xref: /openbmc/qemu/hw/virtio/vhost.c (revision 19ac7b29)
1 /*
2  * vhost support
3  *
4  * Copyright Red Hat, Inc. 2010
5  *
6  * Authors:
7  *  Michael S. Tsirkin <mst@redhat.com>
8  *
9  * This work is licensed under the terms of the GNU GPL, version 2.  See
10  * the COPYING file in the top-level directory.
11  *
12  * Contributions after 2012-01-13 are licensed under the terms of the
13  * GNU GPL, version 2 or (at your option) any later version.
14  */
15 
16 #include "qemu/osdep.h"
17 #include "qapi/error.h"
18 #include "hw/virtio/vhost.h"
19 #include "qemu/atomic.h"
20 #include "qemu/range.h"
21 #include "qemu/error-report.h"
22 #include "qemu/memfd.h"
23 #include "qemu/log.h"
24 #include "standard-headers/linux/vhost_types.h"
25 #include "hw/virtio/virtio-bus.h"
26 #include "migration/blocker.h"
27 #include "migration/qemu-file-types.h"
28 #include "sysemu/dma.h"
29 #include "trace.h"
30 
31 /* enabled until disconnected backend stabilizes */
32 #define _VHOST_DEBUG 1
33 
34 #ifdef _VHOST_DEBUG
35 #define VHOST_OPS_DEBUG(retval, fmt, ...) \
36     do { \
37         error_report(fmt ": %s (%d)", ## __VA_ARGS__, \
38                      strerror(-retval), -retval); \
39     } while (0)
40 #else
41 #define VHOST_OPS_DEBUG(retval, fmt, ...) \
42     do { } while (0)
43 #endif
44 
45 static struct vhost_log *vhost_log;
46 static struct vhost_log *vhost_log_shm;
47 
48 static unsigned int used_memslots;
49 static QLIST_HEAD(, vhost_dev) vhost_devices =
50     QLIST_HEAD_INITIALIZER(vhost_devices);
51 
52 bool vhost_has_free_slot(void)
53 {
54     unsigned int slots_limit = ~0U;
55     struct vhost_dev *hdev;
56 
57     QLIST_FOREACH(hdev, &vhost_devices, entry) {
58         unsigned int r = hdev->vhost_ops->vhost_backend_memslots_limit(hdev);
59         slots_limit = MIN(slots_limit, r);
60     }
61     return slots_limit > used_memslots;
62 }
63 
64 static void vhost_dev_sync_region(struct vhost_dev *dev,
65                                   MemoryRegionSection *section,
66                                   uint64_t mfirst, uint64_t mlast,
67                                   uint64_t rfirst, uint64_t rlast)
68 {
69     vhost_log_chunk_t *log = dev->log->log;
70 
71     uint64_t start = MAX(mfirst, rfirst);
72     uint64_t end = MIN(mlast, rlast);
73     vhost_log_chunk_t *from = log + start / VHOST_LOG_CHUNK;
74     vhost_log_chunk_t *to = log + end / VHOST_LOG_CHUNK + 1;
75     uint64_t addr = QEMU_ALIGN_DOWN(start, VHOST_LOG_CHUNK);
76 
77     if (end < start) {
78         return;
79     }
80     assert(end / VHOST_LOG_CHUNK < dev->log_size);
81     assert(start / VHOST_LOG_CHUNK < dev->log_size);
82 
83     for (;from < to; ++from) {
84         vhost_log_chunk_t log;
85         /* We first check with non-atomic: much cheaper,
86          * and we expect non-dirty to be the common case. */
87         if (!*from) {
88             addr += VHOST_LOG_CHUNK;
89             continue;
90         }
91         /* Data must be read atomically. We don't really need barrier semantics
92          * but it's easier to use atomic_* than roll our own. */
93         log = qatomic_xchg(from, 0);
94         while (log) {
95             int bit = ctzl(log);
96             hwaddr page_addr;
97             hwaddr section_offset;
98             hwaddr mr_offset;
99             page_addr = addr + bit * VHOST_LOG_PAGE;
100             section_offset = page_addr - section->offset_within_address_space;
101             mr_offset = section_offset + section->offset_within_region;
102             memory_region_set_dirty(section->mr, mr_offset, VHOST_LOG_PAGE);
103             log &= ~(0x1ull << bit);
104         }
105         addr += VHOST_LOG_CHUNK;
106     }
107 }
108 
109 bool vhost_dev_has_iommu(struct vhost_dev *dev)
110 {
111     VirtIODevice *vdev = dev->vdev;
112 
113     /*
114      * For vhost, VIRTIO_F_IOMMU_PLATFORM means the backend support
115      * incremental memory mapping API via IOTLB API. For platform that
116      * does not have IOMMU, there's no need to enable this feature
117      * which may cause unnecessary IOTLB miss/update transactions.
118      */
119     if (vdev) {
120         return virtio_bus_device_iommu_enabled(vdev) &&
121             virtio_host_has_feature(vdev, VIRTIO_F_IOMMU_PLATFORM);
122     } else {
123         return false;
124     }
125 }
126 
127 static int vhost_sync_dirty_bitmap(struct vhost_dev *dev,
128                                    MemoryRegionSection *section,
129                                    hwaddr first,
130                                    hwaddr last)
131 {
132     int i;
133     hwaddr start_addr;
134     hwaddr end_addr;
135 
136     if (!dev->log_enabled || !dev->started) {
137         return 0;
138     }
139     start_addr = section->offset_within_address_space;
140     end_addr = range_get_last(start_addr, int128_get64(section->size));
141     start_addr = MAX(first, start_addr);
142     end_addr = MIN(last, end_addr);
143 
144     for (i = 0; i < dev->mem->nregions; ++i) {
145         struct vhost_memory_region *reg = dev->mem->regions + i;
146         vhost_dev_sync_region(dev, section, start_addr, end_addr,
147                               reg->guest_phys_addr,
148                               range_get_last(reg->guest_phys_addr,
149                                              reg->memory_size));
150     }
151     for (i = 0; i < dev->nvqs; ++i) {
152         struct vhost_virtqueue *vq = dev->vqs + i;
153 
154         if (!vq->used_phys && !vq->used_size) {
155             continue;
156         }
157 
158         if (vhost_dev_has_iommu(dev)) {
159             IOMMUTLBEntry iotlb;
160             hwaddr used_phys = vq->used_phys, used_size = vq->used_size;
161             hwaddr phys, s, offset;
162 
163             while (used_size) {
164                 rcu_read_lock();
165                 iotlb = address_space_get_iotlb_entry(dev->vdev->dma_as,
166                                                       used_phys,
167                                                       true,
168                                                       MEMTXATTRS_UNSPECIFIED);
169                 rcu_read_unlock();
170 
171                 if (!iotlb.target_as) {
172                     qemu_log_mask(LOG_GUEST_ERROR, "translation "
173                                   "failure for used_iova %"PRIx64"\n",
174                                   used_phys);
175                     return -EINVAL;
176                 }
177 
178                 offset = used_phys & iotlb.addr_mask;
179                 phys = iotlb.translated_addr + offset;
180 
181                 /*
182                  * Distance from start of used ring until last byte of
183                  * IOMMU page.
