xref: /openbmc/qemu/accel/kvm/kvm-all.c (revision 6598f0cd)
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 #include "sysemu/sev.h"
42 
43 #include "hw/boards.h"
44 
45 /* This check must be after config-host.h is included */
46 #ifdef CONFIG_EVENTFD
47 #include <sys/eventfd.h>
48 #endif
49 
50 /* KVM uses PAGE_SIZE in its definition of KVM_COALESCED_MMIO_MAX. We
51  * need to use the real host PAGE_SIZE, as that's what KVM will use.
52  */
53 #define PAGE_SIZE getpagesize()
54 
55 //#define DEBUG_KVM
56 
57 #ifdef DEBUG_KVM
58 #define DPRINTF(fmt, ...) \
59     do { fprintf(stderr, fmt, ## __VA_ARGS__); } while (0)
60 #else
61 #define DPRINTF(fmt, ...) \
62     do { } while (0)
63 #endif
64 
65 #define KVM_MSI_HASHTAB_SIZE    256
66 
67 struct KVMParkedVcpu {
68     unsigned long vcpu_id;
69     int kvm_fd;
70     QLIST_ENTRY(KVMParkedVcpu) node;
71 };
72 
73 struct KVMState
74 {
75     AccelState parent_obj;
76 
77     int nr_slots;
78     int fd;
79     int vmfd;
80     int coalesced_mmio;
81     struct kvm_coalesced_mmio_ring *coalesced_mmio_ring;
82     bool coalesced_flush_in_progress;
83     int vcpu_events;
84     int robust_singlestep;
85     int debugregs;
86 #ifdef KVM_CAP_SET_GUEST_DEBUG
87     struct kvm_sw_breakpoint_head kvm_sw_breakpoints;
88 #endif
89     int many_ioeventfds;
90     int intx_set_mask;
91     bool sync_mmu;
92     /* The man page (and posix) say ioctl numbers are signed int, but
93      * they're not.  Linux, glibc and *BSD all treat ioctl numbers as
94      * unsigned, and treating them as signed here can break things */
95     unsigned irq_set_ioctl;
96     unsigned int sigmask_len;
97     GHashTable *gsimap;
98 #ifdef KVM_CAP_IRQ_ROUTING
99     struct kvm_irq_routing *irq_routes;
100     int nr_allocated_irq_routes;
101     unsigned long *used_gsi_bitmap;
102     unsigned int gsi_count;
103     QTAILQ_HEAD(msi_hashtab, KVMMSIRoute) msi_hashtab[KVM_MSI_HASHTAB_SIZE];
104 #endif
105     KVMMemoryListener memory_listener;
106     QLIST_HEAD(, KVMParkedVcpu) kvm_parked_vcpus;
107 
108     /* memory encryption */
109     void *memcrypt_handle;
110     int (*memcrypt_encrypt_data)(void *handle, uint8_t *ptr, uint64_t len);
111 };
112 
113 KVMState *kvm_state;
114 bool kvm_kernel_irqchip;
115 bool kvm_split_irqchip;
116 bool kvm_async_interrupts_allowed;
117 bool kvm_halt_in_kernel_allowed;
118 bool kvm_eventfds_allowed;
119 bool kvm_irqfds_allowed;
120 bool kvm_resamplefds_allowed;
121 bool kvm_msi_via_irqfd_allowed;
122 bool kvm_gsi_routing_allowed;
123 bool kvm_gsi_direct_mapping;
124 bool kvm_allowed;
125 bool kvm_readonly_mem_allowed;
126 bool kvm_vm_attributes_allowed;
127 bool kvm_direct_msi_allowed;
128 bool kvm_ioeventfd_any_length_allowed;
129 bool kvm_msi_use_devid;
130 static bool kvm_immediate_exit;
131 
132 static const KVMCapabilityInfo kvm_required_capabilites[] = {
133     KVM_CAP_INFO(USER_MEMORY),
134     KVM_CAP_INFO(DESTROY_MEMORY_REGION_WORKS),
135     KVM_CAP_INFO(JOIN_MEMORY_REGIONS_WORKS),
136     KVM_CAP_LAST_INFO
137 };
138 
139 int kvm_get_max_memslots(void)
140 {
141     KVMState *s = KVM_STATE(current_machine->accelerator);
142 
143     return s->nr_slots;
144 }
145 
146 bool kvm_memcrypt_enabled(void)
147 {
148     if (kvm_state && kvm_state->memcrypt_handle) {
149         return true;
150     }
151 
152     return false;
153 }
154 
155 int kvm_memcrypt_encrypt_data(uint8_t *ptr, uint64_t len)
156 {
157     if (kvm_state->memcrypt_handle &&
158         kvm_state->memcrypt_encrypt_data) {
159         return kvm_state->memcrypt_encrypt_data(kvm_state->memcrypt_handle,
160                                               ptr, len);
161     }
162 
163     return 1;
164 }
165 
166 static KVMSlot *kvm_get_free_slot(KVMMemoryListener *kml)
167 {
168     KVMState *s = kvm_state;
169     int i;
170 
171     for (i = 0; i < s->nr_slots; i++) {
172         if (kml->slots[i].memory_size == 0) {
173             return &kml->slots[i];
174         }
175     }
176 
177     return NULL;
178 }
179 
180 bool kvm_has_free_slot(MachineState *ms)
181 {
182     KVMState *s = KVM_STATE(ms->accelerator);
183 
184     return kvm_get_free_slot(&s->memory_listener);
185 }
186 
187 static KVMSlot *kvm_alloc_slot(KVMMemoryListener *kml)
188 {
189     KVMSlot *slot = kvm_get_free_slot(kml);
190 
191     if (slot) {
192         return slot;
193     }
194 
195     fprintf(stderr, "%s: no free slot available\n", __func__);
196     abort();
197 }
198 
199 static KVMSlot *kvm_lookup_matching_slot(KVMMemoryListener *kml,
200                                          hwaddr start_addr,
201                                          hwaddr size)
202 {
203     KVMState *s = kvm_state;
204     int i;
205 
206     for (i = 0; i < s->nr_slots; i++) {
207         KVMSlot *mem = &kml->slots[i];
208 
209         if (start_addr == mem->start_addr && size == mem->memory_size) {
210             return mem;
211         }
212     }
213 
214     return NULL;
215 }
216 
217 /*
218  * Calculate and align the start address and the size of the section.
219  * Return the size. If the size is 0, the aligned section is empty.
220  */
221 static hwaddr kvm_align_section(MemoryRegionSection *section,
222                                 hwaddr *start)
223 {
224     hwaddr size = int128_get64(section->size);
225     hwaddr delta, aligned;
226 
227     /* kvm works in page size chunks, but the function may be called
228        with sub-page size and unaligned start address. Pad the start
229        address to next and truncate size to previous page boundary. */
230     aligned = ROUND_UP(section->offset_within_address_space,
231                        qemu_real_host_page_size);
232     delta = aligned - section->offset_within_address_space;
233     *start = aligned;
234     if (delta > size) {
235         return 0;
236     }
237 
238     return (size - delta) & qemu_real_host_page_mask;
239 }
240 
241 int kvm_physical_memory_addr_from_host(KVMState *s, void *ram,
242                                        hwaddr *phys_addr)
243 {
244     KVMMemoryListener *kml = &s->memory_listener;
245     int i;
246 
247     for (i = 0; i < s->nr_slots; i++) {
248         KVMSlot *mem = &kml->slots[i];
249 
250         if (ram >= mem->ram && ram < mem->ram + mem->memory_size) {
251             *phys_addr = mem->start_addr + (ram - mem->ram);
252             return 1;
253         }
254     }
255 
256     return 0;
257 }
258 
259 static int kvm_set_user_memory_region(KVMMemoryListener *kml, KVMSlot *slot, bool new)
260 {
261     KVMState *s = kvm_state;
262     struct kvm_userspace_memory_region mem;
263     int ret;
264 
265     mem.slot = slot->slot | (kml->as_id << 16);
266     mem.guest_phys_addr = slot->start_addr;
267     mem.userspace_addr = (unsigned long)slot->ram;
268     mem.flags = slot->flags;
269 
270     if (slot->memory_size && !new && (mem.flags ^ slot->old_flags) & KVM_MEM_READONLY) {
271         /* Set the slot size to 0 before setting the slot to the desired
272          * value. This is needed based on KVM commit 75d61fbc. */
273         mem.memory_size = 0;
274         kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
275     }
276     mem.memory_size = slot->memory_size;
277     ret = kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
278     slot->old_flags = mem.flags;
279     trace_kvm_set_user_memory(mem.slot, mem.flags, mem.guest_phys_addr,
280                               mem.memory_size, mem.userspace_addr, ret);
281     return ret;
282 }
283 
284 int kvm_destroy_vcpu(CPUState *cpu)
285 {
286     KVMState *s = kvm_state;
287     long mmap_size;
288     struct KVMParkedVcpu *vcpu = NULL;
289     int ret = 0;
290 
291     DPRINTF("kvm_destroy_vcpu\n");
292 
293     mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
294     if (mmap_size < 0) {
295         ret = mmap_size;
296         DPRINTF("KVM_GET_VCPU_MMAP_SIZE failed\n");
297         goto err;
298     }
299 
300     ret = munmap(cpu->kvm_run, mmap_size);
301     if (ret < 0) {
302         goto err;
303     }
304 
305     vcpu = g_malloc0(sizeof(*vcpu));
306     vcpu->vcpu_id = kvm_arch_vcpu_id(cpu);
307     vcpu->kvm_fd = cpu->kvm_fd;
308     QLIST_INSERT_HEAD(&kvm_state->kvm_parked_vcpus, vcpu, node);
309 err:
310     return ret;
311 }
312 
313 static int kvm_get_vcpu(KVMState *s, unsigned long vcpu_id)
314 {
315     struct KVMParkedVcpu *cpu;
316 
317     QLIST_FOREACH(cpu, &s->kvm_parked_vcpus, node) {
318         if (cpu->vcpu_id == vcpu_id) {
319             int kvm_fd;
320 
321             QLIST_REMOVE(cpu, node);
322             kvm_fd = cpu->kvm_fd;
323             g_free(cpu);
324             return kvm_fd;
325         }
326     }
327 
328     return kvm_vm_ioctl(s, KVM_CREATE_VCPU, (void *)vcpu_id);
329 }
330 
331 int kvm_init_vcpu(CPUState *cpu)
332 {
333     KVMState *s = kvm_state;
334     long mmap_size;
335     int ret;
336 
337     DPRINTF("kvm_init_vcpu\n");
338 
339     ret = kvm_get_vcpu(s, kvm_arch_vcpu_id(cpu));
340     if (ret < 0) {
341         DPRINTF("kvm_create_vcpu failed\n");
342         goto err;
343     }
344 
345     cpu->kvm_fd = ret;
346     cpu->kvm_state = s;
347     cpu->vcpu_dirty = true;
348 
349     mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
350     if (mmap_size < 0) {
351         ret = mmap_size;
352         DPRINTF("KVM_GET_VCPU_MMAP_SIZE failed\n");
353         goto err;
354     }
355 
356     cpu->kvm_run = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
357                         cpu->kvm_fd, 0);
358     if (cpu->kvm_run == MAP_FAILED) {
359         ret = -errno;
360         DPRINTF("mmap'ing vcpu state failed\n");
361         goto err;
362     }
363 
364     if (s->coalesced_mmio && !s->coalesced_mmio_ring) {
