xref: /qemu/hw/virtio/vhost-vdpa.c (revision 74e76c7d5bf074b3dbb5ec71109d6ba80f21e0a9)
1 /*
2  * vhost-vdpa
3  *
4  *  Copyright(c) 2017-2018 Intel Corporation.
5  *  Copyright(c) 2020 Red Hat, Inc.
6  *
7  * This work is licensed under the terms of the GNU GPL, version 2 or later.
8  * See the COPYING file in the top-level directory.
9  *
10  */
11 
12 #include "qemu/osdep.h"
13 #include <linux/vhost.h>
14 #include <linux/vfio.h>
15 #include <sys/eventfd.h>
16 #include <sys/ioctl.h>
17 #include "exec/target_page.h"
18 #include "hw/virtio/vhost.h"
19 #include "hw/virtio/vhost-backend.h"
20 #include "hw/virtio/virtio-net.h"
21 #include "hw/virtio/vhost-shadow-virtqueue.h"
22 #include "hw/virtio/vhost-vdpa.h"
23 #include "exec/address-spaces.h"
24 #include "migration/blocker.h"
25 #include "qemu/cutils.h"
26 #include "qemu/main-loop.h"
27 #include "trace.h"
28 #include "qapi/error.h"
29 
30 /*
31  * Return one past the end of the end of section. Be careful with uint64_t
32  * conversions!
33  */
34 static Int128 vhost_vdpa_section_end(const MemoryRegionSection *section,
35                                      int page_mask)
36 {
37     Int128 llend = int128_make64(section->offset_within_address_space);
38     llend = int128_add(llend, section->size);
39     llend = int128_and(llend, int128_exts64(page_mask));
40 
41     return llend;
42 }
43 
44 static bool vhost_vdpa_listener_skipped_section(MemoryRegionSection *section,
45                                                 uint64_t iova_min,
46                                                 uint64_t iova_max,
47                                                 int page_mask)
48 {
49     Int128 llend;
50 
51     if ((!memory_region_is_ram(section->mr) &&
52          !memory_region_is_iommu(section->mr)) ||
53         memory_region_is_protected(section->mr) ||
54         /* vhost-vDPA doesn't allow MMIO to be mapped  */
55         memory_region_is_ram_device(section->mr)) {
56         return true;
57     }
58 
59     if (section->offset_within_address_space < iova_min) {
60         error_report("RAM section out of device range (min=0x%" PRIx64
61                      ", addr=0x%" HWADDR_PRIx ")",
62                      iova_min, section->offset_within_address_space);
63         return true;
64     }
65     /*
66      * While using vIOMMU, sometimes the section will be larger than iova_max,
67      * but the memory that actually maps is smaller, so move the check to
68      * function vhost_vdpa_iommu_map_notify(). That function will use the actual
69      * size that maps to the kernel
70      */
71 
72     if (!memory_region_is_iommu(section->mr)) {
73         llend = vhost_vdpa_section_end(section, page_mask);
74         if (int128_gt(llend, int128_make64(iova_max))) {
75             error_report("RAM section out of device range (max=0x%" PRIx64
76                          ", end addr=0x%" PRIx64 ")",
77                          iova_max, int128_get64(llend));
78             return true;
79         }
80     }
81 
82     return false;
83 }
84 
85 /*
86  * The caller must set asid = 0 if the device does not support asid.
87  * This is not an ABI break since it is set to 0 by the initializer anyway.
88  */
89 int vhost_vdpa_dma_map(struct vhost_vdpa *v, uint32_t asid, hwaddr iova,
90                        hwaddr size, void *vaddr, bool readonly)
91 {
92     struct vhost_msg_v2 msg = {};
93     int fd = v->shared->device_fd;
94     int ret = 0;
95 
96     msg.type = VHOST_IOTLB_MSG_V2;
97     msg.asid = asid;
98     msg.iotlb.iova = iova;
99     msg.iotlb.size = size;
100     msg.iotlb.uaddr = (uint64_t)(uintptr_t)vaddr;
101     msg.iotlb.perm = readonly ? VHOST_ACCESS_RO : VHOST_ACCESS_RW;
102     msg.iotlb.type = VHOST_IOTLB_UPDATE;
103 
104     trace_vhost_vdpa_dma_map(v->shared, fd, msg.type, msg.asid, msg.iotlb.iova,
105                              msg.iotlb.size, msg.iotlb.uaddr, msg.iotlb.perm,
106                              msg.iotlb.type);
107 
108     if (write(fd, &msg, sizeof(msg)) != sizeof(msg)) {
109         error_report("failed to write, fd=%d, errno=%d (%s)",
110             fd, errno, strerror(errno));
111         return -EIO ;
112     }
113 
114     return ret;
115 }
116 
117 /*
118  * The caller must set asid = 0 if the device does not support asid.
119  * This is not an ABI break since it is set to 0 by the initializer anyway.
120  */
121 int vhost_vdpa_dma_unmap(struct vhost_vdpa *v, uint32_t asid, hwaddr iova,
122                          hwaddr size)
123 {
124     struct vhost_msg_v2 msg = {};
125     int fd = v->shared->device_fd;
126     int ret = 0;
127 
128     msg.type = VHOST_IOTLB_MSG_V2;
129     msg.asid = asid;
130     msg.iotlb.iova = iova;
131     msg.iotlb.size = size;
132     msg.iotlb.type = VHOST_IOTLB_INVALIDATE;
133 
134     trace_vhost_vdpa_dma_unmap(v->shared, fd, msg.type, msg.asid,
135                                msg.iotlb.iova, msg.iotlb.size, msg.iotlb.type);
136 
137     if (write(fd, &msg, sizeof(msg)) != sizeof(msg)) {
138         error_report("failed to write, fd=%d, errno=%d (%s)",
139             fd, errno, strerror(errno));
140         return -EIO ;
141     }
142 
143     return ret;
144 }
145 
146 static void vhost_vdpa_listener_begin_batch(struct vhost_vdpa *v)
147 {
148     int fd = v->shared->device_fd;
149     struct vhost_msg_v2 msg = {
150         .type = VHOST_IOTLB_MSG_V2,
151         .iotlb.type = VHOST_IOTLB_BATCH_BEGIN,
152     };
153 
154     trace_vhost_vdpa_listener_begin_batch(v->shared, fd, msg.type,
155                                           msg.iotlb.type);
156     if (write(fd, &msg, sizeof(msg)) != sizeof(msg)) {
157         error_report("failed to write, fd=%d, errno=%d (%s)",
158                      fd, errno, strerror(errno));
159     }
160 }
161 
162 static void vhost_vdpa_iotlb_batch_begin_once(struct vhost_vdpa *v)
163 {
164     if (v->shared->backend_cap & (0x1ULL << VHOST_BACKEND_F_IOTLB_BATCH) &&
165         !v->shared->iotlb_batch_begin_sent) {
166         vhost_vdpa_listener_begin_batch(v);
167     }
168 
169     v->shared->iotlb_batch_begin_sent = true;
170 }
171 
172 static void vhost_vdpa_listener_commit(MemoryListener *listener)
173 {
174     struct vhost_vdpa *v = container_of(listener, struct vhost_vdpa, listener);
175     struct vhost_msg_v2 msg = {};
176     int fd = v->shared->device_fd;
177 
178     if (!(v->shared->backend_cap & (0x1ULL << VHOST_BACKEND_F_IOTLB_BATCH))) {
179         return;
180     }
181 
182     if (!v->shared->iotlb_batch_begin_sent) {
183         return;
184     }
185 
186     msg.type = VHOST_IOTLB_MSG_V2;
187     msg.iotlb.type = VHOST_IOTLB_BATCH_END;
188 
189     trace_vhost_vdpa_listener_commit(v->shared, fd, msg.type, msg.iotlb.type);
190     if (write(fd, &msg, sizeof(msg)) != sizeof(msg)) {
191         error_report("failed to write, fd=%d, errno=%d (%s)",
192                      fd, errno, strerror(errno));
193     }
194 
195     v->shared->iotlb_batch_begin_sent = false;
196 }
197 
198 static void vhost_vdpa_iommu_map_notify(IOMMUNotifier *n, IOMMUTLBEntry *iotlb)
199 {
200     struct vdpa_iommu *iommu = container_of(n, struct vdpa_iommu, n);
201 
202     hwaddr iova = iotlb->iova + iommu->iommu_offset;
203     struct vhost_vdpa *v = iommu->dev;
204     void *vaddr;
205     int ret;
206     Int128 llend;
207 
