Certain data accesses in the kernel should have volatile semantics to be correct and not rely on the whims of the compiler. Change queue clearing to simply catch read up to write rather than reset both to 0 to ensure sane results for queue_count and queue_empty with concurrency. Binsize may or may not increase a bit depending upon whether the output was as intended in all places; wrong stuff was already unlikely to cause any issue.

git-svn-id: svn://svn.rockbox.org/rockbox/trunk@28909 a1c6a512-1295-4272-9138-f99709370657

3 files changed, 54 insertions, 38 deletions

diff --git a/firmware/export/kernel.h b/firmware/export/kernel.h
index c7fcd93284..4656d87fb2 100644
--- a/firmware/export/kernel.h
+++ b/firmware/export/kernel.h
@@ -106,7 +106,7 @@ struct queue_sender_list
     struct thread_entry *senders[QUEUE_LENGTH]; /* message->thread map */
     struct thread_entry *list;                  /* list of senders in map */
     /* Send info for last message dequeued or NULL if replied or not sent */
-    struct thread_entry *curr_sender;
+    struct thread_entry * volatile curr_sender;
 #ifdef HAVE_PRIORITY_SCHEDULING
     struct blocker blocker;
 #endif
@@ -126,10 +126,10 @@ struct event_queue
 {
     struct thread_entry *queue;         /* waiter list */
     struct queue_event events[QUEUE_LENGTH]; /* list of events */
-    unsigned int read;                  /* head of queue */
-    unsigned int write;                 /* tail of queue */
+    unsigned int volatile read;         /* head of queue */
+    unsigned int volatile write;        /* tail of queue */
 #ifdef HAVE_EXTENDED_MESSAGING_AND_NAME
-    struct queue_sender_list *send;     /* list of threads waiting for
+    struct queue_sender_list * volatile send; /* list of threads waiting for
                                            reply to an event */
 #ifdef HAVE_PRIORITY_SCHEDULING
     struct blocker *blocker_p;          /* priority inheritance info
@@ -171,7 +171,7 @@ struct semaphore
 struct wakeup
 {
     struct thread_entry *queue;         /* waiter list */
-    bool signalled;                     /* signalled status */
+    bool volatile signalled;            /* signalled status */
     IF_COP( struct corelock cl; )       /* multiprocessor sync */
 };
 #endif
diff --git a/firmware/export/thread.h b/firmware/export/thread.h
index 87c2d2d709..ba777dc3d1 100644
--- a/firmware/export/thread.h
+++ b/firmware/export/thread.h
@@ -269,7 +269,7 @@ struct thread_entry
     /* Only enabled when using queue_send for now */
 #endif
 #if defined(HAVE_EXTENDED_MESSAGING_AND_NAME) || NUM_CORES > 1
-    intptr_t retval;           /* Return value from a blocked operation/
+    volatile intptr_t retval;  /* Return value from a blocked operation/
                                   misc. use */
 #endif
 #ifdef HAVE_PRIORITY_SCHEDULING
diff --git a/firmware/kernel.c b/firmware/kernel.c
index aaa675241b..41d1d00689 100644
--- a/firmware/kernel.c
+++ b/firmware/kernel.c
@@ -274,7 +274,7 @@ void yield(void)
  * more efficent to reject the majority of cases that don't need this
  * called.
  */
-static void queue_release_sender(struct thread_entry **sender,
+static void queue_release_sender(struct thread_entry * volatile * sender,
                                  intptr_t retval)
 {
     struct thread_entry *thread = *sender;
@@ -427,8 +427,11 @@ void queue_init(struct event_queue *q, bool register_queue)
 
     corelock_init(&q->cl);
     q->queue = NULL;
-    q->read = 0;
-    q->write = 0;
+    /* What garbage is in write is irrelevant because of the masking design-
+     * any other functions the empty the queue do this as well so that
+     * queue_count and queue_empty return sane values in the case of a
+     * concurrent change without locking inside them. */
+    q->read = q->write;
 #ifdef HAVE_EXTENDED_MESSAGING_AND_NAME
     q->send = NULL; /* No message sending by default */
     IF_PRIO( q->blocker_p = NULL; )
@@ -484,8 +487,7 @@ void queue_delete(struct event_queue *q)
     }
 #endif
 
-    q->read = 0;
-    q->write = 0;
+    q->read = q->write;
 
     corelock_unlock(&q->cl);
     restore_irq(oldlevel);
@@ -510,29 +512,31 @@ void queue_wait(struct event_queue *q, struct queue_event *ev)
 
     /* auto-reply */
     queue_do_auto_reply(q->send);
-    
-    if (q->read == q->write)
+
+    while(1)
     {
-        struct thread_entry *current = thread_id_entry(THREAD_ID_CURRENT);
+        struct thread_entry *current;
 
