*      Now using anonymous inode source.
  *      Thanks to Oleg Nesterov for useful code review and suggestions.
  *      More comments and suggestions from Arnd Bergmann.
- * Sat May 19, 2007: Davi E. M. Arnaut <davi@haxent.com.br>
+ *  Sat May 19, 2007: Davi E. M. Arnaut <davi@haxent.com.br>
  *      Retrieve multiple signals with one read() call
+ *  Sun Jul 15, 2007: Davide Libenzi <davidel@xmailserver.org>
+ *      Attach to the sighand only during read() and poll().
  */
 
 #include <linux/file.h>
 #include <linux/signalfd.h>
 
 struct signalfd_ctx {
-       struct list_head lnk;
-       wait_queue_head_t wqh;
        sigset_t sigmask;
-       struct task_struct *tsk;
 };
 
-struct signalfd_lockctx {
-       struct task_struct *tsk;
-       unsigned long flags;
-};
-
-/*
- * Tries to acquire the sighand lock. We do not increment the sighand
- * use count, and we do not even pin the task struct, so we need to
- * do it inside an RCU read lock, and we must be prepared for the
- * ctx->tsk going to NULL (in signalfd_deliver()), and for the sighand
- * being detached. We return 0 if the sighand has been detached, or
- * 1 if we were able to pin the sighand lock.
- */
-static int signalfd_lock(struct signalfd_ctx *ctx, struct signalfd_lockctx *lk)
-{
-       struct sighand_struct *sighand = NULL;
-
-       rcu_read_lock();
-       lk->tsk = rcu_dereference(ctx->tsk);
-       if (likely(lk->tsk != NULL))
-               sighand = lock_task_sighand(lk->tsk, &lk->flags);
-       rcu_read_unlock();
-
-       if (!sighand)
-               return 0;
-
-       if (!ctx->tsk) {
-               unlock_task_sighand(lk->tsk, &lk->flags);
-               return 0;
-       }
-
-       if (lk->tsk->tgid == current->tgid)
-               lk->tsk = current;
-
-       return 1;
-}
-
-static void signalfd_unlock(struct signalfd_lockctx *lk)
-{
-       unlock_task_sighand(lk->tsk, &lk->flags);
-}
-
-/*
- * This must be called with the sighand lock held.
- */
-void signalfd_deliver(struct task_struct *tsk, int sig)
-{
-       struct sighand_struct *sighand = tsk->sighand;
-       struct signalfd_ctx *ctx, *tmp;
-
-       BUG_ON(!sig);
-       list_for_each_entry_safe(ctx, tmp, &sighand->signalfd_list, lnk) {
-               /*
-                * We use a negative signal value as a way to broadcast that the
-                * sighand has been orphaned, so that we can notify all the
-                * listeners about this. Remember the ctx->sigmask is inverted,
-                * so if the user is interested in a signal, that corresponding
-                * bit will be zero.
-                */
-               if (sig < 0) {
-                       if (ctx->tsk == tsk) {
-                               ctx->tsk = NULL;
-                               list_del_init(&ctx->lnk);
-                               wake_up(&ctx->wqh);
-                       }
-               } else {
-                       if (!sigismember(&ctx->sigmask, sig))
-                               wake_up(&ctx->wqh);
-               }
-       }
-}
-
-static void signalfd_cleanup(struct signalfd_ctx *ctx)
-{
-       struct signalfd_lockctx lk;
-
-       /*
-        * This is tricky. If the sighand is gone, we do not need to remove
-        * context from the list, the list itself won't be there anymore.
-        */
-       if (signalfd_lock(ctx, &lk)) {
-               list_del(&ctx->lnk);
-               signalfd_unlock(&lk);
-       }
-       kfree(ctx);
-}
-
 static int signalfd_release(struct inode *inode, struct file *file)
 {
-       signalfd_cleanup(file->private_data);
+       kfree(file->private_data);
        return 0;
 }
 
 {
        struct signalfd_ctx *ctx = file->private_data;
        unsigned int events = 0;
-       struct signalfd_lockctx lk;
 
