On Fri, Feb 02, 2024 at 03:11:46PM +0100, Laurent Vivier wrote: > Add vhost_user.c and vhost_user.h that define the functions needed > to implement vhost-user backend. > > Signed-off-by: Laurent Vivier > --- > Makefile | 4 +- > passt.c | 2 + > passt.h | 8 + > tap.c | 2 +- > tap.h | 3 + > vhost_user.c | 1050 ++++++++++++++++++++++++++++++++++++++++++++++++++ > vhost_user.h | 139 +++++++ > 7 files changed, 1205 insertions(+), 3 deletions(-) > create mode 100644 vhost_user.c > create mode 100644 vhost_user.h > > diff --git a/Makefile b/Makefile > index ae1daa6b2b50..2016b071ddf2 100644 > --- a/Makefile > +++ b/Makefile > @@ -47,7 +47,7 @@ FLAGS += -DDUAL_STACK_SOCKETS=$(DUAL_STACK_SOCKETS) > PASST_SRCS = arch.c arp.c checksum.c conf.c dhcp.c dhcpv6.c flow.c icmp.c \ > igmp.c isolation.c lineread.c log.c mld.c ndp.c netlink.c packet.c \ > passt.c pasta.c pcap.c pif.c port_fwd.c tap.c tcp.c tcp_splice.c \ > - tcp_buf.c udp.c util.c iov.c ip.c virtio.c > + tcp_buf.c udp.c util.c iov.c ip.c virtio.c vhost_user.c > QRAP_SRCS = qrap.c > SRCS = $(PASST_SRCS) $(QRAP_SRCS) > > @@ -57,7 +57,7 @@ PASST_HEADERS = arch.h arp.h checksum.h conf.h dhcp.h dhcpv6.h flow.h \ > flow_table.h icmp.h inany.h isolation.h lineread.h log.h ndp.h \ > netlink.h packet.h passt.h pasta.h pcap.h pif.h port_fwd.h siphash.h \ > tap.h tcp.h tcp_conn.h tcp_splice.h tcp_buf.h tcp_internal.h udp.h \ > - util.h iov.h ip.h virtio.h > + util.h iov.h ip.h virtio.h vhost_user.h > HEADERS = $(PASST_HEADERS) seccomp.h > > C := \#include \nstruct tcp_info x = { .tcpi_snd_wnd = 0 }; > diff --git a/passt.c b/passt.c > index 5056a49dec95..95034d73381f 100644 > --- a/passt.c > +++ b/passt.c > @@ -72,6 +72,8 @@ char *epoll_type_str[] = { > [EPOLL_TYPE_TAP_PASTA] = "/dev/net/tun device", > [EPOLL_TYPE_TAP_PASST] = "connected qemu socket", > [EPOLL_TYPE_TAP_LISTEN] = "listening qemu socket", > + [EPOLL_TYPE_VHOST_CMD] = "vhost-user command socket", > + [EPOLL_TYPE_VHOST_KICK] = "vhost-user kick socket", > }; > static_assert(ARRAY_SIZE(epoll_type_str) == EPOLL_NUM_TYPES, > "epoll_type_str[] doesn't match enum epoll_type"); > diff --git a/passt.h b/passt.h > index a9e8f15af0e1..6ed1d0b19e82 100644 > --- a/passt.h > +++ b/passt.h > @@ -42,6 +42,7 @@ union epoll_ref; > #include "port_fwd.h" > #include "tcp.h" > #include "udp.h" > +#include "vhost_user.h" > > /** > * enum epoll_type - Different types of fds we poll over > @@ -71,6 +72,10 @@ enum epoll_type { > EPOLL_TYPE_TAP_PASST, > /* socket listening for qemu socket connections */ > EPOLL_TYPE_TAP_LISTEN, > + /* vhost-user command socket */ > + EPOLL_TYPE_VHOST_CMD, > + /* vhost-user kick event socket */ > + EPOLL_TYPE_VHOST_KICK, > > EPOLL_NUM_TYPES, > }; > @@ -303,6 +308,9 @@ struct ctx { > > int low_wmem; > int low_rmem; > + > + /* vhost-user */ > + struct VuDev vdev; > }; > > void proto_update_l2_buf(const unsigned char *eth_d, > diff --git a/tap.c b/tap.c > index ebe52247ad87..936206e53637 100644 > --- a/tap.c > +++ b/tap.c > @@ -954,7 +954,7 @@ void packet_add_all_do(struct ctx *c, ssize_t len, char *p, > * tap_sock_reset() - Handle closing or failure of connect AF_UNIX socket > * @c: Execution context > */ > -static void tap_sock_reset(struct ctx *c) > +void tap_sock_reset(struct ctx *c) > { > if (c->one_off) { > info("Client closed connection, exiting"); > diff --git a/tap.h b/tap.h > index 7157ef37ee6e..ee839d4f09dc 100644 > --- a/tap.h > +++ b/tap.h > @@ -81,12 +81,15 @@ void tap_handler_pasta(struct ctx *c, uint32_t events, > const struct timespec *now); > void tap_handler_passt(struct ctx *c, uint32_t events, > const struct timespec *now); > +void tap_sock_reset(struct ctx *c); > void tap_sock_init(struct ctx *c); > void pool_flush_all(void); > void tap_handler_all(struct ctx *c, const struct timespec *now); > > void packet_add_do(struct pool *p, size_t len, const char *start, > const char *func, int line); > +void packet_add_all_do(struct ctx *c, ssize_t len, char *p, > + const char *func, int line); This looks like it should be a fixup for the earlier patch which introduced the macro below. > #define packet_add_all(p, len, start) \ > packet_add_all_do(p, len, start, __func__, __LINE__) > > diff --git a/vhost_user.c b/vhost_user.c > new file mode 100644 > index 000000000000..2acd72398e3a > --- /dev/null > +++ b/vhost_user.c > @@ -0,0 +1,1050 @@ > +// SPDX-License-Identifier: GPL-2.0-or-later > + > +/* some parts from QEMU subprojects/libvhost-user/libvhost-user.c */ > + > +#include > +#include > +#include > +#include > +#include > +#include > +#include > +#include > +#include > +#include > +#include > +#include > +#include > +#include > +#include > +#include > +#include > +#include > + > +#include "util.h" > +#include "passt.h" > +#include "tap.h" > +#include "vhost_user.h" > + > +#define