Re: Re: [RFC PATCH 0/3] Fix errors on DT overlay removal with devlinks
On Thu, Oct 29, 2020 at 03:54:21PM -0500, Frank Rowand wrote: > On 10/28/20 11:25 AM, Michael Auchter wrote: > > Hey Saravana, > > > > Thanks for taking the time to look into this! > > > > On Mon, Oct 26, 2020 at 12:10:33PM -0700, Saravana Kannan wrote: > >> On Wed, Oct 21, 2020 at 2:02 PM Frank Rowand > >> wrote: > >>> > >>> Hi Saravana, > >>> > >>> Michael found an issue related to the removal of a devicetree node > >>> which involves devlinks: > >>> > >>> On 10/14/20 2:36 PM, Michael Auchter wrote: > After updating to v5.9, I've started seeing errors in the kernel log > when using device tree overlays. Specifically, the problem seems to > happen when removing a device tree overlay that contains two devices > with some dependency between them (e.g., a device that provides a clock > and a device that consumes that clock). Removing such an overlay results > in: > > OF: ERROR: memory leak, expected refcount 1 instead of 2, > of_node_get()/of_node_put() unbalanced - destroy > OF: ERROR: memory leak, expected refcount 1 instead of 2, > of_node_get()/of_node_put() unbalanced - destroy > > followed by hitting some REFCOUNT_WARNs in refcount.c > > In the first patch, I've included a unittest that can be used to > reproduce this when built with CONFIG_OF_UNITTEST [1]. > > I believe the issue is caused by the cleanup performed when releasing > the devlink device that's created to represent the dependency between > devices. The devlink device has references to the consumer and supplier > devices, which it drops in device_link_free; the devlink device's > release callback calls device_link_free via call_srcu. > > When the overlay is being removed, all devices are removed, and > eventually the release callback for the devlink device run, and > schedules cleanup using call_srcu. Before device_link_free can and call > put_device on the consumer/supplier, the rest of the overlay removal > process runs, resulting in the error traces above. > >>> > >>> When a devicetree node in an overlay is removed, the remove code expects > >>> all previous users of the related device to have done the appropriate put > >>> of the device and to have no later references. > >>> > >>> As Michael described above, the devlink release callback defers the > >>> put_device(). The cleanup via srcu was implemented in commit > >>> 843e600b8a2b01463c4d873a90b2c2ea8033f1f6 "driver core: Fix sleeping > >>> in invalid context during device link deletion" to solve yet another > >>> issue. > >>> > >>> > > Patches 2 and 3 are an attempt at fixing this: call srcu_barrier to wait > for any pending device_link_free's to execute before continuing on with > the removal process. > > These patches resolve the issue, but probably not in the best way. In > particular, it seems strange to need to leak details of devlinks into > the device tree overlay code. So, I'd be curious to get some feedback or > hear any other ideas for how to resolve this issue. > >>> > >>> I agree with Michael that adding an indirect call of > >>> srcu_barrier(_links_srcu) > >>> into the devicetree overlay code is not an appropriate solution. > >> > >> I kind of see your point too. I wonder if the srcu_barrier() should > >> happen inside like so: > >> device_del() -> device_links_purge()->srcu_barrier() > >> > >> I don't know what contention the use of SRCUs in device links was > >> trying to avoid, but I think the srcu_barrier() call path I suggested > >> above shouldn't be a problem. If that fixes the issue, the best way to > >> know if it's an issue is to send out a patch and see if Rafael has any > >> problem with it :) > > > > I was able to test this by adding the srcu_barrier() at the end of > > device_links_purge(), and that does seem to have fixed the issue. > > > >>> Is there some other way to fix the problem that 843e600b8a2b solves > >>> without > >>> deferring the put_device() done by the devlink release callback? > >> > >> Ok I finally got some time to look into this closely. > >> > >> Even if you revert 843e600b8a2b, you'll see that device_link_free() > >> (which drops the reference to the consumer and supplier devices) was > >> scheduled to run when the SRCU clean up occurs. So I think this issue > >> was present even before 843e600b8a2b, but commit 843e600b8a2b just > >> made it more likely to hit this scenario because it introduces some > >> delay in dropping the ref count of the supplier and consumer by going > >> through the device link device's release path. So, I think this issue > >> isn't related to 843e600b8a2b. > >> > >> As to why 843e600b8a2b had to be written to call call_srcu() from the > >> device link device's release path, it's a mess of dependencies/delays: > >> 1. The device link device is part of the struct device_link. So we > >> can't free device_link before
