xref: /openbmc/qemu/hw/vfio/common.c (revision d0bf492f)
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 
366 static unsigned int vfio_migratable_device_num(void)
367 {
368     VFIOGroup *group;
369     VFIODevice *vbasedev;
370     unsigned int device_num = 0;
371 
372     QLIST_FOREACH(group, &vfio_group_list, next) {
373         QLIST_FOREACH(vbasedev, &group->device_list, next) {
374             if (vbasedev->migration) {
375                 device_num++;
376             }
377         }
378     }
379 
380     return device_num;
381 }
382 
383 int vfio_block_multiple_devices_migration(VFIODevice *vbasedev, Error **errp)
384 {
385     int ret;
386 
387     if (multiple_devices_migration_blocker ||
388         vfio_migratable_device_num() <= 1) {
389         return 0;
390     }
391 
392     if (vbasedev->enable_migration == ON_OFF_AUTO_ON) {
393         error_setg(errp, "Migration is currently not supported with multiple "
394                          "VFIO devices");
395         return -EINVAL;
396     }
397 
398     error_setg(&multiple_devices_migration_blocker,
399                "Migration is currently not supported with multiple "
400                "VFIO devices");
401     ret = migrate_add_blocker(multiple_devices_migration_blocker, errp);
402     if (ret < 0) {
403         error_free(multiple_devices_migration_blocker);
404         multiple_devices_migration_blocker = NULL;
405     }
406 
407     return ret;
408 }
409 
410 void vfio_unblock_multiple_devices_migration(void)
411 {
412     if (!multiple_devices_migration_blocker ||
413         vfio_migratable_device_num() > 1) {
414         return;
415     }
416 
417     migrate_del_blocker(multiple_devices_migration_blocker);
418     error_free(multiple_devices_migration_blocker);
419     multiple_devices_migration_blocker = NULL;
420 }
421 
422 bool vfio_viommu_preset(VFIODevice *vbasedev)
423 {
424     return vbasedev->group->container->space->as != &address_space_memory;
425 }
426 
427 static void vfio_set_migration_error(int err)
428 {
429     MigrationState *ms = migrate_get_current();
430 
431     if (migration_is_setup_or_active(ms->state)) {
432         WITH_QEMU_LOCK_GUARD(&ms->qemu_file_lock) {
433             if (ms->to_dst_file) {
434                 qemu_file_set_error(ms->to_dst_file, err);
435             }
436         }
437     }
438 }
439 
440 static bool vfio_devices_all_dirty_tracking(VFIOContainer *container)
441 {
442     VFIOGroup *group;
443     VFIODevice *vbasedev;
444     MigrationState *ms = migrate_get_current();
445 
446     if (ms->state != MIGRATION_STATUS_ACTIVE &&
447         ms->state != MIGRATION_STATUS_DEVICE) {
448         return false;
449     }
450 
451     QLIST_FOREACH(group, &container->group_list, container_next) {
452         QLIST_FOREACH(vbasedev, &group->device_list, next) {
453             VFIOMigration *migration = vbasedev->migration;
454 
455             if (!migration) {
456                 return false;
457             }
458 
459             if (vbasedev->pre_copy_dirty_page_tracking == ON_OFF_AUTO_OFF &&
460                 (migration->device_state == VFIO_DEVICE_STATE_RUNNING ||
461                  migration->device_state == VFIO_DEVICE_STATE_PRE_COPY)) {
462                 return false;
463             }
464         }
465     }
466     return true;
467 }
468 
469 static bool vfio_devices_all_device_dirty_tracking(VFIOContainer *container)
470 {
471     VFIOGroup *group;
472     VFIODevice *vbasedev;
473 
474     QLIST_FOREACH(group, &container->group_list, container_next) {
475         QLIST_FOREACH(vbasedev, &group->device_list, next) {
476             if (!vbasedev->dirty_pages_supported) {
477                 return false;
478             }
479         }
480     }
481 
482     return true;
483 }
484 
485 /*
486  * Check if all VFIO devices are running and migration is active, which is
487  * essentially equivalent to the migration being in pre-copy phase.
488  */
489 static bool vfio_devices_all_running_and_mig_active(VFIOContainer *container)
490 {
491     VFIOGroup *group;
492     VFIODevice *vbasedev;
493 
494     if (!migration_is_active(migrate_get_current())) {
495         return false;
496     }
497 
498     QLIST_FOREACH(group, &container->group_list, container_next) {
499         QLIST_FOREACH(vbasedev, &group->device_list, next) {
500             VFIOMigration *migration = vbasedev->migration;
501 
502             if (!migration) {
503                 return false;
504             }
505 
506             if (migration->device_state == VFIO_DEVICE_STATE_RUNNING ||
507                 migration->device_state == VFIO_DEVICE_STATE_PRE_COPY) {
508                 continue;
509             } else {
510                 return false;
511             }
512         }
513     }
514     return true;
515 }
516 
517 static int vfio_dma_unmap_bitmap(VFIOContainer *container,
518                                  hwaddr iova, ram_addr_t size,
519                                  IOMMUTLBEntry *iotlb)
520 {
521     struct vfio_iommu_type1_dma_unmap *unmap;
522     struct vfio_bitmap *bitmap;
523     VFIOBitmap vbmap;
524     int ret;
525 
526     ret = vfio_bitmap_alloc(&vbmap, size);
527     if (ret) {
528         return ret;
529     }
530 
531     unmap = g_malloc0(sizeof(*unmap) + sizeof(*bitmap));
532 
533     unmap->argsz = sizeof(*unmap) + sizeof(*bitmap);
534     unmap->iova = iova;
535     unmap->size = size;
536     unmap->flags |= VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP;
537     bitmap = (struct vfio_bitmap *)&unmap->data;
538 
539     /*
540      * cpu_physical_memory_set_dirty_lebitmap() supports pages in bitmap of
541      * qemu_real_host_page_size to mark those dirty. Hence set bitmap_pgsize
542      * to qemu_real_host_page_size.
543      */
544     bitmap->pgsize = qemu_real_host_page_size();
545     bitmap->size = vbmap.size;
546     bitmap->data = (__u64 *)vbmap.bitmap;
547 
548     if (vbmap.size > container->max_dirty_bitmap_size) {
549         error_report("UNMAP: Size of bitmap too big 0x%"PRIx64, vbmap.size);
550         ret = -E2BIG;
551         goto unmap_exit;
552     }
553 
554     ret = ioctl(container->fd, VFIO_IOMMU_UNMAP_DMA, unmap);
555     if (!ret) {
556         cpu_physical_memory_set_dirty_lebitmap(vbmap.bitmap,
557                 iotlb->translated_addr, vbmap.pages);
558     } else {
559         error_report("VFIO_UNMAP_DMA with DIRTY_BITMAP : %m");
560     }
561 
562 unmap_exit:
563     g_free(unmap);
564     g_free(vbmap.bitmap);
565 
566     return ret;
567 }
568 
569 /*
570  * DMA - Mapping and unmapping for the "type1" IOMMU interface used on x86
571  */
572 static int vfio_dma_unmap(VFIOContainer *container,
573                           hwaddr iova, ram_addr_t size,
574                           IOMMUTLBEntry *iotlb)
575 {
576     struct vfio_iommu_type1_dma_unmap unmap = {
577         .argsz = sizeof(unmap),
578         .flags = 0,
579         .iova = iova,
580         .size = size,
581     };
582     bool need_dirty_sync = false;
583     int ret;
584 
585     if (iotlb && vfio_devices_all_running_and_mig_active(container)) {
586         if (!vfio_devices_all_device_dirty_tracking(container) &&
587             container->dirty_pages_supported) {
588             return vfio_dma_unmap_bitmap(container, iova, size, iotlb);
589         }
590 
591         need_dirty_sync = true;
592     }
593 
594     while (ioctl(container->fd, VFIO_IOMMU_UNMAP_DMA, &unmap)) {
595         /*
596          * The type1 backend has an off-by-one bug in the kernel (71a7d3d78e3c
597          * v4.15) where an overflow in its wrap-around check prevents us from
598          * unmapping the last page of the address space.  Test for the error
599          * condition and re-try the unmap excluding the last page.  The
600          * expectation is that we've never mapped the last page anyway and this
601          * unmap request comes via vIOMMU support which also makes it unlikely
602          * that this page is used.  This bug was introduced well after type1 v2
603          * support was introduced, so we shouldn't need to test for v1.  A fix
604          * is queued for kernel v5.0 so this workaround can be removed once
605          * affected kernels are sufficiently deprecated.
606          */
607         if (errno == EINVAL && unmap.size && !(unmap.iova + unmap.size) &&
608             container->iommu_type == VFIO_TYPE1v2_IOMMU) {
609             trace_vfio_dma_unmap_overflow_workaround();
610             unmap.size -= 1ULL << ctz64(container->pgsizes);
611             continue;
612         }
613         error_report("VFIO_UNMAP_DMA failed: %s", strerror(errno));
614         return -errno;
615     }
616 
617     if (need_dirty_sync) {
618         ret = vfio_get_dirty_bitmap(container, iova, size,
619                                     iotlb->translated_addr);
620         if (ret) {
621             return ret;
622         }
623     }
624 
625     return 0;
626 }
627 
628 static int vfio_dma_map(VFIOContainer *container, hwaddr iova,
629                         ram_addr_t size, void *vaddr, bool readonly)
630 {
631     struct vfio_iommu_type1_dma_map map = {
632         .argsz = sizeof(map),
633         .flags = VFIO_DMA_MAP_FLAG_READ,
634         .vaddr = (__u64)(uintptr_t)vaddr,
635         .iova = iova,
636         .size = size,
637     };
638 
639     if (!readonly) {
640         map.flags |= VFIO_DMA_MAP_FLAG_WRITE;
641     }
642 
643     /*
644      * Try the mapping, if it fails with EBUSY, unmap the region and try
645      * again.  This shouldn't be necessary, but we sometimes see it in
646      * the VGA ROM space.
647      */
648     if (ioctl(container->fd, VFIO_IOMMU_MAP_DMA, &map) == 0 ||
649         (errno == EBUSY && vfio_dma_unmap(container, iova, size, NULL) == 0 &&
650          ioctl(container->fd, VFIO_IOMMU_MAP_DMA, &map) == 0)) {
651         return 0;
652     }
653 
654     error_report("VFIO_MAP_DMA failed: %s", strerror(errno));
655     return -errno;
656 }
657 
658 static void vfio_host_win_add(VFIOContainer *container,
659                               hwaddr min_iova, hwaddr max_iova,
660                               uint64_t iova_pgsizes)
661 {
662     VFIOHostDMAWindow *hostwin;
663 
664     QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) {
665         if (ranges_overlap(hostwin->min_iova,
666                            hostwin->max_iova - hostwin->min_iova + 1,
667                            min_iova,
668                            max_iova - min_iova + 1)) {
669             hw_error("%s: Overlapped IOMMU are not enabled", __func__);
670         }
671     }
672 
673     hostwin = g_malloc0(sizeof(*hostwin));
674 
675     hostwin->min_iova = min_iova;
676     hostwin->max_iova = max_iova;
677     hostwin->iova_pgsizes = iova_pgsizes;
678     QLIST_INSERT_HEAD(&container->hostwin_list, hostwin, hostwin_next);
679 }
680 
681 static int vfio_host_win_del(VFIOContainer *container, hwaddr min_iova,
682                              hwaddr max_iova)
683 {
684     VFIOHostDMAWindow *hostwin;
685 
686     QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) {
687         if (hostwin->min_iova == min_iova && hostwin->max_iova == max_iova) {
688             QLIST_REMOVE(hostwin, hostwin_next);
689             g_free(hostwin);
690             return 0;
691         }
692     }
693 
694     return -1;
695 }
696 
697 static bool vfio_listener_skipped_section(MemoryRegionSection *section)
698 {
699     return (!memory_region_is_ram(section->mr) &&
700             !memory_region_is_iommu(section->mr)) ||
701            memory_region_is_protected(section->mr) ||
702            /*
703             * Sizing an enabled 64-bit BAR can cause spurious mappings to
704             * addresses in the upper part of the 64-bit address space.  These
705             * are never accessed by the CPU and beyond the address width of
706             * some IOMMU hardware.  TODO: VFIO should tell us the IOMMU width.
707             */
708            section->offset_within_address_space & (1ULL << 63);
709 }
710 
711 /* Called with rcu_read_lock held.  */
712 static bool vfio_get_xlat_addr(IOMMUTLBEntry *iotlb, void **vaddr,
713                                ram_addr_t *ram_addr, bool *read_only)
714 {
715     bool ret, mr_has_discard_manager;
716 
717     ret = memory_get_xlat_addr(iotlb, vaddr, ram_addr, read_only,
718                                &mr_has_discard_manager);
719     if (ret && mr_has_discard_manager) {
720         /*
721          * Malicious VMs might trigger discarding of IOMMU-mapped memory. The
722          * pages will remain pinned inside vfio until unmapped, resulting in a
723          * higher memory consumption than expected. If memory would get
724          * populated again later, there would be an inconsistency between pages
725          * pinned by vfio and pages seen by QEMU. This is the case until
726          * unmapped from the IOMMU (e.g., during device reset).
