[PATCH 0/2 V3] xen-netback: switch to NAPI + kthread 1:1 model

[PATCH 0/2 V3] xen-netback: switch to NAPI + kthread 1:1 model

[Wei Liu]

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 first two patches are ground work for the third patch. First one
simplifies code in netback, second one can reduce memory footprint if we
switch to 1:1 model.

The third 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 done on a Dell T3400 workstation with 4 cores, running 4
DomUs. Netserver running 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).

Comments?

----------------------------------------------------
Change since V2:
 - Remove per-cpu scratch space.

Change since V1:
 - No page pool in this version. Instead page tracking facility is
   removed.

Wei Liu (2):
  xen-netback: remove page tracking facility
  xen-netback: switch to NAPI + kthread 1:1 model

 drivers/net/xen-netback/common.h    |  105 +++--
 drivers/net/xen-netback/interface.c |  122 ++++--
 drivers/net/xen-netback/netback.c   |  817 +++++++++++------------------------
 3 files changed, 390 insertions(+), 654 deletions(-)

-- 
1.7.10.4

[PATCH 1/2] xen-netback: remove page tracking facility

[Wei Liu]

The data flow from DomU to DomU on the same host in current copying
scheme with tracking facility:

       copy
DomU --------> Dom0          DomU
 |                            ^
 |____________________________|
             copy

The page in Dom0 is a page with valid MFN. So we can always copy from
page Dom0, thus removing the need for a tracking facility.

       copy           copy
DomU --------> Dom0 -------> DomU

Simple iperf test shows no performance regression (obviously we do two
copy's either way):

  W/  tracking: ~5.3Gb/s
  W/o tracking: ~5.4Gb/s

Signed-off-by: Wei Liu <wei.liu2@citrix.com>
---
 drivers/net/xen-netback/netback.c |   77 +------------------------------------
 1 file changed, 2 insertions(+), 75 deletions(-)

diff --git a/drivers/net/xen-netback/netback.c b/drivers/net/xen-netback/netback.c
index 82576ff..54853be 100644
--- a/drivers/net/xen-netback/netback.c
+++ b/drivers/net/xen-netback/netback.c
@@ -95,21 +95,6 @@ struct netbk_rx_meta {
 
 #define MAX_BUFFER_OFFSET PAGE_SIZE
 
-/* extra field used in struct page */
-union page_ext {
-	struct {
-#if BITS_PER_LONG < 64
-#define IDX_WIDTH   8
-#define GROUP_WIDTH (BITS_PER_LONG - IDX_WIDTH)
-		unsigned int group:GROUP_WIDTH;
-		unsigned int idx:IDX_WIDTH;
-#else
-		unsigned int group, idx;
-#endif
-	} e;
-	void *mapping;
-};
-
 struct xen_netbk {
 	wait_queue_head_t wq;
 	struct task_struct *task;
@@ -214,45 +199,6 @@ static inline unsigned long idx_to_kaddr(struct xen_netbk *netbk,
 	return (unsigned long)pfn_to_kaddr(idx_to_pfn(netbk, idx));
 }
 
-/* extra field used in struct page */
-static inline void set_page_ext(struct page *pg, struct xen_netbk *netbk,
-				unsigned int idx)
-{
-	unsigned int group = netbk - xen_netbk;
-	union page_ext ext = { .e = { .group = group + 1, .idx = idx } };
-
-	BUILD_BUG_ON(sizeof(ext) > sizeof(ext.mapping));
-	pg->mapping = ext.mapping;
-}
-
-static int get_page_ext(struct page *pg,
-			unsigned int *pgroup, unsigned int *pidx)
-{
-	union page_ext ext = { .mapping = pg->mapping };
-	struct xen_netbk *netbk;
-	unsigned int group, idx;
-
-	group = ext.e.group - 1;
-
-	if (group < 0 || group >= xen_netbk_group_nr)
-		return 0;
-
-	netbk = &xen_netbk[group];
-
-	idx = ext.e.idx;
-
-	if ((idx < 0) || (idx >= MAX_PENDING_REQS))
-		return 0;
-
-	if (netbk->mmap_pages[idx] != pg)
-		return 0;
-
-	*pgroup = group;
-	*pidx = idx;
-
-	return 1;
-}
-
 /*
  * This is the amount of packet we copy rather than map, so that the
  * guest can't fiddle with the contents of the headers while we do
@@ -453,12 +399,6 @@ static void netbk_gop_frag_copy(struct xenvif *vif, struct sk_buff *skb,
 {
 	struct gnttab_copy *copy_gop;
 	struct netbk_rx_meta *meta;
-	/*
-	 * These variables are used iff get_page_ext returns true,
-	 * in which case they are guaranteed to be initialized.
-	 */
-	unsigned int uninitialized_var(group), uninitialized_var(idx);
-	int foreign = get_page_ext(page, &group, &idx);
 	unsigned long bytes;
 
 	/* Data must not cross a page boundary. */
@@ -494,20 +434,9 @@ static void netbk_gop_frag_copy(struct xenvif *vif, struct sk_buff *skb,
 
 		copy_gop = npo->copy + npo->copy_prod++;
 		copy_gop->flags = GNTCOPY_dest_gref;
-		if (foreign) {
-			struct xen_netbk *netbk = &xen_netbk[group];
-			struct pending_tx_info *src_pend;
+		copy_gop->source.domid = DOMID_SELF;
+		copy_gop->source.u.gmfn = virt_to_mfn(page_address(page));
 
-			src_pend = &netbk->pending_tx_info[idx];
-
-			copy_gop->source.domid = src_pend->vif->domid;
-			copy_gop->source.u.ref = src_pend->req.gref;
-			copy_gop->flags |= GNTCOPY_source_gref;
-		} else {
-			void *vaddr = page_address(page);
-			copy_gop->source.domid = DOMID_SELF;
-			copy_gop->source.u.gmfn = virt_to_mfn(vaddr);
-		}
 		copy_gop->source.offset = offset;
 		copy_gop->dest.domid = vif->domid;
 
@@ -1045,7 +974,6 @@ static struct page *xen_netbk_alloc_page(struct xen_netbk *netbk,
 	page = alloc_page(GFP_KERNEL|__GFP_COLD);
 	if (!page)
 		return NULL;
-	set_page_ext(page, netbk, pending_idx);
 	netbk->mmap_pages[pending_idx] = page;
 	return page;
 }
@@ -1153,7 +1081,6 @@ static struct gnttab_copy *xen_netbk_get_requests(struct xen_netbk *netbk,
 		first->req.offset = 0;
 		first->req.size = dst_offset;
 		first->head = start_idx;
-		set_page_ext(page, netbk, head_idx);
 		netbk->mmap_pages[head_idx] = page;
 		frag_set_pending_idx(&frags[shinfo->nr_frags], head_idx);
 	}
-- 
1.7.10.4

Re: [PATCH 1/2] xen-netback: remove page tracking facility

[Ian Campbell]

On Wed, 2013-05-29 at 12:43 +0100, Wei Liu wrote:
> The data flow from DomU to DomU on the same host in current copying
> scheme with tracking facility:
> 
>        copy
> DomU --------> Dom0          DomU
>  |                            ^
>  |____________________________|
>              copy

Um, wow, that was certainly unintentional!

> The page in Dom0 is a page with valid MFN. So we can always copy from
> page Dom0, thus removing the need for a tracking facility.
> 
>        copy           copy
> DomU --------> Dom0 -------> DomU

Much nicer!

> Simple iperf test shows no performance regression (obviously we do two
> copy's either way):
> 
>   W/  tracking: ~5.3Gb/s
>   W/o tracking: ~5.4Gb/s
> 
> Signed-off-by: Wei Liu <wei.liu2@citrix.com>

Acked-by: Ian Campbell <ian.campbell@citrix.com>

> ---
>  drivers/net/xen-netback/netback.c |   77 +------------------------------------
>  1 file changed, 2 insertions(+), 75 deletions(-)
> 
> diff --git a/drivers/net/xen-netback/netback.c b/drivers/net/xen-netback/netback.c
> index 82576ff..54853be 100644
> --- a/drivers/net/xen-netback/netback.c
> +++ b/drivers/net/xen-netback/netback.c
> @@ -95,21 +95,6 @@ struct netbk_rx_meta {
>  
>  #define MAX_BUFFER_OFFSET PAGE_SIZE
>  
> -/* extra field used in struct page */
> -union page_ext {
> -	struct {
> -#if BITS_PER_LONG < 64
> -#define IDX_WIDTH   8
> -#define GROUP_WIDTH (BITS_PER_LONG - IDX_WIDTH)
> -		unsigned int group:GROUP_WIDTH;
> -		unsigned int idx:IDX_WIDTH;
> -#else
> -		unsigned int group, idx;
> -#endif
> -	} e;
> -	void *mapping;
> -};
> -
>  struct xen_netbk {
>  	wait_queue_head_t wq;
>  	struct task_struct *task;
> @@ -214,45 +199,6 @@ static inline unsigned long idx_to_kaddr(struct xen_netbk *netbk,
>  	return (unsigned long)pfn_to_kaddr(idx_to_pfn(netbk, idx));
>  }
>  
> -/* extra field used in struct page */
> -static inline void set_page_ext(struct page *pg, struct xen_netbk *netbk,
> -				unsigned int idx)
> -{
> -	unsigned int group = netbk - xen_netbk;
> -	union page_ext ext = { .e = { .group = group + 1, .idx = idx } };
> -
> -	BUILD_BUG_ON(sizeof(ext) > sizeof(ext.mapping));
> -	pg->mapping = ext.mapping;
> -}
> -
> -static int get_page_ext(struct page *pg,
> -			unsigned int *pgroup, unsigned int *pidx)
> -{
> -	union page_ext ext = { .mapping = pg->mapping };
> -	struct xen_netbk *netbk;
> -	unsigned int group, idx;
> -
> -	group = ext.e.group - 1;
> -
> -	if (group < 0 || group >= xen_netbk_group_nr)
> -		return 0;
> -
> -	netbk = &xen_netbk[group];
> -
> -	idx = ext.e.idx;
> -
> -	if ((idx < 0) || (idx >= MAX_PENDING_REQS))
> -		return 0;
> -
> -	if (netbk->mmap_pages[idx] != pg)
> -		return 0;
> -
> -	*pgroup = group;
> -	*pidx = idx;
> -
> -	return 1;
> -}
> -
>  /*
>   * This is the amount of packet we copy rather than map, so that the
>   * guest can't fiddle with the contents of the headers while we do
> @@ -453,12 +399,6 @@ static void netbk_gop_frag_copy(struct xenvif *vif, struct sk_buff *skb,
>  {
>  	struct gnttab_copy *copy_gop;
>  	struct netbk_rx_meta *meta;
> -	/*
> -	 * These variables are used iff get_page_ext returns true,
> -	 * in which case they are guaranteed to be initialized.
> -	 */
> -	unsigned int uninitialized_var(group), uninitialized_var(idx);
> -	int foreign = get_page_ext(page, &group, &idx);
>  	unsigned long bytes;
>  
>  	/* Data must not cross a page boundary. */
> @@ -494,20 +434,9 @@ static void netbk_gop_frag_copy(struct xenvif *vif, struct sk_buff *skb,
>  
>  		copy_gop = npo->copy + npo->copy_prod++;
>  		copy_gop->flags = GNTCOPY_dest_gref;
> -		if (foreign) {
> -			struct xen_netbk *netbk = &xen_netbk[group];
> -			struct pending_tx_info *src_pend;
> +		copy_gop->source.domid = DOMID_SELF;
> +		copy_gop->source.u.gmfn = virt_to_mfn(page_address(page));
>  
> -			src_pend = &netbk->pending_tx_info[idx];
> -
> -			copy_gop->source.domid = src_pend->vif->domid;
> -			copy_gop->source.u.ref = src_pend->req.gref;
> -			copy_gop->flags |= GNTCOPY_source_gref;
> -		} else {
> -			void *vaddr = page_address(page);
> -			copy_gop->source.domid = DOMID_SELF;
> -			copy_gop->source.u.gmfn = virt_to_mfn(vaddr);
> -		}
>  		copy_gop->source.offset = offset;
>  		copy_gop->dest.domid = vif->domid;
>  
> @@ -1045,7 +974,6 @@ static struct page *xen_netbk_alloc_page(struct xen_netbk *netbk,
>  	page = alloc_page(GFP_KERNEL|__GFP_COLD);
>  	if (!page)
>  		return NULL;
> -	set_page_ext(page, netbk, pending_idx);
>  	netbk->mmap_pages[pending_idx] = page;
>  	return page;
>  }
> @@ -1153,7 +1081,6 @@ static struct gnttab_copy *xen_netbk_get_requests(struct xen_netbk *netbk,
>  		first->req.offset = 0;
>  		first->req.size = dst_offset;
>  		first->head = start_idx;
> -		set_page_ext(page, netbk, head_idx);
>  		netbk->mmap_pages[head_idx] = page;
>  		frag_set_pending_idx(&frags[shinfo->nr_frags], head_idx);
>  	}

[PATCH 2/2] xen-netback: switch to NAPI + kthread 1:1 model

[Wei Liu]

This patch implements 1:1 model netback. NAPI and kthread are utilized
to do the weight-lifting job:

 - NAPI is used for guest side TX (host side RX)
 - kthread is used for guest side RX (host side TX)

Xenvif and xen_netbk are made into one structure to reduce code size.

