1 /* 2 * QEMU KVM support 3 * 4 * Copyright IBM, Corp. 2008 5 * Red Hat, Inc. 2008 6 * 7 * Authors: 8 * Anthony Liguori <aliguori@us.ibm.com> 9 * Glauber Costa <gcosta@redhat.com> 10 * 11 * This work is licensed under the terms of the GNU GPL, version 2 or later. 12 * See the COPYING file in the top-level directory. 13 * 14 */ 15 16 #include "qemu/osdep.h" 17 #include <sys/ioctl.h> 18 19 #include <linux/kvm.h> 20 21 #include "qemu-common.h" 22 #include "qemu/atomic.h" 23 #include "qemu/option.h" 24 #include "qemu/config-file.h" 25 #include "qemu/error-report.h" 26 #include "qapi/error.h" 27 #include "hw/hw.h" 28 #include "hw/pci/msi.h" 29 #include "hw/pci/msix.h" 30 #include "hw/s390x/adapter.h" 31 #include "exec/gdbstub.h" 32 #include "sysemu/kvm_int.h" 33 #include "sysemu/cpus.h" 34 #include "qemu/bswap.h" 35 #include "exec/memory.h" 36 #include "exec/ram_addr.h" 37 #include "exec/address-spaces.h" 38 #include "qemu/event_notifier.h" 39 #include "trace.h" 40 #include "hw/irq.h" 41 42 #include "hw/boards.h" 43 44 /* This check must be after config-host.h is included */ 45 #ifdef CONFIG_EVENTFD 46 #include <sys/eventfd.h> 47 #endif 48 49 /* KVM uses PAGE_SIZE in its definition of KVM_COALESCED_MMIO_MAX. We 50 * need to use the real host PAGE_SIZE, as that's what KVM will use. 51 */ 52 #define PAGE_SIZE getpagesize() 53 54 //#define DEBUG_KVM 55 56 #ifdef DEBUG_KVM 57 #define DPRINTF(fmt, ...) \ 58 do { fprintf(stderr, fmt, ## __VA_ARGS__); } while (0) 59 #else 60 #define DPRINTF(fmt, ...) \ 61 do { } while (0) 62 #endif 63 64 #define KVM_MSI_HASHTAB_SIZE 256 65 66 struct KVMParkedVcpu { 67 unsigned long vcpu_id; 68 int kvm_fd; 69 QLIST_ENTRY(KVMParkedVcpu) node; 70 }; 71 72 struct KVMState 73 { 74 AccelState parent_obj; 75 76 int nr_slots; 77 int fd; 78 int vmfd; 79 int coalesced_mmio; 80 struct kvm_coalesced_mmio_ring *coalesced_mmio_ring; 81 bool coalesced_flush_in_progress; 82 int vcpu_events; 83 int robust_singlestep; 84 int debugregs; 85 #ifdef KVM_CAP_SET_GUEST_DEBUG 86 struct kvm_sw_breakpoint_head kvm_sw_breakpoints; 87 #endif 88 int many_ioeventfds; 89 int intx_set_mask; 90 bool sync_mmu; 91 /* The man page (and posix) say ioctl numbers are signed int, but 92 * they're not. Linux, glibc and *BSD all treat ioctl numbers as 93 * unsigned, and treating them as signed here can break things */ 94 unsigned irq_set_ioctl; 95 unsigned int sigmask_len; 96 GHashTable *gsimap; 97 #ifdef KVM_CAP_IRQ_ROUTING 98 struct kvm_irq_routing *irq_routes; 99 int nr_allocated_irq_routes; 100 unsigned long *used_gsi_bitmap; 101 unsigned int gsi_count; 102 QTAILQ_HEAD(msi_hashtab, KVMMSIRoute) msi_hashtab[KVM_MSI_HASHTAB_SIZE]; 103 #endif 104 KVMMemoryListener memory_listener; 105 QLIST_HEAD(, KVMParkedVcpu) kvm_parked_vcpus; 106 }; 107 108 KVMState *kvm_state; 109 bool kvm_kernel_irqchip; 110 bool kvm_split_irqchip; 111 bool kvm_async_interrupts_allowed; 112 bool kvm_halt_in_kernel_allowed; 113 bool kvm_eventfds_allowed; 114 bool kvm_irqfds_allowed; 115 bool kvm_resamplefds_allowed; 116 bool kvm_msi_via_irqfd_allowed; 117 bool kvm_gsi_routing_allowed; 118 bool kvm_gsi_direct_mapping; 119 bool kvm_allowed; 120 bool kvm_readonly_mem_allowed; 121 bool kvm_vm_attributes_allowed; 122 bool kvm_direct_msi_allowed; 123 bool kvm_ioeventfd_any_length_allowed; 124 bool kvm_msi_use_devid; 125 static bool kvm_immediate_exit; 126 127 static const KVMCapabilityInfo kvm_required_capabilites[] = { 128 KVM_CAP_INFO(USER_MEMORY), 129 KVM_CAP_INFO(DESTROY_MEMORY_REGION_WORKS), 130 KVM_CAP_INFO(JOIN_MEMORY_REGIONS_WORKS), 131 KVM_CAP_LAST_INFO 132 }; 133 134 int kvm_get_max_memslots(void) 135 { 136 KVMState *s = KVM_STATE(current_machine->accelerator); 137 138 return s->nr_slots; 139 } 140 141 static KVMSlot *kvm_get_free_slot(KVMMemoryListener *kml) 142 { 143 KVMState *s = kvm_state; 144 int i; 145 146 for (i = 0; i < s->nr_slots; i++) { 147 if (kml->slots[i].memory_size == 0) { 148 return &kml->slots[i]; 149 } 150 } 151 152 return NULL; 153 } 154 155 bool kvm_has_free_slot(MachineState *ms) 156 { 157 KVMState *s = KVM_STATE(ms->accelerator); 158 159 return kvm_get_free_slot(&s->memory_listener); 160 } 161 162 static KVMSlot *kvm_alloc_slot(KVMMemoryListener *kml) 163 { 164 KVMSlot *slot = kvm_get_free_slot(kml); 165 166 if (slot) { 167 return slot; 168 } 169 170 fprintf(stderr, "%s: no free slot available\n", __func__); 171 abort(); 172 } 173 174 static KVMSlot *kvm_lookup_matching_slot(KVMMemoryListener *kml, 175 hwaddr start_addr, 176 hwaddr size) 177 { 178 KVMState *s = kvm_state; 179 int i; 180 181 for (i = 0; i < s->nr_slots; i++) { 182 KVMSlot *mem = &kml->slots[i]; 183 184 if (start_addr == mem->start_addr && size == mem->memory_size) { 185 return mem; 186 } 187 } 188 189 return NULL; 190 } 191 192 /* 193 * Calculate and align the start address and the size of the section. 194 * Return the size. If the size is 0, the aligned section is empty. 195 */ 196 static hwaddr kvm_align_section(MemoryRegionSection *section, 197 hwaddr *start) 198 { 199 hwaddr size = int128_get64(section->size); 200 hwaddr delta; 201 202 *start = section->offset_within_address_space; 203 204 /* kvm works in page size chunks, but the function may be called 205 with sub-page size and unaligned start address. Pad the start 206 address to next and truncate size to previous page boundary. */ 207 delta = qemu_real_host_page_size - (*start & ~qemu_real_host_page_mask); 208 delta &= ~qemu_real_host_page_mask; 209 *start += delta; 210 if (delta > size) { 211 return 0; 212 } 213 size -= delta; 214 size &= qemu_real_host_page_mask; 215 if (*start & ~qemu_real_host_page_mask) { 216 return 0; 217 } 218 219 return size; 220 } 221 222 int kvm_physical_memory_addr_from_host(KVMState *s, void *ram, 223 hwaddr *phys_addr) 224 { 225 KVMMemoryListener *kml = &s->memory_listener; 226 int i; 227 228 for (i = 0; i < s->nr_slots; i++) { 229 KVMSlot *mem = &kml->slots[i]; 230 231 if (ram >= mem->ram && ram < mem->ram + mem->memory_size) { 232 *phys_addr = mem->start_addr + (ram - mem->ram); 233 return 1; 234 } 235 } 236 237 return 0; 238 } 239 240 static int kvm_set_user_memory_region(KVMMemoryListener *kml, KVMSlot *slot) 241 { 242 KVMState *s = kvm_state; 243 struct kvm_userspace_memory_region mem; 244 245 mem.slot = slot->slot | (kml->as_id << 16); 246 mem.guest_phys_addr = slot->start_addr; 247 mem.userspace_addr = (unsigned long)slot->ram; 248 mem.flags = slot->flags; 249 250 if (slot->memory_size && mem.flags & KVM_MEM_READONLY) { 251 /* Set the slot size to 0 before setting the slot to the desired 252 * value. This is needed based on KVM commit 75d61fbc. */ 253 mem.memory_size = 0; 254 kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem); 255 } 256 mem.memory_size = slot->memory_size; 257 return kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem); 258 } 259 260 int kvm_destroy_vcpu(CPUState *cpu) 261 { 262 KVMState *s = kvm_state; 263 long mmap_size; 264 struct KVMParkedVcpu *vcpu = NULL; 265 int ret = 0; 266 267 DPRINTF("kvm_destroy_vcpu\n"); 268 269 mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0); 270 if (mmap_size < 0) { 271 ret = mmap_size; 272 DPRINTF("KVM_GET_VCPU_MMAP_SIZE failed\n"); 273 goto err; 274 } 275 276 ret = munmap(cpu->kvm_run, mmap_size); 277 if (ret < 0) { 278 goto err; 279 } 280 281 vcpu = g_malloc0(sizeof(*vcpu)); 282 vcpu->vcpu_id = kvm_arch_vcpu_id(cpu); 283 vcpu->kvm_fd = cpu->kvm_fd; 284 QLIST_INSERT_HEAD(&kvm_state->kvm_parked_vcpus, vcpu, node); 285 err: 286 return ret; 287 } 288 289 static int kvm_get_vcpu(KVMState *s, unsigned long vcpu_id) 290 { 291 struct KVMParkedVcpu *cpu; 292 293 QLIST_FOREACH(cpu, &s->kvm_parked_vcpus, node) { 294 if (cpu->vcpu_id == vcpu_id) { 295 int kvm_fd; 296 297 QLIST_REMOVE(cpu, node); 298 kvm_fd = cpu->kvm_fd; 299 g_free(cpu); 300 return kvm_fd; 301 } 302 } 303 304 return kvm_vm_ioctl(s, KVM_CREATE_VCPU, (void *)vcpu_id); 305 } 306 307 int kvm_init_vcpu(CPUState *cpu) 308 { 309 KVMState *s = kvm_state; 310 long mmap_size; 311 int ret; 312 313 DPRINTF("kvm_init_vcpu\n"); 314 315 ret = kvm_get_vcpu(s, kvm_arch_vcpu_id(cpu)); 316 if (ret < 0) { 317 DPRINTF("kvm_create_vcpu failed\n"); 318 goto err; 