arm: mvebu: Enable L2 cache also on Armada 38x

For some unknown reason when L2 cache is disabled on Armada 385 then loadb,
loadx and loady commands do not work with higher baudrates than 115200
(they just abort transfer) and lzmadec command with lzma image of size
0x7000000 (maybe even smaller, we tested this one) is doing decompression
for more than 2 minutes. After enabling L2 cache decompression takes only
30s and loadb, loadx and loady are stable and working fine.

git bisect identified problematic commit 3308933d2fe9 ("arm: mvebu: Avoid
reading MVEBU_REG_PCIE_DEVID register too many times"). Before this commit
above issues were not present.

But investigation showed that above issue was possible to reproduce also by
reverting that commit and forcing compiler to do inline optimization of
mvebu_soc_family() function. Which seems that the root of this issue is in
caches and position of instruction of segments. So currently it is unknown
what is or was broken, but code movement, code inlining or other compiler
optimization triggered it.

Commit 3e5ce7ceeb94 ("arm: mvebu: Enable L2 cache on Armada XP") mentioned
that enabling L2 cache on Armada XP improved performance and that Armada
38x has L2 disabled (which is default state) and if needed it has to be
enabled in separate patch. As enabling L2 cache also improve performance
on Armada 38x, enable it.

Note that Aurora cache in no outer mode is available only on Armada XP,
hence it is not touched for Armada 38x code.

Fixes: 3308933d2fe9 ("arm: mvebu: Avoid reading MVEBU_REG_PCIE_DEVID register too many times")
Reported-by: Marek Behún <kabel@kernel.org>
Signed-off-by: Pali Rohár <pali@kernel.org>
Signed-off-by: Marek Behún <kabel@kernel.org>
Reviewed-by: Stefan Roese <sr@denx.de>

1 files changed, 15 insertions, 7 deletions

diff --git a/arch/arm/mach-mvebu/cpu.c b/arch/arm/mach-mvebu/cpu.c
index 949bf21c9b..77e403ba73 100644
--- a/arch/arm/mach-mvebu/cpu.c
+++ b/arch/arm/mach-mvebu/cpu.c
@@ -671,13 +671,21 @@ void enable_caches(void)
 
 void v7_outer_cache_enable(void)
 {
+	struct pl310_regs *const pl310 =
+		(struct pl310_regs *)CONFIG_SYS_PL310_BASE;
+
+	/* The L2 cache is already disabled at this point */
+
+	/*
+	 * For now L2 cache will be enabled only for Armada XP and Armada 38x.
+	 * It can be enabled also for other SoCs after testing that it works fine.
+	 */
+	if (!IS_ENABLED(CONFIG_ARMADA_XP) && !IS_ENABLED(CONFIG_ARMADA_38X))
+		return;
+
 	if (IS_ENABLED(CONFIG_ARMADA_XP)) {
-		struct pl310_regs *const pl310 =
-			(struct pl310_regs *)CONFIG_SYS_PL310_BASE;
 		u32 u;
 
-		/* The L2 cache is already disabled at this point */
-
 		/*
 		 * For Aurora cache in no outer mode, enable via the CP15
 		 * coprocessor broadcasting of cache commands to L2.
@@ -687,10 +695,10 @@ void v7_outer_cache_enable(void)
 		asm volatile("mcr p15, 1, %0, c15, c2, 0" : : "r" (u));
 
 		isb();
-
-		/* Enable the L2 cache */
-		setbits_le32(&pl310->pl310_ctrl, L2X0_CTRL_EN);
 	}
+
+	/* Enable the L2 cache */
+	setbits_le32(&pl310->pl310_ctrl, L2X0_CTRL_EN);
 }
 
 void v7_outer_cache_disable(void)