]> git.proxmox.com Git - mirror_ubuntu-bionic-kernel.git/commitdiff
UBUNTU: linux/dim: import DIMLIB (lib/dim/)
authorKamal Mostafa <kamal@canonical.com>
Thu, 7 Nov 2019 17:52:08 +0000 (09:52 -0800)
committerStefan Bader <stefan.bader@canonical.com>
Thu, 28 Nov 2019 15:44:25 +0000 (16:44 +0100)
BugLink: https://bugs.launchpad.net/bugs/1852637
Import DIMLIB (lib/dim/) from mainline v5.3 as of
commit f8be17b81d44 ("lib/dim: Fix -Wunused-const-variable warnings").

Signed-off-by: Kamal Mostafa <kamal@canonical.com>
Acked-by: Khaled Elmously <khalid.elmously@canonical.com>
Acked-by: Andrea Righi <andrea.righi@canonical.com>
Signed-off-by: Stefan Bader <stefan.bader@canonical.com>
include/linux/dim.h [new file with mode: 0644]
lib/Kconfig
lib/Makefile
lib/dim/Makefile [new file with mode: 0644]
lib/dim/dim.c [new file with mode: 0644]
lib/dim/net_dim.c [new file with mode: 0644]
lib/dim/rdma_dim.c [new file with mode: 0644]

diff --git a/include/linux/dim.h b/include/linux/dim.h
new file mode 100644 (file)
index 0000000..4b42b77
--- /dev/null
@@ -0,0 +1,346 @@
+/* SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB */
+/* Copyright (c) 2019 Mellanox Technologies. */
+
+#ifndef DIM_H
+#define DIM_H
+
+#include <linux/module.h>
+
+/**
+ * Number of events between DIM iterations.
+ * Causes a moderation of the algorithm run.
+ */
+#define DIM_NEVENTS 64
+
+/**
+ * Is a difference between values justifies taking an action.
+ * We consider 10% difference as significant.
+ */
+#define IS_SIGNIFICANT_DIFF(val, ref) \
+       (((100UL * abs((val) - (ref))) / (ref)) > 10)
+
+/**
+ * Calculate the gap between two values.
+ * Take wrap-around and variable size into consideration.
+ */
+#define BIT_GAP(bits, end, start) ((((end) - (start)) + BIT_ULL(bits)) \
+               & (BIT_ULL(bits) - 1))
+
+/**
+ * Structure for CQ moderation values.
+ * Used for communications between DIM and its consumer.
+ *
+ * @usec: CQ timer suggestion (by DIM)
+ * @pkts: CQ packet counter suggestion (by DIM)
+ * @cq_period_mode: CQ priod count mode (from CQE/EQE)
+ */
+struct dim_cq_moder {
+       u16 usec;
+       u16 pkts;
+       u16 comps;
+       u8 cq_period_mode;
+};
+
+/**
+ * Structure for DIM sample data.
+ * Used for communications between DIM and its consumer.
+ *
+ * @time: Sample timestamp
+ * @pkt_ctr: Number of packets
+ * @byte_ctr: Number of bytes
+ * @event_ctr: Number of events
+ */
+struct dim_sample {
+       ktime_t time;
+       u32 pkt_ctr;
+       u32 byte_ctr;
+       u16 event_ctr;
+       u32 comp_ctr;
+};
+
+/**
+ * Structure for DIM stats.
+ * Used for holding current measured rates.
+ *
+ * @ppms: Packets per msec
+ * @bpms: Bytes per msec
+ * @epms: Events per msec
+ */
+struct dim_stats {
+       int ppms; /* packets per msec */
+       int bpms; /* bytes per msec */
+       int epms; /* events per msec */
+       int cpms; /* completions per msec */
+       int cpe_ratio; /* ratio of completions to events */
+};
+
+/**
+ * Main structure for dynamic interrupt moderation (DIM).
+ * Used for holding all information about a specific DIM instance.
