Re: [PATCH v4 01/11] ASoC: SOF: core: add 'no_wq' probe and remove callbacks
Hi,
On Wed, 30 Aug 2023, Maarten Lankhorst wrote:
With the upcoming changes for i915/Xe driver relying on the -EPROBE_DEFER mechanism, we need to have a first pass of the probe which cannot be pushed to a workqueue. Introduce 2 new optional callbacks.
[...]
diff --git a/sound/soc/sof/core.c b/sound/soc/sof/core.c index 30db685cc5f4b..54c384a5d6140 100644 --- a/sound/soc/sof/core.c +++ b/sound/soc/sof/core.c @@ -327,8 +327,6 @@ static int sof_probe_continue(struct snd_sof_dev *sdev) dsp_err: snd_sof_remove(sdev); probe_err:
- sof_ops_free(sdev);
this seems a bit out-of-place in this patch. It seems a valid change, but not really related to this patch, right?
We seem to have a related fix waiting to be sent to alsa-devel, by Peter: "ASoC: SOF: core: Only call sof_ops_free() on remove if the probe wa" https://github.com/thesofproject/linux/pull/4515
... not yet in Mark's tree.
Otherwise patch looks good to me.
Br, Kai
On 01/09/2023 15:15, Kai Vehmanen wrote:
Hi,
On Wed, 30 Aug 2023, Maarten Lankhorst wrote:
With the upcoming changes for i915/Xe driver relying on the -EPROBE_DEFER mechanism, we need to have a first pass of the probe which cannot be pushed to a workqueue. Introduce 2 new optional callbacks.
[...]
diff --git a/sound/soc/sof/core.c b/sound/soc/sof/core.c index 30db685cc5f4b..54c384a5d6140 100644 --- a/sound/soc/sof/core.c +++ b/sound/soc/sof/core.c @@ -327,8 +327,6 @@ static int sof_probe_continue(struct snd_sof_dev *sdev) dsp_err: snd_sof_remove(sdev); probe_err:
- sof_ops_free(sdev);
this seems a bit out-of-place in this patch. It seems a valid change, but not really related to this patch, right?
The ops needs to be preserved even if the wq fails since the patch wants to call snd_sof_remove_no_wq() unconditionally on remove.
We seem to have a related fix waiting to be sent to alsa-devel, by Peter: "ASoC: SOF: core: Only call sof_ops_free() on remove if the probe wa" https://github.com/thesofproject/linux/pull/4515
I guess we can revert that in sof-dev, if this is the preferred way?
... not yet in Mark's tree.
Otherwise patch looks good to me.
I would have not created the snd_sof_remove_no_wq() as it makes not much functional sense. It might be even better if the remove in the wq would do the hda_codec_i915_exit() as the module will remain in there until the user removes it.
Br, Kai
On 9/1/23 08:44, Péter Ujfalusi wrote:
On 01/09/2023 15:15, Kai Vehmanen wrote:
Hi,
On Wed, 30 Aug 2023, Maarten Lankhorst wrote:
With the upcoming changes for i915/Xe driver relying on the -EPROBE_DEFER mechanism, we need to have a first pass of the probe which cannot be pushed to a workqueue. Introduce 2 new optional callbacks.
[...]
diff --git a/sound/soc/sof/core.c b/sound/soc/sof/core.c index 30db685cc5f4b..54c384a5d6140 100644 --- a/sound/soc/sof/core.c +++ b/sound/soc/sof/core.c @@ -327,8 +327,6 @@ static int sof_probe_continue(struct snd_sof_dev *sdev) dsp_err: snd_sof_remove(sdev); probe_err:
- sof_ops_free(sdev);
this seems a bit out-of-place in this patch. It seems a valid change, but not really related to this patch, right?
The ops needs to be preserved even if the wq fails since the patch wants to call snd_sof_remove_no_wq() unconditionally on remove.
We seem to have a related fix waiting to be sent to alsa-devel, by Peter: "ASoC: SOF: core: Only call sof_ops_free() on remove if the probe wa" https://github.com/thesofproject/linux/pull/4515
I guess we can revert that in sof-dev, if this is the preferred way?
... not yet in Mark's tree.
Otherwise patch looks good to me.
I would have not created the snd_sof_remove_no_wq() as it makes not much functional sense. It might be even better if the remove in the wq would do the hda_codec_i915_exit() as the module will remain in there until the user removes it.
