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/atomic.h" 22 #include "qemu/option.h" 23 #include "qemu/config-file.h" 24 #include "qemu/error-report.h" 25 #include "qapi/error.h" 26 #include "hw/pci/msi.h" 27 #include "hw/pci/msix.h" 28 #include "hw/s390x/adapter.h" 29 #include "exec/gdbstub.h" 30 #include "sysemu/kvm_int.h" 31 #include "sysemu/runstate.h" 32 #include "sysemu/cpus.h" 33 #include "sysemu/sysemu.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 "qemu/main-loop.h" 40 #include "trace.h" 41 #include "hw/irq.h" 42 #include "sysemu/sev.h" 43 #include "qapi/visitor.h" 44 #include "qapi/qapi-types-common.h" 45 #include "qapi/qapi-visit-common.h" 46 #include "sysemu/reset.h" 47 #include "qemu/guest-random.h" 48 #include "sysemu/hw_accel.h" 49 #include "kvm-cpus.h" 50 51 #include "hw/boards.h" 52 53 /* This check must be after config-host.h is included */ 54 #ifdef CONFIG_EVENTFD 55 #include <sys/eventfd.h> 56 #endif 57 58 /* KVM uses PAGE_SIZE in its definition of KVM_COALESCED_MMIO_MAX. We 59 * need to use the real host PAGE_SIZE, as that's what KVM will use. 60 */ 61 #define PAGE_SIZE qemu_real_host_page_size 62 63 //#define DEBUG_KVM 64 65 #ifdef DEBUG_KVM 66 #define DPRINTF(fmt, ...) \ 67 do { fprintf(stderr, fmt, ## __VA_ARGS__); } while (0) 68 #else 69 #define DPRINTF(fmt, ...) \ 70 do { } while (0) 71 #endif 72 73 #define KVM_MSI_HASHTAB_SIZE 256 74 75 struct KVMParkedVcpu { 76 unsigned long vcpu_id; 77 int kvm_fd; 78 QLIST_ENTRY(KVMParkedVcpu) node; 79 }; 80 81 struct KVMState 82 { 83 AccelState parent_obj; 84 85 int nr_slots; 86 int fd; 87 int vmfd; 88 int coalesced_mmio; 89 int coalesced_pio; 90 struct kvm_coalesced_mmio_ring *coalesced_mmio_ring; 91 bool coalesced_flush_in_progress; 92 int vcpu_events; 93 int robust_singlestep; 94 int debugregs; 95 #ifdef KVM_CAP_SET_GUEST_DEBUG 96 QTAILQ_HEAD(, kvm_sw_breakpoint) kvm_sw_breakpoints; 97 #endif 98 int max_nested_state_len; 99 int many_ioeventfds; 100 int intx_set_mask; 101 int kvm_shadow_mem; 102 bool kernel_irqchip_allowed; 103 bool kernel_irqchip_required; 104 OnOffAuto kernel_irqchip_split; 105 bool sync_mmu; 106 uint64_t manual_dirty_log_protect; 107 /* The man page (and posix) say ioctl numbers are signed int, but 108 * they're not. Linux, glibc and *BSD all treat ioctl numbers as 109 * unsigned, and treating them as signed here can break things */ 110 unsigned irq_set_ioctl; 111 unsigned int sigmask_len; 112 GHashTable *gsimap; 113 #ifdef KVM_CAP_IRQ_ROUTING 114 struct kvm_irq_routing *irq_routes; 115 int nr_allocated_irq_routes; 116 unsigned long *used_gsi_bitmap; 117 unsigned int gsi_count; 118 QTAILQ_HEAD(, KVMMSIRoute) msi_hashtab[KVM_MSI_HASHTAB_SIZE]; 119 #endif 120 KVMMemoryListener memory_listener; 121 QLIST_HEAD(, KVMParkedVcpu) kvm_parked_vcpus; 122 123 /* memory encryption */ 124 void *memcrypt_handle; 125 int (*memcrypt_encrypt_data)(void *handle, uint8_t *ptr, uint64_t len); 126 127 /* For "info mtree -f" to tell if an MR is registered in KVM */ 128 int nr_as; 129 struct KVMAs { 130 KVMMemoryListener *ml; 131 AddressSpace *as; 132 } *as; 133 }; 134 135 KVMState *kvm_state; 136 bool kvm_kernel_irqchip; 137 bool kvm_split_irqchip; 138 bool kvm_async_interrupts_allowed; 139 bool kvm_halt_in_kernel_allowed; 140 bool kvm_eventfds_allowed; 141 bool kvm_irqfds_allowed; 142 bool kvm_resamplefds_allowed; 143 bool kvm_msi_via_irqfd_allowed; 144 bool kvm_gsi_routing_allowed; 145 bool kvm_gsi_direct_mapping; 146 bool kvm_allowed; 147 bool kvm_readonly_mem_allowed; 148 bool kvm_vm_attributes_allowed; 149 bool kvm_direct_msi_allowed; 150 bool kvm_ioeventfd_any_length_allowed; 151 bool kvm_msi_use_devid; 152 static bool kvm_immediate_exit; 153 static hwaddr kvm_max_slot_size = ~0; 154 155 static const KVMCapabilityInfo kvm_required_capabilites[] = { 156 KVM_CAP_INFO(USER_MEMORY), 157 KVM_CAP_INFO(DESTROY_MEMORY_REGION_WORKS), 158 KVM_CAP_INFO(JOIN_MEMORY_REGIONS_WORKS), 159 KVM_CAP_LAST_INFO 160 }; 161 162 static NotifierList kvm_irqchip_change_notifiers = 163 NOTIFIER_LIST_INITIALIZER(kvm_irqchip_change_notifiers); 164 165 struct KVMResampleFd { 166 int gsi; 167 EventNotifier *resample_event; 168 QLIST_ENTRY(KVMResampleFd) node; 169 }; 170 typedef struct KVMResampleFd KVMResampleFd; 171 172 /* 173 * Only used with split irqchip where we need to do the resample fd 174 * kick for the kernel from userspace. 175 */ 176 static QLIST_HEAD(, KVMResampleFd) kvm_resample_fd_list = 177 QLIST_HEAD_INITIALIZER(kvm_resample_fd_list); 178 179 #define kvm_slots_lock(kml) qemu_mutex_lock(&(kml)->slots_lock) 180 #define kvm_slots_unlock(kml) qemu_mutex_unlock(&(kml)->slots_lock) 181 182 static inline void kvm_resample_fd_remove(int gsi) 183 { 184 KVMResampleFd *rfd; 185 186 QLIST_FOREACH(rfd, &kvm_resample_fd_list, node) { 187 if (rfd->gsi == gsi) { 188 QLIST_REMOVE(rfd, node); 189 g_free(rfd); 190 break; 191 } 192 } 193 } 194 195 static inline void kvm_resample_fd_insert(int gsi, EventNotifier *event) 196 { 197 KVMResampleFd *rfd = g_new0(KVMResampleFd, 1); 198 199 rfd->gsi = gsi; 200 rfd->resample_event = event; 201 202 QLIST_INSERT_HEAD(&kvm_resample_fd_list, rfd, node); 203 } 204 205 void kvm_resample_fd_notify(int gsi) 206 { 207 KVMResampleFd *rfd; 208 209 QLIST_FOREACH(rfd, &kvm_resample_fd_list, node) { 210 if (rfd->gsi == gsi) { 211 event_notifier_set(rfd->resample_event); 212 trace_kvm_resample_fd_notify(gsi); 213 return; 214 } 215 } 216 } 217 218 int kvm_get_max_memslots(void) 219 { 220 KVMState *s = KVM_STATE(current_accel()); 221 222 return s->nr_slots; 223 } 224 225 bool kvm_memcrypt_enabled(void) 226 { 227 if (kvm_state && kvm_state->memcrypt_handle) { 228 return true; 229 } 230 231 return false; 232 } 233 234 int kvm_memcrypt_encrypt_data(uint8_t *ptr, uint64_t len) 235 { 236 if (kvm_state->memcrypt_handle && 237 kvm_state->memcrypt_encrypt_data) { 238 return kvm_state->memcrypt_encrypt_data(kvm_state->memcrypt_handle, 239 ptr, len); 240 } 241 242 return 1; 243 } 244 245 /* Called with KVMMemoryListener.slots_lock held */ 246 static KVMSlot *kvm_get_free_slot(KVMMemoryListener *kml) 247 { 248 KVMState *s = kvm_state; 249 int i; 250 251 for (i = 0; i < s->nr_slots; i++) { 252 if (kml->slots[i].memory_size == 0) { 253 return &kml->slots[i]; 254 } 255 } 256 257 return NULL; 258 } 259 260 bool kvm_has_free_slot(MachineState *ms) 261 { 262 KVMState *s = KVM_STATE(ms->accelerator); 263 bool result; 264 KVMMemoryListener *kml = &s->memory_listener; 265 266 kvm_slots_lock(kml); 267 result = !!kvm_get_free_slot(kml); 268 kvm_slots_unlock(kml); 269 270 return result; 271 } 272 273 /* Called with KVMMemoryListener.slots_lock held */ 274 static KVMSlot *kvm_alloc_slot(KVMMemoryListener *kml) 275 { 276 KVMSlot *slot = kvm_get_free_slot(kml); 277 278 if (slot) { 279 return slot; 280 } 281 282 fprintf(stderr, "%s: no free slot available\n", __func__); 283 abort(); 284 } 285 286 static KVMSlot *kvm_lookup_matching_slot(KVMMemoryListener *kml, 287 hwaddr start_addr, 288 hwaddr size) 289 { 290 KVMState *s = kvm_state; 291 int i; 292 293 for (i = 0; i < s->nr_slots; i++) { 294 KVMSlot *mem = &kml->slots[i]; 295 296 if (start_addr == mem->start_addr && size == mem->memory_size) { 297 return mem; 298 } 299 } 300 301 return NULL; 302 } 303 304 /* 305 * Calculate and align the start address and the size of the section. 306 * Return the size. If the size is 0, the aligned section is empty. 307 */ 308 static hwaddr kvm_align_section(MemoryRegionSection *section, 309 hwaddr *start) 310 { 311 hwaddr size = int128_get64(section->size); 312 hwaddr delta, aligned; 313 314 /* kvm works in page size chunks, but the function may be called 315 with sub-page size and unaligned start address. Pad the start 316 address to next and truncate size to previous page boundary. */ 317 aligned = ROUND_UP(section->offset_within_address_space, 318 qemu_real_host_page_size); 319 delta = aligned - section->offset_within_address_space; 320 *start = aligned; 321 if (delta > size) { 322 return 0; 323 } 324 325 return (size - delta) & qemu_real_host_page_mask; 326 } 327 328 int kvm_physical_memory_addr_from_host(KVMState *s, void *ram, 329 hwaddr *phys_addr) 330 { 331 KVMMemoryListener *kml = &s->memory_listener; 332 int i, ret = 0; 333 334 kvm_slots_lock(kml); 335 for (i = 0; i < s->nr_slots; i++) { 336 KVMSlot *mem = &kml->slots[i]; 337 338 if (ram >= mem->ram && ram < mem->ram + mem->memory_size) { 339 *phys_addr = mem->start_addr + (ram - mem->ram); 340 ret = 1; 341 break; 342 } 343 } 344 kvm_slots_unlock(kml); 345 346 return ret; 347 } 348 349 static int kvm_set_user_memory_region(KVMMemoryListener *kml, KVMSlot *slot, bool new) 350 { 351 KVMState *s = kvm_state; 352 struct kvm_userspace_memory_region mem; 353 int ret; 354 355 mem.slot = slot->slot | (kml->as_id << 16); 356 mem.guest_phys_addr = slot->start_addr; 357 mem.userspace_addr = (unsigned long)slot->ram; 358 mem.flags = slot->flags; 359 360 if (slot->memory_size && !new && (mem.flags ^ slot->old_flags) & KVM_MEM_READONLY) { 361 /* Set the slot size to 0 before setting the slot to the desired 362 * value. This is needed based on KVM commit 75d61fbc. */ 363 mem.memory_size = 0; 364 ret = kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem); 365 if (ret < 0) { 366 goto err; 367 } 368 } 369 mem.memory_size = slot->memory_size; 370 ret = kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem); 371 slot->old_flags = mem.flags; 372 err: 373 trace_kvm_set_user_memory(mem.slot, mem.flags, mem.guest_phys_addr, 374 mem.memory_size, mem.userspace_addr, ret); 375 if (ret < 0) { 376 error_report("%s: KVM_SET_USER_MEMORY_REGION failed, slot=%d," 377 " start=0x%" PRIx64 ", size=0x%" PRIx64 ": %s", 378 __func__, mem.slot, slot->start_addr, 379 (uint64_t)mem.memory_size, strerror(errno)); 380 } 381 return ret; 382 } 383 384 static int do_kvm_destroy_vcpu(CPUState *cpu) 385 { 386 KVMState *s = kvm_state; 387 long mmap_size; 388 struct KVMParkedVcpu *vcpu = NULL; 389 int ret = 0; 390 391 DPRINTF("kvm_destroy_vcpu\n"); 392 393 ret = kvm_arch_destroy_vcpu(cpu); 394 if (ret < 0) { 395 goto err; 396 } 397 398 mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0); 399 if (mmap_size < 0) { 400 ret = mmap_size; 401 DPRINTF("KVM_GET_VCPU_MMAP_SIZE failed\n"); 402 goto err; 403 } 404 405 ret = munmap(cpu->kvm_run, mmap_size); 406 if (ret < 0) { 407 goto err; 408 } 409 410 vcpu = g_malloc0(sizeof(*vcpu)); 411 vcpu->vcpu_id = kvm_arch_vcpu_id(cpu); 412 vcpu->kvm_fd = cpu->kvm_fd; 413 QLIST_INSERT_HEAD(&kvm_state->kvm_parked_vcpus, vcpu, node); 414 err: 415 return ret; 416 } 417 418 void kvm_destroy_vcpu(CPUState *cpu) 419 { 420 if (do_kvm_destroy_vcpu(cpu) < 0) { 421 error_report("kvm_destroy_vcpu failed"); 422 exit(EXIT_FAILURE); 423 } 424 } 425 426 static int