1*6c300a71STang Yizhou.. SPDX-License-Identifier: GPL-2.0 2*6c300a71STang Yizhou.. include:: ../disclaimer-zh_CN.rst 3*6c300a71STang Yizhou 4*6c300a71STang Yizhou:Original: Documentation/locking/spinlocks.rst 5*6c300a71STang Yizhou 6*6c300a71STang Yizhou:翻译: 7*6c300a71STang Yizhou 8*6c300a71STang Yizhou 唐艺舟 Tang Yizhou <tangyeechou@gmail.com> 9*6c300a71STang Yizhou 10*6c300a71STang Yizhou========== 11*6c300a71STang Yizhou加锁的教训 12*6c300a71STang Yizhou========== 13*6c300a71STang Yizhou 14*6c300a71STang Yizhou教训 1:自旋锁 15*6c300a71STang Yizhou============== 16*6c300a71STang Yizhou 17*6c300a71STang Yizhou加锁最基本的原语是自旋锁(spinlock):: 18*6c300a71STang Yizhou 19*6c300a71STang Yizhou static DEFINE_SPINLOCK(xxx_lock); 20*6c300a71STang Yizhou 21*6c300a71STang Yizhou unsigned long flags; 22*6c300a71STang Yizhou 23*6c300a71STang Yizhou spin_lock_irqsave(&xxx_lock, flags); 24*6c300a71STang Yizhou ... 这里是临界区 .. 25*6c300a71STang Yizhou spin_unlock_irqrestore(&xxx_lock, flags); 26*6c300a71STang Yizhou 27*6c300a71STang Yizhou上述代码总是安全的。自旋锁将在 _本地_ 禁用中断,但它本身将保证全局锁定。所以它 28*6c300a71STang Yizhou将保证在该锁保护的区域内只有一个控制线程。即使在单处理器(UP)下也能很好的工作, 29*6c300a71STang Yizhou所以代码 _不_ 需要担心UP还是SMP的问题:自旋锁在两种情况下都能正常工作。 30*6c300a71STang Yizhou 31*6c300a71STang Yizhou 注意!自旋锁对内存的潜在影响由下述文档进一步描述: 32*6c300a71STang Yizhou 33*6c300a71STang Yizhou Documentation/memory-barriers.txt 34*6c300a71STang Yizhou 35*6c300a71STang Yizhou (5) ACQUIRE operations. 36*6c300a71STang Yizhou 37*6c300a71STang Yizhou (6) RELEASE operations. 38*6c300a71STang Yizhou 39*6c300a71STang Yizhou上述代码通常非常简单(对大部分情况,你通常需要并且只希望有一个自旋锁——使用多个 40*6c300a71STang Yizhou自旋锁会使事情变得更复杂,甚至更慢,而且通常仅仅在你 **理解的** 序列有被拆分的 41*6c300a71STang Yizhou需求时才值得这么做:如果你不确定的话,请不惜一切代价避免这样做)。 42*6c300a71STang Yizhou 43*6c300a71STang Yizhou这是关于自旋锁的唯一真正困难的部分:一旦你开始使用自旋锁,它们往往会扩展到你以前 44*6c300a71STang Yizhou可能没有注意到的领域,因为你必须确保自旋锁正确地保护共享数据结构 **每一处** 被 45*6c300a71STang Yizhou使用的地方。自旋锁是最容易被添加到完全独立于其它代码的地方(例如,没有人访问的 46*6c300a71STang Yizhou内部驱动数据结构)的。 47*6c300a71STang Yizhou 48*6c300a71STang Yizhou 注意!仅当你在跨CPU核访问时使用 **同一把** 自旋锁,对它的使用才是安全的。 49*6c300a71STang Yizhou 这意味着所有访问共享变量的代码必须对它们想使用的自旋锁达成一致。 50*6c300a71STang Yizhou 51*6c300a71STang Yizhou---- 52*6c300a71STang Yizhou 53*6c300a71STang Yizhou教训 2:读-写自旋锁 54*6c300a71STang Yizhou=================== 55*6c300a71STang Yizhou 56*6c300a71STang Yizhou如果你的数据访问有一个非常自然的模式,倾向于从共享变量中读取数据,读-写自旋锁 57*6c300a71STang Yizhou(rw_lock)有时是有用的。它们允许多个读者同时出现在同一个临界区,但是如果有人想 58*6c300a71STang Yizhou改变变量,它必须获得一个独占的写锁。 59*6c300a71STang Yizhou 60*6c300a71STang Yizhou 注意!读-写自旋锁比原始自旋锁需要更多的原子内存操作。除非读者的临界区很长, 61*6c300a71STang Yizhou 否则你最好只使用原始自旋锁。 62*6c300a71STang Yizhou 63*6c300a71STang Yizhou例程看起来和上面一样:: 64*6c300a71STang Yizhou 65*6c300a71STang Yizhou rwlock_t xxx_lock = __RW_LOCK_UNLOCKED(xxx_lock); 66*6c300a71STang Yizhou 67*6c300a71STang Yizhou unsigned long flags; 68*6c300a71STang Yizhou 69*6c300a71STang Yizhou read_lock_irqsave(&xxx_lock, flags); 70*6c300a71STang Yizhou .. 仅读取信息的临界区 ... 71*6c300a71STang Yizhou read_unlock_irqrestore(&xxx_lock, flags); 72*6c300a71STang Yizhou 73*6c300a71STang Yizhou write_lock_irqsave(&xxx_lock, flags); 74*6c300a71STang Yizhou .. 读取和独占写信息 ... 