xref: /openbmc/qemu/accel/kvm/kvm-all.c (revision 1ef8185a)
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 
239     mem.slot = slot->slot | (kml->as_id << 16);
240     mem.guest_phys_addr = slot->start_addr;
241     mem.userspace_addr = (unsigned long)slot->ram;
242     mem.flags = slot->flags;
243 
244     if (slot->memory_size && mem.flags & KVM_MEM_READONLY) {
245         /* Set the slot size to 0 before setting the slot to the desired
246          * value. This is needed based on KVM commit 75d61fbc. */
247         mem.memory_size = 0;
248         kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
249     }
250     mem.memory_size = slot->memory_size;
251     return kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
252 }
253 
254 int kvm_destroy_vcpu(CPUState *cpu)
255 {
256     KVMState *s = kvm_state;
257     long mmap_size;
258     struct KVMParkedVcpu *vcpu = NULL;
259     int ret = 0;
260 
261     DPRINTF("kvm_destroy_vcpu\n");
262 
263     mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
264     if (mmap_size < 0) {
265         ret = mmap_size;
266         DPRINTF("KVM_GET_VCPU_MMAP_SIZE failed\n");
267         goto err;
268     }
269 
270     ret = munmap(cpu->kvm_run, mmap_size);
271     if (ret < 0) {
272         goto err;
273     }
274 
275     vcpu = g_malloc0(sizeof(*vcpu));
276     vcpu->vcpu_id = kvm_arch_vcpu_id(cpu);
277     vcpu->kvm_fd = cpu->kvm_fd;
278     QLIST_INSERT_HEAD(&kvm_state->kvm_parked_vcpus, vcpu, node);
279 err:
280     return ret;
281 }
282 
283 static int kvm_get_vcpu(KVMState *s, unsigned long vcpu_id)
284 {
285     struct KVMParkedVcpu *cpu;
286 
287     QLIST_FOREACH(cpu, &s->kvm_parked_vcpus, node) {
288         if (cpu->vcpu_id == vcpu_id) {
289             int kvm_fd;
290 
291             QLIST_REMOVE(cpu, node);
292             kvm_fd = cpu->kvm_fd;
293             g_free(cpu);
294             return kvm_fd;
295         }
296     }
297 
298     return kvm_vm_ioctl(s, KVM_CREATE_VCPU, (void *)vcpu_id);
299 }
300 
301 int kvm_init_vcpu(CPUState *cpu)
302 {
303     KVMState *s = kvm_state;
304     long mmap_size;
305     int ret;
306 
307     DPRINTF("kvm_init_vcpu\n");
308 
309     ret = kvm_get_vcpu(s, kvm_arch_vcpu_id(cpu));
310     if (ret < 0) {
311         DPRINTF("kvm_create_vcpu failed\n");
312         goto err;
313     }
314 
315     cpu->kvm_fd = ret;
316     cpu->kvm_state = s;
317     cpu->vcpu_dirty = true;
318 
319     mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
320     if (mmap_size < 0) {
321         ret = mmap_size;
322         DPRINTF("KVM_GET_VCPU_MMAP_SIZE failed\n");
323         goto err;
324     }
325 
326     cpu->kvm_run = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
327                         cpu->kvm_fd, 0);
328     if (cpu->kvm_run == MAP_FAILED) {
329         ret = -errno;
330         DPRINTF("mmap'ing vcpu state failed\n");
331         goto err;
332     }
333 
334     if (s->coalesced_mmio && !s->coalesced_mmio_ring) {
335         s->coalesced_mmio_ring =
336             (void *)cpu->kvm_run + s->coalesced_mmio * PAGE_SIZE;
337     }
338 
339     ret = kvm_arch_init_vcpu(cpu);
340 err:
341     return ret;
342 }
343 
344 /*
345  * dirty pages logging control
346  */
347 
348 static int kvm_mem_flags(MemoryRegion *mr)
349 {
350     bool readonly = mr->readonly || memory_region_is_romd(mr);
351     int flags = 0;
352 
353     if (memory_region_get_dirty_log_mask(mr) != 0) {
354         flags |= KVM_MEM_LOG_DIRTY_PAGES;
355     }
356     if (readonly && kvm_readonly_mem_allowed) {
357         flags |= KVM_MEM_READONLY;
358     }
359     return flags;
360 }
361 
362 static int kvm_slot_update_flags(KVMMemoryListener *kml, KVMSlot *mem,
363                                  MemoryRegion *mr)
364 {
365     int old_flags;
366 
367     old_flags = mem->flags;
368     mem->flags = kvm_mem_flags(mr);
369 
370     /* If nothing changed effectively, no need to issue ioctl */
371     if (mem->flags == old_flags) {
372         return 0;
373     }
374 
375     return kvm_set_user_memory_region(kml, mem);
376 }
377 
378 static int kvm_section_update_flags(KVMMemoryListener *kml,
379                                     MemoryRegionSection *section)
380 {
381     hwaddr start_addr, size;
382     KVMSlot *mem;
383 
384     size = kvm_align_section(section, &start_addr);
385     if (!size) {
386         return 0;
387     }
388 
389     mem = kvm_lookup_matching_slot(kml, start_addr, size);
390     if (!mem) {
391         /* We don't have a slot if we want to trap every access. */
392         return 0;
393     }
394 
395     return kvm_slot_update_flags(kml, mem, section->mr);
396 }
397 
398 static void kvm_log_start(MemoryListener *listener,
399                           MemoryRegionSection *section,
400                           int old, int new)
401 {
402     KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
403     int r;
404 
405     if (old != 0) {
406         return;
407     }
408 
409     r = kvm_section_update_flags(kml, section);
410     if (r < 0) {
411         abort();
412     }
413 }
414 
415 static void kvm_log_stop(MemoryListener *listener,
416                           MemoryRegionSection *section,
417                           int old, int new)
418 {
419     KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
420     int r;
421 
422     if (new != 0) {
423         return;
424     }
425 
426     r = kvm_section_update_flags(kml, section);
427     if (r < 0) {
428         abort();
429     }
430 }
431 
432 /* get kvm's dirty pages bitmap and update qemu's */
433 static int kvm_get_dirty_pages_log_range(MemoryRegionSection *section,
434                                          unsigned long *bitmap)
435 {
436     ram_addr_t start = section->offset_within_region +
437                        memory_region_get_ram_addr(section->mr);
438     ram_addr_t pages = int128_get64(section->size) / getpagesize();
439 
440     cpu_physical_memory_set_dirty_lebitmap(bitmap, start, pages);
441     return 0;
442 }
443 
444 #define ALIGN(x, y)  (((x)+(y)-1) & ~((y)-1))
445 
446 /**
447  * kvm_physical_sync_dirty_bitmap - Grab dirty bitmap from kernel space
448  * This function updates qemu's dirty bitmap using
449  * memory_region_set_dirty().  This means all bits are set
450  * to dirty.
451  *
452  * @start_add: start of logged region.
453  * @end_addr: end of logged region.
454  */
455 static int kvm_physical_sync_dirty_bitmap(KVMMemoryListener *kml,
456                                           MemoryRegionSection *section)
457 {
458     KVMState *s = kvm_state;
459     struct kvm_dirty_log d = {};
460     KVMSlot *mem;
461     hwaddr start_addr, size;
462 
463     size = kvm_align_section(section, &start_addr);
464     if (size) {
465         mem = kvm_lookup_matching_slot(kml, start_addr, size);
466         if (!mem) {
467             /* We don't have a slot if we want to trap every access. */
468             return 0;
469         }
470 
471         /* XXX bad kernel interface alert
472          * For dirty bitmap, kernel allocates array of size aligned to
473          * bits-per-long.  But for case when the kernel is 64bits and
474          * the userspace is 32bits, userspace can't align to the same
475          * bits-per-long, since sizeof(long) is different between kernel
476          * and user space.  This way, userspace will provide buffer which
477          * may be 4 bytes less than the kernel will use, resulting in
478          * userspace memory corruption (which is not detectable by valgrind
479          * too, in most cases).
480          * So for now, let's align to 64 instead of HOST_LONG_BITS here, in
481          * a hope that sizeof(long) won't become >8 any time soon.
482          */
483         size = ALIGN(((mem->memory_size) >> TARGET_PAGE_BITS),
484                      /*HOST_LONG_BITS*/ 64) / 8;
485         d.dirty_bitmap = g_malloc0(size);
486 
487         d.slot = mem->slot | (kml->as_id << 16);
488         if (kvm_vm_ioctl(s, KVM_GET_DIRTY_LOG, &d) == -1) {
489             DPRINTF("ioctl failed %d\n", errno);
490             g_free(d.dirty_bitmap);
491             return -1;
492         }
493 
494         kvm_get_dirty_pages_log_range(section, d.dirty_bitmap);
495         g_free(d.dirty_bitmap);
496     }
497 
498     return 0;
499 }
500 
501 static void kvm_coalesce_mmio_region(MemoryListener *listener,
502                                      MemoryRegionSection *secion,
503                                      hwaddr start, hwaddr size)
504 {
505     KVMState *s = kvm_state;
506 
507     if (s->coalesced_mmio) {
508         struct kvm_coalesced_mmio_zone zone;
509 
510         zone.addr = start;
511         zone.size = size;
512         zone.pad = 0;
513 
514         (void)kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone);
515     }
516 }
517 
518 static void kvm_uncoalesce_mmio_region(MemoryListener *listener,
519                                        MemoryRegionSection *secion,
520                                        hwaddr start, hwaddr size)
521 {
522     KVMState *s = kvm_state;
523 
524     if (s->coalesced_mmio) {
525         struct kvm_coalesced_mmio_zone zone;
526 
527         zone.addr = start;
528         zone.size = size;
529         zone.pad = 0;
530 
531         (void)kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone);
532     }
533 }
534 
535 int kvm_check_extension(KVMState *s, unsigned int extension)
536 {
537     int ret;
538 
539     ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, extension);
540     if (ret < 0) {
541         ret = 0;
542     }
543 
544     return ret;
545 }
546 
547 int kvm_vm_check_extension(KVMState *s, unsigned int extension)
548 {
549     int ret;
550 
551     ret = kvm_vm_ioctl(s, KVM_CHECK_EXTENSION, extension);
552     if (ret < 0) {
553         /* VM wide version not implemented, use global one instead */
554         ret = kvm_check_extension(s, extension);
555     }
556 
557     return ret;
558 }
559 
560 static uint32_t adjust_ioeventfd_endianness(uint32_t val, uint32_t size)
561 {
562 #if defined(HOST_WORDS_BIGENDIAN) != defined(TARGET_WORDS_BIGENDIAN)
563     /* The kernel expects ioeventfd values in HOST_WORDS_BIGENDIAN
564      * endianness, but the memory core hands them in target endianness.
