xref: /openbmc/qemu/hw/virtio/vhost-vdpa.c (revision a4069309)
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     bool is_ram = memory_region_is_ram(section->mr);
51     bool is_iommu = memory_region_is_iommu(section->mr);
52     bool is_protected = memory_region_is_protected(section->mr);
53 
54     /* vhost-vDPA doesn't allow MMIO to be mapped  */
55     bool is_ram_device = memory_region_is_ram_device(section->mr);
56 
57     if ((!is_ram && !is_iommu) || is_protected || is_ram_device) {
58         trace_vhost_vdpa_skipped_memory_section(is_ram, is_iommu, is_protected,
59                                                 is_ram_device, iova_min,
60                                                 iova_max, page_mask);
61         return true;
62     }
63 
64     if (section->offset_within_address_space < iova_min) {
65         error_report("RAM section out of device range (min=0x%" PRIx64
66                      ", addr=0x%" HWADDR_PRIx ")",
67                      iova_min, section->offset_within_address_space);
68         return true;
69     }
70     /*
71      * While using vIOMMU, sometimes the section will be larger than iova_max,
72      * but the memory that actually maps is smaller, so move the check to
73      * function vhost_vdpa_iommu_map_notify(). That function will use the actual
74      * size that maps to the kernel
75      */
76 
77     if (!is_iommu) {
78         llend = vhost_vdpa_section_end(section, page_mask);
79         if (int128_gt(llend, int128_make64(iova_max))) {
80             error_report("RAM section out of device range (max=0x%" PRIx64
81                          ", end addr=0x%" PRIx64 ")",
82                          iova_max, int128_get64(llend));
83             return true;
84         }
85     }
86 
87     return false;
88 }
89 
90 /*
91  * The caller must set asid = 0 if the device does not support asid.
92  * This is not an ABI break since it is set to 0 by the initializer anyway.
93  */
94 int vhost_vdpa_dma_map(VhostVDPAShared *s, uint32_t asid, hwaddr iova,
95                        hwaddr size, void *vaddr, bool readonly)
96 {
97     struct vhost_msg_v2 msg = {};
98     int fd = s->device_fd;
99     int ret = 0;
100 
101     msg.type = VHOST_IOTLB_MSG_V2;
102     msg.asid = asid;
103     msg.iotlb.iova = iova;
104     msg.iotlb.size = size;
105     msg.iotlb.uaddr = (uint64_t)(uintptr_t)vaddr;
106     msg.iotlb.perm = readonly ? VHOST_ACCESS_RO : VHOST_ACCESS_RW;
107     msg.iotlb.type = VHOST_IOTLB_UPDATE;
108 
109     trace_vhost_vdpa_dma_map(s, fd, msg.type, msg.asid, msg.iotlb.iova,
110                              msg.iotlb.size, msg.iotlb.uaddr, msg.iotlb.perm,
111                              msg.iotlb.type);
112 
113     if (write(fd, &msg, sizeof(msg)) != sizeof(msg)) {
114         error_report("failed to write, fd=%d, errno=%d (%s)",
115             fd, errno, strerror(errno));
116         return -EIO ;
117     }
118 
119     return ret;
120 }
121 
122 /*
123  * The caller must set asid = 0 if the device does not support asid.
124  * This is not an ABI break since it is set to 0 by the initializer anyway.
125  */
126 int vhost_vdpa_dma_unmap(VhostVDPAShared *s, uint32_t asid, hwaddr iova,
127                          hwaddr size)
128 {
129     struct vhost_msg_v2 msg = {};
130     int fd = s->device_fd;
131     int ret = 0;
132 
133     msg.type = VHOST_IOTLB_MSG_V2;
134     msg.asid = asid;
135     msg.iotlb.iova = iova;
136     msg.iotlb.size = size;
137     msg.iotlb.type = VHOST_IOTLB_INVALIDATE;
138 
139     trace_vhost_vdpa_dma_unmap(s, fd, msg.type, msg.asid, msg.iotlb.iova,
140                                msg.iotlb.size, msg.iotlb.type);
141 
142     if (write(fd, &msg, sizeof(msg)) != sizeof(msg)) {
143         error_report("failed to write, fd=%d, errno=%d (%s)",
144             fd, errno, strerror(errno));
145         return -EIO ;
146     }
147 
148     return ret;
149 }
150 
151 static void vhost_vdpa_listener_begin_batch(VhostVDPAShared *s)
152 {
153     int fd = s->device_fd;
154     struct vhost_msg_v2 msg = {
155         .type = VHOST_IOTLB_MSG_V2,
156         .iotlb.type = VHOST_IOTLB_BATCH_BEGIN,
157     };
158 
159     trace_vhost_vdpa_listener_begin_batch(s, fd, msg.type, msg.iotlb.type);
160     if (write(fd, &msg, sizeof(msg)) != sizeof(msg)) {
161         error_report("failed to write, fd=%d, errno=%d (%s)",
162                      fd, errno, strerror(errno));
163     }
164 }
165 
166 static void vhost_vdpa_iotlb_batch_begin_once(VhostVDPAShared *s)
167 {
168     if (s->backend_cap & (0x1ULL << VHOST_BACKEND_F_IOTLB_BATCH) &&
169         !s->iotlb_batch_begin_sent) {
170         vhost_vdpa_listener_begin_batch(s);
171     }
172 
173     s->iotlb_batch_begin_sent = true;
174 }
175 
176 static void vhost_vdpa_listener_commit(MemoryListener *listener)
177 {
178     VhostVDPAShared *s = container_of(listener, VhostVDPAShared, listener);
179     struct vhost_msg_v2 msg = {};
180     int fd = s->device_fd;
181 
182     if (!(s->backend_cap & (0x1ULL << VHOST_BACKEND_F_IOTLB_BATCH))) {
183         return;
184     }
185 
186     if (!s->iotlb_batch_begin_sent) {
187         return;
188     }
189 
190     msg.type = VHOST_IOTLB_MSG_V2;
191     msg.iotlb.type = VHOST_IOTLB_BATCH_END;
192 
193     trace_vhost_vdpa_listener_commit(s, fd, msg.type, msg.iotlb.type);
194     if (write(fd, &msg, sizeof(msg)) != sizeof(msg)) {
195         error_report("failed to write, fd=%d, errno=%d (%s)",
196                      fd, errno, strerror(errno));
197     }
198 
199     s->iotlb_batch_begin_sent = false;
200 }
201 
202 static void vhost_vdpa_iommu_map_notify(IOMMUNotifier *n, IOMMUTLBEntry *iotlb)
203 {
204     struct vdpa_iommu *iommu = container_of(n, struct vdpa_iommu, n);
205 
206     hwaddr iova = iotlb->iova + iommu->iommu_offset;
207     VhostVDPAShared *s = iommu->dev_shared;
208     void *vaddr;
209     int ret;
210     Int128 llend;
211 
212     if (iotlb->target_as != &address_space_memory) {
213         error_report("Wrong target AS \"%s\", only system memory is allowed",
214                      iotlb->target_as->name ? iotlb->target_as->name : "none");
215         return;
216     }
217     RCU_READ_LOCK_GUARD();
218     /* check if RAM section out of device range */
219     llend = int128_add(int128_makes64(iotlb->addr_mask), int128_makes64(iova));
220     if (int128_gt(llend, int128_make64(s->iova_range.last))) {
221         error_report("RAM section out of device range (max=0x%" PRIx64
222                      ", end addr=0x%" PRIx64 ")",
