1 | /* |
2 | * QEMU KVM support |
3 | * |
4 | * Copyright IBM, Corp. 2008 |
5 | * Red Hat, Inc. 2008 |
6 | * |
7 | * Authors: |
8 | * Anthony Liguori <aliguori@us.ibm.com> |
9 | * Glauber Costa <gcosta@redhat.com> |
10 | * |
11 | * This work is licensed under the terms of the GNU GPL, version 2 or later. |
12 | * See the COPYING file in the top-level directory. |
13 | * |
14 | */ |
15 | |
16 | #include "qemu/osdep.h" |
17 | #include <sys/ioctl.h> |
18 | |
19 | #include <linux/kvm.h> |
20 | |
21 | #include "qemu/atomic.h" |
22 | #include "qemu/option.h" |
23 | #include "qemu/config-file.h" |
24 | #include "qemu/error-report.h" |
25 | #include "qapi/error.h" |
26 | #include "hw/pci/msi.h" |
27 | #include "hw/pci/msix.h" |
28 | #include "hw/s390x/adapter.h" |
29 | #include "exec/gdbstub.h" |
30 | #include "sysemu/kvm_int.h" |
31 | #include "sysemu/runstate.h" |
32 | #include "sysemu/cpus.h" |
33 | #include "sysemu/sysemu.h" |
34 | #include "qemu/bswap.h" |
35 | #include "exec/memory.h" |
36 | #include "exec/ram_addr.h" |
37 | #include "exec/address-spaces.h" |
38 | #include "qemu/event_notifier.h" |
39 | #include "qemu/main-loop.h" |
40 | #include "trace.h" |
41 | #include "hw/irq.h" |
42 | #include "sysemu/sev.h" |
43 | #include "sysemu/balloon.h" |
44 | |
45 | #include "hw/boards.h" |
46 | |
47 | /* This check must be after config-host.h is included */ |
48 | #ifdef CONFIG_EVENTFD |
49 | #include <sys/eventfd.h> |
50 | #endif |
51 | |
52 | /* KVM uses PAGE_SIZE in its definition of KVM_COALESCED_MMIO_MAX. We |
53 | * need to use the real host PAGE_SIZE, as that's what KVM will use. |
54 | */ |
55 | #define PAGE_SIZE getpagesize() |
56 | |
57 | //#define DEBUG_KVM |
58 | |
59 | #ifdef DEBUG_KVM |
60 | #define DPRINTF(fmt, ...) \ |
61 | do { fprintf(stderr, fmt, ## __VA_ARGS__); } while (0) |
62 | #else |
63 | #define DPRINTF(fmt, ...) \ |
64 | do { } while (0) |
65 | #endif |
66 | |
67 | #define KVM_MSI_HASHTAB_SIZE 256 |
68 | |
69 | struct KVMParkedVcpu { |
70 | unsigned long vcpu_id; |
71 | int kvm_fd; |
72 | QLIST_ENTRY(KVMParkedVcpu) node; |
73 | }; |
74 | |
75 | struct KVMState |
76 | { |
77 | AccelState parent_obj; |
78 | |
79 | int nr_slots; |
80 | int fd; |
81 | int vmfd; |
82 | int coalesced_mmio; |
83 | int coalesced_pio; |
84 | struct kvm_coalesced_mmio_ring *coalesced_mmio_ring; |
85 | bool coalesced_flush_in_progress; |
86 | int vcpu_events; |
87 | int robust_singlestep; |
88 | int debugregs; |
89 | #ifdef KVM_CAP_SET_GUEST_DEBUG |
90 | QTAILQ_HEAD(, kvm_sw_breakpoint) kvm_sw_breakpoints; |
91 | #endif |
92 | int max_nested_state_len; |
93 | int many_ioeventfds; |
94 | int intx_set_mask; |
95 | bool sync_mmu; |
96 | bool manual_dirty_log_protect; |
97 | /* The man page (and posix) say ioctl numbers are signed int, but |
98 | * they're not. Linux, glibc and *BSD all treat ioctl numbers as |
99 | * unsigned, and treating them as signed here can break things */ |
100 | unsigned irq_set_ioctl; |
101 | unsigned int sigmask_len; |
102 | GHashTable *gsimap; |
103 | #ifdef KVM_CAP_IRQ_ROUTING |
104 | struct kvm_irq_routing *irq_routes; |
105 | int nr_allocated_irq_routes; |
106 | unsigned long *used_gsi_bitmap; |
107 | unsigned int gsi_count; |
108 | QTAILQ_HEAD(, KVMMSIRoute) msi_hashtab[KVM_MSI_HASHTAB_SIZE]; |
109 | #endif |
110 | KVMMemoryListener memory_listener; |
111 | QLIST_HEAD(, KVMParkedVcpu) kvm_parked_vcpus; |
112 | |
113 | /* memory encryption */ |
114 | void *memcrypt_handle; |
115 | int (*memcrypt_encrypt_data)(void *handle, uint8_t *ptr, uint64_t len); |
116 | |
117 | /* For "info mtree -f" to tell if an MR is registered in KVM */ |
118 | int nr_as; |
119 | struct KVMAs { |
120 | KVMMemoryListener *ml; |
121 | AddressSpace *as; |
122 | } *as; |
123 | }; |
124 | |
125 | KVMState *kvm_state; |
126 | bool kvm_kernel_irqchip; |
127 | bool kvm_split_irqchip; |
128 | bool kvm_async_interrupts_allowed; |
129 | bool kvm_halt_in_kernel_allowed; |
130 | bool kvm_eventfds_allowed; |
131 | bool kvm_irqfds_allowed; |
132 | bool kvm_resamplefds_allowed; |
133 | bool kvm_msi_via_irqfd_allowed; |
134 | bool kvm_gsi_routing_allowed; |
135 | bool kvm_gsi_direct_mapping; |
136 | bool kvm_allowed; |
137 | bool kvm_readonly_mem_allowed; |
138 | bool kvm_vm_attributes_allowed; |
139 | bool kvm_direct_msi_allowed; |
140 | bool kvm_ioeventfd_any_length_allowed; |
141 | bool kvm_msi_use_devid; |
142 | static bool kvm_immediate_exit; |
143 | |
144 | static const KVMCapabilityInfo kvm_required_capabilites[] = { |
145 | KVM_CAP_INFO(USER_MEMORY), |
146 | KVM_CAP_INFO(DESTROY_MEMORY_REGION_WORKS), |
147 | KVM_CAP_INFO(JOIN_MEMORY_REGIONS_WORKS), |
148 | KVM_CAP_LAST_INFO |
149 | }; |
150 | |
151 | #define kvm_slots_lock(kml) qemu_mutex_lock(&(kml)->slots_lock) |
152 | #define kvm_slots_unlock(kml) qemu_mutex_unlock(&(kml)->slots_lock) |
153 | |
154 | int kvm_get_max_memslots(void) |
155 | { |
156 | KVMState *s = KVM_STATE(current_machine->accelerator); |
157 | |
158 | return s->nr_slots; |
159 | } |
160 | |
161 | bool kvm_memcrypt_enabled(void) |
162 | { |
163 | if (kvm_state && kvm_state->memcrypt_handle) { |
164 | return true; |
165 | } |
166 | |
167 | return false; |
168 | } |
169 | |
170 | int kvm_memcrypt_encrypt_data(uint8_t *ptr, uint64_t len) |
171 | { |
172 | if (kvm_state->memcrypt_handle && |
173 | kvm_state->memcrypt_encrypt_data) { |
174 | return kvm_state->memcrypt_encrypt_data(kvm_state->memcrypt_handle, |
175 | ptr, len); |
176 | } |
177 | |
178 | return 1; |
179 | } |
180 | |
181 | /* Called with KVMMemoryListener.slots_lock held */ |
182 | static KVMSlot *kvm_get_free_slot(KVMMemoryListener *kml) |
183 | { |
184 | KVMState *s = kvm_state; |
185 | int i; |
186 | |
187 | for (i = 0; i < s->nr_slots; i++) { |
188 | if (kml->slots[i].memory_size == 0) { |
189 | return &kml->slots[i]; |
190 | } |
191 | } |
192 | |
193 | return NULL; |
194 | } |
195 | |
196 | bool kvm_has_free_slot(MachineState *ms) |
197 | { |
198 | KVMState *s = KVM_STATE(ms->accelerator); |
199 | bool result; |
200 | KVMMemoryListener *kml = &s->memory_listener; |
201 | |
202 | kvm_slots_lock(kml); |
203 | result = !!kvm_get_free_slot(kml); |
204 | kvm_slots_unlock(kml); |
205 | |
206 | return result; |
207 | } |
208 | |
209 | /* Called with KVMMemoryListener.slots_lock held */ |
210 | static KVMSlot *kvm_alloc_slot(KVMMemoryListener *kml) |
211 | { |
212 | KVMSlot *slot = kvm_get_free_slot(kml); |
213 | |
214 | if (slot) { |
215 | return slot; |
216 | } |
217 | |
218 | fprintf(stderr, "%s: no free slot available\n" , __func__); |
219 | abort(); |
220 | } |
221 | |
222 | static KVMSlot *kvm_lookup_matching_slot(KVMMemoryListener *kml, |
223 | hwaddr start_addr, |
224 | hwaddr size) |
225 | { |
226 | KVMState *s = kvm_state; |
227 | int i; |
228 | |
229 | for (i = 0; i < s->nr_slots; i++) { |
230 | KVMSlot *mem = &kml->slots[i]; |
231 | |
232 | if (start_addr == mem->start_addr && size == mem->memory_size) { |
233 | return mem; |
234 | } |
235 | } |
236 | |
237 | return NULL; |
238 | } |
239 | |
240 | /* |
241 | * Calculate and align the start address and the size of the section. |
242 | * Return the size. If the size is 0, the aligned section is empty. |
243 | */ |
244 | static hwaddr kvm_align_section(MemoryRegionSection *section, |
245 | hwaddr *start) |
246 | { |
247 | hwaddr size = int128_get64(section->size); |
248 | hwaddr delta, aligned; |
249 | |
250 | /* kvm works in page size chunks, but the function may be called |
251 | with sub-page size and unaligned start address. Pad the start |
252 | address to next and truncate size to previous page boundary. */ |
253 | aligned = ROUND_UP(section->offset_within_address_space, |
254 | qemu_real_host_page_size); |
255 | delta = aligned - section->offset_within_address_space; |
256 | *start = aligned; |
257 | if (delta > size) { |
258 | return 0; |
259 | } |
260 | |
261 | return (size - delta) & qemu_real_host_page_mask; |
262 | } |
263 | |
264 | int kvm_physical_memory_addr_from_host(KVMState *s, void *ram, |
265 | hwaddr *phys_addr) |
266 | { |
267 | KVMMemoryListener *kml = &s->memory_listener; |
268 | int i, ret = 0; |
269 | |
270 | kvm_slots_lock(kml); |
271 | for (i = 0; i < s->nr_slots; i++) { |
272 | KVMSlot *mem = &kml->slots[i]; |
273 | |
274 | if (ram >= mem->ram && ram < mem->ram + mem->memory_size) { |
275 | *phys_addr = mem->start_addr + (ram - mem->ram); |
276 | ret = 1; |
277 | break; |
278 | } |
279 | } |
280 | kvm_slots_unlock(kml); |
281 | |
282 | return ret; |
283 | } |
284 | |
285 | static int kvm_set_user_memory_region(KVMMemoryListener *kml, KVMSlot *slot, bool new) |
286 | { |
287 | KVMState *s = kvm_state; |
288 | struct kvm_userspace_memory_region mem; |
289 | int ret; |
290 | |
291 | mem.slot = slot->slot | (kml->as_id << 16); |
292 | mem.guest_phys_addr = slot->start_addr; |
293 | mem.userspace_addr = (unsigned long)slot->ram; |
294 | mem.flags = slot->flags; |
295 | |
296 | if (slot->memory_size && !new && (mem.flags ^ slot->old_flags) & KVM_MEM_READONLY) { |
297 | /* Set the slot size to 0 before setting the slot to the desired |
298 | * value. This is needed based on KVM commit 75d61fbc. */ |
299 | mem.memory_size = 0; |
300 | kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem); |
301 | } |
302 | mem.memory_size = slot->memory_size; |
303 | ret = kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem); |
304 | slot->old_flags = mem.flags; |
305 | trace_kvm_set_user_memory(mem.slot, mem.flags, mem.guest_phys_addr, |
306 | mem.memory_size, mem.userspace_addr, ret); |
307 | return ret; |
308 | } |
309 | |
310 | int kvm_destroy_vcpu(CPUState *cpu) |
311 | { |
312 | KVMState *s = kvm_state; |
313 | long mmap_size; |
314 | struct KVMParkedVcpu *vcpu = NULL; |
315 | int ret = 0; |
316 | |
317 | DPRINTF("kvm_destroy_vcpu\n" ); |
318 | |
319 | ret = kvm_arch_destroy_vcpu(cpu); |
320 | if (ret < 0) { |
321 | goto err; |
322 | } |
323 | |
324 | mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0); |
325 | if (mmap_size < 0) { |
326 | ret = mmap_size; |
327 | DPRINTF("KVM_GET_VCPU_MMAP_SIZE failed\n" ); |
328 | goto err; |
329 | } |
330 | |
331 | ret = munmap(cpu->kvm_run, mmap_size); |
332 | if (ret < 0) { |
333 | goto err; |
334 | } |
335 | |
336 | vcpu = g_malloc0(sizeof(*vcpu)); |
337 | vcpu->vcpu_id = kvm_arch_vcpu_id(cpu); |
338 | vcpu->kvm_fd = cpu->kvm_fd; |
339 | QLIST_INSERT_HEAD(&kvm_state->kvm_parked_vcpus, vcpu, node); |
340 | err: |
341 | return ret; |
342 | } |
343 | |
344 | static int kvm_get_vcpu(KVMState *s, unsigned long vcpu_id) |
345 | { |
346 | struct KVMParkedVcpu *cpu; |
347 | |
348 | QLIST_FOREACH(cpu, &s->kvm_parked_vcpus, node) { |
349 | if (cpu->vcpu_id == vcpu_id) { |
350 | int kvm_fd; |
351 | |
352 | QLIST_REMOVE(cpu, node); |
353 | kvm_fd = cpu->kvm_fd; |
354 | g_free(cpu); |
355 | return kvm_fd; |
356 | } |
357 | } |
358 | |
359 | return kvm_vm_ioctl(s, KVM_CREATE_VCPU, (void *)vcpu_id); |
360 | } |
361 | |
362 | int kvm_init_vcpu(CPUState *cpu) |
363 | { |
364 | KVMState *s = kvm_state; |
365 | long mmap_size; |
366 | int ret; |
367 | |