+ *
+ * @state: Algorithm state (see below)
+ * @prev_stats: Measured rates from previous iteration (for comparison)
+ * @start_sample: Sampled data at start of current iteration
+ * @work: Work to perform on action required
+ * @priv: A pointer to the struct that points to dim
+ * @profile_ix: Current moderation profile
+ * @mode: CQ period count mode
+ * @tune_state: Algorithm tuning state (see below)
+ * @steps_right: Number of steps taken towards higher moderation
+ * @steps_left: Number of steps taken towards lower moderation
+ * @tired: Parking depth counter
+ */
+struct dim {
+       u8 state;
+       struct dim_stats prev_stats;
+       struct dim_sample start_sample;
+       struct dim_sample measuring_sample;
+       struct work_struct work;
+       void *priv;
+       u8 profile_ix;
+       u8 mode;
+       u8 tune_state;
+       u8 steps_right;
+       u8 steps_left;
+       u8 tired;
+};
+
+/**
+ * enum dim_cq_period_mode
+ *
+ * These are the modes for CQ period count.
+ *
+ * @DIM_CQ_PERIOD_MODE_START_FROM_EQE: Start counting from EQE
+ * @DIM_CQ_PERIOD_MODE_START_FROM_CQE: Start counting from CQE (implies timer reset)
+ * @DIM_CQ_PERIOD_NUM_MODES: Number of modes
+ */
+enum {
+       DIM_CQ_PERIOD_MODE_START_FROM_EQE = 0x0,
+       DIM_CQ_PERIOD_MODE_START_FROM_CQE = 0x1,
+       DIM_CQ_PERIOD_NUM_MODES
+};
+
+/**
+ * enum dim_state
+ *
+ * These are the DIM algorithm states.
+ * These will determine if the algorithm is in a valid state to start an iteration.
+ *
+ * @DIM_START_MEASURE: This is the first iteration (also after applying a new profile)
+ * @DIM_MEASURE_IN_PROGRESS: Algorithm is already in progress - check if
+ * need to perform an action
+ * @DIM_APPLY_NEW_PROFILE: DIM consumer is currently applying a profile - no need to measure
+ */
+enum {
+       DIM_START_MEASURE,
+       DIM_MEASURE_IN_PROGRESS,
+       DIM_APPLY_NEW_PROFILE,
+};
+
+/**
+ * enum dim_tune_state
+ *
+ * These are the DIM algorithm tune states.
+ * These will determine which action the algorithm should perform.
+ *
+ * @DIM_PARKING_ON_TOP: Algorithm found a local top point - exit on significant difference
+ * @DIM_PARKING_TIRED: Algorithm found a deep top point - don't exit if tired > 0
+ * @DIM_GOING_RIGHT: Algorithm is currently trying higher moderation levels
+ * @DIM_GOING_LEFT: Algorithm is currently trying lower moderation levels
+ */
+enum {
+       DIM_PARKING_ON_TOP,
+       DIM_PARKING_TIRED,
+       DIM_GOING_RIGHT,
+       DIM_GOING_LEFT,
+};
+
+/**
+ * enum dim_stats_state
+ *
+ * These are the DIM algorithm statistics states.
+ * These will determine the verdict of current iteration.
+ *
+ * @DIM_STATS_WORSE: Current iteration shows worse performance than before
+ * @DIM_STATS_WORSE: Current iteration shows same performance than before
+ * @DIM_STATS_WORSE: Current iteration shows better performance than before
+ */
+enum {
+       DIM_STATS_WORSE,
+       DIM_STATS_SAME,
+       DIM_STATS_BETTER,
+};
+
+/**
+ * enum dim_step_result
+ *
+ * These are the DIM algorithm step results.
+ * These describe the result of a step.
+ *
+ * @DIM_STEPPED: Performed a regular step
+ * @DIM_TOO_TIRED: Same kind of step was done multiple times - should go to
+ * tired parking
+ * @DIM_ON_EDGE: Stepped to the most left/right profile
+ */
+enum {
+       DIM_STEPPED,
+       DIM_TOO_TIRED,
+       DIM_ON_EDGE,
+};
+
+/**
+ *     dim_on_top - check if current state is a good place to stop (top location)
+ *     @dim: DIM context
+ *
+ * Check if current profile is a good place to park at.
+ * This will result in reducing the DIM checks frequency as we assume we
+ * shouldn't probably change profiles, unless traffic pattern wasn't changed.