223                      s->iova_range.last, int128_get64(llend));
224         return;
225     }
226 
227     if ((iotlb->perm & IOMMU_RW) != IOMMU_NONE) {
228         bool read_only;
229 
230         if (!memory_get_xlat_addr(iotlb, &vaddr, NULL, &read_only, NULL)) {
231             return;
232         }
233         ret = vhost_vdpa_dma_map(s, VHOST_VDPA_GUEST_PA_ASID, iova,
234                                  iotlb->addr_mask + 1, vaddr, read_only);
235         if (ret) {
236             error_report("vhost_vdpa_dma_map(%p, 0x%" HWADDR_PRIx ", "
237                          "0x%" HWADDR_PRIx ", %p) = %d (%m)",
238                          s, iova, iotlb->addr_mask + 1, vaddr, ret);
239         }
240     } else {
241         ret = vhost_vdpa_dma_unmap(s, VHOST_VDPA_GUEST_PA_ASID, iova,
242                                    iotlb->addr_mask + 1);
243         if (ret) {
244             error_report("vhost_vdpa_dma_unmap(%p, 0x%" HWADDR_PRIx ", "
245                          "0x%" HWADDR_PRIx ") = %d (%m)",
246                          s, iova, iotlb->addr_mask + 1, ret);
247         }
248     }
249 }
250 
251 static void vhost_vdpa_iommu_region_add(MemoryListener *listener,
252                                         MemoryRegionSection *section)
253 {
254     VhostVDPAShared *s = container_of(listener, VhostVDPAShared, listener);
255 
256     struct vdpa_iommu *iommu;
257     Int128 end;
258     int iommu_idx;
259     IOMMUMemoryRegion *iommu_mr;
260     int ret;
261 
262     iommu_mr = IOMMU_MEMORY_REGION(section->mr);
263 
264     iommu = g_malloc0(sizeof(*iommu));
265     end = int128_add(int128_make64(section->offset_within_region),
266                      section->size);
267     end = int128_sub(end, int128_one());
268     iommu_idx = memory_region_iommu_attrs_to_index(iommu_mr,
269                                                    MEMTXATTRS_UNSPECIFIED);
270     iommu->iommu_mr = iommu_mr;
271     iommu_notifier_init(&iommu->n, vhost_vdpa_iommu_map_notify,
272                         IOMMU_NOTIFIER_IOTLB_EVENTS,
273                         section->offset_within_region,
274                         int128_get64(end),
275                         iommu_idx);
276     iommu->iommu_offset = section->offset_within_address_space -
277                           section->offset_within_region;
278     iommu->dev_shared = s;
279 
280     ret = memory_region_register_iommu_notifier(section->mr, &iommu->n, NULL);
281     if (ret) {
282         g_free(iommu);
283         return;
284     }
285 
286     QLIST_INSERT_HEAD(&s->iommu_list, iommu, iommu_next);
287     memory_region_iommu_replay(iommu->iommu_mr, &iommu->n);
288 
289     return;
290 }
291 
292 static void vhost_vdpa_iommu_region_del(MemoryListener *listener,
293                                         MemoryRegionSection *section)
294 {
295     VhostVDPAShared *s = container_of(listener, VhostVDPAShared, listener);
296 
297     struct vdpa_iommu *iommu;
298 
299     QLIST_FOREACH(iommu, &s->iommu_list, iommu_next)
300     {
301         if (MEMORY_REGION(iommu->iommu_mr) == section->mr &&
302             iommu->n.start == section->offset_within_region) {
303             memory_region_unregister_iommu_notifier(section->mr, &iommu->n);
304             QLIST_REMOVE(iommu, iommu_next);
305             g_free(iommu);
306             break;
307         }
308     }
309 }
310 
311 static void vhost_vdpa_listener_region_add(MemoryListener *listener,
312                                            MemoryRegionSection *section)
313 {
314     DMAMap mem_region = {};
315     VhostVDPAShared *s = container_of(listener, VhostVDPAShared, listener);
316     hwaddr iova;
317     Int128 llend, llsize;
318     void *vaddr;
319     int ret;
320     int page_size = qemu_target_page_size();
321     int page_mask = -page_size;
322 
323     if (vhost_vdpa_listener_skipped_section(section, s->iova_range.first,
324                                             s->iova_range.last, page_mask)) {
325         return;
326     }
327     if (memory_region_is_iommu(section->mr)) {
328         vhost_vdpa_iommu_region_add(listener, section);
329         return;
330     }
331 
332     if (unlikely((section->offset_within_address_space & ~page_mask) !=
333                  (section->offset_within_region & ~page_mask))) {
334         trace_vhost_vdpa_listener_region_add_unaligned(s, section->mr->name,
335                        section->offset_within_address_space & ~page_mask,
336                        section->offset_within_region & ~page_mask);
337         return;
338     }
339 
340     iova = ROUND_UP(section->offset_within_address_space, page_size);
341     llend = vhost_vdpa_section_end(section, page_mask);
342     if (int128_ge(int128_make64(iova), llend)) {
343         return;
344     }
345 
346     memory_region_ref(section->mr);
347 
348     /* Here we assume that memory_region_is_ram(section->mr)==true */
349 
350     vaddr = memory_region_get_ram_ptr(section->mr) +
351             section->offset_within_region +
352             (iova - section->offset_within_address_space);
353 
354     trace_vhost_vdpa_listener_region_add(s, iova, int128_get64(llend),
355                                          vaddr, section->readonly);
356 
357     llsize = int128_sub(llend, int128_make64(iova));
358     if (s->shadow_data) {
359         int r;
360 
361         mem_region.translated_addr = (hwaddr)(uintptr_t)vaddr,
362         mem_region.size = int128_get64(llsize) - 1,
363         mem_region.perm = IOMMU_ACCESS_FLAG(true, section->readonly),
364 
365         r = vhost_iova_tree_map_alloc(s->iova_tree, &mem_region);
366         if (unlikely(r != IOVA_OK)) {
367             error_report("Can't allocate a mapping (%d)", r);
368             goto fail;
369         }
370 
371         iova = mem_region.iova;
372     }
373 
374     vhost_vdpa_iotlb_batch_begin_once(s);
375     ret = vhost_vdpa_dma_map(s, VHOST_VDPA_GUEST_PA_ASID, iova,
376                              int128_get64(llsize), vaddr, section->readonly);
377     if (ret) {
378         error_report("vhost vdpa map fail!");
379         goto fail_map;
380     }
381 
382     return;
383 
384 fail_map:
385     if (s->shadow_data) {
386         vhost_iova_tree_remove(s->iova_tree, mem_region);
387     }
388 
389 fail:
390     /*
391      * On the initfn path, store the first error in the container so we
392      * can gracefully fail.  Runtime, there's not much we can do other
393      * than throw a hardware error.
