xref: /openbmc/qemu/accel/kvm/kvm-all.c (revision 69430111)
1 /*
2  * QEMU KVM support
3  *
4  * Copyright IBM, Corp. 2008
5  *           Red Hat, Inc. 2008
6  *
7  * Authors:
8  *  Anthony Liguori   <aliguori@us.ibm.com>
9  *  Glauber Costa     <gcosta@redhat.com>
10  *
11  * This work is licensed under the terms of the GNU GPL, version 2 or later.
12  * See the COPYING file in the top-level directory.
13  *
14  */
15 
16 #include "qemu/osdep.h"
17 #include <sys/ioctl.h>
18 
19 #include <linux/kvm.h>
20 
21 #include "qemu/atomic.h"
22 #include "qemu/option.h"
23 #include "qemu/config-file.h"
24 #include "qemu/error-report.h"
25 #include "qapi/error.h"
26 #include "hw/pci/msi.h"
27 #include "hw/pci/msix.h"
28 #include "hw/s390x/adapter.h"
29 #include "exec/gdbstub.h"
30 #include "sysemu/kvm_int.h"
31 #include "sysemu/runstate.h"
32 #include "sysemu/cpus.h"
33 #include "sysemu/sysemu.h"
34 #include "qemu/bswap.h"
35 #include "exec/memory.h"
36 #include "exec/ram_addr.h"
37 #include "exec/address-spaces.h"
38 #include "qemu/event_notifier.h"
39 #include "qemu/main-loop.h"
40 #include "trace.h"
41 #include "hw/irq.h"
42 #include "sysemu/sev.h"
43 #include "qapi/visitor.h"
44 #include "qapi/qapi-types-common.h"
45 #include "qapi/qapi-visit-common.h"
46 #include "sysemu/reset.h"
47 #include "qemu/guest-random.h"
48 #include "sysemu/hw_accel.h"
49 #include "kvm-cpus.h"
50 
51 #include "hw/boards.h"
52 
53 /* This check must be after config-host.h is included */
54 #ifdef CONFIG_EVENTFD
55 #include <sys/eventfd.h>
56 #endif
57 
58 /* KVM uses PAGE_SIZE in its definition of KVM_COALESCED_MMIO_MAX. We
59  * need to use the real host PAGE_SIZE, as that's what KVM will use.
60  */
61 #define PAGE_SIZE qemu_real_host_page_size
62 
63 //#define DEBUG_KVM
64 
65 #ifdef DEBUG_KVM
66 #define DPRINTF(fmt, ...) \
67     do { fprintf(stderr, fmt, ## __VA_ARGS__); } while (0)
68 #else
69 #define DPRINTF(fmt, ...) \
70     do { } while (0)
71 #endif
72 
73 #define KVM_MSI_HASHTAB_SIZE    256
74 
75 struct KVMParkedVcpu {
76     unsigned long vcpu_id;
77     int kvm_fd;
78     QLIST_ENTRY(KVMParkedVcpu) node;
79 };
80 
81 struct KVMState
82 {
83     AccelState parent_obj;
84 
85     int nr_slots;
86     int fd;
87     int vmfd;
88     int coalesced_mmio;
89     int coalesced_pio;
90     struct kvm_coalesced_mmio_ring *coalesced_mmio_ring;
91     bool coalesced_flush_in_progress;
92     int vcpu_events;
93     int robust_singlestep;
94     int debugregs;
95 #ifdef KVM_CAP_SET_GUEST_DEBUG
96     QTAILQ_HEAD(, kvm_sw_breakpoint) kvm_sw_breakpoints;
97 #endif
98     int max_nested_state_len;
99     int many_ioeventfds;
100     int intx_set_mask;
101     int kvm_shadow_mem;
102     bool kernel_irqchip_allowed;
103     bool kernel_irqchip_required;
104     OnOffAuto kernel_irqchip_split;
105     bool sync_mmu;
106     uint64_t manual_dirty_log_protect;
107     /* The man page (and posix) say ioctl numbers are signed int, but
108      * they're not.  Linux, glibc and *BSD all treat ioctl numbers as
109      * unsigned, and treating them as signed here can break things */
110     unsigned irq_set_ioctl;
111     unsigned int sigmask_len;
112     GHashTable *gsimap;
113 #ifdef KVM_CAP_IRQ_ROUTING
114     struct kvm_irq_routing *irq_routes;
115     int nr_allocated_irq_routes;
116     unsigned long *used_gsi_bitmap;
117     unsigned int gsi_count;
118     QTAILQ_HEAD(, KVMMSIRoute) msi_hashtab[KVM_MSI_HASHTAB_SIZE];
119 #endif
120     KVMMemoryListener memory_listener;
121     QLIST_HEAD(, KVMParkedVcpu) kvm_parked_vcpus;
122 
123     /* memory encryption */
124     void *memcrypt_handle;
125     int (*memcrypt_encrypt_data)(void *handle, uint8_t *ptr, uint64_t len);
126 
127     /* For "info mtree -f" to tell if an MR is registered in KVM */
128     int nr_as;
129     struct KVMAs {
130         KVMMemoryListener *ml;
131         AddressSpace *as;
132     } *as;
133 };
134 
135 KVMState *kvm_state;
136 bool kvm_kernel_irqchip;
137 bool kvm_split_irqchip;
138 bool kvm_async_interrupts_allowed;
139 bool kvm_halt_in_kernel_allowed;
140 bool kvm_eventfds_allowed;
141 bool kvm_irqfds_allowed;
142 bool kvm_resamplefds_allowed;
143 bool kvm_msi_via_irqfd_allowed;
144 bool kvm_gsi_routing_allowed;
145 bool kvm_gsi_direct_mapping;
146 bool kvm_allowed;
147 bool kvm_readonly_mem_allowed;
148 bool kvm_vm_attributes_allowed;
149 bool kvm_direct_msi_allowed;
150 bool kvm_ioeventfd_any_length_allowed;
151 bool kvm_msi_use_devid;
152 static bool kvm_immediate_exit;
153 static hwaddr kvm_max_slot_size = ~0;
154 
155 static const KVMCapabilityInfo kvm_required_capabilites[] = {
156     KVM_CAP_INFO(USER_MEMORY),
157     KVM_CAP_INFO(DESTROY_MEMORY_REGION_WORKS),
158     KVM_CAP_INFO(JOIN_MEMORY_REGIONS_WORKS),
159     KVM_CAP_LAST_INFO
160 };
161 
162 static NotifierList kvm_irqchip_change_notifiers =
163     NOTIFIER_LIST_INITIALIZER(kvm_irqchip_change_notifiers);
164 
165 struct KVMResampleFd {
166     int gsi;
167     EventNotifier *resample_event;
168     QLIST_ENTRY(KVMResampleFd) node;
169 };
170 typedef struct KVMResampleFd KVMResampleFd;
171 
172 /*
173  * Only used with split irqchip where we need to do the resample fd
174  * kick for the kernel from userspace.
175  */
176 static QLIST_HEAD(, KVMResampleFd) kvm_resample_fd_list =
177     QLIST_HEAD_INITIALIZER(kvm_resample_fd_list);
178 
179 #define kvm_slots_lock(kml)      qemu_mutex_lock(&(kml)->slots_lock)
180 #define kvm_slots_unlock(kml)    qemu_mutex_unlock(&(kml)->slots_lock)
181 
182 static inline void kvm_resample_fd_remove(int gsi)
183 {
184     KVMResampleFd *rfd;
185 
186     QLIST_FOREACH(rfd, &kvm_resample_fd_list, node) {
187         if (rfd->gsi == gsi) {
188             QLIST_REMOVE(rfd, node);
189             g_free(rfd);
190             break;
191         }
192     }
193 }
194 
195 static inline void kvm_resample_fd_insert(int gsi, EventNotifier *event)
196 {
197     KVMResampleFd *rfd = g_new0(KVMResampleFd, 1);
198 
199     rfd->gsi = gsi;
200     rfd->resample_event = event;
201 
202     QLIST_INSERT_HEAD(&kvm_resample_fd_list, rfd, node);
203 }
204 
205 void kvm_resample_fd_notify(int gsi)
206 {
207     KVMResampleFd *rfd;
208 
209     QLIST_FOREACH(rfd, &kvm_resample_fd_list, node) {
210         if (rfd->gsi == gsi) {
211             event_notifier_set(rfd->resample_event);
212             trace_kvm_resample_fd_notify(gsi);
213             return;
214         }
215     }
216 }
217 
218 int kvm_get_max_memslots(void)
219 {
220     KVMState *s = KVM_STATE(current_accel());
221 
222     return s->nr_slots;
223 }
224 
225 bool kvm_memcrypt_enabled(void)
226 {
227     if (kvm_state && kvm_state->memcrypt_handle) {
228         return true;
229     }
230 
231     return false;
232 }
233 
234 int kvm_memcrypt_encrypt_data(uint8_t *ptr, uint64_t len)
235 {
236     if (kvm_state->memcrypt_handle &&
237         kvm_state->memcrypt_encrypt_data) {
238         return kvm_state->memcrypt_encrypt_data(kvm_state->memcrypt_handle,
239                                               ptr, len);
240     }
241 
242     return 1;
243 }
244 
245 /* Called with KVMMemoryListener.slots_lock held */
246 static KVMSlot *kvm_get_free_slot(KVMMemoryListener *kml)
247 {
248     KVMState *s = kvm_state;
249     int i;
250 
251     for (i = 0; i < s->nr_slots; i++) {
252         if (kml->slots[i].memory_size == 0) {
253             return &kml->slots[i];
254         }
255     }
256 
257     return NULL;
258 }
259 
260 bool kvm_has_free_slot(MachineState *ms)
261 {
262     KVMState *s = KVM_STATE(ms->accelerator);
263     bool result;
264     KVMMemoryListener *kml = &s->memory_listener;
265 
266     kvm_slots_lock(kml);
267     result = !!kvm_get_free_slot(kml);
268     kvm_slots_unlock(kml);
269 
270     return result;
271 }
272 
273 /* Called with KVMMemoryListener.slots_lock held */
274 static KVMSlot *kvm_alloc_slot(KVMMemoryListener *kml)
275 {
276     KVMSlot *slot = kvm_get_free_slot(kml);
277 
278     if (slot) {
279         return slot;
280     }
281 
282     fprintf(stderr, "%s: no free slot available\n", __func__);
283     abort();
284 }
285 
286 static KVMSlot *kvm_lookup_matching_slot(KVMMemoryListener *kml,
287                                          hwaddr start_addr,
288                                          hwaddr size)
289 {
290     KVMState *s = kvm_state;
291     int i;
292 
293     for (i = 0; i < s->nr_slots; i++) {
294         KVMSlot *mem = &kml->slots[i];
295 
296         if (start_addr == mem->start_addr && size == mem->memory_size) {
297             return mem;
298         }
299     }
300 
301     return NULL;
302 }
303 
304 /*
305  * Calculate and align the start address and the size of the section.
306  * Return the size. If the size is 0, the aligned section is empty.
