xref: /openbmc/qemu/accel/kvm/kvm-all.c (revision 761a9c58)
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 #include <poll.h>
19 
20 #include <linux/kvm.h>
21 
22 #include "qemu/atomic.h"
23 #include "qemu/option.h"
24 #include "qemu/config-file.h"
25 #include "qemu/error-report.h"
26 #include "qapi/error.h"
27 #include "hw/pci/msi.h"
28 #include "hw/pci/msix.h"
29 #include "hw/s390x/adapter.h"
30 #include "gdbstub/enums.h"
31 #include "sysemu/kvm_int.h"
32 #include "sysemu/runstate.h"
33 #include "sysemu/cpus.h"
34 #include "sysemu/accel-blocker.h"
35 #include "qemu/bswap.h"
36 #include "exec/memory.h"
37 #include "exec/ram_addr.h"
38 #include "qemu/event_notifier.h"
39 #include "qemu/main-loop.h"
40 #include "trace.h"
41 #include "hw/irq.h"
42 #include "qapi/visitor.h"
43 #include "qapi/qapi-types-common.h"
44 #include "qapi/qapi-visit-common.h"
45 #include "sysemu/reset.h"
46 #include "qemu/guest-random.h"
47 #include "sysemu/hw_accel.h"
48 #include "kvm-cpus.h"
49 #include "sysemu/dirtylimit.h"
50 #include "qemu/range.h"
51 
52 #include "hw/boards.h"
53 #include "sysemu/stats.h"
54 
55 /* This check must be after config-host.h is included */
56 #ifdef CONFIG_EVENTFD
57 #include <sys/eventfd.h>
58 #endif
59 
60 /* KVM uses PAGE_SIZE in its definition of KVM_COALESCED_MMIO_MAX. We
61  * need to use the real host PAGE_SIZE, as that's what KVM will use.
62  */
63 #ifdef PAGE_SIZE
64 #undef PAGE_SIZE
65 #endif
66 #define PAGE_SIZE qemu_real_host_page_size()
67 
68 #ifndef KVM_GUESTDBG_BLOCKIRQ
69 #define KVM_GUESTDBG_BLOCKIRQ 0
70 #endif
71 
72 struct KVMParkedVcpu {
73     unsigned long vcpu_id;
74     int kvm_fd;
75     QLIST_ENTRY(KVMParkedVcpu) node;
76 };
77 
78 KVMState *kvm_state;
79 bool kvm_kernel_irqchip;
80 bool kvm_split_irqchip;
81 bool kvm_async_interrupts_allowed;
82 bool kvm_halt_in_kernel_allowed;
83 bool kvm_resamplefds_allowed;
84 bool kvm_msi_via_irqfd_allowed;
85 bool kvm_gsi_routing_allowed;
86 bool kvm_gsi_direct_mapping;
87 bool kvm_allowed;
88 bool kvm_readonly_mem_allowed;
89 bool kvm_vm_attributes_allowed;
90 bool kvm_msi_use_devid;
91 static bool kvm_has_guest_debug;
92 static int kvm_sstep_flags;
93 static bool kvm_immediate_exit;
94 static uint64_t kvm_supported_memory_attributes;
95 static bool kvm_guest_memfd_supported;
96 static hwaddr kvm_max_slot_size = ~0;
97 
98 static const KVMCapabilityInfo kvm_required_capabilites[] = {
99     KVM_CAP_INFO(USER_MEMORY),
100     KVM_CAP_INFO(DESTROY_MEMORY_REGION_WORKS),
101     KVM_CAP_INFO(JOIN_MEMORY_REGIONS_WORKS),
102     KVM_CAP_INFO(INTERNAL_ERROR_DATA),
103     KVM_CAP_INFO(IOEVENTFD),
104     KVM_CAP_INFO(IOEVENTFD_ANY_LENGTH),
105     KVM_CAP_LAST_INFO
106 };
107 
108 static NotifierList kvm_irqchip_change_notifiers =
109     NOTIFIER_LIST_INITIALIZER(kvm_irqchip_change_notifiers);
110 
111 struct KVMResampleFd {
112     int gsi;
113     EventNotifier *resample_event;
114     QLIST_ENTRY(KVMResampleFd) node;
115 };
116 typedef struct KVMResampleFd KVMResampleFd;
117 
118 /*
119  * Only used with split irqchip where we need to do the resample fd
120  * kick for the kernel from userspace.
121  */
122 static QLIST_HEAD(, KVMResampleFd) kvm_resample_fd_list =
123     QLIST_HEAD_INITIALIZER(kvm_resample_fd_list);
124 
125 static QemuMutex kml_slots_lock;
126 
127 #define kvm_slots_lock()    qemu_mutex_lock(&kml_slots_lock)
128 #define kvm_slots_unlock()  qemu_mutex_unlock(&kml_slots_lock)
129 
130 static void kvm_slot_init_dirty_bitmap(KVMSlot *mem);
131 
132 static inline void kvm_resample_fd_remove(int gsi)
133 {
134     KVMResampleFd *rfd;
135 
136     QLIST_FOREACH(rfd, &kvm_resample_fd_list, node) {
137         if (rfd->gsi == gsi) {
138             QLIST_REMOVE(rfd, node);
139             g_free(rfd);
140             break;
141         }
142     }
143 }
144 
145 static inline void kvm_resample_fd_insert(int gsi, EventNotifier *event)
146 {
147     KVMResampleFd *rfd = g_new0(KVMResampleFd, 1);
148 
149     rfd->gsi = gsi;
150     rfd->resample_event = event;
151 
152     QLIST_INSERT_HEAD(&kvm_resample_fd_list, rfd, node);
153 }
154 
155 void kvm_resample_fd_notify(int gsi)
156 {
157     KVMResampleFd *rfd;
158 
159     QLIST_FOREACH(rfd, &kvm_resample_fd_list, node) {
160         if (rfd->gsi == gsi) {
161             event_notifier_set(rfd->resample_event);
162             trace_kvm_resample_fd_notify(gsi);
163             return;
164         }
165     }
166 }
167 
168 unsigned int kvm_get_max_memslots(void)
169 {
170     KVMState *s = KVM_STATE(current_accel());
171 
172     return s->nr_slots;
173 }
174 
175 unsigned int kvm_get_free_memslots(void)
176 {
177     unsigned int used_slots = 0;
178     KVMState *s = kvm_state;
179     int i;
180 
181     kvm_slots_lock();
182     for (i = 0; i < s->nr_as; i++) {
183         if (!s->as[i].ml) {
184             continue;
185         }
186         used_slots = MAX(used_slots, s->as[i].ml->nr_used_slots);
187     }
188     kvm_slots_unlock();
189 
190     return s->nr_slots - used_slots;
191 }
192 
193 /* Called with KVMMemoryListener.slots_lock held */
194 static KVMSlot *kvm_get_free_slot(KVMMemoryListener *kml)
195 {
196     KVMState *s = kvm_state;
197     int i;
198 
199     for (i = 0; i < s->nr_slots; i++) {
200         if (kml->slots[i].memory_size == 0) {
201             return &kml->slots[i];
202         }
203     }
204 
205     return NULL;
206 }
207 
208 /* Called with KVMMemoryListener.slots_lock held */
209 static KVMSlot *kvm_alloc_slot(KVMMemoryListener *kml)
210 {
211     KVMSlot *slot = kvm_get_free_slot(kml);
212 
213     if (slot) {
214         return slot;
215     }
216 
217     fprintf(stderr, "%s: no free slot available\n", __func__);
218     abort();
219 }
220 
221 static KVMSlot *kvm_lookup_matching_slot(KVMMemoryListener *kml,
222                                          hwaddr start_addr,
223                                          hwaddr size)
224 {
225     KVMState *s = kvm_state;
226     int i;
227 
228     for (i = 0; i < s->nr_slots; i++) {
229         KVMSlot *mem = &kml->slots[i];
230 
231         if (start_addr == mem->start_addr && size == mem->memory_size) {
232             return mem;
233         }
234     }
235 
236     return NULL;
237 }
238 
239 /*
240  * Calculate and align the start address and the size of the section.
241  * Return the size. If the size is 0, the aligned section is empty.
242  */
243 static hwaddr kvm_align_section(MemoryRegionSection *section,
244                                 hwaddr *start)
245 {
246     hwaddr size = int128_get64(section->size);
247     hwaddr delta, aligned;
248 
249     /* kvm works in page size chunks, but the function may be called
250        with sub-page size and unaligned start address. Pad the start
251        address to next and truncate size to previous page boundary. */
252     aligned = ROUND_UP(section->offset_within_address_space,
253                        qemu_real_host_page_size());
254     delta = aligned - section->offset_within_address_space;
255     *start = aligned;
256     if (delta > size) {
257         return 0;
258     }
259 
260     return (size - delta) & qemu_real_host_page_mask();
261 }
262 
263 int kvm_physical_memory_addr_from_host(KVMState *s, void *ram,
264                                        hwaddr *phys_addr)
265 {
266     KVMMemoryListener *kml = &s->memory_listener;
267     int i, ret = 0;
268 
269     kvm_slots_lock();
270     for (i = 0; i < s->nr_slots; i++) {
271         KVMSlot *mem = &kml->slots[i];
272 
273         if (ram >= mem->ram && ram < mem->ram + mem->memory_size) {
274             *phys_addr = mem->start_addr + (ram - mem->ram);
275             ret = 1;
276             break;
277         }
278     }
279     kvm_slots_unlock();
280 
281     return ret;
282 }
283 
284 static int kvm_set_user_memory_region(KVMMemoryListener *kml, KVMSlot *slot, bool new)
285 {
286     KVMState *s = kvm_state;
287     struct kvm_userspace_memory_region2 mem;
288     int ret;
289 
290     mem.slot = slot->slot | (kml->as_id << 16);
291     mem.guest_phys_addr = slot->start_addr;
292     mem.userspace_addr = (unsigned long)slot->ram;
293     mem.flags = slot->flags;
294     mem.guest_memfd = slot->guest_memfd;
295     mem.guest_memfd_offset = slot->guest_memfd_offset;
296 
297     if (slot->memory_size && !new && (mem.flags ^ slot->old_flags) & KVM_MEM_READONLY) {
298         /* Set the slot size to 0 before setting the slot to the desired
299          * value. This is needed based on KVM commit 75d61fbc. */
300         mem.memory_size = 0;
301 
302         if (kvm_guest_memfd_supported) {
303             ret = kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION2, &mem);
304         } else {
305             ret = kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
306         }
307         if (ret < 0) {
308             goto err;
309         }
310     }
311     mem.memory_size = slot->memory_size;
312     if (kvm_guest_memfd_supported) {
313         ret = kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION2, &mem);
314     } else {
315         ret = kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
316     }
317     slot->old_flags = mem.flags;
318 err:
319     trace_kvm_set_user_memory(mem.slot >> 16, (uint16_t)mem.slot, mem.flags,
320                               mem.guest_phys_addr, mem.memory_size,
321                               mem.userspace_addr, mem.guest_memfd,
322                               mem.guest_memfd_offset, ret);
323     if (ret < 0) {
324         if (kvm_guest_memfd_supported) {
325                 error_report("%s: KVM_SET_USER_MEMORY_REGION2 failed, slot=%d,"
326                         " start=0x%" PRIx64 ", size=0x%" PRIx64 ","
327                         " flags=0x%" PRIx32 ", guest_memfd=%" PRId32 ","
328                         " guest_memfd_offset=0x%" PRIx64 ": %s",
329                         __func__, mem.slot, slot->start_addr,
330                         (uint64_t)mem.memory_size, mem.flags,
331                         mem.guest_memfd, (uint64_t)mem.guest_memfd_offset,
332                         strerror(errno));
333         } else {
334                 error_report("%s: KVM_SET_USER_MEMORY_REGION failed, slot=%d,"
335                             " start=0x%" PRIx64 ", size=0x%" PRIx64 ": %s",
336                             __func__, mem.slot, slot->start_addr,
337                             (uint64_t)mem.memory_size, strerror(errno));
338         }
339     }
340     return ret;
341 }
342 
343 void kvm_park_vcpu(CPUState *cpu)
344 {
345     struct KVMParkedVcpu *vcpu;
346 
347     trace_kvm_park_vcpu(cpu->cpu_index, kvm_arch_vcpu_id(cpu));
348 
349     vcpu = g_malloc0(sizeof(*vcpu));
350     vcpu->vcpu_id = kvm_arch_vcpu_id(cpu);
351     vcpu->kvm_fd = cpu->kvm_fd;
352     QLIST_INSERT_HEAD(&kvm_state->kvm_parked_vcpus, vcpu, node);
353 }
354 
355 int kvm_unpark_vcpu(KVMState *s, unsigned long vcpu_id)
356 {
357     struct KVMParkedVcpu *cpu;
358     int kvm_fd = -ENOENT;
359 
360     QLIST_FOREACH(cpu, &s->kvm_parked_vcpus, node) {
361         if (cpu->vcpu_id == vcpu_id) {
362             QLIST_REMOVE(cpu, node);
363             kvm_fd = cpu->kvm_fd;
364             g_free(cpu);
365             break;
366         }
367     }
368 
369     trace_kvm_unpark_vcpu(vcpu_id, kvm_fd > 0 ? "unparked" : "!found parked");
370 
371     return kvm_fd;
372 }
373 
374 int kvm_create_vcpu(CPUState *cpu)
375 {
376     unsigned long vcpu_id = kvm_arch_vcpu_id(cpu);
377     KVMState *s = kvm_state;
378     int kvm_fd;
379 
380     /* check if the KVM vCPU already exist but is parked */
381     kvm_fd = kvm_unpark_vcpu(s, vcpu_id);
382     if (kvm_fd < 0) {
383         /* vCPU not parked: create a new KVM vCPU */
384         kvm_fd = kvm_vm_ioctl(s, KVM_CREATE_VCPU, vcpu_id);
385         if (kvm_fd < 0) {
386             error_report("KVM_CREATE_VCPU IOCTL failed for vCPU %lu", vcpu_id);
387             return kvm_fd;
388         }
389     }
390 
391     cpu->kvm_fd = kvm_fd;
392     cpu->kvm_state = s;
393     cpu->vcpu_dirty = true;
394     cpu->dirty_pages = 0;
395     cpu->throttle_us_per_full = 0;
396 
397     trace_kvm_create_vcpu(cpu->cpu_index, vcpu_id, kvm_fd);
398 
399     return 0;
400 }
401 
402 int kvm_create_and_park_vcpu(CPUState *cpu)
403 {
404     int ret = 0;
405 
406     ret = kvm_create_vcpu(cpu);
407     if (!ret) {
408         kvm_park_vcpu(cpu);
409     }
410 
411     return ret;
412 }
413 
414 static int do_kvm_destroy_vcpu(CPUState *cpu)
415 {
416     KVMState *s = kvm_state;
417     int mmap_size;
418     int ret = 0;
419 
420     trace_kvm_destroy_vcpu(cpu->cpu_index, kvm_arch_vcpu_id(cpu));
421 
422     ret = kvm_arch_destroy_vcpu(cpu);
423     if (ret < 0) {
424         goto err;
425     }
426 
427     mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
428     if (mmap_size < 0) {
429         ret = mmap_size;
430         trace_kvm_failed_get_vcpu_mmap_size();
431         goto err;
432     }
433 
434     ret = munmap(cpu->kvm_run, mmap_size);
435     if (ret < 0) {
436         goto err;
437     }
438 
439     if (cpu->kvm_dirty_gfns) {
440         ret = munmap(cpu->kvm_dirty_gfns, s->kvm_dirty_ring_bytes);
441         if (ret < 0) {
442             goto err;
443         }
444     }
445 
446     kvm_park_vcpu(cpu);
447 err:
448     return ret;
449 }
450 
451 void kvm_destroy_vcpu(CPUState *cpu)
452 {
453     if (do_kvm_destroy_vcpu(cpu) < 0) {
454         error_report("kvm_destroy_vcpu failed");
455         exit(EXIT_FAILURE);
456     }
457 }
458 
459 int kvm_init_vcpu(CPUState *cpu, Error **errp)
460 {
461     KVMState *s = kvm_state;
462     int mmap_size;
463     int ret;
464 
465     trace_kvm_init_vcpu(cpu->cpu_index, kvm_arch_vcpu_id(cpu));
466 
467     ret = kvm_create_vcpu(cpu);
468     if (ret < 0) {
469         error_setg_errno(errp, -ret,
470                          "kvm_init_vcpu: kvm_create_vcpu failed (%lu)",
471                          kvm_arch_vcpu_id(cpu));
472         goto err;
473     }
474 
475     mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
476     if (mmap_size < 0) {
477         ret = mmap_size;
478         error_setg_errno(errp, -mmap_size,
479                          "kvm_init_vcpu: KVM_GET_VCPU_MMAP_SIZE failed");
480         goto err;
481     }
482 
483     cpu->kvm_run = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
484                         cpu->kvm_fd, 0);
485     if (cpu->kvm_run == MAP_FAILED) {
486         ret = -errno;
487         error_setg_errno(errp, ret,
488                          "kvm_init_vcpu: mmap'ing vcpu state failed (%lu)",
489                          kvm_arch_vcpu_id(cpu));
490         goto err;
491     }
492 
493     if (s->coalesced_mmio && !s->coalesced_mmio_ring) {
494         s->coalesced_mmio_ring =
495             (void *)cpu->kvm_run + s->coalesced_mmio * PAGE_SIZE;
496     }
497 
498     if (s->kvm_dirty_ring_size) {
499         /* Use MAP_SHARED to share pages with the kernel */
500         cpu->kvm_dirty_gfns = mmap(NULL, s->kvm_dirty_ring_bytes,
501                                    PROT_READ | PROT_WRITE, MAP_SHARED,
502                                    cpu->kvm_fd,
503                                    PAGE_SIZE * KVM_DIRTY_LOG_PAGE_OFFSET);
504         if (cpu->kvm_dirty_gfns == MAP_FAILED) {
505             ret = -errno;
506             goto err;
507         }
508     }
509 
510     ret = kvm_arch_init_vcpu(cpu);
511     if (ret < 0) {
512         error_setg_errno(errp, -ret,
513                          "kvm_init_vcpu: kvm_arch_init_vcpu failed (%lu)",
514                          kvm_arch_vcpu_id(cpu));
515     }
516     cpu->kvm_vcpu_stats_fd = kvm_vcpu_ioctl(cpu, KVM_GET_STATS_FD, NULL);
517 
518 err:
519     return ret;
520 }
521 
522 /*
523  * dirty pages logging control
524  */
525 
526 static int kvm_mem_flags(MemoryRegion *mr)
527 {
528     bool readonly = mr->readonly || memory_region_is_romd(mr);
529     int flags = 0;
530 
531     if (memory_region_get_dirty_log_mask(mr) != 0) {
532         flags |= KVM_MEM_LOG_DIRTY_PAGES;
533     }
534     if (readonly && kvm_readonly_mem_allowed) {
535         flags |= KVM_MEM_READONLY;
536     }
537     if (memory_region_has_guest_memfd(mr)) {
538         assert(kvm_guest_memfd_supported);
539         flags |= KVM_MEM_GUEST_MEMFD;
540     }
541     return flags;
542 }
543 
544 /* Called with KVMMemoryListener.slots_lock held */
545 static int kvm_slot_update_flags(KVMMemoryListener *kml, KVMSlot *mem,
546                                  MemoryRegion *mr)
547 {
548     mem->flags = kvm_mem_flags(mr);
549 
550     /* If nothing changed effectively, no need to issue ioctl */
551     if (mem->flags == mem->old_flags) {
552         return 0;
553     }
554 
555     kvm_slot_init_dirty_bitmap(mem);
556     return kvm_set_user_memory_region(kml, mem, false);
557 }
558 
559 static int kvm_section_update_flags(KVMMemoryListener *kml,
560                                     MemoryRegionSection *section)
561 {
562     hwaddr start_addr, size, slot_size;
563     KVMSlot *mem;
564     int ret = 0;
565 
566     size = kvm_align_section(section, &start_addr);
567     if (!size) {
568         return 0;
569     }
570 
571     kvm_slots_lock();
572 
573     while (size && !ret) {
574         slot_size = MIN(kvm_max_slot_size, size);
575         mem = kvm_lookup_matching_slot(kml, start_addr, slot_size);
576         if (!mem) {
577             /* We don't have a slot if we want to trap every access. */
578             goto out;
579         }
580 
581         ret = kvm_slot_update_flags(kml, mem, section->mr);
582         start_addr += slot_size;
583         size -= slot_size;
584     }
585 
586 out:
587     kvm_slots_unlock();
588     return ret;
589 }
590 
591 static void kvm_log_start(MemoryListener *listener,
592                           MemoryRegionSection *section,
593                           int old, int new)
594 {
595     KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
596     int r;
597 
598     if (old != 0) {
599         return;
600     }
601 
602     r = kvm_section_update_flags(kml, section);
603     if (r < 0) {
604         abort();
605     }
606 }
607 
608 static void kvm_log_stop(MemoryListener *listener,
609                           MemoryRegionSection *section,
610                           int old, int new)
611 {
612     KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
613     int r;
614 
615     if (new != 0) {
616         return;
617     }
618 
619     r = kvm_section_update_flags(kml, section);
620     if (r < 0) {
621         abort();
622     }
623 }
624 
625 /* get kvm's dirty pages bitmap and update qemu's */
626 static void kvm_slot_sync_dirty_pages(KVMSlot *slot)
627 {
628     ram_addr_t start = slot->ram_start_offset;
629     ram_addr_t pages = slot->memory_size / qemu_real_host_page_size();
630 
631     cpu_physical_memory_set_dirty_lebitmap(slot->dirty_bmap, start, pages);
632 }
633 
634 static void kvm_slot_reset_dirty_pages(KVMSlot *slot)
635 {
636     memset(slot->dirty_bmap, 0, slot->dirty_bmap_size);
637 }
638 
639 #define ALIGN(x, y)  (((x)+(y)-1) & ~((y)-1))
640 
641 /* Allocate the dirty bitmap for a slot  */
642 static void kvm_slot_init_dirty_bitmap(KVMSlot *mem)
643 {
644     if (!(mem->flags & KVM_MEM_LOG_DIRTY_PAGES) || mem->dirty_bmap) {
645         return;
646     }
647 
648     /*
649      * XXX bad kernel interface alert
650      * For dirty bitmap, kernel allocates array of size aligned to
651      * bits-per-long.  But for case when the kernel is 64bits and
652      * the userspace is 32bits, userspace can't align to the same
653      * bits-per-long, since sizeof(long) is different between kernel
654      * and user space.  This way, userspace will provide buffer which
655      * may be 4 bytes less than the kernel will use, resulting in
656      * userspace memory corruption (which is not detectable by valgrind
657      * too, in most cases).
