net/ice: fix Tx threshold setup
[dpdk.git] / drivers / event / octeontx / timvf_worker.h
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright(c) 2017 Cavium, Inc
3  */
4
5 #include <rte_common.h>
6 #include <rte_branch_prediction.h>
7
8 #include "timvf_evdev.h"
9
10 static inline int16_t
11 timr_bkt_fetch_rem(uint64_t w1)
12 {
13         return (w1 >> TIM_BUCKET_W1_S_CHUNK_REMAINDER) &
14                 TIM_BUCKET_W1_M_CHUNK_REMAINDER;
15 }
16
17 static inline int16_t
18 timr_bkt_get_rem(struct tim_mem_bucket *bktp)
19 {
20         return __atomic_load_n(&bktp->chunk_remainder,
21                         __ATOMIC_ACQUIRE);
22 }
23
24 static inline void
25 timr_bkt_set_rem(struct tim_mem_bucket *bktp, uint16_t v)
26 {
27         __atomic_store_n(&bktp->chunk_remainder, v,
28                         __ATOMIC_RELEASE);
29 }
30
31 static inline void
32 timr_bkt_sub_rem(struct tim_mem_bucket *bktp, uint16_t v)
33 {
34         __atomic_fetch_sub(&bktp->chunk_remainder, v,
35                         __ATOMIC_RELEASE);
36 }
37
38 static inline uint8_t
39 timr_bkt_get_sbt(uint64_t w1)
40 {
41         return (w1 >> TIM_BUCKET_W1_S_SBT) & TIM_BUCKET_W1_M_SBT;
42 }
43
44 static inline uint64_t
45 timr_bkt_set_sbt(struct tim_mem_bucket *bktp)
46 {
47         const uint64_t v = TIM_BUCKET_W1_M_SBT << TIM_BUCKET_W1_S_SBT;
48         return __atomic_fetch_or(&bktp->w1, v, __ATOMIC_ACQ_REL);
49 }
50
51 static inline uint64_t
52 timr_bkt_clr_sbt(struct tim_mem_bucket *bktp)
53 {
54         const uint64_t v = ~(TIM_BUCKET_W1_M_SBT << TIM_BUCKET_W1_S_SBT);
55         return __atomic_fetch_and(&bktp->w1, v, __ATOMIC_ACQ_REL);
56 }
57
58 static inline uint8_t
59 timr_bkt_get_shbt(uint64_t w1)
60 {
61         return ((w1 >> TIM_BUCKET_W1_S_HBT) & TIM_BUCKET_W1_M_HBT) |
62                 ((w1 >> TIM_BUCKET_W1_S_SBT) & TIM_BUCKET_W1_M_SBT);
63 }
64
65 static inline uint8_t
66 timr_bkt_get_hbt(uint64_t w1)
67 {
68         return (w1 >> TIM_BUCKET_W1_S_HBT) & TIM_BUCKET_W1_M_HBT;
69 }
70
71 static inline uint8_t
72 timr_bkt_get_bsk(uint64_t w1)
73 {
74         return (w1 >> TIM_BUCKET_W1_S_BSK) & TIM_BUCKET_W1_M_BSK;
75 }
76
77 static inline uint64_t
78 timr_bkt_clr_bsk(struct tim_mem_bucket *bktp)
79 {
80         /*Clear everything except lock. */
81         const uint64_t v = TIM_BUCKET_W1_M_LOCK << TIM_BUCKET_W1_S_LOCK;
82         return __atomic_fetch_and(&bktp->w1, v, __ATOMIC_ACQ_REL);
83 }
84
85 static inline uint64_t
86 timr_bkt_fetch_sema_lock(struct tim_mem_bucket *bktp)
87 {
88         return __atomic_fetch_add(&bktp->w1, TIM_BUCKET_SEMA_WLOCK,
89                         __ATOMIC_ACQ_REL);
90 }
91
92 static inline uint64_t
93 timr_bkt_fetch_sema(struct tim_mem_bucket *bktp)
94 {
95         return __atomic_fetch_add(&bktp->w1, TIM_BUCKET_SEMA,
96                         __ATOMIC_RELAXED);
97 }
98
99 static inline uint64_t
100 timr_bkt_inc_lock(struct tim_mem_bucket *bktp)
101 {
102         const uint64_t v = 1ull << TIM_BUCKET_W1_S_LOCK;
103         return __atomic_fetch_add(&bktp->w1, v, __ATOMIC_ACQ_REL);
104 }
105
106 static inline void
107 timr_bkt_dec_lock(struct tim_mem_bucket *bktp)
108 {
109         __atomic_add_fetch(&bktp->lock, 0xff, __ATOMIC_ACQ_REL);
110 }
111
112 static inline uint32_t
113 timr_bkt_get_nent(uint64_t w1)
114 {
115         return (w1 >> TIM_BUCKET_W1_S_NUM_ENTRIES) &
116                 TIM_BUCKET_W1_M_NUM_ENTRIES;
117 }
118
119 static inline void
120 timr_bkt_inc_nent(struct tim_mem_bucket *bktp)
121 {
122         __atomic_add_fetch(&bktp->nb_entry, 1, __ATOMIC_RELAXED);
123 }
124
125 static inline void
126 timr_bkt_add_nent(struct tim_mem_bucket *bktp, uint32_t v)
127 {
128         __atomic_add_fetch(&bktp->nb_entry, v, __ATOMIC_RELAXED);
129 }
130
131 static inline uint64_t
132 timr_bkt_clr_nent(struct tim_mem_bucket *bktp)
133 {
134         const uint64_t v = ~(TIM_BUCKET_W1_M_NUM_ENTRIES <<
135                         TIM_BUCKET_W1_S_NUM_ENTRIES);
136         return __atomic_and_fetch(&bktp->w1, v, __ATOMIC_ACQ_REL);