394      */
395     error_report("vhost-vdpa: DMA mapping failed, unable to continue");
396     return;
397 
398 }
399 
400 static void vhost_vdpa_listener_region_del(MemoryListener *listener,
401                                            MemoryRegionSection *section)
402 {
403     VhostVDPAShared *s = container_of(listener, VhostVDPAShared, listener);
404     hwaddr iova;
405     Int128 llend, llsize;
406     int ret;
407     int page_size = qemu_target_page_size();
408     int page_mask = -page_size;
409 
410     if (vhost_vdpa_listener_skipped_section(section, s->iova_range.first,
411                                             s->iova_range.last, page_mask)) {
412         return;
413     }
414     if (memory_region_is_iommu(section->mr)) {
415         vhost_vdpa_iommu_region_del(listener, section);
416     }
417 
418     if (unlikely((section->offset_within_address_space & ~page_mask) !=
419                  (section->offset_within_region & ~page_mask))) {
420         trace_vhost_vdpa_listener_region_del_unaligned(s, section->mr->name,
421                        section->offset_within_address_space & ~page_mask,
422                        section->offset_within_region & ~page_mask);
423         return;
424     }
425 
426     iova = ROUND_UP(section->offset_within_address_space, page_size);
427     llend = vhost_vdpa_section_end(section, page_mask);
428 
429     trace_vhost_vdpa_listener_region_del(s, iova,
430         int128_get64(int128_sub(llend, int128_one())));
431 
432     if (int128_ge(int128_make64(iova), llend)) {
433         return;
434     }
435 
436     llsize = int128_sub(llend, int128_make64(iova));
437 
438     if (s->shadow_data) {
439         const DMAMap *result;
440         const void *vaddr = memory_region_get_ram_ptr(section->mr) +
441             section->offset_within_region +
442             (iova - section->offset_within_address_space);
443         DMAMap mem_region = {
444             .translated_addr = (hwaddr)(uintptr_t)vaddr,
445             .size = int128_get64(llsize) - 1,
446         };
447 
448         result = vhost_iova_tree_find_iova(s->iova_tree, &mem_region);
449         if (!result) {
450             /* The memory listener map wasn't mapped */
451             return;
452         }
453         iova = result->iova;
454         vhost_iova_tree_remove(s->iova_tree, *result);
455     }
456     vhost_vdpa_iotlb_batch_begin_once(s);
457     /*
458      * The unmap ioctl doesn't accept a full 64-bit. need to check it
459      */
460     if (int128_eq(llsize, int128_2_64())) {
461         llsize = int128_rshift(llsize, 1);
462         ret = vhost_vdpa_dma_unmap(s, VHOST_VDPA_GUEST_PA_ASID, iova,
463                                    int128_get64(llsize));
464 
465         if (ret) {
466             error_report("vhost_vdpa_dma_unmap(%p, 0x%" HWADDR_PRIx ", "
467                          "0x%" HWADDR_PRIx ") = %d (%m)",
468                          s, iova, int128_get64(llsize), ret);
469         }
470         iova += int128_get64(llsize);
471     }
472     ret = vhost_vdpa_dma_unmap(s, VHOST_VDPA_GUEST_PA_ASID, iova,
473                                int128_get64(llsize));
474 
475     if (ret) {
476         error_report("vhost_vdpa_dma_unmap(%p, 0x%" HWADDR_PRIx ", "
477                      "0x%" HWADDR_PRIx ") = %d (%m)",
478                      s, iova, int128_get64(llsize), ret);
479     }
480 
481     memory_region_unref(section->mr);
482 }
483 /*
484  * IOTLB API is used by vhost-vdpa which requires incremental updating
485  * of the mapping. So we can not use generic vhost memory listener which
486  * depends on the addnop().
487  */
488 static const MemoryListener vhost_vdpa_memory_listener = {
489     .name = "vhost-vdpa",
490     .commit = vhost_vdpa_listener_commit,
491     .region_add = vhost_vdpa_listener_region_add,
492     .region_del = vhost_vdpa_listener_region_del,
493 };
494 
495 static int vhost_vdpa_call(struct vhost_dev *dev, unsigned long int request,
496                              void *arg)
497 {
498     struct vhost_vdpa *v = dev->opaque;
499     int fd = v->shared->device_fd;
500     int ret;
501 
502     assert(dev->vhost_ops->backend_type == VHOST_BACKEND_TYPE_VDPA);
503 
504     ret = ioctl(fd, request, arg);
505     return ret < 0 ? -errno : ret;
506 }
507 
508 static int vhost_vdpa_add_status(struct vhost_dev *dev, uint8_t status)
509 {
510     uint8_t s;
511     int ret;
512 
513     trace_vhost_vdpa_add_status(dev, status);
514     ret = vhost_vdpa_call(dev, VHOST_VDPA_GET_STATUS, &s);
515     if (ret < 0) {
516         return ret;
517     }
518     if ((s & status) == status) {
519         /* Don't set bits already set */
520         return 0;
521     }
522 
523     s |= status;
524 
525     ret = vhost_vdpa_call(dev, VHOST_VDPA_SET_STATUS, &s);
526     if (ret < 0) {
527         return ret;
528     }
529 
530     ret = vhost_vdpa_call(dev, VHOST_VDPA_GET_STATUS, &s);
531     if (ret < 0) {
532         return ret;
533     }
534 
535     if (!(s & status)) {
536         return -EIO;
537     }
538 
539     return 0;
540 }
541 
542 int vhost_vdpa_get_iova_range(int fd, struct vhost_vdpa_iova_range *iova_range)
543 {
544     int ret = ioctl(fd, VHOST_VDPA_GET_IOVA_RANGE, iova_range);
545 
546     return ret < 0 ? -errno : 0;
547 }
548 
549 /*
550  * The use of this function is for requests that only need to be
551  * applied once. Typically such request occurs at the beginning
552  * of operation, and before setting up queues. It should not be
553  * used for request that performs operation until all queues are
554  * set, which would need to check dev->vq_index_end instead.
555  */
556 static bool vhost_vdpa_first_dev(struct vhost_dev *dev)
557 {
558     struct vhost_vdpa *v = dev->opaque;
559 
560     return v->index == 0;
561 }
562 
563 static bool vhost_vdpa_last_dev(struct vhost_dev *dev)
564 {
565     return dev->vq_index + dev->nvqs == dev->vq_index_end;
566 }
567 
568 static int vhost_vdpa_get_dev_features(struct vhost_dev *dev,
569                                        uint64_t *features)
570 {
571     int ret;
572 
573     ret = vhost_vdpa_call(dev, VHOST_GET_FEATURES, features);
574     trace_vhost_vdpa_get_features(dev, *features);
575     return ret;
576 }
577 
578 static void vhost_vdpa_init_svq(struct vhost_dev *hdev, struct vhost_vdpa *v)
579 {
580     g_autoptr(GPtrArray) shadow_vqs = NULL;
581 
582     shadow_vqs = g_ptr_array_new_full(hdev->nvqs, vhost_svq_free);
583     for (unsigned n = 0; n < hdev->nvqs; ++n) {
584         VhostShadowVirtqueue *svq;
585 
586         svq = vhost_svq_new(v->shadow_vq_ops, v->shadow_vq_ops_opaque);
587         g_ptr_array_add(shadow_vqs, svq);
588     }
589 
590     v->shadow_vqs = g_steal_pointer(&shadow_vqs);
591 }
592 
593 static int vhost_vdpa_init(struct vhost_dev *dev, void *opaque, Error **errp)
594 {
595     struct vhost_vdpa *v = opaque;
596     assert(dev->vhost_ops->backend_type == VHOST_BACKEND_TYPE_VDPA);
597     trace_vhost_vdpa_init(dev, v->shared, opaque);
598     int ret;
599 
600     v->dev = dev;
601     dev->opaque =  opaque ;
602     v->shared->listener = vhost_vdpa_memory_listener;
603     vhost_vdpa_init_svq(dev, v);
604 
605     error_propagate(&dev->migration_blocker, v->migration_blocker);
606     if (!vhost_vdpa_first_dev(dev)) {
607         return 0;
608     }
609 
610     /*
611      * If dev->shadow_vqs_enabled at initialization that means the device has
612      * been started with x-svq=on, so don't block migration
613      */
614     if (dev->migration_blocker == NULL && !v->shadow_vqs_enabled) {
615         /* We don't have dev->features yet */
616         uint64_t features;
617         ret = vhost_vdpa_get_dev_features(dev, &features);
618         if (unlikely(ret)) {
619             error_setg_errno(errp, -ret, "Could not get device features");
620             return ret;
621         }
622         vhost_svq_valid_features(features, &dev->migration_blocker);
623     }
624 
625     /*
626      * Similar to VFIO, we end up pinning all guest memory and have to
627      * disable discarding of RAM.
