xref: /openbmc/qemu/hw/virtio/virtio-iommu.c (revision 243975c0)
1 /*
2  * virtio-iommu device
3  *
4  * Copyright (c) 2020 Red Hat, Inc.
5  *
6  * This program is free software; you can redistribute it and/or modify it
7  * under the terms and conditions of the GNU General Public License,
8  * version 2 or later, as published by the Free Software Foundation.
9  *
10  * This program is distributed in the hope it will be useful, but WITHOUT
11  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
12  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
13  * more details.
14  *
15  * You should have received a copy of the GNU General Public License along with
16  * this program.  If not, see <http://www.gnu.org/licenses/>.
17  *
18  */
19 
20 #include "qemu/osdep.h"
21 #include "qemu/log.h"
22 #include "qemu/iov.h"
23 #include "exec/target_page.h"
24 #include "hw/qdev-properties.h"
25 #include "hw/virtio/virtio.h"
26 #include "sysemu/kvm.h"
27 #include "sysemu/reset.h"
28 #include "sysemu/sysemu.h"
29 #include "qapi/error.h"
30 #include "qemu/error-report.h"
31 #include "trace.h"
32 
33 #include "standard-headers/linux/virtio_ids.h"
34 
35 #include "hw/virtio/virtio-bus.h"
36 #include "hw/virtio/virtio-iommu.h"
37 #include "hw/pci/pci_bus.h"
38 #include "hw/pci/pci.h"
39 
40 /* Max size */
41 #define VIOMMU_DEFAULT_QUEUE_SIZE 256
42 #define VIOMMU_PROBE_SIZE 512
43 
44 typedef struct VirtIOIOMMUDomain {
45     uint32_t id;
46     bool bypass;
47     GTree *mappings;
48     QLIST_HEAD(, VirtIOIOMMUEndpoint) endpoint_list;
49 } VirtIOIOMMUDomain;
50 
51 typedef struct VirtIOIOMMUEndpoint {
52     uint32_t id;
53     VirtIOIOMMUDomain *domain;
54     IOMMUMemoryRegion *iommu_mr;
55     QLIST_ENTRY(VirtIOIOMMUEndpoint) next;
56 } VirtIOIOMMUEndpoint;
57 
58 typedef struct VirtIOIOMMUInterval {
59     uint64_t low;
60     uint64_t high;
61 } VirtIOIOMMUInterval;
62 
63 typedef struct VirtIOIOMMUMapping {
64     uint64_t phys_addr;
65     uint32_t flags;
66 } VirtIOIOMMUMapping;
67 
68 static inline uint16_t virtio_iommu_get_bdf(IOMMUDevice *dev)
69 {
70     return PCI_BUILD_BDF(pci_bus_num(dev->bus), dev->devfn);
71 }
72 
73 static bool virtio_iommu_device_bypassed(IOMMUDevice *sdev)
74 {
75     uint32_t sid;
76     bool bypassed;
77     VirtIOIOMMU *s = sdev->viommu;
78     VirtIOIOMMUEndpoint *ep;
79 
80     sid = virtio_iommu_get_bdf(sdev);
81 
82     qemu_rec_mutex_lock(&s->mutex);
83     /* need to check bypass before system reset */
84     if (!s->endpoints) {
85         bypassed = s->config.bypass;
86         goto unlock;
87     }
88 
89     ep = g_tree_lookup(s->endpoints, GUINT_TO_POINTER(sid));
90     if (!ep || !ep->domain) {
91         bypassed = s->config.bypass;
92     } else {
93         bypassed = ep->domain->bypass;
94     }
95 
96 unlock:
97     qemu_rec_mutex_unlock(&s->mutex);
98     return bypassed;
99 }
100 
101 /* Return whether the device is using IOMMU translation. */
102 static bool virtio_iommu_switch_address_space(IOMMUDevice *sdev)
103 {
104     bool use_remapping;
105 
106     assert(sdev);
107 
108     use_remapping = !virtio_iommu_device_bypassed(sdev);
109 
110     trace_virtio_iommu_switch_address_space(pci_bus_num(sdev->bus),
111                                             PCI_SLOT(sdev->devfn),
112                                             PCI_FUNC(sdev->devfn),
113                                             use_remapping);
114 
115     /* Turn off first then on the other */
116     if (use_remapping) {
117         memory_region_set_enabled(&sdev->bypass_mr, false);
118         memory_region_set_enabled(MEMORY_REGION(&sdev->iommu_mr), true);
119     } else {
120         memory_region_set_enabled(MEMORY_REGION(&sdev->iommu_mr), false);
121         memory_region_set_enabled(&sdev->bypass_mr, true);
122     }
123 
124     return use_remapping;
125 }
126 
127 static void virtio_iommu_switch_address_space_all(VirtIOIOMMU *s)
128 {
129     GHashTableIter iter;
130     IOMMUPciBus *iommu_pci_bus;
131     int i;
132 
133     g_hash_table_iter_init(&iter, s->as_by_busptr);
134     while (g_hash_table_iter_next(&iter, NULL, (void **)&iommu_pci_bus)) {
135         for (i = 0; i < PCI_DEVFN_MAX; i++) {
136             if (!iommu_pci_bus->pbdev[i]) {
137                 continue;
138             }
139             virtio_iommu_switch_address_space(iommu_pci_bus->pbdev[i]);
140         }
141     }
142 }
143 
144 /**
145  * The bus number is used for lookup when SID based operations occur.
146  * In that case we lazily populate the IOMMUPciBus array from the bus hash
147  * table. At the time the IOMMUPciBus is created (iommu_find_add_as), the bus
148  * numbers may not be always initialized yet.
