xref: /openbmc/qemu/hw/virtio/virtio-iommu.c (revision a158c63b)
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 "qemu/range.h"
24 #include "qemu/reserved-region.h"
25 #include "exec/target_page.h"
26 #include "hw/qdev-properties.h"
27 #include "hw/virtio/virtio.h"
28 #include "sysemu/kvm.h"
29 #include "sysemu/reset.h"
30 #include "sysemu/sysemu.h"
31 #include "qemu/reserved-region.h"
32 #include "qemu/units.h"
33 #include "qapi/error.h"
34 #include "qemu/error-report.h"
35 #include "trace.h"
36 
37 #include "standard-headers/linux/virtio_ids.h"
38 
39 #include "hw/virtio/virtio-bus.h"
40 #include "hw/virtio/virtio-iommu.h"
41 #include "hw/pci/pci_bus.h"
42 #include "hw/pci/pci.h"
43 
44 /* Max size */
45 #define VIOMMU_DEFAULT_QUEUE_SIZE 256
46 #define VIOMMU_PROBE_SIZE 512
47 
48 typedef struct VirtIOIOMMUDomain {
49     uint32_t id;
50     bool bypass;
51     GTree *mappings;
52     QLIST_HEAD(, VirtIOIOMMUEndpoint) endpoint_list;
53 } VirtIOIOMMUDomain;
54 
55 typedef struct VirtIOIOMMUEndpoint {
56     uint32_t id;
57     VirtIOIOMMUDomain *domain;
58     IOMMUMemoryRegion *iommu_mr;
59     QLIST_ENTRY(VirtIOIOMMUEndpoint) next;
60 } VirtIOIOMMUEndpoint;
61 
62 typedef struct VirtIOIOMMUInterval {
63     uint64_t low;
64     uint64_t high;
65 } VirtIOIOMMUInterval;
66 
67 typedef struct VirtIOIOMMUMapping {
68     uint64_t phys_addr;
69     uint32_t flags;
70 } VirtIOIOMMUMapping;
71 
72 static inline uint16_t virtio_iommu_get_bdf(IOMMUDevice *dev)
73 {
74     return PCI_BUILD_BDF(pci_bus_num(dev->bus), dev->devfn);
75 }
76 
77 static bool virtio_iommu_device_bypassed(IOMMUDevice *sdev)
78 {
79     uint32_t sid;
80     bool bypassed;
81     VirtIOIOMMU *s = sdev->viommu;
82     VirtIOIOMMUEndpoint *ep;
83 
84     sid = virtio_iommu_get_bdf(sdev);
85 
86     qemu_rec_mutex_lock(&s->mutex);
87     /* need to check bypass before system reset */
88     if (!s->endpoints) {
89         bypassed = s->config.bypass;
90         goto unlock;
91     }
92 
93     ep = g_tree_lookup(s->endpoints, GUINT_TO_POINTER(sid));
94     if (!ep || !ep->domain) {
95         bypassed = s->config.bypass;
96     } else {
97         bypassed = ep->domain->bypass;
98     }
99 
100 unlock:
101     qemu_rec_mutex_unlock(&s->mutex);
102     return bypassed;
103 }
104 
105 /* Return whether the device is using IOMMU translation. */
106 static bool virtio_iommu_switch_address_space(IOMMUDevice *sdev)
107 {
108     bool use_remapping;
109 
110     assert(sdev);
111 
112     use_remapping = !virtio_iommu_device_bypassed(sdev);
113 
114     trace_virtio_iommu_switch_address_space(pci_bus_num(sdev->bus),
115                                             PCI_SLOT(sdev->devfn),
116                                             PCI_FUNC(sdev->devfn),
117                                             use_remapping);
118 
119     /* Turn off first then on the other */
120     if (use_remapping) {
121         memory_region_set_enabled(&sdev->bypass_mr, false);
122         memory_region_set_enabled(MEMORY_REGION(&sdev->iommu_mr), true);
123     } else {
124         memory_region_set_enabled(MEMORY_REGION(&sdev->iommu_mr), false);
125         memory_region_set_enabled(&sdev->bypass_mr, true);
126     }
127 
128     return use_remapping;
129 }
130 
131 static void virtio_iommu_switch_address_space_all(VirtIOIOMMU *s)
132 {
133     GHashTableIter iter;
134     IOMMUPciBus *iommu_pci_bus;
135     int i;
136 
137     g_hash_table_iter_init(&iter, s->as_by_busptr);
138     while (g_hash_table_iter_next(&iter, NULL, (void **)&iommu_pci_bus)) {
139         for (i = 0; i < PCI_DEVFN_MAX; i++) {
140             if (!iommu_pci_bus->pbdev[i]) {
141                 continue;
142             }
143             virtio_iommu_switch_address_space(iommu_pci_bus->pbdev[i]);
144         }
145     }
146 }
147 
148 /**
149  * The bus number is used for lookup when SID based operations occur.
150  * In that case we lazily populate the IOMMUPciBus array from the bus hash
151  * table. At the time the IOMMUPciBus is created (iommu_find_add_as), the bus
152  * numbers may not be always initialized yet.
153  */
154 static IOMMUPciBus *iommu_find_iommu_pcibus(VirtIOIOMMU *s, uint8_t bus_num)
155 {
156     IOMMUPciBus *iommu_pci_bus = s->iommu_pcibus_by_bus_num[bus_num];
157 
158     if (!iommu_pci_bus) {
159         GHashTableIter iter;
160 
161         g_hash_table_iter_init(&iter, s->as_by_busptr);
162         while (g_hash_table_iter_next(&iter, NULL, (void **)&iommu_pci_bus)) {
163             if (pci_bus_num(iommu_pci_bus->bus) == bus_num) {
164                 s->iommu_pcibus_by_bus_num[bus_num] = iommu_pci_bus;
165                 return iommu_pci_bus;
166             }
167         }
168         return NULL;
169     }
170     return iommu_pci_bus;
171 }
172 
173 static IOMMUMemoryRegion *virtio_iommu_mr(VirtIOIOMMU *s, uint32_t sid)
174 {
175     uint8_t bus_n, devfn;
176     IOMMUPciBus *iommu_pci_bus;
177     IOMMUDevice *dev;
178 
179     bus_n = PCI_BUS_NUM(sid);
180     iommu_pci_bus = iommu_find_iommu_pcibus(s, bus_n);
181     if (iommu_pci_bus) {
182         devfn = sid & (PCI_DEVFN_MAX - 1);
183         dev = iommu_pci_bus->pbdev[devfn];
184         if (dev) {
185             return &dev->iommu_mr;
186         }
187     }
188     return NULL;
189 }
190 
191 static gint interval_cmp(gconstpointer a, gconstpointer b, gpointer user_data)
192 {
193     VirtIOIOMMUInterval *inta = (VirtIOIOMMUInterval *)a;
194     VirtIOIOMMUInterval *intb = (VirtIOIOMMUInterval *)b;
195 
196     if (inta->high < intb->low) {
197         return -1;
198     } else if (intb->high < inta->low) {
199         return 1;
200     } else {
201         return 0;
202     }
203 }
204 
205 static void virtio_iommu_notify_map_unmap(IOMMUMemoryRegion *mr,
206                                           IOMMUTLBEvent *event,
207                                           hwaddr virt_start, hwaddr virt_end)
208 {
209     uint64_t delta = virt_end - virt_start;
210 
211     event->entry.iova = virt_start;
212     event->entry.addr_mask = delta;
213 
214     if (delta == UINT64_MAX) {
215         memory_region_notify_iommu(mr, 0, *event);
216     }
217 
218     while (virt_start != virt_end + 1) {
219         uint64_t mask = dma_aligned_pow2_mask(virt_start, virt_end, 64);
220 
221         event->entry.addr_mask = mask;
222         event->entry.iova = virt_start;
223         memory_region_notify_iommu(mr, 0, *event);
224         virt_start += mask + 1;
225         if (event->entry.perm != IOMMU_NONE) {
226             event->entry.translated_addr += mask + 1;
227         }
228     }
229 }
230 
231 static void virtio_iommu_notify_map(IOMMUMemoryRegion *mr, hwaddr virt_start,
232                                     hwaddr virt_end, hwaddr paddr,
233                                     uint32_t flags)
234 {
235     IOMMUTLBEvent event;
236     IOMMUAccessFlags perm = IOMMU_ACCESS_FLAG(flags & VIRTIO_IOMMU_MAP_F_READ,
237                                               flags & VIRTIO_IOMMU_MAP_F_WRITE);
238 
239     if (!(mr->iommu_notify_flags & IOMMU_NOTIFIER_MAP) ||
