1/* SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB */
2/* Copyright (c) 2019 Mellanox Technologies. */
3
4#ifndef DIM_H
5#define DIM_H
6
7#include <linux/bits.h>
8#include <linux/kernel.h>
9#include <linux/module.h>
10#include <linux/types.h>
11#include <linux/workqueue.h>
12
13struct net_device;
14
15/* Number of DIM profiles and period mode. */
16#define NET_DIM_PARAMS_NUM_PROFILES 5
17#define NET_DIM_DEFAULT_RX_CQ_PKTS_FROM_EQE 256
18#define NET_DIM_DEFAULT_TX_CQ_PKTS_FROM_EQE 128
19#define NET_DIM_DEF_PROFILE_CQE 1
20#define NET_DIM_DEF_PROFILE_EQE 1
21
22/*
23 * Number of events between DIM iterations.
24 * Causes a moderation of the algorithm run.
25 */
26#define DIM_NEVENTS 64
27
28/*
29 * Is a difference between values justifies taking an action.
30 * We consider 10% difference as significant.
31 */
32#define IS_SIGNIFICANT_DIFF(val, ref) \
33 ((ref) && (((100UL * abs((val) - (ref))) / (ref)) > 10))
34
35/*
36 * Calculate the gap between two values.
37 * Take wrap-around and variable size into consideration.
38 */
39#define BIT_GAP(bits, end, start) ((((end) - (start)) + BIT_ULL(bits)) \
40 & (BIT_ULL(bits) - 1))
41
42/**
43 * struct dim_cq_moder - Structure for CQ moderation values.
44 * Used for communications between DIM and its consumer.
45 *
46 * @usec: CQ timer suggestion (by DIM)
47 * @pkts: CQ packet counter suggestion (by DIM)
48 * @comps: Completion counter
49 * @cq_period_mode: CQ period count mode (from CQE/EQE)
50 * @rcu: for asynchronous kfree_rcu
51 */
52struct dim_cq_moder {
53 u16 usec;
54 u16 pkts;
55 u16 comps;
56 u8 cq_period_mode;
57 struct rcu_head rcu;
58};
59
60#define DIM_PROFILE_RX BIT(0) /* support rx profile modification */
61#define DIM_PROFILE_TX BIT(1) /* support tx profile modification */
62
63#define DIM_COALESCE_USEC BIT(0) /* support usec field modification */
64#define DIM_COALESCE_PKTS BIT(1) /* support pkts field modification */
65#define DIM_COALESCE_COMPS BIT(2) /* support comps field modification */
66
67/**
68 * struct dim_irq_moder - Structure for irq moderation information.
69 * Used to collect irq moderation related information.
70 *
71 * @profile_flags: DIM_PROFILE_*
72 * @coal_flags: DIM_COALESCE_* for Rx and Tx
73 * @dim_rx_mode: Rx DIM period count mode: CQE or EQE
74 * @dim_tx_mode: Tx DIM period count mode: CQE or EQE
75 * @rx_profile: DIM profile list for Rx
76 * @tx_profile: DIM profile list for Tx
77 * @rx_dim_work: Rx DIM worker scheduled by net_dim()
78 * @tx_dim_work: Tx DIM worker scheduled by net_dim()
79 */
80struct dim_irq_moder {
81 u8 profile_flags;
82 u8 coal_flags;
83 u8 dim_rx_mode;
84 u8 dim_tx_mode;
85 struct dim_cq_moder __rcu *rx_profile;
86 struct dim_cq_moder __rcu *tx_profile;
87 void (*rx_dim_work)(struct work_struct *work);
88 void (*tx_dim_work)(struct work_struct *work);
89};
90
91/**
92 * struct dim_sample - Structure for DIM sample data.
93 * Used for communications between DIM and its consumer.
94 *
95 * @time: Sample timestamp
96 * @pkt_ctr: Number of packets
97 * @byte_ctr: Number of bytes
98 * @event_ctr: Number of events
99 * @comp_ctr: Current completion counter
100 */
101struct dim_sample {
102 ktime_t time;
103 u32 pkt_ctr;
104 u32 byte_ctr;
105 u16 event_ctr;
106 u32 comp_ctr;
107};
108
109/**
110 * struct dim_stats - Structure for DIM stats.
