xref: /openbmc/qemu/accel/kvm/kvm-all.c (revision 5de5b99b)
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 
48 #include "hw/boards.h"
49 
50 /* This check must be after config-host.h is included */
51 #ifdef CONFIG_EVENTFD
52 #include <sys/eventfd.h>
53 #endif
54 
55 /* KVM uses PAGE_SIZE in its definition of KVM_COALESCED_MMIO_MAX. We
56  * need to use the real host PAGE_SIZE, as that's what KVM will use.
57  */
58 #define PAGE_SIZE qemu_real_host_page_size
59 
60 //#define DEBUG_KVM
61 
62 #ifdef DEBUG_KVM
63 #define DPRINTF(fmt, ...) \
64     do { fprintf(stderr, fmt, ## __VA_ARGS__); } while (0)
65 #else
66 #define DPRINTF(fmt, ...) \
67     do { } while (0)
68 #endif
69 
70 #define KVM_MSI_HASHTAB_SIZE    256
71 
72 struct KVMParkedVcpu {
73     unsigned long vcpu_id;
74     int kvm_fd;
75     QLIST_ENTRY(KVMParkedVcpu) node;
76 };
77 
78 struct KVMState
79 {
80     AccelState parent_obj;
81 
82     int nr_slots;
83     int fd;
84     int vmfd;
85     int coalesced_mmio;
86     int coalesced_pio;
87     struct kvm_coalesced_mmio_ring *coalesced_mmio_ring;
88     bool coalesced_flush_in_progress;
89     int vcpu_events;
90     int robust_singlestep;
91     int debugregs;
92 #ifdef KVM_CAP_SET_GUEST_DEBUG
93     QTAILQ_HEAD(, kvm_sw_breakpoint) kvm_sw_breakpoints;
94 #endif
95     int max_nested_state_len;
96     int many_ioeventfds;
97     int intx_set_mask;
98     int kvm_shadow_mem;
99     bool kernel_irqchip_allowed;
100     bool kernel_irqchip_required;
101     OnOffAuto kernel_irqchip_split;
102     bool sync_mmu;
103     uint64_t manual_dirty_log_protect;
104     /* The man page (and posix) say ioctl numbers are signed int, but
105      * they're not.  Linux, glibc and *BSD all treat ioctl numbers as
106      * unsigned, and treating them as signed here can break things */
107     unsigned irq_set_ioctl;
108     unsigned int sigmask_len;
109     GHashTable *gsimap;
110 #ifdef KVM_CAP_IRQ_ROUTING
111     struct kvm_irq_routing *irq_routes;
112     int nr_allocated_irq_routes;
113     unsigned long *used_gsi_bitmap;
114     unsigned int gsi_count;
115     QTAILQ_HEAD(, KVMMSIRoute) msi_hashtab[KVM_MSI_HASHTAB_SIZE];
116 #endif
117     KVMMemoryListener memory_listener;
118     QLIST_HEAD(, KVMParkedVcpu) kvm_parked_vcpus;
119 
120     /* memory encryption */
121     void *memcrypt_handle;
122     int (*memcrypt_encrypt_data)(void *handle, uint8_t *ptr, uint64_t len);
123 
124     /* For "info mtree -f" to tell if an MR is registered in KVM */
125     int nr_as;
126     struct KVMAs {
127         KVMMemoryListener *ml;
128         AddressSpace *as;
129     } *as;
130 };
131 
132 KVMState *kvm_state;
133 bool kvm_kernel_irqchip;
134 bool kvm_split_irqchip;
135 bool kvm_async_interrupts_allowed;
136 bool kvm_halt_in_kernel_allowed;
137 bool kvm_eventfds_allowed;
138 bool kvm_irqfds_allowed;
139 bool kvm_resamplefds_allowed;
140 bool kvm_msi_via_irqfd_allowed;
141 bool kvm_gsi_routing_allowed;
142 bool kvm_gsi_direct_mapping;
143 bool kvm_allowed;
144 bool kvm_readonly_mem_allowed;
145 bool kvm_vm_attributes_allowed;
146 bool kvm_direct_msi_allowed;
147 bool kvm_ioeventfd_any_length_allowed;
148 bool kvm_msi_use_devid;
149 static bool kvm_immediate_exit;
150 static hwaddr kvm_max_slot_size = ~0;
151 
152 static const KVMCapabilityInfo kvm_required_capabilites[] = {
153     KVM_CAP_INFO(USER_MEMORY),
154     KVM_CAP_INFO(DESTROY_MEMORY_REGION_WORKS),
155     KVM_CAP_INFO(JOIN_MEMORY_REGIONS_WORKS),
156     KVM_CAP_LAST_INFO
157 };
158 
159 static NotifierList kvm_irqchip_change_notifiers =
160     NOTIFIER_LIST_INITIALIZER(kvm_irqchip_change_notifiers);
161 
162 struct KVMResampleFd {
163     int gsi;
164     EventNotifier *resample_event;
165     QLIST_ENTRY(KVMResampleFd) node;
166 };
167 typedef struct KVMResampleFd KVMResampleFd;
168 
169 /*
170  * Only used with split irqchip where we need to do the resample fd
171  * kick for the kernel from userspace.
172  */
173 static QLIST_HEAD(, KVMResampleFd) kvm_resample_fd_list =
174     QLIST_HEAD_INITIALIZER(kvm_resample_fd_list);
175 
176 #define kvm_slots_lock(kml)      qemu_mutex_lock(&(kml)->slots_lock)
177 #define kvm_slots_unlock(kml)    qemu_mutex_unlock(&(kml)->slots_lock)
178 
179 static inline void kvm_resample_fd_remove(int gsi)
180 {
181     KVMResampleFd *rfd;
182 
183     QLIST_FOREACH(rfd, &kvm_resample_fd_list, node) {
184         if (rfd->gsi == gsi) {
185             QLIST_REMOVE(rfd, node);
186             g_free(rfd);
187             break;
188         }
189     }
190 }
191 
192 static inline void kvm_resample_fd_insert(int gsi, EventNotifier *event)
193 {
194     KVMResampleFd *rfd = g_new0(KVMResampleFd, 1);
195 
196     rfd->gsi = gsi;
197     rfd->resample_event = event;
198 
199     QLIST_INSERT_HEAD(&kvm_resample_fd_list, rfd, node);
200 }
201 
202 void kvm_resample_fd_notify(int gsi)
203 {
204     KVMResampleFd *rfd;
205 
206     QLIST_FOREACH(rfd, &kvm_resample_fd_list, node) {
207         if (rfd->gsi == gsi) {
208             event_notifier_set(rfd->resample_event);
209             trace_kvm_resample_fd_notify(gsi);
210             return;
211         }
212     }
213 }
214 
215 int kvm_get_max_memslots(void)
216 {
217     KVMState *s = KVM_STATE(current_accel());
218 
219     return s->nr_slots;
220 }
221 
222 bool kvm_memcrypt_enabled(void)
223 {
224     if (kvm_state && kvm_state->memcrypt_handle) {
225         return true;
226     }
227 
228     return false;
229 }
230 
231 int kvm_memcrypt_encrypt_data(uint8_t *ptr, uint64_t len)
232 {
233     if (kvm_state->memcrypt_handle &&
234         kvm_state->memcrypt_encrypt_data) {
235         return kvm_state->memcrypt_encrypt_data(kvm_state->memcrypt_handle,
236                                               ptr, len);
237     }
238 
239     return 1;
240 }
241 
242 /* Called with KVMMemoryListener.slots_lock held */
243 static KVMSlot *kvm_get_free_slot(KVMMemoryListener *kml)
244 {
245     KVMState *s = kvm_state;
246     int i;
247 
248     for (i = 0; i < s->nr_slots; i++) {
249         if (kml->slots[i].memory_size == 0) {
250             return &kml->slots[i];
251         }
252     }
253 
254     return NULL;
255 }
256 
257 bool kvm_has_free_slot(MachineState *ms)
258 {
259     KVMState *s = KVM_STATE(ms->accelerator);
260     bool result;
261     KVMMemoryListener *kml = &s->memory_listener;
262 
263     kvm_slots_lock(kml);
264     result = !!kvm_get_free_slot(kml);
265     kvm_slots_unlock(kml);
266 
267     return result;
268 }
269 
270 /* Called with KVMMemoryListener.slots_lock held */
271 static KVMSlot *kvm_alloc_slot(KVMMemoryListener *kml)
272 {
273     KVMSlot *slot = kvm_get_free_slot(kml);
274 
275     if (slot) {
276         return slot;
277     }
278 
279     fprintf(stderr, "%s: no free slot available\n", __func__);
280     abort();
281 }
282 
283 static KVMSlot *kvm_lookup_matching_slot(KVMMemoryListener *kml,
284                                          hwaddr start_addr,
285                                          hwaddr size)
286 {
287     KVMState *s = kvm_state;
288     int i;
289 
290     for (i = 0; i < s->nr_slots; i++) {
291         KVMSlot *mem = &kml->slots[i];
292 
293         if (start_addr == mem->start_addr && size == mem->memory_size) {
294             return mem;
295         }
296     }
297 
298     return NULL;
299 }
300 
301 /*
302  * Calculate and align the start address and the size of the section.
303  * Return the size. If the size is 0, the aligned section is empty.
304  */
305 static hwaddr kvm_align_section(MemoryRegionSection *section,
306                                 hwaddr *start)
307 {
308     hwaddr size = int128_get64(section->size);
309     hwaddr delta, aligned;
310 
311     /* kvm works in page size chunks, but the function may be called
312        with sub-page size and unaligned start address. Pad the start
313        address to next and truncate size to previous page boundary. */
314     aligned = ROUND_UP(section->offset_within_address_space,
315                        qemu_real_host_page_size);
316     delta = aligned - section->offset_within_address_space;
317     *start = aligned;
318     if (delta > size) {
319         return 0;
320     }
321 
322     return (size - delta) & qemu_real_host_page_mask;
323 }
324 
325 int kvm_physical_memory_addr_from_host(KVMState *s, void *ram,
326                                        hwaddr *phys_addr)
327 {
328     KVMMemoryListener *kml = &s->memory_listener;
329     int i, ret = 0;
330 
331     kvm_slots_lock(kml);
332     for (i = 0; i < s->nr_slots; i++) {
333         KVMSlot *mem = &kml->slots[i];
334 
335         if (ram >= mem->ram && ram < mem->ram + mem->memory_size) {
336             *phys_addr = mem->start_addr + (ram - mem->ram);
337             ret = 1;
338             break;
339         }
340     }
341     kvm_slots_unlock(kml);
342 
343     return ret;
344 }
345 
346 static int kvm_set_user_memory_region(KVMMemoryListener *kml, KVMSlot *slot, bool new)
347 {
348     KVMState *s = kvm_state;
349     struct kvm_userspace_memory_region mem;
350     int ret;
351 
352     mem.slot = slot->slot | (kml->as_id << 16);
353     mem.guest_phys_addr = slot->start_addr;
354     mem.userspace_addr = (unsigned long)slot->ram;
355     mem.flags = slot->flags;
356 
357     if (slot->memory_size && !new && (mem.flags ^ slot->old_flags) & KVM_MEM_READONLY) {
358         /* Set the slot size to 0 before setting the slot to the desired
359          * value. This is needed based on KVM commit 75d61fbc. */
360         mem.memory_size = 0;
361         ret = kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
362         if (ret < 0) {
363             goto err;
364         }
365     }
366     mem.memory_size = slot->memory_size;
367     ret = kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
368     slot->old_flags = mem.flags;
369 err:
370     trace_kvm_set_user_memory(mem.slot, mem.flags, mem.guest_phys_addr,
371                               mem.memory_size, mem.userspace_addr, ret);
372     if (ret < 0) {
373         error_report("%s: KVM_SET_USER_MEMORY_REGION failed, slot=%d,"
374                      " start=0x%" PRIx64 ", size=0x%" PRIx64 ": %s",
375                      __func__, mem.slot, slot->start_addr,
376                      (uint64_t)mem.memory_size, strerror(errno));
377     }
378     return ret;
379 }
380 
381 int kvm_destroy_vcpu(CPUState *cpu)
382 {
383     KVMState *s = kvm_state;
384     long mmap_size;
385     struct KVMParkedVcpu *vcpu = NULL;
386     int ret = 0;
387 
388     DPRINTF("kvm_destroy_vcpu\n");
389 
390     ret = kvm_arch_destroy_vcpu(cpu);
391     if (ret < 0) {
392         goto err;
393     }
394 
395     mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
396     if (mmap_size < 0) {
397         ret = mmap_size;
398         DPRINTF("KVM_GET_VCPU_MMAP_SIZE failed\n");
399         goto err;
400     }
401 
402     ret = munmap(cpu->kvm_run, mmap_size);
403     if (ret < 0) {
404         goto err;
405     }
406 
407     vcpu = g_malloc0(sizeof(*vcpu));
408     vcpu->vcpu_id = kvm_arch_vcpu_id(cpu);
409     vcpu->kvm_fd = cpu->kvm_fd;
410     QLIST_INSERT_HEAD(&kvm_state->kvm_parked_vcpus, vcpu, node);
411 err:
412     return ret;
413 }
414 
415 static int kvm_get_vcpu(KVMState *s, unsigned long vcpu_id)
416 {
417     struct KVMParkedVcpu *cpu;
418 
419     QLIST_FOREACH(cpu, &s->kvm_parked_vcpus, node) {
420         if (cpu->vcpu_id == vcpu_id) {
421             int kvm_fd;
422 
423             QLIST_REMOVE(cpu, node);
424             kvm_fd = cpu->kvm_fd;
425             g_free(cpu);
426             return kvm_fd;
427         }
428     }
429 
430     return kvm_vm_ioctl(s, KVM_CREATE_VCPU, (void *)vcpu_id);
431 }
432 
433 int kvm_init_vcpu(CPUState *cpu)
434 {
435     KVMState *s = kvm_state;
436     long mmap_size;
437     int ret;
438 
439     DPRINTF("kvm_init_vcpu\n");
440 
441     ret = kvm_get_vcpu(s, kvm_arch_vcpu_id(cpu));
442     if (ret < 0) {
443         DPRINTF("kvm_create_vcpu failed\n");
444         goto err;
445     }
446 
447     cpu->kvm_fd = ret;
448     cpu->kvm_state = s;
449     cpu->vcpu_dirty = true;