319 } 320 321 cpu->kvm_fd = ret; 322 cpu->kvm_state = s; 323 cpu->vcpu_dirty = true; 324 325 mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0); 326 if (mmap_size < 0) { 327 ret = mmap_size; 328 DPRINTF("KVM_GET_VCPU_MMAP_SIZE failed\n"); 329 goto err; 330 } 331 332 cpu->kvm_run = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED, 333 cpu->kvm_fd, 0); 334 if (cpu->kvm_run == MAP_FAILED) { 335 ret = -errno; 336 DPRINTF("mmap'ing vcpu state failed\n"); 337 goto err; 338 } 339 340 if (s->coalesced_mmio && !s->coalesced_mmio_ring) { 341 s->coalesced_mmio_ring = 342 (void *)cpu->kvm_run + s->coalesced_mmio * PAGE_SIZE; 343 } 344 345 ret = kvm_arch_init_vcpu(cpu); 346 err: 347 return ret; 348 } 349 350 /* 351 * dirty pages logging control 352 */ 353 354 static int kvm_mem_flags(MemoryRegion *mr) 355 { 356 bool readonly = mr->readonly || memory_region_is_romd(mr); 357 int flags = 0; 358 359 if (memory_region_get_dirty_log_mask(mr) != 0) { 360 flags |= KVM_MEM_LOG_DIRTY_PAGES; 361 } 362 if (readonly && kvm_readonly_mem_allowed) { 363 flags |= KVM_MEM_READONLY; 364 } 365 return flags; 366 } 367 368 static int kvm_slot_update_flags(KVMMemoryListener *kml, KVMSlot *mem, 369 MemoryRegion *mr) 370 { 371 int old_flags; 372 373 old_flags = mem->flags; 374 mem->flags = kvm_mem_flags(mr); 375 376 /* If nothing changed effectively, no need to issue ioctl */ 377 if (mem->flags == old_flags) { 378 return 0; 379 } 380 381 return kvm_set_user_memory_region(kml, mem); 382 } 383 384 static int kvm_section_update_flags(KVMMemoryListener *kml, 385 MemoryRegionSection *section) 386 { 387 hwaddr start_addr, size; 388 KVMSlot *mem; 389 390 size = kvm_align_section(section, &start_addr); 391 if (!size) { 392 return 0; 393 } 394 395 mem = kvm_lookup_matching_slot(kml, start_addr, size); 396 if (!mem) { 397 fprintf(stderr, "%s: error finding slot\n", __func__); 398 abort(); 399 } 400 401 return kvm_slot_update_flags(kml, mem, section->mr); 402 } 403 404 static void kvm_log_start(MemoryListener *listener, 405 MemoryRegionSection *section, 406 int old, int new) 407 { 408 KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener); 409 int r; 410 411 if (old != 0) { 412 return; 413 } 414 415 r = kvm_section_update_flags(kml, section); 416 if (r < 0) { 417 abort(); 418 } 419 } 420 421 static void kvm_log_stop(MemoryListener *listener, 422 MemoryRegionSection *section, 423 int old, int new) 424 { 425 KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener); 426 int r; 427 428 if (new != 0) { 429 return; 430 } 431 432 r = kvm_section_update_flags(kml, section); 433 if (r < 0) { 434 abort(); 435 } 436 } 437 438 /* get kvm's dirty pages bitmap and update qemu's */ 439 static int kvm_get_dirty_pages_log_range(MemoryRegionSection *section, 440 unsigned long *bitmap) 441 { 442 ram_addr_t start = section->offset_within_region + 443 memory_region_get_ram_addr(section->mr); 444 ram_addr_t pages = int128_get64(section->size) / getpagesize(); 445 446 cpu_physical_memory_set_dirty_lebitmap(bitmap, start, pages); 447 return 0; 448 } 449 450 #define ALIGN(x, y) (((x)+(y)-1) & ~((y)-1)) 451 452 /** 453 * kvm_physical_sync_dirty_bitmap - Grab dirty bitmap from kernel space 454 * This function updates qemu's dirty bitmap using 455 * memory_region_set_dirty(). This means all bits are set 456 * to dirty. 457 * 458 * @start_add: start of logged region. 459 * @end_addr: end of logged region. 460 */ 461 static int kvm_physical_sync_dirty_bitmap(KVMMemoryListener *kml, 462 MemoryRegionSection *section) 463 { 464 KVMState *s = kvm_state; 465 struct kvm_dirty_log d = {}; 466 KVMSlot *mem; 467 hwaddr start_addr, size; 468 469 size = kvm_align_section(section, &start_addr); 470 if (size) { 471 mem = kvm_lookup_matching_slot(kml, start_addr, size); 472 if (!mem) { 473 fprintf(stderr, "%s: error finding slot\n", __func__); 474 abort(); 475 } 476 477 /* XXX bad kernel interface alert 478 * For dirty bitmap, kernel allocates array of size aligned to 479 * bits-per-long. But for case when the kernel is 64bits and 480 * the userspace is 32bits, userspace can't align to the same 481 * bits-per-long, since sizeof(long) is different between kernel 482 * and user space. This way, userspace will provide buffer which 483 * may be 4 bytes less than the kernel will use, resulting in 484 * userspace memory corruption (which is not detectable by valgrind 485 * too, in most cases). 486 * So for now, let's align to 64 instead of HOST_LONG_BITS here, in 487 * a hope that sizeof(long) won't become >8 any time soon. 488 */ 489 size = ALIGN(((mem->memory_size) >> TARGET_PAGE_BITS), 490 /*HOST_LONG_BITS*/ 64) / 8; 491 d.dirty_bitmap = g_malloc0(size); 492 493 d.slot = mem->slot | (kml->as_id << 16); 494 if (kvm_vm_ioctl(s, KVM_GET_DIRTY_LOG, &d) == -1) { 495 DPRINTF("ioctl failed %d\n", errno); 496 g_free(d.dirty_bitmap); 497 return -1; 498 } 499 500 kvm_get_dirty_pages_log_range(section, d.dirty_bitmap); 501 g_free(d.dirty_bitmap); 502 } 503 504 return 0; 505 } 506 507 static void kvm_coalesce_mmio_region(MemoryListener *listener, 508 MemoryRegionSection *secion, 509 hwaddr start, hwaddr size) 510 { 511 KVMState *s = kvm_state; 512 513 if (s->coalesced_mmio) { 514 struct kvm_coalesced_mmio_zone zone; 515 516 zone.addr = start; 517 zone.size = size; 518 zone.pad = 0; 519 520 (void)kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone); 521 } 522 } 523 524 static void kvm_uncoalesce_mmio_region(MemoryListener *listener, 525 MemoryRegionSection *secion, 526 hwaddr start, hwaddr size) 527 { 528 KVMState *s = kvm_state; 529 530 if (s->coalesced_mmio) { 531 struct kvm_coalesced_mmio_zone zone; 532 533 zone.addr = start; 534 zone.size = size; 535 zone.pad = 0; 536 537 (void)kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone); 538 } 539 } 540 541 int kvm_check_extension(KVMState *s, unsigned int extension) 542 { 543 int ret; 544 545 ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, extension); 546 if (ret < 0) { 547 ret = 0; 548 } 549 550 return ret; 551 } 552 553 int kvm_vm_check_extension(KVMState *s, unsigned int extension) 554 { 555 int ret; 556 557 ret = kvm_vm_ioctl(s, KVM_CHECK_EXTENSION, extension); 558 if (ret < 0) { 559 /* VM wide version not implemented, use global one instead */ 560 ret = kvm_check_extension(s, extension); 561 } 562 563 return ret; 564 } 565 566 static uint32_t adjust_ioeventfd_endianness(uint32_t val, uint32_t size) 567 { 568 #if defined(HOST_WORDS_BIGENDIAN) != defined(TARGET_WORDS_BIGENDIAN) 569 /* The kernel expects ioeventfd values in HOST_WORDS_BIGENDIAN 570 * endianness, but the memory core hands them in target endianness. 571 * For example, PPC is always treated as big-endian even if running 572 * on KVM and on PPC64LE. Correct here. 573 */ 574 switch (size) { 575 case 2: 576 val = bswap16(val); 577 break; 578 case 4: 579 val = bswap32(val); 580 break; 581 } 582 #endif 583 return val; 584 } 585 586 static int kvm_set_ioeventfd_mmio(int fd, hwaddr addr, uint32_t val, 587 bool assign, uint32_t size, bool datamatch) 588 { 589 int ret; 590 struct kvm_ioeventfd iofd = { 591 .datamatch = datamatch ? adjust_ioeventfd_endianness(val, size) : 0, 592 .addr = addr, 593 .len = size, 594 .flags = 0, 595 .fd = fd, 596 }; 597 598 if (!kvm_enabled()) { 599 return -ENOSYS; 600 } 601 602 if (datamatch) { 603 iofd.flags |= KVM_IOEVENTFD_FLAG_DATAMATCH; 604 } 605 if (!assign) { 606 iofd.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN; 607 } 608 609 ret = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &iofd); 610 611 if (ret < 0) { 612 return -errno; 613 } 614 615 return 0; 616 } 617 618 static int kvm_set_ioeventfd_pio(int fd, uint16_t addr, uint16_t val, 619 bool assign, uint32_t size, bool datamatch) 620 { 621 struct kvm_ioeventfd kick = { 622 .datamatch = datamatch ? adjust_ioeventfd_endianness(val, size) : 0, 623 .addr = addr, 624 .flags = KVM_IOEVENTFD_FLAG_PIO, 625 .len = size, 626 .fd = fd, 627 }; 628 int r; 629 if (!kvm_enabled()) { 630 return -ENOSYS; 631 } 632 if (datamatch) { 633 kick.flags |= KVM_IOEVENTFD_FLAG_DATAMATCH; 634 } 635 if (!assign) { 636 kick.