diff options
Diffstat (limited to 'vnet/vnet/devices/virtio/vhost-user.c')
| -rw-r--r-- | vnet/vnet/devices/virtio/vhost-user.c | 2391 |
1 files changed, 1313 insertions, 1078 deletions
diff --git a/vnet/vnet/devices/virtio/vhost-user.c b/vnet/vnet/devices/virtio/vhost-user.c index dca60f14c67..51b0c409d2a 100644 --- a/vnet/vnet/devices/virtio/vhost-user.c +++ b/vnet/vnet/devices/virtio/vhost-user.c @@ -1,4 +1,4 @@ -/* +/* *------------------------------------------------------------------ * vhost.c - vhost-user * @@ -66,16 +66,17 @@ vlib_node_registration_t vhost_user_input_node; _(PKT_DROP_NOBUF, "tx packet drops (no available descriptors)") \ _(MMAP_FAIL, "mmap failure") -typedef enum { +typedef enum +{ #define _(f,s) VHOST_USER_TX_FUNC_ERROR_##f, foreach_vhost_user_tx_func_error #undef _ - VHOST_USER_TX_FUNC_N_ERROR, + VHOST_USER_TX_FUNC_N_ERROR, } vhost_user_tx_func_error_t; -static char * vhost_user_tx_func_error_strings[] = { +static char *vhost_user_tx_func_error_strings[] = { #define _(n,s) s, - foreach_vhost_user_tx_func_error + foreach_vhost_user_tx_func_error #undef _ }; @@ -85,19 +86,21 @@ static char * vhost_user_tx_func_error_strings[] = { _(MMAP_FAIL, "mmap failure") \ _(UNDERSIZED_FRAME, "undersized ethernet frame received (< 14 bytes)") -typedef enum { +typedef enum +{ #define _(f,s) VHOST_USER_INPUT_FUNC_ERROR_##f, foreach_vhost_user_input_func_error #undef _ - VHOST_USER_INPUT_FUNC_N_ERROR, + VHOST_USER_INPUT_FUNC_N_ERROR, } vhost_user_input_func_error_t; -static char * vhost_user_input_func_error_strings[] = { +static char *vhost_user_input_func_error_strings[] = { #define _(n,s) s, - foreach_vhost_user_input_func_error + foreach_vhost_user_input_func_error #undef _ }; +/* *INDENT-OFF* */ static vhost_user_main_t vhost_user_main = { .mtu_bytes = 1518, }; @@ -105,12 +108,14 @@ static vhost_user_main_t vhost_user_main = { VNET_HW_INTERFACE_CLASS (vhost_interface_class, static) = { .name = "vhost-user", }; +/* *INDENT-ON* */ -static u8 * format_vhost_user_interface_name (u8 * s, va_list * args) +static u8 * +format_vhost_user_interface_name (u8 * s, va_list * args) { u32 i = va_arg (*args, u32); u32 show_dev_instance = ~0; - vhost_user_main_t * vum = &vhost_user_main; + vhost_user_main_t *vum = &vhost_user_main; if (i < vec_len (vum->show_dev_instance_by_real_dev_instance)) show_dev_instance = vum->show_dev_instance_by_real_dev_instance[i]; @@ -122,141 +127,168 @@ static u8 * format_vhost_user_interface_name (u8 * s, va_list * args) return s; } -static int vhost_user_name_renumber (vnet_hw_interface_t * hi, - u32 new_dev_instance) +static int +vhost_user_name_renumber (vnet_hw_interface_t * hi, u32 new_dev_instance) { - vhost_user_main_t * vum = &vhost_user_main; + vhost_user_main_t *vum = &vhost_user_main; vec_validate_init_empty (vum->show_dev_instance_by_real_dev_instance, - hi->dev_instance, ~0); + hi->dev_instance, ~0); - vum->show_dev_instance_by_real_dev_instance [hi->dev_instance] = + vum->show_dev_instance_by_real_dev_instance[hi->dev_instance] = new_dev_instance; - DBG_SOCK("renumbered vhost-user interface dev_instance %d to %d", - hi->dev_instance, new_dev_instance); + DBG_SOCK ("renumbered vhost-user interface dev_instance %d to %d", + hi->dev_instance, new_dev_instance); return 0; } -static inline void * map_guest_mem(vhost_user_intf_t * vui, uword addr) +static inline void * +map_guest_mem (vhost_user_intf_t * vui, uword addr) { int i; - for (i=0; i<vui->nregions; i++) { - if ((vui->regions[i].guest_phys_addr <= addr) && - ((vui->regions[i].guest_phys_addr + vui->regions[i].memory_size) > addr)) { - return (void *) (vui->region_mmap_addr[i] + addr - vui->regions[i].guest_phys_addr); - } - } - DBG_VQ("failed to map guest mem addr %llx", addr); + for (i = 0; i < vui->nregions; i++) + { + if ((vui->regions[i].guest_phys_addr <= addr) && + ((vui->regions[i].guest_phys_addr + vui->regions[i].memory_size) > + addr)) + { + return (void *) (vui->region_mmap_addr[i] + addr - + vui->regions[i].guest_phys_addr); + } + } + DBG_VQ ("failed to map guest mem addr %llx", addr); return 0; } -static inline void * map_user_mem(vhost_user_intf_t * vui, uword addr) +static inline void * +map_user_mem (vhost_user_intf_t * vui, uword addr) { int i; - for (i=0; i<vui->nregions; i++) { - if ((vui->regions[i].userspace_addr <= addr) && - ((vui->regions[i].userspace_addr + vui->regions[i].memory_size) > addr)) { - return (void *) (vui->region_mmap_addr[i] + addr - vui->regions[i].userspace_addr); - } - } + for (i = 0; i < vui->nregions; i++) + { + if ((vui->regions[i].userspace_addr <= addr) && + ((vui->regions[i].userspace_addr + vui->regions[i].memory_size) > + addr)) + { + return (void *) (vui->region_mmap_addr[i] + addr - + vui->regions[i].userspace_addr); + } + } return 0; } -static long get_huge_page_size(int fd) +static long +get_huge_page_size (int fd) { struct statfs s; - fstatfs(fd, &s); + fstatfs (fd, &s); return s.f_bsize; } -static void unmap_all_mem_regions(vhost_user_intf_t * vui) +static void +unmap_all_mem_regions (vhost_user_intf_t * vui) { - int i,r; - for (i=0; i<vui->nregions; i++) { - if (vui->region_mmap_addr[i] != (void *) -1) { + int i, r; + for (i = 0; i < vui->nregions; i++) + { + if (vui->region_mmap_addr[i] != (void *) -1) + { - long page_sz = get_huge_page_size(vui->region_mmap_fd[i]); + long page_sz = get_huge_page_size (vui->region_mmap_fd[i]); - ssize_t map_sz = (vui->regions[i].memory_size + - vui->regions[i].mmap_offset + page_sz) & ~(page_sz - 1); + ssize_t map_sz = (vui->regions[i].memory_size + + vui->regions[i].mmap_offset + + page_sz) & ~(page_sz - 1); - r = munmap(vui->region_mmap_addr[i] - vui->regions[i].mmap_offset, map_sz); + r = + munmap (vui->region_mmap_addr[i] - vui->regions[i].mmap_offset, + map_sz); - DBG_SOCK("unmap memory region %d addr 0x%lx len 0x%lx page_sz 0x%x", i, - vui->region_mmap_addr[i], map_sz, page_sz); + DBG_SOCK + ("unmap memory region %d addr 0x%lx len 0x%lx page_sz 0x%x", i, + vui->region_mmap_addr[i], map_sz, page_sz); - vui->region_mmap_addr[i]= (void *) -1; + vui->region_mmap_addr[i] = (void *) -1; - if (r == -1) { - clib_warning("failed to unmap memory region (errno %d)", errno); - } - close(vui->region_mmap_fd[i]); + if (r == -1) + { + clib_warning ("failed to unmap memory region (errno %d)", + errno); + } + close (vui->region_mmap_fd[i]); + } } - } vui->nregions = 0; } -static clib_error_t * vhost_user_callfd_read_ready (unix_file_t * uf) +static clib_error_t * +vhost_user_callfd_read_ready (unix_file_t * uf) { - __attribute__((unused)) int n; + __attribute__ ((unused)) int n; u8 buff[8]; - n = read(uf->file_descriptor, ((char*)&buff), 8); + n = read (uf->file_descriptor, ((char *) &buff), 8); return 0; } -static inline void vhost_user_if_disconnect(vhost_user_intf_t * vui) +static inline void +vhost_user_if_disconnect (vhost_user_intf_t * vui) { - vhost_user_main_t * vum = &vhost_user_main; - vnet_main_t * vnm = vnet_get_main(); + vhost_user_main_t *vum = &vhost_user_main; + vnet_main_t *vnm = vnet_get_main (); int q; - vnet_hw_interface_set_flags (vnm, vui->hw_if_index, 0); + vnet_hw_interface_set_flags (vnm, vui->hw_if_index, 0); - if (vui->unix_file_index != ~0) { + if (vui->unix_file_index != ~0) + { unix_file_del (&unix_main, unix_main.file_pool + vui->unix_file_index); vui->unix_file_index = ~0; - } + } - hash_unset(vum->vhost_user_interface_index_by_sock_fd, vui->unix_fd); - hash_unset(vum->vhost_user_interface_index_by_listener_fd, vui->unix_fd); - close(vui->unix_fd); + hash_unset (vum->vhost_user_interface_index_by_sock_fd, vui->unix_fd); + hash_unset (vum->vhost_user_interface_index_by_listener_fd, vui->unix_fd); + close (vui->unix_fd); vui->unix_fd = -1; vui->is_up = 0; - for (q = 0; q < vui->num_vrings; q++) { - vui->vrings[q].desc = NULL; - vui->vrings[q].avail = NULL; - vui->vrings[q].used = NULL; - vui->vrings[q].log_guest_addr = 0; - vui->vrings[q].log_used = 0; - } + for (q = 0; q < vui->num_vrings; q++) + { + vui->vrings[q].desc = NULL; + vui->vrings[q].avail = NULL; + vui->vrings[q].used = NULL; + vui->vrings[q].log_guest_addr = 0; + vui->vrings[q].log_used = 0; + } - unmap_all_mem_regions(vui); - DBG_SOCK("interface ifindex %d disconnected", vui->sw_if_index); + unmap_all_mem_regions (vui); + DBG_SOCK ("interface ifindex %d disconnected", vui->sw_if_index); } #define VHOST_LOG_PAGE 0x1000 -always_inline void vhost_user_log_dirty_pages(vhost_user_intf_t * vui, - u64 addr, u64 len) +always_inline void +vhost_user_log_dirty_pages (vhost_user_intf_t * vui, u64 addr, u64 len) { - if (PREDICT_TRUE(vui->log_base_addr == 0 - || !(vui->features & (1 << FEAT_VHOST_F_LOG_ALL)))) { - return; - } - if (PREDICT_FALSE((addr + len - 1) / VHOST_LOG_PAGE / 8 >= vui->log_size)) { - DBG_SOCK("vhost_user_log_dirty_pages(): out of range\n"); - return; - } + if (PREDICT_TRUE (vui->log_base_addr == 0 + || !(vui->features & (1 << FEAT_VHOST_F_LOG_ALL)))) + { + return; + } + if (PREDICT_FALSE ((addr + len - 1) / VHOST_LOG_PAGE / 8 >= vui->log_size)) + { + DBG_SOCK ("vhost_user_log_dirty_pages(): out of range\n"); + return; + } - CLIB_MEMORY_BARRIER(); + CLIB_MEMORY_BARRIER (); u64 page = addr / VHOST_LOG_PAGE; - while (page * VHOST_LOG_PAGE < addr + len) { - ((u8*)vui->log_base_addr)[page / 8] |= 1 << page % 8; - page++; - } + while (page * VHOST_LOG_PAGE < addr + len) + { + ((u8 *) vui->log_base_addr)[page / 8] |= 1 << page % 8; + page++; + } } #define vhost_user_log_dirty_ring(vui, vq, member) \ @@ -265,7 +297,8 @@ always_inline void vhost_user_log_dirty_pages(vhost_user_intf_t * vui, sizeof(vq->used->member)); \ } -static clib_error_t * vhost_user_socket_read (unix_file_t * uf) +static clib_error_t * +vhost_user_socket_read (unix_file_t * uf) { int n, i; int fd, number_of_fds = 0; @@ -273,30 +306,30 @@ static clib_error_t * vhost_user_socket_read (unix_file_t * uf) vhost_user_msg_t msg; struct msghdr mh; struct iovec iov[1]; - vhost_user_main_t * vum = &vhost_user_main; - vhost_user_intf_t * vui; + vhost_user_main_t *vum = &vhost_user_main; + vhost_user_intf_t *vui; struct cmsghdr *cmsg; - uword * p; + uword *p; u8 q; - unix_file_t template = {0}; - vnet_main_t * vnm = vnet_get_main(); + unix_file_t template = { 0 }; + vnet_main_t *vnm = vnet_get_main (); p = hash_get (vum->vhost_user_interface_index_by_sock_fd, - uf->file_descriptor); - if (p == 0) { - DBG_SOCK ("FD %d doesn't belong to any interface", - uf->file_descriptor); + uf->file_descriptor); + if (p == 0) + { + DBG_SOCK ("FD %d doesn't belong to any interface", uf->file_descriptor); return 0; } else vui = vec_elt_at_index (vum->vhost_user_interfaces, p[0]); - char control[CMSG_SPACE(VHOST_MEMORY_MAX_NREGIONS * sizeof(int))]; + char control[CMSG_SPACE (VHOST_MEMORY_MAX_NREGIONS * sizeof (int))]; - memset(&mh, 0, sizeof(mh)); - memset(control, 0, sizeof(control)); + memset (&mh, 0, sizeof (mh)); + memset (control, 0, sizeof (control)); - for (i=0; i < VHOST_MEMORY_MAX_NREGIONS; i++) + for (i = 0; i < VHOST_MEMORY_MAX_NREGIONS; i++) fds[i] = -1; /* set the payload */ @@ -306,420 +339,445 @@ static clib_error_t * vhost_user_socket_read (unix_file_t * uf) mh.msg_iov = iov; mh.msg_iovlen = 1; mh.msg_control = control; - mh.msg_controllen = sizeof(control); + mh.msg_controllen = sizeof (control); - n = recvmsg(uf->file_descriptor, &mh, 0); + n = recvmsg (uf->file_descriptor, &mh, 0); if (n != VHOST_USER_MSG_HDR_SZ) goto close_socket; - if (mh.msg_flags & MSG_CTRUNC) { - goto close_socket; - } + if (mh.msg_flags & MSG_CTRUNC) + { + goto close_socket; + } - cmsg = CMSG_FIRSTHDR(&mh); + cmsg = CMSG_FIRSTHDR (&mh); if (cmsg && (cmsg->cmsg_len > 0) && (cmsg->cmsg_level == SOL_SOCKET) && (cmsg->cmsg_type == SCM_RIGHTS) && - (cmsg->cmsg_len - CMSG_LEN(0) <= VHOST_MEMORY_MAX_NREGIONS * sizeof(int))) { - number_of_fds = (cmsg->cmsg_len - CMSG_LEN(0)) / sizeof(int); - clib_memcpy(fds, CMSG_DATA(cmsg), number_of_fds * sizeof(int)); - } + (cmsg->cmsg_len - CMSG_LEN (0) <= + VHOST_MEMORY_MAX_NREGIONS * sizeof (int))) + { + number_of_fds = (cmsg->cmsg_len - CMSG_LEN (0)) / sizeof (int); + clib_memcpy (fds, CMSG_DATA (cmsg), number_of_fds * sizeof (int)); + } - /* version 1, no reply bit set*/ - if ((msg.flags & 7) != 1) { - DBG_SOCK("malformed message received. closing socket"); - goto close_socket; - } + /* version 1, no reply bit set */ + if ((msg.flags & 7) != 1) + { + DBG_SOCK ("malformed message received. closing socket"); + goto close_socket; + } { - int rv __attribute__((unused)); - /* $$$$ pay attention to rv */ - rv = read(uf->file_descriptor, ((char*)&msg) + n, msg.size); + int rv __attribute__ ((unused)); + /* $$$$ pay attention to rv */ + rv = read (uf->file_descriptor, ((char *) &msg) + n, msg.size); } - switch (msg.request) { + switch (msg.request) + { case VHOST_USER_GET_FEATURES: - DBG_SOCK("if %d msg VHOST_USER_GET_FEATURES", - vui->hw_if_index); + DBG_SOCK ("if %d msg VHOST_USER_GET_FEATURES", vui->hw_if_index); msg.flags |= 4; msg.u64 = (1 << FEAT_VIRTIO_NET_F_MRG_RXBUF) | - (1 << FEAT_VIRTIO_F_ANY_LAYOUT) | - (1 << FEAT_VHOST_F_LOG_ALL) | - (1 << FEAT_VIRTIO_NET_F_GUEST_ANNOUNCE) | - (1 << FEAT_VHOST_USER_F_PROTOCOL_FEATURES); + (1 << FEAT_VIRTIO_F_ANY_LAYOUT) | + (1 << FEAT_VHOST_F_LOG_ALL) | + (1 << FEAT_VIRTIO_NET_F_GUEST_ANNOUNCE) | + (1 << FEAT_VHOST_USER_F_PROTOCOL_FEATURES); msg.u64 &= vui->feature_mask; - msg.size = sizeof(msg.u64); + msg.size = sizeof (msg.u64); break; case VHOST_USER_SET_FEATURES: - DBG_SOCK("if %d msg VHOST_USER_SET_FEATURES features 0x%016llx", - vui->hw_if_index, msg.u64); + DBG_SOCK ("if %d msg VHOST_USER_SET_FEATURES features 0x%016llx", + vui->hw_if_index, msg.u64); vui->features = msg.u64; if (vui->features & (1 << FEAT_VIRTIO_NET_F_MRG_RXBUF)) - vui->virtio_net_hdr_sz = 12; + vui->virtio_net_hdr_sz = 12; else - vui->virtio_net_hdr_sz = 10; + vui->virtio_net_hdr_sz = 10; - vui->is_any_layout = (vui->features & (1 << FEAT_VIRTIO_F_ANY_LAYOUT)) ? 1 : 0; + vui->is_any_layout = + (vui->features & (1 << FEAT_VIRTIO_F_ANY_LAYOUT)) ? 1 : 0; ASSERT (vui->virtio_net_hdr_sz < VLIB_BUFFER_PRE_DATA_SIZE); - vnet_hw_interface_set_flags (vnm, vui->hw_if_index, 0); + vnet_hw_interface_set_flags (vnm, vui->hw_if_index, 0); vui->is_up = 0; - for (q = 0; q < 2; q++) { - vui->vrings[q].desc = 0; - vui->vrings[q].avail = 0; - vui->vrings[q].used = 0; - vui->vrings[q].log_guest_addr = 0; - vui->vrings[q].log_used = 0; - } + for (q = 0; q < 2; q++) + { + vui->vrings[q].desc = 0; + vui->vrings[q].avail = 0; + vui->vrings[q].used = 0; + vui->vrings[q].log_guest_addr = 0; + vui->vrings[q].log_used = 0; + } - DBG_SOCK("interface %d disconnected", vui->sw_if_index); + DBG_SOCK ("interface %d disconnected", vui->sw_if_index); break; case VHOST_USER_SET_MEM_TABLE: - DBG_SOCK("if %d msg VHOST_USER_SET_MEM_TABLE nregions %d", - vui->hw_if_index, msg.memory.nregions); + DBG_SOCK ("if %d msg VHOST_USER_SET_MEM_TABLE nregions %d", + vui->hw_if_index, msg.memory.nregions); if ((msg.memory.nregions < 1) || - (msg.memory.nregions > VHOST_MEMORY_MAX_NREGIONS)) { - - DBG_SOCK("number of mem regions must be between 1 and %i", - VHOST_MEMORY_MAX_NREGIONS); - - goto close_socket; - } - - if (msg.memory.nregions != number_of_fds) { - DBG_SOCK("each memory region must have FD"); - goto close_socket; - } - unmap_all_mem_regions(vui); - for(i=0; i < msg.memory.nregions; i++) { - clib_memcpy(&(vui->regions[i]), &msg.memory.regions[i], - sizeof(vhost_user_memory_region_t)); - - long page_sz = get_huge_page_size(fds[i]); - - /* align size to 2M page */ - ssize_t map_sz = (vui->regions[i].memory_size + - vui->regions[i].mmap_offset + page_sz) & ~(page_sz - 1); - - vui->region_mmap_addr[i] = mmap(0, map_sz, PROT_READ | PROT_WRITE, - MAP_SHARED, fds[i], 0); - - DBG_SOCK("map memory region %d addr 0 len 0x%lx fd %d mapped 0x%lx " - "page_sz 0x%x", i, map_sz, fds[i], vui->region_mmap_addr[i], page_sz); - - if (vui->region_mmap_addr[i] == MAP_FAILED) { - clib_warning("failed to map memory. errno is %d", errno); - goto close_socket; - } - vui->region_mmap_addr[i] += vui->regions[i].mmap_offset; - vui->region_mmap_fd[i] = fds[i]; - } + (msg.memory.nregions > VHOST_MEMORY_MAX_NREGIONS)) + { + + DBG_SOCK ("number of mem regions must be between 1 and %i", + VHOST_MEMORY_MAX_NREGIONS); + + goto close_socket; + } + + if (msg.memory.nregions != number_of_fds) + { + DBG_SOCK ("each memory region must have FD"); + goto close_socket; + } + unmap_all_mem_regions (vui); + for (i = 0; i < msg.memory.nregions; i++) + { + clib_memcpy (&(vui->regions[i]), &msg.memory.regions[i], + sizeof (vhost_user_memory_region_t)); + + long page_sz = get_huge_page_size (fds[i]); + + /* align size to 2M page */ + ssize_t map_sz = (vui->regions[i].memory_size + + vui->regions[i].mmap_offset + + page_sz) & ~(page_sz - 1); + + vui->region_mmap_addr[i] = mmap (0, map_sz, PROT_READ | PROT_WRITE, + MAP_SHARED, fds[i], 0); + + DBG_SOCK + ("map memory region %d addr 0 len 0x%lx fd %d mapped 0x%lx " + "page_sz 0x%x", i, map_sz, fds[i], vui->region_mmap_addr[i], + page_sz); + + if (vui->region_mmap_addr[i] == MAP_FAILED) + { + clib_warning ("failed to map memory. errno is %d", errno); + goto close_socket; + } + vui->region_mmap_addr[i] += vui->regions[i].mmap_offset; + vui->region_mmap_fd[i] = fds[i]; + } vui->nregions = msg.memory.nregions; break; case VHOST_USER_SET_VRING_NUM: - DBG_SOCK("if %d msg VHOST_USER_SET_VRING_NUM idx %d num %d", - vui->hw_if_index, msg.state.index, msg.state.num); + DBG_SOCK ("if %d msg VHOST_USER_SET_VRING_NUM idx %d num %d", + vui->hw_if_index, msg.state.index, msg.state.num); - if ((msg.state.num > 32768) || /* maximum ring size is 32768 */ - (msg.state.num == 0) || /* it cannot be zero */ - (msg.state.num % 2)) /* must be power of 2 */ - goto close_socket; + if ((msg.state.num > 32768) || /* maximum ring size is 32768 */ + (msg.state.num == 0) || /* it cannot be zero */ + (msg.state.num % 2)) /* must be power of 2 */ + goto close_socket; vui->vrings[msg.state.index].qsz = msg.state.num; break; case VHOST_USER_SET_VRING_ADDR: - DBG_SOCK("if %d msg VHOST_USER_SET_VRING_ADDR idx %d", - vui->hw_if_index, msg.state.index); + DBG_SOCK ("if %d msg VHOST_USER_SET_VRING_ADDR idx %d", + vui->hw_if_index, msg.state.index); - vui->vrings[msg.state.index].desc = (vring_desc_t *) - map_user_mem(vui, msg.addr.desc_user_addr); + vui->vrings[msg.state.index].desc = (vring_desc_t *) + map_user_mem (vui, msg.addr.desc_user_addr); vui->vrings[msg.state.index].used = (vring_used_t *) - map_user_mem(vui, msg.addr.used_user_addr); + map_user_mem (vui, msg.addr.used_user_addr); vui->vrings[msg.state.index].avail = (vring_avail_t *) - map_user_mem(vui, msg.addr.avail_user_addr); + map_user_mem (vui, msg.addr.avail_user_addr); if ((vui->vrings[msg.state.index].desc == NULL) || - (vui->vrings[msg.state.index].used == NULL) || - (vui->vrings[msg.state.index].avail == NULL)) { - DBG_SOCK("failed to map user memory for hw_if_index %d", - vui->hw_if_index); - goto close_socket; - } + (vui->vrings[msg.state.index].used == NULL) || + (vui->vrings[msg.state.index].avail == NULL)) + { + DBG_SOCK ("failed to map user memory for hw_if_index %d", + vui->hw_if_index); + goto close_socket; + } vui->vrings[msg.state.index].log_guest_addr = msg.addr.log_guest_addr; vui->vrings[msg.state.index].log_used = - (msg.addr.flags & (1 << VHOST_VRING_F_LOG)) ? 1 : 0; + (msg.addr.flags & (1 << VHOST_VRING_F_LOG)) ? 1 : 0; /* Spec says: If VHOST_USER_F_PROTOCOL_FEATURES has not been negotiated, - the ring is initialized in an enabled state. */ + the ring is initialized in an enabled state. */ - if (!(vui->features & (1 << FEAT_VHOST_USER_F_PROTOCOL_FEATURES))) { - vui->vrings[msg.state.index].enabled = 1; - } + if (!(vui->features & (1 << FEAT_VHOST_USER_F_PROTOCOL_FEATURES))) + { + vui->vrings[msg.state.index].enabled = 1; + } vui->vrings[msg.state.index].last_used_idx = - vui->vrings[msg.state.index].used->idx; + vui->vrings[msg.state.index].used->idx; /* tell driver that we don't want interrupts */ vui->vrings[msg.state.index].used->flags |= 1; break; case VHOST_USER_SET_OWNER: - DBG_SOCK("if %d msg VHOST_USER_SET_OWNER", - vui->hw_if_index); + DBG_SOCK ("if %d msg VHOST_USER_SET_OWNER", vui->hw_if_index); break; case VHOST_USER_RESET_OWNER: - DBG_SOCK("if %d msg VHOST_USER_RESET_OWNER", - vui->hw_if_index); + DBG_SOCK ("if %d msg VHOST_USER_RESET_OWNER", vui->hw_if_index); break; case VHOST_USER_SET_VRING_CALL: - DBG_SOCK("if %d msg VHOST_USER_SET_VRING_CALL u64 %d", - vui->hw_if_index, msg.u64); + DBG_SOCK ("if %d msg VHOST_USER_SET_VRING_CALL u64 %d", + vui->hw_if_index, msg.u64); q = (u8) (msg.u64 & 0xFF); if (!(msg.u64 & 0x100)) - { - if (number_of_fds != 1) - goto close_socket; - - /* if there is old fd, delete it */ - if (vui->vrings[q].callfd) { - unix_file_t * uf = pool_elt_at_index (unix_main.file_pool, - vui->vrings[q].callfd_idx); - unix_file_del (&unix_main, uf); - } - vui->vrings[q].callfd = fds[0]; - template.read_function = vhost_user_callfd_read_ready; - template.file_descriptor = fds[0]; - vui->vrings[q].callfd_idx = unix_file_add (&unix_main, &template); - } + { + if (number_of_fds != 1) + goto close_socket; + + /* if there is old fd, delete it */ + if (vui->vrings[q].callfd) + { + unix_file_t *uf = pool_elt_at_index (unix_main.file_pool, + vui->vrings[q].callfd_idx); + unix_file_del (&unix_main, uf); + } + vui->vrings[q].callfd = fds[0]; + template.read_function = vhost_user_callfd_read_ready; + template.file_descriptor = fds[0]; + vui->vrings[q].callfd_idx = unix_file_add (&unix_main, &template); + } else - vui->vrings[q].callfd = -1; + vui->vrings[q].callfd = -1; break; case VHOST_USER_SET_VRING_KICK: - DBG_SOCK("if %d msg VHOST_USER_SET_VRING_KICK u64 %d", - vui->hw_if_index, msg.u64); + DBG_SOCK ("if %d msg VHOST_USER_SET_VRING_KICK u64 %d", + vui->hw_if_index, msg.u64); q = (u8) (msg.u64 & 0xFF); if (!(msg.u64 & 0x100)) - { - if (number_of_fds != 1) - goto close_socket; + { + if (number_of_fds != 1) + goto close_socket; - vui->vrings[q].kickfd = fds[0]; - } + vui->vrings[q].kickfd = fds[0]; + } else - vui->vrings[q].kickfd = -1; + vui->vrings[q].kickfd = -1; break; case VHOST_USER_SET_VRING_ERR: - DBG_SOCK("if %d msg VHOST_USER_SET_VRING_ERR u64 %d", - vui->hw_if_index, msg.u64); + DBG_SOCK ("if %d msg VHOST_USER_SET_VRING_ERR u64 %d", + vui->hw_if_index, msg.u64); q = (u8) (msg.u64 & 0xFF); if (!(msg.u64 & 0x100)) - { - if (number_of_fds != 1) - goto close_socket; + { + if (number_of_fds != 1) + goto close_socket; - fd = fds[0]; - } + fd = fds[0]; + } else - fd = -1; + fd = -1; vui->vrings[q].errfd = fd; break; case VHOST_USER_SET_VRING_BASE: - DBG_SOCK("if %d msg VHOST_USER_SET_VRING_BASE idx %d num %d", - vui->hw_if_index, msg.state.index, msg.state.num); + DBG_SOCK ("if %d msg VHOST_USER_SET_VRING_BASE idx %d num %d", + vui->hw_if_index, msg.state.index, msg.state.num); vui->vrings[msg.state.index].last_avail_idx = msg.state.num; break; case VHOST_USER_GET_VRING_BASE: - DBG_SOCK("if %d msg VHOST_USER_GET_VRING_BASE idx %d num %d", - vui->hw_if_index, msg.state.index, msg.state.num); + DBG_SOCK ("if %d msg VHOST_USER_GET_VRING_BASE idx %d num %d", + vui->hw_if_index, msg.state.index, msg.state.num); /* Spec says: Client must [...] stop ring upon receiving VHOST_USER_GET_VRING_BASE. */ vui->vrings[msg.state.index].enabled = 0; msg.state.num = vui->vrings[msg.state.index].last_avail_idx; msg.flags |= 4; - msg.size = sizeof(msg.state); + msg.size = sizeof (msg.state); break; case VHOST_USER_NONE: - DBG_SOCK("if %d msg VHOST_USER_NONE", - vui->hw_if_index); + DBG_SOCK ("if %d msg VHOST_USER_NONE", vui->hw_if_index); break; case VHOST_USER_SET_LOG_BASE: - { - DBG_SOCK("if %d msg VHOST_USER_SET_LOG_BASE", - vui->hw_if_index); - - if (msg.size != sizeof(msg.log)) { - DBG_SOCK("invalid msg size for VHOST_USER_SET_LOG_BASE: %d instead of %d", - msg.size, sizeof(msg.log)); - goto close_socket; - } - - if (!