1 // SPDX-License-Identifier: GPL-2.0-or-later 2 /* 3 * Copyright The Asahi Linux Contributors 4 * 5 * Based on irq-lpc32xx: 6 * Copyright 2015-2016 Vladimir Zapolskiy <vz@mleia.com> 7 * Based on irq-bcm2836: 8 * Copyright 2015 Broadcom 9 */ 10 11 /* 12 * AIC is a fairly simple interrupt controller with the following features: 13 * 14 * - 896 level-triggered hardware IRQs 15 * - Single mask bit per IRQ 16 * - Per-IRQ affinity setting 17 * - Automatic masking on event delivery (auto-ack) 18 * - Software triggering (ORed with hw line) 19 * - 2 per-CPU IPIs (meant as "self" and "other", but they are 20 * interchangeable if not symmetric) 21 * - Automatic prioritization (single event/ack register per CPU, lower IRQs = 22 * higher priority) 23 * - Automatic masking on ack 24 * - Default "this CPU" register view and explicit per-CPU views 25 * 26 * In addition, this driver also handles FIQs, as these are routed to the same 27 * IRQ vector. These are used for Fast IPIs, the ARMv8 timer IRQs, and 28 * performance counters (TODO). 29 * 30 * Implementation notes: 31 * 32 * - This driver creates two IRQ domains, one for HW IRQs and internal FIQs, 33 * and one for IPIs. 34 * - Since Linux needs more than 2 IPIs, we implement a software IRQ controller 35 * and funnel all IPIs into one per-CPU IPI (the second "self" IPI is unused). 36 * - FIQ hwirq numbers are assigned after true hwirqs, and are per-cpu. 37 * - DT bindings use 3-cell form (like GIC): 38 * - <0 nr flags> - hwirq #nr 39 * - <1 nr flags> - FIQ #nr 40 * - nr=0 Physical HV timer 41 * - nr=1 Virtual HV timer 42 * - nr=2 Physical guest timer 43 * - nr=3 Virtual guest timer 44 */ 45 46 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt 47 48 #include <linux/bits.h> 49 #include <linux/bitfield.h> 50 #include <linux/cpuhotplug.h> 51 #include <linux/io.h> 52 #include <linux/irqchip.h> 53 #include <linux/irqchip/arm-vgic-info.h> 54 #include <linux/irqdomain.h> 55 #include <linux/jump_label.h> 56 #include <linux/limits.h> 57 #include <linux/of_address.h> 58 #include <linux/slab.h> 59 #include <asm/apple_m1_pmu.h> 60 #include <asm/cputype.h> 61 #include <asm/exception.h> 62 #include <asm/sysreg.h> 63 #include <asm/virt.h> 64 65 #include <dt-bindings/interrupt-controller/apple-aic.h> 66 67 /* 68 * AIC v1 registers (MMIO) 69 */ 70 71 #define AIC_INFO 0x0004 72 #define AIC_INFO_NR_IRQ GENMASK(15, 0) 73 74 #define AIC_CONFIG 0x0010 75 76 #define AIC_WHOAMI 0x2000 77 #define AIC_EVENT 0x2004 78 #define AIC_EVENT_DIE GENMASK(31, 24) 79 #define AIC_EVENT_TYPE GENMASK(23, 16) 80 #define AIC_EVENT_NUM GENMASK(15, 0) 81 82 #define AIC_EVENT_TYPE_FIQ 0 /* Software use */ 83 #define AIC_EVENT_TYPE_IRQ 1 84 #define AIC_EVENT_TYPE_IPI 4 85 #define AIC_EVENT_IPI_OTHER 1 86 #define AIC_EVENT_IPI_SELF 2 87 88 #define AIC_IPI_SEND 0x2008 89 #define AIC_IPI_ACK 0x200c 90 #define AIC_IPI_MASK_SET 0x2024 91 #define AIC_IPI_MASK_CLR 0x2028 92 93 #define AIC_IPI_SEND_CPU(cpu) BIT(cpu) 94 95 #define AIC_IPI_OTHER BIT(0) 96 #define AIC_IPI_SELF BIT(31) 97 98 #define AIC_TARGET_CPU 0x3000 99 100 #define AIC_CPU_IPI_SET(cpu) (0x5008 + ((cpu) << 7)) 101 #define AIC_CPU_IPI_CLR(cpu) (0x500c + ((cpu) << 7)) 102 #define AIC_CPU_IPI_MASK_SET(cpu) (0x5024 + ((cpu) << 7)) 103 #define AIC_CPU_IPI_MASK_CLR(cpu) (0x5028 + ((cpu) << 7)) 104 105 #define AIC_MAX_IRQ 0x400 106 107 /* 108 * AIC v2 registers (MMIO) 109 */ 110 111 #define AIC2_VERSION 0x0000 112 #define AIC2_VERSION_VER GENMASK(7, 0) 113 114 #define AIC2_INFO1 0x0004 115 #define AIC2_INFO1_NR_IRQ GENMASK(15, 0) 116 #define AIC2_INFO1_LAST_DIE GENMASK(27, 24) 117 118 #define AIC2_INFO2 0x0008 119 120 #define AIC2_INFO3 0x000c 121 #define AIC2_INFO3_MAX_IRQ GENMASK(15, 0) 122 #define AIC2_INFO3_MAX_DIE GENMASK(27, 24) 123 124 #define AIC2_RESET 0x0010 125 #define AIC2_RESET_RESET BIT(0) 126 127 #define AIC2_CONFIG 0x0014 128 #define AIC2_CONFIG_ENABLE BIT(0) 129 #define AIC2_CONFIG_PREFER_PCPU BIT(28) 130 131 #define AIC2_TIMEOUT 0x0028 132 #define AIC2_CLUSTER_PRIO 0x0030 133 #define AIC2_DELAY_GROUPS 0x0100 134 135 #define AIC2_IRQ_CFG 0x2000 136 137 /* 138 * AIC2 registers are laid out like this, starting at AIC2_IRQ_CFG: 139 * 140 * Repeat for each die: 141 * IRQ_CFG: u32 * MAX_IRQS 142 * SW_SET: u32 * (MAX_IRQS / 32) 143 * SW_CLR: u32 * (MAX_IRQS / 32) 144 * MASK_SET: u32 * (MAX_IRQS / 32) 145 * MASK_CLR: u32 * (MAX_IRQS / 32) 146 * HW_STATE: u32 * (MAX_IRQS / 32) 147 * 148 * This is followed by a set of event registers, each 16K page aligned. 