Skip to contentSkip to conversation

A proposal to check the device in generic way

DISCUSSION Shevchenko, Andriy 4 participants 5 messages Mar 11, 2013

lore
Share & tools
lore
#0 lore raw

Hello.

Currently in linux-next we have the following things:

$ git grep -n 'chan->device->dev->driver' drivers/dma/

drivers/dma/amba-pl08x.c:1594: if (chan->device->dev->driver != &pl08x_amba_driver.drv)
drivers/dma/dmaengine.c:190: return chan->device->dev->driver->owner;
drivers/dma/edma.c:609: if (chan->device->dev->driver == &edma_driver.driver) {
drivers/dma/omap-dma.c:654: if (chan->device->dev->driver ==
&omap_dma_driver.driver) {
drivers/dma/pl330.c:2374: if (chan->device->dev->driver !=
&pl330_driver.drv)
drivers/dma/sa11x0-dma.c:1080: if (chan->device->dev->driver == &sa11x0_dma_driver.driver)

I think it's a non-generic way to check which driver provides a channel into filter function. First of all, I don't get why that comparison goes as deep as driver structure. Isn't clearer to check chan->device->dev against the struct dev passed in the custom parameter structure? Like:

struct filter_params {
struct dev *dev;
void *param;
};

bool filter_fn(struct dma_chan *chan, void *fparams) {
struct filter_params *p = fparams;
if (chan->device->dev != p->dev)
return false;
...
}

In case my idea has a right to live, what about to move such check inside DMA engine code?

Opinions, comments?

Earlier I tried to discuss this with Arnd here: http://www.spinics.net/lists/arm-kernel/msg220716.html

Signature · 14 lines
Andy Shevchenko <[email protected]>
Intel Finland Oy

---------------------------------------------------------------------
Intel Finland Oy
Registered Address: PL 281, 00181 Helsinki 
Business Identity Code: 0357606 - 4 
Domiciled in Helsinki 

This e-mail and any attachments may contain confidential material for
the sole use of the intended recipient(s). Any review or distribution
by others is strictly prohibited. If you are not the intended
recipient, please contact the sender and delete all copies.
���{.n�+�������+%�����ݶ��w��{.n�+����{��G�����{ay�ʇڙ�,j��f���h��������z_�(�階�ݢj"���m������G������?����&���~��iO��z��v�^�m����������?�I�
#1 lore raw

Hi Andriy

(adding Russell to CC)

Quoted 23 lines · Andriy

Hello.

Currently in linux-next we have the following things:

$ git grep -n 'chan->device->dev->driver' drivers/dma/

drivers/dma/amba-pl08x.c:1594: if (chan->device->dev->driver != &pl08x_amba_driver.drv)
drivers/dma/dmaengine.c:190: return chan->device->dev->driver->owner;
drivers/dma/edma.c:609: if (chan->device->dev->driver == &edma_driver.driver) {
drivers/dma/omap-dma.c:654: if (chan->device->dev->driver ==
&omap_dma_driver.driver) {
drivers/dma/pl330.c:2374: if (chan->device->dev->driver !=
&pl330_driver.drv)
drivers/dma/sa11x0-dma.c:1080: if (chan->device->dev->driver == &sa11x0_dma_driver.driver)

I think it's a non-generic way to check which driver provides a channel into filter function. First of all, I don't get why that comparison goes as deep as driver structure. Isn't clearer to check chan->device->dev against the struct dev passed in the custom parameter structure? Like:

Quoted 4 lines

struct filter_params {
struct dev *dev;
void *param;
};

I don't think you always know which DMA device you want to use with this DMA client - sometimes there are several DMA engine devices, that can be used with your DMA client, or even if it's only one, you don't necessarily have a pointer to it in your DMA client.

Thanks
Guennadi

Quoted 8 lines

bool filter_fn(struct dma_chan *chan, void *fparams) {
struct filter_params *p = fparams;
if (chan->device->dev != p->dev)
return false;
...
}

Quoted 6 lines

In case my idea has a right to live, what about to move such check inside DMA engine code?

Opinions, comments?

Earlier I tried to discuss this with Arnd here: http://www.spinics.net/lists/arm-kernel/msg220716.html

---
Guennadi Liakhovetski, Ph.D.
Freelance Open-Source Software Developer
http://www.open-technology.de/

#2 lore raw
Quoted 26 lines · Guennadi Liakhovetski
Quoted 21 lines

Currently in linux-next we have the following things:

$ git grep -n 'chan->device->dev->driver' drivers/dma/

drivers/dma/amba-pl08x.c:1594: if (chan->device->dev->driver != &pl08x_amba_driver.drv)
drivers/dma/dmaengine.c:190: return chan->device->dev->driver->owner;
drivers/dma/edma.c:609: if (chan->device->dev->driver == &edma_driver.driver) {
drivers/dma/omap-dma.c:654: if (chan->device->dev->driver ==
&omap_dma_driver.driver) {
drivers/dma/pl330.c:2374: if (chan->device->dev->driver !=
&pl330_driver.drv)
drivers/dma/sa11x0-dma.c:1080: if (chan->device->dev->driver == &sa11x0_dma_driver.driver)

I think it's a non-generic way to check which driver provides a channel into filter function. First of all, I don't get why that comparison goes as deep as driver structure. Isn't clearer to check chan->device->dev against the struct dev passed in the custom parameter structure? Like:

Quoted 4 lines

struct filter_params {
struct dev *dev;
void *param;
};

Quoted 4 lines

I don't think you always know which DMA device you want to use with this DMA client - sometimes there are several DMA engine devices, that can be used with your DMA client, or even if it's only one, you don't necessarily have a pointer to it in your DMA client.