240         (flags & VIRTIO_IOMMU_MAP_F_MMIO) || !perm) {
241         return;
242     }
243 
244     trace_virtio_iommu_notify_map(mr->parent_obj.name, virt_start, virt_end,
245                                   paddr, perm);
246 
247     event.type = IOMMU_NOTIFIER_MAP;
248     event.entry.target_as = &address_space_memory;
249     event.entry.perm = perm;
250     event.entry.translated_addr = paddr;
251 
252     virtio_iommu_notify_map_unmap(mr, &event, virt_start, virt_end);
253 }
254 
255 static void virtio_iommu_notify_unmap(IOMMUMemoryRegion *mr, hwaddr virt_start,
256                                       hwaddr virt_end)
257 {
258     IOMMUTLBEvent event;
259 
260     if (!(mr->iommu_notify_flags & IOMMU_NOTIFIER_UNMAP)) {
261         return;
262     }
263 
264     trace_virtio_iommu_notify_unmap(mr->parent_obj.name, virt_start, virt_end);
265 
266     event.type = IOMMU_NOTIFIER_UNMAP;
267     event.entry.target_as = &address_space_memory;
268     event.entry.perm = IOMMU_NONE;
269     event.entry.translated_addr = 0;
270 
271     virtio_iommu_notify_map_unmap(mr, &event, virt_start, virt_end);
272 }
273 
274 static gboolean virtio_iommu_notify_unmap_cb(gpointer key, gpointer value,
275                                              gpointer data)
276 {
277     VirtIOIOMMUInterval *interval = (VirtIOIOMMUInterval *) key;
278     IOMMUMemoryRegion *mr = (IOMMUMemoryRegion *) data;
279 
280     virtio_iommu_notify_unmap(mr, interval->low, interval->high);
281 
282     return false;
283 }
284 
285 static gboolean virtio_iommu_notify_map_cb(gpointer key, gpointer value,
286                                            gpointer data)
287 {
288     VirtIOIOMMUMapping *mapping = (VirtIOIOMMUMapping *) value;
289     VirtIOIOMMUInterval *interval = (VirtIOIOMMUInterval *) key;
290     IOMMUMemoryRegion *mr = (IOMMUMemoryRegion *) data;
291 
292     virtio_iommu_notify_map(mr, interval->low, interval->high,
293                             mapping->phys_addr, mapping->flags);
294 
295     return false;
296 }
297 
298 static void virtio_iommu_detach_endpoint_from_domain(VirtIOIOMMUEndpoint *ep)
299 {
300     VirtIOIOMMUDomain *domain = ep->domain;
301     IOMMUDevice *sdev = container_of(ep->iommu_mr, IOMMUDevice, iommu_mr);
302 
303     if (!ep->domain) {
304         return;
305     }
306     g_tree_foreach(domain->mappings, virtio_iommu_notify_unmap_cb,
307                    ep->iommu_mr);
308     QLIST_REMOVE(ep, next);
309     ep->domain = NULL;
310     virtio_iommu_switch_address_space(sdev);
311 }
312 
313 static VirtIOIOMMUEndpoint *virtio_iommu_get_endpoint(VirtIOIOMMU *s,
314                                                       uint32_t ep_id)
315 {
316     VirtIOIOMMUEndpoint *ep;
317     IOMMUMemoryRegion *mr;
318 
319     ep = g_tree_lookup(s->endpoints, GUINT_TO_POINTER(ep_id));
320     if (ep) {
321         return ep;
322     }
323     mr = virtio_iommu_mr(s, ep_id);
324     if (!mr) {
325         return NULL;
326     }
327     ep = g_malloc0(sizeof(*ep));
328     ep->id = ep_id;
329     ep->iommu_mr = mr;
330     trace_virtio_iommu_get_endpoint(ep_id);
331     g_tree_insert(s->endpoints, GUINT_TO_POINTER(ep_id), ep);
332     return ep;
333 }
334 
335 static void virtio_iommu_put_endpoint(gpointer data)
336 {
337     VirtIOIOMMUEndpoint *ep = (VirtIOIOMMUEndpoint *)data;
338 
339     if (ep->domain) {
340         virtio_iommu_detach_endpoint_from_domain(ep);
341     }
342 
343     trace_virtio_iommu_put_endpoint(ep->id);
344     g_free(ep);
345 }
346 
347 static VirtIOIOMMUDomain *virtio_iommu_get_domain(VirtIOIOMMU *s,
348                                                   uint32_t domain_id,
349                                                   bool bypass)
350 {
351     VirtIOIOMMUDomain *domain;
352 
353     domain = g_tree_lookup(s->domains, GUINT_TO_POINTER(domain_id));
354     if (domain) {
355         if (domain->bypass != bypass) {
356             return NULL;
357         }
358         return domain;
359     }
360     domain = g_malloc0(sizeof(*domain));
361     domain->id = domain_id;
362     domain->mappings = g_tree_new_full((GCompareDataFunc)interval_cmp,
363                                    NULL, (GDestroyNotify)g_free,
364                                    (GDestroyNotify)g_free);
365     domain->bypass = bypass;
366     g_tree_insert(s->domains, GUINT_TO_POINTER(domain_id), domain);
367     QLIST_INIT(&domain->endpoint_list);
368     trace_virtio_iommu_get_domain(domain_id);
369     return domain;
370 }
371 
372 static void virtio_iommu_put_domain(gpointer data)
373 {
374     VirtIOIOMMUDomain *domain = (VirtIOIOMMUDomain *)data;
375     VirtIOIOMMUEndpoint *iter, *tmp;
376 
377     QLIST_FOREACH_SAFE(iter, &domain->endpoint_list, next, tmp) {
378         virtio_iommu_detach_endpoint_from_domain(iter);
379     }
380     g_tree_destroy(domain->mappings);
381     trace_virtio_iommu_put_domain(domain->id);
382     g_free(domain);
383 }
384 
385 static void add_prop_resv_regions(IOMMUDevice *sdev)
386 {
387     VirtIOIOMMU *s = sdev->viommu;
388     int i;
389 
390     for (i = 0; i < s->nr_prop_resv_regions; i++) {
391         ReservedRegion *reg = g_new0(ReservedRegion, 1);
392 
393         *reg = s->prop_resv_regions[i];
394         sdev->resv_regions = resv_region_list_insert(sdev->resv_regions, reg);
395     }
396 }
397 
398 static AddressSpace *virtio_iommu_find_add_as(PCIBus *bus, void *opaque,
399                                               int devfn)
400 {
401     VirtIOIOMMU *s = opaque;
402     IOMMUPciBus *sbus = g_hash_table_lookup(s->as_by_busptr, bus);
403     static uint32_t mr_index;
404     IOMMUDevice *sdev;
405 
406     if (!sbus) {
407         sbus = g_malloc0(sizeof(IOMMUPciBus) +
408                          sizeof(IOMMUDevice *) * PCI_DEVFN_MAX);
409         sbus->bus = bus;
410         g_hash_table_insert(s->as_by_busptr, bus, sbus);
411     }
412 
413     sdev = sbus->pbdev[devfn];
414     if (!sdev) {
415         char *name = g_strdup_printf("%s-%d-%d",
416                                      TYPE_VIRTIO_IOMMU_MEMORY_REGION,
417                                      mr_index++, devfn);
418         sdev = sbus->pbdev[devfn] = g_new0(IOMMUDevice, 1);
419 
420         sdev->viommu = s;
421         sdev->bus = bus;
422         sdev->devfn = devfn;
423 
424         trace_virtio_iommu_init_iommu_mr(name);
425 
426         memory_region_init(&sdev->root, OBJECT(s), name, UINT64_MAX);
427         address_space_init(&sdev->as, &sdev->root, TYPE_VIRTIO_IOMMU);
428         add_prop_resv_regions(sdev);
429 
430         /*
431          * Build the IOMMU disabled container with aliases to the
432          * shared MRs.  Note that aliasing to a shared memory region
433          * could help the memory API to detect same FlatViews so we
434          * can have devices to share the same FlatView when in bypass
435          * mode. (either by not configuring virtio-iommu driver or with
436          * "iommu=pt").  It will greatly reduce the total number of
437          * FlatViews of the system hence VM runs faster.
