Re: [dpdk-dev] 答复: [PATCH v4] vhost: support inflight share memory protocol feature

[Maxime Coquelin <maxime.coquelin@redhat.com>]

On 5/22/19 1:04 PM, Li,Lin(ACG Cloud) wrote:
> 
> 
>> -----邮件原件-----
>> 发件人: Maxime Coquelin [mailto:maxime.coquelin@redhat.com]
>> 发送时间: 2019年5月17日 23:47
>> 收件人: Li Lin <chuckylinchuckylin@gmail.com>; tiwei.bie@intel.com;
>> zhihong.wang@intel.com
>> 抄送: dev@dpdk.org; dariusz.stojaczyk@intel.com; changpeng.liu@intel.com;
>> james.r.harris@intel.com; Xie,Yongji <xieyongji@baidu.com>; Li,Lin(ACG Cloud)
>> <lilin24@baidu.com>; Ni,Xun <nixun@baidu.com>; Zhang,Yu(ACG Cloud)
>> <zhangyu31@baidu.com>
>> 主题: Re: [PATCH v4] vhost: support inflight share memory protocol feature
>>
>>
>>
>> On 5/5/19 11:02 AM, Li Lin wrote:
>>> From: Li Lin <lilin24@baidu.com>
>>>
>>> This patch introduces two new messages VHOST_USER_GET_INFLIGHT_FD
>> and
>>> VHOST_USER_SET_INFLIGHT_FD to support transferring a shared buffer
>>> between qemu and backend.
>>>
>>> Firstly, qemu uses VHOST_USER_GET_INFLIGHT_FD to get the shared buffer
>>> from backend. Then qemu should send it back through
>>> VHOST_USER_SET_INFLIGHT_FD each time we start vhost-user.
>>>
>>> This shared buffer is used to process inflight I/O when backend
>>> reconnect.
>>>
>>> Signed-off-by: Li Lin <lilin24@baidu.com>
>>> Signed-off-by: Ni Xun <nixun@baidu.com>
>>> Signed-off-by: Zhang Yu <zhangyu31@baidu.com>
>>> ---
>>> v2 - add set&clr inflight desc functions
>>> v3 - simplified some functions implementation
>>> v4 - rework some data structures according to qemu doc
>>>
>>>    lib/librte_vhost/rte_vhost.h           |  83 ++++++++++
>>>    lib/librte_vhost/rte_vhost_version.map |   3 +
>>>    lib/librte_vhost/vhost.c               | 105 ++++++++++++
>>>    lib/librte_vhost/vhost.h               |  12 ++
>>>    lib/librte_vhost/vhost_user.c          | 292
>> +++++++++++++++++++++++++++++++++
>>>    lib/librte_vhost/vhost_user.h          |  13 +-
>>>    6 files changed, 507 insertions(+), 1 deletion(-)
>>>
>>> diff --git a/lib/librte_vhost/rte_vhost.h
>>> b/lib/librte_vhost/rte_vhost.h index d2c0c93f4..d4d709717 100644
>>> --- a/lib/librte_vhost/rte_vhost.h
>>> +++ b/lib/librte_vhost/rte_vhost.h
>>> @@ -71,6 +71,10 @@ extern "C" {
>>>    #define VHOST_USER_PROTOCOL_F_HOST_NOTIFIER 11
>>>    #endif
>>>
>>> +#ifndef VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD
>>> +#define VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD 12 #endif
>>> +
>>>    /** Indicate whether protocol features negotiation is supported. */
>>>    #ifndef VHOST_USER_F_PROTOCOL_FEATURES
>>>    #define VHOST_USER_F_PROTOCOL_FEATURES	30
>>> @@ -98,12 +102,41 @@ struct rte_vhost_memory {
>>>    	struct rte_vhost_mem_region regions[];
>>>    };
>>>
>>> +struct inflight_desc {
>>> +	uint8_t inflight;
>>> +	uint8_t padding[5];
>>> +	uint16_t next;
>>> +	uint64_t counter;
>>> +};
>>> +
>>> +struct inflight_info {
>>> +	uint64_t features;
>>> +	uint16_t version;
>>> +	uint16_t desc_num;
>>> +	uint16_t last_inflight_io;
