[FFmpeg-devel] Add protocol for Android content providers

[Matthieu Bouron]

On Tue, Feb 27, 2024 at 09:14:04PM +0800, Zhao Zhili wrote:
> 
> 
> > On Feb 27, 2024, at 15:17, Matthieu Bouron <matthieu.bouron at gmail.com> wrote:
> > 
> > On Sat, Feb 24, 2024 at 12:29:24PM +0100, Matthieu Bouron wrote:
> >> On Thu, Feb 15, 2024 at 10:13:03AM +0100, Matthieu Bouron wrote:
> >>> Le jeu. 15 févr. 2024, 9:46 AM, Zhao Zhili <quinkblack at foxmail.com> a
> >>> écrit :
> >>> 
> >>>> 
> >>>>> 在 2024年2月15日，下午3:57，Matthieu Bouron <matthieu.bouron at gmail.com> 写道：
> >>>>> 
> >>>>> On Thu, Feb 15, 2024 at 12:13:59PM +0800, Zhao Zhili wrote:
> >>>>>> 
> >>>>>> 
> >>>>>>>> On Feb 14, 2024, at 06:50, Matthieu Bouron <matthieu.bouron at gmail.com>
> >>>> wrote:
> >>>>>>> 
> >>>>>>> Hi,
> >>>>>>> 
> >>>>>>> On Android, content providers are used for accessing files through
> >>>> shared
> >>>>>>> mechanisms. One typical case would be an app willing to open a video
> >>>> from
> >>>>>>> Google Photos, gallery apps, TikTok, Instagram or some other providers.
> >>>>>>> A content URI looks something like "content://authority/path/id", see:
> >>>>>>> https://developer.android.com/reference/android/content/ContentUris
> >>>>>>> 
> >>>> https://developer.android.com/guide/topics/providers/content-provider-basics
> >>>>>>> 
> >>>>>>> It can currently be somehow managed through clumsy means such as using
> >>>> a "fd:"
> >>>>>>> filename and crafting a special AVOption, which also has the drawback
> >>>> of
> >>>>>>> requiring the third party to carry around opened file descriptors
> >>>> (with the
> >>>>>>> multiple opened file limitations implied). Custom AVIOContexts are
> >>>> also an
> >>>>>> 
> >>>>>> File descriptor is a general abstraction layer, it target more
> >>>> platforms than
> >>>>>> Android specific content provider. Android provided getFd() API since
> >>>> API
> >>>>>> level 12, I guess that’s the default method to deal with content
> >>>> provider in
> >>>>>> native code. It’s a few lines of code to get native fd in Java, but
> >>>> dozens of code
> >>>>>> in C with JNI, which is what this patchset done.
> >>>>>> 
> >>>>>> For multiple opened file limitations issue, they can close the file
> >>>> descriptor after
> >>>>>> open. It’s unlikely to reach the limit in normal case without leak.
> >>>>>> 
> >>>>>> I’m OK to provide this android_content_protocol helper if user requests.
> >>>>> 
> >>>>> I've been doing this kind of work for 3/4 users (including myself) at
> >>>> this
> >>>>> point and have to do it another time, this is what motivated me to
> >>>> propose
> >>>>> this patchset.
> >>>>> 
> >>>>>> 
> >>>>>>> option. Both options will have to deal with the JNI though and end
> >>>> users will
> >>>>>>> have to re-implement the same exact thing.
> >>>>>> 
> >>>>>> User still need to deal with JNI with the new android_content_protocol,
> >>>> more or
> >>>>>> less, it’s unavoidable.
> >>>>> 
> >>>>> The advantage I see of using this protocol is that the user only need to
> >>>>> call av_jni_set_jvm() + av_jni_set_android_app_ctx() at the start of the
> >>>>> application and FFmpeg will handle the content-uri transparently. This is
> >>>>> especially helpful if the Android application rely on multiple libraries
> >>>>> that in turn rely on FFmpeg to read medias.
> >>>> 
> >>>> The url still need to be passed from Java to C via JNI, it’s not much
> >>>> different compared to pass fd.
> >>>> 
> >>> 
> >>> It's not that much different I agree. But let's say you have a rendering
> >>> engine (in C) where you need to pass hundreds of media (from the user) to
> >>> render a scene, each media is used at different time during the rendering.
> >>> And Ffmpeg is not a direct dependency and can be called from different
> >>> libraries/places used by the rendering engine. Calling
> >>> av_jni_set_android_app_ctx() and you're done, you can pass the content URI
> >>> to the engine (passing fd at this stage is not an option imho). You still
> >>> need to convert the uri from java string to c before calling the c code,
> >>> but it's a direct translation which is typically part of a binding.
> >>> 
> >>> 
> >>> 
> >>>>> 
> >>>>>> 
> >>>>>>> 
> >>>>>>> This patchset addresses this by adding a content provider protocol,
> >>>> which has
> >>>>>>> an API fairly similar to fopen. Android 11 appears to provide something
> >>>>>>> transparent within fopen(), but FFmpeg doesn't use it in the file
> >>>> protocol, and
> >>>>>>> Android < 11 are still widely used.
> >>>>>>> 
> >>>>>>> The first part move the JNI infrastructure from avcodec to avutil (it
> >>>> remains
> >>>>>>> internally shared, there is little user implication),
> >>>>>> 
> >>>>>> OK. JNI infrastructure should belong to avutil at the first place, so
> >>>> hwcontext_mediacodec
> >>>>>> and so on can use it. Unfortunately for those new avpriv_.
> >>>>> 
> >>>>> What do you mean by "Unfortunately" ? Would you like to make the JNI API
> >>>>> public ?
> >>>> 
> >>>> I think it’s our target to reduce the number of avpriv API, not increase
> >>>> it. Does duplicate the compile unit work in this case so we don’t need to
> >>>> export the symbols?
> >>>> 
> >>> 
> >>> Directly including ffjni.c from libavformat/file.c works. We still need to
> >>> pass the application context though (could be added to avcodec/jni.h)
> >> 
> >> So what would be the preferred way forward ? including libavformat/file.c or
> >> migrating the code to avutil (avpriv_*) ?
> > 
> > Ping (sorry to ping this early, I'd like to not miss the 7.0 window,
> > especially if we choose the avpriv_ route).
> 
> I prefer the solution to dup compile unit and use as less avpriv_ as possible.

Thanks. I'll adjust the patchset accordingly.

[...]