xref: /openbmc/qemu/accel/kvm/kvm-all.c (revision dc5bd18f)
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-common.h"
22 #include "qemu/atomic.h"
23 #include "qemu/option.h"
24 #include "qemu/config-file.h"
25 #include "qemu/error-report.h"
26 #include "qapi/error.h"
27 #include "hw/hw.h"
28 #include "hw/pci/msi.h"
29 #include "hw/pci/msix.h"
30 #include "hw/s390x/adapter.h"
31 #include "exec/gdbstub.h"
32 #include "sysemu/kvm_int.h"
33 #include "sysemu/cpus.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 "trace.h"
40 #include "hw/irq.h"
41 
42 #include "hw/boards.h"
43 
44 /* This check must be after config-host.h is included */
45 #ifdef CONFIG_EVENTFD
46 #include <sys/eventfd.h>
47 #endif
48 
49 /* KVM uses PAGE_SIZE in its definition of KVM_COALESCED_MMIO_MAX. We
50  * need to use the real host PAGE_SIZE, as that's what KVM will use.
51  */
52 #define PAGE_SIZE getpagesize()
53 
54 //#define DEBUG_KVM
55 
56 #ifdef DEBUG_KVM
57 #define DPRINTF(fmt, ...) \
58     do { fprintf(stderr, fmt, ## __VA_ARGS__); } while (0)
59 #else
60 #define DPRINTF(fmt, ...) \
61     do { } while (0)
62 #endif
63 
64 #define KVM_MSI_HASHTAB_SIZE    256
65 
66 struct KVMParkedVcpu {
67     unsigned long vcpu_id;
68     int kvm_fd;
69     QLIST_ENTRY(KVMParkedVcpu) node;
70 };
71 
72 struct KVMState
73 {
74     AccelState parent_obj;
75 
76     int nr_slots;
77     int fd;
78     int vmfd;
79     int coalesced_mmio;
80     struct kvm_coalesced_mmio_ring *coalesced_mmio_ring;
81     bool coalesced_flush_in_progress;
82     int vcpu_events;
83     int robust_singlestep;
84     int debugregs;
85 #ifdef KVM_CAP_SET_GUEST_DEBUG
86     struct kvm_sw_breakpoint_head kvm_sw_breakpoints;
87 #endif
88     int many_ioeventfds;
89     int intx_set_mask;
90     bool sync_mmu;
91     /* The man page (and posix) say ioctl numbers are signed int, but
92      * they're not.  Linux, glibc and *BSD all treat ioctl numbers as
93      * unsigned, and treating them as signed here can break things */
94     unsigned irq_set_ioctl;
95     unsigned int sigmask_len;
96     GHashTable *gsimap;
97 #ifdef KVM_CAP_IRQ_ROUTING
98     struct kvm_irq_routing *irq_routes;
99     int nr_allocated_irq_routes;
100     unsigned long *used_gsi_bitmap;
101     unsigned int gsi_count;
102     QTAILQ_HEAD(msi_hashtab, KVMMSIRoute) msi_hashtab[KVM_MSI_HASHTAB_SIZE];
103 #endif
104     KVMMemoryListener memory_listener;
105     QLIST_HEAD(, KVMParkedVcpu) kvm_parked_vcpus;
106 };
107 
108 KVMState *kvm_state;
109 bool kvm_kernel_irqchip;
110 bool kvm_split_irqchip;
111 bool kvm_async_interrupts_allowed;
112 bool kvm_halt_in_kernel_allowed;
113 bool kvm_eventfds_allowed;
114 bool kvm_irqfds_allowed;
115 bool kvm_resamplefds_allowed;
116 bool kvm_msi_via_irqfd_allowed;
117 bool kvm_gsi_routing_allowed;
118 bool kvm_gsi_direct_mapping;
119 bool kvm_allowed;
120 bool kvm_readonly_mem_allowed;
121 bool kvm_vm_attributes_allowed;
122 bool kvm_direct_msi_allowed;
123 bool kvm_ioeventfd_any_length_allowed;
124 bool kvm_msi_use_devid;
125 static bool kvm_immediate_exit;
126 
127 static const KVMCapabilityInfo kvm_required_capabilites[] = {
128     KVM_CAP_INFO(USER_MEMORY),
129     KVM_CAP_INFO(DESTROY_MEMORY_REGION_WORKS),
130     KVM_CAP_INFO(JOIN_MEMORY_REGIONS_WORKS),
131     KVM_CAP_LAST_INFO
132 };
133 
134 int kvm_get_max_memslots(void)
135 {
136     KVMState *s = KVM_STATE(current_machine->accelerator);
137 
138     return s->nr_slots;
139 }
140 
141 static KVMSlot *kvm_get_free_slot(KVMMemoryListener *kml)
142 {
143     KVMState *s = kvm_state;
144     int i;
145 
146     for (i = 0; i < s->nr_slots; i++) {
147         if (kml->slots[i].memory_size == 0) {
148             return &kml->slots[i];
149         }
150     }
151 
152     return NULL;
153 }
154 
155 bool kvm_has_free_slot(MachineState *ms)
156 {
157     KVMState *s = KVM_STATE(ms->accelerator);
158 
159     return kvm_get_free_slot(&s->memory_listener);
160 }
161 
162 static KVMSlot *kvm_alloc_slot(KVMMemoryListener *kml)
163 {
164     KVMSlot *slot = kvm_get_free_slot(kml);
165 
166     if (slot) {
167         return slot;
168     }
169 
170     fprintf(stderr, "%s: no free slot available\n", __func__);
171     abort();
172 }
173 
174 static KVMSlot *kvm_lookup_matching_slot(KVMMemoryListener *kml,
175                                          hwaddr start_addr,
176                                          hwaddr size)
177 {
178     KVMState *s = kvm_state;
179     int i;
180 
181     for (i = 0; i < s->nr_slots; i++) {
182         KVMSlot *mem = &kml->slots[i];
183 
184         if (start_addr == mem->start_addr && size == mem->memory_size) {
185             return mem;
186         }
187     }
188 
189     return NULL;
190 }
191 
192 /*
193  * Calculate and align the start address and the size of the section.
194  * Return the size. If the size is 0, the aligned section is empty.
195  */
196 static hwaddr kvm_align_section(MemoryRegionSection *section,
197                                 hwaddr *start)
198 {
199     hwaddr size = int128_get64(section->size);
200     hwaddr delta, aligned;
201 
202     /* kvm works in page size chunks, but the function may be called
203        with sub-page size and unaligned start address. Pad the start
204        address to next and truncate size to previous page boundary. */
205     aligned = ROUND_UP(section->offset_within_address_space,
206                        qemu_real_host_page_size);
207     delta = aligned - section->offset_within_address_space;
208     *start = aligned;
209     if (delta > size) {
210         return 0;
211     }
212 
213     return (size - delta) & qemu_real_host_page_mask;
214 }
215 
216 int kvm_physical_memory_addr_from_host(KVMState *s, void *ram,
217                                        hwaddr *phys_addr)
218 {
219     KVMMemoryListener *kml = &s->memory_listener;
220     int i;
221 
222     for (i = 0; i < s->nr_slots; i++) {
223         KVMSlot *mem = &kml->slots[i];
224 
225         if (ram >= mem->ram && ram < mem->ram + mem->memory_size) {
226             *phys_addr = mem->start_addr + (ram - mem->ram);
227             return 1;
228         }
229     }
230 
231     return 0;
232 }
233 
234 static int kvm_set_user_memory_region(KVMMemoryListener *kml, KVMSlot *slot)
235 {
236     KVMState *s = kvm_state;
237     struct kvm_userspace_memory_region mem;
238     int ret;
239 
240     mem.slot = slot->slot | (kml->as_id << 16);
241     mem.guest_phys_addr = slot->start_addr;
242     mem.userspace_addr = (unsigned long)slot->ram;
243     mem.flags = slot->flags;
244 
245     if (slot->memory_size && mem.flags & KVM_MEM_READONLY) {
246         /* Set the slot size to 0 before setting the slot to the desired
247          * value. This is needed based on KVM commit 75d61fbc. */
248         mem.memory_size = 0;
249         kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
250     }
251     mem.memory_size = slot->memory_size;
252     ret = kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
253     trace_kvm_set_user_memory(mem.slot, mem.flags, mem.guest_phys_addr,
254                               mem.memory_size, mem.userspace_addr, ret);
255     return ret;
256 }
257 
258 int kvm_destroy_vcpu(CPUState *cpu)
259 {
260     KVMState *s = kvm_state;
261     long mmap_size;
262     struct KVMParkedVcpu *vcpu = NULL;
263     int ret = 0;
264 
265     DPRINTF("kvm_destroy_vcpu\n");
266 
267     mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
268     if (mmap_size < 0) {
269         ret = mmap_size;
270         DPRINTF("KVM_GET_VCPU_MMAP_SIZE failed\n");
271         goto err;
272     }
273 
274     ret = munmap(cpu->kvm_run, mmap_size);
275     if (ret < 0) {
276         goto err;
277     }
278 
279     vcpu = g_malloc0(sizeof(*vcpu));
280     vcpu->vcpu_id = kvm_arch_vcpu_id(cpu);
281     vcpu->kvm_fd = cpu->kvm_fd;
282     QLIST_INSERT_HEAD(&kvm_state->kvm_parked_vcpus, vcpu, node);
283 err:
284     return ret;
285 }
286 
287 static int kvm_get_vcpu(KVMState *s, unsigned long vcpu_id)
288 {
289     struct KVMParkedVcpu *cpu;
290 
291     QLIST_FOREACH(cpu, &s->kvm_parked_vcpus, node) {
292         if (cpu->vcpu_id == vcpu_id) {
293             int kvm_fd;
294 
295             QLIST_REMOVE(cpu, node);
296             kvm_fd = cpu->kvm_fd;
297             g_free(cpu);
298             return kvm_fd;
299         }
300     }
301 
302     return kvm_vm_ioctl(s, KVM_CREATE_VCPU, (void *)vcpu_id);
303 }
304 
305 int kvm_init_vcpu(CPUState *cpu)
306 {
307     KVMState *s = kvm_state;
308     long mmap_size;
309     int ret;
310 
311     DPRINTF("kvm_init_vcpu\n");
312 
313     ret = kvm_get_vcpu(s, kvm_arch_vcpu_id(cpu));
314     if (ret < 0) {
315         DPRINTF("kvm_create_vcpu failed\n");
316         goto err;
317     }
318 
319     cpu->kvm_fd = ret;
320     cpu->kvm_state = s;
321     cpu->vcpu_dirty = true;
322 
323     mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
324     if (mmap_size < 0) {
325         ret = mmap_size;
326         DPRINTF("KVM_GET_VCPU_MMAP_SIZE failed\n");
327         goto err;
328     }
329 
330     cpu->kvm_run = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
331                         cpu->kvm_fd, 0);
332     if (cpu->kvm_run == MAP_FAILED) {
333         ret = -errno;
334         DPRINTF("mmap'ing vcpu state failed\n");
335         goto err;
336     }
337 
338     if (s->coalesced_mmio && !s->coalesced_mmio_ring) {
339         s->coalesced_mmio_ring =
340             (void *)cpu->kvm_run + s->coalesced_mmio * PAGE_SIZE;
341     }
342 
343     ret = kvm_arch_init_vcpu(cpu);
344 err:
345     return ret;
346 }
347 
348 /*
349  * dirty pages logging control
350  */
351 
352 static int kvm_mem_flags(MemoryRegion *mr)
353 {
354     bool readonly = mr->readonly || memory_region_is_romd(mr);
355     int flags = 0;
356 
357     if (memory_region_get_dirty_log_mask(mr) != 0) {
358         flags |= KVM_MEM_LOG_DIRTY_PAGES;
359     }
360     if (readonly && kvm_readonly_mem_allowed) {
361         flags |= KVM_MEM_READONLY;
362     }
363     return flags;
364 }
365 
366 static int kvm_slot_update_flags(KVMMemoryListener *kml, KVMSlot *mem,
367                                  MemoryRegion *mr)
368 {
369     int old_flags;
370 
371     old_flags = mem->flags;
372     mem->flags = kvm_mem_flags(mr);
373 
374     /* If nothing changed effectively, no need to issue ioctl */
375     if (mem->flags == old_flags) {
376         return 0;
377     }
378 
379     return kvm_set_user_memory_region(kml, mem);
380 }
381 
382 static int kvm_section_update_flags(KVMMemoryListener *kml,
383                                     MemoryRegionSection *section)
384 {
385     hwaddr start_addr, size;
386     KVMSlot *mem;
387 
388     size = kvm_align_section(section, &start_addr);
389     if (!size) {
390         return 0;
391     }
392 
393     mem = kvm_lookup_matching_slot(kml, start_addr, size);
394     if (!mem) {
395         /* We don't have a slot if we want to trap every access. */
396         return 0;
397     }
398 
399     return kvm_slot_update_flags(kml, mem, section->mr);
400 }
401 
402 static void kvm_log_start(MemoryListener *listener,
403                           MemoryRegionSection *section,
404                           int old, int new)
405 {
406     KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
407     int r;
408 
409     if (old != 0) {
410         return;
411     }
412 
413     r = kvm_section_update_flags(kml, section);
414     if (r < 0) {
415         abort();
416     }
417 }
418 
419 static void kvm_log_stop(MemoryListener *listener,
420                           MemoryRegionSection *section,
421                           int old, int new)
422 {
423     KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
424     int r;
425 
426     if (new != 0) {
427         return;
428     }
429 
430     r = kvm_section_update_flags(kml, section);
431     if (r < 0) {
432         abort();
433     }
434 }
435 
436 /* get kvm's dirty pages bitmap and update qemu's */
437 static int kvm_get_dirty_pages_log_range(MemoryRegionSection *section,
438                                          unsigned long *bitmap)
439 {
440     ram_addr_t start = section->offset_within_region +
441                        memory_region_get_ram_addr(section->mr);
442     ram_addr_t pages = int128_get64(section->size) / getpagesize();
443 
444     cpu_physical_memory_set_dirty_lebitmap(bitmap, start, pages);
445     return 0;
446 }
447 
448 #define ALIGN(x, y)  (((x)+(y)-1) & ~((y)-1))
449 
450 /**
451  * kvm_physical_sync_dirty_bitmap - Grab dirty bitmap from kernel space
452  * This function updates qemu's dirty bitmap using
453  * memory_region_set_dirty().  This means all bits are set
454  * to dirty.
