Commit cb3c7fd4 authored by Gil Rockah's avatar Gil Rockah Committed by David S. Miller
Browse files

net/mlx5e: Support adaptive RX coalescing



Striving for high message rate and low interrupt rate.

Usage:
        ethtool -C <interface> adaptive-rx on/off

Signed-off-by: default avatarGil Rockah <gilr@mellanox.com>
Signed-off-by: default avatarAchiad Shochat <achiad@mellanox.com>
Signed-off-by: default avatarSaeed Mahameed <saeedm@mellanox.com>
CC: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: default avatarDavid S. Miller <davem@davemloft.net>
parent 9908aa29
......@@ -7,6 +7,7 @@ mlx5_core-y := main.o cmd.o debugfs.o fw.o eq.o uar.o pagealloc.o \
mlx5_core-$(CONFIG_MLX5_CORE_EN) += wq.o eswitch.o \
en_main.o en_fs.o en_ethtool.o en_tx.o en_rx.o \
en_txrx.o en_clock.o vxlan.o en_tc.o en_arfs.o
en_rx_am.o en_txrx.o en_clock.o vxlan.o en_tc.o \
en_arfs.o
mlx5_core-$(CONFIG_MLX5_CORE_EN_DCB) += en_dcbnl.o
......@@ -195,6 +195,7 @@ struct mlx5e_params {
#ifdef CONFIG_MLX5_CORE_EN_DCB
struct ieee_ets ets;
#endif
bool rx_am_enabled;
};
struct mlx5e_tstamp {
......@@ -213,6 +214,7 @@ struct mlx5e_tstamp {
enum {
MLX5E_RQ_STATE_POST_WQES_ENABLE,
MLX5E_RQ_STATE_UMR_WQE_IN_PROGRESS,
MLX5E_RQ_STATE_AM,
};
struct mlx5e_cq {
......@@ -220,6 +222,7 @@ struct mlx5e_cq {
struct mlx5_cqwq wq;
/* data path - accessed per napi poll */
u16 event_ctr;
struct napi_struct *napi;
struct mlx5_core_cq mcq;
struct mlx5e_channel *channel;
......@@ -247,6 +250,30 @@ struct mlx5e_dma_info {
dma_addr_t addr;
};
struct mlx5e_rx_am_stats {
int ppms; /* packets per msec */
int epms; /* events per msec */
};
struct mlx5e_rx_am_sample {
ktime_t time;
unsigned int pkt_ctr;
u16 event_ctr;
};
struct mlx5e_rx_am { /* Adaptive Moderation */
u8 state;
struct mlx5e_rx_am_stats prev_stats;
struct mlx5e_rx_am_sample start_sample;
struct work_struct work;
u8 profile_ix;
u8 mode;
u8 tune_state;
u8 steps_right;
u8 steps_left;
u8 tired;
};
struct mlx5e_rq {
/* data path */
struct mlx5_wq_ll wq;
......@@ -267,6 +294,8 @@ struct mlx5e_rq {
unsigned long state;
int ix;
struct mlx5e_rx_am am; /* Adaptive Moderation */
/* control */
struct mlx5_wq_ctrl wq_ctrl;
u8 wq_type;
......@@ -637,6 +666,10 @@ void mlx5e_free_rx_fragmented_mpwqe(struct mlx5e_rq *rq,
struct mlx5e_mpw_info *wi);
struct mlx5_cqe64 *mlx5e_get_cqe(struct mlx5e_cq *cq);
void mlx5e_rx_am(struct mlx5e_rq *rq);
void mlx5e_rx_am_work(struct work_struct *work);
struct mlx5e_cq_moder mlx5e_am_get_def_profile(u8 rx_cq_period_mode);
void mlx5e_update_stats(struct mlx5e_priv *priv);
int mlx5e_create_flow_steering(struct mlx5e_priv *priv);
......
