| Message ID | 20210716101602.1891-1-laurentiu.tudor@nxp.com |
|---|---|
| State | Superseded |
| Headers | show |
| Series | software node: balance refcount for managed sw nodes | expand |
On 7/16/2021 8:21 PM, Jon Nettleton wrote: > On Fri, Jul 16, 2021 at 2:17 PM Andy Shevchenko > <andriy.shevchenko@linux.intel.com> wrote: >> >> On Fri, Jul 16, 2021 at 01:16:02PM +0300, laurentiu.tudor@nxp.com wrote: >>> From: Laurentiu Tudor <laurentiu.tudor@nxp.com> >>> >>> software_node_notify(), on KOBJ_REMOVE drops the refcount twice on managed >>> software nodes, thus leading to underflow errors. Balance the refcount by >>> bumping it in the device_create_managed_software_node() function. >>> >>> The error [1] was encountered after adding a .shutdown() op to our >>> fsl-mc-bus driver. >> >> Looking into the history of adding ->shutdown() to dwc3 driver (it got reverted >> later on), I can tell that probably something is wrong in the ->shutdown() >> method itself. >> >> -- >> With Best Regards, >> Andy Shevchenko >> >> > > Isn't the other alternative to just remove the second kobject_put from > KOBJ_REMOVE ? > Or maybe on top of Heikki's suggestion, replace the calls to sysfs_create_link() from KOBJ_ADD with sysfs_create_link_nowarn()? --- Best Regards, Laurentiu
On Mon, Jul 19, 2021 at 03:00:17PM +0300, Laurentiu Tudor wrote: > On 7/16/2021 8:21 PM, Jon Nettleton wrote: > > On Fri, Jul 16, 2021 at 2:17 PM Andy Shevchenko > > <andriy.shevchenko@linux.intel.com> wrote: > >> > >> On Fri, Jul 16, 2021 at 01:16:02PM +0300, laurentiu.tudor@nxp.com wrote: > >>> From: Laurentiu Tudor <laurentiu.tudor@nxp.com> > >>> > >>> software_node_notify(), on KOBJ_REMOVE drops the refcount twice on managed > >>> software nodes, thus leading to underflow errors. Balance the refcount by > >>> bumping it in the device_create_managed_software_node() function. > >>> > >>> The error [1] was encountered after adding a .shutdown() op to our > >>> fsl-mc-bus driver. > >> > >> Looking into the history of adding ->shutdown() to dwc3 driver (it got reverted > >> later on), I can tell that probably something is wrong in the ->shutdown() > >> method itself. > > > > Isn't the other alternative to just remove the second kobject_put from > > KOBJ_REMOVE ? > > > > Or maybe on top of Heikki's suggestion, replace the calls to > sysfs_create_link() from KOBJ_ADD with sysfs_create_link_nowarn()? _noearn will hide the problem. It was there, it was removed from there. Perhaps we have to understand the root cause better (some specific flow?). Any insight from you on the flow when the issue appears? I.o.w. what happened on the big picture that we got into the warning you see? -- With Best Regards, Andy Shevchenko
On 7/19/2021 3:22 PM, Andy Shevchenko wrote: > On Mon, Jul 19, 2021 at 03:00:17PM +0300, Laurentiu Tudor wrote: >> On 7/16/2021 8:21 PM, Jon Nettleton wrote: >>> On Fri, Jul 16, 2021 at 2:17 PM Andy Shevchenko >>> <andriy.shevchenko@linux.intel.com> wrote: >>>> >>>> On Fri, Jul 16, 2021 at 01:16:02PM +0300, laurentiu.tudor@nxp.com wrote: >>>>> From: Laurentiu Tudor <laurentiu.tudor@nxp.com> >>>>> >>>>> software_node_notify(), on KOBJ_REMOVE drops the refcount twice on managed >>>>> software nodes, thus leading to underflow errors. Balance the refcount by >>>>> bumping it in the device_create_managed_software_node() function. >>>>> >>>>> The error [1] was encountered after adding a .shutdown() op to our >>>>> fsl-mc-bus driver. >>>> >>>> Looking into the history of adding ->shutdown() to dwc3 driver (it got reverted >>>> later on), I can tell that probably something is wrong in the ->shutdown() >>>> method itself. >>> >>> Isn't the other alternative to just remove the second kobject_put from >>> KOBJ_REMOVE ? >>> >> >> Or maybe on top of Heikki's suggestion, replace the calls to >> sysfs_create_link() from KOBJ_ADD with sysfs_create_link_nowarn()? > > _noearn will hide the problem. It was there, it was removed from there. > Perhaps we have to understand the root cause better (some specific flow?). > > Any insight from you on the flow when the issue appears? I.o.w. what happened > on the big picture that we got into the warning you see? > I encountered the initial issue when trying to shut down a system booted with ACPI but only after adding a .shutdown() callback to our bus driver so that the devices are properly taken down. The problem was that software_node_notify(), on KOBJ_REMOVE was dropping the reference count twice leading to an underflow error. My initial proposal was to just bump the refcount in device_create_managed_software_node(). The device properties that triggered the problem are created here [1]. Heikko suggested that instead of manually incrementing the refcount to use software_node_notify(KOBJ_ADD). This triggered the second issue, a duplicated sysfs entry warning originating in the usb subsystem: device_create_managed_software_node() ends up being called twice, once here [2] and secondly, the place I previous mentioned [1]. [1] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/acpi/arm64/iort.c#n952 [2] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/usb/dwc3/host.c#n111 --- Best Regards, Laurentiu
On Tue, Jul 20, 2021 at 12:22 PM Laurentiu Tudor <laurentiu.tudor@nxp.com> wrote: > On 7/19/2021 3:22 PM, Andy Shevchenko wrote: > > On Mon, Jul 19, 2021 at 03:00:17PM +0300, Laurentiu Tudor wrote: > >> On 7/16/2021 8:21 PM, Jon Nettleton wrote: > >>> On Fri, Jul 16, 2021 at 2:17 PM Andy Shevchenko > >>> <andriy.shevchenko@linux.intel.com> wrote: > >>>> > >>>> On Fri, Jul 16, 2021 at 01:16:02PM +0300, laurentiu.tudor@nxp.com wrote: > >>>>> From: Laurentiu Tudor <laurentiu.tudor@nxp.com> > >>>>> > >>>>> software_node_notify(), on KOBJ_REMOVE drops the refcount twice on managed > >>>>> software nodes, thus leading to underflow errors. Balance the refcount by > >>>>> bumping it in the device_create_managed_software_node() function. > >>>>> > >>>>> The error [1] was encountered after adding a .shutdown() op to our > >>>>> fsl-mc-bus driver. > >>>> > >>>> Looking into the history of adding ->shutdown() to dwc3 driver (it got reverted > >>>> later on), I can tell that probably something is wrong in the ->shutdown() > >>>> method itself. > >>> > >>> Isn't the other alternative to just remove the second kobject_put from > >>> KOBJ_REMOVE ? > >> > >> Or maybe on top of Heikki's suggestion, replace the calls to > >> sysfs_create_link() from KOBJ_ADD with sysfs_create_link_nowarn()? > > > > _noearn will hide the problem. It was there, it was removed from there. > > Perhaps we have to understand the root cause better (some specific flow?). > > > > Any insight from you on the flow when the issue appears? I.o.w. what happened > > on the big picture that we got into the warning you see? > > I encountered the initial issue when trying to shut down a system booted > with ACPI but only after adding a .shutdown() callback to our bus driver > so that the devices are properly taken down. The problem was that > software_node_notify(), on KOBJ_REMOVE was dropping the reference count > twice leading to an underflow error. My initial proposal was to just > bump the refcount in device_create_managed_software_node(). The device > properties that triggered the problem are created here [1]. > > Heikko suggested that instead of manually incrementing the refcount to > use software_node_notify(KOBJ_ADD). This triggered the second issue, a > duplicated sysfs entry warning originating in the usb subsystem: > device_create_managed_software_node() ends up being called twice, once > here [2] and secondly, the place I previous mentioned [1]. This [3] is what I have reported against DWC3 when ->shutdown() has been added there. And here [4] is another thread about the issue with that callback. The ->release() callback is called at put_device() [5] and ->shutdown() is called before that [6]. That said, can you inspect your ->shutdown() implementation once more time and perhaps see if there is anything that can be amended? > [1] > https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/acpi/arm64/iort.c#n952 > [2] > https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/usb/dwc3/host.c#n111 [3]: https://lore.kernel.org/linux-usb/CAHp75Vd-5U5zgtDfM5C3Jsx51HVYB+rNcHYC2XP=G7dOd=cdTg@mail.gmail.com/ [4]: https://lore.kernel.org/linux-usb/c3c75895-313a-5be7-6421-b32bac741a88@arm.com/T/#u [5]: https://elixir.bootlin.com/linux/latest/source/drivers/base/core.c#L2216 [6]: https://elixir.bootlin.com/linux/latest/source/drivers/base/core.c#L4447 -- With Best Regards, Andy Shevchenko
On 7/20/2021 1:27 PM, Andy Shevchenko wrote: > On Tue, Jul 20, 2021 at 12:22 PM Laurentiu Tudor > <laurentiu.tudor@nxp.com> wrote: >> On 7/19/2021 3:22 PM, Andy Shevchenko wrote: >>> On Mon, Jul 19, 2021 at 03:00:17PM +0300, Laurentiu Tudor wrote: >>>> On 7/16/2021 8:21 PM, Jon Nettleton wrote: >>>>> On Fri, Jul 16, 2021 at 2:17 PM Andy Shevchenko >>>>> <andriy.shevchenko@linux.intel.com> wrote: >>>>>> >>>>>> On Fri, Jul 16, 2021 at 01:16:02PM +0300, laurentiu.tudor@nxp.com wrote: >>>>>>> From: Laurentiu Tudor <laurentiu.tudor@nxp.com> >>>>>>> >>>>>>> software_node_notify(), on KOBJ_REMOVE drops the refcount twice on managed >>>>>>> software nodes, thus leading to underflow errors. Balance the refcount by >>>>>>> bumping it in the device_create_managed_software_node() function. >>>>>>> >>>>>>> The error [1] was encountered after adding a .shutdown() op to our >>>>>>> fsl-mc-bus driver. >>>>>> >>>>>> Looking into the history of adding ->shutdown() to dwc3 driver (it got reverted >>>>>> later on), I can tell that probably something is wrong in the ->shutdown() >>>>>> method itself. >>>>> >>>>> Isn't the other alternative to just remove the second kobject_put from >>>>> KOBJ_REMOVE ? >>>> >>>> Or maybe on top of Heikki's suggestion, replace the calls to >>>> sysfs_create_link() from KOBJ_ADD with sysfs_create_link_nowarn()? >>> >>> _noearn will hide the problem. It was there, it was removed from there. >>> Perhaps we have to understand the root cause better (some specific flow?). >>> >>> Any insight from you on the flow when the issue appears? I.o.w. what happened >>> on the big picture that we got into the warning you see? >> >> I encountered the initial issue when trying to shut down a system booted >> with ACPI but only after adding a .shutdown() callback to our bus driver >> so that the devices are properly taken down. The problem was that >> software_node_notify(), on KOBJ_REMOVE was dropping the reference count >> twice leading to an underflow error. My initial proposal was to just >> bump the refcount in device_create_managed_software_node(). The device >> properties that triggered the problem are created here [1]. >> >> Heikko suggested that instead of manually incrementing the refcount to >> use software_node_notify(KOBJ_ADD). This triggered the second issue, a >> duplicated sysfs entry warning originating in the usb subsystem: >> device_create_managed_software_node() ends up being called twice, once >> here [2] and secondly, the place I previous mentioned [1]. > > This [3] is what I have reported against DWC3 when ->shutdown() has > been added there. And here [4] is another thread about the issue with > that callback. The ->release() callback is called at put_device() [5] > and ->shutdown() is called before that [6]. That said, can you inspect > your ->shutdown() implementation once more time and perhaps see if > there is anything that can be amended? > Will do, thanks for the pointers. It could be that we mess something out in how we use the driver model. --- Best Regards, Laurentiu
On 7/26/2021 10:59 AM, Laurentiu Tudor wrote: > > > On 7/20/2021 1:27 PM, Andy Shevchenko wrote: >> On Tue, Jul 20, 2021 at 12:22 PM Laurentiu Tudor >> <laurentiu.tudor@nxp.com> wrote: >>> On 7/19/2021 3:22 PM, Andy Shevchenko wrote: >>>> On Mon, Jul 19, 2021 at 03:00:17PM +0300, Laurentiu Tudor wrote: >>>>> On 7/16/2021 8:21 PM, Jon Nettleton wrote: >>>>>> On Fri, Jul 16, 2021 at 2:17 PM Andy Shevchenko >>>>>> <andriy.shevchenko@linux.intel.com> wrote: >>>>>>> >>>>>>> On Fri, Jul 16, 2021 at 01:16:02PM +0300, laurentiu.tudor@nxp.com wrote: >>>>>>>> From: Laurentiu Tudor <laurentiu.tudor@nxp.com> >>>>>>>> >>>>>>>> software_node_notify(), on KOBJ_REMOVE drops the refcount twice on managed >>>>>>>> software nodes, thus leading to underflow errors. Balance the refcount by >>>>>>>> bumping it in the device_create_managed_software_node() function. >>>>>>>> >>>>>>>> The error [1] was encountered after adding a .shutdown() op to our >>>>>>>> fsl-mc-bus driver. >>>>>>> >>>>>>> Looking into the history of adding ->shutdown() to dwc3 driver (it got reverted >>>>>>> later on), I can tell that probably something is wrong in the ->shutdown() >>>>>>> method itself. >>>>>> >>>>>> Isn't the other alternative to just remove the second kobject_put from >>>>>> KOBJ_REMOVE ? >>>>> >>>>> Or maybe on top of Heikki's suggestion, replace the calls to >>>>> sysfs_create_link() from KOBJ_ADD with sysfs_create_link_nowarn()? >>>> >>>> _noearn will hide the problem. It was there, it was removed from there. >>>> Perhaps we have to understand the root cause better (some specific flow?). >>>> >>>> Any insight from you on the flow when the issue appears? I.o.w. what happened >>>> on the big picture that we got into the warning you see? >>> >>> I encountered the initial issue when trying to shut down a system booted >>> with ACPI but only after adding a .shutdown() callback to our bus driver >>> so that the devices are properly taken down. The problem was that >>> software_node_notify(), on KOBJ_REMOVE was dropping the reference count >>> twice leading to an underflow error. My initial proposal was to just >>> bump the refcount in device_create_managed_software_node(). The device >>> properties that triggered the problem are created here [1]. >>> >>> Heikko suggested that instead of manually incrementing the refcount to >>> use software_node_notify(KOBJ_ADD). This triggered the second issue, a >>> duplicated sysfs entry warning originating in the usb subsystem: >>> device_create_managed_software_node() ends up being called twice, once >>> here [2] and secondly, the place I previous mentioned [1]. >> >> This [3] is what I have reported against DWC3 when ->shutdown() has >> been added there. And here [4] is another thread about the issue with >> that callback. The ->release() callback is called at put_device() [5] >> and ->shutdown() is called before that [6]. That said, can you inspect >> your ->shutdown() implementation