?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
hw/arm/armsse: Move PPUs into data-driven framework
[qemu/ar7.git] / accel / kvm / kvm-all.c
blobf88a52393fe0bd04636855d56895397fcad30d9c
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 "qapi/visitor.h"
43 #include "qapi/qapi-types-common.h"
44 #include "qapi/qapi-visit-common.h"
45 #include "sysemu/reset.h"
46 #include "qemu/guest-random.h"
47 #include "sysemu/hw_accel.h"
48 #include "kvm-cpus.h"
49
50 #include "hw/boards.h"
51
52 /* This check must be after config-host.h is included */
53 #ifdef CONFIG_EVENTFD
54 #include <sys/eventfd.h>
55 #endif
56
57 /* KVM uses PAGE_SIZE in its definition of KVM_COALESCED_MMIO_MAX. We
58 * need to use the real host PAGE_SIZE, as that's what KVM will use.
59 */
60 #ifdef PAGE_SIZE
61 #undef PAGE_SIZE
62 #endif
63 #define PAGE_SIZE qemu_real_host_page_size
64
65 //#define DEBUG_KVM
66
67 #ifdef DEBUG_KVM
68 #define DPRINTF(fmt, ...) \
69 do { fprintf(stderr, fmt, ## __VA_ARGS__); } while (0)
70 #else
71 #define DPRINTF(fmt, ...) \
72 do { } while (0)
73 #endif
74
75 #define KVM_MSI_HASHTAB_SIZE 256
76
77 struct KVMParkedVcpu {
78 unsigned long vcpu_id;
79 int kvm_fd;
80 QLIST_ENTRY(KVMParkedVcpu) node;
81 };
82
83 struct KVMState
84 {
85 AccelState parent_obj;
86
87 int nr_slots;
88 int fd;
89 int vmfd;
90 int coalesced_mmio;
91 int coalesced_pio;
92 struct kvm_coalesced_mmio_ring *coalesced_mmio_ring;
93 bool coalesced_flush_in_progress;
94 int vcpu_events;
95 int robust_singlestep;
96 int debugregs;
97 #ifdef KVM_CAP_SET_GUEST_DEBUG
98 QTAILQ_HEAD(, kvm_sw_breakpoint) kvm_sw_breakpoints;
99 #endif
100 int max_nested_state_len;
101 int many_ioeventfds;
102 int intx_set_mask;
103 int kvm_shadow_mem;
104 bool kernel_irqchip_allowed;
105 bool kernel_irqchip_required;
106 OnOffAuto kernel_irqchip_split;
107 bool sync_mmu;
108 uint64_t manual_dirty_log_protect;
109 /* The man page (and posix) say ioctl numbers are signed int, but
110 * they're not. Linux, glibc and *BSD all treat ioctl numbers as
111 * unsigned, and treating them as signed here can break things */
112 unsigned irq_set_ioctl;
113 unsigned int sigmask_len;
114 GHashTable *gsimap;
115 #ifdef KVM_CAP_IRQ_ROUTING
116 struct kvm_irq_routing *irq_routes;
117 int nr_allocated_irq_routes;
118 unsigned long *used_gsi_bitmap;
119 unsigned int gsi_count;
120 QTAILQ_HEAD(, KVMMSIRoute) msi_hashtab[KVM_MSI_HASHTAB_SIZE];
121 #endif
122 KVMMemoryListener memory_listener;
123 QLIST_HEAD(, KVMParkedVcpu) kvm_parked_vcpus;
124
125 /* For "info mtree -f" to tell if an MR is registered in KVM */
126 int nr_as;
127 struct KVMAs {
128 KVMMemoryListener *ml;
129 AddressSpace *as;
130 } *as;
131 };
132
133 KVMState *kvm_state;
134 bool kvm_kernel_irqchip;
135 bool kvm_split_irqchip;
136 bool kvm_async_interrupts_allowed;
137 bool kvm_halt_in_kernel_allowed;
138 bool kvm_eventfds_allowed;
139 bool kvm_irqfds_allowed;
140 bool kvm_resamplefds_allowed;
141 bool kvm_msi_via_irqfd_allowed;
142 bool kvm_gsi_routing_allowed;
143 bool kvm_gsi_direct_mapping;
144 bool kvm_allowed;
145 bool kvm_readonly_mem_allowed;
146 bool kvm_vm_attributes_allowed;
147 bool kvm_direct_msi_allowed;
148 bool kvm_ioeventfd_any_length_allowed;
149 bool kvm_msi_use_devid;
150 static bool kvm_immediate_exit;
151 static hwaddr kvm_max_slot_size = ~0;
152
153 static const KVMCapabilityInfo kvm_required_capabilites[] = {
154 KVM_CAP_INFO(USER_MEMORY),
155 KVM_CAP_INFO(DESTROY_MEMORY_REGION_WORKS),
156 KVM_CAP_INFO(JOIN_MEMORY_REGIONS_WORKS),
157 KVM_CAP_LAST_INFO
158 };
159
160 static NotifierList kvm_irqchip_change_notifiers =
161 NOTIFIER_LIST_INITIALIZER(kvm_irqchip_change_notifiers);
162
163 struct KVMResampleFd {
164 int gsi;
165 EventNotifier *resample_event;
166 QLIST_ENTRY(KVMResampleFd) node;
167 };
168 typedef struct KVMResampleFd KVMResampleFd;
169
170 /*
171 * Only used with split irqchip where we need to do the resample fd
172 * kick for the kernel from userspace.
173 */
174 static QLIST_HEAD(, KVMResampleFd) kvm_resample_fd_list =
175 QLIST_HEAD_INITIALIZER(kvm_resample_fd_list);
176
177 #define kvm_slots_lock(kml) qemu_mutex_lock(&(kml)->slots_lock)
178 #define kvm_slots_unlock(kml) qemu_mutex_unlock(&(kml)->slots_lock)
179
180 static inline void kvm_resample_fd_remove(int gsi)
181 {
182 KVMResampleFd *rfd;
183
184 QLIST_FOREACH(rfd, &kvm_resample_fd_list, node) {
185 if (rfd->gsi == gsi) {
186 QLIST_REMOVE(rfd, node);
187 g_free(rfd);
188 break;
189 }
190 }
191 }
192
193 static inline void kvm_resample_fd_insert(int gsi, EventNotifier *event)
194 {
195 KVMResampleFd *rfd = g_new0(KVMResampleFd, 1);
196
197 rfd->gsi = gsi;
198 rfd->resample_event = event;
199
200 QLIST_INSERT_HEAD(&kvm_resample_fd_list, rfd, node);
201 }
202
203 void kvm_resample_fd_notify(int gsi)
204 {
205 KVMResampleFd *rfd;
206
207 QLIST_FOREACH(rfd, &kvm_resample_fd_list, node) {
208 if (rfd->gsi == gsi) {
209 event_notifier_set(rfd->resample_event);
210 trace_kvm_resample_fd_notify(gsi);
211 return;
212 }
213 }
214 }
215
216 int kvm_get_max_memslots(void)
217 {
218 KVMState *s = KVM_STATE(current_accel());
219
220 return s->nr_slots;
221 }
222
223 /* Called with KVMMemoryListener.slots_lock held */
224 static KVMSlot *kvm_get_free_slot(KVMMemoryListener *kml)
225 {
226 KVMState *s = kvm_state;
227 int i;
228
229 for (i = 0; i < s->nr_slots; i++) {
230 if (kml->slots[i].memory_size == 0) {
231 return &kml->slots[i];
232 }
233 }
234
235 return NULL;
236 }
237
238 bool kvm_has_free_slot(MachineState *ms)
239 {
240 KVMState *s = KVM_STATE(ms->accelerator);
241 bool result;
242 KVMMemoryListener *kml = &s->memory_listener;
243
244 kvm_slots_lock(kml);
245 result = !!kvm_get_free_slot(kml);
246 kvm_slots_unlock(kml);
247
248 return result;
249 }
250
251 /* Called with KVMMemoryListener.slots_lock held */
252 static KVMSlot *kvm_alloc_slot(KVMMemoryListener *kml)
253 {
254 KVMSlot *slot = kvm_get_free_slot(kml);
255
256 if (slot) {
257 return slot;
258 }
259
260 fprintf(stderr, "%s: no free slot available\n", __func__);
261 abort();
262 }
263
264 static KVMSlot *kvm_lookup_matching_slot(KVMMemoryListener *kml,
265 hwaddr start_addr,
266 hwaddr size)
267 {
268 KVMState *s = kvm_state;
269 int i;
270
271 for (i = 0; i < s->nr_slots; i++) {
272 KVMSlot *mem = &kml->slots[i];
273
274 if (start_addr == mem->start_addr && size == mem->memory_size) {
275 return mem;
276 }
277 }
278
279 return NULL;
280 }
281
282 /*
283 * Calculate and align the start address and the size of the section.
284 * Return the size. If the size is 0, the aligned section is empty.
285 */
286 static hwaddr kvm_align_section(MemoryRegionSection *section,
287 hwaddr *start)
288 {
289 hwaddr size = int128_get64(section->size);
290 hwaddr delta, aligned;
291
292 /* kvm works in page size chunks, but the function may be called
293 with sub-page size and unaligned start address. Pad the start
294 address to next and truncate size to previous page boundary. */
295 aligned = ROUND_UP(section->offset_within_address_space,
296 qemu_real_host_page_size);
297 delta = aligned - section->offset_within_address_space;
298 *start = aligned;
299 if (delta > size) {
300 return 0;
301 }
302
303 return (size - delta) & qemu_real_host_page_mask;
304 }
305
306 int kvm_physical_memory_addr_from_host(KVMState *s, void *ram,
307 hwaddr *phys_addr)
308 {
309 KVMMemoryListener *kml = &s->memory_listener;
310 int i, ret = 0;
311
312 kvm_slots_lock(kml);
313 for (i = 0; i < s->nr_slots; i++) {
314 KVMSlot *mem = &kml->slots[i];
315
316 if (ram >= mem->ram && ram < mem->ram + mem->memory_size) {
317 *phys_addr = mem->start_addr + (ram - mem->ram);
318 ret = 1;
319 break;
320 }
321 }
322 kvm_slots_unlock(kml);
323
324 return ret;
325 }
326
327 static int kvm_set_user_memory_region(KVMMemoryListener *kml, KVMSlot *slot, bool new)
328 {
329 KVMState *s = kvm_state;
330 struct kvm_userspace_memory_region mem;
331 int ret;
332
333 mem.slot = slot->slot | (kml->as_id << 16);
334 mem.guest_phys_addr = slot->start_addr;
335 mem.userspace_addr = (unsigned long)slot->ram;
336 mem.flags = slot->flags;
337
338 if (slot->memory_size && !new && (mem.flags ^ slot->old_flags) & KVM_MEM_READONLY) {
339 /* Set the slot size to 0 before setting the slot to the desired
340 * value. This is needed based on KVM commit 75d61fbc. */
341 mem.memory_size = 0;
342 ret = kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
343 if (ret < 0) {
344 goto err;
345 }
346 }
347 mem.memory_size = slot->memory_size;
348 ret = kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
349 slot->old_flags = mem.flags;
350 err:
351 trace_kvm_set_user_memory(mem.slot, mem.flags, mem.guest_phys_addr,
352 mem.memory_size, mem.userspace_addr, ret);
353 if (ret < 0) {
354 error_report("%s: KVM_SET_USER_MEMORY_REGION failed, slot=%d,"
355 " start=0x%" PRIx64 ", size=0x%" PRIx64 ": %s",
356 __func__, mem.slot, slot->start_addr,
357 (uint64_t)mem.memory_size, strerror(errno));
358 }
359 return ret;
360 }
361
362 static int do_kvm_destroy_vcpu(CPUState *cpu)
363 {
364 KVMState *s = kvm_state;
365 long mmap_size;
366 struct KVMParkedVcpu *vcpu = NULL;
367 int ret = 0;
368
369 DPRINTF("kvm_destroy_vcpu\n");
370
371 ret = kvm_arch_destroy_vcpu(cpu);
372 if (ret < 0) {
373 goto err;
374 }
375
376 mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
377 if (mmap_size < 0) {
378 ret = mmap_size;
379 DPRINTF("KVM_GET_VCPU_MMAP_SIZE failed\n");
380 goto err;
381 }
382
383 ret = munmap(cpu->kvm_run, mmap_size);
384 if (ret < 0) {
385 goto err;
386 }
387
388 vcpu = g_malloc0(sizeof(*vcpu));
389 vcpu->vcpu_id = kvm_arch_vcpu_id(cpu);
390 vcpu->kvm_fd = cpu->kvm_fd;
391 QLIST_INSERT_HEAD(&kvm_state->kvm_parked_vcpus, vcpu, node);
392 err:
393 return ret;
394 }
395
396 void kvm_destroy_vcpu(CPUState *cpu)
397 {
398 if (do_kvm_destroy_vcpu(cpu) < 0) {
399 error_report("kvm_destroy_vcpu failed");
400 exit(EXIT_FAILURE);
401 }
402 }
403
404 static int kvm_get_vcpu(KVMState *s, unsigned long vcpu_id)
405 {
406 struct KVMParkedVcpu *cpu;
407
408 QLIST_FOREACH(cpu, &s->kvm_parked_vcpus, node) {
409 if (cpu->vcpu_id == vcpu_id) {
410 int kvm_fd;
411
412 QLIST_REMOVE(cpu, node);
413 kvm_fd = cpu->kvm_fd;
414 g_free(cpu);
415 return kvm_fd;
416 }
417 }
418
419 return kvm_vm_ioctl(s, KVM_CREATE_VCPU, (void *)vcpu_id);
420 }
421
422 int kvm_init_vcpu(CPUState *cpu, Error **errp)
423 {
424 KVMState *s = kvm_state;
425 long mmap_size;
426 int ret;
427
428 trace_kvm_init_vcpu(cpu->cpu_index, kvm_arch_vcpu_id(cpu));
429
430 ret = kvm_get_vcpu(s, kvm_arch_vcpu_id(cpu));
431 if (ret < 0) {
432 error_setg_errno(errp, -ret, "kvm_init_vcpu: kvm_get_vcpu failed (%lu)",
433 kvm_arch_vcpu_id(cpu));
434 goto err;
435 }
436
437 cpu->kvm_fd = ret;
438 cpu->kvm_state = s;
439 cpu->vcpu_dirty = true;
440
441 mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
442 if (mmap_size < 0) {
443 ret = mmap_size;