658      * So for now, let's align to 64 instead of HOST_LONG_BITS here, in
659      * a hope that sizeof(long) won't become >8 any time soon.
660      *
661      * Note: the granule of kvm dirty log is qemu_real_host_page_size.
662      * And mem->memory_size is aligned to it (otherwise this mem can't
663      * be registered to KVM).
664      */
665     hwaddr bitmap_size = ALIGN(mem->memory_size / qemu_real_host_page_size(),
666                                         /*HOST_LONG_BITS*/ 64) / 8;
667     mem->dirty_bmap = g_malloc0(bitmap_size);
668     mem->dirty_bmap_size = bitmap_size;
669 }
670 
671 /*
672  * Sync dirty bitmap from kernel to KVMSlot.dirty_bmap, return true if
673  * succeeded, false otherwise
674  */
675 static bool kvm_slot_get_dirty_log(KVMState *s, KVMSlot *slot)
676 {
677     struct kvm_dirty_log d = {};
678     int ret;
679 
680     d.dirty_bitmap = slot->dirty_bmap;
681     d.slot = slot->slot | (slot->as_id << 16);
682     ret = kvm_vm_ioctl(s, KVM_GET_DIRTY_LOG, &d);
683 
684     if (ret == -ENOENT) {
685         /* kernel does not have dirty bitmap in this slot */
686         ret = 0;
687     }
688     if (ret) {
689         error_report_once("%s: KVM_GET_DIRTY_LOG failed with %d",
690                           __func__, ret);
691     }
692     return ret == 0;
693 }
694 
695 /* Should be with all slots_lock held for the address spaces. */
696 static void kvm_dirty_ring_mark_page(KVMState *s, uint32_t as_id,
697                                      uint32_t slot_id, uint64_t offset)
698 {
699     KVMMemoryListener *kml;
700     KVMSlot *mem;
701 
702     if (as_id >= s->nr_as) {
703         return;
704     }
705 
706     kml = s->as[as_id].ml;
707     mem = &kml->slots[slot_id];
708 
709     if (!mem->memory_size || offset >=
710         (mem->memory_size / qemu_real_host_page_size())) {
711         return;
712     }
713 
714     set_bit(offset, mem->dirty_bmap);
715 }
716 
717 static bool dirty_gfn_is_dirtied(struct kvm_dirty_gfn *gfn)
718 {
719     /*
720      * Read the flags before the value.  Pairs with barrier in
721      * KVM's kvm_dirty_ring_push() function.
722      */
723     return qatomic_load_acquire(&gfn->flags) == KVM_DIRTY_GFN_F_DIRTY;
724 }
725 
726 static void dirty_gfn_set_collected(struct kvm_dirty_gfn *gfn)
727 {
728     /*
729      * Use a store-release so that the CPU that executes KVM_RESET_DIRTY_RINGS
730      * sees the full content of the ring:
731      *
732      * CPU0                     CPU1                         CPU2
733      * ------------------------------------------------------------------------------
734      *                                                       fill gfn0
735      *                                                       store-rel flags for gfn0
736      * load-acq flags for gfn0
737      * store-rel RESET for gfn0
738      *                          ioctl(RESET_RINGS)
739      *                            load-acq flags for gfn0
740      *                            check if flags have RESET
741      *
742      * The synchronization goes from CPU2 to CPU0 to CPU1.
743      */
744     qatomic_store_release(&gfn->flags, KVM_DIRTY_GFN_F_RESET);
745 }
746 
747 /*
748  * Should be with all slots_lock held for the address spaces.  It returns the
749  * dirty page we've collected on this dirty ring.
750  */
751 static uint32_t kvm_dirty_ring_reap_one(KVMState *s, CPUState *cpu)
752 {
753     struct kvm_dirty_gfn *dirty_gfns = cpu->kvm_dirty_gfns, *cur;
754     uint32_t ring_size = s->kvm_dirty_ring_size;
755     uint32_t count = 0, fetch = cpu->kvm_fetch_index;
756 
757     /*
758      * It's possible that we race with vcpu creation code where the vcpu is
759      * put onto the vcpus list but not yet initialized the dirty ring
760      * structures.  If so, skip it.
761      */
762     if (!cpu->created) {
763         return 0;
764     }
765 
766     assert(dirty_gfns && ring_size);
767     trace_kvm_dirty_ring_reap_vcpu(cpu->cpu_index);
768 
769     while (true) {
770         cur = &dirty_gfns[fetch % ring_size];
771         if (!dirty_gfn_is_dirtied(cur)) {
772             break;
773         }
774         kvm_dirty_ring_mark_page(s, cur->slot >> 16, cur->slot & 0xffff,
775                                  cur->offset);
776         dirty_gfn_set_collected(cur);
777         trace_kvm_dirty_ring_page(cpu->cpu_index, fetch, cur->offset);
778         fetch++;
779         count++;
780     }
781     cpu->kvm_fetch_index = fetch;
782     cpu->dirty_pages += count;
783 
784     return count;
785 }
786 
787 /* Must be with slots_lock held */
788 static uint64_t kvm_dirty_ring_reap_locked(KVMState *s, CPUState* cpu)
789 {
790     int ret;
791     uint64_t total = 0;
792     int64_t stamp;
793 
794     stamp = get_clock();
795 
796     if (cpu) {
797         total = kvm_dirty_ring_reap_one(s, cpu);
798     } else {
799         CPU_FOREACH(cpu) {
800             total += kvm_dirty_ring_reap_one(s, cpu);
801         }
802     }
803 
804     if (total) {
805         ret = kvm_vm_ioctl(s, KVM_RESET_DIRTY_RINGS);
806         assert(ret == total);
807     }
808 
809     stamp = get_clock() - stamp;
810 
811     if (total) {
812         trace_kvm_dirty_ring_reap(total, stamp / 1000);
813     }
814 
815     return total;
816 }
817 
818 /*
819  * Currently for simplicity, we must hold BQL before calling this.  We can
820  * consider to drop the BQL if we're clear with all the race conditions.
821  */
822 static uint64_t kvm_dirty_ring_reap(KVMState *s, CPUState *cpu)
823 {
824     uint64_t total;
825 
826     /*
827      * We need to lock all kvm slots for all address spaces here,
828      * because:
829      *
830      * (1) We need to mark dirty for dirty bitmaps in multiple slots
831      *     and for tons of pages, so it's better to take the lock here
832      *     once rather than once per page.  And more importantly,
833      *
834      * (2) We must _NOT_ publish dirty bits to the other threads
835      *     (e.g., the migration thread) via the kvm memory slot dirty
836      *     bitmaps before correctly re-protect those dirtied pages.
837      *     Otherwise we can have potential risk of data corruption if
838      *     the page data is read in the other thread before we do
839      *     reset below.
840      */
841     kvm_slots_lock();
842     total = kvm_dirty_ring_reap_locked(s, cpu);
843     kvm_slots_unlock();
844 
845     return total;
846 }
847 
848 static void do_kvm_cpu_synchronize_kick(CPUState *cpu, run_on_cpu_data arg)
849 {
850     /* No need to do anything */
851 }
852 
853 /*
854  * Kick all vcpus out in a synchronized way.  When returned, we
855  * guarantee that every vcpu has been kicked and at least returned to
856  * userspace once.
857  */
858 static void kvm_cpu_synchronize_kick_all(void)
859 {
860     CPUState *cpu;
861 
862     CPU_FOREACH(cpu) {
863         run_on_cpu(cpu, do_kvm_cpu_synchronize_kick, RUN_ON_CPU_NULL);
864     }
865 }
866 
867 /*
868  * Flush all the existing dirty pages to the KVM slot buffers.  When
869  * this call returns, we guarantee that all the touched dirty pages
870  * before calling this function have been put into the per-kvmslot
871  * dirty bitmap.
872  *
873  * This function must be called with BQL held.
874  */
875 static void kvm_dirty_ring_flush(void)
876 {
877     trace_kvm_dirty_ring_flush(0);
878     /*
879      * The function needs to be serialized.  Since this function
880      * should always be with BQL held, serialization is guaranteed.
881      * However, let's be sure of it.
882      */
883     assert(bql_locked());
884     /*
885      * First make sure to flush the hardware buffers by kicking all
886      * vcpus out in a synchronous way.
887      */
888     kvm_cpu_synchronize_kick_all();
889     kvm_dirty_ring_reap(kvm_state, NULL);
890     trace_kvm_dirty_ring_flush(1);
891 }
892 
893 /**
894  * kvm_physical_sync_dirty_bitmap - Sync dirty bitmap from kernel space
895  *
896  * This function will first try to fetch dirty bitmap from the kernel,
897  * and then updates qemu's dirty bitmap.
898  *
899  * NOTE: caller must be with kml->slots_lock held.
900  *
901  * @kml: the KVM memory listener object
902  * @section: the memory section to sync the dirty bitmap with
903  */
904 static void kvm_physical_sync_dirty_bitmap(KVMMemoryListener *kml,
905                                            MemoryRegionSection *section)
906 {
907     KVMState *s = kvm_state;
908     KVMSlot *mem;
909     hwaddr start_addr, size;
910     hwaddr slot_size;
911 
912     size = kvm_align_section(section, &start_addr);
913     while (size) {
914         slot_size = MIN(kvm_max_slot_size, size);
915         mem = kvm_lookup_matching_slot(kml, start_addr, slot_size);
916         if (!mem) {
917             /* We don't have a slot if we want to trap every access. */
918             return;
919         }
920         if (kvm_slot_get_dirty_log(s, mem)) {
921             kvm_slot_sync_dirty_pages(mem);
922         }
923         start_addr += slot_size;
924         size -= slot_size;
925     }
926 }
927 
928 /* Alignment requirement for KVM_CLEAR_DIRTY_LOG - 64 pages */
929 #define KVM_CLEAR_LOG_SHIFT  6
930 #define KVM_CLEAR_LOG_ALIGN  (qemu_real_host_page_size() << KVM_CLEAR_LOG_SHIFT)
931 #define KVM_CLEAR_LOG_MASK   (-KVM_CLEAR_LOG_ALIGN)
932 
933 static int kvm_log_clear_one_slot(KVMSlot *mem, int as_id, uint64_t start,
934                                   uint64_t size)
935 {
936     KVMState *s = kvm_state;
937     uint64_t end, bmap_start, start_delta, bmap_npages;
938     struct kvm_clear_dirty_log d;
939     unsigned long *bmap_clear = NULL, psize = qemu_real_host_page_size();
940     int ret;
941 
942     /*
943      * We need to extend either the start or the size or both to
944      * satisfy the KVM interface requirement.  Firstly, do the start
945      * page alignment on 64 host pages
946      */
947     bmap_start = start & KVM_CLEAR_LOG_MASK;
948     start_delta = start - bmap_start;
949     bmap_start /= psize;
950 
951     /*
952      * The kernel interface has restriction on the size too, that either:
953      *
954      * (1) the size is 64 host pages aligned (just like the start), or
955      * (2) the size fills up until the end of the KVM memslot.
956      */
957     bmap_npages = DIV_ROUND_UP(size + start_delta, KVM_CLEAR_LOG_ALIGN)
958         << KVM_CLEAR_LOG_SHIFT;
959     end = mem->memory_size / psize;
960     if (bmap_npages > end - bmap_start) {
961         bmap_npages = end - bmap_start;
962     }
963     start_delta /= psize;
964 
965     /*
966      * Prepare the bitmap to clear dirty bits.  Here we must guarantee
967      * that we won't clear any unknown dirty bits otherwise we might
968      * accidentally clear some set bits which are not yet synced from
969      * the kernel into QEMU's bitmap, then we'll lose track of the
970      * guest modifications upon those pages (which can directly lead
971      * to guest data loss or panic after migration).
972      *
973      * Layout of the KVMSlot.dirty_bmap:
974      *
975      *                   |<-------- bmap_npages -----------..>|
976      *                                                     [1]
977      *                     start_delta         size
978      *  |----------------|-------------|------------------|------------|
979      *  ^                ^             ^                               ^
980      *  |                |             |                               |
981      * start          bmap_start     (start)                         end
982      * of memslot                                             of memslot
983      *
984      * [1] bmap_npages can be aligned to either 64 pages or the end of slot
985      */
986 
987     assert(bmap_start % BITS_PER_LONG == 0);
988     /* We should never do log_clear before log_sync */
989     assert(mem->dirty_bmap);
990     if (start_delta || bmap_npages - size / psize) {
991         /* Slow path - we need to manipulate a temp bitmap */
992         bmap_clear = bitmap_new(bmap_npages);
993         bitmap_copy_with_src_offset(bmap_clear, mem->dirty_bmap,
994                                     bmap_start, start_delta + size / psize);
995         /*
996          * We need to fill the holes at start because that was not
997          * specified by the caller and we extended the bitmap only for
998          * 64 pages alignment
999          */
1000         bitmap_clear(bmap_clear, 0, start_delta);
1001         d.dirty_bitmap = bmap_clear;
1002     } else {
1003         /*
1004          * Fast path - both start and size align well with BITS_PER_LONG
1005          * (or the end of memory slot)
1006          */
1007         d.dirty_bitmap = mem->dirty_bmap + BIT_WORD(bmap_start);
1008     }
1009 
1010     d.first_page = bmap_start;
1011     /* It should never overflow.  If it happens, say something */
1012     assert(bmap_npages <= UINT32_MAX);
1013     d.num_pages = bmap_npages;
1014     d.slot = mem->slot | (as_id << 16);
1015 
1016     ret = kvm_vm_ioctl(s, KVM_CLEAR_DIRTY_LOG, &d);
1017     if (ret < 0 && ret != -ENOENT) {
1018         error_report("%s: KVM_CLEAR_DIRTY_LOG failed, slot=%d, "
1019                      "start=0x%"PRIx64", size=0x%"PRIx32", errno=%d",
1020                      __func__, d.slot, (uint64_t)d.first_page,
1021                      (uint32_t)d.num_pages, ret);
1022     } else {
1023         ret = 0;
1024         trace_kvm_clear_dirty_log(d.slot, d.first_page, d.num_pages);
1025     }
1026 
1027     /*
1028      * After we have updated the remote dirty bitmap, we update the
1029      * cached bitmap as well for the memslot, then if another user
1030      * clears the same region we know we shouldn't clear it again on
1031      * the remote otherwise it's data loss as well.
1032      */
1033     bitmap_clear(mem->dirty_bmap, bmap_start + start_delta,
1034                  size / psize);
1035     /* This handles the NULL case well */
1036     g_free(bmap_clear);
1037     return ret;
1038 }
1039 
1040 
1041 /**
1042  * kvm_physical_log_clear - Clear the kernel's dirty bitmap for range
1043  *
1044  * NOTE: this will be a no-op if we haven't enabled manual dirty log
1045  * protection in the host kernel because in that case this operation
1046  * will be done within log_sync().