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN; 637 } 638 r = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &kick); 639 if (r < 0) { 640 return r; 641 } 642 return 0; 643 } 644 645 646 static int kvm_check_many_ioeventfds(void) 647 { 648 /* Userspace can use ioeventfd for io notification. This requires a host 649 * that supports eventfd(2) and an I/O thread; since eventfd does not 650 * support SIGIO it cannot interrupt the vcpu. 651 * 652 * Older kernels have a 6 device limit on the KVM io bus. Find out so we 653 * can avoid creating too many ioeventfds. 654 */ 655 #if defined(CONFIG_EVENTFD) 656 int ioeventfds[7]; 657 int i, ret = 0; 658 for (i = 0; i < ARRAY_SIZE(ioeventfds); i++) { 659 ioeventfds[i] = eventfd(0, EFD_CLOEXEC); 660 if (ioeventfds[i] < 0) { 661 break; 662 } 663 ret = kvm_set_ioeventfd_pio(ioeventfds[i], 0, i, true, 2, true); 664 if (ret < 0) { 665 close(ioeventfds[i]); 666 break; 667 } 668 } 669 670 /* Decide whether many devices are supported or not */ 671 ret = i == ARRAY_SIZE(ioeventfds); 672 673 while (i-- > 0) { 674 kvm_set_ioeventfd_pio(ioeventfds[i], 0, i, false, 2, true); 675 close(ioeventfds[i]); 676 } 677 return ret; 678 #else 679 return 0; 680 #endif 681 } 682 683 static const KVMCapabilityInfo * 684 kvm_check_extension_list(KVMState *s, const KVMCapabilityInfo *list) 685 { 686 while (list->name) { 687 if (!kvm_check_extension(s, list->value)) { 688 return list; 689 } 690 list++; 691 } 692 return NULL; 693 } 694 695 static void kvm_set_phys_mem(KVMMemoryListener *kml, 696 MemoryRegionSection *section, bool add) 697 { 698 KVMSlot *mem; 699 int err; 700 MemoryRegion *mr = section->mr; 701 bool writeable = !mr->readonly && !mr->rom_device; 702 hwaddr start_addr, size; 703 void *ram; 704 705 if (!memory_region_is_ram(mr)) { 706 if (writeable || !kvm_readonly_mem_allowed) { 707 return; 708 } else if (!mr->romd_mode) { 709 /* If the memory device is not in romd_mode, then we actually want 710 * to remove the kvm memory slot so all accesses will trap. */ 711 add = false; 712 } 713 } 714 715 size = kvm_align_section(section, &start_addr); 716 if (!size) { 717 return; 718 } 719 720 ram = memory_region_get_ram_ptr(mr) + section->offset_within_region + 721 (section->offset_within_address_space - start_addr); 722 723 mem = kvm_lookup_matching_slot(kml, start_addr, size); 724 if (!add) { 725 if (!mem) { 726 return; 727 } 728 if (mem->flags & KVM_MEM_LOG_DIRTY_PAGES) { 729 kvm_physical_sync_dirty_bitmap(kml, section); 730 } 731 732 /* unregister the slot */ 733 mem->memory_size = 0; 734 err = kvm_set_user_memory_region(kml, mem); 735 if (err) { 736 fprintf(stderr, "%s: error unregistering overlapping slot: %s\n", 737 __func__, strerror(-err)); 738 abort(); 739 } 740 return; 741 } 742 743 if (mem) { 744 /* update the slot */ 745 kvm_slot_update_flags(kml, mem, mr); 746 return; 747 } 748 749 /* register the new slot */ 750 mem = kvm_alloc_slot(kml); 751 mem->memory_size = size; 752 mem->start_addr = start_addr; 753 mem->ram = ram; 754 mem->flags = kvm_mem_flags(mr); 755 756 err = kvm_set_user_memory_region(kml, mem); 757 if (err) { 758 fprintf(stderr, "%s: error registering slot: %s\n", __func__, 759 strerror(-err)); 760 abort(); 761 } 762 } 763 764 static void kvm_region_add(MemoryListener *listener, 765 MemoryRegionSection *section) 766 { 767 KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener); 768 769 memory_region_ref(section->mr); 770 kvm_set_phys_mem(kml, section, true); 771 } 772 773 static void kvm_region_del(MemoryListener *listener, 774 MemoryRegionSection *section) 775 { 776 KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener); 777 778 kvm_set_phys_mem(kml, section, false); 779 memory_region_unref(section->mr); 780 } 781 782 static void kvm_log_sync(MemoryListener *listener, 783 MemoryRegionSection *section) 784 { 785 KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener); 786 int r; 787 788 r = kvm_physical_sync_dirty_bitmap(kml, section); 789 if (r < 0) { 790 abort(); 791 } 792 } 793 794 static void kvm_mem_ioeventfd_add(MemoryListener *listener, 795 MemoryRegionSection *section, 796 bool match_data, uint64_t data, 797 EventNotifier *e) 798 { 799 int fd = event_notifier_get_fd(e); 800 int r; 801 802 r = kvm_set_ioeventfd_mmio(fd, section->offset_within_address_space, 803 data, true, int128_get64(section->size), 804 match_data); 805 if (r < 0) { 806 fprintf(stderr, "%s: error adding ioeventfd: %s\n", 807 __func__, strerror(-r)); 808 abort(); 809 } 810 } 811 812 static void kvm_mem_ioeventfd_del(MemoryListener *listener, 813 MemoryRegionSection *section, 814 bool match_data, uint64_t data, 815 EventNotifier *e) 816 { 817 int fd = event_notifier_get_fd(e); 818 int r; 819 820 r = kvm_set_ioeventfd_mmio(fd, section->offset_within_address_space, 821 data, false, int128_get64(section->size), 822 match_data); 823 if (r < 0) { 824 abort(); 825 } 826 } 827 828 static void kvm_io_ioeventfd_add(MemoryListener *listener, 829 MemoryRegionSection *section, 830 bool match_data, uint64_t data, 831 EventNotifier *e) 832 { 833 int fd = event_notifier_get_fd(e); 834 int r; 835 836 r = kvm_set_ioeventfd_pio(fd, section->offset_within_address_space, 837 data, true, int128_get64(section->size), 838 match_data); 839 if (r < 0) { 840 fprintf(stderr, "%s: error adding ioeventfd: %s\n", 841 __func__, strerror(-r)); 842 abort(); 843 } 844 } 845 846 static void kvm_io_ioeventfd_del(MemoryListener *listener, 847 MemoryRegionSection *section, 848 bool match_data, uint64_t data, 849 EventNotifier *e) 850 851 { 852 int fd = event_notifier_get_fd(e); 853 int r; 854 855 r = kvm_set_ioeventfd_pio(fd, section->offset_within_address_space, 856 data, false, int128_get64(section->size), 857 match_data); 858 if (r < 0) { 859 abort(); 860 } 861 } 862 863 void kvm_memory_listener_register(KVMState *s, KVMMemoryListener *kml, 864 AddressSpace *as, int as_id) 865 { 866 int i; 867 868 kml->slots = g_malloc0(s->nr_slots * sizeof(KVMSlot)); 869 kml->as_id = as_id; 870 871 for (i = 0; i < s->nr_slots; i++) { 872 kml->slots[i].slot = i; 873 } 874 875 kml->listener.region_add = kvm_region_add; 876 kml->listener.region_del = kvm_region_del; 877 kml->listener.log_start = kvm_log_start; 878 kml->listener.log_stop = kvm_log_stop; 879 kml->listener.log_sync = kvm_log_sync; 880 kml->listener.priority = 10; 881 882 memory_listener_register(&kml->listener, as); 883 } 884 885 static MemoryListener kvm_io_listener = { 886 .eventfd_add = kvm_io_ioeventfd_add, 887 .eventfd_del = kvm_io_ioeventfd_del, 888 .priority = 10, 889 }; 890 891 int kvm_set_irq(KVMState *s, int irq, int level) 892 { 893 struct kvm_irq_level event; 894 int ret; 895 896 assert(kvm_async_interrupts_enabled()); 897 898 event.level = level; 899 event.irq = irq; 900 ret = kvm_vm_ioctl(s, s->irq_set_ioctl, &event); 901 if (ret < 0) { 902 perror("kvm_set_irq"); 903 abort(); 904 } 905 906 return (s->irq_set_ioctl == KVM_IRQ_LINE) ? 1 : event.status; 907 } 908 909 #ifdef KVM_CAP_IRQ_ROUTING 910 typedef struct KVMMSIRoute { 911 struct kvm_irq_routing_entry kroute; 912 QTAILQ_ENTRY(KVMMSIRoute) entry; 913 } KVMMSIRoute; 914 915 static void set_gsi(KVMState *s, unsigned int gsi) 916 { 917 set_bit(gsi, s->used_gsi_bitmap); 918 } 919 920 static void clear_gsi(KVMState *s, unsigned int gsi) 921 { 922 clear_bit(gsi, s->used_gsi_bitmap); 923 } 924 925 void kvm_init_irq_routing(KVMState *s) 926 { 927 int gsi_count, i; 928 929 gsi_count = kvm_check_extension(s, KVM_CAP_IRQ_ROUTING) - 1; 930 if (gsi_count > 0) { 931 /* Round up so we can search ints using ffs */ 932 s->used_gsi_bitmap = bitmap_new(gsi_count); 933 s->gsi_count = gsi_count; 934 } 935 936 s->irq_routes = g_malloc0(sizeof(*s->irq_routes)); 937 s->nr_allocated_irq_routes = 0; 938 939 if (!kvm_direct_msi_allowed) { 940 for (i = 0; i < KVM_MSI_HASHTAB_SIZE; i++) { 941 QTAILQ_INIT(&s->msi_hashtab[i]); 942 } 943 } 944 945 kvm_arch_init_irq_routing(s); 946 } 947 948 void kvm_irqchip_commit_routes(KVMState *s) 949 { 950 int ret; 951 952 if (kvm_gsi_direct_mapping()) { 953 return; 954 } 955 956 if (!kvm_gsi_routing_enabled()) { 957 return; 958 } 959 960 s->irq_routes->flags = 0; 961 trace_kvm_irqchip_commit_routes(); 962 ret = kvm_vm_ioctl(s, KVM_SET_GSI_ROUTING, s->irq_routes); 