xref: /openbmc/qemu/hw/vfio/common.c (revision c2b38b27)
1 /*
2  * generic functions used by VFIO devices
3  *
4  * Copyright Red Hat, Inc. 2012
5  *
6  * Authors:
7  *  Alex Williamson <alex.williamson@redhat.com>
8  *
9  * This work is licensed under the terms of the GNU GPL, version 2.  See
10  * the COPYING file in the top-level directory.
11  *
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)
19  */
20 
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>
27 
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"
38 
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);
43 
44 #ifdef CONFIG_KVM
45 /*
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.
51  */
52 static int vfio_kvm_device_fd = -1;
53 #endif
54 
55 /*
56  * Common VFIO interrupt disable
57  */
58 void vfio_disable_irqindex(VFIODevice *vbasedev, int index)
59 {
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,
66     };
67 
68     ioctl(vbasedev->fd, VFIO_DEVICE_SET_IRQS, &irq_set);
69 }
70 
71 void vfio_unmask_single_irqindex(VFIODevice *vbasedev, int index)
72 {
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,
79     };
80 
81     ioctl(vbasedev->fd, VFIO_DEVICE_SET_IRQS, &irq_set);
82 }
83 
84 void vfio_mask_single_irqindex(VFIODevice *vbasedev, int index)
85 {
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,
92     };
93 
94     ioctl(vbasedev->fd, VFIO_DEVICE_SET_IRQS, &irq_set);
95 }
96 
97 /*
98  * IO Port/MMIO - Beware of the endians, VFIO is always little endian
99  */
100 void vfio_region_write(void *opaque, hwaddr addr,
101                        uint64_t data, unsigned size)
102 {
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;
111 
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     default:
123         hw_error("vfio: unsupported write size, %d bytes", size);
124         break;
125     }
126 
127     if (pwrite(vbasedev->fd, &buf, size, region->fd_offset + addr) != size) {
128         error_report("%s(%s:region%d+0x%"HWADDR_PRIx", 0x%"PRIx64
129                      ",%d) failed: %m",
130                      __func__, vbasedev->name, region->nr,
131                      addr, data, size);
132     }
133 
134     trace_vfio_region_write(vbasedev->name, region->nr, addr, data, size);
135 
136     /*
137      * A read or write to a BAR always signals an INTx EOI.  This will
138      * do nothing if not pending (including not in INTx mode).  We assume
139      * that a BAR access is in response to an interrupt and that BAR
140      * accesses will service the interrupt.  Unfortunately, we don't know
141      * which access will service the interrupt, so we're potentially
142      * getting quite a few host interrupts per guest interrupt.
143      */
144     vbasedev->ops->vfio_eoi(vbasedev);
145 }
146 
147 uint64_t vfio_region_read(void *opaque,
148                           hwaddr addr, unsigned size)
149 {
150     VFIORegion *region = opaque;
151     VFIODevice *vbasedev = region->vbasedev;
152     union {
153         uint8_t byte;
154         uint16_t word;
155         uint32_t dword;
156         uint64_t qword;
157     } buf;
158     uint64_t data = 0;
159 
160     if (pread(vbasedev->fd, &buf, size, region->fd_offset + addr) != size) {
161         error_report("%s(%s:region%d+0x%"HWADDR_PRIx", %d) failed: %m",
162                      __func__, vbasedev->name, region->nr,
163                      addr, size);
164         return (uint64_t)-1;
165     }
166     switch (size) {
167     case 1:
168         data = buf.byte;
169         break;
170     case 2:
171         data = le16_to_cpu(buf.word);
172         break;
173     case 4:
174         data = le32_to_cpu(buf.dword);
175         break;
176     default:
177         hw_error("vfio: unsupported read size, %d bytes", size);
178         break;
179     }
180 
181     trace_vfio_region_read(vbasedev->name, region->nr, addr, size, data);
182 
183     /* Same as write above */
184     vbasedev->ops->vfio_eoi(vbasedev);
185 
186     return data;
187 }
188 
189 const MemoryRegionOps vfio_region_ops = {
190     .read = vfio_region_read,
191     .write = vfio_region_write,
192     .endianness = DEVICE_LITTLE_ENDIAN,
193 };
194 
195 /*
196  * DMA - Mapping and unmapping for the "type1" IOMMU interface used on x86
197  */
198 static int vfio_dma_unmap(VFIOContainer *container,
199                           hwaddr iova, ram_addr_t size)
200 {
201     struct vfio_iommu_type1_dma_unmap unmap = {
202         .argsz = sizeof(unmap),
203         .flags = 0,
204         .iova = iova,
205         .size = size,
206     };
207 
208     if (ioctl(container->fd, VFIO_IOMMU_UNMAP_DMA, &unmap)) {
209         error_report("VFIO_UNMAP_DMA: %d", -errno);
210         return -errno;
211     }
212 
213     return 0;
214 }
215 
216 static int vfio_dma_map(VFIOContainer *container, hwaddr iova,
217                         ram_addr_t size, void *vaddr, bool readonly)
218 {
219     struct vfio_iommu_type1_dma_map map = {
220         .argsz = sizeof(map),
221         .flags = VFIO_DMA_MAP_FLAG_READ,
222         .vaddr = (__u64)(uintptr_t)vaddr,
223         .iova = iova,
224         .size = size,
225     };
226 
227     if (!readonly) {
228         map.flags |= VFIO_DMA_MAP_FLAG_WRITE;
229     }
230 
231     /*
232      * Try the mapping, if it fails with EBUSY, unmap the region and try
233      * again.  This shouldn't be necessary, but we sometimes see it in
234      * the VGA ROM space.