307  */
308 static hwaddr kvm_align_section(MemoryRegionSection *section,
309                                 hwaddr *start)
310 {
311     hwaddr size = int128_get64(section->size);
312     hwaddr delta, aligned;
313 
314     /* kvm works in page size chunks, but the function may be called
315        with sub-page size and unaligned start address. Pad the start
316        address to next and truncate size to previous page boundary. */
317     aligned = ROUND_UP(section->offset_within_address_space,
318                        qemu_real_host_page_size);
319     delta = aligned - section->offset_within_address_space;
320     *start = aligned;
321     if (delta > size) {
322         return 0;
323     }
324 
325     return (size - delta) & qemu_real_host_page_mask;
326 }
327 
328 int kvm_physical_memory_addr_from_host(KVMState *s, void *ram,
329                                        hwaddr *phys_addr)
330 {
331     KVMMemoryListener *kml = &s->memory_listener;
332     int i, ret = 0;
333 
334     kvm_slots_lock(kml);
335     for (i = 0; i < s->nr_slots; i++) {
336         KVMSlot *mem = &kml->slots[i];
337 
338         if (ram >= mem->ram && ram < mem->ram + mem->memory_size) {
339             *phys_addr = mem->start_addr + (ram - mem->ram);
340             ret = 1;
341             break;
342         }
343     }
344     kvm_slots_unlock(kml);
345 
346     return ret;
347 }
348 
349 static int kvm_set_user_memory_region(KVMMemoryListener *kml, KVMSlot *slot, bool new)
350 {
351     KVMState *s = kvm_state;
352     struct kvm_userspace_memory_region mem;
353     int ret;
354 
355     mem.slot = slot->slot | (kml->as_id << 16);
356     mem.guest_phys_addr = slot->start_addr;
357     mem.userspace_addr = (unsigned long)slot->ram;
358     mem.flags = slot->flags;
359 
360     if (slot->memory_size && !new && (mem.flags ^ slot->old_flags) & KVM_MEM_READONLY) {
361         /* Set the slot size to 0 before setting the slot to the desired
362          * value. This is needed based on KVM commit 75d61fbc. */
363         mem.memory_size = 0;
364         ret = kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
365         if (ret < 0) {
366             goto err;
367         }
368     }
369     mem.memory_size = slot->memory_size;
370     ret = kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
371     slot->old_flags = mem.flags;
372 err:
373     trace_kvm_set_user_memory(mem.slot, mem.flags, mem.guest_phys_addr,
374                               mem.memory_size, mem.userspace_addr, ret);
375     if (ret < 0) {
376         error_report("%s: KVM_SET_USER_MEMORY_REGION failed, slot=%d,"
377                      " start=0x%" PRIx64 ", size=0x%" PRIx64 ": %s",
378                      __func__, mem.slot, slot->start_addr,
379                      (uint64_t)mem.memory_size, strerror(errno));
380     }
381     return ret;
382 }
383 
384 static int do_kvm_destroy_vcpu(CPUState *cpu)
385 {
386     KVMState *s = kvm_state;
387     long mmap_size;
388     struct KVMParkedVcpu *vcpu = NULL;
389     int ret = 0;
390 
391     DPRINTF("kvm_destroy_vcpu\n");
392 
393     ret = kvm_arch_destroy_vcpu(cpu);
394     if (ret < 0) {
395         goto err;
396     }
397 
398     mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
399     if (mmap_size < 0) {
400         ret = mmap_size;
401         DPRINTF("KVM_GET_VCPU_MMAP_SIZE failed\n");
402         goto err;
403     }
404 
405     ret = munmap(cpu->kvm_run, mmap_size);
406     if (ret < 0) {
407         goto err;
408     }
409 
410     vcpu = g_malloc0(sizeof(*vcpu));
411     vcpu->vcpu_id = kvm_arch_vcpu_id(cpu);
412     vcpu->kvm_fd = cpu->kvm_fd;
413     QLIST_INSERT_HEAD(&kvm_state->kvm_parked_vcpus, vcpu, node);
414 err:
415     return ret;
416 }
417 
418 void kvm_destroy_vcpu(CPUState *cpu)
419 {
420     if (do_kvm_destroy_vcpu(cpu) < 0) {
421         error_report("kvm_destroy_vcpu failed");
422         exit(EXIT_FAILURE);
423     }
424 }
425 
426 static int kvm_get_vcpu(KVMState *s, unsigned long vcpu_id)
427 {
428     struct KVMParkedVcpu *cpu;
429 
430     QLIST_FOREACH(cpu, &s->kvm_parked_vcpus, node) {
431         if (cpu->vcpu_id == vcpu_id) {
432             int kvm_fd;
433 
434             QLIST_REMOVE(cpu, node);
435             kvm_fd = cpu->kvm_fd;
436             g_free(cpu);
437             return kvm_fd;
438         }
439     }
440 
441     return kvm_vm_ioctl(s, KVM_CREATE_VCPU, (void *)vcpu_id);
442 }
443 
444 int kvm_init_vcpu(CPUState *cpu, Error **errp)
445 {
446     KVMState *s = kvm_state;
447     long mmap_size;
448     int ret;
449 
450     trace_kvm_init_vcpu(cpu->cpu_index, kvm_arch_vcpu_id(cpu));
451 
452     ret = kvm_get_vcpu(s, kvm_arch_vcpu_id(cpu));
453     if (ret < 0) {
454         error_setg_errno(errp, -ret, "kvm_init_vcpu: kvm_get_vcpu failed (%lu)",
455                          kvm_arch_vcpu_id(cpu));
456         goto err;
457     }
458 
459     cpu->kvm_fd = ret;
460     cpu->kvm_state = s;
461     cpu->vcpu_dirty = true;
462 
463     mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
464     if (mmap_size < 0) {
465         ret = mmap_size;
466         error_setg_errno(errp, -mmap_size,
467                          "kvm_init_vcpu: KVM_GET_VCPU_MMAP_SIZE failed");
468         goto err;
469     }
470 
471     cpu->kvm_run = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
472                         cpu->kvm_fd, 0);
473     if (cpu->kvm_run == MAP_FAILED) {
474         ret = -errno;
475         error_setg_errno(errp, ret,
476                          "kvm_init_vcpu: mmap'ing vcpu state failed (%lu)",
477                          kvm_arch_vcpu_id(cpu));
478         goto err;
479     }
480 
481     if (s->coalesced_mmio && !s->coalesced_mmio_ring) {
482         s->coalesced_mmio_ring =
483             (void *)cpu->kvm_run + s->coalesced_mmio * PAGE_SIZE;
484     }
485 
486     ret = kvm_arch_init_vcpu(cpu);
487     if (ret < 0) {
488         error_setg_errno(errp, -ret,
489                          "kvm_init_vcpu: kvm_arch_init_vcpu failed (%lu)",
490                          kvm_arch_vcpu_id(cpu));
491     }
492 err:
493     return ret;
494 }
495 
496 /*
497  * dirty pages logging control
498  */
499 
500 static int kvm_mem_flags(MemoryRegion *mr)
501 {
502     bool readonly = mr->readonly || memory_region_is_romd(mr);
503     int flags = 0;
504 
505     if (memory_region_get_dirty_log_mask(mr) != 0) {
506         flags |= KVM_MEM_LOG_DIRTY_PAGES;
507     }
508     if (readonly && kvm_readonly_mem_allowed) {
509         flags |= KVM_MEM_READONLY;
510     }
511     return flags;
512 }
513 
514 /* Called with KVMMemoryListener.slots_lock held */
515 static int kvm_slot_update_flags(KVMMemoryListener *kml, KVMSlot *mem,
516                                  MemoryRegion *mr)
517 {
518     mem->flags = kvm_mem_flags(mr);
519 
520     /* If nothing changed effectively, no need to issue ioctl */
521     if (mem->flags == mem->old_flags) {
522         return 0;
523     }
524 
525     return kvm_set_user_memory_region(kml, mem, false);
526 }
527 
528 static int kvm_section_update_flags(KVMMemoryListener *kml,
529                                     MemoryRegionSection *section)
530 {
531     hwaddr start_addr, size, slot_size;
532     KVMSlot *mem;
533     int ret = 0;
534 
535     size = kvm_align_section(section, &start_addr);
536     if (!size) {
537         return 0;
538     }
539 
540     kvm_slots_lock(kml);
541 
542     while (size && !ret) {
543         slot_size = MIN(kvm_max_slot_size, size);
544         mem = kvm_lookup_matching_slot(kml, start_addr, slot_size);
545         if (!mem) {
546             /* We don't have a slot if we want to trap every access. */
547             goto out;
548         }
549 
550         ret = kvm_slot_update_flags(kml, mem, section->mr);
551         start_addr += slot_size;
552         size -= slot_size;
553     }
554 
555 out:
556     kvm_slots_unlock(kml);
557     return ret;
558 }
559 
560 static void kvm_log_start(MemoryListener *listener,
561                           MemoryRegionSection *section,
562                           int old, int new)
563 {
564     KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
565     int r;
566 
567     if (old != 0) {
568         return;
569     }
570 
571     r = kvm_section_update_flags(kml, section);
572     if (r < 0) {
573         abort();
574     }
575 }
576 
577 static void kvm_log_stop(MemoryListener *listener,
578                           MemoryRegionSection *section,
579                           int old, int new)
580 {
581     KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
582     int r;
583 
584     if (new != 0) {
585         return;
586     }
587 
588     r = kvm_section_update_flags(kml, section);
589     if (r < 0) {
590         abort();
591     }
592 }
593 
594 /* get kvm's dirty pages bitmap and update qemu's */
595 static int kvm_get_dirty_pages_log_range(MemoryRegionSection *section,
596                                          unsigned long *bitmap)
597 {
598     ram_addr_t start = section->offset_within_region +
599                        memory_region_get_ram_addr(section->mr);
600     ram_addr_t pages = int128_get64(section->size) / qemu_real_host_page_size;
601 
602     cpu_physical_memory_set_dirty_lebitmap(bitmap, start, pages);
603     return 0;
604 }
605 
606 #define ALIGN(x, y)  (((x)+(y)-1) & ~((y)-1))
607 
608 /* Allocate the dirty bitmap for a slot  */
609 static void kvm_memslot_init_dirty_bitmap(KVMSlot *mem)
610 {
611     /*
612      * XXX bad kernel interface alert
613      * For dirty bitmap, kernel allocates array of size aligned to
614      * bits-per-long.  But for case when the kernel is 64bits and
615      * the userspace is 32bits, userspace can't align to the same
616      * bits-per-long, since sizeof(long) is different between kernel
617      * and user space.  This way, userspace will provide buffer which
618      * may be 4 bytes less than the kernel will use, resulting in
619      * userspace memory corruption (which is not detectable by valgrind
620      * too, in most cases).
621      * So for now, let's align to 64 instead of HOST_LONG_BITS here, in
622      * a hope that sizeof(long) won't become >8 any time soon.
623      */
624     hwaddr bitmap_size = ALIGN(((mem->memory_size) >> TARGET_PAGE_BITS),
625                                         /*HOST_LONG_BITS*/ 64) / 8;
626     mem->dirty_bmap = g_malloc0(bitmap_size);
627 }
628 
629 /**
630  * kvm_physical_sync_dirty_bitmap - Sync dirty bitmap from kernel space
631  *
632  * This function will first try to fetch dirty bitmap from the kernel,
633  * and then updates qemu's dirty bitmap.
634  *
635  * NOTE: caller must be with kml->slots_lock held.
636  *
637  * @kml: the KVM memory listener object
638  * @section: the memory section to sync the dirty bitmap with
639  */
640 static int kvm_physical_sync_dirty_bitmap(KVMMemoryListener *kml,
641                                           MemoryRegionSection *section)
642 {
643     KVMState *s = kvm_state;
644     struct kvm_dirty_log d = {};
645     KVMSlot *mem;
646     hwaddr start_addr, size;
647     hwaddr slot_size, slot_offset = 0;
648     int ret = 0;
649 
650     size = kvm_align_section(section, &start_addr);
651     while (size) {
652         MemoryRegionSection subsection = *section;
653 
654         slot_size = MIN(kvm_max_slot_size, size);
655         mem = kvm_lookup_matching_slot(kml, start_addr, slot_size);
656         if (!mem) {
657             /* We don't have a slot if we want to trap every access. */
658             goto out;
659         }
660 
661         if (!mem->dirty_bmap) {
662             /* Allocate on the first log_sync, once and for all */
663             kvm_memslot_init_dirty_bitmap(mem);
664         }
665 
666         d.dirty_bitmap = mem->dirty_bmap;
667         d.slot = mem->slot | (kml->as_id << 16);
668         if (kvm_vm_ioctl(s, KVM_GET_DIRTY_LOG, &d) == -1) {
669             DPRINTF("ioctl failed %d\n", errno);
670             ret = -1;
671             goto out;
672         }
673 
674         subsection.offset_within_region += slot_offset;
675         subsection.size = int128_make64(slot_size);
676         kvm_get_dirty_pages_log_range(&subsection, d.dirty_bitmap);
677 
678         slot_offset += slot_size;
679         start_addr += slot_size;
680         size -= slot_size;
681     }
682 out:
683     return ret;
684 }
685 
686 /* Alignment requirement for KVM_CLEAR_DIRTY_LOG - 64 pages */
687 #define KVM_CLEAR_LOG_SHIFT  6
688 #define KVM_CLEAR_LOG_ALIGN  (qemu_real_host_page_size << KVM_CLEAR_LOG_SHIFT)
689 #define KVM_CLEAR_LOG_MASK   (-KVM_CLEAR_LOG_ALIGN)
690 
691 static int kvm_log_clear_one_slot(KVMSlot *mem, int as_id, uint64_t start,
692                                   uint64_t size)
693 {
694     KVMState *s = kvm_state;
695     uint64_t end, bmap_start, start_delta, bmap_npages;
696     struct kvm_clear_dirty_log d;
697     unsigned long *bmap_clear = NULL, psize = qemu_real_host_page_size;
698     int ret;
699 
700     /*
701      * We need to extend either the start or the size or both to
702      * satisfy the KVM interface requirement.  Firstly, do the start
703      * page alignment on 64 host pages
704      */
705     bmap_start = start & KVM_CLEAR_LOG_MASK;
706     start_delta = start - bmap_start;
707     bmap_start /= psize;
708 
709     /*
710      * The kernel interface has restriction on the size too, that either:
711      *
712      * (1) the size is 64 host pages aligned (just like the start), or
713      * (2) the size fills up until the end of the KVM memslot.
714      */
715     bmap_npages = DIV_ROUND_UP(size + start_delta, KVM_CLEAR_LOG_ALIGN)
716         << KVM_CLEAR_LOG_SHIFT;
717     end = mem->memory_size / psize;
718     if (bmap_npages > end - bmap_start) {
719         bmap_npages = end - bmap_start;
720     }
721     start_delta /= psize;
722 
723     /*
724      * Prepare the bitmap to clear dirty bits.  Here we must guarantee
725      * that we won't clear any unknown dirty bits otherwise we might
726      * accidentally clear some set bits which are not yet synced from
727      * the kernel into QEMU's bitmap, then we'll lose track of the
728      * guest modifications upon those pages (which can directly lead
729      * to guest data loss or panic after migration).
730      *
731      * Layout of the KVMSlot.dirty_bmap:
732      *
733      *                   |<-------- bmap_npages -----------..>|
734      *                                                     [1]
735      *                     start_delta         size
736      *  |----------------|-------------|------------------|------------|
737      *  ^                ^             ^                               ^
738      *  |                |             |                               |
739      * start          bmap_start     (start)                         end
740      * of memslot                                             of memslot
741      *
742      * [1] bmap_npages can be aligned to either 64 pages or the end of slot
743      */
744 
745     assert(bmap_start % BITS_PER_LONG == 0);
746     /* We should never do log_clear before log_sync */
747     assert(mem->dirty_bmap);
748     if (start_delta) {
749         /* Slow path - we need to manipulate a temp bitmap */
750         bmap_clear = bitmap_new(bmap_npages);
751         bitmap_copy_with_src_offset(bmap_clear, mem->dirty_bmap,
752                                     bmap_start, start_delta + size / psize);
753         /*
754          * We need to fill the holes at start because that was not
755          * specified by the caller and we extended the bitmap only for
756          * 64 pages alignment
757          */
758         bitmap_clear(bmap_clear, 0, start_delta);
759         d.dirty_bitmap = bmap_clear;
760     } else {
761         /* Fast path - start address aligns well with BITS_PER_LONG */
762         d.dirty_bitmap = mem->dirty_bmap + BIT_WORD(bmap_start);
763     }
764 
765     d.first_page = bmap_start;
766     /* It should never overflow.  If it happens, say something */
767     assert(bmap_npages <= UINT32_MAX);
768     d.num_pages = bmap_npages;
769     d.slot = mem->slot | (as_id << 16);
770 
771     if (kvm_vm_ioctl(s, KVM_CLEAR_DIRTY_LOG, &d) == -1) {
772         ret = -errno;
773         error_report("%s: KVM_CLEAR_DIRTY_LOG failed, slot=%d, "
774                      "start=0x%"PRIx64", size=0x%"PRIx32", errno=%d",
775                      __func__, d.slot, (uint64_t)d.first_page,
776                      (uint32_t)d.num_pages, ret);
777     } else {
778         ret = 0;
779         trace_kvm_clear_dirty_log(d.slot, d.first_page, d.num_pages);
780     }
781 
782     /*
783      * After we have updated the remote dirty bitmap, we update the
784      * cached bitmap as well for the memslot, then if another user
785      * clears the same region we know we shouldn't clear it again on
786      * the remote otherwise it's data loss as well.
787      */
788     bitmap_clear(mem->dirty_bmap, bmap_start + start_delta,
789                  size / psize);
790     /* This handles the NULL case well */
791     g_free(bmap_clear);
792     return ret;
793 }
794 
795 
796 /**
797  * kvm_physical_log_clear - Clear the kernel's dirty bitmap for range
798  *
799  * NOTE: this will be a no-op if we haven't enabled manual dirty log
800  * protection in the host kernel because in that case this operation
801  * will be done within log_sync().