963 assert(ret == 0); 964 } 965 966 static void kvm_add_routing_entry(KVMState *s, 967 struct kvm_irq_routing_entry *entry) 968 { 969 struct kvm_irq_routing_entry *new; 970 int n, size; 971 972 if (s->irq_routes->nr == s->nr_allocated_irq_routes) { 973 n = s->nr_allocated_irq_routes * 2; 974 if (n < 64) { 975 n = 64; 976 } 977 size = sizeof(struct kvm_irq_routing); 978 size += n * sizeof(*new); 979 s->irq_routes = g_realloc(s->irq_routes, size); 980 s->nr_allocated_irq_routes = n; 981 } 982 n = s->irq_routes->nr++; 983 new = &s->irq_routes->entries[n]; 984 985 *new = *entry; 986 987 set_gsi(s, entry->gsi); 988 } 989 990 static int kvm_update_routing_entry(KVMState *s, 991 struct kvm_irq_routing_entry *new_entry) 992 { 993 struct kvm_irq_routing_entry *entry; 994 int n; 995 996 for (n = 0; n < s->irq_routes->nr; n++) { 997 entry = &s->irq_routes->entries[n]; 998 if (entry->gsi != new_entry->gsi) { 999 continue; 1000 } 1001 1002 if(!memcmp(entry, new_entry, sizeof *entry)) { 1003 return 0; 1004 } 1005 1006 *entry = *new_entry; 1007 1008 return 0; 1009 } 1010 1011 return -ESRCH; 1012 } 1013 1014 void kvm_irqchip_add_irq_route(KVMState *s, int irq, int irqchip, int pin) 1015 { 1016 struct kvm_irq_routing_entry e = {}; 1017 1018 assert(pin < s->gsi_count); 1019 1020 e.gsi = irq; 1021 e.type = KVM_IRQ_ROUTING_IRQCHIP; 1022 e.flags = 0; 1023 e.u.irqchip.irqchip = irqchip; 1024 e.u.irqchip.pin = pin; 1025 kvm_add_routing_entry(s, &e); 1026 } 1027 1028 void kvm_irqchip_release_virq(KVMState *s, int virq) 1029 { 1030 struct kvm_irq_routing_entry *e; 1031 int i; 1032 1033 if (kvm_gsi_direct_mapping()) { 1034 return; 1035 } 1036 1037 for (i = 0; i < s->irq_routes->nr; i++) { 1038 e = &s->irq_routes->entries[i]; 1039 if (e->gsi == virq) { 1040 s->irq_routes->nr--; 1041 *e = s->irq_routes->entries[s->irq_routes->nr]; 1042 } 1043 } 1044 clear_gsi(s, virq); 1045 kvm_arch_release_virq_post(virq); 1046 trace_kvm_irqchip_release_virq(virq); 1047 } 1048 1049 static unsigned int kvm_hash_msi(uint32_t data) 1050 { 1051 /* This is optimized for IA32 MSI layout. However, no other arch shall 1052 * repeat the mistake of not providing a direct MSI injection API. */ 1053 return data & 0xff; 1054 } 1055 1056 static void kvm_flush_dynamic_msi_routes(KVMState *s) 1057 { 1058 KVMMSIRoute *route, *next; 1059 unsigned int hash; 1060 1061 for (hash = 0; hash < KVM_MSI_HASHTAB_SIZE; hash++) { 1062 QTAILQ_FOREACH_SAFE(route, &s->msi_hashtab[hash], entry, next) { 1063 kvm_irqchip_release_virq(s, route->kroute.gsi); 1064 QTAILQ_REMOVE(&s->msi_hashtab[hash], route, entry); 1065 g_free(route); 1066 } 1067 } 1068 } 1069 1070 static int kvm_irqchip_get_virq(KVMState *s) 1071 { 1072 int next_virq; 1073 1074 /* 1075 * PIC and IOAPIC share the first 16 GSI numbers, thus the available 1076 * GSI numbers are more than the number of IRQ route. Allocating a GSI 1077 * number can succeed even though a new route entry cannot be added. 1078 * When this happens, flush dynamic MSI entries to free IRQ route entries. 1079 */ 1080 if (!kvm_direct_msi_allowed && s->irq_routes->nr == s->gsi_count) { 1081 kvm_flush_dynamic_msi_routes(s); 1082 } 1083 1084 /* Return the lowest unused GSI in the bitmap */ 1085 next_virq = find_first_zero_bit(s->used_gsi_bitmap, s->gsi_count); 1086 if (next_virq >= s->gsi_count) { 1087 return -ENOSPC; 1088 } else { 1089 return next_virq; 1090 } 1091 } 1092 1093 static KVMMSIRoute *kvm_lookup_msi_route(KVMState *s, MSIMessage msg) 1094 { 1095 unsigned int hash = kvm_hash_msi(msg.data); 1096 KVMMSIRoute *route; 1097 1098 QTAILQ_FOREACH(route, &s->msi_hashtab[hash], entry) { 1099 if (route->kroute.u.msi.address_lo == (uint32_t)msg.address && 1100 route->kroute.u.msi.address_hi == (msg.address >> 32) && 1101 route->kroute.u.msi.data == le32_to_cpu(msg.data)) { 1102 return route; 1103 } 1104 } 1105 return NULL; 1106 } 1107 1108 int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg) 1109 { 1110 struct kvm_msi msi; 1111 KVMMSIRoute *route; 1112 1113 if (kvm_direct_msi_allowed) { 1114 msi.address_lo = (uint32_t)msg.address; 1115 msi.address_hi = msg.address >> 32; 1116 msi.data = le32_to_cpu(msg.data); 1117 msi.flags = 0; 1118 memset(msi.pad, 0, sizeof(msi.pad)); 1119 1120 return kvm_vm_ioctl(s, KVM_SIGNAL_MSI, &msi); 1121 } 1122 1123 route = kvm_lookup_msi_route(s, msg); 1124 if (!route) { 1125 int virq; 1126 1127 virq = kvm_irqchip_get_virq(s); 1128 if (virq < 0) { 1129 return virq; 1130 } 1131 1132 route = g_malloc0(sizeof(KVMMSIRoute)); 1133 route->kroute.gsi = virq; 1134 route->kroute.type = KVM_IRQ_ROUTING_MSI; 1135 route->kroute.flags = 0; 1136 route->kroute.u.msi.address_lo = (uint32_t)msg.address; 1137 route->kroute.u.msi.address_hi = msg.address >> 32; 1138 route->kroute.u.msi.data = le32_to_cpu(msg.data); 1139 1140 kvm_add_routing_entry(s, &route->kroute); 1141 kvm_irqchip_commit_routes(s); 1142 1143 QTAILQ_INSERT_TAIL(&s->msi_hashtab[kvm_hash_msi(msg.data)], route, 1144 entry); 1145 } 1146 1147 assert(route->kroute.type == KVM_IRQ_ROUTING_MSI); 1148 1149 return kvm_set_irq(s, route->kroute.gsi, 1); 1150 } 1151 1152 int kvm_irqchip_add_msi_route(KVMState *s, int vector, PCIDevice *dev) 1153 { 1154 struct kvm_irq_routing_entry kroute = {}; 1155 int virq; 1156 MSIMessage msg = {0, 0}; 1157 1158 if (pci_available && dev) { 1159 msg = pci_get_msi_message(dev, vector); 1160 } 1161 1162 if (kvm_gsi_direct_mapping()) { 1163 return kvm_arch_msi_data_to_gsi(msg.data); 1164 } 1165 1166 if (!kvm_gsi_routing_enabled()) { 1167 return -ENOSYS; 1168 } 1169 1170 virq = kvm_irqchip_get_virq(s); 1171 if (virq < 0) { 1172 return virq; 1173 } 1174 1175 kroute.gsi = virq; 1176 kroute.type = KVM_IRQ_ROUTING_MSI; 1177 kroute.flags = 0; 1178 kroute.u.msi.address_lo = (uint32_t)msg.address; 1179 kroute.u.msi.address_hi = msg.address >> 32; 1180 kroute.u.msi.data = le32_to_cpu(msg.data); 1181 if (pci_available && kvm_msi_devid_required()) { 1182 kroute.flags = KVM_MSI_VALID_DEVID; 1183 kroute.u.msi.devid = pci_requester_id(dev); 1184 } 1185 if (kvm_arch_fixup_msi_route(&kroute, msg.address, msg.data, dev)) { 1186 kvm_irqchip_release_virq(s, virq); 1187 return -EINVAL; 1188 } 1189 1190 trace_kvm_irqchip_add_msi_route(dev ? dev->name : (char *)"N/A", 1191 vector, virq); 1192 1193 kvm_add_routing_entry(s, &kroute); 1194 kvm_arch_add_msi_route_post(&kroute, vector, dev); 1195 kvm_irqchip_commit_routes(s); 1196 1197 return virq; 1198 } 1199 1200 int kvm_irqchip_update_msi_route(KVMState *s, int virq, MSIMessage msg, 1201 PCIDevice *dev) 1202 { 1203 struct kvm_irq_routing_entry kroute = {}; 1204 1205 if (kvm_gsi_direct_mapping()) { 1206 return 0; 1207 } 1208 1209 if (!kvm_irqchip_in_kernel()) { 1210 return -ENOSYS; 1211 } 1212 1213 kroute.gsi = virq; 1214 kroute.type = KVM_IRQ_ROUTING_MSI; 1215 kroute.flags = 0; 1216 kroute.u.msi.address_lo = (uint32_t)msg.address; 1217 kroute.u.msi.address_hi = msg.address >> 32; 1218 kroute.u.msi.data = le32_to_cpu(msg.data); 1219 if (pci_available && kvm_msi_devid_required()) { 1220 kroute.flags = KVM_MSI_VALID_DEVID; 1221 kroute.u.msi.devid = pci_requester_id(dev); 1222 } 1223 if (kvm_arch_fixup_msi_route(&kroute, msg.address, msg.data, dev)) { 1224 return -EINVAL; 1225 } 1226 1227 trace_kvm_irqchip_update_msi_route(virq); 1228 1229 return kvm_update_routing_entry(s, &kroute); 1230 } 1231 1232 static int kvm_irqchip_assign_irqfd(KVMState *s, int fd, int rfd, int virq, 1233 bool assign) 1234 { 1235 struct kvm_irqfd irqfd = { 1236 .fd = fd, 1237 .gsi = virq, 1238 .flags = assign ? 