184                  */
185                 s = iotlb.addr_mask - offset;
186                 /*
187                  * Size of used ring, or of the part of it until end
188                  * of IOMMU page. To avoid zero result, do the adding
189                  * outside of MIN().
190                  */
191                 s = MIN(s, used_size - 1) + 1;
192 
193                 vhost_dev_sync_region(dev, section, start_addr, end_addr, phys,
194                                       range_get_last(phys, s));
195                 used_size -= s;
196                 used_phys += s;
197             }
198         } else {
199             vhost_dev_sync_region(dev, section, start_addr,
200                                   end_addr, vq->used_phys,
201                                   range_get_last(vq->used_phys, vq->used_size));
202         }
203     }
204     return 0;
205 }
206 
207 static void vhost_log_sync(MemoryListener *listener,
208                           MemoryRegionSection *section)
209 {
210     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
211                                          memory_listener);
212     vhost_sync_dirty_bitmap(dev, section, 0x0, ~0x0ULL);
213 }
214 
215 static void vhost_log_sync_range(struct vhost_dev *dev,
216                                  hwaddr first, hwaddr last)
217 {
218     int i;
219     /* FIXME: this is N^2 in number of sections */
220     for (i = 0; i < dev->n_mem_sections; ++i) {
221         MemoryRegionSection *section = &dev->mem_sections[i];
222         vhost_sync_dirty_bitmap(dev, section, first, last);
223     }
224 }
225 
226 static uint64_t vhost_get_log_size(struct vhost_dev *dev)
227 {
228     uint64_t log_size = 0;
229     int i;
230     for (i = 0; i < dev->mem->nregions; ++i) {
231         struct vhost_memory_region *reg = dev->mem->regions + i;
232         uint64_t last = range_get_last(reg->guest_phys_addr,
233                                        reg->memory_size);
234         log_size = MAX(log_size, last / VHOST_LOG_CHUNK + 1);
235     }
236     return log_size;
237 }
238 
239 static int vhost_set_backend_type(struct vhost_dev *dev,
240                                   VhostBackendType backend_type)
241 {
242     int r = 0;
243 
244     switch (backend_type) {
245 #ifdef CONFIG_VHOST_KERNEL
246     case VHOST_BACKEND_TYPE_KERNEL:
247         dev->vhost_ops = &kernel_ops;
248         break;
249 #endif
250 #ifdef CONFIG_VHOST_USER
251     case VHOST_BACKEND_TYPE_USER:
252         dev->vhost_ops = &user_ops;
253         break;
254 #endif
255 #ifdef CONFIG_VHOST_VDPA
256     case VHOST_BACKEND_TYPE_VDPA:
257         dev->vhost_ops = &vdpa_ops;
258         break;
259 #endif
260     default:
261         error_report("Unknown vhost backend type");
262         r = -1;
263     }
264 
265     return r;
266 }
267 
268 static struct vhost_log *vhost_log_alloc(uint64_t size, bool share)
269 {
270     Error *err = NULL;
271     struct vhost_log *log;
272     uint64_t logsize = size * sizeof(*(log->log));
273     int fd = -1;
274 
275     log = g_new0(struct vhost_log, 1);
276     if (share) {
277         log->log = qemu_memfd_alloc("vhost-log", logsize,
278                                     F_SEAL_GROW | F_SEAL_SHRINK | F_SEAL_SEAL,
279                                     &fd, &err);
280         if (err) {
281             error_report_err(err);
282             g_free(log);
283             return NULL;
284         }
285         memset(log->log, 0, logsize);
286     } else {
287         log->log = g_malloc0(logsize);
288     }
289 
290     log->size = size;
291     log->refcnt = 1;
292     log->fd = fd;
293 
294     return log;
295 }
296 
297 static struct vhost_log *vhost_log_get(uint64_t size, bool share)
298 {
299     struct vhost_log *log = share ? vhost_log_shm : vhost_log;
300 
301     if (!log || log->size != size) {
302         log = vhost_log_alloc(size, share);
303         if (share) {
304             vhost_log_shm = log;
305         } else {
306             vhost_log = log;
307         }
308     } else {
309         ++log->refcnt;
310     }
311 
312     return log;
313 }
314 
315 static void vhost_log_put(struct vhost_dev *dev, bool sync)
316 {
317     struct vhost_log *log = dev->log;
318 
319     if (!log) {
320         return;
321     }
322 
323     --log->refcnt;
324     if (log->refcnt == 0) {
325         /* Sync only the range covered by the old log */
326         if (dev->log_size && sync) {
327             vhost_log_sync_range(dev, 0, dev->log_size * VHOST_LOG_CHUNK - 1);
328         }
329 
330         if (vhost_log == log) {
331             g_free(log->log);
332             vhost_log = NULL;
333         } else if (vhost_log_shm == log) {
334             qemu_memfd_free(log->log, log->size * sizeof(*(log->log)),
335                             log->fd);
336             vhost_log_shm = NULL;
337         }
338 
339         g_free(log);
340     }
341 
342     dev->log = NULL;
343     dev->log_size = 0;
344 }
345 
346 static bool vhost_dev_log_is_shared(struct vhost_dev *dev)
347 {
348     return dev->vhost_ops->vhost_requires_shm_log &&
