On 25.01.2021 17:59, Guennadi Liakhovetski wrote:
> On Sun, 24 Jan 2021, Anton Yakovlev wrote:
>
> [snip]
>
>> +/**
>> + * virtsnd_pcm_release() - Release the PCM substream on the device
side.
>> + * @substream: VirtIO substream.
>> + *
>> + * Context: Any context that permits to sleep.
>> + * Return: 0 on success, -errno on failure.
>> + */
>> +static inline bool virtsnd_pcm_released(struct virtio_pcm_substream
>> *substream)
>> +{
>> + /*
>> + * The spec states that upon receipt of the RELEASE command "the
>> device
>> + * MUST complete all pending I/O messages for the specified
>> stream ID".
>> + * Thus, we consider the absence of I/O messages in the queue
as an
>> + * indication that the substream has been released.
>> + */
>> + return atomic_read(&substream->msg_count) == 0;
>
> Also here having it atomic doesn't really seem to help. This just means,
> that at some point of time it was == 0.
Technically, you're right. In practice, everything looks like this:
I/O messages are added to the virtqueue either at the start of the
substream or in the interrupt handler (and only as long as .xfer_enabled
is true). In general, this means that the .msg_count can only be
incremented in the interrupt handler. As soon as the substream stops,
the .xfer_enabled becomes false and the .msg_count no longer increases.
This means that the .msg_count was either already 0, or we need to wait
for it to become 0.
>> +}
>> +
>> +static int virtsnd_pcm_release(struct virtio_pcm_substream *substream)
>
> kernel-doc missing
Yeap, thanks!
>> +{
>> + struct virtio_snd *snd = substream->snd;
>> + struct virtio_snd_msg *msg;
>> + unsigned int js = msecs_to_jiffies(msg_timeout_ms);
>> + int rc;
>> +
>> + msg = virtsnd_pcm_ctl_msg_alloc(substream,
>> VIRTIO_SND_R_PCM_RELEASE,
>> + GFP_KERNEL);
>> + if (IS_ERR(msg))
>> + return PTR_ERR(msg);
>> +
>> + rc = virtsnd_ctl_msg_send_sync(snd, msg);
>> + if (rc)
>> + return rc;
>> +
>> + return wait_event_interruptible_timeout(substream->msg_empty,
>> +
>> virtsnd_pcm_released(substream),
>> + js);
>
> wait_event_interruptible_timeout() will return a positive number in
> success cases, 0 means a timeout and condition still false. Whereas when
> you call this function you interpret 0 as success and you expect any != 0
> to be a negative error. Wondering how this worked during your tests?
Yeah, that's actually a bug. We haven't hit a timeout on that control path.
>> +}
>> +
>> +/**
>> + * virtsnd_pcm_open() - Open the PCM substream.
>> + * @substream: Kernel ALSA substream.
>> + *
>> + * Context: Any context.
>> + * Return: 0 on success, -errno on failure.
>> + */
>> +static int virtsnd_pcm_open(struct snd_pcm_substream *substream)
>> +{
>> + struct virtio_pcm *pcm = snd_pcm_substream_chip(substream);
>> + struct virtio_pcm_substream *ss = NULL;
>> +
>> + if (pcm) {
>> + switch (substream->stream) {
>> + case SNDRV_PCM_STREAM_PLAYBACK:
>> + case SNDRV_PCM_STREAM_CAPTURE: {
>> + struct virtio_pcm_stream *stream =
>> + &pcm->streams[substream->stream];
>> +
>> + if (substream->number < stream->nsubstreams)
>
> Can this condition ever be false?
Hard to tell. But there may be some bug. In general, I try to adhere to
the rule that if an array element is referenced by index, it is better
to check the index value first.
>> + ss =
stream->substreams[substream->number];
>> + break;
>> + }
>> + }
>> + }
>> +
>> + if (!ss)
>> + return -EBADFD;
>> +
>> + substream->runtime->hw = ss->hw;
>> + substream->private_data = ss;
>> +
>> + return 0;
>> +}
>> +
>> +/**
>> + * virtsnd_pcm_close() - Close the PCM substream.
>> + * @substream: Kernel ALSA substream.
>> + *
>> + * Context: Any context.
>> + * Return: 0.
