xref: /openbmc/qemu/hw/vfio/common.c (revision a56de056)
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/balloon.h"
36 #include "sysemu/kvm.h"
37 #include "trace.h"
38 #include "qapi/error.h"
39 
40 VFIOGroupList vfio_group_list =
41     QLIST_HEAD_INITIALIZER(vfio_group_list);
42 static QLIST_HEAD(, VFIOAddressSpace) vfio_address_spaces =
43     QLIST_HEAD_INITIALIZER(vfio_address_spaces);
44 
45 #ifdef CONFIG_KVM
46 /*
47  * We have a single VFIO pseudo device per KVM VM.  Once created it lives
48  * for the life of the VM.  Closing the file descriptor only drops our
49  * reference to it and the device's reference to kvm.  Therefore once
50  * initialized, this file descriptor is only released on QEMU exit and
51  * we'll re-use it should another vfio device be attached before then.
52  */
53 static int vfio_kvm_device_fd = -1;
54 #endif
55 
56 /*
57  * Common VFIO interrupt disable
58  */
59 void vfio_disable_irqindex(VFIODevice *vbasedev, int index)
60 {
61     struct vfio_irq_set irq_set = {
62         .argsz = sizeof(irq_set),
63         .flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_TRIGGER,
64         .index = index,
65         .start = 0,
66         .count = 0,
67     };
68 
69     ioctl(vbasedev->fd, VFIO_DEVICE_SET_IRQS, &irq_set);
70 }
71 
72 void vfio_unmask_single_irqindex(VFIODevice *vbasedev, int index)
73 {
74     struct vfio_irq_set irq_set = {
75         .argsz = sizeof(irq_set),
76         .flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_UNMASK,
77         .index = index,
78         .start = 0,
79         .count = 1,
80     };
81 
82     ioctl(vbasedev->fd, VFIO_DEVICE_SET_IRQS, &irq_set);
83 }
84 
85 void vfio_mask_single_irqindex(VFIODevice *vbasedev, int index)
86 {
87     struct vfio_irq_set irq_set = {
88         .argsz = sizeof(irq_set),
89         .flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_MASK,
90         .index = index,
91         .start = 0,
92         .count = 1,
93     };
94 
95     ioctl(vbasedev->fd, VFIO_DEVICE_SET_IRQS, &irq_set);
96 }
97 
98 /*
99  * IO Port/MMIO - Beware of the endians, VFIO is always little endian
100  */
101 void vfio_region_write(void *opaque, hwaddr addr,
102                        uint64_t data, unsigned size)
103 {
104     VFIORegion *region = opaque;
105     VFIODevice *vbasedev = region->vbasedev;
106     union {
107         uint8_t byte;
108         uint16_t word;
109         uint32_t dword;
110         uint64_t qword;
111     } buf;
112 
113     switch (size) {
114     case 1:
115         buf.byte = data;
116         break;
117     case 2:
118         buf.word = cpu_to_le16(data);
119         break;
120     case 4:
121         buf.dword = cpu_to_le32(data);
122         break;
123     case 8:
124         buf.qword = cpu_to_le64(data);
125         break;
126     default:
127         hw_error("vfio: unsupported write size, %d bytes", size);
128         break;
129     }
130 
131     if (pwrite(vbasedev->fd, &buf, size, region->fd_offset + addr) != size) {
132         error_report("%s(%s:region%d+0x%"HWADDR_PRIx", 0x%"PRIx64
133                      ",%d) failed: %m",
134                      __func__, vbasedev->name, region->nr,
135                      addr, data, size);
136     }
137 
138     trace_vfio_region_write(vbasedev->name, region->nr, addr, data, size);
139 
140     /*
141      * A read or write to a BAR always signals an INTx EOI.  This will
142      * do nothing if not pending (including not in INTx mode).  We assume
143      * that a BAR access is in response to an interrupt and that BAR
144      * accesses will service the interrupt.  Unfortunately, we don't know
145      * which access will service the interrupt, so we're potentially
146      * getting quite a few host interrupts per guest interrupt.
147      */
148     vbasedev->ops->vfio_eoi(vbasedev);
149 }
150 
151 uint64_t vfio_region_read(void *opaque,
152                           hwaddr addr, unsigned size)
153 {
154     VFIORegion *region = opaque;
155     VFIODevice *vbasedev = region->vbasedev;
156     union {
157         uint8_t byte;
158         uint16_t word;
159         uint32_t dword;
160         uint64_t qword;
161     } buf;
162     uint64_t data = 0;
163 
164     if (pread(vbasedev->fd, &buf, size, region->fd_offset + addr) != size) {
165         error_report("%s(%s:region%d+0x%"HWADDR_PRIx", %d) failed: %m",
166                      __func__, vbasedev->name, region->nr,
167                      addr, size);
168         return (uint64_t)-1;
169     }
170     switch (size) {
171     case 1:
172         data = buf.byte;
173         break;
174     case 2:
175         data = le16_to_cpu(buf.word);
176         break;
177     case 4:
178         data = le32_to_cpu(buf.dword);
179         break;
180     case 8:
181         data = le64_to_cpu(buf.qword);
182         break;
183     default:
184         hw_error("vfio: unsupported read size, %d bytes", size);
185         break;
186     }
187 
188     trace_vfio_region_read(vbasedev->name, region->nr, addr, size, data);
189 
190     /* Same as write above */
191     vbasedev->ops->vfio_eoi(vbasedev);
192 
193     return data;
194 }
195 
196 const MemoryRegionOps vfio_region_ops = {
197     .read = vfio_region_read,
198     .write = vfio_region_write,
199     .endianness = DEVICE_LITTLE_ENDIAN,
200     .valid = {
201         .min_access_size = 1,
202         .max_access_size = 8,
203     },
204     .impl = {
205         .min_access_size = 1,
206         .max_access_size = 8,
207     },
208 };
209 
210 /*
211  * DMA - Mapping and unmapping for the "type1" IOMMU interface used on x86
212  */
213 static int vfio_dma_unmap(VFIOContainer *container,
214                           hwaddr iova, ram_addr_t size)
215 {
216     struct vfio_iommu_type1_dma_unmap unmap = {
217         .argsz = sizeof(unmap),
218         .flags = 0,
219         .iova = iova,
220         .size = size,
221     };
222 
223     if (ioctl(container->fd, VFIO_IOMMU_UNMAP_DMA, &unmap)) {
224         error_report("VFIO_UNMAP_DMA: %d", -errno);
225         return -errno;
226     }
227 
228     return 0;
229 }
230 
231 static int vfio_dma_map(VFIOContainer *container, hwaddr iova,
232                         ram_addr_t size, void *vaddr, bool readonly)
233 {
234     struct vfio_iommu_type1_dma_map map = {
235         .argsz = sizeof(map),
236         .flags = VFIO_DMA_MAP_FLAG_READ,
237         .vaddr = (__u64)(uintptr_t)vaddr,
238         .iova = iova,
239         .size = size,
240     };
241 
242     if (!readonly) {
243         map.flags |= VFIO_DMA_MAP_FLAG_WRITE;
244     }
245 
246     /*
247      * Try the mapping, if it fails with EBUSY, unmap the region and try
248      * again.  This shouldn't be necessary, but we sometimes see it in
249      * the VGA ROM space.
