sun4u_iommu: update to reflect IOMMU is no longer part of the APB device
[qemu/ar7.git] / hw / vfio / common.c
blobb77be3a8b36becfeb733f82dbded9a645ea2dd40
1 /*
2 * generic functions used by VFIO devices
4 * Copyright Red Hat, Inc. 2012
6 * Authors:
7 * Alex Williamson <alex.williamson@redhat.com>
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.
14 * Copyright (c) 2007, Neocleus, Alex Novik (alex@neocleus.com)
15 * Copyright (c) 2007, Neocleus, Guy Zana (guy@neocleus.com)
16 * Copyright (C) 2008, Qumranet, Amit Shah (amit.shah@qumranet.com)
17 * Copyright (C) 2008, Red Hat, Amit Shah (amit.shah@redhat.com)
18 * Copyright (C) 2008, IBM, Muli Ben-Yehuda (muli@il.ibm.com)
21 #include "qemu/osdep.h"
22 #include <sys/ioctl.h>
23 #ifdef CONFIG_KVM
24 #include <linux/kvm.h>
25 #endif
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"
32 #include "hw/hw.h"
33 #include "qemu/error-report.h"
34 #include "qemu/range.h"
35 #include "sysemu/kvm.h"
36 #include "trace.h"
37 #include "qapi/error.h"
39 struct vfio_group_head vfio_group_list =
40 QLIST_HEAD_INITIALIZER(vfio_group_list);
41 struct vfio_as_head vfio_address_spaces =
42 QLIST_HEAD_INITIALIZER(vfio_address_spaces);
44 #ifdef CONFIG_KVM
46 * We have a single VFIO pseudo device per KVM VM. Once created it lives
47 * for the life of the VM. Closing the file descriptor only drops our
48 * reference to it and the device's reference to kvm. Therefore once
49 * initialized, this file descriptor is only released on QEMU exit and
50 * we'll re-use it should another vfio device be attached before then.
52 static int vfio_kvm_device_fd = -1;
53 #endif
56 * Common VFIO interrupt disable
58 void vfio_disable_irqindex(VFIODevice *vbasedev, int index)
60 struct vfio_irq_set irq_set = {
61 .argsz = sizeof(irq_set),
62 .flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_TRIGGER,
63 .index = index,
64 .start = 0,
65 .count = 0,
68 ioctl(vbasedev->fd, VFIO_DEVICE_SET_IRQS, &irq_set);
71 void vfio_unmask_single_irqindex(VFIODevice *vbasedev, int index)
73 struct vfio_irq_set irq_set = {
74 .argsz = sizeof(irq_set),
75 .flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_UNMASK,
76 .index = index,
77 .start = 0,
78 .count = 1,
81 ioctl(vbasedev->fd, VFIO_DEVICE_SET_IRQS, &irq_set);
84 void vfio_mask_single_irqindex(VFIODevice *vbasedev, int index)
86 struct vfio_irq_set irq_set = {
87 .argsz = sizeof(irq_set),
88 .flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_MASK,
89 .index = index,
90 .start = 0,
91 .count = 1,
94 ioctl(vbasedev->fd, VFIO_DEVICE_SET_IRQS, &irq_set);
98 * IO Port/MMIO - Beware of the endians, VFIO is always little endian
100 void vfio_region_write(void *opaque, hwaddr addr,
101 uint64_t data, unsigned size)
103 VFIORegion *region = opaque;
104 VFIODevice *vbasedev = region->vbasedev;
105 union {
106 uint8_t byte;
107 uint16_t word;
108 uint32_t dword;
109 uint64_t qword;
110 } buf;
112 switch (size) {
113 case 1:
114 buf.byte = data;
115 break;
116 case 2:
117 buf.word = cpu_to_le16(data);
118 break;
119 case 4:
120 buf.dword = cpu_to_le32(data);
121 break;
122 case 8:
123 buf.qword = cpu_to_le64(data);
124 break;
125 default:
126 hw_error("vfio: unsupported write size, %d bytes", size);
127 break;
130 if (pwrite(vbasedev->fd, &buf, size, region->fd_offset + addr) != size) {
131 error_report("%s(%s:region%d+0x%"HWADDR_PRIx", 0x%"PRIx64
132 ",%d) failed: %m",
133 __func__, vbasedev->name, region->nr,
134 addr, data, size);
137 trace_vfio_region_write(vbasedev->name, region->nr, addr, data, size);
140 * A read or write to a BAR always signals an INTx EOI. This will
141 * do nothing if not pending (including not in INTx mode). We assume
142 * that a BAR access is in response to an interrupt and that BAR
143 * accesses will service the interrupt. Unfortunately, we don't know
144 * which access will service the interrupt, so we're potentially
145 * getting quite a few host interrupts per guest interrupt.
