Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]
Groups > linux.kernel > #1545847 > unrolled thread
| Started by | Hans de Goede <hdegoede@redhat.com> |
|---|---|
| First post | 2016-12-21 17:00 +0100 |
| Last post | 2016-12-27 06:10 +0100 |
| Articles | 3 — 2 participants |
Back to article view | Back to linux.kernel
Unregistering extcon providers while they are in use leads to kernel crashes Hans de Goede <hdegoede@redhat.com> - 2016-12-21 17:00 +0100
Re: Unregistering extcon providers while they are in use leads to kernel crashes Hans de Goede <hdegoede@redhat.com> - 2016-12-21 17:00 +0100
Re: Unregistering extcon providers while they are in use leads to kernel crashes Chanwoo Choi <cw00.choi@samsung.com> - 2016-12-27 06:10 +0100
| From | Hans de Goede <hdegoede@redhat.com> |
|---|---|
| Date | 2016-12-21 17:00 +0100 |
| Subject | Unregistering extcon providers while they are in use leads to kernel crashes |
| Message-ID | <sQRt7-5jR-3@gated-at.bofh.it> |
Hi, With the recent extcon work I've been doing I noticed that if I want to rmmod and then insmod say extcon_axp288 I can do so without problems even if axp288_charger is holding a reference to the extcon device returned by extcon_get_extcon_dev. The problem is that extcon_get_extcon_dev simply looks up the extcon-device in the list of current registered extcon-s and then returns a pointer to it, without any reference counting. The rmmod scenario can be fixed by doing a module_get from extcon_get_extcon_dev, but that still leaves the same problem when root manually unbinds the driver through sysfs. A possible way fix this would be: 1) Make all extcon providers use devm_extcon_dev_allocate and document using this to allocate an extcon_dev mandatory 2) Add a refcount to struct extcon_dev and introduce extcon_dev_get and extcon_dev_put helpers which modify the refcount and only free the memory on the final put (and make the evm_extcon_dev_allocate cleanup function call extcon_dev_put) 3) On extcon_dev_unregister set a flag in the extcon_dev that it has been free-ed, make all extcon consumer functions which take an extcon_dev (extcon_get_state, extcon_register_notifier, etc.) check this flag and return -ENODEV when the extcon has been unregistered 4) Make extcon_get_extcon_dev call extcon_dev_get on the returned edev before returning it From here on we've fixed the crash, but we now leak the extcon_dev when the consumer gets unbound. 5) Add a devm_extcon_get_extcon_dev which calls extcon_dev_put as the devm cleanup function 6) Convert all extcon consumers to use devm_extcon_get_extcon_dev Regards, Hans
[toc] | [next] | [standalone]
| From | Hans de Goede <hdegoede@redhat.com> |
|---|---|
| Date | 2016-12-21 17:00 +0100 |
| Subject | Re: Unregistering extcon providers while they are in use leads to kernel crashes |
| Message-ID | <sQRt8-5jR-9@gated-at.bofh.it> |
| In reply to | #1545847 |
Hi, On 21-12-16 16:54, Hans de Goede wrote: > Hi, > > With the recent extcon work I've been doing I noticed that > if I want to rmmod and then insmod say extcon_axp288 I can > do so without problems even if axp288_charger is holding > a reference to the extcon device returned by extcon_get_extcon_dev. > > The problem is that extcon_get_extcon_dev simply looks up > the extcon-device in the list of current registered extcon-s > and then returns a pointer to it, without any reference > counting. > > The rmmod scenario can be fixed by doing a module_get from > extcon_get_extcon_dev, but that still leaves the same problem > when root manually unbinds the driver through sysfs. > > A possible way fix this would be: > > 1) Make all extcon providers use devm_extcon_dev_allocate and document > using this to allocate an extcon_dev mandatory > > 2) Add a refcount to struct extcon_dev and introduce extcon_dev_get > and extcon_dev_put helpers which modify the refcount and only free > the memory on the final put (and make the evm_extcon_dev_allocate > cleanup function call extcon_dev_put) > > 3) On extcon_dev_unregister set a flag in the extcon_dev that it > has been free-ed, make all extcon consumer functions which take > an extcon_dev (extcon_get_state, extcon_register_notifier, etc.) > check this flag and return -ENODEV when the extcon has been unregistered > > 4) Make extcon_get_extcon_dev call extcon_dev_get on the returned edev > before returning it > > From here on we've fixed the crash, but we now leak the extcon_dev > when the consumer gets unbound. > > 5) Add a devm_extcon_get_extcon_dev which calls extcon_dev_put as the devm > cleanup function > > 6) Convert all extcon consumers to use devm_extcon_get_extcon_dev p.s. In case it was not clear, I noticed this, but I don't have the time to fix it, which is why I wrote the above plan with the hope that someone can use it as a base to actual fix this. Regards, Hans
[toc] | [prev] | [next] | [standalone]
| From | Chanwoo Choi <cw00.choi@samsung.com> |
|---|---|
| Date | 2016-12-27 06:10 +0100 |
| Subject | Re: Unregistering extcon providers while they are in use leads to kernel crashes |
| Message-ID | <sSSbn-6bf-1@gated-at.bofh.it> |
| In reply to | #1545847 |
Hi Hans, Thanks for your report. I'll check this problem and try to resolve it. On 2016년 12월 22일 00:54, Hans de Goede wrote: > Hi, > > With the recent extcon work I've been doing I noticed that > if I want to rmmod and then insmod say extcon_axp288 I can > do so without problems even if axp288_charger is holding > a reference to the extcon device returned by extcon_get_extcon_dev. > > The problem is that extcon_get_extcon_dev simply looks up > the extcon-device in the list of current registered extcon-s > and then returns a pointer to it, without any reference > counting. > > The rmmod scenario can be fixed by doing a module_get from > extcon_get_extcon_dev, but that still leaves the same problem > when root manually unbinds the driver through sysfs. > > A possible way fix this would be: > > 1) Make all extcon providers use devm_extcon_dev_allocate and document > using this to allocate an extcon_dev mandatory > > 2) Add a refcount to struct extcon_dev and introduce extcon_dev_get > and extcon_dev_put helpers which modify the refcount and only free > the memory on the final put (and make the evm_extcon_dev_allocate > cleanup function call extcon_dev_put) > > 3) On extcon_dev_unregister set a flag in the extcon_dev that it > has been free-ed, make all extcon consumer functions which take > an extcon_dev (extcon_get_state, extcon_register_notifier, etc.) > check this flag and return -ENODEV when the extcon has been unregistered > > 4) Make extcon_get_extcon_dev call extcon_dev_get on the returned edev > before returning it > > From here on we've fixed the crash, but we now leak the extcon_dev > when the consumer gets unbound. > > 5) Add a devm_extcon_get_extcon_dev which calls extcon_dev_put as the devm > cleanup function > > 6) Convert all extcon consumers to use devm_extcon_get_extcon_dev > > Regards, > > Hans > > > -- Regards, Chanwoo Choi
[toc] | [prev] | [standalone]
Back to top | Article view | linux.kernel
csiph-web