Ah, right. But let's assume we will use filter_fn in conjunction with xlate_fn. In that case we will have explicit understanding for which device we registered specific xlate_fn. So, xlate_fn could provide struct device pointer to be compared with, though I prefer to have more suitable mechanism instead of filter_fn.

Signature · 13 lines
Andy Shevchenko <[email protected]>
Intel Finland Oy
---------------------------------------------------------------------
Intel Finland Oy
Registered Address: PL 281, 00181 Helsinki 
Business Identity Code: 0357606 - 4 
Domiciled in Helsinki 

This e-mail and any attachments may contain confidential material for
the sole use of the intended recipient(s). Any review or distribution
by others is strictly prohibited. If you are not the intended
recipient, please contact the sender and delete all copies.
���{.n�+�������+%�����ݶ��w��{.n�+����{��G�����{ay�ʇڙ�,j��f���h��������z_�(�階�ݢj"���m������G������?����&���~��iO��z��v�^�m����������?�I�
#3 lore raw
In reply to Guennadi Liakhovetski
Quoted 34 lines · Guennadi Liakhovetski

Hi Andriy

(adding Russell to CC)

Quoted 23 lines · Andriy

Hello.

Currently in linux-next we have the following things:

$ git grep -n 'chan->device->dev->driver' drivers/dma/

drivers/dma/amba-pl08x.c:1594: if (chan->device->dev->driver != &pl08x_amba_driver.drv)
drivers/dma/dmaengine.c:190: return chan->device->dev->driver->owner;
drivers/dma/edma.c:609: if (chan->device->dev->driver == &edma_driver.driver) {
drivers/dma/omap-dma.c:654: if (chan->device->dev->driver ==
&omap_dma_driver.driver) {
drivers/dma/pl330.c:2374: if (chan->device->dev->driver !=
&pl330_driver.drv)
drivers/dma/sa11x0-dma.c:1080: if (chan->device->dev->driver == &sa11x0_dma_driver.driver)

I think it's a non-generic way to check which driver provides a channel into filter function. First of all, I don't get why that comparison goes as deep as driver structure. Isn't clearer to check chan->device->dev against the struct dev passed in the custom parameter structure? Like:

Quoted 4 lines

struct filter_params {
struct dev *dev;
void *param;
};

Quoted 4 lines

I don't think you always know which DMA device you want to use with this DMA client - sometimes there are several DMA engine devices, that can be used with your DMA client, or even if it's only one, you don't necessarily have a pointer to it in your DMA client.

Correct - take a look at what gets passed - for example, sometimes the filter parameter is just a string. Sometimes, even in the static platform code, we may not have an address for the DMA device structure.

So, how would you supply the struct device for the DMA? You can't.

But, think about it a moment. What we're actually really interested in
is: do we know the structure of the driver's channel specific data? What determines that? The device? Or the driver? It's the driver.

So for safe and correct type checking, we need to make sure that the dma_chan is really associated with _this_ driver before we dereference anything in the driver-private data structures surrounding dma_chan.

Hence, the struct device_driver is the right thing to use. I would however agree that it would be nice if the device model provided us a way to do that without having to delve that deeply ourselves.

#4 lore raw
In reply to Shevchenko, Andriy
Quoted 18 lines · Andriy

Hello.

Currently in linux-next we have the following things:

$ git grep -n 'chan->device->dev->driver' drivers/dma/

drivers/dma/amba-pl08x.c:1594: if (chan->device->dev->driver != &pl08x_amba_driver.drv)
drivers/dma/dmaengine.c:190: return chan->device->dev->driver->owner;
drivers/dma/edma.c:609: if (chan->device->dev->driver == &edma_driver.driver) {
drivers/dma/omap-dma.c:654: if (chan->device->dev->driver ==
&omap_dma_driver.driver) {
drivers/dma/pl330.c:2374: if (chan->device->dev->driver !=
&pl330_driver.drv)
drivers/dma/sa11x0-dma.c:1080: if (chan->device->dev->driver == &sa11x0_dma_driver.driver)

Quoted 4 lines

I think it's a non-generic way to check which driver provides a channel into filter function. First of all, I don't get why that comparison goes as deep as driver structure. Isn't clearer to check chan->device->dev against the struct dev passed in the custom parameter structure? Like:

I agree that checking the driver is wrong and that we should check the device instead, if we have a filter function called from ->xlate(), which in turn gets called by dma_request_slave_channel().

Some of the filter functions above are passed to device drivers and are not used in combination with the DT binding, which implies that there is nothing to check the device against, so it would not work.

If you look for instance at the pl330 driver, there are two filter functions: pl330_dt_filter correctly compares the device pointer, while pl330_filter only gets an integer argument and cannot compare that to the device.

Quoted 13 lines

struct filter_params {
struct dev *dev;
void *param;
};

bool filter_fn(struct dma_chan *chan, void *fparams) {
struct filter_params *p = fparams;
if (chan->device->dev != p->dev)
return false;
...
}

Quoted 6 lines

In case my idea has a right to live, what about to move such check inside DMA engine code?

Opinions, comments?

Earlier I tried to discuss this with Arnd here: http://www.spinics.net/lists/arm-kernel/msg220716.html

I think the right solution is to provide an interface in the dmaengine common code that lets the ->xlate() callback get right at the channel it wants, rather than call dma_request_channel(). It is rather silly d a dmaengine driver that has all the information about the dma channel it wants to return, we have to call back into a common function that then searches all the channel known to the system.

Arnd