Merge branch 'xen-netback'

Wei Liu says:

====================
xen-netback: switch to NAPI + kthread 1:1 model

This series implements NAPI + kthread 1:1 model for Xen netback.

This model
 - provides better scheduling fairness among vifs
 - is prerequisite for implementing multiqueue for Xen network driver

The second patch has the real meat:
 - make use of NAPI to mitigate interrupt
 - kthreads are not bound to CPUs any more, so that we can take
   advantage of backend scheduler and trust it to do the right thing

Benchmark is done on a Dell T3400 workstation with 4 cores, running 4
DomUs. Netserver runs in Dom0. DomUs do netperf to Dom0 with
following command: /root/netperf -H Dom0 -fm -l120

IRQs are distributed to 4 cores by hand in the new model, while in the
old model vifs are automatically distributed to 4 kthreads.

* New model
%Cpu0  :  0.5 us, 20.3 sy,  0.0 ni, 28.9 id,  0.0 wa,  0.0 hi, 24.4 si, 25.9 st
%Cpu1  :  0.5 us, 17.8 sy,  0.0 ni, 28.8 id,  0.0 wa,  0.0 hi, 27.7 si, 25.1 st
%Cpu2  :  0.5 us, 18.8 sy,  0.0 ni, 30.7 id,  0.0 wa,  0.0 hi, 22.9 si, 27.1 st
%Cpu3  :  0.0 us, 20.1 sy,  0.0 ni, 30.4 id,  0.0 wa,  0.0 hi, 22.7 si, 26.8 st
Throughputs: 2027.89 2025.95 2018.57 2016.23 aggregated: 8088.64

* Old model
%Cpu0  :  0.5 us, 68.8 sy,  0.0 ni, 16.1 id,  0.5 wa,  0.0 hi,  2.8 si, 11.5 st
%Cpu1  :  0.4 us, 45.1 sy,  0.0 ni, 31.1 id,  0.4 wa,  0.0 hi,  2.1 si, 20.9 st
%Cpu2  :  0.9 us, 44.8 sy,  0.0 ni, 30.9 id,  0.0 wa,  0.0 hi,  1.3 si, 22.2 st
%Cpu3  :  0.8 us, 46.4 sy,  0.0 ni, 28.3 id,  1.3 wa,  0.0 hi,  2.1 si, 21.1 st
Throughputs: 1899.14 2280.43 1963.33 1893.47 aggregated: 8036.37

We can see that the impact is mainly on CPU usage. The new model moves
processing from kthread to NAPI (software interrupt).
====================
Signed-off-by: David S. Miller <davem@davemloft.net>

drivers/net/xen-netback/common.h

@@ -45,31 +45,109 @@
 #include <xen/grant_table.h>
 #include <xen/xenbus.h>
 
-struct xen_netbk;
+typedef unsigned int pending_ring_idx_t;
+#define INVALID_PENDING_RING_IDX (~0U)
+
+/* For the head field in pending_tx_info: it is used to indicate
+ * whether this tx info is the head of one or more coalesced requests.
+ *
+ * When head != INVALID_PENDING_RING_IDX, it means the start of a new
+ * tx requests queue and the end of previous queue.
+ *
+ * An example sequence of head fields (I = INVALID_PENDING_RING_IDX):
+ *
+ * ...|0 I I I|5 I|9 I I I|...
+ * -->|<-INUSE----------------
+ *
+ * After consuming the first slot(s) we have:
+ *
+ * ...|V V V V|5 I|9 I I I|...
+ * -----FREE->|<-INUSE--------
+ *
+ * where V stands for "valid pending ring index". Any number other
+ * than INVALID_PENDING_RING_IDX is OK. These entries are considered
+ * free and can contain any number other than
+ * INVALID_PENDING_RING_IDX. In practice we use 0.
+ *
+ * The in use non-INVALID_PENDING_RING_IDX (say 0, 5 and 9 in the
+ * above example) number is the index into pending_tx_info and
+ * mmap_pages arrays.
+ */
+struct pending_tx_info {
+	struct xen_netif_tx_request req; /* coalesced tx request */
+	pending_ring_idx_t head; /* head != INVALID_PENDING_RING_IDX
+				  * if it is head of one or more tx
+				  * reqs
+				  */
+};
+
+#define XEN_NETIF_TX_RING_SIZE __CONST_RING_SIZE(xen_netif_tx, PAGE_SIZE)
+#define XEN_NETIF_RX_RING_SIZE __CONST_RING_SIZE(xen_netif_rx, PAGE_SIZE)
+
+struct xenvif_rx_meta {
+	int id;
+	int size;
+	int gso_size;
+};
+
+/* Discriminate from any valid pending_idx value. */
+#define INVALID_PENDING_IDX 0xFFFF
+
+#define MAX_BUFFER_OFFSET PAGE_SIZE
+
+#define MAX_PENDING_REQS 256
 
