From mboxrd@z Thu Jan 1 00:00:00 1970 Return-path: Received: from lb1-smtp-cloud9.xs4all.net ([194.109.24.22]:43196 "EHLO lb1-smtp-cloud9.xs4all.net" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1726986AbeJPVNQ (ORCPT ); Tue, 16 Oct 2018 17:13:16 -0400 Subject: Re: [RFP] Which V4L2 ioctls could be replaced by better versions? To: Tomasz Figa , nicolas@ndufresne.ca Cc: Linux Media Mailing List , Laurent Pinchart References: <6efdab2da3e4263a49a6a2630df7f79511302088.camel@ndufresne.ca> From: Hans Verkuil Message-ID: Date: Tue, 16 Oct 2018 15:22:41 +0200 MIME-Version: 1.0 In-Reply-To: Content-Type: text/plain; charset=utf-8 Content-Language: en-US Content-Transfer-Encoding: 8bit Sender: linux-media-owner@vger.kernel.org List-ID: On 10/03/18 10:24, Tomasz Figa wrote: > On Fri, Sep 21, 2018 at 3:14 AM Nicolas Dufresne wrote: >> >> Le jeudi 20 septembre 2018 à 16:42 +0200, Hans Verkuil a écrit : >>> Some parts of the V4L2 API are awkward to use and I think it would be >>> a good idea to look at possible candidates for that. >>> >>> Examples are the ioctls that use struct v4l2_buffer: the multiplanar support is >>> really horrible, and writing code to support both single and multiplanar is hard. >>> We are also running out of fields and the timeval isn't y2038 compliant. >>> >>> A proof-of-concept is here: >>> >>> https://git.linuxtv.org/hverkuil/media_tree.git/commit/?h=v4l2-buffer&id=a95549df06d9900f3559afdbb9da06bd4b22d1f3 >>> >>> It's a bit old, but it gives a good impression of what I have in mind. >>> >>> Another candidate is VIDIOC_SUBDEV_ENUM_FRAME_INTERVAL/VIDIOC_ENUM_FRAMEINTERVALS: >>> expressing frame intervals as a fraction is really awkward and so is the fact >>> that the subdev and 'normal' ioctls are not the same. >>> >>> Would using nanoseconds or something along those lines for intervals be better? >> >> This one is not a good idea, because you cannot represent well known >> rates used a lot in the field. Like 60000/1001 also known as 59.94 Hz. >> You could endup with drift issues. >> >> For me, what is the most difficult with this API is the fact that it >> uses frame internal (duration) instead of frame rate. >> >>> >>> I have similar concerns with VIDIOC_SUBDEV_ENUM_FRAME_SIZE where there is no >>> stepwise option, making it different from VIDIOC_ENUM_FRAMESIZES. But it should >>> be possible to extend VIDIOC_SUBDEV_ENUM_FRAME_SIZE with stepwise support, I >>> think. >> >> One of the thing to fix, maybe it's doable now, is the differentiation >> between allocation size and display size. Pretty much all video capture >> code assumes this is display size and ignores the selection API. This >> should be documented explicitly. > > Technically, there is already a distinction between allocation and > display size at format level - width and height could be interpreted > as the rectangle containing the captured video, while bytesperline and > sizeimage would match to the allocation size. The behavior between > drivers seems to be extremely variable - some would just enforce > bytesperline = bpp*width and sizeimage = bytesperline*height, while > others would actually give control over them to the user space. > > It's a bit more complicated with video decoders, because the stream, > in addition to the 2 sizes mentioned before, is also characterized by > coded size, which corresponds to the actual number of pixels encoded > in the stream, even if not all contain pixels to be displayed. This is > something that needs to be specified explicitly (and is, in my > documentation patches) in the Codec Interface. > >> >> In fact, the display/allocation dimension isn't very nice, as both >> information overlaps in structures. As an example, you call S_FMT with >> a display size you want, and it will return you an allocation size >> (which yes, can be smaller, because we always round to the middle). >> >>> >>> Do we have more ioctls that could use a refresh? S/G/TRY_FMT perhaps, again in >>> order to improve single vs multiplanar handling. >> >> Yes, but that would fall in a complete redesign I guess. The buffer >> allocation scheme is very inflexible. You can't have buffers of two >> dimensions allocated at the same time for the same queue. Worst, you >> cannot leave even 1 buffer as your scannout buffer while reallocating >> new buffers, this is not permitted by the framework (in software). As a >> side effect, there is no way to optimize the resolution changes, you >> even have to copy your scannout buffer on the CPU, to free it in order >> to proceed. Resolution changes are thus painfully slow, by design. > > Hmm? There is VIDIOC_CREATEBUFS, which can allows you to allocate > buffers for explicitly specified format, with other buffers already > existing in the queue. Of course, we are missing the VIDIOC_DELETEBUFS ioctl. Also, CREATEBUFS is rather awful. Using v4l2_format in the struct was a major mistake. > > Also, I fail to understand the scanout issue. If one exports a vb2 > buffer to a DMA-buf and import it to the scanout engine, it can keep > scanning out from it as long as it want, because the DMA-buf will hold > a reference on the buffer, even if it's removed from the vb2 queue. > >> >> You also cannot switch from internal buffers to importing buffers >> easily (in some case you, like encoder, you cannot do that without >> flushing the encoder state). > > Hmm, technically VIDIOC_CREATEBUFS accepts the "memory" type value, > but I'm not sure what happens if the queue already has buffers > requested with different one. It is not allowed to mix memory types, that will return -EINVAL. I have to say that this is the first time I had this request. It is probably doable, but the use-case is not clear to me. Regards, Hans