250      */
251     if (ioctl(container->fd, VFIO_IOMMU_MAP_DMA, &map) == 0 ||
252         (errno == EBUSY && vfio_dma_unmap(container, iova, size) == 0 &&
253          ioctl(container->fd, VFIO_IOMMU_MAP_DMA, &map) == 0)) {
254         return 0;
255     }
256 
257     error_report("VFIO_MAP_DMA: %d", -errno);
258     return -errno;
259 }
260 
261 static void vfio_host_win_add(VFIOContainer *container,
262                               hwaddr min_iova, hwaddr max_iova,
263                               uint64_t iova_pgsizes)
264 {
265     VFIOHostDMAWindow *hostwin;
266 
267     QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) {
268         if (ranges_overlap(hostwin->min_iova,
269                            hostwin->max_iova - hostwin->min_iova + 1,
270                            min_iova,
271                            max_iova - min_iova + 1)) {
272             hw_error("%s: Overlapped IOMMU are not enabled", __func__);
273         }
274     }
275 
276     hostwin = g_malloc0(sizeof(*hostwin));
277 
278     hostwin->min_iova = min_iova;
279     hostwin->max_iova = max_iova;
280     hostwin->iova_pgsizes = iova_pgsizes;
281     QLIST_INSERT_HEAD(&container->hostwin_list, hostwin, hostwin_next);
282 }
283 
284 static int vfio_host_win_del(VFIOContainer *container, hwaddr min_iova,
285                              hwaddr max_iova)
286 {
287     VFIOHostDMAWindow *hostwin;
288 
289     QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) {
290         if (hostwin->min_iova == min_iova && hostwin->max_iova == max_iova) {
291             QLIST_REMOVE(hostwin, hostwin_next);
292             return 0;
293         }
294     }
295 
296     return -1;
297 }
298 
299 static bool vfio_listener_skipped_section(MemoryRegionSection *section)
300 {
301     return (!memory_region_is_ram(section->mr) &&
302             !memory_region_is_iommu(section->mr)) ||
303            /*
304             * Sizing an enabled 64-bit BAR can cause spurious mappings to
305             * addresses in the upper part of the 64-bit address space.  These
306             * are never accessed by the CPU and beyond the address width of
307             * some IOMMU hardware.  TODO: VFIO should tell us the IOMMU width.
308             */
309            section->offset_within_address_space & (1ULL << 63);
310 }
311 
312 /* Called with rcu_read_lock held.  */
313 static bool vfio_get_vaddr(IOMMUTLBEntry *iotlb, void **vaddr,
314                            bool *read_only)
315 {
316     MemoryRegion *mr;
317     hwaddr xlat;
318     hwaddr len = iotlb->addr_mask + 1;
319     bool writable = iotlb->perm & IOMMU_WO;
320 
321     /*
322      * The IOMMU TLB entry we have just covers translation through
323      * this IOMMU to its immediate target.  We need to translate
324      * it the rest of the way through to memory.
325      */
326     mr = address_space_translate(&address_space_memory,
327                                  iotlb->translated_addr,
328                                  &xlat, &len, writable,
329                                  MEMTXATTRS_UNSPECIFIED);
330     if (!memory_region_is_ram(mr)) {
331         error_report("iommu map to non memory area %"HWADDR_PRIx"",
332                      xlat);
333         return false;
334     }
335 
336     /*
337      * Translation truncates length to the IOMMU page size,
338      * check that it did not truncate too much.
339      */
340     if (len & iotlb->addr_mask) {
341         error_report("iommu has granularity incompatible with target AS");
342         return false;
343     }
344 
345     *vaddr = memory_region_get_ram_ptr(mr) + xlat;
346     *read_only = !writable || mr->readonly;
347 
348     return true;
349 }
350 
351 static void vfio_iommu_map_notify(IOMMUNotifier *n, IOMMUTLBEntry *iotlb)
352 {
353     VFIOGuestIOMMU *giommu = container_of(n, VFIOGuestIOMMU, n);
354     VFIOContainer *container = giommu->container;
355     hwaddr iova = iotlb->iova + giommu->iommu_offset;
356     bool read_only;
357     void *vaddr;
358     int ret;
359 
360     trace_vfio_iommu_map_notify(iotlb->perm == IOMMU_NONE ? "UNMAP" : "MAP",
361                                 iova, iova + iotlb->addr_mask);
362 
363     if (iotlb->target_as != &address_space_memory) {
364         error_report("Wrong target AS \"%s\", only system memory is allowed",
365                      iotlb->target_as->name ? iotlb->target_as->name : "none");
366         return;
367     }
368 
369     rcu_read_lock();
370 
371     if ((iotlb->perm & IOMMU_RW) != IOMMU_NONE) {
372         if (!vfio_get_vaddr(iotlb, &vaddr, &read_only)) {
373             goto out;
374         }
375         /*
376          * vaddr is only valid until rcu_read_unlock(). But after
377          * vfio_dma_map has set up the mapping the pages will be
378          * pinned by the kernel. This makes sure that the RAM backend
379          * of vaddr will always be there, even if the memory object is
380          * destroyed and its backing memory munmap-ed.
381          */
382         ret = vfio_dma_map(container, iova,
383                            iotlb->addr_mask + 1, vaddr,
384                            read_only);
385         if (ret) {
386             error_report("vfio_dma_map(%p, 0x%"HWADDR_PRIx", "
387                          "0x%"HWADDR_PRIx", %p) = %d (%m)",
388                          container, iova,
389                          iotlb->addr_mask + 1, vaddr, ret);
390         }
391     } else {
392         ret = vfio_dma_unmap(container, iova, iotlb->addr_mask + 1);
393         if (ret) {
394             error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx", "
395                          "0x%"HWADDR_PRIx") = %d (%m)",
396                          container, iova,
397                          iotlb->addr_mask + 1, ret);
398         }
399     }
400 out:
401     rcu_read_unlock();
402 }
403 
404 static void vfio_listener_region_add(MemoryListener *listener,
405                                      MemoryRegionSection *section)
406 {
407     VFIOContainer *container = container_of(listener, VFIOContainer, listener);
408     hwaddr iova, end;
409     Int128 llend, llsize;
410     void *vaddr;
411     int ret;
412     VFIOHostDMAWindow *hostwin;
413     bool hostwin_found;
414 
415     if (vfio_listener_skipped_section(section)) {
416         trace_vfio_listener_region_add_skip(
417                 section->offset_within_address_space,
418                 section->offset_within_address_space +
419                 int128_get64(int128_sub(section->size, int128_one())));
420         return;
421     }
422 
423     if (unlikely((section->offset_within_address_space & ~TARGET_PAGE_MASK) !=
424                  (section->offset_within_region & ~TARGET_PAGE_MASK))) {
425         error_report("%s received unaligned region", __func__);
426         return;
427     }
428 
429     iova = TARGET_PAGE_ALIGN(section->offset_within_address_space);
430     llend = int128_make64(section->offset_within_address_space);
431     llend = int128_add(llend, section->size);
432     llend = int128_and(llend, int128_exts64(TARGET_PAGE_MASK));
433 
434     if (int128_ge(int128_make64(iova), llend)) {
435         return;
436     }
437     end = int128_get64(int128_sub(llend, int128_one()));
438 
439     if (container->iommu_type == VFIO_SPAPR_TCE_v2_IOMMU) {
440         hwaddr pgsize = 0;
441 
442         /* For now intersections are not allowed, we may relax this later */
443         QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) {
444             if (ranges_overlap(hostwin->min_iova,
445                                hostwin->max_iova - hostwin->min_iova + 1,
446                                section->offset_within_address_space,
447                                int128_get64(section->size))) {
448                 ret = -1;
449                 goto fail;
450             }
451         }
452 
