refactored gc_jd_queue_data, faster mutex_unlock

git-svn-id: http://gnuradio.org/svn/gnuradio/trunk@8558 221aa14e-8319-0410-a670-987f0aec2ac5

3 files changed, 70 insertions, 29 deletions

diff --git a/gcell/src/include/gc_jd_queue_data.h b/gcell/src/include/gc_jd_queue_data.h
index d48591bd2..e5fa87499 100644
--- a/gcell/src/include/gc_jd_queue_data.h
+++ b/gcell/src/include/gc_jd_queue_data.h
@@ -1,6 +1,6 @@
 /* -*- c++ -*- */
 /*
- * Copyright 2007 Free Software Foundation, Inc.
+ * Copyright 2007,2008 Free Software Foundation, Inc.
  * 
  * This file is part of GNU Radio
  * 
@@ -36,13 +36,31 @@ __GC_BEGIN_DECLS
  *
  * FIXME make it lock free ;)  For now, use a spin lock.
  */
-typedef struct gc_jd_queue
+
+typedef struct gc_jd_q_links
 {
   gc_eaddr_t	head _AL16;
   gc_eaddr_t	tail _AL16;
+} gc_jd_q_links_t;
+
+typedef struct gc_jd_q_mutex
+{
   uint32_t	mutex;		// libsync mutex (spin lock)
-} gc_jd_queue_t;
+  uint32_t	_pad[31];	// pad to cache line so we can use putlluc on SPE
+} _AL128 gc_jd_q_mutex_t;
 
+typedef struct gc_jd_q_flag
+{
+  uint32_t	flag;		// host writes this after enqueuing
+  uint32_t	_pad[31];	// pad to cache line
+} _AL128 gc_jd_q_flag_t;
+
+typedef struct gc_jd_queue
+{
+  gc_jd_q_links_t	l;
+  gc_jd_q_mutex_t	m;
+  gc_jd_q_flag_t	f;
+} _AL128 gc_jd_queue_t;
 
 __GC_END_DECLS
 
diff --git a/gcell/src/lib/runtime/gc_jd_queue.c b/gcell/src/lib/runtime/gc_jd_queue.c
index b5cdcac9b..29b74c29d 100644
--- a/gcell/src/lib/runtime/gc_jd_queue.c
+++ b/gcell/src/lib/runtime/gc_jd_queue.c
@@ -28,9 +28,10 @@
 void 
 gc_jd_queue_init(gc_jd_queue_t *q)
 {
-  _mutex_init(ptr_to_ea(&q->mutex));
-  q->head = 0;
-  q->tail = 0;
+  _mutex_init(ptr_to_ea(&q->m.mutex));
+  q->l.head = 0;
+  q->l.tail = 0;
+  q->f.flag = 0;
   smp_wmb();
 }
   
@@ -38,41 +39,44 @@ void
 gc_jd_queue_enqueue(gc_jd_queue_t *q, gc_job_desc_t *item)
 {
   item->sys.next = 0;
-  _mutex_lock(ptr_to_ea(&q->mutex));
+  _mutex_lock(ptr_to_ea(&q->m.mutex));
   smp_rmb();		// import barrier
 
-  if (q->tail == 0){    // currently empty
-    q->tail = q->head = jdp_to_ea(item);
+  if (q->l.tail == 0){    // currently empty
+    q->l.tail = q->l.head = jdp_to_ea(item);
   }
   else {		// not empty, append
-    ea_to_jdp(q->tail)->sys.next = jdp_to_ea(item);
-    q->tail = jdp_to_ea(item);
+    ea_to_jdp(q->l.tail)->sys.next = jdp_to_ea(item);
+    q->l.tail = jdp_to_ea(item);
   }
 
   smp_wmb();		// orders stores above before clearing of mutex
-  _mutex_unlock(ptr_to_ea(&q->mutex));
+  _mutex_unlock(ptr_to_ea(&q->m.mutex));
+
+  // let SPE's know we wrote something if they've got a lock-line reservation
+  q->f.flag = 1;
 }
 
 gc_job_desc_t *
 gc_jd_queue_dequeue(gc_jd_queue_t *q)
 {
-  _mutex_lock(ptr_to_ea(&q->mutex));
+  _mutex_lock(ptr_to_ea(&q->m.mutex));
   smp_rmb();		// import barrier
   
-  gc_eaddr_t item_ea = q->head;
+  gc_eaddr_t item_ea = q->l.head;
   if (item_ea == 0){	// empty
-    _mutex_unlock(ptr_to_ea(&q->mutex));
+    _mutex_unlock(ptr_to_ea(&q->m.mutex));
     return 0;
   }
 