565      * For example, PPC is always treated as big-endian even if running
566      * on KVM and on PPC64LE.  Correct here.
567      */
568     switch (size) {
569     case 2:
570         val = bswap16(val);
571         break;
572     case 4:
573         val = bswap32(val);
574         break;
575     }
576 #endif
577     return val;
578 }
579 
580 static int kvm_set_ioeventfd_mmio(int fd, hwaddr addr, uint32_t val,
581                                   bool assign, uint32_t size, bool datamatch)
582 {
583     int ret;
584     struct kvm_ioeventfd iofd = {
585         .datamatch = datamatch ? adjust_ioeventfd_endianness(val, size) : 0,
586         .addr = addr,
587         .len = size,
588         .flags = 0,
589         .fd = fd,
590     };
591 
592     if (!kvm_enabled()) {
593         return -ENOSYS;
594     }
595 
596     if (datamatch) {
597         iofd.flags |= KVM_IOEVENTFD_FLAG_DATAMATCH;
598     }
599     if (!assign) {
600         iofd.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
601     }
602 
603     ret = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &iofd);
604 
605     if (ret < 0) {
606         return -errno;
607     }
608 
609     return 0;
610 }
611 
612 static int kvm_set_ioeventfd_pio(int fd, uint16_t addr, uint16_t val,
613                                  bool assign, uint32_t size, bool datamatch)
614 {
615     struct kvm_ioeventfd kick = {
616         .datamatch = datamatch ? adjust_ioeventfd_endianness(val, size) : 0,
617         .addr = addr,
618         .flags = KVM_IOEVENTFD_FLAG_PIO,
619         .len = size,
620         .fd = fd,
621     };
622     int r;
623     if (!kvm_enabled()) {
624         return -ENOSYS;
625     }
626     if (datamatch) {
627         kick.flags |= KVM_IOEVENTFD_FLAG_DATAMATCH;
628     }
629     if (!assign) {
630         kick.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
631     }
632     r = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &kick);
633     if (r < 0) {
634         return r;
635     }
636     return 0;
637 }
638 
639 
640 static int kvm_check_many_ioeventfds(void)
641 {
642     /* Userspace can use ioeventfd for io notification.  This requires a host
643      * that supports eventfd(2) and an I/O thread; since eventfd does not
644      * support SIGIO it cannot interrupt the vcpu.
645      *
646      * Older kernels have a 6 device limit on the KVM io bus.  Find out so we
647      * can avoid creating too many ioeventfds.
648      */
649 #if defined(CONFIG_EVENTFD)
650     int ioeventfds[7];
651     int i, ret = 0;
652     for (i = 0; i < ARRAY_SIZE(ioeventfds); i++) {
653         ioeventfds[i] = eventfd(0, EFD_CLOEXEC);
654         if (ioeventfds[i] < 0) {
655             break;
656         }
657         ret = kvm_set_ioeventfd_pio(ioeventfds[i], 0, i, true, 2, true);
658         if (ret < 0) {
659             close(ioeventfds[i]);
660             break;
661         }
662     }
663 
664     /* Decide whether many devices are supported or not */
665     ret = i == ARRAY_SIZE(ioeventfds);
666 
667     while (i-- > 0) {
668         kvm_set_ioeventfd_pio(ioeventfds[i], 0, i, false, 2, true);
669         close(ioeventfds[i]);
670     }
671     return ret;
672 #else
673     return 0;
674 #endif
675 }
676 
677 static const KVMCapabilityInfo *
678 kvm_check_extension_list(KVMState *s, const KVMCapabilityInfo *list)
679 {
680     while (list->name) {
681         if (!kvm_check_extension(s, list->value)) {
682             return list;
683         }
684         list++;
685     }
686     return NULL;
687 }
688 
689 static void kvm_set_phys_mem(KVMMemoryListener *kml,
690                              MemoryRegionSection *section, bool add)
691 {
692     KVMSlot *mem;
693     int err;
694     MemoryRegion *mr = section->mr;
695     bool writeable = !mr->readonly && !mr->rom_device;
696     hwaddr start_addr, size;
697     void *ram;
698 
699     if (!memory_region_is_ram(mr)) {
700         if (writeable || !kvm_readonly_mem_allowed) {
701             return;
702         } else if (!mr->romd_mode) {
703             /* If the memory device is not in romd_mode, then we actually want
704              * to remove the kvm memory slot so all accesses will trap. */
705             add = false;
706         }
707     }
708 
709     size = kvm_align_section(section, &start_addr);
710     if (!size) {
711         return;
712     }
713 
714     /* use aligned delta to align the ram address */
715     ram = memory_region_get_ram_ptr(mr) + section->offset_within_region +
716           (start_addr - section->offset_within_address_space);
717 
718     if (!add) {
719         mem = kvm_lookup_matching_slot(kml, start_addr, size);
720         if (!mem) {
721             return;
722         }
723         if (mem->flags & KVM_MEM_LOG_DIRTY_PAGES) {
724             kvm_physical_sync_dirty_bitmap(kml, section);
725         }
726 
727         /* unregister the slot */
728         mem->memory_size = 0;
729         err = kvm_set_user_memory_region(kml, mem);
730         if (err) {
731             fprintf(stderr, "%s: error unregistering slot: %s\n",
732                     __func__, strerror(-err));
733             abort();
734         }
735         return;
736     }
737 
738     /* register the new slot */
739     mem = kvm_alloc_slot(kml);
740     mem->memory_size = size;
741     mem->start_addr = start_addr;
742     mem->ram = ram;
743     mem->flags = kvm_mem_flags(mr);
744 
745     err = kvm_set_user_memory_region(kml, mem);
746     if (err) {
747         fprintf(stderr, "%s: error registering slot: %s\n", __func__,
748                 strerror(-err));
749         abort();
750     }
751 }
752 
753 static void kvm_region_add(MemoryListener *listener,
754                            MemoryRegionSection *section)
755 {
756     KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
757 
758     memory_region_ref(section->mr);
759     kvm_set_phys_mem(kml, section, true);
760 }
761 
762 static void kvm_region_del(MemoryListener *listener,
763                            MemoryRegionSection *section)
764 {
765     KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
766 
767     kvm_set_phys_mem(kml, section, false);
768     memory_region_unref(section->mr);
769 }
770 
771 static void kvm_log_sync(MemoryListener *listener,
772                          MemoryRegionSection *section)
773 {
774     KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
775     int r;
776 
777     r = kvm_physical_sync_dirty_bitmap(kml, section);
778     if (r < 0) {
779         abort();
780     }
781 }
782 
783 static void kvm_mem_ioeventfd_add(MemoryListener *listener,
784                                   MemoryRegionSection *section,
785                                   bool match_data, uint64_t data,
786                                   EventNotifier *e)
787 {
788     int fd = event_notifier_get_fd(e);
789     int r;
790 
791     r = kvm_set_ioeventfd_mmio(fd, section->offset_within_address_space,
792                                data, true, int128_get64(section->size),
793                                match_data);
794     if (r < 0) {
795         fprintf(stderr, "%s: error adding ioeventfd: %s\n",
796                 __func__, strerror(-r));
797         abort();
798     }
799 }
800 
801 static void kvm_mem_ioeventfd_del(MemoryListener *listener,
802                                   MemoryRegionSection *section,
803                                   bool match_data, uint64_t data,
804                                   EventNotifier *e)
805 {
806     int fd = event_notifier_get_fd(e);
807     int r;
808 
809     r = kvm_set_ioeventfd_mmio(fd, section->offset_within_address_space,
810                                data, false, int128_get64(section->size),
811                                match_data);
812     if (r < 0) {
813         abort();
814     }
815 }
816 
817 static void kvm_io_ioeventfd_add(MemoryListener *listener,
818                                  MemoryRegionSection *section,
819                                  bool match_data, uint64_t data,
820                                  EventNotifier *e)
821 {
822     int fd = event_notifier_get_fd(e);
823     int r;
824 
825     r = kvm_set_ioeventfd_pio(fd, section->offset_within_address_space,
826                               data, true, int128_get64(section->size),
827                               match_data);
828     if (r < 0) {
829         fprintf(stderr, "%s: error adding ioeventfd: %s\n",
830                 __func__, strerror(-r));
831         abort();
832     }
833 }
834 
835 static void kvm_io_ioeventfd_del(MemoryListener *listener,
836                                  MemoryRegionSection *section,
837                                  bool match_data, uint64_t data,
838                                  EventNotifier *e)
839 
840 {
841     int fd = event_notifier_get_fd(e);
842     int r;
843 
844     r = kvm_set_ioeventfd_pio(fd, section->offset_within_address_space,