628      */
629     ret = ram_block_discard_disable(true);
630     if (ret) {
631         error_report("Cannot set discarding of RAM broken");
632         return ret;
633     }
634 
635     vhost_vdpa_add_status(dev, VIRTIO_CONFIG_S_ACKNOWLEDGE |
636                                VIRTIO_CONFIG_S_DRIVER);
637 
638     return 0;
639 }
640 
641 static void vhost_vdpa_host_notifier_uninit(struct vhost_dev *dev,
642                                             int queue_index)
643 {
644     size_t page_size = qemu_real_host_page_size();
645     struct vhost_vdpa *v = dev->opaque;
646     VirtIODevice *vdev = dev->vdev;
647     VhostVDPAHostNotifier *n;
648 
649     n = &v->notifier[queue_index];
650 
651     if (n->addr) {
652         virtio_queue_set_host_notifier_mr(vdev, queue_index, &n->mr, false);
653         object_unparent(OBJECT(&n->mr));
654         munmap(n->addr, page_size);
655         n->addr = NULL;
656     }
657 }
658 
659 static int vhost_vdpa_host_notifier_init(struct vhost_dev *dev, int queue_index)
660 {
661     size_t page_size = qemu_real_host_page_size();
662     struct vhost_vdpa *v = dev->opaque;
663     VirtIODevice *vdev = dev->vdev;
664     VhostVDPAHostNotifier *n;
665     int fd = v->shared->device_fd;
666     void *addr;
667     char *name;
668 
669     vhost_vdpa_host_notifier_uninit(dev, queue_index);
670 
671     n = &v->notifier[queue_index];
672 
673     addr = mmap(NULL, page_size, PROT_WRITE, MAP_SHARED, fd,
674                 queue_index * page_size);
675     if (addr == MAP_FAILED) {
676         goto err;
677     }
678 
679     name = g_strdup_printf("vhost-vdpa/host-notifier@%p mmaps[%d]",
680                            v, queue_index);
681     memory_region_init_ram_device_ptr(&n->mr, OBJECT(vdev), name,
682                                       page_size, addr);
683     g_free(name);
684 
685     if (virtio_queue_set_host_notifier_mr(vdev, queue_index, &n->mr, true)) {
686         object_unparent(OBJECT(&n->mr));
687         munmap(addr, page_size);
688         goto err;
689     }
690     n->addr = addr;
691 
692     return 0;
693 
694 err:
695     return -1;
696 }
697 
698 static void vhost_vdpa_host_notifiers_uninit(struct vhost_dev *dev, int n)
699 {
700     int i;
701 
702     /*
703      * Pack all the changes to the memory regions in a single
704      * transaction to avoid a few updating of the address space
705      * topology.
706      */
707     memory_region_transaction_begin();
708 
709     for (i = dev->vq_index; i < dev->vq_index + n; i++) {
710         vhost_vdpa_host_notifier_uninit(dev, i);
711     }
712 
713     memory_region_transaction_commit();
714 }
715 
716 static void vhost_vdpa_host_notifiers_init(struct vhost_dev *dev)
717 {
718     struct vhost_vdpa *v = dev->opaque;
719     int i;
720 
721     if (v->shadow_vqs_enabled) {
722         /* FIXME SVQ is not compatible with host notifiers mr */
723         return;
724     }
725 
726     /*
727      * Pack all the changes to the memory regions in a single
728      * transaction to avoid a few updating of the address space
729      * topology.
730      */
731     memory_region_transaction_begin();
732 
733     for (i = dev->vq_index; i < dev->vq_index + dev->nvqs; i++) {
734         if (vhost_vdpa_host_notifier_init(dev, i)) {
735             vhost_vdpa_host_notifiers_uninit(dev, i - dev->vq_index);
736             break;
737         }
738     }
739 
740     memory_region_transaction_commit();
741 }
742 
743 static void vhost_vdpa_svq_cleanup(struct vhost_dev *dev)
744 {
745     struct vhost_vdpa *v = dev->opaque;
746     size_t idx;
747 
748     for (idx = 0; idx < v->shadow_vqs->len; ++idx) {
749         vhost_svq_stop(g_ptr_array_index(v->shadow_vqs, idx));
750     }
751     g_ptr_array_free(v->shadow_vqs, true);
752 }
753 
754 static int vhost_vdpa_cleanup(struct vhost_dev *dev)
755 {
756     struct vhost_vdpa *v;
757     assert(dev->vhost_ops->backend_type == VHOST_BACKEND_TYPE_VDPA);
758     v = dev->opaque;
759     trace_vhost_vdpa_cleanup(dev, v);
760     if (vhost_vdpa_first_dev(dev)) {
761         ram_block_discard_disable(false);
762         memory_listener_unregister(&v->shared->listener);
763     }
764 
765     vhost_vdpa_host_notifiers_uninit(dev, dev->nvqs);
766     vhost_vdpa_svq_cleanup(dev);
767 
768     dev->opaque = NULL;
769 
770     return 0;
771 }
772 
773 static int vhost_vdpa_memslots_limit(struct vhost_dev *dev)
774 {
775     trace_vhost_vdpa_memslots_limit(dev, INT_MAX);
776     return INT_MAX;
777 }
778 
779 static int vhost_vdpa_set_mem_table(struct vhost_dev *dev,
780                                     struct vhost_memory *mem)
781 {
782     if (!vhost_vdpa_first_dev(dev)) {
783         return 0;
784     }
785 
786     trace_vhost_vdpa_set_mem_table(dev, mem->nregions, mem->padding);
787     if (trace_event_get_state_backends(TRACE_VHOST_VDPA_SET_MEM_TABLE) &&
788         trace_event_get_state_backends(TRACE_VHOST_VDPA_DUMP_REGIONS)) {
789         int i;
790         for (i = 0; i < mem->nregions; i++) {
791             trace_vhost_vdpa_dump_regions(dev, i,
792                                           mem->regions[i].guest_phys_addr,
793                                           mem->regions[i].memory_size,
794                                           mem->regions[i].userspace_addr,
795                                           mem->regions[i].flags_padding);
796         }
797     }
798     if (mem->padding) {
799         return -EINVAL;
800     }
801 
802     return 0;
803 }
804 
805 static int vhost_vdpa_set_features(struct vhost_dev *dev,
806                                    uint64_t features)
807 {
808     struct vhost_vdpa *v = dev->opaque;
809     int ret;
810 
811     if (!vhost_vdpa_first_dev(dev)) {
812         return 0;
813     }
814 
815     if (v->shadow_vqs_enabled) {
816         if ((v->acked_features ^ features) == BIT_ULL(VHOST_F_LOG_ALL)) {
817             /*
818              * QEMU is just trying to enable or disable logging. SVQ handles
819              * this sepparately, so no need to forward this.