once more time and perhaps see if >> there is anything that can be amended? >> > > Will do, thanks for the pointers. It could be that we mess something out > in how we use the driver model. > Quick (and late, sorry) update from my side. I've spent time on debugging our bus, did found some issues but, at least for now, none are related to sw node. In the mean time, I noticed in the swnode code that device_add_software_node() calls software_node_notify(KOBJ_ADD) while device_create_managed_software_node() doesn't. Updating [1] the later with the call to software_node_notify(KOBJ_ADD) does seem to fix the issue I'm seeing. Could this be a problem? Any comments appreciated. One more thing perhaps worth mentioning is that, at least for now, there are few uses for this device_create_managed_software_node() api, mentioning here a couple of them: - arm64 iort code - this seems to be triggering the issue i'm getting - dwc3 usb - Andy reported similar issues here, maybe the issue is common? [1] @@ -1113,6 +1125,15 @@ int device_create_managed_software_node(struct device *dev, to_swnode(fwnode)->managed = true; set_secondary_fwnode(dev, fwnode); + /* + * If the device has been fully registered by the time this function is + * called, software_node_notify() must be called separately so that the + * symlinks get created and the reference count of the node is kept in + * balance. + */ + if (device_is_registered(dev)) + software_node_notify(dev, KOBJ_ADD); + return 0; } --- Thanks & Best Regards, Laurentiu
On Tue, Sep 07, 2021 at 06:59:18PM +0300, Laurentiu Tudor wrote: > > > On 7/26/2021 10:59 AM, Laurentiu Tudor wrote: > > > > > > On 7/20/2021 1:27 PM, Andy Shevchenko wrote: > >> On Tue, Jul 20, 2021 at 12:22 PM Laurentiu Tudor > >> <laurentiu.tudor@nxp.com> wrote: > >>> On 7/19/2021 3:22 PM, Andy Shevchenko wrote: > >>>> On Mon, Jul 19, 2021 at 03:00:17PM +0300, Laurentiu Tudor wrote: > >>>>> On 7/16/2021 8:21 PM, Jon Nettleton wrote: > >>>>>> On Fri, Jul 16, 2021 at 2:17 PM Andy Shevchenko > >>>>>> <andriy.shevchenko@linux.intel.com> wrote: > >>>>>>> > >>>>>>> On Fri, Jul 16, 2021 at 01:16:02PM +0300, laurentiu.tudor@nxp.com wrote: > >>>>>>>> From: Laurentiu Tudor <laurentiu.tudor@nxp.com> > >>>>>>>> > >>>>>>>> software_node_notify(), on KOBJ_REMOVE drops the refcount twice on managed > >>>>>>>> software nodes, thus leading to underflow errors. Balance the refcount by > >>>>>>>> bumping it in the device_create_managed_software_node() function. > >>>>>>>> > >>>>>>>> The error [1] was encountered after adding a .shutdown() op to our > >>>>>>>> fsl-mc-bus driver. > >>>>>>> > >>>>>>> Looking into the history of adding ->shutdown() to dwc3 driver (it got reverted > >>>>>>> later on), I can tell that probably something is wrong in the ->shutdown() > >>>>>>> method itself. > >>>>>> > >>>>>> Isn't the other alternative to just remove the second kobject_put from > >>>>>> KOBJ_REMOVE ? > >>>>> > >>>>> Or maybe on top of Heikki's suggestion, replace the calls to > >>>>> sysfs_create_link() from KOBJ_ADD with sysfs_create_link_nowarn()? > >>>> > >>>> _noearn will hide the problem. It was there, it was removed from there. > >>>> Perhaps we have to understand the root cause better (some specific flow?). > >>>> > >>>> Any insight from you on the flow when the issue appears? I.o.w. what happened > >>>> on the big picture that we got into the warning you see? > >>> > >>> I encountered the initial issue when trying to shut down a system booted > >>> with ACPI but only after adding a .shutdown() callback to our bus driver > >>> so that the devices are properly taken down. The problem was that > >>> software_node_notify(), on KOBJ_REMOVE was dropping the reference count > >>> twice leading to an underflow error. My initial proposal was to just > >>> bump the refcount in device_create_managed_software_node(). The device > >>> properties that triggered the problem are created here [1]. > >>> > >>> Heikko suggested that instead of manually incrementing the refcount to > >>> use software_node_notify(KOBJ_ADD). This triggered the second issue, a > >>> duplicated sysfs entry warning originating in the usb subsystem: > >>> device_create_managed_software_node() ends up being called twice, once > >>> here [2] and secondly, the place I previous mentioned [1]. > >> > >> This [3] is what I have reported against DWC3 when ->shutdown() has > >> been added there. And here [4] is another thread about the issue with > >> that callback. The ->release() callback is called at put_device() [5] > >> and ->shutdown() is called before that [6]. That said, can you inspect > >> your ->shutdown() implementation once more time and perhaps see if > >> there is anything that can be amended? > >> > > > > Will do, thanks for the pointers. It could be that we mess something out > > in how we use the driver model. > > > > Quick (and late, sorry) update from my side. I've spent time on > debugging our bus, did found some issues but, at least for now, none are > related to sw node. > In the mean time, I noticed in the swnode code that > device_add_software_node() calls software_node_notify(KOBJ_ADD) while > device_create_managed_software_node() doesn't. Updating [1] the later > with the call to software_node_notify(KOBJ_ADD) does seem to fix the > issue I'm seeing. > > Could this be a problem? Any comments appreciated. > > One more thing perhaps worth mentioning is that, at least for now, there > are few uses for this device_create_managed_software_node() api, > mentioning here a couple of them: > - arm64 iort code - this seems to be triggering the issue i'm getting > - dwc3 usb - Andy reported similar issues here, maybe the issue is common? > > [1] > @@ -1113,6 +1125,15 @@ int device_create_managed_software_node(struct > device *dev, > to_swnode(fwnode)->managed = true; > set_secondary_fwnode(dev, fwnode); > > + /* > + * If the device has been fully registered by the time this > function is > + * called, software_node_notify() must be called separately so > that the > + * symlinks get created and the reference count of the node is > kept in > + * balance. > + */ > + if (device_is_registered(dev)) > + software_node_notify(dev, KOBJ_ADD); > + > return 0; > } That should be fixed indeed. Please send that after -rc1 is out. thanks, -- heikki
