TB852 3BSC950263R1 ABB TB852 RCU链路终结器

上图中，每一个读者、更新者表示一个独立的线程，总共4个读线程，一个写线程。 RCU更新操作分为两个阶段：移除阶段和回收阶段。两个阶段通过宽限期隔开。更新者在移除元素后，通过synchronize_rcu()原语，初始化一个宽限期，并等待宽限期结束后，回收移除的元素。 1. 移除阶段：RCU更新通过rcu_assign_pointer()等函数移除或插入元素。现代CPU的指针操作都是原子的， rcu_assign_pointer()原语在大多数系统上编译为一个简单的指针赋值操作。TB852 3BSC950263R1 移除的元素仅可被移除阶段（以灰色显示）前的读者访问。 2. 回收阶段：一个宽限期后， 宽限期开始前的原有读者都完成读操作，因此，此阶段可安全释放由删除阶段删除的元素。一个宽限期可以用于多个删除阶段，即可由多个更新程序执行更新。 此外，跟踪RCU宽限期的开销可能会均摊到现有流程调度上，因此开销较小。对于某些常见的工作负载，宽限期跟踪开销可以被多个RCU更新操作均摊，从而使每个RCU更新的平均开销接近零。

临界区指的是一个访问共用资源（例如：共用设备或是共用存储器）的程序片段，而这些共用资源又无法同时被多个线程访问的特性。当有线程进入临界区段时，其他线程或是进程必须等待（例如：bounded waiting 等待法），有一些同步的机制必须在临界区段的进入点与离开点实现，以确保这些共用资源是被互斥获得使用，例如：semaphore。只能被单一线程访问的设备，例如：打印机。

读者提供一个信号告诉写者什么时候可以安全执行销毁操作，但是这个信号可能是被延时的，允许多个读侧临界区使用一个信号。RCU通常通过一个非原子地增加本地计数器告诉写者，这种操作开销非常小。另外，RCU中没有指定写者是否并行。

TB852 3BSC950263R1

TB852 3BSC950263R1 What is RCU? It is an abbreviation of Read, Copy-Update, which means read-copy update. It’s a synchronization mechanism. It takes the asymmetric distribution of synchronization overhead to its logical limit,
Basic concepts of RCU
TB852 3BSC950263R1 read side critical sections: Each critical section begins with rcu_read_lock() and ends with rcu_read_unlock(), and may contain functions such as rcu_dereference() that access RCU-protected data structures. These pointer functions implement the concept of dependent sequential loading, called memory_order_consume loading.

TB852 3BSC950263R1 Write side critical section: In order to accommodate the read side critical section, the writer delays destroying and maintaining multiple versions of the data structure, which has a lot of synchronization overhead. In addition, the writer must use some kind of synchronization mechanism, such as locking, to provide ordered updates.

TB852 3BSC950263R1 quiescent state: When a thread is not running the read-side critical section, it is silent. A state of silence that lasts for a considerable period of time is called an extended quiescent state.

TB852 3BSC950263R1 Grace period: The grace period is the time during which all threads go silent at least once. All readers reading the critical section before the grace period will end after the grace period. Different grace periods may overlap in part or in whole.

TB852 3BSC950263R1 Reader reading critical section traverses RCU data. If the writer removes an element from this data, a grace period is required before the memory reclaim operation can be performed. The diagram of the above operations is shown below, where the box marked read is a read critical section.
In the figure above, each reader and updater represents a separate thread, a total of four read threads and one write thread. The RCU update operation is divided into two phases: the remove phase and the reclaim phase. The two phases are separated by a grace period. After the element is removed, the updater initializes a grace period using the synchronize_rcu() primitive, waits for the grace period to end, and reclaims the removed element. 1. Removal phase: RCU updates remove or insert elements with functions such as rcu_assign_pointer(). Pointer operations on modern cpus are atomic, and the rcu_assign_pointer() primitive compiles to a simple pointer assignment operation on most systems. Elements removed by TB852 3BSC950263R1 can only be accessed by readers prior to the removal stage (shown in gray). 2. Recycle stage: After a grace period, the original readers before the grace period have completed the reading operation, so this stage can safely release the elements deleted by the delete stage. A grace period can be used for multiple deletion phases, and updates can be performed by multiple updaters. In addition, the overhead of tracking the RCU grace period may be spread evenly over the existing process schedule, so there is less overhead. For some common workloads, grace trace overhead can be spread evenly across multiple RCU update operations, bringing the average overhead per RCU update close to zero.
A critical section is a feature of a program segment that accesses a shared resource (such as a shared device or shared memory) that cannot be accessed by multiple threads at the same time. When a thread enters a critical section, other threads or processes must wait (for example, bounded waiting method), and some synchronization mechanism must be implemented at the entry and exit points of the critical section to ensure that the shared resources are mutually exclusive and available, such as semaphore. A device that can only be accessed by a single thread, such as a printer.

The reader provides a signal telling the writer when it is safe to perform the destroy operation, but this signal may be delayed to allow multiple read-side critical sections to use a signal. RCU usually tells the writer by a non-atomic increment of the local counter that this operation has very little overhead. In addition, RCU does not specify whether the writer is parallel.