 struct xenvif {
 	/* Unique identifier for this interface. */
 	domid_t          domid;
 	unsigned int     handle;
 
-	/* Reference to netback processing backend. */
-	struct xen_netbk *netbk;
+	/* Use NAPI for guest TX */
+	struct napi_struct napi;
+	/* When feature-split-event-channels = 0, tx_irq = rx_irq. */
+	unsigned int tx_irq;
+	/* Only used when feature-split-event-channels = 1 */
+	char tx_irq_name[IFNAMSIZ+4]; /* DEVNAME-tx */
+	struct xen_netif_tx_back_ring tx;
+	struct sk_buff_head tx_queue;
+	struct page *mmap_pages[MAX_PENDING_REQS];
+	pending_ring_idx_t pending_prod;
+	pending_ring_idx_t pending_cons;
+	u16 pending_ring[MAX_PENDING_REQS];
+	struct pending_tx_info pending_tx_info[MAX_PENDING_REQS];
+
+	/* Coalescing tx requests before copying makes number of grant
+	 * copy ops greater or equal to number of slots required. In
+	 * worst case a tx request consumes 2 gnttab_copy.
+	 */
+	struct gnttab_copy tx_copy_ops[2*MAX_PENDING_REQS];
 
-	u8               fe_dev_addr[6];
 
+	/* Use kthread for guest RX */
+	struct task_struct *task;
+	wait_queue_head_t wq;
 	/* When feature-split-event-channels = 0, tx_irq = rx_irq. */
-	unsigned int tx_irq;
 	unsigned int rx_irq;
 	/* Only used when feature-split-event-channels = 1 */
-	char tx_irq_name[IFNAMSIZ+4]; /* DEVNAME-tx */
 	char rx_irq_name[IFNAMSIZ+4]; /* DEVNAME-rx */
+	struct xen_netif_rx_back_ring rx;
+	struct sk_buff_head rx_queue;
 
-	/* List of frontends to notify after a batch of frames sent. */
-	struct list_head notify_list;
+	/* Allow xenvif_start_xmit() to peek ahead in the rx request
+	 * ring.  This is a prediction of what rx_req_cons will be
+	 * once all queued skbs are put on the ring.
+	 */
+	RING_IDX rx_req_cons_peek;
+
+	/* Given MAX_BUFFER_OFFSET of 4096 the worst case is that each
+	 * head/fragment page uses 2 copy operations because it
+	 * straddles two buffers in the frontend.
+	 */
+	struct gnttab_copy grant_copy_op[2*XEN_NETIF_RX_RING_SIZE];
+	struct xenvif_rx_meta meta[2*XEN_NETIF_RX_RING_SIZE];
 
-	/* The shared rings and indexes. */
-	struct xen_netif_tx_back_ring tx;
-	struct xen_netif_rx_back_ring rx;
+
+	u8               fe_dev_addr[6];
 
 	/* Frontend feature information. */
 	u8 can_sg:1;
@@ -80,13 +158,6 @@ struct xenvif {
 	/* Internal feature information. */
 	u8 can_queue:1;	    /* can queue packets for receiver? */
 