365         s->coalesced_mmio_ring =
366             (void *)cpu->kvm_run + s->coalesced_mmio * PAGE_SIZE;
367     }
368 
369     ret = kvm_arch_init_vcpu(cpu);
370 err:
371     return ret;
372 }
373 
374 /*
375  * dirty pages logging control
376  */
377 
378 static int kvm_mem_flags(MemoryRegion *mr)
379 {
380     bool readonly = mr->readonly || memory_region_is_romd(mr);
381     int flags = 0;
382 
383     if (memory_region_get_dirty_log_mask(mr) != 0) {
384         flags |= KVM_MEM_LOG_DIRTY_PAGES;
385     }
386     if (readonly && kvm_readonly_mem_allowed) {
387         flags |= KVM_MEM_READONLY;
388     }
389     return flags;
390 }
391 
392 static int kvm_slot_update_flags(KVMMemoryListener *kml, KVMSlot *mem,
393                                  MemoryRegion *mr)
394 {
395     mem->flags = kvm_mem_flags(mr);
396 
397     /* If nothing changed effectively, no need to issue ioctl */
398     if (mem->flags == mem->old_flags) {
399         return 0;
400     }
401 
402     return kvm_set_user_memory_region(kml, mem, false);
403 }
404 
405 static int kvm_section_update_flags(KVMMemoryListener *kml,
406                                     MemoryRegionSection *section)
407 {
408     hwaddr start_addr, size;
409     KVMSlot *mem;
410 
411     size = kvm_align_section(section, &start_addr);
412     if (!size) {
413         return 0;
414     }
415 
416     mem = kvm_lookup_matching_slot(kml, start_addr, size);
417     if (!mem) {
418         /* We don't have a slot if we want to trap every access. */
419         return 0;
420     }
421 
422     return kvm_slot_update_flags(kml, mem, section->mr);
423 }
424 
425 static void kvm_log_start(MemoryListener *listener,
426                           MemoryRegionSection *section,
427                           int old, int new)
428 {
429     KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
430     int r;
431 
432     if (old != 0) {
433         return;
434     }
435 
436     r = kvm_section_update_flags(kml, section);
437     if (r < 0) {
438         abort();
439     }
440 }
441 
442 static void kvm_log_stop(MemoryListener *listener,
443                           MemoryRegionSection *section,
444                           int old, int new)
445 {
446     KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
447     int r;
448 
449     if (new != 0) {
450         return;
451     }
452 
453     r = kvm_section_update_flags(kml, section);
454     if (r < 0) {
455         abort();
456     }
457 }
458 
459 /* get kvm's dirty pages bitmap and update qemu's */
460 static int kvm_get_dirty_pages_log_range(MemoryRegionSection *section,
461                                          unsigned long *bitmap)
462 {
463     ram_addr_t start = section->offset_within_region +
464                        memory_region_get_ram_addr(section->mr);
465     ram_addr_t pages = int128_get64(section->size) / getpagesize();
466 
467     cpu_physical_memory_set_dirty_lebitmap(bitmap, start, pages);
468     return 0;
469 }
470 
471 #define ALIGN(x, y)  (((x)+(y)-1) & ~((y)-1))
472 
473 /**
474  * kvm_physical_sync_dirty_bitmap - Grab dirty bitmap from kernel space
475  * This function updates qemu's dirty bitmap using
476  * memory_region_set_dirty().  This means all bits are set
477  * to dirty.
478  *
479  * @start_add: start of logged region.
480  * @end_addr: end of logged region.
481  */
482 static int kvm_physical_sync_dirty_bitmap(KVMMemoryListener *kml,
483                                           MemoryRegionSection *section)
484 {
485     KVMState *s = kvm_state;
486     struct kvm_dirty_log d = {};
487     KVMSlot *mem;
488     hwaddr start_addr, size;
489 
490     size = kvm_align_section(section, &start_addr);
491     if (size) {
492         mem = kvm_lookup_matching_slot(kml, start_addr, size);
493         if (!mem) {
494             /* We don't have a slot if we want to trap every access. */
495             return 0;
496         }
497 
498         /* XXX bad kernel interface alert
499          * For dirty bitmap, kernel allocates array of size aligned to
500          * bits-per-long.  But for case when the kernel is 64bits and
501          * the userspace is 32bits, userspace can't align to the same
502          * bits-per-long, since sizeof(long) is different between kernel
503          * and user space.  This way, userspace will provide buffer which
504          * may be 4 bytes less than the kernel will use, resulting in
505          * userspace memory corruption (which is not detectable by valgrind
506          * too, in most cases).
507          * So for now, let's align to 64 instead of HOST_LONG_BITS here, in
508          * a hope that sizeof(long) won't become >8 any time soon.
509          */
510         size = ALIGN(((mem->memory_size) >> TARGET_PAGE_BITS),
511                      /*HOST_LONG_BITS*/ 64) / 8;
512         d.dirty_bitmap = g_malloc0(size);
513 
514         d.slot = mem->slot | (kml->as_id << 16);
515         if (kvm_vm_ioctl(s, KVM_GET_DIRTY_LOG, &d) == -1) {
516             DPRINTF("ioctl failed %d\n", errno);
517             g_free(d.dirty_bitmap);
518             return -1;
519         }
520 
521         kvm_get_dirty_pages_log_range(section, d.dirty_bitmap);
522         g_free(d.dirty_bitmap);
523     }
524 
525     return 0;
526 }
527 
528 static void kvm_coalesce_mmio_region(MemoryListener *listener,
529                                      MemoryRegionSection *secion,
530                                      hwaddr start, hwaddr size)
531 {
532     KVMState *s = kvm_state;
533 
534     if (s->coalesced_mmio) {
535         struct kvm_coalesced_mmio_zone zone;
536 
537         zone.addr = start;
538         zone.size = size;
539         zone.pad = 0;
540 
541         (void)kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone);
542     }
543 }
544 
545 static void kvm_uncoalesce_mmio_region(MemoryListener *listener,
546                                        MemoryRegionSection *secion,
547                                        hwaddr start, hwaddr size)
548 {
549     KVMState *s = kvm_state;
550 
551     if (s->coalesced_mmio) {
552         struct kvm_coalesced_mmio_zone zone;
553 
554         zone.addr = start;
555         zone.size = size;
556         zone.pad = 0;
557 
558         (void)kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone);
559     }
560 }
561 
562 int kvm_check_extension(KVMState *s, unsigned int extension)
563 {
564     int ret;
565 
566     ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, extension);
567     if (ret < 0) {
568         ret = 0;
569     }
570 
571     return ret;
572 }
573 
574 int kvm_vm_check_extension(KVMState *s, unsigned int extension)
575 {
576     int ret;
577 
578     ret = kvm_vm_ioctl(s, KVM_CHECK_EXTENSION, extension);
579     if (ret < 0) {
580         /* VM wide version not implemented, use global one instead */
581         ret = kvm_check_extension(s, extension);
582     }
583 
584     return ret;
585 }
586 
587 static uint32_t adjust_ioeventfd_endianness(uint32_t val, uint32_t size)
588 {
589 #if defined(HOST_WORDS_BIGENDIAN) != defined(TARGET_WORDS_BIGENDIAN)
590     /* The kernel expects ioeventfd values in HOST_WORDS_BIGENDIAN
591      * endianness, but the memory core hands them in target endianness.
592      * For example, PPC is always treated as big-endian even if running
593      * on KVM and on PPC64LE.  Correct here.
594      */
595     switch (size) {
596     case 2:
597         val = bswap16(val);
598         break;
599     case 4:
600         val = bswap32(val);
601         break;
602     }
603 #endif
604     return val;
605 }
606 
607 static int kvm_set_ioeventfd_mmio(int fd, hwaddr addr, uint32_t val,
608                                   bool assign, uint32_t size, bool datamatch)
609 {
610     int ret;
611     struct kvm_ioeventfd iofd = {
612         .datamatch = datamatch ? adjust_ioeventfd_endianness(val, size) : 0,
613         .addr = addr,
614         .len = size,
615         .flags = 0,
616         .fd = fd,
617     };
618 
619     if (!kvm_enabled()) {
620         return -ENOSYS;
621     }
622 
623     if (datamatch) {
624         iofd.flags |= KVM_IOEVENTFD_FLAG_DATAMATCH;
625     }
626     if (!assign) {
627         iofd.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
628     }
629 
630     ret = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &iofd);
631 
632     if (ret < 0) {
633         return -errno;
634     }
635 
636     return 0;
637 }
638 
639 static int kvm_set_ioeventfd_pio(int fd, uint16_t addr, uint16_t val,
640                                  bool assign, uint32_t size, bool datamatch)
641 {
642     struct kvm_ioeventfd kick = {
643         .datamatch = datamatch ? adjust_ioeventfd_endianness(val, size) : 0,
644         .addr = addr,
645         .flags = KVM_IOEVENTFD_FLAG_PIO,
646         .len = size,
647         .fd = fd,
648     };
649     int r;
650     if (!kvm_enabled()) {
651         return -ENOSYS;
652     }
653     if (datamatch) {
654         kick.flags |= KVM_IOEVENTFD_FLAG_DATAMATCH;
655     }
656     if (!assign) {
657         kick.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
658     }
659     r = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &kick);
660     if (r < 0) {
661         return r;
662     }
663     return 0;
664 }
665 
666 
667 static int kvm_check_many_ioeventfds(void)
668 {
669     /* Userspace can use ioeventfd for io notification.  This requires a host
670      * that supports eventfd(2) and an I/O thread; since eventfd does not
671      * support SIGIO it cannot interrupt the vcpu.
672      *
673      * Older kernels have a 6 device limit on the KVM io bus.  Find out so we
674      * can avoid creating too many ioeventfds.