VHOST_USER_VERSION 1 > + > +static unsigned char buffer[65536][VHOST_USER_MAX_QUEUES]; > + > +void vu_print_capabilities(void) > +{ > + printf("{\n"); > + printf(" \"type\": \"net\"\n"); > + printf("}\n"); > + exit(EXIT_SUCCESS); It's not clear to me what we need this for. > +} > + > +static const char * > +vu_request_to_string(unsigned int req) > +{ > +#define REQ(req) [req] = #req > + static const char *vu_request_str[] = { > + REQ(VHOST_USER_NONE), > + REQ(VHOST_USER_GET_FEATURES), > + REQ(VHOST_USER_SET_FEATURES), > + REQ(VHOST_USER_SET_OWNER), > + REQ(VHOST_USER_RESET_OWNER), > + REQ(VHOST_USER_SET_MEM_TABLE), > + REQ(VHOST_USER_SET_LOG_BASE), > + REQ(VHOST_USER_SET_LOG_FD), > + REQ(VHOST_USER_SET_VRING_NUM), > + REQ(VHOST_USER_SET_VRING_ADDR), > + REQ(VHOST_USER_SET_VRING_BASE), > + REQ(VHOST_USER_GET_VRING_BASE), > + REQ(VHOST_USER_SET_VRING_KICK), > + REQ(VHOST_USER_SET_VRING_CALL), > + REQ(VHOST_USER_SET_VRING_ERR), > + REQ(VHOST_USER_GET_PROTOCOL_FEATURES), > + REQ(VHOST_USER_SET_PROTOCOL_FEATURES), > + REQ(VHOST_USER_GET_QUEUE_NUM), > + REQ(VHOST_USER_SET_VRING_ENABLE), > + REQ(VHOST_USER_SEND_RARP), > + REQ(VHOST_USER_NET_SET_MTU), > + REQ(VHOST_USER_SET_BACKEND_REQ_FD), > + REQ(VHOST_USER_IOTLB_MSG), > + REQ(VHOST_USER_SET_VRING_ENDIAN), > + REQ(VHOST_USER_GET_CONFIG), > + REQ(VHOST_USER_SET_CONFIG), > + REQ(VHOST_USER_POSTCOPY_ADVISE), > + REQ(VHOST_USER_POSTCOPY_LISTEN), > + REQ(VHOST_USER_POSTCOPY_END), > + REQ(VHOST_USER_GET_INFLIGHT_FD), > + REQ(VHOST_USER_SET_INFLIGHT_FD), > + REQ(VHOST_USER_GPU_SET_SOCKET), > + REQ(VHOST_USER_VRING_KICK), > + REQ(VHOST_USER_GET_MAX_MEM_SLOTS), > + REQ(VHOST_USER_ADD_MEM_REG), > + REQ(VHOST_USER_REM_MEM_REG), > + REQ(VHOST_USER_MAX), > + }; > +#undef REQ > + > + if (req < VHOST_USER_MAX) { > + return vu_request_str[req]; > + } else { > + return "unknown"; > + } > +} > + > +/* Translate qemu virtual address to our virtual address. */ The meaning of this comment in its new context is rather unclear. > +static void *qva_to_va(VuDev *dev, uint64_t qemu_addr) > +{ > + unsigned int i; > + > + /* Find matching memory region. */ > + for (i = 0; i < dev->nregions; i++) { > + VuDevRegion *r = &dev->regions[i]; > + > + if ((qemu_addr >= r->qva) && (qemu_addr < (r->qva + r->size))) { > + return (void *)(uintptr_t) > + (qemu_addr - r->qva + r->mmap_addr + r->mmap_offset); > + } > + } > + > + return NULL; > +} > + > +static void > +vmsg_close_fds(VhostUserMsg *vmsg) > +{ > + int i; > + > + for (i = 0; i < vmsg->fd_num; i++) > + close(vmsg->fds[i]); > +} > + > +static void vu_remove_watch(VuDev *vdev, int fd) > +{ > + struct ctx *c = (struct ctx *) ((char *)vdev - offsetof(struct ctx, vdev)); > + > + epoll_ctl(c->epollfd, EPOLL_CTL_DEL, fd, NULL); > +} > + > +/* Set reply payload.u64 and clear request flags and fd_num */ > +static void vmsg_set_reply_u64(struct VhostUserMsg *vmsg, uint64_t val) > +{ > + vmsg->hdr.flags = 0; /* defaults will be set by vu_send_reply() */ > + vmsg->hdr.size = sizeof(vmsg->payload.u64); > + vmsg->payload.u64 = val; > + vmsg->fd_num = 0; > +} > + > +static ssize_t vu_message_read_default(VuDev *dev, int conn_fd, struct VhostUserMsg *vmsg) This only appears to return 0, 1 or a negative errno, which makes ssize_t an odd choice. > +{ > + char control[CMSG_SPACE(VHOST_MEMORY_BASELINE_NREGIONS * > + sizeof(int))] = { 0 }; > + struct iovec iov = { > + .iov_base = (char *)vmsg, > + .iov_len = VHOST_USER_HDR_SIZE, > + }; > + struct msghdr msg = { > + .msg_iov = &iov, > + .msg_iovlen = 1, > + .msg_control = control, > + .msg_controllen = sizeof(control), > + }; > + size_t fd_size; > + struct cmsghdr *cmsg; > + ssize_t ret, sz_payload; > + > + ret = recvmsg(conn_fd, &msg, MSG_DONTWAIT); > + if (ret < 0) { > + if (errno == EINTR || errno == EAGAIN || errno == EWOULDBLOCK) > + return 0; > + vu_panic(dev, "Error while recvmsg: %s", strerror(errno)); > + goto out; > + } > + > + vmsg->fd_num = 0; > + for (cmsg = CMSG_FIRSTHDR(&msg); cmsg != NULL; > + cmsg = CMSG_NXTHDR(&msg, cmsg)) { > + if (cmsg->cmsg_level == SOL_SOCKET && > + cmsg->cmsg_type == SCM_RIGHTS) { > + fd_size = cmsg->cmsg_len - CMSG_LEN(0); > + vmsg->fd_num = fd_size / sizeof(int); > + memcpy(vmsg->fds, CMSG_DATA(cmsg), fd_size); > + break; > + } > + } > + > + sz_payload = vmsg->hdr.size; > + if ((size_t)sz_payload > sizeof(vmsg->payload)) { > + vu_panic(dev, > + "Error: too big message request: %d, size: vmsg->size: %zd, " > + "while sizeof(vmsg->payload) = %zu", > + vmsg->hdr.request, sz_payload, sizeof(vmsg->payload)); > + goto out; > + } > + > + if (sz_payload) { > + do { > + ret = recv(conn_fd, &vmsg->payload, sz_payload, 0); > + } while (ret < 0 && (errno == EINTR || errno == EAGAIN)); > + > + if (ret < sz_payload) { > + vu_panic(dev, "Error while reading: %s", strerror(errno)); > + goto