This model provides better scheduling fairness among vifs. It is also
prerequisite for implementing multiqueue for Xen netback.

Signed-off-by: Wei Liu <wei.liu2@citrix.com>
---
 drivers/net/xen-netback/common.h    |  105 +++--
 drivers/net/xen-netback/interface.c |  122 +++---
 drivers/net/xen-netback/netback.c   |  742 ++++++++++++-----------------------
 3 files changed, 389 insertions(+), 580 deletions(-)

diff --git a/drivers/net/xen-netback/common.h b/drivers/net/xen-netback/common.h
index 8a4d77e..2e270cf 100644
--- a/drivers/net/xen-netback/common.h
+++ b/drivers/net/xen-netback/common.h
@@ -45,15 +45,43 @@
 #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)
+
+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;
+	/* Use kthread for guest RX */
+	struct task_struct *task;
+	wait_queue_head_t wq;
 
 	u8               fe_dev_addr[6];
 
@@ -64,9 +92,6 @@ struct xenvif {
 	char tx_irq_name[IFNAMSIZ+4]; /* DEVNAME-tx */
 	char rx_irq_name[IFNAMSIZ+4]; /* DEVNAME-rx */
 
-	/* List of frontends to notify after a batch of frames sent. */
-	struct list_head notify_list;
-
 	/* The shared rings and indexes. */
 	struct xen_netif_tx_back_ring tx;
 	struct xen_netif_rx_back_ring rx;
@@ -96,12 +121,33 @@ struct xenvif {
 	/* Statistics */
 	unsigned long rx_gso_checksum_fixup;
 
+	struct sk_buff_head rx_queue;
+	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];
+
+	/*
+	 * 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];
+
 	/* 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 +155,6 @@ 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 +164,26 @@ 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,
-				 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);
+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);
 
 /* 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 +191,12 @@ 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;
 
diff --git a/drivers/net/xen-netback/interface.c b/drivers/net/xen-netback/interface.c
index 087d2db..3d30e93 100644
--- a/drivers/net/xen-netback/interface.c
+++ b/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 +48,46 @@ 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)
-		return IRQ_HANDLED;
-
-	xen_netbk_schedule_xenvif(vif);
+	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;
+
+	work_done = xenvif_tx_action(vif, budget);
+
+	if (work_done < budget) {
+		int more_to_do = 0;
+		unsigned long flags;
+
+		local_irq_save(flags);
+
+		RING_FINAL_CHECK_FOR_REQUESTS(&vif->tx, more_to_do);
+		if (!more_to_do || work_done < 0)
+			__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 +108,8 @@ 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 +117,12 @@ 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 +132,6 @@ 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 +146,20 @@ 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 +275,13 @@ 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");
+		printk(KERN_WARNING "xen-netback: Could not allocate netdev for vif%d.%d\n",
+		       domid, handle);
 		return ERR_PTR(-ENOMEM);
 	}
 
@@ -285,14 +290,9 @@ 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 +307,16 @@ 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 +326,8 @@ 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 +353,7 @@ 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 +389,16 @@ 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%d.%d", vif->domid, vif->handle);
+	if (IS_ERR(vif->task)) {
+		printk(KERN_WARNING "xen-netback: Could not allocate kthread for vif%d.%d\n",
+		       vif->domid, vif->handle);
+		err = PTR_ERR(vif->task);
+		goto err_rx_unbind;
+	}
 
 	rtnl_lock();
 	if (!vif->can_sg && vif->dev->mtu > ETH_DATA_LEN)
@@ -388,12 +409,18 @@ 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 +435,6 @@ 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 +448,6 @@ 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 +461,14 @@ 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);
 
diff --git a/drivers/net/xen-netback/netback.c b/drivers/net/xen-netback/netback.c
index 54853be..09f32f2 100644
--- a/drivers/net/xen-netback/netback.c
+++ b/drivers/net/xen-netback/netback.c
@@ -70,116 +70,25 @@ module_param(fatal_skb_slots, uint, 0444);
  */
 #define XEN_NETBK_LEGACY_SLOTS_MAX XEN_NETIF_NR_SLOTS_MIN
 
-typedef unsigned int pending_ring_idx_t;
-#define INVALID_PENDING_RING_IDX (~0U)
-
-struct pending_tx_info {
-	struct xen_netif_tx_request req; /* coalesced tx request */
-	struct xenvif *vif;
-	pending_ring_idx_t head; /* head != INVALID_PENDING_RING_IDX
-				  * if it is head of one or more tx
-				  * reqs
-				  */
-};
-
-struct netbk_rx_meta {
-	int id;
-	int size;
-	int gso_size;
-};
-
-#define MAX_PENDING_REQS 256
-
-/* Discriminate from any valid pending_idx value. */
-#define INVALID_PENDING_IDX 0xFFFF
-
-#define MAX_BUFFER_OFFSET PAGE_SIZE
-
-struct xen_netbk {
-	wait_queue_head_t wq;
-	struct task_struct *task;
-
-	struct sk_buff_head rx_queue;
-	struct sk_buff_head tx_queue;
-
-	struct timer_list net_timer;
-
-	struct page *mmap_pages[MAX_PENDING_REQS];
-
-	pending_ring_idx_t pending_prod;
-	pending_ring_idx_t pending_cons;
-	struct list_head net_schedule_list;
-
-	/* Protect the net_schedule_list in netif. */
-	spinlock_t net_schedule_list_lock;
-
-	atomic_t netfront_count;
-
-	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];
-
-	u16 pending_ring[MAX_PENDING_REQS];
-
-	/*
-	 * 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 netbk_rx_meta meta[2*XEN_NETIF_RX_RING_SIZE];
-};
-
-static struct xen_netbk *xen_netbk;
-static int xen_netbk_group_nr;
-
 /*
  * If head != INVALID_PENDING_RING_IDX, it means this tx request is head of
  * one or more merged tx requests, otherwise it is the continuation of
  * previous tx request.
  */
-static inline int pending_tx_is_head(struct xen_netbk *netbk, RING_IDX idx)
+static inline int pending_tx_is_head(struct xenvif *vif, RING_IDX idx)
 {
-	return netbk->pending_tx_info[idx].head != INVALID_PENDING_RING_IDX;
+	return vif->pending_tx_info[idx].head != INVALID_PENDING_RING_IDX;
 }
 
-void xen_netbk_add_xenvif(struct xenvif *vif)
-{
-	int i;
-	int min_netfront_count;
-	int min_group = 0;
-	struct xen_netbk *netbk;
-
-	min_netfront_count = atomic_read(&xen_netbk[0].netfront_count);
-	for (i = 0; i < xen_netbk_group_nr; i++) {
-		int netfront_count = atomic_read(&xen_netbk[i].netfront_count);
-		if (netfront_count < min_netfront_count) {
-			min_group = i;
-			min_netfront_count = netfront_count;
-		}
-	}
-
-	netbk = &xen_netbk[min_group];
-
-	vif->netbk = netbk;
-	atomic_inc(&netbk->netfront_count);
-}
-
-void xen_netbk_remove_xenvif(struct xenvif *vif)
-{
-	struct xen_netbk *netbk = vif->netbk;
-	vif->netbk = NULL;
-	atomic_dec(&netbk->netfront_count);
-}
-
-static void xen_netbk_idx_release(struct xen_netbk *netbk, u16 pending_idx,
-				  u8 status);
+static void xenvif_idx_release(struct xenvif *vif, u16 pending_idx,
+			       u8 status);
 static void make_tx_response(struct xenvif *vif,
 			     struct xen_netif_tx_request *txp,
 			     s8       st);
+
+static inline int tx_work_todo(struct xenvif *vif);
+static inline int rx_work_todo(struct xenvif *vif);
+
 static struct xen_netif_rx_response *make_rx_response(struct xenvif *vif,
 					     u16      id,
 					     s8       st,
@@ -187,16 +96,16 @@ static struct xen_netif_rx_response *make_rx_response(struct xenvif *vif,
 					     u16      size,
 					     u16      flags);
 
-static inline unsigned long idx_to_pfn(struct xen_netbk *netbk,
+static inline unsigned long idx_to_pfn(struct xenvif *vif,
 				       u16 idx)
 {
-	return page_to_pfn(netbk->mmap_pages[idx]);
+	return page_to_pfn(vif->mmap_pages[idx]);
 }
 
-static inline unsigned long idx_to_kaddr(struct xen_netbk *netbk,
+static inline unsigned long idx_to_kaddr(struct xenvif *vif,
 					 u16 idx)
 {
-	return (unsigned long)pfn_to_kaddr(idx_to_pfn(netbk, idx));
+	return (unsigned long)pfn_to_kaddr(idx_to_pfn(vif, idx));
 }
 
 /*
@@ -224,15 +133,10 @@ static inline pending_ring_idx_t pending_index(unsigned i)
 	return i & (MAX_PENDING_REQS-1);
 }
 
-static inline pending_ring_idx_t nr_pending_reqs(struct xen_netbk *netbk)
+static inline pending_ring_idx_t nr_pending_reqs(struct xenvif *vif)
 {
 	return MAX_PENDING_REQS -
-		netbk->pending_prod + netbk->pending_cons;
-}
-
-static void xen_netbk_kick_thread(struct xen_netbk *netbk)
-{
-	wake_up(&netbk->wq);
+		vif->pending_prod + vif->pending_cons;
 }
 
 static int max_required_rx_slots(struct xenvif *vif)
@@ -246,7 +150,7 @@ static int max_required_rx_slots(struct xenvif *vif)
 	return max;
 }
 
-int xen_netbk_rx_ring_full(struct xenvif *vif)
+int xenvif_rx_ring_full(struct xenvif *vif)
 {
 	RING_IDX peek   = vif->rx_req_cons_peek;
 	RING_IDX needed = max_required_rx_slots(vif);
@@ -255,16 +159,16 @@ int xen_netbk_rx_ring_full(struct xenvif *vif)
 	       ((vif->rx.rsp_prod_pvt + XEN_NETIF_RX_RING_SIZE - peek) < needed);
 }
 
-int xen_netbk_must_stop_queue(struct xenvif *vif)
+int xenvif_must_stop_queue(struct xenvif *vif)
 {
-	if (!xen_netbk_rx_ring_full(vif))
+	if (!xenvif_rx_ring_full(vif))
 		return 0;
 
 	vif->rx.sring->req_event = vif->rx_req_cons_peek +
 		max_required_rx_slots(vif);
 	mb(); /* request notification /then/ check the queue */
 
-	return xen_netbk_rx_ring_full(vif);
+	return xenvif_rx_ring_full(vif);
 }
 
 /*
@@ -310,9 +214,9 @@ static bool start_new_rx_buffer(int offset, unsigned long size, int head)
 /*
  * Figure out how many ring slots we're going to need to send @skb to
  * the guest. This function is essentially a dry run of
- * netbk_gop_frag_copy.
+ * xenvif_gop_frag_copy.
  */
-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)
 {
 	unsigned int count;
 	int i, copy_off;
@@ -364,15 +268,15 @@ struct netrx_pending_operations {
 	unsigned copy_prod, copy_cons;
 	unsigned meta_prod, meta_cons;
 	struct gnttab_copy *copy;
-	struct netbk_rx_meta *meta;
+	struct xenvif_rx_meta *meta;
 	int copy_off;
 	grant_ref_t copy_gref;
 };
 
