[lkml]   [2021]   [Jun]   [3]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: [PATCH v1 4/4] PCI: brcmstb: add shutdown call to driver
On 6/3/21 1:58 PM, Bjorn Helgaas wrote:
> On Thu, Jun 03, 2021 at 10:30:37AM -0700, Florian Fainelli wrote:
>> On 6/3/21 10:23 AM, Bjorn Helgaas wrote:
>>> On Wed, May 26, 2021 at 10:03:47AM -0700, Florian Fainelli wrote:
>>>> On 5/25/21 2:18 PM, Bjorn Helgaas wrote:
>>>>> On Tue, Apr 27, 2021 at 01:51:39PM -0400, Jim Quinlan wrote:
>>>>>> The shutdown() call is similar to the remove() call except the former does
>>>>>> not need to invoke pci_{stop,remove}_root_bus(), and besides, errors occur
>>>>>> if it does.
>>>>> This doesn't explain why shutdown() is necessary. "errors occur"
>>>>> might be a hint, except that AFAICT, many similar drivers do invoke
>>>>> pci_stop_root_bus() and pci_remove_root_bus() (several of them while
>>>>> holding pci_lock_rescan_remove()), without implementing .shutdown().
>>>> We have to implement .shutdown() in order to meet a certain power budget
>>>> while the chip is being put into S5 (soft off) state and still support
>>>> Wake-on-WLAN, for our latest chips this translates into roughly 200mW of
>>>> power savings at the wall. We could probably add a word or two in a v2
>>>> that indicates this is done for power savings.
>>> "Saving power" is a great reason to do this. But we still need to
>>> connect this to the driver model and the system-level behavior
>>> somehow.
>>> The pci_driver comment says @shutdown is to "stop idling DMA
>>> operations" and it hooks into reboot_notifier_list in kernel/sys.c.
>>> That's incorrect or at least incomplete because reboot_notifier_list
>>> isn't mentioned at all in kernel/sys.c, and I don't see the connection
>>> between @shutdown and reboot_notifier_list.
>>> AFAICT, @shutdown is currently used in this path:
>>> kernel_restart_prepare or kernel_shutdown_prepare
>>> device_shutdown
>>> dev->bus->shutdown
>>> pci_device_shutdown # pci_bus_type.shutdown
>>> drv->shutdown
>>> so we're going to either reboot or halt/power-off the entire system,
>>> and we're not going to use this device again until we're in a
>>> brand-new kernel and we re-enumerate the device and re-register the
>>> driver.
>>> I'm not quite sure how either of those fits into the power-saving
>>> reason. I guess going to S5 is probably via the kernel_power_off()
>>> path and that by itself doesn't turn off as much power to the PCIe
>>> controller as it could? And this new .shutdown() method will get
>>> called in that path and will turn off more power, but will still leave
>>> enough for wake-on-LAN to work? And when we *do* wake from S5,
>>> obviously that means a complete boot with a new kernel.
>> Correct, the S5 shutdown is via kernel_power_off() and will turn off all
>> that we can in the PCIe root complex and its PHY, drop the PCIe link to
>> the end-point which signals that the end-point can enter its own suspend
>> logic, too. And yes, when we do wake-up from S5 it means booting a
>> completely new kernel. S5 is typically implemented in our chips by
>> keeping just a little bit of logic active to service wake-up events
>> (infrared remotes, GPIOs, RTC, etc.).
> Which part of that does this patch change? Is it that the new
> .shutdown() turns off more power than machine_power_off() does by
> itself?

Yes, with pcie-brcmstb.c providing a .shutdow() callback we have a
chance to turn off our PCIe PHY and the RC's digital clock which would
not be able to do otherwise.
[unhandled content-type:application/pkcs7-signature]
 \ /
  Last update: 2021-06-03 23:02    [W:0.069 / U:0.152 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site