453         ret = vfio_spapr_create_window(container, section, &pgsize);
454         if (ret) {
455             goto fail;
456         }
457 
458         vfio_host_win_add(container, section->offset_within_address_space,
459                           section->offset_within_address_space +
460                           int128_get64(section->size) - 1, pgsize);
461 #ifdef CONFIG_KVM
462         if (kvm_enabled()) {
463             VFIOGroup *group;
464             IOMMUMemoryRegion *iommu_mr = IOMMU_MEMORY_REGION(section->mr);
465             struct kvm_vfio_spapr_tce param;
466             struct kvm_device_attr attr = {
467                 .group = KVM_DEV_VFIO_GROUP,
468                 .attr = KVM_DEV_VFIO_GROUP_SET_SPAPR_TCE,
469                 .addr = (uint64_t)(unsigned long)&param,
470             };
471 
472             if (!memory_region_iommu_get_attr(iommu_mr, IOMMU_ATTR_SPAPR_TCE_FD,
473                                               &param.tablefd)) {
474                 QLIST_FOREACH(group, &container->group_list, container_next) {
475                     param.groupfd = group->fd;
476                     if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) {
477                         error_report("vfio: failed to setup fd %d "
478                                      "for a group with fd %d: %s",
479                                      param.tablefd, param.groupfd,
480                                      strerror(errno));
481                         return;
482                     }
483                     trace_vfio_spapr_group_attach(param.groupfd, param.tablefd);
484                 }
485             }
486         }
487 #endif
488     }
489 
490     hostwin_found = false;
491     QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) {
492         if (hostwin->min_iova <= iova && end <= hostwin->max_iova) {
493             hostwin_found = true;
494             break;
495         }
496     }
497 
498     if (!hostwin_found) {
499         error_report("vfio: IOMMU container %p can't map guest IOVA region"
500                      " 0x%"HWADDR_PRIx"..0x%"HWADDR_PRIx,
501                      container, iova, end);
502         ret = -EFAULT;
503         goto fail;
504     }
505 
506     memory_region_ref(section->mr);
507 
508     if (memory_region_is_iommu(section->mr)) {
509         VFIOGuestIOMMU *giommu;
510         IOMMUMemoryRegion *iommu_mr = IOMMU_MEMORY_REGION(section->mr);
511         int iommu_idx;
512 
513         trace_vfio_listener_region_add_iommu(iova, end);
514         /*
515          * FIXME: For VFIO iommu types which have KVM acceleration to
516          * avoid bouncing all map/unmaps through qemu this way, this
517          * would be the right place to wire that up (tell the KVM
518          * device emulation the VFIO iommu handles to use).
519          */
520         giommu = g_malloc0(sizeof(*giommu));
521         giommu->iommu = iommu_mr;
522         giommu->iommu_offset = section->offset_within_address_space -
523                                section->offset_within_region;
524         giommu->container = container;
525         llend = int128_add(int128_make64(section->offset_within_region),
526                            section->size);
527         llend = int128_sub(llend, int128_one());
528         iommu_idx = memory_region_iommu_attrs_to_index(iommu_mr,
529                                                        MEMTXATTRS_UNSPECIFIED);
530         iommu_notifier_init(&giommu->n, vfio_iommu_map_notify,
531                             IOMMU_NOTIFIER_ALL,
532                             section->offset_within_region,
533                             int128_get64(llend),
534                             iommu_idx);
535         QLIST_INSERT_HEAD(&container->giommu_list, giommu, giommu_next);
536 
537         memory_region_register_iommu_notifier(section->mr, &giommu->n);
538         memory_region_iommu_replay(giommu->iommu, &giommu->n);
539 
540         return;
541     }
542 
543     /* Here we assume that memory_region_is_ram(section->mr)==true */
544 
545     vaddr = memory_region_get_ram_ptr(section->mr) +
546             section->offset_within_region +
547             (iova - section->offset_within_address_space);
548 
549     trace_vfio_listener_region_add_ram(iova, end, vaddr);
550 
551     llsize = int128_sub(llend, int128_make64(iova));
552 
553     if (memory_region_is_ram_device(section->mr)) {
554         hwaddr pgmask = (1ULL << ctz64(hostwin->iova_pgsizes)) - 1;
555 
556         if ((iova & pgmask) || (int128_get64(llsize) & pgmask)) {
557             trace_vfio_listener_region_add_no_dma_map(
558                 memory_region_name(section->mr),
559                 section->offset_within_address_space,
560                 int128_getlo(section->size),
561                 pgmask + 1);
562             return;
563         }
564     }
565 
566     ret = vfio_dma_map(container, iova, int128_get64(llsize),
567                        vaddr, section->readonly);
568     if (ret) {
569         error_report("vfio_dma_map(%p, 0x%"HWADDR_PRIx", "
570                      "0x%"HWADDR_PRIx", %p) = %d (%m)",
571                      container, iova, int128_get64(llsize), vaddr, ret);
572         if (memory_region_is_ram_device(section->mr)) {
573             /* Allow unexpected mappings not to be fatal for RAM devices */
574             return;
575         }
576         goto fail;
577     }
578 
579     return;
580 
581 fail:
582     if (memory_region_is_ram_device(section->mr)) {
583         error_report("failed to vfio_dma_map. pci p2p may not work");
584         return;
585     }
586     /*
587      * On the initfn path, store the first error in the container so we
588      * can gracefully fail.  Runtime, there's not much we can do other
589      * than throw a hardware error.
590      */
591     if (!container->initialized) {
592         if (!container->error) {
593             container->error = ret;
594         }
595     } else {
596         hw_error("vfio: DMA mapping failed, unable to continue");
597     }
598 }
599 
600 static void vfio_listener_region_del(MemoryListener *listener,
601                                      MemoryRegionSection *section)
602 {
603     VFIOContainer *container = container_of(listener, VFIOContainer, listener);
604     hwaddr iova, end;
605     Int128 llend, llsize;
606     int ret;
607     bool try_unmap = true;
608 
609     if (vfio_listener_skipped_section(section)) {
610         trace_vfio_listener_region_del_skip(
611                 section->offset_within_address_space,
612                 section->offset_within_address_space +
613                 int128_get64(int128_sub(section->size, int128_one())));
614         return;
615     }
616 
617     if (unlikely((section->offset_within_address_space & ~TARGET_PAGE_MASK) !=
618                  (section->offset_within_region & ~TARGET_PAGE_MASK))) {
619         error_report("%s received unaligned region", __func__);
620         return;
621     }
622 
623     if (memory_region_is_iommu(section->mr)) {
624         VFIOGuestIOMMU *giommu;
625 
626         QLIST_FOREACH(giommu, &container->giommu_list, giommu_next) {
627             if (MEMORY_REGION(giommu->iommu) == section->mr &&
628                 giommu->n.start == section->offset_within_region) {
629                 memory_region_unregister_iommu_notifier(section->mr,
630                                                         &giommu->n);
631                 QLIST_REMOVE(giommu, giommu_next);
632                 g_free(giommu);
633                 break;
634             }
635         }
636 
637         /*
638          * FIXME: We assume the one big unmap below is adequate to
639          * remove any individual page mappings in the IOMMU which
640          * might have been copied into VFIO. This works for a page table
641          * based IOMMU where a big unmap flattens a large range of IO-PTEs.
