]> git.kernelconcepts.de Git - karo-tx-linux.git/blobdiff - drivers/dma/edma.c
dma: edma: Find missed events and issue them
[karo-tx-linux.git] / drivers / dma / edma.c
index 4c1c258a5b5474090ad088fe74a60c6948a83910..2966ef06c4775c7ece0ac0c2cf076931072863d4 100644 (file)
@@ -70,6 +70,7 @@ struct edma_chan {
        int                             ch_num;
        bool                            alloced;
        int                             slot[EDMA_MAX_SLOTS];
+       int                             missed;
        struct dma_slave_config         cfg;
 };
 
@@ -170,6 +171,20 @@ static void edma_execute(struct edma_chan *echan)
                dev_dbg(dev, "first transfer starting %d\n", echan->ch_num);
                edma_start(echan->ch_num);
        }
+
+       /*
+        * This happens due to setup times between intermediate transfers
+        * in long SG lists which have to be broken up into transfers of
+        * MAX_NR_SG
+        */
+       if (echan->missed) {
+               dev_dbg(dev, "missed event in execute detected\n");
+               edma_clean_channel(echan->ch_num);
+               edma_stop(echan->ch_num);
+               edma_start(echan->ch_num);
+               edma_trigger_channel(echan->ch_num);
+               echan->missed = 0;
+       }
 }
 
 static int edma_terminate_all(struct edma_chan *echan)
@@ -387,6 +402,7 @@ static void edma_callback(unsigned ch_num, u16 ch_status, void *data)
        struct device *dev = echan->vchan.chan.device->dev;
        struct edma_desc *edesc;
        unsigned long flags;
+       struct edmacc_param p;
 
        /* Pause the channel */
        edma_pause(echan->ch_num);
@@ -412,7 +428,39 @@ static void edma_callback(unsigned ch_num, u16 ch_status, void *data)
 
                break;
        case DMA_CC_ERROR:
-               dev_dbg(dev, "transfer error on channel %d\n", ch_num);
+               spin_lock_irqsave(&echan->vchan.lock, flags);
+
+               edma_read_slot(EDMA_CHAN_SLOT(echan->slot[0]), &p);
+
+               /*
+                * Issue later based on missed flag which will be sure
+                * to happen as:
+                * (1) we finished transmitting an intermediate slot and
+                *     edma_execute is coming up.
+                * (2) or we finished current transfer and issue will
+                *     call edma_execute.
+                *
+                * Important note: issuing can be dangerous here and
+                * lead to some nasty recursion when we are in a NULL
+                * slot. So we avoid doing so and set the missed flag.
+                */
+               if (p.a_b_cnt == 0 && p.ccnt == 0) {
+                       dev_dbg(dev, "Error occurred, looks like slot is null, just setting miss\n");
+                       echan->missed = 1;
+               } else {
+                       /*
+                        * The slot is already programmed but the event got
+                        * missed, so its safe to issue it here.
+                        */
+                       dev_dbg(dev, "Error occurred but slot is non-null, TRIGGERING\n");
+                       edma_clean_channel(echan->ch_num);
+                       edma_stop(echan->ch_num);
+                       edma_start(echan->ch_num);
+                       edma_trigger_channel(echan->ch_num);
+               }
+
+               spin_unlock_irqrestore(&echan->vchan.lock, flags);
+
                break;
        default:
                break;