149  */
150 static IOMMUPciBus *iommu_find_iommu_pcibus(VirtIOIOMMU *s, uint8_t bus_num)
151 {
152     IOMMUPciBus *iommu_pci_bus = s->iommu_pcibus_by_bus_num[bus_num];
153 
154     if (!iommu_pci_bus) {
155         GHashTableIter iter;
156 
157         g_hash_table_iter_init(&iter, s->as_by_busptr);
158         while (g_hash_table_iter_next(&iter, NULL, (void **)&iommu_pci_bus)) {
159             if (pci_bus_num(iommu_pci_bus->bus) == bus_num) {
160                 s->iommu_pcibus_by_bus_num[bus_num] = iommu_pci_bus;
161                 return iommu_pci_bus;
162             }
163         }
164         return NULL;
165     }
166     return iommu_pci_bus;
167 }
168 
169 static IOMMUMemoryRegion *virtio_iommu_mr(VirtIOIOMMU *s, uint32_t sid)
170 {
171     uint8_t bus_n, devfn;
172     IOMMUPciBus *iommu_pci_bus;
173     IOMMUDevice *dev;
174 
175     bus_n = PCI_BUS_NUM(sid);
176     iommu_pci_bus = iommu_find_iommu_pcibus(s, bus_n);
177     if (iommu_pci_bus) {
178         devfn = sid & (PCI_DEVFN_MAX - 1);
179         dev = iommu_pci_bus->pbdev[devfn];
180         if (dev) {
181             return &dev->iommu_mr;
182         }
183     }
184     return NULL;
185 }
186 
187 static gint interval_cmp(gconstpointer a, gconstpointer b, gpointer user_data)
188 {
189     VirtIOIOMMUInterval *inta = (VirtIOIOMMUInterval *)a;
190     VirtIOIOMMUInterval *intb = (VirtIOIOMMUInterval *)b;
191 
192     if (inta->high < intb->low) {
193         return -1;
194     } else if (intb->high < inta->low) {
195         return 1;
196     } else {
197         return 0;
198     }
199 }
200 
201 static void virtio_iommu_notify_map_unmap(IOMMUMemoryRegion *mr,
202                                           IOMMUTLBEvent *event,
203                                           hwaddr virt_start, hwaddr virt_end)
204 {
205     uint64_t delta = virt_end - virt_start;
206 
207     event->entry.iova = virt_start;
208     event->entry.addr_mask = delta;
209 
210     if (delta == UINT64_MAX) {
211         memory_region_notify_iommu(mr, 0, *event);
212     }
213 
214     while (virt_start != virt_end + 1) {
215         uint64_t mask = dma_aligned_pow2_mask(virt_start, virt_end, 64);
216 
217         event->entry.addr_mask = mask;
218         event->entry.iova = virt_start;
219         memory_region_notify_iommu(mr, 0, *event);
220         virt_start += mask + 1;
221         if (event->entry.perm != IOMMU_NONE) {
222             event->entry.translated_addr += mask + 1;
223         }
224     }
225 }
226 
227 static void virtio_iommu_notify_map(IOMMUMemoryRegion *mr, hwaddr virt_start,
228                                     hwaddr virt_end, hwaddr paddr,
229                                     uint32_t flags)
230 {
231     IOMMUTLBEvent event;
232     IOMMUAccessFlags perm = IOMMU_ACCESS_FLAG(flags & VIRTIO_IOMMU_MAP_F_READ,
233                                               flags & VIRTIO_IOMMU_MAP_F_WRITE);
234 
235     if (!(mr->iommu_notify_flags & IOMMU_NOTIFIER_MAP) ||
236         (flags & VIRTIO_IOMMU_MAP_F_MMIO) || !perm) {
237         return;
238     }
239 
240     trace_virtio_iommu_notify_map(mr->parent_obj.name, virt_start, virt_end,
241                                   paddr, perm);
242 
243     event.type = IOMMU_NOTIFIER_MAP;
244     event.entry.target_as = &address_space_memory;
245     event.entry.perm = perm;
246     event.entry.translated_addr = paddr;
247 
248     virtio_iommu_notify_map_unmap(mr, &event, virt_start, virt_end);
249 }
250 
251 static void virtio_iommu_notify_unmap(IOMMUMemoryRegion *mr, hwaddr virt_start,
252                                       hwaddr virt_end)
253 {
254     IOMMUTLBEvent event;
255 
256     if (!(mr->iommu_notify_flags & IOMMU_NOTIFIER_UNMAP)) {
257         return;
258     }
259 
260     trace_virtio_iommu_notify_unmap(mr->parent_obj.name, virt_start, virt_end);
261 
262     event.type = IOMMU_NOTIFIER_UNMAP;
263     event.entry.target_as = &address_space_memory;
264     event.entry.perm = IOMMU_NONE;
265     event.entry.translated_addr = 0;
266 
267     virtio_iommu_notify_map_unmap(mr, &event, virt_start, virt_end);
268 }
269 
270 static gboolean virtio_iommu_notify_unmap_cb(gpointer key, gpointer value,
271                                              gpointer data)
272 {
273     VirtIOIOMMUInterval *interval = (VirtIOIOMMUInterval *) key;
274     IOMMUMemoryRegion *mr = (IOMMUMemoryRegion *) data;
275 
276     virtio_iommu_notify_unmap(mr, interval->low, interval->high);
277 
278     return false;
279 }
280 
281 static gboolean virtio_iommu_notify_map_cb(gpointer key, gpointer value,
282                                            gpointer data)
283 {
284     VirtIOIOMMUMapping *mapping = (VirtIOIOMMUMapping *) value;
285     VirtIOIOMMUInterval *interval = (VirtIOIOMMUInterval *) key;
286     IOMMUMemoryRegion *mr = (IOMMUMemoryRegion *) data;
287 
288     virtio_iommu_notify_map(mr, interval->low, interval->high,
289                             mapping->phys_addr, mapping->flags);
290 
291     return false;
292 }
293 
294 static void virtio_iommu_detach_endpoint_from_domain(VirtIOIOMMUEndpoint *ep)
295 {
296     VirtIOIOMMUDomain *domain = ep->domain;
297     IOMMUDevice *sdev = container_of(ep->iommu_mr, IOMMUDevice, iommu_mr);
298 
299     if (!ep->domain) {
300         return;
301     }
302     g_tree_foreach(domain->mappings, virtio_iommu_notify_unmap_cb,
303                    ep->iommu_mr);
304     QLIST_REMOVE(ep, next);
305     ep->domain = NULL;
306     virtio_iommu_switch_address_space(sdev);
307 }
308 
309 static VirtIOIOMMUEndpoint *virtio_iommu_get_endpoint(VirtIOIOMMU *s,
310                                                       uint32_t ep_id)
311 {
312     VirtIOIOMMUEndpoint *ep;
313     IOMMUMemoryRegion *mr;
314 
315     ep = g_tree_lookup(s->endpoints, GUINT_TO_POINTER(ep_id));
316     if (ep) {
317         return ep;
318     }
319     mr = virtio_iommu_mr(s, ep_id);
320     if (!mr) {
321         return NULL;
322     }
323     ep = g_malloc0(sizeof(*ep));
324     ep->id = ep_id;
325     ep->iommu_mr = mr;
326     trace_virtio_iommu_get_endpoint(ep_id);
327     g_tree_insert(s->endpoints, GUINT_TO_POINTER(ep_id), ep);
328     return ep;
329 }
330 
331 static void virtio_iommu_put_endpoint(gpointer data)
332 {
333     VirtIOIOMMUEndpoint *ep = (VirtIOIOMMUEndpoint *)data;
334 
335     if (ep->domain) {
336         virtio_iommu_detach_endpoint_from_domain(ep);
337     }
338 
339     trace_virtio_iommu_put_endpoint(ep->id);
340     g_free(ep);
341 }
342 
343 static VirtIOIOMMUDomain *virtio_iommu_get_domain(VirtIOIOMMU *s,
344                                                   uint32_t domain_id,
345                                                   bool bypass)
346 {
347     VirtIOIOMMUDomain *domain;
348 
349     domain = g_tree_lookup(s->domains, GUINT_TO_POINTER(domain_id));
350     if (domain) {
351         if (domain->bypass != bypass) {
352             return NULL;
353         }
354         return domain;
355     }
356     domain = g_malloc0(sizeof(*domain));
357     domain->id = domain_id;
358     domain->mappings = g_tree_new_full((GCompareDataFunc)interval_cmp,
359                                    NULL, (GDestroyNotify)g_free,
360                                    (GDestroyNotify)g_free);
361     domain->bypass = bypass;
362     g_tree_insert(s->domains, GUINT_TO_POINTER(domain_id), domain);
363     QLIST_INIT(&domain->endpoint_list);
364     trace_virtio_iommu_get_domain(domain_id);
365     return domain;
366 }
367 
368 static void virtio_iommu_put_domain(gpointer data)
369 {
370     VirtIOIOMMUDomain *domain = (VirtIOIOMMUDomain *)data;
371     VirtIOIOMMUEndpoint *iter, *tmp;
372 
373     QLIST_FOREACH_SAFE(iter, &domain->endpoint_list, next, tmp) {
374         virtio_iommu_detach_endpoint_from_domain(iter);
375     }
376     g_tree_destroy(domain->mappings);
377     trace_virtio_iommu_put_domain(domain->id);
378     g_free(domain);
379 }
380 
381 static AddressSpace *virtio_iommu_find_add_as(PCIBus *bus, void *opaque,
382                                               int devfn)
383 {
384     VirtIOIOMMU *s = opaque;
385     IOMMUPciBus *sbus = g_hash_table_lookup(s->as_by_busptr, bus);
386     static uint32_t mr_index;
387     IOMMUDevice *sdev;
388 
389     if (!sbus) {
390         sbus = g_malloc0(sizeof(IOMMUPciBus) +
391                          sizeof(IOMMUDevice *) * PCI_DEVFN_MAX);
392         sbus->bus = bus;
393         g_hash_table_insert(s->as_by_busptr, bus, sbus);
394     }
395 
396     sdev = sbus->pbdev[devfn];
397     if (!sdev) {
398         char *name = g_strdup_printf("%s-%d-%d",
399                                      TYPE_VIRTIO_IOMMU_MEMORY_REGION,
400                                      mr_index++, devfn);
401         sdev = sbus->pbdev[devfn] = g_new0(IOMMUDevice, 1);
402 
403         sdev->viommu = s;
404         sdev->bus = bus;
405         sdev->devfn = devfn;
406 
407         trace_virtio_iommu_init_iommu_mr(name);
408 
409         memory_region_init(&sdev->root, OBJECT(s), name, UINT64_MAX);
410         address_space_init(&sdev->as, &sdev->root, TYPE_VIRTIO_IOMMU);
411 
412         /*
413          * Build the IOMMU disabled container with aliases to the
414          * shared MRs.  Note that aliasing to a shared memory region
415          * could help the memory API to detect same FlatViews so we
416          * can have devices to share the same FlatView when in bypass
417          * mode. (either by not configuring virtio-iommu driver or with
418          * "iommu=pt").  It will greatly reduce the total number of
419          * FlatViews of the system hence VM runs faster.
