cs46xx: once again SMP fixes, new aproach
PATCH 1/1 cs46xx: once again SMP fixes, new aproach
Hi!
This was hard nut break with good solution, with this solution I can only find two issues:
1 - for the moment I'm totally convinced that it's a excelent solution -;) 2- as I dont have any SMP machine (yet) I cant test it, untested code is condemned to fail.
I will try to explain my theory, this diagram shows the SCB tree structure in overview:
FGTaskTree
¦
|
+-----------------+ +-----------------+
| TimingMasterSCB |---next_scb-- | WriteBackSCB |--nextSCB ... (misc functionality)
+-----------------+ +-----------------+
| |
sub_list_ptr (sub tree used for capture)
| ...
+-----------------+ | CodecOutSCB |--nullSCB
+-----------------+
|
sub_list_ptr
|
+-----------------+ | MasterMixSCB |--nullSCB
+-----------------+
|
sub_list_ptr
|
+-----------------+ +-----------------+ +-----------------+
| SRCTask_SPDIF |---nextSCB---| SRCTask_SCB0 |--nextSCB--| SRCTask_SCB1 |- ... (etc)
+-----------------+ +-----------------+ +-----------------+
|
| |
sub_list_ptr sub_list_ptr
sub_list_ptr
|
| |
+-----------------+ +------------+ +-----------+ +-----------+ +-----------+
| AsynchFGRxSCB | | PCMReader1 |-nextSCB-| PCMReader2| | PCMReader3|-nextSCB-| PCMReader4|-nextSCB- ... (etc)
+-----------------+ +------------+ +-----------+ +-----------+ +-----------+
| |
| | |
nullSCB nullSCB
nullSCB nullSCB nullSCB(i hope you have a wide terminal and fixed font width)
- The pcm_link/unlink only access SCB under it's related SRCTask_SCB*, it only touches the nex_scb_ptr, parent_ptr and sub_list_ptr on SRCTask.. so it can safely be running paralel with any other simultaneous operation on other parts of the tree, and it can be running paralell with others pcm_link/unlink related to other SRCTask*.
- In create_pcm_channel the channel is initially created unlinked, which mean pcm_link must be called before playback start, invoked by playback_trigger. Anyway pcm_unlink was called in playback_open, which now can be removed. In this way the pcm_reader SCB will be initially floating, with no parent SCB and therefor we dont need any spin_lock in create_pcm_channel.
- But, the _same_ PCM stream, playback_trigger cant run before playback_open has completed. Which make no sense, I supose the ALSA PCM core prevent a process from doing this to happen.
- In pcm_link the PCMReaderSCB is always linked in after SRCTask* sub_list_ptr, then therefor is no need loop searching for the last free SCB in sublist.
Conclusion what we got is:
- No mutexes in playback_trigger
- Independent spin_locks for each SRCTask*
- spin_lock only needed in pcm_link/unlink and in cs46xx_dsp_remove_scb.
With fewer and short spin_locks, theorically we manage to have good performance on SMP machines -:)
That's all, ... i think ...
Question, suggestion or whatever you got in mind about this is welcome, i'll try to be open.
/Benny
At Thu, 22 Aug 2002 23:29:20 +0200,
Quoted 15 lines · Benny Sjostrand
- The pcm_link/unlink only access SCB under it's related SRCTask_SCB*, it only touches the
nex_scb_ptr,parent_ptrandsub_list_ptron SRCTask.. so it can safely be running paralel with any other simultaneous operation on other parts of the tree, and it can be running paralell with others pcm_link/unlink related to other SRCTask*.- In
create_pcm_channelthe channel is initially created unlinked, which meanpcm_linkmust be called before playback start, invoked byplayback_trigger. Anywaypcm_unlinkwas called inplayback_open, which now can be removed. In this way thepcm_readerSCB will be initially floating, with no parent SCB and therefor we dont need anyspin_lockincreate_pcm_channel.
Quoted 3 lines
- But, the _same_ PCM stream,
playback_triggercant run beforeplayback_openhas completed. Which make no sense, I supose the ALSA PCM core prevent a process from doing this to happen.
this must be ok.
Quoted 3 lines
- In
pcm_linkthe PCMReaderSCB is always linked in after SRCTask*sub_list_ptr, then therefor is no need loop searching for the last free SCB in sublist.
good, that's somewhat i thought of.
Quoted 4 lines
Conclusion what we got is:
- No mutexes inplayback_trigger
- Independentspin_locksfor each SRCTask*
-spin_lockonly needed in pcm_link/unlink and incs46xx_dsp_remove_scb.
great!
the patch looks more cleaner than before.
Quoted 6 lines
With fewer and short
spin_locks, theorically we manage to have good performance on SMP machines -:)That's all, ... i think ...
Question, suggestion or whatever you got in mind about this is welcome, i'll try to be open.
i'll try to run and merge to cvs later.
thanks!
Takashi
-------------------------------------------------------
This sf.net email is sponsored by: OSDN - Tired of that same old cell phone? Get a new here for FREE!
https://www.inphonic.com/r.asp?r=sourceforge1&refcode1=vs3390