444 error_setg_errno(errp, -mmap_size,
445 "kvm_init_vcpu: KVM_GET_VCPU_MMAP_SIZE failed");
446 goto err;
447 }
448
449 cpu->kvm_run = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
450 cpu->kvm_fd, 0);
451 if (cpu->kvm_run == MAP_FAILED) {
452 ret = -errno;
453 error_setg_errno(errp, ret,
454 "kvm_init_vcpu: mmap'ing vcpu state failed (%lu)",
455 kvm_arch_vcpu_id(cpu));
456 goto err;
457 }
458
459 if (s->coalesced_mmio && !s->coalesced_mmio_ring) {
460 s->coalesced_mmio_ring =
461 (void *)cpu->kvm_run + s->coalesced_mmio * PAGE_SIZE;
462 }
463
464 ret = kvm_arch_init_vcpu(cpu);
465 if (ret < 0) {
466 error_setg_errno(errp, -ret,
467 "kvm_init_vcpu: kvm_arch_init_vcpu failed (%lu)",
468 kvm_arch_vcpu_id(cpu));
469 }
470 err:
471 return ret;
472 }
473
474 /*
475 * dirty pages logging control
476 */
477
478 static int kvm_mem_flags(MemoryRegion *mr)
479 {
480 bool readonly = mr->readonly || memory_region_is_romd(mr);
481 int flags = 0;
482
483 if (memory_region_get_dirty_log_mask(mr) != 0) {
484 flags |= KVM_MEM_LOG_DIRTY_PAGES;
485 }
486 if (readonly && kvm_readonly_mem_allowed) {
487 flags |= KVM_MEM_READONLY;
488 }
489 return flags;
490 }
491
492 /* Called with KVMMemoryListener.slots_lock held */
493 static int kvm_slot_update_flags(KVMMemoryListener *kml, KVMSlot *mem,
494 MemoryRegion *mr)
495 {
496 mem->flags = kvm_mem_flags(mr);
497
498 /* If nothing changed effectively, no need to issue ioctl */
499 if (mem->flags == mem->old_flags) {
500 return 0;
501 }
502
503 return kvm_set_user_memory_region(kml, mem, false);
504 }
505
506 static int kvm_section_update_flags(KVMMemoryListener *kml,
507 MemoryRegionSection *section)
508 {
509 hwaddr start_addr, size, slot_size;
510 KVMSlot *mem;
511 int ret = 0;
512
513 size = kvm_align_section(section, &start_addr);
514 if (!size) {
515 return 0;
516 }
517
518 kvm_slots_lock(kml);
519
520 while (size && !ret) {
521 slot_size = MIN(kvm_max_slot_size, size);
522 mem = kvm_lookup_matching_slot(kml, start_addr, slot_size);
523 if (!mem) {
524 /* We don't have a slot if we want to trap every access. */
525 goto out;
526 }
527
528 ret = kvm_slot_update_flags(kml, mem, section->mr);
529 start_addr += slot_size;
530 size -= slot_size;
531 }
532
533 out:
534 kvm_slots_unlock(kml);
535 return ret;
536 }
537
538 static void kvm_log_start(MemoryListener *listener,
539 MemoryRegionSection *section,
540 int old, int new)
541 {
542 KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
543 int r;
544
545 if (old != 0) {
546 return;
547 }
548
549 r = kvm_section_update_flags(kml, section);
550 if (r < 0) {
551 abort();
552 }
553 }
554
555 static void kvm_log_stop(MemoryListener *listener,
556 MemoryRegionSection *section,
557 int old, int new)
558 {
559 KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
560 int r;
561
562 if (new != 0) {
563 return;
564 }
565
566 r = kvm_section_update_flags(kml, section);
567 if (r < 0) {
568 abort();
569 }
570 }
571
572 /* get kvm's dirty pages bitmap and update qemu's */
573 static int kvm_get_dirty_pages_log_range(MemoryRegionSection *section,
574 unsigned long *bitmap)
575 {
576 ram_addr_t start = section->offset_within_region +
577 memory_region_get_ram_addr(section->mr);
578 ram_addr_t pages = int128_get64(section->size) / qemu_real_host_page_size;
579
580 cpu_physical_memory_set_dirty_lebitmap(bitmap, start, pages);
581 return 0;
582 }
583
584 #define ALIGN(x, y) (((x)+(y)-1) & ~((y)-1))
585
586 /* Allocate the dirty bitmap for a slot */
587 static void kvm_memslot_init_dirty_bitmap(KVMSlot *mem)
588 {
589 /*
590 * XXX bad kernel interface alert
591 * For dirty bitmap, kernel allocates array of size aligned to
592 * bits-per-long. But for case when the kernel is 64bits and
593 * the userspace is 32bits, userspace can't align to the same
594 * bits-per-long, since sizeof(long) is different between kernel
595 * and user space. This way, userspace will provide buffer which
596 * may be 4 bytes less than the kernel will use, resulting in
597 * userspace memory corruption (which is not detectable by valgrind
598 * too, in most cases).
599 * So for now, let's align to 64 instead of HOST_LONG_BITS here, in
600 * a hope that sizeof(long) won't become >8 any time soon.
601 *
602 * Note: the granule of kvm dirty log is qemu_real_host_page_size.
603 * And mem->memory_size is aligned to it (otherwise this mem can't
604 * be registered to KVM).
605 */
606 hwaddr bitmap_size = ALIGN(mem->memory_size / qemu_real_host_page_size,
607 /*HOST_LONG_BITS*/ 64) / 8;
608 mem->dirty_bmap = g_malloc0(bitmap_size);
609 }
610
611 /**
612 * kvm_physical_sync_dirty_bitmap - Sync dirty bitmap from kernel space
613 *
614 * This function will first try to fetch dirty bitmap from the kernel,
615 * and then updates qemu's dirty bitmap.
616 *
617 * NOTE: caller must be with kml->slots_lock held.
618 *
619 * @kml: the KVM memory listener object
620 * @section: the memory section to sync the dirty bitmap with
621 */
622 static int kvm_physical_sync_dirty_bitmap(KVMMemoryListener *kml,
623 MemoryRegionSection *section)
624 {
625 KVMState *s = kvm_state;
626 struct kvm_dirty_log d = {};
627 KVMSlot *mem;
628 hwaddr start_addr, size;
629 hwaddr slot_size, slot_offset = 0;
630 int ret = 0;
631
632 size = kvm_align_section(section, &start_addr);
633 while (size) {
634 MemoryRegionSection subsection = *section;
635
636 slot_size = MIN(kvm_max_slot_size, size);
637 mem = kvm_lookup_matching_slot(kml, start_addr, slot_size);
638 if (!mem) {
639 /* We don't have a slot if we want to trap every access. */
640 goto out;
641 }
642
643 if (!mem->dirty_bmap) {
644 /* Allocate on the first log_sync, once and for all */
645 kvm_memslot_init_dirty_bitmap(mem);
646 }
647
648 d.dirty_bitmap = mem->dirty_bmap;
649 d.slot = mem->slot | (kml->as_id << 16);
650 ret = kvm_vm_ioctl(s, KVM_GET_DIRTY_LOG, &d);
651 if (ret == -ENOENT) {
652 /* kernel does not have dirty bitmap in this slot */
653 ret = 0;
654 } else if (ret < 0) {
655 error_report("ioctl KVM_GET_DIRTY_LOG failed: %d", errno);
656 goto out;
657 } else {
658 subsection.offset_within_region += slot_offset;
659 subsection.size = int128_make64(slot_size);
660 kvm_get_dirty_pages_log_range(&subsection, d.dirty_bitmap);
661 }
662
663 slot_offset += slot_size;
664 start_addr += slot_size;
665 size -= slot_size;
666 }
667 out:
668 return ret;
669 }
670
671 /* Alignment requirement for KVM_CLEAR_DIRTY_LOG - 64 pages */
672 #define KVM_CLEAR_LOG_SHIFT 6
673 #define KVM_CLEAR_LOG_ALIGN (qemu_real_host_page_size << KVM_CLEAR_LOG_SHIFT)
674 #define KVM_CLEAR_LOG_MASK (-KVM_CLEAR_LOG_ALIGN)
675
676 /*
677 * As the granule of kvm dirty log is qemu_real_host_page_size,
678 * @start and @size are expected and restricted to align to it.
679 */
680 static int kvm_log_clear_one_slot(KVMSlot *mem, int as_id, uint64_t start,
681 uint64_t size)
682 {
683 KVMState *s = kvm_state;
684 uint64_t end, bmap_start, start_delta, bmap_npages;
685 struct kvm_clear_dirty_log d;
686 unsigned long *bmap_clear = NULL, psize = qemu_real_host_page_size;
687 int ret;
688
689 /* Make sure start and size are qemu_real_host_page_size aligned */
690 assert(QEMU_IS_ALIGNED(start | size, psize));
691
692 /*
693 * We need to extend either the start or the size or both to
694 * satisfy the KVM interface requirement. Firstly, do the start
695 * page alignment on 64 host pages
696 */
697 bmap_start = start & KVM_CLEAR_LOG_MASK;
698 start_delta = start - bmap_start;
699 bmap_start /= psize;
700
701 /*
702 * The kernel interface has restriction on the size too, that either:
703 *
704 * (1) the size is 64 host pages aligned (just like the start), or
705 * (2) the size fills up until the end of the KVM memslot.
706 */
707 bmap_npages = DIV_ROUND_UP(size + start_delta, KVM_CLEAR_LOG_ALIGN)
708 << KVM_CLEAR_LOG_SHIFT;
709 end = mem->memory_size / psize;
710 if (bmap_npages > end - bmap_start) {
711 bmap_npages = end - bmap_start;
712 }
713 start_delta /= psize;
714
715 /*
716 * Prepare the bitmap to clear dirty bits. Here we must guarantee
717 * that we won't clear any unknown dirty bits otherwise we might
718 * accidentally clear some set bits which are not yet synced from
719 * the kernel into QEMU's bitmap, then we'll lose track of the
720 * guest modifications upon those pages (which can directly lead
721 * to guest data loss or panic after migration).
722 *
723 * Layout of the KVMSlot.dirty_bmap:
724 *
725 * |<-------- bmap_npages -----------..>|
726 * [1]
727 * start_delta size
728 * |----------------|-------------|------------------|------------|
729 * ^ ^ ^ ^
730 * | | | |
731 * start bmap_start (start) end
732 * of memslot of memslot
733 *
734 * [1] bmap_npages can be aligned to either 64 pages or the end of slot
735 */
736
737 assert(bmap_start % BITS_PER_LONG == 0);
738 /* We should never do log_clear before log_sync */
739 assert(mem->dirty_bmap);
740 if (start_delta || bmap_npages - size / psize) {
741 /* Slow path - we need to manipulate a temp bitmap */
742 bmap_clear = bitmap_new(bmap_npages);
743 bitmap_copy_with_src_offset(bmap_clear, mem->dirty_bmap,
744 bmap_start, start_delta + size / psize);
745 /*
746 * We need to fill the holes at start because that was not
747 * specified by the caller and we extended the bitmap only for
748 * 64 pages alignment
749 */
750 bitmap_clear(bmap_clear, 0, start_delta);
751 d.dirty_bitmap = bmap_clear;
752 } else {
753 /*
754 * Fast path - both start and size align well with BITS_PER_LONG
755 * (or the end of memory slot)
756 */
757 d.dirty_bitmap = mem->dirty_bmap + BIT_WORD(bmap_start);
758 }
759
760 d.first_page = bmap_start;
761 /* It should never overflow. If it happens, say something */
762 assert(bmap_npages <= UINT32_MAX);
763 d.num_pages = bmap_npages;
764 d.slot = mem->slot | (as_id << 16);
765
766 ret = kvm_vm_ioctl(s, KVM_CLEAR_DIRTY_LOG, &d);
767 if (ret < 0 && ret != -ENOENT) {
768 error_report("%s: KVM_CLEAR_DIRTY_LOG failed, slot=%d, "
769 "start=0x%"PRIx64", size=0x%"PRIx32", errno=%d",
770 __func__, d.slot, (uint64_t)d.first_page,
771 (uint32_t)d.num_pages, ret);
772 } else {
773 ret = 0;
774 trace_kvm_clear_dirty_log(d.slot, d.first_page, d.num_pages);
775 }
776
777 /*
778 * After we have updated the remote dirty bitmap, we update the
779 * cached bitmap as well for the memslot, then if another user
780 * clears the same region we know we shouldn't clear it again on
781 * the remote otherwise it's data loss as well.
782 */
783 bitmap_clear(mem->dirty_bmap, bmap_start + start_delta,
784 size / psize);
785 /* This handles the NULL case well */
786 g_free(bmap_clear);
787 return ret;
788 }
789
790
791 /**
792 * kvm_physical_log_clear - Clear the kernel's dirty bitmap for range
793 *
794 * NOTE: this will be a no-op if we haven't enabled manual dirty log
795 * protection in the host kernel because in that case this operation
796 * will be done within log_sync().