I think find all this very confusing, because there is no workqueue used in the remove steps. The workqueue is only used ONCE during the probe.
On 9/5/23 08:37, Pierre-Louis Bossart wrote:
On 9/1/23 08:44, Péter Ujfalusi wrote:
On 01/09/2023 15:15, Kai Vehmanen wrote:
Hi,
On Wed, 30 Aug 2023, Maarten Lankhorst wrote:
With the upcoming changes for i915/Xe driver relying on the -EPROBE_DEFER mechanism, we need to have a first pass of the probe which cannot be pushed to a workqueue. Introduce 2 new optional callbacks.
[...]
diff --git a/sound/soc/sof/core.c b/sound/soc/sof/core.c index 30db685cc5f4b..54c384a5d6140 100644 --- a/sound/soc/sof/core.c +++ b/sound/soc/sof/core.c @@ -327,8 +327,6 @@ static int sof_probe_continue(struct snd_sof_dev *sdev) dsp_err: snd_sof_remove(sdev); probe_err:
- sof_ops_free(sdev);
this seems a bit out-of-place in this patch. It seems a valid change, but not really related to this patch, right?
The ops needs to be preserved even if the wq fails since the patch wants to call snd_sof_remove_no_wq() unconditionally on remove.
We seem to have a related fix waiting to be sent to alsa-devel, by Peter: "ASoC: SOF: core: Only call sof_ops_free() on remove if the probe wa" https://github.com/thesofproject/linux/pull/4515
I guess we can revert that in sof-dev, if this is the preferred way?
... not yet in Mark's tree.
Otherwise patch looks good to me.
I would have not created the snd_sof_remove_no_wq() as it makes not much functional sense. It might be even better if the remove in the wq would do the hda_codec_i915_exit() as the module will remain in there until the user removes it.
I think find all this very confusing, because there is no workqueue used in the remove steps. The workqueue is only used ONCE during the probe.
Maybe we should just remove any references to workqueues, and have
probe_start (cannot run in a wq) probe (may run in a wq) remove (cannot run in a wq, needs to call cancel_work_sync() if the probe runs in a wq) remove_last (cannot run in a wq, releases all resources acquired in probe_start)
Or something similar that shows the symmetry between steps and when the wq is allowed.
On 07/09/2023 20:29, Pierre-Louis Bossart wrote:
I think find all this very confusing, because there is no workqueue used in the remove steps. The workqueue is only used ONCE during the probe.
Maybe we should just remove any references to workqueues, and have
probe_start (cannot run in a wq) probe (may run in a wq) remove (cannot run in a wq, needs to call cancel_work_sync() if the probe runs in a wq) remove_last (cannot run in a wq, releases all resources acquired in probe_start)
Or something similar that shows the symmetry between steps and when the wq is allowed.
What we have atm: snd_sof_probe - might be called from wq snd_sof_remove - might be called from wq (cleans up the snd_sof_probe step)
We want a callbacks for hardware/device probing, right, split the snd_sof_probe (and remove) to be able to support a sane level of deferred probing support.
With that in mind: snd_sof_device_probe - Not called from wq (to handle deferred probing) snd_sof_probe - might be called from wq
snd_sof_remove - might be called from wq (cleans up the snd_sof_probe step) snd_sof_device_remove - Not called from wq (to up the snd_sof_device_probe step)
Naming option: s/device/hardware
However, I think the snd_sof_device_remove itself is redundant and we might not need it at all as in case we have wq and there is a failure in there we do want to release resources as much as possible. The module will be kept loaded (no deferred handling in wq) and that might block PM, other devices to behave correctly. Iow, if the wq has failure we should do a cleanup to the best effort to reach a level like the driver is not even loaded.
Doable? I thin it is.
What we have atm: snd_sof_probe - might be called from wq snd_sof_remove - might be called from wq (cleans up the snd_sof_probe step)
I don't think it's correct, snd_sof_remove cannot be called from a wq. The device core knows nothing about workqueues.
We want a callbacks for hardware/device probing, right, split the snd_sof_probe (and remove) to be able to support a sane level of deferred probing support.
With that in mind: snd_sof_device_probe - Not called from wq (to handle deferred probing) snd_sof_probe - might be called from wq
snd_sof_remove - might be called from wq (cleans up the snd_sof_probe step) snd_sof_device_remove - Not called from wq (to up the snd_sof_device_probe step)
Naming option: s/device/hardware
I like the 'device' hint since it's directly related to the device (or subsystem) callbacks.