kvm_get_vcpu(KVMState *s, unsigned long vcpu_id) 427 { 428 struct KVMParkedVcpu *cpu; 429 430 QLIST_FOREACH(cpu, &s->kvm_parked_vcpus, node) { 431 if (cpu->vcpu_id == vcpu_id) { 432 int kvm_fd; 433 434 QLIST_REMOVE(cpu, node); 435 kvm_fd = cpu->kvm_fd; 436 g_free(cpu); 437 return kvm_fd; 438 } 439 } 440 441 return kvm_vm_ioctl(s, KVM_CREATE_VCPU, (void *)vcpu_id); 442 } 443 444 int kvm_init_vcpu(CPUState *cpu) 445 { 446 KVMState *s = kvm_state; 447 long mmap_size; 448 int ret; 449 450 DPRINTF("kvm_init_vcpu\n"); 451 452 ret = kvm_get_vcpu(s, kvm_arch_vcpu_id(cpu)); 453 if (ret < 0) { 454 DPRINTF("kvm_create_vcpu failed\n"); 455 goto err; 456 } 457 458 cpu->kvm_fd = ret; 459 cpu->kvm_state = s; 460 cpu->vcpu_dirty = true; 461 462 mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0); 463 if (mmap_size < 0) { 464 ret = mmap_size; 465 DPRINTF("KVM_GET_VCPU_MMAP_SIZE failed\n"); 466 goto err; 467 } 468 469 cpu->kvm_run = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED, 470 cpu->kvm_fd, 0); 471 if (cpu->kvm_run == MAP_FAILED) { 472 ret = -errno; 473 DPRINTF("mmap'ing vcpu state failed\n"); 474 goto err; 475 } 476 477 if (s->coalesced_mmio && !s->coalesced_mmio_ring) { 478 s->coalesced_mmio_ring = 479 (void *)cpu->kvm_run + s->coalesced_mmio * PAGE_SIZE; 480 } 481 482 ret = kvm_arch_init_vcpu(cpu); 483 err: 484 return ret; 485 } 486 487 /* 488 * dirty pages logging control 489 */ 490 491 static int kvm_mem_flags(MemoryRegion *mr) 492 { 493 bool readonly = mr->readonly || memory_region_is_romd(mr); 494 int flags = 0; 495 496 if (memory_region_get_dirty_log_mask(mr) != 0) { 497 flags |= KVM_MEM_LOG_DIRTY_PAGES; 498 } 499 if (readonly && kvm_readonly_mem_allowed) { 500 flags |= KVM_MEM_READONLY; 501 } 502 return flags; 503 } 504 505 /* Called with KVMMemoryListener.slots_lock held */ 506 static int kvm_slot_update_flags(KVMMemoryListener *kml, KVMSlot *mem, 507 MemoryRegion *mr) 508 { 509 mem->flags = kvm_mem_flags(mr); 510 511 /* If nothing changed effectively, no need to issue ioctl */ 512 if (mem->flags == mem->old_flags) { 513 return 0; 514 } 515 516 return kvm_set_user_memory_region(kml, mem, false); 517 } 518 519 static int kvm_section_update_flags(KVMMemoryListener *kml, 520 MemoryRegionSection *section) 521 { 522 hwaddr start_addr, size, slot_size; 523 KVMSlot *mem; 524 int ret = 0; 525 526 size = kvm_align_section(section, &start_addr); 527 if (!size) { 528 return 0; 529 } 530 531 kvm_slots_lock(kml); 532 533 while (size && !ret) { 534 slot_size = MIN(kvm_max_slot_size, size); 535 mem = kvm_lookup_matching_slot(kml, start_addr, slot_size); 536 if (!mem) { 537 /* We don't have a slot if we want to trap every access. */ 538 goto out; 539 } 540 541 ret = kvm_slot_update_flags(kml, mem, section->mr); 542 start_addr += slot_size; 543 size -= slot_size; 544 } 545 546 out: 547 kvm_slots_unlock(kml); 548 return ret; 549 } 550 551 static void kvm_log_start(MemoryListener *listener, 552 MemoryRegionSection *section, 553 int old, int new) 554 { 555 KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener); 556 int r; 557 558 if (old != 0) { 559 return; 560 } 561 562 r = kvm_section_update_flags(kml, section); 563 if (r < 0) { 564 abort(); 565 } 566 } 567 568 static void kvm_log_stop(MemoryListener *listener, 569 MemoryRegionSection *section, 570 int old, int new) 571 { 572 KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener); 573 int r; 574 575 if (new != 0) { 576 return; 577 } 578 579 r = kvm_section_update_flags(kml, section); 580 if (r < 0) { 581 abort(); 582 } 583 } 584 585 /* get kvm's dirty pages bitmap and update qemu's */ 586 static int kvm_get_dirty_pages_log_range(MemoryRegionSection *section, 587 unsigned long *bitmap) 588 { 589 ram_addr_t start = section->offset_within_region + 590 memory_region_get_ram_addr(section->mr); 591 ram_addr_t pages = int128_get64(section->size) / qemu_real_host_page_size; 592 593 cpu_physical_memory_set_dirty_lebitmap(bitmap, start, pages); 594 return 0; 595 } 596 597 #define ALIGN(x, y) (((x)+(y)-1) & ~((y)-1)) 598 599 /* Allocate the dirty bitmap for a slot */ 600 static void kvm_memslot_init_dirty_bitmap(KVMSlot *mem) 601 { 602 /* 603 * XXX bad kernel interface alert 604 * For dirty bitmap, kernel allocates array of size aligned to 605 * bits-per-long. But for case when the kernel is 64bits and 606 * the userspace is 32bits, userspace can't align to the same 607 * bits-per-long, since sizeof(long) is different between kernel 608 * and user space. This way, userspace will provide buffer which 609 * may be 4 bytes less than the kernel will use, resulting in 610 * userspace memory corruption (which is not detectable by valgrind 611 * too, in most cases). 612 * So for now, let's align to 64 instead of HOST_LONG_BITS here, in 613 * a hope that sizeof(long) won't become >8 any time soon. 614 */ 615 hwaddr bitmap_size = ALIGN(((mem->memory_size) >> TARGET_PAGE_BITS), 616 /*HOST_LONG_BITS*/ 64) / 8; 617 mem->dirty_bmap = g_malloc0(bitmap_size); 618 } 619 620 /** 621 * kvm_physical_sync_dirty_bitmap - Sync dirty bitmap from kernel space 622 * 623 * This function will first try to fetch dirty bitmap from the kernel, 624 * and then updates qemu's dirty bitmap. 625 * 626 * NOTE: caller must be with kml->slots_lock held. 627 * 628 * @kml: the KVM memory listener object 629 * @section: the memory section to sync the dirty bitmap with 630 */ 631 static int kvm_physical_sync_dirty_bitmap(KVMMemoryListener *kml, 632 MemoryRegionSection *section) 633 { 634 KVMState *s = kvm_state; 635 struct kvm_dirty_log d = {}; 636 KVMSlot *mem; 637 hwaddr start_addr, size; 638 hwaddr slot_size, slot_offset = 0; 639 int ret = 0; 640 641 size = kvm_align_section(section, &start_addr); 642 while (size) { 643 MemoryRegionSection subsection = *section; 644 645 slot_size = MIN(kvm_max_slot_size, size); 646 mem = kvm_lookup_matching_slot(kml, start_addr, slot_size); 647 if (!mem) { 648 /* We don't have a slot if we want to trap every access. */ 649 goto out; 650 } 651 652 if (!mem->dirty_bmap) { 653 /* Allocate on the first log_sync, once and for all */ 654 kvm_memslot_init_dirty_bitmap(mem); 655 } 656 657 d.dirty_bitmap = mem->dirty_bmap; 658 d.slot = mem->slot | (kml->as_id << 16); 659 if (kvm_vm_ioctl(s, KVM_GET_DIRTY_LOG, &d) == -1) { 660 DPRINTF("ioctl failed %d\n", errno); 661 ret = -1; 662 goto out; 663 } 664 665 subsection.offset_within_region += slot_offset; 666 subsection.size = int128_make64(slot_size); 667 kvm_get_dirty_pages_log_range(&subsection, d.dirty_bitmap); 668 669 slot_offset += slot_size; 670 start_addr += slot_size; 671 size -= slot_size; 672 } 673 out: 674 return ret; 675 } 676 677 /* Alignment requirement for KVM_CLEAR_DIRTY_LOG - 64 pages */ 678 #define KVM_CLEAR_LOG_SHIFT 6 679 #define KVM_CLEAR_LOG_ALIGN (qemu_real_host_page_size << KVM_CLEAR_LOG_SHIFT) 680 #define KVM_CLEAR_LOG_MASK (-KVM_CLEAR_LOG_ALIGN) 681 682 static int kvm_log_clear_one_slot(KVMSlot *mem, int as_id, uint64_t start, 683 uint64_t size) 684 { 685 KVMState *s = kvm_state; 686 uint64_t end, bmap_start, start_delta, bmap_npages; 687 struct kvm_clear_dirty_log d; 688 unsigned long *bmap_clear = NULL, psize = qemu_real_host_page_size; 689 int ret; 690 691 /* 692 * We need to extend either the start or the size or both to 693 * satisfy the KVM interface requirement. Firstly, do the start 694 * page alignment on 64 host pages 695 */ 696 bmap_start = start & KVM_CLEAR_LOG_MASK; 697 start_delta = start - bmap_start; 698 bmap_start /= psize; 699 700 /* 701 * The kernel interface has restriction on the size too, that either: 702 * 703 * (1) the size is 64 host pages aligned (just like the start), or 704 * (2) the size fills up until the end of the KVM memslot. 705 */ 706 bmap_npages = DIV_ROUND_UP(size + start_delta, KVM_CLEAR_LOG_ALIGN) 707 << KVM_CLEAR_LOG_SHIFT; 708 end = mem->memory_size / psize; 709 if (bmap_npages > end - bmap_start) { 710 bmap_npages = end - bmap_start; 711 } 712 start_delta /= psize; 713 714 /* 715 * Prepare the bitmap to clear dirty bits. Here we must guarantee 716 * that we won't clear any unknown dirty bits otherwise we might 717 * accidentally clear some set bits which are not yet synced from 718 * the kernel into QEMU's bitmap, then we'll lose track of the 719 * guest modifications upon those pages (which can directly lead 720 * to guest data loss or panic after migration). 721 * 722 * Layout of the KVMSlot.dirty_bmap: 723 * 724 * |<-------- bmap_npages -----------..>| 725 * [1] 726 * start_delta size 727 * |----------------|-------------|------------------|------------| 728 * ^ ^ ^ ^ 729 * | | | | 730 * start bmap_start (start) end 731 * of memslot of memslot 732 * 733 * [1] bmap_npages can be aligned to either 64 pages or the end of slot 734 */ 735 736 assert(bmap_start % BITS_PER_LONG == 0); 737 /* We should never do log_clear before log_sync */ 738 assert(mem->dirty_bmap); 739 if (start_delta) { 740 /* Slow path - we need to manipulate a temp bitmap */ 741 bmap_clear = bitmap_new(bmap_npages); 742 bitmap_copy_with_src_offset(bmap_clear, mem->dirty_bmap, 743 bmap_start, start_delta + size / psize); 744 /* 745 * We need to fill the holes at start because that was not 746 * specified by the caller and we extended the bitmap only for 747 * 64 pages alignment 748 */ 749 bitmap_clear(bmap_clear, 0, start_delta); 750 d.dirty_bitmap = bmap_clear; 751 } else { 752 /* Fast path - start address aligns well with BITS_PER_LONG */ 753 d.dirty_bitmap = mem->dirty_bmap + BIT_WORD(bmap_start); 754 } 755 756 d.first_page = bmap_start; 757 /* It should never overflow. If it happens, say something */ 758 assert(bmap_npages <= UINT32_MAX); 759 d.num_pages = bmap_npages; 760 d.slot = mem->slot | (as_id << 16); 761 762 if (kvm_vm_ioctl(s, KVM_CLEAR_DIRTY_LOG, &d) == -1) { 763 ret = -errno; 764 error_report("%s: KVM_CLEAR_DIRTY_LOG failed, slot=%d, " 765 "start=0x%"PRIx64", size=0x%"PRIx32", errno=%d", 766 __func__, d.slot, (uint64_t)d.first_page, 767 (uint32_t)d.num_pages, ret); 768 } else { 769 ret = 0; 770 trace_kvm_clear_dirty_log(d.slot, d.first_page, d.num_pages); 771 } 772 773 /* 774 * After we have updated the remote dirty bitmap, we update the 775 * cached bitmap as well for the memslot, then if another user 776 * clears the same region we know we shouldn't clear it again on 777 * the remote otherwise it's data loss as well. 778 */ 779 bitmap_clear(mem->dirty_bmap, bmap_start + start_delta, 780 size / psize); 781 /* This handles the NULL case well */ 782 g_free(bmap_clear); 783 return ret; 784 } 785 786 787 /** 788 * kvm_physical_log_clear - Clear the kernel's dirty bitmap for range 789 * 790 * NOTE: this will be a no-op if we haven't enabled manual dirty log 791 * protection in the host kernel because in that case this operation 792 * will be done within log_sync(). 