On Thu, Sep 09, 2021 at 03:13:47PM +0300, Heikki Krogerus wrote: > On Tue, Sep 07, 2021 at 06:59:18PM +0300, Laurentiu Tudor wrote: > > > > > > On 7/26/2021 10:59 AM, Laurentiu Tudor wrote: > > > > > > > > > On 7/20/2021 1:27 PM, Andy Shevchenko wrote: > > >> On Tue, Jul 20, 2021 at 12:22 PM Laurentiu Tudor > > >> <laurentiu.tudor@nxp.com> wrote: > > >>> On 7/19/2021 3:22 PM, Andy Shevchenko wrote: > > >>>> On Mon, Jul 19, 2021 at 03:00:17PM +0300, Laurentiu Tudor wrote: > > >>>>> On 7/16/2021 8:21 PM, Jon Nettleton wrote: > > >>>>>> On Fri, Jul 16, 2021 at 2:17 PM Andy Shevchenko > > >>>>>> <andriy.shevchenko@linux.intel.com> wrote: > > >>>>>>> > > >>>>>>> On Fri, Jul 16, 2021 at 01:16:02PM +0300, laurentiu.tudor@nxp.com wrote: > > >>>>>>>> From: Laurentiu Tudor <laurentiu.tudor@nxp.com> > > >>>>>>>> > > >>>>>>>> software_node_notify(), on KOBJ_REMOVE drops the refcount twice on managed > > >>>>>>>> software nodes, thus leading to underflow errors. Balance the refcount by > > >>>>>>>> bumping it in the device_create_managed_software_node() function. > > >>>>>>>> > > >>>>>>>> The error [1] was encountered after adding a .shutdown() op to our > > >>>>>>>> fsl-mc-bus driver. > > >>>>>>> > > >>>>>>> Looking into the history of adding ->shutdown() to dwc3 driver (it got reverted > > >>>>>>> later on), I can tell that probably something is wrong in the ->shutdown() > > >>>>>>> method itself. > > >>>>>> > > >>>>>> Isn't the other alternative to just remove the second kobject_put from > > >>>>>> KOBJ_REMOVE ? > > >>>>> > > >>>>> Or maybe on top of Heikki's suggestion, replace the calls to > > >>>>> sysfs_create_link() from KOBJ_ADD with sysfs_create_link_nowarn()? > > >>>> > > >>>> _noearn will hide the problem. It was there, it was removed from there. > > >>>> Perhaps we have to understand the root cause better (some specific flow?). > > >>>> > > >>>> Any insight from you on the flow when the issue appears? I.o.w. what happened > > >>>> on the big picture that we got into the warning you see? > > >>> > > >>> I encountered the initial issue when trying to shut down a system booted > > >>> with ACPI but only after adding a .shutdown() callback to our bus driver > > >>> so that the devices are properly taken down. The problem was that > > >>> software_node_notify(), on KOBJ_REMOVE was dropping the reference count > > >>> twice leading to an underflow error. My initial proposal was to just > > >>> bump the refcount in device_create_managed_software_node(). The device > > >>> properties that triggered the problem are created here [1]. > > >>> > > >>> Heikko suggested that instead of manually incrementing the refcount to > > >>> use software_node_notify(KOBJ_ADD). This triggered the second issue, a > > >>> duplicated sysfs entry warning originating in the usb subsystem: > > >>> device_create_managed_software_node() ends up being called twice, once > > >>> here [2] and secondly, the place I previous mentioned [1]. > > >> > > >> This [3] is what I have reported against DWC3 when ->shutdown() has > > >> been added there. And here [4] is another thread about the issue with > > >> that callback. The ->release() callback is called at put_device() [5] > > >> and ->shutdown() is called before that [6]. That said, can you inspect > > >> your ->shutdown() implementation once more time and perhaps see if > > >> there is anything that can be amended? > > >> > > > > > > Will do, thanks for the pointers. It could be that we mess something out > > > in how we use the driver model. > > > > > > > Quick (and late, sorry) update from my side. I've spent time on > > debugging our bus, did found some issues but, at least for now, none are > > related to sw node. > > In the mean time, I noticed in the swnode code that > > device_add_software_node() calls software_node_notify(KOBJ_ADD) while > > device_create_managed_software_node() doesn't. Updating [1] the later > > with the call to software_node_notify(KOBJ_ADD) does seem to fix the > > issue I'm seeing. > > > > Could this be a problem? Any comments appreciated. > > > > One more thing perhaps worth mentioning is that, at least for now, there > > are few uses for this device_create_managed_software_node() api, > > mentioning here a couple of them: > > - arm64 iort code - this seems to be triggering the issue i'm getting > > - dwc3 usb - Andy reported similar issues here, maybe the issue is common? > > > > [1] > > @@ -1113,6 +1125,15 @@ int device_create_managed_software_node(struct > > device *dev, > > to_swnode(fwnode)->managed = true; > > set_secondary_fwnode(dev, fwnode); > > > > + /* > > + * If the device has been fully registered by the time this > > function is > > + * called, software_node_notify() must be called separately so > > that the > > + * symlinks get created and the reference count of the node is > > kept in > > + * balance. > > + */ > > + if (device_is_registered(dev)) > > + software_node_notify(dev, KOBJ_ADD); > > + > > return 0; > > } > > That should be fixed indeed. Please send that after -rc1 is out. I mean, resend :-) -- heikki
On 9/9/2021 3:16 PM, Heikki Krogerus wrote: > On Thu, Sep 09, 2021 at 03:13:47PM +0300, Heikki Krogerus wrote: >> On Tue, Sep 07, 2021 at 06:59:18PM +0300, Laurentiu Tudor wrote: >>> >>> >>> On 7/26/2021 10:59 AM, Laurentiu Tudor wrote: >>>> >>>> >>>> On 7/20/2021 1:27 PM, Andy Shevchenko wrote: >>>>> On Tue, Jul 20, 2021 at 12:22 PM Laurentiu Tudor >>>>> <laurentiu.tudor@nxp.com> wrote: >>>>>> On 7/19/2021 3:22 PM, Andy Shevchenko wrote: >>>>>>> On Mon, Jul 19, 2021 at 03:00:17PM +0300, Laurentiu Tudor wrote: >>>>>>>> On 7/16/2021 8:21 PM, Jon Nettleton wrote: >>>>>>>>> On Fri, Jul 16, 2021 at 2:17 PM Andy Shevchenko >>>>>>>>> <andriy.shevchenko@linux.intel.com> wrote: >>>>>>>>>> >>>>>>>>>> On Fri, Jul 16, 2021 at 01:16:02PM +0300, laurentiu.tudor@nxp.com wrote: >>>>>>>>>>> From: Laurentiu Tudor <laurentiu.tudor@nxp.com> >>>>>>>>>>> >>>>>>>>>>> software_node_notify(), on KOBJ_REMOVE drops the refcount twice on managed >>>>>>>>>>> software nodes, thus leading to underflow errors. Balance the refcount by >>>>>>>>>>> bumping it in the device_create_managed_software_node() function. >>>>>>>>>>> >>>>>>>>>>> The error [1] was encountered after adding a .shutdown() op to our >>>>>>>>>>> fsl-mc-bus driver. >>>>>>>>>> >>>>>>>>>> Looking into the history of adding ->shutdown() to dwc3 driver (it got reverted >>>>>>>>>> later on), I can tell that probably something is wrong in the ->shutdown() >>>>>>>>>> method itself. >>>>>>>>> >>>>>>>>> Isn't the other alternative to just remove the second kobject_put from >>>>>>>>> KOBJ_REMOVE ? >>>>>>>> >>>>>>>> Or maybe on top of Heikki's suggestion, replace the calls to >>>>>>>> sysfs_create_link() from KOBJ_ADD with sysfs_create_link_nowarn()? >>>>>>> >>>>>>> _noearn will hide the problem. It was there, it was removed from there. >>>>>>> Perhaps we have to understand the root cause better (some specific flow?). >>>>>>> >>>>>>> Any insight from you on the flow when the issue appears? I.o.w. what happened >>>>>>> on the big picture that we got into the warning you see? >>>>>> >>>>>> I encountered the initial issue when trying to shut down a system booted >>>>>> with ACPI but only after adding a .shutdown() callback to our bus driver >>>>>> so that the devices are properly taken down. The problem was that >>>>>> software_node_notify(), on KOBJ_REMOVE was dropping the reference count >>>>>> twice leading to an underflow error. My initial proposal was to just >>>>>> bump the refcount in device_create_managed_software_node(). The device >>>>>> properties that triggered the problem are created here [1]. >>>>>> >>>>>> Heikko suggested that instead of manually incrementing the refcount to >>>>>> use software_node_notify(KOBJ_ADD). This triggered the second issue, a >>>>>> duplicated sysfs entry warning originating in the usb subsystem: >>>>>> device_create_managed_software_node() ends up being called twice, once >>>>>> here [2] and secondly, the place I previous mentioned [1]. >>>>> >>>>> This [3] is what I have reported against DWC3 when ->shutdown() has >>>>> been added there. And here [4] is another thread about the issue with >>>>> that callback. The ->release() callback is called at put_device() [5] >>>>> and ->shutdown() is called before that [6]. That said, can you inspect >>>>> your ->shutdown() implementation once more time and perhaps see if >>>>> there is anything that can be amended? >>>>> >>>> >>>> Will do, thanks for the pointers. It could be that we mess something out >>>> in how we use the driver model. >>>> >>> >>> Quick (and late, sorry) update from my side. I've spent time on >>> debugging our bus, did found some issues but, at least for now, none are >>> related to sw node. >>> In the mean time, I noticed in the swnode code that >>> device_add_software_node() calls software_node_notify(KOBJ_ADD) while >>> device_create_managed_software_node() doesn't. Updating [1] the later >>> with the call to software_node_notify(KOBJ_ADD) does seem to fix the >>> issue I'm seeing. >>> >>> Could this be a problem? Any comments appreciated. >>> >>> One more thing perhaps worth mentioning is that, at least for now, there >>> are few uses for this device_create_managed_software_node() api, >>> mentioning here a couple of them: >>> - arm64 iort code - this seems to be triggering the issue i'm getting >>> - dwc3 usb - Andy reported similar issues here, maybe the issue is common? >>> >>> [1] >>> @@ -1113,6 +1125,15 @@ int device_create_managed_software_node(struct >>> device *dev, >>> to_swnode(fwnode)->managed = true; >>> set_secondary_fwnode(dev, fwnode); >>> >>> + /* >>> + * If the device has been fully registered by the time this >>> function is >>> + * called, software_node_notify() must be called separately so >>> that the >>> + * symlinks get created and the reference count of the node is >>> kept in >>> + * balance. >>> + */ >>> + if (device_is_registered(dev)) >>> + software_node_notify(dev, KOBJ_ADD); >>> + >>> return 0; >>> } >> >> That should be fixed indeed. Please send that after -rc1 is out. > > I mean, resend :-) > Right, actually I just noticed that this is the fix you suggested last time we discussed. :-) I forgot about it, sorry. There's still the WARN_ON() [1] triggered by the usb subsys, apparently happening only (!) in ACPI boot scenario, so +Lorenzo. I'll delay the sending a bit to try to understand what's going on. [1] [ 11.760346] sysfs: cannot create duplicate filename '/devices/platform/808622B7:01/xhci-hcd.2.auto/software_node' [ 11.770612] CPU: 9 PID: 1 Comm: swapper/0 Tainted: G W 5.14.0-rc1-00214-gbf7f1083ebd3-dirty #62 [ 11.780611] Hardware name: NXP NXP LX2160ARDB Platform, BIOS EDK II Apr 16 2021 [ 11.787913] Call trace: [ 11.790351] dump_backtrace+0x0/0x2a4 [ 11.794017] show_stack+0x1c/0x30 [ 11.797331] dump_stack_lvl+0x68/0x84 [ 11.800991] dump_stack+0x20/0x3c [ 11.804302] sysfs_warn_dup+0x88/0xac [ 11.807965] sysfs_do_create_link_sd+0xf8/0x100 [ 11.812492] sysfs_create_link+0x48/0x80 [ 11.816411] software_node_notify+0x1a8/0x35c [ 11.820769] device_create_managed_software_node+0x158/0x1b0 [ 11.826428] iort_named_component_init+0xe0/0x140 [ 11.831131] iort_iommu_configure_id+0xf4/0x270 [ 11.835660] acpi_dma_configure_id+0x160/0x200 [ 11.840101] platform_dma_configure+0xa0/0xa4 [ 11.844457] really_probe.part.0+0x84/0x480 [ 11.848639] __driver_probe_device+0xd4/0x180 [ 11.852994] driver_probe_device+0xf8/0x1e0 [ 11.857174] __driver_attach+0x108/0x220 [ 11.861095] bus_for_each_dev+0xe4/0x154 [ 11.865014] driver_attach+0x38/0x50 [ 11.868587] bus_add_driver+0x1bc/0x2c4 [ 11.872419] driver_register+0xf0/0x210 [ 11.876253] __platform_driver_register+0x48/0x60 [ 11.880956] xhci_plat_init+0x34/0x40 [ 11.884616] do_one_initcall+0xa8/0x270 [ 11.888449] kernel_init_freeable+0x2c0/0x348 [ 11.892806] kernel_init+0x28/0x140 [ 11.896295] ret_from_fork+0x10/0x18 [ 11.900062] xhci-hcd xhci-hcd.2.auto: Adding to iommu group 6 [ 11.906044] xhci-hcd xhci-hcd.2.auto: xHCI Host Controller [ 11.911566] xhci-hcd xhci-hcd.2.auto: new USB bus registered, assigned bus number 3 [ 11.919702] xhci-hcd xhci-hcd.2.auto: hcc params 0x0220f66d hci version 0x100 quirks 0x0000000000010010 [ 11.929187] xhci-hcd xhci-hcd.2.auto: irq 106, io mem 0x03110000 --- Best Regards, Laurentiu
On 9/9/2021 5:01 PM, Laurentiu Tudor wrote: > > > On 9/9/2021 3:16 PM, Heikki Krogerus wrote: >> On Thu, Sep 09, 2021 at 03:13:47PM +0300, Heikki Krogerus wrote: >>> On Tue, Sep 07, 2021 at 06:59:18PM +0300, Laurentiu Tudor wrote: >>>> >>>> >>>> On 7/26/2021 10:59 AM, Laurentiu Tudor wrote: >>>>> >>>>> >>>>> On 7/20/2021 1:27 PM, Andy Shevchenko wrote: >>>>>> On Tue, Jul 20, 2021 at 12:22 PM Laurentiu Tudor >>>>>> <laurentiu.tudor@nxp.com> wrote: >>>>>>> On 7/19/2021 3:22 PM, Andy Shevchenko wrote: >>>>>>>> On Mon, Jul 19, 2021 at 03:00:17PM +0300, Laurentiu Tudor wrote: >>>>>>>>> On 7/16/2021 8:21 PM, Jon Nettleton wrote: >>>>>>>>>> On Fri, Jul 16, 2021 at 2:17 PM Andy Shevchenko >>>>>>>>>> <andriy.shevchenko@linux.intel.com> wrote: >>>>>>>>>>> >>>>>>>>>>> On Fri, Jul 16, 2021 at 01:16:02PM +0300, laurentiu.tudor@nxp.com wrote: >>>>>>>>>>>> From: Laurentiu Tudor <laurentiu.tudor@nxp.com> >>>>>>>>>>>> >>>>>>>>>>>> software_node_notify(), on KOBJ_REMOVE drops the refcount twice on managed >>>>>>>>>>>> software nodes, thus leading to underflow errors. Balance the refcount by >>>>>>>>>>>> bumping it in the device_create_managed_software_node() function. >>>>>>>>>>>> >>>>>>>>>>>> The error [1] was encountered after adding a .shutdown() op to our >>>>>>>>>>>> fsl-mc-bus driver. >>>>>>>>>>> >>>>>>>>>>> Looking into the history of adding ->shutdown() to dwc3 driver (it got reverted >>>>>>>>>>> later on), I can tell that probably something is wrong in the ->shutdown() >>>>>>>>>>> method itself. >>>>>>>>>> >>>>>>>>>> Isn't the other alternative to just remove the second kobject_put from >>>>>>>>>> KOBJ_REMOVE ? >>>>>>>>> >>>>>>>>> Or maybe on top of Heikki's suggestion, replace the calls to >>>>>>>>> sysfs_create_link() from KOBJ_ADD with sysfs_create_link_nowarn()? >>>>>>>> >>>>>>>> _noearn will hide the problem. It was there, it was removed from there. >>>>>>>> Perhaps we have to understand the root cause better (some specific flow?). >>>>>>>> >>>>>>>> Any insight from you on the flow when the issue appears? I.o.w. what happened >>>>>>>> on the big picture that we got into the warning you see? >>>>>>> >>>>>>> I encountered the initial issue when trying to shut down a system booted >>>>>>> with ACPI but only after adding a .shutdown() callback to our bus driver >>>>>>> so that the devices are properly taken down. The problem was that >>>>>>> software_node_notify(), on KOBJ_REMOVE was dropping the reference count >>>>>>> twice leading to an underflow error. My initial proposal was to just >>>>>>> bump the refcount in device_create_managed_software_node(). The device >>>>>>> properties that triggered the problem are created here [1]. >>>>>>> >>>>>>> Heikko suggested that instead of manually incrementing the refcount to >>>>>>> use software_node_notify(KOBJ_ADD). This triggered the second issue, a >>>>>>> duplicated sysfs entry warning originating in the usb subsystem: >>>>>>> device_create_managed_software_node() ends up being called twice, once >>>>>>> here [2] and secondly, the place I previous mentioned [1]. >>>>>> >>>>>> This [3] is what I have reported against DWC3 when ->shutdown() has >>>>>> been added there. And here [4] is another thread about the issue with >>>>>> that callback. The ->release() callback is called at put_device() [5] >>>>>> and ->shutdown() is called before that [6]. That said, can you inspect >>>>>> your ->shutdown() implementation once more time and perhaps see if >>>>>> there is anything that can be amended? >>>>>> >>>>> >>>>> Will do, thanks for the pointers. It could be that we mess something out >>>>> in how we use the driver model. >>>>> >>>> >>>> Quick (and late, sorry) update from my side. I've spent time on >>>> debugging our bus, did found some issues but, at least for now, none are >>>> related to sw node. >>>> In the mean time, I noticed in the swnode code that >>>> device_add_software_node() calls software_node_notify(KOBJ_ADD) while >>>> device_create_managed_software_node() doesn't. Updating [1] the later >>>> with the call to software_node_notify(KOBJ_ADD) does seem to fix the >>>> issue I'm seeing. >>>> >>>> Could this be a problem? Any comments appreciated. >>>> >>>> One more thing perhaps worth mentioning is that, at least for now, there >>>> are few uses for this device_create_managed_software_node() api, >>>> mentioning here a couple of them: >>>> - arm64 iort code - this seems to be triggering the issue i'm getting >>>> - dwc3 usb - Andy reported similar issues here, maybe the issue is common? >>>> >>>> [1] >>>> @@ -1113,6 +1125,15 @@ int device_create_managed_software_node(struct >>>> device *dev, >>>> to_swnode(fwnode)->managed = true; >>>> set_secondary_fwnode(dev, fwnode); >>>> >>>> + /* >>>> + * If the device has been fully registered by the time this >>>> function is >>>> + * called, software_node_notify() must be called separately so >>>> that the >>>> + * symlinks get created and the reference count of the node is >>>> kept in >>>> + * balance. >>>> + */ >>>> + if (device_is_registered(dev)) >>>> + software_node_notify(dev, KOBJ_ADD); >>>> + >>>> return 0; >>>> } >>> >>> That should be fixed indeed. Please send that after -rc1 is out. >> >> I mean, resend :-) >> > > Right, actually I just noticed that this is the fix you suggested last > time we discussed. :-) I forgot about it, sorry. > There's still the WARN_ON() [1] triggered by the usb subsys, apparently > happening only (!) in ACPI boot scenario, so +Lorenzo. > I'll delay the sending a bit to try to understand what's going on. I've spent some time looking into this and it turns out that in the ACPI case, device_create_managed_software_node() ends up being called twice, first here [1] and after that, in the IORT code here [2]. With the proposed patch this causes software_node_notify(KOBJ_ADD) being called twice thus triggering the dup sysfs entry warning. Any comments / ideas welcomed. [1] https://elixir.bootlin.com/linux/latest/source/drivers/usb/dwc3/host.c#L111 [2] https://elixir.bootlin.com/linux/latest/source/drivers/acpi/arm64/iort.c#L952 --- Best Regards, Laurentiu
On Fri, Sep 10, 2021 at 03:05:16PM +0300, Laurentiu Tudor wrote: > > > On 9/9/2021 5:01 PM, Laurentiu Tudor wrote: > > > > > > On 9/9/2021 3:16 PM, Heikki Krogerus wrote: > >> On Thu, Sep 09, 2021 at 03:13:47PM +0300, Heikki Krogerus wrote: > >>> On Tue, Sep 07, 2021 at 06:59:18PM +0300, Laurentiu Tudor wrote: > >>>> > >>>> > >>>> On 7/26/2021 10:59 AM, Laurentiu Tudor wrote: > >>>>> > >>>>> > >>>>> On 7/20/2021 1:27 PM, Andy Shevchenko wrote: > >>>>>> On Tue, Jul 20, 2021 at 12:22 PM Laurentiu Tudor > >>>>>> <laurentiu.tudor@nxp.com> wrote: > >>>>>>> On 7/19/2021 3:22 PM, Andy Shevchenko wrote: > >>>>>>>> On Mon, Jul 19, 2021 at 03:00:17PM +0300, Laurentiu Tudor wrote: > >>>>>>>>> On 7/16/2021 8:21 PM, Jon Nettleton wrote: > >>>>>>>>>> On Fri, Jul 16, 2021 at 2:17 PM Andy Shevchenko > >>>>>>>>>> <andriy.shevchenko@linux.intel.com> wrote: > >>>>>>>>>>> > >>>>>>>>>>> On Fri, Jul 16, 2021 at 01:16:02PM +0300, laurentiu.tudor@nxp.com wrote: > >>>>>>>>>>>> From: Laurentiu Tudor <laurentiu.tudor@nxp.com> > >>>>>>>>>>>> > >>>>>>>>>>>> software_node_notify(), on KOBJ_REMOVE drops the refcount twice on managed > >>>>>>>>>>>> software nodes, thus leading to underflow errors. Balance the refcount by > >>>>>>>>>>>> bumping it in the device_create_managed_software_node() function. > >>>>>>>>>>>> > >>>>>>>>>>>> The error [1] was encountered after adding a .shutdown() op to our > >>>>>>>>>>>> fsl-mc-bus driver. > >>>>>>>>>>> > >>>>>>>>>>> Looking into the history of adding ->shutdown() to dwc3 driver (it got reverted > >>>>>>>>>>> later on), I can tell that probably something is wrong in the ->shutdown() > >>>>>>>>>>> method itself. > >>>>>>>>>> > >>>>>>>>>> Isn't the other alternative to just remove the second kobject_put from > >>>>>>>>>> KOBJ_REMOVE ? > >>>>>>>>> > >>>>>>>>> Or maybe on top of Heikki's suggestion, replace the calls to > >>>>>>>>> sysfs_create_link() from KOBJ_ADD with sysfs_create_link_nowarn()? > >>>>>>>> > >>>>>>>> _noearn will hide the problem. It was there, it was removed from there. > >>>>>>>> Perhaps we have to understand the root cause better (some specific flow?). > >>>>>>>> > >>>>>>>> Any insight from you on the flow when the issue appears? I.o.w. what happened > >>>>>>>> on the big picture that we got into the warning you see? > >>>>>>> > >>>>>>> I encountered the initial issue when trying to shut down a system booted > >>>>>>> with ACPI but only after adding a .shutdown() callback to our bus driver > >>>>>>> so that the devices are properly taken down. The problem was that > >>>>>>> software_node_notify(), on KOBJ_REMOVE was dropping the reference count > >>>>>>> twice leading to an underflow error. My initial proposal was to just > >>>>>>> bump the refcount in device_create_managed_software_node(). The device > >>>>>>> properties that triggered the problem are created here [1]. > >>>>>>> > >>>>>>> Heikko suggested that instead of manually incrementing the refcount to > >>>>>>> use software_node_notify(KOBJ_ADD). This triggered the second issue, a > >>>>>>> duplicated sysfs entry warning originating in the usb subsystem: > >>>>>>> device_create_managed_software_node() ends up being called twice, once > >>>>>>> here [2] and secondly, the place I previous mentioned [1]. > >>>>>> > >>>>>> This [3] is what I have reported against DWC3 when ->shutdown() has > >>>>>> been added there. And here [4] is another thread about the issue with > >>>>>> that callback. The ->release() callback is called at put_device() [5] > >>>>>> and ->shutdown() is called before that [6]. That said, can you inspect > >>>>>> your ->shutdown() implementation once more time and perhaps see if > >>>>>> there is anything that can be amended? > >>>>>> > >>>>> > >>>>> Will do, thanks for the pointers. It could be that we mess something out > >>>>> in how we use the driver model. > >>>>> > >>>> > >>>> Quick (and late, sorry) update from my side. I've spent time on > >>>> debugging our bus, did found some issues but, at least for now, none are > >>>> related to sw node. > >>>> In the mean time, I noticed in the swnode code that > >>>> device_add_software_node() calls software_node_notify(KOBJ_ADD) while > >>>> device_create_managed_software_node() doesn't. Updating [1] the later > >>>> with the call to software_node_notify(KOBJ_ADD) does seem to fix the > >>>> issue I'm seeing. > >>>> > >>>> Could this be a problem? Any comments appreciated. > >>>> > >>>> One