820              */
821             v->acked_features = features;
822             return 0;
823         }
824 
825         v->acked_features = features;
826 
827         /* We must not ack _F_LOG if SVQ is enabled */
828         features &= ~BIT_ULL(VHOST_F_LOG_ALL);
829     }
830 
831     trace_vhost_vdpa_set_features(dev, features);
832     ret = vhost_vdpa_call(dev, VHOST_SET_FEATURES, &features);
833     if (ret) {
834         return ret;
835     }
836 
837     return vhost_vdpa_add_status(dev, VIRTIO_CONFIG_S_FEATURES_OK);
838 }
839 
840 static int vhost_vdpa_set_backend_cap(struct vhost_dev *dev)
841 {
842     struct vhost_vdpa *v = dev->opaque;
843 
844     uint64_t features;
845     uint64_t f = 0x1ULL << VHOST_BACKEND_F_IOTLB_MSG_V2 |
846         0x1ULL << VHOST_BACKEND_F_IOTLB_BATCH |
847         0x1ULL << VHOST_BACKEND_F_IOTLB_ASID |
848         0x1ULL << VHOST_BACKEND_F_SUSPEND;
849     int r;
850 
851     if (vhost_vdpa_call(dev, VHOST_GET_BACKEND_FEATURES, &features)) {
852         return -EFAULT;
853     }
854 
855     features &= f;
856 
857     if (vhost_vdpa_first_dev(dev)) {
858         r = vhost_vdpa_call(dev, VHOST_SET_BACKEND_FEATURES, &features);
859         if (r) {
860             return -EFAULT;
861         }
862     }
863 
864     dev->backend_cap = features;
865     v->shared->backend_cap = features;
866 
867     return 0;
868 }
869 
870 static int vhost_vdpa_get_device_id(struct vhost_dev *dev,
871                                     uint32_t *device_id)
872 {
873     int ret;
874     ret = vhost_vdpa_call(dev, VHOST_VDPA_GET_DEVICE_ID, device_id);
875     trace_vhost_vdpa_get_device_id(dev, *device_id);
876     return ret;
877 }
878 
879 static int vhost_vdpa_reset_device(struct vhost_dev *dev)
880 {
881     struct vhost_vdpa *v = dev->opaque;
882     int ret;
883     uint8_t status = 0;
884 
885     ret = vhost_vdpa_call(dev, VHOST_VDPA_SET_STATUS, &status);
886     trace_vhost_vdpa_reset_device(dev);
887     v->suspended = false;
888     return ret;
889 }
890 
891 static int vhost_vdpa_get_vq_index(struct vhost_dev *dev, int idx)
892 {
893     assert(idx >= dev->vq_index && idx < dev->vq_index + dev->nvqs);
894 
895     trace_vhost_vdpa_get_vq_index(dev, idx, idx);
896     return idx;
897 }
898 
899 int vhost_vdpa_set_vring_ready(struct vhost_vdpa *v, unsigned idx)
900 {
901     struct vhost_dev *dev = v->dev;
902     struct vhost_vring_state state = {
903         .index = idx,
904         .num = 1,
905     };
906     int r = vhost_vdpa_call(dev, VHOST_VDPA_SET_VRING_ENABLE, &state);
907 
908     trace_vhost_vdpa_set_vring_ready(dev, idx, r);
909     return r;
910 }
911 
912 static int vhost_vdpa_set_config_call(struct vhost_dev *dev,
913                                        int fd)
914 {
915     trace_vhost_vdpa_set_config_call(dev, fd);
916     return vhost_vdpa_call(dev, VHOST_VDPA_SET_CONFIG_CALL, &fd);
917 }
918 
919 static void vhost_vdpa_dump_config(struct vhost_dev *dev, const uint8_t *config,
920                                    uint32_t config_len)
921 {
922     int b, len;
923     char line[QEMU_HEXDUMP_LINE_LEN];
924 
925     for (b = 0; b < config_len; b += 16) {
926         len = config_len - b;
927         qemu_hexdump_line(line, b, config, len, false);
928         trace_vhost_vdpa_dump_config(dev, line);
929     }
930 }
931 
932 static int vhost_vdpa_set_config(struct vhost_dev *dev, const uint8_t *data,
933                                    uint32_t offset, uint32_t size,
934                                    uint32_t flags)
935 {
936     struct vhost_vdpa_config *config;
937     int ret;
938     unsigned long config_size = offsetof(struct vhost_vdpa_config, buf);
939 
940     trace_vhost_vdpa_set_config(dev, offset, size, flags);
941     config = g_malloc(size + config_size);
942     config->off = offset;
943     config->len = size;
944     memcpy(config->buf, data, size);
945     if (trace_event_get_state_backends(TRACE_VHOST_VDPA_SET_CONFIG) &&
946         trace_event_get_state_backends(TRACE_VHOST_VDPA_DUMP_CONFIG)) {
947         vhost_vdpa_dump_config(dev, data, size);
948     }
949     ret = vhost_vdpa_call(dev, VHOST_VDPA_SET_CONFIG, config);
950     g_free(config);
951     return ret;
952 }
953 
954 static int vhost_vdpa_get_config(struct vhost_dev *dev, uint8_t *config,
955                                    uint32_t config_len, Error **errp)
956 {
957     struct vhost_vdpa_config *v_config;
958     unsigned long config_size = offsetof(struct vhost_vdpa_config, buf);
959     int ret;
960 
961     trace_vhost_vdpa_get_config(dev, config, config_len);
962     v_config = g_malloc(config_len + config_size);
963     v_config->len = config_len;
964     v_config->off = 0;
965     ret = vhost_vdpa_call(dev, VHOST_VDPA_GET_CONFIG, v_config);
966     memcpy(config, v_config->buf, config_len);
967     g_free(v_config);
968     if (trace_event_get_state_backends(TRACE_VHOST_VDPA_GET_CONFIG) &&
969         trace_event_get_state_backends(TRACE_VHOST_VDPA_DUMP_CONFIG)) {
970         vhost_vdpa_dump_config(dev, config, config_len);
971     }
972     return ret;
973  }
974 
975 static int vhost_vdpa_set_dev_vring_base(struct vhost_dev *dev,
976                                          struct vhost_vring_state *ring)
977 {
978     struct vhost_vdpa *v = dev->opaque;
979 
980     trace_vhost_vdpa_set_dev_vring_base(dev, ring->index, ring->num,
981                                         v->shadow_vqs_enabled);
982     return vhost_vdpa_call(dev, VHOST_SET_VRING_BASE, ring);
983 }
984 
985 static int vhost_vdpa_set_vring_dev_kick(struct vhost_dev *dev,
986                                          struct vhost_vring_file *file)
987 {
988     trace_vhost_vdpa_set_vring_kick(dev, file->index, file->fd);
989     return vhost_vdpa_call(dev, VHOST_SET_VRING_KICK, file);
990 }
991 
992 static int vhost_vdpa_set_vring_dev_call(struct vhost_dev *dev,
993                                          struct vhost_vring_file *file)
994 {
995     trace_vhost_vdpa_set_vring_call(dev, file->index, file->fd);
996     return vhost_vdpa_call(dev, VHOST_SET_VRING_CALL, file);
997 }
998 
999 static int vhost_vdpa_set_vring_dev_addr(struct vhost_dev *dev,
1000                                          struct vhost_vring_addr *addr)
1001 {
1002     trace_vhost_vdpa_set_vring_addr(dev, addr->index, addr->flags,
1003                                 addr->desc_user_addr, addr->used_user_addr,
1004                                 addr->avail_user_addr,
1005                                 addr->log_guest_addr);
1006 
1007     return vhost_vdpa_call(dev, VHOST_SET_VRING_ADDR, addr);
1008 
1009 }
1010 
1011 /**
1012  * Set the shadow virtqueue descriptors to the device
1013  *
1014  * @dev: The vhost device model
1015  * @svq: The shadow virtqueue
1016  * @idx: The index of the virtqueue in the vhost device
1017  * @errp: Error
1018  *
1019  * Note that this function does not rewind kick file descriptor if cannot set
1020  * call one.