111 * Used for holding current measured rates.
112 *
113 * @ppms: Packets per msec
114 * @bpms: Bytes per msec
115 * @epms: Events per msec
116 * @cpms: Completions per msec
117 * @cpe_ratio: Ratio of completions to events
118 */
119struct dim_stats {
120 int ppms; /* packets per msec */
121 int bpms; /* bytes per msec */
122 int epms; /* events per msec */
123 int cpms; /* completions per msec */
124 int cpe_ratio; /* ratio of completions to events */
125};
126
127/**
128 * struct dim - Main structure for dynamic interrupt moderation (DIM).
129 * Used for holding all information about a specific DIM instance.
130 *
131 * @state: Algorithm state (see below)
132 * @prev_stats: Measured rates from previous iteration (for comparison)
133 * @start_sample: Sampled data at start of current iteration
134 * @measuring_sample: A &dim_sample that is used to update the current events
135 * @work: Work to perform on action required
136 * @priv: A pointer to the struct that points to dim
137 * @profile_ix: Current moderation profile
138 * @mode: CQ period count mode
139 * @tune_state: Algorithm tuning state (see below)
140 * @steps_right: Number of steps taken towards higher moderation
141 * @steps_left: Number of steps taken towards lower moderation
142 * @tired: Parking depth counter
143 */
144struct dim {
145 u8 state;
146 struct dim_stats prev_stats;
147 struct dim_sample start_sample;
148 struct dim_sample measuring_sample;
149 struct work_struct work;
150 void *priv;
151 u8 profile_ix;
152 u8 mode;
153 u8 tune_state;
154 u8 steps_right;
155 u8 steps_left;
156 u8 tired;
157};
158
159/**
160 * enum dim_cq_period_mode - Modes for CQ period count
161 *
162 * @DIM_CQ_PERIOD_MODE_START_FROM_EQE: Start counting from EQE
163 * @DIM_CQ_PERIOD_MODE_START_FROM_CQE: Start counting from CQE (implies timer reset)
164 * @DIM_CQ_PERIOD_NUM_MODES: Number of modes
165 */
166enum dim_cq_period_mode {
167 DIM_CQ_PERIOD_MODE_START_FROM_EQE = 0x0,
168 DIM_CQ_PERIOD_MODE_START_FROM_CQE = 0x1,
169 DIM_CQ_PERIOD_NUM_MODES
170};
171
172/**
173 * enum dim_state - DIM algorithm states
174 *
175 * These will determine if the algorithm is in a valid state to start an iteration.
176 *
177 * @DIM_START_MEASURE: This is the first iteration (also after applying a new profile)
178 * @DIM_MEASURE_IN_PROGRESS: Algorithm is already in progress - check if
179 * need to perform an action
180 * @DIM_APPLY_NEW_PROFILE: DIM consumer is currently applying a profile - no need to measure
181 */
182enum dim_state {
183 DIM_START_MEASURE,
184 DIM_MEASURE_IN_PROGRESS,
185 DIM_APPLY_NEW_PROFILE,
186};
187
188/**
189 * enum dim_tune_state - DIM algorithm tune states
190 *
191 * These will determine which action the algorithm should perform.
192 *
193 * @DIM_PARKING_ON_TOP: Algorithm found a local top point - exit on significant difference
194 * @DIM_PARKING_TIRED: Algorithm found a deep top point - don't exit if tired > 0
195 * @DIM_GOING_RIGHT: Algorithm is currently trying higher moderation levels
196 * @DIM_GOING_LEFT: Algorithm is currently trying lower moderation levels
197 */
198enum dim_tune_state {
199 DIM_PARKING_ON_TOP,
200 DIM_PARKING_TIRED,
201 DIM_GOING_RIGHT,
202 DIM_GOING_LEFT,
203};
204
205/**
206 * enum dim_stats_state - DIM algorithm statistics states
207 *
208 * These will determine the verdict of current iteration.