208     if (iotlb->target_as != &address_space_memory) {
209         error_report("Wrong target AS \"%s\", only system memory is allowed",
210                      iotlb->target_as->name ? iotlb->target_as->name : "none");
211         return;
212     }
213     RCU_READ_LOCK_GUARD();
214     /* check if RAM section out of device range */
215     llend = int128_add(int128_makes64(iotlb->addr_mask), int128_makes64(iova));
216     if (int128_gt(llend, int128_make64(v->shared->iova_range.last))) {
217         error_report("RAM section out of device range (max=0x%" PRIx64
218                      ", end addr=0x%" PRIx64 ")",
219                      v->shared->iova_range.last, int128_get64(llend));
220         return;
221     }
222 
223     if ((iotlb->perm & IOMMU_RW) != IOMMU_NONE) {
224         bool read_only;
225 
226         if (!memory_get_xlat_addr(iotlb, &vaddr, NULL, &read_only, NULL)) {
227             return;
228         }
229         ret = vhost_vdpa_dma_map(v, VHOST_VDPA_GUEST_PA_ASID, iova,
230                                  iotlb->addr_mask + 1, vaddr, read_only);
231         if (ret) {
232             error_report("vhost_vdpa_dma_map(%p, 0x%" HWADDR_PRIx ", "
233                          "0x%" HWADDR_PRIx ", %p) = %d (%m)",
234                          v, iova, iotlb->addr_mask + 1, vaddr, ret);
235         }
236     } else {
237         ret = vhost_vdpa_dma_unmap(v, VHOST_VDPA_GUEST_PA_ASID, iova,
238                                    iotlb->addr_mask + 1);
239         if (ret) {
240             error_report("vhost_vdpa_dma_unmap(%p, 0x%" HWADDR_PRIx ", "
241                          "0x%" HWADDR_PRIx ") = %d (%m)",
242                          v, iova, iotlb->addr_mask + 1, ret);
243         }
244     }
245 }
246 
247 static void vhost_vdpa_iommu_region_add(MemoryListener *listener,
248                                         MemoryRegionSection *section)
249 {
250     struct vhost_vdpa *v = container_of(listener, struct vhost_vdpa, listener);
251 
252     struct vdpa_iommu *iommu;
253     Int128 end;
254     int iommu_idx;
255     IOMMUMemoryRegion *iommu_mr;
256     int ret;
257 
258     iommu_mr = IOMMU_MEMORY_REGION(section->mr);
259 
260     iommu = g_malloc0(sizeof(*iommu));
261     end = int128_add(int128_make64(section->offset_within_region),
262                      section->size);
263     end = int128_sub(end, int128_one());
264     iommu_idx = memory_region_iommu_attrs_to_index(iommu_mr,
265                                                    MEMTXATTRS_UNSPECIFIED);
266     iommu->iommu_mr = iommu_mr;
267     iommu_notifier_init(&iommu->n, vhost_vdpa_iommu_map_notify,
268                         IOMMU_NOTIFIER_IOTLB_EVENTS,
269                         section->offset_within_region,
270                         int128_get64(end),
271                         iommu_idx);
272     iommu->iommu_offset = section->offset_within_address_space -
273                           section->offset_within_region;
274     iommu->dev = v;
275 
276     ret = memory_region_register_iommu_notifier(section->mr, &iommu->n, NULL);
277     if (ret) {
278         g_free(iommu);
279         return;
280     }
281 
282     QLIST_INSERT_HEAD(&v->iommu_list, iommu, iommu_next);
283     memory_region_iommu_replay(iommu->iommu_mr, &iommu->n);
284 
285     return;
286 }
287 
288 static void vhost_vdpa_iommu_region_del(MemoryListener *listener,
289                                         MemoryRegionSection *section)
290 {
291     struct vhost_vdpa *v = container_of(listener, struct vhost_vdpa, listener);
292 
293     struct vdpa_iommu *iommu;
294 
295     QLIST_FOREACH(iommu, &v->iommu_list, iommu_next)
296     {
297         if (MEMORY_REGION(iommu->iommu_mr) == section->mr &&
298             iommu->n.start == section->offset_within_region) {
299             memory_region_unregister_iommu_notifier(section->mr, &iommu->n);
300             QLIST_REMOVE(iommu, iommu_next);
301             g_free(iommu);
302             break;
303         }
304     }
305 }
306 
307 static void vhost_vdpa_listener_region_add(MemoryListener *listener,
308                                            MemoryRegionSection *section)
309 {
310     DMAMap mem_region = {};
311     struct vhost_vdpa *v = container_of(listener, struct vhost_vdpa, listener);
312     hwaddr iova;
313     Int128 llend, llsize;
314     void *vaddr;
315     int ret;
316     int page_size = qemu_target_page_size();
317     int page_mask = -page_size;
318 
319     if (vhost_vdpa_listener_skipped_section(section,
320                                             v->shared->iova_range.first,
321                                             v->shared->iova_range.last,
322                                             page_mask)) {
323         return;
324     }
325     if (memory_region_is_iommu(section->mr)) {
326         vhost_vdpa_iommu_region_add(listener, section);
327         return;
328     }
329 
330     if (unlikely((section->offset_within_address_space & ~page_mask) !=
331                  (section->offset_within_region & ~page_mask))) {
332         trace_vhost_vdpa_listener_region_add_unaligned(v->shared,
333                        section->mr->name,
334                        section->offset_within_address_space & ~page_mask,
335                        section->offset_within_region & ~page_mask);
336         return;
337     }
338 
339     iova = ROUND_UP(section->offset_within_address_space, page_size);
340     llend = vhost_vdpa_section_end(section, page_mask);
341     if (int128_ge(int128_make64(iova), llend)) {
342         return;
343     }
344 
345     memory_region_ref(section->mr);
346 
347     /* Here we assume that memory_region_is_ram(section->mr)==true */
348 
349     vaddr = memory_region_get_ram_ptr(section->mr) +
350             section->offset_within_region +
351             (iova - section->offset_within_address_space);
352 
353     trace_vhost_vdpa_listener_region_add(v->shared, iova, int128_get64(llend),
354                                          vaddr, section->readonly);
355 
356     llsize = int128_sub(llend, int128_make64(iova));
357     if (v->shared->shadow_data) {
358         int r;
359 
360         mem_region.translated_addr = (hwaddr)(uintptr_t)vaddr,
361         mem_region.size = int128_get64(llsize) - 1,
362         mem_region.perm = IOMMU_ACCESS_FLAG(true, section->readonly),
363 
364         r = vhost_iova_tree_map_alloc(v->shared->iova_tree, &mem_region);
365         if (unlikely(r != IOVA_OK)) {
366             error_report("Can't allocate a mapping (%d)", r);
367             goto fail;
368         }
369 
370         iova = mem_region.iova;
371     }
372 
373     vhost_vdpa_iotlb_batch_begin_once(v);
374     ret = vhost_vdpa_dma_map(v, VHOST_VDPA_GUEST_PA_ASID, iova,
375                              int128_get64(llsize), vaddr, section->readonly);
376     if (ret) {
377         error_report("vhost vdpa map fail!");
378         goto fail_map;
379     }
380 
381     return;
382 
383 fail_map:
384     if (v->shared->shadow_data) {
385         vhost_iova_tree_remove(v->shared->iova_tree, mem_region);
386     }
387 
388 fail:
389     /*
390      * On the initfn path, store the first error in the container so we
391      * can gracefully fail.  Runtime, there's not much we can do other
392      * than throw a hardware error.
