trap signals for "-serial mon:stdio"
[qemu.git] / kvm-all.c
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 <sys/types.h>
17 #include <sys/ioctl.h>
18 #include <sys/mman.h>
19 #include <stdarg.h>
20
21 #include <linux/kvm.h>
22
23 #include "qemu-common.h"
24 #include "qemu/atomic.h"
25 #include "qemu/option.h"
26 #include "qemu/config-file.h"
27 #include "sysemu/sysemu.h"
28 #include "hw/hw.h"
29 #include "hw/pci/msi.h"
30 #include "exec/gdbstub.h"
31 #include "sysemu/kvm.h"
32 #include "qemu/bswap.h"
33 #include "exec/memory.h"
34 #include "exec/address-spaces.h"
35 #include "qemu/event_notifier.h"
36 #include "trace.h"
37
38 /* This check must be after config-host.h is included */
39 #ifdef CONFIG_EVENTFD
40 #include <sys/eventfd.h>
41 #endif
42
43 #ifdef CONFIG_VALGRIND_H
44 #include <valgrind/memcheck.h>
45 #endif
46
47 /* KVM uses PAGE_SIZE in its definition of COALESCED_MMIO_MAX */
48 #define PAGE_SIZE TARGET_PAGE_SIZE
49
50 //#define DEBUG_KVM
51
52 #ifdef DEBUG_KVM
53 #define DPRINTF(fmt, ...) \
54 do { fprintf(stderr, fmt, ## __VA_ARGS__); } while (0)
55 #else
56 #define DPRINTF(fmt, ...) \
57 do { } while (0)
58 #endif
59
60 #define KVM_MSI_HASHTAB_SIZE 256
61
62 typedef struct KVMSlot
63 {
64 hwaddr start_addr;
65 ram_addr_t memory_size;
66 void *ram;
67 int slot;
68 int flags;
69 } KVMSlot;
70
71 typedef struct kvm_dirty_log KVMDirtyLog;
72
73 struct KVMState
74 {
75 KVMSlot slots[32];
76 int fd;
77 int vmfd;
78 int coalesced_mmio;
79 struct kvm_coalesced_mmio_ring *coalesced_mmio_ring;
80 bool coalesced_flush_in_progress;
81 int broken_set_mem_region;
82 int migration_log;
83 int vcpu_events;
84 int robust_singlestep;
85 int debugregs;
86 #ifdef KVM_CAP_SET_GUEST_DEBUG
87 struct kvm_sw_breakpoint_head kvm_sw_breakpoints;
88 #endif
89 int pit_state2;
90 int xsave, xcrs;
91 int many_ioeventfds;
92 int intx_set_mask;
93 /* The man page (and posix) say ioctl numbers are signed int, but
94 * they're not. Linux, glibc and *BSD all treat ioctl numbers as
95 * unsigned, and treating them as signed here can break things */
96 unsigned irq_set_ioctl;
97 #ifdef KVM_CAP_IRQ_ROUTING
98 struct kvm_irq_routing *irq_routes;
99 int nr_allocated_irq_routes;
100 uint32_t *used_gsi_bitmap;
101 unsigned int gsi_count;
102 QTAILQ_HEAD(msi_hashtab, KVMMSIRoute) msi_hashtab[KVM_MSI_HASHTAB_SIZE];
103 bool direct_msi;
104 #endif
105 };
106
107 KVMState *kvm_state;
108 bool kvm_kernel_irqchip;
109 bool kvm_async_interrupts_allowed;
110 bool kvm_halt_in_kernel_allowed;
111 bool kvm_irqfds_allowed;
112 bool kvm_msi_via_irqfd_allowed;
113 bool kvm_gsi_routing_allowed;
114 bool kvm_allowed;
115 bool kvm_readonly_mem_allowed;
116
117 static const KVMCapabilityInfo kvm_required_capabilites[] = {
118 KVM_CAP_INFO(USER_MEMORY),
119 KVM_CAP_INFO(DESTROY_MEMORY_REGION_WORKS),
120 KVM_CAP_LAST_INFO
121 };
122
123 static KVMSlot *kvm_alloc_slot(KVMState *s)
124 {
125 int i;
126
127 for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
128 if (s->slots[i].memory_size == 0) {
129 return &s->slots[i];
130 }
131 }
132
133 fprintf(stderr, "%s: no free slot available\n", __func__);
134 abort();
135 }
136
137 static KVMSlot *kvm_lookup_matching_slot(KVMState *s,
138 hwaddr start_addr,
139 hwaddr end_addr)
140 {
141 int i;
142
143 for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
144 KVMSlot *mem = &s->slots[i];
145
146 if (start_addr == mem->start_addr &&
147 end_addr == mem->start_addr + mem->memory_size) {
148 return mem;
149 }
150 }
151
152 return NULL;
153 }
154
155 /*
156 * Find overlapping slot with lowest start address
157 */
158 static KVMSlot *kvm_lookup_overlapping_slot(KVMState *s,
159 hwaddr start_addr,
160 hwaddr end_addr)
161 {
162 KVMSlot *found = NULL;
163 int i;
164
165 for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
166 KVMSlot *mem = &s->slots[i];
167
168 if (mem->memory_size == 0 ||
169 (found && found->start_addr < mem->start_addr)) {
170 continue;
171 }
172
173 if (end_addr > mem->start_addr &&
174 start_addr < mem->start_addr + mem->memory_size) {
175 found = mem;
176 }
177 }
178
179 return found;
180 }
181
182 int kvm_physical_memory_addr_from_host(KVMState *s, void *ram,
183 hwaddr *phys_addr)
184 {
185 int i;
186
187 for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
188 KVMSlot *mem = &s->slots[i];
189
190 if (ram >= mem->ram && ram < mem->ram + mem->memory_size) {
191 *phys_addr = mem->start_addr + (ram - mem->ram);
192 return 1;
193 }
194 }
195
196 return 0;
197 }
198
199 static int kvm_set_user_memory_region(KVMState *s, KVMSlot *slot)
200 {
201 struct kvm_userspace_memory_region mem;
202
203 mem.slot = slot->slot;
204 mem.guest_phys_addr = slot->start_addr;
205 mem.userspace_addr = (unsigned long)slot->ram;
206 mem.flags = slot->flags;
207 if (s->migration_log) {
208 mem.flags |= KVM_MEM_LOG_DIRTY_PAGES;
209 }
210
211 if (slot->memory_size && mem.flags & KVM_MEM_READONLY) {
212 /* Set the slot size to 0 before setting the slot to the desired
213 * value. This is needed based on KVM commit 75d61fbc. */
214 mem.memory_size = 0;
215 kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
216 }
217 mem.memory_size = slot->memory_size;
218 return kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
219 }
220
221 static void kvm_reset_vcpu(void *opaque)
222 {
223 CPUState *cpu = opaque;
224
225 kvm_arch_reset_vcpu(cpu);
226 }
227
228 int kvm_init_vcpu(CPUState *cpu)
229 {
230 KVMState *s = kvm_state;
231 long mmap_size;
232 int ret;
233
234 DPRINTF("kvm_init_vcpu\n");
235
236 ret = kvm_vm_ioctl(s, KVM_CREATE_VCPU, (void *)kvm_arch_vcpu_id(cpu));
237 if (ret < 0) {
238 DPRINTF("kvm_create_vcpu failed\n");
239 goto err;
240 }
241
242 cpu->kvm_fd = ret;
243 cpu->kvm_state = s;
244 cpu->kvm_vcpu_dirty = true;
245
246 mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
247 if (mmap_size < 0) {
248 ret = mmap_size;
249 DPRINTF("KVM_GET_VCPU_MMAP_SIZE failed\n");
250 goto err;
251 }
252
253 cpu->kvm_run = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
254 cpu->kvm_fd, 0);
255 if (cpu->kvm_run == MAP_FAILED) {
256 ret = -errno;
257 DPRINTF("mmap'ing vcpu state failed\n");
258 goto err;
259 }
260
261 if (s->coalesced_mmio && !s->coalesced_mmio_ring) {
262 s->coalesced_mmio_ring =
263 (void *)cpu->kvm_run + s->coalesced_mmio * PAGE_SIZE;
264 }
265
266 ret = kvm_arch_init_vcpu(cpu);
267 if (ret == 0) {
268 qemu_register_reset(kvm_reset_vcpu, cpu);
269 kvm_arch_reset_vcpu(cpu);
270 }
271 err:
272 return ret;
273 }
274
275 /*
276 * dirty pages logging control
277 */
278
279 static int kvm_mem_flags(KVMState *s, bool log_dirty, bool readonly)
280 {
281 int flags = 0;
282 flags = log_dirty ? KVM_MEM_LOG_DIRTY_PAGES : 0;
283 if (readonly && kvm_readonly_mem_allowed) {
284 flags |= KVM_MEM_READONLY;
285 }
286 return flags;
287 }
288
289 static int kvm_slot_dirty_pages_log_change(KVMSlot *mem, bool log_dirty)
290 {
291 KVMState *s = kvm_state;
292 int flags, mask = KVM_MEM_LOG_DIRTY_PAGES;
293 int old_flags;
294
295 old_flags = mem->flags;
296
297 flags = (mem->flags & ~mask) | kvm_mem_flags(s, log_dirty, false);
298 mem->flags = flags;
299
300 /* If nothing changed effectively, no need to issue ioctl */
301 if (s->migration_log) {
302 flags |= KVM_MEM_LOG_DIRTY_PAGES;
303 }
304
305 if (flags == old_flags) {
306 return 0;
307 }
308
309 return kvm_set_user_memory_region(s, mem);
310 }
311
