1 /* 2 * Copyright (c) 2015-2021 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_napi.c 21 * 22 * HIF NAPI interface implementation 23 */ 24 25 #include <linux/string.h> /* memset */ 26 27 /* Linux headers */ 28 #include <linux/cpumask.h> 29 #include <linux/cpufreq.h> 30 #include <linux/cpu.h> 31 #include <linux/topology.h> 32 #include <linux/interrupt.h> 33 #ifdef CONFIG_SCHED_CORE_CTL 34 #include <linux/sched/core_ctl.h> 35 #endif 36 #include <pld_common.h> 37 #include <linux/pm.h> 38 39 /* Driver headers */ 40 #include <hif_napi.h> 41 #include <hif_debug.h> 42 #include <hif_io32.h> 43 #include <ce_api.h> 44 #include <ce_internal.h> 45 #include <hif_irq_affinity.h> 46 #include "qdf_cpuhp.h" 47 #include "qdf_module.h" 48 #include "qdf_net_if.h" 49 #include "qdf_dev.h" 50 #include "qdf_irq.h" 51 52 enum napi_decision_vector { 53 HIF_NAPI_NOEVENT = 0, 54 HIF_NAPI_INITED = 1, 55 HIF_NAPI_CONF_UP = 2 56 }; 57 #define ENABLE_NAPI_MASK (HIF_NAPI_INITED | HIF_NAPI_CONF_UP) 58 59 #ifdef RECEIVE_OFFLOAD 60 /** 61 * hif_rxthread_napi_poll() - dummy napi poll for rx_thread NAPI 62 * @napi: Rx_thread NAPI 63 * @budget: NAPI BUDGET 64 * 65 * Return: 0 as it is not supposed to be polled at all as it is not scheduled. 66 */ 67 static int hif_rxthread_napi_poll(struct napi_struct *napi, int budget) 68 { 69 hif_err("This napi_poll should not be polled as we don't schedule it"); 70 QDF_ASSERT(0); 71 return 0; 72 } 73 74 /** 75 * hif_init_rx_thread_napi() - Initialize dummy Rx_thread NAPI 76 * @napii: Handle to napi_info holding rx_thread napi 77 * 78 * Return: None 79 */ 80 static void hif_init_rx_thread_napi(struct qca_napi_info *napii) 81 { 82 struct qdf_net_if *nd = (struct qdf_net_if *)&napii->rx_thread_netdev; 83 84 qdf_net_if_create_dummy_if(nd); 85 netif_napi_add(&napii->rx_thread_netdev, &napii->rx_thread_napi, 86 hif_rxthread_napi_poll, 64); 87 napi_enable(&napii->rx_thread_napi); 88 } 89 90 /** 91 * hif_deinit_rx_thread_napi() - Deinitialize dummy Rx_thread NAPI 92 * @napii: Handle to napi_info holding rx_thread napi 93 * 94 * Return: None 95 */ 96 static void hif_deinit_rx_thread_napi(struct qca_napi_info *napii) 97 { 98 netif_napi_del(&napii->rx_thread_napi); 99 } 100 #else /* RECEIVE_OFFLOAD */ 101 static void hif_init_rx_thread_napi(struct qca_napi_info *napii) 102 { 103 } 104 105 static void hif_deinit_rx_thread_napi(struct qca_napi_info *napii) 106 { 107 } 108 #endif 109 110 /** 111 * hif_napi_create() - creates the NAPI structures for a given CE 112 * @hif : pointer to hif context 113 * @pipe_id: the CE id on which the instance will be created 114 * @poll : poll function to be used for this NAPI instance 115 * @budget : budget to be registered with the NAPI instance 116 * @scale : scale factor on the weight (to scaler budget to 1000) 117 * @flags : feature flags 118 * 119 * Description: 120 * Creates NAPI instances. This function is called 121 * unconditionally during initialization. It creates 122 * napi structures through the proper HTC/HIF calls. 123 * The structures are disabled on creation. 124 * Note that for each NAPI instance a separate dummy netdev is used 125 * 126 * Return: 127 * < 0: error 128 * = 0: <should never happen> 129 * > 0: id of the created object (for multi-NAPI, number of objects created) 130 */ 131 int hif_napi_create(struct hif_opaque_softc *hif_ctx, 132 int (*poll)(struct napi_struct *, int), 133 int budget, 134 int scale, 135 uint8_t flags) 136 { 137 int i; 138 struct qca_napi_data *napid; 139 struct qca_napi_info *napii; 140 struct CE_state *ce_state; 141 struct hif_softc *hif = HIF_GET_SOFTC(hif_ctx); 142 int rc = 0; 143 144 NAPI_DEBUG("-->(budget=%d, scale=%d)", 145 budget, scale); 146 NAPI_DEBUG("hif->napi_data.state = 0x%08x", 147 hif->napi_data.state); 148 NAPI_DEBUG("hif->napi_data.ce_map = 0x%08x", 149 hif->napi_data.ce_map); 150 151 napid = &(hif->napi_data); 152 if (0 == (napid->state & HIF_NAPI_INITED)) { 153 memset(napid, 0, sizeof(struct qca_napi_data)); 154 qdf_spinlock_create(&(napid->lock)); 155 156 napid->state |= HIF_NAPI_INITED; 157 napid->flags = flags; 158 159 rc = hif_napi_cpu_init(hif_ctx); 160 if (rc != 0 && rc != -EALREADY) { 161 hif_err("NAPI_initialization failed(rc=%d)", rc); 162 rc = napid->ce_map; 163 goto hnc_err; 164 } else 165 rc = 0; 166 167 hif_debug("NAPI structures initialized, rc=%d", rc); 168 } 169 for (i = 0; i < hif->ce_count; i++) { 170 ce_state = hif->ce_id_to_state[i]; 171 NAPI_DEBUG("ce %d: htt_rx=%d htt_tx=%d", 172 i, ce_state->htt_rx_data, 173 ce_state->htt_tx_data); 174 if (ce_srng_based(hif)) 175 continue; 176 177 if (!ce_state->htt_rx_data) 178 continue; 179 180 /* Now this is a CE where we need NAPI on */ 181 NAPI_DEBUG("Creating NAPI on pipe %d", i); 182 napii = qdf_mem_malloc(sizeof(*napii)); 183 napid->napis[i] = napii; 184 if (!napii) { 185 rc = -ENOMEM; 186 goto napii_free; 187 } 188 } 189 190 for (i = 0; i < hif->ce_count; i++) { 191 napii = napid->napis[i]; 192 if (!napii) 193 continue; 194 195 NAPI_DEBUG("initializing NAPI for pipe %d", i); 196 memset(napii, 0, sizeof(struct qca_napi_info)); 197 napii->scale = scale; 198 napii->id = NAPI_PIPE2ID(i); 199 napii->hif_ctx = hif_ctx; 200 napii->irq = pld_get_irq(hif->qdf_dev->dev, i); 201 202 if (napii->irq < 0) 203 hif_warn("bad IRQ value for CE %d: %d", i, napii->irq); 204 205 qdf_net_if_create_dummy_if((struct qdf_net_if *)&napii->netdev); 206 207 NAPI_DEBUG("adding napi=%pK to netdev=%pK (poll=%pK, bdgt=%d)", 208 &(napii->napi), &(napii->netdev), poll, budget); 209 netif_napi_add(&(napii->netdev), &(napii->napi), poll, budget); 210 211 NAPI_DEBUG("after napi_add"); 212 NAPI_DEBUG("napi=0x%pK, netdev=0x%pK", 213 &(napii->napi), &(napii->netdev)); 214 NAPI_DEBUG("napi.dev_list.prev=0x%pK, next=0x%pK", 215 napii->napi.dev_list.prev, 216 napii->napi.dev_list.next); 217 NAPI_DEBUG("dev.napi_list.prev=0x%pK, next=0x%pK", 218 napii->netdev.napi_list.prev, 219 napii->netdev.napi_list.next); 220 221 hif_init_rx_thread_napi(napii); 222 napii->lro_ctx = qdf_lro_init(); 223 NAPI_DEBUG("Registering LRO for ce_id %d NAPI callback for %d lro_ctx %pK\n", 224 i, napii->id, napii->lro_ctx); 225 226 /* It is OK to change the state variable below without 227 * protection as there should be no-one around yet 228 */ 229 napid->ce_map |= (0x01 << i); 230 hif_debug("NAPI id %d created for pipe %d", napii->id, i); 231 } 232 233 /* no ces registered with the napi */ 234 if (!ce_srng_based(hif) && napid->ce_map == 0) { 235 hif_warn("no napis created for copy engines"); 236 rc = -EFAULT; 237 goto napii_free; 238 } 239 240 NAPI_DEBUG("napi map = %x", napid->ce_map); 241 