393      */
394     error_report("vhost-vdpa: DMA mapping failed, unable to continue");
395     return;
396 
397 }
398 
399 static void vhost_vdpa_listener_region_del(MemoryListener *listener,
400                                            MemoryRegionSection *section)
401 {
402     struct vhost_vdpa *v = container_of(listener, struct vhost_vdpa, listener);
403     hwaddr iova;
404     Int128 llend, llsize;
405     int ret;
406     int page_size = qemu_target_page_size();
407     int page_mask = -page_size;
408 
409     if (vhost_vdpa_listener_skipped_section(section,
410                                             v->shared->iova_range.first,
411                                             v->shared->iova_range.last,
412                                             page_mask)) {
413         return;
414     }
415     if (memory_region_is_iommu(section->mr)) {
416         vhost_vdpa_iommu_region_del(listener, section);
417     }
418 
419     if (unlikely((section->offset_within_address_space & ~page_mask) !=
420                  (section->offset_within_region & ~page_mask))) {
421         trace_vhost_vdpa_listener_region_del_unaligned(v->shared,
422                        section->mr->name,
423                        section->offset_within_address_space & ~page_mask,
424                        section->offset_within_region & ~page_mask);
425         return;
426     }
427 
428     iova = ROUND_UP(section->offset_within_address_space, page_size);
429     llend = vhost_vdpa_section_end(section, page_mask);
430 
431     trace_vhost_vdpa_listener_region_del(v->shared, iova,
432         int128_get64(int128_sub(llend, int128_one())));
433 
434     if (int128_ge(int128_make64(iova), llend)) {
435         return;
436     }
437 
438     llsize = int128_sub(llend, int128_make64(iova));
439 
440     if (v->shared->shadow_data) {
441         const DMAMap *result;
442         const void *vaddr = memory_region_get_ram_ptr(section->mr) +
443             section->offset_within_region +
444             (iova - section->offset_within_address_space);
445         DMAMap mem_region = {
446             .translated_addr = (hwaddr)(uintptr_t)vaddr,
447             .size = int128_get64(llsize) - 1,
448         };
449 
450         result = vhost_iova_tree_find_iova(v->shared->iova_tree, &mem_region);
451         if (!result) {
452             /* The memory listener map wasn't mapped */
453             return;
454         }
455         iova = result->iova;
456         vhost_iova_tree_remove(v->shared->iova_tree, *result);
457     }
458     vhost_vdpa_iotlb_batch_begin_once(v);
459     /*
460      * The unmap ioctl doesn't accept a full 64-bit. need to check it
461      */
462     if (int128_eq(llsize, int128_2_64())) {
463         llsize = int128_rshift(llsize, 1);
464         ret = vhost_vdpa_dma_unmap(v, VHOST_VDPA_GUEST_PA_ASID, iova,
465                                    int128_get64(llsize));
466 
467         if (ret) {
468             error_report("vhost_vdpa_dma_unmap(%p, 0x%" HWADDR_PRIx ", "
469                          "0x%" HWADDR_PRIx ") = %d (%m)",
470                          v, iova, int128_get64(llsize), ret);
471         }
472         iova += int128_get64(llsize);
473     }
474     ret = vhost_vdpa_dma_unmap(v, VHOST_VDPA_GUEST_PA_ASID, iova,
475                                int128_get64(llsize));
476 
477     if (ret) {
478         error_report("vhost_vdpa_dma_unmap(%p, 0x%" HWADDR_PRIx ", "
479                      "0x%" HWADDR_PRIx ") = %d (%m)",
480                      v, iova, int128_get64(llsize), ret);
481     }
482 
483     memory_region_unref(section->mr);
484 }
485 /*
486  * IOTLB API is used by vhost-vdpa which requires incremental updating
487  * of the mapping. So we can not use generic vhost memory listener which
488  * depends on the addnop().
489  */
490 static const MemoryListener vhost_vdpa_memory_listener = {
491     .name = "vhost-vdpa",
492     .commit = vhost_vdpa_listener_commit,
493     .region_add = vhost_vdpa_listener_region_add,
494     .region_del = vhost_vdpa_listener_region_del,
495 };
496 
497 static int vhost_vdpa_call(struct vhost_dev *dev, unsigned long int request,
498                              void *arg)
499 {
500     struct vhost_vdpa *v = dev->opaque;
501     int fd = v->shared->device_fd;
502     int ret;
503 
504     assert(dev->vhost_ops->backend_type == VHOST_BACKEND_TYPE_VDPA);
505 
506     ret = ioctl(fd, request, arg);
507     return ret < 0 ? -errno : ret;
508 }
509 
510 static int vhost_vdpa_add_status(struct vhost_dev *dev, uint8_t status)
511 {
512     uint8_t s;
513     int ret;
514 
515     trace_vhost_vdpa_add_status(dev, status);
516     ret = vhost_vdpa_call(dev, VHOST_VDPA_GET_STATUS, &s);
517     if (ret < 0) {
518         return ret;
519     }
520     if ((s & status) == status) {
521         /* Don't set bits already set */
522         return 0;
523     }
524 
525     s |= status;
526 
527     ret = vhost_vdpa_call(dev, VHOST_VDPA_SET_STATUS, &s);
528     if (ret < 0) {
529         return ret;
530     }
531 
532     ret = vhost_vdpa_call(dev, VHOST_VDPA_GET_STATUS, &s);
533     if (ret < 0) {
534         return ret;
535     }
536 
537     if (!(s & status)) {
538         return -EIO;
539     }
540 
541     return 0;
542 }
543 
544 int vhost_vdpa_get_iova_range(int fd, struct vhost_vdpa_iova_range *iova_range)
545 {
546     int ret = ioctl(fd, VHOST_VDPA_GET_IOVA_RANGE, iova_range);
547 
548     return ret < 0 ? -errno : 0;
549 }
550 
551 /*
552  * The use of this function is for requests that only need to be
553  * applied once. Typically such request occurs at the beginning
554  * of operation, and before setting up queues. It should not be
555  * used for request that performs operation until all queues are
556  * set, which would need to check dev->vq_index_end instead.