438          */
439         memory_region_init_alias(&sdev->bypass_mr, OBJECT(s),
440                                  "system", get_system_memory(), 0,
441                                  memory_region_size(get_system_memory()));
442 
443         memory_region_init_iommu(&sdev->iommu_mr, sizeof(sdev->iommu_mr),
444                                  TYPE_VIRTIO_IOMMU_MEMORY_REGION,
445                                  OBJECT(s), name,
446                                  UINT64_MAX);
447 
448         /*
449          * Hook both the containers under the root container, we
450          * switch between iommu & bypass MRs by enable/disable
451          * corresponding sub-containers
452          */
453         memory_region_add_subregion_overlap(&sdev->root, 0,
454                                             MEMORY_REGION(&sdev->iommu_mr),
455                                             0);
456         memory_region_add_subregion_overlap(&sdev->root, 0,
457                                             &sdev->bypass_mr, 0);
458 
459         virtio_iommu_switch_address_space(sdev);
460         g_free(name);
461     }
462     return &sdev->as;
463 }
464 
465 static const PCIIOMMUOps virtio_iommu_ops = {
466     .get_address_space = virtio_iommu_find_add_as,
467 };
468 
469 static int virtio_iommu_attach(VirtIOIOMMU *s,
470                                struct virtio_iommu_req_attach *req)
471 {
472     uint32_t domain_id = le32_to_cpu(req->domain);
473     uint32_t ep_id = le32_to_cpu(req->endpoint);
474     uint32_t flags = le32_to_cpu(req->flags);
475     VirtIOIOMMUDomain *domain;
476     VirtIOIOMMUEndpoint *ep;
477     IOMMUDevice *sdev;
478 
479     trace_virtio_iommu_attach(domain_id, ep_id);
480 
481     if (flags & ~VIRTIO_IOMMU_ATTACH_F_BYPASS) {
482         return VIRTIO_IOMMU_S_INVAL;
483     }
484 
485     ep = virtio_iommu_get_endpoint(s, ep_id);
486     if (!ep) {
487         return VIRTIO_IOMMU_S_NOENT;
488     }
489 
490     if (ep->domain) {
491         VirtIOIOMMUDomain *previous_domain = ep->domain;
492         /*
493          * the device is already attached to a domain,
494          * detach it first
495          */
496         virtio_iommu_detach_endpoint_from_domain(ep);
497         if (QLIST_EMPTY(&previous_domain->endpoint_list)) {
498             g_tree_remove(s->domains, GUINT_TO_POINTER(previous_domain->id));
499         }
500     }
501 
502     domain = virtio_iommu_get_domain(s, domain_id,
503                                      flags & VIRTIO_IOMMU_ATTACH_F_BYPASS);
504     if (!domain) {
505         /* Incompatible bypass flag */
506         return VIRTIO_IOMMU_S_INVAL;
507     }
508     QLIST_INSERT_HEAD(&domain->endpoint_list, ep, next);
509 
510     ep->domain = domain;
511     sdev = container_of(ep->iommu_mr, IOMMUDevice, iommu_mr);
512     virtio_iommu_switch_address_space(sdev);
513 
514     /* Replay domain mappings on the associated memory region */
515     g_tree_foreach(domain->mappings, virtio_iommu_notify_map_cb,
516                    ep->iommu_mr);
517 
518     return VIRTIO_IOMMU_S_OK;
519 }
520 
521 static int virtio_iommu_detach(VirtIOIOMMU *s,
522                                struct virtio_iommu_req_detach *req)
523 {
524     uint32_t domain_id = le32_to_cpu(req->domain);
525     uint32_t ep_id = le32_to_cpu(req->endpoint);
526     VirtIOIOMMUDomain *domain;
527     VirtIOIOMMUEndpoint *ep;
528 
529     trace_virtio_iommu_detach(domain_id, ep_id);
530 
531     ep = g_tree_lookup(s->endpoints, GUINT_TO_POINTER(ep_id));
532     if (!ep) {
533         return VIRTIO_IOMMU_S_NOENT;
534     }
535 
536     domain = ep->domain;
537 
538     if (!domain || domain->id != domain_id) {
539         return VIRTIO_IOMMU_S_INVAL;
540     }
541 
542     virtio_iommu_detach_endpoint_from_domain(ep);
543 
544     if (QLIST_EMPTY(&domain->endpoint_list)) {
545         g_tree_remove(s->domains, GUINT_TO_POINTER(domain->id));
546     }
547     return VIRTIO_IOMMU_S_OK;
548 }
549 
550 static int virtio_iommu_map(VirtIOIOMMU *s,
551                             struct virtio_iommu_req_map *req)
552 {
553     uint32_t domain_id = le32_to_cpu(req->domain);
554     uint64_t phys_start = le64_to_cpu(req->phys_start);
555     uint64_t virt_start = le64_to_cpu(req->virt_start);
556     uint64_t virt_end = le64_to_cpu(req->virt_end);
557     uint32_t flags = le32_to_cpu(req->flags);
558     VirtIOIOMMUDomain *domain;
559     VirtIOIOMMUInterval *interval;
560     VirtIOIOMMUMapping *mapping;
561     VirtIOIOMMUEndpoint *ep;
562 
563     if (flags & ~VIRTIO_IOMMU_MAP_F_MASK) {
564         return VIRTIO_IOMMU_S_INVAL;
565     }
566 
567     domain = g_tree_lookup(s->domains, GUINT_TO_POINTER(domain_id));
568     if (!domain) {
569         return VIRTIO_IOMMU_S_NOENT;
570     }
571 
572     if (domain->bypass) {
573         return VIRTIO_IOMMU_S_INVAL;
574     }
575 
576     interval = g_malloc0(sizeof(*interval));
577 
578     interval->low = virt_start;
579     interval->high = virt_end;
580 
581     mapping = g_tree_lookup(domain->mappings, (gpointer)interval);
582     if (mapping) {
583         g_free(interval);
584         return VIRTIO_IOMMU_S_INVAL;
585     }
586 
587     trace_virtio_iommu_map(domain_id, virt_start, virt_end, phys_start, flags);
588 
589     mapping = g_malloc0(sizeof(*mapping));
590     mapping->phys_addr = phys_start;
591     mapping->flags = flags;
592 
593     g_tree_insert(domain->mappings, interval, mapping);
594 
595     QLIST_FOREACH(ep, &domain->endpoint_list, next) {
596         virtio_iommu_notify_map(ep->iommu_mr, virt_start, virt_end, phys_start,
597                                 flags);
598     }
599 
600     return VIRTIO_IOMMU_S_OK;
601 }
602 
603 static int virtio_iommu_unmap(VirtIOIOMMU *s,
604                               struct virtio_iommu_req_unmap *req)
605 {
606     uint32_t domain_id = le32_to_cpu(req->domain);
607     uint64_t virt_start = le64_to_cpu(req->virt_start);
608     uint64_t virt_end = le64_to_cpu(req->virt_end);
609     VirtIOIOMMUMapping *iter_val;
610     VirtIOIOMMUInterval interval, *iter_key;
611     VirtIOIOMMUDomain *domain;
612     VirtIOIOMMUEndpoint *ep;
613     int ret = VIRTIO_IOMMU_S_OK;
614 
615     trace_virtio_iommu_unmap(domain_id, virt_start, virt_end);
616 
617     domain = g_tree_lookup(s->domains, GUINT_TO_POINTER(domain_id));
618     if (!domain) {
619         return VIRTIO_IOMMU_S_NOENT;
620     }
621 
622     if (domain->bypass) {
623         return VIRTIO_IOMMU_S_INVAL;
624     }
625 
626     interval.low = virt_start;
627     interval.high = virt_end;
628 
629     while (g_tree_lookup_extended(domain->mappings, &interval,
630                                   (void **)&iter_key, (void**)&iter_val)) {
631         uint64_t current_low = iter_key->low;
632         uint64_t current_high = iter_key->high;
633 
634         if (interval.low <= current_low && interval.high >= current_high) {
635             QLIST_FOREACH(ep, &domain->endpoint_list, next) {
636                 virtio_iommu_notify_unmap(ep->iommu_mr, current_low,
637                                           current_high);
638             }
639             g_tree_remove(domain->mappings, iter_key);