455  *
456  * @start_add: start of logged region.
457  * @end_addr: end of logged region.
458  */
459 static int kvm_physical_sync_dirty_bitmap(KVMMemoryListener *kml,
460                                           MemoryRegionSection *section)
461 {
462     KVMState *s = kvm_state;
463     struct kvm_dirty_log d = {};
464     KVMSlot *mem;
465     hwaddr start_addr, size;
466 
467     size = kvm_align_section(section, &start_addr);
468     if (size) {
469         mem = kvm_lookup_matching_slot(kml, start_addr, size);
470         if (!mem) {
471             /* We don't have a slot if we want to trap every access. */
472             return 0;
473         }
474 
475         /* XXX bad kernel interface alert
476          * For dirty bitmap, kernel allocates array of size aligned to
477          * bits-per-long.  But for case when the kernel is 64bits and
478          * the userspace is 32bits, userspace can't align to the same
479          * bits-per-long, since sizeof(long) is different between kernel
480          * and user space.  This way, userspace will provide buffer which
481          * may be 4 bytes less than the kernel will use, resulting in
482          * userspace memory corruption (which is not detectable by valgrind
483          * too, in most cases).
484          * So for now, let's align to 64 instead of HOST_LONG_BITS here, in
485          * a hope that sizeof(long) won't become >8 any time soon.
486          */
487         size = ALIGN(((mem->memory_size) >> TARGET_PAGE_BITS),
488                      /*HOST_LONG_BITS*/ 64) / 8;
489         d.dirty_bitmap = g_malloc0(size);
490 
491         d.slot = mem->slot | (kml->as_id << 16);
492         if (kvm_vm_ioctl(s, KVM_GET_DIRTY_LOG, &d) == -1) {
493             DPRINTF("ioctl failed %d\n", errno);
494             g_free(d.dirty_bitmap);
495             return -1;
496         }
497 
498         kvm_get_dirty_pages_log_range(section, d.dirty_bitmap);
499         g_free(d.dirty_bitmap);
500     }
501 
502     return 0;
503 }
504 
505 static void kvm_coalesce_mmio_region(MemoryListener *listener,
506                                      MemoryRegionSection *secion,
507                                      hwaddr start, hwaddr size)
508 {
509     KVMState *s = kvm_state;
510 
511     if (s->coalesced_mmio) {
512         struct kvm_coalesced_mmio_zone zone;
513 
514         zone.addr = start;
515         zone.size = size;
516         zone.pad = 0;
517 
518         (void)kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone);
519     }
520 }
521 
522 static void kvm_uncoalesce_mmio_region(MemoryListener *listener,
523                                        MemoryRegionSection *secion,
524                                        hwaddr start, hwaddr size)
525 {
526     KVMState *s = kvm_state;
527 
528     if (s->coalesced_mmio) {
529         struct kvm_coalesced_mmio_zone zone;
530 
531         zone.addr = start;
532         zone.size = size;
533         zone.pad = 0;
534 
535         (void)kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone);
536     }
537 }
538 
539 int kvm_check_extension(KVMState *s, unsigned int extension)
540 {
541     int ret;
542 
543     ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, extension);
544     if (ret < 0) {
545         ret = 0;
546     }
547 
548     return ret;
549 }
550 
551 int kvm_vm_check_extension(KVMState *s, unsigned int extension)
552 {
553     int ret;
554 
555     ret = kvm_vm_ioctl(s, KVM_CHECK_EXTENSION, extension);
556     if (ret < 0) {
557         /* VM wide version not implemented, use global one instead */
558         ret = kvm_check_extension(s, extension);
559     }
560 
561     return ret;
562 }
563 
564 static uint32_t adjust_ioeventfd_endianness(uint32_t val, uint32_t size)
565 {
566 #if defined(HOST_WORDS_BIGENDIAN) != defined(TARGET_WORDS_BIGENDIAN)
567     /* The kernel expects ioeventfd values in HOST_WORDS_BIGENDIAN
568      * endianness, but the memory core hands them in target endianness.
569      * For example, PPC is always treated as big-endian even if running
570      * on KVM and on PPC64LE.  Correct here.
571      */
572     switch (size) {
573     case 2:
574         val = bswap16(val);
575         break;
576     case 4:
577         val = bswap32(val);
578         break;
579     }
580 #endif
581     return val;
582 }
583 
584 static int kvm_set_ioeventfd_mmio(int fd, hwaddr addr, uint32_t val,
585                                   bool assign, uint32_t size, bool datamatch)
586 {
587     int ret;
588     struct kvm_ioeventfd iofd = {
589         .datamatch = datamatch ? adjust_ioeventfd_endianness(val, size) : 0,
590         .addr = addr,
591         .len = size,
592         .flags = 0,
593         .fd = fd,
594     };
595 
596     if (!kvm_enabled()) {
597         return -ENOSYS;
598     }
599 
600     if (datamatch) {
601         iofd.flags |= KVM_IOEVENTFD_FLAG_DATAMATCH;
602     }
603     if (!assign) {
604         iofd.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
605     }
606 
607     ret = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &iofd);
608 
609     if (ret < 0) {
610         return -errno;
611     }
612 
613     return 0;
614 }
615 
616 static int kvm_set_ioeventfd_pio(int fd, uint16_t addr, uint16_t val,
617                                  bool assign, uint32_t size, bool datamatch)
618 {
619     struct kvm_ioeventfd kick = {
620         .datamatch = datamatch ? adjust_ioeventfd_endianness(val, size) : 0,
621         .addr = addr,
622         .flags = KVM_IOEVENTFD_FLAG_PIO,
623         .len = size,
624         .fd = fd,
625     };
626     int r;
627     if (!kvm_enabled()) {
628         return -ENOSYS;
629     }
630     if (datamatch) {
631         kick.flags |= KVM_IOEVENTFD_FLAG_DATAMATCH;
632     }
633     if (!assign) {
634         kick.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
635     }
636     r = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &kick);
637     if (r < 0) {
638         return r;
639     }
640     return 0;
641 }
642 
643 
644 static int kvm_check_many_ioeventfds(void)
645 {
646     /* Userspace can use ioeventfd for io notification.  This requires a host
647      * that supports eventfd(2) and an I/O thread; since eventfd does not
648      * support SIGIO it cannot interrupt the vcpu.
649      *
650      * Older kernels have a 6 device limit on the KVM io bus.  Find out so we
651      * can avoid creating too many ioeventfds.
