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 #include <poll.h> 19 20 #include <linux/kvm.h> 21 22 #include "qemu/atomic.h" 23 #include "qemu/option.h" 24 #include "qemu/config-file.h" 25 #include "qemu/error-report.h" 26 #include "qapi/error.h" 27 #include "hw/pci/msi.h" 28 #include "hw/pci/msix.h" 29 #include "hw/s390x/adapter.h" 30 #include "exec/gdbstub.h" 31 #include "sysemu/kvm_int.h" 32 #include "sysemu/runstate.h" 33 #include "sysemu/cpus.h" 34 #include "qemu/bswap.h" 35 #include "exec/memory.h" 36 #include "exec/ram_addr.h" 37 #include "qemu/event_notifier.h" 38 #include "qemu/main-loop.h" 39 #include "trace.h" 40 #include "hw/irq.h" 41 #include "qapi/visitor.h" 42 #include "qapi/qapi-types-common.h" 43 #include "qapi/qapi-visit-common.h" 44 #include "sysemu/reset.h" 45 #include "qemu/guest-random.h" 46 #include "sysemu/hw_accel.h" 47 #include "kvm-cpus.h" 48 #include "sysemu/dirtylimit.h" 49 50 #include "hw/boards.h" 51 #include "monitor/stats.h" 52 53 /* This check must be after config-host.h is included */ 54 #ifdef CONFIG_EVENTFD 55 #include <sys/eventfd.h> 56 #endif 57 58 /* KVM uses PAGE_SIZE in its definition of KVM_COALESCED_MMIO_MAX. We 59 * need to use the real host PAGE_SIZE, as that's what KVM will use. 60 */ 61 #ifdef PAGE_SIZE 62 #undef PAGE_SIZE 63 #endif 64 #define PAGE_SIZE qemu_real_host_page_size() 65 66 #ifndef KVM_GUESTDBG_BLOCKIRQ 67 #define KVM_GUESTDBG_BLOCKIRQ 0 68 #endif 69 70 //#define DEBUG_KVM 71 72 #ifdef DEBUG_KVM 73 #define DPRINTF(fmt, ...) \ 74 do { fprintf(stderr, fmt, ## __VA_ARGS__); } while (0) 75 #else 76 #define DPRINTF(fmt, ...) \ 77 do { } while (0) 78 #endif 79 80 #define KVM_MSI_HASHTAB_SIZE 256 81 82 struct KVMParkedVcpu { 83 unsigned long vcpu_id; 84 int kvm_fd; 85 QLIST_ENTRY(KVMParkedVcpu) node; 86 }; 87 88 enum KVMDirtyRingReaperState { 89 KVM_DIRTY_RING_REAPER_NONE = 0, 90 /* The reaper is sleeping */ 91 KVM_DIRTY_RING_REAPER_WAIT, 92 /* The reaper is reaping for dirty pages */ 93 KVM_DIRTY_RING_REAPER_REAPING, 94 }; 95 96 /* 97 * KVM reaper instance, responsible for collecting the KVM dirty bits 98 * via the dirty ring. 99 */ 100 struct KVMDirtyRingReaper { 101 /* The reaper thread */ 102 QemuThread reaper_thr; 103 volatile uint64_t reaper_iteration; /* iteration number of reaper thr */ 104 volatile enum KVMDirtyRingReaperState reaper_state; /* reap thr state */ 105 }; 106 107 struct KVMState 108 { 109 AccelState parent_obj; 110 111 int nr_slots; 112 int fd; 113 int vmfd; 114 int coalesced_mmio; 115 int coalesced_pio; 116 struct kvm_coalesced_mmio_ring *coalesced_mmio_ring; 117 bool coalesced_flush_in_progress; 118 int vcpu_events; 119 int robust_singlestep; 120 int debugregs; 121 #ifdef KVM_CAP_SET_GUEST_DEBUG 122 QTAILQ_HEAD(, kvm_sw_breakpoint) kvm_sw_breakpoints; 123 #endif 124 int max_nested_state_len; 125 int many_ioeventfds; 126 int intx_set_mask; 127 int kvm_shadow_mem; 128 bool kernel_irqchip_allowed; 129 bool kernel_irqchip_required; 130 OnOffAuto kernel_irqchip_split; 131 bool sync_mmu; 132 uint64_t manual_dirty_log_protect; 133 /* The man page (and posix) say ioctl numbers are signed int, but 134 * they're not. Linux, glibc and *BSD all treat ioctl numbers as 135 * unsigned, and treating them as signed here can break things */ 136 unsigned irq_set_ioctl; 137 unsigned int sigmask_len; 138 GHashTable *gsimap; 139 #ifdef KVM_CAP_IRQ_ROUTING 140 struct kvm_irq_routing *irq_routes; 141 int nr_allocated_irq_routes; 142 unsigned long *used_gsi_bitmap; 143 unsigned int gsi_count; 144 QTAILQ_HEAD(, KVMMSIRoute) msi_hashtab[KVM_MSI_HASHTAB_SIZE]; 145 #endif 146 KVMMemoryListener memory_listener; 147 QLIST_HEAD(, KVMParkedVcpu) kvm_parked_vcpus; 148 149 /* For "info mtree -f" to tell if an MR is registered in KVM */ 150 int nr_as; 151 struct KVMAs { 152 KVMMemoryListener *ml; 153 AddressSpace *as; 154 } *as; 155 uint64_t kvm_dirty_ring_bytes; /* Size of the per-vcpu dirty ring */ 156 uint32_t kvm_dirty_ring_size; /* Number of dirty GFNs per ring */ 157 struct KVMDirtyRingReaper reaper; 158 }; 159 160 KVMState *kvm_state; 161 bool kvm_kernel_irqchip; 162 bool kvm_split_irqchip; 163 bool kvm_async_interrupts_allowed; 164 bool kvm_halt_in_kernel_allowed; 165 bool kvm_eventfds_allowed; 166 bool kvm_irqfds_allowed; 167 bool kvm_resamplefds_allowed; 168 bool kvm_msi_via_irqfd_allowed; 169 bool kvm_gsi_routing_allowed; 170 bool kvm_gsi_direct_mapping; 171 bool kvm_allowed; 172 bool kvm_readonly_mem_allowed; 173 bool kvm_vm_attributes_allowed; 174 bool kvm_direct_msi_allowed; 175 bool kvm_ioeventfd_any_length_allowed; 176 bool kvm_msi_use_devid; 177 bool kvm_has_guest_debug; 178 int kvm_sstep_flags; 179 static bool kvm_immediate_exit; 180 static hwaddr kvm_max_slot_size = ~0; 181 182 static const KVMCapabilityInfo kvm_required_capabilites[] = { 183 KVM_CAP_INFO(USER_MEMORY), 184 KVM_CAP_INFO(DESTROY_MEMORY_REGION_WORKS), 185 KVM_CAP_INFO(JOIN_MEMORY_REGIONS_WORKS), 186 KVM_CAP_LAST_INFO 187 }; 188 189 static NotifierList kvm_irqchip_change_notifiers = 190 NOTIFIER_LIST_INITIALIZER(kvm_irqchip_change_notifiers); 191 192 struct KVMResampleFd { 193 int gsi; 194 EventNotifier *resample_event; 195 QLIST_ENTRY(KVMResampleFd) node; 196 }; 197 typedef struct KVMResampleFd KVMResampleFd; 198 199 /* 200 * Only used with split irqchip where we need to do the resample fd 201 * kick for the kernel from userspace. 202 */ 203 static QLIST_HEAD(, KVMResampleFd) kvm_resample_fd_list = 204 QLIST_HEAD_INITIALIZER(kvm_resample_fd_list); 205 206 static QemuMutex kml_slots_lock; 207 208 #define kvm_slots_lock() qemu_mutex_lock(&kml_slots_lock) 209 #define kvm_slots_unlock() qemu_mutex_unlock(&kml_slots_lock) 210 211 static void kvm_slot_init_dirty_bitmap(KVMSlot *mem); 212 213 static inline void kvm_resample_fd_remove(int gsi) 214 { 215 KVMResampleFd *rfd; 216 217 QLIST_FOREACH(rfd, &kvm_resample_fd_list, node) { 218 if (rfd->gsi == gsi) { 219 QLIST_REMOVE(rfd, node); 220 g_free(rfd); 221 break; 222 } 223 } 224 } 225 226 static inline void kvm_resample_fd_insert(int gsi, EventNotifier *event) 227 { 228 KVMResampleFd *rfd = g_new0(KVMResampleFd, 1); 229 230 rfd->gsi = gsi; 231 rfd->resample_event = event; 232 233 QLIST_INSERT_HEAD(&kvm_resample_fd_list, rfd, node); 234 } 235 236 void kvm_resample_fd_notify(int gsi) 237 { 238 KVMResampleFd *rfd; 239 240 QLIST_FOREACH(rfd, &kvm_resample_fd_list, node) { 241 if (rfd->gsi == gsi) { 242 event_notifier_set(rfd->resample_event); 243 trace_kvm_resample_fd_notify(gsi); 244 return; 245 } 246 } 247 } 248 249 int kvm_get_max_memslots(void) 250 { 251 KVMState *s = KVM_STATE(current_accel()); 252 253 return s->nr_slots; 254 } 255 256 /* Called with KVMMemoryListener.slots_lock held */ 257 static KVMSlot *kvm_get_free_slot(KVMMemoryListener *kml) 258 { 259 KVMState *s = kvm_state; 260 int i; 261 262 for (i = 0; i < s->nr_slots; i++) { 263 if (kml->slots[i].memory_size == 0) { 264 return &kml->slots[i]; 265 } 266 } 267 268 return NULL; 269 } 270 271 bool kvm_has_free_slot(MachineState *ms) 272 { 273 KVMState *s = KVM_STATE(ms->accelerator); 274 bool result; 275 KVMMemoryListener *kml = &s->memory_listener; 276 277 kvm_slots_lock(); 278 result = !!kvm_get_free_slot(kml); 279 kvm_slots_unlock(); 280 281 return result; 282 } 283 284 /* Called with KVMMemoryListener.slots_lock held */ 285 static KVMSlot *kvm_alloc_slot(KVMMemoryListener *kml) 286 { 287 KVMSlot *slot = kvm_get_free_slot(kml); 288 289 if (slot) { 290 return slot; 291 } 292 293 fprintf(stderr, "%s: no free slot available\n", __func__); 294 abort(); 295 } 296 297 static KVMSlot *kvm_lookup_matching_slot(KVMMemoryListener *kml, 298 hwaddr start_addr, 299 hwaddr size) 300 { 301 KVMState *s = kvm_state; 302 int i; 303 304 for (i = 0; i < s->nr_slots; i++) { 305 KVMSlot *mem = &kml->slots[i]; 306 307 if (start_addr == mem->start_addr && size == mem->memory_size) { 308 return mem; 309 } 310 } 311 312 return NULL; 313 } 314 315 /* 316 * Calculate and align the start address and the size of the section. 317 * Return the size. If the size is 0, the aligned section is empty. 318 */ 319 static hwaddr kvm_align_section(MemoryRegionSection *section, 320 hwaddr *start) 321 { 322 hwaddr size = int128_get64(section->size); 323 hwaddr delta, aligned; 324 325 /* kvm works in page size chunks, but the function may be called 326 with sub-page size and unaligned start address. Pad the start 327 address to next and truncate size to previous page boundary. */ 328 aligned = ROUND_UP(section->offset_within_address_space, 329 qemu_real_host_page_size()); 330 delta = aligned - section->offset_within_address_space; 331 *start = aligned; 332 if (delta > size) { 333 return 0; 334 } 335 336 return (size - delta) & qemu_real_host_page_mask(); 337 } 338 339 int kvm_physical_memory_addr_from_host(KVMState *s, void *ram, 340 hwaddr *phys_addr) 341 { 342 KVMMemoryListener *kml = &s->memory_listener; 343 int i, ret = 0; 344 345 kvm_slots_lock(); 346 for (i = 0; i < s->nr_slots; i++) { 347 KVMSlot *mem = &kml->slots[i]; 348 349 if (ram >= mem->ram && ram < mem->ram + mem->memory_size) { 350 *phys_addr = mem->start_addr + (ram - mem->ram); 351 ret = 1; 352 break; 353 } 354 } 355 kvm_slots_unlock(); 356 357 return ret; 358 } 359 360 static int kvm_set_user_memory_region(KVMMemoryListener *kml, KVMSlot *slot, bool new) 361 { 362 KVMState *s = kvm_state; 363 struct kvm_userspace_memory_region mem; 364 int ret; 365 366 mem.slot = slot->slot | (kml->as_id << 16); 367 mem.guest_phys_addr = slot->start_addr; 368 mem.userspace_addr = (unsigned long)slot->ram; 369 mem.flags = slot->flags; 370 371 if (slot->memory_size && !new && (mem.flags ^ slot->old_flags) & KVM_MEM_READONLY) { 372 /* Set the slot size to 0 before setting the slot to the desired 373 * value. This is needed based on KVM commit 75d61fbc. */ 374 mem.memory_size = 0; 375 ret = kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem); 376 if (ret < 0) { 377 goto err; 378 } 379 } 380 mem.memory_size = slot->memory_size; 381 ret = kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem); 382 slot->old_flags = mem.flags; 383 err: 384 trace_kvm_set_user_memory(mem.slot, mem.flags, mem.guest_phys_addr, 385 mem.memory_size, mem.userspace_addr, ret); 386 if (ret < 0) { 387 error_report("%s: KVM_SET_USER_MEMORY_REGION failed, slot=%d," 388 " start=0x%" PRIx64 ", size=0x%" PRIx64 ": %s", 389 __func__, mem.slot, slot->start_addr, 390 (uint64_t)mem.memory_size, strerror(errno)); 391 } 392 return ret; 393 } 394 395 static int do_kvm_destroy_vcpu(CPUState *cpu) 396 { 397 KVMState *s = kvm_state; 398 long mmap_size; 399 struct KVMParkedVcpu *vcpu = NULL; 400 int ret = 0; 401 402 DPRINTF("kvm_destroy_vcpu\n"); 403 404 ret = kvm_arch_destroy_vcpu(cpu); 405 if (ret < 0) { 406 goto err; 407 } 408 409 mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0); 410 if (mmap_size < 0) { 411 ret = mmap_size; 412 DPRINTF("KVM_GET_VCPU_MMAP_SIZE failed\n"); 413 goto err; 414 } 415 416 ret = munmap(cpu->kvm_run, mmap_size); 417 if (ret < 0) { 418 goto err; 419 } 420 421 if (cpu->kvm_dirty_gfns) { 422 ret = munmap(cpu->kvm_dirty_gfns, s->kvm_dirty_ring_bytes); 423 if (ret < 0) { 424 goto err; 425 } 426 } 427 428 vcpu = g_malloc0(sizeof(*vcpu)); 429 vcpu->vcpu_id = kvm_arch_vcpu_id(cpu); 430 vcpu->kvm_fd = cpu->kvm_fd; 431 QLIST_INSERT_HEAD(&kvm_state->kvm_parked_vcpus, vcpu, node); 432 err: 433 return ret; 434 } 435 436 void kvm_destroy_vcpu(CPUState *cpu) 437 { 438 if (do_kvm_destroy_vcpu(cpu) < 0) { 439 error_report("kvm_destroy_vcpu failed"); 440 exit(EXIT_FAILURE); 441 } 442 } 443 444 static int kvm_get_vcpu(KVMState *s, unsigned long vcpu_id) 445 { 446 struct KVMParkedVcpu *cpu; 447 448 QLIST_FOREACH(cpu, &s->kvm_parked_vcpus, node) { 449 if (cpu->vcpu_id == vcpu_id) { 450 int kvm_fd; 451 452 QLIST_REMOVE(cpu, node); 453 kvm_fd = cpu->kvm_fd; 454 g_free(cpu); 455 return kvm_fd; 456 } 457 } 458 459 return kvm_vm_ioctl(s, KVM_CREATE_VCPU, (void *)vcpu_id); 460 } 461 462 int kvm_init_vcpu(CPUState *cpu, Error **errp) 463 { 464 KVMState *s = kvm_state; 465 long mmap_size; 466 int ret; 467 468 trace_kvm_init_vcpu(cpu->cpu_index, kvm_arch_vcpu_id(cpu)); 469 470 ret = kvm_get_vcpu(s, kvm_arch_vcpu_id(cpu)); 471 if (ret < 0) { 472 error_setg_errno(errp, -ret, "kvm_init_vcpu: kvm_get_vcpu failed (%lu)", 473 kvm_arch_vcpu_id(cpu)); 474 goto err; 475 } 476 477 cpu->kvm_fd = ret; 478 cpu->kvm_state = s; 479 cpu->vcpu_dirty = true; 480 cpu->dirty_pages = 0; 481 cpu->throttle_us_per_full = 0; 482 483 mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0); 484 if (mmap_size < 0) { 485 ret = mmap_size; 486 error_setg_errno(errp, -mmap_size, 487 "kvm_init_vcpu: KVM_GET_VCPU_MMAP_SIZE failed"); 488 goto err; 489 } 490 491 cpu->kvm_run = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED, 492 cpu->kvm_fd, 0); 493 if (cpu->kvm_run == MAP_FAILED) { 494 ret = -errno; 495 error_setg_errno(errp, ret, 496 "kvm_init_vcpu: mmap'ing vcpu state failed (%lu)", 497 kvm_arch_vcpu_id(cpu)); 498 goto err; 499 } 500 501 if (s->coalesced_mmio && !s->coalesced_mmio_ring) { 502 s->coalesced_mmio_ring = 503 (void *)cpu->kvm_run + s->coalesced_mmio * PAGE_SIZE; 504 } 505 506 if (s->kvm_dirty_ring_size) { 507 /* Use MAP_SHARED to share pages with the kernel */ 508 cpu->kvm_dirty_gfns = mmap(NULL, s->kvm_dirty_ring_bytes, 509 PROT_READ | PROT_WRITE, MAP_SHARED, 510 cpu->kvm_fd, 511 PAGE_SIZE * KVM_DIRTY_LOG_PAGE_OFFSET); 512 if (cpu->kvm_dirty_gfns == MAP_FAILED) { 513 ret = -errno; 514 DPRINTF("mmap'ing vcpu dirty gfns failed: %d\n", ret); 515 goto err; 516 } 517 } 518 519 ret = kvm_arch_init_vcpu(cpu); 520 if (ret < 0) { 521 error_setg_errno(errp, -ret, 522 "kvm_init_vcpu: kvm_arch_init_vcpu failed (%lu)", 523 kvm_arch_vcpu_id(cpu)); 524 } 525 err: 526 return ret; 527 } 528 529 /* 530 * dirty pages logging control 531 */ 532 533 static int kvm_mem_flags(MemoryRegion *mr) 534 { 535 bool readonly = mr->readonly || memory_region_is_romd(mr); 536 int flags = 0; 537 538 if (memory_region_get_dirty_log_mask(mr) != 0) { 539 flags |= KVM_MEM_LOG_DIRTY_PAGES; 540 } 541 if (readonly && kvm_readonly_mem_allowed) { 542 flags |= KVM_MEM_READONLY; 543 } 544 return flags; 545 } 546 547 /* Called with KVMMemoryListener.slots_lock held */ 548 static int kvm_slot_update_flags(KVMMemoryListener *kml, KVMSlot *mem, 549 MemoryRegion *mr) 550 { 551 mem->flags = kvm_mem_flags(mr); 552 553 /* If nothing changed effectively, no need to issue ioctl */ 554 if (mem->flags == mem->old_flags) { 555 return 0; 556 } 557 558 kvm_slot_init_dirty_bitmap(mem); 559 return kvm_set_user_memory_region(kml, mem, false); 560 } 561 562 static int kvm_section_update_flags(KVMMemoryListener *kml, 563 MemoryRegionSection *section) 564 { 565 hwaddr start_addr, size, slot_size; 566 KVMSlot *mem; 567 int ret = 0; 568 569 size = kvm_align_section(section, &start_addr); 570 if (!size) { 571 return 0; 572 } 573 574 kvm_slots_lock(); 575 576 while (size && !ret) { 577 slot_size = MIN(kvm_max_slot_size, size); 578 mem = kvm_lookup_matching_slot(kml, start_addr, slot_size); 579 if (!mem) { 580 /* We don't have a slot if we want to trap every access. */ 581 goto out; 582 } 583 584 ret = kvm_slot_update_flags(kml, mem, section->mr); 585 start_addr += slot_size; 586 size -= slot_size; 587 } 588 589 out: 590 kvm_slots_unlock(); 591 return ret; 592 } 593 594 static void kvm_log_start(MemoryListener *listener, 595 MemoryRegionSection *section, 596 int old, int new) 597 { 598 KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener); 599 int r; 600 601 if (old != 0) { 602 return; 603 } 604 605 r = kvm_section_update_flags(kml, section); 606 if (r < 0) { 607 abort(); 608 } 609 } 610 611 static void kvm_log_stop(MemoryListener *listener, 612 MemoryRegionSection *section, 613 int old, int new) 614 { 615 KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener); 616 int r; 617 618 if (new != 0) { 619 return; 620 } 621 622 r = kvm_section_update_flags(kml, section); 623 if (r < 0) { 624 abort(); 625 } 626 } 627 628 /* get kvm's dirty pages bitmap and update qemu's */ 629 static void kvm_slot_sync_dirty_pages(KVMSlot *slot) 630 { 631 ram_addr_t start = slot->ram_start_offset; 632 ram_addr_t pages = slot->memory_size / qemu_real_host_page_size(); 633 634 cpu_physical_memory_set_dirty_lebitmap(slot->dirty_bmap, start, pages); 635 } 636 637 static void kvm_slot_reset_dirty_pages(KVMSlot *slot) 638 { 639 memset(slot->dirty_bmap, 0, slot->dirty_bmap_size); 640 } 641 642 #define ALIGN(x, y) (((x)+(y)-1) & ~((y)-1)) 643 644 /* Allocate the dirty bitmap for a slot */ 645 static void kvm_slot_init_dirty_bitmap(KVMSlot *mem) 646 { 647 if (!(mem->flags & KVM_MEM_LOG_DIRTY_PAGES) || mem->dirty_bmap) { 648 return; 649 } 650 651 /* 652 * XXX bad kernel interface alert 653 * For dirty bitmap, kernel allocates array of size aligned to 654 * bits-per-long. But for case when the kernel is 64bits and 655 * the userspace is 32bits, userspace can't align to the same 656 * bits-per-long, since sizeof(long) is different between kernel 657 * and user space. This way, userspace will provide buffer which 658 * may be 4 bytes less than the kernel will use, resulting in 659 * userspace memory corruption (which is not detectable by valgrind 660 * too, in most cases). 661 * So for now, let's align to 64 instead of HOST_LONG_BITS here, in 662 * a hope that sizeof(long) won't become >8 any time soon. 663 * 664 * Note: the granule of kvm dirty log is qemu_real_host_page_size. 665 * And mem->memory_size is aligned to it (otherwise this mem can't 666 * be registered to KVM). 667 */ 668 hwaddr bitmap_size = ALIGN(mem->memory_size / qemu_real_host_page_size(), 669 /*HOST_LONG_BITS*/ 64) / 8; 670 mem->dirty_bmap = g_malloc0(bitmap_size); 671 mem->dirty_bmap_size = bitmap_size; 672 } 673 674 /* 675 * Sync dirty bitmap from kernel to KVMSlot.dirty_bmap, return true if 676 * succeeded, false otherwise 677 */ 678 static bool kvm_slot_get_dirty_log(KVMState *s, KVMSlot *slot) 679 { 680 struct kvm_dirty_log d = {}; 681 int ret; 682 683 d.dirty_bitmap = slot->dirty_bmap; 684 d.slot = slot->slot | (slot->as_id << 16); 685 ret = kvm_vm_ioctl(s, KVM_GET_DIRTY_LOG, &d); 686 687 if (ret == -ENOENT) { 688 /* kernel does not have dirty bitmap in this slot */ 689 ret = 0; 690 } 691 if (ret) { 692 error_report_once("%s: KVM_GET_DIRTY_LOG failed with %d", 693 __func__, ret); 694 } 695 return ret == 0; 696 } 697 698 /* Should be with all slots_lock held for the address spaces. */ 699 static void kvm_dirty_ring_mark_page(KVMState *s, uint32_t as_id, 700 uint32_t slot_id, uint64_t offset) 701 { 702 KVMMemoryListener *kml; 703 KVMSlot *mem; 704 705 if (as_id >= s->nr_as) { 706 return; 707 } 708 709 kml = s->as[as_id].ml; 710 mem = &kml->slots[slot_id]; 711 712 if (!mem->memory_size || offset >= 713 (mem->memory_size / qemu_real_host_page_size())) { 714 return; 715 } 716 717 set_bit(offset, mem->dirty_bmap); 718 } 719 720 static bool dirty_gfn_is_dirtied(struct kvm_dirty_gfn *gfn) 721 { 722 /* 723 * Read the flags before the value. Pairs with barrier in 724 * KVM's kvm_dirty_ring_push() function. 725 */ 726 return qatomic_load_acquire(&gfn->flags) == KVM_DIRTY_GFN_F_DIRTY; 727 } 728 729 static void dirty_gfn_set_collected(struct kvm_dirty_gfn *gfn) 730 { 731 gfn->flags = KVM_DIRTY_GFN_F_RESET; 732 } 733 734 /* 735 * Should be with all slots_lock held for the address spaces. It returns the 736 * dirty page we've collected on this dirty ring. 737 */ 738 static uint32_t kvm_dirty_ring_reap_one(KVMState *s, CPUState *cpu) 739 { 740 struct kvm_dirty_gfn *dirty_gfns = cpu->kvm_dirty_gfns, *cur; 741 uint32_t ring_size = s->kvm_dirty_ring_size; 742 uint32_t count = 0, fetch = cpu->kvm_fetch_index; 743 744 assert(dirty_gfns && ring_size); 745 trace_kvm_dirty_ring_reap_vcpu(cpu->cpu_index); 746 747 while (true) { 748 cur = &dirty_gfns[fetch % ring_size]; 749 if (!dirty_gfn_is_dirtied(cur)) { 750 break; 751 } 752 kvm_dirty_ring_mark_page(s, cur->slot >> 16, cur->slot & 0xffff, 753 cur->offset); 754 dirty_gfn_set_collected(cur); 755 trace_kvm_dirty_ring_page(cpu->cpu_index, fetch, cur->offset); 756 fetch++; 757 count++; 758 } 759 cpu->kvm_fetch_index = fetch; 760 cpu->dirty_pages += count; 761 762 return count; 763 } 764 765 /* Must be with slots_lock held */ 766 static uint64_t kvm_dirty_ring_reap_locked(KVMState *s, CPUState* cpu) 767 { 768 int ret; 769 uint64_t total = 0; 770 int64_t stamp; 771 772 stamp = get_clock(); 773 774 if (cpu) { 775 total = kvm_dirty_ring_reap_one(s, cpu); 776 } else { 777 CPU_FOREACH(cpu) { 778 total += kvm_dirty_ring_reap_one(s, cpu); 779 } 780 } 781 782 if (total) { 783 ret = kvm_vm_ioctl(s, KVM_RESET_DIRTY_RINGS); 784 assert(ret == total); 785 } 786 787 stamp = get_clock() - stamp; 788 789 if (total) { 790 trace_kvm_dirty_ring_reap(total, stamp / 1000); 791 } 792 793 return total; 794 } 795 796 /* 797 * Currently for simplicity, we must hold BQL before calling this. We can 798 * consider to drop the BQL if we're clear with all the race conditions. 799 */ 800 static uint64_t kvm_dirty_ring_reap(KVMState *s, CPUState *cpu) 801 { 802 uint64_t total; 803 804 /* 805 * We need to lock all kvm slots for all address spaces here, 806 * because: 807 * 808 * (1) We need to mark dirty for dirty bitmaps in multiple slots 809 * and for tons of pages, so it's better to take the lock here 810 * once rather than once per page. And more importantly, 811 * 812 * (2) We must _NOT_ publish dirty bits to the other threads 813 * (e.g., the migration thread) via the kvm memory slot dirty 814 * bitmaps before correctly re-protect those dirtied pages. 815 * Otherwise we can have potential risk of data corruption if 816 * the page data is read in the other thread before we do 817 * reset below. 818 */ 819 kvm_slots_lock(); 820 total = kvm_dirty_ring_reap_locked(s, cpu); 821 kvm_slots_unlock(); 822 823 return total; 824 } 825 826 static void do_kvm_cpu_synchronize_kick(CPUState *cpu, run_on_cpu_data arg) 827 { 828 /* No need to do anything */ 829 } 830 831 /* 832 * Kick all vcpus out in a synchronized way. When returned, we 833 * guarantee that every vcpu has been kicked and at least returned to 834 * userspace once. 835 */ 836 static void kvm_cpu_synchronize_kick_all(void) 837 { 838 CPUState *cpu; 839 840 CPU_FOREACH(cpu) { 841 run_on_cpu(cpu, do_kvm_cpu_synchronize_kick, RUN_ON_CPU_NULL); 842 } 843 } 844 845 /* 846 * Flush all the existing dirty pages to the KVM slot buffers. When 847 * this call returns, we guarantee that all the touched dirty pages 848 * before calling this function have been put into the per-kvmslot 849 * dirty bitmap. 850 * 851 * This function must be called with BQL held. 852 */ 853 static void kvm_dirty_ring_flush(void) 854 { 855 trace_kvm_dirty_ring_flush(0); 856 /* 857 * The function needs to be serialized. Since this function 858 * should always be with BQL held, serialization is guaranteed. 859 * However, let's be sure of it. 860 */ 861 assert(qemu_mutex_iothread_locked()); 862 /* 863 * First make sure to flush the hardware buffers by kicking all 864 * vcpus out in a synchronous way. 865 */ 866 kvm_cpu_synchronize_kick_all(); 867 kvm_dirty_ring_reap(kvm_state, NULL); 868 trace_kvm_dirty_ring_flush(1); 869 } 870 871 /** 872 * kvm_physical_sync_dirty_bitmap - Sync dirty bitmap from kernel space 873 * 874 * This function will first try to fetch dirty bitmap from the kernel, 875 * and then updates qemu's dirty bitmap. 876 * 877 * NOTE: caller must be with kml->slots_lock held. 878 * 879 * @kml: the KVM memory listener object 880 * @section: the memory section to sync the dirty bitmap with 881 */ 882 static void kvm_physical_sync_dirty_bitmap(KVMMemoryListener *kml, 883 MemoryRegionSection *section) 884 { 885 KVMState *s = kvm_state; 886 KVMSlot *mem; 887 hwaddr start_addr, size; 888 hwaddr slot_size; 889 890 size = kvm_align_section(section, &start_addr); 891 while (size) { 892 slot_size = MIN(kvm_max_slot_size, size); 893 mem = kvm_lookup_matching_slot(kml, start_addr, slot_size); 894 if (!mem) { 895 /* We don't have a slot if we want to trap every access. */ 896 return; 897 } 898 if (kvm_slot_get_dirty_log(s, mem)) { 899 kvm_slot_sync_dirty_pages(mem); 900 } 901 start_addr += slot_size; 902 size -= slot_size; 903 } 904 } 905 906 /* Alignment requirement for KVM_CLEAR_DIRTY_LOG - 64 pages */ 907 #define KVM_CLEAR_LOG_SHIFT 6 908 #define KVM_CLEAR_LOG_ALIGN (qemu_real_host_page_size() << KVM_CLEAR_LOG_SHIFT) 909 #define KVM_CLEAR_LOG_MASK (-KVM_CLEAR_LOG_ALIGN) 910 911 static int kvm_log_clear_one_slot(KVMSlot *mem, int as_id, uint64_t start, 912 uint64_t size) 913 { 914 KVMState *s = kvm_state; 915 uint64_t end, bmap_start, start_delta, bmap_npages; 916 struct kvm_clear_dirty_log d; 917 unsigned long *bmap_clear = NULL, psize = qemu_real_host_page_size(); 918 int ret; 919 920 /* 921 * We need to extend either the start or the size or both to 922 * satisfy the KVM interface requirement. Firstly, do the start 923 * page alignment on 64 host pages 924 */ 925 bmap_start = start & KVM_CLEAR_LOG_MASK; 926 start_delta = start - bmap_start; 927 bmap_start /= psize; 928 929 /* 930 * The kernel interface has restriction on the size too, that either: 931 * 932 * (1) the size is 64 host pages aligned (just like the start), or 933 * (2) the size fills up until the end of the KVM memslot. 934 */ 935 bmap_npages = DIV_ROUND_UP(size + start_delta, KVM_CLEAR_LOG_ALIGN) 936 << KVM_CLEAR_LOG_SHIFT; 937 end = mem->memory_size / psize; 938 if (bmap_npages > end - bmap_start) { 939 bmap_npages = end - bmap_start; 940 } 941 start_delta /= psize; 942 943 /* 944 * Prepare the bitmap to clear dirty bits. Here we must guarantee 945 * that we won't clear any unknown dirty bits otherwise we might 946 * accidentally clear some set bits which are not yet synced from 947 * the kernel into QEMU's bitmap, then we'll lose track of the 948 * guest modifications upon those pages (which can directly lead 949 * to guest data loss or panic after migration). 950 * 951 * Layout of the KVMSlot.dirty_bmap: 952 * 953 * |<-------- bmap_npages -----------..>| 954 * [1] 955 * start_delta size 956 * |----------------|-------------|------------------|------------| 957 * ^ ^ ^ ^ 958 * | | | | 959 * start bmap_start (start) end 960 * of memslot of memslot 961 * 962 * [1] bmap_npages can be aligned to either 64 pages or the end of slot 963 */ 964 965 assert(bmap_start % BITS_PER_LONG == 0); 966 /* We should never do log_clear before log_sync */ 967 assert(mem->dirty_bmap); 968 if (start_delta || bmap_npages - size / psize) { 969 /* Slow path - we need to manipulate a temp bitmap */ 970 bmap_clear = bitmap_new(bmap_npages); 971 bitmap_copy_with_src_offset(bmap_clear, mem->dirty_bmap, 972 bmap_start, start_delta + size / psize); 973 /* 974 * We need to fill the holes at start because that was not 975 * specified by the caller and we extended the bitmap only for 976 * 64 pages alignment 977 */ 978 bitmap_clear(bmap_clear, 0, start_delta); 979 d.dirty_bitmap = bmap_clear; 980 } else { 981 /* 982 * Fast path - both start and size align well with BITS_PER_LONG 983 * (or the end of memory slot) 984 */ 985 d.dirty_bitmap = mem->dirty_bmap + BIT_WORD(bmap_start); 986 } 987 988 d.first_page = bmap_start; 989 /* It should never overflow. If it happens, say something */ 990 assert(bmap_npages <= UINT32_MAX); 991 d.num_pages = bmap_npages; 992 d.slot = mem->slot | (as_id << 16); 993 994 ret = kvm_vm_ioctl(s, KVM_CLEAR_DIRTY_LOG, &d); 995 if (ret < 0 && ret != -ENOENT) { 996 error_report("%s: KVM_CLEAR_DIRTY_LOG failed, slot=%d, " 997 "start=0x%"PRIx64", size=0x%"PRIx32", errno=%d", 998 __func__, d.slot, (uint64_t)d.first_page, 999 (uint32_t)d.num_pages, ret); 1000 } else { 1001 ret = 0; 1002 trace_kvm_clear_dirty_log(d.slot, d.first_page, d.num_pages); 1003 } 1004 1005 /* 1006 * After we have updated the remote dirty bitmap, we update the 1007 * cached bitmap as well for the memslot, then if another user 1008 * clears the same region we know we shouldn't clear it again on 1009 * the remote otherwise it's data loss as well. 1010 */ 1011 bitmap_clear(mem->dirty_bmap, bmap_start + start_delta, 1012 size / psize); 1013 /* This handles the NULL case well */ 1014 g_free(bmap_clear); 1015 return ret; 1016 } 1017 1018 1019 /** 1020 * kvm_physical_log_clear - Clear the kernel's dirty bitmap for range 1021 * 1022 * NOTE: this will be a no-op if we haven't enabled manual dirty log 1023 * protection in the host kernel because in that case this operation 1024 * will be done within log_sync(). 