Re: [PATCH 4/7] mm: defer vmalloc from atomic context

[Joel Fernandes]

On Mon, Nov 7, 2016 at 7:01 AM, Andrey Ryabinin <aryabinin@xxxxxxxxxxxxx> wrote:
> On 11/05/2016 06:43 AM, Joel Fernandes wrote:
>> On Mon, Oct 24, 2016 at 8:44 AM, Andrey Ryabinin
>> <aryabinin@xxxxxxxxxxxxx> wrote:
>>>
>>>
>>> On 10/22/2016 06:17 PM, Christoph Hellwig wrote:
>>>> We want to be able to use a sleeping lock for freeing vmap to keep
>>>> latency down.  For this we need to use the deferred vfree mechanisms
>>>> no only from interrupt, but from any atomic context.
>>>>
>>>> Signed-off-by: Christoph Hellwig <hch@xxxxxx>
>>>> ---
>>>>  mm/vmalloc.c | 2 +-
>>>>  1 file changed, 1 insertion(+), 1 deletion(-)
>>>>
>>>> diff --git a/mm/vmalloc.c b/mm/vmalloc.c
>>>> index a4e2cec..bcc1a64 100644
>>>> --- a/mm/vmalloc.c
>>>> +++ b/mm/vmalloc.c
>>>> @@ -1509,7 +1509,7 @@ void vfree(const void *addr)
>>>>
>>>>       if (!addr)
>>>>               return;
>>>> -     if (unlikely(in_interrupt())) {
>>>> +     if (unlikely(in_atomic())) {
>>>
>>> in_atomic() cannot always detect atomic context, thus it shouldn't be used here.
>>> You can add something like vfree_in_atomic() and use it in atomic call sites.
>>>
>>
>> So because in_atomic doesn't work for !CONFIG_PREEMPT kernels, can we
>> always defer the work in these cases?
>>
>> So for non-preemptible kernels, we always defer:
>>
>> if (!IS_ENABLED(CONFIG_PREEMPT) || in_atomic()) {
>>   // defer
>> }
>>
>> Is this fine? Or any other ideas?
>>
>
> What's wrong with my idea?
> We can add vfree_in_atomic() and use it to free vmapped stacks
> and for any other places where vfree() used 'in_atomict() && !in_interrupt()' context.

Yes, this sounds like a better idea as there may not be that many
callers and my idea may hurt perf.

Thanks,

Joel