368 | DPRINTF("kvm_init_vcpu\n" ); |
369 | |
370 | ret = kvm_get_vcpu(s, kvm_arch_vcpu_id(cpu)); |
371 | if (ret < 0) { |
372 | DPRINTF("kvm_create_vcpu failed\n" ); |
373 | goto err; |
374 | } |
375 | |
376 | cpu->kvm_fd = ret; |
377 | cpu->kvm_state = s; |
378 | cpu->vcpu_dirty = true; |
379 | |
380 | mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0); |
381 | if (mmap_size < 0) { |
382 | ret = mmap_size; |
383 | DPRINTF("KVM_GET_VCPU_MMAP_SIZE failed\n" ); |
384 | goto err; |
385 | } |
386 | |
387 | cpu->kvm_run = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED, |
388 | cpu->kvm_fd, 0); |
389 | if (cpu->kvm_run == MAP_FAILED) { |
390 | ret = -errno; |
391 | DPRINTF("mmap'ing vcpu state failed\n" ); |
392 | goto err; |
393 | } |
394 | |
395 | if (s->coalesced_mmio && !s->coalesced_mmio_ring) { |
396 | s->coalesced_mmio_ring = |
397 | (void *)cpu->kvm_run + s->coalesced_mmio * PAGE_SIZE; |
398 | } |
399 | |
400 | ret = kvm_arch_init_vcpu(cpu); |
401 | err: |
402 | return ret; |
403 | } |
404 | |
405 | /* |
406 | * dirty pages logging control |
407 | */ |
408 | |
409 | static int kvm_mem_flags(MemoryRegion *mr) |
410 | { |
411 | bool readonly = mr->readonly || memory_region_is_romd(mr); |
412 | int flags = 0; |
413 | |
414 | if (memory_region_get_dirty_log_mask(mr) != 0) { |
415 | flags |= KVM_MEM_LOG_DIRTY_PAGES; |
416 | } |
417 | if (readonly && kvm_readonly_mem_allowed) { |
418 | flags |= KVM_MEM_READONLY; |
419 | } |
420 | return flags; |
421 | } |
422 | |
423 | /* Called with KVMMemoryListener.slots_lock held */ |
424 | static int kvm_slot_update_flags(KVMMemoryListener *kml, KVMSlot *mem, |
425 | MemoryRegion *mr) |
426 | { |
427 | mem->flags = kvm_mem_flags(mr); |
428 | |
429 | /* If nothing changed effectively, no need to issue ioctl */ |
430 | if (mem->flags == mem->old_flags) { |
431 | return 0; |
432 | } |
433 | |
434 | return kvm_set_user_memory_region(kml, mem, false); |
435 | } |
436 | |
437 | static int kvm_section_update_flags(KVMMemoryListener *kml, |
438 | MemoryRegionSection *section) |
439 | { |
440 | hwaddr start_addr, size; |
441 | KVMSlot *mem; |
442 | int ret = 0; |
443 | |
444 | size = kvm_align_section(section, &start_addr); |
445 | if (!size) { |
446 | return 0; |
447 | } |
448 | |
449 | kvm_slots_lock(kml); |
450 | |
451 | mem = kvm_lookup_matching_slot(kml, start_addr, size); |
452 | if (!mem) { |
453 | /* We don't have a slot if we want to trap every access. */ |
454 | goto out; |
455 | } |
456 | |
457 | ret = kvm_slot_update_flags(kml, mem, section->mr); |
458 | |
459 | out: |
460 | kvm_slots_unlock(kml); |
461 | return ret; |
462 | } |
463 | |
464 | static void kvm_log_start(MemoryListener *listener, |
465 | MemoryRegionSection *section, |
466 | int old, int new) |
467 | { |
468 | KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener); |
469 | int r; |
470 | |
471 | if (old != 0) { |
472 | return; |
473 | } |
474 | |
475 | r = kvm_section_update_flags(kml, section); |
476 | if (r < 0) { |
477 | abort(); |
478 | } |
479 | } |
480 | |
481 | static void kvm_log_stop(MemoryListener *listener, |
482 | MemoryRegionSection *section, |
483 | int old, int new) |
484 | { |
485 | KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener); |
486 | int r; |
487 | |
488 | if (new != 0) { |
489 | return; |
490 | } |
491 | |
492 | r = kvm_section_update_flags(kml, section); |
493 | if (r < 0) { |
494 | abort(); |
495 | } |
496 | } |
497 | |
498 | /* get kvm's dirty pages bitmap and update qemu's */ |
499 | static int kvm_get_dirty_pages_log_range(MemoryRegionSection *section, |
500 | unsigned long *bitmap) |
501 | { |
502 | ram_addr_t start = section->offset_within_region + |
503 | memory_region_get_ram_addr(section->mr); |
504 | ram_addr_t pages = int128_get64(section->size) / getpagesize(); |
505 | |
506 | cpu_physical_memory_set_dirty_lebitmap(bitmap, start, pages); |
507 | return 0; |
508 | } |
509 | |
510 | #define ALIGN(x, y) (((x)+(y)-1) & ~((y)-1)) |
511 | |
512 | /** |
513 | * kvm_physical_sync_dirty_bitmap - Sync dirty bitmap from kernel space |
514 | * |
515 | * This function will first try to fetch dirty bitmap from the kernel, |
516 | * and then updates qemu's dirty bitmap. |
517 | * |
518 | * NOTE: caller must be with kml->slots_lock held. |
519 | * |
520 | * @kml: the KVM memory listener object |
521 | * @section: the memory section to sync the dirty bitmap with |
522 | */ |
523 | static int kvm_physical_sync_dirty_bitmap(KVMMemoryListener *kml, |
524 | MemoryRegionSection *section) |
525 | { |
526 | KVMState *s = kvm_state; |
527 | struct kvm_dirty_log d = {}; |
528 | KVMSlot *mem; |
529 | hwaddr start_addr, size; |
530 | int ret = 0; |
531 | |
532 | size = kvm_align_section(section, &start_addr); |
533 | if (size) { |
534 | mem = kvm_lookup_matching_slot(kml, start_addr, size); |
535 | if (!mem) { |
536 | /* We don't have a slot if we want to trap every access. */ |
537 | goto out; |
538 | } |
539 | |
540 | /* XXX bad kernel interface alert |
541 | * For dirty bitmap, kernel allocates array of size aligned to |
542 | * bits-per-long. But for case when the kernel is 64bits and |
543 | * the userspace is 32bits, userspace can't align to the same |
544 | * bits-per-long, since sizeof(long) is different between kernel |
545 | * and user space. This way, userspace will provide buffer which |
546 | * may be 4 bytes less than the kernel will use, resulting in |
547 | * userspace memory corruption (which is not detectable by valgrind |
548 | * too, in most cases). |
549 | * So for now, let's align to 64 instead of HOST_LONG_BITS here, in |
550 | * a hope that sizeof(long) won't become >8 any time soon. |
551 | */ |
552 | size = ALIGN(((mem->memory_size) >> TARGET_PAGE_BITS), |
553 | /*HOST_LONG_BITS*/ 64) / 8; |
554 | if (!mem->dirty_bmap) { |
555 | /* Allocate on the first log_sync, once and for all */ |
556 | mem->dirty_bmap = g_malloc0(size); |
557 | } |
558 | |
559 | d.dirty_bitmap = mem->dirty_bmap; |
560 | d.slot = mem->slot | (kml->as_id << 16); |
561 | if (kvm_vm_ioctl(s, KVM_GET_DIRTY_LOG, &d) == -1) { |
562 | DPRINTF("ioctl failed %d\n" , errno); |
563 | ret = -1; |
564 | goto out; |
565 | } |
566 | |
567 | kvm_get_dirty_pages_log_range(section, d.dirty_bitmap); |
568 | } |
569 | out: |
570 | return ret; |
571 | } |
572 | |
573 | /* Alignment requirement for KVM_CLEAR_DIRTY_LOG - 64 pages */ |
574 | #define KVM_CLEAR_LOG_SHIFT 6 |
575 | #define KVM_CLEAR_LOG_ALIGN (qemu_real_host_page_size << KVM_CLEAR_LOG_SHIFT) |
576 | #define KVM_CLEAR_LOG_MASK (-KVM_CLEAR_LOG_ALIGN) |
577 | |
578 | /** |
579 | * kvm_physical_log_clear - Clear the kernel's dirty bitmap for range |
580 | * |
581 | * NOTE: this will be a no-op if we haven't enabled manual dirty log |
582 | * protection in the host kernel because in that case this operation |
583 | * will be done within log_sync(). |
584 | * |
585 | * @kml: the kvm memory listener |
586 | * @section: the memory range to clear dirty bitmap |
587 | */ |
588 | static int kvm_physical_log_clear(KVMMemoryListener *kml, |
589 | MemoryRegionSection *section) |
590 | { |
591 | KVMState *s = kvm_state; |
592 | struct kvm_clear_dirty_log d; |
593 | uint64_t start, end, bmap_start, start_delta, bmap_npages, size; |
594 | unsigned long *bmap_clear = NULL, psize = qemu_real_host_page_size; |
595 | KVMSlot *mem = NULL; |
596 | int ret, i; |
597 | |
598 | if (!s->manual_dirty_log_protect) { |
599 | /* No need to do explicit clear */ |
600 | return 0; |
601 | } |
602 | |
603 | start = section->offset_within_address_space; |
604 | size = int128_get64(section->size); |
605 | |
606 | if (!size) { |
607 | /* Nothing more we can do... */ |
608 | return 0; |
609 | } |
610 | |
611 | kvm_slots_lock(kml); |
612 | |
613 | /* Find any possible slot that covers the section */ |
614 | for (i = 0; i < s->nr_slots; i++) { |
615 | mem = &kml->slots[i]; |
616 | if (mem->start_addr <= start && |
617 | start + size <= mem->start_addr + mem->memory_size) { |
618 | break; |
619 | } |
620 | } |
621 | |
622 | /* |
623 | * We should always find one memslot until this point, otherwise |
624 | * there could be something wrong from the upper layer |
625 | */ |
626 | assert(mem && i != s->nr_slots); |
627 | |
628 | /* |
629 | * We need to extend either the start or the size or both to |
630 | * satisfy the KVM interface requirement. Firstly, do the start |
631 | * page alignment on 64 host pages |
632 | */ |
633 | bmap_start = (start - mem->start_addr) & KVM_CLEAR_LOG_MASK; |
634 | start_delta = start - mem->start_addr - bmap_start; |
635 | bmap_start /= psize; |
636 | |
637 | /* |
638 | * The kernel interface has restriction on the size too, that either: |
639 | * |
640 | * (1) the size is 64 host pages aligned (just like the start), or |
641 | * (2) the size fills up until the end of the KVM memslot. |
642 | */ |
643 | bmap_npages = DIV_ROUND_UP(size + start_delta, KVM_CLEAR_LOG_ALIGN) |
644 | << KVM_CLEAR_LOG_SHIFT; |
645 | end = mem->memory_size / psize; |
646 | if (bmap_npages > end - bmap_start) { |
647 | bmap_npages = end - bmap_start; |
648 | } |
649 | start_delta /= psize; |
650 | |
651 | /* |
652 | * Prepare the bitmap to clear dirty bits. Here we must guarantee |
653 | * that we won't clear any unknown dirty bits otherwise we might |
654 | * accidentally clear some set bits which are not yet synced from |
655 | * the kernel into QEMU's bitmap, then we'll lose track of the |
656 | * guest modifications upon those pages (which can directly lead |
657 | * to guest data loss or panic after migration). |
658 | * |
659 | * Layout of the KVMSlot.dirty_bmap: |
660 | * |
661 | * |<-------- bmap_npages -----------..>| |
662 | * [1] |
663 | * start_delta size |
664 | * |----------------|-------------|------------------|------------| |
665 | * ^ ^ ^ ^ |
666 | * | | | | |
667 | * start bmap_start (start) end |
668 | * of memslot of memslot |
669 | * |
670 | * [1] bmap_npages can be aligned to either 64 pages or the end of slot |
671 | */ |
672 | |
673 | assert(bmap_start % BITS_PER_LONG == 0); |
674 | /* We should never do log_clear before log_sync */ |
675 | assert(mem->dirty_bmap); |
676 | if (start_delta) { |
677 | /* Slow path - we need to manipulate a temp bitmap */ |
678 | bmap_clear = bitmap_new(bmap_npages); |
679 | bitmap_copy_with_src_offset(bmap_clear, mem->dirty_bmap, |
680 | bmap_start, start_delta + size / psize); |
681 | /* |
682 | * We need to fill the holes at start because that was not |
683 | * specified by the caller and we extended the bitmap only for |
684 | * 64 pages alignment |
685 | */ |
686 | bitmap_clear(bmap_clear, 0, start_delta); |
687 | d.dirty_bitmap = bmap_clear; |
688 | } else { |
689 | /* Fast path - start address aligns well with BITS_PER_LONG */ |
690 | d.dirty_bitmap = mem->dirty_bmap + BIT_WORD(bmap_start); |
691 | } |
692 | |
693 | d.first_page = bmap_start; |
694 | /* It should never overflow. If it happens, say something */ |
695 | assert(bmap_npages <= UINT32_MAX); |
696 | d.num_pages = bmap_npages; |
697 | d.slot = mem->slot | (kml->as_id << 16); |
698 | |