0 : KVM_IRQFD_FLAG_DEASSIGN, 1239 }; 1240 1241 if (rfd != -1) { 1242 irqfd.flags |= KVM_IRQFD_FLAG_RESAMPLE; 1243 irqfd.resamplefd = rfd; 1244 } 1245 1246 if (!kvm_irqfds_enabled()) { 1247 return -ENOSYS; 1248 } 1249 1250 return kvm_vm_ioctl(s, KVM_IRQFD, &irqfd); 1251 } 1252 1253 int kvm_irqchip_add_adapter_route(KVMState *s, AdapterInfo *adapter) 1254 { 1255 struct kvm_irq_routing_entry kroute = {}; 1256 int virq; 1257 1258 if (!kvm_gsi_routing_enabled()) { 1259 return -ENOSYS; 1260 } 1261 1262 virq = kvm_irqchip_get_virq(s); 1263 if (virq < 0) { 1264 return virq; 1265 } 1266 1267 kroute.gsi = virq; 1268 kroute.type = KVM_IRQ_ROUTING_S390_ADAPTER; 1269 kroute.flags = 0; 1270 kroute.u.adapter.summary_addr = adapter->summary_addr; 1271 kroute.u.adapter.ind_addr = adapter->ind_addr; 1272 kroute.u.adapter.summary_offset = adapter->summary_offset; 1273 kroute.u.adapter.ind_offset = adapter->ind_offset; 1274 kroute.u.adapter.adapter_id = adapter->adapter_id; 1275 1276 kvm_add_routing_entry(s, &kroute); 1277 1278 return virq; 1279 } 1280 1281 int kvm_irqchip_add_hv_sint_route(KVMState *s, uint32_t vcpu, uint32_t sint) 1282 { 1283 struct kvm_irq_routing_entry kroute = {}; 1284 int virq; 1285 1286 if (!kvm_gsi_routing_enabled()) { 1287 return -ENOSYS; 1288 } 1289 if (!kvm_check_extension(s, KVM_CAP_HYPERV_SYNIC)) { 1290 return -ENOSYS; 1291 } 1292 virq = kvm_irqchip_get_virq(s); 1293 if (virq < 0) { 1294 return virq; 1295 } 1296 1297 kroute.gsi = virq; 1298 kroute.type = KVM_IRQ_ROUTING_HV_SINT; 1299 kroute.flags = 0; 1300 kroute.u.hv_sint.vcpu = vcpu; 1301 kroute.u.hv_sint.sint = sint; 1302 1303 kvm_add_routing_entry(s, &kroute); 1304 kvm_irqchip_commit_routes(s); 1305 1306 return virq; 1307 } 1308 1309 #else /* !KVM_CAP_IRQ_ROUTING */ 1310 1311 void kvm_init_irq_routing(KVMState *s) 1312 { 1313 } 1314 1315 void kvm_irqchip_release_virq(KVMState *s, int virq) 1316 { 1317 } 1318 1319 int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg) 1320 { 1321 abort(); 1322 } 1323 1324 int kvm_irqchip_add_msi_route(KVMState *s, int vector, PCIDevice *dev) 1325 { 1326 return -ENOSYS; 1327 } 1328 1329 int kvm_irqchip_add_adapter_route(KVMState *s, AdapterInfo *adapter) 1330 { 1331 return -ENOSYS; 1332 } 1333 1334 int kvm_irqchip_add_hv_sint_route(KVMState *s, uint32_t vcpu, uint32_t sint) 1335 { 1336 return -ENOSYS; 1337 } 1338 1339 static int kvm_irqchip_assign_irqfd(KVMState *s, int fd, int virq, bool assign) 1340 { 1341 abort(); 1342 } 1343 1344 int kvm_irqchip_update_msi_route(KVMState *s, int virq, MSIMessage msg) 1345 { 1346 return -ENOSYS; 1347 } 1348 #endif /* !KVM_CAP_IRQ_ROUTING */ 1349 1350 int kvm_irqchip_add_irqfd_notifier_gsi(KVMState *s, EventNotifier *n, 1351 EventNotifier *rn, int virq) 1352 { 1353 return kvm_irqchip_assign_irqfd(s, event_notifier_get_fd(n), 1354 rn ? event_notifier_get_fd(rn) : -1, virq, true); 1355 } 1356 1357 int kvm_irqchip_remove_irqfd_notifier_gsi(KVMState *s, EventNotifier *n, 1358 int virq) 1359 { 1360 return kvm_irqchip_assign_irqfd(s, event_notifier_get_fd(n), -1, virq, 1361 false); 1362 } 1363 1364 int kvm_irqchip_add_irqfd_notifier(KVMState *s, EventNotifier *n, 1365 EventNotifier *rn, qemu_irq irq) 1366 { 1367 gpointer key, gsi; 1368 gboolean found = g_hash_table_lookup_extended(s->gsimap, irq, &key, &gsi); 1369 1370 if (!found) { 1371 return -ENXIO; 1372 } 1373 return kvm_irqchip_add_irqfd_notifier_gsi(s, n, rn, GPOINTER_TO_INT(gsi)); 1374 } 1375 1376 int kvm_irqchip_remove_irqfd_notifier(KVMState *s, EventNotifier *n, 1377 qemu_irq irq) 1378 { 1379 gpointer key, gsi; 1380 gboolean found = g_hash_table_lookup_extended(s->gsimap, irq, &key, &gsi); 1381 1382 if (!found) { 1383 return -ENXIO; 1384 } 1385 return kvm_irqchip_remove_irqfd_notifier_gsi(s, n, GPOINTER_TO_INT(gsi)); 1386 } 1387 1388 void kvm_irqchip_set_qemuirq_gsi(KVMState *s, qemu_irq irq, int gsi) 1389 { 1390 g_hash_table_insert(s->gsimap, irq, GINT_TO_POINTER(gsi)); 1391 } 1392 1393 static void kvm_irqchip_create(MachineState *machine, KVMState *s) 1394 { 1395 int ret; 1396 1397 if (kvm_check_extension(s, KVM_CAP_IRQCHIP)) { 1398 ; 1399 } else if (kvm_check_extension(s, KVM_CAP_S390_IRQCHIP)) { 1400 ret = kvm_vm_enable_cap(s, KVM_CAP_S390_IRQCHIP, 0); 1401 if (ret < 0) { 1402 fprintf(stderr, "Enable kernel irqchip failed: %s\n", strerror(-ret)); 1403 exit(1); 1404 } 1405 } else { 1406 return; 1407 } 1408 1409 /* First probe and see if there's a arch-specific hook to create the 1410 * in-kernel irqchip for us */ 1411 ret = kvm_arch_irqchip_create(machine, s); 1412 if (ret == 0) { 1413 if (machine_kernel_irqchip_split(machine)) { 1414 perror("Split IRQ chip mode not supported."); 1415 exit(1); 1416 } else { 1417 ret = kvm_vm_ioctl(s, KVM_CREATE_IRQCHIP); 1418 } 1419 } 1420 if (ret < 0) { 1421 fprintf(stderr, "Create kernel irqchip failed: %s\n", strerror(-ret)); 1422 exit(1); 1423 } 1424 1425 kvm_kernel_irqchip = true; 1426 /* If we have an in-kernel IRQ chip then we must have asynchronous 1427 * interrupt delivery (though the reverse is not necessarily true) 1428 */ 1429 kvm_async_interrupts_allowed = true; 1430 kvm_halt_in_kernel_allowed = true; 1431 1432 kvm_init_irq_routing(s); 1433 1434 s->gsimap = g_hash_table_new(g_direct_hash, g_direct_equal); 1435 } 1436 1437 /* Find number of supported CPUs using the recommended 1438 * procedure from the kernel API documentation to cope with 1439 * older kernels that may be missing capabilities. 1440 */ 1441 static int kvm_recommended_vcpus(KVMState *s) 1442 { 1443 int ret = kvm_vm_check_extension(s, KVM_CAP_NR_VCPUS); 1444 return (ret) ? ret : 4; 1445 } 1446 1447 static int kvm_max_vcpus(KVMState *s) 1448 { 1449 int ret = kvm_check_extension(s, KVM_CAP_MAX_VCPUS); 1450 return (ret) ? ret : kvm_recommended_vcpus(s); 1451 } 1452 1453 static int kvm_max_vcpu_id(KVMState *s) 1454 { 1455 int ret = kvm_check_extension(s, KVM_CAP_MAX_VCPU_ID); 1456 return (ret) ? ret : kvm_max_vcpus(s); 1457 } 1458 1459 bool kvm_vcpu_id_is_valid(int vcpu_id) 1460 { 1461 KVMState *s = KVM_STATE(current_machine->accelerator); 1462 return vcpu_id >= 0 && vcpu_id < kvm_max_vcpu_id(s); 1463 } 1464 1465 static int kvm_init(MachineState *ms) 1466 { 1467 MachineClass *mc = MACHINE_GET_CLASS(ms); 1468 static const char upgrade_note[] = 1469 "Please upgrade to at least kernel 2.6.29 or recent kvm-kmod\n" 1470 "(see http://sourceforge.net/projects/kvm).\n"; 1471 struct { 1472 const char *name; 1473 int num; 1474 } num_cpus[] = { 1475 { "SMP", smp_cpus }, 1476 { "hotpluggable", max_cpus }, 1477 { NULL, } 1478 }, *nc = num_cpus; 1479 int soft_vcpus_limit, hard_vcpus_limit; 1480 KVMState *s; 1481 const KVMCapabilityInfo *missing_cap; 1482 int ret; 1483 int type = 0; 1484 const char *kvm_type; 1485 1486 s = KVM_STATE(ms->accelerator); 1487 1488 /* 1489 * On systems where the kernel can support different base page 1490 * sizes, host page size may be different from TARGET_PAGE_SIZE, 1491 * even with KVM. TARGET_PAGE_SIZE is assumed to be the minimum 1492 * page size for the system though. 1493 */ 1494 assert(TARGET_PAGE_SIZE <= getpagesize()); 1495 1496 s->sigmask_len = 8; 1497 1498 #ifdef KVM_CAP_SET_GUEST_DEBUG 1499 QTAILQ_INIT(&s->kvm_sw_breakpoints); 1500 #endif 1501 QLIST_INIT(&s->kvm_parked_vcpus); 1502 s->vmfd = -1; 1503 s->fd = qemu_open("/dev/kvm", O_RDWR); 1504 if (s->fd == -1) { 1505 fprintf(stderr, "Could not access KVM kernel module: %m\n"); 1506 ret = -errno; 1507 goto err; 1508 } 1509 1510 ret = kvm_ioctl(s, KVM_GET_API_VERSION, 0); 1511 if (ret < KVM_API_VERSION) { 1512 if (ret >= 0) { 1513 ret = -EINVAL; 1514 } 1515 fprintf(stderr, "kvm version too old\n"); 1516 goto err; 1517 } 1518 1519 if (ret > KVM_API_VERSION) { 1520 ret = -EINVAL; 1521 fprintf(stderr, "kvm version not supported\n"); 1522 goto err; 1523 } 1524 1525 kvm_immediate_exit = kvm_check_extension(s, KVM_CAP_IMMEDIATE_EXIT); 1526 s->nr_slots = kvm_check_extension(s, KVM_CAP_NR_MEMSLOTS); 1527 1528 /* If unspecified, use the default value */ 1529 if (!s->nr_slots) { 1530 s->nr_slots = 32; 1531 } 1532 1533 kvm_type = qemu_opt_get(qemu_get_machine_opts(), "kvm-type"); 1534 if (mc->kvm_type) { 1535 type = mc->kvm_type(kvm_type); 1536 } else if (kvm_type) { 1537 ret = -EINVAL; 1538 fprintf(stderr, "Invalid argument kvm-type=%s\n", kvm_type); 1539 goto err; 1540 } 1541 1542 do { 1543 ret = kvm_ioctl(s, KVM_CREATE_VM, type); 1544 } while (ret == -EINTR); 1545 1546 if (ret < 0) { 1547 fprintf(stderr, "ioctl(KVM_CREATE_VM) failed: %d %s\n", -ret, 1548 strerror(-ret)); 1549 1550 #ifdef TARGET_S390X 1551 if (ret == -EINVAL) { 1552 fprintf(stderr, 1553 "Host kernel setup problem detected. Please verify:\n"); 1554 fprintf(stderr, "- for kernels supporting the switch_amode or" 1555 " user_mode parameters, whether\n"); 1556 fprintf(stderr, 1557 " user space is running in