147 vbasedev->ops->vfio_eoi(vbasedev);
150 uint64_t vfio_region_read(void *opaque,
151 hwaddr addr, unsigned size)
153 VFIORegion *region = opaque;
154 VFIODevice *vbasedev = region->vbasedev;
155 union {
156 uint8_t byte;
157 uint16_t word;
158 uint32_t dword;
159 uint64_t qword;
160 } buf;
161 uint64_t data = 0;
163 if (pread(vbasedev->fd, &buf, size, region->fd_offset + addr) != size) {
164 error_report("%s(%s:region%d+0x%"HWADDR_PRIx", %d) failed: %m",
165 __func__, vbasedev->name, region->nr,
166 addr, size);
167 return (uint64_t)-1;
169 switch (size) {
170 case 1:
171 data = buf.byte;
172 break;
173 case 2:
174 data = le16_to_cpu(buf.word);
175 break;
176 case 4:
177 data = le32_to_cpu(buf.dword);
178 break;
179 case 8:
180 data = le64_to_cpu(buf.qword);
181 break;
182 default:
183 hw_error("vfio: unsupported read size, %d bytes", size);
184 break;
187 trace_vfio_region_read(vbasedev->name, region->nr, addr, size, data);
189 /* Same as write above */
190 vbasedev->ops->vfio_eoi(vbasedev);
192 return data;
195 const MemoryRegionOps vfio_region_ops = {
196 .read = vfio_region_read,
197 .write = vfio_region_write,
198 .endianness = DEVICE_LITTLE_ENDIAN,
199 .valid = {
200 .min_access_size = 1,
201 .max_access_size = 8,
203 .impl = {
204 .min_access_size = 1,
205 .max_access_size = 8,
210 * DMA - Mapping and unmapping for the "type1" IOMMU interface used on x86
212 static int vfio_dma_unmap(VFIOContainer *container,
213 hwaddr iova, ram_addr_t size)
215 struct vfio_iommu_type1_dma_unmap unmap = {
216 .argsz = sizeof(unmap),
217 .flags = 0,
218 .iova = iova,
219 .size = size,
222 if (ioctl(container->fd, VFIO_IOMMU_UNMAP_DMA, &unmap)) {
223 error_report("VFIO_UNMAP_DMA: %d", -errno);
224 return -errno;
227 return 0;
230 static int vfio_dma_map(VFIOContainer *container, hwaddr iova,
231 ram_addr_t size, void *vaddr, bool readonly)
233 struct vfio_iommu_type1_dma_map map = {
234 .argsz = sizeof(map),
235 .flags = VFIO_DMA_MAP_FLAG_READ,
236 .vaddr = (__u64)(uintptr_t)vaddr,
237 .iova = iova,
238 .size = size,
241 if (!readonly) {
242 map.flags |= VFIO_DMA_MAP_FLAG_WRITE;
246 * Try the mapping, if it fails with EBUSY, unmap the region and try
247 * again. This shouldn't be necessary, but we sometimes see it in
248 * the VGA ROM space.
250 if (ioctl(container->fd, VFIO_IOMMU_MAP_DMA, &map) == 0 ||
251 (errno == EBUSY && vfio_dma_unmap(container, iova, size) == 0 &&
252 ioctl(container->fd, VFIO_IOMMU_MAP_DMA, &map) == 0)) {
253 return 0;
256 error_report("VFIO_MAP_DMA: %d", -errno);
257 return -errno;
260 static void vfio_host_win_add(VFIOContainer *container,
261 hwaddr min_iova, hwaddr max_iova,
262 uint64_t iova_pgsizes)
264 VFIOHostDMAWindow *hostwin;
266 QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) {
267 if (ranges_overlap(hostwin->min_iova,
268 hostwin->max_iova - hostwin->min_iova + 1,
269 min_iova,
270 max_iova - min_iova + 1)) {
271 hw_error("%s: Overlapped IOMMU are not enabled", __func__);
275 hostwin = g_malloc0(sizeof(*hostwin));
277 hostwin->min_iova = min_iova;
278 hostwin->max_iova = max_iova;
279 hostwin->iova_pgsizes = iova_pgsizes;
280 QLIST_INSERT_HEAD(&container->hostwin_list, hostwin, hostwin_next);
283 static int vfio_host_win_del(VFIOContainer *container, hwaddr min_iova,
284 hwaddr max_iova)
286 VFIOHostDMAWindow *hostwin;
288 QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) {
289 if (hostwin->min_iova == min_iova && hostwin->max_iova == max_iova) {
290 QLIST_REMOVE(hostwin, hostwin_next);
291 return 0;
295 return -1;
298 static bool vfio_listener_skipped_section(MemoryRegionSection *section)
300 return (!memory_region_is_ram(section->mr) &&
301 !memory_region_is_iommu(section->mr)) ||
303 * Sizing an enabled 64-bit BAR can cause spurious mappings to
304 * addresses in the upper part of the 64-bit address space. These
305 * are never accessed by the CPU and beyond the address width of
306 * some IOMMU hardware. TODO: VFIO should tell us the IOMMU width.
308 section->offset_within_address_space & (1ULL << 63);
311 /* Called with rcu_read_lock held. */
312 static bool vfio_get_vaddr(IOMMUTLBEntry *iotlb, void **vaddr,
313 bool *read_only)
315 MemoryRegion *mr;
316 hwaddr xlat;
317 hwaddr len = iotlb->addr_mask + 1;
318 bool writable = iotlb->perm & IOMMU_WO;
321 * The IOMMU TLB entry we have just covers translation through
322 * this IOMMU to its immediate target. We need to translate
323 * it the rest of the way through to memory.
325 mr = address_space_translate(&address_space_memory,
326 iotlb->translated_addr,
327 &xlat, &len, writable);
328 if (!memory_region_is_ram(mr)) {
329 error_report("iommu map to non memory area %"HWADDR_PRIx"",
330 xlat);
331 return false;
335 * Translation truncates length to the IOMMU page size,
336 * check that it did not truncate too much.
338 if (len & iotlb->addr_mask) {
339 error_report("iommu has granularity incompatible with target AS");
340 return false;
343 *vaddr = memory_region_get_ram_ptr(mr) + xlat;
344 *read_only = !writable || mr->readonly;
346 return true;
349 static void vfio_iommu_map_notify(IOMMUNotifier *n, IOMMUTLBEntry *iotlb)
351 VFIOGuestIOMMU *giommu = container_of(n, VFIOGuestIOMMU, n);
352 VFIOContainer *container = giommu->container;
353 hwaddr iova = iotlb->iova + giommu->iommu_offset;
354 bool read_only;
355 void *vaddr;
356 int ret;
358 trace_vfio_iommu_map_notify(iotlb->perm == IOMMU_NONE ? "UNMAP" : "MAP",
359 iova, iova + iotlb->addr_mask);
361 if (iotlb->target_as != &address_space_memory) {
362 error_report("Wrong target AS \"%s\", only system memory is allowed",
363 iotlb->target_as->name ? iotlb->target_as->name : "none");
364 return;
367 rcu_read_lock();
369 if ((iotlb->perm & IOMMU_RW) != IOMMU_NONE) {
370 if (!vfio_get_vaddr(iotlb, &vaddr, &read_only)) {
371 goto out;
374 * vaddr is only valid until rcu_read_unlock(). But after
375 * vfio_dma_map has set up the mapping the pages will be
376 * pinned by the kernel. This makes sure that the RAM backend
377 * of vaddr will always be there, even if the memory object is
378 * destroyed and its backing memory munmap-ed.