149 * The first one is the AP event register we will use. Unfortunately, 150 * the actual implemented die count is not specified anywhere in the 151 * capability registers, so we have to explicitly specify the event 152 * register as a second reg entry in the device tree to remain 153 * forward-compatible. 154 */ 155 156 #define AIC2_IRQ_CFG_TARGET GENMASK(3, 0) 157 #define AIC2_IRQ_CFG_DELAY_IDX GENMASK(7, 5) 158 159 #define MASK_REG(x) (4 * ((x) >> 5)) 160 #define MASK_BIT(x) BIT((x) & GENMASK(4, 0)) 161 162 /* 163 * IMP-DEF sysregs that control FIQ sources 164 */ 165 166 /* IPI request registers */ 167 #define SYS_IMP_APL_IPI_RR_LOCAL_EL1 sys_reg(3, 5, 15, 0, 0) 168 #define SYS_IMP_APL_IPI_RR_GLOBAL_EL1 sys_reg(3, 5, 15, 0, 1) 169 #define IPI_RR_CPU GENMASK(7, 0) 170 /* Cluster only used for the GLOBAL register */ 171 #define IPI_RR_CLUSTER GENMASK(23, 16) 172 #define IPI_RR_TYPE GENMASK(29, 28) 173 #define IPI_RR_IMMEDIATE 0 174 #define IPI_RR_RETRACT 1 175 #define IPI_RR_DEFERRED 2 176 #define IPI_RR_NOWAKE 3 177 178 /* IPI status register */ 179 #define SYS_IMP_APL_IPI_SR_EL1 sys_reg(3, 5, 15, 1, 1) 180 #define IPI_SR_PENDING BIT(0) 181 182 /* Guest timer FIQ enable register */ 183 #define SYS_IMP_APL_VM_TMR_FIQ_ENA_EL2 sys_reg(3, 5, 15, 1, 3) 184 #define VM_TMR_FIQ_ENABLE_V BIT(0) 185 #define VM_TMR_FIQ_ENABLE_P BIT(1) 186 187 /* Deferred IPI countdown register */ 188 #define SYS_IMP_APL_IPI_CR_EL1 sys_reg(3, 5, 15, 3, 1) 189 190 /* Uncore PMC control register */ 191 #define SYS_IMP_APL_UPMCR0_EL1 sys_reg(3, 7, 15, 0, 4) 192 #define UPMCR0_IMODE GENMASK(18, 16) 193 #define UPMCR0_IMODE_OFF 0 194 #define UPMCR0_IMODE_AIC 2 195 #define UPMCR0_IMODE_HALT 3 196 #define UPMCR0_IMODE_FIQ 4 197 198 /* Uncore PMC status register */ 199 #define SYS_IMP_APL_UPMSR_EL1 sys_reg(3, 7, 15, 6, 4) 200 #define UPMSR_IACT BIT(0) 201 202 /* MPIDR fields */ 203 #define MPIDR_CPU(x) MPIDR_AFFINITY_LEVEL(x, 0) 204 #define MPIDR_CLUSTER(x) MPIDR_AFFINITY_LEVEL(x, 1) 205 206 #define AIC_IRQ_HWIRQ(die, irq) (FIELD_PREP(AIC_EVENT_DIE, die) | \ 207 FIELD_PREP(AIC_EVENT_TYPE, AIC_EVENT_TYPE_IRQ) | \ 208 FIELD_PREP(AIC_EVENT_NUM, irq)) 209 #define AIC_FIQ_HWIRQ(x) (FIELD_PREP(AIC_EVENT_TYPE, AIC_EVENT_TYPE_FIQ) | \ 210 FIELD_PREP(AIC_EVENT_NUM, x)) 211 #define AIC_HWIRQ_IRQ(x) FIELD_GET(AIC_EVENT_NUM, x) 212 #define AIC_HWIRQ_DIE(x) FIELD_GET(AIC_EVENT_DIE, x) 213 #define AIC_NR_FIQ 6 214 #define AIC_NR_SWIPI 32 215 216 /* 217 * FIQ hwirq index definitions: FIQ sources use the DT binding defines 218 * directly, except that timers are special. At the irqchip level, the 219 * two timer types are represented by their access method: _EL0 registers 220 * or _EL02 registers. In the DT binding, the timers are represented 221 * by their purpose (HV or guest). This mapping is for when the kernel is 222 * running at EL2 (with VHE). When the kernel is running at EL1, the 223 * mapping differs and aic_irq_domain_translate() performs the remapping. 224 */ 225 226 #define AIC_TMR_EL0_PHYS AIC_TMR_HV_PHYS 227 #define AIC_TMR_EL0_VIRT AIC_TMR_HV_VIRT 228 #define AIC_TMR_EL02_PHYS AIC_TMR_GUEST_PHYS 229 #define AIC_TMR_EL02_VIRT AIC_TMR_GUEST_VIRT 230 231 static DEFINE_STATIC_KEY_TRUE(use_fast_ipi); 232 233 struct aic_info { 234 int version; 235 236 /* Register offsets */ 237 u32 event; 238 u32 target_cpu; 239 u32 irq_cfg; 240 u32 sw_set; 241 u32 sw_clr; 242 u32 mask_set; 243 u32 mask_clr; 244 245 u32 die_stride; 246 247 /* Features */ 248 bool fast_ipi; 249 }; 250 251 static const struct aic_info aic1_info = { 252 .version = 1, 253 254 .event = AIC_EVENT, 255 .target_cpu = AIC_TARGET_CPU, 256 }; 257 258 static const struct aic_info aic1_fipi_info = { 259 .version = 1, 260 261 .event = AIC_EVENT, 262 .target_cpu = AIC_TARGET_CPU, 263 264 .fast_ipi = true, 265 }; 266 267 static const struct aic_info aic2_info = { 268 .version = 2, 269 270 .irq_cfg = AIC2_IRQ_CFG, 271 272 .fast_ipi = true, 273 }; 274 275 static const struct of_device_id aic_info_match[] = { 276 { 277 .compatible = "apple,t8103-aic", 278 .data = &aic1_fipi_info, 279 }, 280 { 281 .compatible = "apple,aic", 282 .data = &aic1_info, 283 }, 284 { 285 .compatible = "apple,aic2", 286 .data = &aic2_info, 287 }, 288 {} 289 }; 290 291 struct aic_irq_chip { 292 void __iomem *base; 293 