>>> +	uint16_t used_idx;
>>> +	struct inflight_desc desc[0];
>>> +};
>>> +
>>> +struct resubmit_desc {
>>> +	uint16_t index;
>>> +	uint64_t counter;
>>> +};
>>> +
>>> +struct resubmit_info {
>>> +	struct resubmit_desc *resubmit_list;
>>> +	uint16_t resubmit_num;
>>> +};
>>> +
>>>    struct rte_vhost_vring {
>>>    	struct vring_desc	*desc;
>>>    	struct vring_avail	*avail;
>>>    	struct vring_used	*used;
>>>    	uint64_t		log_guest_addr;
>>>
>>> +	struct inflight_info	*inflight;
>>> +	struct resubmit_info	*resubmit;
>>> +
>>>    	/** Deprecated, use rte_vhost_vring_call() instead. */
>>>    	int			callfd;
>>>
>>> @@ -632,6 +665,56 @@ int rte_vhost_get_vhost_vring(int vid, uint16_t
>> vring_idx,
>>>    int rte_vhost_vring_call(int vid, uint16_t vring_idx);
>>>
>>>    /**
>>> + * set inflight flag for a desc.
>>> + *
>>> + * @param vid
>>> + *  vhost device ID
>>> + * @param vring_idx
>>> + *  vring index
>>> + * @param idx
>>> + *  inflight entry index
>>> + * @return
>>> + *  0 on success, -1 on failure
>>> + */
>>> +int __rte_experimental
>>> +rte_vhost_set_inflight_desc(int vid, uint16_t vring_idx,
>>> +		uint16_t idx);
>>> +
>>> +/**
>>> + * clear inflight flag for a desc.
>>> + *
>>> + * @param vid
>>> + *  vhost device ID
>>> + * @param vring_idx
>>> + *  vring index
>>> + * @param last_used_idx
>>> + *  next free used_idx
>>> + * @param idx
>>> + *  inflight entry index
>>> + * @return
>>> + *  0 on success, -1 on failure
>>> + */
>>> +int __rte_experimental
>>> +rte_vhost_clr_inflight_desc(int vid, uint16_t vring_idx,
>>> +		uint16_t last_used_idx, uint16_t idx);
>>> +
>>> +/**
>>> + * set last inflight io index.
>>> + *
>>> + * @param vid
>>> + *  vhost device ID
>>> + * @param vring_idx
>>> + *  vring index
>>> + * @param idx
>>> + *  inflight entry index
>>> + * @return
>>> + *  0 on success, -1 on failure
>>> + */
>>> +int __rte_experimental
>>> +rte_vhost_set_last_inflight_io(int vid, uint16_t vring_idx,
>>> +		uint16_t idx);
>>> +
>>> +/**
>>>     * Get vhost RX queue avail count.
>>>     *
>>>     * @param vid
>>> diff --git a/lib/librte_vhost/rte_vhost_version.map
>>> b/lib/librte_vhost/rte_vhost_version.map
>>> index 5f1d4a75c..a992b3935 100644
>>> --- a/lib/librte_vhost/rte_vhost_version.map
>>> +++ b/lib/librte_vhost/rte_vhost_version.map
>>> @@ -87,4 +87,7 @@ EXPERIMENTAL {
>>>    	rte_vdpa_relay_vring_used;
>>>    	rte_vhost_extern_callback_register;
>>>    	rte_vhost_driver_set_protocol_features;
>>> +	rte_vhost_set_inflight_desc;
>>> +	rte_vhost_clr_inflight_desc;
>>> +	rte_vhost_set_last_inflight_io;
>>>    };
>>> diff --git a/lib/librte_vhost/vhost.c b/lib/librte_vhost/vhost.c index
>>> 163f4595e..03eab3551 100644
>>> --- a/lib/librte_vhost/vhost.c
>>> +++ b/lib/librte_vhost/vhost.c
>>> @@ -76,6 +76,16 @@ cleanup_vq(struct vhost_virtqueue *vq, int destroy)
>>>    		close(vq->callfd);
>>>    	if (vq->kickfd >= 0)
>>>    		close(vq->kickfd);
>>> +	if (vq->inflight)
>>> +		vq->inflight = NULL;
>>> +	if (vq->resubmit->resubmit_list) {
>>> +		free(vq->resubmit->resubmit_list);