675      */
676 #if defined(CONFIG_EVENTFD)
677     int ioeventfds[7];
678     int i, ret = 0;
679     for (i = 0; i < ARRAY_SIZE(ioeventfds); i++) {
680         ioeventfds[i] = eventfd(0, EFD_CLOEXEC);
681         if (ioeventfds[i] < 0) {
682             break;
683         }
684         ret = kvm_set_ioeventfd_pio(ioeventfds[i], 0, i, true, 2, true);
685         if (ret < 0) {
686             close(ioeventfds[i]);
687             break;
688         }
689     }
690 
691     /* Decide whether many devices are supported or not */
692     ret = i == ARRAY_SIZE(ioeventfds);
693 
694     while (i-- > 0) {
695         kvm_set_ioeventfd_pio(ioeventfds[i], 0, i, false, 2, true);
696         close(ioeventfds[i]);
697     }
698     return ret;
699 #else
700     return 0;
701 #endif
702 }
703 
704 static const KVMCapabilityInfo *
705 kvm_check_extension_list(KVMState *s, const KVMCapabilityInfo *list)
706 {
707     while (list->name) {
708         if (!kvm_check_extension(s, list->value)) {
709             return list;
710         }
711         list++;
712     }
713     return NULL;
714 }
715 
716 static void kvm_set_phys_mem(KVMMemoryListener *kml,
717                              MemoryRegionSection *section, bool add)
718 {
719     KVMSlot *mem;
720     int err;
721     MemoryRegion *mr = section->mr;
722     bool writeable = !mr->readonly && !mr->rom_device;
723     hwaddr start_addr, size;
724     void *ram;
725 
726     if (!memory_region_is_ram(mr)) {
727         if (writeable || !kvm_readonly_mem_allowed) {
728             return;
729         } else if (!mr->romd_mode) {
730             /* If the memory device is not in romd_mode, then we actually want
731              * to remove the kvm memory slot so all accesses will trap. */
732             add = false;
733         }
734     }
735 
736     size = kvm_align_section(section, &start_addr);
737     if (!size) {
738         return;
739     }
740 
741     /* use aligned delta to align the ram address */
742     ram = memory_region_get_ram_ptr(mr) + section->offset_within_region +
743           (start_addr - section->offset_within_address_space);
744 
745     if (!add) {
746         mem = kvm_lookup_matching_slot(kml, start_addr, size);
747         if (!mem) {
748             return;
749         }
750         if (mem->flags & KVM_MEM_LOG_DIRTY_PAGES) {
751             kvm_physical_sync_dirty_bitmap(kml, section);
752         }
753 
754         /* unregister the slot */
755         mem->memory_size = 0;
756         mem->flags = 0;
757         err = kvm_set_user_memory_region(kml, mem, false);
758         if (err) {
759             fprintf(stderr, "%s: error unregistering slot: %s\n",
760                     __func__, strerror(-err));
761             abort();
762         }
763         return;
764     }
765 
766     /* register the new slot */
767     mem = kvm_alloc_slot(kml);
768     mem->memory_size = size;
769     mem->start_addr = start_addr;
770     mem->ram = ram;
771     mem->flags = kvm_mem_flags(mr);
772 
773     err = kvm_set_user_memory_region(kml, mem, true);
774     if (err) {
775         fprintf(stderr, "%s: error registering slot: %s\n", __func__,
776                 strerror(-err));
777         abort();
778     }
779 }
780 
781 static void kvm_region_add(MemoryListener *listener,
782                            MemoryRegionSection *section)
783 {
784     KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
785 
786     memory_region_ref(section->mr);
787     kvm_set_phys_mem(kml, section, true);
788 }
789 
790 static void kvm_region_del(MemoryListener *listener,
791                            MemoryRegionSection *section)
792 {
793     KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
794 
795     kvm_set_phys_mem(kml, section, false);
796     memory_region_unref(section->mr);
797 }
798 
799 static void kvm_log_sync(MemoryListener *listener,
800                          MemoryRegionSection *section)
801 {
802     KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
803     int r;
804 
805     r = kvm_physical_sync_dirty_bitmap(kml, section);
806     if (r < 0) {
807         abort();
808     }
809 }
810 
811 static void kvm_mem_ioeventfd_add(MemoryListener *listener,
812                                   MemoryRegionSection *section,
813                                   bool match_data, uint64_t data,
814                                   EventNotifier *e)
815 {
816     int fd = event_notifier_get_fd(e);
817     int r;
818 
819     r = kvm_set_ioeventfd_mmio(fd, section->offset_within_address_space,
820                                data, true, int128_get64(section->size),
821                                match_data);
822     if (r < 0) {
823         fprintf(stderr, "%s: error adding ioeventfd: %s\n",
824                 __func__, strerror(-r));
825         abort();
826     }
827 }
828 
829 static void kvm_mem_ioeventfd_del(MemoryListener *listener,
830                                   MemoryRegionSection *section,
831                                   bool match_data, uint64_t data,
832                                   EventNotifier *e)
833 {
834     int fd = event_notifier_get_fd(e);
835     int r;
836 
837     r = kvm_set_ioeventfd_mmio(fd, section->offset_within_address_space,
838                                data, false, int128_get64(section->size),
839                                match_data);
840     if (r < 0) {
841         abort();
842     }
843 }
844 
845 static void kvm_io_ioeventfd_add(MemoryListener *listener,
846                                  MemoryRegionSection *section,
847                                  bool match_data, uint64_t data,
848                                  EventNotifier *e)
849 {
850     int fd = event_notifier_get_fd(e);
851     int r;
852 
853     r = kvm_set_ioeventfd_pio(fd, section->offset_within_address_space,
854                               data, true, int128_get64(section->size),
855                               match_data);
856     if (r < 0) {
857         fprintf(stderr, "%s: error adding ioeventfd: %s\n",
858                 __func__, strerror(-r));
859         abort();
860     }
861 }
862 
863 static void kvm_io_ioeventfd_del(MemoryListener *listener,
864                                  MemoryRegionSection *section,
865                                  bool match_data, uint64_t data,
866                                  EventNotifier *e)
867 
868 {
869     int fd = event_notifier_get_fd(e);
870     int r;
871 
872     r = kvm_set_ioeventfd_pio(fd, section->offset_within_address_space,
873                               data, false, int128_get64(section->size),
874                               match_data);
875     if (r < 0) {
876         abort();
877     }
878 }
879 
880 void kvm_memory_listener_register(KVMState *s, KVMMemoryListener *kml,
881                                   AddressSpace *as, int as_id)
882 {
883     int i;
884 
885     kml->slots = g_malloc0(s->nr_slots * sizeof(KVMSlot));
886     kml->as_id = as_id;
887 
888     for (i = 0; i < s->nr_slots; i++) {
889         kml->slots[i].slot = i;
890     }
891 
892     kml->listener.region_add = kvm_region_add;
893     kml->listener.region_del = kvm_region_del;
894     kml->listener.log_start = kvm_log_start;
895     kml->listener.log_stop = kvm_log_stop;
896     kml->listener.log_sync = kvm_log_sync;
897     kml->listener.priority = 10;
898 
899     memory_listener_register(&kml->listener, as);
900 }
901 
902 static MemoryListener kvm_io_listener = {
903     .eventfd_add = kvm_io_ioeventfd_add,
904     .eventfd_del = kvm_io_ioeventfd_del,
905     .priority = 10,
906 };
907 
908 int kvm_set_irq(KVMState *s, int irq, int level)
909 {
910     struct kvm_irq_level event;
911     int ret;
912 
913     assert(kvm_async_interrupts_enabled());
914 
915     event.level = level;
916     event.irq = irq;
917     ret = kvm_vm_ioctl(s, s->irq_set_ioctl, &event);
918     if (ret < 0) {
919         perror("kvm_set_irq");
920         abort();
921     }
922 
923     return (s->irq_set_ioctl == KVM_IRQ_LINE) ? 1 : event.status;
924 }
925 
926 #ifdef KVM_CAP_IRQ_ROUTING
927 typedef struct KVMMSIRoute {
928     struct kvm_irq_routing_entry kroute;
929     QTAILQ_ENTRY(KVMMSIRoute) entry;
930 } KVMMSIRoute;
931 
932 static void set_gsi(KVMState *s, unsigned int gsi)
933 {
934     set_bit(gsi, s->used_gsi_bitmap);
935 }
936 
937 static void clear_gsi(KVMState *s, unsigned int gsi)
938 {
939     clear_bit(gsi, s->used_gsi_bitmap);
940 }
941 
942 void kvm_init_irq_routing(KVMState *s)
943 {
944     int gsi_count, i;
945 
946     gsi_count = kvm_check_extension(s, KVM_CAP_IRQ_ROUTING) - 1;
947     if (gsi_count > 0) {
948         /* Round up so we can search ints using ffs */
949         s->used_gsi_bitmap = bitmap_new(gsi_count);
950         s->gsi_count = gsi_count;
951     }
952 
953     s->irq_routes = g_malloc0(sizeof(*s->irq_routes));
954     s->nr_allocated_irq_routes = 0;
955 
956     if (!kvm_direct_msi_allowed) {
957         for (i = 0; i < KVM_MSI_HASHTAB_SIZE; i++) {
958             QTAILQ_INIT(&s->msi_hashtab[i]);
959         }
960     }
961 
962     kvm_arch_init_irq_routing(s);
963 }
964 
965 void kvm_irqchip_commit_routes(KVMState *s)
966 {
967     int ret;
968 
969     if (kvm_gsi_direct_mapping()) {
970         return;
971     }
972 
973     if (!kvm_gsi_routing_enabled()) {
974         return;
975     }
976 
977     s->irq_routes->flags = 0;
978     trace_kvm_irqchip_commit_routes();
979     ret = kvm_vm_ioctl(s, KVM_SET_GSI_ROUTING, s->irq_routes);
980     assert(ret == 0);
981 }
982 
983 static void kvm_add_routing_entry(KVMState *s,
984                                   struct kvm_irq_routing_entry *entry)
985 {
986     struct kvm_irq_routing_entry *new;
987     int n, size;
988 
989     if (s->irq_routes->nr == s->nr_allocated_irq_routes) {
990         n = s->nr_allocated_irq_routes * 2;
991         if (n < 64) {
992             n = 64;
993         }
994         size = sizeof(struct kvm_irq_routing);
995         size += n * sizeof(*new);
996         s->irq_routes = g_realloc(s->irq_routes, size);
997         s->nr_allocated_irq_routes = n;
998     }
999     n = s->irq_routes->nr++;
1000     new = &s->irq_routes->entries[n];
1001 
1002     *new = *entry;
1003 
1004     set_gsi(s, entry->gsi);
1005 }
1006 
1007 static int kvm_update_routing_entry(KVMState *s,
1008                                     struct kvm_irq_routing_entry *new_entry)
1009 {
1010     struct kvm_irq_routing_entry *entry;
1011     int n;
1012 
1013     for (n = 0; n < s->irq_routes->nr; n++) {
1014         entry = &s->irq_routes->entries[n];
1015         if (entry->gsi != new_entry->gsi) {
1016             continue;
1017         }
1018 
1019         if(!memcmp(entry, new_entry, sizeof *entry)) {
1020             return 0;
1021         }
1022 
1023         *entry = *new_entry;
1024 
1025         return 0;
1026     }
1027 
1028     return -ESRCH;
1029 }
1030 
1031 void kvm_irqchip_add_irq_route(KVMState *s, int irq, int irqchip, int pin)
1032 {
1033     struct kvm_irq_routing_entry e = {};
1034 
1035     assert(pin < s->gsi_count);
1036 
1037     e.gsi = irq;
1038     e.type = KVM_IRQ_ROUTING_IRQCHIP;
1039     e.flags = 0;
1040     e.u.irqchip.irqchip = irqchip;
1041     e.u.irqchip.pin = pin;
1042     kvm_add_routing_entry(s, &e);
1043 }
1044 
1045 void kvm_irqchip_release_virq(KVMState *s, int virq)
1046 {
1047     struct kvm_irq_routing_entry *e;
1048     int i;
1049 
1050     if (kvm_gsi_direct_mapping()) {
1051         return;
1052     }
1053 
1054     for (i = 0; i < s->irq_routes->nr; i++) {
1055         e = &s->irq_routes->entries[i];
1056         if (e->gsi == virq) {
1057             s->irq_routes->nr--;
1058             *e = s->irq_routes->entries[s->irq_routes->nr];
1059         }
1060     }
1061     clear_gsi(s, virq);
1062     kvm_arch_release_virq_post(virq);
1063     trace_kvm_irqchip_release_virq(virq);
1064 }
1065 
1066 static unsigned int kvm_hash_msi(uint32_t data)
1067 {
1068     /* This is optimized for IA32 MSI layout. However, no other arch shall
1069      * repeat the mistake of not providing a direct MSI injection API. */
1070     return data & 0xff;
1071 }
1072 
1073 static void kvm_flush_dynamic_msi_routes(KVMState *s)
1074 {
1075     KVMMSIRoute *route, *next;
1076     unsigned int hash;
1077 
1078     for (hash = 0; hash < KVM_MSI_HASHTAB_SIZE; hash++) {
1079         QTAILQ_FOREACH_SAFE(route, &s->msi_hashtab[hash], entry, next) {
1080             kvm_irqchip_release_virq(s, route->kroute.gsi);
1081             QTAILQ_REMOVE(&s->msi_hashtab[hash], route, entry);
1082             g_free(route);
1083         }
1084     }
1085 }
1086 
1087 static int kvm_irqchip_get_virq(KVMState *s)
1088 {
1089     int next_virq;
1090 
1091     /*
1092      * PIC and IOAPIC share the first 16 GSI numbers, thus the available
1093      * GSI numbers are more than the number of IRQ route. Allocating a GSI
1094      * number can succeed even though a new route entry cannot be added.
