte)：如果给定一个由先前的 malloc 返回的指针，那么该函数会将分配的空间归还给进程的“空闲空间”。 malloc_init 将是初始化内存分配程序的函数。它要完成以下三件事：将分配程序标识为已经初始化，找到系统中最后一个有效内存地址，然后建立起指向我们管理的内存的指针。这三个变量都是全局变量：清单 1. 我们的简单分配程序的全局变量 int has_initialized = 0; void *managed_memory_start; void *last_valid_address; 如前所述，被映射的内存的边界（最后一个有效地址）常被称为系统中断点或者当前中断点。在很多 UNIX® 系统中，为了指出当前系统中断点，必须使用 sbrk(0) 函数。 sbrk 根据参数中给出的字节数移动当前系统中断点，然后返回新的系统中断点。使用参数 0 只是返回当前中断点。这里是我们的 malloc 初始化代码，它将找到当前中断点并初始化我们的变量：清单 2. 分配程序初始化函数 /* Include the sbrk function */ #include void malloc_init() { /* grab the last valid address from the OS */ last_valid_address = sbrk(0); /* we don't have any memory to manage yet, so *just set the beginning to be last_valid_address */ managed_memory_start = last_valid_address; /* Okay, we're initialized and ready to go */ has_initialized = 1; } 现在，为了完全地管理内存，我们需要能够追踪要分配和回收哪些内存。在对内存块进行了 free 调用之后，我们需要做的是诸如将它们标记为未被使用的等事情，并且，在调用 malloc 时，我们要能够定位未被使用的内存块。因此， malloc 返回的每块内存的起始处首先要有这个结构：清单 3. 内存控制块结构定义 struct mem_control_block { int is_available; int size; }; 现在，您可能会认为当程序调用 malloc 时这会引发问题 —— 它们如何知道这个结构？答案是它们不必知道；在返回指针之前，我们会将其移动到这个结构之后，把它隐藏起来。这使得返回的指针指向没有用于任何其他用途的内存。那样，从调用程序的角度来看，它们所得到的全部是空闲的、开放的内存。然后，当通过 free() 将该指针传递回来时，我们只需要倒退几个内存字节就可以再次找到这个结构。在讨论分配内存之前，我们将先讨论释放，因为它更简单。为了释放内存，我们必须要做的惟一一件事情就是，获得我们给出的指针，回退 sizeof(struct mem_control_block) 个字节，并将其标记为可用的。这里是对应的代码：清单 4. 解除分配函数 void free(void *firstbyte) { struct mem_control_block *mcb; /* Backup from the given pointer to find the * mem_control_block */ mcb = firstbyte - sizeof(struct mem_control_block); /* Mark the block as being available */ mcb->is_available = 1; /* That's It! We're done. */ return; } 如您所见，在这个分配程序中，内存的释放使用了一个非常简单的机制，在固定时间内完成内存释放。分配内存稍微困难一些。以下是该算法的略述：清单 5. 主分配程序的伪代码 1. If our allocator has not been initialized, initialize it. 2. Add sizeof(struct mem_control_block) to the size requested. 3. start at managed_memory_start. 4. Are we at last_valid address? 5. If we are: A. We didn't find any existing space that was large enough -- ask the operating system for more and return that. 6. Otherwise: A. Is the current space available (check is_available from the mem_control_block)? B. If it is: i) Is it large enough (check "size" from the mem_control_block)? ii) If so: a. Mark it as unavailable b. Move past mem_control_block and return the pointer iii) Otherwise: a. Move forward "size" bytes b. Go back go step 4 C. Otherwise: i) Move forward "size" bytes ii) Go back to step 4 我们主要使用连接的指针遍历内存来寻找开放的内存块。这里是代码：清单 6. 