845                               data, false, int128_get64(section->size),
846                               match_data);
847     if (r < 0) {
848         abort();
849     }
850 }
851 
852 void kvm_memory_listener_register(KVMState *s, KVMMemoryListener *kml,
853                                   AddressSpace *as, int as_id)
854 {
855     int i;
856 
857     kml->slots = g_malloc0(s->nr_slots * sizeof(KVMSlot));
858     kml->as_id = as_id;
859 
860     for (i = 0; i < s->nr_slots; i++) {
861         kml->slots[i].slot = i;
862     }
863 
864     kml->listener.region_add = kvm_region_add;
865     kml->listener.region_del = kvm_region_del;
866     kml->listener.log_start = kvm_log_start;
867     kml->listener.log_stop = kvm_log_stop;
868     kml->listener.log_sync = kvm_log_sync;
869     kml->listener.priority = 10;
870 
871     memory_listener_register(&kml->listener, as);
872 }
873 
874 static MemoryListener kvm_io_listener = {
875     .eventfd_add = kvm_io_ioeventfd_add,
876     .eventfd_del = kvm_io_ioeventfd_del,
877     .priority = 10,
878 };
879 
880 int kvm_set_irq(KVMState *s, int irq, int level)
881 {
882     struct kvm_irq_level event;
883     int ret;
884 
885     assert(kvm_async_interrupts_enabled());
886 
887     event.level = level;
888     event.irq = irq;
889     ret = kvm_vm_ioctl(s, s->irq_set_ioctl, &event);
890     if (ret < 0) {
891         perror("kvm_set_irq");
892         abort();
893     }
894 
895     return (s->irq_set_ioctl == KVM_IRQ_LINE) ? 1 : event.status;
896 }
897 
898 #ifdef KVM_CAP_IRQ_ROUTING
899 typedef struct KVMMSIRoute {
900     struct kvm_irq_routing_entry kroute;
901     QTAILQ_ENTRY(KVMMSIRoute) entry;
902 } KVMMSIRoute;
903 
904 static void set_gsi(KVMState *s, unsigned int gsi)
905 {
906     set_bit(gsi, s->used_gsi_bitmap);
907 }
908 
909 static void clear_gsi(KVMState *s, unsigned int gsi)
910 {
911     clear_bit(gsi, s->used_gsi_bitmap);
912 }
913 
914 void kvm_init_irq_routing(KVMState *s)
915 {
916     int gsi_count, i;
917 
918     gsi_count = kvm_check_extension(s, KVM_CAP_IRQ_ROUTING) - 1;
919     if (gsi_count > 0) {
920         /* Round up so we can search ints using ffs */
921         s->used_gsi_bitmap = bitmap_new(gsi_count);
922         s->gsi_count = gsi_count;
923     }
924 
925     s->irq_routes = g_malloc0(sizeof(*s->irq_routes));
926     s->nr_allocated_irq_routes = 0;
927 
928     if (!kvm_direct_msi_allowed) {
929         for (i = 0; i < KVM_MSI_HASHTAB_SIZE; i++) {
930             QTAILQ_INIT(&s->msi_hashtab[i]);
931         }
932     }
933 
934     kvm_arch_init_irq_routing(s);
935 }
936 
937 void kvm_irqchip_commit_routes(KVMState *s)
938 {
939     int ret;
940 
941     if (kvm_gsi_direct_mapping()) {
942         return;
943     }
944 
945     if (!kvm_gsi_routing_enabled()) {
946         return;
947     }
948 
949     s->irq_routes->flags = 0;
950     trace_kvm_irqchip_commit_routes();
951     ret = kvm_vm_ioctl(s, KVM_SET_GSI_ROUTING, s->irq_routes);
952     assert(ret == 0);
953 }
954 
955 static void kvm_add_routing_entry(KVMState *s,
956                                   struct kvm_irq_routing_entry *entry)
957 {
958     struct kvm_irq_routing_entry *new;
959     int n, size;
960 
961     if (s->irq_routes->nr == s->nr_allocated_irq_routes) {
962         n = s->nr_allocated_irq_routes * 2;
963         if (n < 64) {
964             n = 64;
965         }
966         size = sizeof(struct kvm_irq_routing);
967         size += n * sizeof(*new);
968         s->irq_routes = g_realloc(s->irq_routes, size);
969         s->nr_allocated_irq_routes = n;
970     }
971     n = s->irq_routes->nr++;
972     new = &s->irq_routes->entries[n];
973 
974     *new = *entry;
975 
976     set_gsi(s, entry->gsi);
977 }
978 
979 static int kvm_update_routing_entry(KVMState *s,
980                                     struct kvm_irq_routing_entry *new_entry)
981 {
982     struct kvm_irq_routing_entry *entry;
983     int n;
984 
985     for (n = 0; n < s->irq_routes->nr; n++) {
986         entry = &s->irq_routes->entries[n];
987         if (entry->gsi != new_entry->gsi) {
988             continue;
989         }
990 
991         if(!memcmp(entry, new_entry, sizeof *entry)) {
992             return 0;
993         }
994 
995         *entry = *new_entry;
996 
997         return 0;
998     }
999 
1000     return -ESRCH;
1001 }
1002 
1003 void kvm_irqchip_add_irq_route(KVMState *s, int irq, int irqchip, int pin)
1004 {
1005     struct kvm_irq_routing_entry e = {};
1006 
1007     assert(pin < s->gsi_count);
1008 
1009     e.gsi = irq;
1010     e.type = KVM_IRQ_ROUTING_IRQCHIP;
1011     e.flags = 0;
1012     e.u.irqchip.irqchip = irqchip;
1013     e.u.irqchip.pin = pin;
1014     kvm_add_routing_entry(s, &e);
1015 }
1016 
1017 void kvm_irqchip_release_virq(KVMState *s, int virq)
1018 {
1019     struct kvm_irq_routing_entry *e;
1020     int i;
1021 
1022     if (kvm_gsi_direct_mapping()) {
1023         return;
1024     }
1025 
1026     for (i = 0; i < s->irq_routes->nr; i++) {
1027         e = &s->irq_routes->entries[i];
1028         if (e->gsi == virq) {
1029             s->irq_routes->nr--;
1030             *e = s->irq_routes->entries[s->irq_routes->nr];
1031         }
1032     }
1033     clear_gsi(s, virq);
1034     kvm_arch_release_virq_post(virq);
1035     trace_kvm_irqchip_release_virq(virq);
1036 }
1037 
1038 static unsigned int kvm_hash_msi(uint32_t data)
1039 {
1040     /* This is optimized for IA32 MSI layout. However, no other arch shall
1041      * repeat the mistake of not providing a direct MSI injection API. */
1042     return data & 0xff;
1043 }
1044 
1045 static void kvm_flush_dynamic_msi_routes(KVMState *s)
1046 {
1047     KVMMSIRoute *route, *next;
1048     unsigned int hash;
1049 
1050     for (hash = 0; hash < KVM_MSI_HASHTAB_SIZE; hash++) {
1051         QTAILQ_FOREACH_SAFE(route, &s->msi_hashtab[hash], entry, next) {
1052             kvm_irqchip_release_virq(s, route->kroute.gsi);
1053             QTAILQ_REMOVE(&s->msi_hashtab[hash], route, entry);
1054             g_free(route);
1055         }
1056     }
1057 }
1058 
1059 static int kvm_irqchip_get_virq(KVMState *s)
1060 {
1061     int next_virq;
1062 
1063     /*
1064      * PIC and IOAPIC share the first 16 GSI numbers, thus the available
1065      * GSI numbers are more than the number of IRQ route. Allocating a GSI
1066      * number can succeed even though a new route entry cannot be added.
1067      * When this happens, flush dynamic MSI entries to free IRQ route entries.
1068      */
1069     if (!kvm_direct_msi_allowed && s->irq_routes->nr == s->gsi_count) {
1070         kvm_flush_dynamic_msi_routes(s);
1071     }
1072 
1073     /* Return the lowest unused GSI in the bitmap */
1074     next_virq = find_first_zero_bit(s->used_gsi_bitmap, s->gsi_count);
1075     if (next_virq >= s->gsi_count) {
1076         return -ENOSPC;
1077     } else {
1078         return next_virq;
1079     }
1080 }
1081 
1082 static KVMMSIRoute *kvm_lookup_msi_route(KVMState *s, MSIMessage msg)
1083 {
1084     unsigned int hash = kvm_hash_msi(msg.data);
1085     KVMMSIRoute *route;
1086 
1087     QTAILQ_FOREACH(route, &s->msi_hashtab[hash], entry) {
1088         if (route->kroute.u.msi.address_lo == (uint32_t)msg.address &&
1089             route->kroute.u.msi.address_hi == (msg.address >> 32) &&
1090             route->kroute.u.msi.data == le32_to_cpu(msg.data)) {
1091             return route;
1092         }
1093     }
1094     return NULL;
1095 }
1096 
1097 int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg)
1098 {
1099     struct kvm_msi msi;
1100     KVMMSIRoute *route;
1101 
1102     if (kvm_direct_msi_allowed) {
1103         msi.address_lo = (uint32_t)msg.address;
1104         msi.address_hi = msg.address >> 32;
1105         msi.data = le32_to_cpu(msg.data);
1106         msi.flags = 0;
1107         memset(msi.pad, 0, sizeof(msi.pad));
1108 
1109         return kvm_vm_ioctl(s, KVM_SIGNAL_MSI, &msi);
1110     }
1111 
1112     route = kvm_lookup_msi_route(s, msg);
1113     if (!route) {
1114         int virq;
1115 
1116         virq = kvm_irqchip_get_virq(s);
1117         if (virq < 0) {
1118             return virq;
1119         }
1120 
1121         route = g_malloc0(sizeof(KVMMSIRoute));
1122         route->kroute.gsi = virq;
1123         route->kroute.type = KVM_IRQ_ROUTING_MSI;
1124         route->kroute.flags = 0;
1125         route->kroute.u.msi.address_lo = (uint32_t)msg.address;
1126         route->kroute.u.msi.address_hi = msg.address >> 32;
1127         route->kroute.u.msi.data = le32_to_cpu(msg.data);
1128 
1129         kvm_add_routing_entry(s, &route->kroute);
1130         kvm_irqchip_commit_routes(s);
1131 
1132         QTAILQ_INSERT_TAIL(&s->msi_hashtab[kvm_hash_msi(msg.data)], route,
1133                            entry);
1134     }
1135 
1136     assert(route->kroute.type == KVM_IRQ_ROUTING_MSI);
1137 
1138     return kvm_set_irq(s, route->kroute.gsi, 1);
1139 }
1140 
1141 int kvm_irqchip_add_msi_route(KVMState *s, int vector, PCIDevice *dev)
