1 /* 2 * QEMU PowerPC sPAPR IRQ interface 3 * 4 * Copyright (c) 2018, IBM Corporation. 5 * 6 * This code is licensed under the GPL version 2 or later. See the 7 * COPYING file in the top-level directory. 8 */ 9 10 #include "qemu/osdep.h" 11 #include "qemu/log.h" 12 #include "qemu/error-report.h" 13 #include "qapi/error.h" 14 #include "hw/irq.h" 15 #include "hw/ppc/spapr.h" 16 #include "hw/ppc/spapr_cpu_core.h" 17 #include "hw/ppc/spapr_xive.h" 18 #include "hw/ppc/xics.h" 19 #include "hw/ppc/xics_spapr.h" 20 #include "hw/qdev-properties.h" 21 #include "cpu-models.h" 22 #include "sysemu/kvm.h" 23 24 #include "trace.h" 25 26 static const TypeInfo spapr_intc_info = { 27 .name = TYPE_SPAPR_INTC, 28 .parent = TYPE_INTERFACE, 29 .class_size = sizeof(SpaprInterruptControllerClass), 30 }; 31 32 void spapr_irq_msi_init(SpaprMachineState *spapr, uint32_t nr_msis) 33 { 34 spapr->irq_map_nr = nr_msis; 35 spapr->irq_map = bitmap_new(spapr->irq_map_nr); 36 } 37 38 int spapr_irq_msi_alloc(SpaprMachineState *spapr, uint32_t num, bool align, 39 Error **errp) 40 { 41 int irq; 42 43 /* 44 * The 'align_mask' parameter of bitmap_find_next_zero_area() 45 * should be one less than a power of 2; 0 means no 46 * alignment. Adapt the 'align' value of the former allocator 47 * to fit the requirements of bitmap_find_next_zero_area() 48 */ 49 align -= 1; 50 51 irq = bitmap_find_next_zero_area(spapr->irq_map, spapr->irq_map_nr, 0, num, 52 align); 53 if (irq == spapr->irq_map_nr) { 54 error_setg(errp, "can't find a free %d-IRQ block", num); 55 return -1; 56 } 57 58 bitmap_set(spapr->irq_map, irq, num); 59 60 return irq + SPAPR_IRQ_MSI; 61 } 62 63 void spapr_irq_msi_free(SpaprMachineState *spapr, int irq, uint32_t num) 64 { 65 bitmap_clear(spapr->irq_map, irq - SPAPR_IRQ_MSI, num); 66 } 67 68 static void spapr_irq_init_kvm(SpaprMachineState *spapr, 69 SpaprIrq *irq, Error **errp) 70 { 71 MachineState *machine = MACHINE(spapr); 72 Error *local_err = NULL; 73 74 if (kvm_enabled() && machine_kernel_irqchip_allowed(machine)) { 75 irq->init_kvm(spapr, &local_err); 76 if (local_err && machine_kernel_irqchip_required(machine)) { 77 error_prepend(&local_err, 78 "kernel_irqchip requested but unavailable: "); 79 error_propagate(errp, local_err); 80 return; 81 } 82 83 if (!local_err) { 84 return; 85 } 86 87 /* 88 * We failed to initialize the KVM device, fallback to 89 * emulated mode 90 */ 91 error_prepend(&local_err, "kernel_irqchip allowed but unavailable: "); 92 error_append_hint(&local_err, "Falling back to kernel-irqchip=off\n"); 93 warn_report_err(local_err); 94 } 95 } 96 97 /* 98 * XICS IRQ backend. 