1021  */
1022 static int vhost_vdpa_svq_set_fds(struct vhost_dev *dev,
1023                                   VhostShadowVirtqueue *svq, unsigned idx,
1024                                   Error **errp)
1025 {
1026     struct vhost_vring_file file = {
1027         .index = dev->vq_index + idx,
1028     };
1029     const EventNotifier *event_notifier = &svq->hdev_kick;
1030     int r;
1031 
1032     r = event_notifier_init(&svq->hdev_kick, 0);
1033     if (r != 0) {
1034         error_setg_errno(errp, -r, "Couldn't create kick event notifier");
1035         goto err_init_hdev_kick;
1036     }
1037 
1038     r = event_notifier_init(&svq->hdev_call, 0);
1039     if (r != 0) {
1040         error_setg_errno(errp, -r, "Couldn't create call event notifier");
1041         goto err_init_hdev_call;
1042     }
1043 
1044     file.fd = event_notifier_get_fd(event_notifier);
1045     r = vhost_vdpa_set_vring_dev_kick(dev, &file);
1046     if (unlikely(r != 0)) {
1047         error_setg_errno(errp, -r, "Can't set device kick fd");
1048         goto err_init_set_dev_fd;
1049     }
1050 
1051     event_notifier = &svq->hdev_call;
1052     file.fd = event_notifier_get_fd(event_notifier);
1053     r = vhost_vdpa_set_vring_dev_call(dev, &file);
1054     if (unlikely(r != 0)) {
1055         error_setg_errno(errp, -r, "Can't set device call fd");
1056         goto err_init_set_dev_fd;
1057     }
1058 
1059     return 0;
1060 
1061 err_init_set_dev_fd:
1062     event_notifier_set_handler(&svq->hdev_call, NULL);
1063 
1064 err_init_hdev_call:
1065     event_notifier_cleanup(&svq->hdev_kick);
1066 
1067 err_init_hdev_kick:
1068     return r;
1069 }
1070 
1071 /**
1072  * Unmap a SVQ area in the device
1073  */
1074 static void vhost_vdpa_svq_unmap_ring(struct vhost_vdpa *v, hwaddr addr)
1075 {
1076     const DMAMap needle = {
1077         .translated_addr = addr,
1078     };
1079     const DMAMap *result = vhost_iova_tree_find_iova(v->shared->iova_tree,
1080                                                      &needle);
1081     hwaddr size;
1082     int r;
1083 
1084     if (unlikely(!result)) {
1085         error_report("Unable to find SVQ address to unmap");
1086         return;
1087     }
1088 
1089     size = ROUND_UP(result->size, qemu_real_host_page_size());
1090     r = vhost_vdpa_dma_unmap(v->shared, v->address_space_id, result->iova,
1091                              size);
1092     if (unlikely(r < 0)) {
1093         error_report("Unable to unmap SVQ vring: %s (%d)", g_strerror(-r), -r);
1094         return;
1095     }
1096 
1097     vhost_iova_tree_remove(v->shared->iova_tree, *result);
1098 }
1099 
1100 static void vhost_vdpa_svq_unmap_rings(struct vhost_dev *dev,
1101                                        const VhostShadowVirtqueue *svq)
1102 {
1103     struct vhost_vdpa *v = dev->opaque;
1104     struct vhost_vring_addr svq_addr;
1105 
1106     vhost_svq_get_vring_addr(svq, &svq_addr);
1107 
1108     vhost_vdpa_svq_unmap_ring(v, svq_addr.desc_user_addr);
1109 
1110     vhost_vdpa_svq_unmap_ring(v, svq_addr.used_user_addr);
1111 }
1112 
1113 /**
1114  * Map the SVQ area in the device
1115  *
1116  * @v: Vhost-vdpa device
1117  * @needle: The area to search iova
1118  * @errorp: Error pointer
1119  */
1120 static bool vhost_vdpa_svq_map_ring(struct vhost_vdpa *v, DMAMap *needle,
1121                                     Error **errp)
1122 {
1123     int r;
1124 
1125     r = vhost_iova_tree_map_alloc(v->shared->iova_tree, needle);
1126     if (unlikely(r != IOVA_OK)) {
1127         error_setg(errp, "Cannot allocate iova (%d)", r);
1128         return false;
1129     }
1130 
1131     r = vhost_vdpa_dma_map(v->shared, v->address_space_id, needle->iova,
1132                            needle->size + 1,
1133                            (void *)(uintptr_t)needle->translated_addr,
1134                            needle->perm == IOMMU_RO);
1135     if (unlikely(r != 0)) {
1136         error_setg_errno(errp, -r, "Cannot map region to device");
1137         vhost_iova_tree_remove(v->shared->iova_tree, *needle);
1138     }
1139 
1140     return r == 0;
1141 }
1142 
1143 /**
1144  * Map the shadow virtqueue rings in the device
1145  *
1146  * @dev: The vhost device
1147  * @svq: The shadow virtqueue
1148  * @addr: Assigned IOVA addresses
1149  * @errp: Error pointer
1150  */
1151 static bool vhost_vdpa_svq_map_rings(struct vhost_dev *dev,
1152                                      const VhostShadowVirtqueue *svq,
1153                                      struct vhost_vring_addr *addr,
1154                                      Error **errp)
1155 {
1156     ERRP_GUARD();
1157     DMAMap device_region, driver_region;
1158     struct vhost_vring_addr svq_addr;
1159     struct vhost_vdpa *v = dev->opaque;
1160     size_t device_size = vhost_svq_device_area_size(svq);
1161     size_t driver_size = vhost_svq_driver_area_size(svq);
1162     size_t avail_offset;
1163     bool ok;
1164 
1165     vhost_svq_get_vring_addr(svq, &svq_addr);
1166 
1167     driver_region = (DMAMap) {
1168         .translated_addr = svq_addr.desc_user_addr,
1169         .size = driver_size - 1,
1170         .perm = IOMMU_RO,
1171     };
1172     ok = vhost_vdpa_svq_map_ring(v, &driver_region, errp);
1173     if (unlikely(!ok)) {
1174         error_prepend(errp, "Cannot create vq driver region: ");
1175         return false;
1176     }
1177     addr->desc_user_addr = driver_region.iova;
1178     avail_offset = svq_addr.avail_user_addr - svq_addr.desc_user_addr;
1179     addr->avail_user_addr = driver_region.iova + avail_offset;
1180 
1181     device_region = (DMAMap) {
1182         .translated_addr = svq_addr.used_user_addr,
1183         .size = device_size - 1,
1184         .perm = IOMMU_RW,
1185     };
1186     ok = vhost_vdpa_svq_map_ring(v, &device_region, errp);
1187     if (unlikely(!ok)) {
1188         error_prepend(errp, "Cannot create vq device region: ");
1189         vhost_vdpa_svq_unmap_ring(v, driver_region.translated_addr);
1190     }
1191     addr->used_user_addr = device_region.iova;
1192 
1193     return ok;
1194 }
1195 