420          */
421         memory_region_init_alias(&sdev->bypass_mr, OBJECT(s),
422                                  "system", get_system_memory(), 0,
423                                  memory_region_size(get_system_memory()));
424 
425         memory_region_init_iommu(&sdev->iommu_mr, sizeof(sdev->iommu_mr),
426                                  TYPE_VIRTIO_IOMMU_MEMORY_REGION,
427                                  OBJECT(s), name,
428                                  UINT64_MAX);
429 
430         /*
431          * Hook both the containers under the root container, we
432          * switch between iommu & bypass MRs by enable/disable
433          * corresponding sub-containers
434          */
435         memory_region_add_subregion_overlap(&sdev->root, 0,
436                                             MEMORY_REGION(&sdev->iommu_mr),
437                                             0);
438         memory_region_add_subregion_overlap(&sdev->root, 0,
439                                             &sdev->bypass_mr, 0);
440 
441         virtio_iommu_switch_address_space(sdev);
442         g_free(name);
443     }
444     return &sdev->as;
445 }
446 
447 static int virtio_iommu_attach(VirtIOIOMMU *s,
448                                struct virtio_iommu_req_attach *req)
449 {
450     uint32_t domain_id = le32_to_cpu(req->domain);
451     uint32_t ep_id = le32_to_cpu(req->endpoint);
452     uint32_t flags = le32_to_cpu(req->flags);
453     VirtIOIOMMUDomain *domain;
454     VirtIOIOMMUEndpoint *ep;
455     IOMMUDevice *sdev;
456 
457     trace_virtio_iommu_attach(domain_id, ep_id);
458 
459     if (flags & ~VIRTIO_IOMMU_ATTACH_F_BYPASS) {
460         return VIRTIO_IOMMU_S_INVAL;
461     }
462 
463     ep = virtio_iommu_get_endpoint(s, ep_id);
464     if (!ep) {
465         return VIRTIO_IOMMU_S_NOENT;
466     }
467 
468     if (ep->domain) {
469         VirtIOIOMMUDomain *previous_domain = ep->domain;
470         /*
471          * the device is already attached to a domain,
472          * detach it first
473          */
474         virtio_iommu_detach_endpoint_from_domain(ep);
475         if (QLIST_EMPTY(&previous_domain->endpoint_list)) {
476             g_tree_remove(s->domains, GUINT_TO_POINTER(previous_domain->id));
477         }
478     }
479 
480     domain = virtio_iommu_get_domain(s, domain_id,
481                                      flags & VIRTIO_IOMMU_ATTACH_F_BYPASS);
482     if (!domain) {
483         /* Incompatible bypass flag */
484         return VIRTIO_IOMMU_S_INVAL;
485     }
486     QLIST_INSERT_HEAD(&domain->endpoint_list, ep, next);
487 
488     ep->domain = domain;
489     sdev = container_of(ep->iommu_mr, IOMMUDevice, iommu_mr);
490     virtio_iommu_switch_address_space(sdev);
491 
492     /* Replay domain mappings on the associated memory region */
493     g_tree_foreach(domain->mappings, virtio_iommu_notify_map_cb,
494                    ep->iommu_mr);
495 
496     return VIRTIO_IOMMU_S_OK;
497 }
498 
499 static int virtio_iommu_detach(VirtIOIOMMU *s,
500                                struct virtio_iommu_req_detach *req)
501 {
502     uint32_t domain_id = le32_to_cpu(req->domain);
503     uint32_t ep_id = le32_to_cpu(req->endpoint);
504     VirtIOIOMMUDomain *domain;
505     VirtIOIOMMUEndpoint *ep;
506 
507     trace_virtio_iommu_detach(domain_id, ep_id);
508 
509     ep = g_tree_lookup(s->endpoints, GUINT_TO_POINTER(ep_id));
510     if (!ep) {
511         return VIRTIO_IOMMU_S_NOENT;
512     }
513 
514     domain = ep->domain;
515 
516     if (!domain || domain->id != domain_id) {
517         return VIRTIO_IOMMU_S_INVAL;
518     }
519 
520     virtio_iommu_detach_endpoint_from_domain(ep);
521 
522     if (QLIST_EMPTY(&domain->endpoint_list)) {
523         g_tree_remove(s->domains, GUINT_TO_POINTER(domain->id));
524     }
525     return VIRTIO_IOMMU_S_OK;
526 }
527 
528 static int virtio_iommu_map(VirtIOIOMMU *s,
529                             struct virtio_iommu_req_map *req)
530 {
531     uint32_t domain_id = le32_to_cpu(req->domain);
532     uint64_t phys_start = le64_to_cpu(req->phys_start);
533     uint64_t virt_start = le64_to_cpu(req->virt_start);
534     uint64_t virt_end = le64_to_cpu(req->virt_end);
535     uint32_t flags = le32_to_cpu(req->flags);
536     VirtIOIOMMUDomain *domain;
537     VirtIOIOMMUInterval *interval;
538     VirtIOIOMMUMapping *mapping;
539     VirtIOIOMMUEndpoint *ep;
540 
541     if (flags & ~VIRTIO_IOMMU_MAP_F_MASK) {
542         return VIRTIO_IOMMU_S_INVAL;
543     }
544 
545     domain = g_tree_lookup(s->domains, GUINT_TO_POINTER(domain_id));
546     if (!domain) {
547         return VIRTIO_IOMMU_S_NOENT;
548     }
549 
550     if (domain->bypass) {
551         return VIRTIO_IOMMU_S_INVAL;
552     }
553 
554     interval = g_malloc0(sizeof(*interval));
555 
556     interval->low = virt_start;
557     interval->high = virt_end;
558 
559     mapping = g_tree_lookup(domain->mappings, (gpointer)interval);
560     if (mapping) {
561         g_free(interval);
562         return VIRTIO_IOMMU_S_INVAL;
563     }
564 
565     trace_virtio_iommu_map(domain_id, virt_start, virt_end, phys_start, flags);
566 
567     mapping = g_malloc0(sizeof(*mapping));
568     mapping->phys_addr = phys_start;
569     mapping->flags = flags;
570 
571     g_tree_insert(domain->mappings, interval, mapping);
572 
573     QLIST_FOREACH(ep, &domain->endpoint_list, next) {
574         virtio_iommu_notify_map(ep->iommu_mr, virt_start, virt_end, phys_start,
575                                 flags);
576     }
577 
578     return VIRTIO_IOMMU_S_OK;
579 }
580 
581 static int virtio_iommu_unmap(VirtIOIOMMU *s,
582                               struct virtio_iommu_req_unmap *req)
583 {
584     uint32_t domain_id = le32_to_cpu(req->domain);
585     uint64_t virt_start = le64_to_cpu(req->virt_start);
586     uint64_t virt_end = le64_to_cpu(req->virt_end);
587     VirtIOIOMMUMapping *iter_val;
588     VirtIOIOMMUInterval interval, *iter_key;
589     VirtIOIOMMUDomain *domain;
590     VirtIOIOMMUEndpoint *ep;
591     int ret = VIRTIO_IOMMU_S_OK;
592 
593     trace_virtio_iommu_unmap(domain_id, virt_start, virt_end);
594 
595     domain = g_tree_lookup(s->domains, GUINT_TO_POINTER(domain_id));
596     if (!domain) {
597         return VIRTIO_IOMMU_S_NOENT;
598     }
599 
600     if (domain->bypass) {
601         return VIRTIO_IOMMU_S_INVAL;
602     }
603 
604     interval.low = virt_start;
605     interval.high = virt_end;
606 
607     while (g_tree_lookup_extended(domain->mappings, &interval,