1047  *
1048  * @kml:     the kvm memory listener
1049  * @section: the memory range to clear dirty bitmap
1050  */
1051 static int kvm_physical_log_clear(KVMMemoryListener *kml,
1052                                   MemoryRegionSection *section)
1053 {
1054     KVMState *s = kvm_state;
1055     uint64_t start, size, offset, count;
1056     KVMSlot *mem;
1057     int ret = 0, i;
1058 
1059     if (!s->manual_dirty_log_protect) {
1060         /* No need to do explicit clear */
1061         return ret;
1062     }
1063 
1064     start = section->offset_within_address_space;
1065     size = int128_get64(section->size);
1066 
1067     if (!size) {
1068         /* Nothing more we can do... */
1069         return ret;
1070     }
1071 
1072     kvm_slots_lock();
1073 
1074     for (i = 0; i < s->nr_slots; i++) {
1075         mem = &kml->slots[i];
1076         /* Discard slots that are empty or do not overlap the section */
1077         if (!mem->memory_size ||
1078             mem->start_addr > start + size - 1 ||
1079             start > mem->start_addr + mem->memory_size - 1) {
1080             continue;
1081         }
1082 
1083         if (start >= mem->start_addr) {
1084             /* The slot starts before section or is aligned to it.  */
1085             offset = start - mem->start_addr;
1086             count = MIN(mem->memory_size - offset, size);
1087         } else {
1088             /* The slot starts after section.  */
1089             offset = 0;
1090             count = MIN(mem->memory_size, size - (mem->start_addr - start));
1091         }
1092         ret = kvm_log_clear_one_slot(mem, kml->as_id, offset, count);
1093         if (ret < 0) {
1094             break;
1095         }
1096     }
1097 
1098     kvm_slots_unlock();
1099 
1100     return ret;
1101 }
1102 
1103 static void kvm_coalesce_mmio_region(MemoryListener *listener,
1104                                      MemoryRegionSection *secion,
1105                                      hwaddr start, hwaddr size)
1106 {
1107     KVMState *s = kvm_state;
1108 
1109     if (s->coalesced_mmio) {
1110         struct kvm_coalesced_mmio_zone zone;
1111 
1112         zone.addr = start;
1113         zone.size = size;
1114         zone.pad = 0;
1115 
1116         (void)kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone);
1117     }
1118 }
1119 
1120 static void kvm_uncoalesce_mmio_region(MemoryListener *listener,
1121                                        MemoryRegionSection *secion,
1122                                        hwaddr start, hwaddr size)
1123 {
1124     KVMState *s = kvm_state;
1125 
1126     if (s->coalesced_mmio) {
1127         struct kvm_coalesced_mmio_zone zone;
1128 
1129         zone.addr = start;
1130         zone.size = size;
1131         zone.pad = 0;
1132 
1133         (void)kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone);
1134     }
1135 }
1136 
1137 static void kvm_coalesce_pio_add(MemoryListener *listener,
1138                                 MemoryRegionSection *section,
1139                                 hwaddr start, hwaddr size)
1140 {
1141     KVMState *s = kvm_state;
1142 
1143     if (s->coalesced_pio) {
1144         struct kvm_coalesced_mmio_zone zone;
1145 
1146         zone.addr = start;
1147         zone.size = size;
1148         zone.pio = 1;
1149 
1150         (void)kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone);
1151     }
1152 }
1153 
1154 static void kvm_coalesce_pio_del(MemoryListener *listener,
1155                                 MemoryRegionSection *section,
1156                                 hwaddr start, hwaddr size)
1157 {
1158     KVMState *s = kvm_state;
1159 
1160     if (s->coalesced_pio) {
1161         struct kvm_coalesced_mmio_zone zone;
1162 
1163         zone.addr = start;
1164         zone.size = size;
1165         zone.pio = 1;
1166 
1167         (void)kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone);
1168      }
1169 }
1170 
1171 int kvm_check_extension(KVMState *s, unsigned int extension)
1172 {
1173     int ret;
1174 
1175     ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, extension);
1176     if (ret < 0) {
1177         ret = 0;
1178     }
1179 
1180     return ret;
1181 }
1182 
1183 int kvm_vm_check_extension(KVMState *s, unsigned int extension)
1184 {
1185     int ret;
1186 
1187     ret = kvm_vm_ioctl(s, KVM_CHECK_EXTENSION, extension);
1188     if (ret < 0) {
1189         /* VM wide version not implemented, use global one instead */
1190         ret = kvm_check_extension(s, extension);
1191     }
1192 
1193     return ret;
1194 }
1195 
1196 /*
1197  * We track the poisoned pages to be able to:
1198  * - replace them on VM reset
1199  * - block a migration for a VM with a poisoned page
1200  */
1201 typedef struct HWPoisonPage {
1202     ram_addr_t ram_addr;
1203     QLIST_ENTRY(HWPoisonPage) list;
1204 } HWPoisonPage;
1205 
1206 static QLIST_HEAD(, HWPoisonPage) hwpoison_page_list =
1207     QLIST_HEAD_INITIALIZER(hwpoison_page_list);
1208 
1209 static void kvm_unpoison_all(void *param)
1210 {
1211     HWPoisonPage *page, *next_page;
1212 
1213     QLIST_FOREACH_SAFE(page, &hwpoison_page_list, list, next_page) {
1214         QLIST_REMOVE(page, list);
1215         qemu_ram_remap(page->ram_addr, TARGET_PAGE_SIZE);
1216         g_free(page);
1217     }
1218 }
1219 
1220 void kvm_hwpoison_page_add(ram_addr_t ram_addr)
1221 {
1222     HWPoisonPage *page;
1223 
1224     QLIST_FOREACH(page, &hwpoison_page_list, list) {
1225         if (page->ram_addr == ram_addr) {
1226             return;
1227         }
1228     }
1229     page = g_new(HWPoisonPage, 1);
1230     page->ram_addr = ram_addr;
1231     QLIST_INSERT_HEAD(&hwpoison_page_list, page, list);
1232 }
1233 
1234 bool kvm_hwpoisoned_mem(void)
1235 {
1236     return !QLIST_EMPTY(&hwpoison_page_list);
1237 }
1238 
1239 static uint32_t adjust_ioeventfd_endianness(uint32_t val, uint32_t size)
1240 {
1241 #if HOST_BIG_ENDIAN != TARGET_BIG_ENDIAN
1242     /* The kernel expects ioeventfd values in HOST_BIG_ENDIAN
1243      * endianness, but the memory core hands them in target endianness.
1244      * For example, PPC is always treated as big-endian even if running
1245      * on KVM and on PPC64LE.  Correct here.
1246      */
1247     switch (size) {
1248     case 2:
1249         val = bswap16(val);
1250         break;
1251     case 4:
1252         val = bswap32(val);
1253         break;
1254     }
1255 #endif
1256     return val;
1257 }
1258 
1259 static int kvm_set_ioeventfd_mmio(int fd, hwaddr addr, uint32_t val,
1260                                   bool assign, uint32_t size, bool datamatch)
1261 {
1262     int ret;
1263     struct kvm_ioeventfd iofd = {
1264         .datamatch = datamatch ? adjust_ioeventfd_endianness(val, size) : 0,
1265         .addr = addr,
1266         .len = size,
1267         .flags = 0,
1268         .fd = fd,
1269     };
1270 
1271     trace_kvm_set_ioeventfd_mmio(fd, (uint64_t)addr, val, assign, size,
1272                                  datamatch);
1273     if (!kvm_enabled()) {
1274         return -ENOSYS;
1275     }
1276 
1277     if (datamatch) {
1278         iofd.flags |= KVM_IOEVENTFD_FLAG_DATAMATCH;
1279     }
1280     if (!assign) {
1281         iofd.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
1282     }
1283 
1284     ret = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &iofd);
1285 
1286     if (ret < 0) {
1287         return -errno;
1288     }
1289 
1290     return 0;
1291 }
1292 
1293 static int kvm_set_ioeventfd_pio(int fd, uint16_t addr, uint16_t val,
1294                                  bool assign, uint32_t size, bool datamatch)
1295 {
1296     struct kvm_ioeventfd kick = {
1297         .datamatch = datamatch ? adjust_ioeventfd_endianness(val, size) : 0,
1298         .addr = addr,
1299         .flags = KVM_IOEVENTFD_FLAG_PIO,
1300         .len = size,
1301         .fd = fd,
1302     };
1303     int r;
1304     trace_kvm_set_ioeventfd_pio(fd, addr, val, assign, size, datamatch);
1305     if (!kvm_enabled()) {
1306         return -ENOSYS;
1307     }
1308     if (datamatch) {
1309         kick.flags |= KVM_IOEVENTFD_FLAG_DATAMATCH;
1310     }
1311     if (!assign) {
1312         kick.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
1313     }
1314     r = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &kick);
1315     if (r < 0) {
1316         return r;
1317     }
1318     return 0;
1319 }
1320 
1321 
1322 static const KVMCapabilityInfo *
1323 kvm_check_extension_list(KVMState *s, const KVMCapabilityInfo *list)
1324 {
1325     while (list->name) {
1326         if (!kvm_check_extension(s, list->value)) {
1327             return list;
1328         }
1329         list++;
1330     }
1331     return NULL;
1332 }
1333 
1334 void kvm_set_max_memslot_size(hwaddr max_slot_size)
1335 {
1336     g_assert(
1337         ROUND_UP(max_slot_size, qemu_real_host_page_size()) == max_slot_size
1338     );
1339     kvm_max_slot_size = max_slot_size;
1340 }
1341 
1342 static int kvm_set_memory_attributes(hwaddr start, uint64_t size, uint64_t attr)
1343 {
1344     struct kvm_memory_attributes attrs;
1345     int r;
1346 
1347     assert((attr & kvm_supported_memory_attributes) == attr);
1348     attrs.attributes = attr;
1349     attrs.address = start;
1350     attrs.size = size;
1351     attrs.flags = 0;
1352 
1353     r = kvm_vm_ioctl(kvm_state, KVM_SET_MEMORY_ATTRIBUTES, &attrs);
1354     if (r) {
1355         error_report("failed to set memory (0x%" HWADDR_PRIx "+0x%" PRIx64 ") "
1356                      "with attr 0x%" PRIx64 " error '%s'",
1357                      start, size, attr, strerror(errno));
1358     }
1359     return r;
1360 }
1361 
1362 int kvm_set_memory_attributes_private(hwaddr start, uint64_t size)
1363 {
1364     return kvm_set_memory_attributes(start, size, KVM_MEMORY_ATTRIBUTE_PRIVATE);
1365 }
1366 
1367 int kvm_set_memory_attributes_shared(hwaddr start, uint64_t size)
1368 {
1369     return kvm_set_memory_attributes(start, size, 0);
1370 }
1371 
1372 /* Called with KVMMemoryListener.slots_lock held */
1373 static void kvm_set_phys_mem(KVMMemoryListener *kml,
1374                              MemoryRegionSection *section, bool add)
1375 {
1376     KVMSlot *mem;
1377     int err;
1378     MemoryRegion *mr = section->mr;
1379     bool writable = !mr->readonly && !mr->rom_device;
1380     hwaddr start_addr, size, slot_size, mr_offset;
1381     ram_addr_t ram_start_offset;
1382     void *ram;
1383 
1384     if (!memory_region_is_ram(mr)) {
1385         if (writable || !kvm_readonly_mem_allowed) {
1386             return;
1387         } else if (!mr->romd_mode) {
1388             /* If the memory device is not in romd_mode, then we actually want
1389              * to remove the kvm memory slot so all accesses will trap. */
1390             add = false;
1391         }
1392     }
1393 
1394     size = kvm_align_section(section, &start_addr);
1395     if (!size) {
1396         return;
1397     }
1398 
1399     /* The offset of the kvmslot within the memory region */
1400     mr_offset = section->offset_within_region + start_addr -
1401         section->offset_within_address_space;
1402 
1403     /* use aligned delta to align the ram address and offset */
1404     ram = memory_region_get_ram_ptr(mr) + mr_offset;
1405     ram_start_offset = memory_region_get_ram_addr(mr) + mr_offset;
1406 
1407     if (!add) {
1408         do {
1409             slot_size = MIN(kvm_max_slot_size, size);
1410             mem = kvm_lookup_matching_slot(kml, start_addr, slot_size);
1411             if (!mem) {
1412                 return;
1413             }
1414             if (mem->flags & KVM_MEM_LOG_DIRTY_PAGES) {
1415                 /*
1416                  * NOTE: We should be aware of the fact that here we're only
1417                  * doing a best effort to sync dirty bits.  No matter whether
1418                  * we're using dirty log or dirty ring, we ignored two facts:
1419                  *
1420                  * (1) dirty bits can reside in hardware buffers (PML)
1421                  *
1422                  * (2) after we collected dirty bits here, pages can be dirtied
1423                  * again before we do the final KVM_SET_USER_MEMORY_REGION to
1424                  * remove the slot.
1425                  *
1426                  * Not easy.  Let's cross the fingers until it's fixed.
1427                  */
1428                 if (kvm_state->kvm_dirty_ring_size) {
1429                     kvm_dirty_ring_reap_locked(kvm_state, NULL);
1430                     if (kvm_state->kvm_dirty_ring_with_bitmap) {
1431                         kvm_slot_sync_dirty_pages(mem);
1432                         kvm_slot_get_dirty_log(kvm_state, mem);
1433                     }
1434                 } else {
1435                     kvm_slot_get_dirty_log(kvm_state, mem);
1436                 }
1437                 kvm_slot_sync_dirty_pages(mem);
1438             }
1439 
1440             /* unregister the slot */
1441             g_free(mem->dirty_bmap);
1442             mem->dirty_bmap = NULL;
1443             mem->memory_size = 0;
1444             mem->flags = 0;
1445             err = kvm_set_user_memory_region(kml, mem, false);
1446             if (err) {
1447                 fprintf(stderr, "%s: error unregistering slot: %s\n",
1448                         __func__, strerror(-err));
1449                 abort();
1450             }
1451             start_addr += slot_size;
1452             size -= slot_size;
1453             kml->nr_used_slots--;
1454         } while (size);
1455         return;
1456     }
1457 
1458     /* register the new slot */
1459     do {
1460         slot_size = MIN(kvm_max_slot_size, size);
1461         mem = kvm_alloc_slot(kml);
1462         mem->as_id = kml->as_id;
1463         mem->memory_size = slot_size;
1464         mem->start_addr = start_addr;
1465         mem->ram_start_offset = ram_start_offset;
1466         mem->ram = ram;
1467         mem->flags = kvm_mem_flags(mr);
1468         mem->guest_memfd = mr->ram_block->guest_memfd;
1469         mem->guest_memfd_offset = (uint8_t*)ram - mr->ram_block->host;
1470 
1471         kvm_slot_init_dirty_bitmap(mem);
1472         err = kvm_set_user_memory_region(kml, mem, true);
1473         if (err) {
1474             fprintf(stderr, "%s: error registering slot: %s\n", __func__,
1475                     strerror(-err));
1476             abort();
1477         }
1478 
1479         if (memory_region_has_guest_memfd(mr)) {
1480             err = kvm_set_memory_attributes_private(start_addr, slot_size);
1481             if (err) {
1482                 error_report("%s: failed to set memory attribute private: %s",
1483                              __func__, strerror(-err));
1484                 exit(1);
1485             }
1486         }
1487 
1488         start_addr += slot_size;
1489         ram_start_offset += slot_size;
1490         ram += slot_size;
1491         size -= slot_size;
1492         kml->nr_used_slots++;
1493     } while (size);
1494 }
1495 
1496 static void *kvm_dirty_ring_reaper_thread(void *data)
1497 {
1498     KVMState *s = data;
1499     struct KVMDirtyRingReaper *r = &s->reaper;
1500 
1501     rcu_register_thread();
1502 
1503     trace_kvm_dirty_ring_reaper("init");
1504 
1505     while (true) {
1506         r->reaper_state = KVM_DIRTY_RING_REAPER_WAIT;
1507         trace_kvm_dirty_ring_reaper("wait");
1508         /*
1509          * TODO: provide a smarter timeout rather than a constant?
1510          */
1511         sleep(1);
1512 
1513         /* keep sleeping so that dirtylimit not be interfered by reaper */
1514         if (dirtylimit_in_service()) {
1515             continue;
1516         }
1517 
1518         trace_kvm_dirty_ring_reaper("wakeup");
1519         r->reaper_state = KVM_DIRTY_RING_REAPER_REAPING;
1520 
1521         bql_lock();
1522         kvm_dirty_ring_reap(s, NULL);
1523         bql_unlock();
1524 
1525         r->reaper_iteration++;
1526     }
1527 
1528     g_assert_not_reached();
1529 }
1530 
1531 static void kvm_dirty_ring_reaper_init(KVMState *s)
1532 {
1533     struct KVMDirtyRingReaper *r = &s->reaper;
1534 
1535     qemu_thread_create(&r->reaper_thr, "kvm-reaper",
1536                        kvm_dirty_ring_reaper_thread,
1537                        s, QEMU_THREAD_JOINABLE);
1538 }
1539 
1540 static int kvm_dirty_ring_init(KVMState *s)
1541 {
1542     uint32_t ring_size = s->kvm_dirty_ring_size;
1543     uint64_t ring_bytes = ring_size * sizeof(struct kvm_dirty_gfn);
1544     unsigned int capability = KVM_CAP_DIRTY_LOG_RING;
1545     int ret;
1546 
1547     s->kvm_dirty_ring_size = 0;
1548     s->kvm_dirty_ring_bytes = 0;
1549 
1550     /* Bail if the dirty ring size isn't specified */
1551     if (!ring_size) {
1552         return 0;
1553     }
1554 
1555     /*
1556      * Read the max supported pages. Fall back to dirty logging mode
1557      * if the dirty ring isn't supported.
1558      */
1559     ret = kvm_vm_check_extension(s, capability);
1560     if (ret <= 0) {
1561         capability = KVM_CAP_DIRTY_LOG_RING_ACQ_REL;
1562         ret = kvm_vm_check_extension(s, capability);
1563     }
1564 
1565     if (ret <= 0) {
1566         warn_report("KVM dirty ring not available, using bitmap method");
1567         return 0;
1568     }
1569 
1570     if (ring_bytes > ret) {
1571         error_report("KVM dirty ring size %" PRIu32 " too big "
1572                      "(maximum is %ld).  Please use a smaller value.",
1573                      ring_size, (long)ret / sizeof(struct kvm_dirty_gfn));
1574         return -EINVAL;
1575     }
1576 
1577     ret = kvm_vm_enable_cap(s, capability, 0, ring_bytes);
1578     if (ret) {
1579         error_report("Enabling of KVM dirty ring failed: %s. "
1580                      "Suggested minimum value is 1024.", strerror(-ret));
1581         return -EIO;
1582     }
1583 
1584     /* Enable the backup bitmap if it is supported */
1585     ret = kvm_vm_check_extension(s, KVM_CAP_DIRTY_LOG_RING_WITH_BITMAP);
1586     if (ret > 0) {
1587         ret = kvm_vm_enable_cap(s, KVM_CAP_DIRTY_LOG_RING_WITH_BITMAP, 0);
1588         if (ret) {
1589             error_report("Enabling of KVM dirty ring's backup bitmap failed: "
1590                          "%s. ", strerror(-ret));
1591             return -EIO;
1592         }
1593 
1594         s->kvm_dirty_ring_with_bitmap = true;
1595     }
1596 
1597     s->kvm_dirty_ring_size = ring_size;
1598     s->kvm_dirty_ring_bytes = ring_bytes;
1599 
1600     return 0;
1601 }
1602 
1603 static void kvm_region_add(MemoryListener *listener,
1604                            MemoryRegionSection *section)
1605 {
1606     KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
1607     KVMMemoryUpdate *update;
1608 
1609     update = g_new0(KVMMemoryUpdate, 1);
1610     update->section = *section;
1611 
1612     QSIMPLEQ_INSERT_TAIL(&kml->transaction_add, update, next);
1613 }
1614 
1615 static void kvm_region_del(MemoryListener *listener,
1616                            MemoryRegionSection *section)
1617 {
1618     KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
1619     KVMMemoryUpdate *update;
1620 
1621     update = g_new0(KVMMemoryUpdate, 1);
1622     update->section = *section;
1623 
1624     QSIMPLEQ_INSERT_TAIL(&kml->transaction_del, update, next);
1625 }
1626 
1627 static void kvm_region_commit(MemoryListener *listener)
1628 {
1629     KVMMemoryListener *kml = container_of(listener, KVMMemoryListener,
1630                                           listener);
1631     KVMMemoryUpdate *u1, *u2;
1632     bool need_inhibit = false;
1633 
1634     if (QSIMPLEQ_EMPTY(&kml->transaction_add) &&
1635         QSIMPLEQ_EMPTY(&kml->transaction_del)) {
1636         return;
1637     }
1638 
1639     /*
1640      * We have to be careful when regions to add overlap with ranges to remove.
1641      * We have to simulate atomic KVM memslot updates by making sure no ioctl()
1642      * is currently active.
1643      *
1644      * The lists are order by addresses, so it's easy to find overlaps.
1645      */
1646     u1 = QSIMPLEQ_FIRST(&kml->transaction_del);
1647     u2 = QSIMPLEQ_FIRST(&kml->transaction_add);
1648     while (u1 && u2) {
1649         Range r1, r2;
1650 
1651         range_init_nofail(&r1, u1->section.offset_within_address_space,
1652                           int128_get64(u1->section.size));
1653         range_init_nofail(&r2, u2->section.offset_within_address_space,
1654                           int128_get64(u2->section.size));
1655 
1656         if (range_overlaps_range(&r1, &r2)) {
1657             need_inhibit = true;
1658             break;
1659         }
1660         if (range_lob(&r1) < range_lob(&r2)) {
1661             u1 = QSIMPLEQ_NEXT(u1, next);
1662         } else {
1663             u2 = QSIMPLEQ_NEXT(u2, next);
1664         }
1665     }
1666 
1667     kvm_slots_lock();
1668     if (need_inhibit) {
1669         accel_ioctl_inhibit_begin();
1670     }
1671 
1672     /* Remove all memslots before adding the new ones. */
1673     while (!QSIMPLEQ_EMPTY(&kml->transaction_del)) {
1674         u1 = QSIMPLEQ_FIRST(&kml->transaction_del);
1675         QSIMPLEQ_REMOVE_HEAD(&kml->transaction_del, next);
1676 
1677         kvm_set_phys_mem(kml, &u1->section, false);
1678         memory_region_unref(u1->section.mr);
1679 
1680         g_free(u1);
1681     }
1682     while (!QSIMPLEQ_EMPTY(&kml->transaction_add)) {
1683         u1 = QSIMPLEQ_FIRST(&kml->transaction_add);
1684         QSIMPLEQ_REMOVE_HEAD(&kml->transaction_add, next);
1685 
1686         memory_region_ref(u1->section.mr);
1687         kvm_set_phys_mem(kml, &u1->section, true);
1688 
1689         g_free(u1);
1690     }
1691 
1692     if (need_inhibit) {
1693         accel_ioctl_inhibit_end();
1694     }
1695     kvm_slots_unlock();
1696 }
1697 
1698 static void kvm_log_sync(MemoryListener *listener,
1699                          MemoryRegionSection *section)
1700 {
1701     KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
1702 
1703     kvm_slots_lock();
1704     kvm_physical_sync_dirty_bitmap(kml, section);
1705     kvm_slots_unlock();
1706 }
1707 
1708 static void kvm_log_sync_global(MemoryListener *l, bool last_stage)
1709 {
1710     KVMMemoryListener *kml = container_of(l, KVMMemoryListener, listener);
1711     KVMState *s = kvm_state;
1712     KVMSlot *mem;
1713     int i;
1714 
1715     /* Flush all kernel dirty addresses into KVMSlot dirty bitmap */
1716     kvm_dirty_ring_flush();
1717 
1718     /*
1719      * TODO: make this faster when nr_slots is big while there are
1720      * only a few used slots (small VMs).