797 *
798 * @kml: the kvm memory listener
799 * @section: the memory range to clear dirty bitmap
800 */
801 static int kvm_physical_log_clear(KVMMemoryListener *kml,
802 MemoryRegionSection *section)
803 {
804 KVMState *s = kvm_state;
805 uint64_t start, size, offset, count;
806 KVMSlot *mem;
807 int ret = 0, i;
808
809 if (!s->manual_dirty_log_protect) {
810 /* No need to do explicit clear */
811 return ret;
812 }
813
814 start = section->offset_within_address_space;
815 size = int128_get64(section->size);
816
817 if (!size) {
818 /* Nothing more we can do... */
819 return ret;
820 }
821
822 kvm_slots_lock(kml);
823
824 for (i = 0; i < s->nr_slots; i++) {
825 mem = &kml->slots[i];
826 /* Discard slots that are empty or do not overlap the section */
827 if (!mem->memory_size ||
828 mem->start_addr > start + size - 1 ||
829 start > mem->start_addr + mem->memory_size - 1) {
830 continue;
831 }
832
833 if (start >= mem->start_addr) {
834 /* The slot starts before section or is aligned to it. */
835 offset = start - mem->start_addr;
836 count = MIN(mem->memory_size - offset, size);
837 } else {
838 /* The slot starts after section. */
839 offset = 0;
840 count = MIN(mem->memory_size, size - (mem->start_addr - start));
841 }
842 ret = kvm_log_clear_one_slot(mem, kml->as_id, offset, count);
843 if (ret < 0) {
844 break;
845 }
846 }
847
848 kvm_slots_unlock(kml);
849
850 return ret;
851 }
852
853 static void kvm_coalesce_mmio_region(MemoryListener *listener,
854 MemoryRegionSection *secion,
855 hwaddr start, hwaddr size)
856 {
857 KVMState *s = kvm_state;
858
859 if (s->coalesced_mmio) {
860 struct kvm_coalesced_mmio_zone zone;
861
862 zone.addr = start;
863 zone.size = size;
864 zone.pad = 0;
865
866 (void)kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone);
867 }
868 }
869
870 static void kvm_uncoalesce_mmio_region(MemoryListener *listener,
871 MemoryRegionSection *secion,
872 hwaddr start, hwaddr size)
873 {
874 KVMState *s = kvm_state;
875
876 if (s->coalesced_mmio) {
877 struct kvm_coalesced_mmio_zone zone;
878
879 zone.addr = start;
880 zone.size = size;
881 zone.pad = 0;
882
883 (void)kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone);
884 }
885 }
886
887 static void kvm_coalesce_pio_add(MemoryListener *listener,
888 MemoryRegionSection *section,
889 hwaddr start, hwaddr size)
890 {
891 KVMState *s = kvm_state;
892
893 if (s->coalesced_pio) {
894 struct kvm_coalesced_mmio_zone zone;
895
896 zone.addr = start;
897 zone.size = size;
898 zone.pio = 1;
899
900 (void)kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone);
901 }
902 }
903
904 static void kvm_coalesce_pio_del(MemoryListener *listener,
905 MemoryRegionSection *section,
906 hwaddr start, hwaddr size)
907 {
908 KVMState *s = kvm_state;
909
910 if (s->coalesced_pio) {
911 struct kvm_coalesced_mmio_zone zone;
912
913 zone.addr = start;
914 zone.size = size;
915 zone.pio = 1;
916
917 (void)kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone);
918 }
919 }
920
921 static MemoryListener kvm_coalesced_pio_listener = {
922 .coalesced_io_add = kvm_coalesce_pio_add,
923 .coalesced_io_del = kvm_coalesce_pio_del,
924 };
925
926 int kvm_check_extension(KVMState *s, unsigned int extension)
927 {
928 int ret;
929
930 ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, extension);
931 if (ret < 0) {
932 ret = 0;
933 }
934
935 return ret;
936 }
937
938 int kvm_vm_check_extension(KVMState *s, unsigned int extension)
939 {
940 int ret;
941
942 ret = kvm_vm_ioctl(s, KVM_CHECK_EXTENSION, extension);
943 if (ret < 0) {
944 /* VM wide version not implemented, use global one instead */
945 ret = kvm_check_extension(s, extension);
946 }
947
948 return ret;
949 }
950
951 typedef struct HWPoisonPage {
952 ram_addr_t ram_addr;
953 QLIST_ENTRY(HWPoisonPage) list;
954 } HWPoisonPage;
955
956 static QLIST_HEAD(, HWPoisonPage) hwpoison_page_list =
957 QLIST_HEAD_INITIALIZER(hwpoison_page_list);
958
959 static void kvm_unpoison_all(void *param)
960 {
961 HWPoisonPage *page, *next_page;
962
963 QLIST_FOREACH_SAFE(page, &hwpoison_page_list, list, next_page) {
964 QLIST_REMOVE(page, list);
965 qemu_ram_remap(page->ram_addr, TARGET_PAGE_SIZE);
966 g_free(page);
967 }
968 }
969
970 void kvm_hwpoison_page_add(ram_addr_t ram_addr)
971 {
972 HWPoisonPage *page;
973
974 QLIST_FOREACH(page, &hwpoison_page_list, list) {
975 if (page->ram_addr == ram_addr) {
976 return;
977 }
978 }
979 page = g_new(HWPoisonPage, 1);
980 page->ram_addr = ram_addr;
981 QLIST_INSERT_HEAD(&hwpoison_page_list, page, list);
982 }
983
984 static uint32_t adjust_ioeventfd_endianness(uint32_t val, uint32_t size)
985 {
986 #if defined(HOST_WORDS_BIGENDIAN) != defined(TARGET_WORDS_BIGENDIAN)
987 /* The kernel expects ioeventfd values in HOST_WORDS_BIGENDIAN
988 * endianness, but the memory core hands them in target endianness.
989 * For example, PPC is always treated as big-endian even if running
990 * on KVM and on PPC64LE. Correct here.
991 */
992 switch (size) {
993 case 2:
994 val = bswap16(val);
995 break;
996 case 4:
997 val = bswap32(val);
998 break;
999 }
1000 #endif
1001 return val;
1002 }
1003
1004 static int kvm_set_ioeventfd_mmio(int fd, hwaddr addr, uint32_t val,
1005 bool assign, uint32_t size, bool datamatch)
1006 {
1007 int ret;
1008 struct kvm_ioeventfd iofd = {
1009 .datamatch = datamatch ? adjust_ioeventfd_endianness(val, size) : 0,
1010 .addr = addr,
1011 .len = size,
1012 .flags = 0,
1013 .fd = fd,
1014 };
1015
1016 trace_kvm_set_ioeventfd_mmio(fd, (uint64_t)addr, val, assign, size,
1017 datamatch);
1018 if (!kvm_enabled()) {
1019 return -ENOSYS;
1020 }
1021
1022 if (datamatch) {
1023 iofd.flags |= KVM_IOEVENTFD_FLAG_DATAMATCH;
1024 }
1025 if (!assign) {
1026 iofd.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
1027 }
1028
1029 ret = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &iofd);
1030
1031 if (ret < 0) {
1032 return -errno;
1033 }
1034
1035 return 0;
1036 }
1037
1038 static int kvm_set_ioeventfd_pio(int fd, uint16_t addr, uint16_t val,
1039 bool assign, uint32_t size, bool datamatch)
1040 {
1041 struct kvm_ioeventfd kick = {
1042 .datamatch = datamatch ? adjust_ioeventfd_endianness(val, size) : 0,
1043 .addr = addr,
1044 .flags = KVM_IOEVENTFD_FLAG_PIO,
1045 .len = size,
1046 .fd = fd,
1047 };
1048 int r;
1049 trace_kvm_set_ioeventfd_pio(fd, addr, val, assign, size, datamatch);
1050 if (!kvm_enabled()) {
1051 return -ENOSYS;
1052 }
1053 if (datamatch) {
1054 kick.flags |= KVM_IOEVENTFD_FLAG_DATAMATCH;
1055 }
1056 if (!assign) {
1057 kick.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
1058 }
1059 r = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &kick);
1060 if (r < 0) {
1061 return r;
1062 }
1063 return 0;
1064 }
1065
1066
1067 static int kvm_check_many_ioeventfds(void)
1068 {
1069 /* Userspace can use ioeventfd for io notification. This requires a host
1070 * that supports eventfd(2) and an I/O thread; since eventfd does not
1071 * support SIGIO it cannot interrupt the vcpu.
1072 *
1073 * Older kernels have a 6 device limit on the KVM io bus. Find out so we
1074 * can avoid creating too many ioeventfds.
1075 */
1076 #if defined(CONFIG_EVENTFD)
1077 int ioeventfds[7];
1078 int i, ret = 0;
1079 for (i = 0; i < ARRAY_SIZE(ioeventfds); i++) {
1080 ioeventfds[i] = eventfd(0, EFD_CLOEXEC);
1081 if (ioeventfds[i] < 0) {
1082 break;
1083 }
1084 ret = kvm_set_ioeventfd_pio(ioeventfds[i], 0, i, true, 2, true);
1085 if (ret < 0) {
1086 close(ioeventfds[i]);
1087 break;
1088 }
1089 }
1090
1091 /* Decide whether many devices are supported or not */
1092 ret = i == ARRAY_SIZE(ioeventfds);
1093
1094 while (i-- > 0) {
1095 kvm_set_ioeventfd_pio(ioeventfds[i], 0, i, false, 2, true);
1096 close(ioeventfds[i]);
1097 }
1098 return ret;
1099 #else
1100 return 0;
1101 #endif
1102 }
1103
1104 static const KVMCapabilityInfo *
1105 kvm_check_extension_list(KVMState *s, const KVMCapabilityInfo *list)
1106 {
1107 while (list->name) {
1108 if (!kvm_check_extension(s, list->value)) {
1109 return list;
1110 }
1111 list++;
1112 }
1113 return NULL;
1114 }
1115
1116 void kvm_set_max_memslot_size(hwaddr max_slot_size)
1117 {
1118 g_assert(
1119 ROUND_UP(max_slot_size, qemu_real_host_page_size) == max_slot_size
1120 );
1121 kvm_max_slot_size = max_slot_size;
1122 }
1123
1124 static void kvm_set_phys_mem(KVMMemoryListener *kml,
1125 MemoryRegionSection *section, bool add)
1126 {
1127 KVMSlot *mem;
1128 int err;
1129 MemoryRegion *mr = section->mr;
1130 bool writeable = !mr->readonly && !mr->rom_device;
1131 hwaddr start_addr, size, slot_size;
1132 void *ram;
1133
1134 if (!memory_region_is_ram(mr)) {
1135 if (writeable || !kvm_readonly_mem_allowed) {
1136 return;
1137 } else if (!mr->romd_mode) {
1138 /* If the memory device is not in romd_mode, then we actually want
1139 * to remove the kvm memory slot so all accesses will trap. */
1140 add = false;
1141 }
1142 }
1143
1144 size = kvm_align_section(section, &start_addr);
1145 if (!size) {
1146 return;
1147 }
1148
1149 /* use aligned delta to align the ram address */
1150 ram = memory_region_get_ram_ptr(mr) + section->offset_within_region +
1151 (start_addr - section->offset_within_address_space);
1152
1153 kvm_slots_lock(kml);
1154
1155 if (!add) {
1156 do {
1157 slot_size = MIN(kvm_max_slot_size, size);
1158 mem = kvm_lookup_matching_slot(kml, start_addr, slot_size);
1159 if (!mem) {
1160 goto out;
1161 }
1162 if (mem->flags & KVM_MEM_LOG_DIRTY_PAGES) {
1163 kvm_physical_sync_dirty_bitmap(kml, section);
1164 }
1165
1166 /* unregister the slot */
1167 g_free(mem->dirty_bmap);
1168 mem->dirty_bmap = NULL;
1169 mem->memory_size = 0;
1170 mem->flags = 0;
1171 err = kvm_set_user_memory_region(kml, mem, false);
1172 if (err) {
1173 fprintf(stderr, "%s: error unregistering slot: %s\n",
1174 __func__, strerror(-err));
1175 abort();
1176 }
1177 start_addr += slot_size;
1178 size -= slot_size;
1179 } while (size);
1180 goto out;
1181 }
1182
1183 /* register the new slot */
1184 do {
1185 slot_size = MIN(kvm_max_slot_size, size);
1186 mem = kvm_alloc_slot(kml);
1187 mem->memory_size = slot_size;
1188 mem->start_addr = start_addr;
1189 mem->ram = ram;
1190 mem->flags = kvm_mem_flags(mr);
1191
1192 if (mem->flags & KVM_MEM_LOG_DIRTY_PAGES) {
1193 /*
1194 * Reallocate the bmap; it means it doesn't disappear in
1195 * middle of a migrate.