802  *
803  * @kml:     the kvm memory listener
804  * @section: the memory range to clear dirty bitmap
805  */
806 static int kvm_physical_log_clear(KVMMemoryListener *kml,
807                                   MemoryRegionSection *section)
808 {
809     KVMState *s = kvm_state;
810     uint64_t start, size, offset, count;
811     KVMSlot *mem;
812     int ret = 0, i;
813 
814     if (!s->manual_dirty_log_protect) {
815         /* No need to do explicit clear */
816         return ret;
817     }
818 
819     start = section->offset_within_address_space;
820     size = int128_get64(section->size);
821 
822     if (!size) {
823         /* Nothing more we can do... */
824         return ret;
825     }
826 
827     kvm_slots_lock(kml);
828 
829     for (i = 0; i < s->nr_slots; i++) {
830         mem = &kml->slots[i];
831         /* Discard slots that are empty or do not overlap the section */
832         if (!mem->memory_size ||
833             mem->start_addr > start + size - 1 ||
834             start > mem->start_addr + mem->memory_size - 1) {
835             continue;
836         }
837 
838         if (start >= mem->start_addr) {
839             /* The slot starts before section or is aligned to it.  */
840             offset = start - mem->start_addr;
841             count = MIN(mem->memory_size - offset, size);
842         } else {
843             /* The slot starts after section.  */
844             offset = 0;
845             count = MIN(mem->memory_size, size - (mem->start_addr - start));
846         }
847         ret = kvm_log_clear_one_slot(mem, kml->as_id, offset, count);
848         if (ret < 0) {
849             break;
850         }
851     }
852 
853     kvm_slots_unlock(kml);
854 
855     return ret;
856 }
857 
858 static void kvm_coalesce_mmio_region(MemoryListener *listener,
859                                      MemoryRegionSection *secion,
860                                      hwaddr start, hwaddr size)
861 {
862     KVMState *s = kvm_state;
863 
864     if (s->coalesced_mmio) {
865         struct kvm_coalesced_mmio_zone zone;
866 
867         zone.addr = start;
868         zone.size = size;
869         zone.pad = 0;
870 
871         (void)kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone);
872     }
873 }
874 
875 static void kvm_uncoalesce_mmio_region(MemoryListener *listener,
876                                        MemoryRegionSection *secion,
877                                        hwaddr start, hwaddr size)
878 {
879     KVMState *s = kvm_state;
880 
881     if (s->coalesced_mmio) {
882         struct kvm_coalesced_mmio_zone zone;
883 
884         zone.addr = start;
885         zone.size = size;
886         zone.pad = 0;
887 
888         (void)kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone);
889     }
890 }
891 
892 static void kvm_coalesce_pio_add(MemoryListener *listener,
893                                 MemoryRegionSection *section,
894                                 hwaddr start, hwaddr size)
895 {
896     KVMState *s = kvm_state;
897 
898     if (s->coalesced_pio) {
899         struct kvm_coalesced_mmio_zone zone;
900 
901         zone.addr = start;
902         zone.size = size;
903         zone.pio = 1;
904 
905         (void)kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone);
906     }
907 }
908 
909 static void kvm_coalesce_pio_del(MemoryListener *listener,
910                                 MemoryRegionSection *section,
911                                 hwaddr start, hwaddr size)
912 {
913     KVMState *s = kvm_state;
914 
915     if (s->coalesced_pio) {
916         struct kvm_coalesced_mmio_zone zone;
917 
918         zone.addr = start;
919         zone.size = size;
920         zone.pio = 1;
921 
922         (void)kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone);
923      }
924 }
925 
926 static MemoryListener kvm_coalesced_pio_listener = {
927     .coalesced_io_add = kvm_coalesce_pio_add,
928     .coalesced_io_del = kvm_coalesce_pio_del,
929 };
930 
931 int kvm_check_extension(KVMState *s, unsigned int extension)
932 {
933     int ret;
934 
935     ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, extension);
936     if (ret < 0) {
937         ret = 0;
938     }
939 
940     return ret;
941 }
942 
943 int kvm_vm_check_extension(KVMState *s, unsigned int extension)
944 {
945     int ret;
946 
947     ret = kvm_vm_ioctl(s, KVM_CHECK_EXTENSION, extension);
948     if (ret < 0) {
949         /* VM wide version not implemented, use global one instead */
950         ret = kvm_check_extension(s, extension);
951     }
952 
953     return ret;
954 }
955 
956 typedef struct HWPoisonPage {
957     ram_addr_t ram_addr;
958     QLIST_ENTRY(HWPoisonPage) list;
959 } HWPoisonPage;
960 
961 static QLIST_HEAD(, HWPoisonPage) hwpoison_page_list =
962     QLIST_HEAD_INITIALIZER(hwpoison_page_list);
963 
964 static void kvm_unpoison_all(void *param)
965 {
966     HWPoisonPage *page, *next_page;
967 
968     QLIST_FOREACH_SAFE(page, &hwpoison_page_list, list, next_page) {
969         QLIST_REMOVE(page, list);
970         qemu_ram_remap(page->ram_addr, TARGET_PAGE_SIZE);
971         g_free(page);
972     }
973 }
974 
975 void kvm_hwpoison_page_add(ram_addr_t ram_addr)
976 {
977     HWPoisonPage *page;
978 
979     QLIST_FOREACH(page, &hwpoison_page_list, list) {
980         if (page->ram_addr == ram_addr) {
981             return;
982         }
983     }
984     page = g_new(HWPoisonPage, 1);
985     page->ram_addr = ram_addr;
986     QLIST_INSERT_HEAD(&hwpoison_page_list, page, list);
987 }
988 
989 static uint32_t adjust_ioeventfd_endianness(uint32_t val, uint32_t size)
990 {
991 #if defined(HOST_WORDS_BIGENDIAN) != defined(TARGET_WORDS_BIGENDIAN)
992     /* The kernel expects ioeventfd values in HOST_WORDS_BIGENDIAN
993      * endianness, but the memory core hands them in target endianness.
994      * For example, PPC is always treated as big-endian even if running
995      * on KVM and on PPC64LE.  Correct here.
996      */
997     switch (size) {
998     case 2:
999         val = bswap16(val);
1000         break;
1001     case 4:
1002         val = bswap32(val);
1003         break;
1004     }
1005 #endif
1006     return val;
1007 }
1008 
1009 static int kvm_set_ioeventfd_mmio(int fd, hwaddr addr, uint32_t val,
1010                                   bool assign, uint32_t size, bool datamatch)
1011 {
1012     int ret;
1013     struct kvm_ioeventfd iofd = {
1014         .datamatch = datamatch ? adjust_ioeventfd_endianness(val, size) : 0,
1015         .addr = addr,
1016         .len = size,
1017         .flags = 0,
1018         .fd = fd,
1019     };
1020 
1021     trace_kvm_set_ioeventfd_mmio(fd, (uint64_t)addr, val, assign, size,
1022                                  datamatch);
1023     if (!kvm_enabled()) {
1024         return -ENOSYS;
1025     }
1026 
1027     if (datamatch) {
1028         iofd.flags |= KVM_IOEVENTFD_FLAG_DATAMATCH;
1029     }
1030     if (!assign) {
1031         iofd.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
1032     }
1033 
1034     ret = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &iofd);
1035 
1036     if (ret < 0) {
1037         return -errno;
1038     }
1039 
1040     return 0;
1041 }
1042 
1043 static int kvm_set_ioeventfd_pio(int fd, uint16_t addr, uint16_t val,
1044                                  bool assign, uint32_t size, bool datamatch)
1045 {
1046     struct kvm_ioeventfd kick = {
1047         .datamatch = datamatch ? adjust_ioeventfd_endianness(val, size) : 0,
1048         .addr = addr,
1049         .flags = KVM_IOEVENTFD_FLAG_PIO,
1050         .len = size,
1051         .fd = fd,
1052     };
1053     int r;
1054     trace_kvm_set_ioeventfd_pio(fd, addr, val, assign, size, datamatch);
1055     if (!kvm_enabled()) {
1056         return -ENOSYS;
1057     }
1058     if (datamatch) {
1059         kick.flags |= KVM_IOEVENTFD_FLAG_DATAMATCH;
1060     }
1061     if (!assign) {
1062         kick.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
1063     }
1064     r = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &kick);
1065     if (r < 0) {
1066         return r;
1067     }
1068     return 0;
1069 }
1070 
1071 
1072 static int kvm_check_many_ioeventfds(void)
1073 {
1074     /* Userspace can use ioeventfd for io notification.  This requires a host
1075      * that supports eventfd(2) and an I/O thread; since eventfd does not
1076      * support SIGIO it cannot interrupt the vcpu.
1077      *
1078      * Older kernels have a 6 device limit on the KVM io bus.  Find out so we
1079      * can avoid creating too many ioeventfds.
1080      */
1081 #if defined(CONFIG_EVENTFD)
1082     int ioeventfds[7];
1083     int i, ret = 0;
1084     for (i = 0; i < ARRAY_SIZE(ioeventfds); i++) {
1085         ioeventfds[i] = eventfd(0, EFD_CLOEXEC);
1086         if (ioeventfds[i] < 0) {
1087             break;
1088         }
1089         ret = kvm_set_ioeventfd_pio(ioeventfds[i], 0, i, true, 2, true);
1090         if (ret < 0) {
1091             close(ioeventfds[i]);
1092             break;
1093         }
1094     }
1095 
1096     /* Decide whether many devices are supported or not */
1097     ret = i == ARRAY_SIZE(ioeventfds);
1098 
1099     while (i-- > 0) {
1100         kvm_set_ioeventfd_pio(ioeventfds[i], 0, i, false, 2, true);
1101         close(ioeventfds[i]);
1102     }
1103     return ret;
1104 #else
1105     return 0;
1106 #endif
1107 }
1108 
1109 static const KVMCapabilityInfo *
1110 kvm_check_extension_list(KVMState *s, const KVMCapabilityInfo *list)
1111 {
1112     while (list->name) {
1113         if (!kvm_check_extension(s, list->value)) {
1114             return list;
1115         }
1116         list++;
1117     }
1118     return NULL;
1119 }
1120 
1121 void kvm_set_max_memslot_size(hwaddr max_slot_size)
1122 {
1123     g_assert(
1124         ROUND_UP(max_slot_size, qemu_real_host_page_size) == max_slot_size
1125     );
1126     kvm_max_slot_size = max_slot_size;
1127 }
1128 
1129 static void kvm_set_phys_mem(KVMMemoryListener *kml,
1130                              MemoryRegionSection *section, bool add)
1131 {
1132     KVMSlot *mem;
1133     int err;
1134     MemoryRegion *mr = section->mr;
1135     bool writeable = !mr->readonly && !mr->rom_device;
1136     hwaddr start_addr, size, slot_size;
1137     void *ram;
1138 
1139     if (!memory_region_is_ram(mr)) {
1140         if (writeable || !kvm_readonly_mem_allowed) {
1141             return;
1142         } else if (!mr->romd_mode) {
1143             /* If the memory device is not in romd_mode, then we actually want
1144              * to remove the kvm memory slot so all accesses will trap. */
1145             add = false;
1146         }
1147     }
1148 
1149     size = kvm_align_section(section, &start_addr);
1150     if (!size) {
1151         return;
1152     }
1153 
1154     /* use aligned delta to align the ram address */
1155     ram = memory_region_get_ram_ptr(mr) + section->offset_within_region +
1156           (start_addr - section->offset_within_address_space);
1157 
1158     kvm_slots_lock(kml);
1159 
1160     if (!add) {
1161         do {
1162             slot_size = MIN(kvm_max_slot_size, size);
1163             mem = kvm_lookup_matching_slot(kml, start_addr, slot_size);
1164             if (!mem) {
1165                 goto out;
1166             }
1167             if (mem->flags & KVM_MEM_LOG_DIRTY_PAGES) {
1168                 kvm_physical_sync_dirty_bitmap(kml, section);
1169             }
1170 
1171             /* unregister the slot */
1172             g_free(mem->dirty_bmap);
1173             mem->dirty_bmap = NULL;
1174             mem->memory_size = 0;
1175             mem->flags = 0;
1176             err = kvm_set_user_memory_region(kml, mem, false);
1177             if (err) {
1178                 fprintf(stderr, "%s: error unregistering slot: %s\n",
1179                         __func__, strerror(-err));
1180                 abort();
1181             }
1182             start_addr += slot_size;
1183             size -= slot_size;
1184         } while (size);
1185         goto out;
1186     }
1187 
1188     /* register the new slot */
1189     do {
1190         slot_size = MIN(kvm_max_slot_size, size);
1191         mem = kvm_alloc_slot(kml);
1192         mem->memory_size = slot_size;
1193         mem->start_addr = start_addr;
1194         mem->ram = ram;
1195         mem->flags = kvm_mem_flags(mr);
1196 
1197         if (mem->flags & KVM_MEM_LOG_DIRTY_PAGES) {
1198             /*
1199              * Reallocate the bmap; it means it doesn't disappear in
1200              * middle of a migrate.