1025 * 1026 * @kml: the kvm memory listener 1027 * @section: the memory range to clear dirty bitmap 1028 */ 1029 static int kvm_physical_log_clear(KVMMemoryListener *kml, 1030 MemoryRegionSection *section) 1031 { 1032 KVMState *s = kvm_state; 1033 uint64_t start, size, offset, count; 1034 KVMSlot *mem; 1035 int ret = 0, i; 1036 1037 if (!s->manual_dirty_log_protect) { 1038 /* No need to do explicit clear */ 1039 return ret; 1040 } 1041 1042 start = section->offset_within_address_space; 1043 size = int128_get64(section->size); 1044 1045 if (!size) { 1046 /* Nothing more we can do... */ 1047 return ret; 1048 } 1049 1050 kvm_slots_lock(); 1051 1052 for (i = 0; i < s->nr_slots; i++) { 1053 mem = &kml->slots[i]; 1054 /* Discard slots that are empty or do not overlap the section */ 1055 if (!mem->memory_size || 1056 mem->start_addr > start + size - 1 || 1057 start > mem->start_addr + mem->memory_size - 1) { 1058 continue; 1059 } 1060 1061 if (start >= mem->start_addr) { 1062 /* The slot starts before section or is aligned to it. */ 1063 offset = start - mem->start_addr; 1064 count = MIN(mem->memory_size - offset, size); 1065 } else { 1066 /* The slot starts after section. */ 1067 offset = 0; 1068 count = MIN(mem->memory_size, size - (mem->start_addr - start)); 1069 } 1070 ret = kvm_log_clear_one_slot(mem, kml->as_id, offset, count); 1071 if (ret < 0) { 1072 break; 1073 } 1074 } 1075 1076 kvm_slots_unlock(); 1077 1078 return ret; 1079 } 1080 1081 static void kvm_coalesce_mmio_region(MemoryListener *listener, 1082 MemoryRegionSection *secion, 1083 hwaddr start, hwaddr size) 1084 { 1085 KVMState *s = kvm_state; 1086 1087 if (s->coalesced_mmio) { 1088 struct kvm_coalesced_mmio_zone zone; 1089 1090 zone.addr = start; 1091 zone.size = size; 1092 zone.pad = 0; 1093 1094 (void)kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone); 1095 } 1096 } 1097 1098 static void kvm_uncoalesce_mmio_region(MemoryListener *listener, 1099 MemoryRegionSection *secion, 1100 hwaddr start, hwaddr size) 1101 { 1102 KVMState *s = kvm_state; 1103 1104 if (s->coalesced_mmio) { 1105 struct kvm_coalesced_mmio_zone zone; 1106 1107 zone.addr = start; 1108 zone.size = size; 1109 zone.pad = 0; 1110 1111 (void)kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone); 1112 } 1113 } 1114 1115 static void kvm_coalesce_pio_add(MemoryListener *listener, 1116 MemoryRegionSection *section, 1117 hwaddr start, hwaddr size) 1118 { 1119 KVMState *s = kvm_state; 1120 1121 if (s->coalesced_pio) { 1122 struct kvm_coalesced_mmio_zone zone; 1123 1124 zone.addr = start; 1125 zone.size = size; 1126 zone.pio = 1; 1127 1128 (void)kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone); 1129 } 1130 } 1131 1132 static void kvm_coalesce_pio_del(MemoryListener *listener, 1133 MemoryRegionSection *section, 1134 hwaddr start, hwaddr size) 1135 { 1136 KVMState *s = kvm_state; 1137 1138 if (s->coalesced_pio) { 1139 struct kvm_coalesced_mmio_zone zone; 1140 1141 zone.addr = start; 1142 zone.size = size; 1143 zone.pio = 1; 1144 1145 (void)kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone); 1146 } 1147 } 1148 1149 static MemoryListener kvm_coalesced_pio_listener = { 1150 .name = "kvm-coalesced-pio", 1151 .coalesced_io_add = kvm_coalesce_pio_add, 1152 .coalesced_io_del = kvm_coalesce_pio_del, 1153 }; 1154 1155 int kvm_check_extension(KVMState *s, unsigned int extension) 1156 { 1157 int ret; 1158 1159 ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, extension); 1160 if (ret < 0) { 1161 ret = 0; 1162 } 1163 1164 return ret; 1165 } 1166 1167 int kvm_vm_check_extension(KVMState *s, unsigned int extension) 1168 { 1169 int ret; 1170 1171 ret = kvm_vm_ioctl(s, KVM_CHECK_EXTENSION, extension); 1172 if (ret < 0) { 1173 /* VM wide version not implemented, use global one instead */ 1174 ret = kvm_check_extension(s, extension); 1175 } 1176 1177 return ret; 1178 } 1179 1180 typedef struct HWPoisonPage { 1181 ram_addr_t ram_addr; 1182 QLIST_ENTRY(HWPoisonPage) list; 1183 } HWPoisonPage; 1184 1185 static QLIST_HEAD(, HWPoisonPage) hwpoison_page_list = 1186 QLIST_HEAD_INITIALIZER(hwpoison_page_list); 1187 1188 static void kvm_unpoison_all(void *param) 1189 { 1190 HWPoisonPage *page, *next_page; 1191 1192 QLIST_FOREACH_SAFE(page, &hwpoison_page_list, list, next_page) { 1193 QLIST_REMOVE(page, list); 1194 qemu_ram_remap(page->ram_addr, TARGET_PAGE_SIZE); 1195 g_free(page); 1196 } 1197 } 1198 1199 void kvm_hwpoison_page_add(ram_addr_t ram_addr) 1200 { 1201 HWPoisonPage *page; 1202 1203 QLIST_FOREACH(page, &hwpoison_page_list, list) { 1204 if (page->ram_addr == ram_addr) { 1205 return; 1206 } 1207 } 1208 page = g_new(HWPoisonPage, 1); 1209 page->ram_addr = ram_addr; 1210 QLIST_INSERT_HEAD(&hwpoison_page_list, page, list); 1211 } 1212 1213 static uint32_t adjust_ioeventfd_endianness(uint32_t val, uint32_t size) 1214 { 1215 #if HOST_BIG_ENDIAN != TARGET_BIG_ENDIAN 1216 /* The kernel expects ioeventfd values in HOST_BIG_ENDIAN 1217 * endianness, but the memory core hands them in target endianness. 1218 * For example, PPC is always treated as big-endian even if running 1219 * on KVM and on PPC64LE. Correct here. 1220 */ 1221 switch (size) { 1222 case 2: 1223 val = bswap16(val); 1224 break; 1225 case 4: 1226 val = bswap32(val); 1227 break; 1228 } 1229 #endif 1230 return val; 1231 } 1232 1233 static int kvm_set_ioeventfd_mmio(int fd, hwaddr addr, uint32_t val, 1234 bool assign, uint32_t size, bool datamatch) 1235 { 1236 int ret; 1237 struct kvm_ioeventfd iofd = { 1238 .datamatch = datamatch ? adjust_ioeventfd_endianness(val, size) : 0, 1239 .addr = addr, 1240 .len = size, 1241 .flags = 0, 1242 .fd = fd, 1243 }; 1244 1245 trace_kvm_set_ioeventfd_mmio(fd, (uint64_t)addr, val, assign, size, 1246 datamatch); 1247 if (!kvm_enabled()) { 1248 return -ENOSYS; 1249 } 1250 1251 if (datamatch) { 1252 iofd.flags |= KVM_IOEVENTFD_FLAG_DATAMATCH; 1253 } 1254 if (!assign) { 1255 iofd.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN; 1256 } 1257 1258 ret = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &iofd); 1259 1260 if (ret < 0) { 1261 return -errno; 1262 } 1263 1264 return 0; 1265 } 1266 1267 static int kvm_set_ioeventfd_pio(int fd, uint16_t addr, uint16_t val, 1268 bool assign, uint32_t size, bool datamatch) 1269 { 1270 struct kvm_ioeventfd kick = { 1271 .datamatch = datamatch ? adjust_ioeventfd_endianness(val, size) : 0, 1272 .addr = addr, 1273 .flags = KVM_IOEVENTFD_FLAG_PIO, 1274 .len = size, 1275 .fd = fd, 1276 }; 1277 int r; 1278 trace_kvm_set_ioeventfd_pio(fd, addr, val, assign, size, datamatch); 1279 if (!kvm_enabled()) { 1280 return -ENOSYS; 1281 } 1282 if (datamatch) { 1283 kick.flags |= KVM_IOEVENTFD_FLAG_DATAMATCH; 1284 } 1285 if (!assign) { 1286 kick.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN; 1287 } 1288 r = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &kick); 1289 if (r < 0) { 1290 return r; 1291 } 1292 return 0; 1293 } 1294 1295 1296 static int kvm_check_many_ioeventfds(void) 1297 { 1298 /* Userspace can use ioeventfd for io notification. This requires a host 1299 * that supports eventfd(2) and an I/O thread; since eventfd does not 1300 * support SIGIO it cannot interrupt the vcpu. 1301 * 1302 * Older kernels have a 6 device limit on the KVM io bus. Find out so we 1303 * can avoid creating too many ioeventfds. 1304 */ 1305 #if defined(CONFIG_EVENTFD) 1306 int ioeventfds[7]; 1307 int i, ret = 0; 1308 for (i = 0; i < ARRAY_SIZE(ioeventfds); i++) { 1309 ioeventfds[i] = eventfd(0, EFD_CLOEXEC); 1310 if (ioeventfds[i] < 0) { 1311 break; 1312 } 1313 ret = kvm_set_ioeventfd_pio(ioeventfds[i], 0, i, true, 2, true); 1314 if (ret < 0) { 1315 close(ioeventfds[i]); 1316 break; 1317 } 1318 } 1319 1320 /* Decide whether many devices are supported or not */ 1321 ret = i == ARRAY_SIZE(ioeventfds); 1322 1323 while (i-- > 0) { 1324 kvm_set_ioeventfd_pio(ioeventfds[i], 0, i, false, 2, true); 1325 close(ioeventfds[i]); 1326 } 1327 return ret; 1328 #else 1329 return 0; 1330 #endif 1331 } 1332 1333 static const KVMCapabilityInfo * 1334 kvm_check_extension_list(KVMState *s, const KVMCapabilityInfo *list) 1335 { 1336 while (list->name) { 1337 if (!kvm_check_extension(s, list->value)) { 1338 return list; 1339 } 1340 list++; 1341 } 1342 return NULL; 1343 } 1344 1345 void kvm_set_max_memslot_size(hwaddr max_slot_size) 1346 { 1347 g_assert( 1348 ROUND_UP(max_slot_size, qemu_real_host_page_size()) == max_slot_size 1349 ); 1350 kvm_max_slot_size = max_slot_size; 1351 } 1352 1353 static void kvm_set_phys_mem(KVMMemoryListener *kml, 1354 MemoryRegionSection *section, bool add) 1355 { 1356 KVMSlot *mem; 1357 int err; 1358 MemoryRegion *mr = section->mr; 1359 bool writable = !mr->readonly && !mr->rom_device; 1360 hwaddr start_addr, size, slot_size, mr_offset; 1361 ram_addr_t ram_start_offset; 1362 void *ram; 1363 1364 if (!memory_region_is_ram(mr)) { 1365 if (writable || !kvm_readonly_mem_allowed) { 1366 return; 1367 } else if (!mr->romd_mode) { 1368 /* If the memory device is not in romd_mode, then we actually want 1369 * to remove the kvm memory slot so all accesses will trap. */ 1370 add = false; 1371 } 1372 } 1373 1374 size = kvm_align_section(section, &start_addr); 1375 if (!size) { 1376 return; 1377 } 1378 1379 /* The offset of the kvmslot within the memory region */ 1380 mr_offset = section->offset_within_region + start_addr - 1381 section->offset_within_address_space; 1382 1383 /* use aligned delta to align the ram address and offset */ 1384 ram = memory_region_get_ram_ptr(mr) + mr_offset; 1385 ram_start_offset = memory_region_get_ram_addr(mr) + mr_offset; 1386 1387 kvm_slots_lock(); 1388 1389 if (!add) { 1390 do { 1391 slot_size = MIN(kvm_max_slot_size, size); 1392 mem = kvm_lookup_matching_slot(kml, start_addr, slot_size); 1393 if (!mem) { 1394 goto out; 1395 } 1396 if (mem->flags & KVM_MEM_LOG_DIRTY_PAGES) { 1397 /* 1398 * NOTE: We should be aware of the fact that here we're only 1399 * doing a best effort to sync dirty bits. No matter whether 1400 * we're using dirty log or dirty ring, we ignored two facts: 1401 * 1402 * (1) dirty bits can reside in hardware buffers (PML) 1403 * 1404 * (2) after we collected dirty bits here, pages can be dirtied 1405 * again before we do the final KVM_SET_USER_MEMORY_REGION to 1406 * remove the slot. 1407 * 1408 * Not easy. Let's cross the fingers until it's fixed. 1409 */ 1410 if (kvm_state->kvm_dirty_ring_size) { 1411 kvm_dirty_ring_reap_locked(kvm_state, NULL); 1412 } else { 1413 kvm_slot_get_dirty_log(kvm_state, mem); 1414 } 1415 kvm_slot_sync_dirty_pages(mem); 1416 } 1417 1418 /* unregister the slot */ 1419 g_free(mem->dirty_bmap); 1420 mem->dirty_bmap = NULL; 1421 mem->memory_size = 0; 1422 mem->flags = 0; 1423 err = kvm_set_user_memory_region(kml, mem, false); 1424 if (err) { 1425 fprintf(stderr, "%s: error unregistering slot: %s\n", 1426 __func__, strerror(-err)); 1427 abort(); 1428 } 1429 start_addr += slot_size; 1430 size -= slot_size; 1431 } while (size); 1432 goto out; 1433 } 1434 1435 /* register the new slot */ 1436 do { 1437 slot_size = MIN(kvm_max_slot_size, size); 1438 mem = kvm_alloc_slot(kml); 1439 mem->as_id = kml->as_id; 1440 mem->memory_size = slot_size; 1441 mem->start_addr = start_addr; 1442 mem->ram_start_offset = ram_start_offset; 1443 mem->ram = ram; 1444 mem->flags = kvm_mem_flags(mr); 1445 kvm_slot_init_dirty_bitmap(mem); 1446 err = kvm_set_user_memory_region(kml, mem, true); 1447 if (err) { 1448 fprintf(stderr, "%s: error registering slot: %s\n", __func__, 1449 strerror(-err)); 1450 abort(); 1451 } 1452 start_addr += slot_size; 1453 ram_start_offset += slot_size; 1454 ram += slot_size; 1455 size -= slot_size; 1456 } while (size); 1457 1458 out: 1459 kvm_slots_unlock(); 1460 } 1461 1462 static void *kvm_dirty_ring_reaper_thread(void *data) 1463 { 1464 KVMState *s = data; 1465 struct KVMDirtyRingReaper *r = &s->reaper; 1466 1467 rcu_register_thread(); 1468 1469 trace_kvm_dirty_ring_reaper("init"); 1470 1471 while (true) { 1472 r->reaper_state = KVM_DIRTY_RING_REAPER_WAIT; 1473 trace_kvm_dirty_ring_reaper("wait"); 1474 /* 1475 * TODO: provide a smarter timeout rather than a constant? 1476 */ 1477 sleep(1); 1478 1479 /* keep sleeping so that dirtylimit not be interfered by reaper */ 1480 if (dirtylimit_in_service()) { 1481 continue; 1482 } 1483 1484 trace_kvm_dirty_ring_reaper("wakeup"); 1485 r->reaper_state = KVM_DIRTY_RING_REAPER_REAPING; 1486 1487 qemu_mutex_lock_iothread(); 1488 kvm_dirty_ring_reap(s, NULL); 1489 qemu_mutex_unlock_iothread(); 1490 1491 r->reaper_iteration++; 1492 } 1493 1494 trace_kvm_dirty_ring_reaper("exit"); 1495 1496 rcu_unregister_thread(); 1497 1498 return NULL; 1499 } 1500 1501 static int kvm_dirty_ring_reaper_init(KVMState *s) 1502 { 1503 struct KVMDirtyRingReaper *r = &s->reaper; 1504 1505 qemu_thread_create(&r->reaper_thr, "kvm-reaper", 1506 kvm_dirty_ring_reaper_thread, 1507 s, QEMU_THREAD_JOINABLE); 1508 1509 return 0; 1510 } 1511 1512 static void kvm_region_add(MemoryListener *listener, 1513 MemoryRegionSection *section) 1514 { 1515 KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener); 1516 1517 memory_region_ref(section->mr); 1518 kvm_set_phys_mem(kml, section, true); 1519 } 1520 1521 static void kvm_region_del(MemoryListener *listener, 1522 MemoryRegionSection *section) 1523 { 1524 KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener); 1525 1526 kvm_set_phys_mem(kml, section, false); 1527 memory_region_unref(section->mr); 1528 } 1529 1530 static void kvm_log_sync(MemoryListener *listener, 1531 MemoryRegionSection *section) 1532 { 1533 KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener); 1534 1535 kvm_slots_lock(); 1536 kvm_physical_sync_dirty_bitmap(kml, section); 1537 kvm_slots_unlock(); 1538 } 1539 1540 static void kvm_log_sync_global(MemoryListener *l) 1541 { 1542 KVMMemoryListener *kml = container_of(l, KVMMemoryListener, listener); 1543 KVMState *s = kvm_state; 1544 KVMSlot *mem; 1545 int i; 1546 1547 /* Flush all kernel dirty addresses into KVMSlot dirty bitmap */ 1548 kvm_dirty_ring_flush(); 1549 1550 /* 1551 * TODO: make this faster when nr_slots is big while there are 1552 * only a few used slots (small VMs). 1553 */ 1554 kvm_slots_lock(); 1555 for (i = 0; i < s->nr_slots; i++) { 1556 mem = &kml->slots[i]; 1557 if (mem->memory_size && mem->flags & KVM_MEM_LOG_DIRTY_PAGES) { 1558 kvm_slot_sync_dirty_pages(mem); 1559 /* 1560 * This is not needed by KVM_GET_DIRTY_LOG because the 1561 * ioctl will unconditionally overwrite the whole region. 