However, I think the snd_sof_device_remove itself is redundant and we might not need it at all as in case we have wq and there is a failure in there we do want to release resources as much as possible. The module will be kept loaded (no deferred handling in wq) and that might block PM, other devices to behave correctly. Iow, if the wq has failure we should do a cleanup to the best effort to reach a level like the driver is not even loaded.
If we have a failure in a workqueue used for probe, then we have to clean-up everything since nothing in the device core will do so for us.
On 12/09/2023 03:25, Pierre-Louis Bossart wrote:
What we have atm: snd_sof_probe - might be called from wq snd_sof_remove - might be called from wq (cleans up the snd_sof_probe step)
I don't think it's correct, snd_sof_remove cannot be called from a wq. The device core knows nothing about workqueues.
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/soun...
it is called on the error path of sof_probe_continue(), which can be run in a workque.
We want a callbacks for hardware/device probing, right, split the snd_sof_probe (and remove) to be able to support a sane level of deferred probing support.
With that in mind: snd_sof_device_probe - Not called from wq (to handle deferred probing) snd_sof_probe - might be called from wq
snd_sof_remove - might be called from wq (cleans up the snd_sof_probe step) snd_sof_device_remove - Not called from wq (to up the snd_sof_device_probe step)
Naming option: s/device/hardware
I like the 'device' hint since it's directly related to the device (or subsystem) callbacks.
However, I think the snd_sof_device_remove itself is redundant and we might not need it at all as in case we have wq and there is a failure in there we do want to release resources as much as possible. The module will be kept loaded (no deferred handling in wq) and that might block PM, other devices to behave correctly. Iow, if the wq has failure we should do a cleanup to the best effort to reach a level like the driver is not even loaded.
If we have a failure in a workqueue used for probe, then we have to clean-up everything since nothing in the device core will do so for us.
Yes, this makes the snd_sof_device_remove() redundant or at least the definition of it is no longer a mirror of snd_sof_device_probe():
snd_sof_device_remove - might be called from wq (cleans up the snd_sof_device_probe step)
Any failure in sof_probe_continue() should execute the snd_sof_device_remove(), snd_sof_remove() is only involved after the snd_sof_probe() have returned with success.
I think this makes actually makes sense and it is well defined. On module remove we need to take into account the case when we have failed in wq similarly as we do currently (the resources have been freed up already).
On 9/12/23 02:10, Péter Ujfalusi wrote:
On 12/09/2023 03:25, Pierre-Louis Bossart wrote:
What we have atm: snd_sof_probe - might be called from wq snd_sof_remove - might be called from wq (cleans up the snd_sof_probe step)
I don't think it's correct, snd_sof_remove cannot be called from a wq. The device core knows nothing about workqueues.
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/soun...
it is called on the error path of sof_probe_continue(), which can be run in a workque.
We want a callbacks for hardware/device probing, right, split the snd_sof_probe (and remove) to be able to support a sane level of deferred probing support.
With that in mind: snd_sof_device_probe - Not called from wq (to handle deferred probing) snd_sof_probe - might be called from wq
snd_sof_remove - might be called from wq (cleans up the snd_sof_probe step) snd_sof_device_remove - Not called from wq (to up the snd_sof_device_probe step)
Naming option: s/device/hardware
I like the 'device' hint since it's directly related to the device (or subsystem) callbacks.
However, I think the snd_sof_device_remove itself is redundant and we might not need it at all as in case we have wq and there is a failure in there we do want to release resources as much as possible. The module will be kept loaded (no deferred handling in wq) and that might block PM, other devices to behave correctly. Iow, if the wq has failure we should do a cleanup to the best effort to reach a level like the driver is not even loaded.
If we have a failure in a workqueue used for probe, then we have to clean-up everything since nothing in the device core will do so for us.
Yes, this makes the snd_sof_device_remove() redundant or at least the definition of it is no longer a mirror of snd_sof_device_probe():
snd_sof_device_remove - might be called from wq (cleans up the snd_sof_device_probe step)
Any failure in sof_probe_continue() should execute the snd_sof_device_remove(), snd_sof_remove() is only involved after the snd_sof_probe() have returned with success.
I think this makes actually makes sense and it is well defined. On module remove we need to take into account the case when we have failed in wq similarly as we do currently (the resources have been freed up already).
Agree, I stand corrected, thanks Peter.
participants (3)
-
Kai Vehmanen -
Pierre-Louis Bossart -
Péter Ujfalusi