RE: [EXT] Re: [PATCH 1/2] net/virtio-user: improve kick performance with notification area mapping

[Srujana Challa <schalla@marvell.com>]

> Hi Srujana,
> 
> Thanks for your contribution!
> Is it possible to provide information on which hardware it can be tested on?
It can be tested on Marvell's Octeon DPU.

> 
> On 12/8/23 06:31, Srujana Challa wrote:
> > This patch introduces new virtio-user callback to map the vq
> > notification area and implements it for the vhost-vDPA backend.
> > This is simply done by using mmap()/munmap() for the vhost-vDPA fd.
> >
> > This patch also adds a parameter for configuring feature bit
> > VIRTIO_NET_F_NOTIFICATION_DATA. If feature is disabled, also
> 
> VIRTIO_F_NOTIFICATION_DATA*
> 
> > update corresponding unsupported feature bit. And also adds code to
> > write to queue notify address in notify callback.
> > This will help in increasing the kick performance.
> 
> Do you have any number about the performance improvement?

We don't have exact comparison with the performance data, we are mainly supporting
notification data with notify area which enables direct data path notifications and
avoids intermediate checks in-between.

> 
> > Signed-off-by: Srujana Challa <schalla@marvell.com>
> > ---
> >   doc/guides/nics/virtio.rst                    |  5 ++
> >   drivers/net/virtio/virtio_user/vhost.h        |  1 +
> >   drivers/net/virtio/virtio_user/vhost_vdpa.c   | 56 ++++++++++++++++++
> >   .../net/virtio/virtio_user/virtio_user_dev.c  | 52 +++++++++++++++--
> >   .../net/virtio/virtio_user/virtio_user_dev.h  |  5 +-
> >   drivers/net/virtio/virtio_user_ethdev.c       | 57 ++++++++++++++++---
> >   6 files changed, 162 insertions(+), 14 deletions(-)
> >
> > diff --git a/doc/guides/nics/virtio.rst b/doc/guides/nics/virtio.rst
> > index c22ce56a02..11dd6359e5 100644
> > --- a/doc/guides/nics/virtio.rst
> > +++ b/doc/guides/nics/virtio.rst
> > @@ -349,6 +349,11 @@ Below devargs are supported by the virtio-user
> vdev:
> >       election.
> >       (Default: 0 (disabled))
> >
> > +#.  ``notification_data``:
> > +
> > +    It is used to enable virtio device notification data feature.
> > +    (Default: 1 (enabled))
> 
> Is there any reason currently to make it configurable?
> 
> As it is enabled by default, I guess we want it to be negociated if the device
> supports it.
> 
> Let's remove the devarg for now, and we can revisit if the need arise?
Ack.