793 * 794 * @kml: the kvm memory listener 795 * @section: the memory range to clear dirty bitmap 796 */ 797 static int kvm_physical_log_clear(KVMMemoryListener *kml, 798 MemoryRegionSection *section) 799 { 800 KVMState *s = kvm_state; 801 uint64_t start, size, offset, count; 802 KVMSlot *mem; 803 int ret = 0, i; 804 805 if (!s->manual_dirty_log_protect) { 806 /* No need to do explicit clear */ 807 return ret; 808 } 809 810 start = section->offset_within_address_space; 811 size = int128_get64(section->size); 812 813 if (!size) { 814 /* Nothing more we can do... */ 815 return ret; 816 } 817 818 kvm_slots_lock(kml); 819 820 for (i = 0; i < s->nr_slots; i++) { 821 mem = &kml->slots[i]; 822 /* Discard slots that are empty or do not overlap the section */ 823 if (!mem->memory_size || 824 mem->start_addr > start + size - 1 || 825 start > mem->start_addr + mem->memory_size - 1) { 826 continue; 827 } 828 829 if (start >= mem->start_addr) { 830 /* The slot starts before section or is aligned to it. */ 831 offset = start - mem->start_addr; 832 count = MIN(mem->memory_size - offset, size); 833 } else { 834 /* The slot starts after section. */ 835 offset = 0; 836 count = MIN(mem->memory_size, size - (mem->start_addr - start)); 837 } 838 ret = kvm_log_clear_one_slot(mem, kml->as_id, offset, count); 839 if (ret < 0) { 840 break; 841 } 842 } 843 844 kvm_slots_unlock(kml); 845 846 return ret; 847 } 848 849 static void kvm_coalesce_mmio_region(MemoryListener *listener, 850 MemoryRegionSection *secion, 851 hwaddr start, hwaddr size) 852 { 853 KVMState *s = kvm_state; 854 855 if (s->coalesced_mmio) { 856 struct kvm_coalesced_mmio_zone zone; 857 858 zone.addr = start; 859 zone.size = size; 860 zone.pad = 0; 861 862 (void)kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone); 863 } 864 } 865 866 static void kvm_uncoalesce_mmio_region(MemoryListener *listener, 867 MemoryRegionSection *secion, 868 hwaddr start, hwaddr size) 869 { 870 KVMState *s = kvm_state; 871 872 if (s->coalesced_mmio) { 873 struct kvm_coalesced_mmio_zone zone; 874 875 zone.addr = start; 876 zone.size = size; 877 zone.pad = 0; 878 879 (void)kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone); 880 } 881 } 882 883 static void kvm_coalesce_pio_add(MemoryListener *listener, 884 MemoryRegionSection *section, 885 hwaddr start, hwaddr size) 886 { 887 KVMState *s = kvm_state; 888 889 if (s->coalesced_pio) { 890 struct kvm_coalesced_mmio_zone zone; 891 892 zone.addr = start; 893 zone.size = size; 894 zone.pio = 1; 895 896 (void)kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone); 897 } 898 } 899 900 static void kvm_coalesce_pio_del(MemoryListener *listener, 901 MemoryRegionSection *section, 902 hwaddr start, hwaddr size) 903 { 904 KVMState *s = kvm_state; 905 906 if (s->coalesced_pio) { 907 struct kvm_coalesced_mmio_zone zone; 908 909 zone.addr = start; 910 zone.size = size; 911 zone.pio = 1; 912 913 (void)kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone); 914 } 915 } 916 917 static MemoryListener kvm_coalesced_pio_listener = { 918 .coalesced_io_add = kvm_coalesce_pio_add, 919 .coalesced_io_del = kvm_coalesce_pio_del, 920 }; 921 922 int kvm_check_extension(KVMState *s, unsigned int extension) 923 { 924 int ret; 925 926 ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, extension); 927 if (ret < 0) { 928 ret = 0; 929 } 930 931 return ret; 932 } 933 934 int kvm_vm_check_extension(KVMState *s, unsigned int extension) 935 { 936 int ret; 937 938 ret = kvm_vm_ioctl(s, KVM_CHECK_EXTENSION, extension); 939 if (ret < 0) { 940 /* VM wide version not implemented, use global one instead */ 941 ret = kvm_check_extension(s, extension); 942 } 943 944 return ret; 945 } 946 947 typedef struct HWPoisonPage { 948 ram_addr_t ram_addr; 949 QLIST_ENTRY(HWPoisonPage) list; 950 } HWPoisonPage; 951 952 static QLIST_HEAD(, HWPoisonPage) hwpoison_page_list = 953 QLIST_HEAD_INITIALIZER(hwpoison_page_list); 954 955 static void kvm_unpoison_all(void *param) 956 { 957 HWPoisonPage *page, *next_page; 958 959 QLIST_FOREACH_SAFE(page, &hwpoison_page_list, list, next_page) { 960 QLIST_REMOVE(page, list); 961 qemu_ram_remap(page->ram_addr, TARGET_PAGE_SIZE); 962 g_free(page); 963 } 964 } 965 966 void kvm_hwpoison_page_add(ram_addr_t ram_addr) 967 { 968 HWPoisonPage *page; 969 970 QLIST_FOREACH(page, &hwpoison_page_list, list) { 971 if (page->ram_addr == ram_addr) { 972 return; 973 } 974 } 975 page = g_new(HWPoisonPage, 1); 976 page->ram_addr = ram_addr; 977 QLIST_INSERT_HEAD(&hwpoison_page_list, page, list); 978 } 979 980 static uint32_t adjust_ioeventfd_endianness(uint32_t val, uint32_t size) 981 { 982 #if defined(HOST_WORDS_BIGENDIAN) != defined(TARGET_WORDS_BIGENDIAN) 983 /* The kernel expects ioeventfd values in HOST_WORDS_BIGENDIAN 984 * endianness, but the memory core hands them in target endianness. 985 * For example, PPC is always treated as big-endian even if running 986 * on KVM and on PPC64LE. Correct here. 987 */ 988 switch (size) { 989 case 2: 990 val = bswap16(val); 991 break; 992 case 4: 993 val = bswap32(val); 994 break; 995 } 996 #endif 997 return val; 998 } 999 1000 static int kvm_set_ioeventfd_mmio(int fd, hwaddr addr, uint32_t val, 1001 bool assign, uint32_t size, bool datamatch) 1002 { 1003 int ret; 1004 struct kvm_ioeventfd iofd = { 1005 .datamatch = datamatch ? adjust_ioeventfd_endianness(val, size) : 0, 1006 .addr = addr, 1007 .len = size, 1008 .flags = 0, 1009 .fd = fd, 1010 }; 1011 1012 trace_kvm_set_ioeventfd_mmio(fd, (uint64_t)addr, val, assign, size, 1013 datamatch); 1014 if (!kvm_enabled()) { 1015 return -ENOSYS; 1016 } 1017 1018 if (datamatch) { 1019 iofd.flags |= KVM_IOEVENTFD_FLAG_DATAMATCH; 1020 } 1021 if (!assign) { 1022 iofd.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN; 1023 } 1024 1025 ret = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &iofd); 1026 1027 if (ret < 0) { 1028 return -errno; 1029 } 1030 1031 return 0; 1032 } 1033 1034 static int kvm_set_ioeventfd_pio(int fd, uint16_t addr, uint16_t val, 1035 bool assign, uint32_t size, bool datamatch) 1036 { 1037 struct kvm_ioeventfd kick = { 1038 .datamatch = datamatch ? adjust_ioeventfd_endianness(val, size) : 0, 1039 .addr = addr, 1040 .flags = KVM_IOEVENTFD_FLAG_PIO, 1041 .len = size, 1042 .fd = fd, 1043 }; 1044 int r; 1045 trace_kvm_set_ioeventfd_pio(fd, addr, val, assign, size, datamatch); 1046 if (!kvm_enabled()) { 1047 return -ENOSYS; 1048 } 1049 if (datamatch) { 1050 kick.flags |= KVM_IOEVENTFD_FLAG_DATAMATCH; 1051 } 1052 if (!assign) { 1053 kick.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN; 1054 } 1055 r = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &kick); 1056 if (r < 0) { 1057 return r; 1058 } 1059 return 0; 1060 } 1061 1062 1063 static int kvm_check_many_ioeventfds(void) 1064 { 1065 /* Userspace can use ioeventfd for io notification. This requires a host 1066 * that supports eventfd(2) and an I/O thread; since eventfd does not 1067 * support SIGIO it cannot interrupt the vcpu. 1068 * 1069 * Older kernels have a 6 device limit on the KVM io bus. Find out so we 1070 * can avoid creating too many ioeventfds. 1071 */ 1072 #if defined(CONFIG_EVENTFD) 1073 int ioeventfds[7]; 1074 int i, ret = 0; 1075 for (i = 0; i < ARRAY_SIZE(ioeventfds); i++) { 1076 ioeventfds[i] = eventfd(0, EFD_CLOEXEC); 1077 if (ioeventfds[i] < 0) { 1078 break; 1079 } 1080 ret = kvm_set_ioeventfd_pio(ioeventfds[i], 0, i, true, 2, true); 1081 if (ret < 0) { 1082 close(ioeventfds[i]); 1083 break; 1084 } 1085 } 1086 1087 /* Decide whether many devices are supported or not */ 1088 ret = i == ARRAY_SIZE(ioeventfds); 1089 1090 while (i-- > 0) { 1091 kvm_set_ioeventfd_pio(ioeventfds[i], 0, i, false, 2, true); 1092 close(ioeventfds[i]); 1093 } 1094 return ret; 1095 #else 1096 return 0; 1097 #endif 1098 } 1099 1100 static const KVMCapabilityInfo * 1101 kvm_check_extension_list(KVMState *s, const KVMCapabilityInfo *list) 1102 { 1103 while (list->name) { 1104 if (!kvm_check_extension(s, list->value)) { 1105 return list; 1106 } 1107 list++; 1108 } 1109 return NULL; 1110 } 1111 1112 void kvm_set_max_memslot_size(hwaddr max_slot_size) 1113 { 1114 g_assert( 1115 ROUND_UP(max_slot_size, qemu_real_host_page_size) == max_slot_size 1116 ); 1117 kvm_max_slot_size = max_slot_size; 1118 } 1119 1120 static void kvm_set_phys_mem(KVMMemoryListener *kml, 1121 MemoryRegionSection *section, bool add) 1122 { 1123 KVMSlot *mem; 1124 int err; 1125 MemoryRegion *mr = section->mr; 1126 bool writeable = !mr->readonly && !mr->rom_device; 1127 hwaddr start_addr, size, slot_size; 1128 void *ram; 1129 1130 if (!memory_region_is_ram(mr)) { 1131 if (writeable || !kvm_readonly_mem_allowed) { 1132 return; 1133 } else if (!mr->romd_mode) { 1134 /* If the memory device is not in romd_mode, then we actually want 1135 * to remove the kvm memory slot so all accesses will trap. */ 1136 add = false; 1137 } 1138 } 1139 1140 size = kvm_align_section(section, &start_addr); 1141 if (!size) { 1142 return; 1143 } 1144 1145 /* use aligned delta to align the ram address */ 1146 ram = memory_region_get_ram_ptr(mr) + section->offset_within_region + 1147 (start_addr - section->offset_within_address_space); 1148 1149 kvm_slots_lock(kml); 1150 1151 if (!add) { 1152 do { 1153 slot_size = MIN(kvm_max_slot_size, size); 1154 mem = kvm_lookup_matching_slot(kml, start_addr, slot_size); 1155 if (!mem) { 1156 goto out; 1157 } 1158 if (mem->flags & KVM_MEM_LOG_DIRTY_PAGES) { 1159 kvm_physical_sync_dirty_bitmap(kml, section); 1160 } 1161 1162 /* unregister the slot */ 1163 g_free(mem->dirty_bmap); 1164 mem->dirty_bmap = NULL; 1165 mem->memory_size = 0; 1166 mem->flags = 0; 1167 err = kvm_set_user_memory_region(kml, mem, false); 1168 if (err) { 1169 fprintf(stderr, "%s: error unregistering slot: %s\n", 1170 __func__, strerror(-err)); 1171 abort(); 1172 } 1173 start_addr += slot_size; 1174 size -= slot_size; 1175 } while (size); 1176 goto out; 1177 } 1178 1179 /* register the new slot */ 1180 do { 1181 slot_size = MIN(kvm_max_slot_size, size); 1182 mem = kvm_alloc_slot(kml); 1183 mem->memory_size = slot_size; 1184 mem->start_addr = start_addr; 1185 mem->ram = ram; 1186 mem->flags = kvm_mem_flags(mr); 1187 1188 if (mem->flags & KVM_MEM_LOG_DIRTY_PAGES) { 1189 /* 1190 * Reallocate the bmap; it means it doesn't disappear in 1191 * middle of a migrate. 