......@@ -528,6 +528,7 @@ static int mlx5e_get_coalesce(struct net_device *netdev,
coal->rx_max_coalesced_frames = priv->params.rx_cq_moderation.pkts;
coal->tx_coalesce_usecs = priv->params.tx_cq_moderation.usec;
coal->tx_max_coalesced_frames = priv->params.tx_cq_moderation.pkts;
coal->use_adaptive_rx_coalesce = priv->params.rx_am_enabled;
return 0;
}
......@@ -538,6 +539,10 @@ static int mlx5e_set_coalesce(struct net_device *netdev,
struct mlx5e_priv *priv = netdev_priv(netdev);
struct mlx5_core_dev *mdev = priv->mdev;
struct mlx5e_channel *c;
bool restart =
!!coal->use_adaptive_rx_coalesce != priv->params.rx_am_enabled;
bool was_opened;
int err = 0;
int tc;
int i;
......@@ -546,12 +551,18 @@ static int mlx5e_set_coalesce(struct net_device *netdev,
mutex_lock(&priv->state_lock);
was_opened = test_bit(MLX5E_STATE_OPENED, &priv->state);
if (was_opened && restart) {
mlx5e_close_locked(netdev);
priv->params.rx_am_enabled = !!coal->use_adaptive_rx_coalesce;
}
priv->params.tx_cq_moderation.usec = coal->tx_coalesce_usecs;
priv->params.tx_cq_moderation.pkts = coal->tx_max_coalesced_frames;
priv->params.rx_cq_moderation.usec = coal->rx_coalesce_usecs;
priv->params.rx_cq_moderation.pkts = coal->rx_max_coalesced_frames;
if (!test_bit(MLX5E_STATE_OPENED, &priv->state))
if (!was_opened || restart)
goto out;
for (i = 0; i < priv->params.num_channels; ++i) {
......@@ -570,8 +581,11 @@ static int mlx5e_set_coalesce(struct net_device *netdev,
}
out:
if (was_opened && restart)
err = mlx5e_open_locked(netdev);
mutex_unlock(&priv->state_lock);
return 0;
return err;
}
static u32 ptys2ethtool_supported_link(u32 eth_proto_cap)
......
......@@ -40,8 +40,9 @@
#include "vxlan.h"
struct mlx5e_rq_param {
u32 rqc[MLX5_ST_SZ_DW(rqc)];
struct mlx5_wq_param wq;
u32 rqc[MLX5_ST_SZ_DW(rqc)];
struct mlx5_wq_param wq;
bool am_enabled;
};
struct mlx5e_sq_param {
......@@ -337,6 +338,9 @@ static int mlx5e_create_rq(struct mlx5e_channel *c,
wqe->data.byte_count = cpu_to_be32(byte_count);
}
INIT_WORK(&rq->am.work, mlx5e_rx_am_work);
rq->am.mode = priv->params.rx_cq_period_mode;
rq->wq_type = priv->params.rq_wq_type;
rq->pdev = c->pdev;
rq->netdev = c->netdev;
......@@ -509,6 +513,9 @@ static int mlx5e_open_rq(struct mlx5e_channel *c,
if (err)
goto err_disable_rq;
if (param->am_enabled)
set_bit(MLX5E_RQ_STATE_AM, &c->rq.state);
set_bit(MLX5E_RQ_STATE_POST_WQES_ENABLE, &rq->state);
sq->ico_wqe_info[pi].opcode = MLX5_OPCODE_NOP;
......@@ -537,6 +544,8 @@ static void mlx5e_close_rq(struct mlx5e_rq *rq)
/* avoid destroying rq before mlx5e_poll_rx_cq() is done with it */
napi_synchronize(&rq->channel->napi);
cancel_work_sync(&rq->am.work);
mlx5e_disable_rq(rq);
mlx5e_destroy_rq(rq);
}
