@@ -12,6 +12,7 @@
#include "rmnet_handlers.h"
#include "rmnet_vnd.h"
#include "rmnet_private.h"
+#include "rmnet_map.h"
/* Local Definitions and Declarations */
@@ -39,6 +40,8 @@ static int rmnet_unregister_real_device(struct net_device *real_dev)
if (port->nr_rmnet_devs)
return -EINVAL;
+ rmnet_map_tx_aggregate_exit(port);
+
netdev_rx_handler_unregister(real_dev);
kfree(port);
@@ -79,6 +82,8 @@ static int rmnet_register_real_device(struct net_device *real_dev,
for (entry = 0; entry < RMNET_MAX_LOGICAL_EP; entry++)
INIT_HLIST_HEAD(&port->muxed_ep[entry]);
+ rmnet_map_tx_aggregate_init(port);
+
netdev_dbg(real_dev, "registered with rmnet\n");
return 0;
}
@@ -6,6 +6,7 @@
*/
#include <linux/skbuff.h>
+#include <linux/time.h>
#include <net/gro_cells.h>
#ifndef _RMNET_CONFIG_H_
@@ -19,6 +20,12 @@ struct rmnet_endpoint {
struct hlist_node hlnode;
};
+struct rmnet_egress_agg_params {
+ u32 bytes;
+ u32 count;
+ u64 time_nsec;
+};
+
/* One instance of this structure is instantiated for each real_dev associated
* with rmnet.
*/
@@ -30,6 +37,19 @@ struct rmnet_port {
struct hlist_head muxed_ep[RMNET_MAX_LOGICAL_EP];
struct net_device *bridge_ep;
struct net_device *rmnet_dev;
+
+ /* Egress aggregation information */
+ struct rmnet_egress_agg_params egress_agg_params;
+ /* Protect aggregation related elements */
+ spinlock_t agg_lock;
+ struct sk_buff *skbagg_head;
+ struct sk_buff *skbagg_tail;
+ int agg_state;
+ u8 agg_count;
+ struct timespec64 agg_time;
+ struct timespec64 agg_last;
+ struct hrtimer hrtimer;
+ struct work_struct agg_wq;
};
extern struct rtnl_link_ops rmnet_link_ops;
@@ -164,8 +164,18 @@ static int rmnet_map_egress_handler(struct sk_buff *skb,
map_header->mux_id = mux_id;
- skb->protocol = htons(ETH_P_MAP);
+ if (READ_ONCE(port->egress_agg_params.count) > 1) {
+ unsigned int len;
+
+ len = rmnet_map_tx_aggregate(skb, port, orig_dev);
+ if (likely(len)) {
+ rmnet_vnd_tx_fixup_len(len, orig_dev);
+ return -EINPROGRESS;
+ }
+ return -ENOMEM;
+ }
+ skb->protocol = htons(ETH_P_MAP);
return 0;
}
@@ -235,6 +245,7 @@ void rmnet_egress_handler(struct sk_buff *skb)
struct rmnet_port *port;
struct rmnet_priv *priv;
u8 mux_id;
+ int err;
sk_pacing_shift_update(skb->sk, 8);
@@ -247,8 +258,11 @@ void rmnet_egress_handler(struct sk_buff *skb)
if (!port)
goto drop;
- if (rmnet_map_egress_handler(skb, port, mux_id, orig_dev))
+ err = rmnet_map_egress_handler(skb, port, mux_id, orig_dev);
+ if (err == -ENOMEM)
goto drop;
+ else if (err == -EINPROGRESS)
+ return;
rmnet_vnd_tx_fixup(skb, orig_dev);
@@ -53,5 +53,11 @@ void rmnet_map_checksum_uplink_packet(struct sk_buff *skb,