1095      * When this happens, flush dynamic MSI entries to free IRQ route entries.
1096      */
1097     if (!kvm_direct_msi_allowed && s->irq_routes->nr == s->gsi_count) {
1098         kvm_flush_dynamic_msi_routes(s);
1099     }
1100 
1101     /* Return the lowest unused GSI in the bitmap */
1102     next_virq = find_first_zero_bit(s->used_gsi_bitmap, s->gsi_count);
1103     if (next_virq >= s->gsi_count) {
1104         return -ENOSPC;
1105     } else {
1106         return next_virq;
1107     }
1108 }
1109 
1110 static KVMMSIRoute *kvm_lookup_msi_route(KVMState *s, MSIMessage msg)
1111 {
1112     unsigned int hash = kvm_hash_msi(msg.data);
1113     KVMMSIRoute *route;
1114 
1115     QTAILQ_FOREACH(route, &s->msi_hashtab[hash], entry) {
1116         if (route->kroute.u.msi.address_lo == (uint32_t)msg.address &&
1117             route->kroute.u.msi.address_hi == (msg.address >> 32) &&
1118             route->kroute.u.msi.data == le32_to_cpu(msg.data)) {
1119             return route;
1120         }
1121     }
1122     return NULL;
1123 }
1124 
1125 int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg)
1126 {
1127     struct kvm_msi msi;
1128     KVMMSIRoute *route;
1129 
1130     if (kvm_direct_msi_allowed) {
1131         msi.address_lo = (uint32_t)msg.address;
1132         msi.address_hi = msg.address >> 32;
1133         msi.data = le32_to_cpu(msg.data);
1134         msi.flags = 0;
1135         memset(msi.pad, 0, sizeof(msi.pad));
1136 
1137         return kvm_vm_ioctl(s, KVM_SIGNAL_MSI, &msi);
1138     }
1139 
1140     route = kvm_lookup_msi_route(s, msg);
1141     if (!route) {
1142         int virq;
1143 
1144         virq = kvm_irqchip_get_virq(s);
1145         if (virq < 0) {
1146             return virq;
1147         }
1148 
1149         route = g_malloc0(sizeof(KVMMSIRoute));
1150         route->kroute.gsi = virq;
1151         route->kroute.type = KVM_IRQ_ROUTING_MSI;
1152         route->kroute.flags = 0;
1153         route->kroute.u.msi.address_lo = (uint32_t)msg.address;
1154         route->kroute.u.msi.address_hi = msg.address >> 32;
1155         route->kroute.u.msi.data = le32_to_cpu(msg.data);
1156 
1157         kvm_add_routing_entry(s, &route->kroute);
1158         kvm_irqchip_commit_routes(s);
1159 
1160         QTAILQ_INSERT_TAIL(&s->msi_hashtab[kvm_hash_msi(msg.data)], route,
1161                            entry);
1162     }
1163 
1164     assert(route->kroute.type == KVM_IRQ_ROUTING_MSI);
1165 
1166     return kvm_set_irq(s, route->kroute.gsi, 1);
1167 }
1168 
1169 int kvm_irqchip_add_msi_route(KVMState *s, int vector, PCIDevice *dev)
1170 {
1171     struct kvm_irq_routing_entry kroute = {};
1172     int virq;
1173     MSIMessage msg = {0, 0};
1174 
1175     if (pci_available && dev) {
1176         msg = pci_get_msi_message(dev, vector);
1177     }
1178 
1179     if (kvm_gsi_direct_mapping()) {
1180         return kvm_arch_msi_data_to_gsi(msg.data);
1181     }
1182 
1183     if (!kvm_gsi_routing_enabled()) {
1184         return -ENOSYS;
1185     }
1186 
1187     virq = kvm_irqchip_get_virq(s);
1188     if (virq < 0) {
1189         return virq;
1190     }
1191 
1192     kroute.gsi = virq;
1193     kroute.type = KVM_IRQ_ROUTING_MSI;
1194     kroute.flags = 0;
1195     kroute.u.msi.address_lo = (uint32_t)msg.address;
1196     kroute.u.msi.address_hi = msg.address >> 32;
1197     kroute.u.msi.data = le32_to_cpu(msg.data);
1198     if (pci_available && kvm_msi_devid_required()) {
1199         kroute.flags = KVM_MSI_VALID_DEVID;
1200         kroute.u.msi.devid = pci_requester_id(dev);
1201     }
1202     if (kvm_arch_fixup_msi_route(&kroute, msg.address, msg.data, dev)) {
1203         kvm_irqchip_release_virq(s, virq);
1204         return -EINVAL;
1205     }
1206 
1207     trace_kvm_irqchip_add_msi_route(dev ? dev->name : (char *)"N/A",
1208                                     vector, virq);
1209 
1210     kvm_add_routing_entry(s, &kroute);
1211     kvm_arch_add_msi_route_post(&kroute, vector, dev);
1212     kvm_irqchip_commit_routes(s);
1213 
1214     return virq;
1215 }
1216 
1217 int kvm_irqchip_update_msi_route(KVMState *s, int virq, MSIMessage msg,
1218                                  PCIDevice *dev)
1219 {
1220     struct kvm_irq_routing_entry kroute = {};
1221 
1222     if (kvm_gsi_direct_mapping()) {
1223         return 0;
1224     }
1225 
1226     if (!kvm_irqchip_in_kernel()) {
1227         return -ENOSYS;
1228     }
1229 
1230     kroute.gsi = virq;
1231     kroute.type = KVM_IRQ_ROUTING_MSI;
1232     kroute.flags = 0;
1233     kroute.u.msi.address_lo = (uint32_t)msg.address;
1234     kroute.u.msi.address_hi = msg.address >> 32;
1235     kroute.u.msi.data = le32_to_cpu(msg.data);
1236     if (pci_available && kvm_msi_devid_required()) {
1237         kroute.flags = KVM_MSI_VALID_DEVID;
1238         kroute.u.msi.devid = pci_requester_id(dev);
1239     }
1240     if (kvm_arch_fixup_msi_route(&kroute, msg.address, msg.data, dev)) {
1241         return -EINVAL;
1242     }
1243 
1244     trace_kvm_irqchip_update_msi_route(virq);
1245 
1246     return kvm_update_routing_entry(s, &kroute);
1247 }
1248 
1249 static int kvm_irqchip_assign_irqfd(KVMState *s, int fd, int rfd, int virq,
1250                                     bool assign)
1251 {
1252     struct kvm_irqfd irqfd = {
1253         .fd = fd,
1254         .gsi = virq,
1255         .flags = assign ? 0 : KVM_IRQFD_FLAG_DEASSIGN,
1256     };
1257 
1258     if (rfd != -1) {
1259         irqfd.flags |= KVM_IRQFD_FLAG_RESAMPLE;
1260         irqfd.resamplefd = rfd;
1261     }
1262 
1263     if (!kvm_irqfds_enabled()) {
1264         return -ENOSYS;
1265     }
1266 
1267     return kvm_vm_ioctl(s, KVM_IRQFD, &irqfd);
1268 }
1269 
1270 int kvm_irqchip_add_adapter_route(KVMState *s, AdapterInfo *adapter)
1271 {
1272     struct kvm_irq_routing_entry kroute = {};
1273     int virq;
1274 
1275     if (!kvm_gsi_routing_enabled()) {
1276         return -ENOSYS;
1277     }
1278 
1279     virq = kvm_irqchip_get_virq(s);
1280     if (virq < 0) {
1281         return virq;
1282     }
1283 
1284     kroute.gsi = virq;
1285     kroute.type = KVM_IRQ_ROUTING_S390_ADAPTER;
1286     kroute.flags = 0;
1287     kroute.u.adapter.summary_addr = adapter->summary_addr;
1288     kroute.u.adapter.ind_addr = adapter->ind_addr;
1289     kroute.u.adapter.summary_offset = adapter->summary_offset;
1290     kroute.u.adapter.ind_offset = adapter->ind_offset;
1291     kroute.u.adapter.adapter_id = adapter->adapter_id;
1292 
1293     kvm_add_routing_entry(s, &kroute);
1294 
1295     return virq;
1296 }
1297 
1298 int kvm_irqchip_add_hv_sint_route(KVMState *s, uint32_t vcpu, uint32_t sint)
1299 {
1300     struct kvm_irq_routing_entry kroute = {};
1301     int virq;
1302 
1303     if (!kvm_gsi_routing_enabled()) {
1304         return -ENOSYS;
1305     }
1306     if (!kvm_check_extension(s, KVM_CAP_HYPERV_SYNIC)) {
1307         return -ENOSYS;
1308     }
1309     virq = kvm_irqchip_get_virq(s);
1310     if (virq < 0) {
1311         return virq;
1312     }
1313 
1314     kroute.gsi = virq;
1315     kroute.type = KVM_IRQ_ROUTING_HV_SINT;
1316     kroute.flags = 0;
1317     kroute.u.hv_sint.vcpu = vcpu;
1318     kroute.u.hv_sint.sint = sint;
1319 
1320     kvm_add_routing_entry(s, &kroute);
1321     kvm_irqchip_commit_routes(s);
1322 
1323     return virq;
1324 }
1325 
1326 #else /* !KVM_CAP_IRQ_ROUTING */
1327 
1328 void kvm_init_irq_routing(KVMState *s)
1329 {
1330 }
1331 
1332 void kvm_irqchip_release_virq(KVMState *s, int virq)
1333 {
1334 }
1335 
1336 int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg)
1337 {
1338     abort();
1339 }
1340 
1341 int kvm_irqchip_add_msi_route(KVMState *s, int vector, PCIDevice *dev)
1342 {
1343     return -ENOSYS;
1344 }
1345 
1346 int kvm_irqchip_add_adapter_route(KVMState *s, AdapterInfo *adapter)
1347 {
1348     return -ENOSYS;
1349 }
1350 
1351 int kvm_irqchip_add_hv_sint_route(KVMState *s, uint32_t vcpu, uint32_t sint)
1352 {
1353     return -ENOSYS;
1354 }
1355 
1356 static int kvm_irqchip_assign_irqfd(KVMState *s, int fd, int virq, bool assign)
1357 {
1358     abort();
1359 }
1360 
1361 int kvm_irqchip_update_msi_route(KVMState *s, int virq, MSIMessage msg)
1362 {
1363     return -ENOSYS;
1364 }
1365 #endif /* !KVM_CAP_IRQ_ROUTING */
1366 
1367 int kvm_irqchip_add_irqfd_notifier_gsi(KVMState *s, EventNotifier *n,
1368                                        EventNotifier *rn, int virq)
1369 {
1370     return kvm_irqchip_assign_irqfd(s, event_notifier_get_fd(n),
1371            rn ? event_notifier_get_fd(rn) : -1, virq, true);
1372 }
1373 
1374 int kvm_irqchip_remove_irqfd_notifier_gsi(KVMState *s, EventNotifier *n,
1375                                           int virq)
1376 {
1377     return kvm_irqchip_assign_irqfd(s, event_notifier_get_fd(n), -1, virq,
1378            false);
1379 }
1380 
1381 int kvm_irqchip_add_irqfd_notifier(KVMState *s, EventNotifier *n,
1382                                    EventNotifier *rn, qemu_irq irq)
1383 {
1384     gpointer key, gsi;
1385     gboolean found = g_hash_table_lookup_extended(s->gsimap, irq, &key, &gsi);
1386 
1387     if (!found) {
1388         return -ENXIO;
1389     }
1390     return kvm_irqchip_add_irqfd_notifier_gsi(s, n, rn, GPOINTER_TO_INT(gsi));
1391 }
1392 
1393 int kvm_irqchip_remove_irqfd_notifier(KVMState *s, EventNotifier *n,