312 static int kvm_dirty_pages_log_change(hwaddr phys_addr,
313 ram_addr_t size, bool log_dirty)
314 {
315 KVMState *s = kvm_state;
316 KVMSlot *mem = kvm_lookup_matching_slot(s, phys_addr, phys_addr + size);
317
318 if (mem == NULL) {
319 fprintf(stderr, "BUG: %s: invalid parameters " TARGET_FMT_plx "-"
320 TARGET_FMT_plx "\n", __func__, phys_addr,
321 (hwaddr)(phys_addr + size - 1));
322 return -EINVAL;
323 }
324 return kvm_slot_dirty_pages_log_change(mem, log_dirty);
325 }
326
327 static void kvm_log_start(MemoryListener *listener,
328 MemoryRegionSection *section)
329 {
330 int r;
331
332 r = kvm_dirty_pages_log_change(section->offset_within_address_space,
333 int128_get64(section->size), true);
334 if (r < 0) {
335 abort();
336 }
337 }
338
339 static void kvm_log_stop(MemoryListener *listener,
340 MemoryRegionSection *section)
341 {
342 int r;
343
344 r = kvm_dirty_pages_log_change(section->offset_within_address_space,
345 int128_get64(section->size), false);
346 if (r < 0) {
347 abort();
348 }
349 }
350
351 static int kvm_set_migration_log(int enable)
352 {
353 KVMState *s = kvm_state;
354 KVMSlot *mem;
355 int i, err;
356
357 s->migration_log = enable;
358
359 for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
360 mem = &s->slots[i];
361
362 if (!mem->memory_size) {
363 continue;
364 }
365 if (!!(mem->flags & KVM_MEM_LOG_DIRTY_PAGES) == enable) {
366 continue;
367 }
368 err = kvm_set_user_memory_region(s, mem);
369 if (err) {
370 return err;
371 }
372 }
373 return 0;
374 }
375
376 /* get kvm's dirty pages bitmap and update qemu's */
377 static int kvm_get_dirty_pages_log_range(MemoryRegionSection *section,
378 unsigned long *bitmap)
379 {
380 unsigned int i, j;
381 unsigned long page_number, c;
382 hwaddr addr, addr1;
383 unsigned int pages = int128_get64(section->size) / getpagesize();
384 unsigned int len = (pages + HOST_LONG_BITS - 1) / HOST_LONG_BITS;
385 unsigned long hpratio = getpagesize() / TARGET_PAGE_SIZE;
386
387 /*
388 * bitmap-traveling is faster than memory-traveling (for addr...)
389 * especially when most of the memory is not dirty.
390 */
391 for (i = 0; i < len; i++) {
392 if (bitmap[i] != 0) {
393 c = leul_to_cpu(bitmap[i]);
394 do {
395 j = ffsl(c) - 1;
396 c &= ~(1ul << j);
397 page_number = (i * HOST_LONG_BITS + j) * hpratio;
398 addr1 = page_number * TARGET_PAGE_SIZE;
399 addr = section->offset_within_region + addr1;
400 memory_region_set_dirty(section->mr, addr,
401 TARGET_PAGE_SIZE * hpratio);
402 } while (c != 0);
403 }
404 }
405 return 0;
406 }
407
408 #define ALIGN(x, y) (((x)+(y)-1) & ~((y)-1))
409
410 /**
411 * kvm_physical_sync_dirty_bitmap - Grab dirty bitmap from kernel space
412 * This function updates qemu's dirty bitmap using
413 * memory_region_set_dirty(). This means all bits are set
414 * to dirty.
415 *
416 * @start_add: start of logged region.
417 * @end_addr: end of logged region.
418 */
419 static int kvm_physical_sync_dirty_bitmap(MemoryRegionSection *section)
420 {
421 KVMState *s = kvm_state;
422 unsigned long size, allocated_size = 0;
423 KVMDirtyLog d;
424 KVMSlot *mem;
425 int ret = 0;
426 hwaddr start_addr = section->offset_within_address_space;
427 hwaddr end_addr = start_addr + int128_get64(section->size);
428
429 d.dirty_bitmap = NULL;
430 while (start_addr < end_addr) {
431 mem = kvm_lookup_overlapping_slot(s, start_addr, end_addr);
432 if (mem == NULL) {
433 break;
434 }
435
436 /* XXX bad kernel interface alert
437 * For dirty bitmap, kernel allocates array of size aligned to
438 * bits-per-long. But for case when the kernel is 64bits and
439 * the userspace is 32bits, userspace can't align to the same
440 * bits-per-long, since sizeof(long) is different between kernel
441 * and user space. This way, userspace will provide buffer which
442 * may be 4 bytes less than the kernel will use, resulting in
443 * userspace memory corruption (which is not detectable by valgrind
444 * too, in most cases).
445 * So for now, let's align to 64 instead of HOST_LONG_BITS here, in
446 * a hope that sizeof(long) wont become >8 any time soon.
447 */
448 size = ALIGN(((mem->memory_size) >> TARGET_PAGE_BITS),
449 /*HOST_LONG_BITS*/ 64) / 8;
450 if (!d.dirty_bitmap) {
451 d.dirty_bitmap = g_malloc(size);
452 } else if (size > allocated_size) {
453 d.dirty_bitmap = g_realloc(d.dirty_bitmap, size);
454 }
455 allocated_size = size;
456 memset(d.dirty_bitmap, 0, allocated_size);
457
458 d.slot = mem->slot;
459
460 if (kvm_vm_ioctl(s, KVM_GET_DIRTY_LOG, &d) == -1) {
461 DPRINTF("ioctl failed %d\n", errno);
462 ret = -1;
463 break;
464 }
465
466 kvm_get_dirty_pages_log_range(section, d.dirty_bitmap);
467 start_addr = mem->start_addr + mem->memory_size;
468 }
469 g_free(d.dirty_bitmap);
470
471 return ret;
472 }
473
474 static void kvm_coalesce_mmio_region(MemoryListener *listener,
475 MemoryRegionSection *secion,
476 hwaddr start, hwaddr size)
477 {
478 KVMState *s = kvm_state;
479
480 if (s->coalesced_mmio) {
481 struct kvm_coalesced_mmio_zone zone;
482
483 zone.addr = start;
484 zone.size = size;
485 zone.pad = 0;
486
487 (void)kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone);
488 }
489 }
490
491 static void kvm_uncoalesce_mmio_region(MemoryListener *listener,
492 MemoryRegionSection *secion,
493 hwaddr start, hwaddr size)
494 {
495 KVMState *s = kvm_state;
496
497 if (s->coalesced_mmio) {
498 struct kvm_coalesced_mmio_zone zone;
499
500 zone.addr = start;
501 zone.size = size;
502 zone.pad = 0;
503
504 (void)kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone);
505 }
506 }
507
508 int kvm_check_extension(KVMState *s, unsigned int extension)
509 {
510 int ret;
511
512 ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, extension);
513 if (ret < 0) {
514 ret = 0;
515 }
516
517 return ret;
518 }
519
520 static int kvm_set_ioeventfd_mmio(int fd, uint32_t addr, uint32_t val,
521 bool assign, uint32_t size, bool datamatch)
522 {
523 int ret;
524 struct kvm_ioeventfd iofd;
525
526 iofd.datamatch = datamatch ? val : 0;
527 iofd.addr = addr;
528 iofd.len = size;
529 iofd.flags = 0;
530 iofd.fd = fd;
531
532 if (!kvm_enabled()) {
533 return -ENOSYS;
534 }
535
536 if (datamatch) {
537 iofd.flags |= KVM_IOEVENTFD_FLAG_DATAMATCH;
538 }
539 if (!assign) {
540 iofd.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
541 }
542
543 ret = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &iofd);
544
545 if (ret < 0) {
546 return -errno;
547 }
548
549 return 0;
550 }
551
552 static int kvm_set_ioeventfd_pio(int fd, uint16_t addr, uint16_t val,
553 bool assign, uint32_t size, bool datamatch)
554 {
555 struct kvm_ioeventfd kick = {
556 .datamatch = datamatch ? val : 0,
557 .addr = addr,
558 .flags = KVM_IOEVENTFD_FLAG_PIO,
559 .len = size,
560 .fd = fd,
561 };
562 int r;
563 if (!kvm_enabled()) {
564 return -ENOSYS;
565 }
566 if (datamatch) {
567 kick.flags |= KVM_IOEVENTFD_FLAG_DATAMATCH;
568 }
569 if (!assign) {
570 kick.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
571 }
572 r = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &kick);
573 if (r < 0) {
574 return r;
575 }
576 return 0;
577 }
578
579
580 static int kvm_check_many_ioeventfds(void)
581 {
582 /* Userspace can use ioeventfd for io notification. This requires a host
583 * that supports eventfd(2) and an I/O thread; since eventfd does not
584 * support SIGIO it cannot interrupt the vcpu.