more thing perhaps worth mentioning is that, at least for now, there > >>>> are few uses for this device_create_managed_software_node() api, > >>>> mentioning here a couple of them: > >>>> - arm64 iort code - this seems to be triggering the issue i'm getting > >>>> - dwc3 usb - Andy reported similar issues here, maybe the issue is common? > >>>> > >>>> [1] > >>>> @@ -1113,6 +1125,15 @@ int device_create_managed_software_node(struct > >>>> device *dev, > >>>> to_swnode(fwnode)->managed = true; > >>>> set_secondary_fwnode(dev, fwnode); > >>>> > >>>> + /* > >>>> + * If the device has been fully registered by the time this > >>>> function is > >>>> + * called, software_node_notify() must be called separately so > >>>> that the > >>>> + * symlinks get created and the reference count of the node is > >>>> kept in > >>>> + * balance. > >>>> + */ > >>>> + if (device_is_registered(dev)) > >>>> + software_node_notify(dev, KOBJ_ADD); > >>>> + > >>>> return 0; > >>>> } > >>> > >>> That should be fixed indeed. Please send that after -rc1 is out. > >> > >> I mean, resend :-) > >> > > > > Right, actually I just noticed that this is the fix you suggested last > > time we discussed. :-) I forgot about it, sorry. > > There's still the WARN_ON() [1] triggered by the usb subsys, apparently > > happening only (!) in ACPI boot scenario, so +Lorenzo. > > I'll delay the sending a bit to try to understand what's going on. > > I've spent some time looking into this and it turns out that in the > ACPI case, device_create_managed_software_node() ends up being called > twice, first here [1] and after that, in the IORT code here [2]. With > the proposed patch this causes software_node_notify(KOBJ_ADD) being > called twice thus triggering the dup sysfs entry warning. > Any comments / ideas welcomed. > > [1] https://elixir.bootlin.com/linux/latest/source/drivers/usb/dwc3/host.c#L111 I think the problem here is that the secondary fwnode get's replaced because the primary fwnode is shared. Can you test it with this, just to see if the problem goes away: diff --git a/drivers/usb/dwc3/host.c b/drivers/usb/dwc3/host.c index f29a264635aa1..e4b40f8b8f242 100644 --- a/drivers/usb/dwc3/host.c +++ b/drivers/usb/dwc3/host.c @@ -76,7 +76,6 @@ int dwc3_host_init(struct dwc3 *dwc) } xhci->dev.parent = dwc->dev; - ACPI_COMPANION_SET(&xhci->dev, ACPI_COMPANION(dwc->dev)); dwc->xhci = xhci; > [2] https://elixir.bootlin.com/linux/latest/source/drivers/acpi/arm64/iort.c#L952 I didn't yet look at that one. thanks, -- heikki
Hi Heikki, On 9/10/2021 3:38 PM, Heikki Krogerus wrote: > On Fri, Sep 10, 2021 at 03:05:16PM +0300, Laurentiu Tudor wrote: >> >> >> On 9/9/2021 5:01 PM, Laurentiu Tudor wrote: >>> >>> >>> On 9/9/2021 3:16 PM, Heikki Krogerus wrote: >>>> On Thu, Sep 09, 2021 at 03:13:47PM +0300, Heikki Krogerus wrote: >>>>> On Tue, Sep 07, 2021 at 06:59:18PM +0300, Laurentiu Tudor wrote: >>>>>> >>>>>> >>>>>> On 7/26/2021 10:59 AM, Laurentiu Tudor wrote: >>>>>>> >>>>>>> >>>>>>> On 7/20/2021 1:27 PM, Andy Shevchenko wrote: >>>>>>>> On Tue, Jul 20, 2021 at 12:22 PM Laurentiu Tudor >>>>>>>> <laurentiu.tudor@nxp.com> wrote: >>>>>>>>> On 7/19/2021 3:22 PM, Andy Shevchenko wrote: >>>>>>>>>> On Mon, Jul 19, 2021 at 03:00:17PM +0300, Laurentiu Tudor wrote: >>>>>>>>>>> On 7/16/2021 8:21 PM, Jon Nettleton wrote: >>>>>>>>>>>> On Fri, Jul 16, 2021 at 2:17 PM Andy Shevchenko >>>>>>>>>>>> <andriy.shevchenko@linux.intel.com> wrote: >>>>>>>>>>>>> >>>>>>>>>>>>> On Fri, Jul 16, 2021 at 01:16:02PM +0300, laurentiu.tudor@nxp.com wrote: >>>>>>>>>>>>>> From: Laurentiu Tudor <laurentiu.tudor@nxp.com> >>>>>>>>>>>>>> >>>>>>>>>>>>>> software_node_notify(), on KOBJ_REMOVE drops the refcount twice on managed >>>>>>>>>>>>>> software nodes, thus leading to underflow errors. Balance the refcount by >>>>>>>>>>>>>> bumping it in the device_create_managed_software_node() function. >>>>>>>>>>>>>> >>>>>>>>>>>>>> The error [1] was encountered after adding a .shutdown() op to our >>>>>>>>>>>>>> fsl-mc-bus driver. >>>>>>>>>>>>> >>>>>>>>>>>>> Looking into the history of adding ->shutdown() to dwc3 driver (it got reverted >>>>>>>>>>>>> later on), I can tell that probably something is wrong in the ->shutdown() >>>>>>>>>>>>> method itself. >>>>>>>>>>>> >>>>>>>>>>>> Isn't the other alternative to just remove the second kobject_put from >>>>>>>>>>>> KOBJ_REMOVE ? >>>>>>>>>>> >>>>>>>>>>> Or maybe on top of Heikki's suggestion, replace the calls to >>>>>>>>>>> sysfs_create_link() from KOBJ_ADD with sysfs_create_link_nowarn()? >>>>>>>>>> >>>>>>>>>> _noearn will hide the problem. It was there, it was removed from there. >>>>>>>>>> Perhaps we have to understand the root cause better (some specific flow?). >>>>>>>>>> >>>>>>>>>> Any insight from you on the flow when the issue appears? I.o.w. what happened >>>>>>>>>> on the big picture that we got into the warning you see? >>>>>>>>> >>>>>>>>> I encountered the initial issue when trying to shut down a system booted >>>>>>>>> with ACPI but only after adding a .shutdown() callback to our bus driver >>>>>>>>> so that the devices are properly taken down. The problem was that >>>>>>>>> software_node_notify(), on KOBJ_REMOVE was dropping the reference count >>>>>>>>> twice leading to an underflow error. My initial proposal was to just >>>>>>>>> bump the refcount in device_create_managed_software_node(). The device >>>>>>>>> properties that triggered the problem are created here [1]. >>>>>>>>> >>>>>>>>> Heikko suggested that instead of manually incrementing the refcount to >>>>>>>>> use software_node_notify(KOBJ_ADD). This triggered the second issue, a >>>>>>>>> duplicated sysfs entry warning originating in the usb subsystem: >>>>>>>>> device_create_managed_software_node() ends up being called twice, once >>>>>>>>> here [2] and secondly, the place I previous mentioned [1]. >>>>>>>> >>>>>>>> This [3] is what I have reported against DWC3 when ->shutdown() has >>>>>>>> been added there. And here [4] is another thread about the issue with >>>>>>>> that callback. The ->release() callback is called at put_device() [5] >>>>>>>> and ->shutdown() is called before that [6]. That said, can you inspect >>>>>>>> your ->shutdown() implementation once more time and perhaps see if >>>>>>>> there is anything that can be amended? >>>>>>>> >>>>>>> >>>>>>> Will do, thanks for the pointers. It could be that we mess something out >>>>>>> in how we use the driver model. >>>>>>> >>>>>> >>>>>> Quick (and late, sorry) update from my side. I've spent time on >>>>>> debugging our bus, did found some issues but, at least for now, none are >>>>>> related to sw node. >>>>>> In the mean time, I noticed in the swnode code that >>>>>> device_add_software_node() calls software_node_notify(KOBJ_ADD) while >>>>>> device_create_managed_software_node() doesn't. Updating [1] the later >>>>>> with the call to software_node_notify(KOBJ_ADD) does seem to fix the >>>>>> issue I'm seeing. >>>>>> >>>>>> Could this be a problem? Any comments appreciated. >>>>>> >>>>>> One more thing perhaps worth mentioning is that, at least for now, there >>>>>> are few uses for this device_create_managed_software_node() api, >>>>>> mentioning here a couple of them: >>>>>> - arm64 iort code - this seems to be triggering the issue i'm getting >>>>>> - dwc3 