640             trace_virtio_iommu_unmap_done(domain_id, current_low, current_high);
641         } else {
642             ret = VIRTIO_IOMMU_S_RANGE;
643             break;
644         }
645     }
646     return ret;
647 }
648 
649 static ssize_t virtio_iommu_fill_resv_mem_prop(IOMMUDevice *sdev, uint32_t ep,
650                                                uint8_t *buf, size_t free)
651 {
652     struct virtio_iommu_probe_resv_mem prop = {};
653     size_t size = sizeof(prop), length = size - sizeof(prop.head), total;
654     GList *l;
655 
656     total = size * g_list_length(sdev->resv_regions);
657     if (total > free) {
658         return -ENOSPC;
659     }
660 
661     for (l = sdev->resv_regions; l; l = l->next) {
662         ReservedRegion *reg = l->data;
663         unsigned subtype = reg->type;
664         Range *range = &reg->range;
665 
666         assert(subtype == VIRTIO_IOMMU_RESV_MEM_T_RESERVED ||
667                subtype == VIRTIO_IOMMU_RESV_MEM_T_MSI);
668         prop.head.type = cpu_to_le16(VIRTIO_IOMMU_PROBE_T_RESV_MEM);
669         prop.head.length = cpu_to_le16(length);
670         prop.subtype = subtype;
671         prop.start = cpu_to_le64(range_lob(range));
672         prop.end = cpu_to_le64(range_upb(range));
673 
674         memcpy(buf, &prop, size);
675 
676         trace_virtio_iommu_fill_resv_property(ep, prop.subtype,
677                                               prop.start, prop.end);
678         buf += size;
679     }
680     return total;
681 }
682 
683 /**
684  * virtio_iommu_probe - Fill the probe request buffer with
685  * the properties the device is able to return
686  */
687 static int virtio_iommu_probe(VirtIOIOMMU *s,
688                               struct virtio_iommu_req_probe *req,
689                               uint8_t *buf)
690 {
691     uint32_t ep_id = le32_to_cpu(req->endpoint);
692     IOMMUMemoryRegion *iommu_mr = virtio_iommu_mr(s, ep_id);
693     size_t free = VIOMMU_PROBE_SIZE;
694     IOMMUDevice *sdev;
695     ssize_t count;
696 
697     if (!iommu_mr) {
698         return VIRTIO_IOMMU_S_NOENT;
699     }
700 
701     sdev = container_of(iommu_mr, IOMMUDevice, iommu_mr);
702 
703     count = virtio_iommu_fill_resv_mem_prop(sdev, ep_id, buf, free);
704     if (count < 0) {
705         return VIRTIO_IOMMU_S_INVAL;
706     }
707     buf += count;
708     free -= count;
709     sdev->probe_done = true;
710 
711     return VIRTIO_IOMMU_S_OK;
712 }
713 
714 static int virtio_iommu_iov_to_req(struct iovec *iov,
715                                    unsigned int iov_cnt,
716                                    void *req, size_t payload_sz)
717 {
718     size_t sz = iov_to_buf(iov, iov_cnt, 0, req, payload_sz);
719 
720     if (unlikely(sz != payload_sz)) {
721         return VIRTIO_IOMMU_S_INVAL;
722     }
723     return 0;
724 }
725 
726 #define virtio_iommu_handle_req(__req)                                  \
727 static int virtio_iommu_handle_ ## __req(VirtIOIOMMU *s,                \
728                                          struct iovec *iov,             \
729                                          unsigned int iov_cnt)          \
730 {                                                                       \
731     struct virtio_iommu_req_ ## __req req;                              \
732     int ret = virtio_iommu_iov_to_req(iov, iov_cnt, &req,               \
733                     sizeof(req) - sizeof(struct virtio_iommu_req_tail));\
734                                                                         \
735     return ret ? ret : virtio_iommu_ ## __req(s, &req);                 \
736 }
737 
738 virtio_iommu_handle_req(attach)
739 virtio_iommu_handle_req(detach)
740 virtio_iommu_handle_req(map)
741 virtio_iommu_handle_req(unmap)
742 
743 static int virtio_iommu_handle_probe(VirtIOIOMMU *s,
744                                      struct iovec *iov,
745                                      unsigned int iov_cnt,
746                                      uint8_t *buf)
747 {
748     struct virtio_iommu_req_probe req;
749     int ret = virtio_iommu_iov_to_req(iov, iov_cnt, &req, sizeof(req));
750 
751     return ret ? ret : virtio_iommu_probe(s, &req, buf);
752 }
753 
754 static void virtio_iommu_handle_command(VirtIODevice *vdev, VirtQueue *vq)
755 {
756     VirtIOIOMMU *s = VIRTIO_IOMMU(vdev);
757     struct virtio_iommu_req_head head;
758     struct virtio_iommu_req_tail tail = {};
759     VirtQueueElement *elem;
760     unsigned int iov_cnt;
761     struct iovec *iov;
762     void *buf = NULL;
763     size_t sz;
764 
765     for (;;) {
766         size_t output_size = sizeof(tail);
767 
768         elem = virtqueue_pop(vq, sizeof(VirtQueueElement));
769         if (!elem) {
770             return;
771         }
772 
773         if (iov_size(elem->in_sg, elem->in_num) < sizeof(tail) ||
774             iov_size(elem->out_sg, elem->out_num) < sizeof(head)) {
775             virtio_error(vdev, "virtio-iommu bad head/tail size");
776             virtqueue_detach_element(vq, elem, 0);
777             g_free(elem);
778             break;
779         }
780 
781         iov_cnt = elem->out_num;
782         iov = elem->out_sg;
783         sz = iov_to_buf(iov, iov_cnt, 0, &head, sizeof(head));
784         if (unlikely(sz != sizeof(head))) {
785             tail.status = VIRTIO_IOMMU_S_DEVERR;
786             goto out;
787         }
788         qemu_rec_mutex_lock(&s->mutex);
789         switch (head.type) {
790         case VIRTIO_IOMMU_T_ATTACH:
791             tail.status = virtio_iommu_handle_attach(s, iov, iov_cnt);
792             break;
793         case VIRTIO_IOMMU_T_DETACH:
794             tail.status = virtio_iommu_handle_detach(s, iov, iov_cnt);
795             break;
796         case VIRTIO_IOMMU_T_MAP:
797             tail.status = virtio_iommu_handle_map(s, iov, iov_cnt);
798             break;
799         case VIRTIO_IOMMU_T_UNMAP:
800             tail.status = virtio_iommu_handle_unmap(s, iov, iov_cnt);
801             break;
802         case VIRTIO_IOMMU_T_PROBE:
803         {
804             struct virtio_iommu_req_tail *ptail;
805 
806             output_size = s->config.probe_size + sizeof(tail);
807             buf = g_malloc0(output_size);
808 
809             ptail = buf + s->config.probe_size;
810             ptail->status = virtio_iommu_handle_probe(s, iov, iov_cnt, buf);
811             break;
812         }
813         default:
814             tail.status = VIRTIO_IOMMU_S_UNSUPP;
815         }
816         qemu_rec_mutex_unlock(&s->mutex);
817 
818 out:
819         sz = iov_from_buf(elem->in_sg, elem->in_num, 0,
820                           buf ? buf : &tail, output_size);
821         assert(sz == output_size);
822 
823         virtqueue_push(vq, elem, sz);
824         virtio_notify(vdev, vq);
825         g_free(elem);
826         g_free(buf);
827         buf = NULL;
828     }
829 }
830 
831 static void virtio_iommu_report_fault(VirtIOIOMMU *viommu, uint8_t reason,
832                                       int flags, uint32_t endpoint,
833                                       uint64_t address)
834 {
835     VirtIODevice *vdev = &viommu->parent_obj;