642          * That may not be true for all IOMMU types.
643          */
644     }
645 
646     iova = TARGET_PAGE_ALIGN(section->offset_within_address_space);
647     llend = int128_make64(section->offset_within_address_space);
648     llend = int128_add(llend, section->size);
649     llend = int128_and(llend, int128_exts64(TARGET_PAGE_MASK));
650 
651     if (int128_ge(int128_make64(iova), llend)) {
652         return;
653     }
654     end = int128_get64(int128_sub(llend, int128_one()));
655 
656     llsize = int128_sub(llend, int128_make64(iova));
657 
658     trace_vfio_listener_region_del(iova, end);
659 
660     if (memory_region_is_ram_device(section->mr)) {
661         hwaddr pgmask;
662         VFIOHostDMAWindow *hostwin;
663         bool hostwin_found = false;
664 
665         QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) {
666             if (hostwin->min_iova <= iova && end <= hostwin->max_iova) {
667                 hostwin_found = true;
668                 break;
669             }
670         }
671         assert(hostwin_found); /* or region_add() would have failed */
672 
673         pgmask = (1ULL << ctz64(hostwin->iova_pgsizes)) - 1;
674         try_unmap = !((iova & pgmask) || (int128_get64(llsize) & pgmask));
675     }
676 
677     if (try_unmap) {
678         ret = vfio_dma_unmap(container, iova, int128_get64(llsize));
679         if (ret) {
680             error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx", "
681                          "0x%"HWADDR_PRIx") = %d (%m)",
682                          container, iova, int128_get64(llsize), ret);
683         }
684     }
685 
686     memory_region_unref(section->mr);
687 
688     if (container->iommu_type == VFIO_SPAPR_TCE_v2_IOMMU) {
689         vfio_spapr_remove_window(container,
690                                  section->offset_within_address_space);
691         if (vfio_host_win_del(container,
692                               section->offset_within_address_space,
693                               section->offset_within_address_space +
694                               int128_get64(section->size) - 1) < 0) {
695             hw_error("%s: Cannot delete missing window at %"HWADDR_PRIx,
696                      __func__, section->offset_within_address_space);
697         }
698     }
699 }
700 
701 static const MemoryListener vfio_memory_listener = {
702     .region_add = vfio_listener_region_add,
703     .region_del = vfio_listener_region_del,
704 };
705 
706 static void vfio_listener_release(VFIOContainer *container)
707 {
708     memory_listener_unregister(&container->listener);
709     if (container->iommu_type == VFIO_SPAPR_TCE_v2_IOMMU) {
710         memory_listener_unregister(&container->prereg_listener);
711     }
712 }
713 
714 static struct vfio_info_cap_header *
715 vfio_get_region_info_cap(struct vfio_region_info *info, uint16_t id)
716 {
717     struct vfio_info_cap_header *hdr;
718     void *ptr = info;
719 
720     if (!(info->flags & VFIO_REGION_INFO_FLAG_CAPS)) {
721         return NULL;
722     }
723 
724     for (hdr = ptr + info->cap_offset; hdr != ptr; hdr = ptr + hdr->next) {
725         if (hdr->id == id) {
726             return hdr;
727         }
728     }
729 
730     return NULL;
731 }
732 
733 static int vfio_setup_region_sparse_mmaps(VFIORegion *region,
734                                           struct vfio_region_info *info)
735 {
736     struct vfio_info_cap_header *hdr;
737     struct vfio_region_info_cap_sparse_mmap *sparse;
738     int i, j;
739 
740     hdr = vfio_get_region_info_cap(info, VFIO_REGION_INFO_CAP_SPARSE_MMAP);
741     if (!hdr) {
742         return -ENODEV;
743     }
744 
745     sparse = container_of(hdr, struct vfio_region_info_cap_sparse_mmap, header);
746 
747     trace_vfio_region_sparse_mmap_header(region->vbasedev->name,
748                                          region->nr, sparse->nr_areas);
749 
750     region->mmaps = g_new0(VFIOMmap, sparse->nr_areas);
751 
752     for (i = 0, j = 0; i < sparse->nr_areas; i++) {
753         trace_vfio_region_sparse_mmap_entry(i, sparse->areas[i].offset,
754                                             sparse->areas[i].offset +
755                                             sparse->areas[i].size);
756 
757         if (sparse->areas[i].size) {
758             region->mmaps[j].offset = sparse->areas[i].offset;
759             region->mmaps[j].size = sparse->areas[i].size;
760             j++;
761         }
762     }
763 
764     region->nr_mmaps = j;
765     region->mmaps = g_realloc(region->mmaps, j * sizeof(VFIOMmap));
766 
767     return 0;
768 }
769 
770 int vfio_region_setup(Object *obj, VFIODevice *vbasedev, VFIORegion *region,
771                       int index, const char *name)
772 {
773     struct vfio_region_info *info;
774     int ret;
775 
776     ret = vfio_get_region_info(vbasedev, index, &info);
777     if (ret) {
778         return ret;
779     }
780 
781     region->vbasedev = vbasedev;
782     region->flags = info->flags;
783     region->size = info->size;
784     region->fd_offset = info->offset;
785     region->nr = index;
786 
787     if (region->size) {
788         region->mem = g_new0(MemoryRegion, 1);
789         memory_region_init_io(region->mem, obj, &vfio_region_ops,
790                               region, name, region->size);
791 
792         if (!vbasedev->no_mmap &&
793             region->flags & VFIO_REGION_INFO_FLAG_MMAP) {
794 
795             ret = vfio_setup_region_sparse_mmaps(region, info);
796 
797             if (ret) {
798                 region->nr_mmaps = 1;
799                 region->mmaps = g_new0(VFIOMmap, region->nr_mmaps);
800                 region->mmaps[0].offset = 0;
801                 region->mmaps[0].size = region->size;
802             }
803         }
804     }
805 
806     g_free(info);
807 
808     trace_vfio_region_setup(vbasedev->name, index, name,
809                             region->flags, region->fd_offset, region->size);
810     return 0;
811 }
812 
813 int vfio_region_mmap(VFIORegion *region)
814 {
815     int i, prot = 0;
816     char *name;
817 
818     if (!region->mem) {
819         return 0;
820     }
821 
822     prot |= region->flags & VFIO_REGION_INFO_FLAG_READ ? PROT_READ : 0;
823     prot |= region->flags & VFIO_REGION_INFO_FLAG_WRITE ? PROT_WRITE : 0;
824 
825     for (i = 0; i < region->nr_mmaps; i++) {
826         region->mmaps[i].mmap = mmap(NULL, region->mmaps[i].size, prot,
827                                      MAP_SHARED, region->vbasedev->fd,
828                                      region->fd_offset +
829                                      region->mmaps[i].offset);
830         if (region->mmaps[i].mmap == MAP_FAILED) {
831             int ret = -errno;
832 
833             trace_vfio_region_mmap_fault(memory_region_name(region->mem), i,
834                                          region->fd_offset +
835                                          region->mmaps[i].offset,
836                                          region->fd_offset +
837                                          region->mmaps[i].offset +
838                                          region->mmaps[i].size - 1, ret);
839 
