Re: [PATCH 4/4] media: pxa_camera: conversion to dmaengine

From: Robert Jarzmik
Date: Sun Jun 21 2015 - 13:18:31 EST


Guennadi Liakhovetski <g.liakhovetski@xxxxxx> writes:

> Hi Robert,
>
> On Sun, 22 Mar 2015, Robert Jarzmik wrote:
>
>> From: Robert Jarzmik <robert.jarzmik@xxxxxxxxx>
>>
>> Convert pxa_camera to dmaengine. This removes all DMA registers
>> manipulation in favor of the more generic dmaengine API.
>>
>> The functional level should be the same as before. The biggest change is
>> in the videobuf_sg_splice() function, which splits a videobuf-dma into
>
> As also commented below, I'm not sure "splice" is a good word for
> splitting.
Ok. I'm all ears for your best candidate :)

>> several scatterlists for 3 planes captures (Y, U, V).
>
> "Several" is actually exactly 3, isn't it?
Yup, it's 3, definitely. I can amend the commit message accordingly.

>> +static struct scatterlist *videobuf_sg_splice(struct scatterlist *sglist,
>> + int sglen, int offset, int size,
>> + int *new_sg_len)
>> {
>> - int i, offset, dma_len, xfer_len;
>> - struct scatterlist *sg;
>> + struct scatterlist *sg0, *sg, *sg_first = NULL;
>> + int i, dma_len, dropped_xfer_len, dropped_remain, remain;
>> + int nfirst = -1, nfirst_offset = 0, xfer_len;
>>
>> - offset = sg_first_ofs;
>> + *new_sg_len = 0;
>> + dropped_remain = offset;
>> + remain = size;
>> for_each_sg(sglist, sg, sglen, i) {
>
> Ok, after something-that-felt-like-hours of looking at this code, I think,
> I understand now, that first you calculate what sg elements have been used
> for offset bytes, which is either 0, or the size of the Y plain, or size
> of Y + U plains.
You can say it that way. I'd say that I calculate how to "malloc and fill" a new
scatter-gather to represent [offset, offset + size [ interval of the original
sglist.

>
>> dma_len = sg_dma_len(sg);
>> -
>> /* PXA27x Developer's Manual 27.4.4.1: round up to 8 bytes */
>> - xfer_len = roundup(min(dma_len - offset, size), 8);
>> + dropped_xfer_len = roundup(min(dma_len, dropped_remain), 8);
>> + if (dropped_remain)
>> + dropped_remain -= dropped_xfer_len;
>> + xfer_len = dma_len - dropped_xfer_len;
>> +
>> + if ((nfirst < 0) && (xfer_len > 0)) {
>
> Superfluous parentheses
Got it.

>
>> + sg_first = sg;
>> + nfirst = i;
>> + nfirst_offset = dropped_xfer_len;
>> + }
>> + if (xfer_len > 0) {
>> + *new_sg_len = *new_sg_len + 1;
>
> make it
> + (*new_sg_len)++;
Got it.

>> static void pxa_camera_dma_irq(struct pxa_camera_dev *pcdev,
>> enum pxa_camera_active_dma act_dma);
>>
>> @@ -343,93 +379,59 @@ static void pxa_camera_dma_irq_v(void *data)
>> * @channel: dma channel (0 => 'Y', 1 => 'U', 2 => 'V')
>> * @cibr: camera Receive Buffer Register
>> * @size: bytes to transfer
>> - * @sg_first: first element of sg_list
>> - * @sg_first_ofs: offset in first element of sg_list
>> + * @offset: offset in videobuffer of the first byte to transfer
>> *
>> * Prepares the pxa dma descriptors to transfer one camera channel.
>> - * Beware sg_first and sg_first_ofs are both input and output parameters.
>> *
>> - * Returns 0 or -ENOMEM if no coherent memory is available
>> + * Returns 0 if success or -ENOMEM if no memory is available
>> */
>> static int pxa_init_dma_channel(struct pxa_camera_dev *pcdev,
>> struct pxa_buffer *buf,
>> struct videobuf_dmabuf *dma, int channel,
>> - int cibr, int size,
>> - struct scatterlist **sg_first, int *sg_first_ofs)
>> + int cibr, int size, int offset)
>> {
>
> Hmm, ok, can you, please, explain, why you have to change this so much?
> IIUC, the functionality, that you're implementing now is rather similar to
> the present one - you split the global videobuf SG list into 3 SG lists
> for YUV formats. Of course, details are different, you don't use
> pxa_dma_desc and all the low-level values directly, you prepare a standard
> SG list for your dmaengine driver. But the splitting is the same, isn't
> it?
The overall splitting is the same, yes : split one global SG list into 3 SG
lists.