1196 */
1197 kvm_memslot_init_dirty_bitmap(mem);
1198 }
1199 err = kvm_set_user_memory_region(kml, mem, true);
1200 if (err) {
1201 fprintf(stderr, "%s: error registering slot: %s\n", __func__,
1202 strerror(-err));
1203 abort();
1204 }
1205 start_addr += slot_size;
1206 ram += slot_size;
1207 size -= slot_size;
1208 } while (size);
1209
1210 out:
1211 kvm_slots_unlock(kml);
1212 }
1213
1214 static void kvm_region_add(MemoryListener *listener,
1215 MemoryRegionSection *section)
1216 {
1217 KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
1218
1219 memory_region_ref(section->mr);
1220 kvm_set_phys_mem(kml, section, true);
1221 }
1222
1223 static void kvm_region_del(MemoryListener *listener,
1224 MemoryRegionSection *section)
1225 {
1226 KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
1227
1228 kvm_set_phys_mem(kml, section, false);
1229 memory_region_unref(section->mr);
1230 }
1231
1232 static void kvm_log_sync(MemoryListener *listener,
1233 MemoryRegionSection *section)
1234 {
1235 KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
1236 int r;
1237
1238 kvm_slots_lock(kml);
1239 r = kvm_physical_sync_dirty_bitmap(kml, section);
1240 kvm_slots_unlock(kml);
1241 if (r < 0) {
1242 abort();
1243 }
1244 }
1245
1246 static void kvm_log_clear(MemoryListener *listener,
1247 MemoryRegionSection *section)
1248 {
1249 KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
1250 int r;
1251
1252 r = kvm_physical_log_clear(kml, section);
1253 if (r < 0) {
1254 error_report_once("%s: kvm log clear failed: mr=%s "
1255 "offset=%"HWADDR_PRIx" size=%"PRIx64, __func__,
1256 section->mr->name, section->offset_within_region,
1257 int128_get64(section->size));
1258 abort();
1259 }
1260 }
1261
1262 static void kvm_mem_ioeventfd_add(MemoryListener *listener,
1263 MemoryRegionSection *section,
1264 bool match_data, uint64_t data,
1265 EventNotifier *e)
1266 {
1267 int fd = event_notifier_get_fd(e);
1268 int r;
1269
1270 r = kvm_set_ioeventfd_mmio(fd, section->offset_within_address_space,
1271 data, true, int128_get64(section->size),
1272 match_data);
1273 if (r < 0) {
1274 fprintf(stderr, "%s: error adding ioeventfd: %s (%d)\n",
1275 __func__, strerror(-r), -r);
1276 abort();
1277 }
1278 }
1279
1280 static void kvm_mem_ioeventfd_del(MemoryListener *listener,
1281 MemoryRegionSection *section,
1282 bool match_data, uint64_t data,
1283 EventNotifier *e)
1284 {
1285 int fd = event_notifier_get_fd(e);
1286 int r;
1287
1288 r = kvm_set_ioeventfd_mmio(fd, section->offset_within_address_space,
1289 data, false, int128_get64(section->size),
1290 match_data);
1291 if (r < 0) {
1292 fprintf(stderr, "%s: error deleting ioeventfd: %s (%d)\n",
1293 __func__, strerror(-r), -r);
1294 abort();
1295 }
1296 }
1297
1298 static void kvm_io_ioeventfd_add(MemoryListener *listener,
1299 MemoryRegionSection *section,
1300 bool match_data, uint64_t data,
1301 EventNotifier *e)
1302 {
1303 int fd = event_notifier_get_fd(e);
1304 int r;
1305
1306 r = kvm_set_ioeventfd_pio(fd, section->offset_within_address_space,
1307 data, true, int128_get64(section->size),
1308 match_data);
1309 if (r < 0) {
1310 fprintf(stderr, "%s: error adding ioeventfd: %s (%d)\n",
1311 __func__, strerror(-r), -r);
1312 abort();
1313 }
1314 }
1315
1316 static void kvm_io_ioeventfd_del(MemoryListener *listener,
1317 MemoryRegionSection *section,
1318 bool match_data, uint64_t data,
1319 EventNotifier *e)
1320
1321 {
1322 int fd = event_notifier_get_fd(e);
1323 int r;
1324
1325 r = kvm_set_ioeventfd_pio(fd, section->offset_within_address_space,
1326 data, false, int128_get64(section->size),
1327 match_data);
1328 if (r < 0) {
1329 fprintf(stderr, "%s: error deleting ioeventfd: %s (%d)\n",
1330 __func__, strerror(-r), -r);
1331 abort();
1332 }
1333 }
1334
1335 void kvm_memory_listener_register(KVMState *s, KVMMemoryListener *kml,
1336 AddressSpace *as, int as_id)
1337 {
1338 int i;
1339
1340 qemu_mutex_init(&kml->slots_lock);
1341 kml->slots = g_malloc0(s->nr_slots * sizeof(KVMSlot));
1342 kml->as_id = as_id;
1343
1344 for (i = 0; i < s->nr_slots; i++) {
1345 kml->slots[i].slot = i;
1346 }
1347
1348 kml->listener.region_add = kvm_region_add;
1349 kml->listener.region_del = kvm_region_del;
1350 kml->listener.log_start = kvm_log_start;
1351 kml->listener.log_stop = kvm_log_stop;
1352 kml->listener.log_sync = kvm_log_sync;
1353 kml->listener.log_clear = kvm_log_clear;
1354 kml->listener.priority = 10;
1355
1356 memory_listener_register(&kml->listener, as);
1357
1358 for (i = 0; i < s->nr_as; ++i) {
1359 if (!s->as[i].as) {
1360 s->as[i].as = as;
1361 s->as[i].ml = kml;
1362 break;
1363 }
1364 }
1365 }
1366
1367 static MemoryListener kvm_io_listener = {
1368 .eventfd_add = kvm_io_ioeventfd_add,
1369 .eventfd_del = kvm_io_ioeventfd_del,
1370 .priority = 10,
1371 };
1372
1373 int kvm_set_irq(KVMState *s, int irq, int level)
1374 {
1375 struct kvm_irq_level event;
1376 int ret;
1377
1378 assert(kvm_async_interrupts_enabled());
1379
1380 event.level = level;
1381 event.irq = irq;
1382 ret = kvm_vm_ioctl(s, s->irq_set_ioctl, &event);
1383 if (ret < 0) {
1384 perror("kvm_set_irq");
1385 abort();
1386 }
1387
1388 return (s->irq_set_ioctl == KVM_IRQ_LINE) ? 1 : event.status;
1389 }
1390
1391 #ifdef KVM_CAP_IRQ_ROUTING
1392 typedef struct KVMMSIRoute {
1393 struct kvm_irq_routing_entry kroute;
1394 QTAILQ_ENTRY(KVMMSIRoute) entry;
1395 } KVMMSIRoute;
1396
1397 static void set_gsi(KVMState *s, unsigned int gsi)
1398 {
1399 set_bit(gsi, s->used_gsi_bitmap);
1400 }
1401
1402 static void clear_gsi(KVMState *s, unsigned int gsi)
1403 {
1404 clear_bit(gsi, s->used_gsi_bitmap);
1405 }
1406
1407 void kvm_init_irq_routing(KVMState *s)
1408 {
1409 int gsi_count, i;
1410
1411 gsi_count = kvm_check_extension(s, KVM_CAP_IRQ_ROUTING) - 1;
1412 if (gsi_count > 0) {
1413 /* Round up so we can search ints using ffs */
1414 s->used_gsi_bitmap = bitmap_new(gsi_count);
1415 s->gsi_count = gsi_count;
1416 }
1417
1418 s->irq_routes = g_malloc0(sizeof(*s->irq_routes));
1419 s->nr_allocated_irq_routes = 0;
1420
1421 if (!kvm_direct_msi_allowed) {
1422 for (i = 0; i < KVM_MSI_HASHTAB_SIZE; i++) {
1423 QTAILQ_INIT(&s->msi_hashtab[i]);
1424 }
1425 }
1426
1427 kvm_arch_init_irq_routing(s);
1428 }
1429
1430 void kvm_irqchip_commit_routes(KVMState *s)
1431 {
1432 int ret;
1433
1434 if (kvm_gsi_direct_mapping()) {
1435 return;
1436 }
1437
1438 if (!kvm_gsi_routing_enabled()) {
1439 return;
1440 }
1441
1442 s->irq_routes->flags = 0;
1443 trace_kvm_irqchip_commit_routes();
1444 ret = kvm_vm_ioctl(s, KVM_SET_GSI_ROUTING, s->irq_routes);
1445 assert(ret == 0);
1446 }
1447
1448 static void kvm_add_routing_entry(KVMState *s,
1449 struct kvm_irq_routing_entry *entry)
1450 {
1451 struct kvm_irq_routing_entry *new;
1452 int n, size;
1453
1454 if (s->irq_routes->nr == s->nr_allocated_irq_routes) {
1455 n = s->nr_allocated_irq_routes * 2;
1456 if (n < 64) {
1457 n = 64;
1458 }
1459 size = sizeof(struct kvm_irq_routing);
1460 size += n * sizeof(*new);
1461 s->irq_routes = g_realloc(s->irq_routes, size);
1462 s->nr_allocated_irq_routes = n;
1463 }
1464 n = s->irq_routes->nr++;
1465 new = &s->irq_routes->entries[n];
1466
1467 *new = *entry;
1468
1469 set_gsi(s, entry->gsi);
1470 }
1471
1472 static int kvm_update_routing_entry(KVMState *s,
1473 struct kvm_irq_routing_entry *new_entry)
1474 {
1475 struct kvm_irq_routing_entry *entry;
1476 int n;
1477
1478 for (n = 0; n < s->irq_routes->nr; n++) {
1479 entry = &s->irq_routes->entries[n];
1480 if (entry->gsi != new_entry->gsi) {
1481 continue;
1482 }
1483
1484 if(!memcmp(entry, new_entry, sizeof *entry)) {
1485 return 0;
1486 }
1487
1488 *entry = *new_entry;
1489
1490 return 0;
1491 }
1492
1493 return -ESRCH;
1494 }
1495
1496 void kvm_irqchip_add_irq_route(KVMState *s, int irq, int irqchip, int pin)
1497 {
1498 struct kvm_irq_routing_entry e = {};
1499
1500 assert(pin < s->gsi_count);
1501
1502 e.gsi = irq;
1503 e.type = KVM_IRQ_ROUTING_IRQCHIP;
1504 e.flags = 0;
1505 e.u.irqchip.irqchip = irqchip;
1506 e.u.irqchip.pin = pin;
1507 kvm_add_routing_entry(s, &e);
1508 }
1509
1510 void kvm_irqchip_release_virq(KVMState *s, int virq)
1511 {
1512 struct kvm_irq_routing_entry *e;
1513 int i;
1514
1515 if (kvm_gsi_direct_mapping()) {
1516 return;
1517 }
1518
1519 for (i = 0; i < s->irq_routes->nr; i++) {
1520 e = &s->irq_routes->entries[i];
1521 if (e->gsi == virq) {
1522 s->irq_routes->nr--;
1523 *e = s->irq_routes->entries[s->irq_routes->nr];
1524 }
1525 }
1526 clear_gsi(s, virq);
1527 kvm_arch_release_virq_post(virq);
1528 trace_kvm_irqchip_release_virq(virq);
1529 }
1530
1531 void kvm_irqchip_add_change_notifier(Notifier *n)
1532 {
1533 notifier_list_add(&kvm_irqchip_change_notifiers, n);
1534 }
1535
1536 void kvm_irqchip_remove_change_notifier(Notifier *n)
1537 {
1538 notifier_remove(n);
1539 }
1540
1541 void kvm_irqchip_change_notify(void)
1542 {
1543 notifier_list_notify(&kvm_irqchip_change_notifiers, NULL);
1544 }
1545
1546 static unsigned int kvm_hash_msi(uint32_t data)
1547 {
1548 /* This is optimized for IA32 MSI layout. However, no other arch shall
1549 * repeat the mistake of not providing a direct MSI injection API. */
1550 return data & 0xff;
1551 }
1552
1553 static void kvm_flush_dynamic_msi_routes(KVMState *s)
1554 {
1555 KVMMSIRoute *route, *next;
1556 unsigned int hash;
1557
1558 for (hash = 0; hash < KVM_MSI_HASHTAB_SIZE; hash++) {
1559 QTAILQ_FOREACH_SAFE(route, &s->msi_hashtab[hash], entry, next) {
1560 kvm_irqchip_release_virq(s, route->kroute.gsi);
1561 QTAILQ_REMOVE(&s->msi_hashtab[hash], route, entry);
1562 g_free(route);
1563 }
1564 }
1565 }
1566
1567 static int kvm_irqchip_get_virq(KVMState *s)
1568 {
1569 int next_virq;
1570
1571 /*
1572 * PIC and IOAPIC share the first 16 GSI numbers, thus the available
1573 * GSI numbers are more than the number of IRQ route. Allocating a GSI
1574 * number can succeed even though a new route entry cannot be added.
1575 * When this happens, flush dynamic MSI entries to free IRQ route entries.
1576 */
1577 if (!kvm_direct_msi_allowed && s->irq_routes->nr == s->gsi_count) {
1578 kvm_flush_dynamic_msi_routes(s);
1579 }
1580
1581 /* Return the lowest unused GSI in the bitmap */
1582 next_virq = find_first_zero_bit(s->used_gsi_bitmap, s->gsi_count);
1583 if (next_virq >= s->gsi_count) {
1584 return -ENOSPC;
1585 } else {
1586 return next_virq;
1587 }
1588 }
1589
1590 static KVMMSIRoute *kvm_lookup_msi_route(KVMState *s, MSIMessage msg)
1591 {
1592 unsigned int hash = kvm_hash_msi(msg.data);
1593 KVMMSIRoute *route;
1594
1595 QTAILQ_FOREACH(route, &s->msi_hashtab[hash], entry) {
1596 if (route->kroute.u.msi.address_lo == (uint32_t)msg.address &&
1597 route->kroute.u.msi.address_hi == (msg.address >> 32) &&
1598 route->kroute.u.msi.data == le32_to_cpu(msg.data)) {
1599 return route;
1600 }
1601 }
1602 return NULL;
1603 }
1604
1605 int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg)
1606 {
1607 struct kvm_msi msi;
1608 KVMMSIRoute *route;
1609
1610 if (kvm_direct_msi_allowed) {
1611 msi.address_lo = (uint32_t)msg.address;
1612 msi.address_hi = msg.address >> 32;
1613 msi.data = le32_to_cpu(msg.data);
1614 msi.flags = 0;
1615 memset(msi.pad, 0, sizeof(msi.pad));
1616
1617 return kvm_vm_ioctl(s, KVM_SIGNAL_MSI, &msi);
1618 }
1619
1620 route = kvm_lookup_msi_route(s, msg);
1621 if (!route) {
1622 int virq;
1623
1624 virq = kvm_irqchip_get_virq(s);
1625 if (virq < 0) {
1626 return virq;
1627 }
1628
1629 route = g_malloc0(sizeof(KVMMSIRoute));
1630 route->kroute.gsi = virq;
1631 route->kroute.type = KVM_IRQ_ROUTING_MSI;
1632 route->kroute.flags = 0;
1633 route->kroute.u.msi.address_lo = (uint32_t)msg.address;
1634 route->kroute.u.msi.address_hi = msg.address >> 32;
1635 route->kroute.u.msi.data = le32_to_cpu(msg.data);
1636
1637 kvm_add_routing_entry(s, &route->kroute);
1638 kvm_irqchip_commit_routes(s);
1639
1640 QTAILQ_INSERT_TAIL(&s->msi_hashtab[kvm_hash_msi(msg.data)], route,
1641 entry);
1642 }
1643
1644 assert(route->kroute.type == KVM_IRQ_ROUTING_MSI);
1645
1646 return kvm_set_irq(s, route->kroute.gsi, 1);
1647 }
1648
1649 int kvm_irqchip_add_msi_route(KVMState *s, int vector, PCIDevice *dev)
1650 {
1651 struct kvm_irq_routing_entry kroute = {};
1652 int virq;
1653 MSIMessage msg = {0, 0};
1654
1655 if (pci_available && dev) {
1656 msg = pci_get_msi_message(dev, vector);
1657 }
1658
1659 if (kvm_gsi_direct_mapping()) {
1660 return kvm_arch_msi_data_to_gsi(msg.data);
1661 }
1662
1663 if (!kvm_gsi_routing_enabled()) {
1664 return -ENOSYS;
1665 }
1666
1667 virq = kvm_irqchip_get_virq(s);
1668 if (virq < 0) {
1669 return virq;
1670 }
1671
1672 kroute.gsi = virq;
1673 kroute.type = KVM_IRQ_ROUTING_MSI;
1674 kroute.flags = 0;
1675 kroute.u.msi.address_lo = (uint32_t)msg.address;
1676 kroute.u.msi.address_hi = msg.address >> 32;
1677 kroute.u.msi.data = le32_to_cpu(msg.data);
1678 if (pci_available && kvm_msi_devid_required()) {
1679 kroute.flags = KVM_MSI_VALID_DEVID;
1680 kroute.u.msi.devid = pci_requester_id(dev);
1681 }
1682 if (kvm_arch_fixup_msi_route(&kroute, msg.address, msg.data, dev)) {
1683 kvm_irqchip_release_virq(s, virq);
1684 return -EINVAL;
1685 }
1686
1687 trace_kvm_irqchip_add_msi_route(dev ? dev->name : (char *)"N/A",
1688 vector, virq);
1689
1690 kvm_add_routing_entry(s, &kroute);
1691 kvm_arch_add_msi_route_post(&kroute, vector, dev);
1692 kvm_irqchip_commit_routes(s);
1693
1694 return virq;
1695 }
1696
1697 int kvm_irqchip_update_msi_route(KVMState *s, int virq, MSIMessage msg,
1698 PCIDevice *dev)
1699 {
1700 struct kvm_irq_routing_entry kroute = {};
1701
1702 if (kvm_gsi_direct_mapping()) {
1703 return 0;
1704 }
1705
1706 if (!kvm_irqchip_in_kernel()) {
1707 return -ENOSYS;
1708 }
1709
1710 kroute.gsi = virq;
1711 kroute.type = KVM_IRQ_ROUTING_MSI;
1712 kroute.flags = 0;
1713 kroute.u.msi.address_lo = (uint32_t)msg.address;
1714 kroute.u.msi.address_hi = msg.address >> 32;
1715 kroute.u.msi.data = le32_to_cpu(msg.data);
1716 if (pci_available && kvm_msi_devid_required()) {
1717 kroute.flags = KVM_MSI_VALID_DEVID;
1718 kroute.u.msi.devid = pci_requester_id(dev);
1719 }
1720 if (kvm_arch_fixup_msi_route(&kroute, msg.address, msg.data, dev)) {
1721 return -EINVAL;
1722 }
1723
1724 trace_kvm_irqchip_update_msi_route(virq);
1725
1726 return kvm_update_routing_entry(s, &kroute);
1727 }
1728
1729 static int kvm_irqchip_assign_irqfd(KVMState *s, EventNotifier *event,
1730 EventNotifier *resample, int virq,
1731 bool assign)
1732 {
1733 int fd = event_notifier_get_fd(event);
1734 int rfd = resample ? event_notifier_get_fd(resample) : -1;
1735
1736 struct kvm_irqfd irqfd = {
1737 .fd = fd,
1738 .gsi = virq,
1739 .flags = assign ? 0 : KVM_IRQFD_FLAG_DEASSIGN,
1740 };
1741
1742 if (rfd != -1) {
1743 assert(assign);
1744 if (kvm_irqchip_is_split()) {
1745 /*
1746 * When the slow irqchip (e.g. IOAPIC) is in the
1747 * userspace, KVM kernel resamplefd will not work because
1748 * the EOI of the interrupt will be delivered to userspace
1749 * instead, so the KVM kernel resamplefd kick will be
1750 * skipped. The userspace here mimics what the kernel
1751 * provides with resamplefd, remember the resamplefd and
1752 * kick it when we receive EOI of this IRQ.