NAPI_DEBUG("NAPI ids created for all applicable pipes"); 242 return napid->ce_map; 243 244 napii_free: 245 for (i = 0; i < hif->ce_count; i++) { 246 napii = napid->napis[i]; 247 napid->napis[i] = NULL; 248 if (napii) 249 qdf_mem_free(napii); 250 } 251 252 hnc_err: 253 NAPI_DEBUG("<--napi_instances_map=%x]", napid->ce_map); 254 return rc; 255 } 256 qdf_export_symbol(hif_napi_create); 257 258 #ifdef RECEIVE_OFFLOAD 259 void hif_napi_rx_offld_flush_cb_register(struct hif_opaque_softc *hif_hdl, 260 void (offld_flush_handler)(void *)) 261 { 262 int i; 263 struct CE_state *ce_state; 264 struct hif_softc *scn = HIF_GET_SOFTC(hif_hdl); 265 struct qca_napi_data *napid; 266 struct qca_napi_info *napii; 267 268 if (!scn) { 269 hif_err("hif_state NULL!"); 270 QDF_ASSERT(0); 271 return; 272 } 273 274 napid = hif_napi_get_all(hif_hdl); 275 for (i = 0; i < scn->ce_count; i++) { 276 ce_state = scn->ce_id_to_state[i]; 277 if (ce_state && (ce_state->htt_rx_data)) { 278 napii = napid->napis[i]; 279 napii->offld_flush_cb = offld_flush_handler; 280 hif_debug("Registering offload for ce_id %d NAPI callback for %d flush_cb %pK", 281 i, napii->id, napii->offld_flush_cb); 282 } 283 } 284 } 285 286 void hif_napi_rx_offld_flush_cb_deregister(struct hif_opaque_softc *hif_hdl) 287 { 288 int i; 289 struct CE_state *ce_state; 290 struct hif_softc *scn = HIF_GET_SOFTC(hif_hdl); 291 struct qca_napi_data *napid; 292 struct qca_napi_info *napii; 293 294 if (!scn) { 295 hif_err("hif_state NULL!"); 296 QDF_ASSERT(0); 297 return; 298 } 299 300 napid = hif_napi_get_all(hif_hdl); 301 for (i = 0; i < scn->ce_count; i++) { 302 ce_state = scn->ce_id_to_state[i]; 303 if (ce_state && (ce_state->htt_rx_data)) { 304 napii = napid->napis[i]; 305 hif_debug("deRegistering offld for ce_id %d NAPI callback for %d flush_cb %pK", 306 i, napii->id, napii->offld_flush_cb); 307 /* Not required */ 308 napii->offld_flush_cb = NULL; 309 } 310 } 311 } 312 #endif /* RECEIVE_OFFLOAD */ 313 314 /** 315 * 316 * hif_napi_destroy() - destroys the NAPI structures for a given instance 317 * @hif : pointer to hif context 318 * @ce_id : the CE id whose napi instance will be destroyed 319 * @force : if set, will destroy even if entry is active (de-activates) 320 * 321 * Description: 322 * Destroy a given NAPI instance. This function is called 323 * unconditionally during cleanup. 324 * Refuses to destroy an entry of it is still enabled (unless force=1) 325 * Marks the whole napi_data invalid if all instances are destroyed. 326 * 327 * Return: 328 * -EINVAL: specific entry has not been created 329 * -EPERM : specific entry is still active 330 * 0 < : error 331 * 0 = : success 332 */ 333 int hif_napi_destroy(struct hif_opaque_softc *hif_ctx, 334 uint8_t id, 335 int force) 336 { 337 uint8_t ce = NAPI_ID2PIPE(id); 338 int rc = 0; 339 struct hif_softc *hif = HIF_GET_SOFTC(hif_ctx); 340 341 NAPI_DEBUG("-->(id=%d, force=%d)", id, force); 342 343 if (0 == (hif->napi_data.state & HIF_NAPI_INITED)) { 344 hif_err("NAPI not initialized or entry %d not created", id); 345 rc = -EINVAL; 346 } else if (0 == (hif->napi_data.ce_map & (0x01 << ce))) { 347 hif_err("NAPI instance %d (pipe %d) not created", id, ce); 348 if (hif->napi_data.napis[ce]) 349 hif_err("memory allocated but ce_map not set %d (pipe %d)", 350 id, ce); 351 rc = -EINVAL; 352 } else { 353 struct qca_napi_data *napid; 354 struct qca_napi_info *napii; 355 356 napid = &(hif->napi_data); 357 napii = napid->napis[ce]; 358 if (!napii) { 359 if (napid->ce_map & (0x01 << ce)) 360 hif_err("napii & ce_map out of sync(ce %d)", ce); 361 return -EINVAL; 362 } 363 364 365 if (hif->napi_data.state == HIF_NAPI_CONF_UP) { 366 if (force) { 367 napi_disable(&(napii->napi)); 368 hif_debug("NAPI entry %d force disabled", id); 369 NAPI_DEBUG("NAPI %d force disabled", id); 370 } else { 371 hif_err("Cannot destroy active NAPI %d", id); 372 rc = -EPERM; 373 } 374 } 375 if (0 == rc) { 376 NAPI_DEBUG("before napi_del"); 377 NAPI_DEBUG("napi.dlist.prv=0x%pK, next=0x%pK", 378 napii->napi.dev_list.prev, 379 napii->napi.dev_list.next); 380 NAPI_DEBUG("dev.napi_l.prv=0x%pK, next=0x%pK", 381 napii->netdev.napi_list.prev, 382 napii->netdev.napi_list.next); 383 384 qdf_lro_deinit(napii->lro_ctx); 385 netif_napi_del(&(napii->napi)); 386 hif_deinit_rx_thread_napi(napii); 387 388 napid->ce_map &= ~(0x01 << ce); 389 napid->napis[ce] = NULL; 390 napii->scale = 0; 391 qdf_mem_free(napii); 392 hif_debug("NAPI %d destroyed", id); 393 394 /* if there are no active instances and 395 * if they are all destroyed, 396 * set the whole structure to uninitialized state 397 */ 398 if (napid->ce_map == 0) { 399 rc = hif_napi_cpu_deinit(hif_ctx); 400 /* caller is tolerant to receiving !=0 rc */ 401 402 qdf_spinlock_destroy(&(napid->lock)); 403 memset(napid, 404 0, sizeof(struct qca_napi_data)); 405 hif_debug("no NAPI instances. Zapped"); 406 } 407 } 408 } 409 410 return rc; 411 } 412 qdf_export_symbol(hif_napi_destroy); 413 414 #ifdef FEATURE_LRO 415 void *hif_napi_get_lro_info(struct hif_opaque_softc *hif_hdl, int napi_id) 416 { 417 struct hif_softc *scn = HIF_GET_SOFTC(hif_hdl); 418 struct qca_napi_data *napid; 419 struct qca_napi_info *napii; 420 421 napid = &(scn->napi_data); 422 napii = napid->napis[NAPI_ID2PIPE(napi_id)]; 423 424 if (napii) 425 return napii->lro_ctx; 426 return 0; 427 } 428 #endif 429 430 /** 431 * 432 * hif_napi_get_all() - returns the address of the whole HIF NAPI structure 433 * @hif: pointer to hif context 434 * 435 * Description: 436 * Returns the address of the whole structure 437 * 438 * Return: 439 * <addr>: address of the whole HIF NAPI structure 440 */ 441 inline struct qca_napi_data *hif_napi_get_all(struct hif_opaque_softc *hif_ctx) 442 { 443 struct hif_softc *hif = HIF_GET_SOFTC(hif_ctx); 444 445 return &(hif->napi_data); 446 } 447 448 struct qca_napi_info *hif_get_napi(int napi_id, struct qca_napi_data *napid) 449 { 450 int id = NAPI_ID2PIPE(napi_id); 451 452 return napid->napis[id]; 453 } 454 455 /** 456 * 457 * hif_napi_event() - reacts to events that impact NAPI 458 * @hif : pointer to hif context 459 * @evnt: event that has been detected 460 * @data: more data regarding the event 461 * 462 * Description: 463 * This function handles two types of events: 464 * 1- Events that change the state of NAPI (enabled/disabled): 465 * {NAPI_EVT_INI_FILE, NAPI_EVT_CMD_STATE} 466 * The state is retrievable by "hdd_napi_enabled(-1)" 467 * - NAPI will be on if either INI file is on and it has not been disabled 468 * by a subsequent vendor CMD, 469 * or it has been enabled by a vendor CMD. 470 * 2- Events that change the CPU affinity of a NAPI instance/IRQ: 471 * {NAPI_EVT_TPUT_STATE, NAPI_EVT_CPU_STATE} 472 * - NAPI will support a throughput mode (HI/LO), kept at napid->napi_mode 473 * - NAPI will switch throughput mode based on hdd_napi_throughput_policy() 474 * - In LO tput mode, NAPI will yield control if its interrupts to the system 475 * management functions. However in HI throughput mode, NAPI will actively 476 * manage its interrupts/instances (by trying to disperse them out to 477 * separate performance cores). 