699 | if (kvm_vm_ioctl(s, KVM_CLEAR_DIRTY_LOG, &d) == -1) { |
700 | ret = -errno; |
701 | error_report("%s: KVM_CLEAR_DIRTY_LOG failed, slot=%d, " |
702 | "start=0x%" PRIx64", size=0x%" PRIx32", errno=%d" , |
703 | __func__, d.slot, (uint64_t)d.first_page, |
704 | (uint32_t)d.num_pages, ret); |
705 | } else { |
706 | ret = 0; |
707 | trace_kvm_clear_dirty_log(d.slot, d.first_page, d.num_pages); |
708 | } |
709 | |
710 | /* |
711 | * After we have updated the remote dirty bitmap, we update the |
712 | * cached bitmap as well for the memslot, then if another user |
713 | * clears the same region we know we shouldn't clear it again on |
714 | * the remote otherwise it's data loss as well. |
715 | */ |
716 | bitmap_clear(mem->dirty_bmap, bmap_start + start_delta, |
717 | size / psize); |
718 | /* This handles the NULL case well */ |
719 | g_free(bmap_clear); |
720 | |
721 | kvm_slots_unlock(kml); |
722 | |
723 | return ret; |
724 | } |
725 | |
726 | static void kvm_coalesce_mmio_region(MemoryListener *listener, |
727 | MemoryRegionSection *secion, |
728 | hwaddr start, hwaddr size) |
729 | { |
730 | KVMState *s = kvm_state; |
731 | |
732 | if (s->coalesced_mmio) { |
733 | struct kvm_coalesced_mmio_zone zone; |
734 | |
735 | zone.addr = start; |
736 | zone.size = size; |
737 | zone.pad = 0; |
738 | |
739 | (void)kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone); |
740 | } |
741 | } |
742 | |
743 | static void kvm_uncoalesce_mmio_region(MemoryListener *listener, |
744 | MemoryRegionSection *secion, |
745 | hwaddr start, hwaddr size) |
746 | { |
747 | KVMState *s = kvm_state; |
748 | |
749 | if (s->coalesced_mmio) { |
750 | struct kvm_coalesced_mmio_zone zone; |
751 | |
752 | zone.addr = start; |
753 | zone.size = size; |
754 | zone.pad = 0; |
755 | |
756 | (void)kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone); |
757 | } |
758 | } |
759 | |
760 | static void kvm_coalesce_pio_add(MemoryListener *listener, |
761 | MemoryRegionSection *section, |
762 | hwaddr start, hwaddr size) |
763 | { |
764 | KVMState *s = kvm_state; |
765 | |
766 | if (s->coalesced_pio) { |
767 | struct kvm_coalesced_mmio_zone zone; |
768 | |
769 | zone.addr = start; |
770 | zone.size = size; |
771 | zone.pio = 1; |
772 | |
773 | (void)kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone); |
774 | } |
775 | } |
776 | |
777 | static void kvm_coalesce_pio_del(MemoryListener *listener, |
778 | MemoryRegionSection *section, |
779 | hwaddr start, hwaddr size) |
780 | { |
781 | KVMState *s = kvm_state; |
782 | |
783 | if (s->coalesced_pio) { |
784 | struct kvm_coalesced_mmio_zone zone; |
785 | |
786 | zone.addr = start; |
787 | zone.size = size; |
788 | zone.pio = 1; |
789 | |
790 | (void)kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone); |
791 | } |
792 | } |
793 | |
794 | static MemoryListener kvm_coalesced_pio_listener = { |
795 | .coalesced_io_add = kvm_coalesce_pio_add, |
796 | .coalesced_io_del = kvm_coalesce_pio_del, |
797 | }; |
798 | |
799 | int kvm_check_extension(KVMState *s, unsigned int extension) |
800 | { |
801 | int ret; |
802 | |
803 | ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, extension); |
804 | if (ret < 0) { |
805 | ret = 0; |
806 | } |
807 | |
808 | return ret; |
809 | } |
810 | |
811 | int kvm_vm_check_extension(KVMState *s, unsigned int extension) |
812 | { |
813 | int ret; |
814 | |
815 | ret = kvm_vm_ioctl(s, KVM_CHECK_EXTENSION, extension); |
816 | if (ret < 0) { |
817 | /* VM wide version not implemented, use global one instead */ |
818 | ret = kvm_check_extension(s, extension); |
819 | } |
820 | |
821 | return ret; |
822 | } |
823 | |
824 | static uint32_t adjust_ioeventfd_endianness(uint32_t val, uint32_t size) |
825 | { |
826 | #if defined(HOST_WORDS_BIGENDIAN) != defined(TARGET_WORDS_BIGENDIAN) |
827 | /* The kernel expects ioeventfd values in HOST_WORDS_BIGENDIAN |
828 | * endianness, but the memory core hands them in target endianness. |
829 | * For example, PPC is always treated as big-endian even if running |
830 | * on KVM and on PPC64LE. Correct here. |
831 | */ |
832 | switch (size) { |
833 | case 2: |
834 | val = bswap16(val); |
835 | break; |
836 | case 4: |
837 | val = bswap32(val); |
838 | break; |
839 | } |
840 | #endif |
841 | return val; |
842 | } |
843 | |
844 | static int kvm_set_ioeventfd_mmio(int fd, hwaddr addr, uint32_t val, |
845 | bool assign, uint32_t size, bool datamatch) |
846 | { |
847 | int ret; |
848 | struct kvm_ioeventfd iofd = { |
849 | .datamatch = datamatch ? adjust_ioeventfd_endianness(val, size) : 0, |
850 | .addr = addr, |
851 | .len = size, |
852 | .flags = 0, |
853 | .fd = fd, |
854 | }; |
855 | |
856 | trace_kvm_set_ioeventfd_mmio(fd, (uint64_t)addr, val, assign, size, |
857 | datamatch); |
858 | if (!kvm_enabled()) { |
859 | return -ENOSYS; |
860 | } |
861 | |
862 | if (datamatch) { |
863 | iofd.flags |= KVM_IOEVENTFD_FLAG_DATAMATCH; |
864 | } |
865 | if (!assign) { |
866 | iofd.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN; |
867 | } |
868 | |
869 | ret = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &iofd); |
870 | |
871 | if (ret < 0) { |
872 | return -errno; |
873 | } |
874 | |
875 | return 0; |
876 | } |
877 | |
878 | static int kvm_set_ioeventfd_pio(int fd, uint16_t addr, uint16_t val, |
879 | bool assign, uint32_t size, bool datamatch) |
880 | { |
881 | struct kvm_ioeventfd kick = { |
882 | .datamatch = datamatch ? adjust_ioeventfd_endianness(val, size) : 0, |
883 | .addr = addr, |
884 | .flags = KVM_IOEVENTFD_FLAG_PIO, |
885 | .len = size, |
886 | .fd = fd, |
887 | }; |
888 | int r; |
889 | trace_kvm_set_ioeventfd_pio(fd, addr, val, assign, size, datamatch); |
890 | if (!kvm_enabled()) { |
891 | return -ENOSYS; |
892 | } |
893 | if (datamatch) { |
894 | kick.flags |= KVM_IOEVENTFD_FLAG_DATAMATCH; |
895 | } |
896 | if (!assign) { |
897 | kick.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN; |
898 | } |
899 | r = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &kick); |
900 | if (r < 0) { |
901 | return r; |
902 | } |
903 | return 0; |
904 | } |
905 | |
906 | |
907 | static int kvm_check_many_ioeventfds(void) |
908 | { |
909 | /* Userspace can use ioeventfd for io notification. This requires a host |
910 | * that supports eventfd(2) and an I/O thread; since eventfd does not |
911 | * support SIGIO it cannot interrupt the vcpu. |
912 | * |
913 | * Older kernels have a 6 device limit on the KVM io bus. Find out so we |
914 | * can avoid creating too many ioeventfds. |
915 | */ |
916 | #if defined(CONFIG_EVENTFD) |
917 | int ioeventfds[7]; |
918 | int i, ret = 0; |
919 | for (i = 0; i < ARRAY_SIZE(ioeventfds); i++) { |
920 | ioeventfds[i] = eventfd(0, EFD_CLOEXEC); |
921 | if (ioeventfds[i] < 0) { |
922 | break; |
923 | } |
924 | ret = kvm_set_ioeventfd_pio(ioeventfds[i], 0, i, true, 2, true); |
925 | if (ret < 0) { |
926 | close(ioeventfds[i]); |
927 | break; |
928 | } |
929 | } |
930 | |
931 | /* Decide whether many devices are supported or not */ |
932 | ret = i == ARRAY_SIZE(ioeventfds); |
933 | |
934 | while (i-- > 0) { |
935 | kvm_set_ioeventfd_pio(ioeventfds[i], 0, i, false, 2, true); |
936 | close(ioeventfds[i]); |
937 | } |
938 | return ret; |
939 | #else |
940 | return 0; |
941 | #endif |
942 | } |
943 | |
944 | static const KVMCapabilityInfo * |
945 | kvm_check_extension_list(KVMState *s, const KVMCapabilityInfo *list) |
946 | { |
947 | while (list->name) { |
948 | if (!kvm_check_extension(s, list->value)) { |
949 | return list; |
950 | } |
951 | list++; |
952 | } |
953 | return NULL; |
954 | } |
955 | |
956 | static void kvm_set_phys_mem(KVMMemoryListener *kml, |
957 | MemoryRegionSection *section, bool add) |
958 | { |
959 | KVMSlot *mem; |
960 | int err; |
961 | MemoryRegion *mr = section->mr; |
962 | bool writeable = !mr->readonly && !mr->rom_device; |
963 | hwaddr start_addr, size; |
964 | void *ram; |
965 | |
966 | if (!memory_region_is_ram(mr)) { |
967 | if (writeable || !kvm_readonly_mem_allowed) { |
968 | return; |
969 | } else if (!mr->romd_mode) { |
970 | /* If the memory device is not in romd_mode, then we actually want |
971 | * to remove the kvm memory slot so all accesses will trap. */ |
972 | add = false; |
973 | } |
974 | } |
975 | |
976 | size = kvm_align_section(section, &start_addr); |
977 | if (!size) { |
978 | return; |
979 | } |
980 | |
981 | /* use aligned delta to align the ram address */ |
982 | ram = memory_region_get_ram_ptr(mr) + section->offset_within_region + |
983 | (start_addr - section->offset_within_address_space); |
984 | |
985 | kvm_slots_lock(kml); |
986 | |
987 | if (!add) { |
988 | mem = kvm_lookup_matching_slot(kml, start_addr, size); |
989 | if (!mem) { |
990 | goto out; |
991 | } |
992 | if (mem->flags & KVM_MEM_LOG_DIRTY_PAGES) { |
993 | kvm_physical_sync_dirty_bitmap(kml, section); |
994 | } |
995 | |
996 | /* unregister the slot */ |
997 | g_free(mem->dirty_bmap); |
998 | mem->dirty_bmap = NULL; |
999 | mem->memory_size = 0; |
1000 | mem->flags = 0; |
1001 | err = kvm_set_user_memory_region(kml, mem, false); |
1002 | if (err) { |
1003 | fprintf(stderr, "%s: error unregistering slot: %s\n" , |
1004 | __func__, strerror(-err)); |
1005 | abort(); |
1006 | } |
1007 | goto out; |
1008 | } |
1009 | |
1010 | /* register the new slot */ |
1011 | mem = kvm_alloc_slot(kml); |
1012 | mem->memory_size = size; |
1013 | mem->start_addr = start_addr; |
1014 | mem->ram = ram; |
1015 | mem->flags = kvm_mem_flags(mr); |
1016 | |
1017 | err = kvm_set_user_memory_region(kml, mem, true); |
1018 | if (err) { |
1019 | fprintf(stderr, "%s: error registering slot: %s\n" , __func__, |
1020 | strerror(-err)); |
1021 | abort(); |
1022 | } |
1023 | |
1024 | out: |
1025 | kvm_slots_unlock(kml); |
1026 | } |
1027 | |
1028 | static void kvm_region_add(MemoryListener *listener, |
1029 | MemoryRegionSection *section) |
1030 | { |
1031 | KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener); |
1032 | |
1033 | memory_region_ref(section->mr); |
1034 | kvm_set_phys_mem(kml, section, true); |
1035 | } |
1036 | |
1037 | static void kvm_region_del(MemoryListener *listener, |
1038 | MemoryRegionSection *section) |
1039 | { |
1040 | KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener); |
1041 | |
1042 | kvm_set_phys_mem(kml, section, false); |
1043 | memory_region_unref(section->mr); |
1044 | } |
1045 | |
1046 | static void kvm_log_sync(MemoryListener *listener, |
1047 | MemoryRegionSection *section) |
1048 | { |
1049 | KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener); |
1050 | int r; |
1051 | |
1052 | kvm_slots_lock(kml); |
1053 | r = kvm_physical_sync_dirty_bitmap(kml, section); |
1054 | kvm_slots_unlock(kml); |
1055 | if (r < 0) { |
1056 | abort(); |
1057 | } |
1058 | } |
1059 | |
1060 | static void kvm_log_clear(MemoryListener *listener, |
1061 | MemoryRegionSection *section) |
1062 | { |
1063 | KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener); |
1064 | int r; |
1065 | |
1066 | r = kvm_physical_log_clear(kml, section); |
1067 | if (r < 0) { |
1068 | error_report_once("%s: kvm log clear failed: mr=%s " |
