Despite the fact that the Linux Kernel is mostly written in C, it makes broad use of some techniques from the field of object-oriented programming. Developers wanting to use these object-oriented techniques receive little support or guidance from the language and so are left to fend for themselves. As is often the case, this is a double-edged sword. The developer has enough flexibility to do really cool things, and equally the flexibility to do really stupid things, and it isn't always clear at first glance which is which, or more accurately: where on the spectrum a particular approach sits. Instead of looking to the language to provide guidance, a software engineer must look to established practice to find out what works well and what is best avoided. Interpreting established practice is not always as easy as one might like and the effort, once made, is worth preserving. To preserve that effort on your author's part, this article brings another installment in an occasional series on Linux Kernel Design Patterns and attempts to set out - with examples - the design patterns in the Linux Kernel which effect an object-oriented style of programming. Rather than providing a brief introduction to the object-oriented style, tempting though that is, we will assume the reader has a basic knowledge of objects, classes, methods, inheritance, and similar terms. For those as yet unfamiliar with these, there are plenty of resources to be found elsewhere on the web. Over two weeks we will look for patterns in just two areas: method dispatch and data inheritance. Despite their apparent simplicity they lead to some rich veins for investigation. This first article will focus on method dispatch.
Method DispatchThe large variety of styles of inheritance and rules for its usage in languages today seems to suggest that there is no uniform understanding of what "object-oriented" really means. The term is a bit like "love": everyone thinks they know what it means but when you get down to details people can find they have very different ideas. While what it means to be "oriented" might not be clear, what we mean by an "object" does seem to be uniformly agreed upon. It is simply an abstraction comprising both state and behavior. An object is like a record (Pascal) or struct (C), except that some of the names of members refer to functions which act on the other fields in the object. These function members are sometimes referred to a "methods". The most obvious way to implement objects in C is to declare a "struct" where some fields are pointers to functions which take a pointer to the struct itself as their first argument. The calling convention for method "foo" in object "bar" would simply be: bar->foo(bar, ...args); While this pattern is used in the Linux kernel it is not the dominant pattern so we will leave discussion of it until a little later. As methods (unlike state) are not normally changed on a per-object basis, a more common and only slightly less obvious approach is to collect all the methods for a particular class of objects into a separate structure, sometimes known as a "virtual function table" or vtable. The object then has a single pointer to this table rather than a separate pointer for each method, and consequently uses less memory. Read more: LWN.net Part 1, part 2
QR:
Method DispatchThe large variety of styles of inheritance and rules for its usage in languages today seems to suggest that there is no uniform understanding of what "object-oriented" really means. The term is a bit like "love": everyone thinks they know what it means but when you get down to details people can find they have very different ideas. While what it means to be "oriented" might not be clear, what we mean by an "object" does seem to be uniformly agreed upon. It is simply an abstraction comprising both state and behavior. An object is like a record (Pascal) or struct (C), except that some of the names of members refer to functions which act on the other fields in the object. These function members are sometimes referred to a "methods". The most obvious way to implement objects in C is to declare a "struct" where some fields are pointers to functions which take a pointer to the struct itself as their first argument. The calling convention for method "foo" in object "bar" would simply be: bar->foo(bar, ...args); While this pattern is used in the Linux kernel it is not the dominant pattern so we will leave discussion of it until a little later. As methods (unlike state) are not normally changed on a per-object basis, a more common and only slightly less obvious approach is to collect all the methods for a particular class of objects into a separate structure, sometimes known as a "virtual function table" or vtable. The object then has a single pointer to this table rather than a separate pointer for each method, and consequently uses less memory. Read more: LWN.net Part 1, part 2
QR:
Posts
0 comments:
Post a Comment