xref: /openbmc/qemu/hw/virtio/virtio-iommu.c (revision 57b3a7d8)
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-common.h"
24 #include "hw/qdev-properties.h"
25 #include "hw/virtio/virtio.h"
26 #include "sysemu/kvm.h"
27 #include "qapi/error.h"
28 #include "qemu/error-report.h"
29 #include "trace.h"
30 
31 #include "standard-headers/linux/virtio_ids.h"
32 
33 #include "hw/virtio/virtio-bus.h"
34 #include "hw/virtio/virtio-access.h"
35 #include "hw/virtio/virtio-iommu.h"
36 #include "hw/pci/pci_bus.h"
37 #include "hw/pci/pci.h"
38 
39 /* Max size */
40 #define VIOMMU_DEFAULT_QUEUE_SIZE 256
41 #define VIOMMU_PROBE_SIZE 512
42 
43 typedef struct VirtIOIOMMUDomain {
44     uint32_t id;
45     GTree *mappings;
46     QLIST_HEAD(, VirtIOIOMMUEndpoint) endpoint_list;
47 } VirtIOIOMMUDomain;
48 
49 typedef struct VirtIOIOMMUEndpoint {
50     uint32_t id;
51     VirtIOIOMMUDomain *domain;
52     IOMMUMemoryRegion *iommu_mr;
53     QLIST_ENTRY(VirtIOIOMMUEndpoint) next;
54 } VirtIOIOMMUEndpoint;
55 
56 typedef struct VirtIOIOMMUInterval {
57     uint64_t low;
58     uint64_t high;
59 } VirtIOIOMMUInterval;
60 
61 typedef struct VirtIOIOMMUMapping {
62     uint64_t phys_addr;
63     uint32_t flags;
64 } VirtIOIOMMUMapping;
65 
66 static inline uint16_t virtio_iommu_get_bdf(IOMMUDevice *dev)
67 {
68     return PCI_BUILD_BDF(pci_bus_num(dev->bus), dev->devfn);
69 }
70 
71 /**
72  * The bus number is used for lookup when SID based operations occur.
73  * In that case we lazily populate the IOMMUPciBus array from the bus hash
74  * table. At the time the IOMMUPciBus is created (iommu_find_add_as), the bus
75  * numbers may not be always initialized yet.
76  */
77 static IOMMUPciBus *iommu_find_iommu_pcibus(VirtIOIOMMU *s, uint8_t bus_num)
78 {
79     IOMMUPciBus *iommu_pci_bus = s->iommu_pcibus_by_bus_num[bus_num];
80 
81     if (!iommu_pci_bus) {
82         GHashTableIter iter;
83 
84         g_hash_table_iter_init(&iter, s->as_by_busptr);
85         while (g_hash_table_iter_next(&iter, NULL, (void **)&iommu_pci_bus)) {
86             if (pci_bus_num(iommu_pci_bus->bus) == bus_num) {
87                 s->iommu_pcibus_by_bus_num[bus_num] = iommu_pci_bus;
88                 return iommu_pci_bus;
89             }
90         }
91         return NULL;
92     }
93     return iommu_pci_bus;
94 }
95 
96 static IOMMUMemoryRegion *virtio_iommu_mr(VirtIOIOMMU *s, uint32_t sid)
97 {
98     uint8_t bus_n, devfn;
99     IOMMUPciBus *iommu_pci_bus;
100     IOMMUDevice *dev;
101 
102     bus_n = PCI_BUS_NUM(sid);
103     iommu_pci_bus = iommu_find_iommu_pcibus(s, bus_n);
104     if (iommu_pci_bus) {
105         devfn = sid & (PCI_DEVFN_MAX - 1);
106         dev = iommu_pci_bus->pbdev[devfn];
107         if (dev) {
108             return &dev->iommu_mr;
109         }
110     }
111     return NULL;
112 }
113 
114 static gint interval_cmp(gconstpointer a, gconstpointer b, gpointer user_data)
115 {
116     VirtIOIOMMUInterval *inta = (VirtIOIOMMUInterval *)a;
117     VirtIOIOMMUInterval *intb = (VirtIOIOMMUInterval *)b;
118 
119     if (inta->high < intb->low) {
120         return -1;
121     } else if (intb->high < inta->low) {
122         return 1;
123     } else {
124         return 0;
125     }
126 }
127 
128 static void virtio_iommu_notify_map(IOMMUMemoryRegion *mr, hwaddr virt_start,
129                                     hwaddr virt_end, hwaddr paddr,
130                                     uint32_t flags)
131 {
132     IOMMUTLBEntry entry;
133     IOMMUAccessFlags perm = IOMMU_ACCESS_FLAG(flags & VIRTIO_IOMMU_MAP_F_READ,
134                                               flags & VIRTIO_IOMMU_MAP_F_WRITE);
135 
136     if (!(mr->iommu_notify_flags & IOMMU_NOTIFIER_MAP) ||
137         (flags & VIRTIO_IOMMU_MAP_F_MMIO) || !perm) {
138         return;
139     }
140 
141     trace_virtio_iommu_notify_map(mr->parent_obj.name, virt_start, virt_end,
142                                   paddr, perm);
143 
144     entry.target_as = &address_space_memory;
145     entry.addr_mask = virt_end - virt_start;
146     entry.iova = virt_start;
147     entry.perm = perm;
148     entry.translated_addr = paddr;
149 
150     memory_region_notify_iommu(mr, 0, entry);
151 }
152 
153 static void virtio_iommu_notify_unmap(IOMMUMemoryRegion *mr, hwaddr virt_start,
154                                       hwaddr virt_end)
155 {
156     IOMMUTLBEntry entry;
157 
158     if (!(mr->iommu_notify_flags & IOMMU_NOTIFIER_UNMAP)) {
159         return;
160     }
161 
162     trace_virtio_iommu_notify_unmap(mr->parent_obj.name, virt_start, virt_end);
163 
164     entry.target_as = &address_space_memory;
165     entry.addr_mask = virt_end - virt_start;
166     entry.iova = virt_start;
167     entry.perm = IOMMU_NONE;
168     entry.translated_addr = 0;
169 
170     memory_region_notify_iommu(mr, 0, entry);
171 }
172 
173 static gboolean virtio_iommu_notify_unmap_cb(gpointer key, gpointer value,
174                                              gpointer data)
175 {
176     VirtIOIOMMUInterval *interval = (VirtIOIOMMUInterval *) key;
177     IOMMUMemoryRegion *mr = (IOMMUMemoryRegion *) data;
178 
179     virtio_iommu_notify_unmap(mr, interval->low, interval->high);
180 
181     return false;
182 }
183 
184 static gboolean virtio_iommu_notify_map_cb(gpointer key, gpointer value,
185                                            gpointer data)
186 {
187     VirtIOIOMMUMapping *mapping = (VirtIOIOMMUMapping *) value;
