DPDK中无锁环形队列实现

因为最近在研究高性能方面的技术,突然想起上一份工作从事抗D的项目,在项目中使用到的dpdk组件,其中之一有无锁相关的技术,就重新翻了下源码,便写下这篇。

lock-free相关的技术也蛮久的了,与之有关的ABA问题,CAS相关的,还有wait-free这个,有兴趣的自行研究。

本篇主要介绍dpdk中无锁环形队列的实现原理,比如在单进程多线程框架中,有个主线程把接收到的请求均匀的分发到多个工作线程,要高效的,尽量减少同步互斥,阻塞的问题;这一类的问题也有成熟高效的实现方案如muduo库中的代码;这里dpdk中的实现和kfifo实现原理差不多。

先看下实现说明:

  73  * The Ring Manager is a fixed-size queue, implemented as a table of
  74  * pointers. Head and tail pointers are modified atomically, allowing
  75  * concurrent access to it. It has the following features:
  76  *
  77  * - FIFO (First In First Out)
  78  * - Maximum size is fixed; the pointers are stored in a table.
  79  * - Lockless implementation.
  80  * - Multi- or single-consumer dequeue.
  81  * - Multi- or single-producer enqueue.
  82  * - Bulk dequeue.
  83  * - Bulk enqueue.

ring的基本数据结构:

 152 struct rte_ring {
 159     int flags;                       /**< Flags supplied at creation. */
 160     const struct rte_memzone *memzone;
 161             /**< Memzone, if any, containing the rte_ring */
 162 
 163     /** Ring producer status. */
 164     struct prod {
 165         uint32_t watermark;      /**< Maximum items before EDQUOT. */
 166         uint32_t sp_enqueue;     /**< True, if single producer. */
 167         uint32_t size;           /**< Size of ring. */
 168         uint32_t mask;           /**< Mask (size-1) of ring. */
 169         volatile uint32_t head;  /**< Producer head. */
 170         volatile uint32_t tail;  /**< Producer tail. */
 171     } prod __rte_cache_aligned;
 172 
 173     /** Ring consumer status. */
 174     struct cons {
 175         uint32_t sc_dequeue;     /**< True, if single consumer. */
 176         uint32_t size;           /**< Size of the ring. */
 177         uint32_t mask;           /**< Mask (size-1) of ring. */
 178         volatile uint32_t head;  /**< Consumer head. */
 179         volatile uint32_t tail;  /**< Consumer tail. */
 180 #ifdef RTE_RING_SPLIT_PROD_CONS
 181     } cons __rte_cache_aligned;
 182 #else
 183     } cons;
 184 #endif
 190     void * ring[0] __rte_cache_aligned; /**< Memory space of ring starts here.
 191                                          * not volatile so need to be careful
 192                                          * about compiler re-ordering */
 193 };

其中prod和cons分别表示生产者和消费者数据类型声明,都cache line对齐,保证cpu读取数据时很快,且做到生产者和消费者线程的数据隔离,竞争不同的cache line;flags表示是单生产单消费还是其他,对应实现需cas保证正确性,这里假设是单生单消的情况;memzone用于记录在哪块mem分配的到时释放的时候使用;ring数组保存可用空间的地址;这里使用void * ring[0][不占任何空间,分配时连续,方便内存释放,能提高速度];其中prod和cons结构中的mask字段是2的幂次方减1,用于索引下标,这里不用担心会数组越界什么的,这样使用“index & mask”,比起取模要快得多,而且当index累加到uint32_t最大值后,再加一又回绕到0了;

这里需要注意的是prodcons都有headtail,作用会在下面解释。

ring的创建及初始化:

160 /* create the ring */
161 struct rte_ring *
162 rte_ring_create(const char *name, unsigned count, int socket_id,
163         unsigned flags)
164 {
166     struct rte_ring *r;
167     struct rte_tailq_entry *te;
168     const struct rte_memzone *mz;
169     ssize_t ring_size;
170     int mz_flags = 0;
171     struct rte_ring_list* ring_list = NULL;
172     int ret;
173 
174     ring_list = RTE_TAILQ_CAST(rte_ring_tailq.head, rte_ring_list);
175 
176     ring_size = rte_ring_get_memsize(count);
177     if (ring_size < 0) {
178         rte_errno = ring_size;
179         return NULL;
180     }
181     ...more code
189     te = rte_zmalloc("RING_TAILQ_ENTRY", sizeof(*te), 0);
190     if (te == NULL) {
191         RTE_LOG(ERR, RING, "Cannot reserve memory for tailq\n");
192         rte_errno = ENOMEM;
193         return NULL;
194     }
195 
196     rte_rwlock_write_lock(RTE_EAL_TAILQ_RWLOCK);
197 
198     /* reserve a memory zone for this ring. If we can't get rte_config or
199      * we are secondary process, the memzone_reserve function will set
200      * rte_errno for us appropriately - hence no check in this this function */
201     mz = rte_memzone_reserve(mz_name, ring_size, socket_id, mz_flags);
202     if (mz != NULL) {
203         r = mz->addr;
204         /* no need to check return value here, we already checked the
205          * arguments above */
206         rte_ring_init(r, name, count, flags);
207 
208         te->data = (void *) r;
209         r->memzone = mz;
210 
211         TAILQ_INSERT_TAIL(ring_list, te, next);
212     } else {
213         r = NULL;
214         RTE_LOG(ERR, RING, "Cannot reserve memory\n");
215         rte_free(te);
216     }
217     rte_rwlock_write_unlock(RTE_EAL_TAILQ_RWLOCK);
218 
219     return r;
220 }

rte_ring_create实现中,rte_ring_get_memsize先判断count有没有是2的幂次方和是否超过宏定义的最大size[#define RTE_RING_SZ_MASK (unsigned)(0x0fffffff)],然后调整大小至对齐cache line

153 #define RTE_ALIGN_CEIL(val, align) \
154     RTE_ALIGN_FLOOR(((val) + ((typeof(val)) (align) - 1)), align)

135 #define RTE_ALIGN_FLOOR(val, align) \
136     (typeof(val))((val) & (~((typeof(val))((align) - 1))))

计算总共需要的内存空间;分配保存ring地址的struct rte_tailq_entry后期用于释放,加写锁并设置ring相关数据成员,然后释放锁,这部分逻辑是比较简单的;但是里面也使用到了其他复杂的接口如rte_memzone_reserve,主要功能是就近socket id node开辟一块空间[rte_memzone_reserve_thread_safe],有些额外的信息,这里主要使用mz->addr用于ring,源码在后面如果有时间和篇幅再作分析,之后进行rte_ring_init

rte_ring_init 初始化中的RTE_BUILD_BUG_ON宏是编译期类型cache line对齐检查,具体作用也在前面的分析中说明过;剩下的是初始化ring各个成员变量:

121 int
122 rte_ring_init(struct rte_ring *r, const char *name, unsigned count,
123     unsigned flags)
124 {   
125     int ret;
126     
127     /* compilation-time checks */
128     RTE_BUILD_BUG_ON((sizeof(struct rte_ring) &
129               RTE_CACHE_LINE_MASK) != 0);
130 #ifdef RTE_RING_SPLIT_PROD_CONS
131     RTE_BUILD_BUG_ON((offsetof(struct rte_ring, cons) &
132               RTE_CACHE_LINE_MASK) != 0);
133 #endif
134     RTE_BUILD_BUG_ON((offsetof(struct rte_ring, prod) &
135               RTE_CACHE_LINE_MASK) != 0);
142     
143     /* init the ring structure */
144     memset(r, 0, sizeof(*r));
148     r->flags = flags; 
149     r->prod.watermark = count;
150     r->prod.sp_enqueue = !!(flags & RING_F_SP_ENQ);
151     r->cons.sc_dequeue = !!(flags & RING_F_SC_DEQ);
152     r->prod.size = r->cons.size = count;
153     r->prod.mask = r->cons.mask = count-1;
154     r->prod.head = r->cons.head = 0;
155     r->prod.tail = r->cons.tail = 0;
156 
157     return 0;
158 }

#define RTE_BUILD_BUG_ON(condition) ((void)sizeof(char[1 - 2*!!(condition)]))
#define RTE_CACHE_LINE_MASK (RTE_CACHE_LINE_SIZE-1)
#define offsetof(t, m) ((size_t) &((t *)0)->m)

ring的释放:

223 void
224 rte_ring_free(struct rte_ring *r)
225 {
226     struct rte_ring_list *ring_list = NULL;
227     struct rte_tailq_entry *te;
228 
229     if (r == NULL)
230         return;
231 
232     /*
233      * Ring was not created with rte_ring_create,
234      * therefore, there is no memzone to free.
235      */
236     if (r->memzone == NULL) {
238         return;
239     }
240 
241     if (rte_memzone_free(r->memzone) != 0) {
243         return;
244     }
245 
246     ring_list = RTE_TAILQ_CAST(rte_ring_tailq.head, rte_ring_list);
247     rte_rwlock_write_lock(RTE_EAL_TAILQ_RWLOCK);
248 
249     /* find out tailq entry */
250     TAILQ_FOREACH(te, ring_list, next) {
251         if (te->data == (void *) r)
252             break;
253     }
254 
255     if (te == NULL) {
256         rte_rwlock_write_unlock(RTE_EAL_TAILQ_RWLOCK);
257         return;
258     }
259 
260     TAILQ_REMOVE(ring_list, te, next);
262     rte_rwlock_write_unlock(RTE_EAL_TAILQ_RWLOCK);
264     rte_free(te);
265 }

释放没什么好分析的,详细释放过程不作过多分析,这里先进行free(r->memzone),再把自己从全局链表中摘除,并释放自己;

先说明下入ring有两种模式rte_ring_queue_behavior

RTE_RING_QUEUE_FIXED:    Enqueue a fixed number of items from a ring
RTE_RING_QUEUE_VARIABLE: Enqueue as many items a possible from ring

下面具体先分析单生产者和单消费者的实现过程,如何做到无锁高性能,由RING_F_SP_ENQRING_F_SC_DEQ表示,以下是生产者过程:

 539 static inline int __attribute__((always_inline))
 540 __rte_ring_sp_do_enqueue(struct rte_ring *r, void * const *obj_table,
 541              unsigned n, enum rte_ring_queue_behavior behavior)
 542 {
 543     uint32_t prod_head, cons_tail;
 544     uint32_t prod_next, free_entries;
 545     unsigned i;
 546     uint32_t mask = r->prod.mask;
 547     int ret;
 548 
 549     prod_head = r->prod.head;
 550     cons_tail = r->cons.tail;
 551     /* The subtraction is done between two unsigned 32bits value
 552      * (the result is always modulo 32 bits even if we have
 553      * prod_head > cons_tail). So 'free_entries' is always between 0
 554      * and size(ring)-1. */
 555     free_entries = mask + cons_tail - prod_head;
 557     /* check that we have enough room in ring */
 558     if (unlikely(n > free_entries)) {
 559         if (behavior == RTE_RING_QUEUE_FIXED) {
 560             __RING_STAT_ADD(r, enq_fail, n);
 561             return -ENOBUFS;
 562         }
 563         else {
 564             /* No free entry available */
 565             if (unlikely(free_entries == 0)) {
 566                 __RING_STAT_ADD(r, enq_fail, n);
 567                 return 0;
 568             }
 569 
 570             n = free_entries;
 571         }
 572     }
 573 
 574     prod_next = prod_head + n;
 575     r->prod.head = prod_next;
 577     /* write entries in ring */
 578     ENQUEUE_PTRS();
 579     rte_smp_wmb();
 580 
 581     /* if we exceed the watermark */
 582     if (unlikely(((mask + 1) - free_entries + n) > r->prod.watermark)) {
 583         ret = (behavior == RTE_RING_QUEUE_FIXED) ? -EDQUOT :
 584             (int)(n | RTE_RING_QUOT_EXCEED);
 585         __RING_STAT_ADD(r, enq_quota, n);
 586     }
 587     else {
 588         ret = (behavior == RTE_RING_QUEUE_FIXED) ? 0 : n;
 589         __RING_STAT_ADD(r, enq_success, n);
 590     }
 591 
 592     r->prod.tail = prod_next;
 593     return ret;
 594 }

其中代码行555〜572是判断要入队列ring的元素个数是否大于可用空间,取决于free_entries的情况,这些只是检查;其中free_entries = mask + cons_tail - prod_head算出来的值区间总是在[0,size-1]

一开始r->prod.head和r->cons.tail都为0,此时执行prod_next = prod_head + n,接着修改prod.head,即r->prod.head = prod_next(A1);接着入队元素ENQUEUE_PTRS()(A2):