836     VirtQueue *vq = viommu->event_vq;
837     struct virtio_iommu_fault fault;
838     VirtQueueElement *elem;
839     size_t sz;
840 
841     memset(&fault, 0, sizeof(fault));
842     fault.reason = reason;
843     fault.flags = cpu_to_le32(flags);
844     fault.endpoint = cpu_to_le32(endpoint);
845     fault.address = cpu_to_le64(address);
846 
847     elem = virtqueue_pop(vq, sizeof(VirtQueueElement));
848 
849     if (!elem) {
850         error_report_once(
851             "no buffer available in event queue to report event");
852         return;
853     }
854 
855     if (iov_size(elem->in_sg, elem->in_num) < sizeof(fault)) {
856         virtio_error(vdev, "error buffer of wrong size");
857         virtqueue_detach_element(vq, elem, 0);
858         g_free(elem);
859         return;
860     }
861 
862     sz = iov_from_buf(elem->in_sg, elem->in_num, 0,
863                       &fault, sizeof(fault));
864     assert(sz == sizeof(fault));
865 
866     trace_virtio_iommu_report_fault(reason, flags, endpoint, address);
867     virtqueue_push(vq, elem, sz);
868     virtio_notify(vdev, vq);
869     g_free(elem);
870 
871 }
872 
873 static IOMMUTLBEntry virtio_iommu_translate(IOMMUMemoryRegion *mr, hwaddr addr,
874                                             IOMMUAccessFlags flag,
875                                             int iommu_idx)
876 {
877     IOMMUDevice *sdev = container_of(mr, IOMMUDevice, iommu_mr);
878     VirtIOIOMMUInterval interval, *mapping_key;
879     VirtIOIOMMUMapping *mapping_value;
880     VirtIOIOMMU *s = sdev->viommu;
881     bool read_fault, write_fault;
882     VirtIOIOMMUEndpoint *ep;
883     uint32_t sid, flags;
884     bool bypass_allowed;
885     int granule;
886     bool found;
887     GList *l;
888 
889     interval.low = addr;
890     interval.high = addr + 1;
891     granule = ctz64(s->config.page_size_mask);
892 
893     IOMMUTLBEntry entry = {
894         .target_as = &address_space_memory,
895         .iova = addr,
896         .translated_addr = addr,
897         .addr_mask = BIT_ULL(granule) - 1,
898         .perm = IOMMU_NONE,
899     };
900 
901     bypass_allowed = s->config.bypass;
902 
903     sid = virtio_iommu_get_bdf(sdev);
904 
905     trace_virtio_iommu_translate(mr->parent_obj.name, sid, addr, flag);
906     qemu_rec_mutex_lock(&s->mutex);
907 
908     ep = g_tree_lookup(s->endpoints, GUINT_TO_POINTER(sid));
909 
910     if (bypass_allowed)
911         assert(ep && ep->domain && !ep->domain->bypass);
912 
913     if (!ep) {
914         if (!bypass_allowed) {
915             error_report_once("%s sid=%d is not known!!", __func__, sid);
916             virtio_iommu_report_fault(s, VIRTIO_IOMMU_FAULT_R_UNKNOWN,
917                                       VIRTIO_IOMMU_FAULT_F_ADDRESS,
918                                       sid, addr);
919         } else {
920             entry.perm = flag;
921         }
922         goto unlock;
923     }
924 
925     for (l = sdev->resv_regions; l; l = l->next) {
926         ReservedRegion *reg = l->data;
927 
928         if (range_contains(&reg->range, addr)) {
929             switch (reg->type) {
930             case VIRTIO_IOMMU_RESV_MEM_T_MSI:
931                 entry.perm = flag;
932                 break;
933             case VIRTIO_IOMMU_RESV_MEM_T_RESERVED:
934             default:
935                 virtio_iommu_report_fault(s, VIRTIO_IOMMU_FAULT_R_MAPPING,
936                                           VIRTIO_IOMMU_FAULT_F_ADDRESS,
937                                           sid, addr);
938                 break;
939             }
940             goto unlock;
941         }
942     }
943 
944     if (!ep->domain) {
945         if (!bypass_allowed) {
946             error_report_once("%s %02x:%02x.%01x not attached to any domain",
947                               __func__, PCI_BUS_NUM(sid),
948                               PCI_SLOT(sid), PCI_FUNC(sid));
949             virtio_iommu_report_fault(s, VIRTIO_IOMMU_FAULT_R_DOMAIN,
950                                       VIRTIO_IOMMU_FAULT_F_ADDRESS,
951                                       sid, addr);
952         } else {
953             entry.perm = flag;
954         }
955         goto unlock;
956     } else if (ep->domain->bypass) {
957         entry.perm = flag;
958         goto unlock;
959     }
960 
961     found = g_tree_lookup_extended(ep->domain->mappings, (gpointer)(&interval),
962                                    (void **)&mapping_key,
963                                    (void **)&mapping_value);
964     if (!found) {
965         error_report_once("%s no mapping for 0x%"PRIx64" for sid=%d",
966                           __func__, addr, sid);
967         virtio_iommu_report_fault(s, VIRTIO_IOMMU_FAULT_R_MAPPING,
968                                   VIRTIO_IOMMU_FAULT_F_ADDRESS,
969                                   sid, addr);
970         goto unlock;
971     }
972 
973     read_fault = (flag & IOMMU_RO) &&
974                     !(mapping_value->flags & VIRTIO_IOMMU_MAP_F_READ);
975     write_fault = (flag & IOMMU_WO) &&
976                     !(mapping_value->flags & VIRTIO_IOMMU_MAP_F_WRITE);
977 
978     flags = read_fault ? VIRTIO_IOMMU_FAULT_F_READ : 0;
979     flags |= write_fault ? VIRTIO_IOMMU_FAULT_F_WRITE : 0;
980     if (flags) {
981         error_report_once("%s permission error on 0x%"PRIx64"(%d): allowed=%d",
982                           __func__, addr, flag, mapping_value->flags);
983         flags |= VIRTIO_IOMMU_FAULT_F_ADDRESS;
984         virtio_iommu_report_fault(s, VIRTIO_IOMMU_FAULT_R_MAPPING,
985                                   flags | VIRTIO_IOMMU_FAULT_F_ADDRESS,
986                                   sid, addr);
987         goto unlock;
988     }
989     entry.translated_addr = addr - mapping_key->low + mapping_value->phys_addr;
990     entry.perm = flag;
991     trace_virtio_iommu_translate_out(addr, entry.translated_addr, sid);
992 
993 unlock:
994     qemu_rec_mutex_unlock(&s->mutex);
995     return entry;
996 }
997 
998 static void virtio_iommu_get_config(VirtIODevice *vdev, uint8_t *config_data)
999 {
1000     VirtIOIOMMU *dev = VIRTIO_IOMMU(vdev);
1001     struct virtio_iommu_config *dev_config = &dev->config;
1002     struct virtio_iommu_config *out_config = (void *)config_data;
1003 
1004     out_config->page_size_mask = cpu_to_le64(dev_config->page_size_mask);
1005     out_config->input_range.start = cpu_to_le64(dev_config->input_range.start);
1006     out_config->input_range.end = cpu_to_le64(dev_config->input_range.end);
1007     out_config->domain_range.start = cpu_to_le32(dev_config->domain_range.start);
1008     out_config->domain_range.end = cpu_to_le32(dev_config->domain_range.end);
1009     out_config->probe_size = cpu_to_le32(dev_config->probe_size);
1010     out_config->bypass = dev_config->bypass;
1011 
1012     trace_virtio_iommu_get_config(dev_config->page_size_mask,
1013                                   dev_config->input_range.start,
1014                                   dev_config->input_range.end,
1015                                   dev_config->domain_range.start,