-static struct netbk_rx_meta *get_next_rx_buffer(struct xenvif *vif,
-						struct netrx_pending_operations *npo)
+static struct xenvif_rx_meta *get_next_rx_buffer(struct xenvif *vif,
+						 struct netrx_pending_operations *npo)
 {
-	struct netbk_rx_meta *meta;
+	struct xenvif_rx_meta *meta;
 	struct xen_netif_rx_request *req;
 
 	req = RING_GET_REQUEST(&vif->rx, vif->rx.req_cons++);
@@ -392,13 +296,13 @@ static struct netbk_rx_meta *get_next_rx_buffer(struct xenvif *vif,
  * Set up the grant operations for this fragment. If it's a flipping
  * interface, we also set up the unmap request from here.
  */
-static void netbk_gop_frag_copy(struct xenvif *vif, struct sk_buff *skb,
-				struct netrx_pending_operations *npo,
-				struct page *page, unsigned long size,
-				unsigned long offset, int *head)
+static void xenvif_gop_frag_copy(struct xenvif *vif, struct sk_buff *skb,
+				 struct netrx_pending_operations *npo,
+				 struct page *page, unsigned long size,
+				 unsigned long offset, int *head)
 {
 	struct gnttab_copy *copy_gop;
-	struct netbk_rx_meta *meta;
+	struct xenvif_rx_meta *meta;
 	unsigned long bytes;
 
 	/* Data must not cross a page boundary. */
@@ -434,9 +338,9 @@ static void netbk_gop_frag_copy(struct xenvif *vif, struct sk_buff *skb,
 
 		copy_gop = npo->copy + npo->copy_prod++;
 		copy_gop->flags = GNTCOPY_dest_gref;
+
 		copy_gop->source.domid = DOMID_SELF;
 		copy_gop->source.u.gmfn = virt_to_mfn(page_address(page));
-
 		copy_gop->source.offset = offset;
 		copy_gop->dest.domid = vif->domid;
 
@@ -478,14 +382,14 @@ static void netbk_gop_frag_copy(struct xenvif *vif, struct sk_buff *skb,
  * zero GSO descriptors (for non-GSO packets) or one descriptor (for
  * frontend-side LRO).
  */
-static int netbk_gop_skb(struct sk_buff *skb,
-			 struct netrx_pending_operations *npo)
+static int xenvif_gop_skb(struct sk_buff *skb,
+			  struct netrx_pending_operations *npo)
 {
 	struct xenvif *vif = netdev_priv(skb->dev);
 	int nr_frags = skb_shinfo(skb)->nr_frags;
 	int i;
 	struct xen_netif_rx_request *req;
-	struct netbk_rx_meta *meta;
+	struct xenvif_rx_meta *meta;
 	unsigned char *data;
 	int head = 1;
 	int old_meta_prod;
@@ -522,30 +426,30 @@ static int netbk_gop_skb(struct sk_buff *skb,
 		if (data + len > skb_tail_pointer(skb))
 			len = skb_tail_pointer(skb) - data;
 
-		netbk_gop_frag_copy(vif, skb, npo,
-				    virt_to_page(data), len, offset, &head);
+		xenvif_gop_frag_copy(vif, skb, npo,
+				     virt_to_page(data), len, offset, &head);
 		data += len;
 	}
 
 	for (i = 0; i < nr_frags; i++) {
-		netbk_gop_frag_copy(vif, skb, npo,
-				    skb_frag_page(&skb_shinfo(skb)->frags[i]),
-				    skb_frag_size(&skb_shinfo(skb)->frags[i]),
-				    skb_shinfo(skb)->frags[i].page_offset,
-				    &head);
+		xenvif_gop_frag_copy(vif, skb, npo,
+				     skb_frag_page(&skb_shinfo(skb)->frags[i]),
+				     skb_frag_size(&skb_shinfo(skb)->frags[i]),
+				     skb_shinfo(skb)->frags[i].page_offset,
+				     &head);
 	}
 
 	return npo->meta_prod - old_meta_prod;
 }
 
 /*
- * This is a twin to netbk_gop_skb.  Assume that netbk_gop_skb was
+ * This is a twin to xenvif_gop_skb.  Assume that xenvif_gop_skb was
  * used to set up the operations on the top of
  * netrx_pending_operations, which have since been done.  Check that
  * they didn't give any errors and advance over them.
  */
-static int netbk_check_gop(struct xenvif *vif, int nr_meta_slots,
-			   struct netrx_pending_operations *npo)
+static int xenvif_check_gop(struct xenvif *vif, int nr_meta_slots,
+			    struct netrx_pending_operations *npo)
 {
 	struct gnttab_copy     *copy_op;
 	int status = XEN_NETIF_RSP_OKAY;
@@ -564,9 +468,9 @@ static int netbk_check_gop(struct xenvif *vif, int nr_meta_slots,
 	return status;
 }
 
-static void netbk_add_frag_responses(struct xenvif *vif, int status,
-				     struct netbk_rx_meta *meta,
-				     int nr_meta_slots)
+static void xenvif_add_frag_responses(struct xenvif *vif, int status,
+				      struct xenvif_rx_meta *meta,
+				      int nr_meta_slots)
 {
 	int i;
 	unsigned long offset;
@@ -594,9 +498,13 @@ struct skb_cb_overlay {
 	int meta_slots_used;
 };
 
-static void xen_netbk_rx_action(struct xen_netbk *netbk)
+static void xenvif_kick_thread(struct xenvif *vif)
+{
+	wake_up(&vif->wq);
+}
+
+void xenvif_rx_action(struct xenvif *vif)
 {
-	struct xenvif *vif = NULL, *tmp;
 	s8 status;
 	u16 flags;
 	struct xen_netif_rx_response *resp;
@@ -608,22 +516,23 @@ static void xen_netbk_rx_action(struct xen_netbk *netbk)
 	int count;
 	unsigned long offset;
 	struct skb_cb_overlay *sco;
+	int need_to_notify = 0;
 
 	struct netrx_pending_operations npo = {
-		.copy  = netbk->grant_copy_op,
-		.meta  = netbk->meta,
+		.copy  = vif->grant_copy_op,
+		.meta  = vif->meta,
 	};
 
 	skb_queue_head_init(&rxq);
 
 	count = 0;
 
-	while ((skb = skb_dequeue(&netbk->rx_queue)) != NULL) {
+	while ((skb = skb_dequeue(&vif->rx_queue)) != NULL) {
 		vif = netdev_priv(skb->dev);
 		nr_frags = skb_shinfo(skb)->nr_frags;
 
 		sco = (struct skb_cb_overlay *)skb->cb;
-		sco->meta_slots_used = netbk_gop_skb(skb, &npo);
+		sco->meta_slots_used = xenvif_gop_skb(skb, &npo);
 
 		count += nr_frags + 1;
 
@@ -635,27 +544,27 @@ static void xen_netbk_rx_action(struct xen_netbk *netbk)
 			break;
 	}
 
-	BUG_ON(npo.meta_prod > ARRAY_SIZE(netbk->meta));
+	BUG_ON(npo.meta_prod > ARRAY_SIZE(vif->meta));
 
 	if (!npo.copy_prod)
 		return;
 
-	BUG_ON(npo.copy_prod > ARRAY_SIZE(netbk->grant_copy_op));
-	gnttab_batch_copy(netbk->grant_copy_op, npo.copy_prod);
+	BUG_ON(npo.copy_prod > ARRAY_SIZE(vif->grant_copy_op));
+	gnttab_batch_copy(vif->grant_copy_op, npo.copy_prod);
 
 	while ((skb = __skb_dequeue(&rxq)) != NULL) {
 		sco = (struct skb_cb_overlay *)skb->cb;
 
 		vif = netdev_priv(skb->dev);
 
-		if (netbk->meta[npo.meta_cons].gso_size && vif->gso_prefix) {
+		if (vif->meta[npo.meta_cons].gso_size && vif->gso_prefix) {
 			resp = RING_GET_RESPONSE(&vif->rx,
-						vif->rx.rsp_prod_pvt++);
+						 vif->rx.rsp_prod_pvt++);
 
 			resp->flags = XEN_NETRXF_gso_prefix | XEN_NETRXF_more_data;
 
-			resp->offset = netbk->meta[npo.meta_cons].gso_size;
-			resp->id = netbk->meta[npo.meta_cons].id;
+			resp->offset = vif->meta[npo.meta_cons].gso_size;
+			resp->id = vif->meta[npo.meta_cons].id;
 			resp->status = sco->meta_slots_used;
 
 			npo.meta_cons++;
@@ -666,7 +575,7 @@ static void xen_netbk_rx_action(struct xen_netbk *netbk)
 		vif->dev->stats.tx_bytes += skb->len;
 		vif->dev->stats.tx_packets++;
 
-		status = netbk_check_gop(vif, sco->meta_slots_used, &npo);
+		status = xenvif_check_gop(vif, sco->meta_slots_used, &npo);
 
 		if (sco->meta_slots_used == 1)
 			flags = 0;
@@ -680,12 +589,12 @@ static void xen_netbk_rx_action(struct xen_netbk *netbk)
 			flags |= XEN_NETRXF_data_validated;
 
 		offset = 0;
-		resp = make_rx_response(vif, netbk->meta[npo.meta_cons].id,
+		resp = make_rx_response(vif, vif->meta[npo.meta_cons].id,
 					status, offset,
-					netbk->meta[npo.meta_cons].size,
+					vif->meta[npo.meta_cons].size,
 					flags);
 
-		if (netbk->meta[npo.meta_cons].gso_size && !vif->gso_prefix) {
+		if (vif->meta[npo.meta_cons].gso_size && !vif->gso_prefix) {
 			struct xen_netif_extra_info *gso =
 				(struct xen_netif_extra_info *)
 				RING_GET_RESPONSE(&vif->rx,
@@ -693,7 +602,7 @@ static void xen_netbk_rx_action(struct xen_netbk *netbk)
 
 			resp->flags |= XEN_NETRXF_extra_info;
 
-			gso->u.gso.size = netbk->meta[npo.meta_cons].gso_size;
+			gso->u.gso.size = vif->meta[npo.meta_cons].gso_size;
 			gso->u.gso.type = XEN_NETIF_GSO_TYPE_TCPV4;
 			gso->u.gso.pad = 0;
 			gso->u.gso.features = 0;
@@ -702,121 +611,43 @@ static void xen_netbk_rx_action(struct xen_netbk *netbk)
 			gso->flags = 0;
 		}
 
-		netbk_add_frag_responses(vif, status,
-					 netbk->meta + npo.meta_cons + 1,
-					 sco->meta_slots_used);
+		xenvif_add_frag_responses(vif, status,
+					  vif->meta + npo.meta_cons + 1,
+					  sco->meta_slots_used);
 
 		RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&vif->rx, ret);
-		if (ret && list_empty(&vif->notify_list))
-			list_add_tail(&vif->notify_list, &notify);
+		if (ret)
+			need_to_notify = 1;
 
 		xenvif_notify_tx_completion(vif);
 
-		xenvif_put(vif);
 		npo.meta_cons += sco->meta_slots_used;
 		dev_kfree_skb(skb);
 	}
 
-	list_for_each_entry_safe(vif, tmp, &notify, notify_list) {
+	if (need_to_notify)
 		notify_remote_via_irq(vif->rx_irq);
-		list_del_init(&vif->notify_list);
-	}
 
 	/* More work to do? */
-	if (!skb_queue_empty(&netbk->rx_queue) &&
-			!timer_pending(&netbk->net_timer))
-		xen_netbk_kick_thread(netbk);
-}
-
-void xen_netbk_queue_tx_skb(struct xenvif *vif, struct sk_buff *skb)
-{
-	struct xen_netbk *netbk = vif->netbk;
-
-	skb_queue_tail(&netbk->rx_queue, skb);
-
-	xen_netbk_kick_thread(netbk);
-}
-
-static void xen_netbk_alarm(unsigned long data)
-{
-	struct xen_netbk *netbk = (struct xen_netbk *)data;
-	xen_netbk_kick_thread(netbk);
-}
-
-static int __on_net_schedule_list(struct xenvif *vif)
-{
-	return !list_empty(&vif->schedule_list);
-}
-
-/* Must be called with net_schedule_list_lock held */
-static void remove_from_net_schedule_list(struct xenvif *vif)
-{
-	if (likely(__on_net_schedule_list(vif))) {
-		list_del_init(&vif->schedule_list);
-		xenvif_put(vif);
-	}
+	if (!skb_queue_empty(&vif->rx_queue))
+		xenvif_kick_thread(vif);
 }
 
