Date 29.11.2013 10:08, Takashi Iwai wrote:
At Fri, 29 Nov 2013 13:14:59 +0530, Vinod Koul wrote:
On Fri, Nov 29, 2013 at 08:46:03AM +0100, Takashi Iwai wrote:
At Fri, 29 Nov 2013 09:59:57 +0530, Vinod Koul wrote:
As discussed in the last Audio uConf we need to improve byte controls to allow larger size data to be sent to DSPs The modern DSPs require alsa byte controls size which far exceeds the today 512bytes limit. In order for usermode to send larger sizes (few 100s of KBs) along with size information we add extended byte control interface which sends any size bytes parameter buffer for DSPs to use Obviosly the size must be supported by the device and would be required to inform the max size allowed for the control.
My primary question is -- must this be a control element?
I think yes. With controls we have an easy way to send parameter bytes and a good support from both kernel and usermode, so leveraging that would make sense.
But it means that we have to extend / fix allover places using the control elements. And the variable size isn't handled there, so far. It's the biggest concern.
For example, how would you read such a control element? Currently, all the control element values have the same static size, so the data is just there to read. OTOH, if you read data with an unknown size, you have to query the size at first, allocate the buffer, then read the data. It's a completely different flow.
Mostly here the limit of 512 is hitting folks and IMO arbitrarily increasing size doesn't help. For DSPs the algorithm coefficients can be larger
Well, it's basically a kind of abuse of control elements, IMO...
I basically agree, but... I believe that these chunks can be divided to the 512 limit using continuous indexes (kcontrol->count) and a simple rule in the driver "write all to a DSP when the last control (index) is touched" may be enough. No API extensions are required. The question is: Do you rellay need 100+KB for coefficients? Do you expect to handle these data in standard tools like alsactl?
Here i was thinking of just adding a new API in libraries to access this new control type so that users can send large parameter blobs.
mixer_ctl_set_bytes_ext(struct mixer_ctl *ctl, const void *data, size_t len) { struct snd_ctl_elem_value ev;
/*usual init code */
switch (ctl->info->type) { case SNDRV_CTL_ELEM_TYPE_BYTES_EXT: ev.value.bytes_ext.size = len; ev.value.bytes_ext.data = (__u64)data; break; default: return -EINVAL; } return ioctl(ctl->mixer->fd, SNDRV_CTL_IOCTL_ELEM_WRITE, &ev); }
in kernel-side, the .put handler for kcontrol implemented in drivers would then copy the userdata and send it to DSPs.
If you create a new API function, why we need to stick with the control elements?
So in that respect this becomes quite simple to implement without adding lots of parallel code and uses the infrastruture we already have.
Why it can't be implemented in hwdep, for example? In the past, we already implemented DSP firmware loader in the hwdep side. It's a bit complicated (and limited) for a few driver usages, but it was more or less standardized for them.
My concern extending control element is that it'd need the rewrite in the whole control API in user-space side, too. alsactl accesses the whole control elements and save/restore them, for example.
I am not sure I got your comment on rewriting control API. I though the union here makes the ABI itself not to change so existing implementations which don't use this new control type will not be impacted and for newer DSP we can add new APIs like above to take benifit.
You'll change SNDRV_CTL_ELEM_TYPE_LAST. It means you'll need to look over all codes that refer to this and fix them. (Not only in alsa-lib.) That is, it's difficult to know how an old app reacts when it received an unknown data type.
I just don't want to go the way "let's break and wait to see whether people start whining". At least for the old rock solid stuff.
For save-restore part (though i havent looked at alsactl so might be wrong here) yes, for this new control type, we would have to allocate buffers and then destroy on restore. Eventually we will need to do so if we need alsactl suppport for such devices. But shouldnt the change be less intrusive as it would be limited to one control type.
So, you'd suggest to do save/restore unlimited size of data via alsactl?
Also, sharing scalar data and a reference pointer in the same union is a bit error-prone.
Can you elobrate a bit here and if we go down this path then how do we go about it
Currently the data are stored in the fixed size of union. You can read the all data there. If a program sets some wrong type or value in the data, nothing severe would happen. Just a wrong data is transferred. But, with a reference pointer, a wrong data triggers the segfault, crashes the whole program.
One orthogonal question: why do we have bunch of deprecated pointers in the unions and is anyone still using it? I guess this would be bit of history lesson :)
No it's never used in reality, and this was a big pain in the kernel side because we had to convert the data for the 32/64bit compatible ioctls. That's why it was deprecated. The data field remains there just for ABI compatibility reason.
I and Abramo though it may be useful, but until 2008 no one required these big data so you removed this without complains.. But ... the room for this extension is available or it may be implemented using another WRITE_LARGE_BYTES ioctl to satisfy the 32/64 bit conversions.
Jaroslav