188     VirtIOIOMMUInterval *interval = (VirtIOIOMMUInterval *) key;
189     IOMMUMemoryRegion *mr = (IOMMUMemoryRegion *) data;
190 
191     virtio_iommu_notify_map(mr, interval->low, interval->high,
192                             mapping->phys_addr, mapping->flags);
193 
194     return false;
195 }
196 
197 static void virtio_iommu_detach_endpoint_from_domain(VirtIOIOMMUEndpoint *ep)
198 {
199     VirtIOIOMMUDomain *domain = ep->domain;
200 
201     if (!ep->domain) {
202         return;
203     }
204     g_tree_foreach(domain->mappings, virtio_iommu_notify_unmap_cb,
205                    ep->iommu_mr);
206     QLIST_REMOVE(ep, next);
207     ep->domain = NULL;
208 }
209 
210 static VirtIOIOMMUEndpoint *virtio_iommu_get_endpoint(VirtIOIOMMU *s,
211                                                       uint32_t ep_id)
212 {
213     VirtIOIOMMUEndpoint *ep;
214     IOMMUMemoryRegion *mr;
215 
216     ep = g_tree_lookup(s->endpoints, GUINT_TO_POINTER(ep_id));
217     if (ep) {
218         return ep;
219     }
220     mr = virtio_iommu_mr(s, ep_id);
221     if (!mr) {
222         return NULL;
223     }
224     ep = g_malloc0(sizeof(*ep));
225     ep->id = ep_id;
226     ep->iommu_mr = mr;
227     trace_virtio_iommu_get_endpoint(ep_id);
228     g_tree_insert(s->endpoints, GUINT_TO_POINTER(ep_id), ep);
229     return ep;
230 }
231 
232 static void virtio_iommu_put_endpoint(gpointer data)
233 {
234     VirtIOIOMMUEndpoint *ep = (VirtIOIOMMUEndpoint *)data;
235 
236     if (ep->domain) {
237         virtio_iommu_detach_endpoint_from_domain(ep);
238     }
239 
240     trace_virtio_iommu_put_endpoint(ep->id);
241     g_free(ep);
242 }
243 
244 static VirtIOIOMMUDomain *virtio_iommu_get_domain(VirtIOIOMMU *s,
245                                                   uint32_t domain_id)
246 {
247     VirtIOIOMMUDomain *domain;
248 
249     domain = g_tree_lookup(s->domains, GUINT_TO_POINTER(domain_id));
250     if (domain) {
251         return domain;
252     }
253     domain = g_malloc0(sizeof(*domain));
254     domain->id = domain_id;
255     domain->mappings = g_tree_new_full((GCompareDataFunc)interval_cmp,
256                                    NULL, (GDestroyNotify)g_free,
257                                    (GDestroyNotify)g_free);
258     g_tree_insert(s->domains, GUINT_TO_POINTER(domain_id), domain);
259     QLIST_INIT(&domain->endpoint_list);
260     trace_virtio_iommu_get_domain(domain_id);
261     return domain;
262 }
263 
264 static void virtio_iommu_put_domain(gpointer data)
265 {
266     VirtIOIOMMUDomain *domain = (VirtIOIOMMUDomain *)data;
267     VirtIOIOMMUEndpoint *iter, *tmp;
268 
269     QLIST_FOREACH_SAFE(iter, &domain->endpoint_list, next, tmp) {
270         virtio_iommu_detach_endpoint_from_domain(iter);
271     }
272     g_tree_destroy(domain->mappings);
273     trace_virtio_iommu_put_domain(domain->id);
274     g_free(domain);
275 }
276 
277 static AddressSpace *virtio_iommu_find_add_as(PCIBus *bus, void *opaque,
278                                               int devfn)
279 {
280     VirtIOIOMMU *s = opaque;
281     IOMMUPciBus *sbus = g_hash_table_lookup(s->as_by_busptr, bus);
282     static uint32_t mr_index;
283     IOMMUDevice *sdev;
284 
285     if (!sbus) {
286         sbus = g_malloc0(sizeof(IOMMUPciBus) +
287                          sizeof(IOMMUDevice *) * PCI_DEVFN_MAX);
288         sbus->bus = bus;
289         g_hash_table_insert(s->as_by_busptr, bus, sbus);
290     }
291 
292     sdev = sbus->pbdev[devfn];
293     if (!sdev) {
294         char *name = g_strdup_printf("%s-%d-%d",
295                                      TYPE_VIRTIO_IOMMU_MEMORY_REGION,
296                                      mr_index++, devfn);
297         sdev = sbus->pbdev[devfn] = g_malloc0(sizeof(IOMMUDevice));
298 
299         sdev->viommu = s;
300         sdev->bus = bus;
301         sdev->devfn = devfn;
302 
303         trace_virtio_iommu_init_iommu_mr(name);
304 
305         memory_region_init_iommu(&sdev->iommu_mr, sizeof(sdev->iommu_mr),
306                                  TYPE_VIRTIO_IOMMU_MEMORY_REGION,
307                                  OBJECT(s), name,
308                                  UINT64_MAX);
309         address_space_init(&sdev->as,
310                            MEMORY_REGION(&sdev->iommu_mr), TYPE_VIRTIO_IOMMU);
311         g_free(name);
312     }
313     return &sdev->as;
314 }
315 
316 static int virtio_iommu_attach(VirtIOIOMMU *s,
317                                struct virtio_iommu_req_attach *req)
318 {
319     uint32_t domain_id = le32_to_cpu(req->domain);
320     uint32_t ep_id = le32_to_cpu(req->endpoint);
321     VirtIOIOMMUDomain *domain;
322     VirtIOIOMMUEndpoint *ep;
323 
324     trace_virtio_iommu_attach(domain_id, ep_id);
325 
326     ep = virtio_iommu_get_endpoint(s, ep_id);
327     if (!ep) {
328         return VIRTIO_IOMMU_S_NOENT;
329     }
330 
331     if (ep->domain) {
332         VirtIOIOMMUDomain *previous_domain = ep->domain;
333         /*
334          * the device is already attached to a domain,
335          * detach it first
336          */
337         virtio_iommu_detach_endpoint_from_domain(ep);
338         if (QLIST_EMPTY(&previous_domain->endpoint_list)) {
339             g_tree_remove(s->domains, GUINT_TO_POINTER(previous_domain->id));
340         }
341     }
342 
343     domain = virtio_iommu_get_domain(s, domain_id);
344     QLIST_INSERT_HEAD(&domain->endpoint_list, ep, next);
345 
346     ep->domain = domain;