Re: [RFC PATCH 0/3] Fix errors on DT overlay removal with devlinks
On 10/28/20 11:25 AM, Michael Auchter wrote: > Hey Saravana, > > Thanks for taking the time to look into this! > > On Mon, Oct 26, 2020 at 12:10:33PM -0700, Saravana Kannan wrote: >> On Wed, Oct 21, 2020 at 2:02 PM Frank Rowand wrote: >>> >>> Hi Saravana, >>> >>> Michael found an issue related to the removal of a devicetree node >>> which involves devlinks: >>> >>> On 10/14/20 2:36 PM, Michael Auchter wrote: After updating to v5.9, I've started seeing errors in the kernel log when using device tree overlays. Specifically, the problem seems to happen when removing a device tree overlay that contains two devices with some dependency between them (e.g., a device that provides a clock and a device that consumes that clock). Removing such an overlay results in: OF: ERROR: memory leak, expected refcount 1 instead of 2, of_node_get()/of_node_put() unbalanced - destroy OF: ERROR: memory leak, expected refcount 1 instead of 2, of_node_get()/of_node_put() unbalanced - destroy followed by hitting some REFCOUNT_WARNs in refcount.c In the first patch, I've included a unittest that can be used to reproduce this when built with CONFIG_OF_UNITTEST [1]. I believe the issue is caused by the cleanup performed when releasing the devlink device that's created to represent the dependency between devices. The devlink device has references to the consumer and supplier devices, which it drops in device_link_free; the devlink device's release callback calls device_link_free via call_srcu. When the overlay is being removed, all devices are removed, and eventually the release callback for the devlink device run, and schedules cleanup using call_srcu. Before device_link_free can and call put_device on the consumer/supplier, the rest of the overlay removal process runs, resulting in the error traces above. >>> >>> When a devicetree node in an overlay is removed, the remove code expects >>> all previous users of the related device to have done the appropriate put >>> of the device and to have no later references. >>> >>> As Michael described above, the devlink release callback defers the >>> put_device(). The cleanup via srcu was implemented in commit >>> 843e600b8a2b01463c4d873a90b2c2ea8033f1f6 "driver core: Fix sleeping >>> in invalid context during device link deletion" to solve yet another >>> issue. >>> >>> Patches 2 and 3 are an attempt at fixing this: call srcu_barrier to wait for any pending device_link_free's to execute before continuing on with the removal process. These patches resolve the issue, but probably not in the best way. In particular, it seems strange to need to leak details of devlinks into the device tree overlay code. So, I'd be curious to get some feedback or hear any other ideas for how to resolve this issue. >>> >>> I agree with Michael that adding an indirect call of >>> srcu_barrier(_links_srcu) >>> into the devicetree overlay code is not an appropriate solution. >> >> I kind of see your point too. I wonder if the srcu_barrier() should >> happen inside like so: >> device_del() -> device_links_purge()->srcu_barrier() >> >> I don't know what contention the use of SRCUs in device links was >> trying to avoid, but I think the srcu_barrier() call path I suggested >> above shouldn't be a problem. If that fixes the issue, the best way to >> know if it's an issue is to send out a patch and see if Rafael has any >> problem with it :) > > I was able to test this by adding the srcu_barrier() at the end of > device_links_purge(), and that does seem to have fixed the issue. > >>> Is there some other way to fix the problem that 843e600b8a2b solves without >>> deferring the put_device() done by the devlink release callback? >> >> Ok I finally got some time to look into this closely. >> >> Even if you revert 843e600b8a2b, you'll see that device_link_free() >> (which drops the reference to the consumer and supplier devices) was >> scheduled to run when the SRCU clean up occurs. So I think this issue >> was present even before 843e600b8a2b, but commit 843e600b8a2b just >> made