235      */
236     if (ioctl(container->fd, VFIO_IOMMU_MAP_DMA, &map) == 0 ||
237         (errno == EBUSY && vfio_dma_unmap(container, iova, size) == 0 &&
238          ioctl(container->fd, VFIO_IOMMU_MAP_DMA, &map) == 0)) {
239         return 0;
240     }
241 
242     error_report("VFIO_MAP_DMA: %d", -errno);
243     return -errno;
244 }
245 
246 static void vfio_host_win_add(VFIOContainer *container,
247                               hwaddr min_iova, hwaddr max_iova,
248                               uint64_t iova_pgsizes)
249 {
250     VFIOHostDMAWindow *hostwin;
251 
252     QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) {
253         if (ranges_overlap(hostwin->min_iova,
254                            hostwin->max_iova - hostwin->min_iova + 1,
255                            min_iova,
256                            max_iova - min_iova + 1)) {
257             hw_error("%s: Overlapped IOMMU are not enabled", __func__);
258         }
259     }
260 
261     hostwin = g_malloc0(sizeof(*hostwin));
262 
263     hostwin->min_iova = min_iova;
264     hostwin->max_iova = max_iova;
265     hostwin->iova_pgsizes = iova_pgsizes;
266     QLIST_INSERT_HEAD(&container->hostwin_list, hostwin, hostwin_next);
267 }
268 
269 static int vfio_host_win_del(VFIOContainer *container, hwaddr min_iova,
270                              hwaddr max_iova)
271 {
272     VFIOHostDMAWindow *hostwin;
273 
274     QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) {
275         if (hostwin->min_iova == min_iova && hostwin->max_iova == max_iova) {
276             QLIST_REMOVE(hostwin, hostwin_next);
277             return 0;
278         }
279     }
280 
281     return -1;
282 }
283 
284 static bool vfio_listener_skipped_section(MemoryRegionSection *section)
285 {
286     return (!memory_region_is_ram(section->mr) &&
287             !memory_region_is_iommu(section->mr)) ||
288            /*
289             * Sizing an enabled 64-bit BAR can cause spurious mappings to
290             * addresses in the upper part of the 64-bit address space.  These
291             * are never accessed by the CPU and beyond the address width of
292             * some IOMMU hardware.  TODO: VFIO should tell us the IOMMU width.
293             */
294            section->offset_within_address_space & (1ULL << 63);
295 }
296 
297 /* Called with rcu_read_lock held.  */
298 static bool vfio_get_vaddr(IOMMUTLBEntry *iotlb, void **vaddr,
299                            bool *read_only)
300 {
301     MemoryRegion *mr;
302     hwaddr xlat;
303     hwaddr len = iotlb->addr_mask + 1;
304     bool writable = iotlb->perm & IOMMU_WO;
305 
306     /*
307      * The IOMMU TLB entry we have just covers translation through
308      * this IOMMU to its immediate target.  We need to translate
309      * it the rest of the way through to memory.
310      */
311     mr = address_space_translate(&address_space_memory,
312                                  iotlb->translated_addr,
313                                  &xlat, &len, writable);
314     if (!memory_region_is_ram(mr)) {
315         error_report("iommu map to non memory area %"HWADDR_PRIx"",
316                      xlat);
317         return false;
318     }
319 
320     /*
321      * Translation truncates length to the IOMMU page size,
322      * check that it did not truncate too much.
323      */
324     if (len & iotlb->addr_mask) {
325         error_report("iommu has granularity incompatible with target AS");
326         return false;
327     }
328 
329     *vaddr = memory_region_get_ram_ptr(mr) + xlat;
330     *read_only = !writable || mr->readonly;
331 
332     return true;
333 }
334 
335 static void vfio_iommu_map_notify(IOMMUNotifier *n, IOMMUTLBEntry *iotlb)
336 {
337     VFIOGuestIOMMU *giommu = container_of(n, VFIOGuestIOMMU, n);
338     VFIOContainer *container = giommu->container;
339     hwaddr iova = iotlb->iova + giommu->iommu_offset;
340     bool read_only;
341     void *vaddr;
342     int ret;
343 
344     trace_vfio_iommu_map_notify(iotlb->perm == IOMMU_NONE ? "UNMAP" : "MAP",
345                                 iova, iova + iotlb->addr_mask);
346 
347     if (iotlb->target_as != &address_space_memory) {
348         error_report("Wrong target AS \"%s\", only system memory is allowed",
349                      iotlb->target_as->name ? iotlb->target_as->name : "none");
350         return;
351     }
352 
353     rcu_read_lock();
354 
355     if ((iotlb->perm & IOMMU_RW) != IOMMU_NONE) {
356         if (!vfio_get_vaddr(iotlb, &vaddr, &read_only)) {
357             goto out;
358         }
359         /*
360          * vaddr is only valid until rcu_read_unlock(). But after
361          * vfio_dma_map has set up the mapping the pages will be
362          * pinned by the kernel. This makes sure that the RAM backend
363          * of vaddr will always be there, even if the memory object is
364          * destroyed and its backing memory munmap-ed.