652      */
653 #if defined(CONFIG_EVENTFD)
654     int ioeventfds[7];
655     int i, ret = 0;
656     for (i = 0; i < ARRAY_SIZE(ioeventfds); i++) {
657         ioeventfds[i] = eventfd(0, EFD_CLOEXEC);
658         if (ioeventfds[i] < 0) {
659             break;
660         }
661         ret = kvm_set_ioeventfd_pio(ioeventfds[i], 0, i, true, 2, true);
662         if (ret < 0) {
663             close(ioeventfds[i]);
664             break;
665         }
666     }
667 
668     /* Decide whether many devices are supported or not */
669     ret = i == ARRAY_SIZE(ioeventfds);
670 
671     while (i-- > 0) {
672         kvm_set_ioeventfd_pio(ioeventfds[i], 0, i, false, 2, true);
673         close(ioeventfds[i]);
674     }
675     return ret;
676 #else
677     return 0;
678 #endif
679 }
680 
681 static const KVMCapabilityInfo *
682 kvm_check_extension_list(KVMState *s, const KVMCapabilityInfo *list)
683 {
684     while (list->name) {
685         if (!kvm_check_extension(s, list->value)) {
686             return list;
687         }
688         list++;
689     }
690     return NULL;
691 }
692 
693 static void kvm_set_phys_mem(KVMMemoryListener *kml,
694                              MemoryRegionSection *section, bool add)
695 {
696     KVMSlot *mem;
697     int err;
698     MemoryRegion *mr = section->mr;
699     bool writeable = !mr->readonly && !mr->rom_device;
700     hwaddr start_addr, size;
701     void *ram;
702 
703     if (!memory_region_is_ram(mr)) {
704         if (writeable || !kvm_readonly_mem_allowed) {
705             return;
706         } else if (!mr->romd_mode) {
707             /* If the memory device is not in romd_mode, then we actually want
708              * to remove the kvm memory slot so all accesses will trap. */
709             add = false;
710         }
711     }
712 
713     size = kvm_align_section(section, &start_addr);
714     if (!size) {
715         return;
716     }
717 
718     /* use aligned delta to align the ram address */
719     ram = memory_region_get_ram_ptr(mr) + section->offset_within_region +
720           (start_addr - section->offset_within_address_space);
721 
722     if (!add) {
723         mem = kvm_lookup_matching_slot(kml, start_addr, size);
724         if (!mem) {
725             return;
726         }
727         if (mem->flags & KVM_MEM_LOG_DIRTY_PAGES) {
728             kvm_physical_sync_dirty_bitmap(kml, section);
729         }
730 
731         /* unregister the slot */
732         mem->memory_size = 0;
733         err = kvm_set_user_memory_region(kml, mem);
734         if (err) {
735             fprintf(stderr, "%s: error unregistering slot: %s\n",
736                     __func__, strerror(-err));
737             abort();
738         }
739         return;
740     }
741 
742     /* register the new slot */
743     mem = kvm_alloc_slot(kml);
744     mem->memory_size = size;
745     mem->start_addr = start_addr;
746     mem->ram = ram;
747     mem->flags = kvm_mem_flags(mr);
748 
749     err = kvm_set_user_memory_region(kml, mem);
750     if (err) {
751         fprintf(stderr, "%s: error registering slot: %s\n", __func__,
752                 strerror(-err));
753         abort();
754     }
755 }
756 
757 static void kvm_region_add(MemoryListener *listener,
758                            MemoryRegionSection *section)
759 {
760     KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
761 
762     memory_region_ref(section->mr);
763     kvm_set_phys_mem(kml, section, true);
764 }
765 
766 static void kvm_region_del(MemoryListener *listener,
767                            MemoryRegionSection *section)
768 {
769     KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
770 
771     kvm_set_phys_mem(kml, section, false);
772     memory_region_unref(section->mr);
773 }
774 
775 static void kvm_log_sync(MemoryListener *listener,
776                          MemoryRegionSection *section)
777 {
778     KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
779     int r;
780 
781     r = kvm_physical_sync_dirty_bitmap(kml, section);
782     if (r < 0) {
783         abort();
784     }
785 }
786 
787 static void kvm_mem_ioeventfd_add(MemoryListener *listener,
788                                   MemoryRegionSection *section,
789                                   bool match_data, uint64_t data,
790                                   EventNotifier *e)
791 {
792     int fd = event_notifier_get_fd(e);
793     int r;
794 
795     r = kvm_set_ioeventfd_mmio(fd, section->offset_within_address_space,
796                                data, true, int128_get64(section->size),
797                                match_data);
798     if (r < 0) {
799         fprintf(stderr, "%s: error adding ioeventfd: %s\n",
800                 __func__, strerror(-r));
801         abort();
802     }
803 }
804 
805 static void kvm_mem_ioeventfd_del(MemoryListener *listener,
806                                   MemoryRegionSection *section,
807                                   bool match_data, uint64_t data,
808                                   EventNotifier *e)
809 {
810     int fd = event_notifier_get_fd(e);
811     int r;
812 
813     r = kvm_set_ioeventfd_mmio(fd, section->offset_within_address_space,
814                                data, false, int128_get64(section->size),
815                                match_data);
816     if (r < 0) {
817         abort();
818     }
819 }
820 
821 static void kvm_io_ioeventfd_add(MemoryListener *listener,
822                                  MemoryRegionSection *section,
823                                  bool match_data, uint64_t data,
824                                  EventNotifier *e)
825 {
826     int fd = event_notifier_get_fd(e);
827     int r;
828 
829     r = kvm_set_ioeventfd_pio(fd, section->offset_within_address_space,
830                               data, true, int128_get64(section->size),
831                               match_data);
832     if (r < 0) {
833         fprintf(stderr, "%s: error adding ioeventfd: %s\n",
834                 __func__, strerror(-r));
835         abort();
836     }
837 }
838 
839 static void kvm_io_ioeventfd_del(MemoryListener *listener,
840                                  MemoryRegionSection *section,
841                                  bool match_data, uint64_t data,
842                                  EventNotifier *e)
843 
844 {
845     int fd = event_notifier_get_fd(e);
846     int r;
847 
848     r = kvm_set_ioeventfd_pio(fd, section->offset_within_address_space,
849                               data, false, int128_get64(section->size),
850                               match_data);
851     if (r < 0) {
852         abort();
853     }
854 }
855 
856 void kvm_memory_listener_register(KVMState *s, KVMMemoryListener *kml,
857                                   AddressSpace *as, int as_id)
858 {
859     int i;
860 
861     kml->slots = g_malloc0(s->nr_slots * sizeof(KVMSlot));
862     kml->as_id = as_id;
863 
864     for (i = 0; i < s->nr_slots; i++) {
865         kml->slots[i].slot = i;
866     }
867 
868     kml->listener.region_add = kvm_region_add;
869     kml->listener.region_del = kvm_region_del;
870     kml->listener.log_start = kvm_log_start;
871     kml->listener.log_stop = kvm_log_stop;
872     kml->listener.log_sync = kvm_log_sync;
873     kml->listener.priority = 10;
874 
875     memory_listener_register(&kml->listener, as);
876 }
877 
878 static MemoryListener kvm_io_listener = {
879     .eventfd_add = kvm_io_ioeventfd_add,
880     .eventfd_del = kvm_io_ioeventfd_del,
881     .priority = 10,
882 };
883 
884 int kvm_set_irq(KVMState *s, int irq, int level)
885 {
886     struct kvm_irq_level event;
887     int ret;
888 
889     assert(kvm_async_interrupts_enabled());
890 
891     event.level = level;
892     event.irq = irq;
893     ret = kvm_vm_ioctl(s, s->irq_set_ioctl, &event);
894     if (ret < 0) {
895         perror("kvm_set_irq");
896         abort();
897     }
898 
899     return (s->irq_set_ioctl == KVM_IRQ_LINE) ? 1 : event.status;
900 }
901 
902 #ifdef KVM_CAP_IRQ_ROUTING
903 typedef struct KVMMSIRoute {
904     struct kvm_irq_routing_entry kroute;
905     QTAILQ_ENTRY(KVMMSIRoute) entry;
906 } KVMMSIRoute;
907 
908 static void set_gsi(KVMState *s, unsigned int gsi)
909 {
910     set_bit(gsi, s->used_gsi_bitmap);
911 }
912 
913 static void clear_gsi(KVMState *s, unsigned int gsi)
914 {
915     clear_bit(gsi, s->used_gsi_bitmap);
916 }
917 
918 void kvm_init_irq_routing(KVMState *s)
919 {
920     int gsi_count, i;
921 
922     gsi_count = kvm_check_extension(s, KVM_CAP_IRQ_ROUTING) - 1;
923     if (gsi_count > 0) {
924         /* Round up so we can search ints using ffs */
925         s->used_gsi_bitmap = bitmap_new(gsi_count);
926         s->gsi_count = gsi_count;
927     }
928 
929     s->irq_routes = g_malloc0(sizeof(*s->irq_routes));
930     s->nr_allocated_irq_routes = 0;
931 
932     if (!kvm_direct_msi_allowed) {
933         for (i = 0; i < KVM_MSI_HASHTAB_SIZE; i++) {
934             QTAILQ_INIT(&s->msi_hashtab[i]);
935         }
936     }
937 
938     kvm_arch_init_irq_routing(s);
939 }
940 
941 void kvm_irqchip_commit_routes(KVMState *s)
942 {
943     int ret;
944 
945     if (kvm_gsi_direct_mapping()) {
946         return;
947     }
948 
949     if (!kvm_gsi_routing_enabled()) {
950         return;
951     }
952 
953     s->irq_routes->flags = 0;
954     trace_kvm_irqchip_commit_routes();
955     ret = kvm_vm_ioctl(s, KVM_SET_GSI_ROUTING, s->irq_routes);
956     assert(ret == 0);
957 }
958 
959 static void kvm_add_routing_entry(KVMState *s,
960                                   struct kvm_irq_routing_entry *entry)
961 {
962     struct kvm_irq_routing_entry *new;
963     int n, size;
964 
965     if (s->irq_routes->nr == s->nr_allocated_irq_routes) {
966         n = s->nr_allocated_irq_routes * 2;
967         if (n < 64) {
968             n = 64;
969         }
970         size = sizeof(struct kvm_irq_routing);
971         size += n * sizeof(*new);
972         s->irq_routes = g_realloc(s->irq_routes, size);
973         s->nr_allocated_irq_routes = n;
974     }
975     n = s->irq_routes->nr++;
976     new = &s->irq_routes->entries[n];
977 
978     *new = *entry;
979 
980     set_gsi(s, entry->gsi);
981 }
982 
983 static int kvm_update_routing_entry(KVMState *s,
984                                     struct kvm_irq_routing_entry *new_entry)
985 {
986     struct kvm_irq_routing_entry *entry;
987     int n;
988 
989     for (n = 0; n < s->irq_routes->nr; n++) {
990         entry = &s->irq_routes->entries[n];
991         if (entry->gsi != new_entry->gsi) {
992             continue;
993         }
994 
995         if(!memcmp(entry, new_entry, sizeof *entry)) {
996             return 0;
997         }
998 
999         *entry = *new_entry;
1000 
1001         return 0;
1002     }
1003 
1004     return -ESRCH;
1005 }
1006 
1007 void kvm_irqchip_add_irq_route(KVMState *s, int irq, int irqchip, int pin)
1008 {
1009     struct kvm_irq_routing_entry e = {};
1010 
1011     assert(pin < s->gsi_count);
1012 
1013     e.gsi = irq;
1014     e.type = KVM_IRQ_ROUTING_IRQCHIP;
1015     e.flags = 0;
1016     e.u.irqchip.irqchip = irqchip;
1017     e.u.irqchip.pin = pin;
1018     kvm_add_routing_entry(s, &e);
1019 }
1020 
1021 void kvm_irqchip_release_virq(KVMState *s, int virq)
1022 {
1023     struct kvm_irq_routing_entry *e;
1024     int i;
1025 
1026     if (kvm_gsi_direct_mapping()) {
1027         return;
1028     }
1029 
1030     for (i = 0; i < s->irq_routes->nr; i++) {
1031         e = &s->irq_routes->entries[i];
1032         if (e->gsi == virq) {
1033             s->irq_routes->nr--;
1034             *e = s->irq_routes->entries[s->irq_routes->nr];
1035         }
1036     }
1037     clear_gsi(s, virq);
1038     kvm_arch_release_virq_post(virq);
1039     trace_kvm_irqchip_release_virq(virq);
1040 }
1041 
1042 static unsigned int kvm_hash_msi(uint32_t data)
1043 {
1044     /* This is optimized for IA32 MSI layout. However, no other arch shall
1045      * repeat the mistake of not providing a direct MSI injection API. */
1046     return data & 0xff;
1047 }
1048 
1049 static void kvm_flush_dynamic_msi_routes(KVMState *s)
1050 {
1051     KVMMSIRoute *route, *next;
1052     unsigned int hash;
1053 
1054     for (hash = 0; hash < KVM_MSI_HASHTAB_SIZE; hash++) {
1055         QTAILQ_FOREACH_SAFE(route, &s->msi_hashtab[hash], entry, next) {
1056             kvm_irqchip_release_virq(s, route->kroute.gsi);
1057             QTAILQ_REMOVE(&s->msi_hashtab[hash], route, entry);
1058             g_free(route);
1059         }
1060     }
1061 }
1062 
1063 static int kvm_irqchip_get_virq(KVMState *s)
1064 {
1065     int next_virq;
1066 
1067     /*
1068      * PIC and IOAPIC share the first 16 GSI numbers, thus the available
1069      * GSI numbers are more than the number of IRQ route. Allocating a GSI
1070      * number can succeed even though a new route entry cannot be added.