727          *
728          * With malicious guests, we really only care about pinning more memory
729          * than expected. RLIMIT_MEMLOCK set for the user/process can never be
730          * exceeded and can be used to mitigate this problem.
731          */
732         warn_report_once("Using vfio with vIOMMUs and coordinated discarding of"
733                          " RAM (e.g., virtio-mem) works, however, malicious"
734                          " guests can trigger pinning of more memory than"
735                          " intended via an IOMMU. It's possible to mitigate "
736                          " by setting/adjusting RLIMIT_MEMLOCK.");
737     }
738     return ret;
739 }
740 
741 static void vfio_iommu_map_notify(IOMMUNotifier *n, IOMMUTLBEntry *iotlb)
742 {
743     VFIOGuestIOMMU *giommu = container_of(n, VFIOGuestIOMMU, n);
744     VFIOContainer *container = giommu->container;
745     hwaddr iova = iotlb->iova + giommu->iommu_offset;
746     void *vaddr;
747     int ret;
748 
749     trace_vfio_iommu_map_notify(iotlb->perm == IOMMU_NONE ? "UNMAP" : "MAP",
750                                 iova, iova + iotlb->addr_mask);
751 
752     if (iotlb->target_as != &address_space_memory) {
753         error_report("Wrong target AS \"%s\", only system memory is allowed",
754                      iotlb->target_as->name ? iotlb->target_as->name : "none");
755         vfio_set_migration_error(-EINVAL);
756         return;
757     }
758 
759     rcu_read_lock();
760 
761     if ((iotlb->perm & IOMMU_RW) != IOMMU_NONE) {
762         bool read_only;
763 
764         if (!vfio_get_xlat_addr(iotlb, &vaddr, NULL, &read_only)) {
765             goto out;
766         }
767         /*
768          * vaddr is only valid until rcu_read_unlock(). But after
769          * vfio_dma_map has set up the mapping the pages will be
770          * pinned by the kernel. This makes sure that the RAM backend
771          * of vaddr will always be there, even if the memory object is
772          * destroyed and its backing memory munmap-ed.
773          */
774         ret = vfio_dma_map(container, iova,
775                            iotlb->addr_mask + 1, vaddr,
776                            read_only);
777         if (ret) {
778             error_report("vfio_dma_map(%p, 0x%"HWADDR_PRIx", "
779                          "0x%"HWADDR_PRIx", %p) = %d (%s)",
780                          container, iova,
781                          iotlb->addr_mask + 1, vaddr, ret, strerror(-ret));
782         }
783     } else {
784         ret = vfio_dma_unmap(container, iova, iotlb->addr_mask + 1, iotlb);
785         if (ret) {
786             error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx", "
787                          "0x%"HWADDR_PRIx") = %d (%s)",
788                          container, iova,
789                          iotlb->addr_mask + 1, ret, strerror(-ret));
790             vfio_set_migration_error(ret);
791         }
792     }
793 out:
794     rcu_read_unlock();
795 }
796 
797 static void vfio_ram_discard_notify_discard(RamDiscardListener *rdl,
798                                             MemoryRegionSection *section)
799 {
800     VFIORamDiscardListener *vrdl = container_of(rdl, VFIORamDiscardListener,
801                                                 listener);
802     const hwaddr size = int128_get64(section->size);
803     const hwaddr iova = section->offset_within_address_space;
804     int ret;
805 
806     /* Unmap with a single call. */
807     ret = vfio_dma_unmap(vrdl->container, iova, size , NULL);
808     if (ret) {
809         error_report("%s: vfio_dma_unmap() failed: %s", __func__,
810                      strerror(-ret));
811     }
812 }
813 
814 static int vfio_ram_discard_notify_populate(RamDiscardListener *rdl,
815                                             MemoryRegionSection *section)
816 {
817     VFIORamDiscardListener *vrdl = container_of(rdl, VFIORamDiscardListener,
818                                                 listener);
819     const hwaddr end = section->offset_within_region +
820                        int128_get64(section->size);
821     hwaddr start, next, iova;
822     void *vaddr;
823     int ret;
824 
825     /*
826      * Map in (aligned within memory region) minimum granularity, so we can
827      * unmap in minimum granularity later.
828      */
829     for (start = section->offset_within_region; start < end; start = next) {
830         next = ROUND_UP(start + 1, vrdl->granularity);
831         next = MIN(next, end);
832 
833         iova = start - section->offset_within_region +
834                section->offset_within_address_space;
835         vaddr = memory_region_get_ram_ptr(section->mr) + start;
836 
837         ret = vfio_dma_map(vrdl->container, iova, next - start,
838                            vaddr, section->readonly);
839         if (ret) {
840             /* Rollback */
841             vfio_ram_discard_notify_discard(rdl, section);
842             return ret;
843         }
844     }
845     return 0;
846 }
847 
848 static void vfio_register_ram_discard_listener(VFIOContainer *container,
849                                                MemoryRegionSection *section)
850 {
851     RamDiscardManager *rdm = memory_region_get_ram_discard_manager(section->mr);
852     VFIORamDiscardListener *vrdl;
853 
854     /* Ignore some corner cases not relevant in practice. */
855     g_assert(QEMU_IS_ALIGNED(section->offset_within_region, TARGET_PAGE_SIZE));
856     g_assert(QEMU_IS_ALIGNED(section->offset_within_address_space,
857                              TARGET_PAGE_SIZE));
858     g_assert(QEMU_IS_ALIGNED(int128_get64(section->size), TARGET_PAGE_SIZE));
859 
860     vrdl = g_new0(VFIORamDiscardListener, 1);
861     vrdl->container = container;
862     vrdl->mr = section->mr;
863     vrdl->offset_within_address_space = section->offset_within_address_space;
864     vrdl->size = int128_get64(section->size);
865     vrdl->granularity = ram_discard_manager_get_min_granularity(rdm,
866                                                                 section->mr);
867 
868     g_assert(vrdl->granularity && is_power_of_2(vrdl->granularity));
869     g_assert(container->pgsizes &&
870              vrdl->granularity >= 1ULL << ctz64(container->pgsizes));
871 
872     ram_discard_listener_init(&vrdl->listener,
873                               vfio_ram_discard_notify_populate,
874                               vfio_ram_discard_notify_discard, true);
875     ram_discard_manager_register_listener(rdm, &vrdl->listener, section);
876     QLIST_INSERT_HEAD(&container->vrdl_list, vrdl, next);
877 
878     /*
879      * Sanity-check if we have a theoretically problematic setup where we could
880      * exceed the maximum number of possible DMA mappings over time. We assume
881      * that each mapped section in the same address space as a RamDiscardManager
882      * section consumes exactly one DMA mapping, with the exception of
883      * RamDiscardManager sections; i.e., we don't expect to have gIOMMU sections
884      * in the same address space as RamDiscardManager sections.
885      *
886      * We assume that each section in the address space consumes one memslot.
887      * We take the number of KVM memory slots as a best guess for the maximum
888      * number of sections in the address space we could have over time,
889      * also consuming DMA mappings.
890      */
891     if (container->dma_max_mappings) {
892         unsigned int vrdl_count = 0, vrdl_mappings = 0, max_memslots = 512;
893 
894 #ifdef CONFIG_KVM
895         if (kvm_enabled()) {
896             max_memslots = kvm_get_max_memslots();
897         }
898 #endif
899 
900         QLIST_FOREACH(vrdl, &container->vrdl_list, next) {
901             hwaddr start, end;
902 
903             start = QEMU_ALIGN_DOWN(vrdl->offset_within_address_space,
904                                     vrdl->granularity);
905             end = ROUND_UP(vrdl->offset_within_address_space + vrdl->size,
906                            vrdl->granularity);
907             vrdl_mappings += (end - start) / vrdl->granularity;
908             vrdl_count++;
909         }
910 
911         if (vrdl_mappings + max_memslots - vrdl_count >
912             container->dma_max_mappings) {
913             warn_report("%s: possibly running out of DMA mappings. E.g., try"
914                         " increasing the 'block-size' of virtio-mem devies."
915                         " Maximum possible DMA mappings: %d, Maximum possible"
916                         " memslots: %d", __func__, container->dma_max_mappings,
917                         max_memslots);
918         }
919     }
920 }
921 
922 static void vfio_unregister_ram_discard_listener(VFIOContainer *container,
923                                                  MemoryRegionSection *section)
924 {
925     RamDiscardManager *rdm = memory_region_get_ram_discard_manager(section->mr);
926     VFIORamDiscardListener *vrdl = NULL;
927 
928     QLIST_FOREACH(vrdl, &container->vrdl_list, next) {
929         if (vrdl->mr == section->mr &&
930             vrdl->offset_within_address_space ==
931             section->offset_within_address_space) {
932             break;
933         }
934     }
935 
936     if (!vrdl) {
937         hw_error("vfio: Trying to unregister missing RAM discard listener");
938     }
939 
940     ram_discard_manager_unregister_listener(rdm, &vrdl->listener);
941     QLIST_REMOVE(vrdl, next);
942     g_free(vrdl);
943 }
944 
945 static VFIOHostDMAWindow *vfio_find_hostwin(VFIOContainer *container,
946                                             hwaddr iova, hwaddr end)
947 {
948     VFIOHostDMAWindow *hostwin;
949     bool hostwin_found = false;
950 
951     QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) {
952         if (hostwin->min_iova <= iova && end <= hostwin->max_iova) {
953             hostwin_found = true;
954             break;
955         }
956     }
957 
958     return hostwin_found ? hostwin : NULL;
959 }
960 
961 static bool vfio_known_safe_misalignment(MemoryRegionSection *section)
962 {
963     MemoryRegion *mr = section->mr;
964 
965     if (!TPM_IS_CRB(mr->owner)) {
966         return false;
967     }
968 
969     /* this is a known safe misaligned region, just trace for debug purpose */
970     trace_vfio_known_safe_misalignment(memory_region_name(mr),
971                                        section->offset_within_address_space,
972                                        section->offset_within_region,
973                                        qemu_real_host_page_size());
974     return true;
975 }
976 
977 static bool vfio_listener_valid_section(MemoryRegionSection *section,
978                                         const char *name)
979 {
980     if (vfio_listener_skipped_section(section)) {
981         trace_vfio_listener_region_skip(name,
982                 section->offset_within_address_space,
983                 section->offset_within_address_space +
984                 int128_get64(int128_sub(section->size, int128_one())));
985         return false;
986     }
987 
988     if (unlikely((section->offset_within_address_space &
989                   ~qemu_real_host_page_mask()) !=
990                  (section->offset_within_region & ~qemu_real_host_page_mask()))) {
991         if (!vfio_known_safe_misalignment(section)) {
992             error_report("%s received unaligned region %s iova=0x%"PRIx64
993                          " offset_within_region=0x%"PRIx64
994                          " qemu_real_host_page_size=0x%"PRIxPTR,
995                          __func__, memory_region_name(section->mr),
996                          section->offset_within_address_space,
997                          section->offset_within_region,
998                          qemu_real_host_page_size());
999         }
1000         return false;
1001     }
1002 
1003     return true;
1004 }
1005 
1006 static bool vfio_get_section_iova_range(VFIOContainer *container,
1007                                         MemoryRegionSection *section,
1008                                         hwaddr *out_iova, hwaddr *out_end,
1009                                         Int128 *out_llend)
1010 {
1011     Int128 llend;
1012     hwaddr iova;
1013 
1014     iova = REAL_HOST_PAGE_ALIGN(section->offset_within_address_space);
1015     llend = int128_make64(section->offset_within_address_space);
1016     llend = int128_add(llend, section->size);
1017     llend = int128_and(llend, int128_exts64(qemu_real_host_page_mask()));
1018 
1019     if (int128_ge(int128_make64(iova), llend)) {
1020         return false;
1021     }
1022 
1023     *out_iova = iova;
1024     *out_end = int128_get64(int128_sub(llend, int128_one()));
1025     if (out_llend) {
1026         *out_llend = llend;
1027     }
1028     return true;
1029 }
1030 
1031 static void vfio_listener_region_add(MemoryListener *listener,
1032                                      MemoryRegionSection *section)
1033 {
1034     VFIOContainer *container = container_of(listener, VFIOContainer, listener);
1035     hwaddr iova, end;
1036     Int128 llend, llsize;
1037     void *vaddr;
1038     int ret;
1039     VFIOHostDMAWindow *hostwin;
1040     Error *err = NULL;
1041 
1042     if (!vfio_listener_valid_section(section, "region_add")) {
1043         return;
1044     }
1045 
1046     if (!vfio_get_section_iova_range(container, section, &iova, &end, &llend)) {
1047         if (memory_region_is_ram_device(section->mr)) {
1048             trace_vfio_listener_region_add_no_dma_map(
1049                 memory_region_name(section->mr),
1050                 section->offset_within_address_space,
1051                 int128_getlo(section->size),
1052                 qemu_real_host_page_size());
1053         }
1054         return;
1055     }
1056 
1057     if (container->iommu_type == VFIO_SPAPR_TCE_v2_IOMMU) {
1058         hwaddr pgsize = 0;
1059 
1060         /* For now intersections are not allowed, we may relax this later */
1061         QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) {
1062             if (ranges_overlap(hostwin->min_iova,