> 
> > +
> >   Virtio paths Selection and Usage
> >   --------------------------------
> >
> > diff --git a/drivers/net/virtio/virtio_user/vhost.h
> > b/drivers/net/virtio/virtio_user/vhost.h
> > index f817cab77a..1bce00c7ac 100644
> > --- a/drivers/net/virtio/virtio_user/vhost.h
> > +++ b/drivers/net/virtio/virtio_user/vhost.h
> > @@ -90,6 +90,7 @@ struct virtio_user_backend_ops {
> >   	int (*server_disconnect)(struct virtio_user_dev *dev);
> >   	int (*server_reconnect)(struct virtio_user_dev *dev);
> >   	int (*get_intr_fd)(struct virtio_user_dev *dev);
> > +	int (*map_notification_area)(struct virtio_user_dev *dev, bool
> map);
> >   };
> >
> >   extern struct virtio_user_backend_ops virtio_ops_user; diff --git
> > a/drivers/net/virtio/virtio_user/vhost_vdpa.c
> > b/drivers/net/virtio/virtio_user/vhost_vdpa.c
> > index 2c36b26224..1eb0f9ec48 100644
> > --- a/drivers/net/virtio/virtio_user/vhost_vdpa.c
> > +++ b/drivers/net/virtio/virtio_user/vhost_vdpa.c
> > @@ -5,6 +5,7 @@
> >   #include <sys/ioctl.h>
> >   #include <sys/types.h>
> >   #include <sys/stat.h>
> > +#include <sys/mman.h>
> >   #include <fcntl.h>
> >   #include <stdlib.h>
> >   #include <unistd.h>
> > @@ -622,6 +623,60 @@ vhost_vdpa_get_intr_fd(struct virtio_user_dev
> *dev __rte_unused)
> >   	return -1;
> >   }
> >
> > +static int
> > +unmap_notification_area(struct virtio_user_dev *dev, int nr_vrings) {
> > +	int i;
> > +
> > +	for (i = 0; i < nr_vrings; i++) {
> > +		if (dev->notify_area[i])
> > +			munmap(dev->notify_area[i], getpagesize());
> > +	}
> > +	free(dev->notify_area);
> > +
> > +	return 0;
> > +}
> > +
> > +static int
> > +vhost_vdpa_map_notification_area(struct virtio_user_dev *dev, bool
> > +map) {
> > +	struct vhost_vdpa_data *data = dev->backend_data;
> > +	int nr_vrings, i, page_size = getpagesize();
> > +	uint16_t **notify_area;
> > +
> > +	nr_vrings = dev->max_queue_pairs * 2;
> > +	if (dev->device_features & (1ull << VIRTIO_NET_F_CTRL_VQ))
> > +		nr_vrings++;
> > +
> > +	if (!map)
> > +		return unmap_notification_area(dev, nr_vrings);
> > +
> > +	notify_area = malloc(nr_vrings * sizeof(*notify_area));
> > +	if (!notify_area) {
> > +		PMD_DRV_LOG(ERR, "(%s) Failed to allocate notify area
> array", dev->path);
> > +		return -1;
> > +	}
> 
> Add new line here.
> 
> > +	for (i = 0; i < nr_vrings; i++) {
> > +		notify_area[i] = mmap(NULL, page_size, PROT_WRITE,
> MAP_SHARED | MAP_FILE,
> > +				      data->vhostfd, i * page_size);
> > +		if (notify_area[i] == MAP_FAILED) {
> > +			PMD_DRV_LOG(ERR, "(%s) Map failed for notify
> address of queue %d\n",
> > +				    dev->path, i);
> > +			goto map_err;
> > +		}
> > +	}
> > +	dev->notify_area = notify_area;
> > +
> > +	return 0;
> > +
> > +map_err:
> > +	i--;
> 
> I would move this before the goto map_err;
> 
> > +	for (; i >= 0; i--)
> > +		munmap(notify_area[i], page_size);
> > +	free(notify_area);
> 
> New line here.
> 
> > +	return -1;
> > +}
> > +
> >   struct virtio_user_backend_ops virtio_ops_vdpa = {
> >   	.setup = vhost_vdpa_setup,
> >   	.destroy = vhost_vdpa_destroy,
> > @@ -646,4 +701,5 @@ struct virtio_user_backend_ops virtio_ops_vdpa = {
> >   	.dma_unmap = vhost_vdpa_dma_unmap_batch,
> >   	.update_link_state = vhost_vdpa_update_link_state,
> >   	.get_intr_fd = vhost_vdpa_get_intr_fd,
> > +	.map_notification_area = vhost_vdpa_map_notification_area,
> 
> I wonder if it wouldn't be cleaner to have one cb for map, one for unmap.
> WDYT?
I will add one more cb for unmap in the next version.