137 }
138
139 static inline struct tim_mem_entry *
140 timr_clr_bkt(struct timvf_ring * const timr, struct tim_mem_bucket * const bkt)
141 {
142         struct tim_mem_entry *chunk;
143         struct tim_mem_entry *pnext;
144         chunk = ((struct tim_mem_entry *)(uintptr_t)bkt->first_chunk);
145         chunk = (struct tim_mem_entry *)(uintptr_t)(chunk + nb_chunk_slots)->w0;
146
147         while (chunk) {
148                 pnext = (struct tim_mem_entry *)(uintptr_t)
149                         ((chunk + nb_chunk_slots)->w0);
150                 rte_mempool_put(timr->chunk_pool, chunk);
151                 chunk = pnext;
152         }
153         return (struct tim_mem_entry *)(uintptr_t)bkt->first_chunk;
154 }
155
156 static inline int
157 timvf_rem_entry(struct rte_event_timer *tim)
158 {
159         uint64_t lock_sema;
160         struct tim_mem_entry *entry;
161         struct tim_mem_bucket *bkt;
162         if (tim->impl_opaque[1] == 0 ||
163                         tim->impl_opaque[0] == 0)
164                 return -ENOENT;
165
166         entry = (struct tim_mem_entry *)(uintptr_t)tim->impl_opaque[0];
167         if (entry->wqe != tim->ev.u64) {
168                 tim->impl_opaque[1] = tim->impl_opaque[0] = 0;
169                 return -ENOENT;
170         }
171         bkt = (struct tim_mem_bucket *)(uintptr_t)tim->impl_opaque[1];
172         lock_sema = timr_bkt_inc_lock(bkt);
173         if (timr_bkt_get_shbt(lock_sema)
174                         || !timr_bkt_get_nent(lock_sema)) {
175                 timr_bkt_dec_lock(bkt);
176                 tim->impl_opaque[1] = tim->impl_opaque[0] = 0;
177                 return -ENOENT;
178         }
179
180         entry->w0 = entry->wqe = 0;
181         timr_bkt_dec_lock(bkt);
182
183         tim->state = RTE_EVENT_TIMER_CANCELED;
184         tim->impl_opaque[1] = tim->impl_opaque[0] = 0;
185         return 0;
186 }
187
188 static inline struct tim_mem_entry *
189 timvf_refill_chunk_generic(struct tim_mem_bucket * const bkt,
190                 struct timvf_ring * const timr)
191 {
192         struct tim_mem_entry *chunk;
193
194         if (bkt->nb_entry || !bkt->first_chunk) {
195                 if (unlikely(rte_mempool_get(timr->chunk_pool,
196                                                 (void **)&chunk))) {
197                         return NULL;
198                 }
199                 if (bkt->nb_entry) {
200                         *(uint64_t *)(((struct tim_mem_entry *)(uintptr_t)
201                                         bkt->current_chunk) +
202                                         nb_chunk_slots) =
203                                 (uintptr_t) chunk;
204                 } else {
205                         bkt->first_chunk = (uintptr_t) chunk;
206                 }
207         } else {
208                 chunk = timr_clr_bkt(timr, bkt);
209                 bkt->first_chunk = (uintptr_t)chunk;
210         }
211         *(uint64_t *)(chunk + nb_chunk_slots) = 0;
212
213         return chunk;
214 }
215
216 static inline struct tim_mem_entry *
217 timvf_refill_chunk_fpa(struct tim_mem_bucket * const bkt,
218                 struct timvf_ring * const timr)
219 {
220         struct tim_mem_entry *chunk;
221
222         if (unlikely(rte_mempool_get(timr->chunk_pool, (void **)&chunk)))
223                 return NULL;
224
225         *(uint64_t *)(chunk + nb_chunk_slots) = 0;
226         if (bkt->nb_entry) {
227                 *(uint64_t *)(((struct tim_mem_entry *)(uintptr_t)
228                                 bkt->current_chunk) +
229                                 nb_chunk_slots) =
230                         (uintptr_t) chunk;
231         } else {
232                 bkt->first_chunk = (uintptr_t) chunk;
233         }
234
235         return chunk;
236 }
237
238 static inline struct tim_mem_bucket *
239 timvf_get_target_bucket(struct timvf_ring * const timr, const uint32_t rel_bkt)
240 {
