2 * generic functions used by VFIO devices
4 * Copyright Red Hat, Inc. 2012
9 * This work is licensed under the terms of the GNU GPL, version 2. See
10 * the COPYING file in the top-level directory.
12 * Based on qemu-kvm device-assignment:
13 * Adapted for KVM by Qumranet.
21 #include "qemu/osdep.h"
22 #include <sys/ioctl.h>
24 #include <linux/kvm.h>
26 #include <linux/vfio.h>
28 #include "hw/vfio/vfio-common.h"
29 #include "hw/vfio/vfio.h"
30 #include "exec/address-spaces.h"
31 #include "exec/memory.h"
33 #include "qemu/error-report.h"
34 #include "qemu/range.h"
35 #include "sysemu/kvm.h"
38 struct vfio_group_head vfio_group_list =
39 QLIST_HEAD_INITIALIZER(vfio_group_list);
40 struct vfio_as_head vfio_address_spaces =
41 QLIST_HEAD_INITIALIZER(vfio_address_spaces);
45 * We have a single VFIO pseudo device per KVM VM. Once created it lives
46 * for the life of the VM. Closing the file descriptor only drops our
47 * reference to it and the device's reference to kvm. Therefore once
48 * initialized, this file descriptor is only released on QEMU exit and
49 * we'll re-use it should another vfio device be attached before then.
51 static int vfio_kvm_device_fd = -1;
55 * Common VFIO interrupt disable
57 void vfio_disable_irqindex(VFIODevice *vbasedev, int index)
59 struct vfio_irq_set irq_set = {
60 .argsz = sizeof(irq_set),
61 .flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_TRIGGER,
67 ioctl(vbasedev->fd, VFIO_DEVICE_SET_IRQS, &irq_set);
70 void vfio_unmask_single_irqindex(VFIODevice *vbasedev, int index)
72 struct vfio_irq_set irq_set = {
73 .argsz = sizeof(irq_set),
74 .flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_UNMASK,
80 ioctl(vbasedev->fd, VFIO_DEVICE_SET_IRQS, &irq_set);
83 void vfio_mask_single_irqindex(VFIODevice *vbasedev, int index)
85 struct vfio_irq_set irq_set = {
86 .argsz = sizeof(irq_set),
87 .flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_MASK,
93 ioctl(vbasedev->fd, VFIO_DEVICE_SET_IRQS, &irq_set);
97 * IO Port/MMIO - Beware of the endians, VFIO is always little endian
99 void vfio_region_write(void *opaque, hwaddr addr,
100 uint64_t data, unsigned size)
102 VFIORegion *region = opaque;
103 VFIODevice *vbasedev = region->vbasedev;
116 buf.word = cpu_to_le16(data);
119 buf.dword = cpu_to_le32(data);
122 hw_error("vfio: unsupported write size, %d bytes", size);
126 if (pwrite(vbasedev->fd, &buf, size, region->fd_offset + addr) != size) {
127 error_report("%s(%s:region%d+0x%"HWADDR_PRIx", 0x%"PRIx64
129 __func__, vbasedev->name, region->nr,
133 trace_vfio_region_write(vbasedev->name, region->nr, addr, data, size);
136 * A read or write to a BAR always signals an INTx EOI. This will
137 * do nothing if not pending (including not in INTx mode). We assume
138 * that a BAR access is in response to an interrupt and that BAR
139 * accesses will service the interrupt. Unfortunately, we don't know
140 * which access will service the interrupt, so we're potentially
141 * getting quite a few host interrupts per guest interrupt.