1562 * However kvm dirty ring has no such side effect. 1563 */ 1564 kvm_slot_reset_dirty_pages(mem); 1565 } 1566 } 1567 kvm_slots_unlock(); 1568 } 1569 1570 static void kvm_log_clear(MemoryListener *listener, 1571 MemoryRegionSection *section) 1572 { 1573 KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener); 1574 int r; 1575 1576 r = kvm_physical_log_clear(kml, section); 1577 if (r < 0) { 1578 error_report_once("%s: kvm log clear failed: mr=%s " 1579 "offset=%"HWADDR_PRIx" size=%"PRIx64, __func__, 1580 section->mr->name, section->offset_within_region, 1581 int128_get64(section->size)); 1582 abort(); 1583 } 1584 } 1585 1586 static void kvm_mem_ioeventfd_add(MemoryListener *listener, 1587 MemoryRegionSection *section, 1588 bool match_data, uint64_t data, 1589 EventNotifier *e) 1590 { 1591 int fd = event_notifier_get_fd(e); 1592 int r; 1593 1594 r = kvm_set_ioeventfd_mmio(fd, section->offset_within_address_space, 1595 data, true, int128_get64(section->size), 1596 match_data); 1597 if (r < 0) { 1598 fprintf(stderr, "%s: error adding ioeventfd: %s (%d)\n", 1599 __func__, strerror(-r), -r); 1600 abort(); 1601 } 1602 } 1603 1604 static void kvm_mem_ioeventfd_del(MemoryListener *listener, 1605 MemoryRegionSection *section, 1606 bool match_data, uint64_t data, 1607 EventNotifier *e) 1608 { 1609 int fd = event_notifier_get_fd(e); 1610 int r; 1611 1612 r = kvm_set_ioeventfd_mmio(fd, section->offset_within_address_space, 1613 data, false, int128_get64(section->size), 1614 match_data); 1615 if (r < 0) { 1616 fprintf(stderr, "%s: error deleting ioeventfd: %s (%d)\n", 1617 __func__, strerror(-r), -r); 1618 abort(); 1619 } 1620 } 1621 1622 static void kvm_io_ioeventfd_add(MemoryListener *listener, 1623 MemoryRegionSection *section, 1624 bool match_data, uint64_t data, 1625 EventNotifier *e) 1626 { 1627 int fd = event_notifier_get_fd(e); 1628 int r; 1629 1630 r = kvm_set_ioeventfd_pio(fd, section->offset_within_address_space, 1631 data, true, int128_get64(section->size), 1632 match_data); 1633 if (r < 0) { 1634 fprintf(stderr, "%s: error adding ioeventfd: %s (%d)\n", 1635 __func__, strerror(-r), -r); 1636 abort(); 1637 } 1638 } 1639 1640 static void kvm_io_ioeventfd_del(MemoryListener *listener, 1641 MemoryRegionSection *section, 1642 bool match_data, uint64_t data, 1643 EventNotifier *e) 1644 1645 { 1646 int fd = event_notifier_get_fd(e); 1647 int r; 1648 1649 r = kvm_set_ioeventfd_pio(fd, section->offset_within_address_space, 1650 data, false, int128_get64(section->size), 1651 match_data); 1652 if (r < 0) { 1653 fprintf(stderr, "%s: error deleting ioeventfd: %s (%d)\n", 1654 __func__, strerror(-r), -r); 1655 abort(); 1656 } 1657 } 1658 1659 void kvm_memory_listener_register(KVMState *s, KVMMemoryListener *kml, 1660 AddressSpace *as, int as_id, const char *name) 1661 { 1662 int i; 1663 1664 kml->slots = g_new0(KVMSlot, s->nr_slots); 1665 kml->as_id = as_id; 1666 1667 for (i = 0; i < s->nr_slots; i++) { 1668 kml->slots[i].slot = i; 1669 } 1670 1671 kml->listener.region_add = kvm_region_add; 1672 kml->listener.region_del = kvm_region_del; 1673 kml->listener.log_start = kvm_log_start; 1674 kml->listener.log_stop = kvm_log_stop; 1675 kml->listener.priority = 10; 1676 kml->listener.name = name; 1677 1678 if (s->kvm_dirty_ring_size) { 1679 kml->listener.log_sync_global = kvm_log_sync_global; 1680 } else { 1681 kml->listener.log_sync = kvm_log_sync; 1682 kml->listener.log_clear = kvm_log_clear; 1683 } 1684 1685 memory_listener_register(&kml->listener, as); 1686 1687 for (i = 0; i < s->nr_as; ++i) { 1688 if (!s->as[i].as) { 1689 s->as[i].as = as; 1690 s->as[i].ml = kml; 1691 break; 1692 } 1693 } 1694 } 1695 1696 static MemoryListener kvm_io_listener = { 1697 .name = "kvm-io", 1698 .eventfd_add = kvm_io_ioeventfd_add, 1699 .eventfd_del = kvm_io_ioeventfd_del, 1700 .priority = 10, 1701 }; 1702 1703 int kvm_set_irq(KVMState *s, int irq, int level) 1704 { 1705 struct kvm_irq_level event; 1706 int ret; 1707 1708 assert(kvm_async_interrupts_enabled()); 1709 1710 event.level = level; 1711 event.irq = irq; 1712 ret = kvm_vm_ioctl(s, s->irq_set_ioctl, &event); 1713 if (ret < 0) { 1714 perror("kvm_set_irq"); 1715 abort(); 1716 } 1717 1718 return (s->irq_set_ioctl == KVM_IRQ_LINE) ? 1 : event.status; 1719 } 1720 1721 #ifdef KVM_CAP_IRQ_ROUTING 1722 typedef struct KVMMSIRoute { 1723 struct kvm_irq_routing_entry kroute; 1724 QTAILQ_ENTRY(KVMMSIRoute) entry; 1725 } KVMMSIRoute; 1726 1727 static void set_gsi(KVMState *s, unsigned int gsi) 1728 { 1729 set_bit(gsi, s->used_gsi_bitmap); 1730 } 1731 1732 static void clear_gsi(KVMState *s, unsigned int gsi) 1733 { 1734 clear_bit(gsi, s->used_gsi_bitmap); 1735 } 1736 1737 void kvm_init_irq_routing(KVMState *s) 1738 { 1739 int gsi_count, i; 1740 1741 gsi_count = kvm_check_extension(s, KVM_CAP_IRQ_ROUTING) - 1; 1742 if (gsi_count > 0) { 1743 /* Round up so we can search ints using ffs */ 1744 s->used_gsi_bitmap = bitmap_new(gsi_count); 1745 s->gsi_count = gsi_count; 1746 } 1747 1748 s->irq_routes = g_malloc0(sizeof(*s->irq_routes)); 1749 s->nr_allocated_irq_routes = 0; 1750 1751 if (!kvm_direct_msi_allowed) { 1752 for (i = 0; i < KVM_MSI_HASHTAB_SIZE; i++) { 1753 QTAILQ_INIT(&s->msi_hashtab[i]); 1754 } 1755 } 1756 1757 kvm_arch_init_irq_routing(s); 1758 } 1759 1760 void kvm_irqchip_commit_routes(KVMState *s) 1761 { 1762 int ret; 1763 1764 if (kvm_gsi_direct_mapping()) { 1765 return; 1766 } 1767 1768 if (!kvm_gsi_routing_enabled()) { 1769 return; 1770 } 1771 1772 s->irq_routes->flags = 0; 1773 trace_kvm_irqchip_commit_routes(); 1774 ret = kvm_vm_ioctl(s, KVM_SET_GSI_ROUTING, s->irq_routes); 1775 assert(ret == 0); 1776 } 1777 1778 static void kvm_add_routing_entry(KVMState *s, 1779 struct kvm_irq_routing_entry *entry) 1780 { 1781 struct kvm_irq_routing_entry *new; 1782 int n, size; 1783 1784 if (s->irq_routes->nr == s->nr_allocated_irq_routes) { 1785 n = s->nr_allocated_irq_routes * 2; 1786 if (n < 64) { 1787 n = 64; 1788 } 1789 size = sizeof(struct kvm_irq_routing); 1790 size += n * sizeof(*new); 1791 s->irq_routes = g_realloc(s->irq_routes, size); 1792 s->nr_allocated_irq_routes = n; 1793 } 1794 n = s->irq_routes->nr++; 1795 new = &s->irq_routes->entries[n]; 1796 1797 *new = *entry; 1798 1799 set_gsi(s, entry->gsi); 1800 } 1801 1802 static int kvm_update_routing_entry(KVMState *s, 1803 struct kvm_irq_routing_entry *new_entry) 1804 { 1805 struct kvm_irq_routing_entry *entry; 1806 int n; 1807 1808 for (n = 0; n < s->irq_routes->nr; n++) { 1809 entry = &s->irq_routes->entries[n]; 1810 if (entry->gsi != new_entry->gsi) { 1811 continue; 1812 } 1813 1814 if(!memcmp(entry, new_entry, sizeof *entry)) { 1815 return 0; 1816 } 1817 1818 *entry = *new_entry; 1819 1820 return 0; 1821 } 1822 1823 return -ESRCH; 1824 } 1825 1826 void kvm_irqchip_add_irq_route(KVMState *s, int irq, int irqchip, int pin) 1827 { 1828 struct kvm_irq_routing_entry e = {}; 1829 1830 assert(pin < s->gsi_count); 1831 1832 e.gsi = irq; 1833 e.type = KVM_IRQ_ROUTING_IRQCHIP; 1834 e.flags = 0; 1835 e.u.irqchip.irqchip = irqchip; 1836 e.u.irqchip.pin = pin; 1837 kvm_add_routing_entry(s, &e); 1838 } 1839 1840 void kvm_irqchip_release_virq(KVMState *s, int virq) 1841 { 1842 struct kvm_irq_routing_entry *e; 1843 int i; 1844 1845 if (kvm_gsi_direct_mapping()) { 1846 return; 1847 } 1848 1849 for (i = 0; i < s->irq_routes->nr; i++) { 1850 e = &s->irq_routes->entries[i]; 1851 if (e->gsi == virq) { 1852 s->irq_routes->nr--; 1853 *e = s->irq_routes->entries[s->irq_routes->nr]; 1854 } 1855 } 1856 clear_gsi(s, virq); 1857 kvm_arch_release_virq_post(virq); 1858 trace_kvm_irqchip_release_virq(virq); 1859 } 1860 1861 void kvm_irqchip_add_change_notifier(Notifier *n) 1862 { 1863 notifier_list_add(&kvm_irqchip_change_notifiers, n); 1864 } 1865 1866 void kvm_irqchip_remove_change_notifier(Notifier *n) 1867 { 1868 notifier_remove(n); 1869 } 1870 1871 void kvm_irqchip_change_notify(void) 1872 { 1873 notifier_list_notify(&kvm_irqchip_change_notifiers, NULL); 1874 } 1875 1876 static unsigned int kvm_hash_msi(uint32_t data) 1877 { 1878 /* This is optimized for IA32 MSI layout. However, no other arch shall 1879 * repeat the mistake of not providing a direct MSI injection API. */ 1880 return data & 0xff; 1881 } 1882 1883 static void kvm_flush_dynamic_msi_routes(KVMState *s) 1884 { 1885 KVMMSIRoute *route, *next; 1886 unsigned int hash; 1887 1888 for (hash = 0; hash < KVM_MSI_HASHTAB_SIZE; hash++) { 1889 QTAILQ_FOREACH_SAFE(route, &s->msi_hashtab[hash], entry, next) { 1890 kvm_irqchip_release_virq(s, route->kroute.gsi); 1891 QTAILQ_REMOVE(&s->msi_hashtab[hash], route, entry); 1892 g_free(route); 1893 } 1894 } 1895 } 1896 1897 static int kvm_irqchip_get_virq(KVMState *s) 1898 { 1899 int next_virq; 1900 1901 /* 1902 * PIC and IOAPIC share the first 16 GSI numbers, thus the available 1903 * GSI numbers are more than the number of IRQ route. Allocating a GSI 1904 * number can succeed even though a new route entry cannot be added. 1905 * When this happens, flush dynamic MSI entries to free IRQ route entries. 1906 */ 1907 if (!kvm_direct_msi_allowed && s->irq_routes->nr == s->gsi_count) { 1908 kvm_flush_dynamic_msi_routes(s); 1909 } 1910 1911 /* Return the lowest unused GSI in the bitmap */ 1912 next_virq = find_first_zero_bit(s->used_gsi_bitmap, s->gsi_count); 1913 if (next_virq >= s->gsi_count) { 1914 return -ENOSPC; 1915 } else { 1916 return next_virq; 1917 } 1918 } 1919 1920 static KVMMSIRoute *kvm_lookup_msi_route(KVMState *s, MSIMessage msg) 1921 { 1922 unsigned int hash = kvm_hash_msi(msg.data); 1923 KVMMSIRoute *route; 1924 1925 QTAILQ_FOREACH(route, &s->msi_hashtab[hash], entry) { 1926 if (route->kroute.u.msi.address_lo == (uint32_t)msg.address && 1927 route->kroute.u.msi.address_hi == (msg.address >> 32) && 1928 route->kroute.u.msi.data == le32_to_cpu(msg.data)) { 1929 return route; 1930 } 1931 } 1932 return NULL; 1933 } 1934 1935 int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg) 1936 { 1937 struct kvm_msi msi; 1938 KVMMSIRoute *route; 1939 1940 if (kvm_direct_msi_allowed) { 1941 msi.address_lo = (uint32_t)msg.address; 1942 msi.address_hi = msg.address >> 32; 1943 msi.data = le32_to_cpu(msg.data); 1944 msi.flags = 0; 1945 memset(msi.pad, 0, sizeof(msi.pad)); 1946 1947 return kvm_vm_ioctl(s, KVM_SIGNAL_MSI, &msi); 1948 } 1949 1950 route = kvm_lookup_msi_route(s, msg); 1951 if (!route) { 1952 int virq; 1953 1954 virq = kvm_irqchip_get_virq(s); 1955 if (virq < 0) { 1956 return virq; 1957 } 1958 1959 route = g_new0(KVMMSIRoute, 1); 1960 route->kroute.gsi = virq; 1961 route->kroute.type = KVM_IRQ_ROUTING_MSI; 1962 route->kroute.flags = 0; 1963 route->kroute.u.msi.address_lo = (uint32_t)msg.address; 1964 route->kroute.u.msi.address_hi = msg.address >> 32; 1965 route->kroute.u.msi.data = le32_to_cpu(msg.data); 1966 1967 kvm_add_routing_entry(s, &route->kroute); 1968 kvm_irqchip_commit_routes(s); 1969 1970 QTAILQ_INSERT_TAIL(&s->msi_hashtab[kvm_hash_msi(msg.data)], route, 1971 entry); 1972 } 1973 1974 assert(route->kroute.type == KVM_IRQ_ROUTING_MSI); 1975 1976 return kvm_set_irq(s, route->kroute.gsi, 1); 1977 } 1978 1979 int kvm_irqchip_add_msi_route(KVMRouteChange *c, int vector, PCIDevice *dev) 1980 { 1981 struct kvm_irq_routing_entry kroute = {}; 1982 int virq; 1983 KVMState *s = c->s; 1984 MSIMessage msg = {0, 0}; 1985 1986 if (pci_available && dev) { 1987 msg = pci_get_msi_message(dev, vector); 1988 } 1989 1990 if (kvm_gsi_direct_mapping()) { 1991 return kvm_arch_msi_data_to_gsi(msg.data); 1992 } 1993 1994 if (!kvm_gsi_routing_enabled()) { 1995 return -ENOSYS; 1996 } 1997 1998 virq = kvm_irqchip_get_virq(s); 1999 if (virq < 0) { 2000 return virq; 2001 } 2002 2003 kroute.gsi = virq; 2004 kroute.type = KVM_IRQ_ROUTING_MSI; 2005 kroute.flags = 0; 2006 kroute.u.msi.address_lo = (uint32_t)msg.address; 2007 kroute.u.msi.address_hi = msg.address >> 32; 2008 kroute.u.msi.data = le32_to_cpu(msg.data); 2009 if (pci_available && kvm_msi_devid_required()) { 2010 kroute.flags = KVM_MSI_VALID_DEVID; 2011 kroute.u.msi.devid = pci_requester_id(dev); 2012 } 2013 if (kvm_arch_fixup_msi_route(&kroute, msg.address, msg.data, dev)) { 2014 kvm_irqchip_release_virq(s, virq); 2015 return -EINVAL; 2016 } 2017 2018 trace_kvm_irqchip_add_msi_route(dev ? dev->name : (char *)"N/A", 2019 vector, virq); 2020 2021 kvm_add_routing_entry(s, &kroute); 2022 kvm_arch_add_msi_route_post(&kroute, vector, dev); 2023 c->changes++; 2024 2025 return virq; 2026 } 2027 2028 int kvm_irqchip_update_msi_route(KVMState *s, int virq, MSIMessage msg, 2029 PCIDevice *dev) 2030 { 2031 struct kvm_irq_routing_entry kroute = {}; 2032 2033 if (kvm_gsi_direct_mapping()) { 2034 return 0; 2035 } 2036 2037 if (!kvm_irqchip_in_kernel()) { 2038 return -ENOSYS; 2039 } 2040 2041 kroute.gsi = virq; 2042 kroute.type = KVM_IRQ_ROUTING_MSI; 2043 kroute.flags = 0; 2044 kroute.u.msi.address_lo = (uint32_t)msg.address; 2045 kroute.u.msi.address_hi = msg.address >> 32; 2046 kroute.u.msi.data = le32_to_cpu(msg.data); 2047 if (pci_available && kvm_msi_devid_required()) { 2048 kroute.flags = KVM_MSI_VALID_DEVID; 2049 kroute.u.msi.devid = pci_requester_id(dev); 2050 } 2051 if (kvm_arch_fixup_msi_route(&kroute, msg.address, msg.data, dev)) { 2052 return -EINVAL; 2053 } 2054 2055 trace_kvm_irqchip_update_msi_route(virq); 2056 2057 return kvm_update_routing_entry(s, &kroute); 2058 } 2059 2060 static int kvm_irqchip_assign_irqfd(KVMState *s, EventNotifier *event, 2061 EventNotifier *resample, int virq, 2062 bool assign) 2063 { 2064 int fd = event_notifier_get_fd(event); 2065 int rfd = resample ? event_notifier_get_fd(resample) : -1; 2066 2067 struct kvm_irqfd irqfd = { 2068 .fd = fd, 2069 .gsi = virq, 2070 .flags = assign ? 