1069 | "offset=%" HWADDR_PRIx" size=%" PRIx64, __func__, |
1070 | section->mr->name, section->offset_within_region, |
1071 | int128_get64(section->size)); |
1072 | abort(); |
1073 | } |
1074 | } |
1075 | |
1076 | static void kvm_mem_ioeventfd_add(MemoryListener *listener, |
1077 | MemoryRegionSection *section, |
1078 | bool match_data, uint64_t data, |
1079 | EventNotifier *e) |
1080 | { |
1081 | int fd = event_notifier_get_fd(e); |
1082 | int r; |
1083 | |
1084 | r = kvm_set_ioeventfd_mmio(fd, section->offset_within_address_space, |
1085 | data, true, int128_get64(section->size), |
1086 | match_data); |
1087 | if (r < 0) { |
1088 | fprintf(stderr, "%s: error adding ioeventfd: %s (%d)\n" , |
1089 | __func__, strerror(-r), -r); |
1090 | abort(); |
1091 | } |
1092 | } |
1093 | |
1094 | static void kvm_mem_ioeventfd_del(MemoryListener *listener, |
1095 | MemoryRegionSection *section, |
1096 | bool match_data, uint64_t data, |
1097 | EventNotifier *e) |
1098 | { |
1099 | int fd = event_notifier_get_fd(e); |
1100 | int r; |
1101 | |
1102 | r = kvm_set_ioeventfd_mmio(fd, section->offset_within_address_space, |
1103 | data, false, int128_get64(section->size), |
1104 | match_data); |
1105 | if (r < 0) { |
1106 | fprintf(stderr, "%s: error deleting ioeventfd: %s (%d)\n" , |
1107 | __func__, strerror(-r), -r); |
1108 | abort(); |
1109 | } |
1110 | } |
1111 | |
1112 | static void kvm_io_ioeventfd_add(MemoryListener *listener, |
1113 | MemoryRegionSection *section, |
1114 | bool match_data, uint64_t data, |
1115 | EventNotifier *e) |
1116 | { |
1117 | int fd = event_notifier_get_fd(e); |
1118 | int r; |
1119 | |
1120 | r = kvm_set_ioeventfd_pio(fd, section->offset_within_address_space, |
1121 | data, true, int128_get64(section->size), |
1122 | match_data); |
1123 | if (r < 0) { |
1124 | fprintf(stderr, "%s: error adding ioeventfd: %s (%d)\n" , |
1125 | __func__, strerror(-r), -r); |
1126 | abort(); |
1127 | } |
1128 | } |
1129 | |
1130 | static void kvm_io_ioeventfd_del(MemoryListener *listener, |
1131 | MemoryRegionSection *section, |
1132 | bool match_data, uint64_t data, |
1133 | EventNotifier *e) |
1134 | |
1135 | { |
1136 | int fd = event_notifier_get_fd(e); |
1137 | int r; |
1138 | |
1139 | r = kvm_set_ioeventfd_pio(fd, section->offset_within_address_space, |
1140 | data, false, int128_get64(section->size), |
1141 | match_data); |
1142 | if (r < 0) { |
1143 | fprintf(stderr, "%s: error deleting ioeventfd: %s (%d)\n" , |
1144 | __func__, strerror(-r), -r); |
1145 | abort(); |
1146 | } |
1147 | } |
1148 | |
1149 | void kvm_memory_listener_register(KVMState *s, KVMMemoryListener *kml, |
1150 | AddressSpace *as, int as_id) |
1151 | { |
1152 | int i; |
1153 | |
1154 | qemu_mutex_init(&kml->slots_lock); |
1155 | kml->slots = g_malloc0(s->nr_slots * sizeof(KVMSlot)); |
1156 | kml->as_id = as_id; |
1157 | |
1158 | for (i = 0; i < s->nr_slots; i++) { |
1159 | kml->slots[i].slot = i; |
1160 | } |
1161 | |
1162 | kml->listener.region_add = kvm_region_add; |
1163 | kml->listener.region_del = kvm_region_del; |
1164 | kml->listener.log_start = kvm_log_start; |
1165 | kml->listener.log_stop = kvm_log_stop; |
1166 | kml->listener.log_sync = kvm_log_sync; |
1167 | kml->listener.log_clear = kvm_log_clear; |
1168 | kml->listener.priority = 10; |
1169 | |
1170 | memory_listener_register(&kml->listener, as); |
1171 | |
1172 | for (i = 0; i < s->nr_as; ++i) { |
1173 | if (!s->as[i].as) { |
1174 | s->as[i].as = as; |
1175 | s->as[i].ml = kml; |
1176 | break; |
1177 | } |
1178 | } |
1179 | } |
1180 | |
1181 | static MemoryListener kvm_io_listener = { |
1182 | .eventfd_add = kvm_io_ioeventfd_add, |
1183 | .eventfd_del = kvm_io_ioeventfd_del, |
1184 | .priority = 10, |
1185 | }; |
1186 | |
1187 | int kvm_set_irq(KVMState *s, int irq, int level) |
1188 | { |
1189 | struct kvm_irq_level event; |
1190 | int ret; |
1191 | |
1192 | assert(kvm_async_interrupts_enabled()); |
1193 | |
1194 | event.level = level; |
1195 | event.irq = irq; |
1196 | ret = kvm_vm_ioctl(s, s->irq_set_ioctl, &event); |
1197 | if (ret < 0) { |
1198 | perror("kvm_set_irq" ); |
1199 | abort(); |
1200 | } |
1201 | |
1202 | return (s->irq_set_ioctl == KVM_IRQ_LINE) ? 1 : event.status; |
1203 | } |
1204 | |
1205 | #ifdef KVM_CAP_IRQ_ROUTING |
1206 | typedef struct KVMMSIRoute { |
1207 | struct kvm_irq_routing_entry kroute; |
1208 | QTAILQ_ENTRY(KVMMSIRoute) entry; |
1209 | } KVMMSIRoute; |
1210 | |
1211 | static void set_gsi(KVMState *s, unsigned int gsi) |
1212 | { |
1213 | set_bit(gsi, s->used_gsi_bitmap); |
1214 | } |
1215 | |
1216 | static void clear_gsi(KVMState *s, unsigned int gsi) |
1217 | { |
1218 | clear_bit(gsi, s->used_gsi_bitmap); |
1219 | } |
1220 | |
1221 | void kvm_init_irq_routing(KVMState *s) |
1222 | { |
1223 | int gsi_count, i; |
1224 | |
1225 | gsi_count = kvm_check_extension(s, KVM_CAP_IRQ_ROUTING) - 1; |
1226 | if (gsi_count > 0) { |
1227 | /* Round up so we can search ints using ffs */ |
1228 | s->used_gsi_bitmap = bitmap_new(gsi_count); |
1229 | s->gsi_count = gsi_count; |
1230 | } |
1231 | |
1232 | s->irq_routes = g_malloc0(sizeof(*s->irq_routes)); |
1233 | s->nr_allocated_irq_routes = 0; |
1234 | |
1235 | if (!kvm_direct_msi_allowed) { |
1236 | for (i = 0; i < KVM_MSI_HASHTAB_SIZE; i++) { |
1237 | QTAILQ_INIT(&s->msi_hashtab[i]); |
1238 | } |
1239 | } |
1240 | |
1241 | kvm_arch_init_irq_routing(s); |
1242 | } |
1243 | |
1244 | void kvm_irqchip_commit_routes(KVMState *s) |
1245 | { |
1246 | int ret; |
1247 | |
1248 | if (kvm_gsi_direct_mapping()) { |
1249 | return; |
1250 | } |
1251 | |
1252 | if (!kvm_gsi_routing_enabled()) { |
1253 | return; |
1254 | } |
1255 | |
1256 | s->irq_routes->flags = 0; |
1257 | trace_kvm_irqchip_commit_routes(); |
1258 | ret = kvm_vm_ioctl(s, KVM_SET_GSI_ROUTING, s->irq_routes); |
1259 | assert(ret == 0); |
1260 | } |
1261 | |
1262 | static void kvm_add_routing_entry(KVMState *s, |
1263 | struct kvm_irq_routing_entry *entry) |
1264 | { |
1265 | struct kvm_irq_routing_entry *new; |
1266 | int n, size; |
1267 | |
1268 | if (s->irq_routes->nr == s->nr_allocated_irq_routes) { |
1269 | n = s->nr_allocated_irq_routes * 2; |
1270 | if (n < 64) { |
1271 | n = 64; |
1272 | } |
1273 | size = sizeof(struct kvm_irq_routing); |
1274 | size += n * sizeof(*new); |
1275 | s->irq_routes = g_realloc(s->irq_routes, size); |
1276 | s->nr_allocated_irq_routes = n; |
1277 | } |
1278 | n = s->irq_routes->nr++; |
1279 | new = &s->irq_routes->entries[n]; |
1280 | |
1281 | *new = *entry; |
1282 | |
1283 | set_gsi(s, entry->gsi); |
1284 | } |
1285 | |
1286 | static int kvm_update_routing_entry(KVMState *s, |
1287 | struct kvm_irq_routing_entry *new_entry) |
1288 | { |
1289 | struct kvm_irq_routing_entry *entry; |
1290 | int n; |
1291 | |
1292 | for (n = 0; n < s->irq_routes->nr; n++) { |
1293 | entry = &s->irq_routes->entries[n]; |
1294 | if (entry->gsi != new_entry->gsi) { |
1295 | continue; |
1296 | } |
1297 | |
1298 | if(!memcmp(entry, new_entry, sizeof *entry)) { |
1299 | return 0; |
1300 | } |
1301 | |
1302 | *entry = *new_entry; |
1303 | |
1304 | return 0; |
1305 | } |
1306 | |
1307 | return -ESRCH; |
1308 | } |
1309 | |
1310 | void kvm_irqchip_add_irq_route(KVMState *s, int irq, int irqchip, int pin) |
1311 | { |
1312 | struct kvm_irq_routing_entry e = {}; |
1313 | |
1314 | assert(pin < s->gsi_count); |
1315 | |
1316 | e.gsi = irq; |
1317 | e.type = KVM_IRQ_ROUTING_IRQCHIP; |
1318 | e.flags = 0; |
1319 | e.u.irqchip.irqchip = irqchip; |
1320 | e.u.irqchip.pin = pin; |
1321 | kvm_add_routing_entry(s, &e); |
1322 | } |
1323 | |
1324 | void kvm_irqchip_release_virq(KVMState *s, int virq) |
1325 | { |
1326 | struct kvm_irq_routing_entry *e; |
1327 | int i; |
1328 | |
1329 | if (kvm_gsi_direct_mapping()) { |
1330 | return; |
1331 | } |
1332 | |
1333 | for (i = 0; i < s->irq_routes->nr; i++) { |
1334 | e = &s->irq_routes->entries[i]; |
1335 | if (e->gsi == virq) { |
1336 | s->irq_routes->nr--; |
1337 | *e = s->irq_routes->entries[s->irq_routes->nr]; |
1338 | } |
1339 | } |
1340 | clear_gsi(s, virq); |
1341 | kvm_arch_release_virq_post(virq); |
1342 | trace_kvm_irqchip_release_virq(virq); |
1343 | } |
1344 | |
1345 | static unsigned int kvm_hash_msi(uint32_t data) |
1346 | { |
1347 | /* This is optimized for IA32 MSI layout. However, no other arch shall |
1348 | * repeat the mistake of not providing a direct MSI injection API. */ |
1349 | return data & 0xff; |
1350 | } |
1351 | |
1352 | static void kvm_flush_dynamic_msi_routes(KVMState *s) |
1353 | { |
1354 | KVMMSIRoute *route, *next; |
1355 | unsigned int hash; |
1356 | |
1357 | for (hash = 0; hash < KVM_MSI_HASHTAB_SIZE; hash++) { |
1358 | QTAILQ_FOREACH_SAFE(route, &s->msi_hashtab[hash], entry, next) { |
1359 | kvm_irqchip_release_virq(s, route->kroute.gsi); |
1360 | QTAILQ_REMOVE(&s->msi_hashtab[hash], route, entry); |
1361 | g_free(route); |
1362 | } |
1363 | } |
1364 | } |
1365 | |
1366 | static int kvm_irqchip_get_virq(KVMState *s) |
1367 | { |
1368 | int next_virq; |
1369 | |
1370 | /* |
1371 | * PIC and IOAPIC share the first 16 GSI numbers, thus the available |
1372 | * GSI numbers are more than the number of IRQ route. Allocating a GSI |
1373 | * number can succeed even though a new route entry cannot be added. |
1374 | * When this happens, flush dynamic MSI entries to free IRQ route entries. |
1375 | */ |
1376 | if (!kvm_direct_msi_allowed && s->irq_routes->nr == s->gsi_count) { |
1377 | kvm_flush_dynamic_msi_routes(s); |
1378 | } |
1379 | |
1380 | /* Return the lowest unused GSI in the bitmap */ |
1381 | next_virq = find_first_zero_bit(s->used_gsi_bitmap, s->gsi_count); |
1382 | if (next_virq >= s->gsi_count) { |
1383 | return -ENOSPC; |
1384 | } else { |
1385 | return next_virq; |
1386 | } |
1387 | } |
1388 | |
1389 | static KVMMSIRoute *kvm_lookup_msi_route(KVMState *s, MSIMessage msg) |
1390 | { |
1391 | unsigned int hash = kvm_hash_msi(msg.data); |
1392 | KVMMSIRoute *route; |
1393 | |
1394 | QTAILQ_FOREACH(route, &s->msi_hashtab[hash], entry) { |
1395 | if (route->kroute.u.msi.address_lo == (uint32_t)msg.address && |
1396 | route->kroute.u.msi.address_hi == (msg.address >> 32) && |
1397 | route->kroute.u.msi.data == le32_to_cpu(msg.data)) { |
1398 | return route; |
1399 | } |
1400 | } |
1401 | return NULL; |
1402 | } |
1403 | |
1404 | int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg) |
1405 | { |
1406 | struct kvm_msi msi; |
1407 | KVMMSIRoute *route; |
1408 | |
1409 | if (kvm_direct_msi_allowed) { |
1410 | msi.address_lo = (uint32_t)msg.address; |
1411 | msi.address_hi = msg.address >> 32; |
1412 | msi.data = le32_to_cpu(msg.data); |
1413 | msi.flags = 0; |
1414 | memset(msi.pad, 0, sizeof(msi.pad)); |
1415 | |
1416 | return kvm_vm_ioctl(s, KVM_SIGNAL_MSI, &msi); |
1417 | } |
1418 | |
1419 | route = kvm_lookup_msi_route(s, msg); |
1420 | if (!route) { |
1421 | int virq; |
1422 | |
1423 | virq = kvm_irqchip_get_virq(s); |
1424 | if (virq < 0) { |
1425 | return virq; |
1426 | } |
1427 | |
1428 | route = g_malloc0(sizeof(KVMMSIRoute)); |
1429 | route->kroute.gsi = virq; |