450 
451     mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
452     if (mmap_size < 0) {
453         ret = mmap_size;
454         DPRINTF("KVM_GET_VCPU_MMAP_SIZE failed\n");
455         goto err;
456     }
457 
458     cpu->kvm_run = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
459                         cpu->kvm_fd, 0);
460     if (cpu->kvm_run == MAP_FAILED) {
461         ret = -errno;
462         DPRINTF("mmap'ing vcpu state failed\n");
463         goto err;
464     }
465 
466     if (s->coalesced_mmio && !s->coalesced_mmio_ring) {
467         s->coalesced_mmio_ring =
468             (void *)cpu->kvm_run + s->coalesced_mmio * PAGE_SIZE;
469     }
470 
471     ret = kvm_arch_init_vcpu(cpu);
472 err:
473     return ret;
474 }
475 
476 /*
477  * dirty pages logging control
478  */
479 
480 static int kvm_mem_flags(MemoryRegion *mr)
481 {
482     bool readonly = mr->readonly || memory_region_is_romd(mr);
483     int flags = 0;
484 
485     if (memory_region_get_dirty_log_mask(mr) != 0) {
486         flags |= KVM_MEM_LOG_DIRTY_PAGES;
487     }
488     if (readonly && kvm_readonly_mem_allowed) {
489         flags |= KVM_MEM_READONLY;
490     }
491     return flags;
492 }
493 
494 /* Called with KVMMemoryListener.slots_lock held */
495 static int kvm_slot_update_flags(KVMMemoryListener *kml, KVMSlot *mem,
496                                  MemoryRegion *mr)
497 {
498     mem->flags = kvm_mem_flags(mr);
499 
500     /* If nothing changed effectively, no need to issue ioctl */
501     if (mem->flags == mem->old_flags) {
502         return 0;
503     }
504 
505     return kvm_set_user_memory_region(kml, mem, false);
506 }
507 
508 static int kvm_section_update_flags(KVMMemoryListener *kml,
509                                     MemoryRegionSection *section)
510 {
511     hwaddr start_addr, size, slot_size;
512     KVMSlot *mem;
513     int ret = 0;
514 
515     size = kvm_align_section(section, &start_addr);
516     if (!size) {
517         return 0;
518     }
519 
520     kvm_slots_lock(kml);
521 
522     while (size && !ret) {
523         slot_size = MIN(kvm_max_slot_size, size);
524         mem = kvm_lookup_matching_slot(kml, start_addr, slot_size);
525         if (!mem) {
526             /* We don't have a slot if we want to trap every access. */
527             goto out;
528         }
529 
530         ret = kvm_slot_update_flags(kml, mem, section->mr);
531         start_addr += slot_size;
532         size -= slot_size;
533     }
534 
535 out:
536     kvm_slots_unlock(kml);
537     return ret;
538 }
539 
540 static void kvm_log_start(MemoryListener *listener,
541                           MemoryRegionSection *section,
542                           int old, int new)
543 {
544     KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
545     int r;
546 
547     if (old != 0) {
548         return;
549     }
550 
551     r = kvm_section_update_flags(kml, section);
552     if (r < 0) {
553         abort();
554     }
555 }
556 
557 static void kvm_log_stop(MemoryListener *listener,
558                           MemoryRegionSection *section,
559                           int old, int new)
560 {
561     KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
562     int r;
563 
564     if (new != 0) {
565         return;
566     }
567 
568     r = kvm_section_update_flags(kml, section);
569     if (r < 0) {
570         abort();
571     }
572 }
573 
574 /* get kvm's dirty pages bitmap and update qemu's */
575 static int kvm_get_dirty_pages_log_range(MemoryRegionSection *section,
576                                          unsigned long *bitmap)
577 {
578     ram_addr_t start = section->offset_within_region +
579                        memory_region_get_ram_addr(section->mr);
580     ram_addr_t pages = int128_get64(section->size) / qemu_real_host_page_size;
581 
582     cpu_physical_memory_set_dirty_lebitmap(bitmap, start, pages);
583     return 0;
584 }
585 
586 #define ALIGN(x, y)  (((x)+(y)-1) & ~((y)-1))
587 
588 /* Allocate the dirty bitmap for a slot  */
589 static void kvm_memslot_init_dirty_bitmap(KVMSlot *mem)
590 {
591     /*
592      * XXX bad kernel interface alert
593      * For dirty bitmap, kernel allocates array of size aligned to
594      * bits-per-long.  But for case when the kernel is 64bits and
595      * the userspace is 32bits, userspace can't align to the same
596      * bits-per-long, since sizeof(long) is different between kernel
597      * and user space.  This way, userspace will provide buffer which
598      * may be 4 bytes less than the kernel will use, resulting in
599      * userspace memory corruption (which is not detectable by valgrind
600      * too, in most cases).
601      * So for now, let's align to 64 instead of HOST_LONG_BITS here, in
602      * a hope that sizeof(long) won't become >8 any time soon.
603      */
604     hwaddr bitmap_size = ALIGN(((mem->memory_size) >> TARGET_PAGE_BITS),
605                                         /*HOST_LONG_BITS*/ 64) / 8;
606     mem->dirty_bmap = g_malloc0(bitmap_size);
607 }
608 
609 /**
610  * kvm_physical_sync_dirty_bitmap - Sync dirty bitmap from kernel space
611  *
612  * This function will first try to fetch dirty bitmap from the kernel,
613  * and then updates qemu's dirty bitmap.
614  *
615  * NOTE: caller must be with kml->slots_lock held.
616  *
617  * @kml: the KVM memory listener object
618  * @section: the memory section to sync the dirty bitmap with
619  */
620 static int kvm_physical_sync_dirty_bitmap(KVMMemoryListener *kml,
621                                           MemoryRegionSection *section)
622 {
623     KVMState *s = kvm_state;
624     struct kvm_dirty_log d = {};
625     KVMSlot *mem;
626     hwaddr start_addr, size;
627     hwaddr slot_size, slot_offset = 0;
628     int ret = 0;
629 
630     size = kvm_align_section(section, &start_addr);
631     while (size) {
632         MemoryRegionSection subsection = *section;
633 
634         slot_size = MIN(kvm_max_slot_size, size);
635         mem = kvm_lookup_matching_slot(kml, start_addr, slot_size);
636         if (!mem) {
637             /* We don't have a slot if we want to trap every access. */
638             goto out;
639         }
640 
641         if (!mem->dirty_bmap) {
642             /* Allocate on the first log_sync, once and for all */
643             kvm_memslot_init_dirty_bitmap(mem);
644         }
645 
646         d.dirty_bitmap = mem->dirty_bmap;
647         d.slot = mem->slot | (kml->as_id << 16);
648         if (kvm_vm_ioctl(s, KVM_GET_DIRTY_LOG, &d) == -1) {
649             DPRINTF("ioctl failed %d\n", errno);
650             ret = -1;
651             goto out;
652         }
653 
654         subsection.offset_within_region += slot_offset;
655         subsection.size = int128_make64(slot_size);
656         kvm_get_dirty_pages_log_range(&subsection, d.dirty_bitmap);
657 
658         slot_offset += slot_size;
659         start_addr += slot_size;
660         size -= slot_size;
661     }
662 out:
663     return ret;
664 }
665 
666 /* Alignment requirement for KVM_CLEAR_DIRTY_LOG - 64 pages */
667 #define KVM_CLEAR_LOG_SHIFT  6
668 #define KVM_CLEAR_LOG_ALIGN  (qemu_real_host_page_size << KVM_CLEAR_LOG_SHIFT)
669 #define KVM_CLEAR_LOG_MASK   (-KVM_CLEAR_LOG_ALIGN)
670 
671 static int kvm_log_clear_one_slot(KVMSlot *mem, int as_id, uint64_t start,
672                                   uint64_t size)
673 {
674     KVMState *s = kvm_state;
675     uint64_t end, bmap_start, start_delta, bmap_npages;
676     struct kvm_clear_dirty_log d;
677     unsigned long *bmap_clear = NULL, psize = qemu_real_host_page_size;
678     int ret;
679 
680     /*
681      * We need to extend either the start or the size or both to
682      * satisfy the KVM interface requirement.  Firstly, do the start
683      * page alignment on 64 host pages
684      */
685     bmap_start = start & KVM_CLEAR_LOG_MASK;
686     start_delta = start - bmap_start;
687     bmap_start /= psize;
688 
689     /*
690      * The kernel interface has restriction on the size too, that either:
691      *
692      * (1) the size is 64 host pages aligned (just like the start), or
693      * (2) the size fills up until the end of the KVM memslot.
694      */
695     bmap_npages = DIV_ROUND_UP(size + start_delta, KVM_CLEAR_LOG_ALIGN)
696         << KVM_CLEAR_LOG_SHIFT;
697     end = mem->memory_size / psize;
698     if (bmap_npages > end - bmap_start) {
699         bmap_npages = end - bmap_start;
700     }
701     start_delta /= psize;
702 
703     /*
704      * Prepare the bitmap to clear dirty bits.  Here we must guarantee
705      * that we won't clear any unknown dirty bits otherwise we might
706      * accidentally clear some set bits which are not yet synced from
707      * the kernel into QEMU's bitmap, then we'll lose track of the
708      * guest modifications upon those pages (which can directly lead
709      * to guest data loss or panic after migration).
710      *
711      * Layout of the KVMSlot.dirty_bmap:
712      *
713      *                   |<-------- bmap_npages -----------..>|
714      *                                                     [1]
715      *                     start_delta         size
716      *  |----------------|-------------|------------------|------------|
717      *  ^                ^             ^                               ^
718      *  |                |             |                               |
719      * start          bmap_start     (start)                         end
720      * of memslot                                             of memslot
721      *
722      * [1] bmap_npages can be aligned to either 64 pages or the end of slot
723      */
724 
725     assert(bmap_start % BITS_PER_LONG == 0);
726     /* We should never do log_clear before log_sync */
727     assert(mem->dirty_bmap);
728     if (start_delta) {
729         /* Slow path - we need to manipulate a temp bitmap */
730         bmap_clear = bitmap_new(bmap_npages);
731         bitmap_copy_with_src_offset(bmap_clear, mem->dirty_bmap,
732                                     bmap_start, start_delta + size / psize);
733         /*
734          * We need to fill the holes at start because that was not
735          * specified by the caller and we extended the bitmap only for
736          * 64 pages alignment
737          */
738         bitmap_clear(bmap_clear, 0, start_delta);
739         d.dirty_bitmap = bmap_clear;
740     } else {
741         /* Fast path - start address aligns well with BITS_PER_LONG */
742         d.dirty_bitmap = mem->dirty_bmap + BIT_WORD(bmap_start);
743     }
744 
745     d.first_page = bmap_start;
746     /* It should never overflow.  If it happens, say something */
747     assert(bmap_npages <= UINT32_MAX);
748     d.num_pages = bmap_npages;
749     d.slot = mem->slot | (as_id << 16);
750 
751     if (kvm_vm_ioctl(s, KVM_CLEAR_DIRTY_LOG, &d) == -1) {
752         ret = -errno;
753         error_report("%s: KVM_CLEAR_DIRTY_LOG failed, slot=%d, "
754                      "start=0x%"PRIx64", size=0x%"PRIx32", errno=%d",
755                      __func__, d.slot, (uint64_t)d.first_page,
756                      (uint32_t)d.num_pages, ret);
757     } else {
758         ret = 0;
759         trace_kvm_clear_dirty_log(d.slot, d.first_page, d.num_pages);
760     }
761 
762     /*
763      * After we have updated the remote dirty bitmap, we update the
764      * cached bitmap as well for the memslot, then if another user
765      * clears the same region we know we shouldn't clear it again on
766      * the remote otherwise it's data loss as well.
767      */
768     bitmap_clear(mem->dirty_bmap, bmap_start + start_delta,
769                  size / psize);
770     /* This handles the NULL case well */
771     g_free(bmap_clear);
772     return ret;
773 }
774 
775 
776 /**
777  * kvm_physical_log_clear - Clear the kernel's dirty bitmap for range
778  *
779  * NOTE: this will be a no-op if we haven't enabled manual dirty log
780  * protection in the host kernel because in that case this operation
781  * will be done within log_sync().