主分配程序 void *malloc(long numbytes) { /* Holds where we are looking in memory */ void *current_location; /* This is the same as current_location, but cast to a * memory_control_block */ struct mem_control_block *current_location_mcb; /* This is the memory location we will return. It will * be set to 0 until we find something suitable */ void *memory_location; /* Initialize if we haven't already done so */ if(! has_initialized) { malloc_init(); } /* The memory we search for has to include the memory * control block, but the users of malloc don't need * to know this, so we'll just add it in for them. */ numbytes = numbytes + sizeof(struct mem_control_block); /* Set memory_location to 0 until we find a suitable * location */ memory_location = 0; /* Begin searching at the start of managed memory */ current_location = managed_memory_start; /* Keep going until we have searched all allocated space */ while(current_location != last_valid_address) { /* current_location and current_location_mcb point * to the same address. However, current_location_mcb * is of the correct type, so we can use it as a struct. * current_location is a void pointer so we can use it * to calculate addresses. */ current_location_mcb = (struct mem_control_block *)current_location; if(current_location_mcb->is_available) { if(current_location_mcb->size >= numbytes) { /* Woohoo! We've found an open, * appropriately-size location. */ /* It is no longer available */ current_location_mcb->is_available = 0; /* We own it */ memory_location = current_location; /* Leave the loop */ break; } } /* If we made it here, it's because the Current memory * block not suitable; move to the next one */ current_location = current_location + current_location_mcb->size; } /* If we still don't have a valid location, we'll * have to ask the operating system for more memory */ if(! memory_location) { /* Move the program break numbytes further */ sbrk(numbytes); /* The new memory will be where the last valid * address left off */ memory_location = last_valid_address; /* We'll move the last valid address forward * numbytes */ last_valid_address = last_valid_address + numbytes; /* We need to initialize the mem_control_block */ current_location_mcb = memory_location; current_location_mcb->is_available = 0; current_location_mcb->size = numbytes; } /* Now, no matter what (well, except for error conditions), * memory_location has the address of the memory, including * the mem_control_block */ /* Move the pointer past the mem_control_block */ memory_location = memory_location + sizeof(struct mem_control_block); /* Return the pointer */ return memory_location; } 这就是我们的内存管理器。