1063                                hostwin->max_iova - hostwin->min_iova + 1,
1064                                section->offset_within_address_space,
1065                                int128_get64(section->size))) {
1066                 error_setg(&err,
1067                     "region [0x%"PRIx64",0x%"PRIx64"] overlaps with existing"
1068                     "host DMA window [0x%"PRIx64",0x%"PRIx64"]",
1069                     section->offset_within_address_space,
1070                     section->offset_within_address_space +
1071                         int128_get64(section->size) - 1,
1072                     hostwin->min_iova, hostwin->max_iova);
1073                 goto fail;
1074             }
1075         }
1076 
1077         ret = vfio_spapr_create_window(container, section, &pgsize);
1078         if (ret) {
1079             error_setg_errno(&err, -ret, "Failed to create SPAPR window");
1080             goto fail;
1081         }
1082 
1083         vfio_host_win_add(container, section->offset_within_address_space,
1084                           section->offset_within_address_space +
1085                           int128_get64(section->size) - 1, pgsize);
1086 #ifdef CONFIG_KVM
1087         if (kvm_enabled()) {
1088             VFIOGroup *group;
1089             IOMMUMemoryRegion *iommu_mr = IOMMU_MEMORY_REGION(section->mr);
1090             struct kvm_vfio_spapr_tce param;
1091             struct kvm_device_attr attr = {
1092                 .group = KVM_DEV_VFIO_GROUP,
1093                 .attr = KVM_DEV_VFIO_GROUP_SET_SPAPR_TCE,
1094                 .addr = (uint64_t)(unsigned long)&param,
1095             };
1096 
1097             if (!memory_region_iommu_get_attr(iommu_mr, IOMMU_ATTR_SPAPR_TCE_FD,
1098                                               &param.tablefd)) {
1099                 QLIST_FOREACH(group, &container->group_list, container_next) {
1100                     param.groupfd = group->fd;
1101                     if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) {
1102                         error_report("vfio: failed to setup fd %d "
1103                                      "for a group with fd %d: %s",
1104                                      param.tablefd, param.groupfd,
1105                                      strerror(errno));
1106                         return;
1107                     }
1108                     trace_vfio_spapr_group_attach(param.groupfd, param.tablefd);
1109                 }
1110             }
1111         }
1112 #endif
1113     }
1114 
1115     hostwin = vfio_find_hostwin(container, iova, end);
1116     if (!hostwin) {
1117         error_setg(&err, "Container %p can't map guest IOVA region"
1118                    " 0x%"HWADDR_PRIx"..0x%"HWADDR_PRIx, container, iova, end);
1119         goto fail;
1120     }
1121 
1122     memory_region_ref(section->mr);
1123 
1124     if (memory_region_is_iommu(section->mr)) {
1125         VFIOGuestIOMMU *giommu;
1126         IOMMUMemoryRegion *iommu_mr = IOMMU_MEMORY_REGION(section->mr);
1127         int iommu_idx;
1128 
1129         trace_vfio_listener_region_add_iommu(iova, end);
1130         /*
1131          * FIXME: For VFIO iommu types which have KVM acceleration to
1132          * avoid bouncing all map/unmaps through qemu this way, this
1133          * would be the right place to wire that up (tell the KVM
1134          * device emulation the VFIO iommu handles to use).
1135          */
1136         giommu = g_malloc0(sizeof(*giommu));
1137         giommu->iommu_mr = iommu_mr;
1138         giommu->iommu_offset = section->offset_within_address_space -
1139                                section->offset_within_region;
1140         giommu->container = container;
1141         llend = int128_add(int128_make64(section->offset_within_region),
1142                            section->size);
1143         llend = int128_sub(llend, int128_one());
1144         iommu_idx = memory_region_iommu_attrs_to_index(iommu_mr,
1145                                                        MEMTXATTRS_UNSPECIFIED);
1146         iommu_notifier_init(&giommu->n, vfio_iommu_map_notify,
1147                             IOMMU_NOTIFIER_IOTLB_EVENTS,
1148                             section->offset_within_region,
1149                             int128_get64(llend),
1150                             iommu_idx);
1151 
1152         ret = memory_region_iommu_set_page_size_mask(giommu->iommu_mr,
1153                                                      container->pgsizes,
1154                                                      &err);
1155         if (ret) {
1156             g_free(giommu);
1157             goto fail;
1158         }
1159 
1160         ret = memory_region_register_iommu_notifier(section->mr, &giommu->n,
1161                                                     &err);
1162         if (ret) {
1163             g_free(giommu);
1164             goto fail;
1165         }
1166         QLIST_INSERT_HEAD(&container->giommu_list, giommu, giommu_next);
1167         memory_region_iommu_replay(giommu->iommu_mr, &giommu->n);
1168 
1169         return;
1170     }
1171 
1172     /* Here we assume that memory_region_is_ram(section->mr)==true */
1173 
1174     /*
1175      * For RAM memory regions with a RamDiscardManager, we only want to map the
1176      * actually populated parts - and update the mapping whenever we're notified
1177      * about changes.
1178      */
1179     if (memory_region_has_ram_discard_manager(section->mr)) {
1180         vfio_register_ram_discard_listener(container, section);
1181         return;
1182     }
1183 
1184     vaddr = memory_region_get_ram_ptr(section->mr) +
1185             section->offset_within_region +
1186             (iova - section->offset_within_address_space);
1187 
1188     trace_vfio_listener_region_add_ram(iova, end, vaddr);
1189 
1190     llsize = int128_sub(llend, int128_make64(iova));
1191 
1192     if (memory_region_is_ram_device(section->mr)) {
1193         hwaddr pgmask = (1ULL << ctz64(hostwin->iova_pgsizes)) - 1;
1194 
1195         if ((iova & pgmask) || (int128_get64(llsize) & pgmask)) {
1196             trace_vfio_listener_region_add_no_dma_map(
1197                 memory_region_name(section->mr),
1198                 section->offset_within_address_space,
1199                 int128_getlo(section->size),
1200                 pgmask + 1);
1201             return;
1202         }
1203     }
1204 
1205     ret = vfio_dma_map(container, iova, int128_get64(llsize),
1206                        vaddr, section->readonly);
1207     if (ret) {
1208         error_setg(&err, "vfio_dma_map(%p, 0x%"HWADDR_PRIx", "
1209                    "0x%"HWADDR_PRIx", %p) = %d (%s)",
1210                    container, iova, int128_get64(llsize), vaddr, ret,
1211                    strerror(-ret));
1212         if (memory_region_is_ram_device(section->mr)) {
1213             /* Allow unexpected mappings not to be fatal for RAM devices */
1214             error_report_err(err);
1215             return;
1216         }
1217         goto fail;
1218     }
1219 
1220     return;
1221 
1222 fail:
1223     if (memory_region_is_ram_device(section->mr)) {
1224         error_report("failed to vfio_dma_map. pci p2p may not work");
1225         return;
1226     }
1227     /*
1228      * On the initfn path, store the first error in the container so we
1229      * can gracefully fail.  Runtime, there's not much we can do other
1230      * than throw a hardware error.
1231      */
1232     if (!container->initialized) {
1233         if (!container->error) {
1234             error_propagate_prepend(&container->error, err,
1235                                     "Region %s: ",
1236                                     memory_region_name(section->mr));
1237         } else {
1238             error_free(err);
1239         }
1240     } else {
1241         error_report_err(err);
1242         hw_error("vfio: DMA mapping failed, unable to continue");
1243     }
1244 }
1245 
1246 static void vfio_listener_region_del(MemoryListener *listener,
1247                                      MemoryRegionSection *section)
1248 {
1249     VFIOContainer *container = container_of(listener, VFIOContainer, listener);
1250     hwaddr iova, end;
1251     Int128 llend, llsize;
1252     int ret;
1253     bool try_unmap = true;
1254 
1255     if (!vfio_listener_valid_section(section, "region_del")) {
1256         return;
1257     }
1258 
1259     if (memory_region_is_iommu(section->mr)) {
1260         VFIOGuestIOMMU *giommu;
1261 
1262         QLIST_FOREACH(giommu, &container->giommu_list, giommu_next) {
1263             if (MEMORY_REGION(giommu->iommu_mr) == section->mr &&
1264                 giommu->n.start == section->offset_within_region) {
1265                 memory_region_unregister_iommu_notifier(section->mr,
1266                                                         &giommu->n);
1267                 QLIST_REMOVE(giommu, giommu_next);
1268                 g_free(giommu);
1269                 break;
1270             }
1271         }
1272 
1273         /*
1274          * FIXME: We assume the one big unmap below is adequate to
1275          * remove any individual page mappings in the IOMMU which
1276          * might have been copied into VFIO. This works for a page table
1277          * based IOMMU where a big unmap flattens a large range of IO-PTEs.
1278          * That may not be true for all IOMMU types.
1279          */
1280     }
1281 
1282     if (!vfio_get_section_iova_range(container, section, &iova, &end, &llend)) {
1283         return;
1284     }
1285 
1286     llsize = int128_sub(llend, int128_make64(iova));
1287 
1288     trace_vfio_listener_region_del(iova, end);
1289 
1290     if (memory_region_is_ram_device(section->mr)) {
1291         hwaddr pgmask;
1292         VFIOHostDMAWindow *hostwin;
1293 
1294         hostwin = vfio_find_hostwin(container, iova, end);
1295         assert(hostwin); /* or region_add() would have failed */
1296 
1297         pgmask = (1ULL << ctz64(hostwin->iova_pgsizes)) - 1;
1298         try_unmap = !((iova & pgmask) || (int128_get64(llsize) & pgmask));
1299     } else if (memory_region_has_ram_discard_manager(section->mr)) {
1300         vfio_unregister_ram_discard_listener(container, section);
1301         /* Unregistering will trigger an unmap. */
1302         try_unmap = false;
1303     }
1304 
1305     if (try_unmap) {
1306         if (int128_eq(llsize, int128_2_64())) {
1307             /* The unmap ioctl doesn't accept a full 64-bit span. */
1308             llsize = int128_rshift(llsize, 1);
1309             ret = vfio_dma_unmap(container, iova, int128_get64(llsize), NULL);
1310             if (ret) {
1311                 error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx", "
1312                              "0x%"HWADDR_PRIx") = %d (%s)",
1313                              container, iova, int128_get64(llsize), ret,
1314                              strerror(-ret));
1315             }
1316             iova += int128_get64(llsize);
1317         }
1318         ret = vfio_dma_unmap(container, iova, int128_get64(llsize), NULL);
1319         if (ret) {
1320             error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx", "
1321                          "0x%"HWADDR_PRIx") = %d (%s)",
1322                          container, iova, int128_get64(llsize), ret,
1323                          strerror(-ret));
1324         }
1325     }
1326 
1327     memory_region_unref(section->mr);
1328 
1329     if (container->iommu_type == VFIO_SPAPR_TCE_v2_IOMMU) {
1330         vfio_spapr_remove_window(container,
1331                                  section->offset_within_address_space);
1332         if (vfio_host_win_del(container,
1333                               section->offset_within_address_space,
1334                               section->offset_within_address_space +
1335                               int128_get64(section->size) - 1) < 0) {
1336             hw_error("%s: Cannot delete missing window at %"HWADDR_PRIx,
1337                      __func__, section->offset_within_address_space);
1338         }
1339     }
1340 }
1341 
1342 static int vfio_set_dirty_page_tracking(VFIOContainer *container, bool start)
1343 {
1344     int ret;
1345     struct vfio_iommu_type1_dirty_bitmap dirty = {
1346         .argsz = sizeof(dirty),
1347     };
1348 
1349     if (!container->dirty_pages_supported) {
1350         return 0;
1351     }
1352 
1353     if (start) {
1354         dirty.flags = VFIO_IOMMU_DIRTY_PAGES_FLAG_START;
1355     } else {
1356         dirty.flags = VFIO_IOMMU_DIRTY_PAGES_FLAG_STOP;
1357     }
1358 
1359     ret = ioctl(container->fd, VFIO_IOMMU_DIRTY_PAGES, &dirty);
1360     if (ret) {
1361         ret = -errno;
1362         error_report("Failed to set dirty tracking flag 0x%x errno: %d",
1363                      dirty.flags, errno);
1364     }
1365 
1366     return ret;
1367 }
1368 
1369 typedef struct VFIODirtyRanges {
1370     hwaddr min32;
1371     hwaddr max32;
1372     hwaddr min64;
1373     hwaddr max64;
1374 } VFIODirtyRanges;
1375 
1376 typedef struct VFIODirtyRangesListener {
1377     VFIOContainer *container;
1378     VFIODirtyRanges ranges;
1379     MemoryListener listener;
1380 } VFIODirtyRangesListener;
1381 
1382 static void vfio_dirty_tracking_update(MemoryListener *listener,
1383                                        MemoryRegionSection *section)
1384 {
1385     VFIODirtyRangesListener *dirty = container_of(listener,
1386                                                   VFIODirtyRangesListener,
1387                                                   listener);
1388     VFIODirtyRanges *range = &dirty->ranges;
1389     hwaddr iova, end, *min, *max;
1390 
1391     if (!vfio_listener_valid_section(section, "tracking_update") ||
1392         !vfio_get_section_iova_range(dirty->container, section,
1393                                      &iova, &end, NULL)) {
1394         return;
1395     }
1396 
1397     /*
1398      * The address space passed to the dirty tracker is reduced to two ranges:
1399      * one for 32-bit DMA ranges, and another one for 64-bit DMA ranges.
