Re: [PATCH v2] soc: mediatek: mtk-infracfg: Disable ACP on MT8192

From: Matthias Brugger
Date: Tue Mar 01 2022 - 02:25:50 EST


Hi all,

On 17/02/2022 11:40, AngeloGioacchino Del Regno wrote:
Il 15/02/22 19:46, Alyssa Rosenzweig ha scritto:
MT8192 contains an experimental Accelerator Coherency Port
implementation, which does not work correctly but was unintentionally
enabled by default. For correct operation of the GPU, we must set a
chicken bit disabling ACP on MT8192.

Adapted from the following downstream change to the out-of-tree, legacy
Mali GPU driver:

https://chromium-review.googlesource.com/c/chromiumos/third_party/kernel/+/2781271/5

Note this change is required for both Panfrost and the legacy kernel
driver.

Hello Alyssa,
the v2 note should not get inside the commit message, even though they should be
on the patch to provide context.
Look at the example...


I think the most obvious would be to disable the ACP in the boot firmware. But as we can't control that, we will need to work around it. So:
Patch applied. I fixed the commit message by deleting the v2 note.

Thanks everybody for figuring out what to do about that.

Regards,
Matthias


v2: Move the change from clk-mt8192.c to mtk-infracfg.c (Robin).
Although it does not make sense to add this platform-specific hack to
the GPU driver, it has nothing to do with clocks. We already have
mtk-infracfg.c to manage other infracfg bits; the ACP disable should
live there too.

Co-developed-by: Robin Murphy <robin.murphy@xxxxxxx>
Signed-off-by: Robin Murphy <robin.murphy@xxxxxxx>
Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@xxxxxxxxxxxxx>
Cc: Nick Fan <Nick.Fan@xxxxxxxxxxxx>
Cc: Nicolas Boichat <drinkcat@xxxxxxxxxxxx>
Cc: Chen-Yu Tsai <wenst@xxxxxxxxxxxx>
Cc: Stephen Boyd <sboyd@xxxxxxxxxx>
Cc: AngeloGioacchino Del Regno <angelogioacchino.delregno@xxxxxxxxxxxxx>
---

v2: Move the change from clk-mt8192.c to mtk-infracfg.c (Robin).

^^^ put it here, after the "---" :)

Anyway,
Tested-by: AngeloGioacchino Del Regno <angelogioacchino.delregno@xxxxxxxxxxxxx>

  drivers/soc/mediatek/mtk-infracfg.c   | 19 +++++++++++++++++++
  include/linux/soc/mediatek/infracfg.h |  3 +++
  2 files changed, 22 insertions(+)

diff --git a/drivers/soc/mediatek/mtk-infracfg.c b/drivers/soc/mediatek/mtk-infracfg.c
index 0590b68e0d78..2acf19676af2 100644
--- a/drivers/soc/mediatek/mtk-infracfg.c
+++ b/drivers/soc/mediatek/mtk-infracfg.c
@@ -6,6 +6,7 @@
  #include <linux/export.h>
  #include <linux/jiffies.h>
  #include <linux/regmap.h>
+#include <linux/mfd/syscon.h>
  #include <linux/soc/mediatek/infracfg.h>
  #include <asm/processor.h>
@@ -72,3 +73,21 @@ int mtk_infracfg_clear_bus_protection(struct regmap *infracfg, u32 mask,
      return ret;
  }
+
+static int __init mtk_infracfg_init(void)
+{
+    struct regmap *infracfg;
+
+    /*
+     * MT8192 has an experimental path to route GPU traffic to the DSU's
+     * Accelerator Coherency Port, which is inadvertently enabled by
+     * default. It turns out not to work, so disable it to prevent spurious
+     * GPU faults.
+     */
+    infracfg = syscon_regmap_lookup_by_compatible("mediatek,mt8192-infracfg");
+    if (!IS_ERR(infracfg))
+        regmap_set_bits(infracfg, MT8192_INFRA_CTRL,
+                MT8192_INFRA_CTRL_DISABLE_MFG2ACP);
+    return 0;
+}
+postcore_initcall(mtk_infracfg_init);
diff --git a/include/linux/soc/mediatek/infracfg.h b/include/linux/soc/mediatek/infracfg.h
index d858e0bab7a2..fcbbd0dd5e55 100644
--- a/include/linux/soc/mediatek/infracfg.h
+++ b/include/linux/soc/mediatek/infracfg.h
@@ -229,6 +229,9 @@
  #define INFRA_TOPAXI_PROTECTEN_SET        0x0260
  #define INFRA_TOPAXI_PROTECTEN_CLR        0x0264
+#define MT8192_INFRA_CTRL            0x290
+#define MT8192_INFRA_CTRL_DISABLE_MFG2ACP    BIT(9)
+
  #define REG_INFRA_MISC                0xf00
  #define F_DDR_4GB_SUPPORT_EN            BIT(13)