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 "qemu/bswap.h" 34 #include "exec/memory.h" 35 #include "exec/ram_addr.h" 36 #include "qemu/event_notifier.h" 37 #include "qemu/main-loop.h" 38 #include "trace.h" 39 #include "hw/irq.h" 40 #include "qapi/visitor.h" 41 #include "qapi/qapi-types-common.h" 42 #include "qapi/qapi-visit-common.h" 43 #include "sysemu/reset.h" 44 #include "qemu/guest-random.h" 45 #include "sysemu/hw_accel.h" 46 #include "kvm-cpus.h" 47 48 #include "hw/boards.h" 49 50 /* This check must be after config-host.h is included */ 51 #ifdef CONFIG_EVENTFD 52 #include <sys/eventfd.h> 53 #endif 54 55 /* KVM uses PAGE_SIZE in its definition of KVM_COALESCED_MMIO_MAX. We 56 * need to use the real host PAGE_SIZE, as that's what KVM will use. 57 */ 58 #ifdef PAGE_SIZE 59 #undef PAGE_SIZE 60 #endif 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 /* For "info mtree -f" to tell if an MR is registered in KVM */ 124 int nr_as; 125 struct KVMAs { 126 KVMMemoryListener *ml; 127 AddressSpace *as; 128 } *as; 129 }; 130 131 KVMState *kvm_state; 132 bool kvm_kernel_irqchip; 133 bool kvm_split_irqchip; 134 bool kvm_async_interrupts_allowed; 135 bool kvm_halt_in_kernel_allowed; 136 bool kvm_eventfds_allowed; 137 bool kvm_irqfds_allowed; 138 bool kvm_resamplefds_allowed; 139 bool kvm_msi_via_irqfd_allowed; 140 bool kvm_gsi_routing_allowed; 141 bool kvm_gsi_direct_mapping; 142 bool kvm_allowed; 143 bool kvm_readonly_mem_allowed; 144 bool kvm_vm_attributes_allowed; 145 bool kvm_direct_msi_allowed; 146 bool kvm_ioeventfd_any_length_allowed; 147 bool kvm_msi_use_devid; 148 static bool kvm_immediate_exit; 149 static hwaddr kvm_max_slot_size = ~0; 150 151 static const KVMCapabilityInfo kvm_required_capabilites[] = { 152 KVM_CAP_INFO(USER_MEMORY), 153 KVM_CAP_INFO(DESTROY_MEMORY_REGION_WORKS), 154 KVM_CAP_INFO(JOIN_MEMORY_REGIONS_WORKS), 155 KVM_CAP_LAST_INFO 156 }; 157 158 static NotifierList kvm_irqchip_change_notifiers = 159 NOTIFIER_LIST_INITIALIZER(kvm_irqchip_change_notifiers); 160 161 struct KVMResampleFd { 162 int gsi; 163 EventNotifier *resample_event; 164 QLIST_ENTRY(KVMResampleFd) node; 165 }; 166 typedef struct KVMResampleFd KVMResampleFd; 167 168 /* 169 * Only used with split irqchip where we need to do the resample fd 170 * kick for the kernel from userspace. 171 */ 172 static QLIST_HEAD(, KVMResampleFd) kvm_resample_fd_list = 173 QLIST_HEAD_INITIALIZER(kvm_resample_fd_list); 174 175 #define kvm_slots_lock(kml) qemu_mutex_lock(&(kml)->slots_lock) 176 #define kvm_slots_unlock(kml) qemu_mutex_unlock(&(kml)->slots_lock) 177 178 static inline void kvm_resample_fd_remove(int gsi) 179 { 180 KVMResampleFd *rfd; 181 182 QLIST_FOREACH(rfd, &kvm_resample_fd_list, node) { 183 if (rfd->gsi == gsi) { 184 QLIST_REMOVE(rfd, node); 185 g_free(rfd); 186 break; 187 } 188 } 189 } 190 191 static inline void kvm_resample_fd_insert(int gsi, EventNotifier *event) 192 { 193 KVMResampleFd *rfd = g_new0(KVMResampleFd, 1); 194 195 rfd->gsi = gsi; 196 rfd->resample_event = event; 197 198 QLIST_INSERT_HEAD(&kvm_resample_fd_list, rfd, node); 199 } 200 201 void kvm_resample_fd_notify(int gsi) 202 { 203 KVMResampleFd *rfd; 204 205 QLIST_FOREACH(rfd, &kvm_resample_fd_list, node) { 206 if (rfd->gsi == gsi) { 207 event_notifier_set(rfd->resample_event); 208 trace_kvm_resample_fd_notify(gsi); 209 return; 210 } 211 } 212 } 213 214 int kvm_get_max_memslots(void) 215 { 216 KVMState *s = KVM_STATE(current_accel()); 217 218 return s->nr_slots; 219 } 220 221 /* Called with KVMMemoryListener.slots_lock held */ 222 static KVMSlot *kvm_get_free_slot(KVMMemoryListener *kml) 223 { 224 KVMState *s = kvm_state; 225 int i; 226 227 for (i = 0; i < s->nr_slots; i++) { 228 if (kml->slots[i].memory_size == 0) { 229 return &kml->slots[i]; 230 } 231 } 232 233 return NULL; 234 } 235 236 bool kvm_has_free_slot(MachineState *ms) 237 { 238 KVMState *s = KVM_STATE(ms->accelerator); 239 bool result; 240 KVMMemoryListener *kml = &s->memory_listener; 241 242 kvm_slots_lock(kml); 243 result = !!kvm_get_free_slot(kml); 244 kvm_slots_unlock(kml); 245 246 return result; 247 } 248 249 /* Called with KVMMemoryListener.slots_lock held */ 250 static KVMSlot *kvm_alloc_slot(KVMMemoryListener *kml) 251 { 252 KVMSlot *slot = kvm_get_free_slot(kml); 253 254 if (slot) { 255 return slot; 256 } 257 258 fprintf(stderr, "%s: no free slot available\n", __func__); 259 abort(); 260 } 261 262 static KVMSlot *kvm_lookup_matching_slot(KVMMemoryListener *kml, 263 hwaddr start_addr, 264 hwaddr size) 265 { 266 KVMState *s = kvm_state; 267 int i; 268 269 for (i = 0; i < s->nr_slots; i++) { 270 KVMSlot *mem = &kml->slots[i]; 271 272 if (start_addr == mem->start_addr && size == mem->memory_size) { 273 return mem; 274 } 275 } 276 277 return NULL; 278 } 279 280 /* 281 * Calculate and align the start address and the size of the section. 282 * Return the size. If the size is 0, the aligned section is empty. 283 */ 284 static hwaddr kvm_align_section(MemoryRegionSection *section, 285 hwaddr *start) 286 { 287 hwaddr size = int128_get64(section->size); 288 hwaddr delta, aligned; 289 290 /* kvm works in page size chunks, but the function may be called 291 with sub-page size and unaligned start address. Pad the start 292 address to next and truncate size to previous page boundary. */ 293 aligned = ROUND_UP(section->offset_within_address_space, 294 qemu_real_host_page_size); 295 delta = aligned - section->offset_within_address_space; 296 *start = aligned; 297 if (delta > size) { 298 return 0; 299 } 300 301 return (size - delta) & qemu_real_host_page_mask; 302 } 303 304 int kvm_physical_memory_addr_from_host(KVMState *s, void *ram, 305 hwaddr *phys_addr) 306 { 307 KVMMemoryListener *kml = &s->memory_listener; 308 int i, ret = 0; 309 310 kvm_slots_lock(kml); 311 for (i = 0; i < s->nr_slots; i++) { 312 KVMSlot *mem = &kml->slots[i]; 313 314 if (ram >= mem->ram && ram < mem->ram + mem->memory_size) { 315 *phys_addr = mem->start_addr + (ram - mem->ram); 316 ret = 1; 317 break; 318 } 319 } 320 kvm_slots_unlock(kml); 321 322 return ret; 323 } 324 325 static int kvm_set_user_memory_region(KVMMemoryListener *kml, KVMSlot *slot, bool new) 326 { 327 KVMState *s = kvm_state; 328 struct kvm_userspace_memory_region mem; 329 int ret; 330 331 mem.slot = slot->slot | (kml->as_id << 16); 332 mem.guest_phys_addr = slot->start_addr; 333 mem.userspace_addr = (unsigned long)slot->ram; 334 mem.flags = slot->flags; 335 336 if (slot->memory_size && !new && (mem.flags ^ slot->old_flags) & KVM_MEM_READONLY) { 337 /* Set the slot size to 0 before setting the slot to the desired 338 * value. This is needed based on KVM commit 75d61fbc. */ 339 mem.memory_size = 0; 340 ret = kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem); 341 if (ret < 0) { 342 goto err; 343 } 344 } 345 mem.memory_size = slot->memory_size; 346 ret = kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem); 347 slot->old_flags = mem.flags; 348 err: 349 trace_kvm_set_user_memory(mem.slot, mem.flags, mem.guest_phys_addr, 350 mem.memory_size, mem.userspace_addr, ret); 351 if (ret < 0) { 352 error_report("%s: KVM_SET_USER_MEMORY_REGION failed, slot=%d," 353 " start=0x%" PRIx64 ", size=0x%" PRIx64 ": %s", 354 __func__, mem.slot, slot->start_addr, 355 (uint64_t)mem.memory_size, strerror(errno)); 356 } 357 return ret; 358 } 359 360 static int do_kvm_destroy_vcpu(CPUState *cpu) 361 { 362 KVMState *s = kvm_state; 363 long mmap_size; 364 struct KVMParkedVcpu *vcpu = NULL; 365 int ret = 0; 366 367 DPRINTF("kvm_destroy_vcpu\n"); 368 369 ret = kvm_arch_destroy_vcpu(cpu); 370 if (ret < 0) { 371 goto err; 372 } 373 374 mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0); 375 if (mmap_size < 0) { 376 ret = mmap_size; 377 DPRINTF("KVM_GET_VCPU_MMAP_SIZE failed\n"); 378 goto err; 379 } 380 381 ret = munmap(cpu->kvm_run, mmap_size); 382 if (ret < 0) { 383 goto err; 384 } 385 386 vcpu = g_malloc0(sizeof(*vcpu)); 387 vcpu->vcpu_id = kvm_arch_vcpu_id(cpu); 388 vcpu->kvm_fd = cpu->kvm_fd; 389 QLIST_INSERT_HEAD(&kvm_state->kvm_parked_vcpus, vcpu, node); 390 err: 391 return ret; 392 } 393 394 void kvm_destroy_vcpu(CPUState *cpu) 395 { 396 if (do_kvm_destroy_vcpu(cpu) < 0) { 397 error_report("kvm_destroy_vcpu failed"); 398 exit(EXIT_FAILURE); 399 } 400 } 401 402 static int kvm_get_vcpu(KVMState *s, unsigned long vcpu_id) 403 { 404 struct KVMParkedVcpu *cpu; 405 406 QLIST_FOREACH(cpu, &s->kvm_parked_vcpus, node) { 407 if (cpu->vcpu_id == vcpu_id) { 408 int kvm_fd; 409 410 QLIST_REMOVE(cpu, node); 411 kvm_fd = cpu->kvm_fd; 412 g_free(cpu); 413 return kvm_fd; 414 } 415 } 416 417 return kvm_vm_ioctl(s, KVM_CREATE_VCPU, (void *)vcpu_id); 418 } 419 420 int kvm_init_vcpu(CPUState *cpu, Error **errp) 421 { 422 KVMState *s = kvm_state; 423 long mmap_size; 424 int ret; 425 426 trace_kvm_init_vcpu(cpu->cpu_index, kvm_arch_vcpu_id(cpu)); 427 428 ret = kvm_get_vcpu(s, kvm_arch_vcpu_id(cpu)); 429 if (ret < 0) { 430 error_setg_errno(errp, -ret, "kvm_init_vcpu: kvm_get_vcpu failed (%lu)", 431 kvm_arch_vcpu_id(cpu)); 432 goto err; 433 } 434 435 cpu->kvm_fd = ret; 436 cpu->kvm_state = s; 437 cpu->vcpu_dirty = true; 438 439 mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0); 440 if (mmap_size < 0) { 441 ret = mmap_size; 442 error_setg_errno(errp, -mmap_size, 443 "kvm_init_vcpu: KVM_GET_VCPU_MMAP_SIZE failed"); 444 goto err; 445 } 446 447 cpu->kvm_run = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED, 448 cpu->kvm_fd, 0); 449 if (cpu->kvm_run == MAP_FAILED) { 450 ret = -errno; 451 error_setg_errno(errp, ret, 452 "kvm_init_vcpu: mmap'ing vcpu state failed (%lu)", 453 kvm_arch_vcpu_id(cpu)); 454 goto err; 455 } 456 457 if (s->coalesced_mmio && !s->coalesced_mmio_ring) { 458 s->coalesced_mmio_ring = 459 (void *)cpu->kvm_run + s->coalesced_mmio * PAGE_SIZE; 460 } 461 462 ret = kvm_arch_init_vcpu(cpu); 463 if (ret < 0) { 464 error_setg_errno(errp, -ret, 465 "kvm_init_vcpu: kvm_arch_init_vcpu failed (%lu)", 466 kvm_arch_vcpu_id(cpu)); 467 } 468 err: 469 return ret; 470 } 471 472 /* 473 * dirty pages logging control 474 */ 475 476 static int kvm_mem_flags(MemoryRegion *mr) 477 { 478 bool readonly = mr->readonly || memory_region_is_romd(mr); 479 int flags = 0; 480 481 if (memory_region_get_dirty_log_mask(mr) != 0) { 482 flags |= KVM_MEM_LOG_DIRTY_PAGES; 483 } 484 if (readonly && kvm_readonly_mem_allowed) { 485 flags |= KVM_MEM_READONLY; 486 } 487 return flags; 488 } 489 490 /* Called with KVMMemoryListener.slots_lock held */ 491 static int kvm_slot_update_flags(KVMMemoryListener *kml, KVMSlot *mem, 492 MemoryRegion *mr) 493 { 494 mem->flags = kvm_mem_flags(mr); 495 496 /* If nothing changed effectively, no need to issue ioctl */ 497 if (mem->flags == mem->old_flags) { 498 return 0; 499 } 500 501 return kvm_set_user_memory_region(kml, mem, false); 502 } 503 504 static int kvm_section_update_flags(KVMMemoryListener *kml, 505 MemoryRegionSection *section) 506 { 507 hwaddr start_addr, size, slot_size; 508 KVMSlot *mem; 509 int ret = 0; 510 511 size = kvm_align_section(section, &start_addr); 512 if (!size) { 513 return 0; 514 } 515 516 kvm_slots_lock(kml); 517 518 while (size && !ret) { 519 slot_size = MIN(kvm_max_slot_size, size); 520 mem = kvm_lookup_matching_slot(kml, start_addr, slot_size); 521 if (!mem) { 522 /* We don't have a slot if we want to trap every access. */ 523 goto out; 524 } 525 526 ret = kvm_slot_update_flags(kml, mem, section->mr); 527 start_addr += slot_size; 528 size -= slot_size; 529 } 530 531 out: 532 kvm_slots_unlock(kml); 533 return ret; 534 } 535 536 static void kvm_log_start(MemoryListener *listener, 537 MemoryRegionSection *section, 538 int old, int new) 539 { 540 KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener); 541 int r; 542 543 if (old != 0) { 544 return; 545 } 546 547 r = kvm_section_update_flags(kml, section); 548 if (r < 0) { 549 abort(); 550 } 551 } 552 553 static void kvm_log_stop(MemoryListener *listener, 554 MemoryRegionSection *section, 555 int old, int new) 556 { 557 KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener); 558 int r; 559 560 if (new != 0) { 561 return; 562 } 563 564 r = kvm_section_update_flags(kml, section); 565 if (r < 0) { 566 abort(); 567 } 568 } 569 570 /* get kvm's dirty pages bitmap and update qemu's */ 571 static int kvm_get_dirty_pages_log_range(MemoryRegionSection *section, 572 unsigned long *bitmap) 573 { 574 ram_addr_t start = section->offset_within_region + 575 memory_region_get_ram_addr(section->mr); 576 ram_addr_t pages = int128_get64(section->size) / qemu_real_host_page_size; 577 578 cpu_physical_memory_set_dirty_lebitmap(bitmap, start, pages); 579 return 0; 580 } 581 582 #define ALIGN(x, y) (((x)+(y)-1) & ~((y)-1)) 583 584 /* Allocate the dirty bitmap for a slot */ 585 static void kvm_memslot_init_dirty_bitmap(KVMSlot *mem) 586 { 587 /* 588 * XXX bad kernel interface alert 589 * For dirty bitmap, kernel allocates array of size aligned to 590 * bits-per-long. But for case when the kernel is 64bits and 591 * the userspace is 32bits, userspace can't align to the same 592 * bits-per-long, since sizeof(long) is different between kernel 593 * and user space. This way, userspace will provide buffer which 594 * may be 4 bytes less than the kernel will use, resulting in 595 * userspace memory corruption (which is not detectable by valgrind 596 * too, in most cases). 597 * So for now, let's align to 64 instead of HOST_LONG_BITS here, in 598 * a hope that sizeof(long) won't become >8 any time soon. 599 * 600 * Note: the granule of kvm dirty log is qemu_real_host_page_size. 601 * And mem->memory_size is aligned to it (otherwise this mem can't 602 * be registered to KVM). 603 */ 604 hwaddr bitmap_size = ALIGN(mem->memory_size / qemu_real_host_page_size, 605 /*HOST_LONG_BITS*/ 64) / 8; 606 mem->dirty_bmap = g_malloc0(bitmap_size); 607 } 608 609 /** 610 * kvm_physical_sync_dirty_bitmap - Sync dirty bitmap from kernel space 611 * 612 * This function will first try to fetch dirty bitmap from the kernel, 613 * and then updates qemu's dirty bitmap. 614 * 615 * NOTE: caller must be with kml->slots_lock held. 616 * 617 * @kml: the KVM memory listener object 618 * @section: the memory section to sync the dirty bitmap with 619 */ 620 static int kvm_physical_sync_dirty_bitmap(KVMMemoryListener *kml, 621 MemoryRegionSection *section) 622 { 623 KVMState *s = kvm_state; 624 struct kvm_dirty_log d = {}; 625 KVMSlot *mem; 626 hwaddr start_addr, size; 627 hwaddr slot_size, slot_offset = 0; 628 int ret = 0; 629 630 size = kvm_align_section(section, &start_addr); 631 while (size) { 632 MemoryRegionSection subsection = *section; 633 634 slot_size = MIN(kvm_max_slot_size, size); 635 mem = kvm_lookup_matching_slot(kml, start_addr, slot_size); 636 if (!mem) { 637 /* We don't have a slot if we want to trap every access. */ 638 goto out; 639 } 640 641 if (!mem->dirty_bmap) { 642 /* Allocate on the first log_sync, once and for all */ 643 kvm_memslot_init_dirty_bitmap(mem); 644 } 645 646 d.dirty_bitmap = mem->dirty_bmap; 647 d.slot = mem->slot | (kml->as_id << 16); 648 ret = kvm_vm_ioctl(s, KVM_GET_DIRTY_LOG, &d); 649 if (ret == -ENOENT) { 650 /* kernel does not have dirty bitmap in this slot */ 651 ret = 0; 652 } else if (ret < 0) { 653 error_report("ioctl KVM_GET_DIRTY_LOG failed: %d", errno); 654 goto out; 655 } else { 656 subsection.offset_within_region += slot_offset; 657 subsection.size = int128_make64(slot_size); 658 kvm_get_dirty_pages_log_range(&subsection, d.dirty_bitmap); 659 } 660 661 slot_offset += slot_size; 662 start_addr += slot_size; 663 size -= slot_size; 664 } 665 out: 666 return ret; 667 } 668 669 /* Alignment requirement for KVM_CLEAR_DIRTY_LOG - 64 pages */ 670 #define KVM_CLEAR_LOG_SHIFT 6 671 #define KVM_CLEAR_LOG_ALIGN (qemu_real_host_page_size << KVM_CLEAR_LOG_SHIFT) 672 #define KVM_CLEAR_LOG_MASK (-KVM_CLEAR_LOG_ALIGN) 673 674 static int kvm_log_clear_one_slot(KVMSlot *mem, int as_id, uint64_t start, 675 uint64_t size) 676 { 677 KVMState *s = kvm_state; 678 uint64_t end, bmap_start, start_delta, bmap_npages; 679 struct kvm_clear_dirty_log d; 680 unsigned long *bmap_clear = NULL, psize = qemu_real_host_page_size; 681 int ret; 682 683 /* 684 * We need to extend either the start or the size or both to 685 * satisfy the KVM interface requirement. Firstly, do the start 686 * page alignment on 64 host pages 687 */ 688 bmap_start = start & KVM_CLEAR_LOG_MASK; 689 start_delta = start - bmap_start; 690 bmap_start /= psize; 691 692 /* 693 * The kernel interface has restriction on the size too, that either: 694 * 695 * (1) the size is 64 host pages aligned (just like the start), or 696 * (2) the size fills up until the end of the KVM memslot. 697 */ 698 bmap_npages = DIV_ROUND_UP(size + start_delta, KVM_CLEAR_LOG_ALIGN) 699 << KVM_CLEAR_LOG_SHIFT; 700 end = mem->memory_size / psize; 701 if (bmap_npages > end - bmap_start) { 702 bmap_npages = end - bmap_start; 703 } 704 start_delta /= psize; 705 706 /* 707 * Prepare the bitmap to clear dirty bits. Here we must guarantee 708 * that we won't clear any unknown dirty bits otherwise we might 709 * accidentally clear some set bits which are not yet synced from 710 * the kernel into QEMU's bitmap, then we'll lose track of the 711 * guest modifications upon those pages (which can directly lead 712 * to guest data loss or panic after migration). 713 * 714 * Layout of the KVMSlot.dirty_bmap: 715 * 716 * |<-------- bmap_npages -----------..>| 717 * [1] 718 * start_delta size 719 * |----------------|-------------|------------------|------------| 720 * ^ ^ ^ ^ 721 * | | | | 722 * start bmap_start (start) end 723 * of memslot of memslot 724 * 725 * [1] bmap_npages can be aligned to either 64 pages or the end of slot 726 */ 727 728 assert(bmap_start % BITS_PER_LONG == 0); 729 /* We should never do log_clear before log_sync */ 730 assert(mem->dirty_bmap); 731 if (start_delta || bmap_npages - size / psize) { 732 /* Slow path - we need to manipulate a temp bitmap */ 733 bmap_clear = bitmap_new(bmap_npages); 734 bitmap_copy_with_src_offset(bmap_clear, mem->dirty_bmap, 735 bmap_start, start_delta + size / psize); 736 /* 737 * We need to fill the holes at start because that was not 738 * specified by the caller and we extended the bitmap only for 739 * 64 pages alignment 740 */ 741 bitmap_clear(bmap_clear, 0, start_delta); 742 d.dirty_bitmap = bmap_clear; 743 } else { 744 /* 745 * Fast path - both start and size align well with BITS_PER_LONG 746 * (or the end of memory slot) 747 */ 748 d.dirty_bitmap = mem->dirty_bmap + BIT_WORD(bmap_start); 749 } 750 751 d.first_page = bmap_start; 752 /* It should never overflow. If it happens, say something */ 753 assert(bmap_npages <= UINT32_MAX); 754 d.num_pages = bmap_npages; 755 d.slot = mem->slot | (as_id << 16); 756 757 ret = kvm_vm_ioctl(s, KVM_CLEAR_DIRTY_LOG, &d); 758 if (ret < 0 && ret != -ENOENT) { 759 error_report("%s: KVM_CLEAR_DIRTY_LOG failed, slot=%d, " 760 "start=0x%"PRIx64", size=0x%"PRIx32", errno=%d", 761 __func__, d.slot, (uint64_t)d.first_page, 762 (uint32_t)d.num_pages, ret); 763 } else { 764 ret = 0; 765 trace_kvm_clear_dirty_log(d.slot, d.first_page, d.num_pages); 766 } 767 768 /* 769 * After we have updated the remote dirty bitmap, we update the 770 * cached bitmap as well for the memslot, then if another user 771 * clears the same region we know we shouldn't clear it again on 772 * the remote otherwise it's data loss as well. 773 */ 774 bitmap_clear(mem->dirty_bmap, bmap_start + start_delta, 775 size / psize); 776 /* This handles the NULL case well */ 777 g_free(bmap_clear); 778 return ret; 779 } 780 781 782 /** 783 * kvm_physical_log_clear - Clear the kernel's dirty bitmap for range 784 * 785 * NOTE: this will be a no-op if we haven't enabled manual dirty log 786 * protection in the host kernel because in that case this operation 787 * will be done within log_sync(). 