xref: /openbmc/qemu/accel/kvm/kvm-all.c (revision 62dd4eda)
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;
201 
202     *start = section->offset_within_address_space;
203 
204     /* kvm works in page size chunks, but the function may be called
205        with sub-page size and unaligned start address. Pad the start
206        address to next and truncate size to previous page boundary. */
207     delta = qemu_real_host_page_size - (*start & ~qemu_real_host_page_mask);
208     delta &= ~qemu_real_host_page_mask;
209     *start += delta;
210     if (delta > size) {
211         return 0;
212     }
213     size -= delta;
214     size &= qemu_real_host_page_mask;
215     if (*start & ~qemu_real_host_page_mask) {
216         return 0;
217     }
218 
219     return size;
220 }
221 
222 int kvm_physical_memory_addr_from_host(KVMState *s, void *ram,
223                                        hwaddr *phys_addr)
224 {
225     KVMMemoryListener *kml = &s->memory_listener;
226     int i;
227 
228     for (i = 0; i < s->nr_slots; i++) {
229         KVMSlot *mem = &kml->slots[i];
230 
231         if (ram >= mem->ram && ram < mem->ram + mem->memory_size) {
232             *phys_addr = mem->start_addr + (ram - mem->ram);
233             return 1;
234         }
235     }
236 
237     return 0;
238 }
239 
240 static int kvm_set_user_memory_region(KVMMemoryListener *kml, KVMSlot *slot)
241 {
242     KVMState *s = kvm_state;
243     struct kvm_userspace_memory_region mem;
244 
245     mem.slot = slot->slot | (kml->as_id << 16);
246     mem.guest_phys_addr = slot->start_addr;
247     mem.userspace_addr = (unsigned long)slot->ram;
248     mem.flags = slot->flags;
249 
250     if (slot->memory_size && mem.flags & KVM_MEM_READONLY) {
251         /* Set the slot size to 0 before setting the slot to the desired
252          * value. This is needed based on KVM commit 75d61fbc. */
253         mem.memory_size = 0;
254         kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
255     }
256     mem.memory_size = slot->memory_size;
257     return kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
258 }
259 
260 int kvm_destroy_vcpu(CPUState *cpu)
261 {
262     KVMState *s = kvm_state;
263     long mmap_size;
264     struct KVMParkedVcpu *vcpu = NULL;
265     int ret = 0;
266 
267     DPRINTF("kvm_destroy_vcpu\n");
268 
269     mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
270     if (mmap_size < 0) {
271         ret = mmap_size;
272         DPRINTF("KVM_GET_VCPU_MMAP_SIZE failed\n");
273         goto err;
274     }
275 
276     ret = munmap(cpu->kvm_run, mmap_size);
277     if (ret < 0) {
278         goto err;
279     }
280 
281     vcpu = g_malloc0(sizeof(*vcpu));
282     vcpu->vcpu_id = kvm_arch_vcpu_id(cpu);
283     vcpu->kvm_fd = cpu->kvm_fd;
284     QLIST_INSERT_HEAD(&kvm_state->kvm_parked_vcpus, vcpu, node);
285 err:
286     return ret;
287 }
288 
289 static int kvm_get_vcpu(KVMState *s, unsigned long vcpu_id)
290 {
291     struct KVMParkedVcpu *cpu;
292 
293     QLIST_FOREACH(cpu, &s->kvm_parked_vcpus, node) {
294         if (cpu->vcpu_id == vcpu_id) {
295             int kvm_fd;
296 
297             QLIST_REMOVE(cpu, node);
298             kvm_fd = cpu->kvm_fd;
299             g_free(cpu);
300             return kvm_fd;
301         }
302     }
303 
304     return kvm_vm_ioctl(s, KVM_CREATE_VCPU, (void *)vcpu_id);
305 }
306 
307 int kvm_init_vcpu(CPUState *cpu)
308 {
309     KVMState *s = kvm_state;
310     long mmap_size;
311     int ret;
312 
313     DPRINTF("kvm_init_vcpu\n");
314 
315     ret = kvm_get_vcpu(s, kvm_arch_vcpu_id(cpu));
316     if (ret < 0) {
317         DPRINTF("kvm_create_vcpu failed\n");
318         goto err;
319     }
320 
321     cpu->kvm_fd = ret;
322     cpu->kvm_state = s;
323     cpu->vcpu_dirty = true;
324 
325     mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
326     if (mmap_size < 0) {
327         ret = mmap_size;
328         DPRINTF("KVM_GET_VCPU_MMAP_SIZE failed\n");
329         goto err;
330     }
331 
332     cpu->kvm_run = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
333                         cpu->kvm_fd, 0);
334     if (cpu->kvm_run == MAP_FAILED) {
335         ret = -errno;
336         DPRINTF("mmap'ing vcpu state failed\n");
337         goto err;
338     }
339 
340     if (s->coalesced_mmio && !s->coalesced_mmio_ring) {
341         s->coalesced_mmio_ring =
342             (void *)cpu->kvm_run + s->coalesced_mmio * PAGE_SIZE;
343     }
344 
345     ret = kvm_arch_init_vcpu(cpu);
346 err:
347     return ret;
348 }
349 
350 /*
351  * dirty pages logging control
352  */
353 
354 static int kvm_mem_flags(MemoryRegion *mr)
355 {
356     bool readonly = mr->readonly || memory_region_is_romd(mr);
357     int flags = 0;
358 
359     if (memory_region_get_dirty_log_mask(mr) != 0) {
360         flags |= KVM_MEM_LOG_DIRTY_PAGES;
361     }
362     if (readonly && kvm_readonly_mem_allowed) {
363         flags |= KVM_MEM_READONLY;
364     }
365     return flags;
366 }
367 
368 static int kvm_slot_update_flags(KVMMemoryListener *kml, KVMSlot *mem,
369                                  MemoryRegion *mr)
370 {
371     int old_flags;
372 
373     old_flags = mem->flags;
374     mem->flags = kvm_mem_flags(mr);
375 
376     /* If nothing changed effectively, no need to issue ioctl */
377     if (mem->flags == old_flags) {
378         return 0;
379     }
380 
381     return kvm_set_user_memory_region(kml, mem);
382 }
383 
384 static int kvm_section_update_flags(KVMMemoryListener *kml,
385                                     MemoryRegionSection *section)
386 {
387     hwaddr start_addr, size;
388     KVMSlot *mem;
389 
390     size = kvm_align_section(section, &start_addr);
391     if (!size) {
392         return 0;
393     }
394 
395     mem = kvm_lookup_matching_slot(kml, start_addr, size);
396     if (!mem) {
397         fprintf(stderr, "%s: error finding slot\n", __func__);
398         abort();
399     }
400 
401     return kvm_slot_update_flags(kml, mem, section->mr);
402 }
403 
404 static void kvm_log_start(MemoryListener *listener,
405                           MemoryRegionSection *section,
406                           int old, int new)
407 {
408     KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
409     int r;
410 
411     if (old != 0) {
412         return;
413     }
414 
415     r = kvm_section_update_flags(kml, section);
416     if (r < 0) {
417         abort();
418     }
419 }
420 
421 static void kvm_log_stop(MemoryListener *listener,
422                           MemoryRegionSection *section,
423                           int old, int new)
424 {
425     KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
426     int r;
427 
428     if (new != 0) {
429         return;
430     }
431 
432     r = kvm_section_update_flags(kml, section);
433     if (r < 0) {
434         abort();
435     }
436 }
437 
438 /* get kvm's dirty pages bitmap and update qemu's */
439 static int kvm_get_dirty_pages_log_range(MemoryRegionSection *section,
440                                          unsigned long *bitmap)
441 {
442     ram_addr_t start = section->offset_within_region +
443                        memory_region_get_ram_addr(section->mr);
444     ram_addr_t pages = int128_get64(section->size) / getpagesize();
445 
446     cpu_physical_memory_set_dirty_lebitmap(bitmap, start, pages);
447     return 0;
448 }
449 
450 #define ALIGN(x, y)  (((x)+(y)-1) & ~((y)-1))
451 
452 /**
453  * kvm_physical_sync_dirty_bitmap - Grab dirty bitmap from kernel space
454  * This function updates qemu's dirty bitmap using
455  * memory_region_set_dirty().  This means all bits are set
456  * to dirty.
457  *
458  * @start_add: start of logged region.
459  * @end_addr: end of logged region.
460  */
461 static int kvm_physical_sync_dirty_bitmap(KVMMemoryListener *kml,
462                                           MemoryRegionSection *section)
463 {
464     KVMState *s = kvm_state;
465     struct kvm_dirty_log d = {};
466     KVMSlot *mem;
467     hwaddr start_addr, size;
468 
469     size = kvm_align_section(section, &start_addr);
470     if (size) {
471         mem = kvm_lookup_matching_slot(kml, start_addr, size);
472         if (!mem) {
473             fprintf(stderr, "%s: error finding slot\n", __func__);
474             abort();
475         }
476 
477         /* XXX bad kernel interface alert
478          * For dirty bitmap, kernel allocates array of size aligned to
479          * bits-per-long.  But for case when the kernel is 64bits and
480          * the userspace is 32bits, userspace can't align to the same
481          * bits-per-long, since sizeof(long) is different between kernel
482          * and user space.  This way, userspace will provide buffer which
483          * may be 4 bytes less than the kernel will use, resulting in
484          * userspace memory corruption (which is not detectable by valgrind
485          * too, in most cases).
486          * So for now, let's align to 64 instead of HOST_LONG_BITS here, in
487          * a hope that sizeof(long) won't become >8 any time soon.