primary address space\n"); 1558 fprintf(stderr, 1559 "- for kernels supporting the vm.allocate_pgste sysctl, " 1560 "whether it is enabled\n"); 1561 } 1562 #endif 1563 goto err; 1564 } 1565 1566 s->vmfd = ret; 1567 1568 /* check the vcpu limits */ 1569 soft_vcpus_limit = kvm_recommended_vcpus(s); 1570 hard_vcpus_limit = kvm_max_vcpus(s); 1571 1572 while (nc->name) { 1573 if (nc->num > soft_vcpus_limit) { 1574 warn_report("Number of %s cpus requested (%d) exceeds " 1575 "the recommended cpus supported by KVM (%d)", 1576 nc->name, nc->num, soft_vcpus_limit); 1577 1578 if (nc->num > hard_vcpus_limit) { 1579 fprintf(stderr, "Number of %s cpus requested (%d) exceeds " 1580 "the maximum cpus supported by KVM (%d)\n", 1581 nc->name, nc->num, hard_vcpus_limit); 1582 exit(1); 1583 } 1584 } 1585 nc++; 1586 } 1587 1588 missing_cap = kvm_check_extension_list(s, kvm_required_capabilites); 1589 if (!missing_cap) { 1590 missing_cap = 1591 kvm_check_extension_list(s, kvm_arch_required_capabilities); 1592 } 1593 if (missing_cap) { 1594 ret = -EINVAL; 1595 fprintf(stderr, "kvm does not support %s\n%s", 1596 missing_cap->name, upgrade_note); 1597 goto err; 1598 } 1599 1600 s->coalesced_mmio = kvm_check_extension(s, KVM_CAP_COALESCED_MMIO); 1601 1602 #ifdef KVM_CAP_VCPU_EVENTS 1603 s->vcpu_events = kvm_check_extension(s, KVM_CAP_VCPU_EVENTS); 1604 #endif 1605 1606 s->robust_singlestep = 1607 kvm_check_extension(s, KVM_CAP_X86_ROBUST_SINGLESTEP); 1608 1609 #ifdef KVM_CAP_DEBUGREGS 1610 s->debugregs = kvm_check_extension(s, KVM_CAP_DEBUGREGS); 1611 #endif 1612 1613 #ifdef KVM_CAP_IRQ_ROUTING 1614 kvm_direct_msi_allowed = (kvm_check_extension(s, KVM_CAP_SIGNAL_MSI) > 0); 1615 #endif 1616 1617 s->intx_set_mask = kvm_check_extension(s, KVM_CAP_PCI_2_3); 1618 1619 s->irq_set_ioctl = KVM_IRQ_LINE; 1620 if (kvm_check_extension(s, KVM_CAP_IRQ_INJECT_STATUS)) { 1621 s->irq_set_ioctl = KVM_IRQ_LINE_STATUS; 1622 } 1623 1624 #ifdef KVM_CAP_READONLY_MEM 1625 kvm_readonly_mem_allowed = 1626 (kvm_check_extension(s, KVM_CAP_READONLY_MEM) > 0); 1627 #endif 1628 1629 kvm_eventfds_allowed = 1630 (kvm_check_extension(s, KVM_CAP_IOEVENTFD) > 0); 1631 1632 kvm_irqfds_allowed = 1633 (kvm_check_extension(s, KVM_CAP_IRQFD) > 0); 1634 1635 kvm_resamplefds_allowed = 1636 (kvm_check_extension(s, KVM_CAP_IRQFD_RESAMPLE) > 0); 1637 1638 kvm_vm_attributes_allowed = 1639 (kvm_check_extension(s, KVM_CAP_VM_ATTRIBUTES) > 0); 1640 1641 kvm_ioeventfd_any_length_allowed = 1642 (kvm_check_extension(s, KVM_CAP_IOEVENTFD_ANY_LENGTH) > 0); 1643 1644 kvm_state = s; 1645 1646 ret = kvm_arch_init(ms, s); 1647 if (ret < 0) { 1648 goto err; 1649 } 1650 1651 if (machine_kernel_irqchip_allowed(ms)) { 1652 kvm_irqchip_create(ms, s); 1653 } 1654 1655 if (kvm_eventfds_allowed) { 1656 s->memory_listener.listener.eventfd_add = kvm_mem_ioeventfd_add; 1657 s->memory_listener.listener.eventfd_del = kvm_mem_ioeventfd_del; 1658 } 1659 s->memory_listener.listener.coalesced_mmio_add = kvm_coalesce_mmio_region; 1660 s->memory_listener.listener.coalesced_mmio_del = kvm_uncoalesce_mmio_region; 1661 1662 kvm_memory_listener_register(s, &s->memory_listener, 1663 &address_space_memory, 0); 1664 memory_listener_register(&kvm_io_listener, 1665 &address_space_io); 1666 1667 s->many_ioeventfds = kvm_check_many_ioeventfds(); 1668 1669 s->sync_mmu = !!kvm_vm_check_extension(kvm_state, KVM_CAP_SYNC_MMU); 1670 1671 return 0; 1672 1673 err: 1674 assert(ret < 0); 1675 if (s->vmfd >= 0) { 1676 close(s->vmfd); 1677 } 1678 if (s->fd != -1) { 1679 close(s->fd); 1680 } 1681 g_free(s->memory_listener.slots); 1682 1683 return ret; 1684 } 1685 1686 void kvm_set_sigmask_len(KVMState *s, unsigned int sigmask_len) 1687 { 1688 s->sigmask_len = sigmask_len; 1689 } 1690 1691 static void kvm_handle_io(uint16_t port, MemTxAttrs attrs, void *data, int direction, 1692 int size, uint32_t count) 1693 { 1694 int i; 1695 uint8_t *ptr = data; 1696 1697 for (i = 0; i < count; i++) { 1698 address_space_rw(&address_space_io, port, attrs, 1699 ptr, size, 1700 direction == KVM_EXIT_IO_OUT); 1701 ptr += size; 1702 } 1703 } 1704 1705 static int kvm_handle_internal_error(CPUState *cpu, struct kvm_run *run) 1706 { 1707 fprintf(stderr, "KVM internal error. Suberror: %d\n", 1708 run->internal.suberror); 1709 1710 if (kvm_check_extension(kvm_state, KVM_CAP_INTERNAL_ERROR_DATA)) { 1711 int i; 1712 1713 for (i = 0; i < run->internal.ndata; ++i) { 1714 fprintf(stderr, "extra data[%d]: %"PRIx64"\n", 1715 i, (uint64_t)run->internal.data[i]); 1716 } 1717 } 1718 if (run->internal.suberror == KVM_INTERNAL_ERROR_EMULATION) { 1719 fprintf(stderr, "emulation failure\n"); 1720 if (!kvm_arch_stop_on_emulation_error(cpu)) { 1721 cpu_dump_state(cpu, stderr, fprintf, CPU_DUMP_CODE); 1722 return EXCP_INTERRUPT; 1723 } 1724 } 1725 /* FIXME: Should trigger a qmp message to let management know 1726 * something went wrong. 1727 */ 1728 return -1; 1729 } 1730 1731 void kvm_flush_coalesced_mmio_buffer(void) 1732 { 1733 KVMState *s = kvm_state; 1734 1735 if (s->coalesced_flush_in_progress) { 1736 return; 1737 } 1738 1739 s->coalesced_flush_in_progress = true; 1740 1741 if (s->coalesced_mmio_ring) { 1742 struct kvm_coalesced_mmio_ring *ring = s->coalesced_mmio_ring; 1743 while (ring->first != ring->last) { 1744 struct kvm_coalesced_mmio *ent; 1745 1746 ent = &ring->coalesced_mmio[ring->first]; 1747 1748 cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len); 1749 smp_wmb(); 1750 ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX; 1751 } 1752 } 1753 1754 s->coalesced_flush_in_progress = false; 1755 } 1756 1757 static void do_kvm_cpu_synchronize_state(CPUState *cpu, run_on_cpu_data arg) 1758 { 1759 if (!cpu->vcpu_dirty) { 1760 kvm_arch_get_registers(cpu); 1761 cpu->vcpu_dirty = true; 1762 } 1763 } 1764 1765 void kvm_cpu_synchronize_state(CPUState *cpu) 1766 { 1767 if (!cpu->vcpu_dirty) { 1768 run_on_cpu(cpu, do_kvm_cpu_synchronize_state, RUN_ON_CPU_NULL); 1769 } 1770 } 1771 1772 static void do_kvm_cpu_synchronize_post_reset(CPUState *cpu, run_on_cpu_data arg) 1773 { 1774 kvm_arch_put_registers(cpu, KVM_PUT_RESET_STATE); 1775 cpu->vcpu_dirty = false; 1776 } 1777 1778 void kvm_cpu_synchronize_post_reset(CPUState *cpu) 1779 { 1780 run_on_cpu(cpu, do_kvm_cpu_synchronize_post_reset, RUN_ON_CPU_NULL); 1781 } 1782 1783 static void do_kvm_cpu_synchronize_post_init(CPUState *cpu, run_on_cpu_data arg) 1784 { 1785 kvm_arch_put_registers(cpu, KVM_PUT_FULL_STATE); 1786 cpu->vcpu_dirty = false; 1787 } 1788 1789 void kvm_cpu_synchronize_post_init(CPUState *cpu) 1790 { 1791 run_on_cpu(cpu, do_kvm_cpu_synchronize_post_init, RUN_ON_CPU_NULL); 1792 } 1793 1794 static void do_kvm_cpu_synchronize_pre_loadvm(CPUState *cpu, run_on_cpu_data arg) 1795 { 1796 cpu->vcpu_dirty = true; 1797 } 1798 1799 void kvm_cpu_synchronize_pre_loadvm(CPUState *cpu) 1800 { 1801 run_on_cpu(cpu, do_kvm_cpu_synchronize_pre_loadvm, RUN_ON_CPU_NULL); 1802 } 1803 1804 #ifdef KVM_HAVE_MCE_INJECTION 1805 static __thread void *pending_sigbus_addr; 1806 static __thread int pending_sigbus_code; 1807 static __thread bool have_sigbus_pending; 1808 #endif 1809 1810 static void kvm_cpu_kick(CPUState *cpu) 1811 { 1812 atomic_set(&cpu->kvm_run->immediate_exit, 1); 1813 } 1814 1815 static void kvm_cpu_kick_self(void) 1816 { 1817 if (kvm_immediate_exit) { 1818 kvm_cpu_kick(current_cpu); 1819 } else { 1820 qemu_cpu_kick_self(); 1821 } 1822 } 1823 1824 static void kvm_eat_signals(CPUState *cpu) 1825 { 1826 struct timespec ts = { 0, 0 }; 1827 siginfo_t siginfo; 1828 sigset_t waitset; 1829 sigset_t chkset; 1830 int r; 1831 1832 if (kvm_immediate_exit) { 1833 atomic_set(&cpu->kvm_run->immediate_exit, 0); 1834 /* Write kvm_run->immediate_exit before the cpu->exit_request 1835 * write in kvm_cpu_exec. 1836 */ 1837 smp_wmb(); 1838 return; 1839 } 1840 1841 sigemptyset(&waitset); 1842 sigaddset(&waitset, SIG_IPI); 1843 1844 do { 1845 r = sigtimedwait(&waitset, &siginfo, &ts); 1846 if (r == -1 && !