788 * 789 * @kml: the kvm memory listener 790 * @section: the memory range to clear dirty bitmap 791 */ 792 static int kvm_physical_log_clear(KVMMemoryListener *kml, 793 MemoryRegionSection *section) 794 { 795 KVMState *s = kvm_state; 796 uint64_t start, size, offset, count; 797 KVMSlot *mem; 798 int ret = 0, i; 799 800 if (!s->manual_dirty_log_protect) { 801 /* No need to do explicit clear */ 802 return ret; 803 } 804 805 start = section->offset_within_address_space; 806 size = int128_get64(section->size); 807 808 if (!size) { 809 /* Nothing more we can do... */ 810 return ret; 811 } 812 813 kvm_slots_lock(kml); 814 815 for (i = 0; i < s->nr_slots; i++) { 816 mem = &kml->slots[i]; 817 /* Discard slots that are empty or do not overlap the section */ 818 if (!mem->memory_size || 819 mem->start_addr > start + size - 1 || 820 start > mem->start_addr + mem->memory_size - 1) { 821 continue; 822 } 823 824 if (start >= mem->start_addr) { 825 /* The slot starts before section or is aligned to it. */ 826 offset = start - mem->start_addr; 827 count = MIN(mem->memory_size - offset, size); 828 } else { 829 /* The slot starts after section. */ 830 offset = 0; 831 count = MIN(mem->memory_size, size - (mem->start_addr - start)); 832 } 833 ret = kvm_log_clear_one_slot(mem, kml->as_id, offset, count); 834 if (ret < 0) { 835 break; 836 } 837 } 838 839 kvm_slots_unlock(kml); 840 841 return ret; 842 } 843 844 static void kvm_coalesce_mmio_region(MemoryListener *listener, 845 MemoryRegionSection *secion, 846 hwaddr start, hwaddr size) 847 { 848 KVMState *s = kvm_state; 849 850 if (s->coalesced_mmio) { 851 struct kvm_coalesced_mmio_zone zone; 852 853 zone.addr = start; 854 zone.size = size; 855 zone.pad = 0; 856 857 (void)kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone); 858 } 859 } 860 861 static void kvm_uncoalesce_mmio_region(MemoryListener *listener, 862 MemoryRegionSection *secion, 863 hwaddr start, hwaddr size) 864 { 865 KVMState *s = kvm_state; 866 867 if (s->coalesced_mmio) { 868 struct kvm_coalesced_mmio_zone zone; 869 870 zone.addr = start; 871 zone.size = size; 872 zone.pad = 0; 873 874 (void)kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone); 875 } 876 } 877 878 static void kvm_coalesce_pio_add(MemoryListener *listener, 879 MemoryRegionSection *section, 880 hwaddr start, hwaddr size) 881 { 882 KVMState *s = kvm_state; 883 884 if (s->coalesced_pio) { 885 struct kvm_coalesced_mmio_zone zone; 886 887 zone.addr = start; 888 zone.size = size; 889 zone.pio = 1; 890 891 (void)kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone); 892 } 893 } 894 895 static void kvm_coalesce_pio_del(MemoryListener *listener, 896 MemoryRegionSection *section, 897 hwaddr start, hwaddr size) 898 { 899 KVMState *s = kvm_state; 900 901 if (s->coalesced_pio) { 902 struct kvm_coalesced_mmio_zone zone; 903 904 zone.addr = start; 905 zone.size = size; 906 zone.pio = 1; 907 908 (void)kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone); 909 } 910 } 911 912 static MemoryListener kvm_coalesced_pio_listener = { 913 .coalesced_io_add = kvm_coalesce_pio_add, 914 .coalesced_io_del = kvm_coalesce_pio_del, 915 }; 916 917 int kvm_check_extension(KVMState *s, unsigned int extension) 918 { 919 int ret; 920 921 ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, extension); 922 if (ret < 0) { 923 ret = 0; 924 } 925 926 return ret; 927 } 928 929 int kvm_vm_check_extension(KVMState *s, unsigned int extension) 930 { 931 int ret; 932 933 ret = kvm_vm_ioctl(s, KVM_CHECK_EXTENSION, extension); 934 if (ret < 0) { 935 /* VM wide version not implemented, use global one instead */ 936 ret = kvm_check_extension(s, extension); 937 } 938 939 return ret; 940 } 941 942 typedef struct HWPoisonPage { 943 ram_addr_t ram_addr; 944 QLIST_ENTRY(HWPoisonPage) list; 945 } HWPoisonPage; 946 947 static QLIST_HEAD(, HWPoisonPage) hwpoison_page_list = 948 QLIST_HEAD_INITIALIZER(hwpoison_page_list); 949 950 static void kvm_unpoison_all(void *param) 951 { 952 HWPoisonPage *page, *next_page; 953 954 QLIST_FOREACH_SAFE(page, &hwpoison_page_list, list, next_page) { 955 QLIST_REMOVE(page, list); 956 qemu_ram_remap(page->ram_addr, TARGET_PAGE_SIZE); 957 g_free(page); 958 } 959 } 960 961 void kvm_hwpoison_page_add(ram_addr_t ram_addr) 962 { 963 HWPoisonPage *page; 964 965 QLIST_FOREACH(page, &hwpoison_page_list, list) { 966 if (page->ram_addr == ram_addr) { 967 return; 968 } 969 } 970 page = g_new(HWPoisonPage, 1); 971 page->ram_addr = ram_addr; 972 QLIST_INSERT_HEAD(&hwpoison_page_list, page, list); 973 } 974 975 static uint32_t adjust_ioeventfd_endianness(uint32_t val, uint32_t size) 976 { 977 #if defined(HOST_WORDS_BIGENDIAN) != defined(TARGET_WORDS_BIGENDIAN) 978 /* The kernel expects ioeventfd values in HOST_WORDS_BIGENDIAN 979 * endianness, but the memory core hands them in target endianness. 980 * For example, PPC is always treated as big-endian even if running 981 * on KVM and on PPC64LE. Correct here. 982 */ 983 switch (size) { 984 case 2: 985 val = bswap16(val); 986 break; 987 case 4: 988 val = bswap32(val); 989 break; 990 } 991 #endif 992 return val; 993 } 994 995 static int kvm_set_ioeventfd_mmio(int fd, hwaddr addr, uint32_t val, 996 bool assign, uint32_t size, bool datamatch) 997 { 998 int ret; 999 struct kvm_ioeventfd iofd = { 1000 .datamatch = datamatch ? adjust_ioeventfd_endianness(val, size) : 0, 1001 .addr = addr, 1002 .len = size, 1003 .flags = 0, 1004 .fd = fd, 1005 }; 1006 1007 trace_kvm_set_ioeventfd_mmio(fd, (uint64_t)addr, val, assign, size, 1008 datamatch); 1009 if (!kvm_enabled()) { 1010 return -ENOSYS; 1011 } 1012 1013 if (datamatch) { 1014 iofd.flags |= KVM_IOEVENTFD_FLAG_DATAMATCH; 1015 } 1016 if (!assign) { 1017 iofd.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN; 1018 } 1019 1020 ret = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &iofd); 1021 1022 if (ret < 0) { 1023 return -errno; 1024 } 1025 1026 return 0; 1027 } 1028 1029 static int kvm_set_ioeventfd_pio(int fd, uint16_t addr, uint16_t val, 1030 bool assign, uint32_t size, bool datamatch) 1031 { 1032 struct kvm_ioeventfd kick = { 1033 .datamatch = datamatch ? adjust_ioeventfd_endianness(val, size) : 0, 1034 .addr = addr, 1035 .flags = KVM_IOEVENTFD_FLAG_PIO, 1036 .len = size, 1037 .fd = fd, 1038 }; 1039 int r; 1040 trace_kvm_set_ioeventfd_pio(fd, addr, val, assign, size, datamatch); 1041 if (!kvm_enabled()) { 1042 return -ENOSYS; 1043 } 1044 if (datamatch) { 1045 kick.flags |= KVM_IOEVENTFD_FLAG_DATAMATCH; 1046 } 1047 if (!assign) { 1048 kick.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN; 1049 } 1050 r = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &kick); 1051 if (r < 0) { 1052 return r; 1053 } 1054 return 0; 1055 } 1056 1057 1058 static int kvm_check_many_ioeventfds(void) 1059 { 1060 /* Userspace can use ioeventfd for io notification. This requires a host 1061 * that supports eventfd(2) and an I/O thread; since eventfd does not 1062 * support SIGIO it cannot interrupt the vcpu. 1063 * 1064 * Older kernels have a 6 device limit on the KVM io bus. Find out so we 1065 * can avoid creating too many ioeventfds. 1066 */ 1067 #if defined(CONFIG_EVENTFD) 1068 int ioeventfds[7]; 1069 int i, ret = 0; 1070 for (i = 0; i < ARRAY_SIZE(ioeventfds); i++) { 1071 ioeventfds[i] = eventfd(0, EFD_CLOEXEC); 1072 if (ioeventfds[i] < 0) { 1073 break; 1074 } 1075 ret = kvm_set_ioeventfd_pio(ioeventfds[i], 0, i, true, 2, true); 1076 if (ret < 0) { 1077 close(ioeventfds[i]); 1078 break; 1079 } 1080 } 1081 1082 /* Decide whether many devices are supported or not */ 1083 ret = i == ARRAY_SIZE(ioeventfds); 1084 1085 while (i-- > 0) { 1086 kvm_set_ioeventfd_pio(ioeventfds[i], 0, i, false, 2, true); 1087 close(ioeventfds[i]); 1088 } 1089 return ret; 1090 #else 1091 return 0; 1092 #endif 1093 } 1094 1095 static const KVMCapabilityInfo * 1096 kvm_check_extension_list(KVMState *s, const KVMCapabilityInfo *list) 1097 { 1098 while (list->name) { 1099 if (!kvm_check_extension(s, list->value)) { 1100 return list; 1101 } 1102 list++; 1103 } 1104 return NULL; 1105 } 1106 1107 void kvm_set_max_memslot_size(hwaddr max_slot_size) 1108 { 1109 g_assert( 1110 ROUND_UP(max_slot_size, qemu_real_host_page_size) == max_slot_size 1111 ); 1112 kvm_max_slot_size = max_slot_size; 1113 } 1114 1115 static void kvm_set_phys_mem(KVMMemoryListener *kml, 1116 MemoryRegionSection *section, bool add) 1117 { 1118 KVMSlot *mem; 1119 int err; 1120 MemoryRegion *mr = section->mr; 1121 bool writeable = !mr->readonly && !mr->rom_device; 1122 hwaddr start_addr, size, slot_size; 1123 void *ram; 1124 1125 if (!memory_region_is_ram(mr)) { 1126 if (writeable || !kvm_readonly_mem_allowed) { 1127 return; 1128 } else if (!mr->romd_mode) { 1129 /* If the memory device is not in romd_mode, then we actually want 1130 * to remove the kvm memory slot so all accesses will trap. */ 1131 add = false; 1132 } 1133 } 1134 1135 size = kvm_align_section(section, &start_addr); 1136 if (!size) { 1137 return; 1138 } 1139 1140 /* use aligned delta to align the ram address */ 1141 ram = memory_region_get_ram_ptr(mr) + section->offset_within_region + 1142 (start_addr - section->offset_within_address_space); 1143 1144 kvm_slots_lock(kml); 1145 1146 if (!add) { 1147 do { 1148 slot_size = MIN(kvm_max_slot_size, size); 1149 mem = kvm_lookup_matching_slot(kml, start_addr, slot_size); 1150 if (!mem) { 1151 goto out; 1152 } 1153 if (mem->flags & KVM_MEM_LOG_DIRTY_PAGES) { 1154 kvm_physical_sync_dirty_bitmap(kml, section); 1155 } 1156 1157 /* unregister the slot */ 1158 g_free(mem->dirty_bmap); 1159 mem->dirty_bmap = NULL; 1160 mem->memory_size = 0; 1161 mem->flags = 0; 1162 err = kvm_set_user_memory_region(kml, mem, false); 1163 if (err) { 1164 fprintf(stderr, "%s: error unregistering slot: %s\n", 1165 __func__, strerror(-err)); 1166 abort(); 1167 } 1168 start_addr += slot_size; 1169 size -= slot_size; 1170 } while (size); 1171 goto out; 1172 } 1173 1174 /* register the new slot */ 1175 do { 1176 slot_size = MIN(kvm_max_slot_size, size); 1177 mem = kvm_alloc_slot(kml); 1178 mem->memory_size = slot_size; 1179 mem->start_addr = start_addr; 1180 mem->ram = ram; 1181 mem->flags = kvm_mem_flags(mr); 1182 1183 if (mem->flags & KVM_MEM_LOG_DIRTY_PAGES) { 1184 /* 1185 * Reallocate the bmap; it means it doesn't disappear in 1186 * middle of a migrate. 