488          */
489         size = ALIGN(((mem->memory_size) >> TARGET_PAGE_BITS),
490                      /*HOST_LONG_BITS*/ 64) / 8;
491         d.dirty_bitmap = g_malloc0(size);
492 
493         d.slot = mem->slot | (kml->as_id << 16);
494         if (kvm_vm_ioctl(s, KVM_GET_DIRTY_LOG, &d) == -1) {
495             DPRINTF("ioctl failed %d\n", errno);
496             g_free(d.dirty_bitmap);
497             return -1;
498         }
499 
500         kvm_get_dirty_pages_log_range(section, d.dirty_bitmap);
501         g_free(d.dirty_bitmap);
502     }
503 
504     return 0;
505 }
506 
507 static void kvm_coalesce_mmio_region(MemoryListener *listener,
508                                      MemoryRegionSection *secion,
509                                      hwaddr start, hwaddr size)
510 {
511     KVMState *s = kvm_state;
512 
513     if (s->coalesced_mmio) {
514         struct kvm_coalesced_mmio_zone zone;
515 
516         zone.addr = start;
517         zone.size = size;
518         zone.pad = 0;
519 
520         (void)kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone);
521     }
522 }
523 
524 static void kvm_uncoalesce_mmio_region(MemoryListener *listener,
525                                        MemoryRegionSection *secion,
526                                        hwaddr start, hwaddr size)
527 {
528     KVMState *s = kvm_state;
529 
530     if (s->coalesced_mmio) {
531         struct kvm_coalesced_mmio_zone zone;
532 
533         zone.addr = start;
534         zone.size = size;
535         zone.pad = 0;
536 
537         (void)kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone);
538     }
539 }
540 
541 int kvm_check_extension(KVMState *s, unsigned int extension)
542 {
543     int ret;
544 
545     ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, extension);
546     if (ret < 0) {
547         ret = 0;
548     }
549 
550     return ret;
551 }
552 
553 int kvm_vm_check_extension(KVMState *s, unsigned int extension)
554 {
555     int ret;
556 
557     ret = kvm_vm_ioctl(s, KVM_CHECK_EXTENSION, extension);
558     if (ret < 0) {
559         /* VM wide version not implemented, use global one instead */
560         ret = kvm_check_extension(s, extension);
561     }
562 
563     return ret;
564 }
565 
566 static uint32_t adjust_ioeventfd_endianness(uint32_t val, uint32_t size)
567 {
568 #if defined(HOST_WORDS_BIGENDIAN) != defined(TARGET_WORDS_BIGENDIAN)
569     /* The kernel expects ioeventfd values in HOST_WORDS_BIGENDIAN
570      * endianness, but the memory core hands them in target endianness.
571      * For example, PPC is always treated as big-endian even if running
572      * on KVM and on PPC64LE.  Correct here.
573      */
574     switch (size) {
575     case 2:
576         val = bswap16(val);
577         break;
578     case 4:
579         val = bswap32(val);
580         break;
581     }
582 #endif
583     return val;
584 }
585 
586 static int kvm_set_ioeventfd_mmio(int fd, hwaddr addr, uint32_t val,
587                                   bool assign, uint32_t size, bool datamatch)
588 {
589     int ret;
590     struct kvm_ioeventfd iofd = {
591         .datamatch = datamatch ? adjust_ioeventfd_endianness(val, size) : 0,
592         .addr = addr,
593         .len = size,
594         .flags = 0,
595         .fd = fd,
596     };
597 
598     if (!kvm_enabled()) {
599         return -ENOSYS;
600     }
601 
602     if (datamatch) {
603         iofd.flags |= KVM_IOEVENTFD_FLAG_DATAMATCH;
604     }
605     if (!assign) {
606         iofd.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
607     }
608 
609     ret = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &iofd);
610 
611     if (ret < 0) {
612         return -errno;
613     }
614 
615     return 0;
616 }
617 
618 static int kvm_set_ioeventfd_pio(int fd, uint16_t addr, uint16_t val,
619                                  bool assign, uint32_t size, bool datamatch)
620 {
621     struct kvm_ioeventfd kick = {
622         .datamatch = datamatch ? adjust_ioeventfd_endianness(val, size) : 0,
623         .addr = addr,
624         .flags = KVM_IOEVENTFD_FLAG_PIO,
625         .len = size,
626         .fd = fd,
627     };
628     int r;
629     if (!kvm_enabled()) {
630         return -ENOSYS;
631     }
632     if (datamatch) {
633         kick.flags |= KVM_IOEVENTFD_FLAG_DATAMATCH;
634     }
635     if (!assign) {
636         kick.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
637     }
638     r = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &kick);
639     if (r < 0) {
640         return r;
641     }
642     return 0;
643 }
644 
645 
646 static int kvm_check_many_ioeventfds(void)
647 {
648     /* Userspace can use ioeventfd for io notification.  This requires a host
649      * that supports eventfd(2) and an I/O thread; since eventfd does not
650      * support SIGIO it cannot interrupt the vcpu.
651      *
652      * Older kernels have a 6 device limit on the KVM io bus.  Find out so we
653      * can avoid creating too many ioeventfds.
654      */
655 #if defined(CONFIG_EVENTFD)
656     int ioeventfds[7];
657     int i, ret = 0;
658     for (i = 0; i < ARRAY_SIZE(ioeventfds); i++) {
659         ioeventfds[i] = eventfd(0, EFD_CLOEXEC);
660         if (ioeventfds[i] < 0) {
661             break;
662         }
663         ret = kvm_set_ioeventfd_pio(ioeventfds[i], 0, i, true, 2, true);
664         if (ret < 0) {
665             close(ioeventfds[i]);
666             break;
667         }
668     }
669 
670     /* Decide whether many devices are supported or not */
671     ret = i == ARRAY_SIZE(ioeventfds);
672 
673     while (i-- > 0) {
674         kvm_set_ioeventfd_pio(ioeventfds[i], 0, i, false, 2, true);
675         close(ioeventfds[i]);
676     }
677     return ret;
678 #else
679     return 0;
680 #endif
681 }
682 
683 static const KVMCapabilityInfo *
684 kvm_check_extension_list(KVMState *s, const KVMCapabilityInfo *list)
685 {
686     while (list->name) {
687         if (!kvm_check_extension(s, list->value)) {
688             return list;
689         }
690         list++;
691     }
692     return NULL;
693 }
694 
695 static void kvm_set_phys_mem(KVMMemoryListener *kml,
696                              MemoryRegionSection *section, bool add)
697 {
698     KVMSlot *mem;
699     int err;
700     MemoryRegion *mr = section->mr;
701     bool writeable = !mr->readonly && !mr->rom_device;
702     hwaddr start_addr, size;
703     void *ram;
704 
705     if (!memory_region_is_ram(mr)) {
706         if (writeable || !kvm_readonly_mem_allowed) {
707             return;
708         } else if (!mr->romd_mode) {
709             /* If the memory device is not in romd_mode, then we actually want
710              * to remove the kvm memory slot so all accesses will trap. */
711             add = false;
712         }
713     }
714 
715     size = kvm_align_section(section, &start_addr);
716     if (!size) {
717         return;
718     }
719 
720     ram = memory_region_get_ram_ptr(mr) + section->offset_within_region +
721           (section->offset_within_address_space - start_addr);
722 
723     mem = kvm_lookup_matching_slot(kml, start_addr, size);
724     if (!add) {
725         if (!mem) {
726             return;
727         }
728         if (mem->flags & KVM_MEM_LOG_DIRTY_PAGES) {
729             kvm_physical_sync_dirty_bitmap(kml, section);
730         }
731 
732         /* unregister the slot */
733         mem->memory_size = 0;
734         err = kvm_set_user_memory_region(kml, mem);
735         if (err) {
736             fprintf(stderr, "%s: error unregistering overlapping slot: %s\n",
737                     __func__, strerror(-err));
738             abort();
739         }
740         return;
741     }
742 
743     if (mem) {
744         /* update the slot */
745         kvm_slot_update_flags(kml, mem, mr);
746         return;
747     }
748 
749     /* register the new slot */
750     mem = kvm_alloc_slot(kml);
751     mem->memory_size = size;
752     mem->start_addr = start_addr;
753     mem->ram = ram;
754     mem->flags = kvm_mem_flags(mr);
755 
756     err = kvm_set_user_memory_region(kml, mem);
757     if (err) {
758         fprintf(stderr, "%s: error registering slot: %s\n", __func__,
759                 strerror(-err));
760         abort();
761     }
762 }
763 
764 static void kvm_region_add(MemoryListener *listener,
765                            MemoryRegionSection *section)
766 {
767     KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
768 
769     memory_region_ref(section->mr);
770     kvm_set_phys_mem(kml, section, true);
771 }
772 
773 static void kvm_region_del(MemoryListener *listener,
774                            MemoryRegionSection *section)
775 {
776     KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
777 
778     kvm_set_phys_mem(kml, section, false);
779     memory_region_unref(section->mr);
780 }
781 
782 static void kvm_log_sync(MemoryListener *listener,
783                          MemoryRegionSection *section)
784 {
785     KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