1394                                       qemu_irq irq)
1395 {
1396     gpointer key, gsi;
1397     gboolean found = g_hash_table_lookup_extended(s->gsimap, irq, &key, &gsi);
1398 
1399     if (!found) {
1400         return -ENXIO;
1401     }
1402     return kvm_irqchip_remove_irqfd_notifier_gsi(s, n, GPOINTER_TO_INT(gsi));
1403 }
1404 
1405 void kvm_irqchip_set_qemuirq_gsi(KVMState *s, qemu_irq irq, int gsi)
1406 {
1407     g_hash_table_insert(s->gsimap, irq, GINT_TO_POINTER(gsi));
1408 }
1409 
1410 static void kvm_irqchip_create(MachineState *machine, KVMState *s)
1411 {
1412     int ret;
1413 
1414     if (kvm_check_extension(s, KVM_CAP_IRQCHIP)) {
1415         ;
1416     } else if (kvm_check_extension(s, KVM_CAP_S390_IRQCHIP)) {
1417         ret = kvm_vm_enable_cap(s, KVM_CAP_S390_IRQCHIP, 0);
1418         if (ret < 0) {
1419             fprintf(stderr, "Enable kernel irqchip failed: %s\n", strerror(-ret));
1420             exit(1);
1421         }
1422     } else {
1423         return;
1424     }
1425 
1426     /* First probe and see if there's a arch-specific hook to create the
1427      * in-kernel irqchip for us */
1428     ret = kvm_arch_irqchip_create(machine, s);
1429     if (ret == 0) {
1430         if (machine_kernel_irqchip_split(machine)) {
1431             perror("Split IRQ chip mode not supported.");
1432             exit(1);
1433         } else {
1434             ret = kvm_vm_ioctl(s, KVM_CREATE_IRQCHIP);
1435         }
1436     }
1437     if (ret < 0) {
1438         fprintf(stderr, "Create kernel irqchip failed: %s\n", strerror(-ret));
1439         exit(1);
1440     }
1441 
1442     kvm_kernel_irqchip = true;
1443     /* If we have an in-kernel IRQ chip then we must have asynchronous
1444      * interrupt delivery (though the reverse is not necessarily true)
1445      */
1446     kvm_async_interrupts_allowed = true;
1447     kvm_halt_in_kernel_allowed = true;
1448 
1449     kvm_init_irq_routing(s);
1450 
1451     s->gsimap = g_hash_table_new(g_direct_hash, g_direct_equal);
1452 }
1453 
1454 /* Find number of supported CPUs using the recommended
1455  * procedure from the kernel API documentation to cope with
1456  * older kernels that may be missing capabilities.
1457  */
1458 static int kvm_recommended_vcpus(KVMState *s)
1459 {
1460     int ret = kvm_vm_check_extension(s, KVM_CAP_NR_VCPUS);
1461     return (ret) ? ret : 4;
1462 }
1463 
1464 static int kvm_max_vcpus(KVMState *s)
1465 {
1466     int ret = kvm_check_extension(s, KVM_CAP_MAX_VCPUS);
1467     return (ret) ? ret : kvm_recommended_vcpus(s);
1468 }
1469 
1470 static int kvm_max_vcpu_id(KVMState *s)
1471 {
1472     int ret = kvm_check_extension(s, KVM_CAP_MAX_VCPU_ID);
1473     return (ret) ? ret : kvm_max_vcpus(s);
1474 }
1475 
1476 bool kvm_vcpu_id_is_valid(int vcpu_id)
1477 {
1478     KVMState *s = KVM_STATE(current_machine->accelerator);
1479     return vcpu_id >= 0 && vcpu_id < kvm_max_vcpu_id(s);
1480 }
1481 
1482 static int kvm_init(MachineState *ms)
1483 {
1484     MachineClass *mc = MACHINE_GET_CLASS(ms);
1485     static const char upgrade_note[] =
1486         "Please upgrade to at least kernel 2.6.29 or recent kvm-kmod\n"
1487         "(see http://sourceforge.net/projects/kvm).\n";
1488     struct {
1489         const char *name;
1490         int num;
1491     } num_cpus[] = {
1492         { "SMP",          smp_cpus },
1493         { "hotpluggable", max_cpus },
1494         { NULL, }
1495     }, *nc = num_cpus;
1496     int soft_vcpus_limit, hard_vcpus_limit;
1497     KVMState *s;
1498     const KVMCapabilityInfo *missing_cap;
1499     int ret;
1500     int type = 0;
1501     const char *kvm_type;
1502 
1503     s = KVM_STATE(ms->accelerator);
1504 
1505     /*
1506      * On systems where the kernel can support different base page
1507      * sizes, host page size may be different from TARGET_PAGE_SIZE,
1508      * even with KVM.  TARGET_PAGE_SIZE is assumed to be the minimum
1509      * page size for the system though.
1510      */
1511     assert(TARGET_PAGE_SIZE <= getpagesize());
1512 
1513     s->sigmask_len = 8;
1514 
1515 #ifdef KVM_CAP_SET_GUEST_DEBUG
1516     QTAILQ_INIT(&s->kvm_sw_breakpoints);
1517 #endif
1518     QLIST_INIT(&s->kvm_parked_vcpus);
1519     s->vmfd = -1;
1520     s->fd = qemu_open("/dev/kvm", O_RDWR);
1521     if (s->fd == -1) {
1522         fprintf(stderr, "Could not access KVM kernel module: %m\n");
1523         ret = -errno;
1524         goto err;
1525     }
1526 
1527     ret = kvm_ioctl(s, KVM_GET_API_VERSION, 0);
1528     if (ret < KVM_API_VERSION) {
1529         if (ret >= 0) {
1530             ret = -EINVAL;
1531         }
1532         fprintf(stderr, "kvm version too old\n");
1533         goto err;
1534     }
1535 
1536     if (ret > KVM_API_VERSION) {
1537         ret = -EINVAL;
1538         fprintf(stderr, "kvm version not supported\n");
1539         goto err;
1540     }
1541 
1542     kvm_immediate_exit = kvm_check_extension(s, KVM_CAP_IMMEDIATE_EXIT);
1543     s->nr_slots = kvm_check_extension(s, KVM_CAP_NR_MEMSLOTS);
1544 
1545     /* If unspecified, use the default value */
1546     if (!s->nr_slots) {
1547         s->nr_slots = 32;
1548     }
1549 
1550     kvm_type = qemu_opt_get(qemu_get_machine_opts(), "kvm-type");
1551     if (mc->kvm_type) {
1552         type = mc->kvm_type(kvm_type);
1553     } else if (kvm_type) {
1554         ret = -EINVAL;
1555         fprintf(stderr, "Invalid argument kvm-type=%s\n", kvm_type);
1556         goto err;
1557     }
1558 
1559     do {
1560         ret = kvm_ioctl(s, KVM_CREATE_VM, type);
1561     } while (ret == -EINTR);
1562 
1563     if (ret < 0) {
1564         fprintf(stderr, "ioctl(KVM_CREATE_VM) failed: %d %s\n", -ret,
1565                 strerror(-ret));
1566 
1567 #ifdef TARGET_S390X
1568         if (ret == -EINVAL) {
1569             fprintf(stderr,
1570                     "Host kernel setup problem detected. Please verify:\n");
1571             fprintf(stderr, "- for kernels supporting the switch_amode or"
1572                     " user_mode parameters, whether\n");
1573             fprintf(stderr,
1574                     "  user space is running in primary address space\n");
1575             fprintf(stderr,
1576                     "- for kernels supporting the vm.allocate_pgste sysctl, "
1577                     "whether it is enabled\n");
1578         }
1579 #endif
1580         goto err;
1581     }
1582 
1583     s->vmfd = ret;
1584 
1585     /* check the vcpu limits */
1586     soft_vcpus_limit = kvm_recommended_vcpus(s);
1587     hard_vcpus_limit = kvm_max_vcpus(s);
1588 
1589     while (nc->name) {
1590         if (nc->num > soft_vcpus_limit) {
1591             warn_report("Number of %s cpus requested (%d) exceeds "
1592                         "the recommended cpus supported by KVM (%d)",
1593                         nc->name, nc->num, soft_vcpus_limit);
1594 
1595             if (nc->num > hard_vcpus_limit) {
1596                 fprintf(stderr, "Number of %s cpus requested (%d) exceeds "
1597                         "the maximum cpus supported by KVM (%d)\n",
1598                         nc->name, nc->num, hard_vcpus_limit);
1599                 exit(1);
1600             }
1601         }
1602         nc++;
1603     }
1604 
1605     missing_cap = kvm_check_extension_list(s, kvm_required_capabilites);
1606     if (!missing_cap) {
1607         missing_cap =
1608             kvm_check_extension_list(s, kvm_arch_required_capabilities);
1609     }
1610     if (missing_cap) {
1611         ret = -EINVAL;
1612         fprintf(stderr, "kvm does not support %s\n%s",
1613                 missing_cap->name, upgrade_note);
1614         goto err;
1615     }
1616 
1617     s->coalesced_mmio = kvm_check_extension(s, KVM_CAP_COALESCED_MMIO);
1618 
1619 #ifdef KVM_CAP_VCPU_EVENTS
1620     s->vcpu_events = kvm_check_extension(s, KVM_CAP_VCPU_EVENTS);
1621 #endif
1622 
1623     s->robust_singlestep =
1624         kvm_check_extension(s, KVM_CAP_X86_ROBUST_SINGLESTEP);
1625 
1626 #ifdef KVM_CAP_DEBUGREGS
1627     s->debugregs = kvm_check_extension(s, KVM_CAP_DEBUGREGS);
1628 #endif
1629 
1630 #ifdef KVM_CAP_IRQ_ROUTING
1631     kvm_direct_msi_allowed = (kvm_check_extension(s, KVM_CAP_SIGNAL_MSI) > 0);
1632 #endif
1633 
1634     s->intx_set_mask = kvm_check_extension(s, KVM_CAP_PCI_2_3);
1635 
1636     s->irq_set_ioctl = KVM_IRQ_LINE;
1637     if (kvm_check_extension(s, KVM_CAP_IRQ_INJECT_STATUS)) {
1638         s->irq_set_ioctl = KVM_IRQ_LINE_STATUS;
1639     }
1640 
1641 #ifdef KVM_CAP_READONLY_MEM
1642     kvm_readonly_mem_allowed =
1643         (kvm_check_extension(s, KVM_CAP_READONLY_MEM) > 0);
1644 #endif
1645 
1646     kvm_eventfds_allowed =
1647         (kvm_check_extension(s, KVM_CAP_IOEVENTFD) > 0);
1648 
1649     kvm_irqfds_allowed =
1650         (kvm_check_extension(s, KVM_CAP_IRQFD) > 0);
1651 
1652     kvm_resamplefds_allowed =
1653         (kvm_check_extension(s, KVM_CAP_IRQFD_RESAMPLE) > 0);
1654 
1655     kvm_vm_attributes_allowed =
1656         (kvm_check_extension(s, KVM_CAP_VM_ATTRIBUTES) > 0);
1657 
1658     kvm_ioeventfd_any_length_allowed =
1659         (kvm_check_extension(s, KVM_CAP_IOEVENTFD_ANY_LENGTH) > 0);
1660 
1661     kvm_state = s;
1662 
1663     /*
1664      * if memory encryption object is specified then initialize the memory
1665      * encryption context.
