xref: /openbmc/qemu/hw/vfio/common.c (revision 284d697c)
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 "exec/ram_addr.h"
33 #include "hw/hw.h"
34 #include "qemu/error-report.h"
35 #include "qemu/main-loop.h"
36 #include "qemu/range.h"
37 #include "sysemu/kvm.h"
38 #include "sysemu/reset.h"
39 #include "trace.h"
40 #include "qapi/error.h"
41 #include "migration/migration.h"
42 
43 VFIOGroupList vfio_group_list =
44     QLIST_HEAD_INITIALIZER(vfio_group_list);
45 static QLIST_HEAD(, VFIOAddressSpace) vfio_address_spaces =
46     QLIST_HEAD_INITIALIZER(vfio_address_spaces);
47 
48 #ifdef CONFIG_KVM
49 /*
50  * We have a single VFIO pseudo device per KVM VM.  Once created it lives
51  * for the life of the VM.  Closing the file descriptor only drops our
52  * reference to it and the device's reference to kvm.  Therefore once
53  * initialized, this file descriptor is only released on QEMU exit and
54  * we'll re-use it should another vfio device be attached before then.
55  */
56 static int vfio_kvm_device_fd = -1;
57 #endif
58 
59 /*
60  * Common VFIO interrupt disable
61  */
62 void vfio_disable_irqindex(VFIODevice *vbasedev, int index)
63 {
64     struct vfio_irq_set irq_set = {
65         .argsz = sizeof(irq_set),
66         .flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_TRIGGER,
67         .index = index,
68         .start = 0,
69         .count = 0,
70     };
71 
72     ioctl(vbasedev->fd, VFIO_DEVICE_SET_IRQS, &irq_set);
73 }
74 
75 void vfio_unmask_single_irqindex(VFIODevice *vbasedev, int index)
76 {
77     struct vfio_irq_set irq_set = {
78         .argsz = sizeof(irq_set),
79         .flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_UNMASK,
80         .index = index,
81         .start = 0,
82         .count = 1,
83     };
84 
85     ioctl(vbasedev->fd, VFIO_DEVICE_SET_IRQS, &irq_set);
86 }
87 
88 void vfio_mask_single_irqindex(VFIODevice *vbasedev, int index)
89 {
90     struct vfio_irq_set irq_set = {
91         .argsz = sizeof(irq_set),
92         .flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_MASK,
93         .index = index,
94         .start = 0,
95         .count = 1,
96     };
97 
98     ioctl(vbasedev->fd, VFIO_DEVICE_SET_IRQS, &irq_set);
99 }
100 
101 static inline const char *action_to_str(int action)
102 {
103     switch (action) {
104     case VFIO_IRQ_SET_ACTION_MASK:
105         return "MASK";
106     case VFIO_IRQ_SET_ACTION_UNMASK:
107         return "UNMASK";
108     case VFIO_IRQ_SET_ACTION_TRIGGER:
109         return "TRIGGER";
110     default:
111         return "UNKNOWN ACTION";
112     }
113 }
114 
115 static const char *index_to_str(VFIODevice *vbasedev, int index)
116 {
117     if (vbasedev->type != VFIO_DEVICE_TYPE_PCI) {
118         return NULL;
119     }
120 
121     switch (index) {
122     case VFIO_PCI_INTX_IRQ_INDEX:
123         return "INTX";
124     case VFIO_PCI_MSI_IRQ_INDEX:
125         return "MSI";
126     case VFIO_PCI_MSIX_IRQ_INDEX:
127         return "MSIX";
128     case VFIO_PCI_ERR_IRQ_INDEX:
129         return "ERR";
130     case VFIO_PCI_REQ_IRQ_INDEX:
131         return "REQ";
132     default:
133         return NULL;
134     }
135 }
136 
137 int vfio_set_irq_signaling(VFIODevice *vbasedev, int index, int subindex,
138                            int action, int fd, Error **errp)
139 {
140     struct vfio_irq_set *irq_set;
141     int argsz, ret = 0;
142     const char *name;
143     int32_t *pfd;
144 
145     argsz = sizeof(*irq_set) + sizeof(*pfd);
146 
147     irq_set = g_malloc0(argsz);
148     irq_set->argsz = argsz;
149     irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD | action;
150     irq_set->index = index;
151     irq_set->start = subindex;
152     irq_set->count = 1;
153     pfd = (int32_t *)&irq_set->data;
154     *pfd = fd;
155 
156     if (ioctl(vbasedev->fd, VFIO_DEVICE_SET_IRQS, irq_set)) {
157         ret = -errno;
158     }
159     g_free(irq_set);
160 
161     if (!ret) {
162         return 0;
163     }
164 
165     error_setg_errno(errp, -ret, "VFIO_DEVICE_SET_IRQS failure");
166 
167     name = index_to_str(vbasedev, index);
168     if (name) {
169         error_prepend(errp, "%s-%d: ", name, subindex);
170     } else {
171         error_prepend(errp, "index %d-%d: ", index, subindex);
172     }
173     error_prepend(errp,
174                   "Failed to %s %s eventfd signaling for interrupt ",
175                   fd < 0 ? "tear down" : "set up", action_to_str(action));
176     return ret;
177 }
178 
179 /*
180  * IO Port/MMIO - Beware of the endians, VFIO is always little endian
181  */
182 void vfio_region_write(void *opaque, hwaddr addr,
183                        uint64_t data, unsigned size)
184 {
185     VFIORegion *region = opaque;
186     VFIODevice *vbasedev = region->vbasedev;
187     union {
188         uint8_t byte;
189         uint16_t word;
190         uint32_t dword;
191         uint64_t qword;
192     } buf;
193 
194     switch (size) {
195     case 1:
196         buf.byte = data;
197         break;
198     case 2:
199         buf.word = cpu_to_le16(data);
200         break;
201     case 4:
202         buf.dword = cpu_to_le32(data);
203         break;
204     case 8:
205         buf.qword = cpu_to_le64(data);
206         break;
207     default:
208         hw_error("vfio: unsupported write size, %u bytes", size);
209         break;
210     }
211 
212     if (pwrite(vbasedev->fd, &buf, size, region->fd_offset + addr) != size) {
213         error_report("%s(%s:region%d+0x%"HWADDR_PRIx", 0x%"PRIx64
214                      ",%d) failed: %m",
215                      __func__, vbasedev->name, region->nr,
216                      addr, data, size);
217     }
218 
219     trace_vfio_region_write(vbasedev->name, region->nr, addr, data, size);
220 
221     /*
222      * A read or write to a BAR always signals an INTx EOI.  This will
223      * do nothing if not pending (including not in INTx mode).  We assume
224      * that a BAR access is in response to an interrupt and that BAR
225      * accesses will service the interrupt.  Unfortunately, we don't know
226      * which access will service the interrupt, so we're potentially
227      * getting quite a few host interrupts per guest interrupt.
228      */
229     vbasedev->ops->vfio_eoi(vbasedev);
230 }
231 
232 uint64_t vfio_region_read(void *opaque,
233                           hwaddr addr, unsigned size)
234 {
235     VFIORegion *region = opaque;
236     VFIODevice *vbasedev = region->vbasedev;
237     union {
238         uint8_t byte;
239         uint16_t word;
240         uint32_t dword;
241         uint64_t qword;
242     } buf;
243     uint64_t data = 0;
244 
245     if (pread(vbasedev->fd, &buf, size, region->fd_offset + addr) != size) {
246         error_report("%s(%s:region%d+0x%"HWADDR_PRIx", %d) failed: %m",
247                      __func__, vbasedev->name, region->nr,
248                      addr, size);
249         return (uint64_t)-1;
250     }
251     switch (size) {
252     case 1:
253         data = buf.byte;
254         break;
255     case 2:
256         data = le16_to_cpu(buf.word);
257         break;
258     case 4:
259         data = le32_to_cpu(buf.dword);
260         break;
261     case 8:
262         data = le64_to_cpu(buf.qword);
263         break;
264     default:
265         hw_error("vfio: unsupported read size, %u bytes", size);
266         break;
267     }
268 
269     trace_vfio_region_read(vbasedev->name, region->nr, addr, size, data);
270 
271     /* Same as write above */
272     vbasedev->ops->vfio_eoi(vbasedev);
273 
274     return data;
275 }
276 
277 const MemoryRegionOps vfio_region_ops = {
278     .read = vfio_region_read,
279     .write = vfio_region_write,
280     .endianness = DEVICE_LITTLE_ENDIAN,
281     .valid = {
282         .min_access_size = 1,
283         .max_access_size = 8,
284     },
285     .impl = {
286         .min_access_size = 1,
287         .max_access_size = 8,
288     },
289 };
290 
291 /*
292  * Device state interfaces
293  */
294 
295 bool vfio_mig_active(void)
296 {
297     VFIOGroup *group;
298     VFIODevice *vbasedev;
299 
300     if (QLIST_EMPTY(&vfio_group_list)) {
301         return false;
302     }
303 
304     QLIST_FOREACH(group, &vfio_group_list, next) {
305         QLIST_FOREACH(vbasedev, &group->device_list, next) {
306             if (vbasedev->migration_blocker) {
307                 return false;
308             }
309         }
310     }
311     return true;
312 }
313 
314 static bool vfio_devices_all_stopped_and_saving(VFIOContainer *container)
315 {
316     VFIOGroup *group;
317     VFIODevice *vbasedev;
318     MigrationState *ms = migrate_get_current();
319 
320     if (!migration_is_setup_or_active(ms->state)) {
321         return false;
322     }
323 
324     QLIST_FOREACH(group, &container->group_list, container_next) {
325         QLIST_FOREACH(vbasedev, &group->device_list, next) {
326             VFIOMigration *migration = vbasedev->migration;
327 
328             if (!migration) {
329                 return false;
330             }
331 
332             if ((migration->device_state & VFIO_DEVICE_STATE_SAVING) &&
333                 !(migration->device_state & VFIO_DEVICE_STATE_RUNNING)) {
334                 continue;
335             } else {
336                 return false;
337             }
338         }
339     }
340     return true;
341 }
342 
343 static bool vfio_devices_all_running_and_saving(VFIOContainer *container)
344 {
345     VFIOGroup *group;
346     VFIODevice *vbasedev;
347     MigrationState *ms = migrate_get_current();
348 
349     if (!migration_is_setup_or_active(ms->state)) {
350         return false;
351     }
352 
353     QLIST_FOREACH(group, &container->group_list, container_next) {
354         QLIST_FOREACH(vbasedev, &group->device_list, next) {
355             VFIOMigration *migration = vbasedev->migration;
356 
357             if (!migration) {
358                 return false;
359             }
360 
361             if ((migration->device_state & VFIO_DEVICE_STATE_SAVING) &&
362                 (migration->device_state & VFIO_DEVICE_STATE_RUNNING)) {
363                 continue;
364             } else {
365                 return false;
366             }
367         }
368     }
369     return true;
370 }
371 
372 static int vfio_dma_unmap_bitmap(VFIOContainer *container,
373                                  hwaddr iova, ram_addr_t size,
374                                  IOMMUTLBEntry *iotlb)
375 {
376     struct vfio_iommu_type1_dma_unmap *unmap;
377     struct vfio_bitmap *bitmap;
378     uint64_t pages = TARGET_PAGE_ALIGN(size) >> TARGET_PAGE_BITS;
379     int ret;
380 
381     unmap = g_malloc0(sizeof(*unmap) + sizeof(*bitmap));
382 
383     unmap->argsz = sizeof(*unmap) + sizeof(*bitmap);
384     unmap->iova = iova;
385     unmap->size = size;
386     unmap->flags |= VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP;
387     bitmap = (struct vfio_bitmap *)&unmap->data;
388 
389     /*
390      * cpu_physical_memory_set_dirty_lebitmap() expects pages in bitmap of
391      * TARGET_PAGE_SIZE to mark those dirty. Hence set bitmap_pgsize to
392      * TARGET_PAGE_SIZE.