143 vbasedev->ops->vfio_eoi(vbasedev);
146 uint64_t vfio_region_read(void *opaque,
147 hwaddr addr, unsigned size)
149 VFIORegion *region = opaque;
150 VFIODevice *vbasedev = region->vbasedev;
159 if (pread(vbasedev->fd, &buf, size, region->fd_offset + addr) != size) {
160 error_report("%s(%s:region%d+0x%"HWADDR_PRIx", %d) failed: %m",
161 __func__, vbasedev->name, region->nr,
170 data = le16_to_cpu(buf.word);
173 data = le32_to_cpu(buf.dword);
176 hw_error("vfio: unsupported read size, %d bytes", size);
180 trace_vfio_region_read(vbasedev->name, region->nr, addr, size, data);
182 /* Same as write above */
183 vbasedev->ops->vfio_eoi(vbasedev);
188 const MemoryRegionOps vfio_region_ops = {
189 .read = vfio_region_read,
190 .write = vfio_region_write,
191 .endianness = DEVICE_LITTLE_ENDIAN,
195 * DMA - Mapping and unmapping for the "type1" IOMMU interface used on x86
197 static int vfio_dma_unmap(VFIOContainer *container,
198 hwaddr iova, ram_addr_t size)
200 struct vfio_iommu_type1_dma_unmap unmap = {
201 .argsz = sizeof(unmap),
207 if (ioctl(container->fd, VFIO_IOMMU_UNMAP_DMA, &unmap)) {
208 error_report("VFIO_UNMAP_DMA: %d", -errno);
215 static int vfio_dma_map(VFIOContainer *container, hwaddr iova,
216 ram_addr_t size, void *vaddr, bool readonly)
218 struct vfio_iommu_type1_dma_map map = {
219 .argsz = sizeof(map),
220 .flags = VFIO_DMA_MAP_FLAG_READ,
221 .vaddr = (__u64)(uintptr_t)vaddr,
227 map.flags |= VFIO_DMA_MAP_FLAG_WRITE;
231 * Try the mapping, if it fails with EBUSY, unmap the region and try
232 * again. This shouldn't be necessary, but we sometimes see it in
235 if (ioctl(container->fd, VFIO_IOMMU_MAP_DMA, &map) == 0 ||
236 (errno == EBUSY && vfio_dma_unmap(container, iova, size) == 0 &&
237 ioctl(container->fd, VFIO_IOMMU_MAP_DMA, &map) == 0)) {
241 error_report("VFIO_MAP_DMA: %d", -errno);
245 static void vfio_host_win_add(VFIOContainer *container,
246 hwaddr min_iova, hwaddr max_iova,
247 uint64_t iova_pgsizes)
249 VFIOHostDMAWindow *hostwin;
251 QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) {
252 if (ranges_overlap(hostwin->min_iova,
253 hostwin->max_iova - hostwin->min_iova + 1,
255 max_iova - min_iova + 1)) {
256 hw_error("%s: Overlapped IOMMU are not enabled", __func__);
260 hostwin = g_malloc0(sizeof(*hostwin));
262 hostwin->min_iova = min_iova;
263 hostwin->max_iova = max_iova;
264 hostwin->iova_pgsizes = iova_pgsizes;
265 QLIST_INSERT_HEAD(&container->hostwin_list, hostwin, hostwin_next);
268 static int vfio_host_win_del(VFIOContainer *container, hwaddr min_iova,
271 VFIOHostDMAWindow *hostwin;
273 QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) {
274 if (hostwin->min_iova == min_iova && hostwin->max_iova == max_iova) {
275 QLIST_REMOVE(hostwin, hostwin_next);
283 static bool vfio_listener_skipped_section(MemoryRegionSection *section)
285 return (!memory_region_is_ram(section->mr) &&
286 !memory_region_is_iommu(section->mr)) ||
288 * Sizing an enabled 64-bit BAR can cause spurious mappings to
289 * addresses in the upper part of the 64-bit address space. These
290 * are never accessed by the CPU and beyond the address width of
291 * some IOMMU hardware. TODO: VFIO should tell us the IOMMU width.
293 section->offset_within_address_space & (1ULL << 63);
296 static void vfio_iommu_map_notify(Notifier *n, void *data)
298 VFIOGuestIOMMU *giommu = container_of(n, VFIOGuestIOMMU, n);
299 VFIOContainer *container = giommu->container;
300 IOMMUTLBEntry *iotlb = data;
301 hwaddr iova = iotlb->iova + giommu->iommu_offset;
304 hwaddr len = iotlb->addr_mask + 1;
308 trace_vfio_iommu_map_notify(iova, iova + iotlb->addr_mask);
310 if (iotlb->target_as != &address_space_memory) {
311 error_report("Wrong target AS \"%s\", only system memory is allowed",
312 iotlb->target_as->name ? iotlb->target_as->name : "none");
317 * The IOMMU TLB entry we have just covers translation through
318 * this IOMMU to its immediate target. We need to translate
319 * it the rest of the way through to memory.
322 mr = address_space_translate(&address_space_memory,
323 iotlb->translated_addr,
324 &xlat, &len, iotlb->perm & IOMMU_WO);
325 if (!memory_region_is_ram(mr)) {
326 error_report("iommu map to non memory area %"HWADDR_PRIx"",
331 * Translation truncates length to the IOMMU page size,
332 * check that it did not truncate too much.