>>> +		vq->resubmit->resubmit_list = NULL;
>>> +	}
>>> +	if (vq->resubmit) {
>>> +		free(vq->resubmit);
>>> +		vq->resubmit = NULL;
>>> +	}
>>>    }
>>>
>>>    /*
>>> @@ -589,6 +599,9 @@ rte_vhost_get_vhost_vring(int vid, uint16_t
>> vring_idx,
>>>    	vring->kickfd  = vq->kickfd;
>>>    	vring->size    = vq->size;
>>>
>>> +	vring->inflight = vq->inflight;
>>> +	vring->resubmit = vq->resubmit;
>>> +
>>>    	return 0;
>>>    }
>>>
>>> @@ -617,6 +630,98 @@ rte_vhost_vring_call(int vid, uint16_t vring_idx)
>>>    	return 0;
>>>    }
>>>
>>> +int
>>> +rte_vhost_set_inflight_desc(int vid, uint16_t vring_idx, uint16_t
>>> +idx) {
>>> +	struct virtio_net *dev;
>>> +	struct vhost_virtqueue *vq;
>>> +
>>> +	dev = get_device(vid);
>>> +	if (!dev)
>>> +		return -1;
>>> +
>>> +	if (!(dev->features &
>>> +		(1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD)))
>>> +		return 0;
>>
>> Shouldn't -1 be returned here also?
>> I think this function should be called only if the feature negotiation succeeded.
>>
>> Returning an error here would help the calling application to know something
>> went wrong.
>>
>> Same comment applies for the clear function.
> 
> If the feature is not set，set & clr function returns directly.
> They are similar to empty functions.
> The return value is - 1, which represents an error in function execution
> and unexpected behavior.
> 
>>> +
>>> +	if (vring_idx >= VHOST_MAX_VRING)
>>> +		return -1;
>>> +
>>> +	vq = dev->virtqueue[vring_idx];
>>> +	if (!vq)
>>> +		return -1;
>>> +
>>> +	if (unlikely(!vq->inflight))
>>> +		return -1;
>>> +
>>> +	vq->inflight->desc[idx].counter = vq->counter++;
>>> +	vq->inflight->desc[idx].inflight = 1;
>>> +	return 0;
>>> +}
>>> +
>>> +int
>>> +rte_vhost_clr_inflight_desc(int vid, uint16_t vring_idx,
>>> +	uint16_t last_used_idx, uint16_t idx) {
>>> +	struct virtio_net *dev;
>>> +	struct vhost_virtqueue *vq;
>>> +
>>> +	dev = get_device(vid);
>>> +	if (!dev)
>>> +		return -1;
>>> +
>>> +	if (!(dev->features &
>>> +		(1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD)))
>>> +		return 0;
>>> +
>>> +	if (vring_idx >= VHOST_MAX_VRING)
>>> +		return -1;
>>> +
>>> +	vq = dev->virtqueue[vring_idx];
>>> +	if (!vq)
>>> +		return -1;
>>> +
>>> +	if (unlikely(!vq->inflight))
>>> +		return -1;
>>> +
>>> +	rte_compiler_barrier();
>>> +
>>> +	vq->inflight->desc[idx].inflight = 0;
>>> +
>>> +	rte_compiler_barrier();
>>> +
>>> +	vq->inflight->used_idx = last_used_idx;
>>> +	return 0;
>>> +}
>>> +
>>> +int
>>> +rte_vhost_set_last_inflight_io(int vid, uint16_t vring_idx, uint16_t
>>> +idx) {
>>> +	struct virtio_net *dev;
>>> +	struct vhost_virtqueue *vq;
>>> +
>>> +	dev = get_device(vid);
>>> +	if (!dev)
>>> +		return -1;
>>> +
>>> +	if (!(dev->features &
>>> +		(1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD)))
>>> +		return 0;
>>> +
>>> +	if (vring_idx >= VHOST_MAX_VRING)
>>> +		return -1;
>>> +
>>> +	vq = dev->virtqueue[vring_idx];
>>> +	if (!vq)
>>> +		return -1;
>>> +
>>> +	if (unlikely(!vq->inflight))