1016                                   dev_config->domain_range.end,
1017                                   dev_config->probe_size,
1018                                   dev_config->bypass);
1019 }
1020 
1021 static void virtio_iommu_set_config(VirtIODevice *vdev,
1022                                     const uint8_t *config_data)
1023 {
1024     VirtIOIOMMU *dev = VIRTIO_IOMMU(vdev);
1025     struct virtio_iommu_config *dev_config = &dev->config;
1026     const struct virtio_iommu_config *in_config = (void *)config_data;
1027 
1028     if (in_config->bypass != dev_config->bypass) {
1029         if (!virtio_vdev_has_feature(vdev, VIRTIO_IOMMU_F_BYPASS_CONFIG)) {
1030             virtio_error(vdev, "cannot set config.bypass");
1031             return;
1032         } else if (in_config->bypass != 0 && in_config->bypass != 1) {
1033             virtio_error(vdev, "invalid config.bypass value '%u'",
1034                          in_config->bypass);
1035             return;
1036         }
1037         dev_config->bypass = in_config->bypass;
1038         virtio_iommu_switch_address_space_all(dev);
1039     }
1040 
1041     trace_virtio_iommu_set_config(in_config->bypass);
1042 }
1043 
1044 static uint64_t virtio_iommu_get_features(VirtIODevice *vdev, uint64_t f,
1045                                           Error **errp)
1046 {
1047     VirtIOIOMMU *dev = VIRTIO_IOMMU(vdev);
1048 
1049     f |= dev->features;
1050     trace_virtio_iommu_get_features(f);
1051     return f;
1052 }
1053 
1054 static gint int_cmp(gconstpointer a, gconstpointer b, gpointer user_data)
1055 {
1056     guint ua = GPOINTER_TO_UINT(a);
1057     guint ub = GPOINTER_TO_UINT(b);
1058     return (ua > ub) - (ua < ub);
1059 }
1060 
1061 static gboolean virtio_iommu_remap(gpointer key, gpointer value, gpointer data)
1062 {
1063     VirtIOIOMMUMapping *mapping = (VirtIOIOMMUMapping *) value;
1064     VirtIOIOMMUInterval *interval = (VirtIOIOMMUInterval *) key;
1065     IOMMUMemoryRegion *mr = (IOMMUMemoryRegion *) data;
1066 
1067     trace_virtio_iommu_remap(mr->parent_obj.name, interval->low, interval->high,
1068                              mapping->phys_addr);
1069     virtio_iommu_notify_map(mr, interval->low, interval->high,
1070                             mapping->phys_addr, mapping->flags);
1071     return false;
1072 }
1073 
1074 static void virtio_iommu_replay(IOMMUMemoryRegion *mr, IOMMUNotifier *n)
1075 {
1076     IOMMUDevice *sdev = container_of(mr, IOMMUDevice, iommu_mr);
1077     VirtIOIOMMU *s = sdev->viommu;
1078     uint32_t sid;
1079     VirtIOIOMMUEndpoint *ep;
1080 
1081     sid = virtio_iommu_get_bdf(sdev);
1082 
1083     qemu_rec_mutex_lock(&s->mutex);
1084 
1085     if (!s->endpoints) {
1086         goto unlock;
1087     }
1088 
1089     ep = g_tree_lookup(s->endpoints, GUINT_TO_POINTER(sid));
1090     if (!ep || !ep->domain) {
1091         goto unlock;
1092     }
1093 
1094     g_tree_foreach(ep->domain->mappings, virtio_iommu_remap, mr);
1095 
1096 unlock:
1097     qemu_rec_mutex_unlock(&s->mutex);
1098 }
1099 
1100 static int virtio_iommu_notify_flag_changed(IOMMUMemoryRegion *iommu_mr,
1101                                             IOMMUNotifierFlag old,
1102                                             IOMMUNotifierFlag new,
1103                                             Error **errp)
1104 {
1105     if (new & IOMMU_NOTIFIER_DEVIOTLB_UNMAP) {
1106         error_setg(errp, "Virtio-iommu does not support dev-iotlb yet");
1107         return -EINVAL;
1108     }
1109 
1110     if (old == IOMMU_NOTIFIER_NONE) {
1111         trace_virtio_iommu_notify_flag_add(iommu_mr->parent_obj.name);
1112     } else if (new == IOMMU_NOTIFIER_NONE) {
1113         trace_virtio_iommu_notify_flag_del(iommu_mr->parent_obj.name);
1114     }
1115     return 0;
1116 }
1117 
1118 /*
1119  * The default mask depends on the "granule" property. For example, with
1120  * 4k granule, it is -(4 * KiB). When an assigned device has page size
1121  * restrictions due to the hardware IOMMU configuration, apply this restriction
1122  * to the mask.
1123  */
1124 static int virtio_iommu_set_page_size_mask(IOMMUMemoryRegion *mr,
1125                                            uint64_t new_mask,
1126                                            Error **errp)
1127 {
1128     IOMMUDevice *sdev = container_of(mr, IOMMUDevice, iommu_mr);
1129     VirtIOIOMMU *s = sdev->viommu;
1130     uint64_t cur_mask = s->config.page_size_mask;
1131 
1132     trace_virtio_iommu_set_page_size_mask(mr->parent_obj.name, cur_mask,
1133                                           new_mask);
1134 
1135     if ((cur_mask & new_mask) == 0) {
1136         error_setg(errp, "virtio-iommu %s reports a page size mask 0x%"PRIx64
1137                    " incompatible with currently supported mask 0x%"PRIx64,
1138                    mr->parent_obj.name, new_mask, cur_mask);
1139         return -1;
1140     }
1141 
1142     /*
1143      * Once the granule is frozen we can't change the mask anymore. If by
1144      * chance the hotplugged device supports the same granule, we can still
1145      * accept it.
1146      */
1147     if (s->granule_frozen) {
1148         int cur_granule = ctz64(cur_mask);
1149 
1150         if (!(BIT_ULL(cur_granule) & new_mask)) {
1151             error_setg(errp, "virtio-iommu %s does not support frozen granule 0x%llx",
1152                        mr->parent_obj.name, BIT_ULL(cur_granule));
1153             return -1;
1154         }
1155         return 0;
1156     }
1157 
1158     s->config.page_size_mask &= new_mask;
1159     return 0;
1160 }
1161 
1162 /**
1163  * rebuild_resv_regions: rebuild resv regions with both the
1164  * info of host resv ranges and property set resv ranges
1165  */
1166 static int rebuild_resv_regions(IOMMUDevice *sdev)
1167 {
1168     GList *l;
1169     int i = 0;
1170 
1171     /* free the existing list and rebuild it from scratch */
1172     g_list_free_full(sdev->resv_regions, g_free);
1173     sdev->resv_regions = NULL;
1174 
1175     /* First add host reserved regions if any, all tagged as RESERVED */
1176     for (l = sdev->host_resv_ranges; l; l = l->next) {
1177         ReservedRegion *reg = g_new0(ReservedRegion, 1);
1178         Range *r = (Range *)l->data;
1179 
1180         reg->type = VIRTIO_IOMMU_RESV_MEM_T_RESERVED;
1181         range_set_bounds(&reg->range, range_lob(r), range_upb(r));
1182         sdev->resv_regions = resv_region_list_insert(sdev->resv_regions, reg);
1183         trace_virtio_iommu_host_resv_regions(sdev->iommu_mr.parent_obj.name, i,
1184                                              range_lob(&reg->range),
1185                                              range_upb(&reg->range));
1186         i++;
1187     }
1188     /*
1189      * then add higher priority reserved regions set by the machine
1190      * through properties
1191      */
1192     add_prop_resv_regions(sdev);
1193     return 0;
1194 }
1195 
1196 /**
1197  * virtio_iommu_set_iova_ranges: Conveys the usable IOVA ranges
1198  *
1199  * The function turns those into reserved ranges. Once some
1200  * reserved ranges have been set, new reserved regions cannot be
1201  * added outside of the original ones.