608                                   (void **)&iter_key, (void**)&iter_val)) {
609         uint64_t current_low = iter_key->low;
610         uint64_t current_high = iter_key->high;
611 
612         if (interval.low <= current_low && interval.high >= current_high) {
613             QLIST_FOREACH(ep, &domain->endpoint_list, next) {
614                 virtio_iommu_notify_unmap(ep->iommu_mr, current_low,
615                                           current_high);
616             }
617             g_tree_remove(domain->mappings, iter_key);
618             trace_virtio_iommu_unmap_done(domain_id, current_low, current_high);
619         } else {
620             ret = VIRTIO_IOMMU_S_RANGE;
621             break;
622         }
623     }
624     return ret;
625 }
626 
627 static ssize_t virtio_iommu_fill_resv_mem_prop(VirtIOIOMMU *s, uint32_t ep,
628                                                uint8_t *buf, size_t free)
629 {
630     struct virtio_iommu_probe_resv_mem prop = {};
631     size_t size = sizeof(prop), length = size - sizeof(prop.head), total;
632     int i;
633 
634     total = size * s->nb_reserved_regions;
635 
636     if (total > free) {
637         return -ENOSPC;
638     }
639 
640     for (i = 0; i < s->nb_reserved_regions; i++) {
641         unsigned subtype = s->reserved_regions[i].type;
642 
643         assert(subtype == VIRTIO_IOMMU_RESV_MEM_T_RESERVED ||
644                subtype == VIRTIO_IOMMU_RESV_MEM_T_MSI);
645         prop.head.type = cpu_to_le16(VIRTIO_IOMMU_PROBE_T_RESV_MEM);
646         prop.head.length = cpu_to_le16(length);
647         prop.subtype = subtype;
648         prop.start = cpu_to_le64(s->reserved_regions[i].low);
649         prop.end = cpu_to_le64(s->reserved_regions[i].high);
650 
651         memcpy(buf, &prop, size);
652 
653         trace_virtio_iommu_fill_resv_property(ep, prop.subtype,
654                                               prop.start, prop.end);
655         buf += size;
656     }
657     return total;
658 }
659 
660 /**
661  * virtio_iommu_probe - Fill the probe request buffer with
662  * the properties the device is able to return
663  */
664 static int virtio_iommu_probe(VirtIOIOMMU *s,
665                               struct virtio_iommu_req_probe *req,
666                               uint8_t *buf)
667 {
668     uint32_t ep_id = le32_to_cpu(req->endpoint);
669     size_t free = VIOMMU_PROBE_SIZE;
670     ssize_t count;
671 
672     if (!virtio_iommu_mr(s, ep_id)) {
673         return VIRTIO_IOMMU_S_NOENT;
674     }
675 
676     count = virtio_iommu_fill_resv_mem_prop(s, ep_id, buf, free);
677     if (count < 0) {
678         return VIRTIO_IOMMU_S_INVAL;
679     }
680     buf += count;
681     free -= count;
682 
683     return VIRTIO_IOMMU_S_OK;
684 }
685 
686 static int virtio_iommu_iov_to_req(struct iovec *iov,
687                                    unsigned int iov_cnt,
688                                    void *req, size_t payload_sz)
689 {
690     size_t sz = iov_to_buf(iov, iov_cnt, 0, req, payload_sz);
691 
692     if (unlikely(sz != payload_sz)) {
693         return VIRTIO_IOMMU_S_INVAL;
694     }
695     return 0;
696 }
697 
698 #define virtio_iommu_handle_req(__req)                                  \
699 static int virtio_iommu_handle_ ## __req(VirtIOIOMMU *s,                \
700                                          struct iovec *iov,             \
701                                          unsigned int iov_cnt)          \
702 {                                                                       \
703     struct virtio_iommu_req_ ## __req req;                              \
704     int ret = virtio_iommu_iov_to_req(iov, iov_cnt, &req,               \
705                     sizeof(req) - sizeof(struct virtio_iommu_req_tail));\
706                                                                         \
707     return ret ? ret : virtio_iommu_ ## __req(s, &req);                 \
708 }
709 
710 virtio_iommu_handle_req(attach)
711 virtio_iommu_handle_req(detach)
712 virtio_iommu_handle_req(map)
713 virtio_iommu_handle_req(unmap)
714 
715 static int virtio_iommu_handle_probe(VirtIOIOMMU *s,
716                                      struct iovec *iov,
717                                      unsigned int iov_cnt,
718                                      uint8_t *buf)
719 {
720     struct virtio_iommu_req_probe req;
721     int ret = virtio_iommu_iov_to_req(iov, iov_cnt, &req, sizeof(req));
722 
723     return ret ? ret : virtio_iommu_probe(s, &req, buf);
724 }
725 
726 static void virtio_iommu_handle_command(VirtIODevice *vdev, VirtQueue *vq)
727 {
728     VirtIOIOMMU *s = VIRTIO_IOMMU(vdev);
729     struct virtio_iommu_req_head head;
730     struct virtio_iommu_req_tail tail = {};
731     VirtQueueElement *elem;
732     unsigned int iov_cnt;
733     struct iovec *iov;
734     void *buf = NULL;
735     size_t sz;
736 
737     for (;;) {
738         size_t output_size = sizeof(tail);
739 
740         elem = virtqueue_pop(vq, sizeof(VirtQueueElement));
741         if (!elem) {
742             return;
743         }
744 
745         if (iov_size(elem->in_sg, elem->in_num) < sizeof(tail) ||
746             iov_size(elem->out_sg, elem->out_num) < sizeof(head)) {
747             virtio_error(vdev, "virtio-iommu bad head/tail size");
748             virtqueue_detach_element(vq, elem, 0);
749             g_free(elem);
750             break;
751         }
752 
753         iov_cnt = elem->out_num;
754         iov = elem->out_sg;
755         sz = iov_to_buf(iov, iov_cnt, 0, &head, sizeof(head));
756         if (unlikely(sz != sizeof(head))) {
757             tail.status = VIRTIO_IOMMU_S_DEVERR;
758             goto out;
759         }
760         qemu_rec_mutex_lock(&s->mutex);
761         switch (head.type) {
762         case VIRTIO_IOMMU_T_ATTACH:
763             tail.status = virtio_iommu_handle_attach(s, iov, iov_cnt);
764             break;
765         case VIRTIO_IOMMU_T_DETACH:
766             tail.status = virtio_iommu_handle_detach(s, iov, iov_cnt);
767             break;
768         case VIRTIO_IOMMU_T_MAP:
769             tail.status = virtio_iommu_handle_map(s, iov, iov_cnt);
770             break;
771         case VIRTIO_IOMMU_T_UNMAP:
772             tail.status = virtio_iommu_handle_unmap(s, iov, iov_cnt);
773             break;
774         case VIRTIO_IOMMU_T_PROBE:
775         {
776             struct virtio_iommu_req_tail *ptail;
777 
778             output_size = s->config.probe_size + sizeof(tail);
779             buf = g_malloc0(output_size);
780 
781             ptail = buf + s->config.probe_size;
