[PATCH] mm/memory_hotplug: fix kfree() of bootmem memory

[PATCH] mm/memory_hotplug: fix kfree() of bootmem memory

[Miaohe Lin]

We can't use kfree() to release the resource as it might come from bootmem.
Use release_mem_region() instead.

Fixes: ebff7d8f270d ("mem hotunplug: fix kfree() of bootmem memory")
Signed-off-by: Miaohe Lin <linmiaohe@huawei.com>
---
 mm/memory_hotplug.c | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

diff --git a/mm/memory_hotplug.c b/mm/memory_hotplug.c
index 7dc7e12302db..dc570772b4b1 100644
--- a/mm/memory_hotplug.c
+++ b/mm/memory_hotplug.c
@@ -216,8 +216,7 @@ static void release_memory_resource(struct resource *res)
 {
 	if (!res)
 		return;
-	release_resource(res);
-	kfree(res);
+	release_mem_region(res->start, resource_size(res));
 }
 
 static int check_pfn_span(unsigned long pfn, unsigned long nr_pages,
-- 
2.23.0

Re: [PATCH] mm/memory_hotplug: fix kfree() of bootmem memory

[David Hildenbrand]

On 07.02.22 14:56, Miaohe Lin wrote:
> We can't use kfree() to release the resource as it might come from bootmem.
> Use release_mem_region() instead.

How can this happen? release_mem_region() is called either from
__add_memory() or from add_memory_driver_managed(), where we allocated
the region via register_memory_resource(). Both functions shouldn't ever
be called before the buddy is up an running.

Do you have a backtrace of an actual instance of this issue? Or was this
identified as possibly broken by code inspection?

-- 
Thanks,

David / dhildenb

Re: [PATCH] mm/memory_hotplug: fix kfree() of bootmem memory

[Miaohe Lin]

Hi:
On 2022/2/7 22:33, David Hildenbrand wrote:
> On 07.02.22 14:56, Miaohe Lin wrote:
>> We can't use kfree() to release the resource as it might come from bootmem.
>> Use release_mem_region() instead.
> 
> How can this happen? release_mem_region() is called either from
> __add_memory() or from add_memory_driver_managed(), where we allocated
> the region via register_memory_resource(). Both functions shouldn't ever
> be called before the buddy is up an running.
> 
> Do you have a backtrace of an actual instance of this issue? Or was this
> identified as possibly broken by code inspection?
> 

This is identified as possibly broken by code inspection. IIUC, alloc_resource
is always used to allocate the resource. It has the below logic:

  if (bootmem_resource_free) {
	res = bootmem_resource_free;
	bootmem_resource_free = res->sibling;
  }

where bootmem_resource_free is used to reusing the resource entries allocated by boot
mem after the system is up:

/*
 * For memory hotplug, there is no way to free resource entries allocated
 * by boot mem after the system is up. So for reusing the resource entry
 * we need to remember the resource.
 */
static struct resource *bootmem_resource_free;

So I think register_memory_resource() can reuse the resource allocated by bootmem.
Or am I miss anything?

Thanks.

Re: [PATCH] mm/memory_hotplug: fix kfree() of bootmem memory

[David Hildenbrand]

On 08.02.22 02:59, Miaohe Lin wrote:
> Hi:
> On 2022/2/7 22:33, David Hildenbrand wrote:
>> On 07.02.22 14:56, Miaohe Lin wrote:
>>> We can't use kfree() to release the resource as it might come from bootmem.
>>> Use release_mem_region() instead.
>>
>> How can this happen? release_mem_region() is called either from
>> __add_memory() or from add_memory_driver_managed(), where we allocated
>> the region via register_memory_resource(). Both functions shouldn't ever
>> be called before the buddy is up an running.
>>
>> Do you have a backtrace of an actual instance of this issue? Or was this
>> identified as possibly broken by code inspection?
>>
> 
> This is identified as possibly broken by code inspection. IIUC, alloc_resource
> is always used to allocate the resource. It has the below logic:
> 
>   if (bootmem_resource_free) {
> 	res = bootmem_resource_free;
> 	bootmem_resource_free = res->sibling;
>   }
> 
> where bootmem_resource_free is used to reusing the resource entries allocated by boot
> mem after the system is up:
> 
> /*
>  * For memory hotplug, there is no way to free resource entries allocated
>  * by boot mem after the system is up. So for reusing the resource entry
>  * we need to remember the resource.
>  */
> static struct resource *bootmem_resource_free;
> 
> So I think register_memory_resource() can reuse the resource allocated by bootmem.
> Or am I miss anything?