usb - Andy reported similar issues here, maybe the issue is common? >>>>>> >>>>>> [1] >>>>>> @@ -1113,6 +1125,15 @@ int device_create_managed_software_node(struct >>>>>> device *dev, >>>>>> to_swnode(fwnode)->managed = true; >>>>>> set_secondary_fwnode(dev, fwnode); >>>>>> >>>>>> + /* >>>>>> + * If the device has been fully registered by the time this >>>>>> function is >>>>>> + * called, software_node_notify() must be called separately so >>>>>> that the >>>>>> + * symlinks get created and the reference count of the node is >>>>>> kept in >>>>>> + * balance. >>>>>> + */ >>>>>> + if (device_is_registered(dev)) >>>>>> + software_node_notify(dev, KOBJ_ADD); >>>>>> + >>>>>> return 0; >>>>>> } >>>>> >>>>> That should be fixed indeed. Please send that after -rc1 is out. >>>> >>>> I mean, resend :-) >>>> >>> >>> Right, actually I just noticed that this is the fix you suggested last >>> time we discussed. :-) I forgot about it, sorry. >>> There's still the WARN_ON() [1] triggered by the usb subsys, apparently >>> happening only (!) in ACPI boot scenario, so +Lorenzo. >>> I'll delay the sending a bit to try to understand what's going on. >> >> I've spent some time looking into this and it turns out that in the >> ACPI case, device_create_managed_software_node() ends up being called >> twice, first here [1] and after that, in the IORT code here [2]. With >> the proposed patch this causes software_node_notify(KOBJ_ADD) being >> called twice thus triggering the dup sysfs entry warning. >> Any comments / ideas welcomed. >> >> [1] https://elixir.bootlin.com/linux/latest/source/drivers/usb/dwc3/host.c#L111 > > I think the problem here is that the secondary fwnode get's replaced > because the primary fwnode is shared. Can you test it with this, just > to see if the problem goes away: > > diff --git a/drivers/usb/dwc3/host.c b/drivers/usb/dwc3/host.c > index f29a264635aa1..e4b40f8b8f242 100644 > --- a/drivers/usb/dwc3/host.c > +++ b/drivers/usb/dwc3/host.c > @@ -76,7 +76,6 @@ int dwc3_host_init(struct dwc3 *dwc) > } > > xhci->dev.parent = dwc->dev; > - ACPI_COMPANION_SET(&xhci->dev, ACPI_COMPANION(dwc->dev)); > > dwc->xhci = xhci; Thanks for looking into this! Yes, this does make the issue go away. --- Best Regards, Laurentiu
On Fri, Jul 16, 2021 at 01:16:02PM +0300, laurentiu.tudor@nxp.com wrote: > From: Laurentiu Tudor <laurentiu.tudor@nxp.com> > > software_node_notify(), on KOBJ_REMOVE drops the refcount twice on managed > software nodes, thus leading to underflow errors. Balance the refcount by > bumping it in the device_create_managed_software_node() function. > > The error [1] was encountered after adding a .shutdown() op to our > fsl-mc-bus driver. > > [1] > pc : refcount_warn_saturate+0xf8/0x150 > lr : refcount_warn_saturate+0xf8/0x150 > sp : ffff80001009b920 > x29: ffff80001009b920 x28: ffff1a2420318000 x27: 0000000000000000 > x26: ffffccac15e7a038 x25: 0000000000000008 x24: ffffccac168e0030 > x23: ffff1a2428a82000 x22: 0000000000080000 x21: ffff1a24287b5000 > x20: 0000000000000001 x19: ffff1a24261f4400 x18: ffffffffffffffff > x17: 6f72645f726f7272 x16: 0000000000000000 x15: ffff80009009b607 > x14: 0000000000000000 x13: ffffccac16602670 x12: 0000000000000a17 > x11: 000000000000035d x10: ffffccac16602670 x9 : ffffccac16602670 > x8 : 00000000ffffefff x7 : ffffccac1665a670 x6 : ffffccac1665a670 > x5 : 0000000000000000 x4 : 0000000000000000 x3 : 00000000ffffffff > x2 : 0000000000000000 x1 : 0000000000000000 x0 : ffff1a2420318000 > Call trace: > refcount_warn_saturate+0xf8/0x150 > kobject_put+0x10c/0x120 > software_node_notify+0xd8/0x140 > device_platform_notify+0x4c/0xb4 > device_del+0x188/0x424 > fsl_mc_device_remove+0x2c/0x4c > rebofind sp.c__fsl_mc_device_remove+0x14/0x2c > device_for_each_child+0x5c/0xac > dprc_remove+0x9c/0xc0 > fsl_mc_driver_remove+0x28/0x64 > __device_release_driver+0x188/0x22c > device_release_driver+0x30/0x50 > bus_remove_device+0x128/0x134 > device_del+0x16c/0x424 > fsl_mc_bus_remove+0x8c/0x114 > fsl_mc_bus_shutdown+0x14/0x20 > platform_shutdown+0x28/0x40 > device_shutdown+0x15c/0x330 > __do_sys_reboot+0x218/0x2a0 > __arm64_sys_reboot+0x28/0x34 > invoke_syscall+0x48/0x114 > el0_svc_common+0x40/0xdc > do_el0_svc+0x2c/0x94 > el0_svc+0x2c/0x54 > el0t_64_sync_handler+0xa8/0x12c > el0t_64_sync+0x198/0x19c > ---[ end trace 32eb1c71c7d86821 ]--- > > Reported-by: Jon Nettleton <jon@solid-run.com> > Suggested-by: Heikki Krogerus <heikki.krogerus@linux.intel.com> > Signed-off-by: Laurentiu Tudor <laurentiu.tudor@nxp.com> FWIW: Reviewed-by: Heikki Krogerus <heikki.krogerus@linux.intel.com> > --- > Changes since RFC: > - use software_node_notify(KOBJ_ADD) instead of directly bumping > refcount (Heikki) > > drivers/base/swnode.c | 3 +++ > 1 file changed, 3 insertions(+) > > diff --git a/drivers/base/swnode.c b/drivers/base/swnode.c > index d1f1a8240120..bdb50a06c82a 100644 > --- a/drivers/base/swnode.c > +++ b/drivers/base/swnode.c > @@ -1113,6 +1113,9 @@ int device_create_managed_software_node(struct device *dev, > to_swnode(fwnode)->managed = true; > set_secondary_fwnode(dev, fwnode); > > + if (device_is_registered(dev)) > + software_node_notify(dev, KOBJ_ADD); > + > return 0; > } > EXPORT_SYMBOL_GPL(device_create_managed_software_node); > -- > 2.17.1 thanks, -- heikki
On Fri, Sep 10, 2021 at 04:00:49PM +0300, Laurentiu Tudor wrote: > >> I've spent some time looking into this and it turns out that in the > >> ACPI case, device_create_managed_software_node() ends up being called > >> twice, first here [1] and after that, in the IORT code here [2]. With > >> the proposed patch this causes software_node_notify(KOBJ_ADD) being > >> called twice thus triggering the dup sysfs entry warning. > >> Any comments / ideas welcomed. > >> > >> [1] https://elixir.bootlin.com/linux/latest/source/drivers/usb/dwc3/host.c#L111 > > > > I think the problem here is that the secondary fwnode get's replaced > > because the primary fwnode is shared. Can you test it with this, just > > to see if the problem goes away: > > > > diff --git a/drivers/usb/dwc3/host.c b/drivers/usb/dwc3/host.c > > index f29a264635aa1..e4b40f8b8f242 100644 > > --- a/drivers/usb/dwc3/host.c > > +++ b/drivers/usb/dwc3/host.c > > @@ -76,7 +76,6 @@ int dwc3_host_init(struct dwc3 *dwc) > > } > > > > xhci->dev.parent = dwc->dev; > > - ACPI_COMPANION_SET(&xhci->dev, ACPI_COMPANION(dwc->dev)); > > > > dwc->xhci = xhci; > > > Thanks for looking into this! Yes, this does make the issue go away. We need to think about how to solve this one. The problem is that we have to share the ACPI node between the parent dwc3 device and child xHCI, but at the same xHCI needs to have its own software node. The fwnode->secondary pointer does not quite bend to this. If the primary fwnode is shared, the secondary fwnode has to be shared as well. thanks, -- heikki
diff --git a/drivers/base/swnode.c b/drivers/base/swnode.c index d1f1a8240120..bdb50a06c82a 100644 --- a/drivers/base/swnode.c +++ b/drivers/base/swnode.c @@ -1113,6 +1113,9 @@ int device_create_managed_software_node(struct device *dev, to_swnode(fwnode)->managed = true; set_secondary_fwnode(dev, fwnode); + if (device_is_registered(dev)) + software_node_notify(dev, KOBJ_ADD); + return 0; } EXPORT_SYMBOL_GPL(device_create_managed_software_node);