786     int r;
787 
788     r = kvm_physical_sync_dirty_bitmap(kml, section);
789     if (r < 0) {
790         abort();
791     }
792 }
793 
794 static void kvm_mem_ioeventfd_add(MemoryListener *listener,
795                                   MemoryRegionSection *section,
796                                   bool match_data, uint64_t data,
797                                   EventNotifier *e)
798 {
799     int fd = event_notifier_get_fd(e);
800     int r;
801 
802     r = kvm_set_ioeventfd_mmio(fd, section->offset_within_address_space,
803                                data, true, int128_get64(section->size),
804                                match_data);
805     if (r < 0) {
806         fprintf(stderr, "%s: error adding ioeventfd: %s\n",
807                 __func__, strerror(-r));
808         abort();
809     }
810 }
811 
812 static void kvm_mem_ioeventfd_del(MemoryListener *listener,
813                                   MemoryRegionSection *section,
814                                   bool match_data, uint64_t data,
815                                   EventNotifier *e)
816 {
817     int fd = event_notifier_get_fd(e);
818     int r;
819 
820     r = kvm_set_ioeventfd_mmio(fd, section->offset_within_address_space,
821                                data, false, int128_get64(section->size),
822                                match_data);
823     if (r < 0) {
824         abort();
825     }
826 }
827 
828 static void kvm_io_ioeventfd_add(MemoryListener *listener,
829                                  MemoryRegionSection *section,
830                                  bool match_data, uint64_t data,
831                                  EventNotifier *e)
832 {
833     int fd = event_notifier_get_fd(e);
834     int r;
835 
836     r = kvm_set_ioeventfd_pio(fd, section->offset_within_address_space,
837                               data, true, int128_get64(section->size),
838                               match_data);
839     if (r < 0) {
840         fprintf(stderr, "%s: error adding ioeventfd: %s\n",
841                 __func__, strerror(-r));
842         abort();
843     }
844 }
845 
846 static void kvm_io_ioeventfd_del(MemoryListener *listener,
847                                  MemoryRegionSection *section,
848                                  bool match_data, uint64_t data,
849                                  EventNotifier *e)
850 
851 {
852     int fd = event_notifier_get_fd(e);
853     int r;
854 
855     r = kvm_set_ioeventfd_pio(fd, section->offset_within_address_space,
856                               data, false, int128_get64(section->size),
857                               match_data);
858     if (r < 0) {
859         abort();
860     }
861 }
862 
863 void kvm_memory_listener_register(KVMState *s, KVMMemoryListener *kml,
864                                   AddressSpace *as, int as_id)
865 {
866     int i;
867 
868     kml->slots = g_malloc0(s->nr_slots * sizeof(KVMSlot));
869     kml->as_id = as_id;
870 
871     for (i = 0; i < s->nr_slots; i++) {
872         kml->slots[i].slot = i;
873     }
874 
875     kml->listener.region_add = kvm_region_add;
876     kml->listener.region_del = kvm_region_del;
877     kml->listener.log_start = kvm_log_start;
878     kml->listener.log_stop = kvm_log_stop;
879     kml->listener.log_sync = kvm_log_sync;
880     kml->listener.priority = 10;
881 
882     memory_listener_register(&kml->listener, as);
883 }
884 
885 static MemoryListener kvm_io_listener = {
886     .eventfd_add = kvm_io_ioeventfd_add,
887     .eventfd_del = kvm_io_ioeventfd_del,
888     .priority = 10,
889 };
890 
891 int kvm_set_irq(KVMState *s, int irq, int level)
892 {
893     struct kvm_irq_level event;
894     int ret;
895 
896     assert(kvm_async_interrupts_enabled());
897 
898     event.level = level;
899     event.irq = irq;
900     ret = kvm_vm_ioctl(s, s->irq_set_ioctl, &event);
901     if (ret < 0) {
902         perror("kvm_set_irq");
903         abort();
904     }
905 
906     return (s->irq_set_ioctl == KVM_IRQ_LINE) ? 1 : event.status;
907 }
908 
909 #ifdef KVM_CAP_IRQ_ROUTING
910 typedef struct KVMMSIRoute {
911     struct kvm_irq_routing_entry kroute;
912     QTAILQ_ENTRY(KVMMSIRoute) entry;
913 } KVMMSIRoute;
914 
915 static void set_gsi(KVMState *s, unsigned int gsi)
916 {
917     set_bit(gsi, s->used_gsi_bitmap);
918 }
919 
920 static void clear_gsi(KVMState *s, unsigned int gsi)
921 {
922     clear_bit(gsi, s->used_gsi_bitmap);
923 }
924 
925 void kvm_init_irq_routing(KVMState *s)
926 {
927     int gsi_count, i;
928 
929     gsi_count = kvm_check_extension(s, KVM_CAP_IRQ_ROUTING) - 1;
930     if (gsi_count > 0) {
931         /* Round up so we can search ints using ffs */
932         s->used_gsi_bitmap = bitmap_new(gsi_count);
933         s->gsi_count = gsi_count;
934     }
935 
936     s->irq_routes = g_malloc0(sizeof(*s->irq_routes));
937     s->nr_allocated_irq_routes = 0;
938 
939     if (!kvm_direct_msi_allowed) {
940         for (i = 0; i < KVM_MSI_HASHTAB_SIZE; i++) {
941             QTAILQ_INIT(&s->msi_hashtab[i]);
942         }
943     }
944 
945     kvm_arch_init_irq_routing(s);
946 }
947 
948 void kvm_irqchip_commit_routes(KVMState *s)
949 {
950     int ret;
951 
952     if (kvm_gsi_direct_mapping()) {
953         return;
954     }
955 
956     if (!kvm_gsi_routing_enabled()) {
957         return;
958     }
959 
960     s->irq_routes->flags = 0;
961     trace_kvm_irqchip_commit_routes();
962     ret = kvm_vm_ioctl(s, KVM_SET_GSI_ROUTING, s->irq_routes);
963     assert(ret == 0);
964 }
965 
966 static void kvm_add_routing_entry(KVMState *s,
967                                   struct kvm_irq_routing_entry *entry)
968 {
969     struct kvm_irq_routing_entry *new;
970     int n, size;
971 
972     if (s->irq_routes->nr == s->nr_allocated_irq_routes) {
973         n = s->nr_allocated_irq_routes * 2;
974         if (n < 64) {
975             n = 64;
976         }
977         size = sizeof(struct kvm_irq_routing);
978         size += n * sizeof(*new);
979         s->irq_routes = g_realloc(s->irq_routes, size);
980         s->nr_allocated_irq_routes = n;
981     }
982     n = s->irq_routes->nr++;
983     new = &s->irq_routes->entries[n];
984 
985     *new = *entry;
986 
987     set_gsi(s, entry->gsi);
988 }
989 
990 static int kvm_update_routing_entry(KVMState *s,
991                                     struct kvm_irq_routing_entry *new_entry)
992 {
993     struct kvm_irq_routing_entry *entry;
994     int n;
995 
996     for (n = 0; n < s->irq_routes->nr; n++) {
997         entry = &s->irq_routes->entries[n];
998         if (entry->gsi != new_entry->gsi) {
999             continue;
1000         }
1001 
1002         if(!memcmp(entry, new_entry, sizeof *entry)) {
1003             return 0;
1004         }
1005 
1006         *entry = *new_entry;
1007 
1008         return 0;
1009     }
1010 
1011     return -ESRCH;
1012 }
1013 
1014 void kvm_irqchip_add_irq_route(KVMState *s, int irq, int irqchip, int pin)
1015 {
1016     struct kvm_irq_routing_entry e = {};
1017 
1018     assert(pin < s->gsi_count);
1019 
1020     e.gsi = irq;
1021     e.type = KVM_IRQ_ROUTING_IRQCHIP;
1022     e.flags = 0;
1023     e.u.irqchip.irqchip = irqchip;
1024     e.u.irqchip.pin = pin;
1025     kvm_add_routing_entry(s, &e);
1026 }
1027 
1028 void kvm_irqchip_release_virq(KVMState *s, int virq)
1029 {
1030     struct kvm_irq_routing_entry *e;
1031     int i;
1032 
1033     if (kvm_gsi_direct_mapping()) {
1034         return;
1035     }
1036 
1037     for (i = 0; i < s->irq_routes->nr; i++) {
1038         e = &s->irq_routes->entries[i];
1039         if (e->gsi == virq) {
1040             s->irq_routes->nr--;
1041             *e = s->irq_routes->entries[s->irq_routes->nr];
1042         }
1043     }
1044     clear_gsi(s, virq);
1045     kvm_arch_release_virq_post(virq);
1046     trace_kvm_irqchip_release_virq(virq);
1047 }
1048 
1049 static unsigned int kvm_hash_msi(uint32_t data)
1050 {
1051     /* This is optimized for IA32 MSI layout. However, no other arch shall
1052      * repeat the mistake of not providing a direct MSI injection API. */
1053     return data & 0xff;
1054 }
1055 
1056 static void kvm_flush_dynamic_msi_routes(KVMState *s)
1057 {
1058     KVMMSIRoute *route, *next;
1059     unsigned int hash;
1060 
1061     for (hash = 0; hash < KVM_MSI_HASHTAB_SIZE; hash++) {
1062         QTAILQ_FOREACH_SAFE(route, &s->msi_hashtab[hash], entry, next) {
1063             kvm_irqchip_release_virq(s, route->kroute.gsi);
1064             QTAILQ_REMOVE(&s->msi_hashtab[hash], route, entry);
1065             g_free(route);
1066         }
1067     }
1068 }
1069 
1070 static int kvm_irqchip_get_virq(KVMState *s)
1071 {
1072     int next_virq;
1073 
1074     /*
1075      * PIC and IOAPIC share the first 16 GSI numbers, thus the available
1076      * GSI numbers are more than the number of IRQ route. Allocating a GSI
1077      * number can succeed even though a new route entry cannot be added.
1078      * When this happens, flush dynamic MSI entries to free IRQ route entries.