1721      */
1722     kvm_slots_lock();
1723     for (i = 0; i < s->nr_slots; i++) {
1724         mem = &kml->slots[i];
1725         if (mem->memory_size && mem->flags & KVM_MEM_LOG_DIRTY_PAGES) {
1726             kvm_slot_sync_dirty_pages(mem);
1727 
1728             if (s->kvm_dirty_ring_with_bitmap && last_stage &&
1729                 kvm_slot_get_dirty_log(s, mem)) {
1730                 kvm_slot_sync_dirty_pages(mem);
1731             }
1732 
1733             /*
1734              * This is not needed by KVM_GET_DIRTY_LOG because the
1735              * ioctl will unconditionally overwrite the whole region.
1736              * However kvm dirty ring has no such side effect.
1737              */
1738             kvm_slot_reset_dirty_pages(mem);
1739         }
1740     }
1741     kvm_slots_unlock();
1742 }
1743 
1744 static void kvm_log_clear(MemoryListener *listener,
1745                           MemoryRegionSection *section)
1746 {
1747     KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
1748     int r;
1749 
1750     r = kvm_physical_log_clear(kml, section);
1751     if (r < 0) {
1752         error_report_once("%s: kvm log clear failed: mr=%s "
1753                           "offset=%"HWADDR_PRIx" size=%"PRIx64, __func__,
1754                           section->mr->name, section->offset_within_region,
1755                           int128_get64(section->size));
1756         abort();
1757     }
1758 }
1759 
1760 static void kvm_mem_ioeventfd_add(MemoryListener *listener,
1761                                   MemoryRegionSection *section,
1762                                   bool match_data, uint64_t data,
1763                                   EventNotifier *e)
1764 {
1765     int fd = event_notifier_get_fd(e);
1766     int r;
1767 
1768     r = kvm_set_ioeventfd_mmio(fd, section->offset_within_address_space,
1769                                data, true, int128_get64(section->size),
1770                                match_data);
1771     if (r < 0) {
1772         fprintf(stderr, "%s: error adding ioeventfd: %s (%d)\n",
1773                 __func__, strerror(-r), -r);
1774         abort();
1775     }
1776 }
1777 
1778 static void kvm_mem_ioeventfd_del(MemoryListener *listener,
1779                                   MemoryRegionSection *section,
1780                                   bool match_data, uint64_t data,
1781                                   EventNotifier *e)
1782 {
1783     int fd = event_notifier_get_fd(e);
1784     int r;
1785 
1786     r = kvm_set_ioeventfd_mmio(fd, section->offset_within_address_space,
1787                                data, false, int128_get64(section->size),
1788                                match_data);
1789     if (r < 0) {
1790         fprintf(stderr, "%s: error deleting ioeventfd: %s (%d)\n",
1791                 __func__, strerror(-r), -r);
1792         abort();
1793     }
1794 }
1795 
1796 static void kvm_io_ioeventfd_add(MemoryListener *listener,
1797                                  MemoryRegionSection *section,
1798                                  bool match_data, uint64_t data,
1799                                  EventNotifier *e)
1800 {
1801     int fd = event_notifier_get_fd(e);
1802     int r;
1803 
1804     r = kvm_set_ioeventfd_pio(fd, section->offset_within_address_space,
1805                               data, true, int128_get64(section->size),
1806                               match_data);
1807     if (r < 0) {
1808         fprintf(stderr, "%s: error adding ioeventfd: %s (%d)\n",
1809                 __func__, strerror(-r), -r);
1810         abort();
1811     }
1812 }
1813 
1814 static void kvm_io_ioeventfd_del(MemoryListener *listener,
1815                                  MemoryRegionSection *section,
1816                                  bool match_data, uint64_t data,
1817                                  EventNotifier *e)
1818 
1819 {
1820     int fd = event_notifier_get_fd(e);
1821     int r;
1822 
1823     r = kvm_set_ioeventfd_pio(fd, section->offset_within_address_space,
1824                               data, false, int128_get64(section->size),
1825                               match_data);
1826     if (r < 0) {
1827         fprintf(stderr, "%s: error deleting ioeventfd: %s (%d)\n",
1828                 __func__, strerror(-r), -r);
1829         abort();
1830     }
1831 }
1832 
1833 void kvm_memory_listener_register(KVMState *s, KVMMemoryListener *kml,
1834                                   AddressSpace *as, int as_id, const char *name)
1835 {
1836     int i;
1837 
1838     kml->slots = g_new0(KVMSlot, s->nr_slots);
1839     kml->as_id = as_id;
1840 
1841     for (i = 0; i < s->nr_slots; i++) {
1842         kml->slots[i].slot = i;
1843     }
1844 
1845     QSIMPLEQ_INIT(&kml->transaction_add);
1846     QSIMPLEQ_INIT(&kml->transaction_del);
1847 
1848     kml->listener.region_add = kvm_region_add;
1849     kml->listener.region_del = kvm_region_del;
1850     kml->listener.commit = kvm_region_commit;
1851     kml->listener.log_start = kvm_log_start;
1852     kml->listener.log_stop = kvm_log_stop;
1853     kml->listener.priority = MEMORY_LISTENER_PRIORITY_ACCEL;
1854     kml->listener.name = name;
1855 
1856     if (s->kvm_dirty_ring_size) {
1857         kml->listener.log_sync_global = kvm_log_sync_global;
1858     } else {
1859         kml->listener.log_sync = kvm_log_sync;
1860         kml->listener.log_clear = kvm_log_clear;
1861     }
1862 
1863     memory_listener_register(&kml->listener, as);
1864 
1865     for (i = 0; i < s->nr_as; ++i) {
1866         if (!s->as[i].as) {
1867             s->as[i].as = as;
1868             s->as[i].ml = kml;
1869             break;
1870         }
1871     }
1872 }
1873 
1874 static MemoryListener kvm_io_listener = {
1875     .name = "kvm-io",
1876     .coalesced_io_add = kvm_coalesce_pio_add,
1877     .coalesced_io_del = kvm_coalesce_pio_del,
1878     .eventfd_add = kvm_io_ioeventfd_add,
1879     .eventfd_del = kvm_io_ioeventfd_del,
1880     .priority = MEMORY_LISTENER_PRIORITY_DEV_BACKEND,
1881 };
1882 
1883 int kvm_set_irq(KVMState *s, int irq, int level)
1884 {
1885     struct kvm_irq_level event;
1886     int ret;
1887 
1888     assert(kvm_async_interrupts_enabled());
1889 
1890     event.level = level;
1891     event.irq = irq;
1892     ret = kvm_vm_ioctl(s, s->irq_set_ioctl, &event);
1893     if (ret < 0) {
1894         perror("kvm_set_irq");
1895         abort();
1896     }
1897 
1898     return (s->irq_set_ioctl == KVM_IRQ_LINE) ? 1 : event.status;
1899 }
1900 
1901 #ifdef KVM_CAP_IRQ_ROUTING
1902 typedef struct KVMMSIRoute {
1903     struct kvm_irq_routing_entry kroute;
1904     QTAILQ_ENTRY(KVMMSIRoute) entry;
1905 } KVMMSIRoute;
1906 
1907 static void set_gsi(KVMState *s, unsigned int gsi)
1908 {
1909     set_bit(gsi, s->used_gsi_bitmap);
1910 }
1911 
1912 static void clear_gsi(KVMState *s, unsigned int gsi)
1913 {
1914     clear_bit(gsi, s->used_gsi_bitmap);
1915 }
1916 
1917 void kvm_init_irq_routing(KVMState *s)
1918 {
1919     int gsi_count;
1920 
1921     gsi_count = kvm_check_extension(s, KVM_CAP_IRQ_ROUTING) - 1;
1922     if (gsi_count > 0) {
1923         /* Round up so we can search ints using ffs */
1924         s->used_gsi_bitmap = bitmap_new(gsi_count);
1925         s->gsi_count = gsi_count;
1926     }
1927 
1928     s->irq_routes = g_malloc0(sizeof(*s->irq_routes));
1929     s->nr_allocated_irq_routes = 0;
1930 
1931     kvm_arch_init_irq_routing(s);
1932 }
1933 
1934 void kvm_irqchip_commit_routes(KVMState *s)
1935 {
1936     int ret;
1937 
1938     if (kvm_gsi_direct_mapping()) {
1939         return;
1940     }
1941 
1942     if (!kvm_gsi_routing_enabled()) {
1943         return;
1944     }
1945 
1946     s->irq_routes->flags = 0;
1947     trace_kvm_irqchip_commit_routes();
1948     ret = kvm_vm_ioctl(s, KVM_SET_GSI_ROUTING, s->irq_routes);
1949     assert(ret == 0);
1950 }
1951 
1952 void kvm_add_routing_entry(KVMState *s,
1953                            struct kvm_irq_routing_entry *entry)
1954 {
1955     struct kvm_irq_routing_entry *new;
1956     int n, size;
1957 
1958     if (s->irq_routes->nr == s->nr_allocated_irq_routes) {
1959         n = s->nr_allocated_irq_routes * 2;
1960         if (n < 64) {
1961             n = 64;
1962         }
1963         size = sizeof(struct kvm_irq_routing);
1964         size += n * sizeof(*new);
1965         s->irq_routes = g_realloc(s->irq_routes, size);
1966         s->nr_allocated_irq_routes = n;
1967     }
1968     n = s->irq_routes->nr++;
1969     new = &s->irq_routes->entries[n];
1970 
1971     *new = *entry;
1972 
1973     set_gsi(s, entry->gsi);
1974 }
1975 
1976 static int kvm_update_routing_entry(KVMState *s,
1977                                     struct kvm_irq_routing_entry *new_entry)
1978 {
1979     struct kvm_irq_routing_entry *entry;
1980     int n;
1981 
1982     for (n = 0; n < s->irq_routes->nr; n++) {
1983         entry = &s->irq_routes->entries[n];
1984         if (entry->gsi != new_entry->gsi) {
1985             continue;
1986         }
1987 
1988         if(!memcmp(entry, new_entry, sizeof *entry)) {
1989             return 0;
1990         }
1991 
1992         *entry = *new_entry;
1993 
1994         return 0;
1995     }
1996 
1997     return -ESRCH;
1998 }
1999 
2000 void kvm_irqchip_add_irq_route(KVMState *s, int irq, int irqchip, int pin)
2001 {
2002     struct kvm_irq_routing_entry e = {};
2003 
2004     assert(pin < s->gsi_count);
2005 
2006     e.gsi = irq;
2007     e.type = KVM_IRQ_ROUTING_IRQCHIP;
2008     e.flags = 0;
2009     e.u.irqchip.irqchip = irqchip;
2010     e.u.irqchip.pin = pin;
2011     kvm_add_routing_entry(s, &e);
2012 }
2013 
2014 void kvm_irqchip_release_virq(KVMState *s, int virq)
2015 {
2016     struct kvm_irq_routing_entry *e;
2017     int i;
2018 
2019     if (kvm_gsi_direct_mapping()) {
2020         return;
2021     }
2022 
2023     for (i = 0; i < s->irq_routes->nr; i++) {
2024         e = &s->irq_routes->entries[i];
2025         if (e->gsi == virq) {
2026             s->irq_routes->nr--;
2027             *e = s->irq_routes->entries[s->irq_routes->nr];
2028         }
2029     }
2030     clear_gsi(s, virq);
2031     kvm_arch_release_virq_post(virq);
2032     trace_kvm_irqchip_release_virq(virq);
2033 }
2034 
2035 void kvm_irqchip_add_change_notifier(Notifier *n)
2036 {
2037     notifier_list_add(&kvm_irqchip_change_notifiers, n);
2038 }
2039 
2040 void kvm_irqchip_remove_change_notifier(Notifier *n)
2041 {
2042     notifier_remove(n);
2043 }
2044 
2045 void kvm_irqchip_change_notify(void)
2046 {
2047     notifier_list_notify(&kvm_irqchip_change_notifiers, NULL);
2048 }
2049 
2050 int kvm_irqchip_get_virq(KVMState *s)
2051 {
2052     int next_virq;
2053 
2054     /* Return the lowest unused GSI in the bitmap */
2055     next_virq = find_first_zero_bit(s->used_gsi_bitmap, s->gsi_count);
2056     if (next_virq >= s->gsi_count) {
2057         return -ENOSPC;
2058     } else {
2059         return next_virq;
2060     }
2061 }
2062 
2063 int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg)
2064 {
2065     struct kvm_msi msi;
2066 
2067     msi.address_lo = (uint32_t)msg.address;
2068     msi.address_hi = msg.address >> 32;
2069     msi.data = le32_to_cpu(msg.data);
2070     msi.flags = 0;
2071     memset(msi.pad, 0, sizeof(msi.pad));
2072 
2073     return kvm_vm_ioctl(s, KVM_SIGNAL_MSI, &msi);
2074 }
2075 
2076 int kvm_irqchip_add_msi_route(KVMRouteChange *c, int vector, PCIDevice *dev)
2077 {
2078     struct kvm_irq_routing_entry kroute = {};
2079     int virq;
2080     KVMState *s = c->s;
2081     MSIMessage msg = {0, 0};
2082 
2083     if (pci_available && dev) {
2084         msg = pci_get_msi_message(dev, vector);
2085     }
2086 
2087     if (kvm_gsi_direct_mapping()) {
2088         return kvm_arch_msi_data_to_gsi(msg.data);
2089     }
2090 
2091     if (!kvm_gsi_routing_enabled()) {
2092         return -ENOSYS;
2093     }
2094 
2095     virq = kvm_irqchip_get_virq(s);
2096     if (virq < 0) {
2097         return virq;
2098     }
2099 
2100     kroute.gsi = virq;
2101     kroute.type = KVM_IRQ_ROUTING_MSI;
2102     kroute.flags = 0;
2103     kroute.u.msi.address_lo = (uint32_t)msg.address;
2104     kroute.u.msi.address_hi = msg.address >> 32;
2105     kroute.u.msi.data = le32_to_cpu(msg.data);
2106     if (pci_available && kvm_msi_devid_required()) {
2107         kroute.flags = KVM_MSI_VALID_DEVID;
2108         kroute.u.msi.devid = pci_requester_id(dev);
2109     }
2110     if (kvm_arch_fixup_msi_route(&kroute, msg.address, msg.data, dev)) {
2111         kvm_irqchip_release_virq(s, virq);
2112         return -EINVAL;
2113     }
2114 
2115     if (s->irq_routes->nr < s->gsi_count) {
2116         trace_kvm_irqchip_add_msi_route(dev ? dev->name : (char *)"N/A",
2117                                         vector, virq);
2118 
2119         kvm_add_routing_entry(s, &kroute);
2120         kvm_arch_add_msi_route_post(&kroute, vector, dev);
2121         c->changes++;
2122     } else {
2123         kvm_irqchip_release_virq(s, virq);
2124         return -ENOSPC;
2125     }
2126 
2127     return virq;
2128 }
2129 
2130 int kvm_irqchip_update_msi_route(KVMState *s, int virq, MSIMessage msg,
2131                                  PCIDevice *dev)
2132 {
2133     struct kvm_irq_routing_entry kroute = {};
2134 
2135     if (kvm_gsi_direct_mapping()) {
2136         return 0;
2137     }
2138 
2139     if (!kvm_irqchip_in_kernel()) {
2140         return -ENOSYS;
2141     }
2142 
2143     kroute.gsi = virq;
2144     kroute.type = KVM_IRQ_ROUTING_MSI;
2145     kroute.flags = 0;
2146     kroute.u.msi.address_lo = (uint32_t)msg.address;
2147     kroute.u.msi.address_hi = msg.address >> 32;
2148     kroute.u.msi.data = le32_to_cpu(msg.data);
2149     if (pci_available && kvm_msi_devid_required()) {
2150         kroute.flags = KVM_MSI_VALID_DEVID;
2151         kroute.u.msi.devid = pci_requester_id(dev);
2152     }
2153     if (kvm_arch_fixup_msi_route(&kroute, msg.address, msg.data, dev)) {
2154         return -EINVAL;
2155     }
2156 
2157     trace_kvm_irqchip_update_msi_route(virq);
2158 
2159     return kvm_update_routing_entry(s, &kroute);
2160 }
2161 
2162 static int kvm_irqchip_assign_irqfd(KVMState *s, EventNotifier *event,
2163                                     EventNotifier *resample, int virq,
2164                                     bool assign)
2165 {
2166     int fd = event_notifier_get_fd(event);
2167     int rfd = resample ? event_notifier_get_fd(resample) : -1;
2168 
2169     struct kvm_irqfd irqfd = {
2170         .fd = fd,
2171         .gsi = virq,
2172         .flags = assign ? 0 : KVM_IRQFD_FLAG_DEASSIGN,
2173     };
2174 
2175     if (rfd != -1) {
2176         assert(assign);
2177         if (kvm_irqchip_is_split()) {
2178             /*
2179              * When the slow irqchip (e.g. IOAPIC) is in the
2180              * userspace, KVM kernel resamplefd will not work because
2181              * the EOI of the interrupt will be delivered to userspace
2182              * instead, so the KVM kernel resamplefd kick will be
2183              * skipped.  The userspace here mimics what the kernel
2184              * provides with resamplefd, remember the resamplefd and
2185              * kick it when we receive EOI of this IRQ.
2186              *
2187              * This is hackery because IOAPIC is mostly bypassed
2188              * (except EOI broadcasts) when irqfd is used.  However
2189              * this can bring much performance back for split irqchip
2190              * with INTx IRQs (for VFIO, this gives 93% perf of the
2191              * full fast path, which is 46% perf boost comparing to
2192              * the INTx slow path).