1142 {
1143     struct kvm_irq_routing_entry kroute = {};
1144     int virq;
1145     MSIMessage msg = {0, 0};
1146 
1147     if (pci_available && dev) {
1148         msg = pci_get_msi_message(dev, vector);
1149     }
1150 
1151     if (kvm_gsi_direct_mapping()) {
1152         return kvm_arch_msi_data_to_gsi(msg.data);
1153     }
1154 
1155     if (!kvm_gsi_routing_enabled()) {
1156         return -ENOSYS;
1157     }
1158 
1159     virq = kvm_irqchip_get_virq(s);
1160     if (virq < 0) {
1161         return virq;
1162     }
1163 
1164     kroute.gsi = virq;
1165     kroute.type = KVM_IRQ_ROUTING_MSI;
1166     kroute.flags = 0;
1167     kroute.u.msi.address_lo = (uint32_t)msg.address;
1168     kroute.u.msi.address_hi = msg.address >> 32;
1169     kroute.u.msi.data = le32_to_cpu(msg.data);
1170     if (pci_available && kvm_msi_devid_required()) {
1171         kroute.flags = KVM_MSI_VALID_DEVID;
1172         kroute.u.msi.devid = pci_requester_id(dev);
1173     }
1174     if (kvm_arch_fixup_msi_route(&kroute, msg.address, msg.data, dev)) {
1175         kvm_irqchip_release_virq(s, virq);
1176         return -EINVAL;
1177     }
1178 
1179     trace_kvm_irqchip_add_msi_route(dev ? dev->name : (char *)"N/A",
1180                                     vector, virq);
1181 
1182     kvm_add_routing_entry(s, &kroute);
1183     kvm_arch_add_msi_route_post(&kroute, vector, dev);
1184     kvm_irqchip_commit_routes(s);
1185 
1186     return virq;
1187 }
1188 
1189 int kvm_irqchip_update_msi_route(KVMState *s, int virq, MSIMessage msg,
1190                                  PCIDevice *dev)
1191 {
1192     struct kvm_irq_routing_entry kroute = {};
1193 
1194     if (kvm_gsi_direct_mapping()) {
1195         return 0;
1196     }
1197 
1198     if (!kvm_irqchip_in_kernel()) {
1199         return -ENOSYS;
1200     }
1201 
1202     kroute.gsi = virq;
1203     kroute.type = KVM_IRQ_ROUTING_MSI;
1204     kroute.flags = 0;
1205     kroute.u.msi.address_lo = (uint32_t)msg.address;
1206     kroute.u.msi.address_hi = msg.address >> 32;
1207     kroute.u.msi.data = le32_to_cpu(msg.data);
1208     if (pci_available && kvm_msi_devid_required()) {
1209         kroute.flags = KVM_MSI_VALID_DEVID;
1210         kroute.u.msi.devid = pci_requester_id(dev);
1211     }
1212     if (kvm_arch_fixup_msi_route(&kroute, msg.address, msg.data, dev)) {
1213         return -EINVAL;
1214     }
1215 
1216     trace_kvm_irqchip_update_msi_route(virq);
1217 
1218     return kvm_update_routing_entry(s, &kroute);
1219 }
1220 
1221 static int kvm_irqchip_assign_irqfd(KVMState *s, int fd, int rfd, int virq,
1222                                     bool assign)
1223 {
1224     struct kvm_irqfd irqfd = {
1225         .fd = fd,
1226         .gsi = virq,
1227         .flags = assign ? 0 : KVM_IRQFD_FLAG_DEASSIGN,
1228     };
1229 
1230     if (rfd != -1) {
1231         irqfd.flags |= KVM_IRQFD_FLAG_RESAMPLE;
1232         irqfd.resamplefd = rfd;
1233     }
1234 
1235     if (!kvm_irqfds_enabled()) {
1236         return -ENOSYS;
1237     }
1238 
1239     return kvm_vm_ioctl(s, KVM_IRQFD, &irqfd);
1240 }
1241 
1242 int kvm_irqchip_add_adapter_route(KVMState *s, AdapterInfo *adapter)
1243 {
1244     struct kvm_irq_routing_entry kroute = {};
1245     int virq;
1246 
1247     if (!kvm_gsi_routing_enabled()) {
1248         return -ENOSYS;
1249     }
1250 
1251     virq = kvm_irqchip_get_virq(s);
1252     if (virq < 0) {
1253         return virq;
1254     }
1255 
1256     kroute.gsi = virq;
1257     kroute.type = KVM_IRQ_ROUTING_S390_ADAPTER;
1258     kroute.flags = 0;
1259     kroute.u.adapter.summary_addr = adapter->summary_addr;
1260     kroute.u.adapter.ind_addr = adapter->ind_addr;
1261     kroute.u.adapter.summary_offset = adapter->summary_offset;
1262     kroute.u.adapter.ind_offset = adapter->ind_offset;
1263     kroute.u.adapter.adapter_id = adapter->adapter_id;
1264 
1265     kvm_add_routing_entry(s, &kroute);
1266 
1267     return virq;
1268 }
1269 
1270 int kvm_irqchip_add_hv_sint_route(KVMState *s, uint32_t vcpu, uint32_t sint)
1271 {
1272     struct kvm_irq_routing_entry kroute = {};
1273     int virq;
1274 
1275     if (!kvm_gsi_routing_enabled()) {
1276         return -ENOSYS;
1277     }
1278     if (!kvm_check_extension(s, KVM_CAP_HYPERV_SYNIC)) {
1279         return -ENOSYS;
1280     }
1281     virq = kvm_irqchip_get_virq(s);
1282     if (virq < 0) {
1283         return virq;
1284     }
1285 
1286     kroute.gsi = virq;
1287     kroute.type = KVM_IRQ_ROUTING_HV_SINT;
1288     kroute.flags = 0;
1289     kroute.u.hv_sint.vcpu = vcpu;
1290     kroute.u.hv_sint.sint = sint;
1291 
1292     kvm_add_routing_entry(s, &kroute);
1293     kvm_irqchip_commit_routes(s);
1294 
1295     return virq;
1296 }
1297 
1298 #else /* !KVM_CAP_IRQ_ROUTING */
1299 
1300 void kvm_init_irq_routing(KVMState *s)
1301 {
1302 }
1303 
1304 void kvm_irqchip_release_virq(KVMState *s, int virq)
1305 {
1306 }
1307 
1308 int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg)
1309 {
1310     abort();
1311 }
1312 
1313 int kvm_irqchip_add_msi_route(KVMState *s, int vector, PCIDevice *dev)
1314 {
1315     return -ENOSYS;
1316 }
1317 
1318 int kvm_irqchip_add_adapter_route(KVMState *s, AdapterInfo *adapter)
1319 {
1320     return -ENOSYS;
1321 }
1322 
1323 int kvm_irqchip_add_hv_sint_route(KVMState *s, uint32_t vcpu, uint32_t sint)
1324 {
1325     return -ENOSYS;
1326 }
1327 
1328 static int kvm_irqchip_assign_irqfd(KVMState *s, int fd, int virq, bool assign)
1329 {
1330     abort();
1331 }
1332 
1333 int kvm_irqchip_update_msi_route(KVMState *s, int virq, MSIMessage msg)
1334 {
1335     return -ENOSYS;
1336 }
1337 #endif /* !KVM_CAP_IRQ_ROUTING */
1338 
1339 int kvm_irqchip_add_irqfd_notifier_gsi(KVMState *s, EventNotifier *n,
1340                                        EventNotifier *rn, int virq)
1341 {
1342     return kvm_irqchip_assign_irqfd(s, event_notifier_get_fd(n),
1343            rn ? event_notifier_get_fd(rn) : -1, virq, true);
1344 }
1345 
1346 int kvm_irqchip_remove_irqfd_notifier_gsi(KVMState *s, EventNotifier *n,
1347                                           int virq)
1348 {
1349     return kvm_irqchip_assign_irqfd(s, event_notifier_get_fd(n), -1, virq,
1350            false);
1351 }
1352 
1353 int kvm_irqchip_add_irqfd_notifier(KVMState *s, EventNotifier *n,
1354                                    EventNotifier *rn, qemu_irq irq)
1355 {
1356     gpointer key, gsi;
1357     gboolean found = g_hash_table_lookup_extended(s->gsimap, irq, &key, &gsi);
1358 
1359     if (!found) {
1360         return -ENXIO;
1361     }
1362     return kvm_irqchip_add_irqfd_notifier_gsi(s, n, rn, GPOINTER_TO_INT(gsi));
1363 }
1364 
1365 int kvm_irqchip_remove_irqfd_notifier(KVMState *s, EventNotifier *n,
1366                                       qemu_irq irq)
1367 {
1368     gpointer key, gsi;
1369     gboolean found = g_hash_table_lookup_extended(s->gsimap, irq, &key, &gsi);
1370 
1371     if (!found) {
1372         return -ENXIO;
1373     }
1374     return kvm_irqchip_remove_irqfd_notifier_gsi(s, n, GPOINTER_TO_INT(gsi));
1375 }
1376 
1377 void kvm_irqchip_set_qemuirq_gsi(KVMState *s, qemu_irq irq, int gsi)
1378 {
1379     g_hash_table_insert(s->gsimap, irq, GINT_TO_POINTER(gsi));
1380 }
1381 
1382 static void kvm_irqchip_create(MachineState *machine, KVMState *s)
1383 {
1384     int ret;
1385 
1386     if (kvm_check_extension(s, KVM_CAP_IRQCHIP)) {
1387         ;
1388     } else if (kvm_check_extension(s, KVM_CAP_S390_IRQCHIP)) {
1389         ret = kvm_vm_enable_cap(s, KVM_CAP_S390_IRQCHIP, 0);
1390         if (ret < 0) {
1391             fprintf(stderr, "Enable kernel irqchip failed: %s\n", strerror(-ret));
1392             exit(1);
1393         }
1394     } else {
1395         return;
1396     }
1397 
1398     /* First probe and see if there's a arch-specific hook to create the
1399      * in-kernel irqchip for us */
1400     ret = kvm_arch_irqchip_create(machine, s);
1401     if (ret == 0) {
1402         if (machine_kernel_irqchip_split(machine)) {
1403             perror("Split IRQ chip mode not supported.");
1404             exit(1);
1405         } else {
1406             ret = kvm_vm_ioctl(s, KVM_CREATE_IRQCHIP);
1407         }
1408     }
1409     if (ret < 0) {
1410         fprintf(stderr, "Create kernel irqchip failed: %s\n", strerror(-ret));
1411         exit(1);
1412     }
1413 
1414     kvm_kernel_irqchip = true;
1415     /* If we have an in-kernel IRQ chip then we must have asynchronous
1416      * interrupt delivery (though the reverse is not necessarily true)
1417      */
1418     kvm_async_interrupts_allowed = true;
1419     kvm_halt_in_kernel_allowed = true;
1420 
1421     kvm_init_irq_routing(s);
1422 
1423     s->gsimap = g_hash_table_new(g_direct_hash, g_direct_equal);
1424 }
1425 
1426 /* Find number of supported CPUs using the recommended
1427  * procedure from the kernel API documentation to cope with
1428  * older kernels that may be missing capabilities.