1400      * The underlying reports of dirty will query a sub-interval of each of
1401      * these ranges.
1402      *
1403      * The purpose of the dual range handling is to handle known cases of big
1404      * holes in the address space, like the x86 AMD 1T hole. The alternative
1405      * would be an IOVATree but that has a much bigger runtime overhead and
1406      * unnecessary complexity.
1407      */
1408     min = (end <= UINT32_MAX) ? &range->min32 : &range->min64;
1409     max = (end <= UINT32_MAX) ? &range->max32 : &range->max64;
1410 
1411     if (*min > iova) {
1412         *min = iova;
1413     }
1414     if (*max < end) {
1415         *max = end;
1416     }
1417 
1418     trace_vfio_device_dirty_tracking_update(iova, end, *min, *max);
1419     return;
1420 }
1421 
1422 static const MemoryListener vfio_dirty_tracking_listener = {
1423     .name = "vfio-tracking",
1424     .region_add = vfio_dirty_tracking_update,
1425 };
1426 
1427 static void vfio_dirty_tracking_init(VFIOContainer *container,
1428                                      VFIODirtyRanges *ranges)
1429 {
1430     VFIODirtyRangesListener dirty;
1431 
1432     memset(&dirty, 0, sizeof(dirty));
1433     dirty.ranges.min32 = UINT32_MAX;
1434     dirty.ranges.min64 = UINT64_MAX;
1435     dirty.listener = vfio_dirty_tracking_listener;
1436     dirty.container = container;
1437 
1438     memory_listener_register(&dirty.listener,
1439                              container->space->as);
1440 
1441     *ranges = dirty.ranges;
1442 
1443     /*
1444      * The memory listener is synchronous, and used to calculate the range
1445      * to dirty tracking. Unregister it after we are done as we are not
1446      * interested in any follow-up updates.
1447      */
1448     memory_listener_unregister(&dirty.listener);
1449 }
1450 
1451 static void vfio_devices_dma_logging_stop(VFIOContainer *container)
1452 {
1453     uint64_t buf[DIV_ROUND_UP(sizeof(struct vfio_device_feature),
1454                               sizeof(uint64_t))] = {};
1455     struct vfio_device_feature *feature = (struct vfio_device_feature *)buf;
1456     VFIODevice *vbasedev;
1457     VFIOGroup *group;
1458 
1459     feature->argsz = sizeof(buf);
1460     feature->flags = VFIO_DEVICE_FEATURE_SET |
1461                      VFIO_DEVICE_FEATURE_DMA_LOGGING_STOP;
1462 
1463     QLIST_FOREACH(group, &container->group_list, container_next) {
1464         QLIST_FOREACH(vbasedev, &group->device_list, next) {
1465             if (!vbasedev->dirty_tracking) {
1466                 continue;
1467             }
1468 
1469             if (ioctl(vbasedev->fd, VFIO_DEVICE_FEATURE, feature)) {
1470                 warn_report("%s: Failed to stop DMA logging, err %d (%s)",
1471                              vbasedev->name, -errno, strerror(errno));
1472             }
1473             vbasedev->dirty_tracking = false;
1474         }
1475     }
1476 }
1477 
1478 static struct vfio_device_feature *
1479 vfio_device_feature_dma_logging_start_create(VFIOContainer *container,
1480                                              VFIODirtyRanges *tracking)
1481 {
1482     struct vfio_device_feature *feature;
1483     size_t feature_size;
1484     struct vfio_device_feature_dma_logging_control *control;
1485     struct vfio_device_feature_dma_logging_range *ranges;
1486 
1487     feature_size = sizeof(struct vfio_device_feature) +
1488                    sizeof(struct vfio_device_feature_dma_logging_control);
1489     feature = g_try_malloc0(feature_size);
1490     if (!feature) {
1491         errno = ENOMEM;
1492         return NULL;
1493     }
1494     feature->argsz = feature_size;
1495     feature->flags = VFIO_DEVICE_FEATURE_SET |
1496                      VFIO_DEVICE_FEATURE_DMA_LOGGING_START;
1497 
1498     control = (struct vfio_device_feature_dma_logging_control *)feature->data;
1499     control->page_size = qemu_real_host_page_size();
1500 
1501     /*
1502      * DMA logging uAPI guarantees to support at least a number of ranges that
1503      * fits into a single host kernel base page.
1504      */
1505     control->num_ranges = !!tracking->max32 + !!tracking->max64;
1506     ranges = g_try_new0(struct vfio_device_feature_dma_logging_range,
1507                         control->num_ranges);
1508     if (!ranges) {
1509         g_free(feature);
1510         errno = ENOMEM;
1511 
1512         return NULL;
1513     }
1514 
1515     control->ranges = (__u64)(uintptr_t)ranges;
1516     if (tracking->max32) {
1517         ranges->iova = tracking->min32;
1518         ranges->length = (tracking->max32 - tracking->min32) + 1;
1519         ranges++;
1520     }
1521     if (tracking->max64) {
1522         ranges->iova = tracking->min64;
1523         ranges->length = (tracking->max64 - tracking->min64) + 1;
1524     }
1525 
1526     trace_vfio_device_dirty_tracking_start(control->num_ranges,
1527                                            tracking->min32, tracking->max32,
1528                                            tracking->min64, tracking->max64);
1529 
1530     return feature;
1531 }
1532 
1533 static void vfio_device_feature_dma_logging_start_destroy(
1534     struct vfio_device_feature *feature)
1535 {
1536     struct vfio_device_feature_dma_logging_control *control =
1537         (struct vfio_device_feature_dma_logging_control *)feature->data;
1538     struct vfio_device_feature_dma_logging_range *ranges =
1539         (struct vfio_device_feature_dma_logging_range *)(uintptr_t)control->ranges;
1540 
1541     g_free(ranges);
1542     g_free(feature);
1543 }
1544 
1545 static int vfio_devices_dma_logging_start(VFIOContainer *container)
1546 {
1547     struct vfio_device_feature *feature;
1548     VFIODirtyRanges ranges;
1549     VFIODevice *vbasedev;
1550     VFIOGroup *group;
1551     int ret = 0;
1552 
1553     vfio_dirty_tracking_init(container, &ranges);
1554     feature = vfio_device_feature_dma_logging_start_create(container,
1555                                                            &ranges);
1556     if (!feature) {
1557         return -errno;
1558     }
1559 
1560     QLIST_FOREACH(group, &container->group_list, container_next) {
1561         QLIST_FOREACH(vbasedev, &group->device_list, next) {
1562             if (vbasedev->dirty_tracking) {
1563                 continue;
1564             }
1565 
1566             ret = ioctl(vbasedev->fd, VFIO_DEVICE_FEATURE, feature);
1567             if (ret) {
1568                 ret = -errno;
1569                 error_report("%s: Failed to start DMA logging, err %d (%s)",
1570                              vbasedev->name, ret, strerror(errno));
1571                 goto out;
1572             }
1573             vbasedev->dirty_tracking = true;
1574         }
1575     }
1576 
1577 out:
1578     if (ret) {
1579         vfio_devices_dma_logging_stop(container);
1580     }
1581 
1582     vfio_device_feature_dma_logging_start_destroy(feature);
1583 
1584     return ret;
1585 }
1586 
1587 static void vfio_listener_log_global_start(MemoryListener *listener)
1588 {
1589     VFIOContainer *container = container_of(listener, VFIOContainer, listener);
1590     int ret;
1591 
1592     if (vfio_devices_all_device_dirty_tracking(container)) {
1593         ret = vfio_devices_dma_logging_start(container);
1594     } else {
1595         ret = vfio_set_dirty_page_tracking(container, true);
1596     }
1597 
1598     if (ret) {
1599         error_report("vfio: Could not start dirty page tracking, err: %d (%s)",
1600                      ret, strerror(-ret));
1601         vfio_set_migration_error(ret);
1602     }
1603 }
1604 
1605 static void vfio_listener_log_global_stop(MemoryListener *listener)
1606 {
1607     VFIOContainer *container = container_of(listener, VFIOContainer, listener);
1608     int ret = 0;
1609 
1610     if (vfio_devices_all_device_dirty_tracking(container)) {
1611         vfio_devices_dma_logging_stop(container);
1612     } else {
1613         ret = vfio_set_dirty_page_tracking(container, false);
1614     }
1615 
1616     if (ret) {
1617         error_report("vfio: Could not stop dirty page tracking, err: %d (%s)",
1618                      ret, strerror(-ret));
1619         vfio_set_migration_error(ret);
1620     }
1621 }
1622 
1623 static int vfio_device_dma_logging_report(VFIODevice *vbasedev, hwaddr iova,
1624                                           hwaddr size, void *bitmap)
1625 {
1626     uint64_t buf[DIV_ROUND_UP(sizeof(struct vfio_device_feature) +
1627                         sizeof(struct vfio_device_feature_dma_logging_report),
1628                         sizeof(__u64))] = {};
1629     struct vfio_device_feature *feature = (struct vfio_device_feature *)buf;
1630     struct vfio_device_feature_dma_logging_report *report =
1631         (struct vfio_device_feature_dma_logging_report *)feature->data;
1632 
1633     report->iova = iova;
1634     report->length = size;
1635     report->page_size = qemu_real_host_page_size();
1636     report->bitmap = (__u64)(uintptr_t)bitmap;
1637 
1638     feature->argsz = sizeof(buf);
1639     feature->flags = VFIO_DEVICE_FEATURE_GET |
1640                      VFIO_DEVICE_FEATURE_DMA_LOGGING_REPORT;
1641 
1642     if (ioctl(vbasedev->fd, VFIO_DEVICE_FEATURE, feature)) {
1643         return -errno;
1644     }
1645 
1646     return 0;
1647 }
1648 
1649 static int vfio_devices_query_dirty_bitmap(VFIOContainer *container,
1650                                            VFIOBitmap *vbmap, hwaddr iova,
1651                                            hwaddr size)
1652 {
1653     VFIODevice *vbasedev;
1654     VFIOGroup *group;
1655     int ret;
1656 
1657     QLIST_FOREACH(group, &container->group_list, container_next) {
1658         QLIST_FOREACH(vbasedev, &group->device_list, next) {
1659             ret = vfio_device_dma_logging_report(vbasedev, iova, size,
1660                                                  vbmap->bitmap);
1661             if (ret) {
1662                 error_report("%s: Failed to get DMA logging report, iova: "
1663                              "0x%" HWADDR_PRIx ", size: 0x%" HWADDR_PRIx
1664                              ", err: %d (%s)",
1665                              vbasedev->name, iova, size, ret, strerror(-ret));
1666 
1667                 return ret;
1668             }
1669         }
1670     }
1671 
1672     return 0;
1673 }
1674 
1675 static int vfio_query_dirty_bitmap(VFIOContainer *container, VFIOBitmap *vbmap,
1676                                    hwaddr iova, hwaddr size)
1677 {
1678     struct vfio_iommu_type1_dirty_bitmap *dbitmap;
1679     struct vfio_iommu_type1_dirty_bitmap_get *range;
1680     int ret;
1681 
1682     dbitmap = g_malloc0(sizeof(*dbitmap) + sizeof(*range));
1683 
1684     dbitmap->argsz = sizeof(*dbitmap) + sizeof(*range);
1685     dbitmap->flags = VFIO_IOMMU_DIRTY_PAGES_FLAG_GET_BITMAP;
1686     range = (struct vfio_iommu_type1_dirty_bitmap_get *)&dbitmap->data;
1687     range->iova = iova;
1688     range->size = size;
1689 
1690     /*
1691      * cpu_physical_memory_set_dirty_lebitmap() supports pages in bitmap of
1692      * qemu_real_host_page_size to mark those dirty. Hence set bitmap's pgsize
1693      * to qemu_real_host_page_size.