I think you're right, if we did a previous free_resource() of a resource allocated
during boot we could end up reusing that here. My best guess is that this never
really happens.

Wow, that's ugly. It affects essentially anybody reserving+freeing a resource.

E.g., dax/kmem.c similarly does a release_resource(res)+kfree(res)


We could either

a) Expose free_resource() and replace all kfree(res) instances by it

b) Just simplify that. I don't think we care about saving a couple of 
   bytes in corner cases. I might be wrong (IIRC primarily ppc64 really 
   succeeds in unplugging boot memory)


diff --git a/kernel/resource.c b/kernel/resource.c
index 9c08d6e9eef2..fe91a72fd951 100644
--- a/kernel/resource.c
+++ b/kernel/resource.c
@@ -56,14 +56,6 @@ struct resource_constraint {
 
 static DEFINE_RWLOCK(resource_lock);
 
-/*
- * For memory hotplug, there is no way to free resource entries allocated
- * by boot mem after the system is up. So for reusing the resource entry
- * we need to remember the resource.
- */
-static struct resource *bootmem_resource_free;
-static DEFINE_SPINLOCK(bootmem_resource_lock);
-
 static struct resource *next_resource(struct resource *p)
 {
        if (p->child)
@@ -160,36 +152,19 @@ __initcall(ioresources_init);
 
 static void free_resource(struct resource *res)
 {
-       if (!res)
-               return;
-
-       if (!PageSlab(virt_to_head_page(res))) {
-               spin_lock(&bootmem_resource_lock);
-               res->sibling = bootmem_resource_free;
-               bootmem_resource_free = res;
-               spin_unlock(&bootmem_resource_lock);
-       } else {
+       /*
+        * If the resource was allocated using memblock early during boot
+        * we'll leak it here: we can only return full pages back to the
+        * buddy and trying to be smart and reusing them eventually in
+        * alloc_resource() overcomplicates resource handling.
+        */
+       if (res && PageSlab(virt_to_head_page(res)))
                kfree(res);
-       }
 }
 
 static struct resource *alloc_resource(gfp_t flags)
 {
-       struct resource *res = NULL;
-
-       spin_lock(&bootmem_resource_lock);
-       if (bootmem_resource_free) {
-               res = bootmem_resource_free;
-               bootmem_resource_free = res->sibling;
-       }
-       spin_unlock(&bootmem_resource_lock);
-
-       if (res)
-               memset(res, 0, sizeof(struct resource));
-       else
-               res = kzalloc(sizeof(struct resource), flags);
-
-       return res;
+       return kzalloc(sizeof(struct resource), flags);
 }
 
 /* Return the conflict entry if you can't request it */


-- 
Thanks,

David / dhildenb

Re: [PATCH] mm/memory_hotplug: fix kfree() of bootmem memory

[Miaohe Lin]