1079      */
1080     if (!kvm_direct_msi_allowed && s->irq_routes->nr == s->gsi_count) {
1081         kvm_flush_dynamic_msi_routes(s);
1082     }
1083 
1084     /* Return the lowest unused GSI in the bitmap */
1085     next_virq = find_first_zero_bit(s->used_gsi_bitmap, s->gsi_count);
1086     if (next_virq >= s->gsi_count) {
1087         return -ENOSPC;
1088     } else {
1089         return next_virq;
1090     }
1091 }
1092 
1093 static KVMMSIRoute *kvm_lookup_msi_route(KVMState *s, MSIMessage msg)
1094 {
1095     unsigned int hash = kvm_hash_msi(msg.data);
1096     KVMMSIRoute *route;
1097 
1098     QTAILQ_FOREACH(route, &s->msi_hashtab[hash], entry) {
1099         if (route->kroute.u.msi.address_lo == (uint32_t)msg.address &&
1100             route->kroute.u.msi.address_hi == (msg.address >> 32) &&
1101             route->kroute.u.msi.data == le32_to_cpu(msg.data)) {
1102             return route;
1103         }
1104     }
1105     return NULL;
1106 }
1107 
1108 int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg)
1109 {
1110     struct kvm_msi msi;
1111     KVMMSIRoute *route;
1112 
1113     if (kvm_direct_msi_allowed) {
1114         msi.address_lo = (uint32_t)msg.address;
1115         msi.address_hi = msg.address >> 32;
1116         msi.data = le32_to_cpu(msg.data);
1117         msi.flags = 0;
1118         memset(msi.pad, 0, sizeof(msi.pad));
1119 
1120         return kvm_vm_ioctl(s, KVM_SIGNAL_MSI, &msi);
1121     }
1122 
1123     route = kvm_lookup_msi_route(s, msg);
1124     if (!route) {
1125         int virq;
1126 
1127         virq = kvm_irqchip_get_virq(s);
1128         if (virq < 0) {
1129             return virq;
1130         }
1131 
1132         route = g_malloc0(sizeof(KVMMSIRoute));
1133         route->kroute.gsi = virq;
1134         route->kroute.type = KVM_IRQ_ROUTING_MSI;
1135         route->kroute.flags = 0;
1136         route->kroute.u.msi.address_lo = (uint32_t)msg.address;
1137         route->kroute.u.msi.address_hi = msg.address >> 32;
1138         route->kroute.u.msi.data = le32_to_cpu(msg.data);
1139 
1140         kvm_add_routing_entry(s, &route->kroute);
1141         kvm_irqchip_commit_routes(s);
1142 
1143         QTAILQ_INSERT_TAIL(&s->msi_hashtab[kvm_hash_msi(msg.data)], route,
1144                            entry);
1145     }
1146 
1147     assert(route->kroute.type == KVM_IRQ_ROUTING_MSI);
1148 
1149     return kvm_set_irq(s, route->kroute.gsi, 1);
1150 }
1151 
1152 int kvm_irqchip_add_msi_route(KVMState *s, int vector, PCIDevice *dev)
1153 {
1154     struct kvm_irq_routing_entry kroute = {};
1155     int virq;
1156     MSIMessage msg = {0, 0};
1157 
1158     if (pci_available && dev) {
1159         msg = pci_get_msi_message(dev, vector);
1160     }
1161 
1162     if (kvm_gsi_direct_mapping()) {
1163         return kvm_arch_msi_data_to_gsi(msg.data);
1164     }
1165 
1166     if (!kvm_gsi_routing_enabled()) {
1167         return -ENOSYS;
1168     }
1169 
1170     virq = kvm_irqchip_get_virq(s);
1171     if (virq < 0) {
1172         return virq;
1173     }
1174 
1175     kroute.gsi = virq;
1176     kroute.type = KVM_IRQ_ROUTING_MSI;
1177     kroute.flags = 0;
1178     kroute.u.msi.address_lo = (uint32_t)msg.address;
1179     kroute.u.msi.address_hi = msg.address >> 32;
1180     kroute.u.msi.data = le32_to_cpu(msg.data);
1181     if (pci_available && kvm_msi_devid_required()) {
1182         kroute.flags = KVM_MSI_VALID_DEVID;
1183         kroute.u.msi.devid = pci_requester_id(dev);
1184     }
1185     if (kvm_arch_fixup_msi_route(&kroute, msg.address, msg.data, dev)) {
1186         kvm_irqchip_release_virq(s, virq);
1187         return -EINVAL;
1188     }
1189 
1190     trace_kvm_irqchip_add_msi_route(dev ? dev->name : (char *)"N/A",
1191                                     vector, virq);
1192 
1193     kvm_add_routing_entry(s, &kroute);
1194     kvm_arch_add_msi_route_post(&kroute, vector, dev);
1195     kvm_irqchip_commit_routes(s);
1196 
1197     return virq;
1198 }
1199 
1200 int kvm_irqchip_update_msi_route(KVMState *s, int virq, MSIMessage msg,
1201                                  PCIDevice *dev)
1202 {
1203     struct kvm_irq_routing_entry kroute = {};
1204 
1205     if (kvm_gsi_direct_mapping()) {
1206         return 0;
1207     }
1208 
1209     if (!kvm_irqchip_in_kernel()) {
1210         return -ENOSYS;
1211     }
1212 
1213     kroute.gsi = virq;
1214     kroute.type = KVM_IRQ_ROUTING_MSI;
1215     kroute.flags = 0;
1216     kroute.u.msi.address_lo = (uint32_t)msg.address;
1217     kroute.u.msi.address_hi = msg.address >> 32;
1218     kroute.u.msi.data = le32_to_cpu(msg.data);
1219     if (pci_available && kvm_msi_devid_required()) {
1220         kroute.flags = KVM_MSI_VALID_DEVID;
1221         kroute.u.msi.devid = pci_requester_id(dev);
1222     }
1223     if (kvm_arch_fixup_msi_route(&kroute, msg.address, msg.data, dev)) {
1224         return -EINVAL;
1225     }
1226 
1227     trace_kvm_irqchip_update_msi_route(virq);
1228 
1229     return kvm_update_routing_entry(s, &kroute);
1230 }
1231 
1232 static int kvm_irqchip_assign_irqfd(KVMState *s, int fd, int rfd, int virq,
1233                                     bool assign)
1234 {
1235     struct kvm_irqfd irqfd = {
1236         .fd = fd,
1237         .gsi = virq,
1238         .flags = assign ? 0 : KVM_IRQFD_FLAG_DEASSIGN,
1239     };
1240 
1241     if (rfd != -1) {
1242         irqfd.flags |= KVM_IRQFD_FLAG_RESAMPLE;
1243         irqfd.resamplefd = rfd;
1244     }
1245 
1246     if (!kvm_irqfds_enabled()) {
1247         return -ENOSYS;
1248     }
1249 
1250     return kvm_vm_ioctl(s, KVM_IRQFD, &irqfd);
1251 }
1252 
1253 int kvm_irqchip_add_adapter_route(KVMState *s, AdapterInfo *adapter)
1254 {
1255     struct kvm_irq_routing_entry kroute = {};
1256     int virq;
1257 
1258     if (!kvm_gsi_routing_enabled()) {
1259         return -ENOSYS;
1260     }
1261 
1262     virq = kvm_irqchip_get_virq(s);
1263     if (virq < 0) {
1264         return virq;
1265     }
1266 
1267     kroute.gsi = virq;
1268     kroute.type = KVM_IRQ_ROUTING_S390_ADAPTER;
1269     kroute.flags = 0;
1270     kroute.u.adapter.summary_addr = adapter->summary_addr;
1271     kroute.u.adapter.ind_addr = adapter->ind_addr;
1272     kroute.u.adapter.summary_offset = adapter->summary_offset;
1273     kroute.u.adapter.ind_offset = adapter->ind_offset;
1274     kroute.u.adapter.adapter_id = adapter->adapter_id;
1275 
1276     kvm_add_routing_entry(s, &kroute);
1277 
1278     return virq;
1279 }
1280 
1281 int kvm_irqchip_add_hv_sint_route(KVMState *s, uint32_t vcpu, uint32_t sint)
1282 {
1283     struct kvm_irq_routing_entry kroute = {};
1284     int virq;
1285 
1286     if (!kvm_gsi_routing_enabled()) {
1287         return -ENOSYS;
1288     }
1289     if (!kvm_check_extension(s, KVM_CAP_HYPERV_SYNIC)) {
1290         return -ENOSYS;
1291     }
1292     virq = kvm_irqchip_get_virq(s);
1293     if (virq < 0) {
1294         return virq;
1295     }
1296 
1297     kroute.gsi = virq;
1298     kroute.type = KVM_IRQ_ROUTING_HV_SINT;
1299     kroute.flags = 0;
1300     kroute.u.hv_sint.vcpu = vcpu;
1301     kroute.u.hv_sint.sint = sint;
1302 
1303     kvm_add_routing_entry(s, &kroute);
1304     kvm_irqchip_commit_routes(s);
1305 
1306     return virq;
1307 }
1308 
1309 #else /* !KVM_CAP_IRQ_ROUTING */
1310 
1311 void kvm_init_irq_routing(KVMState *s)
1312 {
1313 }
1314 
1315 void kvm_irqchip_release_virq(KVMState *s, int virq)
1316 {
1317 }
1318 
1319 int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg)
1320 {
1321     abort();
1322 }
1323 
1324 int kvm_irqchip_add_msi_route(KVMState *s, int vector, PCIDevice *dev)
1325 {
1326     return -ENOSYS;
1327 }
1328 
1329 int kvm_irqchip_add_adapter_route(KVMState *s, AdapterInfo *adapter)
1330 {
1331     return -ENOSYS;
1332 }
1333 
1334 int kvm_irqchip_add_hv_sint_route(KVMState *s, uint32_t vcpu, uint32_t sint)
1335 {
1336     return -ENOSYS;
1337 }
1338 
1339 static int kvm_irqchip_assign_irqfd(KVMState *s, int fd, int virq, bool assign)
1340 {
1341     abort();
1342 }
1343 
1344 int kvm_irqchip_update_msi_route(KVMState *s, int virq, MSIMessage msg)
1345 {
1346     return -ENOSYS;
1347 }
1348 #endif /* !KVM_CAP_IRQ_ROUTING */
1349 
1350 int kvm_irqchip_add_irqfd_notifier_gsi(KVMState *s, EventNotifier *n,
1351                                        EventNotifier *rn, int virq)
1352 {
1353     return kvm_irqchip_assign_irqfd(s, event_notifier_get_fd(n),
1354            rn ? event_notifier_get_fd(rn) : -1, virq, true);
1355 }
1356 
1357 int kvm_irqchip_remove_irqfd_notifier_gsi(KVMState *s, EventNotifier *n,
1358                                           int virq)
1359 {
1360     return kvm_irqchip_assign_irqfd(s, event_notifier_get_fd(n), -1, virq,
1361            false);
1362 }
1363 
1364 int kvm_irqchip_add_irqfd_notifier(KVMState *s, EventNotifier *n,
1365                                    EventNotifier *rn, qemu_irq irq)
1366 {
1367     gpointer key, gsi;
1368     gboolean found = g_hash_table_lookup_extended(s->gsimap, irq, &key, &gsi);
1369 
1370     if (!found) {
1371         return -ENXIO;
1372     }
1373     return kvm_irqchip_add_irqfd_notifier_gsi(s, n, rn, GPOINTER_TO_INT(gsi));
1374 }
1375 
1376 int kvm_irqchip_remove_irqfd_notifier(KVMState *s, EventNotifier *n,
1377                                       qemu_irq irq)
1378 {
1379     gpointer key, gsi;
1380     gboolean found = g_hash_table_lookup_extended(s->gsimap, irq, &key, &gsi);
1381 
1382     if (!found) {
1383         return -ENXIO;
1384     }
1385     return kvm_irqchip_remove_irqfd_notifier_gsi(s, n, GPOINTER_TO_INT(gsi));
1386 }
1387 
1388 void kvm_irqchip_set_qemuirq_gsi(KVMState *s, qemu_irq irq, int gsi)
1389 {
1390     g_hash_table_insert(s->gsimap, irq, GINT_TO_POINTER(gsi));
1391 }
1392 
1393 static void kvm_irqchip_create(MachineState *machine, KVMState *s)
1394 {
1395     int ret;
1396 
1397     if (kvm_check_extension(s, KVM_CAP_IRQCHIP)) {
1398         ;
1399     } else if (kvm_check_extension(s, KVM_CAP_S390_IRQCHIP)) {
1400         ret = kvm_vm_enable_cap(s, KVM_CAP_S390_IRQCHIP, 0);
1401         if (ret < 0) {
1402             fprintf(stderr, "Enable kernel irqchip failed: %s\n", strerror(-ret));
1403             exit(1);
1404         }
1405     } else {
1406         return;
1407     }
1408 
1409     /* First probe and see if there's a arch-specific hook to create the
1410      * in-kernel irqchip for us */
1411     ret = kvm_arch_irqchip_create(machine, s);
1412     if (ret == 0) {
1413         if (machine_kernel_irqchip_split(machine)) {
1414             perror("Split IRQ chip mode not supported.");
1415             exit(1);
1416         } else {
1417             ret = kvm_vm_ioctl(s, KVM_CREATE_IRQCHIP);
1418         }
1419     }
1420     if (ret < 0) {
1421         fprintf(stderr, "Create kernel irqchip failed: %s\n", strerror(-ret));
1422         exit(1);
1423     }
1424 
1425     kvm_kernel_irqchip = true;
1426     /* If we have an in-kernel IRQ chip then we must have asynchronous
1427      * interrupt delivery (though the reverse is not necessarily true)
1428      */
1429     kvm_async_interrupts_allowed = true;
1430     kvm_halt_in_kernel_allowed = true;
1431 
1432     kvm_init_irq_routing(s);
1433 
1434     s->gsimap = g_hash_table_new(g_direct_hash, g_direct_equal);
1435 }
1436 
1437 /* Find number of supported CPUs using the recommended
1438  * procedure from the kernel API documentation to cope with
1439  * older kernels that may be missing capabilities.