> And the current one seems rather good to me, because it preserves and
> re-uses partially consumed pointers instead of recalculating them every
> time, like you do it in your new version. What's the reason for that? Is
> the current version buggy or the current approach unsuitable for your new
> version?

Let's say "unsuitable", or to be more correct, let's say that I didn't found any
better idea yet. As I find that piece of code a bit complicated too, I'll tell
you what was my need for doing it, and maybe you'll/we'll come up with a better
idea.

The previous version had the good fortune to rely on a _single_ scatter-gather
list. Only the DMA descriptors list was computed to be 3 lists
(dma[channe].sg_cpu[0..n]).

The new version must have 3 separated SG lists, each one fed to a different
dmaengine channel.

So the big goal here is to compute the 3 SGs lists, which was not done
previously. That's the purpose of all that, and I couldn't find any easier
method. Would a "new_sg = sg_extract(sglist, offset, len)" have existed, this code
would have been a breeze ...

>> + struct dma_chan *dma_chan = pcdev->dma_chans[channel];
>> + struct scatterlist *sg = NULL;
>
> Superfluous initialisation
Got it.

>> + int sglen;
>> + struct dma_async_tx_descriptor *tx;
>> +
>> + sg = videobuf_sg_splice(dma->sglist, dma->sglen, offset, size, &sglen);
>
> Isn't it rather a cut, than a splice function?
Yeah, actually I meant "slice", but wrote "splice" ...

>> @@ -550,11 +548,7 @@ static int pxa_videobuf_prepare(struct videobuf_queue *vq,
>> return 0;
>>
>> fail_v:
>> - dma_free_coherent(dev, buf->dmas[1].sg_size,
>> - buf->dmas[1].sg_cpu, buf->dmas[1].sg_dma);
>> fail_u:
>> - dma_free_coherent(dev, buf->dmas[0].sg_size,
>> - buf->dmas[0].sg_cpu, buf->dmas[0].sg_dma);
>> fail:
>
> You don't need 3 exit labels any more.
Yes, I still need "fail:" I think, the other 2 will be done in next iteration.

>> @@ -741,50 +723,41 @@ static void pxa_camera_wakeup(struct pxa_camera_dev *pcdev,
>> *
>> * Context: should only be called within the dma irq handler
>> */
>> -static void pxa_camera_check_link_miss(struct pxa_camera_dev *pcdev)
>> +static void pxa_camera_check_link_miss(struct pxa_camera_dev *pcdev,
>> + dma_cookie_t last_submitted,
>> + dma_cookie_t last_issued)
>> {
>> - int i, is_dma_stopped = 1;
>> + int is_dma_stopped;
>>
>> - for (i = 0; i < pcdev->channels; i++)
>> - if (DDADR(pcdev->dma_chans[i]) != DDADR_STOP)
>> - is_dma_stopped = 0;
>> + is_dma_stopped = last_submitted > last_issued;
>
> IIUC, actually last_issued should be == last_submitted, right?
No.
last_issued == last_submitted is only true if the last descriptor "hot-chaining"
succeeded. If it didn't succeed, and the dma stopped, last_submitted >
last_issued, for as long as dma_async_issue_pending() is not called.

> And your dmaengine driver's .device_tx_status() method returns the last
> "really" submitted cookie,
If you mean the last tx cookie for which tx->submit() was called, then yes.