-static struct xenvif *poll_net_schedule_list(struct xen_netbk *netbk)
+void xenvif_queue_tx_skb(struct xenvif *vif, struct sk_buff *skb)
 {
-	struct xenvif *vif = NULL;
-
-	spin_lock_irq(&netbk->net_schedule_list_lock);
-	if (list_empty(&netbk->net_schedule_list))
-		goto out;
-
-	vif = list_first_entry(&netbk->net_schedule_list,
-			       struct xenvif, schedule_list);
-	if (!vif)
-		goto out;
-
-	xenvif_get(vif);
+	skb_queue_tail(&vif->rx_queue, skb);
 
-	remove_from_net_schedule_list(vif);
-out:
-	spin_unlock_irq(&netbk->net_schedule_list_lock);
-	return vif;
+	xenvif_kick_thread(vif);
 }
 
-void xen_netbk_schedule_xenvif(struct xenvif *vif)
-{
-	unsigned long flags;
-	struct xen_netbk *netbk = vif->netbk;
-
-	if (__on_net_schedule_list(vif))
-		goto kick;
-
-	spin_lock_irqsave(&netbk->net_schedule_list_lock, flags);
-	if (!__on_net_schedule_list(vif) &&
-	    likely(xenvif_schedulable(vif))) {
-		list_add_tail(&vif->schedule_list, &netbk->net_schedule_list);
-		xenvif_get(vif);
-	}
-	spin_unlock_irqrestore(&netbk->net_schedule_list_lock, flags);
-
-kick:
-	smp_mb();
-	if ((nr_pending_reqs(netbk) < (MAX_PENDING_REQS/2)) &&
-	    !list_empty(&netbk->net_schedule_list))
-		xen_netbk_kick_thread(netbk);
-}
-
-void xen_netbk_deschedule_xenvif(struct xenvif *vif)
-{
-	struct xen_netbk *netbk = vif->netbk;
-	spin_lock_irq(&netbk->net_schedule_list_lock);
-	remove_from_net_schedule_list(vif);
-	spin_unlock_irq(&netbk->net_schedule_list_lock);
-}
-
-void xen_netbk_check_rx_xenvif(struct xenvif *vif)
+void xenvif_check_rx_xenvif(struct xenvif *vif)
 {
 	int more_to_do;
 
 	RING_FINAL_CHECK_FOR_REQUESTS(&vif->tx, more_to_do);
 
 	if (more_to_do)
-		xen_netbk_schedule_xenvif(vif);
+		napi_schedule(&vif->napi);
 }
 
 static void tx_add_credit(struct xenvif *vif)
@@ -843,11 +674,11 @@ static void tx_credit_callback(unsigned long data)
 {
 	struct xenvif *vif = (struct xenvif *)data;
 	tx_add_credit(vif);
-	xen_netbk_check_rx_xenvif(vif);
+	xenvif_check_rx_xenvif(vif);
 }
 
-static void netbk_tx_err(struct xenvif *vif,
-			 struct xen_netif_tx_request *txp, RING_IDX end)
+static void xenvif_tx_err(struct xenvif *vif,
+			  struct xen_netif_tx_request *txp, RING_IDX end)
 {
 	RING_IDX cons = vif->tx.req_cons;
 
@@ -858,21 +689,18 @@ static void netbk_tx_err(struct xenvif *vif,
 		txp = RING_GET_REQUEST(&vif->tx, cons++);
 	} while (1);
 	vif->tx.req_cons = cons;
-	xen_netbk_check_rx_xenvif(vif);
-	xenvif_put(vif);
 }
 
-static void netbk_fatal_tx_err(struct xenvif *vif)
+static void xenvif_fatal_tx_err(struct xenvif *vif)
 {
 	netdev_err(vif->dev, "fatal error; disabling device\n");
 	xenvif_carrier_off(vif);
-	xenvif_put(vif);
 }
 
-static int netbk_count_requests(struct xenvif *vif,
-				struct xen_netif_tx_request *first,
-				struct xen_netif_tx_request *txp,
-				int work_to_do)
+static int xenvif_count_requests(struct xenvif *vif,
+				 struct xen_netif_tx_request *first,
+				 struct xen_netif_tx_request *txp,
+				 int work_to_do)
 {
 	RING_IDX cons = vif->tx.req_cons;
 	int slots = 0;
@@ -889,7 +717,7 @@ static int netbk_count_requests(struct xenvif *vif,
 			netdev_err(vif->dev,
 				   "Asked for %d slots but exceeds this limit\n",
 				   work_to_do);
-			netbk_fatal_tx_err(vif);
+			xenvif_fatal_tx_err(vif);
 			return -ENODATA;
 		}
 
@@ -900,7 +728,7 @@ static int netbk_count_requests(struct xenvif *vif,
 			netdev_err(vif->dev,
 				   "Malicious frontend using %d slots, threshold %u\n",
 				   slots, fatal_skb_slots);
-			netbk_fatal_tx_err(vif);
+			xenvif_fatal_tx_err(vif);
 			return -E2BIG;
 		}
 
@@ -948,7 +776,7 @@ static int netbk_count_requests(struct xenvif *vif,
 		if (unlikely((txp->offset + txp->size) > PAGE_SIZE)) {
 			netdev_err(vif->dev, "Cross page boundary, txp->offset: %x, size: %u\n",
 				 txp->offset, txp->size);
-			netbk_fatal_tx_err(vif);
+			xenvif_fatal_tx_err(vif);
 			return -EINVAL;
 		}
 
@@ -960,29 +788,30 @@ static int netbk_count_requests(struct xenvif *vif,
 	} while (more_data);
 
 	if (drop_err) {
-		netbk_tx_err(vif, first, cons + slots);
+		xenvif_tx_err(vif, first, cons + slots);
 		return drop_err;
 	}
 
 	return slots;
 }
 
-static struct page *xen_netbk_alloc_page(struct xen_netbk *netbk,
-					 u16 pending_idx)
+static struct page *xenvif_alloc_page(struct xenvif *vif,
+				      u16 pending_idx)
 {
 	struct page *page;
+
 	page = alloc_page(GFP_KERNEL|__GFP_COLD);
 	if (!page)
 		return NULL;
-	netbk->mmap_pages[pending_idx] = page;
+	vif->mmap_pages[pending_idx] = page;
+
 	return page;
 }
 
-static struct gnttab_copy *xen_netbk_get_requests(struct xen_netbk *netbk,
-						  struct xenvif *vif,
-						  struct sk_buff *skb,
-						  struct xen_netif_tx_request *txp,
-						  struct gnttab_copy *gop)
+static struct gnttab_copy *xenvif_get_requests(struct xenvif *vif,
+					       struct sk_buff *skb,
+					       struct xen_netif_tx_request *txp,
+					       struct gnttab_copy *gop)
 {
 	struct skb_shared_info *shinfo = skb_shinfo(skb);
 	skb_frag_t *frags = shinfo->frags;
@@ -1005,12 +834,12 @@ static struct gnttab_copy *xen_netbk_get_requests(struct xen_netbk *netbk,
 
 	/* Coalesce tx requests, at this point the packet passed in
 	 * should be <= 64K. Any packets larger than 64K have been
-	 * handled in netbk_count_requests().
+	 * handled in xenvif_count_requests().
 	 */
 	for (shinfo->nr_frags = slot = start; slot < nr_slots;
 	     shinfo->nr_frags++) {
 		struct pending_tx_info *pending_tx_info =
-			netbk->pending_tx_info;
+			vif->pending_tx_info;
 
 		page = alloc_page(GFP_KERNEL|__GFP_COLD);
 		if (!page)
@@ -1047,21 +876,18 @@ static struct gnttab_copy *xen_netbk_get_requests(struct xen_netbk *netbk,
 				gop->len = txp->size;
 				dst_offset += gop->len;
 
-				index = pending_index(netbk->pending_cons++);
+				index = pending_index(vif->pending_cons++);
 
-				pending_idx = netbk->pending_ring[index];
+				pending_idx = vif->pending_ring[index];
 
 				memcpy(&pending_tx_info[pending_idx].req, txp,
 				       sizeof(*txp));
-				xenvif_get(vif);
-
-				pending_tx_info[pending_idx].vif = vif;
 
 				/* Poison these fields, corresponding
 				 * fields for head tx req will be set
 				 * to correct values after the loop.
 				 */
-				netbk->mmap_pages[pending_idx] = (void *)(~0UL);
+				vif->mmap_pages[pending_idx] = (void *)(~0UL);
 				pending_tx_info[pending_idx].head =
 					INVALID_PENDING_RING_IDX;
 
@@ -1081,7 +907,7 @@ static struct gnttab_copy *xen_netbk_get_requests(struct xen_netbk *netbk,
 		first->req.offset = 0;
 		first->req.size = dst_offset;
 		first->head = start_idx;
-		netbk->mmap_pages[head_idx] = page;
+		vif->mmap_pages[head_idx] = page;
 		frag_set_pending_idx(&frags[shinfo->nr_frags], head_idx);
 	}
 
@@ -1091,20 +917,20 @@ static struct gnttab_copy *xen_netbk_get_requests(struct xen_netbk *netbk,
 err:
 	/* Unwind, freeing all pages and sending error responses. */
 	while (shinfo->nr_frags-- > start) {
-		xen_netbk_idx_release(netbk,
+		xenvif_idx_release(vif,
 				frag_get_pending_idx(&frags[shinfo->nr_frags]),
 				XEN_NETIF_RSP_ERROR);
 	}
 	/* The head too, if necessary. */
 	if (start)
-		xen_netbk_idx_release(netbk, pending_idx, XEN_NETIF_RSP_ERROR);
+		xenvif_idx_release(vif, pending_idx, XEN_NETIF_RSP_ERROR);
 
 	return NULL;
 }
 
-static int xen_netbk_tx_check_gop(struct xen_netbk *netbk,
-				  struct sk_buff *skb,
-				  struct gnttab_copy **gopp)
+static int xenvif_tx_check_gop(struct xenvif *vif,
+			       struct sk_buff *skb,
+			       struct gnttab_copy **gopp)
 {
 	struct gnttab_copy *gop = *gopp;
 	u16 pending_idx = *((u16 *)skb->data);
@@ -1117,7 +943,7 @@ static int xen_netbk_tx_check_gop(struct xen_netbk *netbk,
 	/* Check status of header. */
 	err = gop->status;
 	if (unlikely(err))
-		xen_netbk_idx_release(netbk, pending_idx, XEN_NETIF_RSP_ERROR);
+		xenvif_idx_release(vif, pending_idx, XEN_NETIF_RSP_ERROR);
 
 	/* Skip first skb fragment if it is on same page as header fragment. */
 	start = (frag_get_pending_idx(&shinfo->frags[0]) == pending_idx);
@@ -1127,7 +953,7 @@ static int xen_netbk_tx_check_gop(struct xen_netbk *netbk,
 		pending_ring_idx_t head;
 
 		pending_idx = frag_get_pending_idx(&shinfo->frags[i]);
-		tx_info = &netbk->pending_tx_info[pending_idx];
+		tx_info = &vif->pending_tx_info[pending_idx];
 		head = tx_info->head;
 
 		/* Check error status: if okay then remember grant handle. */
@@ -1135,18 +961,18 @@ static int xen_netbk_tx_check_gop(struct xen_netbk *netbk,
 			newerr = (++gop)->status;
 			if (newerr)
 				break;
-			peek = netbk->pending_ring[pending_index(++head)];
-		} while (!pending_tx_is_head(netbk, peek));
+			peek = vif->pending_ring[pending_index(++head)];
+		} while (!pending_tx_is_head(vif, peek));
 
 		if (likely(!newerr)) {
 			/* Had a previous error? Invalidate this fragment. */
 			if (unlikely(err))
-				xen_netbk_idx_release(netbk, pending_idx, XEN_NETIF_RSP_OKAY);
+				xenvif_idx_release(vif, pending_idx, XEN_NETIF_RSP_OKAY);
 			continue;
 		}
 
 		/* Error on this fragment: respond to client with an error. */
-		xen_netbk_idx_release(netbk, pending_idx, XEN_NETIF_RSP_ERROR);
+		xenvif_idx_release(vif, pending_idx, XEN_NETIF_RSP_ERROR);
 