0 : KVM_IRQFD_FLAG_DEASSIGN, 2071 }; 2072 2073 if (rfd != -1) { 2074 assert(assign); 2075 if (kvm_irqchip_is_split()) { 2076 /* 2077 * When the slow irqchip (e.g. IOAPIC) is in the 2078 * userspace, KVM kernel resamplefd will not work because 2079 * the EOI of the interrupt will be delivered to userspace 2080 * instead, so the KVM kernel resamplefd kick will be 2081 * skipped. The userspace here mimics what the kernel 2082 * provides with resamplefd, remember the resamplefd and 2083 * kick it when we receive EOI of this IRQ. 2084 * 2085 * This is hackery because IOAPIC is mostly bypassed 2086 * (except EOI broadcasts) when irqfd is used. However 2087 * this can bring much performance back for split irqchip 2088 * with INTx IRQs (for VFIO, this gives 93% perf of the 2089 * full fast path, which is 46% perf boost comparing to 2090 * the INTx slow path). 2091 */ 2092 kvm_resample_fd_insert(virq, resample); 2093 } else { 2094 irqfd.flags |= KVM_IRQFD_FLAG_RESAMPLE; 2095 irqfd.resamplefd = rfd; 2096 } 2097 } else if (!assign) { 2098 if (kvm_irqchip_is_split()) { 2099 kvm_resample_fd_remove(virq); 2100 } 2101 } 2102 2103 if (!kvm_irqfds_enabled()) { 2104 return -ENOSYS; 2105 } 2106 2107 return kvm_vm_ioctl(s, KVM_IRQFD, &irqfd); 2108 } 2109 2110 int kvm_irqchip_add_adapter_route(KVMState *s, AdapterInfo *adapter) 2111 { 2112 struct kvm_irq_routing_entry kroute = {}; 2113 int virq; 2114 2115 if (!kvm_gsi_routing_enabled()) { 2116 return -ENOSYS; 2117 } 2118 2119 virq = kvm_irqchip_get_virq(s); 2120 if (virq < 0) { 2121 return virq; 2122 } 2123 2124 kroute.gsi = virq; 2125 kroute.type = KVM_IRQ_ROUTING_S390_ADAPTER; 2126 kroute.flags = 0; 2127 kroute.u.adapter.summary_addr = adapter->summary_addr; 2128 kroute.u.adapter.ind_addr = adapter->ind_addr; 2129 kroute.u.adapter.summary_offset = adapter->summary_offset; 2130 kroute.u.adapter.ind_offset = adapter->ind_offset; 2131 kroute.u.adapter.adapter_id = adapter->adapter_id; 2132 2133 kvm_add_routing_entry(s, &kroute); 2134 2135 return virq; 2136 } 2137 2138 int kvm_irqchip_add_hv_sint_route(KVMState *s, uint32_t vcpu, uint32_t sint) 2139 { 2140 struct kvm_irq_routing_entry kroute = {}; 2141 int virq; 2142 2143 if (!kvm_gsi_routing_enabled()) { 2144 return -ENOSYS; 2145 } 2146 if (!kvm_check_extension(s, KVM_CAP_HYPERV_SYNIC)) { 2147 return -ENOSYS; 2148 } 2149 virq = kvm_irqchip_get_virq(s); 2150 if (virq < 0) { 2151 return virq; 2152 } 2153 2154 kroute.gsi = virq; 2155 kroute.type = KVM_IRQ_ROUTING_HV_SINT; 2156 kroute.flags = 0; 2157 kroute.u.hv_sint.vcpu = vcpu; 2158 kroute.u.hv_sint.sint = sint; 2159 2160 kvm_add_routing_entry(s, &kroute); 2161 kvm_irqchip_commit_routes(s); 2162 2163 return virq; 2164 } 2165 2166 #else /* !KVM_CAP_IRQ_ROUTING */ 2167 2168 void kvm_init_irq_routing(KVMState *s) 2169 { 2170 } 2171 2172 void kvm_irqchip_release_virq(KVMState *s, int virq) 2173 { 2174 } 2175 2176 int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg) 2177 { 2178 abort(); 2179 } 2180 2181 int kvm_irqchip_add_msi_route(KVMRouteChange *c, int vector, PCIDevice *dev) 2182 { 2183 return -ENOSYS; 2184 } 2185 2186 int kvm_irqchip_add_adapter_route(KVMState *s, AdapterInfo *adapter) 2187 { 2188 return -ENOSYS; 2189 } 2190 2191 int kvm_irqchip_add_hv_sint_route(KVMState *s, uint32_t vcpu, uint32_t sint) 2192 { 2193 return -ENOSYS; 2194 } 2195 2196 static int kvm_irqchip_assign_irqfd(KVMState *s, EventNotifier *event, 2197 EventNotifier *resample, int virq, 2198 bool assign) 2199 { 2200 abort(); 2201 } 2202 2203 int kvm_irqchip_update_msi_route(KVMState *s, int virq, MSIMessage msg) 2204 { 2205 return -ENOSYS; 2206 } 2207 #endif /* !KVM_CAP_IRQ_ROUTING */ 2208 2209 int kvm_irqchip_add_irqfd_notifier_gsi(KVMState *s, EventNotifier *n, 2210 EventNotifier *rn, int virq) 2211 { 2212 return kvm_irqchip_assign_irqfd(s, n, rn, virq, true); 2213 } 2214 2215 int kvm_irqchip_remove_irqfd_notifier_gsi(KVMState *s, EventNotifier *n, 2216 int virq) 2217 { 2218 return kvm_irqchip_assign_irqfd(s, n, NULL, virq, false); 2219 } 2220 2221 int kvm_irqchip_add_irqfd_notifier(KVMState *s, EventNotifier *n, 2222 EventNotifier *rn, qemu_irq irq) 2223 { 2224 gpointer key, gsi; 2225 gboolean found = g_hash_table_lookup_extended(s->gsimap, irq, &key, &gsi); 2226 2227 if (!found) { 2228 return -ENXIO; 2229 } 2230 return kvm_irqchip_add_irqfd_notifier_gsi(s, n, rn, GPOINTER_TO_INT(gsi)); 2231 } 2232 2233 int kvm_irqchip_remove_irqfd_notifier(KVMState *s, EventNotifier *n, 2234 qemu_irq irq) 2235 { 2236 gpointer key, gsi; 2237 gboolean found = g_hash_table_lookup_extended(s->gsimap, irq, &key, &gsi); 2238 2239 if (!found) { 2240 return -ENXIO; 2241 } 2242 return kvm_irqchip_remove_irqfd_notifier_gsi(s, n, GPOINTER_TO_INT(gsi)); 2243 } 2244 2245 void kvm_irqchip_set_qemuirq_gsi(KVMState *s, qemu_irq irq, int gsi) 2246 { 2247 g_hash_table_insert(s->gsimap, irq, GINT_TO_POINTER(gsi)); 2248 } 2249 2250 static void kvm_irqchip_create(KVMState *s) 2251 { 2252 int ret; 2253 2254 assert(s->kernel_irqchip_split != ON_OFF_AUTO_AUTO); 2255 if (kvm_check_extension(s, KVM_CAP_IRQCHIP)) { 2256 ; 2257 } else if (kvm_check_extension(s, KVM_CAP_S390_IRQCHIP)) { 2258 ret = kvm_vm_enable_cap(s, KVM_CAP_S390_IRQCHIP, 0); 2259 if (ret < 0) { 2260 fprintf(stderr, "Enable kernel irqchip failed: %s\n", strerror(-ret)); 2261 exit(1); 2262 } 2263 } else { 2264 return; 2265 } 2266 2267 /* First probe and see if there's a arch-specific hook to create the 2268 * in-kernel irqchip for us */ 2269 ret = kvm_arch_irqchip_create(s); 2270 if (ret == 0) { 2271 if (s->kernel_irqchip_split == ON_OFF_AUTO_ON) { 2272 error_report("Split IRQ chip mode not supported."); 2273 exit(1); 2274 } else { 2275 ret = kvm_vm_ioctl(s, KVM_CREATE_IRQCHIP); 2276 } 2277 } 2278 if (ret < 0) { 2279 fprintf(stderr, "Create kernel irqchip failed: %s\n", strerror(-ret)); 2280 exit(1); 2281 } 2282 2283 kvm_kernel_irqchip = true; 2284 /* If we have an in-kernel IRQ chip then we must have asynchronous 2285 * interrupt delivery (though the reverse is not necessarily true) 2286 */ 2287 kvm_async_interrupts_allowed = true; 2288 kvm_halt_in_kernel_allowed = true; 2289 2290 kvm_init_irq_routing(s); 2291 2292 s->gsimap = g_hash_table_new(g_direct_hash, g_direct_equal); 2293 } 2294 2295 /* Find number of supported CPUs using the recommended 2296 * procedure from the kernel API documentation to cope with 2297 * older kernels that may be missing capabilities. 2298 */ 2299 static int kvm_recommended_vcpus(KVMState *s) 2300 { 2301 int ret = kvm_vm_check_extension(s, KVM_CAP_NR_VCPUS); 2302 return (ret) ? ret : 4; 2303 } 2304 2305 static int kvm_max_vcpus(KVMState *s) 2306 { 2307 int ret = kvm_check_extension(s, KVM_CAP_MAX_VCPUS); 2308 return (ret) ? ret : kvm_recommended_vcpus(s); 2309 } 2310 2311 static int kvm_max_vcpu_id(KVMState *s) 2312 { 2313 int ret = kvm_check_extension(s, KVM_CAP_MAX_VCPU_ID); 2314 return (ret) ? ret : kvm_max_vcpus(s); 2315 } 2316 2317 bool kvm_vcpu_id_is_valid(int vcpu_id) 2318 { 2319 KVMState *s = KVM_STATE(current_accel()); 2320 return vcpu_id >= 0 && vcpu_id < kvm_max_vcpu_id(s); 2321 } 2322 2323 bool kvm_dirty_ring_enabled(void) 2324 { 2325 return kvm_state->kvm_dirty_ring_size ? true : false; 2326 } 2327 2328 static void query_stats_cb(StatsResultList **result, StatsTarget target, 2329 strList *names, strList *targets, Error **errp); 2330 static void query_stats_schemas_cb(StatsSchemaList **result, Error **errp); 2331 2332 uint32_t kvm_dirty_ring_size(void) 2333 { 2334 return kvm_state->kvm_dirty_ring_size; 2335 } 2336 2337 static int kvm_init(MachineState *ms) 2338 { 2339 MachineClass *mc = MACHINE_GET_CLASS(ms); 2340 static const char upgrade_note[] = 2341 "Please upgrade to at least kernel 2.6.29 or recent kvm-kmod\n" 2342 "(see http://sourceforge.net/projects/kvm).\n"; 2343 struct { 2344 const char *name; 2345 int num; 2346 } num_cpus[] = { 2347 { "SMP", ms->smp.cpus }, 2348 { "hotpluggable", ms->smp.max_cpus }, 2349 { NULL, } 2350 }, *nc = num_cpus; 2351 int soft_vcpus_limit, hard_vcpus_limit; 2352 KVMState *s; 2353 const KVMCapabilityInfo *missing_cap; 2354 int ret; 2355 int type = 0; 2356 uint64_t dirty_log_manual_caps; 2357 2358 qemu_mutex_init(&kml_slots_lock); 2359 2360 s = KVM_STATE(ms->accelerator); 2361 2362 /* 2363 * On systems where the kernel can support different base page 2364 * sizes, host page size may be different from TARGET_PAGE_SIZE, 2365 * even with KVM. TARGET_PAGE_SIZE is assumed to be the minimum 2366 * page size for the system though. 2367 */ 2368 assert(TARGET_PAGE_SIZE <= qemu_real_host_page_size()); 2369 2370 s->sigmask_len = 8; 2371 2372 #ifdef KVM_CAP_SET_GUEST_DEBUG 2373 QTAILQ_INIT(&s->kvm_sw_breakpoints); 2374 #endif 2375 QLIST_INIT(&s->kvm_parked_vcpus); 2376 s->fd = qemu_open_old("/dev/kvm", O_RDWR); 2377 if (s->fd == -1) { 2378 fprintf(stderr, "Could not access KVM kernel module: %m\n"); 2379 ret = -errno; 2380 goto err; 2381 } 2382 2383 ret = kvm_ioctl(s, KVM_GET_API_VERSION, 0); 2384 if (ret < KVM_API_VERSION) { 2385 if (ret >= 0) { 2386 ret = -EINVAL; 2387 } 2388 fprintf(stderr, "kvm version too old\n"); 2389 goto err; 2390 } 2391 2392 if (ret > KVM_API_VERSION) { 2393 ret = -EINVAL; 2394 fprintf(stderr, "kvm version not supported\n"); 2395 goto err; 2396 } 2397 2398 kvm_immediate_exit = kvm_check_extension(s, KVM_CAP_IMMEDIATE_EXIT); 2399 s->nr_slots = kvm_check_extension(s, KVM_CAP_NR_MEMSLOTS); 2400 2401 /* If unspecified, use the default value */ 2402 if (!s->nr_slots) { 2403 s->nr_slots = 32; 2404 } 2405 2406 s->nr_as = kvm_check_extension(s, KVM_CAP_MULTI_ADDRESS_SPACE); 2407 if (s->nr_as <= 1) { 2408 s->nr_as = 1; 2409 } 2410 s->as = g_new0(struct KVMAs, s->nr_as); 2411 2412 if (object_property_find(OBJECT(current_machine), "kvm-type")) { 2413 g_autofree char *kvm_type = object_property_get_str(OBJECT(current_machine), 2414 "kvm-type", 2415 &error_abort); 2416 type = mc->kvm_type(ms, kvm_type); 2417 } else if (mc->kvm_type) { 2418 type = mc->kvm_type(ms, NULL); 2419 } 2420 2421 do { 2422 ret = kvm_ioctl(s, KVM_CREATE_VM, type); 2423 } while (ret == -EINTR); 2424 2425 if (ret < 0) { 2426 fprintf(stderr, "ioctl(KVM_CREATE_VM) failed: %d %s\n", -ret, 2427 strerror(-ret)); 2428 2429 #ifdef TARGET_S390X 2430 if (ret == -EINVAL) { 2431 fprintf(stderr, 2432 "Host kernel setup problem detected. Please verify:\n"); 2433 fprintf(stderr, "- for kernels supporting the switch_amode or" 2434 " user_mode parameters, whether\n"); 2435 fprintf(stderr, 2436 " user space is running in primary address space\n"); 2437 fprintf(stderr, 2438 "- for kernels supporting the vm.allocate_pgste sysctl, " 2439 "whether it is enabled\n"); 2440 } 2441 #elif defined(TARGET_PPC) 2442 if (ret == -EINVAL) { 2443 fprintf(stderr, 2444 "PPC KVM module is not loaded. Try modprobe kvm_%s.\n", 2445 (type == 2) ? "pr" : "hv"); 2446 } 2447 #endif 2448 goto err; 2449 } 2450 2451 s->vmfd = ret; 2452 2453 /* check the vcpu limits */ 2454 soft_vcpus_limit = kvm_recommended_vcpus(s); 2455 hard_vcpus_limit = kvm_max_vcpus(s); 2456 2457 while (nc->name) { 2458 if (nc->num > soft_vcpus_limit) { 2459 warn_report("Number of %s cpus requested (%d) exceeds " 2460 "the recommended cpus supported by KVM (%d)", 2461 nc->name, nc->num, soft_vcpus_limit); 2462 2463 if (nc->num > hard_vcpus_limit) { 2464 fprintf(stderr, "Number of %s cpus requested (%d) exceeds " 2465 "the maximum cpus supported by KVM (%d)\n", 2466 nc->name, nc->num, hard_vcpus_limit); 2467 exit(1); 2468 } 2469 } 2470 nc++; 2471 } 2472 2473 missing_cap = kvm_check_extension_list(s, kvm_required_capabilites); 2474 if (!missing_cap) { 2475 missing_cap = 2476 kvm_check_extension_list(s, kvm_arch_required_capabilities); 2477 } 2478 if (missing_cap) { 2479 ret = -EINVAL; 2480 fprintf(stderr, "kvm does not support %s\n%s", 2481 missing_cap->name, upgrade_note); 2482 goto err; 2483 } 2484 2485 s->coalesced_mmio = kvm_check_extension(s, KVM_CAP_COALESCED_MMIO); 2486 s->coalesced_pio = s->coalesced_mmio && 2487 kvm_check_extension(s, KVM_CAP_COALESCED_PIO); 2488 2489 /* 2490 * Enable KVM dirty ring if supported, otherwise fall back to 2491 * dirty logging mode 2492 */ 2493 if (s->kvm_dirty_ring_size > 0) { 2494 uint64_t ring_bytes; 2495 2496 ring_bytes = s->kvm_dirty_ring_size * sizeof(struct kvm_dirty_gfn); 2497 2498 /* Read the max supported pages */ 2499 ret = kvm_vm_check_extension(s, KVM_CAP_DIRTY_LOG_RING); 2500 if (ret > 0) { 2501 if (ring_bytes > ret) { 2502 error_report("KVM dirty ring size %" PRIu32 " too big " 2503 "(maximum is %ld). Please use a smaller value.", 2504 s->kvm_dirty_ring_size, 2505 (long)ret / sizeof(struct kvm_dirty_gfn)); 2506 ret = -EINVAL; 2507 goto err; 2508 } 2509 2510 ret = kvm_vm_enable_cap(s, KVM_CAP_DIRTY_LOG_RING, 0, ring_bytes); 2511 if (ret) { 2512 error_report("Enabling of KVM dirty ring failed: %s. " 2513 "Suggested minimum value is 1024.", strerror(-ret)); 2514 goto err; 2515 } 2516 2517 s->kvm_dirty_ring_bytes = ring_bytes; 2518 } else { 2519 warn_report("KVM dirty ring not available, using bitmap method"); 2520 s->kvm_dirty_ring_size = 0; 2521 } 2522 } 2523 2524 /* 2525 * KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2 is not needed when dirty ring is 2526 * enabled. More importantly, KVM_DIRTY_LOG_INITIALLY_SET will assume no 2527 * page is wr-protected initially, which is against how kvm dirty ring is 2528 * usage - kvm dirty ring requires all pages are wr-protected at the very 2529 * beginning. Enabling this feature for dirty ring causes data corruption. 2530 * 2531 * TODO: Without KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2 and kvm clear dirty log, 2532 * we may expect a higher stall time when starting the migration. In the 2533 * future we can enable KVM_CLEAR_DIRTY_LOG to work with dirty ring too: 2534 * instead of clearing dirty bit, it can be a way to explicitly wr-protect 2535 * guest pages. 