1429  */
1430 static int kvm_recommended_vcpus(KVMState *s)
1431 {
1432     int ret = kvm_vm_check_extension(s, KVM_CAP_NR_VCPUS);
1433     return (ret) ? ret : 4;
1434 }
1435 
1436 static int kvm_max_vcpus(KVMState *s)
1437 {
1438     int ret = kvm_check_extension(s, KVM_CAP_MAX_VCPUS);
1439     return (ret) ? ret : kvm_recommended_vcpus(s);
1440 }
1441 
1442 static int kvm_max_vcpu_id(KVMState *s)
1443 {
1444     int ret = kvm_check_extension(s, KVM_CAP_MAX_VCPU_ID);
1445     return (ret) ? ret : kvm_max_vcpus(s);
1446 }
1447 
1448 bool kvm_vcpu_id_is_valid(int vcpu_id)
1449 {
1450     KVMState *s = KVM_STATE(current_machine->accelerator);
1451     return vcpu_id >= 0 && vcpu_id < kvm_max_vcpu_id(s);
1452 }
1453 
1454 static int kvm_init(MachineState *ms)
1455 {
1456     MachineClass *mc = MACHINE_GET_CLASS(ms);
1457     static const char upgrade_note[] =
1458         "Please upgrade to at least kernel 2.6.29 or recent kvm-kmod\n"
1459         "(see http://sourceforge.net/projects/kvm).\n";
1460     struct {
1461         const char *name;
1462         int num;
1463     } num_cpus[] = {
1464         { "SMP",          smp_cpus },
1465         { "hotpluggable", max_cpus },
1466         { NULL, }
1467     }, *nc = num_cpus;
1468     int soft_vcpus_limit, hard_vcpus_limit;
1469     KVMState *s;
1470     const KVMCapabilityInfo *missing_cap;
1471     int ret;
1472     int type = 0;
1473     const char *kvm_type;
1474 
1475     s = KVM_STATE(ms->accelerator);
1476 
1477     /*
1478      * On systems where the kernel can support different base page
1479      * sizes, host page size may be different from TARGET_PAGE_SIZE,
1480      * even with KVM.  TARGET_PAGE_SIZE is assumed to be the minimum
1481      * page size for the system though.
1482      */
1483     assert(TARGET_PAGE_SIZE <= getpagesize());
1484 
1485     s->sigmask_len = 8;
1486 
1487 #ifdef KVM_CAP_SET_GUEST_DEBUG
1488     QTAILQ_INIT(&s->kvm_sw_breakpoints);
1489 #endif
1490     QLIST_INIT(&s->kvm_parked_vcpus);
1491     s->vmfd = -1;
1492     s->fd = qemu_open("/dev/kvm", O_RDWR);
1493     if (s->fd == -1) {
1494         fprintf(stderr, "Could not access KVM kernel module: %m\n");
1495         ret = -errno;
1496         goto err;
1497     }
1498 
1499     ret = kvm_ioctl(s, KVM_GET_API_VERSION, 0);
1500     if (ret < KVM_API_VERSION) {
1501         if (ret >= 0) {
1502             ret = -EINVAL;
1503         }
1504         fprintf(stderr, "kvm version too old\n");
1505         goto err;
1506     }
1507 
1508     if (ret > KVM_API_VERSION) {
1509         ret = -EINVAL;
1510         fprintf(stderr, "kvm version not supported\n");
1511         goto err;
1512     }
1513 
1514     kvm_immediate_exit = kvm_check_extension(s, KVM_CAP_IMMEDIATE_EXIT);
1515     s->nr_slots = kvm_check_extension(s, KVM_CAP_NR_MEMSLOTS);
1516 
1517     /* If unspecified, use the default value */
1518     if (!s->nr_slots) {
1519         s->nr_slots = 32;
1520     }
1521 
1522     kvm_type = qemu_opt_get(qemu_get_machine_opts(), "kvm-type");
1523     if (mc->kvm_type) {
1524         type = mc->kvm_type(kvm_type);
1525     } else if (kvm_type) {
1526         ret = -EINVAL;
1527         fprintf(stderr, "Invalid argument kvm-type=%s\n", kvm_type);
1528         goto err;
1529     }
1530 
1531     do {
1532         ret = kvm_ioctl(s, KVM_CREATE_VM, type);
1533     } while (ret == -EINTR);
1534 
1535     if (ret < 0) {
1536         fprintf(stderr, "ioctl(KVM_CREATE_VM) failed: %d %s\n", -ret,
1537                 strerror(-ret));
1538 
1539 #ifdef TARGET_S390X
1540         if (ret == -EINVAL) {
1541             fprintf(stderr,
1542                     "Host kernel setup problem detected. Please verify:\n");
1543             fprintf(stderr, "- for kernels supporting the switch_amode or"
1544                     " user_mode parameters, whether\n");
1545             fprintf(stderr,
1546                     "  user space is running in primary address space\n");
1547             fprintf(stderr,
1548                     "- for kernels supporting the vm.allocate_pgste sysctl, "
1549                     "whether it is enabled\n");
1550         }
1551 #endif
1552         goto err;
1553     }
1554 
1555     s->vmfd = ret;
1556 
1557     /* check the vcpu limits */
1558     soft_vcpus_limit = kvm_recommended_vcpus(s);
1559     hard_vcpus_limit = kvm_max_vcpus(s);
1560 
1561     while (nc->name) {
1562         if (nc->num > soft_vcpus_limit) {
1563             warn_report("Number of %s cpus requested (%d) exceeds "
1564                         "the recommended cpus supported by KVM (%d)",
1565                         nc->name, nc->num, soft_vcpus_limit);
1566 
1567             if (nc->num > hard_vcpus_limit) {
1568                 fprintf(stderr, "Number of %s cpus requested (%d) exceeds "
1569                         "the maximum cpus supported by KVM (%d)\n",
1570                         nc->name, nc->num, hard_vcpus_limit);
1571                 exit(1);
1572             }
1573         }
1574         nc++;
1575     }
1576 
1577     missing_cap = kvm_check_extension_list(s, kvm_required_capabilites);
1578     if (!missing_cap) {
1579         missing_cap =
1580             kvm_check_extension_list(s, kvm_arch_required_capabilities);
1581     }
1582     if (missing_cap) {
1583         ret = -EINVAL;
1584         fprintf(stderr, "kvm does not support %s\n%s",
1585                 missing_cap->name, upgrade_note);
1586         goto err;
1587     }
1588 
1589     s->coalesced_mmio = kvm_check_extension(s, KVM_CAP_COALESCED_MMIO);
1590 
1591 #ifdef KVM_CAP_VCPU_EVENTS
1592     s->vcpu_events = kvm_check_extension(s, KVM_CAP_VCPU_EVENTS);
1593 #endif
1594 
1595     s->robust_singlestep =
1596         kvm_check_extension(s, KVM_CAP_X86_ROBUST_SINGLESTEP);
1597 
1598 #ifdef KVM_CAP_DEBUGREGS
1599     s->debugregs = kvm_check_extension(s, KVM_CAP_DEBUGREGS);
1600 #endif
1601 
1602 #ifdef KVM_CAP_IRQ_ROUTING
1603     kvm_direct_msi_allowed = (kvm_check_extension(s, KVM_CAP_SIGNAL_MSI) > 0);
1604 #endif
1605 
1606     s->intx_set_mask = kvm_check_extension(s, KVM_CAP_PCI_2_3);
1607 
1608     s->irq_set_ioctl = KVM_IRQ_LINE;
1609     if (kvm_check_extension(s, KVM_CAP_IRQ_INJECT_STATUS)) {
1610         s->irq_set_ioctl = KVM_IRQ_LINE_STATUS;
1611     }
1612 
1613 #ifdef KVM_CAP_READONLY_MEM
1614     kvm_readonly_mem_allowed =
1615         (kvm_check_extension(s, KVM_CAP_READONLY_MEM) > 0);
1616 #endif
1617 
1618     kvm_eventfds_allowed =
1619         (kvm_check_extension(s, KVM_CAP_IOEVENTFD) > 0);
1620 
1621     kvm_irqfds_allowed =
1622         (kvm_check_extension(s, KVM_CAP_IRQFD) > 0);
1623 
1624     kvm_resamplefds_allowed =
1625         (kvm_check_extension(s, KVM_CAP_IRQFD_RESAMPLE) > 0);
1626 
1627     kvm_vm_attributes_allowed =
1628         (kvm_check_extension(s, KVM_CAP_VM_ATTRIBUTES) > 0);
1629 
1630     kvm_ioeventfd_any_length_allowed =
1631         (kvm_check_extension(s, KVM_CAP_IOEVENTFD_ANY_LENGTH) > 0);
1632 
1633     kvm_state = s;
1634 
1635     ret = kvm_arch_init(ms, s);
1636     if (ret < 0) {
1637         goto err;
1638     }
1639 
1640     if (machine_kernel_irqchip_allowed(ms)) {
1641         kvm_irqchip_create(ms, s);
1642     }
1643 
1644     if (kvm_eventfds_allowed) {
1645         s->memory_listener.listener.eventfd_add = kvm_mem_ioeventfd_add;
1646         s->memory_listener.listener.eventfd_del = kvm_mem_ioeventfd_del;
1647     }
1648     s->memory_listener.listener.coalesced_mmio_add = kvm_coalesce_mmio_region;
1649     s->memory_listener.listener.coalesced_mmio_del = kvm_uncoalesce_mmio_region;
1650 
1651     kvm_memory_listener_register(s, &s->memory_listener,
1652                                  &address_space_memory, 0);
1653     memory_listener_register(&kvm_io_listener,
1654                              &address_space_io);
1655 
1656     s->many_ioeventfds = kvm_check_many_ioeventfds();
1657 
1658     s->sync_mmu = !!kvm_vm_check_extension(kvm_state, KVM_CAP_SYNC_MMU);
1659 
1660     return 0;
1661 
1662 err:
1663     assert(ret < 0);
1664     if (s->vmfd >= 0) {
1665         close(s->vmfd);
1666     }
1667     if (s->fd != -1) {
1668         close(s->fd);
1669     }
1670     g_free(s->memory_listener.slots);
1671 
1672     return ret;
1673 }
1674 
1675 void kvm_set_sigmask_len(KVMState *s, unsigned int sigmask_len)
1676 {
1677     s->sigmask_len = sigmask_len;
1678 }
1679 
1680 static void kvm_handle_io(uint16_t port, MemTxAttrs attrs, void *data, int direction,
1681                           int size, uint32_t count)
1682 {
1683     int i;
1684     uint8_t *ptr = data;
1685 
1686     for (i = 0; i < count; i++) {
1687         address_space_rw(&address_space_io, port, attrs,
1688                          ptr, size,
1689                          direction == KVM_EXIT_IO_OUT);
1690         ptr += size;
1691     }
1692 }
1693 
1694 static int kvm_handle_internal_error(CPUState *cpu, struct kvm_run *run)
1695 {
1696     fprintf(stderr, "KVM internal error. Suberror: %d\n",
1697             run->internal.suberror);
1698 
1699     if (kvm_check_extension(kvm_state, KVM_CAP_INTERNAL_ERROR_DATA)) {
1700         int i;
1701 
1702         for (i = 0; i < run->internal.ndata; ++i) {
1703             fprintf(stderr, "extra data[%d]: %"PRIx64"\n",
1704                     i, (uint64_t)run->internal.data[i]);
1705         }
1706     }
1707     if (run->internal.suberror == KVM_INTERNAL_ERROR_EMULATION) {
1708         fprintf(stderr, "emulation failure\n");
1709         if (!kvm_arch_stop_on_emulation_error(cpu)) {
1710             cpu_dump_state(cpu, stderr, fprintf, CPU_DUMP_CODE);
1711             return EXCP_INTERRUPT;
1712         }
1713     }
1714     /* FIXME: Should trigger a qmp message to let management know
1715      * something went wrong.