struct net_device *orig_dev,
int csum_type);
int rmnet_map_process_next_hdr_packet(struct sk_buff *skb, u16 len);
+unsigned int rmnet_map_tx_aggregate(struct sk_buff *skb, struct rmnet_port *port,
+ struct net_device *orig_dev);
+void rmnet_map_tx_aggregate_init(struct rmnet_port *port);
+void rmnet_map_tx_aggregate_exit(struct rmnet_port *port);
+void rmnet_map_update_ul_agg_config(struct rmnet_port *port, u32 size,
+ u32 count, u32 time);
#endif /* _RMNET_MAP_H_ */
@@ -12,6 +12,7 @@
#include "rmnet_config.h"
#include "rmnet_map.h"
#include "rmnet_private.h"
+#include "rmnet_vnd.h"
#define RMNET_MAP_DEAGGR_SPACING 64
#define RMNET_MAP_DEAGGR_HEADROOM (RMNET_MAP_DEAGGR_SPACING / 2)
@@ -518,3 +519,193 @@ int rmnet_map_process_next_hdr_packet(struct sk_buff *skb,
return 0;
}
+
+#define RMNET_AGG_BYPASS_TIME_NSEC 10000000L
+
+static void reset_aggr_params(struct rmnet_port *port)
+{
+ port->skbagg_head = NULL;
+ port->agg_count = 0;
+ port->agg_state = 0;
+ memset(&port->agg_time, 0, sizeof(struct timespec64));
+}
+
+static void rmnet_send_skb(struct rmnet_port *port, struct sk_buff *skb)
+{
+ if (skb_needs_linearize(skb, port->dev->features)) {
+ if (unlikely(__skb_linearize(skb))) {
+ struct rmnet_priv *priv;
+
+ priv = netdev_priv(port->rmnet_dev);
+ this_cpu_inc(priv->pcpu_stats->stats.tx_drops);
+ dev_kfree_skb_any(skb);
+ return;
+ }
+ }
+
+ dev_queue_xmit(skb);
+}
+
+static void rmnet_map_flush_tx_packet_work(struct work_struct *work)
+{
+ struct sk_buff *skb = NULL;
+ struct rmnet_port *port;
+
+ port = container_of(work, struct rmnet_port, agg_wq);
+
+ spin_lock_bh(&port->agg_lock);
+ if (likely(port->agg_state == -EINPROGRESS)) {
+ /* Buffer may have already been shipped out */
+ if (likely(port->skbagg_head)) {
+ skb = port->skbagg_head;
+ reset_aggr_params(port);
+ }
+ port->agg_state = 0;
+ }
+
+ spin_unlock_bh(&port->agg_lock);
+ if (skb)
+ rmnet_send_skb(port, skb);
+}
+
+static enum hrtimer_restart rmnet_map_flush_tx_packet_queue(struct hrtimer *t)
+{
+ struct rmnet_port *port;
+
+ port = container_of(t, struct rmnet_port, hrtimer);
+
+ schedule_work(&port->agg_wq);
+
+ return HRTIMER_NORESTART;
+}
+
+unsigned int rmnet_map_tx_aggregate(struct sk_buff *skb, struct rmnet_port *port,
+ struct net_device *orig_dev)
+{
+ struct timespec64 diff, last;
+ unsigned int len = skb->len;
+ struct sk_buff *agg_skb;
+ int size;
+
+ spin_lock_bh(&port->agg_lock);
+ memcpy(&last, &port->agg_last, sizeof(struct timespec64));
+ ktime_get_real_ts64(&port->agg_last);
+
+ if (!port->skbagg_head) {
+ /* Check to see if we should agg first. If the traffic is very
+ * sparse, don't aggregate.