1440  */
1441 static int kvm_recommended_vcpus(KVMState *s)
1442 {
1443     int ret = kvm_check_extension(s, KVM_CAP_NR_VCPUS);
1444     return (ret) ? ret : 4;
1445 }
1446 
1447 static int kvm_max_vcpus(KVMState *s)
1448 {
1449     int ret = kvm_check_extension(s, KVM_CAP_MAX_VCPUS);
1450     return (ret) ? ret : kvm_recommended_vcpus(s);
1451 }
1452 
1453 static int kvm_max_vcpu_id(KVMState *s)
1454 {
1455     int ret = kvm_check_extension(s, KVM_CAP_MAX_VCPU_ID);
1456     return (ret) ? ret : kvm_max_vcpus(s);
1457 }
1458 
1459 bool kvm_vcpu_id_is_valid(int vcpu_id)
1460 {
1461     KVMState *s = KVM_STATE(current_machine->accelerator);
1462     return vcpu_id >= 0 && vcpu_id < kvm_max_vcpu_id(s);
1463 }
1464 
1465 static int kvm_init(MachineState *ms)
1466 {
1467     MachineClass *mc = MACHINE_GET_CLASS(ms);
1468     static const char upgrade_note[] =
1469         "Please upgrade to at least kernel 2.6.29 or recent kvm-kmod\n"
1470         "(see http://sourceforge.net/projects/kvm).\n";
1471     struct {
1472         const char *name;
1473         int num;
1474     } num_cpus[] = {
1475         { "SMP",          smp_cpus },
1476         { "hotpluggable", max_cpus },
1477         { NULL, }
1478     }, *nc = num_cpus;
1479     int soft_vcpus_limit, hard_vcpus_limit;
1480     KVMState *s;
1481     const KVMCapabilityInfo *missing_cap;
1482     int ret;
1483     int type = 0;
1484     const char *kvm_type;
1485 
1486     s = KVM_STATE(ms->accelerator);
1487 
1488     /*
1489      * On systems where the kernel can support different base page
1490      * sizes, host page size may be different from TARGET_PAGE_SIZE,
1491      * even with KVM.  TARGET_PAGE_SIZE is assumed to be the minimum
1492      * page size for the system though.
1493      */
1494     assert(TARGET_PAGE_SIZE <= getpagesize());
1495 
1496     s->sigmask_len = 8;
1497 
1498 #ifdef KVM_CAP_SET_GUEST_DEBUG
1499     QTAILQ_INIT(&s->kvm_sw_breakpoints);
1500 #endif
1501     QLIST_INIT(&s->kvm_parked_vcpus);
1502     s->vmfd = -1;
1503     s->fd = qemu_open("/dev/kvm", O_RDWR);
1504     if (s->fd == -1) {
1505         fprintf(stderr, "Could not access KVM kernel module: %m\n");
1506         ret = -errno;
1507         goto err;
1508     }
1509 
1510     ret = kvm_ioctl(s, KVM_GET_API_VERSION, 0);
1511     if (ret < KVM_API_VERSION) {
1512         if (ret >= 0) {
1513             ret = -EINVAL;
1514         }
1515         fprintf(stderr, "kvm version too old\n");
1516         goto err;
1517     }
1518 
1519     if (ret > KVM_API_VERSION) {
1520         ret = -EINVAL;
1521         fprintf(stderr, "kvm version not supported\n");
1522         goto err;
1523     }
1524 
1525     kvm_immediate_exit = kvm_check_extension(s, KVM_CAP_IMMEDIATE_EXIT);
1526     s->nr_slots = kvm_check_extension(s, KVM_CAP_NR_MEMSLOTS);
1527 
1528     /* If unspecified, use the default value */
1529     if (!s->nr_slots) {
1530         s->nr_slots = 32;
1531     }
1532 
1533     /* check the vcpu limits */
1534     soft_vcpus_limit = kvm_recommended_vcpus(s);
1535     hard_vcpus_limit = kvm_max_vcpus(s);
1536 
1537     while (nc->name) {
1538         if (nc->num > soft_vcpus_limit) {
1539             warn_report("Number of %s cpus requested (%d) exceeds "
1540                         "the recommended cpus supported by KVM (%d)",
1541                         nc->name, nc->num, soft_vcpus_limit);
1542 
1543             if (nc->num > hard_vcpus_limit) {
1544                 fprintf(stderr, "Number of %s cpus requested (%d) exceeds "
1545                         "the maximum cpus supported by KVM (%d)\n",
1546                         nc->name, nc->num, hard_vcpus_limit);
1547                 exit(1);
1548             }
1549         }
1550         nc++;
1551     }
1552 
1553     kvm_type = qemu_opt_get(qemu_get_machine_opts(), "kvm-type");
1554     if (mc->kvm_type) {
1555         type = mc->kvm_type(kvm_type);
1556     } else if (kvm_type) {
1557         ret = -EINVAL;
1558         fprintf(stderr, "Invalid argument kvm-type=%s\n", kvm_type);
1559         goto err;
1560     }
1561 
1562     do {
1563         ret = kvm_ioctl(s, KVM_CREATE_VM, type);
1564     } while (ret == -EINTR);
1565 
1566     if (ret < 0) {
1567         fprintf(stderr, "ioctl(KVM_CREATE_VM) failed: %d %s\n", -ret,
1568                 strerror(-ret));
1569 
1570 #ifdef TARGET_S390X
1571         if (ret == -EINVAL) {
1572             fprintf(stderr,
1573                     "Host kernel setup problem detected. Please verify:\n");
1574             fprintf(stderr, "- for kernels supporting the switch_amode or"
1575                     " user_mode parameters, whether\n");
1576             fprintf(stderr,
1577                     "  user space is running in primary address space\n");
1578             fprintf(stderr,
1579                     "- for kernels supporting the vm.allocate_pgste sysctl, "
1580                     "whether it is enabled\n");
1581         }
1582 #endif
1583         goto err;
1584     }
1585 
1586     s->vmfd = ret;
1587     missing_cap = kvm_check_extension_list(s, kvm_required_capabilites);
1588     if (!missing_cap) {
1589         missing_cap =
1590             kvm_check_extension_list(s, kvm_arch_required_capabilities);
1591     }
1592     if (missing_cap) {
1593         ret = -EINVAL;
1594         fprintf(stderr, "kvm does not support %s\n%s",
1595                 missing_cap->name, upgrade_note);
1596         goto err;
1597     }
1598 
1599     s->coalesced_mmio = kvm_check_extension(s, KVM_CAP_COALESCED_MMIO);
1600 
1601 #ifdef KVM_CAP_VCPU_EVENTS
1602     s->vcpu_events = kvm_check_extension(s, KVM_CAP_VCPU_EVENTS);
1603 #endif
1604 
1605     s->robust_singlestep =
1606         kvm_check_extension(s, KVM_CAP_X86_ROBUST_SINGLESTEP);
1607 
1608 #ifdef KVM_CAP_DEBUGREGS
1609     s->debugregs = kvm_check_extension(s, KVM_CAP_DEBUGREGS);
1610 #endif
1611 
1612 #ifdef KVM_CAP_IRQ_ROUTING
1613     kvm_direct_msi_allowed = (kvm_check_extension(s, KVM_CAP_SIGNAL_MSI) > 0);
1614 #endif
1615 
1616     s->intx_set_mask = kvm_check_extension(s, KVM_CAP_PCI_2_3);
1617 
1618     s->irq_set_ioctl = KVM_IRQ_LINE;
1619     if (kvm_check_extension(s, KVM_CAP_IRQ_INJECT_STATUS)) {
1620         s->irq_set_ioctl = KVM_IRQ_LINE_STATUS;
1621     }
1622 
1623 #ifdef KVM_CAP_READONLY_MEM
1624     kvm_readonly_mem_allowed =
1625         (kvm_check_extension(s, KVM_CAP_READONLY_MEM) > 0);
1626 #endif
1627 
1628     kvm_eventfds_allowed =
1629         (kvm_check_extension(s, KVM_CAP_IOEVENTFD) > 0);
1630 
1631     kvm_irqfds_allowed =
1632         (kvm_check_extension(s, KVM_CAP_IRQFD) > 0);
1633 
1634     kvm_resamplefds_allowed =
1635         (kvm_check_extension(s, KVM_CAP_IRQFD_RESAMPLE) > 0);
1636 
1637     kvm_vm_attributes_allowed =
1638         (kvm_check_extension(s, KVM_CAP_VM_ATTRIBUTES) > 0);
1639 
1640     kvm_ioeventfd_any_length_allowed =
1641         (kvm_check_extension(s, KVM_CAP_IOEVENTFD_ANY_LENGTH) > 0);
1642 
1643     kvm_state = s;
1644 
1645     ret = kvm_arch_init(ms, s);
1646     if (ret < 0) {
1647         goto err;
1648     }
1649 
1650     if (machine_kernel_irqchip_allowed(ms)) {
1651         kvm_irqchip_create(ms, s);
1652     }
1653 
1654     if (kvm_eventfds_allowed) {
1655         s->memory_listener.listener.eventfd_add = kvm_mem_ioeventfd_add;
1656         s->memory_listener.listener.eventfd_del = kvm_mem_ioeventfd_del;
1657     }
1658     s->memory_listener.listener.coalesced_mmio_add = kvm_coalesce_mmio_region;
1659     s->memory_listener.listener.coalesced_mmio_del = kvm_uncoalesce_mmio_region;
1660 
1661     kvm_memory_listener_register(s, &s->memory_listener,
1662                                  &address_space_memory, 0);
1663     memory_listener_register(&kvm_io_listener,
1664                              &address_space_io);
1665 
1666     s->many_ioeventfds = kvm_check_many_ioeventfds();
1667 
1668     s->sync_mmu = !!kvm_vm_check_extension(kvm_state, KVM_CAP_SYNC_MMU);
1669 
1670     return 0;
1671 
1672 err:
1673     assert(ret < 0);
1674     if (s->vmfd >= 0) {
1675         close(s->vmfd);
1676     }
1677     if (s->fd != -1) {
1678         close(s->fd);
1679     }
1680     g_free(s->memory_listener.slots);
1681 
1682     return ret;
1683 }
1684 
1685 void kvm_set_sigmask_len(KVMState *s, unsigned int sigmask_len)
1686 {
1687     s->sigmask_len = sigmask_len;
1688 }
1689 
1690 static void kvm_handle_io(uint16_t port, MemTxAttrs attrs, void *data, int direction,
1691                           int size, uint32_t count)
1692 {
1693     int i;
1694     uint8_t *ptr = data;
1695 
1696     for (i = 0; i < count; i++) {
1697         address_space_rw(&address_space_io, port, attrs,
1698                          ptr, size,
1699                          direction == KVM_EXIT_IO_OUT);
1700         ptr += size;
1701     }
1702 }
1703 
1704 static int kvm_handle_internal_error(CPUState *cpu, struct kvm_run *run)
1705 {
1706     fprintf(stderr, "KVM internal error. Suberror: %d\n",
1707             run->internal.suberror);
1708 
1709     if (kvm_check_extension(kvm_state, KVM_CAP_INTERNAL_ERROR_DATA)) {
1710         int i;
1711 
1712         for (i = 0; i < run->internal.ndata; ++i) {
1713             fprintf(stderr, "extra data[%d]: %"PRIx64"\n",
1714                     i, (uint64_t)run->internal.data[i]);
1715         }
1716     }
1717     if (run->internal.suberror == KVM_INTERNAL_ERROR_EMULATION) {
1718         fprintf(stderr, "emulation failure\n");
1719         if (!kvm_arch_stop_on_emulation_error(cpu)) {
1720             cpu_dump_state(cpu, stderr, fprintf, CPU_DUMP_CODE);
1721             return EXCP_INTERRUPT;
1722         }
1723     }
1724     /* FIXME: Should trigger a qmp message to let management know
1725      * something went wrong.