1201              */
1202             kvm_memslot_init_dirty_bitmap(mem);
1203         }
1204         err = kvm_set_user_memory_region(kml, mem, true);
1205         if (err) {
1206             fprintf(stderr, "%s: error registering slot: %s\n", __func__,
1207                     strerror(-err));
1208             abort();
1209         }
1210         start_addr += slot_size;
1211         ram += slot_size;
1212         size -= slot_size;
1213     } while (size);
1214 
1215 out:
1216     kvm_slots_unlock(kml);
1217 }
1218 
1219 static void kvm_region_add(MemoryListener *listener,
1220                            MemoryRegionSection *section)
1221 {
1222     KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
1223 
1224     memory_region_ref(section->mr);
1225     kvm_set_phys_mem(kml, section, true);
1226 }
1227 
1228 static void kvm_region_del(MemoryListener *listener,
1229                            MemoryRegionSection *section)
1230 {
1231     KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
1232 
1233     kvm_set_phys_mem(kml, section, false);
1234     memory_region_unref(section->mr);
1235 }
1236 
1237 static void kvm_log_sync(MemoryListener *listener,
1238                          MemoryRegionSection *section)
1239 {
1240     KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
1241     int r;
1242 
1243     kvm_slots_lock(kml);
1244     r = kvm_physical_sync_dirty_bitmap(kml, section);
1245     kvm_slots_unlock(kml);
1246     if (r < 0) {
1247         abort();
1248     }
1249 }
1250 
1251 static void kvm_log_clear(MemoryListener *listener,
1252                           MemoryRegionSection *section)
1253 {
1254     KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
1255     int r;
1256 
1257     r = kvm_physical_log_clear(kml, section);
1258     if (r < 0) {
1259         error_report_once("%s: kvm log clear failed: mr=%s "
1260                           "offset=%"HWADDR_PRIx" size=%"PRIx64, __func__,
1261                           section->mr->name, section->offset_within_region,
1262                           int128_get64(section->size));
1263         abort();
1264     }
1265 }
1266 
1267 static void kvm_mem_ioeventfd_add(MemoryListener *listener,
1268                                   MemoryRegionSection *section,
1269                                   bool match_data, uint64_t data,
1270                                   EventNotifier *e)
1271 {
1272     int fd = event_notifier_get_fd(e);
1273     int r;
1274 
1275     r = kvm_set_ioeventfd_mmio(fd, section->offset_within_address_space,
1276                                data, true, int128_get64(section->size),
1277                                match_data);
1278     if (r < 0) {
1279         fprintf(stderr, "%s: error adding ioeventfd: %s (%d)\n",
1280                 __func__, strerror(-r), -r);
1281         abort();
1282     }
1283 }
1284 
1285 static void kvm_mem_ioeventfd_del(MemoryListener *listener,
1286                                   MemoryRegionSection *section,
1287                                   bool match_data, uint64_t data,
1288                                   EventNotifier *e)
1289 {
1290     int fd = event_notifier_get_fd(e);
1291     int r;
1292 
1293     r = kvm_set_ioeventfd_mmio(fd, section->offset_within_address_space,
1294                                data, false, int128_get64(section->size),
1295                                match_data);
1296     if (r < 0) {
1297         fprintf(stderr, "%s: error deleting ioeventfd: %s (%d)\n",
1298                 __func__, strerror(-r), -r);
1299         abort();
1300     }
1301 }
1302 
1303 static void kvm_io_ioeventfd_add(MemoryListener *listener,
1304                                  MemoryRegionSection *section,
1305                                  bool match_data, uint64_t data,
1306                                  EventNotifier *e)
1307 {
1308     int fd = event_notifier_get_fd(e);
1309     int r;
1310 
1311     r = kvm_set_ioeventfd_pio(fd, section->offset_within_address_space,
1312                               data, true, int128_get64(section->size),
1313                               match_data);
1314     if (r < 0) {
1315         fprintf(stderr, "%s: error adding ioeventfd: %s (%d)\n",
1316                 __func__, strerror(-r), -r);
1317         abort();
1318     }
1319 }
1320 
1321 static void kvm_io_ioeventfd_del(MemoryListener *listener,
1322                                  MemoryRegionSection *section,
1323                                  bool match_data, uint64_t data,
1324                                  EventNotifier *e)
1325 
1326 {
1327     int fd = event_notifier_get_fd(e);
1328     int r;
1329 
1330     r = kvm_set_ioeventfd_pio(fd, section->offset_within_address_space,
1331                               data, false, int128_get64(section->size),
1332                               match_data);
1333     if (r < 0) {
1334         fprintf(stderr, "%s: error deleting ioeventfd: %s (%d)\n",
1335                 __func__, strerror(-r), -r);
1336         abort();
1337     }
1338 }
1339 
1340 void kvm_memory_listener_register(KVMState *s, KVMMemoryListener *kml,
1341                                   AddressSpace *as, int as_id)
1342 {
1343     int i;
1344 
1345     qemu_mutex_init(&kml->slots_lock);
1346     kml->slots = g_malloc0(s->nr_slots * sizeof(KVMSlot));
1347     kml->as_id = as_id;
1348 
1349     for (i = 0; i < s->nr_slots; i++) {
1350         kml->slots[i].slot = i;
1351     }
1352 
1353     kml->listener.region_add = kvm_region_add;
1354     kml->listener.region_del = kvm_region_del;
1355     kml->listener.log_start = kvm_log_start;
1356     kml->listener.log_stop = kvm_log_stop;
1357     kml->listener.log_sync = kvm_log_sync;
1358     kml->listener.log_clear = kvm_log_clear;
1359     kml->listener.priority = 10;
1360 
1361     memory_listener_register(&kml->listener, as);
1362 
1363     for (i = 0; i < s->nr_as; ++i) {
1364         if (!s->as[i].as) {
1365             s->as[i].as = as;
1366             s->as[i].ml = kml;
1367             break;
1368         }
1369     }
1370 }
1371 
1372 static MemoryListener kvm_io_listener = {
1373     .eventfd_add = kvm_io_ioeventfd_add,
1374     .eventfd_del = kvm_io_ioeventfd_del,
1375     .priority = 10,
1376 };
1377 
1378 int kvm_set_irq(KVMState *s, int irq, int level)
1379 {
1380     struct kvm_irq_level event;
1381     int ret;
1382 
1383     assert(kvm_async_interrupts_enabled());
1384 
1385     event.level = level;
1386     event.irq = irq;
1387     ret = kvm_vm_ioctl(s, s->irq_set_ioctl, &event);
1388     if (ret < 0) {
1389         perror("kvm_set_irq");
1390         abort();
1391     }
1392 
1393     return (s->irq_set_ioctl == KVM_IRQ_LINE) ? 1 : event.status;
1394 }
1395 
1396 #ifdef KVM_CAP_IRQ_ROUTING
1397 typedef struct KVMMSIRoute {
1398     struct kvm_irq_routing_entry kroute;
1399     QTAILQ_ENTRY(KVMMSIRoute) entry;
1400 } KVMMSIRoute;
1401 
1402 static void set_gsi(KVMState *s, unsigned int gsi)
1403 {
1404     set_bit(gsi, s->used_gsi_bitmap);
1405 }
1406 
1407 static void clear_gsi(KVMState *s, unsigned int gsi)
1408 {
1409     clear_bit(gsi, s->used_gsi_bitmap);
1410 }
1411 
1412 void kvm_init_irq_routing(KVMState *s)
1413 {
1414     int gsi_count, i;
1415 
1416     gsi_count = kvm_check_extension(s, KVM_CAP_IRQ_ROUTING) - 1;
1417     if (gsi_count > 0) {
1418         /* Round up so we can search ints using ffs */
1419         s->used_gsi_bitmap = bitmap_new(gsi_count);
1420         s->gsi_count = gsi_count;
1421     }
1422 
1423     s->irq_routes = g_malloc0(sizeof(*s->irq_routes));
1424     s->nr_allocated_irq_routes = 0;
1425 
1426     if (!kvm_direct_msi_allowed) {
1427         for (i = 0; i < KVM_MSI_HASHTAB_SIZE; i++) {
1428             QTAILQ_INIT(&s->msi_hashtab[i]);
1429         }
1430     }
1431 
1432     kvm_arch_init_irq_routing(s);
1433 }
1434 
1435 void kvm_irqchip_commit_routes(KVMState *s)
1436 {
1437     int ret;
1438 
1439     if (kvm_gsi_direct_mapping()) {
1440         return;
1441     }
1442 
1443     if (!kvm_gsi_routing_enabled()) {
1444         return;
1445     }
1446 
1447     s->irq_routes->flags = 0;
1448     trace_kvm_irqchip_commit_routes();
1449     ret = kvm_vm_ioctl(s, KVM_SET_GSI_ROUTING, s->irq_routes);
1450     assert(ret == 0);
1451 }
1452 
1453 static void kvm_add_routing_entry(KVMState *s,
1454                                   struct kvm_irq_routing_entry *entry)
1455 {
1456     struct kvm_irq_routing_entry *new;
1457     int n, size;
1458 
1459     if (s->irq_routes->nr == s->nr_allocated_irq_routes) {
1460         n = s->nr_allocated_irq_routes * 2;
1461         if (n < 64) {
1462             n = 64;
1463         }
1464         size = sizeof(struct kvm_irq_routing);
1465         size += n * sizeof(*new);
1466         s->irq_routes = g_realloc(s->irq_routes, size);
1467         s->nr_allocated_irq_routes = n;
1468     }
1469     n = s->irq_routes->nr++;
1470     new = &s->irq_routes->entries[n];
1471 
1472     *new = *entry;
1473 
1474     set_gsi(s, entry->gsi);
1475 }
1476 
1477 static int kvm_update_routing_entry(KVMState *s,
1478                                     struct kvm_irq_routing_entry *new_entry)
1479 {
1480     struct kvm_irq_routing_entry *entry;
1481     int n;
1482 
1483     for (n = 0; n < s->irq_routes->nr; n++) {
1484         entry = &s->irq_routes->entries[n];
1485         if (entry->gsi != new_entry->gsi) {
1486             continue;
1487         }
1488 
1489         if(!memcmp(entry, new_entry, sizeof *entry)) {
1490             return 0;
1491         }
1492 
1493         *entry = *new_entry;
1494 
1495         return 0;
1496     }
1497 
1498     return -ESRCH;
1499 }
1500 
1501 void kvm_irqchip_add_irq_route(KVMState *s, int irq, int irqchip, int pin)
1502 {
1503     struct kvm_irq_routing_entry e = {};
1504 
1505     assert(pin < s->gsi_count);
1506 
1507     e.gsi = irq;
1508     e.type = KVM_IRQ_ROUTING_IRQCHIP;
1509     e.flags = 0;
1510     e.u.irqchip.irqchip = irqchip;
1511     e.u.irqchip.pin = pin;
1512     kvm_add_routing_entry(s, &e);
1513 }
1514 
1515 void kvm_irqchip_release_virq(KVMState *s, int virq)
1516 {
1517     struct kvm_irq_routing_entry *e;
1518     int i;
1519 
1520     if (kvm_gsi_direct_mapping()) {
1521         return;
1522     }
1523 
1524     for (i = 0; i < s->irq_routes->nr; i++) {
1525         e = &s->irq_routes->entries[i];
1526         if (e->gsi == virq) {
1527             s->irq_routes->nr--;
1528             *e = s->irq_routes->entries[s->irq_routes->nr];
1529         }
1530     }
1531     clear_gsi(s, virq);
1532     kvm_arch_release_virq_post(virq);
1533     trace_kvm_irqchip_release_virq(virq);
1534 }
1535 
1536 void kvm_irqchip_add_change_notifier(Notifier *n)
1537 {
1538     notifier_list_add(&kvm_irqchip_change_notifiers, n);
1539 }
1540 
1541 void kvm_irqchip_remove_change_notifier(Notifier *n)
1542 {
1543     notifier_remove(n);
1544 }
1545 
1546 void kvm_irqchip_change_notify(void)
1547 {
1548     notifier_list_notify(&kvm_irqchip_change_notifiers, NULL);
1549 }
1550 
1551 static unsigned int kvm_hash_msi(uint32_t data)
1552 {
1553     /* This is optimized for IA32 MSI layout. However, no other arch shall
1554      * repeat the mistake of not providing a direct MSI injection API. */
1555     return data & 0xff;
1556 }
1557 
1558 static void kvm_flush_dynamic_msi_routes(KVMState *s)
1559 {
1560     KVMMSIRoute *route, *next;
1561     unsigned int hash;
1562 
1563     for (hash = 0; hash < KVM_MSI_HASHTAB_SIZE; hash++) {
1564         QTAILQ_FOREACH_SAFE(route, &s->msi_hashtab[hash], entry, next) {
1565             kvm_irqchip_release_virq(s, route->kroute.gsi);
1566             QTAILQ_REMOVE(&s->msi_hashtab[hash], route, entry);
1567             g_free(route);
1568         }
1569     }
1570 }
1571 
1572 static int kvm_irqchip_get_virq(KVMState *s)
1573 {
1574     int next_virq;
1575 
1576     /*
1577      * PIC and IOAPIC share the first 16 GSI numbers, thus the available
1578      * GSI numbers are more than the number of IRQ route. Allocating a GSI
1579      * number can succeed even though a new route entry cannot be added.
1580      * When this happens, flush dynamic MSI entries to free IRQ route entries.
1581      */
1582     if (!kvm_direct_msi_allowed && s->irq_routes->nr == s->gsi_count) {
1583         kvm_flush_dynamic_msi_routes(s);
1584     }
1585 
1586     /* Return the lowest unused GSI in the bitmap */
1587     next_virq = find_first_zero_bit(s->used_gsi_bitmap, s->gsi_count);
1588     if (next_virq >= s->gsi_count) {
1589         return -ENOSPC;
1590     } else {
1591         return next_virq;
1592     }
1593 }
1594 
1595 static KVMMSIRoute *kvm_lookup_msi_route(KVMState *s, MSIMessage msg)
1596 {
1597     unsigned int hash = kvm_hash_msi(msg.data);
1598     KVMMSIRoute *route;
1599 
1600     QTAILQ_FOREACH(route, &s->msi_hashtab[hash], entry) {
1601         if (route->kroute.u.msi.address_lo == (uint32_t)msg.address &&
1602             route->kroute.u.msi.address_hi == (msg.address >> 32) &&
1603             route->kroute.u.msi.data == le32_to_cpu(msg.data)) {
1604             return route;
1605         }
1606     }
1607     return NULL;
1608 }
1609 
1610 int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg)
1611 {
1612     struct kvm_msi msi;
1613     KVMMSIRoute *route;
1614 
1615     if (kvm_direct_msi_allowed) {
1616         msi.address_lo = (uint32_t)msg.address;
1617         msi.address_hi = msg.address >> 32;
1618         msi.data = le32_to_cpu(msg.data);
1619         msi.flags = 0;
1620         memset(msi.pad, 0, sizeof(msi.pad));
1621 
1622         return kvm_vm_ioctl(s, KVM_SIGNAL_MSI, &msi);
1623     }
1624 
1625     route = kvm_lookup_msi_route(s, msg);
1626     if (!route) {
1627         int virq;
1628 
1629         virq = kvm_irqchip_get_virq(s);
1630         if (virq < 0) {
1631             return virq;
1632         }
1633 
1634         route = g_malloc0(sizeof(KVMMSIRoute));
1635         route->kroute.gsi = virq;
1636         route->kroute.type = KVM_IRQ_ROUTING_MSI;
1637         route->kroute.flags = 0;
1638         route->kroute.u.msi.address_lo = (uint32_t)msg.address;
1639         route->kroute.u.msi.address_hi = msg.address >> 32;
1640         route->kroute.u.msi.data = le32_to_cpu(msg.data);
1641 
1642         kvm_add_routing_entry(s, &route->kroute);
1643         kvm_irqchip_commit_routes(s);
1644 
1645         QTAILQ_INSERT_TAIL(&s->msi_hashtab[kvm_hash_msi(msg.data)], route,
1646                            entry);
1647     }
1648 
1649     assert(route->kroute.type == KVM_IRQ_ROUTING_MSI);
1650 
1651     return kvm_set_irq(s, route->kroute.gsi, 1);
1652 }
1653 
1654 int kvm_irqchip_add_msi_route(KVMState *s, int vector, PCIDevice *dev)
1655 {
1656     struct kvm_irq_routing_entry kroute = {};
1657     int virq;
1658     MSIMessage msg = {0, 0};
1659 
1660     if (pci_available && dev) {
1661         msg = pci_get_msi_message(dev, vector);
1662     }
1663 
1664     if (kvm_gsi_direct_mapping()) {
1665         return kvm_arch_msi_data_to_gsi(msg.data);
1666     }
1667 
1668     if (!kvm_gsi_routing_enabled()) {
1669         return -ENOSYS;
1670     }
1671 
1672     virq = kvm_irqchip_get_virq(s);
1673     if (virq < 0) {
1674         return virq;
1675     }
1676 
1677     kroute.gsi = virq;
1678     kroute.type = KVM_IRQ_ROUTING_MSI;
1679     kroute.flags = 0;
1680     kroute.u.msi.address_lo = (uint32_t)msg.address;
1681     kroute.u.msi.address_hi = msg.address >> 32;
1682     kroute.u.msi.data = le32_to_cpu(msg.data);
1683     if (pci_available && kvm_msi_devid_required()) {
1684         kroute.flags = KVM_MSI_VALID_DEVID;
1685         kroute.u.msi.devid = pci_requester_id(dev);
1686     }
1687     if (kvm_arch_fixup_msi_route(&kroute, msg.address, msg.data, dev)) {
1688         kvm_irqchip_release_virq(s, virq);
1689         return -EINVAL;
1690     }
1691 
1692     trace_kvm_irqchip_add_msi_route(dev ? dev->name : (char *)"N/A",
1693                                     vector, virq);
1694 
1695     kvm_add_routing_entry(s, &kroute);
1696     kvm_arch_add_msi_route_post(&kroute, vector, dev);
1697     kvm_irqchip_commit_routes(s);
1698 
1699     return virq;
1700 }
1701 
1702 int kvm_irqchip_update_msi_route(KVMState *s, int virq, MSIMessage msg,
1703                                  PCIDevice *dev)
1704 {
1705     struct kvm_irq_routing_entry kroute = {};
1706 
1707     if (kvm_gsi_direct_mapping()) {
1708         return 0;
1709     }
1710 
1711     if (!kvm_irqchip_in_kernel()) {
1712         return -ENOSYS;
1713     }
1714 
1715     kroute.gsi = virq;
1716     kroute.type = KVM_IRQ_ROUTING_MSI;
1717     kroute.flags = 0;
1718     kroute.u.msi.address_lo = (uint32_t)msg.address;
1719     kroute.u.msi.address_hi = msg.address >> 32;
1720     kroute.u.msi.data = le32_to_cpu(msg.data);
1721     if (pci_available && kvm_msi_devid_required()) {
1722         kroute.flags = KVM_MSI_VALID_DEVID;
1723         kroute.u.msi.devid = pci_requester_id(dev);
1724     }
1725     if (kvm_arch_fixup_msi_route(&kroute, msg.address, msg.data, dev)) {
1726         return -EINVAL;
1727     }
1728 
1729     trace_kvm_irqchip_update_msi_route(virq);
1730 
1731     return kvm_update_routing_entry(s, &kroute);
1732 }
1733 
1734 static int kvm_irqchip_assign_irqfd(KVMState *s, EventNotifier *event,
1735                                     EventNotifier *resample, int virq,
1736                                     bool assign)
1737 {
1738     int fd = event_notifier_get_fd(event);
1739     int rfd = resample ? event_notifier_get_fd(resample) : -1;
1740 
1741     struct kvm_irqfd irqfd = {
1742         .fd = fd,
1743         .gsi = virq,
1744         .flags = assign ? 0 : KVM_IRQFD_FLAG_DEASSIGN,
1745     };
1746 
1747     if (rfd != -1) {
1748         assert(assign);
1749         if (kvm_irqchip_is_split()) {
1750             /*
1751              * When the slow irqchip (e.g. IOAPIC) is in the
1752              * userspace, KVM kernel resamplefd will not work because
1753              * the EOI of the interrupt will be delivered to userspace
1754              * instead, so the KVM kernel resamplefd kick will be
1755              * skipped.  The userspace here mimics what the kernel
1756              * provides with resamplefd, remember the resamplefd and
1757              * kick it when we receive EOI of this IRQ.