1071      * When this happens, flush dynamic MSI entries to free IRQ route entries.
1072      */
1073     if (!kvm_direct_msi_allowed && s->irq_routes->nr == s->gsi_count) {
1074         kvm_flush_dynamic_msi_routes(s);
1075     }
1076 
1077     /* Return the lowest unused GSI in the bitmap */
1078     next_virq = find_first_zero_bit(s->used_gsi_bitmap, s->gsi_count);
1079     if (next_virq >= s->gsi_count) {
1080         return -ENOSPC;
1081     } else {
1082         return next_virq;
1083     }
1084 }
1085 
1086 static KVMMSIRoute *kvm_lookup_msi_route(KVMState *s, MSIMessage msg)
1087 {
1088     unsigned int hash = kvm_hash_msi(msg.data);
1089     KVMMSIRoute *route;
1090 
1091     QTAILQ_FOREACH(route, &s->msi_hashtab[hash], entry) {
1092         if (route->kroute.u.msi.address_lo == (uint32_t)msg.address &&
1093             route->kroute.u.msi.address_hi == (msg.address >> 32) &&
1094             route->kroute.u.msi.data == le32_to_cpu(msg.data)) {
1095             return route;
1096         }
1097     }
1098     return NULL;
1099 }
1100 
1101 int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg)
1102 {
1103     struct kvm_msi msi;
1104     KVMMSIRoute *route;
1105 
1106     if (kvm_direct_msi_allowed) {
1107         msi.address_lo = (uint32_t)msg.address;
1108         msi.address_hi = msg.address >> 32;
1109         msi.data = le32_to_cpu(msg.data);
1110         msi.flags = 0;
1111         memset(msi.pad, 0, sizeof(msi.pad));
1112 
1113         return kvm_vm_ioctl(s, KVM_SIGNAL_MSI, &msi);
1114     }
1115 
1116     route = kvm_lookup_msi_route(s, msg);
1117     if (!route) {
1118         int virq;
1119 
1120         virq = kvm_irqchip_get_virq(s);
1121         if (virq < 0) {
1122             return virq;
1123         }
1124 
1125         route = g_malloc0(sizeof(KVMMSIRoute));
1126         route->kroute.gsi = virq;
1127         route->kroute.type = KVM_IRQ_ROUTING_MSI;
1128         route->kroute.flags = 0;
1129         route->kroute.u.msi.address_lo = (uint32_t)msg.address;
1130         route->kroute.u.msi.address_hi = msg.address >> 32;
1131         route->kroute.u.msi.data = le32_to_cpu(msg.data);
1132 
1133         kvm_add_routing_entry(s, &route->kroute);
1134         kvm_irqchip_commit_routes(s);
1135 
1136         QTAILQ_INSERT_TAIL(&s->msi_hashtab[kvm_hash_msi(msg.data)], route,
1137                            entry);
1138     }
1139 
1140     assert(route->kroute.type == KVM_IRQ_ROUTING_MSI);
1141 
1142     return kvm_set_irq(s, route->kroute.gsi, 1);
1143 }
1144 
1145 int kvm_irqchip_add_msi_route(KVMState *s, int vector, PCIDevice *dev)
1146 {
1147     struct kvm_irq_routing_entry kroute = {};
1148     int virq;
1149     MSIMessage msg = {0, 0};
1150 
1151     if (pci_available && dev) {
1152         msg = pci_get_msi_message(dev, vector);
1153     }
1154 
1155     if (kvm_gsi_direct_mapping()) {
1156         return kvm_arch_msi_data_to_gsi(msg.data);
1157     }
1158 
1159     if (!kvm_gsi_routing_enabled()) {
1160         return -ENOSYS;
1161     }
1162 
1163     virq = kvm_irqchip_get_virq(s);
1164     if (virq < 0) {
1165         return virq;
1166     }
1167 
1168     kroute.gsi = virq;
1169     kroute.type = KVM_IRQ_ROUTING_MSI;
1170     kroute.flags = 0;
1171     kroute.u.msi.address_lo = (uint32_t)msg.address;
1172     kroute.u.msi.address_hi = msg.address >> 32;
1173     kroute.u.msi.data = le32_to_cpu(msg.data);
1174     if (pci_available && kvm_msi_devid_required()) {
1175         kroute.flags = KVM_MSI_VALID_DEVID;
1176         kroute.u.msi.devid = pci_requester_id(dev);
1177     }
1178     if (kvm_arch_fixup_msi_route(&kroute, msg.address, msg.data, dev)) {
1179         kvm_irqchip_release_virq(s, virq);
1180         return -EINVAL;
1181     }
1182 
1183     trace_kvm_irqchip_add_msi_route(dev ? dev->name : (char *)"N/A",
1184                                     vector, virq);
1185 
1186     kvm_add_routing_entry(s, &kroute);
1187     kvm_arch_add_msi_route_post(&kroute, vector, dev);
1188     kvm_irqchip_commit_routes(s);
1189 
1190     return virq;
1191 }
1192 
1193 int kvm_irqchip_update_msi_route(KVMState *s, int virq, MSIMessage msg,
1194                                  PCIDevice *dev)
1195 {
1196     struct kvm_irq_routing_entry kroute = {};
1197 
1198     if (kvm_gsi_direct_mapping()) {
1199         return 0;
1200     }
1201 
1202     if (!kvm_irqchip_in_kernel()) {
1203         return -ENOSYS;
1204     }
1205 
1206     kroute.gsi = virq;
1207     kroute.type = KVM_IRQ_ROUTING_MSI;
1208     kroute.flags = 0;
1209     kroute.u.msi.address_lo = (uint32_t)msg.address;
1210     kroute.u.msi.address_hi = msg.address >> 32;
1211     kroute.u.msi.data = le32_to_cpu(msg.data);
1212     if (pci_available && kvm_msi_devid_required()) {
1213         kroute.flags = KVM_MSI_VALID_DEVID;
1214         kroute.u.msi.devid = pci_requester_id(dev);
1215     }
1216     if (kvm_arch_fixup_msi_route(&kroute, msg.address, msg.data, dev)) {
1217         return -EINVAL;
1218     }
1219 
1220     trace_kvm_irqchip_update_msi_route(virq);
1221 
1222     return kvm_update_routing_entry(s, &kroute);
1223 }
1224 
1225 static int kvm_irqchip_assign_irqfd(KVMState *s, int fd, int rfd, int virq,
1226                                     bool assign)
1227 {
1228     struct kvm_irqfd irqfd = {
1229         .fd = fd,
1230         .gsi = virq,
1231         .flags = assign ? 0 : KVM_IRQFD_FLAG_DEASSIGN,
1232     };
1233 
1234     if (rfd != -1) {
1235         irqfd.flags |= KVM_IRQFD_FLAG_RESAMPLE;
1236         irqfd.resamplefd = rfd;
1237     }
1238 
1239     if (!kvm_irqfds_enabled()) {
1240         return -ENOSYS;
1241     }
1242 
1243     return kvm_vm_ioctl(s, KVM_IRQFD, &irqfd);
1244 }
1245 
1246 int kvm_irqchip_add_adapter_route(KVMState *s, AdapterInfo *adapter)
1247 {
1248     struct kvm_irq_routing_entry kroute = {};
1249     int virq;
1250 
1251     if (!kvm_gsi_routing_enabled()) {
1252         return -ENOSYS;
1253     }
1254 
1255     virq = kvm_irqchip_get_virq(s);
1256     if (virq < 0) {
1257         return virq;
1258     }
1259 
1260     kroute.gsi = virq;
1261     kroute.type = KVM_IRQ_ROUTING_S390_ADAPTER;
1262     kroute.flags = 0;
1263     kroute.u.adapter.summary_addr = adapter->summary_addr;
1264     kroute.u.adapter.ind_addr = adapter->ind_addr;
1265     kroute.u.adapter.summary_offset = adapter->summary_offset;
1266     kroute.u.adapter.ind_offset = adapter->ind_offset;
1267     kroute.u.adapter.adapter_id = adapter->adapter_id;
1268 
1269     kvm_add_routing_entry(s, &kroute);
1270 
1271     return virq;
1272 }
1273 
1274 int kvm_irqchip_add_hv_sint_route(KVMState *s, uint32_t vcpu, uint32_t sint)
1275 {
1276     struct kvm_irq_routing_entry kroute = {};
1277     int virq;
1278 
1279     if (!kvm_gsi_routing_enabled()) {
1280         return -ENOSYS;
1281     }
1282     if (!kvm_check_extension(s, KVM_CAP_HYPERV_SYNIC)) {
1283         return -ENOSYS;
1284     }
1285     virq = kvm_irqchip_get_virq(s);
1286     if (virq < 0) {
1287         return virq;
1288     }
1289 
1290     kroute.gsi = virq;
1291     kroute.type = KVM_IRQ_ROUTING_HV_SINT;
1292     kroute.flags = 0;
1293     kroute.u.hv_sint.vcpu = vcpu;
1294     kroute.u.hv_sint.sint = sint;
1295 
1296     kvm_add_routing_entry(s, &kroute);
1297     kvm_irqchip_commit_routes(s);
1298 
1299     return virq;
1300 }
1301 
1302 #else /* !KVM_CAP_IRQ_ROUTING */
1303 
1304 void kvm_init_irq_routing(KVMState *s)
1305 {
1306 }
1307 
1308 void kvm_irqchip_release_virq(KVMState *s, int virq)
1309 {
1310 }
1311 
1312 int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg)
1313 {
1314     abort();
1315 }
1316 
1317 int kvm_irqchip_add_msi_route(KVMState *s, int vector, PCIDevice *dev)
1318 {
1319     return -ENOSYS;
1320 }
1321 
1322 int kvm_irqchip_add_adapter_route(KVMState *s, AdapterInfo *adapter)
1323 {
1324     return -ENOSYS;
1325 }
1326 
1327 int kvm_irqchip_add_hv_sint_route(KVMState *s, uint32_t vcpu, uint32_t sint)
1328 {
1329     return -ENOSYS;
1330 }
1331 
1332 static int kvm_irqchip_assign_irqfd(KVMState *s, int fd, int virq, bool assign)
1333 {
1334     abort();
1335 }
1336 
1337 int kvm_irqchip_update_msi_route(KVMState *s, int virq, MSIMessage msg)
1338 {
1339     return -ENOSYS;
1340 }
1341 #endif /* !KVM_CAP_IRQ_ROUTING */
1342 
1343 int kvm_irqchip_add_irqfd_notifier_gsi(KVMState *s, EventNotifier *n,
1344                                        EventNotifier *rn, int virq)
1345 {
1346     return kvm_irqchip_assign_irqfd(s, event_notifier_get_fd(n),
1347            rn ? event_notifier_get_fd(rn) : -1, virq, true);
1348 }
1349 
1350 int kvm_irqchip_remove_irqfd_notifier_gsi(KVMState *s, EventNotifier *n,
1351                                           int virq)
1352 {