1694      */
1695     range->bitmap.pgsize = qemu_real_host_page_size();
1696     range->bitmap.size = vbmap->size;
1697     range->bitmap.data = (__u64 *)vbmap->bitmap;
1698 
1699     ret = ioctl(container->fd, VFIO_IOMMU_DIRTY_PAGES, dbitmap);
1700     if (ret) {
1701         ret = -errno;
1702         error_report("Failed to get dirty bitmap for iova: 0x%"PRIx64
1703                 " size: 0x%"PRIx64" err: %d", (uint64_t)range->iova,
1704                 (uint64_t)range->size, errno);
1705     }
1706 
1707     g_free(dbitmap);
1708 
1709     return ret;
1710 }
1711 
1712 static int vfio_get_dirty_bitmap(VFIOContainer *container, uint64_t iova,
1713                                  uint64_t size, ram_addr_t ram_addr)
1714 {
1715     bool all_device_dirty_tracking =
1716         vfio_devices_all_device_dirty_tracking(container);
1717     uint64_t dirty_pages;
1718     VFIOBitmap vbmap;
1719     int ret;
1720 
1721     if (!container->dirty_pages_supported && !all_device_dirty_tracking) {
1722         cpu_physical_memory_set_dirty_range(ram_addr, size,
1723                                             tcg_enabled() ? DIRTY_CLIENTS_ALL :
1724                                             DIRTY_CLIENTS_NOCODE);
1725         return 0;
1726     }
1727 
1728     ret = vfio_bitmap_alloc(&vbmap, size);
1729     if (ret) {
1730         return ret;
1731     }
1732 
1733     if (all_device_dirty_tracking) {
1734         ret = vfio_devices_query_dirty_bitmap(container, &vbmap, iova, size);
1735     } else {
1736         ret = vfio_query_dirty_bitmap(container, &vbmap, iova, size);
1737     }
1738 
1739     if (ret) {
1740         goto out;
1741     }
1742 
1743     dirty_pages = cpu_physical_memory_set_dirty_lebitmap(vbmap.bitmap, ram_addr,
1744                                                          vbmap.pages);
1745 
1746     trace_vfio_get_dirty_bitmap(container->fd, iova, size, vbmap.size,
1747                                 ram_addr, dirty_pages);
1748 out:
1749     g_free(vbmap.bitmap);
1750 
1751     return ret;
1752 }
1753 
1754 typedef struct {
1755     IOMMUNotifier n;
1756     VFIOGuestIOMMU *giommu;
1757 } vfio_giommu_dirty_notifier;
1758 
1759 static void vfio_iommu_map_dirty_notify(IOMMUNotifier *n, IOMMUTLBEntry *iotlb)
1760 {
1761     vfio_giommu_dirty_notifier *gdn = container_of(n,
1762                                                 vfio_giommu_dirty_notifier, n);
1763     VFIOGuestIOMMU *giommu = gdn->giommu;
1764     VFIOContainer *container = giommu->container;
1765     hwaddr iova = iotlb->iova + giommu->iommu_offset;
1766     ram_addr_t translated_addr;
1767     int ret = -EINVAL;
1768 
1769     trace_vfio_iommu_map_dirty_notify(iova, iova + iotlb->addr_mask);
1770 
1771     if (iotlb->target_as != &address_space_memory) {
1772         error_report("Wrong target AS \"%s\", only system memory is allowed",
1773                      iotlb->target_as->name ? iotlb->target_as->name : "none");
1774         goto out;
1775     }
1776 
1777     rcu_read_lock();
1778     if (vfio_get_xlat_addr(iotlb, NULL, &translated_addr, NULL)) {
1779         ret = vfio_get_dirty_bitmap(container, iova, iotlb->addr_mask + 1,
1780                                     translated_addr);
1781         if (ret) {
1782             error_report("vfio_iommu_map_dirty_notify(%p, 0x%"HWADDR_PRIx", "
1783                          "0x%"HWADDR_PRIx") = %d (%s)",
1784                          container, iova, iotlb->addr_mask + 1, ret,
1785                          strerror(-ret));
1786         }
1787     }
1788     rcu_read_unlock();
1789 
1790 out:
1791     if (ret) {
1792         vfio_set_migration_error(ret);
1793     }
1794 }
1795 
1796 static int vfio_ram_discard_get_dirty_bitmap(MemoryRegionSection *section,
1797                                              void *opaque)
1798 {
1799     const hwaddr size = int128_get64(section->size);
1800     const hwaddr iova = section->offset_within_address_space;
1801     const ram_addr_t ram_addr = memory_region_get_ram_addr(section->mr) +
1802                                 section->offset_within_region;
1803     VFIORamDiscardListener *vrdl = opaque;
1804 
1805     /*
1806      * Sync the whole mapped region (spanning multiple individual mappings)
1807      * in one go.
1808      */
1809     return vfio_get_dirty_bitmap(vrdl->container, iova, size, ram_addr);
1810 }
1811 
1812 static int vfio_sync_ram_discard_listener_dirty_bitmap(VFIOContainer *container,
1813                                                    MemoryRegionSection *section)
1814 {
1815     RamDiscardManager *rdm = memory_region_get_ram_discard_manager(section->mr);
1816     VFIORamDiscardListener *vrdl = NULL;
1817 
1818     QLIST_FOREACH(vrdl, &container->vrdl_list, next) {
1819         if (vrdl->mr == section->mr &&
1820             vrdl->offset_within_address_space ==
1821             section->offset_within_address_space) {
1822             break;
1823         }
1824     }
1825 
1826     if (!vrdl) {
1827         hw_error("vfio: Trying to sync missing RAM discard listener");
1828     }
1829 
1830     /*
1831      * We only want/can synchronize the bitmap for actually mapped parts -
1832      * which correspond to populated parts. Replay all populated parts.
1833      */
1834     return ram_discard_manager_replay_populated(rdm, section,
1835                                               vfio_ram_discard_get_dirty_bitmap,
1836                                                 &vrdl);
1837 }
1838 
1839 static int vfio_sync_dirty_bitmap(VFIOContainer *container,
1840                                   MemoryRegionSection *section)
1841 {
1842     ram_addr_t ram_addr;
1843 
1844     if (memory_region_is_iommu(section->mr)) {
1845         VFIOGuestIOMMU *giommu;
1846 
1847         QLIST_FOREACH(giommu, &container->giommu_list, giommu_next) {
1848             if (MEMORY_REGION(giommu->iommu_mr) == section->mr &&
1849                 giommu->n.start == section->offset_within_region) {
1850                 Int128 llend;
1851                 vfio_giommu_dirty_notifier gdn = { .giommu = giommu };
1852                 int idx = memory_region_iommu_attrs_to_index(giommu->iommu_mr,
1853                                                        MEMTXATTRS_UNSPECIFIED);
1854 
1855                 llend = int128_add(int128_make64(section->offset_within_region),
1856                                    section->size);
1857                 llend = int128_sub(llend, int128_one());
1858 
1859                 iommu_notifier_init(&gdn.n,
1860                                     vfio_iommu_map_dirty_notify,
1861                                     IOMMU_NOTIFIER_MAP,
1862                                     section->offset_within_region,
1863                                     int128_get64(llend),
1864                                     idx);
1865                 memory_region_iommu_replay(giommu->iommu_mr, &gdn.n);
1866                 break;
1867             }
1868         }
1869         return 0;
1870     } else if (memory_region_has_ram_discard_manager(section->mr)) {
1871         return vfio_sync_ram_discard_listener_dirty_bitmap(container, section);
1872     }
1873 
1874     ram_addr = memory_region_get_ram_addr(section->mr) +
1875                section->offset_within_region;
1876 
1877     return vfio_get_dirty_bitmap(container,
1878                    REAL_HOST_PAGE_ALIGN(section->offset_within_address_space),
1879                    int128_get64(section->size), ram_addr);
1880 }
1881 
1882 static void vfio_listener_log_sync(MemoryListener *listener,
1883         MemoryRegionSection *section)
1884 {
1885     VFIOContainer *container = container_of(listener, VFIOContainer, listener);
1886     int ret;
1887 
1888     if (vfio_listener_skipped_section(section)) {
1889         return;
1890     }
1891 
1892     if (vfio_devices_all_dirty_tracking(container)) {
1893         ret = vfio_sync_dirty_bitmap(container, section);
1894         if (ret) {
1895             error_report("vfio: Failed to sync dirty bitmap, err: %d (%s)", ret,
1896                          strerror(-ret));
1897             vfio_set_migration_error(ret);
1898         }
1899     }
1900 }
1901 
1902 static const MemoryListener vfio_memory_listener = {
1903     .name = "vfio",
1904     .region_add = vfio_listener_region_add,
1905     .region_del = vfio_listener_region_del,
1906     .log_global_start = vfio_listener_log_global_start,
1907     .log_global_stop = vfio_listener_log_global_stop,
1908     .log_sync = vfio_listener_log_sync,
1909 };
1910 
1911 static void vfio_listener_release(VFIOContainer *container)
1912 {
1913     memory_listener_unregister(&container->listener);
1914     if (container->iommu_type == VFIO_SPAPR_TCE_v2_IOMMU) {
1915         memory_listener_unregister(&container->prereg_listener);
1916     }
1917 }
1918 
1919 static struct vfio_info_cap_header *
1920 vfio_get_cap(void *ptr, uint32_t cap_offset, uint16_t id)
1921 {
1922     struct vfio_info_cap_header *hdr;
1923 
1924     for (hdr = ptr + cap_offset; hdr != ptr; hdr = ptr + hdr->next) {
1925         if (hdr->id == id) {
1926             return hdr;
1927         }
1928     }
1929 
1930     return NULL;
1931 }
1932 
1933 struct vfio_info_cap_header *
1934 vfio_get_region_info_cap(struct vfio_region_info *info, uint16_t id)
1935 {
1936     if (!(info->flags & VFIO_REGION_INFO_FLAG_CAPS)) {
1937         return NULL;
1938     }
1939 
1940     return vfio_get_cap((void *)info, info->cap_offset, id);
1941 }
1942 
1943 static struct vfio_info_cap_header *
1944 vfio_get_iommu_type1_info_cap(struct vfio_iommu_type1_info *info, uint16_t id)
1945 {
1946     if (!(info->flags & VFIO_IOMMU_INFO_CAPS)) {
1947         return NULL;
1948     }
1949 
1950     return vfio_get_cap((void *)info, info->cap_offset, id);
1951 }
1952 
1953 struct vfio_info_cap_header *
1954 vfio_get_device_info_cap(struct vfio_device_info *info, uint16_t id)
1955 {
1956     if (!(info->flags & VFIO_DEVICE_FLAGS_CAPS)) {
1957         return NULL;
1958     }
1959 
1960     return vfio_get_cap((void *)info, info->cap_offset, id);
1961 }
1962 
1963 bool vfio_get_info_dma_avail(struct vfio_iommu_type1_info *info,
1964                              unsigned int *avail)
1965 {
1966     struct vfio_info_cap_header *hdr;
1967     struct vfio_iommu_type1_info_dma_avail *cap;
1968 
1969     /* If the capability cannot be found, assume no DMA limiting */
1970     hdr = vfio_get_iommu_type1_info_cap(info,
1971                                         VFIO_IOMMU_TYPE1_INFO_DMA_AVAIL);
1972     if (hdr == NULL) {
1973         return false;
1974     }
1975 
1976     if (avail != NULL) {
1977         cap = (void *) hdr;
1978         *avail = cap->avail;
1979     }
1980 
1981     return true;
1982 }
1983 
1984 static int vfio_setup_region_sparse_mmaps(VFIORegion *region,
1985                                           struct vfio_region_info *info)
1986 {
1987     struct vfio_info_cap_header *hdr;
1988     struct vfio_region_info_cap_sparse_mmap *sparse;
1989     int i, j;
1990 
1991     hdr = vfio_get_region_info_cap(info, VFIO_REGION_INFO_CAP_SPARSE_MMAP);
1992     if (!hdr) {
1993         return -ENODEV;
1994     }
1995 
1996     sparse = container_of(hdr, struct vfio_region_info_cap_sparse_mmap, header);
1997 
1998     trace_vfio_region_sparse_mmap_header(region->vbasedev->name,
1999                                          region->nr, sparse->nr_areas);
2000 
2001     region->mmaps = g_new0(VFIOMmap, sparse->nr_areas);
2002 
2003     for (i = 0, j = 0; i < sparse->nr_areas; i++) {
2004         if (sparse->areas[i].size) {
2005             trace_vfio_region_sparse_mmap_entry(i, sparse->areas[i].offset,