现在，我们只需要构建它，并在程序中使用它即可。运行下面的命令来构建 malloc 兼容的分配程序（实际上，我们忽略了 realloc() 等一些函数，不过， malloc() 和 free() 才是最主要的函数）：清单 7. 编译分配程序 gcc -shared -fpic malloc.c -o malloc.so 该程序将生成一个名为 malloc.so 的文件，它是一个包含有我们的代码的共享库。在 UNIX 系统中，现在您可以用您的分配程序来取代系统的 malloc()，做法如下：清单 8. 替换您的标准的 malloc LD_PRELOAD=/path/to/malloc.so export LD_PRELOAD LD_PRELOAD 环境变量使动态链接器在加载任何可执行程序之前，先加载给定的共享库的符号。它还为特定库中的符号赋予优先权。因此，从现在起，该会话中的任何应用程序都将使用我们的 malloc()，而不是只有系统的应用程序能够使用。有一些应用程序不使用 malloc()，不过它们是例外。其他使用 realloc() 等其他内存管理函数的应用程序，或者错误地假定 malloc() 内部行为的那些应用程序，很可能会崩溃。ash shell 似乎可以使用我们的新 malloc() 很好地工作。如果您想确保 malloc() 正在被使用，那么您应该通过向函数的入口点添加 write() 调用来进行测试。我们的内存管理器在很多方面都还存在欠缺，但它可以有效地展示内存管理需要做什么事情。它的某些缺点包括：由于它对系统中断点（一个全局变量）进行操作，所以它不能与其他分配程序或者 mmap 一起使用。当分配内存时，在最坏的情形下，它将不得不遍历全部进程内存；其中可能包括位于硬盘上的很多内存，这意味着操作系统将不得不花时间去向硬盘移入数据和从硬盘中移出数据。没有很好的内存不足处理方案（ malloc 只假定内存分配是成功的）。它没有实现很多其他的内存函数，比如 realloc()。由于 sbrk() 可能会交回比我们请求的更多的内存，所以在堆（heap）的末端会遗漏一些内存。虽然 is_available 标记只包含一位信息，但它要使用完整的 4-字节的字。分配程序不是线程安全的。分配程序不能将空闲空间拼合为更大的内存块。分配程序的过于简单的匹配算法会导致产生很多潜在的内存碎片。我确信还有很多其他问题。这就是为什么它只是一个例子! 其他 malloc 实现 malloc() 的实现有很多，这些实现各有优点与缺点。在设计一个分配程序时，要面临许多需要折衷的选择，其中包括：分配的速度。回收的速度。有线程的环境的行为。内存将要被用光时的行为。局部缓存。簿记（Bookkeeping）内存开销。虚拟内存环境中的行为。小的或者大的对象。实时保证。每一个实现都有其自身的优缺点集合。在我们的简单的分配程序中，分配非常慢，而回收非常快。另外，由于它在使用虚拟内存系统方面较差，所以它最适于处理大的对象。还有其他许多分配程序可以使用。其中包括： Doug Lea Malloc：Doug Lea Malloc 实际上是完整的一组分配程序，其中包括 Doug Lea 的原始分配程序，GNU libc 分配程序和 ptmalloc。 Doug Lea 的分配程序有着与我们的版本非常类似的基本结构，但是它加入了索引，这使得搜索速度更快，并且可以将多个没有被使用的块组合为一个大的块。它还支持缓存，以便更快地再次使用最近释放的内存。 ptmalloc 是 Doug Lea Malloc 的一个扩展版本，支持多线程。在本文后面的参考资料部分中，有一篇描述 Doug Lea 的 Malloc 实现的文章。 BSD Malloc：BSD Malloc 是随 4.2 BSD 发行的实现，包含在 FreeBSD 之中，这个分配程序可以从预先确实大小的对象构成的池中分配对象。它有一些用于对象大小的 size 类，这些对象的大小为 2 的若干次幂减去某一常数。所以，如果您请求给定大小的一个对象，它就简单地分配一个与之匹配的 size 类。这样就提供了一个快速的实现，但是可能会浪费内存。在参考资料部分中，有一篇描述该实现的文章。 Hoard：编写 Hoard 的目标是使内存分配在多线程环境中进行得非常快。因此，它的构造以锁的使用为中心，从而使所有进程不必等待分配内存。它可以显著地加快那些进行很多分配和回收的多线程进程的速度。在参考资料部分中，有一篇描述该实现的文章。众多可用的分配程序中最有名的就是上述这些分配程序。如果您的程序有特别的分配需求，那么您可能更愿意编写一个定制的能匹配您的程序内存分配方式的分配程序。不过，如果不熟悉分配程序的设计，那么定制分配程序通常会带来比它们解决的问题更多的问题。要获得关于该主题的适当的介绍，请参阅 Donald Knuth 撰写的 The Art of Computer Programming Volume 1: Fundamental Algorithms 中的第 2.5 节“Dynamic Storage Allocation”（请参阅参考资料中的链接）。它有点过时，因为它没有考虑虚拟内存环境，不过大部分算法都是基于前面给出的函数。在 C++ 中，通过重载 operator new()，您可以以每个类或者每个模板为单位实现自己的分配程序。在 Andrei Alexandrescu 撰写的 Modern C++ Design 的第 4 章（“Small Object Allocation”）中，描述了一个小对象分配程序（请参阅参考资料中的链接）。基于 malloc() 的内存管理的缺点不只是我们的内存管理器有缺点，基于 malloc() 的内存管理器仍然也有很多缺点，不管您使用的是哪个分配程序。对于那些需要保持长期存储的程序使用 malloc() 来管理内存可能会非常令人失望。