585 *
586 * Older kernels have a 6 device limit on the KVM io bus. Find out so we
587 * can avoid creating too many ioeventfds.
588 */
589 #if defined(CONFIG_EVENTFD)
590 int ioeventfds[7];
591 int i, ret = 0;
592 for (i = 0; i < ARRAY_SIZE(ioeventfds); i++) {
593 ioeventfds[i] = eventfd(0, EFD_CLOEXEC);
594 if (ioeventfds[i] < 0) {
595 break;
596 }
597 ret = kvm_set_ioeventfd_pio(ioeventfds[i], 0, i, true, 2, true);
598 if (ret < 0) {
599 close(ioeventfds[i]);
600 break;
601 }
602 }
603
604 /* Decide whether many devices are supported or not */
605 ret = i == ARRAY_SIZE(ioeventfds);
606
607 while (i-- > 0) {
608 kvm_set_ioeventfd_pio(ioeventfds[i], 0, i, false, 2, true);
609 close(ioeventfds[i]);
610 }
611 return ret;
612 #else
613 return 0;
614 #endif
615 }
616
617 static const KVMCapabilityInfo *
618 kvm_check_extension_list(KVMState *s, const KVMCapabilityInfo *list)
619 {
620 while (list->name) {
621 if (!kvm_check_extension(s, list->value)) {
622 return list;
623 }
624 list++;
625 }
626 return NULL;
627 }
628
629 static void kvm_set_phys_mem(MemoryRegionSection *section, bool add)
630 {
631 KVMState *s = kvm_state;
632 KVMSlot *mem, old;
633 int err;
634 MemoryRegion *mr = section->mr;
635 bool log_dirty = memory_region_is_logging(mr);
636 bool writeable = !mr->readonly && !mr->rom_device;
637 bool readonly_flag = mr->readonly || memory_region_is_romd(mr);
638 hwaddr start_addr = section->offset_within_address_space;
639 ram_addr_t size = int128_get64(section->size);
640 void *ram = NULL;
641 unsigned delta;
642
643 /* kvm works in page size chunks, but the function may be called
644 with sub-page size and unaligned start address. */
645 delta = TARGET_PAGE_ALIGN(size) - size;
646 if (delta > size) {
647 return;
648 }
649 start_addr += delta;
650 size -= delta;
651 size &= TARGET_PAGE_MASK;
652 if (!size || (start_addr & ~TARGET_PAGE_MASK)) {
653 return;
654 }
655
656 if (!memory_region_is_ram(mr)) {
657 if (writeable || !kvm_readonly_mem_allowed) {
658 return;
659 } else if (!mr->romd_mode) {
660 /* If the memory device is not in romd_mode, then we actually want
661 * to remove the kvm memory slot so all accesses will trap. */
662 add = false;
663 }
664 }
665
666 ram = memory_region_get_ram_ptr(mr) + section->offset_within_region + delta;
667
668 while (1) {
669 mem = kvm_lookup_overlapping_slot(s, start_addr, start_addr + size);
670 if (!mem) {
671 break;
672 }
673
674 if (add && start_addr >= mem->start_addr &&
675 (start_addr + size <= mem->start_addr + mem->memory_size) &&
676 (ram - start_addr == mem->ram - mem->start_addr)) {
677 /* The new slot fits into the existing one and comes with
678 * identical parameters - update flags and done. */
679 kvm_slot_dirty_pages_log_change(mem, log_dirty);
680 return;
681 }
682
683 old = *mem;
684
685 if (mem->flags & KVM_MEM_LOG_DIRTY_PAGES) {
686 kvm_physical_sync_dirty_bitmap(section);
687 }
688
689 /* unregister the overlapping slot */
690 mem->memory_size = 0;
691 err = kvm_set_user_memory_region(s, mem);
692 if (err) {
693 fprintf(stderr, "%s: error unregistering overlapping slot: %s\n",
694 __func__, strerror(-err));
695 abort();
696 }
697
698 /* Workaround for older KVM versions: we can't join slots, even not by
699 * unregistering the previous ones and then registering the larger
700 * slot. We have to maintain the existing fragmentation. Sigh.
701 *
702 * This workaround assumes that the new slot starts at the same
703 * address as the first existing one. If not or if some overlapping
704 * slot comes around later, we will fail (not seen in practice so far)
705 * - and actually require a recent KVM version. */
706 if (s->broken_set_mem_region &&
707 old.start_addr == start_addr && old.memory_size < size && add) {
708 mem = kvm_alloc_slot(s);
709 mem->memory_size = old.memory_size;
710 mem->start_addr = old.start_addr;
711 mem->ram = old.ram;
712 mem->flags = kvm_mem_flags(s, log_dirty, readonly_flag);
713
714 err = kvm_set_user_memory_region(s, mem);
715 if (err) {
716 fprintf(stderr, "%s: error updating slot: %s\n", __func__,
717 strerror(-err));
718 abort();
719 }
720
721 start_addr += old.memory_size;
722 ram += old.memory_size;
723 size -= old.memory_size;
724 continue;
725 }
726
727 /* register prefix slot */
728 if (old.start_addr < start_addr) {
729 mem = kvm_alloc_slot(s);
730 mem->memory_size = start_addr - old.start_addr;
731 mem->start_addr = old.start_addr;
732 mem->ram = old.ram;
733 mem->flags = kvm_mem_flags(s, log_dirty, readonly_flag);
734
735 err = kvm_set_user_memory_region(s, mem);
736 if (err) {
737 fprintf(stderr, "%s: error registering prefix slot: %s\n",
738 __func__, strerror(-err));
739 #ifdef TARGET_PPC
740 fprintf(stderr, "%s: This is probably because your kernel's " \
741 "PAGE_SIZE is too big. Please try to use 4k " \
742 "PAGE_SIZE!\n", __func__);
743 #endif
744 abort();
745 }
746 }
747
748 /* register suffix slot */
749 if (old.start_addr + old.memory_size > start_addr + size) {
750 ram_addr_t size_delta;
751
752 mem = kvm_alloc_slot(s);
753 mem->start_addr = start_addr + size;
754 size_delta = mem->start_addr - old.start_addr;
755 mem->memory_size = old.memory_size - size_delta;
756 mem->ram = old.ram + size_delta;
757 mem->flags = kvm_mem_flags(s, log_dirty, readonly_flag);
758
759 err = kvm_set_user_memory_region(s, mem);
760 if (err) {
761 fprintf(stderr, "%s: error registering suffix slot: %s\n",
762 __func__, strerror(-err));
763 abort();
764 }
765 }
766 }
767
768 /* in case the KVM bug workaround already "consumed" the new slot */
769 if (!size) {
770 return;
771 }
772 if (!add) {
773 return;
774 }
775 mem = kvm_alloc_slot(s);
776 mem->memory_size = size;
777 mem->start_addr = start_addr;
778 mem->ram = ram;
779 mem->flags = kvm_mem_flags(s, log_dirty, readonly_flag);
780
781 err = kvm_set_user_memory_region(s, mem);
782 if (err) {
783 fprintf(stderr, "%s: error registering slot: %s\n", __func__,
784 strerror(-err));
785 abort();
786 }
787 }
788
789 static void kvm_region_add(MemoryListener *listener,
790 MemoryRegionSection *section)
791 {
792 memory_region_ref(section->mr);
793 kvm_set_phys_mem(section, true);
794 }
795
796 static void kvm_region_del(MemoryListener *listener,
797 MemoryRegionSection *section)
798 {
799 kvm_set_phys_mem(section, false);
800 memory_region_unref(section->mr);
801 }
802
803 static void kvm_log_sync(MemoryListener *listener,
804 MemoryRegionSection *section)
805 {
806 int r;
807
808 r = kvm_physical_sync_dirty_bitmap(section);
809 if (r < 0) {
810 abort();
811 }
812 }
813