 356 #define ENQUEUE_PTRS() do { \
 357     const uint32_t size = r->prod.size; \
 358     uint32_t idx = prod_head & mask; \
 359     if (likely(idx + n < size)) { \
 360         for (i = 0; i < (n & ((~(unsigned)0x3))); i+=4, idx+=4) { \
 361             r->ring[idx] = obj_table[i]; \
 362             r->ring[idx+1] = obj_table[i+1]; \
 363             r->ring[idx+2] = obj_table[i+2]; \
 364             r->ring[idx+3] = obj_table[i+3]; \
 365         } \
 366         switch (n & 0x3) { \
 367             case 3: r->ring[idx++] = obj_table[i++]; \
 368             case 2: r->ring[idx++] = obj_table[i++]; \
 369             case 1: r->ring[idx++] = obj_table[i++]; \
 370         } \
 371     } else { \
 372         for (i = 0; idx < size; i++, idx++)\
 373             r->ring[idx] = obj_table[i]; \
 374         for (idx = 0; i < n; i++, idx++) \
 375             r->ring[idx] = obj_table[i]; \
 376     } \
 377 } while(0)

这里likely成立的情况当要入队的元素不需要回绕队列起始处的时候,然后循环步长是四,依次赋值,相当于循环展开优化,减少计算循环索引和预测条件分支判断等优化;如果需要回绕,那么先入队[idx,size),再入队[0,idx)索引处;

入完队后执行rte_smp_wmb()(A3),即do { asm volatile ("dmb st" : : : "memory"); } while (0),这里表示要等待前面所有存储内存的指令执行完后再执行后面的存储内存的指令;这语句作用如注释:

 54 /**
 55  * Write memory barrier.
 56  *
 57  * Guarantees that the STORE operations generated before the barrier
 58  * occur before the STORE operations generated after.
 59  */

内存屏障相关的知识点会单独开篇博客整理下,这里先不作过错说明,只要知道在编译器和cpu会给我们的代码进行优化乱序执行等,为了保证正确性,wmb表示在此之前的写不能放在后面执行,在此之后的写不能放到前面执行;rmb是针对读的,mb两都限制;
代码行582〜590判断有没有超过最大watermark[Quota exceeded. The objects have been enqueued, but the high water mark is exceeded.];不过这里有个疑问是什么情况下if会成立,因为这里用了unlikely,表示很大可能if不会成立;
最后执行r->prod.tail = prod_next(A4);
(A1)和(A4)位于(A3)前后。

再来看消费者过程:

 722 static inline int __attribute__((always_inline))
 723 __rte_ring_sc_do_dequeue(struct rte_ring *r, void **obj_table,
 724          unsigned n, enum rte_ring_queue_behavior behavior)
 725 {
 726     uint32_t cons_head, prod_tail;
 727     uint32_t cons_next, entries;
 728     unsigned i;
 729     uint32_t mask = r->prod.mask;
 730 
 731     cons_head = r->cons.head;
 732     prod_tail = r->prod.tail;
 733     /* The subtraction is done between two unsigned 32bits value
 734      * (the result is always modulo 32 bits even if we have
 735      * cons_head > prod_tail). So 'entries' is always between 0
 736      * and size(ring)-1. */
 737     entries = prod_tail - cons_head;
 738 
 739     if (n > entries) {
 740         if (behavior == RTE_RING_QUEUE_FIXED) {
 741             __RING_STAT_ADD(r, deq_fail, n);
 742             return -ENOENT;
 743         }
 744         else {
 745             if (unlikely(entries == 0)){
 746                 __RING_STAT_ADD(r, deq_fail, n);
 747                 return 0;
 748             }
 749 
 750             n = entries;
 751         }
 752     }
 754     cons_next = cons_head + n;
 755     r->cons.head = cons_next;
 756     
 757     /* copy in table */
 758     DEQUEUE_PTRS();
 759     rte_smp_rmb();
 760     
 761     __RING_STAT_ADD(r, deq_success, n);
 762     r->cons.tail = cons_next;
 763     return behavior == RTE_RING_QUEUE_FIXED ? 0 : n;
 764 }
 731     cons_head = r->cons.head;
 732     prod_tail = r->prod.tail;

上面两行拷贝生产者的tail索引,并计算有多少个entries,这里由于类型是uint32_t,故始终entries >=0
代码行739〜752根据behavior值判断能不能pop出n个元素等;此后增加cons_next,并执行r->cons.head = cons_next(B1),然后执行DEQUEUE_PTRS,是ENQUEUE_PTRS反向操作:

 382 #define DEQUEUE_PTRS() do { \
 383     uint32_t idx = cons_head & mask; \
 384     const uint32_t size = r->cons.size; \
 385     if (likely(idx + n < size)) { \
 386         for (i = 0; i < (n & (~(unsigned)0x3)); i+=4, idx+=4) {\
 387             obj_table[i] = r->ring[idx]; \
 388             obj_table[i+1] = r->ring[idx+1]; \
 389             obj_table[i+2] = r->ring[idx+2]; \
 390             obj_table[i+3] = r->ring[idx+3]; \
 391         } \
 392         switch (n & 0x3) { \
 393             case 3: obj_table[i++] = r->ring[idx++]; \
 394             case 2: obj_table[i++] = r->ring[idx++]; \
 395             case 1: obj_table[i++] = r->ring[idx++]; \
 396         } \
 397     } else { \
 398         for (i = 0; idx < size; i++, idx++) \
 399             obj_table[i] = r->ring[idx]; \
 400         for (idx = 0; i < n; i++, idx++) \
 401             obj_table[i] = r->ring[idx]; \
 402     } \
 403 } while (0)

设置内存读屏障rte_smp_rmb()__sync_synchronize;最近更新r->cons.tail = cons_next(B3);同样,(B2)中间;

简单考虑push和pop数据,假设有生产者A和消费者B两个线程,分别指向ring,有几个问题:B去pop数据的时候会不会因为判断有数据而取不到的情况?(A在push数据时不小心先更新了prod.tail后push数据);A在push数据时是否会因为B先更新cons.tail而导致覆盖B还没来得及pop的数据?还有一些其他情况...

因为这里有内存屏障,并不会出现以上问题,以上列的实现在单生产者和单消费者的情况下比较简单,但对于多读多写一个ring可能比较复杂;

以下分析下复杂情况,即多个线程push和pop数据,这里仅分析生产和消费的源码,其他的一些比较细节且不是很重要的代码不在这里分析。

以下是多生产者实现,有段说明:

 408  * This function uses a "compare and set" instruction to move the
 409  * producer index atomically.
 430 static inline int __attribute__((always_inline))
 431 __rte_ring_mp_do_enqueue(struct rte_ring *r, void * const *obj_table,
 432              unsigned n, enum rte_ring_queue_behavior behavior)
 433 {
 434     uint32_t prod_head, prod_next;
 435     uint32_t cons_tail, free_entries;
 436     const unsigned max = n;
 437     int success;
 438     unsigned i, rep = 0;
 439     uint32_t mask = r->prod.mask;
 440     int ret;
 442     /* Avoid the unnecessary cmpset operation below, which is also
 443      * potentially harmful when n equals 0. */
 444     if (n == 0)
 445         return 0;
 446 
 447     /* move prod.head atomically */
 448     do {
 449         /* Reset n to the initial burst count */
 450         n = max;
 451 
 452         prod_head = r->prod.head;
 453         cons_tail = r->cons.tail;
 454         /* The subtraction is done between two unsigned 32bits value
 455          * (the result is always modulo 32 bits even if we have
 456          * prod_head > cons_tail). So 'free_entries' is always between 0
 457          * and size(ring)-1. */
 458         free_entries = (mask + cons_tail - prod_head);
 459 
 460         /* check that we have enough room in ring */
 461         if (unlikely(n > free_entries)) {
 462             if (behavior == RTE_RING_QUEUE_FIXED) {
 463                 __RING_STAT_ADD(r, enq_fail, n);
 464                 return -ENOBUFS;
 465             }
 466             else {
 467                 /* No free entry available */
 468                 if (unlikely(free_entries == 0)) {
 469                     __RING_STAT_ADD(r, enq_fail, n);
 470                     return 0;
 471                 }
 472 
 473                 n = free_entries;
 474             }
 475         }
 477         prod_next = prod_head + n;
 478         success = rte_atomic32_cmpset(&r->prod.head, prod_head,
 479                           prod_next);
 480     } while (unlikely(success == 0));
 481 
 482     /* write entries in ring */
 483     ENQUEUE_PTRS();
 484     rte_smp_wmb();
 485 
 486     /* if we exceed the watermark */
 487     if (unlikely(((mask + 1) - free_entries + n) > r->prod.watermark)) {
 488         ret = (behavior == RTE_RING_QUEUE_FIXED) ? -EDQUOT :
 489                 (int)(n | RTE_RING_QUOT_EXCEED);
 490         __RING_STAT_ADD(r, enq_quota, n);
 491     }
 492     else {
 493         ret = (behavior == RTE_RING_QUEUE_FIXED) ? 0 : n;
 494         __RING_STAT_ADD(r, enq_success, n);
 495     }
 497     /*
 498      * If there are other enqueues in progress that preceded us,
 499      * we need to wait for them to complete
 500      */
 501     while (unlikely(r->prod.tail != prod_head)) {
 502         rte_pause();
 503 
 504         /* Set RTE_RING_PAUSE_REP_COUNT to avoid spin too long waiting
 505          * for other thread finish. It gives pre-empted thread a chance
 506          * to proceed and finish with ring dequeue operation. */
 507         if (RTE_RING_PAUSE_REP_COUNT &&
 508             ++rep == RTE_RING_PAUSE_REP_COUNT) {
 509             rep = 0;
 510             sched_yield();
 511         }
 512     }
 513     r->prod.tail = prod_next;
 514     return ret;
 515 }