365          */
366         ret = vfio_dma_map(container, iova,
367                            iotlb->addr_mask + 1, vaddr,
368                            read_only);
369         if (ret) {
370             error_report("vfio_dma_map(%p, 0x%"HWADDR_PRIx", "
371                          "0x%"HWADDR_PRIx", %p) = %d (%m)",
372                          container, iova,
373                          iotlb->addr_mask + 1, vaddr, ret);
374         }
375     } else {
376         ret = vfio_dma_unmap(container, iova, iotlb->addr_mask + 1);
377         if (ret) {
378             error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx", "
379                          "0x%"HWADDR_PRIx") = %d (%m)",
380                          container, iova,
381                          iotlb->addr_mask + 1, ret);
382         }
383     }
384 out:
385     rcu_read_unlock();
386 }
387 
388 static void vfio_listener_region_add(MemoryListener *listener,
389                                      MemoryRegionSection *section)
390 {
391     VFIOContainer *container = container_of(listener, VFIOContainer, listener);
392     hwaddr iova, end;
393     Int128 llend, llsize;
394     void *vaddr;
395     int ret;
396     VFIOHostDMAWindow *hostwin;
397     bool hostwin_found;
398 
399     if (vfio_listener_skipped_section(section)) {
400         trace_vfio_listener_region_add_skip(
401                 section->offset_within_address_space,
402                 section->offset_within_address_space +
403                 int128_get64(int128_sub(section->size, int128_one())));
404         return;
405     }
406 
407     if (unlikely((section->offset_within_address_space & ~TARGET_PAGE_MASK) !=
408                  (section->offset_within_region & ~TARGET_PAGE_MASK))) {
409         error_report("%s received unaligned region", __func__);
410         return;
411     }
412 
413     iova = TARGET_PAGE_ALIGN(section->offset_within_address_space);
414     llend = int128_make64(section->offset_within_address_space);
415     llend = int128_add(llend, section->size);
416     llend = int128_and(llend, int128_exts64(TARGET_PAGE_MASK));
417 
418     if (int128_ge(int128_make64(iova), llend)) {
419         return;
420     }
421     end = int128_get64(int128_sub(llend, int128_one()));
422 
423     if (container->iommu_type == VFIO_SPAPR_TCE_v2_IOMMU) {
424         VFIOHostDMAWindow *hostwin;
425         hwaddr pgsize = 0;
426 
427         /* For now intersections are not allowed, we may relax this later */
428         QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) {
429             if (ranges_overlap(hostwin->min_iova,
430                                hostwin->max_iova - hostwin->min_iova + 1,
431                                section->offset_within_address_space,
432                                int128_get64(section->size))) {
433                 ret = -1;
434                 goto fail;
435             }
436         }
437 
438         ret = vfio_spapr_create_window(container, section, &pgsize);
439         if (ret) {
440             goto fail;
441         }
442 
443         vfio_host_win_add(container, section->offset_within_address_space,
444                           section->offset_within_address_space +
445                           int128_get64(section->size) - 1, pgsize);
446     }
447 
448     hostwin_found = false;
449     QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) {
450         if (hostwin->min_iova <= iova && end <= hostwin->max_iova) {
451             hostwin_found = true;
452             break;
453         }
454     }
455 
456     if (!hostwin_found) {
457         error_report("vfio: IOMMU container %p can't map guest IOVA region"
458                      " 0x%"HWADDR_PRIx"..0x%"HWADDR_PRIx,
459                      container, iova, end);
460         ret = -EFAULT;
461         goto fail;
462     }
463 
464     memory_region_ref(section->mr);
465 
466     if (memory_region_is_iommu(section->mr)) {
467         VFIOGuestIOMMU *giommu;
468 
469         trace_vfio_listener_region_add_iommu(iova, end);
470         /*
471          * FIXME: For VFIO iommu types which have KVM acceleration to
472          * avoid bouncing all map/unmaps through qemu this way, this
473          * would be the right place to wire that up (tell the KVM
474          * device emulation the VFIO iommu handles to use).
475          */
476         giommu = g_malloc0(sizeof(*giommu));
477         giommu->iommu = section->mr;
478         giommu->iommu_offset = section->offset_within_address_space -
479                                section->offset_within_region;
480         giommu->container = container;
481         giommu->n.notify = vfio_iommu_map_notify;
482         giommu->n.notifier_flags = IOMMU_NOTIFIER_ALL;
483         QLIST_INSERT_HEAD(&container->giommu_list, giommu, giommu_next);
484 
485         memory_region_register_iommu_notifier(giommu->iommu, &giommu->n);
486         memory_region_iommu_replay(giommu->iommu, &giommu->n, false);
487 
488         return;
489     }
490 
491     /* Here we assume that memory_region_is_ram(section->mr)==true */
492 
493     vaddr = memory_region_get_ram_ptr(section->mr) +
494             section->offset_within_region +
495             (iova - section->offset_within_address_space);
496 
497     trace_vfio_listener_region_add_ram(iova, end, vaddr);
498 
499     llsize = int128_sub(llend, int128_make64(iova));
500 
501     ret = vfio_dma_map(container, iova, int128_get64(llsize),
502                        vaddr, section->readonly);
503     if (ret) {
504         error_report("vfio_dma_map(%p, 0x%"HWADDR_PRIx", "
505                      "0x%"HWADDR_PRIx", %p) = %d (%m)",
506                      container, iova, int128_get64(llsize), vaddr, ret);
507         goto fail;
508     }
509 
510     return;
511 
512 fail:
513     /*
514      * On the initfn path, store the first error in the container so we
515      * can gracefully fail.  Runtime, there's not much we can do other
516      * than throw a hardware error.
517      */
518     if (!container->initialized) {
519         if (!container->error) {
520             container->error = ret;
521         }
522     } else {
523         hw_error("vfio: DMA mapping failed, unable to continue");
524     }
525 }
526 
527 static void vfio_listener_region_del(MemoryListener *listener,
528                                      MemoryRegionSection *section)
529 {
530     VFIOContainer *container = container_of(listener, VFIOContainer, listener);
531     hwaddr iova, end;
532     Int128 llend, llsize;
533     int ret;
534 
535     if (vfio_listener_skipped_section(section)) {
536         trace_vfio_listener_region_del_skip(
537                 section->offset_within_address_space,
538                 section->offset_within_address_space +
539                 int128_get64(int128_sub(section->size, int128_one())));
540         return;
541     }
542 
543     if (unlikely((section->offset_within_address_space & ~TARGET_PAGE_MASK) !=
544                  (section->offset_within_region & ~TARGET_PAGE_MASK))) {
545         error_report("%s received unaligned region", __func__);
546         return;
547     }
548 
549     if (memory_region_is_iommu(section->mr)) {
550         VFIOGuestIOMMU *giommu;
551 
552         QLIST_FOREACH(giommu, &container->giommu_list, giommu_next) {
553             if (giommu->iommu == section->mr) {
554                 memory_region_unregister_iommu_notifier(giommu->iommu,
555                                                         &giommu->n);
556                 QLIST_REMOVE(giommu, giommu_next);
557                 g_free(giommu);
558                 break;
559             }
560         }
561 
562         /*
563          * FIXME: We assume the one big unmap below is adequate to
564          * remove any individual page mappings in the IOMMU which
565          * might have been copied into VFIO. This works for a page table
566          * based IOMMU where a big unmap flattens a large range of IO-PTEs.
567          * That may not be true for all IOMMU types.