......@@ -1112,6 +1121,7 @@ static int mlx5e_open_channel(struct mlx5e_priv *priv, int ix,
{
struct mlx5e_cq_moder icosq_cq_moder = {0, 0};
struct net_device *netdev = priv->netdev;
struct mlx5e_cq_moder rx_cq_profile;
int cpu = mlx5e_get_cpu(priv, ix);
struct mlx5e_channel *c;
struct mlx5e_sq *sq;
......@@ -1130,6 +1140,11 @@ static int mlx5e_open_channel(struct mlx5e_priv *priv, int ix,
c->mkey_be = cpu_to_be32(priv->mkey.key);
c->num_tc = priv->params.num_tc;
if (priv->params.rx_am_enabled)
rx_cq_profile = mlx5e_am_get_def_profile(priv->params.rx_cq_period_mode);
else
rx_cq_profile = priv->params.rx_cq_moderation;
mlx5e_build_channeltc_to_txq_map(priv, ix);
netif_napi_add(netdev, &c->napi, mlx5e_napi_poll, 64);
......@@ -1143,7 +1158,7 @@ static int mlx5e_open_channel(struct mlx5e_priv *priv, int ix,
goto err_close_icosq_cq;
err = mlx5e_open_cq(c, &cparam->rx_cq, &c->rq.cq,
priv->params.rx_cq_moderation);
rx_cq_profile);
if (err)
goto err_close_tx_cqs;
......@@ -1243,6 +1258,8 @@ static void mlx5e_build_rq_param(struct mlx5e_priv *priv,
param->wq.buf_numa_node = dev_to_node(&priv->mdev->pdev->dev);
param->wq.linear = 1;
param->am_enabled = priv->params.rx_am_enabled;
}
static void mlx5e_build_drop_rq_param(struct mlx5e_rq_param *param)
......@@ -2883,6 +2900,9 @@ static void mlx5e_build_netdev_priv(struct mlx5_core_dev *mdev,
struct mlx5e_priv *priv = netdev_priv(netdev);
u32 link_speed = 0;
u32 pci_bw = 0;
u8 cq_period_mode = MLX5_CAP_GEN(mdev, cq_period_start_from_cqe) ?
MLX5_CQ_PERIOD_MODE_START_FROM_CQE :
MLX5_CQ_PERIOD_MODE_START_FROM_EQE;
priv->params.log_sq_size =
MLX5E_PARAMS_DEFAULT_LOG_SQ_SIZE;
......@@ -2929,8 +2949,8 @@ static void mlx5e_build_netdev_priv(struct mlx5_core_dev *mdev,
priv->params.min_rx_wqes = mlx5_min_rx_wqes(priv->params.rq_wq_type,
BIT(priv->params.log_rq_size));
mlx5e_set_rx_cq_mode_params(&priv->params,
MLX5_CQ_PERIOD_MODE_START_FROM_EQE);
priv->params.rx_am_enabled = MLX5_CAP_GEN(mdev, cq_moderation);
mlx5e_set_rx_cq_mode_params(&priv->params, cq_period_mode);
priv->params.tx_cq_moderation.usec =
MLX5E_PARAMS_DEFAULT_TX_CQ_MODERATION_USEC;
......
/*
* Copyright (c) 2016, Mellanox Technologies. All rights reserved.
*
* This software is available to you under a choice of one of two
* licenses. You may choose to be licensed under the terms of the GNU
* General Public License (GPL) Version 2, available from the file
* COPYING in the main directory of this source tree, or the
* OpenIB.org BSD license below:
*
* Redistribution and use in source and binary forms, with or
* without modification, are permitted provided that the following
* conditions are met:
*
* - Redistributions of source code must retain the above
* copyright notice, this list of conditions and the following
* disclaimer.