1187 */ 1188 kvm_memslot_init_dirty_bitmap(mem); 1189 } 1190 err = kvm_set_user_memory_region(kml, mem, true); 1191 if (err) { 1192 fprintf(stderr, "%s: error registering slot: %s\n", __func__, 1193 strerror(-err)); 1194 abort(); 1195 } 1196 start_addr += slot_size; 1197 ram += slot_size; 1198 size -= slot_size; 1199 } while (size); 1200 1201 out: 1202 kvm_slots_unlock(kml); 1203 } 1204 1205 static void kvm_region_add(MemoryListener *listener, 1206 MemoryRegionSection *section) 1207 { 1208 KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener); 1209 1210 memory_region_ref(section->mr); 1211 kvm_set_phys_mem(kml, section, true); 1212 } 1213 1214 static void kvm_region_del(MemoryListener *listener, 1215 MemoryRegionSection *section) 1216 { 1217 KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener); 1218 1219 kvm_set_phys_mem(kml, section, false); 1220 memory_region_unref(section->mr); 1221 } 1222 1223 static void kvm_log_sync(MemoryListener *listener, 1224 MemoryRegionSection *section) 1225 { 1226 KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener); 1227 int r; 1228 1229 kvm_slots_lock(kml); 1230 r = kvm_physical_sync_dirty_bitmap(kml, section); 1231 kvm_slots_unlock(kml); 1232 if (r < 0) { 1233 abort(); 1234 } 1235 } 1236 1237 static void kvm_log_clear(MemoryListener *listener, 1238 MemoryRegionSection *section) 1239 { 1240 KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener); 1241 int r; 1242 1243 r = kvm_physical_log_clear(kml, section); 1244 if (r < 0) { 1245 error_report_once("%s: kvm log clear failed: mr=%s " 1246 "offset=%"HWADDR_PRIx" size=%"PRIx64, __func__, 1247 section->mr->name, section->offset_within_region, 1248 int128_get64(section->size)); 1249 abort(); 1250 } 1251 } 1252 1253 static void kvm_mem_ioeventfd_add(MemoryListener *listener, 1254 MemoryRegionSection *section, 1255 bool match_data, uint64_t data, 1256 EventNotifier *e) 1257 { 1258 int fd = event_notifier_get_fd(e); 1259 int r; 1260 1261 r = kvm_set_ioeventfd_mmio(fd, section->offset_within_address_space, 1262 data, true, int128_get64(section->size), 1263 match_data); 1264 if (r < 0) { 1265 fprintf(stderr, "%s: error adding ioeventfd: %s (%d)\n", 1266 __func__, strerror(-r), -r); 1267 abort(); 1268 } 1269 } 1270 1271 static void kvm_mem_ioeventfd_del(MemoryListener *listener, 1272 MemoryRegionSection *section, 1273 bool match_data, uint64_t data, 1274 EventNotifier *e) 1275 { 1276 int fd = event_notifier_get_fd(e); 1277 int r; 1278 1279 r = kvm_set_ioeventfd_mmio(fd, section->offset_within_address_space, 1280 data, false, int128_get64(section->size), 1281 match_data); 1282 if (r < 0) { 1283 fprintf(stderr, "%s: error deleting ioeventfd: %s (%d)\n", 1284 __func__, strerror(-r), -r); 1285 abort(); 1286 } 1287 } 1288 1289 static void kvm_io_ioeventfd_add(MemoryListener *listener, 1290 MemoryRegionSection *section, 1291 bool match_data, uint64_t data, 1292 EventNotifier *e) 1293 { 1294 int fd = event_notifier_get_fd(e); 1295 int r; 1296 1297 r = kvm_set_ioeventfd_pio(fd, section->offset_within_address_space, 1298 data, true, int128_get64(section->size), 1299 match_data); 1300 if (r < 0) { 1301 fprintf(stderr, "%s: error adding ioeventfd: %s (%d)\n", 1302 __func__, strerror(-r), -r); 1303 abort(); 1304 } 1305 } 1306 1307 static void kvm_io_ioeventfd_del(MemoryListener *listener, 1308 MemoryRegionSection *section, 1309 bool match_data, uint64_t data, 1310 EventNotifier *e) 1311 1312 { 1313 int fd = event_notifier_get_fd(e); 1314 int r; 1315 1316 r = kvm_set_ioeventfd_pio(fd, section->offset_within_address_space, 1317 data, false, int128_get64(section->size), 1318 match_data); 1319 if (r < 0) { 1320 fprintf(stderr, "%s: error deleting ioeventfd: %s (%d)\n", 1321 __func__, strerror(-r), -r); 1322 abort(); 1323 } 1324 } 1325 1326 void kvm_memory_listener_register(KVMState *s, KVMMemoryListener *kml, 1327 AddressSpace *as, int as_id) 1328 { 1329 int i; 1330 1331 qemu_mutex_init(&kml->slots_lock); 1332 kml->slots = g_malloc0(s->nr_slots * sizeof(KVMSlot)); 1333 kml->as_id = as_id; 1334 1335 for (i = 0; i < s->nr_slots; i++) { 1336 kml->slots[i].slot = i; 1337 } 1338 1339 kml->listener.region_add = kvm_region_add; 1340 kml->listener.region_del = kvm_region_del; 1341 kml->listener.log_start = kvm_log_start; 1342 kml->listener.log_stop = kvm_log_stop; 1343 kml->listener.log_sync = kvm_log_sync; 1344 kml->listener.log_clear = kvm_log_clear; 1345 kml->listener.priority = 10; 1346 1347 memory_listener_register(&kml->listener, as); 1348 1349 for (i = 0; i < s->nr_as; ++i) { 1350 if (!s->as[i].as) { 1351 s->as[i].as = as; 1352 s->as[i].ml = kml; 1353 break; 1354 } 1355 } 1356 } 1357 1358 static MemoryListener kvm_io_listener = { 1359 .eventfd_add = kvm_io_ioeventfd_add, 1360 .eventfd_del = kvm_io_ioeventfd_del, 1361 .priority = 10, 1362 }; 1363 1364 int kvm_set_irq(KVMState *s, int irq, int level) 1365 { 1366 struct kvm_irq_level event; 1367 int ret; 1368 1369 assert(kvm_async_interrupts_enabled()); 1370 1371 event.level = level; 1372 event.irq = irq; 1373 ret = kvm_vm_ioctl(s, s->irq_set_ioctl, &event); 1374 if (ret < 0) { 1375 perror("kvm_set_irq"); 1376 abort(); 1377 } 1378 1379 return (s->irq_set_ioctl == KVM_IRQ_LINE) ? 1 : event.status; 1380 } 1381 1382 #ifdef KVM_CAP_IRQ_ROUTING 1383 typedef struct KVMMSIRoute { 1384 struct kvm_irq_routing_entry kroute; 1385 QTAILQ_ENTRY(KVMMSIRoute) entry; 1386 } KVMMSIRoute; 1387 1388 static void set_gsi(KVMState *s, unsigned int gsi) 1389 { 1390 set_bit(gsi, s->used_gsi_bitmap); 1391 } 1392 1393 static void clear_gsi(KVMState *s, unsigned int gsi) 1394 { 1395 clear_bit(gsi, s->used_gsi_bitmap); 1396 } 1397 1398 void kvm_init_irq_routing(KVMState *s) 1399 { 1400 int gsi_count, i; 1401 1402 gsi_count = kvm_check_extension(s, KVM_CAP_IRQ_ROUTING) - 1; 1403 if (gsi_count > 0) { 1404 /* Round up so we can search ints using ffs */ 1405 s->used_gsi_bitmap = bitmap_new(gsi_count); 1406 s->gsi_count = gsi_count; 1407 } 1408 1409 s->irq_routes = g_malloc0(sizeof(*s->irq_routes)); 1410 s->nr_allocated_irq_routes = 0; 1411 1412 if (!kvm_direct_msi_allowed) { 1413 for (i = 0; i < KVM_MSI_HASHTAB_SIZE; i++) { 1414 QTAILQ_INIT(&s->msi_hashtab[i]); 1415 } 1416 } 1417 1418 kvm_arch_init_irq_routing(s); 1419 } 1420 1421 void kvm_irqchip_commit_routes(KVMState *s) 1422 { 1423 int ret; 1424 1425 if (kvm_gsi_direct_mapping()) { 1426 return; 1427 } 1428 1429 if (!kvm_gsi_routing_enabled()) { 1430 return; 1431 } 1432 1433 s->irq_routes->flags = 0; 1434 trace_kvm_irqchip_commit_routes(); 1435 ret = kvm_vm_ioctl(s, KVM_SET_GSI_ROUTING, s->irq_routes); 1436 assert(ret == 0); 1437 } 1438 1439 static void kvm_add_routing_entry(KVMState *s, 1440 struct kvm_irq_routing_entry *entry) 1441 { 1442 struct kvm_irq_routing_entry *new; 1443 int n, size; 1444 1445 if (s->irq_routes->nr == s->nr_allocated_irq_routes) { 1446 n = s->nr_allocated_irq_routes * 2; 1447 if (n < 64) { 1448 n = 64; 1449 } 1450 size = sizeof(struct kvm_irq_routing); 1451 size += n * sizeof(*new); 1452 s->irq_routes = g_realloc(s->irq_routes, size); 1453 s->nr_allocated_irq_routes = n; 1454 } 1455 n = s->irq_routes->nr++; 1456 new = &s->irq_routes->entries[n]; 1457 1458 *new = *entry; 1459 1460 set_gsi(s, entry->gsi); 1461 } 1462 1463 static int kvm_update_routing_entry(KVMState *s, 1464 struct kvm_irq_routing_entry *new_entry) 1465 { 1466 struct kvm_irq_routing_entry *entry; 1467 int n; 1468 1469 for (n = 0; n < s->irq_routes->nr; n++) { 1470 entry = &s->irq_routes->entries[n]; 1471 if (entry->gsi != new_entry->gsi) { 1472 continue; 1473 } 1474 1475 if(!memcmp(entry, new_entry, sizeof *entry)) { 1476 return 0; 1477 } 1478 1479 *entry = *new_entry; 1480 1481 return 0; 1482 } 1483 1484 return -ESRCH; 1485 } 1486 1487 void kvm_irqchip_add_irq_route(KVMState *s, int irq, int irqchip, int pin) 1488 { 1489 struct kvm_irq_routing_entry e = {}; 1490 1491 assert(pin < s->gsi_count); 1492 1493 e.gsi = irq; 1494 e.type = KVM_IRQ_ROUTING_IRQCHIP; 1495 e.flags = 0; 1496 e.u.irqchip.irqchip = irqchip; 1497 e.u.irqchip.pin = pin; 1498 kvm_add_routing_entry(s, &e); 1499 } 1500 1501 void kvm_irqchip_release_virq(KVMState *s, int virq) 1502 { 1503 struct kvm_irq_routing_entry *e; 1504 int i; 1505 1506 if (kvm_gsi_direct_mapping()) { 1507 return; 1508 } 1509 1510 for (i = 0; i < s->irq_routes->nr; i++) { 1511 e = &s->irq_routes->entries[i]; 1512 if (e->gsi == virq) { 1513 s->irq_routes->nr--; 1514 *e = s->irq_routes->entries[s->irq_routes->nr]; 1515 } 1516 } 1517 clear_gsi(s, virq); 1518 kvm_arch_release_virq_post(virq); 1519 trace_kvm_irqchip_release_virq(virq); 1520 } 1521 1522 void kvm_irqchip_add_change_notifier(Notifier *n) 1523 { 1524 notifier_list_add(&kvm_irqchip_change_notifiers, n); 1525 } 1526 1527 void kvm_irqchip_remove_change_notifier(Notifier *n) 1528 { 1529 notifier_remove(n); 1530 } 1531 1532 void kvm_irqchip_change_notify(void) 1533 { 1534 notifier_list_notify(&kvm_irqchip_change_notifiers, NULL); 1535 } 1536 1537 static unsigned int kvm_hash_msi(uint32_t data) 1538 { 1539 /* This is optimized for IA32 MSI layout. However, no other arch shall 1540 * repeat the mistake of not providing a direct MSI injection API. */ 1541 return data & 0xff; 1542 } 1543 1544 static void kvm_flush_dynamic_msi_routes(KVMState *s) 1545 { 1546 KVMMSIRoute *route, *next; 1547 unsigned int hash; 1548 1549 for (hash = 0; hash < KVM_MSI_HASHTAB_SIZE; hash++) { 1550 QTAILQ_FOREACH_SAFE(route, &s->msi_hashtab[hash], entry, next) { 1551 kvm_irqchip_release_virq(s, route->kroute.gsi); 1552 QTAILQ_REMOVE(&s->msi_hashtab[hash], route, entry); 1553 g_free(route); 1554 } 1555 } 1556 } 1557 1558 static int kvm_irqchip_get_virq(KVMState *s) 1559 { 1560 int next_virq; 1561 1562 /* 1563 * PIC and IOAPIC share the first 16 GSI numbers, thus the available 1564 * GSI numbers are more than the number of IRQ route. Allocating a GSI 1565 * number can succeed even though a new route entry cannot be added. 1566 * When this happens, flush dynamic MSI entries to free IRQ route entries. 1567 */ 1568 if (!kvm_direct_msi_allowed && s->irq_routes->nr == s->gsi_count) { 1569 kvm_flush_dynamic_msi_routes(s); 1570 } 1571 1572 /* Return the lowest unused GSI in the bitmap */ 1573 next_virq = find_first_zero_bit(s->used_gsi_bitmap, s->gsi_count); 1574 if (next_virq >= s->gsi_count) { 1575 return -ENOSPC; 1576 } else { 1577 return next_virq; 1578 } 1579 } 1580 1581 static KVMMSIRoute *kvm_lookup_msi_route(KVMState *s, MSIMessage msg) 1582 { 1583 unsigned int hash = kvm_hash_msi(msg.data); 1584 KVMMSIRoute *route; 1585 1586 QTAILQ_FOREACH(route, &s->msi_hashtab[hash], entry) { 1587 if (route->kroute.u.msi.address_lo == (uint32_t)msg.address && 1588 route->kroute.u.msi.address_hi == (msg.address >> 32) && 1589 route->kroute.u.msi.data == le32_to_cpu(msg.data)) { 1590 return route; 1591 } 1592 } 1593 return NULL; 1594 } 1595 1596 int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg) 1597 { 1598 struct kvm_msi msi; 1599 KVMMSIRoute *route; 1600 1601 if (kvm_direct_msi_allowed) { 1602 msi.address_lo = (uint32_t)msg.address; 1603 msi.address_hi = msg.address >> 32; 1604 msi.data = le32_to_cpu(msg.data); 1605 msi.flags = 0; 1606 memset(msi.pad, 0, sizeof(msi.pad)); 1607 1608 return kvm_vm_ioctl(s, KVM_SIGNAL_MSI, &msi); 1609 } 1610 1611 route = kvm_lookup_msi_route(s, msg); 1612 if (!route) { 1613 int virq; 1614 1615 virq = kvm_irqchip_get_virq(s); 1616 if (virq < 0) { 1617 return virq; 1618 } 1619 1620 route = g_malloc0(sizeof(KVMMSIRoute)); 1621 route->kroute.gsi = virq; 1622 route->kroute.type = KVM_IRQ_ROUTING_MSI; 1623 route->kroute.flags = 0; 1624 route->kroute.u.msi.address_lo = (uint32_t)msg.address; 1625 route->kroute.u.msi.address_hi = msg.address >> 32; 1626 route->kroute.u.msi.data = le32_to_cpu(msg.data); 1627 1628 kvm_add_routing_entry(s, &route->kroute); 1629 kvm_irqchip_commit_routes(s); 1630 1631 QTAILQ_INSERT_TAIL(&s->msi_hashtab[kvm_hash_msi(msg.data)], route, 