349            dev->vhost_ops->vhost_requires_shm_log(dev);
350 }
351 
352 static inline void vhost_dev_log_resize(struct vhost_dev *dev, uint64_t size)
353 {
354     struct vhost_log *log = vhost_log_get(size, vhost_dev_log_is_shared(dev));
355     uint64_t log_base = (uintptr_t)log->log;
356     int r;
357 
358     /* inform backend of log switching, this must be done before
359        releasing the current log, to ensure no logging is lost */
360     r = dev->vhost_ops->vhost_set_log_base(dev, log_base, log);
361     if (r < 0) {
362         VHOST_OPS_DEBUG(r, "vhost_set_log_base failed");
363     }
364 
365     vhost_log_put(dev, true);
366     dev->log = log;
367     dev->log_size = size;
368 }
369 
370 static void *vhost_memory_map(struct vhost_dev *dev, hwaddr addr,
371                               hwaddr *plen, bool is_write)
372 {
373     if (!vhost_dev_has_iommu(dev)) {
374         return cpu_physical_memory_map(addr, plen, is_write);
375     } else {
376         return (void *)(uintptr_t)addr;
377     }
378 }
379 
380 static void vhost_memory_unmap(struct vhost_dev *dev, void *buffer,
381                                hwaddr len, int is_write,
382                                hwaddr access_len)
383 {
384     if (!vhost_dev_has_iommu(dev)) {
385         cpu_physical_memory_unmap(buffer, len, is_write, access_len);
386     }
387 }
388 
389 static int vhost_verify_ring_part_mapping(void *ring_hva,
390                                           uint64_t ring_gpa,
391                                           uint64_t ring_size,
392                                           void *reg_hva,
393                                           uint64_t reg_gpa,
394                                           uint64_t reg_size)
395 {
396     uint64_t hva_ring_offset;
397     uint64_t ring_last = range_get_last(ring_gpa, ring_size);
398     uint64_t reg_last = range_get_last(reg_gpa, reg_size);
399 
400     if (ring_last < reg_gpa || ring_gpa > reg_last) {
401         return 0;
402     }
403     /* check that whole ring's is mapped */
404     if (ring_last > reg_last) {
405         return -ENOMEM;
406     }
407     /* check that ring's MemoryRegion wasn't replaced */
408     hva_ring_offset = ring_gpa - reg_gpa;
409     if (ring_hva != reg_hva + hva_ring_offset) {
410         return -EBUSY;
411     }
412 
413     return 0;
414 }
415 
416 static int vhost_verify_ring_mappings(struct vhost_dev *dev,
417                                       void *reg_hva,
418                                       uint64_t reg_gpa,
419                                       uint64_t reg_size)
420 {
421     int i, j;
422     int r = 0;
423     const char *part_name[] = {
424         "descriptor table",
425         "available ring",
426         "used ring"
427     };
428 
429     if (vhost_dev_has_iommu(dev)) {
430         return 0;
431     }
432 
433     for (i = 0; i < dev->nvqs; ++i) {
434         struct vhost_virtqueue *vq = dev->vqs + i;
435 
436         if (vq->desc_phys == 0) {
437             continue;
438         }
439 
440         j = 0;
441         r = vhost_verify_ring_part_mapping(
442                 vq->desc, vq->desc_phys, vq->desc_size,
443                 reg_hva, reg_gpa, reg_size);
444         if (r) {
445             break;
446         }
447 
448         j++;
449         r = vhost_verify_ring_part_mapping(
450                 vq->avail, vq->avail_phys, vq->avail_size,
451                 reg_hva, reg_gpa, reg_size);
452         if (r) {
453             break;
454         }
455 
456         j++;
457         r = vhost_verify_ring_part_mapping(
458                 vq->used, vq->used_phys, vq->used_size,
459                 reg_hva, reg_gpa, reg_size);
460         if (r) {
461             break;
462         }
463     }
464 
465     if (r == -ENOMEM) {
466         error_report("Unable to map %s for ring %d", part_name[j], i);
467     } else if (r == -EBUSY) {
468         error_report("%s relocated for ring %d", part_name[j], i);
469     }
470     return r;
471 }
472 
473 /*
474  * vhost_section: identify sections needed for vhost access
475  *
476  * We only care about RAM sections here (where virtqueue and guest
477  * internals accessed by virtio might live). If we find one we still
478  * allow the backend to potentially filter it out of our list.
479  */
480 static bool vhost_section(struct vhost_dev *dev, MemoryRegionSection *section)
481 {
482     MemoryRegion *mr = section->mr;
483 
484     if (memory_region_is_ram(mr) && !memory_region_is_rom(mr)) {
485         uint8_t dirty_mask = memory_region_get_dirty_log_mask(mr);
486         uint8_t handled_dirty;
487 
488         /*
489          * Kernel based vhost doesn't handle any block which is doing
490          * dirty-tracking other than migration for which it has
491          * specific logging support. However for TCG the kernel never
492          * gets involved anyway so we can also ignore it's
493          * self-modiying code detection flags. However a vhost-user
494          * client could still confuse a TCG guest if it re-writes
495          * executable memory that has already been translated.