1726      */
1727     return -1;
1728 }
1729 
1730 void kvm_flush_coalesced_mmio_buffer(void)
1731 {
1732     KVMState *s = kvm_state;
1733 
1734     if (s->coalesced_flush_in_progress) {
1735         return;
1736     }
1737 
1738     s->coalesced_flush_in_progress = true;
1739 
1740     if (s->coalesced_mmio_ring) {
1741         struct kvm_coalesced_mmio_ring *ring = s->coalesced_mmio_ring;
1742         while (ring->first != ring->last) {
1743             struct kvm_coalesced_mmio *ent;
1744 
1745             ent = &ring->coalesced_mmio[ring->first];
1746 
1747             cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len);
1748             smp_wmb();
1749             ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX;
1750         }
1751     }
1752 
1753     s->coalesced_flush_in_progress = false;
1754 }
1755 
1756 static void do_kvm_cpu_synchronize_state(CPUState *cpu, run_on_cpu_data arg)
1757 {
1758     if (!cpu->vcpu_dirty) {
1759         kvm_arch_get_registers(cpu);
1760         cpu->vcpu_dirty = true;
1761     }
1762 }
1763 
1764 void kvm_cpu_synchronize_state(CPUState *cpu)
1765 {
1766     if (!cpu->vcpu_dirty) {
1767         run_on_cpu(cpu, do_kvm_cpu_synchronize_state, RUN_ON_CPU_NULL);
1768     }
1769 }
1770 
1771 static void do_kvm_cpu_synchronize_post_reset(CPUState *cpu, run_on_cpu_data arg)
1772 {
1773     kvm_arch_put_registers(cpu, KVM_PUT_RESET_STATE);
1774     cpu->vcpu_dirty = false;
1775 }
1776 
1777 void kvm_cpu_synchronize_post_reset(CPUState *cpu)
1778 {
1779     run_on_cpu(cpu, do_kvm_cpu_synchronize_post_reset, RUN_ON_CPU_NULL);
1780 }
1781 
1782 static void do_kvm_cpu_synchronize_post_init(CPUState *cpu, run_on_cpu_data arg)
1783 {
1784     kvm_arch_put_registers(cpu, KVM_PUT_FULL_STATE);
1785     cpu->vcpu_dirty = false;
1786 }
1787 
1788 void kvm_cpu_synchronize_post_init(CPUState *cpu)
1789 {
1790     run_on_cpu(cpu, do_kvm_cpu_synchronize_post_init, RUN_ON_CPU_NULL);
1791 }
1792 
1793 static void do_kvm_cpu_synchronize_pre_loadvm(CPUState *cpu, run_on_cpu_data arg)
1794 {
1795     cpu->vcpu_dirty = true;
1796 }
1797 
1798 void kvm_cpu_synchronize_pre_loadvm(CPUState *cpu)
1799 {
1800     run_on_cpu(cpu, do_kvm_cpu_synchronize_pre_loadvm, RUN_ON_CPU_NULL);
1801 }
1802 
1803 #ifdef KVM_HAVE_MCE_INJECTION
1804 static __thread void *pending_sigbus_addr;
1805 static __thread int pending_sigbus_code;
1806 static __thread bool have_sigbus_pending;
1807 #endif
1808 
1809 static void kvm_cpu_kick(CPUState *cpu)
1810 {
1811     atomic_set(&cpu->kvm_run->immediate_exit, 1);
1812 }
1813 
1814 static void kvm_cpu_kick_self(void)
1815 {
1816     if (kvm_immediate_exit) {
1817         kvm_cpu_kick(current_cpu);
1818     } else {
1819         qemu_cpu_kick_self();
1820     }
1821 }
1822 
1823 static void kvm_eat_signals(CPUState *cpu)
1824 {
1825     struct timespec ts = { 0, 0 };
1826     siginfo_t siginfo;
1827     sigset_t waitset;
1828     sigset_t chkset;
1829     int r;
1830 
1831     if (kvm_immediate_exit) {
1832         atomic_set(&cpu->kvm_run->immediate_exit, 0);
1833         /* Write kvm_run->immediate_exit before the cpu->exit_request
1834          * write in kvm_cpu_exec.
1835          */
1836         smp_wmb();
1837         return;
1838     }
1839 
1840     sigemptyset(&waitset);
1841     sigaddset(&waitset, SIG_IPI);
1842 
1843     do {
1844         r = sigtimedwait(&waitset, &siginfo, &ts);
1845         if (r == -1 && !(errno == EAGAIN || errno == EINTR)) {
1846             perror("sigtimedwait");
1847             exit(1);
1848         }
1849 
1850         r = sigpending(&chkset);
1851         if (r == -1) {
1852             perror("sigpending");
1853             exit(1);
1854         }
1855     } while (sigismember(&chkset, SIG_IPI));
1856 }
1857 
1858 int kvm_cpu_exec(CPUState *cpu)
1859 {
1860     struct kvm_run *run = cpu->kvm_run;
1861     int ret, run_ret;
1862 
1863     DPRINTF("kvm_cpu_exec()\n");
1864 
1865     if (kvm_arch_process_async_events(cpu)) {
1866         atomic_set(&cpu->exit_request, 0);
1867         return EXCP_HLT;
1868     }
1869 
1870     qemu_mutex_unlock_iothread();
1871     cpu_exec_start(cpu);
1872 
1873     do {
1874         MemTxAttrs attrs;
1875 
1876         if (cpu->vcpu_dirty) {
1877             kvm_arch_put_registers(cpu, KVM_PUT_RUNTIME_STATE);
1878             cpu->vcpu_dirty = false;
1879         }
1880 
1881         kvm_arch_pre_run(cpu, run);
1882         if (atomic_read(&cpu->exit_request)) {
1883             DPRINTF("interrupt exit requested\n");
1884             /*
1885              * KVM requires us to reenter the kernel after IO exits to complete
1886              * instruction emulation. This self-signal will ensure that we
1887              * leave ASAP again.
1888              */
1889             kvm_cpu_kick_self();
1890         }
1891 
1892         /* Read cpu->exit_request before KVM_RUN reads run->immediate_exit.
1893          * Matching barrier in kvm_eat_signals.