347 
348     /* Replay domain mappings on the associated memory region */
349     g_tree_foreach(domain->mappings, virtio_iommu_notify_map_cb,
350                    ep->iommu_mr);
351 
352     return VIRTIO_IOMMU_S_OK;
353 }
354 
355 static int virtio_iommu_detach(VirtIOIOMMU *s,
356                                struct virtio_iommu_req_detach *req)
357 {
358     uint32_t domain_id = le32_to_cpu(req->domain);
359     uint32_t ep_id = le32_to_cpu(req->endpoint);
360     VirtIOIOMMUDomain *domain;
361     VirtIOIOMMUEndpoint *ep;
362 
363     trace_virtio_iommu_detach(domain_id, ep_id);
364 
365     ep = g_tree_lookup(s->endpoints, GUINT_TO_POINTER(ep_id));
366     if (!ep) {
367         return VIRTIO_IOMMU_S_NOENT;
368     }
369 
370     domain = ep->domain;
371 
372     if (!domain || domain->id != domain_id) {
373         return VIRTIO_IOMMU_S_INVAL;
374     }
375 
376     virtio_iommu_detach_endpoint_from_domain(ep);
377 
378     if (QLIST_EMPTY(&domain->endpoint_list)) {
379         g_tree_remove(s->domains, GUINT_TO_POINTER(domain->id));
380     }
381     return VIRTIO_IOMMU_S_OK;
382 }
383 
384 static int virtio_iommu_map(VirtIOIOMMU *s,
385                             struct virtio_iommu_req_map *req)
386 {
387     uint32_t domain_id = le32_to_cpu(req->domain);
388     uint64_t phys_start = le64_to_cpu(req->phys_start);
389     uint64_t virt_start = le64_to_cpu(req->virt_start);
390     uint64_t virt_end = le64_to_cpu(req->virt_end);
391     uint32_t flags = le32_to_cpu(req->flags);
392     VirtIOIOMMUDomain *domain;
393     VirtIOIOMMUInterval *interval;
394     VirtIOIOMMUMapping *mapping;
395     VirtIOIOMMUEndpoint *ep;
396 
397     if (flags & ~VIRTIO_IOMMU_MAP_F_MASK) {
398         return VIRTIO_IOMMU_S_INVAL;
399     }
400 
401     domain = g_tree_lookup(s->domains, GUINT_TO_POINTER(domain_id));
402     if (!domain) {
403         return VIRTIO_IOMMU_S_NOENT;
404     }
405 
406     interval = g_malloc0(sizeof(*interval));
407 
408     interval->low = virt_start;
409     interval->high = virt_end;
410 
411     mapping = g_tree_lookup(domain->mappings, (gpointer)interval);
412     if (mapping) {
413         g_free(interval);
414         return VIRTIO_IOMMU_S_INVAL;
415     }
416 
417     trace_virtio_iommu_map(domain_id, virt_start, virt_end, phys_start, flags);
418 
419     mapping = g_malloc0(sizeof(*mapping));
420     mapping->phys_addr = phys_start;
421     mapping->flags = flags;
422 
423     g_tree_insert(domain->mappings, interval, mapping);
424 
425     QLIST_FOREACH(ep, &domain->endpoint_list, next) {
426         virtio_iommu_notify_map(ep->iommu_mr, virt_start, virt_end, phys_start,
427                                 flags);
428     }
429 
430     return VIRTIO_IOMMU_S_OK;
431 }
432 
433 static int virtio_iommu_unmap(VirtIOIOMMU *s,
434                               struct virtio_iommu_req_unmap *req)
435 {
436     uint32_t domain_id = le32_to_cpu(req->domain);
437     uint64_t virt_start = le64_to_cpu(req->virt_start);
438     uint64_t virt_end = le64_to_cpu(req->virt_end);
439     VirtIOIOMMUMapping *iter_val;
440     VirtIOIOMMUInterval interval, *iter_key;
441     VirtIOIOMMUDomain *domain;
442     VirtIOIOMMUEndpoint *ep;
443     int ret = VIRTIO_IOMMU_S_OK;
444 
445     trace_virtio_iommu_unmap(domain_id, virt_start, virt_end);
446 
447     domain = g_tree_lookup(s->domains, GUINT_TO_POINTER(domain_id));
448     if (!domain) {
449         return VIRTIO_IOMMU_S_NOENT;
450     }
451     interval.low = virt_start;
452     interval.high = virt_end;
453 
454     while (g_tree_lookup_extended(domain->mappings, &interval,
455                                   (void **)&iter_key, (void**)&iter_val)) {
456         uint64_t current_low = iter_key->low;
457         uint64_t current_high = iter_key->high;
458 
459         if (interval.low <= current_low && interval.high >= current_high) {
460             QLIST_FOREACH(ep, &domain->endpoint_list, next) {
461                 virtio_iommu_notify_unmap(ep->iommu_mr, current_low,
462                                           current_high);
463             }
464             g_tree_remove(domain->mappings, iter_key);
465             trace_virtio_iommu_unmap_done(domain_id, current_low, current_high);
466         } else {
467             ret = VIRTIO_IOMMU_S_RANGE;
468             break;
469         }
470     }
471     return ret;
472 }
473 
474 static ssize_t virtio_iommu_fill_resv_mem_prop(VirtIOIOMMU *s, uint32_t ep,
475                                                uint8_t *buf, size_t free)
476 {
477     struct virtio_iommu_probe_resv_mem prop = {};
478     size_t size = sizeof(prop), length = size - sizeof(prop.head), total;
479     int i;
480 
481     total = size * s->nb_reserved_regions;
482 
483     if (total > free) {
484         return -ENOSPC;
485     }
486 
487     for (i = 0; i < s->nb_reserved_regions; i++) {
488         unsigned subtype = s->reserved_regions[i].type;
489 
490         assert(subtype == VIRTIO_IOMMU_RESV_MEM_T_RESERVED ||
491                subtype == VIRTIO_IOMMU_RESV_MEM_T_MSI);
492         prop.head.type = cpu_to_le16(VIRTIO_IOMMU_PROBE_T_RESV_MEM);
493         prop.head.length = cpu_to_le16(length);
494         prop.subtype = subtype;
495         prop.start = cpu_to_le64(s->reserved_regions[i].low);
496         prop.end = cpu_to_le64(s->reserved_regions[i].high);
497 
498         memcpy(buf, &prop, size);
499 
500         trace_virtio_iommu_fill_resv_property(ep, prop.subtype,