*
* - Redistributions in binary form must reproduce the above
* copyright notice, this list of conditions and the following
* disclaimer in the documentation and/or other materials
* provided with the distribution.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
* EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
* MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
* NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
* BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
* ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
* CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
#include "en.h"
/* Adaptive moderation profiles */
#define MLX5E_AM_DEFAULT_RX_CQ_MODERATION_PKTS_FROM_EQE 256
#define MLX5E_RX_AM_DEF_PROFILE_CQE 1
#define MLX5E_RX_AM_DEF_PROFILE_EQE 1
#define MLX5E_PARAMS_AM_NUM_PROFILES 5
/* All profiles sizes must be MLX5E_PARAMS_AM_NUM_PROFILES */
#define MLX5_AM_EQE_PROFILES { \
{1, MLX5E_AM_DEFAULT_RX_CQ_MODERATION_PKTS_FROM_EQE}, \
{8, MLX5E_AM_DEFAULT_RX_CQ_MODERATION_PKTS_FROM_EQE}, \
{64, MLX5E_AM_DEFAULT_RX_CQ_MODERATION_PKTS_FROM_EQE}, \
{128, MLX5E_AM_DEFAULT_RX_CQ_MODERATION_PKTS_FROM_EQE}, \
{256, MLX5E_AM_DEFAULT_RX_CQ_MODERATION_PKTS_FROM_EQE}, \
}
#define MLX5_AM_CQE_PROFILES { \
{2, 256}, \
{8, 128}, \
{16, 64}, \
{32, 64}, \
{64, 64} \
}
static const struct mlx5e_cq_moder
profile[MLX5_CQ_PERIOD_NUM_MODES][MLX5E_PARAMS_AM_NUM_PROFILES] = {
MLX5_AM_EQE_PROFILES,
MLX5_AM_CQE_PROFILES,
};
static inline struct mlx5e_cq_moder mlx5e_am_get_profile(u8 cq_period_mode, int ix)
{
return profile[cq_period_mode][ix];
}
struct mlx5e_cq_moder mlx5e_am_get_def_profile(u8 rx_cq_period_mode)
{
int default_profile_ix;
if (rx_cq_period_mode == MLX5_CQ_PERIOD_MODE_START_FROM_CQE)
default_profile_ix = MLX5E_RX_AM_DEF_PROFILE_CQE;
else /* MLX5_CQ_PERIOD_MODE_START_FROM_EQE */
default_profile_ix = MLX5E_RX_AM_DEF_PROFILE_EQE;
return profile[rx_cq_period_mode][default_profile_ix];
}
/* Adaptive moderation logic */
enum {
MLX5E_AM_START_MEASURE,
MLX5E_AM_MEASURE_IN_PROGRESS,
MLX5E_AM_APPLY_NEW_PROFILE,
};
enum {
MLX5E_AM_PARKING_ON_TOP,
MLX5E_AM_PARKING_TIRED,
MLX5E_AM_GOING_RIGHT,
MLX5E_AM_GOING_LEFT,
};
enum {
MLX5E_AM_STATS_WORSE,
MLX5E_AM_STATS_SAME,
MLX5E_AM_STATS_BETTER,
};
enum {
MLX5E_AM_STEPPED,
MLX5E_AM_TOO_TIRED,
MLX5E_AM_ON_EDGE,
};
static bool mlx5e_am_on_top(struct mlx5e_rx_am *am)
{
switch (am->tune_state) {
case MLX5E_AM_PARKING_ON_TOP:
case MLX5E_AM_PARKING_TIRED:
WARN_ONCE(true, "mlx5e_am_on_top: PARKING\n");
return true;
case MLX5E_AM_GOING_RIGHT:
return (am->steps_left > 1) && (am->steps_right == 1);
default: /* MLX5E_AM_GOING_LEFT */
return (am->steps_right > 1) && (am->steps_left == 1);
}
}
static void mlx5e_am_turn(struct mlx5e_rx_am *am)