it more likely to hit this scenario because it introduces some >> delay in dropping the ref count of the supplier and consumer by going >> through the device link device's release path. So, I think this issue >> isn't related to 843e600b8a2b. >> >> As to why 843e600b8a2b had to be written to call call_srcu() from the >> device link device's release path, it's a mess of dependencies/delays: >> 1. The device link device is part of the struct device_link. So we >> can't free device_link before the device_link.link_dev refcount goes >> to 0. >> 2. But I can't assume device_link.link_dev's refcount will go to 0 as >> soon as I call put_device() on it because of >> CONFIG_DEBUG_KOBJECT_RELEASE which frees up the kobject after a random >> delay. >> 3. The use of
Re: Re: [RFC PATCH 0/3] Fix errors on DT overlay removal with devlinks
On Wed, Oct 28, 2020 at 9:25 AM Michael Auchter wrote: > > Hey Saravana, > > Thanks for taking the time to look into this! > > On Mon, Oct 26, 2020 at 12:10:33PM -0700, Saravana Kannan wrote: > > On Wed, Oct 21, 2020 at 2:02 PM Frank Rowand wrote: > > > > > > Hi Saravana, > > > > > > Michael found an issue related to the removal of a devicetree node > > > which involves devlinks: > > > > > > On 10/14/20 2:36 PM, Michael Auchter wrote: > > > > After updating to v5.9, I've started seeing errors in the kernel log > > > > when using device tree overlays. Specifically, the problem seems to > > > > happen when removing a device tree overlay that contains two devices > > > > with some dependency between them (e.g., a device that provides a clock > > > > and a device that consumes that clock). Removing such an overlay results > > > > in: > > > > > > > > OF: ERROR: memory leak, expected refcount 1 instead of 2, > > > > of_node_get()/of_node_put() unbalanced - destroy > > > > OF: ERROR: memory leak, expected refcount 1 instead of 2, > > > > of_node_get()/of_node_put() unbalanced - destroy > > > > > > > > followed by hitting some REFCOUNT_WARNs in refcount.c > > > > > > > > In the first patch, I've included a unittest that can be used to > > > > reproduce this when built with CONFIG_OF_UNITTEST [1]. > > > > > > > > I believe the issue is caused by the cleanup performed when releasing > > > > the devlink device that's created to represent the dependency between > > > > devices. The devlink device has references to the consumer and supplier > > > > devices, which it drops in device_link_free; the devlink device's > > > > release callback calls device_link_free via call_srcu. > > > > > > > > When the overlay is being removed, all devices are removed, and > > > > eventually the release callback for the devlink device run, and > > > > schedules cleanup using call_srcu. Before device_link_free can and call > > > > put_device on the consumer/supplier, the rest of the overlay removal > > > > process runs, resulting in the error traces above. > > > > > > When a devicetree node in an overlay is removed, the remove code expects > > > all previous users of the related device to have done the appropriate put > > > of the device and to have no later references. > > > > > > As Michael described above, the devlink release callback defers the > > > put_device(). The cleanup via srcu was implemented in commit > > > 843e600b8a2b01463c4d873a90b2c2ea8033f1f6 "driver core: Fix sleeping > > > in invalid context during device link deletion" to solve yet another > > > issue. > > > > > > > > > > > > > > Patches 2 and 3 are an attempt at fixing this: call srcu_barrier to wait > > > > for any pending device_link_free's to execute before continuing on with > > > > the removal process. > > > > > > > > These patches resolve the issue, but probably not in the best way. In > > > > particular, it seems strange to need to leak details of devlinks into > > > > the device tree overlay code. So, I'd be curious to get some feedback or > > > > hear any other ideas for how to resolve this issue. > > > > > > I agree with Michael that adding an indirect call of > > > srcu_barrier(_links_srcu) > > > into the devicetree overlay code is not an appropriate solution. > > > > I kind of see your point too. I wonder if the srcu_barrier() should > > happen inside like so: > > device_del() -> device_links_purge()->srcu_barrier() > > > > I don't know what contention the use of SRCUs in device links was > > trying to avoid, but I think the srcu_barrier() call path I suggested > > above shouldn't be a problem. If that fixes the issue, the best