1192 */ 1193 kvm_memslot_init_dirty_bitmap(mem); 1194 } 1195 err = kvm_set_user_memory_region(kml, mem, true); 1196 if (err) { 1197 fprintf(stderr, "%s: error registering slot: %s\n", __func__, 1198 strerror(-err)); 1199 abort(); 1200 } 1201 start_addr += slot_size; 1202 ram += slot_size; 1203 size -= slot_size; 1204 } while (size); 1205 1206 out: 1207 kvm_slots_unlock(kml); 1208 } 1209 1210 static void kvm_region_add(MemoryListener *listener, 1211 MemoryRegionSection *section) 1212 { 1213 KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener); 1214 1215 memory_region_ref(section->mr); 1216 kvm_set_phys_mem(kml, section, true); 1217 } 1218 1219 static void kvm_region_del(MemoryListener *listener, 1220 MemoryRegionSection *section) 1221 { 1222 KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener); 1223 1224 kvm_set_phys_mem(kml, section, false); 1225 memory_region_unref(section->mr); 1226 } 1227 1228 static void kvm_log_sync(MemoryListener *listener, 1229 MemoryRegionSection *section) 1230 { 1231 KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener); 1232 int r; 1233 1234 kvm_slots_lock(kml); 1235 r = kvm_physical_sync_dirty_bitmap(kml, section); 1236 kvm_slots_unlock(kml); 1237 if (r < 0) { 1238 abort(); 1239 } 1240 } 1241 1242 static void kvm_log_clear(MemoryListener *listener, 1243 MemoryRegionSection *section) 1244 { 1245 KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener); 1246 int r; 1247 1248 r = kvm_physical_log_clear(kml, section); 1249 if (r < 0) { 1250 error_report_once("%s: kvm log clear failed: mr=%s " 1251 "offset=%"HWADDR_PRIx" size=%"PRIx64, __func__, 1252 section->mr->name, section->offset_within_region, 1253 int128_get64(section->size)); 1254 abort(); 1255 } 1256 } 1257 1258 static void kvm_mem_ioeventfd_add(MemoryListener *listener, 1259 MemoryRegionSection *section, 1260 bool match_data, uint64_t data, 1261 EventNotifier *e) 1262 { 1263 int fd = event_notifier_get_fd(e); 1264 int r; 1265 1266 r = kvm_set_ioeventfd_mmio(fd, section->offset_within_address_space, 1267 data, true, int128_get64(section->size), 1268 match_data); 1269 if (r < 0) { 1270 fprintf(stderr, "%s: error adding ioeventfd: %s (%d)\n", 1271 __func__, strerror(-r), -r); 1272 abort(); 1273 } 1274 } 1275 1276 static void kvm_mem_ioeventfd_del(MemoryListener *listener, 1277 MemoryRegionSection *section, 1278 bool match_data, uint64_t data, 1279 EventNotifier *e) 1280 { 1281 int fd = event_notifier_get_fd(e); 1282 int r; 1283 1284 r = kvm_set_ioeventfd_mmio(fd, section->offset_within_address_space, 1285 data, false, int128_get64(section->size), 1286 match_data); 1287 if (r < 0) { 1288 fprintf(stderr, "%s: error deleting ioeventfd: %s (%d)\n", 1289 __func__, strerror(-r), -r); 1290 abort(); 1291 } 1292 } 1293 1294 static void kvm_io_ioeventfd_add(MemoryListener *listener, 1295 MemoryRegionSection *section, 1296 bool match_data, uint64_t data, 1297 EventNotifier *e) 1298 { 1299 int fd = event_notifier_get_fd(e); 1300 int r; 1301 1302 r = kvm_set_ioeventfd_pio(fd, section->offset_within_address_space, 1303 data, true, int128_get64(section->size), 1304 match_data); 1305 if (r < 0) { 1306 fprintf(stderr, "%s: error adding ioeventfd: %s (%d)\n", 1307 __func__, strerror(-r), -r); 1308 abort(); 1309 } 1310 } 1311 1312 static void kvm_io_ioeventfd_del(MemoryListener *listener, 1313 MemoryRegionSection *section, 1314 bool match_data, uint64_t data, 1315 EventNotifier *e) 1316 1317 { 1318 int fd = event_notifier_get_fd(e); 1319 int r; 1320 1321 r = kvm_set_ioeventfd_pio(fd, section->offset_within_address_space, 1322 data, false, int128_get64(section->size), 1323 match_data); 1324 if (r < 0) { 1325 fprintf(stderr, "%s: error deleting ioeventfd: %s (%d)\n", 1326 __func__, strerror(-r), -r); 1327 abort(); 1328 } 1329 } 1330 1331 void kvm_memory_listener_register(KVMState *s, KVMMemoryListener *kml, 1332 AddressSpace *as, int as_id) 1333 { 1334 int i; 1335 1336 qemu_mutex_init(&kml->slots_lock); 1337 kml->slots = g_malloc0(s->nr_slots * sizeof(KVMSlot)); 1338 kml->as_id = as_id; 1339 1340 for (i = 0; i < s->nr_slots; i++) { 1341 kml->slots[i].slot = i; 1342 } 1343 1344 kml->listener.region_add = kvm_region_add; 1345 kml->listener.region_del = kvm_region_del; 1346 kml->listener.log_start = kvm_log_start; 1347 kml->listener.log_stop = kvm_log_stop; 1348 kml->listener.log_sync = kvm_log_sync; 1349 kml->listener.log_clear = kvm_log_clear; 1350 kml->listener.priority = 10; 1351 1352 memory_listener_register(&kml->listener, as); 1353 1354 for (i = 0; i < s->nr_as; ++i) { 1355 if (!s->as[i].as) { 1356 s->as[i].as = as; 1357 s->as[i].ml = kml; 1358 break; 1359 } 1360 } 1361 } 1362 1363 static MemoryListener kvm_io_listener = { 1364 .eventfd_add = kvm_io_ioeventfd_add, 1365 .eventfd_del = kvm_io_ioeventfd_del, 1366 .priority = 10, 1367 }; 1368 1369 int kvm_set_irq(KVMState *s, int irq, int level) 1370 { 1371 struct kvm_irq_level event; 1372 int ret; 1373 1374 assert(kvm_async_interrupts_enabled()); 1375 1376 event.level = level; 1377 event.irq = irq; 1378 ret = kvm_vm_ioctl(s, s->irq_set_ioctl, &event); 1379 if (ret < 0) { 1380 perror("kvm_set_irq"); 1381 abort(); 1382 } 1383 1384 return (s->irq_set_ioctl == KVM_IRQ_LINE) ? 1 : event.status; 1385 } 1386 1387 #ifdef KVM_CAP_IRQ_ROUTING 1388 typedef struct KVMMSIRoute { 1389 struct kvm_irq_routing_entry kroute; 1390 QTAILQ_ENTRY(KVMMSIRoute) entry; 1391 } KVMMSIRoute; 1392 1393 static void set_gsi(KVMState *s, unsigned int gsi) 1394 { 1395 set_bit(gsi, s->used_gsi_bitmap); 1396 } 1397 1398 static void clear_gsi(KVMState *s, unsigned int gsi) 1399 { 1400 clear_bit(gsi, s->used_gsi_bitmap); 1401 } 1402 1403 void kvm_init_irq_routing(KVMState *s) 1404 { 1405 int gsi_count, i; 1406 1407 gsi_count = kvm_check_extension(s, KVM_CAP_IRQ_ROUTING) - 1; 1408 if (gsi_count > 0) { 1409 /* Round up so we can search ints using ffs */ 1410 s->used_gsi_bitmap = bitmap_new(gsi_count); 1411 s->gsi_count = gsi_count; 1412 } 1413 1414 s->irq_routes = g_malloc0(sizeof(*s->irq_routes)); 1415 s->nr_allocated_irq_routes = 0; 1416 1417 if (!kvm_direct_msi_allowed) { 1418 for (i = 0; i < KVM_MSI_HASHTAB_SIZE; i++) { 1419 QTAILQ_INIT(&s->msi_hashtab[i]); 1420 } 1421 } 1422 1423 kvm_arch_init_irq_routing(s); 1424 } 1425 1426 void kvm_irqchip_commit_routes(KVMState *s) 1427 { 1428 int ret; 1429 1430 if (kvm_gsi_direct_mapping()) { 1431 return; 1432 } 1433 1434 if (!kvm_gsi_routing_enabled()) { 1435 return; 1436 } 1437 1438 s->irq_routes->flags = 0; 1439 trace_kvm_irqchip_commit_routes(); 1440 ret = kvm_vm_ioctl(s, KVM_SET_GSI_ROUTING, s->irq_routes); 1441 assert(ret == 0); 1442 } 1443 1444 static void kvm_add_routing_entry(KVMState *s, 1445 struct kvm_irq_routing_entry *entry) 1446 { 1447 struct kvm_irq_routing_entry *new; 1448 int n, size; 1449 1450 if (s->irq_routes->nr == s->nr_allocated_irq_routes) { 1451 n = s->nr_allocated_irq_routes * 2; 1452 if (n < 64) { 1453 n = 64; 1454 } 1455 size = sizeof(struct kvm_irq_routing); 1456 size += n * sizeof(*new); 1457 s->irq_routes = g_realloc(s->irq_routes, size); 1458 s->nr_allocated_irq_routes = n; 1459 } 1460 n = s->irq_routes->nr++; 1461 new = &s->irq_routes->entries[n]; 1462 1463 *new = *entry; 1464 1465 set_gsi(s, entry->gsi); 1466 } 1467 1468 static int kvm_update_routing_entry(KVMState *s, 1469 struct kvm_irq_routing_entry *new_entry) 1470 { 1471 struct kvm_irq_routing_entry *entry; 1472 int n; 1473 1474 for (n = 0; n < s->irq_routes->nr; n++) { 1475 entry = &s->irq_routes->entries[n]; 1476 if (entry->gsi != new_entry->gsi) { 1477 continue; 1478 } 1479 1480 if(!memcmp(entry, new_entry, sizeof *entry)) { 1481 return 0; 1482 } 1483 1484 *entry = *new_entry; 1485 1486 return 0; 1487 } 1488 1489 return -ESRCH; 1490 } 1491 1492 void kvm_irqchip_add_irq_route(KVMState *s, int irq, int irqchip, int pin) 1493 { 1494 struct kvm_irq_routing_entry e = {}; 1495 1496 assert(pin < s->gsi_count); 1497 1498 e.gsi = irq; 1499 e.type = KVM_IRQ_ROUTING_IRQCHIP; 1500 e.flags = 0; 1501 e.u.irqchip.irqchip = irqchip; 1502 e.u.irqchip.pin = pin; 1503 kvm_add_routing_entry(s, &e); 1504 } 1505 1506 void kvm_irqchip_release_virq(KVMState *s, int virq) 1507 { 1508 struct kvm_irq_routing_entry *e; 1509 int i; 1510 1511 if (kvm_gsi_direct_mapping()) { 1512 return; 1513 } 1514 1515 for (i = 0; i < s->irq_routes->nr; i++) { 1516 e = &s->irq_routes->entries[i]; 1517 if (e->gsi == virq) { 1518 s->irq_routes->nr--; 1519 *e = s->irq_routes->entries[s->irq_routes->nr]; 1520 } 1521 } 1522 clear_gsi(s, virq); 1523 kvm_arch_release_virq_post(virq); 1524 trace_kvm_irqchip_release_virq(virq); 1525 } 1526 1527 void kvm_irqchip_add_change_notifier(Notifier *n) 1528 { 1529 notifier_list_add(&kvm_irqchip_change_notifiers, n); 1530 } 1531 1532 void kvm_irqchip_remove_change_notifier(Notifier *n) 1533 { 1534 notifier_remove(n); 1535 } 1536 1537 void kvm_irqchip_change_notify(void) 1538 { 1539 notifier_list_notify(&kvm_irqchip_change_notifiers, NULL); 1540 } 1541 1542 static unsigned int kvm_hash_msi(uint32_t data) 1543 { 1544 /* This is optimized for IA32 MSI layout. However, no other arch shall 1545 * repeat the mistake of not providing a direct MSI injection API. */ 1546 return data & 0xff; 1547 } 1548 1549 static void kvm_flush_dynamic_msi_routes(KVMState *s) 1550 { 1551 KVMMSIRoute *route, *next; 1552 unsigned int hash; 1553 1554 for (hash = 0; hash < KVM_MSI_HASHTAB_SIZE; hash++) { 1555 QTAILQ_FOREACH_SAFE(route, &s->msi_hashtab[hash], entry, next) { 1556 kvm_irqchip_release_virq(s, route->kroute.gsi); 1557 QTAILQ_REMOVE(&s->msi_hashtab[hash], route, entry); 1558 g_free(route); 1559 } 1560 } 1561 } 1562 1563 static int kvm_irqchip_get_virq(KVMState *s) 1564 { 1565 int next_virq; 1566 1567 /* 1568 * PIC and IOAPIC share the first 16 GSI numbers, thus the available 1569 * GSI numbers are more than the number of IRQ route. Allocating a GSI 1570 * number can succeed even though a new route entry cannot be added. 1571 * When this happens, flush dynamic MSI entries to free IRQ route entries. 