2006                                             sparse->areas[i].offset +
2007                                             sparse->areas[i].size - 1);
2008             region->mmaps[j].offset = sparse->areas[i].offset;
2009             region->mmaps[j].size = sparse->areas[i].size;
2010             j++;
2011         }
2012     }
2013 
2014     region->nr_mmaps = j;
2015     region->mmaps = g_realloc(region->mmaps, j * sizeof(VFIOMmap));
2016 
2017     return 0;
2018 }
2019 
2020 int vfio_region_setup(Object *obj, VFIODevice *vbasedev, VFIORegion *region,
2021                       int index, const char *name)
2022 {
2023     struct vfio_region_info *info;
2024     int ret;
2025 
2026     ret = vfio_get_region_info(vbasedev, index, &info);
2027     if (ret) {
2028         return ret;
2029     }
2030 
2031     region->vbasedev = vbasedev;
2032     region->flags = info->flags;
2033     region->size = info->size;
2034     region->fd_offset = info->offset;
2035     region->nr = index;
2036 
2037     if (region->size) {
2038         region->mem = g_new0(MemoryRegion, 1);
2039         memory_region_init_io(region->mem, obj, &vfio_region_ops,
2040                               region, name, region->size);
2041 
2042         if (!vbasedev->no_mmap &&
2043             region->flags & VFIO_REGION_INFO_FLAG_MMAP) {
2044 
2045             ret = vfio_setup_region_sparse_mmaps(region, info);
2046 
2047             if (ret) {
2048                 region->nr_mmaps = 1;
2049                 region->mmaps = g_new0(VFIOMmap, region->nr_mmaps);
2050                 region->mmaps[0].offset = 0;
2051                 region->mmaps[0].size = region->size;
2052             }
2053         }
2054     }
2055 
2056     g_free(info);
2057 
2058     trace_vfio_region_setup(vbasedev->name, index, name,
2059                             region->flags, region->fd_offset, region->size);
2060     return 0;
2061 }
2062 
2063 static void vfio_subregion_unmap(VFIORegion *region, int index)
2064 {
2065     trace_vfio_region_unmap(memory_region_name(&region->mmaps[index].mem),
2066                             region->mmaps[index].offset,
2067                             region->mmaps[index].offset +
2068                             region->mmaps[index].size - 1);
2069     memory_region_del_subregion(region->mem, &region->mmaps[index].mem);
2070     munmap(region->mmaps[index].mmap, region->mmaps[index].size);
2071     object_unparent(OBJECT(&region->mmaps[index].mem));
2072     region->mmaps[index].mmap = NULL;
2073 }
2074 
2075 int vfio_region_mmap(VFIORegion *region)
2076 {
2077     int i, prot = 0;
2078     char *name;
2079 
2080     if (!region->mem) {
2081         return 0;
2082     }
2083 
2084     prot |= region->flags & VFIO_REGION_INFO_FLAG_READ ? PROT_READ : 0;
2085     prot |= region->flags & VFIO_REGION_INFO_FLAG_WRITE ? PROT_WRITE : 0;
2086 
2087     for (i = 0; i < region->nr_mmaps; i++) {
2088         region->mmaps[i].mmap = mmap(NULL, region->mmaps[i].size, prot,
2089                                      MAP_SHARED, region->vbasedev->fd,
2090                                      region->fd_offset +
2091                                      region->mmaps[i].offset);
2092         if (region->mmaps[i].mmap == MAP_FAILED) {
2093             int ret = -errno;
2094 
2095             trace_vfio_region_mmap_fault(memory_region_name(region->mem), i,
2096                                          region->fd_offset +
2097                                          region->mmaps[i].offset,
2098                                          region->fd_offset +
2099                                          region->mmaps[i].offset +
2100                                          region->mmaps[i].size - 1, ret);
2101 
2102             region->mmaps[i].mmap = NULL;
2103 
2104             for (i--; i >= 0; i--) {
2105                 vfio_subregion_unmap(region, i);
2106             }
2107 
2108             return ret;
2109         }
2110 
2111         name = g_strdup_printf("%s mmaps[%d]",
2112                                memory_region_name(region->mem), i);
2113         memory_region_init_ram_device_ptr(&region->mmaps[i].mem,
2114                                           memory_region_owner(region->mem),
2115                                           name, region->mmaps[i].size,
2116                                           region->mmaps[i].mmap);
2117         g_free(name);
2118         memory_region_add_subregion(region->mem, region->mmaps[i].offset,
2119                                     &region->mmaps[i].mem);
2120 
2121         trace_vfio_region_mmap(memory_region_name(&region->mmaps[i].mem),
2122                                region->mmaps[i].offset,
2123                                region->mmaps[i].offset +
2124                                region->mmaps[i].size - 1);
2125     }
2126 
2127     return 0;
2128 }
2129 
2130 void vfio_region_unmap(VFIORegion *region)
2131 {
2132     int i;
2133 
2134     if (!region->mem) {
2135         return;
2136     }
2137 
2138     for (i = 0; i < region->nr_mmaps; i++) {
2139         if (region->mmaps[i].mmap) {
2140             vfio_subregion_unmap(region, i);
2141         }
2142     }
2143 }
2144 
2145 void vfio_region_exit(VFIORegion *region)
2146 {
2147     int i;
2148 
2149     if (!region->mem) {
2150         return;
2151     }
2152 
2153     for (i = 0; i < region->nr_mmaps; i++) {
2154         if (region->mmaps[i].mmap) {
2155             memory_region_del_subregion(region->mem, &region->mmaps[i].mem);
2156         }
2157     }
2158 
2159     trace_vfio_region_exit(region->vbasedev->name, region->nr);
2160 }
2161 
2162 void vfio_region_finalize(VFIORegion *region)
2163 {
2164     int i;
2165 
2166     if (!region->mem) {
2167         return;
2168     }
2169 
2170     for (i = 0; i < region->nr_mmaps; i++) {
2171         if (region->mmaps[i].mmap) {
2172             munmap(region->mmaps[i].mmap, region->mmaps[i].size);
2173             object_unparent(OBJECT(&region->mmaps[i].mem));
2174         }
2175     }
2176 
2177     object_unparent(OBJECT(region->mem));
2178 
2179     g_free(region->mem);
2180     g_free(region->mmaps);
2181 
2182     trace_vfio_region_finalize(region->vbasedev->name, region->nr);
2183 
2184     region->mem = NULL;
2185     region->mmaps = NULL;
2186     region->nr_mmaps = 0;
2187     region->size = 0;
2188     region->flags = 0;
2189     region->nr = 0;
2190 }
2191 
2192 void vfio_region_mmaps_set_enabled(VFIORegion *region, bool enabled)
2193 {
2194     int i;
2195 
2196     if (!region->mem) {
2197         return;
2198     }
2199 
2200     for (i = 0; i < region->nr_mmaps; i++) {
2201         if (region->mmaps[i].mmap) {
2202             memory_region_set_enabled(&region->mmaps[i].mem, enabled);
2203         }
2204     }
2205 
2206     trace_vfio_region_mmaps_set_enabled(memory_region_name(region->mem),
2207                                         enabled);
2208 }
2209 
2210 void vfio_reset_handler(void *opaque)
2211 {
2212     VFIOGroup *group;
2213     VFIODevice *vbasedev;
2214 
2215     QLIST_FOREACH(group, &vfio_group_list, next) {
2216         QLIST_FOREACH(vbasedev, &group->device_list, next) {
2217             if (vbasedev->dev->realized) {
2218                 vbasedev->ops->vfio_compute_needs_reset(vbasedev);
2219             }
2220         }
2221     }
2222 
2223     QLIST_FOREACH(group, &vfio_group_list, next) {
2224         QLIST_FOREACH(vbasedev, &group->device_list, next) {
2225             if (vbasedev->dev->realized && vbasedev->needs_reset) {
2226                 vbasedev->ops->vfio_hot_reset_multi(vbasedev);
2227             }
2228         }
2229     }
2230 }
2231 
2232 static void vfio_kvm_device_add_group(VFIOGroup *group)
2233 {
2234 #ifdef CONFIG_KVM
2235     struct kvm_device_attr attr = {
2236         .group = KVM_DEV_VFIO_GROUP,
2237         .attr = KVM_DEV_VFIO_GROUP_ADD,
2238         .addr = (uint64_t)(unsigned long)&group->fd,
2239     };
2240 
2241     if (!kvm_enabled()) {
2242         return;
2243     }
2244 
2245     if (vfio_kvm_device_fd < 0) {
2246         struct kvm_create_device cd = {
2247             .type = KVM_DEV_TYPE_VFIO,
2248         };
2249 
2250         if (kvm_vm_ioctl(kvm_state, KVM_CREATE_DEVICE, &cd)) {
2251             error_report("Failed to create KVM VFIO device: %m");
2252             return;
2253         }
2254 
2255         vfio_kvm_device_fd = cd.fd;
2256     }
2257 
2258     if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) {
2259         error_report("Failed to add group %d to KVM VFIO device: %m",
2260                      group->groupid);
2261     }
2262 #endif
2263 }
2264 
2265 static void vfio_kvm_device_del_group(VFIOGroup *group)
2266 {
2267 #ifdef CONFIG_KVM
2268     struct kvm_device_attr attr = {
2269         .group = KVM_DEV_VFIO_GROUP,
2270         .attr = KVM_DEV_VFIO_GROUP_DEL,
2271         .addr = (uint64_t)(unsigned long)&group->fd,
2272     };
2273 
2274     if (vfio_kvm_device_fd < 0) {
2275         return;
2276     }
2277 
2278     if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) {
2279         error_report("Failed to remove group %d from KVM VFIO device: %m",
2280                      group->groupid);
2281     }
2282 #endif
2283 }
2284 
2285 static VFIOAddressSpace *vfio_get_address_space(AddressSpace *as)
2286 {
2287     VFIOAddressSpace *space;
2288 
2289     QLIST_FOREACH(space, &vfio_address_spaces, list) {
2290         if (space->as == as) {
2291             return space;
2292         }
2293     }
2294 
2295     /* No suitable VFIOAddressSpace, create a new one */
2296     space = g_malloc0(sizeof(*space));
2297     space->as = as;
2298     QLIST_INIT(&space->containers);
2299 
2300     QLIST_INSERT_HEAD(&vfio_address_spaces, space, list);
2301 
2302     return space;
2303 }
2304 
2305 static void vfio_put_address_space(VFIOAddressSpace *space)
2306 {
2307     if (QLIST_EMPTY(&space->containers)) {
2308         QLIST_REMOVE(space, list);
2309         g_free(space);
2310     }
2311 }
2312 
2313 /*
2314  * vfio_get_iommu_type - selects the richest iommu_type (v2 first)
2315  */
2316 static int vfio_get_iommu_type(VFIOContainer *container,
2317                                Error **errp)
2318 {
2319     int iommu_types[] = { VFIO_TYPE1v2_IOMMU, VFIO_TYPE1_IOMMU,
2320                           VFIO_SPAPR_TCE_v2_IOMMU, VFIO_SPAPR_TCE_IOMMU };
2321     int i;
2322 
2323     for (i = 0; i < ARRAY_SIZE(iommu_types); i++) {
2324         if (ioctl(container->fd, VFIO_CHECK_EXTENSION, iommu_types[i])) {
2325             return iommu_types[i];
2326         }
2327     }
2328     error_setg(errp, "No available IOMMU models");
2329     return -EINVAL;
2330 }
2331 
2332 static int vfio_init_container(VFIOContainer *container, int group_fd,
2333                                Error **errp)
2334 {
2335     int iommu_type, ret;
2336 
2337     iommu_type = vfio_get_iommu_type(container, errp);
2338     if (iommu_type < 0) {
2339         return iommu_type;
2340     }
2341 
2342     ret = ioctl(group_fd, VFIO_GROUP_SET_CONTAINER, &container->fd);
2343     if (ret) {
2344         error_setg_errno(errp, errno, "Failed to set group container");
2345         return -errno;
2346     }
2347 
2348     while (ioctl(container->fd, VFIO_SET_IOMMU, iommu_type)) {
2349         if (iommu_type == VFIO_SPAPR_TCE_v2_IOMMU) {
2350             /*
2351              * On sPAPR, despite the IOMMU subdriver always advertises v1 and
2352              * v2, the running platform may not support v2 and there is no
2353              * way to guess it until an IOMMU group gets added to the container.
2354              * So in case it fails with v2, try v1 as a fallback.