334 if (len & iotlb->addr_mask) {
335 error_report("iommu has granularity incompatible with target AS");
339 if ((iotlb->perm & IOMMU_RW) != IOMMU_NONE) {
340 vaddr = memory_region_get_ram_ptr(mr) + xlat;
341 ret = vfio_dma_map(container, iova,
342 iotlb->addr_mask + 1, vaddr,
343 !(iotlb->perm & IOMMU_WO) || mr->readonly);
345 error_report("vfio_dma_map(%p, 0x%"HWADDR_PRIx", "
346 "0x%"HWADDR_PRIx", %p) = %d (%m)",
348 iotlb->addr_mask + 1, vaddr, ret);
351 ret = vfio_dma_unmap(container, iova, iotlb->addr_mask + 1);
353 error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx", "
354 "0x%"HWADDR_PRIx") = %d (%m)",
356 iotlb->addr_mask + 1, ret);
363 static void vfio_listener_region_add(MemoryListener *listener,
364 MemoryRegionSection *section)
366 VFIOContainer *container = container_of(listener, VFIOContainer, listener);
368 Int128 llend, llsize;
371 VFIOHostDMAWindow *hostwin;
374 if (vfio_listener_skipped_section(section)) {
375 trace_vfio_listener_region_add_skip(
376 section->offset_within_address_space,
377 section->offset_within_address_space +
378 int128_get64(int128_sub(section->size, int128_one())));
382 if (unlikely((section->offset_within_address_space & ~TARGET_PAGE_MASK) !=
383 (section->offset_within_region & ~TARGET_PAGE_MASK))) {
384 error_report("%s received unaligned region", __func__);
388 iova = TARGET_PAGE_ALIGN(section->offset_within_address_space);
389 llend = int128_make64(section->offset_within_address_space);
390 llend = int128_add(llend, section->size);
391 llend = int128_and(llend, int128_exts64(TARGET_PAGE_MASK));
393 if (int128_ge(int128_make64(iova), llend)) {
396 end = int128_get64(int128_sub(llend, int128_one()));
398 if (container->iommu_type == VFIO_SPAPR_TCE_v2_IOMMU) {
399 VFIOHostDMAWindow *hostwin;
402 /* For now intersections are not allowed, we may relax this later */
403 QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) {
404 if (ranges_overlap(hostwin->min_iova,
405 hostwin->max_iova - hostwin->min_iova + 1,
406 section->offset_within_address_space,
407 int128_get64(section->size))) {
413 ret = vfio_spapr_create_window(container, section, &pgsize);
418 vfio_host_win_add(container, section->offset_within_address_space,
419 section->offset_within_address_space +
420 int128_get64(section->size) - 1, pgsize);
423 hostwin_found = false;
424 QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) {
425 if (hostwin->min_iova <= iova && end <= hostwin->max_iova) {
426 hostwin_found = true;
431 if (!hostwin_found) {
432 error_report("vfio: IOMMU container %p can't map guest IOVA region"
433 " 0x%"HWADDR_PRIx"..0x%"HWADDR_PRIx,
434 container, iova, end);
439 memory_region_ref(section->mr);
441 if (memory_region_is_iommu(section->mr)) {
442 VFIOGuestIOMMU *giommu;
444 trace_vfio_listener_region_add_iommu(iova, end);
446 * FIXME: For VFIO iommu types which have KVM acceleration to
447 * avoid bouncing all map/unmaps through qemu this way, this
448 * would be the right place to wire that up (tell the KVM
449 * device emulation the VFIO iommu handles to use).
451 giommu = g_malloc0(sizeof(*giommu));
452 giommu->iommu = section->mr;
453 giommu->iommu_offset = section->offset_within_address_space -
454 section->offset_within_region;
455 giommu->container = container;
456 giommu->n.notify = vfio_iommu_map_notify;
457 QLIST_INSERT_HEAD(&container->giommu_list, giommu, giommu_next);
459 memory_region_register_iommu_notifier(giommu->iommu, &giommu->n);
460 memory_region_iommu_replay(giommu->iommu, &giommu->n, false);
465 /* Here we assume that memory_region_is_ram(section->mr)==true */
467 vaddr = memory_region_get_ram_ptr(section->mr) +
468 section->offset_within_region +
469 (iova - section->offset_within_address_space);
471 trace_vfio_listener_region_add_ram(iova, end, vaddr);
473 llsize = int128_sub(llend, int128_make64(iova));
475 ret = vfio_dma_map(container, iova, int128_get64(llsize),
476 vaddr, section->readonly);
478 error_report("vfio_dma_map(%p, 0x%"HWADDR_PRIx", "
479 "0x%"HWADDR_PRIx", %p) = %d (%m)",
480 container, iova, int128_get64(llsize), vaddr, ret);
488 * On the initfn path, store the first error in the container so we
489 * can gracefully fail. Runtime, there's not much we can do other
490 * than throw a hardware error.
492 if (!container->initialized) {
493 if (!container->error) {
494 container->error = ret;
497 hw_error("vfio: DMA mapping failed, unable to continue");
501 static void vfio_listener_region_del(MemoryListener *listener,
502 MemoryRegionSection *section)
504 VFIOContainer *container = container_of(listener, VFIOContainer, listener);
506 Int128 llend, llsize;
509 if (vfio_listener_skipped_section(section)) {
510 trace_vfio_listener_region_del_skip(
511 section->offset_within_address_space,
512 section->offset_within_address_space +
513 int128_get64(int128_sub(section->size, int128_one())));
517 if (unlikely((section->offset_within_address_space & ~TARGET_PAGE_MASK) !=
518 (section->offset_within_region & ~TARGET_PAGE_MASK))) {
519 error_report("%s received unaligned region", __func__);
523 if (memory_region_is_iommu(section->mr)) {
524 VFIOGuestIOMMU *giommu;
526 QLIST_FOREACH(giommu, &container->giommu_list, giommu_next) {
527 if (giommu->iommu == section->mr) {
528 memory_region_unregister_iommu_notifier(giommu->iommu,
530 QLIST_REMOVE(giommu, giommu_next);
537 * FIXME: We assume the one big unmap below is adequate to
538 * remove any individual page mappings in the IOMMU which
539 * might have been copied into VFIO. This works for a page table
540 * based IOMMU where a big unmap flattens a large range of IO-PTEs.