478 * - CPU eligibility is kept up-to-date by NAPI_EVT_CPU_STATE events. 479 * 480 * + In some cases (roaming peer management is the only case so far), a 481 * a client can trigger a "SERIALIZE" event. Basically, this means that the 482 * users is asking NAPI to go into a truly single execution context state. 483 * So, NAPI indicates to msm-irqbalancer that it wants to be blacklisted, 484 * (if called for the first time) and then moves all IRQs (for NAPI 485 * instances) to be collapsed to a single core. If called multiple times, 486 * it will just re-collapse the CPUs. This is because blacklist-on() API 487 * is reference-counted, and because the API has already been called. 488 * 489 * Such a user, should call "DESERIALIZE" (NORMAL) event, to set NAPI to go 490 * to its "normal" operation. Optionally, they can give a timeout value (in 491 * multiples of BusBandwidthCheckPeriod -- 100 msecs by default). In this 492 * case, NAPI will just set the current throughput state to uninitialized 493 * and set the delay period. Once policy handler is called, it would skip 494 * applying the policy delay period times, and otherwise apply the policy. 495 * 496 * Return: 497 * < 0: some error 498 * = 0: event handled successfully 499 */ 500 int hif_napi_event(struct hif_opaque_softc *hif_ctx, enum qca_napi_event event, 501 void *data) 502 { 503 int rc = 0; 504 uint32_t prev_state; 505 int i; 506 bool state_changed; 507 struct napi_struct *napi; 508 struct hif_softc *hif = HIF_GET_SOFTC(hif_ctx); 509 struct qca_napi_data *napid = &(hif->napi_data); 510 enum qca_napi_tput_state tput_mode = QCA_NAPI_TPUT_UNINITIALIZED; 511 enum { 512 BLACKLIST_NOT_PENDING, 513 BLACKLIST_ON_PENDING, 514 BLACKLIST_OFF_PENDING 515 } blacklist_pending = BLACKLIST_NOT_PENDING; 516 517 NAPI_DEBUG("%s: -->(event=%d, aux=%pK)", __func__, event, data); 518 519 if (ce_srng_based(hif)) 520 return hif_exec_event(hif_ctx, event, data); 521 522 if ((napid->state & HIF_NAPI_INITED) == 0) { 523 NAPI_DEBUG("%s: got event when NAPI not initialized", 524 __func__); 525 return -EINVAL; 526 } 527 qdf_spin_lock_bh(&(napid->lock)); 528 prev_state = napid->state; 529 switch (event) { 530 case NAPI_EVT_INI_FILE: 531 case NAPI_EVT_CMD_STATE: 532 case NAPI_EVT_INT_STATE: { 533 int on = (data != ((void *)0)); 534 535 hif_debug("recved evnt: STATE_CMD %d; v = %d (state=0x%0x)", 536 event, on, prev_state); 537 if (on) 538 if (prev_state & HIF_NAPI_CONF_UP) { 539 hif_debug("Duplicate NAPI conf ON msg"); 540 } else { 541 hif_debug("Setting state to ON"); 542 napid->state |= HIF_NAPI_CONF_UP; 543 } 544 else /* off request */ 545 if (prev_state & HIF_NAPI_CONF_UP) { 546 hif_debug("Setting state to OFF"); 547 napid->state &= ~HIF_NAPI_CONF_UP; 548 } else { 549 hif_debug("Duplicate NAPI conf OFF msg"); 550 } 551 break; 552 } 553 /* case NAPI_INIT_FILE/CMD_STATE */ 554 555 case NAPI_EVT_CPU_STATE: { 556 int cpu = ((unsigned long int)data >> 16); 557 int val = ((unsigned long int)data & 0x0ff); 558 559 NAPI_DEBUG("%s: evt=CPU_STATE on CPU %d value=%d", 560 __func__, cpu, val); 561 562 /* state has already been set by hnc_cpu_notify_cb */ 563 if ((val == QCA_NAPI_CPU_DOWN) && 564 (napid->napi_mode == QCA_NAPI_TPUT_HI) && /* we manage */ 565 (napid->napi_cpu[cpu].napis != 0)) { 566 NAPI_DEBUG("%s: Migrating NAPIs out of cpu %d", 567 __func__, cpu); 568 rc = hif_napi_cpu_migrate(napid, 569 cpu, 570 HNC_ACT_RELOCATE); 571 napid->napi_cpu[cpu].napis = 0; 572 } 573 /* in QCA_NAPI_TPUT_LO case, napis MUST == 0 */ 574 break; 575 } 576 577 case NAPI_EVT_TPUT_STATE: { 578 tput_mode = (enum qca_napi_tput_state)data; 579 if (tput_mode == QCA_NAPI_TPUT_LO) { 580 /* from TPUT_HI -> TPUT_LO */ 581 NAPI_DEBUG("%s: Moving to napi_tput_LO state", 582 __func__); 583 blacklist_pending = BLACKLIST_OFF_PENDING; 584 /* 585 * Ideally we should "collapse" interrupts here, since 586 * we are "dispersing" interrupts in the "else" case. 587 * This allows the possibility that our interrupts may 588 * still be on the perf cluster the next time we enter 589 * high tput mode. However, the irq_balancer is free 590 * to move our interrupts to power cluster once 591 * blacklisting has been turned off in the "else" case. 592 */ 593 } else { 594 /* from TPUT_LO -> TPUT->HI */ 595 NAPI_DEBUG("%s: Moving to napi_tput_HI state", 596 __func__); 597 rc = hif_napi_cpu_migrate(napid, 598 HNC_ANY_CPU, 599 HNC_ACT_DISPERSE); 600 601 blacklist_pending = BLACKLIST_ON_PENDING; 602 } 603 napid->napi_mode = tput_mode; 604 break; 605 } 606 607 case NAPI_EVT_USR_SERIAL: { 608 unsigned long users = (unsigned long)data; 609 610 NAPI_DEBUG("%s: User forced SERIALIZATION; users=%ld", 611 __func__, users); 612 613 rc = hif_napi_cpu_migrate(napid, 614 HNC_ANY_CPU, 615 HNC_ACT_COLLAPSE); 616 if ((users == 0) && (rc == 0)) 617 blacklist_pending = BLACKLIST_ON_PENDING; 618 break; 619 } 620 case NAPI_EVT_USR_NORMAL: { 621 NAPI_DEBUG("%s: User forced DE-SERIALIZATION", __func__); 622 if (!napid->user_cpu_affin_mask) 623 blacklist_pending = BLACKLIST_OFF_PENDING; 624 /* 625 * Deserialization timeout is handled at hdd layer; 626 * just mark current mode to uninitialized to ensure 627 * it will be set when the delay is over 628 */ 629 napid->napi_mode = QCA_NAPI_TPUT_UNINITIALIZED; 630 break; 631 } 632 default: { 633 hif_err("Unknown event: %d (data=0x%0lx)", 634 event, (unsigned long) data); 635 break; 636 } /* default */ 637 }; /* switch */ 638 639 640 switch (blacklist_pending) { 641 case BLACKLIST_ON_PENDING: 642 /* assume the control of WLAN IRQs */ 643 hif_napi_cpu_blacklist(napid, BLACKLIST_ON); 644 break; 645 case BLACKLIST_OFF_PENDING: 646 /* yield the control of WLAN IRQs */ 647 hif_napi_cpu_blacklist(napid, BLACKLIST_OFF); 648 break; 649 default: /* nothing to do */ 650 break; 651 } /* switch blacklist_pending */ 652 653 /* we want to perform the comparison in lock: 654 * there is a possiblity of hif_napi_event get called 655 * from two different contexts (driver unload and cpu hotplug 656 * notification) and napid->state get changed 657 * in driver unload context and can lead to race condition 658 * in cpu hotplug context. Therefore, perform