782             ptail->status = virtio_iommu_handle_probe(s, iov, iov_cnt, buf);
783             break;
784         }
785         default:
786             tail.status = VIRTIO_IOMMU_S_UNSUPP;
787         }
788         qemu_rec_mutex_unlock(&s->mutex);
789 
790 out:
791         sz = iov_from_buf(elem->in_sg, elem->in_num, 0,
792                           buf ? buf : &tail, output_size);
793         assert(sz == output_size);
794 
795         virtqueue_push(vq, elem, sz);
796         virtio_notify(vdev, vq);
797         g_free(elem);
798         g_free(buf);
799         buf = NULL;
800     }
801 }
802 
803 static void virtio_iommu_report_fault(VirtIOIOMMU *viommu, uint8_t reason,
804                                       int flags, uint32_t endpoint,
805                                       uint64_t address)
806 {
807     VirtIODevice *vdev = &viommu->parent_obj;
808     VirtQueue *vq = viommu->event_vq;
809     struct virtio_iommu_fault fault;
810     VirtQueueElement *elem;
811     size_t sz;
812 
813     memset(&fault, 0, sizeof(fault));
814     fault.reason = reason;
815     fault.flags = cpu_to_le32(flags);
816     fault.endpoint = cpu_to_le32(endpoint);
817     fault.address = cpu_to_le64(address);
818 
819     elem = virtqueue_pop(vq, sizeof(VirtQueueElement));
820 
821     if (!elem) {
822         error_report_once(
823             "no buffer available in event queue to report event");
824         return;
825     }
826 
827     if (iov_size(elem->in_sg, elem->in_num) < sizeof(fault)) {
828         virtio_error(vdev, "error buffer of wrong size");
829         virtqueue_detach_element(vq, elem, 0);
830         g_free(elem);
831         return;
832     }
833 
834     sz = iov_from_buf(elem->in_sg, elem->in_num, 0,
835                       &fault, sizeof(fault));
836     assert(sz == sizeof(fault));
837 
838     trace_virtio_iommu_report_fault(reason, flags, endpoint, address);
839     virtqueue_push(vq, elem, sz);
840     virtio_notify(vdev, vq);
841     g_free(elem);
842 
843 }
844 
845 static IOMMUTLBEntry virtio_iommu_translate(IOMMUMemoryRegion *mr, hwaddr addr,
846                                             IOMMUAccessFlags flag,
847                                             int iommu_idx)
848 {
849     IOMMUDevice *sdev = container_of(mr, IOMMUDevice, iommu_mr);
850     VirtIOIOMMUInterval interval, *mapping_key;
851     VirtIOIOMMUMapping *mapping_value;
852     VirtIOIOMMU *s = sdev->viommu;
853     bool read_fault, write_fault;
854     VirtIOIOMMUEndpoint *ep;
855     uint32_t sid, flags;
856     bool bypass_allowed;
857     int granule;
858     bool found;
859     int i;
860 
861     interval.low = addr;
862     interval.high = addr + 1;
863     granule = ctz64(s->config.page_size_mask);
864 
865     IOMMUTLBEntry entry = {
866         .target_as = &address_space_memory,
867         .iova = addr,
868         .translated_addr = addr,
869         .addr_mask = BIT_ULL(granule) - 1,
870         .perm = IOMMU_NONE,
871     };
872 
873     bypass_allowed = s->config.bypass;
874 
875     sid = virtio_iommu_get_bdf(sdev);
876 
877     trace_virtio_iommu_translate(mr->parent_obj.name, sid, addr, flag);
878     qemu_rec_mutex_lock(&s->mutex);
879 
880     ep = g_tree_lookup(s->endpoints, GUINT_TO_POINTER(sid));
881 
882     if (bypass_allowed)
883         assert(ep && ep->domain && !ep->domain->bypass);
884 
885     if (!ep) {
886         if (!bypass_allowed) {
887             error_report_once("%s sid=%d is not known!!", __func__, sid);
888             virtio_iommu_report_fault(s, VIRTIO_IOMMU_FAULT_R_UNKNOWN,
889                                       VIRTIO_IOMMU_FAULT_F_ADDRESS,
890                                       sid, addr);
891         } else {
892             entry.perm = flag;
893         }
894         goto unlock;
895     }
896 
897     for (i = 0; i < s->nb_reserved_regions; i++) {
898         ReservedRegion *reg = &s->reserved_regions[i];
899 
900         if (addr >= reg->low && addr <= reg->high) {
901             switch (reg->type) {
902             case VIRTIO_IOMMU_RESV_MEM_T_MSI:
903                 entry.perm = flag;
904                 break;
905             case VIRTIO_IOMMU_RESV_MEM_T_RESERVED:
906             default:
907                 virtio_iommu_report_fault(s, VIRTIO_IOMMU_FAULT_R_MAPPING,
908                                           VIRTIO_IOMMU_FAULT_F_ADDRESS,
909                                           sid, addr);
910                 break;
911             }
912             goto unlock;
913         }
914     }
915 
916     if (!ep->domain) {
917         if (!bypass_allowed) {
918             error_report_once("%s %02x:%02x.%01x not attached to any domain",
919                               __func__, PCI_BUS_NUM(sid),
920                               PCI_SLOT(sid), PCI_FUNC(sid));
921             virtio_iommu_report_fault(s, VIRTIO_IOMMU_FAULT_R_DOMAIN,
922                                       VIRTIO_IOMMU_FAULT_F_ADDRESS,
923                                       sid, addr);
924         } else {
925             entry.perm = flag;
926         }
927         goto unlock;
928     } else if (ep->domain->bypass) {
929         entry.perm = flag;
930         goto unlock;
931     }
932 
933     found = g_tree_lookup_extended(ep->domain->mappings, (gpointer)(&interval),
934                                    (void **)&mapping_key,
935                                    (void **)&mapping_value);
936     if (!found) {
937         error_report_once("%s no mapping for 0x%"PRIx64" for sid=%d",
938                           __func__, addr, sid);
939         virtio_iommu_report_fault(s, VIRTIO_IOMMU_FAULT_R_MAPPING,
940                                   VIRTIO_IOMMU_FAULT_F_ADDRESS,
941                                   sid, addr);
942         goto unlock;
943     }
944 
945     read_fault = (flag & IOMMU_RO) &&
946                     !(mapping_value->flags & VIRTIO_IOMMU_MAP_F_READ);
947     write_fault = (flag & IOMMU_WO) &&
948                     !(mapping_value->flags & VIRTIO_IOMMU_MAP_F_WRITE);
949 
950     flags = read_fault ? VIRTIO_IOMMU_FAULT_F_READ : 0;
951     flags |= write_fault ? VIRTIO_IOMMU_FAULT_F_WRITE : 0;
952     if (flags) {
953         error_report_once("%s permission error on 0x%"PRIx64"(%d): allowed=%d",
954                           __func__, addr, flag, mapping_value->flags);
955         flags |= VIRTIO_IOMMU_FAULT_F_ADDRESS;
956         virtio_iommu_report_fault(s, VIRTIO_IOMMU_FAULT_R_MAPPING,
957                                   flags | VIRTIO_IOMMU_FAULT_F_ADDRESS,