2355              */
2356             iommu_type = VFIO_SPAPR_TCE_IOMMU;
2357             continue;
2358         }
2359         error_setg_errno(errp, errno, "Failed to set iommu for container");
2360         return -errno;
2361     }
2362 
2363     container->iommu_type = iommu_type;
2364     return 0;
2365 }
2366 
2367 static int vfio_get_iommu_info(VFIOContainer *container,
2368                                struct vfio_iommu_type1_info **info)
2369 {
2370 
2371     size_t argsz = sizeof(struct vfio_iommu_type1_info);
2372 
2373     *info = g_new0(struct vfio_iommu_type1_info, 1);
2374 again:
2375     (*info)->argsz = argsz;
2376 
2377     if (ioctl(container->fd, VFIO_IOMMU_GET_INFO, *info)) {
2378         g_free(*info);
2379         *info = NULL;
2380         return -errno;
2381     }
2382 
2383     if (((*info)->argsz > argsz)) {
2384         argsz = (*info)->argsz;
2385         *info = g_realloc(*info, argsz);
2386         goto again;
2387     }
2388 
2389     return 0;
2390 }
2391 
2392 static struct vfio_info_cap_header *
2393 vfio_get_iommu_info_cap(struct vfio_iommu_type1_info *info, uint16_t id)
2394 {
2395     struct vfio_info_cap_header *hdr;
2396     void *ptr = info;
2397 
2398     if (!(info->flags & VFIO_IOMMU_INFO_CAPS)) {
2399         return NULL;
2400     }
2401 
2402     for (hdr = ptr + info->cap_offset; hdr != ptr; hdr = ptr + hdr->next) {
2403         if (hdr->id == id) {
2404             return hdr;
2405         }
2406     }
2407 
2408     return NULL;
2409 }
2410 
2411 static void vfio_get_iommu_info_migration(VFIOContainer *container,
2412                                          struct vfio_iommu_type1_info *info)
2413 {
2414     struct vfio_info_cap_header *hdr;
2415     struct vfio_iommu_type1_info_cap_migration *cap_mig;
2416 
2417     hdr = vfio_get_iommu_info_cap(info, VFIO_IOMMU_TYPE1_INFO_CAP_MIGRATION);
2418     if (!hdr) {
2419         return;
2420     }
2421 
2422     cap_mig = container_of(hdr, struct vfio_iommu_type1_info_cap_migration,
2423                             header);
2424 
2425     /*
2426      * cpu_physical_memory_set_dirty_lebitmap() supports pages in bitmap of
2427      * qemu_real_host_page_size to mark those dirty.
2428      */
2429     if (cap_mig->pgsize_bitmap & qemu_real_host_page_size()) {
2430         container->dirty_pages_supported = true;
2431         container->max_dirty_bitmap_size = cap_mig->max_dirty_bitmap_size;
2432         container->dirty_pgsizes = cap_mig->pgsize_bitmap;
2433     }
2434 }
2435 
2436 static int vfio_connect_container(VFIOGroup *group, AddressSpace *as,
2437                                   Error **errp)
2438 {
2439     VFIOContainer *container;
2440     int ret, fd;
2441     VFIOAddressSpace *space;
2442 
2443     space = vfio_get_address_space(as);
2444 
2445     /*
2446      * VFIO is currently incompatible with discarding of RAM insofar as the
2447      * madvise to purge (zap) the page from QEMU's address space does not
2448      * interact with the memory API and therefore leaves stale virtual to
2449      * physical mappings in the IOMMU if the page was previously pinned.  We
2450      * therefore set discarding broken for each group added to a container,
2451      * whether the container is used individually or shared.  This provides
2452      * us with options to allow devices within a group to opt-in and allow
2453      * discarding, so long as it is done consistently for a group (for instance
2454      * if the device is an mdev device where it is known that the host vendor
2455      * driver will never pin pages outside of the working set of the guest
2456      * driver, which would thus not be discarding candidates).
2457      *
2458      * The first opportunity to induce pinning occurs here where we attempt to
2459      * attach the group to existing containers within the AddressSpace.  If any
2460      * pages are already zapped from the virtual address space, such as from
2461      * previous discards, new pinning will cause valid mappings to be
2462      * re-established.  Likewise, when the overall MemoryListener for a new
2463      * container is registered, a replay of mappings within the AddressSpace
2464      * will occur, re-establishing any previously zapped pages as well.
2465      *
2466      * Especially virtio-balloon is currently only prevented from discarding
2467      * new memory, it will not yet set ram_block_discard_set_required() and
2468      * therefore, neither stops us here or deals with the sudden memory
2469      * consumption of inflated memory.
2470      *
2471      * We do support discarding of memory coordinated via the RamDiscardManager
2472      * with some IOMMU types. vfio_ram_block_discard_disable() handles the
2473      * details once we know which type of IOMMU we are using.
2474      */
2475 
2476     QLIST_FOREACH(container, &space->containers, next) {
2477         if (!ioctl(group->fd, VFIO_GROUP_SET_CONTAINER, &container->fd)) {
2478             ret = vfio_ram_block_discard_disable(container, true);
2479             if (ret) {
2480                 error_setg_errno(errp, -ret,
2481                                  "Cannot set discarding of RAM broken");
2482                 if (ioctl(group->fd, VFIO_GROUP_UNSET_CONTAINER,
2483                           &container->fd)) {
2484                     error_report("vfio: error disconnecting group %d from"
2485                                  " container", group->groupid);
2486                 }
2487                 return ret;
2488             }
2489             group->container = container;
2490             QLIST_INSERT_HEAD(&container->group_list, group, container_next);
2491             vfio_kvm_device_add_group(group);
2492             return 0;
2493         }
2494     }
2495 
2496     fd = qemu_open_old("/dev/vfio/vfio", O_RDWR);
2497     if (fd < 0) {
2498         error_setg_errno(errp, errno, "failed to open /dev/vfio/vfio");
2499         ret = -errno;
2500         goto put_space_exit;
2501     }
2502 
2503     ret = ioctl(fd, VFIO_GET_API_VERSION);
2504     if (ret != VFIO_API_VERSION) {
2505         error_setg(errp, "supported vfio version: %d, "
2506                    "reported version: %d", VFIO_API_VERSION, ret);
2507         ret = -EINVAL;
2508         goto close_fd_exit;
2509     }
2510 
2511     container = g_malloc0(sizeof(*container));
2512     container->space = space;
2513     container->fd = fd;
2514     container->error = NULL;
2515     container->dirty_pages_supported = false;
2516     container->dma_max_mappings = 0;
2517     QLIST_INIT(&container->giommu_list);
2518     QLIST_INIT(&container->hostwin_list);
2519     QLIST_INIT(&container->vrdl_list);
2520 
2521     ret = vfio_init_container(container, group->fd, errp);
2522     if (ret) {
2523         goto free_container_exit;
2524     }
2525 
2526     ret = vfio_ram_block_discard_disable(container, true);
2527     if (ret) {
2528         error_setg_errno(errp, -ret, "Cannot set discarding of RAM broken");
2529         goto free_container_exit;
2530     }
2531 
2532     switch (container->iommu_type) {
2533     case VFIO_TYPE1v2_IOMMU:
2534     case VFIO_TYPE1_IOMMU:
2535     {
2536         struct vfio_iommu_type1_info *info;
2537 
2538         ret = vfio_get_iommu_info(container, &info);
2539         if (ret) {
2540             error_setg_errno(errp, -ret, "Failed to get VFIO IOMMU info");
2541             goto enable_discards_exit;
2542         }
2543 
2544         if (info->flags & VFIO_IOMMU_INFO_PGSIZES) {
2545             container->pgsizes = info->iova_pgsizes;
2546         } else {
2547             container->pgsizes = qemu_real_host_page_size();
2548         }
2549 
2550         if (!vfio_get_info_dma_avail(info, &container->dma_max_mappings)) {
2551             container->dma_max_mappings = 65535;
2552         }
2553         vfio_get_iommu_info_migration(container, info);
2554         g_free(info);
2555 
2556         /*
2557          * FIXME: We should parse VFIO_IOMMU_TYPE1_INFO_CAP_IOVA_RANGE
2558          * information to get the actual window extent rather than assume
2559          * a 64-bit IOVA address space.
2560          */
2561         vfio_host_win_add(container, 0, (hwaddr)-1, container->pgsizes);
2562 
2563         break;
2564     }
2565     case VFIO_SPAPR_TCE_v2_IOMMU:
2566     case VFIO_SPAPR_TCE_IOMMU:
2567     {
2568         struct vfio_iommu_spapr_tce_info info;
2569         bool v2 = container->iommu_type == VFIO_SPAPR_TCE_v2_IOMMU;
2570 
2571         /*
2572          * The host kernel code implementing VFIO_IOMMU_DISABLE is called
2573          * when container fd is closed so we do not call it explicitly
2574          * in this file.
2575          */
2576         if (!v2) {
2577             ret = ioctl(fd, VFIO_IOMMU_ENABLE);
2578             if (ret) {
2579                 error_setg_errno(errp, errno, "failed to enable container");
2580                 ret = -errno;
2581                 goto enable_discards_exit;
2582             }
2583         } else {
2584             container->prereg_listener = vfio_prereg_listener;
2585 
2586             memory_listener_register(&container->prereg_listener,
2587                                      &address_space_memory);
2588             if (container->error) {
2589                 memory_listener_unregister(&container->prereg_listener);
2590                 ret = -1;
2591                 error_propagate_prepend(errp, container->error,
2592                     "RAM memory listener initialization failed: ");
2593                 goto enable_discards_exit;
2594             }
2595         }
2596 
2597         info.argsz = sizeof(info);
2598         ret = ioctl(fd, VFIO_IOMMU_SPAPR_TCE_GET_INFO, &info);
2599         if (ret) {
2600             error_setg_errno(errp, errno,
2601                              "VFIO_IOMMU_SPAPR_TCE_GET_INFO failed");
2602             ret = -errno;
2603             if (v2) {
2604                 memory_listener_unregister(&container->prereg_listener);
2605             }
2606             goto enable_discards_exit;
2607         }
2608 
2609         if (v2) {
2610             container->pgsizes = info.ddw.pgsizes;
2611             /*
2612              * There is a default window in just created container.
2613              * To make region_add/del simpler, we better remove this
2614              * window now and let those iommu_listener callbacks
2615              * create/remove them when needed.
2616              */
2617             ret = vfio_spapr_remove_window(container, info.dma32_window_start);
2618             if (ret) {
2619                 error_setg_errno(errp, -ret,
2620                                  "failed to remove existing window");
2621                 goto enable_discards_exit;
2622             }
2623         } else {
2624             /* The default table uses 4K pages */
2625             container->pgsizes = 0x1000;
2626             vfio_host_win_add(container, info.dma32_window_start,
2627                               info.dma32_window_start +
2628                               info.dma32_window_size - 1,
2629                               0x1000);
2630         }
2631     }
2632     }
2633 
2634     vfio_kvm_device_add_group(group);
2635 
2636     QLIST_INIT(&container->group_list);
2637     QLIST_INSERT_HEAD(&space->containers, container, next);
2638 
2639     group->container = container;
2640     QLIST_INSERT_HEAD(&container->group_list, group, container_next);
2641 
2642     container->listener = vfio_memory_listener;
2643 
2644     memory_listener_register(&container->listener, container->space->as);
2645 
2646     if (container->error) {
2647         ret = -1;
2648         error_propagate_prepend(errp, container->error,
2649             "memory listener initialization failed: ");
2650         goto listener_release_exit;
2651     }
2652 
2653     container->initialized = true;
2654 
2655     return 0;
2656 listener_release_exit:
2657     QLIST_REMOVE(group, container_next);
2658     QLIST_REMOVE(container, next);
2659     vfio_kvm_device_del_group(group);
2660     vfio_listener_release(container);
2661 
2662 enable_discards_exit:
2663     vfio_ram_block_discard_disable(container, false);
2664 
2665 free_container_exit:
2666     g_free(container);
2667 
2668 close_fd_exit:
2669     close(fd);
2670 
2671 put_space_exit:
2672     vfio_put_address_space(space);
2673 
2674     return ret;
2675 }
2676 
2677 static void vfio_disconnect_container(VFIOGroup *group)
2678 {
2679     VFIOContainer *container = group->container;
2680 
2681     QLIST_REMOVE(group, container_next);
2682     group->container = NULL;
2683 
2684     /*
2685      * Explicitly release the listener first before unset container,
2686      * since unset may destroy the backend container if it's the last
2687      * group.