(vui->protocol_features & (1 << VHOST_USER_PROTOCOL_F_LOG_SHMFD))) { - DBG_SOCK("VHOST_USER_PROTOCOL_F_LOG_SHMFD not set but VHOST_USER_SET_LOG_BASE received"); - goto close_socket; - } - - fd = fds[0]; - /* align size to 2M page */ - long page_sz = get_huge_page_size(fd); - ssize_t map_sz = (msg.log.size + msg.log.offset + page_sz) & ~(page_sz - 1); - - vui->log_base_addr = mmap(0, map_sz, PROT_READ | PROT_WRITE, - MAP_SHARED, fd, 0); - - DBG_SOCK("map log region addr 0 len 0x%lx off 0x%lx fd %d mapped 0x%lx", - map_sz, msg.log.offset, fd, vui->log_base_addr); - - if (vui->log_base_addr == MAP_FAILED) { - clib_warning("failed to map memory. errno is %d", errno); - goto close_socket; + { + DBG_SOCK ("if %d msg VHOST_USER_SET_LOG_BASE", vui->hw_if_index); + + if (msg.size != sizeof (msg.log)) + { + DBG_SOCK + ("invalid msg size for VHOST_USER_SET_LOG_BASE: %d instead of %d", + msg.size, sizeof (msg.log)); + goto close_socket; + } + + if (! + (vui->protocol_features & (1 << VHOST_USER_PROTOCOL_F_LOG_SHMFD))) + { + DBG_SOCK + ("VHOST_USER_PROTOCOL_F_LOG_SHMFD not set but VHOST_USER_SET_LOG_BASE received"); + goto close_socket; + } + + fd = fds[0]; + /* align size to 2M page */ + long page_sz = get_huge_page_size (fd); + ssize_t map_sz = + (msg.log.size + msg.log.offset + page_sz) & ~(page_sz - 1); + + vui->log_base_addr = mmap (0, map_sz, PROT_READ | PROT_WRITE, + MAP_SHARED, fd, 0); + + DBG_SOCK + ("map log region addr 0 len 0x%lx off 0x%lx fd %d mapped 0x%lx", + map_sz, msg.log.offset, fd, vui->log_base_addr); + + if (vui->log_base_addr == MAP_FAILED) + { + clib_warning ("failed to map memory. errno is %d", errno); + goto close_socket; + } + + vui->log_base_addr += msg.log.offset; + vui->log_size = msg.log.size; + + msg.flags |= 4; + msg.size = sizeof (msg.u64); + + break; } - vui->log_base_addr += msg.log.offset; - vui->log_size = msg.log.size; - - msg.flags |= 4; - msg.size = sizeof(msg.u64); - - break; - } - case VHOST_USER_SET_LOG_FD: - DBG_SOCK("if %d msg VHOST_USER_SET_LOG_FD", - vui->hw_if_index); + DBG_SOCK ("if %d msg VHOST_USER_SET_LOG_FD", vui->hw_if_index); break; case VHOST_USER_GET_PROTOCOL_FEATURES: - DBG_SOCK("if %d msg VHOST_USER_GET_PROTOCOL_FEATURES", vui->hw_if_index); + DBG_SOCK ("if %d msg VHOST_USER_GET_PROTOCOL_FEATURES", + vui->hw_if_index); msg.flags |= 4; msg.u64 = (1 << VHOST_USER_PROTOCOL_F_LOG_SHMFD); - msg.size = sizeof(msg.u64); + msg.size = sizeof (msg.u64); break; case VHOST_USER_SET_PROTOCOL_FEATURES: - DBG_SOCK("if %d msg VHOST_USER_SET_PROTOCOL_FEATURES features 0x%lx", - vui->hw_if_index, msg.u64); + DBG_SOCK ("if %d msg VHOST_USER_SET_PROTOCOL_FEATURES features 0x%lx", + vui->hw_if_index, msg.u64); vui->protocol_features = msg.u64; break; case VHOST_USER_SET_VRING_ENABLE: - DBG_SOCK("if %d VHOST_USER_SET_VRING_ENABLE, enable: %d", - vui->hw_if_index, msg.state.num); + DBG_SOCK ("if %d VHOST_USER_SET_VRING_ENABLE, enable: %d", + vui->hw_if_index, msg.state.num); vui->vrings[msg.state.index].enabled = msg.state.num; break; default: - DBG_SOCK("unknown vhost-user message %d received. closing socket", - msg.request); + DBG_SOCK ("unknown vhost-user message %d received. closing socket", + msg.request); goto close_socket; - } + } - /* if we have pointers to descriptor table, go up*/ + /* if we have pointers to descriptor table, go up */ if (!vui->is_up && - vui->vrings[VHOST_NET_VRING_IDX_TX].desc && - vui->vrings[VHOST_NET_VRING_IDX_RX].desc) { + vui->vrings[VHOST_NET_VRING_IDX_TX].desc && + vui->vrings[VHOST_NET_VRING_IDX_RX].desc) + { - DBG_SOCK("interface %d connected", vui->sw_if_index); + DBG_SOCK ("interface %d connected", vui->sw_if_index); - vnet_hw_interface_set_flags (vnm, vui->hw_if_index, VNET_HW_INTERFACE_FLAG_LINK_UP); + vnet_hw_interface_set_flags (vnm, vui->hw_if_index, + VNET_HW_INTERFACE_FLAG_LINK_UP); vui->is_up = 1; - } + } /* if we need to reply */ if (msg.flags & 4) - { - n = send(uf->file_descriptor, &msg, VHOST_USER_MSG_HDR_SZ + msg.size, 0); + { + n = + send (uf->file_descriptor, &msg, VHOST_USER_MSG_HDR_SZ + msg.size, 0); if (n != (msg.size + VHOST_USER_MSG_HDR_SZ)) - goto close_socket; - } + goto close_socket; + } return 0; close_socket: - vhost_user_if_disconnect(vui); + vhost_user_if_disconnect (vui); return 0; } -static clib_error_t * vhost_user_socket_error (unix_file_t * uf) +static clib_error_t * +vhost_user_socket_error (unix_file_t * uf) { - vhost_user_main_t * vum = &vhost_user_main; - vhost_user_intf_t * vui; - uword * p; + vhost_user_main_t *vum = &vhost_user_main; + vhost_user_intf_t *vui; + uword *p; p = hash_get (vum->vhost_user_interface_index_by_sock_fd, - uf->file_descriptor); - if (p == 0) { - DBG_SOCK ("fd %d doesn't belong to any interface", - uf->file_descriptor); + uf->file_descriptor); + if (p == 0) + { + DBG_SOCK ("fd %d doesn't belong to any interface", uf->file_descriptor); return 0; } else vui = vec_elt_at_index (vum->vhost_user_interfaces, p[0]); - vhost_user_if_disconnect(vui); + vhost_user_if_disconnect (vui); return 0; } -static clib_error_t * vhost_user_socksvr_accept_ready (unix_file_t * uf) +static clib_error_t * +vhost_user_socksvr_accept_ready (unix_file_t * uf) { int client_fd, client_len; struct sockaddr_un client; - unix_file_t template = {0}; - vhost_user_main_t * vum = &vhost_user_main; - vhost_user_intf_t * vui; - uword * p; + unix_file_t template = { 0 }; + vhost_user_main_t *vum = &vhost_user_main; + vhost_user_intf_t *vui; + uword *p; p = hash_get (vum->vhost_user_interface_index_by_listener_fd, - uf->file_descriptor); - if (p == 0) { - DBG_SOCK ("fd %d doesn't belong to any interface", - uf->file_descriptor); + uf->file_descriptor); + if (p == 0) + { + DBG_SOCK ("fd %d doesn't belong to any interface", uf->file_descriptor); return 0; } else vui = vec_elt_at_index (vum->vhost_user_interfaces, p[0]); - client_len = sizeof(client); + client_len = sizeof (client); client_fd = accept (uf->file_descriptor, - (struct sockaddr *)&client, - (socklen_t *)&client_len); + (struct sockaddr *) &client, + (socklen_t *) & client_len); if (client_fd < 0) - return clib_error_return_unix (0, "accept"); + return clib_error_return_unix (0, "accept"); template.read_function = vhost_user_socket_read; template.error_function = vhost_user_socket_error; @@ -728,7 +786,7 @@ static clib_error_t * vhost_user_socksvr_accept_ready (unix_file_t * uf) vui->client_fd = client_fd; hash_set (vum->vhost_user_interface_index_by_sock_fd, vui->client_fd, - vui - vum->vhost_user_interfaces); + vui - vum->vhost_user_interfaces); return 0; } @@ -736,22 +794,25 @@ static clib_error_t * vhost_user_socksvr_accept_ready (unix_file_t * uf) static clib_error_t * vhost_user_init (vlib_main_t * vm) { - clib_error_t * error; - vhost_user_main_t * vum = &vhost_user_main; - vlib_thread_main_t * tm = vlib_get_thread_main(); + clib_error_t *error; + vhost_user_main_t *vum = &vhost_user_main; + vlib_thread_main_t *tm = vlib_get_thread_main (); error = vlib_call_init_function (vm, ip4_init); if (error) return error; - vum->vhost_user_interface_index_by_listener_fd = hash_create (0, sizeof (uword)); - vum->vhost_user_interface_index_by_sock_fd = hash_create (0, sizeof (uword)); - vum->vhost_user_interface_index_by_sw_if_index = hash_create (0, sizeof (uword)); + vum->vhost_user_interface_index_by_listener_fd = + hash_create (0, sizeof (uword)); + vum->vhost_user_interface_index_by_sock_fd = + hash_create (0, sizeof (uword)); + vum->vhost_user_interface_index_by_sw_if_index = + hash_create (0, sizeof (uword)); vum->coalesce_frames = 32; vum->coalesce_time = 1e-3; vec_validate_aligned (vum->rx_buffers, tm->n_vlib_mains - 1, - CLIB_CACHE_LINE_BYTES); + CLIB_CACHE_LINE_BYTES); return 0; } @@ -767,14 +828,16 @@ vhost_user_exit (vlib_main_t * vm) VLIB_MAIN_LOOP_EXIT_FUNCTION (vhost_user_exit); -enum { +enum +{ VHOST_USER_RX_NEXT_ETHERNET_INPUT, VHOST_USER_RX_NEXT_DROP, VHOST_USER_RX_N_NEXT, }; -typedef struct { +typedef struct +{ u16 virtqueue; u16 device_index; #if VHOST_USER_COPY_TX_HDR == 1 @@ -782,94 +845,93 @@ typedef struct { #endif } vhost_user_input_trace_t; -static u8 * format_vhost_user_input_trace (u8 * s, va_list * va) +static u8 * +format_vhost_user_input_trace (u8 * s, va_list * va) { CLIB_UNUSED (vlib_main_t * vm) = va_arg (*va, vlib_main_t *); CLIB_UNUSED (vlib_node_t * node) = va_arg (*va, vlib_node_t *); - CLIB_UNUSED (vnet_main_t * vnm) = vnet_get_main(); - vhost_user_main_t * vum = &vhost_user_main; - vhost_user_input_trace_t * t = va_arg (*va, vhost_user_input_trace_t *); - vhost_user_intf_t * vui = vec_elt_at_index (vum->vhost_user_interfaces, - t->device_index); + CLIB_UNUSED (vnet_main_t * vnm) = vnet_get_main (); + vhost_user_main_t *vum = &vhost_user_main; + vhost_user_input_trace_t *t = va_arg (*va, vhost_user_input_trace_t *); + vhost_user_intf_t *vui = vec_elt_at_index (vum->vhost_user_interfaces, + t->device_index); - vnet_sw_interface_t * sw = vnet_get_sw_interface (vnm, vui->sw_if_index); + vnet_sw_interface_t *sw = vnet_get_sw_interface (vnm, vui->sw_if_index); #if VHOST_USER_COPY_TX_HDR == 1 uword indent = format_get_indent (s); #endif s = format (s, "%U virtqueue %d", - format_vnet_sw_interface_name, vnm, sw, - t->virtqueue); + format_vnet_sw_interface_name, vnm, sw, t->virtqueue); #if VHOST_USER_COPY_TX_HDR == 1 s = format (s, "\n%Uvirtio_net_hdr flags 0x%02x gso_type %u hdr_len %u", - format_white_space, indent, - t->hdr.flags, - t->hdr.gso_type, - t->hdr.hdr_len); + format_white_space, indent, + t->hdr.flags, t->hdr.gso_type, t->hdr.hdr_len); #endif return s; } -void vhost_user_rx_trace (vlib_main_t * vm, - vlib_node_runtime_t * node, - vhost_user_intf_t *vui, - i16 virtqueue) +void +vhost_user_rx_trace (vlib_main_t * vm, + vlib_node_runtime_t * node, + vhost_user_intf_t * vui, i16 virtqueue) { - u32 * b, n_left; - vhost_user_main_t * vum = &vhost_user_main; + u32 *b, n_left; + vhost_user_main_t *vum = &vhost_user_main; u32 next_index = VHOST_USER_RX_NEXT_ETHERNET_INPUT; - n_left = vec_len(vui->d_trace_buffers); + n_left = vec_len (vui->d_trace_buffers); b = vui->d_trace_buffers; while (n_left >= 1) - { - u32 bi0; - vlib_buffer_t * b0; - vhost_user_input_trace_t * t0; - - bi0 = b[0]; - n_left -= 1; - - b0 = vlib_get_buffer (vm, bi0); - vlib_trace_buffer (vm, node, next_index, b0, /* follow_chain */ 0); - t0 = vlib_add_trace (vm, node, b0, sizeof (t0[0])); - t0->virtqueue = virtqueue; - t0->device_index = vui - vum->vhost_user_interfaces; + { + u32 bi0; + vlib_buffer_t *b0; + vhost_user_input_trace_t *t0; + + bi0 = b[0]; + n_left -= 1; + + b0 = vlib_get_buffer (vm, bi0); + vlib_trace_buffer (vm, node, next_index, b0, /* follow_chain */ 0); + t0 = vlib_add_trace (vm, node, b0, sizeof (t0[0])); + t0->virtqueue = virtqueue; + t0->device_index = vui - vum->vhost_user_interfaces; #if VHOST_USER_COPY_TX_HDR == 1 - clib_memcpy(&t0->hdr, b0->pre_data, sizeof(virtio_net_hdr_t)); + clib_memcpy (&t0->hdr, b0->pre_data, sizeof (virtio_net_hdr_t)); #endif - b+=1; - } + b += 1; + } } -static inline void vhost_user_send_call(vlib_main_t * vm, vhost_user_vring_t * vq) +static inline void +vhost_user_send_call (vlib_main_t * vm, vhost_user_vring_t * vq) { - vhost_user_main_t * vum = &vhost_user_main; - u64 x = 1; - int rv __attribute__((unused)); - /* $$$$ pay attention to rv */ - rv = write(vq->callfd, &x, sizeof(x)); - vq->n_since_last_int = 0; - vq->int_deadline = vlib_time_now(vm) + vum->coalesce_time; + vhost_user_main_t *vum = &vhost_user_main; + u64 x = 1; + int rv __attribute__ ((unused)); + /* $$$$ pay attention to rv */ + rv = write (vq->callfd, &x, sizeof (x)); + vq->n_since_last_int = 0; + vq->int_deadline = vlib_time_now (vm) + vum->coalesce_time; } -static u32 vhost_user_if_input ( vlib_main_t * vm, - vhost_user_main_t * vum, - vhost_user_intf_t * vui, - vlib_node_runtime_t * node) +static u32 +vhost_user_if_input (vlib_main_t * vm, + vhost_user_main_t * vum, + vhost_user_intf_t * vui, vlib_node_runtime_t * node) { - vhost_user_vring_t * txvq = &vui->vrings[VHOST_NET_VRING_IDX_TX]; - vhost_user_vring_t * rxvq = &vui->vrings[VHOST_NET_VRING_IDX_RX]; + vhost_user_vring_t *txvq = &vui->vrings[VHOST_NET_VRING_IDX_TX]; + vhost_user_vring_t *rxvq = &vui->vrings[VHOST_NET_VRING_IDX_RX]; uword n_rx_packets = 0, n_rx_bytes = 0; uword n_left; - u32 n_left_to_next, * to_next; + u32 n_left_to_next, *to_next; u32 next_index = 0; u32 next0; uword n_trace = vlib_get_trace_count (vm, node); @@ -880,45 +942,46 @@ static u32 vhost_user_if_input ( vlib_main_t * vm, vec_reset_length (vui->d_trace_buffers); /* no descriptor ptr - bail out */ - if (PREDICT_FALSE(!txvq->desc || !txvq->avail || !txvq->enabled)) + if (PREDICT_FALSE (!txvq->desc || !txvq->avail || !txvq->enabled)) return 0; /* do we have pending intterupts ? */ if ((txvq->n_since_last_int) && (txvq->int_deadline < now)) - vhost_user_send_call(vm, txvq); + vhost_user_send_call (vm, txvq); if ((rxvq->n_since_last_int) && (rxvq->int_deadline < now)) - vhost_user_send_call(vm, rxvq); + vhost_user_send_call (vm, rxvq); /* only bit 0 of avail.flags is used so we don't want to deal with this interface if any other bit is set */ - if (PREDICT_FALSE(txvq->avail->flags & 0xFFFE)) + if (PREDICT_FALSE (txvq->avail->flags & 0xFFFE)) return 0; /* nothing to do */ if (txvq->avail->idx == txvq->last_avail_idx) return 0; - if (PREDICT_TRUE(txvq->avail->idx > txvq->last_avail_idx)) + if (PREDICT_TRUE (txvq->avail->idx > txvq->last_avail_idx)) n_left = txvq->avail->idx - txvq->last_avail_idx; - else /* wrapped */ - n_left = (u16) -1 - txvq->last_avail_idx + txvq->avail->idx; - - if (PREDICT_FALSE(!vui->admin_up)) { - /* if intf is admin down, just drop all packets waiting in the ring */ - txvq->last_avail_idx = txvq->last_used_idx = txvq->avail->idx; - CLIB_MEMORY_BARRIER(); - txvq->used->idx = txvq->last_used_idx; - vhost_user_log_dirty_ring(vui, txvq, idx); - vhost_user_send_call(vm, txvq); - return 0; - } + else /* wrapped */ + n_left = (u16) - 1 - txvq->last_avail_idx + txvq->avail->idx; - if (PREDICT_FALSE(n_left > txvq->qsz)) + if (PREDICT_FALSE (!vui->admin_up)) + { + /* if intf is admin down, just drop all packets waiting in the ring */ + txvq->last_avail_idx = txvq->last_used_idx = txvq->avail->idx; + CLIB_MEMORY_BARRIER (); + txvq->used->idx = txvq->last_used_idx; + vhost_user_log_dirty_ring (vui, txvq, idx); + vhost_user_send_call (vm, txvq); + return 0; + } + + if (PREDICT_FALSE (n_left > txvq->qsz)) return 0; qsz_mask = txvq->qsz - 1; - cpu_index = os_get_cpu_number(); + cpu_index = os_get_cpu_number (); drops = 0; flush = 0; @@ -932,200 +995,239 @@ static u32 vhost_user_if_input ( vlib_main_t * vm, * to cycle through the descriptors without having to check for errors. * For jumbo frames, the bottleneck is memory copy anyway. */ - if (PREDICT_FALSE(!vum->rx_buffers[cpu_index])) { - vec_alloc (vum->rx_buffers[cpu_index], VLIB_FRAME_SIZE); - - if (PREDICT_FALSE(!vum->rx_buffers[cpu_index])) - flush = n_left; //Drop all input - } - - if (PREDICT_FALSE(_vec_len(vum->rx_buffers[cpu_index]) < n_left)) { - _vec_len(vum->rx_buffers[cpu_index]) += - vlib_buffer_alloc_from_free_list(vm, vum->rx_buffers[cpu_index] + _vec_len(vum->rx_buffers[cpu_index]), - VLIB_FRAME_SIZE - _vec_len(vum->rx_buffers[cpu_index]), - VLIB_BUFFER_DEFAULT_FREE_LIST_INDEX); - - if (PREDICT_FALSE(n_left > _vec_len(vum->rx_buffers[cpu_index]))) - flush = n_left - _vec_len(vum->rx_buffers[cpu_index]); - } + if (PREDICT_FALSE (!vum->rx_buffers[cpu_index])) + { + vec_alloc (vum->rx_buffers[cpu_index], VLIB_FRAME_SIZE); - if (PREDICT_FALSE(flush)) { - //Remove some input buffers - drops += flush; - n_left -= flush; - vlib_error_count(vm, vhost_user_input_node.index, - VHOST_USER_INPUT_FUNC_ERROR_NO_BUFFER, flush); - while (flush) { - u16 desc_chain_head = txvq->avail->ring[txvq->last_avail_idx & qsz_mask]; - txvq->last_avail_idx++; - txvq->used->ring[txvq->last_used_idx & qsz_mask].id = desc_chain_head; - txvq->used->ring[txvq->last_used_idx & qsz_mask].len = 0; - vhost_user_log_dirty_ring(vui, txvq, ring[txvq->last_used_idx & qsz_mask]); - txvq->last_used_idx++; - flush--; + if (PREDICT_FALSE (!vum->rx_buffers[cpu_index])) + flush = n_left; //Drop all input } - } - - rx_len = vec_len(vum->rx_buffers[cpu_index]); //vector might be null - while (n_left > 0) { - vlib_get_next_frame (vm, node, next_index, to_next, n_left_to_next); - while (n_left > 0 && n_left_to_next > 0) { - vlib_buffer_t *b_head, *b_current; - u32 bi_head, bi_current; - u16 desc_chain_head, desc_current; - u8 error = VHOST_USER_INPUT_FUNC_ERROR_NO_ERROR; - - desc_chain_head = desc_current = txvq->avail->ring[txvq->last_avail_idx & qsz_mask]; - bi_head = bi_current = vum->rx_buffers[cpu_index][--rx_len]; - b_head = b_current = vlib_get_buffer (vm, bi_head); - vlib_buffer_chain_init(b_head); + if (PREDICT_FALSE (_vec_len (vum->rx_buffers[cpu_index]) < n_left)) + { + _vec_len (vum->rx_buffers[cpu_index]) += + vlib_buffer_alloc_from_free_list (vm, + vum->rx_buffers[cpu_index] + + _vec_len (vum->rx_buffers + [cpu_index]), + VLIB_FRAME_SIZE - + _vec_len (vum->rx_buffers + [cpu_index]), + VLIB_BUFFER_DEFAULT_FREE_LIST_INDEX); + + if (PREDICT_FALSE (n_left > _vec_len (vum->rx_buffers[cpu_index]))) + flush = n_left - _vec_len (vum->rx_buffers[cpu_index]); + } - uword offset; - if (PREDICT_TRUE(vui->is_any_layout) || - !(txvq->desc[desc_current].flags & VIRTQ_DESC_F_NEXT)) { - /* ANYLAYOUT or single buffer */ - offset = vui->virtio_net_hdr_sz; - } else { - /* CSR case without ANYLAYOUT, skip 1st buffer */ - offset = txvq->desc[desc_current].len; - } + if (PREDICT_FALSE (flush)) + { + //Remove some input buffers + drops += flush; + n_left -= flush; + vlib_error_count (vm, vhost_user_input_node.index, + VHOST_USER_INPUT_FUNC_ERROR_NO_BUFFER, flush); + while (flush) + { + u16 desc_chain_head = + txvq->avail->ring[txvq->last_avail_idx & qsz_mask]; + txvq->last_avail_idx++; + txvq->used->ring[txvq->last_used_idx & qsz_mask].id = + desc_chain_head; + txvq->used->ring[txvq->last_used_idx & qsz_mask].len = 0; + vhost_user_log_dirty_ring (vui, txvq, + ring[txvq->last_used_idx & qsz_mask]); + txvq->last_used_idx++; + flush--; + } + } - while(1) { - void * buffer_addr = map_guest_mem(vui, txvq->desc[desc_current].addr); - if (PREDICT_FALSE(buffer_addr == 0)) { - error = VHOST_USER_INPUT_FUNC_ERROR_MMAP_FAIL; - break; - } + rx_len = vec_len (vum->rx_buffers[cpu_index]); //vector might be null + while (n_left > 0) + { + vlib_get_next_frame (vm, node, next_index, to_next, n_left_to_next); + + while (n_left > 0 && n_left_to_next > 0) + { + vlib_buffer_t *b_head, *b_current; + u32 bi_head, bi_current; + u16 desc_chain_head, desc_current; + u8 error = VHOST_USER_INPUT_FUNC_ERROR_NO_ERROR; + + desc_chain_head = desc_current = + txvq->avail->ring[txvq->last_avail_idx & qsz_mask]; + bi_head = bi_current = vum->rx_buffers[cpu_index][--rx_len]; + b_head = b_current = vlib_get_buffer (vm, bi_head); + vlib_buffer_chain_init (b_head); + + uword offset; + if (PREDICT_TRUE (vui->is_any_layout) || + !(txvq->desc[desc_current].flags & VIRTQ_DESC_F_NEXT)) + { + /* ANYLAYOUT or single buffer */ + offset = vui->virtio_net_hdr_sz; + } + else + { + /* CSR case without ANYLAYOUT, skip 1st buffer */ + offset = txvq->desc[desc_current].len; + } + + while (1) + { + void *buffer_addr = + map_guest_mem (vui, txvq->desc[desc_current].addr); + if (PREDICT_FALSE (buffer_addr == 0)) + { + error = VHOST_USER_INPUT_FUNC_ERROR_MMAP_FAIL; + break; + } #if VHOST_USER_COPY_TX_HDR == 1 - if (PREDICT_TRUE(offset)) - clib_memcpy(b->pre_data, buffer_addr, sizeof(virtio_net_hdr_t)); /* 12 byte hdr is not used on tx */ + if (PREDICT_TRUE (offset)) + clib_memcpy (b->pre_data, buffer_addr, sizeof (virtio_net_hdr_t)); /* 12 byte hdr is not used on tx */ #endif - if (txvq->desc[desc_current].len > offset) { - u16 len = txvq->desc[desc_current].len - offset; - u16 copied = vlib_buffer_chain_append_data_with_alloc(vm, VLIB_BUFFER_DEFAULT_FREE_LIST_INDEX, - b_head, &b_current, buffer_addr + offset, len); - - if (copied != len) { - error = VHOST_USER_INPUT_FUNC_ERROR_NO_BUFFER; - break; - } - } - offset = 0; - - /* if next flag is set, take next desc in the chain */ - if (txvq->desc[desc_current].flags & VIRTQ_DESC_F_NEXT ) - desc_current = txvq->desc[desc_current].next; - else - break; - } - - /* consume the descriptor and return it as used */ - txvq->last_avail_idx++; - txvq->used->ring[txvq->last_used_idx & qsz_mask].id = desc_chain_head; - txvq->used->ring[txvq->last_used_idx & qsz_mask].len = 0; - vhost_user_log_dirty_ring(vui, txvq, ring[txvq->last_used_idx & qsz_mask]); - txvq->last_used_idx++; - - if(PREDICT_FALSE(b_head->current_length < 14 && - error == VHOST_USER_INPUT_FUNC_ERROR_NO_ERROR)) { - error = VHOST_USER_INPUT_FUNC_ERROR_UNDERSIZED_FRAME; - } - - VLIB_BUFFER_TRACE_TRAJECTORY_INIT(b_head); - - vnet_buffer (b_head)->sw_if_index[VLIB_RX] = vui->sw_if_index; - vnet_buffer (b_head)->sw_if_index[VLIB_TX] = (u32)~0; - b_head->error = node->errors[error]; - - if (PREDICT_FALSE (n_trace > n_rx_packets)) - vec_add1 (vui->d_trace_buffers, bi_head); - - if (PREDICT_FALSE(error)) { - drops++; - next0 = VHOST_USER_RX_NEXT_DROP; - } else { - n_rx_bytes += b_head->current_length + b_head->total_length_not_including_first_buffer; - n_rx_packets++; - next0 = VHOST_USER_RX_NEXT_ETHERNET_INPUT; - } - - to_next[0] = bi_head; - to_next++; - n_left_to_next--; - vlib_validate_buffer_enqueue_x1 (vm, node, next_index, - to_next, n_left_to_next, - bi_head, next0); - n_left--; + if (txvq->desc[desc_current].len > offset) + { + u16 len = txvq->desc[desc_current].len - offset; + u16 copied = vlib_buffer_chain_append_data_with_alloc (vm, + VLIB_BUFFER_DEFAULT_FREE_LIST_INDEX, + b_head, + &b_current, + buffer_addr + + + offset, + len); + + if (copied != len) + { + error = VHOST_USER_INPUT_FUNC_ERROR_NO_BUFFER; + break; + } + } + offset = 0; + + /* if next flag is set, take next desc in the chain */ + if (txvq->desc[desc_current].flags & VIRTQ_DESC_F_NEXT) + desc_current = txvq->desc[desc_current].next; + else + break; + } + + /* consume the descriptor and return it as used */ + txvq->last_avail_idx++; + txvq->used->ring[txvq->last_used_idx & qsz_mask].id = + desc_chain_head; + txvq->used->ring[txvq->last_used_idx & qsz_mask].len = 0; + vhost_user_log_dirty_ring (vui, txvq, + ring[txvq->last_used_idx & qsz_mask]); + txvq->last_used_idx++; + + if (PREDICT_FALSE (b_head->current_length < 14 && + error == VHOST_USER_INPUT_FUNC_ERROR_NO_ERROR)) + { + error = VHOST_USER_INPUT_FUNC_ERROR_UNDERSIZED_FRAME; + } + + VLIB_BUFFER_TRACE_TRAJECTORY_INIT (b_head); + + vnet_buffer (b_head)->sw_if_index[VLIB_RX] = vui->sw_if_index; + vnet_buffer (b_head)->sw_if_index[VLIB_TX] = (u32) ~ 0; + b_head->error = node->errors[error]; + + if (PREDICT_FALSE (n_trace > n_rx_packets)) + vec_add1 (vui->d_trace_buffers, bi_head); + + if (PREDICT_FALSE (error)) + { + drops++; + next0 = VHOST_USER_RX_NEXT_DROP; + } + else + { + n_rx_bytes += + b_head->current_length + + b_head->total_length_not_including_first_buffer; + n_rx_packets++; + next0 = VHOST_USER_RX_NEXT_ETHERNET_INPUT; + } + + to_next[0] = bi_head; + to_next++; + n_left_to_next--; + vlib_validate_buffer_enqueue_x1 (vm, node, next_index, + to_next, n_left_to_next, + bi_head, next0); + n_left--; + } + + vlib_put_next_frame (vm, node, next_index, n_left_to_next); } - vlib_put_next_frame (vm, node, next_index, n_left_to_next); - } - - if (PREDICT_TRUE(vum->rx_buffers[cpu_index] != 0)) - _vec_len(vum->rx_buffers[cpu_index]) = rx_len; + if (PREDICT_TRUE (vum->rx_buffers[cpu_index] != 0)) + _vec_len (vum->rx_buffers[cpu_index]) = rx_len; /* give buffers back to driver */ - CLIB_MEMORY_BARRIER(); + CLIB_MEMORY_BARRIER (); txvq->used->idx = txvq->last_used_idx; - vhost_user_log_dirty_ring(vui, txvq, idx); + vhost_user_log_dirty_ring (vui, txvq, idx); if (PREDICT_FALSE (vec_len (vui->d_trace_buffers) > 0)) - { - vhost_user_rx_trace (vm, node, vui, VHOST_NET_VRING_IDX_TX); - vlib_set_trace_count (vm, node, n_trace - vec_len (vui->d_trace_buffers)); - } + { + vhost_user_rx_trace (vm, node, vui, VHOST_NET_VRING_IDX_TX); + vlib_set_trace_count (vm, node, + n_trace - vec_len (vui->d_trace_buffers)); + } /* interrupt (call) handling */ - if((txvq->callfd > 0) && !