[dpdk-dev] [PATCH v3] vhost: support inflight share memory protocol feature

[Tiwei Bie]

On Mon, Apr 29, 2019 at 12:07:05PM +0800, lin li wrote:
> Tiwei Bie <tiwei.bie at intel.com> 于2019年4月28日周日 下午7:18写道：
> > On Fri, Apr 26, 2019 at 05:40:21AM -0400, Li Lin wrote:
[...]
> > > @@ -98,12 +102,26 @@ struct rte_vhost_memory {
> > >       struct rte_vhost_mem_region regions[];
> > >  };
> > >
> > > +typedef struct VhostUserInflightEntry {
> > > +     uint8_t inflight;
> > > +} VhostUserInflightEntry;
> > > +
> > > +typedef struct VhostInflightInfo {
> > > +     uint16_t version;
> > > +     uint16_t last_inflight_io;
> > > +     uint16_t used_idx;
> > > +     VhostUserInflightEntry desc[0];
> > > +} VhostInflightInfo;
> >
> > Is there any details on above structure? Why does it not match
> > QueueRegionSplit or QueueRegionPacked structures described in
> > qemu/docs/interop/vhost-user.txt?
> 
> Qemu have its vhost-user backend，

Do you mean contrib/libvhost-user in QEMU?

> qemu did the submission of IO in it.

What does this mean?

> The implementation of dpdk is more general. It is just to mark inflight entry.

Based on the discussion in below threads:

https://patchwork.kernel.org/patch/10753893/
https://patchwork.kernel.org/patch/10817735/

IIUC, above structure is the extra buffer allocated by slave to
store the information of inflight descriptors and share with master
for persistent. All vhost-user backends' implementation should
follow the vhost-user protocol (including the format of above
structure) defined in qemu/docs/interop/vhost-user.txt. Right?

> The submission of inflight entry is handle over to different backends.
> They have their own ways to handle it, such as spdk.
> So there are some differences in data structure.
> 
> >
[...]
> > > +static int
> > > +vhost_user_set_inflight_fd(struct virtio_net **pdev, VhostUserMsg *msg,
> > > +             int main_fd __rte_unused)
> > > +{
> > > +     int fd, i;
> > > +     uint64_t mmap_size, mmap_offset;
> > > +     uint16_t num_queues, queue_size;
> > > +     uint32_t pervq_inflight_size;
> > > +     void *rc;
> > > +     struct vhost_virtqueue *vq;
> > > +     struct virtio_net *dev = *pdev;
> > > +
> > > +     fd = msg->fds[0];
> > > +     if (msg->size != sizeof(msg->payload.inflight) || fd < 0) {
> > > +             RTE_LOG(ERR, VHOST_CONFIG, "Invalid set_inflight_fd message size is %d,fd is %d\n",
> > > +                     msg->size, fd);
> > > +             return -1;
> >
> > Ditto.
> >
> > > +     }
> > > +
> > > +     mmap_size = msg->payload.inflight.mmap_size;
> > > +     mmap_offset = msg->payload.inflight.mmap_offset;
> > > +     num_queues = msg->payload.inflight.num_queues;
> > > +     queue_size = msg->payload.inflight.queue_size;
> > > +     pervq_inflight_size = get_pervq_shm_size(queue_size);
> > > +
> > > +     RTE_LOG(INFO, VHOST_CONFIG,
> > > +             "set_inflight_fd mmap_size: %lu\n", mmap_size);
> > > +     RTE_LOG(INFO, VHOST_CONFIG,
> > > +             "set_inflight_fd mmap_offset: %lu\n", mmap_offset);
> > > +     RTE_LOG(INFO, VHOST_CONFIG,
> > > +             "set_inflight_fd num_queues: %u\n", num_queues);
> > > +     RTE_LOG(INFO, VHOST_CONFIG,
> > > +             "set_inflight_fd queue_size: %u\n", queue_size);
> > > +     RTE_LOG(INFO, VHOST_CONFIG,
> > > +             "set_inflight_fd fd: %d\n", fd);
> > > +     RTE_LOG(INFO, VHOST_CONFIG,
> > > +             "set_inflight_fd pervq_inflight_size: %d\n",
> > > +             pervq_inflight_size);
> > > +
> > > +     if (dev->inflight_info.addr)
> > > +             munmap(dev->inflight_info.addr, dev->inflight_info.size);
> > > +
> > > +     rc = mmap(0, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
> > > +                     fd, mmap_offset);
> >
> > Why call it rc? Maybe addr is a better name?
> 
> In some scenarios, shared memory is reallocated or resized by qemu, so
> again mmap is needed.

In which case the shared memory will be reallocated or resized by QEMU?

> 
> >
> > > +     if (rc == MAP_FAILED) {
> > > +             RTE_LOG(ERR, VHOST_CONFIG, "failed to mmap share memory.\n");
> > > +             return -1;
> >
> > Should always return RTE_VHOST_MSG_RESULT_* in
> > message handler.
> >
> > > +     }
> > > +
> > > +     if (dev->inflight_info.fd)
> > > +             close(dev->inflight_info.fd);
> > > +
> > > +     dev->inflight_info.fd = fd;
> > > +     dev->inflight_info.addr = rc;
> > > +     dev->inflight_info.size = mmap_size;
> > > +
> > > +     for (i = 0; i < num_queues; i++) {
> > > +             vq = dev->virtqueue[i];
> > > +             vq->inflight = (VhostInflightInfo *)rc;
> > > +             rc = (void *)((char *)rc + pervq_inflight_size);
> > > +     }
> > > +
> > > +     return RTE_VHOST_MSG_RESULT_OK;
> > > +}
[...]
> > > diff --git a/lib/librte_vhost/vhost_user.h b/lib/librte_vhost/vhost_user.h
> > > index 2a650fe4b..35f969b1b 100644
> > > --- a/lib/librte_vhost/vhost_user.h
> > > +++ b/lib/librte_vhost/vhost_user.h
> > > @@ -23,7 +23,8 @@
> > >                                        (1ULL << VHOST_USER_PROTOCOL_F_CRYPTO_SESSION) | \
> > >                                        (1ULL << VHOST_USER_PROTOCOL_F_SLAVE_SEND_FD) | \
> > >                                        (1ULL << VHOST_USER_PROTOCOL_F_HOST_NOTIFIER) | \
> > > -                                      (1ULL << VHOST_USER_PROTOCOL_F_PAGEFAULT))
> > > +                                     (1ULL << VHOST_USER_PROTOCOL_F_PAGEFAULT) | \
> > > +                                     (1ULL << VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD))
> >
> > It will advertise this feature for builtin net and crypto
> > backends. It's probably not what you intended.
> >
> 
> Indeed, this feature is mainly used for spdk-like backends. You mean
> this function is disabled by default？

External backends should use rte_vhost_driver_set_protocol_features()
to advertise the protocol features they support.