Dne 22. 06. 20 v 11:31 Mark Hills napsal(a):
On Fri, 19 Jun 2020, Takashi Iwai wrote:
The recent overlooked bug about the unconditional semaphore usage in the dmix implies that basically we've had no problem with the locking in the practical usages over years. Although the lockless operation has a clear merit, it's a much higher CPU usage (especially on some uncached pages), and it might lead to a potential deadlock in theory (which is hard to reproduce at will, though).
This patch introduces a new configure option "--enable-lockless-dmix" or "--disable-lockless-dmix" to let user choose the default dmix operation mode, then the default value for the dmix "direct_memory_access" option is set based on it.
A big question is which operation mode should be default: it's hard to decide, but in this patch, I bet for disabling the lockless in the end as the performance loss is significant.
But the user can enable the lockless operation at any time; at build time, with the configure option above, and at run time, by specifying the dmix "direct_memory_access" option, too.
I would like to express some caution here.
Seems it must be essential (not just choice) that the semaphore implementation is the default. As per Maarten's information, there are libasound embedded in applications and containers. Differing defaults results in broken audio between applications.
Sadly there is, in effect, an ABI here; a practical risk that users suffering the consequences eventually.
Because of the history of sepahores, an application would need to signal its intent to use atomics, which is not a good thing as that is complex.
Instead I think it is smart here to consider the opportunity which Maarten has come here with.
Patches here are just the beginning to bring alive a lot of dormant functionality. It assumes hand-written assembly code will run concurrently that appears to not yet have been tested in that way. It is a joy to see hand written assembly, but my worry is that is influencing the decision making.
I am only recently looking at dmix/snoop code, so I do not aim to stand in the way. But I think it would be prudent to consider that bringing alive dormant functionality (vs. opportunity to remove code) as if it were adding the code explicitly. Would ALSA developers review and accept a 1000+ line patch adding architecture-specific assembly, changes to the ABI, based on the benchmarks which Maarten has presented?
Where I am more certain is: if options are to be provided to users then it should be because a user is in the best position to decide. In this case I think ALSA developers must equip users in understanding the pros/cons. That's why ideally there's no option and code just does the right thing. If not, at very least documentation must explain the tradeoff (and I think a better name should be chosen.)
I can certainly see interesting positives for mixing based on atomics. But there are many years without it, and this feels hasty and there are risks.
And Maarten has presented and benchmarked an excellent opportunity to simplify, which could be missed. It is one thing to leave the code dormant until a decision or clearer picture. But these patches risk meeting that opportunity and transforming it into complexity for developers and users.
Some comments: This code was designed when we worked with SMP machines with 2 CPUs (not cores - physical CPUs). It's old, but I feel it's worth to keep it at least for the reference and testing. The one instruction racy time window described by Maarten seems true, but as I wrote, the errors were mostly zero (no hearable).
The locked variant (without atomic instructions) should be the default without any questions. It's more effective for the multi-cpu-core architectures used nowadays.
I would propose this:
1) use the locked variant as default (no atomic code even for x86) 2) add a configuration option to select the mixing code 3) change SND_PCM_DIRECT_MAGIC which will follow the mixing code selection to avoid locking / atomic mixing code mismatch / misuse
Jaroslav