1196 static bool vhost_vdpa_svq_setup(struct vhost_dev *dev,
1197                                  VhostShadowVirtqueue *svq, unsigned idx,
1198                                  Error **errp)
1199 {
1200     uint16_t vq_index = dev->vq_index + idx;
1201     struct vhost_vring_state s = {
1202         .index = vq_index,
1203     };
1204     int r;
1205 
1206     r = vhost_vdpa_set_dev_vring_base(dev, &s);
1207     if (unlikely(r)) {
1208         error_setg_errno(errp, -r, "Cannot set vring base");
1209         return false;
1210     }
1211 
1212     r = vhost_vdpa_svq_set_fds(dev, svq, idx, errp);
1213     return r == 0;
1214 }
1215 
1216 static bool vhost_vdpa_svqs_start(struct vhost_dev *dev)
1217 {
1218     struct vhost_vdpa *v = dev->opaque;
1219     Error *err = NULL;
1220     unsigned i;
1221 
1222     if (!v->shadow_vqs_enabled) {
1223         return true;
1224     }
1225 
1226     for (i = 0; i < v->shadow_vqs->len; ++i) {
1227         VirtQueue *vq = virtio_get_queue(dev->vdev, dev->vq_index + i);
1228         VhostShadowVirtqueue *svq = g_ptr_array_index(v->shadow_vqs, i);
1229         struct vhost_vring_addr addr = {
1230             .index = dev->vq_index + i,
1231         };
1232         int r;
1233         bool ok = vhost_vdpa_svq_setup(dev, svq, i, &err);
1234         if (unlikely(!ok)) {
1235             goto err;
1236         }
1237 
1238         vhost_svq_start(svq, dev->vdev, vq, v->shared->iova_tree);
1239         ok = vhost_vdpa_svq_map_rings(dev, svq, &addr, &err);
1240         if (unlikely(!ok)) {
1241             goto err_map;
1242         }
1243 
1244         /* Override vring GPA set by vhost subsystem */
1245         r = vhost_vdpa_set_vring_dev_addr(dev, &addr);
1246         if (unlikely(r != 0)) {
1247             error_setg_errno(&err, -r, "Cannot set device address");
1248             goto err_set_addr;
1249         }
1250     }
1251 
1252     return true;
1253 
1254 err_set_addr:
1255     vhost_vdpa_svq_unmap_rings(dev, g_ptr_array_index(v->shadow_vqs, i));
1256 
1257 err_map:
1258     vhost_svq_stop(g_ptr_array_index(v->shadow_vqs, i));
1259 
1260 err:
1261     error_reportf_err(err, "Cannot setup SVQ %u: ", i);
1262     for (unsigned j = 0; j < i; ++j) {
1263         VhostShadowVirtqueue *svq = g_ptr_array_index(v->shadow_vqs, j);
1264         vhost_vdpa_svq_unmap_rings(dev, svq);
1265         vhost_svq_stop(svq);
1266     }
1267 
1268     return false;
1269 }
1270 
1271 static void vhost_vdpa_svqs_stop(struct vhost_dev *dev)
1272 {
1273     struct vhost_vdpa *v = dev->opaque;
1274 
1275     if (!v->shadow_vqs_enabled) {
1276         return;
1277     }
1278 
1279     for (unsigned i = 0; i < v->shadow_vqs->len; ++i) {
1280         VhostShadowVirtqueue *svq = g_ptr_array_index(v->shadow_vqs, i);
1281 
1282         vhost_svq_stop(svq);
1283         vhost_vdpa_svq_unmap_rings(dev, svq);
1284 
1285         event_notifier_cleanup(&svq->hdev_kick);
1286         event_notifier_cleanup(&svq->hdev_call);
1287     }
1288 }
1289 
1290 static void vhost_vdpa_suspend(struct vhost_dev *dev)
1291 {
1292     struct vhost_vdpa *v = dev->opaque;
1293     int r;
1294 
1295     if (!vhost_vdpa_first_dev(dev)) {
1296         return;
1297     }
1298 
1299     if (dev->backend_cap & BIT_ULL(VHOST_BACKEND_F_SUSPEND)) {
1300         trace_vhost_vdpa_suspend(dev);
1301         r = ioctl(v->shared->device_fd, VHOST_VDPA_SUSPEND);
1302         if (unlikely(r)) {
1303             error_report("Cannot suspend: %s(%d)", g_strerror(errno), errno);
1304         } else {
1305             v->suspended = true;
1306             return;
1307         }
1308     }
1309 
1310     vhost_vdpa_reset_device(dev);
1311 }
1312 
1313 static int vhost_vdpa_dev_start(struct vhost_dev *dev, bool started)
1314 {
1315     struct vhost_vdpa *v = dev->opaque;
1316     bool ok;
1317     trace_vhost_vdpa_dev_start(dev, started);
1318 
1319     if (started) {
1320         vhost_vdpa_host_notifiers_init(dev);
1321         ok = vhost_vdpa_svqs_start(dev);
1322         if (unlikely(!ok)) {
1323             return -1;
1324         }
1325     } else {
1326         vhost_vdpa_suspend(dev);
1327         vhost_vdpa_svqs_stop(dev);
1328         vhost_vdpa_host_notifiers_uninit(dev, dev->nvqs);
1329     }
1330 
1331     if (!vhost_vdpa_last_dev(dev)) {
1332         return 0;
1333     }
1334 
1335     if (started) {
1336         if (vhost_dev_has_iommu(dev) && (v->shadow_vqs_enabled)) {
1337             error_report("SVQ can not work while IOMMU enable, please disable"
1338                          "IOMMU and try again");
1339             return -1;
1340         }
1341         memory_listener_register(&v->shared->listener, dev->vdev->dma_as);
1342 
1343         return vhost_vdpa_add_status(dev, VIRTIO_CONFIG_S_DRIVER_OK);
1344     }
1345 
1346     return 0;
1347 }
1348 
1349 static void vhost_vdpa_reset_status(struct vhost_dev *dev)
1350 {
1351     struct vhost_vdpa *v = dev->opaque;
1352 
1353     if (!vhost_vdpa_last_dev(dev)) {
1354         return;
1355     }
1356 
1357     vhost_vdpa_reset_device(dev);
1358     vhost_vdpa_add_status(dev, VIRTIO_CONFIG_S_ACKNOWLEDGE |
1359                                VIRTIO_CONFIG_S_DRIVER);
1360     memory_listener_unregister(&v->shared->listener);
1361 }
1362 
1363 static int vhost_vdpa_set_log_base(struct vhost_dev *dev, uint64_t base,
1364                                      struct vhost_log *log)
1365 {
1366     struct vhost_vdpa *v = dev->opaque;
1367     if (v->shadow_vqs_enabled || !vhost_vdpa_first_dev(dev)) {
1368         return 0;
1369     }
1370 
1371     trace_vhost_vdpa_set_log_base(dev, base, log->size, log->refcnt, log->fd,
1372                                   log->log);
1373     return vhost_vdpa_call(dev, VHOST_SET_LOG_BASE, &base);
1374 }
1375 
1376 static int vhost_vdpa_set_vring_addr(struct vhost_dev *dev,
1377                                        struct vhost_vring_addr *addr)
1378 {
1379     struct vhost_vdpa *v = dev->opaque;
1380 
1381     if (v->shadow_vqs_enabled) {
1382         /*
1383          * Device vring addr was set at device start. SVQ base is handled by
1384          * VirtQueue code.
