From: Peter Dinda Date: Thu, 9 Aug 2012 18:52:52 +0000 (-0500) Subject: Virtio NIC's guest-driven, vmm-driven, or adaptive modes can now X-Git-Url: http://v3vee.org/palacios/gitweb/gitweb.cgi?a=commitdiff_plain;h=357764d1d3bc432b149e8864c183c3a39ee4d474;p=palacios.git Virtio NIC's guest-driven, vmm-driven, or adaptive modes can now be set from a configuration block: ... --- diff --git a/palacios/src/devices/lnx_virtio_nic.c b/palacios/src/devices/lnx_virtio_nic.c index 5e7d4f2..2f5c2b8 100644 --- a/palacios/src/devices/lnx_virtio_nic.c +++ b/palacios/src/devices/lnx_virtio_nic.c @@ -107,14 +107,19 @@ struct virtio_net_config } __attribute__((packed)); struct virtio_dev_state { + struct vm_device * pci_bus; struct list_head dev_list; struct v3_vm_info *vm; + enum {GUEST_DRIVEN=0, VMM_DRIVEN, ADAPTIVE} model; + uint64_t lower_thresh_pps, upper_thresh_pps, period_us; + uint8_t mac[ETH_ALEN]; }; struct virtio_net_state { + struct virtio_net_config net_cfg; struct virtio_config virtio_cfg; @@ -188,6 +193,14 @@ static int virtio_init_state(struct virtio_net_state * virtio) return 0; } +static int virtio_deinit_state(struct guest_info *core, struct virtio_net_state *ns) +{ + if (ns->timer) { + v3_remove_timer(core,ns->timer); + } + return 0; +} + static int tx_one_pkt(struct guest_info * core, struct virtio_net_state * virtio, struct vring_desc * buf_desc) @@ -735,14 +748,13 @@ static int virtio_free(struct virtio_dev_state * virtio) { list_for_each_entry_safe(backend, tmp, &(virtio->dev_list), dev_link) { - - // unregister from PCI - + virtio_deinit_state(&(virtio->vm->cores[0]),backend); list_del(&(backend->dev_link)); V3_Free(backend); } V3_Free(virtio); + return 0; } @@ -838,10 +850,9 @@ static int register_dev(struct virtio_dev_state * virtio, return 0; } -#if 0 -#define RATE_UPPER_THRESHOLD 10 /* 10000 pkts per second, around 100Mbits */ -#define RATE_LOWER_THRESHOLD 1 -#define PROFILE_PERIOD 10000 /*us*/ +#define RATE_UPPER_THRESHOLD_DEFAULT 10000 /* 10000 pkts per second, around 100Mbits */ +#define RATE_LOWER_THRESHOLD_DEFAULT 1000 /* 1000 pkts per second, around 10Mbits */ +#define PROFILE_PERIOD_DEFAULT 10000 /* us */ static void virtio_nic_timer(struct guest_info * core, uint64_t cpu_cycles, uint64_t cpu_freq, @@ -849,6 +860,10 @@ static void virtio_nic_timer(struct guest_info * core, struct virtio_net_state * net_state = (struct virtio_net_state *)priv_data; uint64_t period_us; static int profile_ms = 0; + uint64_t target_period_us = net_state->virtio_dev->period_us; + uint64_t upper_thresh_pps = net_state->virtio_dev->upper_thresh_pps; + uint64_t lower_thresh_pps = net_state->virtio_dev->lower_thresh_pps; + if(!net_state->status){ /* VNIC is not in working status */ return; @@ -857,33 +872,39 @@ static void virtio_nic_timer(struct guest_info * core, period_us = (1000*cpu_cycles)/cpu_freq; net_state->past_us += period_us; - if(net_state->past_us > PROFILE_PERIOD){ - uint32_t tx_rate, rx_rate; - - tx_rate = (net_state->stats.tx_pkts - net_state->tx_pkts)/(net_state->past_us/1000); /* pkts/per ms */ - rx_rate = (net_state->stats.rx_pkts - net_state->rx_pkts)/(net_state->past_us/1000); + if (net_state->past_us > target_period_us) { + + uint64_t tx_count, rx_count; + uint64_t lb_tx_count, lb_rx_count; + uint64_t ub_tx_count, ub_rx_count; + + lb_tx_count = lb_rx_count = (lower_thresh_pps * 1000000) / net_state->past_us; // packets expected in this interval + ub_tx_count = ub_rx_count = (upper_thresh_pps * 1000000) / net_state->past_us; + + tx_count = net_state->stats.tx_pkts - net_state->tx_pkts; + rx_count = net_state->stats.rx_pkts - net_state->rx_pkts; net_state->tx_pkts = net_state->stats.tx_pkts; net_state->rx_pkts = net_state->stats.rx_pkts; - if(tx_rate > RATE_UPPER_THRESHOLD && net_state->tx_notify == 1){ - V3_Print("Virtio NIC: Switch TX to VMM driven mode\n"); + if(tx_count > ub_tx_count && net_state->tx_notify == 1) { + PrintDebug("Virtio NIC: Switch TX to VMM driven mode\n"); disable_cb(&(net_state->tx_vq)); net_state->tx_notify = 0; } - if(tx_rate < RATE_LOWER_THRESHOLD && net_state->tx_notify == 0){ - V3_Print("Virtio NIC: Switch TX to Guest driven mode\n"); + if(tx_count < lb_tx_count && net_state->tx_notify == 0) { + PrintDebug("Virtio NIC: Switch TX to Guest driven mode\n"); enable_cb(&(net_state->tx_vq)); net_state->tx_notify = 1; } - if(rx_rate > RATE_UPPER_THRESHOLD && net_state->rx_notify == 1){ - V3_Print("Virtio NIC: Switch RX to VMM None notify mode\n"); + if(rx_count > ub_rx_count && net_state->rx_notify == 1) { + PrintDebug("Virtio NIC: Switch RX to VMM None notify mode\n"); net_state->rx_notify = 0; } - if(rx_rate < RATE_LOWER_THRESHOLD && net_state->rx_notify == 0){ + if(rx_count < lb_rx_count && net_state->rx_notify == 0) { V3_Print("Virtio NIC: Switch RX to VMM notify mode\n"); net_state->rx_notify = 1; } @@ -893,7 +914,7 @@ static void virtio_nic_timer(struct guest_info * core, profile_ms += period_us/1000; if(profile_ms > 20000){ - V3_Net_Print(1, "Virtio NIC: TX: Pkt: %lld, Bytes: %lld\n\t\tRX Pkt: %lld. Bytes: %lld\n\t\tDropped: tx %lld, rx %lld\nInterrupts: tx %d, rx %d\nTotal Exit: %lld\n", + PrintDebug("Virtio NIC: TX: Pkt: %lld, Bytes: %lld\n\t\tRX Pkt: %lld. Bytes: %lld\n\t\tDropped: tx %lld, rx %lld\nInterrupts: tx %d, rx %d\nTotal Exit: %lld\n", net_state->stats.tx_pkts, net_state->stats.tx_bytes, net_state->stats.rx_pkts, net_state->stats.rx_bytes, net_state->stats.tx_dropped, net_state->stats.rx_dropped, @@ -906,7 +927,6 @@ static void virtio_nic_timer(struct guest_info * core, static struct v3_timer_ops timer_ops = { .update_timer = virtio_nic_timer, }; -#endif static int connect_fn(struct v3_vm_info * info, void * frontend_data, @@ -929,10 +949,34 @@ static int connect_fn(struct v3_vm_info * info, net_state->backend_data = private_data; net_state->virtio_dev = virtio; - net_state->tx_notify = 1; - net_state->rx_notify = 1; - - //net_state->timer = v3_add_timer(&(info->cores[0]), &timer_ops,net_state); + switch (virtio->model) { + case GUEST_DRIVEN: + V3_Print("Virtio NIC: Guest-driven operation\n"); + net_state->tx_notify = 1; + net_state->rx_notify = 1; + break; + case VMM_DRIVEN: + V3_Print("Virtio NIC: VMM-driven operation\n"); + net_state->tx_notify = 0; + net_state->rx_notify = 0; + break; + case ADAPTIVE: { + V3_Print("Virtio NIC: Adaptive operation (begins in guest-driven mode)\n"); + net_state->tx_notify = 1; + net_state->rx_notify = 1; + + net_state->timer = v3_add_timer(&(info->cores[0]), &timer_ops,net_state); + + } + break; + + default: + V3_Print("Virtio NIC: Unknown model, using GUEST_DRIVEN\n"); + net_state->tx_notify = 1; + net_state->rx_notify = 1; + break; + } + ops->recv = virtio_rx; ops->poll = virtio_poll; @@ -952,14 +996,77 @@ static int connect_fn(struct v3_vm_info * info, return 0; } +static int setup_perf_model(struct virtio_dev_state *virtio_state, v3_cfg_tree_t *t) +{ + char *mode = v3_cfg_val(t,"mode"); + + // defaults + virtio_state->model = GUEST_DRIVEN; + virtio_state->lower_thresh_pps = RATE_LOWER_THRESHOLD_DEFAULT; + virtio_state->upper_thresh_pps = RATE_UPPER_THRESHOLD_DEFAULT; + virtio_state->period_us = PROFILE_PERIOD_DEFAULT; + + + // overrides + if (mode) { + if (!strcasecmp(mode,"guest-driven")) { + V3_Print("Virtio NIC: Setting static GUEST_DRIVEN mode of operation (latency optimized)\n"); + virtio_state->model=GUEST_DRIVEN; + } else if (!strcasecmp(mode, "vmm-driven")) { + V3_Print("Virtio NIC: Setting static VMM_DRIVEN mode of operation (throughput optimized)\n"); + virtio_state->model=VMM_DRIVEN; + } else if (!strcasecmp(mode, "adaptive")) { + char *s; + + V3_Print("Virtio NIC: Setting dynamic ADAPTIVE mode of operation\n"); + virtio_state->model=ADAPTIVE; + + if (!