如果您有大量的不固定的内存引用，经常难以知道它们何时被释放。生存期局限于当前函数的内存非常容易管理，但是对于生存期超出该范围的内存来说，管理内存则困难得多。而且，关于内存管理是由进行调用的程序还是由被调用的函数来负责这一问题，很多 API 都不是很明确。因为管理内存的问题，很多程序倾向于使用它们自己的内存管理规则。C++ 的异常处理使得这项任务更成问题。有时好像致力于管理内存分配和清理的代码比实际完成计算任务的代码还要多！因此，我们将研究内存管理的其他选择。回页首半自动内存管理策略引用计数引用计数是一种半自动（semi-automated）的内存管理技术，这表示它需要一些编程支持，但是它不需要您确切知道某一对象何时不再被使用。引用计数机制为您完成内存管理任务。在引用计数中，所有共享的数据结构都有一个域来包含当前活动“引用”结构的次数。当向一个程序传递一个指向某个数据结构指针时，该程序会将引用计数增加 1。实质上，您是在告诉数据结构，它正在被存储在多少个位置上。然后，当您的进程完成对它的使用后，该程序就会将引用计数减少 1。结束这个动作之后，它还会检查计数是否已经减到零。如果是，那么它将释放内存。这样做的好处是，您不必追踪程序中某个给定的数据结构可能会遵循的每一条路径。每次对其局部的引用，都将导致计数的适当增加或减少。这样可以防止在使用数据结构时释放该结构。不过，当您使用某个采用引用计数的数据结构时，您必须记得运行引用计数函数。另外，内置函数和第三方的库不会知道或者可以使用您的引用计数机制。引用计数也难以处理发生循环引用的数据结构。要实现引用计数，您只需要两个函数 —— 一个增加引用计数，一个减少引用计数并当计数减少到零时释放内存。一个示例引用计数函数集可能看起来如下所示：清单 9. 基本的引用计数函数 /* Structure Definitions*/ /* Base structure that holds a refcount */ struct refcountedstruct { int refcount; } /* All refcounted structures must mirror struct * refcountedstruct for their first variables */ /* Refcount maintenance functions */ /* Increase reference count */ void REF(void *data) { struct refcountedstruct *rstruct; rstruct = (struct refcountedstruct *) data; rstruct->refcount++; } /* Decrease reference count */ void UNREF(void *data) { struct refcountedstruct *rstruct; rstruct = (struct refcountedstruct *) data; rstruct->refcount--; /* Free the structure if there are no more users */ if(rstruct->refcount == 0) { free(rstruct); } } REF 和 UNREF 可能会更复杂，这取决于您想要做的事情。例如，您可能想要为多线程程序增加锁，那么您可能想扩展 refcountedstruct，使它同样包含一个指向某个在释放内存之前要调用的函数的指针（类似于面向对象语言中的析构函数 —— 如果您的结构中包含这些指针，那么这是必需的）。当使用 REF 和 UNREF 时，您需要遵守这些指针的分配规则： UNREF 分配前左端指针（left-hand-side pointer）指向的值。 REF 分配后左端指针（left-hand-side pointer）指向的值。在传递使用引用计数的结构的函数中，函数需要遵循以下这些规则：在函数的起始处 REF 每一个指针。在函数的结束处 UNREF 第一个指针。以下是一个使用引用计数的生动的代码示例：清单 10. 使用引用计数的示例 /* EXAMPLES OF USAGE */ /* Data type to be refcounted */ struct mydata { int refcount; /* same as refcountedstruct */ int datafield1; /* Fields specific to this struct */ int datafield2; /* other declarations would go here as appropriate */ }; /* Use the functions in code */ void dosomething(struct mydata *data) { REF(data); /* Process data */ /* when we are through */ UNREF(data); } struct mydata *globalvar1; /* Note that in this one, we don't decrease the * refcount since we are maintaining the reference * past the end of the function call through the * global variable */ void storesomething(struct mydata *data) { REF(data); /* passed as a parameter */ globalvar1 = data; REF(data); /* ref because of Assignment */ UNREF(data); /* Function finished */ } 由于引用计数是如此简单，大部分程序员都自已去实现它，而不是使用库。