782  *
783  * @kml:     the kvm memory listener
784  * @section: the memory range to clear dirty bitmap
785  */
786 static int kvm_physical_log_clear(KVMMemoryListener *kml,
787                                   MemoryRegionSection *section)
788 {
789     KVMState *s = kvm_state;
790     uint64_t start, size, offset, count;
791     KVMSlot *mem;
792     int ret = 0, i;
793 
794     if (!s->manual_dirty_log_protect) {
795         /* No need to do explicit clear */
796         return ret;
797     }
798 
799     start = section->offset_within_address_space;
800     size = int128_get64(section->size);
801 
802     if (!size) {
803         /* Nothing more we can do... */
804         return ret;
805     }
806 
807     kvm_slots_lock(kml);
808 
809     for (i = 0; i < s->nr_slots; i++) {
810         mem = &kml->slots[i];
811         /* Discard slots that are empty or do not overlap the section */
812         if (!mem->memory_size ||
813             mem->start_addr > start + size - 1 ||
814             start > mem->start_addr + mem->memory_size - 1) {
815             continue;
816         }
817 
818         if (start >= mem->start_addr) {
819             /* The slot starts before section or is aligned to it.  */
820             offset = start - mem->start_addr;
821             count = MIN(mem->memory_size - offset, size);
822         } else {
823             /* The slot starts after section.  */
824             offset = 0;
825             count = MIN(mem->memory_size, size - (mem->start_addr - start));
826         }
827         ret = kvm_log_clear_one_slot(mem, kml->as_id, offset, count);
828         if (ret < 0) {
829             break;
830         }
831     }
832 
833     kvm_slots_unlock(kml);
834 
835     return ret;
836 }
837 
838 static void kvm_coalesce_mmio_region(MemoryListener *listener,
839                                      MemoryRegionSection *secion,
840                                      hwaddr start, hwaddr size)
841 {
842     KVMState *s = kvm_state;
843 
844     if (s->coalesced_mmio) {
845         struct kvm_coalesced_mmio_zone zone;
846 
847         zone.addr = start;
848         zone.size = size;
849         zone.pad = 0;
850 
851         (void)kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone);
852     }
853 }
854 
855 static void kvm_uncoalesce_mmio_region(MemoryListener *listener,
856                                        MemoryRegionSection *secion,
857                                        hwaddr start, hwaddr size)
858 {
859     KVMState *s = kvm_state;
860 
861     if (s->coalesced_mmio) {
862         struct kvm_coalesced_mmio_zone zone;
863 
864         zone.addr = start;
865         zone.size = size;
866         zone.pad = 0;
867 
868         (void)kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone);
869     }
870 }
871 
872 static void kvm_coalesce_pio_add(MemoryListener *listener,
873                                 MemoryRegionSection *section,
874                                 hwaddr start, hwaddr size)
875 {
876     KVMState *s = kvm_state;
877 
878     if (s->coalesced_pio) {
879         struct kvm_coalesced_mmio_zone zone;
880 
881         zone.addr = start;
882         zone.size = size;
883         zone.pio = 1;
884 
885         (void)kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone);
886     }
887 }
888 
889 static void kvm_coalesce_pio_del(MemoryListener *listener,
890                                 MemoryRegionSection *section,
891                                 hwaddr start, hwaddr size)
892 {
893     KVMState *s = kvm_state;
894 
895     if (s->coalesced_pio) {
896         struct kvm_coalesced_mmio_zone zone;
897 
898         zone.addr = start;
899         zone.size = size;
900         zone.pio = 1;
901 
902         (void)kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone);
903      }
904 }
905 
906 static MemoryListener kvm_coalesced_pio_listener = {
907     .coalesced_io_add = kvm_coalesce_pio_add,
908     .coalesced_io_del = kvm_coalesce_pio_del,
909 };
910 
911 int kvm_check_extension(KVMState *s, unsigned int extension)
912 {
913     int ret;
914 
915     ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, extension);
916     if (ret < 0) {
917         ret = 0;
918     }
919 
920     return ret;
921 }
922 
923 int kvm_vm_check_extension(KVMState *s, unsigned int extension)
924 {
925     int ret;
926 
927     ret = kvm_vm_ioctl(s, KVM_CHECK_EXTENSION, extension);
928     if (ret < 0) {
929         /* VM wide version not implemented, use global one instead */
930         ret = kvm_check_extension(s, extension);
931     }
932 
933     return ret;
934 }
935 
936 typedef struct HWPoisonPage {
937     ram_addr_t ram_addr;
938     QLIST_ENTRY(HWPoisonPage) list;
939 } HWPoisonPage;
940 
941 static QLIST_HEAD(, HWPoisonPage) hwpoison_page_list =
942     QLIST_HEAD_INITIALIZER(hwpoison_page_list);
943 
944 static void kvm_unpoison_all(void *param)
945 {
946     HWPoisonPage *page, *next_page;
947 
948     QLIST_FOREACH_SAFE(page, &hwpoison_page_list, list, next_page) {
949         QLIST_REMOVE(page, list);
950         qemu_ram_remap(page->ram_addr, TARGET_PAGE_SIZE);
951         g_free(page);
952     }
953 }
954 
955 void kvm_hwpoison_page_add(ram_addr_t ram_addr)
956 {
957     HWPoisonPage *page;
958 
959     QLIST_FOREACH(page, &hwpoison_page_list, list) {
960         if (page->ram_addr == ram_addr) {
961             return;
962         }
963     }
964     page = g_new(HWPoisonPage, 1);
965     page->ram_addr = ram_addr;
966     QLIST_INSERT_HEAD(&hwpoison_page_list, page, list);
967 }
968 
969 static uint32_t adjust_ioeventfd_endianness(uint32_t val, uint32_t size)
970 {
971 #if defined(HOST_WORDS_BIGENDIAN) != defined(TARGET_WORDS_BIGENDIAN)
972     /* The kernel expects ioeventfd values in HOST_WORDS_BIGENDIAN
973      * endianness, but the memory core hands them in target endianness.
974      * For example, PPC is always treated as big-endian even if running
975      * on KVM and on PPC64LE.  Correct here.
976      */
977     switch (size) {
978     case 2:
979         val = bswap16(val);
980         break;
981     case 4:
982         val = bswap32(val);
983         break;
984     }
985 #endif
986     return val;
987 }
988 
989 static int kvm_set_ioeventfd_mmio(int fd, hwaddr addr, uint32_t val,
990                                   bool assign, uint32_t size, bool datamatch)
991 {
992     int ret;
993     struct kvm_ioeventfd iofd = {
994         .datamatch = datamatch ? adjust_ioeventfd_endianness(val, size) : 0,
995         .addr = addr,
996         .len = size,
997         .flags = 0,
998         .fd = fd,
999     };
1000 
1001     trace_kvm_set_ioeventfd_mmio(fd, (uint64_t)addr, val, assign, size,
1002                                  datamatch);
1003     if (!kvm_enabled()) {
1004         return -ENOSYS;
1005     }
1006 
1007     if (datamatch) {
1008         iofd.flags |= KVM_IOEVENTFD_FLAG_DATAMATCH;
1009     }
1010     if (!assign) {
1011         iofd.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
1012     }
1013 
1014     ret = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &iofd);
1015 
1016     if (ret < 0) {
1017         return -errno;
1018     }
1019 
1020     return 0;
1021 }
1022 
1023 static int kvm_set_ioeventfd_pio(int fd, uint16_t addr, uint16_t val,
1024                                  bool assign, uint32_t size, bool datamatch)
1025 {
1026     struct kvm_ioeventfd kick = {
1027         .datamatch = datamatch ? adjust_ioeventfd_endianness(val, size) : 0,
1028         .addr = addr,
1029         .flags = KVM_IOEVENTFD_FLAG_PIO,
1030         .len = size,
1031         .fd = fd,
1032     };
1033     int r;
1034     trace_kvm_set_ioeventfd_pio(fd, addr, val, assign, size, datamatch);
1035     if (!kvm_enabled()) {
1036         return -ENOSYS;
1037     }
1038     if (datamatch) {
1039         kick.flags |= KVM_IOEVENTFD_FLAG_DATAMATCH;
1040     }
1041     if (!assign) {
1042         kick.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
1043     }
1044     r = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &kick);
1045     if (r < 0) {
1046         return r;
1047     }
1048     return 0;
1049 }
1050 
1051 
1052 static int kvm_check_many_ioeventfds(void)
1053 {
1054     /* Userspace can use ioeventfd for io notification.  This requires a host
1055      * that supports eventfd(2) and an I/O thread; since eventfd does not
1056      * support SIGIO it cannot interrupt the vcpu.
1057      *
1058      * Older kernels have a 6 device limit on the KVM io bus.  Find out so we
1059      * can avoid creating too many ioeventfds.
1060      */
1061 #if defined(CONFIG_EVENTFD)
1062     int ioeventfds[7];
1063     int i, ret = 0;
1064     for (i = 0; i < ARRAY_SIZE(ioeventfds); i++) {
1065         ioeventfds[i] = eventfd(0, EFD_CLOEXEC);
1066         if (ioeventfds[i] < 0) {
1067             break;
1068         }
1069         ret = kvm_set_ioeventfd_pio(ioeventfds[i], 0, i, true, 2, true);
1070         if (ret < 0) {
1071             close(ioeventfds[i]);
1072             break;
1073         }
1074     }
1075 
1076     /* Decide whether many devices are supported or not */
1077     ret = i == ARRAY_SIZE(ioeventfds);
1078 
1079     while (i-- > 0) {
1080         kvm_set_ioeventfd_pio(ioeventfds[i], 0, i, false, 2, true);
1081         close(ioeventfds[i]);
1082     }
1083     return ret;
1084 #else
1085     return 0;
1086 #endif
1087 }
1088 
1089 static const KVMCapabilityInfo *
1090 kvm_check_extension_list(KVMState *s, const KVMCapabilityInfo *list)
1091 {
1092     while (list->name) {
1093         if (!kvm_check_extension(s, list->value)) {
1094             return list;
1095         }
1096         list++;
1097     }
1098     return NULL;
1099 }
1100 
1101 void kvm_set_max_memslot_size(hwaddr max_slot_size)
1102 {
1103     g_assert(
1104         ROUND_UP(max_slot_size, qemu_real_host_page_size) == max_slot_size
1105     );
1106     kvm_max_slot_size = max_slot_size;
1107 }
1108 
1109 static void kvm_set_phys_mem(KVMMemoryListener *kml,
1110                              MemoryRegionSection *section, bool add)
1111 {
1112     KVMSlot *mem;
1113     int err;
1114     MemoryRegion *mr = section->mr;
1115     bool writeable = !mr->readonly && !mr->rom_device;
1116     hwaddr start_addr, size, slot_size;
1117     void *ram;
1118 
1119     if (!memory_region_is_ram(mr)) {
1120         if (writeable || !kvm_readonly_mem_allowed) {
1121             return;
1122         } else if (!mr->romd_mode) {
1123             /* If the memory device is not in romd_mode, then we actually want
1124              * to remove the kvm memory slot so all accesses will trap. */
1125             add = false;
1126         }
1127     }
1128 
1129     size = kvm_align_section(section, &start_addr);
1130     if (!size) {
1131         return;
1132     }
1133 
1134     /* use aligned delta to align the ram address */
1135     ram = memory_region_get_ram_ptr(mr) + section->offset_within_region +
1136           (start_addr - section->offset_within_address_space);
1137 
1138     kvm_slots_lock(kml);
1139 
1140     if (!add) {
1141         do {
1142             slot_size = MIN(kvm_max_slot_size, size);
1143             mem = kvm_lookup_matching_slot(kml, start_addr, slot_size);
1144             if (!mem) {
1145                 goto out;
1146             }
1147             if (mem->flags & KVM_MEM_LOG_DIRTY_PAGES) {
1148                 kvm_physical_sync_dirty_bitmap(kml, section);
1149             }
1150 
1151             /* unregister the slot */
1152             g_free(mem->dirty_bmap);
1153             mem->dirty_bmap = NULL;
1154             mem->memory_size = 0;
1155             mem->flags = 0;
1156             err = kvm_set_user_memory_region(kml, mem, false);
1157             if (err) {
1158                 fprintf(stderr, "%s: error unregistering slot: %s\n",
1159                         __func__, strerror(-err));
1160                 abort();
1161             }
1162             start_addr += slot_size;
1163             size -= slot_size;
1164         } while (size);
1165         goto out;
1166     }
1167 
1168     /* register the new slot */
1169     do {
1170         slot_size = MIN(kvm_max_slot_size, size);
1171         mem = kvm_alloc_slot(kml);
1172         mem->memory_size = slot_size;
1173         mem->start_addr = start_addr;
1174         mem->ram = ram;
1175         mem->flags = kvm_mem_flags(mr);
1176 
1177         if (mem->flags & KVM_MEM_LOG_DIRTY_PAGES) {
1178             /*
1179              * Reallocate the bmap; it means it doesn't disappear in
1180              * middle of a migrate.