501                                               prop.start, prop.end);
502         buf += size;
503     }
504     return total;
505 }
506 
507 /**
508  * virtio_iommu_probe - Fill the probe request buffer with
509  * the properties the device is able to return
510  */
511 static int virtio_iommu_probe(VirtIOIOMMU *s,
512                               struct virtio_iommu_req_probe *req,
513                               uint8_t *buf)
514 {
515     uint32_t ep_id = le32_to_cpu(req->endpoint);
516     size_t free = VIOMMU_PROBE_SIZE;
517     ssize_t count;
518 
519     if (!virtio_iommu_mr(s, ep_id)) {
520         return VIRTIO_IOMMU_S_NOENT;
521     }
522 
523     count = virtio_iommu_fill_resv_mem_prop(s, ep_id, buf, free);
524     if (count < 0) {
525         return VIRTIO_IOMMU_S_INVAL;
526     }
527     buf += count;
528     free -= count;
529 
530     return VIRTIO_IOMMU_S_OK;
531 }
532 
533 static int virtio_iommu_iov_to_req(struct iovec *iov,
534                                    unsigned int iov_cnt,
535                                    void *req, size_t req_sz)
536 {
537     size_t sz, payload_sz = req_sz - sizeof(struct virtio_iommu_req_tail);
538 
539     sz = iov_to_buf(iov, iov_cnt, 0, req, payload_sz);
540     if (unlikely(sz != payload_sz)) {
541         return VIRTIO_IOMMU_S_INVAL;
542     }
543     return 0;
544 }
545 
546 #define virtio_iommu_handle_req(__req)                                  \
547 static int virtio_iommu_handle_ ## __req(VirtIOIOMMU *s,                \
548                                          struct iovec *iov,             \
549                                          unsigned int iov_cnt)          \
550 {                                                                       \
551     struct virtio_iommu_req_ ## __req req;                              \
552     int ret = virtio_iommu_iov_to_req(iov, iov_cnt, &req, sizeof(req)); \
553                                                                         \
554     return ret ? ret : virtio_iommu_ ## __req(s, &req);                 \
555 }
556 
557 virtio_iommu_handle_req(attach)
558 virtio_iommu_handle_req(detach)
559 virtio_iommu_handle_req(map)
560 virtio_iommu_handle_req(unmap)
561 
562 static int virtio_iommu_handle_probe(VirtIOIOMMU *s,
563                                      struct iovec *iov,
564                                      unsigned int iov_cnt,
565                                      uint8_t *buf)
566 {
567     struct virtio_iommu_req_probe req;
568     int ret = virtio_iommu_iov_to_req(iov, iov_cnt, &req, sizeof(req));
569 
570     return ret ? ret : virtio_iommu_probe(s, &req, buf);
571 }
572 
573 static void virtio_iommu_handle_command(VirtIODevice *vdev, VirtQueue *vq)
574 {
575     VirtIOIOMMU *s = VIRTIO_IOMMU(vdev);
576     struct virtio_iommu_req_head head;
577     struct virtio_iommu_req_tail tail = {};
578     size_t output_size = sizeof(tail), sz;
579     VirtQueueElement *elem;
580     unsigned int iov_cnt;
581     struct iovec *iov;
582     void *buf = NULL;
583 
584     for (;;) {
585         elem = virtqueue_pop(vq, sizeof(VirtQueueElement));
586         if (!elem) {
587             return;
588         }
589 
590         if (iov_size(elem->in_sg, elem->in_num) < sizeof(tail) ||
591             iov_size(elem->out_sg, elem->out_num) < sizeof(head)) {
592             virtio_error(vdev, "virtio-iommu bad head/tail size");
593             virtqueue_detach_element(vq, elem, 0);
594             g_free(elem);
595             break;
596         }
597 
598         iov_cnt = elem->out_num;
599         iov = elem->out_sg;
600         sz = iov_to_buf(iov, iov_cnt, 0, &head, sizeof(head));
601         if (unlikely(sz != sizeof(head))) {
602             tail.status = VIRTIO_IOMMU_S_DEVERR;
603             goto out;
604         }
605         qemu_mutex_lock(&s->mutex);
606         switch (head.type) {
607         case VIRTIO_IOMMU_T_ATTACH:
608             tail.status = virtio_iommu_handle_attach(s, iov, iov_cnt);
609             break;
610         case VIRTIO_IOMMU_T_DETACH:
611             tail.status = virtio_iommu_handle_detach(s, iov, iov_cnt);
612             break;
613         case VIRTIO_IOMMU_T_MAP:
614             tail.status = virtio_iommu_handle_map(s, iov, iov_cnt);
615             break;
616         case VIRTIO_IOMMU_T_UNMAP:
617             tail.status = virtio_iommu_handle_unmap(s, iov, iov_cnt);
618             break;
619         case VIRTIO_IOMMU_T_PROBE:
620         {
621             struct virtio_iommu_req_tail *ptail;
622 
623             output_size = s->config.probe_size + sizeof(tail);
624             buf = g_malloc0(output_size);
625 
626             ptail = (struct virtio_iommu_req_tail *)
627                         (buf + s->config.probe_size);
628             ptail->status = virtio_iommu_handle_probe(s, iov, iov_cnt, buf);
629             break;
630         }
631         default:
632             tail.status = VIRTIO_IOMMU_S_UNSUPP;
633         }
634         qemu_mutex_unlock(&s->mutex);
635 
636 out:
637         sz = iov_from_buf(elem->in_sg, elem->in_num, 0,
638                           buf ? buf : &tail, output_size);
639         assert(sz == output_size);
640 
641         virtqueue_push(vq, elem, sz);
642         virtio_notify(vdev, vq);
643         g_free(elem);
644         g_free(buf);
645     }
646 }
647 
648 static void virtio_iommu_report_fault(VirtIOIOMMU *viommu, uint8_t reason,