840             region->mmaps[i].mmap = NULL;
841 
842             for (i--; i >= 0; i--) {
843                 memory_region_del_subregion(region->mem, &region->mmaps[i].mem);
844                 munmap(region->mmaps[i].mmap, region->mmaps[i].size);
845                 object_unparent(OBJECT(&region->mmaps[i].mem));
846                 region->mmaps[i].mmap = NULL;
847             }
848 
849             return ret;
850         }
851 
852         name = g_strdup_printf("%s mmaps[%d]",
853                                memory_region_name(region->mem), i);
854         memory_region_init_ram_device_ptr(&region->mmaps[i].mem,
855                                           memory_region_owner(region->mem),
856                                           name, region->mmaps[i].size,
857                                           region->mmaps[i].mmap);
858         g_free(name);
859         memory_region_add_subregion(region->mem, region->mmaps[i].offset,
860                                     &region->mmaps[i].mem);
861 
862         trace_vfio_region_mmap(memory_region_name(&region->mmaps[i].mem),
863                                region->mmaps[i].offset,
864                                region->mmaps[i].offset +
865                                region->mmaps[i].size - 1);
866     }
867 
868     return 0;
869 }
870 
871 void vfio_region_exit(VFIORegion *region)
872 {
873     int i;
874 
875     if (!region->mem) {
876         return;
877     }
878 
879     for (i = 0; i < region->nr_mmaps; i++) {
880         if (region->mmaps[i].mmap) {
881             memory_region_del_subregion(region->mem, &region->mmaps[i].mem);
882         }
883     }
884 
885     trace_vfio_region_exit(region->vbasedev->name, region->nr);
886 }
887 
888 void vfio_region_finalize(VFIORegion *region)
889 {
890     int i;
891 
892     if (!region->mem) {
893         return;
894     }
895 
896     for (i = 0; i < region->nr_mmaps; i++) {
897         if (region->mmaps[i].mmap) {
898             munmap(region->mmaps[i].mmap, region->mmaps[i].size);
899             object_unparent(OBJECT(&region->mmaps[i].mem));
900         }
901     }
902 
903     object_unparent(OBJECT(region->mem));
904 
905     g_free(region->mem);
906     g_free(region->mmaps);
907 
908     trace_vfio_region_finalize(region->vbasedev->name, region->nr);
909 
910     region->mem = NULL;
911     region->mmaps = NULL;
912     region->nr_mmaps = 0;
913     region->size = 0;
914     region->flags = 0;
915     region->nr = 0;
916 }
917 
918 void vfio_region_mmaps_set_enabled(VFIORegion *region, bool enabled)
919 {
920     int i;
921 
922     if (!region->mem) {
923         return;
924     }
925 
926     for (i = 0; i < region->nr_mmaps; i++) {
927         if (region->mmaps[i].mmap) {
928             memory_region_set_enabled(&region->mmaps[i].mem, enabled);
929         }
930     }
931 
932     trace_vfio_region_mmaps_set_enabled(memory_region_name(region->mem),
933                                         enabled);
934 }
935 
936 void vfio_reset_handler(void *opaque)
937 {
938     VFIOGroup *group;
939     VFIODevice *vbasedev;
940 
941     QLIST_FOREACH(group, &vfio_group_list, next) {
942         QLIST_FOREACH(vbasedev, &group->device_list, next) {
943             if (vbasedev->dev->realized) {
944                 vbasedev->ops->vfio_compute_needs_reset(vbasedev);
945             }
946         }
947     }
948 
949     QLIST_FOREACH(group, &vfio_group_list, next) {
950         QLIST_FOREACH(vbasedev, &group->device_list, next) {
951             if (vbasedev->dev->realized && vbasedev->needs_reset) {
952                 vbasedev->ops->vfio_hot_reset_multi(vbasedev);
953             }
954         }
955     }
956 }
957 
958 static void vfio_kvm_device_add_group(VFIOGroup *group)
959 {
960 #ifdef CONFIG_KVM
961     struct kvm_device_attr attr = {
962         .group = KVM_DEV_VFIO_GROUP,
963         .attr = KVM_DEV_VFIO_GROUP_ADD,
964         .addr = (uint64_t)(unsigned long)&group->fd,
965     };
966 
967     if (!kvm_enabled()) {
968         return;
969     }
970 
971     if (vfio_kvm_device_fd < 0) {
972         struct kvm_create_device cd = {
973             .type = KVM_DEV_TYPE_VFIO,
974         };
975 
976         if (kvm_vm_ioctl(kvm_state, KVM_CREATE_DEVICE, &cd)) {
977             error_report("Failed to create KVM VFIO device: %m");
978             return;
979         }
980 
981         vfio_kvm_device_fd = cd.fd;
982     }
983 
984     if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) {
985         error_report("Failed to add group %d to KVM VFIO device: %m",
986                      group->groupid);
987     }
988 #endif
989 }
990 
991 static void vfio_kvm_device_del_group(VFIOGroup *group)
992 {
993 #ifdef CONFIG_KVM
994     struct kvm_device_attr attr = {
995         .group = KVM_DEV_VFIO_GROUP,
996         .attr = KVM_DEV_VFIO_GROUP_DEL,
997         .addr = (uint64_t)(unsigned long)&group->fd,
998     };
999 
1000     if (vfio_kvm_device_fd < 0) {
1001         return;
1002     }
1003 
1004     if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) {
1005         error_report("Failed to remove group %d from KVM VFIO device: %m",
1006                      group->groupid);
1007     }
1008 #endif
1009 }
1010 
1011 static VFIOAddressSpace *vfio_get_address_space(AddressSpace *as)
1012 {
1013     VFIOAddressSpace *space;
1014 
1015     QLIST_FOREACH(space, &vfio_address_spaces, list) {
1016         if (space->as == as) {
1017             return space;
1018         }
1019     }
1020 
1021     /* No suitable VFIOAddressSpace, create a new one */
1022     space = g_malloc0(sizeof(*space));
1023     space->as = as;
1024     QLIST_INIT(&space->containers);
1025 
1026     QLIST_INSERT_HEAD(&vfio_address_spaces, space, list);
1027 
1028     return space;
1029 }
1030 
1031 static void vfio_put_address_space(VFIOAddressSpace *space)
1032 {
1033     if (QLIST_EMPTY(&space->containers)) {
1034         QLIST_REMOVE(space, list);
1035         g_free(space);
1036     }
1037 }
1038 
1039 static int vfio_connect_container(VFIOGroup *group, AddressSpace *as,
1040                                   Error **errp)
1041 {
1042     VFIOContainer *container;
1043     int ret, fd;
1044     VFIOAddressSpace *space;
1045 
1046     space = vfio_get_address_space(as);
1047 
1048     /*
1049      * VFIO is currently incompatible with memory ballooning insofar as the
1050      * madvise to purge (zap) the page from QEMU's address space does not
1051      * interact with the memory API and therefore leaves stale virtual to
1052      * physical mappings in the IOMMU if the page was previously pinned.  We
1053      * therefore add a balloon inhibit for each group added to a container,
1054      * whether the container is used individually or shared.  This provides
1055      * us with options to allow devices within a group to opt-in and allow
1056      * ballooning, so long as it is done consistently for a group (for instance
1057      * if the device is an mdev device where it is known that the host vendor
1058      * driver will never pin pages outside of the working set of the guest
1059      * driver, which would thus not be ballooning candidates).
