Skip to content

Commit

Permalink
Merge branch 'xen-netback'
Browse files Browse the repository at this point in the history 
Wei Liu says:

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

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

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

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

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

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

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

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

We can see that the impact is mainly on CPU usage. The new model moves
processing from kthread to NAPI (software interrupt).
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
  • Loading branch information
David S. Miller committed Aug 29, 2013
2 parents 5b2941b + 7376419 commit 823a19e
Show file tree
Hide file tree
Showing 3 changed files with 445 additions and 673 deletions.
150 changes: 103 additions & 47 deletions drivers/net/xen-netback/common.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -45,31 +45,109 @@
#include <xen/grant_table.h>
#include <xen/xenbus.h>

struct xen_netbk;
typedef unsigned int pending_ring_idx_t;
#define INVALID_PENDING_RING_IDX (~0U)

/* For the head field in pending_tx_info: it is used to indicate
* whether this tx info is the head of one or more coalesced requests.
*
* When head != INVALID_PENDING_RING_IDX, it means the start of a new
* tx requests queue and the end of previous queue.
*
* An example sequence of head fields (I = INVALID_PENDING_RING_IDX):
*
* ...|0 I I I|5 I|9 I I I|...
* -->|<-INUSE----------------
*
* After consuming the first slot(s) we have:
*
* ...|V V V V|5 I|9 I I I|...
* -----FREE->|<-INUSE--------
*
* where V stands for "valid pending ring index". Any number other
* than INVALID_PENDING_RING_IDX is OK. These entries are considered
* free and can contain any number other than
* INVALID_PENDING_RING_IDX. In practice we use 0.
*
* The in use non-INVALID_PENDING_RING_IDX (say 0, 5 and 9 in the
* above example) number is the index into pending_tx_info and
* mmap_pages arrays.
*/
struct pending_tx_info {
struct xen_netif_tx_request req; /* coalesced tx request */
pending_ring_idx_t head; /* head != INVALID_PENDING_RING_IDX
* if it is head of one or more tx
* reqs
*/
};

#define XEN_NETIF_TX_RING_SIZE __CONST_RING_SIZE(xen_netif_tx, PAGE_SIZE)
#define XEN_NETIF_RX_RING_SIZE __CONST_RING_SIZE(xen_netif_rx, PAGE_SIZE)

struct xenvif_rx_meta {
int id;
int size;
int gso_size;
};

/* Discriminate from any valid pending_idx value. */
#define INVALID_PENDING_IDX 0xFFFF

#define MAX_BUFFER_OFFSET PAGE_SIZE

#define MAX_PENDING_REQS 256

struct xenvif {
/* Unique identifier for this interface. */
domid_t domid;
unsigned int handle;

/* Reference to netback processing backend. */
struct xen_netbk *netbk;
/* Use NAPI for guest TX */
struct napi_struct napi;
/* When feature-split-event-channels = 0, tx_irq = rx_irq. */
unsigned int tx_irq;
/* Only used when feature-split-event-channels = 1 */
char tx_irq_name[IFNAMSIZ+4]; /* DEVNAME-tx */
struct xen_netif_tx_back_ring tx;
struct sk_buff_head tx_queue;
struct page *mmap_pages[MAX_PENDING_REQS];
pending_ring_idx_t pending_prod;
pending_ring_idx_t pending_cons;
u16 pending_ring[MAX_PENDING_REQS];
struct pending_tx_info pending_tx_info[MAX_PENDING_REQS];

/* Coalescing tx requests before copying makes number of grant
* copy ops greater or equal to number of slots required. In
* worst case a tx request consumes 2 gnttab_copy.
*/
struct gnttab_copy tx_copy_ops[2*MAX_PENDING_REQS];

u8 fe_dev_addr[6];

/* Use kthread for guest RX */
struct task_struct *task;
wait_queue_head_t wq;
/* When feature-split-event-channels = 0, tx_irq = rx_irq. */
unsigned int tx_irq;
unsigned int rx_irq;
/* Only used when feature-split-event-channels = 1 */
char tx_irq_name[IFNAMSIZ+4]; /* DEVNAME-tx */
char rx_irq_name[IFNAMSIZ+4]; /* DEVNAME-rx */
struct xen_netif_rx_back_ring rx;
struct sk_buff_head rx_queue;

/* List of frontends to notify after a batch of frames sent. */
struct list_head notify_list;
/* Allow xenvif_start_xmit() to peek ahead in the rx request
* ring. This is a prediction of what rx_req_cons will be
* once all queued skbs are put on the ring.
*/
RING_IDX rx_req_cons_peek;

/* Given MAX_BUFFER_OFFSET of 4096 the worst case is that each
* head/fragment page uses 2 copy operations because it
* straddles two buffers in the frontend.
*/
struct gnttab_copy grant_copy_op[2*XEN_NETIF_RX_RING_SIZE];
struct xenvif_rx_meta meta[2*XEN_NETIF_RX_RING_SIZE];

/* The shared rings and indexes. */
struct xen_netif_tx_back_ring tx;
struct xen_netif_rx_back_ring rx;

u8 fe_dev_addr[6];

/* Frontend feature information. */
u8 can_sg:1;
Expand All @@ -80,13 +158,6 @@ struct xenvif {
/* Internal feature information. */
u8 can_queue:1; /* can queue packets for receiver? */

/*
* Allow xenvif_start_xmit() to peek ahead in the rx request
* ring. This is a prediction of what rx_req_cons will be
* once all queued skbs are put on the ring.
*/
RING_IDX rx_req_cons_peek;

/* Transmit shaping: allow 'credit_bytes' every 'credit_usec'. */
unsigned long credit_bytes;
unsigned long credit_usec;
Expand All @@ -97,21 +168,14 @@ struct xenvif {
unsigned long rx_gso_checksum_fixup;

/* Miscellaneous private stuff. */
struct list_head schedule_list;
atomic_t refcnt;
struct net_device *dev;

wait_queue_head_t waiting_to_free;
};

static inline struct xenbus_device *xenvif_to_xenbus_device(struct xenvif *vif)
{
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);
Expand All @@ -121,47 +185,39 @@ 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);

/* Prevent the device from generating any further traffic. */
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;

Expand Down
Loading

0 comments on commit 823a19e

Please sign in to comment.