[dpdk-dev] [PATCH 2/3] vhost: move dirty logging cache out of the virtqueue

[Xia, Chenbo]

Hi Maxime,

> -----Original Message-----
> From: Maxime Coquelin <maxime.coquelin at redhat.com>
> Sent: Tuesday, December 22, 2020 9:57 PM
> To: dev at dpdk.org; Xia, Chenbo <chenbo.xia at intel.com>; amorenoz at redhat.com
> Cc: Maxime Coquelin <maxime.coquelin at redhat.com>
> Subject: [PATCH 2/3] vhost: move dirty logging cache out of the virtqueue
> 
> This patch moves the per-virtqueue's dirty logging cache
> out of the virtqueue struct, by allocating it dynamically
> only when live-migration is enabled.
> 
> It saves 8 cachelines in vhost_virtqueue struct.
> 
> Signed-off-by: Maxime Coquelin <maxime.coquelin at redhat.com>
> ---
>  lib/librte_vhost/vhost.c      | 12 ++++++++++++
>  lib/librte_vhost/vhost.h      |  2 +-
>  lib/librte_vhost/vhost_user.c | 25 +++++++++++++++++++++++++
>  3 files changed, 38 insertions(+), 1 deletion(-)
> 
> diff --git a/lib/librte_vhost/vhost.c b/lib/librte_vhost/vhost.c
> index 4e5df862aa..ec6459b2d1 100644
> --- a/lib/librte_vhost/vhost.c
> +++ b/lib/librte_vhost/vhost.c
> @@ -144,6 +144,10 @@ __vhost_log_cache_sync(struct virtio_net *dev, struct
> vhost_virtqueue *vq)
>  	if (unlikely(!dev->log_base))
>  		return;
> 
> +	/* No cache, nothing to sync */
> +	if (unlikely(!vq->log_cache))
> +		return;
> +
>  	rte_smp_wmb();
> 
>  	log_base = (unsigned long *)(uintptr_t)dev->log_base;
> @@ -176,6 +180,14 @@ vhost_log_cache_page(struct virtio_net *dev, struct
> vhost_virtqueue *vq,
>  	uint32_t offset = page / (sizeof(unsigned long) << 3);
>  	int i;
> 
> +	if (unlikely(!vq->log_cache)) {
> +		/* No logging cache allocated, write dirty log map directly */
> +		rte_smp_wmb();
> +		vhost_log_page((uint8_t *)(uintptr_t)dev->log_base, page);
> +
> +		return;
> +	}
> +
>  	for (i = 0; i < vq->log_cache_nb_elem; i++) {
>  		struct log_cache_entry *elem = vq->log_cache + i;
> 
> diff --git a/lib/librte_vhost/vhost.h b/lib/librte_vhost/vhost.h
> index d132e4ae54..e2f14034b4 100644
> --- a/lib/librte_vhost/vhost.h
> +++ b/lib/librte_vhost/vhost.h
> @@ -183,7 +183,7 @@ struct vhost_virtqueue {
>  	bool			used_wrap_counter;
>  	bool			avail_wrap_counter;
> 
> -	struct log_cache_entry log_cache[VHOST_LOG_CACHE_NR];
> +	struct log_cache_entry *log_cache;
>  	uint16_t log_cache_nb_elem;
> 
>  	rte_rwlock_t	iotlb_lock;
> diff --git a/lib/librte_vhost/vhost_user.c b/lib/librte_vhost/vhost_user.c
> index 45c8ac09da..7ac3963a07 100644
> --- a/lib/librte_vhost/vhost_user.c
> +++ b/lib/librte_vhost/vhost_user.c
> @@ -1978,6 +1978,11 @@ vhost_user_get_vring_base(struct virtio_net **pdev,
>  	rte_free(vq->batch_copy_elems);
>  	vq->batch_copy_elems = NULL;
> 
> +	if (vq->log_cache) {
> +		rte_free(vq->log_cache);
> +		vq->log_cache = NULL;
> +	}
> +

I think we'd better check and free the log cache in function free_vq() too?
It is possible that during migration, vhost destroys, right?

Thanks,
Chenbo

>  	msg->size = sizeof(msg->payload.state);
>  	msg->fd_num = 0;
> 
> @@ -2077,6 +2082,7 @@ vhost_user_set_log_base(struct virtio_net **pdev, struct
> VhostUserMsg *msg,
>  	int fd = msg->fds[0];
>  	uint64_t size, off;
>  	void *addr;
> +	uint32_t i;
> 
>  	if (validate_msg_fds(msg, 1) != 0)
>  		return RTE_VHOST_MSG_RESULT_ERR;
> @@ -2130,6 +2136,25 @@ vhost_user_set_log_base(struct virtio_net **pdev,
> struct VhostUserMsg *msg,
>  	dev->log_base = dev->log_addr + off;
>  	dev->log_size = size;
> 
> +	for (i = 0; i < dev->nr_vring; i++) {
> +		struct vhost_virtqueue *vq = dev->virtqueue[i];
> +
> +		if (vq->log_cache) {
> +			rte_free(vq->log_cache);
> +			vq->log_cache = NULL;
> +		}
> +		vq->log_cache_nb_elem = 0;
> +		vq->log_cache = rte_zmalloc("vq log cache",
> +				sizeof(struct log_cache_entry) * VHOST_LOG_CACHE_NR,
> +				0);
> +		/*
> +		 * If log cache alloc fail, don't fail migration, but no
> +		 * caching will be done, which will impact performance
> +		 */
> +		if (!vq->log_cache)
> +			VHOST_LOG_CONFIG(ERR, "Failed to allocate VQ logging
> cache\n");
> +	}
> +
>  	/*
>  	 * The spec is not clear about it (yet), but QEMU doesn't expect
>  	 * any payload in the reply.
> --
> 2.29.2