1060      *
1061      * The first opportunity to induce pinning occurs here where we attempt to
1062      * attach the group to existing containers within the AddressSpace.  If any
1063      * pages are already zapped from the virtual address space, such as from a
1064      * previous ballooning opt-in, new pinning will cause valid mappings to be
1065      * re-established.  Likewise, when the overall MemoryListener for a new
1066      * container is registered, a replay of mappings within the AddressSpace
1067      * will occur, re-establishing any previously zapped pages as well.
1068      *
1069      * NB. Balloon inhibiting does not currently block operation of the
1070      * balloon driver or revoke previously pinned pages, it only prevents
1071      * calling madvise to modify the virtual mapping of ballooned pages.
1072      */
1073     qemu_balloon_inhibit(true);
1074 
1075     QLIST_FOREACH(container, &space->containers, next) {
1076         if (!ioctl(group->fd, VFIO_GROUP_SET_CONTAINER, &container->fd)) {
1077             group->container = container;
1078             QLIST_INSERT_HEAD(&container->group_list, group, container_next);
1079             vfio_kvm_device_add_group(group);
1080             return 0;
1081         }
1082     }
1083 
1084     fd = qemu_open("/dev/vfio/vfio", O_RDWR);
1085     if (fd < 0) {
1086         error_setg_errno(errp, errno, "failed to open /dev/vfio/vfio");
1087         ret = -errno;
1088         goto put_space_exit;
1089     }
1090 
1091     ret = ioctl(fd, VFIO_GET_API_VERSION);
1092     if (ret != VFIO_API_VERSION) {
1093         error_setg(errp, "supported vfio version: %d, "
1094                    "reported version: %d", VFIO_API_VERSION, ret);
1095         ret = -EINVAL;
1096         goto close_fd_exit;
1097     }
1098 
1099     container = g_malloc0(sizeof(*container));
1100     container->space = space;
1101     container->fd = fd;
1102     QLIST_INIT(&container->giommu_list);
1103     QLIST_INIT(&container->hostwin_list);
1104     if (ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_TYPE1_IOMMU) ||
1105         ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_TYPE1v2_IOMMU)) {
1106         bool v2 = !!ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_TYPE1v2_IOMMU);
1107         struct vfio_iommu_type1_info info;
1108 
1109         ret = ioctl(group->fd, VFIO_GROUP_SET_CONTAINER, &fd);
1110         if (ret) {
1111             error_setg_errno(errp, errno, "failed to set group container");
1112             ret = -errno;
1113             goto free_container_exit;
1114         }
1115 
1116         container->iommu_type = v2 ? VFIO_TYPE1v2_IOMMU : VFIO_TYPE1_IOMMU;
1117         ret = ioctl(fd, VFIO_SET_IOMMU, container->iommu_type);
1118         if (ret) {
1119             error_setg_errno(errp, errno, "failed to set iommu for container");
1120             ret = -errno;
1121             goto free_container_exit;
1122         }
1123 
1124         /*
1125          * FIXME: This assumes that a Type1 IOMMU can map any 64-bit
1126          * IOVA whatsoever.  That's not actually true, but the current
1127          * kernel interface doesn't tell us what it can map, and the
1128          * existing Type1 IOMMUs generally support any IOVA we're
1129          * going to actually try in practice.
1130          */
1131         info.argsz = sizeof(info);
1132         ret = ioctl(fd, VFIO_IOMMU_GET_INFO, &info);
1133         /* Ignore errors */
1134         if (ret || !(info.flags & VFIO_IOMMU_INFO_PGSIZES)) {
1135             /* Assume 4k IOVA page size */
1136             info.iova_pgsizes = 4096;
1137         }
1138         vfio_host_win_add(container, 0, (hwaddr)-1, info.iova_pgsizes);
1139         container->pgsizes = info.iova_pgsizes;
1140     } else if (ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_SPAPR_TCE_IOMMU) ||
1141                ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_SPAPR_TCE_v2_IOMMU)) {
1142         struct vfio_iommu_spapr_tce_info info;
1143         bool v2 = !!ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_SPAPR_TCE_v2_IOMMU);
1144 
1145         ret = ioctl(group->fd, VFIO_GROUP_SET_CONTAINER, &fd);
1146         if (ret) {
1147             error_setg_errno(errp, errno, "failed to set group container");
1148             ret = -errno;
1149             goto free_container_exit;
1150         }
1151         container->iommu_type =
1152             v2 ? VFIO_SPAPR_TCE_v2_IOMMU : VFIO_SPAPR_TCE_IOMMU;
1153         ret = ioctl(fd, VFIO_SET_IOMMU, container->iommu_type);
1154         if (ret) {
1155             container->iommu_type = VFIO_SPAPR_TCE_IOMMU;
1156             v2 = false;
1157             ret = ioctl(fd, VFIO_SET_IOMMU, container->iommu_type);
1158         }
1159         if (ret) {
1160             error_setg_errno(errp, errno, "failed to set iommu for container");
1161             ret = -errno;
1162             goto free_container_exit;
1163         }
1164 
1165         /*
1166          * The host kernel code implementing VFIO_IOMMU_DISABLE is called
1167          * when container fd is closed so we do not call it explicitly
1168          * in this file.
1169          */
1170         if (!v2) {
1171             ret = ioctl(fd, VFIO_IOMMU_ENABLE);
1172             if (ret) {
1173                 error_setg_errno(errp, errno, "failed to enable container");
1174                 ret = -errno;
1175                 goto free_container_exit;
1176             }
1177         } else {
1178             container->prereg_listener = vfio_prereg_listener;
1179 
1180             memory_listener_register(&container->prereg_listener,
1181                                      &address_space_memory);
1182             if (container->error) {
1183                 memory_listener_unregister(&container->prereg_listener);
1184                 ret = container->error;
1185                 error_setg(errp,
1186                     "RAM memory listener initialization failed for container");
1187                 goto free_container_exit;
1188             }
1189         }
1190 
1191         info.argsz = sizeof(info);
1192         ret = ioctl(fd, VFIO_IOMMU_SPAPR_TCE_GET_INFO, &info);
1193         if (ret) {
1194             error_setg_errno(errp, errno,
1195                              "VFIO_IOMMU_SPAPR_TCE_GET_INFO failed");
1196             ret = -errno;
1197             if (v2) {
1198                 memory_listener_unregister(&container->prereg_listener);
1199             }
1200             goto free_container_exit;
1201         }
1202 
1203         if (v2) {
1204             container->pgsizes = info.ddw.pgsizes;
1205             /*
1206              * There is a default window in just created container.
1207              * To make region_add/del simpler, we better remove this
1208              * window now and let those iommu_listener callbacks
1209              * create/remove them when needed.
