Michael S. Tsirkin wrote:
> On Tue, Sep 02, 2025 at 10:09:54AM +0200, Simon Schippers wrote:
>> The implementation is inspired by ptr_ring_empty.
>>
>> Co-developed-by: Tim Gebauer <tim.geba...@tu-dortmund.de>
>> Signed-off-by: Tim Gebauer <tim.geba...@tu-dortmund.de>
>> Signed-off-by: Simon Schippers <simon.schipp...@tu-dortmund.de>
>> ---
>> include/linux/ptr_ring.h | 71 ++++++++++++++++++++++++++++++++++++++++
>> 1 file changed, 71 insertions(+)
>>
>> diff --git a/include/linux/ptr_ring.h b/include/linux/ptr_ring.h
>> index 551329220e4f..6b8cfaecf478 100644
>> --- a/include/linux/ptr_ring.h
>> +++ b/include/linux/ptr_ring.h
>> @@ -243,6 +243,77 @@ static inline bool ptr_ring_empty_bh(struct ptr_ring *r)
>> return ret;
>> }
>>
>> +/*
>> + * Check if a spare capacity of cnt is available without taking any locks.
>
> Not sure what "spare" means here. I think you mean
>
> Check if the ring has enough space to produce a given
> number of entries.
>
>> + *
>> + * If cnt==0 or cnt > r->size it acts the same as __ptr_ring_empty.
>
> Logically, cnt = 0 should always be true, cnt > size should always be
> false then?
>
> Why do you want it to act as __ptr_ring_empty?
>
>
>> + *
>> + * The same requirements apply as described for __ptr_ring_empty.
>
>
> Which is:
>
> * However, if some other CPU consumes ring entries at the same time, the
> value
> * returned is not guaranteed to be correct.
>
>
> but it's not right here yes? consuming entries will just add more
> space ...
>
> Also:
> * In this case - to avoid incorrectly detecting the ring
> * as empty - the CPU consuming the ring entries is responsible
> * for either consuming all ring entries until the ring is empty,
> * or synchronizing with some other CPU and causing it to
> * re-test __ptr_ring_empty and/or consume the ring enteries
> * after the synchronization point.
>
> how would you apply this here?
>
>
>> + */
>> +static inline bool __ptr_ring_spare(struct ptr_ring *r, int cnt)
>> +{
>> + int size = r->size;
>> + int to_check;
>> +
>> + if (unlikely(!size || cnt < 0))
>> + return true;
>> +
>> + if (cnt > size)
>> + cnt = 0;
>> +
>> + to_check = READ_ONCE(r->consumer_head) - cnt;
>> +
>> + if (to_check < 0)
>> + to_check += size;
>> +
>> + return !r->queue[to_check];
>> +}
>> +
>
> I will have to look at how this is used to understand if it's
> correct. But I think we need better documentation.
>
>
>> +static inline bool ptr_ring_spare(struct ptr_ring *r, int cnt)
>> +{
>> + bool ret;
>> +
>> + spin_lock(&r->consumer_lock);
>> + ret = __ptr_ring_spare(r, cnt);
>> + spin_unlock(&r->consumer_lock);
>> +
>> + return ret;
>
>
> I don't understand why you take the consumer lock here.
> If a producer is running it will make the value wrong,
> if consumer is running it will just create more space.
>
>
I agree, I messed up the ptr_ring helper.
Your proposed approach is way superior and I will use that one instead.
The idea behind the cnt was to have an option if the producer may produce
multiple entries like tap_handle_frame with GSO. But of course this should
be in a different patch since I will not cover tap_handle_frame, which is
used by ipvtap and macvtap, in this patch series.
