基于select、poll、epoll的区别详解

int select(int nfds, fd_set *readfds, fd_set *writefds, fd_set *exceptfds, struct timeval *timeout);
int poll(struct pollfd *fds, nfds_t nfds, int timeout);
int epoll_wait(int epfd, struct epoll_event *events, int maxevents, int timeout);
 

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;
}

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;
}

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,
};

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);

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;
    }
}