1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
69 #define	VM_LOWMEM_LIMIT	0
98 	memset(&vmc, 0, sizeof(vmc));  in vm_ctl_create()
101 		return (-1);  in vm_ctl_create()
112 	if (modfind("vmm") < 0) {  in vm_create()
114 		if (error != 0)  in vm_create()
115 			return (-1);  in vm_create()
118 	fd = open("/dev/vmmctl", O_RDWR, 0);  in vm_create()
119 	if (fd < 0)  in vm_create()
122 	if (error != 0) {  in vm_create()
126 		return (-1);  in vm_create()
129 	return (0);  in vm_create()
135 	return (vm_openf(name, 0));  in vm_open()
150 	vm->fd = vm->ctlfd = -1;  in vm_openf()
151 	vm->memflags = 0;  in vm_openf()
152 	vm->name = (char *)(vm + 1);  in vm_openf()
153 	strcpy(vm->name, name);  in vm_openf()
154 	memset(vm->memsegs, 0, sizeof(vm->memsegs));  in vm_openf()
156 	if ((vm->ctlfd = open("/dev/vmmctl", O_RDWR, 0)) < 0)  in vm_openf()
159 	vm->fd = vm_device_open(vm->name);  in vm_openf()
160 	if (vm->fd < 0 && errno == ENOENT) {  in vm_openf()
162 			if (vm_ctl_create(vm->name, vm->ctlfd) != 0)  in vm_openf()
164 			vm->fd = vm_device_open(vm->name);  in vm_openf()
168 	if (vm->fd < 0)  in vm_openf()
171 	if (!created && (flags & VMMAPI_OPEN_REINIT) != 0 && vm_reinit(vm) != 0)  in vm_openf()
190 	if (vm->fd >= 0)  in vm_close()
191 		(void)close(vm->fd);  in vm_close()
192 	if (vm->ctlfd >= 0)  in vm_close()
193 		(void)close(vm->ctlfd);  in vm_close()
202 	memset(&vmd, 0, sizeof(vmd));  in vm_destroy()
203 	(void)strlcpy(vmd.name, vm->name, sizeof(vmd.name));  in vm_destroy()
204 	if (ioctl(vm->ctlfd, VMMCTL_VM_DESTROY, &vmd) != 0)  in vm_destroy()
216 	vcpu->ctx = ctx;  in vm_vcpu_open()
217 	vcpu->vcpuid = vcpuid;  in vm_vcpu_open()
230 	return (vcpu->vcpuid);  in vcpu_id()
240 	optval = strtoul(opt, &endptr, 0);  in vm_parse_memsize()
241 	if (*opt != '\0' && *endptr == '\0') {  in vm_parse_memsize()
250 		error = 0;  in vm_parse_memsize()
268 	ctx->memflags = flags;  in vm_set_memflags()
275 	return (ctx->memflags);  in vm_get_memflags()
279  * Map segment 'segid' starting at 'off' into guest address range [gpa,gpa+len).
282 vm_mmap_memseg(struct vmctx *ctx, vm_paddr_t gpa, int segid, vm_ooffset_t off,  in vm_mmap_memseg()  argument
288 	memmap.gpa = gpa;  in vm_mmap_memseg()
293 	memmap.flags = 0;  in vm_mmap_memseg()
295 	if (ctx->memflags & VM_MEM_F_WIRED)  in vm_mmap_memseg()
302 	error = vm_mmap_getnext(ctx, &gpa, &segid, &off, &len, &prot, &flags);  in vm_mmap_memseg()
303 	if (error == 0 && gpa == memmap.gpa) {  in vm_mmap_memseg()
307 			return (-1);  in vm_mmap_memseg()
309 			return (0);  in vm_mmap_memseg()
313 	error = ioctl(ctx->fd, VM_MMAP_MEMSEG, &memmap);  in vm_mmap_memseg()
322 	*guest_baseaddr = ctx->baseaddr;  in vm_get_guestmem_from_ctx()
323 	*lowmem_size = ctx->memsegs[VM_MEMSEG_LOW].size;  in vm_get_guestmem_from_ctx()
324 	*highmem_size = ctx->memsegs[VM_MEMSEG_HIGH].size;  in vm_get_guestmem_from_ctx()
325 	return (0);  in vm_get_guestmem_from_ctx()
329 vm_munmap_memseg(struct vmctx *ctx, vm_paddr_t gpa, size_t len)  in vm_munmap_memseg()  argument
334 	munmap.gpa = gpa;  in vm_munmap_memseg()
337 	error = ioctl(ctx->fd, VM_MUNMAP_MEMSEG, &munmap);  in vm_munmap_memseg()
342 vm_mmap_getnext(struct vmctx *ctx, vm_paddr_t *gpa, int *segid,  in vm_mmap_getnext()  argument
349 	memmap.gpa = *gpa;  in vm_mmap_getnext()
350 	error = ioctl(ctx->fd, VM_MMAP_GETNEXT, &memmap);  in vm_mmap_getnext()
351 	if (error == 0) {  in vm_mmap_getnext()
352 		*gpa = memmap.gpa;  in vm_mmap_getnext()
363  * Return 0 if the segments are identical and non-zero otherwise.
