Re: [RFC PATCH] ALSA: core: Add DMA share buffer support

[Baolin Wang]

On Fri, 25 Jan 2019 at 21:03, Takashi Iwai <tiwai@xxxxxxx> wrote:
>
> On Fri, 25 Jan 2019 12:11:43 +0100,
> Baolin Wang wrote:
> >
> > Hi Takashi,
> > On Fri, 25 Jan 2019 at 18:10, Takashi Iwai <tiwai@xxxxxxx> wrote:
> > >
> > > On Fri, 25 Jan 2019 10:25:37 +0100,
> > > Baolin Wang wrote:
> > > >
> > > > Hi Jaroslav,
> > > > On Thu, 24 Jan 2019 at 21:43, Jaroslav Kysela <perex@xxxxxxxx> wrote:
> > > > >
> > > > > Dne 23.1.2019 v 13:46 Leo Yan napsal(a):
> > > > > > Hi all,
> > > > > >
> > > > > > On Wed, Jan 23, 2019 at 12:58:51PM +0100, Takashi Iwai wrote:
> > > > > >> On Tue, 22 Jan 2019 21:25:35 +0100,
> > > > > >> Mark Brown wrote:
> > > > > >>>
> > > > > >>> On Mon, Jan 21, 2019 at 03:15:43PM +0100, Jaroslav Kysela wrote:
> > > > > >>>> Dne 21.1.2019 v 13:40 Mark Brown napsal(a):
> > > > > >>>
> > > > > >>>>> It was the bit about adding more extended permission control that I was
> > > > > >>>>> worried about there, not the initial O_APPEND bit.  Indeed the O_APPEND
> > > > > >>>>> bit sounds like it might also work from the base buffer sharing point of
> > > > > >>>>> view, I have to confess I'd not heard of that feature before (it didn't
> > > > > >>>>> come up in the discussion when Eric raised this in Prague).
> > > > > >>>
> > > > > >>>> With permissions, I meant to make possible to restrict the file
> > > > > >>>> descriptor operations (ioctls) for the depending task (like access to
> > > > > >>>> the DMA buffer, synchronize it for the non-coherent platforms and maybe
> > > > > >>>> read/write the actual position, delay etc.). It should be relatively
> > > > > >>>> easy to implement using the snd_pcm_file structure.
> > > > > >>>
> > > > > >>> Right, that's what I understood you to mean.  If you want to have a
> > > > > >>> policy saying "it's OK to export a PCM file descriptor if it's only got
> > > > > >>> permissions X and Y" the security module is going to need to know about
> > > > > >>> the mechanism for setting those permissions.  With dma_buf that's all a
> > > > > >>> bit easier as there's less new stuff, though I've no real idea how much
> > > > > >>> of a big deal that actually is.
> > > > > >>
> > > > > >> There are many ways to implement such a thing, yeah.  If we'd need an
> > > > > >> implementation that is done solely in the sound driver layer, I can
> > > > > >> imagine to introduce either a new ioctl or an open flag (like O_EXCL)
> > > > > >> to specify the restricted sharing.  That is, a kind of master / slave
> > > > > >> model where only the master is allowed to manipulate the stream while
> > > > > >> the slave can mmap, read/write and get status.
> > > > > >
> > > > > > In order to support EXCLUSIVE mode, it is necessary to convert the
> > > > > > /dev/snd/ descriptor to an anon_inode:dmabuffer file descriptor.
> > > > > > SELinux allows that file descriptor to be passed to the client. It can
> > > > > > also be used by the AAudioService.
> > > > >
> > > > > Okay, so this is probably the only point which we should resolve for the
> > > > > already available DMA buffer sharing in ALSA (the O_APPEND flag).
> > > > >
> > > > > I had another glance to your dma-buf implementation and I see many
> > > > > things which might cause problems:
> > > > >
> > > > > - allow to call dma-buf ioctls only when the audio device is in specific
> > > > > state (stream is not running)
> > > >
> > > > Right. Will fix.
> > > >
> > > > > - as Takashi mentioned, if we return another file-descriptor (dma-buf
> > > > > export) to the user space and the server closes the main pcm
> > > > > file-descriptor (the client does not) - the result will be a crash (dma
> > > > > buffer will be freed, but referenced through the dma-buf interface)
> > > >
> > > > Yes, will fix.
> > >
> > > There are a few more overlooked problems.  A part of them was already
> > > mentioned in my previous reply, but let me repeat:
> > >
> > > - The racy ioctls have to be considered; you can perform this export
> > >   ioctl concurrently, and both of them write and mix up the setup,
> > >   which is obviously broken.
> >
> > Yes, I think I missed the snd_pcm_stream_lock, and will add.
>
> Beware that it's not so trivial.  The stream lock is usually
> spinlock.

Right. Thanks for your reminding.

>
> In addition, we need to be careful about the PCM state, as Jaroslav
> mentioned.  Basically SNDRV_PCM_STATE_SETUP is already too late for
> attaching the buffer, since another buffer is already assigned to the
> stream.  Similarly, after detaching, the stream state must go to
> SNDRV_PCM_STATE_OPEN.

Make sense.

>
> > > - What happens to the PCM buffer that has been allocated before
> > >   attaching, if it's not the pre-allocated one?
> > >   It should be released properly beforehand, otherwise leaks.
> >
> > I am not sure I understood you correctly. If the PCM buffer has been
> > allocated, the platform driver should handle it? Since we always use
> > substream->dma_buffer.
>
> substream->dma_buffer is merely a pre-allocated buffer, and not every
> driver sets it up.  The actual PCM buffer is found in substream's
> runtime, and this implementation isn't always with the preallocated
> buffer.  It can even be a fixed IO-mapped buffer.

OK. Thanks for your explanation.

>
> > > - There is no validation of the attached dma-buf pages; most drivers
> > >   set coherent DMA mask, and they rely on it.  e.g. if a page over the
> > >   DMA mask is passed, it will break silently.
> >
> > Sorry maybe I did not get your point here. We have validate the
> > dma_map_sg_attr() funtion, in this fucntion it will validate the DMA
> > mask by dma_capable().
>
> OK, then this should be fine -- at least about DMA mask.  But, how
> about other setups, e.g. coherency?
>
> Imagine that a buffer allocated for chip A is exported to another chip
> B.  If chip B requires some special setup of the pages while A isn't,
> this won't work.  For example, some HD-audio chips require the
> non-cached pages while some HD-audio chips allow normal pages.

OK. That's one case need to validate. Thanks for your comments.

-- 
Baolin Wang
Best Regards