Change async-io to use a pipe to signal I/O completetion instead of

relying on rapid polling.

The actual completetion events are still polled. Only the notification
that there is completed events is done via the pipe to have select/poll
react in a timely fashion.

diff --git a/src/fs/aufs/aiops.cc b/src/fs/aufs/aiops.cc
index 7929d6803b755b9fea3567016cc95c81c743b497..19e94e7b591cb588a18fbe0bb8f9becae0e8564f 100644 (file)
--- a/src/fs/aufs/aiops.cc
+++ b/src/fs/aufs/aiops.cc
@@ -1,5 +1,5 @@
 /*
- * $Id: aiops.cc,v 1.18 2002/11/10 02:29:58 hno Exp $
+ * $Id: aiops.cc,v 1.19 2003/01/04 01:28:13 hno Exp $
  *
  * DEBUG: section 43    AIOPS
  * AUTHOR: Stewart Forster <slf@connect.com.au>
@@ -146,6 +146,8 @@ static struct {
 
     NULL, &done_requests.head
 };
+static int done_fd = 0;
+static int done_signalled = 0;
 static pthread_attr_t globattr;
 #if HAVE_SCHED_H
 static struct sched_param globsched;
@@ -221,10 +223,20 @@ squidaio_xstrfree(char *str)
        xfree(str);
 }
 
+static void
+squidaio_fdhandler(int fd, void *data)
+{
+    char buf[256];
+    done_signalled = 0;
+    read(fd, buf, sizeof(buf));
+    commSetSelect(fd, COMM_SELECT_READ, squidaio_fdhandler, NULL, 0);
+}
+
 static void
 squidaio_init(void)
 {
     int i;
+    int done_pipe[2];
     squidaio_thread_t *threadp;
 
     if (squidaio_initialised)
@@ -268,6 +280,15 @@ squidaio_init(void)
     done_queue.requests = 0;
     done_queue.blocked = 0;
 
+    /* Initialize done pipe signal */
+    pipe(done_pipe);
+    done_fd = done_pipe[1];
+    fd_open(done_pipe[0], FD_PIPE, "async-io completetion event: main");
+    fd_open(done_pipe[1], FD_PIPE, "async-io completetion event: threads");
+    commSetNonBlocking(done_pipe[0]);
+    commSetNonBlocking(done_pipe[1]);
+    commSetSelect(done_pipe[0], COMM_SELECT_READ, squidaio_fdhandler, NULL, 0);
+
     /* Create threads and get them to sit in their wait loop */
     squidaio_thread_pool = memPoolCreate("aio_thread", sizeof(squidaio_thread_t));
     assert(NUMTHREADS);
@@ -389,6 +410,10 @@ squidaio_thread_loop(void *ptr)
        *done_queue.tailp = request;
        done_queue.tailp = &request->next;
        pthread_mutex_unlock(&done_queue.mutex);
+       if (!done_signalled) {
+           done_signalled = 1;
+           write(done_fd, "!", 1);
+       }
        threadp->requests++;
     }                          /* while forever */
     return NULL;
@@ -790,19 +815,6 @@ squidaio_poll_queues(void)
        }
        done_requests.tailp = &requests->next;
     }
-#if HAVE_SCHED_H
-    /* Give up the CPU to allow the threads to do their work */
-    /*
-     * For Andres thoughts about yield(), see
-     * http://www.squid-cache.org/mail-archive/squid-dev/200012/0001.html
-     */
-    if (done_queue.head || request_queue.head)
-#ifndef _SQUID_SOLARIS_
-       sched_yield();
-#else
-       yield();
-#endif
-#endif
 }
 
 squidaio_result_t *

diff --git a/src/fs/aufs/async_io.cc b/src/fs/aufs/async_io.cc
index 3f4fdf8ad1962f8dc5c6f34e46d49d611d4d58e6..0733881745a40a1626eef85676de903b186ac4ab 100644 (file)
--- a/src/fs/aufs/async_io.cc
+++ b/src/fs/aufs/async_io.cc
@@ -1,6 +1,6 @@
 
 /*
- * $Id: async_io.cc,v 1.20 2002/12/27 10:26:35 robertc Exp $
+ * $Id: async_io.cc,v 1.21 2003/01/04 01:28:13 hno Exp $
  *
  * DEBUG: section 32    Asynchronous Disk I/O
  * AUTHOR: Pete Bentley <pete@demon.net>
@@ -95,6 +95,7 @@ aioFDWasClosed(int fd)
        fd_close(fd);
 }
 
+
 void
 aioInit(void)
 {
@@ -104,7 +105,6 @@ aioInit(void)
     cachemgrRegister("squidaio_counts", "Async IO Function Counters",
        aioStats, 0, 1);
     initialised = 1;
-    comm_quick_poll_required();
 }
 
 void