xref: /openbmc/qemu/hw/vfio/common.c (revision 25657fc6)
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 #include "migration/misc.h"
44 #include "migration/blocker.h"
45 #include "migration/qemu-file.h"
46 #include "sysemu/tpm.h"
47 
48 VFIOGroupList vfio_group_list =
49     QLIST_HEAD_INITIALIZER(vfio_group_list);
50 static QLIST_HEAD(, VFIOAddressSpace) vfio_address_spaces =
51     QLIST_HEAD_INITIALIZER(vfio_address_spaces);
52 
53 #ifdef CONFIG_KVM
54 /*
55  * We have a single VFIO pseudo device per KVM VM.  Once created it lives
56  * for the life of the VM.  Closing the file descriptor only drops our
57  * reference to it and the device's reference to kvm.  Therefore once
58  * initialized, this file descriptor is only released on QEMU exit and
59  * we'll re-use it should another vfio device be attached before then.
60  */
61 static int vfio_kvm_device_fd = -1;
62 #endif
63 
64 /*
65  * Common VFIO interrupt disable
66  */
67 void vfio_disable_irqindex(VFIODevice *vbasedev, int index)
68 {
69     struct vfio_irq_set irq_set = {
70         .argsz = sizeof(irq_set),
71         .flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_TRIGGER,
72         .index = index,
73         .start = 0,
74         .count = 0,
75     };
76 
77     ioctl(vbasedev->fd, VFIO_DEVICE_SET_IRQS, &irq_set);
78 }
79 
80 void vfio_unmask_single_irqindex(VFIODevice *vbasedev, int index)
81 {
82     struct vfio_irq_set irq_set = {
83         .argsz = sizeof(irq_set),
84         .flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_UNMASK,
85         .index = index,
86         .start = 0,
87         .count = 1,
88     };
89 
90     ioctl(vbasedev->fd, VFIO_DEVICE_SET_IRQS, &irq_set);
91 }
92 
93 void vfio_mask_single_irqindex(VFIODevice *vbasedev, int index)
94 {
95     struct vfio_irq_set irq_set = {
96         .argsz = sizeof(irq_set),
97         .flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_MASK,
98         .index = index,
99         .start = 0,
100         .count = 1,
101     };
102 
103     ioctl(vbasedev->fd, VFIO_DEVICE_SET_IRQS, &irq_set);
104 }
105 
106 static inline const char *action_to_str(int action)
107 {
108     switch (action) {
109     case VFIO_IRQ_SET_ACTION_MASK:
110         return "MASK";
111     case VFIO_IRQ_SET_ACTION_UNMASK:
112         return "UNMASK";
113     case VFIO_IRQ_SET_ACTION_TRIGGER:
114         return "TRIGGER";
115     default:
116         return "UNKNOWN ACTION";
117     }
118 }
119 
120 static const char *index_to_str(VFIODevice *vbasedev, int index)
121 {
122     if (vbasedev->type != VFIO_DEVICE_TYPE_PCI) {
123         return NULL;
124     }
125 
126     switch (index) {
127     case VFIO_PCI_INTX_IRQ_INDEX:
128         return "INTX";
129     case VFIO_PCI_MSI_IRQ_INDEX:
130         return "MSI";
131     case VFIO_PCI_MSIX_IRQ_INDEX:
132         return "MSIX";
133     case VFIO_PCI_ERR_IRQ_INDEX:
134         return "ERR";
135     case VFIO_PCI_REQ_IRQ_INDEX:
136         return "REQ";
137     default:
138         return NULL;
139     }
140 }
141 
142 static int vfio_ram_block_discard_disable(VFIOContainer *container, bool state)
143 {
144     switch (container->iommu_type) {
145     case VFIO_TYPE1v2_IOMMU:
146     case VFIO_TYPE1_IOMMU:
147         /*
148          * We support coordinated discarding of RAM via the RamDiscardManager.
149          */
150         return ram_block_uncoordinated_discard_disable(state);
151     default:
152         /*
153          * VFIO_SPAPR_TCE_IOMMU most probably works just fine with
154          * RamDiscardManager, however, it is completely untested.
155          *
156          * VFIO_SPAPR_TCE_v2_IOMMU with "DMA memory preregistering" does
157          * completely the opposite of managing mapping/pinning dynamically as
158          * required by RamDiscardManager. We would have to special-case sections
159          * with a RamDiscardManager.
160          */
161         return ram_block_discard_disable(state);
162     }
163 }
164 
165 int vfio_set_irq_signaling(VFIODevice *vbasedev, int index, int subindex,
166                            int action, int fd, Error **errp)
167 {
168     struct vfio_irq_set *irq_set;
169     int argsz, ret = 0;
170     const char *name;
171     int32_t *pfd;
172 
173     argsz = sizeof(*irq_set) + sizeof(*pfd);
174 
175     irq_set = g_malloc0(argsz);
176     irq_set->argsz = argsz;
177     irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD | action;
178     irq_set->index = index;
179     irq_set->start = subindex;
180     irq_set->count = 1;
181     pfd = (int32_t *)&irq_set->data;
182     *pfd = fd;
183 
184     if (ioctl(vbasedev->fd, VFIO_DEVICE_SET_IRQS, irq_set)) {
185         ret = -errno;
186     }
187     g_free(irq_set);
188 
189     if (!ret) {
190         return 0;
191     }
192 
193     error_setg_errno(errp, -ret, "VFIO_DEVICE_SET_IRQS failure");
194 
195     name = index_to_str(vbasedev, index);
196     if (name) {
197         error_prepend(errp, "%s-%d: ", name, subindex);
198     } else {
199         error_prepend(errp, "index %d-%d: ", index, subindex);
200     }
201     error_prepend(errp,
202                   "Failed to %s %s eventfd signaling for interrupt ",
203                   fd < 0 ? "tear down" : "set up", action_to_str(action));
204     return ret;
205 }
206 
207 /*
208  * IO Port/MMIO - Beware of the endians, VFIO is always little endian
209  */
210 void vfio_region_write(void *opaque, hwaddr addr,
211                        uint64_t data, unsigned size)
212 {
213     VFIORegion *region = opaque;
214     VFIODevice *vbasedev = region->vbasedev;
215     union {
216         uint8_t byte;
217         uint16_t word;
218         uint32_t dword;
219         uint64_t qword;
220     } buf;
221 
222     switch (size) {
223     case 1:
224         buf.byte = data;
225         break;
226     case 2:
227         buf.word = cpu_to_le16(data);
228         break;
229     case 4:
230         buf.dword = cpu_to_le32(data);
231         break;
232     case 8:
233         buf.qword = cpu_to_le64(data);
234         break;
235     default:
236         hw_error("vfio: unsupported write size, %u bytes", size);
237         break;
238     }
239 
240     if (pwrite(vbasedev->fd, &buf, size, region->fd_offset + addr) != size) {
241         error_report("%s(%s:region%d+0x%"HWADDR_PRIx", 0x%"PRIx64
242                      ",%d) failed: %m",
243                      __func__, vbasedev->name, region->nr,
244                      addr, data, size);
245     }
246 
247     trace_vfio_region_write(vbasedev->name, region->nr, addr, data, size);
248 
249     /*
250      * A read or write to a BAR always signals an INTx EOI.  This will
251      * do nothing if not pending (including not in INTx mode).  We assume
252      * that a BAR access is in response to an interrupt and that BAR
253      * accesses will service the interrupt.  Unfortunately, we don't know
254      * which access will service the interrupt, so we're potentially
255      * getting quite a few host interrupts per guest interrupt.
256      */
257     vbasedev->ops->vfio_eoi(vbasedev);
258 }
259 
260 uint64_t vfio_region_read(void *opaque,
261                           hwaddr addr, unsigned size)
262 {
263     VFIORegion *region = opaque;
264     VFIODevice *vbasedev = region->vbasedev;
265     union {
266         uint8_t byte;
267         uint16_t word;
268         uint32_t dword;
269         uint64_t qword;
270     } buf;
271     uint64_t data = 0;
272 
273     if (pread(vbasedev->fd, &buf, size, region->fd_offset + addr) != size) {
274         error_report("%s(%s:region%d+0x%"HWADDR_PRIx", %d) failed: %m",
275                      __func__, vbasedev->name, region->nr,
276                      addr, size);
277         return (uint64_t)-1;
278     }
279     switch (size) {
280     case 1:
281         data = buf.byte;
282         break;
283     case 2:
284         data = le16_to_cpu(buf.word);
285         break;
286     case 4:
287         data = le32_to_cpu(buf.dword);
288         break;
289     case 8:
290         data = le64_to_cpu(buf.qword);
291         break;
292     default:
293         hw_error("vfio: unsupported read size, %u bytes", size);
294         break;
295     }
296 
297     trace_vfio_region_read(vbasedev->name, region->nr, addr, size, data);
298 
299     /* Same as write above */
300     vbasedev->ops->vfio_eoi(vbasedev);
301 
302     return data;
303 }
304 
305 const MemoryRegionOps vfio_region_ops = {
306     .read = vfio_region_read,
307     .write = vfio_region_write,
308     .endianness = DEVICE_LITTLE_ENDIAN,
309     .valid = {
310         .min_access_size = 1,
311         .max_access_size = 8,
312     },
313     .impl = {
314         .min_access_size = 1,
315         .max_access_size = 8,
316     },
317 };
318 
319 /*
320  * Device state interfaces
321  */
322 
323 typedef struct {
324     unsigned long *bitmap;
325     hwaddr size;
326     hwaddr pages;
327 } VFIOBitmap;
328 
329 static int vfio_bitmap_alloc(VFIOBitmap *vbmap, hwaddr size)
330 {
331     vbmap->pages = REAL_HOST_PAGE_ALIGN(size) / qemu_real_host_page_size();
332     vbmap->size = ROUND_UP(vbmap->pages, sizeof(__u64) * BITS_PER_BYTE) /
333                                          BITS_PER_BYTE;
334     vbmap->bitmap = g_try_malloc0(vbmap->size);
335     if (!vbmap->bitmap) {
336         return -ENOMEM;
337     }
338 
339     return 0;
340 }
341 
342 static int vfio_get_dirty_bitmap(VFIOContainer *container, uint64_t iova,
343                                  uint64_t size, ram_addr_t ram_addr);
344 
345 bool vfio_mig_active(void)
346 {
347     VFIOGroup *group;
348     VFIODevice *vbasedev;
349 
350     if (QLIST_EMPTY(&vfio_group_list)) {
351         return false;
352     }
353 
354     QLIST_FOREACH(group, &vfio_group_list, next) {
355         QLIST_FOREACH(vbasedev, &group->device_list, next) {
356             if (vbasedev->migration_blocker) {
357                 return false;
358             }
359         }
360     }
361     return true;
362 }
363 
364 static Error *multiple_devices_migration_blocker;
365 static Error *giommu_migration_blocker;
366 
367 static unsigned int vfio_migratable_device_num(void)
368 {
369     VFIOGroup *group;
370     VFIODevice *vbasedev;
371     unsigned int device_num = 0;
372 
373     QLIST_FOREACH(group, &vfio_group_list, next) {
374         QLIST_FOREACH(vbasedev, &group->device_list, next) {
375             if (vbasedev->migration) {
376                 device_num++;
377             }
378         }
379     }
380 
381     return device_num;
382 }
383 
384 int vfio_block_multiple_devices_migration(Error **errp)
385 {
386     int ret;
387 
388     if (multiple_devices_migration_blocker ||
389         vfio_migratable_device_num() <= 1) {
390         return 0;
391     }
392 
393     error_setg(&multiple_devices_migration_blocker,
394                "Migration is currently not supported with multiple "
395                "VFIO devices");
396     ret = migrate_add_blocker(multiple_devices_migration_blocker, errp);
397     if (ret < 0) {
398         error_free(multiple_devices_migration_blocker);
399         multiple_devices_migration_blocker = NULL;
400     }
401 
402     return ret;
403 }
404 
405 void vfio_unblock_multiple_devices_migration(void)
406 {
407     if (!multiple_devices_migration_blocker ||
408         vfio_migratable_device_num() > 1) {
409         return;
410     }
411 
412     migrate_del_blocker(multiple_devices_migration_blocker);
413     error_free(multiple_devices_migration_blocker);
414     multiple_devices_migration_blocker = NULL;
415 }
416 
417 static bool vfio_viommu_preset(void)
418 {
419     VFIOAddressSpace *space;
420 
421     QLIST_FOREACH(space, &vfio_address_spaces, list) {
422         if (space->as != &address_space_memory) {
423             return true;
424         }
425     }
426 
427     return false;
428 }
429 
430 int vfio_block_giommu_migration(Error **errp)
431 {
432     int ret;
433 
434     if (giommu_migration_blocker ||
435         !vfio_viommu_preset()) {
436         return 0;
437     }
438 
439     error_setg(&giommu_migration_blocker,
440                "Migration is currently not supported with vIOMMU enabled");
441     ret = migrate_add_blocker(giommu_migration_blocker, errp);
442     if (ret < 0) {
443         error_free(giommu_migration_blocker);
444         giommu_migration_blocker = NULL;
445     }
446 
447     return ret;
448 }
449 
450 void vfio_migration_finalize(void)
451 {
452     if (!giommu_migration_blocker ||
453         vfio_viommu_preset()) {
454         return;
455     }
456 
457     migrate_del_blocker(giommu_migration_blocker);
458     error_free(giommu_migration_blocker);
459     giommu_migration_blocker = NULL;
460 }
461 
462 static void vfio_set_migration_error(int err)
463 {
464     MigrationState *ms = migrate_get_current();
465 
466     if (migration_is_setup_or_active(ms->state)) {
467         WITH_QEMU_LOCK_GUARD(&ms->qemu_file_lock) {
468             if (ms->to_dst_file) {
469                 qemu_file_set_error(ms->to_dst_file, err);
470             }
471         }
472     }
473 }
474 
475 static bool vfio_devices_all_dirty_tracking(VFIOContainer *container)
476 {
477     VFIOGroup *group;
478     VFIODevice *vbasedev;
479     MigrationState *ms = migrate_get_current();
480 
481     if (!migration_is_setup_or_active(ms->state)) {
482         return false;
483     }
484 
485     QLIST_FOREACH(group, &container->group_list, container_next) {
486         QLIST_FOREACH(vbasedev, &group->device_list, next) {
487             VFIOMigration *migration = vbasedev->migration;
488 
489             if (!migration) {
490                 return false;
491             }
492 
493             if (vbasedev->pre_copy_dirty_page_tracking == ON_OFF_AUTO_OFF &&
494                 migration->device_state == VFIO_DEVICE_STATE_RUNNING) {
495                 return false;
496             }
497         }
498     }
499     return true;
500 }
501 
502 static bool vfio_devices_all_device_dirty_tracking(VFIOContainer *container)
503 {
504     VFIOGroup *group;
505     VFIODevice *vbasedev;
506 
507     QLIST_FOREACH(group, &container->group_list, container_next) {
508         QLIST_FOREACH(vbasedev, &group->device_list, next) {
509             if (!vbasedev->dirty_pages_supported) {
510                 return false;
511             }
512         }
513     }
514 
515     return true;
516 }
517 
518 /*
519  * Check if all VFIO devices are running and migration is active, which is
520  * essentially equivalent to the migration being in pre-copy phase.