1572 */ 1573 if (!kvm_direct_msi_allowed && s->irq_routes->nr == s->gsi_count) { 1574 kvm_flush_dynamic_msi_routes(s); 1575 } 1576 1577 /* Return the lowest unused GSI in the bitmap */ 1578 next_virq = find_first_zero_bit(s->used_gsi_bitmap, s->gsi_count); 1579 if (next_virq >= s->gsi_count) { 1580 return -ENOSPC; 1581 } else { 1582 return next_virq; 1583 } 1584 } 1585 1586 static KVMMSIRoute *kvm_lookup_msi_route(KVMState *s, MSIMessage msg) 1587 { 1588 unsigned int hash = kvm_hash_msi(msg.data); 1589 KVMMSIRoute *route; 1590 1591 QTAILQ_FOREACH(route, &s->msi_hashtab[hash], entry) { 1592 if (route->kroute.u.msi.address_lo == (uint32_t)msg.address && 1593 route->kroute.u.msi.address_hi == (msg.address >> 32) && 1594 route->kroute.u.msi.data == le32_to_cpu(msg.data)) { 1595 return route; 1596 } 1597 } 1598 return NULL; 1599 } 1600 1601 int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg) 1602 { 1603 struct kvm_msi msi; 1604 KVMMSIRoute *route; 1605 1606 if (kvm_direct_msi_allowed) { 1607 msi.address_lo = (uint32_t)msg.address; 1608 msi.address_hi = msg.address >> 32; 1609 msi.data = le32_to_cpu(msg.data); 1610 msi.flags = 0; 1611 memset(msi.pad, 0, sizeof(msi.pad)); 1612 1613 return kvm_vm_ioctl(s, KVM_SIGNAL_MSI, &msi); 1614 } 1615 1616 route = kvm_lookup_msi_route(s, msg); 1617 if (!route) { 1618 int virq; 1619 1620 virq = kvm_irqchip_get_virq(s); 1621 if (virq < 0) { 1622 return virq; 1623 } 1624 1625 route = g_malloc0(sizeof(KVMMSIRoute)); 1626 route->kroute.gsi = virq; 1627 route->kroute.type = KVM_IRQ_ROUTING_MSI; 1628 route->kroute.flags = 0; 1629 route->kroute.u.msi.address_lo = (uint32_t)msg.address; 1630 route->kroute.u.msi.address_hi = msg.address >> 32; 1631 route->kroute.u.msi.data = le32_to_cpu(msg.data); 1632 1633 kvm_add_routing_entry(s, &route->kroute); 1634 kvm_irqchip_commit_routes(s); 1635 1636 QTAILQ_INSERT_TAIL(&s->msi_hashtab[kvm_hash_msi(msg.data)], route, 1637 entry); 1638 } 1639 1640 assert(route->kroute.type == KVM_IRQ_ROUTING_MSI); 1641 1642 return kvm_set_irq(s, route->kroute.gsi, 1); 1643 } 1644 1645 int kvm_irqchip_add_msi_route(KVMState *s, int vector, PCIDevice *dev) 1646 { 1647 struct kvm_irq_routing_entry kroute = {}; 1648 int virq; 1649 MSIMessage msg = {0, 0}; 1650 1651 if (pci_available && dev) { 1652 msg = pci_get_msi_message(dev, vector); 1653 } 1654 1655 if (kvm_gsi_direct_mapping()) { 1656 return kvm_arch_msi_data_to_gsi(msg.data); 1657 } 1658 1659 if (!kvm_gsi_routing_enabled()) { 1660 return -ENOSYS; 1661 } 1662 1663 virq = kvm_irqchip_get_virq(s); 1664 if (virq < 0) { 1665 return virq; 1666 } 1667 1668 kroute.gsi = virq; 1669 kroute.type = KVM_IRQ_ROUTING_MSI; 1670 kroute.flags = 0; 1671 kroute.u.msi.address_lo = (uint32_t)msg.address; 1672 kroute.u.msi.address_hi = msg.address >> 32; 1673 kroute.u.msi.data = le32_to_cpu(msg.data); 1674 if (pci_available && kvm_msi_devid_required()) { 1675 kroute.flags = KVM_MSI_VALID_DEVID; 1676 kroute.u.msi.devid = pci_requester_id(dev); 1677 } 1678 if (kvm_arch_fixup_msi_route(&kroute, msg.address, msg.data, dev)) { 1679 kvm_irqchip_release_virq(s, virq); 1680 return -EINVAL; 1681 } 1682 1683 trace_kvm_irqchip_add_msi_route(dev ? dev->name : (char *)"N/A", 1684 vector, virq); 1685 1686 kvm_add_routing_entry(s, &kroute); 1687 kvm_arch_add_msi_route_post(&kroute, vector, dev); 1688 kvm_irqchip_commit_routes(s); 1689 1690 return virq; 1691 } 1692 1693 int kvm_irqchip_update_msi_route(KVMState *s, int virq, MSIMessage msg, 1694 PCIDevice *dev) 1695 { 1696 struct kvm_irq_routing_entry kroute = {}; 1697 1698 if (kvm_gsi_direct_mapping()) { 1699 return 0; 1700 } 1701 1702 if (!kvm_irqchip_in_kernel()) { 1703 return -ENOSYS; 1704 } 1705 1706 kroute.gsi = virq; 1707 kroute.type = KVM_IRQ_ROUTING_MSI; 1708 kroute.flags = 0; 1709 kroute.u.msi.address_lo = (uint32_t)msg.address; 1710 kroute.u.msi.address_hi = msg.address >> 32; 1711 kroute.u.msi.data = le32_to_cpu(msg.data); 1712 if (pci_available && kvm_msi_devid_required()) { 1713 kroute.flags = KVM_MSI_VALID_DEVID; 1714 kroute.u.msi.devid = pci_requester_id(dev); 1715 } 1716 if (kvm_arch_fixup_msi_route(&kroute, msg.address, msg.data, dev)) { 1717 return -EINVAL; 1718 } 1719 1720 trace_kvm_irqchip_update_msi_route(virq); 1721 1722 return kvm_update_routing_entry(s, &kroute); 1723 } 1724 1725 static int kvm_irqchip_assign_irqfd(KVMState *s, EventNotifier *event, 1726 EventNotifier *resample, int virq, 1727 bool assign) 1728 { 1729 int fd = event_notifier_get_fd(event); 1730 int rfd = resample ? event_notifier_get_fd(resample) : -1; 1731 1732 struct kvm_irqfd irqfd = { 1733 .fd = fd, 1734 .gsi = virq, 1735 .flags = assign ? 0 : KVM_IRQFD_FLAG_DEASSIGN, 1736 }; 1737 1738 if (rfd != -1) { 1739 assert(assign); 1740 if (kvm_irqchip_is_split()) { 1741 /* 1742 * When the slow irqchip (e.g. IOAPIC) is in the 1743 * userspace, KVM kernel resamplefd will not work because 1744 * the EOI of the interrupt will be delivered to userspace 1745 * instead, so the KVM kernel resamplefd kick will be 1746 * skipped. The userspace here mimics what the kernel 1747 * provides with resamplefd, remember the resamplefd and 1748 * kick it when we receive EOI of this IRQ. 1749 * 1750 * This is hackery because IOAPIC is mostly bypassed 1751 * (except EOI broadcasts) when irqfd is used. However 1752 * this can bring much performance back for split irqchip 1753 * with INTx IRQs (for VFIO, this gives 93% perf of the 1754 * full fast path, which is 46% perf boost comparing to 1755 * the INTx slow path). 1756 */ 1757 kvm_resample_fd_insert(virq, resample); 1758 } else { 1759 irqfd.flags |= KVM_IRQFD_FLAG_RESAMPLE; 1760 irqfd.resamplefd = rfd; 1761 } 1762 } else if (!assign) { 1763 if (kvm_irqchip_is_split()) { 1764 kvm_resample_fd_remove(virq); 1765 } 1766 } 1767 1768 if (!kvm_irqfds_enabled()) { 1769 return -ENOSYS; 1770 } 1771 1772 return kvm_vm_ioctl(s, KVM_IRQFD, &irqfd); 1773 } 1774 1775 int kvm_irqchip_add_adapter_route(KVMState *s, AdapterInfo *adapter) 1776 { 1777 struct kvm_irq_routing_entry kroute = {}; 1778 int virq; 1779 1780 if (!kvm_gsi_routing_enabled()) { 1781 return -ENOSYS; 1782 } 1783 1784 virq = kvm_irqchip_get_virq(s); 1785 if (virq < 0) { 1786 return virq; 1787 } 1788 1789 kroute.gsi = virq; 1790 kroute.type = KVM_IRQ_ROUTING_S390_ADAPTER; 1791 kroute.flags = 0; 1792 kroute.u.adapter.summary_addr = adapter->summary_addr; 1793 kroute.u.adapter.ind_addr = adapter->ind_addr; 1794 kroute.u.adapter.summary_offset = adapter->summary_offset; 1795 kroute.u.adapter.ind_offset = adapter->ind_offset; 1796 kroute.u.adapter.adapter_id = adapter->adapter_id; 1797 1798 kvm_add_routing_entry(s, &kroute); 1799 1800 return virq; 1801 } 1802 1803 int kvm_irqchip_add_hv_sint_route(KVMState *s, uint32_t vcpu, uint32_t sint) 1804 { 1805 struct kvm_irq_routing_entry kroute = {}; 1806 int virq; 1807 1808 if (!kvm_gsi_routing_enabled()) { 1809 return -ENOSYS; 1810 } 1811 if (!kvm_check_extension(s, KVM_CAP_HYPERV_SYNIC)) { 1812 return -ENOSYS; 1813 } 1814 virq = kvm_irqchip_get_virq(s); 1815 if (virq < 0) { 1816 return virq; 1817 } 1818 1819 kroute.gsi = virq; 1820 kroute.type = KVM_IRQ_ROUTING_HV_SINT; 1821 kroute.flags = 0; 1822 kroute.u.hv_sint.vcpu = vcpu; 1823 kroute.u.hv_sint.sint = sint; 1824 1825 kvm_add_routing_entry(s, &kroute); 1826 kvm_irqchip_commit_routes(s); 1827 1828 return virq; 1829 } 1830 1831 #else /* !KVM_CAP_IRQ_ROUTING */ 1832 1833 void kvm_init_irq_routing(KVMState *s) 1834 { 1835 } 1836 1837 void kvm_irqchip_release_virq(KVMState *s, int virq) 1838 { 1839 } 1840 1841 int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg) 1842 { 1843 abort(); 1844 } 1845 1846 int kvm_irqchip_add_msi_route(KVMState *s, int vector, PCIDevice *dev) 1847 { 1848 return -ENOSYS; 1849 } 1850 1851 int kvm_irqchip_add_adapter_route(KVMState *s, AdapterInfo *adapter) 1852 { 1853 return -ENOSYS; 1854 } 1855 1856 int kvm_irqchip_add_hv_sint_route(KVMState *s, uint32_t vcpu, uint32_t sint) 1857 { 1858 return -ENOSYS; 1859 } 1860 1861 static int kvm_irqchip_assign_irqfd(KVMState *s, EventNotifier *event, 1862 EventNotifier *resample, int virq, 1863 bool assign) 1864 { 1865 abort(); 1866 } 1867 1868 int kvm_irqchip_update_msi_route(KVMState *s, int virq, MSIMessage msg) 1869 { 1870 return -ENOSYS; 1871 } 1872 #endif /* !KVM_CAP_IRQ_ROUTING */ 1873 1874 int kvm_irqchip_add_irqfd_notifier_gsi(KVMState *s, EventNotifier *n, 1875 EventNotifier *rn, int virq) 1876 { 1877 return kvm_irqchip_assign_irqfd(s, n, rn, virq, true); 1878 } 1879 1880 int kvm_irqchip_remove_irqfd_notifier_gsi(KVMState *s, EventNotifier *n, 1881 int virq) 1882 { 1883 return kvm_irqchip_assign_irqfd(s, n, NULL, virq, false); 1884 } 1885 1886 int kvm_irqchip_add_irqfd_notifier(KVMState *s, EventNotifier *n, 1887 EventNotifier *rn, qemu_irq irq) 1888 { 1889 gpointer key, gsi; 1890 gboolean found = g_hash_table_lookup_extended(s->gsimap, irq, &key, &gsi); 1891 1892 if (!found) { 1893 return -ENXIO; 1894 } 1895 return kvm_irqchip_add_irqfd_notifier_gsi(s, n, rn, GPOINTER_TO_INT(gsi)); 1896 } 1897 1898 int kvm_irqchip_remove_irqfd_notifier(KVMState *s, EventNotifier *n, 1899 qemu_irq irq) 1900 { 1901 gpointer key, gsi; 1902 gboolean found = g_hash_table_lookup_extended(s->gsimap, irq, &key, &gsi); 1903 1904 if (!found) { 1905 return -ENXIO; 1906 } 1907 return kvm_irqchip_remove_irqfd_notifier_gsi(s, n, GPOINTER_TO_INT(gsi)); 1908 } 1909 1910 void kvm_irqchip_set_qemuirq_gsi(KVMState *s, qemu_irq irq, int gsi) 1911 { 1912 g_hash_table_insert(s->gsimap, irq, GINT_TO_POINTER(gsi)); 1913 } 1914 1915 static void kvm_irqchip_create(KVMState *s) 1916 { 1917 int ret; 1918 1919 assert(s->kernel_irqchip_split != ON_OFF_AUTO_AUTO); 1920 if (kvm_check_extension(s, KVM_CAP_IRQCHIP)) { 1921 ; 1922 } else if (kvm_check_extension(s, KVM_CAP_S390_IRQCHIP)) { 1923 ret = kvm_vm_enable_cap(s, KVM_CAP_S390_IRQCHIP, 0); 1924 if (ret < 0) { 1925 fprintf(stderr, "Enable kernel irqchip failed: %s\n", strerror(-ret)); 1926 exit(1); 1927 } 1928 } else { 1929 return; 1930 } 1931 1932 /* First probe and see if there's a arch-specific hook to create the 1933 * in-kernel irqchip for us */ 1934 ret = kvm_arch_irqchip_create(s); 1935 if (ret == 0) { 1936 if (s->kernel_irqchip_split == ON_OFF_AUTO_ON) { 1937 perror("Split IRQ chip mode not supported."); 1938 exit(1); 1939 } else { 1940 ret = kvm_vm_ioctl(s, KVM_CREATE_IRQCHIP); 1941 } 1942 } 1943 if (ret < 0) { 1944 fprintf(stderr, "Create kernel irqchip failed: %s\n", strerror(-ret)); 1945 exit(1); 1946 } 1947 1948 kvm_kernel_irqchip = true; 1949 /* If we have an in-kernel IRQ chip then we must have asynchronous 1950 * interrupt delivery (though the reverse is not necessarily true) 1951 */ 1952 kvm_async_interrupts_allowed = true; 1953 kvm_halt_in_kernel_allowed = true; 1954 1955 kvm_init_irq_routing(s); 1956 1957 s->gsimap = g_hash_table_new(g_direct_hash, g_direct_equal); 1958 } 1959 1960 /* Find number of supported CPUs using the recommended 1961 * procedure from the kernel API documentation to cope with 1962 * older kernels that may be missing capabilities. 1963 */ 1964 static int kvm_recommended_vcpus(KVMState *s) 1965 { 1966 int ret = kvm_vm_check_extension(s, KVM_CAP_NR_VCPUS); 1967 return (ret) ? ret : 4; 1968 } 1969 1970 static int kvm_max_vcpus(KVMState *s) 1971 { 1972 int ret = kvm_check_extension(s, KVM_CAP_MAX_VCPUS); 1973 return (ret) ? ret : kvm_recommended_vcpus(s); 1974 } 1975 1976 static int kvm_max_vcpu_id(KVMState *s) 1977 { 1978 int ret = kvm_check_extension(s, KVM_CAP_MAX_VCPU_ID); 1979 return (ret) ? ret : kvm_max_vcpus(s); 1980 } 1981 1982 bool kvm_vcpu_id_is_valid(int vcpu_id) 1983 { 1984 KVMState *s = KVM_STATE(current_accel()); 1985 return vcpu_id >= 0 && vcpu_id < kvm_max_vcpu_id(s); 1986 } 1987 1988 static int kvm_init(MachineState *ms) 1989 { 1990 MachineClass *mc = MACHINE_GET_CLASS(ms); 1991 static const char upgrade_note[] = 1992 "Please upgrade to at least kernel 2.6.29 or recent kvm-kmod\n" 1993 "(see http://sourceforge.net/projects/kvm).