1758              *
1759              * This is hackery because IOAPIC is mostly bypassed
1760              * (except EOI broadcasts) when irqfd is used.  However
1761              * this can bring much performance back for split irqchip
1762              * with INTx IRQs (for VFIO, this gives 93% perf of the
1763              * full fast path, which is 46% perf boost comparing to
1764              * the INTx slow path).
1765              */
1766             kvm_resample_fd_insert(virq, resample);
1767         } else {
1768             irqfd.flags |= KVM_IRQFD_FLAG_RESAMPLE;
1769             irqfd.resamplefd = rfd;
1770         }
1771     } else if (!assign) {
1772         if (kvm_irqchip_is_split()) {
1773             kvm_resample_fd_remove(virq);
1774         }
1775     }
1776 
1777     if (!kvm_irqfds_enabled()) {
1778         return -ENOSYS;
1779     }
1780 
1781     return kvm_vm_ioctl(s, KVM_IRQFD, &irqfd);
1782 }
1783 
1784 int kvm_irqchip_add_adapter_route(KVMState *s, AdapterInfo *adapter)
1785 {
1786     struct kvm_irq_routing_entry kroute = {};
1787     int virq;
1788 
1789     if (!kvm_gsi_routing_enabled()) {
1790         return -ENOSYS;
1791     }
1792 
1793     virq = kvm_irqchip_get_virq(s);
1794     if (virq < 0) {
1795         return virq;
1796     }
1797 
1798     kroute.gsi = virq;
1799     kroute.type = KVM_IRQ_ROUTING_S390_ADAPTER;
1800     kroute.flags = 0;
1801     kroute.u.adapter.summary_addr = adapter->summary_addr;
1802     kroute.u.adapter.ind_addr = adapter->ind_addr;
1803     kroute.u.adapter.summary_offset = adapter->summary_offset;
1804     kroute.u.adapter.ind_offset = adapter->ind_offset;
1805     kroute.u.adapter.adapter_id = adapter->adapter_id;
1806 
1807     kvm_add_routing_entry(s, &kroute);
1808 
1809     return virq;
1810 }
1811 
1812 int kvm_irqchip_add_hv_sint_route(KVMState *s, uint32_t vcpu, uint32_t sint)
1813 {
1814     struct kvm_irq_routing_entry kroute = {};
1815     int virq;
1816 
1817     if (!kvm_gsi_routing_enabled()) {
1818         return -ENOSYS;
1819     }
1820     if (!kvm_check_extension(s, KVM_CAP_HYPERV_SYNIC)) {
1821         return -ENOSYS;
1822     }
1823     virq = kvm_irqchip_get_virq(s);
1824     if (virq < 0) {
1825         return virq;
1826     }
1827 
1828     kroute.gsi = virq;
1829     kroute.type = KVM_IRQ_ROUTING_HV_SINT;
1830     kroute.flags = 0;
1831     kroute.u.hv_sint.vcpu = vcpu;
1832     kroute.u.hv_sint.sint = sint;
1833 
1834     kvm_add_routing_entry(s, &kroute);
1835     kvm_irqchip_commit_routes(s);
1836 
1837     return virq;
1838 }
1839 
1840 #else /* !KVM_CAP_IRQ_ROUTING */
1841 
1842 void kvm_init_irq_routing(KVMState *s)
1843 {
1844 }
1845 
1846 void kvm_irqchip_release_virq(KVMState *s, int virq)
1847 {
1848 }
1849 
1850 int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg)
1851 {
1852     abort();
1853 }
1854 
1855 int kvm_irqchip_add_msi_route(KVMState *s, int vector, PCIDevice *dev)
1856 {
1857     return -ENOSYS;
1858 }
1859 
1860 int kvm_irqchip_add_adapter_route(KVMState *s, AdapterInfo *adapter)
1861 {
1862     return -ENOSYS;
1863 }
1864 
1865 int kvm_irqchip_add_hv_sint_route(KVMState *s, uint32_t vcpu, uint32_t sint)
1866 {
1867     return -ENOSYS;
1868 }
1869 
1870 static int kvm_irqchip_assign_irqfd(KVMState *s, EventNotifier *event,
1871                                     EventNotifier *resample, int virq,
1872                                     bool assign)
1873 {
1874     abort();
1875 }
1876 
1877 int kvm_irqchip_update_msi_route(KVMState *s, int virq, MSIMessage msg)
1878 {
1879     return -ENOSYS;
1880 }
1881 #endif /* !KVM_CAP_IRQ_ROUTING */
1882 
1883 int kvm_irqchip_add_irqfd_notifier_gsi(KVMState *s, EventNotifier *n,
1884                                        EventNotifier *rn, int virq)
1885 {
1886     return kvm_irqchip_assign_irqfd(s, n, rn, virq, true);
1887 }
1888 
1889 int kvm_irqchip_remove_irqfd_notifier_gsi(KVMState *s, EventNotifier *n,
1890                                           int virq)
1891 {
1892     return kvm_irqchip_assign_irqfd(s, n, NULL, virq, false);
1893 }
1894 
1895 int kvm_irqchip_add_irqfd_notifier(KVMState *s, EventNotifier *n,
1896                                    EventNotifier *rn, qemu_irq irq)
1897 {
1898     gpointer key, gsi;
1899     gboolean found = g_hash_table_lookup_extended(s->gsimap, irq, &key, &gsi);
1900 
1901     if (!found) {
1902         return -ENXIO;
1903     }
1904     return kvm_irqchip_add_irqfd_notifier_gsi(s, n, rn, GPOINTER_TO_INT(gsi));
1905 }
1906 
1907 int kvm_irqchip_remove_irqfd_notifier(KVMState *s, EventNotifier *n,
1908                                       qemu_irq irq)
1909 {
1910     gpointer key, gsi;
1911     gboolean found = g_hash_table_lookup_extended(s->gsimap, irq, &key, &gsi);
1912 
1913     if (!found) {
1914         return -ENXIO;
1915     }
1916     return kvm_irqchip_remove_irqfd_notifier_gsi(s, n, GPOINTER_TO_INT(gsi));
1917 }
1918 
1919 void kvm_irqchip_set_qemuirq_gsi(KVMState *s, qemu_irq irq, int gsi)
1920 {
1921     g_hash_table_insert(s->gsimap, irq, GINT_TO_POINTER(gsi));
1922 }
1923 
1924 static void kvm_irqchip_create(KVMState *s)
1925 {
1926     int ret;
1927 
1928     assert(s->kernel_irqchip_split != ON_OFF_AUTO_AUTO);
1929     if (kvm_check_extension(s, KVM_CAP_IRQCHIP)) {
1930         ;
1931     } else if (kvm_check_extension(s, KVM_CAP_S390_IRQCHIP)) {
1932         ret = kvm_vm_enable_cap(s, KVM_CAP_S390_IRQCHIP, 0);
1933         if (ret < 0) {
1934             fprintf(stderr, "Enable kernel irqchip failed: %s\n", strerror(-ret));
1935             exit(1);
1936         }
1937     } else {
1938         return;
1939     }
1940 
1941     /* First probe and see if there's a arch-specific hook to create the
1942      * in-kernel irqchip for us */
1943     ret = kvm_arch_irqchip_create(s);
1944     if (ret == 0) {
1945         if (s->kernel_irqchip_split == ON_OFF_AUTO_ON) {
1946             perror("Split IRQ chip mode not supported.");
1947             exit(1);
1948         } else {
1949             ret = kvm_vm_ioctl(s, KVM_CREATE_IRQCHIP);
1950         }
1951     }
1952     if (ret < 0) {
1953         fprintf(stderr, "Create kernel irqchip failed: %s\n", strerror(-ret));
1954         exit(1);
1955     }
1956 
1957     kvm_kernel_irqchip = true;
1958     /* If we have an in-kernel IRQ chip then we must have asynchronous
1959      * interrupt delivery (though the reverse is not necessarily true)
1960      */
1961     kvm_async_interrupts_allowed = true;
1962     kvm_halt_in_kernel_allowed = true;
1963 
1964     kvm_init_irq_routing(s);
1965 
1966     s->gsimap = g_hash_table_new(g_direct_hash, g_direct_equal);
1967 }
1968 
1969 /* Find number of supported CPUs using the recommended
1970  * procedure from the kernel API documentation to cope with
1971  * older kernels that may be missing capabilities.
1972  */
1973 static int kvm_recommended_vcpus(KVMState *s)
1974 {
1975     int ret = kvm_vm_check_extension(s, KVM_CAP_NR_VCPUS);
1976     return (ret) ? ret : 4;
1977 }
1978 
1979 static int kvm_max_vcpus(KVMState *s)
1980 {
1981     int ret = kvm_check_extension(s, KVM_CAP_MAX_VCPUS);
1982     return (ret) ? ret : kvm_recommended_vcpus(s);
1983 }
1984 
1985 static int kvm_max_vcpu_id(KVMState *s)
1986 {
1987     int ret = kvm_check_extension(s, KVM_CAP_MAX_VCPU_ID);
1988     return (ret) ? ret : kvm_max_vcpus(s);
1989 }
1990 
1991 bool kvm_vcpu_id_is_valid(int vcpu_id)
1992 {
1993     KVMState *s = KVM_STATE(current_accel());
1994     return vcpu_id >= 0 && vcpu_id < kvm_max_vcpu_id(s);
1995 }
1996 
1997 static int kvm_init(MachineState *ms)
1998 {
1999     MachineClass *mc = MACHINE_GET_CLASS(ms);
2000     static const char upgrade_note[] =
2001         "Please upgrade to at least kernel 2.6.29 or recent kvm-kmod\n"
2002         "(see http://sourceforge.net/projects/kvm).\n";
2003     struct {
2004         const char *name;
2005         int num;
2006     } num_cpus[] = {
2007         { "SMP",          ms->smp.cpus },
2008         { "hotpluggable", ms->smp.max_cpus },
2009         { NULL, }
2010     }, *nc = num_cpus;
2011     int soft_vcpus_limit, hard_vcpus_limit;
2012     KVMState *s;
2013     const KVMCapabilityInfo *missing_cap;
2014     int ret;
2015     int type = 0;
2016     const char *kvm_type;
2017     uint64_t dirty_log_manual_caps;
2018 
2019     s = KVM_STATE(ms->accelerator);
2020 
2021     /*
2022      * On systems where the kernel can support different base page
2023      * sizes, host page size may be different from TARGET_PAGE_SIZE,
2024      * even with KVM.  TARGET_PAGE_SIZE is assumed to be the minimum
2025      * page size for the system though.