+ */
+bool dim_on_top(struct dim *dim);
+
+/**
+ *     dim_turn - change profile alterning direction
+ *     @dim: DIM context
+ *
+ * Go left if we were going right and vice-versa.
+ * Do nothing if currently parking.
+ */
+void dim_turn(struct dim *dim);
+
+/**
+ *     dim_park_on_top - enter a parking state on a top location
+ *     @dim: DIM context
+ *
+ * Enter parking state.
+ * Clear all movement history.
+ */
+void dim_park_on_top(struct dim *dim);
+
+/**
+ *     dim_park_tired - enter a tired parking state
+ *     @dim: DIM context
+ *
+ * Enter parking state.
+ * Clear all movement history and cause DIM checks frequency to reduce.
+ */
+void dim_park_tired(struct dim *dim);
+
+/**
+ *     dim_calc_stats - calculate the difference between two samples
+ *     @start: start sample
+ *     @end: end sample
+ *     @curr_stats: delta between samples
+ *
+ * Calculate the delta between two samples (in data rates).
+ * Takes into consideration counter wrap-around.
+ */
+void dim_calc_stats(struct dim_sample *start, struct dim_sample *end,
+                   struct dim_stats *curr_stats);
+
+/**
+ *     dim_update_sample - set a sample's fields with give values
+ *     @event_ctr: number of events to set
+ *     @packets: number of packets to set
+ *     @bytes: number of bytes to set
+ *     @s: DIM sample
+ */
+static inline void
+dim_update_sample(u16 event_ctr, u64 packets, u64 bytes, struct dim_sample *s)
+{
+       s->time      = ktime_get();
+       s->pkt_ctr   = packets;
+       s->byte_ctr  = bytes;
+       s->event_ctr = event_ctr;
+}
+
+/**
+ *     dim_update_sample_with_comps - set a sample's fields with given
+ *     values including the completion parameter
+ *     @event_ctr: number of events to set
+ *     @packets: number of packets to set
+ *     @bytes: number of bytes to set
+ *     @comps: number of completions to set
+ *     @s: DIM sample
+ */
+static inline void
+dim_update_sample_with_comps(u16 event_ctr, u64 packets, u64 bytes, u64 comps,
+                            struct dim_sample *s)
+{
+       dim_update_sample(event_ctr, packets, bytes, s);
+       s->comp_ctr = comps;
+}
+
+/* Net DIM */
+
+/**
+ *     net_dim_get_rx_moderation - provide a CQ moderation object for the given RX profile
+ *     @cq_period_mode: CQ period mode
+ *     @ix: Profile index
+ */
+struct dim_cq_moder net_dim_get_rx_moderation(u8 cq_period_mode, int ix);
+
+/**
+ *     net_dim_get_def_rx_moderation - provide the default RX moderation
+ *     @cq_period_mode: CQ period mode
+ */
+struct dim_cq_moder net_dim_get_def_rx_moderation(u8 cq_period_mode);
+
+/**
+ *     net_dim_get_tx_moderation - provide a CQ moderation object for the given TX profile
+ *     @cq_period_mode: CQ period mode
+ *     @ix: Profile index
+ */
+struct dim_cq_moder net_dim_get_tx_moderation(u8 cq_period_mode, int ix);
+
+/**
+ *     net_dim_get_def_tx_moderation - provide the default TX moderation
+ *     @cq_period_mode: CQ period mode
+ */
+struct dim_cq_moder net_dim_get_def_tx_moderation(u8 cq_period_mode);
+
+/**
+ *     net_dim - main DIM algorithm entry point
+ *     @dim: DIM instance information
+ *     @end_sample: Current data measurement
+ *
+ * Called by the consumer.
+ * This is the main logic of the algorithm, where data is processed in order to decide on next
+ * required action.
+ */
+void net_dim(struct dim *dim, struct dim_sample end_sample);
+
+/* RDMA DIM */
+
+/*
+ * RDMA DIM profile:
+ * profile size must be of RDMA_DIM_PARAMS_NUM_PROFILES.