-	/*
-	 * Allow xenvif_start_xmit() to peek ahead in the rx request
-	 * ring.  This is a prediction of what rx_req_cons will be
-	 * once all queued skbs are put on the ring.
-	 */
-	RING_IDX rx_req_cons_peek;
-
 	/* Transmit shaping: allow 'credit_bytes' every 'credit_usec'. */
 	unsigned long   credit_bytes;
 	unsigned long   credit_usec;
@@ -97,11 +168,7 @@ struct xenvif {
 	unsigned long rx_gso_checksum_fixup;
 
 	/* Miscellaneous private stuff. */
-	struct list_head schedule_list;
-	atomic_t         refcnt;
 	struct net_device *dev;
-
-	wait_queue_head_t waiting_to_free;
 };
 
 static inline struct xenbus_device *xenvif_to_xenbus_device(struct xenvif *vif)
@@ -109,9 +176,6 @@ static inline struct xenbus_device *xenvif_to_xenbus_device(struct xenvif *vif)
 	return to_xenbus_device(vif->dev->dev.parent);
 }
 
-#define XEN_NETIF_TX_RING_SIZE __CONST_RING_SIZE(xen_netif_tx, PAGE_SIZE)
-#define XEN_NETIF_RX_RING_SIZE __CONST_RING_SIZE(xen_netif_rx, PAGE_SIZE)
-
 struct xenvif *xenvif_alloc(struct device *parent,
 			    domid_t domid,
 			    unsigned int handle);
@@ -121,39 +185,26 @@ int xenvif_connect(struct xenvif *vif, unsigned long tx_ring_ref,
 		   unsigned int rx_evtchn);
 void xenvif_disconnect(struct xenvif *vif);
 
-void xenvif_get(struct xenvif *vif);
-void xenvif_put(struct xenvif *vif);
-
 int xenvif_xenbus_init(void);
 void xenvif_xenbus_fini(void);
 
 int xenvif_schedulable(struct xenvif *vif);
 
-int xen_netbk_rx_ring_full(struct xenvif *vif);
+int xenvif_rx_ring_full(struct xenvif *vif);
 
-int xen_netbk_must_stop_queue(struct xenvif *vif);
+int xenvif_must_stop_queue(struct xenvif *vif);
 
 /* (Un)Map communication rings. */
-void xen_netbk_unmap_frontend_rings(struct xenvif *vif);
-int xen_netbk_map_frontend_rings(struct xenvif *vif,
+void xenvif_unmap_frontend_rings(struct xenvif *vif);
+int xenvif_map_frontend_rings(struct xenvif *vif,
 			      grant_ref_t tx_ring_ref,
 			      grant_ref_t rx_ring_ref);
 
-/* (De)Register a xenvif with the netback backend. */
-void xen_netbk_add_xenvif(struct xenvif *vif);
-void xen_netbk_remove_xenvif(struct xenvif *vif);
-
-/* (De)Schedule backend processing for a xenvif */
-void xen_netbk_schedule_xenvif(struct xenvif *vif);
-void xen_netbk_deschedule_xenvif(struct xenvif *vif);
-
 /* Check for SKBs from frontend and schedule backend processing */
-void xen_netbk_check_rx_xenvif(struct xenvif *vif);
-/* Receive an SKB from the frontend */
-void xenvif_receive_skb(struct xenvif *vif, struct sk_buff *skb);
+void xenvif_check_rx_xenvif(struct xenvif *vif);
 
 /* Queue an SKB for transmission to the frontend */
-void xen_netbk_queue_tx_skb(struct xenvif *vif, struct sk_buff *skb);
+void xenvif_queue_tx_skb(struct xenvif *vif, struct sk_buff *skb);
 /* Notify xenvif that ring now has space to send an skb to the frontend */
 void xenvif_notify_tx_completion(struct xenvif *vif);
 
@@ -161,7 +212,12 @@ void xenvif_notify_tx_completion(struct xenvif *vif);
 void xenvif_carrier_off(struct xenvif *vif);
 