99 */ 100 101 static int spapr_irq_post_load_xics(SpaprMachineState *spapr, int version_id) 102 { 103 if (!kvm_irqchip_in_kernel()) { 104 CPUState *cs; 105 CPU_FOREACH(cs) { 106 PowerPCCPU *cpu = POWERPC_CPU(cs); 107 icp_resend(spapr_cpu_state(cpu)->icp); 108 } 109 } 110 return 0; 111 } 112 113 static void spapr_irq_reset_xics(SpaprMachineState *spapr, Error **errp) 114 { 115 Error *local_err = NULL; 116 117 spapr_irq_init_kvm(spapr, &spapr_irq_xics, &local_err); 118 if (local_err) { 119 error_propagate(errp, local_err); 120 return; 121 } 122 } 123 124 static void spapr_irq_init_kvm_xics(SpaprMachineState *spapr, Error **errp) 125 { 126 if (kvm_enabled()) { 127 xics_kvm_connect(SPAPR_INTC(spapr->ics), errp); 128 } 129 } 130 131 SpaprIrq spapr_irq_xics = { 132 .nr_xirqs = SPAPR_NR_XIRQS, 133 .nr_msis = SPAPR_NR_MSIS, 134 .xics = true, 135 .xive = false, 136 137 .post_load = spapr_irq_post_load_xics, 138 .reset = spapr_irq_reset_xics, 139 .init_kvm = spapr_irq_init_kvm_xics, 140 }; 141 142 /* 143 * XIVE IRQ backend. 144 */ 145 146 static int spapr_irq_post_load_xive(SpaprMachineState *spapr, int version_id) 147 { 148 return spapr_xive_post_load(spapr->xive, version_id); 149 } 150 151 static void spapr_irq_reset_xive(SpaprMachineState *spapr, Error **errp) 152 { 153 CPUState *cs; 154 Error *local_err = NULL; 155 156 CPU_FOREACH(cs) { 157 PowerPCCPU *cpu = POWERPC_CPU(cs); 158 159 /* (TCG) Set the OS CAM line of the thread interrupt context. */ 160 spapr_xive_set_tctx_os_cam(spapr_cpu_state(cpu)->tctx); 161 } 162 163 spapr_irq_init_kvm(spapr, &spapr_irq_xive, &local_err); 164 if (local_err) { 165 error_propagate(errp, local_err); 166 return; 167 } 168 169 /* Activate the XIVE MMIOs */ 170 spapr_xive_mmio_set_enabled(spapr->xive, true); 171 } 172 173 static void spapr_irq_init_kvm_xive(SpaprMachineState *spapr, Error **errp) 174 { 175 if (kvm_enabled()) { 176 kvmppc_xive_connect(SPAPR_INTC(spapr->xive), errp); 177 } 178 } 179 180 SpaprIrq spapr_irq_xive = { 181 .nr_xirqs = SPAPR_NR_XIRQS, 182 .nr_msis = SPAPR_NR_MSIS, 183 .xics = false, 184 .xive = true, 185 186 .post_load = spapr_irq_post_load_xive, 187 .reset = spapr_irq_reset_xive, 188 .init_kvm = spapr_irq_init_kvm_xive, 189 }; 190 191 /* 192 * Dual XIVE and XICS IRQ backend. 193 * 194 * Both interrupt mode, XIVE and XICS, objects are created but the 195 * machine starts in legacy interrupt mode (XICS). It can be changed 196 * by the CAS negotiation process and, in that case, the new mode is 197 * activated after an extra machine reset. 198 */ 199 200 /* 201 * Returns the sPAPR IRQ backend negotiated by CAS. XICS is the 202 * default. 203 */ 204 static SpaprIrq *spapr_irq_current(SpaprMachineState *spapr) 205 { 206 return spapr_ovec_test(spapr->ov5_cas, OV5_XIVE_EXPLOIT) ? 207 &spapr_irq_xive : &spapr_irq_xics; 208 } 209 210 static int spapr_irq_post_load_dual(SpaprMachineState *spapr, int version_id) 211 { 212 /* 213 * Force a reset of the XIVE backend after migration. The machine 214 * defaults to XICS at startup. 215 */ 216 if (spapr_ovec_test(spapr->ov5_cas, OV5_XIVE_EXPLOIT)) { 217 if (kvm_irqchip_in_kernel()) { 218 xics_kvm_disconnect(SPAPR_INTC(spapr->ics)); 219 } 220 spapr_irq_xive.reset(spapr, &error_fatal); 221 } 222 223 return spapr_irq_current(spapr)->post_load(spapr, version_id); 224 } 225 226 static void spapr_irq_reset_dual(SpaprMachineState *spapr, Error **errp) 227 { 228 /* 229 * Deactivate the XIVE MMIOs. The XIVE backend will reenable them 230 * if selected. 