the napid->state 659 * comparison before releasing lock. 660 */ 661 state_changed = (prev_state != napid->state); 662 qdf_spin_unlock_bh(&(napid->lock)); 663 664 if (state_changed) { 665 if (napid->state == ENABLE_NAPI_MASK) { 666 rc = 1; 667 for (i = 0; i < CE_COUNT_MAX; i++) { 668 struct qca_napi_info *napii = napid->napis[i]; 669 if (napii) { 670 napi = &(napii->napi); 671 NAPI_DEBUG("%s: enabling NAPI %d", 672 __func__, i); 673 napi_enable(napi); 674 } 675 } 676 } else { 677 rc = 0; 678 for (i = 0; i < CE_COUNT_MAX; i++) { 679 struct qca_napi_info *napii = napid->napis[i]; 680 if (napii) { 681 napi = &(napii->napi); 682 NAPI_DEBUG("%s: disabling NAPI %d", 683 __func__, i); 684 napi_disable(napi); 685 /* in case it is affined, remove it */ 686 qdf_dev_set_irq_affinity(napii->irq, 687 NULL); 688 } 689 } 690 } 691 } else { 692 hif_debug("no change in hif napi state (still %d)", prev_state); 693 } 694 695 NAPI_DEBUG("<--[rc=%d]", rc); 696 return rc; 697 } 698 qdf_export_symbol(hif_napi_event); 699 700 /** 701 * hif_napi_enabled() - checks whether NAPI is enabled for given ce or not 702 * @hif: hif context 703 * @ce : CE instance (or -1, to check if any CEs are enabled) 704 * 705 * Return: bool 706 */ 707 int hif_napi_enabled(struct hif_opaque_softc *hif_ctx, int ce) 708 { 709 int rc; 710 struct hif_softc *hif = HIF_GET_SOFTC(hif_ctx); 711 712 if (-1 == ce) 713 rc = ((hif->napi_data.state == ENABLE_NAPI_MASK)); 714 else 715 rc = ((hif->napi_data.state == ENABLE_NAPI_MASK) && 716 (hif->napi_data.ce_map & (0x01 << ce))); 717 return rc; 718 } 719 qdf_export_symbol(hif_napi_enabled); 720 721 /** 722 * hif_napi_created() - checks whether NAPI is created for given ce or not 723 * @hif: hif context 724 * @ce : CE instance 725 * 726 * Return: bool 727 */ 728 bool hif_napi_created(struct hif_opaque_softc *hif_ctx, int ce) 729 { 730 int rc; 731 struct hif_softc *hif = HIF_GET_SOFTC(hif_ctx); 732 733 rc = (hif->napi_data.ce_map & (0x01 << ce)); 734 735 return !!rc; 736 } 737 qdf_export_symbol(hif_napi_created); 738 739 /** 740 * hif_napi_enable_irq() - enables bus interrupts after napi_complete 741 * 742 * @hif: hif context 743 * @id : id of NAPI instance calling this (used to determine the CE) 744 * 745 * Return: void 746 */ 747 inline void hif_napi_enable_irq(struct hif_opaque_softc *hif, int id) 748 { 749 struct hif_softc *scn = HIF_GET_SOFTC(hif); 750 751 hif_irq_enable(scn, NAPI_ID2PIPE(id)); 752 } 753 754 755 /** 756 * hif_napi_schedule() - schedules napi, updates stats 757 * @scn: hif context 758 * @ce_id: index of napi instance 759 * 760 * Return: false if napi didn't enable or already scheduled, otherwise true 761 */ 762 bool hif_napi_schedule(struct hif_opaque_softc *hif_ctx, int ce_id) 763 { 764 int cpu = smp_processor_id(); 765 struct hif_softc *scn = HIF_GET_SOFTC(hif_ctx); 766 struct qca_napi_info *napii; 767 768 napii = scn->napi_data.napis[ce_id]; 769 if (qdf_unlikely(!napii)) { 770 hif_err("scheduling unallocated napi (ce:%d)", ce_id); 771 qdf_atomic_dec(&scn->active_tasklet_cnt); 772 return false; 773 } 774 775 if (test_bit(NAPI_STATE_SCHED, &napii->napi.state)) { 776 NAPI_DEBUG("napi scheduled, return"); 777 qdf_atomic_dec(&scn->active_tasklet_cnt); 778 return false; 779 } 780 781 hif_record_ce_desc_event(scn, ce_id, NAPI_SCHEDULE, 782 NULL, NULL, 0, 0); 783 napii->stats[cpu].napi_schedules++; 784 NAPI_DEBUG("scheduling napi %d (ce:%d)", napii->id, ce_id); 785 napi_schedule(&(napii->napi)); 786 787 return true; 788 } 789 qdf_export_symbol(hif_napi_schedule); 790 791 /** 792 * hif_napi_correct_cpu() - correct the interrupt affinity for napi if needed 793 * @napi_info: pointer to qca_napi_info for the napi instance 794 * 795 * Return: true => interrupt already on correct cpu, no correction needed 796 * false => interrupt on wrong cpu, correction done for cpu affinity 797 * of the interrupt 798 */ 799 static inline 800 bool hif_napi_correct_cpu(struct qca_napi_info *napi_info) 801 { 802 bool right_cpu = true; 803 int rc = 0; 804 int cpu; 805 struct qca_napi_data *napid; 806 QDF_STATUS ret; 807 808 napid = hif_napi_get_all(GET_HIF_OPAQUE_HDL(napi_info->hif_ctx)); 809 810 if (napid->flags & QCA_NAPI_FEATURE_CPU_CORRECTION) { 811 812 cpu = qdf_get_cpu(); 813 if (unlikely((hif_napi_cpu_blacklist(napid, 814 BLACKLIST_QUERY) > 0) && 815 (cpu != napi_info->cpu))) { 816 right_cpu = false; 817 818 NAPI_DEBUG("interrupt on wrong CPU, correcting"); 819 napi_info->cpumask.bits[0] = (0x01 << napi_info->cpu); 820 821 qdf_dev_modify_irq_status(napi_info->irq, 822 QDF_IRQ_NO_BALANCING, 0); 823 ret = qdf_dev_set_irq_affinity(napi_info->irq, 824 (struct qdf_cpu_mask *) 825 &napi_info->cpumask); 826 rc = qdf_status_to_os_return(ret); 827 qdf_dev_modify_irq_status(napi_info->irq, 0, 828 QDF_IRQ_NO_BALANCING); 829 830 if (rc) 831 hif_err("Setting irq affinity hint: %d", rc); 832 else 833 napi_info->stats[cpu].cpu_corrected++; 834 } 835 } 836 return right_cpu; 837 } 838 839 #ifdef RECEIVE_OFFLOAD 840 /** 841 * hif_napi_offld_flush_cb() - Call upper layer flush callback 842 * @napi_info: Handle to hif_napi_info 843 * 844 * Return: None 845 */ 846 static void hif_napi_offld_flush_cb(struct qca_napi_info *napi_info) 847 { 848 if (napi_info->offld_flush_cb) 849 napi_info->offld_flush_cb(napi_info); 850 } 851 #else 852 static void hif_napi_offld_flush_cb(struct qca_napi_info *napi_info) 853 { 854 } 855 #endif 856 857 /** 858 * hif_napi_poll() - NAPI poll routine 859 * @napi : pointer to NAPI struct as kernel holds it 860 * @budget: 861 * 862 * This is the body of the poll function. 863 * The poll function is called by kernel. So, there is a wrapper 864 * function in HDD, which in turn calls this function. 865 * Two main reasons why the whole thing is not implemented in HDD: 866 * a) references to things like ce_service that HDD is not aware of 867 * b) proximity to the implementation of ce_tasklet, which the body 868 * of this function should be very close to. 869 * 870 * NOTE TO THE MAINTAINER: 871 * Consider this function and ce_tasklet very tightly coupled pairs. 872 * Any changes to ce_tasklet or this function may likely need to be 873 * reflected in the counterpart. 