496          */
497         handled_dirty = (1 << DIRTY_MEMORY_MIGRATION) |
498             (1 << DIRTY_MEMORY_CODE);
499 
500         if (dirty_mask & ~handled_dirty) {
501             trace_vhost_reject_section(mr->name, 1);
502             return false;
503         }
504 
505         if (dev->vhost_ops->vhost_backend_mem_section_filter &&
506             !dev->vhost_ops->vhost_backend_mem_section_filter(dev, section)) {
507             trace_vhost_reject_section(mr->name, 2);
508             return false;
509         }
510 
511         trace_vhost_section(mr->name);
512         return true;
513     } else {
514         trace_vhost_reject_section(mr->name, 3);
515         return false;
516     }
517 }
518 
519 static void vhost_begin(MemoryListener *listener)
520 {
521     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
522                                          memory_listener);
523     dev->tmp_sections = NULL;
524     dev->n_tmp_sections = 0;
525 }
526 
527 static void vhost_commit(MemoryListener *listener)
528 {
529     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
530                                          memory_listener);
531     MemoryRegionSection *old_sections;
532     int n_old_sections;
533     uint64_t log_size;
534     size_t regions_size;
535     int r;
536     int i;
537     bool changed = false;
538 
539     /* Note we can be called before the device is started, but then
540      * starting the device calls set_mem_table, so we need to have
541      * built the data structures.
542      */
543     old_sections = dev->mem_sections;
544     n_old_sections = dev->n_mem_sections;
545     dev->mem_sections = dev->tmp_sections;
546     dev->n_mem_sections = dev->n_tmp_sections;
547 
548     if (dev->n_mem_sections != n_old_sections) {
549         changed = true;
550     } else {
551         /* Same size, lets check the contents */
552         for (int i = 0; i < n_old_sections; i++) {
553             if (!MemoryRegionSection_eq(&old_sections[i],
554                                         &dev->mem_sections[i])) {
555                 changed = true;
556                 break;
557             }
558         }
559     }
560 
561     trace_vhost_commit(dev->started, changed);
562     if (!changed) {
563         goto out;
564     }
565 
566     /* Rebuild the regions list from the new sections list */
567     regions_size = offsetof(struct vhost_memory, regions) +
568                        dev->n_mem_sections * sizeof dev->mem->regions[0];
569     dev->mem = g_realloc(dev->mem, regions_size);
570     dev->mem->nregions = dev->n_mem_sections;
571     used_memslots = dev->mem->nregions;
572     for (i = 0; i < dev->n_mem_sections; i++) {
573         struct vhost_memory_region *cur_vmr = dev->mem->regions + i;
574         struct MemoryRegionSection *mrs = dev->mem_sections + i;
575 
576         cur_vmr->guest_phys_addr = mrs->offset_within_address_space;
577         cur_vmr->memory_size     = int128_get64(mrs->size);
578         cur_vmr->userspace_addr  =
579             (uintptr_t)memory_region_get_ram_ptr(mrs->mr) +
580             mrs->offset_within_region;
581         cur_vmr->flags_padding   = 0;
582     }
583 
584     if (!dev->started) {
585         goto out;
586     }
587 
588     for (i = 0; i < dev->mem->nregions; i++) {
589         if (vhost_verify_ring_mappings(dev,
590                        (void *)(uintptr_t)dev->mem->regions[i].userspace_addr,
591                        dev->mem->regions[i].guest_phys_addr,
592                        dev->mem->regions[i].memory_size)) {
593             error_report("Verify ring failure on region %d", i);
594             abort();
595         }
596     }
597 
598     if (!dev->log_enabled) {
599         r = dev->vhost_ops->vhost_set_mem_table(dev, dev->mem);
600         if (r < 0) {
601             VHOST_OPS_DEBUG(r, "vhost_set_mem_table failed");
602         }
603         goto out;
604     }
605     log_size = vhost_get_log_size(dev);
606     /* We allocate an extra 4K bytes to log,
607      * to reduce the * number of reallocations. */
608 #define VHOST_LOG_BUFFER (0x1000 / sizeof *dev->log)
609     /* To log more, must increase log size before table update. */
610     if (dev->log_size < log_size) {
611         vhost_dev_log_resize(dev, log_size + VHOST_LOG_BUFFER);
612     }
613     r = dev->vhost_ops->vhost_set_mem_table(dev, dev->mem);
614     if (r < 0) {
615         VHOST_OPS_DEBUG(r, "vhost_set_mem_table failed");
616     }
617     /* To log less, can only decrease log size after table update. */
618     if (dev->log_size > log_size + VHOST_LOG_BUFFER) {
619         vhost_dev_log_resize(dev, log_size);
620     }
621 
622 out:
623     /* Deref the old list of sections, this must happen _after_ the
624      * vhost_set_mem_table to ensure the client isn't still using the
625      * section we're about to unref.
626      */
627     while (n_old_sections--) {
628         memory_region_unref(old_sections[n_old_sections].mr);
629     }
630     g_free(old_sections);
631     return;
632 }
633 
634 /* Adds the section data to the tmp_section structure.
635  * It relies on the listener calling us in memory address order
636  * and for each region (via the _add and _nop methods) to
637  * join neighbours.
638  */
639 static void vhost_region_add_section(struct vhost_dev *dev,
640                                      MemoryRegionSection *section)
641 {
642     bool need_add = true;
643     uint64_t mrs_size = int128_get64(section->size);
644     uint64_t mrs_gpa = section->offset_within_address_space;
645     uintptr_t mrs_host = (uintptr_t)memory_region_get_ram_ptr(section->mr) +
646                          section->offset_within_region;
647     RAMBlock *mrs_rb = section->mr->ram_block;
648 
649     trace_vhost_region_add_section(section->mr->name, mrs_gpa, mrs_size,
650                                    mrs_host);
651 
652     if (dev->vhost_ops->backend_type == VHOST_BACKEND_TYPE_USER) {
653         /* Round the section to it's page size */
654         /* First align the start down to a page boundary */
655         size_t mrs_page = qemu_ram_pagesize(mrs_rb);
656         uint64_t alignage = mrs_host & (mrs_page - 1);
657         if (alignage) {
658             mrs_host -= alignage;
659             mrs_size += alignage;
660             mrs_gpa  -= alignage;
661         }
662         /* Now align the size up to a page boundary */
663         alignage = mrs_size & (mrs_page - 1);
664         if (alignage) {
665             mrs_size += mrs_page - alignage;
666         }
667         trace_vhost_region_add_section_aligned(section->mr->name, mrs_gpa,
668                                                mrs_size, mrs_host);
669     }
670 
671     if (dev->n_tmp_sections) {
672         /* Since we already have at least one section, lets see if
673          * this extends it; since we're scanning in order, we only
674          * have to look at the last one, and the FlatView that calls
675          * us shouldn't have overlaps.