393      */
394 
395     bitmap->pgsize = TARGET_PAGE_SIZE;
396     bitmap->size = ROUND_UP(pages, sizeof(__u64) * BITS_PER_BYTE) /
397                    BITS_PER_BYTE;
398 
399     if (bitmap->size > container->max_dirty_bitmap_size) {
400         error_report("UNMAP: Size of bitmap too big 0x%"PRIx64,
401                      (uint64_t)bitmap->size);
402         ret = -E2BIG;
403         goto unmap_exit;
404     }
405 
406     bitmap->data = g_try_malloc0(bitmap->size);
407     if (!bitmap->data) {
408         ret = -ENOMEM;
409         goto unmap_exit;
410     }
411 
412     ret = ioctl(container->fd, VFIO_IOMMU_UNMAP_DMA, unmap);
413     if (!ret) {
414         cpu_physical_memory_set_dirty_lebitmap((unsigned long *)bitmap->data,
415                 iotlb->translated_addr, pages);
416     } else {
417         error_report("VFIO_UNMAP_DMA with DIRTY_BITMAP : %m");
418     }
419 
420     g_free(bitmap->data);
421 unmap_exit:
422     g_free(unmap);
423     return ret;
424 }
425 
426 /*
427  * DMA - Mapping and unmapping for the "type1" IOMMU interface used on x86
428  */
429 static int vfio_dma_unmap(VFIOContainer *container,
430                           hwaddr iova, ram_addr_t size,
431                           IOMMUTLBEntry *iotlb)
432 {
433     struct vfio_iommu_type1_dma_unmap unmap = {
434         .argsz = sizeof(unmap),
435         .flags = 0,
436         .iova = iova,
437         .size = size,
438     };
439 
440     if (iotlb && container->dirty_pages_supported &&
441         vfio_devices_all_running_and_saving(container)) {
442         return vfio_dma_unmap_bitmap(container, iova, size, iotlb);
443     }
444 
445     while (ioctl(container->fd, VFIO_IOMMU_UNMAP_DMA, &unmap)) {
446         /*
447          * The type1 backend has an off-by-one bug in the kernel (71a7d3d78e3c
448          * v4.15) where an overflow in its wrap-around check prevents us from
449          * unmapping the last page of the address space.  Test for the error
450          * condition and re-try the unmap excluding the last page.  The
451          * expectation is that we've never mapped the last page anyway and this
452          * unmap request comes via vIOMMU support which also makes it unlikely
453          * that this page is used.  This bug was introduced well after type1 v2
454          * support was introduced, so we shouldn't need to test for v1.  A fix
455          * is queued for kernel v5.0 so this workaround can be removed once
456          * affected kernels are sufficiently deprecated.
457          */
458         if (errno == EINVAL && unmap.size && !(unmap.iova + unmap.size) &&
459             container->iommu_type == VFIO_TYPE1v2_IOMMU) {
460             trace_vfio_dma_unmap_overflow_workaround();
461             unmap.size -= 1ULL << ctz64(container->pgsizes);
462             continue;
463         }
464         error_report("VFIO_UNMAP_DMA failed: %s", strerror(errno));
465         return -errno;
466     }
467 
468     return 0;
469 }
470 
471 static int vfio_dma_map(VFIOContainer *container, hwaddr iova,
472                         ram_addr_t size, void *vaddr, bool readonly)
473 {
474     struct vfio_iommu_type1_dma_map map = {
475         .argsz = sizeof(map),
476         .flags = VFIO_DMA_MAP_FLAG_READ,
477         .vaddr = (__u64)(uintptr_t)vaddr,
478         .iova = iova,
479         .size = size,
480     };
481 
482     if (!readonly) {
483         map.flags |= VFIO_DMA_MAP_FLAG_WRITE;
484     }
485 
486     /*
487      * Try the mapping, if it fails with EBUSY, unmap the region and try
488      * again.  This shouldn't be necessary, but we sometimes see it in
489      * the VGA ROM space.
490      */
491     if (ioctl(container->fd, VFIO_IOMMU_MAP_DMA, &map) == 0 ||
492         (errno == EBUSY && vfio_dma_unmap(container, iova, size, NULL) == 0 &&
493          ioctl(container->fd, VFIO_IOMMU_MAP_DMA, &map) == 0)) {
494         return 0;
495     }
496 
497     error_report("VFIO_MAP_DMA failed: %s", strerror(errno));
498     return -errno;
499 }
500 
501 static void vfio_host_win_add(VFIOContainer *container,
502                               hwaddr min_iova, hwaddr max_iova,
503                               uint64_t iova_pgsizes)
504 {
505     VFIOHostDMAWindow *hostwin;
506 
507     QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) {
508         if (ranges_overlap(hostwin->min_iova,
509                            hostwin->max_iova - hostwin->min_iova + 1,
510                            min_iova,
511                            max_iova - min_iova + 1)) {
512             hw_error("%s: Overlapped IOMMU are not enabled", __func__);
513         }
514     }
515 
516     hostwin = g_malloc0(sizeof(*hostwin));
517 
518     hostwin->min_iova = min_iova;
519     hostwin->max_iova = max_iova;
520     hostwin->iova_pgsizes = iova_pgsizes;
521     QLIST_INSERT_HEAD(&container->hostwin_list, hostwin, hostwin_next);
522 }
523 
524 static int vfio_host_win_del(VFIOContainer *container, hwaddr min_iova,
525                              hwaddr max_iova)
526 {
527     VFIOHostDMAWindow *hostwin;
528 
529     QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) {
530         if (hostwin->min_iova == min_iova && hostwin->max_iova == max_iova) {
531             QLIST_REMOVE(hostwin, hostwin_next);
532             return 0;
533         }
534     }
535 
536     return -1;
537 }
538 
539 static bool vfio_listener_skipped_section(MemoryRegionSection *section)
540 {
541     return (!memory_region_is_ram(section->mr) &&
542             !memory_region_is_iommu(section->mr)) ||
543            /*
544             * Sizing an enabled 64-bit BAR can cause spurious mappings to
545             * addresses in the upper part of the 64-bit address space.  These
546             * are never accessed by the CPU and beyond the address width of
547             * some IOMMU hardware.  TODO: VFIO should tell us the IOMMU width.
548             */
549            section->offset_within_address_space & (1ULL << 63);
550 }
551 
552 /* Called with rcu_read_lock held.  */
553 static bool vfio_get_xlat_addr(IOMMUTLBEntry *iotlb, void **vaddr,
554                                ram_addr_t *ram_addr, bool *read_only)
555 {
556     MemoryRegion *mr;
557     hwaddr xlat;
558     hwaddr len = iotlb->addr_mask + 1;
559     bool writable = iotlb->perm & IOMMU_WO;
560 
561     /*
562      * The IOMMU TLB entry we have just covers translation through
563      * this IOMMU to its immediate target.  We need to translate
564      * it the rest of the way through to memory.
565      */
566     mr = address_space_translate(&address_space_memory,
567                                  iotlb->translated_addr,
568                                  &xlat, &len, writable,
569                                  MEMTXATTRS_UNSPECIFIED);
570     if (!memory_region_is_ram(mr)) {
571         error_report("iommu map to non memory area %"HWADDR_PRIx"",
572                      xlat);
573         return false;
574     }
575 
576     /*
577      * Translation truncates length to the IOMMU page size,
578      * check that it did not truncate too much.