557  */
558 static bool vhost_vdpa_first_dev(struct vhost_dev *dev)
559 {
560     struct vhost_vdpa *v = dev->opaque;
561 
562     return v->index == 0;
563 }
564 
565 static int vhost_vdpa_get_dev_features(struct vhost_dev *dev,
566                                        uint64_t *features)
567 {
568     int ret;
569 
570     ret = vhost_vdpa_call(dev, VHOST_GET_FEATURES, features);
571     trace_vhost_vdpa_get_features(dev, *features);
572     return ret;
573 }
574 
575 static void vhost_vdpa_init_svq(struct vhost_dev *hdev, struct vhost_vdpa *v)
576 {
577     g_autoptr(GPtrArray) shadow_vqs = NULL;
578 
579     shadow_vqs = g_ptr_array_new_full(hdev->nvqs, vhost_svq_free);
580     for (unsigned n = 0; n < hdev->nvqs; ++n) {
581         VhostShadowVirtqueue *svq;
582 
583         svq = vhost_svq_new(v->shadow_vq_ops, v->shadow_vq_ops_opaque);
584         g_ptr_array_add(shadow_vqs, svq);
585     }
586 
587     v->shadow_vqs = g_steal_pointer(&shadow_vqs);
588 }
589 
590 static int vhost_vdpa_init(struct vhost_dev *dev, void *opaque, Error **errp)
591 {
592     struct vhost_vdpa *v = opaque;
593     assert(dev->vhost_ops->backend_type == VHOST_BACKEND_TYPE_VDPA);
594     trace_vhost_vdpa_init(dev, v->shared, opaque);
595     int ret;
596 
597     v->dev = dev;
598     dev->opaque =  opaque ;
599     v->listener = vhost_vdpa_memory_listener;
600     vhost_vdpa_init_svq(dev, v);
601 
602     error_propagate(&dev->migration_blocker, v->migration_blocker);
603     if (!vhost_vdpa_first_dev(dev)) {
604         return 0;
605     }
606 
607     /*
608      * If dev->shadow_vqs_enabled at initialization that means the device has
609      * been started with x-svq=on, so don't block migration
610      */
611     if (dev->migration_blocker == NULL && !v->shadow_vqs_enabled) {
612         /* We don't have dev->features yet */
613         uint64_t features;
614         ret = vhost_vdpa_get_dev_features(dev, &features);
615         if (unlikely(ret)) {
616             error_setg_errno(errp, -ret, "Could not get device features");
617             return ret;
618         }
619         vhost_svq_valid_features(features, &dev->migration_blocker);
620     }
621 
622     /*
623      * Similar to VFIO, we end up pinning all guest memory and have to
624      * disable discarding of RAM.
625      */
626     ret = ram_block_discard_disable(true);
627     if (ret) {
628         error_report("Cannot set discarding of RAM broken");
629         return ret;
630     }
631 
632     vhost_vdpa_add_status(dev, VIRTIO_CONFIG_S_ACKNOWLEDGE |
633                                VIRTIO_CONFIG_S_DRIVER);
634 
635     return 0;
636 }
637 
638 static void vhost_vdpa_host_notifier_uninit(struct vhost_dev *dev,
639                                             int queue_index)
640 {
641     size_t page_size = qemu_real_host_page_size();
642     struct vhost_vdpa *v = dev->opaque;
643     VirtIODevice *vdev = dev->vdev;
644     VhostVDPAHostNotifier *n;
645 
646     n = &v->notifier[queue_index];
647 
648     if (n->addr) {
649         virtio_queue_set_host_notifier_mr(vdev, queue_index, &n->mr, false);
650         object_unparent(OBJECT(&n->mr));
651         munmap(n->addr, page_size);
652         n->addr = NULL;
653     }
654 }
655 
656 static int vhost_vdpa_host_notifier_init(struct vhost_dev *dev, int queue_index)
657 {
658     size_t page_size = qemu_real_host_page_size();
659     struct vhost_vdpa *v = dev->opaque;
660     VirtIODevice *vdev = dev->vdev;
661     VhostVDPAHostNotifier *n;
662     int fd = v->shared->device_fd;
663     void *addr;
664     char *name;
665 
666     vhost_vdpa_host_notifier_uninit(dev, queue_index);
667 
668     n = &v->notifier[queue_index];
669 
670     addr = mmap(NULL, page_size, PROT_WRITE, MAP_SHARED, fd,
671                 queue_index * page_size);
672     if (addr == MAP_FAILED) {
673         goto err;
674     }
675 
676     name = g_strdup_printf("vhost-vdpa/host-notifier@%p mmaps[%d]",
677                            v, queue_index);
678     memory_region_init_ram_device_ptr(&n->mr, OBJECT(vdev), name,
679                                       page_size, addr);
680     g_free(name);
681 
682     if (virtio_queue_set_host_notifier_mr(vdev, queue_index, &n->mr, true)) {
683         object_unparent(OBJECT(&n->mr));
684         munmap(addr, page_size);
685         goto err;
686     }
687     n->addr = addr;
688 
689     return 0;
690 
691 err:
692     return -1;
693 }
694 
695 static void vhost_vdpa_host_notifiers_uninit(struct vhost_dev *dev, int n)
696 {
697     int i;
698 
699     /*
700      * Pack all the changes to the memory regions in a single
701      * transaction to avoid a few updating of the address space
702      * topology.
703      */
704     memory_region_transaction_begin();
705 
706     for (i = dev->vq_index; i < dev->vq_index + n; i++) {
707         vhost_vdpa_host_notifier_uninit(dev, i);
708     }
709 
710     memory_region_transaction_commit();
711 }
712 
713 static void vhost_vdpa_host_notifiers_init(struct vhost_dev *dev)
714 {
715     struct vhost_vdpa *v = dev->opaque;
716     int i;
717 
718     if (v->shadow_vqs_enabled) {
719         /* FIXME SVQ is not compatible with host notifiers mr */
720         return;
721     }
722 
723     /*
724      * Pack all the changes to the memory regions in a single
725      * transaction to avoid a few updating of the address space
726      * topology.
727      */
728     memory_region_transaction_begin();
729 
730     for (i = dev->vq_index; i < dev->vq_index + dev->nvqs; i++) {
731         if (vhost_vdpa_host_notifier_init(dev, i)) {
732             vhost_vdpa_host_notifiers_uninit(dev, i - dev->vq_index);
733             break;
734         }
735     }
736 
737     memory_region_transaction_commit();
738 }
739 
740 static void vhost_vdpa_svq_cleanup(struct vhost_dev *dev)
741 {
742     struct vhost_vdpa *v = dev->opaque;
743     size_t idx;
744 
745     for (idx = 0; idx < v->shadow_vqs->len; ++idx) {
746         vhost_svq_stop(g_ptr_array_index(v->shadow_vqs, idx));
747     }
748     g_ptr_array_free(v->shadow_vqs, true);
749 }
750 
751 static int vhost_vdpa_cleanup(struct vhost_dev *dev)
752 {
753     struct vhost_vdpa *v;
754     assert(dev->vhost_ops->backend_type == VHOST_BACKEND_TYPE_VDPA);
755     v = dev->opaque;
756     trace_vhost_vdpa_cleanup(dev, v);
757     if (vhost_vdpa_first_dev(dev)) {
758         ram_block_discard_disable(false);
759     }
760 
761     vhost_vdpa_host_notifiers_uninit(dev, dev->nvqs);
762     memory_listener_unregister(&v->listener);
763     vhost_vdpa_svq_cleanup(dev);
764 
765     dev->opaque = NULL;
766 
767     return 0;
768 }
769 
770 static int vhost_vdpa_memslots_limit(struct vhost_dev *dev)
771 {
772     trace_vhost_vdpa_memslots_limit(dev, INT_MAX);
773     return INT_MAX;
774 }
775 
776 static int vhost_vdpa_set_mem_table(struct vhost_dev *dev,
777                                     struct vhost_memory *mem)
778 {
779     if (!vhost_vdpa_first_dev(dev)) {
780         return 0;
781     }
782 
783     trace_vhost_vdpa_set_mem_table(dev, mem->nregions, mem->padding);
784     if (trace_event_get_state_backends(TRACE_VHOST_VDPA_SET_MEM_TABLE) &&
785         trace_event_get_state_backends(TRACE_VHOST_VDPA_DUMP_REGIONS)) {
786         int i;
787         for (i = 0; i < mem->nregions; i++) {
788             trace_vhost_vdpa_dump_regions(dev, i,
789                                           mem->regions[i].guest_phys_addr,
790                                           mem->regions[i].memory_size,
791                                           mem->regions[i].userspace_addr,
792                                           mem->regions[i].flags_padding);
793         }
794     }
795     if (mem->padding) {
796         return -EINVAL;
797     }
798 
799     return 0;
800 }
801 
802 static int vhost_vdpa_set_features(struct vhost_dev *dev,
803                                    uint64_t features)
804 {
805     struct vhost_vdpa *v = dev->opaque;
806     int ret;
807 
808     if (!vhost_vdpa_first_dev(dev)) {
809         return 0;
810     }
811 
812     if (v->shadow_vqs_enabled) {
813         if ((v->acked_features ^ features) == BIT_ULL(VHOST_F_LOG_ALL)) {
814             /*
815              * QEMU is just trying to enable or disable logging. SVQ handles
816              * this sepparately, so no need to forward this.