out; > + } > + } > + > + return 1; > +out: > + vmsg_close_fds(vmsg); > + > + return -ECONNRESET; > +} > + > +static int vu_message_write(VuDev *dev, int conn_fd, struct VhostUserMsg *vmsg) > +{ > + int rc; > + uint8_t *p = (uint8_t *)vmsg; > + char control[CMSG_SPACE(VHOST_MEMORY_BASELINE_NREGIONS * sizeof(int))] = { 0 }; > + struct iovec iov = { > + .iov_base = (char *)vmsg, > + .iov_len = VHOST_USER_HDR_SIZE, > + }; > + struct msghdr msg = { > + .msg_iov = &iov, > + .msg_iovlen = 1, > + .msg_control = control, > + }; > + struct cmsghdr *cmsg; > + > + memset(control, 0, sizeof(control)); Redudant with the initializer AFAICT. > + assert(vmsg->fd_num <= VHOST_MEMORY_BASELINE_NREGIONS); > + if (vmsg->fd_num > 0) { > + size_t fdsize = vmsg->fd_num * sizeof(int); > + msg.msg_controllen = CMSG_SPACE(fdsize); > + cmsg = CMSG_FIRSTHDR(&msg); > + cmsg->cmsg_len = CMSG_LEN(fdsize); > + cmsg->cmsg_level = SOL_SOCKET; > + cmsg->cmsg_type = SCM_RIGHTS; > + memcpy(CMSG_DATA(cmsg), vmsg->fds, fdsize); > + } else { > + msg.msg_controllen = 0; > + } > + > + do { > + rc = sendmsg(conn_fd, &msg, 0); > + } while (rc < 0 && (errno == EINTR || errno == EAGAIN)); Hrm.. if this first sendmsg() returns a (non EINTER, non EGAIN) error, couldn't the sending of the payload clobber rc before we actually handle it and print an error? > + if (vmsg->hdr.size) { > + do { > + if (vmsg->data) { > + rc = write(conn_fd, vmsg->data, vmsg->hdr.size); > + } else { > + rc = write(conn_fd, p + VHOST_USER_HDR_SIZE, vmsg->hdr.size); > + } > + } while (rc < 0 && (errno == EINTR || errno == EAGAIN)); > + } > + > + if (rc <= 0) { > + vu_panic(dev, "Error while writing: %s", strerror(errno)); > + return false; > + } > + > + return true; > +} > + > +static int vu_send_reply(VuDev *dev, int conn_fd, struct VhostUserMsg *msg) > +{ > + msg->hdr.flags &= ~VHOST_USER_VERSION_MASK; > + msg->hdr.flags |= VHOST_USER_VERSION; > + msg->hdr.flags |= VHOST_USER_REPLY_MASK; > + > + return vu_message_write(dev, conn_fd, msg); > +} > + > +static bool vu_get_features_exec(struct VhostUserMsg *msg) > +{ > + uint64_t features = > + 1ULL << VIRTIO_F_VERSION_1 | > + 1ULL << VIRTIO_NET_F_MRG_RXBUF | > + 1ULL << VHOST_USER_F_PROTOCOL_FEATURES; > + > + vmsg_set_reply_u64(msg, features); > + > + debug("Sending back to guest u64: 0x%016"PRIx64, msg->payload.u64); > + > + return true; > +} > + > +static void > +vu_set_enable_all_rings(VuDev *vdev, bool enabled) > +{ > + uint16_t i; > + > + for (i = 0; i < VHOST_USER_MAX_QUEUES; i++) { > + vdev->vq[i].enable = enabled; > + } > +} > + > +static bool > +vu_set_features_exec(VuDev *vdev, struct VhostUserMsg *msg) > +{ > + debug("u64: 0x%016"PRIx64, msg->payload.u64); I think a lot of these debug messages need some looking over to make sure they're actually clear if you don't already know they're related to vhost-user. > + > + vdev->features = msg->payload.u64; > + if (!vu_has_feature(vdev, VIRTIO_F_VERSION_1)) { > + /* > + * We only support devices conforming to VIRTIO 1.0 or > + * later > + */ > + vu_panic(vdev, "virtio legacy devices aren't supported by passt"); > + return false; > + } > + > + if (!vu_has_feature(vdev, VHOST_USER_F_PROTOCOL_FEATURES)) { > + vu_set_enable_all_rings(vdev, true); > + } > + > + /* virtio-net features */ > + > + if (vu_has_feature(vdev, VIRTIO_F_VERSION_1) || Won't this always be true given the test above? > + vu_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF)) { > + vdev->hdrlen = sizeof(struct virtio_net_hdr_mrg_rxbuf); > + } else { > + vdev->hdrlen = sizeof(struct virtio_net_hdr); > + } > + > + return false; > +} > + > +static bool > +vu_set_owner_exec(void) > +{ > + return false; > +} > + > +static bool map_ring(VuDev *vdev, VuVirtq *vq) > +{ > + vq->vring.desc = qva_to_va(vdev, vq->vra.desc_user_addr); > + vq->vring.used = qva_to_va(vdev, vq->vra.used_user_addr); > + vq->vring.avail = qva_to_va(vdev, vq->vra.avail_user_addr); > + > + debug("Setting virtq addresses:"); > + debug(" vring_desc at %p", (void *)vq->vring.desc); > + debug(" vring_used at %p", (void *)vq->vring.used); > + debug(" vring_avail at %p", (void *)vq->vring.avail); > + > + return !