> 
> >   };
> > diff --git a/drivers/net/virtio/virtio_user/virtio_user_dev.c
> > b/drivers/net/virtio/virtio_user/virtio_user_dev.c
> > index af1f8c8237..578877d089 100644
> > --- a/drivers/net/virtio/virtio_user/virtio_user_dev.c
> > +++ b/drivers/net/virtio/virtio_user/virtio_user_dev.c
> > @@ -18,6 +18,7 @@
> >   #include <rte_string_fns.h>
> >   #include <rte_eal_memconfig.h>
> >   #include <rte_malloc.h>
> > +#include <rte_io.h>
> >
> >   #include "vhost.h"
> >   #include "virtio_user_dev.h"
> > @@ -413,6 +414,12 @@ virtio_user_dev_init_notify(struct virtio_user_dev
> *dev)
> >   		dev->callfds[i] = callfd;
> >   		dev->kickfds[i] = kickfd;
> >   	}
> > +	dev->notify_area_mapped = false;
> Can't we just check on dev->notify_area instead of introducing a new field?
> 
> > +	if (dev->ops->map_notification_area) {
> > +		if (dev->ops->map_notification_area(dev, true))
> > +			goto err;
> > +		dev->notify_area_mapped = true;
> > +	}
> >
> >   	return 0;
> >   err:
> > @@ -445,6 +452,11 @@ virtio_user_dev_uninit_notify(struct
> virtio_user_dev *dev)
> >   			dev->callfds[i] = -1;
> >   		}
> >   	}
> > +	if (dev->ops->map_notification_area && dev-
> >notify_area_mapped) {
> > +		/* Unmap notification area */
> > +		dev->ops->map_notification_area(dev, false);
> > +		dev->notify_area_mapped = false;
> > +	}
> >   }
> >
> >   static int
> > @@ -674,12 +686,14 @@ virtio_user_free_vrings(struct virtio_user_dev
> *dev)
> >   	 1ULL << VIRTIO_NET_F_GUEST_TSO6	|	\
> >   	 1ULL << VIRTIO_F_IN_ORDER		|	\
> >   	 1ULL << VIRTIO_F_VERSION_1		|	\
> > -	 1ULL << VIRTIO_F_RING_PACKED)
> > +	 1ULL << VIRTIO_F_RING_PACKED		|	\
> > +	 1ULL << VIRTIO_F_NOTIFICATION_DATA)
> >
> >   int
> >   virtio_user_dev_init(struct virtio_user_dev *dev, char *path, uint16_t
> queues,
> >   		     int cq, int queue_size, const char *mac, char **ifname,
> >   		     int server, int mrg_rxbuf, int in_order, int packed_vq,
> > +		     int notification_data,
> >   		     enum virtio_user_backend_type backend_type)
> >   {
> >   	uint64_t backend_features;
> > @@ -737,6 +751,9 @@ virtio_user_dev_init(struct virtio_user_dev *dev,
> char *path, uint16_t queues,
> >   	if (!packed_vq)
> >   		dev->unsupported_features |= (1ull <<
> VIRTIO_F_RING_PACKED);
> >
> > +	if (!notification_data)
> > +		dev->unsupported_features |= (1ull <<
> VIRTIO_F_NOTIFICATION_DATA);
> > +
> >   	if (dev->mac_specified)
> >   		dev->frontend_features |= (1ull << VIRTIO_NET_F_MAC);
> >   	else
> > @@ -1039,11 +1056,35 @@ static void
> >   virtio_user_control_queue_notify(struct virtqueue *vq, void *cookie)
> >   {
> >   	struct virtio_user_dev *dev = cookie;
> > -	uint64_t buf = 1;
> > +	uint64_t notify_data = 1;
> > +
> > +	if (!dev->notify_area_mapped) {
> > +		if (write(dev->kickfds[vq->vq_queue_index], &notify_data,
> sizeof(notify_data)) < 0)
> > +			PMD_DRV_LOG(ERR, "failed to kick backend: %s",
> > +				    strerror(errno));
> > +		return;
> > +	} else if (!virtio_with_feature(&dev->hw,
> VIRTIO_F_NOTIFICATION_DATA)) {
> > +		rte_write16(vq->vq_queue_index, vq->notify_addr);
> > +		return;
> > +	}
> >
> > -	if (write(dev->kickfds[vq->vq_queue_index], &buf, sizeof(buf)) < 0)
> > -		PMD_DRV_LOG(ERR, "failed to kick backend: %s",
> > -			    strerror(errno));
> > +	if (virtio_with_packed_queue(&dev->hw)) {
> > +		/* Bit[0:15]: vq queue index
