Re: [EXT] about DMA API updates proposal


Hake Huang
 

Hi Gala,

I try to find some reference from Linux code, and below is my finding.

https://docs.huihoo.com/doxygen/linux/kernel/3.7/dmatest_8c_source.html
614 static bool filter(struct dma_chan *chan, void *param)
615 {
616 if (!dmatest_match_channel(chan) || !dmatest_match_device(chan->device))
617 return false;
618 else
619 return true;
620 }
621
622 static int __init dmatest_init(void)
623 {
624 dma_cap_mask_t mask;
625 struct dma_chan *chan;
626 int err = 0;
627
628 dma_cap_zero(mask);
629 dma_cap_set(DMA_MEMCPY, mask);
630 for (;;) {
631 chan = dma_request_channel(mask, filter, NULL);
632 if (chan) {
633 err = dmatest_add_channel(chan);
634 if (err) {
635 dma_release_channel(chan);
636 break; /* add_channel failed, punt */
637 }
638 } else
639 break; /* no more channels available */
640 if (max_channels && nr_channels >= max_channels)
641 break; /* we have all we need */
642 }
643
644 return err;
645 }

In Linux the dma api use a filter function point as its parameter, so the zephyr request API can be

static int dma_edma_request_channel(const struct device *dev,
channel_filter_func filter, void * filter_param)
{
int i = 0;
int channel = -EINVAL;

k_mutex_lock(&DEV_DATA(dev)->dma_mutex, K_FOREVER);

for (i = 0; i < DEV_CFG(dev)->dma_channels; i++) {
if (!atomic_test_and_set_bit(DEV_DATA(dev)->dma_channels, i)) {
channel = i;
if (fn && !fn(filter_param)) {
continue;
}
break;
}
}
k_mutex_unlock(&DEV_DATA(dev)->dma_mutex);

return channel;
}

And each SOC vender response to provide the filter function, if no filter then just give NULL.


How about this solution?

Regards,
Hake

-----Original Message-----
From: Kumar Gala <kumar.gala@linaro.org>
Sent: 2021年3月4日 23:00
To: Hake Huang <hake.huang@nxp.com>
Cc: devel@lists.zephyrproject.org; Maureen Helm <maureen.helm@nxp.com>; Erwan Gouriou <erwan.gouriou@linaro.org>
Subject: Re: [Zephyr-devel] [EXT] about DMA API updates proposal

Caution: EXT Email

1. I think we handle that on a case by case basis and wait and see how it develops over time. If we keep these APIs vendor specific for now we can refactor or introduce a general API for ‘filter’ when we see the pattern.

2. The ‘if’ has to exist somewhere. The ‘if’ is either in the API or the caller.

The concern I have is we don’t know what kinda of ‘filters’ we need today beyond the ‘periodic’ one which is specific to NXP hardware.

- k

On Mar 3, 2021, at 9:27 PM, Hake Huang <hake.huang@nxp.com> wrote:

Hi Gala,

According to the usage of these two APIs, I have two concerns

1. What if there are other special DMA attribute than periodic? We will have to add more APIs?
2. in the peripheral driver, the DMA request can be periodic or not periodic, then we have to use if else instead of a simple call with filter.

Regards,
Hake
-----Original Message-----
From: Kumar Gala <kumar.gala@linaro.org>
Sent: 2021年3月3日 22:49
To: Kumar Gala <kumar.gala@linaro.org>
Cc: Hake Huang <hake.huang@nxp.com>; devel@lists.zephyrproject.org;
Maureen Helm <maureen.helm@nxp.com>; Erwan Gouriou
<erwan.gouriou@linaro.org>
Subject: Re: [Zephyr-devel] [EXT] about DMA API updates proposal

Caution: EXT Email

Hake,

So I was thinking about the following idea:

Having 2 general purpose APIs:

dma_request_channel(const struct device *dev)
dma_release_channel(const struct device *dev)

Than 2 eDMA specific APIs:

mcux_edma_request_periodic_channel(const struct device *dev)
mcux_edma_release_periodic_channel(const struct device *dev)

- k

On Mar 2, 2021, at 9:28 PM, Kumar Gala via lists.zephyrproject.org <kumar.gala=linaro.org@lists.zephyrproject.org> wrote:

Hake,

The ‘filter’ feature I’m still trying to figure out how we handle. I’m curious if other vendors have channels that have specific features like ‘periodic’ that only are supported on specific channels or not.

