Re: [PATCH v2 1/2] mm/vmalloc: Add a safer version of find_vm_area() for debug

[Joel Fernandes]

> On Sep 1, 2023, at 8:48 AM, Uladzislau Rezki <urezki@xxxxxxxxx> wrote:
> 
> On Fri, Sep 01, 2023 at 12:33:21AM +0000, Joel Fernandes wrote:
>>> On Fri, Sep 01, 2023 at 12:19:17AM +0000, Joel Fernandes wrote:
>>> On Thu, Aug 31, 2023 at 09:47:52PM +0200, Uladzislau Rezki wrote:
>>>> On Thu, Aug 31, 2023 at 05:18:25PM +0000, Joel Fernandes (Google) wrote:
>>>>> It is unsafe to dump vmalloc area information when trying to do so from
>>>>> some contexts. Add a safer trylock version of the same function to do a
>>>>> best-effort VMA finding and use it from vmalloc_dump_obj().
>>>>> 
>>>>> [apply test robot feedback on unused function fix.]
>>>>> 
>>>>> Reported-by: Zhen Lei <thunder.leizhen@xxxxxxxxxxxxxxx>
>>>>> Cc: Paul E. McKenney <paulmck@xxxxxxxxxx>
>>>>> Cc: rcu@xxxxxxxxxxxxxxx
>>>>> Cc: Zqiang <qiang.zhang1211@xxxxxxxxx>
>>>>> Reviewed-by: Matthew Wilcox (Oracle) <willy@xxxxxxxxxxxxx>
>>>>> Signed-off-by: Joel Fernandes (Google) <joel@xxxxxxxxxxxxxxxxx>
>>>>> ---
>>>>> v1->v2: Apply review tags and test robot feedback.
>>>>> 
>>>>> mm/vmalloc.c | 39 ++++++++++++++++++++++++++++++++++++++-
>>>>> 1 file changed, 38 insertions(+), 1 deletion(-)
>>>>> 
>>>>> diff --git a/mm/vmalloc.c b/mm/vmalloc.c
>>>>> index 93cf99aba335..f09e882ae3b8 100644
>>>>> --- a/mm/vmalloc.c
>>>>> +++ b/mm/vmalloc.c
>>>>> @@ -1865,6 +1865,20 @@ struct vmap_area *find_vmap_area(unsigned long addr)
>>>>>    return va;
>>>>> }
>>>>> 
>>>>> +#ifdef CONFIG_PRINTK
>>>>> +static struct vmap_area *find_vmap_area_trylock(unsigned long addr)
>>>>> +{
>>>>> +    struct vmap_area *va;
>>>>> +
>>>>> +    if (!spin_trylock(&vmap_area_lock))
>>>>> +        return NULL;
>>>>> +    va = __find_vmap_area(addr, &vmap_area_root);
>>>>> +    spin_unlock(&vmap_area_lock);
>>>>> +
>>>>> +    return va;
>>>>> +}
>>>>> +#endif
>>>>> +
>>>>> static struct vmap_area *find_unlink_vmap_area(unsigned long addr)
>>>>> {
>>>>>    struct vmap_area *va;
>>>>> @@ -2671,6 +2685,29 @@ struct vm_struct *find_vm_area(const void *addr)
>>>>>    return va->vm;
>>>>> }
>>>>> 
>>>>> +/**
>>>>> + * try_to_find_vm_area - find a continuous kernel virtual area
>>>>> + * @addr:      base address
>>>>> + *
>>>>> + * This function is the same as find_vm_area() except that it is
>>>>> + * safe to call if vmap_area_lock is already held and returns NULL
>>>>> + * if it is. See comments in find_vmap_area() for other details.
>>>>> + *
>>>>> + * Return: the area descriptor on success or %NULL on failure.
>>>>> + */
>>>>> +#ifdef CONFIG_PRINTK
>>>>> +static struct vm_struct *try_to_find_vm_area(const void *addr)
>>>>> +{
>>>>> +    struct vmap_area *va;
>>>>> +
>>>>> +    va = find_vmap_area_trylock((unsigned long)addr);
>>>>> +    if (!va)
>>>>> +        return NULL;
>>>>> +
>>>>> +    return va->vm;
>>>>> +}
>>>>> +#endif
>>>>> +
>>>>> /**
>>>>>  * remove_vm_area - find and remove a continuous kernel virtual area
