[U-Boot] [PATCH 1/2] common: make bouncebuf work for non-DMA transfers

[Marek Vasut]

On 11/15/2017 03:17 PM, Goldschmidt Simon wrote:
> Bounce buffer may be used for CPU-only transfers (this is currently
> the case for cadence_qspi). However, in this case, invalidating the
> data cache might throw away copied data that is still in the cache
> only.
> 
> To make CPU-only transfers work with bouncebuf (but still take
> advantage of having aligned buffers), a new flag 'GEN_BB_NODMA' is
> introduced. If this flag is set, cache flushing/invalidation is
> skipped and only the alignment part of bouncebuf is active.
> 
> Signed-off-by: Simon Goldschmidt <sgoldschmidt at de.pepperl-fuchs.com>

Why don't you just fix the cache operations in the driver ?
This bounce-buffer for only CPU operations is just wasting CPU cycles
and degrades performance and I just cannot find a good reason for it.

> ---
>  common/bouncebuf.c  | 9 +++++----
>  include/bouncebuf.h | 9 +++++++++
>  2 files changed, 14 insertions(+), 4 deletions(-)
> 
> diff --git a/common/bouncebuf.c b/common/bouncebuf.c
> index 054d9e0302..0d18477f13 100644
> --- a/common/bouncebuf.c
> +++ b/common/bouncebuf.c
> @@ -55,16 +55,17 @@ int bounce_buffer_start(struct bounce_buffer *state, void *data,
>  	 * Flush data to RAM so DMA reads can pick it up,
>  	 * and any CPU writebacks don't race with DMA writes
>  	 */
> -	flush_dcache_range((unsigned long)state->bounce_buffer,
> -				(unsigned long)(state->bounce_buffer) +
> -					state->len_aligned);
> +	if (!(state->flags & GEN_BB_NODMA))
> +		flush_dcache_range((unsigned long)state->bounce_buffer,
> +					(unsigned long)(state->bounce_buffer) +
> +						state->len_aligned);
>  
>  	return 0;
>  }
>  
>  int bounce_buffer_stop(struct bounce_buffer *state)
>  {
> -	if (state->flags & GEN_BB_WRITE) {
> +	if ((state->flags & (GEN_BB_WRITE|GEN_BB_NODMA)) == GEN_BB_WRITE) {
>  		/* Invalidate cache so that CPU can see any newly DMA'd data */
>  		invalidate_dcache_range((unsigned long)state->bounce_buffer,
>  					(unsigned long)(state->bounce_buffer) +
> diff --git a/include/bouncebuf.h b/include/bouncebuf.h
> index 5ffa99bc13..c6720b3b2e 100644
> --- a/include/bouncebuf.h
> +++ b/include/bouncebuf.h
> @@ -37,6 +37,15 @@
>   */
>  #define GEN_BB_RW	(GEN_BB_READ | GEN_BB_WRITE)
>  
> +/*
> + * GEN_BB_NODMA -- Data is read/written by CPU only (no DMA), so no cache
> + * flushing and invalidating is required. Not passing this for GEN_BB_WRITE
> + * transfers done by the CPU (not DMA) may result in invalid data as the data
> + * written into the cache is lost by invalidating the dcache after the transfer
> + * (unless a writethrough cache is used).
> + */
> +#define GEN_BB_NODMA	(1 << 2)
> +
>  struct bounce_buffer {
>  	/* Copy of data parameter passed to start() */
>  	void *user_buffer;
> 


-- 
Best regards,
Marek Vasut