diff --git a/drivers/mmc/host/dw_mmc-rockchip.c b/drivers/mmc/host/dw_mmc-rockchip.c
index 2b4bcd212767..358b0dc853b0 100644
--- a/drivers/mmc/host/dw_mmc-rockchip.c
+++ b/drivers/mmc/host/dw_mmc-rockchip.c
@@ -66,6 +66,70 @@ static void dw_mci_rk3288_set_ios(struct dw_mci *host, struct mmc_ios *ios)
 	/* Make sure we use phases which we can enumerate with */
 	if (!IS_ERR(priv->sample_clk))
 		clk_set_phase(priv->sample_clk, priv->default_sample_phase);
+
+	/*
+	 * Set the drive phase offset based on speed mode to achieve hold times.
+	 *
+	 * NOTE: this is _not_ a value that is dynamically tuned and is also
+	 * _not_ a value that will vary from board to board.  It is a value
+	 * that could vary between different SoC models if they had massively
+	 * different output clock delays inside their dw_mmc IP block (delay_o),
+	 * but since it's OK to overshoot a little we don't need to do complex
+	 * calculations and can pick values that will just work for everyone.
+	 *
+	 * When picking values we'll stick with picking 0/90/180/270 since
+	 * those can be made very accurately on all known Rockchip SoCs.
+	 *
+	 * Note that these values match values from the DesignWare Databook
+	 * tables for the most part except for SDR12 and "ID mode".  For those
+	 * two modes the databook calculations assume a clock in of 50MHz.  As
+	 * seen above, we always use a clock in rate that is exactly the
+	 * card's input clock (times RK3288_CLKGEN_DIV, but that gets divided
+	 * back out before the controller sees it).
+	 *
+	 * From measurement of a single device, it appears that delay_o is
+	 * about .5 ns.  Since we try to leave a bit of margin, it's expected
+	 * that numbers here will be fine even with much larger delay_o
+	 * (the 1.4 ns assumed by the DesignWare Databook would result in the
+	 * same results, for instance).
+	 */
+	if (!IS_ERR(priv->drv_clk)) {
+		int phase;
+
+		/*
+		 * In almost all cases a 90 degree phase offset will provide
+		 * sufficient hold times across all valid input clock rates
+		 * assuming delay_o is not absurd for a given SoC.  We'll use
+		 * that as a default.
+		 */
+		phase = 90;
+
+		switch (ios->timing) {
+		case MMC_TIMING_MMC_DDR52:
+			/*
+			 * Since clock in rate with MMC_DDR52 is doubled when
+			 * bus width is 8 we need to double the phase offset
+			 * to get the same timings.
+			 */
+			if (ios->bus_width == MMC_BUS_WIDTH_8)
+				phase = 180;
+			break;
+		case MMC_TIMING_UHS_SDR104:
+		case MMC_TIMING_MMC_HS200:
+			/*
+			 * In the case of 150 MHz clock (typical max for
+			 * Rockchip SoCs), 90 degree offset will add a delay
+			 * of 1.67 ns.  That will meet min hold time of .8 ns
+			 * as long as clock output delay is < .87 ns.  On
+			 * SoCs measured this seems to be OK, but it doesn't
+			 * hurt to give margin here, so we use 180.
+			 */
+			phase = 180;
+			break;
+		}
+
+		clk_set_phase(priv->drv_clk, phase);
+	}
 }
 
 #define NUM_PHASES			360