+ */
+#define RDMA_DIM_PARAMS_NUM_PROFILES 9
+#define RDMA_DIM_START_PROFILE 0
+
+static const struct dim_cq_moder
+rdma_dim_prof[RDMA_DIM_PARAMS_NUM_PROFILES] = {
+       {1,   0, 1,  0},
+       {1,   0, 4,  0},
+       {2,   0, 4,  0},
+       {2,   0, 8,  0},
+       {4,   0, 8,  0},
+       {16,  0, 8,  0},
+       {16,  0, 16, 0},
+       {32,  0, 16, 0},
+       {32,  0, 32, 0},
+};
+
+/**
+ * rdma_dim - Runs the adaptive moderation.
+ * @dim: The moderation struct.
+ * @completions: The number of completions collected in this round.
+ *
+ * Each call to rdma_dim takes the latest amount of completions that
+ * have been collected and counts them as a new event.
+ * Once enough events have been collected the algorithm decides a new
+ * moderation level.
+ */
+void rdma_dim(struct dim *dim, u64 completions);
+
+#endif /* DIM_H */
index 5bdfbb270889b08a79ffbef053dacd3f0dcacbb7..5ced0f0e2d92dc2cce1a94ff9c6a8aa4f264d7e6 100644 (file)
@@ -543,6 +543,14 @@ config SIGNATURE
          Digital signature verification. Currently only RSA is supported.
          Implementation is done using GnuPG MPI library
 
+config DIMLIB
+       bool "DIM library"
+       default y
+       help
+         Dynamic Interrupt Moderation library.
+         Implements an algorithm for dynamically change CQ modertion values
+         according to run time performance.
+
 #
 # libfdt files, only selected if needed.
 #
index 62ae814c1f91811b5cc61b9a2aae4d395bca466a..f3ed869e35738ebca0e63369b276885f7d6005c1 100644 (file)
@@ -192,6 +192,7 @@ obj-$(CONFIG_GLOB) += glob.o
 obj-$(CONFIG_GLOB_SELFTEST) += globtest.o
 
 obj-$(CONFIG_MPILIB) += mpi/
+obj-$(CONFIG_DIMLIB) += dim/
 obj-$(CONFIG_SIGNATURE) += digsig.o
 
 lib-$(CONFIG_CLZ_TAB) += clz_tab.o
diff --git a/lib/dim/Makefile b/lib/dim/Makefile
new file mode 100644 (file)
index 0000000..1d6858a
--- /dev/null
@@ -0,0 +1,7 @@
+#
+# DIM Dynamic Interrupt Moderation library
+#
+
+obj-$(CONFIG_DIMLIB) += dim.o
+
+dim-y := dim.o net_dim.o rdma_dim.o
diff --git a/lib/dim/dim.c b/lib/dim/dim.c
new file mode 100644 (file)
index 0000000..38045d6
--- /dev/null
@@ -0,0 +1,83 @@
+// SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
+/*
+ * Copyright (c) 2019, Mellanox Technologies inc.  All rights reserved.
+ */
+
+#include <linux/dim.h>
+
+bool dim_on_top(struct dim *dim)
+{
+       switch (dim->tune_state) {
+       case DIM_PARKING_ON_TOP:
+       case DIM_PARKING_TIRED:
+               return true;
+       case DIM_GOING_RIGHT:
+               return (dim->steps_left > 1) && (dim->steps_right == 1);
+       default: /* DIM_GOING_LEFT */
+               return (dim->steps_right > 1) && (dim->steps_left == 1);
+       }
+}
+EXPORT_SYMBOL(dim_on_top);
+
+void dim_turn(struct dim *dim)
+{
+       switch (dim->tune_state) {
+       case DIM_PARKING_ON_TOP:
+       case DIM_PARKING_TIRED:
+               break;
+       case DIM_GOING_RIGHT:
+               dim->tune_state = DIM_GOING_LEFT;
+               dim->steps_left = 0;
+               break;
+       case DIM_GOING_LEFT:
+               dim->tune_state = DIM_GOING_RIGHT;
+               dim->steps_right = 0;
+               break;
+       }
+}
+EXPORT_SYMBOL(dim_turn);
+
+void dim_park_on_top(struct dim *dim)
+{
+       dim->steps_right  = 0;
+       dim->steps_left   = 0;
+       dim->tired        = 0;