-       poll_wait(file, &ctx->wqh, wait);
+       poll_wait(file, ¤t->sighand->signalfd_wqh, wait);
 
-       /*
-        * Let the caller get a POLLIN in this case, ala socket recv() when
-        * the peer disconnects.
-        */
-       if (signalfd_lock(ctx, &lk)) {
-               if ((lk.tsk == current &&
-                    next_signal(&lk.tsk->pending, &ctx->sigmask) > 0) ||
-                   next_signal(&lk.tsk->signal->shared_pending,
-                               &ctx->sigmask) > 0)
-                       events |= POLLIN;
-               signalfd_unlock(&lk);
-       } else
+       spin_lock_irq(¤t->sighand->siglock);
+       if (next_signal(¤t->pending, &ctx->sigmask) ||
+           next_signal(¤t->signal->shared_pending,
+                       &ctx->sigmask))
                events |= POLLIN;
+       spin_unlock_irq(¤t->sighand->siglock);
 
        return events;
 }
                                int nonblock)
 {
        ssize_t ret;
-       struct signalfd_lockctx lk;
        DECLARE_WAITQUEUE(wait, current);
 
-       if (!signalfd_lock(ctx, &lk))
-               return 0;
-
-       ret = dequeue_signal(lk.tsk, &ctx->sigmask, info);
+       spin_lock_irq(¤t->sighand->siglock);
+       ret = dequeue_signal(current, &ctx->sigmask, info);
        switch (ret) {
        case 0:
                if (!nonblock)
                        break;
                ret = -EAGAIN;
        default:
-               signalfd_unlock(&lk);
+               spin_unlock_irq(¤t->sighand->siglock);
                return ret;
        }
 
-       add_wait_queue(&ctx->wqh, &wait);
+       add_wait_queue(¤t->sighand->signalfd_wqh, &wait);
        for (;;) {
                set_current_state(TASK_INTERRUPTIBLE);
-               ret = dequeue_signal(lk.tsk, &ctx->sigmask, info);
-               signalfd_unlock(&lk);
+               ret = dequeue_signal(current, &ctx->sigmask, info);
                if (ret != 0)
                        break;
                if (signal_pending(current)) {
                        ret = -ERESTARTSYS;
                        break;
                }
+               spin_unlock_irq(¤t->sighand->siglock);
                schedule();
-               ret = signalfd_lock(ctx, &lk);
-               if (unlikely(!ret)) {
-                       /*
-                        * Let the caller read zero byte, ala socket
-                        * recv() when the peer disconnect. This test
-                        * must be done before doing a dequeue_signal(),
-                        * because if the sighand has been orphaned,
-                        * the dequeue_signal() call is going to crash
-                        * because ->sighand will be long gone.
-                        */
-                        break;
-               }
+               spin_lock_irq(¤t->sighand->siglock);
        }
+       spin_unlock_irq(¤t->sighand->siglock);
 
-       remove_wait_queue(&ctx->wqh, &wait);
+       remove_wait_queue(¤t->sighand->signalfd_wqh, &wait);
        __set_current_state(TASK_RUNNING);
 
        return ret;
 }
 
 /*
- * Returns either the size of a "struct signalfd_siginfo", or zero if the
- * sighand we are attached to, has been orphaned. The "count" parameter
- * must be at least the size of a "struct signalfd_siginfo".
+ * Returns a multiple of the size of a "struct signalfd_siginfo", or a negative
+ * error code. The "count" parameter must be at least the size of a
+ * "struct signalfd_siginfo".
  */
 static ssize_t signalfd_read(struct file *file, char __user *buf, size_t count,
                             loff_t *ppos)
                return -EINVAL;
 
        siginfo = (struct signalfd_siginfo __user *) buf;
-
        do {
                ret = signalfd_dequeue(ctx, &info, nonblock);
                if (unlikely(ret <= 0))
                nonblock = 1;
        } while (--count);
 
-       return total ? total : ret;
+       return total ? total: ret;
 }
 
 static const struct file_operations signalfd_fops = {
        .read           = signalfd_read,
 };
 