> > +		 * Bit[16:30]: avail index
> > +		 * Bit[31]: avail wrap counter
> > +		 */
> > +		notify_data = ((uint32_t)(!!(vq->vq_packed.cached_flags &
> > +				VRING_PACKED_DESC_F_AVAIL)) << 31) |
> > +				((uint32_t)vq->vq_avail_idx << 16) |
> > +				vq->vq_queue_index;
> > +	} else {
> > +		/* Bit[0:15]: vq queue index
> > +		 * Bit[16:31]: avail index
> > +		 */
> > +		notify_data = ((uint32_t)vq->vq_avail_idx << 16) |
> > +				vq->vq_queue_index;
> > +	}
> > +	rte_write32(notify_data, vq->notify_addr);
> >   }
> >
> >   int
> > @@ -1062,6 +1103,7 @@ virtio_user_dev_create_shadow_cvq(struct
> > virtio_user_dev *dev, struct virtqueue
> >
> >   	scvq->cq.notify_queue = &virtio_user_control_queue_notify;
> >   	scvq->cq.notify_cookie = dev;
> > +	scvq->notify_addr = vq->notify_addr;
> >   	dev->scvq = scvq;
> >
> >   	return 0;
> > diff --git a/drivers/net/virtio/virtio_user/virtio_user_dev.h
> > b/drivers/net/virtio/virtio_user/virtio_user_dev.h
> > index 7323d88302..29ec386da5 100644
> > --- a/drivers/net/virtio/virtio_user/virtio_user_dev.h
> > +++ b/drivers/net/virtio/virtio_user/virtio_user_dev.h
> > @@ -64,6 +64,9 @@ struct virtio_user_dev {
> >   	struct virtqueue	*scvq;
> >
> >   	void *backend_data;
> > +
> > +	bool notify_area_mapped;
> > +	uint16_t **notify_area;
> >   };
> >
> >   int virtio_user_dev_set_features(struct virtio_user_dev *dev); @@
> > -72,7 +75,7 @@ int virtio_user_stop_device(struct virtio_user_dev *dev);
> >   int virtio_user_dev_init(struct virtio_user_dev *dev, char *path, uint16_t
> queues,
> >   			 int cq, int queue_size, const char *mac, char
> **ifname,
> >   			 int server, int mrg_rxbuf, int in_order,
> > -			 int packed_vq,
> > +			 int packed_vq, int notification_data,
> >   			 enum virtio_user_backend_type backend_type);
> >   void virtio_user_dev_uninit(struct virtio_user_dev *dev);
> >   void virtio_user_handle_cq(struct virtio_user_dev *dev, uint16_t
> > queue_idx); diff --git a/drivers/net/virtio/virtio_user_ethdev.c
> > b/drivers/net/virtio/virtio_user_ethdev.c
> > index 3a31642899..241465ecdd 100644
> > --- a/drivers/net/virtio/virtio_user_ethdev.c
> > +++ b/drivers/net/virtio/virtio_user_ethdev.c
> > @@ -18,6 +18,7 @@
> >   #include <bus_vdev_driver.h>
> >   #include <rte_alarm.h>
> >   #include <rte_cycles.h>
> > +#include <rte_io.h>
> >
> >   #include "virtio_ethdev.h"
> >   #include "virtio_logs.h"
> > @@ -232,6 +233,7 @@ virtio_user_setup_queue(struct virtio_hw *hw,
> struct virtqueue *vq)
> >   	else
> >   		virtio_user_setup_queue_split(vq, dev);
> >
> > +	vq->notify_addr = dev->notify_area[vq->vq_queue_index];
> >   	if (dev->hw_cvq && hw->cvq && (virtnet_cq_to_vq(hw->cvq) ==
> vq))
> >   		return virtio_user_dev_create_shadow_cvq(dev, vq);
> >
> > @@ -262,8 +264,8 @@ virtio_user_del_queue(struct virtio_hw *hw, struct
> virtqueue *vq)
> >   static void
> >   virtio_user_notify_queue(struct virtio_hw *hw, struct virtqueue *vq)
> >   {
> > -	uint64_t buf = 1;
> >   	struct virtio_user_dev *dev = virtio_user_get_dev(hw);
> > +	uint64_t notify_data = 1;
> >
> >   	if (hw->cvq && (virtnet_cq_to_vq(hw->cvq) == vq)) {
> >   		virtio_user_handle_cq(dev, vq->vq_queue_index); @@ -
> 271,9 +273,34
> > @@ virtio_user_notify_queue(struct virtio_hw *hw, struct virtqueue *vq)
> >   		return;
> >   	}
> >