way to > > know if it's an issue is to send out a patch and see if Rafael has any > > problem with it :) > > I was able to test this by adding the srcu_barrier() at the end of > device_links_purge(), and that does seem to have fixed the issue. Thanks for testing my suggestion. If you send out a patch for that, I'd appreciated a Suggested-by: tag. > > > Is there some other way to fix the problem that 843e600b8a2b solves > > > without > > > deferring the put_device() done by the devlink release callback? > > > > Ok I finally got some time to look into this closely. > > > > Even if you revert 843e600b8a2b, you'll see that device_link_free() > > (which drops the reference to the consumer and supplier devices) was > > scheduled to run when the SRCU clean up occurs. So I think this issue > > was present even before 843e600b8a2b, but commit 843e600b8a2b just > > made it more likely to hit this scenario because it introduces some > > delay in dropping the ref count of the supplier and consumer by going > > through the device link device's release path. So, I think this issue > > isn't related to 843e600b8a2b. > > > > As to why 843e600b8a2b had to be written to call call_srcu() from the > > device link device's release path, it's a mess of dependencies/delays: > > 1. The device link device is part of the struct device_link. So
Re: Re: [RFC PATCH 0/3] Fix errors on DT overlay removal with devlinks
Hey Saravana, Thanks for taking the time to look into this! On Mon, Oct 26, 2020 at 12:10:33PM -0700, Saravana Kannan wrote: > On Wed, Oct 21, 2020 at 2:02 PM Frank Rowand wrote: > > > > Hi Saravana, > > > > Michael found an issue related to the removal of a devicetree node > > which involves devlinks: > > > > On 10/14/20 2:36 PM, Michael Auchter wrote: > > > After updating to v5.9, I've started seeing errors in the kernel log > > > when using device tree overlays. Specifically, the problem seems to > > > happen when removing a device tree overlay that contains two devices > > > with some dependency between them (e.g., a device that provides a clock > > > and a device that consumes that clock). Removing such an overlay results > > > in: > > > > > > OF: ERROR: memory leak, expected refcount 1 instead of 2, > > > of_node_get()/of_node_put() unbalanced - destroy > > > OF: ERROR: memory leak, expected refcount 1 instead of 2, > > > of_node_get()/of_node_put() unbalanced - destroy > > > > > > followed by hitting some REFCOUNT_WARNs in refcount.c > > > > > > In the first patch, I've included a unittest that can be used to > > > reproduce this when built with CONFIG_OF_UNITTEST [1]. > > > > > > I believe the issue is caused by the cleanup performed when releasing > > > the devlink device that's created to represent the dependency between > > > devices. The devlink device has references to the consumer and supplier > > > devices, which it drops in device_link_free; the devlink device's > > > release callback calls device_link_free via call_srcu. > > > > > > When the overlay is being removed, all devices are removed, and > > > eventually the release callback for the devlink device run, and > > > schedules cleanup using call_srcu. Before device_link_free can and call > > > put_device on the consumer/supplier, the rest of the overlay removal > > > process runs, resulting in the error traces above. > > > > When a devicetree node in an overlay is removed, the remove code expects > > all previous users of the related device to have done the appropriate put > > of the device and to have no later references. > > > > As Michael described above, the devlink release callback defers the > > put_device(). The cleanup via srcu was implemented in commit > > 843e600b8a2b01463c4d873a90b2c2ea8033f1f6 "driver core: Fix sleeping > > in invalid context during device link deletion" to solve yet another > > issue. > > > > > > > > > > Patches 2 and 3 are an attempt at fixing this: call srcu_barrier to wait > > > for any pending device_link_free's to execute before continuing on with > > > the removal process. > > > > > > These patches resolve the issue, but probably not in the best way. In > > > particular, it seems strange to need to leak details of devlinks into > > > the device tree overlay code. So, I'd be curious to get some feedback or > > > hear any other ideas for how to resolve this issue. > > > > I agree with Michael that adding an indirect call of > > srcu_barrier(_links_srcu) > > into the devicetree overlay code is not an appropriate