541 * That may not be true for all IOMMU types.
545 iova = TARGET_PAGE_ALIGN(section->offset_within_address_space);
546 llend = int128_make64(section->offset_within_address_space);
547 llend = int128_add(llend, section->size);
548 llend = int128_and(llend, int128_exts64(TARGET_PAGE_MASK));
550 if (int128_ge(int128_make64(iova), llend)) {
553 end = int128_get64(int128_sub(llend, int128_one()));
555 llsize = int128_sub(llend, int128_make64(iova));
557 trace_vfio_listener_region_del(iova, end);
559 ret = vfio_dma_unmap(container, iova, int128_get64(llsize));
560 memory_region_unref(section->mr);
562 error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx", "
563 "0x%"HWADDR_PRIx") = %d (%m)",
564 container, iova, int128_get64(llsize), ret);
567 if (container->iommu_type == VFIO_SPAPR_TCE_v2_IOMMU) {
568 vfio_spapr_remove_window(container,
569 section->offset_within_address_space);
570 if (vfio_host_win_del(container,
571 section->offset_within_address_space,
572 section->offset_within_address_space +
573 int128_get64(section->size) - 1) < 0) {
574 hw_error("%s: Cannot delete missing window at %"HWADDR_PRIx,
575 __func__, section->offset_within_address_space);
580 static const MemoryListener vfio_memory_listener = {
581 .region_add = vfio_listener_region_add,
582 .region_del = vfio_listener_region_del,
585 static void vfio_listener_release(VFIOContainer *container)
587 memory_listener_unregister(&container->listener);
588 if (container->iommu_type == VFIO_SPAPR_TCE_v2_IOMMU) {
589 memory_listener_unregister(&container->prereg_listener);
593 static struct vfio_info_cap_header *
594 vfio_get_region_info_cap(struct vfio_region_info *info, uint16_t id)
596 struct vfio_info_cap_header *hdr;
599 if (!(info->flags & VFIO_REGION_INFO_FLAG_CAPS)) {
603 for (hdr = ptr + info->cap_offset; hdr != ptr; hdr = ptr + hdr->next) {
612 static void vfio_setup_region_sparse_mmaps(VFIORegion *region,
613 struct vfio_region_info *info)
615 struct vfio_info_cap_header *hdr;
616 struct vfio_region_info_cap_sparse_mmap *sparse;
619 hdr = vfio_get_region_info_cap(info, VFIO_REGION_INFO_CAP_SPARSE_MMAP);
624 sparse = container_of(hdr, struct vfio_region_info_cap_sparse_mmap, header);
626 trace_vfio_region_sparse_mmap_header(region->vbasedev->name,
627 region->nr, sparse->nr_areas);
629 region->nr_mmaps = sparse->nr_areas;
630 region->mmaps = g_new0(VFIOMmap, region->nr_mmaps);
632 for (i = 0; i < region->nr_mmaps; i++) {
633 region->mmaps[i].offset = sparse->areas[i].offset;
634 region->mmaps[i].size = sparse->areas[i].size;
635 trace_vfio_region_sparse_mmap_entry(i, region->mmaps[i].offset,
636 region->mmaps[i].offset +
637 region->mmaps[i].size);
641 int vfio_region_setup(Object *obj, VFIODevice *vbasedev, VFIORegion *region,
642 int index, const char *name)
644 struct vfio_region_info *info;
647 ret = vfio_get_region_info(vbasedev, index, &info);
652 region->vbasedev = vbasedev;
653 region->flags = info->flags;
654 region->size = info->size;
655 region->fd_offset = info->offset;
659 region->mem = g_new0(MemoryRegion, 1);
660 memory_region_init_io(region->mem, obj, &vfio_region_ops,
661 region, name, region->size);
663 if (!vbasedev->no_mmap &&
664 region->flags & VFIO_REGION_INFO_FLAG_MMAP &&
665 !(region->size & ~qemu_real_host_page_mask)) {
667 vfio_setup_region_sparse_mmaps(region, info);
669 if (!region->nr_mmaps) {
670 region->nr_mmaps = 1;
671 region->mmaps = g_new0(VFIOMmap, region->nr_mmaps);
672 region->mmaps[0].offset = 0;
673 region->mmaps[0].size = region->size;
680 trace_vfio_region_setup(vbasedev->name, index, name,
681 region->flags, region->fd_offset, region->size);
685 int vfio_region_mmap(VFIORegion *region)
694 prot |= region->flags & VFIO_REGION_INFO_FLAG_READ ? PROT_READ : 0;
695 prot |= region->flags & VFIO_REGION_INFO_FLAG_WRITE ? PROT_WRITE : 0;