958                                   sid, addr);
959         goto unlock;
960     }
961     entry.translated_addr = addr - mapping_key->low + mapping_value->phys_addr;
962     entry.perm = flag;
963     trace_virtio_iommu_translate_out(addr, entry.translated_addr, sid);
964 
965 unlock:
966     qemu_rec_mutex_unlock(&s->mutex);
967     return entry;
968 }
969 
970 static void virtio_iommu_get_config(VirtIODevice *vdev, uint8_t *config_data)
971 {
972     VirtIOIOMMU *dev = VIRTIO_IOMMU(vdev);
973     struct virtio_iommu_config *dev_config = &dev->config;
974     struct virtio_iommu_config *out_config = (void *)config_data;
975 
976     out_config->page_size_mask = cpu_to_le64(dev_config->page_size_mask);
977     out_config->input_range.start = cpu_to_le64(dev_config->input_range.start);
978     out_config->input_range.end = cpu_to_le64(dev_config->input_range.end);
979     out_config->domain_range.start = cpu_to_le32(dev_config->domain_range.start);
980     out_config->domain_range.end = cpu_to_le32(dev_config->domain_range.end);
981     out_config->probe_size = cpu_to_le32(dev_config->probe_size);
982     out_config->bypass = dev_config->bypass;
983 
984     trace_virtio_iommu_get_config(dev_config->page_size_mask,
985                                   dev_config->input_range.start,
986                                   dev_config->input_range.end,
987                                   dev_config->domain_range.start,
988                                   dev_config->domain_range.end,
989                                   dev_config->probe_size,
990                                   dev_config->bypass);
991 }
992 
993 static void virtio_iommu_set_config(VirtIODevice *vdev,
994                                     const uint8_t *config_data)
995 {
996     VirtIOIOMMU *dev = VIRTIO_IOMMU(vdev);
997     struct virtio_iommu_config *dev_config = &dev->config;
998     const struct virtio_iommu_config *in_config = (void *)config_data;
999 
1000     if (in_config->bypass != dev_config->bypass) {
1001         if (!virtio_vdev_has_feature(vdev, VIRTIO_IOMMU_F_BYPASS_CONFIG)) {
1002             virtio_error(vdev, "cannot set config.bypass");
1003             return;
1004         } else if (in_config->bypass != 0 && in_config->bypass != 1) {
1005             virtio_error(vdev, "invalid config.bypass value '%u'",
1006                          in_config->bypass);
1007             return;
1008         }
1009         dev_config->bypass = in_config->bypass;
1010         virtio_iommu_switch_address_space_all(dev);
1011     }
1012 
1013     trace_virtio_iommu_set_config(in_config->bypass);
1014 }
1015 
1016 static uint64_t virtio_iommu_get_features(VirtIODevice *vdev, uint64_t f,
1017                                           Error **errp)
1018 {
1019     VirtIOIOMMU *dev = VIRTIO_IOMMU(vdev);
1020 
1021     f |= dev->features;
1022     trace_virtio_iommu_get_features(f);
1023     return f;
1024 }
1025 
1026 static gint int_cmp(gconstpointer a, gconstpointer b, gpointer user_data)
1027 {
1028     guint ua = GPOINTER_TO_UINT(a);
1029     guint ub = GPOINTER_TO_UINT(b);
1030     return (ua > ub) - (ua < ub);
1031 }
1032 
1033 static gboolean virtio_iommu_remap(gpointer key, gpointer value, gpointer data)
1034 {
1035     VirtIOIOMMUMapping *mapping = (VirtIOIOMMUMapping *) value;
1036     VirtIOIOMMUInterval *interval = (VirtIOIOMMUInterval *) key;
1037     IOMMUMemoryRegion *mr = (IOMMUMemoryRegion *) data;
1038 
1039     trace_virtio_iommu_remap(mr->parent_obj.name, interval->low, interval->high,
1040                              mapping->phys_addr);
1041     virtio_iommu_notify_map(mr, interval->low, interval->high,
1042                             mapping->phys_addr, mapping->flags);
1043     return false;
1044 }
1045 
1046 static void virtio_iommu_replay(IOMMUMemoryRegion *mr, IOMMUNotifier *n)
1047 {
1048     IOMMUDevice *sdev = container_of(mr, IOMMUDevice, iommu_mr);
1049     VirtIOIOMMU *s = sdev->viommu;
1050     uint32_t sid;
1051     VirtIOIOMMUEndpoint *ep;
1052 
1053     sid = virtio_iommu_get_bdf(sdev);
1054 
1055     qemu_rec_mutex_lock(&s->mutex);
1056 
1057     if (!s->endpoints) {
1058         goto unlock;
1059     }
1060 
1061     ep = g_tree_lookup(s->endpoints, GUINT_TO_POINTER(sid));
1062     if (!ep || !ep->domain) {
1063         goto unlock;
1064     }
1065 
1066     g_tree_foreach(ep->domain->mappings, virtio_iommu_remap, mr);
1067 
1068 unlock:
1069     qemu_rec_mutex_unlock(&s->mutex);
1070 }
1071 
1072 static int virtio_iommu_notify_flag_changed(IOMMUMemoryRegion *iommu_mr,
1073                                             IOMMUNotifierFlag old,
1074                                             IOMMUNotifierFlag new,
1075                                             Error **errp)
1076 {
1077     if (new & IOMMU_NOTIFIER_DEVIOTLB_UNMAP) {
1078         error_setg(errp, "Virtio-iommu does not support dev-iotlb yet");
1079         return -EINVAL;
1080     }
1081 
1082     if (old == IOMMU_NOTIFIER_NONE) {
1083         trace_virtio_iommu_notify_flag_add(iommu_mr->parent_obj.name);
1084     } else if (new == IOMMU_NOTIFIER_NONE) {
1085         trace_virtio_iommu_notify_flag_del(iommu_mr->parent_obj.name);
1086     }
1087     return 0;
1088 }
1089 
1090 /*
1091  * The default mask (TARGET_PAGE_MASK) is the smallest supported guest granule,
1092  * for example 0xfffffffffffff000. When an assigned device has page size
1093  * restrictions due to the hardware IOMMU configuration, apply this restriction
1094  * to the mask.
1095  */
1096 static int virtio_iommu_set_page_size_mask(IOMMUMemoryRegion *mr,
1097                                            uint64_t new_mask,
1098                                            Error **errp)
1099 {
1100     IOMMUDevice *sdev = container_of(mr, IOMMUDevice, iommu_mr);
1101     VirtIOIOMMU *s = sdev->viommu;
1102     uint64_t cur_mask = s->config.page_size_mask;
1103 
1104     trace_virtio_iommu_set_page_size_mask(mr->parent_obj.name, cur_mask,
1105                                           new_mask);
1106 
1107     if ((cur_mask & new_mask) == 0) {
1108         error_setg(errp, "virtio-iommu %s reports a page size mask 0x%"PRIx64
1109                    " incompatible with currently supported mask 0x%"PRIx64,
1110                    mr->parent_obj.name, new_mask, cur_mask);
1111         return -1;
1112     }
1113 
1114     /*
1115      * Once the granule is frozen we can't change the mask anymore. If by
1116      * chance the hotplugged device supports the same granule, we can still
1117      * accept it.