1353     return kvm_irqchip_assign_irqfd(s, event_notifier_get_fd(n), -1, virq,
1354            false);
1355 }
1356 
1357 int kvm_irqchip_add_irqfd_notifier(KVMState *s, EventNotifier *n,
1358                                    EventNotifier *rn, qemu_irq irq)
1359 {
1360     gpointer key, gsi;
1361     gboolean found = g_hash_table_lookup_extended(s->gsimap, irq, &key, &gsi);
1362 
1363     if (!found) {
1364         return -ENXIO;
1365     }
1366     return kvm_irqchip_add_irqfd_notifier_gsi(s, n, rn, GPOINTER_TO_INT(gsi));
1367 }
1368 
1369 int kvm_irqchip_remove_irqfd_notifier(KVMState *s, EventNotifier *n,
1370                                       qemu_irq irq)
1371 {
1372     gpointer key, gsi;
1373     gboolean found = g_hash_table_lookup_extended(s->gsimap, irq, &key, &gsi);
1374 
1375     if (!found) {
1376         return -ENXIO;
1377     }
1378     return kvm_irqchip_remove_irqfd_notifier_gsi(s, n, GPOINTER_TO_INT(gsi));
1379 }
1380 
1381 void kvm_irqchip_set_qemuirq_gsi(KVMState *s, qemu_irq irq, int gsi)
1382 {
1383     g_hash_table_insert(s->gsimap, irq, GINT_TO_POINTER(gsi));
1384 }
1385 
1386 static void kvm_irqchip_create(MachineState *machine, KVMState *s)
1387 {
1388     int ret;
1389 
1390     if (kvm_check_extension(s, KVM_CAP_IRQCHIP)) {
1391         ;
1392     } else if (kvm_check_extension(s, KVM_CAP_S390_IRQCHIP)) {
1393         ret = kvm_vm_enable_cap(s, KVM_CAP_S390_IRQCHIP, 0);
1394         if (ret < 0) {
1395             fprintf(stderr, "Enable kernel irqchip failed: %s\n", strerror(-ret));
1396             exit(1);
1397         }
1398     } else {
1399         return;
1400     }
1401 
1402     /* First probe and see if there's a arch-specific hook to create the
1403      * in-kernel irqchip for us */
1404     ret = kvm_arch_irqchip_create(machine, s);
1405     if (ret == 0) {
1406         if (machine_kernel_irqchip_split(machine)) {
1407             perror("Split IRQ chip mode not supported.");
1408             exit(1);
1409         } else {
1410             ret = kvm_vm_ioctl(s, KVM_CREATE_IRQCHIP);
1411         }
1412     }
1413     if (ret < 0) {
1414         fprintf(stderr, "Create kernel irqchip failed: %s\n", strerror(-ret));
1415         exit(1);
1416     }
1417 
1418     kvm_kernel_irqchip = true;
1419     /* If we have an in-kernel IRQ chip then we must have asynchronous
1420      * interrupt delivery (though the reverse is not necessarily true)
1421      */
1422     kvm_async_interrupts_allowed = true;
1423     kvm_halt_in_kernel_allowed = true;
1424 
1425     kvm_init_irq_routing(s);
1426 
1427     s->gsimap = g_hash_table_new(g_direct_hash, g_direct_equal);
1428 }
1429 
1430 /* Find number of supported CPUs using the recommended
1431  * procedure from the kernel API documentation to cope with
1432  * older kernels that may be missing capabilities.
1433  */
1434 static int kvm_recommended_vcpus(KVMState *s)
1435 {
1436     int ret = kvm_vm_check_extension(s, KVM_CAP_NR_VCPUS);
1437     return (ret) ? ret : 4;
1438 }
1439 
1440 static int kvm_max_vcpus(KVMState *s)
1441 {
1442     int ret = kvm_check_extension(s, KVM_CAP_MAX_VCPUS);
1443     return (ret) ? ret : kvm_recommended_vcpus(s);
1444 }
1445 
1446 static int kvm_max_vcpu_id(KVMState *s)
1447 {
1448     int ret = kvm_check_extension(s, KVM_CAP_MAX_VCPU_ID);
1449     return (ret) ? ret : kvm_max_vcpus(s);
1450 }
1451 
1452 bool kvm_vcpu_id_is_valid(int vcpu_id)
1453 {
1454     KVMState *s = KVM_STATE(current_machine->accelerator);
1455     return vcpu_id >= 0 && vcpu_id < kvm_max_vcpu_id(s);
1456 }
1457 
1458 static int kvm_init(MachineState *ms)
1459 {
1460     MachineClass *mc = MACHINE_GET_CLASS(ms);
1461     static const char upgrade_note[] =
1462         "Please upgrade to at least kernel 2.6.29 or recent kvm-kmod\n"
1463         "(see http://sourceforge.net/projects/kvm).\n";
1464     struct {
1465         const char *name;
1466         int num;
1467     } num_cpus[] = {
1468         { "SMP",          smp_cpus },
1469         { "hotpluggable", max_cpus },
1470         { NULL, }
1471     }, *nc = num_cpus;
1472     int soft_vcpus_limit, hard_vcpus_limit;
1473     KVMState *s;
1474     const KVMCapabilityInfo *missing_cap;
1475     int ret;
1476     int type = 0;
1477     const char *kvm_type;
1478 
1479     s = KVM_STATE(ms->accelerator);
1480 
1481     /*
1482      * On systems where the kernel can support different base page
1483      * sizes, host page size may be different from TARGET_PAGE_SIZE,
1484      * even with KVM.  TARGET_PAGE_SIZE is assumed to be the minimum
1485      * page size for the system though.
1486      */
1487     assert(TARGET_PAGE_SIZE <= getpagesize());
1488 
1489     s->sigmask_len = 8;
1490 
1491 #ifdef KVM_CAP_SET_GUEST_DEBUG
1492     QTAILQ_INIT(&s->kvm_sw_breakpoints);
1493 #endif
1494     QLIST_INIT(&s->kvm_parked_vcpus);
1495     s->vmfd = -1;
1496     s->fd = qemu_open("/dev/kvm", O_RDWR);
1497     if (s->fd == -1) {
1498         fprintf(stderr, "Could not access KVM kernel module: %m\n");
1499         ret = -errno;
1500         goto err;
1501     }
1502 
1503     ret = kvm_ioctl(s, KVM_GET_API_VERSION, 0);
1504     if (ret < KVM_API_VERSION) {
1505         if (ret >= 0) {
1506             ret = -EINVAL;
1507         }
1508         fprintf(stderr, "kvm version too old\n");
1509         goto err;
1510     }
1511 
1512     if (ret > KVM_API_VERSION) {
1513         ret = -EINVAL;
1514         fprintf(stderr, "kvm version not supported\n");
1515         goto err;
1516     }
1517 
1518     kvm_immediate_exit = kvm_check_extension(s, KVM_CAP_IMMEDIATE_EXIT);
1519     s->nr_slots = kvm_check_extension(s, KVM_CAP_NR_MEMSLOTS);
1520 
1521     /* If unspecified, use the default value */
1522     if (!s->nr_slots) {
1523         s->nr_slots = 32;
1524     }
1525 
1526     kvm_type = qemu_opt_get(qemu_get_machine_opts(), "kvm-type");
1527     if (mc->kvm_type) {
1528         type = mc->kvm_type(kvm_type);
1529     } else if (kvm_type) {
1530         ret = -EINVAL;
1531         fprintf(stderr, "Invalid argument kvm-type=%s\n", kvm_type);
1532         goto err;
1533     }
1534 
1535     do {
1536         ret = kvm_ioctl(s, KVM_CREATE_VM, type);
1537     } while (ret == -EINTR);
1538 
1539     if (ret < 0) {
1540         fprintf(stderr, "ioctl(KVM_CREATE_VM) failed: %d %s\n", -ret,
1541                 strerror(-ret));
1542 
1543 #ifdef TARGET_S390X
1544         if (ret == -EINVAL) {
1545             fprintf(stderr,
1546                     "Host kernel setup problem detected. Please verify:\n");
1547             fprintf(stderr, "- for kernels supporting the switch_amode or"
1548                     " user_mode parameters, whether\n");
1549             fprintf(stderr,
1550                     "  user space is running in primary address space\n");
1551             fprintf(stderr,
1552                     "- for kernels supporting the vm.allocate_pgste sysctl, "
1553                     "whether it is enabled\n");
1554         }
1555 #endif
1556         goto err;
1557     }
1558 
1559     s->vmfd = ret;
1560 
1561     /* check the vcpu limits */
1562     soft_vcpus_limit = kvm_recommended_vcpus(s);
1563     hard_vcpus_limit = kvm_max_vcpus(s);
1564 
1565     while (nc->name) {
1566         if (nc->num > soft_vcpus_limit) {
1567             warn_report("Number of %s cpus requested (%d) exceeds "
1568                         "the recommended cpus supported by KVM (%d)",
1569                         nc->name, nc->num, soft_vcpus_limit);
1570 
1571             if (nc->num > hard_vcpus_limit) {
1572                 fprintf(stderr, "Number of %s cpus requested (%d) exceeds "
1573                         "the maximum cpus supported by KVM (%d)\n",
1574                         nc->name, nc->num, hard_vcpus_limit);
1575                 exit(1);
1576             }
1577         }
1578         nc++;
1579     }
1580 
1581     missing_cap = kvm_check_extension_list(s, kvm_required_capabilites);
1582     if (!missing_cap) {
1583         missing_cap =
1584             kvm_check_extension_list(s, kvm_arch_required_capabilities);
1585     }
1586     if (missing_cap) {
1587         ret = -EINVAL;
1588         fprintf(stderr, "kvm does not support %s\n%s",
1589                 missing_cap->name, upgrade_note);
1590         goto err;
1591     }
1592 
1593     s->coalesced_mmio = kvm_check_extension(s, KVM_CAP_COALESCED_MMIO);
1594 
1595 #ifdef KVM_CAP_VCPU_EVENTS
1596     s->vcpu_events = kvm_check_extension(s, KVM_CAP_VCPU_EVENTS);
1597 #endif
1598 
1599     s->robust_singlestep =
1600         kvm_check_extension(s, KVM_CAP_X86_ROBUST_SINGLESTEP);
1601 
1602 #ifdef KVM_CAP_DEBUGREGS
1603     s->debugregs = kvm_check_extension(s, KVM_CAP_DEBUGREGS);
1604 #endif
1605 
1606 #ifdef KVM_CAP_IRQ_ROUTING