1430 | route->kroute.type = KVM_IRQ_ROUTING_MSI; |
1431 | route->kroute.flags = 0; |
1432 | route->kroute.u.msi.address_lo = (uint32_t)msg.address; |
1433 | route->kroute.u.msi.address_hi = msg.address >> 32; |
1434 | route->kroute.u.msi.data = le32_to_cpu(msg.data); |
1435 | |
1436 | kvm_add_routing_entry(s, &route->kroute); |
1437 | kvm_irqchip_commit_routes(s); |
1438 | |
1439 | QTAILQ_INSERT_TAIL(&s->msi_hashtab[kvm_hash_msi(msg.data)], route, |
1440 | entry); |
1441 | } |
1442 | |
1443 | assert(route->kroute.type == KVM_IRQ_ROUTING_MSI); |
1444 | |
1445 | return kvm_set_irq(s, route->kroute.gsi, 1); |
1446 | } |
1447 | |
1448 | int kvm_irqchip_add_msi_route(KVMState *s, int vector, PCIDevice *dev) |
1449 | { |
1450 | struct kvm_irq_routing_entry kroute = {}; |
1451 | int virq; |
1452 | MSIMessage msg = {0, 0}; |
1453 | |
1454 | if (pci_available && dev) { |
1455 | msg = pci_get_msi_message(dev, vector); |
1456 | } |
1457 | |
1458 | if (kvm_gsi_direct_mapping()) { |
1459 | return kvm_arch_msi_data_to_gsi(msg.data); |
1460 | } |
1461 | |
1462 | if (!kvm_gsi_routing_enabled()) { |
1463 | return -ENOSYS; |
1464 | } |
1465 | |
1466 | virq = kvm_irqchip_get_virq(s); |
1467 | if (virq < 0) { |
1468 | return virq; |
1469 | } |
1470 | |
1471 | kroute.gsi = virq; |
1472 | kroute.type = KVM_IRQ_ROUTING_MSI; |
1473 | kroute.flags = 0; |
1474 | kroute.u.msi.address_lo = (uint32_t)msg.address; |
1475 | kroute.u.msi.address_hi = msg.address >> 32; |
1476 | kroute.u.msi.data = le32_to_cpu(msg.data); |
1477 | if (pci_available && kvm_msi_devid_required()) { |
1478 | kroute.flags = KVM_MSI_VALID_DEVID; |
1479 | kroute.u.msi.devid = pci_requester_id(dev); |
1480 | } |
1481 | if (kvm_arch_fixup_msi_route(&kroute, msg.address, msg.data, dev)) { |
1482 | kvm_irqchip_release_virq(s, virq); |
1483 | return -EINVAL; |
1484 | } |
1485 | |
1486 | trace_kvm_irqchip_add_msi_route(dev ? dev->name : (char *)"N/A" , |
1487 | vector, virq); |
1488 | |
1489 | kvm_add_routing_entry(s, &kroute); |
1490 | kvm_arch_add_msi_route_post(&kroute, vector, dev); |
1491 | kvm_irqchip_commit_routes(s); |
1492 | |
1493 | return virq; |
1494 | } |
1495 | |
1496 | int kvm_irqchip_update_msi_route(KVMState *s, int virq, MSIMessage msg, |
1497 | PCIDevice *dev) |
1498 | { |
1499 | struct kvm_irq_routing_entry kroute = {}; |
1500 | |
1501 | if (kvm_gsi_direct_mapping()) { |
1502 | return 0; |
1503 | } |
1504 | |
1505 | if (!kvm_irqchip_in_kernel()) { |
1506 | return -ENOSYS; |
1507 | } |
1508 | |
1509 | kroute.gsi = virq; |
1510 | kroute.type = KVM_IRQ_ROUTING_MSI; |
1511 | kroute.flags = 0; |
1512 | kroute.u.msi.address_lo = (uint32_t)msg.address; |
1513 | kroute.u.msi.address_hi = msg.address >> 32; |
1514 | kroute.u.msi.data = le32_to_cpu(msg.data); |
1515 | if (pci_available && kvm_msi_devid_required()) { |
1516 | kroute.flags = KVM_MSI_VALID_DEVID; |
1517 | kroute.u.msi.devid = pci_requester_id(dev); |
1518 | } |
1519 | if (kvm_arch_fixup_msi_route(&kroute, msg.address, msg.data, dev)) { |
1520 | return -EINVAL; |
1521 | } |
1522 | |
1523 | trace_kvm_irqchip_update_msi_route(virq); |
1524 | |
1525 | return kvm_update_routing_entry(s, &kroute); |
1526 | } |
1527 | |
1528 | static int kvm_irqchip_assign_irqfd(KVMState *s, int fd, int rfd, int virq, |
1529 | bool assign) |
1530 | { |
1531 | struct kvm_irqfd irqfd = { |
1532 | .fd = fd, |
1533 | .gsi = virq, |
1534 | .flags = assign ? 0 : KVM_IRQFD_FLAG_DEASSIGN, |
1535 | }; |
1536 | |
1537 | if (rfd != -1) { |
1538 | irqfd.flags |= KVM_IRQFD_FLAG_RESAMPLE; |
1539 | irqfd.resamplefd = rfd; |
1540 | } |
1541 | |
1542 | if (!kvm_irqfds_enabled()) { |
1543 | return -ENOSYS; |
1544 | } |
1545 | |
1546 | return kvm_vm_ioctl(s, KVM_IRQFD, &irqfd); |
1547 | } |
1548 | |
1549 | int kvm_irqchip_add_adapter_route(KVMState *s, AdapterInfo *adapter) |
1550 | { |
1551 | struct kvm_irq_routing_entry kroute = {}; |
1552 | int virq; |
1553 | |
1554 | if (!kvm_gsi_routing_enabled()) { |
1555 | return -ENOSYS; |
1556 | } |
1557 | |
1558 | virq = kvm_irqchip_get_virq(s); |
1559 | if (virq < 0) { |
1560 | return virq; |
1561 | } |
1562 | |
1563 | kroute.gsi = virq; |
1564 | kroute.type = KVM_IRQ_ROUTING_S390_ADAPTER; |
1565 | kroute.flags = 0; |
1566 | kroute.u.adapter.summary_addr = adapter->summary_addr; |
1567 | kroute.u.adapter.ind_addr = adapter->ind_addr; |
1568 | kroute.u.adapter.summary_offset = adapter->summary_offset; |
1569 | kroute.u.adapter.ind_offset = adapter->ind_offset; |
1570 | kroute.u.adapter.adapter_id = adapter->adapter_id; |
1571 | |
1572 | kvm_add_routing_entry(s, &kroute); |
1573 | |
1574 | return virq; |
1575 | } |
1576 | |
1577 | int kvm_irqchip_add_hv_sint_route(KVMState *s, uint32_t vcpu, uint32_t sint) |
1578 | { |
1579 | struct kvm_irq_routing_entry kroute = {}; |
1580 | int virq; |
1581 | |
1582 | if (!kvm_gsi_routing_enabled()) { |
1583 | return -ENOSYS; |
1584 | } |
1585 | if (!kvm_check_extension(s, KVM_CAP_HYPERV_SYNIC)) { |
1586 | return -ENOSYS; |
1587 | } |
1588 | virq = kvm_irqchip_get_virq(s); |
1589 | if (virq < 0) { |
1590 | return virq; |
1591 | } |
1592 | |
1593 | kroute.gsi = virq; |
1594 | kroute.type = KVM_IRQ_ROUTING_HV_SINT; |
1595 | kroute.flags = 0; |
1596 | kroute.u.hv_sint.vcpu = vcpu; |
1597 | kroute.u.hv_sint.sint = sint; |
1598 | |
1599 | kvm_add_routing_entry(s, &kroute); |
1600 | kvm_irqchip_commit_routes(s); |
1601 | |
1602 | return virq; |
1603 | } |
1604 | |
1605 | #else /* !KVM_CAP_IRQ_ROUTING */ |
1606 | |
1607 | void kvm_init_irq_routing(KVMState *s) |
1608 | { |
1609 | } |
1610 | |
1611 | void kvm_irqchip_release_virq(KVMState *s, int virq) |
1612 | { |
1613 | } |
1614 | |
1615 | int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg) |
1616 | { |
1617 | abort(); |
1618 | } |
1619 | |
1620 | int kvm_irqchip_add_msi_route(KVMState *s, int vector, PCIDevice *dev) |
1621 | { |
1622 | return -ENOSYS; |
1623 | } |
1624 | |
1625 | int kvm_irqchip_add_adapter_route(KVMState *s, AdapterInfo *adapter) |
1626 | { |
1627 | return -ENOSYS; |
1628 | } |
1629 | |
1630 | int kvm_irqchip_add_hv_sint_route(KVMState *s, uint32_t vcpu, uint32_t sint) |
1631 | { |
1632 | return -ENOSYS; |
1633 | } |
1634 | |
1635 | static int kvm_irqchip_assign_irqfd(KVMState *s, int fd, int virq, bool assign) |
1636 | { |
1637 | abort(); |
1638 | } |
1639 | |
1640 | int kvm_irqchip_update_msi_route(KVMState *s, int virq, MSIMessage msg) |
1641 | { |
1642 | return -ENOSYS; |
1643 | } |
1644 | #endif /* !KVM_CAP_IRQ_ROUTING */ |
1645 | |
1646 | int kvm_irqchip_add_irqfd_notifier_gsi(KVMState *s, EventNotifier *n, |
1647 | EventNotifier *rn, int virq) |
1648 | { |
1649 | return kvm_irqchip_assign_irqfd(s, event_notifier_get_fd(n), |
1650 | rn ? event_notifier_get_fd(rn) : -1, virq, true); |
1651 | } |
1652 | |
1653 | int kvm_irqchip_remove_irqfd_notifier_gsi(KVMState *s, EventNotifier *n, |
1654 | int virq) |
1655 | { |
1656 | return kvm_irqchip_assign_irqfd(s, event_notifier_get_fd(n), -1, virq, |
1657 | false); |
1658 | } |
1659 | |
1660 | int kvm_irqchip_add_irqfd_notifier(KVMState *s, EventNotifier *n, |
1661 | EventNotifier *rn, qemu_irq irq) |
1662 | { |
1663 | gpointer key, gsi; |
1664 | gboolean found = g_hash_table_lookup_extended(s->gsimap, irq, &key, &gsi); |
1665 | |
1666 | if (!found) { |
1667 | return -ENXIO; |
1668 | } |
1669 | return kvm_irqchip_add_irqfd_notifier_gsi(s, n, rn, GPOINTER_TO_INT(gsi)); |
1670 | } |
1671 | |
1672 | int kvm_irqchip_remove_irqfd_notifier(KVMState *s, EventNotifier *n, |
1673 | qemu_irq irq) |
1674 | { |
1675 | gpointer key, gsi; |
1676 | gboolean found = g_hash_table_lookup_extended(s->gsimap, irq, &key, &gsi); |
1677 | |
1678 | if (!found) { |
1679 | return -ENXIO; |
1680 | } |
1681 | return kvm_irqchip_remove_irqfd_notifier_gsi(s, n, GPOINTER_TO_INT(gsi)); |
1682 | } |
1683 | |
1684 | void kvm_irqchip_set_qemuirq_gsi(KVMState *s, qemu_irq irq, int gsi) |
1685 | { |
1686 | g_hash_table_insert(s->gsimap, irq, GINT_TO_POINTER(gsi)); |
1687 | } |
1688 | |
1689 | static void kvm_irqchip_create(MachineState *machine, KVMState *s) |
1690 | { |
1691 | int ret; |
1692 | |
1693 | if (kvm_check_extension(s, KVM_CAP_IRQCHIP)) { |
1694 | ; |
1695 | } else if (kvm_check_extension(s, KVM_CAP_S390_IRQCHIP)) { |
1696 | ret = kvm_vm_enable_cap(s, KVM_CAP_S390_IRQCHIP, 0); |
1697 | if (ret < 0) { |
1698 | fprintf(stderr, "Enable kernel irqchip failed: %s\n" , strerror(-ret)); |
1699 | exit(1); |
1700 | } |
1701 | } else { |
1702 | return; |
1703 | } |
1704 | |
1705 | /* First probe and see if there's a arch-specific hook to create the |
1706 | * in-kernel irqchip for us */ |
1707 | ret = kvm_arch_irqchip_create(machine, s); |
1708 | if (ret == 0) { |
1709 | if (machine_kernel_irqchip_split(machine)) { |
1710 | perror("Split IRQ chip mode not supported." ); |
1711 | exit(1); |
1712 | } else { |
1713 | ret = kvm_vm_ioctl(s, KVM_CREATE_IRQCHIP); |
1714 | } |
1715 | } |
1716 | if (ret < 0) { |
1717 | fprintf(stderr, "Create kernel irqchip failed: %s\n" , strerror(-ret)); |
1718 | exit(1); |
1719 | } |
1720 | |
1721 | kvm_kernel_irqchip = true; |
1722 | /* If we have an in-kernel IRQ chip then we must have asynchronous |
1723 | * interrupt delivery (though the reverse is not necessarily true) |
1724 | */ |
1725 | kvm_async_interrupts_allowed = true; |
1726 | kvm_halt_in_kernel_allowed = true; |
1727 | |
1728 | kvm_init_irq_routing(s); |
1729 | |
1730 | s->gsimap = g_hash_table_new(g_direct_hash, g_direct_equal); |
1731 | } |
1732 | |
1733 | /* Find number of supported CPUs using the recommended |
1734 | * procedure from the kernel API documentation to cope with |
1735 | * older kernels that may be missing capabilities. |
1736 | */ |
1737 | static int kvm_recommended_vcpus(KVMState *s) |
1738 | { |
1739 | int ret = kvm_vm_check_extension(s, KVM_CAP_NR_VCPUS); |
1740 | return (ret) ? ret : 4; |
1741 | } |
1742 | |
1743 | static int kvm_max_vcpus(KVMState *s) |
1744 | { |
1745 | int ret = kvm_check_extension(s, KVM_CAP_MAX_VCPUS); |
1746 | return (ret) ? ret : kvm_recommended_vcpus(s); |
1747 | } |
1748 | |
1749 | static int kvm_max_vcpu_id(KVMState *s) |
1750 | { |
1751 | int ret = kvm_check_extension(s, KVM_CAP_MAX_VCPU_ID); |
1752 | return (ret) ? ret : kvm_max_vcpus(s); |
1753 | } |
1754 | |
1755 | bool kvm_vcpu_id_is_valid(int vcpu_id) |
1756 | { |
1757 | KVMState *s = KVM_STATE(current_machine->accelerator); |
1758 | return vcpu_id >= 0 && vcpu_id < kvm_max_vcpu_id(s); |
1759 | } |
1760 | |
1761 | static int kvm_init(MachineState *ms) |
1762 | { |
1763 | MachineClass *mc = MACHINE_GET_CLASS(ms); |
1764 | static const char upgrade_note[] = |
1765 | "Please upgrade to at least kernel 2.6.29 or recent kvm-kmod\n" |
1766 | "(see http://sourceforge.net/projects/kvm).\n" ; |
1767 | struct { |
1768 | const char *name; |
1769 | int num; |
1770 | } num_cpus[] = { |
1771 | { "SMP" , ms->smp.cpus }, |
1772 | { "hotpluggable" , ms->smp.max_cpus }, |
1773 | { NULL, } |
1774 | }, *nc = num_cpus; |
1775 | int soft_vcpus_limit, hard_vcpus_limit; |
1776 | KVMState *s; |
1777 | const KVMCapabilityInfo *missing_cap; |