1118      */
1119     if (s->granule_frozen) {
1120         int cur_granule = ctz64(cur_mask);
1121 
1122         if (!(BIT_ULL(cur_granule) & new_mask)) {
1123             error_setg(errp, "virtio-iommu %s does not support frozen granule 0x%llx",
1124                        mr->parent_obj.name, BIT_ULL(cur_granule));
1125             return -1;
1126         }
1127         return 0;
1128     }
1129 
1130     s->config.page_size_mask &= new_mask;
1131     return 0;
1132 }
1133 
1134 static void virtio_iommu_system_reset(void *opaque)
1135 {
1136     VirtIOIOMMU *s = opaque;
1137 
1138     trace_virtio_iommu_system_reset();
1139 
1140     /*
1141      * config.bypass is sticky across device reset, but should be restored on
1142      * system reset
1143      */
1144     s->config.bypass = s->boot_bypass;
1145     virtio_iommu_switch_address_space_all(s);
1146 
1147 }
1148 
1149 static void virtio_iommu_freeze_granule(Notifier *notifier, void *data)
1150 {
1151     VirtIOIOMMU *s = container_of(notifier, VirtIOIOMMU, machine_done);
1152     int granule;
1153 
1154     if (likely(s->config.bypass)) {
1155         /*
1156          * Transient IOMMU MR enable to collect page_size_mask requirements
1157          * through memory_region_iommu_set_page_size_mask() called by
1158          * VFIO region_add() callback
1159          */
1160          s->config.bypass = false;
1161          virtio_iommu_switch_address_space_all(s);
1162          /* restore default */
1163          s->config.bypass = true;
1164          virtio_iommu_switch_address_space_all(s);
1165     }
1166     s->granule_frozen = true;
1167     granule = ctz64(s->config.page_size_mask);
1168     trace_virtio_iommu_freeze_granule(BIT_ULL(granule));
1169 }
1170 
1171 static void virtio_iommu_device_realize(DeviceState *dev, Error **errp)
1172 {
1173     VirtIODevice *vdev = VIRTIO_DEVICE(dev);
1174     VirtIOIOMMU *s = VIRTIO_IOMMU(dev);
1175 
1176     virtio_init(vdev, VIRTIO_ID_IOMMU, sizeof(struct virtio_iommu_config));
1177 
1178     memset(s->iommu_pcibus_by_bus_num, 0, sizeof(s->iommu_pcibus_by_bus_num));
1179 
1180     s->req_vq = virtio_add_queue(vdev, VIOMMU_DEFAULT_QUEUE_SIZE,
1181                              virtio_iommu_handle_command);
1182     s->event_vq = virtio_add_queue(vdev, VIOMMU_DEFAULT_QUEUE_SIZE, NULL);
1183 
1184     /*
1185      * config.bypass is needed to get initial address space early, such as
1186      * in vfio realize
1187      */
1188     s->config.bypass = s->boot_bypass;
1189     s->config.page_size_mask = qemu_target_page_mask();
1190     s->config.input_range.end = UINT64_MAX;
1191     s->config.domain_range.end = UINT32_MAX;
1192     s->config.probe_size = VIOMMU_PROBE_SIZE;
1193 
1194     virtio_add_feature(&s->features, VIRTIO_RING_F_EVENT_IDX);
1195     virtio_add_feature(&s->features, VIRTIO_RING_F_INDIRECT_DESC);
1196     virtio_add_feature(&s->features, VIRTIO_F_VERSION_1);
1197     virtio_add_feature(&s->features, VIRTIO_IOMMU_F_INPUT_RANGE);
1198     virtio_add_feature(&s->features, VIRTIO_IOMMU_F_DOMAIN_RANGE);
1199     virtio_add_feature(&s->features, VIRTIO_IOMMU_F_MAP_UNMAP);
1200     virtio_add_feature(&s->features, VIRTIO_IOMMU_F_MMIO);
1201     virtio_add_feature(&s->features, VIRTIO_IOMMU_F_PROBE);
1202     virtio_add_feature(&s->features, VIRTIO_IOMMU_F_BYPASS_CONFIG);
1203 
1204     qemu_rec_mutex_init(&s->mutex);
1205 
1206     s->as_by_busptr = g_hash_table_new_full(NULL, NULL, NULL, g_free);
1207 
1208     if (s->primary_bus) {
1209         pci_setup_iommu(s->primary_bus, virtio_iommu_find_add_as, s);
1210     } else {
1211         error_setg(errp, "VIRTIO-IOMMU is not attached to any PCI bus!");
1212     }
1213 
1214     s->machine_done.notify = virtio_iommu_freeze_granule;
1215     qemu_add_machine_init_done_notifier(&s->machine_done);
1216 
1217     qemu_register_reset(virtio_iommu_system_reset, s);
1218 }
1219 
1220 static void virtio_iommu_device_unrealize(DeviceState *dev)
1221 {
1222     VirtIODevice *vdev = VIRTIO_DEVICE(dev);
1223     VirtIOIOMMU *s = VIRTIO_IOMMU(dev);
1224 
1225     qemu_unregister_reset(virtio_iommu_system_reset, s);
1226     qemu_remove_machine_init_done_notifier(&s->machine_done);
1227 
1228     g_hash_table_destroy(s->as_by_busptr);
1229     if (s->domains) {
1230         g_tree_destroy(s->domains);
1231     }
1232     if (s->endpoints) {
1233         g_tree_destroy(s->endpoints);
1234     }
1235 
1236     qemu_rec_mutex_destroy(&s->mutex);
1237 
1238     virtio_delete_queue(s->req_vq);
1239     virtio_delete_queue(s->event_vq);
1240     virtio_cleanup(vdev);
1241 }
1242 
1243 static void virtio_iommu_device_reset(VirtIODevice *vdev)
1244 {
1245     VirtIOIOMMU *s = VIRTIO_IOMMU(vdev);
1246 
1247     trace_virtio_iommu_device_reset();
1248 
1249     if (s->domains) {
1250         g_tree_destroy(s->domains);
1251     }
1252     if (s->endpoints) {
1253         g_tree_destroy(s->endpoints);
1254     }
1255     s->domains = g_tree_new_full((GCompareDataFunc)int_cmp,
1256                                  NULL, NULL, virtio_iommu_put_domain);
1257     s->endpoints = g_tree_new_full((GCompareDataFunc)int_cmp,
1258                                    NULL, NULL, virtio_iommu_put_endpoint);
1259 }
1260 
1261 static void virtio_iommu_set_status(VirtIODevice *vdev, uint8_t status)
1262 {
1263     trace_virtio_iommu_device_status(status);
1264 }
1265 
1266 static void virtio_iommu_instance_init(Object *obj)
1267 {
1268 }
1269 
1270 #define VMSTATE_INTERVAL                               \
1271 {                                                      \
1272     .name = "interval",                                \
1273     .version_id = 1,                                   \
1274     .minimum_version_id = 1,                           \
1275     .fields = (VMStateField[]) {                       \
1276         VMSTATE_UINT64(low, VirtIOIOMMUInterval),      \
1277         VMSTATE_UINT64(high, VirtIOIOMMUInterval),     \
1278         VMSTATE_END_OF_LIST()                          \
1279     }                                                  \
1280 }
1281 
1282 #define VMSTATE_MAPPING                               \
1283 {                                                     \
1284     .name = "mapping",                                \