不过，它们依赖于 malloc 和 free 等低层的分配程序来实际地分配和释放它们的内存。在 Perl 等高级语言中，进行内存管理时使用引用计数非常广泛。在这些语言中，引用计数由语言自动地处理，所以您根本不必担心它，除非要编写扩展模块。由于所有内容都必须进行引用计数，所以这会对速度产生一些影响，但它极大地提高了编程的安全性和方便性。以下是引用计数的益处：实现简单。易于使用。由于引用是数据结构的一部分，所以它有一个好的缓存位置。不过，它也有其不足之处：要求您永远不要忘记调用引用计数函数。无法释放作为循环数据结构的一部分的结构。减缓几乎每一个指针的分配。尽管所使用的对象采用了引用计数，但是当使用异常处理（比如 try 或 setjmp()/ longjmp()）时，您必须采取其他方法。需要额外的内存来处理引用。引用计数占用了结构中的第一个位置，在大部分机器中最快可以访问到的就是这个位置。在多线程环境中更慢也更难以使用。 C++ 可以通过使用智能指针（smart pointers）来容忍程序员所犯的一些错误，智能指针可以为您处理引用计数等指针处理细节。不过，如果不得不使用任何先前的不能处理智能指针的代码（比如对 C 库的联接），实际上，使用它们的后果通实比不使用它们更为困难和复杂。因此，它通常只是有益于纯 C++ 项目。如果您想使用智能指针，那么您实在应该去阅读 Alexandrescu 撰写的 Modern C++ Design 一书中的“Smart Pointers”那一章。内存池内存池是另一种半自动内存管理方法。内存池帮助某些程序进行自动内存管理，这些程序会经历一些特定的阶段，而且每个阶段中都有分配给进程的特定阶段的内存。例如，很多网络服务器进程都会分配很多针对每个连接的内存 —— 内存的最大生存期限为当前连接的存在期。Apache 使用了池式内存（pooled memory），将其连接拆分为各个阶段，每个阶段都有自己的内存池。在结束每个阶段时，会一次释放所有内存。在池式内存管理中，每次内存分配都会指定内存池，从中分配内存。每个内存池都有不同的生存期限。在 Apache 中，有一个持续时间为服务器存在期的内存池，还有一个持续时间为连接的存在期的内存池，以及一个持续时间为请求的存在期的池，另外还有其他一些内存池。因此，如果我的一系列函数不会生成比连接持续时间更长的数据，那么我就可以完全从连接池中分配内存，并知道在连接结束时，这些内存会被自动释放。另外，有一些实现允许注册清除函数（cleanup functions），在清除内存池之前，恰好可以调用它，来完成在内存被清理前需要完成的其他所有任务（类似于面向对象中的析构函数）。要在自己的程序中使用池，您既可以使用 GNU libc 的 obstack 实现，也可以使用 Apache 的 Apache Portable Runtime。GNU obstack 的好处在于，基于 GNU 的 Linux 发行版本中默认会包括它们。Apache Portable Runtime 的好处在于它有很多其他工具，可以处理编写多平台服务器软件所有方面的事情。要深入了解 GNU obstack 和 Apache 的池式内存实现，请参阅参考资料部分中指向这些实现的文档的链接。下面的假想代码列表展示了如何使用 obstack：清单 11. obstack 的示例代码 #include #include /* Example code listing for using obstacks */ /* Used for obstack macros (xmalloc is a malloc function that exits if memory is exhausted */ #define obstack_chunk_alloc xmalloc #define obstack_chunk_free free /* Pools */ /* Only permanent allocations should go in this pool */ struct obstack *global_pool; /* This pool is for per-connection data */ struct obstack *connection_pool; /* This pool is for per-request data */ struct obstack *request_pool; void allocation_failed() { exit(1); } int main() { /* Initialize Pools */ global_pool = (struct obstack *) xmalloc (sizeof (struct obstack)); obstack_init(global_pool); connection_pool = (struct obstack *) xmalloc (sizeof (struct obstack)); obstack_init(connection_pool); request_pool = (struct obstack *) xmalloc (sizeof (struct obstack)); obstack_init(request_pool); /* Set the error handling function */ obstack_alloc_failed_handler = &allocation_failed; /* Server main loop */ while(1) { wait_for_connection(); /* We are in a connection */ while(more_requests_available()) { /* Handle request */ handle_request(); /* Free all of the memory allocated * in the request pool */ obstack_free(request_pool, NULL); } /* We're finished with the connection, time * to free that pool */ obstack_free(connection_pool, NULL); } } int handle_request() { /* Be sure that all object allocations are allocated * from the request pool */ int bytes_i_need = 400; void *data1 = obstack_alloc(request_pool, bytes_i_need); /* Do stuff to process the request */ /* return */ return 0; } 基本上，在操作的每一个主要阶段结束之后，这个阶段的 obstack 会被释放。不过，要注意的是，如果一个过程需要分配持续时间比当前阶段更长的内存，那么它也可以使用更长期限的 obstack，比如连接或者全局内存。传递给 obstack_free() 的 NULL 指出它应该释放 obstack 的全部内容。可以用其他的值，但是它们通常不怎么实用。使用池式内存分配的益处如下所示：应用程序可以简单地管理内存。内存分配和回收更快，因为每次都是在一个池中完成的。分配可以在 O(1) 时间内完成，释放内存池所需时间也差不多（实际上是 O(n) 时间，不过在大部分情况下会除以一个大的因数，使其变成 O(1)）。可以预先分配错误处理池（Error-handling pools），以便程序在常规内存被耗尽时仍可以恢复。有非常易于使用的标准实现。池式内存的缺点是：内存池只适用于操作可以分阶段的程序。内存池通常不能与第三方库很好地合作。如果程序的结构发生变化，则不得不修改内存池，这可能会导致内存管理系统的重新设计。您必须记住需要从哪个池进行分配。另外，如果在这里出错，就很难捕获该内存池。回页首垃圾收集垃圾收集（Garbage collection）是全自动地检测并移除不再使用的数据对象。垃圾收集器通常会在当可用内存减少到少于一个具体的阈值时运行。通常，它们以程序所知的可用的一组“基本”数据 —— 栈数据、全局变量、寄存器 —— 作为出发点。然后它们尝试去追踪通过这些数据连接到每一块数据。收集器找到的都是有用的数据；它没有找到的就是垃圾，可以被销毁并重新使用这些无用的数据。为了有效地管理内存，很多类型的垃圾收集器都需要知道数据结构内部指针的规划，所以，为了正确运行垃圾收集器，它们必须是语言本身的一部分。收集器的类型复制（copying）：这些收集器将内存存储器分为两部分，只允许数据驻留在其中一部分上。它们定时地从“基本”的元素开始将数据从一部分复制到另一部分。内存新近被占用的部分现在成为活动的，另一部分上的所有内容都认为是垃圾。另外，当进行这项复制操作时，所有指针都必须被更新为指向每个内存条目的新位置。因此，为使用这种垃圾收集方法，垃圾收集器必须与编程语言集成在一起。标记并清理（Mark and sweep）：每一块数据都被加上一个标签。不定期的，所有标签都被设置为 0，收集器从“基本”的元素开始遍历数据。当它遇到内存时，就将标签标记为 1。最后没有被标记为 1 的所有内容都认为是垃圾，以后分配内存时会重新使用它们。增量的（Incremental）：增量垃圾收集器不需要遍历全部数据对象。因为在收集期间的突然等待，也因为与访问所有当前数据相关的缓存问题（所有内容都不得不被页入（page-in）），遍历所有内存会引发问题。增量收集器避免了这些问题。保守的（Conservative）：保守的垃圾收集器在管理内存时不需要知道与数据结构相关的任何信息。它们只查看所有数据类型，并假定它们可以全部都是指针。所以，如果一个字节序列可以是一个指向一块被分配的内存的指针，那么收集器就将其标记为正在被引用。有时没有被引用的内存会被收集，这样会引发问题，例如，如果一个整数域中包含一个值，该值是已分配内存的地址。不过，这种情况极少发生，而且它只会浪费少量内存。保守的收集器的优势是，它们可以与任何编程语言相集成。 Hans Boehm 的保守垃圾收集器是可用的最流行的垃圾收集器之一，因为它是免费的，而且既是保守的又是增量的，可以使用 --enable-redirect-malloc 选项来构建它，并且可以将它用作系统分配程序的简易替代者（drop-in replacement）（用 malloc/ free 代替它自己的 API）。实际上，如果这样做，您就可以使用与我们在示例分配程序中所使用的相同的 LD_PRELOAD 技巧，在系统上的几乎任何程序中启用垃圾收集。如果您怀疑某个程序正在泄漏内存，那么您可以使用这个垃圾收集器来控制进程。在早期，当 Mozilla 严重地泄漏内存时，很多人在其中使用了这项技术。这种垃圾收集器既可以在 Windows® 下运行，也可以在 UNIX 下运行。垃圾收集的一些优点：您永远不必担心内存的双重释放或者对象的生命周期。使用某些收集器，您可以使用与常规分配相同的 API。其缺点包括：使用大部分收集器时，您都无法干涉何时释放内存。在多数情况下，垃圾收集比其他形式的内存管理更慢。垃圾收集错误引发的缺陷难于调试。如果您忘记将不再使用的指针设置为 null，那么仍然会有内存泄漏。回页首结束语一切都需要折衷：性能、易用、易于实现、支持线程的能力等，这里只列出了其中的一些。为了满足项目的要求，有很多内存管理模式可以供您使用。每种模式都有大量的实现，各有其优缺点。