1778 | int ret; |
1779 | int type = 0; |
1780 | const char *kvm_type; |
1781 | |
1782 | s = KVM_STATE(ms->accelerator); |
1783 | |
1784 | /* |
1785 | * On systems where the kernel can support different base page |
1786 | * sizes, host page size may be different from TARGET_PAGE_SIZE, |
1787 | * even with KVM. TARGET_PAGE_SIZE is assumed to be the minimum |
1788 | * page size for the system though. |
1789 | */ |
1790 | assert(TARGET_PAGE_SIZE <= getpagesize()); |
1791 | |
1792 | s->sigmask_len = 8; |
1793 | |
1794 | #ifdef KVM_CAP_SET_GUEST_DEBUG |
1795 | QTAILQ_INIT(&s->kvm_sw_breakpoints); |
1796 | #endif |
1797 | QLIST_INIT(&s->kvm_parked_vcpus); |
1798 | s->vmfd = -1; |
1799 | s->fd = qemu_open("/dev/kvm" , O_RDWR); |
1800 | if (s->fd == -1) { |
1801 | fprintf(stderr, "Could not access KVM kernel module: %m\n" ); |
1802 | ret = -errno; |
1803 | goto err; |
1804 | } |
1805 | |
1806 | ret = kvm_ioctl(s, KVM_GET_API_VERSION, 0); |
1807 | if (ret < KVM_API_VERSION) { |
1808 | if (ret >= 0) { |
1809 | ret = -EINVAL; |
1810 | } |
1811 | fprintf(stderr, "kvm version too old\n" ); |
1812 | goto err; |
1813 | } |
1814 | |
1815 | if (ret > KVM_API_VERSION) { |
1816 | ret = -EINVAL; |
1817 | fprintf(stderr, "kvm version not supported\n" ); |
1818 | goto err; |
1819 | } |
1820 | |
1821 | kvm_immediate_exit = kvm_check_extension(s, KVM_CAP_IMMEDIATE_EXIT); |
1822 | s->nr_slots = kvm_check_extension(s, KVM_CAP_NR_MEMSLOTS); |
1823 | |
1824 | /* If unspecified, use the default value */ |
1825 | if (!s->nr_slots) { |
1826 | s->nr_slots = 32; |
1827 | } |
1828 | |
1829 | s->nr_as = kvm_check_extension(s, KVM_CAP_MULTI_ADDRESS_SPACE); |
1830 | if (s->nr_as <= 1) { |
1831 | s->nr_as = 1; |
1832 | } |
1833 | s->as = g_new0(struct KVMAs, s->nr_as); |
1834 | |
1835 | kvm_type = qemu_opt_get(qemu_get_machine_opts(), "kvm-type" ); |
1836 | if (mc->kvm_type) { |
1837 | type = mc->kvm_type(ms, kvm_type); |
1838 | } else if (kvm_type) { |
1839 | ret = -EINVAL; |
1840 | fprintf(stderr, "Invalid argument kvm-type=%s\n" , kvm_type); |
1841 | goto err; |
1842 | } |
1843 | |
1844 | do { |
1845 | ret = kvm_ioctl(s, KVM_CREATE_VM, type); |
1846 | } while (ret == -EINTR); |
1847 | |
1848 | if (ret < 0) { |
1849 | fprintf(stderr, "ioctl(KVM_CREATE_VM) failed: %d %s\n" , -ret, |
1850 | strerror(-ret)); |
1851 | |
1852 | #ifdef TARGET_S390X |
1853 | if (ret == -EINVAL) { |
1854 | fprintf(stderr, |
1855 | "Host kernel setup problem detected. Please verify:\n" ); |
1856 | fprintf(stderr, "- for kernels supporting the switch_amode or" |
1857 | " user_mode parameters, whether\n" ); |
1858 | fprintf(stderr, |
1859 | " user space is running in primary address space\n" ); |
1860 | fprintf(stderr, |
1861 | "- for kernels supporting the vm.allocate_pgste sysctl, " |
1862 | "whether it is enabled\n" ); |
1863 | } |
1864 | #endif |
1865 | goto err; |
1866 | } |
1867 | |
1868 | s->vmfd = ret; |
1869 | |
1870 | /* check the vcpu limits */ |
1871 | soft_vcpus_limit = kvm_recommended_vcpus(s); |
1872 | hard_vcpus_limit = kvm_max_vcpus(s); |
1873 | |
1874 | while (nc->name) { |
1875 | if (nc->num > soft_vcpus_limit) { |
1876 | warn_report("Number of %s cpus requested (%d) exceeds " |
1877 | "the recommended cpus supported by KVM (%d)" , |
1878 | nc->name, nc->num, soft_vcpus_limit); |
1879 | |
1880 | if (nc->num > hard_vcpus_limit) { |
1881 | fprintf(stderr, "Number of %s cpus requested (%d) exceeds " |
1882 | "the maximum cpus supported by KVM (%d)\n" , |
1883 | nc->name, nc->num, hard_vcpus_limit); |
1884 | exit(1); |
1885 | } |
1886 | } |
1887 | nc++; |
1888 | } |
1889 | |
1890 | missing_cap = kvm_check_extension_list(s, kvm_required_capabilites); |
1891 | if (!missing_cap) { |
1892 | missing_cap = |
1893 | kvm_check_extension_list(s, kvm_arch_required_capabilities); |
1894 | } |
1895 | if (missing_cap) { |
1896 | ret = -EINVAL; |
1897 | fprintf(stderr, "kvm does not support %s\n%s" , |
1898 | missing_cap->name, upgrade_note); |
1899 | goto err; |
1900 | } |
1901 | |
1902 | s->coalesced_mmio = kvm_check_extension(s, KVM_CAP_COALESCED_MMIO); |
1903 | s->coalesced_pio = s->coalesced_mmio && |
1904 | kvm_check_extension(s, KVM_CAP_COALESCED_PIO); |
1905 | |
1906 | s->manual_dirty_log_protect = |
1907 | kvm_check_extension(s, KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2); |
1908 | if (s->manual_dirty_log_protect) { |
1909 | ret = kvm_vm_enable_cap(s, KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2, 0, 1); |
1910 | if (ret) { |
1911 | warn_report("Trying to enable KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2 " |
1912 | "but failed. Falling back to the legacy mode. " ); |
1913 | s->manual_dirty_log_protect = false; |
1914 | } |
1915 | } |
1916 | |
1917 | #ifdef KVM_CAP_VCPU_EVENTS |
1918 | s->vcpu_events = kvm_check_extension(s, KVM_CAP_VCPU_EVENTS); |
1919 | #endif |
1920 | |
1921 | s->robust_singlestep = |
1922 | kvm_check_extension(s, KVM_CAP_X86_ROBUST_SINGLESTEP); |
1923 | |
1924 | #ifdef KVM_CAP_DEBUGREGS |
1925 | s->debugregs = kvm_check_extension(s, KVM_CAP_DEBUGREGS); |
1926 | #endif |
1927 | |
1928 | s->max_nested_state_len = kvm_check_extension(s, KVM_CAP_NESTED_STATE); |
1929 | |
1930 | #ifdef KVM_CAP_IRQ_ROUTING |
1931 | kvm_direct_msi_allowed = (kvm_check_extension(s, KVM_CAP_SIGNAL_MSI) > 0); |
1932 | #endif |
1933 | |
1934 | s->intx_set_mask = kvm_check_extension(s, KVM_CAP_PCI_2_3); |
1935 | |
1936 | s->irq_set_ioctl = KVM_IRQ_LINE; |
1937 | if (kvm_check_extension(s, KVM_CAP_IRQ_INJECT_STATUS)) { |
1938 | s->irq_set_ioctl = KVM_IRQ_LINE_STATUS; |
1939 | } |
1940 | |
1941 | kvm_readonly_mem_allowed = |
1942 | (kvm_check_extension(s, KVM_CAP_READONLY_MEM) > 0); |
1943 | |
1944 | kvm_eventfds_allowed = |
1945 | (kvm_check_extension(s, KVM_CAP_IOEVENTFD) > 0); |
1946 | |
1947 | kvm_irqfds_allowed = |
1948 | (kvm_check_extension(s, KVM_CAP_IRQFD) > 0); |
1949 | |
1950 | kvm_resamplefds_allowed = |
1951 | (kvm_check_extension(s, KVM_CAP_IRQFD_RESAMPLE) > 0); |
1952 | |
1953 | kvm_vm_attributes_allowed = |
1954 | (kvm_check_extension(s, KVM_CAP_VM_ATTRIBUTES) > 0); |
1955 | |
1956 | kvm_ioeventfd_any_length_allowed = |
1957 | (kvm_check_extension(s, KVM_CAP_IOEVENTFD_ANY_LENGTH) > 0); |
1958 | |
1959 | kvm_state = s; |
1960 | |
1961 | /* |
1962 | * if memory encryption object is specified then initialize the memory |
1963 | * encryption context. |
1964 | */ |
1965 | if (ms->memory_encryption) { |
1966 | kvm_state->memcrypt_handle = sev_guest_init(ms->memory_encryption); |
1967 | if (!kvm_state->memcrypt_handle) { |
1968 | ret = -1; |
1969 | goto err; |
1970 | } |
1971 | |
1972 | kvm_state->memcrypt_encrypt_data = sev_encrypt_data; |
1973 | } |
1974 | |
1975 | ret = kvm_arch_init(ms, s); |
1976 | if (ret < 0) { |
1977 | goto err; |
1978 | } |
1979 | |
1980 | if (machine_kernel_irqchip_allowed(ms)) { |
1981 | kvm_irqchip_create(ms, s); |
1982 | } |
1983 | |
1984 | if (kvm_eventfds_allowed) { |
1985 | s->memory_listener.listener.eventfd_add = kvm_mem_ioeventfd_add; |
1986 | s->memory_listener.listener.eventfd_del = kvm_mem_ioeventfd_del; |
1987 | } |
1988 | s->memory_listener.listener.coalesced_io_add = kvm_coalesce_mmio_region; |
1989 | s->memory_listener.listener.coalesced_io_del = kvm_uncoalesce_mmio_region; |
1990 | |
1991 | kvm_memory_listener_register(s, &s->memory_listener, |
1992 | &address_space_memory, 0); |
1993 | memory_listener_register(&kvm_io_listener, |
1994 | &address_space_io); |
1995 | memory_listener_register(&kvm_coalesced_pio_listener, |
1996 | &address_space_io); |
1997 | |
1998 | s->many_ioeventfds = kvm_check_many_ioeventfds(); |
1999 | |
2000 | s->sync_mmu = !!kvm_vm_check_extension(kvm_state, KVM_CAP_SYNC_MMU); |
2001 | if (!s->sync_mmu) { |
2002 | qemu_balloon_inhibit(true); |
2003 | } |
2004 | |
2005 | return 0; |
2006 | |
2007 | err: |
2008 | assert(ret < 0); |
2009 | if (s->vmfd >= 0) { |
2010 | close(s->vmfd); |
2011 | } |
2012 | if (s->fd != -1) { |
2013 | close(s->fd); |
2014 | } |
2015 | g_free(s->memory_listener.slots); |
2016 | |
2017 | return ret; |
2018 | } |
2019 | |
2020 | void kvm_set_sigmask_len(KVMState *s, unsigned int sigmask_len) |
2021 | { |
2022 | s->sigmask_len = sigmask_len; |
2023 | } |
2024 | |
2025 | static void kvm_handle_io(uint16_t port, MemTxAttrs attrs, void *data, int direction, |
2026 | int size, uint32_t count) |
2027 | { |
2028 | int i; |
2029 | uint8_t *ptr = data; |
2030 | |
2031 | for (i = 0; i < count; i++) { |
2032 | address_space_rw(&address_space_io, port, attrs, |
2033 | ptr, size, |
2034 | direction == KVM_EXIT_IO_OUT); |
2035 | ptr += size; |
2036 | } |
2037 | } |
2038 | |
2039 | static int kvm_handle_internal_error(CPUState *cpu, struct kvm_run *run) |
2040 | { |
2041 | fprintf(stderr, "KVM internal error. Suberror: %d\n" , |
2042 | run->internal.suberror); |
2043 | |
2044 | if (kvm_check_extension(kvm_state, KVM_CAP_INTERNAL_ERROR_DATA)) { |
2045 | int i; |
2046 | |
2047 | for (i = 0; i < run->internal.ndata; ++i) { |
2048 | fprintf(stderr, "extra data[%d]: %" PRIx64"\n" , |
2049 | i, (uint64_t)run->internal.data[i]); |
2050 | } |
2051 | } |
2052 | if (run->internal.suberror == KVM_INTERNAL_ERROR_EMULATION) { |
2053 | fprintf(stderr, "emulation failure\n" ); |
2054 | if (!kvm_arch_stop_on_emulation_error(cpu)) { |
2055 | cpu_dump_state(cpu, stderr, CPU_DUMP_CODE); |
2056 | return EXCP_INTERRUPT; |
2057 | } |
2058 | } |
2059 | /* FIXME: Should trigger a qmp message to let management know |
2060 | * something went wrong. |
2061 | */ |
2062 | return -1; |
2063 | } |
2064 | |
2065 | void kvm_flush_coalesced_mmio_buffer(void) |
2066 | { |
2067 | KVMState *s = kvm_state; |
2068 | |
2069 | if (s->coalesced_flush_in_progress) { |
2070 | return; |
2071 | } |
2072 | |
2073 | s->coalesced_flush_in_progress = true; |
2074 | |
2075 | if (s->coalesced_mmio_ring) { |
2076 | struct kvm_coalesced_mmio_ring *ring = s->coalesced_mmio_ring; |
2077 | while (ring->first != ring->last) { |
2078 | struct kvm_coalesced_mmio *ent; |
2079 | |
2080 | ent = &ring->coalesced_mmio[ring->first]; |
2081 | |
2082 | if (ent->pio == 1) { |
2083 | address_space_rw(&address_space_io, ent->phys_addr, |
2084 | MEMTXATTRS_UNSPECIFIED, ent->data, |
2085 | ent->len, true); |
2086 | } else { |
2087 | cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len); |
2088 | } |
2089 | smp_wmb(); |
2090 | ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX; |
2091 | } |
2092 | } |
2093 | |
2094 | s->coalesced_flush_in_progress = false; |
2095 | } |
2096 | |
2097 | static void do_kvm_cpu_synchronize_state(CPUState *cpu, run_on_cpu_data arg) |
2098 | { |
2099 | if (!cpu->vcpu_dirty) { |
2100 | kvm_arch_get_registers(cpu); |
2101 | cpu->vcpu_dirty = true; |
2102 | } |
2103 | } |
2104 | |
2105 | void kvm_cpu_synchronize_state(CPUState *cpu) |
2106 | { |
2107 | if (!cpu->vcpu_dirty) { |
2108 | run_on_cpu(cpu, do_kvm_cpu_synchronize_state, RUN_ON_CPU_NULL); |
2109 | } |
2110 | } |
2111 | |
2112 | static void do_kvm_cpu_synchronize_post_reset(CPUState *cpu, run_on_cpu_data arg) |
2113 | { |
2114 | kvm_arch_put_registers(cpu, KVM_PUT_RESET_STATE); |