solution. > > I kind of see your point too. I wonder if the srcu_barrier() should > happen inside like so: > device_del() -> device_links_purge()->srcu_barrier() > > I don't know what contention the use of SRCUs in device links was > trying to avoid, but I think the srcu_barrier() call path I suggested > above shouldn't be a problem. If that fixes the issue, the best way to > know if it's an issue is to send out a patch and see if Rafael has any > problem with it :) I was able to test this by adding the srcu_barrier() at the end of device_links_purge(), and that does seem to have fixed the issue. > > Is there some other way to fix the problem that 843e600b8a2b solves without > > deferring the put_device() done by the devlink release callback? > > Ok I finally got some time to look into this closely. > > Even if you revert 843e600b8a2b, you'll see that device_link_free() > (which drops the reference to the consumer and supplier devices) was > scheduled to run when the SRCU clean up occurs. So I think this issue > was present even before 843e600b8a2b, but commit 843e600b8a2b just > made it more likely to hit this scenario because it introduces some > delay in dropping the ref count of the supplier and consumer by going > through the device link device's release path. So, I think this issue > isn't related to 843e600b8a2b. > > As to why 843e600b8a2b had to be written to call call_srcu() from the > device link device's release path, it's a mess of dependencies/delays: > 1. The device link device is part of the struct device_link. So we > can't free device_link before the device_link.link_dev refcount goes > to 0. > 2. But I can't assume device_link.link_dev's refcount will go to 0 as > soon as I call put_device() on it because of > CONFIG_DEBUG_KOBJECT_RELEASE which frees up the kobject after a random > delay. > 3. The use of SRCU also means I can't free device_link until the SRCU
Re: [RFC PATCH 0/3] Fix errors on DT overlay removal with devlinks
On Wed, Oct 21, 2020 at 2:02 PM Frank Rowand wrote: > > Hi Saravana, > > Michael found an issue related to the removal of a devicetree node > which involves devlinks: > > On 10/14/20 2:36 PM, Michael Auchter wrote: > > After updating to v5.9, I've started seeing errors in the kernel log > > when using device tree overlays. Specifically, the problem seems to > > happen when removing a device tree overlay that contains two devices > > with some dependency between them (e.g., a device that provides a clock > > and a device that consumes that clock). Removing such an overlay results > > in: > > > > OF: ERROR: memory leak, expected refcount 1 instead of 2, > > of_node_get()/of_node_put() unbalanced - destroy > > OF: ERROR: memory leak, expected refcount 1 instead of 2, > > of_node_get()/of_node_put() unbalanced - destroy > > > > followed by hitting some REFCOUNT_WARNs in refcount.c > > > > In the first patch, I've included a unittest that can be used to > > reproduce this when built with CONFIG_OF_UNITTEST [1]. > > > > I believe the issue is caused by the cleanup performed when releasing > > the devlink device that's created to represent the dependency between > > devices. The devlink device has references to the consumer and supplier > > devices, which it drops in device_link_free; the devlink device's > > release callback calls device_link_free via call_srcu. > > > > When the overlay is being removed, all devices are removed, and > > eventually the release callback for the devlink device run, and > > schedules cleanup using call_srcu. Before device_link_free can and call > > put_device on the consumer/supplier, the rest of the overlay removal > > process runs, resulting in the error traces above. > > When a devicetree node in an overlay is removed, the remove code expects > all previous users of the related device to have done the appropriate put > of the device and to have no later references. > > As Michael described above, the devlink release callback defers the > put_device(). The cleanup via srcu was implemented in commit > 843e600b8a2b01463c4d873a90b2c2ea8033f1f6 "driver core: Fix sleeping > in invalid context during device link deletion" to solve yet another > issue. > > > > > > Patches 2 and 3 are an attempt at fixing this: call srcu_barrier to wait > > for any pending device_link_free's to execute before continuing on with > > the removal process. > > > > These patches resolve the issue, but probably not in the best way. In > > particular, it seems strange