209 *
210 * @DIM_STATS_WORSE: Current iteration shows worse performance than before
211 * @DIM_STATS_SAME: Current iteration shows same performance than before
212 * @DIM_STATS_BETTER: Current iteration shows better performance than before
213 */
214enum dim_stats_state {
215 DIM_STATS_WORSE,
216 DIM_STATS_SAME,
217 DIM_STATS_BETTER,
218};
219
220/**
221 * enum dim_step_result - DIM algorithm step results
222 *
223 * These describe the result of a step.
224 *
225 * @DIM_STEPPED: Performed a regular step
226 * @DIM_TOO_TIRED: Same kind of step was done multiple times - should go to
227 * tired parking
228 * @DIM_ON_EDGE: Stepped to the most left/right profile
229 */
230enum dim_step_result {
231 DIM_STEPPED,
232 DIM_TOO_TIRED,
233 DIM_ON_EDGE,
234};
235
236/**
237 * net_dim_init_irq_moder - collect information to initialize irq moderation
238 * @dev: target network device
239 * @profile_flags: Rx or Tx profile modification capability
240 * @coal_flags: irq moderation params flags
241 * @rx_mode: CQ period mode for Rx
242 * @tx_mode: CQ period mode for Tx
243 * @rx_dim_work: Rx worker called after dim decision
244 * @tx_dim_work: Tx worker called after dim decision
245 *
246 * Return: 0 on success or a negative error code.
247 */
248int net_dim_init_irq_moder(struct net_device *dev, u8 profile_flags,
249 u8 coal_flags, u8 rx_mode, u8 tx_mode,
250 void (*rx_dim_work)(struct work_struct *work),
251 void (*tx_dim_work)(struct work_struct *work));
252
253/**
254 * net_dim_free_irq_moder - free fields for irq moderation
255 * @dev: target network device
256 */
257void net_dim_free_irq_moder(struct net_device *dev);
258
259/**
260 * net_dim_setting - initialize DIM's cq mode and schedule worker
261 * @dev: target network device
262 * @dim: DIM context
263 * @is_tx: true indicates the tx direction, false indicates the rx direction
264 */
265void net_dim_setting(struct net_device *dev, struct dim *dim, bool is_tx);
266
267/**
268 * net_dim_work_cancel - synchronously cancel dim's worker
269 * @dim: DIM context
270 */
271void net_dim_work_cancel(struct dim *dim);
272
273/**
274 * net_dim_get_rx_irq_moder - get DIM rx results based on profile_ix
275 * @dev: target network device
276 * @dim: DIM context
277 *
278 * Return: DIM irq moderation
279 */
280struct dim_cq_moder
281net_dim_get_rx_irq_moder(struct net_device *dev, struct dim *dim);
282
283/**
284 * net_dim_get_tx_irq_moder - get DIM tx results based on profile_ix
285 * @dev: target network device
286 * @dim: DIM context
287 *
288 * Return: DIM irq moderation
289 */
290struct dim_cq_moder
291net_dim_get_tx_irq_moder(struct net_device *dev, struct dim *dim);
292
293/**
294 * net_dim_set_rx_mode - set DIM rx cq mode
295 * @dev: target network device
296 * @rx_mode: target rx cq mode
297 */
298void net_dim_set_rx_mode(struct net_device *dev, u8 rx_mode);
299
300/**
301 * net_dim_set_tx_mode - set DIM tx cq mode
302 * @dev: target network device
303 * @tx_mode: target tx cq mode
304 */
305void net_dim_set_tx_mode(struct net_device *dev, u8 tx_mode);
306
307/**
308 * dim_on_top - check if current state is a good place to stop (top location)
309 * @dim: DIM context
310 *
311 * Check if current profile is a good place to park at.
312 * This will result in reducing the DIM checks frequency as we assume we
313 * shouldn't probably change profiles, unless traffic pattern wasn't changed.
314 */
315bool dim_on_top(struct dim *dim);
316
317/**
318 * dim_turn - change profile altering direction
319 * @dim: DIM context
320 *
321 * Go left if we were going right and vice-versa.
322 * Do nothing if currently parking.
323 */
324void dim_turn(struct dim *dim);
325
326/**
327 * dim_park_on_top - enter a parking state on a top location
328 * @dim: DIM context
329 *
330 * Enter parking state.