874 * 875 * Returns: 876 * int: the amount of work done in this poll (<= budget) 877 */ 878 int hif_napi_poll(struct hif_opaque_softc *hif_ctx, 879 struct napi_struct *napi, 880 int budget) 881 { 882 int rc = 0; /* default: no work done, also takes care of error */ 883 int normalized = 0; 884 int bucket; 885 int cpu = smp_processor_id(); 886 bool poll_on_right_cpu; 887 struct hif_softc *hif = HIF_GET_SOFTC(hif_ctx); 888 struct qca_napi_info *napi_info; 889 struct CE_state *ce_state = NULL; 890 891 if (unlikely(!hif)) { 892 hif_err("hif context is NULL"); 893 QDF_ASSERT(0); 894 goto out; 895 } 896 897 napi_info = (struct qca_napi_info *) 898 container_of(napi, struct qca_napi_info, napi); 899 900 NAPI_DEBUG("%s -->(napi(%d, irq=%d), budget=%d)", 901 __func__, napi_info->id, napi_info->irq, budget); 902 903 napi_info->stats[cpu].napi_polls++; 904 905 hif_record_ce_desc_event(hif, NAPI_ID2PIPE(napi_info->id), 906 NAPI_POLL_ENTER, NULL, NULL, cpu, 0); 907 908 rc = ce_per_engine_service(hif, NAPI_ID2PIPE(napi_info->id)); 909 NAPI_DEBUG("%s: ce_per_engine_service processed %d msgs", 910 __func__, rc); 911 912 hif_napi_offld_flush_cb(napi_info); 913 914 /* do not return 0, if there was some work done, 915 * even if it is below the scale 916 */ 917 if (rc) { 918 napi_info->stats[cpu].napi_workdone += rc; 919 normalized = (rc / napi_info->scale); 920 if (normalized == 0) 921 normalized++; 922 bucket = (normalized - 1) / 923 (QCA_NAPI_BUDGET / QCA_NAPI_NUM_BUCKETS); 924 if (bucket >= QCA_NAPI_NUM_BUCKETS) { 925 bucket = QCA_NAPI_NUM_BUCKETS - 1; 926 hif_err("Bad bucket#(%d) > QCA_NAPI_NUM_BUCKETS(%d)" 927 " normalized %d, napi budget %d", 928 bucket, QCA_NAPI_NUM_BUCKETS, 929 normalized, QCA_NAPI_BUDGET); 930 } 931 napi_info->stats[cpu].napi_budget_uses[bucket]++; 932 } else { 933 /* if ce_per engine reports 0, then poll should be terminated */ 934 NAPI_DEBUG("%s:%d: nothing processed by CE. Completing NAPI", 935 __func__, __LINE__); 936 } 937 938 ce_state = hif->ce_id_to_state[NAPI_ID2PIPE(napi_info->id)]; 939 940 /* 941 * Not using the API hif_napi_correct_cpu directly in the if statement 942 * below since the API may not get evaluated if put at the end if any 943 * prior condition would evaluate to be true. The CPU correction 944 * check should kick in every poll. 945 */ 946 #ifdef NAPI_YIELD_BUDGET_BASED 947 if (ce_state && (ce_state->force_break || 0 == rc)) { 948 #else 949 poll_on_right_cpu = hif_napi_correct_cpu(napi_info); 950 if ((ce_state) && 951 (!ce_check_rx_pending(ce_state) || (0 == rc) || 952 !poll_on_right_cpu)) { 953 #endif 954 napi_info->stats[cpu].napi_completes++; 955 #ifdef NAPI_YIELD_BUDGET_BASED 956 ce_state->force_break = 0; 957 #endif 958 959 hif_record_ce_desc_event(hif, ce_state->id, NAPI_COMPLETE, 960 NULL, NULL, 0, 0); 961 if (normalized >= budget) 962 normalized = budget - 1; 963 964 napi_complete(napi); 965 /* enable interrupts */ 966 hif_napi_enable_irq(hif_ctx, napi_info->id); 967 /* support suspend/resume */ 968 qdf_atomic_dec(&(hif->active_tasklet_cnt)); 969 970 NAPI_DEBUG("%s:%d: napi_complete + enabling the interrupts", 971 __func__, __LINE__); 972 } else { 973 /* 4.4 kernel NAPI implementation requires drivers to 974 * return full work when they ask to be re-scheduled, 975 * or napi_complete and re-start with a fresh interrupt 976 */ 977 normalized = budget; 978 } 979 980 hif_record_ce_desc_event(hif, NAPI_ID2PIPE(napi_info->id), 981 NAPI_POLL_EXIT, NULL, NULL, normalized, 0); 982 983 NAPI_DEBUG("%s <--[normalized=%d]", __func__, normalized); 984 return normalized; 985 out: 986 return rc; 987 } 988 qdf_export_symbol(hif_napi_poll); 989 990 void hif_update_napi_max_poll_time(struct CE_state *ce_state, 991 int ce_id, 992 int cpu_id) 993 { 994 struct hif_softc *hif; 995 struct qca_napi_info *napi_info; 996 unsigned long long napi_poll_time = qdf_time_sched_clock() - 997 ce_state->ce_service_start_time; 998 999 hif = ce_state->scn; 1000 napi_info = hif->napi_data.napis[ce_id]; 1001 if (napi_poll_time > 1002 napi_info->stats[cpu_id].napi_max_poll_time) 1003 napi_info->stats[cpu_id].napi_max_poll_time = napi_poll_time; 1004 } 1005 qdf_export_symbol(hif_update_napi_max_poll_time); 1006 1007 #ifdef HIF_IRQ_AFFINITY 1008 /** 1009 * 1010 * hif_napi_update_yield_stats() - update NAPI yield related stats 1011 * @cpu_id: CPU ID for which stats needs to be updates 1012 * @ce_id: Copy Engine ID for which yield stats needs to be updates 1013 * @time_limit_reached: indicates whether the time limit was reached 1014 * @rxpkt_thresh_reached: indicates whether rx packet threshold was reached 1015 * 1016 * Return: None 1017 */ 1018 void hif_napi_update_yield_stats(struct CE_state *ce_state, 1019 bool time_limit_reached, 1020 bool rxpkt_thresh_reached) 1021 { 1022 struct hif_softc *hif; 1023 struct qca_napi_data *napi_data = NULL; 1024 int ce_id = 0; 1025 int cpu_id = 0; 1026 1027 if (unlikely(!ce_state)) { 1028 QDF_ASSERT(ce_state); 1029 return; 1030 } 1031 1032 hif = ce_state->scn; 1033 1034 if (unlikely(!hif)) { 1035 QDF_ASSERT(hif); 1036 return; 1037 } 1038 napi_data = &(hif->napi_data); 1039 if (unlikely(!napi_data)) { 1040 QDF_ASSERT(napi_data); 1041 return; 1042 } 1043 1044 ce_id = ce_state->id; 1045 cpu_id = qdf_get_cpu(); 1046 1047 if (unlikely(!napi_data->napis[ce_id])) { 1048 return; 1049 } 1050 1051 if (time_limit_reached) 1052 napi_data->napis[ce_id]->stats[cpu_id].time_limit_reached++; 1053 else 1054 napi_data->napis[ce_id]->stats[cpu_id].rxpkt_thresh_reached++; 1055 1056 hif_update_napi_max_poll_time(ce_state, ce_id, 1057 cpu_id); 1058 } 1059 1060 /** 1061 * 1062 * hif_napi_stats() - display NAPI CPU statistics 1063 * @napid: pointer to qca_napi_data 1064 * 1065 * Description: 1066 * Prints the various CPU cores on which the NAPI instances /CEs interrupts 1067 * are being executed. Can be called from outside NAPI layer. 1068 * 1069 * Return: None 1070 */ 1071 void hif_napi_stats(struct qca_napi_data *napid) 1072 { 1073 int i; 1074 struct qca_napi_cpu *cpu; 1075 1076 if (!napid) { 1077 qdf_debug("%s: napiid struct is null", __func__); 1078 return; 1079 } 1080 1081 cpu = napid->napi_cpu; 1082 qdf_debug("NAPI CPU TABLE"); 1083 qdf_debug("lilclhead=%d, bigclhead=%d", 1084 napid->lilcl_head, napid->bigcl_head); 1085 for (i = 0; i < NR_CPUS; i++) { 1086 qdf_debug("CPU[%02d]: state:%d crid=%02d clid=%02d crmk:0x%0lx thmk:0x%0lx frq:%d napi = 0x%08x lnk:%d", 1087 i, 1088 cpu[i].state, cpu[i].core_id, cpu[i].cluster_id, 1089 cpu[i].core_mask.bits[0], 1090 cpu[i].thread_mask.bits[0], 1091 cpu[i].max_freq, cpu[i].napis, 1092 cpu[i].cluster_nxt); 1093 } 1094 } 1095 1096 #ifdef FEATURE_NAPI_DEBUG 1097 /* 1098 * Local functions 1099 * - no argument checks, all internal/trusted callers 1100 */ 1101 static void hnc_dump_cpus(struct qca_napi_data *napid) 1102 { 1103 hif_napi_stats(napid); 1104 } 1105 #else 1106 static void hnc_dump_cpus(struct qca_napi_data *napid) { /* no-op */ }; 1107 #endif /* FEATURE_NAPI_DEBUG */ 1108 /** 1109 * hnc_link_clusters() - partitions to cpu table into clusters 1110 * @napid: pointer to NAPI data 1111 * 1112 * Takes in a CPU topology table and builds two linked lists 1113 * (big cluster cores, list-head at bigcl_head, and little cluster 1114 * cores, list-head at lilcl_head) out of it. 1115 * 1116 * If there are more than two clusters: 1117 * - bigcl_head and lilcl_head will be different, 1118 * - the cluster with highest cpufreq will be considered the "big" cluster. 