to need to leak details of devlinks into > > the device tree overlay code. So, I'd be curious to get some feedback or > > hear any other ideas for how to resolve this issue. > > I agree with Michael that adding an indirect call of > srcu_barrier(_links_srcu) > into the devicetree overlay code is not an appropriate solution. I kind of see your point too. I wonder if the srcu_barrier() should happen inside like so: device_del() -> device_links_purge()->srcu_barrier() I don't know what contention the use of SRCUs in device links was trying to avoid, but I think the srcu_barrier() call path I suggested above shouldn't be a problem. If that fixes the issue, the best way to know if it's an issue is to send out a patch and see if Rafael has any problem with it :) > Is there some other way to fix the problem that 843e600b8a2b solves without > deferring the put_device() done by the devlink release callback? Ok I finally got some time to look into this closely. Even if you revert 843e600b8a2b, you'll see that device_link_free() (which drops the reference to the consumer and supplier devices) was scheduled to run when the SRCU clean up occurs. So I think this issue was present even before 843e600b8a2b, but commit 843e600b8a2b just made it more likely to hit this scenario because it introduces some delay in dropping the ref count of the supplier and consumer by going through the device link device's release path. So, I think this issue isn't related to 843e600b8a2b. As to why 843e600b8a2b had to be written to call call_srcu() from the device link device's release path, it's a mess of dependencies/delays: 1. The device link device is part of the struct device_link. So we can't free device_link before the device_link.link_dev refcount goes to 0. 2. But I can't assume device_link.link_dev's refcount will go to 0 as soon as I call put_device() on it because of CONFIG_DEBUG_KOBJECT_RELEASE which frees up the kobject after a random delay. 3. The use of SRCU also means I can't free device_link until the SRCU is cleaned up. Because of (1), (2) and (3), when the device_link_del() (or any of the other device link deletion APIs are called) I first have to do a put_device(device_link.link_dev) to make sure the device memory is no longer referenced, then trigger an SRCU clean up and then in the scheduled SRCU cleanup I can free struct device_link. And obviously, until struct device_link is ready to be freed up, I can't drop the reference to the supplier
Re: [RFC PATCH 0/3] Fix errors on DT overlay removal with devlinks
Hi Saravana, Michael found an issue related to the removal of a devicetree node which involves devlinks: On 10/14/20 2:36 PM, Michael Auchter wrote: > After updating to v5.9, I've started seeing errors in the kernel log > when using device tree overlays. Specifically, the problem seems to > happen when removing a device tree overlay that contains two devices > with some dependency between them (e.g., a device that provides a clock > and a device that consumes that clock). Removing such an overlay results > in: > > OF: ERROR: memory leak, expected refcount 1 instead of 2, > of_node_get()/of_node_put() unbalanced - destroy > OF: ERROR: memory leak, expected refcount 1 instead of 2, > of_node_get()/of_node_put() unbalanced - destroy > > followed by hitting some REFCOUNT_WARNs in refcount.c > > In the first patch, I've included a unittest that can be used to > reproduce this when built with CONFIG_OF_UNITTEST [1]. > > I believe the issue is caused by the cleanup performed when releasing > the devlink device that's created to represent the dependency between > devices. The devlink device has references to the consumer and supplier > devices, which it drops in device_link_free; the devlink device's > release callback calls device_link_free via call_srcu. > > When the overlay is being removed, all devices are removed, and > eventually the release callback for the devlink device run, and > schedules cleanup using call_srcu. Before device_link_free can and call > put_device on the consumer/supplier, the rest of the overlay removal > process runs, resulting in the error traces above. When a devicetree node in an overlay is removed, the remove code expects all previous users of the related device to have done the appropriate put of the device and to have no later references. As Michael described above, the devlink release callback defers the put_device(). The cleanup via srcu was implemented in commit 