331 * Clear all movement history.
332 */
333void dim_park_on_top(struct dim *dim);
334
335/**
336 * dim_park_tired - enter a tired parking state
337 * @dim: DIM context
338 *
339 * Enter parking state.
340 * Clear all movement history and cause DIM checks frequency to reduce.
341 */
342void dim_park_tired(struct dim *dim);
343
344/**
345 * dim_calc_stats - calculate the difference between two samples
346 * @start: start sample
347 * @end: end sample
348 * @curr_stats: delta between samples
349 *
350 * Calculate the delta between two samples (in data rates).
351 * Takes into consideration counter wrap-around.
352 * Returned boolean indicates whether curr_stats are reliable.
353 */
354bool dim_calc_stats(const struct dim_sample *start,
355 const struct dim_sample *end,
356 struct dim_stats *curr_stats);
357
358/**
359 * dim_update_sample - set a sample's fields with given values
360 * @event_ctr: number of events to set
361 * @packets: number of packets to set
362 * @bytes: number of bytes to set
363 * @s: DIM sample
364 */
365static inline void
366dim_update_sample(u16 event_ctr, u64 packets, u64 bytes, struct dim_sample *s)
367{
368 s->time = ktime_get();
369 s->pkt_ctr = packets;
370 s->byte_ctr = bytes;
371 s->event_ctr = event_ctr;
372}
373
374/**
375 * dim_update_sample_with_comps - set a sample's fields with given
376 * values including the completion parameter
377 * @event_ctr: number of events to set
378 * @packets: number of packets to set
379 * @bytes: number of bytes to set
380 * @comps: number of completions to set
381 * @s: DIM sample
382 */
383static inline void
384dim_update_sample_with_comps(u16 event_ctr, u64 packets, u64 bytes, u64 comps,
385 struct dim_sample *s)
386{
387 dim_update_sample(event_ctr, packets, bytes, s);
388 s->comp_ctr = comps;
389}
390
391/* Net DIM */
392
393/**
394 * net_dim_get_rx_moderation - provide a CQ moderation object for the given RX profile
395 * @cq_period_mode: CQ period mode
396 * @ix: Profile index
397 */
398struct dim_cq_moder net_dim_get_rx_moderation(u8 cq_period_mode, int ix);
399
400/**
401 * net_dim_get_def_rx_moderation - provide the default RX moderation
402 * @cq_period_mode: CQ period mode
403 */
404struct dim_cq_moder net_dim_get_def_rx_moderation(u8 cq_period_mode);
405
406/**
407 * net_dim_get_tx_moderation - provide a CQ moderation object for the given TX profile
408 * @cq_period_mode: CQ period mode
409 * @ix: Profile index
410 */
411struct dim_cq_moder net_dim_get_tx_moderation(u8 cq_period_mode, int ix);
412
413/**
414 * net_dim_get_def_tx_moderation - provide the default TX moderation
415 * @cq_period_mode: CQ period mode
416 */
417struct dim_cq_moder net_dim_get_def_tx_moderation(u8 cq_period_mode);
418
419/**
420 * net_dim - main DIM algorithm entry point
421 * @dim: DIM instance information
422 * @end_sample: Current data measurement
423 *
424 * Called by the consumer.
425 * This is the main logic of the algorithm, where data is processed in order
426 * to decide on next required action.
427 */
428void net_dim(struct dim *dim, const struct dim_sample *end_sample);
429
430/* RDMA DIM */
431
432/*
433 * RDMA DIM profile:
434 * profile size must be of RDMA_DIM_PARAMS_NUM_PROFILES.
435 */
436#define RDMA_DIM_PARAMS_NUM_PROFILES 9
437#define RDMA_DIM_START_PROFILE 0
438
439/**
440 * rdma_dim - Runs the adaptive moderation.
441 * @dim: The moderation struct.
442 * @completions: The number of completions collected in this round.
443 *
444 * Each call to rdma_dim takes the latest amount of completions that
445 * have been collected and counts them as a new event.
446 * Once enough events have been collected the algorithm decides a new
447 * moderation level.
448 */
449void rdma_dim(struct dim *dim, u64 completions);
450
451#endif /* DIM_H */
452