Date:   Fri, 20 Oct 2023 13:33:50 +0200
From:   Takashi Iwai <tiwai@...e.de>
To:     Matias Ezequiel Vara Larsen <mvaralar@...hat.com>
Cc:     Anton Yakovlev <anton.yakovlev@...nsynergy.com>, mst@...hat.com,
        perex@...ex.cz, tiwai@...e.com,
        virtualization@...ts.linux-foundation.org,
        alsa-devel@...a-project.org, linux-kernel@...r.kernel.org,
        pbonzini@...hat.com, stefanha@...hat.com, sgarzare@...hat.com,
        manos.pitsidianakis@...aro.org, mripard@...hat.com
Subject: Re: [PATCH v2] ALSA: virtio: use copy and fill_silence callbacks

On Fri, 20 Oct 2023 11:45:39 +0200,
Matias Ezequiel Vara Larsen wrote:
> 
> Hello Takashi,
> 
> On Thu, Oct 19, 2023 at 09:48:03AM +0200, Takashi Iwai wrote:
> > On Thu, 19 Oct 2023 03:20:19 +0200,
> > Anton Yakovlev wrote:
> > > 
> > > Hi Takashi,
> > > 
> > > On 19.10.2023 03:07, Takashi Iwai wrote:
> > > > On Wed, 18 Oct 2023 12:48:23 +0200,
> > > > Matias Ezequiel Vara Larsen wrote:
> > > >> 
> > > >> This commit replaces the mmap mechanism with the copy() and
> > > >> fill_silence() callbacks for both capturing and playback for the
> > > >> virtio-sound driver. This change is required to prevent the updating of
> > > >> the content of a buffer that is already in the available ring.
> > > >> 
> > > >> The current mechanism splits a dma buffer into descriptors that are
> > > >> exposed to the device. This dma buffer is shared with the user
> > > >> application. When the device consumes a buffer, the driver moves the
> > > >> request from the used ring to available ring.
> > > >> 
> > > >> The driver exposes the buffer to the device without knowing if the
> > > >> content has been updated from the user. The section 2.8.21.1 of the
> > > >> virtio spec states that: "The device MAY access the descriptor chains
> > > >> the driver created and the memory they refer to immediately". If the
> > > >> device picks up buffers from the available ring just after it is
> > > >> notified, it happens that the content may be old.
> > > >> 
> > > >> By providing the copy() callback, the driver first updates the content
> > > >> of the buffer, and then, exposes the buffer to the device by enqueuing
> > > >> it in the available ring. Thus, device always picks up a buffer that is
> > > >> updated. During copy(), the number of requests enqueued depends on the
> > > >> "pos" and "bytes" arguments. The length of each request is period_size
> > > >> bytes.
> > > >> 
> > > >> For capturing, the driver starts by exposing all the available buffers
> > > >> to device. After device updates the content of a buffer, it enqueues it
> > > >> in the used ring. It is only after the copy() for capturing is issued
> > > >> that the driver re-enqueues the buffer in the available ring.
> > > >> 
> > > >> Co-developed-by: Anton Yakovlev <anton.yakovlev@...nsynergy.com>
> > > >> Signed-off-by: Matias Ezequiel Vara Larsen <mvaralar@...hat.com>
> > > >> ---
> > > >> Changelog:
> > > >> v1 -> v2:
> > > >>   * Use snd_pcm_set_managed_buffer_all()for buffer allocation/freeing.
> > > >>   * Make virtsnd_pcm_msg_send() generic by specifying the offset and size
> > > >>     for the modified part of the buffer; this way no assumptions need to
> > > >>     be made.
> > > >>   * Disable SNDRV_PCM_INFO_NO_REWINDS since now only sequential
> > > >>     reading/writing of frames is supported.
> > > >>   * Correct comment at virtsnd_pcm_msg_send().
> > > >>   * v1 patch at:
> > > >>     https://ddec1-0-en-ctp.trendmicro.com:443/wis/clicktime/v1/query?url=https%3a%2f%2flore.kernel.org%2flkml%2f20231016151000.GE119987%40fedora%2ft%2f&umid=2f305b77-83e7-47b6-a461-a8ca67d0bfe2&auth=53c7c7de28b92dfd96e93d9dd61a23e634d2fbec-2d5775265e7e1741ae8eb783a3cb78ed553093c1
> > > >> 
> > > >>   sound/virtio/virtio_pcm.c     |  7 ++-
> > > >>   sound/virtio/virtio_pcm.h     |  9 ++--
> > > >>   sound/virtio/virtio_pcm_msg.c | 93 ++++++++++++++++++++++-------------
> > > >>   sound/virtio/virtio_pcm_ops.c | 81 +++++++++++++++++++++++++-----
> > > >>   4 files changed, 137 insertions(+), 53 deletions(-)
> > > > 
> > > > Most of the code changes look good, but I wonder:
> > > > 
> > > >> 
> > > >> diff --git a/sound/virtio/virtio_pcm.c b/sound/virtio/virtio_pcm.c
> > > >> index c10d91fff2fb..66d67eef1bcc 100644
> > > >> --- a/sound/virtio/virtio_pcm.c
> > > >> +++ b/sound/virtio/virtio_pcm.c
> > > >> @@ -104,12 +104,11 @@ static int virtsnd_pcm_build_hw(struct virtio_pcm_substream *vss,
> > > >>   	 * only message-based transport.
> > > >>   	 */
> > > >>   	vss->hw.info =
> > > >> -		SNDRV_PCM_INFO_MMAP |
> > > >> -		SNDRV_PCM_INFO_MMAP_VALID |
> > > > 
> > > > Do we need the removal of those MMAP features inevitably?
> > > > Usually mmap can still work even if the driver implements the copy
> > > > ops.  Those aren't always mutual exclusive.
> > > 
> > > The driver uses a message queue to communicate with the device. Thus,
> > > the audio buffer is sliced into several I/O requests (= number of
> > > periods) of the same size (= period size).
> > > 
> > > Before this, all such requests were enqueued when the substream started,
> > > and immediately re-enqueued once the request is completed. This approach
> > > made it possible to add mmap support. But for mmap there are no explicit
> > > notifications from the application how many frames were written or read.
> > > Thus, it was assumed that the virtual device should read/write frames to
> > > requests based on timings. And there are some problems here:
> > > 
> > >   1. This was found to violate the virtio specification: if a request is
> > >      already in the queue, the device can safely read/write there at any
> > >      time.
> > >   2. It looks like this breaks the use case with swiotlb. Personally I'm
> > >      not sure how the application handles DMA ownership in the case of
> > >      mmaped buffer.
> > > 
> > > To correctly implement mmap support, instead of transferring data via a
> > > message queue, the driver and device must have a shared memory region.
> > > We can add mmap in the future when we expand the functionality of the
> > > device to support such shared memory.
> > 
> > Ah, then this implementation might be an overkill.  You're still using
> > the (intermediate) vmalloc buffer allocated via PCM managed mode, and
> > the actual data is copied from/to there.  So it doesn't conflict with
> > the mmap operation at all.
> > 
> > I guess that the problem you're trying to solve (the immediate data
> > transfer to the queue) can be implemented rather via PCM ack callback
> > instead.  ALSA PCM core notifies the possible data transfer via PCM
> > ack callback right after each change of appl_ptr or hw_ptr, including
> > each read/write op or mmap commit.  Then the driver can check the
> > change of appl_ptr (or hw_ptr for capture), fetch the newly available
> > data, and queue it immediately.
> > 
> > Usually together with the use of ack callback, the driver sets
> > SNDRV_PCM_INFO_SYNC_APPLPTR flag.  This prevents the mmap of the PCM
> > control record (not the audio data) and enforces the use of
> > SNDRV_PCM_IOCTL_SYNC_PTR ioctl instead (so that the driver always gets
> > the ack callback).
> > 
> > 
> 
> Thanks for your comments. If I understand correctly, we have two
> options:
> 1. Use copy/fill_silence callbacks and use my own buffers thus disabling
> mmap.
> 2. Use mmap and the ack callback to track when appl_ptr changes thus
> moving the content to the queues after it has been updated.
> 
> Am I right? 