 /* Returns number of ring slots required to send an skb to the frontend */
-unsigned int xen_netbk_count_skb_slots(struct xenvif *vif, struct sk_buff *skb);
+unsigned int xenvif_count_skb_slots(struct xenvif *vif, struct sk_buff *skb);
+
+int xenvif_tx_action(struct xenvif *vif, int budget);
+void xenvif_rx_action(struct xenvif *vif);
+
+int xenvif_kthread(void *data);
 
 extern bool separate_tx_rx_irq;
 

drivers/net/xen-netback/interface.c

@@ -30,6 +30,7 @@
 
 #include "common.h"
 
+#include <linux/kthread.h>
 #include <linux/ethtool.h>
 #include <linux/rtnetlink.h>
 #include <linux/if_vlan.h>
@@ -38,17 +39,7 @@
 #include <asm/xen/hypercall.h>
 
 #define XENVIF_QUEUE_LENGTH 32
-
-void xenvif_get(struct xenvif *vif)
-{
-	atomic_inc(&vif->refcnt);
-}
-
-void xenvif_put(struct xenvif *vif)
-{
-	if (atomic_dec_and_test(&vif->refcnt))
-		wake_up(&vif->waiting_to_free);
-}
+#define XENVIF_NAPI_WEIGHT  64
 
 int xenvif_schedulable(struct xenvif *vif)
 {
@@ -57,28 +48,62 @@ int xenvif_schedulable(struct xenvif *vif)
 
 static int xenvif_rx_schedulable(struct xenvif *vif)
 {
-	return xenvif_schedulable(vif) && !xen_netbk_rx_ring_full(vif);
+	return xenvif_schedulable(vif) && !xenvif_rx_ring_full(vif);
 }
 
 static irqreturn_t xenvif_tx_interrupt(int irq, void *dev_id)
 {
 	struct xenvif *vif = dev_id;
 
-	if (vif->netbk == NULL)
+	if (RING_HAS_UNCONSUMED_REQUESTS(&vif->tx))
+		napi_schedule(&vif->napi);
+
 	return IRQ_HANDLED;
+}
+
+static int xenvif_poll(struct napi_struct *napi, int budget)
+{
+	struct xenvif *vif = container_of(napi, struct xenvif, napi);
+	int work_done;
 
-	xen_netbk_schedule_xenvif(vif);
+	work_done = xenvif_tx_action(vif, budget);
 
-	return IRQ_HANDLED;
+	if (work_done < budget) {
+		int more_to_do = 0;
+		unsigned long flags;
+
+		/* It is necessary to disable IRQ before calling
+		 * RING_HAS_UNCONSUMED_REQUESTS. Otherwise we might
+		 * lose event from the frontend.
+		 *
+		 * Consider:
+		 *   RING_HAS_UNCONSUMED_REQUESTS
+		 *   <frontend generates event to trigger napi_schedule>
+		 *   __napi_complete
+		 *
+		 * This handler is still in scheduled state so the
+		 * event has no effect at all. After __napi_complete
+		 * this handler is descheduled and cannot get
+		 * scheduled again. We lose event in this case and the ring
+		 * will be completely stalled.
+		 */
+
+		local_irq_save(flags);
+
+		RING_FINAL_CHECK_FOR_REQUESTS(&vif->tx, more_to_do);
+		if (!more_to_do)
+			__napi_complete(napi);
+
+		local_irq_restore(flags);
+	}
+
+	return work_done;
 }
 
 static irqreturn_t xenvif_rx_interrupt(int irq, void *dev_id)
 {
 	struct xenvif *vif = dev_id;
 
-	if (vif->netbk == NULL)
-		return IRQ_HANDLED;
-
 	if (xenvif_rx_schedulable(vif))
 		netif_wake_queue(vif->dev);
 
@@ -99,7 +124,8 @@ static int xenvif_start_xmit(struct sk_buff *skb, struct net_device *dev)
 
 	BUG_ON(skb->dev != dev);
 
-	if (vif->netbk == NULL)
+	/* Drop the packet if vif is not ready */
+	if (vif->task == NULL)
 		goto drop;
 