> so, if that situation occurs, that you submit a transfer when DMA is off, your
> last_buf->cookie[chan] will contain it, as returned by dmaengine_submit(), but
> .device_tx_status() will not return it?
Here I'm a bit lost, but I feel the first hypothesis "last_issued ==
last_submitted" blurs the discussion, so I'll let you first read my comment
about it, and then you'll tell me if you feel there is still a concern.

> Then, this comparison "last_submitted
> > last_issued" isn't a good test, cookies can overrun. Maybe just check for
> unequal? Otherwise I think you'd have to find a way to use
> dma_async_is_complete().
Yes, agreed, I'll try unequal, I just need a couple of days to think about it if
another corner case doesn't arise.

>> + if (camera_status & overrun &&
>> + last_status != DMA_COMPLETE) {
>
> Isn't the compiler suggesting parentheses here?
Not mine at least. As the "&&" has the lowest precedence, my understanding is
that the expression is correct. But I can add parenthesis for maintainability.

>> @@ -1012,10 +996,7 @@ static void pxa_camera_clock_stop(struct soc_camera_host *ici)
>> __raw_writel(0x3ff, pcdev->base + CICR0);
>>
>> /* Stop DMA engine */
>> - DCSR(pcdev->dma_chans[0]) = 0;
>> - DCSR(pcdev->dma_chans[1]) = 0;
>> - DCSR(pcdev->dma_chans[2]) = 0;
>> -
>> + pxa_dma_stop_channels(pcdev);
>
> Isn't calling pxa_dma_stop_channels() in pxa_camera_stop_capture() enough?
> .clock_stop() should only be called after streaming has been stopped. But
> it has been here since forever, so...

I don't think so, because the buffers are not released, ie. acked. The effect of
pxa_dma_stop_channels() is to call dmaengine_terminate_all(). But this last call
is only entitled to free the dma txs if they are acked, which happens when
free_buffer() is called.

So in pxa_camera_stop_capture(), as the buffers were not released, the dmaengine
txs were not released as well. In pxa_camera_clock_stop(), I think they have
been released, so dmaengine_terminate_all() will actually free the resources of
these txs, which it couldn't do sooner.

>> - DRCMR(68) = pcdev->dma_chans[0] | DRCMR_MAPVLD;
>> - DRCMR(69) = pcdev->dma_chans[1] | DRCMR_MAPVLD;
>> - DRCMR(70) = pcdev->dma_chans[2] | DRCMR_MAPVLD;
>> -
>
> Will resume still work after this? Because of dmaengine resume?
I must admit I haven't tried. And I fear I have not implemented suspend/resume
in pxa_dma dmaengine driver. I'd say it's a dmaengine problem now, and that this
chunk is correct. Yet your point is correct, pxa_dma driver has to handle
suspend/resume.

>> __raw_writel(pcdev->save_cicr[i++] & ~CICR0_ENB, pcdev->base + CICR0);
>> __raw_writel(pcdev->save_cicr[i++], pcdev->base + CICR1);
>> __raw_writel(pcdev->save_cicr[i++], pcdev->base + CICR2);
>> @@ -1738,8 +1715,11 @@ static int pxa_camera_probe(struct platform_device *pdev)
>> struct pxa_camera_dev *pcdev;
>> struct resource *res;
>> void __iomem *base;
>> + struct dma_slave_config config;
>
> I would do
>
> + struct dma_slave_config config = {
> + .direction = DMA_DEV_TO_MEM,
> + .src_maxburst = 8,
> + };
>
> and have all other fields initialised to 0 automatically.
Euh and what would do the "to 0 automatically" initialization ? It's an
automatic variable, not a static one.

>> + dma_cap_zero(mask);
>> + dma_cap_set(DMA_SLAVE, mask);
>
> Also DMA_PRIVATE?
Most certainly.

So thanks for the very detailed review, and I think we'll iterate a couple of
mails to converge. I especially have high hopes about the "slice" thing, two
minds are better than a single coffee broken one :)

Cheers.

--
Robert
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
Please read the FAQ at http://www.tux.org/lkml/