568          */
569     }
570 
571     iova = TARGET_PAGE_ALIGN(section->offset_within_address_space);
572     llend = int128_make64(section->offset_within_address_space);
573     llend = int128_add(llend, section->size);
574     llend = int128_and(llend, int128_exts64(TARGET_PAGE_MASK));
575 
576     if (int128_ge(int128_make64(iova), llend)) {
577         return;
578     }
579     end = int128_get64(int128_sub(llend, int128_one()));
580 
581     llsize = int128_sub(llend, int128_make64(iova));
582 
583     trace_vfio_listener_region_del(iova, end);
584 
585     ret = vfio_dma_unmap(container, iova, int128_get64(llsize));
586     memory_region_unref(section->mr);
587     if (ret) {
588         error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx", "
589                      "0x%"HWADDR_PRIx") = %d (%m)",
590                      container, iova, int128_get64(llsize), ret);
591     }
592 
593     if (container->iommu_type == VFIO_SPAPR_TCE_v2_IOMMU) {
594         vfio_spapr_remove_window(container,
595                                  section->offset_within_address_space);
596         if (vfio_host_win_del(container,
597                               section->offset_within_address_space,
598                               section->offset_within_address_space +
599                               int128_get64(section->size) - 1) < 0) {
600             hw_error("%s: Cannot delete missing window at %"HWADDR_PRIx,
601                      __func__, section->offset_within_address_space);
602         }
603     }
604 }
605 
606 static const MemoryListener vfio_memory_listener = {
607     .region_add = vfio_listener_region_add,
608     .region_del = vfio_listener_region_del,
609 };
610 
611 static void vfio_listener_release(VFIOContainer *container)
612 {
613     memory_listener_unregister(&container->listener);
614     if (container->iommu_type == VFIO_SPAPR_TCE_v2_IOMMU) {
615         memory_listener_unregister(&container->prereg_listener);
616     }
617 }
618 
619 static struct vfio_info_cap_header *
620 vfio_get_region_info_cap(struct vfio_region_info *info, uint16_t id)
621 {
622     struct vfio_info_cap_header *hdr;
623     void *ptr = info;
624 
625     if (!(info->flags & VFIO_REGION_INFO_FLAG_CAPS)) {
626         return NULL;
627     }
628 
629     for (hdr = ptr + info->cap_offset; hdr != ptr; hdr = ptr + hdr->next) {
630         if (hdr->id == id) {
631             return hdr;
632         }
633     }
634 
635     return NULL;
636 }
637 
638 static int vfio_setup_region_sparse_mmaps(VFIORegion *region,
639                                           struct vfio_region_info *info)
640 {
641     struct vfio_info_cap_header *hdr;
642     struct vfio_region_info_cap_sparse_mmap *sparse;
643     int i, j;
644 
645     hdr = vfio_get_region_info_cap(info, VFIO_REGION_INFO_CAP_SPARSE_MMAP);
646     if (!hdr) {
647         return -ENODEV;
648     }
649 
650     sparse = container_of(hdr, struct vfio_region_info_cap_sparse_mmap, header);
651 
652     trace_vfio_region_sparse_mmap_header(region->vbasedev->name,
653                                          region->nr, sparse->nr_areas);
654 
655     region->mmaps = g_new0(VFIOMmap, sparse->nr_areas);
656 
657     for (i = 0, j = 0; i < sparse->nr_areas; i++) {
658         trace_vfio_region_sparse_mmap_entry(i, sparse->areas[i].offset,
659                                             sparse->areas[i].offset +
660                                             sparse->areas[i].size);
661 
662         if (sparse->areas[i].size) {
663             region->mmaps[j].offset = sparse->areas[i].offset;
664             region->mmaps[j].size = sparse->areas[i].size;
665             j++;
666         }
667     }
668 
669     region->nr_mmaps = j;
670     region->mmaps = g_realloc(region->mmaps, j * sizeof(VFIOMmap));
671 
672     return 0;
673 }
674 
675 int vfio_region_setup(Object *obj, VFIODevice *vbasedev, VFIORegion *region,
676                       int index, const char *name)
677 {
678     struct vfio_region_info *info;
679     int ret;
680 
681     ret = vfio_get_region_info(vbasedev, index, &info);
682     if (ret) {
683         return ret;
684     }
685 
686     region->vbasedev = vbasedev;
687     region->flags = info->flags;
688     region->size = info->size;
689     region->fd_offset = info->offset;
690     region->nr = index;
691 
692     if (region->size) {
693         region->mem = g_new0(MemoryRegion, 1);
694         memory_region_init_io(region->mem, obj, &vfio_region_ops,
695                               region, name, region->size);
696 
697         if (!vbasedev->no_mmap &&
698             region->flags & VFIO_REGION_INFO_FLAG_MMAP) {
699 
700             ret = vfio_setup_region_sparse_mmaps(region, info);
701 
702             if (ret) {
703                 region->nr_mmaps = 1;
704                 region->mmaps = g_new0(VFIOMmap, region->nr_mmaps);
705                 region->mmaps[0].offset = 0;
706                 region->mmaps[0].size = region->size;
707             }
708         }
709     }
710 
711     g_free(info);
712 
713     trace_vfio_region_setup(vbasedev->name, index, name,
714                             region->flags, region->fd_offset, region->size);
715     return 0;
716 }
717 
718 int vfio_region_mmap(VFIORegion *region)
719 {
720     int i, prot = 0;
721     char *name;
722 
723     if (!region->mem) {
724         return 0;
725     }
726 
727     prot |= region->flags & VFIO_REGION_INFO_FLAG_READ ? PROT_READ : 0;
728     prot |= region->flags & VFIO_REGION_INFO_FLAG_WRITE ? PROT_WRITE : 0;
729 
730     for (i = 0; i < region->nr_mmaps; i++) {
731         region->mmaps[i].mmap = mmap(NULL, region->mmaps[i].size, prot,
732                                      MAP_SHARED, region->vbasedev->fd,
733                                      region->fd_offset +
734                                      region->mmaps[i].offset);
735         if (region->mmaps[i].mmap == MAP_FAILED) {
736             int ret = -errno;