380 ret = vfio_dma_map(container, iova,
381 iotlb->addr_mask + 1, vaddr,
382 read_only);
383 if (ret) {
384 error_report("vfio_dma_map(%p, 0x%"HWADDR_PRIx", "
385 "0x%"HWADDR_PRIx", %p) = %d (%m)",
386 container, iova,
387 iotlb->addr_mask + 1, vaddr, ret);
389 } else {
390 ret = vfio_dma_unmap(container, iova, iotlb->addr_mask + 1);
391 if (ret) {
392 error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx", "
393 "0x%"HWADDR_PRIx") = %d (%m)",
394 container, iova,
395 iotlb->addr_mask + 1, ret);
398 out:
399 rcu_read_unlock();
402 static void vfio_listener_region_add(MemoryListener *listener,
403 MemoryRegionSection *section)
405 VFIOContainer *container = container_of(listener, VFIOContainer, listener);
406 hwaddr iova, end;
407 Int128 llend, llsize;
408 void *vaddr;
409 int ret;
410 VFIOHostDMAWindow *hostwin;
411 bool hostwin_found;
413 if (vfio_listener_skipped_section(section)) {
414 trace_vfio_listener_region_add_skip(
415 section->offset_within_address_space,
416 section->offset_within_address_space +
417 int128_get64(int128_sub(section->size, int128_one())));
418 return;
421 if (unlikely((section->offset_within_address_space & ~TARGET_PAGE_MASK) !=
422 (section->offset_within_region & ~TARGET_PAGE_MASK))) {
423 error_report("%s received unaligned region", __func__);
424 return;
427 iova = TARGET_PAGE_ALIGN(section->offset_within_address_space);
428 llend = int128_make64(section->offset_within_address_space);
429 llend = int128_add(llend, section->size);
430 llend = int128_and(llend, int128_exts64(TARGET_PAGE_MASK));
432 if (int128_ge(int128_make64(iova), llend)) {
433 return;
435 end = int128_get64(int128_sub(llend, int128_one()));
437 if (container->iommu_type == VFIO_SPAPR_TCE_v2_IOMMU) {
438 VFIOHostDMAWindow *hostwin;
439 hwaddr pgsize = 0;
441 /* For now intersections are not allowed, we may relax this later */
442 QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) {
443 if (ranges_overlap(hostwin->min_iova,
444 hostwin->max_iova - hostwin->min_iova + 1,
445 section->offset_within_address_space,
446 int128_get64(section->size))) {
447 ret = -1;
448 goto fail;
452 ret = vfio_spapr_create_window(container, section, &pgsize);
453 if (ret) {
454 goto fail;
457 vfio_host_win_add(container, section->offset_within_address_space,
458 section->offset_within_address_space +
459 int128_get64(section->size) - 1, pgsize);
462 hostwin_found = false;
463 QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) {
464 if (hostwin->min_iova <= iova && end <= hostwin->max_iova) {
465 hostwin_found = true;
466 break;
470 if (!hostwin_found) {
471 error_report("vfio: IOMMU container %p can't map guest IOVA region"
472 " 0x%"HWADDR_PRIx"..0x%"HWADDR_PRIx,
473 container, iova, end);
474 ret = -EFAULT;
475 goto fail;
478 memory_region_ref(section->mr);
480 if (memory_region_is_iommu(section->mr)) {
481 VFIOGuestIOMMU *giommu;
482 IOMMUMemoryRegion *iommu_mr = IOMMU_MEMORY_REGION(section->mr);
484 trace_vfio_listener_region_add_iommu(iova, end);
486 * FIXME: For VFIO iommu types which have KVM acceleration to
487 * avoid bouncing all map/unmaps through qemu this way, this
488 * would be the right place to wire that up (tell the KVM
489 * device emulation the VFIO iommu handles to use).
491 giommu = g_malloc0(sizeof(*giommu));
492 giommu->iommu = iommu_mr;
493 giommu->iommu_offset = section->offset_within_address_space -
494 section->offset_within_region;
495 giommu->container = container;
496 llend = int128_add(int128_make64(section->offset_within_region),
497 section->size);
498 llend = int128_sub(llend, int128_one());
499 iommu_notifier_init(&giommu->n, vfio_iommu_map_notify,
500 IOMMU_NOTIFIER_ALL,
501 section->offset_within_region,
502 int128_get64(llend));
503 QLIST_INSERT_HEAD(&container->giommu_list, giommu, giommu_next);
505 memory_region_register_iommu_notifier(section->mr, &giommu->n);
506 memory_region_iommu_replay(giommu->iommu, &giommu->n);
508 return;
511 /* Here we assume that memory_region_is_ram(section->mr)==true */
513 vaddr = memory_region_get_ram_ptr(section->mr) +
514 section->offset_within_region +
515 (iova - section->offset_within_address_space);
517 trace_vfio_listener_region_add_ram(iova, end, vaddr);
519 llsize = int128_sub(llend, int128_make64(iova));
521 ret = vfio_dma_map(container, iova, int128_get64(llsize),
522 vaddr, section->readonly);
523 if (ret) {
524 error_report("vfio_dma_map(%p, 0x%"HWADDR_PRIx", "
525 "0x%"HWADDR_PRIx", %p) = %d (%m)",
526 container, iova, int128_get64(llsize), vaddr, ret);
527 goto fail;
530 return;
532 fail:
534 * On the initfn path, store the first error in the container so we
535 * can gracefully fail. Runtime, there's not much we can do other
536 * than throw a hardware error.