1181              */
1182             kvm_memslot_init_dirty_bitmap(mem);
1183         }
1184         err = kvm_set_user_memory_region(kml, mem, true);
1185         if (err) {
1186             fprintf(stderr, "%s: error registering slot: %s\n", __func__,
1187                     strerror(-err));
1188             abort();
1189         }
1190         start_addr += slot_size;
1191         ram += slot_size;
1192         size -= slot_size;
1193     } while (size);
1194 
1195 out:
1196     kvm_slots_unlock(kml);
1197 }
1198 
1199 static void kvm_region_add(MemoryListener *listener,
1200                            MemoryRegionSection *section)
1201 {
1202     KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
1203 
1204     memory_region_ref(section->mr);
1205     kvm_set_phys_mem(kml, section, true);
1206 }
1207 
1208 static void kvm_region_del(MemoryListener *listener,
1209                            MemoryRegionSection *section)
1210 {
1211     KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
1212 
1213     kvm_set_phys_mem(kml, section, false);
1214     memory_region_unref(section->mr);
1215 }
1216 
1217 static void kvm_log_sync(MemoryListener *listener,
1218                          MemoryRegionSection *section)
1219 {
1220     KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
1221     int r;
1222 
1223     kvm_slots_lock(kml);
1224     r = kvm_physical_sync_dirty_bitmap(kml, section);
1225     kvm_slots_unlock(kml);
1226     if (r < 0) {
1227         abort();
1228     }
1229 }
1230 
1231 static void kvm_log_clear(MemoryListener *listener,
1232                           MemoryRegionSection *section)
1233 {
1234     KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
1235     int r;
1236 
1237     r = kvm_physical_log_clear(kml, section);
1238     if (r < 0) {
1239         error_report_once("%s: kvm log clear failed: mr=%s "
1240                           "offset=%"HWADDR_PRIx" size=%"PRIx64, __func__,
1241                           section->mr->name, section->offset_within_region,
1242                           int128_get64(section->size));
1243         abort();
1244     }
1245 }
1246 
1247 static void kvm_mem_ioeventfd_add(MemoryListener *listener,
1248                                   MemoryRegionSection *section,
1249                                   bool match_data, uint64_t data,
1250                                   EventNotifier *e)
1251 {
1252     int fd = event_notifier_get_fd(e);
1253     int r;
1254 
1255     r = kvm_set_ioeventfd_mmio(fd, section->offset_within_address_space,
1256                                data, true, int128_get64(section->size),
1257                                match_data);
1258     if (r < 0) {
1259         fprintf(stderr, "%s: error adding ioeventfd: %s (%d)\n",
1260                 __func__, strerror(-r), -r);
1261         abort();
1262     }
1263 }
1264 
1265 static void kvm_mem_ioeventfd_del(MemoryListener *listener,
1266                                   MemoryRegionSection *section,
1267                                   bool match_data, uint64_t data,
1268                                   EventNotifier *e)
1269 {
1270     int fd = event_notifier_get_fd(e);
1271     int r;
1272 
1273     r = kvm_set_ioeventfd_mmio(fd, section->offset_within_address_space,
1274                                data, false, int128_get64(section->size),
1275                                match_data);
1276     if (r < 0) {
1277         fprintf(stderr, "%s: error deleting ioeventfd: %s (%d)\n",
1278                 __func__, strerror(-r), -r);
1279         abort();
1280     }
1281 }
1282 
1283 static void kvm_io_ioeventfd_add(MemoryListener *listener,
1284                                  MemoryRegionSection *section,
1285                                  bool match_data, uint64_t data,
1286                                  EventNotifier *e)
1287 {
1288     int fd = event_notifier_get_fd(e);
1289     int r;
1290 
1291     r = kvm_set_ioeventfd_pio(fd, section->offset_within_address_space,
1292                               data, true, int128_get64(section->size),
1293                               match_data);
1294     if (r < 0) {
1295         fprintf(stderr, "%s: error adding ioeventfd: %s (%d)\n",
1296                 __func__, strerror(-r), -r);
1297         abort();
1298     }
1299 }
1300 
1301 static void kvm_io_ioeventfd_del(MemoryListener *listener,
1302                                  MemoryRegionSection *section,
1303                                  bool match_data, uint64_t data,
1304                                  EventNotifier *e)
1305 
1306 {
1307     int fd = event_notifier_get_fd(e);
1308     int r;
1309 
1310     r = kvm_set_ioeventfd_pio(fd, section->offset_within_address_space,
1311                               data, false, int128_get64(section->size),
1312                               match_data);
1313     if (r < 0) {
1314         fprintf(stderr, "%s: error deleting ioeventfd: %s (%d)\n",
1315                 __func__, strerror(-r), -r);
1316         abort();
1317     }
1318 }
1319 
1320 void kvm_memory_listener_register(KVMState *s, KVMMemoryListener *kml,
1321                                   AddressSpace *as, int as_id)
1322 {
1323     int i;
1324 
1325     qemu_mutex_init(&kml->slots_lock);
1326     kml->slots = g_malloc0(s->nr_slots * sizeof(KVMSlot));
1327     kml->as_id = as_id;
1328 
1329     for (i = 0; i < s->nr_slots; i++) {
1330         kml->slots[i].slot = i;
1331     }
1332 
1333     kml->listener.region_add = kvm_region_add;
1334     kml->listener.region_del = kvm_region_del;
1335     kml->listener.log_start = kvm_log_start;
1336     kml->listener.log_stop = kvm_log_stop;
1337     kml->listener.log_sync = kvm_log_sync;
1338     kml->listener.log_clear = kvm_log_clear;
1339     kml->listener.priority = 10;
1340 
1341     memory_listener_register(&kml->listener, as);
1342 
1343     for (i = 0; i < s->nr_as; ++i) {
1344         if (!s->as[i].as) {
1345             s->as[i].as = as;
1346             s->as[i].ml = kml;
1347             break;
1348         }
1349     }
1350 }
1351 
1352 static MemoryListener kvm_io_listener = {
1353     .eventfd_add = kvm_io_ioeventfd_add,
1354     .eventfd_del = kvm_io_ioeventfd_del,
1355     .priority = 10,
1356 };
1357 
1358 int kvm_set_irq(KVMState *s, int irq, int level)
1359 {
1360     struct kvm_irq_level event;
1361     int ret;
1362 
1363     assert(kvm_async_interrupts_enabled());
1364 
1365     event.level = level;
1366     event.irq = irq;
1367     ret = kvm_vm_ioctl(s, s->irq_set_ioctl, &event);
1368     if (ret < 0) {
1369         perror("kvm_set_irq");
1370         abort();
1371     }
1372 
1373     return (s->irq_set_ioctl == KVM_IRQ_LINE) ? 1 : event.status;
1374 }
1375 
1376 #ifdef KVM_CAP_IRQ_ROUTING
1377 typedef struct KVMMSIRoute {
1378     struct kvm_irq_routing_entry kroute;
1379     QTAILQ_ENTRY(KVMMSIRoute) entry;
1380 } KVMMSIRoute;
1381 
1382 static void set_gsi(KVMState *s, unsigned int gsi)
1383 {
1384     set_bit(gsi, s->used_gsi_bitmap);
1385 }
1386 
1387 static void clear_gsi(KVMState *s, unsigned int gsi)
1388 {
1389     clear_bit(gsi, s->used_gsi_bitmap);
1390 }
1391 
1392 void kvm_init_irq_routing(KVMState *s)
1393 {
1394     int gsi_count, i;
1395 
1396     gsi_count = kvm_check_extension(s, KVM_CAP_IRQ_ROUTING) - 1;
1397     if (gsi_count > 0) {
1398         /* Round up so we can search ints using ffs */
1399         s->used_gsi_bitmap = bitmap_new(gsi_count);
1400         s->gsi_count = gsi_count;
1401     }
1402 
1403     s->irq_routes = g_malloc0(sizeof(*s->irq_routes));
1404     s->nr_allocated_irq_routes = 0;
1405 
1406     if (!kvm_direct_msi_allowed) {
1407         for (i = 0; i < KVM_MSI_HASHTAB_SIZE; i++) {
1408             QTAILQ_INIT(&s->msi_hashtab[i]);
1409         }
1410     }
1411 
1412     kvm_arch_init_irq_routing(s);
1413 }
1414 
1415 void kvm_irqchip_commit_routes(KVMState *s)
1416 {
1417     int ret;
1418 
1419     if (kvm_gsi_direct_mapping()) {
1420         return;
1421     }
1422 
1423     if (!kvm_gsi_routing_enabled()) {
1424         return;
1425     }
1426 
1427     s->irq_routes->flags = 0;
1428     trace_kvm_irqchip_commit_routes();
1429     ret = kvm_vm_ioctl(s, KVM_SET_GSI_ROUTING, s->irq_routes);
1430     assert(ret == 0);
1431 }
1432 
1433 static void kvm_add_routing_entry(KVMState *s,
1434                                   struct kvm_irq_routing_entry *entry)
1435 {
1436     struct kvm_irq_routing_entry *new;
1437     int n, size;
1438 
1439     if (s->irq_routes->nr == s->nr_allocated_irq_routes) {
1440         n = s->nr_allocated_irq_routes * 2;
1441         if (n < 64) {
1442             n = 64;
1443         }
1444         size = sizeof(struct kvm_irq_routing);
1445         size += n * sizeof(*new);
1446         s->irq_routes = g_realloc(s->irq_routes, size);
1447         s->nr_allocated_irq_routes = n;
1448     }
1449     n = s->irq_routes->nr++;
1450     new = &s->irq_routes->entries[n];
1451 
1452     *new = *entry;
1453 
1454     set_gsi(s, entry->gsi);
1455 }
1456 
1457 static int kvm_update_routing_entry(KVMState *s,
1458                                     struct kvm_irq_routing_entry *new_entry)
1459 {
1460     struct kvm_irq_routing_entry *entry;
1461     int n;
1462 
1463     for (n = 0; n < s->irq_routes->nr; n++) {
1464         entry = &s->irq_routes->entries[n];
1465         if (entry->gsi != new_entry->gsi) {
1466             continue;
1467         }
1468 
1469         if(!memcmp(entry, new_entry, sizeof *entry)) {
1470             return 0;
1471         }
1472 
1473         *entry = *new_entry;
1474 
1475         return 0;
1476     }
1477 
1478     return -ESRCH;
1479 }
1480 
1481 void kvm_irqchip_add_irq_route(KVMState *s, int irq, int irqchip, int pin)
1482 {
1483     struct kvm_irq_routing_entry e = {};
1484 
1485     assert(pin < s->gsi_count);
1486 
1487     e.gsi = irq;
1488     e.type = KVM_IRQ_ROUTING_IRQCHIP;
1489     e.flags = 0;
1490     e.u.irqchip.irqchip = irqchip;
1491     e.u.irqchip.pin = pin;
1492     kvm_add_routing_entry(s, &e);
1493 }
1494 
1495 void kvm_irqchip_release_virq(KVMState *s, int virq)
1496 {
1497     struct kvm_irq_routing_entry *e;
1498     int i;
1499 
1500     if (kvm_gsi_direct_mapping()) {
1501         return;
1502     }
1503 
1504     for (i = 0; i < s->irq_routes->nr; i++) {
1505         e = &s->irq_routes->entries[i];
1506         if (e->gsi == virq) {
1507             s->irq_routes->nr--;
1508             *e = s->irq_routes->entries[s->irq_routes->nr];
1509         }
1510     }
1511     clear_gsi(s, virq);
1512     kvm_arch_release_virq_post(virq);
1513     trace_kvm_irqchip_release_virq(virq);
1514 }
1515 
1516 void kvm_irqchip_add_change_notifier(Notifier *n)
1517 {
1518     notifier_list_add(&kvm_irqchip_change_notifiers, n);
1519 }
1520 
1521 void kvm_irqchip_remove_change_notifier(Notifier *n)
1522 {
1523     notifier_remove(n);
1524 }
1525 
1526 void kvm_irqchip_change_notify(void)
1527 {
1528     notifier_list_notify(&kvm_irqchip_change_notifiers, NULL);
1529 }
1530 
1531 static unsigned int kvm_hash_msi(uint32_t data)
1532 {
1533     /* This is optimized for IA32 MSI layout. However, no other arch shall
1534      * repeat the mistake of not providing a direct MSI injection API. */
1535     return data & 0xff;
1536 }
1537 
1538 static void kvm_flush_dynamic_msi_routes(KVMState *s)
1539 {
1540     KVMMSIRoute *route, *next;
1541     unsigned int hash;
1542 
1543     for (hash = 0; hash < KVM_MSI_HASHTAB_SIZE; hash++) {
1544         QTAILQ_FOREACH_SAFE(route, &s->msi_hashtab[hash], entry, next) {
1545             kvm_irqchip_release_virq(s, route->kroute.gsi);
1546             QTAILQ_REMOVE(&s->msi_hashtab[hash], route, entry);
1547             g_free(route);
1548         }
1549     }
1550 }
1551 
1552 static int kvm_irqchip_get_virq(KVMState *s)
1553 {
1554     int next_virq;
1555 
1556     /*
1557      * PIC and IOAPIC share the first 16 GSI numbers, thus the available
1558      * GSI numbers are more than the number of IRQ route. Allocating a GSI
1559      * number can succeed even though a new route entry cannot be added.
1560      * When this happens, flush dynamic MSI entries to free IRQ route entries.
1561      */
1562     if (!kvm_direct_msi_allowed && s->irq_routes->nr == s->gsi_count) {
1563         kvm_flush_dynamic_msi_routes(s);
1564     }
1565 
1566     /* Return the lowest unused GSI in the bitmap */
1567     next_virq = find_first_zero_bit(s->used_gsi_bitmap, s->gsi_count);
1568     if (next_virq >= s->gsi_count) {
1569         return -ENOSPC;
1570     } else {
1571         return next_virq;
1572     }
1573 }
1574 
1575 static KVMMSIRoute *kvm_lookup_msi_route(KVMState *s, MSIMessage msg)
1576 {
1577     unsigned int hash = kvm_hash_msi(msg.data);
1578     KVMMSIRoute *route;
1579 
1580     QTAILQ_FOREACH(route, &s->msi_hashtab[hash], entry) {
1581         if (route->kroute.u.msi.address_lo == (uint32_t)msg.address &&
1582             route->kroute.u.msi.address_hi == (msg.address >> 32) &&
1583             route->kroute.u.msi.data == le32_to_cpu(msg.data)) {
1584             return route;
1585         }
1586     }
1587     return NULL;
1588 }
1589 
1590 int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg)
1591 {
1592     struct kvm_msi msi;
1593     KVMMSIRoute *route;
1594 
1595     if (kvm_direct_msi_allowed) {
1596         msi.address_lo = (uint32_t)msg.address;
1597         msi.address_hi = msg.address >> 32;
1598         msi.data = le32_to_cpu(msg.data);
1599         msi.flags = 0;
1600         memset(msi.pad, 0, sizeof(msi.pad));
1601 
1602         return kvm_vm_ioctl(s, KVM_SIGNAL_MSI, &msi);
1603     }
1604 
1605     route = kvm_lookup_msi_route(s, msg);
1606     if (!route) {
1607         int virq;
1608 
1609         virq = kvm_irqchip_get_virq(s);
1610         if (virq < 0) {
1611             return virq;
1612         }
1613 
1614         route = g_malloc0(sizeof(KVMMSIRoute));
1615         route->kroute.gsi = virq;
1616         route->kroute.type = KVM_IRQ_ROUTING_MSI;
1617         route->kroute.flags = 0;
1618         route->kroute.u.msi.address_lo = (uint32_t)msg.address;
1619         route->kroute.u.msi.address_hi = msg.address >> 32;
1620         route->kroute.u.msi.data = le32_to_cpu(msg.data);
1621 
1622         kvm_add_routing_entry(s, &route->kroute);
1623         kvm_irqchip_commit_routes(s);
1624 
1625         QTAILQ_INSERT_TAIL(&s->msi_hashtab[kvm_hash_msi(msg.data)], route,
1626                            entry);
1627     }
1628 
1629     assert(route->kroute.type == KVM_IRQ_ROUTING_MSI);
1630 
1631     return kvm_set_irq(s, route->kroute.gsi, 1);
1632 }
1633 
1634 int kvm_irqchip_add_msi_route(KVMState *s, int vector, PCIDevice *dev)
1635 {
1636     struct kvm_irq_routing_entry kroute = {};
1637     int virq;
1638     MSIMessage msg = {0, 0};
1639 
1640     if (pci_available && dev) {
1641         msg = pci_get_msi_message(dev, vector);
1642     }
1643 
1644     if (kvm_gsi_direct_mapping()) {
1645         return kvm_arch_msi_data_to_gsi(msg.data);
1646     }
1647 
1648     if (!kvm_gsi_routing_enabled()) {
1649         return -ENOSYS;
1650     }
1651 
1652     virq = kvm_irqchip_get_virq(s);
1653     if (virq < 0) {
1654         return virq;
1655     }
1656 
1657     kroute.gsi = virq;
1658     kroute.type = KVM_IRQ_ROUTING_MSI;
1659     kroute.flags = 0;
1660     kroute.u.msi.address_lo = (uint32_t)msg.address;
1661     kroute.u.msi.address_hi = msg.address >> 32;
1662     kroute.u.msi.data = le32_to_cpu(msg.data);
1663     if (pci_available && kvm_msi_devid_required()) {
1664         kroute.flags = KVM_MSI_VALID_DEVID;
1665         kroute.u.msi.devid = pci_requester_id(dev);
1666     }
1667     if (kvm_arch_fixup_msi_route(&kroute, msg.address, msg.data, dev)) {
1668         kvm_irqchip_release_virq(s, virq);
1669         return -EINVAL;
1670     }
1671 
1672     trace_kvm_irqchip_add_msi_route(dev ? dev->name : (char *)"N/A",
1673                                     vector, virq);
1674 
1675     kvm_add_routing_entry(s, &kroute);
1676     kvm_arch_add_msi_route_post(&kroute, vector, dev);
1677     kvm_irqchip_commit_routes(s);
1678 
1679     return virq;
1680 }
1681 
1682 int kvm_irqchip_update_msi_route(KVMState *s, int virq, MSIMessage msg,
1683                                  PCIDevice *dev)
1684 {
1685     struct kvm_irq_routing_entry kroute = {};
1686 
1687     if (kvm_gsi_direct_mapping()) {
1688         return 0;
1689     }
1690 
1691     if (!kvm_irqchip_in_kernel()) {
1692         return -ENOSYS;
1693     }
1694 
1695     kroute.gsi = virq;
1696     kroute.type = KVM_IRQ_ROUTING_MSI;
1697     kroute.flags = 0;
1698     kroute.u.msi.address_lo = (uint32_t)msg.address;
1699     kroute.u.msi.address_hi = msg.address >> 32;
1700     kroute.u.msi.data = le32_to_cpu(msg.data);
1701     if (pci_available && kvm_msi_devid_required()) {
1702         kroute.flags = KVM_MSI_VALID_DEVID;
1703         kroute.u.msi.devid = pci_requester_id(dev);
1704     }
1705     if (kvm_arch_fixup_msi_route(&kroute, msg.address, msg.data, dev)) {
1706         return -EINVAL;
1707     }
1708 
1709     trace_kvm_irqchip_update_msi_route(virq);
1710 
1711     return kvm_update_routing_entry(s, &kroute);
1712 }
1713 
1714 static int kvm_irqchip_assign_irqfd(KVMState *s, EventNotifier *event,
1715                                     EventNotifier *resample, int virq,
1716                                     bool assign)
1717 {
1718     int fd = event_notifier_get_fd(event);
1719     int rfd = resample ? event_notifier_get_fd(resample) : -1;
1720 
1721     struct kvm_irqfd irqfd = {
1722         .fd = fd,
1723         .gsi = virq,
1724         .flags = assign ? 0 : KVM_IRQFD_FLAG_DEASSIGN,
1725     };
1726 
1727     if (rfd != -1) {
1728         assert(assign);
1729         if (kvm_irqchip_is_split()) {
1730             /*
1731              * When the slow irqchip (e.g. IOAPIC) is in the
1732              * userspace, KVM kernel resamplefd will not work because
1733              * the EOI of the interrupt will be delivered to userspace
1734              * instead, so the KVM kernel resamplefd kick will be
1735              * skipped.  The userspace here mimics what the kernel
1736              * provides with resamplefd, remember the resamplefd and
1737              * kick it when we receive EOI of this IRQ.
