xref: /wlan-dirver/qca-wifi-host-cmn/hif/src/hif_irq_affinity.c (revision c8e2987f9325baadee03d0265544a08c4a0217b0)
1 /*
2  * Copyright (c) 2015-2018 The Linux Foundation. All rights reserved.
3  *
4  * Permission to use, copy, modify, and/or distribute this software for
5  * any purpose with or without fee is hereby granted, provided that the
6  * above copyright notice and this permission notice appear in all
7  * copies.
8  *
9  * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL
10  * WARRANTIES WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED
11  * WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE
12  * AUTHOR BE LIABLE FOR ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL
13  * DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR
14  * PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR OTHER
15  * TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
16  * PERFORMANCE OF THIS SOFTWARE.
17  */
18 
19 /**
20  * DOC: hif_irq_afinity.c
21  *
22  * This irq afinity implementation is os dependent, so this can be treated as
23  * an abstraction layer...  Should this be moved into a /linux folder?
24  */
25 
26 #include <linux/string.h> /* memset */
27 
28 /* Linux headers */
29 #include <linux/cpumask.h>
30 #include <linux/cpufreq.h>
31 #include <linux/cpu.h>
32 #include <linux/topology.h>
33 #include <linux/interrupt.h>
34 #include <linux/irq.h>
35 #ifdef CONFIG_SCHED_CORE_CTL
36 #include <linux/sched/core_ctl.h>
37 #endif
38 #include <linux/pm.h>
39 #include <hif_napi.h>
40 #include <hif_irq_affinity.h>
41 #include <hif_exec.h>
42 #include <hif_main.h>
43 
44 #if defined(FEATURE_NAPI_DEBUG) && defined(HIF_IRQ_AFFINITY)
45 /*
46  * Local functions
47  * - no argument checks, all internal/trusted callers
48  */
49 static void hnc_dump_cpus(struct qca_napi_data *napid)
50 {
51 	hif_napi_stats(napid);
52 }
53 #else
54 static void hnc_dump_cpus(struct qca_napi_data *napid) { /* no-op */ };
55 #endif /* FEATURE_NAPI_DEBUG */
56 
57 #ifdef HIF_IRQ_AFFINITY
58 /**
59  *
60  * hif_exec_event() - reacts to events that impact irq affinity
61  * @hif : pointer to hif context
62  * @evnt: event that has been detected
63  * @data: more data regarding the event
64  *
65  * Description:
66  *   This function handles two types of events:
67  *   1- Events that change the state of NAPI (enabled/disabled):
68  *      {NAPI_EVT_INI_FILE, NAPI_EVT_CMD_STATE}
69  *      The state is retrievable by "hdd_napi_enabled(-1)"
70  *    - NAPI will be on if either INI file is on and it has not been disabled
71  *                                by a subsequent vendor CMD,
72  *                         or     it has been enabled by a vendor CMD.
73  *   2- Events that change the CPU affinity of a NAPI instance/IRQ:
74  *      {NAPI_EVT_TPUT_STATE, NAPI_EVT_CPU_STATE}
75  *    - NAPI will support a throughput mode (HI/LO), kept at napid->napi_mode
76  *    - NAPI will switch throughput mode based on hdd_napi_throughput_policy()
77  *    - In LO tput mode, NAPI will yield control if its interrupts to the system
78  *      management functions. However in HI throughput mode, NAPI will actively
79  *      manage its interrupts/instances (by trying to disperse them out to
80  *      separate performance cores).
81  *    - CPU eligibility is kept up-to-date by NAPI_EVT_CPU_STATE events.
82  *
83  *    + In some cases (roaming peer management is the only case so far), a
84  *      a client can trigger a "SERIALIZE" event. Basically, this means that the
85  *      users is asking NAPI to go into a truly single execution context state.
86  *      So, NAPI indicates to msm-irqbalancer that it wants to be blacklisted,
87  *      (if called for the first time) and then moves all IRQs (for NAPI
88  *      instances) to be collapsed to a single core. If called multiple times,
89  *      it will just re-collapse the CPUs. This is because blacklist-on() API
90  *      is reference-counted, and because the API has already been called.
91  *
92  *      Such a user, should call "DESERIALIZE" (NORMAL) event, to set NAPI to go
93  *      to its "normal" operation. Optionally, they can give a timeout value (in
94  *      multiples of BusBandwidthCheckPeriod -- 100 msecs by default). In this
95  *      case, NAPI will just set the current throughput state to uninitialized
96  *      and set the delay period. Once policy handler is called, it would skip
97  *      applying the policy delay period times, and otherwise apply the policy.
98  *
99  * Return:
100  *  < 0: some error
101  *  = 0: event handled successfully
102  */
103 int hif_exec_event(struct hif_opaque_softc *hif_ctx, enum qca_napi_event event,
104 		   void *data)
105 {
106 	int      rc = 0;
107 	uint32_t prev_state;
108 	struct hif_softc *hif = HIF_GET_SOFTC(hif_ctx);
109 	struct qca_napi_data *napid = &(hif->napi_data);
110 	enum qca_napi_tput_state tput_mode = QCA_NAPI_TPUT_UNINITIALIZED;
111 	enum {
112 		BLACKLIST_NOT_PENDING,
113 		BLACKLIST_ON_PENDING,
114 		BLACKLIST_OFF_PENDING
115 	     } blacklist_pending = BLACKLIST_NOT_PENDING;
116 
117 	NAPI_DEBUG("%s: -->(event=%d, aux=%pK)", __func__, event, data);
118 
119 	qdf_spin_lock_bh(&(napid->lock));
120 	prev_state = napid->state;
121 	switch (event) {
122 	case NAPI_EVT_INI_FILE:
123 	case NAPI_EVT_CMD_STATE:
124 	case NAPI_EVT_INT_STATE:
125 		/* deprecated */
126 		break;
127 
128 	case NAPI_EVT_CPU_STATE: {
129 		int cpu = ((unsigned long int)data >> 16);
130 		int val = ((unsigned long int)data & 0x0ff);
131 
132 		NAPI_DEBUG("%s: evt=CPU_STATE on CPU %d value=%d",
133 			   __func__, cpu, val);
134 
135 		/* state has already been set by hnc_cpu_notify_cb */
136 		if ((val == QCA_NAPI_CPU_DOWN) &&
137 		    (napid->napi_mode == QCA_NAPI_TPUT_HI) && /* we manage */
138 		    (napid->napi_cpu[cpu].napis != 0)) {
139 			NAPI_DEBUG("%s: Migrating NAPIs out of cpu %d",
140 				   __func__, cpu);
141 			rc = hif_exec_cpu_migrate(napid,
142 						  cpu,
143 						  HNC_ACT_RELOCATE);
144 			napid->napi_cpu[cpu].napis = 0;
145 		}
146 		/* in QCA_NAPI_TPUT_LO case, napis MUST == 0 */
147 		break;
148 	}
149 
150 	case NAPI_EVT_TPUT_STATE: {
151 		tput_mode = (enum qca_napi_tput_state)data;
152 		if (tput_mode == QCA_NAPI_TPUT_LO) {
153 			/* from TPUT_HI -> TPUT_LO */
154 			NAPI_DEBUG("%s: Moving to napi_tput_LO state",
155 				   __func__);
156 			blacklist_pending = BLACKLIST_OFF_PENDING;
157 			/*
158 			 * Ideally we should "collapse" interrupts here, since
159 			 * we are "dispersing" interrupts in the "else" case.
160 			 * This allows the possibility that our interrupts may
161 			 * still be on the perf cluster the next time we enter
162 			 * high tput mode. However, the irq_balancer is free
163 			 * to move our interrupts to power cluster once
164 			 * blacklisting has been turned off in the "else" case.
165 			 */
166 		} else {
167 			/* from TPUT_LO -> TPUT->HI */
168 			NAPI_DEBUG("%s: Moving to napi_tput_HI state",
169 				   __func__);
170 			rc = hif_exec_cpu_migrate(napid,
171 						  HNC_ANY_CPU,
172 						  HNC_ACT_DISPERSE);
173 
174 			blacklist_pending = BLACKLIST_ON_PENDING;
175 		}
176 		napid->napi_mode = tput_mode;
177 		break;
178 	}
179 
180 	case NAPI_EVT_USR_SERIAL: {
181 		unsigned long users = (unsigned long)data;
182 
183 		NAPI_DEBUG("%s: User forced SERIALIZATION; users=%ld",
184 			   __func__, users);
185 
186 		rc = hif_exec_cpu_migrate(napid,
187 					  HNC_ANY_CPU,
188 					  HNC_ACT_COLLAPSE);
189 		if ((users == 0) && (rc == 0))
190 			blacklist_pending = BLACKLIST_ON_PENDING;
191 		break;
192 	}
193 	case NAPI_EVT_USR_NORMAL: {
194 		NAPI_DEBUG("%s: User forced DE-SERIALIZATION", __func__);
195 		/*
196 		 * Deserialization timeout is handled at hdd layer;
197 		 * just mark current mode to uninitialized to ensure
198 		 * it will be set when the delay is over
199 		 */
200 		napid->napi_mode = QCA_NAPI_TPUT_UNINITIALIZED;
201 		break;
202 	}
203 	default: {
204 		HIF_ERROR("%s: unknown event: %d (data=0x%0lx)",
205 			  __func__, event, (unsigned long) data);
206 		break;
207 	} /* default */
208 	}; /* switch */
209 
210 
211 	switch (blacklist_pending) {
212 	case BLACKLIST_ON_PENDING:
213 		/* assume the control of WLAN IRQs */
214 		hif_napi_cpu_blacklist(napid, BLACKLIST_ON);
215 		break;
216 	case BLACKLIST_OFF_PENDING:
217 		/* yield the control of WLAN IRQs */
218 		hif_napi_cpu_blacklist(napid, BLACKLIST_OFF);
219 		break;
220 	default: /* nothing to do */
221 		break;
222 	} /* switch blacklist_pending */
223 
224 	qdf_spin_unlock_bh(&(napid->lock));
225 
226 	NAPI_DEBUG("<--[rc=%d]", rc);
227 	return rc;
228 }
229 
230 #endif
231 
232 /**
233  * hncm_migrate_to() - migrates a NAPI to a CPU
234  * @napid: pointer to NAPI block
235  * @ce_id: CE_id of the NAPI instance
236  * @didx : index in the CPU topology table for the CPU to migrate to
237  *
238  * Migrates NAPI (identified by the CE_id) to the destination core
239  * Updates the napi_map of the destination entry
240  *
241  * Return:
242  *  =0 : success
243  *  <0 : error
244  */
245 static int hncm_exec_migrate_to(struct qca_napi_data *napid, uint8_t ctx_id,
246 				int didx)
247 {
248 	struct hif_exec_context *exec_ctx;
249 	int rc = 0;
250 	int status = 0;
251 	int ind;
252 	cpumask_t cpumask;
253 
254 
255 	NAPI_DEBUG("-->%s(napi_cd=%d, didx=%d)", __func__, napi_ce, didx);
256 
257 	cpumask.bits[0] = (1 << didx);
258 	exec_ctx = hif_exec_get_ctx(&napid->hif_softc->osc, ctx_id);
259 	if (exec_ctx == NULL)
260 		return -EINVAL;
261 
262 	for (ind = 0; ind < exec_ctx->numirq; ind++) {
263 		if (exec_ctx->os_irq[ind]) {
264 			irq_modify_status(exec_ctx->os_irq[ind],
265 					  IRQ_NO_BALANCING, 0);
266 			rc = irq_set_affinity_hint(exec_ctx->os_irq[ind],
267 						   &cpumask);
268 			if (rc)
269 				status = rc;
270 		}
271 	}
272 
273 	/* unmark the napis bitmap in the cpu table */
274 	napid->napi_cpu[exec_ctx->cpu].napis &= ~(0x01 << ctx_id);
275 	/* mark the napis bitmap for the new designated cpu */
276 	napid->napi_cpu[didx].napis |= (0x01 << ctx_id);
277 	exec_ctx->cpu = didx;
278 
279 	NAPI_DEBUG("<--%s[%d]", __func__, rc);
280 	return status;
281 }
282 
283 /**
284  * hncm_dest_cpu() - finds a destination CPU for NAPI
285  * @napid: pointer to NAPI block
286  * @act  : RELOCATE | COLLAPSE | DISPERSE
287  *
288  * Finds the designated destionation for the next IRQ.
289  * RELOCATE: translated to either COLLAPSE or DISPERSE based
290  *           on napid->napi_mode (throughput state)
291  * COLLAPSE: All have the same destination: the first online CPU in lilcl
292  * DISPERSE: One of the CPU in bigcl, which has the smallest number of
293  *           NAPIs on it
294  *
295  * Return: >=0 : index in the cpu topology table
296  *       : < 0 : error
297  */
298 static int hncm_dest_cpu(struct qca_napi_data *napid, int act)
299 {
300 	int destidx = -1;
301 	int head, i;
302 
303 	NAPI_DEBUG("-->%s(act=%d)", __func__, act);
304 	if (act == HNC_ACT_RELOCATE) {
305 		if (napid->napi_mode == QCA_NAPI_TPUT_LO)
306 			act = HNC_ACT_COLLAPSE;
307 		else
308 			act = HNC_ACT_DISPERSE;
309 		NAPI_DEBUG("%s: act changed from HNC_ACT_RELOCATE to %d",
310 			   __func__, act);
311 	}
312 	if (act == HNC_ACT_COLLAPSE) {
313 		head = i = napid->lilcl_head;
314 retry_collapse:
315 		while (i >= 0) {
316 			if (napid->napi_cpu[i].state == QCA_NAPI_CPU_UP) {
317 				destidx = i;
318 				break;
319 			}
320 			i = napid->napi_cpu[i].cluster_nxt;
321 		}
322 		if ((destidx < 0) && (head == napid->lilcl_head)) {
323 			NAPI_DEBUG("%s: COLLAPSE: no lilcl dest, try bigcl",
324 				__func__);
325 			head = i = napid->bigcl_head;
326 			goto retry_collapse;
327 		}
328 	} else { /* HNC_ACT_DISPERSE */
329 		int smallest = 99; /* all 32 bits full */
330 		int smallidx = -1;
331 
332 		head = i = napid->bigcl_head;
333 retry_disperse:
334 		while (i >= 0) {
335 			if ((napid->napi_cpu[i].state == QCA_NAPI_CPU_UP) &&
336 			    (hweight32(napid->napi_cpu[i].napis) <= smallest)) {
337 				smallest = napid->napi_cpu[i].napis;
338 				smallidx = i;
339 			}
340 			i = napid->napi_cpu[i].cluster_nxt;
341 		}
342 		destidx = smallidx;
343 		if ((destidx < 0) && (head == napid->bigcl_head)) {
344 			NAPI_DEBUG("%s: DISPERSE: no bigcl dest, try lilcl",
345 				__func__);
346 			head = i = napid->lilcl_head;
347 			goto retry_disperse;
348 		}
349 	}
350 	NAPI_DEBUG("<--%s[dest=%d]", __func__, destidx);
351 	return destidx;
352 }
353 /**
354  * hif_napi_cpu_migrate() - migrate IRQs away
355  * @cpu: -1: all CPUs <n> specific CPU
356  * @act: COLLAPSE | DISPERSE
357  *
358  * Moves IRQs/NAPIs from specific or all CPUs (specified by @cpu) to eligible
359  * cores. Eligible cores are:
360  * act=COLLAPSE -> the first online core of the little cluster
361  * act=DISPERSE -> separate cores of the big cluster, so that each core will
362  *                 host minimum number of NAPIs/IRQs (napid->cpus[cpu].napis)
363  *
364  * Note that this function is called with a spinlock acquired already.
365  *
366  * Return: =0: success
367  *         <0: error
368  */
369 int hif_exec_cpu_migrate(struct qca_napi_data *napid, int cpu, int action)
370 {
371 	int      rc = 0;
372 	struct qca_napi_cpu *cpup;
373 	int      i, dind;
374 	uint32_t napis;
375 
376 
377 	NAPI_DEBUG("-->%s(.., cpu=%d, act=%d)",
378 		   __func__, cpu, action);
379 
380 	if (napid->exec_map == 0) {
381 		NAPI_DEBUG("%s: datapath contexts to disperse", __func__);
382 		goto hncm_return;
383 	}
384 	cpup = napid->napi_cpu;
385 
386 	switch (action) {
387 	case HNC_ACT_RELOCATE:
388 	case HNC_ACT_DISPERSE:
389 	case HNC_ACT_COLLAPSE: {
390 		/* first find the src napi set */
391 		if (cpu == HNC_ANY_CPU)
392 			napis = napid->exec_map;
393 		else
394 			napis = cpup[cpu].napis;
395 		/* then clear the napi bitmap on each CPU */
396 		for (i = 0; i < NR_CPUS; i++)
397 			cpup[i].napis = 0;
398 		/* then for each of the NAPIs to disperse: */
399 		for (i = 0; i < HIF_MAX_GROUP; i++)
400 			if (napis & (1 << i)) {
401 				/* find a destination CPU */
402 				dind = hncm_dest_cpu(napid, action);
403 				if (dind >= 0) {
404 					NAPI_DEBUG("Migrating NAPI ce%d to %d",
405 						   i, dind);
406 					rc = hncm_exec_migrate_to(napid, i,
407 								  dind);
408 				} else {
409 					NAPI_DEBUG("No dest for NAPI ce%d", i);
410 					hnc_dump_cpus(napid);
411 					rc = -1;
412 				}
413 			}
414 		break;
415 	}
416 	default: {
417 		NAPI_DEBUG("%s: bad action: %d\n", __func__, action);
418 		QDF_BUG(0);
419 		break;
420 	}
421 	} /* switch action */
422 
423 hncm_return:
424 	hnc_dump_cpus(napid);
425 	return rc;
426 }
427 
428 
429 /**
430  * hif_exec_bl_irq() - calls irq_modify_status to enable/disable blacklisting
431  * @napid: pointer to qca_napi_data structure
432  * @bl_flag: blacklist flag to enable/disable blacklisting
433  *
434  * The function enables/disables blacklisting for all the copy engine
435  * interrupts on which NAPI is enabled.
436  *
437  * Return: None
438  */
439 static inline void hif_exec_bl_irq(struct qca_napi_data *napid, bool bl_flag)
440 {
441 	int i, j;
442 	struct hif_exec_context *exec_ctx;
443 
444 	for (i = 0; i < HIF_MAX_GROUP; i++) {
445 		/* check if NAPI is enabled on the CE */
446 		if (!(napid->exec_map & (0x01 << i)))
447 			continue;
448 
449 		/*double check that NAPI is allocated for the CE */
450 		exec_ctx = hif_exec_get_ctx(&napid->hif_softc->osc, i);
451 		if (!(exec_ctx))
452 			continue;
453 
454 		if (bl_flag == true)
455 			for (j = 0; j < exec_ctx->numirq; j++)
456 				irq_modify_status(exec_ctx->os_irq[j],
457 						  0, IRQ_NO_BALANCING);
458 		else
459 			for (j = 0; j < exec_ctx->numirq; j++)
460 				irq_modify_status(exec_ctx->os_irq[j],
461 						  IRQ_NO_BALANCING, 0);
462 		HIF_DBG("%s: bl_flag %d CE %d", __func__, bl_flag, i);
463 	}
464 }
465 
466 #ifdef CONFIG_SCHED_CORE_CTL
467 /* Enable this API only if kernel feature - CONFIG_SCHED_CORE_CTL is defined */
468 static inline int hif_napi_core_ctl_set_boost(bool boost)
469 {
470 	return core_ctl_set_boost(boost);
471 }
472 #else
473 static inline int hif_napi_core_ctl_set_boost(bool boost)
474 {
475 	return 0;
476 }
477 #endif
478 
479 /**
480  * hif_napi_cpu_blacklist() - en(dis)ables blacklisting for NAPI RX interrupts.
481  * @napid: pointer to qca_napi_data structure
482  * @op: blacklist operation to perform
483  *
484  * The function enables/disables/queries blacklisting for all CE RX
485  * interrupts with NAPI enabled. Besides blacklisting, it also enables/disables
486  * core_ctl_set_boost.
487  * Once blacklisting is enabled, the interrupts will not be managed by the IRQ
488  * balancer.
489  *
490  * Return: -EINVAL, in case IRQ_BLACKLISTING and CORE_CTL_BOOST is not enabled
491  *         for BLACKLIST_QUERY op - blacklist refcount
492  *         for BLACKLIST_ON op    - return value from core_ctl_set_boost API
493  *         for BLACKLIST_OFF op   - return value from core_ctl_set_boost API
494  */
495 int hif_exec_cpu_blacklist(struct qca_napi_data *napid,
496 			   enum qca_blacklist_op op)
497 {
498 	int rc = 0;
499 	static int ref_count; /* = 0 by the compiler */
500 	uint8_t flags = napid->flags;
501 	bool bl_en = flags & QCA_NAPI_FEATURE_IRQ_BLACKLISTING;
502 	bool ccb_en = flags & QCA_NAPI_FEATURE_CORE_CTL_BOOST;
503 
504 	NAPI_DEBUG("-->%s(%d %d)", __func__, flags, op);
505 
506 	if (!(bl_en && ccb_en)) {
507 		rc = -EINVAL;
508 		goto out;
509 	}
510 
511 	switch (op) {
512 	case BLACKLIST_QUERY:
513 		rc = ref_count;
514 		break;
515 	case BLACKLIST_ON:
516 		ref_count++;
517 		rc = 0;
518 		if (ref_count == 1) {
519 			rc = hif_napi_core_ctl_set_boost(true);
520 			NAPI_DEBUG("boost_on() returns %d - refcnt=%d",
521 				rc, ref_count);
522 			hif_exec_bl_irq(napid, true);
523 		}
524 		break;
525 	case BLACKLIST_OFF:
526 		if (ref_count)
527 			ref_count--;
528 		rc = 0;
529 		if (ref_count == 0) {
530 			rc = hif_napi_core_ctl_set_boost(false);
531 			NAPI_DEBUG("boost_off() returns %d - refcnt=%d",
532 				   rc, ref_count);
533 			hif_exec_bl_irq(napid, false);
534 		}
535 		break;
536 	default:
537 		NAPI_DEBUG("Invalid blacklist op: %d", op);
538 		rc = -EINVAL;
539 	} /* switch */
540 out:
541 	NAPI_DEBUG("<--%s[%d]", __func__, rc);
542 	return rc;
543 }
544 
545