void __iomem *event; 294 struct irq_domain *hw_domain; 295 struct irq_domain *ipi_domain; 296 struct { 297 cpumask_t aff; 298 } *fiq_aff[AIC_NR_FIQ]; 299 300 int nr_irq; 301 int max_irq; 302 int nr_die; 303 int max_die; 304 305 struct aic_info info; 306 }; 307 308 static DEFINE_PER_CPU(uint32_t, aic_fiq_unmasked); 309 310 static DEFINE_PER_CPU(atomic_t, aic_vipi_flag); 311 static DEFINE_PER_CPU(atomic_t, aic_vipi_enable); 312 313 static struct aic_irq_chip *aic_irqc; 314 315 static void aic_handle_ipi(struct pt_regs *regs); 316 317 static u32 aic_ic_read(struct aic_irq_chip *ic, u32 reg) 318 { 319 return readl_relaxed(ic->base + reg); 320 } 321 322 static void aic_ic_write(struct aic_irq_chip *ic, u32 reg, u32 val) 323 { 324 writel_relaxed(val, ic->base + reg); 325 } 326 327 /* 328 * IRQ irqchip 329 */ 330 331 static void aic_irq_mask(struct irq_data *d) 332 { 333 irq_hw_number_t hwirq = irqd_to_hwirq(d); 334 struct aic_irq_chip *ic = irq_data_get_irq_chip_data(d); 335 336 u32 off = AIC_HWIRQ_DIE(hwirq) * ic->info.die_stride; 337 u32 irq = AIC_HWIRQ_IRQ(hwirq); 338 339 aic_ic_write(ic, ic->info.mask_set + off + MASK_REG(irq), MASK_BIT(irq)); 340 } 341 342 static void aic_irq_unmask(struct irq_data *d) 343 { 344 irq_hw_number_t hwirq = irqd_to_hwirq(d); 345 struct aic_irq_chip *ic = irq_data_get_irq_chip_data(d); 346 347 u32 off = AIC_HWIRQ_DIE(hwirq) * ic->info.die_stride; 348 u32 irq = AIC_HWIRQ_IRQ(hwirq); 349 350 aic_ic_write(ic, ic->info.mask_clr + off + MASK_REG(irq), MASK_BIT(irq)); 351 } 352 353 static void aic_irq_eoi(struct irq_data *d) 354 { 355 /* 356 * Reading the interrupt reason automatically acknowledges and masks 357 * the IRQ, so we just unmask it here if needed. 358 */ 359 if (!irqd_irq_masked(d)) 360 aic_irq_unmask(d); 361 } 362 363 static void __exception_irq_entry aic_handle_irq(struct pt_regs *regs) 364 { 365 struct aic_irq_chip *ic = aic_irqc; 366 u32 event, type, irq; 367 368 do { 369 /* 370 * We cannot use a relaxed read here, as reads from DMA buffers 371 * need to be ordered after the IRQ fires. 372 */ 373 event = readl(ic->event + ic->info.event); 374 type = FIELD_GET(AIC_EVENT_TYPE, event); 375 irq = FIELD_GET(AIC_EVENT_NUM, event); 376 377 if (type == AIC_EVENT_TYPE_IRQ) 378 generic_handle_domain_irq(aic_irqc->hw_domain, event); 379 else if (type == AIC_EVENT_TYPE_IPI && irq == 1) 380 aic_handle_ipi(regs); 381 else if (event != 0) 382 pr_err_ratelimited("Unknown IRQ event %d, %d\n", type, irq); 383 } while (event); 384 385 /* 386 * vGIC maintenance interrupts end up here too, so we need to check 387 * for them separately. This should never trigger if KVM is working 388 * properly, because it will have already taken care of clearing it 389 * on guest exit before this handler runs. 390 */ 391 if (is_kernel_in_hyp_mode() && (read_sysreg_s(SYS_ICH_HCR_EL2) & ICH_HCR_EN) && 392 read_sysreg_s(SYS_ICH_MISR_EL2) != 0) { 393 pr_err_ratelimited("vGIC IRQ fired and not handled by KVM, disabling.\n"); 394 sysreg_clear_set_s(SYS_ICH_HCR_EL2, ICH_HCR_EN, 0); 395 } 396 } 397 398 static int aic_irq_set_affinity(struct irq_data *d, 399 const struct cpumask *mask_val, bool force) 400 { 401 irq_hw_number_t hwirq = irqd_to_hwirq(d); 402 struct aic_irq_chip *ic = irq_data_get_irq_chip_data(d); 403 int cpu; 404 405 BUG_ON(!ic->info.target_cpu); 406 407 if (force) 408 cpu = cpumask_first(mask_val); 409 else 410 cpu = cpumask_any_and(mask_val, cpu_online_mask); 411 412 aic_ic_write(ic, ic->info.target_cpu + AIC_HWIRQ_IRQ(hwirq) * 4, BIT(cpu)); 413 irq_data_update_effective_affinity(d, cpumask_of(cpu)); 414 415 return IRQ_SET_MASK_OK; 416 } 417 418 static int aic_irq_set_type(struct irq_data *d, unsigned int type) 419 { 420 /* 421 * Some IRQs (e.g. MSIs) implicitly have edge semantics, and we don't 422 * have a way to find out the type of any given IRQ, so just allow both. 423 */ 424 return (type == IRQ_TYPE_LEVEL_HIGH || type == IRQ_TYPE_EDGE_RISING) ? 