817              */
818             v->acked_features = features;
819             return 0;
820         }
821 
822         v->acked_features = features;
823 
824         /* We must not ack _F_LOG if SVQ is enabled */
825         features &= ~BIT_ULL(VHOST_F_LOG_ALL);
826     }
827 
828     trace_vhost_vdpa_set_features(dev, features);
829     ret = vhost_vdpa_call(dev, VHOST_SET_FEATURES, &features);
830     if (ret) {
831         return ret;
832     }
833 
834     return vhost_vdpa_add_status(dev, VIRTIO_CONFIG_S_FEATURES_OK);
835 }
836 
837 static int vhost_vdpa_set_backend_cap(struct vhost_dev *dev)
838 {
839     struct vhost_vdpa *v = dev->opaque;
840 
841     uint64_t features;
842     uint64_t f = 0x1ULL << VHOST_BACKEND_F_IOTLB_MSG_V2 |
843         0x1ULL << VHOST_BACKEND_F_IOTLB_BATCH |
844         0x1ULL << VHOST_BACKEND_F_IOTLB_ASID |
845         0x1ULL << VHOST_BACKEND_F_SUSPEND;
846     int r;
847 
848     if (vhost_vdpa_call(dev, VHOST_GET_BACKEND_FEATURES, &features)) {
849         return -EFAULT;
850     }
851 
852     features &= f;
853 
854     if (vhost_vdpa_first_dev(dev)) {
855         r = vhost_vdpa_call(dev, VHOST_SET_BACKEND_FEATURES, &features);
856         if (r) {
857             return -EFAULT;
858         }
859     }
860 
861     dev->backend_cap = features;
862     v->shared->backend_cap = features;
863 
864     return 0;
865 }
866 
867 static int vhost_vdpa_get_device_id(struct vhost_dev *dev,
868                                     uint32_t *device_id)
869 {
870     int ret;
871     ret = vhost_vdpa_call(dev, VHOST_VDPA_GET_DEVICE_ID, device_id);
872     trace_vhost_vdpa_get_device_id(dev, *device_id);
873     return ret;
874 }
875 
876 static int vhost_vdpa_reset_device(struct vhost_dev *dev)
877 {
878     struct vhost_vdpa *v = dev->opaque;
879     int ret;
880     uint8_t status = 0;
881 
882     ret = vhost_vdpa_call(dev, VHOST_VDPA_SET_STATUS, &status);
883     trace_vhost_vdpa_reset_device(dev);
884     v->suspended = false;
885     return ret;
886 }
887 
888 static int vhost_vdpa_get_vq_index(struct vhost_dev *dev, int idx)
889 {
890     assert(idx >= dev->vq_index && idx < dev->vq_index + dev->nvqs);
891 
892     trace_vhost_vdpa_get_vq_index(dev, idx, idx);
893     return idx;
894 }
895 
896 int vhost_vdpa_set_vring_ready(struct vhost_vdpa *v, unsigned idx)
897 {
898     struct vhost_dev *dev = v->dev;
899     struct vhost_vring_state state = {
900         .index = idx,
901         .num = 1,
902     };
903     int r = vhost_vdpa_call(dev, VHOST_VDPA_SET_VRING_ENABLE, &state);
904 
905     trace_vhost_vdpa_set_vring_ready(dev, idx, r);
906     return r;
907 }
908 
909 static int vhost_vdpa_set_config_call(struct vhost_dev *dev,
910                                        int fd)
911 {
912     trace_vhost_vdpa_set_config_call(dev, fd);
913     return vhost_vdpa_call(dev, VHOST_VDPA_SET_CONFIG_CALL, &fd);
914 }
915 
916 static void vhost_vdpa_dump_config(struct vhost_dev *dev, const uint8_t *config,
917                                    uint32_t config_len)
918 {
919     int b, len;
920     char line[QEMU_HEXDUMP_LINE_LEN];
921 
922     for (b = 0; b < config_len; b += 16) {
923         len = config_len - b;
924         qemu_hexdump_line(line, b, config, len, false);
925         trace_vhost_vdpa_dump_config(dev, line);
926     }
927 }
928 
929 static int vhost_vdpa_set_config(struct vhost_dev *dev, const uint8_t *data,
930                                    uint32_t offset, uint32_t size,
931                                    uint32_t flags)
932 {
933     struct vhost_vdpa_config *config;
934     int ret;
935     unsigned long config_size = offsetof(struct vhost_vdpa_config, buf);
936 
937     trace_vhost_vdpa_set_config(dev, offset, size, flags);
938     config = g_malloc(size + config_size);
939     config->off = offset;
940     config->len = size;
941     memcpy(config->buf, data, size);
942     if (trace_event_get_state_backends(TRACE_VHOST_VDPA_SET_CONFIG) &&
943         trace_event_get_state_backends(TRACE_VHOST_VDPA_DUMP_CONFIG)) {
944         vhost_vdpa_dump_config(dev, data, size);
945     }
946     ret = vhost_vdpa_call(dev, VHOST_VDPA_SET_CONFIG, config);
947     g_free(config);
948     return ret;
949 }
950 
951 static int vhost_vdpa_get_config(struct vhost_dev *dev, uint8_t *config,
952                                    uint32_t config_len, Error **errp)
953 {
954     struct vhost_vdpa_config *v_config;
955     unsigned long config_size = offsetof(struct vhost_vdpa_config, buf);
956     int ret;
957 
958     trace_vhost_vdpa_get_config(dev, config, config_len);
959     v_config = g_malloc(config_len + config_size);
960     v_config->len = config_len;
961     v_config->off = 0;
962     ret = vhost_vdpa_call(dev, VHOST_VDPA_GET_CONFIG, v_config);
963     memcpy(config, v_config->buf, config_len);
964     g_free(v_config);
965     if (trace_event_get_state_backends(TRACE_VHOST_VDPA_GET_CONFIG) &&
966         trace_event_get_state_backends(TRACE_VHOST_VDPA_DUMP_CONFIG)) {
967         vhost_vdpa_dump_config(dev, config, config_len);
968     }
969     return ret;
970  }
971 
972 static int vhost_vdpa_set_dev_vring_base(struct vhost_dev *dev,
973                                          struct vhost_vring_state *ring)
974 {
975     trace_vhost_vdpa_set_vring_base(dev, ring->index, ring->num);
976     return vhost_vdpa_call(dev, VHOST_SET_VRING_BASE, ring);
977 }
978 
979 static int vhost_vdpa_set_vring_dev_kick(struct vhost_dev *dev,
980                                          struct vhost_vring_file *file)
981 {
982     trace_vhost_vdpa_set_vring_kick(dev, file->index, file->fd);
983     return vhost_vdpa_call(dev, VHOST_SET_VRING_KICK, file);
984 }
985 
986 static int vhost_vdpa_set_vring_dev_call(struct vhost_dev *dev,
987                                          struct vhost_vring_file *file)
988 {
989     trace_vhost_vdpa_set_vring_call(dev, file->index, file->fd);
990     return vhost_vdpa_call(dev, VHOST_SET_VRING_CALL, file);
991 }
992 
993 static int vhost_vdpa_set_vring_dev_addr(struct vhost_dev *dev,
994                                          struct vhost_vring_addr *addr)
995 {
996     trace_vhost_vdpa_set_vring_addr(dev, addr->index, addr->flags,
997                                 addr->desc_user_addr, addr->used_user_addr,
998                                 addr->avail_user_addr,
999                                 addr->log_guest_addr);
1000 
1001     return vhost_vdpa_call(dev, VHOST_SET_VRING_ADDR, addr);
1002 
1003 }
1004 
1005 /**
1006  * Set the shadow virtqueue descriptors to the device
1007  *
1008  * @dev: The vhost device model
1009  * @svq: The shadow virtqueue
1010  * @idx: The index of the virtqueue in the vhost device
1011  * @errp: Error
1012  *
1013  * Note that this function does not rewind kick file descriptor if cannot set
1014  * call one.