521  */
522 static bool vfio_devices_all_running_and_mig_active(VFIOContainer *container)
523 {
524     VFIOGroup *group;
525     VFIODevice *vbasedev;
526 
527     if (!migration_is_active(migrate_get_current())) {
528         return false;
529     }
530 
531     QLIST_FOREACH(group, &container->group_list, container_next) {
532         QLIST_FOREACH(vbasedev, &group->device_list, next) {
533             VFIOMigration *migration = vbasedev->migration;
534 
535             if (!migration) {
536                 return false;
537             }
538 
539             if (migration->device_state == VFIO_DEVICE_STATE_RUNNING) {
540                 continue;
541             } else {
542                 return false;
543             }
544         }
545     }
546     return true;
547 }
548 
549 static int vfio_dma_unmap_bitmap(VFIOContainer *container,
550                                  hwaddr iova, ram_addr_t size,
551                                  IOMMUTLBEntry *iotlb)
552 {
553     struct vfio_iommu_type1_dma_unmap *unmap;
554     struct vfio_bitmap *bitmap;
555     VFIOBitmap vbmap;
556     int ret;
557 
558     ret = vfio_bitmap_alloc(&vbmap, size);
559     if (ret) {
560         return ret;
561     }
562 
563     unmap = g_malloc0(sizeof(*unmap) + sizeof(*bitmap));
564 
565     unmap->argsz = sizeof(*unmap) + sizeof(*bitmap);
566     unmap->iova = iova;
567     unmap->size = size;
568     unmap->flags |= VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP;
569     bitmap = (struct vfio_bitmap *)&unmap->data;
570 
571     /*
572      * cpu_physical_memory_set_dirty_lebitmap() supports pages in bitmap of
573      * qemu_real_host_page_size to mark those dirty. Hence set bitmap_pgsize
574      * to qemu_real_host_page_size.
575      */
576     bitmap->pgsize = qemu_real_host_page_size();
577     bitmap->size = vbmap.size;
578     bitmap->data = (__u64 *)vbmap.bitmap;
579 
580     if (vbmap.size > container->max_dirty_bitmap_size) {
581         error_report("UNMAP: Size of bitmap too big 0x%"PRIx64, vbmap.size);
582         ret = -E2BIG;
583         goto unmap_exit;
584     }
585 
586     ret = ioctl(container->fd, VFIO_IOMMU_UNMAP_DMA, unmap);
587     if (!ret) {
588         cpu_physical_memory_set_dirty_lebitmap(vbmap.bitmap,
589                 iotlb->translated_addr, vbmap.pages);
590     } else {
591         error_report("VFIO_UNMAP_DMA with DIRTY_BITMAP : %m");
592     }
593 
594 unmap_exit:
595     g_free(unmap);
596     g_free(vbmap.bitmap);
597 
598     return ret;
599 }
600 
601 /*
602  * DMA - Mapping and unmapping for the "type1" IOMMU interface used on x86
603  */
604 static int vfio_dma_unmap(VFIOContainer *container,
605                           hwaddr iova, ram_addr_t size,
606                           IOMMUTLBEntry *iotlb)
607 {
608     struct vfio_iommu_type1_dma_unmap unmap = {
609         .argsz = sizeof(unmap),
610         .flags = 0,
611         .iova = iova,
612         .size = size,
613     };
614     bool need_dirty_sync = false;
615     int ret;
616 
617     if (iotlb && vfio_devices_all_running_and_mig_active(container)) {
618         if (!vfio_devices_all_device_dirty_tracking(container) &&
619             container->dirty_pages_supported) {
620             return vfio_dma_unmap_bitmap(container, iova, size, iotlb);
621         }
622 
623         need_dirty_sync = true;
624     }
625 
626     while (ioctl(container->fd, VFIO_IOMMU_UNMAP_DMA, &unmap)) {
627         /*
628          * The type1 backend has an off-by-one bug in the kernel (71a7d3d78e3c
629          * v4.15) where an overflow in its wrap-around check prevents us from
630          * unmapping the last page of the address space.  Test for the error
631          * condition and re-try the unmap excluding the last page.  The
632          * expectation is that we've never mapped the last page anyway and this
633          * unmap request comes via vIOMMU support which also makes it unlikely
634          * that this page is used.  This bug was introduced well after type1 v2
635          * support was introduced, so we shouldn't need to test for v1.  A fix
636          * is queued for kernel v5.0 so this workaround can be removed once
637          * affected kernels are sufficiently deprecated.
638          */
639         if (errno == EINVAL && unmap.size && !(unmap.iova + unmap.size) &&
640             container->iommu_type == VFIO_TYPE1v2_IOMMU) {
641             trace_vfio_dma_unmap_overflow_workaround();
642             unmap.size -= 1ULL << ctz64(container->pgsizes);
643             continue;
644         }
645         error_report("VFIO_UNMAP_DMA failed: %s", strerror(errno));
646         return -errno;
647     }
648 
649     if (need_dirty_sync) {
650         ret = vfio_get_dirty_bitmap(container, iova, size,
651                                     iotlb->translated_addr);
652         if (ret) {
653             return ret;
654         }
655     }
656 
657     return 0;
658 }
659 
660 static int vfio_dma_map(VFIOContainer *container, hwaddr iova,
661                         ram_addr_t size, void *vaddr, bool readonly)
662 {
663     struct vfio_iommu_type1_dma_map map = {
664         .argsz = sizeof(map),
665         .flags = VFIO_DMA_MAP_FLAG_READ,
666         .vaddr = (__u64)(uintptr_t)vaddr,
667         .iova = iova,
668         .size = size,
669     };
670 
671     if (!readonly) {
672         map.flags |= VFIO_DMA_MAP_FLAG_WRITE;
673     }
674 
675     /*
676      * Try the mapping, if it fails with EBUSY, unmap the region and try
677      * again.  This shouldn't be necessary, but we sometimes see it in
678      * the VGA ROM space.
679      */
680     if (ioctl(container->fd, VFIO_IOMMU_MAP_DMA, &map) == 0 ||
681         (errno == EBUSY && vfio_dma_unmap(container, iova, size, NULL) == 0 &&
682          ioctl(container->fd, VFIO_IOMMU_MAP_DMA, &map) == 0)) {
683         return 0;
684     }
685 
686     error_report("VFIO_MAP_DMA failed: %s", strerror(errno));
687     return -errno;
688 }
689 
690 static void vfio_host_win_add(VFIOContainer *container,
691                               hwaddr min_iova, hwaddr max_iova,
692                               uint64_t iova_pgsizes)
693 {
694     VFIOHostDMAWindow *hostwin;
695 
696     QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) {
697         if (ranges_overlap(hostwin->min_iova,
698                            hostwin->max_iova - hostwin->min_iova + 1,
699                            min_iova,
700                            max_iova - min_iova + 1)) {
701             hw_error("%s: Overlapped IOMMU are not enabled", __func__);
702         }
703     }
704 
705     hostwin = g_malloc0(sizeof(*hostwin));
706 
707     hostwin->min_iova = min_iova;
708     hostwin->max_iova = max_iova;
709     hostwin->iova_pgsizes = iova_pgsizes;
710     QLIST_INSERT_HEAD(&container->hostwin_list, hostwin, hostwin_next);
711 }
712 
713 static int vfio_host_win_del(VFIOContainer *container, hwaddr min_iova,
714                              hwaddr max_iova)
715 {
716     VFIOHostDMAWindow *hostwin;
717 
718     QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) {
719         if (hostwin->min_iova == min_iova && hostwin->max_iova == max_iova) {
720             QLIST_REMOVE(hostwin, hostwin_next);
721             g_free(hostwin);
722             return 0;
723         }
724     }
725 
726     return -1;
727 }
728 
729 static bool vfio_listener_skipped_section(MemoryRegionSection *section)
730 {
731     return (!memory_region_is_ram(section->mr) &&
732             !memory_region_is_iommu(section->mr)) ||
733            memory_region_is_protected(section->mr) ||
734            /*
735             * Sizing an enabled 64-bit BAR can cause spurious mappings to
736             * addresses in the upper part of the 64-bit address space.  These
737             * are never accessed by the CPU and beyond the address width of
738             * some IOMMU hardware.  TODO: VFIO should tell us the IOMMU width.
739             */
740            section->offset_within_address_space & (1ULL << 63);
741 }
742 
743 /* Called with rcu_read_lock held.  */
744 static bool vfio_get_xlat_addr(IOMMUTLBEntry *iotlb, void **vaddr,
745                                ram_addr_t *ram_addr, bool *read_only)
746 {
747     bool ret, mr_has_discard_manager;
748 
749     ret = memory_get_xlat_addr(iotlb, vaddr, ram_addr, read_only,
750                                &mr_has_discard_manager);
751     if (ret && mr_has_discard_manager) {
752         /*
753          * Malicious VMs might trigger discarding of IOMMU-mapped memory. The
754          * pages will remain pinned inside vfio until unmapped, resulting in a
755          * higher memory consumption than expected. If memory would get
756          * populated again later, there would be an inconsistency between pages
757          * pinned by vfio and pages seen by QEMU. This is the case until
758          * unmapped from the IOMMU (e.g., during device reset).
759          *
760          * With malicious guests, we really only care about pinning more memory
761          * than expected. RLIMIT_MEMLOCK set for the user/process can never be
762          * exceeded and can be used to mitigate this problem.
763          */
764         warn_report_once("Using vfio with vIOMMUs and coordinated discarding of"
765                          " RAM (e.g., virtio-mem) works, however, malicious"
766                          " guests can trigger pinning of more memory than"
767                          " intended via an IOMMU. It's possible to mitigate "
768                          " by setting/adjusting RLIMIT_MEMLOCK.");
769     }
770     return ret;
771 }
772 
773 static void vfio_iommu_map_notify(IOMMUNotifier *n, IOMMUTLBEntry *iotlb)
774 {
775     VFIOGuestIOMMU *giommu = container_of(n, VFIOGuestIOMMU, n);
776     VFIOContainer *container = giommu->container;
777     hwaddr iova = iotlb->iova + giommu->iommu_offset;
778     void *vaddr;
779     int ret;
780 
781     trace_vfio_iommu_map_notify(iotlb->perm == IOMMU_NONE ? "UNMAP" : "MAP",
782                                 iova, iova + iotlb->addr_mask);
783 
784     if (iotlb->target_as != &address_space_memory) {
785         error_report("Wrong target AS \"%s\", only system memory is allowed",
786                      iotlb->target_as->name ? iotlb->target_as->name : "none");
787         vfio_set_migration_error(-EINVAL);
788         return;
789     }
790 
791     rcu_read_lock();
792 
793     if ((iotlb->perm & IOMMU_RW) != IOMMU_NONE) {
794         bool read_only;
795 
796         if (!vfio_get_xlat_addr(iotlb, &vaddr, NULL, &read_only)) {
797             goto out;
798         }
799         /*
800          * vaddr is only valid until rcu_read_unlock(). But after
801          * vfio_dma_map has set up the mapping the pages will be
802          * pinned by the kernel. This makes sure that the RAM backend
803          * of vaddr will always be there, even if the memory object is
804          * destroyed and its backing memory munmap-ed.
805          */
806         ret = vfio_dma_map(container, iova,
807                            iotlb->addr_mask + 1, vaddr,
808                            read_only);
809         if (ret) {
810             error_report("vfio_dma_map(%p, 0x%"HWADDR_PRIx", "
811                          "0x%"HWADDR_PRIx", %p) = %d (%s)",
812                          container, iova,
813                          iotlb->addr_mask + 1, vaddr, ret, strerror(-ret));
814         }
815     } else {
816         ret = vfio_dma_unmap(container, iova, iotlb->addr_mask + 1, iotlb);
817         if (ret) {
818             error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx", "
819                          "0x%"HWADDR_PRIx") = %d (%s)",
820                          container, iova,
821                          iotlb->addr_mask + 1, ret, strerror(-ret));
822             vfio_set_migration_error(ret);
823         }
824     }
825 out:
826     rcu_read_unlock();
827 }
828 
829 static void vfio_ram_discard_notify_discard(RamDiscardListener *rdl,
830                                             MemoryRegionSection *section)
831 {
832     VFIORamDiscardListener *vrdl = container_of(rdl, VFIORamDiscardListener,
833                                                 listener);
834     const hwaddr size = int128_get64(section->size);
835     const hwaddr iova = section->offset_within_address_space;
836     int ret;
837 
838     /* Unmap with a single call. */
839     ret = vfio_dma_unmap(vrdl->container, iova, size , NULL);
840     if (ret) {
841         error_report("%s: vfio_dma_unmap() failed: %s", __func__,
842                      strerror(-ret));
843     }
844 }
845 
846 static int vfio_ram_discard_notify_populate(RamDiscardListener *rdl,
847                                             MemoryRegionSection *section)
848 {
849     VFIORamDiscardListener *vrdl = container_of(rdl, VFIORamDiscardListener,
850                                                 listener);
851     const hwaddr end = section->offset_within_region +
852                        int128_get64(section->size);
853     hwaddr start, next, iova;
854     void *vaddr;
855     int ret;
856 
857     /*
858      * Map in (aligned within memory region) minimum granularity, so we can
859      * unmap in minimum granularity later.
