2020-07-22 22:02:56

by Doug Anderson

Subject: [PATCH] soc: qcom-geni-se: Don't use relaxed writes when writing commands

Writing the command is the final step in kicking off a transfer.
Let's use writel() to ensure that any other memory accesses are done
before the command kicks off. It's expected that this is mostly
relevant if we're in DMA mode but since it doesn't appear to regress
performance in a measurable way [1] even in PIO mode and it's easier
to reason about then let's just always use it.

NOTE: this patch came about due to code inspection. No actual
problems were observed that this patch fixes.

[1] Tested by timing "flashrom -p ec" on a Chromebook which stresses
GENI SPI a lot.

Suggested-by: Stephen Boyd <[email protected]>
Signed-off-by: Douglas Anderson <[email protected]>
---

include/linux/qcom-geni-se.h | 4 ++--
1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/include/linux/qcom-geni-se.h b/include/linux/qcom-geni-se.h
index dd464943f717..f50c73be1428 100644
--- a/include/linux/qcom-geni-se.h
+++ b/include/linux/qcom-geni-se.h
@@ -262,7 +262,7 @@ static inline void geni_se_setup_m_cmd(struct geni_se *se, u32 cmd, u32 params)
u32 m_cmd;

m_cmd = (cmd << M_OPCODE_SHFT) | (params & M_PARAMS_MSK);
- writel_relaxed(m_cmd, se->base + SE_GENI_M_CMD0);
+ writel(m_cmd, se->base + SE_GENI_M_CMD0);
}

/**
@@ -282,7 +282,7 @@ static inline void geni_se_setup_s_cmd(struct geni_se *se, u32 cmd, u32 params)
s_cmd &= ~(S_OPCODE_MSK | S_PARAMS_MSK);
s_cmd |= (cmd << S_OPCODE_SHFT);
s_cmd |= (params & S_PARAMS_MSK);
- writel_relaxed(s_cmd, se->base + SE_GENI_S_CMD0);
+ writel(s_cmd, se->base + SE_GENI_S_CMD0);
}

/**
--
2.28.0.rc0.142.g3c755180ce-goog

2020-07-23 00:51:14

by Stephen Boyd

Subject: Re: [PATCH] soc: qcom-geni-se: Don't use relaxed writes when writing commands

Quoting Douglas Anderson (2020-07-22 15:01:20)
> Writing the command is the final step in kicking off a transfer.
> Let's use writel() to ensure that any other memory accesses are done
> before the command kicks off. It's expected that this is mostly
> relevant if we're in DMA mode but since it doesn't appear to regress
> performance in a measurable way [1] even in PIO mode and it's easier
> to reason about then let's just always use it.
>
> NOTE: this patch came about due to code inspection. No actual
> problems were observed that this patch fixes.
>
> [1] Tested by timing "flashrom -p ec" on a Chromebook which stresses
> GENI SPI a lot.
>
> Suggested-by: Stephen Boyd <[email protected]>
> Signed-off-by: Douglas Anderson <[email protected]>
> ---

Reviewed-by: Stephen Boyd <[email protected]>

2020-07-23 05:17:30

by Akash Asthana

Subject: Re: [PATCH] soc: qcom-geni-se: Don't use relaxed writes when writing commands

On 7/23/2020 3:31 AM, Douglas Anderson wrote:
> Writing the command is the final step in kicking off a transfer.
> Let's use writel() to ensure that any other memory accesses are done
> before the command kicks off. It's expected that this is mostly
> relevant if we're in DMA mode but since it doesn't appear to regress
> performance in a measurable way [1] even in PIO mode and it's easier
> to reason about then let's just always use it.
>
> NOTE: this patch came about due to code inspection. No actual
> problems were observed that this patch fixes.
>
> [1] Tested by timing "flashrom -p ec" on a Chromebook which stresses
> GENI SPI a lot.
>
> Suggested-by: Stephen Boyd <[email protected]>
> Signed-off-by: Douglas Anderson <[email protected]>
> ---
Reviewed-by: Akash Asthana <[email protected]>

--
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum,\na Linux Foundation Collaborative Project

2020-07-23 08:58:21

by Mukesh, Savaliya

Subject: Re: [PATCH] soc: qcom-geni-se: Don't use relaxed writes when writing commands

On 7/23/2020 3:31 AM, Douglas Anderson wrote:
> Writing the command is the final step in kicking off a transfer.
> Let's use writel() to ensure that any other memory accesses are done
> before the command kicks off. It's expected that this is mostly
> relevant if we're in DMA mode but since it doesn't appear to regress
> performance in a measurable way [1] even in PIO mode and it's easier
> to reason about then let's just always use it.
>
> NOTE: this patch came about due to code inspection. No actual
> problems were observed that this patch fixes.
>
> [1] Tested by timing "flashrom -p ec" on a Chromebook which stresses
> GENI SPI a lot.
>
> Suggested-by: Stephen Boyd <[email protected]>
> Signed-off-by: Douglas Anderson <[email protected]>
Reviewed-by: Mukesh Kumar Savaliya <[email protected]>
> ---
>
> include/linux/qcom-geni-se.h | 4 ++--
> 1 file changed, 2 insertions(+), 2 deletions(-)
>
> diff --git a/include/linux/qcom-geni-se.h b/include/linux/qcom-geni-se.h
> index dd464943f717..f50c73be1428 100644
> --- a/include/linux/qcom-geni-se.h
> +++ b/include/linux/qcom-geni-se.h
> @@ -262,7 +262,7 @@ static inline void geni_se_setup_m_cmd(struct geni_se *se, u32 cmd, u32 params)
> u32 m_cmd;
>
> m_cmd = (cmd << M_OPCODE_SHFT) | (params & M_PARAMS_MSK);
> - writel_relaxed(m_cmd, se->base + SE_GENI_M_CMD0);
> + writel(m_cmd, se->base + SE_GENI_M_CMD0);
> }
>
> /**
> @@ -282,7 +282,7 @@ static inline void geni_se_setup_s_cmd(struct geni_se *se, u32 cmd, u32 params)
> s_cmd &= ~(S_OPCODE_MSK | S_PARAMS_MSK);
> s_cmd |= (cmd << S_OPCODE_SHFT);
> s_cmd |= (params & S_PARAMS_MSK);
> - writel_relaxed(s_cmd, se->base + SE_GENI_S_CMD0);
> + writel(s_cmd, se->base + SE_GENI_S_CMD0);
> }
>
> /**