对很多项目来说，使用编程环境默认的技术就足够了，不过，当您的项目有特殊的需要时，了解可用的选择将会有帮助。下表对比了本文中涉及的内存管理策略。表 1. 内存分配策略的对比策略分配速度回收速度局部缓存易用性通用性实时可用 SMP 线程友好定制分配程序取决于实现取决于实现取决于实现很难无取决于实现取决于实现简单分配程序内存使用少时较快很快差容易高否否 GNU malloc 中快中容易高否中 Hoard 中中中容易高否是引用计数 N/A N/A 非常好中中是（取决于 malloc 实现）取决于实现池中非常快极好中中是（取决于 malloc 实现）取决于实现垃圾收集中（进行收集时慢）中差中中否几乎不增量垃圾收集中中中中中否几乎不增量保守垃圾收集中中中容易高否几乎不参考资料您可以参阅本文在 developerWorks 全球站点上的英文原文。 Web 上的文档 GNU C Library 手册的 obstacks 部分提供了 obstacks 编程接口。 Apache Portable Runtime 文档描述了它们的池式分配程序的接口。基本的分配程序 Doug Lea 的 Malloc 是最流行的内存分配程序之一。 BSD Malloc 用于大部分基于 BSD 的系统中。 ptmalloc 起源于 Doug Lea 的 malloc，用于 GLIBC 之中。 Hoard 是一个为多线程应用程序优化的 malloc 实现。 GNU Memory-Mapped Malloc（GDB 的组成部分）是一个基于 mmap() 的 malloc 实现。池式分配程序 GNU Obstacks（GNU Libc 的组成部分）是安装最多的池式分配程序，因为在每一个基于 glibc 的系统中都有它。 Apache 的池式分配程序（Apache Portable Runtime 中）是应用最为广泛的池式分配程序。 Squid 有其自己的池式分配程序。 NetBSD 也有其自己的池式分配程序。 talloc 是一个池式分配程序，是 Samba 的组成部分。智能指针和定制分配程序 Loki C++ Library 有很多为 C++ 实现的通用模式，包括智能指针和一个定制的小对象分配程序。垃圾收集器 Hahns Boehm Conservative Garbage Collector 是最流行的开源垃圾收集器，它可以用于常规的 C/C++ 程序。关于现代操作系统中的虚拟内存的文章 Marshall Kirk McKusick 和 Michael J. Karels 合著的 A New Virtual Memory Implementation for Berkeley UNIX 讨论了 BSD 的 VM 系统。 Mel Gorman's Linux VM Documentation 讨论了 Linux VM 系统。关于 malloc 的文章 Poul-Henning Kamp 撰写的 Malloc in Modern Virtual Memory Environments 讨论的是 malloc 以及它如何与 BSD 虚拟内存交互。 Berger、McKinley、Blumofe 和 Wilson 合著的 Hoard -- a Scalable Memory Allocator for Multithreaded Environments 讨论了 Hoard 分配程序的实现。 Marshall Kirk McKusick 和 Michael J. Karels 合著的 Design of a General Purpose Memory Allocator for the 4.3BSD UNIX Kernel 讨论了内核级的分配程序。 Doug Lea 撰写的 A Memory Allocator 给出了一个关于设计和实现分配程序的概述，其中包括设计选择与折衷。 Emery D. Berger 撰写的 Memory Management for High-Performance Applications 讨论的是定制内存管理以及它如何影响高性能应用程序。关于定制分配程序的文章 Doug Lea 撰写的 Some Storage Management Techniques for Container Classes 描述的是为 C++ 类编写定制分配程序。 Berger、Zorn 和 McKinley 合著的 Composing High-Performance Memory Allocators 讨论了如何编写定制分配程序来加快具体工作的速度。 Berger、Zorn 和 McKinley 合著的 Reconsidering Custom Memory Allocation 再次提及了定制分配的主题，看是否真正值得为其费心。关于垃圾收集的文章 Paul R. Wilson 撰写的 Uniprocessor Garbage Collection Techniques 给出了垃圾收集的一个基本概述。 Benjamin Zorn 撰写的 The Measured Cost of Garbage Collection 给出了关于垃圾收集和性能的硬数据（hard data）。 Hans-Juergen Boehm 撰写的 Memory Allocation Myths and Half-Truths 给出了关于垃圾收集的神话（myths）。 Hans-Juergen Boehm 撰写的 Space Efficient Conservative Garbage Collection 是一篇描述他的用于 C/C++ 的垃圾收集器的文章。 Web 上的通用参考资料内存管理参考中有很多关于内存管理参考资料和技术文章的链接。关于内存管理和内存层级的 OOPS Group Papers 是非常好的一组关于此主题的技术文章。 C++ 中的内存管理讨论的是为 C++ 编写定制的分配程序。 