579      */
580     if (len & iotlb->addr_mask) {
581         error_report("iommu has granularity incompatible with target AS");
582         return false;
583     }
584 
585     if (vaddr) {
586         *vaddr = memory_region_get_ram_ptr(mr) + xlat;
587     }
588 
589     if (ram_addr) {
590         *ram_addr = memory_region_get_ram_addr(mr) + xlat;
591     }
592 
593     if (read_only) {
594         *read_only = !writable || mr->readonly;
595     }
596 
597     return true;
598 }
599 
600 static void vfio_iommu_map_notify(IOMMUNotifier *n, IOMMUTLBEntry *iotlb)
601 {
602     VFIOGuestIOMMU *giommu = container_of(n, VFIOGuestIOMMU, n);
603     VFIOContainer *container = giommu->container;
604     hwaddr iova = iotlb->iova + giommu->iommu_offset;
605     void *vaddr;
606     int ret;
607 
608     trace_vfio_iommu_map_notify(iotlb->perm == IOMMU_NONE ? "UNMAP" : "MAP",
609                                 iova, iova + iotlb->addr_mask);
610 
611     if (iotlb->target_as != &address_space_memory) {
612         error_report("Wrong target AS \"%s\", only system memory is allowed",
613                      iotlb->target_as->name ? iotlb->target_as->name : "none");
614         return;
615     }
616 
617     rcu_read_lock();
618 
619     if ((iotlb->perm & IOMMU_RW) != IOMMU_NONE) {
620         bool read_only;
621 
622         if (!vfio_get_xlat_addr(iotlb, &vaddr, NULL, &read_only)) {
623             goto out;
624         }
625         /*
626          * vaddr is only valid until rcu_read_unlock(). But after
627          * vfio_dma_map has set up the mapping the pages will be
628          * pinned by the kernel. This makes sure that the RAM backend
629          * of vaddr will always be there, even if the memory object is
630          * destroyed and its backing memory munmap-ed.
631          */
632         ret = vfio_dma_map(container, iova,
633                            iotlb->addr_mask + 1, vaddr,
634                            read_only);
635         if (ret) {
636             error_report("vfio_dma_map(%p, 0x%"HWADDR_PRIx", "
637                          "0x%"HWADDR_PRIx", %p) = %d (%m)",
638                          container, iova,
639                          iotlb->addr_mask + 1, vaddr, ret);
640         }
641     } else {
642         ret = vfio_dma_unmap(container, iova, iotlb->addr_mask + 1, iotlb);
643         if (ret) {
644             error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx", "
645                          "0x%"HWADDR_PRIx") = %d (%m)",
646                          container, iova,
647                          iotlb->addr_mask + 1, ret);
648         }
649     }
650 out:
651     rcu_read_unlock();
652 }
653 
654 static void vfio_listener_region_add(MemoryListener *listener,
655                                      MemoryRegionSection *section)
656 {
657     VFIOContainer *container = container_of(listener, VFIOContainer, listener);
658     hwaddr iova, end;
659     Int128 llend, llsize;
660     void *vaddr;
661     int ret;
662     VFIOHostDMAWindow *hostwin;
663     bool hostwin_found;
664     Error *err = NULL;
665 
666     if (vfio_listener_skipped_section(section)) {
667         trace_vfio_listener_region_add_skip(
668                 section->offset_within_address_space,
669                 section->offset_within_address_space +
670                 int128_get64(int128_sub(section->size, int128_one())));
671         return;
672     }
673 
674     if (unlikely((section->offset_within_address_space & ~TARGET_PAGE_MASK) !=
675                  (section->offset_within_region & ~TARGET_PAGE_MASK))) {
676         error_report("%s received unaligned region", __func__);
677         return;
678     }
679 
680     iova = TARGET_PAGE_ALIGN(section->offset_within_address_space);
681     llend = int128_make64(section->offset_within_address_space);
682     llend = int128_add(llend, section->size);
683     llend = int128_and(llend, int128_exts64(TARGET_PAGE_MASK));
684 
685     if (int128_ge(int128_make64(iova), llend)) {
686         return;
687     }
688     end = int128_get64(int128_sub(llend, int128_one()));
689 
690     if (container->iommu_type == VFIO_SPAPR_TCE_v2_IOMMU) {
691         hwaddr pgsize = 0;
692 
693         /* For now intersections are not allowed, we may relax this later */
694         QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) {
695             if (ranges_overlap(hostwin->min_iova,
696                                hostwin->max_iova - hostwin->min_iova + 1,
697                                section->offset_within_address_space,
698                                int128_get64(section->size))) {
699                 error_setg(&err,
700                     "region [0x%"PRIx64",0x%"PRIx64"] overlaps with existing"
701                     "host DMA window [0x%"PRIx64",0x%"PRIx64"]",
702                     section->offset_within_address_space,
703                     section->offset_within_address_space +
704                         int128_get64(section->size) - 1,
705                     hostwin->min_iova, hostwin->max_iova);
706                 goto fail;
707             }
708         }
709 
710         ret = vfio_spapr_create_window(container, section, &pgsize);
711         if (ret) {
712             error_setg_errno(&err, -ret, "Failed to create SPAPR window");
713             goto fail;
714         }
715 
716         vfio_host_win_add(container, section->offset_within_address_space,
717                           section->offset_within_address_space +
718                           int128_get64(section->size) - 1, pgsize);
719 #ifdef CONFIG_KVM
720         if (kvm_enabled()) {
721             VFIOGroup *group;
722             IOMMUMemoryRegion *iommu_mr = IOMMU_MEMORY_REGION(section->mr);
723             struct kvm_vfio_spapr_tce param;
724             struct kvm_device_attr attr = {
725                 .group = KVM_DEV_VFIO_GROUP,
726                 .attr = KVM_DEV_VFIO_GROUP_SET_SPAPR_TCE,
727                 .addr = (uint64_t)(unsigned long)&param,
728             };
729 
730             if (!memory_region_iommu_get_attr(iommu_mr, IOMMU_ATTR_SPAPR_TCE_FD,
731                                               &param.tablefd)) {
732                 QLIST_FOREACH(group, &container->group_list, container_next) {
733                     param.groupfd = group->fd;
734                     if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) {
735                         error_report("vfio: failed to setup fd %d "
736                                      "for a group with fd %d: %s",
737                                      param.tablefd, param.groupfd,
738                                      strerror(errno));
739                         return;
740                     }
741                     trace_vfio_spapr_group_attach(param.groupfd, param.tablefd);
742                 }
743             }
744         }
745 #endif
746     }
747 
748     hostwin_found = false;
749     QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) {
750         if (hostwin->min_iova <= iova && end <= hostwin->max_iova) {
751             hostwin_found = true;
752             break;
753         }
754     }
755 
756     if (!hostwin_found) {
757         error_setg(&err, "Container %p can't map guest IOVA region"
758                    " 0x%"HWADDR_PRIx"..0x%"HWADDR_PRIx, container, iova, end);
759         goto fail;
760     }
761 
762     memory_region_ref(section->mr);
763 
764     if (memory_region_is_iommu(section->mr)) {
765         VFIOGuestIOMMU *giommu;
766         IOMMUMemoryRegion *iommu_mr = IOMMU_MEMORY_REGION(section->mr);
767         int iommu_idx;
768 
769         trace_vfio_listener_region_add_iommu(iova, end);
770         /*
771          * FIXME: For VFIO iommu types which have KVM acceleration to
772          * avoid bouncing all map/unmaps through qemu this way, this
773          * would be the right place to wire that up (tell the KVM
774          * device emulation the VFIO iommu handles to use).
775          */
776         giommu = g_malloc0(sizeof(*giommu));
777         giommu->iommu = iommu_mr;
778         giommu->iommu_offset = section->offset_within_address_space -
779                                section->offset_within_region;
780         giommu->container = container;
781         llend = int128_add(int128_make64(section->offset_within_region),
782                            section->size);
783         llend = int128_sub(llend, int128_one());
784         iommu_idx = memory_region_iommu_attrs_to_index(iommu_mr,
785                                                        MEMTXATTRS_UNSPECIFIED);
786         iommu_notifier_init(&giommu->n, vfio_iommu_map_notify,
787                             IOMMU_NOTIFIER_ALL,
788                             section->offset_within_region,
789                             int128_get64(llend),
790                             iommu_idx);
791 
792         ret = memory_region_register_iommu_notifier(section->mr, &giommu->n,
793                                                     &err);
794         if (ret) {
795             g_free(giommu);
796             goto fail;
797         }
798         QLIST_INSERT_HEAD(&container->giommu_list, giommu, giommu_next);
799         memory_region_iommu_replay(giommu->iommu, &giommu->n);
800 
801         return;
802     }
803 
804     /* Here we assume that memory_region_is_ram(section->mr)==true */
805 
806     vaddr = memory_region_get_ram_ptr(section->mr) +
807             section->offset_within_region +
808             (iova - section->offset_within_address_space);
809 
810     trace_vfio_listener_region_add_ram(iova, end, vaddr);
811 
812     llsize = int128_sub(llend, int128_make64(iova));
813 
814     if (memory_region_is_ram_device(section->mr)) {
815         hwaddr pgmask = (1ULL << ctz64(hostwin->iova_pgsizes)) - 1;
816 
817         if ((iova & pgmask) || (int128_get64(llsize) & pgmask)) {
818             trace_vfio_listener_region_add_no_dma_map(
819                 memory_region_name(section->mr),
820                 section->offset_within_address_space,
821                 int128_getlo(section->size),
822                 pgmask + 1);
823             return;
824         }
825     }
826 
827     ret = vfio_dma_map(container, iova, int128_get64(llsize),
828                        vaddr, section->readonly);
829     if (ret) {
830         error_setg(&err, "vfio_dma_map(%p, 0x%"HWADDR_PRIx", "
831                    "0x%"HWADDR_PRIx", %p) = %d (%m)",
832                    container, iova, int128_get64(llsize), vaddr, ret);
833         if (memory_region_is_ram_device(section->mr)) {
834             /* Allow unexpected mappings not to be fatal for RAM devices */
835             error_report_err(err);
836             return;
837         }
838         goto fail;
839     }
840 
841     return;
842 
843 fail:
844     if (memory_region_is_ram_device(section->mr)) {
845         error_report("failed to vfio_dma_map. pci p2p may not work");
846         return;
847     }
848     /*
849      * On the initfn path, store the first error in the container so we
850      * can gracefully fail.  Runtime, there's not much we can do other
851      * than throw a hardware error.