{
switch (am->tune_state) {
case MLX5E_AM_PARKING_ON_TOP:
case MLX5E_AM_PARKING_TIRED:
WARN_ONCE(true, "mlx5e_am_turn: PARKING\n");
break;
case MLX5E_AM_GOING_RIGHT:
am->tune_state = MLX5E_AM_GOING_LEFT;
am->steps_left = 0;
break;
case MLX5E_AM_GOING_LEFT:
am->tune_state = MLX5E_AM_GOING_RIGHT;
am->steps_right = 0;
break;
}
}
static int mlx5e_am_step(struct mlx5e_rx_am *am)
{
if (am->tired == (MLX5E_PARAMS_AM_NUM_PROFILES * 2))
return MLX5E_AM_TOO_TIRED;
switch (am->tune_state) {
case MLX5E_AM_PARKING_ON_TOP:
case MLX5E_AM_PARKING_TIRED:
WARN_ONCE(true, "mlx5e_am_step: PARKING\n");
break;
case MLX5E_AM_GOING_RIGHT:
if (am->profile_ix == (MLX5E_PARAMS_AM_NUM_PROFILES - 1))
return MLX5E_AM_ON_EDGE;
am->profile_ix++;
am->steps_right++;
break;
case MLX5E_AM_GOING_LEFT:
if (am->profile_ix == 0)
return MLX5E_AM_ON_EDGE;
am->profile_ix--;
am->steps_left++;
break;
}
am->tired++;
return MLX5E_AM_STEPPED;
}
static void mlx5e_am_park_on_top(struct mlx5e_rx_am *am)
{
am->steps_right = 0;
am->steps_left = 0;
am->tired = 0;
am->tune_state = MLX5E_AM_PARKING_ON_TOP;
}
static void mlx5e_am_park_tired(struct mlx5e_rx_am *am)
{
am->steps_right = 0;
am->steps_left = 0;
am->tune_state = MLX5E_AM_PARKING_TIRED;
}
static void mlx5e_am_exit_parking(struct mlx5e_rx_am *am)
{
am->tune_state = am->profile_ix ? MLX5E_AM_GOING_LEFT :
MLX5E_AM_GOING_RIGHT;
mlx5e_am_step(am);
}
static int mlx5e_am_stats_compare(struct mlx5e_rx_am_stats *curr,
struct mlx5e_rx_am_stats *prev)
{
int diff;
if (!prev->ppms)
return curr->ppms ? MLX5E_AM_STATS_BETTER :
MLX5E_AM_STATS_SAME;
diff = curr->ppms - prev->ppms;
if (((100 * abs(diff)) / prev->ppms) > 10) /* more than 10% diff */
return (diff > 0) ? MLX5E_AM_STATS_BETTER :
MLX5E_AM_STATS_WORSE;
if (!prev->epms)
return curr->epms ? MLX5E_AM_STATS_WORSE :
MLX5E_AM_STATS_SAME;
diff = curr->epms - prev->epms;
if (((100 * abs(diff)) / prev->epms) > 10) /* more than 10% diff */
return (diff < 0) ? MLX5E_AM_STATS_BETTER :
MLX5E_AM_STATS_WORSE;
return MLX5E_AM_STATS_SAME;
}
static bool mlx5e_am_decision(struct mlx5e_rx_am_stats *curr_stats,
struct mlx5e_rx_am *am)
{
int prev_state = am->tune_state;
int prev_ix = am->profile_ix;
int stats_res;
int step_res;
switch (am->tune_state) {
case MLX5E_AM_PARKING_ON_TOP:
stats_res = mlx5e_am_stats_compare(curr_stats, &am->prev_stats);
if (stats_res != MLX5E_AM_STATS_SAME)
mlx5e_am_exit_parking(am);
break;
case MLX5E_AM_PARKING_TIRED:
am->tired--;
if (!am->tired)
mlx5e_am_exit_parking(am);
break;
case MLX5E_AM_GOING_RIGHT:
case MLX5E_AM_GOING_LEFT:
stats_res = mlx5e_am_stats_compare(curr_stats, &am->prev_stats);
if (stats_res != MLX5E_AM_STATS_BETTER)
mlx5e_am_turn(am);
if (mlx5e_am_on_top(am)) {
mlx5e_am_park_on_top(am);
break;