2193              */
2194             kvm_resample_fd_insert(virq, resample);
2195         } else {
2196             irqfd.flags |= KVM_IRQFD_FLAG_RESAMPLE;
2197             irqfd.resamplefd = rfd;
2198         }
2199     } else if (!assign) {
2200         if (kvm_irqchip_is_split()) {
2201             kvm_resample_fd_remove(virq);
2202         }
2203     }
2204 
2205     return kvm_vm_ioctl(s, KVM_IRQFD, &irqfd);
2206 }
2207 
2208 #else /* !KVM_CAP_IRQ_ROUTING */
2209 
2210 void kvm_init_irq_routing(KVMState *s)
2211 {
2212 }
2213 
2214 void kvm_irqchip_release_virq(KVMState *s, int virq)
2215 {
2216 }
2217 
2218 int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg)
2219 {
2220     abort();
2221 }
2222 
2223 int kvm_irqchip_add_msi_route(KVMRouteChange *c, int vector, PCIDevice *dev)
2224 {
2225     return -ENOSYS;
2226 }
2227 
2228 int kvm_irqchip_add_adapter_route(KVMState *s, AdapterInfo *adapter)
2229 {
2230     return -ENOSYS;
2231 }
2232 
2233 int kvm_irqchip_add_hv_sint_route(KVMState *s, uint32_t vcpu, uint32_t sint)
2234 {
2235     return -ENOSYS;
2236 }
2237 
2238 static int kvm_irqchip_assign_irqfd(KVMState *s, EventNotifier *event,
2239                                     EventNotifier *resample, int virq,
2240                                     bool assign)
2241 {
2242     abort();
2243 }
2244 
2245 int kvm_irqchip_update_msi_route(KVMState *s, int virq, MSIMessage msg)
2246 {
2247     return -ENOSYS;
2248 }
2249 #endif /* !KVM_CAP_IRQ_ROUTING */
2250 
2251 int kvm_irqchip_add_irqfd_notifier_gsi(KVMState *s, EventNotifier *n,
2252                                        EventNotifier *rn, int virq)
2253 {
2254     return kvm_irqchip_assign_irqfd(s, n, rn, virq, true);
2255 }
2256 
2257 int kvm_irqchip_remove_irqfd_notifier_gsi(KVMState *s, EventNotifier *n,
2258                                           int virq)
2259 {
2260     return kvm_irqchip_assign_irqfd(s, n, NULL, virq, false);
2261 }
2262 
2263 int kvm_irqchip_add_irqfd_notifier(KVMState *s, EventNotifier *n,
2264                                    EventNotifier *rn, qemu_irq irq)
2265 {
2266     gpointer key, gsi;
2267     gboolean found = g_hash_table_lookup_extended(s->gsimap, irq, &key, &gsi);
2268 
2269     if (!found) {
2270         return -ENXIO;
2271     }
2272     return kvm_irqchip_add_irqfd_notifier_gsi(s, n, rn, GPOINTER_TO_INT(gsi));
2273 }
2274 
2275 int kvm_irqchip_remove_irqfd_notifier(KVMState *s, EventNotifier *n,
2276                                       qemu_irq irq)
2277 {
2278     gpointer key, gsi;
2279     gboolean found = g_hash_table_lookup_extended(s->gsimap, irq, &key, &gsi);
2280 
2281     if (!found) {
2282         return -ENXIO;
2283     }
2284     return kvm_irqchip_remove_irqfd_notifier_gsi(s, n, GPOINTER_TO_INT(gsi));
2285 }
2286 
2287 void kvm_irqchip_set_qemuirq_gsi(KVMState *s, qemu_irq irq, int gsi)
2288 {
2289     g_hash_table_insert(s->gsimap, irq, GINT_TO_POINTER(gsi));
2290 }
2291 
2292 static void kvm_irqchip_create(KVMState *s)
2293 {
2294     int ret;
2295 
2296     assert(s->kernel_irqchip_split != ON_OFF_AUTO_AUTO);
2297     if (kvm_check_extension(s, KVM_CAP_IRQCHIP)) {
2298         ;
2299     } else if (kvm_check_extension(s, KVM_CAP_S390_IRQCHIP)) {
2300         ret = kvm_vm_enable_cap(s, KVM_CAP_S390_IRQCHIP, 0);
2301         if (ret < 0) {
2302             fprintf(stderr, "Enable kernel irqchip failed: %s\n", strerror(-ret));
2303             exit(1);
2304         }
2305     } else {
2306         return;
2307     }
2308 
2309     if (kvm_check_extension(s, KVM_CAP_IRQFD) <= 0) {
2310         fprintf(stderr, "kvm: irqfd not implemented\n");
2311         exit(1);
2312     }
2313 
2314     /* First probe and see if there's a arch-specific hook to create the
2315      * in-kernel irqchip for us */
2316     ret = kvm_arch_irqchip_create(s);
2317     if (ret == 0) {
2318         if (s->kernel_irqchip_split == ON_OFF_AUTO_ON) {
2319             error_report("Split IRQ chip mode not supported.");
2320             exit(1);
2321         } else {
2322             ret = kvm_vm_ioctl(s, KVM_CREATE_IRQCHIP);
2323         }
2324     }
2325     if (ret < 0) {
2326         fprintf(stderr, "Create kernel irqchip failed: %s\n", strerror(-ret));
2327         exit(1);
2328     }
2329 
2330     kvm_kernel_irqchip = true;
2331     /* If we have an in-kernel IRQ chip then we must have asynchronous
2332      * interrupt delivery (though the reverse is not necessarily true)
2333      */
2334     kvm_async_interrupts_allowed = true;
2335     kvm_halt_in_kernel_allowed = true;
2336 
2337     kvm_init_irq_routing(s);
2338 
2339     s->gsimap = g_hash_table_new(g_direct_hash, g_direct_equal);
2340 }
2341 
2342 /* Find number of supported CPUs using the recommended
2343  * procedure from the kernel API documentation to cope with
2344  * older kernels that may be missing capabilities.
2345  */
2346 static int kvm_recommended_vcpus(KVMState *s)
2347 {
2348     int ret = kvm_vm_check_extension(s, KVM_CAP_NR_VCPUS);
2349     return (ret) ? ret : 4;
2350 }
2351 
2352 static int kvm_max_vcpus(KVMState *s)
2353 {
2354     int ret = kvm_check_extension(s, KVM_CAP_MAX_VCPUS);
2355     return (ret) ? ret : kvm_recommended_vcpus(s);
2356 }
2357 
2358 static int kvm_max_vcpu_id(KVMState *s)
2359 {
2360     int ret = kvm_check_extension(s, KVM_CAP_MAX_VCPU_ID);
2361     return (ret) ? ret : kvm_max_vcpus(s);
2362 }
2363 
2364 bool kvm_vcpu_id_is_valid(int vcpu_id)
2365 {
2366     KVMState *s = KVM_STATE(current_accel());
2367     return vcpu_id >= 0 && vcpu_id < kvm_max_vcpu_id(s);
2368 }
2369 
2370 bool kvm_dirty_ring_enabled(void)
2371 {
2372     return kvm_state && kvm_state->kvm_dirty_ring_size;
2373 }
2374 
2375 static void query_stats_cb(StatsResultList **result, StatsTarget target,
2376                            strList *names, strList *targets, Error **errp);
2377 static void query_stats_schemas_cb(StatsSchemaList **result, Error **errp);
2378 
2379 uint32_t kvm_dirty_ring_size(void)
2380 {
2381     return kvm_state->kvm_dirty_ring_size;
2382 }
2383 
2384 static int kvm_init(MachineState *ms)
2385 {
2386     MachineClass *mc = MACHINE_GET_CLASS(ms);
2387     static const char upgrade_note[] =
2388         "Please upgrade to at least kernel 2.6.29 or recent kvm-kmod\n"
2389         "(see http://sourceforge.net/projects/kvm).\n";
2390     const struct {
2391         const char *name;
2392         int num;
2393     } num_cpus[] = {
2394         { "SMP",          ms->smp.cpus },
2395         { "hotpluggable", ms->smp.max_cpus },
2396         { /* end of list */ }
2397     }, *nc = num_cpus;
2398     int soft_vcpus_limit, hard_vcpus_limit;
2399     KVMState *s;
2400     const KVMCapabilityInfo *missing_cap;
2401     int ret;
2402     int type;
2403     uint64_t dirty_log_manual_caps;
2404 
2405     qemu_mutex_init(&kml_slots_lock);
2406 
2407     s = KVM_STATE(ms->accelerator);
2408 
2409     /*
2410      * On systems where the kernel can support different base page
2411      * sizes, host page size may be different from TARGET_PAGE_SIZE,
2412      * even with KVM.  TARGET_PAGE_SIZE is assumed to be the minimum
2413      * page size for the system though.
2414      */
2415     assert(TARGET_PAGE_SIZE <= qemu_real_host_page_size());
2416 
2417     s->sigmask_len = 8;
2418     accel_blocker_init();
2419 
2420 #ifdef TARGET_KVM_HAVE_GUEST_DEBUG
2421     QTAILQ_INIT(&s->kvm_sw_breakpoints);
2422 #endif
2423     QLIST_INIT(&s->kvm_parked_vcpus);
2424     s->fd = qemu_open_old(s->device ?: "/dev/kvm", O_RDWR);
2425     if (s->fd == -1) {
2426         fprintf(stderr, "Could not access KVM kernel module: %m\n");
2427         ret = -errno;
2428         goto err;
2429     }
2430 
2431     ret = kvm_ioctl(s, KVM_GET_API_VERSION, 0);
2432     if (ret < KVM_API_VERSION) {
2433         if (ret >= 0) {
2434             ret = -EINVAL;
2435         }
2436         fprintf(stderr, "kvm version too old\n");
2437         goto err;
2438     }
2439 
2440     if (ret > KVM_API_VERSION) {
2441         ret = -EINVAL;
2442         fprintf(stderr, "kvm version not supported\n");
2443         goto err;
2444     }
2445 
2446     kvm_supported_memory_attributes = kvm_check_extension(s, KVM_CAP_MEMORY_ATTRIBUTES);
2447     kvm_guest_memfd_supported =
2448         kvm_check_extension(s, KVM_CAP_GUEST_MEMFD) &&
2449         kvm_check_extension(s, KVM_CAP_USER_MEMORY2) &&
2450         (kvm_supported_memory_attributes & KVM_MEMORY_ATTRIBUTE_PRIVATE);
2451 
2452     kvm_immediate_exit = kvm_check_extension(s, KVM_CAP_IMMEDIATE_EXIT);
2453     s->nr_slots = kvm_check_extension(s, KVM_CAP_NR_MEMSLOTS);
2454 
2455     /* If unspecified, use the default value */
2456     if (!s->nr_slots) {
2457         s->nr_slots = 32;
2458     }
2459 
2460     s->nr_as = kvm_check_extension(s, KVM_CAP_MULTI_ADDRESS_SPACE);
2461     if (s->nr_as <= 1) {
2462         s->nr_as = 1;
2463     }
2464     s->as = g_new0(struct KVMAs, s->nr_as);
2465 
2466     if (object_property_find(OBJECT(current_machine), "kvm-type")) {
2467         g_autofree char *kvm_type = object_property_get_str(OBJECT(current_machine),
2468                                                             "kvm-type",
2469                                                             &error_abort);
2470         type = mc->kvm_type(ms, kvm_type);
2471     } else if (mc->kvm_type) {
2472         type = mc->kvm_type(ms, NULL);
2473     } else {
2474         type = kvm_arch_get_default_type(ms);
2475     }
2476 
2477     if (type < 0) {
2478         ret = -EINVAL;
2479         goto err;
2480     }
2481 
2482     do {
2483         ret = kvm_ioctl(s, KVM_CREATE_VM, type);
2484     } while (ret == -EINTR);
2485 
2486     if (ret < 0) {
2487         fprintf(stderr, "ioctl(KVM_CREATE_VM) failed: %d %s\n", -ret,
2488                 strerror(-ret));
2489 
2490 #ifdef TARGET_S390X
2491         if (ret == -EINVAL) {
2492             fprintf(stderr,
2493                     "Host kernel setup problem detected. Please verify:\n");
2494             fprintf(stderr, "- for kernels supporting the switch_amode or"
2495                     " user_mode parameters, whether\n");
2496             fprintf(stderr,
2497                     "  user space is running in primary address space\n");
2498             fprintf(stderr,
2499                     "- for kernels supporting the vm.allocate_pgste sysctl, "
2500                     "whether it is enabled\n");
2501         }
2502 #elif defined(TARGET_PPC)
2503         if (ret == -EINVAL) {
2504             fprintf(stderr,
2505                     "PPC KVM module is not loaded. Try modprobe kvm_%s.\n",
2506                     (type == 2) ? "pr" : "hv");
2507         }
2508 #endif
2509         goto err;
2510     }
2511 
2512     s->vmfd = ret;
2513 
2514     /* check the vcpu limits */
2515     soft_vcpus_limit = kvm_recommended_vcpus(s);
2516     hard_vcpus_limit = kvm_max_vcpus(s);
2517 
2518     while (nc->name) {
2519         if (nc->num > soft_vcpus_limit) {
2520             warn_report("Number of %s cpus requested (%d) exceeds "
2521                         "the recommended cpus supported by KVM (%d)",
2522                         nc->name, nc->num, soft_vcpus_limit);
2523 
2524             if (nc->num > hard_vcpus_limit) {
2525                 fprintf(stderr, "Number of %s cpus requested (%d) exceeds "
2526                         "the maximum cpus supported by KVM (%d)\n",
2527                         nc->name, nc->num, hard_vcpus_limit);
2528                 exit(1);
2529             }
2530         }
2531         nc++;
2532     }
2533 
2534     missing_cap = kvm_check_extension_list(s, kvm_required_capabilites);
2535     if (!missing_cap) {
2536         missing_cap =
2537             kvm_check_extension_list(s, kvm_arch_required_capabilities);
2538     }
2539     if (missing_cap) {
2540         ret = -EINVAL;
2541         fprintf(stderr, "kvm does not support %s\n%s",
2542                 missing_cap->name, upgrade_note);
2543         goto err;
2544     }
2545 
2546     s->coalesced_mmio = kvm_check_extension(s, KVM_CAP_COALESCED_MMIO);
2547     s->coalesced_pio = s->coalesced_mmio &&
2548                        kvm_check_extension(s, KVM_CAP_COALESCED_PIO);
2549 
2550     /*
2551      * Enable KVM dirty ring if supported, otherwise fall back to
2552      * dirty logging mode
2553      */
2554     ret = kvm_dirty_ring_init(s);
2555     if (ret < 0) {
2556         goto err;
2557     }
2558 
2559     /*
2560      * KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2 is not needed when dirty ring is
2561      * enabled.  More importantly, KVM_DIRTY_LOG_INITIALLY_SET will assume no
2562      * page is wr-protected initially, which is against how kvm dirty ring is
2563      * usage - kvm dirty ring requires all pages are wr-protected at the very
2564      * beginning.  Enabling this feature for dirty ring causes data corruption.
2565      *
2566      * TODO: Without KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2 and kvm clear dirty log,
2567      * we may expect a higher stall time when starting the migration.  In the
2568      * future we can enable KVM_CLEAR_DIRTY_LOG to work with dirty ring too:
2569      * instead of clearing dirty bit, it can be a way to explicitly wr-protect
2570      * guest pages.
2571      */
2572     if (!s->kvm_dirty_ring_size) {
2573         dirty_log_manual_caps =
2574             kvm_check_extension(s, KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2);
2575         dirty_log_manual_caps &= (KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE |
2576                                   KVM_DIRTY_LOG_INITIALLY_SET);
2577         s->manual_dirty_log_protect = dirty_log_manual_caps;
2578         if (dirty_log_manual_caps) {
2579             ret = kvm_vm_enable_cap(s, KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2, 0,
2580                                     dirty_log_manual_caps);
2581             if (ret) {
2582                 warn_report("Trying to enable capability %"PRIu64" of "
2583                             "KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2 but failed. "
2584                             "Falling back to the legacy mode. ",
2585                             dirty_log_manual_caps);
2586                 s->manual_dirty_log_protect = 0;
2587             }
2588         }
2589     }
2590 
2591 #ifdef KVM_CAP_VCPU_EVENTS
2592     s->vcpu_events = kvm_check_extension(s, KVM_CAP_VCPU_EVENTS);
2593 #endif
2594     s->max_nested_state_len = kvm_check_extension(s, KVM_CAP_NESTED_STATE);
2595 
2596     s->irq_set_ioctl = KVM_IRQ_LINE;
2597     if (kvm_check_extension(s, KVM_CAP_IRQ_INJECT_STATUS)) {
2598         s->irq_set_ioctl = KVM_IRQ_LINE_STATUS;
2599     }
2600 
2601     kvm_readonly_mem_allowed =
2602         (kvm_check_extension(s, KVM_CAP_READONLY_MEM) > 0);
2603 
2604     kvm_resamplefds_allowed =
2605         (kvm_check_extension(s, KVM_CAP_IRQFD_RESAMPLE) > 0);
2606 
2607     kvm_vm_attributes_allowed =
2608         (kvm_check_extension(s, KVM_CAP_VM_ATTRIBUTES) > 0);
2609 
2610 #ifdef TARGET_KVM_HAVE_GUEST_DEBUG
2611     kvm_has_guest_debug =
2612         (kvm_check_extension(s, KVM_CAP_SET_GUEST_DEBUG) > 0);
2613 #endif
2614 
2615     kvm_sstep_flags = 0;
2616     if (kvm_has_guest_debug) {
2617         kvm_sstep_flags = SSTEP_ENABLE;
2618 
2619 #if defined TARGET_KVM_HAVE_GUEST_DEBUG
2620         int guest_debug_flags =
2621             kvm_check_extension(s, KVM_CAP_SET_GUEST_DEBUG2);
2622 
2623         if (guest_debug_flags & KVM_GUESTDBG_BLOCKIRQ) {
2624             kvm_sstep_flags |= SSTEP_NOIRQ;
2625         }
2626 #endif
2627     }
2628 
2629     kvm_state = s;
2630 
2631     ret = kvm_arch_init(ms, s);
2632     if (ret < 0) {
2633         goto err;
2634     }
2635 
2636     if (s->kernel_irqchip_split == ON_OFF_AUTO_AUTO) {
2637         s->kernel_irqchip_split = mc->default_kernel_irqchip_split ? ON_OFF_AUTO_ON : ON_OFF_AUTO_OFF;
2638     }
2639 
2640     qemu_register_reset(kvm_unpoison_all, NULL);
2641 
2642     if (s->kernel_irqchip_allowed) {
2643         kvm_irqchip_create(s);
2644     }
2645 
2646     s->memory_listener.listener.eventfd_add = kvm_mem_ioeventfd_add;
2647     s->memory_listener.listener.eventfd_del = kvm_mem_ioeventfd_del;
2648     s->memory_listener.listener.coalesced_io_add = kvm_coalesce_mmio_region;
2649     s->memory_listener.listener.coalesced_io_del = kvm_uncoalesce_mmio_region;
2650 
2651     kvm_memory_listener_register(s, &s->memory_listener,
2652                                  &address_space_memory, 0, "kvm-memory");
2653     memory_listener_register(&kvm_io_listener,
2654                              &address_space_io);
2655 
2656     s->sync_mmu = !!kvm_vm_check_extension(kvm_state, KVM_CAP_SYNC_MMU);
2657     if (!s->sync_mmu) {
2658         ret = ram_block_discard_disable(true);
2659         assert(!ret);
2660     }
2661 
2662     if (s->kvm_dirty_ring_size) {
2663         kvm_dirty_ring_reaper_init(s);
2664     }
2665 
2666     if (kvm_check_extension(kvm_state, KVM_CAP_BINARY_STATS_FD)) {
2667         add_stats_callbacks(STATS_PROVIDER_KVM, query_stats_cb,
2668                             query_stats_schemas_cb);
2669     }
2670 
2671     return 0;
2672 
2673 err:
2674     assert(ret < 0);
2675     if (s->vmfd >= 0) {
2676         close(s->vmfd);
2677     }
2678     if (s->fd != -1) {
2679         close(s->fd);
2680     }
2681     g_free(s->as);
2682     g_free(s->memory_listener.slots);
2683 
2684     return ret;
2685 }
2686 
2687 void kvm_set_sigmask_len(KVMState *s, unsigned int sigmask_len)
2688 {
2689     s->sigmask_len = sigmask_len;
2690 }
2691 
2692 static void kvm_handle_io(uint16_t port, MemTxAttrs attrs, void *data, int direction,
2693                           int size, uint32_t count)
2694 {
2695     int i;
2696     uint8_t *ptr = data;
2697 
2698     for (i = 0; i < count; i++) {
2699         address_space_rw(&address_space_io, port, attrs,
2700                          ptr, size,
2701                          direction == KVM_EXIT_IO_OUT);
2702         ptr += size;
2703     }
2704 }
2705 
2706 static int kvm_handle_internal_error(CPUState *cpu, struct kvm_run *run)
2707 {
2708     int i;
2709 
2710     fprintf(stderr, "KVM internal error. Suberror: %d\n",
2711             run->internal.suberror);
2712 
2713     for (i = 0; i < run->internal.ndata; ++i) {
2714         fprintf(stderr, "extra data[%d]: 0x%016"PRIx64"\n",
2715                 i, (uint64_t)run->internal.data[i]);
2716     }
2717     if (run->internal.suberror == KVM_INTERNAL_ERROR_EMULATION) {
2718         fprintf(stderr, "emulation failure\n");
2719         if (!kvm_arch_stop_on_emulation_error(cpu)) {
2720             cpu_dump_state(cpu, stderr, CPU_DUMP_CODE);
2721             return EXCP_INTERRUPT;
2722         }
2723     }
2724     /* FIXME: Should trigger a qmp message to let management know
2725      * something went wrong.