649                                       int flags, uint32_t endpoint,
650                                       uint64_t address)
651 {
652     VirtIODevice *vdev = &viommu->parent_obj;
653     VirtQueue *vq = viommu->event_vq;
654     struct virtio_iommu_fault fault;
655     VirtQueueElement *elem;
656     size_t sz;
657 
658     memset(&fault, 0, sizeof(fault));
659     fault.reason = reason;
660     fault.flags = cpu_to_le32(flags);
661     fault.endpoint = cpu_to_le32(endpoint);
662     fault.address = cpu_to_le64(address);
663 
664     elem = virtqueue_pop(vq, sizeof(VirtQueueElement));
665 
666     if (!elem) {
667         error_report_once(
668             "no buffer available in event queue to report event");
669         return;
670     }
671 
672     if (iov_size(elem->in_sg, elem->in_num) < sizeof(fault)) {
673         virtio_error(vdev, "error buffer of wrong size");
674         virtqueue_detach_element(vq, elem, 0);
675         g_free(elem);
676         return;
677     }
678 
679     sz = iov_from_buf(elem->in_sg, elem->in_num, 0,
680                       &fault, sizeof(fault));
681     assert(sz == sizeof(fault));
682 
683     trace_virtio_iommu_report_fault(reason, flags, endpoint, address);
684     virtqueue_push(vq, elem, sz);
685     virtio_notify(vdev, vq);
686     g_free(elem);
687 
688 }
689 
690 static IOMMUTLBEntry virtio_iommu_translate(IOMMUMemoryRegion *mr, hwaddr addr,
691                                             IOMMUAccessFlags flag,
692                                             int iommu_idx)
693 {
694     IOMMUDevice *sdev = container_of(mr, IOMMUDevice, iommu_mr);
695     VirtIOIOMMUInterval interval, *mapping_key;
696     VirtIOIOMMUMapping *mapping_value;
697     VirtIOIOMMU *s = sdev->viommu;
698     bool read_fault, write_fault;
699     VirtIOIOMMUEndpoint *ep;
700     uint32_t sid, flags;
701     bool bypass_allowed;
702     bool found;
703     int i;
704 
705     interval.low = addr;
706     interval.high = addr + 1;
707 
708     IOMMUTLBEntry entry = {
709         .target_as = &address_space_memory,
710         .iova = addr,
711         .translated_addr = addr,
712         .addr_mask = (1 << ctz32(s->config.page_size_mask)) - 1,
713         .perm = IOMMU_NONE,
714     };
715 
716     bypass_allowed = virtio_vdev_has_feature(&s->parent_obj,
717                                              VIRTIO_IOMMU_F_BYPASS);
718 
719     sid = virtio_iommu_get_bdf(sdev);
720 
721     trace_virtio_iommu_translate(mr->parent_obj.name, sid, addr, flag);
722     qemu_mutex_lock(&s->mutex);
723 
724     ep = g_tree_lookup(s->endpoints, GUINT_TO_POINTER(sid));
725     if (!ep) {
726         if (!bypass_allowed) {
727             error_report_once("%s sid=%d is not known!!", __func__, sid);
728             virtio_iommu_report_fault(s, VIRTIO_IOMMU_FAULT_R_UNKNOWN,
729                                       VIRTIO_IOMMU_FAULT_F_ADDRESS,
730                                       sid, addr);
731         } else {
732             entry.perm = flag;
733         }
734         goto unlock;
735     }
736 
737     for (i = 0; i < s->nb_reserved_regions; i++) {
738         ReservedRegion *reg = &s->reserved_regions[i];
739 
740         if (addr >= reg->low && addr <= reg->high) {
741             switch (reg->type) {
742             case VIRTIO_IOMMU_RESV_MEM_T_MSI:
743                 entry.perm = flag;
744                 break;
745             case VIRTIO_IOMMU_RESV_MEM_T_RESERVED:
746             default:
747                 virtio_iommu_report_fault(s, VIRTIO_IOMMU_FAULT_R_MAPPING,
748                                           VIRTIO_IOMMU_FAULT_F_ADDRESS,
749                                           sid, addr);
750                 break;
751             }
752             goto unlock;
753         }
754     }
755 
756     if (!ep->domain) {
757         if (!bypass_allowed) {
758             error_report_once("%s %02x:%02x.%01x not attached to any domain",
759                               __func__, PCI_BUS_NUM(sid),
760                               PCI_SLOT(sid), PCI_FUNC(sid));
761             virtio_iommu_report_fault(s, VIRTIO_IOMMU_FAULT_R_DOMAIN,
762                                       VIRTIO_IOMMU_FAULT_F_ADDRESS,
763                                       sid, addr);
764         } else {
765             entry.perm = flag;
766         }
767         goto unlock;
768     }
769 
770     found = g_tree_lookup_extended(ep->domain->mappings, (gpointer)(&interval),
771                                    (void **)&mapping_key,
772                                    (void **)&mapping_value);
773     if (!found) {
774         error_report_once("%s no mapping for 0x%"PRIx64" for sid=%d",
775                           __func__, addr, sid);
776         virtio_iommu_report_fault(s, VIRTIO_IOMMU_FAULT_R_MAPPING,
777                                   VIRTIO_IOMMU_FAULT_F_ADDRESS,
778                                   sid, addr);
779         goto unlock;
780     }
781 
782     read_fault = (flag & IOMMU_RO) &&
783                     !(mapping_value->flags & VIRTIO_IOMMU_MAP_F_READ);
784     write_fault = (flag & IOMMU_WO) &&
785                     !(mapping_value->flags & VIRTIO_IOMMU_MAP_F_WRITE);
786 
787     flags = read_fault ? VIRTIO_IOMMU_FAULT_F_READ : 0;
788     flags |= write_fault ? VIRTIO_IOMMU_FAULT_F_WRITE : 0;
789     if (flags) {
790         error_report_once("%s permission error on 0x%"PRIx64"(%d): allowed=%d",