1716      */
1717     return -1;
1718 }
1719 
1720 void kvm_flush_coalesced_mmio_buffer(void)
1721 {
1722     KVMState *s = kvm_state;
1723 
1724     if (s->coalesced_flush_in_progress) {
1725         return;
1726     }
1727 
1728     s->coalesced_flush_in_progress = true;
1729 
1730     if (s->coalesced_mmio_ring) {
1731         struct kvm_coalesced_mmio_ring *ring = s->coalesced_mmio_ring;
1732         while (ring->first != ring->last) {
1733             struct kvm_coalesced_mmio *ent;
1734 
1735             ent = &ring->coalesced_mmio[ring->first];
1736 
1737             cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len);
1738             smp_wmb();
1739             ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX;
1740         }
1741     }
1742 
1743     s->coalesced_flush_in_progress = false;
1744 }
1745 
1746 static void do_kvm_cpu_synchronize_state(CPUState *cpu, run_on_cpu_data arg)
1747 {
1748     if (!cpu->vcpu_dirty) {
1749         kvm_arch_get_registers(cpu);
1750         cpu->vcpu_dirty = true;
1751     }
1752 }
1753 
1754 void kvm_cpu_synchronize_state(CPUState *cpu)
1755 {
1756     if (!cpu->vcpu_dirty) {
1757         run_on_cpu(cpu, do_kvm_cpu_synchronize_state, RUN_ON_CPU_NULL);
1758     }
1759 }
1760 
1761 static void do_kvm_cpu_synchronize_post_reset(CPUState *cpu, run_on_cpu_data arg)
1762 {
1763     kvm_arch_put_registers(cpu, KVM_PUT_RESET_STATE);
1764     cpu->vcpu_dirty = false;
1765 }
1766 
1767 void kvm_cpu_synchronize_post_reset(CPUState *cpu)
1768 {
1769     run_on_cpu(cpu, do_kvm_cpu_synchronize_post_reset, RUN_ON_CPU_NULL);
1770 }
1771 
1772 static void do_kvm_cpu_synchronize_post_init(CPUState *cpu, run_on_cpu_data arg)
1773 {
1774     kvm_arch_put_registers(cpu, KVM_PUT_FULL_STATE);
1775     cpu->vcpu_dirty = false;
1776 }
1777 
1778 void kvm_cpu_synchronize_post_init(CPUState *cpu)
1779 {
1780     run_on_cpu(cpu, do_kvm_cpu_synchronize_post_init, RUN_ON_CPU_NULL);
1781 }
1782 
1783 static void do_kvm_cpu_synchronize_pre_loadvm(CPUState *cpu, run_on_cpu_data arg)
1784 {
1785     cpu->vcpu_dirty = true;
1786 }
1787 
1788 void kvm_cpu_synchronize_pre_loadvm(CPUState *cpu)
1789 {
1790     run_on_cpu(cpu, do_kvm_cpu_synchronize_pre_loadvm, RUN_ON_CPU_NULL);
1791 }
1792 
1793 #ifdef KVM_HAVE_MCE_INJECTION
1794 static __thread void *pending_sigbus_addr;
1795 static __thread int pending_sigbus_code;
1796 static __thread bool have_sigbus_pending;
1797 #endif
1798 
1799 static void kvm_cpu_kick(CPUState *cpu)
1800 {
1801     atomic_set(&cpu->kvm_run->immediate_exit, 1);
1802 }
1803 
1804 static void kvm_cpu_kick_self(void)
1805 {
1806     if (kvm_immediate_exit) {
1807         kvm_cpu_kick(current_cpu);
1808     } else {
1809         qemu_cpu_kick_self();
1810     }
1811 }
1812 
1813 static void kvm_eat_signals(CPUState *cpu)
1814 {
1815     struct timespec ts = { 0, 0 };
1816     siginfo_t siginfo;
1817     sigset_t waitset;
1818     sigset_t chkset;
1819     int r;
1820 
1821     if (kvm_immediate_exit) {
1822         atomic_set(&cpu->kvm_run->immediate_exit, 0);
1823         /* Write kvm_run->immediate_exit before the cpu->exit_request
1824          * write in kvm_cpu_exec.
1825          */
1826         smp_wmb();
1827         return;
1828     }
1829 
1830     sigemptyset(&waitset);
1831     sigaddset(&waitset, SIG_IPI);
1832 
1833     do {
1834         r = sigtimedwait(&waitset, &siginfo, &ts);
1835         if (r == -1 && !(errno == EAGAIN || errno == EINTR)) {
1836             perror("sigtimedwait");
1837             exit(1);
1838         }
1839 
1840         r = sigpending(&chkset);
1841         if (r == -1) {
1842             perror("sigpending");
1843             exit(1);
1844         }
1845     } while (sigismember(&chkset, SIG_IPI));
1846 }
1847 
1848 int kvm_cpu_exec(CPUState *cpu)
1849 {
1850     struct kvm_run *run = cpu->kvm_run;
1851     int ret, run_ret;
1852 
1853     DPRINTF("kvm_cpu_exec()\n");
1854 
1855     if (kvm_arch_process_async_events(cpu)) {
1856         atomic_set(&cpu->exit_request, 0);
1857         return EXCP_HLT;
1858     }
1859 
1860     qemu_mutex_unlock_iothread();
1861     cpu_exec_start(cpu);
1862 
1863     do {
1864         MemTxAttrs attrs;
1865 
1866         if (cpu->vcpu_dirty) {
1867             kvm_arch_put_registers(cpu, KVM_PUT_RUNTIME_STATE);
1868             cpu->vcpu_dirty = false;
1869         }
1870 
1871         kvm_arch_pre_run(cpu, run);
1872         if (atomic_read(&cpu->exit_request)) {
1873             DPRINTF("interrupt exit requested\n");
1874             /*
1875              * KVM requires us to reenter the kernel after IO exits to complete
1876              * instruction emulation. This self-signal will ensure that we
1877              * leave ASAP again.
1878              */
1879             kvm_cpu_kick_self();
1880         }
1881 
1882         /* Read cpu->exit_request before KVM_RUN reads run->immediate_exit.
1883          * Matching barrier in kvm_eat_signals.
