Re: [PATCH] [RFC] dmaengine: add fifo_size member

From: Jon Hunter
Date: Fri Jun 07 2019 - 09:03:10 EST

On 07/06/2019 13:17, Peter Ujfalusi wrote:
> On 07/06/2019 13.27, Jon Hunter wrote:
>>>> Hrm, it is still not clear how all of these fits together.
>>>> What happens if you configure ADMA side:
>>>> BURST = 10
>>>> TX/RXSIZE = 100 (100 * 64 bytes?) /* FIFO_SIZE? */
>>>> *THRES = 5
>>>> And if you change the *THRES to 10?
>>>> And if you change the TX/RXSIZE to 50 (50 * 64 bytes?)
>>>> And if you change the BURST to 5?
>>>> In other words what is the relation between all of these?
>>> So the THRES values are only applicable when the FETCHING_POLICY (bit 31
>>> of the CH_FIFO_CTRL) is set. The FETCHING_POLICY bit defines two modes;
>>> a threshold based transfer mode or a burst based transfer mode. The
>>> burst mode transfer data as and when there is room for a burst in the FIFO.
>>> We use the burst mode and so we really should not be setting the THRES
>>> fields as they are not applicable. Oh well something else to correct,
>>> but this is side issue.
>>>> There must be a rule and constraints around these and if we do really
>>>> need a new parameter for ADMA's FIFO_SIZE I'd like it to be defined in a
>>>> generic way so others could benefit without 'misusing' a fifo_size
>>>> parameter for similar, but not quite fifo_size information.
>>> Yes I see what you are saying. One option would be to define both a
>>> src/dst_maxburst and src/dst_minburst size. Then we could use max for
>>> the FIFO size and min for the actual burst size.
>> Actually, we don't even need to do that. We only use src_maxburst for
>> DEV_TO_MEM and dst_maxburst for MEM_TO_DEV. I don't see any reason why
>> we could not use both the src_maxburst for dst_maxburst for both
>> DEV_TO_MEM and MEM_TO_DEV, where one represents the FIFO size and one
>> represents that DMA burst size.
>> Sorry should have thought of that before. Any objections to using these
>> this way? Obviously we would document is clearly in the driver.
> Imho if you can explain it without using 'HACK' in the sentences it
> might be OK, but it does not feel right.

I don't perceive this as a hack. Although from looking at the
description of the src/dst_maxburst these are burst size with regard to
the device, so maybe it is a stretch.

> However since your ADMA and ADMIF is highly coupled and it does needs
> special maxburst information (burst and allocated FIFO depth) I would
> rather use src_maxburst/dst_maxburst alone for DEV_TO_MEM/MEM_TO_DEV:
> ADMA_BURST_SIZE(maxburst) ((maxburst) & 0xff)
> ADMA_FIFO_SIZE(maxburst) (((maxburst) >> 8) & 0xffffff)
> So lower 1 byte is the burst value you want from ADMA
> the other 3 bytes are the allocated FIFO size for the given ADMAIF channel.
> Sure, you need a header for this to make sure there is no
> misunderstanding between the two sides.

I don't like this because as I mentioned to Dmitry, the ADMA can perform
memory-to-memory transfers where such encoding would not be applicable.

That does not align with the description in the
include/linux/dmaengine.h either.

> Or pass the allocated FIFO size via maxburst and then the ADMA driver
> will pick a 'good/safe' burst value for it.
> Or new member, but do you need two of them for src/dst? Probably
> fifo_depth is better word for it, or allocated_fifo_depth.

Right, so looking at the struct dma_slave_config we have ...

u32 src_maxburst;
u32 dst_maxburst;
u32 src_port_window_size;
u32 dst_port_window_size;

Now if we could make these window sizes a union like the following this
could work ...

diff --git a/include/linux/dmaengine.h b/include/linux/dmaengine.h
index 8fcdee1c0cf9..851251263527 100644
--- a/include/linux/dmaengine.h
+++ b/include/linux/dmaengine.h
@@ -360,8 +360,14 @@ struct dma_slave_config {
enum dma_slave_buswidth dst_addr_width;
u32 src_maxburst;
u32 dst_maxburst;
- u32 src_port_window_size;
- u32 dst_port_window_size;
+ union {
+ u32 port_window_size;
+ u32 port_fifo_size;
+ } src;
+ union {
+ u32 port_window_size;
+ u32 port_fifo_size;
+ } dst;