860      */
861     for (start = section->offset_within_region; start < end; start = next) {
862         next = ROUND_UP(start + 1, vrdl->granularity);
863         next = MIN(next, end);
864 
865         iova = start - section->offset_within_region +
866                section->offset_within_address_space;
867         vaddr = memory_region_get_ram_ptr(section->mr) + start;
868 
869         ret = vfio_dma_map(vrdl->container, iova, next - start,
870                            vaddr, section->readonly);
871         if (ret) {
872             /* Rollback */
873             vfio_ram_discard_notify_discard(rdl, section);
874             return ret;
875         }
876     }
877     return 0;
878 }
879 
880 static void vfio_register_ram_discard_listener(VFIOContainer *container,
881                                                MemoryRegionSection *section)
882 {
883     RamDiscardManager *rdm = memory_region_get_ram_discard_manager(section->mr);
884     VFIORamDiscardListener *vrdl;
885 
886     /* Ignore some corner cases not relevant in practice. */
887     g_assert(QEMU_IS_ALIGNED(section->offset_within_region, TARGET_PAGE_SIZE));
888     g_assert(QEMU_IS_ALIGNED(section->offset_within_address_space,
889                              TARGET_PAGE_SIZE));
890     g_assert(QEMU_IS_ALIGNED(int128_get64(section->size), TARGET_PAGE_SIZE));
891 
892     vrdl = g_new0(VFIORamDiscardListener, 1);
893     vrdl->container = container;
894     vrdl->mr = section->mr;
895     vrdl->offset_within_address_space = section->offset_within_address_space;
896     vrdl->size = int128_get64(section->size);
897     vrdl->granularity = ram_discard_manager_get_min_granularity(rdm,
898                                                                 section->mr);
899 
900     g_assert(vrdl->granularity && is_power_of_2(vrdl->granularity));
901     g_assert(container->pgsizes &&
902              vrdl->granularity >= 1ULL << ctz64(container->pgsizes));
903 
904     ram_discard_listener_init(&vrdl->listener,
905                               vfio_ram_discard_notify_populate,
906                               vfio_ram_discard_notify_discard, true);
907     ram_discard_manager_register_listener(rdm, &vrdl->listener, section);
908     QLIST_INSERT_HEAD(&container->vrdl_list, vrdl, next);
909 
910     /*
911      * Sanity-check if we have a theoretically problematic setup where we could
912      * exceed the maximum number of possible DMA mappings over time. We assume
913      * that each mapped section in the same address space as a RamDiscardManager
914      * section consumes exactly one DMA mapping, with the exception of
915      * RamDiscardManager sections; i.e., we don't expect to have gIOMMU sections
916      * in the same address space as RamDiscardManager sections.
917      *
918      * We assume that each section in the address space consumes one memslot.
919      * We take the number of KVM memory slots as a best guess for the maximum
920      * number of sections in the address space we could have over time,
921      * also consuming DMA mappings.
922      */
923     if (container->dma_max_mappings) {
924         unsigned int vrdl_count = 0, vrdl_mappings = 0, max_memslots = 512;
925 
926 #ifdef CONFIG_KVM
927         if (kvm_enabled()) {
928             max_memslots = kvm_get_max_memslots();
929         }
930 #endif
931 
932         QLIST_FOREACH(vrdl, &container->vrdl_list, next) {
933             hwaddr start, end;
934 
935             start = QEMU_ALIGN_DOWN(vrdl->offset_within_address_space,
936                                     vrdl->granularity);
937             end = ROUND_UP(vrdl->offset_within_address_space + vrdl->size,
938                            vrdl->granularity);
939             vrdl_mappings += (end - start) / vrdl->granularity;
940             vrdl_count++;
941         }
942 
943         if (vrdl_mappings + max_memslots - vrdl_count >
944             container->dma_max_mappings) {
945             warn_report("%s: possibly running out of DMA mappings. E.g., try"
946                         " increasing the 'block-size' of virtio-mem devies."
947                         " Maximum possible DMA mappings: %d, Maximum possible"
948                         " memslots: %d", __func__, container->dma_max_mappings,
949                         max_memslots);
950         }
951     }
952 }
953 
954 static void vfio_unregister_ram_discard_listener(VFIOContainer *container,
955                                                  MemoryRegionSection *section)
956 {
957     RamDiscardManager *rdm = memory_region_get_ram_discard_manager(section->mr);
958     VFIORamDiscardListener *vrdl = NULL;
959 
960     QLIST_FOREACH(vrdl, &container->vrdl_list, next) {
961         if (vrdl->mr == section->mr &&
962             vrdl->offset_within_address_space ==
963             section->offset_within_address_space) {
964             break;
965         }
966     }
967 
968     if (!vrdl) {
969         hw_error("vfio: Trying to unregister missing RAM discard listener");
970     }
971 
972     ram_discard_manager_unregister_listener(rdm, &vrdl->listener);
973     QLIST_REMOVE(vrdl, next);
974     g_free(vrdl);
975 }
976 
977 static VFIOHostDMAWindow *vfio_find_hostwin(VFIOContainer *container,
978                                             hwaddr iova, hwaddr end)
979 {
980     VFIOHostDMAWindow *hostwin;
981     bool hostwin_found = false;
982 
983     QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) {
984         if (hostwin->min_iova <= iova && end <= hostwin->max_iova) {
985             hostwin_found = true;
986             break;
987         }
988     }
989 
990     return hostwin_found ? hostwin : NULL;
991 }
992 
993 static bool vfio_known_safe_misalignment(MemoryRegionSection *section)
994 {
995     MemoryRegion *mr = section->mr;
996 
997     if (!TPM_IS_CRB(mr->owner)) {
998         return false;
999     }
1000 
1001     /* this is a known safe misaligned region, just trace for debug purpose */
1002     trace_vfio_known_safe_misalignment(memory_region_name(mr),
1003                                        section->offset_within_address_space,
1004                                        section->offset_within_region,
1005                                        qemu_real_host_page_size());
1006     return true;
1007 }
1008 
1009 static bool vfio_listener_valid_section(MemoryRegionSection *section,
1010                                         const char *name)
1011 {
1012     if (vfio_listener_skipped_section(section)) {
1013         trace_vfio_listener_region_skip(name,
1014                 section->offset_within_address_space,
1015                 section->offset_within_address_space +
1016                 int128_get64(int128_sub(section->size, int128_one())));
1017         return false;
1018     }
1019 
1020     if (unlikely((section->offset_within_address_space &
1021                   ~qemu_real_host_page_mask()) !=
1022                  (section->offset_within_region & ~qemu_real_host_page_mask()))) {
1023         if (!vfio_known_safe_misalignment(section)) {
1024             error_report("%s received unaligned region %s iova=0x%"PRIx64
1025                          " offset_within_region=0x%"PRIx64
1026                          " qemu_real_host_page_size=0x%"PRIxPTR,
1027                          __func__, memory_region_name(section->mr),
1028                          section->offset_within_address_space,
1029                          section->offset_within_region,
1030                          qemu_real_host_page_size());
1031         }
1032         return false;
1033     }
1034 
1035     return true;
1036 }
1037 
1038 static bool vfio_get_section_iova_range(VFIOContainer *container,
1039                                         MemoryRegionSection *section,
1040                                         hwaddr *out_iova, hwaddr *out_end,
1041                                         Int128 *out_llend)
1042 {
1043     Int128 llend;
1044     hwaddr iova;
1045 
1046     iova = REAL_HOST_PAGE_ALIGN(section->offset_within_address_space);
1047     llend = int128_make64(section->offset_within_address_space);
1048     llend = int128_add(llend, section->size);
1049     llend = int128_and(llend, int128_exts64(qemu_real_host_page_mask()));
1050 
1051     if (int128_ge(int128_make64(iova), llend)) {
1052         return false;
1053     }
1054 
1055     *out_iova = iova;
1056     *out_end = int128_get64(int128_sub(llend, int128_one()));
1057     if (out_llend) {
1058         *out_llend = llend;
1059     }
1060     return true;
1061 }
1062 
1063 static void vfio_listener_region_add(MemoryListener *listener,
1064                                      MemoryRegionSection *section)
1065 {
1066     VFIOContainer *container = container_of(listener, VFIOContainer, listener);
1067     hwaddr iova, end;
1068     Int128 llend, llsize;
1069     void *vaddr;
1070     int ret;
1071     VFIOHostDMAWindow *hostwin;
1072     Error *err = NULL;
1073 
1074     if (!vfio_listener_valid_section(section, "region_add")) {
1075         return;
1076     }
1077 
1078     if (!vfio_get_section_iova_range(container, section, &iova, &end, &llend)) {
1079         if (memory_region_is_ram_device(section->mr)) {
1080             trace_vfio_listener_region_add_no_dma_map(
1081                 memory_region_name(section->mr),
1082                 section->offset_within_address_space,
1083                 int128_getlo(section->size),
1084                 qemu_real_host_page_size());
1085         }
1086         return;
1087     }
1088 
1089     if (container->iommu_type == VFIO_SPAPR_TCE_v2_IOMMU) {
1090         hwaddr pgsize = 0;
1091 
1092         /* For now intersections are not allowed, we may relax this later */
1093         QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) {
1094             if (ranges_overlap(hostwin->min_iova,
1095                                hostwin->max_iova - hostwin->min_iova + 1,
1096                                section->offset_within_address_space,
1097                                int128_get64(section->size))) {
1098                 error_setg(&err,
1099                     "region [0x%"PRIx64",0x%"PRIx64"] overlaps with existing"
1100                     "host DMA window [0x%"PRIx64",0x%"PRIx64"]",
1101                     section->offset_within_address_space,
1102                     section->offset_within_address_space +
1103                         int128_get64(section->size) - 1,
1104                     hostwin->min_iova, hostwin->max_iova);
1105                 goto fail;
1106             }
1107         }
1108 
1109         ret = vfio_spapr_create_window(container, section, &pgsize);
1110         if (ret) {
1111             error_setg_errno(&err, -ret, "Failed to create SPAPR window");
1112             goto fail;
1113         }
1114 
1115         vfio_host_win_add(container, section->offset_within_address_space,
1116                           section->offset_within_address_space +
1117                           int128_get64(section->size) - 1, pgsize);
1118 #ifdef CONFIG_KVM
1119         if (kvm_enabled()) {
1120             VFIOGroup *group;
1121             IOMMUMemoryRegion *iommu_mr = IOMMU_MEMORY_REGION(section->mr);
1122             struct kvm_vfio_spapr_tce param;
1123             struct kvm_device_attr attr = {
1124                 .group = KVM_DEV_VFIO_GROUP,
1125                 .attr = KVM_DEV_VFIO_GROUP_SET_SPAPR_TCE,
1126                 .addr = (uint64_t)(unsigned long)&param,
1127             };
1128 
1129             if (!memory_region_iommu_get_attr(iommu_mr, IOMMU_ATTR_SPAPR_TCE_FD,
1130                                               &param.tablefd)) {
1131                 QLIST_FOREACH(group, &container->group_list, container_next) {
1132                     param.groupfd = group->fd;
1133                     if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) {
1134                         error_report("vfio: failed to setup fd %d "
1135                                      "for a group with fd %d: %s",
1136                                      param.tablefd, param.groupfd,
1137                                      strerror(errno));
1138                         return;
1139                     }
1140                     trace_vfio_spapr_group_attach(param.groupfd, param.tablefd);
1141                 }
1142             }
1143         }
1144 #endif
1145     }
1146 
1147     hostwin = vfio_find_hostwin(container, iova, end);
1148     if (!hostwin) {
1149         error_setg(&err, "Container %p can't map guest IOVA region"
1150                    " 0x%"HWADDR_PRIx"..0x%"HWADDR_PRIx, container, iova, end);
1151         goto fail;
1152     }
1153 
1154     memory_region_ref(section->mr);
1155 
1156     if (memory_region_is_iommu(section->mr)) {
1157         VFIOGuestIOMMU *giommu;
1158         IOMMUMemoryRegion *iommu_mr = IOMMU_MEMORY_REGION(section->mr);
1159         int iommu_idx;
1160 
1161         trace_vfio_listener_region_add_iommu(iova, end);
1162         /*
1163          * FIXME: For VFIO iommu types which have KVM acceleration to
1164          * avoid bouncing all map/unmaps through qemu this way, this
1165          * would be the right place to wire that up (tell the KVM
1166          * device emulation the VFIO iommu handles to use).
1167          */
1168         giommu = g_malloc0(sizeof(*giommu));
1169         giommu->iommu_mr = iommu_mr;
1170         giommu->iommu_offset = section->offset_within_address_space -
1171                                section->offset_within_region;
1172         giommu->container = container;
1173         llend = int128_add(int128_make64(section->offset_within_region),
1174                            section->size);
1175         llend = int128_sub(llend, int128_one());
1176         iommu_idx = memory_region_iommu_attrs_to_index(iommu_mr,
1177                                                        MEMTXATTRS_UNSPECIFIED);
1178         iommu_notifier_init(&giommu->n, vfio_iommu_map_notify,
1179                             IOMMU_NOTIFIER_IOTLB_EVENTS,
1180                             section->offset_within_region,
1181                             int128_get64(llend),
1182                             iommu_idx);
1183 
1184         ret = memory_region_iommu_set_page_size_mask(giommu->iommu_mr,
1185                                                      container->pgsizes,
1186                                                      &err);
1187         if (ret) {
1188             g_free(giommu);
1189             goto fail;
1190         }
1191 
1192         ret = memory_region_register_iommu_notifier(section->mr, &giommu->n,
1193                                                     &err);
1194         if (ret) {
1195             g_free(giommu);
1196             goto fail;
1197         }
1198         QLIST_INSERT_HEAD(&container->giommu_list, giommu, giommu_next);
1199         memory_region_iommu_replay(giommu->iommu_mr, &giommu->n);
1200 
1201         return;
1202     }
1203 
1204     /* Here we assume that memory_region_is_ram(section->mr)==true */
1205 
1206     /*
1207      * For RAM memory regions with a RamDiscardManager, we only want to map the
1208      * actually populated parts - and update the mapping whenever we're notified
1209      * about changes.