>>> +		return -1;
>>> +
>>> +	vq->inflight->last_inflight_io = idx;
>>> +	return 0;
>>> +}
>>
>> Are above functions supposed to be called in the hot path?
>> If so, you might want to use unlikely for the error checks everywhere.
> 
> You're right. I'll modify it in the next version.
> 
>>
>>> +
>>>    uint16_t
>>>    rte_vhost_avail_entries(int vid, uint16_t queue_id)
>>>    {
>>> diff --git a/lib/librte_vhost/vhost.h b/lib/librte_vhost/vhost.h index
>>> e9138dfab..3dace2ec3 100644
>>> --- a/lib/librte_vhost/vhost.h
>>> +++ b/lib/librte_vhost/vhost.h
>>> @@ -128,6 +128,11 @@ struct vhost_virtqueue {
>>>    	/* Physical address of used ring, for logging */
>>>    	uint64_t		log_guest_addr;
>>>
>>> +	/* Inflight share memory info */
>>> +	struct inflight_info	*inflight;
>>> +	struct resubmit_info	*resubmit;
>>> +	uint64_t counter;
>>
>> counter and resubmit are too generic names.
> 
> I'll modify the names in the next version.
> 
>>
>>> +
>>>    	uint16_t		nr_zmbuf;
>>>    	uint16_t		zmbuf_size;
>>>    	uint16_t		last_zmbuf_idx;
>>> @@ -286,6 +291,12 @@ struct guest_page {
>>>    	uint64_t size;
>>>    };
>>>
>>> +struct inflight_mem_info {
>>> +	int fd;
>>> +	void *addr;
>>> +	uint64_t size;
>>> +};
>>> +
>>>    /**
>>>     * Device structure contains all configuration information relating
>>>     * to the device.
>>> @@ -303,6 +314,7 @@ struct virtio_net {
>>>    	uint32_t		nr_vring;
>>>    	int			dequeue_zero_copy;
>>>    	struct vhost_virtqueue	*virtqueue[VHOST_MAX_QUEUE_PAIRS * 2];
>>> +	struct inflight_mem_info inflight_info;
>>>    #define IF_NAME_SZ (PATH_MAX > IFNAMSIZ ? PATH_MAX : IFNAMSIZ)
>>>    	char			ifname[IF_NAME_SZ];
>>>    	uint64_t		log_size;
>>
>> Do you have some code example using these new APIs?
>> It would help for reviewing the patch.
> 
> This patch is only valid for vhost blk.

This does not seem vhost-blk specifics, otherwise the requests names
would reflect that and it should live in spdk only as we have worked
with the spdk team so that they can handle their specific requests 
directly in their lib.

This new feature is for backends handling the ring out-of-order,
including vhost-blk.

> The use of new two messages and the implementation of packed queue are
> described in docs/vhost-user.txt.

Thanks, that's now in QEMU's docs/interop/vhost-user.rst

Regarding packed ring, I understand you don't want to implement it now.
However, the new rte_vhost APIs you introduce need to specify it is
split-ring only in their names, or find a way to to make the API generic
but it seems difficult.

Regarding the feature, I wonder whether it could also be useful in case
of in-order handling to recover from crashes. Maybe not for split ring 
where we have some workarounds since we can get the info from the rings,
but more for packed ring, where the wrap counter values aren't saved
anywhere except the backend.

> I have submitted SPDK-related patches to use new API.

Here is the link:
https://review.gerrithub.io/c/spdk/spdk/+/449418/1
> 
>>
>> Thanks,
>> Maxime
>>
>