75*6c300a71STang Yizhou write_unlock_irqrestore(&xxx_lock, flags); 76*6c300a71STang Yizhou 77*6c300a71STang Yizhou上面这种锁对于复杂的数据结构如链表可能会有用,特别是在不改变链表的情况下搜索其中 78*6c300a71STang Yizhou的条目。读锁允许许多并发的读者。任何希望 **修改** 链表的代码将必须先获取写锁。 79*6c300a71STang Yizhou 80*6c300a71STang Yizhou 注意!RCU锁更适合遍历链表,但需要仔细注意设计细节(见Documentation/RCU/listRCU.rst)。 81*6c300a71STang Yizhou 82*6c300a71STang Yizhou另外,你不能把读锁“升级”为写锁,所以如果你在 _任何_ 时候需要做任何修改 83*6c300a71STang Yizhou(即使你不是每次都这样做),你必须在一开始就获得写锁。 84*6c300a71STang Yizhou 85*6c300a71STang Yizhou 注意!我们正在努力消除大多数情况下的读-写自旋锁的使用,所以请不要在没有达成 86*6c300a71STang Yizhou 共识的情况下增加一个新的(相反,请参阅Documentation/RCU/rcu.rst以获得完整 87*6c300a71STang Yizhou 信息)。 88*6c300a71STang Yizhou 89*6c300a71STang Yizhou---- 90*6c300a71STang Yizhou 91*6c300a71STang Yizhou教训 3:重新审视自旋锁 92*6c300a71STang Yizhou====================== 93*6c300a71STang Yizhou 94*6c300a71STang Yizhou上述的自旋锁原语绝不是唯一的。它们是最安全的,在所有情况下都能正常工作,但部分 95*6c300a71STang Yizhou**因为** 它们是安全的,它们也是相当慢的。它们比原本需要的更慢,因为它们必须要 96*6c300a71STang Yizhou禁用中断(在X86上只是一条指令,但却是一条昂贵的指令——而在其他体系结构上,情况 97*6c300a71STang Yizhou可能更糟)。 98*6c300a71STang Yizhou 99*6c300a71STang Yizhou如果你有必须保护跨CPU访问的数据结构且你想使用自旋锁的场景,你有可能使用代价小的 100*6c300a71STang Yizhou自旋锁版本。当且仅当你知道某自旋锁永远不会在中断处理程序中使用,你可以使用非中断 101*6c300a71STang Yizhou的版本:: 102*6c300a71STang Yizhou 103*6c300a71STang Yizhou spin_lock(&lock); 104*6c300a71STang Yizhou ... 105*6c300a71STang Yizhou spin_unlock(&lock); 106*6c300a71STang Yizhou 107*6c300a71STang Yizhou(当然,也可以使用相应的读-写锁版本)。这种自旋锁将同样可以保证独占访问,而且 108*6c300a71STang Yizhou速度会快很多。如果你知道有关的数据只在“进程上下文”中被存取,即,不涉及中断, 109*6c300a71STang Yizhou这种自旋锁就有用了。 110*6c300a71STang Yizhou 111*6c300a71STang Yizhou当这些版本的自旋锁涉及中断时,你不能使用的原因是会陷入死锁:: 112*6c300a71STang Yizhou 113*6c300a71STang Yizhou spin_lock(&lock); 114*6c300a71STang Yizhou ... 115*6c300a71STang Yizhou <- 中断来临: 116*6c300a71STang Yizhou spin_lock(&lock); 117*6c300a71STang Yizhou 118*6c300a71STang Yizhou一个中断试图对一个已经锁定的变量上锁。如果中断发生在另一个CPU上,不会有问题; 119*6c300a71STang Yizhou但如果中断发生在已经持有自旋锁的同一个CPU上,将 _会_ 有问题,因为该锁显然永远 120*6c300a71STang Yizhou不会被释放(因为中断正在等待该锁,而锁的持有者被中断打断,并且无法继续执行, 121*6c300a71STang Yizhou直到中断处理结束)。 122*6c300a71STang Yizhou 123*6c300a71STang Yizhou(这也是自旋锁的中断版本只需要禁用 _本地_ 中断的原因——在发生于其它CPU的中断中 124*6c300a71STang Yizhou使用同一把自旋锁是没问题的,因为发生于其它CPU的中断不会打断已经持锁的CPU,所以 125*6c300a71STang Yizhou锁的持有者可以继续执行并最终释放锁)。 126*6c300a71STang Yizhou 127*6c300a71STang Yizhou Linus 128*6c300a71STang Yizhou 129*6c300a71STang Yizhou---- 130*6c300a71STang Yizhou 131*6c300a71STang Yizhou参考信息 132*6c300a71STang Yizhou======== 133*6c300a71STang Yizhou 134*6c300a71STang Yizhou对于动态初始化,使用spin_lock_init()或rwlock_init()是合适的:: 135*6c300a71STang Yizhou 136*6c300a71STang Yizhou spinlock_t xxx_lock; 137*6c300a71STang Yizhou rwlock_t xxx_rw_lock; 138*6c300a71STang Yizhou 139*6c300a71STang Yizhou static int __init xxx_init(void) 140*6c300a71STang Yizhou { 141*6c300a71STang Yizhou spin_lock_init(&xxx_lock); 142*6c300a71STang Yizhou rwlock_init(&xxx_rw_lock); 143*6c300a71STang Yizhou ... 144*6c300a71STang Yizhou } 145*6c300a71STang Yizhou 146*6c300a71STang Yizhou module_init(xxx_init); 147*6c300a71STang Yizhou 148*6c300a71STang Yizhou对于静态初始化,使用DEFINE_SPINLOCK() / DEFINE_RWLOCK()或 149*6c300a71STang Yizhou__SPIN_LOCK_UNLOCKED() / __RW_LOCK_UNLOCKED()是合适的。 150