1015  */
1016 static int vhost_vdpa_svq_set_fds(struct vhost_dev *dev,
1017                                   VhostShadowVirtqueue *svq, unsigned idx,
1018                                   Error **errp)
1019 {
1020     struct vhost_vring_file file = {
1021         .index = dev->vq_index + idx,
1022     };
1023     const EventNotifier *event_notifier = &svq->hdev_kick;
1024     int r;
1025 
1026     r = event_notifier_init(&svq->hdev_kick, 0);
1027     if (r != 0) {
1028         error_setg_errno(errp, -r, "Couldn't create kick event notifier");
1029         goto err_init_hdev_kick;
1030     }
1031 
1032     r = event_notifier_init(&svq->hdev_call, 0);
1033     if (r != 0) {
1034         error_setg_errno(errp, -r, "Couldn't create call event notifier");
1035         goto err_init_hdev_call;
1036     }
1037 
1038     file.fd = event_notifier_get_fd(event_notifier);
1039     r = vhost_vdpa_set_vring_dev_kick(dev, &file);
1040     if (unlikely(r != 0)) {
1041         error_setg_errno(errp, -r, "Can't set device kick fd");
1042         goto err_init_set_dev_fd;
1043     }
1044 
1045     event_notifier = &svq->hdev_call;
1046     file.fd = event_notifier_get_fd(event_notifier);
1047     r = vhost_vdpa_set_vring_dev_call(dev, &file);
1048     if (unlikely(r != 0)) {
1049         error_setg_errno(errp, -r, "Can't set device call fd");
1050         goto err_init_set_dev_fd;
1051     }
1052 
1053     return 0;
1054 
1055 err_init_set_dev_fd:
1056     event_notifier_set_handler(&svq->hdev_call, NULL);
1057 
1058 err_init_hdev_call:
1059     event_notifier_cleanup(&svq->hdev_kick);
1060 
1061 err_init_hdev_kick:
1062     return r;
1063 }
1064 
1065 /**
1066  * Unmap a SVQ area in the device
1067  */
1068 static void vhost_vdpa_svq_unmap_ring(struct vhost_vdpa *v, hwaddr addr)
1069 {
1070     const DMAMap needle = {
1071         .translated_addr = addr,
1072     };
1073     const DMAMap *result = vhost_iova_tree_find_iova(v->shared->iova_tree,
1074                                                      &needle);
1075     hwaddr size;
1076     int r;
1077 
1078     if (unlikely(!result)) {
1079         error_report("Unable to find SVQ address to unmap");
1080         return;
1081     }
1082 
1083     size = ROUND_UP(result->size, qemu_real_host_page_size());
1084     r = vhost_vdpa_dma_unmap(v, v->address_space_id, result->iova, size);
1085     if (unlikely(r < 0)) {
1086         error_report("Unable to unmap SVQ vring: %s (%d)", g_strerror(-r), -r);
1087         return;
1088     }
1089 
1090     vhost_iova_tree_remove(v->shared->iova_tree, *result);
1091 }
1092 
1093 static void vhost_vdpa_svq_unmap_rings(struct vhost_dev *dev,
1094                                        const VhostShadowVirtqueue *svq)
1095 {
1096     struct vhost_vdpa *v = dev->opaque;
1097     struct vhost_vring_addr svq_addr;
1098 
1099     vhost_svq_get_vring_addr(svq, &svq_addr);
1100 
1101     vhost_vdpa_svq_unmap_ring(v, svq_addr.desc_user_addr);
1102 
1103     vhost_vdpa_svq_unmap_ring(v, svq_addr.used_user_addr);
1104 }
1105 
1106 /**
1107  * Map the SVQ area in the device
1108  *
1109  * @v: Vhost-vdpa device
1110  * @needle: The area to search iova
1111  * @errorp: Error pointer
1112  */
1113 static bool vhost_vdpa_svq_map_ring(struct vhost_vdpa *v, DMAMap *needle,
1114                                     Error **errp)
1115 {
1116     int r;
1117 
1118     r = vhost_iova_tree_map_alloc(v->shared->iova_tree, needle);
1119     if (unlikely(r != IOVA_OK)) {
1120         error_setg(errp, "Cannot allocate iova (%d)", r);
1121         return false;
1122     }
1123 
1124     r = vhost_vdpa_dma_map(v, v->address_space_id, needle->iova,
1125                            needle->size + 1,
1126                            (void *)(uintptr_t)needle->translated_addr,
1127                            needle->perm == IOMMU_RO);
1128     if (unlikely(r != 0)) {
1129         error_setg_errno(errp, -r, "Cannot map region to device");
1130         vhost_iova_tree_remove(v->shared->iova_tree, *needle);
1131     }
1132 
1133     return r == 0;
1134 }
1135 
1136 /**
1137  * Map the shadow virtqueue rings in the device
1138  *
1139  * @dev: The vhost device
1140  * @svq: The shadow virtqueue
1141  * @addr: Assigned IOVA addresses
1142  * @errp: Error pointer
1143  */
1144 static bool vhost_vdpa_svq_map_rings(struct vhost_dev *dev,
1145                                      const VhostShadowVirtqueue *svq,
1146                                      struct vhost_vring_addr *addr,
1147                                      Error **errp)
1148 {
1149     ERRP_GUARD();
1150     DMAMap device_region, driver_region;
1151     struct vhost_vring_addr svq_addr;
1152     struct vhost_vdpa *v = dev->opaque;
1153     size_t device_size = vhost_svq_device_area_size(svq);
1154     size_t driver_size = vhost_svq_driver_area_size(svq);
1155     size_t avail_offset;
1156     bool ok;
1157 
1158     vhost_svq_get_vring_addr(svq, &svq_addr);
1159 
1160     driver_region = (DMAMap) {
1161         .translated_addr = svq_addr.desc_user_addr,
1162         .size = driver_size - 1,
1163         .perm = IOMMU_RO,
1164     };
1165     ok = vhost_vdpa_svq_map_ring(v, &driver_region, errp);
1166     if (unlikely(!ok)) {
1167         error_prepend(errp, "Cannot create vq driver region: ");
1168         return false;
1169     }
1170     addr->desc_user_addr = driver_region.iova;
1171     avail_offset = svq_addr.avail_user_addr - svq_addr.desc_user_addr;
1172     addr->avail_user_addr = driver_region.iova + avail_offset;
1173 
1174     device_region = (DMAMap) {
1175         .translated_addr = svq_addr.used_user_addr,
1176         .size = device_size - 1,
1177         .perm = IOMMU_RW,
1178     };
1179     ok = vhost_vdpa_svq_map_ring(v, &device_region, errp);
1180     if (unlikely(!ok)) {
1181         error_prepend(errp, "Cannot create vq device region: ");
1182         vhost_vdpa_svq_unmap_ring(v, driver_region.translated_addr);
1183     }
1184     addr->used_user_addr = device_region.iova;
1185 
1186     return ok;
1187 }
1188 
1189 static bool vhost_vdpa_svq_setup(struct vhost_dev *dev,