1894          */
1895         smp_rmb();
1896 
1897         run_ret = kvm_vcpu_ioctl(cpu, KVM_RUN, 0);
1898 
1899         attrs = kvm_arch_post_run(cpu, run);
1900 
1901 #ifdef KVM_HAVE_MCE_INJECTION
1902         if (unlikely(have_sigbus_pending)) {
1903             qemu_mutex_lock_iothread();
1904             kvm_arch_on_sigbus_vcpu(cpu, pending_sigbus_code,
1905                                     pending_sigbus_addr);
1906             have_sigbus_pending = false;
1907             qemu_mutex_unlock_iothread();
1908         }
1909 #endif
1910 
1911         if (run_ret < 0) {
1912             if (run_ret == -EINTR || run_ret == -EAGAIN) {
1913                 DPRINTF("io window exit\n");
1914                 kvm_eat_signals(cpu);
1915                 ret = EXCP_INTERRUPT;
1916                 break;
1917             }
1918             fprintf(stderr, "error: kvm run failed %s\n",
1919                     strerror(-run_ret));
1920 #ifdef TARGET_PPC
1921             if (run_ret == -EBUSY) {
1922                 fprintf(stderr,
1923                         "This is probably because your SMT is enabled.\n"
1924                         "VCPU can only run on primary threads with all "
1925                         "secondary threads offline.\n");
1926             }
1927 #endif
1928             ret = -1;
1929             break;
1930         }
1931 
1932         trace_kvm_run_exit(cpu->cpu_index, run->exit_reason);
1933         switch (run->exit_reason) {
1934         case KVM_EXIT_IO:
1935             DPRINTF("handle_io\n");
1936             /* Called outside BQL */
1937             kvm_handle_io(run->io.port, attrs,
1938                           (uint8_t *)run + run->io.data_offset,
1939                           run->io.direction,
1940                           run->io.size,
1941                           run->io.count);
1942             ret = 0;
1943             break;
1944         case KVM_EXIT_MMIO:
1945             DPRINTF("handle_mmio\n");
1946             /* Called outside BQL */
1947             address_space_rw(&address_space_memory,
1948                              run->mmio.phys_addr, attrs,
1949                              run->mmio.data,
1950                              run->mmio.len,
1951                              run->mmio.is_write);
1952             ret = 0;
1953             break;
1954         case KVM_EXIT_IRQ_WINDOW_OPEN:
1955             DPRINTF("irq_window_open\n");
1956             ret = EXCP_INTERRUPT;
1957             break;
1958         case KVM_EXIT_SHUTDOWN:
1959             DPRINTF("shutdown\n");
1960             qemu_system_reset_request(SHUTDOWN_CAUSE_GUEST_RESET);
1961             ret = EXCP_INTERRUPT;
1962             break;
1963         case KVM_EXIT_UNKNOWN:
1964             fprintf(stderr, "KVM: unknown exit, hardware reason %" PRIx64 "\n",
1965                     (uint64_t)run->hw.hardware_exit_reason);
1966             ret = -1;
1967             break;
1968         case KVM_EXIT_INTERNAL_ERROR:
1969             ret = kvm_handle_internal_error(cpu, run);
1970             break;
1971         case KVM_EXIT_SYSTEM_EVENT:
1972             switch (run->system_event.type) {
1973             case KVM_SYSTEM_EVENT_SHUTDOWN:
1974                 qemu_system_shutdown_request(SHUTDOWN_CAUSE_GUEST_SHUTDOWN);
1975                 ret = EXCP_INTERRUPT;
1976                 break;
1977             case KVM_SYSTEM_EVENT_RESET:
1978                 qemu_system_reset_request(SHUTDOWN_CAUSE_GUEST_RESET);
1979                 ret = EXCP_INTERRUPT;
1980                 break;
1981             case KVM_SYSTEM_EVENT_CRASH:
1982                 kvm_cpu_synchronize_state(cpu);
1983                 qemu_mutex_lock_iothread();
1984                 qemu_system_guest_panicked(cpu_get_crash_info(cpu));
1985                 qemu_mutex_unlock_iothread();
1986                 ret = 0;
1987                 break;
1988             default:
1989                 DPRINTF("kvm_arch_handle_exit\n");
1990                 ret = kvm_arch_handle_exit(cpu, run);
1991                 break;
1992             }
1993             break;
1994         default:
1995             DPRINTF("kvm_arch_handle_exit\n");
1996             ret = kvm_arch_handle_exit(cpu, run);
1997             break;
1998         }
1999     } while (ret == 0);
2000 
2001     cpu_exec_end(cpu);
2002     qemu_mutex_lock_iothread();
2003 
2004     if (ret < 0) {
2005         cpu_dump_state(cpu, stderr, fprintf, CPU_DUMP_CODE);
2006         vm_stop(RUN_STATE_INTERNAL_ERROR);
2007     }
2008 
2009     atomic_set(&cpu->exit_request, 0);
2010     return ret;
2011 }
2012 
2013 int kvm_ioctl(KVMState *s, int type, ...)
2014 {
2015     int ret;
2016     void *arg;
2017     va_list ap;
2018 
2019     va_start(ap, type);
2020     arg = va_arg(ap, void *);
2021     va_end(ap);
2022 
2023     trace_kvm_ioctl(type, arg);
2024     ret = ioctl(s->fd, type, arg);
2025     if (ret == -1) {
2026         ret = -errno;
2027     }
2028     return ret;
2029 }
2030 
2031 int kvm_vm_ioctl(KVMState *s, int type, ...)
2032 {
2033     int ret;
2034     void *arg;
2035     va_list ap;
2036 
2037     va_start(ap, type);
2038     arg = va_arg(ap, void *);
2039     va_end(ap);
2040 
2041     trace_kvm_vm_ioctl(type, arg);
2042     ret = ioctl(s->vmfd, type, arg);
2043     if (ret == -1) {
2044         ret = -errno;
2045     }
2046     return ret;
2047 }
2048 
2049 int kvm_vcpu_ioctl(CPUState *cpu, int type, ...)
2050 {
2051     int ret;
2052     void *arg;
2053     va_list ap;
2054 
2055     va_start(ap, type);
2056     arg = va_arg(ap, void *);
2057     va_end(ap);
2058 
2059     trace_kvm_vcpu_ioctl(cpu->cpu_index, type, arg);
2060     ret = ioctl(cpu->kvm_fd, type, arg);
2061     if (ret == -1) {
2062         ret = -errno;
2063     }
2064     return ret;
2065 }
2066 
2067 int kvm_device_ioctl(int fd, int type, ...)
2068 {
2069     int ret;
2070     void *arg;
2071     va_list ap;
2072 
2073     va_start(ap, type);
2074     arg = va_arg(ap, void *);
2075     va_end(ap);
2076 
2077     trace_kvm_device_ioctl(fd, type, arg);
2078     ret = ioctl(fd, type, arg);
2079     if (ret == -1) {
2080         ret = -errno;
2081     }
2082     return ret;
2083 }
2084 
2085 int kvm_vm_check_attr(KVMState *s, uint32_t group, uint64_t attr)
2086 {
2087     int ret;
2088     struct kvm_device_attr attribute = {
2089         .group = group,
2090         .attr = attr,
2091     };
2092 
2093     if (!kvm_vm_attributes_allowed) {
2094         return 0;
2095     }
2096 
2097     ret = kvm_vm_ioctl(s, KVM_HAS_DEVICE_ATTR, &attribute);
2098     /* kvm returns 0 on success for HAS_DEVICE_ATTR */
2099     return ret ? 0 : 1;
2100 }
2101 
2102 int kvm_device_check_attr(int dev_fd, uint32_t group, uint64_t attr)
2103 {
2104     struct kvm_device_attr attribute = {
2105         .group = group,
2106         .attr = attr,
2107         .flags = 0,
2108     };
2109 
2110     return kvm_device_ioctl(dev_fd, KVM_HAS_DEVICE_ATTR, &attribute) ? 0 : 1;
2111 }
2112 
2113 int kvm_device_access(int fd, int group, uint64_t attr,
2114                       void *val, bool write, Error **errp)
2115 {
2116     struct kvm_device_attr kvmattr;
2117     int err;
2118 
2119     kvmattr.flags = 0;
2120     kvmattr.group = group;
2121     kvmattr.attr = attr;
2122     kvmattr.addr = (uintptr_t)val;
2123 
2124     err = kvm_device_ioctl(fd,
2125                            write ? KVM_SET_DEVICE_ATTR : KVM_GET_DEVICE_ATTR,
2126                            &kvmattr);
2127     if (err < 0) {
2128         error_setg_errno(errp, -err,
2129                          "KVM_%s_DEVICE_ATTR failed: Group %d "
2130                          "attr 0x%016" PRIx64,
2131                          write ? "SET" : "GET", group, attr);
2132     }
2133     return err;
2134 }
2135 
2136 bool kvm_has_sync_mmu(void)
2137 {
2138     return kvm_state->sync_mmu;
2139 }
2140 
2141 int kvm_has_vcpu_events(void)
2142 {
2143     return kvm_state->vcpu_events;
2144 }
2145 
2146 int kvm_has_robust_singlestep(void)
2147 {
2148     return kvm_state->robust_singlestep;
2149 }
2150 
2151 int kvm_has_debugregs(void)
2152 {
2153     return kvm_state->debugregs;
2154 }
2155 
2156 int kvm_has_many_ioeventfds(void)
2157 {
2158     if (!kvm_enabled()) {
2159         return 0;
2160     }
2161     return kvm_state->many_ioeventfds;
2162 }
2163 
2164 int kvm_has_gsi_routing(void)
2165 {
2166 #ifdef KVM_CAP_IRQ_ROUTING
2167     return kvm_check_extension(kvm_state, KVM_CAP_IRQ_ROUTING);
2168 #else
2169     return false;
2170 #endif
2171 }
2172 
2173 int kvm_has_intx_set_mask(void)
2174 {
2175     return kvm_state->intx_set_mask;
2176 }
2177 
2178 bool kvm_arm_supports_user_irq(void)
2179 {
2180     return kvm_check_extension(kvm_state, KVM_CAP_ARM_USER_IRQ);
2181 }
2182 
2183 #ifdef KVM_CAP_SET_GUEST_DEBUG
2184 struct kvm_sw_breakpoint *kvm_find_sw_breakpoint(CPUState *cpu,
2185                                                  target_ulong pc)
2186 {
2187     struct kvm_sw_breakpoint *bp;
2188 
2189     QTAILQ_FOREACH(bp, &cpu->kvm_state->kvm_sw_breakpoints, entry) {
2190         if (bp->pc == pc) {
2191             return bp;
2192         }
2193     }
2194     return NULL;
2195 }
2196 
2197 int kvm_sw_breakpoints_active(CPUState *cpu)
2198 {
2199     return !QTAILQ_EMPTY(&cpu->kvm_state->kvm_sw_breakpoints);
2200 }
2201 
2202 struct kvm_set_guest_debug_data {
2203     struct kvm_guest_debug dbg;