> > -	if (write(dev->kickfds[vq->vq_queue_index], &buf, sizeof(buf)) < 0)
> > -		PMD_DRV_LOG(ERR, "failed to kick backend: %s",
> > -			    strerror(errno));
> > +	if (!dev->notify_area_mapped) {
> > +		if (write(dev->kickfds[vq->vq_queue_index], &notify_data,
> > +			  sizeof(notify_data)) < 0)
> > +			PMD_DRV_LOG(ERR, "failed to kick backend: %s",
> > +				    strerror(errno));
> > +		return;
> > +	} else if (!virtio_with_feature(hw, VIRTIO_F_NOTIFICATION_DATA))
> {
> > +		rte_write16(vq->vq_queue_index, vq->notify_addr);
> > +		return;
> > +	}
> > +
> > +	if (virtio_with_packed_queue(hw)) {
> > +		/* Bit[0:15]: vq queue index
> > +		 * Bit[16:30]: avail index
> > +		 * Bit[31]: avail wrap counter
> > +		 */
> > +		notify_data = ((uint32_t)(!!(vq->vq_packed.cached_flags &
> > +				VRING_PACKED_DESC_F_AVAIL)) << 31) |
> > +				((uint32_t)vq->vq_avail_idx << 16) |
> > +				vq->vq_queue_index;
> > +	} else {
> > +		/* Bit[0:15]: vq queue index
> > +		 * Bit[16:31]: avail index
> > +		 */
> > +		notify_data = ((uint32_t)vq->vq_avail_idx << 16) |
> > +				vq->vq_queue_index;
> > +	}
> > +	rte_write32(notify_data, vq->notify_addr);
> >   }
> >
> >   static int
> > @@ -329,6 +356,8 @@ static const char *valid_args[] = {
> >   	VIRTIO_USER_ARG_SPEED,
> >   #define VIRTIO_USER_ARG_VECTORIZED     "vectorized"
> >   	VIRTIO_USER_ARG_VECTORIZED,
> > +#define VIRTIO_USER_ARG_NOTIFICATION_DATA  "notification_data"
> > +	VIRTIO_USER_ARG_NOTIFICATION_DATA,
> >   	NULL
> >   };
> >
> > @@ -480,6 +509,7 @@ virtio_user_pmd_probe(struct rte_vdev_device
> *vdev)
> >   	uint64_t in_order = 1;
> >   	uint64_t packed_vq = 0;
> >   	uint64_t vectorized = 0;
> > +	uint64_t notification_data = 1;
> >   	char *path = NULL;
> >   	char *ifname = NULL;
> >   	char *mac_addr = NULL;
> > @@ -637,6 +667,15 @@ virtio_user_pmd_probe(struct rte_vdev_device
> *vdev)
> >   		}
> >   	}
> >
> > +	if (rte_kvargs_count(kvlist,
> VIRTIO_USER_ARG_NOTIFICATION_DATA) == 1) {
> > +		if (rte_kvargs_process(kvlist,
> VIRTIO_USER_ARG_NOTIFICATION_DATA,
> > +				       &get_integer_arg, &notification_data) <
> 0) {
> > +			PMD_INIT_LOG(ERR, "error to parse %s",
> > +
> VIRTIO_USER_ARG_NOTIFICATION_DATA);
> > +			goto end;
> > +		}
> > +	}
> > +
> >   	eth_dev = virtio_user_eth_dev_alloc(vdev);
> >   	if (!eth_dev) {
> >   		PMD_INIT_LOG(ERR, "virtio_user fails to alloc device"); @@ -
> 645,9
> > +684,10 @@ virtio_user_pmd_probe(struct rte_vdev_device *vdev)
> >
> >   	dev = eth_dev->data->dev_private;
> >   	hw = &dev->hw;
> > -	if (virtio_user_dev_init(dev, path, (uint16_t)queues, cq,
> > -			 queue_size, mac_addr, &ifname, server_mode,
> > -			 mrg_rxbuf, in_order, packed_vq, backend_type) <
> 0) {
> > +	if (virtio_user_dev_init(dev, path, (uint16_t)queues, cq, queue_size,
> > +				 mac_addr, &ifname, server_mode,
> mrg_rxbuf,
> > +				 in_order, packed_vq, notification_data,
> > +				 backend_type) < 0) {
> >   		PMD_INIT_LOG(ERR, "virtio_user_dev_init fails");
> >   		virtio_user_eth_dev_free(eth_dev);
> >   		goto end;
> > @@ -784,4 +824,5 @@
> RTE_PMD_REGISTER_PARAM_STRING(net_virtio_user,
> >   	"in_order=<0|1> "
> >   	"packed_vq=<0|1> "
> >   	"speed=<int> "
> > -	"vectorized=<0|1>");
> > +	"vectorized=<0|1> "
> > +	"notification_data=<0|1> ");

Thanks for the review and feedback. I will send the next version with suggested changes.

Thanks,
Srujana.