1190                                  VhostShadowVirtqueue *svq, unsigned idx,
1191                                  Error **errp)
1192 {
1193     uint16_t vq_index = dev->vq_index + idx;
1194     struct vhost_vring_state s = {
1195         .index = vq_index,
1196     };
1197     int r;
1198 
1199     r = vhost_vdpa_set_dev_vring_base(dev, &s);
1200     if (unlikely(r)) {
1201         error_setg_errno(errp, -r, "Cannot set vring base");
1202         return false;
1203     }
1204 
1205     r = vhost_vdpa_svq_set_fds(dev, svq, idx, errp);
1206     return r == 0;
1207 }
1208 
1209 static bool vhost_vdpa_svqs_start(struct vhost_dev *dev)
1210 {
1211     struct vhost_vdpa *v = dev->opaque;
1212     Error *err = NULL;
1213     unsigned i;
1214 
1215     if (!v->shadow_vqs_enabled) {
1216         return true;
1217     }
1218 
1219     for (i = 0; i < v->shadow_vqs->len; ++i) {
1220         VirtQueue *vq = virtio_get_queue(dev->vdev, dev->vq_index + i);
1221         VhostShadowVirtqueue *svq = g_ptr_array_index(v->shadow_vqs, i);
1222         struct vhost_vring_addr addr = {
1223             .index = dev->vq_index + i,
1224         };
1225         int r;
1226         bool ok = vhost_vdpa_svq_setup(dev, svq, i, &err);
1227         if (unlikely(!ok)) {
1228             goto err;
1229         }
1230 
1231         vhost_svq_start(svq, dev->vdev, vq, v->shared->iova_tree);
1232         ok = vhost_vdpa_svq_map_rings(dev, svq, &addr, &err);
1233         if (unlikely(!ok)) {
1234             goto err_map;
1235         }
1236 
1237         /* Override vring GPA set by vhost subsystem */
1238         r = vhost_vdpa_set_vring_dev_addr(dev, &addr);
1239         if (unlikely(r != 0)) {
1240             error_setg_errno(&err, -r, "Cannot set device address");
1241             goto err_set_addr;
1242         }
1243     }
1244 
1245     return true;
1246 
1247 err_set_addr:
1248     vhost_vdpa_svq_unmap_rings(dev, g_ptr_array_index(v->shadow_vqs, i));
1249 
1250 err_map:
1251     vhost_svq_stop(g_ptr_array_index(v->shadow_vqs, i));
1252 
1253 err:
1254     error_reportf_err(err, "Cannot setup SVQ %u: ", i);
1255     for (unsigned j = 0; j < i; ++j) {
1256         VhostShadowVirtqueue *svq = g_ptr_array_index(v->shadow_vqs, j);
1257         vhost_vdpa_svq_unmap_rings(dev, svq);
1258         vhost_svq_stop(svq);
1259     }
1260 
1261     return false;
1262 }
1263 
1264 static void vhost_vdpa_svqs_stop(struct vhost_dev *dev)
1265 {
1266     struct vhost_vdpa *v = dev->opaque;
1267 
1268     if (!v->shadow_vqs_enabled) {
1269         return;
1270     }
1271 
1272     for (unsigned i = 0; i < v->shadow_vqs->len; ++i) {
1273         VhostShadowVirtqueue *svq = g_ptr_array_index(v->shadow_vqs, i);
1274 
1275         vhost_svq_stop(svq);
1276         vhost_vdpa_svq_unmap_rings(dev, svq);
1277 
1278         event_notifier_cleanup(&svq->hdev_kick);
1279         event_notifier_cleanup(&svq->hdev_call);
1280     }
1281 }
1282 
1283 static void vhost_vdpa_suspend(struct vhost_dev *dev)
1284 {
1285     struct vhost_vdpa *v = dev->opaque;
1286     int r;
1287 
1288     if (!vhost_vdpa_first_dev(dev)) {
1289         return;
1290     }
1291 
1292     if (dev->backend_cap & BIT_ULL(VHOST_BACKEND_F_SUSPEND)) {
1293         trace_vhost_vdpa_suspend(dev);
1294         r = ioctl(v->shared->device_fd, VHOST_VDPA_SUSPEND);
1295         if (unlikely(r)) {
1296             error_report("Cannot suspend: %s(%d)", g_strerror(errno), errno);
1297         } else {
1298             v->suspended = true;
1299             return;
1300         }
1301     }
1302 
1303     vhost_vdpa_reset_device(dev);
1304 }
1305 
1306 static int vhost_vdpa_dev_start(struct vhost_dev *dev, bool started)
1307 {
1308     struct vhost_vdpa *v = dev->opaque;
1309     bool ok;
1310     trace_vhost_vdpa_dev_start(dev, started);
1311 
1312     if (started) {
1313         vhost_vdpa_host_notifiers_init(dev);
1314         ok = vhost_vdpa_svqs_start(dev);
1315         if (unlikely(!ok)) {
1316             return -1;
1317         }
1318     } else {
1319         vhost_vdpa_suspend(dev);
1320         vhost_vdpa_svqs_stop(dev);
1321         vhost_vdpa_host_notifiers_uninit(dev, dev->nvqs);
1322     }
1323 
1324     if (dev->vq_index + dev->nvqs != dev->vq_index_end) {
1325         return 0;
1326     }
1327 
1328     if (started) {
1329         if (vhost_dev_has_iommu(dev) && (v->shadow_vqs_enabled)) {
1330             error_report("SVQ can not work while IOMMU enable, please disable"
1331                          "IOMMU and try again");
1332             return -1;
1333         }
1334         memory_listener_register(&v->listener, dev->vdev->dma_as);
1335 
1336         return vhost_vdpa_add_status(dev, VIRTIO_CONFIG_S_DRIVER_OK);
1337     }
1338 
1339     return 0;
1340 }
1341 
1342 static void vhost_vdpa_reset_status(struct vhost_dev *dev)
1343 {
1344     struct vhost_vdpa *v = dev->opaque;
1345 
1346     if (dev->vq_index + dev->nvqs != dev->vq_index_end) {
1347         return;
1348     }
1349 
1350     vhost_vdpa_reset_device(dev);
1351     vhost_vdpa_add_status(dev, VIRTIO_CONFIG_S_ACKNOWLEDGE |
1352                                VIRTIO_CONFIG_S_DRIVER);
1353     memory_listener_unregister(&v->listener);
1354 }
1355 
1356 static int vhost_vdpa_set_log_base(struct vhost_dev *dev, uint64_t base,
1357                                      struct vhost_log *log)
1358 {
1359     struct vhost_vdpa *v = dev->opaque;
1360     if (v->shadow_vqs_enabled || !vhost_vdpa_first_dev(dev)) {
1361         return 0;
1362     }
1363 
1364     trace_vhost_vdpa_set_log_base(dev, base, log->size, log->refcnt, log->fd,
1365                                   log->log);
1366     return vhost_vdpa_call(dev, VHOST_SET_LOG_BASE, &base);
1367 }
1368 
1369 static int vhost_vdpa_set_vring_addr(struct vhost_dev *dev,
1370                                        struct vhost_vring_addr *addr)
1371 {
1372     struct vhost_vdpa *v = dev->opaque;
1373 
1374     if (v->shadow_vqs_enabled) {
1375         /*
1376          * Device vring addr was set at device start. SVQ base is handled by
1377          * VirtQueue code.