737 
738             trace_vfio_region_mmap_fault(memory_region_name(region->mem), i,
739                                          region->fd_offset +
740                                          region->mmaps[i].offset,
741                                          region->fd_offset +
742                                          region->mmaps[i].offset +
743                                          region->mmaps[i].size - 1, ret);
744 
745             region->mmaps[i].mmap = NULL;
746 
747             for (i--; i >= 0; i--) {
748                 memory_region_del_subregion(region->mem, &region->mmaps[i].mem);
749                 munmap(region->mmaps[i].mmap, region->mmaps[i].size);
750                 object_unparent(OBJECT(&region->mmaps[i].mem));
751                 region->mmaps[i].mmap = NULL;
752             }
753 
754             return ret;
755         }
756 
757         name = g_strdup_printf("%s mmaps[%d]",
758                                memory_region_name(region->mem), i);
759         memory_region_init_ram_device_ptr(&region->mmaps[i].mem,
760                                           memory_region_owner(region->mem),
761                                           name, region->mmaps[i].size,
762                                           region->mmaps[i].mmap);
763         g_free(name);
764         memory_region_add_subregion(region->mem, region->mmaps[i].offset,
765                                     &region->mmaps[i].mem);
766 
767         trace_vfio_region_mmap(memory_region_name(&region->mmaps[i].mem),
768                                region->mmaps[i].offset,
769                                region->mmaps[i].offset +
770                                region->mmaps[i].size - 1);
771     }
772 
773     return 0;
774 }
775 
776 void vfio_region_exit(VFIORegion *region)
777 {
778     int i;
779 
780     if (!region->mem) {
781         return;
782     }
783 
784     for (i = 0; i < region->nr_mmaps; i++) {
785         if (region->mmaps[i].mmap) {
786             memory_region_del_subregion(region->mem, &region->mmaps[i].mem);
787         }
788     }
789 
790     trace_vfio_region_exit(region->vbasedev->name, region->nr);
791 }
792 
793 void vfio_region_finalize(VFIORegion *region)
794 {
795     int i;
796 
797     if (!region->mem) {
798         return;
799     }
800 
801     for (i = 0; i < region->nr_mmaps; i++) {
802         if (region->mmaps[i].mmap) {
803             munmap(region->mmaps[i].mmap, region->mmaps[i].size);
804             object_unparent(OBJECT(&region->mmaps[i].mem));
805         }
806     }
807 
808     object_unparent(OBJECT(region->mem));
809 
810     g_free(region->mem);
811     g_free(region->mmaps);
812 
813     trace_vfio_region_finalize(region->vbasedev->name, region->nr);
814 }
815 
816 void vfio_region_mmaps_set_enabled(VFIORegion *region, bool enabled)
817 {
818     int i;
819 
820     if (!region->mem) {
821         return;
822     }
823 
824     for (i = 0; i < region->nr_mmaps; i++) {
825         if (region->mmaps[i].mmap) {
826             memory_region_set_enabled(&region->mmaps[i].mem, enabled);
827         }
828     }
829 
830     trace_vfio_region_mmaps_set_enabled(memory_region_name(region->mem),
831                                         enabled);
832 }
833 
834 void vfio_reset_handler(void *opaque)
835 {
836     VFIOGroup *group;
837     VFIODevice *vbasedev;
838 
839     QLIST_FOREACH(group, &vfio_group_list, next) {
840         QLIST_FOREACH(vbasedev, &group->device_list, next) {
841             vbasedev->ops->vfio_compute_needs_reset(vbasedev);
842         }
843     }
844 
845     QLIST_FOREACH(group, &vfio_group_list, next) {
846         QLIST_FOREACH(vbasedev, &group->device_list, next) {
847             if (vbasedev->needs_reset) {
848                 vbasedev->ops->vfio_hot_reset_multi(vbasedev);
849             }
850         }
851     }
852 }
853 
854 static void vfio_kvm_device_add_group(VFIOGroup *group)
855 {
856 #ifdef CONFIG_KVM
857     struct kvm_device_attr attr = {
858         .group = KVM_DEV_VFIO_GROUP,
859         .attr = KVM_DEV_VFIO_GROUP_ADD,
860         .addr = (uint64_t)(unsigned long)&group->fd,
861     };
862 
863     if (!kvm_enabled()) {
864         return;
865     }
866 
867     if (vfio_kvm_device_fd < 0) {
868         struct kvm_create_device cd = {
869             .type = KVM_DEV_TYPE_VFIO,
870         };
871 
872         if (kvm_vm_ioctl(kvm_state, KVM_CREATE_DEVICE, &cd)) {
873             error_report("Failed to create KVM VFIO device: %m");
874             return;
875         }
876 
877         vfio_kvm_device_fd = cd.fd;
878     }
879 
880     if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) {
881         error_report("Failed to add group %d to KVM VFIO device: %m",
882                      group->groupid);
883     }
884 #endif
885 }
886 
887 static void vfio_kvm_device_del_group(VFIOGroup *group)
888 {
889 #ifdef CONFIG_KVM
890     struct kvm_device_attr attr = {
891         .group = KVM_DEV_VFIO_GROUP,
892         .attr = KVM_DEV_VFIO_GROUP_DEL,
893         .addr = (uint64_t)(unsigned long)&group->fd,
894     };
895 
896     if (vfio_kvm_device_fd < 0) {
897         return;
898     }
899 
900     if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) {
901         error_report("Failed to remove group %d from KVM VFIO device: %m",
902                      group->groupid);
903     }
904 #endif
905 }
906 
907 static VFIOAddressSpace *vfio_get_address_space(AddressSpace *as)
908 {
909     VFIOAddressSpace *space;
910 
911     QLIST_FOREACH(space, &vfio_address_spaces, list) {
912         if (space->as == as) {
913             return space;
914         }
915     }
916 
917     /* No suitable VFIOAddressSpace, create a new one */
918     space = g_malloc0(sizeof(*space));
919     space->as = as;
920     QLIST_INIT(&space->containers);
921 
922     QLIST_INSERT_HEAD(&vfio_address_spaces, space, list);
923 
924     return space;
925 }
926 