2536 */ 2537 if (!s->kvm_dirty_ring_size) { 2538 dirty_log_manual_caps = 2539 kvm_check_extension(s, KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2); 2540 dirty_log_manual_caps &= (KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE | 2541 KVM_DIRTY_LOG_INITIALLY_SET); 2542 s->manual_dirty_log_protect = dirty_log_manual_caps; 2543 if (dirty_log_manual_caps) { 2544 ret = kvm_vm_enable_cap(s, KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2, 0, 2545 dirty_log_manual_caps); 2546 if (ret) { 2547 warn_report("Trying to enable capability %"PRIu64" of " 2548 "KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2 but failed. " 2549 "Falling back to the legacy mode. ", 2550 dirty_log_manual_caps); 2551 s->manual_dirty_log_protect = 0; 2552 } 2553 } 2554 } 2555 2556 #ifdef KVM_CAP_VCPU_EVENTS 2557 s->vcpu_events = kvm_check_extension(s, KVM_CAP_VCPU_EVENTS); 2558 #endif 2559 2560 s->robust_singlestep = 2561 kvm_check_extension(s, KVM_CAP_X86_ROBUST_SINGLESTEP); 2562 2563 #ifdef KVM_CAP_DEBUGREGS 2564 s->debugregs = kvm_check_extension(s, KVM_CAP_DEBUGREGS); 2565 #endif 2566 2567 s->max_nested_state_len = kvm_check_extension(s, KVM_CAP_NESTED_STATE); 2568 2569 #ifdef KVM_CAP_IRQ_ROUTING 2570 kvm_direct_msi_allowed = (kvm_check_extension(s, KVM_CAP_SIGNAL_MSI) > 0); 2571 #endif 2572 2573 s->intx_set_mask = kvm_check_extension(s, KVM_CAP_PCI_2_3); 2574 2575 s->irq_set_ioctl = KVM_IRQ_LINE; 2576 if (kvm_check_extension(s, KVM_CAP_IRQ_INJECT_STATUS)) { 2577 s->irq_set_ioctl = KVM_IRQ_LINE_STATUS; 2578 } 2579 2580 kvm_readonly_mem_allowed = 2581 (kvm_check_extension(s, KVM_CAP_READONLY_MEM) > 0); 2582 2583 kvm_eventfds_allowed = 2584 (kvm_check_extension(s, KVM_CAP_IOEVENTFD) > 0); 2585 2586 kvm_irqfds_allowed = 2587 (kvm_check_extension(s, KVM_CAP_IRQFD) > 0); 2588 2589 kvm_resamplefds_allowed = 2590 (kvm_check_extension(s, KVM_CAP_IRQFD_RESAMPLE) > 0); 2591 2592 kvm_vm_attributes_allowed = 2593 (kvm_check_extension(s, KVM_CAP_VM_ATTRIBUTES) > 0); 2594 2595 kvm_ioeventfd_any_length_allowed = 2596 (kvm_check_extension(s, KVM_CAP_IOEVENTFD_ANY_LENGTH) > 0); 2597 2598 #ifdef KVM_CAP_SET_GUEST_DEBUG 2599 kvm_has_guest_debug = 2600 (kvm_check_extension(s, KVM_CAP_SET_GUEST_DEBUG) > 0); 2601 #endif 2602 2603 kvm_sstep_flags = 0; 2604 if (kvm_has_guest_debug) { 2605 kvm_sstep_flags = SSTEP_ENABLE; 2606 2607 #if defined KVM_CAP_SET_GUEST_DEBUG2 2608 int guest_debug_flags = 2609 kvm_check_extension(s, KVM_CAP_SET_GUEST_DEBUG2); 2610 2611 if (guest_debug_flags & KVM_GUESTDBG_BLOCKIRQ) { 2612 kvm_sstep_flags |= SSTEP_NOIRQ; 2613 } 2614 #endif 2615 } 2616 2617 kvm_state = s; 2618 2619 ret = kvm_arch_init(ms, s); 2620 if (ret < 0) { 2621 goto err; 2622 } 2623 2624 if (s->kernel_irqchip_split == ON_OFF_AUTO_AUTO) { 2625 s->kernel_irqchip_split = mc->default_kernel_irqchip_split ? ON_OFF_AUTO_ON : ON_OFF_AUTO_OFF; 2626 } 2627 2628 qemu_register_reset(kvm_unpoison_all, NULL); 2629 2630 if (s->kernel_irqchip_allowed) { 2631 kvm_irqchip_create(s); 2632 } 2633 2634 if (kvm_eventfds_allowed) { 2635 s->memory_listener.listener.eventfd_add = kvm_mem_ioeventfd_add; 2636 s->memory_listener.listener.eventfd_del = kvm_mem_ioeventfd_del; 2637 } 2638 s->memory_listener.listener.coalesced_io_add = kvm_coalesce_mmio_region; 2639 s->memory_listener.listener.coalesced_io_del = kvm_uncoalesce_mmio_region; 2640 2641 kvm_memory_listener_register(s, &s->memory_listener, 2642 &address_space_memory, 0, "kvm-memory"); 2643 if (kvm_eventfds_allowed) { 2644 memory_listener_register(&kvm_io_listener, 2645 &address_space_io); 2646 } 2647 memory_listener_register(&kvm_coalesced_pio_listener, 2648 &address_space_io); 2649 2650 s->many_ioeventfds = kvm_check_many_ioeventfds(); 2651 2652 s->sync_mmu = !!kvm_vm_check_extension(kvm_state, KVM_CAP_SYNC_MMU); 2653 if (!s->sync_mmu) { 2654 ret = ram_block_discard_disable(true); 2655 assert(!ret); 2656 } 2657 2658 if (s->kvm_dirty_ring_size) { 2659 ret = kvm_dirty_ring_reaper_init(s); 2660 if (ret) { 2661 goto err; 2662 } 2663 } 2664 2665 if (kvm_check_extension(kvm_state, KVM_CAP_BINARY_STATS_FD)) { 2666 add_stats_callbacks(STATS_PROVIDER_KVM, query_stats_cb, 2667 query_stats_schemas_cb); 2668 } 2669 2670 return 0; 2671 2672 err: 2673 assert(ret < 0); 2674 if (s->vmfd >= 0) { 2675 close(s->vmfd); 2676 } 2677 if (s->fd != -1) { 2678 close(s->fd); 2679 } 2680 g_free(s->memory_listener.slots); 2681 2682 return ret; 2683 } 2684 2685 void kvm_set_sigmask_len(KVMState *s, unsigned int sigmask_len) 2686 { 2687 s->sigmask_len = sigmask_len; 2688 } 2689 2690 static void kvm_handle_io(uint16_t port, MemTxAttrs attrs, void *data, int direction, 2691 int size, uint32_t count) 2692 { 2693 int i; 2694 uint8_t *ptr = data; 2695 2696 for (i = 0; i < count; i++) { 2697 address_space_rw(&address_space_io, port, attrs, 2698 ptr, size, 2699 direction == KVM_EXIT_IO_OUT); 2700 ptr += size; 2701 } 2702 } 2703 2704 static int kvm_handle_internal_error(CPUState *cpu, struct kvm_run *run) 2705 { 2706 fprintf(stderr, "KVM internal error. Suberror: %d\n", 2707 run->internal.suberror); 2708 2709 if (kvm_check_extension(kvm_state, KVM_CAP_INTERNAL_ERROR_DATA)) { 2710 int i; 2711 2712 for (i = 0; i < run->internal.ndata; ++i) { 2713 fprintf(stderr, "extra data[%d]: 0x%016"PRIx64"\n", 2714 i, (uint64_t)run->internal.data[i]); 2715 } 2716 } 2717 if (run->internal.suberror == KVM_INTERNAL_ERROR_EMULATION) { 2718 fprintf(stderr, "emulation failure\n"); 2719 if (!kvm_arch_stop_on_emulation_error(cpu)) { 2720 cpu_dump_state(cpu, stderr, CPU_DUMP_CODE); 2721 return EXCP_INTERRUPT; 2722 } 2723 } 2724 /* FIXME: Should trigger a qmp message to let management know 2725 * something went wrong. 2726 */ 2727 return -1; 2728 } 2729 2730 void kvm_flush_coalesced_mmio_buffer(void) 2731 { 2732 KVMState *s = kvm_state; 2733 2734 if (s->coalesced_flush_in_progress) { 2735 return; 2736 } 2737 2738 s->coalesced_flush_in_progress = true; 2739 2740 if (s->coalesced_mmio_ring) { 2741 struct kvm_coalesced_mmio_ring *ring = s->coalesced_mmio_ring; 2742 while (ring->first != ring->last) { 2743 struct kvm_coalesced_mmio *ent; 2744 2745 ent = &ring->coalesced_mmio[ring->first]; 2746 2747 if (ent->pio == 1) { 2748 address_space_write(&address_space_io, ent->phys_addr, 2749 MEMTXATTRS_UNSPECIFIED, ent->data, 2750 ent->len); 2751 } else { 2752 cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len); 2753 } 2754 smp_wmb(); 2755 ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX; 2756 } 2757 } 2758 2759 s->coalesced_flush_in_progress = false; 2760 } 2761 2762 bool kvm_cpu_check_are_resettable(void) 2763 { 2764 return kvm_arch_cpu_check_are_resettable(); 2765 } 2766 2767 static void do_kvm_cpu_synchronize_state(CPUState *cpu, run_on_cpu_data arg) 2768 { 2769 if (!cpu->vcpu_dirty) { 2770 kvm_arch_get_registers(cpu); 2771 cpu->vcpu_dirty = true; 2772 } 2773 } 2774 2775 void kvm_cpu_synchronize_state(CPUState *cpu) 2776 { 2777 if (!cpu->vcpu_dirty) { 2778 run_on_cpu(cpu, do_kvm_cpu_synchronize_state, RUN_ON_CPU_NULL); 2779 } 2780 } 2781 2782 static void do_kvm_cpu_synchronize_post_reset(CPUState *cpu, run_on_cpu_data arg) 2783 { 2784 kvm_arch_put_registers(cpu, KVM_PUT_RESET_STATE); 2785 cpu->vcpu_dirty = false; 2786 } 2787 2788 void kvm_cpu_synchronize_post_reset(CPUState *cpu) 2789 { 2790 run_on_cpu(cpu, do_kvm_cpu_synchronize_post_reset, RUN_ON_CPU_NULL); 2791 } 2792 2793 static void do_kvm_cpu_synchronize_post_init(CPUState *cpu, run_on_cpu_data arg) 2794 { 2795 kvm_arch_put_registers(cpu, KVM_PUT_FULL_STATE); 2796 cpu->vcpu_dirty = false; 2797 } 2798 2799 void kvm_cpu_synchronize_post_init(CPUState *cpu) 2800 { 2801 run_on_cpu(cpu, do_kvm_cpu_synchronize_post_init, RUN_ON_CPU_NULL); 2802 } 2803 2804 static void do_kvm_cpu_synchronize_pre_loadvm(CPUState *cpu, run_on_cpu_data arg) 2805 { 2806 cpu->vcpu_dirty = true; 2807 } 2808 2809 void kvm_cpu_synchronize_pre_loadvm(CPUState *cpu) 2810 { 2811 run_on_cpu(cpu, do_kvm_cpu_synchronize_pre_loadvm, RUN_ON_CPU_NULL); 2812 } 2813 2814 #ifdef KVM_HAVE_MCE_INJECTION 2815 static __thread void *pending_sigbus_addr; 2816 static __thread int pending_sigbus_code; 2817 static __thread bool have_sigbus_pending; 2818 #endif 2819 2820 static void kvm_cpu_kick(CPUState *cpu) 2821 { 2822 qatomic_set(&cpu->kvm_run->immediate_exit, 1); 2823 } 2824 2825 static void kvm_cpu_kick_self(void) 2826 { 2827 if (kvm_immediate_exit) { 2828 kvm_cpu_kick(current_cpu); 2829 } else { 2830 qemu_cpu_kick_self(); 2831 } 2832 } 2833 2834 static void kvm_eat_signals(CPUState *cpu) 2835 { 2836 struct timespec ts = { 0, 0 }; 2837 siginfo_t siginfo; 2838 sigset_t waitset; 2839 sigset_t chkset; 2840 int r; 2841 2842 if (kvm_immediate_exit) { 2843 qatomic_set(&cpu->kvm_run->immediate_exit, 0); 2844 /* Write kvm_run->immediate_exit before the cpu->exit_request 2845 * write in kvm_cpu_exec. 2846 */ 2847 smp_wmb(); 2848 return; 2849 } 2850 2851 sigemptyset(&waitset); 2852 sigaddset(&waitset, SIG_IPI); 2853 2854 do { 2855 r = sigtimedwait(&waitset, &siginfo, &ts); 2856 if (r == -1 && !(errno == EAGAIN || errno == EINTR)) { 2857 perror("sigtimedwait"); 2858 exit(1); 2859 } 2860 2861 r = sigpending(&chkset); 2862 if (r == -1) { 2863 perror("sigpending"); 2864 exit(1); 2865 } 2866 } while (sigismember(&chkset, SIG_IPI)); 2867 } 2868 2869 int kvm_cpu_exec(CPUState *cpu) 2870 { 2871 struct kvm_run *run = cpu->kvm_run; 2872 int ret, run_ret; 2873 2874 DPRINTF("kvm_cpu_exec()\n"); 2875 2876 if (kvm_arch_process_async_events(cpu)) { 2877 qatomic_set(&cpu->exit_request, 0); 2878 return EXCP_HLT; 2879 } 2880 2881 qemu_mutex_unlock_iothread(); 2882 cpu_exec_start(cpu); 2883 2884 do { 2885 MemTxAttrs attrs; 2886 2887 if (cpu->vcpu_dirty) { 2888 kvm_arch_put_registers(cpu, KVM_PUT_RUNTIME_STATE); 2889 cpu->vcpu_dirty = false; 2890 } 2891 2892 kvm_arch_pre_run(cpu, run); 2893 if (qatomic_read(&cpu->exit_request)) { 2894 DPRINTF("interrupt exit requested\n"); 2895 /* 2896 * KVM requires us to reenter the kernel after IO exits to complete 2897 * instruction emulation. This self-signal will ensure that we 2898 * leave ASAP again. 2899 */ 2900 kvm_cpu_kick_self(); 2901 } 2902 2903 /* Read cpu->exit_request before KVM_RUN reads run->immediate_exit. 2904 * Matching barrier in kvm_eat_signals. 2905 */ 2906 smp_rmb(); 2907 2908 run_ret = kvm_vcpu_ioctl(cpu, KVM_RUN, 0); 2909 2910 attrs = kvm_arch_post_run(cpu, run); 2911 2912 #ifdef KVM_HAVE_MCE_INJECTION 2913 if (unlikely(have_sigbus_pending)) { 2914 qemu_mutex_lock_iothread(); 2915 kvm_arch_on_sigbus_vcpu(cpu, pending_sigbus_code, 2916 pending_sigbus_addr); 2917 have_sigbus_pending = false; 2918 qemu_mutex_unlock_iothread(); 2919 } 2920 #endif 2921 2922 if (run_ret < 0) { 2923 if (run_ret == -EINTR || run_ret == -EAGAIN) { 2924 DPRINTF("io window exit\n"); 2925 kvm_eat_signals(cpu); 2926 ret = EXCP_INTERRUPT; 2927 break; 2928 } 2929 fprintf(stderr, "error: kvm run failed %s\n", 2930 strerror(-run_ret)); 2931 #ifdef TARGET_PPC 2932 if (run_ret == -EBUSY) { 2933 fprintf(stderr, 2934 "This is probably because your SMT is enabled.\n" 2935 "VCPU can only run on primary threads with all " 2936 "secondary threads offline.\n"); 2937 } 2938 #endif 2939 ret = -1; 2940 break; 2941 } 2942 2943 trace_kvm_run_exit(cpu->cpu_index, run->exit_reason); 2944 switch (run->exit_reason) { 2945 case KVM_EXIT_IO: 2946 DPRINTF("handle_io\n"); 2947 /* Called outside BQL */ 2948 kvm_handle_io(run->io.port, attrs, 2949 (uint8_t *)run + run->io.data_offset, 2950 run->io.direction, 2951 run->io.size, 2952 run->io.count); 2953 ret = 0; 2954 break; 2955 case KVM_EXIT_MMIO: 2956 DPRINTF("handle_mmio\n"); 2957 /* Called outside BQL */ 2958 address_space_rw(&address_space_memory, 2959 run->mmio.phys_addr, attrs, 2960 run->mmio.data, 2961 run->mmio.len, 2962 run->mmio.is_write); 2963 ret = 0; 2964 break; 2965 case KVM_EXIT_IRQ_WINDOW_OPEN: 2966 DPRINTF("irq_window_open\n"); 2967 ret = EXCP_INTERRUPT; 2968 break; 2969 case KVM_EXIT_SHUTDOWN: 2970 DPRINTF("shutdown\n"); 2971 qemu_system_reset_request(SHUTDOWN_CAUSE_GUEST_RESET); 2972 ret = EXCP_INTERRUPT; 2973 break; 2974 case KVM_EXIT_UNKNOWN: 2975 fprintf(stderr, "KVM: unknown exit, hardware reason %" PRIx64 "\n", 2976 (uint64_t)run->hw.hardware_exit_reason); 2977 ret = -1; 2978 break; 2979 case KVM_EXIT_INTERNAL_ERROR: 2980 ret = kvm_handle_internal_error(cpu, run); 2981 break; 2982 case KVM_EXIT_DIRTY_RING_FULL: 2983 /* 2984 * We shouldn't continue if the dirty ring of this vcpu is 2985 * still full. Got kicked by KVM_RESET_DIRTY_RINGS. 2986 */ 2987 trace_kvm_dirty_ring_full(cpu->cpu_index); 2988 qemu_mutex_lock_iothread(); 2989 /* 2990 * We throttle vCPU by making it sleep once it exit from kernel 2991 * due to dirty ring full. In the dirtylimit scenario, reaping 2992 * all vCPUs after a single vCPU dirty ring get full result in 2993 * the miss of sleep, so just reap the ring-fulled vCPU. 2994 */ 2995 if (dirtylimit_in_service()) { 2996 kvm_dirty_ring_reap(kvm_state, cpu); 2997 } else { 2998 kvm_dirty_ring_reap(kvm_state, NULL); 2999 } 3000 qemu_mutex_unlock_iothread(); 3001 dirtylimit_vcpu_execute(cpu); 3002 ret = 0; 3003 break; 3004 case KVM_EXIT_SYSTEM_EVENT: 3005 switch (run->system_event.type) { 3006 case KVM_SYSTEM_EVENT_SHUTDOWN: 3007 qemu_system_shutdown_request(SHUTDOWN_CAUSE_GUEST_SHUTDOWN); 3008 ret = EXCP_INTERRUPT; 3009 break; 3010 case KVM_SYSTEM_EVENT_RESET: 3011 qemu_system_reset_request(SHUTDOWN_CAUSE_GUEST_RESET); 3012 ret = EXCP_INTERRUPT; 3013 break; 3014 case KVM_SYSTEM_EVENT_CRASH: 3015 kvm_cpu_synchronize_state(cpu); 3016 qemu_mutex_lock_iothread(); 3017 qemu_system_guest_panicked(cpu_get_crash_info(cpu)); 3018 qemu_mutex_unlock_iothread(); 3019 ret = 0; 3020 break; 3021 default: 3022 DPRINTF("kvm_arch_handle_exit\n"); 3023 ret = kvm_arch_handle_exit(cpu, run); 3024 break; 3025 } 3026 break; 3027 default: 3028 DPRINTF("kvm_arch_handle_exit\n"); 3029 ret = kvm_arch_handle_exit(cpu, run); 3030 break; 3031 } 3032 } while (ret == 0); 3033 3034 cpu_exec_end(cpu); 3035 qemu_mutex_lock_iothread(); 3036 3037 if (ret < 0) { 3038 cpu_dump_state(cpu, stderr, CPU_DUMP_CODE); 3039 vm_stop(RUN_STATE_INTERNAL_ERROR); 3040 } 3041 3042 qatomic_set(&cpu->exit_request, 0); 3043 return ret; 3044 } 3045 3046 int kvm_ioctl(KVMState *s, int type, ...) 3047 { 3048 int ret; 3049 void *arg; 3050 va_list ap; 3051 3052 va_start(ap, type); 3053 arg = va_arg(ap, void *); 3054 va_end(ap); 3055 3056 trace_kvm_ioctl(type, arg); 3057 ret = ioctl(s->fd, type, arg); 3058 if (ret == -1) { 3059 ret = -errno; 3060 } 3061 return ret; 3062 } 3063 3064 int kvm_vm_ioctl(KVMState *s, int type, ...) 3065 { 3066 int ret; 3067 void *arg; 3068 va_list ap; 3069 3070 va_start(ap, type); 3071 arg = va_arg(ap, void *); 3072 va_end(ap); 3073 3074 trace_kvm_vm_ioctl(type, arg); 3075 ret = ioctl(s->vmfd, type, arg); 3076 if (ret == -1) { 3077 ret = -errno; 3078 } 3079 return ret; 3080 } 3081 3082 int kvm_vcpu_ioctl(CPUState *cpu, int type, ...) 3083 { 3084 int ret; 3085 void *arg; 3086 va_list ap; 3087 3088 va_start(ap, type); 3089 arg = va_arg(ap, void *); 3090 va_end(ap); 3091 3092 trace_kvm_vcpu_ioctl(cpu->cpu_index, type, arg); 3093 ret = ioctl(cpu->kvm_fd, type, arg); 3094 if (ret == -1) { 3095 ret = -errno; 3096 } 3097 return ret; 3098 } 3099 3100 int kvm_device_ioctl(int fd, int type, ...) 