Immediately after the platform_device_unregister() the device will be cleaned up.
Accessing the freed pointer immediately after that will crash the system.
Found this bug when kernel is built with CONFIG_PAGE_POISONING and testing
loading/unloading audio drivers in a loop on Qcom platforms.
Fix this by removing accessing the dev pointer.
Below is the carsh trace:
Unable to handle kernel paging request at virtual address 6b6b6b6b6b6c03
Mem abort info:
ESR = 0x96000021
Exception class = DABT (current EL), IL = 32 bits
SET = 0, FnV = 0
EA = 0, S1PTW = 0
Data abort info:
ISV = 0, ISS = 0x00000021
CM = 0, WnR = 0
[006b6b6b6b6b6c03] address between user and kernel address ranges
Internal error: Oops: 96000021 [#1] PREEMPT SMP
Modules linked in:
CPU: 2 PID: 1784 Comm: sh Tainted: G W 4.17.0-rc7-02230-ge3a63a7ef641-dirty #204
Hardware name: Qualcomm Technologies, Inc. APQ 8016 SBC (DT)
pstate: 80000005 (Nzcv daif -PAN -UAO)
pc : clear_bit+0x18/0x2c
lr : of_platform_device_destroy+0x64/0xb8
sp : ffff00000c9c3930
x29: ffff00000c9c3930 x28: ffff80003d39b200
x27: ffff000008bb1000 x26: 0000000000000040
x25: 0000000000000124 x24: ffff80003a9a3080
x23: 0000000000000060 x22: ffff00000939f518
x21: ffff80003aa79e98 x20: ffff80003aa3dae0
x19: ffff80003aa3c890 x18: ffff800009feb794
x17: 0000000000000000 x16: 0000000000000000
x15: ffff800009feb790 x14: 0000000000000000
x13: ffff80003a058778 x12: ffff80003a058728
x11: ffff80003a058750 x10: 0000000000000000
x9 : 0000000000000006 x8 : ffff80003a825988
x7 : bbbbbbbbbbbbbbbb x6 : 0000000000000001
x5 : 0000000000000000 x4 : 0000000000000001
x3 : 0000000000000008 x2 : 0000000000000001
x1 : 6b6b6b6b6b6b6c03 x0 : 0000000000000000
Process sh (pid: 1784, stack limit = 0x (ptrval))
Call trace:
clear_bit+0x18/0x2c
q6afe_remove+0x20/0x38
apr_device_remove+0x30/0x70
device_release_driver_internal+0x170/0x208
device_release_driver+0x14/0x20
bus_remove_device+0xcc/0x150
device_del+0x10c/0x310
device_unregister+0x1c/0x70
apr_remove_device+0xc/0x18
device_for_each_child+0x50/0x80
apr_remove+0x18/0x20
rpmsg_dev_remove+0x38/0x68
device_release_driver_internal+0x170/0x208
device_release_driver+0x14/0x20
bus_remove_device+0xcc/0x150
device_del+0x10c/0x310
device_unregister+0x1c/0x70
qcom_smd_remove_device+0xc/0x18
device_for_each_child+0x50/0x80
qcom_smd_unregister_edge+0x3c/0x70
smd_subdev_remove+0x18/0x28
rproc_stop+0x48/0xd8
rproc_shutdown+0x60/0xe8
state_store+0xbc/0xf8
dev_attr_store+0x18/0x28
sysfs_kf_write+0x3c/0x50
kernfs_fop_write+0x118/0x1e0
__vfs_write+0x18/0x110
vfs_write+0xa4/0x1a8
ksys_write+0x48/0xb0
sys_write+0xc/0x18
el0_svc_naked+0x30/0x34
Code: d2800022 8b400c21 f9800031 9ac32043 (c85f7c22)
Signed-off-by: Srinivas Kandagatla <[email protected]>
---
Changes since v1:
- fixed issue while reprobing.
drivers/of/platform.c | 7 +++++--
1 file changed, 5 insertions(+), 2 deletions(-)
diff --git a/drivers/of/platform.c b/drivers/of/platform.c
index c00d81dfac0b..84c5c899187b 100644
--- a/drivers/of/platform.c
+++ b/drivers/of/platform.c
@@ -529,10 +529,13 @@ arch_initcall_sync(of_platform_default_populate_init);
int of_platform_device_destroy(struct device *dev, void *data)
{
+ struct device_node *np;
+
/* Do not touch devices not populated from the device tree */
if (!dev->of_node || !of_node_check_flag(dev->of_node, OF_POPULATED))
return 0;
+ np = dev->of_node;
/* Recurse for any nodes that were treated as busses */
if (of_node_check_flag(dev->of_node, OF_POPULATED_BUS))
device_for_each_child(dev, NULL, of_platform_device_destroy);
@@ -544,8 +547,8 @@ int of_platform_device_destroy(struct device *dev, void *data)
amba_device_unregister(to_amba_device(dev));
#endif
- of_node_clear_flag(dev->of_node, OF_POPULATED);
- of_node_clear_flag(dev->of_node, OF_POPULATED_BUS);
+ of_node_clear_flag(np, OF_POPULATED);
+ of_node_clear_flag(np, OF_POPULATED_BUS);
return 0;
}
EXPORT_SYMBOL_GPL(of_platform_device_destroy);
--
2.16.2
On Fri, Jun 1, 2018 at 7:03 PM, Srinivas Kandagatla
<[email protected]> wrote:
> Immediately after the platform_device_unregister() the device will be cleaned up.
