Re: [PATCH 0/2][RFC] ALSA: firewire-lib: change packet processing timing in process context

Takashi Sakamoto <o-takashi@xxxxxxxxxxxxx> · Thu, 13 Sep 2018 18:31:34 +0900

Hi,

On Sep 13 2018 06:21, Takashi Iwai wrote:
On Wed, 12 Sep 2018 17:04:48 +0200,
Takashi Sakamoto wrote:

In IEC 61883-1/6 engine in ALSA firewire stack, a work to process packet
is done in three places:
1. bottom-half of handler for interrupts scheduled on 1394OHCI controller
2. a callback of struct 'snd_pcm_ops.pointer()'
3. a callback of struct 'snd_pcm_ops.ack()'

This patchset adds optimization for 3., and remove 2. for efficiency.

In Linux kernel v2.6.39, 1. and 2. was committed in initial implementation
of this engine at a commit 31ef9134eb52('ALSA: add LaCie FireWire
Speakers/Griffin FireWave Surround driver'). This engine schedules hardware
interrupts 16 packets (=INTERRUPT_INTERVAL) thus 1. appears in software IRQ
context every 2 msec (= 16packets * 1sec / 8,000cycle).

In Linux kernel v3.5, an idea of 2. was introduced at a commit
e9148dddc3c7('ALSA: firewire-lib: flush completed packets when reading
PCM position') to reduce timing gap between queueing/dequeueing PCM
frames and scheduling packet transmission. Additionally, it has an
effect to reduce load of the software IRQ context. It's done in a call
of ioctl(2) with SNDRV_PCM_IOCTL_HWSYNC. As a result, in mmap programming
scenario, a call of snd_pcm_hwsync() can process packets for recent
isochronous cycle.

while (1):
   snd_pcm_hwsync()
     ->ioctl(HWSYNC)
       ->struct snd_pcm_ops.pointer()
       = amdtp_stream_pcm_pointer()
         ->fw_iso_context_flush_completions()
   snd_pcm_mmap_begin()
   (queue/dequeue PCM frames here.)
   snd_pcm_mmap_commit()

In Linux kernel v4.13, ALSA PCM core was improved in an aspect of
updating appl_ptr at a commit 9027c4639ef1('ALSA: pcm: Call ack()
whenever appl_ptr is updated'). This commit ensures callback of
'struct snd_pcm_ops.ack()' when updating appl_ptr. This callback
can be done in below ioctl(2) requests:
  - SNDRV_PCM_IOCTL_REWIND
  - SNDRV_PCM_IOCTL_FORWARD
  - SNDRV_PCM_IOCTL_SYNC_PTR
  - SNDRV_PCM_IOCTL_READ[I|N]_FRAMES
  - SNDRV_PCM_IOCTL_WRITE[I|N]_FRAMES

As I notes in a commit 875becf8412c('ALSA: firewire: process packets
in 'struct snd_pcm_ops.ack' callback'), 3. works less efficiently because
in usual PCM operation the above ioctl(2) requests are not used in event
loop.

At the same time of the appl_ptr improvement, a new flag of PCM
hardware definition was added at a commit 42f945970af9('ALSA: pcm: Add
the explicit appl_ptr sync support'), and relevant changes were added
to ALSA PCM core and alsa-lib v1.1.5. This flag expects userspace to
call ioctl(2) with SNDRV_PCM_IOCTL_SYNC_PTR request per iteration of PCM
frame queueing/dequeueing in mmap programming scenario below:

while (1):
   snd_pcm_hwsync()
     ->ioctl(HWSYNC)
       ->struct snd_pcm_ops.pointer()
       = amdtp_stream_pcm_pointer()
         ->fw_iso_context_flush_completions()
   snd_pcm_mmap_begin()
   (queue/dequeue PCM frames here.)
   snd_pcm_mmap_commit()
     ->ioctl(SYNC_PTR)

This patchset aims to support the new flag, SNDRV_PCM_INFO_SYNC_APPLPTR,
to tell userspace to call ioctl(2) with SNDRV_PCM_IOCTL_SYNC_PTR after
queueing/dequeueing PCM frames in each iteration of event dispatcher. As
a result:

while (1):
   snd_pcm_hwsync()
     ->ioctl(HWSYNC)
       ->struct snd_pcm_ops.pointer()
       = amdtp_stream_pcm_pointer()
         ->fw_iso_context_flush_completions()
   snd_pcm_mmap_begin()
   (queue/dequeue PCM frames here.)
   snd_pcm_mmap_commit()
     ->ioctl(SYNC_PTR)
       ->struct snd_pcm_ops.ack()
       = amdtp_stream_pcm_ack()
         ->fw_iso_context_flush_completions()

This patchset also adds optimization for timing to process packets in
process context. As the above call graph shows, support of the flag brings
closest two calls to process packet against intervals of isochronous cycle
and it is not enough efficient. This patchset also cancel packet processing
in .pointer() callback. As a result:

while (1):
   snd_pcm_hwsync()
   snd_pcm_mmap_begin()
   (queue/dequeue PCM frames here.)
   snd_pcm_mmap_commit()
     ->ioctl(SYNC_PTR)
       ->struct snd_pcm_ops.ack()
       = amdtp_stream_pcm_ack()
         ->fw_iso_context_flush_completions()

Takashi Sakamoto (2):
   ALSA: firewire-lib: adopt SYNC_APPLPTR flag to flush packets after
     queueing/dequeueing PCM frames
   ALSA: firewire-lib: cancel processing packets in struct
     snd_pcm_ops.pointer() call

Both look like a good improvement.  Looking forward to seeing the
official patches once after tests by others pass.

Thanks for your positive feedback.

To ease the others to test this patchset, I prepare for a remote branch
in github.com.
https://github.com/takaswie/snd-firewire-improve/tree/topic/applptr

From the branch, whole codes relevant to ALSA firewire stack can be
backported to Linux kernel v4.17 or later.

This patchset allows userspace applications to queue isochronous packet
with PCM frames without waiting for software IRQ of 1394 OHCI
controller. In this development cycle I have a plan to allow
applications to have a smaller PCM intermediate buffer than current one
(=period-time greater than 5msec). Users look it as 'improvement in a
point of latency', if things go well.

Thanks

Takashi Sakamoto
_______________________________________________
Alsa-devel mailing list
Alsa-devel@xxxxxxxxxxxxxxxx
http://mailman.alsa-project.org/mailman/listinfo/alsa-devel