1210      */
1211     if (memory_region_has_ram_discard_manager(section->mr)) {
1212         vfio_register_ram_discard_listener(container, section);
1213         return;
1214     }
1215 
1216     vaddr = memory_region_get_ram_ptr(section->mr) +
1217             section->offset_within_region +
1218             (iova - section->offset_within_address_space);
1219 
1220     trace_vfio_listener_region_add_ram(iova, end, vaddr);
1221 
1222     llsize = int128_sub(llend, int128_make64(iova));
1223 
1224     if (memory_region_is_ram_device(section->mr)) {
1225         hwaddr pgmask = (1ULL << ctz64(hostwin->iova_pgsizes)) - 1;
1226 
1227         if ((iova & pgmask) || (int128_get64(llsize) & pgmask)) {
1228             trace_vfio_listener_region_add_no_dma_map(
1229                 memory_region_name(section->mr),
1230                 section->offset_within_address_space,
1231                 int128_getlo(section->size),
1232                 pgmask + 1);
1233             return;
1234         }
1235     }
1236 
1237     ret = vfio_dma_map(container, iova, int128_get64(llsize),
1238                        vaddr, section->readonly);
1239     if (ret) {
1240         error_setg(&err, "vfio_dma_map(%p, 0x%"HWADDR_PRIx", "
1241                    "0x%"HWADDR_PRIx", %p) = %d (%s)",
1242                    container, iova, int128_get64(llsize), vaddr, ret,
1243                    strerror(-ret));
1244         if (memory_region_is_ram_device(section->mr)) {
1245             /* Allow unexpected mappings not to be fatal for RAM devices */
1246             error_report_err(err);
1247             return;
1248         }
1249         goto fail;
1250     }
1251 
1252     return;
1253 
1254 fail:
1255     if (memory_region_is_ram_device(section->mr)) {
1256         error_report("failed to vfio_dma_map. pci p2p may not work");
1257         return;
1258     }
1259     /*
1260      * On the initfn path, store the first error in the container so we
1261      * can gracefully fail.  Runtime, there's not much we can do other
1262      * than throw a hardware error.
1263      */
1264     if (!container->initialized) {
1265         if (!container->error) {
1266             error_propagate_prepend(&container->error, err,
1267                                     "Region %s: ",
1268                                     memory_region_name(section->mr));
1269         } else {
1270             error_free(err);
1271         }
1272     } else {
1273         error_report_err(err);
1274         hw_error("vfio: DMA mapping failed, unable to continue");
1275     }
1276 }
1277 
1278 static void vfio_listener_region_del(MemoryListener *listener,
1279                                      MemoryRegionSection *section)
1280 {
1281     VFIOContainer *container = container_of(listener, VFIOContainer, listener);
1282     hwaddr iova, end;
1283     Int128 llend, llsize;
1284     int ret;
1285     bool try_unmap = true;
1286 
1287     if (!vfio_listener_valid_section(section, "region_del")) {
1288         return;
1289     }
1290 
1291     if (memory_region_is_iommu(section->mr)) {
1292         VFIOGuestIOMMU *giommu;
1293 
1294         QLIST_FOREACH(giommu, &container->giommu_list, giommu_next) {
1295             if (MEMORY_REGION(giommu->iommu_mr) == section->mr &&
1296                 giommu->n.start == section->offset_within_region) {
1297                 memory_region_unregister_iommu_notifier(section->mr,
1298                                                         &giommu->n);
1299                 QLIST_REMOVE(giommu, giommu_next);
1300                 g_free(giommu);
1301                 break;
1302             }
1303         }
1304 
1305         /*
1306          * FIXME: We assume the one big unmap below is adequate to
1307          * remove any individual page mappings in the IOMMU which
1308          * might have been copied into VFIO. This works for a page table
1309          * based IOMMU where a big unmap flattens a large range of IO-PTEs.
1310          * That may not be true for all IOMMU types.
1311          */
1312     }
1313 
1314     if (!vfio_get_section_iova_range(container, section, &iova, &end, &llend)) {
1315         return;
1316     }
1317 
1318     llsize = int128_sub(llend, int128_make64(iova));
1319 
1320     trace_vfio_listener_region_del(iova, end);
1321 
1322     if (memory_region_is_ram_device(section->mr)) {
1323         hwaddr pgmask;
1324         VFIOHostDMAWindow *hostwin;
1325 
1326         hostwin = vfio_find_hostwin(container, iova, end);
1327         assert(hostwin); /* or region_add() would have failed */
1328 
1329         pgmask = (1ULL << ctz64(hostwin->iova_pgsizes)) - 1;
1330         try_unmap = !((iova & pgmask) || (int128_get64(llsize) & pgmask));
1331     } else if (memory_region_has_ram_discard_manager(section->mr)) {
1332         vfio_unregister_ram_discard_listener(container, section);
1333         /* Unregistering will trigger an unmap. */
1334         try_unmap = false;
1335     }
1336 
1337     if (try_unmap) {
1338         if (int128_eq(llsize, int128_2_64())) {
1339             /* The unmap ioctl doesn't accept a full 64-bit span. */
1340             llsize = int128_rshift(llsize, 1);
1341             ret = vfio_dma_unmap(container, iova, int128_get64(llsize), NULL);
1342             if (ret) {
1343                 error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx", "
1344                              "0x%"HWADDR_PRIx") = %d (%s)",
1345                              container, iova, int128_get64(llsize), ret,
1346                              strerror(-ret));
1347             }
1348             iova += int128_get64(llsize);
1349         }
1350         ret = vfio_dma_unmap(container, iova, int128_get64(llsize), NULL);
1351         if (ret) {
1352             error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx", "
1353                          "0x%"HWADDR_PRIx") = %d (%s)",
1354                          container, iova, int128_get64(llsize), ret,
1355                          strerror(-ret));
1356         }
1357     }
1358 
1359     memory_region_unref(section->mr);
1360 
1361     if (container->iommu_type == VFIO_SPAPR_TCE_v2_IOMMU) {
1362         vfio_spapr_remove_window(container,
1363                                  section->offset_within_address_space);
1364         if (vfio_host_win_del(container,
1365                               section->offset_within_address_space,
1366                               section->offset_within_address_space +
1367                               int128_get64(section->size) - 1) < 0) {
1368             hw_error("%s: Cannot delete missing window at %"HWADDR_PRIx,
1369                      __func__, section->offset_within_address_space);
1370         }
1371     }
1372 }
1373 
1374 static int vfio_set_dirty_page_tracking(VFIOContainer *container, bool start)
1375 {
1376     int ret;
1377     struct vfio_iommu_type1_dirty_bitmap dirty = {
1378         .argsz = sizeof(dirty),
1379     };
1380 
1381     if (!container->dirty_pages_supported) {
1382         return 0;
1383     }
1384 
1385     if (start) {
1386         dirty.flags = VFIO_IOMMU_DIRTY_PAGES_FLAG_START;
1387     } else {
1388         dirty.flags = VFIO_IOMMU_DIRTY_PAGES_FLAG_STOP;
1389     }
1390 
1391     ret = ioctl(container->fd, VFIO_IOMMU_DIRTY_PAGES, &dirty);
1392     if (ret) {
1393         ret = -errno;
1394         error_report("Failed to set dirty tracking flag 0x%x errno: %d",
1395                      dirty.flags, errno);
1396     }
1397 
1398     return ret;
1399 }
1400 
1401 typedef struct VFIODirtyRanges {
1402     hwaddr min32;
1403     hwaddr max32;
1404     hwaddr min64;
1405     hwaddr max64;
1406 } VFIODirtyRanges;
1407 
1408 typedef struct VFIODirtyRangesListener {
1409     VFIOContainer *container;
1410     VFIODirtyRanges ranges;
1411     MemoryListener listener;
1412 } VFIODirtyRangesListener;
1413 
1414 static void vfio_dirty_tracking_update(MemoryListener *listener,
1415                                        MemoryRegionSection *section)
1416 {
1417     VFIODirtyRangesListener *dirty = container_of(listener,
1418                                                   VFIODirtyRangesListener,
1419                                                   listener);
1420     VFIODirtyRanges *range = &dirty->ranges;
1421     hwaddr iova, end, *min, *max;
1422 
1423     if (!vfio_listener_valid_section(section, "tracking_update") ||
1424         !vfio_get_section_iova_range(dirty->container, section,
1425                                      &iova, &end, NULL)) {
1426         return;
1427     }
1428 
1429     /*
1430      * The address space passed to the dirty tracker is reduced to two ranges:
1431      * one for 32-bit DMA ranges, and another one for 64-bit DMA ranges.
1432      * The underlying reports of dirty will query a sub-interval of each of
1433      * these ranges.
1434      *
1435      * The purpose of the dual range handling is to handle known cases of big
1436      * holes in the address space, like the x86 AMD 1T hole. The alternative
1437      * would be an IOVATree but that has a much bigger runtime overhead and
1438      * unnecessary complexity.
1439      */
1440     min = (end <= UINT32_MAX) ? &range->min32 : &range->min64;
1441     max = (end <= UINT32_MAX) ? &range->max32 : &range->max64;
1442 
1443     if (*min > iova) {
1444         *min = iova;
1445     }
1446     if (*max < end) {
1447         *max = end;
1448     }
1449 
1450     trace_vfio_device_dirty_tracking_update(iova, end, *min, *max);
1451     return;
1452 }
1453 
1454 static const MemoryListener vfio_dirty_tracking_listener = {
1455     .name = "vfio-tracking",
1456     .region_add = vfio_dirty_tracking_update,
1457 };
1458 
1459 static void vfio_dirty_tracking_init(VFIOContainer *container,
1460                                      VFIODirtyRanges *ranges)
1461 {
1462     VFIODirtyRangesListener dirty;
1463 
1464     memset(&dirty, 0, sizeof(dirty));
1465     dirty.ranges.min32 = UINT32_MAX;
1466     dirty.ranges.min64 = UINT64_MAX;
1467     dirty.listener = vfio_dirty_tracking_listener;
1468     dirty.container = container;
1469 
1470     memory_listener_register(&dirty.listener,
1471                              container->space->as);
1472 
1473     *ranges = dirty.ranges;
1474 
1475     /*
1476      * The memory listener is synchronous, and used to calculate the range
1477      * to dirty tracking. Unregister it after we are done as we are not
1478      * interested in any follow-up updates.
1479      */
1480     memory_listener_unregister(&dirty.listener);
1481 }
1482 
1483 static void vfio_devices_dma_logging_stop(VFIOContainer *container)
1484 {
1485     uint64_t buf[DIV_ROUND_UP(sizeof(struct vfio_device_feature),
1486                               sizeof(uint64_t))] = {};
1487     struct vfio_device_feature *feature = (struct vfio_device_feature *)buf;
1488     VFIODevice *vbasedev;
1489     VFIOGroup *group;
1490 
1491     feature->argsz = sizeof(buf);
1492     feature->flags = VFIO_DEVICE_FEATURE_SET |
1493                      VFIO_DEVICE_FEATURE_DMA_LOGGING_STOP;
1494 
1495     QLIST_FOREACH(group, &container->group_list, container_next) {
1496         QLIST_FOREACH(vbasedev, &group->device_list, next) {
1497             if (!vbasedev->dirty_tracking) {
1498                 continue;
1499             }
1500 
1501             if (ioctl(vbasedev->fd, VFIO_DEVICE_FEATURE, feature)) {
1502                 warn_report("%s: Failed to stop DMA logging, err %d (%s)",
1503                              vbasedev->name, -errno, strerror(errno));
1504             }
1505             vbasedev->dirty_tracking = false;
1506         }
1507     }
1508 }
1509 
1510 static struct vfio_device_feature *
1511 vfio_device_feature_dma_logging_start_create(VFIOContainer *container,
1512                                              VFIODirtyRanges *tracking)
1513 {
1514     struct vfio_device_feature *feature;
1515     size_t feature_size;
1516     struct vfio_device_feature_dma_logging_control *control;
1517     struct vfio_device_feature_dma_logging_range *ranges;
1518 
1519     feature_size = sizeof(struct vfio_device_feature) +
1520                    sizeof(struct vfio_device_feature_dma_logging_control);
1521     feature = g_try_malloc0(feature_size);
1522     if (!feature) {
1523         errno = ENOMEM;
1524         return NULL;
1525     }
1526     feature->argsz = feature_size;
1527     feature->flags = VFIO_DEVICE_FEATURE_SET |
1528                      VFIO_DEVICE_FEATURE_DMA_LOGGING_START;
1529 
1530     control = (struct vfio_device_feature_dma_logging_control *)feature->data;
1531     control->page_size = qemu_real_host_page_size();
1532 
1533     /*
1534      * DMA logging uAPI guarantees to support at least a number of ranges that
1535      * fits into a single host kernel base page.