\n"; 1994 struct { 1995 const char *name; 1996 int num; 1997 } num_cpus[] = { 1998 { "SMP", ms->smp.cpus }, 1999 { "hotpluggable", ms->smp.max_cpus }, 2000 { NULL, } 2001 }, *nc = num_cpus; 2002 int soft_vcpus_limit, hard_vcpus_limit; 2003 KVMState *s; 2004 const KVMCapabilityInfo *missing_cap; 2005 int ret; 2006 int type = 0; 2007 const char *kvm_type; 2008 uint64_t dirty_log_manual_caps; 2009 2010 s = KVM_STATE(ms->accelerator); 2011 2012 /* 2013 * On systems where the kernel can support different base page 2014 * sizes, host page size may be different from TARGET_PAGE_SIZE, 2015 * even with KVM. TARGET_PAGE_SIZE is assumed to be the minimum 2016 * page size for the system though. 2017 */ 2018 assert(TARGET_PAGE_SIZE <= qemu_real_host_page_size); 2019 2020 s->sigmask_len = 8; 2021 2022 #ifdef KVM_CAP_SET_GUEST_DEBUG 2023 QTAILQ_INIT(&s->kvm_sw_breakpoints); 2024 #endif 2025 QLIST_INIT(&s->kvm_parked_vcpus); 2026 s->vmfd = -1; 2027 s->fd = qemu_open_old("/dev/kvm", O_RDWR); 2028 if (s->fd == -1) { 2029 fprintf(stderr, "Could not access KVM kernel module: %m\n"); 2030 ret = -errno; 2031 goto err; 2032 } 2033 2034 ret = kvm_ioctl(s, KVM_GET_API_VERSION, 0); 2035 if (ret < KVM_API_VERSION) { 2036 if (ret >= 0) { 2037 ret = -EINVAL; 2038 } 2039 fprintf(stderr, "kvm version too old\n"); 2040 goto err; 2041 } 2042 2043 if (ret > KVM_API_VERSION) { 2044 ret = -EINVAL; 2045 fprintf(stderr, "kvm version not supported\n"); 2046 goto err; 2047 } 2048 2049 kvm_immediate_exit = kvm_check_extension(s, KVM_CAP_IMMEDIATE_EXIT); 2050 s->nr_slots = kvm_check_extension(s, KVM_CAP_NR_MEMSLOTS); 2051 2052 /* If unspecified, use the default value */ 2053 if (!s->nr_slots) { 2054 s->nr_slots = 32; 2055 } 2056 2057 s->nr_as = kvm_check_extension(s, KVM_CAP_MULTI_ADDRESS_SPACE); 2058 if (s->nr_as <= 1) { 2059 s->nr_as = 1; 2060 } 2061 s->as = g_new0(struct KVMAs, s->nr_as); 2062 2063 kvm_type = qemu_opt_get(qemu_get_machine_opts(), "kvm-type"); 2064 if (mc->kvm_type) { 2065 type = mc->kvm_type(ms, kvm_type); 2066 } else if (kvm_type) { 2067 ret = -EINVAL; 2068 fprintf(stderr, "Invalid argument kvm-type=%s\n", kvm_type); 2069 goto err; 2070 } 2071 2072 do { 2073 ret = kvm_ioctl(s, KVM_CREATE_VM, type); 2074 } while (ret == -EINTR); 2075 2076 if (ret < 0) { 2077 fprintf(stderr, "ioctl(KVM_CREATE_VM) failed: %d %s\n", -ret, 2078 strerror(-ret)); 2079 2080 #ifdef TARGET_S390X 2081 if (ret == -EINVAL) { 2082 fprintf(stderr, 2083 "Host kernel setup problem detected. Please verify:\n"); 2084 fprintf(stderr, "- for kernels supporting the switch_amode or" 2085 " user_mode parameters, whether\n"); 2086 fprintf(stderr, 2087 " user space is running in primary address space\n"); 2088 fprintf(stderr, 2089 "- for kernels supporting the vm.allocate_pgste sysctl, " 2090 "whether it is enabled\n"); 2091 } 2092 #endif 2093 goto err; 2094 } 2095 2096 s->vmfd = ret; 2097 2098 /* check the vcpu limits */ 2099 soft_vcpus_limit = kvm_recommended_vcpus(s); 2100 hard_vcpus_limit = kvm_max_vcpus(s); 2101 2102 while (nc->name) { 2103 if (nc->num > soft_vcpus_limit) { 2104 warn_report("Number of %s cpus requested (%d) exceeds " 2105 "the recommended cpus supported by KVM (%d)", 2106 nc->name, nc->num, soft_vcpus_limit); 2107 2108 if (nc->num > hard_vcpus_limit) { 2109 fprintf(stderr, "Number of %s cpus requested (%d) exceeds " 2110 "the maximum cpus supported by KVM (%d)\n", 2111 nc->name, nc->num, hard_vcpus_limit); 2112 exit(1); 2113 } 2114 } 2115 nc++; 2116 } 2117 2118 missing_cap = kvm_check_extension_list(s, kvm_required_capabilites); 2119 if (!missing_cap) { 2120 missing_cap = 2121 kvm_check_extension_list(s, kvm_arch_required_capabilities); 2122 } 2123 if (missing_cap) { 2124 ret = -EINVAL; 2125 fprintf(stderr, "kvm does not support %s\n%s", 2126 missing_cap->name, upgrade_note); 2127 goto err; 2128 } 2129 2130 s->coalesced_mmio = kvm_check_extension(s, KVM_CAP_COALESCED_MMIO); 2131 s->coalesced_pio = s->coalesced_mmio && 2132 kvm_check_extension(s, KVM_CAP_COALESCED_PIO); 2133 2134 dirty_log_manual_caps = 2135 kvm_check_extension(s, KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2); 2136 dirty_log_manual_caps &= (KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE | 2137 KVM_DIRTY_LOG_INITIALLY_SET); 2138 s->manual_dirty_log_protect = dirty_log_manual_caps; 2139 if (dirty_log_manual_caps) { 2140 ret = kvm_vm_enable_cap(s, KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2, 0, 2141 dirty_log_manual_caps); 2142 if (ret) { 2143 warn_report("Trying to enable capability %"PRIu64" of " 2144 "KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2 but failed. " 2145 "Falling back to the legacy mode. ", 2146 dirty_log_manual_caps); 2147 s->manual_dirty_log_protect = 0; 2148 } 2149 } 2150 2151 #ifdef KVM_CAP_VCPU_EVENTS 2152 s->vcpu_events = kvm_check_extension(s, KVM_CAP_VCPU_EVENTS); 2153 #endif 2154 2155 s->robust_singlestep = 2156 kvm_check_extension(s, KVM_CAP_X86_ROBUST_SINGLESTEP); 2157 2158 #ifdef KVM_CAP_DEBUGREGS 2159 s->debugregs = kvm_check_extension(s, KVM_CAP_DEBUGREGS); 2160 #endif 2161 2162 s->max_nested_state_len = kvm_check_extension(s, KVM_CAP_NESTED_STATE); 2163 2164 #ifdef KVM_CAP_IRQ_ROUTING 2165 kvm_direct_msi_allowed = (kvm_check_extension(s, KVM_CAP_SIGNAL_MSI) > 0); 2166 #endif 2167 2168 s->intx_set_mask = kvm_check_extension(s, KVM_CAP_PCI_2_3); 2169 2170 s->irq_set_ioctl = KVM_IRQ_LINE; 2171 if (kvm_check_extension(s, KVM_CAP_IRQ_INJECT_STATUS)) { 2172 s->irq_set_ioctl = KVM_IRQ_LINE_STATUS; 2173 } 2174 2175 kvm_readonly_mem_allowed = 2176 (kvm_check_extension(s, KVM_CAP_READONLY_MEM) > 0); 2177 2178 kvm_eventfds_allowed = 2179 (kvm_check_extension(s, KVM_CAP_IOEVENTFD) > 0); 2180 2181 kvm_irqfds_allowed = 2182 (kvm_check_extension(s, KVM_CAP_IRQFD) > 0); 2183 2184 kvm_resamplefds_allowed = 2185 (kvm_check_extension(s, KVM_CAP_IRQFD_RESAMPLE) > 0); 2186 2187 kvm_vm_attributes_allowed = 2188 (kvm_check_extension(s, KVM_CAP_VM_ATTRIBUTES) > 0); 2189 2190 kvm_ioeventfd_any_length_allowed = 2191 (kvm_check_extension(s, KVM_CAP_IOEVENTFD_ANY_LENGTH) > 0); 2192 2193 kvm_state = s; 2194 2195 /* 2196 * if memory encryption object is specified then initialize the memory 2197 * encryption context. 2198 */ 2199 if (ms->memory_encryption) { 2200 kvm_state->memcrypt_handle = sev_guest_init(ms->memory_encryption); 2201 if (!kvm_state->memcrypt_handle) { 2202 ret = -1; 2203 goto err; 2204 } 2205 2206 kvm_state->memcrypt_encrypt_data = sev_encrypt_data; 2207 } 2208 2209 ret = kvm_arch_init(ms, s); 2210 if (ret < 0) { 2211 goto err; 2212 } 2213 2214 if (s->kernel_irqchip_split == ON_OFF_AUTO_AUTO) { 2215 s->kernel_irqchip_split = mc->default_kernel_irqchip_split ? ON_OFF_AUTO_ON : ON_OFF_AUTO_OFF; 2216 } 2217 2218 qemu_register_reset(kvm_unpoison_all, NULL); 2219 2220 if (s->kernel_irqchip_allowed) { 2221 kvm_irqchip_create(s); 2222 } 2223 2224 if (kvm_eventfds_allowed) { 2225 s->memory_listener.listener.eventfd_add = kvm_mem_ioeventfd_add; 2226 s->memory_listener.listener.eventfd_del = kvm_mem_ioeventfd_del; 2227 } 2228 s->memory_listener.listener.coalesced_io_add = kvm_coalesce_mmio_region; 2229 s->memory_listener.listener.coalesced_io_del = kvm_uncoalesce_mmio_region; 2230 2231 kvm_memory_listener_register(s, &s->memory_listener, 2232 &address_space_memory, 0); 2233 memory_listener_register(&kvm_io_listener, 2234 &address_space_io); 2235 memory_listener_register(&kvm_coalesced_pio_listener, 2236 &address_space_io); 2237 2238 s->many_ioeventfds = kvm_check_many_ioeventfds(); 2239 2240 s->sync_mmu = !!kvm_vm_check_extension(kvm_state, KVM_CAP_SYNC_MMU); 2241 if (!s->sync_mmu) { 2242 ret = ram_block_discard_disable(true); 2243 assert(!ret); 2244 } 2245 2246 cpus_register_accel(&kvm_cpus); 2247 return 0; 2248 2249 err: 2250 assert(ret < 0); 2251 if (s->vmfd >= 0) { 2252 close(s->vmfd); 2253 } 2254 if (s->fd != -1) { 2255 close(s->fd); 2256 } 2257 g_free(s->memory_listener.slots); 2258 2259 return ret; 2260 } 2261 2262 void kvm_set_sigmask_len(KVMState *s, unsigned int sigmask_len) 2263 { 2264 s->sigmask_len = sigmask_len; 2265 } 2266 2267 static void kvm_handle_io(uint16_t port, MemTxAttrs attrs, void *data, int direction, 2268 int size, uint32_t count) 2269 { 2270 int i; 2271 uint8_t *ptr = data; 2272 2273 for (i = 0; i < count; i++) { 2274 address_space_rw(&address_space_io, port, attrs, 2275 ptr, size, 2276 direction == KVM_EXIT_IO_OUT); 2277 ptr += size; 2278 } 2279 } 2280 2281 static int kvm_handle_internal_error(CPUState *cpu, struct kvm_run *run) 2282 { 2283 fprintf(stderr, "KVM internal error. Suberror: %d\n", 2284 run->internal.suberror); 2285 2286 if (kvm_check_extension(kvm_state, KVM_CAP_INTERNAL_ERROR_DATA)) { 2287 int i; 2288 2289 for (i = 0; i < run->internal.ndata; ++i) { 2290 fprintf(stderr, "extra data[%d]: %"PRIx64"\n", 2291 i, (uint64_t)run->internal.data[i]); 2292 } 2293 } 2294 if (run->internal.suberror == KVM_INTERNAL_ERROR_EMULATION) { 2295 fprintf(stderr, "emulation failure\n"); 2296 if (!kvm_arch_stop_on_emulation_error(cpu)) { 2297 cpu_dump_state(cpu, stderr, CPU_DUMP_CODE); 2298 return EXCP_INTERRUPT; 2299 } 2300 } 2301 /* FIXME: Should trigger a qmp message to let management know 2302 * something went wrong. 