On 2022/2/8 17:19, David Hildenbrand wrote:
> On 08.02.22 02:59, Miaohe Lin wrote:
>> Hi:
>> On 2022/2/7 22:33, David Hildenbrand wrote:
>>> On 07.02.22 14:56, Miaohe Lin wrote:
>>>> We can't use kfree() to release the resource as it might come from bootmem.
>>>> Use release_mem_region() instead.
>>>
>>> How can this happen? release_mem_region() is called either from
>>> __add_memory() or from add_memory_driver_managed(), where we allocated
>>> the region via register_memory_resource(). Both functions shouldn't ever
>>> be called before the buddy is up an running.
>>>
>>> Do you have a backtrace of an actual instance of this issue? Or was this
>>> identified as possibly broken by code inspection?
>>>
>>
>> This is identified as possibly broken by code inspection. IIUC, alloc_resource
>> is always used to allocate the resource. It has the below logic:
>>
>>   if (bootmem_resource_free) {
>> 	res = bootmem_resource_free;
>> 	bootmem_resource_free = res->sibling;
>>   }
>>
>> where bootmem_resource_free is used to reusing the resource entries allocated by boot
>> mem after the system is up:
>>
>> /*
>>  * For memory hotplug, there is no way to free resource entries allocated
>>  * by boot mem after the system is up. So for reusing the resource entry
>>  * we need to remember the resource.
>>  */
>> static struct resource *bootmem_resource_free;
>>
>> So I think register_memory_resource() can reuse the resource allocated by bootmem.
>> Or am I miss anything?
> 
> I think you're right, if we did a previous free_resource() of a resource allocated
> during boot we could end up reusing that here. My best guess is that this never
> really happens.
> 

Agree with you. This reusing mechanism is introduced since 2013 and this possible
issue never happens.

> Wow, that's ugly. It affects essentially anybody reserving+freeing a resource.
> 
> E.g., dax/kmem.c similarly does a release_resource(res)+kfree(res)
> 
> 
> We could either
> 
> a) Expose free_resource() and replace all kfree(res) instances by it
> 
> b) Just simplify that. I don't think we care about saving a couple of 
>    bytes in corner cases. I might be wrong (IIRC primarily ppc64 really 
>    succeeds in unplugging boot memory)
> 

I prefer this one or there will be a huge change if we choose a.
I will drop my patch and below patch looks good to me.

Many thanks!

> 
> diff --git a/kernel/resource.c b/kernel/resource.c
> index 9c08d6e9eef2..fe91a72fd951 100644
> --- a/kernel/resource.c
> +++ b/kernel/resource.c
> @@ -56,14 +56,6 @@ struct resource_constraint {
>  
>  static DEFINE_RWLOCK(resource_lock);
>  
> -/*
> - * For memory hotplug, there is no way to free resource entries allocated
> - * by boot mem after the system is up. So for reusing the resource entry
> - * we need to remember the resource.
> - */
> -static struct resource *bootmem_resource_free;
> -static DEFINE_SPINLOCK(bootmem_resource_lock);
> -
>  static struct resource *next_resource(struct resource *p)
>  {
>         if (p->child)
> @@ -160,36 +152,19 @@ __initcall(ioresources_init);
>  
>  static void free_resource(struct resource *res)
>  {
> -       if (!res)
> -               return;
> -
> -       if (!PageSlab(virt_to_head_page(res))) {
> -               spin_lock(&bootmem_resource_lock);
> -               res->sibling = bootmem_resource_free;
> -               bootmem_resource_free = res;
> -               spin_unlock(&bootmem_resource_lock);
> -       } else {
> +       /*
> +        * If the resource was allocated using memblock early during boot
> +        * we'll leak it here: we can only return full pages back to the
> +        * buddy and trying to be smart and reusing them eventually in
> +        * alloc_resource() overcomplicates resource handling.
> +        */
> +       if (res && PageSlab(virt_to_head_page(res)))
>                 kfree(res);
> -       }
>  }
>  
>  static struct resource *alloc_resource(gfp_t flags)
>  {
> -       struct resource *res = NULL;
> -
> -       spin_lock(&bootmem_resource_lock);
> -       if (bootmem_resource_free) {
> -               res = bootmem_resource_free;
> -               bootmem_resource_free = res->sibling;
> -       }
> -       spin_unlock(&bootmem_resource_lock);
> -
> -       if (res)
> -               memset(res, 0, sizeof(struct resource));
> -       else
> -               res = kzalloc(sizeof(struct resource), flags);
> -
> -       return res;
> +       return kzalloc(sizeof(struct resource), flags);
>  }
>  
>  /* Return the conflict entry if you can't request it */
> 
>