Now, PM core supports asynchronous suspend/resume mode for devices
during system suspend/resume, and the power state transition of one
device may be completed in separate kernel thread. PM core ensures
all power state transition timing dependency between devices. This
patch enables SDIO card and function devices to suspend/resume
asynchronously. This will take advantage of multicore and improve
system suspend/resume speed. After enabling the SDIO devices and all
their child devices to suspend/resume asynchronously on ASUS T100TA,
the system suspend-to-idle time is reduced from 1645ms to 1119ms, and
the system resume time is reduced from 940ms to 918ms.
Signed-off-by: Zhonghui Fu <[email protected]>
---
Changes in v3:
- Add test result in commit message
drivers/mmc/core/sdio.c | 4 ++++
1 files changed, 4 insertions(+), 0 deletions(-)
diff --git a/drivers/mmc/core/sdio.c b/drivers/mmc/core/sdio.c
index 16d838e..530ce88 100644
--- a/drivers/mmc/core/sdio.c
+++ b/drivers/mmc/core/sdio.c
@@ -1113,6 +1113,8 @@ int mmc_attach_sdio(struct mmc_host *host)
pm_runtime_enable(&card->dev);
}
+ device_enable_async_suspend(&card->dev);
+
/*
* The number of functions on the card is encoded inside
* the ocr.
@@ -1133,6 +1135,8 @@ int mmc_attach_sdio(struct mmc_host *host)
*/
if (host->caps & MMC_CAP_POWER_OFF_CARD)
pm_runtime_enable(&card->sdio_func[i]->dev);
+
+ device_enable_async_suspend(&card->sdio_func[i]->dev);
}
/*
-- 1.7.1
On 15 November 2015 at 14:53, Fu, Zhonghui <[email protected]> wrote:
> Now, PM core supports asynchronous suspend/resume mode for devices
> during system suspend/resume, and the power state transition of one
> device may be completed in separate kernel thread. PM core ensures
> all power state transition timing dependency between devices. This
> patch enables SDIO card and function devices to suspend/resume
> asynchronously. This will take advantage of multicore and improve
> system suspend/resume speed. After enabling the SDIO devices and all
> their child devices to suspend/resume asynchronously on ASUS T100TA,
> the system suspend-to-idle time is reduced from 1645ms to 1119ms, and
> the system resume time is reduced from 940ms to 918ms.
>
> Signed-off-by: Zhonghui Fu <[email protected]>
I think this is an interesting change, but I wonder if you really
understand how this affects the order of how devices may be
suspended/resumed?
Also, I believe you didn't answer my question for the earlier version
of the patch, so let me try again.
There are a strict dependency chain when suspending/resuming devices
that must be maintained. Currently this is controlled via device
registration/probe order.
An SDIO func driver/device must always be suspended *before* the SDIO
card device. Additionally the corresponding MMC host, must be
suspended after the SDIO card device. Vice verse applies to the resume
sequence.
As this patch enables asynchronous suspend, I am worried that it will
break this dependency chain. What do you think?
Kind regards
Ulf Hansson
> ---
> Changes in v3:
> - Add test result in commit message
>
> drivers/mmc/core/sdio.c | 4 ++++
> 1 files changed, 4 insertions(+), 0 deletions(-)
>
> diff --git a/drivers/mmc/core/sdio.c b/drivers/mmc/core/sdio.c
> index 16d838e..530ce88 100644
> --- a/drivers/mmc/core/sdio.c
> +++ b/drivers/mmc/core/sdio.c
> @@ -1113,6 +1113,8 @@ int mmc_attach_sdio(struct mmc_host *host)
> pm_runtime_enable(&card->dev);
> }
>
> + device_enable_async_suspend(&card->dev);
> +
> /*
> * The number of functions on the card is encoded inside
> * the ocr.