0 : -EINVAL; 425 } 426 427 static struct irq_chip aic_chip = { 428 .name = "AIC", 429 .irq_mask = aic_irq_mask, 430 .irq_unmask = aic_irq_unmask, 431 .irq_eoi = aic_irq_eoi, 432 .irq_set_affinity = aic_irq_set_affinity, 433 .irq_set_type = aic_irq_set_type, 434 }; 435 436 static struct irq_chip aic2_chip = { 437 .name = "AIC2", 438 .irq_mask = aic_irq_mask, 439 .irq_unmask = aic_irq_unmask, 440 .irq_eoi = aic_irq_eoi, 441 .irq_set_type = aic_irq_set_type, 442 }; 443 444 /* 445 * FIQ irqchip 446 */ 447 448 static unsigned long aic_fiq_get_idx(struct irq_data *d) 449 { 450 return AIC_HWIRQ_IRQ(irqd_to_hwirq(d)); 451 } 452 453 static void aic_fiq_set_mask(struct irq_data *d) 454 { 455 /* Only the guest timers have real mask bits, unfortunately. */ 456 switch (aic_fiq_get_idx(d)) { 457 case AIC_TMR_EL02_PHYS: 458 sysreg_clear_set_s(SYS_IMP_APL_VM_TMR_FIQ_ENA_EL2, VM_TMR_FIQ_ENABLE_P, 0); 459 isb(); 460 break; 461 case AIC_TMR_EL02_VIRT: 462 sysreg_clear_set_s(SYS_IMP_APL_VM_TMR_FIQ_ENA_EL2, VM_TMR_FIQ_ENABLE_V, 0); 463 isb(); 464 break; 465 default: 466 break; 467 } 468 } 469 470 static void aic_fiq_clear_mask(struct irq_data *d) 471 { 472 switch (aic_fiq_get_idx(d)) { 473 case AIC_TMR_EL02_PHYS: 474 sysreg_clear_set_s(SYS_IMP_APL_VM_TMR_FIQ_ENA_EL2, 0, VM_TMR_FIQ_ENABLE_P); 475 isb(); 476 break; 477 case AIC_TMR_EL02_VIRT: 478 sysreg_clear_set_s(SYS_IMP_APL_VM_TMR_FIQ_ENA_EL2, 0, VM_TMR_FIQ_ENABLE_V); 479 isb(); 480 break; 481 default: 482 break; 483 } 484 } 485 486 static void aic_fiq_mask(struct irq_data *d) 487 { 488 aic_fiq_set_mask(d); 489 __this_cpu_and(aic_fiq_unmasked, ~BIT(aic_fiq_get_idx(d))); 490 } 491 492 static void aic_fiq_unmask(struct irq_data *d) 493 { 494 aic_fiq_clear_mask(d); 495 __this_cpu_or(aic_fiq_unmasked, BIT(aic_fiq_get_idx(d))); 496 } 497 498 static void aic_fiq_eoi(struct irq_data *d) 499 { 500 /* We mask to ack (where we can), so we need to unmask at EOI. */ 501 if (__this_cpu_read(aic_fiq_unmasked) & BIT(aic_fiq_get_idx(d))) 502 aic_fiq_clear_mask(d); 503 } 504 505 #define TIMER_FIRING(x) \ 506 (((x) & (ARCH_TIMER_CTRL_ENABLE | ARCH_TIMER_CTRL_IT_MASK | \ 507 ARCH_TIMER_CTRL_IT_STAT)) == \ 508 (ARCH_TIMER_CTRL_ENABLE | ARCH_TIMER_CTRL_IT_STAT)) 509 510 static void __exception_irq_entry aic_handle_fiq(struct pt_regs *regs) 511 { 512 /* 513 * It would be really nice if we had a system register that lets us get 514 * the FIQ source state without having to peek down into sources... 515 * but such a register does not seem to exist. 516 * 517 * So, we have these potential sources to test for: 518 * - Fast IPIs (not yet used) 519 * - The 4 timers (CNTP, CNTV for each of HV and guest) 520 * - Per-core PMCs (not yet supported) 521 * - Per-cluster uncore PMCs (not yet supported) 522 * 523 * Since not dealing with any of these results in a FIQ storm, 524 * we check for everything here, even things we don't support yet. 525 */ 526 527 if (read_sysreg_s(SYS_IMP_APL_IPI_SR_EL1) & IPI_SR_PENDING) { 528 if (static_branch_likely(&use_fast_ipi)) { 529 aic_handle_ipi(regs); 530 } else { 531 pr_err_ratelimited("Fast IPI fired. Acking.\n"); 532 write_sysreg_s(IPI_SR_PENDING, SYS_IMP_APL_IPI_SR_EL1); 533 } 534 } 535 536 if (TIMER_FIRING(read_sysreg(cntp_ctl_el0))) 537 generic_handle_domain_irq(aic_irqc->hw_domain, 538 AIC_FIQ_HWIRQ(AIC_TMR_EL0_PHYS)); 539 540 if (TIMER_FIRING(read_sysreg(cntv_ctl_el0))) 541 generic_handle_domain_irq(aic_irqc->hw_domain, 542 AIC_FIQ_HWIRQ(AIC_TMR_EL0_VIRT)); 543 544 if (is_kernel_in_hyp_mode()) { 545 uint64_t enabled = read_sysreg_s(SYS_IMP_APL_VM_TMR_FIQ_ENA_EL2); 546 547 if ((enabled & VM_TMR_FIQ_ENABLE_P) && 548 TIMER_FIRING(read_sysreg_s(SYS_CNTP_CTL_EL02))) 549 generic_handle_domain_irq(aic_irqc->hw_domain, 550 AIC_FIQ_HWIRQ(AIC_TMR_EL02_PHYS)); 551 552 if ((enabled & VM_TMR_FIQ_ENABLE_V) && 553 TIMER_FIRING(read_sysreg_s(SYS_CNTV_CTL_EL02))) 554 generic_handle_domain_irq(aic_irqc->hw_domain, 555 AIC_FIQ_HWIRQ(AIC_TMR_EL02_VIRT)); 556 } 557 558 if (read_sysreg_s(SYS_IMP_APL_PMCR0_EL1) & PMCR0_IACT) { 559 int irq; 560 if (cpumask_test_cpu(smp_processor_id(), 561 &aic_irqc->fiq_aff[AIC_CPU_PMU_P]->aff)) 562 irq = AIC_CPU_PMU_P; 563 else 564 irq = AIC_CPU_PMU_E; 565 generic_handle_domain_irq(aic_irqc->hw_domain, 566 AIC_FIQ_HWIRQ(irq)); 567 } 568 569 if (FIELD_GET(UPMCR0_IMODE, read_sysreg_s(SYS_IMP_APL_UPMCR0_EL1)) == UPMCR0_IMODE_FIQ && 570 (read_sysreg_s(SYS_IMP_APL_UPMSR_EL1) & UPMSR_IACT)) { 571 /* Same story with uncore PMCs */ 572 pr_err_ratelimited("Uncore PMC FIQ fired. Masking.\n"); 573 sysreg_clear_set_s(SYS_IMP_APL_UPMCR0_EL1, UPMCR0_IMODE, 574 FIELD_PREP(UPMCR0_IMODE, UPMCR0_IMODE_OFF)); 575 } 576 } 577 578 static int aic_fiq_set_type(struct irq_data *d, unsigned int type) 579 { 580 return (type == IRQ_TYPE_LEVEL_HIGH) ? 