538 if (!container->initialized) {
539 if (!container->error) {
540 container->error = ret;
542 } else {
543 hw_error("vfio: DMA mapping failed, unable to continue");
547 static void vfio_listener_region_del(MemoryListener *listener,
548 MemoryRegionSection *section)
550 VFIOContainer *container = container_of(listener, VFIOContainer, listener);
551 hwaddr iova, end;
552 Int128 llend, llsize;
553 int ret;
555 if (vfio_listener_skipped_section(section)) {
556 trace_vfio_listener_region_del_skip(
557 section->offset_within_address_space,
558 section->offset_within_address_space +
559 int128_get64(int128_sub(section->size, int128_one())));
560 return;
563 if (unlikely((section->offset_within_address_space & ~TARGET_PAGE_MASK) !=
564 (section->offset_within_region & ~TARGET_PAGE_MASK))) {
565 error_report("%s received unaligned region", __func__);
566 return;
569 if (memory_region_is_iommu(section->mr)) {
570 VFIOGuestIOMMU *giommu;
572 QLIST_FOREACH(giommu, &container->giommu_list, giommu_next) {
573 if (MEMORY_REGION(giommu->iommu) == section->mr &&
574 giommu->n.start == section->offset_within_region) {
575 memory_region_unregister_iommu_notifier(section->mr,
576 &giommu->n);
577 QLIST_REMOVE(giommu, giommu_next);
578 g_free(giommu);
579 break;
584 * FIXME: We assume the one big unmap below is adequate to
585 * remove any individual page mappings in the IOMMU which
586 * might have been copied into VFIO. This works for a page table
587 * based IOMMU where a big unmap flattens a large range of IO-PTEs.
588 * That may not be true for all IOMMU types.
592 iova = TARGET_PAGE_ALIGN(section->offset_within_address_space);
593 llend = int128_make64(section->offset_within_address_space);
594 llend = int128_add(llend, section->size);
595 llend = int128_and(llend, int128_exts64(TARGET_PAGE_MASK));
597 if (int128_ge(int128_make64(iova), llend)) {
598 return;
600 end = int128_get64(int128_sub(llend, int128_one()));
602 llsize = int128_sub(llend, int128_make64(iova));
604 trace_vfio_listener_region_del(iova, end);
606 ret = vfio_dma_unmap(container, iova, int128_get64(llsize));
607 memory_region_unref(section->mr);
608 if (ret) {
609 error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx", "
610 "0x%"HWADDR_PRIx") = %d (%m)",
611 container, iova, int128_get64(llsize), ret);
614 if (container->iommu_type == VFIO_SPAPR_TCE_v2_IOMMU) {
615 vfio_spapr_remove_window(container,
616 section->offset_within_address_space);
617 if (vfio_host_win_del(container,
618 section->offset_within_address_space,
619 section->offset_within_address_space +
620 int128_get64(section->size) - 1) < 0) {
621 hw_error("%s: Cannot delete missing window at %"HWADDR_PRIx,
622 __func__, section->offset_within_address_space);
627 static const MemoryListener vfio_memory_listener = {
628 .region_add = vfio_listener_region_add,
629 .region_del = vfio_listener_region_del,
632 static void vfio_listener_release(VFIOContainer *container)
634 memory_listener_unregister(&container->listener);
635 if (container->iommu_type == VFIO_SPAPR_TCE_v2_IOMMU) {
636 memory_listener_unregister(&container->prereg_listener);
640 static struct vfio_info_cap_header *
641 vfio_get_region_info_cap(struct vfio_region_info *info, uint16_t id)
643 struct vfio_info_cap_header *hdr;
644 void *ptr = info;
646 if (!(info->flags & VFIO_REGION_INFO_FLAG_CAPS)) {
647 return NULL;
650 for (hdr = ptr + info->cap_offset; hdr != ptr; hdr = ptr + hdr->next) {
651 if (hdr->id == id) {
652 return hdr;
656 return NULL;
659 static int vfio_setup_region_sparse_mmaps(VFIORegion *region,
660 struct vfio_region_info *info)
662 struct vfio_info_cap_header *hdr;
663 struct vfio_region_info_cap_sparse_mmap *sparse;
664 int i, j;
666 hdr = vfio_get_region_info_cap(info, VFIO_REGION_INFO_CAP_SPARSE_MMAP);
667 if (!hdr) {
668 return -ENODEV;
671 sparse = container_of(hdr, struct vfio_region_info_cap_sparse_mmap, header);
673 trace_vfio_region_sparse_mmap_header(region->vbasedev->name,
674 region->nr, sparse->nr_areas);
676 region->mmaps = g_new0(VFIOMmap, sparse->nr_areas);
678 for (i = 0, j = 0; i < sparse->nr_areas; i++) {
679 trace_vfio_region_sparse_mmap_entry(i, sparse->areas[i].offset,
680 sparse->areas[i].offset +
681 sparse->areas[i].size);
683 if (sparse->areas[i].size) {
684 region->mmaps[j].offset = sparse->areas[i].offset;
685 region->mmaps[j].size = sparse->areas[i].size;
686 j++;
690 region->nr_mmaps = j;
691 region->mmaps = g_realloc(region->mmaps, j * sizeof(VFIOMmap));
693 return 0;
696 int vfio_region_setup(Object *obj, VFIODevice *vbasedev, VFIORegion *region,
697 int index, const char *name)
699 struct vfio_region_info *info;
700 int ret;
702 ret = vfio_get_region_info(vbasedev, index, &info);
703 if (ret) {
704 return ret;
707 region->vbasedev = vbasedev;
708 region->flags = info->flags;
709 region->size = info->size;
710 region->fd_offset = info->offset;
711 region->nr = index;
713 if (region->size) {
714 region->mem = g_new0(MemoryRegion, 1);