> @@ -1133,6 +1135,8 @@ int mmc_attach_sdio(struct mmc_host *host)
> */
> if (host->caps & MMC_CAP_POWER_OFF_CARD)
> pm_runtime_enable(&card->sdio_func[i]->dev);
> +
> + device_enable_async_suspend(&card->sdio_func[i]->dev);
> }
>
> /*
> -- 1.7.1
>
On 11/16/2015 7:30 PM, Ulf Hansson wrote:
> On 15 November 2015 at 14:53, Fu, Zhonghui <[email protected]> wrote:
>> Now, PM core supports asynchronous suspend/resume mode for devices
>> during system suspend/resume, and the power state transition of one
>> device may be completed in separate kernel thread. PM core ensures
>> all power state transition timing dependency between devices. This
>> patch enables SDIO card and function devices to suspend/resume
>> asynchronously. This will take advantage of multicore and improve
>> system suspend/resume speed. After enabling the SDIO devices and all
>> their child devices to suspend/resume asynchronously on ASUS T100TA,
>> the system suspend-to-idle time is reduced from 1645ms to 1119ms, and
>> the system resume time is reduced from 940ms to 918ms.
>>
>> Signed-off-by: Zhonghui Fu <[email protected]>
> I think this is an interesting change, but I wonder if you really
> understand how this affects the order of how devices may be
> suspended/resumed?
>
> Also, I believe you didn't answer my question for the earlier version
> of the patch, so let me try again.
>
> There are a strict dependency chain when suspending/resuming devices
> that must be maintained. Currently this is controlled via device
> registration/probe order.
>
> An SDIO func driver/device must always be suspended *before* the SDIO
> card device. Additionally the corresponding MMC host, must be
> suspended after the SDIO card device. Vice verse applies to the resume
> sequence.
>
> As this patch enables asynchronous suspend, I am worried that it will
> break this dependency chain. What do you think?
After enabling asynchronous suspend/resume, PM core still ensures the strict suspend/resume dependency between child and parent devices - child must be suspended before its parent, and parent must be resumed before its child. SDIO function is child of SDIO card, and SDIO card is child of MMC host, and MMC host is child of MMC controller. So the dependency chain is not broken. Actually, many devices have been using asynchronous suspend/resume mode now.
Thanks,
Zhonghui
>
> Kind regards
> Ulf Hansson
>
>> ---
>> Changes in v3:
>> - Add test result in commit message
>>
>> drivers/mmc/core/sdio.c | 4 ++++
>> 1 files changed, 4 insertions(+), 0 deletions(-)
>>
>> diff --git a/drivers/mmc/core/sdio.c b/drivers/mmc/core/sdio.c
>> index 16d838e..530ce88 100644
>> --- a/drivers/mmc/core/sdio.c
>> +++ b/drivers/mmc/core/sdio.c
>> @@ -1113,6 +1113,8 @@ int mmc_attach_sdio(struct mmc_host *host)
>> pm_runtime_enable(&card->dev);
>> }
>>
>> + device_enable_async_suspend(&card->dev);
>> +
>> /*
>> * The number of functions on the card is encoded inside
>> * the ocr.
>> @@ -1133,6 +1135,8 @@ int mmc_attach_sdio(struct mmc_host *host)
>> */
>> if (host->caps & MMC_CAP_POWER_OFF_CARD)
>> pm_runtime_enable(&card->sdio_func[i]->dev);
>> +
>> + device_enable_async_suspend(&card->sdio_func[i]->dev);
>> }
>>
>> /*
>> -- 1.7.1
>>
On 11/17/2015 9:48 PM, Fu, Zhonghui wrote:
>
> On 11/16/2015 7:30 PM, Ulf Hansson wrote:
>> On 15 November 2015 at 14:53, Fu, Zhonghui <[email protected]> wrote:
>>> Now, PM core supports asynchronous suspend/resume mode for devices
>>> during system suspend/resume, and the power state transition of one
>>> device may be completed in separate kernel thread. PM core ensures
>>> all power state transition timing dependency between devices. This
>>> patch enables SDIO card and function devices to suspend/resume
>>> asynchronously. This will take advantage of multicore and improve
>>> system suspend/resume speed. After enabling the SDIO devices and all
>>> their child devices to suspend/resume asynchronously on ASUS T100TA,
>>> the system suspend-to-idle time is reduced from 1645ms to 1119ms, and
>>> the system resume time is reduced from 940ms to 918ms.
>>>
>>> Signed-off-by: Zhonghui Fu <[email protected]>
>> I think this is an interesting change, but I wonder if you really
>> understand how this affects the order of how devices may be
>> suspended/resumed?
>>
>> Also, I believe you didn't answer my question for the earlier version
>> of the patch, so let me try again.