1738              *
1739              * This is hackery because IOAPIC is mostly bypassed
1740              * (except EOI broadcasts) when irqfd is used.  However
1741              * this can bring much performance back for split irqchip
1742              * with INTx IRQs (for VFIO, this gives 93% perf of the
1743              * full fast path, which is 46% perf boost comparing to
1744              * the INTx slow path).
1745              */
1746             kvm_resample_fd_insert(virq, resample);
1747         } else {
1748             irqfd.flags |= KVM_IRQFD_FLAG_RESAMPLE;
1749             irqfd.resamplefd = rfd;
1750         }
1751     } else if (!assign) {
1752         if (kvm_irqchip_is_split()) {
1753             kvm_resample_fd_remove(virq);
1754         }
1755     }
1756 
1757     if (!kvm_irqfds_enabled()) {
1758         return -ENOSYS;
1759     }
1760 
1761     return kvm_vm_ioctl(s, KVM_IRQFD, &irqfd);
1762 }
1763 
1764 int kvm_irqchip_add_adapter_route(KVMState *s, AdapterInfo *adapter)
1765 {
1766     struct kvm_irq_routing_entry kroute = {};
1767     int virq;
1768 
1769     if (!kvm_gsi_routing_enabled()) {
1770         return -ENOSYS;
1771     }
1772 
1773     virq = kvm_irqchip_get_virq(s);
1774     if (virq < 0) {
1775         return virq;
1776     }
1777 
1778     kroute.gsi = virq;
1779     kroute.type = KVM_IRQ_ROUTING_S390_ADAPTER;
1780     kroute.flags = 0;
1781     kroute.u.adapter.summary_addr = adapter->summary_addr;
1782     kroute.u.adapter.ind_addr = adapter->ind_addr;
1783     kroute.u.adapter.summary_offset = adapter->summary_offset;
1784     kroute.u.adapter.ind_offset = adapter->ind_offset;
1785     kroute.u.adapter.adapter_id = adapter->adapter_id;
1786 
1787     kvm_add_routing_entry(s, &kroute);
1788 
1789     return virq;
1790 }
1791 
1792 int kvm_irqchip_add_hv_sint_route(KVMState *s, uint32_t vcpu, uint32_t sint)
1793 {
1794     struct kvm_irq_routing_entry kroute = {};
1795     int virq;
1796 
1797     if (!kvm_gsi_routing_enabled()) {
1798         return -ENOSYS;
1799     }
1800     if (!kvm_check_extension(s, KVM_CAP_HYPERV_SYNIC)) {
1801         return -ENOSYS;
1802     }
1803     virq = kvm_irqchip_get_virq(s);
1804     if (virq < 0) {
1805         return virq;
1806     }
1807 
1808     kroute.gsi = virq;
1809     kroute.type = KVM_IRQ_ROUTING_HV_SINT;
1810     kroute.flags = 0;
1811     kroute.u.hv_sint.vcpu = vcpu;
1812     kroute.u.hv_sint.sint = sint;
1813 
1814     kvm_add_routing_entry(s, &kroute);
1815     kvm_irqchip_commit_routes(s);
1816 
1817     return virq;
1818 }
1819 
1820 #else /* !KVM_CAP_IRQ_ROUTING */
1821 
1822 void kvm_init_irq_routing(KVMState *s)
1823 {
1824 }
1825 
1826 void kvm_irqchip_release_virq(KVMState *s, int virq)
1827 {
1828 }
1829 
1830 int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg)
1831 {
1832     abort();
1833 }
1834 
1835 int kvm_irqchip_add_msi_route(KVMState *s, int vector, PCIDevice *dev)
1836 {
1837     return -ENOSYS;
1838 }
1839 
1840 int kvm_irqchip_add_adapter_route(KVMState *s, AdapterInfo *adapter)
1841 {
1842     return -ENOSYS;
1843 }
1844 
1845 int kvm_irqchip_add_hv_sint_route(KVMState *s, uint32_t vcpu, uint32_t sint)
1846 {
1847     return -ENOSYS;
1848 }
1849 
1850 static int kvm_irqchip_assign_irqfd(KVMState *s, EventNotifier *event,
1851                                     EventNotifier *resample, int virq,
1852                                     bool assign)
1853 {
1854     abort();
1855 }
1856 
1857 int kvm_irqchip_update_msi_route(KVMState *s, int virq, MSIMessage msg)
1858 {
1859     return -ENOSYS;
1860 }
1861 #endif /* !KVM_CAP_IRQ_ROUTING */
1862 
1863 int kvm_irqchip_add_irqfd_notifier_gsi(KVMState *s, EventNotifier *n,
1864                                        EventNotifier *rn, int virq)
1865 {
1866     return kvm_irqchip_assign_irqfd(s, n, rn, virq, true);
1867 }
1868 
1869 int kvm_irqchip_remove_irqfd_notifier_gsi(KVMState *s, EventNotifier *n,
1870                                           int virq)
1871 {
1872     return kvm_irqchip_assign_irqfd(s, n, NULL, virq, false);
1873 }
1874 
1875 int kvm_irqchip_add_irqfd_notifier(KVMState *s, EventNotifier *n,
1876                                    EventNotifier *rn, qemu_irq irq)
1877 {
1878     gpointer key, gsi;
1879     gboolean found = g_hash_table_lookup_extended(s->gsimap, irq, &key, &gsi);
1880 
1881     if (!found) {
1882         return -ENXIO;
1883     }
1884     return kvm_irqchip_add_irqfd_notifier_gsi(s, n, rn, GPOINTER_TO_INT(gsi));
1885 }
1886 
1887 int kvm_irqchip_remove_irqfd_notifier(KVMState *s, EventNotifier *n,
1888                                       qemu_irq irq)
1889 {
1890     gpointer key, gsi;
1891     gboolean found = g_hash_table_lookup_extended(s->gsimap, irq, &key, &gsi);
1892 
1893     if (!found) {
1894         return -ENXIO;
1895     }
1896     return kvm_irqchip_remove_irqfd_notifier_gsi(s, n, GPOINTER_TO_INT(gsi));
1897 }
1898 
1899 void kvm_irqchip_set_qemuirq_gsi(KVMState *s, qemu_irq irq, int gsi)
1900 {
1901     g_hash_table_insert(s->gsimap, irq, GINT_TO_POINTER(gsi));
1902 }
1903 
1904 static void kvm_irqchip_create(KVMState *s)
1905 {
1906     int ret;
1907 
1908     assert(s->kernel_irqchip_split != ON_OFF_AUTO_AUTO);
1909     if (kvm_check_extension(s, KVM_CAP_IRQCHIP)) {
1910         ;
1911     } else if (kvm_check_extension(s, KVM_CAP_S390_IRQCHIP)) {
1912         ret = kvm_vm_enable_cap(s, KVM_CAP_S390_IRQCHIP, 0);
1913         if (ret < 0) {
1914             fprintf(stderr, "Enable kernel irqchip failed: %s\n", strerror(-ret));
1915             exit(1);
1916         }
1917     } else {
1918         return;
1919     }
1920 
1921     /* First probe and see if there's a arch-specific hook to create the
1922      * in-kernel irqchip for us */
1923     ret = kvm_arch_irqchip_create(s);
1924     if (ret == 0) {
1925         if (s->kernel_irqchip_split == ON_OFF_AUTO_ON) {
1926             perror("Split IRQ chip mode not supported.");
1927             exit(1);
1928         } else {
1929             ret = kvm_vm_ioctl(s, KVM_CREATE_IRQCHIP);
1930         }
1931     }
1932     if (ret < 0) {
1933         fprintf(stderr, "Create kernel irqchip failed: %s\n", strerror(-ret));
1934         exit(1);
1935     }
1936 
1937     kvm_kernel_irqchip = true;
1938     /* If we have an in-kernel IRQ chip then we must have asynchronous
1939      * interrupt delivery (though the reverse is not necessarily true)
1940      */
1941     kvm_async_interrupts_allowed = true;
1942     kvm_halt_in_kernel_allowed = true;
1943 
1944     kvm_init_irq_routing(s);
1945 
1946     s->gsimap = g_hash_table_new(g_direct_hash, g_direct_equal);
1947 }
1948 
1949 /* Find number of supported CPUs using the recommended
1950  * procedure from the kernel API documentation to cope with
1951  * older kernels that may be missing capabilities.
1952  */
1953 static int kvm_recommended_vcpus(KVMState *s)
1954 {
1955     int ret = kvm_vm_check_extension(s, KVM_CAP_NR_VCPUS);
1956     return (ret) ? ret : 4;
1957 }
1958 
1959 static int kvm_max_vcpus(KVMState *s)
1960 {
1961     int ret = kvm_check_extension(s, KVM_CAP_MAX_VCPUS);
1962     return (ret) ? ret : kvm_recommended_vcpus(s);
1963 }
1964 
1965 static int kvm_max_vcpu_id(KVMState *s)
1966 {
1967     int ret = kvm_check_extension(s, KVM_CAP_MAX_VCPU_ID);
1968     return (ret) ? ret : kvm_max_vcpus(s);
1969 }
1970 
1971 bool kvm_vcpu_id_is_valid(int vcpu_id)
1972 {
1973     KVMState *s = KVM_STATE(current_accel());
1974     return vcpu_id >= 0 && vcpu_id < kvm_max_vcpu_id(s);
1975 }
1976 
1977 static int kvm_init(MachineState *ms)
1978 {
1979     MachineClass *mc = MACHINE_GET_CLASS(ms);
1980     static const char upgrade_note[] =
1981         "Please upgrade to at least kernel 2.6.29 or recent kvm-kmod\n"
1982         "(see http://sourceforge.net/projects/kvm).\n";
1983     struct {
1984         const char *name;
1985         int num;
1986     } num_cpus[] = {
1987         { "SMP",          ms->smp.cpus },
1988         { "hotpluggable", ms->smp.max_cpus },
1989         { NULL, }
1990     }, *nc = num_cpus;
1991     int soft_vcpus_limit, hard_vcpus_limit;
1992     KVMState *s;
1993     const KVMCapabilityInfo *missing_cap;
1994     int ret;
1995     int type = 0;
1996     const char *kvm_type;
1997     uint64_t dirty_log_manual_caps;
1998 
1999     s = KVM_STATE(ms->accelerator);
2000 
2001     /*
2002      * On systems where the kernel can support different base page
2003      * sizes, host page size may be different from TARGET_PAGE_SIZE,
2004      * even with KVM.  TARGET_PAGE_SIZE is assumed to be the minimum
2005      * page size for the system though.
