|
最近有朋友在面试的时候被问了select 和epoll效率差的原因,和一般人一样,大部分都会回答select是轮询、epoll是触发式的,所以效率高。这个答案听上去很完美,大致也说出了二者的主要区别。今天闲来无事,翻看了下内核代码,结合内核代码和大家分享下我的观点。一、连接数我本人也曾经在项目中用过select和epoll,对于select,感触最深的是linux下select最大数目限制(windows 下似乎没有限制),每个进程的select最多能处理FD_SETSIZE个FD(文件句柄),如果要处理超过1024个句柄,只能采用多进程了。常见的使用slect的多进程模型是这样的: 一个进程专门accept,成功后将fd通过unix socket传递给子进程处理,父进程可以根据子进程负载分派。曾经用过1个父进程+4个子进程 承载了超过4000个的负载。这种模型在我们当时的业务运行的非常好。epoll在连接数方面没有限制,当然可能需要用户调用API重现设置进程的资源限制。二、IO差别1、select的实现这段可以结合linux内核代码描述了,我使用的是2.6.28,其他2.6的代码应该差不多吧。先看看select:select系统调用的代码在fs/Select.c下,asmlinkage long sys_select(int n, fd_set __user *inp, fd_set __user *outp, fd_set __user *exp, struct timeval __user *tvp){ struct timespec end_time, *to = NULL; struct timeval tv; int ret; if (tvp) { if (copy_from_user(&tv, tvp, sizeof(tv))) return -EFAULT; to = &end_time; if (poll_select_set_timeout(to, tv.tv_sec + (tv.tv_usec / USEC_PER_SEC), (tv.tv_usec % USEC_PER_SEC) * NSEC_PER_USEC)) return -EINVAL; } ret = core_sys_select(n, inp, outp, exp, to); ret = poll_select_copy_remaining(&end_time, tvp, 1, ret); return ret;} 前面是从用户控件拷贝各个fd_set到内核空间,接下来的具体工作在core_sys_select中,core_sys_select->do_select,真正的核心内容在do_select里:int do_select(int n, fd_set_bits *fds, struct timespec *end_time){ ktime_t expire, *to = NULL; struct poll_wqueues table; poll_table *wait; int retval, i, timed_out = 0; unsigned long slack = 0; rcu_read_lock(); retval = max_select_fd(n, fds); rcu_read_unlock(); if (retval < 0) return retval; n = retval; poll_initwait(&table); wait = &table.pt; if (end_time && !end_time->tv_sec && !end_time->tv_nsec) { wait = NULL; timed_out = 1; } if (end_time && !timed_out) slack = estimate_accuracy(end_time); retval = 0; for (;;) { unsigned long *rinp, *routp, *rexp, *inp, *outp, *exp; set_current_state(TASK_INTERRUPTIBLE); inp = fds->in; outp = fds->out; exp = fds->ex; rinp = fds->res_in; routp = fds->res_out; rexp = fds->res_ex; for (i = 0; i < n; ++rinp, ++routp, ++rexp) { unsigned long in, out, ex, all_bits, bit = 1, mask, j; unsigned long res_in = 0, res_out = 0, res_ex = 0; const struct file_operations *f_op = NULL; struct file *file = NULL; in = *inp++; out = *outp++; ex = *exp++; all_bits = in | out | ex; if (all_bits == 0) { i += __NFDBITS; continue; } for (j = 0; j < __NFDBITS; ++j, ++i, bit <<= 1) { int fput_needed; if (i >= n) break; if (!(bit & all_bits)) continue; file = fget_light(i, &fput_needed); if (file) { f_op = file->f_op; mask = DEFAULT_POLLMASK; if (f_op && f_op->poll) mask = (*f_op->poll)(file, retval ? NULL : wait); fput_light(file, fput_needed); if ((mask & POLLIN_SET) && (in & bit)) { res_in |= bit; retval++; } if ((mask & POLLOUT_SET) && (out & bit)) { res_out |= bit; retval++; } if ((mask & POLLEX_SET) && (ex & bit)) { res_ex |= bit; retval++; } } } if (res_in) *rinp = res_in; if (res_out) *routp = res_out; if (res_ex) *rexp = res_ex; cond_resched(); } wait = NULL; if (retval || timed_out || signal_pending(current)) break; if (table.error) { retval = table.error; break; } /* * If this is the first loop and we have a timeout * given, then we convert to ktime_t and set the to * pointer to the expiry value. */ if (end_time && !to) { expire = timespec_to_ktime(*end_time); to = &expire; } if (!schedule_hrtimeout_range(to, slack, HRTIMER_MODE_ABS)) timed_out = 1; } __set_current_state(TASK_RUNNING); poll_freewait(&table); return retval;} 上面的代码很多,其实真正关键的代码是这一句:mask = (*f_op->poll)(file, retval ? NULL : wait); 这个是调用文件系统的 poll函数,不同的文件系统poll函数自然不同,由于我们这里关注的是tcp连接,而socketfs的注册在 net/Socket.c里。register_filesystem(&sock_fs_type); socket文件系统的函数也是在net/Socket.c里:static const struct file_operations socket_file_ops = { .owner = THIS_MODULE, .llseek = no_llseek, .aio_read = sock_aio_read, .aio_write = sock_aio_write, .poll = sock_poll, .unlocked_ioctl = sock_ioctl,#ifdef CONFIG_COMPAT .compat_ioctl = compat_sock_ioctl,#endif .mmap = sock_mmap, .open = sock_no_open, /* special open code to disallow open via /proc */ .release = sock_close, .fasync = sock_fasync, .sendpage = sock_sendpage, .splice_write = generic_splice_sendpage, .splice_read = sock_splice_read,};从sock_poll跟随下去,最后可以到 net/ipv4/tcp.c的unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait) 这个是最终的查询函数,也就是说select 的核心功能是调用tcp文件系统的poll函数,不停的查询,如果没有想要的数据,主动执行一次调度(防止一直占用cpu),直到有一个连接有想要的消息为止。从这里可以看出select的执行方式基本就是不同的调用poll,直到有需要的消息为止,如果select 处理的socket很多,这其实对整个机器的性能也是一个消耗。2、epoll的实现epoll的实现代码在 fs/EventPoll.c下,由于epoll涉及到几个系统调用,这里不逐个分析了,仅仅分析几个关键点,第一个关键点在static int ep_insert(struct eventpoll *ep, struct epoll_event *event, struct file *tfile, int fd) 这是在我们调用sys_epoll_ctl 添加一个被管理socket的时候调用的函数,关键的几行如下:epq.epi = epi; init_poll_funcptr(&epq.pt, ep_ptable_queue_proc); /* * Attach the item to the poll hooks and get current event bits. * We can safely use the file* here because its usage count has * been increased by the caller of this function. Note that after * this operation completes, the poll callback can start hitting * the new item. */ revents = tfile->f_op->poll(tfile, &epq.pt); 这里也是调用文件系统的poll函数,不过这次初始化了一个结构,这个结构会带有一个poll函数的callback函数:ep_ptable_queue_proc,在调用poll函数的时候,会执行这个callback,这个callback的功能就是将当前进程添加到 socket的等待进程上。static void ep_ptable_queue_proc(struct file *file, wait_queue_head_t *whead, poll_table *pt){ struct epitem *epi = ep_item_from_epqueue(pt); struct eppoll_entry *pwq; if (epi->nwait >= 0 && (pwq = kmem_cache_alloc(pwq_cache, GFP_KERNEL))) { init_waitqueue_func_entry(&pwq->wait, ep_poll_callback); pwq->whead = whead; pwq->base = epi; add_wait_queue(whead, &pwq->wait); list_add_tail(&pwq->llink, &epi->pwqlist); epi->nwait++; } else { /* We have to signal that an error occurred */ epi->nwait = -1; }} 注意到参数 whead 实际上是 sk->sleep,其实就是将当前进程添加到sk的等待队列里,当该socket收到数据或者其他事件触发时,会调用sock_def_readable 或者sock_def_write_space 通知函数来唤醒等待进程,这2个函数都是在socket创建的时候填充在sk结构里的。从前面的分析来看,epoll确实是比select聪明的多、轻松的多,不用再苦哈哈的去轮询了。
|