843e600b8a2b01463c4d873a90b2c2ea8033f1f6 "driver core: Fix sleeping in invalid context during device link deletion" to solve yet another issue. > > Patches 2 and 3 are an attempt at fixing this: call srcu_barrier to wait > for any pending device_link_free's to execute before continuing on with > the removal process. > > These patches resolve the issue, but probably not in the best way. In > particular, it seems strange to need to leak details of devlinks into > the device tree overlay code. So, I'd be curious to get some feedback or > hear any other ideas for how to resolve this issue. I agree with Michael that adding an indirect call of srcu_barrier(_links_srcu) into the devicetree overlay code is not an appropriate solution. Is there some other way to fix the problem that 843e600b8a2b solves without deferring the put_device() done by the devlink release callback? -Frank > > Thanks, > Michael > > 1. Note that this isn't a very good unit test: it will report a "pass" >even if it fails with the aforementioned errors, as these errors >aren't propogated. > > Michael Auchter (3): > of: unittest: add test of overlay with devlinks > driver core: add device_links_barrier > of: dynamic: add device links barrier before detach > > drivers/base/core.c | 10 ++ > drivers/of/dynamic.c| 3 +++ > drivers/of/unittest-data/Makefile | 1 + > drivers/of/unittest-data/overlay_16.dts | 26 + > drivers/of/unittest.c | 16 +++ > include/linux/device.h | 1 + > 6 files changed, 57 insertions(+) > create mode 100644 drivers/of/unittest-data/overlay_16.dts >
Re: [RFC PATCH 0/3] Fix errors on DT overlay removal with devlinks
Hi Michael, On 10/14/20 2:36 PM, Michael Auchter wrote: > After updating to v5.9, I've started seeing errors in the kernel log > when using device tree overlays. Specifically, the problem seems to > happen when removing a device tree overlay that contains two devices > with some dependency between them (e.g., a device that provides a clock > and a device that consumes that clock). Removing such an overlay results > in: > > OF: ERROR: memory leak, expected refcount 1 instead of 2, > of_node_get()/of_node_put() unbalanced - destroy > OF: ERROR: memory leak, expected refcount 1 instead of 2, > of_node_get()/of_node_put() unbalanced - destroy > > followed by hitting some REFCOUNT_WARNs in refcount.c > > In the first patch, I've included a unittest that can be used to > reproduce this when built with CONFIG_OF_UNITTEST [1]. > > I believe the issue is caused by the cleanup performed when releasing > the devlink device that's created to represent the dependency between > devices. The devlink device has references to the consumer and supplier > devices, which it drops in device_link_free; the devlink device's > release callback calls device_link_free via call_srcu. > > When the overlay is being removed, all devices are removed, and > eventually the release callback for the devlink device run, and > schedules cleanup using call_srcu. Before device_link_free can and call > put_device on the consumer/supplier, the rest of the overlay removal > process runs, resulting in the error traces above. > > Patches 2 and 3 are an attempt at fixing this: call srcu_barrier to wait > for any pending device_link_free's to execute before continuing on with > the removal process. > > These patches resolve the issue, but probably not in the best way. In > particular, it seems strange to need to leak details of devlinks into > the device tree overlay code. So, I'd be curious to get some feedback or > hear any other ideas for how to resolve this issue. Thanks for finding the problem, analyzing it, creating a unittest, and creating a fix. I agree with your analysis that there are issues with the implementation of the test and fix. I'll dig into this to see if I can provide some useful improvements. -Frank > > Thanks, > Michael > > 1. Note that this isn't a very good unit test: it will report a "pass" >even if it fails with the aforementioned errors, as these errors >aren't propogated. > > Michael Auchter (3): > of: unittest: add test of overlay with devlinks > driver core: add device_links_barrier > of: dynamic: add device links barrier before detach > > drivers/base/core.c | 10 ++ > drivers/of/dynamic.c| 3 +++ > drivers/of/unittest-data/Makefile | 1 + > drivers/of/unittest-data/overlay_16.dts | 26 + > drivers/of/unittest.c | 16 +++ > include/linux/device.h | 1 + > 6 files changed, 57 insertions(+) > create mode 100644 drivers/of/unittest-data/overlay_16.dts >