+ */
+new_packet:
+ diff = timespec64_sub(port->agg_last, last);
+ size = port->egress_agg_params.bytes - skb->len;
+
+ if (size < 0) {
+ /* dropped */
+ spin_unlock_bh(&port->agg_lock);
+ return 0;
+ }
+
+ if (diff.tv_sec > 0 || diff.tv_nsec > RMNET_AGG_BYPASS_TIME_NSEC ||
+ size == 0)
+ goto no_aggr;
+
+ port->skbagg_head = skb_copy_expand(skb, 0, size, GFP_ATOMIC);
+ if (!port->skbagg_head)
+ goto no_aggr;
+
+ dev_kfree_skb_any(skb);
+ port->skbagg_head->protocol = htons(ETH_P_MAP);
+ port->agg_count = 1;
+ ktime_get_real_ts64(&port->agg_time);
+ skb_frag_list_init(port->skbagg_head);
+ goto schedule;
+ }
+ diff = timespec64_sub(port->agg_last, port->agg_time);
+ size = port->egress_agg_params.bytes - port->skbagg_head->len;
+
+ if (skb->len > size) {
+ agg_skb = port->skbagg_head;
+ reset_aggr_params(port);
+ spin_unlock_bh(&port->agg_lock);
+ hrtimer_cancel(&port->hrtimer);
+ rmnet_send_skb(port, agg_skb);
+ spin_lock_bh(&port->agg_lock);
+ goto new_packet;
+ }
+
+ if (skb_has_frag_list(port->skbagg_head))
+ port->skbagg_tail->next = skb;
+ else
+ skb_shinfo(port->skbagg_head)->frag_list = skb;
+
+ port->skbagg_head->len += skb->len;
+ port->skbagg_head->data_len += skb->len;
+ port->skbagg_head->truesize += skb->truesize;
+ port->skbagg_tail = skb;
+ port->agg_count++;
+
+ if (diff.tv_sec > 0 || diff.tv_nsec > port->egress_agg_params.time_nsec ||
+ port->agg_count >= port->egress_agg_params.count ||
+ port->skbagg_head->len == port->egress_agg_params.bytes) {
+ agg_skb = port->skbagg_head;
+ reset_aggr_params(port);
+ spin_unlock_bh(&port->agg_lock);
+ hrtimer_cancel(&port->hrtimer);
+ rmnet_send_skb(port, agg_skb);
+ return len;
+ }
+
+schedule:
+ if (!hrtimer_active(&port->hrtimer) && port->agg_state != -EINPROGRESS) {
+ port->agg_state = -EINPROGRESS;
+ hrtimer_start(&port->hrtimer,
+ ns_to_ktime(port->egress_agg_params.time_nsec),
+ HRTIMER_MODE_REL);
+ }
+ spin_unlock_bh(&port->agg_lock);
+
+ return len;
+
+no_aggr:
+ spin_unlock_bh(&port->agg_lock);
+ skb->protocol = htons(ETH_P_MAP);
+ dev_queue_xmit(skb);
+
+ return len;
+}
+
+void rmnet_map_update_ul_agg_config(struct rmnet_port *port, u32 size,
+ u32 count, u32 time)
+{
+ spin_lock_bh(&port->agg_lock);
+ port->egress_agg_params.bytes = size;
+ WRITE_ONCE(port->egress_agg_params.count, count);
+ port->egress_agg_params.time_nsec = time * NSEC_PER_USEC;
+ spin_unlock_bh(&port->agg_lock);
+}
+
+void rmnet_map_tx_aggregate_init(struct rmnet_port *port)
+{
+ hrtimer_init(&port->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
+ port->hrtimer.function = rmnet_map_flush_tx_packet_queue;
+ spin_lock_init(&port->agg_lock);
+ rmnet_map_update_ul_agg_config(port, 4096, 1, 800);
+ INIT_WORK(&port->agg_wq, rmnet_map_flush_tx_packet_work);
+}
+
+void rmnet_map_tx_aggregate_exit(struct rmnet_port *port)
+{
+ hrtimer_cancel(&port->hrtimer);
+ cancel_work_sync(&port->agg_wq);
+
+ spin_lock_bh(&port->agg_lock);
+ if (port->agg_state == -EINPROGRESS) {
+ if (port->skbagg_head) {
+ dev_kfree_skb_any(port->skbagg_head);
+ reset_aggr_params(port);
+ }
+
+ port->agg_state = 0;
+ }
+ spin_unlock_bh(&port->agg_lock);
+}
@@ -29,7 +29,7 @@ void rmnet_vnd_rx_fixup(struct sk_buff *skb, struct net_device *dev)
u64_stats_update_end(&pcpu_ptr->syncp);
}
-void rmnet_vnd_tx_fixup(struct sk_buff *skb, struct net_device *dev)
+void rmnet_vnd_tx_fixup_len(unsigned int len, struct net_device *dev)
