signalfd: retrieve multiple signals with one read() call
authorDavi Arnaut <davi@haxent.com.br>
Wed, 23 May 2007 20:58:04 +0000 (13:58 -0700)
committerLinus Torvalds <torvalds@woody.linux-foundation.org>
Thu, 24 May 2007 03:14:14 +0000 (20:14 -0700)
Gathering signals in bulk enables server applications to drain a signal
queue (almost full of realtime signals) more efficiently by reducing the
syscall and file look-up overhead.

Very similar to the sigtimedwait4() call described by Niels Provos, Chuck
Lever, and Stephen Tweedie in a paper entitled "Analyzing the Overload
Behavior of a Simple Web Server".  The paper lists more details and
advantages.

Signed-off-by: Davi E. M. Arnaut <davi@haxent.com.br>
Acked-by: Davide Libenzi <davidel@xmailserver.org>
Cc: Oleg Nesterov <oleg@tv-sign.ru>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
fs/signalfd.c

index 7cfeab412b45ec4ef381caef24a721e966a4f3ad..f1da89203a9aeea3108fe9270169c501df821129 100644 (file)
@@ -11,6 +11,8 @@
  *      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>
+ *      Retrieve multiple signals with one read() call
  */
 
 #include <linux/file.h>
@@ -206,6 +208,59 @@ static int signalfd_copyinfo(struct signalfd_siginfo __user *uinfo,
        return err ? -EFAULT: sizeof(*uinfo);
 }
 
+static ssize_t signalfd_dequeue(struct signalfd_ctx *ctx, siginfo_t *info,
+                               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);
+       switch (ret) {
+       case 0:
+               if (!nonblock)
+                       break;
+               ret = -EAGAIN;
+       default:
+               signalfd_unlock(&lk);
+               return ret;
+       }
+
+       add_wait_queue(&ctx->wqh, &wait);
+       for (;;) {
+               set_current_state(TASK_INTERRUPTIBLE);
+               ret = dequeue_signal(lk.tsk, &ctx->sigmask, info);
+               signalfd_unlock(&lk);
+               if (ret != 0)
+                       break;
+               if (signal_pending(current)) {
+                       ret = -ERESTARTSYS;
+                       break;
+               }
+               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;
+               }
+       }
+
+       remove_wait_queue(&ctx->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
@@ -215,55 +270,30 @@ static ssize_t signalfd_read(struct file *file, char __user *buf, size_t count,
                             loff_t *ppos)
 {
        struct signalfd_ctx *ctx = file->private_data;
-       ssize_t res = 0;
-       int locked, signo;
+       struct signalfd_siginfo __user *siginfo;
+       int nonblock = file->f_flags & O_NONBLOCK;
+       ssize_t ret, total = 0;
        siginfo_t info;
-       struct signalfd_lockctx lk;
-       DECLARE_WAITQUEUE(wait, current);
 
-       if (count < sizeof(struct signalfd_siginfo))
+       count /= sizeof(struct signalfd_siginfo);
+       if (!count)
                return -EINVAL;
-       locked = signalfd_lock(ctx, &lk);
-       if (!locked)
-               return 0;
-       res = -EAGAIN;
-       signo = dequeue_signal(lk.tsk, &ctx->sigmask, &info);
-       if (signo == 0 && !(file->f_flags & O_NONBLOCK)) {
-               add_wait_queue(&ctx->wqh, &wait);
-               for (;;) {
-                       set_current_state(TASK_INTERRUPTIBLE);
-                       signo = dequeue_signal(lk.tsk, &ctx->sigmask, &info);
-                       if (signo != 0)
-                               break;
-                       if (signal_pending(current)) {
-                               res = -ERESTARTSYS;
-                               break;
-                       }
-                       signalfd_unlock(&lk);
-                       schedule();
-                       locked = signalfd_lock(ctx, &lk);
-                       if (unlikely(!locked)) {
-                               /*
-                                * 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.
-                                */
-                               res = 0;
-                               break;
-                       }
-               }
-               remove_wait_queue(&ctx->wqh, &wait);
-               __set_current_state(TASK_RUNNING);
-       }
-       if (likely(locked))
-               signalfd_unlock(&lk);
-       if (likely(signo))
-               res = signalfd_copyinfo((struct signalfd_siginfo __user *) buf,
-                                       &info);
 
-       return res;
+       siginfo = (struct signalfd_siginfo __user *) buf;
+
+       do {
+               ret = signalfd_dequeue(ctx, &info, nonblock);
+               if (unlikely(ret <= 0))
+                       break;
+               ret = signalfd_copyinfo(siginfo, &info);
+               if (ret < 0)
+                       break;
+               siginfo++;
+               total += ret;
+               nonblock = 1;
+       } while (--count);
+
+       return total ? total : ret;
 }
 
 static const struct file_operations signalfd_fops = {