2026      */
2027     assert(TARGET_PAGE_SIZE <= qemu_real_host_page_size);
2028 
2029     s->sigmask_len = 8;
2030 
2031 #ifdef KVM_CAP_SET_GUEST_DEBUG
2032     QTAILQ_INIT(&s->kvm_sw_breakpoints);
2033 #endif
2034     QLIST_INIT(&s->kvm_parked_vcpus);
2035     s->vmfd = -1;
2036     s->fd = qemu_open_old("/dev/kvm", O_RDWR);
2037     if (s->fd == -1) {
2038         fprintf(stderr, "Could not access KVM kernel module: %m\n");
2039         ret = -errno;
2040         goto err;
2041     }
2042 
2043     ret = kvm_ioctl(s, KVM_GET_API_VERSION, 0);
2044     if (ret < KVM_API_VERSION) {
2045         if (ret >= 0) {
2046             ret = -EINVAL;
2047         }
2048         fprintf(stderr, "kvm version too old\n");
2049         goto err;
2050     }
2051 
2052     if (ret > KVM_API_VERSION) {
2053         ret = -EINVAL;
2054         fprintf(stderr, "kvm version not supported\n");
2055         goto err;
2056     }
2057 
2058     kvm_immediate_exit = kvm_check_extension(s, KVM_CAP_IMMEDIATE_EXIT);
2059     s->nr_slots = kvm_check_extension(s, KVM_CAP_NR_MEMSLOTS);
2060 
2061     /* If unspecified, use the default value */
2062     if (!s->nr_slots) {
2063         s->nr_slots = 32;
2064     }
2065 
2066     s->nr_as = kvm_check_extension(s, KVM_CAP_MULTI_ADDRESS_SPACE);
2067     if (s->nr_as <= 1) {
2068         s->nr_as = 1;
2069     }
2070     s->as = g_new0(struct KVMAs, s->nr_as);
2071 
2072     kvm_type = qemu_opt_get(qemu_get_machine_opts(), "kvm-type");
2073     if (mc->kvm_type) {
2074         type = mc->kvm_type(ms, kvm_type);
2075     } else if (kvm_type) {
2076         ret = -EINVAL;
2077         fprintf(stderr, "Invalid argument kvm-type=%s\n", kvm_type);
2078         goto err;
2079     }
2080 
2081     do {
2082         ret = kvm_ioctl(s, KVM_CREATE_VM, type);
2083     } while (ret == -EINTR);
2084 
2085     if (ret < 0) {
2086         fprintf(stderr, "ioctl(KVM_CREATE_VM) failed: %d %s\n", -ret,
2087                 strerror(-ret));
2088 
2089 #ifdef TARGET_S390X
2090         if (ret == -EINVAL) {
2091             fprintf(stderr,
2092                     "Host kernel setup problem detected. Please verify:\n");
2093             fprintf(stderr, "- for kernels supporting the switch_amode or"
2094                     " user_mode parameters, whether\n");
2095             fprintf(stderr,
2096                     "  user space is running in primary address space\n");
2097             fprintf(stderr,
2098                     "- for kernels supporting the vm.allocate_pgste sysctl, "
2099                     "whether it is enabled\n");
2100         }
2101 #endif
2102         goto err;
2103     }
2104 
2105     s->vmfd = ret;
2106 
2107     /* check the vcpu limits */
2108     soft_vcpus_limit = kvm_recommended_vcpus(s);
2109     hard_vcpus_limit = kvm_max_vcpus(s);
2110 
2111     while (nc->name) {
2112         if (nc->num > soft_vcpus_limit) {
2113             warn_report("Number of %s cpus requested (%d) exceeds "
2114                         "the recommended cpus supported by KVM (%d)",
2115                         nc->name, nc->num, soft_vcpus_limit);
2116 
2117             if (nc->num > hard_vcpus_limit) {
2118                 fprintf(stderr, "Number of %s cpus requested (%d) exceeds "
2119                         "the maximum cpus supported by KVM (%d)\n",
2120                         nc->name, nc->num, hard_vcpus_limit);
2121                 exit(1);
2122             }
2123         }
2124         nc++;
2125     }
2126 
2127     missing_cap = kvm_check_extension_list(s, kvm_required_capabilites);
2128     if (!missing_cap) {
2129         missing_cap =
2130             kvm_check_extension_list(s, kvm_arch_required_capabilities);
2131     }
2132     if (missing_cap) {
2133         ret = -EINVAL;
2134         fprintf(stderr, "kvm does not support %s\n%s",
2135                 missing_cap->name, upgrade_note);
2136         goto err;
2137     }
2138 
2139     s->coalesced_mmio = kvm_check_extension(s, KVM_CAP_COALESCED_MMIO);
2140     s->coalesced_pio = s->coalesced_mmio &&
2141                        kvm_check_extension(s, KVM_CAP_COALESCED_PIO);
2142 
2143     dirty_log_manual_caps =
2144         kvm_check_extension(s, KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2);
2145     dirty_log_manual_caps &= (KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE |
2146                               KVM_DIRTY_LOG_INITIALLY_SET);
2147     s->manual_dirty_log_protect = dirty_log_manual_caps;
2148     if (dirty_log_manual_caps) {
2149         ret = kvm_vm_enable_cap(s, KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2, 0,
2150                                    dirty_log_manual_caps);
2151         if (ret) {
2152             warn_report("Trying to enable capability %"PRIu64" of "
2153                         "KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2 but failed. "
2154                         "Falling back to the legacy mode. ",
2155                         dirty_log_manual_caps);
2156             s->manual_dirty_log_protect = 0;
2157         }
2158     }
2159 
2160 #ifdef KVM_CAP_VCPU_EVENTS
2161     s->vcpu_events = kvm_check_extension(s, KVM_CAP_VCPU_EVENTS);
2162 #endif
2163 
2164     s->robust_singlestep =
2165         kvm_check_extension(s, KVM_CAP_X86_ROBUST_SINGLESTEP);
2166 
2167 #ifdef KVM_CAP_DEBUGREGS
2168     s->debugregs = kvm_check_extension(s, KVM_CAP_DEBUGREGS);
2169 #endif
2170 
2171     s->max_nested_state_len = kvm_check_extension(s, KVM_CAP_NESTED_STATE);
2172 
2173 #ifdef KVM_CAP_IRQ_ROUTING
2174     kvm_direct_msi_allowed = (kvm_check_extension(s, KVM_CAP_SIGNAL_MSI) > 0);
2175 #endif
2176 
2177     s->intx_set_mask = kvm_check_extension(s, KVM_CAP_PCI_2_3);
2178 
2179     s->irq_set_ioctl = KVM_IRQ_LINE;
2180     if (kvm_check_extension(s, KVM_CAP_IRQ_INJECT_STATUS)) {
2181         s->irq_set_ioctl = KVM_IRQ_LINE_STATUS;
2182     }
2183 
2184     kvm_readonly_mem_allowed =
2185         (kvm_check_extension(s, KVM_CAP_READONLY_MEM) > 0);
2186 
2187     kvm_eventfds_allowed =
2188         (kvm_check_extension(s, KVM_CAP_IOEVENTFD) > 0);
2189 
2190     kvm_irqfds_allowed =
2191         (kvm_check_extension(s, KVM_CAP_IRQFD) > 0);
2192 
2193     kvm_resamplefds_allowed =
2194         (kvm_check_extension(s, KVM_CAP_IRQFD_RESAMPLE) > 0);
2195 
2196     kvm_vm_attributes_allowed =
2197         (kvm_check_extension(s, KVM_CAP_VM_ATTRIBUTES) > 0);
2198 
2199     kvm_ioeventfd_any_length_allowed =
2200         (kvm_check_extension(s, KVM_CAP_IOEVENTFD_ANY_LENGTH) > 0);
2201 
2202     kvm_state = s;
2203 
2204     /*
2205      * if memory encryption object is specified then initialize the memory
2206      * encryption context.
2207      */
2208     if (ms->memory_encryption) {
2209         kvm_state->memcrypt_handle = sev_guest_init(ms->memory_encryption);
2210         if (!kvm_state->memcrypt_handle) {
2211             ret = -1;
2212             goto err;
2213         }
2214 
2215         kvm_state->memcrypt_encrypt_data = sev_encrypt_data;
2216     }
2217 
2218     ret = kvm_arch_init(ms, s);
2219     if (ret < 0) {
2220         goto err;
2221     }
2222 
2223     if (s->kernel_irqchip_split == ON_OFF_AUTO_AUTO) {
2224         s->kernel_irqchip_split = mc->default_kernel_irqchip_split ? ON_OFF_AUTO_ON : ON_OFF_AUTO_OFF;
2225     }
2226 
2227     qemu_register_reset(kvm_unpoison_all, NULL);
2228 
2229     if (s->kernel_irqchip_allowed) {
2230         kvm_irqchip_create(s);
2231     }
2232 
2233     if (kvm_eventfds_allowed) {
2234         s->memory_listener.listener.eventfd_add = kvm_mem_ioeventfd_add;
2235         s->memory_listener.listener.eventfd_del = kvm_mem_ioeventfd_del;
2236     }
2237     s->memory_listener.listener.coalesced_io_add = kvm_coalesce_mmio_region;
2238     s->memory_listener.listener.coalesced_io_del = kvm_uncoalesce_mmio_region;
2239 
2240     kvm_memory_listener_register(s, &s->memory_listener,
2241                                  &address_space_memory, 0);
2242     memory_listener_register(&kvm_io_listener,
2243                              &address_space_io);
2244     memory_listener_register(&kvm_coalesced_pio_listener,
2245                              &address_space_io);
2246 
2247     s->many_ioeventfds = kvm_check_many_ioeventfds();
2248 
2249     s->sync_mmu = !!kvm_vm_check_extension(kvm_state, KVM_CAP_SYNC_MMU);
2250     if (!s->sync_mmu) {
2251         ret = ram_block_discard_disable(true);
2252         assert(!ret);
2253     }
2254 
2255     cpus_register_accel(&kvm_cpus);
2256     return 0;
2257 
2258 err:
2259     assert(ret < 0);
2260     if (s->vmfd >= 0) {
2261         close(s->vmfd);
2262     }
2263     if (s->fd != -1) {
2264         close(s->fd);
2265     }
2266     g_free(s->memory_listener.slots);
2267 
2268     return ret;
2269 }
2270 
2271 void kvm_set_sigmask_len(KVMState *s, unsigned int sigmask_len)
2272 {
2273     s->sigmask_len = sigmask_len;
2274 }
2275 
2276 static void kvm_handle_io(uint16_t port, MemTxAttrs attrs, void *data, int direction,
2277                           int size, uint32_t count)
2278 {
2279     int i;
2280     uint8_t *ptr = data;
2281 
2282     for (i = 0; i < count; i++) {
2283         address_space_rw(&address_space_io, port, attrs,
2284                          ptr, size,
2285                          direction == KVM_EXIT_IO_OUT);
2286         ptr += size;
2287     }
2288 }
2289 
2290 static int kvm_handle_internal_error(CPUState *cpu, struct kvm_run *run)
2291 {
2292     fprintf(stderr, "KVM internal error. Suberror: %d\n",
2293             run->internal.suberror);
2294 
2295     if (kvm_check_extension(kvm_state, KVM_CAP_INTERNAL_ERROR_DATA)) {
2296         int i;
2297 
2298         for (i = 0; i < run->internal.ndata; ++i) {
2299             fprintf(stderr, "extra data[%d]: %"PRIx64"\n",
2300                     i, (uint64_t)run->internal.data[i]);
2301         }
2302     }
2303     if (run->internal.suberror == KVM_INTERNAL_ERROR_EMULATION) {
2304         fprintf(stderr, "emulation failure\n");
2305         if (!kvm_arch_stop_on_emulation_error(cpu)) {
2306             cpu_dump_state(cpu, stderr, CPU_DUMP_CODE);
2307             return EXCP_INTERRUPT;
2308         }
2309     }
2310     /* FIXME: Should trigger a qmp message to let management know
2311      * something went wrong.
2312      */
2313     return -1;
2314 }
2315 
2316 void kvm_flush_coalesced_mmio_buffer(void)
2317 {
2318     KVMState *s = kvm_state;
2319 
2320     if (s->coalesced_flush_in_progress) {
2321         return;
2322     }
2323 
2324     s->coalesced_flush_in_progress = true;
2325 
2326     if (s->coalesced_mmio_ring) {
2327         struct kvm_coalesced_mmio_ring *ring = s->coalesced_mmio_ring;
2328         while (ring->first != ring->last) {
2329             struct kvm_coalesced_mmio *ent;
2330 
2331             ent = &ring->coalesced_mmio[ring->first];
2332 
2333             if (ent->pio == 1) {
2334                 address_space_write(&address_space_io, ent->phys_addr,
2335                                     MEMTXATTRS_UNSPECIFIED, ent->data,
2336                                     ent->len);
2337             } else {
2338                 cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len);
2339             }
2340             smp_wmb();
2341             ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX;
2342         }
2343     }
2344 
2345     s->coalesced_flush_in_progress = false;
2346 }
2347 
2348 static void do_kvm_cpu_synchronize_state(CPUState *cpu, run_on_cpu_data arg)
2349 {
2350     if (!cpu->vcpu_dirty) {
2351         kvm_arch_get_registers(cpu);
2352         cpu->vcpu_dirty = true;
2353     }
2354 }
2355 
2356 void kvm_cpu_synchronize_state(CPUState *cpu)
2357 {
2358     if (!cpu->vcpu_dirty) {
2359         run_on_cpu(cpu, do_kvm_cpu_synchronize_state, RUN_ON_CPU_NULL);
2360     }
2361 }
2362 
2363 static void do_kvm_cpu_synchronize_post_reset(CPUState *cpu, run_on_cpu_data arg)
2364 {
2365     kvm_arch_put_registers(cpu, KVM_PUT_RESET_STATE);
2366     cpu->vcpu_dirty = false;
2367 }
2368 
2369 void kvm_cpu_synchronize_post_reset(CPUState *cpu)
2370 {
2371     run_on_cpu(cpu, do_kvm_cpu_synchronize_post_reset, RUN_ON_CPU_NULL);
2372 }
2373 
2374 static void do_kvm_cpu_synchronize_post_init(CPUState *cpu, run_on_cpu_data arg)
2375 {
2376     kvm_arch_put_registers(cpu, KVM_PUT_FULL_STATE);
2377     cpu->vcpu_dirty = false;
2378 }
2379 
2380 void kvm_cpu_synchronize_post_init(CPUState *cpu)
2381 {
2382     run_on_cpu(cpu, do_kvm_cpu_synchronize_post_init, RUN_ON_CPU_NULL);
2383 }
2384 
2385 static void do_kvm_cpu_synchronize_pre_loadvm(CPUState *cpu, run_on_cpu_data arg)
2386 {
2387     cpu->vcpu_dirty = true;
2388 }
2389 
2390 void kvm_cpu_synchronize_pre_loadvm(CPUState *cpu)
2391 {
2392     run_on_cpu(cpu, do_kvm_cpu_synchronize_pre_loadvm, RUN_ON_CPU_NULL);
2393 }
2394 
2395 #ifdef KVM_HAVE_MCE_INJECTION
2396 static __thread void *pending_sigbus_addr;
2397 static __thread int pending_sigbus_code;
2398 static __thread bool have_sigbus_pending;
2399 #endif
2400 
2401 static void kvm_cpu_kick(CPUState *cpu)
2402 {
2403     qatomic_set(&cpu->kvm_run->immediate_exit, 1);
2404 }
2405 
2406 static void kvm_cpu_kick_self(void)
2407 {
2408     if (kvm_immediate_exit) {
2409         kvm_cpu_kick(current_cpu);
2410     } else {
2411         qemu_cpu_kick_self();
2412     }
2413 }
2414 
2415 static void kvm_eat_signals(CPUState *cpu)
2416 {
2417     struct timespec ts = { 0, 0 };
2418     siginfo_t siginfo;
2419     sigset_t waitset;
2420     sigset_t chkset;
2421     int r;
2422 
2423     if (kvm_immediate_exit) {
2424         qatomic_set(&cpu->kvm_run->immediate_exit, 0);
2425         /* Write kvm_run->immediate_exit before the cpu->exit_request
2426          * write in kvm_cpu_exec.