2115 | cpu->vcpu_dirty = false; |
2116 | } |
2117 | |
2118 | void kvm_cpu_synchronize_post_reset(CPUState *cpu) |
2119 | { |
2120 | run_on_cpu(cpu, do_kvm_cpu_synchronize_post_reset, RUN_ON_CPU_NULL); |
2121 | } |
2122 | |
2123 | static void do_kvm_cpu_synchronize_post_init(CPUState *cpu, run_on_cpu_data arg) |
2124 | { |
2125 | kvm_arch_put_registers(cpu, KVM_PUT_FULL_STATE); |
2126 | cpu->vcpu_dirty = false; |
2127 | } |
2128 | |
2129 | void kvm_cpu_synchronize_post_init(CPUState *cpu) |
2130 | { |
2131 | run_on_cpu(cpu, do_kvm_cpu_synchronize_post_init, RUN_ON_CPU_NULL); |
2132 | } |
2133 | |
2134 | static void do_kvm_cpu_synchronize_pre_loadvm(CPUState *cpu, run_on_cpu_data arg) |
2135 | { |
2136 | cpu->vcpu_dirty = true; |
2137 | } |
2138 | |
2139 | void kvm_cpu_synchronize_pre_loadvm(CPUState *cpu) |
2140 | { |
2141 | run_on_cpu(cpu, do_kvm_cpu_synchronize_pre_loadvm, RUN_ON_CPU_NULL); |
2142 | } |
2143 | |
2144 | #ifdef KVM_HAVE_MCE_INJECTION |
2145 | static __thread void *pending_sigbus_addr; |
2146 | static __thread int pending_sigbus_code; |
2147 | static __thread bool have_sigbus_pending; |
2148 | #endif |
2149 | |
2150 | static void kvm_cpu_kick(CPUState *cpu) |
2151 | { |
2152 | atomic_set(&cpu->kvm_run->immediate_exit, 1); |
2153 | } |
2154 | |
2155 | static void kvm_cpu_kick_self(void) |
2156 | { |
2157 | if (kvm_immediate_exit) { |
2158 | kvm_cpu_kick(current_cpu); |
2159 | } else { |
2160 | qemu_cpu_kick_self(); |
2161 | } |
2162 | } |
2163 | |
2164 | static void kvm_eat_signals(CPUState *cpu) |
2165 | { |
2166 | struct timespec ts = { 0, 0 }; |
2167 | siginfo_t siginfo; |
2168 | sigset_t waitset; |
2169 | sigset_t chkset; |
2170 | int r; |
2171 | |
2172 | if (kvm_immediate_exit) { |
2173 | atomic_set(&cpu->kvm_run->immediate_exit, 0); |
2174 | /* Write kvm_run->immediate_exit before the cpu->exit_request |
2175 | * write in kvm_cpu_exec. |
2176 | */ |
2177 | smp_wmb(); |
2178 | return; |
2179 | } |
2180 | |
2181 | sigemptyset(&waitset); |
2182 | sigaddset(&waitset, SIG_IPI); |
2183 | |
2184 | do { |
2185 | r = sigtimedwait(&waitset, &siginfo, &ts); |
2186 | if (r == -1 && !(errno == EAGAIN || errno == EINTR)) { |
2187 | perror("sigtimedwait" ); |
2188 | exit(1); |
2189 | } |
2190 | |
2191 | r = sigpending(&chkset); |
2192 | if (r == -1) { |
2193 | perror("sigpending" ); |
2194 | exit(1); |
2195 | } |
2196 | } while (sigismember(&chkset, SIG_IPI)); |
2197 | } |
2198 | |
2199 | int kvm_cpu_exec(CPUState *cpu) |
2200 | { |
2201 | struct kvm_run *run = cpu->kvm_run; |
2202 | int ret, run_ret; |
2203 | |
2204 | DPRINTF("kvm_cpu_exec()\n" ); |
2205 | |
2206 | if (kvm_arch_process_async_events(cpu)) { |
2207 | atomic_set(&cpu->exit_request, 0); |
2208 | return EXCP_HLT; |
2209 | } |
2210 | |
2211 | qemu_mutex_unlock_iothread(); |
2212 | cpu_exec_start(cpu); |
2213 | |
2214 | do { |
2215 | MemTxAttrs attrs; |
2216 | |
2217 | if (cpu->vcpu_dirty) { |
2218 | kvm_arch_put_registers(cpu, KVM_PUT_RUNTIME_STATE); |
2219 | cpu->vcpu_dirty = false; |
2220 | } |
2221 | |
2222 | kvm_arch_pre_run(cpu, run); |
2223 | if (atomic_read(&cpu->exit_request)) { |
2224 | DPRINTF("interrupt exit requested\n" ); |
2225 | /* |
2226 | * KVM requires us to reenter the kernel after IO exits to complete |
2227 | * instruction emulation. This self-signal will ensure that we |
2228 | * leave ASAP again. |
2229 | */ |
2230 | kvm_cpu_kick_self(); |
2231 | } |
2232 | |
2233 | /* Read cpu->exit_request before KVM_RUN reads run->immediate_exit. |
2234 | * Matching barrier in kvm_eat_signals. |
2235 | */ |
2236 | smp_rmb(); |
2237 | |
2238 | run_ret = kvm_vcpu_ioctl(cpu, KVM_RUN, 0); |
2239 | |
2240 | attrs = kvm_arch_post_run(cpu, run); |
2241 | |
2242 | #ifdef KVM_HAVE_MCE_INJECTION |
2243 | if (unlikely(have_sigbus_pending)) { |
2244 | qemu_mutex_lock_iothread(); |
2245 | kvm_arch_on_sigbus_vcpu(cpu, pending_sigbus_code, |
2246 | pending_sigbus_addr); |
2247 | have_sigbus_pending = false; |
2248 | qemu_mutex_unlock_iothread(); |
2249 | } |
2250 | #endif |
2251 | |
2252 | if (run_ret < 0) { |
2253 | if (run_ret == -EINTR || run_ret == -EAGAIN) { |
2254 | DPRINTF("io window exit\n" ); |
2255 | kvm_eat_signals(cpu); |
2256 | ret = EXCP_INTERRUPT; |
2257 | break; |
2258 | } |
2259 | fprintf(stderr, "error: kvm run failed %s\n" , |
2260 | strerror(-run_ret)); |
2261 | #ifdef TARGET_PPC |
2262 | if (run_ret == -EBUSY) { |
2263 | fprintf(stderr, |
2264 | "This is probably because your SMT is enabled.\n" |
2265 | "VCPU can only run on primary threads with all " |
2266 | "secondary threads offline.\n" ); |
2267 | } |
2268 | #endif |
2269 | ret = -1; |
2270 | break; |
2271 | } |
2272 | |
2273 | trace_kvm_run_exit(cpu->cpu_index, run->exit_reason); |
2274 | switch (run->exit_reason) { |
2275 | case KVM_EXIT_IO: |
2276 | DPRINTF("handle_io\n" ); |
2277 | /* Called outside BQL */ |
2278 | kvm_handle_io(run->io.port, attrs, |
2279 | (uint8_t *)run + run->io.data_offset, |
2280 | run->io.direction, |
2281 | run->io.size, |
2282 | run->io.count); |
2283 | ret = 0; |
2284 | break; |
2285 | case KVM_EXIT_MMIO: |
2286 | DPRINTF("handle_mmio\n" ); |
2287 | /* Called outside BQL */ |
2288 | address_space_rw(&address_space_memory, |
2289 | run->mmio.phys_addr, attrs, |
2290 | run->mmio.data, |
2291 | run->mmio.len, |
2292 | run->mmio.is_write); |
2293 | ret = 0; |
2294 | break; |
2295 | case KVM_EXIT_IRQ_WINDOW_OPEN: |
2296 | DPRINTF("irq_window_open\n" ); |
2297 | ret = EXCP_INTERRUPT; |
2298 | break; |
2299 | case KVM_EXIT_SHUTDOWN: |
2300 | DPRINTF("shutdown\n" ); |
2301 | qemu_system_reset_request(SHUTDOWN_CAUSE_GUEST_RESET); |
2302 | ret = EXCP_INTERRUPT; |
2303 | break; |
2304 | case KVM_EXIT_UNKNOWN: |
2305 | fprintf(stderr, "KVM: unknown exit, hardware reason %" PRIx64 "\n" , |
2306 | (uint64_t)run->hw.hardware_exit_reason); |
2307 | ret = -1; |
2308 | break; |
2309 | case KVM_EXIT_INTERNAL_ERROR: |
2310 | ret = kvm_handle_internal_error(cpu, run); |
2311 | break; |
2312 | case KVM_EXIT_SYSTEM_EVENT: |
2313 | switch (run->system_event.type) { |
2314 | case KVM_SYSTEM_EVENT_SHUTDOWN: |
2315 | qemu_system_shutdown_request(SHUTDOWN_CAUSE_GUEST_SHUTDOWN); |
2316 | ret = EXCP_INTERRUPT; |
2317 | break; |
2318 | case KVM_SYSTEM_EVENT_RESET: |
2319 | qemu_system_reset_request(SHUTDOWN_CAUSE_GUEST_RESET); |
2320 | ret = EXCP_INTERRUPT; |
2321 | break; |
2322 | case KVM_SYSTEM_EVENT_CRASH: |
2323 | kvm_cpu_synchronize_state(cpu); |
2324 | qemu_mutex_lock_iothread(); |
2325 | qemu_system_guest_panicked(cpu_get_crash_info(cpu)); |
2326 | qemu_mutex_unlock_iothread(); |
2327 | ret = 0; |
2328 | break; |
2329 | default: |
2330 | DPRINTF("kvm_arch_handle_exit\n" ); |
2331 | ret = kvm_arch_handle_exit(cpu, run); |
2332 | break; |
2333 | } |
2334 | break; |
2335 | default: |
2336 | DPRINTF("kvm_arch_handle_exit\n" ); |
2337 | ret = kvm_arch_handle_exit(cpu, run); |
2338 | break; |
2339 | } |
2340 | } while (ret == 0); |
2341 | |
2342 | cpu_exec_end(cpu); |
2343 | qemu_mutex_lock_iothread(); |
2344 | |
2345 | if (ret < 0) { |
2346 | cpu_dump_state(cpu, stderr, CPU_DUMP_CODE); |
2347 | vm_stop(RUN_STATE_INTERNAL_ERROR); |
2348 | } |
2349 | |
2350 | atomic_set(&cpu->exit_request, 0); |
2351 | return ret; |
2352 | } |
2353 | |
2354 | int kvm_ioctl(KVMState *s, int type, ...) |
2355 | { |
2356 | int ret; |
2357 | void *arg; |
2358 | va_list ap; |
2359 | |
2360 | va_start(ap, type); |
2361 | arg = va_arg(ap, void *); |
2362 | va_end(ap); |
2363 | |
2364 | trace_kvm_ioctl(type, arg); |
2365 | ret = ioctl(s->fd, type, arg); |
2366 | if (ret == -1) { |
2367 | ret = -errno; |
2368 | } |
2369 | return ret; |
2370 | } |
2371 | |
2372 | int kvm_vm_ioctl(KVMState *s, int type, ...) |
2373 | { |
2374 | int ret; |
2375 | void *arg; |
2376 | va_list ap; |
2377 | |
2378 | va_start(ap, type); |
2379 | arg = va_arg(ap, void *); |
2380 | va_end(ap); |
2381 | |
2382 | trace_kvm_vm_ioctl(type, arg); |
2383 | ret = ioctl(s->vmfd, type, arg); |
2384 | if (ret == -1) { |
2385 | ret = -errno; |
2386 | } |
2387 | return ret; |
2388 | } |
2389 | |
2390 | int kvm_vcpu_ioctl(CPUState *cpu, int type, ...) |
2391 | { |
2392 | int ret; |
2393 | void *arg; |
2394 | va_list ap; |
2395 | |
2396 | va_start(ap, type); |
2397 | arg = va_arg(ap, void *); |
2398 | va_end(ap); |
2399 | |
2400 | trace_kvm_vcpu_ioctl(cpu->cpu_index, type, arg); |
2401 | ret = ioctl(cpu->kvm_fd, type, arg); |
2402 | if (ret == -1) { |
2403 | ret = -errno; |
2404 | } |
2405 | return ret; |
2406 | } |
2407 | |
2408 | int kvm_device_ioctl(int fd, int type, ...) |
2409 | { |
2410 | int ret; |
2411 | void *arg; |
2412 | va_list ap; |
2413 | |
2414 | va_start(ap, type); |
2415 | arg = va_arg(ap, void *); |
2416 | va_end(ap); |
2417 | |
2418 | trace_kvm_device_ioctl(fd, type, arg); |
2419 | ret = ioctl(fd, type, arg); |
2420 | if (ret == -1) { |
2421 | ret = -errno; |
2422 | } |
2423 | return ret; |
2424 | } |
2425 | |
2426 | int kvm_vm_check_attr(KVMState *s, uint32_t group, uint64_t attr) |
2427 | { |
2428 | int ret; |
2429 | struct kvm_device_attr attribute = { |
2430 | .group = group, |
2431 | .attr = attr, |
2432 | }; |
2433 | |
2434 | if (!kvm_vm_attributes_allowed) { |
2435 | return 0; |
2436 | } |
2437 | |
2438 | ret = kvm_vm_ioctl(s, KVM_HAS_DEVICE_ATTR, &attribute); |
2439 | /* kvm returns 0 on success for HAS_DEVICE_ATTR */ |
2440 | return ret ? 0 : 1; |
2441 | } |
2442 | |
2443 | int kvm_device_check_attr(int dev_fd, uint32_t group, uint64_t attr) |
2444 | { |
2445 | struct kvm_device_attr attribute = { |
2446 | .group = group, |
2447 | .attr = attr, |
2448 | .flags = 0, |
2449 | }; |
2450 | |
2451 | return kvm_device_ioctl(dev_fd, KVM_HAS_DEVICE_ATTR, &attribute) ? 0 : 1; |
2452 | } |
2453 | |
2454 | int kvm_device_access(int fd, int group, uint64_t attr, |
2455 | void *val, bool write, Error **errp) |
2456 | { |
2457 | struct kvm_device_attr kvmattr; |
2458 | int err; |
2459 | |
2460 | kvmattr.flags = 0; |
2461 | kvmattr.group = group; |
2462 | kvmattr.attr = attr; |
2463 | kvmattr.addr = (uintptr_t)val; |
2464 | |
2465 | err = kvm_device_ioctl(fd, |
2466 | write ? KVM_SET_DEVICE_ATTR : KVM_GET_DEVICE_ATTR, |
2467 | &kvmattr); |
2468 | if (err < 0) { |
2469 | error_setg_errno(errp, -err, |
2470 | "KVM_%s_DEVICE_ATTR failed: Group %d " |
2471 | "attr 0x%016" PRIx64, |
2472 | write ? "SET" : "GET" , group, attr); |
2473 | } |
2474 | return err; |
2475 | } |
2476 | |
2477 | bool kvm_has_sync_mmu(void) |
2478 | { |
2479 | return kvm_state->sync_mmu; |
2480 | } |
2481 | |
2482 | int kvm_has_vcpu_events(void) |
2483 | { |
2484 | return kvm_state->vcpu_events; |
2485 | } |
2486 | |
2487 | int kvm_has_robust_singlestep(void) |
2488 | { |
2489 | return kvm_state->robust_singlestep; |
2490 | } |
2491 | |
2492 | int kvm_has_debugregs(void) |
2493 | { |
2494 | return kvm_state->debugregs; |
2495 | } |
2496 | |
2497 | int kvm_max_nested_state_length(void) |
2498 | { |
2499 | return kvm_state->max_nested_state_len; |
2500 | } |
2501 | |
2502 | int kvm_has_many_ioeventfds(void) |
2503 | { |
2504 | if (!kvm_enabled()) { |
