The function mem_dump_obj() can sometimes provide valuable debugging
information, but it cannot be called in an interrupt context because
spinlock vmap_area_lock has not been protected against IRQs. If the
current task has held the lock before hard/soft interrupt handler calls
mem_dump_obj(), simply abandoning the dump operation can avoid deadlock.
That is, no deadlock occurs in extreme cases, and dump succeeds in most
cases.

Signed-off-by: Zhen Lei <thunder.leizhen@huawei.com>
---
 mm/vmalloc.c | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/mm/vmalloc.c b/mm/vmalloc.c
index ccaa461998f3c37..cdd36c5a1aa16f8 100644
--- a/mm/vmalloc.c
+++ b/mm/vmalloc.c
@@ -4034,6 +4034,9 @@ bool vmalloc_dump_obj(void *object)
 	struct vm_struct *vm;
 	void *objp = (void *)PAGE_ALIGN((unsigned long)object);
 
+	if (unlikely(spin_is_locked(&vmap_area_lock)))
+		return false;
+
 	vm = find_vm_area(objp);
 	if (!vm)
 		return false;
-- 
2.25.1

On Sat, 12 Nov 2022 20:15:37 +0800 Zhen Lei <thunder.leizhen@huawei.com> wrote:

> The function mem_dump_obj() can sometimes provide valuable debugging
> information, but it cannot be called in an interrupt context because
> spinlock vmap_area_lock has not been protected against IRQs. If the
> current task has held the lock before hard/soft interrupt handler calls
> mem_dump_obj(), simply abandoning the dump operation can avoid deadlock.
> That is, no deadlock occurs in extreme cases, and dump succeeds in most
> cases.
> 
> ...
>
> --- a/mm/vmalloc.c
> +++ b/mm/vmalloc.c
> @@ -4034,6 +4034,9 @@ bool vmalloc_dump_obj(void *object)
>  	struct vm_struct *vm;
>  	void *objp = (void *)PAGE_ALIGN((unsigned long)object);
>  
> +	if (unlikely(spin_is_locked(&vmap_area_lock)))
> +		return false;
> +
>  	vm = find_vm_area(objp);
>  	if (!vm)
>  		return false;

Yes, but this will worsen the current uses of this function.  Consider
the case where task A wants to call vmalloc_dump_obj() but task B holds
vmap_area_lock.  No problem, task A will simply spin until task B is
done.

But after this patch, task A's call to vmalloc_dump_obj() will return
without having done anything.

On 2022/11/15 8:54, Andrew Morton wrote:
> On Sat, 12 Nov 2022 20:15:37 +0800 Zhen Lei <thunder.leizhen@huawei.com> wrote:
> 
>> The function mem_dump_obj() can sometimes provide valuable debugging
>> information, but it cannot be called in an interrupt context because
>> spinlock vmap_area_lock has not been protected against IRQs. If the
>> current task has held the lock before hard/soft interrupt handler calls
>> mem_dump_obj(), simply abandoning the dump operation can avoid deadlock.
>> That is, no deadlock occurs in extreme cases, and dump succeeds in most
>> cases.
>>
>> ...
>>
>> --- a/mm/vmalloc.c
>> +++ b/mm/vmalloc.c
>> @@ -4034,6 +4034,9 @@ bool vmalloc_dump_obj(void *object)
>>  	struct vm_struct *vm;
>>  	void *objp = (void *)PAGE_ALIGN((unsigned long)object);
>>  
>> +	if (unlikely(spin_is_locked(&vmap_area_lock)))
>> +		return false;
>> +
>>  	vm = find_vm_area(objp);
>>  	if (!vm)
>>  		return false;
> 
> Yes, but this will worsen the current uses of this function.  Consider
> the case where task A wants to call vmalloc_dump_obj() but task B holds
> vmap_area_lock.  No problem, task A will simply spin until task B is
> done.
> 
> But after this patch, task A's call to vmalloc_dump_obj() will return
> without having done anything.

Oh, right, this problem occurs when task A and task B run on
two different cores.


> 
> .
> 

-- 
Regards,
  Zhen Lei

On 2022/11/15 10:01, Leizhen (ThunderTown) wrote:
> 
> 
> On 2022/11/15 8:54, Andrew Morton wrote:
>> On Sat, 12 Nov 2022 20:15:37 +0800 Zhen Lei <thunder.leizhen@huawei.com> wrote:
>>
>>> The function mem_dump_obj() can sometimes provide valuable debugging
>>> information, but it cannot be called in an interrupt context because
>>> spinlock vmap_area_lock has not been protected against IRQs. If the
>>> current task has held the lock before hard/soft interrupt handler calls
>>> mem_dump_obj(), simply abandoning the dump operation can avoid deadlock.
>>> That is, no deadlock occurs in extreme cases, and dump succeeds in most
>>> cases.
>>>
>>> ...
>>>
>>> --- a/mm/vmalloc.c
>>> +++ b/mm/vmalloc.c
>>> @@ -4034,6 +4034,9 @@ bool vmalloc_dump_obj(void *object)
>>>  	struct vm_struct *vm;
>>>  	void *objp = (void *)PAGE_ALIGN((unsigned long)object);
>>>  
>>> +	if (unlikely(spin_is_locked(&vmap_area_lock)))
>>> +		return false;
>>> +
>>>  	vm = find_vm_area(objp);
>>>  	if (!vm)
>>>  		return false;
>>
>> Yes, but this will worsen the current uses of this function.  Consider
>> the case where task A wants to call vmalloc_dump_obj() but task B holds
>> vmap_area_lock.  No problem, task A will simply spin until task B is
>> done.
>>
>> But after this patch, task A's call to vmalloc_dump_obj() will return
>> without having done anything.
> 
> Oh, right, this problem occurs when task A and task B run on
> two different cores.

I've rethought it, this can be solved by adding in_interrupt().

> 
> 
>>
>> .
>>
> 

-- 
Regards,
  Zhen Lei