697 for (i = 0; i < region->nr_mmaps; i++) {
698 region->mmaps[i].mmap = mmap(NULL, region->mmaps[i].size, prot,
699 MAP_SHARED, region->vbasedev->fd,
701 region->mmaps[i].offset);
702 if (region->mmaps[i].mmap == MAP_FAILED) {
705 trace_vfio_region_mmap_fault(memory_region_name(region->mem), i,
707 region->mmaps[i].offset,
709 region->mmaps[i].offset +
710 region->mmaps[i].size - 1, ret);
712 region->mmaps[i].mmap = NULL;
714 for (i--; i >= 0; i--) {
715 memory_region_del_subregion(region->mem, ®ion->mmaps[i].mem);
716 munmap(region->mmaps[i].mmap, region->mmaps[i].size);
717 object_unparent(OBJECT(®ion->mmaps[i].mem));
718 region->mmaps[i].mmap = NULL;
724 name = g_strdup_printf("%s mmaps[%d]",
725 memory_region_name(region->mem), i);
726 memory_region_init_ram_ptr(®ion->mmaps[i].mem,
727 memory_region_owner(region->mem),
728 name, region->mmaps[i].size,
729 region->mmaps[i].mmap);
731 memory_region_set_skip_dump(®ion->mmaps[i].mem);
732 memory_region_add_subregion(region->mem, region->mmaps[i].offset,
733 ®ion->mmaps[i].mem);
735 trace_vfio_region_mmap(memory_region_name(®ion->mmaps[i].mem),
736 region->mmaps[i].offset,
737 region->mmaps[i].offset +
738 region->mmaps[i].size - 1);
744 void vfio_region_exit(VFIORegion *region)
752 for (i = 0; i < region->nr_mmaps; i++) {
753 if (region->mmaps[i].mmap) {
754 memory_region_del_subregion(region->mem, ®ion->mmaps[i].mem);
758 trace_vfio_region_exit(region->vbasedev->name, region->nr);
761 void vfio_region_finalize(VFIORegion *region)
769 for (i = 0; i < region->nr_mmaps; i++) {
770 if (region->mmaps[i].mmap) {
771 munmap(region->mmaps[i].mmap, region->mmaps[i].size);
772 object_unparent(OBJECT(®ion->mmaps[i].mem));
776 object_unparent(OBJECT(region->mem));
779 g_free(region->mmaps);
781 trace_vfio_region_finalize(region->vbasedev->name, region->nr);
784 void vfio_region_mmaps_set_enabled(VFIORegion *region, bool enabled)
792 for (i = 0; i < region->nr_mmaps; i++) {
793 if (region->mmaps[i].mmap) {
794 memory_region_set_enabled(®ion->mmaps[i].mem, enabled);
798 trace_vfio_region_mmaps_set_enabled(memory_region_name(region->mem),
802 void vfio_reset_handler(void *opaque)
805 VFIODevice *vbasedev;
807 QLIST_FOREACH(group, &vfio_group_list, next) {
808 QLIST_FOREACH(vbasedev, &group->device_list, next) {
809 vbasedev->ops->vfio_compute_needs_reset(vbasedev);
813 QLIST_FOREACH(group, &vfio_group_list, next) {
814 QLIST_FOREACH(vbasedev, &group->device_list, next) {
815 if (vbasedev->needs_reset) {
816 vbasedev->ops->vfio_hot_reset_multi(vbasedev);
822 static void vfio_kvm_device_add_group(VFIOGroup *group)
825 struct kvm_device_attr attr = {
826 .group = KVM_DEV_VFIO_GROUP,
827 .attr = KVM_DEV_VFIO_GROUP_ADD,
828 .addr = (uint64_t)(unsigned long)&group->fd,
831 if (!kvm_enabled()) {
835 if (vfio_kvm_device_fd < 0) {
836 struct kvm_create_device cd = {
837 .type = KVM_DEV_TYPE_VFIO,
840 if (kvm_vm_ioctl(kvm_state, KVM_CREATE_DEVICE, &cd)) {
841 error_report("Failed to create KVM VFIO device: %m");
845 vfio_kvm_device_fd = cd.fd;
848 if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) {
849 error_report("Failed to add group %d to KVM VFIO device: %m",
855 static void vfio_kvm_device_del_group(VFIOGroup *group)
858 struct kvm_device_attr attr = {
859 .group = KVM_DEV_VFIO_GROUP,
860 .attr = KVM_DEV_VFIO_GROUP_DEL,
861 .addr = (uint64_t)(unsigned long)&group->fd,
864 if (vfio_kvm_device_fd < 0) {
868 if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) {
869 error_report("Failed to remove group %d from KVM VFIO device: %m",
875 static VFIOAddressSpace *vfio_get_address_space(AddressSpace *as)
877 VFIOAddressSpace *space;
879 QLIST_FOREACH(space, &vfio_address_spaces, list) {
880 if (space->as == as) {
885 /* No suitable VFIOAddressSpace, create a new one */
886 space = g_malloc0(sizeof(*space));
888 QLIST_INIT(&space->containers);