241         const uint64_t bkt_cyc = rte_rdtsc() - timr->ring_start_cyc;
242         const uint32_t bucket = rte_reciprocal_divide_u64(bkt_cyc,
243                         &timr->fast_div) + rel_bkt;
244         const uint32_t tbkt_id = timr->get_target_bkt(bucket,
245                         timr->nb_bkts);
246         return &timr->bkt[tbkt_id];
247 }
248
249 /* Single producer functions. */
250 static inline int
251 timvf_add_entry_sp(struct timvf_ring * const timr, const uint32_t rel_bkt,
252                 struct rte_event_timer * const tim,
253                 const struct tim_mem_entry * const pent)
254 {
255         int16_t rem;
256         uint64_t lock_sema;
257         struct tim_mem_bucket *bkt;
258         struct tim_mem_entry *chunk;
259
260
261         bkt = timvf_get_target_bucket(timr, rel_bkt);
262 __retry:
263         /*Get Bucket sema*/
264         lock_sema = timr_bkt_fetch_sema(bkt);
265         /* Bucket related checks. */
266         if (unlikely(timr_bkt_get_hbt(lock_sema)))
267                 goto __retry;
268
269         /* Insert the work. */
270         rem = timr_bkt_fetch_rem(lock_sema);
271
272         if (!rem) {
273                 chunk = timr->refill_chunk(bkt, timr);
274                 if (unlikely(chunk == NULL)) {
275                         timr_bkt_set_rem(bkt, 0);
276                         tim->impl_opaque[0] = tim->impl_opaque[1] = 0;
277                         tim->state = RTE_EVENT_TIMER_ERROR;
278                         return -ENOMEM;
279                 }
280                 bkt->current_chunk = (uintptr_t) chunk;
281                 timr_bkt_set_rem(bkt, nb_chunk_slots - 1);
282         } else {
283                 chunk = (struct tim_mem_entry *)(uintptr_t)bkt->current_chunk;
284                 chunk += nb_chunk_slots - rem;
285         }
286         /* Copy work entry. */
287         *chunk = *pent;
288         timr_bkt_inc_nent(bkt);
289
290         tim->impl_opaque[0] = (uintptr_t)chunk;
291         tim->impl_opaque[1] = (uintptr_t)bkt;
292         tim->state = RTE_EVENT_TIMER_ARMED;
293         return 0;
294 }
295
296 /* Multi producer functions. */
297 static inline int
298 timvf_add_entry_mp(struct timvf_ring * const timr, const uint32_t rel_bkt,
299                 struct rte_event_timer * const tim,
300                 const struct tim_mem_entry * const pent)
301 {
302         int16_t rem;
303         uint64_t lock_sema;
304         struct tim_mem_bucket *bkt;
305         struct tim_mem_entry *chunk;
306
307 __retry:
308         bkt = timvf_get_target_bucket(timr, rel_bkt);
309         /* Bucket related checks. */
310         /*Get Bucket sema*/
311         lock_sema = timr_bkt_fetch_sema_lock(bkt);
312         if (unlikely(timr_bkt_get_shbt(lock_sema))) {
313                 timr_bkt_dec_lock(bkt);
314                 goto __retry;
315         }
316
317         rem = timr_bkt_fetch_rem(lock_sema);
318
319         if (rem < 0) {
320                 /* goto diff bucket. */
321                 timr_bkt_dec_lock(bkt);
322                 goto __retry;
323         } else if (!rem) {
324                 /*Only one thread can be here*/
325                 chunk = timr->refill_chunk(bkt, timr);
326                 if (unlikely(chunk == NULL)) {
327                         timr_bkt_set_rem(bkt, 0);
328                         timr_bkt_dec_lock(bkt);
329                         tim->impl_opaque[0] = tim->impl_opaque[1] = 0;
330                         tim->state = RTE_EVENT_TIMER_ERROR;
331                         return -ENOMEM;
332                 }
333                 bkt->current_chunk = (uintptr_t) chunk;
334                 timr_bkt_set_rem(bkt, nb_chunk_slots - 1);
335         } else {
336                 chunk = (struct tim_mem_entry *)(uintptr_t)bkt->current_chunk;
337                 chunk += nb_chunk_slots - rem;
338         }
339         /* Copy work entry. */
340         *chunk = *pent;
341         timr_bkt_inc_nent(bkt);
342         timr_bkt_dec_lock(bkt);
343
344         tim->impl_opaque[0] = (uintptr_t)chunk;
345         tim->impl_opaque[1] = (uintptr_t)bkt;
346         tim->state = RTE_EVENT_TIMER_ARMED;
347         return 0;
348 }
349
350 static inline uint16_t
351 timvf_cpy_wrk(uint16_t index, uint16_t cpy_lmt,
352                 struct tim_mem_entry *chunk,
353                 struct rte_event_timer ** const tim,
354                 const struct tim_mem_entry * const ents,
355                 const struct tim_mem_bucket * const bkt)
356 {
357         for (; index < cpy_lmt; index++) {
358                 *chunk = *(ents + index);
359                 tim[index]->impl_opaque[0] = (uintptr_t)chunk++;
360                 tim[index]->impl_opaque[1] = (uintptr_t)bkt;
361                 tim[index]->state = RTE_EVENT_TIMER_ARMED;
362         }
363
364         return index;
365 }
366
367 /* Burst mode functions */
368 static inline int
369 timvf_add_entry_brst(struct timvf_ring * const timr, const uint16_t rel_bkt,
370                 struct rte_event_timer ** const tim,
371                 const struct tim_mem_entry *ents,
372                 const uint16_t nb_timers)
373 {
374         int16_t rem;
375         int16_t crem;
376         uint8_t lock_cnt;
377         uint16_t index = 0;
378         uint16_t chunk_remainder;
379         uint64_t lock_sema;
380         struct tim_mem_bucket *bkt;
381         struct tim_mem_entry *chunk;
382
383 __retry:
384         bkt = timvf_get_target_bucket(timr, rel_bkt);
385
386         /* Only one thread beyond this. */
387         lock_sema = timr_bkt_inc_lock(bkt);
388         lock_cnt = (uint8_t)
389                 ((lock_sema >> TIM_BUCKET_W1_S_LOCK) & TIM_BUCKET_W1_M_LOCK);
390
391         if (lock_cnt) {
392                 timr_bkt_dec_lock(bkt);
393                 goto __retry;
394         }
395
396         /* Bucket related checks. */
397         if (unlikely(timr_bkt_get_hbt(lock_sema))) {
398                 timr_bkt_dec_lock(bkt);
399                 goto __retry;
400         }
401
402         chunk_remainder = timr_bkt_fetch_rem(lock_sema);
403         rem = chunk_remainder - nb_timers;
404         if (rem < 0) {
405                 crem = nb_chunk_slots - chunk_remainder;
406                 if (chunk_remainder && crem) {
407                         chunk = ((struct tim_mem_entry *)
408                                         (uintptr_t)bkt->current_chunk) + crem;
409
410                         index = timvf_cpy_wrk(index, chunk_remainder,
411                                         chunk, tim, ents, bkt);
412                         timr_bkt_sub_rem(bkt, chunk_remainder);
413                         timr_bkt_add_nent(bkt, chunk_remainder);
414                 }
415                 rem = nb_timers - chunk_remainder;
416                 ents = ents + chunk_remainder;
417
418                 chunk = timr->refill_chunk(bkt, timr);
419                 if (unlikely(chunk == NULL)) {
420                         timr_bkt_dec_lock(bkt);
421                         rte_errno = ENOMEM;
422                         tim[index]->state = RTE_EVENT_TIMER_ERROR;
423                         return crem;
424                 }
425                 *(uint64_t *)(chunk + nb_chunk_slots) = 0;
426                 bkt->current_chunk = (uintptr_t) chunk;
427
428                 index = timvf_cpy_wrk(index, nb_timers, chunk, tim, ents, bkt);
429                 timr_bkt_set_rem(bkt, nb_chunk_slots - rem);
430                 timr_bkt_add_nent(bkt, rem);
431         } else {
432                 chunk = (struct tim_mem_entry *)(uintptr_t)bkt->current_chunk;
433                 chunk += (nb_chunk_slots - chunk_remainder);
434
435                 index = timvf_cpy_wrk(index, nb_timers,
436                                 chunk, tim, ents, bkt);
437                 timr_bkt_sub_rem(bkt, nb_timers);
438                 timr_bkt_add_nent(bkt, nb_timers);
439         }
440
441         timr_bkt_dec_lock(bkt);
442         return nb_timers;
443 }