xref: /openbmc/qemu/hw/vfio/common.c (revision 37094b6d)
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 "sysemu/runstate.h"
40 #include "trace.h"
41 #include "qapi/error.h"
42 #include "migration/migration.h"
43 
44 VFIOGroupList vfio_group_list =
45     QLIST_HEAD_INITIALIZER(vfio_group_list);
46 static QLIST_HEAD(, VFIOAddressSpace) vfio_address_spaces =
47     QLIST_HEAD_INITIALIZER(vfio_address_spaces);
48 
49 #ifdef CONFIG_KVM
50 /*
51  * We have a single VFIO pseudo device per KVM VM.  Once created it lives
52  * for the life of the VM.  Closing the file descriptor only drops our
53  * reference to it and the device's reference to kvm.  Therefore once
54  * initialized, this file descriptor is only released on QEMU exit and
55  * we'll re-use it should another vfio device be attached before then.
56  */
57 static int vfio_kvm_device_fd = -1;
58 #endif
59 
60 /*
61  * Common VFIO interrupt disable
62  */
63 void vfio_disable_irqindex(VFIODevice *vbasedev, int index)
64 {
65     struct vfio_irq_set irq_set = {
66         .argsz = sizeof(irq_set),
67         .flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_TRIGGER,
68         .index = index,
69         .start = 0,
70         .count = 0,
71     };
72 
73     ioctl(vbasedev->fd, VFIO_DEVICE_SET_IRQS, &irq_set);
74 }
75 
76 void vfio_unmask_single_irqindex(VFIODevice *vbasedev, int index)
77 {
78     struct vfio_irq_set irq_set = {
79         .argsz = sizeof(irq_set),
80         .flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_UNMASK,
81         .index = index,
82         .start = 0,
83         .count = 1,
84     };
85 
86     ioctl(vbasedev->fd, VFIO_DEVICE_SET_IRQS, &irq_set);
87 }
88 
89 void vfio_mask_single_irqindex(VFIODevice *vbasedev, int index)
90 {
91     struct vfio_irq_set irq_set = {
92         .argsz = sizeof(irq_set),
93         .flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_MASK,
94         .index = index,
95         .start = 0,
96         .count = 1,
97     };
98 
99     ioctl(vbasedev->fd, VFIO_DEVICE_SET_IRQS, &irq_set);
100 }
101 
102 static inline const char *action_to_str(int action)
103 {
104     switch (action) {
105     case VFIO_IRQ_SET_ACTION_MASK:
106         return "MASK";
107     case VFIO_IRQ_SET_ACTION_UNMASK:
108         return "UNMASK";
109     case VFIO_IRQ_SET_ACTION_TRIGGER:
110         return "TRIGGER";
111     default:
112         return "UNKNOWN ACTION";
113     }
114 }
115 
116 static const char *index_to_str(VFIODevice *vbasedev, int index)
117 {
118     if (vbasedev->type != VFIO_DEVICE_TYPE_PCI) {
119         return NULL;
120     }
121 
122     switch (index) {
123     case VFIO_PCI_INTX_IRQ_INDEX:
124         return "INTX";
125     case VFIO_PCI_MSI_IRQ_INDEX:
126         return "MSI";
127     case VFIO_PCI_MSIX_IRQ_INDEX:
128         return "MSIX";
129     case VFIO_PCI_ERR_IRQ_INDEX:
130         return "ERR";
131     case VFIO_PCI_REQ_IRQ_INDEX:
132         return "REQ";
133     default:
134         return NULL;
135     }
136 }
137 
138 static int vfio_ram_block_discard_disable(VFIOContainer *container, bool state)
139 {
140     switch (container->iommu_type) {
141     case VFIO_TYPE1v2_IOMMU:
142     case VFIO_TYPE1_IOMMU:
143         /*
144          * We support coordinated discarding of RAM via the RamDiscardManager.
145          */
146         return ram_block_uncoordinated_discard_disable(state);
147     default:
148         /*
149          * VFIO_SPAPR_TCE_IOMMU most probably works just fine with
150          * RamDiscardManager, however, it is completely untested.
151          *
152          * VFIO_SPAPR_TCE_v2_IOMMU with "DMA memory preregistering" does
153          * completely the opposite of managing mapping/pinning dynamically as
154          * required by RamDiscardManager. We would have to special-case sections
155          * with a RamDiscardManager.
156          */
157         return ram_block_discard_disable(state);
158     }
159 }
160 
161 int vfio_set_irq_signaling(VFIODevice *vbasedev, int index, int subindex,
162                            int action, int fd, Error **errp)
163 {
164     struct vfio_irq_set *irq_set;
165     int argsz, ret = 0;
166     const char *name;
167     int32_t *pfd;
168 
169     argsz = sizeof(*irq_set) + sizeof(*pfd);
170 
171     irq_set = g_malloc0(argsz);
172     irq_set->argsz = argsz;
173     irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD | action;
174     irq_set->index = index;
175     irq_set->start = subindex;
176     irq_set->count = 1;
177     pfd = (int32_t *)&irq_set->data;
178     *pfd = fd;
179 
180     if (ioctl(vbasedev->fd, VFIO_DEVICE_SET_IRQS, irq_set)) {
181         ret = -errno;
182     }
183     g_free(irq_set);
184 
185     if (!ret) {
186         return 0;
187     }
188 
189     error_setg_errno(errp, -ret, "VFIO_DEVICE_SET_IRQS failure");
190 
191     name = index_to_str(vbasedev, index);
192     if (name) {
193         error_prepend(errp, "%s-%d: ", name, subindex);
194     } else {
195         error_prepend(errp, "index %d-%d: ", index, subindex);
196     }
197     error_prepend(errp,
198                   "Failed to %s %s eventfd signaling for interrupt ",
199                   fd < 0 ? "tear down" : "set up", action_to_str(action));
200     return ret;
201 }
202 
203 /*
204  * IO Port/MMIO - Beware of the endians, VFIO is always little endian
205  */
206 void vfio_region_write(void *opaque, hwaddr addr,
207                        uint64_t data, unsigned size)
208 {
209     VFIORegion *region = opaque;
210     VFIODevice *vbasedev = region->vbasedev;
211     union {
212         uint8_t byte;
213         uint16_t word;
214         uint32_t dword;
215         uint64_t qword;
216     } buf;
217 
218     switch (size) {
219     case 1:
220         buf.byte = data;
221         break;
222     case 2:
223         buf.word = cpu_to_le16(data);
224         break;
225     case 4:
226         buf.dword = cpu_to_le32(data);
227         break;
228     case 8:
229         buf.qword = cpu_to_le64(data);
230         break;
231     default:
232         hw_error("vfio: unsupported write size, %u bytes", size);
233         break;
234     }
235 
236     if (pwrite(vbasedev->fd, &buf, size, region->fd_offset + addr) != size) {
237         error_report("%s(%s:region%d+0x%"HWADDR_PRIx", 0x%"PRIx64
238                      ",%d) failed: %m",
239                      __func__, vbasedev->name, region->nr,
240                      addr, data, size);
241     }
242 
243     trace_vfio_region_write(vbasedev->name, region->nr, addr, data, size);
244 
245     /*
246      * A read or write to a BAR always signals an INTx EOI.  This will
247      * do nothing if not pending (including not in INTx mode).  We assume
248      * that a BAR access is in response to an interrupt and that BAR
249      * accesses will service the interrupt.  Unfortunately, we don't know
250      * which access will service the interrupt, so we're potentially
251      * getting quite a few host interrupts per guest interrupt.
252      */
253     vbasedev->ops->vfio_eoi(vbasedev);
254 }
255 
256 uint64_t vfio_region_read(void *opaque,
257                           hwaddr addr, unsigned size)
258 {
259     VFIORegion *region = opaque;
260     VFIODevice *vbasedev = region->vbasedev;
261     union {
262         uint8_t byte;
263         uint16_t word;
264         uint32_t dword;
265         uint64_t qword;
266     } buf;
267     uint64_t data = 0;
268 
269     if (pread(vbasedev->fd, &buf, size, region->fd_offset + addr) != size) {
270         error_report("%s(%s:region%d+0x%"HWADDR_PRIx", %d) failed: %m",
271                      __func__, vbasedev->name, region->nr,
272                      addr, size);
273         return (uint64_t)-1;
274     }
275     switch (size) {
276     case 1:
277         data = buf.byte;
278         break;
279     case 2:
280         data = le16_to_cpu(buf.word);
281         break;
282     case 4:
283         data = le32_to_cpu(buf.dword);
284         break;
285     case 8:
286         data = le64_to_cpu(buf.qword);
287         break;
288     default:
289         hw_error("vfio: unsupported read size, %u bytes", size);
290         break;
291     }
292 
293     trace_vfio_region_read(vbasedev->name, region->nr, addr, size, data);
294 
295     /* Same as write above */
296     vbasedev->ops->vfio_eoi(vbasedev);
297 
298     return data;
299 }
300 
301 const MemoryRegionOps vfio_region_ops = {
302     .read = vfio_region_read,
303     .write = vfio_region_write,
304     .endianness = DEVICE_LITTLE_ENDIAN,
305     .valid = {
306         .min_access_size = 1,
307         .max_access_size = 8,
308     },
309     .impl = {
310         .min_access_size = 1,
311         .max_access_size = 8,
312     },
313 };
314 
315 /*
316  * Device state interfaces
317  */
318 
319 bool vfio_mig_active(void)
320 {
321     VFIOGroup *group;
322     VFIODevice *vbasedev;
323 
324     if (QLIST_EMPTY(&vfio_group_list)) {
325         return false;
326     }
327 
328     QLIST_FOREACH(group, &vfio_group_list, next) {
329         QLIST_FOREACH(vbasedev, &group->device_list, next) {
330             if (vbasedev->migration_blocker) {
331                 return false;
332             }
333         }
334     }
335     return true;
336 }
337 
338 static bool vfio_devices_all_dirty_tracking(VFIOContainer *container)
339 {
340     VFIOGroup *group;
341     VFIODevice *vbasedev;
342     MigrationState *ms = migrate_get_current();
343 
344     if (!migration_is_setup_or_active(ms->state)) {
345         return false;
346     }
347 
348     QLIST_FOREACH(group, &container->group_list, container_next) {
349         QLIST_FOREACH(vbasedev, &group->device_list, next) {
350             VFIOMigration *migration = vbasedev->migration;
351 
352             if (!migration) {
353                 return false;
354             }
355 
356             if ((vbasedev->pre_copy_dirty_page_tracking == ON_OFF_AUTO_OFF)
357                 && (migration->device_state & VFIO_DEVICE_STATE_RUNNING)) {
358                 return false;
359             }
360         }
361     }
362     return true;
363 }
364 
365 static bool vfio_devices_all_running_and_saving(VFIOContainer *container)
366 {
367     VFIOGroup *group;
368     VFIODevice *vbasedev;
369     MigrationState *ms = migrate_get_current();
370 
371     if (!migration_is_setup_or_active(ms->state)) {
372         return false;
373     }
374 
375     QLIST_FOREACH(group, &container->group_list, container_next) {
376         QLIST_FOREACH(vbasedev, &group->device_list, next) {
377             VFIOMigration *migration = vbasedev->migration;
378 
379             if (!migration) {
380                 return false;
381             }
382 
383             if ((migration->device_state & VFIO_DEVICE_STATE_SAVING) &&
384                 (migration->device_state & VFIO_DEVICE_STATE_RUNNING)) {
385                 continue;
386             } else {
387                 return false;
388             }
389         }
390     }
391     return true;
392 }
393 
394 static int vfio_dma_unmap_bitmap(VFIOContainer *container,
395                                  hwaddr iova, ram_addr_t size,
396                                  IOMMUTLBEntry *iotlb)
397 {
398     struct vfio_iommu_type1_dma_unmap *unmap;
399     struct vfio_bitmap *bitmap;
400     uint64_t pages = REAL_HOST_PAGE_ALIGN(size) / qemu_real_host_page_size();
401     int ret;
402 
403     unmap = g_malloc0(sizeof(*unmap) + sizeof(*bitmap));
404 
405     unmap->argsz = sizeof(*unmap) + sizeof(*bitmap);
406     unmap->iova = iova;
407     unmap->size = size;
408     unmap->flags |= VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP;
409     bitmap = (struct vfio_bitmap *)&unmap->data;
410 
411     /*
412      * cpu_physical_memory_set_dirty_lebitmap() supports pages in bitmap of
413      * qemu_real_host_page_size to mark those dirty. Hence set bitmap_pgsize
414      * to qemu_real_host_page_size.
415      */
416 
417     bitmap->pgsize = qemu_real_host_page_size();
418     bitmap->size = ROUND_UP(pages, sizeof(__u64) * BITS_PER_BYTE) /
419                    BITS_PER_BYTE;
420 
421     if (bitmap->size > container->max_dirty_bitmap_size) {
422         error_report("UNMAP: Size of bitmap too big 0x%"PRIx64,
423                      (uint64_t)bitmap->size);
424         ret = -E2BIG;
425         goto unmap_exit;
426     }
427 
428     bitmap->data = g_try_malloc0(bitmap->size);
429     if (!bitmap->data) {
430         ret = -ENOMEM;
431         goto unmap_exit;
432     }
433 
434     ret = ioctl(container->fd, VFIO_IOMMU_UNMAP_DMA, unmap);
435     if (!ret) {
436         cpu_physical_memory_set_dirty_lebitmap((unsigned long *)bitmap->data,
437                 iotlb->translated_addr, pages);
438     } else {
439         error_report("VFIO_UNMAP_DMA with DIRTY_BITMAP : %m");
440     }
441 
442     g_free(bitmap->data);
443 unmap_exit:
444     g_free(unmap);
445     return ret;
446 }
447 
448 /*
449  * DMA - Mapping and unmapping for the "type1" IOMMU interface used on x86
450  */
451 static int vfio_dma_unmap(VFIOContainer *container,
452                           hwaddr iova, ram_addr_t size,
453                           IOMMUTLBEntry *iotlb)
454 {
455     struct vfio_iommu_type1_dma_unmap unmap = {
456         .argsz = sizeof(unmap),
457         .flags = 0,
458         .iova = iova,
459         .size = size,
460     };
461 
462     if (iotlb && container->dirty_pages_supported &&
463         vfio_devices_all_running_and_saving(container)) {
464         return vfio_dma_unmap_bitmap(container, iova, size, iotlb);
465     }
466 
467     while (ioctl(container->fd, VFIO_IOMMU_UNMAP_DMA, &unmap)) {
468         /*
469          * The type1 backend has an off-by-one bug in the kernel (71a7d3d78e3c
470          * v4.15) where an overflow in its wrap-around check prevents us from
471          * unmapping the last page of the address space.  Test for the error
472          * condition and re-try the unmap excluding the last page.  The
473          * expectation is that we've never mapped the last page anyway and this
474          * unmap request comes via vIOMMU support which also makes it unlikely
475          * that this page is used.  This bug was introduced well after type1 v2
476          * support was introduced, so we shouldn't need to test for v1.  A fix
477          * is queued for kernel v5.0 so this workaround can be removed once
478          * affected kernels are sufficiently deprecated.
479          */
480         if (errno == EINVAL && unmap.size && !(unmap.iova + unmap.size) &&
481             container->iommu_type == VFIO_TYPE1v2_IOMMU) {
482             trace_vfio_dma_unmap_overflow_workaround();
483             unmap.size -= 1ULL << ctz64(container->pgsizes);
484             continue;
485         }
486         error_report("VFIO_UNMAP_DMA failed: %s", strerror(errno));
487         return -errno;
488     }
489 
490     return 0;
491 }
492 
493 static int vfio_dma_map(VFIOContainer *container, hwaddr iova,
494                         ram_addr_t size, void *vaddr, bool readonly)
495 {
496     struct vfio_iommu_type1_dma_map map = {
497         .argsz = sizeof(map),
498         .flags = VFIO_DMA_MAP_FLAG_READ,
499         .vaddr = (__u64)(uintptr_t)vaddr,
500         .iova = iova,
501         .size = size,
502     };
503 
504     if (!readonly) {
505         map.flags |= VFIO_DMA_MAP_FLAG_WRITE;
506     }
507 
508     /*
509      * Try the mapping, if it fails with EBUSY, unmap the region and try
510      * again.  This shouldn't be necessary, but we sometimes see it in
511      * the VGA ROM space.
512      */
513     if (ioctl(container->fd, VFIO_IOMMU_MAP_DMA, &map) == 0 ||
514         (errno == EBUSY && vfio_dma_unmap(container, iova, size, NULL) == 0 &&
515          ioctl(container->fd, VFIO_IOMMU_MAP_DMA, &map) == 0)) {
516         return 0;
517     }
518 
519     error_report("VFIO_MAP_DMA failed: %s", strerror(errno));
520     return -errno;
521 }
522 
523 static void vfio_host_win_add(VFIOContainer *container,
524                               hwaddr min_iova, hwaddr max_iova,
525                               uint64_t iova_pgsizes)
526 {
527     VFIOHostDMAWindow *hostwin;
528 
529     QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) {
530         if (ranges_overlap(hostwin->min_iova,
531                            hostwin->max_iova - hostwin->min_iova + 1,
532                            min_iova,
533                            max_iova - min_iova + 1)) {
534             hw_error("%s: Overlapped IOMMU are not enabled", __func__);
535         }
536     }
537 
538     hostwin = g_malloc0(sizeof(*hostwin));
539 
540     hostwin->min_iova = min_iova;
541     hostwin->max_iova = max_iova;
542     hostwin->iova_pgsizes = iova_pgsizes;
543     QLIST_INSERT_HEAD(&container->hostwin_list, hostwin, hostwin_next);
544 }
545 
546 static int vfio_host_win_del(VFIOContainer *container, hwaddr min_iova,
547                              hwaddr max_iova)
548 {
549     VFIOHostDMAWindow *hostwin;
550 
551     QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) {
552         if (hostwin->min_iova == min_iova && hostwin->max_iova == max_iova) {
553             QLIST_REMOVE(hostwin, hostwin_next);
554             g_free(hostwin);
555             return 0;
556         }
557     }
558 
559     return -1;
560 }
561 
562 static bool vfio_listener_skipped_section(MemoryRegionSection *section)
563 {
564     return (!memory_region_is_ram(section->mr) &&
565             !memory_region_is_iommu(section->mr)) ||
566            memory_region_is_protected(section->mr) ||
567            /*
568             * Sizing an enabled 64-bit BAR can cause spurious mappings to
569             * addresses in the upper part of the 64-bit address space.  These
570             * are never accessed by the CPU and beyond the address width of
571             * some IOMMU hardware.  TODO: VFIO should tell us the IOMMU width.
572             */
573            section->offset_within_address_space & (1ULL << 63);
574 }
575 
576 /* Called with rcu_read_lock held.  */
577 static bool vfio_get_xlat_addr(IOMMUTLBEntry *iotlb, void **vaddr,
578                                ram_addr_t *ram_addr, bool *read_only)
579 {
580     MemoryRegion *mr;
581     hwaddr xlat;
582     hwaddr len = iotlb->addr_mask + 1;
583     bool writable = iotlb->perm & IOMMU_WO;
584 
585     /*
586      * The IOMMU TLB entry we have just covers translation through
587      * this IOMMU to its immediate target.  We need to translate
588      * it the rest of the way through to memory.
589      */
590     mr = address_space_translate(&address_space_memory,
591                                  iotlb->translated_addr,
592                                  &xlat, &len, writable,
593                                  MEMTXATTRS_UNSPECIFIED);
594     if (!memory_region_is_ram(mr)) {
595         error_report("iommu map to non memory area %"HWADDR_PRIx"",
596                      xlat);
597         return false;
598     } else if (memory_region_has_ram_discard_manager(mr)) {
599         RamDiscardManager *rdm = memory_region_get_ram_discard_manager(mr);
600         MemoryRegionSection tmp = {
601             .mr = mr,
602             .offset_within_region = xlat,
603             .size = int128_make64(len),
604         };
605 
606         /*
607          * Malicious VMs can map memory into the IOMMU, which is expected
608          * to remain discarded. vfio will pin all pages, populating memory.
609          * Disallow that. vmstate priorities make sure any RamDiscardManager
610          * were already restored before IOMMUs are restored.
611          */
612         if (!ram_discard_manager_is_populated(rdm, &tmp)) {
613             error_report("iommu map to discarded memory (e.g., unplugged via"
614                          " virtio-mem): %"HWADDR_PRIx"",
615                          iotlb->translated_addr);
616             return false;
617         }
618 
619         /*
620          * Malicious VMs might trigger discarding of IOMMU-mapped memory. The
621          * pages will remain pinned inside vfio until unmapped, resulting in a
622          * higher memory consumption than expected. If memory would get
623          * populated again later, there would be an inconsistency between pages
624          * pinned by vfio and pages seen by QEMU. This is the case until
625          * unmapped from the IOMMU (e.g., during device reset).
626          *
627          * With malicious guests, we really only care about pinning more memory
628          * than expected. RLIMIT_MEMLOCK set for the user/process can never be
629          * exceeded and can be used to mitigate this problem.
630          */
631         warn_report_once("Using vfio with vIOMMUs and coordinated discarding of"
632                          " RAM (e.g., virtio-mem) works, however, malicious"
633                          " guests can trigger pinning of more memory than"
634                          " intended via an IOMMU. It's possible to mitigate "
635                          " by setting/adjusting RLIMIT_MEMLOCK.");
636     }
637 
638     /*
639      * Translation truncates length to the IOMMU page size,
640      * check that it did not truncate too much.
641      */
642     if (len & iotlb->addr_mask) {
643         error_report("iommu has granularity incompatible with target AS");
644         return false;
645     }
646 
647     if (vaddr) {
648         *vaddr = memory_region_get_ram_ptr(mr) + xlat;
649     }
650 
651     if (ram_addr) {
652         *ram_addr = memory_region_get_ram_addr(mr) + xlat;
653     }
654 
655     if (read_only) {
656         *read_only = !writable || mr->readonly;
657     }
658 
659     return true;
660 }
661 
662 static void vfio_iommu_map_notify(IOMMUNotifier *n, IOMMUTLBEntry *iotlb)
663 {
664     VFIOGuestIOMMU *giommu = container_of(n, VFIOGuestIOMMU, n);
665     VFIOContainer *container = giommu->container;
666     hwaddr iova = iotlb->iova + giommu->iommu_offset;
667     void *vaddr;
668     int ret;
669 
670     trace_vfio_iommu_map_notify(iotlb->perm == IOMMU_NONE ? "UNMAP" : "MAP",
671                                 iova, iova + iotlb->addr_mask);
672 
673     if (iotlb->target_as != &address_space_memory) {
674         error_report("Wrong target AS \"%s\", only system memory is allowed",
675                      iotlb->target_as->name ? iotlb->target_as->name : "none");
676         return;
677     }
678 
679     rcu_read_lock();
680 
681     if ((iotlb->perm & IOMMU_RW) != IOMMU_NONE) {
682         bool read_only;
683 
684         if (!vfio_get_xlat_addr(iotlb, &vaddr, NULL, &read_only)) {
685             goto out;
686         }
687         /*
688          * vaddr is only valid until rcu_read_unlock(). But after
689          * vfio_dma_map has set up the mapping the pages will be
690          * pinned by the kernel. This makes sure that the RAM backend
691          * of vaddr will always be there, even if the memory object is
692          * destroyed and its backing memory munmap-ed.
693          */
694         ret = vfio_dma_map(container, iova,
695                            iotlb->addr_mask + 1, vaddr,
696                            read_only);
697         if (ret) {
698             error_report("vfio_dma_map(%p, 0x%"HWADDR_PRIx", "
699                          "0x%"HWADDR_PRIx", %p) = %d (%m)",
700                          container, iova,
701                          iotlb->addr_mask + 1, vaddr, ret);
702         }
703     } else {
704         ret = vfio_dma_unmap(container, iova, iotlb->addr_mask + 1, iotlb);
705         if (ret) {
706             error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx", "
707                          "0x%"HWADDR_PRIx") = %d (%m)",
708                          container, iova,
709                          iotlb->addr_mask + 1, ret);
710         }
711     }
712 out:
713     rcu_read_unlock();
714 }
715 
716 static void vfio_ram_discard_notify_discard(RamDiscardListener *rdl,
717                                             MemoryRegionSection *section)
718 {
719     VFIORamDiscardListener *vrdl = container_of(rdl, VFIORamDiscardListener,
720                                                 listener);
721     const hwaddr size = int128_get64(section->size);
722     const hwaddr iova = section->offset_within_address_space;
723     int ret;
724 
725     /* Unmap with a single call. */
726     ret = vfio_dma_unmap(vrdl->container, iova, size , NULL);
727     if (ret) {
728         error_report("%s: vfio_dma_unmap() failed: %s", __func__,
729                      strerror(-ret));
730     }
731 }
732 
733 static int vfio_ram_discard_notify_populate(RamDiscardListener *rdl,
734                                             MemoryRegionSection *section)
735 {
736     VFIORamDiscardListener *vrdl = container_of(rdl, VFIORamDiscardListener,
737                                                 listener);
738     const hwaddr end = section->offset_within_region +
739                        int128_get64(section->size);
740     hwaddr start, next, iova;
741     void *vaddr;
742     int ret;
743 
744     /*
745      * Map in (aligned within memory region) minimum granularity, so we can
746      * unmap in minimum granularity later.
747      */
748     for (start = section->offset_within_region; start < end; start = next) {
749         next = ROUND_UP(start + 1, vrdl->granularity);
750         next = MIN(next, end);
751 
752         iova = start - section->offset_within_region +
753                section->offset_within_address_space;
754         vaddr = memory_region_get_ram_ptr(section->mr) + start;
755 
756         ret = vfio_dma_map(vrdl->container, iova, next - start,
757                            vaddr, section->readonly);
758         if (ret) {
759             /* Rollback */
760             vfio_ram_discard_notify_discard(rdl, section);
761             return ret;
762         }
763     }
764     return 0;
765 }
766 
767 static void vfio_register_ram_discard_listener(VFIOContainer *container,
768                                                MemoryRegionSection *section)
769 {
770     RamDiscardManager *rdm = memory_region_get_ram_discard_manager(section->mr);
771     VFIORamDiscardListener *vrdl;
772 
773     /* Ignore some corner cases not relevant in practice. */
774     g_assert(QEMU_IS_ALIGNED(section->offset_within_region, TARGET_PAGE_SIZE));
775     g_assert(QEMU_IS_ALIGNED(section->offset_within_address_space,
776                              TARGET_PAGE_SIZE));
777     g_assert(QEMU_IS_ALIGNED(int128_get64(section->size), TARGET_PAGE_SIZE));
778 
779     vrdl = g_new0(VFIORamDiscardListener, 1);
780     vrdl->container = container;
781     vrdl->mr = section->mr;
782     vrdl->offset_within_address_space = section->offset_within_address_space;
783     vrdl->size = int128_get64(section->size);
784     vrdl->granularity = ram_discard_manager_get_min_granularity(rdm,
785                                                                 section->mr);
786 
787     g_assert(vrdl->granularity && is_power_of_2(vrdl->granularity));
788     g_assert(container->pgsizes &&
789              vrdl->granularity >= 1ULL << ctz64(container->pgsizes));
790 
791     ram_discard_listener_init(&vrdl->listener,
792                               vfio_ram_discard_notify_populate,
793                               vfio_ram_discard_notify_discard, true);
794     ram_discard_manager_register_listener(rdm, &vrdl->listener, section);
795     QLIST_INSERT_HEAD(&container->vrdl_list, vrdl, next);
796 
797     /*
798      * Sanity-check if we have a theoretically problematic setup where we could
799      * exceed the maximum number of possible DMA mappings over time. We assume
800      * that each mapped section in the same address space as a RamDiscardManager
801      * section consumes exactly one DMA mapping, with the exception of
802      * RamDiscardManager sections; i.e., we don't expect to have gIOMMU sections
803      * in the same address space as RamDiscardManager sections.
804      *
805      * We assume that each section in the address space consumes one memslot.
806      * We take the number of KVM memory slots as a best guess for the maximum
807      * number of sections in the address space we could have over time,
808      * also consuming DMA mappings.
809      */
810     if (container->dma_max_mappings) {
811         unsigned int vrdl_count = 0, vrdl_mappings = 0, max_memslots = 512;
812 
813 #ifdef CONFIG_KVM
814         if (kvm_enabled()) {
815             max_memslots = kvm_get_max_memslots();
816         }
817 #endif
818 
819         QLIST_FOREACH(vrdl, &container->vrdl_list, next) {
820             hwaddr start, end;
821 
822             start = QEMU_ALIGN_DOWN(vrdl->offset_within_address_space,
823                                     vrdl->granularity);
824             end = ROUND_UP(vrdl->offset_within_address_space + vrdl->size,
825                            vrdl->granularity);
826             vrdl_mappings += (end - start) / vrdl->granularity;
827             vrdl_count++;
828         }
829 
830         if (vrdl_mappings + max_memslots - vrdl_count >
831             container->dma_max_mappings) {
832             warn_report("%s: possibly running out of DMA mappings. E.g., try"
833                         " increasing the 'block-size' of virtio-mem devies."
834                         " Maximum possible DMA mappings: %d, Maximum possible"
835                         " memslots: %d", __func__, container->dma_max_mappings,
836                         max_memslots);
837         }
838     }
839 }
840 
841 static void vfio_unregister_ram_discard_listener(VFIOContainer *container,
842                                                  MemoryRegionSection *section)
843 {
844     RamDiscardManager *rdm = memory_region_get_ram_discard_manager(section->mr);
845     VFIORamDiscardListener *vrdl = NULL;
846 
847     QLIST_FOREACH(vrdl, &container->vrdl_list, next) {
848         if (vrdl->mr == section->mr &&
849             vrdl->offset_within_address_space ==
850             section->offset_within_address_space) {
851             break;
852         }
853     }
854 
855     if (!vrdl) {
856         hw_error("vfio: Trying to unregister missing RAM discard listener");
857     }
858 
859     ram_discard_manager_unregister_listener(rdm, &vrdl->listener);
860     QLIST_REMOVE(vrdl, next);
861     g_free(vrdl);
862 }
863 
864 static void vfio_listener_region_add(MemoryListener *listener,
865                                      MemoryRegionSection *section)
866 {
867     VFIOContainer *container = container_of(listener, VFIOContainer, listener);
868     hwaddr iova, end;
869     Int128 llend, llsize;
870     void *vaddr;
871     int ret;
872     VFIOHostDMAWindow *hostwin;
873     bool hostwin_found;
874     Error *err = NULL;
875 
876     if (vfio_listener_skipped_section(section)) {
877         trace_vfio_listener_region_add_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 &
885                   ~qemu_real_host_page_mask()) !=
886                  (section->offset_within_region & ~qemu_real_host_page_mask()))) {
887         error_report("%s received unaligned region", __func__);
888         return;
889     }
890 
891     iova = REAL_HOST_PAGE_ALIGN(section->offset_within_address_space);
892     llend = int128_make64(section->offset_within_address_space);
893     llend = int128_add(llend, section->size);
894     llend = int128_and(llend, int128_exts64(qemu_real_host_page_mask()));
895 
896     if (int128_ge(int128_make64(iova), llend)) {
897         if (memory_region_is_ram_device(section->mr)) {
898             trace_vfio_listener_region_add_no_dma_map(
899                 memory_region_name(section->mr),
900                 section->offset_within_address_space,
901                 int128_getlo(section->size),
902                 qemu_real_host_page_size());
903         }
904         return;
905     }
906     end = int128_get64(int128_sub(llend, int128_one()));
907 
908     if (container->iommu_type == VFIO_SPAPR_TCE_v2_IOMMU) {
909         hwaddr pgsize = 0;
910 
911         /* For now intersections are not allowed, we may relax this later */
912         QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) {
913             if (ranges_overlap(hostwin->min_iova,
914                                hostwin->max_iova - hostwin->min_iova + 1,
915                                section->offset_within_address_space,
916                                int128_get64(section->size))) {
917                 error_setg(&err,
918                     "region [0x%"PRIx64",0x%"PRIx64"] overlaps with existing"
919                     "host DMA window [0x%"PRIx64",0x%"PRIx64"]",
920                     section->offset_within_address_space,
921                     section->offset_within_address_space +
922                         int128_get64(section->size) - 1,
923                     hostwin->min_iova, hostwin->max_iova);
924                 goto fail;
925             }
926         }
927 
928         ret = vfio_spapr_create_window(container, section, &pgsize);
929         if (ret) {
930             error_setg_errno(&err, -ret, "Failed to create SPAPR window");
931             goto fail;
932         }
933 
934         vfio_host_win_add(container, section->offset_within_address_space,
935                           section->offset_within_address_space +
936                           int128_get64(section->size) - 1, pgsize);
937 #ifdef CONFIG_KVM
938         if (kvm_enabled()) {
939             VFIOGroup *group;
940             IOMMUMemoryRegion *iommu_mr = IOMMU_MEMORY_REGION(section->mr);
941             struct kvm_vfio_spapr_tce param;
942             struct kvm_device_attr attr = {
943                 .group = KVM_DEV_VFIO_GROUP,
944                 .attr = KVM_DEV_VFIO_GROUP_SET_SPAPR_TCE,
945                 .addr = (uint64_t)(unsigned long)&param,
946             };
947 
948             if (!memory_region_iommu_get_attr(iommu_mr, IOMMU_ATTR_SPAPR_TCE_FD,
949                                               &param.tablefd)) {
950                 QLIST_FOREACH(group, &container->group_list, container_next) {
951                     param.groupfd = group->fd;
952                     if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) {
953                         error_report("vfio: failed to setup fd %d "
954                                      "for a group with fd %d: %s",
955                                      param.tablefd, param.groupfd,
956                                      strerror(errno));
957                         return;
958                     }
959                     trace_vfio_spapr_group_attach(param.groupfd, param.tablefd);
960                 }
961             }
962         }
963 #endif
964     }
965 
966     hostwin_found = false;
967     QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) {
968         if (hostwin->min_iova <= iova && end <= hostwin->max_iova) {
969             hostwin_found = true;
970             break;
971         }
972     }
973 
974     if (!hostwin_found) {
975         error_setg(&err, "Container %p can't map guest IOVA region"
976                    " 0x%"HWADDR_PRIx"..0x%"HWADDR_PRIx, container, iova, end);
977         goto fail;
978     }
979 
980     memory_region_ref(section->mr);
981 
982     if (memory_region_is_iommu(section->mr)) {
983         VFIOGuestIOMMU *giommu;
984         IOMMUMemoryRegion *iommu_mr = IOMMU_MEMORY_REGION(section->mr);
985         int iommu_idx;
986 
987         trace_vfio_listener_region_add_iommu(iova, end);
988         /*
989          * FIXME: For VFIO iommu types which have KVM acceleration to
990          * avoid bouncing all map/unmaps through qemu this way, this
991          * would be the right place to wire that up (tell the KVM
992          * device emulation the VFIO iommu handles to use).
993          */
994         giommu = g_malloc0(sizeof(*giommu));
995         giommu->iommu = iommu_mr;
996         giommu->iommu_offset = section->offset_within_address_space -
997                                section->offset_within_region;
998         giommu->container = container;
999         llend = int128_add(int128_make64(section->offset_within_region),
1000                            section->size);
1001         llend = int128_sub(llend, int128_one());
1002         iommu_idx = memory_region_iommu_attrs_to_index(iommu_mr,
1003                                                        MEMTXATTRS_UNSPECIFIED);
1004         iommu_notifier_init(&giommu->n, vfio_iommu_map_notify,
1005                             IOMMU_NOTIFIER_IOTLB_EVENTS,
1006                             section->offset_within_region,
1007                             int128_get64(llend),
1008                             iommu_idx);
1009 
1010         ret = memory_region_iommu_set_page_size_mask(giommu->iommu,
1011                                                      container->pgsizes,
1012                                                      &err);
1013         if (ret) {
1014             g_free(giommu);
1015             goto fail;
1016         }
1017 
1018         ret = memory_region_register_iommu_notifier(section->mr, &giommu->n,
1019                                                     &err);
1020         if (ret) {
1021             g_free(giommu);
1022             goto fail;
1023         }
1024         QLIST_INSERT_HEAD(&container->giommu_list, giommu, giommu_next);
1025         memory_region_iommu_replay(giommu->iommu, &giommu->n);
1026 
1027         return;
1028     }
1029 
1030     /* Here we assume that memory_region_is_ram(section->mr)==true */
1031 
1032     /*
1033      * For RAM memory regions with a RamDiscardManager, we only want to map the
1034      * actually populated parts - and update the mapping whenever we're notified
1035      * about changes.
1036      */
1037     if (memory_region_has_ram_discard_manager(section->mr)) {
1038         vfio_register_ram_discard_listener(container, section);
1039         return;
1040     }
1041 
1042     vaddr = memory_region_get_ram_ptr(section->mr) +
1043             section->offset_within_region +
1044             (iova - section->offset_within_address_space);
1045 
1046     trace_vfio_listener_region_add_ram(iova, end, vaddr);
1047 
1048     llsize = int128_sub(llend, int128_make64(iova));
1049 
1050     if (memory_region_is_ram_device(section->mr)) {
1051         hwaddr pgmask = (1ULL << ctz64(hostwin->iova_pgsizes)) - 1;
1052 
1053         if ((iova & pgmask) || (int128_get64(llsize) & pgmask)) {
1054             trace_vfio_listener_region_add_no_dma_map(
1055                 memory_region_name(section->mr),
1056                 section->offset_within_address_space,
1057                 int128_getlo(section->size),
1058                 pgmask + 1);
1059             return;
1060         }
1061     }
1062 
1063     ret = vfio_dma_map(container, iova, int128_get64(llsize),
1064                        vaddr, section->readonly);
1065     if (ret) {
1066         error_setg(&err, "vfio_dma_map(%p, 0x%"HWADDR_PRIx", "
1067                    "0x%"HWADDR_PRIx", %p) = %d (%m)",
1068                    container, iova, int128_get64(llsize), vaddr, ret);
1069         if (memory_region_is_ram_device(section->mr)) {
1070             /* Allow unexpected mappings not to be fatal for RAM devices */
1071             error_report_err(err);
1072             return;
1073         }
1074         goto fail;
1075     }
1076 
1077     return;
1078 
1079 fail:
1080     if (memory_region_is_ram_device(section->mr)) {
1081         error_report("failed to vfio_dma_map. pci p2p may not work");
1082         return;
1083     }
1084     /*
1085      * On the initfn path, store the first error in the container so we
1086      * can gracefully fail.  Runtime, there's not much we can do other
1087      * than throw a hardware error.
1088      */
1089     if (!container->initialized) {
1090         if (!container->error) {
1091             error_propagate_prepend(&container->error, err,
1092                                     "Region %s: ",
1093                                     memory_region_name(section->mr));
1094         } else {
1095             error_free(err);
1096         }
1097     } else {
1098         error_report_err(err);
1099         hw_error("vfio: DMA mapping failed, unable to continue");
1100     }
1101 }
1102 
1103 static void vfio_listener_region_del(MemoryListener *listener,
1104                                      MemoryRegionSection *section)
1105 {
1106     VFIOContainer *container = container_of(listener, VFIOContainer, listener);
1107     hwaddr iova, end;
1108     Int128 llend, llsize;
1109     int ret;
1110     bool try_unmap = true;
1111 
1112     if (vfio_listener_skipped_section(section)) {
1113         trace_vfio_listener_region_del_skip(
1114                 section->offset_within_address_space,
1115                 section->offset_within_address_space +
1116                 int128_get64(int128_sub(section->size, int128_one())));
1117         return;
1118     }
1119 
1120     if (unlikely((section->offset_within_address_space &
1121                   ~qemu_real_host_page_mask()) !=
1122                  (section->offset_within_region & ~qemu_real_host_page_mask()))) {
1123         error_report("%s received unaligned region", __func__);
1124         return;
1125     }
1126 
1127     if (memory_region_is_iommu(section->mr)) {
1128         VFIOGuestIOMMU *giommu;
1129 
1130         QLIST_FOREACH(giommu, &container->giommu_list, giommu_next) {
1131             if (MEMORY_REGION(giommu->iommu) == section->mr &&
1132                 giommu->n.start == section->offset_within_region) {
1133                 memory_region_unregister_iommu_notifier(section->mr,
1134                                                         &giommu->n);
1135                 QLIST_REMOVE(giommu, giommu_next);
1136                 g_free(giommu);
1137                 break;
1138             }
1139         }
1140 
1141         /*
1142          * FIXME: We assume the one big unmap below is adequate to
1143          * remove any individual page mappings in the IOMMU which
1144          * might have been copied into VFIO. This works for a page table
1145          * based IOMMU where a big unmap flattens a large range of IO-PTEs.
1146          * That may not be true for all IOMMU types.
1147          */
1148     }
1149 
1150     iova = REAL_HOST_PAGE_ALIGN(section->offset_within_address_space);
1151     llend = int128_make64(section->offset_within_address_space);
1152     llend = int128_add(llend, section->size);
1153     llend = int128_and(llend, int128_exts64(qemu_real_host_page_mask()));
1154 
1155     if (int128_ge(int128_make64(iova), llend)) {
1156         return;
1157     }
1158     end = int128_get64(int128_sub(llend, int128_one()));
1159 
1160     llsize = int128_sub(llend, int128_make64(iova));
1161 
1162     trace_vfio_listener_region_del(iova, end);
1163 
1164     if (memory_region_is_ram_device(section->mr)) {
1165         hwaddr pgmask;
1166         VFIOHostDMAWindow *hostwin;
1167         bool hostwin_found = false;
1168 
1169         QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) {
1170             if (hostwin->min_iova <= iova && end <= hostwin->max_iova) {
1171                 hostwin_found = true;
1172                 break;
1173             }
1174         }
1175         assert(hostwin_found); /* or region_add() would have failed */
1176 
1177         pgmask = (1ULL << ctz64(hostwin->iova_pgsizes)) - 1;
1178         try_unmap = !((iova & pgmask) || (int128_get64(llsize) & pgmask));
1179     } else if (memory_region_has_ram_discard_manager(section->mr)) {
1180         vfio_unregister_ram_discard_listener(container, section);
1181         /* Unregistering will trigger an unmap. */
1182         try_unmap = false;
1183     }
1184 
1185     if (try_unmap) {
1186         if (int128_eq(llsize, int128_2_64())) {
1187             /* The unmap ioctl doesn't accept a full 64-bit span. */
1188             llsize = int128_rshift(llsize, 1);
1189             ret = vfio_dma_unmap(container, iova, int128_get64(llsize), NULL);
1190             if (ret) {
1191                 error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx", "
1192                              "0x%"HWADDR_PRIx") = %d (%m)",
1193                              container, iova, int128_get64(llsize), ret);
1194             }
1195             iova += int128_get64(llsize);
1196         }
1197         ret = vfio_dma_unmap(container, iova, int128_get64(llsize), NULL);
1198         if (ret) {
1199             error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx", "
1200                          "0x%"HWADDR_PRIx") = %d (%m)",
1201                          container, iova, int128_get64(llsize), ret);
1202         }
1203     }
1204 
1205     memory_region_unref(section->mr);
1206 
1207     if (container->iommu_type == VFIO_SPAPR_TCE_v2_IOMMU) {
1208         vfio_spapr_remove_window(container,
1209                                  section->offset_within_address_space);
1210         if (vfio_host_win_del(container,
1211                               section->offset_within_address_space,
1212                               section->offset_within_address_space +
1213                               int128_get64(section->size) - 1) < 0) {
1214             hw_error("%s: Cannot delete missing window at %"HWADDR_PRIx,
1215                      __func__, section->offset_within_address_space);
1216         }
1217     }
1218 }
1219 
1220 static void vfio_set_dirty_page_tracking(VFIOContainer *container, bool start)
1221 {
1222     int ret;
1223     struct vfio_iommu_type1_dirty_bitmap dirty = {
1224         .argsz = sizeof(dirty),
1225     };
1226 
1227     if (start) {
1228         dirty.flags = VFIO_IOMMU_DIRTY_PAGES_FLAG_START;
1229     } else {
1230         dirty.flags = VFIO_IOMMU_DIRTY_PAGES_FLAG_STOP;
1231     }
1232 
1233     ret = ioctl(container->fd, VFIO_IOMMU_DIRTY_PAGES, &dirty);
1234     if (ret) {
1235         error_report("Failed to set dirty tracking flag 0x%x errno: %d",
1236                      dirty.flags, errno);
1237     }
1238 }
1239 
1240 static void vfio_listener_log_global_start(MemoryListener *listener)
1241 {
1242     VFIOContainer *container = container_of(listener, VFIOContainer, listener);
1243 
1244     vfio_set_dirty_page_tracking(container, true);
1245 }
1246 
1247 static void vfio_listener_log_global_stop(MemoryListener *listener)
1248 {
1249     VFIOContainer *container = container_of(listener, VFIOContainer, listener);
1250 
1251     vfio_set_dirty_page_tracking(container, false);
1252 }
1253 
1254 static int vfio_get_dirty_bitmap(VFIOContainer *container, uint64_t iova,
1255                                  uint64_t size, ram_addr_t ram_addr)
1256 {
1257     struct vfio_iommu_type1_dirty_bitmap *dbitmap;
1258     struct vfio_iommu_type1_dirty_bitmap_get *range;
1259     uint64_t pages;
1260     int ret;
1261 
1262     dbitmap = g_malloc0(sizeof(*dbitmap) + sizeof(*range));
1263 
1264     dbitmap->argsz = sizeof(*dbitmap) + sizeof(*range);
1265     dbitmap->flags = VFIO_IOMMU_DIRTY_PAGES_FLAG_GET_BITMAP;
1266     range = (struct vfio_iommu_type1_dirty_bitmap_get *)&dbitmap->data;
1267     range->iova = iova;
1268     range->size = size;
1269 
1270     /*
1271      * cpu_physical_memory_set_dirty_lebitmap() supports pages in bitmap of
1272      * qemu_real_host_page_size to mark those dirty. Hence set bitmap's pgsize
1273      * to qemu_real_host_page_size.
1274      */
1275     range->bitmap.pgsize = qemu_real_host_page_size();
1276 
1277     pages = REAL_HOST_PAGE_ALIGN(range->size) / qemu_real_host_page_size();
1278     range->bitmap.size = ROUND_UP(pages, sizeof(__u64) * BITS_PER_BYTE) /
1279                                          BITS_PER_BYTE;
1280     range->bitmap.data = g_try_malloc0(range->bitmap.size);
1281     if (!range->bitmap.data) {
1282         ret = -ENOMEM;
1283         goto err_out;
1284     }
1285 
1286     ret = ioctl(container->fd, VFIO_IOMMU_DIRTY_PAGES, dbitmap);
1287     if (ret) {
1288         error_report("Failed to get dirty bitmap for iova: 0x%"PRIx64
1289                 " size: 0x%"PRIx64" err: %d", (uint64_t)range->iova,
1290                 (uint64_t)range->size, errno);
1291         goto err_out;
1292     }
1293 
1294     cpu_physical_memory_set_dirty_lebitmap((unsigned long *)range->bitmap.data,
1295                                             ram_addr, pages);
1296 
1297     trace_vfio_get_dirty_bitmap(container->fd, range->iova, range->size,
1298                                 range->bitmap.size, ram_addr);
1299 err_out:
1300     g_free(range->bitmap.data);
1301     g_free(dbitmap);
1302 
1303     return ret;
1304 }
1305 
1306 typedef struct {
1307     IOMMUNotifier n;
1308     VFIOGuestIOMMU *giommu;
1309 } vfio_giommu_dirty_notifier;
1310 
1311 static void vfio_iommu_map_dirty_notify(IOMMUNotifier *n, IOMMUTLBEntry *iotlb)
1312 {
1313     vfio_giommu_dirty_notifier *gdn = container_of(n,
1314                                                 vfio_giommu_dirty_notifier, n);
1315     VFIOGuestIOMMU *giommu = gdn->giommu;
1316     VFIOContainer *container = giommu->container;
1317     hwaddr iova = iotlb->iova + giommu->iommu_offset;
1318     ram_addr_t translated_addr;
1319 
1320     trace_vfio_iommu_map_dirty_notify(iova, iova + iotlb->addr_mask);
1321 
1322     if (iotlb->target_as != &address_space_memory) {
1323         error_report("Wrong target AS \"%s\", only system memory is allowed",
1324                      iotlb->target_as->name ? iotlb->target_as->name : "none");
1325         return;
1326     }
1327 
1328     rcu_read_lock();
1329     if (vfio_get_xlat_addr(iotlb, NULL, &translated_addr, NULL)) {
1330         int ret;
1331 
1332         ret = vfio_get_dirty_bitmap(container, iova, iotlb->addr_mask + 1,
1333                                     translated_addr);
1334         if (ret) {
1335             error_report("vfio_iommu_map_dirty_notify(%p, 0x%"HWADDR_PRIx", "
1336                          "0x%"HWADDR_PRIx") = %d (%m)",
1337                          container, iova,
1338                          iotlb->addr_mask + 1, ret);
1339         }
1340     }
1341     rcu_read_unlock();
1342 }
1343 
1344 static int vfio_ram_discard_get_dirty_bitmap(MemoryRegionSection *section,
1345                                              void *opaque)
1346 {
1347     const hwaddr size = int128_get64(section->size);
1348     const hwaddr iova = section->offset_within_address_space;
1349     const ram_addr_t ram_addr = memory_region_get_ram_addr(section->mr) +
1350                                 section->offset_within_region;
1351     VFIORamDiscardListener *vrdl = opaque;
1352 
1353     /*
1354      * Sync the whole mapped region (spanning multiple individual mappings)
1355      * in one go.
1356      */
1357     return vfio_get_dirty_bitmap(vrdl->container, iova, size, ram_addr);
1358 }
1359 
1360 static int vfio_sync_ram_discard_listener_dirty_bitmap(VFIOContainer *container,
1361                                                    MemoryRegionSection *section)
1362 {
1363     RamDiscardManager *rdm = memory_region_get_ram_discard_manager(section->mr);
1364     VFIORamDiscardListener *vrdl = NULL;
1365 
1366     QLIST_FOREACH(vrdl, &container->vrdl_list, next) {
1367         if (vrdl->mr == section->mr &&
1368             vrdl->offset_within_address_space ==
1369             section->offset_within_address_space) {
1370             break;
1371         }
1372     }
1373 
1374     if (!vrdl) {
1375         hw_error("vfio: Trying to sync missing RAM discard listener");
1376     }
1377 
1378     /*
1379      * We only want/can synchronize the bitmap for actually mapped parts -
1380      * which correspond to populated parts. Replay all populated parts.
1381      */
1382     return ram_discard_manager_replay_populated(rdm, section,
1383                                               vfio_ram_discard_get_dirty_bitmap,
1384                                                 &vrdl);
1385 }
1386 
1387 static int vfio_sync_dirty_bitmap(VFIOContainer *container,
1388                                   MemoryRegionSection *section)
1389 {
1390     ram_addr_t ram_addr;
1391 
1392     if (memory_region_is_iommu(section->mr)) {
1393         VFIOGuestIOMMU *giommu;
1394 
1395         QLIST_FOREACH(giommu, &container->giommu_list, giommu_next) {
1396             if (MEMORY_REGION(giommu->iommu) == section->mr &&
1397                 giommu->n.start == section->offset_within_region) {
1398                 Int128 llend;
1399                 vfio_giommu_dirty_notifier gdn = { .giommu = giommu };
1400                 int idx = memory_region_iommu_attrs_to_index(giommu->iommu,
1401                                                        MEMTXATTRS_UNSPECIFIED);
1402 
1403                 llend = int128_add(int128_make64(section->offset_within_region),
1404                                    section->size);
1405                 llend = int128_sub(llend, int128_one());
1406 
1407                 iommu_notifier_init(&gdn.n,
1408                                     vfio_iommu_map_dirty_notify,
1409                                     IOMMU_NOTIFIER_MAP,
1410                                     section->offset_within_region,
1411                                     int128_get64(llend),
1412                                     idx);
1413                 memory_region_iommu_replay(giommu->iommu, &gdn.n);
1414                 break;
1415             }
1416         }
1417         return 0;
1418     } else if (memory_region_has_ram_discard_manager(section->mr)) {
1419         return vfio_sync_ram_discard_listener_dirty_bitmap(container, section);
1420     }
1421 
1422     ram_addr = memory_region_get_ram_addr(section->mr) +
1423                section->offset_within_region;
1424 
1425     return vfio_get_dirty_bitmap(container,
1426                    REAL_HOST_PAGE_ALIGN(section->offset_within_address_space),
1427                    int128_get64(section->size), ram_addr);
1428 }
1429 
1430 static void vfio_listener_log_sync(MemoryListener *listener,
1431         MemoryRegionSection *section)
1432 {
1433     VFIOContainer *container = container_of(listener, VFIOContainer, listener);
1434 
1435     if (vfio_listener_skipped_section(section) ||
1436         !container->dirty_pages_supported) {
1437         return;
1438     }
1439 
1440     if (vfio_devices_all_dirty_tracking(container)) {
1441         vfio_sync_dirty_bitmap(container, section);
1442     }
1443 }
1444 
1445 static const MemoryListener vfio_memory_listener = {
1446     .name = "vfio",
1447     .region_add = vfio_listener_region_add,
1448     .region_del = vfio_listener_region_del,
1449     .log_global_start = vfio_listener_log_global_start,
1450     .log_global_stop = vfio_listener_log_global_stop,
1451     .log_sync = vfio_listener_log_sync,
1452 };
1453 
1454 static void vfio_listener_release(VFIOContainer *container)
1455 {
1456     memory_listener_unregister(&container->listener);
1457     if (container->iommu_type == VFIO_SPAPR_TCE_v2_IOMMU) {
1458         memory_listener_unregister(&container->prereg_listener);
1459     }
1460 }
1461 
1462 static struct vfio_info_cap_header *
1463 vfio_get_cap(void *ptr, uint32_t cap_offset, uint16_t id)
1464 {
1465     struct vfio_info_cap_header *hdr;
1466 
1467     for (hdr = ptr + cap_offset; hdr != ptr; hdr = ptr + hdr->next) {
1468         if (hdr->id == id) {
1469             return hdr;
1470         }
1471     }
1472 
1473     return NULL;
1474 }
1475 
1476 struct vfio_info_cap_header *
1477 vfio_get_region_info_cap(struct vfio_region_info *info, uint16_t id)
1478 {
1479     if (!(info->flags & VFIO_REGION_INFO_FLAG_CAPS)) {
1480         return NULL;
1481     }
1482 
1483     return vfio_get_cap((void *)info, info->cap_offset, id);
1484 }
1485 
1486 static struct vfio_info_cap_header *
1487 vfio_get_iommu_type1_info_cap(struct vfio_iommu_type1_info *info, uint16_t id)
1488 {
1489     if (!(info->flags & VFIO_IOMMU_INFO_CAPS)) {
1490         return NULL;
1491     }
1492 
1493     return vfio_get_cap((void *)info, info->cap_offset, id);
1494 }
1495 
1496 struct vfio_info_cap_header *
1497 vfio_get_device_info_cap(struct vfio_device_info *info, uint16_t id)
1498 {
1499     if (!(info->flags & VFIO_DEVICE_FLAGS_CAPS)) {
1500         return NULL;
1501     }
1502 
1503     return vfio_get_cap((void *)info, info->cap_offset, id);
1504 }
1505 
1506 bool vfio_get_info_dma_avail(struct vfio_iommu_type1_info *info,
1507                              unsigned int *avail)
1508 {
1509     struct vfio_info_cap_header *hdr;
1510     struct vfio_iommu_type1_info_dma_avail *cap;
1511 
1512     /* If the capability cannot be found, assume no DMA limiting */
1513     hdr = vfio_get_iommu_type1_info_cap(info,
1514                                         VFIO_IOMMU_TYPE1_INFO_DMA_AVAIL);
1515     if (hdr == NULL) {
1516         return false;
1517     }
1518 
1519     if (avail != NULL) {
1520         cap = (void *) hdr;
1521         *avail = cap->avail;
1522     }
1523 
1524     return true;
1525 }
1526 
1527 static int vfio_setup_region_sparse_mmaps(VFIORegion *region,
1528                                           struct vfio_region_info *info)
1529 {
1530     struct vfio_info_cap_header *hdr;
1531     struct vfio_region_info_cap_sparse_mmap *sparse;
1532     int i, j;
1533 
1534     hdr = vfio_get_region_info_cap(info, VFIO_REGION_INFO_CAP_SPARSE_MMAP);
1535     if (!hdr) {
1536         return -ENODEV;
1537     }
1538 
1539     sparse = container_of(hdr, struct vfio_region_info_cap_sparse_mmap, header);
1540 
1541     trace_vfio_region_sparse_mmap_header(region->vbasedev->name,
1542                                          region->nr, sparse->nr_areas);
1543 
1544     region->mmaps = g_new0(VFIOMmap, sparse->nr_areas);
1545 
1546     for (i = 0, j = 0; i < sparse->nr_areas; i++) {
1547         trace_vfio_region_sparse_mmap_entry(i, sparse->areas[i].offset,
1548                                             sparse->areas[i].offset +
1549                                             sparse->areas[i].size);
1550 
1551         if (sparse->areas[i].size) {
1552             region->mmaps[j].offset = sparse->areas[i].offset;
1553             region->mmaps[j].size = sparse->areas[i].size;
1554             j++;
1555         }
1556     }
1557 
1558     region->nr_mmaps = j;
1559     region->mmaps = g_realloc(region->mmaps, j * sizeof(VFIOMmap));
1560 
1561     return 0;
1562 }
1563 
1564 int vfio_region_setup(Object *obj, VFIODevice *vbasedev, VFIORegion *region,
1565                       int index, const char *name)
1566 {
1567     struct vfio_region_info *info;
1568     int ret;
1569 
1570     ret = vfio_get_region_info(vbasedev, index, &info);
1571     if (ret) {
1572         return ret;
1573     }
1574 
1575     region->vbasedev = vbasedev;
1576     region->flags = info->flags;
1577     region->size = info->size;
1578     region->fd_offset = info->offset;
1579     region->nr = index;
1580 
1581     if (region->size) {
1582         region->mem = g_new0(MemoryRegion, 1);
1583         memory_region_init_io(region->mem, obj, &vfio_region_ops,
1584                               region, name, region->size);
1585 
1586         if (!vbasedev->no_mmap &&
1587             region->flags & VFIO_REGION_INFO_FLAG_MMAP) {
1588 
1589             ret = vfio_setup_region_sparse_mmaps(region, info);
1590 
1591             if (ret) {
1592                 region->nr_mmaps = 1;
1593                 region->mmaps = g_new0(VFIOMmap, region->nr_mmaps);
1594                 region->mmaps[0].offset = 0;
1595                 region->mmaps[0].size = region->size;
1596             }
1597         }
1598     }
1599 
1600     g_free(info);
1601 
1602     trace_vfio_region_setup(vbasedev->name, index, name,
1603                             region->flags, region->fd_offset, region->size);
1604     return 0;
1605 }
1606 
1607 static void vfio_subregion_unmap(VFIORegion *region, int index)
1608 {
1609     trace_vfio_region_unmap(memory_region_name(&region->mmaps[index].mem),
1610                             region->mmaps[index].offset,
1611                             region->mmaps[index].offset +
1612                             region->mmaps[index].size - 1);
1613     memory_region_del_subregion(region->mem, &region->mmaps[index].mem);
1614     munmap(region->mmaps[index].mmap, region->mmaps[index].size);
1615     object_unparent(OBJECT(&region->mmaps[index].mem));
1616     region->mmaps[index].mmap = NULL;
1617 }
1618 
1619 int vfio_region_mmap(VFIORegion *region)
1620 {
1621     int i, prot = 0;
1622     char *name;
1623 
1624     if (!region->mem) {
1625         return 0;
1626     }
1627 
1628     prot |= region->flags & VFIO_REGION_INFO_FLAG_READ ? PROT_READ : 0;
1629     prot |= region->flags & VFIO_REGION_INFO_FLAG_WRITE ? PROT_WRITE : 0;
1630 
1631     for (i = 0; i < region->nr_mmaps; i++) {
1632         region->mmaps[i].mmap = mmap(NULL, region->mmaps[i].size, prot,
1633                                      MAP_SHARED, region->vbasedev->fd,
1634                                      region->fd_offset +
1635                                      region->mmaps[i].offset);
1636         if (region->mmaps[i].mmap == MAP_FAILED) {
1637             int ret = -errno;
1638 
1639             trace_vfio_region_mmap_fault(memory_region_name(region->mem), i,
1640                                          region->fd_offset +
1641                                          region->mmaps[i].offset,
1642                                          region->fd_offset +
1643                                          region->mmaps[i].offset +
1644                                          region->mmaps[i].size - 1, ret);
1645 
1646             region->mmaps[i].mmap = NULL;
1647 
1648             for (i--; i >= 0; i--) {
1649                 vfio_subregion_unmap(region, i);
1650             }
1651 
1652             return ret;
1653         }
1654 
1655         name = g_strdup_printf("%s mmaps[%d]",
1656                                memory_region_name(region->mem), i);
1657         memory_region_init_ram_device_ptr(&region->mmaps[i].mem,
1658                                           memory_region_owner(region->mem),
1659                                           name, region->mmaps[i].size,
1660                                           region->mmaps[i].mmap);
1661         g_free(name);
1662         memory_region_add_subregion(region->mem, region->mmaps[i].offset,
1663                                     &region->mmaps[i].mem);
1664 
1665         trace_vfio_region_mmap(memory_region_name(&region->mmaps[i].mem),
1666                                region->mmaps[i].offset,
1667                                region->mmaps[i].offset +
1668                                region->mmaps[i].size - 1);
1669     }
1670 
1671     return 0;
1672 }
1673 
1674 void vfio_region_unmap(VFIORegion *region)
1675 {
1676     int i;
1677 
1678     if (!region->mem) {
1679         return;
1680     }
1681 
1682     for (i = 0; i < region->nr_mmaps; i++) {
1683         if (region->mmaps[i].mmap) {
1684             vfio_subregion_unmap(region, i);
1685         }
1686     }
1687 }
1688 
1689 void vfio_region_exit(VFIORegion *region)
1690 {
1691     int i;
1692 
1693     if (!region->mem) {
1694         return;
1695     }
1696 
1697     for (i = 0; i < region->nr_mmaps; i++) {
1698         if (region->mmaps[i].mmap) {
1699             memory_region_del_subregion(region->mem, &region->mmaps[i].mem);
1700         }
1701     }
1702 
1703     trace_vfio_region_exit(region->vbasedev->name, region->nr);
1704 }
1705 
1706 void vfio_region_finalize(VFIORegion *region)
1707 {
1708     int i;
1709 
1710     if (!region->mem) {
1711         return;
1712     }
1713 
1714     for (i = 0; i < region->nr_mmaps; i++) {
1715         if (region->mmaps[i].mmap) {
1716             munmap(region->mmaps[i].mmap, region->mmaps[i].size);
1717             object_unparent(OBJECT(&region->mmaps[i].mem));
1718         }
1719     }
1720 
1721     object_unparent(OBJECT(region->mem));
1722 
1723     g_free(region->mem);
1724     g_free(region->mmaps);
1725 
1726     trace_vfio_region_finalize(region->vbasedev->name, region->nr);
1727 
1728     region->mem = NULL;
1729     region->mmaps = NULL;
1730     region->nr_mmaps = 0;
1731     region->size = 0;
1732     region->flags = 0;
1733     region->nr = 0;
1734 }
1735 
1736 void vfio_region_mmaps_set_enabled(VFIORegion *region, bool enabled)
1737 {
1738     int i;
1739 
1740     if (!region->mem) {
1741         return;
1742     }
1743 
1744     for (i = 0; i < region->nr_mmaps; i++) {
1745         if (region->mmaps[i].mmap) {
1746             memory_region_set_enabled(&region->mmaps[i].mem, enabled);
1747         }
1748     }
1749 
1750     trace_vfio_region_mmaps_set_enabled(memory_region_name(region->mem),
1751                                         enabled);
1752 }
1753 
1754 void vfio_reset_handler(void *opaque)
1755 {
1756     VFIOGroup *group;
1757     VFIODevice *vbasedev;
1758 
1759     QLIST_FOREACH(group, &vfio_group_list, next) {
1760         QLIST_FOREACH(vbasedev, &group->device_list, next) {
1761             if (vbasedev->dev->realized) {
1762                 vbasedev->ops->vfio_compute_needs_reset(vbasedev);
1763             }
1764         }
1765     }
1766 
1767     QLIST_FOREACH(group, &vfio_group_list, next) {
1768         QLIST_FOREACH(vbasedev, &group->device_list, next) {
1769             if (vbasedev->dev->realized && vbasedev->needs_reset) {
1770                 vbasedev->ops->vfio_hot_reset_multi(vbasedev);
1771             }
1772         }
1773     }
1774 }
1775 
1776 static void vfio_kvm_device_add_group(VFIOGroup *group)
1777 {
1778 #ifdef CONFIG_KVM
1779     struct kvm_device_attr attr = {
1780         .group = KVM_DEV_VFIO_GROUP,
1781         .attr = KVM_DEV_VFIO_GROUP_ADD,
1782         .addr = (uint64_t)(unsigned long)&group->fd,
1783     };
1784 
1785     if (!kvm_enabled()) {
1786         return;
1787     }
1788 
1789     if (vfio_kvm_device_fd < 0) {
1790         struct kvm_create_device cd = {
1791             .type = KVM_DEV_TYPE_VFIO,
1792         };
1793 
1794         if (kvm_vm_ioctl(kvm_state, KVM_CREATE_DEVICE, &cd)) {
1795             error_report("Failed to create KVM VFIO device: %m");
1796             return;
1797         }
1798 
1799         vfio_kvm_device_fd = cd.fd;
1800     }
1801 
1802     if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) {
1803         error_report("Failed to add group %d to KVM VFIO device: %m",
1804                      group->groupid);
1805     }
1806 #endif
1807 }
1808 
1809 static void vfio_kvm_device_del_group(VFIOGroup *group)
1810 {
1811 #ifdef CONFIG_KVM
1812     struct kvm_device_attr attr = {
1813         .group = KVM_DEV_VFIO_GROUP,
1814         .attr = KVM_DEV_VFIO_GROUP_DEL,
1815         .addr = (uint64_t)(unsigned long)&group->fd,
1816     };
1817 
1818     if (vfio_kvm_device_fd < 0) {
1819         return;
1820     }
1821 
1822     if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) {
1823         error_report("Failed to remove group %d from KVM VFIO device: %m",
1824                      group->groupid);
1825     }
1826 #endif
1827 }
1828 
1829 static VFIOAddressSpace *vfio_get_address_space(AddressSpace *as)
1830 {
1831     VFIOAddressSpace *space;
1832 
1833     QLIST_FOREACH(space, &vfio_address_spaces, list) {
1834         if (space->as == as) {
1835             return space;
1836         }
1837     }
1838 
1839     /* No suitable VFIOAddressSpace, create a new one */
1840     space = g_malloc0(sizeof(*space));
1841     space->as = as;
1842     QLIST_INIT(&space->containers);
1843 
1844     QLIST_INSERT_HEAD(&vfio_address_spaces, space, list);
1845 
1846     return space;
1847 }
1848 
1849 static void vfio_put_address_space(VFIOAddressSpace *space)
1850 {
1851     if (QLIST_EMPTY(&space->containers)) {
1852         QLIST_REMOVE(space, list);
1853         g_free(space);
1854     }
1855 }
1856 
1857 /*
1858  * vfio_get_iommu_type - selects the richest iommu_type (v2 first)
1859  */
1860 static int vfio_get_iommu_type(VFIOContainer *container,
1861                                Error **errp)
1862 {
1863     int iommu_types[] = { VFIO_TYPE1v2_IOMMU, VFIO_TYPE1_IOMMU,
1864                           VFIO_SPAPR_TCE_v2_IOMMU, VFIO_SPAPR_TCE_IOMMU };
1865     int i;
1866 
1867     for (i = 0; i < ARRAY_SIZE(iommu_types); i++) {
1868         if (ioctl(container->fd, VFIO_CHECK_EXTENSION, iommu_types[i])) {
1869             return iommu_types[i];
1870         }
1871     }
1872     error_setg(errp, "No available IOMMU models");
1873     return -EINVAL;
1874 }
1875 
1876 static int vfio_init_container(VFIOContainer *container, int group_fd,
1877                                Error **errp)
1878 {
1879     int iommu_type, ret;
1880 
1881     iommu_type = vfio_get_iommu_type(container, errp);
1882     if (iommu_type < 0) {
1883         return iommu_type;
1884     }
1885 
1886     ret = ioctl(group_fd, VFIO_GROUP_SET_CONTAINER, &container->fd);
1887     if (ret) {
1888         error_setg_errno(errp, errno, "Failed to set group container");
1889         return -errno;
1890     }
1891 
1892     while (ioctl(container->fd, VFIO_SET_IOMMU, iommu_type)) {
1893         if (iommu_type == VFIO_SPAPR_TCE_v2_IOMMU) {
1894             /*
1895              * On sPAPR, despite the IOMMU subdriver always advertises v1 and
1896              * v2, the running platform may not support v2 and there is no
1897              * way to guess it until an IOMMU group gets added to the container.
1898              * So in case it fails with v2, try v1 as a fallback.
1899              */
1900             iommu_type = VFIO_SPAPR_TCE_IOMMU;
1901             continue;
1902         }
1903         error_setg_errno(errp, errno, "Failed to set iommu for container");
1904         return -errno;
1905     }
1906 
1907     container->iommu_type = iommu_type;
1908     return 0;
1909 }
1910 
1911 static int vfio_get_iommu_info(VFIOContainer *container,
1912                                struct vfio_iommu_type1_info **info)
1913 {
1914 
1915     size_t argsz = sizeof(struct vfio_iommu_type1_info);
1916 
1917     *info = g_new0(struct vfio_iommu_type1_info, 1);
1918 again:
1919     (*info)->argsz = argsz;
1920 
1921     if (ioctl(container->fd, VFIO_IOMMU_GET_INFO, *info)) {
1922         g_free(*info);
1923         *info = NULL;
1924         return -errno;
1925     }
1926 
1927     if (((*info)->argsz > argsz)) {
1928         argsz = (*info)->argsz;
1929         *info = g_realloc(*info, argsz);
1930         goto again;
1931     }
1932 
1933     return 0;
1934 }
1935 
1936 static struct vfio_info_cap_header *
1937 vfio_get_iommu_info_cap(struct vfio_iommu_type1_info *info, uint16_t id)
1938 {
1939     struct vfio_info_cap_header *hdr;
1940     void *ptr = info;
1941 
1942     if (!(info->flags & VFIO_IOMMU_INFO_CAPS)) {
1943         return NULL;
1944     }
1945 
1946     for (hdr = ptr + info->cap_offset; hdr != ptr; hdr = ptr + hdr->next) {
1947         if (hdr->id == id) {
1948             return hdr;
1949         }
1950     }
1951 
1952     return NULL;
1953 }
1954 
1955 static void vfio_get_iommu_info_migration(VFIOContainer *container,
1956                                          struct vfio_iommu_type1_info *info)
1957 {
1958     struct vfio_info_cap_header *hdr;
1959     struct vfio_iommu_type1_info_cap_migration *cap_mig;
1960 
1961     hdr = vfio_get_iommu_info_cap(info, VFIO_IOMMU_TYPE1_INFO_CAP_MIGRATION);
1962     if (!hdr) {
1963         return;
1964     }
1965 
1966     cap_mig = container_of(hdr, struct vfio_iommu_type1_info_cap_migration,
1967                             header);
1968 
1969     /*
1970      * cpu_physical_memory_set_dirty_lebitmap() supports pages in bitmap of
1971      * qemu_real_host_page_size to mark those dirty.
1972      */
1973     if (cap_mig->pgsize_bitmap & qemu_real_host_page_size()) {
1974         container->dirty_pages_supported = true;
1975         container->max_dirty_bitmap_size = cap_mig->max_dirty_bitmap_size;
1976         container->dirty_pgsizes = cap_mig->pgsize_bitmap;
1977     }
1978 }
1979 
1980 static int vfio_connect_container(VFIOGroup *group, AddressSpace *as,
1981                                   Error **errp)
1982 {
1983     VFIOContainer *container;
1984     int ret, fd;
1985     VFIOAddressSpace *space;
1986 
1987     space = vfio_get_address_space(as);
1988 
1989     /*
1990      * VFIO is currently incompatible with discarding of RAM insofar as the
1991      * madvise to purge (zap) the page from QEMU's address space does not
1992      * interact with the memory API and therefore leaves stale virtual to
1993      * physical mappings in the IOMMU if the page was previously pinned.  We
1994      * therefore set discarding broken for each group added to a container,
1995      * whether the container is used individually or shared.  This provides
1996      * us with options to allow devices within a group to opt-in and allow
1997      * discarding, so long as it is done consistently for a group (for instance
1998      * if the device is an mdev device where it is known that the host vendor
1999      * driver will never pin pages outside of the working set of the guest
2000      * driver, which would thus not be discarding candidates).
2001      *
2002      * The first opportunity to induce pinning occurs here where we attempt to
2003      * attach the group to existing containers within the AddressSpace.  If any
2004      * pages are already zapped from the virtual address space, such as from
2005      * previous discards, new pinning will cause valid mappings to be
2006      * re-established.  Likewise, when the overall MemoryListener for a new
2007      * container is registered, a replay of mappings within the AddressSpace
2008      * will occur, re-establishing any previously zapped pages as well.
2009      *
2010      * Especially virtio-balloon is currently only prevented from discarding
2011      * new memory, it will not yet set ram_block_discard_set_required() and
2012      * therefore, neither stops us here or deals with the sudden memory
2013      * consumption of inflated memory.
2014      *
2015      * We do support discarding of memory coordinated via the RamDiscardManager
2016      * with some IOMMU types. vfio_ram_block_discard_disable() handles the
2017      * details once we know which type of IOMMU we are using.
2018      */
2019 
2020     QLIST_FOREACH(container, &space->containers, next) {
2021         if (!ioctl(group->fd, VFIO_GROUP_SET_CONTAINER, &container->fd)) {
2022             ret = vfio_ram_block_discard_disable(container, true);
2023             if (ret) {
2024                 error_setg_errno(errp, -ret,
2025                                  "Cannot set discarding of RAM broken");
2026                 if (ioctl(group->fd, VFIO_GROUP_UNSET_CONTAINER,
2027                           &container->fd)) {
2028                     error_report("vfio: error disconnecting group %d from"
2029                                  " container", group->groupid);
2030                 }
2031                 return ret;
2032             }
2033             group->container = container;
2034             QLIST_INSERT_HEAD(&container->group_list, group, container_next);
2035             vfio_kvm_device_add_group(group);
2036             return 0;
2037         }
2038     }
2039 
2040     fd = qemu_open_old("/dev/vfio/vfio", O_RDWR);
2041     if (fd < 0) {
2042         error_setg_errno(errp, errno, "failed to open /dev/vfio/vfio");
2043         ret = -errno;
2044         goto put_space_exit;
2045     }
2046 
2047     ret = ioctl(fd, VFIO_GET_API_VERSION);
2048     if (ret != VFIO_API_VERSION) {
2049         error_setg(errp, "supported vfio version: %d, "
2050                    "reported version: %d", VFIO_API_VERSION, ret);
2051         ret = -EINVAL;
2052         goto close_fd_exit;
2053     }
2054 
2055     container = g_malloc0(sizeof(*container));
2056     container->space = space;
2057     container->fd = fd;
2058     container->error = NULL;
2059     container->dirty_pages_supported = false;
2060     container->dma_max_mappings = 0;
2061     QLIST_INIT(&container->giommu_list);
2062     QLIST_INIT(&container->hostwin_list);
2063     QLIST_INIT(&container->vrdl_list);
2064 
2065     ret = vfio_init_container(container, group->fd, errp);
2066     if (ret) {
2067         goto free_container_exit;
2068     }
2069 
2070     ret = vfio_ram_block_discard_disable(container, true);
2071     if (ret) {
2072         error_setg_errno(errp, -ret, "Cannot set discarding of RAM broken");
2073         goto free_container_exit;
2074     }
2075 
2076     switch (container->iommu_type) {
2077     case VFIO_TYPE1v2_IOMMU:
2078     case VFIO_TYPE1_IOMMU:
2079     {
2080         struct vfio_iommu_type1_info *info;
2081 
2082         /*
2083          * FIXME: This assumes that a Type1 IOMMU can map any 64-bit
2084          * IOVA whatsoever.  That's not actually true, but the current
2085          * kernel interface doesn't tell us what it can map, and the
2086          * existing Type1 IOMMUs generally support any IOVA we're
2087          * going to actually try in practice.
2088          */
2089         ret = vfio_get_iommu_info(container, &info);
2090 
2091         if (ret || !(info->flags & VFIO_IOMMU_INFO_PGSIZES)) {
2092             /* Assume 4k IOVA page size */
2093             info->iova_pgsizes = 4096;
2094         }
2095         vfio_host_win_add(container, 0, (hwaddr)-1, info->iova_pgsizes);
2096         container->pgsizes = info->iova_pgsizes;
2097 
2098         /* The default in the kernel ("dma_entry_limit") is 65535. */
2099         container->dma_max_mappings = 65535;
2100         if (!ret) {
2101             vfio_get_info_dma_avail(info, &container->dma_max_mappings);
2102             vfio_get_iommu_info_migration(container, info);
2103         }
2104         g_free(info);
2105         break;
2106     }
2107     case VFIO_SPAPR_TCE_v2_IOMMU:
2108     case VFIO_SPAPR_TCE_IOMMU:
2109     {
2110         struct vfio_iommu_spapr_tce_info info;
2111         bool v2 = container->iommu_type == VFIO_SPAPR_TCE_v2_IOMMU;
2112 
2113         /*
2114          * The host kernel code implementing VFIO_IOMMU_DISABLE is called
2115          * when container fd is closed so we do not call it explicitly
2116          * in this file.
2117          */
2118         if (!v2) {
2119             ret = ioctl(fd, VFIO_IOMMU_ENABLE);
2120             if (ret) {
2121                 error_setg_errno(errp, errno, "failed to enable container");
2122                 ret = -errno;
2123                 goto enable_discards_exit;
2124             }
2125         } else {
2126             container->prereg_listener = vfio_prereg_listener;
2127 
2128             memory_listener_register(&container->prereg_listener,
2129                                      &address_space_memory);
2130             if (container->error) {
2131                 memory_listener_unregister(&container->prereg_listener);
2132                 ret = -1;
2133                 error_propagate_prepend(errp, container->error,
2134                     "RAM memory listener initialization failed: ");
2135                 goto enable_discards_exit;
2136             }
2137         }
2138 
2139         info.argsz = sizeof(info);
2140         ret = ioctl(fd, VFIO_IOMMU_SPAPR_TCE_GET_INFO, &info);
2141         if (ret) {
2142             error_setg_errno(errp, errno,
2143                              "VFIO_IOMMU_SPAPR_TCE_GET_INFO failed");
2144             ret = -errno;
2145             if (v2) {
2146                 memory_listener_unregister(&container->prereg_listener);
2147             }
2148             goto enable_discards_exit;
2149         }
2150 
2151         if (v2) {
2152             container->pgsizes = info.ddw.pgsizes;
2153             /*
2154              * There is a default window in just created container.
2155              * To make region_add/del simpler, we better remove this
2156              * window now and let those iommu_listener callbacks
2157              * create/remove them when needed.
2158              */
2159             ret = vfio_spapr_remove_window(container, info.dma32_window_start);
2160             if (ret) {
2161                 error_setg_errno(errp, -ret,
2162                                  "failed to remove existing window");
2163                 goto enable_discards_exit;
2164             }
2165         } else {
2166             /* The default table uses 4K pages */
2167             container->pgsizes = 0x1000;
2168             vfio_host_win_add(container, info.dma32_window_start,
2169                               info.dma32_window_start +
2170                               info.dma32_window_size - 1,
2171                               0x1000);
2172         }
2173     }
2174     }
2175 
2176     vfio_kvm_device_add_group(group);
2177 
2178     QLIST_INIT(&container->group_list);
2179     QLIST_INSERT_HEAD(&space->containers, container, next);
2180 
2181     group->container = container;
2182     QLIST_INSERT_HEAD(&container->group_list, group, container_next);
2183 
2184     container->listener = vfio_memory_listener;
2185 
2186     memory_listener_register(&container->listener, container->space->as);
2187 
2188     if (container->error) {
2189         ret = -1;
2190         error_propagate_prepend(errp, container->error,
2191             "memory listener initialization failed: ");
2192         goto listener_release_exit;
2193     }
2194 
2195     container->initialized = true;
2196 
2197     return 0;
2198 listener_release_exit:
2199     QLIST_REMOVE(group, container_next);
2200     QLIST_REMOVE(container, next);
2201     vfio_kvm_device_del_group(group);
2202     vfio_listener_release(container);
2203 
2204 enable_discards_exit:
2205     vfio_ram_block_discard_disable(container, false);
2206 
2207 free_container_exit:
2208     g_free(container);
2209 
2210 close_fd_exit:
2211     close(fd);
2212 
2213 put_space_exit:
2214     vfio_put_address_space(space);
2215 
2216     return ret;
2217 }
2218 
2219 static void vfio_disconnect_container(VFIOGroup *group)
2220 {
2221     VFIOContainer *container = group->container;
2222 
2223     QLIST_REMOVE(group, container_next);
2224     group->container = NULL;
2225 
2226     /*
2227      * Explicitly release the listener first before unset container,
2228      * since unset may destroy the backend container if it's the last
2229      * group.
2230      */
2231     if (QLIST_EMPTY(&container->group_list)) {
2232         vfio_listener_release(container);
2233     }
2234 
2235     if (ioctl(group->fd, VFIO_GROUP_UNSET_CONTAINER, &container->fd)) {
2236         error_report("vfio: error disconnecting group %d from container",
2237                      group->groupid);
2238     }
2239 
2240     if (QLIST_EMPTY(&container->group_list)) {
2241         VFIOAddressSpace *space = container->space;
2242         VFIOGuestIOMMU *giommu, *tmp;
2243         VFIOHostDMAWindow *hostwin, *next;
2244 
2245         QLIST_REMOVE(container, next);
2246 
2247         QLIST_FOREACH_SAFE(giommu, &container->giommu_list, giommu_next, tmp) {
2248             memory_region_unregister_iommu_notifier(
2249                     MEMORY_REGION(giommu->iommu), &giommu->n);
2250             QLIST_REMOVE(giommu, giommu_next);
2251             g_free(giommu);
2252         }
2253 
2254         QLIST_FOREACH_SAFE(hostwin, &container->hostwin_list, hostwin_next,
2255                            next) {
2256             QLIST_REMOVE(hostwin, hostwin_next);
2257             g_free(hostwin);
2258         }
2259 
2260         trace_vfio_disconnect_container(container->fd);
2261         close(container->fd);
2262         g_free(container);
2263 
2264         vfio_put_address_space(space);
2265     }
2266 }
2267 
2268 VFIOGroup *vfio_get_group(int groupid, AddressSpace *as, Error **errp)
2269 {
2270     VFIOGroup *group;
2271     char path[32];
2272     struct vfio_group_status status = { .argsz = sizeof(status) };
2273 
2274     QLIST_FOREACH(group, &vfio_group_list, next) {
2275         if (group->groupid == groupid) {
2276             /* Found it.  Now is it already in the right context? */
2277             if (group->container->space->as == as) {
2278                 return group;
2279             } else {
2280                 error_setg(errp, "group %d used in multiple address spaces",
2281                            group->groupid);
2282                 return NULL;
2283             }
2284         }
2285     }
2286 
2287     group = g_malloc0(sizeof(*group));
2288 
2289     snprintf(path, sizeof(path), "/dev/vfio/%d", groupid);
2290     group->fd = qemu_open_old(path, O_RDWR);
2291     if (group->fd < 0) {
2292         error_setg_errno(errp, errno, "failed to open %s", path);
2293         goto free_group_exit;
2294     }
2295 
2296     if (ioctl(group->fd, VFIO_GROUP_GET_STATUS, &status)) {
2297         error_setg_errno(errp, errno, "failed to get group %d status", groupid);
2298         goto close_fd_exit;
2299     }
2300 
2301     if (!(status.flags & VFIO_GROUP_FLAGS_VIABLE)) {
2302         error_setg(errp, "group %d is not viable", groupid);
2303         error_append_hint(errp,
2304                           "Please ensure all devices within the iommu_group "
2305                           "are bound to their vfio bus driver.\n");
2306         goto close_fd_exit;
2307     }
2308 
2309     group->groupid = groupid;
2310     QLIST_INIT(&group->device_list);
2311 
2312     if (vfio_connect_container(group, as, errp)) {
2313         error_prepend(errp, "failed to setup container for group %d: ",
2314                       groupid);
2315         goto close_fd_exit;
2316     }
2317 
2318     if (QLIST_EMPTY(&vfio_group_list)) {
2319         qemu_register_reset(vfio_reset_handler, NULL);
2320     }
2321 
2322     QLIST_INSERT_HEAD(&vfio_group_list, group, next);
2323 
2324     return group;
2325 
2326 close_fd_exit:
2327     close(group->fd);
2328 
2329 free_group_exit:
2330     g_free(group);
2331 
2332     return NULL;
2333 }
2334 
2335 void vfio_put_group(VFIOGroup *group)
2336 {
2337     if (!group || !QLIST_EMPTY(&group->device_list)) {
2338         return;
2339     }
2340 
2341     if (!group->ram_block_discard_allowed) {
2342         vfio_ram_block_discard_disable(group->container, false);
2343     }
2344     vfio_kvm_device_del_group(group);
2345     vfio_disconnect_container(group);
2346     QLIST_REMOVE(group, next);
2347     trace_vfio_put_group(group->fd);
2348     close(group->fd);
2349     g_free(group);
2350 
2351     if (QLIST_EMPTY(&vfio_group_list)) {
2352         qemu_unregister_reset(vfio_reset_handler, NULL);
2353     }
2354 }
2355 
2356 int vfio_get_device(VFIOGroup *group, const char *name,
2357                     VFIODevice *vbasedev, Error **errp)
2358 {
2359     struct vfio_device_info dev_info = { .argsz = sizeof(dev_info) };
2360     int ret, fd;
2361 
2362     fd = ioctl(group->fd, VFIO_GROUP_GET_DEVICE_FD, name);
2363     if (fd < 0) {
2364         error_setg_errno(errp, errno, "error getting device from group %d",
2365                          group->groupid);
2366         error_append_hint(errp,
2367                       "Verify all devices in group %d are bound to vfio-<bus> "
2368                       "or pci-stub and not already in use\n", group->groupid);
2369         return fd;
2370     }
2371 
2372     ret = ioctl(fd, VFIO_DEVICE_GET_INFO, &dev_info);
2373     if (ret) {
2374         error_setg_errno(errp, errno, "error getting device info");
2375         close(fd);
2376         return ret;
2377     }
2378 
2379     /*
2380      * Set discarding of RAM as not broken for this group if the driver knows
2381      * the device operates compatibly with discarding.  Setting must be
2382      * consistent per group, but since compatibility is really only possible
2383      * with mdev currently, we expect singleton groups.
2384      */
2385     if (vbasedev->ram_block_discard_allowed !=
2386         group->ram_block_discard_allowed) {
2387         if (!QLIST_EMPTY(&group->device_list)) {
2388             error_setg(errp, "Inconsistent setting of support for discarding "
2389                        "RAM (e.g., balloon) within group");
2390             close(fd);
2391             return -1;
2392         }
2393 
2394         if (!group->ram_block_discard_allowed) {
2395             group->ram_block_discard_allowed = true;
2396             vfio_ram_block_discard_disable(group->container, false);
2397         }
2398     }
2399 
2400     vbasedev->fd = fd;
2401     vbasedev->group = group;
2402     QLIST_INSERT_HEAD(&group->device_list, vbasedev, next);
2403 
2404     vbasedev->num_irqs = dev_info.num_irqs;
2405     vbasedev->num_regions = dev_info.num_regions;
2406     vbasedev->flags = dev_info.flags;
2407 
2408     trace_vfio_get_device(name, dev_info.flags, dev_info.num_regions,
2409                           dev_info.num_irqs);
2410 
2411     vbasedev->reset_works = !!(dev_info.flags & VFIO_DEVICE_FLAGS_RESET);
2412     return 0;
2413 }
2414 
2415 void vfio_put_base_device(VFIODevice *vbasedev)
2416 {
2417     if (!vbasedev->group) {
2418         return;
2419     }
2420     QLIST_REMOVE(vbasedev, next);
2421     vbasedev->group = NULL;
2422     trace_vfio_put_base_device(vbasedev->fd);
2423     close(vbasedev->fd);
2424 }
2425 
2426 int vfio_get_region_info(VFIODevice *vbasedev, int index,
2427                          struct vfio_region_info **info)
2428 {
2429     size_t argsz = sizeof(struct vfio_region_info);
2430 
2431     *info = g_malloc0(argsz);
2432 
2433     (*info)->index = index;
2434 retry:
2435     (*info)->argsz = argsz;
2436 
2437     if (ioctl(vbasedev->fd, VFIO_DEVICE_GET_REGION_INFO, *info)) {
2438         g_free(*info);
2439         *info = NULL;
2440         return -errno;
2441     }
2442 
2443     if ((*info)->argsz > argsz) {
2444         argsz = (*info)->argsz;
2445         *info = g_realloc(*info, argsz);
2446 
2447         goto retry;
2448     }
2449 
2450     return 0;
2451 }
2452 
2453 int vfio_get_dev_region_info(VFIODevice *vbasedev, uint32_t type,
2454                              uint32_t subtype, struct vfio_region_info **info)
2455 {
2456     int i;
2457 
2458     for (i = 0; i < vbasedev->num_regions; i++) {
2459         struct vfio_info_cap_header *hdr;
2460         struct vfio_region_info_cap_type *cap_type;
2461 
2462         if (vfio_get_region_info(vbasedev, i, info)) {
2463             continue;
2464         }
2465 
2466         hdr = vfio_get_region_info_cap(*info, VFIO_REGION_INFO_CAP_TYPE);
2467         if (!hdr) {
2468             g_free(*info);
2469             continue;
2470         }
2471 
2472         cap_type = container_of(hdr, struct vfio_region_info_cap_type, header);
2473 
2474         trace_vfio_get_dev_region(vbasedev->name, i,
2475                                   cap_type->type, cap_type->subtype);
2476 
2477         if (cap_type->type == type && cap_type->subtype == subtype) {
2478             return 0;
2479         }
2480 
2481         g_free(*info);
2482     }
2483 
2484     *info = NULL;
2485     return -ENODEV;
2486 }
2487 
2488 bool vfio_has_region_cap(VFIODevice *vbasedev, int region, uint16_t cap_type)
2489 {
2490     struct vfio_region_info *info = NULL;
2491     bool ret = false;
2492 
2493     if (!vfio_get_region_info(vbasedev, region, &info)) {
2494         if (vfio_get_region_info_cap(info, cap_type)) {
2495             ret = true;
2496         }
2497         g_free(info);
2498     }
2499 
2500     return ret;
2501 }
2502 
2503 /*
2504  * Interfaces for IBM EEH (Enhanced Error Handling)
2505  */
2506 static bool vfio_eeh_container_ok(VFIOContainer *container)
2507 {
2508     /*
2509      * As of 2016-03-04 (linux-4.5) the host kernel EEH/VFIO
2510      * implementation is broken if there are multiple groups in a
2511      * container.  The hardware works in units of Partitionable
2512      * Endpoints (== IOMMU groups) and the EEH operations naively
2513      * iterate across all groups in the container, without any logic
2514      * to make sure the groups have their state synchronized.  For
2515      * certain operations (ENABLE) that might be ok, until an error
2516      * occurs, but for others (GET_STATE) it's clearly broken.
2517      */
2518 
2519     /*
2520      * XXX Once fixed kernels exist, test for them here
2521      */
2522 
2523     if (QLIST_EMPTY(&container->group_list)) {
2524         return false;
2525     }
2526 
2527     if (QLIST_NEXT(QLIST_FIRST(&container->group_list), container_next)) {
2528         return false;
2529     }
2530 
2531     return true;
2532 }
2533 
2534 static int vfio_eeh_container_op(VFIOContainer *container, uint32_t op)
2535 {
2536     struct vfio_eeh_pe_op pe_op = {
2537         .argsz = sizeof(pe_op),
2538         .op = op,
2539     };
2540     int ret;
2541 
2542     if (!vfio_eeh_container_ok(container)) {
2543         error_report("vfio/eeh: EEH_PE_OP 0x%x: "
2544                      "kernel requires a container with exactly one group", op);
2545         return -EPERM;
2546     }
2547 
2548     ret = ioctl(container->fd, VFIO_EEH_PE_OP, &pe_op);
2549     if (ret < 0) {
2550         error_report("vfio/eeh: EEH_PE_OP 0x%x failed: %m", op);
2551         return -errno;
2552     }
2553 
2554     return ret;
2555 }
2556 
2557 static VFIOContainer *vfio_eeh_as_container(AddressSpace *as)
2558 {
2559     VFIOAddressSpace *space = vfio_get_address_space(as);
2560     VFIOContainer *container = NULL;
2561 
2562     if (QLIST_EMPTY(&space->containers)) {
2563         /* No containers to act on */
2564         goto out;
2565     }
2566 
2567     container = QLIST_FIRST(&space->containers);
2568 
2569     if (QLIST_NEXT(container, next)) {
2570         /* We don't yet have logic to synchronize EEH state across
2571          * multiple containers */
2572         container = NULL;
2573         goto out;
2574     }
2575 
2576 out:
2577     vfio_put_address_space(space);
2578     return container;
2579 }
2580 
2581 bool vfio_eeh_as_ok(AddressSpace *as)
2582 {
2583     VFIOContainer *container = vfio_eeh_as_container(as);
2584 
2585     return (container != NULL) && vfio_eeh_container_ok(container);
2586 }
2587 
2588 int vfio_eeh_as_op(AddressSpace *as, uint32_t op)
2589 {
2590     VFIOContainer *container = vfio_eeh_as_container(as);
2591 
2592     if (!container) {
2593         return -ENODEV;
2594     }
2595     return vfio_eeh_container_op(container, op);
2596 }
2597