(txvq->avail->flags & 1)) { - txvq->n_since_last_int += n_rx_packets; + if ((txvq->callfd > 0) && !(txvq->avail->flags & 1)) + { + txvq->n_since_last_int += n_rx_packets; - if(txvq->n_since_last_int > vum->coalesce_frames) - vhost_user_send_call(vm, txvq); - } + if (txvq->n_since_last_int > vum->coalesce_frames) + vhost_user_send_call (vm, txvq); + } - if (PREDICT_FALSE(drops)) { - vlib_increment_simple_counter - (vnet_main.interface_main.sw_if_counters - + VNET_INTERFACE_COUNTER_DROP, os_get_cpu_number(), - vui->sw_if_index, drops); - } + if (PREDICT_FALSE (drops)) + { + vlib_increment_simple_counter + (vnet_main.interface_main.sw_if_counters + + VNET_INTERFACE_COUNTER_DROP, os_get_cpu_number (), + vui->sw_if_index, drops); + } /* increase rx counters */ vlib_increment_combined_counter - (vnet_main.interface_main.combined_sw_if_counters - + VNET_INTERFACE_COUNTER_RX, - os_get_cpu_number(), - vui->sw_if_index, - n_rx_packets, n_rx_bytes); + (vnet_main.interface_main.combined_sw_if_counters + + VNET_INTERFACE_COUNTER_RX, + os_get_cpu_number (), vui->sw_if_index, n_rx_packets, n_rx_bytes); return n_rx_packets; } static uword vhost_user_input (vlib_main_t * vm, - vlib_node_runtime_t * node, - vlib_frame_t * f) + vlib_node_runtime_t * node, vlib_frame_t * f) { - vhost_user_main_t * vum = &vhost_user_main; - dpdk_main_t * dm = &dpdk_main; - vhost_user_intf_t * vui; + vhost_user_main_t *vum = &vhost_user_main; + dpdk_main_t *dm = &dpdk_main; + vhost_user_intf_t *vui; uword n_rx_packets = 0; - u32 cpu_index = os_get_cpu_number(); + u32 cpu_index = os_get_cpu_number (); int i; - for(i = 0; i < vec_len(vum->vhost_user_interfaces); i++ ) + for (i = 0; i < vec_len (vum->vhost_user_interfaces); i++) { - vui = vec_elt_at_index(vum->vhost_user_interfaces, i); + vui = vec_elt_at_index (vum->vhost_user_interfaces, i); if (vui->is_up && - (i % dm->input_cpu_count) == (cpu_index - dm->input_cpu_first_index)) - n_rx_packets += vhost_user_if_input (vm, vum, vui, node); + (i % dm->input_cpu_count) == + (cpu_index - dm->input_cpu_first_index)) + n_rx_packets += vhost_user_if_input (vm, vum, vui, node); } return n_rx_packets; } +/* *INDENT-OFF* */ VLIB_REGISTER_NODE (vhost_user_input_node) = { .function = vhost_user_input, .type = VLIB_NODE_TYPE_INPUT, @@ -1148,90 +1250,103 @@ VLIB_REGISTER_NODE (vhost_user_input_node) = { }; VLIB_NODE_FUNCTION_MULTIARCH (vhost_user_input_node, vhost_user_input) +/* *INDENT-ON* */ static uword vhost_user_intfc_tx (vlib_main_t * vm, - vlib_node_runtime_t * node, - vlib_frame_t * frame) + vlib_node_runtime_t * node, vlib_frame_t * frame) { - u32 * buffers = vlib_frame_args (frame); + u32 *buffers = vlib_frame_args (frame); u32 n_left = 0; u16 used_index; - vhost_user_main_t * vum = &vhost_user_main; + vhost_user_main_t *vum = &vhost_user_main; uword n_packets = 0; - vnet_interface_output_runtime_t * rd = (void *) node->runtime_data; - vhost_user_intf_t * vui = vec_elt_at_index (vum->vhost_user_interfaces, rd->dev_instance); - vhost_user_vring_t * rxvq = &vui->vrings[VHOST_NET_VRING_IDX_RX]; + vnet_interface_output_runtime_t *rd = (void *) node->runtime_data; + vhost_user_intf_t *vui = + vec_elt_at_index (vum->vhost_user_interfaces, rd->dev_instance); + vhost_user_vring_t *rxvq = &vui->vrings[VHOST_NET_VRING_IDX_RX]; u16 qsz_mask; u8 error = VHOST_USER_TX_FUNC_ERROR_NONE; - if (PREDICT_FALSE(!vui->is_up)) - goto done2; + if (PREDICT_FALSE (!vui->is_up)) + goto done2; - if (PREDICT_FALSE(!rxvq->desc || !rxvq->avail || vui->sock_errno != 0 || !rxvq->enabled)) { - error = VHOST_USER_TX_FUNC_ERROR_NOT_READY; - goto done2; - } + if (PREDICT_FALSE + (!rxvq->desc || !rxvq->avail || vui->sock_errno != 0 || !rxvq->enabled)) + { + error = VHOST_USER_TX_FUNC_ERROR_NOT_READY; + goto done2; + } - if (PREDICT_FALSE(vui->lockp != 0)) + if (PREDICT_FALSE (vui->lockp != 0)) { while (__sync_lock_test_and_set (vui->lockp, 1)) - ; + ; } /* only bit 0 of avail.flags is used so we don't want to deal with this interface if any other bit is set */ - if (PREDICT_FALSE(rxvq->avail->flags & 0xFFFE)) { - error = VHOST_USER_TX_FUNC_ERROR_NOT_READY; - goto done2; - } + if (PREDICT_FALSE (rxvq->avail->flags & 0xFFFE)) + { + error = VHOST_USER_TX_FUNC_ERROR_NOT_READY; + goto done2; + } - if (PREDICT_FALSE((rxvq->avail->idx == rxvq->last_avail_idx))) { - error = VHOST_USER_TX_FUNC_ERROR_PKT_DROP_NOBUF; - goto done2; - } + if (PREDICT_FALSE ((rxvq->avail->idx == rxvq->last_avail_idx))) + { + error = VHOST_USER_TX_FUNC_ERROR_PKT_DROP_NOBUF; + goto done2; + } n_left = n_packets = frame->n_vectors; used_index = rxvq->used->idx; - qsz_mask = rxvq->qsz - 1; /* qsz is always power of 2 */ + qsz_mask = rxvq->qsz - 1; /* qsz is always power of 2 */ while (n_left > 0) - { + { vlib_buffer_t *b0, *current_b0; u16 desc_chain_head, desc_current, desc_len; void *buffer_addr; uword offset; if (n_left >= 2) - vlib_prefetch_buffer_with_index (vm, buffers[1], LOAD); + vlib_prefetch_buffer_with_index (vm, buffers[1], LOAD); b0 = vlib_get_buffer (vm, buffers[0]); buffers++; n_left--; - if (PREDICT_FALSE(rxvq->last_avail_idx == rxvq->avail->idx)) { - error = VHOST_USER_TX_FUNC_ERROR_PKT_DROP_NOBUF; - goto done; - } + if (PREDICT_FALSE (rxvq->last_avail_idx == rxvq->avail->idx)) + { + error = VHOST_USER_TX_FUNC_ERROR_PKT_DROP_NOBUF; + goto done; + } - desc_current = desc_chain_head = rxvq->avail->ring[rxvq->last_avail_idx & qsz_mask]; + desc_current = desc_chain_head = + rxvq->avail->ring[rxvq->last_avail_idx & qsz_mask]; offset = vui->virtio_net_hdr_sz; desc_len = offset; - if (PREDICT_FALSE(!(buffer_addr = map_guest_mem(vui, rxvq->desc[desc_current].addr)))) { - error = VHOST_USER_TX_FUNC_ERROR_MMAP_FAIL; - goto done; - } - CLIB_PREFETCH(buffer_addr, clib_min(rxvq->desc[desc_current].len, - 4*CLIB_CACHE_LINE_BYTES), STORE); - - virtio_net_hdr_mrg_rxbuf_t * hdr = (virtio_net_hdr_mrg_rxbuf_t *) buffer_addr; + if (PREDICT_FALSE + (!(buffer_addr = + map_guest_mem (vui, rxvq->desc[desc_current].addr)))) + { + error = VHOST_USER_TX_FUNC_ERROR_MMAP_FAIL; + goto done; + } + CLIB_PREFETCH (buffer_addr, clib_min (rxvq->desc[desc_current].len, + 4 * CLIB_CACHE_LINE_BYTES), + STORE); + + virtio_net_hdr_mrg_rxbuf_t *hdr = + (virtio_net_hdr_mrg_rxbuf_t *) buffer_addr; hdr->hdr.flags = 0; hdr->hdr.gso_type = 0; - vhost_user_log_dirty_pages(vui, rxvq->desc[desc_current].addr, vui->virtio_net_hdr_sz); + vhost_user_log_dirty_pages (vui, rxvq->desc[desc_current].addr, + vui->virtio_net_hdr_sz); if (vui->virtio_net_hdr_sz == 12) - hdr->num_buffers = 1; + hdr->num_buffers = 1; u16 bytes_left = b0->current_length; buffer_addr += offset; @@ -1239,134 +1354,167 @@ vhost_user_intfc_tx (vlib_main_t * vm, //FIXME: This was in the code but I don't think it is valid /*if (PREDICT_FALSE(!vui->is_any_layout && (rxvq->desc[desc_current].flags & VIRTQ_DESC_F_NEXT))) { - rxvq->desc[desc_current].len = vui->virtio_net_hdr_sz; - }*/ - - while(1) { - if (!bytes_left) { //Get new input - if (current_b0->flags & VLIB_BUFFER_NEXT_PRESENT) { - current_b0 = vlib_get_buffer(vm, current_b0->next_buffer); - bytes_left = current_b0->current_length; - } else { - //End of packet - break; - } - } - - if (rxvq->desc[desc_current].len <= offset) { //Get new output - if (rxvq->desc[desc_current].flags & VIRTQ_DESC_F_NEXT) { - offset = 0; - desc_current = rxvq->desc[desc_current].next; - if (PREDICT_FALSE(!(buffer_addr = map_guest_mem(vui, rxvq->desc[desc_current].addr)))) { - used_index -= hdr->num_buffers - 1; - rxvq->last_avail_idx -= hdr->num_buffers - 1; - error = VHOST_USER_TX_FUNC_ERROR_MMAP_FAIL; - goto done; - } - } else if (vui->virtio_net_hdr_sz == 12) { //MRG is available - - //Move from available to used buffer - rxvq->used->ring[used_index & qsz_mask].id = desc_chain_head; - rxvq->used->ring[used_index & qsz_mask].len = desc_len; - vhost_user_log_dirty_ring(vui, rxvq, ring[used_index & qsz_mask]); - rxvq->last_avail_idx++; - used_index++; - hdr->num_buffers++; - - if (PREDICT_FALSE(rxvq->last_avail_idx == rxvq->avail->idx)) { - //Dequeue queued descriptors for this packet - used_index -= hdr->num_buffers - 1; - rxvq->last_avail_idx -= hdr->num_buffers - 1; - error = VHOST_USER_TX_FUNC_ERROR_PKT_DROP_NOBUF; - goto done; - } - - //Look at next one - desc_chain_head = rxvq->avail->ring[rxvq->last_avail_idx & qsz_mask]; - desc_current = desc_chain_head; - desc_len = 0; - offset = 0; - if (PREDICT_FALSE(!(buffer_addr = map_guest_mem(vui, rxvq->desc[desc_current].addr)))) { - //Dequeue queued descriptors for this packet - used_index -= hdr->num_buffers - 1; - rxvq->last_avail_idx -= hdr->num_buffers - 1; - error = VHOST_USER_TX_FUNC_ERROR_MMAP_FAIL; - goto done; - } - } else { - error = VHOST_USER_TX_FUNC_ERROR_PKT_DROP_NOBUF; - goto done; - } - } - - u16 bytes_to_copy = bytes_left > (rxvq->desc[desc_current].len - offset) ? (rxvq->desc[desc_current].len - offset) : bytes_left; - clib_memcpy(buffer_addr, vlib_buffer_get_current (current_b0) + current_b0->current_length - bytes_left, bytes_to_copy); - - vhost_user_log_dirty_pages(vui, rxvq->desc[desc_current].addr + offset, bytes_to_copy); - bytes_left -= bytes_to_copy; - offset += bytes_to_copy; - buffer_addr += bytes_to_copy; - desc_len += bytes_to_copy; - } + rxvq->desc[desc_current].len = vui->virtio_net_hdr_sz; + } */ + + while (1) + { + if (!bytes_left) + { //Get new input + if (current_b0->flags & VLIB_BUFFER_NEXT_PRESENT) + { + current_b0 = vlib_get_buffer (vm, current_b0->next_buffer); + bytes_left = current_b0->current_length; + } + else + { + //End of packet + break; + } + } + + if (rxvq->desc[desc_current].len <= offset) + { //Get new output + if (rxvq->desc[desc_current].flags & VIRTQ_DESC_F_NEXT) + { + offset = 0; + desc_current = rxvq->desc[desc_current].next; + if (PREDICT_FALSE + (!