diff --git a/sys/dev/ahci/ahci.c b/sys/dev/ahci/ahci.c
index 314bd91d1198..6794f45b70a4 100644
--- a/sys/dev/ahci/ahci.c
+++ b/sys/dev/ahci/ahci.c
@@ -1625,12 +1625,13 @@ ahci_execute_transaction(struct ahci_slot *slot)
 	/* Setup the command list entry */
 	clp = (struct ahci_cmd_list *)
 	    (ch->dma.work + AHCI_CL_OFFSET + (AHCI_CL_SIZE * slot->slot));
-	clp->prd_length = slot->dma.nsegs;
-	clp->cmd_flags = (ccb->ccb_h.flags & CAM_DIR_OUT ? AHCI_CMD_WRITE : 0) |
-		     (ccb->ccb_h.func_code == XPT_SCSI_IO ?
-		      (AHCI_CMD_ATAPI | AHCI_CMD_PREFETCH) : 0) |
-		     (fis_size / sizeof(u_int32_t)) |
-		     (port << 12);
+	clp->cmd_flags = htole16(
+		    (ccb->ccb_h.flags & CAM_DIR_OUT ? AHCI_CMD_WRITE : 0) |
+		    (ccb->ccb_h.func_code == XPT_SCSI_IO ?
+		     (AHCI_CMD_ATAPI | AHCI_CMD_PREFETCH) : 0) |
+		    (fis_size / sizeof(u_int32_t)) |
+		    (port << 12));
+	clp->prd_length = htole16(slot->dma.nsegs);
 	/* Special handling for Soft Reset command. */
 	if ((ccb->ccb_h.func_code == XPT_ATA_IO) &&
 	    (ccb->ataio.cmd.flags & CAM_ATAIO_CONTROL)) {
@@ -1650,7 +1651,7 @@ ahci_execute_transaction(struct ahci_slot *slot)
 	clp->cmd_table_phys = htole64(ch->dma.work_bus + AHCI_CT_OFFSET +
 				  (AHCI_CT_SIZE * slot->slot));
 	bus_dmamap_sync(ch->dma.work_tag, ch->dma.work_map,
-	    BUS_DMASYNC_PREWRITE);
+	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
 	bus_dmamap_sync(ch->dma.rfis_tag, ch->dma.rfis_map,
 	    BUS_DMASYNC_PREREAD);
 	/* Set ACTIVE bit for NCQ commands. */
@@ -1855,10 +1856,13 @@ ahci_end_transaction(struct ahci_slot *slot, enum ahci_err_type et)
 	device_t dev = slot->dev;
 	struct ahci_channel *ch = device_get_softc(dev);
 	union ccb *ccb = slot->ccb;
+	struct ahci_cmd_list *clp;
 	int lastto;
 
 	bus_dmamap_sync(ch->dma.work_tag, ch->dma.work_map,
-	    BUS_DMASYNC_POSTWRITE);
+	    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
+	clp = (struct ahci_cmd_list *)
+	    (ch->dma.work + AHCI_CL_OFFSET + (AHCI_CL_SIZE * slot->slot));
 	/* Read result registers to the result struct
 	 * May be incorrect if several commands finished same time,
 	 * so read only when sure or have to.
@@ -1893,6 +1897,16 @@ ahci_end_transaction(struct ahci_slot *slot, enum ahci_err_type et)
 			res->sector_count_exp = fis[13];
 		} else
 			bzero(res, sizeof(*res));
+		if ((ccb->ataio.cmd.flags & CAM_ATAIO_FPDMA) == 0 &&
+		    (ccb->ccb_h.flags & CAM_DIR_MASK) != CAM_DIR_NONE) {
+			ccb->ataio.resid =
+			    ccb->ataio.dxfer_len - le32toh(clp->bytecount);
+		}
+	} else {
+		if ((ccb->ccb_h.flags & CAM_DIR_MASK) != CAM_DIR_NONE) {
+			ccb->csio.resid =
+			    ccb->csio.dxfer_len - le32toh(clp->bytecount);
+		}
 	}
 	if ((ccb->ccb_h.flags & CAM_DIR_MASK) != CAM_DIR_NONE) {
 		bus_dmamap_sync(ch->dma.data_tag, slot->dma.data_map,
diff --git a/sys/dev/ata/ata-all.c b/sys/dev/ata/ata-all.c
index f40dcc2edcb1..c7fe9baa4deb 100644
--- a/sys/dev/ata/ata-all.c
+++ b/sys/dev/ata/ata-all.c
@@ -1517,6 +1517,15 @@ ata_cam_end_transaction(device_t dev, struct ata_request *request)
 		res->sector_count = request->u.ata.count;
 		res->sector_count_exp = request->u.ata.count >> 8;
 	}
+	if ((ccb->ccb_h.flags & CAM_DIR_MASK) != CAM_DIR_NONE) {
+		if (ccb->ccb_h.func_code == XPT_ATA_IO) {
+			ccb->ataio.resid =
+			    ccb->ataio.dxfer_len - request->donecount;
+		} else {
+			ccb->csio.resid =
+			    ccb->csio.dxfer_len - request->donecount;
+		}
+	}
 	ata_free_request(request);
 	xpt_done(ccb);
 	/* Do error recovery if needed. */
diff --git a/sys/dev/siis/siis.c b/sys/dev/siis/siis.c
index ba10c92c4696..e75fd38b3596 100644
--- a/sys/dev/siis/siis.c
+++ b/sys/dev/siis/siis.c
@@ -1208,6 +1208,17 @@ siis_end_transaction(struct siis_slot *slot, enum siis_err_type et)
 			res->sector_count_exp = ATA_INB(ch->r_mem, offs + 13);
 		} else
 			bzero(res, sizeof(*res));
+		if ((ccb->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_IN &&
+		    ch->numrslots == 1) {
+			ccb->ataio.resid = ccb->ataio.dxfer_len -
+			    ATA_INL(ch->r_mem, SIIS_P_LRAM_SLOT(slot->slot) + 4);
+		}
+	} else {
+		if ((ccb->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_IN &&
+		    ch->numrslots == 1) {
+			ccb->csio.resid = ccb->csio.dxfer_len -
+			    ATA_INL(ch->r_mem, SIIS_P_LRAM_SLOT(slot->slot) + 4);
+		}
 	}
 	if ((ccb->ccb_h.flags & CAM_DIR_MASK) != CAM_DIR_NONE) {
 		bus_dmamap_sync(ch->dma.data_tag, slot->dma.data_map,