Re: [PATCH v4] media: mtk-vpu: Ensure alignment of 8 for DTCM buffer

From: Dafna Hirschfeld
Date: Mon Nov 29 2021 - 09:41:36 EST




On 18.10.21 04:16, Alexandre Courbot wrote:
Hi Hans!

On Mon, Oct 4, 2021 at 6:37 PM Hans Verkuil <hverkuil@xxxxxxxxx> wrote:

On 20/09/2021 19:04, Dafna Hirschfeld wrote:
From: Alexandre Courbot <acourbot@xxxxxxxxxxxx>

When running memcpy_toio:
memcpy_toio(send_obj->share_buf, buf, len);
it was found that errors appear if len is not a multiple of 8:

[58.350841] mtk-mdp 14001000.rdma: processing failed: -22

Why do errors appear? Is that due to a HW bug? Some other reason?

MTK folks would be the best placed to answer this, but since the
failure is reported by the firmware I'd suspect either a firmware or
hardware limitation.



This patch ensures the copy of a multiple of 8 size by calling
round_up(len, 8) when copying

Fixes: e6599adfad30 ("media: mtk-vpu: avoid unaligned access to DTCM buffer.")
Signed-off-by: Alexandre Courbot <acourbot@xxxxxxxxxxxx>
Signed-off-by: Enric Balletbo i Serra <enric.balletbo@xxxxxxxxxxxxx>
Signed-off-by: Dafna Hirschfeld <dafna.hirschfeld@xxxxxxxxxxxxx>
Reviewed-by: Houlong Wei <houlong.wei@xxxxxxxxxxxx>
---
changes since v3:
1. multile -> multiple
2. add inline doc

changes since v2:
1. do the extra copy only if len is not multiple of 8

changes since v1:
1. change sign-off-by tags
2. change values to memset

drivers/media/platform/mtk-vpu/mtk_vpu.c | 15 ++++++++++++++-
1 file changed, 14 insertions(+), 1 deletion(-)

diff --git a/drivers/media/platform/mtk-vpu/mtk_vpu.c b/drivers/media/platform/mtk-vpu/mtk_vpu.c
index ec290dde59cf..1df031716c8f 100644
--- a/drivers/media/platform/mtk-vpu/mtk_vpu.c
+++ b/drivers/media/platform/mtk-vpu/mtk_vpu.c
@@ -349,7 +349,20 @@ int vpu_ipi_send(struct platform_device *pdev,
}
} while (vpu_cfg_readl(vpu, HOST_TO_VPU));

- memcpy_toio(send_obj->share_buf, buf, len);
+ /*
+ * when copying data to the vpu hardware, the memcpy_toio operation must copy
+ * a multiple of 8. Otherwise the processing fails

Same here: it needs to explain why the processing fails.

+ */
+ if (len % 8 != 0) {
+ unsigned char data[SHARE_BUF_SIZE];

Wouldn't it be more robust if you say:

unsigned char data[sizeof(send_obj->share_buf)];

Definitely yes.

won't it actually be better to implement it like this:
(assuming len is always multiply of 4 - which I think it must be since access must be 4 aligned)

void __iomem *to = obj->share_buf;

if (len % 8 != 0) {
memcpy_toio(to, buf, len - 4);
to += len - 4;
buf += len - 4;
writel_relaxed(*(u32 *)buf, to);
} else {
memcpy_toio(obj->share_buf, buf, len);
}
Thanks,
Dafna


I also think that the SHARE_BUF_SIZE define needs a comment stating that it must be a
multiple of 8, otherwise unexpected things can happen.

You also noticed that the current SHARE_BUF_SIZE define is too low, but I saw
no patch correcting this. Shouldn't that be fixed as well?

AFAICT the firmware expects this exact size on its end, so I don't
believe it can be changed that easily. But maybe someone from MTK can
prove me wrong.

Cheers,
Alex.