1884          */
1885         smp_rmb();
1886 
1887         run_ret = kvm_vcpu_ioctl(cpu, KVM_RUN, 0);
1888 
1889         attrs = kvm_arch_post_run(cpu, run);
1890 
1891 #ifdef KVM_HAVE_MCE_INJECTION
1892         if (unlikely(have_sigbus_pending)) {
1893             qemu_mutex_lock_iothread();
1894             kvm_arch_on_sigbus_vcpu(cpu, pending_sigbus_code,
1895                                     pending_sigbus_addr);
1896             have_sigbus_pending = false;
1897             qemu_mutex_unlock_iothread();
1898         }
1899 #endif
1900 
1901         if (run_ret < 0) {
1902             if (run_ret == -EINTR || run_ret == -EAGAIN) {
1903                 DPRINTF("io window exit\n");
1904                 kvm_eat_signals(cpu);
1905                 ret = EXCP_INTERRUPT;
1906                 break;
1907             }
1908             fprintf(stderr, "error: kvm run failed %s\n",
1909                     strerror(-run_ret));
1910 #ifdef TARGET_PPC
1911             if (run_ret == -EBUSY) {
1912                 fprintf(stderr,
1913                         "This is probably because your SMT is enabled.\n"
1914                         "VCPU can only run on primary threads with all "
1915                         "secondary threads offline.\n");
1916             }
1917 #endif
1918             ret = -1;
1919             break;
1920         }
1921 
1922         trace_kvm_run_exit(cpu->cpu_index, run->exit_reason);
1923         switch (run->exit_reason) {
1924         case KVM_EXIT_IO:
1925             DPRINTF("handle_io\n");
1926             /* Called outside BQL */
1927             kvm_handle_io(run->io.port, attrs,
1928                           (uint8_t *)run + run->io.data_offset,
1929                           run->io.direction,
1930                           run->io.size,
1931                           run->io.count);
1932             ret = 0;
1933             break;
1934         case KVM_EXIT_MMIO:
1935             DPRINTF("handle_mmio\n");
1936             /* Called outside BQL */
1937             address_space_rw(&address_space_memory,
1938                              run->mmio.phys_addr, attrs,
1939                              run->mmio.data,
1940                              run->mmio.len,
1941                              run->mmio.is_write);
1942             ret = 0;
1943             break;
1944         case KVM_EXIT_IRQ_WINDOW_OPEN:
1945             DPRINTF("irq_window_open\n");
1946             ret = EXCP_INTERRUPT;
1947             break;
1948         case KVM_EXIT_SHUTDOWN:
1949             DPRINTF("shutdown\n");
1950             qemu_system_reset_request(SHUTDOWN_CAUSE_GUEST_RESET);
1951             ret = EXCP_INTERRUPT;
1952             break;
1953         case KVM_EXIT_UNKNOWN:
1954             fprintf(stderr, "KVM: unknown exit, hardware reason %" PRIx64 "\n",
1955                     (uint64_t)run->hw.hardware_exit_reason);
1956             ret = -1;
1957             break;
1958         case KVM_EXIT_INTERNAL_ERROR:
1959             ret = kvm_handle_internal_error(cpu, run);
1960             break;
1961         case KVM_EXIT_SYSTEM_EVENT:
1962             switch (run->system_event.type) {
1963             case KVM_SYSTEM_EVENT_SHUTDOWN:
1964                 qemu_system_shutdown_request(SHUTDOWN_CAUSE_GUEST_SHUTDOWN);
1965                 ret = EXCP_INTERRUPT;
1966                 break;
1967             case KVM_SYSTEM_EVENT_RESET:
1968                 qemu_system_reset_request(SHUTDOWN_CAUSE_GUEST_RESET);
1969                 ret = EXCP_INTERRUPT;
1970                 break;
1971             case KVM_SYSTEM_EVENT_CRASH:
1972                 kvm_cpu_synchronize_state(cpu);
1973                 qemu_mutex_lock_iothread();
1974                 qemu_system_guest_panicked(cpu_get_crash_info(cpu));
1975                 qemu_mutex_unlock_iothread();
1976                 ret = 0;
1977                 break;
1978             default:
1979                 DPRINTF("kvm_arch_handle_exit\n");
1980                 ret = kvm_arch_handle_exit(cpu, run);
1981                 break;
1982             }
1983             break;
1984         default:
1985             DPRINTF("kvm_arch_handle_exit\n");
1986             ret = kvm_arch_handle_exit(cpu, run);
1987             break;
1988         }
1989     } while (ret == 0);
1990 
1991     cpu_exec_end(cpu);
1992     qemu_mutex_lock_iothread();
1993 
1994     if (ret < 0) {
1995         cpu_dump_state(cpu, stderr, fprintf, CPU_DUMP_CODE);
1996         vm_stop(RUN_STATE_INTERNAL_ERROR);
1997     }
1998 
1999     atomic_set(&cpu->exit_request, 0);
2000     return ret;
2001 }
2002 
2003 int kvm_ioctl(KVMState *s, int type, ...)
2004 {
2005     int ret;
2006     void *arg;
2007     va_list ap;
2008 
2009     va_start(ap, type);
2010     arg = va_arg(ap, void *);
2011     va_end(ap);
2012 
2013     trace_kvm_ioctl(type, arg);
2014     ret = ioctl(s->fd, type, arg);
2015     if (ret == -1) {
2016         ret = -errno;
2017     }
2018     return ret;
2019 }
2020 
2021 int kvm_vm_ioctl(KVMState *s, int type, ...)
2022 {
2023     int ret;
2024     void *arg;
2025     va_list ap;
2026 
2027     va_start(ap, type);
2028     arg = va_arg(ap, void *);
2029     va_end(ap);
2030 
2031     trace_kvm_vm_ioctl(type, arg);
2032     ret = ioctl(s->vmfd, type, arg);
2033     if (ret == -1) {
2034         ret = -errno;
2035     }
2036     return ret;
2037 }
2038 
2039 int kvm_vcpu_ioctl(CPUState *cpu, int type, ...)
2040 {
2041     int ret;
2042     void *arg;
2043     va_list ap;
2044 
2045     va_start(ap, type);
2046     arg = va_arg(ap, void *);
2047     va_end(ap);
2048 
2049     trace_kvm_vcpu_ioctl(cpu->cpu_index, type, arg);
2050     ret = ioctl(cpu->kvm_fd, type, arg);
2051     if (ret == -1) {
2052         ret = -errno;
2053     }
2054     return ret;
2055 }
2056 
2057 int kvm_device_ioctl(int fd, int type, ...)
2058 {
2059     int ret;
2060     void *arg;
2061     va_list ap;
2062 
2063     va_start(ap, type);
2064     arg = va_arg(ap, void *);
2065     va_end(ap);
2066 
2067     trace_kvm_device_ioctl(fd, type, arg);
2068     ret = ioctl(fd, type, arg);
2069     if (ret == -1) {
2070         ret = -errno;
2071     }
2072     return ret;
2073 }
2074 
2075 int kvm_vm_check_attr(KVMState *s, uint32_t group, uint64_t attr)
2076 {
2077     int ret;
2078     struct kvm_device_attr attribute = {
2079         .group = group,
2080         .attr = attr,
2081     };
2082 
2083     if (!kvm_vm_attributes_allowed) {
2084         return 0;
2085     }
2086 
2087     ret = kvm_vm_ioctl(s, KVM_HAS_DEVICE_ATTR, &attribute);
2088     /* kvm returns 0 on success for HAS_DEVICE_ATTR */
2089     return ret ? 0 : 1;
2090 }
2091 
2092 int kvm_device_check_attr(int dev_fd, uint32_t group, uint64_t attr)
2093 {
2094     struct kvm_device_attr attribute = {
2095         .group = group,
2096         .attr = attr,
2097         .flags = 0,
2098     };
2099 
2100     return kvm_device_ioctl(dev_fd, KVM_HAS_DEVICE_ATTR, &attribute) ? 0 : 1;
2101 }
2102 
2103 int kvm_device_access(int fd, int group, uint64_t attr,
2104                       void *val, bool write, Error **errp)
2105 {
2106     struct kvm_device_attr kvmattr;
2107     int err;
2108 
2109     kvmattr.flags = 0;
2110     kvmattr.group = group;
2111     kvmattr.attr = attr;
2112     kvmattr.addr = (uintptr_t)val;
2113 
2114     err = kvm_device_ioctl(fd,
2115                            write ? KVM_SET_DEVICE_ATTR : KVM_GET_DEVICE_ATTR,
2116                            &kvmattr);
2117     if (err < 0) {
2118         error_setg_errno(errp, -err,
2119                          "KVM_%s_DEVICE_ATTR failed: Group %d "
2120                          "attr 0x%016" PRIx64,
2121                          write ? "SET" : "GET", group, attr);
2122     }
2123     return err;
2124 }
2125 
2126 bool kvm_has_sync_mmu(void)
2127 {
2128     return kvm_state->sync_mmu;
2129 }
2130 
2131 int kvm_has_vcpu_events(void)
2132 {
2133     return kvm_state->vcpu_events;
2134 }
2135 
2136 int kvm_has_robust_singlestep(void)
2137 {
2138     return kvm_state->robust_singlestep;
2139 }
2140 
2141 int kvm_has_debugregs(void)
2142 {
2143     return kvm_state->debugregs;
2144 }
2145 
2146 int kvm_has_many_ioeventfds(void)
2147 {
2148     if (!kvm_enabled()) {
2149         return 0;
2150     }
2151     return kvm_state->many_ioeventfds;
2152 }
2153 
2154 int kvm_has_gsi_routing(void)
2155 {
2156 #ifdef KVM_CAP_IRQ_ROUTING
2157     return kvm_check_extension(kvm_state, KVM_CAP_IRQ_ROUTING);
2158 #else
2159     return false;
2160 #endif
2161 }
2162 
2163 int kvm_has_intx_set_mask(void)
2164 {
2165     return kvm_state->intx_set_mask;
2166 }
2167 
2168 bool kvm_arm_supports_user_irq(void)
2169 {
2170     return kvm_check_extension(kvm_state, KVM_CAP_ARM_USER_IRQ);
2171 }
2172 
2173 #ifdef KVM_CAP_SET_GUEST_DEBUG
2174 struct kvm_sw_breakpoint *kvm_find_sw_breakpoint(CPUState *cpu,
2175                                                  target_ulong pc)
2176 {
2177     struct kvm_sw_breakpoint *bp;
2178 
2179     QTAILQ_FOREACH(bp, &cpu->kvm_state->kvm_sw_breakpoints, entry) {
2180         if (bp->pc == pc) {
2181             return bp;
2182         }
2183     }
2184     return NULL;
2185 }
2186 
2187 int kvm_sw_breakpoints_active(CPUState *cpu)
2188 {
2189     return !QTAILQ_EMPTY(&cpu->kvm_state->kvm_sw_breakpoints);
2190 }
2191 
2192 struct kvm_set_guest_debug_data {
2193     struct kvm_guest_debug dbg;