374 			return (0);  in cmpseg()
376 	return (-1);  in cmpseg()
396 	if (memseg.len != 0) {  in vm_alloc_memseg()
399 			return (-1);  in vm_alloc_memseg()
401 			return (0);  in vm_alloc_memseg()
412 			return (-1);  in vm_alloc_memseg()
416 	error = ioctl(ctx->fd, VM_ALLOC_MEMSEG, &memseg);  in vm_alloc_memseg()
430 	error = ioctl(ctx->fd, VM_GET_MEMSEG, &memseg);  in vm_get_memseg()
431 	if (error == 0) {  in vm_get_memseg()
436 			error = -1;  in vm_get_memseg()
443 setup_memory_segment(struct vmctx *ctx, vm_paddr_t gpa, size_t len, char *base)  in setup_memory_segment()  argument
448 	/* Map 'len' bytes starting at 'gpa' in the guest address space */  in setup_memory_segment()
449 	error = vm_mmap_memseg(ctx, gpa, VM_SYSMEM, gpa, len, PROT_ALL);  in setup_memory_segment()
454 	if ((ctx->memflags & VM_MEM_F_INCORE) == 0)  in setup_memory_segment()
458 	ptr = mmap(base + gpa, len, PROT_RW, flags, ctx->fd, gpa);  in setup_memory_segment()
460 		return (-1);  in setup_memory_segment()
462 	return (0);  in setup_memory_segment()
469 	vm_paddr_t gpa;  in vm_setup_memory()  local
480 		ctx->memsegs[VM_MEMSEG_LOW].size = VM_LOWMEM_LIMIT;  in vm_setup_memory()
481 		ctx->memsegs[VM_MEMSEG_HIGH].size = memsize - VM_LOWMEM_LIMIT;  in vm_setup_memory()
482 		objsize = VM_HIGHMEM_BASE + ctx->memsegs[VM_MEMSEG_HIGH].size;  in vm_setup_memory()
484 		ctx->memsegs[VM_MEMSEG_LOW].size = memsize;  in vm_setup_memory()
485 		ctx->memsegs[VM_MEMSEG_HIGH].size = 0;  in vm_setup_memory()
498 	ptr = mmap(NULL, len, PROT_NONE, MAP_GUARD | MAP_ALIGNED_SUPER, -1, 0);  in vm_setup_memory()
500 		return (-1);  in vm_setup_memory()
503 	if (ctx->memsegs[VM_MEMSEG_HIGH].size > 0) {  in vm_setup_memory()
504 		gpa = VM_HIGHMEM_BASE;  in vm_setup_memory()
505 		len = ctx->memsegs[VM_MEMSEG_HIGH].size;  in vm_setup_memory()
506 		error = setup_memory_segment(ctx, gpa, len, baseaddr);  in vm_setup_memory()
511 	if (ctx->memsegs[VM_MEMSEG_LOW].size > 0) {  in vm_setup_memory()
512 		gpa = 0;  in vm_setup_memory()
513 		len = ctx->memsegs[VM_MEMSEG_LOW].size;  in vm_setup_memory()
514 		error = setup_memory_segment(ctx, gpa, len, baseaddr);  in vm_setup_memory()
519 	ctx->baseaddr = baseaddr;  in vm_setup_memory()
521 	return (0);  in vm_setup_memory()
525  * Returns a non-NULL pointer if [gaddr, gaddr+len) is entirely contained in
536 	lowsize = ctx->memsegs[VM_MEMSEG_LOW].size;  in vm_map_gpa()
537 	if (lowsize > 0) {  in vm_map_gpa()
539 			return (ctx->baseaddr + gaddr);  in vm_map_gpa()
542 	highsize = ctx->memsegs[VM_MEMSEG_HIGH].size;  in vm_map_gpa()