1607     kvm_direct_msi_allowed = (kvm_check_extension(s, KVM_CAP_SIGNAL_MSI) > 0);
1608 #endif
1609 
1610     s->intx_set_mask = kvm_check_extension(s, KVM_CAP_PCI_2_3);
1611 
1612     s->irq_set_ioctl = KVM_IRQ_LINE;
1613     if (kvm_check_extension(s, KVM_CAP_IRQ_INJECT_STATUS)) {
1614         s->irq_set_ioctl = KVM_IRQ_LINE_STATUS;
1615     }
1616 
1617 #ifdef KVM_CAP_READONLY_MEM
1618     kvm_readonly_mem_allowed =
1619         (kvm_check_extension(s, KVM_CAP_READONLY_MEM) > 0);
1620 #endif
1621 
1622     kvm_eventfds_allowed =
1623         (kvm_check_extension(s, KVM_CAP_IOEVENTFD) > 0);
1624 
1625     kvm_irqfds_allowed =
1626         (kvm_check_extension(s, KVM_CAP_IRQFD) > 0);
1627 
1628     kvm_resamplefds_allowed =
1629         (kvm_check_extension(s, KVM_CAP_IRQFD_RESAMPLE) > 0);
1630 
1631     kvm_vm_attributes_allowed =
1632         (kvm_check_extension(s, KVM_CAP_VM_ATTRIBUTES) > 0);
1633 
1634     kvm_ioeventfd_any_length_allowed =
1635         (kvm_check_extension(s, KVM_CAP_IOEVENTFD_ANY_LENGTH) > 0);
1636 
1637     kvm_state = s;
1638 
1639     ret = kvm_arch_init(ms, s);
1640     if (ret < 0) {
1641         goto err;
1642     }
1643 
1644     if (machine_kernel_irqchip_allowed(ms)) {
1645         kvm_irqchip_create(ms, s);
1646     }
1647 
1648     if (kvm_eventfds_allowed) {
1649         s->memory_listener.listener.eventfd_add = kvm_mem_ioeventfd_add;
1650         s->memory_listener.listener.eventfd_del = kvm_mem_ioeventfd_del;
1651     }
1652     s->memory_listener.listener.coalesced_mmio_add = kvm_coalesce_mmio_region;
1653     s->memory_listener.listener.coalesced_mmio_del = kvm_uncoalesce_mmio_region;
1654 
1655     kvm_memory_listener_register(s, &s->memory_listener,
1656                                  &address_space_memory, 0);
1657     memory_listener_register(&kvm_io_listener,
1658                              &address_space_io);
1659 
1660     s->many_ioeventfds = kvm_check_many_ioeventfds();
1661 
1662     s->sync_mmu = !!kvm_vm_check_extension(kvm_state, KVM_CAP_SYNC_MMU);
1663 
1664     return 0;
1665 
1666 err:
1667     assert(ret < 0);
1668     if (s->vmfd >= 0) {
1669         close(s->vmfd);
1670     }
1671     if (s->fd != -1) {
1672         close(s->fd);
1673     }
1674     g_free(s->memory_listener.slots);
1675 
1676     return ret;
1677 }
1678 
1679 void kvm_set_sigmask_len(KVMState *s, unsigned int sigmask_len)
1680 {
1681     s->sigmask_len = sigmask_len;
1682 }
1683 
1684 static void kvm_handle_io(uint16_t port, MemTxAttrs attrs, void *data, int direction,
1685                           int size, uint32_t count)
1686 {
1687     int i;
1688     uint8_t *ptr = data;
1689 
1690     for (i = 0; i < count; i++) {
1691         address_space_rw(&address_space_io, port, attrs,
1692                          ptr, size,
1693                          direction == KVM_EXIT_IO_OUT);
1694         ptr += size;
1695     }
1696 }
1697 
1698 static int kvm_handle_internal_error(CPUState *cpu, struct kvm_run *run)
1699 {
1700     fprintf(stderr, "KVM internal error. Suberror: %d\n",
1701             run->internal.suberror);
1702 
1703     if (kvm_check_extension(kvm_state, KVM_CAP_INTERNAL_ERROR_DATA)) {
1704         int i;
1705 
1706         for (i = 0; i < run->internal.ndata; ++i) {
1707             fprintf(stderr, "extra data[%d]: %"PRIx64"\n",
1708                     i, (uint64_t)run->internal.data[i]);
1709         }
1710     }
1711     if (run->internal.suberror == KVM_INTERNAL_ERROR_EMULATION) {
1712         fprintf(stderr, "emulation failure\n");
1713         if (!kvm_arch_stop_on_emulation_error(cpu)) {
1714             cpu_dump_state(cpu, stderr, fprintf, CPU_DUMP_CODE);
1715             return EXCP_INTERRUPT;
1716         }
1717     }
1718     /* FIXME: Should trigger a qmp message to let management know
1719      * something went wrong.
1720      */
1721     return -1;
1722 }
1723 
1724 void kvm_flush_coalesced_mmio_buffer(void)
1725 {
1726     KVMState *s = kvm_state;
1727 
1728     if (s->coalesced_flush_in_progress) {
1729         return;
1730     }
1731 
1732     s->coalesced_flush_in_progress = true;
1733 
1734     if (s->coalesced_mmio_ring) {
1735         struct kvm_coalesced_mmio_ring *ring = s->coalesced_mmio_ring;
1736         while (ring->first != ring->last) {
1737             struct kvm_coalesced_mmio *ent;
1738 
1739             ent = &ring->coalesced_mmio[ring->first];
1740 
1741             cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len);
1742             smp_wmb();
1743             ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX;
1744         }
1745     }
1746 
1747     s->coalesced_flush_in_progress = false;
1748 }
1749 
1750 static void do_kvm_cpu_synchronize_state(CPUState *cpu, run_on_cpu_data arg)
1751 {
1752     if (!cpu->vcpu_dirty) {
1753         kvm_arch_get_registers(cpu);
1754         cpu->vcpu_dirty = true;
1755     }
1756 }
1757 
1758 void kvm_cpu_synchronize_state(CPUState *cpu)
1759 {
1760     if (!cpu->vcpu_dirty) {
1761         run_on_cpu(cpu, do_kvm_cpu_synchronize_state, RUN_ON_CPU_NULL);
1762     }
1763 }
1764 
1765 static void do_kvm_cpu_synchronize_post_reset(CPUState *cpu, run_on_cpu_data arg)
1766 {
1767     kvm_arch_put_registers(cpu, KVM_PUT_RESET_STATE);
1768     cpu->vcpu_dirty = false;
1769 }
1770 
1771 void kvm_cpu_synchronize_post_reset(CPUState *cpu)
1772 {
1773     run_on_cpu(cpu, do_kvm_cpu_synchronize_post_reset, RUN_ON_CPU_NULL);
1774 }
1775 
1776 static void do_kvm_cpu_synchronize_post_init(CPUState *cpu, run_on_cpu_data arg)
1777 {
1778     kvm_arch_put_registers(cpu, KVM_PUT_FULL_STATE);
1779     cpu->vcpu_dirty = false;
1780 }
1781 
1782 void kvm_cpu_synchronize_post_init(CPUState *cpu)
1783 {
1784     run_on_cpu(cpu, do_kvm_cpu_synchronize_post_init, RUN_ON_CPU_NULL);
1785 }
1786 
1787 static void do_kvm_cpu_synchronize_pre_loadvm(CPUState *cpu, run_on_cpu_data arg)
1788 {
1789     cpu->vcpu_dirty = true;
1790 }
1791 
1792 void kvm_cpu_synchronize_pre_loadvm(CPUState *cpu)
1793 {
1794     run_on_cpu(cpu, do_kvm_cpu_synchronize_pre_loadvm, RUN_ON_CPU_NULL);
1795 }
1796 
1797 #ifdef KVM_HAVE_MCE_INJECTION
1798 static __thread void *pending_sigbus_addr;
1799 static __thread int pending_sigbus_code;
1800 static __thread bool have_sigbus_pending;
1801 #endif
1802 
1803 static void kvm_cpu_kick(CPUState *cpu)
1804 {
1805     atomic_set(&cpu->kvm_run->immediate_exit, 1);
1806 }
1807 
1808 static void kvm_cpu_kick_self(void)
1809 {
1810     if (kvm_immediate_exit) {
1811         kvm_cpu_kick(current_cpu);
1812     } else {
1813         qemu_cpu_kick_self();
1814     }
1815 }
1816 
1817 static void kvm_eat_signals(CPUState *cpu)
1818 {
1819     struct timespec ts = { 0, 0 };
1820     siginfo_t siginfo;
1821     sigset_t waitset;
1822     sigset_t chkset;
1823     int r;
1824 
1825     if (kvm_immediate_exit) {
1826         atomic_set(&cpu->kvm_run->immediate_exit, 0);
1827         /* Write kvm_run->immediate_exit before the cpu->exit_request
1828          * write in kvm_cpu_exec.
1829          */
1830         smp_wmb();
1831         return;
1832     }
1833 
1834     sigemptyset(&waitset);
1835     sigaddset(&waitset, SIG_IPI);
1836 
1837     do {
1838         r = sigtimedwait(&waitset, &siginfo, &ts);
1839         if (r == -1 && !(errno == EAGAIN || errno == EINTR)) {
1840             perror("sigtimedwait");
1841             exit(1);
1842         }
1843 
1844         r = sigpending(&chkset);
1845         if (r == -1) {
1846             perror("sigpending");
1847             exit(1);
1848         }
1849     } while (sigismember(&chkset, SIG_IPI));
1850 }
1851 
1852 int kvm_cpu_exec(CPUState *cpu)
1853 {
1854     struct kvm_run *run = cpu->kvm_run;
1855     int ret, run_ret;
1856 
1857     DPRINTF("kvm_cpu_exec()\n");
1858 
1859     if (kvm_arch_process_async_events(cpu)) {
1860         atomic_set(&cpu->exit_request, 0);
1861         return EXCP_HLT;
1862     }
1863 
1864     qemu_mutex_unlock_iothread();
1865     cpu_exec_start(cpu);
1866 
1867     do {
1868         MemTxAttrs attrs;
1869 
1870         if (cpu->vcpu_dirty) {
1871             kvm_arch_put_registers(cpu, KVM_PUT_RUNTIME_STATE);
1872             cpu->vcpu_dirty = false;
1873         }
1874 
1875         kvm_arch_pre_run(cpu, run);
1876         if (atomic_read(&cpu->exit_request)) {
1877             DPRINTF("interrupt exit requested\n");
1878             /*
1879              * KVM requires us to reenter the kernel after IO exits to complete
1880              * instruction emulation. This self-signal will ensure that we
1881              * leave ASAP again.
1882              */
1883             kvm_cpu_kick_self();
1884         }
1885 
1886         /* Read cpu->exit_request before KVM_RUN reads run->immediate_exit.
1887          * Matching barrier in kvm_eat_signals.