(errno == EAGAIN || errno == EINTR)) { 1847 perror("sigtimedwait"); 1848 exit(1); 1849 } 1850 1851 r = sigpending(&chkset); 1852 if (r == -1) { 1853 perror("sigpending"); 1854 exit(1); 1855 } 1856 } while (sigismember(&chkset, SIG_IPI)); 1857 } 1858 1859 int kvm_cpu_exec(CPUState *cpu) 1860 { 1861 struct kvm_run *run = cpu->kvm_run; 1862 int ret, run_ret; 1863 1864 DPRINTF("kvm_cpu_exec()\n"); 1865 1866 if (kvm_arch_process_async_events(cpu)) { 1867 atomic_set(&cpu->exit_request, 0); 1868 return EXCP_HLT; 1869 } 1870 1871 qemu_mutex_unlock_iothread(); 1872 cpu_exec_start(cpu); 1873 1874 do { 1875 MemTxAttrs attrs; 1876 1877 if (cpu->vcpu_dirty) { 1878 kvm_arch_put_registers(cpu, KVM_PUT_RUNTIME_STATE); 1879 cpu->vcpu_dirty = false; 1880 } 1881 1882 kvm_arch_pre_run(cpu, run); 1883 if (atomic_read(&cpu->exit_request)) { 1884 DPRINTF("interrupt exit requested\n"); 1885 /* 1886 * KVM requires us to reenter the kernel after IO exits to complete 1887 * instruction emulation. This self-signal will ensure that we 1888 * leave ASAP again. 1889 */ 1890 kvm_cpu_kick_self(); 1891 } 1892 1893 /* Read cpu->exit_request before KVM_RUN reads run->immediate_exit. 1894 * Matching barrier in kvm_eat_signals. 1895 */ 1896 smp_rmb(); 1897 1898 run_ret = kvm_vcpu_ioctl(cpu, KVM_RUN, 0); 1899 1900 attrs = kvm_arch_post_run(cpu, run); 1901 1902 #ifdef KVM_HAVE_MCE_INJECTION 1903 if (unlikely(have_sigbus_pending)) { 1904 qemu_mutex_lock_iothread(); 1905 kvm_arch_on_sigbus_vcpu(cpu, pending_sigbus_code, 1906 pending_sigbus_addr); 1907 have_sigbus_pending = false; 1908 qemu_mutex_unlock_iothread(); 1909 } 1910 #endif 1911 1912 if (run_ret < 0) { 1913 if (run_ret == -EINTR || run_ret == -EAGAIN) { 1914 DPRINTF("io window exit\n"); 1915 kvm_eat_signals(cpu); 1916 ret = EXCP_INTERRUPT; 1917 break; 1918 } 1919 fprintf(stderr, "error: kvm run failed %s\n", 1920 strerror(-run_ret)); 1921 #ifdef TARGET_PPC 1922 if (run_ret == -EBUSY) { 1923 fprintf(stderr, 1924 "This is probably because your SMT is enabled.\n" 1925 "VCPU can only run on primary threads with all " 1926 "secondary threads offline.\n"); 1927 } 1928 #endif 1929 ret = -1; 1930 break; 1931 } 1932 1933 trace_kvm_run_exit(cpu->cpu_index, run->exit_reason); 1934 switch (run->exit_reason) { 1935 case KVM_EXIT_IO: 1936 DPRINTF("handle_io\n"); 1937 /* Called outside BQL */ 1938 kvm_handle_io(run->io.port, attrs, 1939 (uint8_t *)run + run->io.data_offset, 1940 run->io.direction, 1941 run->io.size, 1942 run->io.count); 1943 ret = 0; 1944 break; 1945 case KVM_EXIT_MMIO: 1946 DPRINTF("handle_mmio\n"); 1947 /* Called outside BQL */ 1948 address_space_rw(&address_space_memory, 1949 run->mmio.phys_addr, attrs, 1950 run->mmio.data, 1951 run->mmio.len, 1952 run->mmio.is_write); 1953 ret = 0; 1954 break; 1955 case KVM_EXIT_IRQ_WINDOW_OPEN: 1956 DPRINTF("irq_window_open\n"); 1957 ret = EXCP_INTERRUPT; 1958 break; 1959 case KVM_EXIT_SHUTDOWN: 1960 DPRINTF("shutdown\n"); 1961 qemu_system_reset_request(SHUTDOWN_CAUSE_GUEST_RESET); 1962 ret = EXCP_INTERRUPT; 1963 break; 1964 case KVM_EXIT_UNKNOWN: 1965 fprintf(stderr, "KVM: unknown exit, hardware reason %" PRIx64 "\n", 1966 (uint64_t)run->hw.hardware_exit_reason); 1967 ret = -1; 1968 break; 1969 case KVM_EXIT_INTERNAL_ERROR: 1970 ret = kvm_handle_internal_error(cpu, run); 1971 break; 1972 case KVM_EXIT_SYSTEM_EVENT: 1973 switch (run->system_event.type) { 1974 case KVM_SYSTEM_EVENT_SHUTDOWN: 1975 qemu_system_shutdown_request(SHUTDOWN_CAUSE_GUEST_SHUTDOWN); 1976 ret = EXCP_INTERRUPT; 1977 break; 1978 case KVM_SYSTEM_EVENT_RESET: 1979 qemu_system_reset_request(SHUTDOWN_CAUSE_GUEST_RESET); 1980 ret = EXCP_INTERRUPT; 1981 break; 1982 case KVM_SYSTEM_EVENT_CRASH: 1983 kvm_cpu_synchronize_state(cpu); 1984 qemu_mutex_lock_iothread(); 1985 qemu_system_guest_panicked(cpu_get_crash_info(cpu)); 1986 qemu_mutex_unlock_iothread(); 1987 ret = 0; 1988 break; 1989 default: 1990 DPRINTF("kvm_arch_handle_exit\n"); 1991 ret = kvm_arch_handle_exit(cpu, run); 1992 break; 1993 } 1994 break; 1995 default: 1996 DPRINTF("kvm_arch_handle_exit\n"); 1997 ret = kvm_arch_handle_exit(cpu, run); 1998 break; 1999 } 2000 } while (ret == 0); 2001 2002 cpu_exec_end(cpu); 2003 qemu_mutex_lock_iothread(); 2004 2005 if (ret < 0) { 2006 cpu_dump_state(cpu, stderr, fprintf, CPU_DUMP_CODE); 2007 vm_stop(RUN_STATE_INTERNAL_ERROR); 2008 } 2009 2010 atomic_set(&cpu->exit_request, 0); 2011 return ret; 2012 } 2013 2014 int kvm_ioctl(KVMState *s, int type, ...) 2015 { 2016 int ret; 2017 void *arg; 2018 va_list ap; 2019 2020 va_start(ap, type); 2021 arg = va_arg(ap, void *); 2022 va_end(ap); 2023 2024 trace_kvm_ioctl(type, arg); 2025 ret = ioctl(s->fd, type, arg); 2026 if (ret == -1) { 2027 ret = -errno; 2028 } 2029 return ret; 2030 } 2031 2032 int kvm_vm_ioctl(KVMState *s, int type, ...) 2033 { 2034 int ret; 2035 void *arg; 2036 va_list ap; 2037 2038 va_start(ap, type); 2039 arg = va_arg(ap, void *); 2040 va_end(ap); 2041 2042 trace_kvm_vm_ioctl(type, arg); 2043 ret = ioctl(s->vmfd, type, arg); 2044 if (ret == -1) { 2045 ret = -errno; 2046 } 2047 return ret; 2048 } 2049 2050 int kvm_vcpu_ioctl(CPUState *cpu, int type, ...) 2051 { 2052 int ret; 2053 void *arg; 2054 va_list ap; 2055 2056 va_start(ap, type); 2057 arg = va_arg(ap, void *); 2058 va_end(ap); 2059 2060 trace_kvm_vcpu_ioctl(cpu->cpu_index, type, arg); 2061 ret = ioctl(cpu->kvm_fd, type, arg); 2062 if (ret == -1) { 2063 ret = -errno; 2064 } 2065 return ret; 2066 } 2067 2068 int kvm_device_ioctl(int fd, int type, ...) 2069 { 2070 int ret; 2071 void *arg; 2072 va_list ap; 2073 2074 va_start(ap, type); 2075 arg = va_arg(ap, void *); 2076 va_end(ap); 2077 2078 trace_kvm_device_ioctl(fd, type, arg); 2079 ret = ioctl(fd, type, arg); 2080 if (ret == -1) { 2081 ret = -errno; 2082 } 2083 return ret; 2084 } 2085 2086 int kvm_vm_check_attr(KVMState *s, uint32_t group, uint64_t attr) 2087 { 2088 int ret; 2089 struct kvm_device_attr attribute = { 2090 .group = group, 2091 .attr = attr, 2092 }; 2093 2094 if (!kvm_vm_attributes_allowed) { 2095 return 0; 2096 } 2097 2098 ret = kvm_vm_ioctl(s, KVM_HAS_DEVICE_ATTR, &attribute); 2099 /* kvm returns 0 on success for HAS_DEVICE_ATTR */ 2100 return ret ? 0 : 1; 2101 } 2102 2103 int kvm_device_check_attr(int dev_fd, uint32_t group, uint64_t attr) 2104 { 2105 struct kvm_device_attr attribute = { 2106 .group = group, 2107 .attr = attr, 2108 .flags = 0, 2109 }; 2110 2111 return kvm_device_ioctl(dev_fd, KVM_HAS_DEVICE_ATTR, &attribute) ? 0 : 1; 2112 } 2113 2114 int kvm_device_access(int fd, int group, uint64_t attr, 2115 void *val, bool write, Error **errp) 2116 { 2117 struct kvm_device_attr kvmattr; 2118 int err; 2119 2120 kvmattr.flags = 0; 2121 kvmattr.group = group; 2122 kvmattr.attr = attr; 2123 kvmattr.addr = (uintptr_t)val; 2124 2125 err = kvm_device_ioctl(fd, 2126 write ? KVM_SET_DEVICE_ATTR : KVM_GET_DEVICE_ATTR, 2127 &kvmattr); 2128 if (err < 0) { 2129 error_setg_errno(errp, -err, 2130 "KVM_%s_DEVICE_ATTR failed: Group %d " 2131 "attr 0x%016" PRIx64, 2132 write ? "SET" : "GET", group, attr); 2133 } 2134 return err; 2135 } 2136 2137 bool kvm_has_sync_mmu(void) 2138 { 2139 return kvm_state->sync_mmu; 2140 } 2141 2142 int kvm_has_vcpu_events(void) 2143 { 2144 return kvm_state->vcpu_events; 2145 } 2146 2147 int kvm_has_robust_singlestep(void) 2148 { 2149 return kvm_state->robust_singlestep; 2150 } 2151 2152 int kvm_has_debugregs(void) 2153 { 2154 return kvm_state->debugregs; 2155 } 2156 2157 int kvm_has_many_ioeventfds(void) 2158 { 2159 if (!kvm_enabled()) { 2160 return 0; 2161 } 2162 return kvm_state->many_ioeventfds; 2163 } 2164 2165 int kvm_has_gsi_routing(void) 2166 { 2167 #ifdef KVM_CAP_IRQ_ROUTING 2168 return kvm_check_extension(kvm_state, KVM_CAP_IRQ_ROUTING); 2169 #else 2170 return false; 2171 #endif 2172 } 2173 2174 int kvm_has_intx_set_mask(void) 2175 { 2176 return kvm_state->intx_set_mask; 2177 } 2178 2179 bool kvm_arm_supports_user_irq(void) 2180 { 2181 return kvm_check_extension(kvm_state, KVM_CAP_ARM_USER_IRQ); 2182 } 2183 2184 #ifdef KVM_CAP_SET_GUEST_DEBUG 2185 struct kvm_sw_breakpoint *kvm_find_sw_breakpoint(CPUState *cpu, 2186 target_ulong pc) 2187 { 2188 struct kvm_sw_breakpoint *bp; 2189 2190 QTAILQ_FOREACH(bp, &cpu->kvm_state->kvm_sw_breakpoints, entry) { 2191 if (bp->pc == pc) { 2192 return bp; 2193 } 2194 } 2195 return NULL; 2196 } 2197 2198 int kvm_sw_breakpoints_active(CPUState *cpu) 2199 { 2200 return !QTAILQ_EMPTY(&cpu->kvm_state->kvm_sw_breakpoints); 2201 } 2202 2203 struct kvm_set_guest_debug_data { 2204 struct kvm_guest_debug dbg; 2205 int err; 2206 }; 2207 2208 static void kvm_invoke_set_guest_debug(CPUState *cpu, run_on_cpu_data data) 2209 { 2210 struct kvm_set_guest_debug_data *dbg_data = 2211 (struct kvm_set_guest_debug_data *) data.host_ptr; 2212 2213 dbg_data->err = kvm_vcpu_ioctl(cpu, KVM_SET_GUEST_DEBUG, 2214 &dbg_data->dbg); 2215 } 2216 2217 int kvm_update_guest_debug(CPUState *cpu, unsigned long reinject_trap) 2218 { 2219 struct kvm_set_guest_debug_data data; 2220 2221 data.dbg.control = reinject_trap; 2222 2223 if (cpu->singlestep_enabled) { 2224 data.dbg.control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_SINGLESTEP; 2225 } 2226 kvm_arch_update_guest_debug(cpu, &data.dbg); 2227 2228 run_on_cpu(cpu, kvm_invoke_set_guest_debug, 2229 RUN_ON_CPU_HOST_PTR(&data)); 2230 return data.err; 2231 } 2232 2233 int kvm_insert_breakpoint(CPUState *cpu, target_ulong addr, 2234 target_ulong len, int type) 2235 { 2236 struct kvm_sw_breakpoint *bp; 2237 int err; 2238 2239 if (type == GDB_BREAKPOINT_SW) { 2240 bp = kvm_find_sw_breakpoint(cpu, addr); 2241 if (bp) { 2242 bp->use_count++; 2243 return 0; 2244 } 2245 2246 bp = g_malloc(sizeof(struct kvm_sw_breakpoint)); 2247 bp->pc = addr; 2248 bp->use_count = 1; 2249 err = kvm_arch_insert_sw_breakpoint(cpu, bp); 2250 if (err) { 2251 g_free(bp); 2252 return err; 2253 } 2254 2255 QTAILQ_INSERT_HEAD(&cpu->kvm_state->kvm_sw_breakpoints, bp, entry); 2256 } else { 2257 err = kvm_arch_insert_hw_breakpoint(addr, len, type); 2258 if (err) { 2259 return err; 2260 } 2261 } 2262 2263 CPU_FOREACH(cpu) { 2264 err = kvm_update_guest_debug(cpu, 0); 2265 if (err) { 2266 return err; 2267 } 2268 } 2269 return 0; 2270 } 2271 2272 int kvm_remove_breakpoint(CPUState *cpu, target_ulong addr, 2273 target_ulong len, int type) 2274 { 2275 struct kvm_sw_breakpoint *bp; 2276 int err; 2277 2278 if (type == GDB_BREAKPOINT_SW) { 2279 bp = kvm_find_sw_breakpoint(cpu, addr); 2280 if (!bp) { 2281 return -ENOENT; 2282 } 2283 2284 if (bp->use_count > 1) { 2285 bp->use_count--; 2286 return 0; 2287 } 2288 2289 err = kvm_arch_remove_sw_breakpoint(cpu, bp); 2290 if (err) { 2291 return err; 2292 } 2293 2294 QTAILQ_REMOVE(&cpu->kvm_state->kvm_sw_breakpoints, bp, entry); 2295 g_free(bp); 2296 } else { 2297 err = kvm_arch_remove_hw_breakpoint(addr, len, type); 2298 if (err) { 2299 return err; 2300 } 2301 } 2302 2303 CPU_FOREACH(cpu) { 2304 err = kvm_update_guest_debug(cpu, 0); 2305 if (err) { 2306 return err; 2307 } 2308 } 2309 return 0; 2310 } 2311 2312 void kvm_remove_all_breakpoints(CPUState *cpu) 2313 { 2314 struct kvm_sw_breakpoint *bp, *next; 2315 KVMState *s = cpu->kvm_state; 2316 CPUState *tmpcpu; 2317 2318 QTAILQ_FOREACH_SAFE(bp, &s->kvm_sw_breakpoints, entry, next) { 2319 if (kvm_arch_remove_sw_breakpoint(cpu, bp) != 0) { 2320 /* Try harder to find a CPU that currently sees the breakpoint. */ 2321 CPU_FOREACH(tmpcpu) { 2322 if (kvm_arch_remove_sw_breakpoint(tmpcpu, bp) == 0) { 2323 break; 2324 } 2325 } 2326 } 2327 QTAILQ_REMOVE(&s->kvm_sw_breakpoints, bp, entry); 2328 g_free(bp); 2329 } 2330 kvm_arch_remove_all_hw_breakpoints(); 2331 2332 CPU_FOREACH(cpu) { 2333 kvm_update_guest_debug(cpu, 0); 2334 } 2335 } 2336 2337 #else /* !KVM_CAP_SET_GUEST_DEBUG */ 2338 2339 int kvm_update_guest_debug(CPUState *cpu, unsigned long reinject_trap) 2340 { 2341 return -EINVAL; 2342 } 2343 2344 int kvm_insert_breakpoint(CPUState *cpu, target_ulong addr, 2345 target_ulong len, int type) 2346 { 2347 return -EINVAL; 2348 } 2349 2350 int kvm_remove_breakpoint(CPUState *cpu, target_ulong addr, 2351 target_ulong len, int type) 2352 { 2353 return -EINVAL; 2354 } 2355 2356 void kvm_remove_all_breakpoints(CPUState *cpu) 2357 { 2358 } 2359 #endif /* !KVM_CAP_SET_GUEST_DEBUG */ 2360 2361 static int kvm_set_signal_mask(CPUState *cpu, const sigset_t *sigset) 2362 { 2363 KVMState *s = kvm_state; 2364 struct kvm_signal_mask *sigmask; 2365 int r; 2366 2367 sigmask = g_malloc(sizeof(*sigmask) + sizeof(*sigset)); 2368 2369 sigmask->len = s->sigmask_len; 2370 memcpy(sigmask->sigset, sigset, sizeof(*sigset)); 2371 r = kvm_vcpu_ioctl(cpu, KVM_SET_SIGNAL_MASK, sigmask); 2372 g_free(sigmask); 2373 2374 return r; 2375 } 2376 2377 static void kvm_ipi_signal(int sig) 2378 { 2379 if (current_cpu) { 2380 assert(kvm_immediate_exit); 2381 kvm_cpu_kick(current_cpu); 2382 } 2383 } 2384 2385 void kvm_init_cpu_signals(CPUState *cpu) 2386 { 2387 int r; 2388 sigset_t set; 2389 struct sigaction sigact; 2390 2391 memset(&sigact, 0, sizeof(sigact)); 2392 sigact.sa_handler = kvm_ipi_signal; 2393 sigaction(SIG_IPI, &sigact, NULL); 2394 2395 pthread_sigmask(SIG_BLOCK, NULL, &set); 2396 #if defined KVM_HAVE_MCE_INJECTION 2397 sigdelset(&set, SIGBUS); 2398 pthread_sigmask(SIG_SETMASK, &set, NULL); 2399 #endif 2400 sigdelset(&set, SIG_IPI); 2401 if (kvm_immediate_exit) { 2402 r = pthread_sigmask(SIG_SETMASK, &set, NULL); 2403 } else { 2404 r = kvm_set_signal_mask(cpu, &set); 2405 } 2406 if (r) { 2407 fprintf(stderr, "kvm_set_signal_mask: %s\n", strerror(-r)); 2408 exit(1); 2409 } 2410 } 2411 2412 /* Called asynchronously in VCPU thread. */ 2413 int kvm_on_sigbus_vcpu(CPUState *cpu, int code, void *addr) 2414 { 2415 #ifdef KVM_HAVE_MCE_INJECTION 2416 if (have_sigbus_pending) { 2417 return 1; 2418 } 2419 have_sigbus_pending = true; 2420 pending_sigbus_addr = addr; 2421 pending_sigbus_code = code; 2422 atomic_set(&cpu->exit_request, 1); 2423 return 0; 2424 #else 2425 return 1; 2426 #endif 2427 } 2428 2429 /* Called synchronously (via signalfd) in main thread. */ 2430 int kvm_on_sigbus(int code, void *addr) 2431 { 2432 #ifdef KVM_HAVE_MCE_INJECTION 2433 /* Action required MCE kills the process if SIGBUS is blocked. Because 2434 * that's what happens in the I/O thread, where we handle MCE via signalfd, 2435 * we can only get action optional here. 2436 */ 2437 assert(code != BUS_MCEERR_AR); 2438 kvm_arch_on_sigbus_vcpu(first_cpu, code, addr); 2439 return 0; 2440 #else 2441 return 1; 2442 #endif 2443 } 2444 2445 int kvm_create_device(KVMState *s, uint64_t type, bool test) 2446 { 2447 int ret; 2448 struct kvm_create_device create_dev; 2449 2450 create_dev.type = type; 2451 create_dev.fd = -1; 2452 create_dev.flags = test ? KVM_CREATE_DEVICE_TEST : 0; 2453 2454 if (!kvm_check_extension(s, KVM_CAP_DEVICE_CTRL)) { 2455 return -ENOTSUP; 2456 } 2457 2458 ret = kvm_vm_ioctl(s, KVM_CREATE_DEVICE, &create_dev); 2459 if (ret) { 2460 return ret; 2461 } 2462 2463 return test ? 0 : create_dev.fd; 2464 } 2465 2466 bool kvm_device_supported(int vmfd, uint64_t type) 2467 { 2468 struct kvm_create_device create_dev = { 2469 .type = type, 2470 .fd = -1, 2471 .flags = KVM_CREATE_DEVICE_TEST, 2472 }; 2473 2474 if (ioctl(vmfd, KVM_CHECK_EXTENSION, KVM_CAP_DEVICE_CTRL) <= 0) { 2475 return false; 2476 } 2477 2478 return (ioctl(vmfd, KVM_CREATE_DEVICE, &create_dev) >= 0); 2479 } 2480 2481 int kvm_set_one_reg(CPUState *cs, uint64_t id, void *source) 2482 { 2483 struct kvm_one_reg reg; 2484 int r; 2485 2486 reg.id = id; 2487 reg.addr = (uintptr_t) source; 2488 r = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, ®); 2489 if (r) { 2490 trace_kvm_failed_reg_set(id, strerror(-r)); 2491 } 2492 return r; 2493 } 2494 2495 int kvm_get_one_reg(CPUState *cs, uint64_t id, void *target) 2496 { 2497 struct kvm_one_reg reg; 2498 int r; 2499 2500 reg.id = id; 2501 reg.addr = (uintptr_t) target; 2502 r = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, ®); 2503 if (r) { 2504 trace_kvm_failed_reg_get(id, strerror(-r)); 2505 } 2506 return r; 2507 } 2508 2509 static void kvm_accel_class_init(ObjectClass *oc, void *data) 2510 { 2511 AccelClass *ac = ACCEL_CLASS(oc); 2512 ac->name = "KVM"; 2513 ac->init_machine = kvm_init; 2514 ac->allowed = &kvm_allowed; 2515 } 2516 2517 static const TypeInfo kvm_accel_type = { 2518 .name = TYPE_KVM_ACCEL, 2519 .parent = TYPE_ACCEL, 2520 .class_init = kvm_accel_class_init, 2521 .instance_size = sizeof(KVMState), 2522 }; 2523 2524 static void kvm_type_init(void) 2525 { 2526 type_register_static(&kvm_accel_type); 2527 } 2528 2529 type_init(kvm_type_init); 2530