{
struct rmnet_priv *priv = netdev_priv(dev);
struct rmnet_pcpu_stats *pcpu_ptr;
@@ -38,10 +38,15 @@ void rmnet_vnd_tx_fixup(struct sk_buff *skb, struct net_device *dev)
u64_stats_update_begin(&pcpu_ptr->syncp);
pcpu_ptr->stats.tx_pkts++;
- pcpu_ptr->stats.tx_bytes += skb->len;
+ pcpu_ptr->stats.tx_bytes += len;
u64_stats_update_end(&pcpu_ptr->syncp);
}
+void rmnet_vnd_tx_fixup(struct sk_buff *skb, struct net_device *dev)
+{
+ rmnet_vnd_tx_fixup_len(skb->len, dev);
+}
+
/* Network Device Operations */
static netdev_tx_t rmnet_vnd_start_xmit(struct sk_buff *skb,
@@ -16,6 +16,7 @@ int rmnet_vnd_newlink(u8 id, struct net_device *rmnet_dev,
int rmnet_vnd_dellink(u8 id, struct rmnet_port *port,
struct rmnet_endpoint *ep);
void rmnet_vnd_rx_fixup(struct sk_buff *skb, struct net_device *dev);
+void rmnet_vnd_tx_fixup_len(unsigned int len, struct net_device *dev);
void rmnet_vnd_tx_fixup(struct sk_buff *skb, struct net_device *dev);
void rmnet_vnd_setup(struct net_device *dev);
int rmnet_vnd_validate_real_dev_mtu(struct net_device *real_dev);
Add tx packets aggregation. Bidirectional TCP throughput tests through iperf with low-cat Thread-x based modems revelead performance issues both in tx and rx. The Windows driver does not show this issue: inspecting USB packets revealed that the only notable change is the driver enabling tx packets aggregation. Tx packets aggregation is by default disabled and can be enabled by increasing the value of ETHTOOL_A_COALESCE_TX_MAX_AGGR_FRAMES. The maximum aggregated size is by default set to a reasonably low value in order to support the majority of modems. This implementation is based on patches available in Code Aurora repositories (msm kernel) whose main authors are Subash Abhinov Kasiviswanathan <subashab@codeaurora.org> Sean Tranchetti <stranche@codeaurora.org> Signed-off-by: Daniele Palmas <dnlplm@gmail.com> --- Hi all, in v4 I should have met all Paolo's remarks, besides the one related to removing the scheduled work, since according to my understanding of the code in dev_queue_xmit it's not possible to avoid the WARN_ONCE when called directly from the timer function. Context at https://lore.kernel.org/netdev/CAGRyCJGHSPO+i_xKHGbNg+Hki5tQC3_6Kc8RNcHWN6pxQdjODw@mail.gmail.com/ v4 - Solved race when accessing egress_agg_params - Removed duplicated code for error in rmnet_map_tx_aggregate v3 - Fixed no previous prototype for rmnet_map_flush_tx_packet_queue warning reported by kernel test robot v2 - Removed icmp packets direct sending - Changed spin_lock_irqsave to spin_lock_bh - Increased the possible maximum size of an aggregated block - Aligned rmnet_egress_agg_params and types to ethtool ones - Changed bypass time from variable to define - Fixed RCT style in rmnet_map_tx_aggregate - Fixed order of skb freeing in rmnet_map_tx_aggregate - rmnet_map_tx_aggregate refactoring - Change aggregation function to use frag_list - Removed RMNET_FLAGS_EGRESS_AGGREGATION --- .../ethernet/qualcomm/rmnet/rmnet_config.c | 5 + .../ethernet/qualcomm/rmnet/rmnet_config.h | 20 ++ .../ethernet/qualcomm/rmnet/rmnet_handlers.c | 18 +- .../net/ethernet/qualcomm/rmnet/rmnet_map.h | 6 + .../ethernet/qualcomm/rmnet/rmnet_map_data.c | 191 ++++++++++++++++++ .../net/ethernet/qualcomm/rmnet/rmnet_vnd.c | 9 +- .../net/ethernet/qualcomm/rmnet/rmnet_vnd.h | 1 + 7 files changed, 246 insertions(+), 4 deletions(-)