Re: [virtio-dev] [PATCH] virtio: support VIRTIO_F_ORDER_PLATFORM

[Jason Wang <jasowang@redhat.com>]

On 2019/1/23 上午11:49, Michael S. Tsirkin wrote:
> On Wed, Jan 23, 2019 at 11:08:04AM +0800, Jason Wang wrote:
>> On 2019/1/23 上午1:03, Tiwei Bie wrote:
>>> This patch introduces the support for VIRTIO_F_ORDER_PLATFORM.
>>> When this feature is negotiated, driver will use the barriers
>>> suitable for hardware devices.
>>>
>>> Signed-off-by: Tiwei Bie <tiwei.bie@intel.com>
>>> ---
>>>    drivers/virtio/virtio_ring.c       | 8 ++++++++
>>>    include/uapi/linux/virtio_config.h | 6 ++++++
>>>    2 files changed, 14 insertions(+)
>>>
>>> diff --git a/drivers/virtio/virtio_ring.c b/drivers/virtio/virtio_ring.c
>>> index cd7e755484e3..27d3f057493e 100644
>>> --- a/drivers/virtio/virtio_ring.c
>>> +++ b/drivers/virtio/virtio_ring.c
>>> @@ -1609,6 +1609,9 @@ static struct virtqueue *vring_create_virtqueue_packed(
>>>    		!context;
>>>    	vq->event = virtio_has_feature(vdev, VIRTIO_RING_F_EVENT_IDX);
>>> +	if (virtio_has_feature(vdev, VIRTIO_F_ORDER_PLATFORM))
>>> +		vq->weak_barriers = false;
>>> +
>>>    	vq->packed.ring_dma_addr = ring_dma_addr;
>>>    	vq->packed.driver_event_dma_addr = driver_event_dma_addr;
>>>    	vq->packed.device_event_dma_addr = device_event_dma_addr;
>>> @@ -2079,6 +2082,9 @@ struct virtqueue *__vring_new_virtqueue(unsigned int index,
>>>    		!context;
>>>    	vq->event = virtio_has_feature(vdev, VIRTIO_RING_F_EVENT_IDX);
>>> +	if (virtio_has_feature(vdev, VIRTIO_F_ORDER_PLATFORM))
>>> +		vq->weak_barriers = false;
>>> +
>>>    	vq->split.queue_dma_addr = 0;
>>>    	vq->split.queue_size_in_bytes = 0;
>>> @@ -2213,6 +2219,8 @@ void vring_transport_features(struct virtio_device *vdev)
>>>    			break;
>>>    		case VIRTIO_F_RING_PACKED:
>>>    			break;
>>> +		case VIRTIO_F_ORDER_PLATFORM:
>>> +			break;
>>>    		default:
>>>    			/* We don't understand this bit. */
>>>    			__virtio_clear_bit(vdev, i);
>>> diff --git a/include/uapi/linux/virtio_config.h b/include/uapi/linux/virtio_config.h
>>> index 1196e1c1d4f6..ff8e7dc9d4dd 100644
>>> --- a/include/uapi/linux/virtio_config.h
>>> +++ b/include/uapi/linux/virtio_config.h
>>> @@ -78,6 +78,12 @@
>>>    /* This feature indicates support for the packed virtqueue layout. */
>>>    #define VIRTIO_F_RING_PACKED		34
>>> +/*
>>> + * This feature indicates that memory accesses by the driver and the
>>> + * device are ordered in a way described by the platform.
>>> + */
>>> +#define VIRTIO_F_ORDER_PLATFORM		36
>>> +
>>>    /*
>>>     * Does the device support Single Root I/O Virtualization?
>>>     */
>>
>> I wonder whether or not this is sufficient. Is dma barrier implies a mmio
>> barrier? Looks not.
> IIUC we don't need an mmio barrier because we are using a
> serializing API: Documentation/memory-barriers.txt says:
>
> 	Note that, when using writel(), a prior
>       wmb() is not needed to guarantee that the cache coherent memory writes
>       have completed before writing to the MMIO region.


Ah, I get this.


>
>
>> See ia64/include/asm/barrier.h:
>>
>>   * Note: "mb()" and its variants cannot be used as a fence to order
>>   * accesses to memory mapped I/O registers.  For that, mf.a needs to
>>   * be used.  However, we don't want to always use mf.a because (a)
>>   * it's (presumably) much slower than mf and (b) mf.a is supported for
>>   * sequential memory pages only.
>>   */
>> #define mb()            ia64_mf()
>> #define rmb()           mb()
>> #define wmb()           mb()
>>
>> #define dma_rmb()       mb()
>> =>efine dma_wmb()       mb()
>>
>> Thanks
> Frankly no idea about ia64.


Neither did me.


>   Sorry. Are any less esoteric platforms
> affected?
>

E.g ppc64?

define dma_wmb()       __asm__ __volatile__ (stringify_in_c(SMPWMB) : : 
:"memo\
ry")

/*
  * Enforce synchronisation of stores vs. spin_unlock
  * (this does it explicitly, though our implementation of spin_unlock
  * does it implicitely too)
  */
static inline void mmiowb(void)
{
         unsigned long tmp;

         __asm__ __volatile__("sync; li %0,0; stb %0,%1(13)"
         : "=&r" (tmp) : "i" (offsetof(struct paca_struct, io_sync))
         : "memory");
}

dma_wmb() is lwsync which is more lightweight than sync I guess?

Thanks