791                           __func__, addr, flag, mapping_value->flags);
792         flags |= VIRTIO_IOMMU_FAULT_F_ADDRESS;
793         virtio_iommu_report_fault(s, VIRTIO_IOMMU_FAULT_R_MAPPING,
794                                   flags | VIRTIO_IOMMU_FAULT_F_ADDRESS,
795                                   sid, addr);
796         goto unlock;
797     }
798     entry.translated_addr = addr - mapping_key->low + mapping_value->phys_addr;
799     entry.perm = flag;
800     trace_virtio_iommu_translate_out(addr, entry.translated_addr, sid);
801 
802 unlock:
803     qemu_mutex_unlock(&s->mutex);
804     return entry;
805 }
806 
807 static void virtio_iommu_get_config(VirtIODevice *vdev, uint8_t *config_data)
808 {
809     VirtIOIOMMU *dev = VIRTIO_IOMMU(vdev);
810     struct virtio_iommu_config *config = &dev->config;
811 
812     trace_virtio_iommu_get_config(config->page_size_mask,
813                                   config->input_range.start,
814                                   config->input_range.end,
815                                   config->domain_range.end,
816                                   config->probe_size);
817     memcpy(config_data, &dev->config, sizeof(struct virtio_iommu_config));
818 }
819 
820 static void virtio_iommu_set_config(VirtIODevice *vdev,
821                                       const uint8_t *config_data)
822 {
823     struct virtio_iommu_config config;
824 
825     memcpy(&config, config_data, sizeof(struct virtio_iommu_config));
826     trace_virtio_iommu_set_config(config.page_size_mask,
827                                   config.input_range.start,
828                                   config.input_range.end,
829                                   config.domain_range.end,
830                                   config.probe_size);
831 }
832 
833 static uint64_t virtio_iommu_get_features(VirtIODevice *vdev, uint64_t f,
834                                           Error **errp)
835 {
836     VirtIOIOMMU *dev = VIRTIO_IOMMU(vdev);
837 
838     f |= dev->features;
839     trace_virtio_iommu_get_features(f);
840     return f;
841 }
842 
843 static gint int_cmp(gconstpointer a, gconstpointer b, gpointer user_data)
844 {
845     guint ua = GPOINTER_TO_UINT(a);
846     guint ub = GPOINTER_TO_UINT(b);
847     return (ua > ub) - (ua < ub);
848 }
849 
850 static gboolean virtio_iommu_remap(gpointer key, gpointer value, gpointer data)
851 {
852     VirtIOIOMMUMapping *mapping = (VirtIOIOMMUMapping *) value;
853     VirtIOIOMMUInterval *interval = (VirtIOIOMMUInterval *) key;
854     IOMMUMemoryRegion *mr = (IOMMUMemoryRegion *) data;
855 
856     trace_virtio_iommu_remap(mr->parent_obj.name, interval->low, interval->high,
857                              mapping->phys_addr);
858     virtio_iommu_notify_map(mr, interval->low, interval->high,
859                             mapping->phys_addr, mapping->flags);
860     return false;
861 }
862 
863 static void virtio_iommu_replay(IOMMUMemoryRegion *mr, IOMMUNotifier *n)
864 {
865     IOMMUDevice *sdev = container_of(mr, IOMMUDevice, iommu_mr);
866     VirtIOIOMMU *s = sdev->viommu;
867     uint32_t sid;
868     VirtIOIOMMUEndpoint *ep;
869 
870     sid = virtio_iommu_get_bdf(sdev);
871 
872     qemu_mutex_lock(&s->mutex);
873 
874     if (!s->endpoints) {
875         goto unlock;
876     }
877 
878     ep = g_tree_lookup(s->endpoints, GUINT_TO_POINTER(sid));
879     if (!ep || !ep->domain) {
880         goto unlock;
881     }
882 
883     g_tree_foreach(ep->domain->mappings, virtio_iommu_remap, mr);
884 
885 unlock:
886     qemu_mutex_unlock(&s->mutex);
887 }
888 
889 static int virtio_iommu_notify_flag_changed(IOMMUMemoryRegion *iommu_mr,
890                                             IOMMUNotifierFlag old,
891                                             IOMMUNotifierFlag new,
892                                             Error **errp)
893 {
894     if (old == IOMMU_NOTIFIER_NONE) {
895         trace_virtio_iommu_notify_flag_add(iommu_mr->parent_obj.name);
896     } else if (new == IOMMU_NOTIFIER_NONE) {
897         trace_virtio_iommu_notify_flag_del(iommu_mr->parent_obj.name);
898     }
899     return 0;
900 }
901 
902 /*
903  * The default mask (TARGET_PAGE_MASK) is the smallest supported guest granule,
904  * for example 0xfffffffffffff000. When an assigned device has page size
905  * restrictions due to the hardware IOMMU configuration, apply this restriction
906  * to the mask.
907  */
908 static int virtio_iommu_set_page_size_mask(IOMMUMemoryRegion *mr,
909                                            uint64_t new_mask,
910                                            Error **errp)
911 {
912     IOMMUDevice *sdev = container_of(mr, IOMMUDevice, iommu_mr);
913     VirtIOIOMMU *s = sdev->viommu;
914     uint64_t cur_mask = s->config.page_size_mask;
915 
916     trace_virtio_iommu_set_page_size_mask(mr->parent_obj.name, cur_mask,
917                                           new_mask);
918 
919     if ((cur_mask & new_mask) == 0) {
920         error_setg(errp, "virtio-iommu page mask 0x%"PRIx64
921                    " is incompatible with mask 0x%"PRIx64, cur_mask, new_mask);
922         return -1;
923     }
924 
925     /*
926      * After the machine is finalized, we can't change the mask anymore. If by
927      * chance the hotplugged device supports the same granule, we can still
928      * accept it. Having a different masks is possible but the guest will use
929      * sub-optimal block sizes, so warn about it.