814 static void kvm_log_global_start(struct MemoryListener *listener)
815 {
816 int r;
817
818 r = kvm_set_migration_log(1);
819 assert(r >= 0);
820 }
821
822 static void kvm_log_global_stop(struct MemoryListener *listener)
823 {
824 int r;
825
826 r = kvm_set_migration_log(0);
827 assert(r >= 0);
828 }
829
830 static void kvm_mem_ioeventfd_add(MemoryListener *listener,
831 MemoryRegionSection *section,
832 bool match_data, uint64_t data,
833 EventNotifier *e)
834 {
835 int fd = event_notifier_get_fd(e);
836 int r;
837
838 r = kvm_set_ioeventfd_mmio(fd, section->offset_within_address_space,
839 data, true, int128_get64(section->size),
840 match_data);
841 if (r < 0) {
842 fprintf(stderr, "%s: error adding ioeventfd: %s\n",
843 __func__, strerror(-r));
844 abort();
845 }
846 }
847
848 static void kvm_mem_ioeventfd_del(MemoryListener *listener,
849 MemoryRegionSection *section,
850 bool match_data, uint64_t data,
851 EventNotifier *e)
852 {
853 int fd = event_notifier_get_fd(e);
854 int r;
855
856 r = kvm_set_ioeventfd_mmio(fd, section->offset_within_address_space,
857 data, false, int128_get64(section->size),
858 match_data);
859 if (r < 0) {
860 abort();
861 }
862 }
863
864 static void kvm_io_ioeventfd_add(MemoryListener *listener,
865 MemoryRegionSection *section,
866 bool match_data, uint64_t data,
867 EventNotifier *e)
868 {
869 int fd = event_notifier_get_fd(e);
870 int r;
871
872 r = kvm_set_ioeventfd_pio(fd, section->offset_within_address_space,
873 data, true, int128_get64(section->size),
874 match_data);
875 if (r < 0) {
876 fprintf(stderr, "%s: error adding ioeventfd: %s\n",
877 __func__, strerror(-r));
878 abort();
879 }
880 }
881
882 static void kvm_io_ioeventfd_del(MemoryListener *listener,
883 MemoryRegionSection *section,
884 bool match_data, uint64_t data,
885 EventNotifier *e)
886
887 {
888 int fd = event_notifier_get_fd(e);
889 int r;
890
891 r = kvm_set_ioeventfd_pio(fd, section->offset_within_address_space,
892 data, false, int128_get64(section->size),
893 match_data);
894 if (r < 0) {
895 abort();
896 }
897 }
898
899 static MemoryListener kvm_memory_listener = {
900 .region_add = kvm_region_add,
901 .region_del = kvm_region_del,
902 .log_start = kvm_log_start,
903 .log_stop = kvm_log_stop,
904 .log_sync = kvm_log_sync,
905 .log_global_start = kvm_log_global_start,
906 .log_global_stop = kvm_log_global_stop,
907 .eventfd_add = kvm_mem_ioeventfd_add,
908 .eventfd_del = kvm_mem_ioeventfd_del,
909 .coalesced_mmio_add = kvm_coalesce_mmio_region,
910 .coalesced_mmio_del = kvm_uncoalesce_mmio_region,
911 .priority = 10,
912 };
913
914 static MemoryListener kvm_io_listener = {
915 .eventfd_add = kvm_io_ioeventfd_add,
916 .eventfd_del = kvm_io_ioeventfd_del,
917 .priority = 10,
918 };
919
920 static void kvm_handle_interrupt(CPUState *cpu, int mask)
921 {
922 cpu->interrupt_request |= mask;
923
924 if (!qemu_cpu_is_self(cpu)) {
925 qemu_cpu_kick(cpu);
926 }
927 }
928
929 int kvm_set_irq(KVMState *s, int irq, int level)
930 {
931 struct kvm_irq_level event;
932 int ret;
933
934 assert(kvm_async_interrupts_enabled());
935
936 event.level = level;
937 event.irq = irq;
938 ret = kvm_vm_ioctl(s, s->irq_set_ioctl, &event);
939 if (ret < 0) {
940 perror("kvm_set_irq");
941 abort();
942 }
943
944 return (s->irq_set_ioctl == KVM_IRQ_LINE) ? 1 : event.status;
945 }
946
947 #ifdef KVM_CAP_IRQ_ROUTING
948 typedef struct KVMMSIRoute {
949 struct kvm_irq_routing_entry kroute;
950 QTAILQ_ENTRY(KVMMSIRoute) entry;
951 } KVMMSIRoute;
952
953 static void set_gsi(KVMState *s, unsigned int gsi)
954 {
955 s->used_gsi_bitmap[gsi / 32] |= 1U << (gsi % 32);
956 }
957
958 static void clear_gsi(KVMState *s, unsigned int gsi)
959 {
960 s->used_gsi_bitmap[gsi / 32] &= ~(1U << (gsi % 32));
961 }
962
963 void kvm_init_irq_routing(KVMState *s)
964 {
965 int gsi_count, i;
966
967 gsi_count = kvm_check_extension(s, KVM_CAP_IRQ_ROUTING);
968 if (gsi_count > 0) {
969 unsigned int gsi_bits, i;
970
971 /* Round up so we can search ints using ffs */
972 gsi_bits = ALIGN(gsi_count, 32);
973 s->used_gsi_bitmap = g_malloc0(gsi_bits / 8);
974 s->gsi_count = gsi_count;
975
976 /* Mark any over-allocated bits as already in use */
977 for (i = gsi_count; i < gsi_bits; i++) {
978 set_gsi(s, i);
979 }
980 }
981
982 s->irq_routes = g_malloc0(sizeof(*s->irq_routes));
983 s->nr_allocated_irq_routes = 0;
984
985 if (!s->direct_msi) {
986 for (i = 0; i < KVM_MSI_HASHTAB_SIZE; i++) {
987 QTAILQ_INIT(&s->msi_hashtab[i]);
988 }
989 }
990
991 kvm_arch_init_irq_routing(s);
992 }
993
994 void kvm_irqchip_commit_routes(KVMState *s)
995 {
996 int ret;
997
998 s->irq_routes->flags = 0;
999 ret = kvm_vm_ioctl(s, KVM_SET_GSI_ROUTING, s->irq_routes);
1000 assert(ret == 0);
1001 }
1002
1003 static void kvm_add_routing_entry(KVMState *s,
1004 struct kvm_irq_routing_entry *entry)
1005 {
1006 struct kvm_irq_routing_entry *new;
1007 int n, size;
1008
1009 if (s->irq_routes->nr == s->nr_allocated_irq_routes) {
1010 n = s->nr_allocated_irq_routes * 2;
1011 if (n < 64) {
1012 n = 64;
1013 }
1014 size = sizeof(struct kvm_irq_routing);
1015 size += n * sizeof(*new);
1016 s->irq_routes = g_realloc(s->irq_routes, size);
1017 s->nr_allocated_irq_routes = n;
1018 }
1019 n = s->irq_routes->nr++;
1020 new = &s->irq_routes->entries[n];
1021
1022 *new = *entry;
1023
1024 set_gsi(s, entry->gsi);
1025 }
1026
1027 static int kvm_update_routing_entry(KVMState *s,
1028 struct kvm_irq_routing_entry *new_entry)
1029 {
1030 struct kvm_irq_routing_entry *entry;
1031 int n;
1032
1033 for (n = 0; n < s->irq_routes->nr; n++) {
1034 entry = &s->irq_routes->entries[n];
1035 if (entry->gsi != new_entry->gsi) {
1036 continue;
1037 }
1038
1039 if(!memcmp(entry, new_entry, sizeof *entry)) {
1040 return 0;
1041 }
1042
1043 *entry = *new_entry;
1044
1045 kvm_irqchip_commit_routes(s);
1046
1047 return 0;
1048 }
1049
1050 return -ESRCH;
1051 }
1052
1053 void kvm_irqchip_add_irq_route(KVMState *s, int irq, int irqchip, int pin)
1054 {
1055 struct kvm_irq_routing_entry e = {};
1056
1057 assert(pin < s->gsi_count);
1058
1059 e.gsi = irq;
1060 e.type = KVM_IRQ_ROUTING_IRQCHIP;
1061 e.flags = 0;
1062 e.u.irqchip.irqchip = irqchip;
1063 e.u.irqchip.pin = pin;
1064 kvm_add_routing_entry(s, &e);
1065 }
1066
1067 void kvm_irqchip_release_virq(KVMState *s, int virq)
1068 {
1069 struct kvm_irq_routing_entry *e;
1070 int i;
1071
1072 for (i = 0; i < s->irq_routes->nr; i++) {
1073 e = &s->irq_routes->entries[i];
1074 if (e->gsi == virq) {
1075 s->irq_routes->nr--;
1076 *e = s->irq_routes->entries[s->irq_routes->nr];
1077 }
1078 }
