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