1632 entry); 1633 } 1634 1635 assert(route->kroute.type == KVM_IRQ_ROUTING_MSI); 1636 1637 return kvm_set_irq(s, route->kroute.gsi, 1); 1638 } 1639 1640 int kvm_irqchip_add_msi_route(KVMState *s, int vector, PCIDevice *dev) 1641 { 1642 struct kvm_irq_routing_entry kroute = {}; 1643 int virq; 1644 MSIMessage msg = {0, 0}; 1645 1646 if (pci_available && dev) { 1647 msg = pci_get_msi_message(dev, vector); 1648 } 1649 1650 if (kvm_gsi_direct_mapping()) { 1651 return kvm_arch_msi_data_to_gsi(msg.data); 1652 } 1653 1654 if (!kvm_gsi_routing_enabled()) { 1655 return -ENOSYS; 1656 } 1657 1658 virq = kvm_irqchip_get_virq(s); 1659 if (virq < 0) { 1660 return virq; 1661 } 1662 1663 kroute.gsi = virq; 1664 kroute.type = KVM_IRQ_ROUTING_MSI; 1665 kroute.flags = 0; 1666 kroute.u.msi.address_lo = (uint32_t)msg.address; 1667 kroute.u.msi.address_hi = msg.address >> 32; 1668 kroute.u.msi.data = le32_to_cpu(msg.data); 1669 if (pci_available && kvm_msi_devid_required()) { 1670 kroute.flags = KVM_MSI_VALID_DEVID; 1671 kroute.u.msi.devid = pci_requester_id(dev); 1672 } 1673 if (kvm_arch_fixup_msi_route(&kroute, msg.address, msg.data, dev)) { 1674 kvm_irqchip_release_virq(s, virq); 1675 return -EINVAL; 1676 } 1677 1678 trace_kvm_irqchip_add_msi_route(dev ? dev->name : (char *)"N/A", 1679 vector, virq); 1680 1681 kvm_add_routing_entry(s, &kroute); 1682 kvm_arch_add_msi_route_post(&kroute, vector, dev); 1683 kvm_irqchip_commit_routes(s); 1684 1685 return virq; 1686 } 1687 1688 int kvm_irqchip_update_msi_route(KVMState *s, int virq, MSIMessage msg, 1689 PCIDevice *dev) 1690 { 1691 struct kvm_irq_routing_entry kroute = {}; 1692 1693 if (kvm_gsi_direct_mapping()) { 1694 return 0; 1695 } 1696 1697 if (!kvm_irqchip_in_kernel()) { 1698 return -ENOSYS; 1699 } 1700 1701 kroute.gsi = virq; 1702 kroute.type = KVM_IRQ_ROUTING_MSI; 1703 kroute.flags = 0; 1704 kroute.u.msi.address_lo = (uint32_t)msg.address; 1705 kroute.u.msi.address_hi = msg.address >> 32; 1706 kroute.u.msi.data = le32_to_cpu(msg.data); 1707 if (pci_available && kvm_msi_devid_required()) { 1708 kroute.flags = KVM_MSI_VALID_DEVID; 1709 kroute.u.msi.devid = pci_requester_id(dev); 1710 } 1711 if (kvm_arch_fixup_msi_route(&kroute, msg.address, msg.data, dev)) { 1712 return -EINVAL; 1713 } 1714 1715 trace_kvm_irqchip_update_msi_route(virq); 1716 1717 return kvm_update_routing_entry(s, &kroute); 1718 } 1719 1720 static int kvm_irqchip_assign_irqfd(KVMState *s, EventNotifier *event, 1721 EventNotifier *resample, int virq, 1722 bool assign) 1723 { 1724 int fd = event_notifier_get_fd(event); 1725 int rfd = resample ? event_notifier_get_fd(resample) : -1; 1726 1727 struct kvm_irqfd irqfd = { 1728 .fd = fd, 1729 .gsi = virq, 1730 .flags = assign ? 0 : KVM_IRQFD_FLAG_DEASSIGN, 1731 }; 1732 1733 if (rfd != -1) { 1734 assert(assign); 1735 if (kvm_irqchip_is_split()) { 1736 /* 1737 * When the slow irqchip (e.g. IOAPIC) is in the 1738 * userspace, KVM kernel resamplefd will not work because 1739 * the EOI of the interrupt will be delivered to userspace 1740 * instead, so the KVM kernel resamplefd kick will be 1741 * skipped. The userspace here mimics what the kernel 1742 * provides with resamplefd, remember the resamplefd and 1743 * kick it when we receive EOI of this IRQ. 1744 * 1745 * This is hackery because IOAPIC is mostly bypassed 1746 * (except EOI broadcasts) when irqfd is used. However 1747 * this can bring much performance back for split irqchip 1748 * with INTx IRQs (for VFIO, this gives 93% perf of the 1749 * full fast path, which is 46% perf boost comparing to 1750 * the INTx slow path). 1751 */ 1752 kvm_resample_fd_insert(virq, resample); 1753 } else { 1754 irqfd.flags |= KVM_IRQFD_FLAG_RESAMPLE; 1755 irqfd.resamplefd = rfd; 1756 } 1757 } else if (!assign) { 1758 if (kvm_irqchip_is_split()) { 1759 kvm_resample_fd_remove(virq); 1760 } 1761 } 1762 1763 if (!kvm_irqfds_enabled()) { 1764 return -ENOSYS; 1765 } 1766 1767 return kvm_vm_ioctl(s, KVM_IRQFD, &irqfd); 1768 } 1769 1770 int kvm_irqchip_add_adapter_route(KVMState *s, AdapterInfo *adapter) 1771 { 1772 struct kvm_irq_routing_entry kroute = {}; 1773 int virq; 1774 1775 if (!kvm_gsi_routing_enabled()) { 1776 return -ENOSYS; 1777 } 1778 1779 virq = kvm_irqchip_get_virq(s); 1780 if (virq < 0) { 1781 return virq; 1782 } 1783 1784 kroute.gsi = virq; 1785 kroute.type = KVM_IRQ_ROUTING_S390_ADAPTER; 1786 kroute.flags = 0; 1787 kroute.u.adapter.summary_addr = adapter->summary_addr; 1788 kroute.u.adapter.ind_addr = adapter->ind_addr; 1789 kroute.u.adapter.summary_offset = adapter->summary_offset; 1790 kroute.u.adapter.ind_offset = adapter->ind_offset; 1791 kroute.u.adapter.adapter_id = adapter->adapter_id; 1792 1793 kvm_add_routing_entry(s, &kroute); 1794 1795 return virq; 1796 } 1797 1798 int kvm_irqchip_add_hv_sint_route(KVMState *s, uint32_t vcpu, uint32_t sint) 1799 { 1800 struct kvm_irq_routing_entry kroute = {}; 1801 int virq; 1802 1803 if (!kvm_gsi_routing_enabled()) { 1804 return -ENOSYS; 1805 } 1806 if (!kvm_check_extension(s, KVM_CAP_HYPERV_SYNIC)) { 1807 return -ENOSYS; 1808 } 1809 virq = kvm_irqchip_get_virq(s); 1810 if (virq < 0) { 1811 return virq; 1812 } 1813 1814 kroute.gsi = virq; 1815 kroute.type = KVM_IRQ_ROUTING_HV_SINT; 1816 kroute.flags = 0; 1817 kroute.u.hv_sint.vcpu = vcpu; 1818 kroute.u.hv_sint.sint = sint; 1819 1820 kvm_add_routing_entry(s, &kroute); 1821 kvm_irqchip_commit_routes(s); 1822 1823 return virq; 1824 } 1825 1826 #else /* !KVM_CAP_IRQ_ROUTING */ 1827 1828 void kvm_init_irq_routing(KVMState *s) 1829 { 1830 } 1831 1832 void kvm_irqchip_release_virq(KVMState *s, int virq) 1833 { 1834 } 1835 1836 int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg) 1837 { 1838 abort(); 1839 } 1840 1841 int kvm_irqchip_add_msi_route(KVMState *s, int vector, PCIDevice *dev) 1842 { 1843 return -ENOSYS; 1844 } 1845 1846 int kvm_irqchip_add_adapter_route(KVMState *s, AdapterInfo *adapter) 1847 { 1848 return -ENOSYS; 1849 } 1850 1851 int kvm_irqchip_add_hv_sint_route(KVMState *s, uint32_t vcpu, uint32_t sint) 1852 { 1853 return -ENOSYS; 1854 } 1855 1856 static int kvm_irqchip_assign_irqfd(KVMState *s, EventNotifier *event, 1857 EventNotifier *resample, int virq, 1858 bool assign) 1859 { 1860 abort(); 1861 } 1862 1863 int kvm_irqchip_update_msi_route(KVMState *s, int virq, MSIMessage msg) 1864 { 1865 return -ENOSYS; 1866 } 1867 #endif /* !KVM_CAP_IRQ_ROUTING */ 1868 1869 int kvm_irqchip_add_irqfd_notifier_gsi(KVMState *s, EventNotifier *n, 1870 EventNotifier *rn, int virq) 1871 { 1872 return kvm_irqchip_assign_irqfd(s, n, rn, virq, true); 1873 } 1874 1875 int kvm_irqchip_remove_irqfd_notifier_gsi(KVMState *s, EventNotifier *n, 1876 int virq) 1877 { 1878 return kvm_irqchip_assign_irqfd(s, n, NULL, virq, false); 1879 } 1880 1881 int kvm_irqchip_add_irqfd_notifier(KVMState *s, EventNotifier *n, 1882 EventNotifier *rn, qemu_irq irq) 1883 { 1884 gpointer key, gsi; 1885 gboolean found = g_hash_table_lookup_extended(s->gsimap, irq, &key, &gsi); 1886 1887 if (!found) { 1888 return -ENXIO; 1889 } 1890 return kvm_irqchip_add_irqfd_notifier_gsi(s, n, rn, GPOINTER_TO_INT(gsi)); 1891 } 1892 1893 int kvm_irqchip_remove_irqfd_notifier(KVMState *s, EventNotifier *n, 1894 qemu_irq irq) 1895 { 1896 gpointer key, gsi; 1897 gboolean found = g_hash_table_lookup_extended(s->gsimap, irq, &key, &gsi); 1898 1899 if (!found) { 1900 return -ENXIO; 1901 } 1902 return kvm_irqchip_remove_irqfd_notifier_gsi(s, n, GPOINTER_TO_INT(gsi)); 1903 } 1904 1905 void kvm_irqchip_set_qemuirq_gsi(KVMState *s, qemu_irq irq, int gsi) 1906 { 1907 g_hash_table_insert(s->gsimap, irq, GINT_TO_POINTER(gsi)); 1908 } 1909 1910 static void kvm_irqchip_create(KVMState *s) 1911 { 1912 int ret; 1913 1914 assert(s->kernel_irqchip_split != ON_OFF_AUTO_AUTO); 1915 if (kvm_check_extension(s, KVM_CAP_IRQCHIP)) { 1916 ; 1917 } else if (kvm_check_extension(s, KVM_CAP_S390_IRQCHIP)) { 1918 ret = kvm_vm_enable_cap(s, KVM_CAP_S390_IRQCHIP, 0); 1919 if (ret < 0) { 1920 fprintf(stderr, "Enable kernel irqchip failed: %s\n", strerror(-ret)); 1921 exit(1); 1922 } 1923 } else { 1924 return; 1925 } 1926 1927 /* First probe and see if there's a arch-specific hook to create the 1928 * in-kernel irqchip for us */ 1929 ret = kvm_arch_irqchip_create(s); 1930 if (ret == 0) { 1931 if (s->kernel_irqchip_split == ON_OFF_AUTO_ON) { 1932 perror("Split IRQ chip mode not supported."); 1933 exit(1); 1934 } else { 1935 ret = kvm_vm_ioctl(s, KVM_CREATE_IRQCHIP); 1936 } 1937 } 1938 if (ret < 0) { 1939 fprintf(stderr, "Create kernel irqchip failed: %s\n", strerror(-ret)); 1940 exit(1); 1941 } 1942 1943 kvm_kernel_irqchip = true; 1944 /* If we have an in-kernel IRQ chip then we must have asynchronous 1945 * interrupt delivery (though the reverse is not necessarily true) 1946 */ 1947 kvm_async_interrupts_allowed = true; 1948 kvm_halt_in_kernel_allowed = true; 1949 1950 kvm_init_irq_routing(s); 1951 1952 s->gsimap = g_hash_table_new(g_direct_hash, g_direct_equal); 1953 } 1954 1955 /* Find number of supported CPUs using the recommended 1956 * procedure from the kernel API documentation to cope with 1957 * older kernels that may be missing capabilities. 1958 */ 1959 static int kvm_recommended_vcpus(KVMState *s) 1960 { 1961 int ret = kvm_vm_check_extension(s, KVM_CAP_NR_VCPUS); 1962 return (ret) ? ret : 4; 1963 } 1964 1965 static int kvm_max_vcpus(KVMState *s) 1966 { 1967 int ret = kvm_check_extension(s, KVM_CAP_MAX_VCPUS); 1968 return (ret) ? ret : kvm_recommended_vcpus(s); 1969 } 1970 1971 static int kvm_max_vcpu_id(KVMState *s) 1972 { 1973 int ret = kvm_check_extension(s, KVM_CAP_MAX_VCPU_ID); 1974 return (ret) ? ret : kvm_max_vcpus(s); 1975 } 1976 1977 bool kvm_vcpu_id_is_valid(int vcpu_id) 1978 { 1979 KVMState *s = KVM_STATE(current_accel()); 1980 return vcpu_id >= 0 && vcpu_id < kvm_max_vcpu_id(s); 1981 } 1982 1983 static int kvm_init(MachineState *ms) 1984 { 1985 MachineClass *mc = MACHINE_GET_CLASS(ms); 1986 static const char upgrade_note[] = 1987 "Please upgrade to at least kernel 2.6.29 or recent kvm-kmod\n" 1988 "(see http://sourceforge.net/projects/kvm).\n"; 1989 struct { 1990 const char *name; 1991 int num; 1992 } num_cpus[] = { 1993 { "SMP", ms->smp.cpus }, 1994 { "hotpluggable", ms->smp.max_cpus }, 1995 { NULL, } 1996 }, *nc = num_cpus; 1997 int soft_vcpus_limit, hard_vcpus_limit; 1998 KVMState *s; 1999 const KVMCapabilityInfo *missing_cap; 2000 int ret; 2001 int type = 0; 2002 uint64_t dirty_log_manual_caps; 2003 2004 s = KVM_STATE(ms->accelerator); 2005 2006 /* 2007 * On systems where the kernel can support different base page 2008 * sizes, host page size may be different from TARGET_PAGE_SIZE, 2009 * even with KVM. TARGET_PAGE_SIZE is assumed to be the minimum 2010 * page size for the system though. 