 		/* Not the first error? Preceding frags already invalidated. */
 		if (err)
@@ -1154,10 +980,10 @@ static int xen_netbk_tx_check_gop(struct xen_netbk *netbk,
 
 		/* First error: invalidate header and preceding fragments. */
 		pending_idx = *((u16 *)skb->data);
-		xen_netbk_idx_release(netbk, pending_idx, XEN_NETIF_RSP_OKAY);
+		xenvif_idx_release(vif, pending_idx, XEN_NETIF_RSP_OKAY);
 		for (j = start; j < i; j++) {
 			pending_idx = frag_get_pending_idx(&shinfo->frags[j]);
-			xen_netbk_idx_release(netbk, pending_idx, XEN_NETIF_RSP_OKAY);
+			xenvif_idx_release(vif, pending_idx, XEN_NETIF_RSP_OKAY);
 		}
 
 		/* Remember the error: invalidate all subsequent fragments. */
@@ -1168,7 +994,7 @@ static int xen_netbk_tx_check_gop(struct xen_netbk *netbk,
 	return err;
 }
 
-static void xen_netbk_fill_frags(struct xen_netbk *netbk, struct sk_buff *skb)
+static void xenvif_fill_frags(struct xenvif *vif, struct sk_buff *skb)
 {
 	struct skb_shared_info *shinfo = skb_shinfo(skb);
 	int nr_frags = shinfo->nr_frags;
@@ -1182,20 +1008,20 @@ static void xen_netbk_fill_frags(struct xen_netbk *netbk, struct sk_buff *skb)
 
 		pending_idx = frag_get_pending_idx(frag);
 
-		txp = &netbk->pending_tx_info[pending_idx].req;
-		page = virt_to_page(idx_to_kaddr(netbk, pending_idx));
+		txp = &vif->pending_tx_info[pending_idx].req;
+		page = virt_to_page(idx_to_kaddr(vif, pending_idx));
 		__skb_fill_page_desc(skb, i, page, txp->offset, txp->size);
 		skb->len += txp->size;
 		skb->data_len += txp->size;
 		skb->truesize += txp->size;
 
-		/* Take an extra reference to offset xen_netbk_idx_release */
-		get_page(netbk->mmap_pages[pending_idx]);
-		xen_netbk_idx_release(netbk, pending_idx, XEN_NETIF_RSP_OKAY);
+		/* Take an extra reference to offset xenvif_idx_release */
+		get_page(vif->mmap_pages[pending_idx]);
+		xenvif_idx_release(vif, pending_idx, XEN_NETIF_RSP_OKAY);
 	}
 }
 
-static int xen_netbk_get_extras(struct xenvif *vif,
+static int xenvif_get_extras(struct xenvif *vif,
 				struct xen_netif_extra_info *extras,
 				int work_to_do)
 {
@@ -1205,7 +1031,7 @@ static int xen_netbk_get_extras(struct xenvif *vif,
 	do {
 		if (unlikely(work_to_do-- <= 0)) {
 			netdev_err(vif->dev, "Missing extra info\n");
-			netbk_fatal_tx_err(vif);
+			xenvif_fatal_tx_err(vif);
 			return -EBADR;
 		}
 
@@ -1216,7 +1042,7 @@ static int xen_netbk_get_extras(struct xenvif *vif,
 			vif->tx.req_cons = ++cons;
 			netdev_err(vif->dev,
 				   "Invalid extra type: %d\n", extra.type);
-			netbk_fatal_tx_err(vif);
+			xenvif_fatal_tx_err(vif);
 			return -EINVAL;
 		}
 
@@ -1227,20 +1053,20 @@ static int xen_netbk_get_extras(struct xenvif *vif,
 	return work_to_do;
 }
 
-static int netbk_set_skb_gso(struct xenvif *vif,
-			     struct sk_buff *skb,
-			     struct xen_netif_extra_info *gso)
+static int xenvif_set_skb_gso(struct xenvif *vif,
+			      struct sk_buff *skb,
+			      struct xen_netif_extra_info *gso)
 {
 	if (!gso->u.gso.size) {
 		netdev_err(vif->dev, "GSO size must not be zero.\n");
-		netbk_fatal_tx_err(vif);
+		xenvif_fatal_tx_err(vif);
 		return -EINVAL;
 	}
 
 	/* Currently only TCPv4 S.O. is supported. */
 	if (gso->u.gso.type != XEN_NETIF_GSO_TYPE_TCPV4) {
 		netdev_err(vif->dev, "Bad GSO type %d.\n", gso->u.gso.type);
-		netbk_fatal_tx_err(vif);
+		xenvif_fatal_tx_err(vif);
 		return -EINVAL;
 	}
 
@@ -1351,16 +1177,14 @@ static bool tx_credit_exceeded(struct xenvif *vif, unsigned size)
 	return false;
 }
 
-static unsigned xen_netbk_tx_build_gops(struct xen_netbk *netbk)
+static unsigned xenvif_tx_build_gops(struct xenvif *vif)
 {
-	struct gnttab_copy *gop = netbk->tx_copy_ops, *request_gop;
+	struct gnttab_copy *gop = vif->tx_copy_ops, *request_gop;
 	struct sk_buff *skb;
 	int ret;
 
-	while ((nr_pending_reqs(netbk) + XEN_NETBK_LEGACY_SLOTS_MAX
-		< MAX_PENDING_REQS) &&
-		!list_empty(&netbk->net_schedule_list)) {
-		struct xenvif *vif;
+	while ((nr_pending_reqs(vif) + XEN_NETBK_LEGACY_SLOTS_MAX
+		< MAX_PENDING_REQS)) {
 		struct xen_netif_tx_request txreq;
 		struct xen_netif_tx_request txfrags[XEN_NETBK_LEGACY_SLOTS_MAX];
 		struct page *page;
@@ -1371,16 +1195,6 @@ static unsigned xen_netbk_tx_build_gops(struct xen_netbk *netbk)
 		unsigned int data_len;
 		pending_ring_idx_t index;
 
-		/* Get a netif from the list with work to do. */
-		vif = poll_net_schedule_list(netbk);
-		/* This can sometimes happen because the test of
-		 * list_empty(net_schedule_list) at the top of the
-		 * loop is unlocked.  Just go back and have another
-		 * look.
-		 */
-		if (!vif)
-			continue;
-
 		if (vif->tx.sring->req_prod - vif->tx.req_cons >
 		    XEN_NETIF_TX_RING_SIZE) {
 			netdev_err(vif->dev,
@@ -1388,15 +1202,13 @@ static unsigned xen_netbk_tx_build_gops(struct xen_netbk *netbk)
 				   "req_prod %d, req_cons %d, size %ld\n",
 				   vif->tx.sring->req_prod, vif->tx.req_cons,
 				   XEN_NETIF_TX_RING_SIZE);
-			netbk_fatal_tx_err(vif);
+			xenvif_fatal_tx_err(vif);
 			continue;
 		}
 
 		RING_FINAL_CHECK_FOR_REQUESTS(&vif->tx, work_to_do);
-		if (!work_to_do) {
-			xenvif_put(vif);
-			continue;
-		}
+		if (!work_to_do)
+			break;
 
 		idx = vif->tx.req_cons;
 		rmb(); /* Ensure that we see the request before we copy it. */
@@ -1404,10 +1216,8 @@ static unsigned xen_netbk_tx_build_gops(struct xen_netbk *netbk)
 
 		/* Credit-based scheduling. */
 		if (txreq.size > vif->remaining_credit &&
-		    tx_credit_exceeded(vif, txreq.size)) {
-			xenvif_put(vif);
-			continue;
-		}
+		    tx_credit_exceeded(vif, txreq.size))
+			break;
 
 		vif->remaining_credit -= txreq.size;
 
@@ -1416,24 +1226,24 @@ static unsigned xen_netbk_tx_build_gops(struct xen_netbk *netbk)
 
 		memset(extras, 0, sizeof(extras));
 		if (txreq.flags & XEN_NETTXF_extra_info) {
-			work_to_do = xen_netbk_get_extras(vif, extras,
+			work_to_do = xenvif_get_extras(vif, extras,
 							  work_to_do);
 			idx = vif->tx.req_cons;
 			if (unlikely(work_to_do < 0))
-				continue;
+				break;
 		}
 
-		ret = netbk_count_requests(vif, &txreq, txfrags, work_to_do);
+		ret = xenvif_count_requests(vif, &txreq, txfrags, work_to_do);
 		if (unlikely(ret < 0))
-			continue;
+			break;
 
 		idx += ret;
 
 		if (unlikely(txreq.size < ETH_HLEN)) {
 			netdev_dbg(vif->dev,
 				   "Bad packet size: %d\n", txreq.size);
-			netbk_tx_err(vif, &txreq, idx);
-			continue;
+			xenvif_tx_err(vif, &txreq, idx);
+			break;
 		}
 
 		/* No crossing a page as the payload mustn't fragment. */
@@ -1442,12 +1252,12 @@ static unsigned xen_netbk_tx_build_gops(struct xen_netbk *netbk)
 				   "txreq.offset: %x, size: %u, end: %lu\n",
 				   txreq.offset, txreq.size,
 				   (txreq.offset&~PAGE_MASK) + txreq.size);
-			netbk_fatal_tx_err(vif);
-			continue;
+			xenvif_fatal_tx_err(vif);
+			break;
 		}
 
-		index = pending_index(netbk->pending_cons);
-		pending_idx = netbk->pending_ring[index];
+		index = pending_index(vif->pending_cons);
+		pending_idx = vif->pending_ring[index];
 
 		data_len = (txreq.size > PKT_PROT_LEN &&
 			    ret < XEN_NETBK_LEGACY_SLOTS_MAX) ?
@@ -1458,7 +1268,7 @@ static unsigned xen_netbk_tx_build_gops(struct xen_netbk *netbk)
 		if (unlikely(skb == NULL)) {
 			netdev_dbg(vif->dev,
 				   "Can't allocate a skb in start_xmit.\n");
-			netbk_tx_err(vif, &txreq, idx);
+			xenvif_tx_err(vif, &txreq, idx);
 			break;
 		}
 
@@ -1469,19 +1279,20 @@ static unsigned xen_netbk_tx_build_gops(struct xen_netbk *netbk)
 			struct xen_netif_extra_info *gso;
 			gso = &extras[XEN_NETIF_EXTRA_TYPE_GSO - 1];
 
-			if (netbk_set_skb_gso(vif, skb, gso)) {
-				/* Failure in netbk_set_skb_gso is fatal. */
+			if (xenvif_set_skb_gso(vif, skb, gso)) {
+				/* Failure in xenvif_set_skb_gso is fatal. */
 				kfree_skb(skb);
-				continue;
+				/* XXX ???? break or continue ?*/
+				break;
 			}
 		}
 
 		/* XXX could copy straight to head */
-		page = xen_netbk_alloc_page(netbk, pending_idx);
+		page = xenvif_alloc_page(vif, pending_idx);
 		if (!page) {
 			kfree_skb(skb);
-			netbk_tx_err(vif, &txreq, idx);
-			continue;
+			xenvif_tx_err(vif, &txreq, idx);
+			break;
 		}
 
 		gop->source.u.ref = txreq.gref;
@@ -1497,10 +1308,9 @@ static unsigned xen_netbk_tx_build_gops(struct xen_netbk *netbk)
 
 		gop++;
 
-		memcpy(&netbk->pending_tx_info[pending_idx].req,
+		memcpy(&vif->pending_tx_info[pending_idx].req,
 		       &txreq, sizeof(txreq));
-		netbk->pending_tx_info[pending_idx].vif = vif;
-		netbk->pending_tx_info[pending_idx].head = index;
+		vif->pending_tx_info[pending_idx].head = index;
 		*((u16 *)skb->data) = pending_idx;
 
 		__skb_put(skb, data_len);
@@ -1515,46 +1325,45 @@ static unsigned xen_netbk_tx_build_gops(struct xen_netbk *netbk)
 					     INVALID_PENDING_IDX);
 		}
 
-		netbk->pending_cons++;
+		vif->pending_cons++;
 
-		request_gop = xen_netbk_get_requests(netbk, vif,
-						     skb, txfrags, gop);
+		request_gop = xenvif_get_requests(vif, skb, txfrags, gop);
 		if (request_gop == NULL) {
 			kfree_skb(skb);
-			netbk_tx_err(vif, &txreq, idx);
-			continue;
+			xenvif_tx_err(vif, &txreq, idx);
+			break;
 		}
 		gop = request_gop;
 