2006      */
2007     assert(TARGET_PAGE_SIZE <= qemu_real_host_page_size);
2008 
2009     s->sigmask_len = 8;
2010 
2011 #ifdef KVM_CAP_SET_GUEST_DEBUG
2012     QTAILQ_INIT(&s->kvm_sw_breakpoints);
2013 #endif
2014     QLIST_INIT(&s->kvm_parked_vcpus);
2015     s->vmfd = -1;
2016     s->fd = qemu_open_old("/dev/kvm", O_RDWR);
2017     if (s->fd == -1) {
2018         fprintf(stderr, "Could not access KVM kernel module: %m\n");
2019         ret = -errno;
2020         goto err;
2021     }
2022 
2023     ret = kvm_ioctl(s, KVM_GET_API_VERSION, 0);
2024     if (ret < KVM_API_VERSION) {
2025         if (ret >= 0) {
2026             ret = -EINVAL;
2027         }
2028         fprintf(stderr, "kvm version too old\n");
2029         goto err;
2030     }
2031 
2032     if (ret > KVM_API_VERSION) {
2033         ret = -EINVAL;
2034         fprintf(stderr, "kvm version not supported\n");
2035         goto err;
2036     }
2037 
2038     kvm_immediate_exit = kvm_check_extension(s, KVM_CAP_IMMEDIATE_EXIT);
2039     s->nr_slots = kvm_check_extension(s, KVM_CAP_NR_MEMSLOTS);
2040 
2041     /* If unspecified, use the default value */
2042     if (!s->nr_slots) {
2043         s->nr_slots = 32;
2044     }
2045 
2046     s->nr_as = kvm_check_extension(s, KVM_CAP_MULTI_ADDRESS_SPACE);
2047     if (s->nr_as <= 1) {
2048         s->nr_as = 1;
2049     }
2050     s->as = g_new0(struct KVMAs, s->nr_as);
2051 
2052     kvm_type = qemu_opt_get(qemu_get_machine_opts(), "kvm-type");
2053     if (mc->kvm_type) {
2054         type = mc->kvm_type(ms, kvm_type);
2055     } else if (kvm_type) {
2056         ret = -EINVAL;
2057         fprintf(stderr, "Invalid argument kvm-type=%s\n", kvm_type);
2058         goto err;
2059     }
2060 
2061     do {
2062         ret = kvm_ioctl(s, KVM_CREATE_VM, type);
2063     } while (ret == -EINTR);
2064 
2065     if (ret < 0) {
2066         fprintf(stderr, "ioctl(KVM_CREATE_VM) failed: %d %s\n", -ret,
2067                 strerror(-ret));
2068 
2069 #ifdef TARGET_S390X
2070         if (ret == -EINVAL) {
2071             fprintf(stderr,
2072                     "Host kernel setup problem detected. Please verify:\n");
2073             fprintf(stderr, "- for kernels supporting the switch_amode or"
2074                     " user_mode parameters, whether\n");
2075             fprintf(stderr,
2076                     "  user space is running in primary address space\n");
2077             fprintf(stderr,
2078                     "- for kernels supporting the vm.allocate_pgste sysctl, "
2079                     "whether it is enabled\n");
2080         }
2081 #endif
2082         goto err;
2083     }
2084 
2085     s->vmfd = ret;
2086 
2087     /* check the vcpu limits */
2088     soft_vcpus_limit = kvm_recommended_vcpus(s);
2089     hard_vcpus_limit = kvm_max_vcpus(s);
2090 
2091     while (nc->name) {
2092         if (nc->num > soft_vcpus_limit) {
2093             warn_report("Number of %s cpus requested (%d) exceeds "
2094                         "the recommended cpus supported by KVM (%d)",
2095                         nc->name, nc->num, soft_vcpus_limit);
2096 
2097             if (nc->num > hard_vcpus_limit) {
2098                 fprintf(stderr, "Number of %s cpus requested (%d) exceeds "
2099                         "the maximum cpus supported by KVM (%d)\n",
2100                         nc->name, nc->num, hard_vcpus_limit);
2101                 exit(1);
2102             }
2103         }
2104         nc++;
2105     }
2106 
2107     missing_cap = kvm_check_extension_list(s, kvm_required_capabilites);
2108     if (!missing_cap) {
2109         missing_cap =
2110             kvm_check_extension_list(s, kvm_arch_required_capabilities);
2111     }
2112     if (missing_cap) {
2113         ret = -EINVAL;
2114         fprintf(stderr, "kvm does not support %s\n%s",
2115                 missing_cap->name, upgrade_note);
2116         goto err;
2117     }
2118 
2119     s->coalesced_mmio = kvm_check_extension(s, KVM_CAP_COALESCED_MMIO);
2120     s->coalesced_pio = s->coalesced_mmio &&
2121                        kvm_check_extension(s, KVM_CAP_COALESCED_PIO);
2122 
2123     dirty_log_manual_caps =
2124         kvm_check_extension(s, KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2);
2125     dirty_log_manual_caps &= (KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE |
2126                               KVM_DIRTY_LOG_INITIALLY_SET);
2127     s->manual_dirty_log_protect = dirty_log_manual_caps;
2128     if (dirty_log_manual_caps) {
2129         ret = kvm_vm_enable_cap(s, KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2, 0,
2130                                    dirty_log_manual_caps);
2131         if (ret) {
2132             warn_report("Trying to enable capability %"PRIu64" of "
2133                         "KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2 but failed. "
2134                         "Falling back to the legacy mode. ",
2135                         dirty_log_manual_caps);
2136             s->manual_dirty_log_protect = 0;
2137         }
2138     }
2139 
2140 #ifdef KVM_CAP_VCPU_EVENTS
2141     s->vcpu_events = kvm_check_extension(s, KVM_CAP_VCPU_EVENTS);
2142 #endif
2143 
2144     s->robust_singlestep =
2145         kvm_check_extension(s, KVM_CAP_X86_ROBUST_SINGLESTEP);
2146 
2147 #ifdef KVM_CAP_DEBUGREGS
2148     s->debugregs = kvm_check_extension(s, KVM_CAP_DEBUGREGS);
2149 #endif
2150 
2151     s->max_nested_state_len = kvm_check_extension(s, KVM_CAP_NESTED_STATE);
2152 
2153 #ifdef KVM_CAP_IRQ_ROUTING
2154     kvm_direct_msi_allowed = (kvm_check_extension(s, KVM_CAP_SIGNAL_MSI) > 0);
2155 #endif
2156 
2157     s->intx_set_mask = kvm_check_extension(s, KVM_CAP_PCI_2_3);
2158 
2159     s->irq_set_ioctl = KVM_IRQ_LINE;
2160     if (kvm_check_extension(s, KVM_CAP_IRQ_INJECT_STATUS)) {
2161         s->irq_set_ioctl = KVM_IRQ_LINE_STATUS;
2162     }
2163 
2164     kvm_readonly_mem_allowed =
2165         (kvm_check_extension(s, KVM_CAP_READONLY_MEM) > 0);
2166 
2167     kvm_eventfds_allowed =
2168         (kvm_check_extension(s, KVM_CAP_IOEVENTFD) > 0);
2169 
2170     kvm_irqfds_allowed =
2171         (kvm_check_extension(s, KVM_CAP_IRQFD) > 0);
2172 
2173     kvm_resamplefds_allowed =
2174         (kvm_check_extension(s, KVM_CAP_IRQFD_RESAMPLE) > 0);
2175 
2176     kvm_vm_attributes_allowed =
2177         (kvm_check_extension(s, KVM_CAP_VM_ATTRIBUTES) > 0);
2178 
2179     kvm_ioeventfd_any_length_allowed =
2180         (kvm_check_extension(s, KVM_CAP_IOEVENTFD_ANY_LENGTH) > 0);
2181 
2182     kvm_state = s;
2183 
2184     /*
2185      * if memory encryption object is specified then initialize the memory
2186      * encryption context.
2187      */
2188     if (ms->memory_encryption) {
2189         kvm_state->memcrypt_handle = sev_guest_init(ms->memory_encryption);
2190         if (!kvm_state->memcrypt_handle) {
2191             ret = -1;
2192             goto err;
2193         }
2194 
2195         kvm_state->memcrypt_encrypt_data = sev_encrypt_data;
2196     }
2197 
2198     ret = kvm_arch_init(ms, s);
2199     if (ret < 0) {
2200         goto err;
2201     }
2202 
2203     if (s->kernel_irqchip_split == ON_OFF_AUTO_AUTO) {
2204         s->kernel_irqchip_split = mc->default_kernel_irqchip_split ? ON_OFF_AUTO_ON : ON_OFF_AUTO_OFF;
2205     }
2206 
2207     qemu_register_reset(kvm_unpoison_all, NULL);
2208 
2209     if (s->kernel_irqchip_allowed) {
2210         kvm_irqchip_create(s);
2211     }
2212 
2213     if (kvm_eventfds_allowed) {
2214         s->memory_listener.listener.eventfd_add = kvm_mem_ioeventfd_add;
2215         s->memory_listener.listener.eventfd_del = kvm_mem_ioeventfd_del;
2216     }
2217     s->memory_listener.listener.coalesced_io_add = kvm_coalesce_mmio_region;
2218     s->memory_listener.listener.coalesced_io_del = kvm_uncoalesce_mmio_region;
2219 
2220     kvm_memory_listener_register(s, &s->memory_listener,
2221                                  &address_space_memory, 0);
2222     memory_listener_register(&kvm_io_listener,
2223                              &address_space_io);
2224     memory_listener_register(&kvm_coalesced_pio_listener,
2225                              &address_space_io);
2226 
2227     s->many_ioeventfds = kvm_check_many_ioeventfds();
2228 
2229     s->sync_mmu = !!kvm_vm_check_extension(kvm_state, KVM_CAP_SYNC_MMU);
2230     if (!s->sync_mmu) {
2231         ret = ram_block_discard_disable(true);
2232         assert(!ret);
2233     }
2234 
2235     return 0;
2236 
2237 err:
2238     assert(ret < 0);
2239     if (s->vmfd >= 0) {
2240         close(s->vmfd);
2241     }
2242     if (s->fd != -1) {
2243         close(s->fd);
2244     }
2245     g_free(s->memory_listener.slots);
2246 
2247     return ret;
2248 }
2249 
2250 void kvm_set_sigmask_len(KVMState *s, unsigned int sigmask_len)
2251 {
2252     s->sigmask_len = sigmask_len;
2253 }
2254 
2255 static void kvm_handle_io(uint16_t port, MemTxAttrs attrs, void *data, int direction,
2256                           int size, uint32_t count)
2257 {
2258     int i;
2259     uint8_t *ptr = data;
2260 
2261     for (i = 0; i < count; i++) {
2262         address_space_rw(&address_space_io, port, attrs,
2263                          ptr, size,
2264                          direction == KVM_EXIT_IO_OUT);
2265         ptr += size;
2266     }
2267 }
2268 
2269 static int kvm_handle_internal_error(CPUState *cpu, struct kvm_run *run)
2270 {
2271     fprintf(stderr, "KVM internal error. Suberror: %d\n",
2272             run->internal.suberror);
2273 
2274     if (kvm_check_extension(kvm_state, KVM_CAP_INTERNAL_ERROR_DATA)) {
2275         int i;
2276 
2277         for (i = 0; i < run->internal.ndata; ++i) {
2278             fprintf(stderr, "extra data[%d]: %"PRIx64"\n",
2279                     i, (uint64_t)run->internal.data[i]);
2280         }
2281     }
2282     if (run->internal.suberror == KVM_INTERNAL_ERROR_EMULATION) {
2283         fprintf(stderr, "emulation failure\n");
2284         if (!kvm_arch_stop_on_emulation_error(cpu)) {
2285             cpu_dump_state(cpu, stderr, CPU_DUMP_CODE);
2286             return EXCP_INTERRUPT;
2287         }
2288     }
2289     /* FIXME: Should trigger a qmp message to let management know
2290      * something went wrong.
2291      */
2292     return -1;
2293 }
2294 
2295 void kvm_flush_coalesced_mmio_buffer(void)
2296 {
2297     KVMState *s = kvm_state;
2298 
2299     if (s->coalesced_flush_in_progress) {
2300         return;
2301     }
2302 
2303     s->coalesced_flush_in_progress = true;
2304 
2305     if (s->coalesced_mmio_ring) {
2306         struct kvm_coalesced_mmio_ring *ring = s->coalesced_mmio_ring;
2307         while (ring->first != ring->last) {
2308             struct kvm_coalesced_mmio *ent;
2309 
2310             ent = &ring->coalesced_mmio[ring->first];
2311 
2312             if (ent->pio == 1) {
2313                 address_space_write(&address_space_io, ent->phys_addr,
2314                                     MEMTXATTRS_UNSPECIFIED, ent->data,
2315                                     ent->len);
2316             } else {
2317                 cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len);
2318             }
2319             smp_wmb();
2320             ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX;
2321         }
2322     }
2323 
2324     s->coalesced_flush_in_progress = false;
2325 }
2326 
2327 static void do_kvm_cpu_synchronize_state(CPUState *cpu, run_on_cpu_data arg)
2328 {
2329     if (!cpu->vcpu_dirty) {
2330         kvm_arch_get_registers(cpu);
2331         cpu->vcpu_dirty = true;
2332     }
2333 }
2334 
2335 void kvm_cpu_synchronize_state(CPUState *cpu)
2336 {
2337     if (!cpu->vcpu_dirty) {
2338         run_on_cpu(cpu, do_kvm_cpu_synchronize_state, RUN_ON_CPU_NULL);
2339     }
2340 }
2341 
2342 static void do_kvm_cpu_synchronize_post_reset(CPUState *cpu, run_on_cpu_data arg)
2343 {
2344     kvm_arch_put_registers(cpu, KVM_PUT_RESET_STATE);
2345     cpu->vcpu_dirty = false;
2346 }
2347 
2348 void kvm_cpu_synchronize_post_reset(CPUState *cpu)
2349 {
2350     run_on_cpu(cpu, do_kvm_cpu_synchronize_post_reset, RUN_ON_CPU_NULL);
2351 }
2352 
2353 static void do_kvm_cpu_synchronize_post_init(CPUState *cpu, run_on_cpu_data arg)
2354 {
2355     kvm_arch_put_registers(cpu, KVM_PUT_FULL_STATE);
2356     cpu->vcpu_dirty = false;
2357 }
2358 
2359 void kvm_cpu_synchronize_post_init(CPUState *cpu)
2360 {
2361     run_on_cpu(cpu, do_kvm_cpu_synchronize_post_init, RUN_ON_CPU_NULL);
2362 }
2363 
2364 static void do_kvm_cpu_synchronize_pre_loadvm(CPUState *cpu, run_on_cpu_data arg)
2365 {
2366     cpu->vcpu_dirty = true;
2367 }
2368 
2369 void kvm_cpu_synchronize_pre_loadvm(CPUState *cpu)
2370 {
2371     run_on_cpu(cpu, do_kvm_cpu_synchronize_pre_loadvm, RUN_ON_CPU_NULL);
2372 }
2373 
2374 #ifdef KVM_HAVE_MCE_INJECTION
2375 static __thread void *pending_sigbus_addr;
2376 static __thread int pending_sigbus_code;
2377 static __thread bool have_sigbus_pending;
2378 #endif
2379 
2380 static void kvm_cpu_kick(CPUState *cpu)
2381 {
2382     atomic_set(&cpu->kvm_run->immediate_exit, 1);
2383 }
2384 
2385 static void kvm_cpu_kick_self(void)
2386 {
2387     if (kvm_immediate_exit) {
2388         kvm_cpu_kick(current_cpu);
2389     } else {
2390         qemu_cpu_kick_self();
2391     }
2392 }
2393 
2394 static void kvm_eat_signals(CPUState *cpu)
2395 {
2396     struct timespec ts = { 0, 0 };
2397     siginfo_t siginfo;
2398     sigset_t waitset;
2399     sigset_t chkset;
2400     int r;
2401 
2402     if (kvm_immediate_exit) {
2403         atomic_set(&cpu->kvm_run->immediate_exit, 0);
2404         /* Write kvm_run->immediate_exit before the cpu->exit_request
2405          * write in kvm_cpu_exec.