(vq->vring.desc && vq->vring.used && vq->vring.avail); > +} > + > +/* > + * #syscalls:passt mmap munmap > + */ > + > +static bool vu_set_mem_table_exec(VuDev *vdev, > + struct VhostUserMsg *msg) > +{ > + unsigned int i; > + struct VhostUserMemory m = msg->payload.memory, *memory = &m; > + > + for (i = 0; i < vdev->nregions; i++) { > + VuDevRegion *r = &vdev->regions[i]; > + void *m = (void *) (uintptr_t) r->mmap_addr; > + > + if (m) > + munmap(m, r->size + r->mmap_offset); > + } > + vdev->nregions = memory->nregions; > + > + debug("Nregions: %u", memory->nregions); > + for (i = 0; i < vdev->nregions; i++) { > + void *mmap_addr; > + VhostUserMemory_region *msg_region = &memory->regions[i]; > + VuDevRegion *dev_region = &vdev->regions[i]; > + > + debug("Region %d", i); > + debug(" guest_phys_addr: 0x%016"PRIx64, > + msg_region->guest_phys_addr); > + debug(" memory_size: 0x%016"PRIx64, > + msg_region->memory_size); > + debug(" userspace_addr 0x%016"PRIx64, > + msg_region->userspace_addr); > + debug(" mmap_offset 0x%016"PRIx64, > + msg_region->mmap_offset); > + > + dev_region->gpa = msg_region->guest_phys_addr; > + dev_region->size = msg_region->memory_size; > + dev_region->qva = msg_region->userspace_addr; > + dev_region->mmap_offset = msg_region->mmap_offset; > + > + /* We don't use offset argument of mmap() since the > + * mapped address has to be page aligned, and we use huge > + * pages. */ Is that accurate in the new context? > + mmap_addr = mmap(0, dev_region->size + dev_region->mmap_offset, > + PROT_READ | PROT_WRITE, MAP_SHARED | MAP_NORESERVE, > + msg->fds[i], 0); > + > + if (mmap_addr == MAP_FAILED) { > + vu_panic(vdev, "region mmap error: %s", strerror(errno)); > + } else { > + dev_region->mmap_addr = (uint64_t)(uintptr_t)mmap_addr; > + debug(" mmap_addr: 0x%016"PRIx64, > + dev_region->mmap_addr); > + } > + > + close(msg->fds[i]); > + } > + > + for (i = 0; i < VHOST_USER_MAX_QUEUES; i++) { > + if (vdev->vq[i].vring.desc) { > + if (map_ring(vdev, &vdev->vq[i])) { > + vu_panic(vdev, "remapping queue %d during setmemtable", i); > + } > + } > + } > + > + return false; > +} > + > +static bool vu_set_vring_num_exec(VuDev *vdev, > + struct VhostUserMsg *msg) > +{ > + unsigned int index = msg->payload.state.index; > + unsigned int num = msg->payload.state.num; > + > + debug("State.index: %u", index); > + debug("State.num: %u", num); > + vdev->vq[index].vring.num = num; > + > + return false; > +} > + > +static bool vu_set_vring_addr_exec(VuDev *vdev, > + struct VhostUserMsg *msg) > +{ > + struct vhost_vring_addr addr = msg->payload.addr, *vra = &addr; > + unsigned int index = vra->index; > + VuVirtq *vq = &vdev->vq[index]; > + > + debug("vhost_vring_addr:"); > + debug(" index: %d", vra->index); > + debug(" flags: %d", vra->flags); > + debug(" desc_user_addr: 0x%016" PRIx64, (uint64_t)vra->desc_user_addr); > + debug(" used_user_addr: 0x%016" PRIx64, (uint64_t)vra->used_user_addr); > + debug(" avail_user_addr: 0x%016" PRIx64, (uint64_t)vra->avail_user_addr); > + debug(" log_guest_addr: 0x%016" PRIx64, (uint64_t)vra->log_guest_addr); > + > + vq->vra = *vra; > + vq->vring.flags = vra->flags; > + vq->vring.log_guest_addr = vra->log_guest_addr; > + > + if (map_ring(vdev, vq)) { > + vu_panic(vdev, "Invalid vring_addr message"); > + return false; > + } > + > + vq->used_idx = le16toh(vq->vring.used->idx); > + > + if (vq->last_avail_idx != vq->used_idx) { > + debug("Last avail index != used index: %u != %u", > + vq->last_avail_idx, vq->used_idx); > + } > + > + return false; > +} > + > +static bool vu_set_vring_base_exec(VuDev *vdev, > + struct VhostUserMsg *msg) > +{ > + unsigned int index = msg->payload.state.index; > + unsigned int num = msg->payload.state.num; > + > + debug("State.index: %u", index); > + debug("State.num: %u", num); > + vdev->vq[index].shadow_avail_idx = vdev->vq[index].last_avail_idx = num; > + > + return false; > +} > + > +static bool vu_get_vring_base_exec(VuDev *vdev, > + struct VhostUserMsg *msg) > +{ > + unsigned int index = msg->payload.state.index; > + > + debug("State.index: %u", index); > + msg->payload.state.num = vdev->vq[index].last_avail_idx; > + msg->hdr.size = sizeof(msg->payload.state); > + > + vdev->vq[index].started = false; > + > + if (vdev->vq[index].call_fd != -1) { > + close(vdev->vq[index].call_fd); > + vdev->vq[index].call_fd = -1; > + } > + if (vdev->vq[index].kick_fd != -1) { > + vu_remove_watch(vdev, vdev->vq[index].kick_fd); > + close(vdev->vq[index].kick_fd); > + vdev->vq[index].kick_fd = -1; > + } > + > + return true; > +} > + > +static void vu_set_watch(VuDev *vdev, int fd) > +{ > + struct ctx *c = (struct ctx *) ((char *)vdev - offsetof(struct ctx, vdev)); > + union epoll_ref ref = { .type = EPOLL_TYPE_VHOST_KICK, .fd = fd }; > + struct epoll_event ev = { 0 }; > + > + ev.data.u64 = ref.u64; > + ev.events = EPOLLIN; > + epoll_ctl(c->epollfd, EPOLL_CTL_ADD, fd, &ev); > +} > + > +int vu_send(const struct ctx *c, const void *buf, size_t size) > +{ > + VuDev *vdev = (VuDev *)&c->vdev; > + size_t hdrlen = vdev->hdrlen; > + VuVirtq *vq = &vdev->vq[VHOST_USER_RX_QUEUE]; > + unsigned int indexes[VIRTQUEUE_MAX_SIZE]; > + size_t lens[VIRTQUEUE_MAX_SIZE]; > + size_t offset; > + int i, j; > + __virtio16 *num_buffers_ptr; > + > + debug("vu_send size %zu hdrlen %zu", size, hdrlen); > + > + if (!vu_queue_enabled(vq) || !vu_queue_started(vq)) { > + err("Got packet, but no available descriptors