930      */
931     if (qdev_hotplug) {
932         int new_granule = ctz64(new_mask);
933         int cur_granule = ctz64(cur_mask);
934 
935         if (new_granule != cur_granule) {
936             error_setg(errp, "virtio-iommu page mask 0x%"PRIx64
937                        " is incompatible with mask 0x%"PRIx64, cur_mask,
938                        new_mask);
939             return -1;
940         } else if (new_mask != cur_mask) {
941             warn_report("virtio-iommu page mask 0x%"PRIx64
942                         " does not match 0x%"PRIx64, cur_mask, new_mask);
943         }
944         return 0;
945     }
946 
947     s->config.page_size_mask &= new_mask;
948     return 0;
949 }
950 
951 static void virtio_iommu_device_realize(DeviceState *dev, Error **errp)
952 {
953     VirtIODevice *vdev = VIRTIO_DEVICE(dev);
954     VirtIOIOMMU *s = VIRTIO_IOMMU(dev);
955 
956     virtio_init(vdev, "virtio-iommu", VIRTIO_ID_IOMMU,
957                 sizeof(struct virtio_iommu_config));
958 
959     memset(s->iommu_pcibus_by_bus_num, 0, sizeof(s->iommu_pcibus_by_bus_num));
960 
961     s->req_vq = virtio_add_queue(vdev, VIOMMU_DEFAULT_QUEUE_SIZE,
962                              virtio_iommu_handle_command);
963     s->event_vq = virtio_add_queue(vdev, VIOMMU_DEFAULT_QUEUE_SIZE, NULL);
964 
965     s->config.page_size_mask = TARGET_PAGE_MASK;
966     s->config.input_range.end = -1UL;
967     s->config.domain_range.end = 32;
968     s->config.probe_size = VIOMMU_PROBE_SIZE;
969 
970     virtio_add_feature(&s->features, VIRTIO_RING_F_EVENT_IDX);
971     virtio_add_feature(&s->features, VIRTIO_RING_F_INDIRECT_DESC);
972     virtio_add_feature(&s->features, VIRTIO_F_VERSION_1);
973     virtio_add_feature(&s->features, VIRTIO_IOMMU_F_INPUT_RANGE);
974     virtio_add_feature(&s->features, VIRTIO_IOMMU_F_DOMAIN_RANGE);
975     virtio_add_feature(&s->features, VIRTIO_IOMMU_F_MAP_UNMAP);
976     virtio_add_feature(&s->features, VIRTIO_IOMMU_F_BYPASS);
977     virtio_add_feature(&s->features, VIRTIO_IOMMU_F_MMIO);
978     virtio_add_feature(&s->features, VIRTIO_IOMMU_F_PROBE);
979 
980     qemu_mutex_init(&s->mutex);
981 
982     s->as_by_busptr = g_hash_table_new_full(NULL, NULL, NULL, g_free);
983 
984     if (s->primary_bus) {
985         pci_setup_iommu(s->primary_bus, virtio_iommu_find_add_as, s);
986     } else {
987         error_setg(errp, "VIRTIO-IOMMU is not attached to any PCI bus!");
988     }
989 }
990 
991 static void virtio_iommu_device_unrealize(DeviceState *dev)
992 {
993     VirtIODevice *vdev = VIRTIO_DEVICE(dev);
994     VirtIOIOMMU *s = VIRTIO_IOMMU(dev);
995 
996     g_hash_table_destroy(s->as_by_busptr);
997     if (s->domains) {
998         g_tree_destroy(s->domains);
999     }
1000     if (s->endpoints) {
1001         g_tree_destroy(s->endpoints);
1002     }
1003 
1004     virtio_delete_queue(s->req_vq);
1005     virtio_delete_queue(s->event_vq);
1006     virtio_cleanup(vdev);
1007 }
1008 
1009 static void virtio_iommu_device_reset(VirtIODevice *vdev)
1010 {
1011     VirtIOIOMMU *s = VIRTIO_IOMMU(vdev);
1012 
1013     trace_virtio_iommu_device_reset();
1014 
1015     if (s->domains) {
1016         g_tree_destroy(s->domains);
1017     }
1018     if (s->endpoints) {
1019         g_tree_destroy(s->endpoints);
1020     }
1021     s->domains = g_tree_new_full((GCompareDataFunc)int_cmp,
1022                                  NULL, NULL, virtio_iommu_put_domain);
1023     s->endpoints = g_tree_new_full((GCompareDataFunc)int_cmp,
1024                                    NULL, NULL, virtio_iommu_put_endpoint);
1025 }
1026 
1027 static void virtio_iommu_set_status(VirtIODevice *vdev, uint8_t status)
1028 {
1029     trace_virtio_iommu_device_status(status);
1030 }
1031 
1032 static void virtio_iommu_instance_init(Object *obj)
1033 {
1034 }
1035 
1036 #define VMSTATE_INTERVAL                               \
1037 {                                                      \
1038     .name = "interval",                                \
1039     .version_id = 1,                                   \
1040     .minimum_version_id = 1,                           \
1041     .fields = (VMStateField[]) {                       \
1042         VMSTATE_UINT64(low, VirtIOIOMMUInterval),      \
1043         VMSTATE_UINT64(high, VirtIOIOMMUInterval),     \
1044         VMSTATE_END_OF_LIST()                          \
1045     }                                                  \
1046 }
1047 
1048 #define VMSTATE_MAPPING                               \
1049 {                                                     \
1050     .name = "mapping",                                \
1051     .version_id = 1,                                  \
1052     .minimum_version_id = 1,                          \
1053     .fields = (VMStateField[]) {                      \
1054         VMSTATE_UINT64(phys_addr, VirtIOIOMMUMapping),\
1055         VMSTATE_UINT32(flags, VirtIOIOMMUMapping),    \
1056         VMSTATE_END_OF_LIST()                         \
1057     },                                                \
1058 }
1059 
1060 static const VMStateDescription vmstate_interval_mapping[2] = {
1061     VMSTATE_MAPPING,   /* value */
1062     VMSTATE_INTERVAL   /* key   */
1063 };
1064 
1065 static int domain_preload(void *opaque)
1066 {
1067     VirtIOIOMMUDomain *domain = opaque;
1068 
1069     domain->mappings = g_tree_new_full((GCompareDataFunc)interval_cmp,
1070                                        NULL, g_free, g_free);
1071     return 0;
1072 }
1073 
1074 static const VMStateDescription vmstate_endpoint = {
1075     .name = "endpoint",
1076     .version_id = 1,
1077     .minimum_version_id = 1,
1078     .fields = (VMStateField[]) {
1079         VMSTATE_UINT32(id, VirtIOIOMMUEndpoint),
1080         VMSTATE_END_OF_LIST()
1081     }
1082 };
1083 
1084 static const VMStateDescription vmstate_domain = {
1085     .name = "domain",
1086     .version_id = 1,
1087     .minimum_version_id = 1,
1088     .pre_load = domain_preload,
1089     .fields = (VMStateField[]) {
1090         VMSTATE_UINT32(id, VirtIOIOMMUDomain),
1091         VMSTATE_GTREE_V(mappings, VirtIOIOMMUDomain, 1,
1092                         vmstate_interval_mapping,
1093                         VirtIOIOMMUInterval, VirtIOIOMMUMapping),
1094         VMSTATE_QLIST_V(endpoint_list, VirtIOIOMMUDomain, 1,
1095                         vmstate_endpoint, VirtIOIOMMUEndpoint, next),
1096         VMSTATE_END_OF_LIST()
1097     }
1098 };
1099 
1100 static gboolean reconstruct_endpoints(gpointer key, gpointer value,
1101                                       gpointer data)
1102 {
1103     VirtIOIOMMU *s = (VirtIOIOMMU *)data;
1104     VirtIOIOMMUDomain *d = (VirtIOIOMMUDomain *)value;
1105     VirtIOIOMMUEndpoint *iter;
1106     IOMMUMemoryRegion *mr;
1107 
1108     QLIST_FOREACH(iter, &d->endpoint_list, next) {
1109         mr = virtio_iommu_mr(s, iter->id);
1110         assert(mr);
1111 
1112         iter->domain = d;
1113         iter->iommu_mr = mr;
1114         g_tree_insert(s->endpoints, GUINT_TO_POINTER(iter->id), iter);
1115     }
1116     return false; /* continue the domain traversal */
1117 }
1118 
1119 static int iommu_post_load(void *opaque, int version_id)
1120 {
1121     VirtIOIOMMU *s = opaque;
1122 
1123     g_tree_foreach(s->domains, reconstruct_endpoints, s);
1124     return 0;
1125 }
1126 
1127 static const VMStateDescription vmstate_virtio_iommu_device = {
1128     .name = "virtio-iommu-device",
1129     .minimum_version_id = 1,
1130     .version_id = 1,
1131     .post_load = iommu_post_load,
1132     .fields = (VMStateField[]) {
1133         VMSTATE_GTREE_DIRECT_KEY_V(domains, VirtIOIOMMU, 1,
1134                                    &vmstate_domain, VirtIOIOMMUDomain),
1135         VMSTATE_END_OF_LIST()
1136     },
1137 };
1138 
1139 static const VMStateDescription vmstate_virtio_iommu = {
1140     .name = "virtio-iommu",
1141     .minimum_version_id = 1,
1142     .priority = MIG_PRI_IOMMU,
1143     .version_id = 1,
1144     .fields = (VMStateField[]) {
1145         VMSTATE_VIRTIO_DEVICE,
1146         VMSTATE_END_OF_LIST()
1147     },
1148 };
1149 
1150 static Property virtio_iommu_properties[] = {
1151     DEFINE_PROP_LINK("primary-bus", VirtIOIOMMU, primary_bus, "PCI", PCIBus *),
1152     DEFINE_PROP_END_OF_LIST(),
1153 };
1154 
1155 static void virtio_iommu_class_init(ObjectClass *klass, void *data)
1156 {
1157     DeviceClass *dc = DEVICE_CLASS(klass);
1158     VirtioDeviceClass *vdc = VIRTIO_DEVICE_CLASS(klass);
1159 
1160     device_class_set_props(dc, virtio_iommu_properties);
1161     dc->vmsd = &vmstate_virtio_iommu;
1162 
1163     set_bit(DEVICE_CATEGORY_MISC, dc->categories);
1164     vdc->realize = virtio_iommu_device_realize;
1165     vdc->unrealize = virtio_iommu_device_unrealize;
1166     vdc->reset = virtio_iommu_device_reset;
1167     vdc->get_config = virtio_iommu_get_config;
1168     vdc->set_config = virtio_iommu_set_config;
1169     vdc->get_features = virtio_iommu_get_features;
1170     vdc->set_status = virtio_iommu_set_status;
1171     vdc->vmsd = &vmstate_virtio_iommu_device;
1172 }
1173 
1174 static void virtio_iommu_memory_region_class_init(ObjectClass *klass,
1175                                                   void *data)
1176 {
1177     IOMMUMemoryRegionClass *imrc = IOMMU_MEMORY_REGION_CLASS(klass);
1178 
1179     imrc->translate = virtio_iommu_translate;
1180     imrc->replay = virtio_iommu_replay;
1181     imrc->notify_flag_changed = virtio_iommu_notify_flag_changed;
1182     imrc->iommu_set_page_size_mask = virtio_iommu_set_page_size_mask;
1183 }
1184 
1185 static const TypeInfo virtio_iommu_info = {
1186     .name = TYPE_VIRTIO_IOMMU,
1187     .parent = TYPE_VIRTIO_DEVICE,
1188     .instance_size = sizeof(VirtIOIOMMU),
1189     .instance_init = virtio_iommu_instance_init,
1190     .class_init = virtio_iommu_class_init,
1191 };
1192 
1193 static const TypeInfo virtio_iommu_memory_region_info = {
1194     .parent = TYPE_IOMMU_MEMORY_REGION,
1195     .name = TYPE_VIRTIO_IOMMU_MEMORY_REGION,
1196     .class_init = virtio_iommu_memory_region_class_init,
1197 };
1198 
1199 static void virtio_register_types(void)
1200 {
1201     type_register_static(&virtio_iommu_info);
1202     type_register_static(&virtio_iommu_memory_region_info);
1203 }
1204 
1205 type_init(virtio_register_types)
1206