890 QLIST_INSERT_HEAD(&vfio_address_spaces, space, list);
895 static void vfio_put_address_space(VFIOAddressSpace *space)
897 if (QLIST_EMPTY(&space->containers)) {
898 QLIST_REMOVE(space, list);
903 static int vfio_connect_container(VFIOGroup *group, AddressSpace *as)
905 VFIOContainer *container;
907 VFIOAddressSpace *space;
909 space = vfio_get_address_space(as);
911 QLIST_FOREACH(container, &space->containers, next) {
912 if (!ioctl(group->fd, VFIO_GROUP_SET_CONTAINER, &container->fd)) {
913 group->container = container;
914 QLIST_INSERT_HEAD(&container->group_list, group, container_next);
919 fd = qemu_open("/dev/vfio/vfio", O_RDWR);
921 error_report("vfio: failed to open /dev/vfio/vfio: %m");
926 ret = ioctl(fd, VFIO_GET_API_VERSION);
927 if (ret != VFIO_API_VERSION) {
928 error_report("vfio: supported vfio version: %d, "
929 "reported version: %d", VFIO_API_VERSION, ret);
934 container = g_malloc0(sizeof(*container));
935 container->space = space;
937 if (ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_TYPE1_IOMMU) ||
938 ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_TYPE1v2_IOMMU)) {
939 bool v2 = !!ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_TYPE1v2_IOMMU);
940 struct vfio_iommu_type1_info info;
942 ret = ioctl(group->fd, VFIO_GROUP_SET_CONTAINER, &fd);
944 error_report("vfio: failed to set group container: %m");
946 goto free_container_exit;
949 container->iommu_type = v2 ? VFIO_TYPE1v2_IOMMU : VFIO_TYPE1_IOMMU;
950 ret = ioctl(fd, VFIO_SET_IOMMU, container->iommu_type);
952 error_report("vfio: failed to set iommu for container: %m");
954 goto free_container_exit;
958 * FIXME: This assumes that a Type1 IOMMU can map any 64-bit
959 * IOVA whatsoever. That's not actually true, but the current
960 * kernel interface doesn't tell us what it can map, and the
961 * existing Type1 IOMMUs generally support any IOVA we're
962 * going to actually try in practice.
964 info.argsz = sizeof(info);
965 ret = ioctl(fd, VFIO_IOMMU_GET_INFO, &info);
967 if (ret || !(info.flags & VFIO_IOMMU_INFO_PGSIZES)) {
968 /* Assume 4k IOVA page size */
969 info.iova_pgsizes = 4096;
971 vfio_host_win_add(container, 0, (hwaddr)-1, info.iova_pgsizes);
972 } else if (ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_SPAPR_TCE_IOMMU) ||
973 ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_SPAPR_TCE_v2_IOMMU)) {
974 struct vfio_iommu_spapr_tce_info info;
975 bool v2 = !!ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_SPAPR_TCE_v2_IOMMU);
977 ret = ioctl(group->fd, VFIO_GROUP_SET_CONTAINER, &fd);
979 error_report("vfio: failed to set group container: %m");
981 goto free_container_exit;
983 container->iommu_type =
984 v2 ? VFIO_SPAPR_TCE_v2_IOMMU : VFIO_SPAPR_TCE_IOMMU;
985 ret = ioctl(fd, VFIO_SET_IOMMU, container->iommu_type);
987 error_report("vfio: failed to set iommu for container: %m");
989 goto free_container_exit;
993 * The host kernel code implementing VFIO_IOMMU_DISABLE is called
994 * when container fd is closed so we do not call it explicitly
998 ret = ioctl(fd, VFIO_IOMMU_ENABLE);
1000 error_report("vfio: failed to enable container: %m");
1002 goto free_container_exit;
1005 container->prereg_listener = vfio_prereg_listener;
1007 memory_listener_register(&container->prereg_listener,
1008 &address_space_memory);
1009 if (container->error) {
1010 memory_listener_unregister(&container->prereg_listener);
1011 error_report("vfio: RAM memory listener initialization failed for container");
1012 goto free_container_exit;
1016 info.argsz = sizeof(info);
1017 ret = ioctl(fd, VFIO_IOMMU_SPAPR_TCE_GET_INFO, &info);
1019 error_report("vfio: VFIO_IOMMU_SPAPR_TCE_GET_INFO failed: %m");
1022 memory_listener_unregister(&container->prereg_listener);
1024 goto free_container_exit;
1029 * There is a default window in just created container.