231 */ 232 spapr_xive_mmio_set_enabled(spapr->xive, false); 233 234 /* Destroy all KVM devices */ 235 if (kvm_irqchip_in_kernel()) { 236 xics_kvm_disconnect(SPAPR_INTC(spapr->ics)); 237 kvmppc_xive_disconnect(SPAPR_INTC(spapr->xive)); 238 } 239 240 spapr_irq_current(spapr)->reset(spapr, errp); 241 } 242 243 /* 244 * Define values in sync with the XIVE and XICS backend 245 */ 246 SpaprIrq spapr_irq_dual = { 247 .nr_xirqs = SPAPR_NR_XIRQS, 248 .nr_msis = SPAPR_NR_MSIS, 249 .xics = true, 250 .xive = true, 251 252 .post_load = spapr_irq_post_load_dual, 253 .reset = spapr_irq_reset_dual, 254 .init_kvm = NULL, /* should not be used */ 255 }; 256 257 258 static int spapr_irq_check(SpaprMachineState *spapr, Error **errp) 259 { 260 MachineState *machine = MACHINE(spapr); 261 262 /* 263 * Sanity checks on non-P9 machines. On these, XIVE is not 264 * advertised, see spapr_dt_ov5_platform_support() 265 */ 266 if (!ppc_type_check_compat(machine->cpu_type, CPU_POWERPC_LOGICAL_3_00, 267 0, spapr->max_compat_pvr)) { 268 /* 269 * If the 'dual' interrupt mode is selected, force XICS as CAS 270 * negotiation is useless. 271 */ 272 if (spapr->irq == &spapr_irq_dual) { 273 spapr->irq = &spapr_irq_xics; 274 return 0; 275 } 276 277 /* 278 * Non-P9 machines using only XIVE is a bogus setup. We have two 279 * scenarios to take into account because of the compat mode: 280 * 281 * 1. POWER7/8 machines should fail to init later on when creating 282 * the XIVE interrupt presenters because a POWER9 exception 283 * model is required. 284 285 * 2. POWER9 machines using the POWER8 compat mode won't fail and 286 * will let the OS boot with a partial XIVE setup : DT 287 * properties but no hcalls. 288 * 289 * To cover both and not confuse the OS, add an early failure in 290 * QEMU. 291 */ 292 if (spapr->irq == &spapr_irq_xive) { 293 error_setg(errp, "XIVE-only machines require a POWER9 CPU"); 294 return -1; 295 } 296 } 297 298 /* 299 * On a POWER9 host, some older KVM XICS devices cannot be destroyed and 300 * re-created. Detect that early to avoid QEMU to exit later when the 301 * guest reboots. 302 */ 303 if (kvm_enabled() && 304 spapr->irq == &spapr_irq_dual && 305 machine_kernel_irqchip_required(machine) && 306 xics_kvm_has_broken_disconnect(spapr)) { 307 error_setg(errp, "KVM is too old to support ic-mode=dual,kernel-irqchip=on"); 308 return -1; 309 } 310 311 return 0; 312 } 313 314 /* 315 * sPAPR IRQ frontend routines for devices 316 */ 317 #define ALL_INTCS(spapr_) \ 318 { SPAPR_INTC((spapr_)->ics), SPAPR_INTC((spapr_)->xive), } 319 320 int spapr_irq_cpu_intc_create(SpaprMachineState *spapr, 321 PowerPCCPU *cpu, Error **errp) 322 { 323 SpaprInterruptController *intcs[] = ALL_INTCS(spapr); 324 int i; 325 int rc; 326 327 for (i = 0; i < ARRAY_SIZE(intcs); i++) { 328 SpaprInterruptController *intc = intcs[i]; 329 if (intc) { 330 SpaprInterruptControllerClass *sicc = SPAPR_INTC_GET_CLASS(intc); 331 rc = sicc->cpu_intc_create(intc, cpu, errp); 332 if (rc < 0) { 333 