927 static void vfio_put_address_space(VFIOAddressSpace *space)
928 {
929     if (QLIST_EMPTY(&space->containers)) {
930         QLIST_REMOVE(space, list);
931         g_free(space);
932     }
933 }
934 
935 static int vfio_connect_container(VFIOGroup *group, AddressSpace *as,
936                                   Error **errp)
937 {
938     VFIOContainer *container;
939     int ret, fd;
940     VFIOAddressSpace *space;
941 
942     space = vfio_get_address_space(as);
943 
944     QLIST_FOREACH(container, &space->containers, next) {
945         if (!ioctl(group->fd, VFIO_GROUP_SET_CONTAINER, &container->fd)) {
946             group->container = container;
947             QLIST_INSERT_HEAD(&container->group_list, group, container_next);
948             return 0;
949         }
950     }
951 
952     fd = qemu_open("/dev/vfio/vfio", O_RDWR);
953     if (fd < 0) {
954         error_setg_errno(errp, errno, "failed to open /dev/vfio/vfio");
955         ret = -errno;
956         goto put_space_exit;
957     }
958 
959     ret = ioctl(fd, VFIO_GET_API_VERSION);
960     if (ret != VFIO_API_VERSION) {
961         error_setg(errp, "supported vfio version: %d, "
962                    "reported version: %d", VFIO_API_VERSION, ret);
963         ret = -EINVAL;
964         goto close_fd_exit;
965     }
966 
967     container = g_malloc0(sizeof(*container));
968     container->space = space;
969     container->fd = fd;
970     if (ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_TYPE1_IOMMU) ||
971         ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_TYPE1v2_IOMMU)) {
972         bool v2 = !!ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_TYPE1v2_IOMMU);
973         struct vfio_iommu_type1_info info;
974 
975         ret = ioctl(group->fd, VFIO_GROUP_SET_CONTAINER, &fd);
976         if (ret) {
977             error_setg_errno(errp, errno, "failed to set group container");
978             ret = -errno;
979             goto free_container_exit;
980         }
981 
982         container->iommu_type = v2 ? VFIO_TYPE1v2_IOMMU : VFIO_TYPE1_IOMMU;
983         ret = ioctl(fd, VFIO_SET_IOMMU, container->iommu_type);
984         if (ret) {
985             error_setg_errno(errp, errno, "failed to set iommu for container");
986             ret = -errno;
987             goto free_container_exit;
988         }
989 
990         /*
991          * FIXME: This assumes that a Type1 IOMMU can map any 64-bit
992          * IOVA whatsoever.  That's not actually true, but the current
993          * kernel interface doesn't tell us what it can map, and the
994          * existing Type1 IOMMUs generally support any IOVA we're
995          * going to actually try in practice.
996          */
997         info.argsz = sizeof(info);
998         ret = ioctl(fd, VFIO_IOMMU_GET_INFO, &info);
999         /* Ignore errors */
1000         if (ret || !(info.flags & VFIO_IOMMU_INFO_PGSIZES)) {
1001             /* Assume 4k IOVA page size */
1002             info.iova_pgsizes = 4096;
1003         }
1004         vfio_host_win_add(container, 0, (hwaddr)-1, info.iova_pgsizes);
1005     } else if (ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_SPAPR_TCE_IOMMU) ||
1006                ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_SPAPR_TCE_v2_IOMMU)) {
1007         struct vfio_iommu_spapr_tce_info info;
1008         bool v2 = !!ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_SPAPR_TCE_v2_IOMMU);
1009 
1010         ret = ioctl(group->fd, VFIO_GROUP_SET_CONTAINER, &fd);
1011         if (ret) {
1012             error_setg_errno(errp, errno, "failed to set group container");
1013             ret = -errno;
1014             goto free_container_exit;
1015         }
1016         container->iommu_type =
1017             v2 ? VFIO_SPAPR_TCE_v2_IOMMU : VFIO_SPAPR_TCE_IOMMU;
1018         ret = ioctl(fd, VFIO_SET_IOMMU, container->iommu_type);
1019         if (ret) {
1020             error_setg_errno(errp, errno, "failed to set iommu for container");
1021             ret = -errno;
1022             goto free_container_exit;
1023         }
1024 
1025         /*
1026          * The host kernel code implementing VFIO_IOMMU_DISABLE is called
1027          * when container fd is closed so we do not call it explicitly
1028          * in this file.
1029          */
1030         if (!v2) {
1031             ret = ioctl(fd, VFIO_IOMMU_ENABLE);
1032             if (ret) {
1033                 error_setg_errno(errp, errno, "failed to enable container");
1034                 ret = -errno;
1035                 goto free_container_exit;
1036             }
1037         } else {
1038             container->prereg_listener = vfio_prereg_listener;
1039 
1040             memory_listener_register(&container->prereg_listener,
1041                                      &address_space_memory);
1042             if (container->error) {
1043                 memory_listener_unregister(&container->prereg_listener);
1044                 ret = container->error;
1045                 error_setg(errp,
1046                     "RAM memory listener initialization failed for container");
1047                 goto free_container_exit;
1048             }
1049         }
1050 
1051         info.argsz = sizeof(info);
1052         ret = ioctl(fd, VFIO_IOMMU_SPAPR_TCE_GET_INFO, &info);
1053         if (ret) {
1054             error_setg_errno(errp, errno,
1055                              "VFIO_IOMMU_SPAPR_TCE_GET_INFO failed");
1056             ret = -errno;
1057             if (v2) {
1058                 memory_listener_unregister(&container->prereg_listener);
1059             }
1060             goto free_container_exit;
1061         }
1062 
1063         if (v2) {
1064             /*
1065              * There is a default window in just created container.
1066              * To make region_add/del simpler, we better remove this
1067              * window now and let those iommu_listener callbacks
1068              * create/remove them when needed.
