From: Dongxiao Xu <dongxiao.xu@intel.com>
Subject: [PATCH 2/3] Netback: Multiple tasklets support.
Patch-mainline: n/a
Now netback uses one pair of tasklets for Tx/Rx data transaction. Netback
tasklet could only run at one CPU at a time, and it is used to serve all the
netfronts. Therefore it has become a performance bottle neck. This patch is to
use multiple tasklet pairs to replace the current single pair in dom0.
Assuming that Dom0 has CPUNR VCPUs, we define CPUNR kinds of tasklets pair
(CPUNR for Tx, and CPUNR for Rx). Each pare of tasklets serve specific group of
netfronts. Also for those global and static variables, we duplicated them for
each group in order to avoid the spinlock.
Signed-off-by: Dongxiao Xu <dongxiao.xu@intel.com>
jb: some cleanups
Acked-by: jbeulich@novell.com
--- sle11sp1-2010-03-05.orig/drivers/xen/netback/common.h 2010-01-14 08:39:00.000000000 +0100
+++ sle11sp1-2010-03-05/drivers/xen/netback/common.h 2010-01-14 08:40:17.000000000 +0100
@@ -58,6 +58,7 @@
typedef struct netif_st {
/* Unique identifier for this interface. */
domid_t domid;
+ unsigned int group;
unsigned int handle;
u8 fe_dev_addr[6];
@@ -99,6 +100,7 @@ typedef struct netif_st {
/* Miscellaneous private stuff. */
struct list_head list; /* scheduling list */
+ struct list_head group_list;
atomic_t refcnt;
struct net_device *dev;
struct net_device_stats stats;
@@ -259,12 +261,15 @@ struct xen_netbk {
struct list_head pending_inuse_head;
struct list_head net_schedule_list;
+ struct list_head group_domain_list;
spinlock_t net_schedule_list_lock;
spinlock_t release_lock;
+ spinlock_t group_domain_list_lock;
struct page **mmap_pages;
+ unsigned int group_domain_nr;
unsigned int alloc_index;
struct page_ext page_extinfo[MAX_PENDING_REQS];
@@ -294,4 +299,8 @@ struct xen_netbk {
unsigned long mfn_list[MAX_MFN_ALLOC];
};
+
+extern struct xen_netbk *xen_netbk;
+extern unsigned int netbk_nr_groups;
+
#endif /* __NETIF__BACKEND__COMMON_H__ */
--- sle11sp1-2010-03-05.orig/drivers/xen/netback/interface.c 2010-01-04 13:31:46.000000000 +0100
+++ sle11sp1-2010-03-05/drivers/xen/netback/interface.c 2010-03-05 10:35:22.000000000 +0100
@@ -54,14 +54,41 @@ module_param_named(queue_length, netbk_q
static void __netif_up(netif_t *netif)
{
+ unsigned int group = 0;
+ unsigned int min_domains = xen_netbk[0].group_domain_nr;
+ unsigned int i;
+
+ /* Find the list which contains least number of domains. */
+ for (i = 1; i < netbk_nr_groups; i++) {
+ if (xen_netbk[i].group_domain_nr < min_domains) {
+ group = i;
+ min_domains = xen_netbk[i].group_domain_nr;
+ }
+ }
+
+ spin_lock(&xen_netbk[group].group_domain_list_lock);
+ list_add_tail(&netif->group_list,
+ &xen_netbk[group].group_domain_list);
+ xen_netbk[group].group_domain_nr++;
+ spin_unlock(&xen_netbk[group].group_domain_list_lock);
+ netif->group = group;
+
enable_irq(netif->irq);
netif_schedule_work(netif);
}
static void __netif_down(netif_t *netif)
{
+ struct xen_netbk *netbk = xen_netbk + netif->group;
+
disable_irq(netif->irq);
netif_deschedule_work(netif);
+
+ netif->group = UINT_MAX;