1202  *
1203  * @mr: IOMMU MR
1204  * @iova_ranges: list of usable IOVA ranges
1205  * @errp: error handle
1206  */
1207 static int virtio_iommu_set_iova_ranges(IOMMUMemoryRegion *mr,
1208                                         GList *iova_ranges,
1209                                         Error **errp)
1210 {
1211     IOMMUDevice *sdev = container_of(mr, IOMMUDevice, iommu_mr);
1212     GList *current_ranges = sdev->host_resv_ranges;
1213     GList *l, *tmp, *new_ranges = NULL;
1214     int ret = -EINVAL;
1215 
1216     /* check that each new resv region is included in an existing one */
1217     if (sdev->host_resv_ranges) {
1218         range_inverse_array(iova_ranges,
1219                             &new_ranges,
1220                             0, UINT64_MAX);
1221 
1222         for (tmp = new_ranges; tmp; tmp = tmp->next) {
1223             Range *newr = (Range *)tmp->data;
1224             bool included = false;
1225 
1226             for (l = current_ranges; l; l = l->next) {
1227                 Range * r = (Range *)l->data;
1228 
1229                 if (range_contains_range(r, newr)) {
1230                     included = true;
1231                     break;
1232                 }
1233             }
1234             if (!included) {
1235                 goto error;
1236             }
1237         }
1238         /* all new reserved ranges are included in existing ones */
1239         ret = 0;
1240         goto out;
1241     }
1242 
1243     if (sdev->probe_done) {
1244         warn_report("%s: Notified about new host reserved regions after probe",
1245                     mr->parent_obj.name);
1246     }
1247 
1248     range_inverse_array(iova_ranges,
1249                         &sdev->host_resv_ranges,
1250                         0, UINT64_MAX);
1251     rebuild_resv_regions(sdev);
1252 
1253     return 0;
1254 error:
1255     error_setg(errp, "IOMMU mr=%s Conflicting host reserved ranges set!",
1256                mr->parent_obj.name);
1257 out:
1258     g_list_free_full(new_ranges, g_free);
1259     return ret;
1260 }
1261 
1262 static void virtio_iommu_system_reset(void *opaque)
1263 {
1264     VirtIOIOMMU *s = opaque;
1265 
1266     trace_virtio_iommu_system_reset();
1267 
1268     memset(s->iommu_pcibus_by_bus_num, 0, sizeof(s->iommu_pcibus_by_bus_num));
1269 
1270     /*
1271      * config.bypass is sticky across device reset, but should be restored on
1272      * system reset
1273      */
1274     s->config.bypass = s->boot_bypass;
1275     virtio_iommu_switch_address_space_all(s);
1276 
1277 }
1278 
1279 static void virtio_iommu_freeze_granule(Notifier *notifier, void *data)
1280 {
1281     VirtIOIOMMU *s = container_of(notifier, VirtIOIOMMU, machine_done);
1282     int granule;
1283 
1284     if (likely(s->config.bypass)) {
1285         /*
1286          * Transient IOMMU MR enable to collect page_size_mask requirements
1287          * through memory_region_iommu_set_page_size_mask() called by
1288          * VFIO region_add() callback
1289          */
1290          s->config.bypass = false;
1291          virtio_iommu_switch_address_space_all(s);
1292          /* restore default */
1293          s->config.bypass = true;
1294          virtio_iommu_switch_address_space_all(s);
1295     }
1296     s->granule_frozen = true;
1297     granule = ctz64(s->config.page_size_mask);
1298     trace_virtio_iommu_freeze_granule(BIT_ULL(granule));
1299 }
1300 
1301 static void virtio_iommu_device_realize(DeviceState *dev, Error **errp)
1302 {
1303     VirtIODevice *vdev = VIRTIO_DEVICE(dev);
1304     VirtIOIOMMU *s = VIRTIO_IOMMU(dev);
1305 
1306     virtio_init(vdev, VIRTIO_ID_IOMMU, sizeof(struct virtio_iommu_config));
1307 
1308     s->req_vq = virtio_add_queue(vdev, VIOMMU_DEFAULT_QUEUE_SIZE,
1309                              virtio_iommu_handle_command);
1310     s->event_vq = virtio_add_queue(vdev, VIOMMU_DEFAULT_QUEUE_SIZE, NULL);
1311 
1312     /*
1313      * config.bypass is needed to get initial address space early, such as
1314      * in vfio realize
1315      */
1316     s->config.bypass = s->boot_bypass;
1317     if (s->aw_bits < 32 || s->aw_bits > 64) {
1318         error_setg(errp, "aw-bits must be within [32,64]");
1319         return;
1320     }
1321     s->config.input_range.end =
1322         s->aw_bits == 64 ? UINT64_MAX : BIT_ULL(s->aw_bits) - 1;
1323 
1324     switch (s->granule_mode) {
1325     case GRANULE_MODE_4K:
1326         s->config.page_size_mask = -(4 * KiB);
1327         break;
1328     case GRANULE_MODE_8K:
1329         s->config.page_size_mask = -(8 * KiB);
1330         break;
1331     case GRANULE_MODE_16K:
1332         s->config.page_size_mask = -(16 * KiB);
1333         break;
1334     case GRANULE_MODE_64K:
1335         s->config.page_size_mask = -(64 * KiB);
1336         break;
1337     case GRANULE_MODE_HOST:
1338         s->config.page_size_mask = qemu_real_host_page_mask();
1339         break;
1340     default:
1341         error_setg(errp, "Unsupported granule mode");
1342     }
1343     s->config.domain_range.end = UINT32_MAX;
1344     s->config.probe_size = VIOMMU_PROBE_SIZE;
1345 
1346     virtio_add_feature(&s->features, VIRTIO_RING_F_EVENT_IDX);
1347     virtio_add_feature(&s->features, VIRTIO_RING_F_INDIRECT_DESC);
1348     virtio_add_feature(&s->features, VIRTIO_F_VERSION_1);
1349     virtio_add_feature(&s->features, VIRTIO_IOMMU_F_INPUT_RANGE);
1350     virtio_add_feature(&s->features, VIRTIO_IOMMU_F_DOMAIN_RANGE);
1351     virtio_add_feature(&s->features, VIRTIO_IOMMU_F_MAP_UNMAP);
1352     virtio_add_feature(&s->features, VIRTIO_IOMMU_F_MMIO);
1353     virtio_add_feature(&s->features, VIRTIO_IOMMU_F_PROBE);
1354     virtio_add_feature(&s->features, VIRTIO_IOMMU_F_BYPASS_CONFIG);
1355 
1356     qemu_rec_mutex_init(&s->mutex);
1357 
1358     s->as_by_busptr = g_hash_table_new_full(NULL, NULL, NULL, g_free);
1359 
1360     if (s->primary_bus) {
1361         pci_setup_iommu(s->primary_bus, &virtio_iommu_ops, s);
1362     } else {
1363         error_setg(errp, "VIRTIO-IOMMU is not attached to any PCI bus!");
1364     }
1365 
1366     s->machine_done.notify = virtio_iommu_freeze_granule;
1367     qemu_add_machine_init_done_notifier(&s->machine_done);
1368 
1369     qemu_register_reset(virtio_iommu_system_reset, s);
1370 }
1371 
1372 static void virtio_iommu_device_unrealize(DeviceState *dev)
1373 {
1374     VirtIODevice *vdev = VIRTIO_DEVICE(dev);
1375     VirtIOIOMMU *s = VIRTIO_IOMMU(dev);
1376 
1377     qemu_unregister_reset(virtio_iommu_system_reset, s);
1378     qemu_remove_machine_init_done_notifier(&s->machine_done);
1379 
1380     g_hash_table_destroy(s->as_by_busptr);
1381     if (s->domains) {
1382         g_tree_destroy(s->domains);
1383     }
1384     if (s->endpoints) {
1385         g_tree_destroy(s->endpoints);
1386     }
1387 
1388     qemu_rec_mutex_destroy(&s->mutex);
1389 
1390     virtio_delete_queue(s->req_vq);
1391     virtio_delete_queue(s->event_vq);
1392     virtio_cleanup(vdev);
1393 }
1394 
1395 static void virtio_iommu_device_reset(VirtIODevice *vdev)
1396 {
1397     VirtIOIOMMU *s = VIRTIO_IOMMU(vdev);
1398 
1399     trace_virtio_iommu_device_reset();
1400 
1401     if (s->domains) {
1402         g_tree_destroy(s->domains);
1403     }
1404     if (s->endpoints) {
1405         g_tree_destroy(s->endpoints);
1406     }
1407     s->domains = g_tree_new_full((GCompareDataFunc)int_cmp,
1408                                  NULL, NULL, virtio_iommu_put_domain);
1409     s->endpoints = g_tree_new_full((GCompareDataFunc)int_cmp,
1410                                    NULL, NULL, virtio_iommu_put_endpoint);