2204     int err;
2205 };
2206 
2207 static void kvm_invoke_set_guest_debug(CPUState *cpu, run_on_cpu_data data)
2208 {
2209     struct kvm_set_guest_debug_data *dbg_data =
2210         (struct kvm_set_guest_debug_data *) data.host_ptr;
2211 
2212     dbg_data->err = kvm_vcpu_ioctl(cpu, KVM_SET_GUEST_DEBUG,
2213                                    &dbg_data->dbg);
2214 }
2215 
2216 int kvm_update_guest_debug(CPUState *cpu, unsigned long reinject_trap)
2217 {
2218     struct kvm_set_guest_debug_data data;
2219 
2220     data.dbg.control = reinject_trap;
2221 
2222     if (cpu->singlestep_enabled) {
2223         data.dbg.control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_SINGLESTEP;
2224     }
2225     kvm_arch_update_guest_debug(cpu, &data.dbg);
2226 
2227     run_on_cpu(cpu, kvm_invoke_set_guest_debug,
2228                RUN_ON_CPU_HOST_PTR(&data));
2229     return data.err;
2230 }
2231 
2232 int kvm_insert_breakpoint(CPUState *cpu, target_ulong addr,
2233                           target_ulong len, int type)
2234 {
2235     struct kvm_sw_breakpoint *bp;
2236     int err;
2237 
2238     if (type == GDB_BREAKPOINT_SW) {
2239         bp = kvm_find_sw_breakpoint(cpu, addr);
2240         if (bp) {
2241             bp->use_count++;
2242             return 0;
2243         }
2244 
2245         bp = g_malloc(sizeof(struct kvm_sw_breakpoint));
2246         bp->pc = addr;
2247         bp->use_count = 1;
2248         err = kvm_arch_insert_sw_breakpoint(cpu, bp);
2249         if (err) {
2250             g_free(bp);
2251             return err;
2252         }
2253 
2254         QTAILQ_INSERT_HEAD(&cpu->kvm_state->kvm_sw_breakpoints, bp, entry);
2255     } else {
2256         err = kvm_arch_insert_hw_breakpoint(addr, len, type);
2257         if (err) {
2258             return err;
2259         }
2260     }
2261 
2262     CPU_FOREACH(cpu) {
2263         err = kvm_update_guest_debug(cpu, 0);
2264         if (err) {
2265             return err;
2266         }
2267     }
2268     return 0;
2269 }
2270 
2271 int kvm_remove_breakpoint(CPUState *cpu, target_ulong addr,
2272                           target_ulong len, int type)
2273 {
2274     struct kvm_sw_breakpoint *bp;
2275     int err;
2276 
2277     if (type == GDB_BREAKPOINT_SW) {
2278         bp = kvm_find_sw_breakpoint(cpu, addr);
2279         if (!bp) {
2280             return -ENOENT;
2281         }
2282 
2283         if (bp->use_count > 1) {
2284             bp->use_count--;
2285             return 0;
2286         }
2287 
2288         err = kvm_arch_remove_sw_breakpoint(cpu, bp);
2289         if (err) {
2290             return err;
2291         }
2292 
2293         QTAILQ_REMOVE(&cpu->kvm_state->kvm_sw_breakpoints, bp, entry);
2294         g_free(bp);
2295     } else {
2296         err = kvm_arch_remove_hw_breakpoint(addr, len, type);
2297         if (err) {
2298             return err;
2299         }
2300     }
2301 
2302     CPU_FOREACH(cpu) {
2303         err = kvm_update_guest_debug(cpu, 0);
2304         if (err) {
2305             return err;
2306         }
2307     }
2308     return 0;
2309 }
2310 
2311 void kvm_remove_all_breakpoints(CPUState *cpu)
2312 {
2313     struct kvm_sw_breakpoint *bp, *next;
2314     KVMState *s = cpu->kvm_state;
2315     CPUState *tmpcpu;
2316 
2317     QTAILQ_FOREACH_SAFE(bp, &s->kvm_sw_breakpoints, entry, next) {
2318         if (kvm_arch_remove_sw_breakpoint(cpu, bp) != 0) {
2319             /* Try harder to find a CPU that currently sees the breakpoint. */
2320             CPU_FOREACH(tmpcpu) {
2321                 if (kvm_arch_remove_sw_breakpoint(tmpcpu, bp) == 0) {
2322                     break;
2323                 }
2324             }
2325         }
2326         QTAILQ_REMOVE(&s->kvm_sw_breakpoints, bp, entry);
2327         g_free(bp);
2328     }
2329     kvm_arch_remove_all_hw_breakpoints();
2330 
2331     CPU_FOREACH(cpu) {
2332         kvm_update_guest_debug(cpu, 0);
2333     }
2334 }
2335 
2336 #else /* !KVM_CAP_SET_GUEST_DEBUG */
2337 
2338 int kvm_update_guest_debug(CPUState *cpu, unsigned long reinject_trap)
2339 {
2340     return -EINVAL;
2341 }
2342 
2343 int kvm_insert_breakpoint(CPUState *cpu, target_ulong addr,
2344                           target_ulong len, int type)
2345 {
2346     return -EINVAL;
2347 }
2348 
2349 int kvm_remove_breakpoint(CPUState *cpu, target_ulong addr,
2350                           target_ulong len, int type)
2351 {
2352     return -EINVAL;
2353 }
2354 
2355 void kvm_remove_all_breakpoints(CPUState *cpu)
2356 {
2357 }
2358 #endif /* !KVM_CAP_SET_GUEST_DEBUG */
2359 
2360 static int kvm_set_signal_mask(CPUState *cpu, const sigset_t *sigset)
2361 {
2362     KVMState *s = kvm_state;
2363     struct kvm_signal_mask *sigmask;
2364     int r;
2365 
2366     sigmask = g_malloc(sizeof(*sigmask) + sizeof(*sigset));
2367 
2368     sigmask->len = s->sigmask_len;
2369     memcpy(sigmask->sigset, sigset, sizeof(*sigset));
2370     r = kvm_vcpu_ioctl(cpu, KVM_SET_SIGNAL_MASK, sigmask);
2371     g_free(sigmask);
2372 
2373     return r;
2374 }
2375 
2376 static void kvm_ipi_signal(int sig)
2377 {
2378     if (current_cpu) {
2379         assert(kvm_immediate_exit);
2380         kvm_cpu_kick(current_cpu);
2381     }
2382 }
2383 
2384 void kvm_init_cpu_signals(CPUState *cpu)
2385 {
2386     int r;
2387     sigset_t set;
2388     struct sigaction sigact;
2389 
2390     memset(&sigact, 0, sizeof(sigact));
2391     sigact.sa_handler = kvm_ipi_signal;
2392     sigaction(SIG_IPI, &sigact, NULL);
2393 
2394     pthread_sigmask(SIG_BLOCK, NULL, &set);
2395 #if defined KVM_HAVE_MCE_INJECTION
2396     sigdelset(&set, SIGBUS);
2397     pthread_sigmask(SIG_SETMASK, &set, NULL);
2398 #endif
2399     sigdelset(&set, SIG_IPI);
2400     if (kvm_immediate_exit) {
2401         r = pthread_sigmask(SIG_SETMASK, &set, NULL);
2402     } else {
2403         r = kvm_set_signal_mask(cpu, &set);
2404     }
2405     if (r) {
2406         fprintf(stderr, "kvm_set_signal_mask: %s\n", strerror(-r));
2407         exit(1);
2408     }
2409 }
2410 
2411 /* Called asynchronously in VCPU thread.  */
2412 int kvm_on_sigbus_vcpu(CPUState *cpu, int code, void *addr)
2413 {
2414 #ifdef KVM_HAVE_MCE_INJECTION
2415     if (have_sigbus_pending) {
2416         return 1;
2417     }
2418     have_sigbus_pending = true;
2419     pending_sigbus_addr = addr;
2420     pending_sigbus_code = code;
2421     atomic_set(&cpu->exit_request, 1);
2422     return 0;
2423 #else
2424     return 1;
2425 #endif
2426 }
2427 
2428 /* Called synchronously (via signalfd) in main thread.  */
2429 int kvm_on_sigbus(int code, void *addr)
2430 {
2431 #ifdef KVM_HAVE_MCE_INJECTION
2432     /* Action required MCE kills the process if SIGBUS is blocked.  Because
2433      * that's what happens in the I/O thread, where we handle MCE via signalfd,
2434      * we can only get action optional here.
2435      */
2436     assert(code != BUS_MCEERR_AR);
2437     kvm_arch_on_sigbus_vcpu(first_cpu, code, addr);
2438     return 0;
2439 #else
2440     return 1;
2441 #endif
2442 }
2443 
2444 int kvm_create_device(KVMState *s, uint64_t type, bool test)
2445 {
2446     int ret;
2447     struct kvm_create_device create_dev;
2448 
2449     create_dev.type = type;
2450     create_dev.fd = -1;
2451     create_dev.flags = test ? KVM_CREATE_DEVICE_TEST : 0;
2452 
2453     if (!kvm_check_extension(s, KVM_CAP_DEVICE_CTRL)) {
2454         return -ENOTSUP;
2455     }
2456 
2457     ret = kvm_vm_ioctl(s, KVM_CREATE_DEVICE, &create_dev);
2458     if (ret) {
2459         return ret;
2460     }
2461 
2462     return test ? 0 : create_dev.fd;
2463 }
2464 
2465 bool kvm_device_supported(int vmfd, uint64_t type)
2466 {
2467     struct kvm_create_device create_dev = {
2468         .type = type,
2469         .fd = -1,
2470         .flags = KVM_CREATE_DEVICE_TEST,
2471     };
2472 
2473     if (ioctl(vmfd, KVM_CHECK_EXTENSION, KVM_CAP_DEVICE_CTRL) <= 0) {
2474         return false;
2475     }
2476 
2477     return (ioctl(vmfd, KVM_CREATE_DEVICE, &create_dev) >= 0);
2478 }
2479 
2480 int kvm_set_one_reg(CPUState *cs, uint64_t id, void *source)
2481 {
2482     struct kvm_one_reg reg;
2483     int r;
2484 
2485     reg.id = id;
2486     reg.addr = (uintptr_t) source;
2487     r = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, &reg);
2488     if (r) {
2489         trace_kvm_failed_reg_set(id, strerror(-r));
2490     }
2491     return r;
2492 }
2493 
2494 int kvm_get_one_reg(CPUState *cs, uint64_t id, void *target)
2495 {
2496     struct kvm_one_reg reg;
2497     int r;
2498 
2499     reg.id = id;
2500     reg.addr = (uintptr_t) target;
2501     r = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &reg);
2502     if (r) {
2503         trace_kvm_failed_reg_get(id, strerror(-r));
2504     }
2505     return r;
2506 }
2507 
2508 static void kvm_accel_class_init(ObjectClass *oc, void *data)
2509 {
2510     AccelClass *ac = ACCEL_CLASS(oc);
2511     ac->name = "KVM";
2512     ac->init_machine = kvm_init;
2513     ac->allowed = &kvm_allowed;
2514 }
2515 
2516 static const TypeInfo kvm_accel_type = {
2517     .name = TYPE_KVM_ACCEL,
2518     .parent = TYPE_ACCEL,
2519     .class_init = kvm_accel_class_init,
2520     .instance_size = sizeof(KVMState),
2521 };
2522 
2523 static void kvm_type_init(void)
2524 {
2525     type_register_static(&kvm_accel_type);
2526 }
2527 
2528 type_init(kvm_type_init);
2529