715 memory_region_init_io(region->mem, obj, &vfio_region_ops,
716 region, name, region->size);
718 if (!vbasedev->no_mmap &&
719 region->flags & VFIO_REGION_INFO_FLAG_MMAP) {
721 ret = vfio_setup_region_sparse_mmaps(region, info);
723 if (ret) {
724 region->nr_mmaps = 1;
725 region->mmaps = g_new0(VFIOMmap, region->nr_mmaps);
726 region->mmaps[0].offset = 0;
727 region->mmaps[0].size = region->size;
732 g_free(info);
734 trace_vfio_region_setup(vbasedev->name, index, name,
735 region->flags, region->fd_offset, region->size);
736 return 0;
739 int vfio_region_mmap(VFIORegion *region)
741 int i, prot = 0;
742 char *name;
744 if (!region->mem) {
745 return 0;
748 prot |= region->flags & VFIO_REGION_INFO_FLAG_READ ? PROT_READ : 0;
749 prot |= region->flags & VFIO_REGION_INFO_FLAG_WRITE ? PROT_WRITE : 0;
751 for (i = 0; i < region->nr_mmaps; i++) {
752 region->mmaps[i].mmap = mmap(NULL, region->mmaps[i].size, prot,
753 MAP_SHARED, region->vbasedev->fd,
754 region->fd_offset +
755 region->mmaps[i].offset);
756 if (region->mmaps[i].mmap == MAP_FAILED) {
757 int ret = -errno;
759 trace_vfio_region_mmap_fault(memory_region_name(region->mem), i,
760 region->fd_offset +
761 region->mmaps[i].offset,
762 region->fd_offset +
763 region->mmaps[i].offset +
764 region->mmaps[i].size - 1, ret);
766 region->mmaps[i].mmap = NULL;
768 for (i--; i >= 0; i--) {
769 memory_region_del_subregion(region->mem, &region->mmaps[i].mem);
770 munmap(region->mmaps[i].mmap, region->mmaps[i].size);
771 object_unparent(OBJECT(&region->mmaps[i].mem));
772 region->mmaps[i].mmap = NULL;
775 return ret;
778 name = g_strdup_printf("%s mmaps[%d]",
779 memory_region_name(region->mem), i);
780 memory_region_init_ram_device_ptr(&region->mmaps[i].mem,
781 memory_region_owner(region->mem),
782 name, region->mmaps[i].size,
783 region->mmaps[i].mmap);
784 g_free(name);
785 memory_region_add_subregion(region->mem, region->mmaps[i].offset,
786 &region->mmaps[i].mem);
788 trace_vfio_region_mmap(memory_region_name(&region->mmaps[i].mem),
789 region->mmaps[i].offset,
790 region->mmaps[i].offset +
791 region->mmaps[i].size - 1);
794 return 0;
797 void vfio_region_exit(VFIORegion *region)
799 int i;
801 if (!region->mem) {
802 return;
805 for (i = 0; i < region->nr_mmaps; i++) {
806 if (region->mmaps[i].mmap) {
807 memory_region_del_subregion(region->mem, &region->mmaps[i].mem);
811 trace_vfio_region_exit(region->vbasedev->name, region->nr);
814 void vfio_region_finalize(VFIORegion *region)
816 int i;
818 if (!region->mem) {
819 return;
822 for (i = 0; i < region->nr_mmaps; i++) {
823 if (region->mmaps[i].mmap) {
824 munmap(region->mmaps[i].mmap, region->mmaps[i].size);
825 object_unparent(OBJECT(&region->mmaps[i].mem));
829 object_unparent(OBJECT(region->mem));
831 g_free(region->mem);
832 g_free(region->mmaps);
834 trace_vfio_region_finalize(region->vbasedev->name, region->nr);
837 void vfio_region_mmaps_set_enabled(VFIORegion *region, bool enabled)
839 int i;
841 if (!region->mem) {
842 return;
845 for (i = 0; i < region->nr_mmaps; i++) {
846 if (region->mmaps[i].mmap) {
847 memory_region_set_enabled(&region->mmaps[i].mem, enabled);
851 trace_vfio_region_mmaps_set_enabled(memory_region_name(region->mem),
852 enabled);
855 void vfio_reset_handler(void *opaque)
857 VFIOGroup *group;
858 VFIODevice *vbasedev;
860 QLIST_FOREACH(group, &vfio_group_list, next) {
861 QLIST_FOREACH(vbasedev, &group->device_list, next) {
862 if (vbasedev->dev->realized) {
863 vbasedev->ops->vfio_compute_needs_reset(vbasedev);
868 QLIST_FOREACH(group, &vfio_group_list, next) {
869 QLIST_FOREACH(vbasedev, &group->device_list, next) {
870 if (vbasedev->dev->realized && vbasedev->needs_reset) {
871 vbasedev->ops->vfio_hot_reset_multi(vbasedev);
877 static void vfio_kvm_device_add_group(VFIOGroup *group)
879 #ifdef CONFIG_KVM
880 struct kvm_device_attr attr = {
881 .group = KVM_DEV_VFIO_GROUP,
882 .attr = KVM_DEV_VFIO_GROUP_ADD,
883 .addr = (uint64_t)(unsigned long)&group->fd,
886 if (!kvm_enabled()) {
887 return;
890 if (vfio_kvm_device_fd < 0) {
891 struct kvm_create_device cd = {
892 .type = KVM_DEV_TYPE_VFIO,
895 if (kvm_vm_ioctl(kvm_state, KVM_CREATE_DEVICE, &cd)) {
896 error_report("Failed to create KVM VFIO device: %m");
897 return;
900 vfio_kvm_device_fd = cd.fd;
903 if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) {
904 error_report("Failed to add group %d to KVM VFIO device: %m",
905 group->groupid);
907 #endif
910 static void vfio_kvm_device_del_group(VFIOGroup *group)
912 #ifdef CONFIG_KVM
913 struct kvm_device_attr attr = {
914 .group = KVM_DEV_VFIO_GROUP,
915 .attr = KVM_DEV_VFIO_GROUP_DEL,
916 .addr = (uint64_t)(unsigned long)&group->fd,
919 if (vfio_kvm_device_fd < 0) {
920 return;
923 if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) {
924 error_report("Failed to remove group %d from KVM VFIO device: %m",
925 group->groupid);
927 #endif
930 static VFIOAddressSpace *vfio_get_address_space(AddressSpace *as)
932 VFIOAddressSpace *space;