>> + */
>> +static int virtsnd_pcm_close(struct snd_pcm_substream *substream)
>> +{
>> + return 0;
>> +}
>> +
>> +/**
>> + * virtsnd_pcm_hw_params() - Set the parameters of the PCM substream.
>> + * @substream: Kernel ALSA substream.
>> + * @hw_params: Hardware parameters (can be NULL).
>> + *
>> + * The function can be called both from the upper level (in this case,
>> + * @hw_params is not NULL) or from the driver itself (in this case,
>> @hw_params
>> + * is NULL, and the parameter values are taken from the runtime
>> structure).
>> + *
>> + * In all cases, the function:
>> + * 1. checks the state of the virtqueue and, if necessary, tries to
>> fix it,
>> + * 2. sets the parameters on the device side,
>> + * 3. allocates a hardware buffer and I/O messages.
>> + *
>> + * Context: Any context that permits to sleep.
>> + * Return: 0 on success, -errno on failure.
>> + */
>> +static int virtsnd_pcm_hw_params(struct snd_pcm_substream *substream,
>> + struct snd_pcm_hw_params *hw_params)
>> +{
>> + struct snd_pcm_runtime *runtime = substream->runtime;
>> + struct virtio_pcm_substream *ss =
>> snd_pcm_substream_chip(substream);
>> + struct virtio_device *vdev = ss->snd->vdev;
>> + struct virtio_snd_msg *msg;
>> + struct virtio_snd_pcm_set_params *request;
>> + snd_pcm_format_t format;
>> + unsigned int channels;
>> + unsigned int rate;
>> + unsigned int buffer_bytes;
>> + unsigned int period_bytes;
>> + unsigned int periods;
>> + unsigned int i;
>> + int vformat = -1;
>> + int vrate = -1;
>> + int rc;
>> +
>> + /*
>> + * If we got here after ops->trigger() was called, the queue may
>> + * still contain messages. In this case, we need to release the
>> + * substream first.
>> + */
>> + if (atomic_read(&ss->msg_count)) {
>> + rc = virtsnd_pcm_release(ss);
>> + if (rc) {
>> + dev_err(&vdev->dev,
>> + "SID %u: invalid I/O queue state\n",
>> + ss->sid);
>> + return rc;
>> + }
>> + }
>> +
>> + /* Set hardware parameters in device */
>> + if (hw_params) {
>> + format = params_format(hw_params);
>> + channels = params_channels(hw_params);
>> + rate = params_rate(hw_params);
>> + buffer_bytes = params_buffer_bytes(hw_params);
>> + period_bytes = params_period_bytes(hw_params);
>> + periods = params_periods(hw_params);
>> + } else {
>> + format = runtime->format;
>> + channels = runtime->channels;
>> + rate = runtime->rate;
>> + buffer_bytes = frames_to_bytes(runtime,
>> runtime->buffer_size);
>> + period_bytes = frames_to_bytes(runtime,
>> runtime->period_size);
>> + periods = runtime->periods;
>> + }
>> +
>> + for (i = 0; i < ARRAY_SIZE(g_a2v_format_map); ++i)
>> + if (g_a2v_format_map[i].alsa_bit == format) {
>> + vformat = g_a2v_format_map[i].vio_bit;
>> +
>> + break;
>> + }
>> +
>> + for (i = 0; i < ARRAY_SIZE(g_a2v_rate_map); ++i)
>> + if (g_a2v_rate_map[i].rate == rate) {
>> + vrate = g_a2v_rate_map[i].vio_bit;
>> +
>> + break;
>> + }
>> +
>> + if (vformat == -1 || vrate == -1)
>> + return -EINVAL;
>> +
>> + msg = virtsnd_pcm_ctl_msg_alloc(ss, VIRTIO_SND_R_PCM_SET_PARAMS,
>> + GFP_KERNEL);
>> + if (IS_ERR(msg))
>> + return PTR_ERR(msg);
>> +
>> + request = sg_virt(&msg->sg_request);
>> +
>> + request->buffer_bytes = cpu_to_virtio32(vdev, buffer_bytes);
>> + request->period_bytes = cpu_to_virtio32(vdev, period_bytes);
>> + request->channels = channels;
>> + request->format = vformat;
>> + request->rate = vrate;
>
> I presume the latter three fields don't have to be endienness-converted,
> perhaps they're 8-bit wide only.
Yes, these three values are u8.