1666      */
1667     if (ms->memory_encryption) {
1668         kvm_state->memcrypt_handle = sev_guest_init(ms->memory_encryption);
1669         if (!kvm_state->memcrypt_handle) {
1670             ret = -1;
1671             goto err;
1672         }
1673 
1674         kvm_state->memcrypt_encrypt_data = sev_encrypt_data;
1675     }
1676 
1677     ret = kvm_arch_init(ms, s);
1678     if (ret < 0) {
1679         goto err;
1680     }
1681 
1682     if (machine_kernel_irqchip_allowed(ms)) {
1683         kvm_irqchip_create(ms, s);
1684     }
1685 
1686     if (kvm_eventfds_allowed) {
1687         s->memory_listener.listener.eventfd_add = kvm_mem_ioeventfd_add;
1688         s->memory_listener.listener.eventfd_del = kvm_mem_ioeventfd_del;
1689     }
1690     s->memory_listener.listener.coalesced_mmio_add = kvm_coalesce_mmio_region;
1691     s->memory_listener.listener.coalesced_mmio_del = kvm_uncoalesce_mmio_region;
1692 
1693     kvm_memory_listener_register(s, &s->memory_listener,
1694                                  &address_space_memory, 0);
1695     memory_listener_register(&kvm_io_listener,
1696                              &address_space_io);
1697 
1698     s->many_ioeventfds = kvm_check_many_ioeventfds();
1699 
1700     s->sync_mmu = !!kvm_vm_check_extension(kvm_state, KVM_CAP_SYNC_MMU);
1701 
1702     return 0;
1703 
1704 err:
1705     assert(ret < 0);
1706     if (s->vmfd >= 0) {
1707         close(s->vmfd);
1708     }
1709     if (s->fd != -1) {
1710         close(s->fd);
1711     }
1712     g_free(s->memory_listener.slots);
1713 
1714     return ret;
1715 }
1716 
1717 void kvm_set_sigmask_len(KVMState *s, unsigned int sigmask_len)
1718 {
1719     s->sigmask_len = sigmask_len;
1720 }
1721 
1722 static void kvm_handle_io(uint16_t port, MemTxAttrs attrs, void *data, int direction,
1723                           int size, uint32_t count)
1724 {
1725     int i;
1726     uint8_t *ptr = data;
1727 
1728     for (i = 0; i < count; i++) {
1729         address_space_rw(&address_space_io, port, attrs,
1730                          ptr, size,
1731                          direction == KVM_EXIT_IO_OUT);
1732         ptr += size;
1733     }
1734 }
1735 
1736 static int kvm_handle_internal_error(CPUState *cpu, struct kvm_run *run)
1737 {
1738     fprintf(stderr, "KVM internal error. Suberror: %d\n",
1739             run->internal.suberror);
1740 
1741     if (kvm_check_extension(kvm_state, KVM_CAP_INTERNAL_ERROR_DATA)) {
1742         int i;
1743 
1744         for (i = 0; i < run->internal.ndata; ++i) {
1745             fprintf(stderr, "extra data[%d]: %"PRIx64"\n",
1746                     i, (uint64_t)run->internal.data[i]);
1747         }
1748     }
1749     if (run->internal.suberror == KVM_INTERNAL_ERROR_EMULATION) {
1750         fprintf(stderr, "emulation failure\n");
1751         if (!kvm_arch_stop_on_emulation_error(cpu)) {
1752             cpu_dump_state(cpu, stderr, fprintf, CPU_DUMP_CODE);
1753             return EXCP_INTERRUPT;
1754         }
1755     }
1756     /* FIXME: Should trigger a qmp message to let management know
1757      * something went wrong.
1758      */
1759     return -1;
1760 }
1761 
1762 void kvm_flush_coalesced_mmio_buffer(void)
1763 {
1764     KVMState *s = kvm_state;
1765 
1766     if (s->coalesced_flush_in_progress) {
1767         return;
1768     }
1769 
1770     s->coalesced_flush_in_progress = true;
1771 
1772     if (s->coalesced_mmio_ring) {
1773         struct kvm_coalesced_mmio_ring *ring = s->coalesced_mmio_ring;
1774         while (ring->first != ring->last) {
1775             struct kvm_coalesced_mmio *ent;
1776 
1777             ent = &ring->coalesced_mmio[ring->first];
1778 
1779             cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len);
1780             smp_wmb();
1781             ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX;
1782         }
1783     }
1784 
1785     s->coalesced_flush_in_progress = false;
1786 }
1787 
1788 static void do_kvm_cpu_synchronize_state(CPUState *cpu, run_on_cpu_data arg)
1789 {
1790     if (!cpu->vcpu_dirty) {
1791         kvm_arch_get_registers(cpu);
1792         cpu->vcpu_dirty = true;
1793     }
1794 }
1795 
1796 void kvm_cpu_synchronize_state(CPUState *cpu)
1797 {
1798     if (!cpu->vcpu_dirty) {
1799         run_on_cpu(cpu, do_kvm_cpu_synchronize_state, RUN_ON_CPU_NULL);
1800     }
1801 }
1802 
1803 static void do_kvm_cpu_synchronize_post_reset(CPUState *cpu, run_on_cpu_data arg)
1804 {
1805     kvm_arch_put_registers(cpu, KVM_PUT_RESET_STATE);
1806     cpu->vcpu_dirty = false;
1807 }
1808 
1809 void kvm_cpu_synchronize_post_reset(CPUState *cpu)
1810 {
1811     run_on_cpu(cpu, do_kvm_cpu_synchronize_post_reset, RUN_ON_CPU_NULL);
1812 }
1813 
1814 static void do_kvm_cpu_synchronize_post_init(CPUState *cpu, run_on_cpu_data arg)
1815 {
1816     kvm_arch_put_registers(cpu, KVM_PUT_FULL_STATE);
1817     cpu->vcpu_dirty = false;
1818 }
1819 
1820 void kvm_cpu_synchronize_post_init(CPUState *cpu)
1821 {
1822     run_on_cpu(cpu, do_kvm_cpu_synchronize_post_init, RUN_ON_CPU_NULL);
1823 }
1824 
1825 static void do_kvm_cpu_synchronize_pre_loadvm(CPUState *cpu, run_on_cpu_data arg)
1826 {
1827     cpu->vcpu_dirty = true;
1828 }
1829 
1830 void kvm_cpu_synchronize_pre_loadvm(CPUState *cpu)
1831 {
1832     run_on_cpu(cpu, do_kvm_cpu_synchronize_pre_loadvm, RUN_ON_CPU_NULL);
1833 }
1834 
1835 #ifdef KVM_HAVE_MCE_INJECTION
1836 static __thread void *pending_sigbus_addr;
1837 static __thread int pending_sigbus_code;
1838 static __thread bool have_sigbus_pending;
1839 #endif
1840 
1841 static void kvm_cpu_kick(CPUState *cpu)
1842 {
1843     atomic_set(&cpu->kvm_run->immediate_exit, 1);
1844 }
1845 
1846 static void kvm_cpu_kick_self(void)
1847 {
1848     if (kvm_immediate_exit) {
1849         kvm_cpu_kick(current_cpu);
1850     } else {
1851         qemu_cpu_kick_self();
1852     }
1853 }
1854 
1855 static void kvm_eat_signals(CPUState *cpu)
1856 {
1857     struct timespec ts = { 0, 0 };
1858     siginfo_t siginfo;
1859     sigset_t waitset;
1860     sigset_t chkset;
1861     int r;
1862 
1863     if (kvm_immediate_exit) {
1864         atomic_set(&cpu->kvm_run->immediate_exit, 0);
1865         /* Write kvm_run->immediate_exit before the cpu->exit_request
1866          * write in kvm_cpu_exec.
1867          */
1868         smp_wmb();
1869         return;
1870     }
1871 
1872     sigemptyset(&waitset);
1873     sigaddset(&waitset, SIG_IPI);
1874 
1875     do {
1876         r = sigtimedwait(&waitset, &siginfo, &ts);
1877         if (r == -1 && !(errno == EAGAIN || errno == EINTR)) {
1878             perror("sigtimedwait");
1879             exit(1);
1880         }
1881 
1882         r = sigpending(&chkset);
1883         if (r == -1) {
1884             perror("sigpending");
1885             exit(1);
1886         }
1887     } while (sigismember(&chkset, SIG_IPI));
1888 }
1889 
1890 int kvm_cpu_exec(CPUState *cpu)
1891 {
1892     struct kvm_run *run = cpu->kvm_run;
1893     int ret, run_ret;
1894 
1895     DPRINTF("kvm_cpu_exec()\n");
1896 
1897     if (kvm_arch_process_async_events(cpu)) {
1898         atomic_set(&cpu->exit_request, 0);
1899         return EXCP_HLT;
1900     }
1901 
1902     qemu_mutex_unlock_iothread();
1903     cpu_exec_start(cpu);
1904 
1905     do {
1906         MemTxAttrs attrs;
1907 
1908         if (cpu->vcpu_dirty) {
1909             kvm_arch_put_registers(cpu, KVM_PUT_RUNTIME_STATE);
1910             cpu->vcpu_dirty = false;
1911         }
1912 
1913         kvm_arch_pre_run(cpu, run);
1914         if (atomic_read(&cpu->exit_request)) {
1915             DPRINTF("interrupt exit requested\n");
1916             /*
1917              * KVM requires us to reenter the kernel after IO exits to complete
1918              * instruction emulation. This self-signal will ensure that we
1919              * leave ASAP again.
1920              */
1921             kvm_cpu_kick_self();
1922         }
1923 
1924         /* Read cpu->exit_request before KVM_RUN reads run->immediate_exit.