>>
>> There are a strict dependency chain when suspending/resuming devices
>> that must be maintained. Currently this is controlled via device
>> registration/probe order.
>>
>> An SDIO func driver/device must always be suspended *before* the SDIO
>> card device. Additionally the corresponding MMC host, must be
>> suspended after the SDIO card device. Vice verse applies to the resume
>> sequence.
>>
>> As this patch enables asynchronous suspend, I am worried that it will
>> break this dependency chain. What do you think?
> After enabling asynchronous suspend/resume, PM core still ensures the strict suspend/resume dependency between child and parent devices - child must be suspended before its parent, and parent must be resumed before its child. SDIO function is child of SDIO card, and SDIO card is child of MMC host, and MMC host is child of MMC controller. So the dependency chain is not broken. Actually, many devices have been using asynchronous suspend/resume mode now.
Any comments are welcome.
Thanks,
Zhonghui
>
> Thanks,
> Zhonghui
>> Kind regards
>> Ulf Hansson
>>
>>> ---
>>> Changes in v3:
>>> - Add test result in commit message
>>>
>>> drivers/mmc/core/sdio.c | 4 ++++
>>> 1 files changed, 4 insertions(+), 0 deletions(-)
>>>
>>> diff --git a/drivers/mmc/core/sdio.c b/drivers/mmc/core/sdio.c
>>> index 16d838e..530ce88 100644
>>> --- a/drivers/mmc/core/sdio.c
>>> +++ b/drivers/mmc/core/sdio.c
>>> @@ -1113,6 +1113,8 @@ int mmc_attach_sdio(struct mmc_host *host)
>>> pm_runtime_enable(&card->dev);
>>> }
>>>
>>> + device_enable_async_suspend(&card->dev);
>>> +
>>> /*
>>> * The number of functions on the card is encoded inside
>>> * the ocr.
>>> @@ -1133,6 +1135,8 @@ int mmc_attach_sdio(struct mmc_host *host)
>>> */
>>> if (host->caps & MMC_CAP_POWER_OFF_CARD)
>>> pm_runtime_enable(&card->sdio_func[i]->dev);
>>> +
>>> + device_enable_async_suspend(&card->sdio_func[i]->dev);
>>> }
>>>
>>> /*
>>> -- 1.7.1
>>>
On 17 November 2015 at 14:48, Fu, Zhonghui <[email protected]> wrote:
>
>
> On 11/16/2015 7:30 PM, Ulf Hansson wrote:
>> On 15 November 2015 at 14:53, Fu, Zhonghui <[email protected]> wrote:
>>> Now, PM core supports asynchronous suspend/resume mode for devices
>>> during system suspend/resume, and the power state transition of one
>>> device may be completed in separate kernel thread. PM core ensures
>>> all power state transition timing dependency between devices. This
What "timing dependency"?
>>> patch enables SDIO card and function devices to suspend/resume
>>> asynchronously. This will take advantage of multicore and improve
>>> system suspend/resume speed. After enabling the SDIO devices and all
>>> their child devices to suspend/resume asynchronously on ASUS T100TA,
>>> the system suspend-to-idle time is reduced from 1645ms to 1119ms, and
>>> the system resume time is reduced from 940ms to 918ms.
Are these improvements achieved by $subject patch on its own or you
need below patches:
[PATCH v3] mmc: enable mmc host device to suspend/resume asynchronously [1]
[PATCH v3] mmc/sdhci-acpi: enable sdhci-acpi device to suspend/resume
asynchronously [2]
Depending if you have SD/(e)MMC card slot(s), the below patch might
affect your results. So it might be a good idea to re-run your test to
get some fresh data.
[PATCH 1/2] mmc: core: Make runtime resume default behavior for MMC/SD [3]
>>>
>>> Signed-off-by: Zhonghui Fu <[email protected]>
>> I think this is an interesting change, but I wonder if you really
>> understand how this affects the order of how devices may be
>> suspended/resumed?
>>
>> Also, I believe you didn't answer my question for the earlier version
>> of the patch, so let me try again.
>>
>> There are a strict dependency chain when suspending/resuming devices
>> that must be maintained. Currently this is controlled via device
>> registration/probe order.