1385          */
1386         return 0;
1387     }
1388 
1389     return vhost_vdpa_set_vring_dev_addr(dev, addr);
1390 }
1391 
1392 static int vhost_vdpa_set_vring_num(struct vhost_dev *dev,
1393                                       struct vhost_vring_state *ring)
1394 {
1395     trace_vhost_vdpa_set_vring_num(dev, ring->index, ring->num);
1396     return vhost_vdpa_call(dev, VHOST_SET_VRING_NUM, ring);
1397 }
1398 
1399 static int vhost_vdpa_set_vring_base(struct vhost_dev *dev,
1400                                        struct vhost_vring_state *ring)
1401 {
1402     struct vhost_vdpa *v = dev->opaque;
1403 
1404     if (v->shadow_vqs_enabled) {
1405         /*
1406          * Device vring base was set at device start. SVQ base is handled by
1407          * VirtQueue code.
1408          */
1409         return 0;
1410     }
1411 
1412     return vhost_vdpa_set_dev_vring_base(dev, ring);
1413 }
1414 
1415 static int vhost_vdpa_get_vring_base(struct vhost_dev *dev,
1416                                        struct vhost_vring_state *ring)
1417 {
1418     struct vhost_vdpa *v = dev->opaque;
1419     int ret;
1420 
1421     if (v->shadow_vqs_enabled) {
1422         ring->num = virtio_queue_get_last_avail_idx(dev->vdev, ring->index);
1423         trace_vhost_vdpa_get_vring_base(dev, ring->index, ring->num, true);
1424         return 0;
1425     }
1426 
1427     if (!v->suspended) {
1428         /*
1429          * Cannot trust in value returned by device, let vhost recover used
1430          * idx from guest.
1431          */
1432         return -1;
1433     }
1434 
1435     ret = vhost_vdpa_call(dev, VHOST_GET_VRING_BASE, ring);
1436     trace_vhost_vdpa_get_vring_base(dev, ring->index, ring->num, false);
1437     return ret;
1438 }
1439 
1440 static int vhost_vdpa_set_vring_kick(struct vhost_dev *dev,
1441                                        struct vhost_vring_file *file)
1442 {
1443     struct vhost_vdpa *v = dev->opaque;
1444     int vdpa_idx = file->index - dev->vq_index;
1445 
1446     if (v->shadow_vqs_enabled) {
1447         VhostShadowVirtqueue *svq = g_ptr_array_index(v->shadow_vqs, vdpa_idx);
1448         vhost_svq_set_svq_kick_fd(svq, file->fd);
1449         return 0;
1450     } else {
1451         return vhost_vdpa_set_vring_dev_kick(dev, file);
1452     }
1453 }
1454 
1455 static int vhost_vdpa_set_vring_call(struct vhost_dev *dev,
1456                                        struct vhost_vring_file *file)
1457 {
1458     struct vhost_vdpa *v = dev->opaque;
1459     int vdpa_idx = file->index - dev->vq_index;
1460     VhostShadowVirtqueue *svq = g_ptr_array_index(v->shadow_vqs, vdpa_idx);
1461 
1462     /* Remember last call fd because we can switch to SVQ anytime. */
1463     vhost_svq_set_svq_call_fd(svq, file->fd);
1464     /*
1465      * When SVQ is transitioning to off, shadow_vqs_enabled has
1466      * not been set back to false yet, but the underlying call fd
1467      * will have to switch back to the guest notifier to signal the
1468      * passthrough virtqueues. In other situations, SVQ's own call
1469      * fd shall be used to signal the device model.
1470      */
1471     if (v->shadow_vqs_enabled &&
1472         v->shared->svq_switching != SVQ_TSTATE_DISABLING) {
1473         return 0;
1474     }
1475 
1476     return vhost_vdpa_set_vring_dev_call(dev, file);
1477 }
1478 
1479 static int vhost_vdpa_get_features(struct vhost_dev *dev,
1480                                      uint64_t *features)
1481 {
1482     int ret = vhost_vdpa_get_dev_features(dev, features);
1483 
1484     if (ret == 0) {
1485         /* Add SVQ logging capabilities */
1486         *features |= BIT_ULL(VHOST_F_LOG_ALL);
1487     }
1488 
1489     return ret;
1490 }
1491 
1492 static int vhost_vdpa_set_owner(struct vhost_dev *dev)
1493 {
1494     if (!vhost_vdpa_first_dev(dev)) {
1495         return 0;
1496     }
1497 
1498     trace_vhost_vdpa_set_owner(dev);
1499     return vhost_vdpa_call(dev, VHOST_SET_OWNER, NULL);
1500 }
1501 
1502 static int vhost_vdpa_vq_get_addr(struct vhost_dev *dev,
1503                     struct vhost_vring_addr *addr, struct vhost_virtqueue *vq)
1504 {
1505     assert(dev->vhost_ops->backend_type == VHOST_BACKEND_TYPE_VDPA);
1506     addr->desc_user_addr = (uint64_t)(unsigned long)vq->desc_phys;
1507     addr->avail_user_addr = (uint64_t)(unsigned long)vq->avail_phys;
1508     addr->used_user_addr = (uint64_t)(unsigned long)vq->used_phys;
1509     trace_vhost_vdpa_vq_get_addr(dev, vq, addr->desc_user_addr,
1510                                  addr->avail_user_addr, addr->used_user_addr);
1511     return 0;
1512 }
1513 
1514 static bool  vhost_vdpa_force_iommu(struct vhost_dev *dev)
1515 {
1516     return true;
1517 }
1518 
1519 const VhostOps vdpa_ops = {
1520         .backend_type = VHOST_BACKEND_TYPE_VDPA,
1521         .vhost_backend_init = vhost_vdpa_init,
1522         .vhost_backend_cleanup = vhost_vdpa_cleanup,
1523         .vhost_set_log_base = vhost_vdpa_set_log_base,
1524         .vhost_set_vring_addr = vhost_vdpa_set_vring_addr,
1525         .vhost_set_vring_num = vhost_vdpa_set_vring_num,
1526         .vhost_set_vring_base = vhost_vdpa_set_vring_base,
1527         .vhost_get_vring_base = vhost_vdpa_get_vring_base,
1528         .vhost_set_vring_kick = vhost_vdpa_set_vring_kick,
1529         .vhost_set_vring_call = vhost_vdpa_set_vring_call,
1530         .vhost_get_features = vhost_vdpa_get_features,
1531         .vhost_set_backend_cap = vhost_vdpa_set_backend_cap,
1532         .vhost_set_owner = vhost_vdpa_set_owner,
1533         .vhost_set_vring_endian = NULL,
1534         .vhost_backend_memslots_limit = vhost_vdpa_memslots_limit,
1535         .vhost_set_mem_table = vhost_vdpa_set_mem_table,
1536         .vhost_set_features = vhost_vdpa_set_features,
1537         .vhost_reset_device = vhost_vdpa_reset_device,
1538         .vhost_get_vq_index = vhost_vdpa_get_vq_index,
1539         .vhost_get_config  = vhost_vdpa_get_config,
1540         .vhost_set_config = vhost_vdpa_set_config,
1541         .vhost_requires_shm_log = NULL,
1542         .vhost_migration_done = NULL,
1543         .vhost_net_set_mtu = NULL,
1544         .vhost_set_iotlb_callback = NULL,
1545         .vhost_send_device_iotlb_msg = NULL,
1546         .vhost_dev_start = vhost_vdpa_dev_start,
1547         .vhost_get_device_id = vhost_vdpa_get_device_id,
1548         .vhost_vq_get_addr = vhost_vdpa_vq_get_addr,
1549         .vhost_force_iommu = vhost_vdpa_force_iommu,
1550         .vhost_set_config_call = vhost_vdpa_set_config_call,
1551         .vhost_reset_status = vhost_vdpa_reset_status,
1552 };
1553