1925          * Matching barrier in kvm_eat_signals.
1926          */
1927         smp_rmb();
1928 
1929         run_ret = kvm_vcpu_ioctl(cpu, KVM_RUN, 0);
1930 
1931         attrs = kvm_arch_post_run(cpu, run);
1932 
1933 #ifdef KVM_HAVE_MCE_INJECTION
1934         if (unlikely(have_sigbus_pending)) {
1935             qemu_mutex_lock_iothread();
1936             kvm_arch_on_sigbus_vcpu(cpu, pending_sigbus_code,
1937                                     pending_sigbus_addr);
1938             have_sigbus_pending = false;
1939             qemu_mutex_unlock_iothread();
1940         }
1941 #endif
1942 
1943         if (run_ret < 0) {
1944             if (run_ret == -EINTR || run_ret == -EAGAIN) {
1945                 DPRINTF("io window exit\n");
1946                 kvm_eat_signals(cpu);
1947                 ret = EXCP_INTERRUPT;
1948                 break;
1949             }
1950             fprintf(stderr, "error: kvm run failed %s\n",
1951                     strerror(-run_ret));
1952 #ifdef TARGET_PPC
1953             if (run_ret == -EBUSY) {
1954                 fprintf(stderr,
1955                         "This is probably because your SMT is enabled.\n"
1956                         "VCPU can only run on primary threads with all "
1957                         "secondary threads offline.\n");
1958             }
1959 #endif
1960             ret = -1;
1961             break;
1962         }
1963 
1964         trace_kvm_run_exit(cpu->cpu_index, run->exit_reason);
1965         switch (run->exit_reason) {
1966         case KVM_EXIT_IO:
1967             DPRINTF("handle_io\n");
1968             /* Called outside BQL */
1969             kvm_handle_io(run->io.port, attrs,
1970                           (uint8_t *)run + run->io.data_offset,
1971                           run->io.direction,
1972                           run->io.size,
1973                           run->io.count);
1974             ret = 0;
1975             break;
1976         case KVM_EXIT_MMIO:
1977             DPRINTF("handle_mmio\n");
1978             /* Called outside BQL */
1979             address_space_rw(&address_space_memory,
1980                              run->mmio.phys_addr, attrs,
1981                              run->mmio.data,
1982                              run->mmio.len,
1983                              run->mmio.is_write);
1984             ret = 0;
1985             break;
1986         case KVM_EXIT_IRQ_WINDOW_OPEN:
1987             DPRINTF("irq_window_open\n");
1988             ret = EXCP_INTERRUPT;
1989             break;
1990         case KVM_EXIT_SHUTDOWN:
1991             DPRINTF("shutdown\n");
1992             qemu_system_reset_request(SHUTDOWN_CAUSE_GUEST_RESET);
1993             ret = EXCP_INTERRUPT;
1994             break;
1995         case KVM_EXIT_UNKNOWN:
1996             fprintf(stderr, "KVM: unknown exit, hardware reason %" PRIx64 "\n",
1997                     (uint64_t)run->hw.hardware_exit_reason);
1998             ret = -1;
1999             break;
2000         case KVM_EXIT_INTERNAL_ERROR:
2001             ret = kvm_handle_internal_error(cpu, run);
2002             break;
2003         case KVM_EXIT_SYSTEM_EVENT:
2004             switch (run->system_event.type) {
2005             case KVM_SYSTEM_EVENT_SHUTDOWN:
2006                 qemu_system_shutdown_request(SHUTDOWN_CAUSE_GUEST_SHUTDOWN);
2007                 ret = EXCP_INTERRUPT;
2008                 break;
2009             case KVM_SYSTEM_EVENT_RESET:
2010                 qemu_system_reset_request(SHUTDOWN_CAUSE_GUEST_RESET);
2011                 ret = EXCP_INTERRUPT;
2012                 break;
2013             case KVM_SYSTEM_EVENT_CRASH:
2014                 kvm_cpu_synchronize_state(cpu);
2015                 qemu_mutex_lock_iothread();
2016                 qemu_system_guest_panicked(cpu_get_crash_info(cpu));
2017                 qemu_mutex_unlock_iothread();
2018                 ret = 0;
2019                 break;
2020             default:
2021                 DPRINTF("kvm_arch_handle_exit\n");
2022                 ret = kvm_arch_handle_exit(cpu, run);
2023                 break;
2024             }
2025             break;
2026         default:
2027             DPRINTF("kvm_arch_handle_exit\n");
2028             ret = kvm_arch_handle_exit(cpu, run);
2029             break;
2030         }
2031     } while (ret == 0);
2032 
2033     cpu_exec_end(cpu);
2034     qemu_mutex_lock_iothread();
2035 
2036     if (ret < 0) {
2037         cpu_dump_state(cpu, stderr, fprintf, CPU_DUMP_CODE);
2038         vm_stop(RUN_STATE_INTERNAL_ERROR);
2039     }
2040 
2041     atomic_set(&cpu->exit_request, 0);
2042     return ret;
2043 }
2044 
2045 int kvm_ioctl(KVMState *s, int type, ...)
2046 {
2047     int ret;
2048     void *arg;
2049     va_list ap;
2050 
2051     va_start(ap, type);
2052     arg = va_arg(ap, void *);
2053     va_end(ap);
2054 
2055     trace_kvm_ioctl(type, arg);
2056     ret = ioctl(s->fd, type, arg);
2057     if (ret == -1) {
2058         ret = -errno;
2059     }
2060     return ret;
2061 }
2062 
2063 int kvm_vm_ioctl(KVMState *s, int type, ...)
2064 {
2065     int ret;
2066     void *arg;
2067     va_list ap;
2068 
2069     va_start(ap, type);
2070     arg = va_arg(ap, void *);
2071     va_end(ap);
2072 
2073     trace_kvm_vm_ioctl(type, arg);
2074     ret = ioctl(s->vmfd, type, arg);
2075     if (ret == -1) {
2076         ret = -errno;
2077     }
2078     return ret;
2079 }
2080 
2081 int kvm_vcpu_ioctl(CPUState *cpu, int type, ...)
2082 {
2083     int ret;
2084     void *arg;
2085     va_list ap;
2086 
2087     va_start(ap, type);
2088     arg = va_arg(ap, void *);
2089     va_end(ap);
2090 
2091     trace_kvm_vcpu_ioctl(cpu->cpu_index, type, arg);
2092     ret = ioctl(cpu->kvm_fd, type, arg);
2093     if (ret == -1) {
2094         ret = -errno;
2095     }
2096     return ret;
2097 }
2098 
2099 int kvm_device_ioctl(int fd, int type, ...)
2100 {
2101     int ret;
2102     void *arg;
2103     va_list ap;
2104 
2105     va_start(ap, type);
2106     arg = va_arg(ap, void *);
2107     va_end(ap);
2108 
2109     trace_kvm_device_ioctl(fd, type, arg);
2110     ret = ioctl(fd, type, arg);
2111     if (ret == -1) {
2112         ret = -errno;
2113     }
2114     return ret;
2115 }
2116 
2117 int kvm_vm_check_attr(KVMState *s, uint32_t group, uint64_t attr)
2118 {
2119     int ret;
2120     struct kvm_device_attr attribute = {
2121         .group = group,
2122         .attr = attr,
2123     };
2124 
2125     if (!kvm_vm_attributes_allowed) {
2126         return 0;
2127     }
2128 
2129     ret = kvm_vm_ioctl(s, KVM_HAS_DEVICE_ATTR, &attribute);
2130     /* kvm returns 0 on success for HAS_DEVICE_ATTR */
2131     return ret ? 0 : 1;
2132 }
2133 
2134 int kvm_device_check_attr(int dev_fd, uint32_t group, uint64_t attr)
2135 {
2136     struct kvm_device_attr attribute = {
2137         .group = group,
2138         .attr = attr,
2139         .flags = 0,
2140     };
2141 
2142     return kvm_device_ioctl(dev_fd, KVM_HAS_DEVICE_ATTR, &attribute) ? 0 : 1;
2143 }
2144 
2145 int kvm_device_access(int fd, int group, uint64_t attr,
2146                       void *val, bool write, Error **errp)
2147 {
2148     struct kvm_device_attr kvmattr;
2149     int err;
2150 
2151     kvmattr.flags = 0;
2152     kvmattr.group = group;
2153     kvmattr.attr = attr;
2154     kvmattr.addr = (uintptr_t)val;
2155 
2156     err = kvm_device_ioctl(fd,
2157                            write ? KVM_SET_DEVICE_ATTR : KVM_GET_DEVICE_ATTR,
2158                            &kvmattr);
2159     if (err < 0) {
2160         error_setg_errno(errp, -err,
2161                          "KVM_%s_DEVICE_ATTR failed: Group %d "
2162                          "attr 0x%016" PRIx64,
2163                          write ? "SET" : "GET", group, attr);
2164     }
2165     return err;
2166 }
2167 
2168 bool kvm_has_sync_mmu(void)
2169 {
2170     return kvm_state->sync_mmu;
2171 }
2172 
2173 int kvm_has_vcpu_events(void)
2174 {
2175     return kvm_state->vcpu_events;
2176 }
2177 
2178 int kvm_has_robust_singlestep(void)
2179 {
2180     return kvm_state->robust_singlestep;
2181 }
2182 
2183 int kvm_has_debugregs(void)
2184 {
2185     return kvm_state->debugregs;
2186 }
2187 
2188 int kvm_has_many_ioeventfds(void)
2189 {
2190     if (!kvm_enabled()) {
2191         return 0;
2192     }
2193     return kvm_state->many_ioeventfds;
2194 }
2195 
2196 int kvm_has_gsi_routing(void)
2197 {
2198 #ifdef KVM_CAP_IRQ_ROUTING
2199     return kvm_check_extension(kvm_state, KVM_CAP_IRQ_ROUTING);
2200 #else
2201     return false;
2202 #endif
2203 }
2204 
2205 int kvm_has_intx_set_mask(void)
2206 {
2207     return kvm_state->intx_set_mask;
2208 }
2209 
2210 bool kvm_arm_supports_user_irq(void)
2211 {
2212     return kvm_check_extension(kvm_state, KVM_CAP_ARM_USER_IRQ);
2213 }
2214 
2215 #ifdef KVM_CAP_SET_GUEST_DEBUG
2216 struct kvm_sw_breakpoint *kvm_find_sw_breakpoint(CPUState *cpu,
2217                                                  target_ulong pc)
2218 {
2219     struct kvm_sw_breakpoint *bp;
2220 
2221     QTAILQ_FOREACH(bp, &cpu->kvm_state->kvm_sw_breakpoints, entry) {
2222         if (bp->pc == pc) {
2223             return bp;
2224         }
2225     }
2226     return NULL;
2227 }
2228 
2229 int kvm_sw_breakpoints_active(CPUState *cpu)
2230 {
2231     return !QTAILQ_EMPTY(&cpu->kvm_state->kvm_sw_breakpoints);
2232 }
2233 
2234 struct kvm_set_guest_debug_data {