1536      */
1537     control->num_ranges = !!tracking->max32 + !!tracking->max64;
1538     ranges = g_try_new0(struct vfio_device_feature_dma_logging_range,
1539                         control->num_ranges);
1540     if (!ranges) {
1541         g_free(feature);
1542         errno = ENOMEM;
1543 
1544         return NULL;
1545     }
1546 
1547     control->ranges = (__u64)(uintptr_t)ranges;
1548     if (tracking->max32) {
1549         ranges->iova = tracking->min32;
1550         ranges->length = (tracking->max32 - tracking->min32) + 1;
1551         ranges++;
1552     }
1553     if (tracking->max64) {
1554         ranges->iova = tracking->min64;
1555         ranges->length = (tracking->max64 - tracking->min64) + 1;
1556     }
1557 
1558     trace_vfio_device_dirty_tracking_start(control->num_ranges,
1559                                            tracking->min32, tracking->max32,
1560                                            tracking->min64, tracking->max64);
1561 
1562     return feature;
1563 }
1564 
1565 static void vfio_device_feature_dma_logging_start_destroy(
1566     struct vfio_device_feature *feature)
1567 {
1568     struct vfio_device_feature_dma_logging_control *control =
1569         (struct vfio_device_feature_dma_logging_control *)feature->data;
1570     struct vfio_device_feature_dma_logging_range *ranges =
1571         (struct vfio_device_feature_dma_logging_range *)(uintptr_t)control->ranges;
1572 
1573     g_free(ranges);
1574     g_free(feature);
1575 }
1576 
1577 static int vfio_devices_dma_logging_start(VFIOContainer *container)
1578 {
1579     struct vfio_device_feature *feature;
1580     VFIODirtyRanges ranges;
1581     VFIODevice *vbasedev;
1582     VFIOGroup *group;
1583     int ret = 0;
1584 
1585     vfio_dirty_tracking_init(container, &ranges);
1586     feature = vfio_device_feature_dma_logging_start_create(container,
1587                                                            &ranges);
1588     if (!feature) {
1589         return -errno;
1590     }
1591 
1592     QLIST_FOREACH(group, &container->group_list, container_next) {
1593         QLIST_FOREACH(vbasedev, &group->device_list, next) {
1594             if (vbasedev->dirty_tracking) {
1595                 continue;
1596             }
1597 
1598             ret = ioctl(vbasedev->fd, VFIO_DEVICE_FEATURE, feature);
1599             if (ret) {
1600                 ret = -errno;
1601                 error_report("%s: Failed to start DMA logging, err %d (%s)",
1602                              vbasedev->name, ret, strerror(errno));
1603                 goto out;
1604             }
1605             vbasedev->dirty_tracking = true;
1606         }
1607     }
1608 
1609 out:
1610     if (ret) {
1611         vfio_devices_dma_logging_stop(container);
1612     }
1613 
1614     vfio_device_feature_dma_logging_start_destroy(feature);
1615 
1616     return ret;
1617 }
1618 
1619 static void vfio_listener_log_global_start(MemoryListener *listener)
1620 {
1621     VFIOContainer *container = container_of(listener, VFIOContainer, listener);
1622     int ret;
1623 
1624     if (vfio_devices_all_device_dirty_tracking(container)) {
1625         ret = vfio_devices_dma_logging_start(container);
1626     } else {
1627         ret = vfio_set_dirty_page_tracking(container, true);
1628     }
1629 
1630     if (ret) {
1631         error_report("vfio: Could not start dirty page tracking, err: %d (%s)",
1632                      ret, strerror(-ret));
1633         vfio_set_migration_error(ret);
1634     }
1635 }
1636 
1637 static void vfio_listener_log_global_stop(MemoryListener *listener)
1638 {
1639     VFIOContainer *container = container_of(listener, VFIOContainer, listener);
1640     int ret = 0;
1641 
1642     if (vfio_devices_all_device_dirty_tracking(container)) {
1643         vfio_devices_dma_logging_stop(container);
1644     } else {
1645         ret = vfio_set_dirty_page_tracking(container, false);
1646     }
1647 
1648     if (ret) {
1649         error_report("vfio: Could not stop dirty page tracking, err: %d (%s)",
1650                      ret, strerror(-ret));
1651         vfio_set_migration_error(ret);
1652     }
1653 }
1654 
1655 static int vfio_device_dma_logging_report(VFIODevice *vbasedev, hwaddr iova,
1656                                           hwaddr size, void *bitmap)
1657 {
1658     uint64_t buf[DIV_ROUND_UP(sizeof(struct vfio_device_feature) +
1659                         sizeof(struct vfio_device_feature_dma_logging_report),
1660                         sizeof(__u64))] = {};
1661     struct vfio_device_feature *feature = (struct vfio_device_feature *)buf;
1662     struct vfio_device_feature_dma_logging_report *report =
1663         (struct vfio_device_feature_dma_logging_report *)feature->data;
1664 
1665     report->iova = iova;
1666     report->length = size;
1667     report->page_size = qemu_real_host_page_size();
1668     report->bitmap = (__u64)(uintptr_t)bitmap;
1669 
1670     feature->argsz = sizeof(buf);
1671     feature->flags = VFIO_DEVICE_FEATURE_GET |
1672                      VFIO_DEVICE_FEATURE_DMA_LOGGING_REPORT;
1673 
1674     if (ioctl(vbasedev->fd, VFIO_DEVICE_FEATURE, feature)) {
1675         return -errno;
1676     }
1677 
1678     return 0;
1679 }
1680 
1681 static int vfio_devices_query_dirty_bitmap(VFIOContainer *container,
1682                                            VFIOBitmap *vbmap, hwaddr iova,
1683                                            hwaddr size)
1684 {
1685     VFIODevice *vbasedev;
1686     VFIOGroup *group;
1687     int ret;
1688 
1689     QLIST_FOREACH(group, &container->group_list, container_next) {
1690         QLIST_FOREACH(vbasedev, &group->device_list, next) {
1691             ret = vfio_device_dma_logging_report(vbasedev, iova, size,
1692                                                  vbmap->bitmap);
1693             if (ret) {
1694                 error_report("%s: Failed to get DMA logging report, iova: "
1695                              "0x%" HWADDR_PRIx ", size: 0x%" HWADDR_PRIx
1696                              ", err: %d (%s)",
1697                              vbasedev->name, iova, size, ret, strerror(-ret));
1698 
1699                 return ret;
1700             }
1701         }
1702     }
1703 
1704     return 0;
1705 }
1706 
1707 static int vfio_query_dirty_bitmap(VFIOContainer *container, VFIOBitmap *vbmap,
1708                                    hwaddr iova, hwaddr size)
1709 {
1710     struct vfio_iommu_type1_dirty_bitmap *dbitmap;
1711     struct vfio_iommu_type1_dirty_bitmap_get *range;
1712     int ret;
1713 
1714     dbitmap = g_malloc0(sizeof(*dbitmap) + sizeof(*range));
1715 
1716     dbitmap->argsz = sizeof(*dbitmap) + sizeof(*range);
1717     dbitmap->flags = VFIO_IOMMU_DIRTY_PAGES_FLAG_GET_BITMAP;
1718     range = (struct vfio_iommu_type1_dirty_bitmap_get *)&dbitmap->data;
1719     range->iova = iova;
1720     range->size = size;
1721 
1722     /*
1723      * cpu_physical_memory_set_dirty_lebitmap() supports pages in bitmap of
1724      * qemu_real_host_page_size to mark those dirty. Hence set bitmap's pgsize
1725      * to qemu_real_host_page_size.
1726      */
1727     range->bitmap.pgsize = qemu_real_host_page_size();
1728     range->bitmap.size = vbmap->size;
1729     range->bitmap.data = (__u64 *)vbmap->bitmap;
1730 
1731     ret = ioctl(container->fd, VFIO_IOMMU_DIRTY_PAGES, dbitmap);
1732     if (ret) {
1733         ret = -errno;
1734         error_report("Failed to get dirty bitmap for iova: 0x%"PRIx64
1735                 " size: 0x%"PRIx64" err: %d", (uint64_t)range->iova,
1736                 (uint64_t)range->size, errno);
1737     }
1738 
1739     g_free(dbitmap);
1740 
1741     return ret;
1742 }
1743 
1744 static int vfio_get_dirty_bitmap(VFIOContainer *container, uint64_t iova,
1745                                  uint64_t size, ram_addr_t ram_addr)
1746 {
1747     bool all_device_dirty_tracking =
1748         vfio_devices_all_device_dirty_tracking(container);
1749     VFIOBitmap vbmap;
1750     int ret;
1751 
1752     if (!container->dirty_pages_supported && !all_device_dirty_tracking) {
1753         cpu_physical_memory_set_dirty_range(ram_addr, size,
1754                                             tcg_enabled() ? DIRTY_CLIENTS_ALL :
1755                                             DIRTY_CLIENTS_NOCODE);
1756         return 0;
1757     }
1758 
1759     ret = vfio_bitmap_alloc(&vbmap, size);
1760     if (ret) {
1761         return ret;
1762     }
1763 
1764     if (all_device_dirty_tracking) {
1765         ret = vfio_devices_query_dirty_bitmap(container, &vbmap, iova, size);
1766     } else {
1767         ret = vfio_query_dirty_bitmap(container, &vbmap, iova, size);
1768     }
1769 
1770     if (ret) {
1771         goto out;
1772     }
1773 
1774     cpu_physical_memory_set_dirty_lebitmap(vbmap.bitmap, ram_addr,
1775                                            vbmap.pages);
1776 
1777     trace_vfio_get_dirty_bitmap(container->fd, iova, size, vbmap.size,
1778                                 ram_addr);
1779 out:
1780     g_free(vbmap.bitmap);
1781 
1782     return ret;
1783 }
1784 
1785 typedef struct {
1786     IOMMUNotifier n;
1787     VFIOGuestIOMMU *giommu;
1788 } vfio_giommu_dirty_notifier;
1789 
1790 static void vfio_iommu_map_dirty_notify(IOMMUNotifier *n, IOMMUTLBEntry *iotlb)
1791 {
1792     vfio_giommu_dirty_notifier *gdn = container_of(n,
1793                                                 vfio_giommu_dirty_notifier, n);
1794     VFIOGuestIOMMU *giommu = gdn->giommu;
1795     VFIOContainer *container = giommu->container;
1796     hwaddr iova = iotlb->iova + giommu->iommu_offset;
1797     ram_addr_t translated_addr;
1798     int ret = -EINVAL;
1799 
1800     trace_vfio_iommu_map_dirty_notify(iova, iova + iotlb->addr_mask);
1801 
1802     if (iotlb->target_as != &address_space_memory) {
1803         error_report("Wrong target AS \"%s\", only system memory is allowed",
1804                      iotlb->target_as->name ? iotlb->target_as->name : "none");
1805         goto out;
1806     }
1807 
1808     rcu_read_lock();
1809     if (vfio_get_xlat_addr(iotlb, NULL, &translated_addr, NULL)) {
1810         ret = vfio_get_dirty_bitmap(container, iova, iotlb->addr_mask + 1,
1811                                     translated_addr);
1812         if (ret) {
1813             error_report("vfio_iommu_map_dirty_notify(%p, 0x%"HWADDR_PRIx", "
1814                          "0x%"HWADDR_PRIx") = %d (%s)",
1815                          container, iova, iotlb->addr_mask + 1, ret,
1816                          strerror(-ret));
1817         }
1818     }
1819     rcu_read_unlock();
1820 
1821 out:
1822     if (ret) {
1823         vfio_set_migration_error(ret);
1824     }
1825 }
1826 
1827 static int vfio_ram_discard_get_dirty_bitmap(MemoryRegionSection *section,
1828                                              void *opaque)
1829 {
1830     const hwaddr size = int128_get64(section->size);
1831     const hwaddr iova = section->offset_within_address_space;
1832     const ram_addr_t ram_addr = memory_region_get_ram_addr(section->mr) +
1833                                 section->offset_within_region;
1834     VFIORamDiscardListener *vrdl = opaque;
1835 
1836     /*
1837      * Sync the whole mapped region (spanning multiple individual mappings)
1838      * in one go.
1839      */
1840     return vfio_get_dirty_bitmap(vrdl->container, iova, size, ram_addr);
1841 }
1842 
1843 static int vfio_sync_ram_discard_listener_dirty_bitmap(VFIOContainer *container,
1844                                                    MemoryRegionSection *section)
1845 {
1846     RamDiscardManager *rdm = memory_region_get_ram_discard_manager(section->mr);
1847     VFIORamDiscardListener *vrdl = NULL;
1848 
1849     QLIST_FOREACH(vrdl, &container->vrdl_list, next) {
1850         if (vrdl->mr == section->mr &&
1851             vrdl->offset_within_address_space ==
1852             section->offset_within_address_space) {
1853             break;
1854         }
1855     }
1856 
1857     if (!vrdl) {
1858         hw_error("vfio: Trying to sync missing RAM discard listener");
1859     }
1860 
1861     /*
1862      * We only want/can synchronize the bitmap for actually mapped parts -
1863      * which correspond to populated parts. Replay all populated parts.
