Workaround strange situation when EDMA_RESQIP register returns zero instead

of proper value. It caused bunch of "EMPTY CRPB" messages and potentially
may cause premature requests completion, which could cause data corruption.
For most cases it seems enough to just reread register to get proper value.
To protect against worse cases - erase processed queue entries with
impossible values and ignore them if problem still happen.

1 files changed, 31 insertions, 15 deletions

diff --git a/sys/dev/mvs/mvs.c b/sys/dev/mvs/mvs.c
index ee04f98..9968938 100644
--- a/sys/dev/mvs/mvs.c
+++ b/sys/dev/mvs/mvs.c
@@ -455,7 +455,7 @@ mvs_setup_edma_queues(device_t dev)
 	bus_dmamap_sync(ch->dma.workrq_tag, ch->dma.workrq_map,
 	    BUS_DMASYNC_PREWRITE);
 	/* Reponses queue. */
-	bzero(ch->dma.workrp, 256);
+	memset(ch->dma.workrp, 0xff, MVS_WORKRP_SIZE);
 	work = ch->dma.workrp_bus;
 	ATA_OUTL(ch->r_mem, EDMA_RESQBAH, work >> 32);
 	ATA_OUTL(ch->r_mem, EDMA_RESQIP, work & 0xffffffff);
@@ -980,38 +980,54 @@ mvs_crbq_intr(device_t dev)
 	struct mvs_channel *ch = device_get_softc(dev);
 	struct mvs_crpb *crpb;
 	union ccb *ccb;
-	int in_idx, cin_idx, slot;
+	int in_idx, fin_idx, cin_idx, slot;
+	uint32_t val;
 	uint16_t flags;
 
-	in_idx = (ATA_INL(ch->r_mem, EDMA_RESQIP) & EDMA_RESQP_ERPQP_MASK) >>
+	val = ATA_INL(ch->r_mem, EDMA_RESQIP);
+	if (val == 0)
+		val = ATA_INL(ch->r_mem, EDMA_RESQIP);
+	in_idx = (val & EDMA_RESQP_ERPQP_MASK) >>
 	    EDMA_RESQP_ERPQP_SHIFT;
 	bus_dmamap_sync(ch->dma.workrp_tag, ch->dma.workrp_map,
 	    BUS_DMASYNC_POSTREAD);
-	cin_idx = ch->in_idx;
+	fin_idx = cin_idx = ch->in_idx;
 	ch->in_idx = in_idx;
 	while (in_idx != cin_idx) {
 		crpb = (struct mvs_crpb *)
-		    (ch->dma.workrp + MVS_CRPB_OFFSET + (MVS_CRPB_SIZE * cin_idx));
+		    (ch->dma.workrp + MVS_CRPB_OFFSET +
+		    (MVS_CRPB_SIZE * cin_idx));
 		slot = le16toh(crpb->id) & MVS_CRPB_TAG_MASK;
 		flags = le16toh(crpb->rspflg);
-//device_printf(dev, "CRPB %d %d %04x\n", cin_idx, slot, flags);
 		/*
 		 * Handle only successfull completions here.
 		 * Errors will be handled by main intr handler.
 		 */
-		if (ch->numtslots != 0 || (flags & EDMA_IE_EDEVERR) == 0) {
-if ((flags >> 8) & ATA_S_ERROR)
-device_printf(dev, "ERROR STATUS CRPB %d %d %04x\n", cin_idx, slot, flags);
+		if (crpb->id == 0xffff && crpb->rspflg == 0xffff) {
+			device_printf(dev, "Unfilled CRPB "
+			    "%d (%d->%d) tag %d flags %04x rs %08x\n",
+			    cin_idx, fin_idx, in_idx, slot, flags, ch->rslots);
+		} else if (ch->numtslots != 0 ||
+		    (flags & EDMA_IE_EDEVERR) == 0) {
+			crpb->id = 0xffff;
+			crpb->rspflg = 0xffff;
 			if (ch->slot[slot].state >= MVS_SLOT_RUNNING) {
 				ccb = ch->slot[slot].ccb;
-				ccb->ataio.res.status = (flags & MVS_CRPB_ATASTS_MASK) >>
+				ccb->ataio.res.status =
+				    (flags & MVS_CRPB_ATASTS_MASK) >>
 				    MVS_CRPB_ATASTS_SHIFT;
 				mvs_end_transaction(&ch->slot[slot], MVS_ERR_NONE);
-			} else 
-device_printf(dev, "EMPTY CRPB %d (->%d) %d %04x\n", cin_idx, in_idx, slot, flags);
-		} else
-device_printf(dev, "ERROR FLAGS CRPB %d %d %04x\n", cin_idx, slot, flags);
-
+			} else {
+				device_printf(dev, "Unused tag in CRPB "
+				    "%d (%d->%d) tag %d flags %04x rs %08x\n",
+				    cin_idx, fin_idx, in_idx, slot, flags,
+				    ch->rslots);
+			}
+		} else {
+			device_printf(dev,
+			    "CRPB with error %d tag %d flags %04x\n",
+			    cin_idx, slot, flags);
+		}
 		cin_idx = (cin_idx + 1) & (MVS_MAX_SLOTS - 1);
 	}
 	bus_dmamap_sync(ch->dma.workrp_tag, ch->dma.workrp_map,