934 QLIST_FOREACH(space, &vfio_address_spaces, list) {
935 if (space->as == as) {
936 return space;
940 /* No suitable VFIOAddressSpace, create a new one */
941 space = g_malloc0(sizeof(*space));
942 space->as = as;
943 QLIST_INIT(&space->containers);
945 QLIST_INSERT_HEAD(&vfio_address_spaces, space, list);
947 return space;
950 static void vfio_put_address_space(VFIOAddressSpace *space)
952 if (QLIST_EMPTY(&space->containers)) {
953 QLIST_REMOVE(space, list);
954 g_free(space);
958 static int vfio_connect_container(VFIOGroup *group, AddressSpace *as,
959 Error **errp)
961 VFIOContainer *container;
962 int ret, fd;
963 VFIOAddressSpace *space;
965 space = vfio_get_address_space(as);
967 QLIST_FOREACH(container, &space->containers, next) {
968 if (!ioctl(group->fd, VFIO_GROUP_SET_CONTAINER, &container->fd)) {
969 group->container = container;
970 QLIST_INSERT_HEAD(&container->group_list, group, container_next);
971 vfio_kvm_device_add_group(group);
972 return 0;
976 fd = qemu_open("/dev/vfio/vfio", O_RDWR);
977 if (fd < 0) {
978 error_setg_errno(errp, errno, "failed to open /dev/vfio/vfio");
979 ret = -errno;
980 goto put_space_exit;
983 ret = ioctl(fd, VFIO_GET_API_VERSION);
984 if (ret != VFIO_API_VERSION) {
985 error_setg(errp, "supported vfio version: %d, "
986 "reported version: %d", VFIO_API_VERSION, ret);
987 ret = -EINVAL;
988 goto close_fd_exit;
991 container = g_malloc0(sizeof(*container));
992 container->space = space;
993 container->fd = fd;
994 QLIST_INIT(&container->giommu_list);
995 QLIST_INIT(&container->hostwin_list);
996 if (ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_TYPE1_IOMMU) ||
997 ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_TYPE1v2_IOMMU)) {
998 bool v2 = !!ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_TYPE1v2_IOMMU);
999 struct vfio_iommu_type1_info info;
1001 ret = ioctl(group->fd, VFIO_GROUP_SET_CONTAINER, &fd);
1002 if (ret) {
1003 error_setg_errno(errp, errno, "failed to set group container");
1004 ret = -errno;
1005 goto free_container_exit;
1008 container->iommu_type = v2 ? VFIO_TYPE1v2_IOMMU : VFIO_TYPE1_IOMMU;
1009 ret = ioctl(fd, VFIO_SET_IOMMU, container->iommu_type);
1010 if (ret) {
1011 error_setg_errno(errp, errno, "failed to set iommu for container");
1012 ret = -errno;
1013 goto free_container_exit;
1017 * FIXME: This assumes that a Type1 IOMMU can map any 64-bit
1018 * IOVA whatsoever. That's not actually true, but the current
1019 * kernel interface doesn't tell us what it can map, and the
1020 * existing Type1 IOMMUs generally support any IOVA we're
1021 * going to actually try in practice.
1023 info.argsz = sizeof(info);
1024 ret = ioctl(fd, VFIO_IOMMU_GET_INFO, &info);
1025 /* Ignore errors */
1026 if (ret || !(info.flags & VFIO_IOMMU_INFO_PGSIZES)) {
1027 /* Assume 4k IOVA page size */
1028 info.iova_pgsizes = 4096;
1030 vfio_host_win_add(container, 0, (hwaddr)-1, info.iova_pgsizes);
1031 } else if (ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_SPAPR_TCE_IOMMU) ||
1032 ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_SPAPR_TCE_v2_IOMMU)) {
1033 struct vfio_iommu_spapr_tce_info info;
1034 bool v2 = !!ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_SPAPR_TCE_v2_IOMMU);
1036 ret = ioctl(group->fd, VFIO_GROUP_SET_CONTAINER, &fd);
1037 if (ret) {
1038 error_setg_errno(errp, errno, "failed to set group container");
1039 ret = -errno;
1040 goto free_container_exit;
1042 container->iommu_type =
1043 v2 ? VFIO_SPAPR_TCE_v2_IOMMU : VFIO_SPAPR_TCE_IOMMU;
1044 ret = ioctl(fd, VFIO_SET_IOMMU, container->iommu_type);
1045 if (ret) {
1046 container->iommu_type = VFIO_SPAPR_TCE_IOMMU;
1047 v2 = false;
1048 ret = ioctl(fd, VFIO_SET_IOMMU, container->iommu_type);
1050 if (ret) {
1051 error_setg_errno(errp, errno, "failed to set iommu for container");
1052 ret = -errno;
1053 goto free_container_exit;
1057 * The host kernel code implementing VFIO_IOMMU_DISABLE is called
1058 * when container fd is closed so we do not call it explicitly
1059 * in this file.
1061 if (!v2) {
1062 ret = ioctl(fd, VFIO_IOMMU_ENABLE);
1063 if (ret) {
1064 error_setg_errno(errp, errno, "failed to enable container");
1065 ret = -errno;
1066 goto free_container_exit;
1068 } else {
1069 container->prereg_listener = vfio_prereg_listener;
1071 memory_listener_register(&container->prereg_listener,
1072 &address_space_memory);
1073 if (container->error) {
1074 memory_listener_unregister(&container->prereg_listener);
1075 ret = container->error;
1076 error_setg(errp,
1077 "RAM memory listener initialization failed for container");
1078 goto free_container_exit;
1082 info.argsz = sizeof(info);
1083 ret = ioctl(fd, VFIO_IOMMU_SPAPR_TCE_GET_INFO, &info);
1084 if (ret) {
1085 error_setg_errno(errp, errno,
1086 "VFIO_IOMMU_SPAPR_TCE_GET_INFO failed");
1087 ret = -errno;
1088 if (v2) {
1089 memory_listener_unregister(&container->prereg_listener);
1091 goto free_container_exit;
1094 if (v2) {
1096 * There is a default window in just created container.