+       dim->tune_state   = DIM_PARKING_ON_TOP;
+}
+EXPORT_SYMBOL(dim_park_on_top);
+
+void dim_park_tired(struct dim *dim)
+{
+       dim->steps_right  = 0;
+       dim->steps_left   = 0;
+       dim->tune_state   = DIM_PARKING_TIRED;
+}
+EXPORT_SYMBOL(dim_park_tired);
+
+void dim_calc_stats(struct dim_sample *start, struct dim_sample *end,
+                   struct dim_stats *curr_stats)
+{
+       /* u32 holds up to 71 minutes, should be enough */
+       u32 delta_us = ktime_us_delta(end->time, start->time);
+       u32 npkts = BIT_GAP(BITS_PER_TYPE(u32), end->pkt_ctr, start->pkt_ctr);
+       u32 nbytes = BIT_GAP(BITS_PER_TYPE(u32), end->byte_ctr,
+                            start->byte_ctr);
+       u32 ncomps = BIT_GAP(BITS_PER_TYPE(u32), end->comp_ctr,
+                            start->comp_ctr);
+
+       if (!delta_us)
+               return;
+
+       curr_stats->ppms = DIV_ROUND_UP(npkts * USEC_PER_MSEC, delta_us);
+       curr_stats->bpms = DIV_ROUND_UP(nbytes * USEC_PER_MSEC, delta_us);
+       curr_stats->epms = DIV_ROUND_UP(DIM_NEVENTS * USEC_PER_MSEC,
+                                       delta_us);
+       curr_stats->cpms = DIV_ROUND_UP(ncomps * USEC_PER_MSEC, delta_us);
+       if (curr_stats->epms != 0)
+               curr_stats->cpe_ratio = DIV_ROUND_DOWN_ULL(
+                       curr_stats->cpms * 100, curr_stats->epms);
+       else
+               curr_stats->cpe_ratio = 0;
+
+}
+EXPORT_SYMBOL(dim_calc_stats);
diff --git a/lib/dim/net_dim.c b/lib/dim/net_dim.c
new file mode 100644 (file)
index 0000000..a4db51c
--- /dev/null
@@ -0,0 +1,246 @@
+// SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
+/*
+ * Copyright (c) 2018, Mellanox Technologies inc.  All rights reserved.
+ */
+
+#include <linux/dim.h>
+
+/*
+ * Net DIM profiles:
+ *        There are different set of profiles for each CQ period mode.
+ *        There are different set of profiles for RX/TX CQs.
+ *        Each profile size must be of NET_DIM_PARAMS_NUM_PROFILES
+ */
+#define NET_DIM_PARAMS_NUM_PROFILES 5
+#define NET_DIM_DEFAULT_RX_CQ_MODERATION_PKTS_FROM_EQE 256
+#define NET_DIM_DEFAULT_TX_CQ_MODERATION_PKTS_FROM_EQE 128
+#define NET_DIM_DEF_PROFILE_CQE 1
+#define NET_DIM_DEF_PROFILE_EQE 1
+
+#define NET_DIM_RX_EQE_PROFILES { \
+       {1,   NET_DIM_DEFAULT_RX_CQ_MODERATION_PKTS_FROM_EQE}, \
+       {8,   NET_DIM_DEFAULT_RX_CQ_MODERATION_PKTS_FROM_EQE}, \
+       {64,  NET_DIM_DEFAULT_RX_CQ_MODERATION_PKTS_FROM_EQE}, \
+       {128, NET_DIM_DEFAULT_RX_CQ_MODERATION_PKTS_FROM_EQE}, \
+       {256, NET_DIM_DEFAULT_RX_CQ_MODERATION_PKTS_FROM_EQE}, \
+}
+
+#define NET_DIM_RX_CQE_PROFILES { \
+       {2,  256},             \
+       {8,  128},             \
+       {16, 64},              \
+       {32, 64},              \
+       {64, 64}               \
+}
+
+#define NET_DIM_TX_EQE_PROFILES { \
+       {1,   NET_DIM_DEFAULT_TX_CQ_MODERATION_PKTS_FROM_EQE},  \
+       {8,   NET_DIM_DEFAULT_TX_CQ_MODERATION_PKTS_FROM_EQE},  \
+       {32,  NET_DIM_DEFAULT_TX_CQ_MODERATION_PKTS_FROM_EQE},  \
+       {64,  NET_DIM_DEFAULT_TX_CQ_MODERATION_PKTS_FROM_EQE},  \
+       {128, NET_DIM_DEFAULT_TX_CQ_MODERATION_PKTS_FROM_EQE}   \
+}
+
+#define NET_DIM_TX_CQE_PROFILES { \
+       {5,  128},  \
+       {8,  64},  \
+       {16, 32},  \
+       {32, 32},  \
+       {64, 32}   \
+}
+