852      */
853     if (!container->initialized) {
854         if (!container->error) {
855             error_propagate_prepend(&container->error, err,
856                                     "Region %s: ",
857                                     memory_region_name(section->mr));
858         } else {
859             error_free(err);
860         }
861     } else {
862         error_report_err(err);
863         hw_error("vfio: DMA mapping failed, unable to continue");
864     }
865 }
866 
867 static void vfio_listener_region_del(MemoryListener *listener,
868                                      MemoryRegionSection *section)
869 {
870     VFIOContainer *container = container_of(listener, VFIOContainer, listener);
871     hwaddr iova, end;
872     Int128 llend, llsize;
873     int ret;
874     bool try_unmap = true;
875 
876     if (vfio_listener_skipped_section(section)) {
877         trace_vfio_listener_region_del_skip(
878                 section->offset_within_address_space,
879                 section->offset_within_address_space +
880                 int128_get64(int128_sub(section->size, int128_one())));
881         return;
882     }
883 
884     if (unlikely((section->offset_within_address_space & ~TARGET_PAGE_MASK) !=
885                  (section->offset_within_region & ~TARGET_PAGE_MASK))) {
886         error_report("%s received unaligned region", __func__);
887         return;
888     }
889 
890     if (memory_region_is_iommu(section->mr)) {
891         VFIOGuestIOMMU *giommu;
892 
893         QLIST_FOREACH(giommu, &container->giommu_list, giommu_next) {
894             if (MEMORY_REGION(giommu->iommu) == section->mr &&
895                 giommu->n.start == section->offset_within_region) {
896                 memory_region_unregister_iommu_notifier(section->mr,
897                                                         &giommu->n);
898                 QLIST_REMOVE(giommu, giommu_next);
899                 g_free(giommu);
900                 break;
901             }
902         }
903 
904         /*
905          * FIXME: We assume the one big unmap below is adequate to
906          * remove any individual page mappings in the IOMMU which
907          * might have been copied into VFIO. This works for a page table
908          * based IOMMU where a big unmap flattens a large range of IO-PTEs.
909          * That may not be true for all IOMMU types.
910          */
911     }
912 
913     iova = TARGET_PAGE_ALIGN(section->offset_within_address_space);
914     llend = int128_make64(section->offset_within_address_space);
915     llend = int128_add(llend, section->size);
916     llend = int128_and(llend, int128_exts64(TARGET_PAGE_MASK));
917 
918     if (int128_ge(int128_make64(iova), llend)) {
919         return;
920     }
921     end = int128_get64(int128_sub(llend, int128_one()));
922 
923     llsize = int128_sub(llend, int128_make64(iova));
924 
925     trace_vfio_listener_region_del(iova, end);
926 
927     if (memory_region_is_ram_device(section->mr)) {
928         hwaddr pgmask;
929         VFIOHostDMAWindow *hostwin;
930         bool hostwin_found = false;
931 
932         QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) {
933             if (hostwin->min_iova <= iova && end <= hostwin->max_iova) {
934                 hostwin_found = true;
935                 break;
936             }
937         }
938         assert(hostwin_found); /* or region_add() would have failed */
939 
940         pgmask = (1ULL << ctz64(hostwin->iova_pgsizes)) - 1;
941         try_unmap = !((iova & pgmask) || (int128_get64(llsize) & pgmask));
942     }
943 
944     if (try_unmap) {
945         ret = vfio_dma_unmap(container, iova, int128_get64(llsize), NULL);
946         if (ret) {
947             error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx", "
948                          "0x%"HWADDR_PRIx") = %d (%m)",
949                          container, iova, int128_get64(llsize), ret);
950         }
951     }
952 
953     memory_region_unref(section->mr);
954 
955     if (container->iommu_type == VFIO_SPAPR_TCE_v2_IOMMU) {
956         vfio_spapr_remove_window(container,
957                                  section->offset_within_address_space);
958         if (vfio_host_win_del(container,
959                               section->offset_within_address_space,
960                               section->offset_within_address_space +
961                               int128_get64(section->size) - 1) < 0) {
962             hw_error("%s: Cannot delete missing window at %"HWADDR_PRIx,
963                      __func__, section->offset_within_address_space);
964         }
965     }
966 }
967 
968 static int vfio_get_dirty_bitmap(VFIOContainer *container, uint64_t iova,
969                                  uint64_t size, ram_addr_t ram_addr)
970 {
971     struct vfio_iommu_type1_dirty_bitmap *dbitmap;
972     struct vfio_iommu_type1_dirty_bitmap_get *range;
973     uint64_t pages;
974     int ret;
975 
976     dbitmap = g_malloc0(sizeof(*dbitmap) + sizeof(*range));
977 
978     dbitmap->argsz = sizeof(*dbitmap) + sizeof(*range);
979     dbitmap->flags = VFIO_IOMMU_DIRTY_PAGES_FLAG_GET_BITMAP;
980     range = (struct vfio_iommu_type1_dirty_bitmap_get *)&dbitmap->data;
981     range->iova = iova;
982     range->size = size;
983 
984     /*
985      * cpu_physical_memory_set_dirty_lebitmap() expects pages in bitmap of
986      * TARGET_PAGE_SIZE to mark those dirty. Hence set bitmap's pgsize to
987      * TARGET_PAGE_SIZE.
988      */
989     range->bitmap.pgsize = TARGET_PAGE_SIZE;
990 
991     pages = TARGET_PAGE_ALIGN(range->size) >> TARGET_PAGE_BITS;
992     range->bitmap.size = ROUND_UP(pages, sizeof(__u64) * BITS_PER_BYTE) /
993                                          BITS_PER_BYTE;
994     range->bitmap.data = g_try_malloc0(range->bitmap.size);
995     if (!range->bitmap.data) {
996         ret = -ENOMEM;
997         goto err_out;
998     }
999 
1000     ret = ioctl(container->fd, VFIO_IOMMU_DIRTY_PAGES, dbitmap);
1001     if (ret) {
1002         error_report("Failed to get dirty bitmap for iova: 0x%"PRIx64
1003                 " size: 0x%"PRIx64" err: %d", (uint64_t)range->iova,
1004                 (uint64_t)range->size, errno);
1005         goto err_out;
1006     }
1007 
1008     cpu_physical_memory_set_dirty_lebitmap((unsigned long *)range->bitmap.data,
1009                                             ram_addr, pages);
1010 
1011     trace_vfio_get_dirty_bitmap(container->fd, range->iova, range->size,
1012                                 range->bitmap.size, ram_addr);
1013 err_out:
1014     g_free(range->bitmap.data);
1015     g_free(dbitmap);
1016 
1017     return ret;
1018 }
1019 
1020 typedef struct {
1021     IOMMUNotifier n;
1022     VFIOGuestIOMMU *giommu;
1023 } vfio_giommu_dirty_notifier;
1024 
1025 static void vfio_iommu_map_dirty_notify(IOMMUNotifier *n, IOMMUTLBEntry *iotlb)
1026 {
1027     vfio_giommu_dirty_notifier *gdn = container_of(n,
1028                                                 vfio_giommu_dirty_notifier, n);
1029     VFIOGuestIOMMU *giommu = gdn->giommu;
1030     VFIOContainer *container = giommu->container;
1031     hwaddr iova = iotlb->iova + giommu->iommu_offset;
1032     ram_addr_t translated_addr;
1033 
1034     trace_vfio_iommu_map_dirty_notify(iova, iova + iotlb->addr_mask);
1035 
1036     if (iotlb->target_as != &address_space_memory) {
1037         error_report("Wrong target AS \"%s\", only system memory is allowed",
1038                      iotlb->target_as->name ? iotlb->target_as->name : "none");
1039         return;
1040     }
1041 
1042     rcu_read_lock();
1043     if (vfio_get_xlat_addr(iotlb, NULL, &translated_addr, NULL)) {
1044         int ret;
1045 
1046         ret = vfio_get_dirty_bitmap(container, iova, iotlb->addr_mask + 1,
1047                                     translated_addr);
1048         if (ret) {
1049             error_report("vfio_iommu_map_dirty_notify(%p, 0x%"HWADDR_PRIx", "
1050                          "0x%"HWADDR_PRIx") = %d (%m)",
1051                          container, iova,
1052                          iotlb->addr_mask + 1, ret);
1053         }
1054     }
1055     rcu_read_unlock();
1056 }
1057 
1058 static int vfio_sync_dirty_bitmap(VFIOContainer *container,
1059                                   MemoryRegionSection *section)
1060 {
1061     ram_addr_t ram_addr;
1062 
1063     if (memory_region_is_iommu(section->mr)) {
1064         VFIOGuestIOMMU *giommu;
1065 
1066         QLIST_FOREACH(giommu, &container->giommu_list, giommu_next) {
1067             if (MEMORY_REGION(giommu->iommu) == section->mr &&
1068                 giommu->n.start == section->offset_within_region) {
1069                 Int128 llend;
1070                 vfio_giommu_dirty_notifier gdn = { .giommu = giommu };
1071                 int idx = memory_region_iommu_attrs_to_index(giommu->iommu,
1072                                                        MEMTXATTRS_UNSPECIFIED);
1073 
1074                 llend = int128_add(int128_make64(section->offset_within_region),
1075                                    section->size);
1076                 llend = int128_sub(llend, int128_one());
1077 
1078                 iommu_notifier_init(&gdn.n,
1079                                     vfio_iommu_map_dirty_notify,
1080                                     IOMMU_NOTIFIER_MAP,
1081                                     section->offset_within_region,
1082                                     int128_get64(llend),
1083                                     idx);
1084                 memory_region_iommu_replay(giommu->iommu, &gdn.n);
1085                 break;
1086             }
1087         }
1088         return 0;
1089     }
1090 
1091     ram_addr = memory_region_get_ram_addr(section->mr) +
1092                section->offset_within_region;
1093 
1094     return vfio_get_dirty_bitmap(container,
1095                        TARGET_PAGE_ALIGN(section->offset_within_address_space),
1096                        int128_get64(section->size), ram_addr);
1097 }
1098 
1099 static void vfio_listerner_log_sync(MemoryListener *listener,
1100         MemoryRegionSection *section)
1101 {
1102     VFIOContainer *container = container_of(listener, VFIOContainer, listener);
1103 
1104     if (vfio_listener_skipped_section(section) ||
1105         !container->dirty_pages_supported) {
1106         return;
1107     }
1108 
1109     if (vfio_devices_all_stopped_and_saving(container)) {
1110         vfio_sync_dirty_bitmap(container, section);
1111     }
1112 }
1113 
1114 static const MemoryListener vfio_memory_listener = {
1115     .region_add = vfio_listener_region_add,
1116     .region_del = vfio_listener_region_del,
1117     .log_sync = vfio_listerner_log_sync,
1118 };
1119 
1120 static void vfio_listener_release(VFIOContainer *container)
1121 {
1122     memory_listener_unregister(&container->listener);
1123     if (container->iommu_type == VFIO_SPAPR_TCE_v2_IOMMU) {
1124         memory_listener_unregister(&container->prereg_listener);
1125     }
1126 }
1127 
1128 static struct vfio_info_cap_header *
1129 vfio_get_cap(void *ptr, uint32_t cap_offset, uint16_t id)
1130 {
1131     struct vfio_info_cap_header *hdr;
1132 
1133     for (hdr = ptr + cap_offset; hdr != ptr; hdr = ptr + hdr->next) {
1134         if (hdr->id == id) {
1135             return hdr;
1136         }
1137     }
1138 
1139     return NULL;
1140 }
1141 
1142 struct vfio_info_cap_header *
1143 vfio_get_region_info_cap(struct vfio_region_info *info, uint16_t id)
1144 {
1145     if (!(info->flags & VFIO_REGION_INFO_FLAG_CAPS)) {
1146         return NULL;
1147     }
1148 
1149     return vfio_get_cap((void *)info, info->cap_offset, id);
1150 }
1151 
1152 static struct vfio_info_cap_header *
1153 vfio_get_iommu_type1_info_cap(struct vfio_iommu_type1_info *info, uint16_t id)
1154 {
1155     if (!(info->flags & VFIO_IOMMU_INFO_CAPS)) {
1156         return NULL;
1157     }
1158 
1159     return vfio_get_cap((void *)info, info->cap_offset, id);
1160 }
1161 
1162 struct vfio_info_cap_header *
1163 vfio_get_device_info_cap(struct vfio_device_info *info, uint16_t id)
1164 {
1165     if (!(info->flags & VFIO_DEVICE_FLAGS_CAPS)) {
1166         return NULL;
1167     }
1168 
1169     return vfio_get_cap((void *)info, info->cap_offset, id);
1170 }
1171 
1172 bool vfio_get_info_dma_avail(struct vfio_iommu_type1_info *info,
1173                              unsigned int *avail)
1174 {
1175     struct vfio_info_cap_header *hdr;
1176     struct vfio_iommu_type1_info_dma_avail *cap;
1177 
1178     /* If the capability cannot be found, assume no DMA limiting */
1179     hdr = vfio_get_iommu_type1_info_cap(info,
1180                                         VFIO_IOMMU_TYPE1_INFO_DMA_AVAIL);
1181     if (hdr == NULL) {
1182         return false;
1183     }
1184 
1185     if (avail != NULL) {
1186         cap = (void *) hdr;
1187         *avail = cap->avail;
1188     }
1189 
1190     return true;
1191 }
1192 
1193 static int vfio_setup_region_sparse_mmaps(VFIORegion *region,
1194                                           struct vfio_region_info *info)
1195 {
1196     struct vfio_info_cap_header *hdr;
1197     struct vfio_region_info_cap_sparse_mmap *sparse;
1198     int i, j;
1199 
1200     hdr = vfio_get_region_info_cap(info, VFIO_REGION_INFO_CAP_SPARSE_MMAP);
1201     if (!hdr) {
1202         return -ENODEV;
1203     }
1204 
1205     sparse = container_of(hdr, struct vfio_region_info_cap_sparse_mmap, header);
1206 
1207     trace_vfio_region_sparse_mmap_header(region->vbasedev->name,
1208                                          region->nr, sparse->nr_areas);
1209 
1210     region->mmaps = g_new0(VFIOMmap, sparse->nr_areas);
1211 
1212     for (i = 0, j = 0; i < sparse->nr_areas; i++) {
1213         trace_vfio_region_sparse_mmap_entry(i, sparse->areas[i].offset,
1214                                             sparse->areas[i].offset +
1215                                             sparse->areas[i].size);
1216 
1217         if (sparse->areas[i].size) {
1218             region->mmaps[j].offset = sparse->areas[i].offset;
1219             region->mmaps[j].size = sparse->areas[i].size;
1220             j++;
1221         }
1222     }
1223 
1224     region->nr_mmaps = j;
1225     region->mmaps = g_realloc(region->mmaps, j * sizeof(VFIOMmap));
1226 
1227     return 0;
1228 }
1229 
1230 int vfio_region_setup(Object *obj, VFIODevice *vbasedev, VFIORegion *region,
1231                       int index, const char *name)
1232 {
1233     struct vfio_region_info *info;
1234     int ret;
1235 
1236     ret = vfio_get_region_info(vbasedev, index, &info);
1237     if (ret) {
1238         return ret;
1239     }
1240 
1241     region->vbasedev = vbasedev;
1242     region->flags = info->flags;
1243     region->size = info->size;
1244     region->fd_offset = info->offset;
1245     region->nr = index;
1246 
1247     if (region->size) {
1248         region->mem = g_new0(MemoryRegion, 1);
1249         memory_region_init_io(region->mem, obj, &vfio_region_ops,
1250                               region, name, region->size);
1251 
1252         if (!vbasedev->no_mmap &&
1253             region->flags & VFIO_REGION_INFO_FLAG_MMAP) {
1254 
1255             ret = vfio_setup_region_sparse_mmaps(region, info);
1256 
1257             if (ret) {
1258                 region->nr_mmaps = 1;
1259                 region->mmaps = g_new0(VFIOMmap, region->nr_mmaps);
1260                 region->mmaps[0].offset = 0;
1261                 region->mmaps[0].size = region->size;
1262             }
1263         }
1264     }
1265 
1266     g_free(info);
1267 
1268     trace_vfio_region_setup(vbasedev->name, index, name,
1269                             region->flags, region->fd_offset, region->size);
1270     return 0;
1271 }
1272 
1273 static void vfio_subregion_unmap(VFIORegion *region, int index)
1274 {
1275     trace_vfio_region_unmap(memory_region_name(&region->mmaps[index].mem),
1276                             region->mmaps[index].offset,
1277                             region->mmaps[index].offset +
1278                             region->mmaps[index].size - 1);
1279     memory_region_del_subregion(region->mem, &region->mmaps[index].mem);
1280     munmap(region->mmaps[index].mmap, region->mmaps[index].size);
1281     object_unparent(OBJECT(&region->mmaps[index].mem));
1282     region->mmaps[index].mmap = NULL;
1283 }
1284 
1285 int vfio_region_mmap(VFIORegion *region)
1286 {
1287     int i, prot = 0;
1288     char *name;
1289 
1290     if (!region->mem) {
1291         return 0;
1292     }
1293 
1294     prot |= region->flags & VFIO_REGION_INFO_FLAG_READ ? PROT_READ : 0;
1295     prot |= region->flags & VFIO_REGION_INFO_FLAG_WRITE ? PROT_WRITE : 0;
1296 
1297     for (i = 0; i < region->nr_mmaps; i++) {
1298         region->mmaps[i].mmap = mmap(NULL, region->mmaps[i].size, prot,
1299                                      MAP_SHARED, region->vbasedev->fd,
1300                                      region->fd_offset +
1301                                      region->mmaps[i].offset);
1302         if (region->mmaps[i].mmap == MAP_FAILED) {
1303             int ret = -errno;
1304 
1305             trace_vfio_region_mmap_fault(memory_region_name(region->mem), i,
1306                                          region->fd_offset +
1307                                          region->mmaps[i].offset,
1308                                          region->fd_offset +
1309                                          region->mmaps[i].offset +
1310                                          region->mmaps[i].size - 1, ret);
1311 
1312             region->mmaps[i].mmap = NULL;
1313 
1314             for (i--; i >= 0; i--) {
1315                 vfio_subregion_unmap(region, i);
1316             }
1317 
1318             return ret;
1319         }
1320 
1321         name = g_strdup_printf("%s mmaps[%d]",
1322                                memory_region_name(region->mem), i);
1323         memory_region_init_ram_device_ptr(&region->mmaps[i].mem,
1324                                           memory_region_owner(region->mem),
1325                                           name, region->mmaps[i].size,
1326                                           region->mmaps[i].mmap);
1327         g_free(name);
1328         memory_region_add_subregion(region->mem, region->mmaps[i].offset,
1329                                     &region->mmaps[i].mem);
1330 
1331         trace_vfio_region_mmap(memory_region_name(&region->mmaps[i].mem),
1332                                region->mmaps[i].offset,
1333                                region->mmaps[i].offset +
1334                                region->mmaps[i].size - 1);
1335     }
1336 
1337     return 0;
1338 }
1339 
1340 void vfio_region_unmap(VFIORegion *region)
1341 {
1342     int i;
1343 
1344     if (!region->mem) {
1345         return;
1346     }
1347 
1348     for (i = 0; i < region->nr_mmaps; i++) {
1349         if (region->mmaps[i].mmap) {
1350             vfio_subregion_unmap(region, i);
1351         }
1352     }
1353 }
1354 
1355 void vfio_region_exit(VFIORegion *region)
1356 {
1357     int i;
1358 
1359     if (!region->mem) {
1360         return;
1361     }
1362 
1363     for (i = 0; i < region->nr_mmaps; i++) {
1364         if (region->mmaps[i].mmap) {
1365             memory_region_del_subregion(region->mem, &region->mmaps[i].mem);
1366         }
1367     }
1368 
1369     trace_vfio_region_exit(region->vbasedev->name, region->nr);
1370 }
1371 
1372 void vfio_region_finalize(VFIORegion *region)
1373 {
1374     int i;
1375 
1376     if (!region->mem) {
1377         return;
1378     }
1379 
1380     for (i = 0; i < region->nr_mmaps; i++) {
1381         if (region->mmaps[i].mmap) {
1382             munmap(region->mmaps[i].mmap, region->mmaps[i].size);
1383             object_unparent(OBJECT(&region->mmaps[i].mem));
1384         }
1385     }
1386 
1387     object_unparent(OBJECT(region->mem));
1388 
1389     g_free(region->mem);
1390     g_free(region->mmaps);
1391 
1392     trace_vfio_region_finalize(region->vbasedev->name, region->nr);
1393 
1394     region->mem = NULL;
1395     region->mmaps = NULL;
1396     region->nr_mmaps = 0;
1397     region->size = 0;
1398     region->flags = 0;
1399     region->nr = 0;
1400 }
1401 
1402 void vfio_region_mmaps_set_enabled(VFIORegion *region, bool enabled)
1403 {
1404     int i;
1405 
1406     if (!region->mem) {
1407         return;
1408     }
1409 
1410     for (i = 0; i < region->nr_mmaps; i++) {
1411         if (region->mmaps[i].mmap) {
1412             memory_region_set_enabled(&region->mmaps[i].mem, enabled);
1413         }
1414     }
1415 
1416     trace_vfio_region_mmaps_set_enabled(memory_region_name(region->mem),
1417                                         enabled);
1418 }
1419 
1420 void vfio_reset_handler(void *opaque)
1421 {
1422     VFIOGroup *group;
1423     VFIODevice *vbasedev;
1424 
1425     QLIST_FOREACH(group, &vfio_group_list, next) {
1426         QLIST_FOREACH(vbasedev, &group->device_list, next) {
1427             if (vbasedev->dev->realized) {
1428                 vbasedev->ops->vfio_compute_needs_reset(vbasedev);
1429             }
1430         }
1431     }
1432 
1433     QLIST_FOREACH(group, &vfio_group_list, next) {
1434         QLIST_FOREACH(vbasedev, &group->device_list, next) {
1435             if (vbasedev->dev->realized && vbasedev->needs_reset) {
1436                 vbasedev->ops->vfio_hot_reset_multi(vbasedev);
1437             }
1438         }
1439     }
1440 }
1441 
1442 static void vfio_kvm_device_add_group(VFIOGroup *group)
1443 {
1444 #ifdef CONFIG_KVM
1445     struct kvm_device_attr attr = {
1446         .group = KVM_DEV_VFIO_GROUP,
1447         .attr = KVM_DEV_VFIO_GROUP_ADD,
1448         .addr = (uint64_t)(unsigned long)&group->fd,
1449     };
1450 
1451     if (!kvm_enabled()) {
1452         return;
1453     }
1454 
1455     if (vfio_kvm_device_fd < 0) {
1456         struct kvm_create_device cd = {
1457             .type = KVM_DEV_TYPE_VFIO,
1458         };
1459 
1460         if (kvm_vm_ioctl(kvm_state, KVM_CREATE_DEVICE, &cd)) {
1461             error_report("Failed to create KVM VFIO device: %m");
1462             return;
1463         }
1464 
1465         vfio_kvm_device_fd = cd.fd;
1466     }
1467 
1468     if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) {
1469         error_report("Failed to add group %d to KVM VFIO device: %m",
1470                      group->groupid);
1471     }
1472 #endif
1473 }
1474 
1475 static void vfio_kvm_device_del_group(VFIOGroup *group)
1476 {
1477 #ifdef CONFIG_KVM
1478     struct kvm_device_attr attr = {
1479         .group = KVM_DEV_VFIO_GROUP,
1480         .attr = KVM_DEV_VFIO_GROUP_DEL,
1481         .addr = (uint64_t)(unsigned long)&group->fd,
1482     };
1483 
1484     if (vfio_kvm_device_fd < 0) {
1485         return;
1486     }
1487 
1488     if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) {
1489         error_report("Failed to remove group %d from KVM VFIO device: %m",
1490                      group->groupid);
1491     }
1492 #endif
1493 }
1494 
1495 static VFIOAddressSpace *vfio_get_address_space(AddressSpace *as)
1496 {
1497     VFIOAddressSpace *space;
1498 
1499     QLIST_FOREACH(space, &vfio_address_spaces, list) {
1500         if (space->as == as) {
1501             return space;
1502         }
1503     }
1504 
1505     /* No suitable VFIOAddressSpace, create a new one */
1506     space = g_malloc0(sizeof(*space));
1507     space->as = as;
1508     QLIST_INIT(&space->containers);
1509 
1510     QLIST_INSERT_HEAD(&vfio_address_spaces, space, list);
1511 
1512     return space;
1513 }
1514 
1515 static void vfio_put_address_space(VFIOAddressSpace *space)
1516 {
1517     if (QLIST_EMPTY(&space->containers)) {
1518         QLIST_REMOVE(space, list);
1519         g_free(space);
1520     }
1521 }
1522 
1523 /*
1524  * vfio_get_iommu_type - selects the richest iommu_type (v2 first)
1525  */
1526 static int vfio_get_iommu_type(VFIOContainer *container,
1527                                Error **errp)
1528 {
1529     int iommu_types[] = { VFIO_TYPE1v2_IOMMU, VFIO_TYPE1_IOMMU,
1530                           VFIO_SPAPR_TCE_v2_IOMMU, VFIO_SPAPR_TCE_IOMMU };
1531     int i;
1532 
1533     for (i = 0; i < ARRAY_SIZE(iommu_types); i++) {
1534         if (ioctl(container->fd, VFIO_CHECK_EXTENSION, iommu_types[i])) {
1535             return iommu_types[i];
1536         }
1537     }
1538     error_setg(errp, "No available IOMMU models");
1539     return -EINVAL;
1540 }
1541 
1542 static int vfio_init_container(VFIOContainer *container, int group_fd,
1543                                Error **errp)
1544 {
1545     int iommu_type, ret;
1546 
1547     iommu_type = vfio_get_iommu_type(container, errp);
1548     if (iommu_type < 0) {
1549         return iommu_type;
1550     }
1551 
1552     ret = ioctl(group_fd, VFIO_GROUP_SET_CONTAINER, &container->fd);
1553     if (ret) {
1554         error_setg_errno(errp, errno, "Failed to set group container");
1555         return -errno;
1556     }
1557 
1558     while (ioctl(container->fd, VFIO_SET_IOMMU, iommu_type)) {
1559         if (iommu_type == VFIO_SPAPR_TCE_v2_IOMMU) {
1560             /*
1561              * On sPAPR, despite the IOMMU subdriver always advertises v1 and
1562              * v2, the running platform may not support v2 and there is no
1563              * way to guess it until an IOMMU group gets added to the container.
1564              * So in case it fails with v2, try v1 as a fallback.
1565              */
1566             iommu_type = VFIO_SPAPR_TCE_IOMMU;
1567             continue;
1568         }
1569         error_setg_errno(errp, errno, "Failed to set iommu for container");
1570         return -errno;
1571     }
1572 
1573     container->iommu_type = iommu_type;
1574     return 0;
1575 }
1576 
1577 static int vfio_get_iommu_info(VFIOContainer *container,
1578                                struct vfio_iommu_type1_info **info)
1579 {
1580 
1581     size_t argsz = sizeof(struct vfio_iommu_type1_info);
1582 
1583     *info = g_new0(struct vfio_iommu_type1_info, 1);
1584 again:
1585     (*info)->argsz = argsz;
1586 
1587     if (ioctl(container->fd, VFIO_IOMMU_GET_INFO, *info)) {
1588         g_free(*info);
1589         *info = NULL;
1590         return -errno;
1591     }
1592 
1593     if (((*info)->argsz > argsz)) {
1594         argsz = (*info)->argsz;
1595         *info = g_realloc(*info, argsz);
1596         goto again;
1597     }
1598 
1599     return 0;
1600 }
1601 
1602 static struct vfio_info_cap_header *
1603 vfio_get_iommu_info_cap(struct vfio_iommu_type1_info *info, uint16_t id)
1604 {
1605     struct vfio_info_cap_header *hdr;
1606     void *ptr = info;
1607 
1608     if (!(info->flags & VFIO_IOMMU_INFO_CAPS)) {
1609         return NULL;
1610     }
1611 
1612     for (hdr = ptr + info->cap_offset; hdr != ptr; hdr = ptr + hdr->next) {
1613         if (hdr->id == id) {
1614             return hdr;
1615         }
1616     }
1617 
1618     return NULL;
1619 }
1620 
1621 static void vfio_get_iommu_info_migration(VFIOContainer *container,
1622                                          struct vfio_iommu_type1_info *info)
1623 {
1624     struct vfio_info_cap_header *hdr;
1625     struct vfio_iommu_type1_info_cap_migration *cap_mig;
1626 
1627     hdr = vfio_get_iommu_info_cap(info, VFIO_IOMMU_TYPE1_INFO_CAP_MIGRATION);
1628     if (!hdr) {
1629         return;
1630     }
1631 
1632     cap_mig = container_of(hdr, struct vfio_iommu_type1_info_cap_migration,
1633                             header);
1634 
1635     /*
1636      * cpu_physical_memory_set_dirty_lebitmap() expects pages in bitmap of
1637      * TARGET_PAGE_SIZE to mark those dirty.
1638      */
1639     if (cap_mig->pgsize_bitmap & TARGET_PAGE_SIZE) {
1640         container->dirty_pages_supported = true;
1641         container->max_dirty_bitmap_size = cap_mig->max_dirty_bitmap_size;
1642         container->dirty_pgsizes = cap_mig->pgsize_bitmap;
1643     }
1644 }
1645 
1646 static int vfio_connect_container(VFIOGroup *group, AddressSpace *as,
1647                                   Error **errp)
1648 {
1649     VFIOContainer *container;
1650     int ret, fd;
1651     VFIOAddressSpace *space;
1652 
1653     space = vfio_get_address_space(as);
1654 
1655     /*
1656      * VFIO is currently incompatible with discarding of RAM insofar as the
1657      * madvise to purge (zap) the page from QEMU's address space does not
1658      * interact with the memory API and therefore leaves stale virtual to
1659      * physical mappings in the IOMMU if the page was previously pinned.  We
1660      * therefore set discarding broken for each group added to a container,
1661      * whether the container is used individually or shared.  This provides
1662      * us with options to allow devices within a group to opt-in and allow
1663      * discarding, so long as it is done consistently for a group (for instance
1664      * if the device is an mdev device where it is known that the host vendor
1665      * driver will never pin pages outside of the working set of the guest
1666      * driver, which would thus not be discarding candidates).
1667      *
1668      * The first opportunity to induce pinning occurs here where we attempt to
1669      * attach the group to existing containers within the AddressSpace.  If any
1670      * pages are already zapped from the virtual address space, such as from
1671      * previous discards, new pinning will cause valid mappings to be
1672      * re-established.  Likewise, when the overall MemoryListener for a new
1673      * container is registered, a replay of mappings within the AddressSpace
1674      * will occur, re-establishing any previously zapped pages as well.
1675      *
1676      * Especially virtio-balloon is currently only prevented from discarding
1677      * new memory, it will not yet set ram_block_discard_set_required() and
1678      * therefore, neither stops us here or deals with the sudden memory
1679      * consumption of inflated memory.
1680      */
1681     ret = ram_block_discard_disable(true);
1682     if (ret) {
1683         error_setg_errno(errp, -ret, "Cannot set discarding of RAM broken");
1684         return ret;
1685     }
1686 
1687     QLIST_FOREACH(container, &space->containers, next) {
1688         if (!ioctl(group->fd, VFIO_GROUP_SET_CONTAINER, &container->fd)) {
1689             group->container = container;
1690             QLIST_INSERT_HEAD(&container->group_list, group, container_next);
1691             vfio_kvm_device_add_group(group);
1692             return 0;
1693         }
1694     }
1695 
1696     fd = qemu_open_old("/dev/vfio/vfio", O_RDWR);
1697     if (fd < 0) {
1698         error_setg_errno(errp, errno, "failed to open /dev/vfio/vfio");
1699         ret = -errno;
1700         goto put_space_exit;
1701     }
1702 
1703     ret = ioctl(fd, VFIO_GET_API_VERSION);
1704     if (ret != VFIO_API_VERSION) {
1705         error_setg(errp, "supported vfio version: %d, "
1706                    "reported version: %d", VFIO_API_VERSION, ret);
1707         ret = -EINVAL;
1708         goto close_fd_exit;
1709     }
1710 
1711     container = g_malloc0(sizeof(*container));
1712     container->space = space;
1713     container->fd = fd;
1714     container->error = NULL;
1715     container->dirty_pages_supported = false;
1716     QLIST_INIT(&container->giommu_list);
1717     QLIST_INIT(&container->hostwin_list);
1718 
1719     ret = vfio_init_container(container, group->fd, errp);
1720     if (ret) {
1721         goto free_container_exit;
1722     }
1723 
1724     switch (container->iommu_type) {
1725     case VFIO_TYPE1v2_IOMMU:
1726     case VFIO_TYPE1_IOMMU:
1727     {
1728         struct vfio_iommu_type1_info *info;
1729 
1730         /*
1731          * FIXME: This assumes that a Type1 IOMMU can map any 64-bit
1732          * IOVA whatsoever.  That's not actually true, but the current
1733          * kernel interface doesn't tell us what it can map, and the
1734          * existing Type1 IOMMUs generally support any IOVA we're
1735          * going to actually try in practice.
1736          */
1737         ret = vfio_get_iommu_info(container, &info);
1738 
1739         if (ret || !(info->flags & VFIO_IOMMU_INFO_PGSIZES)) {
1740             /* Assume 4k IOVA page size */
1741             info->iova_pgsizes = 4096;
1742         }
1743         vfio_host_win_add(container, 0, (hwaddr)-1, info->iova_pgsizes);
1744         container->pgsizes = info->iova_pgsizes;
1745 
1746         if (!ret) {
1747             vfio_get_iommu_info_migration(container, info);
1748         }
1749         g_free(info);
1750         break;
1751     }
1752     case VFIO_SPAPR_TCE_v2_IOMMU:
1753     case VFIO_SPAPR_TCE_IOMMU:
1754     {
1755         struct vfio_iommu_spapr_tce_info info;
1756         bool v2 = container->iommu_type == VFIO_SPAPR_TCE_v2_IOMMU;
1757 
1758         /*
1759          * The host kernel code implementing VFIO_IOMMU_DISABLE is called
1760          * when container fd is closed so we do not call it explicitly
1761          * in this file.
1762          */
1763         if (!v2) {
1764             ret = ioctl(fd, VFIO_IOMMU_ENABLE);
1765             if (ret) {
1766                 error_setg_errno(errp, errno, "failed to enable container");
1767                 ret = -errno;
1768                 goto free_container_exit;
1769             }
1770         } else {
1771             container->prereg_listener = vfio_prereg_listener;
1772 
1773             memory_listener_register(&container->prereg_listener,
1774                                      &address_space_memory);
1775             if (container->error) {
1776                 memory_listener_unregister(&container->prereg_listener);
1777                 ret = -1;
1778                 error_propagate_prepend(errp, container->error,
1779                     "RAM memory listener initialization failed: ");
1780                 goto free_container_exit;
1781             }
1782         }
1783 
1784         info.argsz = sizeof(info);
1785         ret = ioctl(fd, VFIO_IOMMU_SPAPR_TCE_GET_INFO, &info);
1786         if (ret) {
1787             error_setg_errno(errp, errno,
1788                              "VFIO_IOMMU_SPAPR_TCE_GET_INFO failed");
1789             ret = -errno;
1790             if (v2) {
1791                 memory_listener_unregister(&container->prereg_listener);
1792             }
1793             goto free_container_exit;
1794         }
1795 
1796         if (v2) {
1797             container->pgsizes = info.ddw.pgsizes;
1798             /*
1799              * There is a default window in just created container.
1800              * To make region_add/del simpler, we better remove this
1801              * window now and let those iommu_listener callbacks
1802              * create/remove them when needed.
1803              */
1804             ret = vfio_spapr_remove_window(container, info.dma32_window_start);
1805             if (ret) {
1806                 error_setg_errno(errp, -ret,
1807                                  "failed to remove existing window");
1808                 goto free_container_exit;
1809             }
1810         } else {
1811             /* The default table uses 4K pages */
1812             container->pgsizes = 0x1000;
1813             vfio_host_win_add(container, info.dma32_window_start,
1814                               info.dma32_window_start +
1815                               info.dma32_window_size - 1,
1816                               0x1000);
1817         }
1818     }
1819     }
1820 
1821     vfio_kvm_device_add_group(group);
1822 
1823     QLIST_INIT(&container->group_list);
1824     QLIST_INSERT_HEAD(&space->containers, container, next);
1825 
1826     group->container = container;
1827     QLIST_INSERT_HEAD(&container->group_list, group, container_next);
1828 
1829     container->listener = vfio_memory_listener;
1830 
1831     memory_listener_register(&container->listener, container->space->as);
1832 
1833     if (container->error) {
1834         ret = -1;
1835         error_propagate_prepend(errp, container->error,
1836             "memory listener initialization failed: ");
1837         goto listener_release_exit;
1838     }
1839 
1840     container->initialized = true;
1841 
1842     return 0;
1843 listener_release_exit:
1844     QLIST_REMOVE(group, container_next);
1845     QLIST_REMOVE(container, next);
1846     vfio_kvm_device_del_group(group);
1847     vfio_listener_release(container);
1848 
1849 free_container_exit:
1850     g_free(container);
1851 
1852 close_fd_exit:
1853     close(fd);
1854 
1855 put_space_exit:
1856     ram_block_discard_disable(false);
1857     vfio_put_address_space(space);
1858 
1859     return ret;
1860 }
1861 
1862 static void vfio_disconnect_container(VFIOGroup *group)
1863 {
1864     VFIOContainer *container = group->container;
1865 
1866     QLIST_REMOVE(group, container_next);
1867     group->container = NULL;
1868 
1869     /*
1870      * Explicitly release the listener first before unset container,
1871      * since unset may destroy the backend container if it's the last
1872      * group.
1873      */
1874     if (QLIST_EMPTY(&container->group_list)) {
1875         vfio_listener_release(container);
1876     }
1877 
1878     if (ioctl(group->fd, VFIO_GROUP_UNSET_CONTAINER, &container->fd)) {
1879         error_report("vfio: error disconnecting group %d from container",
1880                      group->groupid);
1881     }
1882 
1883     if (QLIST_EMPTY(&container->group_list)) {
1884         VFIOAddressSpace *space = container->space;
1885         VFIOGuestIOMMU *giommu, *tmp;
1886 
1887         QLIST_REMOVE(container, next);
1888 
1889         QLIST_FOREACH_SAFE(giommu, &container->giommu_list, giommu_next, tmp) {
1890             memory_region_unregister_iommu_notifier(
1891                     MEMORY_REGION(giommu->iommu), &giommu->n);
1892             QLIST_REMOVE(giommu, giommu_next);
1893             g_free(giommu);
1894         }
1895 
1896         trace_vfio_disconnect_container(container->fd);
1897         close(container->fd);
1898         g_free(container);
1899 
1900         vfio_put_address_space(space);
1901     }
1902 }
1903 
1904 VFIOGroup *vfio_get_group(int groupid, AddressSpace *as, Error **errp)
1905 {
1906     VFIOGroup *group;
1907     char path[32];
1908     struct vfio_group_status status = { .argsz = sizeof(status) };
1909 
1910     QLIST_FOREACH(group, &vfio_group_list, next) {
1911         if (group->groupid == groupid) {
1912             /* Found it.  Now is it already in the right context? */
1913             if (group->container->space->as == as) {
1914                 return group;
1915             } else {
1916                 error_setg(errp, "group %d used in multiple address spaces",
1917                            group->groupid);
1918                 return NULL;
1919             }
1920         }
1921     }
1922 
1923     group = g_malloc0(sizeof(*group));
1924 
1925     snprintf(path, sizeof(path), "/dev/vfio/%d", groupid);
1926     group->fd = qemu_open_old(path, O_RDWR);
1927     if (group->fd < 0) {
1928         error_setg_errno(errp, errno, "failed to open %s", path);
1929         goto free_group_exit;
1930     }
1931 
1932     if (ioctl(group->fd, VFIO_GROUP_GET_STATUS, &status)) {
1933         error_setg_errno(errp, errno, "failed to get group %d status", groupid);
1934         goto close_fd_exit;
1935     }
1936 
1937     if (!(status.flags & VFIO_GROUP_FLAGS_VIABLE)) {
1938         error_setg(errp, "group %d is not viable", groupid);
1939         error_append_hint(errp,
1940                           "Please ensure all devices within the iommu_group "
1941                           "are bound to their vfio bus driver.\n");
1942         goto close_fd_exit;
1943     }
1944 
1945     group->groupid = groupid;
1946     QLIST_INIT(&group->device_list);
1947 
1948     if (vfio_connect_container(group, as, errp)) {
1949         error_prepend(errp, "failed to setup container for group %d: ",
1950                       groupid);
1951         goto close_fd_exit;
1952     }
1953 
1954     if (QLIST_EMPTY(&vfio_group_list)) {
1955         qemu_register_reset(vfio_reset_handler, NULL);
1956     }
1957 
1958     QLIST_INSERT_HEAD(&vfio_group_list, group, next);
1959 
1960     return group;
1961 
1962 close_fd_exit:
1963     close(group->fd);
1964 
1965 free_group_exit:
1966     g_free(group);
1967 
1968     return NULL;
1969 }
1970 
1971 void vfio_put_group(VFIOGroup *group)
1972 {
1973     if (!group || !QLIST_EMPTY(&group->device_list)) {
1974         return;
1975     }
1976 
1977     if (!group->ram_block_discard_allowed) {
1978         ram_block_discard_disable(false);
1979     }
1980     vfio_kvm_device_del_group(group);
1981     vfio_disconnect_container(group);
1982     QLIST_REMOVE(group, next);
1983     trace_vfio_put_group(group->fd);
1984     close(group->fd);
1985     g_free(group);
1986 
1987     if (QLIST_EMPTY(&vfio_group_list)) {
1988         qemu_unregister_reset(vfio_reset_handler, NULL);
1989     }
1990 }
1991 
1992 int vfio_get_device(VFIOGroup *group, const char *name,
1993                     VFIODevice *vbasedev, Error **errp)
1994 {
1995     struct vfio_device_info dev_info = { .argsz = sizeof(dev_info) };
1996     int ret, fd;
1997 
1998     fd = ioctl(group->fd, VFIO_GROUP_GET_DEVICE_FD, name);
1999     if (fd < 0) {
2000         error_setg_errno(errp, errno, "error getting device from group %d",
2001                          group->groupid);
2002         error_append_hint(errp,
2003                       "Verify all devices in group %d are bound to vfio-<bus> "
2004                       "or pci-stub and not already in use\n", group->groupid);
2005         return fd;
2006     }
2007 
2008     ret = ioctl(fd, VFIO_DEVICE_GET_INFO, &dev_info);
2009     if (ret) {
2010         error_setg_errno(errp, errno, "error getting device info");
2011         close(fd);
2012         return ret;
2013     }
2014 
2015     /*
2016      * Set discarding of RAM as not broken for this group if the driver knows
2017      * the device operates compatibly with discarding.  Setting must be
2018      * consistent per group, but since compatibility is really only possible
2019      * with mdev currently, we expect singleton groups.
2020      */
2021     if (vbasedev->ram_block_discard_allowed !=
2022         group->ram_block_discard_allowed) {
2023         if (!QLIST_EMPTY(&group->device_list)) {
2024             error_setg(errp, "Inconsistent setting of support for discarding "
2025                        "RAM (e.g., balloon) within group");
2026             close(fd);
2027             return -1;
2028         }
2029 
2030         if (!group->ram_block_discard_allowed) {
2031             group->ram_block_discard_allowed = true;
2032             ram_block_discard_disable(false);
2033         }
2034     }
2035 
2036     vbasedev->fd = fd;
2037     vbasedev->group = group;
2038     QLIST_INSERT_HEAD(&group->device_list, vbasedev, next);
2039 
2040     vbasedev->num_irqs = dev_info.num_irqs;
2041     vbasedev->num_regions = dev_info.num_regions;
2042     vbasedev->flags = dev_info.flags;
2043 
2044     trace_vfio_get_device(name, dev_info.flags, dev_info.num_regions,
2045                           dev_info.num_irqs);
2046 
2047     vbasedev->reset_works = !!(dev_info.flags & VFIO_DEVICE_FLAGS_RESET);
2048     return 0;
2049 }
2050 
2051 void vfio_put_base_device(VFIODevice *vbasedev)
2052 {
2053     if (!vbasedev->group) {
2054         return;
2055     }
2056     QLIST_REMOVE(vbasedev, next);
2057     vbasedev->group = NULL;
2058     trace_vfio_put_base_device(vbasedev->fd);
2059     close(vbasedev->fd);
2060 }
2061 
2062 int vfio_get_region_info(VFIODevice *vbasedev, int index,
2063                          struct vfio_region_info **info)
2064 {
2065     size_t argsz = sizeof(struct vfio_region_info);
2066 
2067     *info = g_malloc0(argsz);
2068 
2069     (*info)->index = index;
2070 retry:
2071     (*info)->argsz = argsz;
2072 
2073     if (ioctl(vbasedev->fd, VFIO_DEVICE_GET_REGION_INFO, *info)) {
2074         g_free(*info);
2075         *info = NULL;
2076         return -errno;
2077     }
2078 
2079     if ((*info)->argsz > argsz) {
2080         argsz = (*info)->argsz;
2081         *info = g_realloc(*info, argsz);
2082 
2083         goto retry;
2084     }
2085 
2086     return 0;
2087 }
2088 
2089 int vfio_get_dev_region_info(VFIODevice *vbasedev, uint32_t type,
2090                              uint32_t subtype, struct vfio_region_info **info)
2091 {
2092     int i;
2093 
2094     for (i = 0; i < vbasedev->num_regions; i++) {
2095         struct vfio_info_cap_header *hdr;
2096         struct vfio_region_info_cap_type *cap_type;
2097 
2098         if (vfio_get_region_info(vbasedev, i, info)) {
2099             continue;
2100         }
2101 
2102         hdr = vfio_get_region_info_cap(*info, VFIO_REGION_INFO_CAP_TYPE);
2103         if (!hdr) {
2104             g_free(*info);
2105             continue;
2106         }
2107 
2108         cap_type = container_of(hdr, struct vfio_region_info_cap_type, header);
2109 
2110         trace_vfio_get_dev_region(vbasedev->name, i,
2111                                   cap_type->type, cap_type->subtype);
2112 
2113         if (cap_type->type == type && cap_type->subtype == subtype) {
2114             return 0;
2115         }
2116 
2117         g_free(*info);
2118     }
2119 
2120     *info = NULL;
2121     return -ENODEV;
2122 }
2123 
2124 bool vfio_has_region_cap(VFIODevice *vbasedev, int region, uint16_t cap_type)
2125 {
2126     struct vfio_region_info *info = NULL;
2127     bool ret = false;
2128 
2129     if (!vfio_get_region_info(vbasedev, region, &info)) {
2130         if (vfio_get_region_info_cap(info, cap_type)) {
2131             ret = true;
2132         }
2133         g_free(info);
2134     }
2135 
2136     return ret;
2137 }
2138 
2139 /*
2140  * Interfaces for IBM EEH (Enhanced Error Handling)
2141  */
2142 static bool vfio_eeh_container_ok(VFIOContainer *container)
2143 {
2144     /*
2145      * As of 2016-03-04 (linux-4.5) the host kernel EEH/VFIO
2146      * implementation is broken if there are multiple groups in a
2147      * container.  The hardware works in units of Partitionable
2148      * Endpoints (== IOMMU groups) and the EEH operations naively
2149      * iterate across all groups in the container, without any logic
2150      * to make sure the groups have their state synchronized.  For
2151      * certain operations (ENABLE) that might be ok, until an error
2152      * occurs, but for others (GET_STATE) it's clearly broken.
2153      */
2154 
2155     /*
2156      * XXX Once fixed kernels exist, test for them here
2157      */
2158 
2159     if (QLIST_EMPTY(&container->group_list)) {
2160         return false;
2161     }
2162 
2163     if (QLIST_NEXT(QLIST_FIRST(&container->group_list), container_next)) {
2164         return false;
2165     }
2166 
2167     return true;
2168 }
2169 
2170 static int vfio_eeh_container_op(VFIOContainer *container, uint32_t op)
2171 {
2172     struct vfio_eeh_pe_op pe_op = {
2173         .argsz = sizeof(pe_op),
2174         .op = op,
2175     };
2176     int ret;
2177 
2178     if (!vfio_eeh_container_ok(container)) {
2179         error_report("vfio/eeh: EEH_PE_OP 0x%x: "
2180                      "kernel requires a container with exactly one group", op);
2181         return -EPERM;
2182     }
2183 
2184     ret = ioctl(container->fd, VFIO_EEH_PE_OP, &pe_op);
2185     if (ret < 0) {
2186         error_report("vfio/eeh: EEH_PE_OP 0x%x failed: %m", op);
2187         return -errno;
2188     }
2189 
2190     return ret;
2191 }
2192 
2193 static VFIOContainer *vfio_eeh_as_container(AddressSpace *as)
2194 {
2195     VFIOAddressSpace *space = vfio_get_address_space(as);
2196     VFIOContainer *container = NULL;
2197 
2198     if (QLIST_EMPTY(&space->containers)) {
2199         /* No containers to act on */
2200         goto out;
2201     }
2202 
2203     container = QLIST_FIRST(&space->containers);
2204 
2205     if (QLIST_NEXT(container, next)) {
2206         /* We don't yet have logic to synchronize EEH state across
2207          * multiple containers */
2208         container = NULL;
2209         goto out;
2210     }
2211 
2212 out:
2213     vfio_put_address_space(space);
2214     return container;
2215 }
2216 
2217 bool vfio_eeh_as_ok(AddressSpace *as)
2218 {
2219     VFIOContainer *container = vfio_eeh_as_container(as);
2220 
2221     return (container != NULL) && vfio_eeh_container_ok(container);
2222 }
2223 
2224 int vfio_eeh_as_op(AddressSpace *as, uint32_t op)
2225 {
2226     VFIOContainer *container = vfio_eeh_as_container(as);
2227 
2228     if (!container) {
2229         return -ENODEV;
2230     }
2231     return vfio_eeh_container_op(container, op);
2232 }
2233