2726      */
2727     return -1;
2728 }
2729 
2730 void kvm_flush_coalesced_mmio_buffer(void)
2731 {
2732     KVMState *s = kvm_state;
2733 
2734     if (!s || s->coalesced_flush_in_progress) {
2735         return;
2736     }
2737 
2738     s->coalesced_flush_in_progress = true;
2739 
2740     if (s->coalesced_mmio_ring) {
2741         struct kvm_coalesced_mmio_ring *ring = s->coalesced_mmio_ring;
2742         while (ring->first != ring->last) {
2743             struct kvm_coalesced_mmio *ent;
2744 
2745             ent = &ring->coalesced_mmio[ring->first];
2746 
2747             if (ent->pio == 1) {
2748                 address_space_write(&address_space_io, ent->phys_addr,
2749                                     MEMTXATTRS_UNSPECIFIED, ent->data,
2750                                     ent->len);
2751             } else {
2752                 cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len);
2753             }
2754             smp_wmb();
2755             ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX;
2756         }
2757     }
2758 
2759     s->coalesced_flush_in_progress = false;
2760 }
2761 
2762 static void do_kvm_cpu_synchronize_state(CPUState *cpu, run_on_cpu_data arg)
2763 {
2764     if (!cpu->vcpu_dirty && !kvm_state->guest_state_protected) {
2765         int ret = kvm_arch_get_registers(cpu);
2766         if (ret) {
2767             error_report("Failed to get registers: %s", strerror(-ret));
2768             cpu_dump_state(cpu, stderr, CPU_DUMP_CODE);
2769             vm_stop(RUN_STATE_INTERNAL_ERROR);
2770         }
2771 
2772         cpu->vcpu_dirty = true;
2773     }
2774 }
2775 
2776 void kvm_cpu_synchronize_state(CPUState *cpu)
2777 {
2778     if (!cpu->vcpu_dirty && !kvm_state->guest_state_protected) {
2779         run_on_cpu(cpu, do_kvm_cpu_synchronize_state, RUN_ON_CPU_NULL);
2780     }
2781 }
2782 
2783 static void do_kvm_cpu_synchronize_post_reset(CPUState *cpu, run_on_cpu_data arg)
2784 {
2785     int ret = kvm_arch_put_registers(cpu, KVM_PUT_RESET_STATE);
2786     if (ret) {
2787         error_report("Failed to put registers after reset: %s", strerror(-ret));
2788         cpu_dump_state(cpu, stderr, CPU_DUMP_CODE);
2789         vm_stop(RUN_STATE_INTERNAL_ERROR);
2790     }
2791 
2792     cpu->vcpu_dirty = false;
2793 }
2794 
2795 void kvm_cpu_synchronize_post_reset(CPUState *cpu)
2796 {
2797     run_on_cpu(cpu, do_kvm_cpu_synchronize_post_reset, RUN_ON_CPU_NULL);
2798 }
2799 
2800 static void do_kvm_cpu_synchronize_post_init(CPUState *cpu, run_on_cpu_data arg)
2801 {
2802     int ret = kvm_arch_put_registers(cpu, KVM_PUT_FULL_STATE);
2803     if (ret) {
2804         error_report("Failed to put registers after init: %s", strerror(-ret));
2805         exit(1);
2806     }
2807 
2808     cpu->vcpu_dirty = false;
2809 }
2810 
2811 void kvm_cpu_synchronize_post_init(CPUState *cpu)
2812 {
2813     if (!kvm_state->guest_state_protected) {
2814         /*
2815          * This runs before the machine_init_done notifiers, and is the last
2816          * opportunity to synchronize the state of confidential guests.
2817          */
2818         run_on_cpu(cpu, do_kvm_cpu_synchronize_post_init, RUN_ON_CPU_NULL);
2819     }
2820 }
2821 
2822 static void do_kvm_cpu_synchronize_pre_loadvm(CPUState *cpu, run_on_cpu_data arg)
2823 {
2824     cpu->vcpu_dirty = true;
2825 }
2826 
2827 void kvm_cpu_synchronize_pre_loadvm(CPUState *cpu)
2828 {
2829     run_on_cpu(cpu, do_kvm_cpu_synchronize_pre_loadvm, RUN_ON_CPU_NULL);
2830 }
2831 
2832 #ifdef KVM_HAVE_MCE_INJECTION
2833 static __thread void *pending_sigbus_addr;
2834 static __thread int pending_sigbus_code;
2835 static __thread bool have_sigbus_pending;
2836 #endif
2837 
2838 static void kvm_cpu_kick(CPUState *cpu)
2839 {
2840     qatomic_set(&cpu->kvm_run->immediate_exit, 1);
2841 }
2842 
2843 static void kvm_cpu_kick_self(void)
2844 {
2845     if (kvm_immediate_exit) {
2846         kvm_cpu_kick(current_cpu);
2847     } else {
2848         qemu_cpu_kick_self();
2849     }
2850 }
2851 
2852 static void kvm_eat_signals(CPUState *cpu)
2853 {
2854     struct timespec ts = { 0, 0 };
2855     siginfo_t siginfo;
2856     sigset_t waitset;
2857     sigset_t chkset;
2858     int r;
2859 
2860     if (kvm_immediate_exit) {
2861         qatomic_set(&cpu->kvm_run->immediate_exit, 0);
2862         /* Write kvm_run->immediate_exit before the cpu->exit_request
2863          * write in kvm_cpu_exec.
2864          */
2865         smp_wmb();
2866         return;
2867     }
2868 
2869     sigemptyset(&waitset);
2870     sigaddset(&waitset, SIG_IPI);
2871 
2872     do {
2873         r = sigtimedwait(&waitset, &siginfo, &ts);
2874         if (r == -1 && !(errno == EAGAIN || errno == EINTR)) {
2875             perror("sigtimedwait");
2876             exit(1);
2877         }
2878 
2879         r = sigpending(&chkset);
2880         if (r == -1) {
2881             perror("sigpending");
2882             exit(1);
2883         }
2884     } while (sigismember(&chkset, SIG_IPI));
2885 }
2886 
2887 int kvm_convert_memory(hwaddr start, hwaddr size, bool to_private)
2888 {
2889     MemoryRegionSection section;
2890     ram_addr_t offset;
2891     MemoryRegion *mr;
2892     RAMBlock *rb;
2893     void *addr;
2894     int ret = -1;
2895 
2896     trace_kvm_convert_memory(start, size, to_private ? "shared_to_private" : "private_to_shared");
2897 
2898     if (!QEMU_PTR_IS_ALIGNED(start, qemu_real_host_page_size()) ||
2899         !QEMU_PTR_IS_ALIGNED(size, qemu_real_host_page_size())) {
2900         return -1;
2901     }
2902 
2903     if (!size) {
2904         return -1;
2905     }
2906 
2907     section = memory_region_find(get_system_memory(), start, size);
2908     mr = section.mr;
2909     if (!mr) {
2910         /*
2911          * Ignore converting non-assigned region to shared.
2912          *
2913          * TDX requires vMMIO region to be shared to inject #VE to guest.
2914          * OVMF issues conservatively MapGPA(shared) on 32bit PCI MMIO region,
2915          * and vIO-APIC 0xFEC00000 4K page.
2916          * OVMF assigns 32bit PCI MMIO region to
2917          * [top of low memory: typically 2GB=0xC000000,  0xFC00000)
2918          */
2919         if (!to_private) {
2920             return 0;
2921         }
2922         return -1;
2923     }
2924 
2925     if (!memory_region_has_guest_memfd(mr)) {
2926         /*
2927          * Because vMMIO region must be shared, guest TD may convert vMMIO
2928          * region to shared explicitly.  Don't complain such case.  See
2929          * memory_region_type() for checking if the region is MMIO region.
2930          */
2931         if (!to_private &&
2932             !memory_region_is_ram(mr) &&
2933             !memory_region_is_ram_device(mr) &&
2934             !memory_region_is_rom(mr) &&
2935             !memory_region_is_romd(mr)) {
2936             ret = 0;
2937         } else {
2938             error_report("Convert non guest_memfd backed memory region "
2939                         "(0x%"HWADDR_PRIx" ,+ 0x%"HWADDR_PRIx") to %s",
2940                         start, size, to_private ? "private" : "shared");
2941         }
2942         goto out_unref;
2943     }
2944 
2945     if (to_private) {
2946         ret = kvm_set_memory_attributes_private(start, size);
2947     } else {
2948         ret = kvm_set_memory_attributes_shared(start, size);
2949     }
2950     if (ret) {
2951         goto out_unref;
2952     }
2953 
2954     addr = memory_region_get_ram_ptr(mr) + section.offset_within_region;
2955     rb = qemu_ram_block_from_host(addr, false, &offset);
2956 
2957     if (to_private) {
2958         if (rb->page_size != qemu_real_host_page_size()) {
2959             /*
2960              * shared memory is backed by hugetlb, which is supposed to be
2961              * pre-allocated and doesn't need to be discarded
2962              */
2963             goto out_unref;
2964         }
2965         ret = ram_block_discard_range(rb, offset, size);
2966     } else {
2967         ret = ram_block_discard_guest_memfd_range(rb, offset, size);
2968     }
2969 
2970 out_unref:
2971     memory_region_unref(mr);
2972     return ret;
2973 }
2974 
2975 int kvm_cpu_exec(CPUState *cpu)
2976 {
2977     struct kvm_run *run = cpu->kvm_run;
2978     int ret, run_ret;
2979 
2980     trace_kvm_cpu_exec();
2981 
2982     if (kvm_arch_process_async_events(cpu)) {
2983         qatomic_set(&cpu->exit_request, 0);
2984         return EXCP_HLT;
2985     }
2986 
2987     bql_unlock();
2988     cpu_exec_start(cpu);
2989 
2990     do {
2991         MemTxAttrs attrs;
2992 
2993         if (cpu->vcpu_dirty) {
2994             ret = kvm_arch_put_registers(cpu, KVM_PUT_RUNTIME_STATE);
2995             if (ret) {
2996                 error_report("Failed to put registers after init: %s",
2997                              strerror(-ret));
2998                 ret = -1;
2999                 break;
3000             }
3001 
3002             cpu->vcpu_dirty = false;
3003         }
3004 
3005         kvm_arch_pre_run(cpu, run);
3006         if (qatomic_read(&cpu->exit_request)) {
3007             trace_kvm_interrupt_exit_request();
3008             /*
3009              * KVM requires us to reenter the kernel after IO exits to complete
3010              * instruction emulation. This self-signal will ensure that we
3011              * leave ASAP again.
3012              */
3013             kvm_cpu_kick_self();
3014         }
3015 
3016         /* Read cpu->exit_request before KVM_RUN reads run->immediate_exit.
3017          * Matching barrier in kvm_eat_signals.
3018          */
3019         smp_rmb();
3020 
3021         run_ret = kvm_vcpu_ioctl(cpu, KVM_RUN, 0);
3022 
3023         attrs = kvm_arch_post_run(cpu, run);
3024 
3025 #ifdef KVM_HAVE_MCE_INJECTION
3026         if (unlikely(have_sigbus_pending)) {
3027             bql_lock();
3028             kvm_arch_on_sigbus_vcpu(cpu, pending_sigbus_code,
3029                                     pending_sigbus_addr);
3030             have_sigbus_pending = false;
3031             bql_unlock();
3032         }
3033 #endif
3034 
3035         if (run_ret < 0) {
3036             if (run_ret == -EINTR || run_ret == -EAGAIN) {
3037                 trace_kvm_io_window_exit();
3038                 kvm_eat_signals(cpu);
3039                 ret = EXCP_INTERRUPT;
3040                 break;
3041             }
3042             if (!(run_ret == -EFAULT && run->exit_reason == KVM_EXIT_MEMORY_FAULT)) {
3043                 fprintf(stderr, "error: kvm run failed %s\n",
3044                         strerror(-run_ret));
3045 #ifdef TARGET_PPC
3046                 if (run_ret == -EBUSY) {
3047                     fprintf(stderr,
3048                             "This is probably because your SMT is enabled.\n"
3049                             "VCPU can only run on primary threads with all "
3050                             "secondary threads offline.\n");
3051                 }
3052 #endif
3053                 ret = -1;
3054                 break;
3055             }
3056         }
3057 
3058         trace_kvm_run_exit(cpu->cpu_index, run->exit_reason);
3059         switch (run->exit_reason) {
3060         case KVM_EXIT_IO:
3061             /* Called outside BQL */
3062             kvm_handle_io(run->io.port, attrs,
3063                           (uint8_t *)run + run->io.data_offset,
3064                           run->io.direction,
3065                           run->io.size,
3066                           run->io.count);
3067             ret = 0;
3068             break;
3069         case KVM_EXIT_MMIO:
3070             /* Called outside BQL */
3071             address_space_rw(&address_space_memory,
3072                              run->mmio.phys_addr, attrs,
3073                              run->mmio.data,
3074                              run->mmio.len,
3075                              run->mmio.is_write);
3076             ret = 0;
3077             break;
3078         case KVM_EXIT_IRQ_WINDOW_OPEN:
3079             ret = EXCP_INTERRUPT;
3080             break;
3081         case KVM_EXIT_SHUTDOWN:
3082             qemu_system_reset_request(SHUTDOWN_CAUSE_GUEST_RESET);
3083             ret = EXCP_INTERRUPT;
3084             break;
3085         case KVM_EXIT_UNKNOWN:
3086             fprintf(stderr, "KVM: unknown exit, hardware reason %" PRIx64 "\n",
3087                     (uint64_t)run->hw.hardware_exit_reason);
3088             ret = -1;
3089             break;
3090         case KVM_EXIT_INTERNAL_ERROR:
3091             ret = kvm_handle_internal_error(cpu, run);
3092             break;
3093         case KVM_EXIT_DIRTY_RING_FULL:
3094             /*
3095              * We shouldn't continue if the dirty ring of this vcpu is
3096              * still full.  Got kicked by KVM_RESET_DIRTY_RINGS.
3097              */
3098             trace_kvm_dirty_ring_full(cpu->cpu_index);
3099             bql_lock();
3100             /*
3101              * We throttle vCPU by making it sleep once it exit from kernel
3102              * due to dirty ring full. In the dirtylimit scenario, reaping
3103              * all vCPUs after a single vCPU dirty ring get full result in
3104              * the miss of sleep, so just reap the ring-fulled vCPU.
3105              */
3106             if (dirtylimit_in_service()) {
3107                 kvm_dirty_ring_reap(kvm_state, cpu);
3108             } else {
3109                 kvm_dirty_ring_reap(kvm_state, NULL);
3110             }
3111             bql_unlock();
3112             dirtylimit_vcpu_execute(cpu);
3113             ret = 0;
3114             break;
3115         case KVM_EXIT_SYSTEM_EVENT:
3116             trace_kvm_run_exit_system_event(cpu->cpu_index, run->system_event.type);
3117             switch (run->system_event.type) {
3118             case KVM_SYSTEM_EVENT_SHUTDOWN:
3119                 qemu_system_shutdown_request(SHUTDOWN_CAUSE_GUEST_SHUTDOWN);
3120                 ret = EXCP_INTERRUPT;
3121                 break;
3122             case KVM_SYSTEM_EVENT_RESET:
3123                 qemu_system_reset_request(SHUTDOWN_CAUSE_GUEST_RESET);
3124                 ret = EXCP_INTERRUPT;
3125                 break;
3126             case KVM_SYSTEM_EVENT_CRASH:
3127                 kvm_cpu_synchronize_state(cpu);
3128                 bql_lock();
3129                 qemu_system_guest_panicked(cpu_get_crash_info(cpu));
3130                 bql_unlock();
3131                 ret = 0;
3132                 break;
3133             default:
3134                 ret = kvm_arch_handle_exit(cpu, run);
3135                 break;
3136             }
3137             break;
3138         case KVM_EXIT_MEMORY_FAULT:
3139             trace_kvm_memory_fault(run->memory_fault.gpa,
3140                                    run->memory_fault.size,
3141                                    run->memory_fault.flags);
3142             if (run->memory_fault.flags & ~KVM_MEMORY_EXIT_FLAG_PRIVATE) {
3143                 error_report("KVM_EXIT_MEMORY_FAULT: Unknown flag 0x%" PRIx64,
3144                              (uint64_t)run->memory_fault.flags);
3145                 ret = -1;
3146                 break;
3147             }
3148             ret = kvm_convert_memory(run->memory_fault.gpa, run->memory_fault.size,
3149                                      run->memory_fault.flags & KVM_MEMORY_EXIT_FLAG_PRIVATE);
3150             break;
3151         default:
3152             ret = kvm_arch_handle_exit(cpu, run);
3153             break;
3154         }
3155     } while (ret == 0);
3156 
3157     cpu_exec_end(cpu);
3158     bql_lock();
3159 
3160     if (ret < 0) {
3161         cpu_dump_state(cpu, stderr, CPU_DUMP_CODE);
3162         vm_stop(RUN_STATE_INTERNAL_ERROR);
3163     }
3164 
3165     qatomic_set(&cpu->exit_request, 0);
3166     return ret;
3167 }
3168 
3169 int kvm_ioctl(KVMState *s, unsigned long type, ...)
3170 {
3171     int ret;
3172     void *arg;
3173     va_list ap;
3174 
3175     va_start(ap, type);
3176     arg = va_arg(ap, void *);
3177     va_end(ap);
3178 
3179     trace_kvm_ioctl(type, arg);
3180     ret = ioctl(s->fd, type, arg);
3181     if (ret == -1) {
3182         ret = -errno;
3183     }
3184     return ret;
3185 }
3186 
3187 int kvm_vm_ioctl(KVMState *s, unsigned long type, ...)
3188 {
3189     int ret;
3190     void *arg;
3191     va_list ap;
3192 
3193     va_start(ap, type);
3194     arg = va_arg(ap, void *);
3195     va_end(ap);
3196 
3197     trace_kvm_vm_ioctl(type, arg);
3198     accel_ioctl_begin();
3199     ret = ioctl(s->vmfd, type, arg);
3200     accel_ioctl_end();
3201     if (ret == -1) {
3202         ret = -errno;
3203     }
3204     return ret;
3205 }
3206 
3207 int kvm_vcpu_ioctl(CPUState *cpu, unsigned long type, ...)
3208 {
3209     int ret;
3210     void *arg;
3211     va_list ap;
3212 
3213     va_start(ap, type);
3214     arg = va_arg(ap, void *);
3215     va_end(ap);
3216 
3217     trace_kvm_vcpu_ioctl(cpu->cpu_index, type, arg);
3218     accel_cpu_ioctl_begin(cpu);
3219     ret = ioctl(cpu->kvm_fd, type, arg);
3220     accel_cpu_ioctl_end(cpu);
3221     if (ret == -1) {
3222         ret = -errno;
3223     }
3224     return ret;
3225 }
3226 
3227 int kvm_device_ioctl(int fd, unsigned long type, ...)