}
step_res = mlx5e_am_step(am);
switch (step_res) {
case MLX5E_AM_ON_EDGE:
mlx5e_am_park_on_top(am);
break;
case MLX5E_AM_TOO_TIRED:
mlx5e_am_park_tired(am);
break;
}
break;
}
if ((prev_state != MLX5E_AM_PARKING_ON_TOP) ||
(am->tune_state != MLX5E_AM_PARKING_ON_TOP))
am->prev_stats = *curr_stats;
return am->profile_ix != prev_ix;
}
static void mlx5e_am_sample(struct mlx5e_rq *rq,
struct mlx5e_rx_am_sample *s)
{
s->time = ktime_get();
s->pkt_ctr = rq->stats.packets;
s->event_ctr = rq->cq.event_ctr;
}
#define MLX5E_AM_NEVENTS 64
static void mlx5e_am_calc_stats(struct mlx5e_rx_am_sample *start,
struct mlx5e_rx_am_sample *end,
struct mlx5e_rx_am_stats *curr_stats)
{
/* u32 holds up to 71 minutes, should be enough */
u32 delta_us = ktime_us_delta(end->time, start->time);
unsigned int npkts = end->pkt_ctr - start->pkt_ctr;
if (!delta_us) {
WARN_ONCE(true, "mlx5e_am_calc_stats: delta_us=0\n");
return;
}
curr_stats->ppms = (npkts * USEC_PER_MSEC) / delta_us;
curr_stats->epms = (MLX5E_AM_NEVENTS * USEC_PER_MSEC) / delta_us;
}
void mlx5e_rx_am_work(struct work_struct *work)
{
struct mlx5e_rx_am *am = container_of(work, struct mlx5e_rx_am,
work);
struct mlx5e_rq *rq = container_of(am, struct mlx5e_rq, am);
struct mlx5e_cq_moder cur_profile = profile[am->mode][am->profile_ix];
mlx5_core_modify_cq_moderation(rq->priv->mdev, &rq->cq.mcq,
cur_profile.usec, cur_profile.pkts);
am->state = MLX5E_AM_START_MEASURE;
}
void mlx5e_rx_am(struct mlx5e_rq *rq)
{
struct mlx5e_rx_am *am = &rq->am;
struct mlx5e_rx_am_sample end_sample;
struct mlx5e_rx_am_stats curr_stats;
u16 nevents;
switch (am->state) {
case MLX5E_AM_MEASURE_IN_PROGRESS:
nevents = rq->cq.event_ctr - am->start_sample.event_ctr;
if (nevents < MLX5E_AM_NEVENTS)
break;
mlx5e_am_sample(rq, &end_sample);
mlx5e_am_calc_stats(&am->start_sample, &end_sample,
&curr_stats);
if (mlx5e_am_decision(&curr_stats, am)) {
am->state = MLX5E_AM_APPLY_NEW_PROFILE;
schedule_work(&am->work);
break;
}
/* fall through */
case MLX5E_AM_START_MEASURE:
mlx5e_am_sample(rq, &am->start_sample);
am->state = MLX5E_AM_MEASURE_IN_PROGRESS;
break;
case MLX5E_AM_APPLY_NEW_PROFILE:
break;
}
}
......@@ -136,6 +136,10 @@ int mlx5e_napi_poll(struct napi_struct *napi, int budget)
for (i = 0; i < c->num_tc; i++)
mlx5e_cq_arm(&c->sq[i].cq);
if (test_bit(MLX5E_RQ_STATE_AM, &c->rq.state))
mlx5e_rx_am(&c->rq);
mlx5e_cq_arm(&c->rq.cq);
mlx5e_cq_arm(&c->icosq.cq);
......@@ -146,6 +150,7 @@ void mlx5e_completion_event(struct mlx5_core_cq *mcq)
{
struct mlx5e_cq *cq = container_of(mcq, struct mlx5e_cq, mcq);
cq->event_ctr++;
set_bit(MLX5E_CHANNEL_NAPI_SCHED, &cq->channel->flags);
napi_schedule(cq->napi);
}
......
Supports Markdown
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment