service: use id in runstate function
[dpdk.git] / lib / librte_eal / common / rte_service.c
1 /*-
2  *   BSD LICENSE
3  *
4  *   Copyright(c) 2017 Intel Corporation. All rights reserved.
5  *   All rights reserved.
6  *
7  *   Redistribution and use in source and binary forms, with or without
8  *   modification, are permitted provided that the following conditions
9  *   are met:
10  *
11  *     * Redistributions of source code must retain the above copyright
12  *       notice, this list of conditions and the following disclaimer.
13  *     * Redistributions in binary form must reproduce the above copyright
14  *       notice, this list of conditions and the following disclaimer in
15  *       the documentation and/or other materials provided with the
16  *       distribution.
17  *     * Neither the name of Intel Corporation nor the names of its
18  *       contributors may be used to endorse or promote products derived
19  *       from this software without specific prior written permission.
20  *
21  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
24  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
25  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
26  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
27  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
28  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
29  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
30  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
31  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
32  */
33
34 #include <stdio.h>
35 #include <unistd.h>
36 #include <inttypes.h>
37 #include <limits.h>
38 #include <string.h>
39 #include <dirent.h>
40
41 #include <rte_service.h>
42 #include "include/rte_service_component.h"
43
44 #include <rte_eal.h>
45 #include <rte_lcore.h>
46 #include <rte_common.h>
47 #include <rte_debug.h>
48 #include <rte_cycles.h>
49 #include <rte_atomic.h>
50 #include <rte_memory.h>
51 #include <rte_malloc.h>
52
53 #define RTE_SERVICE_NUM_MAX 64
54
55 #define SERVICE_F_REGISTERED    (1 << 0)
56 #define SERVICE_F_STATS_ENABLED (1 << 1)
57
58 /* runstates for services and lcores, denoting if they are active or not */
59 #define RUNSTATE_STOPPED 0
60 #define RUNSTATE_RUNNING 1
61
62 /* internal representation of a service */
63 struct rte_service_spec_impl {
64         /* public part of the struct */
65         struct rte_service_spec spec;
66
67         /* atomic lock that when set indicates a service core is currently
68          * running this service callback. When not set, a core may take the
69          * lock and then run the service callback.
70          */
71         rte_atomic32_t execute_lock;
72
73         /* API set/get-able variables */
74         int32_t runstate;
75         uint8_t internal_flags;
76
77         /* per service statistics */
78         uint32_t num_mapped_cores;
79         uint64_t calls;
80         uint64_t cycles_spent;
81 } __rte_cache_aligned;
82
83 /* the internal values of a service core */
84 struct core_state {
85         /* map of services IDs are run on this core */
86         uint64_t service_mask;
87         uint8_t runstate; /* running or stopped */
88         uint8_t is_service_core; /* set if core is currently a service core */
89
90         /* extreme statistics */
91         uint64_t calls_per_service[RTE_SERVICE_NUM_MAX];
92 } __rte_cache_aligned;
93
94 static uint32_t rte_service_count;
95 static struct rte_service_spec_impl *rte_services;
96 static struct core_state *lcore_states;
97 static uint32_t rte_service_library_initialized;
98
99 int32_t rte_service_init(void)
100 {
101         if (rte_service_library_initialized) {
102                 printf("service library init() called, init flag %d\n",
103                         rte_service_library_initialized);
104                 return -EALREADY;
105         }
106
107         rte_services = rte_calloc("rte_services", RTE_SERVICE_NUM_MAX,
108                         sizeof(struct rte_service_spec_impl),
109                         RTE_CACHE_LINE_SIZE);
110         if (!rte_services) {
111                 printf("error allocating rte services array\n");
112                 return -ENOMEM;
113         }
114
115         lcore_states = rte_calloc("rte_service_core_states", RTE_MAX_LCORE,
116                         sizeof(struct core_state), RTE_CACHE_LINE_SIZE);
117         if (!lcore_states) {
118                 printf("error allocating core states array\n");
119                 return -ENOMEM;
120         }
121
122         int i;
123         int count = 0;
124         struct rte_config *cfg = rte_eal_get_configuration();
125         for (i = 0; i < RTE_MAX_LCORE; i++) {
126                 if (lcore_config[i].core_role == ROLE_SERVICE) {
127                         if ((unsigned int)i == cfg->master_lcore)
128                                 continue;
129                         rte_service_lcore_add(i);
130                         count++;
131                 }
132         }
133
134         rte_service_library_initialized = 1;
135         return 0;
136 }
137
138 /* returns 1 if service is registered and has not been unregistered
139  * Returns 0 if service never registered, or has been unregistered
140  */
141 static inline int
142 service_valid(uint32_t id)
143 {
144         return !!(rte_services[id].internal_flags & SERVICE_F_REGISTERED);
145 }
146
147 /* validate ID and retrieve service pointer, or return error value */
148 #define SERVICE_VALID_GET_OR_ERR_RET(id, service, retval) do {          \
149         if (id >= RTE_SERVICE_NUM_MAX || !service_valid(id))            \
150                 return retval;                                          \
151         service = &rte_services[id];                                    \
152 } while (0)
153
154 /* returns 1 if statistics should be colleced for service
155  * Returns 0 if statistics should not be collected for service
156  */
157 static inline int
158 service_stats_enabled(struct rte_service_spec_impl *impl)
159 {
160         return !!(impl->internal_flags & SERVICE_F_STATS_ENABLED);
161 }
162
163 static inline int
164 service_mt_safe(struct rte_service_spec_impl *s)
165 {
166         return s->spec.capabilities & RTE_SERVICE_CAP_MT_SAFE;
167 }
168
169 int32_t rte_service_set_stats_enable(struct rte_service_spec *service,
170                                   int32_t enabled)
171 {
172         struct rte_service_spec_impl *impl =
173                 (struct rte_service_spec_impl *)service;
174         if (!impl)
175                 return -EINVAL;
176
177         if (enabled)
178                 impl->internal_flags |= SERVICE_F_STATS_ENABLED;
179         else
180                 impl->internal_flags &= ~(SERVICE_F_STATS_ENABLED);
181
182         return 0;
183 }
184
185 uint32_t
186 rte_service_get_count(void)
187 {
188         return rte_service_count;
189 }
190
191 struct rte_service_spec *
192 rte_service_get_by_id(uint32_t id)
193 {
194         struct rte_service_spec *service = NULL;
195         if (id < rte_service_count)
196                 service = (struct rte_service_spec *)&rte_services[id];
197
198         return service;
199 }
200
201 struct rte_service_spec *rte_service_get_by_name(const char *name)
202 {
203         struct rte_service_spec *service = NULL;
204         int i;
205         for (i = 0; i < RTE_SERVICE_NUM_MAX; i++) {
206                 if (service_valid(i) &&
207                                 strcmp(name, rte_services[i].spec.name) == 0) {
208                         service = (struct rte_service_spec *)&rte_services[i];
209                         break;
210                 }
211         }
212
213         return service;
214 }
215
216 const char *
217 rte_service_get_name(uint32_t id)
218 {
219         struct rte_service_spec_impl *s;
220         SERVICE_VALID_GET_OR_ERR_RET(id, s, 0);
221         return s->spec.name;
222 }
223
224 int32_t
225 rte_service_probe_capability(uint32_t id, uint32_t capability)
226 {
227         struct rte_service_spec_impl *s;
228         SERVICE_VALID_GET_OR_ERR_RET(id, s, -EINVAL);
229         return s->spec.capabilities & capability;
230 }
231
232 int32_t
233 rte_service_component_register(const struct rte_service_spec *spec,
234                                uint32_t *id_ptr)
235 {
236         uint32_t i;
237         int32_t free_slot = -1;
238
239         if (spec->callback == NULL || strlen(spec->name) == 0)
240                 return -EINVAL;
241
242         for (i = 0; i < RTE_SERVICE_NUM_MAX; i++) {
243                 if (!service_valid(i)) {
244                         free_slot = i;
245                         break;
246                 }
247         }
248
249         if ((free_slot < 0) || (i == RTE_SERVICE_NUM_MAX))
250                 return -ENOSPC;
251
252         struct rte_service_spec_impl *s = &rte_services[free_slot];
253         s->spec = *spec;
254         s->internal_flags |= SERVICE_F_REGISTERED;
255
256         rte_smp_wmb();
257         rte_service_count++;
258
259         if (id_ptr)
260                 *id_ptr = free_slot;
261
262         return 0;
263 }
264
265 int32_t
266 rte_service_unregister(struct rte_service_spec *spec)
267 {
268         struct rte_service_spec_impl *s = NULL;
269         struct rte_service_spec_impl *spec_impl =
270                 (struct rte_service_spec_impl *)spec;
271
272         uint32_t i;
273         uint32_t service_id;
274         for (i = 0; i < RTE_SERVICE_NUM_MAX; i++) {
275                 if (&rte_services[i] == spec_impl) {
276                         s = spec_impl;
277                         service_id = i;
278                         break;
279                 }
280         }
281
282         if (!s)
283                 return -EINVAL;
284
285         rte_service_count--;
286         rte_smp_wmb();
287
288         s->internal_flags &= ~(SERVICE_F_REGISTERED);
289
290         for (i = 0; i < RTE_MAX_LCORE; i++)
291                 lcore_states[i].service_mask &= ~(UINT64_C(1) << service_id);
292
293         memset(&rte_services[service_id], 0,
294                         sizeof(struct rte_service_spec_impl));
295
296         return 0;
297 }
298
299 int32_t
300 rte_service_runstate_set(uint32_t id, uint32_t runstate)
301 {
302         struct rte_service_spec_impl *s;
303         SERVICE_VALID_GET_OR_ERR_RET(id, s, -EINVAL);
304
305         if (runstate)
306                 s->runstate = RUNSTATE_RUNNING;
307         else
308                 s->runstate = RUNSTATE_STOPPED;
309
310         rte_smp_wmb();
311         return 0;
312 }
313
314 int32_t
315 rte_service_runstate_get(uint32_t id)
316 {
317         struct rte_service_spec_impl *s;
318         SERVICE_VALID_GET_OR_ERR_RET(id, s, -EINVAL);
319
320         return (s->runstate == RUNSTATE_RUNNING) && (s->num_mapped_cores > 0);
321 }
322
323 static int32_t
324 rte_service_runner_func(void *arg)
325 {
326         RTE_SET_USED(arg);
327         uint32_t i;
328         const int lcore = rte_lcore_id();
329         struct core_state *cs = &lcore_states[lcore];
330
331         while (lcore_states[lcore].runstate == RUNSTATE_RUNNING) {
332                 const uint64_t service_mask = cs->service_mask;
333                 for (i = 0; i < rte_service_count; i++) {
334                         struct rte_service_spec_impl *s = &rte_services[i];
335                         if (s->runstate != RUNSTATE_RUNNING ||
336                                         !(service_mask & (UINT64_C(1) << i)))
337                                 continue;
338
339                         /* check do we need cmpset, if MT safe or <= 1 core
340                          * mapped, atomic ops are not required.
341                          */
342                         const int need_cmpset = !((service_mt_safe(s) == 0) &&
343                                                 (s->num_mapped_cores > 1));
344                         uint32_t *lock = (uint32_t *)&s->execute_lock;
345
346                         if (need_cmpset || rte_atomic32_cmpset(lock, 0, 1)) {
347                                 void *userdata = s->spec.callback_userdata;
348
349                                 if (service_stats_enabled(s)) {
350                                         uint64_t start = rte_rdtsc();
351                                         s->spec.callback(userdata);
352                                         uint64_t end = rte_rdtsc();
353                                         s->cycles_spent += end - start;
354                                         cs->calls_per_service[i]++;
355                                         s->calls++;
356                                 } else
357                                         s->spec.callback(userdata);
358
359                                 if (need_cmpset)
360                                         rte_atomic32_clear(&s->execute_lock);
361                         }
362                 }
363
364                 rte_smp_rmb();
365         }
366
367         lcore_config[lcore].state = WAIT;
368
369         return 0;
370 }
371
372 int32_t
373 rte_service_lcore_count(void)
374 {
375         int32_t count = 0;
376         uint32_t i;
377         for (i = 0; i < RTE_MAX_LCORE; i++)
378                 count += lcore_states[i].is_service_core;
379         return count;
380 }
381
382 int32_t
383 rte_service_lcore_list(uint32_t array[], uint32_t n)
384 {
385         uint32_t count = rte_service_lcore_count();
386         if (count > n)
387                 return -ENOMEM;
388
389         if (!array)
390                 return -EINVAL;
391
392         uint32_t i;
393         uint32_t idx = 0;
394         for (i = 0; i < RTE_MAX_LCORE; i++) {
395                 struct core_state *cs = &lcore_states[i];
396                 if (cs->is_service_core) {
397                         array[idx] = i;
398                         idx++;
399                 }
400         }
401
402         return count;
403 }
404
405 int32_t
406 rte_service_lcore_count_services(uint32_t lcore)
407 {
408         if (lcore >= RTE_MAX_LCORE)
409                 return -EINVAL;
410
411         struct core_state *cs = &lcore_states[lcore];
412         if (!cs->is_service_core)
413                 return -ENOTSUP;
414
415         return __builtin_popcountll(cs->service_mask);
416 }
417
418 int32_t
419 rte_service_start_with_defaults(void)
420 {
421         /* create a default mapping from cores to services, then start the
422          * services to make them transparent to unaware applications.
423          */
424         uint32_t i;
425         int ret;
426         uint32_t count = rte_service_get_count();
427
428         int32_t lcore_iter = 0;
429         uint32_t ids[RTE_MAX_LCORE];
430         int32_t lcore_count = rte_service_lcore_list(ids, RTE_MAX_LCORE);
431
432         if (lcore_count == 0)
433                 return -ENOTSUP;
434
435         for (i = 0; (int)i < lcore_count; i++)
436                 rte_service_lcore_start(ids[i]);
437
438         for (i = 0; i < count; i++) {
439                 struct rte_service_spec *s = rte_service_get_by_id(i);
440                 if (!s)
441                         return -EINVAL;
442
443                 /* do 1:1 core mapping here, with each service getting
444                  * assigned a single core by default. Adding multiple services
445                  * should multiplex to a single core, or 1:1 if there are the
446                  * same amount of services as service-cores
447                  */
448                 ret = rte_service_map_lcore_set(i, ids[lcore_iter], 1);
449                 if (ret)
450                         return -ENODEV;
451
452                 lcore_iter++;
453                 if (lcore_iter >= lcore_count)
454                         lcore_iter = 0;
455
456                 ret = rte_service_runstate_set(i, 1);
457                 if (ret)
458                         return -ENOEXEC;
459         }
460
461         return 0;
462 }
463
464 static int32_t
465 service_update(struct rte_service_spec *service, uint32_t lcore,
466                 uint32_t *set, uint32_t *enabled)
467 {
468         uint32_t i;
469         int32_t sid = -1;
470
471         for (i = 0; i < RTE_SERVICE_NUM_MAX; i++) {
472                 if ((struct rte_service_spec *)&rte_services[i] == service &&
473                                 service_valid(i)) {
474                         sid = i;
475                         break;
476                 }
477         }
478
479         if (sid == -1 || lcore >= RTE_MAX_LCORE)
480                 return -EINVAL;
481
482         if (!lcore_states[lcore].is_service_core)
483                 return -EINVAL;
484
485         uint64_t sid_mask = UINT64_C(1) << sid;
486         if (set) {
487                 if (*set) {
488                         lcore_states[lcore].service_mask |= sid_mask;
489                         rte_services[sid].num_mapped_cores++;
490                 } else {
491                         lcore_states[lcore].service_mask &= ~(sid_mask);
492                         rte_services[sid].num_mapped_cores--;
493                 }
494         }
495
496         if (enabled)
497                 *enabled = (lcore_states[lcore].service_mask & (sid_mask));
498
499         rte_smp_wmb();
500
501         return 0;
502 }
503
504 int32_t
505 rte_service_map_lcore_set(uint32_t id, uint32_t lcore, uint32_t enabled)
506 {
507         struct rte_service_spec_impl *s;
508         SERVICE_VALID_GET_OR_ERR_RET(id, s, -EINVAL);
509         uint32_t on = enabled > 0;
510         return service_update(&s->spec, lcore, &on, 0);
511 }
512
513 int32_t
514 rte_service_map_lcore_get(uint32_t id, uint32_t lcore)
515 {
516         struct rte_service_spec_impl *s;
517         SERVICE_VALID_GET_OR_ERR_RET(id, s, -EINVAL);
518         uint32_t enabled;
519         int ret = service_update(&s->spec, lcore, 0, &enabled);
520         if (ret == 0)
521                 return enabled;
522         return ret;
523 }
524
525 int32_t rte_service_lcore_reset_all(void)
526 {
527         /* loop over cores, reset all to mask 0 */
528         uint32_t i;
529         for (i = 0; i < RTE_MAX_LCORE; i++) {
530                 lcore_states[i].service_mask = 0;
531                 lcore_states[i].is_service_core = 0;
532                 lcore_states[i].runstate = RUNSTATE_STOPPED;
533         }
534         for (i = 0; i < RTE_SERVICE_NUM_MAX; i++)
535                 rte_services[i].num_mapped_cores = 0;
536
537         rte_smp_wmb();
538
539         return 0;
540 }
541
542 static void
543 set_lcore_state(uint32_t lcore, int32_t state)
544 {
545         /* mark core state in hugepage backed config */
546         struct rte_config *cfg = rte_eal_get_configuration();
547         cfg->lcore_role[lcore] = state;
548
549         /* mark state in process local lcore_config */
550         lcore_config[lcore].core_role = state;
551
552         /* update per-lcore optimized state tracking */
553         lcore_states[lcore].is_service_core = (state == ROLE_SERVICE);
554 }
555
556 int32_t
557 rte_service_lcore_add(uint32_t lcore)
558 {
559         if (lcore >= RTE_MAX_LCORE)
560                 return -EINVAL;
561         if (lcore_states[lcore].is_service_core)
562                 return -EALREADY;
563
564         set_lcore_state(lcore, ROLE_SERVICE);
565
566         /* ensure that after adding a core the mask and state are defaults */
567         lcore_states[lcore].service_mask = 0;
568         lcore_states[lcore].runstate = RUNSTATE_STOPPED;
569
570         rte_smp_wmb();
571         return 0;
572 }
573
574 int32_t
575 rte_service_lcore_del(uint32_t lcore)
576 {
577         if (lcore >= RTE_MAX_LCORE)
578                 return -EINVAL;
579
580         struct core_state *cs = &lcore_states[lcore];
581         if (!cs->is_service_core)
582                 return -EINVAL;
583
584         if (cs->runstate != RUNSTATE_STOPPED)
585                 return -EBUSY;
586
587         set_lcore_state(lcore, ROLE_RTE);
588
589         rte_smp_wmb();
590         return 0;
591 }
592
593 int32_t
594 rte_service_lcore_start(uint32_t lcore)
595 {
596         if (lcore >= RTE_MAX_LCORE)
597                 return -EINVAL;
598
599         struct core_state *cs = &lcore_states[lcore];
600         if (!cs->is_service_core)
601                 return -EINVAL;
602
603         if (cs->runstate == RUNSTATE_RUNNING)
604                 return -EALREADY;
605
606         /* set core to run state first, and then launch otherwise it will
607          * return immediately as runstate keeps it in the service poll loop
608          */
609         lcore_states[lcore].runstate = RUNSTATE_RUNNING;
610
611         int ret = rte_eal_remote_launch(rte_service_runner_func, 0, lcore);
612         /* returns -EBUSY if the core is already launched, 0 on success */
613         return ret;
614 }
615
616 int32_t
617 rte_service_lcore_stop(uint32_t lcore)
618 {
619         if (lcore >= RTE_MAX_LCORE)
620                 return -EINVAL;
621
622         if (lcore_states[lcore].runstate == RUNSTATE_STOPPED)
623                 return -EALREADY;
624
625         uint32_t i;
626         for (i = 0; i < RTE_SERVICE_NUM_MAX; i++) {
627                 int32_t enabled =
628                         lcore_states[i].service_mask & (UINT64_C(1) << i);
629                 int32_t service_running = rte_services[i].runstate !=
630                                                 RUNSTATE_STOPPED;
631                 int32_t only_core = rte_services[i].num_mapped_cores == 1;
632
633                 /* if the core is mapped, and the service is running, and this
634                  * is the only core that is mapped, the service would cease to
635                  * run if this core stopped, so fail instead.
636                  */
637                 if (enabled && service_running && only_core)
638                         return -EBUSY;
639         }
640
641         lcore_states[lcore].runstate = RUNSTATE_STOPPED;
642
643         return 0;
644 }
645
646 static void
647 rte_service_dump_one(FILE *f, struct rte_service_spec_impl *s,
648                      uint64_t all_cycles, uint32_t reset)
649 {
650         /* avoid divide by zero */
651         if (all_cycles == 0)
652                 all_cycles = 1;
653
654         int calls = 1;
655         if (s->calls != 0)
656                 calls = s->calls;
657
658         fprintf(f, "  %s: stats %d\tcalls %"PRIu64"\tcycles %"
659                         PRIu64"\tavg: %"PRIu64"\n",
660                         s->spec.name, service_stats_enabled(s), s->calls,
661                         s->cycles_spent, s->cycles_spent / calls);
662
663         if (reset) {
664                 s->cycles_spent = 0;
665                 s->calls = 0;
666         }
667 }
668
669 static void
670 service_dump_calls_per_lcore(FILE *f, uint32_t lcore, uint32_t reset)
671 {
672         uint32_t i;
673         struct core_state *cs = &lcore_states[lcore];
674
675         fprintf(f, "%02d\t", lcore);
676         for (i = 0; i < RTE_SERVICE_NUM_MAX; i++) {
677                 if (!service_valid(i))
678                         continue;
679                 fprintf(f, "%"PRIu64"\t", cs->calls_per_service[i]);
680                 if (reset)
681                         cs->calls_per_service[i] = 0;
682         }
683         fprintf(f, "\n");
684 }
685
686 int32_t rte_service_dump(FILE *f, struct rte_service_spec *service)
687 {
688         uint32_t i;
689
690         uint64_t total_cycles = 0;
691         for (i = 0; i < rte_service_count; i++) {
692                 if (!service_valid(i))
693                         continue;
694                 total_cycles += rte_services[i].cycles_spent;
695         }
696
697         if (service) {
698                 struct rte_service_spec_impl *s =
699                         (struct rte_service_spec_impl *)service;
700                 fprintf(f, "Service %s Summary\n", s->spec.name);
701                 uint32_t reset = 0;
702                 rte_service_dump_one(f, s, total_cycles, reset);
703                 return 0;
704         }
705
706         fprintf(f, "Services Summary\n");
707         for (i = 0; i < rte_service_count; i++) {
708                 uint32_t reset = 1;
709                 rte_service_dump_one(f, &rte_services[i], total_cycles, reset);
710         }
711
712         fprintf(f, "Service Cores Summary\n");
713         for (i = 0; i < RTE_MAX_LCORE; i++) {
714                 if (lcore_config[i].core_role != ROLE_SERVICE)
715                         continue;
716
717                 uint32_t reset = 0;
718                 service_dump_calls_per_lcore(f, i, reset);
719         }
720
721         return 0;
722 }