1069              */
1070             ret = vfio_spapr_remove_window(container, info.dma32_window_start);
1071             if (ret) {
1072                 error_setg_errno(errp, -ret,
1073                                  "failed to remove existing window");
1074                 goto free_container_exit;
1075             }
1076         } else {
1077             /* The default table uses 4K pages */
1078             vfio_host_win_add(container, info.dma32_window_start,
1079                               info.dma32_window_start +
1080                               info.dma32_window_size - 1,
1081                               0x1000);
1082         }
1083     } else {
1084         error_setg(errp, "No available IOMMU models");
1085         ret = -EINVAL;
1086         goto free_container_exit;
1087     }
1088 
1089     container->listener = vfio_memory_listener;
1090 
1091     memory_listener_register(&container->listener, container->space->as);
1092 
1093     if (container->error) {
1094         ret = container->error;
1095         error_setg_errno(errp, -ret,
1096                          "memory listener initialization failed for container");
1097         goto listener_release_exit;
1098     }
1099 
1100     container->initialized = true;
1101 
1102     QLIST_INIT(&container->group_list);
1103     QLIST_INSERT_HEAD(&space->containers, container, next);
1104 
1105     group->container = container;
1106     QLIST_INSERT_HEAD(&container->group_list, group, container_next);
1107 
1108     return 0;
1109 listener_release_exit:
1110     vfio_listener_release(container);
1111 
1112 free_container_exit:
1113     g_free(container);
1114 
1115 close_fd_exit:
1116     close(fd);
1117 
1118 put_space_exit:
1119     vfio_put_address_space(space);
1120 
1121     return ret;
1122 }
1123 
1124 static void vfio_disconnect_container(VFIOGroup *group)
1125 {
1126     VFIOContainer *container = group->container;
1127 
1128     if (ioctl(group->fd, VFIO_GROUP_UNSET_CONTAINER, &container->fd)) {
1129         error_report("vfio: error disconnecting group %d from container",
1130                      group->groupid);
1131     }
1132 
1133     QLIST_REMOVE(group, container_next);
1134     group->container = NULL;
1135 
1136     if (QLIST_EMPTY(&container->group_list)) {
1137         VFIOAddressSpace *space = container->space;
1138         VFIOGuestIOMMU *giommu, *tmp;
1139 
1140         vfio_listener_release(container);
1141         QLIST_REMOVE(container, next);
1142 
1143         QLIST_FOREACH_SAFE(giommu, &container->giommu_list, giommu_next, tmp) {
1144             memory_region_unregister_iommu_notifier(giommu->iommu, &giommu->n);
1145             QLIST_REMOVE(giommu, giommu_next);
1146             g_free(giommu);
1147         }
1148 
1149         trace_vfio_disconnect_container(container->fd);
1150         close(container->fd);
1151         g_free(container);
1152 
1153         vfio_put_address_space(space);
1154     }
1155 }
1156 
1157 VFIOGroup *vfio_get_group(int groupid, AddressSpace *as, Error **errp)
1158 {
1159     VFIOGroup *group;
1160     char path[32];
1161     struct vfio_group_status status = { .argsz = sizeof(status) };
1162 
1163     QLIST_FOREACH(group, &vfio_group_list, next) {
1164         if (group->groupid == groupid) {
1165             /* Found it.  Now is it already in the right context? */
1166             if (group->container->space->as == as) {
1167                 return group;
1168             } else {
1169                 error_setg(errp, "group %d used in multiple address spaces",
1170                            group->groupid);
1171                 return NULL;
1172             }
1173         }
1174     }
1175 
1176     group = g_malloc0(sizeof(*group));
1177 
1178     snprintf(path, sizeof(path), "/dev/vfio/%d", groupid);
1179     group->fd = qemu_open(path, O_RDWR);
1180     if (group->fd < 0) {
1181         error_setg_errno(errp, errno, "failed to open %s", path);
1182         goto free_group_exit;
1183     }
1184 
1185     if (ioctl(group->fd, VFIO_GROUP_GET_STATUS, &status)) {
1186         error_setg_errno(errp, errno, "failed to get group %d status", groupid);
1187         goto close_fd_exit;
1188     }
1189 
1190     if (!(status.flags & VFIO_GROUP_FLAGS_VIABLE)) {
1191         error_setg(errp, "group %d is not viable", groupid);
1192         error_append_hint(errp,
1193                           "Please ensure all devices within the iommu_group "
1194                           "are bound to their vfio bus driver.\n");
1195         goto close_fd_exit;
1196     }
1197 
1198     group->groupid = groupid;
1199     QLIST_INIT(&group->device_list);
1200 
1201     if (vfio_connect_container(group, as, errp)) {
1202         error_prepend(errp, "failed to setup container for group %d: ",
1203                       groupid);
1204         goto close_fd_exit;
1205     }
1206 
1207     if (QLIST_EMPTY(&vfio_group_list)) {
1208         qemu_register_reset(vfio_reset_handler, NULL);
1209     }
1210 
1211     QLIST_INSERT_HEAD(&vfio_group_list, group, next);
1212 
1213     vfio_kvm_device_add_group(group);
1214 
1215     return group;
1216 
1217 close_fd_exit:
1218     close(group->fd);
1219 
1220 free_group_exit:
1221     g_free(group);
1222 
1223     return NULL;
1224 }
1225 
1226 void vfio_put_group(VFIOGroup *group)
1227 {
1228     if (!group || !QLIST_EMPTY(&group->device_list)) {
1229         return;
1230     }
1231 
1232     vfio_kvm_device_del_group(group);
1233     vfio_disconnect_container(group);
1234     QLIST_REMOVE(group, next);
1235     trace_vfio_put_group(group->fd);
1236     close(group->fd);
1237     g_free(group);
1238 
1239     if (QLIST_EMPTY(&vfio_group_list)) {
1240         qemu_unregister_reset(vfio_reset_handler, NULL);
1241     }
1242 }
1243 
1244 int vfio_get_device(VFIOGroup *group, const char *name,
1245                     VFIODevice *vbasedev, Error **errp)
1246 {
1247     struct vfio_device_info dev_info = { .argsz = sizeof(dev_info) };
1248     int ret, fd;
1249 
1250     fd = ioctl(group->fd, VFIO_GROUP_GET_DEVICE_FD, name);
1251     if (fd < 0) {