543 	if (highsize > 0 && gaddr >= VM_HIGHMEM_BASE) {  in vm_map_gpa()
546 			return (ctx->baseaddr + gaddr);  in vm_map_gpa()
558 	offaddr = (char *)addr - ctx->baseaddr;  in vm_rev_map_gpa()
560 	lowsize = ctx->memsegs[VM_MEMSEG_LOW].size;  in vm_rev_map_gpa()
561 	if (lowsize > 0)  in vm_rev_map_gpa()
565 	highsize = ctx->memsegs[VM_MEMSEG_HIGH].size;  in vm_rev_map_gpa()
566 	if (highsize > 0)  in vm_rev_map_gpa()
571 	return ((vm_paddr_t)-1);  in vm_rev_map_gpa()
578 	return (ctx->name);  in vm_get_name()
585 	return (ctx->memsegs[VM_MEMSEG_LOW].size);  in vm_get_lowmem_size()
599 	return (ctx->memsegs[VM_MEMSEG_HIGH].size);  in vm_get_highmem_size()
610 	fd = -1;  in vm_create_devmem()
612 	if (name == NULL || strlen(name) == 0) {  in vm_create_devmem()
622 	strlcat(pathname, ctx->name, sizeof(pathname));  in vm_create_devmem()
627 	if (fd < 0)  in vm_create_devmem()
635 	base = mmap(NULL, len2, PROT_NONE, MAP_GUARD | MAP_ALIGNED_SUPER, -1,  in vm_create_devmem()
636 	    0);  in vm_create_devmem()
641 	if ((ctx->memflags & VM_MEM_F_INCORE) == 0)  in vm_create_devmem()
645 	ptr = mmap(base + VM_MMAP_GUARD_SIZE, len, PROT_RW, flags, fd, 0);  in vm_create_devmem()
647 	if (fd >= 0)  in vm_create_devmem()
660 	*(int *)arg = vcpu->vcpuid;  in vcpu_ioctl()
661 	return (ioctl(vcpu->ctx->fd, cmd, arg));  in vcpu_ioctl()
737 	return (ioctl(ctx->fd, VM_SUSPEND, &vmsuspend));  in vm_suspend()
744 	return (ioctl(ctx->fd, VM_REINIT, 0));  in vm_reinit()
752 	for (i = 0; i < VM_CAP_MAX; i++) {  in vm_capability_name2type()
754 		    strcmp(vm_capstrmap[i], capname) == 0)  in vm_capability_name2type()
758 	return (-1);  in vm_capability_name2type()
764 	if (type >= 0 && type < VM_CAP_MAX)  in vm_capability_type2name()
808 	count = 0;  in vm_get_stats()
809 	for (index = 0;; index += nitems(vmstats.statbuf)) {  in vm_get_stats()
811 		if (vcpu_ioctl(vcpu, VM_STATS, &vmstats) != 0)  in vm_get_stats()
847 	if (ioctl(ctx->fd, VM_STAT_DESC, &statdesc) == 0)  in vm_get_stat_desc()
855 vm_get_gpa_pmap(struct vmctx *ctx, uint64_t gpa, uint64_t *pte, int *num)  in vm_get_gpa_pmap()  argument
861 	gpapte.gpa = gpa;  in vm_get_gpa_pmap()
863 	error = ioctl(ctx->fd, VM_GET_GPA_PMAP, &gpapte);  in vm_get_gpa_pmap()
865 	if (error == 0) {  in vm_get_gpa_pmap()
867 		for (i = 0; i < gpapte.ptenum; i++)  in vm_get_gpa_pmap()
876     uint64_t gla, int prot, uint64_t *gpa, int *fault)  in vm_gla2gpa()  argument
887 	if (error == 0) {  in vm_gla2gpa()
889 		*gpa = gg.gpa;  in vm_gla2gpa()
897     uint64_t gla, int prot, uint64_t *gpa, int *fault)  in vm_gla2gpa_nofault()  argument