1079 clear_gsi(s, virq);
1080 }
1081
1082 static unsigned int kvm_hash_msi(uint32_t data)
1083 {
1084 /* This is optimized for IA32 MSI layout. However, no other arch shall
1085 * repeat the mistake of not providing a direct MSI injection API. */
1086 return data & 0xff;
1087 }
1088
1089 static void kvm_flush_dynamic_msi_routes(KVMState *s)
1090 {
1091 KVMMSIRoute *route, *next;
1092 unsigned int hash;
1093
1094 for (hash = 0; hash < KVM_MSI_HASHTAB_SIZE; hash++) {
1095 QTAILQ_FOREACH_SAFE(route, &s->msi_hashtab[hash], entry, next) {
1096 kvm_irqchip_release_virq(s, route->kroute.gsi);
1097 QTAILQ_REMOVE(&s->msi_hashtab[hash], route, entry);
1098 g_free(route);
1099 }
1100 }
1101 }
1102
1103 static int kvm_irqchip_get_virq(KVMState *s)
1104 {
1105 uint32_t *word = s->used_gsi_bitmap;
1106 int max_words = ALIGN(s->gsi_count, 32) / 32;
1107 int i, bit;
1108 bool retry = true;
1109
1110 again:
1111 /* Return the lowest unused GSI in the bitmap */
1112 for (i = 0; i < max_words; i++) {
1113 bit = ffs(~word[i]);
1114 if (!bit) {
1115 continue;
1116 }
1117
1118 return bit - 1 + i * 32;
1119 }
1120 if (!s->direct_msi && retry) {
1121 retry = false;
1122 kvm_flush_dynamic_msi_routes(s);
1123 goto again;
1124 }
1125 return -ENOSPC;
1126
1127 }
1128
1129 static KVMMSIRoute *kvm_lookup_msi_route(KVMState *s, MSIMessage msg)
1130 {
1131 unsigned int hash = kvm_hash_msi(msg.data);
1132 KVMMSIRoute *route;
1133
1134 QTAILQ_FOREACH(route, &s->msi_hashtab[hash], entry) {
1135 if (route->kroute.u.msi.address_lo == (uint32_t)msg.address &&
1136 route->kroute.u.msi.address_hi == (msg.address >> 32) &&
1137 route->kroute.u.msi.data == le32_to_cpu(msg.data)) {
1138 return route;
1139 }
1140 }
1141 return NULL;
1142 }
1143
1144 int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg)
1145 {
1146 struct kvm_msi msi;
1147 KVMMSIRoute *route;
1148
1149 if (s->direct_msi) {
1150 msi.address_lo = (uint32_t)msg.address;
1151 msi.address_hi = msg.address >> 32;
1152 msi.data = le32_to_cpu(msg.data);
1153 msi.flags = 0;
1154 memset(msi.pad, 0, sizeof(msi.pad));
1155
1156 return kvm_vm_ioctl(s, KVM_SIGNAL_MSI, &msi);
1157 }
1158
1159 route = kvm_lookup_msi_route(s, msg);
1160 if (!route) {
1161 int virq;
1162
1163 virq = kvm_irqchip_get_virq(s);
1164 if (virq < 0) {
1165 return virq;
1166 }
1167
1168 route = g_malloc0(sizeof(KVMMSIRoute));
1169 route->kroute.gsi = virq;
1170 route->kroute.type = KVM_IRQ_ROUTING_MSI;
1171 route->kroute.flags = 0;
1172 route->kroute.u.msi.address_lo = (uint32_t)msg.address;
1173 route->kroute.u.msi.address_hi = msg.address >> 32;
1174 route->kroute.u.msi.data = le32_to_cpu(msg.data);
1175
1176 kvm_add_routing_entry(s, &route->kroute);
1177 kvm_irqchip_commit_routes(s);
1178
1179 QTAILQ_INSERT_TAIL(&s->msi_hashtab[kvm_hash_msi(msg.data)], route,
1180 entry);
1181 }
1182
1183 assert(route->kroute.type == KVM_IRQ_ROUTING_MSI);
1184
1185 return kvm_set_irq(s, route->kroute.gsi, 1);
1186 }
1187
1188 int kvm_irqchip_add_msi_route(KVMState *s, MSIMessage msg)
1189 {
1190 struct kvm_irq_routing_entry kroute = {};
1191 int virq;
1192
1193 if (!kvm_gsi_routing_enabled()) {
1194 return -ENOSYS;
1195 }
1196
1197 virq = kvm_irqchip_get_virq(s);
1198 if (virq < 0) {
1199 return virq;
1200 }
1201
1202 kroute.gsi = virq;
1203 kroute.type = KVM_IRQ_ROUTING_MSI;
1204 kroute.flags = 0;
1205 kroute.u.msi.address_lo = (uint32_t)msg.address;
1206 kroute.u.msi.address_hi = msg.address >> 32;
1207 kroute.u.msi.data = le32_to_cpu(msg.data);
1208
1209 kvm_add_routing_entry(s, &kroute);
1210 kvm_irqchip_commit_routes(s);
1211
1212 return virq;
1213 }
1214
1215 int kvm_irqchip_update_msi_route(KVMState *s, int virq, MSIMessage msg)
1216 {
1217 struct kvm_irq_routing_entry kroute = {};
1218
1219 if (!kvm_irqchip_in_kernel()) {
1220 return -ENOSYS;
1221 }
1222
1223 kroute.gsi = virq;
1224 kroute.type = KVM_IRQ_ROUTING_MSI;
1225 kroute.flags = 0;
1226 kroute.u.msi.address_lo = (uint32_t)msg.address;
1227 kroute.u.msi.address_hi = msg.address >> 32;
1228 kroute.u.msi.data = le32_to_cpu(msg.data);
1229
1230 return kvm_update_routing_entry(s, &kroute);
1231 }
1232
1233 static int kvm_irqchip_assign_irqfd(KVMState *s, int fd, int virq, bool assign)
1234 {
1235 struct kvm_irqfd irqfd = {
1236 .fd = fd,
1237 .gsi = virq,
1238 .flags = assign ? 0 : KVM_IRQFD_FLAG_DEASSIGN,
1239 };
1240
1241 if (!kvm_irqfds_enabled()) {
1242 return -ENOSYS;
1243 }
1244
1245 return kvm_vm_ioctl(s, KVM_IRQFD, &irqfd);
1246 }
1247
1248 #else /* !KVM_CAP_IRQ_ROUTING */
1249
1250 void kvm_init_irq_routing(KVMState *s)
1251 {
1252 }
1253
1254 void kvm_irqchip_release_virq(KVMState *s, int virq)
1255 {
1256 }
1257
1258 int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg)
1259 {
1260 abort();
1261 }
1262
1263 int kvm_irqchip_add_msi_route(KVMState *s, MSIMessage msg)
1264 {
1265 return -ENOSYS;
1266 }
1267
1268 static int kvm_irqchip_assign_irqfd(KVMState *s, int fd, int virq, bool assign)
1269 {
1270 abort();
1271 }
1272
1273 int kvm_irqchip_update_msi_route(KVMState *s, int virq, MSIMessage msg)
1274 {
1275 return -ENOSYS;
1276 }
1277 #endif /* !KVM_CAP_IRQ_ROUTING */
1278
1279 int kvm_irqchip_add_irqfd_notifier(KVMState *s, EventNotifier *n, int virq)
1280 {
1281 return kvm_irqchip_assign_irqfd(s, event_notifier_get_fd(n), virq, true);
1282 }
1283
1284 int kvm_irqchip_remove_irqfd_notifier(KVMState *s, EventNotifier *n, int virq)
1285 {
1286 return kvm_irqchip_assign_irqfd(s, event_notifier_get_fd(n), virq, false);
1287 }
1288
1289 static int kvm_irqchip_create(KVMState *s)
1290 {
1291 QemuOptsList *list = qemu_find_opts("machine");
1292 int ret;
1293
1294 if (QTAILQ_EMPTY(&list->head) ||
1295 !qemu_opt_get_bool(QTAILQ_FIRST(&list->head),
1296 "kernel_irqchip", true) ||
1297 !kvm_check_extension(s, KVM_CAP_IRQCHIP)) {
1298 return 0;
1299 }
1300
1301 ret = kvm_vm_ioctl(s, KVM_CREATE_IRQCHIP);
1302 if (ret < 0) {
1303 fprintf(stderr, "Create kernel irqchip failed\n");
1304 return ret;
1305 }
1306
1307 kvm_kernel_irqchip = true;
1308 /* If we have an in-kernel IRQ chip then we must have asynchronous
1309 * interrupt delivery (though the reverse is not necessarily true)
1310 */
1311 kvm_async_interrupts_allowed = true;
1312 kvm_halt_in_kernel_allowed = true;
1313
1314 kvm_init_irq_routing(s);
1315
1316 return 0;
1317 }
1318
1319 static int kvm_max_vcpus(KVMState *s)
1320 {
1321 int ret;
1322
1323 /* Find number of supported CPUs using the recommended
1324 * procedure from the kernel API documentation to cope with
1325 * older kernels that may be missing capabilities.
