互斥对象有所有权，不像信号量。尽管在互斥量范围内的任何线程都可以获得一个未锁定的互斥量，并锁定对同一关键代码段的访问，但只有锁定了互斥量的线程才应该解锁它。

关于这个主题的好文章:

互斥量与信号量——第1部分:信号量 互斥量与信号量——第2部分:互斥量 互斥量与信号量——第3部分(最后一部分):互斥问题

来自第二部分:

The mutex is similar to the principles of the binary semaphore with one significant difference: the principle of ownership. Ownership is the simple concept that when a task locks (acquires) a mutex only it can unlock (release) it. If a task tries to unlock a mutex it hasn’t locked (thus doesn’t own) then an error condition is encountered and, most importantly, the mutex is not unlocked. If the mutual exclusion object doesn't have ownership then, irrelevant of what it is called, it is not a mutex.

在理论层面上，它们在语义上并无不同。您可以使用信号量实现互斥量，反之亦然(参见这里的示例)。在实践中，实现是不同的，它们提供的服务也略有不同。

实际的区别(就围绕它们的系统服务而言)在于互斥锁的实现旨在成为一种更轻量级的同步机制。在oracle语言中，互斥锁被称为锁存器，而信号量被称为等待。

在最低级别，他们使用某种原子测试和设置机制。它读取内存位置的当前值，计算某种条件，并在一条不能中断的指令中写入该位置的值。这意味着您可以获得一个互斥锁，并测试是否有人在您之前拥有它。

典型的互斥量实现有一个进程或线程执行test-and-set指令，并评估是否有其他东西设置了互斥量。这里的关键点是与调度程序没有交互，因此我们不知道(也不关心)谁设置了锁。然后，我们要么放弃我们的时间片，并在任务重新调度时再次尝试它，要么执行自旋锁。自旋锁是这样一种算法:

Count down from 5000:
     i. Execute the test-and-set instruction
    ii. If the mutex is clear, we have acquired it in the previous instruction 
        so we can exit the loop
   iii. When we get to zero, give up our time slice.

当我们完成执行受保护的代码(称为临界区)时，我们只需将互斥量的值设置为零或其他表示“清除”的值。如果有多个任务试图获取互斥量，那么下一个计划在互斥量释放后的任务将获得对资源的访问权。通常情况下，您可以使用互斥来控制同步资源，在这种资源中，只需要在很短的时间内对其进行独占访问，通常是对共享数据结构进行更新。

A semaphore is a synchronised data structure (typically using a mutex) that has a count and some system call wrappers that interact with the scheduler in a bit more depth than the mutex libraries would. Semaphores are incremented and decremented and used to block tasks until something else is ready. See Producer/Consumer Problem for a simple example of this. Semaphores are initialised to some value - a binary semaphore is just a special case where the semaphore is initialised to 1. Posting to a semaphore has the effect of waking up a waiting process.

一个基本的信号量算法如下所示:

(somewhere in the program startup)
Initialise the semaphore to its start-up value.

Acquiring a semaphore
   i. (synchronised) Attempt to decrement the semaphore value
  ii. If the value would be less than zero, put the task on the tail of the list of tasks waiting on the semaphore and give up the time slice.

Posting a semaphore
   i. (synchronised) Increment the semaphore value
  ii. If the value is greater or equal to the amount requested in the post at the front of the queue, take that task off the queue and make it runnable.  
 iii. Repeat (ii) for all tasks until the posted value is exhausted or there are no more tasks waiting.

在二进制信号量的情况下，两者之间的主要实际区别是围绕实际数据结构的系统服务的性质。

编辑:正如evan正确地指出的那样，自旋锁会降低单个处理器的速度。你只能在多处理器上使用自旋锁，因为在单处理器上，持有互斥锁的进程永远不会在另一个任务运行时重置它。自旋锁只在多处理器架构上有用。

互斥锁

Until recently, the only sleeping lock in the kernel was the semaphore. Most users of semaphores instantiated a semaphore with a count of one and treated them as a mutual exclusion lock—a sleeping version of the spin-lock. Unfortunately, semaphores are rather generic and do not impose any usage constraints. This makes them useful for managing exclusive access in obscure situations, such as complicated dances between the kernel and userspace. But it also means that simpler locking is harder to do, and the lack of enforced rules makes any sort of automated debugging or constraint enforcement impossible. Seeking a simpler sleeping lock, the kernel developers introduced the mutex.Yes, as you are now accustomed to, that is a confusing name. Let’s clarify.The term “mutex” is a generic name to refer to any sleeping lock that enforces mutual exclusion, such as a semaphore with a usage count of one. In recent Linux kernels, the proper noun “mutex” is now also a specific type of sleeping lock that implements mutual exclusion.That is, a mutex is a mutex.

互斥锁的简单性和效率来自于它在信号量要求之外强加给用户的附加约束。信号量是按照Dijkstra的原始设计来实现最基本的行为，而互斥锁则不同，它的用例更严格、更窄: n一次只能有一个任务持有互斥锁。也就是说，互斥锁的使用计数总是1。

Whoever locked a mutex must unlock it. That is, you cannot lock a mutex in one context and then unlock it in another. This means that the mutex isn’t suitable for more complicated synchronizations between kernel and user-space. Most use cases, however, cleanly lock and unlock from the same context. Recursive locks and unlocks are not allowed. That is, you cannot recursively acquire the same mutex, and you cannot unlock an unlocked mutex. A process cannot exit while holding a mutex. A mutex cannot be acquired by an interrupt handler or bottom half, even with mutex_trylock(). A mutex can be managed only via the official API: It must be initialized via the methods described in this section and cannot be copied, hand initialized, or reinitialized.

[1] Linux内核开发，第三版Robert Love

I think most of the answers here were confusing especially those saying that mutex can be released only by the process that holds it but semaphore can be signaled by ay process. The above line is kind of vague in terms of semaphore. To understand we should know that there are two kinds of semaphore one is called counting semaphore and the other is called a binary semaphore. In counting semaphore handles access to n number of resources where n can be defined before the use. Each semaphore has a count variable, which keeps the count of the number of resources in use, initially, it is set to n. Each process that wishes to uses a resource performs a wait() operation on the semaphore (thereby decrementing the count). When a process releases a resource, it performs a release() operation (incrementing the count). When the count becomes 0, all the resources are being used. After that, the process waits until the count becomes more than 0. Now here is the catch only the process that holds the resource can increase the count no other process can increase the count only the processes holding a resource can increase the count and the process waiting for the semaphore again checks and when it sees the resource available it decreases the count again. So in terms of binary semaphore, only the process holding the semaphore can increase the count, and count remains zero until it stops using the semaphore and increases the count and other process gets the chance to access the semaphore.

二进制信号量和互斥量之间的主要区别在于，信号量是一种信号机制，而互斥量是一种锁定机制，但二进制信号量的功能似乎与互斥量类似，这造成了混乱，但两者是适用于不同类型工作的不同概念。

互斥锁用于阻塞关键区域，而信号量用于计数。