>>
>> An SDIO func driver/device must always be suspended *before* the SDIO
>> card device. Additionally the corresponding MMC host, must be
>> suspended after the SDIO card device. Vice verse applies to the resume
>> sequence.
>>
>> As this patch enables asynchronous suspend, I am worried that it will
>> break this dependency chain. What do you think?
> After enabling asynchronous suspend/resume, PM core still ensures the strict suspend/resume dependency between child and parent devices - child must be suspended before its parent, and parent must be resumed before its child. SDIO function is child of SDIO card, and SDIO card is child of MMC host, and MMC host is child of MMC controller. So the dependency chain is not broken. Actually, many devices have been using asynchronous suspend/resume mode now.
I believe your view of how the PM core works for devices that *don't*
use async suspend is wrong! The PM core doesn't respect parent/child
relations during the device system PM phase for these devices.
Instead it relies on that devices in the "dpm list" are ordered
correctly. As I tried to describe earlier, that list is being updated
during device registration and probing (there are some other special
cases as well).
So, by enabling async suspend for a device it will trigger the PM core
durng device system PM, to start caring about device's parent/child
relations. I would appreciate if you could add some of this
information to the change log, as that's *why* it should work nicely
for mmc/sd/sdio to use async suspend.
>
> Thanks,
> Zhonghui
>>
>> Kind regards
>> Ulf Hansson
>>
>>> ---
>>> Changes in v3:
>>> - Add test result in commit message
>>>
>>> drivers/mmc/core/sdio.c | 4 ++++
>>> 1 files changed, 4 insertions(+), 0 deletions(-)
>>>
>>> diff --git a/drivers/mmc/core/sdio.c b/drivers/mmc/core/sdio.c
>>> index 16d838e..530ce88 100644
>>> --- a/drivers/mmc/core/sdio.c
>>> +++ b/drivers/mmc/core/sdio.c
>>> @@ -1113,6 +1113,8 @@ int mmc_attach_sdio(struct mmc_host *host)
>>> pm_runtime_enable(&card->dev);
>>> }
>>>
>>> + device_enable_async_suspend(&card->dev);
>>> +
This change will also affect SDIO combo cards. That means the when
there is a mmc blk device driver bound to the mmc card, its
->suspend() methods will be called asynchronously.
Have you considered that? Especially since there are nothing being
mentioned about it in the change-log?
Also, within this context I am wondering why you *only* enable async
suspend for SDIO cards and not all cards (SD/MMC)? Is there a problem
with doing that?
>>> /*
>>> * The number of functions on the card is encoded inside
>>> * the ocr.
>>> @@ -1133,6 +1135,8 @@ int mmc_attach_sdio(struct mmc_host *host)
>>> */
>>> if (host->caps & MMC_CAP_POWER_OFF_CARD)
>>> pm_runtime_enable(&card->sdio_func[i]->dev);
>>> +
>>> + device_enable_async_suspend(&card->sdio_func[i]->dev);
>>> }
>>>
>>> /*
>>> -- 1.7.1
>>>
>
[1]
https://lkml.org/lkml/2015/11/15/83
[2]
https://lkml.org/lkml/2015/11/16/6
[3]
http://www.spinics.net/lists/linux-mmc/msg34004.html
Kind regards
Ulf Hansson
On 11/23/2015 11:26 PM, Ulf Hansson wrote:
> On 17 November 2015 at 14:48, Fu, Zhonghui <[email protected]> wrote:
>>
>> On 11/16/2015 7:30 PM, Ulf Hansson wrote:
>>> On 15 November 2015 at 14:53, Fu, Zhonghui <[email protected]> wrote:
>>>> Now, PM core supports asynchronous suspend/resume mode for devices
>>>> during system suspend/resume, and the power state transition of one
>>>> device may be completed in separate kernel thread. PM core ensures
>>>> all power state transition timing dependency between devices. This
> What "timing dependency"?
Sorry, "timing" should be needless word.
>
>>>> patch enables SDIO card and function devices to suspend/resume
>>>> asynchronously. This will take advantage of multicore and improve
>>>> system suspend/resume speed. After enabling the SDIO devices and all
>>>> their child devices to suspend/resume asynchronously on ASUS T100TA,
>>>> the system suspend-to-idle time is reduced from 1645ms to 1119ms, and
>>>> the system resume time is reduced from 940ms to 918ms.