1326 */
1327 ret = kvm_check_extension(s, KVM_CAP_MAX_VCPUS);
1328 if (ret) {
1329 return ret;
1330 }
1331 ret = kvm_check_extension(s, KVM_CAP_NR_VCPUS);
1332 if (ret) {
1333 return ret;
1334 }
1335
1336 return 4;
1337 }
1338
1339 int kvm_init(void)
1340 {
1341 static const char upgrade_note[] =
1342 "Please upgrade to at least kernel 2.6.29 or recent kvm-kmod\n"
1343 "(see http://sourceforge.net/projects/kvm).\n";
1344 KVMState *s;
1345 const KVMCapabilityInfo *missing_cap;
1346 int ret;
1347 int i;
1348 int max_vcpus;
1349
1350 s = g_malloc0(sizeof(KVMState));
1351
1352 /*
1353 * On systems where the kernel can support different base page
1354 * sizes, host page size may be different from TARGET_PAGE_SIZE,
1355 * even with KVM. TARGET_PAGE_SIZE is assumed to be the minimum
1356 * page size for the system though.
1357 */
1358 assert(TARGET_PAGE_SIZE <= getpagesize());
1359
1360 #ifdef KVM_CAP_SET_GUEST_DEBUG
1361 QTAILQ_INIT(&s->kvm_sw_breakpoints);
1362 #endif
1363 for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
1364 s->slots[i].slot = i;
1365 }
1366 s->vmfd = -1;
1367 s->fd = qemu_open("/dev/kvm", O_RDWR);
1368 if (s->fd == -1) {
1369 fprintf(stderr, "Could not access KVM kernel module: %m\n");
1370 ret = -errno;
1371 goto err;
1372 }
1373
1374 ret = kvm_ioctl(s, KVM_GET_API_VERSION, 0);
1375 if (ret < KVM_API_VERSION) {
1376 if (ret > 0) {
1377 ret = -EINVAL;
1378 }
1379 fprintf(stderr, "kvm version too old\n");
1380 goto err;
1381 }
1382
1383 if (ret > KVM_API_VERSION) {
1384 ret = -EINVAL;
1385 fprintf(stderr, "kvm version not supported\n");
1386 goto err;
1387 }
1388
1389 max_vcpus = kvm_max_vcpus(s);
1390 if (smp_cpus > max_vcpus) {
1391 ret = -EINVAL;
1392 fprintf(stderr, "Number of SMP cpus requested (%d) exceeds max cpus "
1393 "supported by KVM (%d)\n", smp_cpus, max_vcpus);
1394 goto err;
1395 }
1396
1397 s->vmfd = kvm_ioctl(s, KVM_CREATE_VM, 0);
1398 if (s->vmfd < 0) {
1399 #ifdef TARGET_S390X
1400 fprintf(stderr, "Please add the 'switch_amode' kernel parameter to "
1401 "your host kernel command line\n");
1402 #endif
1403 ret = s->vmfd;
1404 goto err;
1405 }
1406
1407 missing_cap = kvm_check_extension_list(s, kvm_required_capabilites);
1408 if (!missing_cap) {
1409 missing_cap =
1410 kvm_check_extension_list(s, kvm_arch_required_capabilities);
1411 }
1412 if (missing_cap) {
1413 ret = -EINVAL;
1414 fprintf(stderr, "kvm does not support %s\n%s",
1415 missing_cap->name, upgrade_note);
1416 goto err;
1417 }
1418
1419 s->coalesced_mmio = kvm_check_extension(s, KVM_CAP_COALESCED_MMIO);
1420
1421 s->broken_set_mem_region = 1;
1422 ret = kvm_check_extension(s, KVM_CAP_JOIN_MEMORY_REGIONS_WORKS);
1423 if (ret > 0) {
1424 s->broken_set_mem_region = 0;
1425 }
1426
1427 #ifdef KVM_CAP_VCPU_EVENTS
1428 s->vcpu_events = kvm_check_extension(s, KVM_CAP_VCPU_EVENTS);
1429 #endif
1430
1431 s->robust_singlestep =
1432 kvm_check_extension(s, KVM_CAP_X86_ROBUST_SINGLESTEP);
1433
1434 #ifdef KVM_CAP_DEBUGREGS
1435 s->debugregs = kvm_check_extension(s, KVM_CAP_DEBUGREGS);
1436 #endif
1437
1438 #ifdef KVM_CAP_XSAVE
1439 s->xsave = kvm_check_extension(s, KVM_CAP_XSAVE);
1440 #endif
1441
1442 #ifdef KVM_CAP_XCRS
1443 s->xcrs = kvm_check_extension(s, KVM_CAP_XCRS);
1444 #endif
1445
1446 #ifdef KVM_CAP_PIT_STATE2
1447 s->pit_state2 = kvm_check_extension(s, KVM_CAP_PIT_STATE2);
1448 #endif
1449
1450 #ifdef KVM_CAP_IRQ_ROUTING
1451 s->direct_msi = (kvm_check_extension(s, KVM_CAP_SIGNAL_MSI) > 0);
1452 #endif
1453
1454 s->intx_set_mask = kvm_check_extension(s, KVM_CAP_PCI_2_3);
1455
1456 s->irq_set_ioctl = KVM_IRQ_LINE;
1457 if (kvm_check_extension(s, KVM_CAP_IRQ_INJECT_STATUS)) {
1458 s->irq_set_ioctl = KVM_IRQ_LINE_STATUS;
1459 }
1460
1461 #ifdef KVM_CAP_READONLY_MEM
1462 kvm_readonly_mem_allowed =
1463 (kvm_check_extension(s, KVM_CAP_READONLY_MEM) > 0);
1464 #endif
1465
1466 ret = kvm_arch_init(s);
1467 if (ret < 0) {
1468 goto err;
1469 }
1470
1471 ret = kvm_irqchip_create(s);
1472 if (ret < 0) {
1473 goto err;
1474 }
1475
1476 kvm_state = s;
1477 memory_listener_register(&kvm_memory_listener, &address_space_memory);
1478 memory_listener_register(&kvm_io_listener, &address_space_io);
1479
1480 s->many_ioeventfds = kvm_check_many_ioeventfds();
1481
1482 cpu_interrupt_handler = kvm_handle_interrupt;
1483
1484 return 0;
1485
1486 err:
1487 if (s->vmfd >= 0) {
1488 close(s->vmfd);
1489 }
1490 if (s->fd != -1) {
1491 close(s->fd);
1492 }
1493 g_free(s);
1494
1495 return ret;
1496 }
1497
1498 static void kvm_handle_io(uint16_t port, void *data, int direction, int size,
1499 uint32_t count)
1500 {
1501 int i;
1502 uint8_t *ptr = data;
1503
1504 for (i = 0; i < count; i++) {
1505 if (direction == KVM_EXIT_IO_IN) {
1506 switch (size) {
1507 case 1:
1508 stb_p(ptr, cpu_inb(port));
1509 break;
1510 case 2:
1511 stw_p(ptr, cpu_inw(port));
1512 break;
1513 case 4:
1514 stl_p(ptr, cpu_inl(port));
1515 break;
1516 }
1517 } else {
1518 switch (size) {
1519 case 1:
1520 cpu_outb(port, ldub_p(ptr));
1521 break;
1522 case 2:
1523 cpu_outw(port, lduw_p(ptr));
1524 break;
1525 case 4:
1526 cpu_outl(port, ldl_p(ptr));
1527 break;
1528 }
1529 }
1530
1531 ptr += size;
1532 }
1533 }
1534
1535 static int kvm_handle_internal_error(CPUState *cpu, struct kvm_run *run)
1536 {
1537 fprintf(stderr, "KVM internal error.");
1538 if (kvm_check_extension(kvm_state, KVM_CAP_INTERNAL_ERROR_DATA)) {
1539 int i;
1540
1541 fprintf(stderr, " Suberror: %d\n", run->internal.suberror);
1542 for (i = 0; i < run->internal.ndata; ++i) {
1543 fprintf(stderr, "extra data[%d]: %"PRIx64"\n",
1544 i, (uint64_t)run->internal.data[i]);
1545 }
1546 } else {
1547 fprintf(stderr, "\n");
1548 }
1549 if (run->internal.suberror == KVM_INTERNAL_ERROR_EMULATION) {
1550 fprintf(stderr, "emulation failure\n");
1551 if (!kvm_arch_stop_on_emulation_error(cpu)) {
1552 cpu_dump_state(cpu, stderr, fprintf, CPU_DUMP_CODE);
1553 return EXCP_INTERRUPT;
1554 }
1555 }
1556 /* FIXME: Should trigger a qmp message to let management know
1557 * something went wrong.