676          */
677         MemoryRegionSection *prev_sec = dev->tmp_sections +
678                                                (dev->n_tmp_sections - 1);
679         uint64_t prev_gpa_start = prev_sec->offset_within_address_space;
680         uint64_t prev_size = int128_get64(prev_sec->size);
681         uint64_t prev_gpa_end   = range_get_last(prev_gpa_start, prev_size);
682         uint64_t prev_host_start =
683                         (uintptr_t)memory_region_get_ram_ptr(prev_sec->mr) +
684                         prev_sec->offset_within_region;
685         uint64_t prev_host_end   = range_get_last(prev_host_start, prev_size);
686 
687         if (mrs_gpa <= (prev_gpa_end + 1)) {
688             /* OK, looks like overlapping/intersecting - it's possible that
689              * the rounding to page sizes has made them overlap, but they should
690              * match up in the same RAMBlock if they do.
691              */
692             if (mrs_gpa < prev_gpa_start) {
693                 error_report("%s:Section '%s' rounded to %"PRIx64
694                              " prior to previous '%s' %"PRIx64,
695                              __func__, section->mr->name, mrs_gpa,
696                              prev_sec->mr->name, prev_gpa_start);
697                 /* A way to cleanly fail here would be better */
698                 return;
699             }
700             /* Offset from the start of the previous GPA to this GPA */
701             size_t offset = mrs_gpa - prev_gpa_start;
702 
703             if (prev_host_start + offset == mrs_host &&
704                 section->mr == prev_sec->mr &&
705                 (!dev->vhost_ops->vhost_backend_can_merge ||
706                  dev->vhost_ops->vhost_backend_can_merge(dev,
707                     mrs_host, mrs_size,
708                     prev_host_start, prev_size))) {
709                 uint64_t max_end = MAX(prev_host_end, mrs_host + mrs_size);
710                 need_add = false;
711                 prev_sec->offset_within_address_space =
712                     MIN(prev_gpa_start, mrs_gpa);
713                 prev_sec->offset_within_region =
714                     MIN(prev_host_start, mrs_host) -
715                     (uintptr_t)memory_region_get_ram_ptr(prev_sec->mr);
716                 prev_sec->size = int128_make64(max_end - MIN(prev_host_start,
717                                                mrs_host));
718                 trace_vhost_region_add_section_merge(section->mr->name,
719                                         int128_get64(prev_sec->size),
720                                         prev_sec->offset_within_address_space,
721                                         prev_sec->offset_within_region);
722             } else {
723                 /* adjoining regions are fine, but overlapping ones with
724                  * different blocks/offsets shouldn't happen
725                  */
726                 if (mrs_gpa != prev_gpa_end + 1) {
727                     error_report("%s: Overlapping but not coherent sections "
728                                  "at %"PRIx64,
729                                  __func__, mrs_gpa);
730                     return;
731                 }
732             }
733         }
734     }
735 
736     if (need_add) {
737         ++dev->n_tmp_sections;
738         dev->tmp_sections = g_renew(MemoryRegionSection, dev->tmp_sections,
739                                     dev->n_tmp_sections);
740         dev->tmp_sections[dev->n_tmp_sections - 1] = *section;
741         /* The flatview isn't stable and we don't use it, making it NULL
742          * means we can memcmp the list.
743          */
744         dev->tmp_sections[dev->n_tmp_sections - 1].fv = NULL;
745         memory_region_ref(section->mr);
746     }
747 }
748 
749 /* Used for both add and nop callbacks */
750 static void vhost_region_addnop(MemoryListener *listener,
751                                 MemoryRegionSection *section)
752 {
753     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
754                                          memory_listener);
755 
756     if (!vhost_section(dev, section)) {
757         return;
758     }
759     vhost_region_add_section(dev, section);
760 }
761 
762 static void vhost_iommu_unmap_notify(IOMMUNotifier *n, IOMMUTLBEntry *iotlb)
763 {
764     struct vhost_iommu *iommu = container_of(n, struct vhost_iommu, n);
765     struct vhost_dev *hdev = iommu->hdev;
766     hwaddr iova = iotlb->iova + iommu->iommu_offset;
767 
768     if (vhost_backend_invalidate_device_iotlb(hdev, iova,
769                                               iotlb->addr_mask + 1)) {
770         error_report("Fail to invalidate device iotlb");
771     }
772 }
773 
774 static void vhost_iommu_region_add(MemoryListener *listener,
775                                    MemoryRegionSection *section)
776 {
777     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
778                                          iommu_listener);
779     struct vhost_iommu *iommu;
780     Int128 end;
781     int iommu_idx;
782     IOMMUMemoryRegion *iommu_mr;
783 
784     if (!memory_region_is_iommu(section->mr)) {
785         return;
786     }
787 
788     iommu_mr = IOMMU_MEMORY_REGION(section->mr);
789 
790     iommu = g_malloc0(sizeof(*iommu));
791     end = int128_add(int128_make64(section->offset_within_region),
792                      section->size);
793     end = int128_sub(end, int128_one());
794     iommu_idx = memory_region_iommu_attrs_to_index(iommu_mr,
795                                                    MEMTXATTRS_UNSPECIFIED);
796     iommu_notifier_init(&iommu->n, vhost_iommu_unmap_notify,
797                         dev->vdev->device_iotlb_enabled ?
