@@ -123,8 +123,7 @@ int hantro_h1_jpeg_enc_run(struct hantro_ctx *ctx)
| H1_REG_AXI_CTRL_INPUT_SWAP32
| H1_REG_AXI_CTRL_OUTPUT_SWAP8
| H1_REG_AXI_CTRL_INPUT_SWAP8;
- /* Make sure that all registers are written at this point. */
- vepu_write(vpu, reg, H1_REG_AXI_CTRL);
+ vepu_write_relaxed(vpu, reg, H1_REG_AXI_CTRL);
reg = H1_REG_ENC_CTRL_WIDTH(MB_WIDTH(ctx->src_fmt.width))
| H1_REG_ENC_CTRL_HEIGHT(MB_HEIGHT(ctx->src_fmt.height))
@@ -152,8 +152,7 @@ int rockchip_vpu2_jpeg_enc_run(struct hantro_ctx *ctx)
| VEPU_REG_INPUT_SWAP8
| VEPU_REG_INPUT_SWAP16
| VEPU_REG_INPUT_SWAP32;
- /* Make sure that all registers are written at this point. */
- vepu_write(vpu, reg, VEPU_REG_DATA_ENDIAN);
+ vepu_write_relaxed(vpu, reg, VEPU_REG_DATA_ENDIAN);
reg = VEPU_REG_AXI_CTRL_BURST_LEN(16);
vepu_write_relaxed(vpu, reg, VEPU_REG_AXI_CTRL);
In the earlier submissions of the Hantro/Rockchip JPEG encoder driver, a wmb() was inserted before the final register write that starts the encoder. In v11, it was removed and the second-to-last register write was changed to a non-relaxed write, which has an implicit wmb() [1]. The rockchip_vpu2 (then rk3399_vpu) variant is even weirder as there is another writel_relaxed() following the non-relaxed one. Turns out only the last writel() needs to be non-relaxed. Device I/O mappings already guarantee strict ordering to the same endpoint, and the writel() triggering the hardware would force all writes to memory to be observed before the writel() to the hardware is observed. [1] https://lore.kernel.org/linux-media/CAAFQd5ArFG0hU6MgcyLd+_UOP3+T_U-aw2FXv6sE7fGqVCVGqw@mail.gmail.com/ Signed-off-by: Chen-Yu Tsai <wenst@chromium.org> --- drivers/staging/media/hantro/hantro_h1_jpeg_enc.c | 3 +-- drivers/staging/media/hantro/rockchip_vpu2_hw_jpeg_enc.c | 3 +-- 2 files changed, 2 insertions(+), 4 deletions(-)