1558 */
1559 return -1;
1560 }
1561
1562 void kvm_flush_coalesced_mmio_buffer(void)
1563 {
1564 KVMState *s = kvm_state;
1565
1566 if (s->coalesced_flush_in_progress) {
1567 return;
1568 }
1569
1570 s->coalesced_flush_in_progress = true;
1571
1572 if (s->coalesced_mmio_ring) {
1573 struct kvm_coalesced_mmio_ring *ring = s->coalesced_mmio_ring;
1574 while (ring->first != ring->last) {
1575 struct kvm_coalesced_mmio *ent;
1576
1577 ent = &ring->coalesced_mmio[ring->first];
1578
1579 cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len);
1580 smp_wmb();
1581 ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX;
1582 }
1583 }
1584
1585 s->coalesced_flush_in_progress = false;
1586 }
1587
1588 static void do_kvm_cpu_synchronize_state(void *arg)
1589 {
1590 CPUState *cpu = arg;
1591
1592 if (!cpu->kvm_vcpu_dirty) {
1593 kvm_arch_get_registers(cpu);
1594 cpu->kvm_vcpu_dirty = true;
1595 }
1596 }
1597
1598 void kvm_cpu_synchronize_state(CPUState *cpu)
1599 {
1600 if (!cpu->kvm_vcpu_dirty) {
1601 run_on_cpu(cpu, do_kvm_cpu_synchronize_state, cpu);
1602 }
1603 }
1604
1605 void kvm_cpu_synchronize_post_reset(CPUState *cpu)
1606 {
1607 kvm_arch_put_registers(cpu, KVM_PUT_RESET_STATE);
1608 cpu->kvm_vcpu_dirty = false;
1609 }
1610
1611 void kvm_cpu_synchronize_post_init(CPUState *cpu)
1612 {
1613 kvm_arch_put_registers(cpu, KVM_PUT_FULL_STATE);
1614 cpu->kvm_vcpu_dirty = false;
1615 }
1616
1617 int kvm_cpu_exec(CPUState *cpu)
1618 {
1619 struct kvm_run *run = cpu->kvm_run;
1620 int ret, run_ret;
1621
1622 DPRINTF("kvm_cpu_exec()\n");
1623
1624 if (kvm_arch_process_async_events(cpu)) {
1625 cpu->exit_request = 0;
1626 return EXCP_HLT;
1627 }
1628
1629 do {
1630 if (cpu->kvm_vcpu_dirty) {
1631 kvm_arch_put_registers(cpu, KVM_PUT_RUNTIME_STATE);
1632 cpu->kvm_vcpu_dirty = false;
1633 }
1634
1635 kvm_arch_pre_run(cpu, run);
1636 if (cpu->exit_request) {
1637 DPRINTF("interrupt exit requested\n");
1638 /*
1639 * KVM requires us to reenter the kernel after IO exits to complete
1640 * instruction emulation. This self-signal will ensure that we
1641 * leave ASAP again.
1642 */
1643 qemu_cpu_kick_self();
1644 }
1645 qemu_mutex_unlock_iothread();
1646
1647 run_ret = kvm_vcpu_ioctl(cpu, KVM_RUN, 0);
1648
1649 qemu_mutex_lock_iothread();
1650 kvm_arch_post_run(cpu, run);
1651
1652 if (run_ret < 0) {
1653 if (run_ret == -EINTR || run_ret == -EAGAIN) {
1654 DPRINTF("io window exit\n");
1655 ret = EXCP_INTERRUPT;
1656 break;
1657 }
1658 fprintf(stderr, "error: kvm run failed %s\n",
1659 strerror(-run_ret));
1660 abort();
1661 }
1662
1663 trace_kvm_run_exit(cpu->cpu_index, run->exit_reason);
1664 switch (run->exit_reason) {
1665 case KVM_EXIT_IO:
1666 DPRINTF("handle_io\n");
1667 kvm_handle_io(run->io.port,
1668 (uint8_t *)run + run->io.data_offset,
1669 run->io.direction,
1670 run->io.size,
1671 run->io.count);
1672 ret = 0;
1673 break;
1674 case KVM_EXIT_MMIO:
1675 DPRINTF("handle_mmio\n");
1676 cpu_physical_memory_rw(run->mmio.phys_addr,
1677 run->mmio.data,
1678 run->mmio.len,
1679 run->mmio.is_write);
1680 ret = 0;
1681 break;
1682 case KVM_EXIT_IRQ_WINDOW_OPEN:
1683 DPRINTF("irq_window_open\n");
1684 ret = EXCP_INTERRUPT;
1685 break;
1686 case KVM_EXIT_SHUTDOWN:
1687 DPRINTF("shutdown\n");
1688 qemu_system_reset_request();
1689 ret = EXCP_INTERRUPT;
1690 break;
1691 case KVM_EXIT_UNKNOWN:
1692 fprintf(stderr, "KVM: unknown exit, hardware reason %" PRIx64 "\n",
1693 (uint64_t)run->hw.hardware_exit_reason);
1694 ret = -1;
1695 break;
1696 case KVM_EXIT_INTERNAL_ERROR:
1697 ret = kvm_handle_internal_error(cpu, run);
1698 break;
1699 default:
1700 DPRINTF("kvm_arch_handle_exit\n");
1701 ret = kvm_arch_handle_exit(cpu, run);
1702 break;
1703 }
1704 } while (ret == 0);
1705
1706 if (ret < 0) {
1707 cpu_dump_state(cpu, stderr, fprintf, CPU_DUMP_CODE);
1708 vm_stop(RUN_STATE_INTERNAL_ERROR);
1709 }
1710
1711 cpu->exit_request = 0;
1712 return ret;
1713 }
1714
1715 int kvm_ioctl(KVMState *s, int type, ...)
1716 {
1717 int ret;
1718 void *arg;
1719 va_list ap;
1720
1721 va_start(ap, type);
1722 arg = va_arg(ap, void *);
1723 va_end(ap);
1724
1725 trace_kvm_ioctl(type, arg);
1726 ret = ioctl(s->fd, type, arg);
1727 if (ret == -1) {
1728 ret = -errno;
1729 }
1730 return ret;
1731 }
1732
1733 int kvm_vm_ioctl(KVMState *s, int type, ...)
1734 {
1735 int ret;
1736 void *arg;
1737 va_list ap;
1738
1739 va_start(ap, type);
1740 arg = va_arg(ap, void *);
1741 va_end(ap);
1742
1743 trace_kvm_vm_ioctl(type, arg);
1744 ret = ioctl(s->vmfd, type, arg);
1745 if (ret == -1) {
1746 ret = -errno;
1747 }
1748 return ret;
1749 }
1750
1751 int kvm_vcpu_ioctl(CPUState *cpu, int type, ...)