1753 *
1754 * This is hackery because IOAPIC is mostly bypassed
1755 * (except EOI broadcasts) when irqfd is used. However
1756 * this can bring much performance back for split irqchip
1757 * with INTx IRQs (for VFIO, this gives 93% perf of the
1758 * full fast path, which is 46% perf boost comparing to
1759 * the INTx slow path).
1760 */
1761 kvm_resample_fd_insert(virq, resample);
1762 } else {
1763 irqfd.flags |= KVM_IRQFD_FLAG_RESAMPLE;
1764 irqfd.resamplefd = rfd;
1765 }
1766 } else if (!assign) {
1767 if (kvm_irqchip_is_split()) {
1768 kvm_resample_fd_remove(virq);
1769 }
1770 }
1771
1772 if (!kvm_irqfds_enabled()) {
1773 return -ENOSYS;
1774 }
1775
1776 return kvm_vm_ioctl(s, KVM_IRQFD, &irqfd);
1777 }
1778
1779 int kvm_irqchip_add_adapter_route(KVMState *s, AdapterInfo *adapter)
1780 {
1781 struct kvm_irq_routing_entry kroute = {};
1782 int virq;
1783
1784 if (!kvm_gsi_routing_enabled()) {
1785 return -ENOSYS;
1786 }
1787
1788 virq = kvm_irqchip_get_virq(s);
1789 if (virq < 0) {
1790 return virq;
1791 }
1792
1793 kroute.gsi = virq;
1794 kroute.type = KVM_IRQ_ROUTING_S390_ADAPTER;
1795 kroute.flags = 0;
1796 kroute.u.adapter.summary_addr = adapter->summary_addr;
1797 kroute.u.adapter.ind_addr = adapter->ind_addr;
1798 kroute.u.adapter.summary_offset = adapter->summary_offset;
1799 kroute.u.adapter.ind_offset = adapter->ind_offset;
1800 kroute.u.adapter.adapter_id = adapter->adapter_id;
1801
1802 kvm_add_routing_entry(s, &kroute);
1803
1804 return virq;
1805 }
1806
1807 int kvm_irqchip_add_hv_sint_route(KVMState *s, uint32_t vcpu, uint32_t sint)
1808 {
1809 struct kvm_irq_routing_entry kroute = {};
1810 int virq;
1811
1812 if (!kvm_gsi_routing_enabled()) {
1813 return -ENOSYS;
1814 }
1815 if (!kvm_check_extension(s, KVM_CAP_HYPERV_SYNIC)) {
1816 return -ENOSYS;
1817 }
1818 virq = kvm_irqchip_get_virq(s);
1819 if (virq < 0) {
1820 return virq;
1821 }
1822
1823 kroute.gsi = virq;
1824 kroute.type = KVM_IRQ_ROUTING_HV_SINT;
1825 kroute.flags = 0;
1826 kroute.u.hv_sint.vcpu = vcpu;
1827 kroute.u.hv_sint.sint = sint;
1828
1829 kvm_add_routing_entry(s, &kroute);
1830 kvm_irqchip_commit_routes(s);
1831
1832 return virq;
1833 }
1834
1835 #else /* !KVM_CAP_IRQ_ROUTING */
1836
1837 void kvm_init_irq_routing(KVMState *s)
1838 {
1839 }
1840
1841 void kvm_irqchip_release_virq(KVMState *s, int virq)
1842 {
1843 }
1844
1845 int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg)
1846 {
1847 abort();
1848 }
1849
1850 int kvm_irqchip_add_msi_route(KVMState *s, int vector, PCIDevice *dev)
1851 {
1852 return -ENOSYS;
1853 }
1854
1855 int kvm_irqchip_add_adapter_route(KVMState *s, AdapterInfo *adapter)
1856 {
1857 return -ENOSYS;
1858 }
1859
1860 int kvm_irqchip_add_hv_sint_route(KVMState *s, uint32_t vcpu, uint32_t sint)
1861 {
1862 return -ENOSYS;
1863 }
1864
1865 static int kvm_irqchip_assign_irqfd(KVMState *s, EventNotifier *event,
1866 EventNotifier *resample, int virq,
1867 bool assign)
1868 {
1869 abort();
1870 }
1871
1872 int kvm_irqchip_update_msi_route(KVMState *s, int virq, MSIMessage msg)
1873 {
1874 return -ENOSYS;
1875 }
1876 #endif /* !KVM_CAP_IRQ_ROUTING */
1877
1878 int kvm_irqchip_add_irqfd_notifier_gsi(KVMState *s, EventNotifier *n,
1879 EventNotifier *rn, int virq)
1880 {
1881 return kvm_irqchip_assign_irqfd(s, n, rn, virq, true);
1882 }
1883
1884 int kvm_irqchip_remove_irqfd_notifier_gsi(KVMState *s, EventNotifier *n,
1885 int virq)
1886 {
1887 return kvm_irqchip_assign_irqfd(s, n, NULL, virq, false);
1888 }
1889
1890 int kvm_irqchip_add_irqfd_notifier(KVMState *s, EventNotifier *n,
1891 EventNotifier *rn, qemu_irq irq)
1892 {
1893 gpointer key, gsi;
1894 gboolean found = g_hash_table_lookup_extended(s->gsimap, irq, &key, &gsi);
1895
1896 if (!found) {
1897 return -ENXIO;
1898 }
1899 return kvm_irqchip_add_irqfd_notifier_gsi(s, n, rn, GPOINTER_TO_INT(gsi));
1900 }
1901
1902 int kvm_irqchip_remove_irqfd_notifier(KVMState *s, EventNotifier *n,
1903 qemu_irq irq)
1904 {
1905 gpointer key, gsi;
1906 gboolean found = g_hash_table_lookup_extended(s->gsimap, irq, &key, &gsi);
1907
1908 if (!found) {
1909 return -ENXIO;
1910 }
1911 return kvm_irqchip_remove_irqfd_notifier_gsi(s, n, GPOINTER_TO_INT(gsi));
1912 }
1913
1914 void kvm_irqchip_set_qemuirq_gsi(KVMState *s, qemu_irq irq, int gsi)
1915 {
1916 g_hash_table_insert(s->gsimap, irq, GINT_TO_POINTER(gsi));
1917 }
1918
1919 static void kvm_irqchip_create(KVMState *s)
1920 {
1921 int ret;
1922
1923 assert(s->kernel_irqchip_split != ON_OFF_AUTO_AUTO);
1924 if (kvm_check_extension(s, KVM_CAP_IRQCHIP)) {
1925 ;
1926 } else if (kvm_check_extension(s, KVM_CAP_S390_IRQCHIP)) {
1927 ret = kvm_vm_enable_cap(s, KVM_CAP_S390_IRQCHIP, 0);
1928 if (ret < 0) {
1929 fprintf(stderr, "Enable kernel irqchip failed: %s\n", strerror(-ret));
1930 exit(1);
1931 }
1932 } else {
1933 return;
1934 }
1935
1936 /* First probe and see if there's a arch-specific hook to create the
1937 * in-kernel irqchip for us */
1938 ret = kvm_arch_irqchip_create(s);
1939 if (ret == 0) {
1940 if (s->kernel_irqchip_split == ON_OFF_AUTO_ON) {
1941 perror("Split IRQ chip mode not supported.");
1942 exit(1);
1943 } else {
1944 ret = kvm_vm_ioctl(s, KVM_CREATE_IRQCHIP);
1945 }
1946 }
1947 if (ret < 0) {
1948 fprintf(stderr, "Create kernel irqchip failed: %s\n", strerror(-ret));
1949 exit(1);
1950 }
1951
1952 kvm_kernel_irqchip = true;
1953 /* If we have an in-kernel IRQ chip then we must have asynchronous
1954 * interrupt delivery (though the reverse is not necessarily true)
1955 */
1956 kvm_async_interrupts_allowed = true;
1957 kvm_halt_in_kernel_allowed = true;
1958
1959 kvm_init_irq_routing(s);
1960
1961 s->gsimap = g_hash_table_new(g_direct_hash, g_direct_equal);
1962 }
1963
1964 /* Find number of supported CPUs using the recommended
1965 * procedure from the kernel API documentation to cope with
1966 * older kernels that may be missing capabilities.
1967 */
1968 static int kvm_recommended_vcpus(KVMState *s)
1969 {
1970 int ret = kvm_vm_check_extension(s, KVM_CAP_NR_VCPUS);
1971 return (ret) ? ret : 4;
1972 }
1973
1974 static int kvm_max_vcpus(KVMState *s)
1975 {
1976 int ret = kvm_check_extension(s, KVM_CAP_MAX_VCPUS);
1977 return (ret) ? ret : kvm_recommended_vcpus(s);
1978 }
1979
1980 static int kvm_max_vcpu_id(KVMState *s)
1981 {
1982 int ret = kvm_check_extension(s, KVM_CAP_MAX_VCPU_ID);
1983 return (ret) ? ret : kvm_max_vcpus(s);
1984 }
1985
1986 bool kvm_vcpu_id_is_valid(int vcpu_id)
1987 {
1988 KVMState *s = KVM_STATE(current_accel());
1989 return vcpu_id >= 0 && vcpu_id < kvm_max_vcpu_id(s);
1990 }
1991
1992 static int kvm_init(MachineState *ms)
1993 {
1994 MachineClass *mc = MACHINE_GET_CLASS(ms);
1995 static const char upgrade_note[] =
1996 "Please upgrade to at least kernel 2.6.29 or recent kvm-kmod\n"
1997 "(see http://sourceforge.net/projects/kvm).\n";
1998 struct {
1999 const char *name;
2000 int num;
2001 } num_cpus[] = {
2002 { "SMP", ms->smp.cpus },
2003 { "hotpluggable", ms->smp.max_cpus },
2004 { NULL, }
2005 }, *nc = num_cpus;
2006 int soft_vcpus_limit, hard_vcpus_limit;
2007 KVMState *s;
2008 const KVMCapabilityInfo *missing_cap;
2009 int ret;
2010 int type = 0;
2011 uint64_t dirty_log_manual_caps;
2012
2013 s = KVM_STATE(ms->accelerator);
2014
2015 /*
2016 * On systems where the kernel can support different base page
2017 * sizes, host page size may be different from TARGET_PAGE_SIZE,
2018 * even with KVM. TARGET_PAGE_SIZE is assumed to be the minimum
2019 * page size for the system though.