1119 * If there are more than one with the highest frequency, the *last* of such 1120 * clusters will be designated as the "big cluster" 1121 * - the cluster with lowest cpufreq will be considered the "li'l" cluster. 1122 * If there are more than one clusters with the lowest cpu freq, the *first* 1123 * of such clusters will be designated as the "little cluster" 1124 * - We only support up to 32 clusters 1125 * Return: 0 : OK 1126 * !0: error (at least one of lil/big clusters could not be found) 1127 */ 1128 #define HNC_MIN_CLUSTER 0 1129 #define HNC_MAX_CLUSTER 1 1130 static int hnc_link_clusters(struct qca_napi_data *napid) 1131 { 1132 int rc = 0; 1133 1134 int i; 1135 int it = 0; 1136 uint32_t cl_done = 0x0; 1137 int cl, curcl, curclhead = 0; 1138 int more; 1139 unsigned int lilfrq = INT_MAX; 1140 unsigned int bigfrq = 0; 1141 unsigned int clfrq = 0; 1142 int prev = 0; 1143 struct qca_napi_cpu *cpus = napid->napi_cpu; 1144 1145 napid->lilcl_head = napid->bigcl_head = -1; 1146 1147 do { 1148 more = 0; 1149 it++; curcl = -1; 1150 for (i = 0; i < NR_CPUS; i++) { 1151 cl = cpus[i].cluster_id; 1152 NAPI_DEBUG("Processing cpu[%d], cluster=%d\n", 1153 i, cl); 1154 if ((cl < HNC_MIN_CLUSTER) || (cl > HNC_MAX_CLUSTER)) { 1155 NAPI_DEBUG("Bad cluster (%d). SKIPPED\n", cl); 1156 /* continue if ASSERTs are disabled */ 1157 continue; 1158 }; 1159 if (cpumask_weight(&(cpus[i].core_mask)) == 0) { 1160 NAPI_DEBUG("Core mask 0. SKIPPED\n"); 1161 continue; 1162 } 1163 if (cl_done & (0x01 << cl)) { 1164 NAPI_DEBUG("Cluster already processed. SKIPPED\n"); 1165 continue; 1166 } else { 1167 if (more == 0) { 1168 more = 1; 1169 curcl = cl; 1170 curclhead = i; /* row */ 1171 clfrq = cpus[i].max_freq; 1172 prev = -1; 1173 }; 1174 if ((curcl >= 0) && (curcl != cl)) { 1175 NAPI_DEBUG("Entry cl(%d) != curcl(%d). SKIPPED\n", 1176 cl, curcl); 1177 continue; 1178 } 1179 if (cpus[i].max_freq != clfrq) 1180 NAPI_DEBUG("WARN: frq(%d)!=clfrq(%d)\n", 1181 cpus[i].max_freq, clfrq); 1182 if (clfrq >= bigfrq) { 1183 bigfrq = clfrq; 1184 napid->bigcl_head = curclhead; 1185 NAPI_DEBUG("bigcl=%d\n", curclhead); 1186 } 1187 if (clfrq < lilfrq) { 1188 lilfrq = clfrq; 1189 napid->lilcl_head = curclhead; 1190 NAPI_DEBUG("lilcl=%d\n", curclhead); 1191 } 1192 if (prev != -1) 1193 cpus[prev].cluster_nxt = i; 1194 1195 prev = i; 1196 } 1197 } 1198 if (curcl >= 0) 1199 cl_done |= (0x01 << curcl); 1200 1201 } while (more); 1202 1203 if (qdf_unlikely((napid->lilcl_head < 0) && (napid->bigcl_head < 0))) 1204 rc = -EFAULT; 1205 1206 hnc_dump_cpus(napid); /* if NAPI_DEBUG */ 1207 return rc; 1208 } 1209 #undef HNC_MIN_CLUSTER 1210 #undef HNC_MAX_CLUSTER 1211 1212 /* 1213 * hotplug function group 1214 */ 1215 1216 /** 1217 * hnc_cpu_online_cb() - handles CPU hotplug "up" events 1218 * @context: the associated HIF context 1219 * @cpu: the CPU Id of the CPU the event happened on 1220 * 1221 * Return: None 1222 */ 1223 static void hnc_cpu_online_cb(void *context, uint32_t cpu) 1224 { 1225 struct hif_softc *hif = context; 1226 struct qca_napi_data *napid = &hif->napi_data; 1227 1228 if (cpu >= NR_CPUS) 1229 return; 1230 1231 NAPI_DEBUG("-->%s(act=online, cpu=%u)", __func__, cpu); 1232 1233 napid->napi_cpu[cpu].state = QCA_NAPI_CPU_UP; 1234 NAPI_DEBUG("%s: CPU %u marked %d", 1235 __func__, cpu, napid->napi_cpu[cpu].state); 1236 1237 NAPI_DEBUG("<--%s", __func__); 1238 } 1239 1240 /** 1241 * hnc_cpu_before_offline_cb() - handles CPU hotplug "prepare down" events 1242 * @context: the associated HIF context 1243 * @cpu: the CPU Id of the CPU the event happened on 1244 * 1245 * On transtion to offline, we act on PREP events, because we may need to move 1246 * the irqs/NAPIs to another CPU before it is actually off-lined. 1247 * 1248 * Return: None 1249 */ 1250 static void hnc_cpu_before_offline_cb(void *context, uint32_t cpu) 1251 { 1252 struct hif_softc *hif = context; 1253 struct qca_napi_data *napid = &hif->napi_data; 1254 1255 if (cpu >= NR_CPUS) 1256 return; 1257 1258 NAPI_DEBUG("-->%s(act=before_offline, cpu=%u)", __func__, cpu); 1259 1260 napid->napi_cpu[cpu].state = QCA_NAPI_CPU_DOWN; 1261 1262 NAPI_DEBUG("%s: CPU %u marked %d; updating affinity", 1263 __func__, cpu, napid->napi_cpu[cpu].state); 1264 1265 /** 1266 * we need to move any NAPIs on this CPU out. 1267 * if we are in LO throughput mode, then this is valid 1268 * if the CPU is the the low designated CPU. 1269 */ 1270 hif_napi_event(GET_HIF_OPAQUE_HDL(hif), 1271 NAPI_EVT_CPU_STATE, 1272 (void *) 1273 ((size_t)cpu << 16 | napid->napi_cpu[cpu].state)); 1274 1275 NAPI_DEBUG("<--%s", __func__); 1276 } 1277 1278 static int hnc_hotplug_register(struct hif_softc *hif_sc) 1279 { 1280 QDF_STATUS status; 1281 1282 NAPI_DEBUG("-->%s", __func__); 1283 1284 status = qdf_cpuhp_register(&hif_sc->napi_data.cpuhp_handler, 1285 hif_sc, 1286 hnc_cpu_online_cb, 1287 hnc_cpu_before_offline_cb); 1288 1289 NAPI_DEBUG("<--%s [%d]", __func__, status); 1290 1291 return qdf_status_to_os_return(status); 1292 } 1293 1294 static void hnc_hotplug_unregister(struct hif_softc *hif_sc) 1295 { 1296 NAPI_DEBUG("-->%s", __func__); 1297 1298 if (hif_sc->napi_data.cpuhp_handler) 1299 qdf_cpuhp_unregister(&hif_sc->napi_data.cpuhp_handler); 1300 1301 NAPI_DEBUG("<--%s", __func__); 1302 } 1303 1304 /** 1305 * hnc_install_tput() - installs a callback in the throughput detector 1306 * @register: !0 => register; =0: unregister 1307 * 1308 * installs a callback to be called when wifi driver throughput (tx+rx) 1309 * crosses a threshold. Currently, we are using the same criteria as 1310 * TCP ack suppression (500 packets/100ms by default). 1311 * 1312 * Return: 0 : success 1313 * <0: failure 1314 */ 1315 1316 static int hnc_tput_hook(int install) 1317 { 1318 int rc = 0; 1319 1320 /* 1321 * Nothing, until the bw_calculation accepts registration 1322 * it is now hardcoded in the wlan_hdd_main.c::hdd_bus_bw_compute_cbk 1323 * hdd_napi_throughput_policy(...) 