(buffer_addr = + map_guest_mem (vui, rxvq->desc[desc_current].addr)))) + { + used_index -= hdr->num_buffers - 1; + rxvq->last_avail_idx -= hdr->num_buffers - 1; + error = VHOST_USER_TX_FUNC_ERROR_MMAP_FAIL; + goto done; + } + } + else if (vui->virtio_net_hdr_sz == 12) + { //MRG is available + + //Move from available to used buffer + rxvq->used->ring[used_index & qsz_mask].id = + desc_chain_head; + rxvq->used->ring[used_index & qsz_mask].len = desc_len; + vhost_user_log_dirty_ring (vui, rxvq, + ring[used_index & qsz_mask]); + rxvq->last_avail_idx++; + used_index++; + hdr->num_buffers++; + + if (PREDICT_FALSE + (rxvq->last_avail_idx == rxvq->avail->idx)) + { + //Dequeue queued descriptors for this packet + used_index -= hdr->num_buffers - 1; + rxvq->last_avail_idx -= hdr->num_buffers - 1; + error = VHOST_USER_TX_FUNC_ERROR_PKT_DROP_NOBUF; + goto done; + } + + //Look at next one + desc_chain_head = + rxvq->avail->ring[rxvq->last_avail_idx & qsz_mask]; + desc_current = desc_chain_head; + desc_len = 0; + offset = 0; + if (PREDICT_FALSE + (!(buffer_addr = + map_guest_mem (vui, rxvq->desc[desc_current].addr)))) + { + //Dequeue queued descriptors for this packet + used_index -= hdr->num_buffers - 1; + rxvq->last_avail_idx -= hdr->num_buffers - 1; + error = VHOST_USER_TX_FUNC_ERROR_MMAP_FAIL; + goto done; + } + } + else + { + error = VHOST_USER_TX_FUNC_ERROR_PKT_DROP_NOBUF; + goto done; + } + } + + u16 bytes_to_copy = + bytes_left > + (rxvq->desc[desc_current].len - + offset) ? (rxvq->desc[desc_current].len - offset) : bytes_left; + clib_memcpy (buffer_addr, + vlib_buffer_get_current (current_b0) + + current_b0->current_length - bytes_left, + bytes_to_copy); + + vhost_user_log_dirty_pages (vui, + rxvq->desc[desc_current].addr + offset, + bytes_to_copy); + bytes_left -= bytes_to_copy; + offset += bytes_to_copy; + buffer_addr += bytes_to_copy; + desc_len += bytes_to_copy; + } //Move from available to used ring rxvq->used->ring[used_index & qsz_mask].id = desc_chain_head; rxvq->used->ring[used_index & qsz_mask].len = desc_len; - vhost_user_log_dirty_ring(vui, rxvq, ring[used_index & qsz_mask]); + vhost_user_log_dirty_ring (vui, rxvq, ring[used_index & qsz_mask]); rxvq->last_avail_idx++; used_index++; - } + } done: - CLIB_MEMORY_BARRIER(); + CLIB_MEMORY_BARRIER (); rxvq->used->idx = used_index; - vhost_user_log_dirty_ring(vui, rxvq, idx); + vhost_user_log_dirty_ring (vui, rxvq, idx); /* interrupt (call) handling */ - if((rxvq->callfd > 0) && !(rxvq->avail->flags & 1)) { - rxvq->n_since_last_int += n_packets - n_left; + if ((rxvq->callfd > 0) && !(rxvq->avail->flags & 1)) + { + rxvq->n_since_last_int += n_packets - n_left; - if(rxvq->n_since_last_int > vum->coalesce_frames) - vhost_user_send_call(vm, rxvq); - } + if (rxvq->n_since_last_int > vum->coalesce_frames) + vhost_user_send_call (vm, rxvq); + } done2: - if (PREDICT_FALSE(vui->lockp != 0)) - *vui->lockp = 0; - - if (PREDICT_FALSE(n_left && error != VHOST_USER_TX_FUNC_ERROR_NONE)) { - vlib_error_count(vm, node->node_index, error, n_left); - vlib_increment_simple_counter - (vnet_main.interface_main.sw_if_counters - + VNET_INTERFACE_COUNTER_DROP, - os_get_cpu_number(), - vui->sw_if_index, - n_left); - } + if (PREDICT_FALSE (vui->lockp != 0)) + *vui->lockp = 0; + + if (PREDICT_FALSE (n_left && error != VHOST_USER_TX_FUNC_ERROR_NONE)) + { + vlib_error_count (vm, node->node_index, error, n_left); + vlib_increment_simple_counter + (vnet_main.interface_main.sw_if_counters + + VNET_INTERFACE_COUNTER_DROP, + os_get_cpu_number (), vui->sw_if_index, n_left); + } vlib_buffer_free (vm, vlib_frame_args (frame), frame->n_vectors); return frame->n_vectors; } static clib_error_t * -vhost_user_interface_admin_up_down (vnet_main_t * vnm, u32 hw_if_index, u32 flags) +vhost_user_interface_admin_up_down (vnet_main_t * vnm, u32 hw_if_index, + u32 flags) { - vnet_hw_interface_t * hif = vnet_get_hw_interface (vnm, hw_if_index); + vnet_hw_interface_t *hif = vnet_get_hw_interface (vnm, hw_if_index); uword is_up = (flags & VNET_SW_INTERFACE_FLAG_ADMIN_UP) != 0; - vhost_user_main_t * vum = &vhost_user_main; - vhost_user_intf_t * vui = vec_elt_at_index (vum->vhost_user_interfaces, hif->dev_instance); + vhost_user_main_t *vum = &vhost_user_main; + vhost_user_intf_t *vui = + vec_elt_at_index (vum->vhost_user_interfaces, hif->dev_instance); vui->admin_up = is_up; if (is_up) vnet_hw_interface_set_flags (vnm, vui->hw_if_index, - VNET_HW_INTERFACE_FLAG_LINK_UP); + VNET_HW_INTERFACE_FLAG_LINK_UP); return /* no error */ 0; } +/* *INDENT-OFF* */ VNET_DEVICE_CLASS (vhost_user_dev_class,static) = { .name = "vhost-user", .tx_function = vhost_user_intfc_tx, @@ -1380,100 +1528,116 @@ VNET_DEVICE_CLASS (vhost_user_dev_class,static) = { VLIB_DEVICE_TX_FUNCTION_MULTIARCH (vhost_user_dev_class, vhost_user_intfc_tx) +/* *INDENT-ON* */ static uword vhost_user_process (vlib_main_t * vm, - vlib_node_runtime_t * rt, - vlib_frame_t * f) + vlib_node_runtime_t * rt, vlib_frame_t * f) { - vhost_user_main_t * vum = &vhost_user_main; - vhost_user_intf_t * vui; - struct sockaddr_un sun; - int sockfd; - unix_file_t template = {0}; - f64 timeout = 3153600000.0 /* 100 years */; - uword *event_data = 0; + vhost_user_main_t *vum = &vhost_user_main; + vhost_user_intf_t *vui; + struct sockaddr_un sun; + int sockfd; + unix_file_t template = { 0 }; + f64 timeout = 3153600000.0 /* 100 years */ ; + uword *event_data = 0; + + sockfd = socket (AF_UNIX, SOCK_STREAM, 0); + sun.sun_family = AF_UNIX; + template.read_function = vhost_user_socket_read; + template.error_function = vhost_user_socket_error; - sockfd = socket(AF_UNIX, SOCK_STREAM, 0); - sun.sun_family = AF_UNIX; - template.read_function = vhost_user_socket_read; - template.error_function = vhost_user_socket_error; + if (sockfd < 0) + return 0; - if (sockfd < 0) - return 0; + while (1) + { + vlib_process_wait_for_event_or_clock (vm, timeout); + vlib_process_get_events (vm, &event_data); + vec_reset_length (event_data); + + timeout = 3.0; - while (1) { - vlib_process_wait_for_event_or_clock (vm, timeout); - vlib_process_get_events (vm, &event_data); - vec_reset_length (event_data); - - timeout = 3.0; - - vec_foreach (vui, vum->vhost_user_interfaces) { - - if (vui->sock_is_server || !vui->active) - continue; - - if (vui->unix_fd == -1) { - /* try to connect */ - - strncpy(sun.sun_path, (char *) vui->sock_filename, sizeof(sun.sun_path) - 1); - - if (connect(sockfd, (struct sockaddr *) &sun, sizeof(struct sockaddr_un)) == 0) { - vui->sock_errno = 0; - vui->unix_fd = sockfd; - template.file_descriptor = sockfd; - vui->unix_file_index = unix_file_add (&unix_main, &template); - hash_set (vum->vhost_user_interface_index_by_sock_fd, sockfd, vui - vum->vhost_user_interfaces); - - sockfd = socket(AF_UNIX, SOCK_STREAM, 0); - if (sockfd < 0) - return 0; - } - else { - vui->sock_errno = errno; - } - } else { - /* check if socket is alive */ - int error = 0; - socklen_t len = sizeof (error); - int retval = getsockopt(vui->unix_fd, SOL_SOCKET, SO_ERROR, &error, &len); - - if (retval) - vhost_user_if_disconnect(vui); - } - } + vec_foreach (vui, vum->vhost_user_interfaces) + { + + if (vui->sock_is_server || !vui->active) + continue; + + if (vui->unix_fd == -1) + { + /* try to connect */ + + strncpy (sun.sun_path, (char *) vui->sock_filename, + sizeof (sun.sun_path) - 1); + + if (connect + (sockfd, (struct sockaddr *) &sun, + sizeof (struct sockaddr_un)) == 0) + { + vui->sock_errno = 0; + vui->unix_fd = sockfd; + template.file_descriptor = sockfd; + vui->unix_file_index = unix_file_add (&unix_main, &template); + hash_set (vum->vhost_user_interface_index_by_sock_fd, sockfd, + vui - vum->vhost_user_interfaces); + + sockfd = socket (AF_UNIX, SOCK_STREAM, 0); + if (sockfd < 0) + return 0; + } + else + { + vui->sock_errno = errno; + } + } + else + { + /* check if socket is alive */ + int error = 0; + socklen_t len = sizeof (error); + int retval = + getsockopt (vui->unix_fd, SOL_SOCKET, SO_ERROR, &error, &len); + + if (retval) + vhost_user_if_disconnect (vui); + } + } } - return 0; + return 0; } +/* *INDENT-OFF* */ VLIB_REGISTER_NODE (vhost_user_process_node,static) = { .function = vhost_user_process, .type = VLIB_NODE_TYPE_PROCESS, .name = "vhost-user-process", }; +/* *INDENT-ON* */ -int vhost_user_delete_if(vnet_main_t * vnm, vlib_main_t * vm, - u32 sw_if_index) +int +vhost_user_delete_if (vnet_main_t * vnm, vlib_main_t * vm, u32 sw_if_index) { - vhost_user_main_t * vum = &vhost_user_main; - vhost_user_intf_t * vui; + vhost_user_main_t *vum = &vhost_user_main; + vhost_user_intf_t *vui; uword *p = NULL; int rv = 0; - p = hash_get (vum->vhost_user_interface_index_by_sw_if_index, - sw_if_index); - if (p == 0) { - return VNET_API_ERROR_INVALID_SW_IF_INDEX; - } else { - vui = vec_elt_at_index (vum->vhost_user_interfaces, p[0]); - } + p = hash_get (vum->vhost_user_interface_index_by_sw_if_index, sw_if_index); + if (p == 0) + { + return VNET_API_ERROR_INVALID_SW_IF_INDEX; + } + else + { + vui = vec_elt_at_index (vum->vhost_user_interfaces, p[0]); + } // interface is inactive vui->active = 0; // disconnect interface sockets - vhost_user_if_disconnect(vui); + vhost_user_if_disconnect (vui); // add to inactive interface list vec_add1 (vum->vhost_user_inactive_interfaces_index, p[0]); @@ -1488,35 +1652,40 @@ int vhost_user_delete_if(vnet_main_t * vnm, vlib_main_t * vm, } // init server socket on specified sock_filename -static int vhost_user_init_server_sock(const char * sock_filename, int *sockfd) +static int +vhost_user_init_server_sock (const char *sock_filename, int *sockfd) { int rv = 0; - struct sockaddr_un un = {}; + struct sockaddr_un un = { }; int fd; /* create listening socket */ - fd = socket(AF_UNIX, SOCK_STREAM, 0); + fd = socket (AF_UNIX, SOCK_STREAM, 0); - if (fd < 0) { - return VNET_API_ERROR_SYSCALL_ERROR_1; - } + if (fd < 0) + { + return VNET_API_ERROR_SYSCALL_ERROR_1; + } un.sun_family = AF_UNIX; - strncpy((char *) un.sun_path, (char *) sock_filename, sizeof(un.sun_path) - 1); + strncpy ((char *) un.sun_path, (char *) sock_filename, + sizeof (un.sun_path) - 1); /* remove if exists */ - unlink( (char *) sock_filename); + unlink ((char *) sock_filename); - if (bind(fd, (struct sockaddr *) &un, sizeof(un)) == -1) { - rv = VNET_API_ERROR_SYSCALL_ERROR_2; - goto error; - } + if (bind (fd, (struct sockaddr *) &un, sizeof (un)) == -1) + { + rv = VNET_API_ERROR_SYSCALL_ERROR_2; + goto error; + } - if (listen(fd, 1) == -1) { - rv = VNET_API_ERROR_SYSCALL_ERROR_3; - goto error; - } + if (listen (fd, 1) == -1) + { + rv = VNET_API_ERROR_SYSCALL_ERROR_3; + goto error; + } - unix_file_t template = {0}; + unix_file_t template = { 0 }; template.read_function = vhost_user_socksvr_accept_ready; template.file_descriptor = fd; unix_file_add (&unix_main, &template); @@ -1524,27 +1693,32 @@ static int vhost_user_init_server_sock(const char * sock_filename, int *sockfd) return rv; error: - close(fd); + close (fd); return rv; } // get new vhost_user_intf_t from inactive interfaces or create new one -static vhost_user_intf_t *vhost_user_vui_new() +static vhost_user_intf_t * +vhost_user_vui_new () { - vhost_user_main_t * vum = &vhost_user_main; - vhost_user_intf_t * vui = NULL; - int inactive_cnt = vec_len(vum->vhost_user_inactive_interfaces_index); + vhost_user_main_t *vum = &vhost_user_main; + vhost_user_intf_t *vui = NULL; + int inactive_cnt = vec_len (vum->vhost_user_inactive_interfaces_index); // if there are any inactive ifaces - if (inactive_cnt > 0) { - // take last - u32 vui_idx = vum->vhost_user_inactive_interfaces_index[inactive_cnt - 1]; - if (vec_len(vum->vhost_user_interfaces) > vui_idx) { - vui = vec_elt_at_index (vum->vhost_user_interfaces, vui_idx); - DBG_SOCK("reusing inactive vhost-user interface index %d", vui_idx); + if (inactive_cnt > 0) + { + // take last + u32 vui_idx = + vum->vhost_user_inactive_interfaces_index[inactive_cnt - 1]; + if (vec_len (vum->vhost_user_interfaces) > vui_idx) + { + vui = vec_elt_at_index (vum->vhost_user_interfaces, vui_idx); + DBG_SOCK ("reusing inactive vhost-user interface index %d", + vui_idx); + } + // "remove" from inactive list + _vec_len (vum->vhost_user_inactive_interfaces_index) -= 