+static const struct dim_cq_moder
+rx_profile[DIM_CQ_PERIOD_NUM_MODES][NET_DIM_PARAMS_NUM_PROFILES] = {
+       NET_DIM_RX_EQE_PROFILES,
+       NET_DIM_RX_CQE_PROFILES,
+};
+
+static const struct dim_cq_moder
+tx_profile[DIM_CQ_PERIOD_NUM_MODES][NET_DIM_PARAMS_NUM_PROFILES] = {
+       NET_DIM_TX_EQE_PROFILES,
+       NET_DIM_TX_CQE_PROFILES,
+};
+
+struct dim_cq_moder
+net_dim_get_rx_moderation(u8 cq_period_mode, int ix)
+{
+       struct dim_cq_moder cq_moder = rx_profile[cq_period_mode][ix];
+
+       cq_moder.cq_period_mode = cq_period_mode;
+       return cq_moder;
+}
+EXPORT_SYMBOL(net_dim_get_rx_moderation);
+
+struct dim_cq_moder
+net_dim_get_def_rx_moderation(u8 cq_period_mode)
+{
+       u8 profile_ix = cq_period_mode == DIM_CQ_PERIOD_MODE_START_FROM_CQE ?
+                       NET_DIM_DEF_PROFILE_CQE : NET_DIM_DEF_PROFILE_EQE;
+
+       return net_dim_get_rx_moderation(cq_period_mode, profile_ix);
+}
+EXPORT_SYMBOL(net_dim_get_def_rx_moderation);
+
+struct dim_cq_moder
+net_dim_get_tx_moderation(u8 cq_period_mode, int ix)
+{
+       struct dim_cq_moder cq_moder = tx_profile[cq_period_mode][ix];
+
+       cq_moder.cq_period_mode = cq_period_mode;
+       return cq_moder;
+}
+EXPORT_SYMBOL(net_dim_get_tx_moderation);
+
+struct dim_cq_moder
+net_dim_get_def_tx_moderation(u8 cq_period_mode)
+{
+       u8 profile_ix = cq_period_mode == DIM_CQ_PERIOD_MODE_START_FROM_CQE ?
+                       NET_DIM_DEF_PROFILE_CQE : NET_DIM_DEF_PROFILE_EQE;
+
+       return net_dim_get_tx_moderation(cq_period_mode, profile_ix);
+}
+EXPORT_SYMBOL(net_dim_get_def_tx_moderation);
+
+static int net_dim_step(struct dim *dim)
+{
+       if (dim->tired == (NET_DIM_PARAMS_NUM_PROFILES * 2))
+               return DIM_TOO_TIRED;
+
+       switch (dim->tune_state) {
+       case DIM_PARKING_ON_TOP:
+       case DIM_PARKING_TIRED:
+               break;
+       case DIM_GOING_RIGHT:
+               if (dim->profile_ix == (NET_DIM_PARAMS_NUM_PROFILES - 1))
+                       return DIM_ON_EDGE;
+               dim->profile_ix++;
+               dim->steps_right++;
+               break;
+       case DIM_GOING_LEFT:
+               if (dim->profile_ix == 0)
+                       return DIM_ON_EDGE;
+               dim->profile_ix--;
+               dim->steps_left++;
+               break;
+       }
+
+       dim->tired++;
+       return DIM_STEPPED;
+}
+
+static void net_dim_exit_parking(struct dim *dim)
+{
+       dim->tune_state = dim->profile_ix ? DIM_GOING_LEFT : DIM_GOING_RIGHT;
+       net_dim_step(dim);
+}
+
+static int net_dim_stats_compare(struct dim_stats *curr,
+                                struct dim_stats *prev)
+{
+       if (!prev->bpms)
+               return curr->bpms ? DIM_STATS_BETTER : DIM_STATS_SAME;
+
+       if (IS_SIGNIFICANT_DIFF(curr->bpms, prev->bpms))
+               return (curr->bpms > prev->bpms) ? DIM_STATS_BETTER :
+                                                  DIM_STATS_WORSE;
+
+       if (!prev->ppms)
+               return curr->ppms ? DIM_STATS_BETTER :
+                                   DIM_STATS_SAME;
+
+       if (IS_SIGNIFICANT_DIFF(curr->ppms, prev->ppms))
+               return (curr->ppms > prev->ppms) ? DIM_STATS_BETTER :
+                                                  DIM_STATS_WORSE;
+
+       if (!prev->epms)