798                             IOMMU_NOTIFIER_DEVIOTLB_UNMAP :
799                             IOMMU_NOTIFIER_UNMAP,
800                         section->offset_within_region,
801                         int128_get64(end),
802                         iommu_idx);
803     iommu->mr = section->mr;
804     iommu->iommu_offset = section->offset_within_address_space -
805                           section->offset_within_region;
806     iommu->hdev = dev;
807     memory_region_register_iommu_notifier(section->mr, &iommu->n,
808                                           &error_fatal);
809     QLIST_INSERT_HEAD(&dev->iommu_list, iommu, iommu_next);
810     /* TODO: can replay help performance here? */
811 }
812 
813 static void vhost_iommu_region_del(MemoryListener *listener,
814                                    MemoryRegionSection *section)
815 {
816     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
817                                          iommu_listener);
818     struct vhost_iommu *iommu;
819 
820     if (!memory_region_is_iommu(section->mr)) {
821         return;
822     }
823 
824     QLIST_FOREACH(iommu, &dev->iommu_list, iommu_next) {
825         if (iommu->mr == section->mr &&
826             iommu->n.start == section->offset_within_region) {
827             memory_region_unregister_iommu_notifier(iommu->mr,
828                                                     &iommu->n);
829             QLIST_REMOVE(iommu, iommu_next);
830             g_free(iommu);
831             break;
832         }
833     }
834 }
835 
836 void vhost_toggle_device_iotlb(VirtIODevice *vdev)
837 {
838     VirtioDeviceClass *vdc = VIRTIO_DEVICE_GET_CLASS(vdev);
839     struct vhost_dev *dev;
840     struct vhost_iommu *iommu;
841 
842     if (vdev->vhost_started) {
843         dev = vdc->get_vhost(vdev);
844     } else {
845         return;
846     }
847 
848     QLIST_FOREACH(iommu, &dev->iommu_list, iommu_next) {
849         memory_region_unregister_iommu_notifier(iommu->mr, &iommu->n);
850         iommu->n.notifier_flags = vdev->device_iotlb_enabled ?
851                 IOMMU_NOTIFIER_DEVIOTLB_UNMAP : IOMMU_NOTIFIER_UNMAP;
852         memory_region_register_iommu_notifier(iommu->mr, &iommu->n,
853                                               &error_fatal);
854     }
855 }
856 
857 static int vhost_virtqueue_set_addr(struct vhost_dev *dev,
858                                     struct vhost_virtqueue *vq,
859                                     unsigned idx, bool enable_log)
860 {
861     struct vhost_vring_addr addr;
862     int r;
863     memset(&addr, 0, sizeof(struct vhost_vring_addr));
864 
865     if (dev->vhost_ops->vhost_vq_get_addr) {
866         r = dev->vhost_ops->vhost_vq_get_addr(dev, &addr, vq);
867         if (r < 0) {
868             VHOST_OPS_DEBUG(r, "vhost_vq_get_addr failed");
869             return r;
870         }
871     } else {
872         addr.desc_user_addr = (uint64_t)(unsigned long)vq->desc;
873         addr.avail_user_addr = (uint64_t)(unsigned long)vq->avail;
874         addr.used_user_addr = (uint64_t)(unsigned long)vq->used;
875     }
876     addr.index = idx;
877     addr.log_guest_addr = vq->used_phys;
878     addr.flags = enable_log ? (1 << VHOST_VRING_F_LOG) : 0;
879     r = dev->vhost_ops->vhost_set_vring_addr(dev, &addr);
880     if (r < 0) {
881         VHOST_OPS_DEBUG(r, "vhost_set_vring_addr failed");
882     }
883     return r;
884 }
885 
886 static int vhost_dev_set_features(struct vhost_dev *dev,
887                                   bool enable_log)
888 {
889     uint64_t features = dev->acked_features;
890     int r;
891     if (enable_log) {
892         features |= 0x1ULL << VHOST_F_LOG_ALL;
893     }
894     if (!vhost_dev_has_iommu(dev)) {
895         features &= ~(0x1ULL << VIRTIO_F_IOMMU_PLATFORM);
896     }
897     if (dev->vhost_ops->vhost_force_iommu) {
898         if (dev->vhost_ops->vhost_force_iommu(dev) == true) {
899             features |= 0x1ULL << VIRTIO_F_IOMMU_PLATFORM;
900        }
901     }
902     r = dev->vhost_ops->vhost_set_features(dev, features);
903     if (r < 0) {
904         VHOST_OPS_DEBUG(r, "vhost_set_features failed");
905         goto out;
906     }
907     if (dev->vhost_ops->vhost_set_backend_cap) {
908         r = dev->vhost_ops->vhost_set_backend_cap(dev);
909         if (r < 0) {
910             VHOST_OPS_DEBUG(r, "vhost_set_backend_cap failed");
911             goto out;
912         }
913     }
914 
915 out:
916     return r;
917 }
918 
919 static int vhost_dev_set_log(struct vhost_dev *dev, bool enable_log)
920 {
921     int r, i, idx;
922     hwaddr addr;
923 
924     r = vhost_dev_set_features(dev, enable_log);
925     if (r < 0) {
926         goto err_features;
927     }
928     for (i = 0; i < dev->nvqs; ++i) {
929         idx = dev->vhost_ops->vhost_get_vq_index(dev, dev->vq_index + i);
930         addr = virtio_queue_get_desc_addr(dev->vdev, idx);
931         if (!addr) {
932             /*
933              * The queue might not be ready for start. If this
934              * is the case there is no reason to continue the process.
935              * The similar logic is used by the vhost_virtqueue_start()
936              * routine.
937              */
938             continue;
939         }
940         r = vhost_virtqueue_set_addr(dev, dev->vqs + i, idx,
941                                      enable_log);
942         if (r < 0) {
943             goto err_vq;
944         }
945     }
946     return 0;
947 err_vq:
948     for (; i >= 0; --i) {
949         idx = dev->vhost_ops->vhost_get_vq_index(dev, dev->vq_index + i);
950         addr = virtio_queue_get_desc_addr(dev->vdev, idx);
951         if (!addr) {
952             continue;
953         }
954         vhost_virtqueue_set_addr(dev, dev->vqs + i, idx,
955                                  dev->log_enabled);
956     }
957     vhost_dev_set_features(dev, dev->log_enabled);
958 err_features:
959     return r;
960 }
961 
962 static int vhost_migration_log(MemoryListener *listener, bool enable)
963 {
964     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
965                                          memory_listener);
966     int r;
967     if (enable == dev->log_enabled) {
968         return 0;
969     }
970     if (!dev->started) {
971         dev->log_enabled = enable;
972         return 0;
973     }
974 
975     r = 0;
976     if (!enable) {
977         r = vhost_dev_set_log(dev, false);
978         if (r < 0) {
979             goto check_dev_state;
980         }
981         vhost_log_put(dev, false);
982     } else {
983         vhost_dev_log_resize(dev, vhost_get_log_size(dev));
984         r = vhost_dev_set_log(dev, true);
985         if (r < 0) {
986             goto check_dev_state;
987         }
988     }
989 
990 check_dev_state:
991     dev->log_enabled = enable;
992     /*
993      * vhost-user-* devices could change their state during log
994      * initialization due to disconnect. So check dev state after
995      * vhost communication.