-        do
-        {
-            IF_COP( current->obj_cl = &q->cl; )
-            current->bqp = &q->queue;
+        rd = q->read;
+        if (rd != q->write) /* A waking message could disappear */
+            break;
 
-            block_thread(current);
+        current = thread_id_entry(THREAD_ID_CURRENT);
 
-            corelock_unlock(&q->cl);
-            switch_thread();
+        IF_COP( current->obj_cl = &q->cl; )
+        current->bqp = &q->queue;
 
-            oldlevel = disable_irq_save();
-            corelock_lock(&q->cl);
-        }
-        /* A message that woke us could now be gone */
-        while (q->read == q->write);
+        block_thread(current);
+
+        corelock_unlock(&q->cl);
+        switch_thread();
+
+        oldlevel = disable_irq_save();
+        corelock_lock(&q->cl);
     } 
 
-    rd = q->read++ & QUEUE_LENGTH_MASK;
+    q->read = rd + 1;
+    rd &= QUEUE_LENGTH_MASK;
     *ev = q->events[rd];
 
     /* Get data for a waiting thread if one */
@@ -545,6 +549,7 @@ void queue_wait(struct event_queue *q, struct queue_event *ev)
 void queue_wait_w_tmo(struct event_queue *q, struct queue_event *ev, int ticks)
 {
     int oldlevel;
+    unsigned int rd, wr;
 
 #ifdef HAVE_EXTENDED_MESSAGING_AND_NAME
     KERNEL_ASSERT(QUEUE_GET_THREAD(q) == NULL ||
@@ -558,7 +563,9 @@ void queue_wait_w_tmo(struct event_queue *q, struct queue_event *ev, int ticks)
     /* Auto-reply */
     queue_do_auto_reply(q->send);
 
-    if (q->read == q->write && ticks > 0)
+    rd = q->read;
+    wr = q->write;
+    if (rd == wr && ticks > 0)
     {
         struct thread_entry *current = thread_id_entry(THREAD_ID_CURRENT);
 
@@ -572,13 +579,17 @@ void queue_wait_w_tmo(struct event_queue *q, struct queue_event *ev, int ticks)
 
         oldlevel = disable_irq_save();
         corelock_lock(&q->cl);
+
+        rd = q->read;
+        wr = q->write;
     }
 
     /* no worry about a removed message here - status is checked inside
        locks - perhaps verify if timeout or false alarm */
-    if (q->read != q->write)
+    if (rd != wr)
     {
-        unsigned int rd = q->read++ & QUEUE_LENGTH_MASK;
+        q->read = rd + 1;
+        rd &= QUEUE_LENGTH_MASK;
         *ev = q->events[rd];
         /* Get data for a waiting thread if one */
         queue_do_fetch_sender(q->send, rd);
@@ -700,13 +711,16 @@ void queue_reply(struct event_queue *q, intptr_t retval)
 {
     if(q->send && q->send->curr_sender)
     {
+        struct queue_sender_list *sender;
+
         int oldlevel = disable_irq_save();
         corelock_lock(&q->cl);
+
+        sender = q->send;
+
         /* Double-check locking */
-        IF_COP( if(LIKELY(q->send && q->send->curr_sender)) )
-        {
-            queue_release_sender(&q->send->curr_sender, retval);
-        }
+        if(LIKELY(sender && sender->curr_sender))
+            queue_release_sender(&sender->curr_sender, retval);
 
         corelock_unlock(&q->cl);
         restore_irq(oldlevel);
@@ -716,6 +730,8 @@ void queue_reply(struct event_queue *q, intptr_t retval)
 
 bool queue_peek(struct event_queue *q, struct queue_event *ev)
 {
+    unsigned int rd;
+
     if(q->read == q->write)
          return false;
 
@@ -724,9 +740,10 @@ bool queue_peek(struct event_queue *q, struct queue_event *ev)
     int oldlevel = disable_irq_save();
     corelock_lock(&q->cl);
 
-    if(q->read != q->write)
+    rd = q->read;
+    if(rd != q->write)
     {
-        *ev = q->events[q->read & QUEUE_LENGTH_MASK];
+        *ev = q->events[rd & QUEUE_LENGTH_MASK];
         have_msg = true;
     }
 
@@ -756,8 +773,7 @@ void queue_clear(struct event_queue* q)
        dequeued sent message will be handled by owning thread */
     queue_release_all_senders(q);
 
-    q->read = 0;
-    q->write = 0;
+    q->read = q->write;
 
     corelock_unlock(&q->cl);
     restore_irq(oldlevel);