> Accessing the freed pointer immediately after that will crash the system.
>
> Found this bug when kernel is built with CONFIG_PAGE_POISONING and testing
> loading/unloading audio drivers in a loop on Qcom platforms.
Curious, does the unittest not catch this too?
>
> Fix this by removing accessing the dev pointer.
> Below is the carsh trace:
s/carsh/crash/
[...]
> diff --git a/drivers/of/platform.c b/drivers/of/platform.c
> index c00d81dfac0b..84c5c899187b 100644
> --- a/drivers/of/platform.c
> +++ b/drivers/of/platform.c
> @@ -529,10 +529,13 @@ arch_initcall_sync(of_platform_default_populate_init);
>
> int of_platform_device_destroy(struct device *dev, void *data)
> {
> + struct device_node *np;
> +
> /* Do not touch devices not populated from the device tree */
> if (!dev->of_node || !of_node_check_flag(dev->of_node, OF_POPULATED))
> return 0;
>
> + np = dev->of_node;
> /* Recurse for any nodes that were treated as busses */
> if (of_node_check_flag(dev->of_node, OF_POPULATED_BUS))
> device_for_each_child(dev, NULL, of_platform_device_destroy);
> @@ -544,8 +547,8 @@ int of_platform_device_destroy(struct device *dev, void *data)
> amba_device_unregister(to_amba_device(dev));
> #endif
>
> - of_node_clear_flag(dev->of_node, OF_POPULATED);
> - of_node_clear_flag(dev->of_node, OF_POPULATED_BUS);
Just move these 2 lines to before unregister calls.
> + of_node_clear_flag(np, OF_POPULATED);
> + of_node_clear_flag(np, OF_POPULATED_BUS);
> return 0;
> }
> EXPORT_SYMBOL_GPL(of_platform_device_destroy);
> --
> 2.16.2
>
On 04/06/18 14:44, Rob Herring wrote:
> On Fri, Jun 1, 2018 at 7:03 PM, Srinivas Kandagatla
> <[email protected]> wrote:
>> Immediately after the platform_device_unregister() the device will be cleaned up.
>> Accessing the freed pointer immediately after that will crash the system.
>>
>> Found this bug when kernel is built with CONFIG_PAGE_POISONING and testing
>> loading/unloading audio drivers in a loop on Qcom platforms.
>
> Curious, does the unittest not catch this too?
>
Not sure!
>>
>> Fix this by removing accessing the dev pointer.
>> Below is the carsh trace:
>
> s/carsh/crash/
Yep.
>
> [...]
>
>> diff --git a/drivers/of/platform.c b/drivers/of/platform.c
>> index c00d81dfac0b..84c5c899187b 100644
>> --- a/drivers/of/platform.c
>> +++ b/drivers/of/platform.c
>> @@ -529,10 +529,13 @@ arch_initcall_sync(of_platform_default_populate_init);
>>
>> int of_platform_device_destroy(struct device *dev, void *data)
>> {
>> + struct device_node *np;
>> +
>> /* Do not touch devices not populated from the device tree */
>> if (!dev->of_node || !of_node_check_flag(dev->of_node, OF_POPULATED))
>> return 0;
>>
>> + np = dev->of_node;
>> /* Recurse for any nodes that were treated as busses */
>> if (of_node_check_flag(dev->of_node, OF_POPULATED_BUS))
>> device_for_each_child(dev, NULL, of_platform_device_destroy);
>> @@ -544,8 +547,8 @@ int of_platform_device_destroy(struct device *dev, void *data)
>> amba_device_unregister(to_amba_device(dev));
>> #endif
>>
>> - of_node_clear_flag(dev->of_node, OF_POPULATED);
>> - of_node_clear_flag(dev->of_node, OF_POPULATED_BUS);
>
> Just move these 2 lines to before unregister calls.
Make sense.. I will do that in v3.
thanks,
srini
>
>> + of_node_clear_flag(np, OF_POPULATED);
>> + of_node_clear_flag(np, OF_POPULATED_BUS);
>> return 0;
>> }
>> EXPORT_SYMBOL_GPL(of_platform_device_destroy);
>> --
>> 2.16.2
>>