return rc; 334 } 335 } 336 } 337 338 return 0; 339 } 340 341 static void spapr_set_irq(void *opaque, int irq, int level) 342 { 343 SpaprMachineState *spapr = SPAPR_MACHINE(opaque); 344 SpaprInterruptControllerClass *sicc 345 = SPAPR_INTC_GET_CLASS(spapr->active_intc); 346 347 sicc->set_irq(spapr->active_intc, irq, level); 348 } 349 350 void spapr_irq_print_info(SpaprMachineState *spapr, Monitor *mon) 351 { 352 SpaprInterruptControllerClass *sicc 353 = SPAPR_INTC_GET_CLASS(spapr->active_intc); 354 355 sicc->print_info(spapr->active_intc, mon); 356 } 357 358 void spapr_irq_dt(SpaprMachineState *spapr, uint32_t nr_servers, 359 void *fdt, uint32_t phandle) 360 { 361 SpaprInterruptControllerClass *sicc 362 = SPAPR_INTC_GET_CLASS(spapr->active_intc); 363 364 sicc->dt(spapr->active_intc, nr_servers, fdt, phandle); 365 } 366 367 void spapr_irq_init(SpaprMachineState *spapr, Error **errp) 368 { 369 MachineState *machine = MACHINE(spapr); 370 371 if (machine_kernel_irqchip_split(machine)) { 372 error_setg(errp, "kernel_irqchip split mode not supported on pseries"); 373 return; 374 } 375 376 if (!kvm_enabled() && machine_kernel_irqchip_required(machine)) { 377 error_setg(errp, 378 "kernel_irqchip requested but only available with KVM"); 379 return; 380 } 381 382 if (spapr_irq_check(spapr, errp) < 0) { 383 return; 384 } 385 386 /* Initialize the MSI IRQ allocator. */ 387 if (!SPAPR_MACHINE_GET_CLASS(spapr)->legacy_irq_allocation) { 388 spapr_irq_msi_init(spapr, spapr->irq->nr_msis); 389 } 390 391 if (spapr->irq->xics) { 392 Error *local_err = NULL; 393 Object *obj; 394 395 obj = object_new(TYPE_ICS_SPAPR); 396 object_property_add_child(OBJECT(spapr), "ics", obj, &local_err); 397 if (local_err) { 398 error_propagate(errp, local_err); 399 return; 400 } 401 402 object_property_add_const_link(obj, ICS_PROP_XICS, OBJECT(spapr), 403 &local_err); 404 if (local_err) { 405 error_propagate(errp, local_err); 406 return; 407 } 408 409 object_property_set_int(obj, spapr->irq->nr_xirqs, "nr-irqs", 410 &local_err); 411 if (local_err) { 412 error_propagate(errp, local_err); 413 return; 414 } 415 416 object_property_set_bool(obj, true, "realized", &local_err); 417 if (local_err) { 418 error_propagate(errp, local_err); 419 return; 420 } 421 422 spapr->ics = ICS_SPAPR(obj); 423 } 424 425 if (spapr->irq->xive) { 426 uint32_t nr_servers = spapr_max_server_number(spapr); 427 DeviceState *dev; 428 int i; 429 430 dev = qdev_create(NULL, TYPE_SPAPR_XIVE); 431 qdev_prop_set_uint32(dev, "nr-irqs", 432 spapr->irq->nr_xirqs + SPAPR_XIRQ_BASE); 433 /* 434 * 8 XIVE END structures per CPU. One for each available 435 * priority 436 */ 437 qdev_prop_set_uint32(dev, "nr-ends", nr_servers << 3); 438 qdev_init_nofail(dev); 439 440 spapr->xive = SPAPR_XIVE(dev); 441 442 /* Enable the CPU IPIs */ 443 for (i = 0; i < nr_servers; ++i) { 444 SpaprInterruptControllerClass *sicc 445 = SPAPR_INTC_GET_CLASS(spapr->xive); 446 447 if (sicc->claim_irq(SPAPR_INTC(spapr->xive), SPAPR_IRQ_IPI + i, 448 false, errp) < 0) { 449 return; 450 } 451 } 452 453 spapr_xive_hcall_init(spapr); 454 } 455 456 spapr->qirqs = qemu_allocate_irqs(spapr_set_irq, spapr, 457 spapr->irq->nr_xirqs + SPAPR_XIRQ_BASE); 458 } 459 460 int spapr_irq_claim(SpaprMachineState *spapr, int irq, bool lsi, Error **errp) 461 { 462 SpaprInterruptController *intcs[] = ALL_INTCS(spapr); 463 int i; 464 int rc; 465 466 assert(irq >= SPAPR_XIRQ_BASE); 467 assert(irq < (spapr->irq->nr_xirqs + SPAPR_XIRQ_BASE)); 468 469 for (i = 0; i < ARRAY_SIZE(intcs); i++) { 470 SpaprInterruptController *intc = intcs[i]; 471 if (intc) { 472 SpaprInterruptControllerClass *sicc = SPAPR_INTC_GET_CLASS(intc); 473 rc = sicc->claim_irq(intc, irq, lsi, errp); 474 if (rc < 0) { 475 return rc; 476 } 477 } 478 } 479 480 return 0; 481 } 482 483 void spapr_irq_free(SpaprMachineState *spapr, int irq, int num) 484 { 485 SpaprInterruptController *intcs[] = ALL_INTCS(spapr); 486 int i, j; 487 488 assert(irq >= SPAPR_XIRQ_BASE); 489 assert((irq + num) <= (spapr->irq->nr_xirqs + SPAPR_XIRQ_BASE)); 490 491 for (i = irq; i < (irq + num); i++) { 492 for (j = 0; j < ARRAY_SIZE(intcs); j++) { 493 SpaprInterruptController *intc = intcs[j]; 494 495 if (intc) { 496 SpaprInterruptControllerClass *sicc 497 = SPAPR_INTC_GET_CLASS(intc); 498 sicc->free_irq(intc, i); 499 } 500 } 501 } 502 } 503 504 qemu_irq spapr_qirq(SpaprMachineState *spapr, int irq) 505 { 506 /* 507 * This interface is basically for VIO and PHB devices to find the 508 * right qemu_irq to manipulate, so we only allow access to the 509 * external irqs for now. Currently anything which needs to 510 * access the IPIs most naturally gets there via the guest side 511 * interfaces, we can change this if we need to in future. 512 */ 513 assert(irq >= SPAPR_XIRQ_BASE); 514 assert(irq < (spapr->irq->nr_xirqs + SPAPR_XIRQ_BASE)); 515 516 if (spapr->ics) { 517 assert(ics_valid_irq(spapr->ics, irq)); 518 } 519 if (spapr->xive) { 520 assert(irq < spapr->xive->nr_irqs); 521 assert(xive_eas_is_valid(&spapr->xive->eat[irq])); 522 } 523 524 return spapr->qirqs[irq]; 525 } 526 527 int spapr_irq_post_load(SpaprMachineState *spapr, int version_id) 528 { 529 spapr_irq_update_active_intc(spapr); 530 return spapr->irq->post_load(spapr, version_id); 531 } 532 533 void spapr_irq_reset(SpaprMachineState *spapr, Error **errp) 534 { 535 assert(!spapr->irq_map || bitmap_empty(spapr->irq_map, spapr->irq_map_nr)); 536 537 spapr_irq_update_active_intc(spapr); 538 539 if (spapr->irq->reset) { 540 spapr->irq->reset(spapr, errp); 541 } 542 } 543 544 int spapr_irq_get_phandle(SpaprMachineState *spapr, void *fdt, Error **errp) 545 { 546 const char *nodename = "interrupt-controller"; 547 int offset, phandle; 548 549 offset = fdt_subnode_offset(fdt, 0, nodename); 550 if (offset < 0) { 551 error_setg(errp, "Can't find node \"%s\": %s", 552 nodename, fdt_strerror(offset)); 553 return -1; 554 } 555 556 phandle = fdt_get_phandle(fdt, offset); 557 if (!phandle) { 558 error_setg(errp, "Can't get phandle of node \"%s\"", nodename); 559 return -1; 560 } 561 562 return