0 : -EINVAL; 581 } 582 583 static struct irq_chip fiq_chip = { 584 .name = "AIC-FIQ", 585 .irq_mask = aic_fiq_mask, 586 .irq_unmask = aic_fiq_unmask, 587 .irq_ack = aic_fiq_set_mask, 588 .irq_eoi = aic_fiq_eoi, 589 .irq_set_type = aic_fiq_set_type, 590 }; 591 592 /* 593 * Main IRQ domain 594 */ 595 596 static int aic_irq_domain_map(struct irq_domain *id, unsigned int irq, 597 irq_hw_number_t hw) 598 { 599 struct aic_irq_chip *ic = id->host_data; 600 u32 type = FIELD_GET(AIC_EVENT_TYPE, hw); 601 struct irq_chip *chip = &aic_chip; 602 603 if (ic->info.version == 2) 604 chip = &aic2_chip; 605 606 if (type == AIC_EVENT_TYPE_IRQ) { 607 irq_domain_set_info(id, irq, hw, chip, id->host_data, 608 handle_fasteoi_irq, NULL, NULL); 609 irqd_set_single_target(irq_desc_get_irq_data(irq_to_desc(irq))); 610 } else { 611 int fiq = FIELD_GET(AIC_EVENT_NUM, hw); 612 613 switch (fiq) { 614 case AIC_CPU_PMU_P: 615 case AIC_CPU_PMU_E: 616 irq_set_percpu_devid_partition(irq, &ic->fiq_aff[fiq]->aff); 617 break; 618 default: 619 irq_set_percpu_devid(irq); 620 break; 621 } 622 623 irq_domain_set_info(id, irq, hw, &fiq_chip, id->host_data, 624 handle_percpu_devid_irq, NULL, NULL); 625 } 626 627 return 0; 628 } 629 630 static int aic_irq_domain_translate(struct irq_domain *id, 631 struct irq_fwspec *fwspec, 632 unsigned long *hwirq, 633 unsigned int *type) 634 { 635 struct aic_irq_chip *ic = id->host_data; 636 u32 *args; 637 u32 die = 0; 638 639 if (fwspec->param_count < 3 || fwspec->param_count > 4 || 640 !is_of_node(fwspec->fwnode)) 641 return -EINVAL; 642 643 args = &fwspec->param[1]; 644 645 if (fwspec->param_count == 4) { 646 die = args[0]; 647 args++; 648 } 649 650 switch (fwspec->param[0]) { 651 case AIC_IRQ: 652 if (die >= ic->nr_die) 653 return -EINVAL; 654 if (args[0] >= ic->nr_irq) 655 return -EINVAL; 656 *hwirq = AIC_IRQ_HWIRQ(die, args[0]); 657 break; 658 case AIC_FIQ: 659 if (die != 0) 660 return -EINVAL; 661 if (args[0] >= AIC_NR_FIQ) 662 return -EINVAL; 663 *hwirq = AIC_FIQ_HWIRQ(args[0]); 664 665 /* 666 * In EL1 the non-redirected registers are the guest's, 667 * not EL2's, so remap the hwirqs to match. 668 */ 669 if (!is_kernel_in_hyp_mode()) { 670 switch (args[0]) { 671 case AIC_TMR_GUEST_PHYS: 672 *hwirq = AIC_FIQ_HWIRQ(AIC_TMR_EL0_PHYS); 673 break; 674 case AIC_TMR_GUEST_VIRT: 675 *hwirq = AIC_FIQ_HWIRQ(AIC_TMR_EL0_VIRT); 676 break; 677 case AIC_TMR_HV_PHYS: 678 case AIC_TMR_HV_VIRT: 679 return -ENOENT; 680 default: 681 break; 682 } 683 } 684 break; 685 default: 686 return -EINVAL; 687 } 688 689 *type = args[1] & IRQ_TYPE_SENSE_MASK; 690 691 return 0; 692 } 693 694 static int aic_irq_domain_alloc(struct irq_domain *domain, unsigned int virq, 695 unsigned int nr_irqs, void *arg) 696 { 697 unsigned int type = IRQ_TYPE_NONE; 698 struct irq_fwspec *fwspec = arg; 699 irq_hw_number_t hwirq; 700 int i, ret; 701 702 ret = aic_irq_domain_translate(domain, fwspec, &hwirq, &type); 703 if (ret) 704 return ret; 705 706 for (i = 0; i < nr_irqs; i++) { 707 ret = aic_irq_domain_map(domain, virq + i, hwirq + i); 708 if (ret) 709 return ret; 710 } 711 712 return 0; 713 } 714 715 static void aic_irq_domain_free(struct irq_domain *domain, unsigned int virq, 716 unsigned int nr_irqs) 717 { 718 int i; 719 720 for (i = 0; i < nr_irqs; i++) { 721 struct irq_data *d = irq_domain_get_irq_data(domain, virq + i); 722 723 irq_set_handler(virq + i, NULL); 724 irq_domain_reset_irq_data(d); 725 } 726 } 727 728 static const struct irq_domain_ops aic_irq_domain_ops = { 729 .translate = aic_irq_domain_translate, 730 .alloc = aic_irq_domain_alloc, 731 .free = aic_irq_domain_free, 732 }; 733 734 /* 735 * IPI irqchip 736 */ 737 738 static void aic_ipi_send_fast(int cpu) 739 { 740 u64 mpidr = cpu_logical_map(cpu); 741 u64 my_mpidr = read_cpuid_mpidr(); 742 u64 cluster = MPIDR_CLUSTER(mpidr); 743 u64 idx = MPIDR_CPU(mpidr); 744 745 if (MPIDR_CLUSTER(my_mpidr) == cluster) 746 write_sysreg_s(FIELD_PREP(IPI_RR_CPU, idx), 747 SYS_IMP_APL_IPI_RR_LOCAL_EL1); 748 else 749 write_sysreg_s(FIELD_PREP(IPI_RR_CPU, idx) | FIELD_PREP(IPI_RR_CLUSTER, cluster), 750 SYS_IMP_APL_IPI_RR_GLOBAL_EL1); 751 isb(); 752 } 753 754 static void aic_ipi_mask(struct irq_data *d) 755 { 756 u32 irq_bit = BIT(irqd_to_hwirq(d)); 757 758 /* No specific ordering requirements needed here. */ 759 atomic_andnot(irq_bit, this_cpu_ptr(&aic_vipi_enable)); 760 } 761 762 static void aic_ipi_unmask(struct irq_data *d) 763 { 764 struct aic_irq_chip *ic = irq_data_get_irq_chip_data(d); 765 u32 irq_bit = BIT(irqd_to_hwirq(d)); 766 767 atomic_or(irq_bit, this_cpu_ptr(&aic_vipi_enable)); 768 769 /* 770 * The atomic_or() above must complete before the atomic_read() 771 * below to avoid racing aic_ipi_send_mask(). 