From be0d217952222b2bd3ed071de9bb0c66d8cc80d9 Mon Sep 17 00:00:00 2001
From: Andre Przywara <andre.przywara@arm.com>
Date: Wed, 27 Jun 2018 01:42:53 +0100
Subject: [PATCH] arm: timer: sunxi: add Allwinner timer erratum workaround
MIME-Version: 1.0
Content-Type: text/plain; charset=utf8
Content-Transfer-Encoding: 8bit

The Allwinner A64 SoCs suffers from an arch timer implementation erratum,
where sometimes the lower 11 bits of the counter value erroneously
become all 0's or all 1's [1]. This leads to sudden jumps, both forwards and
backwards, with the latter one often showing weird behaviour.
Port the workaround proposed for Linux to U-Boot and activate it for all
A64 boards.
This fixes crashes when accessing MMC devices (SD cards), caused by a
recent change to actually use the counter value for timeout checks.

Fixes: 5ff8e54888e4d26a352453564f7f599d29696dc9 ("sunxi: improve throughput
in the sunxi_mmc driver")

[1] http://lists.infradead.org/pipermail/linux-arm-kernel/2018-May/576886.html

Signed-off-by: Andre Przywara <andre.przywara@arm.com>
Reviewed-by: Philipp Tomsich <philipp.tomsich@theobroma-systems.com>
Tested-by: Jagan Teki <jagan@amarulasolutions.com>
Tested-by: Andreas Färber <afaerber@suse.de>
Tested-by: Guillaume Gardet <guillaume.gardet@free.fr>
---
 arch/arm/cpu/armv8/generic_timer.c | 24 ++++++++++++++++++++++++
 arch/arm/mach-sunxi/Kconfig        |  4 ++++
 2 files changed, 28 insertions(+)

diff --git a/arch/arm/cpu/armv8/generic_timer.c b/arch/arm/cpu/armv8/generic_timer.c
index 3d04fde650..c1706dcec1 100644
--- a/arch/arm/cpu/armv8/generic_timer.c
+++ b/arch/arm/cpu/armv8/generic_timer.c
@@ -46,6 +46,30 @@ unsigned long timer_read_counter(void)
 
 	return cntpct;
 }
+#elif CONFIG_SUNXI_A64_TIMER_ERRATUM
+/*
+ * This erratum sometimes flips the lower 11 bits of the counter value
+ * to all 0's or all 1's, leading to jumps forwards or backwards.
+ * Backwards jumps might be interpreted all roll-overs and be treated as
+ * huge jumps forward.
+ * The workaround is to check whether the lower 11 bits of the counter are
+ * all 0 or all 1, then discard this value and read again.
+ * This occasionally discards valid values, but will catch all erroneous
+ * reads and fixes the problem reliably. Also this mostly requires only a
+ * single read, so does not have any significant overhead.
+ * The algorithm was conceived by Samuel Holland.
+ */
+unsigned long timer_read_counter(void)
+{
+	unsigned long cntpct;
+
+	isb();
+	do {
+		asm volatile("mrs %0, cntpct_el0" : "=r" (cntpct));
+	} while (((cntpct + 1) & GENMASK(10, 0)) <= 1);
+
+	return cntpct;
+}
 #else
 /*
  * timer_read_counter() using the Arm Generic Timer (aka arch timer).
diff --git a/arch/arm/mach-sunxi/Kconfig b/arch/arm/mach-sunxi/Kconfig
index a3f7723028..3624a03947 100644
--- a/arch/arm/mach-sunxi/Kconfig
+++ b/arch/arm/mach-sunxi/Kconfig
@@ -84,6 +84,9 @@ config SUNXI_HIGH_SRAM
 	Chips using the latter setup are supposed to select this option to
 	adjust the addresses accordingly.
 
+config SUNXI_A64_TIMER_ERRATUM
+	bool
+
 # Note only one of these may be selected at a time! But hidden choices are
 # not supported by Kconfig
 config SUNXI_GEN_SUN4I
@@ -270,6 +273,7 @@ config MACH_SUN50I
 	select SUNXI_DRAM_DW_32BIT
 	select FIT
 	select SPL_LOAD_FIT
+	select SUNXI_A64_TIMER_ERRATUM
 
 config MACH_SUN50I_H5
 	bool "sun50i (Allwinner H5)"
-- 
2.39.5