1864      */
1865     return ram_discard_manager_replay_populated(rdm, section,
1866                                               vfio_ram_discard_get_dirty_bitmap,
1867                                                 &vrdl);
1868 }
1869 
1870 static int vfio_sync_dirty_bitmap(VFIOContainer *container,
1871                                   MemoryRegionSection *section)
1872 {
1873     ram_addr_t ram_addr;
1874 
1875     if (memory_region_is_iommu(section->mr)) {
1876         VFIOGuestIOMMU *giommu;
1877 
1878         QLIST_FOREACH(giommu, &container->giommu_list, giommu_next) {
1879             if (MEMORY_REGION(giommu->iommu_mr) == section->mr &&
1880                 giommu->n.start == section->offset_within_region) {
1881                 Int128 llend;
1882                 vfio_giommu_dirty_notifier gdn = { .giommu = giommu };
1883                 int idx = memory_region_iommu_attrs_to_index(giommu->iommu_mr,
1884                                                        MEMTXATTRS_UNSPECIFIED);
1885 
1886                 llend = int128_add(int128_make64(section->offset_within_region),
1887                                    section->size);
1888                 llend = int128_sub(llend, int128_one());
1889 
1890                 iommu_notifier_init(&gdn.n,
1891                                     vfio_iommu_map_dirty_notify,
1892                                     IOMMU_NOTIFIER_MAP,
1893                                     section->offset_within_region,
1894                                     int128_get64(llend),
1895                                     idx);
1896                 memory_region_iommu_replay(giommu->iommu_mr, &gdn.n);
1897                 break;
1898             }
1899         }
1900         return 0;
1901     } else if (memory_region_has_ram_discard_manager(section->mr)) {
1902         return vfio_sync_ram_discard_listener_dirty_bitmap(container, section);
1903     }
1904 
1905     ram_addr = memory_region_get_ram_addr(section->mr) +
1906                section->offset_within_region;
1907 
1908     return vfio_get_dirty_bitmap(container,
1909                    REAL_HOST_PAGE_ALIGN(section->offset_within_address_space),
1910                    int128_get64(section->size), ram_addr);
1911 }
1912 
1913 static void vfio_listener_log_sync(MemoryListener *listener,
1914         MemoryRegionSection *section)
1915 {
1916     VFIOContainer *container = container_of(listener, VFIOContainer, listener);
1917     int ret;
1918 
1919     if (vfio_listener_skipped_section(section)) {
1920         return;
1921     }
1922 
1923     if (vfio_devices_all_dirty_tracking(container)) {
1924         ret = vfio_sync_dirty_bitmap(container, section);
1925         if (ret) {
1926             error_report("vfio: Failed to sync dirty bitmap, err: %d (%s)", ret,
1927                          strerror(-ret));
1928             vfio_set_migration_error(ret);
1929         }
1930     }
1931 }
1932 
1933 static const MemoryListener vfio_memory_listener = {
1934     .name = "vfio",
1935     .region_add = vfio_listener_region_add,
1936     .region_del = vfio_listener_region_del,
1937     .log_global_start = vfio_listener_log_global_start,
1938     .log_global_stop = vfio_listener_log_global_stop,
1939     .log_sync = vfio_listener_log_sync,
1940 };
1941 
1942 static void vfio_listener_release(VFIOContainer *container)
1943 {
1944     memory_listener_unregister(&container->listener);
1945     if (container->iommu_type == VFIO_SPAPR_TCE_v2_IOMMU) {
1946         memory_listener_unregister(&container->prereg_listener);
1947     }
1948 }
1949 
1950 static struct vfio_info_cap_header *
1951 vfio_get_cap(void *ptr, uint32_t cap_offset, uint16_t id)
1952 {
1953     struct vfio_info_cap_header *hdr;
1954 
1955     for (hdr = ptr + cap_offset; hdr != ptr; hdr = ptr + hdr->next) {
1956         if (hdr->id == id) {
1957             return hdr;
1958         }
1959     }
1960 
1961     return NULL;
1962 }
1963 
1964 struct vfio_info_cap_header *
1965 vfio_get_region_info_cap(struct vfio_region_info *info, uint16_t id)
1966 {
1967     if (!(info->flags & VFIO_REGION_INFO_FLAG_CAPS)) {
1968         return NULL;
1969     }
1970 
1971     return vfio_get_cap((void *)info, info->cap_offset, id);
1972 }
1973 
1974 static struct vfio_info_cap_header *
1975 vfio_get_iommu_type1_info_cap(struct vfio_iommu_type1_info *info, uint16_t id)
1976 {
1977     if (!(info->flags & VFIO_IOMMU_INFO_CAPS)) {
1978         return NULL;
1979     }
1980 
1981     return vfio_get_cap((void *)info, info->cap_offset, id);
1982 }
1983 
1984 struct vfio_info_cap_header *
1985 vfio_get_device_info_cap(struct vfio_device_info *info, uint16_t id)
1986 {
1987     if (!(info->flags & VFIO_DEVICE_FLAGS_CAPS)) {
1988         return NULL;
1989     }
1990 
1991     return vfio_get_cap((void *)info, info->cap_offset, id);
1992 }
1993 
1994 bool vfio_get_info_dma_avail(struct vfio_iommu_type1_info *info,
1995                              unsigned int *avail)
1996 {
1997     struct vfio_info_cap_header *hdr;
1998     struct vfio_iommu_type1_info_dma_avail *cap;
1999 
2000     /* If the capability cannot be found, assume no DMA limiting */
2001     hdr = vfio_get_iommu_type1_info_cap(info,
2002                                         VFIO_IOMMU_TYPE1_INFO_DMA_AVAIL);
2003     if (hdr == NULL) {
2004         return false;
2005     }
2006 
2007     if (avail != NULL) {
2008         cap = (void *) hdr;
2009         *avail = cap->avail;
2010     }
2011 
2012     return true;
2013 }
2014 
2015 static int vfio_setup_region_sparse_mmaps(VFIORegion *region,
2016                                           struct vfio_region_info *info)
2017 {
2018     struct vfio_info_cap_header *hdr;
2019     struct vfio_region_info_cap_sparse_mmap *sparse;
2020     int i, j;
2021 
2022     hdr = vfio_get_region_info_cap(info, VFIO_REGION_INFO_CAP_SPARSE_MMAP);
2023     if (!hdr) {
2024         return -ENODEV;
2025     }
2026 
2027     sparse = container_of(hdr, struct vfio_region_info_cap_sparse_mmap, header);
2028 
2029     trace_vfio_region_sparse_mmap_header(region->vbasedev->name,
2030                                          region->nr, sparse->nr_areas);
2031 
2032     region->mmaps = g_new0(VFIOMmap, sparse->nr_areas);
2033 
2034     for (i = 0, j = 0; i < sparse->nr_areas; i++) {
2035         if (sparse->areas[i].size) {
2036             trace_vfio_region_sparse_mmap_entry(i, sparse->areas[i].offset,
2037                                             sparse->areas[i].offset +
2038                                             sparse->areas[i].size - 1);
2039             region->mmaps[j].offset = sparse->areas[i].offset;
2040             region->mmaps[j].size = sparse->areas[i].size;
2041             j++;
2042         }
2043     }
2044 
2045     region->nr_mmaps = j;
2046     region->mmaps = g_realloc(region->mmaps, j * sizeof(VFIOMmap));
2047 
2048     return 0;
2049 }
2050 
2051 int vfio_region_setup(Object *obj, VFIODevice *vbasedev, VFIORegion *region,
2052                       int index, const char *name)
2053 {
2054     struct vfio_region_info *info;
2055     int ret;
2056 
2057     ret = vfio_get_region_info(vbasedev, index, &info);
2058     if (ret) {
2059         return ret;
2060     }
2061 
2062     region->vbasedev = vbasedev;
2063     region->flags = info->flags;
2064     region->size = info->size;
2065     region->fd_offset = info->offset;
2066     region->nr = index;
2067 
2068     if (region->size) {
2069         region->mem = g_new0(MemoryRegion, 1);
2070         memory_region_init_io(region->mem, obj, &vfio_region_ops,
2071                               region, name, region->size);
2072 
2073         if (!vbasedev->no_mmap &&
2074             region->flags & VFIO_REGION_INFO_FLAG_MMAP) {
2075 
2076             ret = vfio_setup_region_sparse_mmaps(region, info);
2077 
2078             if (ret) {
2079                 region->nr_mmaps = 1;
2080                 region->mmaps = g_new0(VFIOMmap, region->nr_mmaps);
2081                 region->mmaps[0].offset = 0;
2082                 region->mmaps[0].size = region->size;
2083             }
2084         }
2085     }
2086 
2087     g_free(info);
2088 
2089     trace_vfio_region_setup(vbasedev->name, index, name,
2090                             region->flags, region->fd_offset, region->size);
2091     return 0;
2092 }
2093 
2094 static void vfio_subregion_unmap(VFIORegion *region, int index)
2095 {
2096     trace_vfio_region_unmap(memory_region_name(&region->mmaps[index].mem),
2097                             region->mmaps[index].offset,
2098                             region->mmaps[index].offset +
2099                             region->mmaps[index].size - 1);
2100     memory_region_del_subregion(region->mem, &region->mmaps[index].mem);
2101     munmap(region->mmaps[index].mmap, region->mmaps[index].size);
2102     object_unparent(OBJECT(&region->mmaps[index].mem));
2103     region->mmaps[index].mmap = NULL;
2104 }
2105 
2106 int vfio_region_mmap(VFIORegion *region)
2107 {
2108     int i, prot = 0;
2109     char *name;
2110 
2111     if (!region->mem) {
2112         return 0;
2113     }
2114 
2115     prot |= region->flags & VFIO_REGION_INFO_FLAG_READ ? PROT_READ : 0;
2116     prot |= region->flags & VFIO_REGION_INFO_FLAG_WRITE ? PROT_WRITE : 0;
2117 
2118     for (i = 0; i < region->nr_mmaps; i++) {
2119         region->mmaps[i].mmap = mmap(NULL, region->mmaps[i].size, prot,
2120                                      MAP_SHARED, region->vbasedev->fd,
2121                                      region->fd_offset +
2122                                      region->mmaps[i].offset);
2123         if (region->mmaps[i].mmap == MAP_FAILED) {
2124             int ret = -errno;
2125 
2126             trace_vfio_region_mmap_fault(memory_region_name(region->mem), i,
2127                                          region->fd_offset +
2128                                          region->mmaps[i].offset,
2129                                          region->fd_offset +
2130                                          region->mmaps[i].offset +
2131                                          region->mmaps[i].size - 1, ret);
2132 
2133             region->mmaps[i].mmap = NULL;
2134 
2135             for (i--; i >= 0; i--) {
2136                 vfio_subregion_unmap(region, i);
2137             }
2138 
2139             return ret;
2140         }
2141 
2142         name = g_strdup_printf("%s mmaps[%d]",
2143                                memory_region_name(region->mem), i);
2144         memory_region_init_ram_device_ptr(&region->mmaps[i].mem,
2145                                           memory_region_owner(region->mem),
2146                                           name, region->mmaps[i].size,
2147                                           region->mmaps[i].mmap);
2148         g_free(name);
2149         memory_region_add_subregion(region->mem, region->mmaps[i].offset,
2150                                     &region->mmaps[i].mem);
2151 
2152         trace_vfio_region_mmap(memory_region_name(&region->mmaps[i].mem),
2153                                region->mmaps[i].offset,
2154                                region->mmaps[i].offset +
2155                                region->mmaps[i].size - 1);
2156     }
2157 
2158     return 0;
2159 }
2160 
2161 void vfio_region_unmap(VFIORegion *region)
2162 {
2163     int i;
2164 
2165     if (!region->mem) {
2166         return;
2167     }
2168 
2169     for (i = 0; i < region->nr_mmaps; i++) {
2170         if (region->mmaps[i].mmap) {
2171             vfio_subregion_unmap(region, i);
2172         }
2173     }
2174 }
2175 
2176 void vfio_region_exit(VFIORegion *region)
2177 {
2178     int i;
2179 
2180     if (!region->mem) {
2181         return;
2182     }
2183 
2184     for (i = 0; i < region->nr_mmaps; i++) {
2185         if (region->mmaps[i].mmap) {
2186             memory_region_del_subregion(region->mem, &region->mmaps[i].mem);
2187         }
2188     }
2189 
2190     trace_vfio_region_exit(region->vbasedev->name, region->nr);
2191 }
2192 
2193 void vfio_region_finalize(VFIORegion *region)
2194 {
2195     int i;
2196 
2197     if (!region->mem) {
2198         return;
2199     }
2200 
2201     for (i = 0; i < region->nr_mmaps; i++) {
2202         if (region->mmaps[i].mmap) {
2203             munmap(region->mmaps[i].mmap, region->mmaps[i].size);
2204             object_unparent(OBJECT(&region->mmaps[i].mem));
2205         }
2206     }
2207 
2208     object_unparent(OBJECT(region->mem));
2209 
2210     g_free(region->mem);
2211     g_free(region->mmaps);
2212 
2213     trace_vfio_region_finalize(region->vbasedev->name, region->nr);
2214 
2215     region->mem = NULL;
2216     region->mmaps = NULL;
2217     region->nr_mmaps = 0;
2218     region->size = 0;
2219     region->flags = 0;
2220     region->nr = 0;
2221 }
2222 
2223 void vfio_region_mmaps_set_enabled(VFIORegion *region, bool enabled)
2224 {
2225     int i;
2226 
2227     if (!region->mem) {
2228         return;
2229     }
2230 
2231     for (i = 0; i < region->nr_mmaps; i++) {
2232         if (region->mmaps[i].mmap) {
2233             memory_region_set_enabled(&region->mmaps[i].mem, enabled);
2234         }
2235     }
2236 
2237     trace_vfio_region_mmaps_set_enabled(memory_region_name(region->mem),
2238                                         enabled);
2239 }
2240 
2241 void vfio_reset_handler(void *opaque)
2242 {
2243     VFIOGroup *group;
2244     VFIODevice *vbasedev;
2245 
2246     QLIST_FOREACH(group, &vfio_group_list, next) {
2247         QLIST_FOREACH(vbasedev, &group->device_list, next) {
2248             if (vbasedev->dev->realized) {
2249                 vbasedev->ops->vfio_compute_needs_reset(vbasedev);
2250             }
2251         }
2252     }
2253 
2254     QLIST_FOREACH(group, &vfio_group_list, next) {
2255         QLIST_FOREACH(vbasedev, &group->device_list, next) {
2256             if (vbasedev->dev->realized && vbasedev->needs_reset) {
2257                 vbasedev->ops->vfio_hot_reset_multi(vbasedev);
2258             }
2259         }
2260     }
2261 }
2262 
2263 static void vfio_kvm_device_add_group(VFIOGroup *group)
2264 {
2265 #ifdef CONFIG_KVM
2266     struct kvm_device_attr attr = {
2267         .group = KVM_DEV_VFIO_GROUP,
2268         .attr = KVM_DEV_VFIO_GROUP_ADD,
2269         .addr = (uint64_t)(unsigned long)&group->fd,
2270     };
2271 
2272     if (!kvm_enabled()) {
2273         return;
2274     }
2275 
2276     if (vfio_kvm_device_fd < 0) {
2277         struct kvm_create_device cd = {
2278             .type = KVM_DEV_TYPE_VFIO,
2279         };
2280 
2281         if (kvm_vm_ioctl(kvm_state, KVM_CREATE_DEVICE, &cd)) {
2282             error_report("Failed to create KVM VFIO device: %m");
2283             return;
2284         }
2285 
2286         vfio_kvm_device_fd = cd.fd;
2287     }
2288 
2289     if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) {
2290         error_report("Failed to add group %d to KVM VFIO device: %m",
2291                      group->groupid);
2292     }
2293 #endif
2294 }
2295 
2296 static void vfio_kvm_device_del_group(VFIOGroup *group)
2297 {
2298 #ifdef CONFIG_KVM
2299     struct kvm_device_attr attr = {
2300         .group = KVM_DEV_VFIO_GROUP,
2301         .attr = KVM_DEV_VFIO_GROUP_DEL,
2302         .addr = (uint64_t)(unsigned long)&group->fd,
2303     };
2304 
2305     if (vfio_kvm_device_fd < 0) {
2306         return;
2307     }
2308 
2309     if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) {
2310         error_report("Failed to remove group %d from KVM VFIO device: %m",
2311                      group->groupid);
2312     }
2313 #endif
2314 }
2315 
2316 static VFIOAddressSpace *vfio_get_address_space(AddressSpace *as)
2317 {
2318     VFIOAddressSpace *space;
2319 
2320     QLIST_FOREACH(space, &vfio_address_spaces, list) {
2321         if (space->as == as) {
2322             return space;
2323         }
2324     }
2325 
2326     /* No suitable VFIOAddressSpace, create a new one */
2327     space = g_malloc0(sizeof(*space));
2328     space->as = as;
2329     QLIST_INIT(&space->containers);
2330 
2331     QLIST_INSERT_HEAD(&vfio_address_spaces, space, list);
2332 
2333     return space;
2334 }
2335 
2336 static void vfio_put_address_space(VFIOAddressSpace *space)
2337 {
2338     if (QLIST_EMPTY(&space->containers)) {
2339         QLIST_REMOVE(space, list);
2340         g_free(space);
2341     }
2342 }
2343 
2344 /*
2345  * vfio_get_iommu_type - selects the richest iommu_type (v2 first)
2346  */
2347 static int vfio_get_iommu_type(VFIOContainer *container,
2348                                Error **errp)
2349 {
2350     int iommu_types[] = { VFIO_TYPE1v2_IOMMU, VFIO_TYPE1_IOMMU,
2351                           VFIO_SPAPR_TCE_v2_IOMMU, VFIO_SPAPR_TCE_IOMMU };
2352     int i;
2353 
2354     for (i = 0; i < ARRAY_SIZE(iommu_types); i++) {
2355         if (ioctl(container->fd, VFIO_CHECK_EXTENSION, iommu_types[i])) {
2356             return iommu_types[i];
2357         }
2358     }
2359     error_setg(errp, "No available IOMMU models");
2360     return -EINVAL;
2361 }
2362 
2363 static int vfio_init_container(VFIOContainer *container, int group_fd,
2364                                Error **errp)
2365 {
2366     int iommu_type, ret;
2367 
2368     iommu_type = vfio_get_iommu_type(container, errp);
2369     if (iommu_type < 0) {
2370         return iommu_type;
2371     }
2372 
2373     ret = ioctl(group_fd, VFIO_GROUP_SET_CONTAINER, &container->fd);
2374     if (ret) {
2375         error_setg_errno(errp, errno, "Failed to set group container");
2376         return -errno;
2377     }
2378 
2379     while (ioctl(container->fd, VFIO_SET_IOMMU, iommu_type)) {
2380         if (iommu_type == VFIO_SPAPR_TCE_v2_IOMMU) {
2381             /*
2382              * On sPAPR, despite the IOMMU subdriver always advertises v1 and
2383              * v2, the running platform may not support v2 and there is no
2384              * way to guess it until an IOMMU group gets added to the container.