1210              */
1211             ret = vfio_spapr_remove_window(container, info.dma32_window_start);
1212             if (ret) {
1213                 error_setg_errno(errp, -ret,
1214                                  "failed to remove existing window");
1215                 goto free_container_exit;
1216             }
1217         } else {
1218             /* The default table uses 4K pages */
1219             container->pgsizes = 0x1000;
1220             vfio_host_win_add(container, info.dma32_window_start,
1221                               info.dma32_window_start +
1222                               info.dma32_window_size - 1,
1223                               0x1000);
1224         }
1225     } else {
1226         error_setg(errp, "No available IOMMU models");
1227         ret = -EINVAL;
1228         goto free_container_exit;
1229     }
1230 
1231     vfio_kvm_device_add_group(group);
1232 
1233     QLIST_INIT(&container->group_list);
1234     QLIST_INSERT_HEAD(&space->containers, container, next);
1235 
1236     group->container = container;
1237     QLIST_INSERT_HEAD(&container->group_list, group, container_next);
1238 
1239     container->listener = vfio_memory_listener;
1240 
1241     memory_listener_register(&container->listener, container->space->as);
1242 
1243     if (container->error) {
1244         ret = container->error;
1245         error_setg_errno(errp, -ret,
1246                          "memory listener initialization failed for container");
1247         goto listener_release_exit;
1248     }
1249 
1250     container->initialized = true;
1251 
1252     return 0;
1253 listener_release_exit:
1254     QLIST_REMOVE(group, container_next);
1255     QLIST_REMOVE(container, next);
1256     vfio_kvm_device_del_group(group);
1257     vfio_listener_release(container);
1258 
1259 free_container_exit:
1260     g_free(container);
1261 
1262 close_fd_exit:
1263     close(fd);
1264 
1265 put_space_exit:
1266     qemu_balloon_inhibit(false);
1267     vfio_put_address_space(space);
1268 
1269     return ret;
1270 }
1271 
1272 static void vfio_disconnect_container(VFIOGroup *group)
1273 {
1274     VFIOContainer *container = group->container;
1275 
1276     QLIST_REMOVE(group, container_next);
1277     group->container = NULL;
1278 
1279     /*
1280      * Explicitly release the listener first before unset container,
1281      * since unset may destroy the backend container if it's the last
1282      * group.
1283      */
1284     if (QLIST_EMPTY(&container->group_list)) {
1285         vfio_listener_release(container);
1286     }
1287 
1288     if (ioctl(group->fd, VFIO_GROUP_UNSET_CONTAINER, &container->fd)) {
1289         error_report("vfio: error disconnecting group %d from container",
1290                      group->groupid);
1291     }
1292 
1293     if (QLIST_EMPTY(&container->group_list)) {
1294         VFIOAddressSpace *space = container->space;
1295         VFIOGuestIOMMU *giommu, *tmp;
1296 
1297         QLIST_REMOVE(container, next);
1298 
1299         QLIST_FOREACH_SAFE(giommu, &container->giommu_list, giommu_next, tmp) {
1300             memory_region_unregister_iommu_notifier(
1301                     MEMORY_REGION(giommu->iommu), &giommu->n);
1302             QLIST_REMOVE(giommu, giommu_next);
1303             g_free(giommu);
1304         }
1305 
1306         trace_vfio_disconnect_container(container->fd);
1307         close(container->fd);
1308         g_free(container);
1309 
1310         vfio_put_address_space(space);
1311     }
1312 }
1313 
1314 VFIOGroup *vfio_get_group(int groupid, AddressSpace *as, Error **errp)
1315 {
1316     VFIOGroup *group;
1317     char path[32];
1318     struct vfio_group_status status = { .argsz = sizeof(status) };
1319 
1320     QLIST_FOREACH(group, &vfio_group_list, next) {
1321         if (group->groupid == groupid) {
1322             /* Found it.  Now is it already in the right context? */
1323             if (group->container->space->as == as) {
1324                 return group;
1325             } else {
1326                 error_setg(errp, "group %d used in multiple address spaces",
1327                            group->groupid);
1328                 return NULL;
1329             }
1330         }
1331     }
1332 
1333     group = g_malloc0(sizeof(*group));
1334 
1335     snprintf(path, sizeof(path), "/dev/vfio/%d", groupid);
1336     group->fd = qemu_open(path, O_RDWR);
1337     if (group->fd < 0) {
1338         error_setg_errno(errp, errno, "failed to open %s", path);
1339         goto free_group_exit;
1340     }
1341 
1342     if (ioctl(group->fd, VFIO_GROUP_GET_STATUS, &status)) {
1343         error_setg_errno(errp, errno, "failed to get group %d status", groupid);
1344         goto close_fd_exit;
1345     }
1346 
1347     if (!(status.flags & VFIO_GROUP_FLAGS_VIABLE)) {
1348         error_setg(errp, "group %d is not viable", groupid);
1349         error_append_hint(errp,
1350                           "Please ensure all devices within the iommu_group "
1351                           "are bound to their vfio bus driver.\n");
1352         goto close_fd_exit;
1353     }
1354 
1355     group->groupid = groupid;
1356     QLIST_INIT(&group->device_list);
1357 
1358     if (vfio_connect_container(group, as, errp)) {
1359         error_prepend(errp, "failed to setup container for group %d: ",
1360                       groupid);
1361         goto close_fd_exit;
1362     }
1363 
1364     if (QLIST_EMPTY(&vfio_group_list)) {
1365         qemu_register_reset(vfio_reset_handler, NULL);
1366     }
1367 
1368     QLIST_INSERT_HEAD(&vfio_group_list, group, next);
1369 
1370     return group;
1371 
1372 close_fd_exit:
1373     close(group->fd);
1374 
1375 free_group_exit:
1376     g_free(group);
1377 
1378     return NULL;
1379 }
1380 
1381 void vfio_put_group(VFIOGroup *group)
1382 {
1383     if (!group || !QLIST_EMPTY(&group->device_list)) {
1384         return;
1385     }
1386 
1387     if (!group->balloon_allowed) {
1388         qemu_balloon_inhibit(false);
1389     }
1390     vfio_kvm_device_del_group(group);
1391     vfio_disconnect_container(group);
1392     QLIST_REMOVE(group, next);
1393     trace_vfio_put_group(group->fd);
1394     close(group->fd);
1395     g_free(group);
1396 
1397     if (QLIST_EMPTY(&vfio_group_list)) {
1398         qemu_unregister_reset(vfio_reset_handler, NULL);
1399     }
1400 }
1401 
1402 int vfio_get_device(VFIOGroup *group, const char *name,
1403                     VFIODevice *vbasedev, Error **errp)
1404 {
1405     struct vfio_device_info dev_info = { .argsz = sizeof(dev_info) };
1406     int ret, fd;
1407 
1408     fd = ioctl(group->fd, VFIO_GROUP_GET_DEVICE_FD, name);
1409     if (fd < 0) {
1410         error_setg_errno(errp, errno, "error getting device from group %d",
1411                          group->groupid);
1412         error_append_hint(errp,
1413                       "Verify all devices in group %d are bound to vfio-<bus> "
1414                       "or pci-stub and not already in use\n", group->groupid);
1415         return fd;
1416     }
1417 
1418     ret = ioctl(fd, VFIO_DEVICE_GET_INFO, &dev_info);
1419     if (ret) {
1420         error_setg_errno(errp, errno, "error getting device info");
1421         close(fd);
1422         return ret;
1423     }
1424 
1425     /*
1426      * Clear the balloon inhibitor for this group if the driver knows the
1427      * device operates compatibly with ballooning.  Setting must be consistent
1428      * per group, but since compatibility is really only possible with mdev
1429      * currently, we expect singleton groups.