1030 * To make region_add/del simpler, we better remove this
1031 * window now and let those iommu_listener callbacks
1032 * create/remove them when needed.
1034 ret = vfio_spapr_remove_window(container, info.dma32_window_start);
1036 goto free_container_exit;
1039 /* The default table uses 4K pages */
1040 vfio_host_win_add(container, info.dma32_window_start,
1041 info.dma32_window_start +
1042 info.dma32_window_size - 1,
1046 error_report("vfio: No available IOMMU models");
1048 goto free_container_exit;
1051 container->listener = vfio_memory_listener;
1053 memory_listener_register(&container->listener, container->space->as);
1055 if (container->error) {
1056 ret = container->error;
1057 error_report("vfio: memory listener initialization failed for container");
1058 goto listener_release_exit;
1061 container->initialized = true;
1063 QLIST_INIT(&container->group_list);
1064 QLIST_INSERT_HEAD(&space->containers, container, next);
1066 group->container = container;
1067 QLIST_INSERT_HEAD(&container->group_list, group, container_next);
1070 listener_release_exit:
1071 vfio_listener_release(container);
1073 free_container_exit:
1080 vfio_put_address_space(space);
1085 static void vfio_disconnect_container(VFIOGroup *group)
1087 VFIOContainer *container = group->container;
1089 if (ioctl(group->fd, VFIO_GROUP_UNSET_CONTAINER, &container->fd)) {
1090 error_report("vfio: error disconnecting group %d from container",
1094 QLIST_REMOVE(group, container_next);
1095 group->container = NULL;
1097 if (QLIST_EMPTY(&container->group_list)) {
1098 VFIOAddressSpace *space = container->space;
1099 VFIOGuestIOMMU *giommu, *tmp;
1101 vfio_listener_release(container);
1102 QLIST_REMOVE(container, next);
1104 QLIST_FOREACH_SAFE(giommu, &container->giommu_list, giommu_next, tmp) {
1105 memory_region_unregister_iommu_notifier(giommu->iommu, &giommu->n);
1106 QLIST_REMOVE(giommu, giommu_next);
1110 trace_vfio_disconnect_container(container->fd);
1111 close(container->fd);
1114 vfio_put_address_space(space);
1118 VFIOGroup *vfio_get_group(int groupid, AddressSpace *as)
1122 struct vfio_group_status status = { .argsz = sizeof(status) };
1124 QLIST_FOREACH(group, &vfio_group_list, next) {
1125 if (group->groupid == groupid) {
1126 /* Found it. Now is it already in the right context? */
1127 if (group->container->space->as == as) {
1130 error_report("vfio: group %d used in multiple address spaces",
1137 group = g_malloc0(sizeof(*group));
1139 snprintf(path, sizeof(path), "/dev/vfio/%d", groupid);
1140 group->fd = qemu_open(path, O_RDWR);
1141 if (group->fd < 0) {
1142 error_report("vfio: error opening %s: %m", path);
1143 goto free_group_exit;
1146 if (ioctl(group->fd, VFIO_GROUP_GET_STATUS, &status)) {
1147 error_report("vfio: error getting group status: %m");
1151 if (!(status.flags & VFIO_GROUP_FLAGS_VIABLE)) {
1152 error_report("vfio: error, group %d is not viable, please ensure "
1153 "all devices within the iommu_group are bound to their "
1154 "vfio bus driver.", groupid);
1158 group->groupid = groupid;
1159 QLIST_INIT(&group->device_list);
1161 if (vfio_connect_container(group, as)) {
1162 error_report("vfio: failed to setup container for group %d", groupid);
1166 if (QLIST_EMPTY(&vfio_group_list)) {
1167 qemu_register_reset(vfio_reset_handler, NULL);
1170 QLIST_INSERT_HEAD(&vfio_group_list, group, next);
1172 vfio_kvm_device_add_group(group);
1185 void vfio_put_group(VFIOGroup *group)
1187 if (!group || !QLIST_EMPTY(&group->device_list)) {
1191 vfio_kvm_device_del_group(group);
1192 vfio_disconnect_container(group);
1193 QLIST_REMOVE(group, next);
1194 trace_vfio_put_group(group->fd);
1198 if (QLIST_EMPTY(&vfio_group_list)) {
1199 qemu_unregister_reset(vfio_reset_handler, NULL);
1203 int vfio_get_device(VFIOGroup *group, const char *name,
1204 VFIODevice *vbasedev)
1206 struct vfio_device_info dev_info = { .argsz = sizeof(dev_info) };
1209 fd = ioctl(group->fd, VFIO_GROUP_GET_DEVICE_FD, name);