+               return DIM_STATS_SAME;
+
+       if (IS_SIGNIFICANT_DIFF(curr->epms, prev->epms))
+               return (curr->epms < prev->epms) ? DIM_STATS_BETTER :
+                                                  DIM_STATS_WORSE;
+
+       return DIM_STATS_SAME;
+}
+
+static bool net_dim_decision(struct dim_stats *curr_stats, struct dim *dim)
+{
+       int prev_state = dim->tune_state;
+       int prev_ix = dim->profile_ix;
+       int stats_res;
+       int step_res;
+
+       switch (dim->tune_state) {
+       case DIM_PARKING_ON_TOP:
+               stats_res = net_dim_stats_compare(curr_stats,
+                                                 &dim->prev_stats);
+               if (stats_res != DIM_STATS_SAME)
+                       net_dim_exit_parking(dim);
+               break;
+
+       case DIM_PARKING_TIRED:
+               dim->tired--;
+               if (!dim->tired)
+                       net_dim_exit_parking(dim);
+               break;
+
+       case DIM_GOING_RIGHT:
+       case DIM_GOING_LEFT:
+               stats_res = net_dim_stats_compare(curr_stats,
+                                                 &dim->prev_stats);
+               if (stats_res != DIM_STATS_BETTER)
+                       dim_turn(dim);
+
+               if (dim_on_top(dim)) {
+                       dim_park_on_top(dim);
+                       break;
+               }
+
+               step_res = net_dim_step(dim);
+               switch (step_res) {
+               case DIM_ON_EDGE:
+                       dim_park_on_top(dim);
+                       break;
+               case DIM_TOO_TIRED:
+                       dim_park_tired(dim);
+                       break;
+               }
+
+               break;
+       }
+
+       if (prev_state != DIM_PARKING_ON_TOP ||
+           dim->tune_state != DIM_PARKING_ON_TOP)
+               dim->prev_stats = *curr_stats;
+
+       return dim->profile_ix != prev_ix;
+}
+
+void net_dim(struct dim *dim, struct dim_sample end_sample)
+{
+       struct dim_stats curr_stats;
+       u16 nevents;
+
+       switch (dim->state) {
+       case DIM_MEASURE_IN_PROGRESS:
+               nevents = BIT_GAP(BITS_PER_TYPE(u16),
+                                 end_sample.event_ctr,
+                                 dim->start_sample.event_ctr);
+               if (nevents < DIM_NEVENTS)
+                       break;
+               dim_calc_stats(&dim->start_sample, &end_sample, &curr_stats);
+               if (net_dim_decision(&curr_stats, dim)) {
+                       dim->state = DIM_APPLY_NEW_PROFILE;
+                       schedule_work(&dim->work);
+                       break;
+               }
+               /* fall through */
+       case DIM_START_MEASURE:
+               dim_update_sample(end_sample.event_ctr, end_sample.pkt_ctr,
+                                 end_sample.byte_ctr, &dim->start_sample);
+               dim->state = DIM_MEASURE_IN_PROGRESS;
+               break;
+       case DIM_APPLY_NEW_PROFILE:
+               break;
+       }
+}
+EXPORT_SYMBOL(net_dim);
diff --git a/lib/dim/rdma_dim.c b/lib/dim/rdma_dim.c
new file mode 100644 (file)
index 0000000..f7e26c7
--- /dev/null
@@ -0,0 +1,108 @@
+// SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
+/*
+ * Copyright (c) 2019, Mellanox Technologies inc.  All rights reserved.