772 */ 773 smp_mb__after_atomic(); 774 775 /* 776 * If a pending vIPI was unmasked, raise a HW IPI to ourselves. 777 * No barriers needed here since this is a self-IPI. 778 */ 779 if (atomic_read(this_cpu_ptr(&aic_vipi_flag)) & irq_bit) { 780 if (static_branch_likely(&use_fast_ipi)) 781 aic_ipi_send_fast(smp_processor_id()); 782 else 783 aic_ic_write(ic, AIC_IPI_SEND, AIC_IPI_SEND_CPU(smp_processor_id())); 784 } 785 } 786 787 static void aic_ipi_send_mask(struct irq_data *d, const struct cpumask *mask) 788 { 789 struct aic_irq_chip *ic = irq_data_get_irq_chip_data(d); 790 u32 irq_bit = BIT(irqd_to_hwirq(d)); 791 u32 send = 0; 792 int cpu; 793 unsigned long pending; 794 795 for_each_cpu(cpu, mask) { 796 /* 797 * This sequence is the mirror of the one in aic_ipi_unmask(); 798 * see the comment there. Additionally, release semantics 799 * ensure that the vIPI flag set is ordered after any shared 800 * memory accesses that precede it. This therefore also pairs 801 * with the atomic_fetch_andnot in aic_handle_ipi(). 802 */ 803 pending = atomic_fetch_or_release(irq_bit, per_cpu_ptr(&aic_vipi_flag, cpu)); 804 805 /* 806 * The atomic_fetch_or_release() above must complete before the 807 * atomic_read() below to avoid racing aic_ipi_unmask(). 808 */ 809 smp_mb__after_atomic(); 810 811 if (!(pending & irq_bit) && 812 (atomic_read(per_cpu_ptr(&aic_vipi_enable, cpu)) & irq_bit)) { 813 if (static_branch_likely(&use_fast_ipi)) 814 aic_ipi_send_fast(cpu); 815 else 816 send |= AIC_IPI_SEND_CPU(cpu); 817 } 818 } 819 820 /* 821 * The flag writes must complete before the physical IPI is issued 822 * to another CPU. This is implied by the control dependency on 823 * the result of atomic_read_acquire() above, which is itself 824 * already ordered after the vIPI flag write. 825 */ 826 if (send) 827 aic_ic_write(ic, AIC_IPI_SEND, send); 828 } 829 830 static struct irq_chip ipi_chip = { 831 .name = "AIC-IPI", 832 .irq_mask = aic_ipi_mask, 833 .irq_unmask = aic_ipi_unmask, 834 .ipi_send_mask = aic_ipi_send_mask, 835 }; 836 837 /* 838 * IPI IRQ domain 839 */ 840 841 static void aic_handle_ipi(struct pt_regs *regs) 842 { 843 int i; 844 unsigned long enabled, firing; 845 846 /* 847 * Ack the IPI. We need to order this after the AIC event read, but 848 * that is enforced by normal MMIO ordering guarantees. 849 * 850 * For the Fast IPI case, this needs to be ordered before the vIPI 851 * handling below, so we need to isb(); 852 */ 853 if (static_branch_likely(&use_fast_ipi)) { 854 write_sysreg_s(IPI_SR_PENDING, SYS_IMP_APL_IPI_SR_EL1); 855 isb(); 856 } else { 857 aic_ic_write(aic_irqc, AIC_IPI_ACK, AIC_IPI_OTHER); 858 } 859 860 /* 861 * The mask read does not need to be ordered. Only we can change 862 * our own mask anyway, so no races are possible here, as long as 863 * we are properly in the interrupt handler (which is covered by 864 * the barrier that is part of the top-level AIC handler's readl()). 865 */ 866 enabled = atomic_read(this_cpu_ptr(&aic_vipi_enable)); 867 868 /* 869 * Clear the IPIs we are about to handle. This pairs with the 870 * atomic_fetch_or_release() in aic_ipi_send_mask(), and needs to be 871 * ordered after the aic_ic_write() above (to avoid dropping vIPIs) and 872 * before IPI handling code (to avoid races handling vIPIs before they 873 * are signaled). The former is taken care of by the release semantics 874 * of the write portion, while the latter is taken care of by the 875 * acquire semantics of the read portion. 876 */ 877 firing = atomic_fetch_andnot(enabled, this_cpu_ptr(&aic_vipi_flag)) & enabled; 878 879 for_each_set_bit(i, &firing, AIC_NR_SWIPI) 880 generic_handle_domain_irq(aic_irqc->ipi_domain, i); 881 882 /* 883 * No ordering needed here; at worst this just changes the timing of 884 * when the next IPI will be delivered. 