hda-dma: exit L1 only once for all DMA channels #2664
Conversation
src/drivers/intel/cavs/hda-dma.c
Outdated
There was a problem hiding this comment.
I'd prefer to have this notifier related piece in if (hda_chan->irq_disabled) rather than double return (just mho).
There was a problem hiding this comment.
It's kernel like approach to avoid large indentation. Most of the reviewers seem to like it this way.
tlauda
force-pushed
the
topic/host-l1-exit-move
branch
from
25313a6 to
586b0f9
Compare
src/schedule/ll_schedule.c
Outdated
There was a problem hiding this comment.
Why are we doing this in the scheduler ?
There was a problem hiding this comment.
It can be done once for all DMA channels, so no point in doing it after every DMA copy separately. We're just wasting cycles. With timer driven pipelines we are sure that it will be executed only once for all channels.
There was a problem hiding this comment.
Ok, but I'm not following the usage here of L1 exit and when we need to use it. i.e. does it apply to all channels together or individually, do we exit L1 after DMA has stopped etc. Can you outline the rules a little more. Thanks
There was a problem hiding this comment.
It applies to all channels on all cores together and we should exit L1 after DMA pointer increment. The usage is exactly the same as before, but now it only happens once after all pipeline_copy tasks has been executed and not individually after every DMA channel copy.
There was a problem hiding this comment.
This PM L1 entry/exit should be handled directly by the HDA-DMA driver and not via LL scheduler state changes (as these can be done for any task unrelated to HDA DMA state). What's stopping us running an unrelated task via the LL scheduler during HDA work when we need to be in L1.
I'm also not following why we can't reference count pm_runtime_get(L1)/put(L1) within the HDA code ?
There was a problem hiding this comment.
@lgirdwood Not true. If there is no LL work that uses HDA DMA, there will be no L1 entry/exit.
There was a problem hiding this comment.
But wont any LL work invoke the notifier? This is regardless of HDA state, it's then a race between HDA pipelines and any LL work.
There was a problem hiding this comment.
They are invoked before processing list of LL tasks and after. Not before and after every single task.
There was a problem hiding this comment.
ok, so my question is still the same, but rather than after/before every LL task, it's running the LL scheduler. Wont this put restrictions on when the scheduler can be used ? i.e. code that wants to schedule work on the LL scheduler will have to know about the state of any pipelines with HDA DMA ?
There was a problem hiding this comment.
We are keeping ref count, so no special knowledge is needed. It should happen automatically.
|
@lgirdwood KWD tests are failing due to the KPB drain timing issue. I think @mrajwa should have some quick fix soon. |
Adds possibility of defining flags, when performing notifier registration. Signed-off-by: Tomasz Lauda <tomasz.lauda@linux.intel.com>
Implements handling of new notifier aggregate flag. With this flag set the notifier will automatically use existing handle instead of creating a new one. It's helpful when we want to register only once for an event, but it's very hard to explicitly keep track of number of registrations. Signed-off-by: Tomasz Lauda <tomasz.lauda@linux.intel.com>
Adds irq_local_disable/enable during notifier_register and notifier_unregister calls to protect notify list from preemption. Signed-off-by: Tomasz Lauda <tomasz.lauda@linux.intel.com>
Saves irq_disabled config flag in private channel data. Signed-off-by: Tomasz Lauda <tomasz.lauda@linux.intel.com>
Improves host component performance by extracting DMA L1 exit to be executed commonly for all active DMA channels as registered callback in low latency scheduler for timer domain. There is no need to wait so many cycles for every channel separately. Signed-off-by: Tomasz Lauda <tomasz.lauda@linux.intel.com>
Allocates cavs_pm_runtime_data as shared to make it usable by multicore code. Signed-off-by: Tomasz Lauda <tomasz.lauda@linux.intel.com>
Implements multicore synchronization mechanism for Host DMA L1 Exit. Simple reference counter is added to avoid a situation, where shorter processing on one core forces Host DMA bus to exit L1, when there is still transfer happening on other core. PM_RUNTIME_HOST_DMA_L1 with get is called in new NOTIFIER_ID_LL_PRE_RUN notification event. Signed-off-by: Tomasz Lauda <tomasz.lauda@linux.intel.com>
tlauda
force-pushed
the
topic/host-l1-exit-move
branch
from
586b0f9 to
de843c9
Compare
|
SOFCI TEST |
|
@lgirdwood Ready to be merged. |
4 participants
Improves host component performance by extracting DMA L1 exit
to be executed commonly for all active DMA channels as registered
callback in low latency scheduler for timer domain. There is no
need to wait so many cycles for every channel separately.
Signed-off-by: Tomasz Lauda tomasz.lauda@linux.intel.com