Support to run vhost-pmd vdev in the secondary process. We obtain
information, like memory regions, kickfd, callfd, through
primary/secondary communication channel.
And by invoking rte_vhost_set_vring_effective_fd, we can set the
kickfd which can be recognized by the secondary process.
Signed-off-by: Jianfeng Tan <jianfeng....@intel.com>
---
drivers/net/vhost/rte_eth_vhost.c | 200 +++++++++++++++++++++++++++++++++++---
1 file changed, 187 insertions(+), 13 deletions(-)
diff --git a/drivers/net/vhost/rte_eth_vhost.c
b/drivers/net/vhost/rte_eth_vhost.c
index 04179b4..9d296aa 100644
--- a/drivers/net/vhost/rte_eth_vhost.c
+++ b/drivers/net/vhost/rte_eth_vhost.c
@@ -33,6 +33,7 @@
#include <unistd.h>
#include <pthread.h>
#include <stdbool.h>
+#include <sys/mman.h>
#include <rte_mbuf.h>
#include <rte_ethdev.h>
@@ -46,6 +47,20 @@
#include "rte_eth_vhost.h"
+#define VHOST_MSG_TYPE_REGIONS 1
+#define VHOST_MSG_TYPE_SET_FDS 2
+#define VHOST_MSG_TYPE_INIT 3
+
+struct vhost_params {
+ int type;
+ union {
+ int vid;
+ int portid;
+ };
+ int vring_idx;
+ struct rte_vhost_mem_region regions[0];
+};
+
enum {VIRTIO_RXQ, VIRTIO_TXQ, VIRTIO_QNUM};
#define ETH_VHOST_IFACE_ARG "iface"
@@ -550,6 +565,66 @@ update_queuing_status(struct rte_eth_dev *dev)
}
static int
+share_device(int vid)
+{
+ uint32_t i, vring_num;
+ int len;
+ int fds[8];
+ struct rte_vhost_memory *mem;
+ struct vhost_params *params;
+ struct rte_vhost_vring vring;
+
+ /* share mem table */
+ if (rte_vhost_get_mem_table(vid, &mem) < 0) {
+ RTE_LOG(ERR, PMD, "Failed to get mem table\n");
+ return 0;
+ }
+ for (i = 0; i < mem->nregions; ++i)
+ fds[i] = mem->regions[i].fd;
+
+ len = sizeof(struct rte_vhost_mem_region) * mem->nregions;
+ params = malloc(sizeof(*params) + len);
+ if (params == NULL) {
+ RTE_LOG(ERR, PMD, "Failed to allocate memory\n");
+ return -1;
+ }
+
+ params->type = VHOST_MSG_TYPE_REGIONS;
+ params->vid = vid;
+ memcpy(params->regions, mem->regions, len);
+
+ if (rte_eal_mp_sendmsg("vhost pmd", params, sizeof(*params) + len,
+ fds, mem->nregions) < 0) {
+ RTE_LOG(ERR, PMD, "Failed to share mem table\n");
+ free(params);
+ return -1;
+ }
+
+ /* share callfd and kickfd */
+ params->type = VHOST_MSG_TYPE_SET_FDS;
+ vring_num = rte_vhost_get_vring_num(vid);
+ for (i = 0; i < vring_num; i++) {
+ if (rte_vhost_get_vhost_vring(vid, i, &vring) < 0) {
+ RTE_LOG(ERR, PMD, "Failed to get vring, idx = %d\n", i);
+ free(params);
+ return -1;
+ }
+
+ params->vring_idx = i;
+ fds[0] = vring.callfd;
+ fds[1] = vring.kickfd;
+ if (rte_eal_mp_sendmsg("vhost pmd", params,
+ sizeof(*params), fds, 2) < 0) {
+ RTE_LOG(ERR, PMD, "Failed to set fds\n");
+ return -1;
+ }
+ }
+
+ free(params);
+ return 0;
+}
+
+static int
new_device(int vid)
{
struct rte_eth_dev *eth_dev;
@@ -610,6 +685,8 @@ new_device(int vid)
_rte_eth_dev_callback_process(eth_dev, RTE_ETH_EVENT_INTR_LSC,
NULL, NULL);
+ share_device(vid);
+
return 0;
}
@@ -1025,13 +1102,6 @@ eth_dev_vhost_create(struct rte_vdev_device *dev, char
*iface_name,
RTE_LOG(INFO, PMD, "Creating VHOST-USER backend on numa socket %u\n",
numa_node);
- /* now do all data allocation - for eth_dev structure and internal
- * (private) data
- */
- data = rte_zmalloc_socket(name, sizeof(*data), 0, numa_node);
- if (data == NULL)
- goto error;
-
list = rte_zmalloc_socket(name, sizeof(*list), 0, numa_node);
if (list == NULL)
goto error;
@@ -1073,11 +1143,7 @@ eth_dev_vhost_create(struct rte_vdev_device *dev, char
*iface_name,
rte_spinlock_init(&vring_state->lock);
vring_states[eth_dev->data->port_id] = vring_state;
- /* We'll replace the 'data' originally allocated by eth_dev. So the
- * vhost PMD resources won't be shared between multi processes.