1097 * To make region_add/del simpler, we better remove this
1098 * window now and let those iommu_listener callbacks
1099 * create/remove them when needed.
1101 ret = vfio_spapr_remove_window(container, info.dma32_window_start);
1102 if (ret) {
1103 error_setg_errno(errp, -ret,
1104 "failed to remove existing window");
1105 goto free_container_exit;
1107 } else {
1108 /* The default table uses 4K pages */
1109 vfio_host_win_add(container, info.dma32_window_start,
1110 info.dma32_window_start +
1111 info.dma32_window_size - 1,
1112 0x1000);
1114 } else {
1115 error_setg(errp, "No available IOMMU models");
1116 ret = -EINVAL;
1117 goto free_container_exit;
1120 vfio_kvm_device_add_group(group);
1122 QLIST_INIT(&container->group_list);
1123 QLIST_INSERT_HEAD(&space->containers, container, next);
1125 group->container = container;
1126 QLIST_INSERT_HEAD(&container->group_list, group, container_next);
1128 container->listener = vfio_memory_listener;
1130 memory_listener_register(&container->listener, container->space->as);
1132 if (container->error) {
1133 ret = container->error;
1134 error_setg_errno(errp, -ret,
1135 "memory listener initialization failed for container");
1136 goto listener_release_exit;
1139 container->initialized = true;
1141 return 0;
1142 listener_release_exit:
1143 QLIST_REMOVE(group, container_next);
1144 QLIST_REMOVE(container, next);
1145 vfio_kvm_device_del_group(group);
1146 vfio_listener_release(container);
1148 free_container_exit:
1149 g_free(container);
1151 close_fd_exit:
1152 close(fd);
1154 put_space_exit:
1155 vfio_put_address_space(space);
1157 return ret;
1160 static void vfio_disconnect_container(VFIOGroup *group)
1162 VFIOContainer *container = group->container;
1164 if (ioctl(group->fd, VFIO_GROUP_UNSET_CONTAINER, &container->fd)) {
1165 error_report("vfio: error disconnecting group %d from container",
1166 group->groupid);
1169 QLIST_REMOVE(group, container_next);
1170 group->container = NULL;
1172 if (QLIST_EMPTY(&container->group_list)) {
1173 VFIOAddressSpace *space = container->space;
1174 VFIOGuestIOMMU *giommu, *tmp;
1176 vfio_listener_release(container);
1177 QLIST_REMOVE(container, next);
1179 QLIST_FOREACH_SAFE(giommu, &container->giommu_list, giommu_next, tmp) {
1180 memory_region_unregister_iommu_notifier(
1181 MEMORY_REGION(giommu->iommu), &giommu->n);
1182 QLIST_REMOVE(giommu, giommu_next);
1183 g_free(giommu);
1186 trace_vfio_disconnect_container(container->fd);
1187 close(container->fd);
1188 g_free(container);
1190 vfio_put_address_space(space);
1194 VFIOGroup *vfio_get_group(int groupid, AddressSpace *as, Error **errp)
1196 VFIOGroup *group;
1197 char path[32];
1198 struct vfio_group_status status = { .argsz = sizeof(status) };
1200 QLIST_FOREACH(group, &vfio_group_list, next) {
1201 if (group->groupid == groupid) {
1202 /* Found it. Now is it already in the right context? */
1203 if (group->container->space->as == as) {
1204 return group;
1205 } else {
1206 error_setg(errp, "group %d used in multiple address spaces",
1207 group->groupid);
1208 return NULL;
1213 group = g_malloc0(sizeof(*group));
1215 snprintf(path, sizeof(path), "/dev/vfio/%d", groupid);
1216 group->fd = qemu_open(path, O_RDWR);
1217 if (group->fd < 0) {
1218 error_setg_errno(errp, errno, "failed to open %s", path);
1219 goto free_group_exit;
1222 if (ioctl(group->fd, VFIO_GROUP_GET_STATUS, &status)) {
1223 error_setg_errno(errp, errno, "failed to get group %d status", groupid);
1224 goto close_fd_exit;
1227 if (!(status.flags & VFIO_GROUP_FLAGS_VIABLE)) {
1228 error_setg(errp, "group %d is not viable", groupid);
1229 error_append_hint(errp,
1230 "Please ensure all devices within the iommu_group "
1231 "are bound to their vfio bus driver.\n");
1232 goto close_fd_exit;
1235 group->groupid = groupid;
1236 QLIST_INIT(&group->device_list);
1238 if (vfio_connect_container(group, as, errp)) {
1239 error_prepend(errp, "failed to setup container for group %d: ",
1240 groupid);
1241 goto close_fd_exit;
1244 if (QLIST_EMPTY(&vfio_group_list)) {
1245 qemu_register_reset(vfio_reset_handler, NULL);
1248 QLIST_INSERT_HEAD(&vfio_group_list, group, next);
1250 return group;
1252 close_fd_exit:
1253 close(group->fd);
1255 free_group_exit:
1256 g_free(group);
1258 return NULL;
1261 void vfio_put_group(VFIOGroup *group)
1263 if (!group || !QLIST_EMPTY(&group->device_list)) {
1264 return;
1267 vfio_kvm_device_del_group(group);
1268 vfio_disconnect_container(group);
1269 QLIST_REMOVE(group, next);
1270 trace_vfio_put_group(group->fd);
1271 close(group->fd);
1272 g_free(group);
1274 if (QLIST_EMPTY(&vfio_group_list)) {
1275 qemu_unregister_reset(vfio_reset_handler, NULL);
1279 int vfio_get_device(VFIOGroup *group, const char *name,
1280 VFIODevice *vbasedev, Error **errp)
1282 struct vfio_device_info dev_info = { .argsz = sizeof(dev_info) };