885 */ 886 if (!static_branch_likely(&use_fast_ipi)) 887 aic_ic_write(aic_irqc, AIC_IPI_MASK_CLR, AIC_IPI_OTHER); 888 } 889 890 static int aic_ipi_alloc(struct irq_domain *d, unsigned int virq, 891 unsigned int nr_irqs, void *args) 892 { 893 int i; 894 895 for (i = 0; i < nr_irqs; i++) { 896 irq_set_percpu_devid(virq + i); 897 irq_domain_set_info(d, virq + i, i, &ipi_chip, d->host_data, 898 handle_percpu_devid_irq, NULL, NULL); 899 } 900 901 return 0; 902 } 903 904 static void aic_ipi_free(struct irq_domain *d, unsigned int virq, unsigned int nr_irqs) 905 { 906 /* Not freeing IPIs */ 907 } 908 909 static const struct irq_domain_ops aic_ipi_domain_ops = { 910 .alloc = aic_ipi_alloc, 911 .free = aic_ipi_free, 912 }; 913 914 static int __init aic_init_smp(struct aic_irq_chip *irqc, struct device_node *node) 915 { 916 struct irq_domain *ipi_domain; 917 int base_ipi; 918 919 ipi_domain = irq_domain_create_linear(irqc->hw_domain->fwnode, AIC_NR_SWIPI, 920 &aic_ipi_domain_ops, irqc); 921 if (WARN_ON(!ipi_domain)) 922 return -ENODEV; 923 924 ipi_domain->flags |= IRQ_DOMAIN_FLAG_IPI_SINGLE; 925 irq_domain_update_bus_token(ipi_domain, DOMAIN_BUS_IPI); 926 927 base_ipi = __irq_domain_alloc_irqs(ipi_domain, -1, AIC_NR_SWIPI, 928 NUMA_NO_NODE, NULL, false, NULL); 929 930 if (WARN_ON(!base_ipi)) { 931 irq_domain_remove(ipi_domain); 932 return -ENODEV; 933 } 934 935 set_smp_ipi_range(base_ipi, AIC_NR_SWIPI); 936 937 irqc->ipi_domain = ipi_domain; 938 939 return 0; 940 } 941 942 static int aic_init_cpu(unsigned int cpu) 943 { 944 /* Mask all hard-wired per-CPU IRQ/FIQ sources */ 945 946 /* Pending Fast IPI FIQs */ 947 write_sysreg_s(IPI_SR_PENDING, SYS_IMP_APL_IPI_SR_EL1); 948 949 /* Timer FIQs */ 950 sysreg_clear_set(cntp_ctl_el0, 0, ARCH_TIMER_CTRL_IT_MASK); 951 sysreg_clear_set(cntv_ctl_el0, 0, ARCH_TIMER_CTRL_IT_MASK); 952 953 /* EL2-only (VHE mode) IRQ sources */ 954 if (is_kernel_in_hyp_mode()) { 955 /* Guest timers */ 956 sysreg_clear_set_s(SYS_IMP_APL_VM_TMR_FIQ_ENA_EL2, 957 VM_TMR_FIQ_ENABLE_V | VM_TMR_FIQ_ENABLE_P, 0); 958 959 /* vGIC maintenance IRQ */ 960 sysreg_clear_set_s(SYS_ICH_HCR_EL2, ICH_HCR_EN, 0); 961 } 962 963 /* PMC FIQ */ 964 sysreg_clear_set_s(SYS_IMP_APL_PMCR0_EL1, PMCR0_IMODE | PMCR0_IACT, 965 FIELD_PREP(PMCR0_IMODE, PMCR0_IMODE_OFF)); 966 967 /* Uncore PMC FIQ */ 968 sysreg_clear_set_s(SYS_IMP_APL_UPMCR0_EL1, UPMCR0_IMODE, 969 FIELD_PREP(UPMCR0_IMODE, UPMCR0_IMODE_OFF)); 970 971 /* Commit all of the above */ 972 isb(); 973 974 if (aic_irqc->info.version == 1) { 975 /* 976 * Make sure the kernel's idea of logical CPU order is the same as AIC's 977 * If we ever end up with a mismatch here, we will have to introduce 978 * a mapping table similar to what other irqchip drivers do. 979 */ 980 WARN_ON(aic_ic_read(aic_irqc, AIC_WHOAMI) != smp_processor_id()); 981 982 /* 983 * Always keep IPIs unmasked at the hardware level (except auto-masking 984 * by AIC during processing). We manage masks at the vIPI level. 985 * These registers only exist on AICv1, AICv2 always uses fast IPIs. 986 */ 987 aic_ic_write(aic_irqc, AIC_IPI_ACK, AIC_IPI_SELF | AIC_IPI_OTHER); 988 if (static_branch_likely(&use_fast_ipi)) { 989 aic_ic_write(aic_irqc, AIC_IPI_MASK_SET, AIC_IPI_SELF | AIC_IPI_OTHER); 990 } else { 991 aic_ic_write(aic_irqc, AIC_IPI_MASK_SET, AIC_IPI_SELF); 992 aic_ic_write(aic_irqc, AIC_IPI_MASK_CLR, AIC_IPI_OTHER); 993 } 994 } 995 996 /* Initialize the local mask state */ 997 __this_cpu_write(aic_fiq_unmasked, 0); 998 999 return 0; 1000 } 1001 1002 static struct gic_kvm_info vgic_info __initdata = { 1003 .type = GIC_V3, 1004 .no_maint_irq_mask = true, 1005 .no_hw_deactivation = true, 1006 }; 1007 1008 static void build_fiq_affinity(struct aic_irq_chip *ic, struct device_node *aff) 1009 { 1010 int i, n; 1011 u32 fiq; 1012 1013 if (of_property_read_u32(aff, "apple,fiq-index", &fiq) || 1014 WARN_ON(fiq >= AIC_NR_FIQ) || ic->fiq_aff[fiq]) 1015 return; 1016 1017 n = of_property_count_elems_of_size(aff, "cpus", sizeof(u32)); 1018 if (WARN_ON(n < 0)) 1019 return; 1020 1021 ic->fiq_aff[fiq] = kzalloc(sizeof(*ic->fiq_aff[fiq]), GFP_KERNEL); 1022 if (!ic->fiq_aff[fiq]) 1023 return; 1024 1025 for (i = 0; i < n; i++) { 1026 struct device_node *cpu_node; 1027 u32 cpu_phandle; 1028 int cpu; 1029 1030 if (of_property_read_u32_index(aff, "cpus", i, &cpu_phandle)) 1031 continue; 1032 1033 cpu_node = of_find_node_by_phandle(cpu_phandle); 1034 if (WARN_ON(!cpu_node)) 1035 continue; 1036 1037 cpu = of_cpu_node_to_id(cpu_node); 1038 of_node_put(cpu_node); 1039 if (WARN_ON(cpu < 0)) 1040 continue; 1041 1042 cpumask_set_cpu(cpu, &ic->fiq_aff[fiq]->aff); 1043 } 1044 } 1045 1046 static int __init aic_of_ic_init(struct device_node *node, struct device_node *parent) 1047 { 1048 int i, die; 