 	/* Drop the packet if the target domain has no receive buffers. */
@@ -107,13 +133,12 @@ static int xenvif_start_xmit(struct sk_buff *skb, struct net_device *dev)
 		goto drop;
 
 	/* Reserve ring slots for the worst-case number of fragments. */
-	vif->rx_req_cons_peek += xen_netbk_count_skb_slots(vif, skb);
-	xenvif_get(vif);
+	vif->rx_req_cons_peek += xenvif_count_skb_slots(vif, skb);
 
-	if (vif->can_queue && xen_netbk_must_stop_queue(vif))
+	if (vif->can_queue && xenvif_must_stop_queue(vif))
 		netif_stop_queue(dev);
 
-	xen_netbk_queue_tx_skb(vif, skb);
+	xenvif_queue_tx_skb(vif, skb);
 
 	return NETDEV_TX_OK;
 
@@ -123,11 +148,6 @@ static int xenvif_start_xmit(struct sk_buff *skb, struct net_device *dev)
 	return NETDEV_TX_OK;
 }
 
-void xenvif_receive_skb(struct xenvif *vif, struct sk_buff *skb)
-{
-	netif_rx_ni(skb);
-}
-
 void xenvif_notify_tx_completion(struct xenvif *vif)
 {
 	if (netif_queue_stopped(vif->dev) && xenvif_rx_schedulable(vif))
@@ -142,21 +162,20 @@ static struct net_device_stats *xenvif_get_stats(struct net_device *dev)
 
 static void xenvif_up(struct xenvif *vif)
 {
-	xen_netbk_add_xenvif(vif);
+	napi_enable(&vif->napi);
 	enable_irq(vif->tx_irq);
 	if (vif->tx_irq != vif->rx_irq)
 		enable_irq(vif->rx_irq);
-	xen_netbk_check_rx_xenvif(vif);
+	xenvif_check_rx_xenvif(vif);
 }
 
 static void xenvif_down(struct xenvif *vif)
 {
+	napi_disable(&vif->napi);
 	disable_irq(vif->tx_irq);
 	if (vif->tx_irq != vif->rx_irq)
 		disable_irq(vif->rx_irq);
 	del_timer_sync(&vif->credit_timeout);
-	xen_netbk_deschedule_xenvif(vif);
-	xen_netbk_remove_xenvif(vif);
 }
 
 static int xenvif_open(struct net_device *dev)
@@ -272,11 +291,12 @@ struct xenvif *xenvif_alloc(struct device *parent, domid_t domid,
 	struct net_device *dev;
 	struct xenvif *vif;
 	char name[IFNAMSIZ] = {};
+	int i;
 
 	snprintf(name, IFNAMSIZ - 1, "vif%u.%u", domid, handle);
 	dev = alloc_netdev(sizeof(struct xenvif), name, ether_setup);
 	if (dev == NULL) {
-		pr_warn("Could not allocate netdev\n");
+		pr_warn("Could not allocate netdev for %s\n", name);
 		return ERR_PTR(-ENOMEM);
 	}
 
@@ -285,14 +305,9 @@ struct xenvif *xenvif_alloc(struct device *parent, domid_t domid,
 	vif = netdev_priv(dev);
 	vif->domid  = domid;
 	vif->handle = handle;
-	vif->netbk  = NULL;
 	vif->can_sg = 1;
 	vif->csum = 1;
-	atomic_set(&vif->refcnt, 1);
-	init_waitqueue_head(&vif->waiting_to_free);
 	vif->dev = dev;
-	INIT_LIST_HEAD(&vif->schedule_list);
-	INIT_LIST_HEAD(&vif->notify_list);
 
 	vif->credit_bytes = vif->remaining_credit = ~0UL;
 	vif->credit_usec  = 0UL;
@@ -307,6 +322,16 @@ struct xenvif *xenvif_alloc(struct device *parent, domid_t domid,
 
 	dev->tx_queue_len = XENVIF_QUEUE_LENGTH;
 