-		__skb_queue_tail(&netbk->tx_queue, skb);
+		__skb_queue_tail(&vif->tx_queue, skb);
 
 		vif->tx.req_cons = idx;
-		xen_netbk_check_rx_xenvif(vif);
 
-		if ((gop-netbk->tx_copy_ops) >= ARRAY_SIZE(netbk->tx_copy_ops))
+		if ((gop-vif->tx_copy_ops) >= ARRAY_SIZE(vif->tx_copy_ops))
 			break;
 	}
 
-	return gop - netbk->tx_copy_ops;
+	return gop - vif->tx_copy_ops;
 }
 
-static void xen_netbk_tx_submit(struct xen_netbk *netbk)
+
+static int xenvif_tx_submit(struct xenvif *vif, int budget)
 {
-	struct gnttab_copy *gop = netbk->tx_copy_ops;
+	struct gnttab_copy *gop = vif->tx_copy_ops;
 	struct sk_buff *skb;
+	int work_done = 0;
 
-	while ((skb = __skb_dequeue(&netbk->tx_queue)) != NULL) {
+	while (work_done < budget &&
+	       (skb = __skb_dequeue(&vif->tx_queue)) != NULL) {
 		struct xen_netif_tx_request *txp;
-		struct xenvif *vif;
 		u16 pending_idx;
 		unsigned data_len;
 
 		pending_idx = *((u16 *)skb->data);
-		vif = netbk->pending_tx_info[pending_idx].vif;
-		txp = &netbk->pending_tx_info[pending_idx].req;
+		txp = &vif->pending_tx_info[pending_idx].req;
 
 		/* Check the remap error code. */
-		if (unlikely(xen_netbk_tx_check_gop(netbk, skb, &gop))) {
+		if (unlikely(xenvif_tx_check_gop(vif, skb, &gop))) {
 			netdev_dbg(vif->dev, "netback grant failed.\n");
 			skb_shinfo(skb)->nr_frags = 0;
 			kfree_skb(skb);
@@ -1563,7 +1372,7 @@ static void xen_netbk_tx_submit(struct xen_netbk *netbk)
 
 		data_len = skb->len;
 		memcpy(skb->data,
-		       (void *)(idx_to_kaddr(netbk, pending_idx)|txp->offset),
+		       (void *)(idx_to_kaddr(vif, pending_idx)|txp->offset),
 		       data_len);
 		if (data_len < txp->size) {
 			/* Append the packet payload as a fragment. */
@@ -1571,7 +1380,7 @@ static void xen_netbk_tx_submit(struct xen_netbk *netbk)
 			txp->size -= data_len;
 		} else {
 			/* Schedule a response immediately. */
-			xen_netbk_idx_release(netbk, pending_idx, XEN_NETIF_RSP_OKAY);
+			xenvif_idx_release(vif, pending_idx, XEN_NETIF_RSP_OKAY);
 		}
 
 		if (txp->flags & XEN_NETTXF_csum_blank)
@@ -1579,7 +1388,7 @@ static void xen_netbk_tx_submit(struct xen_netbk *netbk)
 		else if (txp->flags & XEN_NETTXF_data_validated)
 			skb->ip_summed = CHECKSUM_UNNECESSARY;
 
-		xen_netbk_fill_frags(netbk, skb);
+		xenvif_fill_frags(vif, skb);
 
 		/*
 		 * If the initial fragment was < PKT_PROT_LEN then
@@ -1607,53 +1416,61 @@ static void xen_netbk_tx_submit(struct xen_netbk *netbk)
 		vif->dev->stats.rx_bytes += skb->len;
 		vif->dev->stats.rx_packets++;
 
-		xenvif_receive_skb(vif, skb);
+		work_done++;
+
+		netif_receive_skb(skb);
 	}
+
+	return work_done;
 }
 
 /* Called after netfront has transmitted */
-static void xen_netbk_tx_action(struct xen_netbk *netbk)
+int xenvif_tx_action(struct xenvif *vif, int budget)
 {
 	unsigned nr_gops;
+	int work_done;
+
+	if (unlikely(!tx_work_todo(vif)))
+		return 0;
 
-	nr_gops = xen_netbk_tx_build_gops(netbk);
+	nr_gops = xenvif_tx_build_gops(vif);
 
 	if (nr_gops == 0)
-		return;
+		return 0;
+
+	gnttab_batch_copy(vif->tx_copy_ops, nr_gops);
 
-	gnttab_batch_copy(netbk->tx_copy_ops, nr_gops);
+	work_done = xenvif_tx_submit(vif, nr_gops);
 
-	xen_netbk_tx_submit(netbk);
+	return work_done;
 }
 
-static void xen_netbk_idx_release(struct xen_netbk *netbk, u16 pending_idx,
+static void xenvif_idx_release(struct xenvif *vif, u16 pending_idx,
 				  u8 status)
 {
-	struct xenvif *vif;
 	struct pending_tx_info *pending_tx_info;
 	pending_ring_idx_t head;
 	u16 peek; /* peek into next tx request */
 
-	BUG_ON(netbk->mmap_pages[pending_idx] == (void *)(~0UL));
+	BUG_ON(vif->mmap_pages[pending_idx] == (void *)(~0UL));
 
 	/* Already complete? */
-	if (netbk->mmap_pages[pending_idx] == NULL)
+	if (vif->mmap_pages[pending_idx] == NULL)
 		return;
 
-	pending_tx_info = &netbk->pending_tx_info[pending_idx];
+	pending_tx_info = &vif->pending_tx_info[pending_idx];
 
-	vif = pending_tx_info->vif;
 	head = pending_tx_info->head;
 
-	BUG_ON(!pending_tx_is_head(netbk, head));
-	BUG_ON(netbk->pending_ring[pending_index(head)] != pending_idx);
+	BUG_ON(!pending_tx_is_head(vif, head));
+	BUG_ON(vif->pending_ring[pending_index(head)] != pending_idx);
 
 	do {
 		pending_ring_idx_t index;
 		pending_ring_idx_t idx = pending_index(head);
-		u16 info_idx = netbk->pending_ring[idx];
+		u16 info_idx = vif->pending_ring[idx];
 
-		pending_tx_info = &netbk->pending_tx_info[info_idx];
+		pending_tx_info = &vif->pending_tx_info[info_idx];
 		make_tx_response(vif, &pending_tx_info->req, status);
 
 		/* Setting any number other than
@@ -1662,18 +1479,15 @@ static void xen_netbk_idx_release(struct xen_netbk *netbk, u16 pending_idx,
 		 */
 		pending_tx_info->head = 0;
 
-		index = pending_index(netbk->pending_prod++);
-		netbk->pending_ring[index] = netbk->pending_ring[info_idx];
+		index = pending_index(vif->pending_prod++);
+		vif->pending_ring[index] = vif->pending_ring[info_idx];
 
-		xenvif_put(vif);
+		peek = vif->pending_ring[pending_index(++head)];
 
-		peek = netbk->pending_ring[pending_index(++head)];
+	} while (!pending_tx_is_head(vif, peek));
 
-	} while (!pending_tx_is_head(netbk, peek));
-
-	netbk->mmap_pages[pending_idx]->mapping = 0;
-	put_page(netbk->mmap_pages[pending_idx]);
-	netbk->mmap_pages[pending_idx] = NULL;
+	put_page(vif->mmap_pages[pending_idx]);
+	vif->mmap_pages[pending_idx] = NULL;
 }
 
 
@@ -1721,46 +1535,23 @@ static struct xen_netif_rx_response *make_rx_response(struct xenvif *vif,
 	return resp;
 }
 
-static inline int rx_work_todo(struct xen_netbk *netbk)
+static inline int rx_work_todo(struct xenvif *vif)
 {
-	return !skb_queue_empty(&netbk->rx_queue);
+	return !skb_queue_empty(&vif->rx_queue);
 }
 
-static inline int tx_work_todo(struct xen_netbk *netbk)
+static inline int tx_work_todo(struct xenvif *vif)
 {
 
-	if ((nr_pending_reqs(netbk) + XEN_NETBK_LEGACY_SLOTS_MAX
-	     < MAX_PENDING_REQS) &&
-	     !list_empty(&netbk->net_schedule_list))
+	if (likely(RING_HAS_UNCONSUMED_REQUESTS(&vif->tx)) &&
+	    (nr_pending_reqs(vif) + XEN_NETBK_LEGACY_SLOTS_MAX
+	     < MAX_PENDING_REQS))
 		return 1;
 
 	return 0;
 }
 
-static int xen_netbk_kthread(void *data)
-{
-	struct xen_netbk *netbk = data;
-	while (!kthread_should_stop()) {
-		wait_event_interruptible(netbk->wq,
-				rx_work_todo(netbk) ||
-				tx_work_todo(netbk) ||
-				kthread_should_stop());
-		cond_resched();
-
-		if (kthread_should_stop())
-			break;
-
-		if (rx_work_todo(netbk))
-			xen_netbk_rx_action(netbk);
-
-		if (tx_work_todo(netbk))
-			xen_netbk_tx_action(netbk);
-	}
-
-	return 0;
-}
-
-void xen_netbk_unmap_frontend_rings(struct xenvif *vif)
+void xenvif_unmap_frontend_rings(struct xenvif *vif)
 {
 	if (vif->tx.sring)
 		xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(vif),
@@ -1770,9 +1561,9 @@ void xen_netbk_unmap_frontend_rings(struct xenvif *vif)
 					vif->rx.sring);
 }
 
-int xen_netbk_map_frontend_rings(struct xenvif *vif,
-				 grant_ref_t tx_ring_ref,
-				 grant_ref_t rx_ring_ref)
+int xenvif_map_frontend_rings(struct xenvif *vif,
+			      grant_ref_t tx_ring_ref,
+			      grant_ref_t rx_ring_ref)
 {
 	void *addr;
 	struct xen_netif_tx_sring *txs;
@@ -1801,15 +1592,33 @@ int xen_netbk_map_frontend_rings(struct xenvif *vif,
 	return 0;
 
 err:
-	xen_netbk_unmap_frontend_rings(vif);
+	xenvif_unmap_frontend_rings(vif);
 	return err;
 }
 
+int xenvif_kthread(void *data)
+{
+	struct xenvif *vif = data;
+
+	while (!kthread_should_stop()) {
+		wait_event_interruptible(vif->wq,
+					 rx_work_todo(vif) ||
+					 kthread_should_stop());
+		cond_resched();
+
+		if (kthread_should_stop())
+			break;
+
+		if (rx_work_todo(vif))
+			xenvif_rx_action(vif);
+	}
+
+	return 0;
+}
+
 static int __init netback_init(void)
 {
-	int i;
 	int rc = 0;
-	int group;
 
 	if (!xen_domain())
 		return -ENODEV;
@@ -1821,48 +1630,6 @@ static int __init netback_init(void)
 		fatal_skb_slots = XEN_NETBK_LEGACY_SLOTS_MAX;
 	}
 
-	xen_netbk_group_nr = num_online_cpus();
-	xen_netbk = vzalloc(sizeof(struct xen_netbk) * xen_netbk_group_nr);
-	if (!xen_netbk)
-		return -ENOMEM;
-
-	for (group = 0; group < xen_netbk_group_nr; group++) {
-		struct xen_netbk *netbk = &xen_netbk[group];
-		skb_queue_head_init(&netbk->rx_queue);
-		skb_queue_head_init(&netbk->tx_queue);
-
-		init_timer(&netbk->net_timer);
-		netbk->net_timer.data = (unsigned long)netbk;
-		netbk->net_timer.function = xen_netbk_alarm;
-
-		netbk->pending_cons = 0;
-		netbk->pending_prod = MAX_PENDING_REQS;
-		for (i = 0; i < MAX_PENDING_REQS; i++)
-			netbk->pending_ring[i] = i;
-
-		init_waitqueue_head(&netbk->wq);
-		netbk->task = kthread_create(xen_netbk_kthread,
-					     (void *)netbk,
-					     "netback/%u", group);
-
-		if (IS_ERR(netbk->task)) {
-			printk(KERN_ALERT "kthread_create() fails at netback\n");
-			del_timer(&netbk->net_timer);
-			rc = PTR_ERR(netbk->task);
-			goto failed_init;
-		}
-
-		kthread_bind(netbk->task, group);
-
-		INIT_LIST_HEAD(&netbk->net_schedule_list);
-
-		spin_lock_init(&netbk->net_schedule_list_lock);
-
-		atomic_set(&netbk->netfront_count, 0);
-
-		wake_up_process(netbk->task);
-	}
-
 	rc = xenvif_xenbus_init();
 	if (rc)
 		goto failed_init;
@@ -1870,35 +1637,14 @@ static int __init netback_init(void)
 	return 0;
 
 failed_init:
-	while (--group >= 0) {
-		struct xen_netbk *netbk = &xen_netbk[group];
-		del_timer(&netbk->net_timer);
-		kthread_stop(netbk->task);
-	}
-	vfree(xen_netbk);
 	return rc;
-
 }
 
 module_init(netback_init);
 
 static void __exit netback_fini(void)
 {
-	int i, j;
-
 	xenvif_xenbus_fini();
-
-	for (i = 0; i < xen_netbk_group_nr; i++) {
-		struct xen_netbk *netbk = &xen_netbk[i];
-		del_timer_sync(&netbk->net_timer);
-		kthread_stop(netbk->task);
-		for (j = 0; j < MAX_PENDING_REQS; j++) {
-			if (netbk->mmap_pages[i])
-				__free_page(netbk->mmap_pages[i]);
-		}
-	}
-
-	vfree(xen_netbk);
 }
 module_exit(netback_fini);
 