2427          */
2428         smp_wmb();
2429         return;
2430     }
2431 
2432     sigemptyset(&waitset);
2433     sigaddset(&waitset, SIG_IPI);
2434 
2435     do {
2436         r = sigtimedwait(&waitset, &siginfo, &ts);
2437         if (r == -1 && !(errno == EAGAIN || errno == EINTR)) {
2438             perror("sigtimedwait");
2439             exit(1);
2440         }
2441 
2442         r = sigpending(&chkset);
2443         if (r == -1) {
2444             perror("sigpending");
2445             exit(1);
2446         }
2447     } while (sigismember(&chkset, SIG_IPI));
2448 }
2449 
2450 int kvm_cpu_exec(CPUState *cpu)
2451 {
2452     struct kvm_run *run = cpu->kvm_run;
2453     int ret, run_ret;
2454 
2455     DPRINTF("kvm_cpu_exec()\n");
2456 
2457     if (kvm_arch_process_async_events(cpu)) {
2458         qatomic_set(&cpu->exit_request, 0);
2459         return EXCP_HLT;
2460     }
2461 
2462     qemu_mutex_unlock_iothread();
2463     cpu_exec_start(cpu);
2464 
2465     do {
2466         MemTxAttrs attrs;
2467 
2468         if (cpu->vcpu_dirty) {
2469             kvm_arch_put_registers(cpu, KVM_PUT_RUNTIME_STATE);
2470             cpu->vcpu_dirty = false;
2471         }
2472 
2473         kvm_arch_pre_run(cpu, run);
2474         if (qatomic_read(&cpu->exit_request)) {
2475             DPRINTF("interrupt exit requested\n");
2476             /*
2477              * KVM requires us to reenter the kernel after IO exits to complete
2478              * instruction emulation. This self-signal will ensure that we
2479              * leave ASAP again.
2480              */
2481             kvm_cpu_kick_self();
2482         }
2483 
2484         /* Read cpu->exit_request before KVM_RUN reads run->immediate_exit.
2485          * Matching barrier in kvm_eat_signals.
2486          */
2487         smp_rmb();
2488 
2489         run_ret = kvm_vcpu_ioctl(cpu, KVM_RUN, 0);
2490 
2491         attrs = kvm_arch_post_run(cpu, run);
2492 
2493 #ifdef KVM_HAVE_MCE_INJECTION
2494         if (unlikely(have_sigbus_pending)) {
2495             qemu_mutex_lock_iothread();
2496             kvm_arch_on_sigbus_vcpu(cpu, pending_sigbus_code,
2497                                     pending_sigbus_addr);
2498             have_sigbus_pending = false;
2499             qemu_mutex_unlock_iothread();
2500         }
2501 #endif
2502 
2503         if (run_ret < 0) {
2504             if (run_ret == -EINTR || run_ret == -EAGAIN) {
2505                 DPRINTF("io window exit\n");
2506                 kvm_eat_signals(cpu);
2507                 ret = EXCP_INTERRUPT;
2508                 break;
2509             }
2510             fprintf(stderr, "error: kvm run failed %s\n",
2511                     strerror(-run_ret));
2512 #ifdef TARGET_PPC
2513             if (run_ret == -EBUSY) {
2514                 fprintf(stderr,
2515                         "This is probably because your SMT is enabled.\n"
2516                         "VCPU can only run on primary threads with all "
2517                         "secondary threads offline.\n");
2518             }
2519 #endif
2520             ret = -1;
2521             break;
2522         }
2523 
2524         trace_kvm_run_exit(cpu->cpu_index, run->exit_reason);
2525         switch (run->exit_reason) {
2526         case KVM_EXIT_IO:
2527             DPRINTF("handle_io\n");
2528             /* Called outside BQL */
2529             kvm_handle_io(run->io.port, attrs,
2530                           (uint8_t *)run + run->io.data_offset,
2531                           run->io.direction,
2532                           run->io.size,
2533                           run->io.count);
2534             ret = 0;
2535             break;
2536         case KVM_EXIT_MMIO:
2537             DPRINTF("handle_mmio\n");
2538             /* Called outside BQL */
2539             address_space_rw(&address_space_memory,
2540                              run->mmio.phys_addr, attrs,
2541                              run->mmio.data,
2542                              run->mmio.len,
2543                              run->mmio.is_write);
2544             ret = 0;
2545             break;
2546         case KVM_EXIT_IRQ_WINDOW_OPEN:
2547             DPRINTF("irq_window_open\n");
2548             ret = EXCP_INTERRUPT;
2549             break;
2550         case KVM_EXIT_SHUTDOWN:
2551             DPRINTF("shutdown\n");
2552             qemu_system_reset_request(SHUTDOWN_CAUSE_GUEST_RESET);
2553             ret = EXCP_INTERRUPT;
2554             break;
2555         case KVM_EXIT_UNKNOWN:
2556             fprintf(stderr, "KVM: unknown exit, hardware reason %" PRIx64 "\n",
2557                     (uint64_t)run->hw.hardware_exit_reason);
2558             ret = -1;
2559             break;
2560         case KVM_EXIT_INTERNAL_ERROR:
2561             ret = kvm_handle_internal_error(cpu, run);
2562             break;
2563         case KVM_EXIT_SYSTEM_EVENT:
2564             switch (run->system_event.type) {
2565             case KVM_SYSTEM_EVENT_SHUTDOWN:
2566                 qemu_system_shutdown_request(SHUTDOWN_CAUSE_GUEST_SHUTDOWN);
2567                 ret = EXCP_INTERRUPT;
2568                 break;
2569             case KVM_SYSTEM_EVENT_RESET:
2570                 qemu_system_reset_request(SHUTDOWN_CAUSE_GUEST_RESET);
2571                 ret = EXCP_INTERRUPT;
2572                 break;
2573             case KVM_SYSTEM_EVENT_CRASH:
2574                 kvm_cpu_synchronize_state(cpu);
2575                 qemu_mutex_lock_iothread();
2576                 qemu_system_guest_panicked(cpu_get_crash_info(cpu));
2577                 qemu_mutex_unlock_iothread();
2578                 ret = 0;
2579                 break;
2580             default:
2581                 DPRINTF("kvm_arch_handle_exit\n");
2582                 ret = kvm_arch_handle_exit(cpu, run);
2583                 break;
2584             }
2585             break;
2586         default:
2587             DPRINTF("kvm_arch_handle_exit\n");
2588             ret = kvm_arch_handle_exit(cpu, run);
2589             break;
2590         }
2591     } while (ret == 0);
2592 
2593     cpu_exec_end(cpu);
2594     qemu_mutex_lock_iothread();
2595 
2596     if (ret < 0) {
2597         cpu_dump_state(cpu, stderr, CPU_DUMP_CODE);
2598         vm_stop(RUN_STATE_INTERNAL_ERROR);
2599     }
2600 
2601     qatomic_set(&cpu->exit_request, 0);
2602     return ret;
2603 }
2604 
2605 int kvm_ioctl(KVMState *s, int type, ...)
2606 {
2607     int ret;
2608     void *arg;
2609     va_list ap;
2610 
2611     va_start(ap, type);
2612     arg = va_arg(ap, void *);
2613     va_end(ap);
2614 
2615     trace_kvm_ioctl(type, arg);
2616     ret = ioctl(s->fd, type, arg);
2617     if (ret == -1) {
2618         ret = -errno;
2619     }
2620     return ret;
2621 }
2622 
2623 int kvm_vm_ioctl(KVMState *s, int type, ...)
2624 {
2625     int ret;
2626     void *arg;
2627     va_list ap;
2628 
2629     va_start(ap, type);
2630     arg = va_arg(ap, void *);
2631     va_end(ap);
2632 
2633     trace_kvm_vm_ioctl(type, arg);
2634     ret = ioctl(s->vmfd, type, arg);
2635     if (ret == -1) {
2636         ret = -errno;
2637     }
2638     return ret;
2639 }
2640 
2641 int kvm_vcpu_ioctl(CPUState *cpu, int type, ...)
2642 {
2643     int ret;
2644     void *arg;
2645     va_list ap;
2646 
2647     va_start(ap, type);
2648     arg = va_arg(ap, void *);
2649     va_end(ap);
2650 
2651     trace_kvm_vcpu_ioctl(cpu->cpu_index, type, arg);
2652     ret = ioctl(cpu->kvm_fd, type, arg);
2653     if (ret == -1) {
2654         ret = -errno;
2655     }
2656     return ret;
2657 }
2658 
2659 int kvm_device_ioctl(int fd, int type, ...)
2660 {
2661     int ret;
2662     void *arg;
2663     va_list ap;
2664 
2665     va_start(ap, type);
2666     arg = va_arg(ap, void *);
2667     va_end(ap);
2668 
2669     trace_kvm_device_ioctl(fd, type, arg);
2670     ret = ioctl(fd, type, arg);
2671     if (ret == -1) {
2672         ret = -errno;
2673     }
2674     return ret;
2675 }
2676 
2677 int kvm_vm_check_attr(KVMState *s, uint32_t group, uint64_t attr)
2678 {
2679     int ret;
2680     struct kvm_device_attr attribute = {
2681         .group = group,
2682         .attr = attr,
2683     };
2684 
2685     if (!kvm_vm_attributes_allowed) {
2686         return 0;
2687     }
2688 
2689     ret = kvm_vm_ioctl(s, KVM_HAS_DEVICE_ATTR, &attribute);
2690     /* kvm returns 0 on success for HAS_DEVICE_ATTR */
2691     return ret ? 0 : 1;
2692 }
2693 
2694 int kvm_device_check_attr(int dev_fd, uint32_t group, uint64_t attr)
2695 {
2696     struct kvm_device_attr attribute = {
2697         .group = group,
2698         .attr = attr,
2699         .flags = 0,
2700     };
2701 
2702     return kvm_device_ioctl(dev_fd, KVM_HAS_DEVICE_ATTR, &attribute) ? 0 : 1;
2703 }
2704 
2705 int kvm_device_access(int fd, int group, uint64_t attr,
2706                       void *val, bool write, Error **errp)
2707 {
2708     struct kvm_device_attr kvmattr;
2709     int err;
2710 
2711     kvmattr.flags = 0;
2712     kvmattr.group = group;
2713     kvmattr.attr = attr;
2714     kvmattr.addr = (uintptr_t)val;
2715 
2716     err = kvm_device_ioctl(fd,
2717                            write ? KVM_SET_DEVICE_ATTR : KVM_GET_DEVICE_ATTR,
2718                            &kvmattr);
2719     if (err < 0) {
2720         error_setg_errno(errp, -err,
2721                          "KVM_%s_DEVICE_ATTR failed: Group %d "
2722                          "attr 0x%016" PRIx64,
2723                          write ? "SET" : "GET", group, attr);
2724     }
2725     return err;
2726 }
2727 
2728 bool kvm_has_sync_mmu(void)
2729 {
2730     return kvm_state->sync_mmu;
2731 }
2732 
2733 int kvm_has_vcpu_events(void)
2734 {
2735     return kvm_state->vcpu_events;
2736 }
2737 
2738 int kvm_has_robust_singlestep(void)
2739 {
2740     return kvm_state->robust_singlestep;
2741 }
2742 
2743 int kvm_has_debugregs(void)
2744 {
2745     return kvm_state->debugregs;
2746 }
2747 
2748 int kvm_max_nested_state_length(void)
2749 {
2750     return kvm_state->max_nested_state_len;
2751 }
2752 
2753 int kvm_has_many_ioeventfds(void)
2754 {
2755     if (!kvm_enabled()) {
2756         return 0;
2757     }
2758     return kvm_state->many_ioeventfds;
2759 }
2760 
2761 int kvm_has_gsi_routing(void)
2762 {
2763 #ifdef KVM_CAP_IRQ_ROUTING
2764     return kvm_check_extension(kvm_state, KVM_CAP_IRQ_ROUTING);
2765 #else
2766     return false;
2767 #endif
2768 }
2769 
2770 int kvm_has_intx_set_mask(void)
2771 {
2772     return kvm_state->intx_set_mask;
2773 }
2774 
2775 bool kvm_arm_supports_user_irq(void)
2776 {
2777     return kvm_check_extension(kvm_state, KVM_CAP_ARM_USER_IRQ);
2778 }
2779 
2780 #ifdef KVM_CAP_SET_GUEST_DEBUG
2781 struct kvm_sw_breakpoint *kvm_find_sw_breakpoint(CPUState *cpu,
2782                                                  target_ulong pc)
2783 {
2784     struct kvm_sw_breakpoint *bp;
2785 
2786     QTAILQ_FOREACH(bp, &cpu->kvm_state->kvm_sw_breakpoints, entry) {
2787         if (bp->pc == pc) {
2788             return bp;
2789         }
2790     }
2791     return NULL;
2792 }
2793 
2794 int kvm_sw_breakpoints_active(CPUState *cpu)
2795 {
2796     return !QTAILQ_EMPTY(&cpu->kvm_state->kvm_sw_breakpoints);
2797 }
2798 
2799 struct kvm_set_guest_debug_data {
2800     struct kvm_guest_debug dbg;
2801     int err;
2802 };
2803 
2804 static void kvm_invoke_set_guest_debug(CPUState *cpu, run_on_cpu_data data)
2805 {
2806     struct kvm_set_guest_debug_data *dbg_data =
2807         (struct kvm_set_guest_debug_data *) data.host_ptr;
2808 
2809     dbg_data->err = kvm_vcpu_ioctl(cpu, KVM_SET_GUEST_DEBUG,
2810                                    &dbg_data->dbg);
2811 }
2812 
2813 int kvm_update_guest_debug(CPUState *cpu, unsigned long reinject_trap)
2814 {
2815     struct kvm_set_guest_debug_data data;
2816 
2817     data.dbg.control = reinject_trap;
2818 
2819     if (cpu->singlestep_enabled) {
2820         data.dbg.control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_SINGLESTEP;
2821     }
2822     kvm_arch_update_guest_debug(cpu, &data.dbg);
2823 
2824     run_on_cpu(cpu, kvm_invoke_set_guest_debug,
2825                RUN_ON_CPU_HOST_PTR(&data));
2826     return data.err;
2827 }
2828 
2829 int kvm_insert_breakpoint(CPUState *cpu, target_ulong addr,