1752 {
1753 int ret;
1754 void *arg;
1755 va_list ap;
1756
1757 va_start(ap, type);
1758 arg = va_arg(ap, void *);
1759 va_end(ap);
1760
1761 trace_kvm_vcpu_ioctl(cpu->cpu_index, type, arg);
1762 ret = ioctl(cpu->kvm_fd, type, arg);
1763 if (ret == -1) {
1764 ret = -errno;
1765 }
1766 return ret;
1767 }
1768
1769 int kvm_has_sync_mmu(void)
1770 {
1771 return kvm_check_extension(kvm_state, KVM_CAP_SYNC_MMU);
1772 }
1773
1774 int kvm_has_vcpu_events(void)
1775 {
1776 return kvm_state->vcpu_events;
1777 }
1778
1779 int kvm_has_robust_singlestep(void)
1780 {
1781 return kvm_state->robust_singlestep;
1782 }
1783
1784 int kvm_has_debugregs(void)
1785 {
1786 return kvm_state->debugregs;
1787 }
1788
1789 int kvm_has_xsave(void)
1790 {
1791 return kvm_state->xsave;
1792 }
1793
1794 int kvm_has_xcrs(void)
1795 {
1796 return kvm_state->xcrs;
1797 }
1798
1799 int kvm_has_pit_state2(void)
1800 {
1801 return kvm_state->pit_state2;
1802 }
1803
1804 int kvm_has_many_ioeventfds(void)
1805 {
1806 if (!kvm_enabled()) {
1807 return 0;
1808 }
1809 return kvm_state->many_ioeventfds;
1810 }
1811
1812 int kvm_has_gsi_routing(void)
1813 {
1814 #ifdef KVM_CAP_IRQ_ROUTING
1815 return kvm_check_extension(kvm_state, KVM_CAP_IRQ_ROUTING);
1816 #else
1817 return false;
1818 #endif
1819 }
1820
1821 int kvm_has_intx_set_mask(void)
1822 {
1823 return kvm_state->intx_set_mask;
1824 }
1825
1826 void *kvm_ram_alloc(ram_addr_t size)
1827 {
1828 #ifdef TARGET_S390X
1829 void *mem;
1830
1831 mem = kvm_arch_ram_alloc(size);
1832 if (mem) {
1833 return mem;
1834 }
1835 #endif
1836 return qemu_anon_ram_alloc(size);
1837 }
1838
1839 void kvm_setup_guest_memory(void *start, size_t size)
1840 {
1841 #ifdef CONFIG_VALGRIND_H
1842 VALGRIND_MAKE_MEM_DEFINED(start, size);
1843 #endif
1844 if (!kvm_has_sync_mmu()) {
1845 int ret = qemu_madvise(start, size, QEMU_MADV_DONTFORK);
1846
1847 if (ret) {
1848 perror("qemu_madvise");
1849 fprintf(stderr,
1850 "Need MADV_DONTFORK in absence of synchronous KVM MMU\n");
1851 exit(1);
1852 }
1853 }
1854 }
1855
1856 #ifdef KVM_CAP_SET_GUEST_DEBUG
1857 struct kvm_sw_breakpoint *kvm_find_sw_breakpoint(CPUState *cpu,
1858 target_ulong pc)
1859 {
1860 struct kvm_sw_breakpoint *bp;
1861
1862 QTAILQ_FOREACH(bp, &cpu->kvm_state->kvm_sw_breakpoints, entry) {
1863 if (bp->pc == pc) {
1864 return bp;
1865 }
1866 }
1867 return NULL;
1868 }
1869
1870 int kvm_sw_breakpoints_active(CPUState *cpu)
1871 {
1872 return !QTAILQ_EMPTY(&cpu->kvm_state->kvm_sw_breakpoints);
1873 }
1874
1875 struct kvm_set_guest_debug_data {
1876 struct kvm_guest_debug dbg;
1877 CPUState *cpu;
1878 int err;
1879 };
1880
1881 static void kvm_invoke_set_guest_debug(void *data)
1882 {
1883 struct kvm_set_guest_debug_data *dbg_data = data;
1884
1885 dbg_data->err = kvm_vcpu_ioctl(dbg_data->cpu, KVM_SET_GUEST_DEBUG,
1886 &dbg_data->dbg);
1887 }
1888
1889 int kvm_update_guest_debug(CPUArchState *env, unsigned long reinject_trap)
1890 {
1891 CPUState *cpu = ENV_GET_CPU(env);
1892 struct kvm_set_guest_debug_data data;
1893
1894 data.dbg.control = reinject_trap;
1895
1896 if (env->singlestep_enabled) {
1897 data.dbg.control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_SINGLESTEP;
1898 }
1899 kvm_arch_update_guest_debug(cpu, &data.dbg);
1900 data.cpu = cpu;
1901
1902 run_on_cpu(cpu, kvm_invoke_set_guest_debug, &data);
1903 return data.err;
1904 }
1905
1906 int kvm_insert_breakpoint(CPUArchState *current_env, target_ulong addr,
1907 target_ulong len, int type)
1908 {
1909 CPUState *current_cpu = ENV_GET_CPU(current_env);
1910 struct kvm_sw_breakpoint *bp;
1911 CPUArchState *env;
1912 int err;
1913
1914 if (type == GDB_BREAKPOINT_SW) {
1915 bp = kvm_find_sw_breakpoint(current_cpu, addr);
1916 if (bp) {
1917 bp->use_count++;
1918 return 0;
1919 }
1920
1921 bp = g_malloc(sizeof(struct kvm_sw_breakpoint));
1922 if (!bp) {
1923 return -ENOMEM;
1924 }
1925
1926 bp->pc = addr;
1927 bp->use_count = 1;
1928 err = kvm_arch_insert_sw_breakpoint(current_cpu, bp);
1929 if (err) {
1930 g_free(bp);
1931 return err;
1932 }
1933
1934 QTAILQ_INSERT_HEAD(&current_cpu->kvm_state->kvm_sw_breakpoints,
1935 bp, entry);
1936 } else {
1937 err = kvm_arch_insert_hw_breakpoint(addr, len, type);
1938 if (err) {
1939 return err;
1940 }
1941 }
1942
1943 for (env = first_cpu; env != NULL; env = env->next_cpu) {
1944 err = kvm_update_guest_debug(env, 0);
1945 if (err) {
1946 return err;
1947 }
1948 }
1949 return 0;
1950 }
1951
1952 int kvm_remove_breakpoint(CPUArchState *current_env, target_ulong addr,
1953 target_ulong len, int type)
1954 {
1955 CPUState *current_cpu = ENV_GET_CPU(current_env);
1956 struct kvm_sw_breakpoint *bp;
1957 CPUArchState *env;
1958 int err;
1959
1960 if (type == GDB_BREAKPOINT_SW) {
1961 bp = kvm_find_sw_breakpoint(current_cpu, addr);
1962 if (!bp) {
1963 return -ENOENT;
1964 }
1965
1966 if (bp->use_count > 1) {
1967 bp->use_count--;
1968 return 0;
1969 }
1970
1971 err = kvm_arch_remove_sw_breakpoint(current_cpu, bp);
1972 if (err) {
1973 return err;
1974 }
1975
1976 QTAILQ_REMOVE(&current_cpu->kvm_state->kvm_sw_breakpoints, bp, entry);
1977 g_free(bp);
1978 } else {
1979 err = kvm_arch_remove_hw_breakpoint(addr, len, type);
1980 if (err) {
1981 return err;
1982 }
1983 }
1984
1985 for (env = first_cpu; env != NULL; env = env->next_cpu) {
1986 err = kvm_update_guest_debug(env, 0);
1987 if (err) {
1988 return err;
1989 }
1990 }
1991 return 0;
1992 }
1993
1994 void kvm_remove_all_breakpoints(CPUArchState *current_env)
1995 {
1996 CPUState *current_cpu = ENV_GET_CPU(current_env);
1997 struct kvm_sw_breakpoint *bp, *next;
1998 KVMState *s = current_cpu->kvm_state;
1999 CPUArchState *env;
2000 CPUState *cpu;
2001
2002 QTAILQ_FOREACH_SAFE(bp, &s->kvm_sw_breakpoints, entry, next) {
2003 if (kvm_arch_remove_sw_breakpoint(current_cpu, bp) != 0) {
2004 /* Try harder to find a CPU that currently sees the breakpoint. */
2005 for (env = first_cpu; env != NULL; env = env->next_cpu) {
2006 cpu = ENV_GET_CPU(env);
2007 if (kvm_arch_remove_sw_breakpoint(cpu, bp) == 0) {
2008 break;
2009 }
2010 }
2011 }
2012 QTAILQ_REMOVE(&s->kvm_sw_breakpoints, bp, entry);
2013 g_free(bp);
2014 }
2015 kvm_arch_remove_all_hw_breakpoints();
2016
2017 for (env = first_cpu; env != NULL; env = env->next_cpu) {
2018 kvm_update_guest_debug(env, 0);
2019 }
2020 }
2021
2022 #else /* !KVM_CAP_SET_GUEST_DEBUG */
2023
2024 int kvm_update_guest_debug(CPUArchState *env, unsigned long reinject_trap)
2025 {
2026 return -EINVAL;
2027 }
2028
2029 int kvm_insert_breakpoint(CPUArchState *current_env, target_ulong addr,
2030 target_ulong len, int type)
2031 {
2032 return -EINVAL;
2033 }
2034
2035 int kvm_remove_breakpoint(CPUArchState *current_env, target_ulong addr,
2036 target_ulong len, int type)
2037 {
2038 return -EINVAL;
2039 }
2040
2041 void kvm_remove_all_breakpoints(CPUArchState *current_env)
2042 {
2043 }
2044 #endif /* !KVM_CAP_SET_GUEST_DEBUG */
2045
2046 int kvm_set_signal_mask(CPUState *cpu, const sigset_t *sigset)
2047 {
2048 struct kvm_signal_mask *sigmask;
2049 int r;
2050
2051 if (!sigset) {
2052 return kvm_vcpu_ioctl(cpu, KVM_SET_SIGNAL_MASK, NULL);
2053 }
2054
2055 sigmask = g_malloc(sizeof(*sigmask) + sizeof(*sigset));
2056
2057 sigmask->len = 8;
2058 memcpy(sigmask->sigset, sigset, sizeof(*sigset));
2059 r = kvm_vcpu_ioctl(cpu, KVM_SET_SIGNAL_MASK, sigmask);
2060 g_free(sigmask);
2061
2062 return r;
2063 }
2064 int kvm_on_sigbus_vcpu(CPUState *cpu, int code, void *addr)
2065 {
2066 return kvm_arch_on_sigbus_vcpu(cpu, code, addr);
2067 }
2068
2069 int kvm_on_sigbus(int code, void *addr)
2070 {
2071 return kvm_arch_on_sigbus(code, addr);
2072 }