1324 */ 1325 return rc; 1326 } 1327 1328 /* 1329 * Implementation of hif_napi_cpu API 1330 */ 1331 1332 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4, 4, 0)) 1333 static inline void record_sibling_cpumask(struct qca_napi_cpu *cpus, int i) 1334 { 1335 cpumask_copy(&(cpus[i].thread_mask), 1336 topology_sibling_cpumask(i)); 1337 } 1338 #else 1339 static inline void record_sibling_cpumask(struct qca_napi_cpu *cpus, int i) 1340 { 1341 } 1342 #endif 1343 1344 1345 /** 1346 * hif_napi_cpu_init() - initialization of irq affinity block 1347 * @ctx: pointer to qca_napi_data 1348 * 1349 * called by hif_napi_create, after the first instance is called 1350 * - builds napi_rss_cpus table from cpu topology 1351 * - links cores of the same clusters together 1352 * - installs hot-plug notifier 1353 * - installs throughput trigger notifier (when such mechanism exists) 1354 * 1355 * Return: 0: OK 1356 * <0: error code 1357 */ 1358 int hif_napi_cpu_init(struct hif_opaque_softc *hif) 1359 { 1360 int rc = 0; 1361 int i; 1362 struct qca_napi_data *napid = &HIF_GET_SOFTC(hif)->napi_data; 1363 struct qca_napi_cpu *cpus = napid->napi_cpu; 1364 1365 NAPI_DEBUG("--> "); 1366 1367 if (cpus[0].state != QCA_NAPI_CPU_UNINITIALIZED) { 1368 NAPI_DEBUG("NAPI RSS table already initialized.\n"); 1369 rc = -EALREADY; 1370 goto lab_rss_init; 1371 } 1372 1373 /* build CPU topology table */ 1374 for_each_possible_cpu(i) { 1375 cpus[i].state = ((cpumask_test_cpu(i, cpu_online_mask) 1376 ? QCA_NAPI_CPU_UP 1377 : QCA_NAPI_CPU_DOWN)); 1378 cpus[i].core_id = topology_core_id(i); 1379 cpus[i].cluster_id = topology_physical_package_id(i); 1380 cpumask_copy(&(cpus[i].core_mask), 1381 topology_core_cpumask(i)); 1382 record_sibling_cpumask(cpus, i); 1383 cpus[i].max_freq = cpufreq_quick_get_max(i); 1384 cpus[i].napis = 0x0; 1385 cpus[i].cluster_nxt = -1; /* invalid */ 1386 } 1387 1388 /* link clusters together */ 1389 rc = hnc_link_clusters(napid); 1390 if (0 != rc) 1391 goto lab_err_topology; 1392 1393 /* install hotplug notifier */ 1394 rc = hnc_hotplug_register(HIF_GET_SOFTC(hif)); 1395 if (0 != rc) 1396 goto lab_err_hotplug; 1397 1398 /* install throughput notifier */ 1399 rc = hnc_tput_hook(1); 1400 if (0 == rc) 1401 goto lab_rss_init; 1402 1403 lab_err_hotplug: 1404 hnc_tput_hook(0); 1405 hnc_hotplug_unregister(HIF_GET_SOFTC(hif)); 1406 lab_err_topology: 1407 memset(napid->napi_cpu, 0, sizeof(struct qca_napi_cpu) * NR_CPUS); 1408 lab_rss_init: 1409 NAPI_DEBUG("<-- [rc=%d]", rc); 1410 return rc; 1411 } 1412 1413 /** 1414 * hif_napi_cpu_deinit() - clean-up of irq affinity block 1415 * 1416 * called by hif_napi_destroy, when the last instance is removed 1417 * - uninstalls throughput and hotplug notifiers 1418 * - clears cpu topology table 1419 * Return: 0: OK 1420 */ 1421 int hif_napi_cpu_deinit(struct hif_opaque_softc *hif) 1422 { 1423 int rc = 0; 1424 struct qca_napi_data *napid = &HIF_GET_SOFTC(hif)->napi_data; 1425 1426 NAPI_DEBUG("-->%s(...)", __func__); 1427 1428 /* uninstall tput notifier */ 1429 rc = hnc_tput_hook(0); 1430 1431 /* uninstall hotplug notifier */ 1432 hnc_hotplug_unregister(HIF_GET_SOFTC(hif)); 1433 1434 /* clear the topology table */ 1435 memset(napid->napi_cpu, 0, sizeof(struct qca_napi_cpu) * NR_CPUS); 1436 1437 NAPI_DEBUG("<--%s[rc=%d]", __func__, rc); 1438 1439 return rc; 1440 } 1441 1442 /** 1443 * hncm_migrate_to() - migrates a NAPI to a CPU 1444 * @napid: pointer to NAPI block 1445 * @ce_id: CE_id of the NAPI instance 1446 * @didx : index in the CPU topology table for the CPU to migrate to 1447 * 1448 * Migrates NAPI (identified by the CE_id) to the destination core 1449 * Updates the napi_map of the destination entry 1450 * 1451 * Return: 1452 * =0 : success 1453 * <0 : error 1454 */ 1455 static int hncm_migrate_to(struct qca_napi_data *napid, 1456 int napi_ce, 1457 int didx) 1458 { 1459 int rc = 0; 1460 QDF_STATUS status; 1461 1462 NAPI_DEBUG("-->%s(napi_cd=%d, didx=%d)", __func__, napi_ce, didx); 1463 1464 if (!napid->napis[napi_ce]) 1465 return -EINVAL; 1466 1467 napid->napis[napi_ce]->cpumask.bits[0] = (1 << didx); 1468 1469 qdf_dev_modify_irq_status(napid->napis[napi_ce]->irq, 1470 QDF_IRQ_NO_BALANCING, 0); 1471 status = qdf_dev_set_irq_affinity(napid->napis[napi_ce]->irq, 1472 (struct qdf_cpu_mask *) 1473 &napid->napis[napi_ce]->cpumask); 1474 rc = qdf_status_to_os_return(status); 1475 1476 /* unmark the napis bitmap in the cpu table */ 1477 napid->napi_cpu[napid->napis[napi_ce]->cpu].napis &= ~(0x01 << napi_ce); 1478 /* mark the napis bitmap for the new designated cpu */ 1479 napid->napi_cpu[didx].napis |= (0x01 << napi_ce); 1480 napid->napis[napi_ce]->cpu = didx; 1481 1482 NAPI_DEBUG("<--%s[%d]", __func__, rc); 1483 return rc; 1484 } 1485 /** 1486 * hncm_dest_cpu() - finds a destination CPU for NAPI 1487 * @napid: pointer to NAPI block 1488 * @act : RELOCATE | COLLAPSE | DISPERSE 1489 * 1490 * Finds the designated destionation for the next IRQ. 1491 * RELOCATE: translated to either COLLAPSE or DISPERSE based 1492 * on napid->napi_mode (throughput state) 1493 * COLLAPSE: All have the same destination: the first online CPU in lilcl 1494 * DISPERSE: One of the CPU in bigcl, which has the smallest number of 1495 * NAPIs on it 1496 * 1497 * Return: >=0 : index in the cpu topology table 1498 * : < 0 : error 1499 */ 1500 static int hncm_dest_cpu(struct qca_napi_data *napid, int act) 1501 { 1502 int destidx = -1; 1503 int head, i; 1504 1505 NAPI_DEBUG("-->%s(act=%d)", __func__, act); 1506 if (act == HNC_ACT_RELOCATE) { 1507 if (napid->napi_mode == QCA_NAPI_TPUT_LO) 1508 act = HNC_ACT_COLLAPSE; 1509 else 1510 act = HNC_ACT_DISPERSE; 1511 NAPI_DEBUG("%s: act changed from HNC_ACT_RELOCATE to %d", 1512 __func__, act); 1513 } 1514 if (act == HNC_ACT_COLLAPSE) { 1515 head = i = napid->lilcl_head; 1516 retry_collapse: 1517 while (i >= 0) { 1518 if (napid->napi_cpu[i].state == QCA_NAPI_CPU_UP) { 1519 destidx = i; 1520 break; 1521 } 1522 i = napid->napi_cpu[i].cluster_nxt; 1523 } 1524 if ((destidx < 0) && (head == napid->lilcl_head)) { 1525 NAPI_DEBUG("%s: COLLAPSE: no lilcl dest, try bigcl", 1526 __func__); 1527 head = i = napid->bigcl_head; 1528 goto retry_collapse; 1529 } 1530 } else { /* HNC_ACT_DISPERSE */ 1531 int smallest = 99; /* all 32 bits full */ 1532 int smallidx = -1; 1533 1534 head = i = napid->bigcl_head; 1535 retry_disperse: 1536 while (i >= 0) { 1537 if ((napid->napi_cpu[i].state == QCA_NAPI_CPU_UP) && 1538 (hweight32(napid->napi_cpu[i].napis) <= smallest)) { 1539 smallest = napid->napi_cpu[i].napis; 1540 smallidx = i; 1541 } 1542 i = napid->napi_cpu[i].cluster_nxt; 1543 } 1544 /* Check if matches with user sepecified CPU mask */ 1545 smallidx = ((1 << smallidx) & napid->user_cpu_affin_mask) ? 