1888          */
1889         smp_rmb();
1890 
1891         run_ret = kvm_vcpu_ioctl(cpu, KVM_RUN, 0);
1892 
1893         attrs = kvm_arch_post_run(cpu, run);
1894 
1895 #ifdef KVM_HAVE_MCE_INJECTION
1896         if (unlikely(have_sigbus_pending)) {
1897             qemu_mutex_lock_iothread();
1898             kvm_arch_on_sigbus_vcpu(cpu, pending_sigbus_code,
1899                                     pending_sigbus_addr);
1900             have_sigbus_pending = false;
1901             qemu_mutex_unlock_iothread();
1902         }
1903 #endif
1904 
1905         if (run_ret < 0) {
1906             if (run_ret == -EINTR || run_ret == -EAGAIN) {
1907                 DPRINTF("io window exit\n");
1908                 kvm_eat_signals(cpu);
1909                 ret = EXCP_INTERRUPT;
1910                 break;
1911             }
1912             fprintf(stderr, "error: kvm run failed %s\n",
1913                     strerror(-run_ret));
1914 #ifdef TARGET_PPC
1915             if (run_ret == -EBUSY) {
1916                 fprintf(stderr,
1917                         "This is probably because your SMT is enabled.\n"
1918                         "VCPU can only run on primary threads with all "
1919                         "secondary threads offline.\n");
1920             }
1921 #endif
1922             ret = -1;
1923             break;
1924         }
1925 
1926         trace_kvm_run_exit(cpu->cpu_index, run->exit_reason);
1927         switch (run->exit_reason) {
1928         case KVM_EXIT_IO:
1929             DPRINTF("handle_io\n");
1930             /* Called outside BQL */
1931             kvm_handle_io(run->io.port, attrs,
1932                           (uint8_t *)run + run->io.data_offset,
1933                           run->io.direction,
1934                           run->io.size,
1935                           run->io.count);
1936             ret = 0;
1937             break;
1938         case KVM_EXIT_MMIO:
1939             DPRINTF("handle_mmio\n");
1940             /* Called outside BQL */
1941             address_space_rw(&address_space_memory,
1942                              run->mmio.phys_addr, attrs,
1943                              run->mmio.data,
1944                              run->mmio.len,
1945                              run->mmio.is_write);
1946             ret = 0;
1947             break;
1948         case KVM_EXIT_IRQ_WINDOW_OPEN:
1949             DPRINTF("irq_window_open\n");
1950             ret = EXCP_INTERRUPT;
1951             break;
1952         case KVM_EXIT_SHUTDOWN:
1953             DPRINTF("shutdown\n");
1954             qemu_system_reset_request(SHUTDOWN_CAUSE_GUEST_RESET);
1955             ret = EXCP_INTERRUPT;
1956             break;
1957         case KVM_EXIT_UNKNOWN:
1958             fprintf(stderr, "KVM: unknown exit, hardware reason %" PRIx64 "\n",
1959                     (uint64_t)run->hw.hardware_exit_reason);
1960             ret = -1;
1961             break;
1962         case KVM_EXIT_INTERNAL_ERROR:
1963             ret = kvm_handle_internal_error(cpu, run);
1964             break;
1965         case KVM_EXIT_SYSTEM_EVENT:
1966             switch (run->system_event.type) {
1967             case KVM_SYSTEM_EVENT_SHUTDOWN:
1968                 qemu_system_shutdown_request(SHUTDOWN_CAUSE_GUEST_SHUTDOWN);
1969                 ret = EXCP_INTERRUPT;
1970                 break;
1971             case KVM_SYSTEM_EVENT_RESET:
1972                 qemu_system_reset_request(SHUTDOWN_CAUSE_GUEST_RESET);
1973                 ret = EXCP_INTERRUPT;
1974                 break;
1975             case KVM_SYSTEM_EVENT_CRASH:
1976                 kvm_cpu_synchronize_state(cpu);
1977                 qemu_mutex_lock_iothread();
1978                 qemu_system_guest_panicked(cpu_get_crash_info(cpu));
1979                 qemu_mutex_unlock_iothread();
1980                 ret = 0;
1981                 break;
1982             default:
1983                 DPRINTF("kvm_arch_handle_exit\n");
1984                 ret = kvm_arch_handle_exit(cpu, run);
1985                 break;
1986             }
1987             break;
1988         default:
1989             DPRINTF("kvm_arch_handle_exit\n");
1990             ret = kvm_arch_handle_exit(cpu, run);
1991             break;
1992         }
1993     } while (ret == 0);
1994 
1995     cpu_exec_end(cpu);
1996     qemu_mutex_lock_iothread();
1997 
1998     if (ret < 0) {
1999         cpu_dump_state(cpu, stderr, fprintf, CPU_DUMP_CODE);
2000         vm_stop(RUN_STATE_INTERNAL_ERROR);
2001     }
2002 
2003     atomic_set(&cpu->exit_request, 0);
2004     return ret;
2005 }
2006 
2007 int kvm_ioctl(KVMState *s, int type, ...)
2008 {
2009     int ret;
2010     void *arg;
2011     va_list ap;
2012 
2013     va_start(ap, type);
2014     arg = va_arg(ap, void *);
2015     va_end(ap);
2016 
2017     trace_kvm_ioctl(type, arg);
2018     ret = ioctl(s->fd, type, arg);
2019     if (ret == -1) {
2020         ret = -errno;
2021     }
2022     return ret;
2023 }
2024 
2025 int kvm_vm_ioctl(KVMState *s, int type, ...)
2026 {
2027     int ret;
2028     void *arg;
2029     va_list ap;
2030 
2031     va_start(ap, type);
2032     arg = va_arg(ap, void *);
2033     va_end(ap);
2034 
2035     trace_kvm_vm_ioctl(type, arg);
2036     ret = ioctl(s->vmfd, type, arg);
2037     if (ret == -1) {
2038         ret = -errno;
2039     }
2040     return ret;
2041 }
2042 
2043 int kvm_vcpu_ioctl(CPUState *cpu, int type, ...)
2044 {
2045     int ret;
2046     void *arg;
2047     va_list ap;
2048 
2049     va_start(ap, type);
2050     arg = va_arg(ap, void *);
2051     va_end(ap);
2052 
2053     trace_kvm_vcpu_ioctl(cpu->cpu_index, type, arg);
2054     ret = ioctl(cpu->kvm_fd, type, arg);
2055     if (ret == -1) {
2056         ret = -errno;
2057     }
2058     return ret;
2059 }
2060 
2061 int kvm_device_ioctl(int fd, int type, ...)
2062 {
2063     int ret;
2064     void *arg;
2065     va_list ap;
2066 
2067     va_start(ap, type);
2068     arg = va_arg(ap, void *);
2069     va_end(ap);
2070 
2071     trace_kvm_device_ioctl(fd, type, arg);
2072     ret = ioctl(fd, type, arg);
2073     if (ret == -1) {
2074         ret = -errno;
2075     }
2076     return ret;
2077 }
2078 
2079 int kvm_vm_check_attr(KVMState *s, uint32_t group, uint64_t attr)
2080 {
2081     int ret;
2082     struct kvm_device_attr attribute = {
2083         .group = group,
2084         .attr = attr,
2085     };
2086 
2087     if (!kvm_vm_attributes_allowed) {
2088         return 0;
2089     }
2090 
2091     ret = kvm_vm_ioctl(s, KVM_HAS_DEVICE_ATTR, &attribute);
2092     /* kvm returns 0 on success for HAS_DEVICE_ATTR */
2093     return ret ? 0 : 1;
2094 }
2095 
2096 int kvm_device_check_attr(int dev_fd, uint32_t group, uint64_t attr)
2097 {
2098     struct kvm_device_attr attribute = {
2099         .group = group,
2100         .attr = attr,
2101         .flags = 0,
2102     };
2103 
2104     return kvm_device_ioctl(dev_fd, KVM_HAS_DEVICE_ATTR, &attribute) ? 0 : 1;
2105 }
2106 
2107 int kvm_device_access(int fd, int group, uint64_t attr,
2108                       void *val, bool write, Error **errp)
2109 {
2110     struct kvm_device_attr kvmattr;
2111     int err;
2112 
2113     kvmattr.flags = 0;
2114     kvmattr.group = group;
2115     kvmattr.attr = attr;
2116     kvmattr.addr = (uintptr_t)val;
2117 
2118     err = kvm_device_ioctl(fd,
2119                            write ? KVM_SET_DEVICE_ATTR : KVM_GET_DEVICE_ATTR,
2120                            &kvmattr);
2121     if (err < 0) {
2122         error_setg_errno(errp, -err,
2123                          "KVM_%s_DEVICE_ATTR failed: Group %d "
2124                          "attr 0x%016" PRIx64,
2125                          write ? "SET" : "GET", group, attr);
2126     }
2127     return err;
2128 }
2129 
2130 bool kvm_has_sync_mmu(void)
2131 {
2132     return kvm_state->sync_mmu;
2133 }
2134 
2135 int kvm_has_vcpu_events(void)
2136 {
2137     return kvm_state->vcpu_events;
2138 }
2139 
2140 int kvm_has_robust_singlestep(void)
2141 {
2142     return kvm_state->robust_singlestep;
2143 }
2144 
2145 int kvm_has_debugregs(void)
2146 {
2147     return kvm_state->debugregs;
2148 }
2149 
2150 int kvm_has_many_ioeventfds(void)
2151 {
2152     if (!kvm_enabled()) {
2153         return 0;
2154     }
2155     return kvm_state->many_ioeventfds;
2156 }
2157 
2158 int kvm_has_gsi_routing(void)
2159 {
2160 #ifdef KVM_CAP_IRQ_ROUTING
2161     return kvm_check_extension(kvm_state, KVM_CAP_IRQ_ROUTING);
2162 #else
2163     return false;
2164 #endif
2165 }
2166 
2167 int kvm_has_intx_set_mask(void)
2168 {
2169     return kvm_state->intx_set_mask;
2170 }
2171 
2172 bool kvm_arm_supports_user_irq(void)
2173 {
2174     return kvm_check_extension(kvm_state, KVM_CAP_ARM_USER_IRQ);
2175 }
2176 
2177 #ifdef KVM_CAP_SET_GUEST_DEBUG
2178 struct kvm_sw_breakpoint *kvm_find_sw_breakpoint(CPUState *cpu,
2179                                                  target_ulong pc)
2180 {
2181     struct kvm_sw_breakpoint *bp;
2182 
2183     QTAILQ_FOREACH(bp, &cpu->kvm_state->kvm_sw_breakpoints, entry) {
2184         if (bp->pc == pc) {
2185             return bp;
2186         }
2187     }
2188     return NULL;
2189 }
2190 
2191 int kvm_sw_breakpoints_active(CPUState *cpu)
2192 {
2193     return !QTAILQ_EMPTY(&cpu->kvm_state->kvm_sw_breakpoints);
2194 }
2195 
2196 struct kvm_set_guest_debug_data {
2197     struct kvm_guest_debug dbg;