1252         error_setg_errno(errp, errno, "error getting device from group %d",
1253                          group->groupid);
1254         error_append_hint(errp,
1255                       "Verify all devices in group %d are bound to vfio-<bus> "
1256                       "or pci-stub and not already in use\n", group->groupid);
1257         return fd;
1258     }
1259 
1260     ret = ioctl(fd, VFIO_DEVICE_GET_INFO, &dev_info);
1261     if (ret) {
1262         error_setg_errno(errp, errno, "error getting device info");
1263         close(fd);
1264         return ret;
1265     }
1266 
1267     vbasedev->fd = fd;
1268     vbasedev->group = group;
1269     QLIST_INSERT_HEAD(&group->device_list, vbasedev, next);
1270 
1271     vbasedev->num_irqs = dev_info.num_irqs;
1272     vbasedev->num_regions = dev_info.num_regions;
1273     vbasedev->flags = dev_info.flags;
1274 
1275     trace_vfio_get_device(name, dev_info.flags, dev_info.num_regions,
1276                           dev_info.num_irqs);
1277 
1278     vbasedev->reset_works = !!(dev_info.flags & VFIO_DEVICE_FLAGS_RESET);
1279     return 0;
1280 }
1281 
1282 void vfio_put_base_device(VFIODevice *vbasedev)
1283 {
1284     if (!vbasedev->group) {
1285         return;
1286     }
1287     QLIST_REMOVE(vbasedev, next);
1288     vbasedev->group = NULL;
1289     trace_vfio_put_base_device(vbasedev->fd);
1290     close(vbasedev->fd);
1291 }
1292 
1293 int vfio_get_region_info(VFIODevice *vbasedev, int index,
1294                          struct vfio_region_info **info)
1295 {
1296     size_t argsz = sizeof(struct vfio_region_info);
1297 
1298     *info = g_malloc0(argsz);
1299 
1300     (*info)->index = index;
1301 retry:
1302     (*info)->argsz = argsz;
1303 
1304     if (ioctl(vbasedev->fd, VFIO_DEVICE_GET_REGION_INFO, *info)) {
1305         g_free(*info);
1306         *info = NULL;
1307         return -errno;
1308     }
1309 
1310     if ((*info)->argsz > argsz) {
1311         argsz = (*info)->argsz;
1312         *info = g_realloc(*info, argsz);
1313 
1314         goto retry;
1315     }
1316 
1317     return 0;
1318 }
1319 
1320 int vfio_get_dev_region_info(VFIODevice *vbasedev, uint32_t type,
1321                              uint32_t subtype, struct vfio_region_info **info)
1322 {
1323     int i;
1324 
1325     for (i = 0; i < vbasedev->num_regions; i++) {
1326         struct vfio_info_cap_header *hdr;
1327         struct vfio_region_info_cap_type *cap_type;
1328 
1329         if (vfio_get_region_info(vbasedev, i, info)) {
1330             continue;
1331         }
1332 
1333         hdr = vfio_get_region_info_cap(*info, VFIO_REGION_INFO_CAP_TYPE);
1334         if (!hdr) {
1335             g_free(*info);
1336             continue;
1337         }
1338 
1339         cap_type = container_of(hdr, struct vfio_region_info_cap_type, header);
1340 
1341         trace_vfio_get_dev_region(vbasedev->name, i,
1342                                   cap_type->type, cap_type->subtype);
1343 
1344         if (cap_type->type == type && cap_type->subtype == subtype) {
1345             return 0;
1346         }
1347 
1348         g_free(*info);
1349     }
1350 
1351     *info = NULL;
1352     return -ENODEV;
1353 }
1354 
1355 /*
1356  * Interfaces for IBM EEH (Enhanced Error Handling)
1357  */
1358 static bool vfio_eeh_container_ok(VFIOContainer *container)
1359 {
1360     /*
1361      * As of 2016-03-04 (linux-4.5) the host kernel EEH/VFIO
1362      * implementation is broken if there are multiple groups in a
1363      * container.  The hardware works in units of Partitionable
1364      * Endpoints (== IOMMU groups) and the EEH operations naively
1365      * iterate across all groups in the container, without any logic
1366      * to make sure the groups have their state synchronized.  For
1367      * certain operations (ENABLE) that might be ok, until an error
1368      * occurs, but for others (GET_STATE) it's clearly broken.
1369      */
1370 
1371     /*
1372      * XXX Once fixed kernels exist, test for them here
1373      */
1374 
1375     if (QLIST_EMPTY(&container->group_list)) {
1376         return false;
1377     }
1378 
1379     if (QLIST_NEXT(QLIST_FIRST(&container->group_list), container_next)) {
1380         return false;
1381     }
1382 
1383     return true;
1384 }
1385 
1386 static int vfio_eeh_container_op(VFIOContainer *container, uint32_t op)
1387 {
1388     struct vfio_eeh_pe_op pe_op = {
1389         .argsz = sizeof(pe_op),
1390         .op = op,
1391     };
1392     int ret;
1393 
1394     if (!vfio_eeh_container_ok(container)) {
1395         error_report("vfio/eeh: EEH_PE_OP 0x%x: "
1396                      "kernel requires a container with exactly one group", op);
1397         return -EPERM;
1398     }
1399 
1400     ret = ioctl(container->fd, VFIO_EEH_PE_OP, &pe_op);
1401     if (ret < 0) {
1402         error_report("vfio/eeh: EEH_PE_OP 0x%x failed: %m", op);
1403         return -errno;
1404     }
1405 
1406     return ret;
1407 }
1408 
1409 static VFIOContainer *vfio_eeh_as_container(AddressSpace *as)
1410 {
1411     VFIOAddressSpace *space = vfio_get_address_space(as);
1412     VFIOContainer *container = NULL;
1413 
1414     if (QLIST_EMPTY(&space->containers)) {
1415         /* No containers to act on */
1416         goto out;
1417     }
1418 
1419     container = QLIST_FIRST(&space->containers);
1420 
1421     if (QLIST_NEXT(container, next)) {
1422         /* We don't yet have logic to synchronize EEH state across
1423          * multiple containers */
1424         container = NULL;
1425         goto out;
1426     }
1427 
1428 out:
1429     vfio_put_address_space(space);
1430     return container;
1431 }
1432 
1433 bool vfio_eeh_as_ok(AddressSpace *as)
1434 {
1435     VFIOContainer *container = vfio_eeh_as_container(as);
1436 
1437     return (container != NULL) && vfio_eeh_container_ok(container);
1438 }
1439 
1440 int vfio_eeh_as_op(AddressSpace *as, uint32_t op)
1441 {
1442     VFIOContainer *container = vfio_eeh_as_container(as);
1443 
1444     if (!container) {
1445         return -ENODEV;
1446     }
1447     return vfio_eeh_container_op(container, op);
1448 }
1449