1546 smallidx : -1; 1547 1548 if ((smallidx < 0) && (head == napid->bigcl_head)) { 1549 NAPI_DEBUG("%s: DISPERSE: no bigcl dest, try lilcl", 1550 __func__); 1551 head = i = napid->lilcl_head; 1552 goto retry_disperse; 1553 } 1554 destidx = smallidx; 1555 } 1556 NAPI_DEBUG("<--%s[dest=%d]", __func__, destidx); 1557 return destidx; 1558 } 1559 /** 1560 * hif_napi_cpu_migrate() - migrate IRQs away 1561 * @cpu: -1: all CPUs <n> specific CPU 1562 * @act: COLLAPSE | DISPERSE 1563 * 1564 * Moves IRQs/NAPIs from specific or all CPUs (specified by @cpu) to eligible 1565 * cores. Eligible cores are: 1566 * act=COLLAPSE -> the first online core of the little cluster 1567 * act=DISPERSE -> separate cores of the big cluster, so that each core will 1568 * host minimum number of NAPIs/IRQs (napid->cpus[cpu].napis) 1569 * 1570 * Note that this function is called with a spinlock acquired already. 1571 * 1572 * Return: =0: success 1573 * <0: error 1574 */ 1575 1576 int hif_napi_cpu_migrate(struct qca_napi_data *napid, int cpu, int action) 1577 { 1578 int rc = 0; 1579 struct qca_napi_cpu *cpup; 1580 int i, dind; 1581 uint32_t napis; 1582 1583 NAPI_DEBUG("-->%s(.., cpu=%d, act=%d)", 1584 __func__, cpu, action); 1585 /* the following is really: hif_napi_enabled() with less overhead */ 1586 if (napid->ce_map == 0) { 1587 NAPI_DEBUG("%s: NAPI disabled. Not migrating.", __func__); 1588 goto hncm_return; 1589 } 1590 1591 cpup = napid->napi_cpu; 1592 1593 switch (action) { 1594 case HNC_ACT_RELOCATE: 1595 case HNC_ACT_DISPERSE: 1596 case HNC_ACT_COLLAPSE: { 1597 /* first find the src napi set */ 1598 if (cpu == HNC_ANY_CPU) 1599 napis = napid->ce_map; 1600 else 1601 napis = cpup[cpu].napis; 1602 /* then clear the napi bitmap on each CPU */ 1603 for (i = 0; i < NR_CPUS; i++) 1604 cpup[i].napis = 0; 1605 /* then for each of the NAPIs to disperse: */ 1606 for (i = 0; i < CE_COUNT_MAX; i++) 1607 if (napis & (1 << i)) { 1608 /* find a destination CPU */ 1609 dind = hncm_dest_cpu(napid, action); 1610 if (dind >= 0) { 1611 NAPI_DEBUG("Migrating NAPI ce%d to %d", 1612 i, dind); 1613 rc = hncm_migrate_to(napid, i, dind); 1614 } else { 1615 NAPI_DEBUG("No dest for NAPI ce%d", i); 1616 hnc_dump_cpus(napid); 1617 rc = -1; 1618 } 1619 } 1620 break; 1621 } 1622 default: { 1623 NAPI_DEBUG("%s: bad action: %d\n", __func__, action); 1624 QDF_BUG(0); 1625 break; 1626 } 1627 } /* switch action */ 1628 1629 hncm_return: 1630 hnc_dump_cpus(napid); 1631 return rc; 1632 } 1633 1634 1635 /** 1636 * hif_napi_bl_irq() - calls irq_modify_status to enable/disable blacklisting 1637 * @napid: pointer to qca_napi_data structure 1638 * @bl_flag: blacklist flag to enable/disable blacklisting 1639 * 1640 * The function enables/disables blacklisting for all the copy engine 1641 * interrupts on which NAPI is enabled. 1642 * 1643 * Return: None 1644 */ 1645 static inline void hif_napi_bl_irq(struct qca_napi_data *napid, bool bl_flag) 1646 { 1647 int i; 1648 struct qca_napi_info *napii; 1649 1650 for (i = 0; i < CE_COUNT_MAX; i++) { 1651 /* check if NAPI is enabled on the CE */ 1652 if (!(napid->ce_map & (0x01 << i))) 1653 continue; 1654 1655 /*double check that NAPI is allocated for the CE */ 1656 napii = napid->napis[i]; 1657 if (!(napii)) 1658 continue; 1659 1660 if (bl_flag == true) 1661 qdf_dev_modify_irq_status(napii->irq, 1662 0, QDF_IRQ_NO_BALANCING); 1663 else 1664 qdf_dev_modify_irq_status(napii->irq, 1665 QDF_IRQ_NO_BALANCING, 0); 1666 hif_debug("bl_flag %d CE %d", bl_flag, i); 1667 } 1668 } 1669 1670 /** 1671 * hif_napi_cpu_blacklist() - en(dis)ables blacklisting for NAPI RX interrupts. 1672 * @napid: pointer to qca_napi_data structure 1673 * @op: blacklist operation to perform 1674 * 1675 * The function enables/disables/queries blacklisting for all CE RX 1676 * interrupts with NAPI enabled. Besides blacklisting, it also enables/disables 1677 * core_ctl_set_boost. 1678 * Once blacklisting is enabled, the interrupts will not be managed by the IRQ 1679 * balancer. 1680 * 1681 * Return: -EINVAL, in case IRQ_BLACKLISTING and CORE_CTL_BOOST is not enabled 1682 * for BLACKLIST_QUERY op - blacklist refcount 1683 * for BLACKLIST_ON op - return value from core_ctl_set_boost API 1684 * for BLACKLIST_OFF op - return value from core_ctl_set_boost API 1685 */ 1686 int hif_napi_cpu_blacklist(struct qca_napi_data *napid, 1687 enum qca_blacklist_op op) 1688 { 1689 int rc = 0; 1690 static int ref_count; /* = 0 by the compiler */ 1691 uint8_t flags = napid->flags; 1692 bool bl_en = flags & QCA_NAPI_FEATURE_IRQ_BLACKLISTING; 1693 bool ccb_en = flags & QCA_NAPI_FEATURE_CORE_CTL_BOOST; 1694 1695 NAPI_DEBUG("-->%s(%d %d)", __func__, flags, op); 1696 1697 if (!(bl_en && ccb_en)) { 1698 rc = -EINVAL; 1699 goto out; 1700 } 1701 1702 switch (op) { 1703 case BLACKLIST_QUERY: 1704 rc = ref_count; 1705 break; 1706 case BLACKLIST_ON: 1707 ref_count++; 1708 rc = 0; 1709 if (ref_count == 1) { 1710 rc = hif_napi_core_ctl_set_boost(true); 1711 NAPI_DEBUG("boost_on() returns %d - refcnt=%d", 1712 rc, ref_count); 1713 hif_napi_bl_irq(napid, true); 1714 } 1715 break; 1716 case BLACKLIST_OFF: 1717 if (ref_count) { 1718 ref_count--; 1719 rc = 0; 1720 if (ref_count == 0) { 1721 rc = hif_napi_core_ctl_set_boost(false); 1722 NAPI_DEBUG("boost_off() returns %d - refcnt=%d", 1723 rc, ref_count); 1724 hif_napi_bl_irq(napid, false); 1725 } 1726 } 1727 break; 1728 default: 1729 NAPI_DEBUG("Invalid blacklist op: %d", op); 1730 rc = -EINVAL; 1731 } /* switch */ 1732 out: 1733 NAPI_DEBUG("<--%s[%d]", __func__, rc); 1734 return rc; 1735 } 1736 1737 /** 1738 * hif_napi_serialize() - [de-]serialize NAPI operations 1739 * @hif: context 1740 * @is_on: 1: serialize, 0: deserialize 1741 * 1742 * hif_napi_serialize(hif, 1) can be called multiple times. It will perform the 1743 * following steps (see hif_napi_event for code): 1744 * - put irqs of all NAPI instances on the same CPU 1745 * - only for the first serialize call: blacklist 1746 * 1747 * hif_napi_serialize(hif, 0): 1748 * - start a timer (multiple of BusBandwidthTimer -- default: 100 msec) 1749 * - at the end of the timer, check the current throughput state and 1750 * implement it. 1751 */ 1752 static unsigned long napi_serialize_reqs; 1753 int hif_napi_serialize(struct hif_opaque_softc *hif, int is_on) 1754 { 1755 int rc = -EINVAL; 1756 1757 if (hif) 1758 switch (is_on) { 1759 case 0: { /* de-serialize */ 1760 rc = hif_napi_event(hif, NAPI_EVT_USR_NORMAL, 1761 (void *) 0); 1762 napi_serialize_reqs = 0; 1763 break; 1764 } /* end de-serialize */ 1765 case 1: { /* serialize */ 1766 rc = hif_napi_event(hif, NAPI_EVT_USR_SERIAL, 1767 (void *)napi_serialize_reqs++); 1768 break; 1769 } /* end serialize */ 1770 default: 1771 break; /* no-op */ 1772 } /* switch */ 1773 return rc; 1774 } 1775 1776 #endif /* ifdef HIF_IRQ_AFFINITY */ 1777