2020 */
2021 assert(TARGET_PAGE_SIZE <= qemu_real_host_page_size);
2022
2023 s->sigmask_len = 8;
2024
2025 #ifdef KVM_CAP_SET_GUEST_DEBUG
2026 QTAILQ_INIT(&s->kvm_sw_breakpoints);
2027 #endif
2028 QLIST_INIT(&s->kvm_parked_vcpus);
2029 s->vmfd = -1;
2030 s->fd = qemu_open_old("/dev/kvm", O_RDWR);
2031 if (s->fd == -1) {
2032 fprintf(stderr, "Could not access KVM kernel module: %m\n");
2033 ret = -errno;
2034 goto err;
2035 }
2036
2037 ret = kvm_ioctl(s, KVM_GET_API_VERSION, 0);
2038 if (ret < KVM_API_VERSION) {
2039 if (ret >= 0) {
2040 ret = -EINVAL;
2041 }
2042 fprintf(stderr, "kvm version too old\n");
2043 goto err;
2044 }
2045
2046 if (ret > KVM_API_VERSION) {
2047 ret = -EINVAL;
2048 fprintf(stderr, "kvm version not supported\n");
2049 goto err;
2050 }
2051
2052 kvm_immediate_exit = kvm_check_extension(s, KVM_CAP_IMMEDIATE_EXIT);
2053 s->nr_slots = kvm_check_extension(s, KVM_CAP_NR_MEMSLOTS);
2054
2055 /* If unspecified, use the default value */
2056 if (!s->nr_slots) {
2057 s->nr_slots = 32;
2058 }
2059
2060 s->nr_as = kvm_check_extension(s, KVM_CAP_MULTI_ADDRESS_SPACE);
2061 if (s->nr_as <= 1) {
2062 s->nr_as = 1;
2063 }
2064 s->as = g_new0(struct KVMAs, s->nr_as);
2065
2066 if (object_property_find(OBJECT(current_machine), "kvm-type")) {
2067 g_autofree char *kvm_type = object_property_get_str(OBJECT(current_machine),
2068 "kvm-type",
2069 &error_abort);
2070 type = mc->kvm_type(ms, kvm_type);
2071 }
2072
2073 do {
2074 ret = kvm_ioctl(s, KVM_CREATE_VM, type);
2075 } while (ret == -EINTR);
2076
2077 if (ret < 0) {
2078 fprintf(stderr, "ioctl(KVM_CREATE_VM) failed: %d %s\n", -ret,
2079 strerror(-ret));
2080
2081 #ifdef TARGET_S390X
2082 if (ret == -EINVAL) {
2083 fprintf(stderr,
2084 "Host kernel setup problem detected. Please verify:\n");
2085 fprintf(stderr, "- for kernels supporting the switch_amode or"
2086 " user_mode parameters, whether\n");
2087 fprintf(stderr,
2088 " user space is running in primary address space\n");
2089 fprintf(stderr,
2090 "- for kernels supporting the vm.allocate_pgste sysctl, "
2091 "whether it is enabled\n");
2092 }
2093 #endif
2094 goto err;
2095 }
2096
2097 s->vmfd = ret;
2098
2099 /* check the vcpu limits */
2100 soft_vcpus_limit = kvm_recommended_vcpus(s);
2101 hard_vcpus_limit = kvm_max_vcpus(s);
2102
2103 while (nc->name) {
2104 if (nc->num > soft_vcpus_limit) {
2105 warn_report("Number of %s cpus requested (%d) exceeds "
2106 "the recommended cpus supported by KVM (%d)",
2107 nc->name, nc->num, soft_vcpus_limit);
2108
2109 if (nc->num > hard_vcpus_limit) {
2110 fprintf(stderr, "Number of %s cpus requested (%d) exceeds "
2111 "the maximum cpus supported by KVM (%d)\n",
2112 nc->name, nc->num, hard_vcpus_limit);
2113 exit(1);
2114 }
2115 }
2116 nc++;
2117 }
2118
2119 missing_cap = kvm_check_extension_list(s, kvm_required_capabilites);
2120 if (!missing_cap) {
2121 missing_cap =
2122 kvm_check_extension_list(s, kvm_arch_required_capabilities);
2123 }
2124 if (missing_cap) {
2125 ret = -EINVAL;
2126 fprintf(stderr, "kvm does not support %s\n%s",
2127 missing_cap->name, upgrade_note);
2128 goto err;
2129 }
2130
2131 s->coalesced_mmio = kvm_check_extension(s, KVM_CAP_COALESCED_MMIO);
2132 s->coalesced_pio = s->coalesced_mmio &&
2133 kvm_check_extension(s, KVM_CAP_COALESCED_PIO);
2134
2135 dirty_log_manual_caps =
2136 kvm_check_extension(s, KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2);
2137 dirty_log_manual_caps &= (KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE |
2138 KVM_DIRTY_LOG_INITIALLY_SET);
2139 s->manual_dirty_log_protect = dirty_log_manual_caps;
2140 if (dirty_log_manual_caps) {
2141 ret = kvm_vm_enable_cap(s, KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2, 0,
2142 dirty_log_manual_caps);
2143 if (ret) {
2144 warn_report("Trying to enable capability %"PRIu64" of "
2145 "KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2 but failed. "
2146 "Falling back to the legacy mode. ",
2147 dirty_log_manual_caps);
2148 s->manual_dirty_log_protect = 0;
2149 }
2150 }
2151
2152 #ifdef KVM_CAP_VCPU_EVENTS
2153 s->vcpu_events = kvm_check_extension(s, KVM_CAP_VCPU_EVENTS);
2154 #endif
2155
2156 s->robust_singlestep =
2157 kvm_check_extension(s, KVM_CAP_X86_ROBUST_SINGLESTEP);
2158
2159 #ifdef KVM_CAP_DEBUGREGS
2160 s->debugregs = kvm_check_extension(s, KVM_CAP_DEBUGREGS);
2161 #endif
2162
2163 s->max_nested_state_len = kvm_check_extension(s, KVM_CAP_NESTED_STATE);
2164
2165 #ifdef KVM_CAP_IRQ_ROUTING
2166 kvm_direct_msi_allowed = (kvm_check_extension(s, KVM_CAP_SIGNAL_MSI) > 0);
2167 #endif
2168
2169 s->intx_set_mask = kvm_check_extension(s, KVM_CAP_PCI_2_3);
2170
2171 s->irq_set_ioctl = KVM_IRQ_LINE;
2172 if (kvm_check_extension(s, KVM_CAP_IRQ_INJECT_STATUS)) {
2173 s->irq_set_ioctl = KVM_IRQ_LINE_STATUS;
2174 }
2175
2176 kvm_readonly_mem_allowed =
2177 (kvm_check_extension(s, KVM_CAP_READONLY_MEM) > 0);
2178
2179 kvm_eventfds_allowed =
2180 (kvm_check_extension(s, KVM_CAP_IOEVENTFD) > 0);
2181
2182 kvm_irqfds_allowed =
2183 (kvm_check_extension(s, KVM_CAP_IRQFD) > 0);
2184
2185 kvm_resamplefds_allowed =
2186 (kvm_check_extension(s, KVM_CAP_IRQFD_RESAMPLE) > 0);
2187
2188 kvm_vm_attributes_allowed =
2189 (kvm_check_extension(s, KVM_CAP_VM_ATTRIBUTES) > 0);
2190
2191 kvm_ioeventfd_any_length_allowed =
2192 (kvm_check_extension(s, KVM_CAP_IOEVENTFD_ANY_LENGTH) > 0);
2193
2194 kvm_state = s;
2195
2196 ret = kvm_arch_init(ms, s);
2197 if (ret < 0) {
2198 goto err;
2199 }
2200
2201 if (s->kernel_irqchip_split == ON_OFF_AUTO_AUTO) {
2202 s->kernel_irqchip_split = mc->default_kernel_irqchip_split ? ON_OFF_AUTO_ON : ON_OFF_AUTO_OFF;
2203 }
2204
2205 qemu_register_reset(kvm_unpoison_all, NULL);
2206
2207 if (s->kernel_irqchip_allowed) {
2208 kvm_irqchip_create(s);
2209 }
2210
2211 if (kvm_eventfds_allowed) {
2212 s->memory_listener.listener.eventfd_add = kvm_mem_ioeventfd_add;
2213 s->memory_listener.listener.eventfd_del = kvm_mem_ioeventfd_del;
2214 }
2215 s->memory_listener.listener.coalesced_io_add = kvm_coalesce_mmio_region;
2216 s->memory_listener.listener.coalesced_io_del = kvm_uncoalesce_mmio_region;
2217
2218 kvm_memory_listener_register(s, &s->memory_listener,
2219 &address_space_memory, 0);
2220 if (kvm_eventfds_allowed) {
2221 memory_listener_register(&kvm_io_listener,
2222 &address_space_io);
2223 }
2224 memory_listener_register(&kvm_coalesced_pio_listener,
2225 &address_space_io);
2226
2227 s->many_ioeventfds = kvm_check_many_ioeventfds();
2228
2229 s->sync_mmu = !!kvm_vm_check_extension(kvm_state, KVM_CAP_SYNC_MMU);
2230 if (!s->sync_mmu) {
2231 ret = ram_block_discard_disable(true);
2232 assert(!ret);
2233 }
2234 return 0;
2235
2236 err:
2237 assert(ret < 0);
2238 if (s->vmfd >= 0) {
2239 close(s->vmfd);
2240 }
2241 if (s->fd != -1) {
2242 close(s->fd);
2243 }
2244 g_free(s->memory_listener.slots);
2245
2246 return ret;
2247 }
2248
2249 void kvm_set_sigmask_len(KVMState *s, unsigned int sigmask_len)
2250 {
2251 s->sigmask_len = sigmask_len;
2252 }
2253
2254 static void kvm_handle_io(uint16_t port, MemTxAttrs attrs, void *data, int direction,
2255 int size, uint32_t count)
2256 {
2257 int i;
2258 uint8_t *ptr = data;
2259
2260 for (i = 0; i < count; i++) {
2261 address_space_rw(&address_space_io, port, attrs,
2262 ptr, size,
2263 direction == KVM_EXIT_IO_OUT);
2264 ptr += size;
2265 }
2266 }
2267
2268 static int kvm_handle_internal_error(CPUState *cpu, struct kvm_run *run)
2269 {
2270 fprintf(stderr, "KVM internal error. Suberror: %d\n",
2271 run->internal.suberror);
2272
2273 if (kvm_check_extension(kvm_state, KVM_CAP_INTERNAL_ERROR_DATA)) {
2274 int i;
2275
2276 for (i = 0; i < run->internal.ndata; ++i) {
2277 fprintf(stderr, "extra data[%d]: %"PRIx64"\n",
2278 i, (uint64_t)run->internal.data[i]);
2279 }
2280 }
2281 if (run->internal.suberror == KVM_INTERNAL_ERROR_EMULATION) {
2282 fprintf(stderr, "emulation failure\n");
2283 if (!kvm_arch_stop_on_emulation_error(cpu)) {
2284 cpu_dump_state(cpu, stderr, CPU_DUMP_CODE);
2285 return EXCP_INTERRUPT;
2286 }
2287 }
2288 /* FIXME: Should trigger a qmp message to let management know
2289 * something went wrong.
2290 */
2291 return -1;
2292 }
2293
2294 void kvm_flush_coalesced_mmio_buffer(void)
2295 {
2296 KVMState *s = kvm_state;
2297
2298 if (s->coalesced_flush_in_progress) {
2299 return;
2300 }
2301
2302 s->coalesced_flush_in_progress = true;
2303
2304 if (s->coalesced_mmio_ring) {
2305 struct kvm_coalesced_mmio_ring *ring = s->coalesced_mmio_ring;
2306 while (ring->first != ring->last) {
2307 struct kvm_coalesced_mmio *ent;
2308
2309 ent = &ring->coalesced_mmio[ring->first];
2310
2311 if (ent->pio == 1) {
2312 address_space_write(&address_space_io, ent->phys_addr,
2313 MEMTXATTRS_UNSPECIFIED, ent->data,
2314 ent->len);
2315 } else {
2316 cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len);
2317 }
2318 smp_wmb();
2319 ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX;
2320 }
2321 }
2322
2323 s->coalesced_flush_in_progress = false;
2324 }
2325
2326 bool kvm_cpu_check_are_resettable(void)
2327 {
2328 return kvm_arch_cpu_check_are_resettable();
2329 }
2330
2331 static void do_kvm_cpu_synchronize_state(CPUState *cpu, run_on_cpu_data arg)
2332 {
2333 if (!cpu->vcpu_dirty) {
2334 kvm_arch_get_registers(cpu);
2335 cpu->vcpu_dirty = true;
2336 }
2337 }
2338
2339 void kvm_cpu_synchronize_state(CPUState *cpu)
2340 {
2341 if (!cpu->vcpu_dirty) {
2342 run_on_cpu(cpu, do_kvm_cpu_synchronize_state, RUN_ON_CPU_NULL);
2343 }
2344 }
2345
2346 static void do_kvm_cpu_synchronize_post_reset(CPUState *cpu, run_on_cpu_data arg)
2347 {
2348 kvm_arch_put_registers(cpu, KVM_PUT_RESET_STATE);
2349 cpu->vcpu_dirty = false;
2350 }
2351
2352 void kvm_cpu_synchronize_post_reset(CPUState *cpu)
2353 {
2354 run_on_cpu(cpu, do_kvm_cpu_synchronize_post_reset, RUN_ON_CPU_NULL);
2355 }
2356
2357 static void do_kvm_cpu_synchronize_post_init(CPUState *cpu, run_on_cpu_data arg)
2358 {
2359 kvm_arch_put_registers(cpu, KVM_PUT_FULL_STATE);
2360 cpu->vcpu_dirty = false;
2361 }
2362
2363 void kvm_cpu_synchronize_post_init(CPUState *cpu)
2364 {
2365 run_on_cpu(cpu, do_kvm_cpu_synchronize_post_init, RUN_ON_CPU_NULL);
2366 }
2367
2368 static void do_kvm_cpu_synchronize_pre_loadvm(CPUState *cpu, run_on_cpu_data arg)
2369 {
2370 cpu->vcpu_dirty = true;
2371 }
2372
2373 void kvm_cpu_synchronize_pre_loadvm(CPUState *cpu)
2374 {
2375 run_on_cpu(cpu, do_kvm_cpu_synchronize_pre_loadvm, RUN_ON_CPU_NULL);
2376 }
2377
2378 #ifdef KVM_HAVE_MCE_INJECTION
2379 static __thread void *pending_sigbus_addr;
2380 static __thread int pending_sigbus_code;
2381 static __thread bool have_sigbus_pending;
2382 #endif
2383
2384 static void kvm_cpu_kick(CPUState *cpu)
2385 {
2386 qatomic_set(&cpu->kvm_run->immediate_exit, 1);
2387 }
2388
2389 static void kvm_cpu_kick_self(void)
2390 {
2391 if (kvm_immediate_exit) {
2392 kvm_cpu_kick(current_cpu);
2393 } else {
2394 qemu_cpu_kick_self();
2395 }
2396 }
2397
2398 static void kvm_eat_signals(CPUState *cpu)
2399 {
2400 struct timespec ts = { 0, 0 };
2401 siginfo_t siginfo;
2402 sigset_t waitset;
2403 sigset_t chkset;
2404 int r;
2405
2406 if (kvm_immediate_exit) {
2407 qatomic_set(&cpu->kvm_run->immediate_exit, 0);
2408 /* Write kvm_run->immediate_exit before the cpu->exit_request
2409 * write in kvm_cpu_exec.