+ spin_lock(&netbk->group_domain_list_lock);
+ netbk->group_domain_nr--;
+ list_del(&netif->group_list);
+ spin_unlock(&netbk->group_domain_list_lock);
}
static int net_open(struct net_device *dev)
@@ -203,6 +230,7 @@ netif_t *netif_alloc(struct device *pare
netif = netdev_priv(dev);
memset(netif, 0, sizeof(*netif));
netif->domid = domid;
+ netif->group = UINT_MAX;
netif->handle = handle;
atomic_set(&netif->refcnt, 1);
init_waitqueue_head(&netif->waiting_to_free);
--- sle11sp1-2010-03-05.orig/drivers/xen/netback/netback.c 2010-01-29 12:51:48.000000000 +0100
+++ sle11sp1-2010-03-05/drivers/xen/netback/netback.c 2010-03-08 10:54:19.000000000 +0100
@@ -41,10 +41,10 @@
/*define NETBE_DEBUG_INTERRUPT*/
-static struct xen_netbk *xen_netbk;
-static unsigned int netbk_nr_groups = 1;
+struct xen_netbk *xen_netbk;
+unsigned int netbk_nr_groups;
-#define GET_GROUP_INDEX(netif) (0)
+#define GET_GROUP_INDEX(netif) ((netif)->group)
static void netif_idx_release(struct xen_netbk *, u16 pending_idx);
static void make_tx_response(netif_t *netif,
@@ -93,6 +93,8 @@ MODULE_PARM_DESC(copy_skb, "Copy data re
static int MODPARM_permute_returns = 0;
module_param_named(permute_returns, MODPARM_permute_returns, bool, S_IRUSR|S_IWUSR);
MODULE_PARM_DESC(permute_returns, "Randomly permute the order in which TX responses are sent to the frontend");
+module_param_named(groups, netbk_nr_groups, uint, 0);
+MODULE_PARM_DESC(groups, "Specify the number of tasklet pairs to use");
int netbk_copy_skb_mode;
@@ -1519,9 +1521,20 @@ static void netif_page_release(struct pa
irqreturn_t netif_be_int(int irq, void *dev_id)
{
netif_t *netif = dev_id;
+ unsigned int group = GET_GROUP_INDEX(netif);
+
+ if (unlikely(group >= netbk_nr_groups)) {
+ /*
+ * Short of having a way to bind the IRQ in disabled mode
+ * (IRQ_NOAUTOEN), we have to ignore the first invocation(s)
+ * (before we got assigned to a group).
+ */
+ BUG_ON(group != UINT_MAX);
+ return IRQ_HANDLED;
+ }
add_to_net_schedule_list_tail(netif);
- maybe_schedule_tx_action(GET_GROUP_INDEX(netif));
+ maybe_schedule_tx_action(group);
if (netif_schedulable(netif) && !netbk_queue_full(netif))
netif_wake_queue(netif->dev);
@@ -1637,8 +1650,11 @@ static int __init netback_init(void)
if (!is_running_on_xen())
return -ENODEV;
+ if (!netbk_nr_groups)
+ netbk_nr_groups = (num_online_cpus() + 1) / 2;
+
/* We can increase reservation by this much in net_rx_action(). */
- balloon_update_driver_allowance(NET_RX_RING_SIZE);
+ balloon_update_driver_allowance(netbk_nr_groups * NET_RX_RING_SIZE);
xen_netbk = __vmalloc(netbk_nr_groups * sizeof(*xen_netbk),
GFP_KERNEL|__GFP_HIGHMEM|__GFP_ZERO, PAGE_KERNEL);
@@ -1677,9 +1693,11 @@ static int __init netback_init(void)
INIT_LIST_HEAD(&netbk->pending_inuse_head);
INIT_LIST_HEAD(&netbk->net_schedule_list);
+ INIT_LIST_HEAD(&netbk->group_domain_list);
spin_lock_init(&netbk->net_schedule_list_lock);
spin_lock_init(&netbk->release_lock);
+ spin_lock_init(&netbk->group_domain_list_lock);
for (i = 0; i < MAX_PENDING_REQS; i++) {
page = netbk->mmap_pages[i];