2303 */ 2304 return -1; 2305 } 2306 2307 void kvm_flush_coalesced_mmio_buffer(void) 2308 { 2309 KVMState *s = kvm_state; 2310 2311 if (s->coalesced_flush_in_progress) { 2312 return; 2313 } 2314 2315 s->coalesced_flush_in_progress = true; 2316 2317 if (s->coalesced_mmio_ring) { 2318 struct kvm_coalesced_mmio_ring *ring = s->coalesced_mmio_ring; 2319 while (ring->first != ring->last) { 2320 struct kvm_coalesced_mmio *ent; 2321 2322 ent = &ring->coalesced_mmio[ring->first]; 2323 2324 if (ent->pio == 1) { 2325 address_space_write(&address_space_io, ent->phys_addr, 2326 MEMTXATTRS_UNSPECIFIED, ent->data, 2327 ent->len); 2328 } else { 2329 cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len); 2330 } 2331 smp_wmb(); 2332 ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX; 2333 } 2334 } 2335 2336 s->coalesced_flush_in_progress = false; 2337 } 2338 2339 static void do_kvm_cpu_synchronize_state(CPUState *cpu, run_on_cpu_data arg) 2340 { 2341 if (!cpu->vcpu_dirty) { 2342 kvm_arch_get_registers(cpu); 2343 cpu->vcpu_dirty = true; 2344 } 2345 } 2346 2347 void kvm_cpu_synchronize_state(CPUState *cpu) 2348 { 2349 if (!cpu->vcpu_dirty) { 2350 run_on_cpu(cpu, do_kvm_cpu_synchronize_state, RUN_ON_CPU_NULL); 2351 } 2352 } 2353 2354 static void do_kvm_cpu_synchronize_post_reset(CPUState *cpu, run_on_cpu_data arg) 2355 { 2356 kvm_arch_put_registers(cpu, KVM_PUT_RESET_STATE); 2357 cpu->vcpu_dirty = false; 2358 } 2359 2360 void kvm_cpu_synchronize_post_reset(CPUState *cpu) 2361 { 2362 run_on_cpu(cpu, do_kvm_cpu_synchronize_post_reset, RUN_ON_CPU_NULL); 2363 } 2364 2365 static void do_kvm_cpu_synchronize_post_init(CPUState *cpu, run_on_cpu_data arg) 2366 { 2367 kvm_arch_put_registers(cpu, KVM_PUT_FULL_STATE); 2368 cpu->vcpu_dirty = false; 2369 } 2370 2371 void kvm_cpu_synchronize_post_init(CPUState *cpu) 2372 { 2373 run_on_cpu(cpu, do_kvm_cpu_synchronize_post_init, RUN_ON_CPU_NULL); 2374 } 2375 2376 static void do_kvm_cpu_synchronize_pre_loadvm(CPUState *cpu, run_on_cpu_data arg) 2377 { 2378 cpu->vcpu_dirty = true; 2379 } 2380 2381 void kvm_cpu_synchronize_pre_loadvm(CPUState *cpu) 2382 { 2383 run_on_cpu(cpu, do_kvm_cpu_synchronize_pre_loadvm, RUN_ON_CPU_NULL); 2384 } 2385 2386 #ifdef KVM_HAVE_MCE_INJECTION 2387 static __thread void *pending_sigbus_addr; 2388 static __thread int pending_sigbus_code; 2389 static __thread bool have_sigbus_pending; 2390 #endif 2391 2392 static void kvm_cpu_kick(CPUState *cpu) 2393 { 2394 qatomic_set(&cpu->kvm_run->immediate_exit, 1); 2395 } 2396 2397 static void kvm_cpu_kick_self(void) 2398 { 2399 if (kvm_immediate_exit) { 2400 kvm_cpu_kick(current_cpu); 2401 } else { 2402 qemu_cpu_kick_self(); 2403 } 2404 } 2405 2406 static void kvm_eat_signals(CPUState *cpu) 2407 { 2408 struct timespec ts = { 0, 0 }; 2409 siginfo_t siginfo; 2410 sigset_t waitset; 2411 sigset_t chkset; 2412 int r; 2413 2414 if (kvm_immediate_exit) { 2415 qatomic_set(&cpu->kvm_run->immediate_exit, 0); 2416 /* Write kvm_run->immediate_exit before the cpu->exit_request 2417 * write in kvm_cpu_exec. 2418 */ 2419 smp_wmb(); 2420 return; 2421 } 2422 2423 sigemptyset(&waitset); 2424 sigaddset(&waitset, SIG_IPI); 2425 2426 do { 2427 r = sigtimedwait(&waitset, &siginfo, &ts); 2428 if (r == -1 && !(errno == EAGAIN || errno == EINTR)) { 2429 perror("sigtimedwait"); 2430 exit(1); 2431 } 2432 2433 r = sigpending(&chkset); 2434 if (r == -1) { 2435 perror("sigpending"); 2436 exit(1); 2437 } 2438 } while (sigismember(&chkset, SIG_IPI)); 2439 } 2440 2441 int kvm_cpu_exec(CPUState *cpu) 2442 { 2443 struct kvm_run *run = cpu->kvm_run; 2444 int ret, run_ret; 2445 2446 DPRINTF("kvm_cpu_exec()\n"); 2447 2448 if (kvm_arch_process_async_events(cpu)) { 2449 qatomic_set(&cpu->exit_request, 0); 2450 return EXCP_HLT; 2451 } 2452 2453 qemu_mutex_unlock_iothread(); 2454 cpu_exec_start(cpu); 2455 2456 do { 2457 MemTxAttrs attrs; 2458 2459 if (cpu->vcpu_dirty) { 2460 kvm_arch_put_registers(cpu, KVM_PUT_RUNTIME_STATE); 2461 cpu->vcpu_dirty = false; 2462 } 2463 2464 kvm_arch_pre_run(cpu, run); 2465 if (qatomic_read(&cpu->exit_request)) { 2466 DPRINTF("interrupt exit requested\n"); 2467 /* 2468 * KVM requires us to reenter the kernel after IO exits to complete 2469 * instruction emulation. This self-signal will ensure that we 2470 * leave ASAP again. 2471 */ 2472 kvm_cpu_kick_self(); 2473 } 2474 2475 /* Read cpu->exit_request before KVM_RUN reads run->immediate_exit. 2476 * Matching barrier in kvm_eat_signals. 2477 */ 2478 smp_rmb(); 2479 2480 run_ret = kvm_vcpu_ioctl(cpu, KVM_RUN, 0); 2481 2482 attrs = kvm_arch_post_run(cpu, run); 2483 2484 #ifdef KVM_HAVE_MCE_INJECTION 2485 if (unlikely(have_sigbus_pending)) { 2486 qemu_mutex_lock_iothread(); 2487 kvm_arch_on_sigbus_vcpu(cpu, pending_sigbus_code, 2488 pending_sigbus_addr); 2489 have_sigbus_pending = false; 2490 qemu_mutex_unlock_iothread(); 2491 } 2492 #endif 2493 2494 if (run_ret < 0) { 2495 if (run_ret == -EINTR || run_ret == -EAGAIN) { 2496 DPRINTF("io window exit\n"); 2497 kvm_eat_signals(cpu); 2498 ret = EXCP_INTERRUPT; 2499 break; 2500 } 2501 fprintf(stderr, "error: kvm run failed %s\n", 2502 strerror(-run_ret)); 2503 #ifdef TARGET_PPC 2504 if (run_ret == -EBUSY) { 2505 fprintf(stderr, 2506 "This is probably because your SMT is enabled.\n" 2507 "VCPU can only run on primary threads with all " 2508 "secondary threads offline.\n"); 2509 } 2510 #endif 2511 ret = -1; 2512 break; 2513 } 2514 2515 trace_kvm_run_exit(cpu->cpu_index, run->exit_reason); 2516 switch (run->exit_reason) { 2517 case KVM_EXIT_IO: 2518 DPRINTF("handle_io\n"); 2519 /* Called outside BQL */ 2520 kvm_handle_io(run->io.port, attrs, 2521 (uint8_t *)run + run->io.data_offset, 2522 run->io.direction, 2523 run->io.size, 2524 run->io.count); 2525 ret = 0; 2526 break; 2527 case KVM_EXIT_MMIO: 2528 DPRINTF("handle_mmio\n"); 2529 /* Called outside BQL */ 2530 address_space_rw(&address_space_memory, 2531 run->mmio.phys_addr, attrs, 2532 run->mmio.data, 2533 run->mmio.len, 2534 run->mmio.is_write); 2535 ret = 0; 2536 break; 2537 case KVM_EXIT_IRQ_WINDOW_OPEN: 2538 DPRINTF("irq_window_open\n"); 2539 ret = EXCP_INTERRUPT; 2540 break; 2541 case KVM_EXIT_SHUTDOWN: 2542 DPRINTF("shutdown\n"); 2543 qemu_system_reset_request(SHUTDOWN_CAUSE_GUEST_RESET); 2544 ret = EXCP_INTERRUPT; 2545 break; 2546 case KVM_EXIT_UNKNOWN: 2547 fprintf(stderr, "KVM: unknown exit, hardware reason %" PRIx64 "\n", 2548 (uint64_t)run->hw.hardware_exit_reason); 2549 ret = -1; 2550 break; 2551 case KVM_EXIT_INTERNAL_ERROR: 2552 ret = kvm_handle_internal_error(cpu, run); 2553 break; 2554 case KVM_EXIT_SYSTEM_EVENT: 2555 switch (run->system_event.type) { 2556 case KVM_SYSTEM_EVENT_SHUTDOWN: 2557 qemu_system_shutdown_request(SHUTDOWN_CAUSE_GUEST_SHUTDOWN); 2558 ret = EXCP_INTERRUPT; 2559 break; 2560 case KVM_SYSTEM_EVENT_RESET: 2561 qemu_system_reset_request(SHUTDOWN_CAUSE_GUEST_RESET); 2562 ret = EXCP_INTERRUPT; 2563 break; 2564 case KVM_SYSTEM_EVENT_CRASH: 2565 kvm_cpu_synchronize_state(cpu); 2566 qemu_mutex_lock_iothread(); 2567 qemu_system_guest_panicked(cpu_get_crash_info(cpu)); 2568 qemu_mutex_unlock_iothread(); 2569 ret = 0; 2570 break; 2571 default: 2572 DPRINTF("kvm_arch_handle_exit\n"); 2573 ret = kvm_arch_handle_exit(cpu, run); 2574 break; 2575 } 2576 break; 2577 default: 2578 DPRINTF("kvm_arch_handle_exit\n"); 2579 ret = kvm_arch_handle_exit(cpu, run); 2580 break; 2581 } 2582 } while (ret == 0); 2583 2584 cpu_exec_end(cpu); 2585 qemu_mutex_lock_iothread(); 2586 2587 if (ret < 0) { 2588 cpu_dump_state(cpu, stderr, CPU_DUMP_CODE); 2589 vm_stop(RUN_STATE_INTERNAL_ERROR); 2590 } 2591 2592 qatomic_set(&cpu->exit_request, 0); 2593 return ret; 2594 } 2595 2596 int kvm_ioctl(KVMState *s, int type, ...) 2597 { 2598 int ret; 2599 void *arg; 2600 va_list ap; 2601 2602 va_start(ap, type); 2603 arg = va_arg(ap, void *); 2604 va_end(ap); 2605 2606 trace_kvm_ioctl(type, arg); 2607 ret = ioctl(s->fd, type, arg); 2608 if (ret == -1) { 2609 ret = -errno; 2610 } 2611 return ret; 2612 } 2613 2614 int kvm_vm_ioctl(KVMState *s, int type, ...) 2615 { 2616 int ret; 2617 void *arg; 2618 va_list ap; 2619 2620 va_start(ap, type); 2621 arg = va_arg(ap, void *); 2622 va_end(ap); 2623 2624 trace_kvm_vm_ioctl(type, arg); 2625 ret = ioctl(s->vmfd, type, arg); 2626 if (ret == -1) { 2627 ret = -errno; 2628 } 2629 return ret; 2630 } 2631 2632 int kvm_vcpu_ioctl(CPUState *cpu, int type, ...) 2633 { 2634 int ret; 2635 void *arg; 2636 va_list ap; 2637 2638 va_start(ap, type); 2639 arg = va_arg(ap, void *); 2640 va_end(ap); 2641 2642 trace_kvm_vcpu_ioctl(cpu->cpu_index, type, arg); 2643 ret = ioctl(cpu->kvm_fd, type, arg); 2644 if (ret == -1) { 2645 ret = -errno; 2646 } 2647 return ret; 2648 } 2649 2650 int kvm_device_ioctl(int fd, int type, ...) 2651 { 2652 int ret; 2653 void *arg; 2654 va_list ap; 2655 2656 va_start(ap, type); 2657 arg = va_arg(ap, void *); 2658 va_end(ap); 2659 2660 trace_kvm_device_ioctl(fd, type, arg); 2661 ret = ioctl(fd, type, arg); 2662 if (ret == -1) { 2663 ret = -errno; 2664 } 2665 return ret; 2666 } 2667 2668 int kvm_vm_check_attr(KVMState *s, uint32_t group, uint64_t attr) 2669 { 2670 int ret; 2671 struct kvm_device_attr attribute = { 2672 .group = group, 2673 .attr = attr, 2674 }; 2675 2676 if (!kvm_vm_attributes_allowed) { 2677 return 0; 2678 } 2679 2680 ret = kvm_vm_ioctl(s, KVM_HAS_DEVICE_ATTR, &attribute); 2681 /* kvm returns 0 on success for HAS_DEVICE_ATTR */ 2682 return ret ? 0 : 1; 2683 } 2684 2685 int kvm_device_check_attr(int dev_fd, uint32_t group, uint64_t attr) 2686 { 2687 struct kvm_device_attr attribute = { 2688 .group = group, 2689 .attr = attr, 2690 .flags = 0, 2691 }; 2692 2693 return kvm_device_ioctl(dev_fd, KVM_HAS_DEVICE_ATTR, &attribute) ? 0 : 1; 2694 } 2695 2696 int kvm_device_access(int fd, int group, uint64_t attr, 2697 void *val, bool write, Error **errp) 2698 { 2699 struct kvm_device_attr kvmattr; 2700 int err; 2701 2702 kvmattr.flags = 0; 2703 kvmattr.group = group; 2704 kvmattr.attr = attr; 2705 kvmattr.addr = (uintptr_t)val; 2706 2707 err = kvm_device_ioctl(fd, 2708 write ? KVM_SET_DEVICE_ATTR : KVM_GET_DEVICE_ATTR, 2709 &kvmattr); 2710 if (err < 0) { 2711 error_setg_errno(errp, -err, 2712 "KVM_%s_DEVICE_ATTR failed: Group %d " 2713 "attr 0x%016" PRIx64, 2714 write ? "SET" : "GET", group, attr); 2715 } 2716 return err; 2717 } 2718 2719 bool kvm_has_sync_mmu(void) 2720 { 2721 return kvm_state->sync_mmu; 2722 } 2723 2724 int kvm_has_vcpu_events(void) 2725 { 2726 return kvm_state->vcpu_events; 2727 } 2728 2729 int kvm_has_robust_singlestep(void) 2730 { 2731 return kvm_state->robust_singlestep; 2732 } 2733 2734 int kvm_has_debugregs(void) 2735 { 2736 return kvm_state->debugregs; 2737 } 2738 2739 int kvm_max_nested_state_length(void) 2740 { 2741 return kvm_state->max_nested_state_len; 2742 } 2743 2744 int kvm_has_many_ioeventfds(void) 2745 { 2746 if (!kvm_enabled()) { 2747 return 0; 2748 } 2749 return kvm_state->many_ioeventfds; 2750 } 2751 2752 int kvm_has_gsi_routing(void) 2753 { 2754 #ifdef KVM_CAP_IRQ_ROUTING 2755 return kvm_check_extension(kvm_state, KVM_CAP_IRQ_ROUTING); 2756 #else 2757 return false; 2758 #endif 2759 } 2760 2761 int kvm_has_intx_set_mask(void) 2762 { 2763 return kvm_state->intx_set_mask; 2764 } 2765 2766 bool kvm_arm_supports_user_irq(void) 2767 { 2768 return kvm_check_extension(kvm_state, KVM_CAP_ARM_USER_IRQ); 2769 } 2770 2771 #ifdef KVM_CAP_SET_GUEST_DEBUG 2772 struct kvm_sw_breakpoint *kvm_find_sw_breakpoint(CPUState *cpu, 2773 target_ulong pc) 2774 { 2775 struct kvm_sw_breakpoint *bp; 2776 2777 QTAILQ_FOREACH(bp, &cpu->kvm_state->kvm_sw_breakpoints, entry) { 2778 if (bp->pc == pc) { 2779 return bp; 2780 } 2781 } 2782 return NULL; 2783 } 2784 2785 int kvm_sw_breakpoints_active(CPUState *cpu) 2786 { 2787 return !QTAILQ_EMPTY(&cpu->kvm_state->kvm_sw_breakpoints); 2788 } 2789 2790 struct kvm_set_guest_debug_data { 2791 struct kvm_guest_debug dbg; 2792 int err; 2793 }; 2794 2795 static void kvm_invoke_set_guest_debug(CPUState *cpu, run_on_cpu_data data) 2796 { 2797 struct kvm_set_guest_debug_data *dbg_data = 2798 (struct kvm_set_guest_debug_data *) data.host_ptr; 2799 2800 dbg_data->err = kvm_vcpu_ioctl(cpu, KVM_SET_GUEST_DEBUG, 2801 &dbg_data->dbg); 2802 } 2803 2804 int kvm_update_guest_debug(CPUState *cpu, unsigned long reinject_trap) 