@@ -61,6 +61,27 @@ The maximum number of packets that kernel can handle on a NAPI interrupt,
it's a Per-CPU variable.
Default: 64
+dev_weight_rx_bias
+--------------
+
+RPS (e.g. RFS, aRFS) processing is competing with the registered NAPI poll function
+of the driver for the per softirq cycle netdev_budget. This parameter influences
+the proportion of the configured netdev_budget that is spent on RPS based packet
+processing during RX softirq cycles. It is further meant for making current
+dev_weight adaptable for asymmetric CPU needs on RX/TX side of the network stack.
+(see dev_weight_tx_bias) It is effective on a per CPU basis. Determination is based
+on dev_weight and is calculated multiplicative (dev_weight * dev_weight_rx_bias).
+Default: 1
+
+dev_weight_tx_bias
+--------------
+
+Scales the maximum number of packets that can be processed during a TX softirq cycle.
+Effective on a per CPU basis. Allows scaling of current dev_weight for asymmetric
+net stack processing needs. Be careful to avoid making TX softirq processing a CPU hog.
+Calculation is based on dev_weight (dev_weight * dev_weight_tx_bias).
+Default: 1
+
default_qdisc
--------------
@@ -3795,6 +3795,10 @@ void netdev_stats_to_stats64(struct rtnl_link_stats64 *stats64,
extern int netdev_max_backlog;
extern int netdev_tstamp_prequeue;
extern int weight_p;
+extern int dev_weight_rx_bias;
+extern int dev_weight_tx_bias;
+extern int dev_rx_weight;
+extern int dev_tx_weight;
bool netdev_has_upper_dev(struct net_device *dev, struct net_device *upper_dev);
struct net_device *netdev_upper_get_next_dev_rcu(struct net_device *dev,
@@ -3427,7 +3427,11 @@ EXPORT_SYMBOL(netdev_max_backlog);
int netdev_tstamp_prequeue __read_mostly = 1;
int netdev_budget __read_mostly = 300;
-int weight_p __read_mostly = 64; /* old backlog weight */
+int weight_p __read_mostly = 64; /* old backlog weight */
+int dev_weight_rx_bias __read_mostly = 1; /* bias for backlog weight */
+int dev_weight_tx_bias __read_mostly = 1; /* bias for output_queue quota */
+int dev_rx_weight __read_mostly = 64;
+int dev_tx_weight __read_mostly = 64;
/* Called with irq disabled */
static inline void ____napi_schedule(struct softnet_data *sd,
@@ -4833,7 +4837,7 @@ static int process_backlog(struct napi_struct *napi, int quota)
net_rps_action_and_irq_enable(sd);
}
- napi->weight = weight_p;
+ napi->weight = dev_rx_weight;
while (again) {
struct sk_buff *skb;
@@ -222,6 +222,21 @@ static int set_default_qdisc(struct ctl_table *table, int write,
}
#endif
+static int proc_do_dev_weight(struct ctl_table *table, int write,
+ void __user *buffer, size_t *lenp, loff_t *ppos)
+{
+ int ret;
+
+ ret = proc_dointvec(table, write, buffer, lenp, ppos);
+ if (ret != 0)
+ return ret;
+
+ dev_rx_weight = weight_p * dev_weight_rx_bias;
+ dev_tx_weight = weight_p * dev_weight_tx_bias;
+
+ return ret;
+}
+
static int proc_do_rss_key(struct ctl_table *table, int write,
void __user *buffer, size_t *lenp, loff_t *ppos)
{
@@ -273,7 +288,21 @@ static struct ctl_table net_core_table[] = {
.data = &weight_p,
.maxlen = sizeof(int),
.mode = 0644,
- .proc_handler = proc_dointvec
+ .proc_handler = proc_do_dev_weight,
+ },
+ {
+ .procname = "dev_weight_rx_bias",
+ .data = &dev_weight_rx_bias,
+ .maxlen = sizeof(int),
+ .mode = 0644,
+ .proc_handler = proc_do_dev_weight,
+ },
+ {
+ .procname = "dev_weight_tx_bias",
+ .data = &dev_weight_tx_bias,
+ .maxlen = sizeof(int),
+ .mode = 0644,
+ .proc_handler = proc_do_dev_weight,
},
{
.procname = "netdev_max_backlog",
@@ -247,7 +247,7 @@ static inline int qdisc_restart(struct Qdisc *q, int *packets)
void __qdisc_run(struct Qdisc *q)
{
- int quota = weight_p;
+ int quota = dev_tx_weight;
int packets;
while (qdisc_restart(q, &packets)) {
Oftenly, introducing side effects on packet processing on the other half of the stack by adjusting one of TX/RX via sysctl is not desirable. There are cases of demand for asymmetric, orthogonal configurability. This holds true especially for nodes where RPS for RFS usage on top is configured and therefore use the 'old dev_weight'. This is quite a common base configuration setup nowadays, even with NICs of superior processing support (e.g. aRFS). A good example use case are nodes acting as noSQL data bases with a large number of tiny requests and rather fewer but large packets as responses. It's affordable to have large budget and rx dev_weights for the requests. But as a side effect having this large a number on TX processed in one run can overwhelm drivers. This patch therefore introduces an independent configurability via sysctl to userland. Signed-off-by: Matthias Tafelmeier <matthias.tafelmeier@gmx.net> --- Documentation/sysctl/net.txt | 21 +++++++++++++++++++++ include/linux/netdevice.h | 4 ++++ net/core/dev.c | 8 ++++++-- net/core/sysctl_net_core.c | 31 ++++++++++++++++++++++++++++++- net/sched/sch_generic.c | 2 +- 5 files changed, 62 insertions(+), 4 deletions(-)