Re: [dpdk-dev] [PATCH 05/10] net/virtio: refactor virtqueue structure

[Tiwei Bie <tiwei.bie@intel.com>]

On Tue, Mar 19, 2019 at 02:59:38PM +0000, Kevin Traynor wrote:
> On 19/03/2019 13:50, Maxime Coquelin wrote:
> > 
> > 
> > On 3/19/19 2:47 PM, Jens Freimann wrote:
> >> On Tue, Mar 19, 2019 at 02:28:30PM +0100, Maxime Coquelin wrote:
> >>>
> >>>
> >>> On 3/19/19 11:09 AM, Tiwei Bie wrote:
> >>>> On Tue, Mar 19, 2019 at 10:44:32AM +0100, Jens Freimann wrote:
> >>>>> On Tue, Mar 19, 2019 at 02:43:07PM +0800, Tiwei Bie wrote:
> >>>>>> Put split ring and packed ring specific fields into separate
> >>>>>> sub-structures, and also union them as they won't be available
> >>>>>> at the same time.
> >>>>>>
> >>>>>> Signed-off-by: Tiwei Bie <tiwei.bie@intel.com>
> >>>>>> ---
> >>>>>> drivers/net/virtio/virtio_ethdev.c           | 71 +++++++++---------
> >>>>>> drivers/net/virtio/virtio_rxtx.c             | 66 ++++++++---------
> >>>>>> drivers/net/virtio/virtio_rxtx_simple.h      |  2 +-
> >>>>>> drivers/net/virtio/virtio_rxtx_simple_neon.c |  2 +-
> >>>>>> drivers/net/virtio/virtio_rxtx_simple_sse.c  |  2 +-
> >>>>>> drivers/net/virtio/virtqueue.c               |  6 +-
> >>>>>> drivers/net/virtio/virtqueue.h               | 77 +++++++++++---------
> >>>>>> 7 files changed, 117 insertions(+), 109 deletions(-)
> >>>>>>
> >>>>> [snip]
> >>>>> ...
> >>>>>> diff --git a/drivers/net/virtio/virtqueue.h 
> >>>>>> b/drivers/net/virtio/virtqueue.h
> >>>>>> index 80c0c43c3..48b3912e6 100644
> >>>>>> --- a/drivers/net/virtio/virtqueue.h
> >>>>>> +++ b/drivers/net/virtio/virtqueue.h
> >>>>>> @@ -191,17 +191,22 @@ struct vq_desc_extra {
> >>>>>>
> >>>>>> struct virtqueue {
> >>>>>>     struct virtio_hw  *hw; /**< virtio_hw structure pointer. */
> >>>>>> -    struct vring vq_ring;  /**< vring keeping desc, used and avail */
> >>>>>> -    struct vring_packed ring_packed;  /**< vring keeping descs */
> >>>>>> -    bool used_wrap_counter;
> >>>>>> -    uint16_t cached_flags; /**< cached flags for descs */
> >>>>>> -    uint16_t event_flags_shadow;
> >>>>>> +    union {
> >>>>>> +        struct {
> >>>>>> +            /**< vring keeping desc, used and avail */
> >>>>>> +            struct vring ring;
> >>>>>> +        } vq_split;
> >>>>>>
> >>>>>> -    /**
> >>>>>> -     * Last consumed descriptor in the used table,
> >>>>>> -     * trails vq_ring.used->idx.
> >>>>>> -     */
> >>>>>> -    uint16_t vq_used_cons_idx;
> >>>>>> +        struct {
> >>>>>> +            /**< vring keeping descs and events */
> >>>>>> +            struct vring_packed ring;
> >>>>>> +            bool used_wrap_counter;
> >>>>>> +            uint16_t cached_flags; /**< cached flags for descs */
> >>>>>> +            uint16_t event_flags_shadow;
> >>>>>> +        } vq_packed;
> >>>>>> +    };
> >>>>>> +
> >>>>>> +    uint16_t vq_used_cons_idx; /**< last consumed descriptor */
> >>>>>>     uint16_t vq_nentries;  /**< vring desc numbers */
> >>>>>>     uint16_t vq_free_cnt;  /**< num of desc available */
> >>>>>>     uint16_t vq_avail_idx; /**< sync until needed */
> >>>>>
> >>>>> Honest question: What do we really gain by putting it in a union? We
> >>>>> save a little memory. But we also make code less readable IMHO.
> >>>>
> >>>> I think it will make it clear that fields like used_wrap_counter
> >>>> are only available in packed ring which will make the code more
> >>>> readable.
> >>>>
> >>>>>
> >>>>> If we do this, can we at least shorten some variable names, like drop
> >>>>> the vq_ prefix? (It's used everywhere like vq->vq_packed*, so with
> >>>>> vq->packed* we don't loose any context).
> >>>>
> >>>> I prefer to have consistent prefix like most fields in this
> >>>> structure (although some fields don't really follow this).
> >>>
> >>> As Jens, I tend to agree that the vq_ prefix is quite redundant.
> >>> However, I think it is better to keep it in this patch for consistency.
> >>>
> >>> Maybe it can be remove in a separate patch later?
> >>
> >> I thought it might be convenient to change it now as we are touching
> >> all related code anyway. But I also don't want to block the patch 
> >> because of
> >> this cosmetic thing. So let's defer it to a later patch set.
> > 
> > OK, when I meant later, I meant to remove vq_ prefix for all fields, not
> > only vq_split & vq_packed.
> > 
> > But yes, that's just cosmetic so let's keep it as is for now.
> > 
> 
> I agree the vq_ prefix is not needed and I think the code is more
> readable in general seeing the packed/split name when using the struct.
> 
> Please also consider that cosmetic changes in multiple places likely
> mean backports will not apply cleanly to the stable branches anymore, so
> it does have a cost.

Yeah, agree.

> Although in this case, iirc packed rings are not in
> 18.11, so fixes might need dedicated backports from authors anyway, and
> there haven't been too many virtio backports to date.
> 
> >>
> >> regards,
> >> Jens
>