1; } - // "remove" from inactive list - _vec_len(vum->vhost_user_inactive_interfaces_index) -= 1; - } // vui was not retrieved from inactive ifaces - create new if (!vui) @@ -1553,34 +1727,37 @@ static vhost_user_intf_t *vhost_user_vui_new() } // create ethernet interface for vhost user intf -static void vhost_user_create_ethernet(vnet_main_t * vnm, vlib_main_t * vm, - vhost_user_intf_t *vui, u8 *hwaddress) +static void +vhost_user_create_ethernet (vnet_main_t * vnm, vlib_main_t * vm, + vhost_user_intf_t * vui, u8 * hwaddress) { - vhost_user_main_t * vum = &vhost_user_main; + vhost_user_main_t *vum = &vhost_user_main; u8 hwaddr[6]; - clib_error_t * error; + clib_error_t *error; /* create hw and sw interface */ - if (hwaddress) { - clib_memcpy(hwaddr, hwaddress, 6); - } else { - f64 now = vlib_time_now(vm); - u32 rnd; - rnd = (u32) (now * 1e6); - rnd = random_u32 (&rnd); - - clib_memcpy (hwaddr+2, &rnd, sizeof(rnd)); - hwaddr[0] = 2; - hwaddr[1] = 0xfe; - } + if (hwaddress) + { + clib_memcpy (hwaddr, hwaddress, 6); + } + else + { + f64 now = vlib_time_now (vm); + u32 rnd; + rnd = (u32) (now * 1e6); + rnd = random_u32 (&rnd); + + clib_memcpy (hwaddr + 2, &rnd, sizeof (rnd)); + hwaddr[0] = 2; + hwaddr[1] = 0xfe; + } error = ethernet_register_interface (vnm, vhost_user_dev_class.index, - vui - vum->vhost_user_interfaces /* device instance */, - hwaddr /* ethernet address */, - &vui->hw_if_index, - 0 /* flag change */); + vui - vum->vhost_user_interfaces /* device instance */ , + hwaddr /* ethernet address */ , + &vui->hw_if_index, 0 /* flag change */ ); if (error) clib_error_report (error); @@ -1589,22 +1766,23 @@ static void vhost_user_create_ethernet(vnet_main_t * vnm, vlib_main_t * vm, } // initialize vui with specified attributes -static void vhost_user_vui_init(vnet_main_t * vnm, - vhost_user_intf_t *vui, int sockfd, - const char * sock_filename, - u8 is_server, u64 feature_mask, - u32 * sw_if_index) +static void +vhost_user_vui_init (vnet_main_t * vnm, + vhost_user_intf_t * vui, int sockfd, + const char *sock_filename, + u8 is_server, u64 feature_mask, u32 * sw_if_index) { - vnet_sw_interface_t * sw; + vnet_sw_interface_t *sw; sw = vnet_get_hw_sw_interface (vnm, vui->hw_if_index); - vlib_thread_main_t * tm = vlib_get_thread_main(); + vlib_thread_main_t *tm = vlib_get_thread_main (); int q; vui->unix_fd = sockfd; vui->sw_if_index = sw->sw_if_index; vui->num_vrings = 2; vui->sock_is_server = is_server; - strncpy(vui->sock_filename, sock_filename, ARRAY_LEN(vui->sock_filename)-1); + strncpy (vui->sock_filename, sock_filename, + ARRAY_LEN (vui->sock_filename) - 1); vui->sock_errno = 0; vui->is_up = 0; vui->feature_mask = feature_mask; @@ -1612,128 +1790,138 @@ static void vhost_user_vui_init(vnet_main_t * vnm, vui->unix_file_index = ~0; vui->log_base_addr = 0; - for (q = 0; q < 2; q++) { - vui->vrings[q].enabled = 0; - } + for (q = 0; q < 2; q++) + { + vui->vrings[q].enabled = 0; + } - vnet_hw_interface_set_flags (vnm, vui->hw_if_index, 0); + vnet_hw_interface_set_flags (vnm, vui->hw_if_index, 0); if (sw_if_index) - *sw_if_index = vui->sw_if_index; + *sw_if_index = vui->sw_if_index; if (tm->n_vlib_mains > 1) - { - vui->lockp = clib_mem_alloc_aligned (CLIB_CACHE_LINE_BYTES, - CLIB_CACHE_LINE_BYTES); - memset ((void *) vui->lockp, 0, CLIB_CACHE_LINE_BYTES); - } + { + vui->lockp = clib_mem_alloc_aligned (CLIB_CACHE_LINE_BYTES, + CLIB_CACHE_LINE_BYTES); + memset ((void *) vui->lockp, 0, CLIB_CACHE_LINE_BYTES); + } } // register vui and start polling on it -static void vhost_user_vui_register(vlib_main_t * vm, vhost_user_intf_t *vui) +static void +vhost_user_vui_register (vlib_main_t * vm, vhost_user_intf_t * vui) { - vhost_user_main_t * vum = &vhost_user_main; - dpdk_main_t * dm = &dpdk_main; + vhost_user_main_t *vum = &vhost_user_main; + dpdk_main_t *dm = &dpdk_main; int cpu_index; - vlib_thread_main_t * tm = vlib_get_thread_main(); + vlib_thread_main_t *tm = vlib_get_thread_main (); hash_set (vum->vhost_user_interface_index_by_listener_fd, vui->unix_fd, - vui - vum->vhost_user_interfaces); + vui - vum->vhost_user_interfaces); hash_set (vum->vhost_user_interface_index_by_sw_if_index, vui->sw_if_index, - vui - vum->vhost_user_interfaces); + vui - vum->vhost_user_interfaces); /* start polling */ cpu_index = dm->input_cpu_first_index + - (vui - vum->vhost_user_interfaces) % dm->input_cpu_count; + (vui - vum->vhost_user_interfaces) % dm->input_cpu_count; if (tm->n_vlib_mains == 1) vlib_node_set_state (vm, vhost_user_input_node.index, - VLIB_NODE_STATE_POLLING); + VLIB_NODE_STATE_POLLING); else vlib_node_set_state (vlib_mains[cpu_index], vhost_user_input_node.index, - VLIB_NODE_STATE_POLLING); + VLIB_NODE_STATE_POLLING); /* tell process to start polling for sockets */ - vlib_process_signal_event(vm, vhost_user_process_node.index, 0, 0); + vlib_process_signal_event (vm, vhost_user_process_node.index, 0, 0); } -int vhost_user_create_if(vnet_main_t * vnm, vlib_main_t * vm, - const char * sock_filename, - u8 is_server, - u32 * sw_if_index, - u64 feature_mask, - u8 renumber, u32 custom_dev_instance, - u8 *hwaddr) +int +vhost_user_create_if (vnet_main_t * vnm, vlib_main_t * vm, + const char *sock_filename, + u8 is_server, + u32 * sw_if_index, + u64 feature_mask, + u8 renumber, u32 custom_dev_instance, u8 * hwaddr) { - vhost_user_intf_t * vui = NULL; + vhost_user_intf_t *vui = NULL; u32 sw_if_idx = ~0; int sockfd = -1; int rv = 0; - if (is_server) { - if ((rv = vhost_user_init_server_sock (sock_filename, &sockfd)) != 0) { - return rv; - } - } + if (is_server) + { + if ((rv = vhost_user_init_server_sock (sock_filename, &sockfd)) != 0) + { + return rv; + } + } vui = vhost_user_vui_new (); - ASSERT(vui != NULL); + ASSERT (vui != NULL); vhost_user_create_ethernet (vnm, vm, vui, hwaddr); vhost_user_vui_init (vnm, vui, sockfd, sock_filename, is_server, - feature_mask, &sw_if_idx); + feature_mask, &sw_if_idx); - if (renumber) { - vnet_interface_name_renumber (sw_if_idx, custom_dev_instance); - } + if (renumber) + { + vnet_interface_name_renumber (sw_if_idx, custom_dev_instance); + } vhost_user_vui_register (vm, vui); if (sw_if_index) - *sw_if_index = sw_if_idx; + *sw_if_index = sw_if_idx; return rv; } -int vhost_user_modify_if(vnet_main_t * vnm, vlib_main_t * vm, - const char * sock_filename, - u8 is_server, - u32 sw_if_index, - u64 feature_mask, - u8 renumber, u32 custom_dev_instance) +int +vhost_user_modify_if (vnet_main_t * vnm, vlib_main_t * vm, + const char *sock_filename, + u8 is_server, + u32 sw_if_index, + u64 feature_mask, u8 renumber, u32 custom_dev_instance) { - vhost_user_main_t * vum = &vhost_user_main; - vhost_user_intf_t * vui = NULL; + vhost_user_main_t *vum = &vhost_user_main; + vhost_user_intf_t *vui = NULL; u32 sw_if_idx = ~0; int sockfd = -1; int rv = 0; uword *p = NULL; - p = hash_get (vum->vhost_user_interface_index_by_sw_if_index, - sw_if_index); - if (p == 0) { - return VNET_API_ERROR_INVALID_SW_IF_INDEX; - } else { - vui = vec_elt_at_index (vum->vhost_user_interfaces, p[0]); - } + p = hash_get (vum->vhost_user_interface_index_by_sw_if_index, sw_if_index); + if (p == 0) + { + return VNET_API_ERROR_INVALID_SW_IF_INDEX; + } + else + { + vui = vec_elt_at_index (vum->vhost_user_interfaces, p[0]); + } // interface is inactive vui->active = 0; // disconnect interface sockets - vhost_user_if_disconnect(vui); + vhost_user_if_disconnect (vui); - if (is_server) { - if ((rv = vhost_user_init_server_sock (sock_filename, &sockfd)) != 0) { - return rv; - } - } + if (is_server) + { + if ((rv = vhost_user_init_server_sock (sock_filename, &sockfd)) != 0) + { + return rv; + } + } vhost_user_vui_init (vnm, vui, sockfd, sock_filename, is_server, - feature_mask, &sw_if_idx); + feature_mask, &sw_if_idx); - if (renumber) { - vnet_interface_name_renumber (sw_if_idx, custom_dev_instance); - } + if (renumber) + { + vnet_interface_name_renumber (sw_if_idx, custom_dev_instance); + } vhost_user_vui_register (vm, vui); @@ -1742,124 +1930,135 @@ int vhost_user_modify_if(vnet_main_t * vnm, vlib_main_t * vm, clib_error_t * vhost_user_connect_command_fn (vlib_main_t * vm, - unformat_input_t * input, - vlib_cli_command_t * cmd) + unformat_input_t * input, + vlib_cli_command_t * cmd) { - unformat_input_t _line_input, * line_input = &_line_input; - u8 * sock_filename = NULL; + unformat_input_t _line_input, *line_input = &_line_input; + u8 *sock_filename = NULL; u32 sw_if_index; u8 is_server = 0; - u64 feature_mask = (u64)~0; + u64 feature_mask = (u64) ~ 0; u8 renumber = 0; u32 custom_dev_instance = ~0; u8 hwaddr[6]; u8 *hw = NULL; /* Get a line of input. */ - if (! unformat_user (input, unformat_line_input, line_input)) + if (!unformat_user (input, unformat_line_input, line_input)) return 0; - while (unformat_check_input (line_input) != UNFORMAT_END_OF_INPUT) { - if (unformat (line_input, "socket %s", &sock_filename)) - ; - else if (unformat (line_input, "server")) - is_server = 1; - else if (unformat (line_input, "feature-mask 0x%llx", &feature_mask)) - ; - else if (unformat (line_input, "hwaddr %U", unformat_ethernet_address, hwaddr)) - hw = hwaddr; - else if (unformat (line_input, "renumber %d", &custom_dev_instance)) { - renumber = 1; + while (unformat_check_input (line_input) != UNFORMAT_END_OF_INPUT) + { + if (unformat (line_input, "socket %s", &sock_filename)) + ; + else if (unformat (line_input, "server")) + is_server = 1; + else if (unformat (line_input, "feature-mask 0x%llx", &feature_mask)) + ; + else + if (unformat + (line_input, "hwaddr %U", unformat_ethernet_address, hwaddr)) + hw = hwaddr; + else if (unformat (line_input, "renumber %d", &custom_dev_instance)) + { + renumber = 1; + } + else + return clib_error_return (0, "unknown input `%U'", + format_unformat_error, input); } - else - return clib_error_return (0, "unknown input `%U'", - format_unformat_error, input); - } unformat_free (line_input); - vnet_main_t *vnm = vnet_get_main(); + vnet_main_t *vnm = vnet_get_main (); int rv; - if ((rv = vhost_user_create_if(vnm, vm, (char *)sock_filename, - is_server, &sw_if_index, feature_mask, - renumber, custom_dev_instance, hw))) { - vec_free(sock_filename); + if ((rv = vhost_user_create_if (vnm, vm, (char *) sock_filename, + is_server, &sw_if_index, feature_mask, + renumber, custom_dev_instance, hw))) + { + vec_free (sock_filename); return clib_error_return (0, "vhost_user_create_if returned %d", rv); - } + } - vec_free(sock_filename); - vlib_cli_output(vm, "%U\n", format_vnet_sw_if_index_name, vnet_get_main(), sw_if_index); + vec_free (sock_filename); + vlib_cli_output (vm, "%U\n", format_vnet_sw_if_index_name, vnet_get_main (), + sw_if_index); return 0; } clib_error_t * vhost_user_delete_command_fn (vlib_main_t * vm, - unformat_input_t * input, - vlib_cli_command_t * cmd) + unformat_input_t * input, + vlib_cli_command_t * cmd) { - unformat_input_t _line_input, * line_input = &_line_input; + unformat_input_t _line_input, *line_input = &_line_input; u32 sw_if_index = ~0; /* Get a line of input. */ - if (! unformat_user (input, unformat_line_input, line_input)) + if (!unformat_user (input, unformat_line_input, line_input)) return 0; - while (unformat_check_input (line_input) != UNFORMAT_END_OF_INPUT) { - if (unformat (line_input, "sw_if_index %d", &sw_if_index)) - ; - else - return clib_error_return (0, "unknown input `%U'", - format_unformat_error, input); - } + while (unformat_check_input (line_input) != UNFORMAT_END_OF_INPUT) + { + if (unformat (line_input, "sw_if_index %d", &sw_if_index)) + ; + else + return clib_error_return (0, "unknown input `%U'", + format_unformat_error, input); + } unformat_free (line_input); - vnet_main_t *vnm = vnet_get_main(); + vnet_main_t *vnm = vnet_get_main (); - vhost_user_delete_if(vnm, vm, sw_if_index); + vhost_user_delete_if (vnm, vm, sw_if_index); return 0; } -int vhost_user_dump_ifs(vnet_main_t * vnm, vlib_main_t * vm, vhost_user_intf_details_t **out_vuids) +int +vhost_user_dump_ifs (vnet_main_t * vnm, vlib_main_t * vm, + vhost_user_intf_details_t ** out_vuids) { int rv = 0; - vhost_user_main_t * vum = &vhost_user_main; - vhost_user_intf_t * vui; - vhost_user_intf_details_t * r_vuids = NULL; - vhost_user_intf_details_t * vuid = NULL; - u32 * hw_if_indices = 0; - vnet_hw_interface_t * hi; + vhost_user_main_t *vum = &vhost_user_main; + vhost_user_intf_t *vui; + vhost_user_intf_details_t *r_vuids = NULL; + vhost_user_intf_details_t *vuid = NULL; + u32 *hw_if_indices = 0; + vnet_hw_interface_t *hi; u8 *s = NULL; int i; if (!out_vuids) - return -1; + return -1; - vec_foreach (vui, vum->vhost_user_interfaces) { + vec_foreach (vui, vum->vhost_user_interfaces) + { if (vui->active) - vec_add1(hw_if_indices, vui->hw_if_index); + vec_add1 (hw_if_indices, vui->hw_if_index); } - for (i = 0; i < vec_len (hw_if_indices); i++) { - hi = vnet_get_hw_interface (vnm, hw_if_indices[i]); - vui = vec_elt_at_index (vum->vhost_user_interfaces, hi->dev_instance); - - vec_add2(r_vuids, vuid, 1); - vuid->sw_if_index = vui->sw_if_index; - vuid->virtio_net_hdr_sz = vui->virtio_net_hdr_sz; - vuid->features = vui->features; - vuid->is_server = vui->sock_is_server; - vuid->num_regions = vui->nregions; - vuid->sock_errno = vui->sock_errno; - strncpy((char *)vuid->sock_filename, (char *)vui->sock_filename, - ARRAY_LEN(vuid->sock_filename)-1); - - s = format (s, "%v%c", hi->name, 0); - - strncpy((char *)vuid->if_name, (char *)s, - ARRAY_LEN(vuid->if_name)-1); - _vec_len(s) = 0; - } + for (i = 0; i < vec_len (hw_if_indices); i++) + { + hi = vnet_get_hw_interface (vnm, hw_if_indices[i]); + vui = vec_elt_at_index (vum->vhost_user_interfaces, hi->dev_instance); + + vec_add2 (r_vuids, vuid, 1); + vuid->sw_if_index = vui->sw_if_index; + vuid->virtio_net_hdr_sz = vui->virtio_net_hdr_sz; + vuid->features = vui->features; + vuid->is_server = vui->sock_is_server; + vuid->num_regions = vui->nregions; + vuid->sock_errno = vui->sock_errno; + strncpy ((char *) vuid->sock_filename, (char *) vui->sock_filename, + ARRAY_LEN (vuid->sock_filename) - 1); + + s = format (s, "%v%c", hi->name, 0); + + strncpy ((char *) vuid->if_name, (char *) s, + ARRAY_LEN (vuid->if_name) - 1); + _vec_len (s) = 0; + } vec_free (s); vec_free (hw_if_indices); @@ -1871,124 +2070,152 @@ int vhost_user_dump_ifs(vnet_main_t * vnm, vlib_main_t * vm, vhost_user_intf_det clib_error_t * show_vhost_user_command_fn (vlib_main_t * vm, - unformat_input_t * input, - vlib_cli_command_t * cmd) + unformat_input_t * input, + vlib_cli_command_t * cmd) { - clib_error_t * error = 0; - vnet_main_t * vnm = vnet_get_main(); - vhost_user_main_t * vum = &vhost_user_main; - vhost_user_intf_t * vui; - u32 hw_if_index, * hw_if_indices = 0; - vnet_hw_interface_t * hi; + clib_error_t *error = 0; + vnet_main_t *vnm = vnet_get_main (); + vhost_user_main_t *vum = &vhost_user_main; + vhost_user_intf_t *vui; + u32 hw_if_index, *hw_if_indices = 0; + vnet_hw_interface_t *hi; int i, j, q; int show_descr = 0; - struct feat_struct { u8 bit; char *str;}; + struct feat_struct + { + u8 bit; + char *str; + }; struct feat_struct *feat_entry; static struct feat_struct feat_array[] = { #define _(s,b) { .str = #s, .bit = b, }, - foreach_virtio_net_feature + foreach_virtio_net_feature #undef _ - { .str = NULL } + {.str = NULL} }; - while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT) { - if (unformat (input, "%U", unformat_vnet_hw_interface, vnm, &hw_if_index)) { - vec_add1 (hw_if_indices, hw_if_index); - vlib_cli_output(vm, "add %d", hw_if_index); - } - else if (unformat (input, "descriptors") || unformat (input, "desc") ) - show_descr = 1; - else { - error = clib_error_return (0, "unknown input `%U'", - format_unformat_error, input); - goto done; + while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT) + { + if (unformat + (input, "%U", unformat_vnet_hw_interface, vnm, &hw_if_index)) + { + vec_add1 (hw_if_indices, hw_if_index); + vlib_cli_output (vm, "add %d", hw_if_index); + } + else if (unformat (input, "descriptors") || unformat (input, "desc")) + show_descr = 1; + else + { + error = clib_error_return (0, "unknown input `%U'", + format_unformat_error, input); + goto done; + } } - } - if (vec_len (hw_if_indices) == 0) { - vec_foreach (vui, vum->vhost_user_interfaces) { - if (vui->active) - vec_add1(hw_if_indices, vui->hw_if_index); + if (vec_len (hw_if_indices) == 0) + { + vec_foreach (vui, vum->vhost_user_interfaces) + { + if (vui->active) + vec_add1 (hw_if_indices, vui->hw_if_index); + } } - } vlib_cli_output (vm, "Virtio vhost-user interfaces"); vlib_cli_output (vm, "Global:\n coalesce frames %d time %e\n\n", - vum->coalesce_frames, vum->coalesce_time); - - for (i = 0; i < vec_len (hw_if_indices); i++) { - hi = vnet_get_hw_interface (vnm, hw_if_indices[i]); - vui = vec_elt_at_index (vum->vhost_user_interfaces, hi->dev_instance); - vlib_cli_output (vm, "Interface: %s (ifindex %d)", - hi->name, hw_if_indices[i]); - - vlib_cli_output (vm, "virtio_net_hdr_sz %d\n features (0x%llx): \n", - vui->virtio_net_hdr_sz, vui->features); - - feat_entry = (struct feat_struct *) &feat_array; - while(feat_entry->str) { - if (vui->features & (1 << feat_entry->bit)) - vlib_cli_output (vm, " %s (%d)", feat_entry->str, feat_entry->bit); - feat_entry++; - } - - vlib_cli_output (vm, "\n"); - + vum->coalesce_frames, vum->coalesce_time); - vlib_cli_output (vm, " socket filename %s type %s errno \"%s\"\n\n", - vui->sock_filename, vui->sock_is_server ? "server" : "client", - strerror(vui->sock_errno)); - - vlib_cli_output (vm, " Memory regions (total %d)\n", vui->nregions); - - if (vui->nregions){ - vlib_cli_output(vm, " region fd guest_phys_addr memory_size userspace_addr mmap_offset mmap_addr\n"); - vlib_cli_output(vm, " ====== ===== ================== ================== ================== ================== ==================\n"); - } - for (j = 0; j < vui->nregions; j++) { - vlib_cli_output(vm, " %d %-5d 0x%016lx 0x%016lx 0x%016lx 0x%016lx 0x%016lx\n", j, - vui->region_mmap_fd[j], - vui->regions[j].guest_phys_addr, - vui->regions[j].memory_size, - vui->regions[j].userspace_addr, - vui->regions[j].mmap_offset, - pointer_to_uword( vui->region_mmap_addr[j]) ); - } - for (q = 0; q < vui->num_vrings; q++) { - vlib_cli_output(vm, "\n Virtqueue %d\n", q); - - vlib_cli_output(vm, " qsz %d last_avail_idx %d last_used_idx %d\n", - vui->vrings[q].qsz, - vui->vrings[q].last_avail_idx, - vui->vrings[q].last_used_idx); - - if (vui->vrings[q].avail && vui->vrings[q].used) - vlib_cli_output(vm, " avail.flags %x avail.idx %d used.flags %x used.idx %d\n", - vui->vrings[q].avail->flags, - vui->vrings[q].avail->idx, - vui->vrings[q].used->flags, - vui->vrings[q].used->idx); - - vlib_cli_output(vm, " kickfd %d callfd %d errfd %d\n", - vui->vrings[q].kickfd, - vui->vrings[q].callfd, - vui->vrings[q].errfd); - - if (show_descr) { - vlib_cli_output(vm, "\n descriptor table:\n"); - vlib_cli_output(vm, " id addr len flags next user_addr\n"); - vlib_cli_output(vm, " ===== ================== ===== ====== ===== ==================\n"); - for(j = 0; j < vui->vrings[q].qsz; j++) { - vlib_cli_output(vm, " %-5d 0x%016lx %-5d 0x%04x %-5d 0x%016lx\n", - j, - vui->vrings[q].desc[j].addr, - vui->vrings[q].desc[j].len, - vui->vrings[q].desc[j].flags, - vui->vrings[q].desc[j].next, - pointer_to_uword(map_guest_mem(vui, vui->vrings[q].desc[j].addr)));} - } + for (i = 0; i < vec_len (hw_if_indices); i++) + { + hi = vnet_get_hw_interface (vnm, hw_if_indices[i]); + vui = vec_elt_at_index (vum->vhost_user_interfaces, hi->dev_instance); + vlib_cli_output (vm, "Interface: %s (ifindex %d)", + hi->name, hw_if_indices[i]); + + vlib_cli_output (vm, "virtio_net_hdr_sz %d\n features (0x%llx): \n", + vui->virtio_net_hdr_sz, vui->features); + + feat_entry = (struct feat_struct *) &feat_array; + while (feat_entry->str) + { + if (vui->features & (1 << feat_entry->bit)) + vlib_cli_output (vm, " %s (%d)", feat_entry->str, + feat_entry->bit); + feat_entry++; + } + + vlib_cli_output (vm, "\n"); + + + vlib_cli_output (vm, " socket filename %s type %s errno \"%s\"\n\n", + vui->sock_filename, + vui->sock_is_server ? "server" : "client", + strerror (vui->sock_errno)); + + vlib_cli_output (vm, " Memory regions (total %d)\n", vui->nregions); + + if (vui->nregions) + { + vlib_cli_output (vm, + " region fd guest_phys_addr memory_size userspace_addr mmap_offset mmap_addr\n"); + vlib_cli_output (vm, + " ====== ===== ================== ================== ================== ================== ==================\n"); + } + for (j = 0; j < vui->nregions; j++) + { + vlib_cli_output (vm, + " %d %-5d 0x%016lx 0x%016lx 0x%016lx 0x%016lx 0x%016lx\n", + j, vui->region_mmap_fd[j], + vui->regions[j].guest_phys_addr, + vui->regions[j].memory_size, + vui->regions[j].userspace_addr, + vui->regions[j].mmap_offset, + pointer_to_uword (vui->region_mmap_addr[j])); + } + for (q = 0; q < vui->num_vrings; q++) + { + vlib_cli_output (vm, "\n Virtqueue %d\n", q); + + vlib_cli_output (vm, + " qsz %d last_avail_idx %d last_used_idx %d\n", + vui->vrings[q].qsz, vui->vrings[q].last_avail_idx, + vui->vrings[q].last_used_idx); + + if (vui->vrings[q].avail && vui->vrings[q].used) + vlib_cli_output (vm, + " avail.flags %x avail.idx %d used.flags %x used.idx %d\n", + vui->vrings[q].avail->flags, + vui->vrings[q].avail->idx, + vui->vrings[q].used->flags, + vui->vrings[q].used->idx); + + vlib_cli_output (vm, " kickfd %d callfd %d errfd %d\n", + vui->vrings[q].kickfd, + vui->vrings[q].callfd, vui->vrings[q].errfd); + + if (show_descr) + { + vlib_cli_output (vm, "\n descriptor table:\n"); + vlib_cli_output (vm, + " id addr len flags next user_addr\n"); + vlib_cli_output (vm, + " ===== ================== ===== ====== ===== ==================\n"); + for (j = 0; j < vui->vrings[q].qsz; j++) + { + vlib_cli_output (vm, + " %-5d 0x%016lx %-5d 0x%04x %-5d 0x%016lx\n", + j, vui->vrings[q].desc[j].addr, + vui->vrings[q].desc[j].len, + vui->vrings[q].desc[j].flags, + vui->vrings[q].desc[j].next, + pointer_to_uword (map_guest_mem + (vui, + vui->vrings[q]. + desc[j].addr))); + } + } + } + vlib_cli_output (vm, "\n"); } - vlib_cli_output (vm, "\n"); - } done: vec_free (hw_if_indices); return error; @@ -1997,19 +2224,19 @@ done: static clib_error_t * vhost_user_config (vlib_main_t * vm, unformat_input_t * input) { - vhost_user_main_t * vum = &vhost_user_main; + vhost_user_main_t *vum = &vhost_user_main; while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT) { if (unformat (input, "coalesce-frames %d", &vum->coalesce_frames)) - ; + ; else if (unformat (input, "coalesce-time %f", &vum->coalesce_time)) - ; + ; else if (unformat (input, "dont-dump-memory")) - vum->dont_dump_vhost_user_memory = 1; + vum->dont_dump_vhost_user_memory = 1; else - return clib_error_return (0, "unknown input `%U'", - format_unformat_error, input); + return clib_error_return (0, "unknown input `%U'", + format_unformat_error, input); } return 0; @@ -2021,14 +2248,22 @@ VLIB_CONFIG_FUNCTION (vhost_user_config, "vhost-user"); void vhost_user_unmap_all (void) { - vhost_user_main_t * vum = &vhost_user_main; - vhost_user_intf_t * vui; + vhost_user_main_t *vum = &vhost_user_main; + vhost_user_intf_t *vui; if (vum->dont_dump_vhost_user_memory) { vec_foreach (vui, vum->vhost_user_interfaces) - { - unmap_all_mem_regions(vui); - } + { + unmap_all_mem_regions (vui); + } } } + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ |