diff --git a/drivers/i3c/master/svc-i3c-master.c b/drivers/i3c/master/svc-i3c-master.c
index c0299e3f2cf8..6b0ec74313d3 100644
--- a/drivers/i3c/master/svc-i3c-master.c
+++ b/drivers/i3c/master/svc-i3c-master.c
@@ -113,6 +113,7 @@
 #define SVC_I3C_MWDATAHE     0x0BC
 #define SVC_I3C_MRDATAB      0x0C0
 #define SVC_I3C_MRDATAH      0x0C8
+#define SVC_I3C_MWDATAB1     0x0CC
 #define SVC_I3C_MWMSG_SDR    0x0D0
 #define SVC_I3C_MRMSG_SDR    0x0D4
 #define SVC_I3C_MWMSG_DDR    0x0D8
@@ -133,6 +134,16 @@
 #define SVC_I3C_EVENT_IBI	GENMASK(7, 0)
 #define SVC_I3C_EVENT_HOTJOIN	BIT(31)
 
+/*
+ * SVC_I3C_QUIRK_FIFO_EMPTY:
+ * I3C HW stalls the write transfer if the transmit FIFO becomes empty,
+ * when new data is written to FIFO, I3C HW resumes the transfer but
+ * the first transmitted data bit may have the wrong value.
+ * Workaround:
+ * Fill the FIFO in advance to prevent FIFO from becoming empty.
+ */
+#define SVC_I3C_QUIRK_FIFO_EMPTY	BIT(0)
+
 struct svc_i3c_cmd {
 	u8 addr;
 	bool rnw;
@@ -236,6 +247,11 @@ struct svc_i3c_i2c_dev_data {
 	struct i3c_generic_ibi_pool *ibi_pool;
 };
 
+static inline bool svc_has_quirk(struct svc_i3c_master *master, u32 quirk)
+{
+	return (master->drvdata->quirks & quirk);
+}
+
 static inline bool is_events_enabled(struct svc_i3c_master *master, u32 mask)
 {
 	return !!(master->enabled_events & mask);
@@ -894,7 +910,7 @@ static int svc_i3c_master_do_daa_locked(struct svc_i3c_master *master,
 					u8 *addrs, unsigned int *count)
 {
 	u64 prov_id[SVC_I3C_MAX_DEVS] = {}, nacking_prov_id = 0;
-	unsigned int dev_nb = 0, last_addr = 0;
+	unsigned int dev_nb = 0, last_addr = 0, dyn_addr;
 	u32 reg;
 	int ret, i;
 
@@ -939,6 +955,25 @@ static int svc_i3c_master_do_daa_locked(struct svc_i3c_master *master,
 		if (SVC_I3C_MSTATUS_RXPEND(reg)) {
 			u8 data[6];
 
+			/*
+			 * One slave sends its ID to request for address assignment,
+			 * prefilling the dynamic address can reduce SCL clock stalls
+			 * and also fix the SVC_I3C_QUIRK_FIFO_EMPTY quirk.
+			 *
+			 * Ideally, prefilling before the processDAA command is better.
+			 * However, it requires an additional check to write the dyn_addr
+			 * at the right time because the driver needs to write the processDAA
+			 * command twice for one assignment.
+			 * Prefilling here is safe and efficient because the FIFO starts
+			 * filling within a few hundred nanoseconds, which is significantly
+			 * faster compared to the 64 SCL clock cycles.
+			 */
+			dyn_addr = i3c_master_get_free_addr(&master->base, last_addr + 1);
+			if (dyn_addr < 0)
+				return -ENOSPC;
+
+			writel(dyn_addr, master->regs + SVC_I3C_MWDATAB);
+
 			/*
 			 * We only care about the 48-bit provisioned ID yet to
 			 * be sure a device does not nack an address twice.
@@ -1017,21 +1052,16 @@ static int svc_i3c_master_do_daa_locked(struct svc_i3c_master *master,
 		if (ret)
 			break;
 
-		/* Give the slave device a suitable dynamic address */
-		ret = i3c_master_get_free_addr(&master->base, last_addr + 1);
-		if (ret < 0)
-			break;
-
-		addrs[dev_nb] = ret;
+		addrs[dev_nb] = dyn_addr;
 		dev_dbg(master->dev, "DAA: device %d assigned to 0x%02x\n",
 			dev_nb, addrs[dev_nb]);
-
-		writel(addrs[dev_nb], master->regs + SVC_I3C_MWDATAB);
 		last_addr = addrs[dev_nb++];
 	}
 
 	/* Need manual issue STOP except for Complete condition */
 	svc_i3c_master_emit_stop(master);
+	svc_i3c_master_flush_fifo(master);
+
 	return ret;
 }
 
@@ -1228,6 +1258,24 @@ static int svc_i3c_master_xfer(struct svc_i3c_master *master,
 		       SVC_I3C_MCTRL_RDTERM(*actual_len),
 		       master->regs + SVC_I3C_MCTRL);
 
+		/*
+		 * The entire transaction can consist of multiple write transfers.
+		 * Prefilling before EmitStartAddr causes the data to be emitted
+		 * immediately, becoming part of the previous transfer.
+		 * The only way to work around this hardware issue is to let the
+		 * FIFO start filling as soon as possible after EmitStartAddr.
+		 */
+		if (svc_has_quirk(master, SVC_I3C_QUIRK_FIFO_EMPTY) && !rnw && xfer_len) {
+			u32 end = xfer_len > SVC_I3C_FIFO_SIZE ? 0 : SVC_I3C_MWDATAB_END;
+			u32 len = min_t(u32, xfer_len, SVC_I3C_FIFO_SIZE);
+
+			writesb(master->regs + SVC_I3C_MWDATAB1, out, len - 1);
+			/* Mark END bit if this is the last byte */
+			writel(out[len - 1] | end, master->regs + SVC_I3C_MWDATAB);
+			xfer_len -= len;
+			out += len;
+		}
+
 		ret = readl_poll_timeout(master->regs + SVC_I3C_MSTATUS, reg,
 				 SVC_I3C_MSTATUS_MCTRLDONE(reg), 0, 1000);
 		if (ret)
@@ -1316,6 +1364,7 @@ static int svc_i3c_master_xfer(struct svc_i3c_master *master,
 emit_stop:
 	svc_i3c_master_emit_stop(master);
 	svc_i3c_master_clear_merrwarn(master);
+	svc_i3c_master_flush_fifo(master);
 
 	return ret;
 }
@@ -1970,7 +2019,9 @@ static const struct dev_pm_ops svc_i3c_pm_ops = {
 			   svc_i3c_runtime_resume, NULL)
 };
 
-static const struct svc_i3c_drvdata npcm845_drvdata = {};
+static const struct svc_i3c_drvdata npcm845_drvdata = {
+	.quirks = SVC_I3C_QUIRK_FIFO_EMPTY,
+};
 
 static const struct svc_i3c_drvdata svc_default_drvdata = {};