-/*
- * Create a file descriptor that is associated with our signal
- * state. We can pass it around to others if we want to, but
- * it will always be _our_ signal state.
- */
 asmlinkage long sys_signalfd(int ufd, sigset_t __user *user_mask, size_t sizemask)
 {
        int error;
        sigset_t sigmask;
        struct signalfd_ctx *ctx;
-       struct sighand_struct *sighand;
        struct file *file;
        struct inode *inode;
-       struct signalfd_lockctx lk;
 
        if (sizemask != sizeof(sigset_t) ||
            copy_from_user(&sigmask, user_mask, sizeof(sigmask)))
                if (!ctx)
                        return -ENOMEM;
 
-               init_waitqueue_head(&ctx->wqh);
                ctx->sigmask = sigmask;
-               ctx->tsk = current->group_leader;
-
-               sighand = current->sighand;
-               /*
-                * Add this fd to the list of signal listeners.
-                */
-               spin_lock_irq(&sighand->siglock);
-               list_add_tail(&ctx->lnk, &sighand->signalfd_list);
-               spin_unlock_irq(&sighand->siglock);
 
                /*
                 * When we call this, the initialization must be complete, since
                        fput(file);
                        return -EINVAL;
                }
-               /*
-                * We need to be prepared of the fact that the sighand this fd
-                * is attached to, has been detched. In that case signalfd_lock()
-                * will return 0, and we'll just skip setting the new mask.
-                */
-               if (signalfd_lock(ctx, &lk)) {
-                       ctx->sigmask = sigmask;
-                       signalfd_unlock(&lk);
-               }
-               wake_up(&ctx->wqh);
+               spin_lock_irq(¤t->sighand->siglock);
+               ctx->sigmask = sigmask;
+               spin_unlock_irq(¤t->sighand->siglock);
+
+               wake_up(¤t->sighand->signalfd_wqh);
                fput(file);
        }
 
        return ufd;
 
 err_fdalloc:
-       signalfd_cleanup(ctx);
+       kfree(ctx);
        return error;
 }
 
 
 #ifdef CONFIG_SIGNALFD
 
 /*
- * Deliver the signal to listening signalfd. This must be called
- * with the sighand lock held. Same are the following that end up
- * calling signalfd_deliver().
- */
-void signalfd_deliver(struct task_struct *tsk, int sig);
-
-/*
- * No need to fall inside signalfd_deliver() if no signal listeners
- * are available.
+ * Deliver the signal to listening signalfd.
  */
 static inline void signalfd_notify(struct task_struct *tsk, int sig)
 {
-       if (unlikely(!list_empty(&tsk->sighand->signalfd_list)))
-               signalfd_deliver(tsk, sig);
-}
-
-/*
- * The signal -1 is used to notify the signalfd that the sighand
- * is on its way to be detached.
- */
-static inline void signalfd_detach_locked(struct task_struct *tsk)
-{
-       if (unlikely(!list_empty(&tsk->sighand->signalfd_list)))
-               signalfd_deliver(tsk, -1);
-}
-
-static inline void signalfd_detach(struct task_struct *tsk)
-{
-       struct sighand_struct *sighand = tsk->sighand;
-
-       if (unlikely(!list_empty(&sighand->signalfd_list))) {
-               spin_lock_irq(&sighand->siglock);
-               signalfd_deliver(tsk, -1);
-               spin_unlock_irq(&sighand->siglock);
-       }
+       if (unlikely(waitqueue_active(&tsk->sighand->signalfd_wqh)))
+               wake_up(&tsk->sighand->signalfd_wqh);
 }
 
 #else /* CONFIG_SIGNALFD */
 
-#define signalfd_deliver(t, s) do { } while (0)
-#define signalfd_notify(t, s) do { } while (0)
-#define signalfd_detach_locked(t) do { } while (0)
-#define signalfd_detach(t) do { } while (0)
+static inline void signalfd_notify(struct task_struct *tsk, int sig) { }
 
 #endif /* CONFIG_SIGNALFD */