Re: [RFC PATCH] mm/vmalloc: fix vmalloc which may return null if called with __GFP_NOFAIL

[Gao Xiang <hsiangkao@linux.alibaba.com>]

On 2024/5/9 10:39, Gao Xiang wrote:
> Hi,
> 
> On 2024/5/9 10:20, Barry Song wrote:
>> On Thu, May 9, 2024 at 12:58 AM <hailong.liu@oppo.com> wrote:
>>>
>>> From: "Hailong.Liu" <hailong.liu@oppo.com>
>>>
>>> Commit a421ef303008 ("mm: allow !GFP_KERNEL allocations for kvmalloc")
>>> includes support for __GFP_NOFAIL, but it presents a conflict with
>>> commit dd544141b9eb ("vmalloc: back off when the current task is
>>> OOM-killed"). A possible scenario is as belows:
>>>
>>> process-a
>>> kvcalloc(n, m, GFP_KERNEL | __GFP_NOFAIL)
>>>      __vmalloc_node_range()
>>>          __vmalloc_area_node()
>>>              vm_area_alloc_pages()
>>>              --> oom-killer send SIGKILL to process-a
>>>              if (fatal_signal_pending(current)) break;
>>> --> return NULL;
>>>
>>> to fix this, do not check fatal_signal_pending() in vm_area_alloc_pages()
>>> if __GFP_NOFAIL set.
>>>
>>> Reported-by: Oven <liyangouwen1@oppo.com>
>>> Signed-off-by: Hailong.Liu <hailong.liu@oppo.com>
>>> ---
>>>   mm/vmalloc.c | 2 +-
>>>   1 file changed, 1 insertion(+), 1 deletion(-)
>>>
>>> diff --git a/mm/vmalloc.c b/mm/vmalloc.c
>>> index 6641be0ca80b..2f359d08bf8d 100644
>>> --- a/mm/vmalloc.c
>>> +++ b/mm/vmalloc.c
>>> @@ -3560,7 +3560,7 @@ vm_area_alloc_pages(gfp_t gfp, int nid,
>>>
>>>          /* High-order pages or fallback path if "bulk" fails. */
>>>          while (nr_allocated < nr_pages) {
>>> -               if (fatal_signal_pending(current))
>>> +               if (!(gfp & __GFP_NOFAIL) && fatal_signal_pending(current))
>>>                          break;
>>
>> why not !nofail ?
>>
>> This seems a correct fix, but it undermines the assumption made in
>> commit dd544141b9eb
>>   ("vmalloc: back off when the current task is OOM-killed")
>>
>> "
>>      This may trigger some hidden problems, when caller does not handle
>>      vmalloc failures, or when rollaback after failed vmalloc calls own
>>      vmallocs inside.  However all of these scenarios are incorrect: vmalloc
>>      does not guarantee successful allocation, it has never been called with
>>      __GFP_NOFAIL and threfore either should not be used for any rollbacks or
>>      should handle such errors correctly and not lead to critical failures.
>> "
>>
>> If a significant kvmalloc operation is performed with the NOFAIL flag, it risks
>> reverting the fix intended to address the OOM-killer issue in commit
>> dd544141b9eb.
>> Should we indeed permit the NOFAIL flag for large kvmalloc allocations?
> 
> Just from my perspective, I don't really care about kmalloc, vmalloc
> or kvmalloc (__GFP_NOFAIL).  I even don't care if it returns three
> order-0 pages or a high-order page.   I just would like to need a
> virtual consecutive buffer (even it works slowly.) with __GFP_NOFAIL.
> 
> Because in some cases, writing fallback code may be tough and hard to
> test if such fallback path is correct since it only triggers in extreme
> workloads, and even such buffers are just used in a very short lifetime.

add some words...

    ^ here extreme cases were mostly just generated by syzkaller fuzzing
tests, but if real users try to use some configuration to compress more,
I still think it needs to be handled (even such kvmalloc may be slow if
falling back to order-0 allocations due to memory pressures)

> Also see other FS discussion of __GFP_NOFAIL, e.g.
> https://lore.kernel.org/all/ZcUQfzfQ9R8X0s47@tiehlicka/
> 
> In the worst cases, it usually just needs < 5 order-0 pages (for many
> cases it only needs one page), but with kmalloc it will trigger WARN
> if it occurs to > order-1 allocation. as I mentioned before.
> 
> With my limited understanding I don't see why it could any problem with
> kvmalloc(__GFP_NOFAIL) since it has no difference of kmalloc(GFP_NOFAIL)
> with order-0 allocation.

.. kvmalloc with order-0 pages to form a virtual consecutive buffer
just like several kmalloc(__GFP_NOFAIL) allocations together in the
callers, I don't see any difference of memory pressure here.

Thanks,
Gao Xiang

> 
> 
> Thanks,
> Gao XIang