908 	if (error == 0) {  in vm_gla2gpa_nofault()
910 		*gpa = gg.gpa;  in vm_gla2gpa_nofault()
926 	uint64_t gpa, off;  in vm_copy_setup()  local
929 	for (i = 0; i < iovcnt; i++) {  in vm_copy_setup()
930 		iov[i].iov_base = 0;  in vm_copy_setup()
931 		iov[i].iov_len = 0;  in vm_copy_setup()
935 		assert(iovcnt > 0);  in vm_copy_setup()
936 		error = vm_gla2gpa(vcpu, paging, gla, prot, &gpa, fault);  in vm_copy_setup()
940 		off = gpa & PAGE_MASK;  in vm_copy_setup()
941 		n = MIN(len, PAGE_SIZE - off);  in vm_copy_setup()
943 		va = vm_map_gpa(vcpu->ctx, gpa, n);  in vm_copy_setup()
947 		iov->iov_base = va;  in vm_copy_setup()
948 		iov->iov_len = n;  in vm_copy_setup()
950 		iovcnt--;  in vm_copy_setup()
953 		len -= n;  in vm_copy_setup()
955 	return (0);  in vm_copy_setup()
964 	 * emulation code shared with the kernel.  The in-kernel  in vm_copy_teardown()
965 	 * version of this is non-empty.  in vm_copy_teardown()
978 		assert(iov->iov_len);  in vm_copyin()
979 		n = min(len, iov->iov_len);  in vm_copyin()
980 		src = iov->iov_base;  in vm_copyin()
985 		len -= n;  in vm_copyin()
998 		assert(iov->iov_len);  in vm_copyout()
999 		n = min(len, iov->iov_len);  in vm_copyout()
1000 		dst = iov->iov_base;  in vm_copyout()
1005 		len -= n;  in vm_copyout()
1020 	error = ioctl(ctx->fd, VM_GET_CPUS, &vm_cpuset);  in vm_get_cpus()
1063 	ac.vcpuid = -1;  in vm_suspend_all_cpus()
1064 	error = ioctl(ctx->fd, VM_SUSPEND_CPU, &ac);  in vm_suspend_all_cpus()
1097 	ac.vcpuid = -1;  in vm_resume_all_cpus()
1098 	error = ioctl(ctx->fd, VM_RESUME_CPU, &ac);  in vm_resume_all_cpus()
1111 	if (error == 0) {  in vm_get_intinfo()
1144 	if (ioctl(ctx->fd, VM_SNAPSHOT_REQ, meta) == -1) {  in vm_snapshot_req()
1147 		    __func__, meta->dev_name, errno);  in vm_snapshot_req()
1149 		return (-1);  in vm_snapshot_req()
1151 	return (0);  in vm_snapshot_req()
1159 	dummy = 0;  in vm_restore_time()
1160 	return (ioctl(ctx->fd, VM_RESTORE_TIME, &dummy));  in vm_restore_time()
1175 	return (ioctl(ctx->fd, VM_SET_TOPOLOGY, &topology));  in vm_set_topology()
1186 	error = ioctl(ctx->fd, VM_GET_TOPOLOGY, &topology);  in vm_get_topology()
1187 	if (error == 0) {  in vm_get_topology()
1202 	if (caph_rights_limit(ctx->fd, &rights) != 0)  in vm_limit_rights()
1203 		return (-1);  in vm_limit_rights()
1204 	if (caph_ioctls_limit(ctx->fd, vm_ioctl_cmds, vm_ioctl_ncmds) != 0)  in vm_limit_rights()
1205 		return (-1);  in vm_limit_rights()
1206 	return (0);  in vm_limit_rights()
1217 	return (ctx->fd);  in vm_get_device_fd()
1228 		sz = vm_ioctl_ncmds * sizeof(vm_ioctl_cmds[0]);  in vm_get_ioctls()