2235     struct kvm_guest_debug dbg;
2236     int err;
2237 };
2238 
2239 static void kvm_invoke_set_guest_debug(CPUState *cpu, run_on_cpu_data data)
2240 {
2241     struct kvm_set_guest_debug_data *dbg_data =
2242         (struct kvm_set_guest_debug_data *) data.host_ptr;
2243 
2244     dbg_data->err = kvm_vcpu_ioctl(cpu, KVM_SET_GUEST_DEBUG,
2245                                    &dbg_data->dbg);
2246 }
2247 
2248 int kvm_update_guest_debug(CPUState *cpu, unsigned long reinject_trap)
2249 {
2250     struct kvm_set_guest_debug_data data;
2251 
2252     data.dbg.control = reinject_trap;
2253 
2254     if (cpu->singlestep_enabled) {
2255         data.dbg.control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_SINGLESTEP;
2256     }
2257     kvm_arch_update_guest_debug(cpu, &data.dbg);
2258 
2259     run_on_cpu(cpu, kvm_invoke_set_guest_debug,
2260                RUN_ON_CPU_HOST_PTR(&data));
2261     return data.err;
2262 }
2263 
2264 int kvm_insert_breakpoint(CPUState *cpu, target_ulong addr,
2265                           target_ulong len, int type)
2266 {
2267     struct kvm_sw_breakpoint *bp;
2268     int err;
2269 
2270     if (type == GDB_BREAKPOINT_SW) {
2271         bp = kvm_find_sw_breakpoint(cpu, addr);
2272         if (bp) {
2273             bp->use_count++;
2274             return 0;
2275         }
2276 
2277         bp = g_malloc(sizeof(struct kvm_sw_breakpoint));
2278         bp->pc = addr;
2279         bp->use_count = 1;
2280         err = kvm_arch_insert_sw_breakpoint(cpu, bp);
2281         if (err) {
2282             g_free(bp);
2283             return err;
2284         }
2285 
2286         QTAILQ_INSERT_HEAD(&cpu->kvm_state->kvm_sw_breakpoints, bp, entry);
2287     } else {
2288         err = kvm_arch_insert_hw_breakpoint(addr, len, type);
2289         if (err) {
2290             return err;
2291         }
2292     }
2293 
2294     CPU_FOREACH(cpu) {
2295         err = kvm_update_guest_debug(cpu, 0);
2296         if (err) {
2297             return err;
2298         }
2299     }
2300     return 0;
2301 }
2302 
2303 int kvm_remove_breakpoint(CPUState *cpu, target_ulong addr,
2304                           target_ulong len, int type)
2305 {
2306     struct kvm_sw_breakpoint *bp;
2307     int err;
2308 
2309     if (type == GDB_BREAKPOINT_SW) {
2310         bp = kvm_find_sw_breakpoint(cpu, addr);
2311         if (!bp) {
2312             return -ENOENT;
2313         }
2314 
2315         if (bp->use_count > 1) {
2316             bp->use_count--;
2317             return 0;
2318         }
2319 
2320         err = kvm_arch_remove_sw_breakpoint(cpu, bp);
2321         if (err) {
2322             return err;
2323         }
2324 
2325         QTAILQ_REMOVE(&cpu->kvm_state->kvm_sw_breakpoints, bp, entry);
2326         g_free(bp);
2327     } else {
2328         err = kvm_arch_remove_hw_breakpoint(addr, len, type);
2329         if (err) {
2330             return err;
2331         }
2332     }
2333 
2334     CPU_FOREACH(cpu) {
2335         err = kvm_update_guest_debug(cpu, 0);
2336         if (err) {
2337             return err;
2338         }
2339     }
2340     return 0;
2341 }
2342 
2343 void kvm_remove_all_breakpoints(CPUState *cpu)
2344 {
2345     struct kvm_sw_breakpoint *bp, *next;
2346     KVMState *s = cpu->kvm_state;
2347     CPUState *tmpcpu;
2348 
2349     QTAILQ_FOREACH_SAFE(bp, &s->kvm_sw_breakpoints, entry, next) {
2350         if (kvm_arch_remove_sw_breakpoint(cpu, bp) != 0) {
2351             /* Try harder to find a CPU that currently sees the breakpoint. */
2352             CPU_FOREACH(tmpcpu) {
2353                 if (kvm_arch_remove_sw_breakpoint(tmpcpu, bp) == 0) {
2354                     break;
2355                 }
2356             }
2357         }
2358         QTAILQ_REMOVE(&s->kvm_sw_breakpoints, bp, entry);
2359         g_free(bp);
2360     }
2361     kvm_arch_remove_all_hw_breakpoints();
2362 
2363     CPU_FOREACH(cpu) {
2364         kvm_update_guest_debug(cpu, 0);
2365     }
2366 }
2367 
2368 #else /* !KVM_CAP_SET_GUEST_DEBUG */
2369 
2370 int kvm_update_guest_debug(CPUState *cpu, unsigned long reinject_trap)
2371 {
2372     return -EINVAL;
2373 }
2374 
2375 int kvm_insert_breakpoint(CPUState *cpu, target_ulong addr,
2376                           target_ulong len, int type)
2377 {
2378     return -EINVAL;
2379 }
2380 
2381 int kvm_remove_breakpoint(CPUState *cpu, target_ulong addr,
2382                           target_ulong len, int type)
2383 {
2384     return -EINVAL;
2385 }
2386 
2387 void kvm_remove_all_breakpoints(CPUState *cpu)
2388 {
2389 }
2390 #endif /* !KVM_CAP_SET_GUEST_DEBUG */
2391 
2392 static int kvm_set_signal_mask(CPUState *cpu, const sigset_t *sigset)
2393 {
2394     KVMState *s = kvm_state;
2395     struct kvm_signal_mask *sigmask;
2396     int r;
2397 
2398     sigmask = g_malloc(sizeof(*sigmask) + sizeof(*sigset));
2399 
2400     sigmask->len = s->sigmask_len;
2401     memcpy(sigmask->sigset, sigset, sizeof(*sigset));
2402     r = kvm_vcpu_ioctl(cpu, KVM_SET_SIGNAL_MASK, sigmask);
2403     g_free(sigmask);
2404 
2405     return r;
2406 }
2407 
2408 static void kvm_ipi_signal(int sig)
2409 {
2410     if (current_cpu) {
2411         assert(kvm_immediate_exit);
2412         kvm_cpu_kick(current_cpu);
2413     }
2414 }
2415 
2416 void kvm_init_cpu_signals(CPUState *cpu)
2417 {
2418     int r;
2419     sigset_t set;
2420     struct sigaction sigact;
2421 
2422     memset(&sigact, 0, sizeof(sigact));
2423     sigact.sa_handler = kvm_ipi_signal;
2424     sigaction(SIG_IPI, &sigact, NULL);
2425 
2426     pthread_sigmask(SIG_BLOCK, NULL, &set);
2427 #if defined KVM_HAVE_MCE_INJECTION
2428     sigdelset(&set, SIGBUS);
2429     pthread_sigmask(SIG_SETMASK, &set, NULL);
2430 #endif
2431     sigdelset(&set, SIG_IPI);
2432     if (kvm_immediate_exit) {
2433         r = pthread_sigmask(SIG_SETMASK, &set, NULL);
2434     } else {
2435         r = kvm_set_signal_mask(cpu, &set);
2436     }
2437     if (r) {
2438         fprintf(stderr, "kvm_set_signal_mask: %s\n", strerror(-r));
2439         exit(1);
2440     }
2441 }
2442 
2443 /* Called asynchronously in VCPU thread.  */
2444 int kvm_on_sigbus_vcpu(CPUState *cpu, int code, void *addr)
2445 {
2446 #ifdef KVM_HAVE_MCE_INJECTION
2447     if (have_sigbus_pending) {
2448         return 1;
2449     }
2450     have_sigbus_pending = true;
2451     pending_sigbus_addr = addr;
2452     pending_sigbus_code = code;
2453     atomic_set(&cpu->exit_request, 1);
2454     return 0;
2455 #else
2456     return 1;
2457 #endif
2458 }
2459 
2460 /* Called synchronously (via signalfd) in main thread.  */
2461 int kvm_on_sigbus(int code, void *addr)
2462 {
2463 #ifdef KVM_HAVE_MCE_INJECTION
2464     /* Action required MCE kills the process if SIGBUS is blocked.  Because
2465      * that's what happens in the I/O thread, where we handle MCE via signalfd,
2466      * we can only get action optional here.
2467      */
2468     assert(code != BUS_MCEERR_AR);
2469     kvm_arch_on_sigbus_vcpu(first_cpu, code, addr);
2470     return 0;
2471 #else
2472     return 1;
2473 #endif
2474 }
2475 
2476 int kvm_create_device(KVMState *s, uint64_t type, bool test)
2477 {
2478     int ret;
2479     struct kvm_create_device create_dev;
2480 
2481     create_dev.type = type;
2482     create_dev.fd = -1;
2483     create_dev.flags = test ? KVM_CREATE_DEVICE_TEST : 0;
2484 
2485     if (!kvm_check_extension(s, KVM_CAP_DEVICE_CTRL)) {
2486         return -ENOTSUP;
2487     }
2488 
2489     ret = kvm_vm_ioctl(s, KVM_CREATE_DEVICE, &create_dev);
2490     if (ret) {
2491         return ret;
2492     }
2493 
2494     return test ? 0 : create_dev.fd;
2495 }
2496 
2497 bool kvm_device_supported(int vmfd, uint64_t type)
2498 {
2499     struct kvm_create_device create_dev = {
2500         .type = type,
2501         .fd = -1,
2502         .flags = KVM_CREATE_DEVICE_TEST,
2503     };
2504 
2505     if (ioctl(vmfd, KVM_CHECK_EXTENSION, KVM_CAP_DEVICE_CTRL) <= 0) {
2506         return false;
2507     }
2508 
2509     return (ioctl(vmfd, KVM_CREATE_DEVICE, &create_dev) >= 0);
2510 }
2511 
2512 int kvm_set_one_reg(CPUState *cs, uint64_t id, void *source)
2513 {
2514     struct kvm_one_reg reg;
2515     int r;
2516 
2517     reg.id = id;
2518     reg.addr = (uintptr_t) source;
2519     r = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, &reg);
2520     if (r) {
2521         trace_kvm_failed_reg_set(id, strerror(-r));
2522     }
2523     return r;
2524 }
2525 
2526 int kvm_get_one_reg(CPUState *cs, uint64_t id, void *target)
2527 {
2528     struct kvm_one_reg reg;
2529     int r;
2530 
2531     reg.id = id;
2532     reg.addr = (uintptr_t) target;
2533     r = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &reg);
2534     if (r) {
2535         trace_kvm_failed_reg_get(id, strerror(-r));
2536     }
2537     return r;
2538 }
2539 
2540 static void kvm_accel_class_init(ObjectClass *oc, void *data)
2541 {
2542     AccelClass *ac = ACCEL_CLASS(oc);
2543     ac->name = "KVM";
2544     ac->init_machine = kvm_init;
2545     ac->allowed = &kvm_allowed;
2546 }
2547 
2548 static const TypeInfo kvm_accel_type = {
2549     .name = TYPE_KVM_ACCEL,
2550     .parent = TYPE_ACCEL,
2551     .class_init = kvm_accel_class_init,
2552     .instance_size = sizeof(KVMState),
2553 };
2554 
2555 static void kvm_type_init(void)
2556 {
2557     type_register_static(&kvm_accel_type);
2558 }
2559 
2560 type_init(kvm_type_init);
2561