(s=v3_cfg_val(t,"upper"))) { + V3_Print("Virtio NIC: No upper bound given, using default\n"); + } else { + virtio_state->upper_thresh_pps = atoi(s); + } + if (!(s=v3_cfg_val(t,"lower"))) { + V3_Print("Virtio NIC: No lower bound given, using default\n"); + } else { + virtio_state->lower_thresh_pps = atoi(s); + } + if (!(s=v3_cfg_val(t,"period"))) { + V3_Print("Virtio NIC: No period given, using default\n"); + } else { + virtio_state->period_us = atoi(s); + } + + V3_Print("Virtio NIC: lower_thresh_pps=%llu, upper_thresh_pps=%llu, period_us=%llu\n", + virtio_state->lower_thresh_pps, + virtio_state->upper_thresh_pps, + virtio_state->period_us); + } else { + PrintError("Virtio NIC: Unknown mode of operation '%s', using default (guest-driven)\n",mode); + virtio_state->model=GUEST_DRIVEN; + } + } else { + V3_Print("Virtio NIC: No model given, using default (guest-driven)\n"); + } + + return 0; + +} + +/* + + pci-bus-to-attach-to // required + mac address // if ommited with pic one + + +*/ static int virtio_init(struct v3_vm_info * vm, v3_cfg_tree_t * cfg) { struct vm_device * pci_bus = v3_find_dev(vm, v3_cfg_val(cfg, "bus")); struct virtio_dev_state * virtio_state = NULL; char * dev_id = v3_cfg_val(cfg, "ID"); - char macstr[128]; - char * str = v3_cfg_val(cfg, "mac"); - memcpy(macstr, str, strlen(str)); - + char * mac = v3_cfg_val(cfg, "mac"); + v3_cfg_tree_t *model = v3_cfg_subtree(cfg,"model"); + if (pci_bus == NULL) { PrintError("Virtio NIC: Virtio device require a PCI Bus"); return -1; @@ -978,12 +1085,24 @@ static int virtio_init(struct v3_vm_info * vm, v3_cfg_tree_t * cfg) { virtio_state->pci_bus = pci_bus; virtio_state->vm = vm; - if (macstr != NULL && !str2mac(macstr, virtio_state->mac)) { - PrintDebug("Virtio NIC: Mac specified %s\n", macstr); - }else { + if (mac) { + if (!str2mac(mac, virtio_state->mac)) { + PrintDebug("Virtio NIC: Mac specified %s\n", mac); + } else { + PrintError("Virtio NIC: Mac specified is incorrect, picking a randoom mac\n"); + random_ethaddr(virtio_state->mac); + } + } else { + PrintDebug("Virtio NIC: no mac specified, so picking a random mac\n"); random_ethaddr(virtio_state->mac); } + if (setup_perf_model(virtio_state,model)<0) { + PrintError("Cannnot setup performance model\n"); + V3_Free(virtio_state); + return -1; + } + struct vm_device * dev = v3_add_device(vm, dev_id, &dev_ops, virtio_state); if (dev == NULL) { diff --git a/palacios/src/palacios/vmm_perftune.c b/palacios/src/palacios/vmm_perftune.c index 8ffe3a9..0d58f6f 100644 --- a/palacios/src/palacios/vmm_perftune.c +++ b/palacios/src/palacios/vmm_perftune.c @@ -71,7 +71,7 @@ static void set_yield(struct v3_vm_info *vm, v3_cfg_tree_t *cfg) // now override - t = v3_cfg_val(cfg, "yield_strategy"); + t = v3_cfg_val(cfg, "strategy"); if (t) { if (!strcasecmp(t,"greedy")) { @@ -118,7 +118,7 @@ static void set_yield(struct v3_vm_info *vm, v3_cfg_tree_t *cfg) /* - + greedy,friendly,adaptive us @@ -147,7 +147,7 @@ int v3_setup_performance_tuning(struct v3_vm_info *vm, v3_cfg_tree_t *cfg) if (!id) { V3_Print("Skipping performance parameter group without name\n"); } else { - if (!strcasecmp(id,"yield_strategy")) { + if (!strcasecmp(id,"yield")) { set_yield(vm,t); } else { V3_Print("Skipping unknown performance parameter group\n"); diff --git a/palacios/src/vnet/vnet_core.c b/palacios/src/vnet/vnet_core.c index 7613368..03f412d 100644 --- a/palacios/src/vnet/vnet_core.c +++ b/palacios/src/vnet/vnet_core.c @@ -31,7 +31,7 @@ #define PrintDebug(fmt, args...) #endif -#define VNET_NUM_TX_KICK_THREADS 2 +#define VNET_NUM_TX_KICK_THREADS 1 #define VNET_ADAPTIVE_TX_KICK 0 // set to 1 to try to sleep when there is nothing to do #define VNET_NOPROGRESS_LIMIT 1000 // ... after this many tries