2410 */
2411 smp_wmb();
2412 return;
2413 }
2414
2415 sigemptyset(&waitset);
2416 sigaddset(&waitset, SIG_IPI);
2417
2418 do {
2419 r = sigtimedwait(&waitset, &siginfo, &ts);
2420 if (r == -1 && !(errno == EAGAIN || errno == EINTR)) {
2421 perror("sigtimedwait");
2422 exit(1);
2423 }
2424
2425 r = sigpending(&chkset);
2426 if (r == -1) {
2427 perror("sigpending");
2428 exit(1);
2429 }
2430 } while (sigismember(&chkset, SIG_IPI));
2431 }
2432
2433 int kvm_cpu_exec(CPUState *cpu)
2434 {
2435 struct kvm_run *run = cpu->kvm_run;
2436 int ret, run_ret;
2437
2438 DPRINTF("kvm_cpu_exec()\n");
2439
2440 if (kvm_arch_process_async_events(cpu)) {
2441 qatomic_set(&cpu->exit_request, 0);
2442 return EXCP_HLT;
2443 }
2444
2445 qemu_mutex_unlock_iothread();
2446 cpu_exec_start(cpu);
2447
2448 do {
2449 MemTxAttrs attrs;
2450
2451 if (cpu->vcpu_dirty) {
2452 kvm_arch_put_registers(cpu, KVM_PUT_RUNTIME_STATE);
2453 cpu->vcpu_dirty = false;
2454 }
2455
2456 kvm_arch_pre_run(cpu, run);
2457 if (qatomic_read(&cpu->exit_request)) {
2458 DPRINTF("interrupt exit requested\n");
2459 /*
2460 * KVM requires us to reenter the kernel after IO exits to complete
2461 * instruction emulation. This self-signal will ensure that we
2462 * leave ASAP again.
2463 */
2464 kvm_cpu_kick_self();
2465 }
2466
2467 /* Read cpu->exit_request before KVM_RUN reads run->immediate_exit.
2468 * Matching barrier in kvm_eat_signals.
2469 */
2470 smp_rmb();
2471
2472 run_ret = kvm_vcpu_ioctl(cpu, KVM_RUN, 0);
2473
2474 attrs = kvm_arch_post_run(cpu, run);
2475
2476 #ifdef KVM_HAVE_MCE_INJECTION
2477 if (unlikely(have_sigbus_pending)) {
2478 qemu_mutex_lock_iothread();
2479 kvm_arch_on_sigbus_vcpu(cpu, pending_sigbus_code,
2480 pending_sigbus_addr);
2481 have_sigbus_pending = false;
2482 qemu_mutex_unlock_iothread();
2483 }
2484 #endif
2485
2486 if (run_ret < 0) {
2487 if (run_ret == -EINTR || run_ret == -EAGAIN) {
2488 DPRINTF("io window exit\n");
2489 kvm_eat_signals(cpu);
2490 ret = EXCP_INTERRUPT;
2491 break;
2492 }
2493 fprintf(stderr, "error: kvm run failed %s\n",
2494 strerror(-run_ret));
2495 #ifdef TARGET_PPC
2496 if (run_ret == -EBUSY) {
2497 fprintf(stderr,
2498 "This is probably because your SMT is enabled.\n"
2499 "VCPU can only run on primary threads with all "
2500 "secondary threads offline.\n");
2501 }
2502 #endif
2503 ret = -1;
2504 break;
2505 }
2506
2507 trace_kvm_run_exit(cpu->cpu_index, run->exit_reason);
2508 switch (run->exit_reason) {
2509 case KVM_EXIT_IO:
2510 DPRINTF("handle_io\n");
2511 /* Called outside BQL */
2512 kvm_handle_io(run->io.port, attrs,
2513 (uint8_t *)run + run->io.data_offset,
2514 run->io.direction,
2515 run->io.size,
2516 run->io.count);
2517 ret = 0;
2518 break;
2519 case KVM_EXIT_MMIO:
2520 DPRINTF("handle_mmio\n");
2521 /* Called outside BQL */
2522 address_space_rw(&address_space_memory,
2523 run->mmio.phys_addr, attrs,
2524 run->mmio.data,
2525 run->mmio.len,
2526 run->mmio.is_write);
2527 ret = 0;
2528 break;
2529 case KVM_EXIT_IRQ_WINDOW_OPEN:
2530 DPRINTF("irq_window_open\n");
2531 ret = EXCP_INTERRUPT;
2532 break;
2533 case KVM_EXIT_SHUTDOWN:
2534 DPRINTF("shutdown\n");
2535 qemu_system_reset_request(SHUTDOWN_CAUSE_GUEST_RESET);
2536 ret = EXCP_INTERRUPT;
2537 break;
2538 case KVM_EXIT_UNKNOWN:
2539 fprintf(stderr, "KVM: unknown exit, hardware reason %" PRIx64 "\n",
2540 (uint64_t)run->hw.hardware_exit_reason);
2541 ret = -1;
2542 break;
2543 case KVM_EXIT_INTERNAL_ERROR:
2544 ret = kvm_handle_internal_error(cpu, run);
2545 break;
2546 case KVM_EXIT_SYSTEM_EVENT:
2547 switch (run->system_event.type) {
2548 case KVM_SYSTEM_EVENT_SHUTDOWN:
2549 qemu_system_shutdown_request(SHUTDOWN_CAUSE_GUEST_SHUTDOWN);
2550 ret = EXCP_INTERRUPT;
2551 break;
2552 case KVM_SYSTEM_EVENT_RESET:
2553 qemu_system_reset_request(SHUTDOWN_CAUSE_GUEST_RESET);
2554 ret = EXCP_INTERRUPT;
2555 break;
2556 case KVM_SYSTEM_EVENT_CRASH:
2557 kvm_cpu_synchronize_state(cpu);
2558 qemu_mutex_lock_iothread();
2559 qemu_system_guest_panicked(cpu_get_crash_info(cpu));
2560 qemu_mutex_unlock_iothread();
2561 ret = 0;
2562 break;
2563 default:
2564 DPRINTF("kvm_arch_handle_exit\n");
2565 ret = kvm_arch_handle_exit(cpu, run);
2566 break;
2567 }
2568 break;
2569 default:
2570 DPRINTF("kvm_arch_handle_exit\n");
2571 ret = kvm_arch_handle_exit(cpu, run);
2572 break;
2573 }
2574 } while (ret == 0);
2575
2576 cpu_exec_end(cpu);
2577 qemu_mutex_lock_iothread();
2578
2579 if (ret < 0) {
2580 cpu_dump_state(cpu, stderr, CPU_DUMP_CODE);
2581 vm_stop(RUN_STATE_INTERNAL_ERROR);
2582 }
2583
2584 qatomic_set(&cpu->exit_request, 0);
2585 return ret;
2586 }
2587
2588 int kvm_ioctl(KVMState *s, int type, ...)
2589 {
2590 int ret;
2591 void *arg;
2592 va_list ap;
2593
2594 va_start(ap, type);
2595 arg = va_arg(ap, void *);
2596 va_end(ap);
2597
2598 trace_kvm_ioctl(type, arg);
2599 ret = ioctl(s->fd, type, arg);
2600 if (ret == -1) {
2601 ret = -errno;
2602 }
2603 return ret;
2604 }
2605
2606 int kvm_vm_ioctl(KVMState *s, int type, ...)
2607 {
2608 int ret;
2609 void *arg;
2610 va_list ap;
2611
2612 va_start(ap, type);
2613 arg = va_arg(ap, void *);
2614 va_end(ap);
2615
2616 trace_kvm_vm_ioctl(type, arg);
2617 ret = ioctl(s->vmfd, type, arg);
2618 if (ret == -1) {
2619 ret = -errno;
2620 }
2621 return ret;
2622 }
2623
2624 int kvm_vcpu_ioctl(CPUState *cpu, int type, ...)
2625 {
2626 int ret;
2627 void *arg;
2628 va_list ap;
2629
2630 va_start(ap, type);
2631 arg = va_arg(ap, void *);
2632 va_end(ap);
2633
2634 trace_kvm_vcpu_ioctl(cpu->cpu_index, type, arg);
2635 ret = ioctl(cpu->kvm_fd, type, arg);
2636 if (ret == -1) {
2637 ret = -errno;
2638 }
2639 return ret;
2640 }
2641
2642 int kvm_device_ioctl(int fd, int type, ...)
2643 {
2644 int ret;
2645 void *arg;
2646 va_list ap;
2647
2648 va_start(ap, type);
2649 arg = va_arg(ap, void *);
2650 va_end(ap);
2651
2652 trace_kvm_device_ioctl(fd, type, arg);
2653 ret = ioctl(fd, type, arg);
2654 if (ret == -1) {
2655 ret = -errno;
2656 }
2657 return ret;
2658 }
2659
2660 int kvm_vm_check_attr(KVMState *s, uint32_t group, uint64_t attr)
2661 {
2662 int ret;
2663 struct kvm_device_attr attribute = {
2664 .group = group,
2665 .attr = attr,
2666 };
2667
2668 if (!kvm_vm_attributes_allowed) {
2669 return 0;
2670 }
2671
2672 ret = kvm_vm_ioctl(s, KVM_HAS_DEVICE_ATTR, &attribute);
2673 /* kvm returns 0 on success for HAS_DEVICE_ATTR */
2674 return ret ? 0 : 1;
2675 }
2676
2677 int kvm_device_check_attr(int dev_fd, uint32_t group, uint64_t attr)
2678 {
2679 struct kvm_device_attr attribute = {
2680 .group = group,
2681 .attr = attr,
2682 .flags = 0,
2683 };
2684
2685 return kvm_device_ioctl(dev_fd, KVM_HAS_DEVICE_ATTR, &attribute) ? 0 : 1;
2686 }
2687
2688 int kvm_device_access(int fd, int group, uint64_t attr,
2689 void *val, bool write, Error **errp)
2690 {
2691 struct kvm_device_attr kvmattr;
2692 int err;
2693
2694 kvmattr.flags = 0;
2695 kvmattr.group = group;
2696 kvmattr.attr = attr;
2697 kvmattr.addr = (uintptr_t)val;
2698
2699 err = kvm_device_ioctl(fd,
2700 write ? KVM_SET_DEVICE_ATTR : KVM_GET_DEVICE_ATTR,
2701 &kvmattr);
2702 if (err < 0) {
2703 error_setg_errno(errp, -err,
2704 "KVM_%s_DEVICE_ATTR failed: Group %d "
2705 "attr 0x%016" PRIx64,
2706 write ? "SET" : "GET", group, attr);
2707 }
2708 return err;
2709 }
2710
2711 bool kvm_has_sync_mmu(void)
2712 {
2713 return kvm_state->sync_mmu;
2714 }
2715
2716 int kvm_has_vcpu_events(void)
2717 {
2718 return kvm_state->vcpu_events;
2719 }
2720
2721 int kvm_has_robust_singlestep(void)
2722 {
2723 return kvm_state->robust_singlestep;
2724 }
2725
2726 int kvm_has_debugregs(void)
2727 {
2728 return kvm_state->debugregs;
2729 }
2730
2731 int kvm_max_nested_state_length(void)
2732 {
2733 return kvm_state->max_nested_state_len;
2734 }
2735
2736 int kvm_has_many_ioeventfds(void)
2737 {
2738 if (!kvm_enabled()) {
2739 return 0;
2740 }
2741 return kvm_state->many_ioeventfds;
2742 }
2743
2744 int kvm_has_gsi_routing(void)
2745 {
2746 #ifdef KVM_CAP_IRQ_ROUTING
2747 return kvm_check_extension(kvm_state, KVM_CAP_IRQ_ROUTING);
2748 #else
2749 return false;
2750 #endif
2751 }
2752
2753 int kvm_has_intx_set_mask(void)
2754 {
2755 return kvm_state->intx_set_mask;
2756 }
2757
2758 bool kvm_arm_supports_user_irq(void)
2759 {
2760 return kvm_check_extension(kvm_state, KVM_CAP_ARM_USER_IRQ);
2761 }
2762
2763 #ifdef KVM_CAP_SET_GUEST_DEBUG
2764 struct kvm_sw_breakpoint *kvm_find_sw_breakpoint(CPUState *cpu,
2765 target_ulong pc)
2766 {
2767 struct kvm_sw_breakpoint *bp;
2768
2769 QTAILQ_FOREACH(bp, &cpu->kvm_state->kvm_sw_breakpoints, entry) {
2770 if (bp->pc == pc) {
2771 return bp;
2772 }
2773 }
2774 return NULL;
2775 }
2776
2777 int kvm_sw_breakpoints_active(CPUState *cpu)
2778 {
2779 return !QTAILQ_EMPTY(&cpu->kvm_state->kvm_sw_breakpoints);
2780 }
2781
2782 struct kvm_set_guest_debug_data {
2783 struct kvm_guest_debug dbg;
2784 int err;
2785 };
2786
2787 static void kvm_invoke_set_guest_debug(CPUState *cpu, run_on_cpu_data data)
2788 {
2789 struct kvm_set_guest_debug_data *dbg_data =
2790 (struct kvm_set_guest_debug_data *) data.host_ptr;
2791
2792 dbg_data->err = kvm_vcpu_ioctl(cpu, KVM_SET_GUEST_DEBUG,
2793 &dbg_data->dbg);
2794 }
2795
2796 int kvm_update_guest_debug(CPUState *cpu, unsigned long reinject_trap)
2797 {
2798 struct kvm_set_guest_debug_data data;
2799
2800 data.dbg.control = reinject_trap;
2801
2802 if (cpu->singlestep_enabled) {
2803 data.dbg.control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_SINGLESTEP;
2804 }
2805 kvm_arch_update_guest_debug(cpu, &data.dbg);
2806
2807 run_on_cpu(cpu, kvm_invoke_set_guest_debug,
2808 RUN_ON_CPU_HOST_PTR(&data));
2809 return data.err;
2810 }
2811
2812 int kvm_insert_breakpoint(CPUState *cpu, target_ulong addr,
2813 target_ulong len, int type)
2814 {
2815 struct kvm_sw_breakpoint *bp;
2816 int err;
2817
2818 if (type == GDB_BREAKPOINT_SW) {
2819 bp = kvm_find_sw_breakpoint(cpu, addr);
2820 if (bp) {
2821 bp->use_count++;
2822 return 0;
2823 }
2824
2825 bp = g_malloc(sizeof(struct kvm_sw_breakpoint));
2826 bp->pc = addr;