Hopefully, Erwan might be able to know for STM32.

- k

On Mar 2, 2021, at 9:06 PM, Hake Huang <hake.huang@nxp.com> wrote:

Hi Gala,

I will create a separated PR for DMA new API only.

and one question for below:
"use some of the generic devicetree properties in "dts/bindings/dma/dma-controller.yaml”. The number space allocation code seems pretty generic"

Do you mean, instead of just add API call but move below implement to common code?

static int dma_mcux_edma_request_channel(const struct device *dev,
enum dma_channel_filter filter)
{
int i = 0;
int channel = -EINVAL;

k_mutex_lock(&DEV_DATA(dev)->dma_mutex, K_FOREVER);

for (i = 0; i < DEV_CFG(dev)->dma_channels; i++) {
/* channel 0-3 is Period trigger enabled */
if (filter == DMA_CHANNEL_PERIODIC && i > 3) {
break;
}
if (!atomic_test_and_set_bit(DEV_DATA(dev)->dma_channels, i)) {
channel = i;
break;
}
}

k_mutex_unlock(&DEV_DATA(dev)->dma_mutex);

return channel;
}

If so the filter is hard to aligned, as different SOC has different filter setting.

Regards,
Hake

-----Original Message-----
From: Kumar Gala <kumar.gala@linaro.org>
Sent: 2021年3月3日 1:43
To: Hake Huang <hake.huang@nxp.com>
Cc: devel@lists.zephyrproject.org; Maureen Helm
<maureen.helm@nxp.com>
Subject: [EXT] Re: about DMA API updates proposal

Caution: EXT Email

On Feb 26, 2021, at 5:46 AM, Hake Huang <hake.huang@nxp.com> wrote:

Hi All,

I am working on enable NXP EDMA drivers to Zephyr DMA framework, and I have two DMA API changes need to discussion.
1. https://eur01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fzephyrproject-rtos%2Fzephyr%2Fpull%2F28895&;data=04%7C01%7Chake.huang%40nxp.com%7C9fe84a29cebc431a0e6508d8df1e396a%7C686ea1d3bc2b4c6fa92cd99c5c301635%7C0%7C0%7C637504668195258425%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&amp;sdata=1cYqN9yxrA7SDpGLRLLic%2FIvj4I679VkYcaFm8LHIl0%3D&amp;reserved=0
in this PR, I add chain_link by using the dest_chaining_en(Major) and source_chaining_en(minor).
2.
https://eur01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fg
i
th
ub.com%2Fzephyrproject-rtos%2Fzephyr%2Fpull%2F27737&amp;data=04%7C0
1
%7
Chake.huang%40nxp.com%7C3753d08aba3a4799ddb008d8dda294b0%7C686ea1d3
b
c2
b4c6fa92cd99c5c301635%7C0%7C0%7C637503037637038943%7CUnknown%7CTWFp
b
GZ
sb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn
0
%3
D%7C1000&amp;sdata=3QEP18IvoC9MS3C9o274Z0eMIdbIWTzaoZp18la0OzQ%3D&a
m
p;
reserved=0 in this PR,
https://eur01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fg
i
th
ub.com%2Fzephyrproject-rtos%2Fzephyr%2Fpull%2F27737%2Fcommits%2F25e
1
9f
5b4e4ef07b1598104a02c4722386f2f17b&amp;data=04%7C01%7Chake.huang%40
n
xp
.com%7C3753d08aba3a4799ddb008d8dda294b0%7C686ea1d3bc2b4c6fa92cd99c5c301635%7C0%7C0%7C637503037637038943%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&amp;sdata=wp3iDOFIDkcsLjHRqk9OMWe3yiDfBXko6SHkIDrjQeA%3D&amp;reserved=0 two helper APIs is added dma_api_request_channel request_channel; dma_api_release_channel release_channel; the two APIs are for DMA engine which is capable to dynamically allocate its channels to different DMA requests. and a filter is also added in case of special DMA channel features.

Slack topic discussion is here

Can this topic be scheduled in next API meeting? Thanks.


Regards,
Hake
I wonder if we should pull this into a separate PR. I think a fair amount of the code for these APIs could be generic.

For example I think we can use some of the generic devicetree properties in "dts/bindings/dma/dma-controller.yaml”. The number space allocation code seems pretty generic, even if it just library code used by all the drivers.

- k




Join devel@lists.zephyrproject.org to automatically receive all group messages.