首先从一起多线程无锁算法的事故说起,以下是一个无锁栈的实现测试,但在开-O2以上优化的情况下它却无法正常工作:
#include "lf_stack.h"
#include "kn_list.h"
#include "kn_time.h"
#include "kn_thread.h"
#include "kn_atomic.h"typedef struct lockfree_stack
{volatile kn_list_node *head;
}lockfree_stack,*lockfree_stack_t;static void lfstack_push(lockfree_stack_t ls,kn_list_node *n)
{for( ; ;){ kn_list_node *lhead = ls->head;n->next = lhead;if(COMPARE_AND_SWAP(&ls->head,lhead,n))//if head unchange,set n to be the new headbreak;}
}static kn_list_node* lfstack_pop(lockfree_stack_t ls)
{for( ; ;){ kn_list_node *lhead = ls->head;if(!lhead) return NULL;kn_list_node *next = lhead->next; if(COMPARE_AND_SWAP(&ls->head,lhead,next)){lhead->next = NULL;return lhead;}}
}volatile int count = 0;
atomic_32_t c1 = 0;
atomic_32_t c2 = 0;struct element{kn_list_node node;int value;
};struct element *element_pool1;
struct element *element_pool2;lockfree_stack lf_stack;void *producer1(void *arg)
{printf("producer1\n");int i;while(1){for(i = 0; i < 10000000; ++i){struct element *ele = &element_pool1[i];ATOMIC_INCREASE(&c1);lfstack_push(&lf_stack,(kn_list_node*)ele);}while(c1 > 0){FENCE();kn_sleepms(0);}}printf("producer1 end\n");return NULL;
}void *producer2(void *arg)
{printf("producer2\n");int i;while(1){for(i = 0; i < 10000000; ++i){struct element *ele = &element_pool2[i];ATOMIC_INCREASE(&c2); lfstack_push(&lf_stack,(kn_list_node*)ele);}while(c2 > 0){FENCE();kn_sleepms(0);}}return NULL;
}void *consumer(void *arg)
{printf("consumer\n");volatile struct element *ele;uint32_t tick = 0;while(1){if((ele = (struct element*)lfstack_pop(&lf_stack))){if(count == 0){tick = kn_systemms();}if(++count == 5000000) {uint32_t now = kn_systemms();uint32_t elapse = (uint32_t)(now-tick);printf("pop %d/ms\n",count/elapse*1000);tick = now;count = 0; }if(ele->value == 1)ATOMIC_DECREASE(&c1);else if(ele->value == 2)ATOMIC_DECREASE(&c2);elseprintf("%d\n",ele->value); }}return NULL;
}int main(){element_pool1 = calloc(10000000,sizeof(*element_pool1));element_pool2 = calloc(10000000,sizeof(*element_pool2));int i;for(i = 0; i < 10000000; ++i) element_pool1[i].value = 1;for(i = 0; i < 10000000; ++i) element_pool2[i].value = 2; lf_stack.head = NULL;kn_thread_t t1 = kn_create_thread(THREAD_JOINABLE);kn_thread_t t2 = kn_create_thread(THREAD_JOINABLE); kn_thread_t t3 = kn_create_thread(THREAD_JOINABLE); kn_thread_start_run(t1,producer1,NULL);kn_thread_start_run(t2,producer2,NULL); kn_thread_start_run(t3,consumer,NULL); getchar();return 0;
}
表现就是consumer执行一定次数的pop之后死活也无法再弹出元素.不知道各位看官看出问题在哪没有.
当问题再次出现以后,我用调试器上去中断,consumer线程,断点正好在这行if(!lhead) return NULL;
,lhead为NULL,我回到上一层栈查看实际上lockfree_stack.head
字段并不是空,当我想查看lhead的地址时,显示无法查看寄存器地址.
那么问题就明确了,编译器把lhead存放在了寄存器,导致无法发现head实际已经被改变.那么问题来了,我明明将head标记为volatile的呀.
可是再仔细看看,volatile kn_list_node *head;
修饰符在指针之前,意思是指向的是volatile变量,而我实际要的是,一个指针它本身是volatile的.ok,做相应的调整后kn_list_node * volatile head;
,问题解决.
总之,对指针修饰符关键的一点就是,在*
之前,修饰的是指向的目标.而在*
之后才是修饰指针本身.