3228 {
3229     int ret;
3230     void *arg;
3231     va_list ap;
3232 
3233     va_start(ap, type);
3234     arg = va_arg(ap, void *);
3235     va_end(ap);
3236 
3237     trace_kvm_device_ioctl(fd, type, arg);
3238     accel_ioctl_begin();
3239     ret = ioctl(fd, type, arg);
3240     accel_ioctl_end();
3241     if (ret == -1) {
3242         ret = -errno;
3243     }
3244     return ret;
3245 }
3246 
3247 int kvm_vm_check_attr(KVMState *s, uint32_t group, uint64_t attr)
3248 {
3249     int ret;
3250     struct kvm_device_attr attribute = {
3251         .group = group,
3252         .attr = attr,
3253     };
3254 
3255     if (!kvm_vm_attributes_allowed) {
3256         return 0;
3257     }
3258 
3259     ret = kvm_vm_ioctl(s, KVM_HAS_DEVICE_ATTR, &attribute);
3260     /* kvm returns 0 on success for HAS_DEVICE_ATTR */
3261     return ret ? 0 : 1;
3262 }
3263 
3264 int kvm_device_check_attr(int dev_fd, uint32_t group, uint64_t attr)
3265 {
3266     struct kvm_device_attr attribute = {
3267         .group = group,
3268         .attr = attr,
3269         .flags = 0,
3270     };
3271 
3272     return kvm_device_ioctl(dev_fd, KVM_HAS_DEVICE_ATTR, &attribute) ? 0 : 1;
3273 }
3274 
3275 int kvm_device_access(int fd, int group, uint64_t attr,
3276                       void *val, bool write, Error **errp)
3277 {
3278     struct kvm_device_attr kvmattr;
3279     int err;
3280 
3281     kvmattr.flags = 0;
3282     kvmattr.group = group;
3283     kvmattr.attr = attr;
3284     kvmattr.addr = (uintptr_t)val;
3285 
3286     err = kvm_device_ioctl(fd,
3287                            write ? KVM_SET_DEVICE_ATTR : KVM_GET_DEVICE_ATTR,
3288                            &kvmattr);
3289     if (err < 0) {
3290         error_setg_errno(errp, -err,
3291                          "KVM_%s_DEVICE_ATTR failed: Group %d "
3292                          "attr 0x%016" PRIx64,
3293                          write ? "SET" : "GET", group, attr);
3294     }
3295     return err;
3296 }
3297 
3298 bool kvm_has_sync_mmu(void)
3299 {
3300     return kvm_state->sync_mmu;
3301 }
3302 
3303 int kvm_has_vcpu_events(void)
3304 {
3305     return kvm_state->vcpu_events;
3306 }
3307 
3308 int kvm_max_nested_state_length(void)
3309 {
3310     return kvm_state->max_nested_state_len;
3311 }
3312 
3313 int kvm_has_gsi_routing(void)
3314 {
3315 #ifdef KVM_CAP_IRQ_ROUTING
3316     return kvm_check_extension(kvm_state, KVM_CAP_IRQ_ROUTING);
3317 #else
3318     return false;
3319 #endif
3320 }
3321 
3322 bool kvm_arm_supports_user_irq(void)
3323 {
3324     return kvm_check_extension(kvm_state, KVM_CAP_ARM_USER_IRQ);
3325 }
3326 
3327 #ifdef TARGET_KVM_HAVE_GUEST_DEBUG
3328 struct kvm_sw_breakpoint *kvm_find_sw_breakpoint(CPUState *cpu, vaddr pc)
3329 {
3330     struct kvm_sw_breakpoint *bp;
3331 
3332     QTAILQ_FOREACH(bp, &cpu->kvm_state->kvm_sw_breakpoints, entry) {
3333         if (bp->pc == pc) {
3334             return bp;
3335         }
3336     }
3337     return NULL;
3338 }
3339 
3340 int kvm_sw_breakpoints_active(CPUState *cpu)
3341 {
3342     return !QTAILQ_EMPTY(&cpu->kvm_state->kvm_sw_breakpoints);
3343 }
3344 
3345 struct kvm_set_guest_debug_data {
3346     struct kvm_guest_debug dbg;
3347     int err;
3348 };
3349 
3350 static void kvm_invoke_set_guest_debug(CPUState *cpu, run_on_cpu_data data)
3351 {
3352     struct kvm_set_guest_debug_data *dbg_data =
3353         (struct kvm_set_guest_debug_data *) data.host_ptr;
3354 
3355     dbg_data->err = kvm_vcpu_ioctl(cpu, KVM_SET_GUEST_DEBUG,
3356                                    &dbg_data->dbg);
3357 }
3358 
3359 int kvm_update_guest_debug(CPUState *cpu, unsigned long reinject_trap)
3360 {
3361     struct kvm_set_guest_debug_data data;
3362 
3363     data.dbg.control = reinject_trap;
3364 
3365     if (cpu->singlestep_enabled) {
3366         data.dbg.control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_SINGLESTEP;
3367 
3368         if (cpu->singlestep_enabled & SSTEP_NOIRQ) {
3369             data.dbg.control |= KVM_GUESTDBG_BLOCKIRQ;
3370         }
3371     }
3372     kvm_arch_update_guest_debug(cpu, &data.dbg);
3373 
3374     run_on_cpu(cpu, kvm_invoke_set_guest_debug,
3375                RUN_ON_CPU_HOST_PTR(&data));
3376     return data.err;
3377 }
3378 
3379 bool kvm_supports_guest_debug(void)
3380 {
3381     /* probed during kvm_init() */
3382     return kvm_has_guest_debug;
3383 }
3384 
3385 int kvm_insert_breakpoint(CPUState *cpu, int type, vaddr addr, vaddr len)
3386 {
3387     struct kvm_sw_breakpoint *bp;
3388     int err;
3389 
3390     if (type == GDB_BREAKPOINT_SW) {
3391         bp = kvm_find_sw_breakpoint(cpu, addr);
3392         if (bp) {
3393             bp->use_count++;
3394             return 0;
3395         }
3396 
3397         bp = g_new(struct kvm_sw_breakpoint, 1);
3398         bp->pc = addr;
3399         bp->use_count = 1;
3400         err = kvm_arch_insert_sw_breakpoint(cpu, bp);
3401         if (err) {
3402             g_free(bp);
3403             return err;
3404         }
3405 
3406         QTAILQ_INSERT_HEAD(&cpu->kvm_state->kvm_sw_breakpoints, bp, entry);
3407     } else {
3408         err = kvm_arch_insert_hw_breakpoint(addr, len, type);
3409         if (err) {
3410             return err;
3411         }
3412     }
3413 
3414     CPU_FOREACH(cpu) {
3415         err = kvm_update_guest_debug(cpu, 0);
3416         if (err) {
3417             return err;
3418         }
3419     }
3420     return 0;
3421 }
3422 
3423 int kvm_remove_breakpoint(CPUState *cpu, int type, vaddr addr, vaddr len)
3424 {
3425     struct kvm_sw_breakpoint *bp;
3426     int err;
3427 
3428     if (type == GDB_BREAKPOINT_SW) {
3429         bp = kvm_find_sw_breakpoint(cpu, addr);
3430         if (!bp) {
3431             return -ENOENT;
3432         }
3433 
3434         if (bp->use_count > 1) {
3435             bp->use_count--;
3436             return 0;
3437         }
3438 
3439         err = kvm_arch_remove_sw_breakpoint(cpu, bp);
3440         if (err) {
3441             return err;
3442         }
3443 
3444         QTAILQ_REMOVE(&cpu->kvm_state->kvm_sw_breakpoints, bp, entry);
3445         g_free(bp);
3446     } else {
3447         err = kvm_arch_remove_hw_breakpoint(addr, len, type);
3448         if (err) {
3449             return err;
3450         }
3451     }
3452 
3453     CPU_FOREACH(cpu) {
3454         err = kvm_update_guest_debug(cpu, 0);
3455         if (err) {
3456             return err;
3457         }
3458     }
3459     return 0;
3460 }
3461 
3462 void kvm_remove_all_breakpoints(CPUState *cpu)
3463 {
3464     struct kvm_sw_breakpoint *bp, *next;
3465     KVMState *s = cpu->kvm_state;
3466     CPUState *tmpcpu;
3467 
3468     QTAILQ_FOREACH_SAFE(bp, &s->kvm_sw_breakpoints, entry, next) {
3469         if (kvm_arch_remove_sw_breakpoint(cpu, bp) != 0) {
3470             /* Try harder to find a CPU that currently sees the breakpoint. */
3471             CPU_FOREACH(tmpcpu) {
3472                 if (kvm_arch_remove_sw_breakpoint(tmpcpu, bp) == 0) {
3473                     break;
3474                 }
3475             }
3476         }
3477         QTAILQ_REMOVE(&s->kvm_sw_breakpoints, bp, entry);
3478         g_free(bp);
3479     }
3480     kvm_arch_remove_all_hw_breakpoints();
3481 
3482     CPU_FOREACH(cpu) {
3483         kvm_update_guest_debug(cpu, 0);
3484     }
3485 }
3486 
3487 #endif /* !TARGET_KVM_HAVE_GUEST_DEBUG */
3488 
3489 static int kvm_set_signal_mask(CPUState *cpu, const sigset_t *sigset)
3490 {
3491     KVMState *s = kvm_state;
3492     struct kvm_signal_mask *sigmask;
3493     int r;
3494 
3495     sigmask = g_malloc(sizeof(*sigmask) + sizeof(*sigset));
3496 
3497     sigmask->len = s->sigmask_len;
3498     memcpy(sigmask->sigset, sigset, sizeof(*sigset));
3499     r = kvm_vcpu_ioctl(cpu, KVM_SET_SIGNAL_MASK, sigmask);
3500     g_free(sigmask);
3501 
3502     return r;
3503 }
3504 
3505 static void kvm_ipi_signal(int sig)
3506 {
3507     if (current_cpu) {
3508         assert(kvm_immediate_exit);
3509         kvm_cpu_kick(current_cpu);
3510     }
3511 }
3512 
3513 void kvm_init_cpu_signals(CPUState *cpu)
3514 {
3515     int r;
3516     sigset_t set;
3517     struct sigaction sigact;
3518 
3519     memset(&sigact, 0, sizeof(sigact));
3520     sigact.sa_handler = kvm_ipi_signal;
3521     sigaction(SIG_IPI, &sigact, NULL);
3522 
3523     pthread_sigmask(SIG_BLOCK, NULL, &set);
3524 #if defined KVM_HAVE_MCE_INJECTION
3525     sigdelset(&set, SIGBUS);
3526     pthread_sigmask(SIG_SETMASK, &set, NULL);
3527 #endif
3528     sigdelset(&set, SIG_IPI);
3529     if (kvm_immediate_exit) {
3530         r = pthread_sigmask(SIG_SETMASK, &set, NULL);
3531     } else {
3532         r = kvm_set_signal_mask(cpu, &set);
3533     }
3534     if (r) {
3535         fprintf(stderr, "kvm_set_signal_mask: %s\n", strerror(-r));
3536         exit(1);
3537     }
3538 }
3539 
3540 /* Called asynchronously in VCPU thread.  */
3541 int kvm_on_sigbus_vcpu(CPUState *cpu, int code, void *addr)
3542 {
3543 #ifdef KVM_HAVE_MCE_INJECTION
3544     if (have_sigbus_pending) {
3545         return 1;
3546     }
3547     have_sigbus_pending = true;
3548     pending_sigbus_addr = addr;
3549     pending_sigbus_code = code;
3550     qatomic_set(&cpu->exit_request, 1);
3551     return 0;
3552 #else
3553     return 1;
3554 #endif
3555 }
3556 
3557 /* Called synchronously (via signalfd) in main thread.  */
3558 int kvm_on_sigbus(int code, void *addr)
3559 {
3560 #ifdef KVM_HAVE_MCE_INJECTION
3561     /* Action required MCE kills the process if SIGBUS is blocked.  Because
3562      * that's what happens in the I/O thread, where we handle MCE via signalfd,
3563      * we can only get action optional here.
3564      */
3565     assert(code != BUS_MCEERR_AR);
3566     kvm_arch_on_sigbus_vcpu(first_cpu, code, addr);
3567     return 0;
3568 #else
3569     return 1;
3570 #endif
3571 }
3572 
3573 int kvm_create_device(KVMState *s, uint64_t type, bool test)
3574 {
3575     int ret;
3576     struct kvm_create_device create_dev;
3577 
3578     create_dev.type = type;
3579     create_dev.fd = -1;
3580     create_dev.flags = test ? KVM_CREATE_DEVICE_TEST : 0;
3581 
3582     if (!kvm_check_extension(s, KVM_CAP_DEVICE_CTRL)) {
3583         return -ENOTSUP;
3584     }
3585 
3586     ret = kvm_vm_ioctl(s, KVM_CREATE_DEVICE, &create_dev);
3587     if (ret) {
3588         return ret;
3589     }
3590 
3591     return test ? 0 : create_dev.fd;
3592 }
3593 
3594 bool kvm_device_supported(int vmfd, uint64_t type)
3595 {
3596     struct kvm_create_device create_dev = {
3597         .type = type,
3598         .fd = -1,
3599         .flags = KVM_CREATE_DEVICE_TEST,
3600     };
3601 
3602     if (ioctl(vmfd, KVM_CHECK_EXTENSION, KVM_CAP_DEVICE_CTRL) <= 0) {
3603         return false;
3604     }
3605 
3606     return (ioctl(vmfd, KVM_CREATE_DEVICE, &create_dev) >= 0);
3607 }
3608 
3609 int kvm_set_one_reg(CPUState *cs, uint64_t id, void *source)
3610 {
3611     struct kvm_one_reg reg;
3612     int r;
3613 
3614     reg.id = id;
3615     reg.addr = (uintptr_t) source;
3616     r = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, &reg);
3617     if (r) {
3618         trace_kvm_failed_reg_set(id, strerror(-r));
3619     }
3620     return r;
3621 }
3622 
3623 int kvm_get_one_reg(CPUState *cs, uint64_t id, void *target)
3624 {
3625     struct kvm_one_reg reg;
3626     int r;
3627 
3628     reg.id = id;
3629     reg.addr = (uintptr_t) target;
3630     r = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &reg);
3631     if (r) {
3632         trace_kvm_failed_reg_get(id, strerror(-r));
3633     }
3634     return r;
3635 }
3636 
3637 static bool kvm_accel_has_memory(MachineState *ms, AddressSpace *as,
3638                                  hwaddr start_addr, hwaddr size)
3639 {
3640     KVMState *kvm = KVM_STATE(ms->accelerator);
3641     int i;
3642 
3643     for (i = 0; i < kvm->nr_as; ++i) {
3644         if (kvm->as[i].as == as && kvm->as[i].ml) {
3645             size = MIN(kvm_max_slot_size, size);
3646             return NULL != kvm_lookup_matching_slot(kvm->as[i].ml,
3647                                                     start_addr, size);
3648         }
3649     }
3650 
3651     return false;
3652 }
3653 
3654 static void kvm_get_kvm_shadow_mem(Object *obj, Visitor *v,
3655                                    const char *name, void *opaque,
3656                                    Error **errp)
3657 {
3658     KVMState *s = KVM_STATE(obj);
3659     int64_t value = s->kvm_shadow_mem;
3660 
3661     visit_type_int(v, name, &value, errp);
3662 }
3663 
3664 static void kvm_set_kvm_shadow_mem(Object *obj, Visitor *v,
3665                                    const char *name, void *opaque,
3666                                    Error **errp)
3667 {
3668     KVMState *s = KVM_STATE(obj);
3669     int64_t value;
3670 
3671     if (s->fd != -1) {
3672         error_setg(errp, "Cannot set properties after the accelerator has been initialized");
3673         return;
3674     }
3675 
3676     if (!visit_type_int(v, name, &value, errp)) {
3677         return;
3678     }
3679 
3680     s->kvm_shadow_mem = value;
3681 }
3682 
3683 static void kvm_set_kernel_irqchip(Object *obj, Visitor *v,
3684                                    const char *name, void *opaque,
3685                                    Error **errp)
3686 {
3687     KVMState *s = KVM_STATE(obj);
3688     OnOffSplit mode;
3689 
3690     if (s->fd != -1) {
3691         error_setg(errp, "Cannot set properties after the accelerator has been initialized");
3692         return;
3693     }
3694 
3695     if (!visit_type_OnOffSplit(v, name, &mode, errp)) {
3696         return;
3697     }
3698     switch (mode) {
3699     case ON_OFF_SPLIT_ON:
3700         s->kernel_irqchip_allowed = true;
3701         s->kernel_irqchip_required = true;
3702         s->kernel_irqchip_split = ON_OFF_AUTO_OFF;
3703         break;
3704     case ON_OFF_SPLIT_OFF:
3705         s->kernel_irqchip_allowed = false;
3706         s->kernel_irqchip_required = false;
3707         s->kernel_irqchip_split = ON_OFF_AUTO_OFF;
3708         break;
3709     case ON_OFF_SPLIT_SPLIT:
3710         s->kernel_irqchip_allowed = true;
3711         s->kernel_irqchip_required = true;
3712         s->kernel_irqchip_split = ON_OFF_AUTO_ON;
3713         break;
3714     default:
3715         /* The value was checked in visit_type_OnOffSplit() above. If
3716          * we get here, then something is wrong in QEMU.