3101 { 3102 int ret; 3103 void *arg; 3104 va_list ap; 3105 3106 va_start(ap, type); 3107 arg = va_arg(ap, void *); 3108 va_end(ap); 3109 3110 trace_kvm_device_ioctl(fd, type, arg); 3111 ret = ioctl(fd, type, arg); 3112 if (ret == -1) { 3113 ret = -errno; 3114 } 3115 return ret; 3116 } 3117 3118 int kvm_vm_check_attr(KVMState *s, uint32_t group, uint64_t attr) 3119 { 3120 int ret; 3121 struct kvm_device_attr attribute = { 3122 .group = group, 3123 .attr = attr, 3124 }; 3125 3126 if (!kvm_vm_attributes_allowed) { 3127 return 0; 3128 } 3129 3130 ret = kvm_vm_ioctl(s, KVM_HAS_DEVICE_ATTR, &attribute); 3131 /* kvm returns 0 on success for HAS_DEVICE_ATTR */ 3132 return ret ? 0 : 1; 3133 } 3134 3135 int kvm_device_check_attr(int dev_fd, uint32_t group, uint64_t attr) 3136 { 3137 struct kvm_device_attr attribute = { 3138 .group = group, 3139 .attr = attr, 3140 .flags = 0, 3141 }; 3142 3143 return kvm_device_ioctl(dev_fd, KVM_HAS_DEVICE_ATTR, &attribute) ? 0 : 1; 3144 } 3145 3146 int kvm_device_access(int fd, int group, uint64_t attr, 3147 void *val, bool write, Error **errp) 3148 { 3149 struct kvm_device_attr kvmattr; 3150 int err; 3151 3152 kvmattr.flags = 0; 3153 kvmattr.group = group; 3154 kvmattr.attr = attr; 3155 kvmattr.addr = (uintptr_t)val; 3156 3157 err = kvm_device_ioctl(fd, 3158 write ? KVM_SET_DEVICE_ATTR : KVM_GET_DEVICE_ATTR, 3159 &kvmattr); 3160 if (err < 0) { 3161 error_setg_errno(errp, -err, 3162 "KVM_%s_DEVICE_ATTR failed: Group %d " 3163 "attr 0x%016" PRIx64, 3164 write ? "SET" : "GET", group, attr); 3165 } 3166 return err; 3167 } 3168 3169 bool kvm_has_sync_mmu(void) 3170 { 3171 return kvm_state->sync_mmu; 3172 } 3173 3174 int kvm_has_vcpu_events(void) 3175 { 3176 return kvm_state->vcpu_events; 3177 } 3178 3179 int kvm_has_robust_singlestep(void) 3180 { 3181 return kvm_state->robust_singlestep; 3182 } 3183 3184 int kvm_has_debugregs(void) 3185 { 3186 return kvm_state->debugregs; 3187 } 3188 3189 int kvm_max_nested_state_length(void) 3190 { 3191 return kvm_state->max_nested_state_len; 3192 } 3193 3194 int kvm_has_many_ioeventfds(void) 3195 { 3196 if (!kvm_enabled()) { 3197 return 0; 3198 } 3199 return kvm_state->many_ioeventfds; 3200 } 3201 3202 int kvm_has_gsi_routing(void) 3203 { 3204 #ifdef KVM_CAP_IRQ_ROUTING 3205 return kvm_check_extension(kvm_state, KVM_CAP_IRQ_ROUTING); 3206 #else 3207 return false; 3208 #endif 3209 } 3210 3211 int kvm_has_intx_set_mask(void) 3212 { 3213 return kvm_state->intx_set_mask; 3214 } 3215 3216 bool kvm_arm_supports_user_irq(void) 3217 { 3218 return kvm_check_extension(kvm_state, KVM_CAP_ARM_USER_IRQ); 3219 } 3220 3221 #ifdef KVM_CAP_SET_GUEST_DEBUG 3222 struct kvm_sw_breakpoint *kvm_find_sw_breakpoint(CPUState *cpu, 3223 target_ulong pc) 3224 { 3225 struct kvm_sw_breakpoint *bp; 3226 3227 QTAILQ_FOREACH(bp, &cpu->kvm_state->kvm_sw_breakpoints, entry) { 3228 if (bp->pc == pc) { 3229 return bp; 3230 } 3231 } 3232 return NULL; 3233 } 3234 3235 int kvm_sw_breakpoints_active(CPUState *cpu) 3236 { 3237 return !QTAILQ_EMPTY(&cpu->kvm_state->kvm_sw_breakpoints); 3238 } 3239 3240 struct kvm_set_guest_debug_data { 3241 struct kvm_guest_debug dbg; 3242 int err; 3243 }; 3244 3245 static void kvm_invoke_set_guest_debug(CPUState *cpu, run_on_cpu_data data) 3246 { 3247 struct kvm_set_guest_debug_data *dbg_data = 3248 (struct kvm_set_guest_debug_data *) data.host_ptr; 3249 3250 dbg_data->err = kvm_vcpu_ioctl(cpu, KVM_SET_GUEST_DEBUG, 3251 &dbg_data->dbg); 3252 } 3253 3254 int kvm_update_guest_debug(CPUState *cpu, unsigned long reinject_trap) 3255 { 3256 struct kvm_set_guest_debug_data data; 3257 3258 data.dbg.control = reinject_trap; 3259 3260 if (cpu->singlestep_enabled) { 3261 data.dbg.control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_SINGLESTEP; 3262 3263 if (cpu->singlestep_enabled & SSTEP_NOIRQ) { 3264 data.dbg.control |= KVM_GUESTDBG_BLOCKIRQ; 3265 } 3266 } 3267 kvm_arch_update_guest_debug(cpu, &data.dbg); 3268 3269 run_on_cpu(cpu, kvm_invoke_set_guest_debug, 3270 RUN_ON_CPU_HOST_PTR(&data)); 3271 return data.err; 3272 } 3273 3274 int kvm_insert_breakpoint(CPUState *cpu, target_ulong addr, 3275 target_ulong len, int type) 3276 { 3277 struct kvm_sw_breakpoint *bp; 3278 int err; 3279 3280 if (type == GDB_BREAKPOINT_SW) { 3281 bp = kvm_find_sw_breakpoint(cpu, addr); 3282 if (bp) { 3283 bp->use_count++; 3284 return 0; 3285 } 3286 3287 bp = g_new(struct kvm_sw_breakpoint, 1); 3288 bp->pc = addr; 3289 bp->use_count = 1; 3290 err = kvm_arch_insert_sw_breakpoint(cpu, bp); 3291 if (err) { 3292 g_free(bp); 3293 return err; 3294 } 3295 3296 QTAILQ_INSERT_HEAD(&cpu->kvm_state->kvm_sw_breakpoints, bp, entry); 3297 } else { 3298 err = kvm_arch_insert_hw_breakpoint(addr, len, type); 3299 if (err) { 3300 return err; 3301 } 3302 } 3303 3304 CPU_FOREACH(cpu) { 3305 err = kvm_update_guest_debug(cpu, 0); 3306 if (err) { 3307 return err; 3308 } 3309 } 3310 return 0; 3311 } 3312 3313 int kvm_remove_breakpoint(CPUState *cpu, target_ulong addr, 3314 target_ulong len, int type) 3315 { 3316 struct kvm_sw_breakpoint *bp; 3317 int err; 3318 3319 if (type == GDB_BREAKPOINT_SW) { 3320 bp = kvm_find_sw_breakpoint(cpu, addr); 3321 if (!bp) { 3322 return -ENOENT; 3323 } 3324 3325 if (bp->use_count > 1) { 3326 bp->use_count--; 3327 return 0; 3328 } 3329 3330 err = kvm_arch_remove_sw_breakpoint(cpu, bp); 3331 if (err) { 3332 return err; 3333 } 3334 3335 QTAILQ_REMOVE(&cpu->kvm_state->kvm_sw_breakpoints, bp, entry); 3336 g_free(bp); 3337 } else { 3338 err = kvm_arch_remove_hw_breakpoint(addr, len, type); 3339 if (err) { 3340 return err; 3341 } 3342 } 3343 3344 CPU_FOREACH(cpu) { 3345 err = kvm_update_guest_debug(cpu, 0); 3346 if (err) { 3347 return err; 3348 } 3349 } 3350 return 0; 3351 } 3352 3353 void kvm_remove_all_breakpoints(CPUState *cpu) 3354 { 3355 struct kvm_sw_breakpoint *bp, *next; 3356 KVMState *s = cpu->kvm_state; 3357 CPUState *tmpcpu; 3358 3359 QTAILQ_FOREACH_SAFE(bp, &s->kvm_sw_breakpoints, entry, next) { 3360 if (kvm_arch_remove_sw_breakpoint(cpu, bp) != 0) { 3361 /* Try harder to find a CPU that currently sees the breakpoint. */ 3362 CPU_FOREACH(tmpcpu) { 3363 if (kvm_arch_remove_sw_breakpoint(tmpcpu, bp) == 0) { 3364 break; 3365 } 3366 } 3367 } 3368 QTAILQ_REMOVE(&s->kvm_sw_breakpoints, bp, entry); 3369 g_free(bp); 3370 } 3371 kvm_arch_remove_all_hw_breakpoints(); 3372 3373 CPU_FOREACH(cpu) { 3374 kvm_update_guest_debug(cpu, 0); 3375 } 3376 } 3377 3378 #else /* !KVM_CAP_SET_GUEST_DEBUG */ 3379 3380 int kvm_update_guest_debug(CPUState *cpu, unsigned long reinject_trap) 3381 { 3382 return -EINVAL; 3383 } 3384 3385 int kvm_insert_breakpoint(CPUState *cpu, target_ulong addr, 3386 target_ulong len, int type) 3387 { 3388 return -EINVAL; 3389 } 3390 3391 int kvm_remove_breakpoint(CPUState *cpu, target_ulong addr, 3392 target_ulong len, int type) 3393 { 3394 return -EINVAL; 3395 } 3396 3397 void kvm_remove_all_breakpoints(CPUState *cpu) 3398 { 3399 } 3400 #endif /* !KVM_CAP_SET_GUEST_DEBUG */ 3401 3402 static int kvm_set_signal_mask(CPUState *cpu, const sigset_t *sigset) 3403 { 3404 KVMState *s = kvm_state; 3405 struct kvm_signal_mask *sigmask; 3406 int r; 3407 3408 sigmask = g_malloc(sizeof(*sigmask) + sizeof(*sigset)); 3409 3410 sigmask->len = s->sigmask_len; 3411 memcpy(sigmask->sigset, sigset, sizeof(*sigset)); 3412 r = kvm_vcpu_ioctl(cpu, KVM_SET_SIGNAL_MASK, sigmask); 3413 g_free(sigmask); 3414 3415 return r; 3416 } 3417 3418 static void kvm_ipi_signal(int sig) 3419 { 3420 if (current_cpu) { 3421 assert(kvm_immediate_exit); 3422 kvm_cpu_kick(current_cpu); 3423 } 3424 } 3425 3426 void kvm_init_cpu_signals(CPUState *cpu) 3427 { 3428 int r; 3429 sigset_t set; 3430 struct sigaction sigact; 3431 3432 memset(&sigact, 0, sizeof(sigact)); 3433 sigact.sa_handler = kvm_ipi_signal; 3434 sigaction(SIG_IPI, &sigact, NULL); 3435 3436 pthread_sigmask(SIG_BLOCK, NULL, &set); 3437 #if defined KVM_HAVE_MCE_INJECTION 3438 sigdelset(&set, SIGBUS); 3439 pthread_sigmask(SIG_SETMASK, &set, NULL); 3440 #endif 3441 sigdelset(&set, SIG_IPI); 3442 if (kvm_immediate_exit) { 3443 r = pthread_sigmask(SIG_SETMASK, &set, NULL); 3444 } else { 3445 r = kvm_set_signal_mask(cpu, &set); 3446 } 3447 if (r) { 3448 fprintf(stderr, "kvm_set_signal_mask: %s\n", strerror(-r)); 3449 exit(1); 3450 } 3451 } 3452 3453 /* Called asynchronously in VCPU thread. */ 3454 int kvm_on_sigbus_vcpu(CPUState *cpu, int code, void *addr) 3455 { 3456 #ifdef KVM_HAVE_MCE_INJECTION 3457 if (have_sigbus_pending) { 3458 return 1; 3459 } 3460 have_sigbus_pending = true; 3461 pending_sigbus_addr = addr; 3462 pending_sigbus_code = code; 3463 qatomic_set(&cpu->exit_request, 1); 3464 return 0; 3465 #else 3466 return 1; 3467 #endif 3468 } 3469 3470 /* Called synchronously (via signalfd) in main thread. */ 3471 int kvm_on_sigbus(int code, void *addr) 3472 { 3473 #ifdef KVM_HAVE_MCE_INJECTION 3474 /* Action required MCE kills the process if SIGBUS is blocked. Because 3475 * that's what happens in the I/O thread, where we handle MCE via signalfd, 3476 * we can only get action optional here. 3477 */ 3478 assert(code != BUS_MCEERR_AR); 3479 kvm_arch_on_sigbus_vcpu(first_cpu, code, addr); 3480 return 0; 3481 #else 3482 return 1; 3483 #endif 3484 } 3485 3486 int kvm_create_device(KVMState *s, uint64_t type, bool test) 3487 { 3488 int ret; 3489 struct kvm_create_device create_dev; 3490 3491 create_dev.type = type; 3492 create_dev.fd = -1; 3493 create_dev.flags = test ? KVM_CREATE_DEVICE_TEST : 0; 3494 3495 if (!kvm_check_extension(s, KVM_CAP_DEVICE_CTRL)) { 3496 return -ENOTSUP; 3497 } 3498 3499 ret = kvm_vm_ioctl(s, KVM_CREATE_DEVICE, &create_dev); 3500 if (ret) { 3501 return ret; 3502 } 3503 3504 return test ? 0 : create_dev.fd; 3505 } 3506 3507 bool kvm_device_supported(int vmfd, uint64_t type) 3508 { 3509 struct kvm_create_device create_dev = { 3510 .type = type, 3511 .fd = -1, 3512 .flags = KVM_CREATE_DEVICE_TEST, 3513 }; 3514 3515 if (ioctl(vmfd, KVM_CHECK_EXTENSION, KVM_CAP_DEVICE_CTRL) <= 0) { 3516 return false; 3517 } 3518 3519 return (ioctl(vmfd, KVM_CREATE_DEVICE, &create_dev) >= 0); 3520 } 3521 3522 int kvm_set_one_reg(CPUState *cs, uint64_t id, void *source) 3523 { 3524 struct kvm_one_reg reg; 3525 int r; 3526 3527 reg.id = id; 3528 reg.addr = (uintptr_t) source; 3529 r = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, ®); 3530 if (r) { 3531 trace_kvm_failed_reg_set(id, strerror(-r)); 3532 } 3533 return r; 3534 } 3535 3536 int kvm_get_one_reg(CPUState *cs, uint64_t id, void *target) 3537 { 3538 struct kvm_one_reg reg; 3539 int r; 3540 3541 reg.id = id; 3542 reg.addr = (uintptr_t) target; 3543 r = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, ®); 3544 if (r) { 3545 trace_kvm_failed_reg_get(id, strerror(-r)); 3546 } 3547 return r; 3548 } 3549 3550 static bool kvm_accel_has_memory(MachineState *ms, AddressSpace *as, 3551 hwaddr start_addr, hwaddr size) 3552 { 3553 KVMState *kvm = KVM_STATE(ms->accelerator); 3554 int i; 3555 3556 for (i = 0; i < kvm->nr_as; ++i) { 3557 if (kvm->as[i].as == as && kvm->as[i].ml) { 3558 size = MIN(kvm_max_slot_size, size); 3559 return NULL != kvm_lookup_matching_slot(kvm->as[i].ml, 3560 start_addr, size); 3561 } 3562 } 3563 3564 return false; 3565 } 3566 3567 static void kvm_get_kvm_shadow_mem(Object *obj, Visitor *v, 3568 const char *name, void *opaque, 3569 Error **errp) 3570 { 3571 KVMState *s = KVM_STATE(obj); 3572 int64_t value = s->kvm_shadow_mem; 3573 3574 visit_type_int(v, name, &value, errp); 3575 } 3576 3577 static void kvm_set_kvm_shadow_mem(Object *obj, Visitor *v, 3578 const char *name, void *opaque, 3579 Error **errp) 3580 { 3581 KVMState *s = KVM_STATE(obj); 3582 int64_t value; 3583 3584 if (s->fd != -1) { 3585 error_setg(errp, "Cannot set properties after the accelerator has been initialized"); 3586 return; 3587 } 3588 3589 if (!visit_type_int(v, name, &value, errp)) { 3590 return; 3591 } 3592 3593 s->kvm_shadow_mem = value; 3594 } 3595 3596 static void kvm_set_kernel_irqchip(Object *obj, Visitor *v, 3597 const char *name, void *opaque, 3598 Error **errp) 3599 { 3600 KVMState *s = KVM_STATE(obj); 3601 OnOffSplit mode; 3602 3603 if (s->fd != -1) { 3604 error_setg(errp, "Cannot set properties after the accelerator has been initialized"); 3605 return; 3606 } 3607 3608 if (!visit_type_OnOffSplit(v, name, &mode, errp)) { 3609 return; 3610 } 3611 switch (mode) { 3612 case ON_OFF_SPLIT_ON: 3613 s->kernel_irqchip_allowed = true; 3614 s->kernel_irqchip_required = true; 3615 s->kernel_irqchip_split = ON_OFF_AUTO_OFF; 3616 break; 3617 case ON_OFF_SPLIT_OFF: 3618 s->kernel_irqchip_allowed = false; 3619 s->kernel_irqchip_required = false; 3620 s->kernel_irqchip_split = ON_OFF_AUTO_OFF; 3621 break; 3622 case ON_OFF_SPLIT_SPLIT: 3623 s->kernel_irqchip_allowed = true; 3624 s->kernel_irqchip_required = true; 3625 s->kernel_irqchip_split = ON_OFF_AUTO_ON; 3626 break; 3627 default: 3628 /* The value was checked in visit_type_OnOffSplit() above. If 3629 * we get here, then something is wrong in QEMU. 3630 */ 3631 abort(); 3632 } 3633 } 3634 3635 bool kvm_kernel_irqchip_allowed(void) 3636 { 3637 return kvm_state->kernel_irqchip_allowed; 3638 } 3639 3640 bool kvm_kernel_irqchip_required(void) 3641 { 3642 return kvm_state->kernel_irqchip_required; 3643 } 3644 3645 bool kvm_kernel_irqchip_split(void) 3646 { 3647 return kvm_state->kernel_irqchip_split == ON_OFF_AUTO_ON; 3648 } 3649 3650 static void kvm_get_dirty_ring_size(Object *obj, Visitor *v, 3651 const char *name, void *opaque, 3652 Error **errp) 3653 { 3654 KVMState *s = KVM_STATE(obj); 3655 uint32_t value = s->kvm_dirty_ring_size; 3656 3657 visit_type_uint32(v, name, &value, errp); 3658 } 3659 3660 static void kvm_set_dirty_ring_size(Object *obj, Visitor *v, 3661 const char *name, void *opaque, 3662 Error **errp) 3663 { 3664 KVMState *s = KVM_STATE(obj); 3665 Error *error = NULL; 3666 uint32_t value; 3667 3668 if (s->fd != -1) { 3669 error_setg(errp, "Cannot set properties after the accelerator has been initialized"); 3670 return; 3671 } 3672 3673 visit_type_uint32(v, name, &value, &error); 3674 if (error) { 3675 error_propagate(errp, error); 3676 return; 3677 } 3678 if (value & (value - 1)) { 3679 error_setg(errp, "dirty-ring-size must be a power of two."); 3680 return; 3681 } 3682 3683 s->kvm_dirty_ring_size = value; 3684 } 3685 3686 static void kvm_accel_instance_init(Object *obj) 3687 { 3688 KVMState *s = KVM_STATE(obj); 3689 3690 s->fd = -1; 3691 s->vmfd = -1; 3692 s->kvm_shadow_mem = -1; 3693 s->kernel_irqchip_allowed = true; 3694 s->kernel_irqchip_split = ON_OFF_AUTO_AUTO; 3695 /* KVM dirty ring is by default off */ 3696 s->kvm_dirty_ring_size = 0; 3697 } 3698 3699 static void kvm_accel_class_init(ObjectClass *oc, void *data) 3700 { 3701 AccelClass *ac = ACCEL_CLASS(oc); 3702 ac->name = "KVM"; 3703 ac->init_machine = kvm_init; 3704 ac->has_memory = kvm_accel_has_memory; 3705 ac->allowed = &kvm_allowed; 3706 3707 object_class_property_add(oc, "kernel-irqchip", "on|off|split", 3708 NULL, kvm_set_kernel_irqchip, 3709 NULL, NULL); 3710 object_class_property_set_description(oc, "kernel-irqchip", 3711 "Configure KVM in-kernel irqchip"); 3712 3713 object_class_property_add(oc, "kvm-shadow-mem", "int", 3714 kvm_get_kvm_shadow_mem, kvm_set_kvm_shadow_mem, 3715 NULL, NULL); 3716 object_class_property_set_description(oc, "kvm-shadow-mem", 3717 "KVM shadow MMU size"); 3718 3719 object_class_property_add(oc, "dirty-ring-size", "uint32", 3720 kvm_get_dirty_ring_size, kvm_set_dirty_ring_size, 3721 NULL, NULL); 3722 object_class_property_set_description(oc, "dirty-ring-size", 3723 "Size of KVM dirty page ring buffer (default: 0, i.e. use bitmap)"); 3724 } 3725 3726 static const TypeInfo kvm_accel_type = { 3727 .name = TYPE_KVM_ACCEL, 3728 .parent = TYPE_ACCEL, 3729 .instance_init = kvm_accel_instance_init, 3730 .class_init = kvm_accel_class_init, 3731 .instance_size = sizeof(KVMState), 3732 }; 3733 3734 static void kvm_type_init(void) 3735 { 3736 type_register_static(&kvm_accel_type); 3737 } 3738 3739 type_init(kvm_type_init); 3740 3741 typedef struct StatsArgs { 3742 union StatsResultsType { 3743 StatsResultList **stats; 3744 StatsSchemaList **schema; 3745 } result; 3746 strList *names; 3747 Error **errp; 3748 } StatsArgs; 3749 3750 static StatsList *add_kvmstat_entry(struct kvm_stats_desc *pdesc, 3751 uint64_t *stats_data, 3752 StatsList *stats_list, 3753 Error **errp) 3754 { 3755 3756 Stats *stats; 3757 uint64List *val_list = NULL; 3758 3759 /* Only add stats that we understand. */ 3760 switch (pdesc->flags & KVM_STATS_TYPE_MASK) { 3761 case KVM_STATS_TYPE_CUMULATIVE: 3762 case KVM_STATS_TYPE_INSTANT: 3763 case KVM_STATS_TYPE_PEAK: 3764 case KVM_STATS_TYPE_LINEAR_HIST: 3765 case KVM_STATS_TYPE_LOG_HIST: 3766 break; 3767 default: 3768 return stats_list; 3769 } 3770 3771 switch (pdesc->flags & KVM_STATS_UNIT_MASK) { 3772 case KVM_STATS_UNIT_NONE: 3773 case KVM_STATS_UNIT_BYTES: 3774 case KVM_STATS_UNIT_CYCLES: 3775 case KVM_STATS_UNIT_SECONDS: 3776 case KVM_STATS_UNIT_BOOLEAN: 3777 break; 3778 default: 3779 return stats_list; 3780 } 3781 3782 switch (pdesc->flags & KVM_STATS_BASE_MASK) { 3783 case KVM_STATS_BASE_POW10: 3784 case KVM_STATS_BASE_POW2: 3785 break; 3786 default: 3787 return stats_list; 3788 } 3789 3790 /* Alloc and populate data list */ 3791 stats = g_new0(Stats, 1); 3792 stats->name = g_strdup(pdesc->name); 3793 stats->value = g_new0(StatsValue, 1);; 3794 3795 if ((pdesc->flags & KVM_STATS_UNIT_MASK) == KVM_STATS_UNIT_BOOLEAN) { 3796 stats->value->u.boolean = *stats_data; 3797 stats->value->type = QTYPE_QBOOL; 3798 } else if (pdesc->size == 1) { 3799 stats->value->u.scalar = *stats_data; 3800 stats->value->type = QTYPE_QNUM; 3801 } else { 3802 int i; 3803 for (i = 0; i < pdesc->size; i++) { 3804 QAPI_LIST_PREPEND(val_list, stats_data[i]); 3805 } 3806 stats->value->u.list = val_list; 3807 stats->value->type = QTYPE_QLIST; 3808 } 3809 3810 QAPI_LIST_PREPEND(stats_list, stats); 3811 return stats_list; 3812 } 3813 3814 static StatsSchemaValueList *add_kvmschema_entry(struct kvm_stats_desc *pdesc, 3815 StatsSchemaValueList *list, 3816 Error **errp) 3817 { 3818 StatsSchemaValueList *schema_entry = g_new0(StatsSchemaValueList, 1); 3819 schema_entry->value = g_new0(StatsSchemaValue, 1); 3820 3821 switch (pdesc->flags & KVM_STATS_TYPE_MASK) { 3822 case KVM_STATS_TYPE_CUMULATIVE: 3823 schema_entry->value->type = STATS_TYPE_CUMULATIVE; 3824 break; 3825 case KVM_STATS_TYPE_INSTANT: 3826 schema_entry->value->type = STATS_TYPE_INSTANT; 3827 break; 3828 case KVM_STATS_TYPE_PEAK: 3829 schema_entry->value->type = STATS_TYPE_PEAK; 3830 break; 3831 case KVM_STATS_TYPE_LINEAR_HIST: 3832 schema_entry->value->type = STATS_TYPE_LINEAR_HISTOGRAM; 3833 schema_entry->value->bucket_size = pdesc->bucket_size; 3834 schema_entry->value->has_bucket_size = true; 3835 break; 3836 case KVM_STATS_TYPE_LOG_HIST: 3837 schema_entry->value->type = STATS_TYPE_LOG2_HISTOGRAM; 3838 break; 3839 default: 3840 goto exit; 3841 } 3842 3843 switch (pdesc->flags & KVM_STATS_UNIT_MASK) { 3844 case KVM_STATS_UNIT_NONE: 3845 break; 3846 case KVM_STATS_UNIT_BOOLEAN: 3847 schema_entry->value->has_unit = true; 3848 schema_entry->value->unit = STATS_UNIT_BOOLEAN; 3849 break; 3850 case KVM_STATS_UNIT_BYTES: 3851 schema_entry->value->has_unit = true; 3852 schema_entry->value->unit = STATS_UNIT_BYTES; 3853 break; 3854 case KVM_STATS_UNIT_CYCLES: 3855 schema_entry->value->has_unit = true; 3856 schema_entry->value->unit = STATS_UNIT_CYCLES; 3857 break; 3858 case KVM_STATS_UNIT_SECONDS: 3859 schema_entry->value->has_unit = true; 3860 schema_entry->value->unit = STATS_UNIT_SECONDS; 3861 break; 3862 default: 3863 goto exit; 3864 } 3865 3866 schema_entry->value->exponent = pdesc->exponent; 3867 if (pdesc->exponent) { 3868 switch (pdesc->flags & KVM_STATS_BASE_MASK) { 3869 case KVM_STATS_BASE_POW10: 3870 schema_entry->value->has_base = true; 3871 schema_entry->value->base = 10; 3872 break; 3873 case KVM_STATS_BASE_POW2: 3874 schema_entry->value->has_base = true; 3875 schema_entry->value->base = 2; 3876 break; 3877 default: 3878 goto exit; 3879 } 3880 } 3881 3882 schema_entry->value->name = g_strdup(pdesc->name); 3883 schema_entry->next = list; 3884 return schema_entry; 3885 exit: 3886 g_free(schema_entry->value); 3887 g_free(schema_entry); 3888 return list; 3889 } 3890 3891 /* Cached stats descriptors */ 3892 typedef struct StatsDescriptors { 3893 const char *ident; /* cache key, currently the StatsTarget */ 3894 struct kvm_stats_desc *kvm_stats_desc; 3895 struct kvm_stats_header *kvm_stats_header; 3896 QTAILQ_ENTRY(StatsDescriptors) next; 3897 } StatsDescriptors; 3898 3899 static QTAILQ_HEAD(, StatsDescriptors) stats_descriptors = 3900 QTAILQ_HEAD_INITIALIZER(stats_descriptors); 3901 3902 /* 3903 * Return the descriptors for 'target', that either have already been read 3904 * or are retrieved from 'stats_fd'. 3905 */ 3906 static StatsDescriptors *find_stats_descriptors(StatsTarget target, int stats_fd, 3907 Error **errp) 3908 { 3909 StatsDescriptors *descriptors; 3910 const char *ident; 3911 struct kvm_stats_desc *kvm_stats_desc; 3912 struct kvm_stats_header *kvm_stats_header; 3913 size_t size_desc; 3914 ssize_t ret; 3915 3916 ident = StatsTarget_str(target); 3917 QTAILQ_FOREACH(descriptors, &stats_descriptors, next) { 3918 if (g_str_equal(descriptors->ident, ident)) { 3919 return descriptors; 3920 } 3921 } 3922 3923 descriptors = g_new0(StatsDescriptors, 1); 3924 3925 /* Read stats header */ 3926 kvm_stats_header = g_malloc(sizeof(*kvm_stats_header)); 3927 ret = read(stats_fd, kvm_stats_header, sizeof(*kvm_stats_header)); 3928 if (ret != sizeof(*kvm_stats_header)) { 3929 error_setg(errp, "KVM stats: failed to read stats header: " 3930 "expected %zu actual %zu", 3931 sizeof(*kvm_stats_header), ret); 3932 g_free(descriptors); 3933 return NULL; 3934 } 3935 size_desc = sizeof(*kvm_stats_desc) + kvm_stats_header->name_size; 3936 3937 /* Read stats descriptors */ 3938 kvm_stats_desc = g_malloc0_n(kvm_stats_header->num_desc, size_desc); 3939 ret = pread(stats_fd, kvm_stats_desc, 3940 size_desc * kvm_stats_header->num_desc, 3941 kvm_stats_header->desc_offset); 3942 3943 if (ret != size_desc * kvm_stats_header->num_desc) { 3944 error_setg(errp, "KVM stats: failed to read stats descriptors: " 3945 "expected %zu actual %zu", 3946 size_desc * kvm_stats_header->num_desc, ret); 3947 g_free(descriptors); 3948 g_free(kvm_stats_desc); 3949 return NULL; 3950 } 3951 descriptors->kvm_stats_header = kvm_stats_header; 3952 descriptors->kvm_stats_desc = kvm_stats_desc; 3953 descriptors->ident = ident; 3954 QTAILQ_INSERT_TAIL(&stats_descriptors, descriptors, next); 3955 return descriptors; 3956 } 3957 3958 static void query_stats(StatsResultList **result, StatsTarget target, 3959 strList *names, int stats_fd, Error **errp) 3960 { 3961 struct kvm_stats_desc *kvm_stats_desc; 3962 struct kvm_stats_header *kvm_stats_header; 3963 StatsDescriptors *descriptors; 3964 g_autofree uint64_t *stats_data = NULL; 3965 struct kvm_stats_desc *pdesc; 3966 StatsList *stats_list = NULL; 3967 size_t size_desc, size_data = 0; 3968 ssize_t ret; 3969 int i; 3970 3971 descriptors = find_stats_descriptors(target, stats_fd, errp); 3972 if (!descriptors) { 3973 return; 3974 } 3975 3976 kvm_stats_header = descriptors->kvm_stats_header; 3977 kvm_stats_desc = descriptors->kvm_stats_desc; 3978 size_desc = sizeof(*kvm_stats_desc) + kvm_stats_header->name_size; 3979 3980 /* Tally the total data size; read schema data */ 3981 for (i = 0; i < kvm_stats_header->num_desc; ++i) { 3982 pdesc = (void *)kvm_stats_desc + i * size_desc; 3983 size_data += pdesc->size * sizeof(*stats_data); 3984 } 3985 3986 stats_data = g_malloc0(size_data); 3987 ret = pread(stats_fd, stats_data, size_data, kvm_stats_header->data_offset); 3988 3989 if (ret != size_data) { 3990 error_setg(errp, "KVM stats: failed to read data: " 3991 "expected %zu actual %zu", size_data, ret); 3992 return; 3993 } 3994 3995 for (i = 0; i < kvm_stats_header->num_desc; ++i) { 3996 uint64_t *stats; 3997 pdesc = (void *)kvm_stats_desc + i * size_desc; 3998 3999 /* Add entry to the list */ 4000 stats = (void *)stats_data + pdesc->offset; 4001 if (!apply_str_list_filter(pdesc->name, names)) { 4002 continue; 4003 } 4004 stats_list = add_kvmstat_entry(pdesc, stats, stats_list, errp); 4005 } 4006 4007 if (!stats_list) { 4008 return; 4009 } 4010 4011 switch (target) { 4012 case STATS_TARGET_VM: 4013 add_stats_entry(result, STATS_PROVIDER_KVM, NULL, stats_list); 4014 break; 4015 case STATS_TARGET_VCPU: 4016 add_stats_entry(result, STATS_PROVIDER_KVM, 4017 current_cpu->parent_obj.canonical_path, 4018 stats_list); 4019 break; 4020 default: 4021 g_assert_not_reached(); 4022 } 4023 } 4024 4025 static void query_stats_schema(StatsSchemaList **result, StatsTarget target, 4026 int stats_fd, Error **errp) 4027 { 4028 struct kvm_stats_desc *kvm_stats_desc; 4029 struct kvm_stats_header *kvm_stats_header; 4030 StatsDescriptors *descriptors; 4031 struct kvm_stats_desc *pdesc; 4032 StatsSchemaValueList *stats_list = NULL; 4033 size_t size_desc; 4034 int i; 4035 4036 descriptors = find_stats_descriptors(target, stats_fd, errp); 4037 if (!descriptors) { 4038 return; 4039 } 4040 4041 kvm_stats_header = descriptors->kvm_stats_header; 4042 kvm_stats_desc = descriptors->kvm_stats_desc; 4043 size_desc = sizeof(*kvm_stats_desc) + kvm_stats_header->name_size; 4044 4045 /* Tally the total data size; read schema data */ 4046 for (i = 0; i < kvm_stats_header->num_desc; ++i) { 4047 pdesc = (void *)kvm_stats_desc + i * size_desc; 4048 stats_list = add_kvmschema_entry(pdesc, stats_list, errp); 4049 } 4050 4051 add_stats_schema(result, STATS_PROVIDER_KVM, target, stats_list); 4052 } 4053 4054 static void query_stats_vcpu(CPUState *cpu, run_on_cpu_data data) 4055 { 4056 StatsArgs *kvm_stats_args = (StatsArgs *) data.host_ptr; 4057 int stats_fd = kvm_vcpu_ioctl(cpu, KVM_GET_STATS_FD, NULL); 4058 Error *local_err = NULL; 4059 4060 if (stats_fd == -1) { 4061 error_setg_errno(&local_err, errno, "KVM stats: ioctl failed"); 4062 error_propagate(kvm_stats_args->errp, local_err); 4063 return; 4064 } 4065 query_stats(kvm_stats_args->result.stats, STATS_TARGET_VCPU, 4066 kvm_stats_args->names, stats_fd, kvm_stats_args->errp); 4067 close(stats_fd); 4068 } 4069 4070 static void query_stats_schema_vcpu(CPUState *cpu, run_on_cpu_data data) 4071 { 4072 StatsArgs *kvm_stats_args = (StatsArgs *) data.host_ptr; 4073 int stats_fd = kvm_vcpu_ioctl(cpu, KVM_GET_STATS_FD, NULL); 4074 Error *local_err = NULL; 4075 4076 if (stats_fd == -1) { 4077 error_setg_errno(&local_err, errno, "KVM stats: ioctl failed"); 4078 error_propagate(kvm_stats_args->errp, local_err); 4079 return; 4080 } 4081 query_stats_schema(kvm_stats_args->result.schema, STATS_TARGET_VCPU, stats_fd, 4082 kvm_stats_args->errp); 4083 close(stats_fd); 4084 } 4085 4086 static void query_stats_cb(StatsResultList **result, StatsTarget target, 4087 strList *names, strList *targets, Error **errp) 4088 { 4089 KVMState *s = kvm_state; 4090 CPUState *cpu; 4091 int stats_fd; 4092 4093 switch (target) { 4094 case STATS_TARGET_VM: 4095 { 4096 stats_fd = kvm_vm_ioctl(s, KVM_GET_STATS_FD, NULL); 4097 if (stats_fd == -1) { 4098 error_setg_errno(errp, errno, "KVM stats: ioctl failed"); 4099 return; 4100 } 4101 query_stats(result, target, names, stats_fd, errp); 4102 close(stats_fd); 4103 break; 4104 } 4105 case STATS_TARGET_VCPU: 4106 { 4107 StatsArgs stats_args; 4108 stats_args.result.stats = result; 4109 stats_args.names = names; 4110 stats_args.errp = errp; 4111 CPU_FOREACH(cpu) { 4112 if (!apply_str_list_filter(cpu->parent_obj.canonical_path, targets)) { 4113 continue; 4114 } 4115 run_on_cpu(cpu, query_stats_vcpu, RUN_ON_CPU_HOST_PTR(&stats_args)); 4116 } 4117 break; 4118 } 4119 default: 4120 break; 4121 } 4122 } 4123 4124 void query_stats_schemas_cb(StatsSchemaList **result, Error **errp) 4125 { 4126 StatsArgs stats_args; 4127 KVMState *s = kvm_state; 4128 int stats_fd; 4129 4130 stats_fd = kvm_vm_ioctl(s, KVM_GET_STATS_FD, NULL); 4131 if (stats_fd == -1) { 4132 error_setg_errno(errp, errno, "KVM stats: ioctl failed"); 4133 return; 4134 } 4135 query_stats_schema(result, STATS_TARGET_VM, stats_fd, errp); 4136 close(stats_fd); 4137 4138 if (first_cpu) { 4139 stats_args.result.schema = result; 4140 stats_args.errp = errp; 4141 run_on_cpu(first_cpu, query_stats_schema_vcpu, RUN_ON_CPU_HOST_PTR(&stats_args)); 4142 } 4143 } 4144