on RX virtq."); The message doesn't seem to quite match the condition. > + return 0; > + } > + > + offset = 0; > + i = 0; > + num_buffers_ptr = NULL; > + while (offset < size) { > + VuVirtqElement *elem; > + size_t len; > + int total; > + > + total = 0; > + > + if (i == VIRTQUEUE_MAX_SIZE) { > + err("virtio-net unexpected long buffer chain"); > + goto err; > + } > + > + elem = vu_queue_pop(vdev, vq, sizeof(VuVirtqElement), > + buffer[VHOST_USER_RX_QUEUE]); > + if (!elem) { > + if (!vdev->broken) { > + eventfd_t kick_data; > + ssize_t rc; > + int status; > + > + /* wait the kernel to put new entries in the queue */ > + > + status = fcntl(vq->kick_fd, F_GETFL); > + if (status != -1) { > + fcntl(vq->kick_fd, F_SETFL, status & ~O_NONBLOCK); > + rc = eventfd_read(vq->kick_fd, &kick_data); Is it safe for us to block here? > + fcntl(vq->kick_fd, F_SETFL, status); > + if (rc != -1) > + continue; > + } > + } > + if (i) { > + err("virtio-net unexpected empty queue: " > + "i %d mergeable %d offset %zd, size %zd, " > + "features 0x%" PRIx64, > + i, vu_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF), > + offset, size, vdev->features); > + } > + offset = -1; > + goto err; > + } > + > + if (elem->in_num < 1) { > + err("virtio-net receive queue contains no in buffers"); > + vu_queue_detach_element(vdev, vq, elem->index, 0); > + offset = -1; > + goto err; > + } > + > + if (i == 0) { > + struct virtio_net_hdr hdr = { > + .flags = VIRTIO_NET_HDR_F_DATA_VALID, > + .gso_type = VIRTIO_NET_HDR_GSO_NONE, > + }; > + > + ASSERT(offset == 0); > + ASSERT(elem->in_sg[0].iov_len >= hdrlen); > + > + len = iov_from_buf(elem->in_sg, elem->in_num, 0, &hdr, sizeof hdr); > + > + num_buffers_ptr = (__virtio16 *)((char *)elem->in_sg[0].iov_base + > + len); > + > + total += hdrlen; > + } > + > + len = iov_from_buf(elem->in_sg, elem->in_num, total, (char *)buf + offset, > + size - offset); > + > + total += len; > + offset += len; > + > + /* If buffers can't be merged, at this point we > + * must have consumed the complete packet. > + * Otherwise, drop it. > + */ > + if (!vu_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) && offset < size) { > + vu_queue_unpop(vdev, vq, elem->index, total); > + goto err; > + } > + > + indexes[i] = elem->index; > + lens[i] = total; > + i++; > + } > + > + if (num_buffers_ptr && vu_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF)) { > + *num_buffers_ptr = htole16(i); > + } > + > + for (j = 0; j < i; j++) { > + debug("filling total %zd idx %d", lens[j], j); > + vu_queue_fill_by_index(vdev, vq, indexes[j], lens[j], j); > + } > + > + vu_queue_flush(vdev, vq, i); > + vu_queue_notify(vdev, vq); > + > + debug("sent %zu", offset); > + > + return offset; > +err: > + for (j = 0; j < i; j++) { > + vu_queue_detach_element(vdev, vq, indexes[j], lens[j]); > + } > + > + return offset; > +} > + > +size_t tap_send_frames_vu(const struct ctx *c, const struct iovec *iov, size_t n) > +{ > + size_t i; > + int ret; > + > + debug("tap_send_frames_vu n %zd", n); > + > + for (i = 0; i < n; i++) { > + ret = vu_send(c, iov[i].iov_base, iov[i].iov_len); > + if (ret < 0) > + break; > + } > + debug("count %zd", i); > + return i; > +} > + > +static void vu_handle_tx(VuDev *vdev, int index) > +{ > + struct ctx *c = (struct ctx *) ((char *)vdev - offsetof(struct ctx, vdev)); > + VuVirtq *vq = &vdev->vq[index]; > + int hdrlen = vdev->hdrlen; > + struct timespec now; > + char *p; > + size_t n; > + > + if (index % 2 != VHOST_USER_TX_QUEUE) { > + debug("index %d is not an TX queue", index); When would this situation arise? It seems like it ought to be something stronger than a debug() - either a fatal error in the setup or an ASSERT(). > + return; > + } > + > + clock_gettime(CLOCK_MONOTONIC, &now); If I'm looking ahead at the next patches to see how this is called correctly, I think you can avoid this gettime() call by passing a 'now' value down from the mail epoll loop (the other tap handlers already take one). > + p = pkt_buf; > + > + pool_flush_all(); > + > + while (1) { > + VuVirtqElement *elem; > + unsigned int out_num; > + struct iovec sg[VIRTQUEUE_MAX_SIZE], *out_sg; > + > + ASSERT(index == VHOST_USER_TX_QUEUE); Hrm.. and isn't this redundant with the check at the start of the function? > + elem = vu_queue_pop(vdev, vq, sizeof(VuVirtqElement), buffer[index]); > + if (!elem) { > + break; > + } > + > + out_num = elem->out_num; > + out_sg = elem->out_sg; > + if (out_num < 1) { > + debug("virtio-net header not in first element"); > + break; > + } > + > + if (hdrlen) { > + unsigned sg_num; > + > + sg_num = iov_copy(sg, ARRAY_SIZE(sg), out_sg, out_num, > + hdrlen, -1); > + out_num = sg_num; > + out_sg = sg; > + } > + > + n = iov_to_buf(out_sg, out_num, 0, p, TAP_BUF_FILL); > + > + packet_add_all(c, n, p); > + > + p += n; > + > + vu_queue_push(vdev, vq, elem, 0); > + vu_queue_notify(vdev, vq); > + } > + tap_handler_all(c, &now); > +} > + > +void vu_kick_cb(struct ctx *c, union epoll_ref ref) > +{ > + VuDev *vdev = &c->vdev; > + eventfd_t kick_data; > + ssize_t rc; > + int index; > + > + for (index = 0; index < VHOST_USER_MAX_QUEUES; index++) > + if (c->vdev.vq[index].kick_fd == ref.fd) > + break; > + > + if (index == VHOST_USER_MAX_QUEUES) > + return; > + > + rc = eventfd_read(ref.fd, &kick_data); > + if (rc == -1) { > + vu_panic(vdev, "kick eventfd_read(): %s", strerror(errno)); > + vu_remove_watch(vdev, ref.fd); > + } else { > + debug("Got kick_data: %016"PRIx64" idx:%d", > + kick_data, index); > + if (index % 2 == VHOST_USER_TX_QUEUE) .. and here we seem to have a third check for the same thing. > + vu_handle_tx(vdev, index); > + } > +} > + > +static bool vu_check_queue_msg_file(VuDev *vdev, struct VhostUserMsg *msg) > +{ > + int index = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK; > + bool nofd = msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK; > + > + if (index >= VHOST_USER_MAX_QUEUES) { > + vmsg_close_fds(msg); > + vu_panic(vdev, "Invalid queue index: %u", index); > + return false; > + } > + > + if (nofd) { > + vmsg_close_fds(msg); > + return true; > + } > + > + if (msg->fd_num != 1) { > + vmsg_close_fds(msg); > + vu_panic(vdev, "Invalid fds in request: %d", msg->hdr.request); > + return false; > + } > + > + return true; > +} > + > +static bool vu_set_vring_kick_exec(VuDev *vdev, > + struct VhostUserMsg *msg) > +{ > + int index = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK; > + bool nofd = msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK; > + > + debug("u64: 0x%016"PRIx64, msg->payload.u64); > + > + if (!vu_check_queue_msg_file(vdev, msg)) > + return false; > + > + if (vdev->vq[index].kick_fd != -1) { > + vu_remove_watch(vdev, vdev->vq[index].kick_fd); > + close(vdev->vq[index].kick_fd); > + vdev->vq[index].kick_fd = -1; > + } > + > + vdev->vq[index].kick_fd = nofd ? -1 : msg->fds[0]; > + debug("Got kick_fd: %d for vq: %d", vdev->vq[index].kick_fd, index); > + > + vdev->vq[index].started = true; > + > + if (vdev->vq[index].kick_fd != -1 && index % 2 == VHOST_USER_TX_QUEUE) { > + vu_set_watch(vdev, vdev->vq[index].kick_fd); > + debug("Waiting for kicks on fd: %d for vq: %d", > + vdev->vq[index].kick_fd, index); > + } > + > + return false; > +} > + > +static bool vu_set_vring_call_exec(VuDev *vdev, > + struct VhostUserMsg *msg) > +{ > + int index = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK; > + bool nofd = msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK; > + > + debug("u64: 0x%016"PRIx64, msg->payload.u64); > + > + if (!vu_check_queue_msg_file(vdev, msg)) > + return false; > + > + if (vdev->vq[index].call_fd != -1) { > + close(vdev->vq[index].call_fd); > + vdev->vq[index].call_fd = -1; > + } > + > + vdev->vq[index].call_fd = nofd ? -1 : msg->fds[0]; > + > + /* in case of I/O hang after reconnecting */ > + if (vdev->vq[index].call_fd != -1) { > + eventfd_write(msg->fds[0], 1); > + } > + > + debug("Got call_fd: %d for vq: %d", vdev->vq[index].call_fd, index); > + > + return false; > +} > + > +static bool vu_set_vring_err_exec(VuDev *vdev, > + struct VhostUserMsg *msg) > +{ > + int index = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK; > + bool nofd = msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK; > + > + debug("u64: 0x%016"PRIx64, msg->payload.u64); > + > + if (!vu_check_queue_msg_file(vdev, msg)) > + return false; > + > + if (vdev->vq[index].err_fd != -1) { > + close(vdev->vq[index].err_fd); > + vdev->vq[index].err_fd = -1; > + } > + > + vdev->vq[index].err_fd = nofd ? -1 : msg->fds[0]; > + > + return false; > +} > + > +static bool vu_get_protocol_features_exec(struct VhostUserMsg *msg) > +{ > + uint64_t features = 1ULL << VHOST_USER_PROTOCOL_F_REPLY_ACK; > + > + vmsg_set_reply_u64(msg, features); > + > + return true; > +} > + > +static bool vu_set_protocol_features_exec(VuDev *vdev, struct VhostUserMsg *msg) > +{ > + uint64_t features = msg->payload.u64; > + > + debug("u64: 0x%016"PRIx64, features); > + > + vdev->protocol_features = msg->payload.u64; > + > + if (vu_has_protocol_feature(vdev, > + VHOST_USER_PROTOCOL_F_INBAND_NOTIFICATIONS) && > + (!vu_has_protocol_feature(vdev, VHOST_USER_PROTOCOL_F_BACKEND_REQ) || > + !vu_has_protocol_feature(vdev, VHOST_USER_PROTOCOL_F_REPLY_ACK))) { > + /* > + * The use case for using messages for kick/call is simulation, to make > + * the kick and call synchronous. To actually get that behaviour, both > + * of the other features are required. > + * Theoretically, one could use only kick messages, or do them without > + * having F_REPLY_ACK, but too many (possibly pending) messages on the > + * socket