2830                           target_ulong len, int type)
2831 {
2832     struct kvm_sw_breakpoint *bp;
2833     int err;
2834 
2835     if (type == GDB_BREAKPOINT_SW) {
2836         bp = kvm_find_sw_breakpoint(cpu, addr);
2837         if (bp) {
2838             bp->use_count++;
2839             return 0;
2840         }
2841 
2842         bp = g_malloc(sizeof(struct kvm_sw_breakpoint));
2843         bp->pc = addr;
2844         bp->use_count = 1;
2845         err = kvm_arch_insert_sw_breakpoint(cpu, bp);
2846         if (err) {
2847             g_free(bp);
2848             return err;
2849         }
2850 
2851         QTAILQ_INSERT_HEAD(&cpu->kvm_state->kvm_sw_breakpoints, bp, entry);
2852     } else {
2853         err = kvm_arch_insert_hw_breakpoint(addr, len, type);
2854         if (err) {
2855             return err;
2856         }
2857     }
2858 
2859     CPU_FOREACH(cpu) {
2860         err = kvm_update_guest_debug(cpu, 0);
2861         if (err) {
2862             return err;
2863         }
2864     }
2865     return 0;
2866 }
2867 
2868 int kvm_remove_breakpoint(CPUState *cpu, target_ulong addr,
2869                           target_ulong len, int type)
2870 {
2871     struct kvm_sw_breakpoint *bp;
2872     int err;
2873 
2874     if (type == GDB_BREAKPOINT_SW) {
2875         bp = kvm_find_sw_breakpoint(cpu, addr);
2876         if (!bp) {
2877             return -ENOENT;
2878         }
2879 
2880         if (bp->use_count > 1) {
2881             bp->use_count--;
2882             return 0;
2883         }
2884 
2885         err = kvm_arch_remove_sw_breakpoint(cpu, bp);
2886         if (err) {
2887             return err;
2888         }
2889 
2890         QTAILQ_REMOVE(&cpu->kvm_state->kvm_sw_breakpoints, bp, entry);
2891         g_free(bp);
2892     } else {
2893         err = kvm_arch_remove_hw_breakpoint(addr, len, type);
2894         if (err) {
2895             return err;
2896         }
2897     }
2898 
2899     CPU_FOREACH(cpu) {
2900         err = kvm_update_guest_debug(cpu, 0);
2901         if (err) {
2902             return err;
2903         }
2904     }
2905     return 0;
2906 }
2907 
2908 void kvm_remove_all_breakpoints(CPUState *cpu)
2909 {
2910     struct kvm_sw_breakpoint *bp, *next;
2911     KVMState *s = cpu->kvm_state;
2912     CPUState *tmpcpu;
2913 
2914     QTAILQ_FOREACH_SAFE(bp, &s->kvm_sw_breakpoints, entry, next) {
2915         if (kvm_arch_remove_sw_breakpoint(cpu, bp) != 0) {
2916             /* Try harder to find a CPU that currently sees the breakpoint. */
2917             CPU_FOREACH(tmpcpu) {
2918                 if (kvm_arch_remove_sw_breakpoint(tmpcpu, bp) == 0) {
2919                     break;
2920                 }
2921             }
2922         }
2923         QTAILQ_REMOVE(&s->kvm_sw_breakpoints, bp, entry);
2924         g_free(bp);
2925     }
2926     kvm_arch_remove_all_hw_breakpoints();
2927 
2928     CPU_FOREACH(cpu) {
2929         kvm_update_guest_debug(cpu, 0);
2930     }
2931 }
2932 
2933 #else /* !KVM_CAP_SET_GUEST_DEBUG */
2934 
2935 int kvm_update_guest_debug(CPUState *cpu, unsigned long reinject_trap)
2936 {
2937     return -EINVAL;
2938 }
2939 
2940 int kvm_insert_breakpoint(CPUState *cpu, target_ulong addr,
2941                           target_ulong len, int type)
2942 {
2943     return -EINVAL;
2944 }
2945 
2946 int kvm_remove_breakpoint(CPUState *cpu, target_ulong addr,
2947                           target_ulong len, int type)
2948 {
2949     return -EINVAL;
2950 }
2951 
2952 void kvm_remove_all_breakpoints(CPUState *cpu)
2953 {
2954 }
2955 #endif /* !KVM_CAP_SET_GUEST_DEBUG */
2956 
2957 static int kvm_set_signal_mask(CPUState *cpu, const sigset_t *sigset)
2958 {
2959     KVMState *s = kvm_state;
2960     struct kvm_signal_mask *sigmask;
2961     int r;
2962 
2963     sigmask = g_malloc(sizeof(*sigmask) + sizeof(*sigset));
2964 
2965     sigmask->len = s->sigmask_len;
2966     memcpy(sigmask->sigset, sigset, sizeof(*sigset));
2967     r = kvm_vcpu_ioctl(cpu, KVM_SET_SIGNAL_MASK, sigmask);
2968     g_free(sigmask);
2969 
2970     return r;
2971 }
2972 
2973 static void kvm_ipi_signal(int sig)
2974 {
2975     if (current_cpu) {
2976         assert(kvm_immediate_exit);
2977         kvm_cpu_kick(current_cpu);
2978     }
2979 }
2980 
2981 void kvm_init_cpu_signals(CPUState *cpu)
2982 {
2983     int r;
2984     sigset_t set;
2985     struct sigaction sigact;
2986 
2987     memset(&sigact, 0, sizeof(sigact));
2988     sigact.sa_handler = kvm_ipi_signal;
2989     sigaction(SIG_IPI, &sigact, NULL);
2990 
2991     pthread_sigmask(SIG_BLOCK, NULL, &set);
2992 #if defined KVM_HAVE_MCE_INJECTION
2993     sigdelset(&set, SIGBUS);
2994     pthread_sigmask(SIG_SETMASK, &set, NULL);
2995 #endif
2996     sigdelset(&set, SIG_IPI);
2997     if (kvm_immediate_exit) {
2998         r = pthread_sigmask(SIG_SETMASK, &set, NULL);
2999     } else {
3000         r = kvm_set_signal_mask(cpu, &set);
3001     }
3002     if (r) {
3003         fprintf(stderr, "kvm_set_signal_mask: %s\n", strerror(-r));
3004         exit(1);
3005     }
3006 }
3007 
3008 /* Called asynchronously in VCPU thread.  */
3009 int kvm_on_sigbus_vcpu(CPUState *cpu, int code, void *addr)
3010 {
3011 #ifdef KVM_HAVE_MCE_INJECTION
3012     if (have_sigbus_pending) {
3013         return 1;
3014     }
3015     have_sigbus_pending = true;
3016     pending_sigbus_addr = addr;
3017     pending_sigbus_code = code;
3018     qatomic_set(&cpu->exit_request, 1);
3019     return 0;
3020 #else
3021     return 1;
3022 #endif
3023 }
3024 
3025 /* Called synchronously (via signalfd) in main thread.  */
3026 int kvm_on_sigbus(int code, void *addr)
3027 {
3028 #ifdef KVM_HAVE_MCE_INJECTION
3029     /* Action required MCE kills the process if SIGBUS is blocked.  Because
3030      * that's what happens in the I/O thread, where we handle MCE via signalfd,
3031      * we can only get action optional here.
3032      */
3033     assert(code != BUS_MCEERR_AR);
3034     kvm_arch_on_sigbus_vcpu(first_cpu, code, addr);
3035     return 0;
3036 #else
3037     return 1;
3038 #endif
3039 }
3040 
3041 int kvm_create_device(KVMState *s, uint64_t type, bool test)
3042 {
3043     int ret;
3044     struct kvm_create_device create_dev;
3045 
3046     create_dev.type = type;
3047     create_dev.fd = -1;
3048     create_dev.flags = test ? KVM_CREATE_DEVICE_TEST : 0;
3049 
3050     if (!kvm_check_extension(s, KVM_CAP_DEVICE_CTRL)) {
3051         return -ENOTSUP;
3052     }
3053 
3054     ret = kvm_vm_ioctl(s, KVM_CREATE_DEVICE, &create_dev);
3055     if (ret) {
3056         return ret;
3057     }
3058 
3059     return test ? 0 : create_dev.fd;
3060 }
3061 
3062 bool kvm_device_supported(int vmfd, uint64_t type)
3063 {
3064     struct kvm_create_device create_dev = {
3065         .type = type,
3066         .fd = -1,
3067         .flags = KVM_CREATE_DEVICE_TEST,
3068     };
3069 
3070     if (ioctl(vmfd, KVM_CHECK_EXTENSION, KVM_CAP_DEVICE_CTRL) <= 0) {
3071         return false;
3072     }
3073 
3074     return (ioctl(vmfd, KVM_CREATE_DEVICE, &create_dev) >= 0);
3075 }
3076 
3077 int kvm_set_one_reg(CPUState *cs, uint64_t id, void *source)
3078 {
3079     struct kvm_one_reg reg;
3080     int r;
3081 
3082     reg.id = id;
3083     reg.addr = (uintptr_t) source;
3084     r = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, &reg);
3085     if (r) {
3086         trace_kvm_failed_reg_set(id, strerror(-r));
3087     }
3088     return r;
3089 }
3090 
3091 int kvm_get_one_reg(CPUState *cs, uint64_t id, void *target)
3092 {
3093     struct kvm_one_reg reg;
3094     int r;
3095 
3096     reg.id = id;
3097     reg.addr = (uintptr_t) target;
3098     r = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &reg);
3099     if (r) {
3100         trace_kvm_failed_reg_get(id, strerror(-r));
3101     }
3102     return r;
3103 }
3104 
3105 static bool kvm_accel_has_memory(MachineState *ms, AddressSpace *as,
3106                                  hwaddr start_addr, hwaddr size)
3107 {
3108     KVMState *kvm = KVM_STATE(ms->accelerator);
3109     int i;
3110 
3111     for (i = 0; i < kvm->nr_as; ++i) {
3112         if (kvm->as[i].as == as && kvm->as[i].ml) {
3113             size = MIN(kvm_max_slot_size, size);
3114             return NULL != kvm_lookup_matching_slot(kvm->as[i].ml,
3115                                                     start_addr, size);
3116         }
3117     }
3118 
3119     return false;
3120 }
3121 
3122 static void kvm_get_kvm_shadow_mem(Object *obj, Visitor *v,
3123                                    const char *name, void *opaque,
3124                                    Error **errp)
3125 {
3126     KVMState *s = KVM_STATE(obj);
3127     int64_t value = s->kvm_shadow_mem;
3128 
3129     visit_type_int(v, name, &value, errp);
3130 }
3131 
3132 static void kvm_set_kvm_shadow_mem(Object *obj, Visitor *v,
3133                                    const char *name, void *opaque,
3134                                    Error **errp)
3135 {
3136     KVMState *s = KVM_STATE(obj);
3137     int64_t value;
3138 
3139     if (!visit_type_int(v, name, &value, errp)) {
3140         return;
3141     }
3142 
3143     s->kvm_shadow_mem = value;
3144 }
3145 
3146 static void kvm_set_kernel_irqchip(Object *obj, Visitor *v,
3147                                    const char *name, void *opaque,
3148                                    Error **errp)
3149 {
3150     KVMState *s = KVM_STATE(obj);
3151     OnOffSplit mode;
3152 
3153     if (!visit_type_OnOffSplit(v, name, &mode, errp)) {
3154         return;
3155     }
3156     switch (mode) {
3157     case ON_OFF_SPLIT_ON:
3158         s->kernel_irqchip_allowed = true;
3159         s->kernel_irqchip_required = true;
3160         s->kernel_irqchip_split = ON_OFF_AUTO_OFF;
3161         break;
3162     case ON_OFF_SPLIT_OFF:
3163         s->kernel_irqchip_allowed = false;
3164         s->kernel_irqchip_required = false;
3165         s->kernel_irqchip_split = ON_OFF_AUTO_OFF;
3166         break;
3167     case ON_OFF_SPLIT_SPLIT:
3168         s->kernel_irqchip_allowed = true;
3169         s->kernel_irqchip_required = true;
3170         s->kernel_irqchip_split = ON_OFF_AUTO_ON;
3171         break;
3172     default:
3173         /* The value was checked in visit_type_OnOffSplit() above. If
3174          * we get here, then something is wrong in QEMU.
3175          */
3176         abort();
3177     }
3178 }
3179 
3180 bool kvm_kernel_irqchip_allowed(void)
3181 {
3182     return kvm_state->kernel_irqchip_allowed;
3183 }
3184 
3185 bool kvm_kernel_irqchip_required(void)
3186 {
3187     return kvm_state->kernel_irqchip_required;
3188 }
3189 
3190 bool kvm_kernel_irqchip_split(void)
3191 {
3192     return kvm_state->kernel_irqchip_split == ON_OFF_AUTO_ON;
3193 }
3194 
3195 static void kvm_accel_instance_init(Object *obj)
3196 {
3197     KVMState *s = KVM_STATE(obj);
3198 
3199     s->kvm_shadow_mem = -1;
3200     s->kernel_irqchip_allowed = true;
3201     s->kernel_irqchip_split = ON_OFF_AUTO_AUTO;
3202 }
3203 
3204 static void kvm_accel_class_init(ObjectClass *oc, void *data)
3205 {
3206     AccelClass *ac = ACCEL_CLASS(oc);
3207     ac->name = "KVM";
3208     ac->init_machine = kvm_init;
3209     ac->has_memory = kvm_accel_has_memory;
3210     ac->allowed = &kvm_allowed;
3211 
3212     object_class_property_add(oc, "kernel-irqchip", "on|off|split",
3213         NULL, kvm_set_kernel_irqchip,
3214         NULL, NULL);
3215     object_class_property_set_description(oc, "kernel-irqchip",
3216         "Configure KVM in-kernel irqchip");
3217 
3218     object_class_property_add(oc, "kvm-shadow-mem", "int",
3219         kvm_get_kvm_shadow_mem, kvm_set_kvm_shadow_mem,
3220         NULL, NULL);
3221     object_class_property_set_description(oc, "kvm-shadow-mem",
3222         "KVM shadow MMU size");
3223 }
3224 
3225 static const TypeInfo kvm_accel_type = {
3226     .name = TYPE_KVM_ACCEL,
3227     .parent = TYPE_ACCEL,
3228     .instance_init = kvm_accel_instance_init,
3229     .class_init = kvm_accel_class_init,
3230     .instance_size = sizeof(KVMState),
3231 };
3232 
3233 static void kvm_type_init(void)
3234 {
3235     type_register_static(&kvm_accel_type);
3236 }
3237 
3238 type_init(kvm_type_init);
3239