2688      */
2689     if (QLIST_EMPTY(&container->group_list)) {
2690         vfio_listener_release(container);
2691     }
2692 
2693     if (ioctl(group->fd, VFIO_GROUP_UNSET_CONTAINER, &container->fd)) {
2694         error_report("vfio: error disconnecting group %d from container",
2695                      group->groupid);
2696     }
2697 
2698     if (QLIST_EMPTY(&container->group_list)) {
2699         VFIOAddressSpace *space = container->space;
2700         VFIOGuestIOMMU *giommu, *tmp;
2701         VFIOHostDMAWindow *hostwin, *next;
2702 
2703         QLIST_REMOVE(container, next);
2704 
2705         QLIST_FOREACH_SAFE(giommu, &container->giommu_list, giommu_next, tmp) {
2706             memory_region_unregister_iommu_notifier(
2707                     MEMORY_REGION(giommu->iommu_mr), &giommu->n);
2708             QLIST_REMOVE(giommu, giommu_next);
2709             g_free(giommu);
2710         }
2711 
2712         QLIST_FOREACH_SAFE(hostwin, &container->hostwin_list, hostwin_next,
2713                            next) {
2714             QLIST_REMOVE(hostwin, hostwin_next);
2715             g_free(hostwin);
2716         }
2717 
2718         trace_vfio_disconnect_container(container->fd);
2719         close(container->fd);
2720         g_free(container);
2721 
2722         vfio_put_address_space(space);
2723     }
2724 }
2725 
2726 VFIOGroup *vfio_get_group(int groupid, AddressSpace *as, Error **errp)
2727 {
2728     VFIOGroup *group;
2729     char path[32];
2730     struct vfio_group_status status = { .argsz = sizeof(status) };
2731 
2732     QLIST_FOREACH(group, &vfio_group_list, next) {
2733         if (group->groupid == groupid) {
2734             /* Found it.  Now is it already in the right context? */
2735             if (group->container->space->as == as) {
2736                 return group;
2737             } else {
2738                 error_setg(errp, "group %d used in multiple address spaces",
2739                            group->groupid);
2740                 return NULL;
2741             }
2742         }
2743     }
2744 
2745     group = g_malloc0(sizeof(*group));
2746 
2747     snprintf(path, sizeof(path), "/dev/vfio/%d", groupid);
2748     group->fd = qemu_open_old(path, O_RDWR);
2749     if (group->fd < 0) {
2750         error_setg_errno(errp, errno, "failed to open %s", path);
2751         goto free_group_exit;
2752     }
2753 
2754     if (ioctl(group->fd, VFIO_GROUP_GET_STATUS, &status)) {
2755         error_setg_errno(errp, errno, "failed to get group %d status", groupid);
2756         goto close_fd_exit;
2757     }
2758 
2759     if (!(status.flags & VFIO_GROUP_FLAGS_VIABLE)) {
2760         error_setg(errp, "group %d is not viable", groupid);
2761         error_append_hint(errp,
2762                           "Please ensure all devices within the iommu_group "
2763                           "are bound to their vfio bus driver.\n");
2764         goto close_fd_exit;
2765     }
2766 
2767     group->groupid = groupid;
2768     QLIST_INIT(&group->device_list);
2769 
2770     if (vfio_connect_container(group, as, errp)) {
2771         error_prepend(errp, "failed to setup container for group %d: ",
2772                       groupid);
2773         goto close_fd_exit;
2774     }
2775 
2776     if (QLIST_EMPTY(&vfio_group_list)) {
2777         qemu_register_reset(vfio_reset_handler, NULL);
2778     }
2779 
2780     QLIST_INSERT_HEAD(&vfio_group_list, group, next);
2781 
2782     return group;
2783 
2784 close_fd_exit:
2785     close(group->fd);
2786 
2787 free_group_exit:
2788     g_free(group);
2789 
2790     return NULL;
2791 }
2792 
2793 void vfio_put_group(VFIOGroup *group)
2794 {
2795     if (!group || !QLIST_EMPTY(&group->device_list)) {
2796         return;
2797     }
2798 
2799     if (!group->ram_block_discard_allowed) {
2800         vfio_ram_block_discard_disable(group->container, false);
2801     }
2802     vfio_kvm_device_del_group(group);
2803     vfio_disconnect_container(group);
2804     QLIST_REMOVE(group, next);
2805     trace_vfio_put_group(group->fd);
2806     close(group->fd);
2807     g_free(group);
2808 
2809     if (QLIST_EMPTY(&vfio_group_list)) {
2810         qemu_unregister_reset(vfio_reset_handler, NULL);
2811     }
2812 }
2813 
2814 struct vfio_device_info *vfio_get_device_info(int fd)
2815 {
2816     struct vfio_device_info *info;
2817     uint32_t argsz = sizeof(*info);
2818 
2819     info = g_malloc0(argsz);
2820 
2821 retry:
2822     info->argsz = argsz;
2823 
2824     if (ioctl(fd, VFIO_DEVICE_GET_INFO, info)) {
2825         g_free(info);
2826         return NULL;
2827     }
2828 
2829     if (info->argsz > argsz) {
2830         argsz = info->argsz;
2831         info = g_realloc(info, argsz);
2832         goto retry;
2833     }
2834 
2835     return info;
2836 }
2837 
2838 int vfio_get_device(VFIOGroup *group, const char *name,
2839                     VFIODevice *vbasedev, Error **errp)
2840 {
2841     g_autofree struct vfio_device_info *info = NULL;
2842     int fd;
2843 
2844     fd = ioctl(group->fd, VFIO_GROUP_GET_DEVICE_FD, name);
2845     if (fd < 0) {
2846         error_setg_errno(errp, errno, "error getting device from group %d",
2847                          group->groupid);
2848         error_append_hint(errp,
2849                       "Verify all devices in group %d are bound to vfio-<bus> "
2850                       "or pci-stub and not already in use\n", group->groupid);
2851         return fd;
2852     }
2853 
2854     info = vfio_get_device_info(fd);
2855     if (!info) {
2856         error_setg_errno(errp, errno, "error getting device info");
2857         close(fd);
2858         return -1;
2859     }
2860 
2861     /*
2862      * Set discarding of RAM as not broken for this group if the driver knows
2863      * the device operates compatibly with discarding.  Setting must be
2864      * consistent per group, but since compatibility is really only possible
2865      * with mdev currently, we expect singleton groups.
2866      */
2867     if (vbasedev->ram_block_discard_allowed !=
2868         group->ram_block_discard_allowed) {
2869         if (!QLIST_EMPTY(&group->device_list)) {
2870             error_setg(errp, "Inconsistent setting of support for discarding "
2871                        "RAM (e.g., balloon) within group");
2872             close(fd);
2873             return -1;
2874         }
2875 
2876         if (!group->ram_block_discard_allowed) {
2877             group->ram_block_discard_allowed = true;
2878             vfio_ram_block_discard_disable(group->container, false);
2879         }
2880     }
2881 
2882     vbasedev->fd = fd;
2883     vbasedev->group = group;
2884     QLIST_INSERT_HEAD(&group->device_list, vbasedev, next);
2885 
2886     vbasedev->num_irqs = info->num_irqs;
2887     vbasedev->num_regions = info->num_regions;
2888     vbasedev->flags = info->flags;
2889 
2890     trace_vfio_get_device(name, info->flags, info->num_regions, info->num_irqs);
2891 
2892     vbasedev->reset_works = !!(info->flags & VFIO_DEVICE_FLAGS_RESET);
2893 
2894     return 0;
2895 }
2896 
2897 void vfio_put_base_device(VFIODevice *vbasedev)
2898 {
2899     if (!vbasedev->group) {
2900         return;
2901     }
2902     QLIST_REMOVE(vbasedev, next);
2903     vbasedev->group = NULL;
2904     trace_vfio_put_base_device(vbasedev->fd);
2905     close(vbasedev->fd);
2906 }
2907 
2908 int vfio_get_region_info(VFIODevice *vbasedev, int index,
2909                          struct vfio_region_info **info)
2910 {
2911     size_t argsz = sizeof(struct vfio_region_info);
2912 
2913     *info = g_malloc0(argsz);
2914 
2915     (*info)->index = index;
2916 retry:
2917     (*info)->argsz = argsz;
2918 
2919     if (ioctl(vbasedev->fd, VFIO_DEVICE_GET_REGION_INFO, *info)) {
2920         g_free(*info);
2921         *info = NULL;
2922         return -errno;
2923     }
2924 
2925     if ((*info)->argsz > argsz) {
2926         argsz = (*info)->argsz;
2927         *info = g_realloc(*info, argsz);
2928 
2929         goto retry;
2930     }
2931 
2932     return 0;
2933 }
2934 
2935 int vfio_get_dev_region_info(VFIODevice *vbasedev, uint32_t type,
2936                              uint32_t subtype, struct vfio_region_info **info)
2937 {
2938     int i;
2939 
2940     for (i = 0; i < vbasedev->num_regions; i++) {
2941         struct vfio_info_cap_header *hdr;
2942         struct vfio_region_info_cap_type *cap_type;
2943 
2944         if (vfio_get_region_info(vbasedev, i, info)) {
2945             continue;
2946         }
2947 
2948         hdr = vfio_get_region_info_cap(*info, VFIO_REGION_INFO_CAP_TYPE);
2949         if (!hdr) {
2950             g_free(*info);
2951             continue;
2952         }
2953 
2954         cap_type = container_of(hdr, struct vfio_region_info_cap_type, header);
2955 
2956         trace_vfio_get_dev_region(vbasedev->name, i,
2957                                   cap_type->type, cap_type->subtype);
2958 
2959         if (cap_type->type == type && cap_type->subtype == subtype) {
2960             return 0;
2961         }
2962 
2963         g_free(*info);
2964     }
2965 
2966     *info = NULL;
2967     return -ENODEV;
2968 }
2969 
2970 bool vfio_has_region_cap(VFIODevice *vbasedev, int region, uint16_t cap_type)
2971 {
2972     struct vfio_region_info *info = NULL;
2973     bool ret = false;
2974 
2975     if (!vfio_get_region_info(vbasedev, region, &info)) {
2976         if (vfio_get_region_info_cap(info, cap_type)) {
2977             ret = true;
2978         }
2979         g_free(info);
2980     }
2981 
2982     return ret;
2983 }
2984 
2985 /*
2986  * Interfaces for IBM EEH (Enhanced Error Handling)
2987  */
2988 static bool vfio_eeh_container_ok(VFIOContainer *container)
2989 {
2990     /*
2991      * As of 2016-03-04 (linux-4.5) the host kernel EEH/VFIO
2992      * implementation is broken if there are multiple groups in a
2993      * container.  The hardware works in units of Partitionable
2994      * Endpoints (== IOMMU groups) and the EEH operations naively
2995      * iterate across all groups in the container, without any logic
2996      * to make sure the groups have their state synchronized.  For
2997      * certain operations (ENABLE) that might be ok, until an error
2998      * occurs, but for others (GET_STATE) it's clearly broken.
2999      */
3000 
3001     /*
3002      * XXX Once fixed kernels exist, test for them here
3003      */
3004 
3005     if (QLIST_EMPTY(&container->group_list)) {
3006         return false;
3007     }
3008 
3009     if (QLIST_NEXT(QLIST_FIRST(&container->group_list), container_next)) {
3010         return false;
3011     }
3012 
3013     return true;
3014 }
3015 
3016 static int vfio_eeh_container_op(VFIOContainer *container, uint32_t op)
3017 {
3018     struct vfio_eeh_pe_op pe_op = {
3019         .argsz = sizeof(pe_op),
3020         .op = op,
3021     };
3022     int ret;
3023 
3024     if (!vfio_eeh_container_ok(container)) {
3025         error_report("vfio/eeh: EEH_PE_OP 0x%x: "
3026                      "kernel requires a container with exactly one group", op);
3027         return -EPERM;
3028     }
3029 
3030     ret = ioctl(container->fd, VFIO_EEH_PE_OP, &pe_op);
3031     if (ret < 0) {
3032         error_report("vfio/eeh: EEH_PE_OP 0x%x failed: %m", op);
3033         return -errno;
3034     }
3035 
3036     return ret;
3037 }
3038 
3039 static VFIOContainer *vfio_eeh_as_container(AddressSpace *as)
3040 {
3041     VFIOAddressSpace *space = vfio_get_address_space(as);
3042     VFIOContainer *container = NULL;
3043 
3044     if (QLIST_EMPTY(&space->containers)) {
3045         /* No containers to act on */
3046         goto out;
3047     }
3048 
3049     container = QLIST_FIRST(&space->containers);
3050 
3051     if (QLIST_NEXT(container, next)) {
3052         /* We don't yet have logic to synchronize EEH state across
3053          * multiple containers */
3054         container = NULL;
3055         goto out;
3056     }
3057 
3058 out:
3059     vfio_put_address_space(space);
3060     return container;
3061 }
3062 
3063 bool vfio_eeh_as_ok(AddressSpace *as)
3064 {
3065     VFIOContainer *container = vfio_eeh_as_container(as);
3066 
3067     return (container != NULL) && vfio_eeh_container_ok(container);
3068 }
3069 
3070 int vfio_eeh_as_op(AddressSpace *as, uint32_t op)
3071 {
3072     VFIOContainer *container = vfio_eeh_as_container(as);
3073 
3074     if (!container) {
3075         return -ENODEV;
3076     }
3077     return vfio_eeh_container_op(container, op);
3078 }
3079