will eventually cause the master to hang, to avoid this in > + * scenarios where not desired enforce that the settings are in a way > + * that actually enables the simulation case. > + */ > + vu_panic(vdev, > + "F_IN_BAND_NOTIFICATIONS requires F_BACKEND_REQ && F_REPLY_ACK"); > + return false; > + } > + > + return false; > +} > + > + > +static bool vu_get_queue_num_exec(struct VhostUserMsg *msg) > +{ > + vmsg_set_reply_u64(msg, VHOST_USER_MAX_QUEUES); > + return true; > +} > + > +static bool vu_set_vring_enable_exec(VuDev *vdev, struct VhostUserMsg *msg) > +{ > + unsigned int index = msg->payload.state.index; > + unsigned int enable = msg->payload.state.num; > + > + debug("State.index: %u", index); > + debug("State.enable: %u", enable); > + > + if (index >= VHOST_USER_MAX_QUEUES) { > + vu_panic(vdev, "Invalid vring_enable index: %u", index); > + return false; > + } > + > + vdev->vq[index].enable = enable; > + return false; > +} > + > +void vu_init(struct ctx *c) > +{ > + int i; > + > + c->vdev.hdrlen = 0; > + for (i = 0; i < VHOST_USER_MAX_QUEUES; i++) > + c->vdev.vq[i] = (VuVirtq){ > + .call_fd = -1, > + .kick_fd = -1, > + .err_fd = -1, > + .notification = true, > + }; > +} > + > +static void vu_cleanup(VuDev *vdev) > +{ > + unsigned int i; > + > + for (i = 0; i < VHOST_USER_MAX_QUEUES; i++) { > + VuVirtq *vq = &vdev->vq[i]; > + > + vq->started = false; > + vq->notification = true; > + > + if (vq->call_fd != -1) { > + close(vq->call_fd); > + vq->call_fd = -1; > + } > + if (vq->err_fd != -1) { > + close(vq->err_fd); > + vq->err_fd = -1; > + } > + if (vq->kick_fd != -1) { > + vu_remove_watch(vdev, vq->kick_fd); > + close(vq->kick_fd); > + vq->kick_fd = -1; > + } > + > + vq->vring.desc = 0; > + vq->vring.used = 0; > + vq->vring.avail = 0; > + } > + vdev->hdrlen = 0; > + > + for (i = 0; i < vdev->nregions; i++) { > + VuDevRegion *r = &vdev->regions[i]; > + void *m = (void *) (uintptr_t) r->mmap_addr; > + > + if (m) > + munmap(m, r->size + r->mmap_offset); > + } > + vdev->nregions = 0; > +} > + > +/** > + * tap_handler_vu() - Packet handler for vhost-user > + * @c: Execution context > + * @events: epoll events > + */ > +void tap_handler_vu(struct ctx *c, uint32_t events) > +{ > + VuDev *dev = &c->vdev; > + struct VhostUserMsg msg = { 0 }; > + bool need_reply, reply_requested; > + int ret; > + > + if (events & (EPOLLRDHUP | EPOLLHUP | EPOLLERR)) { > + tap_sock_reset(c); > + return; > + } > + > + > + ret = vu_message_read_default(dev, c->fd_tap, &msg); > + if (ret <= 0) { > + if (errno != EINTR && errno != EAGAIN && errno != EWOULDBLOCK) > + tap_sock_reset(c); > + return; > + } > + debug("================ Vhost user message ================"); > + debug("Request: %s (%d)", vu_request_to_string(msg.hdr.request), > + msg.hdr.request); > + debug("Flags: 0x%x", msg.hdr.flags); > + debug("Size: %u", msg.hdr.size); > + > + need_reply = msg.hdr.flags & VHOST_USER_NEED_REPLY_MASK; > + switch (msg.hdr.request) { > + case VHOST_USER_GET_FEATURES: > + reply_requested = vu_get_features_exec(&msg); > + break; > + case VHOST_USER_SET_FEATURES: > + reply_requested = vu_set_features_exec(dev, &msg); > + break; > + case VHOST_USER_GET_PROTOCOL_FEATURES: > + reply_requested = vu_get_protocol_features_exec(&msg); > + break; > + case VHOST_USER_SET_PROTOCOL_FEATURES: > + reply_requested = vu_set_protocol_features_exec(dev, &msg); > + break; > + case VHOST_USER_GET_QUEUE_NUM: > + reply_requested = vu_get_queue_num_exec(&msg); > + break; > + case VHOST_USER_SET_OWNER: > + reply_requested = vu_set_owner_exec(); > + break; > + case VHOST_USER_SET_MEM_TABLE: > + reply_requested = vu_set_mem_table_exec(dev, &msg); > + break; > + case VHOST_USER_SET_VRING_NUM: > + reply_requested = vu_set_vring_num_exec(dev, &msg); > + break; > + case VHOST_USER_SET_VRING_ADDR: > + reply_requested = vu_set_vring_addr_exec(dev, &msg); > + break; > + case VHOST_USER_SET_VRING_BASE: > + reply_requested = vu_set_vring_base_exec(dev, &msg); > + break; > + case VHOST_USER_GET_VRING_BASE: > + reply_requested = vu_get_vring_base_exec(dev, &msg); > + break; > + case VHOST_USER_SET_VRING_KICK: > + reply_requested = vu_set_vring_kick_exec(dev, &msg); > + break; > + case VHOST_USER_SET_VRING_CALL: > + reply_requested = vu_set_vring_call_exec(dev, &msg); > + break; > + case VHOST_USER_SET_VRING_ERR: > + reply_requested = vu_set_vring_err_exec(dev, &msg); > + break; > + case VHOST_USER_SET_VRING_ENABLE: > + reply_requested = vu_set_vring_enable_exec(dev, &msg); > + break; > + case VHOST_USER_NONE: > + vu_cleanup(dev); > + return; > + default: > + vu_panic(dev, "Unhandled request: %d", msg.hdr.request); > + return; > + } > + > + if (!reply_requested && need_reply) { > + msg.payload.u64 = 0; > + msg.hdr.flags = 0; > + msg.hdr.size = sizeof(msg.payload.u64); > + msg.fd_num = 0; > + reply_requested = true; > + } > + > + if (reply_requested) > + ret = vu_send_reply(dev, c->fd_tap, &msg); > + free(msg.data); > +} > diff --git a/vhost_user.h b/vhost_user.h > new file mode 100644 > index 000000000000..25f0b617ab40 > --- /dev/null > +++ b/vhost_user.h > @@ -0,0 +1,139 @@ > +// SPDX-License-Identifier: GPL-2.0-or-later > + > +/* some parts from subprojects/libvhost-user/libvhost-user.h */ > + > +#ifndef VHOST_USER_H > +#define VHOST_USER_H > + > +#include "virtio.h" > +#include "iov.h" > + > +#define VHOST_USER_F_PROTOCOL_FEATURES 30 > + > +#define VHOST_MEMORY_BASELINE_NREGIONS 8 > + > +enum vhost_user_protocol_feature { > + VHOST_USER_PROTOCOL_F_MQ = 0, > + VHOST_USER_PROTOCOL_F_LOG_SHMFD = 1, > + VHOST_USER_PROTOCOL_F_RARP = 2, > + VHOST_USER_PROTOCOL_F_REPLY_ACK = 3, > + VHOST_USER_PROTOCOL_F_NET_MTU = 4, > + VHOST_USER_PROTOCOL_F_BACKEND_REQ = 5, > + VHOST_USER_PROTOCOL_F_CROSS_ENDIAN = 6, > + VHOST_USER_PROTOCOL_F_CRYPTO_SESSION = 7, > + VHOST_USER_PROTOCOL_F_PAGEFAULT = 8, > + VHOST_USER_PROTOCOL_F_CONFIG = 9, > + VHOST_USER_PROTOCOL_F_SLAVE_SEND_FD = 10, > + VHOST_USER_PROTOCOL_F_HOST_NOTIFIER = 11, > + VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD = 12, > + VHOST_USER_PROTOCOL_F_INBAND_NOTIFICATIONS = 14, > + VHOST_USER_PROTOCOL_F_CONFIGURE_MEM_SLOTS = 15, > + > + VHOST_USER_PROTOCOL_F_MAX > +}; > + > +enum vhost_user_request { > + VHOST_USER_NONE = 0, > + VHOST_USER_GET_FEATURES = 1, > + VHOST_USER_SET_FEATURES = 2, > + VHOST_USER_SET_OWNER = 3, > + VHOST_USER_RESET_OWNER = 4, > + VHOST_USER_SET_MEM_TABLE = 5, > + VHOST_USER_SET_LOG_BASE = 6, > + VHOST_USER_SET_LOG_FD = 7, > + VHOST_USER_SET_VRING_NUM = 8, > + VHOST_USER_SET_VRING_ADDR = 9, > + VHOST_USER_SET_VRING_BASE = 10, > + VHOST_USER_GET_VRING_BASE = 11, > + VHOST_USER_SET_VRING_KICK = 12, > + VHOST_USER_SET_VRING_CALL = 13, > + VHOST_USER_SET_VRING_ERR = 14, > + VHOST_USER_GET_PROTOCOL_FEATURES = 15, > + VHOST_USER_SET_PROTOCOL_FEATURES = 16, > + VHOST_USER_GET_QUEUE_NUM = 17, > + VHOST_USER_SET_VRING_ENABLE = 18, > + VHOST_USER_SEND_RARP = 19, > + VHOST_USER_NET_SET_MTU = 20, > + VHOST_USER_SET_BACKEND_REQ_FD = 21, > + VHOST_USER_IOTLB_MSG = 22, > + VHOST_USER_SET_VRING_ENDIAN = 23, > + VHOST_USER_GET_CONFIG = 24, > + VHOST_USER_SET_CONFIG = 25, > + VHOST_USER_CREATE_CRYPTO_SESSION = 26, > + VHOST_USER_CLOSE_CRYPTO_SESSION = 27, > + VHOST_USER_POSTCOPY_ADVISE = 28, > + VHOST_USER_POSTCOPY_LISTEN = 29, > + VHOST_USER_POSTCOPY_END = 30, > + VHOST_USER_GET_INFLIGHT_FD = 31, > + VHOST_USER_SET_INFLIGHT_FD = 32, > + VHOST_USER_GPU_SET_SOCKET = 33, > + VHOST_USER_VRING_KICK = 35, > + VHOST_USER_GET_MAX_MEM_SLOTS = 36, > + VHOST_USER_ADD_MEM_REG = 37, > + VHOST_USER_REM_MEM_REG = 38, > + VHOST_USER_MAX > +}; > + > +typedef struct { > + enum vhost_user_request request; > + > +#define VHOST_USER_VERSION_MASK 0x3 > +#define VHOST_USER_REPLY_MASK (0x1 << 2) > +#define VHOST_USER_NEED_REPLY_MASK (0x1 << 3) > + uint32_t flags; > + uint32_t size; /* the following payload size */ > +} __attribute__ ((__packed__)) vhost_user_header; > + > +typedef struct VhostUserMemory_region { > + uint64_t guest_phys_addr; > + uint64_t memory_size; > + uint64_t userspace_addr; > + uint64_t mmap_offset; > +} VhostUserMemory_region; > + > +struct VhostUserMemory { > + uint32_t nregions; > + uint32_t padding; > + struct VhostUserMemory_region regions[VHOST_MEMORY_BASELINE_NREGIONS]; > +}; > + > +typedef union { > +#define VHOST_USER_VRING_IDX_MASK 0xff > +#define VHOST_USER_VRING_NOFD_MASK (0x1 << 8) > + uint64_t u64; > + struct vhost_vring_state state; > + struct vhost_vring_addr addr; > + struct VhostUserMemory memory; > +} vhost_user_payload; > + > +typedef struct VhostUserMsg { > + vhost_user_header hdr; > + vhost_user_payload payload; > + > + int fds[VHOST_MEMORY_BASELINE_NREGIONS]; > + int fd_num; > + uint8_t *data; > +} __attribute__ ((__packed__)) VhostUserMsg; > +#define VHOST_USER_HDR_SIZE sizeof(vhost_user_header) > + > +#define VHOST_USER_RX_QUEUE 0 > +#define VHOST_USER_TX_QUEUE 1 > + > +static inline bool vu_queue_enabled(VuVirtq *vq) > +{ > + return vq->enable; > +} > + > +static inline bool vu_queue_started(const VuVirtq *vq) > +{ > + return vq->started; > +} > + > +size_t tap_send_frames_vu(const struct ctx *c, const struct iovec *iov, > + size_t n); > +int vu_send(const struct ctx *c, const void *data, size_t len); > +void vu_print_capabilities(void); > +void vu_init(struct ctx *c); > +void vu_kick_cb(struct ctx *c, union epoll_ref ref); > +void tap_handler_vu(struct ctx *c, uint32_t events); > +#endif /* VHOST_USER_H */ -- David Gibson | I'll have my music baroque, and my code david AT gibson.dropbear.id.au | minimalist, thank you. NOT _the_ _other_ | _way_ _around_! http://www.ozlabs.org/~dgibson