+	skb_queue_head_init(&vif->rx_queue);
+	skb_queue_head_init(&vif->tx_queue);
+
+	vif->pending_cons = 0;
+	vif->pending_prod = MAX_PENDING_REQS;
+	for (i = 0; i < MAX_PENDING_REQS; i++)
+		vif->pending_ring[i] = i;
+	for (i = 0; i < MAX_PENDING_REQS; i++)
+		vif->mmap_pages[i] = NULL;
+
 	/*
 	 * Initialise a dummy MAC address. We choose the numerically
 	 * largest non-broadcast address to prevent the address getting
@@ -316,6 +341,8 @@ struct xenvif *xenvif_alloc(struct device *parent, domid_t domid,
 	memset(dev->dev_addr, 0xFF, ETH_ALEN);
 	dev->dev_addr[0] &= ~0x01;
 
+	netif_napi_add(dev, &vif->napi, xenvif_poll, XENVIF_NAPI_WEIGHT);
+
 	netif_carrier_off(dev);
 
 	err = register_netdev(dev);
@@ -341,7 +368,7 @@ int xenvif_connect(struct xenvif *vif, unsigned long tx_ring_ref,
 
 	__module_get(THIS_MODULE);
 
-	err = xen_netbk_map_frontend_rings(vif, tx_ring_ref, rx_ring_ref);
+	err = xenvif_map_frontend_rings(vif, tx_ring_ref, rx_ring_ref);
 	if (err < 0)
 		goto err;
 
@@ -377,7 +404,14 @@ int xenvif_connect(struct xenvif *vif, unsigned long tx_ring_ref,
 		disable_irq(vif->rx_irq);
 	}
 
-	xenvif_get(vif);
+	init_waitqueue_head(&vif->wq);
+	vif->task = kthread_create(xenvif_kthread,
+				   (void *)vif, vif->dev->name);
+	if (IS_ERR(vif->task)) {
+		pr_warn("Could not allocate kthread for %s\n", vif->dev->name);
+		err = PTR_ERR(vif->task);
+		goto err_rx_unbind;
+	}
 
 	rtnl_lock();
 	if (!vif->can_sg && vif->dev->mtu > ETH_DATA_LEN)
@@ -388,12 +422,18 @@ int xenvif_connect(struct xenvif *vif, unsigned long tx_ring_ref,
 		xenvif_up(vif);
 	rtnl_unlock();
 
+	wake_up_process(vif->task);
+
 	return 0;
+
+err_rx_unbind:
+	unbind_from_irqhandler(vif->rx_irq, vif);
+	vif->rx_irq = 0;
 err_tx_unbind:
 	unbind_from_irqhandler(vif->tx_irq, vif);
 	vif->tx_irq = 0;
 err_unmap:
-	xen_netbk_unmap_frontend_rings(vif);
+	xenvif_unmap_frontend_rings(vif);
 err:
 	module_put(THIS_MODULE);
 	return err;
@@ -408,7 +448,6 @@ void xenvif_carrier_off(struct xenvif *vif)
 	if (netif_running(dev))
 		xenvif_down(vif);
 	rtnl_unlock();
-	xenvif_put(vif);
 }
 
 void xenvif_disconnect(struct xenvif *vif)
@@ -422,9 +461,6 @@ void xenvif_disconnect(struct xenvif *vif)
 	if (netif_carrier_ok(vif->dev))
 		xenvif_carrier_off(vif);
 
-	atomic_dec(&vif->refcnt);
-	wait_event(vif->waiting_to_free, atomic_read(&vif->refcnt) == 0);
-
 	if (vif->tx_irq) {
 		if (vif->tx_irq == vif->rx_irq)
 			unbind_from_irqhandler(vif->tx_irq, vif);
@@ -438,9 +474,14 @@ void xenvif_disconnect(struct xenvif *vif)
 		need_module_put = 1;
 	}
 
+	if (vif->task)
+		kthread_stop(vif->task);
+
+	netif_napi_del(&vif->napi);
+
 	unregister_netdev(vif->dev);
 
-	xen_netbk_unmap_frontend_rings(vif);
+	xenvif_unmap_frontend_rings(vif);
 
 	free_netdev(vif->dev);