1378          */
1379         return 0;
1380     }
1381 
1382     return vhost_vdpa_set_vring_dev_addr(dev, addr);
1383 }
1384 
1385 static int vhost_vdpa_set_vring_num(struct vhost_dev *dev,
1386                                       struct vhost_vring_state *ring)
1387 {
1388     trace_vhost_vdpa_set_vring_num(dev, ring->index, ring->num);
1389     return vhost_vdpa_call(dev, VHOST_SET_VRING_NUM, ring);
1390 }
1391 
1392 static int vhost_vdpa_set_vring_base(struct vhost_dev *dev,
1393                                        struct vhost_vring_state *ring)
1394 {
1395     struct vhost_vdpa *v = dev->opaque;
1396 
1397     if (v->shadow_vqs_enabled) {
1398         /*
1399          * Device vring base was set at device start. SVQ base is handled by
1400          * VirtQueue code.
1401          */
1402         return 0;
1403     }
1404 
1405     return vhost_vdpa_set_dev_vring_base(dev, ring);
1406 }
1407 
1408 static int vhost_vdpa_get_vring_base(struct vhost_dev *dev,
1409                                        struct vhost_vring_state *ring)
1410 {
1411     struct vhost_vdpa *v = dev->opaque;
1412     int ret;
1413 
1414     if (v->shadow_vqs_enabled) {
1415         ring->num = virtio_queue_get_last_avail_idx(dev->vdev, ring->index);
1416         return 0;
1417     }
1418 
1419     if (!v->suspended) {
1420         /*
1421          * Cannot trust in value returned by device, let vhost recover used
1422          * idx from guest.
1423          */
1424         return -1;
1425     }
1426 
1427     ret = vhost_vdpa_call(dev, VHOST_GET_VRING_BASE, ring);
1428     trace_vhost_vdpa_get_vring_base(dev, ring->index, ring->num);
1429     return ret;
1430 }
1431 
1432 static int vhost_vdpa_set_vring_kick(struct vhost_dev *dev,
1433                                        struct vhost_vring_file *file)
1434 {
1435     struct vhost_vdpa *v = dev->opaque;
1436     int vdpa_idx = file->index - dev->vq_index;
1437 
1438     if (v->shadow_vqs_enabled) {
1439         VhostShadowVirtqueue *svq = g_ptr_array_index(v->shadow_vqs, vdpa_idx);
1440         vhost_svq_set_svq_kick_fd(svq, file->fd);
1441         return 0;
1442     } else {
1443         return vhost_vdpa_set_vring_dev_kick(dev, file);
1444     }
1445 }
1446 
1447 static int vhost_vdpa_set_vring_call(struct vhost_dev *dev,
1448                                        struct vhost_vring_file *file)
1449 {
1450     struct vhost_vdpa *v = dev->opaque;
1451     int vdpa_idx = file->index - dev->vq_index;
1452     VhostShadowVirtqueue *svq = g_ptr_array_index(v->shadow_vqs, vdpa_idx);
1453 
1454     /* Remember last call fd because we can switch to SVQ anytime. */
1455     vhost_svq_set_svq_call_fd(svq, file->fd);
1456     if (v->shadow_vqs_enabled) {
1457         return 0;
1458     }
1459 
1460     return vhost_vdpa_set_vring_dev_call(dev, file);
1461 }
1462 
1463 static int vhost_vdpa_get_features(struct vhost_dev *dev,
1464                                      uint64_t *features)
1465 {
1466     int ret = vhost_vdpa_get_dev_features(dev, features);
1467 
1468     if (ret == 0) {
1469         /* Add SVQ logging capabilities */
1470         *features |= BIT_ULL(VHOST_F_LOG_ALL);
1471     }
1472 
1473     return ret;
1474 }
1475 
1476 static int vhost_vdpa_set_owner(struct vhost_dev *dev)
1477 {
1478     if (!vhost_vdpa_first_dev(dev)) {
1479         return 0;
1480     }
1481 
1482     trace_vhost_vdpa_set_owner(dev);
1483     return vhost_vdpa_call(dev, VHOST_SET_OWNER, NULL);
1484 }
1485 
1486 static int vhost_vdpa_vq_get_addr(struct vhost_dev *dev,
1487                     struct vhost_vring_addr *addr, struct vhost_virtqueue *vq)
1488 {
1489     assert(dev->vhost_ops->backend_type == VHOST_BACKEND_TYPE_VDPA);
1490     addr->desc_user_addr = (uint64_t)(unsigned long)vq->desc_phys;
1491     addr->avail_user_addr = (uint64_t)(unsigned long)vq->avail_phys;
1492     addr->used_user_addr = (uint64_t)(unsigned long)vq->used_phys;
1493     trace_vhost_vdpa_vq_get_addr(dev, vq, addr->desc_user_addr,
1494                                  addr->avail_user_addr, addr->used_user_addr);
1495     return 0;
1496 }
1497 
1498 static bool  vhost_vdpa_force_iommu(struct vhost_dev *dev)
1499 {
1500     return true;
1501 }
1502 
1503 const VhostOps vdpa_ops = {
1504         .backend_type = VHOST_BACKEND_TYPE_VDPA,
1505         .vhost_backend_init = vhost_vdpa_init,
1506         .vhost_backend_cleanup = vhost_vdpa_cleanup,
1507         .vhost_set_log_base = vhost_vdpa_set_log_base,
1508         .vhost_set_vring_addr = vhost_vdpa_set_vring_addr,
1509         .vhost_set_vring_num = vhost_vdpa_set_vring_num,
1510         .vhost_set_vring_base = vhost_vdpa_set_vring_base,
1511         .vhost_get_vring_base = vhost_vdpa_get_vring_base,
1512         .vhost_set_vring_kick = vhost_vdpa_set_vring_kick,
1513         .vhost_set_vring_call = vhost_vdpa_set_vring_call,
1514         .vhost_get_features = vhost_vdpa_get_features,
1515         .vhost_set_backend_cap = vhost_vdpa_set_backend_cap,
1516         .vhost_set_owner = vhost_vdpa_set_owner,
1517         .vhost_set_vring_endian = NULL,
1518         .vhost_backend_memslots_limit = vhost_vdpa_memslots_limit,
1519         .vhost_set_mem_table = vhost_vdpa_set_mem_table,
1520         .vhost_set_features = vhost_vdpa_set_features,
1521         .vhost_reset_device = vhost_vdpa_reset_device,
1522         .vhost_get_vq_index = vhost_vdpa_get_vq_index,
1523         .vhost_get_config  = vhost_vdpa_get_config,
1524         .vhost_set_config = vhost_vdpa_set_config,
1525         .vhost_requires_shm_log = NULL,
1526         .vhost_migration_done = NULL,
1527         .vhost_net_set_mtu = NULL,
1528         .vhost_set_iotlb_callback = NULL,
1529         .vhost_send_device_iotlb_msg = NULL,
1530         .vhost_dev_start = vhost_vdpa_dev_start,
1531         .vhost_get_device_id = vhost_vdpa_get_device_id,
1532         .vhost_vq_get_addr = vhost_vdpa_vq_get_addr,
1533         .vhost_force_iommu = vhost_vdpa_force_iommu,
1534         .vhost_set_config_call = vhost_vdpa_set_config_call,
1535         .vhost_reset_status = vhost_vdpa_reset_status,
1536 };
1537