-- 
1.7.10.4

Re: [PATCH 2/2] xen-netback: switch to NAPI + kthread 1:1 model

[Ian Campbell]

On Wed, 2013-05-29 at 12:43 +0100, Wei Liu wrote:
> This patch implements 1:1 model netback. NAPI and kthread are utilized
> to do the weight-lifting job:
> 
>  - NAPI is used for guest side TX (host side RX)
>  - kthread is used for guest side RX (host side TX)
> 
> Xenvif and xen_netbk are made into one structure to reduce code size.
> 
> This model provides better scheduling fairness among vifs. It is also
> prerequisite for implementing multiqueue for Xen netback.

Any interesting numbers from this change?

> Signed-off-by: Wei Liu <wei.liu2@citrix.com>
> ---
>  drivers/net/xen-netback/common.h    |  105 +++--
>  drivers/net/xen-netback/interface.c |  122 +++---
>  drivers/net/xen-netback/netback.c   |  742 ++++++++++++-----------------------
>  3 files changed, 389 insertions(+), 580 deletions(-)
> 
> diff --git a/drivers/net/xen-netback/common.h b/drivers/net/xen-netback/common.h
> index 8a4d77e..2e270cf 100644
> --- a/drivers/net/xen-netback/common.h
> +++ b/drivers/net/xen-netback/common.h
> @@ -45,15 +45,43 @@
>  #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)
> +
> +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

So this sounds like it is only valid for the pending tx corresponding to
the first slot in a frame, right?

It does not contain a link from any arbitrary slot to the slot
containing the head of that frame, right?

So for the head it is effectively that entries index into the pending_tx
array?

> +				  */
> +};
> +
> +#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;
> +	/* Use kthread for guest RX */
> +	struct task_struct *task;
> +	wait_queue_head_t wq;

This isn't new with this patch but I wonder if we shouldn't group tx and
rx related things into to separate regions of the struct, rather than
interleaving them like this. Especially once we start to separate RX and
TX processing onto different processes the current layout is going to
have pretty horrible cache line bouncing properties.

>  	u8               fe_dev_addr[6];
>  
> @@ -64,9 +92,6 @@ struct xenvif {
>  	char tx_irq_name[IFNAMSIZ+4]; /* DEVNAME-tx */
>  	char rx_irq_name[IFNAMSIZ+4]; /* DEVNAME-rx */
>  
> -	/* List of frontends to notify after a batch of frames sent. */
> -	struct list_head notify_list;
> -
>  	/* The shared rings and indexes. */
>  	struct xen_netif_tx_back_ring tx;
>  	struct xen_netif_rx_back_ring rx;
> @@ -96,12 +121,33 @@ struct xenvif {
>  	/* Statistics */
>  	unsigned long rx_gso_checksum_fixup;
>  
> +	struct sk_buff_head rx_queue;
> +	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];
> +
> +	/*
> +	 * 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];
> +
>  	/* 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 +155,6 @@ 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 +164,26 @@ 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);

>From a review PoV it would have been good to do the bulk rename either
before or after the functional changes. I just skipped over anything
which looked to be just renaming at first glance. Hopefully I didn't
miss anything ciritical!

> +static int xenvif_poll(struct napi_struct *napi, int budget)
> +{
> +	struct xenvif *vif = container_of(napi, struct xenvif, napi);
> +	int work_done;
> +
> +	work_done = xenvif_tx_action(vif, budget);
> +
> +	if (work_done < budget) {
> +		int more_to_do = 0;
> +		unsigned long flags;
> +
> +		local_irq_save(flags);

What does this protect against? Maybe it's a napi requirement not an Xen
netif one?

> +
> +		RING_FINAL_CHECK_FOR_REQUESTS(&vif->tx, more_to_do);
> +		if (!more_to_do || work_done < 0)

"work_done < 0" is an error condition from xenvif_tx_action? But
following the chain down through xenvif_tx_action to xenvif_tx_submit it
doesn't look like such a thing is possible?

Perhaps just a BUG_ON(work_done < 0) right after the call to
xenvif_tx_action()?


> @@ -272,11 +275,13 @@ 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");
> +		printk(KERN_WARNING "xen-netback: Could not allocate netdev for vif%d.%d\n",
> +		       domid, handle);

pr_warn can take a fmt string and do the right thing, can't it? Or was
there a different reason for this change?

Rather than repeating vif%d.%d why not reuse the result of the previous
snprintf -- especially since the fmt codes differ (although probably not
in a way which affects the result).

> @@ -377,7 +389,16 @@ 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%d.%d", vif->domid, vif->handle);

%u.%u, or better find a way to reuse the name allocated in xenvif_alloc,
which I suppose you can get via vif->dev->name?

> +	if (IS_ERR(vif->task)) {
> +		printk(KERN_WARNING "xen-netback: Could not allocate kthread for vif%d.%d\n",
> +		       vif->domid, vif->handle);

pr_warn if possible please. and %u vs %d again.

> @@ -434,9 +338,9 @@ static void netbk_gop_frag_copy(struct xenvif *vif, struct sk_buff *skb,
>  
>  		copy_gop = npo->copy + npo->copy_prod++;
>  		copy_gop->flags = GNTCOPY_dest_gref;
> +
>  		copy_gop->source.domid = DOMID_SELF;
>  		copy_gop->source.u.gmfn = virt_to_mfn(page_address(page));
> -

The blank lines in this function seem pretty arbitrary both before and
after this change ;-)

>  		copy_gop->source.offset = offset;
>  		copy_gop->dest.domid = vif->domid;
>  
> @@ -1469,19 +1279,20 @@ static unsigned xen_netbk_tx_build_gops(struct xen_netbk *netbk)
>  			struct xen_netif_extra_info *gso;
>  			gso = &extras[XEN_NETIF_EXTRA_TYPE_GSO - 1];
>  
> -			if (netbk_set_skb_gso(vif, skb, gso)) {
> -				/* Failure in netbk_set_skb_gso is fatal. */
> +			if (xenvif_set_skb_gso(vif, skb, gso)) {
> +				/* Failure in xenvif_set_skb_gso is fatal. */
>  				kfree_skb(skb);
> -				continue;
> +				/* XXX ???? break or continue ?*/

I'd been wondering that a bunch in the previous hunks too ;-)

break seems fine here.
>  
> +int xenvif_kthread(void *data)
> +{
> +	struct xenvif *vif = data;
> +
> +	while (!kthread_should_stop()) {
> +		wait_event_interruptible(vif->wq,
> +					 rx_work_todo(vif) ||
> +					 kthread_should_stop());
> +		cond_resched();
> +
> +		if (kthread_should_stop())
> +			break;

Should this come before the cond resched? Seems odd to potentially wait
a bit before just exiting.

Ian.

Re: [PATCH 2/2] xen-netback: switch to NAPI + kthread 1:1 model

[Wei Liu]

On Fri, Aug 02, 2013 at 01:20:52PM +0100, Ian Campbell wrote:
> On Wed, 2013-05-29 at 12:43 +0100, Wei Liu wrote:
> > This patch implements 1:1 model netback. NAPI and kthread are utilized
> > to do the weight-lifting job:
> > 
> >  - NAPI is used for guest side TX (host side RX)
> >  - kthread is used for guest side RX (host side TX)
> > 
> > Xenvif and xen_netbk are made into one structure to reduce code size.
> > 
> > This model provides better scheduling fairness among vifs. It is also
> > prerequisite for implementing multiqueue for Xen netback.
> 
> Any interesting numbers from this change?
> 

[extracted from 0/2 of this series]

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

> > Signed-off-by: Wei Liu <wei.liu2@citrix.com>
> > ---
> >  drivers/net/xen-netback/common.h    |  105 +++--
> >  drivers/net/xen-netback/interface.c |  122 +++---
> >  drivers/net/xen-netback/netback.c   |  742 ++++++++++++-----------------------
> >  3 files changed, 389 insertions(+), 580 deletions(-)
> > 
> > diff --git a/drivers/net/xen-netback/common.h b/drivers/net/xen-netback/common.h
> > index 8a4d77e..2e270cf 100644
> > --- a/drivers/net/xen-netback/common.h
> > +++ b/drivers/net/xen-netback/common.h
> > @@ -45,15 +45,43 @@
> >  #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)
> > +
> > +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
> 

This is only code movement, no funcational change.

> So this sounds like it is only valid for the pending tx corresponding to
> the first slot in a frame, right?
> 

No. It can also be interpreted as continuation of several merged tx
requests.

> It does not contain a link from any arbitrary slot to the slot
> containing the head of that frame, right?
> 
> So for the head it is effectively that entries index into the pending_tx
> array?
> 

No. It's a flag to tell if this pending_tx_info is the head of one or
more merged tx requests.

> > +				  */
> > +};
> > +
> > +#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;
> > +	/* Use kthread for guest RX */
> > +	struct task_struct *task;
> > +	wait_queue_head_t wq;
> 
> This isn't new with this patch but I wonder if we shouldn't group tx and
> rx related things into to separate regions of the struct, rather than
> interleaving them like this. Especially once we start to separate RX and
> TX processing onto different processes the current layout is going to
> have pretty horrible cache line bouncing properties.
> 

OK, will group them separately.

> >  	u8               fe_dev_addr[6];
> >  
> > @@ -64,9 +92,6 @@ struct xenvif {
> >  	char tx_irq_name[IFNAMSIZ+4]; /* DEVNAME-tx */
> >  	char rx_irq_name[IFNAMSIZ+4]; /* DEVNAME-rx */
> >  
> > -	/* List of frontends to notify after a batch of frames sent. */
> > -	struct list_head notify_list;
> > -
> >  	/* The shared rings and indexes. */
> >  	struct xen_netif_tx_back_ring tx;
> >  	struct xen_netif_rx_back_ring rx;
> > @@ -96,12 +121,33 @@ struct xenvif {
> >  	/* Statistics */
> >  	unsigned long rx_gso_checksum_fixup;
> >  
> > +	struct sk_buff_head rx_queue;
> > +	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];
> > +
> > +	/*
> > +	 * 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];
> > +
> >  	/* 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 +155,6 @@ 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 +164,26 @@ 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);
> 
> >From a review PoV it would have been good to do the bulk rename either
> before or after the functional changes. I just skipped over anything
> which looked to be just renaming at first glance. Hopefully I didn't
> miss anything ciritical!
> 

No problem, will separate rename and functional changes to two patches.

> > +static int xenvif_poll(struct napi_struct *napi, int budget)
> > +{
> > +	struct xenvif *vif = container_of(napi, struct xenvif, napi);
> > +	int work_done;
> > +
> > +	work_done = xenvif_tx_action(vif, budget);
> > +
> > +	if (work_done < budget) {
> > +		int more_to_do = 0;
> > +		unsigned long flags;
> > +
> > +		local_irq_save(flags);
> 
> What does this protect against? Maybe it's a napi requirement not an Xen
> netif one?
> 

This is sort of a NAPI requirement. If we don't do local_irq_save we
would lose event.

