[lkml]   [2019]   [Feb]   [15]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: [PATCH 2/2] driver core: Fix possible supplier PM-usage counter imbalance

On 15/02/2019 13:21, Jon Hunter wrote:
> On 15/02/2019 12:06, Rafael J. Wysocki wrote:
>> On Friday, February 15, 2019 12:00:27 PM CET Jon Hunter wrote:
>>> Hi Rafael,
>>> On 12/02/2019 12:08, Rafael J. Wysocki wrote:
>>>> From: Rafael J. Wysocki <>
>>>> If a stateless device link to a certain supplier with
>>>> DL_FLAG_PM_RUNTIME set in the flags is added and then removed by the
>>>> consumer driver's probe callback, the supplier's PM-runtime usage
>>>> counter will be nonzero after that which effectively causes the
>>>> supplier to remain "always on" going forward.
>>>> Namely, device_link_add() called to add the link invokes
>>>> device_link_rpm_prepare() which notices that the consumer driver is
>>>> probing, so it increments the supplier's PM-runtime usage counter
>>>> with the assumption that the link will stay around until
>>>> pm_runtime_put_suppliers() is called by driver_probe_device(),
>>>> but if the link goes away before that point, the supplier's
>>>> PM-runtime usage counter will remain nonzero.
>>>> To prevent that from happening, first rework pm_runtime_get_suppliers()
>>>> and pm_runtime_put_suppliers() to use the rpm_active refounts of device
>>>> links and make the latter only drop rpm_active and the supplier's
>>>> PM-runtime usage counter for each link by one, unless rpm_active is
>>>> one already for it. Next, modify device_link_add() to bump up the
>>>> new link's rpm_active refcount and the suppliers PM-runtime usage
>>>> counter by two, to prevent pm_runtime_put_suppliers(), if it is
>>>> called subsequently, from suspending the supplier prematurely (in
>>>> case its PM-runtime usage counter goes down to 0 in there).
>>>> Due to the way rpm_put_suppliers() works, this change does not
>>>> affect runtime suspend of the consumer ends of new device links (or,
>>>> generally, device links for which DL_FLAG_PM_RUNTIME has just been
>>>> set).
>>>> Fixes: e2f3cd831a28 ("driver core: Fix handling of runtime PM flags in device_link_add()")
>>>> Reported-by: Ulf Hansson <>
>>>> Signed-off-by: Rafael J. Wysocki <>
>>>> ---
>>>> Note that the issue had been there before commit e2f3cd831a28, but it was
>>>> overlooked by that commit and this change is a fix on top of it, so make
>>>> the Fixes: tag point to commit e2f3cd831a28 (instead of an earlier one
>>>> that the patch will not be applicable to).
>>> I noticed that yesterday's and today's -next were no longer booting on
>>> one of our Tegra boards (Tegra210 Jetson TX2) because networking is
>>> failing. The ethernet chip is a USB device and looking at the bootlogs I
>>> can see that the Tegra XHCI driver is failing ...
>>> tegra-xusb 70090000.usb: xHCI host controller not responding, assume dead
>>> tegra-xusb 70090000.usb: HC died; cleaning up
>>> The Tegra XHCI driver uses multiple power-domains and uses
>>> device_link_add() to attach them. So now I am wondering if there is
>>> something that we have got wrong in our implementation. However, I don't
>>> see the device being probed deferred on boot or anything like that.
>>> The driver in question is drivers/usb/host/xhci-tegra.c and we add the
>>> links in the function tegra_xusb_powerdomain_init() which is before RPM
>>> is enabled. Let me know if you have any thoughts.
>> Please try the appended patch on top of the $subject one (provided that
>> reverting the $subject patch makes the problem go away).
> Thanks and yes to confirm, reverting the $subject patch on top of next
> does make the issue go away.
>> ---
>> drivers/base/power/runtime.c | 9 ++++++---
>> 1 file changed, 6 insertions(+), 3 deletions(-)
>> Index: linux-pm/drivers/base/power/runtime.c
>> ===================================================================
>> --- linux-pm.orig/drivers/base/power/runtime.c
>> +++ linux-pm/drivers/base/power/runtime.c
>> @@ -1675,9 +1675,12 @@ void pm_runtime_put_suppliers(struct dev
>> idx = device_links_read_lock();
>> list_for_each_entry_rcu(link, &dev->links.suppliers, c_node)
>> - if (link->flags & DL_FLAG_PM_RUNTIME &&
>> - refcount_dec_not_one(&link->rpm_active))
>> - pm_runtime_put(link->supplier);
>> + if (link->flags & DL_FLAG_PM_RUNTIME) {
>> + if (refcount_dec_not_one(&link->rpm_active))
>> + pm_runtime_put(link->supplier);
>> + else
>> + pm_request_idle(link->supplier);
>> + }
>> device_links_read_unlock(idx);
>> }
> I will try this now and report back in a bit.

I tried this on top of next, but unfortunately the same issue still
persists and so this did not fix it. Let me know if there is any debug I
can add/enable.



 \ /
  Last update: 2019-02-15 15:15    [W:0.053 / U:5.916 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site