2406          */
2407         smp_wmb();
2408         return;
2409     }
2410 
2411     sigemptyset(&waitset);
2412     sigaddset(&waitset, SIG_IPI);
2413 
2414     do {
2415         r = sigtimedwait(&waitset, &siginfo, &ts);
2416         if (r == -1 && !(errno == EAGAIN || errno == EINTR)) {
2417             perror("sigtimedwait");
2418             exit(1);
2419         }
2420 
2421         r = sigpending(&chkset);
2422         if (r == -1) {
2423             perror("sigpending");
2424             exit(1);
2425         }
2426     } while (sigismember(&chkset, SIG_IPI));
2427 }
2428 
2429 int kvm_cpu_exec(CPUState *cpu)
2430 {
2431     struct kvm_run *run = cpu->kvm_run;
2432     int ret, run_ret;
2433 
2434     DPRINTF("kvm_cpu_exec()\n");
2435 
2436     if (kvm_arch_process_async_events(cpu)) {
2437         atomic_set(&cpu->exit_request, 0);
2438         return EXCP_HLT;
2439     }
2440 
2441     qemu_mutex_unlock_iothread();
2442     cpu_exec_start(cpu);
2443 
2444     do {
2445         MemTxAttrs attrs;
2446 
2447         if (cpu->vcpu_dirty) {
2448             kvm_arch_put_registers(cpu, KVM_PUT_RUNTIME_STATE);
2449             cpu->vcpu_dirty = false;
2450         }
2451 
2452         kvm_arch_pre_run(cpu, run);
2453         if (atomic_read(&cpu->exit_request)) {
2454             DPRINTF("interrupt exit requested\n");
2455             /*
2456              * KVM requires us to reenter the kernel after IO exits to complete
2457              * instruction emulation. This self-signal will ensure that we
2458              * leave ASAP again.
2459              */
2460             kvm_cpu_kick_self();
2461         }
2462 
2463         /* Read cpu->exit_request before KVM_RUN reads run->immediate_exit.
2464          * Matching barrier in kvm_eat_signals.
2465          */
2466         smp_rmb();
2467 
2468         run_ret = kvm_vcpu_ioctl(cpu, KVM_RUN, 0);
2469 
2470         attrs = kvm_arch_post_run(cpu, run);
2471 
2472 #ifdef KVM_HAVE_MCE_INJECTION
2473         if (unlikely(have_sigbus_pending)) {
2474             qemu_mutex_lock_iothread();
2475             kvm_arch_on_sigbus_vcpu(cpu, pending_sigbus_code,
2476                                     pending_sigbus_addr);
2477             have_sigbus_pending = false;
2478             qemu_mutex_unlock_iothread();
2479         }
2480 #endif
2481 
2482         if (run_ret < 0) {
2483             if (run_ret == -EINTR || run_ret == -EAGAIN) {
2484                 DPRINTF("io window exit\n");
2485                 kvm_eat_signals(cpu);
2486                 ret = EXCP_INTERRUPT;
2487                 break;
2488             }
2489             fprintf(stderr, "error: kvm run failed %s\n",
2490                     strerror(-run_ret));
2491 #ifdef TARGET_PPC
2492             if (run_ret == -EBUSY) {
2493                 fprintf(stderr,
2494                         "This is probably because your SMT is enabled.\n"
2495                         "VCPU can only run on primary threads with all "
2496                         "secondary threads offline.\n");
2497             }
2498 #endif
2499             ret = -1;
2500             break;
2501         }
2502 
2503         trace_kvm_run_exit(cpu->cpu_index, run->exit_reason);
2504         switch (run->exit_reason) {
2505         case KVM_EXIT_IO:
2506             DPRINTF("handle_io\n");
2507             /* Called outside BQL */
2508             kvm_handle_io(run->io.port, attrs,
2509                           (uint8_t *)run + run->io.data_offset,
2510                           run->io.direction,
2511                           run->io.size,
2512                           run->io.count);
2513             ret = 0;
2514             break;
2515         case KVM_EXIT_MMIO:
2516             DPRINTF("handle_mmio\n");
2517             /* Called outside BQL */
2518             address_space_rw(&address_space_memory,
2519                              run->mmio.phys_addr, attrs,
2520                              run->mmio.data,
2521                              run->mmio.len,
2522                              run->mmio.is_write);
2523             ret = 0;
2524             break;
2525         case KVM_EXIT_IRQ_WINDOW_OPEN:
2526             DPRINTF("irq_window_open\n");
2527             ret = EXCP_INTERRUPT;
2528             break;
2529         case KVM_EXIT_SHUTDOWN:
2530             DPRINTF("shutdown\n");
2531             qemu_system_reset_request(SHUTDOWN_CAUSE_GUEST_RESET);
2532             ret = EXCP_INTERRUPT;
2533             break;
2534         case KVM_EXIT_UNKNOWN:
2535             fprintf(stderr, "KVM: unknown exit, hardware reason %" PRIx64 "\n",
2536                     (uint64_t)run->hw.hardware_exit_reason);
2537             ret = -1;
2538             break;
2539         case KVM_EXIT_INTERNAL_ERROR:
2540             ret = kvm_handle_internal_error(cpu, run);
2541             break;
2542         case KVM_EXIT_SYSTEM_EVENT:
2543             switch (run->system_event.type) {
2544             case KVM_SYSTEM_EVENT_SHUTDOWN:
2545                 qemu_system_shutdown_request(SHUTDOWN_CAUSE_GUEST_SHUTDOWN);
2546                 ret = EXCP_INTERRUPT;
2547                 break;
2548             case KVM_SYSTEM_EVENT_RESET:
2549                 qemu_system_reset_request(SHUTDOWN_CAUSE_GUEST_RESET);
2550                 ret = EXCP_INTERRUPT;
2551                 break;
2552             case KVM_SYSTEM_EVENT_CRASH:
2553                 kvm_cpu_synchronize_state(cpu);
2554                 qemu_mutex_lock_iothread();
2555                 qemu_system_guest_panicked(cpu_get_crash_info(cpu));
2556                 qemu_mutex_unlock_iothread();
2557                 ret = 0;
2558                 break;
2559             default:
2560                 DPRINTF("kvm_arch_handle_exit\n");
2561                 ret = kvm_arch_handle_exit(cpu, run);
2562                 break;
2563             }
2564             break;
2565         default:
2566             DPRINTF("kvm_arch_handle_exit\n");
2567             ret = kvm_arch_handle_exit(cpu, run);
2568             break;
2569         }
2570     } while (ret == 0);
2571 
2572     cpu_exec_end(cpu);
2573     qemu_mutex_lock_iothread();
2574 
2575     if (ret < 0) {
2576         cpu_dump_state(cpu, stderr, CPU_DUMP_CODE);
2577         vm_stop(RUN_STATE_INTERNAL_ERROR);
2578     }
2579 
2580     atomic_set(&cpu->exit_request, 0);
2581     return ret;
2582 }
2583 
2584 int kvm_ioctl(KVMState *s, int type, ...)
2585 {
2586     int ret;
2587     void *arg;
2588     va_list ap;
2589 
2590     va_start(ap, type);
2591     arg = va_arg(ap, void *);
2592     va_end(ap);
2593 
2594     trace_kvm_ioctl(type, arg);
2595     ret = ioctl(s->fd, type, arg);
2596     if (ret == -1) {
2597         ret = -errno;
2598     }
2599     return ret;
2600 }
2601 
2602 int kvm_vm_ioctl(KVMState *s, int type, ...)
2603 {
2604     int ret;
2605     void *arg;
2606     va_list ap;
2607 
2608     va_start(ap, type);
2609     arg = va_arg(ap, void *);
2610     va_end(ap);
2611 
2612     trace_kvm_vm_ioctl(type, arg);
2613     ret = ioctl(s->vmfd, type, arg);
2614     if (ret == -1) {
2615         ret = -errno;
2616     }
2617     return ret;
2618 }
2619 
2620 int kvm_vcpu_ioctl(CPUState *cpu, int type, ...)
2621 {
2622     int ret;
2623     void *arg;
2624     va_list ap;
2625 
2626     va_start(ap, type);
2627     arg = va_arg(ap, void *);
2628     va_end(ap);
2629 
2630     trace_kvm_vcpu_ioctl(cpu->cpu_index, type, arg);
2631     ret = ioctl(cpu->kvm_fd, type, arg);
2632     if (ret == -1) {
2633         ret = -errno;
2634     }
2635     return ret;
2636 }
2637 
2638 int kvm_device_ioctl(int fd, int type, ...)
2639 {
2640     int ret;
2641     void *arg;
2642     va_list ap;
2643 
2644     va_start(ap, type);
2645     arg = va_arg(ap, void *);
2646     va_end(ap);
2647 
2648     trace_kvm_device_ioctl(fd, type, arg);
2649     ret = ioctl(fd, type, arg);
2650     if (ret == -1) {
2651         ret = -errno;
2652     }
2653     return ret;
2654 }
2655 
2656 int kvm_vm_check_attr(KVMState *s, uint32_t group, uint64_t attr)
2657 {
2658     int ret;
2659     struct kvm_device_attr attribute = {
2660         .group = group,
2661         .attr = attr,
2662     };
2663 
2664     if (!kvm_vm_attributes_allowed) {
2665         return 0;
2666     }
2667 
2668     ret = kvm_vm_ioctl(s, KVM_HAS_DEVICE_ATTR, &attribute);
2669     /* kvm returns 0 on success for HAS_DEVICE_ATTR */
2670     return ret ? 0 : 1;
2671 }
2672 
2673 int kvm_device_check_attr(int dev_fd, uint32_t group, uint64_t attr)
2674 {
2675     struct kvm_device_attr attribute = {
2676         .group = group,
2677         .attr = attr,
2678         .flags = 0,
2679     };
2680 
2681     return kvm_device_ioctl(dev_fd, KVM_HAS_DEVICE_ATTR, &attribute) ? 0 : 1;
2682 }
2683 
2684 int kvm_device_access(int fd, int group, uint64_t attr,
2685                       void *val, bool write, Error **errp)
2686 {
2687     struct kvm_device_attr kvmattr;
2688     int err;
2689 
2690     kvmattr.flags = 0;
2691     kvmattr.group = group;
2692     kvmattr.attr = attr;
2693     kvmattr.addr = (uintptr_t)val;
2694 
2695     err = kvm_device_ioctl(fd,
2696                            write ? KVM_SET_DEVICE_ATTR : KVM_GET_DEVICE_ATTR,
2697                            &kvmattr);
2698     if (err < 0) {
2699         error_setg_errno(errp, -err,
2700                          "KVM_%s_DEVICE_ATTR failed: Group %d "
2701                          "attr 0x%016" PRIx64,
2702                          write ? "SET" : "GET", group, attr);
2703     }
2704     return err;
2705 }
2706 
2707 bool kvm_has_sync_mmu(void)
2708 {
2709     return kvm_state->sync_mmu;
2710 }
2711 
2712 int kvm_has_vcpu_events(void)
2713 {
2714     return kvm_state->vcpu_events;
2715 }
2716 
2717 int kvm_has_robust_singlestep(void)
2718 {
2719     return kvm_state->robust_singlestep;
2720 }
2721 
2722 int kvm_has_debugregs(void)
2723 {
2724     return kvm_state->debugregs;
2725 }
2726 
2727 int kvm_max_nested_state_length(void)
2728 {
2729     return kvm_state->max_nested_state_len;
2730 }
2731 
2732 int kvm_has_many_ioeventfds(void)
2733 {
2734     if (!kvm_enabled()) {
2735         return 0;
2736     }
2737     return kvm_state->many_ioeventfds;
2738 }
2739 
2740 int kvm_has_gsi_routing(void)
2741 {
2742 #ifdef KVM_CAP_IRQ_ROUTING
2743     return kvm_check_extension(kvm_state, KVM_CAP_IRQ_ROUTING);
2744 #else
2745     return false;
2746 #endif
2747 }
2748 
2749 int kvm_has_intx_set_mask(void)
2750 {
2751     return kvm_state->intx_set_mask;
2752 }
2753 
2754 bool kvm_arm_supports_user_irq(void)
2755 {
2756     return kvm_check_extension(kvm_state, KVM_CAP_ARM_USER_IRQ);
2757 }
2758 
2759 #ifdef KVM_CAP_SET_GUEST_DEBUG
2760 struct kvm_sw_breakpoint *kvm_find_sw_breakpoint(CPUState *cpu,
2761                                                  target_ulong pc)
2762 {
2763     struct kvm_sw_breakpoint *bp;
2764 
2765     QTAILQ_FOREACH(bp, &cpu->kvm_state->kvm_sw_breakpoints, entry) {
2766         if (bp->pc == pc) {
2767             return bp;
2768         }
2769     }
2770     return NULL;
2771 }
2772 
2773 int kvm_sw_breakpoints_active(CPUState *cpu)
2774 {
2775     return !QTAILQ_EMPTY(&cpu->kvm_state->kvm_sw_breakpoints);
2776 }
2777 
2778 struct kvm_set_guest_debug_data {
2779     struct kvm_guest_debug dbg;
2780     int err;
2781 };
2782 
2783 static void kvm_invoke_set_guest_debug(CPUState *cpu, run_on_cpu_data data)
2784 {
2785     struct kvm_set_guest_debug_data *dbg_data =
2786         (struct kvm_set_guest_debug_data *) data.host_ptr;
2787 
2788     dbg_data->err = kvm_vcpu_ioctl(cpu, KVM_SET_GUEST_DEBUG,
2789                                    &dbg_data->dbg);
2790 }
2791 
2792 int kvm_update_guest_debug(CPUState *cpu, unsigned long reinject_trap)
2793 {
2794     struct kvm_set_guest_debug_data data;
2795 
2796     data.dbg.control = reinject_trap;
2797 
2798     if (cpu->singlestep_enabled) {
2799         data.dbg.control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_SINGLESTEP;
2800     }
2801     kvm_arch_update_guest_debug(cpu, &data.dbg);
2802 
2803     run_on_cpu(cpu, kvm_invoke_set_guest_debug,
2804                RUN_ON_CPU_HOST_PTR(&data));
2805     return data.err;
2806 }
2807 
2808 int kvm_insert_breakpoint(CPUState *cpu, target_ulong addr,