- */
- rte_memcpy(data, eth_dev->data, sizeof(*data));
- eth_dev->data = data;
+ data = eth_dev->data;
data->nb_rx_queues = queues;
data->nb_tx_queues = queues;
@@ -1125,6 +1191,30 @@ eth_dev_vhost_create(struct rte_vdev_device *dev, char
*iface_name,
return -1;
}
+static int
+eth_dev_vhost_attach(struct rte_vdev_device *dev)
+{
+ struct rte_eth_dev *eth_dev = NULL;
+ struct rte_eth_dev_data *data = NULL;
+
+ RTE_LOG(INFO, PMD, "Attach vhost user port\n");
+
+ /* reserve an ethdev entry */
+ eth_dev = rte_eth_vdev_allocate(dev, sizeof(struct pmd_internal));
+ if (eth_dev == NULL)
+ return -1;
+
+ eth_dev->dev_ops = &ops;
+
+ /* finally assign rx and tx ops */
+ eth_dev->rx_pkt_burst = eth_vhost_rx;
+ eth_dev->tx_pkt_burst = eth_vhost_tx;
+
+ data = eth_dev->data;
+
+ return data->port_id;
+}
+
static inline int
open_iface(const char *key __rte_unused, const char *value, void *extra_args)
{
@@ -1154,10 +1244,84 @@ open_int(const char *key __rte_unused, const char
*value, void *extra_args)
}
static int
+vhost_pmd_action(const void *params, int len, int fds[], int fds_num)
+{
+ int i;
+ int vid;
+ void *base_addr;
+ const struct vhost_params *p = params;
+ const struct rte_vhost_mem_region *regions;
+
+ if (len < (int)sizeof(*p)) {
+ RTE_LOG(ERR, PMD, "message if too short\n");
+ return -1;
+ }
+
+ switch (p->type) {
+ case VHOST_MSG_TYPE_REGIONS:
+ regions = p->regions;
+ for (i = 0; i < fds_num; ++i) {
+ base_addr = mmap(regions[i].mmap_addr,
+ regions[i].mmap_size,
+ PROT_READ | PROT_WRITE,
+ MAP_FIXED | MAP_SHARED, fds[i], 0);
+ if (base_addr != regions[i].mmap_addr) {
+ RTE_LOG(ERR, PMD,
+ "vhost in secondary mmap error: %s\n",
+ strerror(errno));
+ break;
+ }
+ }
+ break;
+ case VHOST_MSG_TYPE_SET_FDS:
+ rte_vhost_set_vring_effective_fd(p->vid,
+ p->vring_idx,
+ fds[0], fds[1]);
+ break;
+ case VHOST_MSG_TYPE_INIT:
+ vid = rte_eth_vhost_get_vid_from_port_id(p->portid);
+ share_device(vid);
+ break;
+ }
+
+ return 0;
+}
+
+static int
+probe_secondary(struct rte_vdev_device *dev)
+{
+ int portid = eth_dev_vhost_attach(dev);
+ struct rte_eth_dev *eth_dev;
+ struct pmd_internal *internal;
+ struct vhost_params p;
+
+ if (portid < 0)
+ return -1;
+
+ eth_dev = &rte_eth_devices[portid];
+ internal = eth_dev->data->dev_private;
+
+ if (!internal ||
+ rte_atomic32_read(&internal->dev_attached) == 0) {
+ RTE_LOG(INFO, PMD, "%s is not ready\n", dev->device.name);
+ return 0;
+ }
+
+ p.type = VHOST_MSG_TYPE_INIT;
+ p.portid = portid;
+ if (rte_eal_mp_sendmsg("vhost pmd", &p, sizeof(p), NULL, 0) < 0) {
+ RTE_LOG(ERR, PMD, "Failed to send request for init\n");
+ return -1;
+ }
+
+ return 0;
+}
+
+static int
rte_pmd_vhost_probe(struct rte_vdev_device *dev)
{
struct rte_kvargs *kvlist = NULL;
- int ret = 0;
+ int ret;
char *iface_name;
uint16_t queues;
uint64_t flags = 0;
@@ -1167,6 +1331,15 @@ rte_pmd_vhost_probe(struct rte_vdev_device *dev)
RTE_LOG(INFO, PMD, "Initializing pmd_vhost for %s\n",
rte_vdev_device_name(dev));
+ ret = rte_eal_mp_action_register("vhost pmd", vhost_pmd_action);
+ if (ret < 0 && ret != -EEXIST) {
+ RTE_LOG(ERR, PMD, "vhost fails to add action\n");
+ return -1;
+ }
+
+ if (rte_eal_process_type() == RTE_PROC_SECONDARY)
+ return probe_secondary(dev);
+
kvlist = rte_kvargs_parse(rte_vdev_device_args(dev), valid_arguments);
if (kvlist == NULL)
return -1;
@@ -1216,6 +1389,7 @@ rte_pmd_vhost_probe(struct rte_vdev_device *dev)
eth_dev_vhost_create(dev, iface_name, queues, dev->device.numa_node,
flags);
+ ret = 0;
out_free:
rte_kvargs_free(kvlist);
return ret;
--
2.7.4