2827 bp->use_count = 1;
2828 err = kvm_arch_insert_sw_breakpoint(cpu, bp);
2829 if (err) {
2830 g_free(bp);
2831 return err;
2832 }
2833
2834 QTAILQ_INSERT_HEAD(&cpu->kvm_state->kvm_sw_breakpoints, bp, entry);
2835 } else {
2836 err = kvm_arch_insert_hw_breakpoint(addr, len, type);
2837 if (err) {
2838 return err;
2839 }
2840 }
2841
2842 CPU_FOREACH(cpu) {
2843 err = kvm_update_guest_debug(cpu, 0);
2844 if (err) {
2845 return err;
2846 }
2847 }
2848 return 0;
2849 }
2850
2851 int kvm_remove_breakpoint(CPUState *cpu, target_ulong addr,
2852 target_ulong len, int type)
2853 {
2854 struct kvm_sw_breakpoint *bp;
2855 int err;
2856
2857 if (type == GDB_BREAKPOINT_SW) {
2858 bp = kvm_find_sw_breakpoint(cpu, addr);
2859 if (!bp) {
2860 return -ENOENT;
2861 }
2862
2863 if (bp->use_count > 1) {
2864 bp->use_count--;
2865 return 0;
2866 }
2867
2868 err = kvm_arch_remove_sw_breakpoint(cpu, bp);
2869 if (err) {
2870 return err;
2871 }
2872
2873 QTAILQ_REMOVE(&cpu->kvm_state->kvm_sw_breakpoints, bp, entry);
2874 g_free(bp);
2875 } else {
2876 err = kvm_arch_remove_hw_breakpoint(addr, len, type);
2877 if (err) {
2878 return err;
2879 }
2880 }
2881
2882 CPU_FOREACH(cpu) {
2883 err = kvm_update_guest_debug(cpu, 0);
2884 if (err) {
2885 return err;
2886 }
2887 }
2888 return 0;
2889 }
2890
2891 void kvm_remove_all_breakpoints(CPUState *cpu)
2892 {
2893 struct kvm_sw_breakpoint *bp, *next;
2894 KVMState *s = cpu->kvm_state;
2895 CPUState *tmpcpu;
2896
2897 QTAILQ_FOREACH_SAFE(bp, &s->kvm_sw_breakpoints, entry, next) {
2898 if (kvm_arch_remove_sw_breakpoint(cpu, bp) != 0) {
2899 /* Try harder to find a CPU that currently sees the breakpoint. */
2900 CPU_FOREACH(tmpcpu) {
2901 if (kvm_arch_remove_sw_breakpoint(tmpcpu, bp) == 0) {
2902 break;
2903 }
2904 }
2905 }
2906 QTAILQ_REMOVE(&s->kvm_sw_breakpoints, bp, entry);
2907 g_free(bp);
2908 }
2909 kvm_arch_remove_all_hw_breakpoints();
2910
2911 CPU_FOREACH(cpu) {
2912 kvm_update_guest_debug(cpu, 0);
2913 }
2914 }
2915
2916 #else /* !KVM_CAP_SET_GUEST_DEBUG */
2917
2918 int kvm_update_guest_debug(CPUState *cpu, unsigned long reinject_trap)
2919 {
2920 return -EINVAL;
2921 }
2922
2923 int kvm_insert_breakpoint(CPUState *cpu, target_ulong addr,
2924 target_ulong len, int type)
2925 {
2926 return -EINVAL;
2927 }
2928
2929 int kvm_remove_breakpoint(CPUState *cpu, target_ulong addr,
2930 target_ulong len, int type)
2931 {
2932 return -EINVAL;
2933 }
2934
2935 void kvm_remove_all_breakpoints(CPUState *cpu)
2936 {
2937 }
2938 #endif /* !KVM_CAP_SET_GUEST_DEBUG */
2939
2940 static int kvm_set_signal_mask(CPUState *cpu, const sigset_t *sigset)
2941 {
2942 KVMState *s = kvm_state;
2943 struct kvm_signal_mask *sigmask;
2944 int r;
2945
2946 sigmask = g_malloc(sizeof(*sigmask) + sizeof(*sigset));
2947
2948 sigmask->len = s->sigmask_len;
2949 memcpy(sigmask->sigset, sigset, sizeof(*sigset));
2950 r = kvm_vcpu_ioctl(cpu, KVM_SET_SIGNAL_MASK, sigmask);
2951 g_free(sigmask);
2952
2953 return r;
2954 }
2955
2956 static void kvm_ipi_signal(int sig)
2957 {
2958 if (current_cpu) {
2959 assert(kvm_immediate_exit);
2960 kvm_cpu_kick(current_cpu);
2961 }
2962 }
2963
2964 void kvm_init_cpu_signals(CPUState *cpu)
2965 {
2966 int r;
2967 sigset_t set;
2968 struct sigaction sigact;
2969
2970 memset(&sigact, 0, sizeof(sigact));
2971 sigact.sa_handler = kvm_ipi_signal;
2972 sigaction(SIG_IPI, &sigact, NULL);
2973
2974 pthread_sigmask(SIG_BLOCK, NULL, &set);
2975 #if defined KVM_HAVE_MCE_INJECTION
2976 sigdelset(&set, SIGBUS);
2977 pthread_sigmask(SIG_SETMASK, &set, NULL);
2978 #endif
2979 sigdelset(&set, SIG_IPI);
2980 if (kvm_immediate_exit) {
2981 r = pthread_sigmask(SIG_SETMASK, &set, NULL);
2982 } else {
2983 r = kvm_set_signal_mask(cpu, &set);
2984 }
2985 if (r) {
2986 fprintf(stderr, "kvm_set_signal_mask: %s\n", strerror(-r));
2987 exit(1);
2988 }
2989 }
2990
2991 /* Called asynchronously in VCPU thread. */
2992 int kvm_on_sigbus_vcpu(CPUState *cpu, int code, void *addr)
2993 {
2994 #ifdef KVM_HAVE_MCE_INJECTION
2995 if (have_sigbus_pending) {
2996 return 1;
2997 }
2998 have_sigbus_pending = true;
2999 pending_sigbus_addr = addr;
3000 pending_sigbus_code = code;
3001 qatomic_set(&cpu->exit_request, 1);
3002 return 0;
3003 #else
3004 return 1;
3005 #endif
3006 }
3007
3008 /* Called synchronously (via signalfd) in main thread. */
3009 int kvm_on_sigbus(int code, void *addr)
3010 {
3011 #ifdef KVM_HAVE_MCE_INJECTION
3012 /* Action required MCE kills the process if SIGBUS is blocked. Because
3013 * that's what happens in the I/O thread, where we handle MCE via signalfd,
3014 * we can only get action optional here.
3015 */
3016 assert(code != BUS_MCEERR_AR);
3017 kvm_arch_on_sigbus_vcpu(first_cpu, code, addr);
3018 return 0;
3019 #else
3020 return 1;
3021 #endif
3022 }
3023
3024 int kvm_create_device(KVMState *s, uint64_t type, bool test)
3025 {
3026 int ret;
3027 struct kvm_create_device create_dev;
3028
3029 create_dev.type = type;
3030 create_dev.fd = -1;
3031 create_dev.flags = test ? KVM_CREATE_DEVICE_TEST : 0;
3032
3033 if (!kvm_check_extension(s, KVM_CAP_DEVICE_CTRL)) {
3034 return -ENOTSUP;
3035 }
3036
3037 ret = kvm_vm_ioctl(s, KVM_CREATE_DEVICE, &create_dev);
3038 if (ret) {
3039 return ret;
3040 }
3041
3042 return test ? 0 : create_dev.fd;
3043 }
3044
3045 bool kvm_device_supported(int vmfd, uint64_t type)
3046 {
3047 struct kvm_create_device create_dev = {
3048 .type = type,
3049 .fd = -1,
3050 .flags = KVM_CREATE_DEVICE_TEST,
3051 };
3052
3053 if (ioctl(vmfd, KVM_CHECK_EXTENSION, KVM_CAP_DEVICE_CTRL) <= 0) {
3054 return false;
3055 }
3056
3057 return (ioctl(vmfd, KVM_CREATE_DEVICE, &create_dev) >= 0);
3058 }
3059
3060 int kvm_set_one_reg(CPUState *cs, uint64_t id, void *source)
3061 {
3062 struct kvm_one_reg reg;
3063 int r;
3064
3065 reg.id = id;
3066 reg.addr = (uintptr_t) source;
3067 r = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, &reg);
3068 if (r) {
3069 trace_kvm_failed_reg_set(id, strerror(-r));
3070 }
3071 return r;
3072 }
3073
3074 int kvm_get_one_reg(CPUState *cs, uint64_t id, void *target)
3075 {
3076 struct kvm_one_reg reg;
3077 int r;
3078
3079 reg.id = id;
3080 reg.addr = (uintptr_t) target;
3081 r = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &reg);
3082 if (r) {
3083 trace_kvm_failed_reg_get(id, strerror(-r));
3084 }
3085 return r;
3086 }
3087
3088 static bool kvm_accel_has_memory(MachineState *ms, AddressSpace *as,
3089 hwaddr start_addr, hwaddr size)
3090 {
3091 KVMState *kvm = KVM_STATE(ms->accelerator);
3092 int i;
3093
3094 for (i = 0; i < kvm->nr_as; ++i) {
3095 if (kvm->as[i].as == as && kvm->as[i].ml) {
3096 size = MIN(kvm_max_slot_size, size);
3097 return NULL != kvm_lookup_matching_slot(kvm->as[i].ml,
3098 start_addr, size);
3099 }
3100 }
3101
3102 return false;
3103 }
3104
3105 static void kvm_get_kvm_shadow_mem(Object *obj, Visitor *v,
3106 const char *name, void *opaque,
3107 Error **errp)
3108 {
3109 KVMState *s = KVM_STATE(obj);
3110 int64_t value = s->kvm_shadow_mem;
3111
3112 visit_type_int(v, name, &value, errp);
3113 }
3114
3115 static void kvm_set_kvm_shadow_mem(Object *obj, Visitor *v,
3116 const char *name, void *opaque,
3117 Error **errp)
3118 {
3119 KVMState *s = KVM_STATE(obj);
3120 int64_t value;
3121
3122 if (!visit_type_int(v, name, &value, errp)) {
3123 return;
3124 }
3125
3126 s->kvm_shadow_mem = value;
3127 }
3128
3129 static void kvm_set_kernel_irqchip(Object *obj, Visitor *v,
3130 const char *name, void *opaque,
3131 Error **errp)
3132 {
3133 KVMState *s = KVM_STATE(obj);
3134 OnOffSplit mode;
3135
3136 if (!visit_type_OnOffSplit(v, name, &mode, errp)) {
3137 return;
3138 }
3139 switch (mode) {
3140 case ON_OFF_SPLIT_ON:
3141 s->kernel_irqchip_allowed = true;
3142 s->kernel_irqchip_required = true;
3143 s->kernel_irqchip_split = ON_OFF_AUTO_OFF;
3144 break;
3145 case ON_OFF_SPLIT_OFF:
3146 s->kernel_irqchip_allowed = false;
3147 s->kernel_irqchip_required = false;
3148 s->kernel_irqchip_split = ON_OFF_AUTO_OFF;
3149 break;
3150 case ON_OFF_SPLIT_SPLIT:
3151 s->kernel_irqchip_allowed = true;
3152 s->kernel_irqchip_required = true;
3153 s->kernel_irqchip_split = ON_OFF_AUTO_ON;
3154 break;
3155 default:
3156 /* The value was checked in visit_type_OnOffSplit() above. If
3157 * we get here, then something is wrong in QEMU.
3158 */
3159 abort();
3160 }
3161 }
3162
3163 bool kvm_kernel_irqchip_allowed(void)
3164 {
3165 return kvm_state->kernel_irqchip_allowed;
3166 }
3167
3168 bool kvm_kernel_irqchip_required(void)
3169 {
3170 return kvm_state->kernel_irqchip_required;
3171 }
3172
3173 bool kvm_kernel_irqchip_split(void)
3174 {
3175 return kvm_state->kernel_irqchip_split == ON_OFF_AUTO_ON;
3176 }
3177
3178 static void kvm_accel_instance_init(Object *obj)
3179 {
3180 KVMState *s = KVM_STATE(obj);
3181
3182 s->kvm_shadow_mem = -1;
3183 s->kernel_irqchip_allowed = true;
3184 s->kernel_irqchip_split = ON_OFF_AUTO_AUTO;
3185 }
3186
3187 static void kvm_accel_class_init(ObjectClass *oc, void *data)
3188 {
3189 AccelClass *ac = ACCEL_CLASS(oc);
3190 ac->name = "KVM";
3191 ac->init_machine = kvm_init;
3192 ac->has_memory = kvm_accel_has_memory;
3193 ac->allowed = &kvm_allowed;
3194
3195 object_class_property_add(oc, "kernel-irqchip", "on|off|split",
3196 NULL, kvm_set_kernel_irqchip,
3197 NULL, NULL);
3198 object_class_property_set_description(oc, "kernel-irqchip",
3199 "Configure KVM in-kernel irqchip");
3200
3201 object_class_property_add(oc, "kvm-shadow-mem", "int",
3202 kvm_get_kvm_shadow_mem, kvm_set_kvm_shadow_mem,
3203 NULL, NULL);
3204 object_class_property_set_description(oc, "kvm-shadow-mem",
3205 "KVM shadow MMU size");
3206 }
3207
3208 static const TypeInfo kvm_accel_type = {
3209 .name = TYPE_KVM_ACCEL,
3210 .parent = TYPE_ACCEL,
3211 .instance_init = kvm_accel_instance_init,
3212 .class_init = kvm_accel_class_init,
3213 .instance_size = sizeof(KVMState),
3214 };
3215
3216 static void kvm_type_init(void)
3217 {
3218 type_register_static(&kvm_accel_type);
3219 }
3220
3221 type_init(kvm_type_init);
AR7 emulation for QEMU