2385              * So in case it fails with v2, try v1 as a fallback.
2386              */
2387             iommu_type = VFIO_SPAPR_TCE_IOMMU;
2388             continue;
2389         }
2390         error_setg_errno(errp, errno, "Failed to set iommu for container");
2391         return -errno;
2392     }
2393 
2394     container->iommu_type = iommu_type;
2395     return 0;
2396 }
2397 
2398 static int vfio_get_iommu_info(VFIOContainer *container,
2399                                struct vfio_iommu_type1_info **info)
2400 {
2401 
2402     size_t argsz = sizeof(struct vfio_iommu_type1_info);
2403 
2404     *info = g_new0(struct vfio_iommu_type1_info, 1);
2405 again:
2406     (*info)->argsz = argsz;
2407 
2408     if (ioctl(container->fd, VFIO_IOMMU_GET_INFO, *info)) {
2409         g_free(*info);
2410         *info = NULL;
2411         return -errno;
2412     }
2413 
2414     if (((*info)->argsz > argsz)) {
2415         argsz = (*info)->argsz;
2416         *info = g_realloc(*info, argsz);
2417         goto again;
2418     }
2419 
2420     return 0;
2421 }
2422 
2423 static struct vfio_info_cap_header *
2424 vfio_get_iommu_info_cap(struct vfio_iommu_type1_info *info, uint16_t id)
2425 {
2426     struct vfio_info_cap_header *hdr;
2427     void *ptr = info;
2428 
2429     if (!(info->flags & VFIO_IOMMU_INFO_CAPS)) {
2430         return NULL;
2431     }
2432 
2433     for (hdr = ptr + info->cap_offset; hdr != ptr; hdr = ptr + hdr->next) {
2434         if (hdr->id == id) {
2435             return hdr;
2436         }
2437     }
2438 
2439     return NULL;
2440 }
2441 
2442 static void vfio_get_iommu_info_migration(VFIOContainer *container,
2443                                          struct vfio_iommu_type1_info *info)
2444 {
2445     struct vfio_info_cap_header *hdr;
2446     struct vfio_iommu_type1_info_cap_migration *cap_mig;
2447 
2448     hdr = vfio_get_iommu_info_cap(info, VFIO_IOMMU_TYPE1_INFO_CAP_MIGRATION);
2449     if (!hdr) {
2450         return;
2451     }
2452 
2453     cap_mig = container_of(hdr, struct vfio_iommu_type1_info_cap_migration,
2454                             header);
2455 
2456     /*
2457      * cpu_physical_memory_set_dirty_lebitmap() supports pages in bitmap of
2458      * qemu_real_host_page_size to mark those dirty.
2459      */
2460     if (cap_mig->pgsize_bitmap & qemu_real_host_page_size()) {
2461         container->dirty_pages_supported = true;
2462         container->max_dirty_bitmap_size = cap_mig->max_dirty_bitmap_size;
2463         container->dirty_pgsizes = cap_mig->pgsize_bitmap;
2464     }
2465 }
2466 
2467 static int vfio_connect_container(VFIOGroup *group, AddressSpace *as,
2468                                   Error **errp)
2469 {
2470     VFIOContainer *container;
2471     int ret, fd;
2472     VFIOAddressSpace *space;
2473 
2474     space = vfio_get_address_space(as);
2475 
2476     /*
2477      * VFIO is currently incompatible with discarding of RAM insofar as the
2478      * madvise to purge (zap) the page from QEMU's address space does not
2479      * interact with the memory API and therefore leaves stale virtual to
2480      * physical mappings in the IOMMU if the page was previously pinned.  We
2481      * therefore set discarding broken for each group added to a container,
2482      * whether the container is used individually or shared.  This provides
2483      * us with options to allow devices within a group to opt-in and allow
2484      * discarding, so long as it is done consistently for a group (for instance
2485      * if the device is an mdev device where it is known that the host vendor
2486      * driver will never pin pages outside of the working set of the guest
2487      * driver, which would thus not be discarding candidates).
2488      *
2489      * The first opportunity to induce pinning occurs here where we attempt to
2490      * attach the group to existing containers within the AddressSpace.  If any
2491      * pages are already zapped from the virtual address space, such as from
2492      * previous discards, new pinning will cause valid mappings to be
2493      * re-established.  Likewise, when the overall MemoryListener for a new
2494      * container is registered, a replay of mappings within the AddressSpace
2495      * will occur, re-establishing any previously zapped pages as well.
2496      *
2497      * Especially virtio-balloon is currently only prevented from discarding
2498      * new memory, it will not yet set ram_block_discard_set_required() and
2499      * therefore, neither stops us here or deals with the sudden memory
2500      * consumption of inflated memory.
2501      *
2502      * We do support discarding of memory coordinated via the RamDiscardManager
2503      * with some IOMMU types. vfio_ram_block_discard_disable() handles the
2504      * details once we know which type of IOMMU we are using.
2505      */
2506 
2507     QLIST_FOREACH(container, &space->containers, next) {
2508         if (!ioctl(group->fd, VFIO_GROUP_SET_CONTAINER, &container->fd)) {
2509             ret = vfio_ram_block_discard_disable(container, true);
2510             if (ret) {
2511                 error_setg_errno(errp, -ret,
2512                                  "Cannot set discarding of RAM broken");
2513                 if (ioctl(group->fd, VFIO_GROUP_UNSET_CONTAINER,
2514                           &container->fd)) {
2515                     error_report("vfio: error disconnecting group %d from"
2516                                  " container", group->groupid);
2517                 }
2518                 return ret;
2519             }
2520             group->container = container;
2521             QLIST_INSERT_HEAD(&container->group_list, group, container_next);
2522             vfio_kvm_device_add_group(group);
2523             return 0;
2524         }
2525     }
2526 
2527     fd = qemu_open_old("/dev/vfio/vfio", O_RDWR);
2528     if (fd < 0) {
2529         error_setg_errno(errp, errno, "failed to open /dev/vfio/vfio");
2530         ret = -errno;
2531         goto put_space_exit;
2532     }
2533 
2534     ret = ioctl(fd, VFIO_GET_API_VERSION);
2535     if (ret != VFIO_API_VERSION) {
2536         error_setg(errp, "supported vfio version: %d, "
2537                    "reported version: %d", VFIO_API_VERSION, ret);
2538         ret = -EINVAL;
2539         goto close_fd_exit;
2540     }
2541 
2542     container = g_malloc0(sizeof(*container));
2543     container->space = space;
2544     container->fd = fd;
2545     container->error = NULL;
2546     container->dirty_pages_supported = false;
2547     container->dma_max_mappings = 0;
2548     QLIST_INIT(&container->giommu_list);
2549     QLIST_INIT(&container->hostwin_list);
2550     QLIST_INIT(&container->vrdl_list);
2551 
2552     ret = vfio_init_container(container, group->fd, errp);
2553     if (ret) {
2554         goto free_container_exit;
2555     }
2556 
2557     ret = vfio_ram_block_discard_disable(container, true);
2558     if (ret) {
2559         error_setg_errno(errp, -ret, "Cannot set discarding of RAM broken");
2560         goto free_container_exit;
2561     }
2562 
2563     switch (container->iommu_type) {
2564     case VFIO_TYPE1v2_IOMMU:
2565     case VFIO_TYPE1_IOMMU:
2566     {
2567         struct vfio_iommu_type1_info *info;
2568 
2569         ret = vfio_get_iommu_info(container, &info);
2570         if (ret) {
2571             error_setg_errno(errp, -ret, "Failed to get VFIO IOMMU info");
2572             goto enable_discards_exit;
2573         }
2574 
2575         if (info->flags & VFIO_IOMMU_INFO_PGSIZES) {
2576             container->pgsizes = info->iova_pgsizes;
2577         } else {
2578             container->pgsizes = qemu_real_host_page_size();
2579         }
2580 
2581         if (!vfio_get_info_dma_avail(info, &container->dma_max_mappings)) {
2582             container->dma_max_mappings = 65535;
2583         }
2584         vfio_get_iommu_info_migration(container, info);
2585         g_free(info);
2586 
2587         /*
2588          * FIXME: We should parse VFIO_IOMMU_TYPE1_INFO_CAP_IOVA_RANGE
2589          * information to get the actual window extent rather than assume
2590          * a 64-bit IOVA address space.
2591          */
2592         vfio_host_win_add(container, 0, (hwaddr)-1, container->pgsizes);
2593 
2594         break;
2595     }
2596     case VFIO_SPAPR_TCE_v2_IOMMU:
2597     case VFIO_SPAPR_TCE_IOMMU:
2598     {
2599         struct vfio_iommu_spapr_tce_info info;
2600         bool v2 = container->iommu_type == VFIO_SPAPR_TCE_v2_IOMMU;
2601 
2602         /*
2603          * The host kernel code implementing VFIO_IOMMU_DISABLE is called
2604          * when container fd is closed so we do not call it explicitly
2605          * in this file.
2606          */
2607         if (!v2) {
2608             ret = ioctl(fd, VFIO_IOMMU_ENABLE);
2609             if (ret) {
2610                 error_setg_errno(errp, errno, "failed to enable container");
2611                 ret = -errno;
2612                 goto enable_discards_exit;
2613             }
2614         } else {
2615             container->prereg_listener = vfio_prereg_listener;
2616 
2617             memory_listener_register(&container->prereg_listener,
2618                                      &address_space_memory);
2619             if (container->error) {
2620                 memory_listener_unregister(&container->prereg_listener);
2621                 ret = -1;
2622                 error_propagate_prepend(errp, container->error,
2623                     "RAM memory listener initialization failed: ");
2624                 goto enable_discards_exit;
2625             }
2626         }
2627 
2628         info.argsz = sizeof(info);
2629         ret = ioctl(fd, VFIO_IOMMU_SPAPR_TCE_GET_INFO, &info);
2630         if (ret) {
2631             error_setg_errno(errp, errno,
2632                              "VFIO_IOMMU_SPAPR_TCE_GET_INFO failed");
2633             ret = -errno;
2634             if (v2) {
2635                 memory_listener_unregister(&container->prereg_listener);
2636             }
2637             goto enable_discards_exit;
2638         }
2639 
2640         if (v2) {
2641             container->pgsizes = info.ddw.pgsizes;
2642             /*
2643              * There is a default window in just created container.
2644              * To make region_add/del simpler, we better remove this
2645              * window now and let those iommu_listener callbacks
2646              * create/remove them when needed.