1411 }
1412 
1413 static void virtio_iommu_set_status(VirtIODevice *vdev, uint8_t status)
1414 {
1415     trace_virtio_iommu_device_status(status);
1416 }
1417 
1418 static void virtio_iommu_instance_init(Object *obj)
1419 {
1420 }
1421 
1422 #define VMSTATE_INTERVAL                               \
1423 {                                                      \
1424     .name = "interval",                                \
1425     .version_id = 1,                                   \
1426     .minimum_version_id = 1,                           \
1427     .fields = (const VMStateField[]) {                 \
1428         VMSTATE_UINT64(low, VirtIOIOMMUInterval),      \
1429         VMSTATE_UINT64(high, VirtIOIOMMUInterval),     \
1430         VMSTATE_END_OF_LIST()                          \
1431     }                                                  \
1432 }
1433 
1434 #define VMSTATE_MAPPING                               \
1435 {                                                     \
1436     .name = "mapping",                                \
1437     .version_id = 1,                                  \
1438     .minimum_version_id = 1,                          \
1439     .fields = (const VMStateField[]) {                \
1440         VMSTATE_UINT64(phys_addr, VirtIOIOMMUMapping),\
1441         VMSTATE_UINT32(flags, VirtIOIOMMUMapping),    \
1442         VMSTATE_END_OF_LIST()                         \
1443     },                                                \
1444 }
1445 
1446 static const VMStateDescription vmstate_interval_mapping[2] = {
1447     VMSTATE_MAPPING,   /* value */
1448     VMSTATE_INTERVAL   /* key   */
1449 };
1450 
1451 static int domain_preload(void *opaque)
1452 {
1453     VirtIOIOMMUDomain *domain = opaque;
1454 
1455     domain->mappings = g_tree_new_full((GCompareDataFunc)interval_cmp,
1456                                        NULL, g_free, g_free);
1457     return 0;
1458 }
1459 
1460 static const VMStateDescription vmstate_endpoint = {
1461     .name = "endpoint",
1462     .version_id = 1,
1463     .minimum_version_id = 1,
1464     .fields = (const VMStateField[]) {
1465         VMSTATE_UINT32(id, VirtIOIOMMUEndpoint),
1466         VMSTATE_END_OF_LIST()
1467     }
1468 };
1469 
1470 static const VMStateDescription vmstate_domain = {
1471     .name = "domain",
1472     .version_id = 2,
1473     .minimum_version_id = 2,
1474     .pre_load = domain_preload,
1475     .fields = (const VMStateField[]) {
1476         VMSTATE_UINT32(id, VirtIOIOMMUDomain),
1477         VMSTATE_GTREE_V(mappings, VirtIOIOMMUDomain, 1,
1478                         vmstate_interval_mapping,
1479                         VirtIOIOMMUInterval, VirtIOIOMMUMapping),
1480         VMSTATE_QLIST_V(endpoint_list, VirtIOIOMMUDomain, 1,
1481                         vmstate_endpoint, VirtIOIOMMUEndpoint, next),
1482         VMSTATE_BOOL_V(bypass, VirtIOIOMMUDomain, 2),
1483         VMSTATE_END_OF_LIST()
1484     }
1485 };
1486 
1487 static gboolean reconstruct_endpoints(gpointer key, gpointer value,
1488                                       gpointer data)
1489 {
1490     VirtIOIOMMU *s = (VirtIOIOMMU *)data;
1491     VirtIOIOMMUDomain *d = (VirtIOIOMMUDomain *)value;
1492     VirtIOIOMMUEndpoint *iter;
1493     IOMMUMemoryRegion *mr;
1494 
1495     QLIST_FOREACH(iter, &d->endpoint_list, next) {
1496         mr = virtio_iommu_mr(s, iter->id);
1497         assert(mr);
1498 
1499         iter->domain = d;
1500         iter->iommu_mr = mr;
1501         g_tree_insert(s->endpoints, GUINT_TO_POINTER(iter->id), iter);
1502     }
1503     return false; /* continue the domain traversal */
1504 }
1505 
1506 static int iommu_post_load(void *opaque, int version_id)
1507 {
1508     VirtIOIOMMU *s = opaque;
1509 
1510     g_tree_foreach(s->domains, reconstruct_endpoints, s);
1511 
1512     /*
1513      * Memory regions are dynamically turned on/off depending on
1514      * 'config.bypass' and attached domain type if there is. After
1515      * migration, we need to make sure the memory regions are
1516      * still correct.
1517      */
1518     virtio_iommu_switch_address_space_all(s);
1519     return 0;
1520 }
1521 
1522 static const VMStateDescription vmstate_virtio_iommu_device = {
1523     .name = "virtio-iommu-device",
1524     .minimum_version_id = 2,
1525     .version_id = 2,
1526     .post_load = iommu_post_load,
1527     .fields = (const VMStateField[]) {
1528         VMSTATE_GTREE_DIRECT_KEY_V(domains, VirtIOIOMMU, 2,
1529                                    &vmstate_domain, VirtIOIOMMUDomain),
1530         VMSTATE_UINT8_V(config.bypass, VirtIOIOMMU, 2),
1531         VMSTATE_END_OF_LIST()
1532     },
1533 };
1534 
1535 static const VMStateDescription vmstate_virtio_iommu = {
1536     .name = "virtio-iommu",
1537     .minimum_version_id = 2,
1538     .priority = MIG_PRI_IOMMU,
1539     .version_id = 2,
1540     .fields = (const VMStateField[]) {
1541         VMSTATE_VIRTIO_DEVICE,
1542         VMSTATE_END_OF_LIST()
1543     },
1544 };
1545 
1546 static Property virtio_iommu_properties[] = {
1547     DEFINE_PROP_LINK("primary-bus", VirtIOIOMMU, primary_bus,
1548                      TYPE_PCI_BUS, PCIBus *),
1549     DEFINE_PROP_BOOL("boot-bypass", VirtIOIOMMU, boot_bypass, true),
1550     DEFINE_PROP_GRANULE_MODE("granule", VirtIOIOMMU, granule_mode,
1551                              GRANULE_MODE_HOST),
1552     DEFINE_PROP_UINT8("aw-bits", VirtIOIOMMU, aw_bits, 64),
1553     DEFINE_PROP_END_OF_LIST(),
1554 };
1555 
1556 static void virtio_iommu_class_init(ObjectClass *klass, void *data)
1557 {
1558     DeviceClass *dc = DEVICE_CLASS(klass);
1559     VirtioDeviceClass *vdc = VIRTIO_DEVICE_CLASS(klass);
1560 
1561     device_class_set_props(dc, virtio_iommu_properties);
1562     dc->vmsd = &vmstate_virtio_iommu;
1563 
1564     set_bit(DEVICE_CATEGORY_MISC, dc->categories);
1565     vdc->realize = virtio_iommu_device_realize;
1566     vdc->unrealize = virtio_iommu_device_unrealize;
1567     vdc->reset = virtio_iommu_device_reset;
1568     vdc->get_config = virtio_iommu_get_config;
1569     vdc->set_config = virtio_iommu_set_config;
1570     vdc->get_features = virtio_iommu_get_features;
1571     vdc->set_status = virtio_iommu_set_status;
1572     vdc->vmsd = &vmstate_virtio_iommu_device;
1573 }
1574 
1575 static void virtio_iommu_memory_region_class_init(ObjectClass *klass,
1576                                                   void *data)
1577 {
1578     IOMMUMemoryRegionClass *imrc = IOMMU_MEMORY_REGION_CLASS(klass);
1579 
1580     imrc->translate = virtio_iommu_translate;
1581     imrc->replay = virtio_iommu_replay;
1582     imrc->notify_flag_changed = virtio_iommu_notify_flag_changed;
1583     imrc->iommu_set_page_size_mask = virtio_iommu_set_page_size_mask;
1584     imrc->iommu_set_iova_ranges = virtio_iommu_set_iova_ranges;
1585 }
1586 
1587 static const TypeInfo virtio_iommu_info = {
1588     .name = TYPE_VIRTIO_IOMMU,
1589     .parent = TYPE_VIRTIO_DEVICE,
1590     .instance_size = sizeof(VirtIOIOMMU),
1591     .instance_init = virtio_iommu_instance_init,
1592     .class_init = virtio_iommu_class_init,
1593 };
1594 
1595 static const TypeInfo virtio_iommu_memory_region_info = {
1596     .parent = TYPE_IOMMU_MEMORY_REGION,
1597     .name = TYPE_VIRTIO_IOMMU_MEMORY_REGION,
1598     .class_init = virtio_iommu_memory_region_class_init,
1599 };
1600 
1601 static void virtio_register_types(void)
1602 {
1603     type_register_static(&virtio_iommu_info);
1604     type_register_static(&virtio_iommu_memory_region_info);
1605 }
1606 
1607 type_init(virtio_register_types)
1608