> > +
> > +		RING_FINAL_CHECK_FOR_REQUESTS(&vif->tx, more_to_do);
> > +		if (!more_to_do || work_done < 0)
> 
> "work_done < 0" is an error condition from xenvif_tx_action? But
> following the chain down through xenvif_tx_action to xenvif_tx_submit it
> doesn't look like such a thing is possible?
> 

Right, this is not necessary.

> Perhaps just a BUG_ON(work_done < 0) right after the call to
> xenvif_tx_action()?
> 
> 
> > @@ -272,11 +275,13 @@ 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");
> > +		printk(KERN_WARNING "xen-netback: Could not allocate netdev for vif%d.%d\n",
> > +		       domid, handle);
> 
> pr_warn can take a fmt string and do the right thing, can't it? Or was
> there a different reason for this change?
> 
> Rather than repeating vif%d.%d why not reuse the result of the previous
> snprintf -- especially since the fmt codes differ (although probably not
> in a way which affects the result).
> 

At that time I was trying to unify all output to use printk. But it
turned out that we went the other way around -- several weeks ago we
changed to pr_*.

This change is not necessary any more.

> > @@ -377,7 +389,16 @@ 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%d.%d", vif->domid, vif->handle);
> 
> %u.%u, or better find a way to reuse the name allocated in xenvif_alloc,
> which I suppose you can get via vif->dev->name?
> 

Yes.

> > +	if (IS_ERR(vif->task)) {
> > +		printk(KERN_WARNING "xen-netback: Could not allocate kthread for vif%d.%d\n",
> > +		       vif->domid, vif->handle);
> 
> pr_warn if possible please. and %u vs %d again.
> 

OK.

> > @@ -434,9 +338,9 @@ static void netbk_gop_frag_copy(struct xenvif *vif, struct sk_buff *skb,
> >  
> >  		copy_gop = npo->copy + npo->copy_prod++;
> >  		copy_gop->flags = GNTCOPY_dest_gref;
> > +
> >  		copy_gop->source.domid = DOMID_SELF;
> >  		copy_gop->source.u.gmfn = virt_to_mfn(page_address(page));
> > -
> 
> The blank lines in this function seem pretty arbitrary both before and
> after this change ;-)
> 
> >  		copy_gop->source.offset = offset;
> >  		copy_gop->dest.domid = vif->domid;
> >  
> > @@ -1469,19 +1279,20 @@ static unsigned xen_netbk_tx_build_gops(struct xen_netbk *netbk)
> >  			struct xen_netif_extra_info *gso;
> >  			gso = &extras[XEN_NETIF_EXTRA_TYPE_GSO - 1];
> >  
> > -			if (netbk_set_skb_gso(vif, skb, gso)) {
> > -				/* Failure in netbk_set_skb_gso is fatal. */
> > +			if (xenvif_set_skb_gso(vif, skb, gso)) {
> > +				/* Failure in xenvif_set_skb_gso is fatal. */
> >  				kfree_skb(skb);
> > -				continue;
> > +				/* XXX ???? break or continue ?*/
> 
> I'd been wondering that a bunch in the previous hunks too ;-)
> 
> break seems fine here.
> >  
> > +int xenvif_kthread(void *data)
> > +{
> > +	struct xenvif *vif = data;
> > +
> > +	while (!kthread_should_stop()) {
> > +		wait_event_interruptible(vif->wq,
> > +					 rx_work_todo(vif) ||
> > +					 kthread_should_stop());
> > +		cond_resched();
> > +
> > +		if (kthread_should_stop())
> > +			break;
> 
> Should this come before the cond resched? Seems odd to potentially wait
> a bit before just exiting.
> 

I didn't think about this. This is just plain copy of existing code. I
will swap those two hunks.


Wei.

> Ian.

Re: [PATCH 2/2] xen-netback: switch to NAPI + kthread 1:1 model

[Ian Campbell]

On Mon, 2013-08-05 at 03:31 +0100, Wei Liu wrote:
> On Fri, Aug 02, 2013 at 01:20:52PM +0100, Ian Campbell wrote:
> > On Wed, 2013-05-29 at 12:43 +0100, Wei Liu wrote:
> > > This patch implements 1:1 model netback. NAPI and kthread are utilized
> > > to do the weight-lifting job:
> > > 
> > >  - NAPI is used for guest side TX (host side RX)
> > >  - kthread is used for guest side RX (host side TX)
> > > 
> > > Xenvif and xen_netbk are made into one structure to reduce code size.
> > > 
> > > This model provides better scheduling fairness among vifs. It is also
> > > prerequisite for implementing multiqueue for Xen netback.
> > 
> > Any interesting numbers from this change?
> > 
> 
> [extracted from 0/2 of this series]

Oops, missed that, sorry.

> 
> 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

The CPU load seems to drop pretty significantly with the new model,
nice. I suppose that is the impact of NAPI rather than the 1:1 side of
things.

> 
> > > Signed-off-by: Wei Liu <wei.liu2@citrix.com>
> > > ---
> > >  drivers/net/xen-netback/common.h    |  105 +++--
> > >  drivers/net/xen-netback/interface.c |  122 +++---
> > >  drivers/net/xen-netback/netback.c   |  742 ++++++++++++-----------------------
> > >  3 files changed, 389 insertions(+), 580 deletions(-)
> > > 
> > > diff --git a/drivers/net/xen-netback/common.h b/drivers/net/xen-netback/common.h
> > > index 8a4d77e..2e270cf 100644
> > > --- a/drivers/net/xen-netback/common.h
> > > +++ b/drivers/net/xen-netback/common.h
> > > @@ -45,15 +45,43 @@
> > >  #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)
> > > +
> > > +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
> > 
> 
> This is only code movement, no funcational change.
> 
> > So this sounds like it is only valid for the pending tx corresponding to
> > the first slot in a frame, right?
> > 
> 
> No. It can also be interpreted as continuation of several merged tx
> requests.
> 
> > It does not contain a link from any arbitrary slot to the slot
> > containing the head of that frame, right?
> > 
> > So for the head it is effectively that entries index into the pending_tx
> > array?
> > 
> 
> No. It's a flag to tell if this pending_tx_info is the head of one or
> more merged tx requests.

OK, then I'm totally confused.

I think a comment with an example of the representation of a single
multislot frame (2-3 slots if that suffices as an illustration) in a
comment would be useful.

> > > +static int xenvif_poll(struct napi_struct *napi, int budget)
> > > +{
> > > +	struct xenvif *vif = container_of(napi, struct xenvif, napi);
> > > +	int work_done;
> > > +
> > > +	work_done = xenvif_tx_action(vif, budget);
> > > +
> > > +	if (work_done < budget) {
> > > +		int more_to_do = 0;
> > > +		unsigned long flags;
> > > +
> > > +		local_irq_save(flags);
> > 
> > What does this protect against? Maybe it's a napi requirement not an Xen
> > netif one?
> > 
> 
> This is sort of a NAPI requirement. If we don't do local_irq_save we
> would lose event.

How come? This is a race against this code vs. a new event coming in and
calling napi_schedule? Do real drivers all disable interrupts in this
way too or do they get this for free via some other means?

I think a comment is at least required.

Ian.

Re: [PATCH 2/2] xen-netback: switch to NAPI + kthread 1:1 model

[Wei Liu]

On Tue, Aug 06, 2013 at 10:26:24AM +0100, Ian Campbell wrote:
> On Mon, 2013-08-05 at 03:31 +0100, Wei Liu wrote:
> > On Fri, Aug 02, 2013 at 01:20:52PM +0100, Ian Campbell wrote:
> > > On Wed, 2013-05-29 at 12:43 +0100, Wei Liu wrote:
> > > > This patch implements 1:1 model netback. NAPI and kthread are utilized
> > > > to do the weight-lifting job:
> > > > 
> > > >  - NAPI is used for guest side TX (host side RX)
> > > >  - kthread is used for guest side RX (host side TX)
> > > > 
> > > > Xenvif and xen_netbk are made into one structure to reduce code size.
> > > > 
> > > > This model provides better scheduling fairness among vifs. It is also
> > > > prerequisite for implementing multiqueue for Xen netback.
> > > 
> > > Any interesting numbers from this change?
> > > 
> > 
> > [extracted from 0/2 of this series]
> 
> Oops, missed that, sorry.
> 
> > 
> > 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
> 
> The CPU load seems to drop pretty significantly with the new model,
> nice. I suppose that is the impact of NAPI rather than the 1:1 side of
> things.
> 

Indeed.

> > 
> > > > Signed-off-by: Wei Liu <wei.liu2@citrix.com>
> > > > ---
> > > >  drivers/net/xen-netback/common.h    |  105 +++--
> > > >  drivers/net/xen-netback/interface.c |  122 +++---
> > > >  drivers/net/xen-netback/netback.c   |  742 ++++++++++++-----------------------
> > > >  3 files changed, 389 insertions(+), 580 deletions(-)
> > > > 
> > > > diff --git a/drivers/net/xen-netback/common.h b/drivers/net/xen-netback/common.h
> > > > index 8a4d77e..2e270cf 100644
> > > > --- a/drivers/net/xen-netback/common.h
> > > > +++ b/drivers/net/xen-netback/common.h
> > > > @@ -45,15 +45,43 @@
> > > >  #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)
> > > > +
> > > > +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
> > > 
> > 
> > This is only code movement, no funcational change.
> > 
> > > So this sounds like it is only valid for the pending tx corresponding to
> > > the first slot in a frame, right?
> > > 
> > 
> > No. It can also be interpreted as continuation of several merged tx
> > requests.
> > 
> > > It does not contain a link from any arbitrary slot to the slot
> > > containing the head of that frame, right?
> > > 
> > > So for the head it is effectively that entries index into the pending_tx
> > > array?
> > > 
> > 
> > No. It's a flag to tell if this pending_tx_info is the head of one or
> > more merged tx requests.
> 
> OK, then I'm totally confused.
> 
> I think a comment with an example of the representation of a single
> multislot frame (2-3 slots if that suffices as an illustration) in a
> comment would be useful.
> 

Ok.

> > > > +static int xenvif_poll(struct napi_struct *napi, int budget)
> > > > +{
> > > > +	struct xenvif *vif = container_of(napi, struct xenvif, napi);
> > > > +	int work_done;
> > > > +
> > > > +	work_done = xenvif_tx_action(vif, budget);
> > > > +
> > > > +	if (work_done < budget) {
> > > > +		int more_to_do = 0;
> > > > +		unsigned long flags;
> > > > +
> > > > +		local_irq_save(flags);
> > > 
> > > What does this protect against? Maybe it's a napi requirement not an Xen
> > > netif one?
> > > 
> > 
> > This is sort of a NAPI requirement. If we don't do local_irq_save we
> > would lose event.
> 
> How come? This is a race against this code vs. a new event coming in and
> calling napi_schedule? 

Yes. That's the case. After we manipulate the event pointer the frontend
might generate a new event. Our NAPI handler is still scheduled at that
time, it won't be scheduled again. 

Do real drivers all disable interrupts in this
> way too or do they get this for free via some other means?
> 

Real drivers often use hardware mechanism to prevent generating events.

> I think a comment is at least required.
> 

Sure.


Wei.

> Ian.

Re: [PATCH 0/2 V3] xen-netback: switch to NAPI + kthread 1:1 model

[Konrad Rzeszutek Wilk]

On Wed, May 29, 2013 at 12:43:40PM +0100, Wei Liu wrote:
> 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 first two patches are ground work for the third patch. First one
> simplifies code in netback, second one can reduce memory footprint if we
> switch to 1:1 model.
> 
> The third 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 done on a Dell T3400 workstation with 4 cores, running 4
> DomUs. Netserver running 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).
> 
> Comments?

I like it!
> 
> ----------------------------------------------------
> Change since V2:
>  - Remove per-cpu scratch space.
> 
> Change since V1:
>  - No page pool in this version. Instead page tracking facility is
>    removed.
> 
> Wei Liu (2):
>   xen-netback: remove page tracking facility
>   xen-netback: switch to NAPI + kthread 1:1 model
> 
>  drivers/net/xen-netback/common.h    |  105 +++--
>  drivers/net/xen-netback/interface.c |  122 ++++--
>  drivers/net/xen-netback/netback.c   |  817 +++++++++++------------------------
>  3 files changed, 390 insertions(+), 654 deletions(-)

:-)

> 
> -- 
> 1.7.10.4
> 
> 
> _______________________________________________
> Xen-devel mailing list
> Xen-devel@lists.xen.org
> http://lists.xen.org/xen-devel
>