1211 error_report("vfio: error getting device %s from group %d: %m",
1212 name, group->groupid);
1213 error_printf("Verify all devices in group %d are bound to vfio-<bus> "
1214 "or pci-stub and not already in use\n", group->groupid);
1218 ret = ioctl(fd, VFIO_DEVICE_GET_INFO, &dev_info);
1220 error_report("vfio: error getting device info: %m");
1226 vbasedev->group = group;
1227 QLIST_INSERT_HEAD(&group->device_list, vbasedev, next);
1229 vbasedev->num_irqs = dev_info.num_irqs;
1230 vbasedev->num_regions = dev_info.num_regions;
1231 vbasedev->flags = dev_info.flags;
1233 trace_vfio_get_device(name, dev_info.flags, dev_info.num_regions,
1236 vbasedev->reset_works = !!(dev_info.flags & VFIO_DEVICE_FLAGS_RESET);
1240 void vfio_put_base_device(VFIODevice *vbasedev)
1242 if (!vbasedev->group) {
1245 QLIST_REMOVE(vbasedev, next);
1246 vbasedev->group = NULL;
1247 trace_vfio_put_base_device(vbasedev->fd);
1248 close(vbasedev->fd);
1251 int vfio_get_region_info(VFIODevice *vbasedev, int index,
1252 struct vfio_region_info **info)
1254 size_t argsz = sizeof(struct vfio_region_info);
1256 *info = g_malloc0(argsz);
1258 (*info)->index = index;
1260 (*info)->argsz = argsz;
1262 if (ioctl(vbasedev->fd, VFIO_DEVICE_GET_REGION_INFO, *info)) {
1268 if ((*info)->argsz > argsz) {
1269 argsz = (*info)->argsz;
1270 *info = g_realloc(*info, argsz);
1278 int vfio_get_dev_region_info(VFIODevice *vbasedev, uint32_t type,
1279 uint32_t subtype, struct vfio_region_info **info)
1283 for (i = 0; i < vbasedev->num_regions; i++) {
1284 struct vfio_info_cap_header *hdr;
1285 struct vfio_region_info_cap_type *cap_type;
1287 if (vfio_get_region_info(vbasedev, i, info)) {
1291 hdr = vfio_get_region_info_cap(*info, VFIO_REGION_INFO_CAP_TYPE);
1297 cap_type = container_of(hdr, struct vfio_region_info_cap_type, header);
1299 trace_vfio_get_dev_region(vbasedev->name, i,
1300 cap_type->type, cap_type->subtype);
1302 if (cap_type->type == type && cap_type->subtype == subtype) {
1314 * Interfaces for IBM EEH (Enhanced Error Handling)
1316 static bool vfio_eeh_container_ok(VFIOContainer *container)
1319 * As of 2016-03-04 (linux-4.5) the host kernel EEH/VFIO
1320 * implementation is broken if there are multiple groups in a
1321 * container. The hardware works in units of Partitionable
1322 * Endpoints (== IOMMU groups) and the EEH operations naively
1323 * iterate across all groups in the container, without any logic
1324 * to make sure the groups have their state synchronized. For
1325 * certain operations (ENABLE) that might be ok, until an error
1326 * occurs, but for others (GET_STATE) it's clearly broken.
1330 * XXX Once fixed kernels exist, test for them here
1333 if (QLIST_EMPTY(&container->group_list)) {
1337 if (QLIST_NEXT(QLIST_FIRST(&container->group_list), container_next)) {
1344 static int vfio_eeh_container_op(VFIOContainer *container, uint32_t op)
1346 struct vfio_eeh_pe_op pe_op = {
1347 .argsz = sizeof(pe_op),
1352 if (!vfio_eeh_container_ok(container)) {
1353 error_report("vfio/eeh: EEH_PE_OP 0x%x: "
1354 "kernel requires a container with exactly one group", op);
1358 ret = ioctl(container->fd, VFIO_EEH_PE_OP, &pe_op);
1360 error_report("vfio/eeh: EEH_PE_OP 0x%x failed: %m", op);
1367 static VFIOContainer *vfio_eeh_as_container(AddressSpace *as)
1369 VFIOAddressSpace *space = vfio_get_address_space(as);
1370 VFIOContainer *container = NULL;
1372 if (QLIST_EMPTY(&space->containers)) {
1373 /* No containers to act on */
1377 container = QLIST_FIRST(&space->containers);
1379 if (QLIST_NEXT(container, next)) {
1380 /* We don't yet have logic to synchronize EEH state across
1381 * multiple containers */
1387 vfio_put_address_space(space);
1391 bool vfio_eeh_as_ok(AddressSpace *as)
1393 VFIOContainer *container = vfio_eeh_as_container(as);
1395 return (container != NULL) && vfio_eeh_container_ok(container);
1398 int vfio_eeh_as_op(AddressSpace *as, uint32_t op)
1400 VFIOContainer *container = vfio_eeh_as_container(as);
1405 return vfio_eeh_container_op(container, op);