2647              */
2648             ret = vfio_spapr_remove_window(container, info.dma32_window_start);
2649             if (ret) {
2650                 error_setg_errno(errp, -ret,
2651                                  "failed to remove existing window");
2652                 goto enable_discards_exit;
2653             }
2654         } else {
2655             /* The default table uses 4K pages */
2656             container->pgsizes = 0x1000;
2657             vfio_host_win_add(container, info.dma32_window_start,
2658                               info.dma32_window_start +
2659                               info.dma32_window_size - 1,
2660                               0x1000);
2661         }
2662     }
2663     }
2664 
2665     vfio_kvm_device_add_group(group);
2666 
2667     QLIST_INIT(&container->group_list);
2668     QLIST_INSERT_HEAD(&space->containers, container, next);
2669 
2670     group->container = container;
2671     QLIST_INSERT_HEAD(&container->group_list, group, container_next);
2672 
2673     container->listener = vfio_memory_listener;
2674 
2675     memory_listener_register(&container->listener, container->space->as);
2676 
2677     if (container->error) {
2678         ret = -1;
2679         error_propagate_prepend(errp, container->error,
2680             "memory listener initialization failed: ");
2681         goto listener_release_exit;
2682     }
2683 
2684     container->initialized = true;
2685 
2686     return 0;
2687 listener_release_exit:
2688     QLIST_REMOVE(group, container_next);
2689     QLIST_REMOVE(container, next);
2690     vfio_kvm_device_del_group(group);
2691     vfio_listener_release(container);
2692 
2693 enable_discards_exit:
2694     vfio_ram_block_discard_disable(container, false);
2695 
2696 free_container_exit:
2697     g_free(container);
2698 
2699 close_fd_exit:
2700     close(fd);
2701 
2702 put_space_exit:
2703     vfio_put_address_space(space);
2704 
2705     return ret;
2706 }
2707 
2708 static void vfio_disconnect_container(VFIOGroup *group)
2709 {
2710     VFIOContainer *container = group->container;
2711 
2712     QLIST_REMOVE(group, container_next);
2713     group->container = NULL;
2714 
2715     /*
2716      * Explicitly release the listener first before unset container,
2717      * since unset may destroy the backend container if it's the last
2718      * group.
2719      */
2720     if (QLIST_EMPTY(&container->group_list)) {
2721         vfio_listener_release(container);
2722     }
2723 
2724     if (ioctl(group->fd, VFIO_GROUP_UNSET_CONTAINER, &container->fd)) {
2725         error_report("vfio: error disconnecting group %d from container",
2726                      group->groupid);
2727     }
2728 
2729     if (QLIST_EMPTY(&container->group_list)) {
2730         VFIOAddressSpace *space = container->space;
2731         VFIOGuestIOMMU *giommu, *tmp;
2732         VFIOHostDMAWindow *hostwin, *next;
2733 
2734         QLIST_REMOVE(container, next);
2735 
2736         QLIST_FOREACH_SAFE(giommu, &container->giommu_list, giommu_next, tmp) {
2737             memory_region_unregister_iommu_notifier(
2738                     MEMORY_REGION(giommu->iommu_mr), &giommu->n);
2739             QLIST_REMOVE(giommu, giommu_next);
2740             g_free(giommu);
2741         }
2742 
2743         QLIST_FOREACH_SAFE(hostwin, &container->hostwin_list, hostwin_next,
2744                            next) {
2745             QLIST_REMOVE(hostwin, hostwin_next);
2746             g_free(hostwin);
2747         }
2748 
2749         trace_vfio_disconnect_container(container->fd);
2750         close(container->fd);
2751         g_free(container);
2752 
2753         vfio_put_address_space(space);
2754     }
2755 }
2756 
2757 VFIOGroup *vfio_get_group(int groupid, AddressSpace *as, Error **errp)
2758 {
2759     VFIOGroup *group;
2760     char path[32];
2761     struct vfio_group_status status = { .argsz = sizeof(status) };
2762 
2763     QLIST_FOREACH(group, &vfio_group_list, next) {
2764         if (group->groupid == groupid) {
2765             /* Found it.  Now is it already in the right context? */
2766             if (group->container->space->as == as) {
2767                 return group;
2768             } else {
2769                 error_setg(errp, "group %d used in multiple address spaces",
2770                            group->groupid);
2771                 return NULL;
2772             }
2773         }
2774     }
2775 
2776     group = g_malloc0(sizeof(*group));
2777 
2778     snprintf(path, sizeof(path), "/dev/vfio/%d", groupid);
2779     group->fd = qemu_open_old(path, O_RDWR);
2780     if (group->fd < 0) {
2781         error_setg_errno(errp, errno, "failed to open %s", path);
2782         goto free_group_exit;
2783     }
2784 
2785     if (ioctl(group->fd, VFIO_GROUP_GET_STATUS, &status)) {
2786         error_setg_errno(errp, errno, "failed to get group %d status", groupid);
2787         goto close_fd_exit;
2788     }
2789 
2790     if (!(status.flags & VFIO_GROUP_FLAGS_VIABLE)) {
2791         error_setg(errp, "group %d is not viable", groupid);
2792         error_append_hint(errp,
2793                           "Please ensure all devices within the iommu_group "
2794                           "are bound to their vfio bus driver.\n");
2795         goto close_fd_exit;
2796     }
2797 
2798     group->groupid = groupid;
2799     QLIST_INIT(&group->device_list);
2800 
2801     if (vfio_connect_container(group, as, errp)) {
2802         error_prepend(errp, "failed to setup container for group %d: ",
2803                       groupid);
2804         goto close_fd_exit;
2805     }
2806 
2807     if (QLIST_EMPTY(&vfio_group_list)) {
2808         qemu_register_reset(vfio_reset_handler, NULL);
2809     }
2810 
2811     QLIST_INSERT_HEAD(&vfio_group_list, group, next);
2812 
2813     return group;
2814 
2815 close_fd_exit:
2816     close(group->fd);
2817 
2818 free_group_exit:
2819     g_free(group);
2820 
2821     return NULL;
2822 }
2823 
2824 void vfio_put_group(VFIOGroup *group)
2825 {
2826     if (!group || !QLIST_EMPTY(&group->device_list)) {
2827         return;
2828     }
2829 
2830     if (!group->ram_block_discard_allowed) {
2831         vfio_ram_block_discard_disable(group->container, false);
2832     }
2833     vfio_kvm_device_del_group(group);
2834     vfio_disconnect_container(group);
2835     QLIST_REMOVE(group, next);
2836     trace_vfio_put_group(group->fd);
2837     close(group->fd);
2838     g_free(group);
2839 
2840     if (QLIST_EMPTY(&vfio_group_list)) {
2841         qemu_unregister_reset(vfio_reset_handler, NULL);
2842     }
2843 }
2844 
2845 int vfio_get_device(VFIOGroup *group, const char *name,
2846                     VFIODevice *vbasedev, Error **errp)
2847 {
2848     struct vfio_device_info dev_info = { .argsz = sizeof(dev_info) };
2849     int ret, fd;
2850 
2851     fd = ioctl(group->fd, VFIO_GROUP_GET_DEVICE_FD, name);
2852     if (fd < 0) {
2853         error_setg_errno(errp, errno, "error getting device from group %d",
2854                          group->groupid);
2855         error_append_hint(errp,
2856                       "Verify all devices in group %d are bound to vfio-<bus> "
2857                       "or pci-stub and not already in use\n", group->groupid);
2858         return fd;
2859     }
2860 
2861     ret = ioctl(fd, VFIO_DEVICE_GET_INFO, &dev_info);
2862     if (ret) {
2863         error_setg_errno(errp, errno, "error getting device info");
2864         close(fd);
2865         return ret;
2866     }
2867 
2868     /*
2869      * Set discarding of RAM as not broken for this group if the driver knows
2870      * the device operates compatibly with discarding.  Setting must be
2871      * consistent per group, but since compatibility is really only possible
2872      * with mdev currently, we expect singleton groups.
2873      */
2874     if (vbasedev->ram_block_discard_allowed !=
2875         group->ram_block_discard_allowed) {
2876         if (!QLIST_EMPTY(&group->device_list)) {
2877             error_setg(errp, "Inconsistent setting of support for discarding "
2878                        "RAM (e.g., balloon) within group");
2879             close(fd);
2880             return -1;
2881         }
2882 
2883         if (!group->ram_block_discard_allowed) {
2884             group->ram_block_discard_allowed = true;
2885             vfio_ram_block_discard_disable(group->container, false);
2886         }
2887     }
2888 
2889     vbasedev->fd = fd;
2890     vbasedev->group = group;
2891     QLIST_INSERT_HEAD(&group->device_list, vbasedev, next);
2892 
2893     vbasedev->num_irqs = dev_info.num_irqs;
2894     vbasedev->num_regions = dev_info.num_regions;
2895     vbasedev->flags = dev_info.flags;
2896 
2897     trace_vfio_get_device(name, dev_info.flags, dev_info.num_regions,
2898                           dev_info.num_irqs);
2899 
2900     vbasedev->reset_works = !!(dev_info.flags & VFIO_DEVICE_FLAGS_RESET);
2901     return 0;
2902 }
2903 
2904 void vfio_put_base_device(VFIODevice *vbasedev)
2905 {
2906     if (!vbasedev->group) {
2907         return;
2908     }
2909     QLIST_REMOVE(vbasedev, next);
2910     vbasedev->group = NULL;
2911     trace_vfio_put_base_device(vbasedev->fd);
2912     close(vbasedev->fd);
2913 }
2914 
2915 int vfio_get_region_info(VFIODevice *vbasedev, int index,
2916                          struct vfio_region_info **info)
2917 {
2918     size_t argsz = sizeof(struct vfio_region_info);
2919 
2920     *info = g_malloc0(argsz);
2921 
2922     (*info)->index = index;
2923 retry:
2924     (*info)->argsz = argsz;
2925 
2926     if (ioctl(vbasedev->fd, VFIO_DEVICE_GET_REGION_INFO, *info)) {
2927         g_free(*info);
2928         *info = NULL;
2929         return -errno;
2930     }
2931 
2932     if ((*info)->argsz > argsz) {
2933         argsz = (*info)->argsz;
2934         *info = g_realloc(*info, argsz);
2935 
2936         goto retry;
2937     }
2938 
2939     return 0;
2940 }
2941 
2942 int vfio_get_dev_region_info(VFIODevice *vbasedev, uint32_t type,
2943                              uint32_t subtype, struct vfio_region_info **info)
2944 {
2945     int i;
2946 
2947     for (i = 0; i < vbasedev->num_regions; i++) {
2948         struct vfio_info_cap_header *hdr;
2949         struct vfio_region_info_cap_type *cap_type;
2950 
2951         if (vfio_get_region_info(vbasedev, i, info)) {
2952             continue;
2953         }
2954 
2955         hdr = vfio_get_region_info_cap(*info, VFIO_REGION_INFO_CAP_TYPE);
2956         if (!hdr) {
2957             g_free(*info);
2958             continue;
2959         }
2960 
2961         cap_type = container_of(hdr, struct vfio_region_info_cap_type, header);
2962 
2963         trace_vfio_get_dev_region(vbasedev->name, i,
2964                                   cap_type->type, cap_type->subtype);
2965 
2966         if (cap_type->type == type && cap_type->subtype == subtype) {
2967             return 0;
2968         }
2969 
2970         g_free(*info);
2971     }
2972 
2973     *info = NULL;
2974     return -ENODEV;
2975 }
2976 
2977 bool vfio_has_region_cap(VFIODevice *vbasedev, int region, uint16_t cap_type)
2978 {
2979     struct vfio_region_info *info = NULL;
2980     bool ret = false;
2981 
2982     if (!vfio_get_region_info(vbasedev, region, &info)) {
2983         if (vfio_get_region_info_cap(info, cap_type)) {
2984             ret = true;
2985         }
2986         g_free(info);
2987     }
2988 
2989     return ret;
2990 }
2991 
2992 /*
2993  * Interfaces for IBM EEH (Enhanced Error Handling)
2994  */
2995 static bool vfio_eeh_container_ok(VFIOContainer *container)
2996 {
2997     /*
2998      * As of 2016-03-04 (linux-4.5) the host kernel EEH/VFIO
2999      * implementation is broken if there are multiple groups in a
3000      * container.  The hardware works in units of Partitionable
3001      * Endpoints (== IOMMU groups) and the EEH operations naively
3002      * iterate across all groups in the container, without any logic
3003      * to make sure the groups have their state synchronized.  For
3004      * certain operations (ENABLE) that might be ok, until an error
3005      * occurs, but for others (GET_STATE) it's clearly broken.
3006      */
3007 
3008     /*
3009      * XXX Once fixed kernels exist, test for them here
3010      */
3011 
3012     if (QLIST_EMPTY(&container->group_list)) {
3013         return false;
3014     }
3015 
3016     if (QLIST_NEXT(QLIST_FIRST(&container->group_list), container_next)) {
3017         return false;
3018     }
3019 
3020     return true;
3021 }
3022 
3023 static int vfio_eeh_container_op(VFIOContainer *container, uint32_t op)
3024 {
3025     struct vfio_eeh_pe_op pe_op = {
3026         .argsz = sizeof(pe_op),
3027         .op = op,
3028     };
3029     int ret;
3030 
3031     if (!vfio_eeh_container_ok(container)) {
3032         error_report("vfio/eeh: EEH_PE_OP 0x%x: "
3033                      "kernel requires a container with exactly one group", op);
3034         return -EPERM;
3035     }
3036 
3037     ret = ioctl(container->fd, VFIO_EEH_PE_OP, &pe_op);
3038     if (ret < 0) {
3039         error_report("vfio/eeh: EEH_PE_OP 0x%x failed: %m", op);
3040         return -errno;
3041     }
3042 
3043     return ret;
3044 }
3045 
3046 static VFIOContainer *vfio_eeh_as_container(AddressSpace *as)
3047 {
3048     VFIOAddressSpace *space = vfio_get_address_space(as);
3049     VFIOContainer *container = NULL;
3050 
3051     if (QLIST_EMPTY(&space->containers)) {
3052         /* No containers to act on */
3053         goto out;
3054     }
3055 
3056     container = QLIST_FIRST(&space->containers);
3057 
3058     if (QLIST_NEXT(container, next)) {
3059         /* We don't yet have logic to synchronize EEH state across
3060          * multiple containers */
3061         container = NULL;
3062         goto out;
3063     }
3064 
3065 out:
3066     vfio_put_address_space(space);
3067     return container;
3068 }
3069 
3070 bool vfio_eeh_as_ok(AddressSpace *as)
3071 {
3072     VFIOContainer *container = vfio_eeh_as_container(as);
3073 
3074     return (container != NULL) && vfio_eeh_container_ok(container);
3075 }
3076 
3077 int vfio_eeh_as_op(AddressSpace *as, uint32_t op)
3078 {
3079     VFIOContainer *container = vfio_eeh_as_container(as);
3080 
3081     if (!container) {
3082         return -ENODEV;
3083     }
3084     return vfio_eeh_container_op(container, op);
3085 }
3086