1430      */
1431     if (vbasedev->balloon_allowed != group->balloon_allowed) {
1432         if (!QLIST_EMPTY(&group->device_list)) {
1433             error_setg(errp,
1434                        "Inconsistent device balloon setting within group");
1435             close(fd);
1436             return -1;
1437         }
1438 
1439         if (!group->balloon_allowed) {
1440             group->balloon_allowed = true;
1441             qemu_balloon_inhibit(false);
1442         }
1443     }
1444 
1445     vbasedev->fd = fd;
1446     vbasedev->group = group;
1447     QLIST_INSERT_HEAD(&group->device_list, vbasedev, next);
1448 
1449     vbasedev->num_irqs = dev_info.num_irqs;
1450     vbasedev->num_regions = dev_info.num_regions;
1451     vbasedev->flags = dev_info.flags;
1452 
1453     trace_vfio_get_device(name, dev_info.flags, dev_info.num_regions,
1454                           dev_info.num_irqs);
1455 
1456     vbasedev->reset_works = !!(dev_info.flags & VFIO_DEVICE_FLAGS_RESET);
1457     return 0;
1458 }
1459 
1460 void vfio_put_base_device(VFIODevice *vbasedev)
1461 {
1462     if (!vbasedev->group) {
1463         return;
1464     }
1465     QLIST_REMOVE(vbasedev, next);
1466     vbasedev->group = NULL;
1467     trace_vfio_put_base_device(vbasedev->fd);
1468     close(vbasedev->fd);
1469 }
1470 
1471 int vfio_get_region_info(VFIODevice *vbasedev, int index,
1472                          struct vfio_region_info **info)
1473 {
1474     size_t argsz = sizeof(struct vfio_region_info);
1475 
1476     *info = g_malloc0(argsz);
1477 
1478     (*info)->index = index;
1479 retry:
1480     (*info)->argsz = argsz;
1481 
1482     if (ioctl(vbasedev->fd, VFIO_DEVICE_GET_REGION_INFO, *info)) {
1483         g_free(*info);
1484         *info = NULL;
1485         return -errno;
1486     }
1487 
1488     if ((*info)->argsz > argsz) {
1489         argsz = (*info)->argsz;
1490         *info = g_realloc(*info, argsz);
1491 
1492         goto retry;
1493     }
1494 
1495     return 0;
1496 }
1497 
1498 int vfio_get_dev_region_info(VFIODevice *vbasedev, uint32_t type,
1499                              uint32_t subtype, struct vfio_region_info **info)
1500 {
1501     int i;
1502 
1503     for (i = 0; i < vbasedev->num_regions; i++) {
1504         struct vfio_info_cap_header *hdr;
1505         struct vfio_region_info_cap_type *cap_type;
1506 
1507         if (vfio_get_region_info(vbasedev, i, info)) {
1508             continue;
1509         }
1510 
1511         hdr = vfio_get_region_info_cap(*info, VFIO_REGION_INFO_CAP_TYPE);
1512         if (!hdr) {
1513             g_free(*info);
1514             continue;
1515         }
1516 
1517         cap_type = container_of(hdr, struct vfio_region_info_cap_type, header);
1518 
1519         trace_vfio_get_dev_region(vbasedev->name, i,
1520                                   cap_type->type, cap_type->subtype);
1521 
1522         if (cap_type->type == type && cap_type->subtype == subtype) {
1523             return 0;
1524         }
1525 
1526         g_free(*info);
1527     }
1528 
1529     *info = NULL;
1530     return -ENODEV;
1531 }
1532 
1533 bool vfio_has_region_cap(VFIODevice *vbasedev, int region, uint16_t cap_type)
1534 {
1535     struct vfio_region_info *info = NULL;
1536     bool ret = false;
1537 
1538     if (!vfio_get_region_info(vbasedev, region, &info)) {
1539         if (vfio_get_region_info_cap(info, cap_type)) {
1540             ret = true;
1541         }
1542         g_free(info);
1543     }
1544 
1545     return ret;
1546 }
1547 
1548 /*
1549  * Interfaces for IBM EEH (Enhanced Error Handling)
1550  */
1551 static bool vfio_eeh_container_ok(VFIOContainer *container)
1552 {
1553     /*
1554      * As of 2016-03-04 (linux-4.5) the host kernel EEH/VFIO
1555      * implementation is broken if there are multiple groups in a
1556      * container.  The hardware works in units of Partitionable
1557      * Endpoints (== IOMMU groups) and the EEH operations naively
1558      * iterate across all groups in the container, without any logic
1559      * to make sure the groups have their state synchronized.  For
1560      * certain operations (ENABLE) that might be ok, until an error
1561      * occurs, but for others (GET_STATE) it's clearly broken.
1562      */
1563 
1564     /*
1565      * XXX Once fixed kernels exist, test for them here
1566      */
1567 
1568     if (QLIST_EMPTY(&container->group_list)) {
1569         return false;
1570     }
1571 
1572     if (QLIST_NEXT(QLIST_FIRST(&container->group_list), container_next)) {
1573         return false;
1574     }
1575 
1576     return true;
1577 }
1578 
1579 static int vfio_eeh_container_op(VFIOContainer *container, uint32_t op)
1580 {
1581     struct vfio_eeh_pe_op pe_op = {
1582         .argsz = sizeof(pe_op),
1583         .op = op,
1584     };
1585     int ret;
1586 
1587     if (!vfio_eeh_container_ok(container)) {
1588         error_report("vfio/eeh: EEH_PE_OP 0x%x: "
1589                      "kernel requires a container with exactly one group", op);
1590         return -EPERM;
1591     }
1592 
1593     ret = ioctl(container->fd, VFIO_EEH_PE_OP, &pe_op);
1594     if (ret < 0) {
1595         error_report("vfio/eeh: EEH_PE_OP 0x%x failed: %m", op);
1596         return -errno;
1597     }
1598 
1599     return ret;
1600 }
1601 
1602 static VFIOContainer *vfio_eeh_as_container(AddressSpace *as)
1603 {
1604     VFIOAddressSpace *space = vfio_get_address_space(as);
1605     VFIOContainer *container = NULL;
1606 
1607     if (QLIST_EMPTY(&space->containers)) {
1608         /* No containers to act on */
1609         goto out;
1610     }
1611 
1612     container = QLIST_FIRST(&space->containers);
1613 
1614     if (QLIST_NEXT(container, next)) {
1615         /* We don't yet have logic to synchronize EEH state across
1616          * multiple containers */
1617         container = NULL;
1618         goto out;
1619     }
1620 
1621 out:
1622     vfio_put_address_space(space);
1623     return container;
1624 }
1625 
1626 bool vfio_eeh_as_ok(AddressSpace *as)
1627 {
1628     VFIOContainer *container = vfio_eeh_as_container(as);
1629 
1630     return (container != NULL) && vfio_eeh_container_ok(container);
1631 }
1632 
1633 int vfio_eeh_as_op(AddressSpace *as, uint32_t op)
1634 {
1635     VFIOContainer *container = vfio_eeh_as_container(as);
1636 
1637     if (!container) {
1638         return -ENODEV;
1639     }
1640     return vfio_eeh_container_op(container, op);
1641 }
1642