>> +
>> + if (ss->features & (1U << VIRTIO_SND_PCM_F_MSG_POLLING))
>> + request->features |=
>> + cpu_to_virtio32(vdev,
>> + 1U <<
>> VIRTIO_SND_PCM_F_MSG_POLLING);
>> +
>> + if (ss->features & (1U << VIRTIO_SND_PCM_F_EVT_XRUNS))
>> + request->features |=
>> + cpu_to_virtio32(vdev,
>> + 1U << VIRTIO_SND_PCM_F_EVT_XRUNS);
>> +
>> + rc = virtsnd_ctl_msg_send_sync(ss->snd, msg);
>
> Wouldn't it be better to only try to send the message after below
> allocations completed successfully?
I thought the reverse logic was better. This message asks the device to
set a specific set of parameters. And if the device returned an error
for some reason, then there is no point in allocating memory.
>> + if (rc)
>> + return rc;
>> +
>> + /* If the buffer was already allocated earlier, do nothing. */
>> + if (runtime->dma_area)
>> + return 0;
>> +
>> + /* Allocate hardware buffer */
>> + rc = snd_pcm_lib_malloc_pages(substream, buffer_bytes);
>> + if (rc < 0)
>> + return rc;
>> +
>> + /* Allocate and initialize I/O messages */
>> + rc = virtsnd_pcm_msg_alloc(ss, periods, runtime->dma_area,
>> + period_bytes);
>> + if (rc)
>> + snd_pcm_lib_free_pages(substream);
>> +
>> + return rc;
>> +}
>> +
>> +/**
>> + * virtsnd_pcm_hw_free() - Reset the parameters of the PCM substream.
>> + * @substream: Kernel ALSA substream.
>> + *
>> + * The function does the following:
>> + * 1. tries to release the PCM substream on the device side,
>> + * 2. frees the hardware buffer.
>> + *
>> + * Context: Any context that permits to sleep.
>> + * Return: 0 on success, -errno on failure.
>> + */
>> +static int virtsnd_pcm_hw_free(struct snd_pcm_substream *substream)
>> +{
>> + struct virtio_pcm_substream *ss =
>> snd_pcm_substream_chip(substream);
>> + int rc;
>> +
>> + rc = virtsnd_pcm_release(ss);
>> +
>> + /*
>> + * Even if we failed to send the RELEASE message or wait for the
>> queue
>> + * flush to complete, we can safely delete the buffer. Because
>> after
>> + * receiving the STOP command, the device must stop all I/O
message
>> + * processing. If there are still pending messages in the queue,
>> the
>> + * next ops->hw_params() call should deal with this.
>> + */
>> + snd_pcm_lib_free_pages(substream);
>> +
>> + return rc;
>> +}
>> +
>> +/**
>> + * virtsnd_pcm_hw_params() - Prepare the PCM substream.
>
> copy-paste: this is virtsnd_pcm_prepare()
Oops... :)
>> + * @substream: Kernel ALSA substream.
>> + *
>> + * The function can be called both from the upper level or from the
>> driver
>> + * itself.
>> + *
>> + * In all cases, the function:
>> + * 1. checks the state of the virtqueue and, if necessary, tries to
>> fix it,
>> + * 2. prepares the substream on the device side.
>> + *
>> + * Context: Any context that permits to sleep. May take and release
>> the tx/rx
>> + * queue spinlock.
>> + * Return: 0 on success, -errno on failure.
>> + */
>> +static int virtsnd_pcm_prepare(struct snd_pcm_substream *substream)
>> +{
>> + struct virtio_pcm_substream *ss =
>> snd_pcm_substream_chip(substream);
>> + struct virtio_snd_queue *queue = virtsnd_pcm_queue(ss);
>> + struct virtio_snd_msg *msg;
>> + unsigned long flags;
>> + int rc;
>> +
>> + /*
>> + * If we got here after ops->trigger() was called, the queue may
>> + * still contain messages. In this case, we need to reset the
>> + * substream first.
>> + */
>> + if (atomic_read(&ss->msg_count)) {
>> + rc = virtsnd_pcm_hw_params(substream, NULL);
>> + if (rc)
>> + return rc;
>> + }
>> +
>> + spin_lock_irqsave(&queue->lock, flags);
>> + ss->msg_last_enqueued = -1;
>> + spin_unlock_irqrestore(&queue->lock, flags);
>> +
>> + /*
>> + * Since I/O messages are asynchronous, they can be completed
>> + * when the runtime structure no longer exists. Since each
>> + * completion implies incrementing the hw_ptr, we cache all the
>> + * current values needed to compute the new hw_ptr value.