-  q->head = ea_to_jdp(item_ea)->sys.next;
-  if (q->head == 0)	// now emtpy
-    q->tail = 0;
+  q->l.head = ea_to_jdp(item_ea)->sys.next;
+  if (q->l.head == 0)	// now emtpy
+    q->l.tail = 0;
 
   gc_job_desc_t *item = ea_to_jdp(item_ea);
   item->sys.next = 0;
 
   smp_wmb();		// orders stores above before clearing of mutex
-  _mutex_unlock(ptr_to_ea(&q->mutex));
+  _mutex_unlock(ptr_to_ea(&q->m.mutex));
   return item;
 }
diff --git a/gcell/src/lib/runtime/spu/gc_spu_jd_queue.c b/gcell/src/lib/runtime/spu/gc_spu_jd_queue.c
index 22752fe68..0dd165fc0 100644
--- a/gcell/src/lib/runtime/spu/gc_spu_jd_queue.c
+++ b/gcell/src/lib/runtime/spu/gc_spu_jd_queue.c
@@ -25,11 +25,30 @@
 
 extern int gc_sys_tag;
 
+/*
+ * ea must be 128-byte aligned, the mutex is in the first int32_t, and
+ * it must be safe to write the remaining 124 bytes with anything at
+ * all.
+ */
+static __inline void _fast_mutex_unlock(mutex_ea_t ea)
+{
+  char _tmp[256];
+  vector signed int *buf
+    = (vector signed int *) ALIGN(_tmp, 128);	// get cache-aligned buffer
+
+  buf[0] = spu_splats(0);	// the value that unlocks the mutex
+
+  mfc_putlluc(buf, ea, 0, 0);	// unconditional put, no reservation reqd
+  spu_readch(MFC_RdAtomicStat);
+}
+
+
+
 bool
 gc_jd_queue_dequeue(gc_eaddr_t q, gc_eaddr_t *item_ea,
 		    int jd_tag, gc_job_desc_t *item)
 {
-  gc_jd_queue_t	local_q;
+  gc_jd_q_links_t	local_q;
 
   // Before aquiring the lock, see if it's possible that there's
   // something in the queue.  Checking in this way makes it easier
@@ -37,7 +56,7 @@ gc_jd_queue_dequeue(gc_eaddr_t q, gc_eaddr_t *item_ea,
   // the lock unless there is something in the queue.
 
   // copy in the queue structure
-  mfc_get(&local_q, q, sizeof(gc_jd_queue_t), gc_sys_tag, 0, 0);
+  mfc_get(&local_q, q, sizeof(local_q), gc_sys_tag, 0, 0);
   mfc_write_tag_mask(1 << gc_sys_tag);	// the tag we're interested in
   mfc_read_tag_status_all();		// wait for DMA to complete
 
@@ -48,15 +67,15 @@ gc_jd_queue_dequeue(gc_eaddr_t q, gc_eaddr_t *item_ea,
   // When we peeked, head was non-zero.  Now grab the
   // lock and do it for real.
 
-  _mutex_lock(q + offsetof(gc_jd_queue_t, mutex));
+  _mutex_lock(q + offsetof(gc_jd_queue_t, m.mutex));
 
   // copy in the queue structure
-  mfc_get(&local_q, q, sizeof(gc_jd_queue_t), gc_sys_tag, 0, 0);
+  mfc_get(&local_q, q, sizeof(local_q), gc_sys_tag, 0, 0);
   mfc_write_tag_mask(1 << gc_sys_tag);	// the tag we're interested in
   mfc_read_tag_status_all();		// wait for DMA to complete
 
   if (local_q.head == 0){		// empty
-    _mutex_unlock(q + offsetof(gc_jd_queue_t, mutex));
+    _fast_mutex_unlock(q + offsetof(gc_jd_queue_t, m.mutex));
     return false;
   }
 
@@ -77,7 +96,7 @@ gc_jd_queue_dequeue(gc_eaddr_t q, gc_eaddr_t *item_ea,
 
 
   // copy the queue structure back out
-  mfc_put(&local_q, q, sizeof(gc_jd_queue_t), gc_sys_tag, 0, 0);
+  mfc_put(&local_q, q, sizeof(local_q), gc_sys_tag, 0, 0);
   mfc_write_tag_mask(1 << gc_sys_tag);	// the tag we're interested in
   mfc_read_tag_status_all();		// wait for DMA to complete
 
@@ -89,7 +108,7 @@ gc_jd_queue_dequeue(gc_eaddr_t q, gc_eaddr_t *item_ea,
   // a normal DMA, and that a putlluc is better than a putllc if
   // you can use it.
 
-  _mutex_unlock(q + offsetof(gc_jd_queue_t, mutex));
+  _fast_mutex_unlock(q + offsetof(gc_jd_queue_t, m.mutex));
   return true;
 }
 
@@ -97,12 +116,12 @@ gc_jd_queue_dequeue(gc_eaddr_t q, gc_eaddr_t *item_ea,
 void
 gc_jd_queue_getllar(gc_eaddr_t q)
 {
-  // get reservation that includes the tail of the queue
-  gc_eaddr_t	tail = q + offsetof(gc_jd_queue_t, tail);
+  // get reservation that includes the flag in the queue
+  gc_eaddr_t	ea = q + offsetof(gc_jd_queue_t, f.flag);
     
   char _tmp[256];
   char *buf = (char *) ALIGN(_tmp, 128);	// get cache-aligned buffer
 
-  mfc_getllar(buf, ALIGN128_EA(tail), 0, 0);
+  mfc_getllar(buf, ALIGN128_EA(ea), 0, 0);
   spu_readch(MFC_RdAtomicStat);
 }