2194     int err;
2195 };
2196 
2197 static void kvm_invoke_set_guest_debug(CPUState *cpu, run_on_cpu_data data)
2198 {
2199     struct kvm_set_guest_debug_data *dbg_data =
2200         (struct kvm_set_guest_debug_data *) data.host_ptr;
2201 
2202     dbg_data->err = kvm_vcpu_ioctl(cpu, KVM_SET_GUEST_DEBUG,
2203                                    &dbg_data->dbg);
2204 }
2205 
2206 int kvm_update_guest_debug(CPUState *cpu, unsigned long reinject_trap)
2207 {
2208     struct kvm_set_guest_debug_data data;
2209 
2210     data.dbg.control = reinject_trap;
2211 
2212     if (cpu->singlestep_enabled) {
2213         data.dbg.control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_SINGLESTEP;
2214     }
2215     kvm_arch_update_guest_debug(cpu, &data.dbg);
2216 
2217     run_on_cpu(cpu, kvm_invoke_set_guest_debug,
2218                RUN_ON_CPU_HOST_PTR(&data));
2219     return data.err;
2220 }
2221 
2222 int kvm_insert_breakpoint(CPUState *cpu, target_ulong addr,
2223                           target_ulong len, int type)
2224 {
2225     struct kvm_sw_breakpoint *bp;
2226     int err;
2227 
2228     if (type == GDB_BREAKPOINT_SW) {
2229         bp = kvm_find_sw_breakpoint(cpu, addr);
2230         if (bp) {
2231             bp->use_count++;
2232             return 0;
2233         }
2234 
2235         bp = g_malloc(sizeof(struct kvm_sw_breakpoint));
2236         bp->pc = addr;
2237         bp->use_count = 1;
2238         err = kvm_arch_insert_sw_breakpoint(cpu, bp);
2239         if (err) {
2240             g_free(bp);
2241             return err;
2242         }
2243 
2244         QTAILQ_INSERT_HEAD(&cpu->kvm_state->kvm_sw_breakpoints, bp, entry);
2245     } else {
2246         err = kvm_arch_insert_hw_breakpoint(addr, len, type);
2247         if (err) {
2248             return err;
2249         }
2250     }
2251 
2252     CPU_FOREACH(cpu) {
2253         err = kvm_update_guest_debug(cpu, 0);
2254         if (err) {
2255             return err;
2256         }
2257     }
2258     return 0;
2259 }
2260 
2261 int kvm_remove_breakpoint(CPUState *cpu, target_ulong addr,
2262                           target_ulong len, int type)
2263 {
2264     struct kvm_sw_breakpoint *bp;
2265     int err;
2266 
2267     if (type == GDB_BREAKPOINT_SW) {
2268         bp = kvm_find_sw_breakpoint(cpu, addr);
2269         if (!bp) {
2270             return -ENOENT;
2271         }
2272 
2273         if (bp->use_count > 1) {
2274             bp->use_count--;
2275             return 0;
2276         }
2277 
2278         err = kvm_arch_remove_sw_breakpoint(cpu, bp);
2279         if (err) {
2280             return err;
2281         }
2282 
2283         QTAILQ_REMOVE(&cpu->kvm_state->kvm_sw_breakpoints, bp, entry);
2284         g_free(bp);
2285     } else {
2286         err = kvm_arch_remove_hw_breakpoint(addr, len, type);
2287         if (err) {
2288             return err;
2289         }
2290     }
2291 
2292     CPU_FOREACH(cpu) {
2293         err = kvm_update_guest_debug(cpu, 0);
2294         if (err) {
2295             return err;
2296         }
2297     }
2298     return 0;
2299 }
2300 
2301 void kvm_remove_all_breakpoints(CPUState *cpu)
2302 {
2303     struct kvm_sw_breakpoint *bp, *next;
2304     KVMState *s = cpu->kvm_state;
2305     CPUState *tmpcpu;
2306 
2307     QTAILQ_FOREACH_SAFE(bp, &s->kvm_sw_breakpoints, entry, next) {
2308         if (kvm_arch_remove_sw_breakpoint(cpu, bp) != 0) {
2309             /* Try harder to find a CPU that currently sees the breakpoint. */
2310             CPU_FOREACH(tmpcpu) {
2311                 if (kvm_arch_remove_sw_breakpoint(tmpcpu, bp) == 0) {
2312                     break;
2313                 }
2314             }
2315         }
2316         QTAILQ_REMOVE(&s->kvm_sw_breakpoints, bp, entry);
2317         g_free(bp);
2318     }
2319     kvm_arch_remove_all_hw_breakpoints();
2320 
2321     CPU_FOREACH(cpu) {
2322         kvm_update_guest_debug(cpu, 0);
2323     }
2324 }
2325 
2326 #else /* !KVM_CAP_SET_GUEST_DEBUG */
2327 
2328 int kvm_update_guest_debug(CPUState *cpu, unsigned long reinject_trap)
2329 {
2330     return -EINVAL;
2331 }
2332 
2333 int kvm_insert_breakpoint(CPUState *cpu, target_ulong addr,
2334                           target_ulong len, int type)
2335 {
2336     return -EINVAL;
2337 }
2338 
2339 int kvm_remove_breakpoint(CPUState *cpu, target_ulong addr,
2340                           target_ulong len, int type)
2341 {
2342     return -EINVAL;
2343 }
2344 
2345 void kvm_remove_all_breakpoints(CPUState *cpu)
2346 {
2347 }
2348 #endif /* !KVM_CAP_SET_GUEST_DEBUG */
2349 
2350 static int kvm_set_signal_mask(CPUState *cpu, const sigset_t *sigset)
2351 {
2352     KVMState *s = kvm_state;
2353     struct kvm_signal_mask *sigmask;
2354     int r;
2355 
2356     sigmask = g_malloc(sizeof(*sigmask) + sizeof(*sigset));
2357 
2358     sigmask->len = s->sigmask_len;
2359     memcpy(sigmask->sigset, sigset, sizeof(*sigset));
2360     r = kvm_vcpu_ioctl(cpu, KVM_SET_SIGNAL_MASK, sigmask);
2361     g_free(sigmask);
2362 
2363     return r;
2364 }
2365 
2366 static void kvm_ipi_signal(int sig)
2367 {
2368     if (current_cpu) {
2369         assert(kvm_immediate_exit);
2370         kvm_cpu_kick(current_cpu);
2371     }
2372 }
2373 
2374 void kvm_init_cpu_signals(CPUState *cpu)
2375 {
2376     int r;
2377     sigset_t set;
2378     struct sigaction sigact;
2379 
2380     memset(&sigact, 0, sizeof(sigact));
2381     sigact.sa_handler = kvm_ipi_signal;
2382     sigaction(SIG_IPI, &sigact, NULL);
2383 
2384     pthread_sigmask(SIG_BLOCK, NULL, &set);
2385 #if defined KVM_HAVE_MCE_INJECTION
2386     sigdelset(&set, SIGBUS);
2387     pthread_sigmask(SIG_SETMASK, &set, NULL);
2388 #endif
2389     sigdelset(&set, SIG_IPI);
2390     if (kvm_immediate_exit) {
2391         r = pthread_sigmask(SIG_SETMASK, &set, NULL);
2392     } else {
2393         r = kvm_set_signal_mask(cpu, &set);
2394     }
2395     if (r) {
2396         fprintf(stderr, "kvm_set_signal_mask: %s\n", strerror(-r));
2397         exit(1);
2398     }
2399 }
2400 
2401 /* Called asynchronously in VCPU thread.  */
2402 int kvm_on_sigbus_vcpu(CPUState *cpu, int code, void *addr)
2403 {
2404 #ifdef KVM_HAVE_MCE_INJECTION
2405     if (have_sigbus_pending) {
2406         return 1;
2407     }
2408     have_sigbus_pending = true;
2409     pending_sigbus_addr = addr;
2410     pending_sigbus_code = code;
2411     atomic_set(&cpu->exit_request, 1);
2412     return 0;
2413 #else
2414     return 1;
2415 #endif
2416 }
2417 
2418 /* Called synchronously (via signalfd) in main thread.  */
2419 int kvm_on_sigbus(int code, void *addr)
2420 {
2421 #ifdef KVM_HAVE_MCE_INJECTION
2422     /* Action required MCE kills the process if SIGBUS is blocked.  Because
2423      * that's what happens in the I/O thread, where we handle MCE via signalfd,
2424      * we can only get action optional here.
2425      */
2426     assert(code != BUS_MCEERR_AR);
2427     kvm_arch_on_sigbus_vcpu(first_cpu, code, addr);
2428     return 0;
2429 #else
2430     return 1;
2431 #endif
2432 }
2433 
2434 int kvm_create_device(KVMState *s, uint64_t type, bool test)
2435 {
2436     int ret;
2437     struct kvm_create_device create_dev;
2438 
2439     create_dev.type = type;
2440     create_dev.fd = -1;
2441     create_dev.flags = test ? KVM_CREATE_DEVICE_TEST : 0;
2442 
2443     if (!kvm_check_extension(s, KVM_CAP_DEVICE_CTRL)) {
2444         return -ENOTSUP;
2445     }
2446 
2447     ret = kvm_vm_ioctl(s, KVM_CREATE_DEVICE, &create_dev);
2448     if (ret) {
2449         return ret;
2450     }
2451 
2452     return test ? 0 : create_dev.fd;
2453 }
2454 
2455 bool kvm_device_supported(int vmfd, uint64_t type)
2456 {
2457     struct kvm_create_device create_dev = {
2458         .type = type,
2459         .fd = -1,
2460         .flags = KVM_CREATE_DEVICE_TEST,
2461     };
2462 
2463     if (ioctl(vmfd, KVM_CHECK_EXTENSION, KVM_CAP_DEVICE_CTRL) <= 0) {
2464         return false;
2465     }
2466 
2467     return (ioctl(vmfd, KVM_CREATE_DEVICE, &create_dev) >= 0);
2468 }
2469 
2470 int kvm_set_one_reg(CPUState *cs, uint64_t id, void *source)
2471 {
2472     struct kvm_one_reg reg;
2473     int r;
2474 
2475     reg.id = id;
2476     reg.addr = (uintptr_t) source;
2477     r = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, &reg);
2478     if (r) {
2479         trace_kvm_failed_reg_set(id, strerror(-r));
2480     }
2481     return r;
2482 }
2483 
2484 int kvm_get_one_reg(CPUState *cs, uint64_t id, void *target)
2485 {
2486     struct kvm_one_reg reg;
2487     int r;
2488 
2489     reg.id = id;
2490     reg.addr = (uintptr_t) target;
2491     r = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &reg);
2492     if (r) {
2493         trace_kvm_failed_reg_get(id, strerror(-r));
2494     }
2495     return r;
2496 }
2497 
2498 static void kvm_accel_class_init(ObjectClass *oc, void *data)
2499 {
2500     AccelClass *ac = ACCEL_CLASS(oc);
2501     ac->name = "KVM";
2502     ac->init_machine = kvm_init;
2503     ac->allowed = &kvm_allowed;
2504 }
2505 
2506 static const TypeInfo kvm_accel_type = {
2507     .name = TYPE_KVM_ACCEL,
2508     .parent = TYPE_ACCEL,
2509     .class_init = kvm_accel_class_init,
2510     .instance_size = sizeof(KVMState),
2511 };
2512 
2513 static void kvm_type_init(void)
2514 {
2515     type_register_static(&kvm_accel_type);
2516 }
2517 
2518 type_init(kvm_type_init);
2519