phandle; 563 } 564 565 static void set_active_intc(SpaprMachineState *spapr, 566 SpaprInterruptController *new_intc) 567 { 568 SpaprInterruptControllerClass *sicc; 569 570 assert(new_intc); 571 572 if (new_intc == spapr->active_intc) { 573 /* Nothing to do */ 574 return; 575 } 576 577 if (spapr->active_intc) { 578 sicc = SPAPR_INTC_GET_CLASS(spapr->active_intc); 579 if (sicc->deactivate) { 580 sicc->deactivate(spapr->active_intc); 581 } 582 } 583 584 sicc = SPAPR_INTC_GET_CLASS(new_intc); 585 if (sicc->activate) { 586 sicc->activate(new_intc, &error_fatal); 587 } 588 589 spapr->active_intc = new_intc; 590 } 591 592 void spapr_irq_update_active_intc(SpaprMachineState *spapr) 593 { 594 SpaprInterruptController *new_intc; 595 596 if (!spapr->ics) { 597 /* 598 * XXX before we run CAS, ov5_cas is initialized empty, which 599 * indicates XICS, even if we have ic-mode=xive. TODO: clean 600 * up the CAS path so that we have a clearer way of handling 601 * this. 602 */ 603 new_intc = SPAPR_INTC(spapr->xive); 604 } else if (spapr_ovec_test(spapr->ov5_cas, OV5_XIVE_EXPLOIT)) { 605 new_intc = SPAPR_INTC(spapr->xive); 606 } else { 607 new_intc = SPAPR_INTC(spapr->ics); 608 } 609 610 set_active_intc(spapr, new_intc); 611 } 612 613 /* 614 * XICS legacy routines - to deprecate one day 615 */ 616 617 static int ics_find_free_block(ICSState *ics, int num, int alignnum) 618 { 619 int first, i; 620 621 for (first = 0; first < ics->nr_irqs; first += alignnum) { 622 if (num > (ics->nr_irqs - first)) { 623 return -1; 624 } 625 for (i = first; i < first + num; ++i) { 626 if (!ics_irq_free(ics, i)) { 627 break; 628 } 629 } 630 if (i == (first + num)) { 631 return first; 632 } 633 } 634 635 return -1; 636 } 637 638 int spapr_irq_find(SpaprMachineState *spapr, int num, bool align, Error **errp) 639 { 640 ICSState *ics = spapr->ics; 641 int first = -1; 642 643 assert(ics); 644 645 /* 646 * MSIMesage::data is used for storing VIRQ so 647 * it has to be aligned to num to support multiple 648 * MSI vectors. MSI-X is not affected by this. 649 * The hint is used for the first IRQ, the rest should 650 * be allocated continuously. 651 */ 652 if (align) { 653 assert((num == 1) || (num == 2) || (num == 4) || 654 (num == 8) || (num == 16) || (num == 32)); 655 first = ics_find_free_block(ics, num, num); 656 } else { 657 first = ics_find_free_block(ics, num, 1); 658 } 659 660 if (first < 0) { 661 error_setg(errp, "can't find a free %d-IRQ block", num); 662 return -1; 663 } 664 665 return first + ics->offset; 666 } 667 668 #define SPAPR_IRQ_XICS_LEGACY_NR_XIRQS 0x400 669 670 SpaprIrq spapr_irq_xics_legacy = { 671 .nr_xirqs = SPAPR_IRQ_XICS_LEGACY_NR_XIRQS, 672 .nr_msis = SPAPR_IRQ_XICS_LEGACY_NR_XIRQS, 673 .xics = true, 674 .xive = false, 675 676 .post_load = spapr_irq_post_load_xics, 677 .reset = spapr_irq_reset_xics, 678 .init_kvm = spapr_irq_init_kvm_xics, 679 }; 680 681 static void spapr_irq_register_types(void) 682 { 683 type_register_static(&spapr_intc_info); 684 } 685 686 type_init(spapr_irq_register_types) 687