代码行434〜446定义和初始化一些变量,其中对n的判断做了优化,如注释说明;
代码行448〜480就是使用cas指令,尝试更新prod.head,直到成功或者没有足够的空间返回,其他的代码计算free_entries和上面分析的一样;
然后执行入队和内存屏障:

 483     ENQUEUE_PTRS();
 484     rte_smp_wmb();

代码行487〜495和上面的分析一样;
而代码501和512的作用是防止另外的生产者线程也在同时更新prod.tail,为了防止覆盖的情况,这里没有用cas,而是短暂的rte_pause一会避免busy waiting,等其他生产者线程完成后再更新,但是这里有次数rep,当到达RTE_RING_PAUSE_REP_COUNT时,会sched_yield让出执行权,不过是个编译选项;
况且这里使用的是unlikely,可能性比较小;
最后更新prod.tail

至于多消费者的实现,这里不分析了,cas实现差不多和上面的多生产者一样,出队和单消费者实现一样,实现大概如下:

do {
    entries = r->prod.tailprod_tail - r->cons.head;
    cons_next = cons_head + n;
    success = rte_atomic32_cmpset(&r->cons.head, cons_head,cons_next);
} while (unlikely(success == 0))
DEQUEUE_PTRS();
rte_smp_rmb();
while (unlikely(r->cons.tail != cons_head)) {
    rte_pause();
}
r->cons.tail = cons_next;

以上是整体无锁环形队列的实现原理,可能更底层的实现没有说清楚,或者跟硬件有关系,或者跟自己水平有限有关,不管怎样,后面会继续多研究。

另外:
关于cas指令的一些说明,引用DPDK开源社区技术文章中的一段话“当两个core同时执行针对同一地址的CAS指令时,其实他们是在试图修改每个core自己持有的Cache line, 假设两个core都持有相同地址对应cacheline,且各自cacheline 状态为S, 这时如果要想成功修改,就首先需要把S转为E或者M, 则需要向其它core invalidate 这个地址的cacheline,则两个core都会向ring bus 发出 invalidate这个操作, 那么在ringbus上就会根据特定的设计协议仲裁是core0,还是core1能赢得这个invalidate,者完成操作,失败者需要接受结果invalidate自己对应的cacheline,再读取胜者修改后的值,回到起点。

到这里, 我们可以发现MESIF协议大大降低了读操作的时延,没有让写操作更慢,同时保持了一致性。那么对于我们的CAS操作来说,其实锁并没有消失,只是转嫁到了ring bus的总线仲裁协议中。而且大量的多核同时针对一个地址的CAS操作会引起反复的互相invalidate 同一cacheline,造成pingpong效应,同样会降低性能。只能说,基于CAS的操作仍然是不能滥用,不到万不得已不用,通常情况下还是使用数据分离模式更好。”

参考:
https://coolshell.cn/articles/8239.html
https://en.wikipedia.org/wiki/Non-blocking_algorithm
http://in355hz.iteye.com/blog/1797829
https://en.wikipedia.org/wiki/Memory_barrier
https://yq.aliyun.com/articles/95441
http://www.lenky.info/archives/2012/11/2028
https://mp.weixin.qq.com/s?__biz=MzI3NDA4ODY4MA==&mid=2653334228&idx=1&sn=8a106aed154ded89283146ddb6a02cf8&chksm=f0cb5d53c7bcd445704592eb7c06407f1b18f1bf94bed3d33345d2443454b15f9c859b64371c&scene=21#wechat_redirect
http://www.man7.org/linux/man-pages/man2/sched_yield.2.html

你可能感兴趣的:(DPDK中无锁环形队列实现)