2011 */ 2012 assert(TARGET_PAGE_SIZE <= qemu_real_host_page_size); 2013 2014 s->sigmask_len = 8; 2015 2016 #ifdef KVM_CAP_SET_GUEST_DEBUG 2017 QTAILQ_INIT(&s->kvm_sw_breakpoints); 2018 #endif 2019 QLIST_INIT(&s->kvm_parked_vcpus); 2020 s->vmfd = -1; 2021 s->fd = qemu_open_old("/dev/kvm", O_RDWR); 2022 if (s->fd == -1) { 2023 fprintf(stderr, "Could not access KVM kernel module: %m\n"); 2024 ret = -errno; 2025 goto err; 2026 } 2027 2028 ret = kvm_ioctl(s, KVM_GET_API_VERSION, 0); 2029 if (ret < KVM_API_VERSION) { 2030 if (ret >= 0) { 2031 ret = -EINVAL; 2032 } 2033 fprintf(stderr, "kvm version too old\n"); 2034 goto err; 2035 } 2036 2037 if (ret > KVM_API_VERSION) { 2038 ret = -EINVAL; 2039 fprintf(stderr, "kvm version not supported\n"); 2040 goto err; 2041 } 2042 2043 kvm_immediate_exit = kvm_check_extension(s, KVM_CAP_IMMEDIATE_EXIT); 2044 s->nr_slots = kvm_check_extension(s, KVM_CAP_NR_MEMSLOTS); 2045 2046 /* If unspecified, use the default value */ 2047 if (!s->nr_slots) { 2048 s->nr_slots = 32; 2049 } 2050 2051 s->nr_as = kvm_check_extension(s, KVM_CAP_MULTI_ADDRESS_SPACE); 2052 if (s->nr_as <= 1) { 2053 s->nr_as = 1; 2054 } 2055 s->as = g_new0(struct KVMAs, s->nr_as); 2056 2057 if (object_property_find(OBJECT(current_machine), "kvm-type")) { 2058 g_autofree char *kvm_type = object_property_get_str(OBJECT(current_machine), 2059 "kvm-type", 2060 &error_abort); 2061 type = mc->kvm_type(ms, kvm_type); 2062 } else if (mc->kvm_type) { 2063 type = mc->kvm_type(ms, NULL); 2064 } 2065 2066 do { 2067 ret = kvm_ioctl(s, KVM_CREATE_VM, type); 2068 } while (ret == -EINTR); 2069 2070 if (ret < 0) { 2071 fprintf(stderr, "ioctl(KVM_CREATE_VM) failed: %d %s\n", -ret, 2072 strerror(-ret)); 2073 2074 #ifdef TARGET_S390X 2075 if (ret == -EINVAL) { 2076 fprintf(stderr, 2077 "Host kernel setup problem detected. Please verify:\n"); 2078 fprintf(stderr, "- for kernels supporting the switch_amode or" 2079 " user_mode parameters, whether\n"); 2080 fprintf(stderr, 2081 " user space is running in primary address space\n"); 2082 fprintf(stderr, 2083 "- for kernels supporting the vm.allocate_pgste sysctl, " 2084 "whether it is enabled\n"); 2085 } 2086 #endif 2087 goto err; 2088 } 2089 2090 s->vmfd = ret; 2091 2092 /* check the vcpu limits */ 2093 soft_vcpus_limit = kvm_recommended_vcpus(s); 2094 hard_vcpus_limit = kvm_max_vcpus(s); 2095 2096 while (nc->name) { 2097 if (nc->num > soft_vcpus_limit) { 2098 warn_report("Number of %s cpus requested (%d) exceeds " 2099 "the recommended cpus supported by KVM (%d)", 2100 nc->name, nc->num, soft_vcpus_limit); 2101 2102 if (nc->num > hard_vcpus_limit) { 2103 fprintf(stderr, "Number of %s cpus requested (%d) exceeds " 2104 "the maximum cpus supported by KVM (%d)\n", 2105 nc->name, nc->num, hard_vcpus_limit); 2106 exit(1); 2107 } 2108 } 2109 nc++; 2110 } 2111 2112 missing_cap = kvm_check_extension_list(s, kvm_required_capabilites); 2113 if (!missing_cap) { 2114 missing_cap = 2115 kvm_check_extension_list(s, kvm_arch_required_capabilities); 2116 } 2117 if (missing_cap) { 2118 ret = -EINVAL; 2119 fprintf(stderr, "kvm does not support %s\n%s", 2120 missing_cap->name, upgrade_note); 2121 goto err; 2122 } 2123 2124 s->coalesced_mmio = kvm_check_extension(s, KVM_CAP_COALESCED_MMIO); 2125 s->coalesced_pio = s->coalesced_mmio && 2126 kvm_check_extension(s, KVM_CAP_COALESCED_PIO); 2127 2128 dirty_log_manual_caps = 2129 kvm_check_extension(s, KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2); 2130 dirty_log_manual_caps &= (KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE | 2131 KVM_DIRTY_LOG_INITIALLY_SET); 2132 s->manual_dirty_log_protect = dirty_log_manual_caps; 2133 if (dirty_log_manual_caps) { 2134 ret = kvm_vm_enable_cap(s, KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2, 0, 2135 dirty_log_manual_caps); 2136 if (ret) { 2137 warn_report("Trying to enable capability %"PRIu64" of " 2138 "KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2 but failed. " 2139 "Falling back to the legacy mode. ", 2140 dirty_log_manual_caps); 2141 s->manual_dirty_log_protect = 0; 2142 } 2143 } 2144 2145 #ifdef KVM_CAP_VCPU_EVENTS 2146 s->vcpu_events = kvm_check_extension(s, KVM_CAP_VCPU_EVENTS); 2147 #endif 2148 2149 s->robust_singlestep = 2150 kvm_check_extension(s, KVM_CAP_X86_ROBUST_SINGLESTEP); 2151 2152 #ifdef KVM_CAP_DEBUGREGS 2153 s->debugregs = kvm_check_extension(s, KVM_CAP_DEBUGREGS); 2154 #endif 2155 2156 s->max_nested_state_len = kvm_check_extension(s, KVM_CAP_NESTED_STATE); 2157 2158 #ifdef KVM_CAP_IRQ_ROUTING 2159 kvm_direct_msi_allowed = (kvm_check_extension(s, KVM_CAP_SIGNAL_MSI) > 0); 2160 #endif 2161 2162 s->intx_set_mask = kvm_check_extension(s, KVM_CAP_PCI_2_3); 2163 2164 s->irq_set_ioctl = KVM_IRQ_LINE; 2165 if (kvm_check_extension(s, KVM_CAP_IRQ_INJECT_STATUS)) { 2166 s->irq_set_ioctl = KVM_IRQ_LINE_STATUS; 2167 } 2168 2169 kvm_readonly_mem_allowed = 2170 (kvm_check_extension(s, KVM_CAP_READONLY_MEM) > 0); 2171 2172 kvm_eventfds_allowed = 2173 (kvm_check_extension(s, KVM_CAP_IOEVENTFD) > 0); 2174 2175 kvm_irqfds_allowed = 2176 (kvm_check_extension(s, KVM_CAP_IRQFD) > 0); 2177 2178 kvm_resamplefds_allowed = 2179 (kvm_check_extension(s, KVM_CAP_IRQFD_RESAMPLE) > 0); 2180 2181 kvm_vm_attributes_allowed = 2182 (kvm_check_extension(s, KVM_CAP_VM_ATTRIBUTES) > 0); 2183 2184 kvm_ioeventfd_any_length_allowed = 2185 (kvm_check_extension(s, KVM_CAP_IOEVENTFD_ANY_LENGTH) > 0); 2186 2187 kvm_state = s; 2188 2189 ret = kvm_arch_init(ms, s); 2190 if (ret < 0) { 2191 goto err; 2192 } 2193 2194 if (s->kernel_irqchip_split == ON_OFF_AUTO_AUTO) { 2195 s->kernel_irqchip_split = mc->default_kernel_irqchip_split ? ON_OFF_AUTO_ON : ON_OFF_AUTO_OFF; 2196 } 2197 2198 qemu_register_reset(kvm_unpoison_all, NULL); 2199 2200 if (s->kernel_irqchip_allowed) { 2201 kvm_irqchip_create(s); 2202 } 2203 2204 if (kvm_eventfds_allowed) { 2205 s->memory_listener.listener.eventfd_add = kvm_mem_ioeventfd_add; 2206 s->memory_listener.listener.eventfd_del = kvm_mem_ioeventfd_del; 2207 } 2208 s->memory_listener.listener.coalesced_io_add = kvm_coalesce_mmio_region; 2209 s->memory_listener.listener.coalesced_io_del = kvm_uncoalesce_mmio_region; 2210 2211 kvm_memory_listener_register(s, &s->memory_listener, 2212 &address_space_memory, 0); 2213 if (kvm_eventfds_allowed) { 2214 memory_listener_register(&kvm_io_listener, 2215 &address_space_io); 2216 } 2217 memory_listener_register(&kvm_coalesced_pio_listener, 2218 &address_space_io); 2219 2220 s->many_ioeventfds = kvm_check_many_ioeventfds(); 2221 2222 s->sync_mmu = !!kvm_vm_check_extension(kvm_state, KVM_CAP_SYNC_MMU); 2223 if (!s->sync_mmu) { 2224 ret = ram_block_discard_disable(true); 2225 assert(!ret); 2226 } 2227 return 0; 2228 2229 err: 2230 assert(ret < 0); 2231 if (s->vmfd >= 0) { 2232 close(s->vmfd); 2233 } 2234 if (s->fd != -1) { 2235 close(s->fd); 2236 } 2237 g_free(s->memory_listener.slots); 2238 2239 return ret; 2240 } 2241 2242 void kvm_set_sigmask_len(KVMState *s, unsigned int sigmask_len) 2243 { 2244 s->sigmask_len = sigmask_len; 2245 } 2246 2247 static void kvm_handle_io(uint16_t port, MemTxAttrs attrs, void *data, int direction, 2248 int size, uint32_t count) 2249 { 2250 int i; 2251 uint8_t *ptr = data; 2252 2253 for (i = 0; i < count; i++) { 2254 address_space_rw(&address_space_io, port, attrs, 2255 ptr, size, 2256 direction == KVM_EXIT_IO_OUT); 2257 ptr += size; 2258 } 2259 } 2260 2261 static int kvm_handle_internal_error(CPUState *cpu, struct kvm_run *run) 2262 { 2263 fprintf(stderr, "KVM internal error. Suberror: %d\n", 2264 run->internal.suberror); 2265 2266 if (kvm_check_extension(kvm_state, KVM_CAP_INTERNAL_ERROR_DATA)) { 2267 int i; 2268 2269 for (i = 0; i < run->internal.ndata; ++i) { 2270 fprintf(stderr, "extra data[%d]: 0x%016"PRIx64"\n", 2271 i, (uint64_t)run->internal.data[i]); 2272 } 2273 } 2274 if (run->internal.suberror == KVM_INTERNAL_ERROR_EMULATION) { 2275 fprintf(stderr, "emulation failure\n"); 2276 if (!kvm_arch_stop_on_emulation_error(cpu)) { 2277 cpu_dump_state(cpu, stderr, CPU_DUMP_CODE); 2278 return EXCP_INTERRUPT; 2279 } 2280 } 2281 /* FIXME: Should trigger a qmp message to let management know 2282 * something went wrong. 2283 */ 2284 return -1; 2285 } 2286 2287 void kvm_flush_coalesced_mmio_buffer(void) 2288 { 2289 KVMState *s = kvm_state; 2290 2291 if (s->coalesced_flush_in_progress) { 2292 return; 2293 } 2294 2295 s->coalesced_flush_in_progress = true; 2296 2297 if (s->coalesced_mmio_ring) { 2298 struct kvm_coalesced_mmio_ring *ring = s->coalesced_mmio_ring; 2299 while (ring->first != ring->last) { 2300 struct kvm_coalesced_mmio *ent; 2301 2302 ent = &ring->coalesced_mmio[ring->first]; 2303 2304 if (ent->pio == 1) { 2305 address_space_write(&address_space_io, ent->phys_addr, 2306 MEMTXATTRS_UNSPECIFIED, ent->data, 2307 ent->len); 2308 } else { 2309 cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len); 2310 } 2311 smp_wmb(); 2312 ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX; 2313 } 2314 } 2315 2316 s->coalesced_flush_in_progress = false; 2317 } 2318 2319 bool kvm_cpu_check_are_resettable(void) 2320 { 2321 return kvm_arch_cpu_check_are_resettable(); 2322 } 2323 2324 static void do_kvm_cpu_synchronize_state(CPUState *cpu, run_on_cpu_data arg) 2325 { 2326 if (!cpu->vcpu_dirty) { 2327 kvm_arch_get_registers(cpu); 2328 cpu->vcpu_dirty = true; 2329 } 2330 } 2331 2332 void kvm_cpu_synchronize_state(CPUState *cpu) 2333 { 2334 if (!cpu->vcpu_dirty) { 2335 run_on_cpu(cpu, do_kvm_cpu_synchronize_state, RUN_ON_CPU_NULL); 2336 } 2337 } 2338 2339 static void do_kvm_cpu_synchronize_post_reset(CPUState *cpu, run_on_cpu_data arg) 2340 { 2341 kvm_arch_put_registers(cpu, KVM_PUT_RESET_STATE); 2342 cpu->vcpu_dirty = false; 2343 } 2344 2345 void kvm_cpu_synchronize_post_reset(CPUState *cpu) 2346 { 2347 run_on_cpu(cpu, do_kvm_cpu_synchronize_post_reset, RUN_ON_CPU_NULL); 2348 } 2349 2350 static void do_kvm_cpu_synchronize_post_init(CPUState *cpu, run_on_cpu_data arg) 2351 { 2352 kvm_arch_put_registers(cpu, KVM_PUT_FULL_STATE); 2353 cpu->vcpu_dirty = false; 2354 } 2355 2356 void kvm_cpu_synchronize_post_init(CPUState *cpu) 2357 { 2358 run_on_cpu(cpu, do_kvm_cpu_synchronize_post_init, RUN_ON_CPU_NULL); 2359 } 2360 2361 static void do_kvm_cpu_synchronize_pre_loadvm(CPUState *cpu, run_on_cpu_data arg) 2362 { 2363 cpu->vcpu_dirty = true; 2364 } 2365 2366 void kvm_cpu_synchronize_pre_loadvm(CPUState *cpu) 2367 { 2368 run_on_cpu(cpu, do_kvm_cpu_synchronize_pre_loadvm, RUN_ON_CPU_NULL); 2369 } 2370 2371 #ifdef KVM_HAVE_MCE_INJECTION 2372 static __thread void *pending_sigbus_addr; 2373 static __thread int pending_sigbus_code; 2374 static __thread bool have_sigbus_pending; 2375 #endif 2376 2377 static void kvm_cpu_kick(CPUState *cpu) 2378 { 2379 qatomic_set(&cpu->kvm_run->immediate_exit, 1); 2380 } 2381 2382 static void kvm_cpu_kick_self(void) 2383 { 2384 if (kvm_immediate_exit) { 2385 kvm_cpu_kick(current_cpu); 2386 } else { 2387 qemu_cpu_kick_self(); 2388 } 2389 } 2390 2391 static void kvm_eat_signals(CPUState *cpu) 2392 { 2393 struct timespec ts = { 0, 0 }; 2394 siginfo_t siginfo; 2395 sigset_t waitset; 2396 sigset_t chkset; 2397 int r; 2398 2399 if (kvm_immediate_exit) { 2400 qatomic_set(&cpu->kvm_run->immediate_exit, 0); 2401 /* Write kvm_run->immediate_exit before the cpu->exit_request 2402 * write in kvm_cpu_exec. 2403 */ 2404 smp_wmb(); 2405 return; 2406 } 2407 2408 sigemptyset(&waitset); 2409 sigaddset(&waitset, SIG_IPI); 2410 2411 do { 2412 r = sigtimedwait(&waitset, &siginfo, &ts); 2413 if (r == -1 && !