2505 | return 0; |
2506 | } |
2507 | return kvm_state->many_ioeventfds; |
2508 | } |
2509 | |
2510 | int kvm_has_gsi_routing(void) |
2511 | { |
2512 | #ifdef KVM_CAP_IRQ_ROUTING |
2513 | return kvm_check_extension(kvm_state, KVM_CAP_IRQ_ROUTING); |
2514 | #else |
2515 | return false; |
2516 | #endif |
2517 | } |
2518 | |
2519 | int kvm_has_intx_set_mask(void) |
2520 | { |
2521 | return kvm_state->intx_set_mask; |
2522 | } |
2523 | |
2524 | bool kvm_arm_supports_user_irq(void) |
2525 | { |
2526 | return kvm_check_extension(kvm_state, KVM_CAP_ARM_USER_IRQ); |
2527 | } |
2528 | |
2529 | #ifdef KVM_CAP_SET_GUEST_DEBUG |
2530 | struct kvm_sw_breakpoint *kvm_find_sw_breakpoint(CPUState *cpu, |
2531 | target_ulong pc) |
2532 | { |
2533 | struct kvm_sw_breakpoint *bp; |
2534 | |
2535 | QTAILQ_FOREACH(bp, &cpu->kvm_state->kvm_sw_breakpoints, entry) { |
2536 | if (bp->pc == pc) { |
2537 | return bp; |
2538 | } |
2539 | } |
2540 | return NULL; |
2541 | } |
2542 | |
2543 | int kvm_sw_breakpoints_active(CPUState *cpu) |
2544 | { |
2545 | return !QTAILQ_EMPTY(&cpu->kvm_state->kvm_sw_breakpoints); |
2546 | } |
2547 | |
2548 | struct kvm_set_guest_debug_data { |
2549 | struct kvm_guest_debug dbg; |
2550 | int err; |
2551 | }; |
2552 | |
2553 | static void kvm_invoke_set_guest_debug(CPUState *cpu, run_on_cpu_data data) |
2554 | { |
2555 | struct kvm_set_guest_debug_data *dbg_data = |
2556 | (struct kvm_set_guest_debug_data *) data.host_ptr; |
2557 | |
2558 | dbg_data->err = kvm_vcpu_ioctl(cpu, KVM_SET_GUEST_DEBUG, |
2559 | &dbg_data->dbg); |
2560 | } |
2561 | |
2562 | int kvm_update_guest_debug(CPUState *cpu, unsigned long reinject_trap) |
2563 | { |
2564 | struct kvm_set_guest_debug_data data; |
2565 | |
2566 | data.dbg.control = reinject_trap; |
2567 | |
2568 | if (cpu->singlestep_enabled) { |
2569 | data.dbg.control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_SINGLESTEP; |
2570 | } |
2571 | kvm_arch_update_guest_debug(cpu, &data.dbg); |
2572 | |
2573 | run_on_cpu(cpu, kvm_invoke_set_guest_debug, |
2574 | RUN_ON_CPU_HOST_PTR(&data)); |
2575 | return data.err; |
2576 | } |
2577 | |
2578 | int kvm_insert_breakpoint(CPUState *cpu, target_ulong addr, |
2579 | target_ulong len, int type) |
2580 | { |
2581 | struct kvm_sw_breakpoint *bp; |
2582 | int err; |
2583 | |
2584 | if (type == GDB_BREAKPOINT_SW) { |
2585 | bp = kvm_find_sw_breakpoint(cpu, addr); |
2586 | if (bp) { |
2587 | bp->use_count++; |
2588 | return 0; |
2589 | } |
2590 | |
2591 | bp = g_malloc(sizeof(struct kvm_sw_breakpoint)); |
2592 | bp->pc = addr; |
2593 | bp->use_count = 1; |
2594 | err = kvm_arch_insert_sw_breakpoint(cpu, bp); |
2595 | if (err) { |
2596 | g_free(bp); |
2597 | return err; |
2598 | } |
2599 | |
2600 | QTAILQ_INSERT_HEAD(&cpu->kvm_state->kvm_sw_breakpoints, bp, entry); |
2601 | } else { |
2602 | err = kvm_arch_insert_hw_breakpoint(addr, len, type); |
2603 | if (err) { |
2604 | return err; |
2605 | } |
2606 | } |
2607 | |
2608 | CPU_FOREACH(cpu) { |
2609 | err = kvm_update_guest_debug(cpu, 0); |
2610 | if (err) { |
2611 | return err; |
2612 | } |
2613 | } |
2614 | return 0; |
2615 | } |
2616 | |
2617 | int kvm_remove_breakpoint(CPUState *cpu, target_ulong addr, |
2618 | target_ulong len, int type) |
2619 | { |
2620 | struct kvm_sw_breakpoint *bp; |
2621 | int err; |
2622 | |
2623 | if (type == GDB_BREAKPOINT_SW) { |
2624 | bp = kvm_find_sw_breakpoint(cpu, addr); |
2625 | if (!bp) { |
2626 | return -ENOENT; |
2627 | } |
2628 | |
2629 | if (bp->use_count > 1) { |
2630 | bp->use_count--; |
2631 | return 0; |
2632 | } |
2633 | |
2634 | err = kvm_arch_remove_sw_breakpoint(cpu, bp); |
2635 | if (err) { |
2636 | return err; |
2637 | } |
2638 | |
2639 | QTAILQ_REMOVE(&cpu->kvm_state->kvm_sw_breakpoints, bp, entry); |
2640 | g_free(bp); |
2641 | } else { |
2642 | err = kvm_arch_remove_hw_breakpoint(addr, len, type); |
2643 | if (err) { |
2644 | return err; |
2645 | } |
2646 | } |
2647 | |
2648 | CPU_FOREACH(cpu) { |
2649 | err = kvm_update_guest_debug(cpu, 0); |
2650 | if (err) { |
2651 | return err; |
2652 | } |
2653 | } |
2654 | return 0; |
2655 | } |
2656 | |
2657 | void kvm_remove_all_breakpoints(CPUState *cpu) |
2658 | { |
2659 | struct kvm_sw_breakpoint *bp, *next; |
2660 | KVMState *s = cpu->kvm_state; |
2661 | CPUState *tmpcpu; |
2662 | |
2663 | QTAILQ_FOREACH_SAFE(bp, &s->kvm_sw_breakpoints, entry, next) { |
2664 | if (kvm_arch_remove_sw_breakpoint(cpu, bp) != 0) { |
2665 | /* Try harder to find a CPU that currently sees the breakpoint. */ |
2666 | CPU_FOREACH(tmpcpu) { |
2667 | if (kvm_arch_remove_sw_breakpoint(tmpcpu, bp) == 0) { |
2668 | break; |
2669 | } |
2670 | } |
2671 | } |
2672 | QTAILQ_REMOVE(&s->kvm_sw_breakpoints, bp, entry); |
2673 | g_free(bp); |
2674 | } |
2675 | kvm_arch_remove_all_hw_breakpoints(); |
2676 | |
2677 | CPU_FOREACH(cpu) { |
2678 | kvm_update_guest_debug(cpu, 0); |
2679 | } |
2680 | } |
2681 | |
2682 | #else /* !KVM_CAP_SET_GUEST_DEBUG */ |
2683 | |
2684 | int kvm_update_guest_debug(CPUState *cpu, unsigned long reinject_trap) |
2685 | { |
2686 | return -EINVAL; |
2687 | } |
2688 | |
2689 | int kvm_insert_breakpoint(CPUState *cpu, target_ulong addr, |
2690 | target_ulong len, int type) |
2691 | { |
2692 | return -EINVAL; |
2693 | } |
2694 | |
2695 | int kvm_remove_breakpoint(CPUState *cpu, target_ulong addr, |
2696 | target_ulong len, int type) |
2697 | { |
2698 | return -EINVAL; |
2699 | } |
2700 | |
2701 | void kvm_remove_all_breakpoints(CPUState *cpu) |
2702 | { |
2703 | } |
2704 | #endif /* !KVM_CAP_SET_GUEST_DEBUG */ |
2705 | |
2706 | static int kvm_set_signal_mask(CPUState *cpu, const sigset_t *sigset) |
2707 | { |
2708 | KVMState *s = kvm_state; |
2709 | struct kvm_signal_mask *sigmask; |
2710 | int r; |
2711 | |
2712 | sigmask = g_malloc(sizeof(*sigmask) + sizeof(*sigset)); |
2713 | |
2714 | sigmask->len = s->sigmask_len; |
2715 | memcpy(sigmask->sigset, sigset, sizeof(*sigset)); |
2716 | r = kvm_vcpu_ioctl(cpu, KVM_SET_SIGNAL_MASK, sigmask); |
2717 | g_free(sigmask); |
2718 | |
2719 | return r; |
2720 | } |
2721 | |
2722 | static void kvm_ipi_signal(int sig) |
2723 | { |
2724 | if (current_cpu) { |
2725 | assert(kvm_immediate_exit); |
2726 | kvm_cpu_kick(current_cpu); |
2727 | } |
2728 | } |
2729 | |
2730 | void kvm_init_cpu_signals(CPUState *cpu) |
2731 | { |
2732 | int r; |
2733 | sigset_t set; |
2734 | struct sigaction sigact; |
2735 | |
2736 | memset(&sigact, 0, sizeof(sigact)); |
2737 | sigact.sa_handler = kvm_ipi_signal; |
2738 | sigaction(SIG_IPI, &sigact, NULL); |
2739 | |
2740 | pthread_sigmask(SIG_BLOCK, NULL, &set); |
2741 | #if defined KVM_HAVE_MCE_INJECTION |
2742 | sigdelset(&set, SIGBUS); |
2743 | pthread_sigmask(SIG_SETMASK, &set, NULL); |
2744 | #endif |
2745 | sigdelset(&set, SIG_IPI); |
2746 | if (kvm_immediate_exit) { |
2747 | r = pthread_sigmask(SIG_SETMASK, &set, NULL); |
2748 | } else { |
2749 | r = kvm_set_signal_mask(cpu, &set); |
2750 | } |
2751 | if (r) { |
2752 | fprintf(stderr, "kvm_set_signal_mask: %s\n" , strerror(-r)); |
2753 | exit(1); |
2754 | } |
2755 | } |
2756 | |
2757 | /* Called asynchronously in VCPU thread. */ |
2758 | int kvm_on_sigbus_vcpu(CPUState *cpu, int code, void *addr) |
2759 | { |
2760 | #ifdef KVM_HAVE_MCE_INJECTION |
2761 | if (have_sigbus_pending) { |
2762 | return 1; |
2763 | } |
2764 | have_sigbus_pending = true; |
2765 | pending_sigbus_addr = addr; |
2766 | pending_sigbus_code = code; |
2767 | atomic_set(&cpu->exit_request, 1); |
2768 | return 0; |
2769 | #else |
2770 | return 1; |
2771 | #endif |
2772 | } |
2773 | |
2774 | /* Called synchronously (via signalfd) in main thread. */ |
2775 | int kvm_on_sigbus(int code, void *addr) |
2776 | { |
2777 | #ifdef KVM_HAVE_MCE_INJECTION |
2778 | /* Action required MCE kills the process if SIGBUS is blocked. Because |
2779 | * that's what happens in the I/O thread, where we handle MCE via signalfd, |
2780 | * we can only get action optional here. |
2781 | */ |
2782 | assert(code != BUS_MCEERR_AR); |
2783 | kvm_arch_on_sigbus_vcpu(first_cpu, code, addr); |
2784 | return 0; |
2785 | #else |
2786 | return 1; |
2787 | #endif |
2788 | } |
2789 | |
2790 | int kvm_create_device(KVMState *s, uint64_t type, bool test) |
2791 | { |
2792 | int ret; |
2793 | struct kvm_create_device create_dev; |
2794 | |
2795 | create_dev.type = type; |
2796 | create_dev.fd = -1; |
2797 | create_dev.flags = test ? KVM_CREATE_DEVICE_TEST : 0; |
2798 | |
2799 | if (!kvm_check_extension(s, KVM_CAP_DEVICE_CTRL)) { |
2800 | return -ENOTSUP; |
2801 | } |
2802 | |
2803 | ret = kvm_vm_ioctl(s, KVM_CREATE_DEVICE, &create_dev); |
2804 | if (ret) { |
2805 | return ret; |
2806 | } |
2807 | |
2808 | return test ? 0 : create_dev.fd; |
2809 | } |
2810 | |
2811 | bool kvm_device_supported(int vmfd, uint64_t type) |
2812 | { |
2813 | struct kvm_create_device create_dev = { |
2814 | .type = type, |
2815 | .fd = -1, |
2816 | .flags = KVM_CREATE_DEVICE_TEST, |
2817 | }; |
2818 | |
2819 | if (ioctl(vmfd, KVM_CHECK_EXTENSION, KVM_CAP_DEVICE_CTRL) <= 0) { |
2820 | return false; |
2821 | } |
2822 | |
2823 | return (ioctl(vmfd, KVM_CREATE_DEVICE, &create_dev) >= 0); |
2824 | } |
2825 | |
2826 | int kvm_set_one_reg(CPUState *cs, uint64_t id, void *source) |
2827 | { |
2828 | struct kvm_one_reg reg; |
2829 | int r; |
2830 | |
2831 | reg.id = id; |
2832 | reg.addr = (uintptr_t) source; |
2833 | r = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, ®); |
2834 | if (r) { |
2835 | trace_kvm_failed_reg_set(id, strerror(-r)); |
2836 | } |
2837 | return r; |
2838 | } |
2839 | |
2840 | int kvm_get_one_reg(CPUState *cs, uint64_t id, void *target) |
2841 | { |
2842 | struct kvm_one_reg reg; |
2843 | int r; |
2844 | |
2845 | reg.id = id; |
2846 | reg.addr = (uintptr_t) target; |
2847 | r = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, ®); |
2848 | if (r) { |
2849 | trace_kvm_failed_reg_get(id, strerror(-r)); |
2850 | } |
2851 | return r; |
2852 | } |
2853 | |
2854 | static bool kvm_accel_has_memory(MachineState *ms, AddressSpace *as, |
2855 | hwaddr start_addr, hwaddr size) |
2856 | { |
2857 | KVMState *kvm = KVM_STATE(ms->accelerator); |
2858 | int i; |
2859 | |
2860 | for (i = 0; i < kvm->nr_as; ++i) { |
2861 | if (kvm->as[i].as == as && kvm->as[i].ml) { |
2862 | return NULL != kvm_lookup_matching_slot(kvm->as[i].ml, |
2863 | start_addr, size); |
2864 | } |
2865 | } |
2866 | |
2867 | return false; |
2868 | } |
2869 | |
2870 | static void kvm_accel_class_init(ObjectClass *oc, void *data) |
2871 | { |
2872 | AccelClass *ac = ACCEL_CLASS(oc); |
2873 | ac->name = "KVM" ; |
2874 | ac->init_machine = kvm_init; |
2875 | ac->has_memory = kvm_accel_has_memory; |
2876 | ac->allowed = &kvm_allowed; |
2877 | } |
2878 | |
2879 | static const TypeInfo kvm_accel_type = { |
2880 | .name = TYPE_KVM_ACCEL, |
2881 | .parent = TYPE_ACCEL, |
2882 | .class_init = kvm_accel_class_init, |
2883 | .instance_size = sizeof(KVMState), |
2884 | }; |
2885 | |
2886 | static void kvm_type_init(void) |
2887 | { |
2888 | type_register_static(&kvm_accel_type); |
2889 | } |
2890 | |
2891 | type_init(kvm_type_init); |
2892 | |