> Are these improvements achieved by $subject patch on its own or you
> need below patches:
>
> [PATCH v3] mmc: enable mmc host device to suspend/resume asynchronously [1]
>
> [PATCH v3] mmc/sdhci-acpi: enable sdhci-acpi device to suspend/resume
> asynchronously [2]
This patch achieves these improvements by its own.
>
> Depending if you have SD/(e)MMC card slot(s), the below patch might
> affect your results. So it might be a good idea to re-run your test to
> get some fresh data.
>
> [PATCH 1/2] mmc: core: Make runtime resume default behavior for MMC/SD [3]
I didn't find this patch in mainline kernel, where is this patch?
>
>>>> Signed-off-by: Zhonghui Fu <[email protected]>
>>> I think this is an interesting change, but I wonder if you really
>>> understand how this affects the order of how devices may be
>>> suspended/resumed?
>>>
>>> Also, I believe you didn't answer my question for the earlier version
>>> of the patch, so let me try again.
>>>
>>> There are a strict dependency chain when suspending/resuming devices
>>> that must be maintained. Currently this is controlled via device
>>> registration/probe order.
>>>
>>> An SDIO func driver/device must always be suspended *before* the SDIO
>>> card device. Additionally the corresponding MMC host, must be
>>> suspended after the SDIO card device. Vice verse applies to the resume
>>> sequence.
>>>
>>> As this patch enables asynchronous suspend, I am worried that it will
>>> break this dependency chain. What do you think?
>> After enabling asynchronous suspend/resume, PM core still ensures the strict suspend/resume dependency between child and parent devices - child must be suspended before its parent, and parent must be resumed before its child. SDIO function is child of SDIO card, and SDIO card is child of MMC host, and MMC host is child of MMC controller. So the dependency chain is not broken. Actually, many devices have been using asynchronous suspend/resume mode now.
>
> I believe your view of how the PM core works for devices that *don't*
> use async suspend is wrong! The PM core doesn't respect parent/child
> relations during the device system PM phase for these devices.
I agree with you for the following description. But, I never described how PM core works for devices that don't use async mode. Where did you get my view about this? I just said that PM core still ensure the dependency between child and parent devices after using async mode, I never said that the method ensuring dependency is the same for sync devices and async devices.
>
> Instead it relies on that devices in the "dpm list" are ordered
> correctly. As I tried to describe earlier, that list is being updated
> during device registration and probing (there are some other special
> cases as well).
>
> So, by enabling async suspend for a device it will trigger the PM core
> durng device system PM, to start caring about device's parent/child
> relations. I would appreciate if you could add some of this
> information to the change log, as that's *why* it should work nicely
> for mmc/sd/sdio to use async suspend.
>
>> Thanks,
>> Zhonghui
>>> Kind regards
>>> Ulf Hansson
>>>
>>>> ---
>>>> Changes in v3:
>>>> - Add test result in commit message
>>>>
>>>> drivers/mmc/core/sdio.c | 4 ++++
>>>> 1 files changed, 4 insertions(+), 0 deletions(-)
>>>>
>>>> diff --git a/drivers/mmc/core/sdio.c b/drivers/mmc/core/sdio.c
>>>> index 16d838e..530ce88 100644
>>>> --- a/drivers/mmc/core/sdio.c
>>>> +++ b/drivers/mmc/core/sdio.c
>>>> @@ -1113,6 +1113,8 @@ int mmc_attach_sdio(struct mmc_host *host)
>>>> pm_runtime_enable(&card->dev);
>>>> }
>>>>
>>>> + device_enable_async_suspend(&card->dev);
>>>> +
> This change will also affect SDIO combo cards. That means the when
> there is a mmc blk device driver bound to the mmc card, its
> ->suspend() methods will be called asynchronously.
>
> Have you considered that? Especially since there are nothing being
> mentioned about it in the change-log?
I have considered this, this patch still work for SDIO combo cards.
>
> Also, within this context I am wondering why you *only* enable async
> suspend for SDIO cards and not all cards (SD/MMC)? Is there a problem
> with doing that?
I am optimizing suspend/resume speed for some Intel's tablet platforms and focusing only on the most time-consuming device path now. I will deliver similar patch for SD/MMC card.