+ */
+
+#include <linux/dim.h>
+
+static int rdma_dim_step(struct dim *dim)
+{
+       if (dim->tune_state == DIM_GOING_RIGHT) {
+               if (dim->profile_ix == (RDMA_DIM_PARAMS_NUM_PROFILES - 1))
+                       return DIM_ON_EDGE;
+               dim->profile_ix++;
+               dim->steps_right++;
+       }
+       if (dim->tune_state == DIM_GOING_LEFT) {
+               if (dim->profile_ix == 0)
+                       return DIM_ON_EDGE;
+               dim->profile_ix--;
+               dim->steps_left++;
+       }
+
+       return DIM_STEPPED;
+}
+
+static int rdma_dim_stats_compare(struct dim_stats *curr,
+                                 struct dim_stats *prev)
+{
+       /* first stat */
+       if (!prev->cpms)
+               return DIM_STATS_SAME;
+
+       if (IS_SIGNIFICANT_DIFF(curr->cpms, prev->cpms))
+               return (curr->cpms > prev->cpms) ? DIM_STATS_BETTER :
+                                               DIM_STATS_WORSE;
+
+       if (IS_SIGNIFICANT_DIFF(curr->cpe_ratio, prev->cpe_ratio))
+               return (curr->cpe_ratio > prev->cpe_ratio) ? DIM_STATS_BETTER :
+                                               DIM_STATS_WORSE;
+
+       return DIM_STATS_SAME;
+}
+
+static bool rdma_dim_decision(struct dim_stats *curr_stats, struct dim *dim)
+{
+       int prev_ix = dim->profile_ix;
+       u8 state = dim->tune_state;
+       int stats_res;
+       int step_res;
+
+       if (state != DIM_PARKING_ON_TOP && state != DIM_PARKING_TIRED) {
+               stats_res = rdma_dim_stats_compare(curr_stats,
+                                                  &dim->prev_stats);
+
+               switch (stats_res) {
+               case DIM_STATS_SAME:
+                       if (curr_stats->cpe_ratio <= 50 * prev_ix)
+                               dim->profile_ix = 0;
+                       break;
+               case DIM_STATS_WORSE:
+                       dim_turn(dim);
+                       /* fall through */
+               case DIM_STATS_BETTER:
+                       step_res = rdma_dim_step(dim);
+                       if (step_res == DIM_ON_EDGE)
+                               dim_turn(dim);
+                       break;
+               }
+       }
+
+       dim->prev_stats = *curr_stats;
+
+       return dim->profile_ix != prev_ix;
+}
+
+void rdma_dim(struct dim *dim, u64 completions)
+{
+       struct dim_sample *curr_sample = &dim->measuring_sample;
+       struct dim_stats curr_stats;
+       u32 nevents;
+
+       dim_update_sample_with_comps(curr_sample->event_ctr + 1, 0, 0,
+                                    curr_sample->comp_ctr + completions,
+                                    &dim->measuring_sample);
+
+       switch (dim->state) {
+       case DIM_MEASURE_IN_PROGRESS:
+               nevents = curr_sample->event_ctr - dim->start_sample.event_ctr;
+               if (nevents < DIM_NEVENTS)
+                       break;
+               dim_calc_stats(&dim->start_sample, curr_sample, &curr_stats);
+               if (rdma_dim_decision(&curr_stats, dim)) {
+                       dim->state = DIM_APPLY_NEW_PROFILE;
+                       schedule_work(&dim->work);
+                       break;
+               }
+               /* fall through */
+       case DIM_START_MEASURE:
+               dim->state = DIM_MEASURE_IN_PROGRESS;
+               dim_update_sample_with_comps(curr_sample->event_ctr, 0, 0,
+                                            curr_sample->comp_ctr,
+                                            &dim->start_sample);
+               break;
+       case DIM_APPLY_NEW_PROFILE:
+               break;
+       }
+}
+EXPORT_SYMBOL(rdma_dim);