Re: [PATCH 2/6] mempolicy: remove all mempolicy sharing

[KOSAKI Motohiro <kosaki.motohiro@gmail.com>]

On Mon, Jun 11, 2012 at 11:02 AM, Christoph Lameter <cl@linux.com> wrote:
> Some more attempts to cleanup changelogs:
>
>> The problem was created by a reference count imbalance. Example, In following case,
>> mbind(addr, len) try to replace mempolicies of vma1 and vma2 and then they will
>> be share the same mempolicy, and the new mempolicy has MPOL_F_SHARED flag.
>
> The bug that we saw <where ? details?> was created by a refcount
> imbalance. If mbind() replaces the memory policies of vma1 and vma and
> they share the same shared mempolicy (MPOL_F_SHARED set) then an imbalance
> may occur.
>
>>   +-------------------+-------------------+
>>   |     vma1          |     vma2(shmem)   |
>>   +-------------------+-------------------+
>>   |                                       |
>>  addr                                 addr+len
>>
>> Look at alloc_pages_vma(), it uses get_vma_policy() and mpol_cond_put() pair
>> for maintaining mempolicy refcount. The current rule is, get_vma_policy() does
>> NOT increase a refcount if the policy is not attached shmem vma and mpol_cond_put()
>> DOES decrease a refcount if mpol has MPOL_F_SHARED.
>
> alloc_pages_vma() uses the two function get_vma_policy() and
> mpol_cond_put() to maintain the refcount on the memory policies. However,
> the current rule is that get_vma_policy() does *not* increase the refcount
> if the policy is not attached to a shm vma. mpol_cond_put *does* decrease
> the refcount if the memory policy has MPOL_F_SHARED set.
>
>> In above case, vma1 is not shmem vma and vma->policy has MPOL_F_SHARED! then,
>> get_vma_policy() doesn't increase a refcount and mpol_cond_put() decrease a
>> refcount whenever alloc_page_vma() is called.
>>
>> The bug was introduced by commit 52cd3b0740 (mempolicy: rework mempolicy Reference
>> Counting) at 4 years ago.
>>
>> More unfortunately mempolicy has one another serious broken. Currently,
>> mempolicy rebind logic (it is called from cpuset rebinding) ignore a refcount
>> of mempolicy and override it forcibly. Thus, any mempolicy sharing may
>> cause mempolicy corruption. The bug was introduced by commit 68860ec10b
>> (cpusets: automatic numa mempolicy rebinding) at 7 years ago.
>
> Memory policies have another issue. Currently the mempolicy rebind logic
> used for cpuset rebinding ignores the refcount of memory policies.
> Therefore, any memory policy sharing can cause refcount mismatches. The
> bug was ...
>
>> To disable policy sharing solves user visible breakage and this patch does it.
>> Maybe, we need to rewrite MPOL_F_SHARED and mempolicy rebinding code and aim
>> to proper cow logic eventually, but I think this is good first step.
>
> Disabling policy sharing solves the breakage and that is how this patch
> fixes the issue for now. Rewriting the shared policy handling with proper
> COW logic support will be necessary to cleanly address the
> problem and allow proper sharing of memory policies.

Thanks, Christoph.
I'll rewrite the description as your suggestion.

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>