>
>>>> /*
>>>> * The number of functions on the card is encoded inside
>>>> * the ocr.
>>>> @@ -1133,6 +1135,8 @@ int mmc_attach_sdio(struct mmc_host *host)
>>>> */
>>>> if (host->caps & MMC_CAP_POWER_OFF_CARD)
>>>> pm_runtime_enable(&card->sdio_func[i]->dev);
>>>> +
>>>> + device_enable_async_suspend(&card->sdio_func[i]->dev);
>>>> }
>>>>
>>>> /*
>>>> -- 1.7.1
>>>>
> [1]
> https://lkml.org/lkml/2015/11/15/83
> [2]
> https://lkml.org/lkml/2015/11/16/6
> [3]
> http://www.spinics.net/lists/linux-mmc/msg34004.html
>
> Kind regards
> Ulf Hansson
> --
> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> the body of a message to [email protected]
> More majordomo info at http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at http://www.tux.org/lkml/
[...]
>>
>> Depending if you have SD/(e)MMC card slot(s), the below patch might
>> affect your results. So it might be a good idea to re-run your test to
>> get some fresh data.
>>
>> [PATCH 1/2] mmc: core: Make runtime resume default behavior for MMC/SD [3]
> I didn't find this patch in mainline kernel, where is this patch?
It's queued for 4.5 on my next branch via my mmc git. Thus also
available in linux-next.
One should know that it's affecting MMC/SD cards and not SDIO.
Future wise I was hoping we could do something similar for SDIO, but I
need to think that through a bit more.
>>
>>>>> Signed-off-by: Zhonghui Fu <[email protected]>
>>>> I think this is an interesting change, but I wonder if you really
>>>> understand how this affects the order of how devices may be
>>>> suspended/resumed?
>>>>
>>>> Also, I believe you didn't answer my question for the earlier version
>>>> of the patch, so let me try again.
>>>>
>>>> There are a strict dependency chain when suspending/resuming devices
>>>> that must be maintained. Currently this is controlled via device
>>>> registration/probe order.
>>>>
>>>> An SDIO func driver/device must always be suspended *before* the SDIO
>>>> card device. Additionally the corresponding MMC host, must be
>>>> suspended after the SDIO card device. Vice verse applies to the resume
>>>> sequence.
>>>>
>>>> As this patch enables asynchronous suspend, I am worried that it will
>>>> break this dependency chain. What do you think?
>>> After enabling asynchronous suspend/resume, PM core still ensures the strict suspend/resume dependency between child and parent devices - child must be suspended before its parent, and parent must be resumed before its child. SDIO function is child of SDIO card, and SDIO card is child of MMC host, and MMC host is child of MMC controller. So the dependency chain is not broken. Actually, many devices have been using asynchronous suspend/resume mode now.
>>
>> I believe your view of how the PM core works for devices that *don't*
>> use async suspend is wrong! The PM core doesn't respect parent/child
>> relations during the device system PM phase for these devices.
> I agree with you for the following description. But, I never described how PM core works for devices that don't use async mode. Where did you get my view about this? I just said that PM core still ensure the dependency between child and parent devices after using async mode, I never said that the method ensuring dependency is the same for sync devices and async devices.
You said it again. "PM core *still* ensures the dependency... after
using async mode". I guess it depends what you mean by "still". :-)
Anyway, it doesn't matter as I think we have sorted it out now.
[...]
>> This change will also affect SDIO combo cards. That means the when
>> there is a mmc blk device driver bound to the mmc card, its
>> ->suspend() methods will be called asynchronously.
>>
>> Have you considered that? Especially since there are nothing being
>> mentioned about it in the change-log?
> I have considered this, this patch still work for SDIO combo cards.
>>
>> Also, within this context I am wondering why you *only* enable async
>> suspend for SDIO cards and not all cards (SD/MMC)? Is there a problem
>> with doing that?
> I am optimizing suspend/resume speed for some Intel's tablet platforms and focusing only on the most time-consuming device path now. I will deliver similar patch for SD/MMC card.
Okay, thanks for sharing you plans!
I would prefer if we instead enables async suspend for all types of
card devices and the sdio func devices within one patch. I suggest we
use mmc_add_card() and sdio_add_func(), to where we enable async
suspend for the devices.
[...]
Kind regards
Uffe