996      */
997     if (!dev->started) {
998         /*
999          * Since device is in the stopped state, it is okay for
1000          * migration. Return success.
1001          */
1002         r = 0;
1003     }
1004     if (r) {
1005         /* An error occurred. */
1006         dev->log_enabled = false;
1007     }
1008 
1009     return r;
1010 }
1011 
1012 static void vhost_log_global_start(MemoryListener *listener)
1013 {
1014     int r;
1015 
1016     r = vhost_migration_log(listener, true);
1017     if (r < 0) {
1018         abort();
1019     }
1020 }
1021 
1022 static void vhost_log_global_stop(MemoryListener *listener)
1023 {
1024     int r;
1025 
1026     r = vhost_migration_log(listener, false);
1027     if (r < 0) {
1028         abort();
1029     }
1030 }
1031 
1032 static void vhost_log_start(MemoryListener *listener,
1033                             MemoryRegionSection *section,
1034                             int old, int new)
1035 {
1036     /* FIXME: implement */
1037 }
1038 
1039 static void vhost_log_stop(MemoryListener *listener,
1040                            MemoryRegionSection *section,
1041                            int old, int new)
1042 {
1043     /* FIXME: implement */
1044 }
1045 
1046 /* The vhost driver natively knows how to handle the vrings of non
1047  * cross-endian legacy devices and modern devices. Only legacy devices
1048  * exposed to a bi-endian guest may require the vhost driver to use a
1049  * specific endianness.
1050  */
1051 static inline bool vhost_needs_vring_endian(VirtIODevice *vdev)
1052 {
1053     if (virtio_vdev_has_feature(vdev, VIRTIO_F_VERSION_1)) {
1054         return false;
1055     }
1056 #if HOST_BIG_ENDIAN
1057     return vdev->device_endian == VIRTIO_DEVICE_ENDIAN_LITTLE;
1058 #else
1059     return vdev->device_endian == VIRTIO_DEVICE_ENDIAN_BIG;
1060 #endif
1061 }
1062 
1063 static int vhost_virtqueue_set_vring_endian_legacy(struct vhost_dev *dev,
1064                                                    bool is_big_endian,
1065                                                    int vhost_vq_index)
1066 {
1067     int r;
1068     struct vhost_vring_state s = {
1069         .index = vhost_vq_index,
1070         .num = is_big_endian
1071     };
1072 
1073     r = dev->vhost_ops->vhost_set_vring_endian(dev, &s);
1074     if (r < 0) {
1075         VHOST_OPS_DEBUG(r, "vhost_set_vring_endian failed");
1076     }
1077     return r;
1078 }
1079 
1080 static int vhost_memory_region_lookup(struct vhost_dev *hdev,
1081                                       uint64_t gpa, uint64_t *uaddr,
1082                                       uint64_t *len)
1083 {
1084     int i;
1085 
1086     for (i = 0; i < hdev->mem->nregions; i++) {
1087         struct vhost_memory_region *reg = hdev->mem->regions + i;
1088 
1089         if (gpa >= reg->guest_phys_addr &&
1090             reg->guest_phys_addr + reg->memory_size > gpa) {
1091             *uaddr = reg->userspace_addr + gpa - reg->guest_phys_addr;
1092             *len = reg->guest_phys_addr + reg->memory_size - gpa;
1093             return 0;
1094         }
1095     }
1096 
1097     return -EFAULT;
1098 }
1099 
1100 int vhost_device_iotlb_miss(struct vhost_dev *dev, uint64_t iova, int write)
1101 {
1102     IOMMUTLBEntry iotlb;
1103     uint64_t uaddr, len;
1104     int ret = -EFAULT;
1105 
1106     RCU_READ_LOCK_GUARD();
1107 
1108     trace_vhost_iotlb_miss(dev, 1);
1109 
1110     iotlb = address_space_get_iotlb_entry(dev->vdev->dma_as,
1111                                           iova, write,
1112                                           MEMTXATTRS_UNSPECIFIED);
1113     if (iotlb.target_as != NULL) {
1114         ret = vhost_memory_region_lookup(dev, iotlb.translated_addr,
1115                                          &uaddr, &len);
1116         if (ret) {
1117             trace_vhost_iotlb_miss(dev, 3);
1118             error_report("Fail to lookup the translated address "
1119                          "%"PRIx64, iotlb.translated_addr);
1120             goto out;
1121         }
1122 
1123         len = MIN(iotlb.addr_mask + 1, len);
1124         iova = iova & ~iotlb.addr_mask;
1125 
1126         ret = vhost_backend_update_device_iotlb(dev, iova, uaddr,
1127                                                 len, iotlb.perm);
1128         if (ret) {
1129             trace_vhost_iotlb_miss(dev, 4);
1130             error_report("Fail to update device iotlb");
1131             goto out;
1132         }
1133     }
1134 
1135     trace_vhost_iotlb_miss(dev, 2);
1136 
1137 out:
1138     return ret;
1139 }
1140 
1141 int vhost_virtqueue_start(struct vhost_dev *dev,
1142                           struct VirtIODevice *vdev,
1143                           struct vhost_virtqueue *vq,
1144                           unsigned idx)
1145 {
1146     BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev)));
1147     VirtioBusState *vbus = VIRTIO_BUS(qbus);
1148     VirtioBusClass *k = VIRTIO_BUS_GET_CLASS(vbus);
1149     hwaddr s, l, a;
1150     int r;
1151     int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, idx);
1152     struct vhost_vring_file file = {
1153         .index = vhost_vq_index
1154     };