3717          */
3718         abort();
3719     }
3720 }
3721 
3722 bool kvm_kernel_irqchip_allowed(void)
3723 {
3724     return kvm_state->kernel_irqchip_allowed;
3725 }
3726 
3727 bool kvm_kernel_irqchip_required(void)
3728 {
3729     return kvm_state->kernel_irqchip_required;
3730 }
3731 
3732 bool kvm_kernel_irqchip_split(void)
3733 {
3734     return kvm_state->kernel_irqchip_split == ON_OFF_AUTO_ON;
3735 }
3736 
3737 static void kvm_get_dirty_ring_size(Object *obj, Visitor *v,
3738                                     const char *name, void *opaque,
3739                                     Error **errp)
3740 {
3741     KVMState *s = KVM_STATE(obj);
3742     uint32_t value = s->kvm_dirty_ring_size;
3743 
3744     visit_type_uint32(v, name, &value, errp);
3745 }
3746 
3747 static void kvm_set_dirty_ring_size(Object *obj, Visitor *v,
3748                                     const char *name, void *opaque,
3749                                     Error **errp)
3750 {
3751     KVMState *s = KVM_STATE(obj);
3752     uint32_t value;
3753 
3754     if (s->fd != -1) {
3755         error_setg(errp, "Cannot set properties after the accelerator has been initialized");
3756         return;
3757     }
3758 
3759     if (!visit_type_uint32(v, name, &value, errp)) {
3760         return;
3761     }
3762     if (value & (value - 1)) {
3763         error_setg(errp, "dirty-ring-size must be a power of two.");
3764         return;
3765     }
3766 
3767     s->kvm_dirty_ring_size = value;
3768 }
3769 
3770 static char *kvm_get_device(Object *obj,
3771                             Error **errp G_GNUC_UNUSED)
3772 {
3773     KVMState *s = KVM_STATE(obj);
3774 
3775     return g_strdup(s->device);
3776 }
3777 
3778 static void kvm_set_device(Object *obj,
3779                            const char *value,
3780                            Error **errp G_GNUC_UNUSED)
3781 {
3782     KVMState *s = KVM_STATE(obj);
3783 
3784     g_free(s->device);
3785     s->device = g_strdup(value);
3786 }
3787 
3788 static void kvm_set_kvm_rapl(Object *obj, bool value, Error **errp)
3789 {
3790     KVMState *s = KVM_STATE(obj);
3791     s->msr_energy.enable = value;
3792 }
3793 
3794 static void kvm_set_kvm_rapl_socket_path(Object *obj,
3795                                          const char *str,
3796                                          Error **errp)
3797 {
3798     KVMState *s = KVM_STATE(obj);
3799     g_free(s->msr_energy.socket_path);
3800     s->msr_energy.socket_path = g_strdup(str);
3801 }
3802 
3803 static void kvm_accel_instance_init(Object *obj)
3804 {
3805     KVMState *s = KVM_STATE(obj);
3806 
3807     s->fd = -1;
3808     s->vmfd = -1;
3809     s->kvm_shadow_mem = -1;
3810     s->kernel_irqchip_allowed = true;
3811     s->kernel_irqchip_split = ON_OFF_AUTO_AUTO;
3812     /* KVM dirty ring is by default off */
3813     s->kvm_dirty_ring_size = 0;
3814     s->kvm_dirty_ring_with_bitmap = false;
3815     s->kvm_eager_split_size = 0;
3816     s->notify_vmexit = NOTIFY_VMEXIT_OPTION_RUN;
3817     s->notify_window = 0;
3818     s->xen_version = 0;
3819     s->xen_gnttab_max_frames = 64;
3820     s->xen_evtchn_max_pirq = 256;
3821     s->device = NULL;
3822     s->msr_energy.enable = false;
3823 }
3824 
3825 /**
3826  * kvm_gdbstub_sstep_flags():
3827  *
3828  * Returns: SSTEP_* flags that KVM supports for guest debug. The
3829  * support is probed during kvm_init()
3830  */
3831 static int kvm_gdbstub_sstep_flags(void)
3832 {
3833     return kvm_sstep_flags;
3834 }
3835 
3836 static void kvm_accel_class_init(ObjectClass *oc, void *data)
3837 {
3838     AccelClass *ac = ACCEL_CLASS(oc);
3839     ac->name = "KVM";
3840     ac->init_machine = kvm_init;
3841     ac->has_memory = kvm_accel_has_memory;
3842     ac->allowed = &kvm_allowed;
3843     ac->gdbstub_supported_sstep_flags = kvm_gdbstub_sstep_flags;
3844 
3845     object_class_property_add(oc, "kernel-irqchip", "on|off|split",
3846         NULL, kvm_set_kernel_irqchip,
3847         NULL, NULL);
3848     object_class_property_set_description(oc, "kernel-irqchip",
3849         "Configure KVM in-kernel irqchip");
3850 
3851     object_class_property_add(oc, "kvm-shadow-mem", "int",
3852         kvm_get_kvm_shadow_mem, kvm_set_kvm_shadow_mem,
3853         NULL, NULL);
3854     object_class_property_set_description(oc, "kvm-shadow-mem",
3855         "KVM shadow MMU size");
3856 
3857     object_class_property_add(oc, "dirty-ring-size", "uint32",
3858         kvm_get_dirty_ring_size, kvm_set_dirty_ring_size,
3859         NULL, NULL);
3860     object_class_property_set_description(oc, "dirty-ring-size",
3861         "Size of KVM dirty page ring buffer (default: 0, i.e. use bitmap)");
3862 
3863     object_class_property_add_str(oc, "device", kvm_get_device, kvm_set_device);
3864     object_class_property_set_description(oc, "device",
3865         "Path to the device node to use (default: /dev/kvm)");
3866 
3867     object_class_property_add_bool(oc, "rapl",
3868                                    NULL,
3869                                    kvm_set_kvm_rapl);
3870     object_class_property_set_description(oc, "rapl",
3871         "Allow energy related MSRs for RAPL interface in Guest");
3872 
3873     object_class_property_add_str(oc, "rapl-helper-socket", NULL,
3874                                   kvm_set_kvm_rapl_socket_path);
3875     object_class_property_set_description(oc, "rapl-helper-socket",
3876         "Socket Path for comminucating with the Virtual MSR helper daemon");
3877 
3878     kvm_arch_accel_class_init(oc);
3879 }
3880 
3881 static const TypeInfo kvm_accel_type = {
3882     .name = TYPE_KVM_ACCEL,
3883     .parent = TYPE_ACCEL,
3884     .instance_init = kvm_accel_instance_init,
3885     .class_init = kvm_accel_class_init,
3886     .instance_size = sizeof(KVMState),
3887 };
3888 
3889 static void kvm_type_init(void)
3890 {
3891     type_register_static(&kvm_accel_type);
3892 }
3893 
3894 type_init(kvm_type_init);
3895 
3896 typedef struct StatsArgs {
3897     union StatsResultsType {
3898         StatsResultList **stats;
3899         StatsSchemaList **schema;
3900     } result;
3901     strList *names;
3902     Error **errp;
3903 } StatsArgs;
3904 
3905 static StatsList *add_kvmstat_entry(struct kvm_stats_desc *pdesc,
3906                                     uint64_t *stats_data,
3907                                     StatsList *stats_list,
3908                                     Error **errp)
3909 {
3910 
3911     Stats *stats;
3912     uint64List *val_list = NULL;
3913 
3914     /* Only add stats that we understand.  */
3915     switch (pdesc->flags & KVM_STATS_TYPE_MASK) {
3916     case KVM_STATS_TYPE_CUMULATIVE:
3917     case KVM_STATS_TYPE_INSTANT:
3918     case KVM_STATS_TYPE_PEAK:
3919     case KVM_STATS_TYPE_LINEAR_HIST:
3920     case KVM_STATS_TYPE_LOG_HIST:
3921         break;
3922     default:
3923         return stats_list;
3924     }
3925 
3926     switch (pdesc->flags & KVM_STATS_UNIT_MASK) {
3927     case KVM_STATS_UNIT_NONE:
3928     case KVM_STATS_UNIT_BYTES:
3929     case KVM_STATS_UNIT_CYCLES:
3930     case KVM_STATS_UNIT_SECONDS:
3931     case KVM_STATS_UNIT_BOOLEAN:
3932         break;
3933     default:
3934         return stats_list;
3935     }
3936 
3937     switch (pdesc->flags & KVM_STATS_BASE_MASK) {
3938     case KVM_STATS_BASE_POW10:
3939     case KVM_STATS_BASE_POW2:
3940         break;
3941     default:
3942         return stats_list;
3943     }
3944 
3945     /* Alloc and populate data list */
3946     stats = g_new0(Stats, 1);
3947     stats->name = g_strdup(pdesc->name);
3948     stats->value = g_new0(StatsValue, 1);
3949 
3950     if ((pdesc->flags & KVM_STATS_UNIT_MASK) == KVM_STATS_UNIT_BOOLEAN) {
3951         stats->value->u.boolean = *stats_data;
3952         stats->value->type = QTYPE_QBOOL;
3953     } else if (pdesc->size == 1) {
3954         stats->value->u.scalar = *stats_data;
3955         stats->value->type = QTYPE_QNUM;
3956     } else {
3957         int i;
3958         for (i = 0; i < pdesc->size; i++) {
3959             QAPI_LIST_PREPEND(val_list, stats_data[i]);
3960         }
3961         stats->value->u.list = val_list;
3962         stats->value->type = QTYPE_QLIST;
3963     }
3964 
3965     QAPI_LIST_PREPEND(stats_list, stats);
3966     return stats_list;
3967 }
3968 
3969 static StatsSchemaValueList *add_kvmschema_entry(struct kvm_stats_desc *pdesc,
3970                                                  StatsSchemaValueList *list,
3971                                                  Error **errp)
3972 {
3973     StatsSchemaValueList *schema_entry = g_new0(StatsSchemaValueList, 1);
3974     schema_entry->value = g_new0(StatsSchemaValue, 1);
3975 
3976     switch (pdesc->flags & KVM_STATS_TYPE_MASK) {
3977     case KVM_STATS_TYPE_CUMULATIVE:
3978         schema_entry->value->type = STATS_TYPE_CUMULATIVE;
3979         break;
3980     case KVM_STATS_TYPE_INSTANT:
3981         schema_entry->value->type = STATS_TYPE_INSTANT;
3982         break;
3983     case KVM_STATS_TYPE_PEAK:
3984         schema_entry->value->type = STATS_TYPE_PEAK;
3985         break;
3986     case KVM_STATS_TYPE_LINEAR_HIST:
3987         schema_entry->value->type = STATS_TYPE_LINEAR_HISTOGRAM;
3988         schema_entry->value->bucket_size = pdesc->bucket_size;
3989         schema_entry->value->has_bucket_size = true;
3990         break;
3991     case KVM_STATS_TYPE_LOG_HIST:
3992         schema_entry->value->type = STATS_TYPE_LOG2_HISTOGRAM;
3993         break;
3994     default:
3995         goto exit;
3996     }
3997 
3998     switch (pdesc->flags & KVM_STATS_UNIT_MASK) {
3999     case KVM_STATS_UNIT_NONE:
4000         break;
4001     case KVM_STATS_UNIT_BOOLEAN:
4002         schema_entry->value->has_unit = true;
4003         schema_entry->value->unit = STATS_UNIT_BOOLEAN;
4004         break;
4005     case KVM_STATS_UNIT_BYTES:
4006         schema_entry->value->has_unit = true;
4007         schema_entry->value->unit = STATS_UNIT_BYTES;
4008         break;
4009     case KVM_STATS_UNIT_CYCLES:
4010         schema_entry->value->has_unit = true;
4011         schema_entry->value->unit = STATS_UNIT_CYCLES;
4012         break;
4013     case KVM_STATS_UNIT_SECONDS:
4014         schema_entry->value->has_unit = true;
4015         schema_entry->value->unit = STATS_UNIT_SECONDS;
4016         break;
4017     default:
4018         goto exit;
4019     }
4020 
4021     schema_entry->value->exponent = pdesc->exponent;
4022     if (pdesc->exponent) {
4023         switch (pdesc->flags & KVM_STATS_BASE_MASK) {
4024         case KVM_STATS_BASE_POW10:
4025             schema_entry->value->has_base = true;
4026             schema_entry->value->base = 10;
4027             break;
4028         case KVM_STATS_BASE_POW2:
4029             schema_entry->value->has_base = true;
4030             schema_entry->value->base = 2;
4031             break;
4032         default:
4033             goto exit;
4034         }
4035     }
4036 
4037     schema_entry->value->name = g_strdup(pdesc->name);
4038     schema_entry->next = list;
4039     return schema_entry;
4040 exit:
4041     g_free(schema_entry->value);
4042     g_free(schema_entry);
4043     return list;
4044 }
4045 
4046 /* Cached stats descriptors */
4047 typedef struct StatsDescriptors {
4048     const char *ident; /* cache key, currently the StatsTarget */
4049     struct kvm_stats_desc *kvm_stats_desc;
4050     struct kvm_stats_header kvm_stats_header;
4051     QTAILQ_ENTRY(StatsDescriptors) next;
4052 } StatsDescriptors;
4053 
4054 static QTAILQ_HEAD(, StatsDescriptors) stats_descriptors =
4055     QTAILQ_HEAD_INITIALIZER(stats_descriptors);
4056 
4057 /*
4058  * Return the descriptors for 'target', that either have already been read
4059  * or are retrieved from 'stats_fd'.
4060  */
4061 static StatsDescriptors *find_stats_descriptors(StatsTarget target, int stats_fd,
4062                                                 Error **errp)
4063 {
4064     StatsDescriptors *descriptors;
4065     const char *ident;
4066     struct kvm_stats_desc *kvm_stats_desc;
4067     struct kvm_stats_header *kvm_stats_header;
4068     size_t size_desc;
4069     ssize_t ret;
4070 
4071     ident = StatsTarget_str(target);
4072     QTAILQ_FOREACH(descriptors, &stats_descriptors, next) {
4073         if (g_str_equal(descriptors->ident, ident)) {
4074             return descriptors;
4075         }
4076     }
4077 
4078     descriptors = g_new0(StatsDescriptors, 1);
4079 
4080     /* Read stats header */
4081     kvm_stats_header = &descriptors->kvm_stats_header;
4082     ret = pread(stats_fd, kvm_stats_header, sizeof(*kvm_stats_header), 0);
4083     if (ret != sizeof(*kvm_stats_header)) {
4084         error_setg(errp, "KVM stats: failed to read stats header: "
4085                    "expected %zu actual %zu",
4086                    sizeof(*kvm_stats_header), ret);
4087         g_free(descriptors);
4088         return NULL;
4089     }
4090     size_desc = sizeof(*kvm_stats_desc) + kvm_stats_header->name_size;
4091 
4092     /* Read stats descriptors */
4093     kvm_stats_desc = g_malloc0_n(kvm_stats_header->num_desc, size_desc);
4094     ret = pread(stats_fd, kvm_stats_desc,
4095                 size_desc * kvm_stats_header->num_desc,
4096                 kvm_stats_header->desc_offset);
4097 
4098     if (ret != size_desc * kvm_stats_header->num_desc) {
4099         error_setg(errp, "KVM stats: failed to read stats descriptors: "
4100                    "expected %zu actual %zu",
4101                    size_desc * kvm_stats_header->num_desc, ret);
4102         g_free(descriptors);
4103         g_free(kvm_stats_desc);
4104         return NULL;
4105     }
4106     descriptors->kvm_stats_desc = kvm_stats_desc;
4107     descriptors->ident = ident;
4108     QTAILQ_INSERT_TAIL(&stats_descriptors, descriptors, next);
4109     return descriptors;
4110 }
4111 
4112 static void query_stats(StatsResultList **result, StatsTarget target,
4113                         strList *names, int stats_fd, CPUState *cpu,
4114                         Error **errp)
4115 {
4116     struct kvm_stats_desc *kvm_stats_desc;
4117     struct kvm_stats_header *kvm_stats_header;
4118     StatsDescriptors *descriptors;
4119     g_autofree uint64_t *stats_data = NULL;
4120     struct kvm_stats_desc *pdesc;
4121     StatsList *stats_list = NULL;
4122     size_t size_desc, size_data = 0;
4123     ssize_t ret;
4124     int i;
4125 
4126     descriptors = find_stats_descriptors(target, stats_fd, errp);
4127     if (!descriptors) {
4128         return;
4129     }
4130 
4131     kvm_stats_header = &descriptors->kvm_stats_header;
4132     kvm_stats_desc = descriptors->kvm_stats_desc;
4133     size_desc = sizeof(*kvm_stats_desc) + kvm_stats_header->name_size;
4134 
4135     /* Tally the total data size; read schema data */
4136     for (i = 0; i < kvm_stats_header->num_desc; ++i) {
4137         pdesc = (void *)kvm_stats_desc + i * size_desc;
4138         size_data += pdesc->size * sizeof(*stats_data);
4139     }
4140 
4141     stats_data = g_malloc0(size_data);
4142     ret = pread(stats_fd, stats_data, size_data, kvm_stats_header->data_offset);
4143 
4144     if (ret != size_data) {
4145         error_setg(errp, "KVM stats: failed to read data: "
4146                    "expected %zu actual %zu", size_data, ret);
4147         return;
4148     }
4149 
4150     for (i = 0; i < kvm_stats_header->num_desc; ++i) {
4151         uint64_t *stats;
4152         pdesc = (void *)kvm_stats_desc + i * size_desc;
4153 
4154         /* Add entry to the list */
4155         stats = (void *)stats_data + pdesc->offset;
4156         if (!apply_str_list_filter(pdesc->name, names)) {
4157             continue;
4158         }
4159         stats_list = add_kvmstat_entry(pdesc, stats, stats_list, errp);
4160     }
4161 
4162     if (!stats_list) {
4163         return;
4164     }
4165 
4166     switch (target) {
4167     case STATS_TARGET_VM:
4168         add_stats_entry(result, STATS_PROVIDER_KVM, NULL, stats_list);
4169         break;
4170     case STATS_TARGET_VCPU:
4171         add_stats_entry(result, STATS_PROVIDER_KVM,
4172                         cpu->parent_obj.canonical_path,
4173                         stats_list);
4174         break;
4175     default:
4176         g_assert_not_reached();
4177     }
4178 }
4179 
4180 static void query_stats_schema(StatsSchemaList **result, StatsTarget target,
4181                                int stats_fd, Error **errp)
4182 {
4183     struct kvm_stats_desc *kvm_stats_desc;
4184     struct kvm_stats_header *kvm_stats_header;
4185     StatsDescriptors *descriptors;
4186     struct kvm_stats_desc *pdesc;
4187     StatsSchemaValueList *stats_list = NULL;
4188     size_t size_desc;
4189     int i;
4190 
4191     descriptors = find_stats_descriptors(target, stats_fd, errp);
4192     if (!descriptors) {
4193         return;
4194     }
4195 
4196     kvm_stats_header = &descriptors->kvm_stats_header;
4197     kvm_stats_desc = descriptors->kvm_stats_desc;
4198     size_desc = sizeof(*kvm_stats_desc) + kvm_stats_header->name_size;
4199 
4200     /* Tally the total data size; read schema data */
4201     for (i = 0; i < kvm_stats_header->num_desc; ++i) {
4202         pdesc = (void *)kvm_stats_desc + i * size_desc;
4203         stats_list = add_kvmschema_entry(pdesc, stats_list, errp);
4204     }
4205 
4206     add_stats_schema(result, STATS_PROVIDER_KVM, target, stats_list);
4207 }
4208 
4209 static void query_stats_vcpu(CPUState *cpu, StatsArgs *kvm_stats_args)
4210 {
4211     int stats_fd = cpu->kvm_vcpu_stats_fd;
4212     Error *local_err = NULL;
4213 
4214     if (stats_fd == -1) {
4215         error_setg_errno(&local_err, errno, "KVM stats: ioctl failed");
4216         error_propagate(kvm_stats_args->errp, local_err);
4217         return;
4218     }
4219     query_stats(kvm_stats_args->result.stats, STATS_TARGET_VCPU,
4220                 kvm_stats_args->names, stats_fd, cpu,
4221                 kvm_stats_args->errp);
4222 }
4223 
4224 static void query_stats_schema_vcpu(CPUState *cpu, StatsArgs *kvm_stats_args)
4225 {
4226     int stats_fd = cpu->kvm_vcpu_stats_fd;
4227     Error *local_err = NULL;
4228 
4229     if (stats_fd == -1) {
4230         error_setg_errno(&local_err, errno, "KVM stats: ioctl failed");
4231         error_propagate(kvm_stats_args->errp, local_err);
4232         return;
4233     }
4234     query_stats_schema(kvm_stats_args->result.schema, STATS_TARGET_VCPU, stats_fd,
4235                        kvm_stats_args->errp);
4236 }
4237 
4238 static void query_stats_cb(StatsResultList **result, StatsTarget target,
4239                            strList *names, strList *targets, Error **errp)
4240 {
4241     KVMState *s = kvm_state;
4242     CPUState *cpu;
4243     int stats_fd;
4244 
4245     switch (target) {
4246     case STATS_TARGET_VM:
4247     {
4248         stats_fd = kvm_vm_ioctl(s, KVM_GET_STATS_FD, NULL);
4249         if (stats_fd == -1) {
4250             error_setg_errno(errp, errno, "KVM stats: ioctl failed");
4251             return;
4252         }
4253         query_stats(result, target, names, stats_fd, NULL, errp);
4254         close(stats_fd);
4255         break;
4256     }
4257     case STATS_TARGET_VCPU:
4258     {
4259         StatsArgs stats_args;
4260         stats_args.result.stats = result;
4261         stats_args.names = names;
4262         stats_args.errp = errp;
4263         CPU_FOREACH(cpu) {
4264             if (!apply_str_list_filter(cpu->parent_obj.canonical_path, targets)) {
4265                 continue;
4266             }
4267             query_stats_vcpu(cpu, &stats_args);
4268         }
4269         break;
4270     }
4271     default:
4272         break;
4273     }
4274 }
4275 
4276 void query_stats_schemas_cb(StatsSchemaList **result, Error **errp)
4277 {
4278     StatsArgs stats_args;
4279     KVMState *s = kvm_state;
4280     int stats_fd;
4281 
4282     stats_fd = kvm_vm_ioctl(s, KVM_GET_STATS_FD, NULL);
4283     if (stats_fd == -1) {
4284         error_setg_errno(errp, errno, "KVM stats: ioctl failed");
4285         return;
4286     }
4287     query_stats_schema(result, STATS_TARGET_VM, stats_fd, errp);
4288     close(stats_fd);
4289 
4290     if (first_cpu) {
4291         stats_args.result.schema = result;
4292         stats_args.errp = errp;
4293         query_stats_schema_vcpu(first_cpu, &stats_args);
4294     }
4295 }
4296 
4297 void kvm_mark_guest_state_protected(void)
4298 {
4299     kvm_state->guest_state_protected = true;
4300 }
4301 
4302 int kvm_create_guest_memfd(uint64_t size, uint64_t flags, Error **errp)
4303 {
4304     int fd;
4305     struct kvm_create_guest_memfd guest_memfd = {
4306         .size = size,
4307         .flags = flags,
4308     };
4309 
4310     if (!kvm_guest_memfd_supported) {
4311         error_setg(errp, "KVM does not support guest_memfd");
4312         return -1;
4313     }
4314 
4315     fd = kvm_vm_ioctl(kvm_state, KVM_CREATE_GUEST_MEMFD, &guest_memfd);
4316     if (fd < 0) {
4317         error_setg_errno(errp, errno, "Error creating KVM guest_memfd");
4318         return -1;
4319     }
4320 
4321     return fd;
4322 }
4323