Not really.  The ack callback is always effective no matter whether
the stream is transferred via mmap or read/write.  So, when you
implement a la packet transfer via ack callback, it works for both
mmap and read/write.  It's used typically for drivers doing a
packet-based transfer, e.g. FireWire drivers or USB-audio.

OTOH, if you'd like to drop the intermediate ring buffer (enabled by
snd_pcm_set_managed_buffer_all() and allocated/freed automatically at
hw_params/hw_free in PCM core), you can implement PCM copy and
fill_ops methods.  But in this case, you can't copy to
runtime->dma_area like your patch; you may need a temporary buffer of
your own, or handle directly the user pointer.

And, even with copy/fill_ops implementations, mmap itself isn't
prohibited when the driver provides its own mmap call.  This can be
seen, for example, when a driver uses the iomem region as the data
ring buffer.


HTH,

Takashi


> Thanks, Matias.
> 
> > thanks,
> > 
> > Takashi
> > 
> > 
> > > 
> > > 
> > > Best regards,
> > > 
> > > > 
> > > > 
> > > > thanks,
> > > > 
> > > > Takashi
> > > 
> > > -- 
> > > Anton Yakovlev
> > > Senior Software Engineer
> > > 
> > > OpenSynergy GmbH
> > > Rotherstr. 20, 10245 Berlin
> > > 
> > 
> 

Powered by blists - more mailing lists