>>>>>  * @addr:        base address
>>>>> @@ -4277,7 +4314,7 @@ bool vmalloc_dump_obj(void *object)
>>>>>    struct vm_struct *vm;
>>>>>    void *objp = (void *)PAGE_ALIGN((unsigned long)object);
>>>>> 
>>>>> -    vm = find_vm_area(objp);
>>>>> +    vm = try_to_find_vm_area(objp);
>>>>>    if (!vm)
>>>>>        return false;
>>>>>    pr_cont(" %u-page vmalloc region starting at %#lx allocated at %pS\n",
>>> 
>>> Hi Vlad,
>>> Thanks for taking a look.
>>> 
>>>> I am not sure if this patch makes a lot of sense. I agree, this is a
>>>> problem and it mitigates it. But it is broken in terms of once you drop
>>>> the lock, the VA should not be accessed.
>>> 
>>> Just to note the lockless-access issue you are referring to is not introduced
>>> by this patch but is rather in the existing code. Also just to note this is
>>> debug code.
>>> 
>>>> Is that a real issue or it gets triggered due to some syntetic test case?
>>> 
>>> It is a real issue. See 2/2.
>>> 
>>>> If i were you, i would go with open-coded version of trylock. Because
>>>> there is only one user so far.
>>> 
>>> Taking your open coding and locking suggestions, I came up with the below
>>> which actually results in a smaller patch. Does it look good to you?
>>> 
>>> diff --git a/mm/vmalloc.c b/mm/vmalloc.c
>>> index 93cf99aba335..aaf6bad997a7 100644
>> 
>> And with some trivial compiler errors fixed (sorry should have build tested
>> but wanted to just share the idea earlier):
>> 
>> diff --git a/mm/vmalloc.c b/mm/vmalloc.c
>> index 93cf99aba335..2c6a0e2ff404 100644
>> --- a/mm/vmalloc.c
>> +++ b/mm/vmalloc.c
>> @@ -4274,14 +4274,32 @@ void pcpu_free_vm_areas(struct vm_struct **vms, int nr_vms)
>> #ifdef CONFIG_PRINTK
>> bool vmalloc_dump_obj(void *object)
>> {
>> -    struct vm_struct *vm;
>>    void *objp = (void *)PAGE_ALIGN((unsigned long)object);
>> +    const void *caller;
>> +    struct vm_struct *vm;
>> +    struct vmap_area *va;
>> +    unsigned long addr;
>> +    unsigned int nr_pages;
>> 
>> -    vm = find_vm_area(objp);
>> -    if (!vm)
>> +    if (!spin_trylock(&vmap_area_lock))
>> +        return false;
>> +    va = __find_vmap_area((unsigned long)objp, &vmap_area_root);
>> +    if (!va) {
>> +        spin_unlock(&vmap_area_lock);
>>        return false;
>> +    }
>> +
>> +    vm = va->vm;
>> +    if (!vm) {
>> +        spin_unlock(&vmap_area_lock);
>> +        return false;
>> +    }
>> +    addr = (unsigned long)vm->addr;
>> +    caller = vm->caller;
>> +    nr_pages = vm->nr_pages;
>> +    spin_unlock(&vmap_area_lock);
>>    pr_cont(" %u-page vmalloc region starting at %#lx allocated at %pS\n",
>> -        vm->nr_pages, (unsigned long)vm->addr, vm->caller);
>> +        nr_pages, addr, caller);
>>    return true;
>> }
>> #endif
>> 
> Looks good to me and thank you for fixing a locking issue :)
> I think you will re-spin and resend it one more time?

Yes. May I add your Reviewed-by tag to both patches after re-spinning as mentioned above?

thanks!

 - Joel

> 
> --
> Uladzislau Rezki