>> + */
>> + ss->frame_bytes = substream->runtime->frame_bits >> 3;
>> + ss->period_size = substream->runtime->period_size;
>> + ss->buffer_size = substream->runtime->buffer_size;
>> +
>> + atomic_set(&ss->hw_ptr, 0);
>> + atomic_set(&ss->xfer_xrun, 0);
>> + atomic_set(&ss->msg_count, 0);
>> +
>> + msg = virtsnd_pcm_ctl_msg_alloc(ss, VIRTIO_SND_R_PCM_PREPARE,
>> + GFP_KERNEL);
>> + if (IS_ERR(msg))
>> + return PTR_ERR(msg);
>> +
>> + return virtsnd_ctl_msg_send_sync(ss->snd, msg);
>> +}
>> +
>> +/**
>> + * virtsnd_pcm_trigger() - Process command for the PCM substream.
>> + * @substream: Kernel ALSA substream.
>> + * @command: Substream command (SNDRV_PCM_TRIGGER_XXX).
>> + *
>> + * Depending on the command, the function does the following:
>> + * 1. enables/disables data transmission,
>> + * 2. starts/stops the substream on the device side.
>> + *
>> + * Context: Atomic context. May take and release the tx/rx queue
>> spinlock.
>
> Really? Cannot .trigger() sleep? E.g. I see mdelay(25) in
> snd_es18xx_playback1_trigger()
Actually, you made a good point here. I didn't know, that it is possible
to disable atomic mode for that callback. But, apparently, it is possible.
And virtio pcm definetly is nonatomic. I need to redo this code, thanks!
>> + * Return: 0 on success, -errno on failure.
>> + */
>> +static int virtsnd_pcm_trigger(struct snd_pcm_substream *substream,
>> int command)
>> +{
>> + struct virtio_pcm_substream *ss =
>> snd_pcm_substream_chip(substream);
>> + struct virtio_snd *snd = ss->snd;
>> + struct virtio_snd_queue *queue = virtsnd_pcm_queue(ss);
>> + struct virtio_snd_msg *msg;
>> +
>> + switch (command) {
>> + case SNDRV_PCM_TRIGGER_START:
>> + case SNDRV_PCM_TRIGGER_PAUSE_RELEASE: {
>> + int rc;
>> +
>> + spin_lock(&queue->lock);
>> + rc = virtsnd_pcm_msg_send(ss);
>> + spin_unlock(&queue->lock);
>
> Maybe it would be good to explain why locking is required here and isn't
> required in most other locations, where messages are sent?
There are two kinds of messages here: control messages and I/O. Functions
for sending control message acquire and release the control virtqueue
spinlock on their own. But we cannot do the same for I/O messages, since
virtsnd_pcm_msg_send is also called from the interrupt handler, which is
already grabbing the lock for the I/O virtqueue.
> Thanks
> Guennadi
>
>> + if (rc)
>> + return rc;
>> +
>> + atomic_set(&ss->xfer_enabled, 1);
>> +
>> + msg = virtsnd_pcm_ctl_msg_alloc(ss,
VIRTIO_SND_R_PCM_START,
>> + GFP_ATOMIC);
>> + if (IS_ERR(msg))
>> + return PTR_ERR(msg);
>> +
>> + return virtsnd_ctl_msg_send(snd, msg);
>> + }
>> + case SNDRV_PCM_TRIGGER_STOP:
>> + case SNDRV_PCM_TRIGGER_PAUSE_PUSH: {
>> + atomic_set(&ss->xfer_enabled, 0);
>> +
>> + msg = virtsnd_pcm_ctl_msg_alloc(ss, VIRTIO_SND_R_PCM_STOP,
>> + GFP_ATOMIC);
>> + if (IS_ERR(msg))
>> + return PTR_ERR(msg);
>> +
>> + return virtsnd_ctl_msg_send(snd, msg);
>> + }
>> + default: {
>> + return -EINVAL;
>> + }
>> + }
>> +}
--
Anton Yakovlev
Senior Software Engineer
OpenSynergy GmbH
Rotherstr. 20, 10245 Berlin
www.opensynergy.com