Programming Alternatives: Memory Management 讨论了程序员进行内存管理时的一些选择。垃圾收集 FAQ 讨论了关于垃圾收集您需要了解的所有内容。 Richard Jones 的 Garbage Collection Bibliography 有指向任何您想要的关于垃圾收集的文章的链接。书籍 Michael Daconta 撰写的 C++ Pointers and Dynamic Memory Management 介绍了关于内存管理的很多技术。 Frantisek Franek 撰写的 Memory as a Programming Concept in C and C++ 讨论了有效使用内存的技术与工具，并给出了在计算机编程中应当引起注意的内存相关错误的角色。 Richard Jones 和 Rafael Lins 合著的 Garbage Collection: Algorithms for Automatic Dynamic Memory Management 描述了当前使用的最常见的垃圾收集算法。在 Donald Knuth 撰写的 The Art of Computer Programming 第 1 卷 Fundamental Algorithms 的第 2.5 节“Dynamic Storage Allocation”中，描述了实现基本的分配程序的一些技术。在 Donald Knuth 撰写的 The Art of Computer Programming 第 1 卷 Fundamental Algorithms 的第 2.3.5 节“Lists and Garbage Collection”中，讨论了用于列表的垃圾收集算法。 Andrei Alexandrescu 撰写的 Modern C++ Design 第 4 章“Small Object Allocation”描述了一个比 C++ 标准分配程序效率高得多的一个高速小对象分配程序。 Andrei Alexandrescu 撰写的 Modern C++ Design 第 7 章“Smart Pointers”描述了在 C++ 中智能指针的实现。 Jonathan 撰写的 Programming from the Ground Up 第 8 章“Intermediate Memory Topics”中有本文使用的简单分配程序的一个汇编语言版本。来自 developerWorks 自我管理数据缓冲区内存（developerWorks，2004 年 1 月）略述了一个用于管理内存的自管理的抽象数据缓存器的伪 C （pseudo-C）实现。 A framework for the user defined malloc replacement feature （developerWorks，2002 年 2 月）展示了如何利用 AIX 中的一个工具，使用自己设计的内存子系统取代原有的内存子系统。掌握 Linux 调试技术（developerWorks，2002 年 8 月）描述了可以使用调试方法的 4 种不同情形：段错误、内存溢出、内存泄漏和挂起。在处理 Java 程序中的内存漏洞（developerWorks，2001 年 2 月）中，了解导致 Java 内存泄漏的原因，以及何时需要考虑它们。在 developerWorks Linux 专区中，可以找到更多为 Linux 开发人员准备的参考资料。从 developerWorks 的 Speed-start your Linux app 专区中，可以下载运行于 Linux 之上的 IBM 中间件产品的免费测试版本，其中包括 WebSphere® Studio Application Developer、WebSphere Application Server、DB2® Universal Database、Tivoli® Access Manager 和 Tivoli Directory Server，查找 how-to 文章和技术支持。通过参与 developerWorks blogs 加入到 developerWorks 社区。可以在 Developer Bookstore Linux 专栏中定购打折出售的 Linux 书籍。关于作者 Jonathan Bartlett 是 Programming from the Ground Up 一书的作者，这本书介绍的是 Linux 汇编语言编程。Jonathan Bartlett 是 New Media Worx 的总开发师，负责为客户开发 Web、视频、kiosk 和桌面应用程序。您可以通过 johnnyb@eskimo.com 与 Jonathan 联系。

Some JVMs put restrictions on the total amount of memory available on the heap. If you are getting OutOfMemoryError s while running Eclipse, the VM can be told to let the heap grow to a larger amount...

1、确定当前系统可以设置的最大值：Below command will help you to identify the maximum memory heap size than can be allocated to a JVMprocess.java -mx[value]m -versionStart with a higher value like 4G or higher and sl...

62,635

社区成员

307,269

社区内容

发帖

与我相关

我的任务

社区管理员

加入社区

近7日
近30日
至今

加载中

查看更多榜单

社区公告

暂无公告

试试用AI创作助手写篇文章吧

+ 用AI写文章