2805 { 2806 struct kvm_set_guest_debug_data data; 2807 2808 data.dbg.control = reinject_trap; 2809 2810 if (cpu->singlestep_enabled) { 2811 data.dbg.control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_SINGLESTEP; 2812 } 2813 kvm_arch_update_guest_debug(cpu, &data.dbg); 2814 2815 run_on_cpu(cpu, kvm_invoke_set_guest_debug, 2816 RUN_ON_CPU_HOST_PTR(&data)); 2817 return data.err; 2818 } 2819 2820 int kvm_insert_breakpoint(CPUState *cpu, target_ulong addr, 2821 target_ulong len, int type) 2822 { 2823 struct kvm_sw_breakpoint *bp; 2824 int err; 2825 2826 if (type == GDB_BREAKPOINT_SW) { 2827 bp = kvm_find_sw_breakpoint(cpu, addr); 2828 if (bp) { 2829 bp->use_count++; 2830 return 0; 2831 } 2832 2833 bp = g_malloc(sizeof(struct kvm_sw_breakpoint)); 2834 bp->pc = addr; 2835 bp->use_count = 1; 2836 err = kvm_arch_insert_sw_breakpoint(cpu, bp); 2837 if (err) { 2838 g_free(bp); 2839 return err; 2840 } 2841 2842 QTAILQ_INSERT_HEAD(&cpu->kvm_state->kvm_sw_breakpoints, bp, entry); 2843 } else { 2844 err = kvm_arch_insert_hw_breakpoint(addr, len, type); 2845 if (err) { 2846 return err; 2847 } 2848 } 2849 2850 CPU_FOREACH(cpu) { 2851 err = kvm_update_guest_debug(cpu, 0); 2852 if (err) { 2853 return err; 2854 } 2855 } 2856 return 0; 2857 } 2858 2859 int kvm_remove_breakpoint(CPUState *cpu, target_ulong addr, 2860 target_ulong len, int type) 2861 { 2862 struct kvm_sw_breakpoint *bp; 2863 int err; 2864 2865 if (type == GDB_BREAKPOINT_SW) { 2866 bp = kvm_find_sw_breakpoint(cpu, addr); 2867 if (!bp) { 2868 return -ENOENT; 2869 } 2870 2871 if (bp->use_count > 1) { 2872 bp->use_count--; 2873 return 0; 2874 } 2875 2876 err = kvm_arch_remove_sw_breakpoint(cpu, bp); 2877 if (err) { 2878 return err; 2879 } 2880 2881 QTAILQ_REMOVE(&cpu->kvm_state->kvm_sw_breakpoints, bp, entry); 2882 g_free(bp); 2883 } else { 2884 err = kvm_arch_remove_hw_breakpoint(addr, len, type); 2885 if (err) { 2886 return err; 2887 } 2888 } 2889 2890 CPU_FOREACH(cpu) { 2891 err = kvm_update_guest_debug(cpu, 0); 2892 if (err) { 2893 return err; 2894 } 2895 } 2896 return 0; 2897 } 2898 2899 void kvm_remove_all_breakpoints(CPUState *cpu) 2900 { 2901 struct kvm_sw_breakpoint *bp, *next; 2902 KVMState *s = cpu->kvm_state; 2903 CPUState *tmpcpu; 2904 2905 QTAILQ_FOREACH_SAFE(bp, &s->kvm_sw_breakpoints, entry, next) { 2906 if (kvm_arch_remove_sw_breakpoint(cpu, bp) != 0) { 2907 /* Try harder to find a CPU that currently sees the breakpoint. */ 2908 CPU_FOREACH(tmpcpu) { 2909 if (kvm_arch_remove_sw_breakpoint(tmpcpu, bp) == 0) { 2910 break; 2911 } 2912 } 2913 } 2914 QTAILQ_REMOVE(&s->kvm_sw_breakpoints, bp, entry); 2915 g_free(bp); 2916 } 2917 kvm_arch_remove_all_hw_breakpoints(); 2918 2919 CPU_FOREACH(cpu) { 2920 kvm_update_guest_debug(cpu, 0); 2921 } 2922 } 2923 2924 #else /* !KVM_CAP_SET_GUEST_DEBUG */ 2925 2926 int kvm_update_guest_debug(CPUState *cpu, unsigned long reinject_trap) 2927 { 2928 return -EINVAL; 2929 } 2930 2931 int kvm_insert_breakpoint(CPUState *cpu, target_ulong addr, 2932 target_ulong len, int type) 2933 { 2934 return -EINVAL; 2935 } 2936 2937 int kvm_remove_breakpoint(CPUState *cpu, target_ulong addr, 2938 target_ulong len, int type) 2939 { 2940 return -EINVAL; 2941 } 2942 2943 void kvm_remove_all_breakpoints(CPUState *cpu) 2944 { 2945 } 2946 #endif /* !KVM_CAP_SET_GUEST_DEBUG */ 2947 2948 static int kvm_set_signal_mask(CPUState *cpu, const sigset_t *sigset) 2949 { 2950 KVMState *s = kvm_state; 2951 struct kvm_signal_mask *sigmask; 2952 int r; 2953 2954 sigmask = g_malloc(sizeof(*sigmask) + sizeof(*sigset)); 2955 2956 sigmask->len = s->sigmask_len; 2957 memcpy(sigmask->sigset, sigset, sizeof(*sigset)); 2958 r = kvm_vcpu_ioctl(cpu, KVM_SET_SIGNAL_MASK, sigmask); 2959 g_free(sigmask); 2960 2961 return r; 2962 } 2963 2964 static void kvm_ipi_signal(int sig) 2965 { 2966 if (current_cpu) { 2967 assert(kvm_immediate_exit); 2968 kvm_cpu_kick(current_cpu); 2969 } 2970 } 2971 2972 void kvm_init_cpu_signals(CPUState *cpu) 2973 { 2974 int r; 2975 sigset_t set; 2976 struct sigaction sigact; 2977 2978 memset(&sigact, 0, sizeof(sigact)); 2979 sigact.sa_handler = kvm_ipi_signal; 2980 sigaction(SIG_IPI, &sigact, NULL); 2981 2982 pthread_sigmask(SIG_BLOCK, NULL, &set); 2983 #if defined KVM_HAVE_MCE_INJECTION 2984 sigdelset(&set, SIGBUS); 2985 pthread_sigmask(SIG_SETMASK, &set, NULL); 2986 #endif 2987 sigdelset(&set, SIG_IPI); 2988 if (kvm_immediate_exit) { 2989 r = pthread_sigmask(SIG_SETMASK, &set, NULL); 2990 } else { 2991 r = kvm_set_signal_mask(cpu, &set); 2992 } 2993 if (r) { 2994 fprintf(stderr, "kvm_set_signal_mask: %s\n", strerror(-r)); 2995 exit(1); 2996 } 2997 } 2998 2999 /* Called asynchronously in VCPU thread. */ 3000 int kvm_on_sigbus_vcpu(CPUState *cpu, int code, void *addr) 3001 { 3002 #ifdef KVM_HAVE_MCE_INJECTION 3003 if (have_sigbus_pending) { 3004 return 1; 3005 } 3006 have_sigbus_pending = true; 3007 pending_sigbus_addr = addr; 3008 pending_sigbus_code = code; 3009 qatomic_set(&cpu->exit_request, 1); 3010 return 0; 3011 #else 3012 return 1; 3013 #endif 3014 } 3015 3016 /* Called synchronously (via signalfd) in main thread. */ 3017 int kvm_on_sigbus(int code, void *addr) 3018 { 3019 #ifdef KVM_HAVE_MCE_INJECTION 3020 /* Action required MCE kills the process if SIGBUS is blocked. Because 3021 * that's what happens in the I/O thread, where we handle MCE via signalfd, 3022 * we can only get action optional here. 3023 */ 3024 assert(code != BUS_MCEERR_AR); 3025 kvm_arch_on_sigbus_vcpu(first_cpu, code, addr); 3026 return 0; 3027 #else 3028 return 1; 3029 #endif 3030 } 3031 3032 int kvm_create_device(KVMState *s, uint64_t type, bool test) 3033 { 3034 int ret; 3035 struct kvm_create_device create_dev; 3036 3037 create_dev.type = type; 3038 create_dev.fd = -1; 3039 create_dev.flags = test ? KVM_CREATE_DEVICE_TEST : 0; 3040 3041 if (!kvm_check_extension(s, KVM_CAP_DEVICE_CTRL)) { 3042 return -ENOTSUP; 3043 } 3044 3045 ret = kvm_vm_ioctl(s, KVM_CREATE_DEVICE, &create_dev); 3046 if (ret) { 3047 return ret; 3048 } 3049 3050 return test ? 0 : create_dev.fd; 3051 } 3052 3053 bool kvm_device_supported(int vmfd, uint64_t type) 3054 { 3055 struct kvm_create_device create_dev = { 3056 .type = type, 3057 .fd = -1, 3058 .flags = KVM_CREATE_DEVICE_TEST, 3059 }; 3060 3061 if (ioctl(vmfd, KVM_CHECK_EXTENSION, KVM_CAP_DEVICE_CTRL) <= 0) { 3062 return false; 3063 } 3064 3065 return (ioctl(vmfd, KVM_CREATE_DEVICE, &create_dev) >= 0); 3066 } 3067 3068 int kvm_set_one_reg(CPUState *cs, uint64_t id, void *source) 3069 { 3070 struct kvm_one_reg reg; 3071 int r; 3072 3073 reg.id = id; 3074 reg.addr = (uintptr_t) source; 3075 r = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, ®); 3076 if (r) { 3077 trace_kvm_failed_reg_set(id, strerror(-r)); 3078 } 3079 return r; 3080 } 3081 3082 int kvm_get_one_reg(CPUState *cs, uint64_t id, void *target) 3083 { 3084 struct kvm_one_reg reg; 3085 int r; 3086 3087 reg.id = id; 3088 reg.addr = (uintptr_t) target; 3089 r = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, ®); 3090 if (r) { 3091 trace_kvm_failed_reg_get(id, strerror(-r)); 3092 } 3093 return r; 3094 } 3095 3096 static bool kvm_accel_has_memory(MachineState *ms, AddressSpace *as, 3097 hwaddr start_addr, hwaddr size) 3098 { 3099 KVMState *kvm = KVM_STATE(ms->accelerator); 3100 int i; 3101 3102 for (i = 0; i < kvm->nr_as; ++i) { 3103 if (kvm->as[i].as == as && kvm->as[i].ml) { 3104 size = MIN(kvm_max_slot_size, size); 3105 return NULL != kvm_lookup_matching_slot(kvm->as[i].ml, 3106 start_addr, size); 3107 } 3108 } 3109 3110 return false; 3111 } 3112 3113 static void kvm_get_kvm_shadow_mem(Object *obj, Visitor *v, 3114 const char *name, void *opaque, 3115 Error **errp) 3116 { 3117 KVMState *s = KVM_STATE(obj); 3118 int64_t value = s->kvm_shadow_mem; 3119 3120 visit_type_int(v, name, &value, errp); 3121 } 3122 3123 static void kvm_set_kvm_shadow_mem(Object *obj, Visitor *v, 3124 const char *name, void *opaque, 3125 Error **errp) 3126 { 3127 KVMState *s = KVM_STATE(obj); 3128 int64_t value; 3129 3130 if (!visit_type_int(v, name, &value, errp)) { 3131 return; 3132 } 3133 3134 s->kvm_shadow_mem = value; 3135 } 3136 3137 static void kvm_set_kernel_irqchip(Object *obj, Visitor *v, 3138 const char *name, void *opaque, 3139 Error **errp) 3140 { 3141 KVMState *s = KVM_STATE(obj); 3142 OnOffSplit mode; 3143 3144 if (!visit_type_OnOffSplit(v, name, &mode, errp)) { 3145 return; 3146 } 3147 switch (mode) { 3148 case ON_OFF_SPLIT_ON: 3149 s->kernel_irqchip_allowed = true; 3150 s->kernel_irqchip_required = true; 3151 s->kernel_irqchip_split = ON_OFF_AUTO_OFF; 3152 break; 3153 case ON_OFF_SPLIT_OFF: 3154 s->kernel_irqchip_allowed = false; 3155 s->kernel_irqchip_required = false; 3156 s->kernel_irqchip_split = ON_OFF_AUTO_OFF; 3157 break; 3158 case ON_OFF_SPLIT_SPLIT: 3159 s->kernel_irqchip_allowed = true; 3160 s->kernel_irqchip_required = true; 3161 s->kernel_irqchip_split = ON_OFF_AUTO_ON; 3162 break; 3163 default: 3164 /* The value was checked in visit_type_OnOffSplit() above. If 3165 * we get here, then something is wrong in QEMU. 3166 */ 3167 abort(); 3168 } 3169 } 3170 3171 bool kvm_kernel_irqchip_allowed(void) 3172 { 3173 return kvm_state->kernel_irqchip_allowed; 3174 } 3175 3176 bool kvm_kernel_irqchip_required(void) 3177 { 3178 return kvm_state->kernel_irqchip_required; 3179 } 3180 3181 bool kvm_kernel_irqchip_split(void) 3182 { 3183 return kvm_state->kernel_irqchip_split == ON_OFF_AUTO_ON; 3184 } 3185 3186 static void kvm_accel_instance_init(Object *obj) 3187 { 3188 KVMState *s = KVM_STATE(obj); 3189 3190 s->kvm_shadow_mem = -1; 3191 s->kernel_irqchip_allowed = true; 3192 s->kernel_irqchip_split = ON_OFF_AUTO_AUTO; 3193 } 3194 3195 static void kvm_accel_class_init(ObjectClass *oc, void *data) 3196 { 3197 AccelClass *ac = ACCEL_CLASS(oc); 3198 ac->name = "KVM"; 3199 ac->init_machine = kvm_init; 3200 ac->has_memory = kvm_accel_has_memory; 3201 ac->allowed = &kvm_allowed; 3202 3203 object_class_property_add(oc, "kernel-irqchip", "on|off|split", 3204 NULL, kvm_set_kernel_irqchip, 3205 NULL, NULL); 3206 object_class_property_set_description(oc, "kernel-irqchip", 3207 "Configure KVM in-kernel irqchip"); 3208 3209 object_class_property_add(oc, "kvm-shadow-mem", "int", 3210 kvm_get_kvm_shadow_mem, kvm_set_kvm_shadow_mem, 3211 NULL, NULL); 3212 object_class_property_set_description(oc, "kvm-shadow-mem", 3213 "KVM shadow MMU size"); 3214 } 3215 3216 static const TypeInfo kvm_accel_type = { 3217 .name = TYPE_KVM_ACCEL, 3218 .parent = TYPE_ACCEL, 3219 .instance_init = kvm_accel_instance_init, 3220 .class_init = kvm_accel_class_init, 3221 .instance_size = sizeof(KVMState), 3222 }; 3223 3224 static void kvm_type_init(void) 3225 { 3226 type_register_static(&kvm_accel_type); 3227 } 3228 3229 type_init(kvm_type_init); 3230