(errno == EAGAIN || errno == EINTR)) { 2414 perror("sigtimedwait"); 2415 exit(1); 2416 } 2417 2418 r = sigpending(&chkset); 2419 if (r == -1) { 2420 perror("sigpending"); 2421 exit(1); 2422 } 2423 } while (sigismember(&chkset, SIG_IPI)); 2424 } 2425 2426 int kvm_cpu_exec(CPUState *cpu) 2427 { 2428 struct kvm_run *run = cpu->kvm_run; 2429 int ret, run_ret; 2430 2431 DPRINTF("kvm_cpu_exec()\n"); 2432 2433 if (kvm_arch_process_async_events(cpu)) { 2434 qatomic_set(&cpu->exit_request, 0); 2435 return EXCP_HLT; 2436 } 2437 2438 qemu_mutex_unlock_iothread(); 2439 cpu_exec_start(cpu); 2440 2441 do { 2442 MemTxAttrs attrs; 2443 2444 if (cpu->vcpu_dirty) { 2445 kvm_arch_put_registers(cpu, KVM_PUT_RUNTIME_STATE); 2446 cpu->vcpu_dirty = false; 2447 } 2448 2449 kvm_arch_pre_run(cpu, run); 2450 if (qatomic_read(&cpu->exit_request)) { 2451 DPRINTF("interrupt exit requested\n"); 2452 /* 2453 * KVM requires us to reenter the kernel after IO exits to complete 2454 * instruction emulation. This self-signal will ensure that we 2455 * leave ASAP again. 2456 */ 2457 kvm_cpu_kick_self(); 2458 } 2459 2460 /* Read cpu->exit_request before KVM_RUN reads run->immediate_exit. 2461 * Matching barrier in kvm_eat_signals. 2462 */ 2463 smp_rmb(); 2464 2465 run_ret = kvm_vcpu_ioctl(cpu, KVM_RUN, 0); 2466 2467 attrs = kvm_arch_post_run(cpu, run); 2468 2469 #ifdef KVM_HAVE_MCE_INJECTION 2470 if (unlikely(have_sigbus_pending)) { 2471 qemu_mutex_lock_iothread(); 2472 kvm_arch_on_sigbus_vcpu(cpu, pending_sigbus_code, 2473 pending_sigbus_addr); 2474 have_sigbus_pending = false; 2475 qemu_mutex_unlock_iothread(); 2476 } 2477 #endif 2478 2479 if (run_ret < 0) { 2480 if (run_ret == -EINTR || run_ret == -EAGAIN) { 2481 DPRINTF("io window exit\n"); 2482 kvm_eat_signals(cpu); 2483 ret = EXCP_INTERRUPT; 2484 break; 2485 } 2486 fprintf(stderr, "error: kvm run failed %s\n", 2487 strerror(-run_ret)); 2488 #ifdef TARGET_PPC 2489 if (run_ret == -EBUSY) { 2490 fprintf(stderr, 2491 "This is probably because your SMT is enabled.\n" 2492 "VCPU can only run on primary threads with all " 2493 "secondary threads offline.\n"); 2494 } 2495 #endif 2496 ret = -1; 2497 break; 2498 } 2499 2500 trace_kvm_run_exit(cpu->cpu_index, run->exit_reason); 2501 switch (run->exit_reason) { 2502 case KVM_EXIT_IO: 2503 DPRINTF("handle_io\n"); 2504 /* Called outside BQL */ 2505 kvm_handle_io(run->io.port, attrs, 2506 (uint8_t *)run + run->io.data_offset, 2507 run->io.direction, 2508 run->io.size, 2509 run->io.count); 2510 ret = 0; 2511 break; 2512 case KVM_EXIT_MMIO: 2513 DPRINTF("handle_mmio\n"); 2514 /* Called outside BQL */ 2515 address_space_rw(&address_space_memory, 2516 run->mmio.phys_addr, attrs, 2517 run->mmio.data, 2518 run->mmio.len, 2519 run->mmio.is_write); 2520 ret = 0; 2521 break; 2522 case KVM_EXIT_IRQ_WINDOW_OPEN: 2523 DPRINTF("irq_window_open\n"); 2524 ret = EXCP_INTERRUPT; 2525 break; 2526 case KVM_EXIT_SHUTDOWN: 2527 DPRINTF("shutdown\n"); 2528 qemu_system_reset_request(SHUTDOWN_CAUSE_GUEST_RESET); 2529 ret = EXCP_INTERRUPT; 2530 break; 2531 case KVM_EXIT_UNKNOWN: 2532 fprintf(stderr, "KVM: unknown exit, hardware reason %" PRIx64 "\n", 2533 (uint64_t)run->hw.hardware_exit_reason); 2534 ret = -1; 2535 break; 2536 case KVM_EXIT_INTERNAL_ERROR: 2537 ret = kvm_handle_internal_error(cpu, run); 2538 break; 2539 case KVM_EXIT_SYSTEM_EVENT: 2540 switch (run->system_event.type) { 2541 case KVM_SYSTEM_EVENT_SHUTDOWN: 2542 qemu_system_shutdown_request(SHUTDOWN_CAUSE_GUEST_SHUTDOWN); 2543 ret = EXCP_INTERRUPT; 2544 break; 2545 case KVM_SYSTEM_EVENT_RESET: 2546 qemu_system_reset_request(SHUTDOWN_CAUSE_GUEST_RESET); 2547 ret = EXCP_INTERRUPT; 2548 break; 2549 case KVM_SYSTEM_EVENT_CRASH: 2550 kvm_cpu_synchronize_state(cpu); 2551 qemu_mutex_lock_iothread(); 2552 qemu_system_guest_panicked(cpu_get_crash_info(cpu)); 2553 qemu_mutex_unlock_iothread(); 2554 ret = 0; 2555 break; 2556 default: 2557 DPRINTF("kvm_arch_handle_exit\n"); 2558 ret = kvm_arch_handle_exit(cpu, run); 2559 break; 2560 } 2561 break; 2562 default: 2563 DPRINTF("kvm_arch_handle_exit\n"); 2564 ret = kvm_arch_handle_exit(cpu, run); 2565 break; 2566 } 2567 } while (ret == 0); 2568 2569 cpu_exec_end(cpu); 2570 qemu_mutex_lock_iothread(); 2571 2572 if (ret < 0) { 2573 cpu_dump_state(cpu, stderr, CPU_DUMP_CODE); 2574 vm_stop(RUN_STATE_INTERNAL_ERROR); 2575 } 2576 2577 qatomic_set(&cpu->exit_request, 0); 2578 return ret; 2579 } 2580 2581 int kvm_ioctl(KVMState *s, int type, ...) 2582 { 2583 int ret; 2584 void *arg; 2585 va_list ap; 2586 2587 va_start(ap, type); 2588 arg = va_arg(ap, void *); 2589 va_end(ap); 2590 2591 trace_kvm_ioctl(type, arg); 2592 ret = ioctl(s->fd, type, arg); 2593 if (ret == -1) { 2594 ret = -errno; 2595 } 2596 return ret; 2597 } 2598 2599 int kvm_vm_ioctl(KVMState *s, int type, ...) 2600 { 2601 int ret; 2602 void *arg; 2603 va_list ap; 2604 2605 va_start(ap, type); 2606 arg = va_arg(ap, void *); 2607 va_end(ap); 2608 2609 trace_kvm_vm_ioctl(type, arg); 2610 ret = ioctl(s->vmfd, type, arg); 2611 if (ret == -1) { 2612 ret = -errno; 2613 } 2614 return ret; 2615 } 2616 2617 int kvm_vcpu_ioctl(CPUState *cpu, int type, ...) 2618 { 2619 int ret; 2620 void *arg; 2621 va_list ap; 2622 2623 va_start(ap, type); 2624 arg = va_arg(ap, void *); 2625 va_end(ap); 2626 2627 trace_kvm_vcpu_ioctl(cpu->cpu_index, type, arg); 2628 ret = ioctl(cpu->kvm_fd, type, arg); 2629 if (ret == -1) { 2630 ret = -errno; 2631 } 2632 return ret; 2633 } 2634 2635 int kvm_device_ioctl(int fd, int type, ...) 2636 { 2637 int ret; 2638 void *arg; 2639 va_list ap; 2640 2641 va_start(ap, type); 2642 arg = va_arg(ap, void *); 2643 va_end(ap); 2644 2645 trace_kvm_device_ioctl(fd, type, arg); 2646 ret = ioctl(fd, type, arg); 2647 if (ret == -1) { 2648 ret = -errno; 2649 } 2650 return ret; 2651 } 2652 2653 int kvm_vm_check_attr(KVMState *s, uint32_t group, uint64_t attr) 2654 { 2655 int ret; 2656 struct kvm_device_attr attribute = { 2657 .group = group, 2658 .attr = attr, 2659 }; 2660 2661 if (!kvm_vm_attributes_allowed) { 2662 return 0; 2663 } 2664 2665 ret = kvm_vm_ioctl(s, KVM_HAS_DEVICE_ATTR, &attribute); 2666 /* kvm returns 0 on success for HAS_DEVICE_ATTR */ 2667 return ret ? 0 : 1; 2668 } 2669 2670 int kvm_device_check_attr(int dev_fd, uint32_t group, uint64_t attr) 2671 { 2672 struct kvm_device_attr attribute = { 2673 .group = group, 2674 .attr = attr, 2675 .flags = 0, 2676 }; 2677 2678 return kvm_device_ioctl(dev_fd, KVM_HAS_DEVICE_ATTR, &attribute) ? 0 : 1; 2679 } 2680 2681 int kvm_device_access(int fd, int group, uint64_t attr, 2682 void *val, bool write, Error **errp) 2683 { 2684 struct kvm_device_attr kvmattr; 2685 int err; 2686 2687 kvmattr.flags = 0; 2688 kvmattr.group = group; 2689 kvmattr.attr = attr; 2690 kvmattr.addr = (uintptr_t)val; 2691 2692 err = kvm_device_ioctl(fd, 2693 write ? KVM_SET_DEVICE_ATTR : KVM_GET_DEVICE_ATTR, 2694 &kvmattr); 2695 if (err < 0) { 2696 error_setg_errno(errp, -err, 2697 "KVM_%s_DEVICE_ATTR failed: Group %d " 2698 "attr 0x%016" PRIx64, 2699 write ? "SET" : "GET", group, attr); 2700 } 2701 return err; 2702 } 2703 2704 bool kvm_has_sync_mmu(void) 2705 { 2706 return kvm_state->sync_mmu; 2707 } 2708 2709 int kvm_has_vcpu_events(void) 2710 { 2711 return kvm_state->vcpu_events; 2712 } 2713 2714 int kvm_has_robust_singlestep(void) 2715 { 2716 return kvm_state->robust_singlestep; 2717 } 2718 2719 int kvm_has_debugregs(void) 2720 { 2721 return kvm_state->debugregs; 2722 } 2723 2724 int kvm_max_nested_state_length(void) 2725 { 2726 return kvm_state->max_nested_state_len; 2727 } 2728 2729 int kvm_has_many_ioeventfds(void) 2730 { 2731 if (!kvm_enabled()) { 2732 return 0; 2733 } 2734 return kvm_state->many_ioeventfds; 2735 } 2736 2737 int kvm_has_gsi_routing(void) 2738 { 2739 #ifdef KVM_CAP_IRQ_ROUTING 2740 return kvm_check_extension(kvm_state, KVM_CAP_IRQ_ROUTING); 2741 #else 2742 return false; 2743 #endif 2744 } 2745 2746 int kvm_has_intx_set_mask(void) 2747 { 2748 return kvm_state->intx_set_mask; 2749 } 2750 2751 bool kvm_arm_supports_user_irq(void) 2752 { 2753 return kvm_check_extension(kvm_state, KVM_CAP_ARM_USER_IRQ); 2754 } 2755 2756 #ifdef KVM_CAP_SET_GUEST_DEBUG 2757 struct kvm_sw_breakpoint *kvm_find_sw_breakpoint(CPUState *cpu, 2758 target_ulong pc) 2759 { 2760 struct kvm_sw_breakpoint *bp; 2761 2762 QTAILQ_FOREACH(bp, &cpu->kvm_state->kvm_sw_breakpoints, entry) { 2763 if (bp->pc == pc) { 2764 return bp; 2765 } 2766 } 2767 return NULL; 2768 } 2769 2770 int kvm_sw_breakpoints_active(CPUState *cpu) 2771 { 2772 return !QTAILQ_EMPTY(&cpu->kvm_state->kvm_sw_breakpoints); 2773 } 2774 2775 struct kvm_set_guest_debug_data { 2776 struct kvm_guest_debug dbg; 2777 int err; 2778 }; 2779 2780 static void kvm_invoke_set_guest_debug(CPUState *cpu, run_on_cpu_data data) 2781 { 2782 struct kvm_set_guest_debug_data *dbg_data = 2783 (struct kvm_set_guest_debug_data *) data.host_ptr; 2784 2785 dbg_data->err = kvm_vcpu_ioctl(cpu, KVM_SET_GUEST_DEBUG, 2786 &dbg_data->dbg); 2787 } 2788 2789 int kvm_update_guest_debug(CPUState *cpu, unsigned long reinject_trap) 2790 { 2791 struct kvm_set_guest_debug_data data; 2792 2793 data.dbg.control = reinject_trap; 2794 2795 if (cpu->singlestep_enabled) { 2796 data.dbg.control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_SINGLESTEP; 2797 } 2798 