2809                           target_ulong len, int type)
2810 {
2811     struct kvm_sw_breakpoint *bp;
2812     int err;
2813 
2814     if (type == GDB_BREAKPOINT_SW) {
2815         bp = kvm_find_sw_breakpoint(cpu, addr);
2816         if (bp) {
2817             bp->use_count++;
2818             return 0;
2819         }
2820 
2821         bp = g_malloc(sizeof(struct kvm_sw_breakpoint));
2822         bp->pc = addr;
2823         bp->use_count = 1;
2824         err = kvm_arch_insert_sw_breakpoint(cpu, bp);
2825         if (err) {
2826             g_free(bp);
2827             return err;
2828         }
2829 
2830         QTAILQ_INSERT_HEAD(&cpu->kvm_state->kvm_sw_breakpoints, bp, entry);
2831     } else {
2832         err = kvm_arch_insert_hw_breakpoint(addr, len, type);
2833         if (err) {
2834             return err;
2835         }
2836     }
2837 
2838     CPU_FOREACH(cpu) {
2839         err = kvm_update_guest_debug(cpu, 0);
2840         if (err) {
2841             return err;
2842         }
2843     }
2844     return 0;
2845 }
2846 
2847 int kvm_remove_breakpoint(CPUState *cpu, target_ulong addr,
2848                           target_ulong len, int type)
2849 {
2850     struct kvm_sw_breakpoint *bp;
2851     int err;
2852 
2853     if (type == GDB_BREAKPOINT_SW) {
2854         bp = kvm_find_sw_breakpoint(cpu, addr);
2855         if (!bp) {
2856             return -ENOENT;
2857         }
2858 
2859         if (bp->use_count > 1) {
2860             bp->use_count--;
2861             return 0;
2862         }
2863 
2864         err = kvm_arch_remove_sw_breakpoint(cpu, bp);
2865         if (err) {
2866             return err;
2867         }
2868 
2869         QTAILQ_REMOVE(&cpu->kvm_state->kvm_sw_breakpoints, bp, entry);
2870         g_free(bp);
2871     } else {
2872         err = kvm_arch_remove_hw_breakpoint(addr, len, type);
2873         if (err) {
2874             return err;
2875         }
2876     }
2877 
2878     CPU_FOREACH(cpu) {
2879         err = kvm_update_guest_debug(cpu, 0);
2880         if (err) {
2881             return err;
2882         }
2883     }
2884     return 0;
2885 }
2886 
2887 void kvm_remove_all_breakpoints(CPUState *cpu)
2888 {
2889     struct kvm_sw_breakpoint *bp, *next;
2890     KVMState *s = cpu->kvm_state;
2891     CPUState *tmpcpu;
2892 
2893     QTAILQ_FOREACH_SAFE(bp, &s->kvm_sw_breakpoints, entry, next) {
2894         if (kvm_arch_remove_sw_breakpoint(cpu, bp) != 0) {
2895             /* Try harder to find a CPU that currently sees the breakpoint. */
2896             CPU_FOREACH(tmpcpu) {
2897                 if (kvm_arch_remove_sw_breakpoint(tmpcpu, bp) == 0) {
2898                     break;
2899                 }
2900             }
2901         }
2902         QTAILQ_REMOVE(&s->kvm_sw_breakpoints, bp, entry);
2903         g_free(bp);
2904     }
2905     kvm_arch_remove_all_hw_breakpoints();
2906 
2907     CPU_FOREACH(cpu) {
2908         kvm_update_guest_debug(cpu, 0);
2909     }
2910 }
2911 
2912 #else /* !KVM_CAP_SET_GUEST_DEBUG */
2913 
2914 int kvm_update_guest_debug(CPUState *cpu, unsigned long reinject_trap)
2915 {
2916     return -EINVAL;
2917 }
2918 
2919 int kvm_insert_breakpoint(CPUState *cpu, target_ulong addr,
2920                           target_ulong len, int type)
2921 {
2922     return -EINVAL;
2923 }
2924 
2925 int kvm_remove_breakpoint(CPUState *cpu, target_ulong addr,
2926                           target_ulong len, int type)
2927 {
2928     return -EINVAL;
2929 }
2930 
2931 void kvm_remove_all_breakpoints(CPUState *cpu)
2932 {
2933 }
2934 #endif /* !KVM_CAP_SET_GUEST_DEBUG */
2935 
2936 static int kvm_set_signal_mask(CPUState *cpu, const sigset_t *sigset)
2937 {
2938     KVMState *s = kvm_state;
2939     struct kvm_signal_mask *sigmask;
2940     int r;
2941 
2942     sigmask = g_malloc(sizeof(*sigmask) + sizeof(*sigset));
2943 
2944     sigmask->len = s->sigmask_len;
2945     memcpy(sigmask->sigset, sigset, sizeof(*sigset));
2946     r = kvm_vcpu_ioctl(cpu, KVM_SET_SIGNAL_MASK, sigmask);
2947     g_free(sigmask);
2948 
2949     return r;
2950 }
2951 
2952 static void kvm_ipi_signal(int sig)
2953 {
2954     if (current_cpu) {
2955         assert(kvm_immediate_exit);
2956         kvm_cpu_kick(current_cpu);
2957     }
2958 }
2959 
2960 void kvm_init_cpu_signals(CPUState *cpu)
2961 {
2962     int r;
2963     sigset_t set;
2964     struct sigaction sigact;
2965 
2966     memset(&sigact, 0, sizeof(sigact));
2967     sigact.sa_handler = kvm_ipi_signal;
2968     sigaction(SIG_IPI, &sigact, NULL);
2969 
2970     pthread_sigmask(SIG_BLOCK, NULL, &set);
2971 #if defined KVM_HAVE_MCE_INJECTION
2972     sigdelset(&set, SIGBUS);
2973     pthread_sigmask(SIG_SETMASK, &set, NULL);
2974 #endif
2975     sigdelset(&set, SIG_IPI);
2976     if (kvm_immediate_exit) {
2977         r = pthread_sigmask(SIG_SETMASK, &set, NULL);
2978     } else {
2979         r = kvm_set_signal_mask(cpu, &set);
2980     }
2981     if (r) {
2982         fprintf(stderr, "kvm_set_signal_mask: %s\n", strerror(-r));
2983         exit(1);
2984     }
2985 }
2986 
2987 /* Called asynchronously in VCPU thread.  */
2988 int kvm_on_sigbus_vcpu(CPUState *cpu, int code, void *addr)
2989 {
2990 #ifdef KVM_HAVE_MCE_INJECTION
2991     if (have_sigbus_pending) {
2992         return 1;
2993     }
2994     have_sigbus_pending = true;
2995     pending_sigbus_addr = addr;
2996     pending_sigbus_code = code;
2997     atomic_set(&cpu->exit_request, 1);
2998     return 0;
2999 #else
3000     return 1;
3001 #endif
3002 }
3003 
3004 /* Called synchronously (via signalfd) in main thread.  */
3005 int kvm_on_sigbus(int code, void *addr)
3006 {
3007 #ifdef KVM_HAVE_MCE_INJECTION
3008     /* Action required MCE kills the process if SIGBUS is blocked.  Because
3009      * that's what happens in the I/O thread, where we handle MCE via signalfd,
3010      * we can only get action optional here.
3011      */
3012     assert(code != BUS_MCEERR_AR);
3013     kvm_arch_on_sigbus_vcpu(first_cpu, code, addr);
3014     return 0;
3015 #else
3016     return 1;
3017 #endif
3018 }
3019 
3020 int kvm_create_device(KVMState *s, uint64_t type, bool test)
3021 {
3022     int ret;
3023     struct kvm_create_device create_dev;
3024 
3025     create_dev.type = type;
3026     create_dev.fd = -1;
3027     create_dev.flags = test ? KVM_CREATE_DEVICE_TEST : 0;
3028 
3029     if (!kvm_check_extension(s, KVM_CAP_DEVICE_CTRL)) {
3030         return -ENOTSUP;
3031     }
3032 
3033     ret = kvm_vm_ioctl(s, KVM_CREATE_DEVICE, &create_dev);
3034     if (ret) {
3035         return ret;
3036     }
3037 
3038     return test ? 0 : create_dev.fd;
3039 }
3040 
3041 bool kvm_device_supported(int vmfd, uint64_t type)
3042 {
3043     struct kvm_create_device create_dev = {
3044         .type = type,
3045         .fd = -1,
3046         .flags = KVM_CREATE_DEVICE_TEST,
3047     };
3048 
3049     if (ioctl(vmfd, KVM_CHECK_EXTENSION, KVM_CAP_DEVICE_CTRL) <= 0) {
3050         return false;
3051     }
3052 
3053     return (ioctl(vmfd, KVM_CREATE_DEVICE, &create_dev) >= 0);
3054 }
3055 
3056 int kvm_set_one_reg(CPUState *cs, uint64_t id, void *source)
3057 {
3058     struct kvm_one_reg reg;
3059     int r;
3060 
3061     reg.id = id;
3062     reg.addr = (uintptr_t) source;
3063     r = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, &reg);
3064     if (r) {
3065         trace_kvm_failed_reg_set(id, strerror(-r));
3066     }
3067     return r;
3068 }
3069 
3070 int kvm_get_one_reg(CPUState *cs, uint64_t id, void *target)
3071 {
3072     struct kvm_one_reg reg;
3073     int r;
3074 
3075     reg.id = id;
3076     reg.addr = (uintptr_t) target;
3077     r = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &reg);
3078     if (r) {
3079         trace_kvm_failed_reg_get(id, strerror(-r));
3080     }
3081     return r;
3082 }
3083 
3084 static bool kvm_accel_has_memory(MachineState *ms, AddressSpace *as,
3085                                  hwaddr start_addr, hwaddr size)
3086 {
3087     KVMState *kvm = KVM_STATE(ms->accelerator);
3088     int i;
3089 
3090     for (i = 0; i < kvm->nr_as; ++i) {
3091         if (kvm->as[i].as == as && kvm->as[i].ml) {
3092             size = MIN(kvm_max_slot_size, size);
3093             return NULL != kvm_lookup_matching_slot(kvm->as[i].ml,
3094                                                     start_addr, size);
3095         }
3096     }
3097 
3098     return false;
3099 }
3100 
3101 static void kvm_get_kvm_shadow_mem(Object *obj, Visitor *v,
3102                                    const char *name, void *opaque,
3103                                    Error **errp)
3104 {
3105     KVMState *s = KVM_STATE(obj);
3106     int64_t value = s->kvm_shadow_mem;
3107 
3108     visit_type_int(v, name, &value, errp);
3109 }
3110 
3111 static void kvm_set_kvm_shadow_mem(Object *obj, Visitor *v,
3112                                    const char *name, void *opaque,
3113                                    Error **errp)
3114 {
3115     KVMState *s = KVM_STATE(obj);
3116     int64_t value;
3117 
3118     if (!visit_type_int(v, name, &value, errp)) {
3119         return;
3120     }
3121 
3122     s->kvm_shadow_mem = value;
3123 }
3124 
3125 static void kvm_set_kernel_irqchip(Object *obj, Visitor *v,
3126                                    const char *name, void *opaque,
3127                                    Error **errp)
3128 {
3129     KVMState *s = KVM_STATE(obj);
3130     OnOffSplit mode;
3131 
3132     if (!visit_type_OnOffSplit(v, name, &mode, errp)) {
3133         return;
3134     }
3135     switch (mode) {
3136     case ON_OFF_SPLIT_ON:
3137         s->kernel_irqchip_allowed = true;
3138         s->kernel_irqchip_required = true;
3139         s->kernel_irqchip_split = ON_OFF_AUTO_OFF;
3140         break;
3141     case ON_OFF_SPLIT_OFF:
3142         s->kernel_irqchip_allowed = false;
3143         s->kernel_irqchip_required = false;
3144         s->kernel_irqchip_split = ON_OFF_AUTO_OFF;
3145         break;
3146     case ON_OFF_SPLIT_SPLIT:
3147         s->kernel_irqchip_allowed = true;
3148         s->kernel_irqchip_required = true;
3149         s->kernel_irqchip_split = ON_OFF_AUTO_ON;
3150         break;
3151     default:
3152         /* The value was checked in visit_type_OnOffSplit() above. If
3153          * we get here, then something is wrong in QEMU.
3154          */
3155         abort();
3156     }
3157 }
3158 
3159 bool kvm_kernel_irqchip_allowed(void)
3160 {
3161     return kvm_state->kernel_irqchip_allowed;
3162 }
3163 
3164 bool kvm_kernel_irqchip_required(void)
3165 {
3166     return kvm_state->kernel_irqchip_required;
3167 }
3168 
3169 bool kvm_kernel_irqchip_split(void)
3170 {
3171     return kvm_state->kernel_irqchip_split == ON_OFF_AUTO_ON;
3172 }
3173 
3174 static void kvm_accel_instance_init(Object *obj)
3175 {
3176     KVMState *s = KVM_STATE(obj);
3177 
3178     s->kvm_shadow_mem = -1;
3179     s->kernel_irqchip_allowed = true;
3180     s->kernel_irqchip_split = ON_OFF_AUTO_AUTO;
3181 }
3182 
3183 static void kvm_accel_class_init(ObjectClass *oc, void *data)
3184 {
3185     AccelClass *ac = ACCEL_CLASS(oc);
3186     ac->name = "KVM";
3187     ac->init_machine = kvm_init;
3188     ac->has_memory = kvm_accel_has_memory;
3189     ac->allowed = &kvm_allowed;
3190 
3191     object_class_property_add(oc, "kernel-irqchip", "on|off|split",
3192         NULL, kvm_set_kernel_irqchip,
3193         NULL, NULL);
3194     object_class_property_set_description(oc, "kernel-irqchip",
3195         "Configure KVM in-kernel irqchip");
3196 
3197     object_class_property_add(oc, "kvm-shadow-mem", "int",
3198         kvm_get_kvm_shadow_mem, kvm_set_kvm_shadow_mem,
3199         NULL, NULL);
3200     object_class_property_set_description(oc, "kvm-shadow-mem",
3201         "KVM shadow MMU size");
3202 }
3203 
3204 static const TypeInfo kvm_accel_type = {
3205     .name = TYPE_KVM_ACCEL,
3206     .parent = TYPE_ACCEL,
3207     .instance_init = kvm_accel_instance_init,
3208     .class_init = kvm_accel_class_init,
3209     .instance_size = sizeof(KVMState),
3210 };
3211 
3212 static void kvm_type_init(void)
3213 {
3214     type_register_static(&kvm_accel_type);
3215 }
3216 
3217 type_init(kvm_type_init);
3218