1283 int ret, fd;
1285 fd = ioctl(group->fd, VFIO_GROUP_GET_DEVICE_FD, name);
1286 if (fd < 0) {
1287 error_setg_errno(errp, errno, "error getting device from group %d",
1288 group->groupid);
1289 error_append_hint(errp,
1290 "Verify all devices in group %d are bound to vfio-<bus> "
1291 "or pci-stub and not already in use\n", group->groupid);
1292 return fd;
1295 ret = ioctl(fd, VFIO_DEVICE_GET_INFO, &dev_info);
1296 if (ret) {
1297 error_setg_errno(errp, errno, "error getting device info");
1298 close(fd);
1299 return ret;
1302 vbasedev->fd = fd;
1303 vbasedev->group = group;
1304 QLIST_INSERT_HEAD(&group->device_list, vbasedev, next);
1306 vbasedev->num_irqs = dev_info.num_irqs;
1307 vbasedev->num_regions = dev_info.num_regions;
1308 vbasedev->flags = dev_info.flags;
1310 trace_vfio_get_device(name, dev_info.flags, dev_info.num_regions,
1311 dev_info.num_irqs);
1313 vbasedev->reset_works = !!(dev_info.flags & VFIO_DEVICE_FLAGS_RESET);
1314 return 0;
1317 void vfio_put_base_device(VFIODevice *vbasedev)
1319 if (!vbasedev->group) {
1320 return;
1322 QLIST_REMOVE(vbasedev, next);
1323 vbasedev->group = NULL;
1324 trace_vfio_put_base_device(vbasedev->fd);
1325 close(vbasedev->fd);
1328 int vfio_get_region_info(VFIODevice *vbasedev, int index,
1329 struct vfio_region_info **info)
1331 size_t argsz = sizeof(struct vfio_region_info);
1333 *info = g_malloc0(argsz);
1335 (*info)->index = index;
1336 retry:
1337 (*info)->argsz = argsz;
1339 if (ioctl(vbasedev->fd, VFIO_DEVICE_GET_REGION_INFO, *info)) {
1340 g_free(*info);
1341 *info = NULL;
1342 return -errno;
1345 if ((*info)->argsz > argsz) {
1346 argsz = (*info)->argsz;
1347 *info = g_realloc(*info, argsz);
1349 goto retry;
1352 return 0;
1355 int vfio_get_dev_region_info(VFIODevice *vbasedev, uint32_t type,
1356 uint32_t subtype, struct vfio_region_info **info)
1358 int i;
1360 for (i = 0; i < vbasedev->num_regions; i++) {
1361 struct vfio_info_cap_header *hdr;
1362 struct vfio_region_info_cap_type *cap_type;
1364 if (vfio_get_region_info(vbasedev, i, info)) {
1365 continue;
1368 hdr = vfio_get_region_info_cap(*info, VFIO_REGION_INFO_CAP_TYPE);
1369 if (!hdr) {
1370 g_free(*info);
1371 continue;
1374 cap_type = container_of(hdr, struct vfio_region_info_cap_type, header);
1376 trace_vfio_get_dev_region(vbasedev->name, i,
1377 cap_type->type, cap_type->subtype);
1379 if (cap_type->type == type && cap_type->subtype == subtype) {
1380 return 0;
1383 g_free(*info);
1386 *info = NULL;
1387 return -ENODEV;
1391 * Interfaces for IBM EEH (Enhanced Error Handling)
1393 static bool vfio_eeh_container_ok(VFIOContainer *container)
1396 * As of 2016-03-04 (linux-4.5) the host kernel EEH/VFIO
1397 * implementation is broken if there are multiple groups in a
1398 * container. The hardware works in units of Partitionable
1399 * Endpoints (== IOMMU groups) and the EEH operations naively
1400 * iterate across all groups in the container, without any logic
1401 * to make sure the groups have their state synchronized. For
1402 * certain operations (ENABLE) that might be ok, until an error
1403 * occurs, but for others (GET_STATE) it's clearly broken.
1407 * XXX Once fixed kernels exist, test for them here
1410 if (QLIST_EMPTY(&container->group_list)) {
1411 return false;
1414 if (QLIST_NEXT(QLIST_FIRST(&container->group_list), container_next)) {
1415 return false;
1418 return true;
1421 static int vfio_eeh_container_op(VFIOContainer *container, uint32_t op)
1423 struct vfio_eeh_pe_op pe_op = {
1424 .argsz = sizeof(pe_op),
1425 .op = op,
1427 int ret;
1429 if (!vfio_eeh_container_ok(container)) {
1430 error_report("vfio/eeh: EEH_PE_OP 0x%x: "
1431 "kernel requires a container with exactly one group", op);
1432 return -EPERM;
1435 ret = ioctl(container->fd, VFIO_EEH_PE_OP, &pe_op);
1436 if (ret < 0) {
1437 error_report("vfio/eeh: EEH_PE_OP 0x%x failed: %m", op);
1438 return -errno;
1441 return ret;
1444 static VFIOContainer *vfio_eeh_as_container(AddressSpace *as)
1446 VFIOAddressSpace *space = vfio_get_address_space(as);
1447 VFIOContainer *container = NULL;
1449 if (QLIST_EMPTY(&space->containers)) {
1450 /* No containers to act on */
1451 goto out;
1454 container = QLIST_FIRST(&space->containers);
1456 if (QLIST_NEXT(container, next)) {
1457 /* We don't yet have logic to synchronize EEH state across
1458 * multiple containers */
1459 container = NULL;
1460 goto out;
1463 out:
1464 vfio_put_address_space(space);
1465 return container;
1468 bool vfio_eeh_as_ok(AddressSpace *as)
1470 VFIOContainer *container = vfio_eeh_as_container(as);
1472 return (container != NULL) && vfio_eeh_container_ok(container);
1475 int vfio_eeh_as_op(AddressSpace *as, uint32_t op)
1477 VFIOContainer *container = vfio_eeh_as_container(as);
1479 if (!container) {
1480 return -ENODEV;
1482 return vfio_eeh_container_op(container, op);