1285     .version_id = 1,                                  \
1286     .minimum_version_id = 1,                          \
1287     .fields = (VMStateField[]) {                      \
1288         VMSTATE_UINT64(phys_addr, VirtIOIOMMUMapping),\
1289         VMSTATE_UINT32(flags, VirtIOIOMMUMapping),    \
1290         VMSTATE_END_OF_LIST()                         \
1291     },                                                \
1292 }
1293 
1294 static const VMStateDescription vmstate_interval_mapping[2] = {
1295     VMSTATE_MAPPING,   /* value */
1296     VMSTATE_INTERVAL   /* key   */
1297 };
1298 
1299 static int domain_preload(void *opaque)
1300 {
1301     VirtIOIOMMUDomain *domain = opaque;
1302 
1303     domain->mappings = g_tree_new_full((GCompareDataFunc)interval_cmp,
1304                                        NULL, g_free, g_free);
1305     return 0;
1306 }
1307 
1308 static const VMStateDescription vmstate_endpoint = {
1309     .name = "endpoint",
1310     .version_id = 1,
1311     .minimum_version_id = 1,
1312     .fields = (VMStateField[]) {
1313         VMSTATE_UINT32(id, VirtIOIOMMUEndpoint),
1314         VMSTATE_END_OF_LIST()
1315     }
1316 };
1317 
1318 static const VMStateDescription vmstate_domain = {
1319     .name = "domain",
1320     .version_id = 2,
1321     .minimum_version_id = 2,
1322     .pre_load = domain_preload,
1323     .fields = (VMStateField[]) {
1324         VMSTATE_UINT32(id, VirtIOIOMMUDomain),
1325         VMSTATE_GTREE_V(mappings, VirtIOIOMMUDomain, 1,
1326                         vmstate_interval_mapping,
1327                         VirtIOIOMMUInterval, VirtIOIOMMUMapping),
1328         VMSTATE_QLIST_V(endpoint_list, VirtIOIOMMUDomain, 1,
1329                         vmstate_endpoint, VirtIOIOMMUEndpoint, next),
1330         VMSTATE_BOOL_V(bypass, VirtIOIOMMUDomain, 2),
1331         VMSTATE_END_OF_LIST()
1332     }
1333 };
1334 
1335 static gboolean reconstruct_endpoints(gpointer key, gpointer value,
1336                                       gpointer data)
1337 {
1338     VirtIOIOMMU *s = (VirtIOIOMMU *)data;
1339     VirtIOIOMMUDomain *d = (VirtIOIOMMUDomain *)value;
1340     VirtIOIOMMUEndpoint *iter;
1341     IOMMUMemoryRegion *mr;
1342 
1343     QLIST_FOREACH(iter, &d->endpoint_list, next) {
1344         mr = virtio_iommu_mr(s, iter->id);
1345         assert(mr);
1346 
1347         iter->domain = d;
1348         iter->iommu_mr = mr;
1349         g_tree_insert(s->endpoints, GUINT_TO_POINTER(iter->id), iter);
1350     }
1351     return false; /* continue the domain traversal */
1352 }
1353 
1354 static int iommu_post_load(void *opaque, int version_id)
1355 {
1356     VirtIOIOMMU *s = opaque;
1357 
1358     g_tree_foreach(s->domains, reconstruct_endpoints, s);
1359 
1360     /*
1361      * Memory regions are dynamically turned on/off depending on
1362      * 'config.bypass' and attached domain type if there is. After
1363      * migration, we need to make sure the memory regions are
1364      * still correct.
1365      */
1366     virtio_iommu_switch_address_space_all(s);
1367     return 0;
1368 }
1369 
1370 static const VMStateDescription vmstate_virtio_iommu_device = {
1371     .name = "virtio-iommu-device",
1372     .minimum_version_id = 2,
1373     .version_id = 2,
1374     .post_load = iommu_post_load,
1375     .fields = (VMStateField[]) {
1376         VMSTATE_GTREE_DIRECT_KEY_V(domains, VirtIOIOMMU, 2,
1377                                    &vmstate_domain, VirtIOIOMMUDomain),
1378         VMSTATE_UINT8_V(config.bypass, VirtIOIOMMU, 2),
1379         VMSTATE_END_OF_LIST()
1380     },
1381 };
1382 
1383 static const VMStateDescription vmstate_virtio_iommu = {
1384     .name = "virtio-iommu",
1385     .minimum_version_id = 2,
1386     .priority = MIG_PRI_IOMMU,
1387     .version_id = 2,
1388     .fields = (VMStateField[]) {
1389         VMSTATE_VIRTIO_DEVICE,
1390         VMSTATE_END_OF_LIST()
1391     },
1392 };
1393 
1394 static Property virtio_iommu_properties[] = {
1395     DEFINE_PROP_LINK("primary-bus", VirtIOIOMMU, primary_bus,
1396                      TYPE_PCI_BUS, PCIBus *),
1397     DEFINE_PROP_BOOL("boot-bypass", VirtIOIOMMU, boot_bypass, true),
1398     DEFINE_PROP_END_OF_LIST(),
1399 };
1400 
1401 static void virtio_iommu_class_init(ObjectClass *klass, void *data)
1402 {
1403     DeviceClass *dc = DEVICE_CLASS(klass);
1404     VirtioDeviceClass *vdc = VIRTIO_DEVICE_CLASS(klass);
1405 
1406     device_class_set_props(dc, virtio_iommu_properties);
1407     dc->vmsd = &vmstate_virtio_iommu;
1408 
1409     set_bit(DEVICE_CATEGORY_MISC, dc->categories);
1410     vdc->realize = virtio_iommu_device_realize;
1411     vdc->unrealize = virtio_iommu_device_unrealize;
1412     vdc->reset = virtio_iommu_device_reset;
1413     vdc->get_config = virtio_iommu_get_config;
1414     vdc->set_config = virtio_iommu_set_config;
1415     vdc->get_features = virtio_iommu_get_features;
1416     vdc->set_status = virtio_iommu_set_status;
1417     vdc->vmsd = &vmstate_virtio_iommu_device;
1418 }
1419 
1420 static void virtio_iommu_memory_region_class_init(ObjectClass *klass,
1421                                                   void *data)
1422 {
1423     IOMMUMemoryRegionClass *imrc = IOMMU_MEMORY_REGION_CLASS(klass);
1424 
1425     imrc->translate = virtio_iommu_translate;
1426     imrc->replay = virtio_iommu_replay;
1427     imrc->notify_flag_changed = virtio_iommu_notify_flag_changed;
1428     imrc->iommu_set_page_size_mask = virtio_iommu_set_page_size_mask;
1429 }
1430 
1431 static const TypeInfo virtio_iommu_info = {
1432     .name = TYPE_VIRTIO_IOMMU,
1433     .parent = TYPE_VIRTIO_DEVICE,
1434     .instance_size = sizeof(VirtIOIOMMU),
1435     .instance_init = virtio_iommu_instance_init,
1436     .class_init = virtio_iommu_class_init,
1437 };
1438 
1439 static const TypeInfo virtio_iommu_memory_region_info = {
1440     .parent = TYPE_IOMMU_MEMORY_REGION,
1441     .name = TYPE_VIRTIO_IOMMU_MEMORY_REGION,
1442     .class_init = virtio_iommu_memory_region_class_init,
1443 };
1444 
1445 static void virtio_register_types(void)
1446 {
1447     type_register_static(&virtio_iommu_info);
1448     type_register_static(&virtio_iommu_memory_region_info);
1449 }
1450 
1451 type_init(virtio_register_types)
1452