1049 u32 off, start_off; 1050 void __iomem *regs; 1051 struct aic_irq_chip *irqc; 1052 struct device_node *affs; 1053 const struct of_device_id *match; 1054 1055 regs = of_iomap(node, 0); 1056 if (WARN_ON(!regs)) 1057 return -EIO; 1058 1059 irqc = kzalloc(sizeof(*irqc), GFP_KERNEL); 1060 if (!irqc) { 1061 iounmap(regs); 1062 return -ENOMEM; 1063 } 1064 1065 irqc->base = regs; 1066 1067 match = of_match_node(aic_info_match, node); 1068 if (!match) 1069 goto err_unmap; 1070 1071 irqc->info = *(struct aic_info *)match->data; 1072 1073 aic_irqc = irqc; 1074 1075 switch (irqc->info.version) { 1076 case 1: { 1077 u32 info; 1078 1079 info = aic_ic_read(irqc, AIC_INFO); 1080 irqc->nr_irq = FIELD_GET(AIC_INFO_NR_IRQ, info); 1081 irqc->max_irq = AIC_MAX_IRQ; 1082 irqc->nr_die = irqc->max_die = 1; 1083 1084 off = start_off = irqc->info.target_cpu; 1085 off += sizeof(u32) * irqc->max_irq; /* TARGET_CPU */ 1086 1087 irqc->event = irqc->base; 1088 1089 break; 1090 } 1091 case 2: { 1092 u32 info1, info3; 1093 1094 info1 = aic_ic_read(irqc, AIC2_INFO1); 1095 info3 = aic_ic_read(irqc, AIC2_INFO3); 1096 1097 irqc->nr_irq = FIELD_GET(AIC2_INFO1_NR_IRQ, info1); 1098 irqc->max_irq = FIELD_GET(AIC2_INFO3_MAX_IRQ, info3); 1099 irqc->nr_die = FIELD_GET(AIC2_INFO1_LAST_DIE, info1) + 1; 1100 irqc->max_die = FIELD_GET(AIC2_INFO3_MAX_DIE, info3); 1101 1102 off = start_off = irqc->info.irq_cfg; 1103 off += sizeof(u32) * irqc->max_irq; /* IRQ_CFG */ 1104 1105 irqc->event = of_iomap(node, 1); 1106 if (WARN_ON(!irqc->event)) 1107 goto err_unmap; 1108 1109 break; 1110 } 1111 } 1112 1113 irqc->info.sw_set = off; 1114 off += sizeof(u32) * (irqc->max_irq >> 5); /* SW_SET */ 1115 irqc->info.sw_clr = off; 1116 off += sizeof(u32) * (irqc->max_irq >> 5); /* SW_CLR */ 1117 irqc->info.mask_set = off; 1118 off += sizeof(u32) * (irqc->max_irq >> 5); /* MASK_SET */ 1119 irqc->info.mask_clr = off; 1120 off += sizeof(u32) * (irqc->max_irq >> 5); /* MASK_CLR */ 1121 off += sizeof(u32) * (irqc->max_irq >> 5); /* HW_STATE */ 1122 1123 if (irqc->info.fast_ipi) 1124 static_branch_enable(&use_fast_ipi); 1125 else 1126 static_branch_disable(&use_fast_ipi); 1127 1128 irqc->info.die_stride = off - start_off; 1129 1130 irqc->hw_domain = irq_domain_create_tree(of_node_to_fwnode(node), 1131 &aic_irq_domain_ops, irqc); 1132 if (WARN_ON(!irqc->hw_domain)) 1133 goto err_unmap; 1134 1135 irq_domain_update_bus_token(irqc->hw_domain, DOMAIN_BUS_WIRED); 1136 1137 if (aic_init_smp(irqc, node)) 1138 goto err_remove_domain; 1139 1140 affs = of_get_child_by_name(node, "affinities"); 1141 if (affs) { 1142 struct device_node *chld; 1143 1144 for_each_child_of_node(affs, chld) 1145 build_fiq_affinity(irqc, chld); 1146 } 1147 of_node_put(affs); 1148 1149 set_handle_irq(aic_handle_irq); 1150 set_handle_fiq(aic_handle_fiq); 1151 1152 off = 0; 1153 for (die = 0; die < irqc->nr_die; die++) { 1154 for (i = 0; i < BITS_TO_U32(irqc->nr_irq); i++) 1155 aic_ic_write(irqc, irqc->info.mask_set + off + i * 4, U32_MAX); 1156 for (i = 0; i < BITS_TO_U32(irqc->nr_irq); i++) 1157 aic_ic_write(irqc, irqc->info.sw_clr + off + i * 4, U32_MAX); 1158 if (irqc->info.target_cpu) 1159 for (i = 0; i < irqc->nr_irq; i++) 1160 aic_ic_write(irqc, irqc->info.target_cpu + off + i * 4, 1); 1161 off += irqc->info.die_stride; 1162 } 1163 1164 if (irqc->info.version == 2) { 1165 u32 config = aic_ic_read(irqc, AIC2_CONFIG); 1166 1167 config |= AIC2_CONFIG_ENABLE; 1168 aic_ic_write(irqc, AIC2_CONFIG, config); 1169 } 1170 1171 if (!is_kernel_in_hyp_mode()) 1172 pr_info("Kernel running in EL1, mapping interrupts"); 1173 1174 if (static_branch_likely(&use_fast_ipi)) 1175 pr_info("Using Fast IPIs"); 1176 1177 cpuhp_setup_state(CPUHP_AP_IRQ_APPLE_AIC_STARTING, 1178 "irqchip/apple-aic/ipi:starting", 1179 aic_init_cpu, NULL); 1180 1181 vgic_set_kvm_info(&vgic_info); 1182 1183 pr_info("Initialized with %d/%d IRQs * %d/%d die(s), %d FIQs, %d vIPIs", 1184 irqc->nr_irq, irqc->max_irq, irqc->nr_die, irqc->max_die, AIC_NR_FIQ, AIC_NR_SWIPI); 1185 1186 return 0; 1187 1188 err_remove_domain: 1189 irq_domain_remove(irqc->hw_domain); 1190 err_unmap: 1191 if (irqc->event && irqc->event != irqc->base) 1192 iounmap(irqc->event); 1193 iounmap(irqc->base); 1194 kfree(irqc); 1195 return -ENODEV; 1196 } 1197 1198 IRQCHIP_DECLARE(apple_aic, "apple,aic", aic_of_ic_init); 1199 IRQCHIP_DECLARE(apple_aic2, "apple,aic2", aic_of_ic_init); 1200