2198     int err;
2199 };
2200 
2201 static void kvm_invoke_set_guest_debug(CPUState *cpu, run_on_cpu_data data)
2202 {
2203     struct kvm_set_guest_debug_data *dbg_data =
2204         (struct kvm_set_guest_debug_data *) data.host_ptr;
2205 
2206     dbg_data->err = kvm_vcpu_ioctl(cpu, KVM_SET_GUEST_DEBUG,
2207                                    &dbg_data->dbg);
2208 }
2209 
2210 int kvm_update_guest_debug(CPUState *cpu, unsigned long reinject_trap)
2211 {
2212     struct kvm_set_guest_debug_data data;
2213 
2214     data.dbg.control = reinject_trap;
2215 
2216     if (cpu->singlestep_enabled) {
2217         data.dbg.control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_SINGLESTEP;
2218     }
2219     kvm_arch_update_guest_debug(cpu, &data.dbg);
2220 
2221     run_on_cpu(cpu, kvm_invoke_set_guest_debug,
2222                RUN_ON_CPU_HOST_PTR(&data));
2223     return data.err;
2224 }
2225 
2226 int kvm_insert_breakpoint(CPUState *cpu, target_ulong addr,
2227                           target_ulong len, int type)
2228 {
2229     struct kvm_sw_breakpoint *bp;
2230     int err;
2231 
2232     if (type == GDB_BREAKPOINT_SW) {
2233         bp = kvm_find_sw_breakpoint(cpu, addr);
2234         if (bp) {
2235             bp->use_count++;
2236             return 0;
2237         }
2238 
2239         bp = g_malloc(sizeof(struct kvm_sw_breakpoint));
2240         bp->pc = addr;
2241         bp->use_count = 1;
2242         err = kvm_arch_insert_sw_breakpoint(cpu, bp);
2243         if (err) {
2244             g_free(bp);
2245             return err;
2246         }
2247 
2248         QTAILQ_INSERT_HEAD(&cpu->kvm_state->kvm_sw_breakpoints, bp, entry);
2249     } else {
2250         err = kvm_arch_insert_hw_breakpoint(addr, len, type);
2251         if (err) {
2252             return err;
2253         }
2254     }
2255 
2256     CPU_FOREACH(cpu) {
2257         err = kvm_update_guest_debug(cpu, 0);
2258         if (err) {
2259             return err;
2260         }
2261     }
2262     return 0;
2263 }
2264 
2265 int kvm_remove_breakpoint(CPUState *cpu, target_ulong addr,
2266                           target_ulong len, int type)
2267 {
2268     struct kvm_sw_breakpoint *bp;
2269     int err;
2270 
2271     if (type == GDB_BREAKPOINT_SW) {
2272         bp = kvm_find_sw_breakpoint(cpu, addr);
2273         if (!bp) {
2274             return -ENOENT;
2275         }
2276 
2277         if (bp->use_count > 1) {
2278             bp->use_count--;
2279             return 0;
2280         }
2281 
2282         err = kvm_arch_remove_sw_breakpoint(cpu, bp);
2283         if (err) {
2284             return err;
2285         }
2286 
2287         QTAILQ_REMOVE(&cpu->kvm_state->kvm_sw_breakpoints, bp, entry);
2288         g_free(bp);
2289     } else {
2290         err = kvm_arch_remove_hw_breakpoint(addr, len, type);
2291         if (err) {
2292             return err;
2293         }
2294     }
2295 
2296     CPU_FOREACH(cpu) {
2297         err = kvm_update_guest_debug(cpu, 0);
2298         if (err) {
2299             return err;
2300         }
2301     }
2302     return 0;
2303 }
2304 
2305 void kvm_remove_all_breakpoints(CPUState *cpu)
2306 {
2307     struct kvm_sw_breakpoint *bp, *next;
2308     KVMState *s = cpu->kvm_state;
2309     CPUState *tmpcpu;
2310 
2311     QTAILQ_FOREACH_SAFE(bp, &s->kvm_sw_breakpoints, entry, next) {
2312         if (kvm_arch_remove_sw_breakpoint(cpu, bp) != 0) {
2313             /* Try harder to find a CPU that currently sees the breakpoint. */
2314             CPU_FOREACH(tmpcpu) {
2315                 if (kvm_arch_remove_sw_breakpoint(tmpcpu, bp) == 0) {
2316                     break;
2317                 }
2318             }
2319         }
2320         QTAILQ_REMOVE(&s->kvm_sw_breakpoints, bp, entry);
2321         g_free(bp);
2322     }
2323     kvm_arch_remove_all_hw_breakpoints();
2324 
2325     CPU_FOREACH(cpu) {
2326         kvm_update_guest_debug(cpu, 0);
2327     }
2328 }
2329 
2330 #else /* !KVM_CAP_SET_GUEST_DEBUG */
2331 
2332 int kvm_update_guest_debug(CPUState *cpu, unsigned long reinject_trap)
2333 {
2334     return -EINVAL;
2335 }
2336 
2337 int kvm_insert_breakpoint(CPUState *cpu, target_ulong addr,
2338                           target_ulong len, int type)
2339 {
2340     return -EINVAL;
2341 }
2342 
2343 int kvm_remove_breakpoint(CPUState *cpu, target_ulong addr,
2344                           target_ulong len, int type)
2345 {
2346     return -EINVAL;
2347 }
2348 
2349 void kvm_remove_all_breakpoints(CPUState *cpu)
2350 {
2351 }
2352 #endif /* !KVM_CAP_SET_GUEST_DEBUG */
2353 
2354 static int kvm_set_signal_mask(CPUState *cpu, const sigset_t *sigset)
2355 {
2356     KVMState *s = kvm_state;
2357     struct kvm_signal_mask *sigmask;
2358     int r;
2359 
2360     sigmask = g_malloc(sizeof(*sigmask) + sizeof(*sigset));
2361 
2362     sigmask->len = s->sigmask_len;
2363     memcpy(sigmask->sigset, sigset, sizeof(*sigset));
2364     r = kvm_vcpu_ioctl(cpu, KVM_SET_SIGNAL_MASK, sigmask);
2365     g_free(sigmask);
2366 
2367     return r;
2368 }
2369 
2370 static void kvm_ipi_signal(int sig)
2371 {
2372     if (current_cpu) {
2373         assert(kvm_immediate_exit);
2374         kvm_cpu_kick(current_cpu);
2375     }
2376 }
2377 
2378 void kvm_init_cpu_signals(CPUState *cpu)
2379 {
2380     int r;
2381     sigset_t set;
2382     struct sigaction sigact;
2383 
2384     memset(&sigact, 0, sizeof(sigact));
2385     sigact.sa_handler = kvm_ipi_signal;
2386     sigaction(SIG_IPI, &sigact, NULL);
2387 
2388     pthread_sigmask(SIG_BLOCK, NULL, &set);
2389 #if defined KVM_HAVE_MCE_INJECTION
2390     sigdelset(&set, SIGBUS);
2391     pthread_sigmask(SIG_SETMASK, &set, NULL);
2392 #endif
2393     sigdelset(&set, SIG_IPI);
2394     if (kvm_immediate_exit) {
2395         r = pthread_sigmask(SIG_SETMASK, &set, NULL);
2396     } else {
2397         r = kvm_set_signal_mask(cpu, &set);
2398     }
2399     if (r) {
2400         fprintf(stderr, "kvm_set_signal_mask: %s\n", strerror(-r));
2401         exit(1);
2402     }
2403 }
2404 
2405 /* Called asynchronously in VCPU thread.  */
2406 int kvm_on_sigbus_vcpu(CPUState *cpu, int code, void *addr)
2407 {
2408 #ifdef KVM_HAVE_MCE_INJECTION
2409     if (have_sigbus_pending) {
2410         return 1;
2411     }
2412     have_sigbus_pending = true;
2413     pending_sigbus_addr = addr;
2414     pending_sigbus_code = code;
2415     atomic_set(&cpu->exit_request, 1);
2416     return 0;
2417 #else
2418     return 1;
2419 #endif
2420 }
2421 
2422 /* Called synchronously (via signalfd) in main thread.  */
2423 int kvm_on_sigbus(int code, void *addr)
2424 {
2425 #ifdef KVM_HAVE_MCE_INJECTION
2426     /* Action required MCE kills the process if SIGBUS is blocked.  Because
2427      * that's what happens in the I/O thread, where we handle MCE via signalfd,
2428      * we can only get action optional here.
2429      */
2430     assert(code != BUS_MCEERR_AR);
2431     kvm_arch_on_sigbus_vcpu(first_cpu, code, addr);
2432     return 0;
2433 #else
2434     return 1;
2435 #endif
2436 }
2437 
2438 int kvm_create_device(KVMState *s, uint64_t type, bool test)
2439 {
2440     int ret;
2441     struct kvm_create_device create_dev;
2442 
2443     create_dev.type = type;
2444     create_dev.fd = -1;
2445     create_dev.flags = test ? KVM_CREATE_DEVICE_TEST : 0;
2446 
2447     if (!kvm_check_extension(s, KVM_CAP_DEVICE_CTRL)) {
2448         return -ENOTSUP;
2449     }
2450 
2451     ret = kvm_vm_ioctl(s, KVM_CREATE_DEVICE, &create_dev);
2452     if (ret) {
2453         return ret;
2454     }
2455 
2456     return test ? 0 : create_dev.fd;
2457 }
2458 
2459 bool kvm_device_supported(int vmfd, uint64_t type)
2460 {
2461     struct kvm_create_device create_dev = {
2462         .type = type,
2463         .fd = -1,
2464         .flags = KVM_CREATE_DEVICE_TEST,
2465     };
2466 
2467     if (ioctl(vmfd, KVM_CHECK_EXTENSION, KVM_CAP_DEVICE_CTRL) <= 0) {
2468         return false;
2469     }
2470 
2471     return (ioctl(vmfd, KVM_CREATE_DEVICE, &create_dev) >= 0);
2472 }
2473 
2474 int kvm_set_one_reg(CPUState *cs, uint64_t id, void *source)
2475 {
2476     struct kvm_one_reg reg;
2477     int r;
2478 
2479     reg.id = id;
2480     reg.addr = (uintptr_t) source;
2481     r = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, &reg);
2482     if (r) {
2483         trace_kvm_failed_reg_set(id, strerror(-r));
2484     }
2485     return r;
2486 }
2487 
2488 int kvm_get_one_reg(CPUState *cs, uint64_t id, void *target)
2489 {
2490     struct kvm_one_reg reg;
2491     int r;
2492 
2493     reg.id = id;
2494     reg.addr = (uintptr_t) target;
2495     r = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &reg);
2496     if (r) {
2497         trace_kvm_failed_reg_get(id, strerror(-r));
2498     }
2499     return r;
2500 }
2501 
2502 static void kvm_accel_class_init(ObjectClass *oc, void *data)
2503 {
2504     AccelClass *ac = ACCEL_CLASS(oc);
2505     ac->name = "KVM";
2506     ac->init_machine = kvm_init;
2507     ac->allowed = &kvm_allowed;
2508 }
2509 
2510 static const TypeInfo kvm_accel_type = {
2511     .name = TYPE_KVM_ACCEL,
2512     .parent = TYPE_ACCEL,
2513     .class_init = kvm_accel_class_init,
2514     .instance_size = sizeof(KVMState),
2515 };
2516 
2517 static void kvm_type_init(void)
2518 {
2519     type_register_static(&kvm_accel_type);
2520 }
2521 
2522 type_init(kvm_type_init);
2523