kvm_arch_update_guest_debug(cpu, &data.dbg); 2799 2800 run_on_cpu(cpu, kvm_invoke_set_guest_debug, 2801 RUN_ON_CPU_HOST_PTR(&data)); 2802 return data.err; 2803 } 2804 2805 int kvm_insert_breakpoint(CPUState *cpu, target_ulong addr, 2806 target_ulong len, int type) 2807 { 2808 struct kvm_sw_breakpoint *bp; 2809 int err; 2810 2811 if (type == GDB_BREAKPOINT_SW) { 2812 bp = kvm_find_sw_breakpoint(cpu, addr); 2813 if (bp) { 2814 bp->use_count++; 2815 return 0; 2816 } 2817 2818 bp = g_malloc(sizeof(struct kvm_sw_breakpoint)); 2819 bp->pc = addr; 2820 bp->use_count = 1; 2821 err = kvm_arch_insert_sw_breakpoint(cpu, bp); 2822 if (err) { 2823 g_free(bp); 2824 return err; 2825 } 2826 2827 QTAILQ_INSERT_HEAD(&cpu->kvm_state->kvm_sw_breakpoints, bp, entry); 2828 } else { 2829 err = kvm_arch_insert_hw_breakpoint(addr, len, type); 2830 if (err) { 2831 return err; 2832 } 2833 } 2834 2835 CPU_FOREACH(cpu) { 2836 err = kvm_update_guest_debug(cpu, 0); 2837 if (err) { 2838 return err; 2839 } 2840 } 2841 return 0; 2842 } 2843 2844 int kvm_remove_breakpoint(CPUState *cpu, target_ulong addr, 2845 target_ulong len, int type) 2846 { 2847 struct kvm_sw_breakpoint *bp; 2848 int err; 2849 2850 if (type == GDB_BREAKPOINT_SW) { 2851 bp = kvm_find_sw_breakpoint(cpu, addr); 2852 if (!bp) { 2853 return -ENOENT; 2854 } 2855 2856 if (bp->use_count > 1) { 2857 bp->use_count--; 2858 return 0; 2859 } 2860 2861 err = kvm_arch_remove_sw_breakpoint(cpu, bp); 2862 if (err) { 2863 return err; 2864 } 2865 2866 QTAILQ_REMOVE(&cpu->kvm_state->kvm_sw_breakpoints, bp, entry); 2867 g_free(bp); 2868 } else { 2869 err = kvm_arch_remove_hw_breakpoint(addr, len, type); 2870 if (err) { 2871 return err; 2872 } 2873 } 2874 2875 CPU_FOREACH(cpu) { 2876 err = kvm_update_guest_debug(cpu, 0); 2877 if (err) { 2878 return err; 2879 } 2880 } 2881 return 0; 2882 } 2883 2884 void kvm_remove_all_breakpoints(CPUState *cpu) 2885 { 2886 struct kvm_sw_breakpoint *bp, *next; 2887 KVMState *s = cpu->kvm_state; 2888 CPUState *tmpcpu; 2889 2890 QTAILQ_FOREACH_SAFE(bp, &s->kvm_sw_breakpoints, entry, next) { 2891 if (kvm_arch_remove_sw_breakpoint(cpu, bp) != 0) { 2892 /* Try harder to find a CPU that currently sees the breakpoint. */ 2893 CPU_FOREACH(tmpcpu) { 2894 if (kvm_arch_remove_sw_breakpoint(tmpcpu, bp) == 0) { 2895 break; 2896 } 2897 } 2898 } 2899 QTAILQ_REMOVE(&s->kvm_sw_breakpoints, bp, entry); 2900 g_free(bp); 2901 } 2902 kvm_arch_remove_all_hw_breakpoints(); 2903 2904 CPU_FOREACH(cpu) { 2905 kvm_update_guest_debug(cpu, 0); 2906 } 2907 } 2908 2909 #else /* !KVM_CAP_SET_GUEST_DEBUG */ 2910 2911 int kvm_update_guest_debug(CPUState *cpu, unsigned long reinject_trap) 2912 { 2913 return -EINVAL; 2914 } 2915 2916 int kvm_insert_breakpoint(CPUState *cpu, target_ulong addr, 2917 target_ulong len, int type) 2918 { 2919 return -EINVAL; 2920 } 2921 2922 int kvm_remove_breakpoint(CPUState *cpu, target_ulong addr, 2923 target_ulong len, int type) 2924 { 2925 return -EINVAL; 2926 } 2927 2928 void kvm_remove_all_breakpoints(CPUState *cpu) 2929 { 2930 } 2931 #endif /* !KVM_CAP_SET_GUEST_DEBUG */ 2932 2933 static int kvm_set_signal_mask(CPUState *cpu, const sigset_t *sigset) 2934 { 2935 KVMState *s = kvm_state; 2936 struct kvm_signal_mask *sigmask; 2937 int r; 2938 2939 sigmask = g_malloc(sizeof(*sigmask) + sizeof(*sigset)); 2940 2941 sigmask->len = s->sigmask_len; 2942 memcpy(sigmask->sigset, sigset, sizeof(*sigset)); 2943 r = kvm_vcpu_ioctl(cpu, KVM_SET_SIGNAL_MASK, sigmask); 2944 g_free(sigmask); 2945 2946 return r; 2947 } 2948 2949 static void kvm_ipi_signal(int sig) 2950 { 2951 if (current_cpu) { 2952 assert(kvm_immediate_exit); 2953 kvm_cpu_kick(current_cpu); 2954 } 2955 } 2956 2957 void kvm_init_cpu_signals(CPUState *cpu) 2958 { 2959 int r; 2960 sigset_t set; 2961 struct sigaction sigact; 2962 2963 memset(&sigact, 0, sizeof(sigact)); 2964 sigact.sa_handler = kvm_ipi_signal; 2965 sigaction(SIG_IPI, &sigact, NULL); 2966 2967 pthread_sigmask(SIG_BLOCK, NULL, &set); 2968 #if defined KVM_HAVE_MCE_INJECTION 2969 sigdelset(&set, SIGBUS); 2970 pthread_sigmask(SIG_SETMASK, &set, NULL); 2971 #endif 2972 sigdelset(&set, SIG_IPI); 2973 if (kvm_immediate_exit) { 2974 r = pthread_sigmask(SIG_SETMASK, &set, NULL); 2975 } else { 2976 r = kvm_set_signal_mask(cpu, &set); 2977 } 2978 if (r) { 2979 fprintf(stderr, "kvm_set_signal_mask: %s\n", strerror(-r)); 2980 exit(1); 2981 } 2982 } 2983 2984 /* Called asynchronously in VCPU thread. */ 2985 int kvm_on_sigbus_vcpu(CPUState *cpu, int code, void *addr) 2986 { 2987 #ifdef KVM_HAVE_MCE_INJECTION 2988 if (have_sigbus_pending) { 2989 return 1; 2990 } 2991 have_sigbus_pending = true; 2992 pending_sigbus_addr = addr; 2993 pending_sigbus_code = code; 2994 qatomic_set(&cpu->exit_request, 1); 2995 return 0; 2996 #else 2997 return 1; 2998 #endif 2999 } 3000 3001 /* Called synchronously (via signalfd) in main thread. */ 3002 int kvm_on_sigbus(int code, void *addr) 3003 { 3004 #ifdef KVM_HAVE_MCE_INJECTION 3005 /* Action required MCE kills the process if SIGBUS is blocked. Because 3006 * that's what happens in the I/O thread, where we handle MCE via signalfd, 3007 * we can only get action optional here. 3008 */ 3009 assert(code != BUS_MCEERR_AR); 3010 kvm_arch_on_sigbus_vcpu(first_cpu, code, addr); 3011 return 0; 3012 #else 3013 return 1; 3014 #endif 3015 } 3016 3017 int kvm_create_device(KVMState *s, uint64_t type, bool test) 3018 { 3019 int ret; 3020 struct kvm_create_device create_dev; 3021 3022 create_dev.type = type; 3023 create_dev.fd = -1; 3024 create_dev.flags = test ? KVM_CREATE_DEVICE_TEST : 0; 3025 3026 if (!kvm_check_extension(s, KVM_CAP_DEVICE_CTRL)) { 3027 return -ENOTSUP; 3028 } 3029 3030 ret = kvm_vm_ioctl(s, KVM_CREATE_DEVICE, &create_dev); 3031 if (ret) { 3032 return ret; 3033 } 3034 3035 return test ? 0 : create_dev.fd; 3036 } 3037 3038 bool kvm_device_supported(int vmfd, uint64_t type) 3039 { 3040 struct kvm_create_device create_dev = { 3041 .type = type, 3042 .fd = -1, 3043 .flags = KVM_CREATE_DEVICE_TEST, 3044 }; 3045 3046 if (ioctl(vmfd, KVM_CHECK_EXTENSION, KVM_CAP_DEVICE_CTRL) <= 0) { 3047 return false; 3048 } 3049 3050 return (ioctl(vmfd, KVM_CREATE_DEVICE, &create_dev) >= 0); 3051 } 3052 3053 int kvm_set_one_reg(CPUState *cs, uint64_t id, void *source) 3054 { 3055 struct kvm_one_reg reg; 3056 int r; 3057 3058 reg.id = id; 3059 reg.addr = (uintptr_t) source; 3060 r = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, ®); 3061 if (r) { 3062 trace_kvm_failed_reg_set(id, strerror(-r)); 3063 } 3064 return r; 3065 } 3066 3067 int kvm_get_one_reg(CPUState *cs, uint64_t id, void *target) 3068 { 3069 struct kvm_one_reg reg; 3070 int r; 3071 3072 reg.id = id; 3073 reg.addr = (uintptr_t) target; 3074 r = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, ®); 3075 if (r) { 3076 trace_kvm_failed_reg_get(id, strerror(-r)); 3077 } 3078 return r; 3079 } 3080 3081 static bool kvm_accel_has_memory(MachineState *ms, AddressSpace *as, 3082 hwaddr start_addr, hwaddr size) 3083 { 3084 KVMState *kvm = KVM_STATE(ms->accelerator); 3085 int i; 3086 3087 for (i = 0; i < kvm->nr_as; ++i) { 3088 if (kvm->as[i].as == as && kvm->as[i].ml) { 3089 size = MIN(kvm_max_slot_size, size); 3090 return NULL != kvm_lookup_matching_slot(kvm->as[i].ml, 3091 start_addr, size); 3092 } 3093 } 3094 3095 return false; 3096 } 3097 3098 static void kvm_get_kvm_shadow_mem(Object *obj, Visitor *v, 3099 const char *name, void *opaque, 3100 Error **errp) 3101 { 3102 KVMState *s = KVM_STATE(obj); 3103 int64_t value = s->kvm_shadow_mem; 3104 3105 visit_type_int(v, name, &value, errp); 3106 } 3107 3108 static void kvm_set_kvm_shadow_mem(Object *obj, Visitor *v, 3109 const char *name, void *opaque, 3110 Error **errp) 3111 { 3112 KVMState *s = KVM_STATE(obj); 3113 int64_t value; 3114 3115 if (!visit_type_int(v, name, &value, errp)) { 3116 return; 3117 } 3118 3119 s->kvm_shadow_mem = value; 3120 } 3121 3122 static void kvm_set_kernel_irqchip(Object *obj, Visitor *v, 3123 const char *name, void *opaque, 3124 Error **errp) 3125 { 3126 KVMState *s = KVM_STATE(obj); 3127 OnOffSplit mode; 3128 3129 if (!visit_type_OnOffSplit(v, name, &mode, errp)) { 3130 return; 3131 } 3132 switch (mode) { 3133 case ON_OFF_SPLIT_ON: 3134 s->kernel_irqchip_allowed = true; 3135 s->kernel_irqchip_required = true; 3136 s->kernel_irqchip_split = ON_OFF_AUTO_OFF; 3137 break; 3138 case ON_OFF_SPLIT_OFF: 3139 s->kernel_irqchip_allowed = false; 3140 s->kernel_irqchip_required = false; 3141 s->kernel_irqchip_split = ON_OFF_AUTO_OFF; 3142 break; 3143 case ON_OFF_SPLIT_SPLIT: 3144 s->kernel_irqchip_allowed = true; 3145 s->kernel_irqchip_required = true; 3146 s->kernel_irqchip_split = ON_OFF_AUTO_ON; 3147 break; 3148 default: 3149 /* The value was checked in visit_type_OnOffSplit() above. If 3150 * we get here, then something is wrong in QEMU. 3151 */ 3152 abort(); 3153 } 3154 } 3155 3156 bool kvm_kernel_irqchip_allowed(void) 3157 { 3158 return kvm_state->kernel_irqchip_allowed; 3159 } 3160 3161 bool kvm_kernel_irqchip_required(void) 3162 { 3163 return kvm_state->kernel_irqchip_required; 3164 } 3165 3166 bool kvm_kernel_irqchip_split(void) 3167 { 3168 return kvm_state->kernel_irqchip_split == ON_OFF_AUTO_ON; 3169 } 3170 3171 static void kvm_accel_instance_init(Object *obj) 3172 { 3173 KVMState *s = KVM_STATE(obj); 3174 3175 s->kvm_shadow_mem = -1; 3176 s->kernel_irqchip_allowed = true; 3177 s->kernel_irqchip_split = ON_OFF_AUTO_AUTO; 3178 } 3179 3180 static void kvm_accel_class_init(ObjectClass *oc, void *data) 3181 { 3182 AccelClass *ac = ACCEL_CLASS(oc); 3183 ac->name = "KVM"; 3184 ac->init_machine = kvm_init; 3185 ac->has_memory = kvm_accel_has_memory; 3186 ac->allowed = &kvm_allowed; 3187 3188 object_class_property_add(oc, "kernel-irqchip", "on|off|split", 3189 NULL, kvm_set_kernel_irqchip, 3190 NULL, NULL); 3191 object_class_property_set_description(oc, "kernel-irqchip", 3192 "Configure KVM in-kernel irqchip"); 3193 3194 object_class_property_add(oc, "kvm-shadow-mem", "int", 3195 kvm_get_kvm_shadow_mem, kvm_set_kvm_shadow_mem, 3196 NULL, NULL); 3197 object_class_property_set_description(oc, "kvm-shadow-mem", 3198 "KVM shadow MMU size"); 3199 } 3200 3201 static const TypeInfo kvm_accel_type = { 3202 .name = TYPE_KVM_ACCEL, 3203 .parent = TYPE_ACCEL, 3204 .instance_init = kvm_accel_instance_init, 3205 .class_init = kvm_accel_class_init, 3206 .instance_size = sizeof(KVMState), 3207 }; 3208 3209 static void kvm_type_init(void) 3210 { 3211 type_register_static(&kvm_accel_type); 3212 } 3213 3214 type_init(kvm_type_init); 3215