1155     struct vhost_vring_state state = {
1156         .index = vhost_vq_index
1157     };
1158     struct VirtQueue *vvq = virtio_get_queue(vdev, idx);
1159 
1160     a = virtio_queue_get_desc_addr(vdev, idx);
1161     if (a == 0) {
1162         /* Queue might not be ready for start */
1163         return 0;
1164     }
1165 
1166     vq->num = state.num = virtio_queue_get_num(vdev, idx);
1167     r = dev->vhost_ops->vhost_set_vring_num(dev, &state);
1168     if (r) {
1169         VHOST_OPS_DEBUG(r, "vhost_set_vring_num failed");
1170         return r;
1171     }
1172 
1173     state.num = virtio_queue_get_last_avail_idx(vdev, idx);
1174     r = dev->vhost_ops->vhost_set_vring_base(dev, &state);
1175     if (r) {
1176         VHOST_OPS_DEBUG(r, "vhost_set_vring_base failed");
1177         return r;
1178     }
1179 
1180     if (vhost_needs_vring_endian(vdev)) {
1181         r = vhost_virtqueue_set_vring_endian_legacy(dev,
1182                                                     virtio_is_big_endian(vdev),
1183                                                     vhost_vq_index);
1184         if (r) {
1185             return r;
1186         }
1187     }
1188 
1189     vq->desc_size = s = l = virtio_queue_get_desc_size(vdev, idx);
1190     vq->desc_phys = a;
1191     vq->desc = vhost_memory_map(dev, a, &l, false);
1192     if (!vq->desc || l != s) {
1193         r = -ENOMEM;
1194         goto fail_alloc_desc;
1195     }
1196     vq->avail_size = s = l = virtio_queue_get_avail_size(vdev, idx);
1197     vq->avail_phys = a = virtio_queue_get_avail_addr(vdev, idx);
1198     vq->avail = vhost_memory_map(dev, a, &l, false);
1199     if (!vq->avail || l != s) {
1200         r = -ENOMEM;
1201         goto fail_alloc_avail;
1202     }
1203     vq->used_size = s = l = virtio_queue_get_used_size(vdev, idx);
1204     vq->used_phys = a = virtio_queue_get_used_addr(vdev, idx);
1205     vq->used = vhost_memory_map(dev, a, &l, true);
1206     if (!vq->used || l != s) {
1207         r = -ENOMEM;
1208         goto fail_alloc_used;
1209     }
1210 
1211     r = vhost_virtqueue_set_addr(dev, vq, vhost_vq_index, dev->log_enabled);
1212     if (r < 0) {
1213         goto fail_alloc;
1214     }
1215 
1216     file.fd = event_notifier_get_fd(virtio_queue_get_host_notifier(vvq));
1217     r = dev->vhost_ops->vhost_set_vring_kick(dev, &file);
1218     if (r) {
1219         VHOST_OPS_DEBUG(r, "vhost_set_vring_kick failed");
1220         goto fail_kick;
1221     }
1222 
1223     /* Clear and discard previous events if any. */
1224     event_notifier_test_and_clear(&vq->masked_notifier);
1225 
1226     /* Init vring in unmasked state, unless guest_notifier_mask
1227      * will do it later.
1228      */
1229     if (!vdev->use_guest_notifier_mask) {
1230         /* TODO: check and handle errors. */
1231         vhost_virtqueue_mask(dev, vdev, idx, false);
1232     }
1233 
1234     if (k->query_guest_notifiers &&
1235         k->query_guest_notifiers(qbus->parent) &&
1236         virtio_queue_vector(vdev, idx) == VIRTIO_NO_VECTOR) {
1237         file.fd = -1;
1238         r = dev->vhost_ops->vhost_set_vring_call(dev, &file);
1239         if (r) {
1240             goto fail_vector;
1241         }
1242     }
1243 
1244     return 0;
1245 
1246 fail_vector:
1247 fail_kick:
1248 fail_alloc:
1249     vhost_memory_unmap(dev, vq->used, virtio_queue_get_used_size(vdev, idx),
1250                        0, 0);
1251 fail_alloc_used:
1252     vhost_memory_unmap(dev, vq->avail, virtio_queue_get_avail_size(vdev, idx),
1253                        0, 0);
1254 fail_alloc_avail:
1255     vhost_memory_unmap(dev, vq->desc, virtio_queue_get_desc_size(vdev, idx),
1256                        0, 0);
1257 fail_alloc_desc:
1258     return r;
1259 }
1260 
1261 void vhost_virtqueue_stop(struct vhost_dev *dev,
1262                           struct VirtIODevice *vdev,
1263                           struct vhost_virtqueue *vq,
1264                           unsigned idx)
1265 {
1266     int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, idx);
1267     struct vhost_vring_state state = {
1268         .index = vhost_vq_index,
1269     };
1270     int r;
1271 
1272     if (virtio_queue_get_desc_addr(vdev, idx) == 0) {
1273         /* Don't stop the virtqueue which might have not been started */
1274         return;
1275     }
1276 
1277     r = dev->vhost_ops->vhost_get_vring_base(dev, &state);
1278     if (r < 0) {
1279         VHOST_OPS_DEBUG(r, "vhost VQ %u ring restore failed: %d", idx, r);
1280         /* Connection to the backend is broken, so let's sync internal
1281          * last avail idx to the device used idx.
1282          */
1283         virtio_queue_restore_last_avail_idx(vdev, idx);
1284     } else {
1285         virtio_queue_set_last_avail_idx(vdev, idx, state.num);
1286     }
1287     virtio_queue_invalidate_signalled_used(vdev, idx);
1288     virtio_queue_update_used_idx(vdev, idx);
1289 
1290     /* In the cross-endian case, we need to reset the vring endianness to
1291      * native as legacy devices expect so by default.
1292      */
1293     if (vhost_needs_vring_endian(vdev)) {
1294         vhost_virtqueue_set_vring_endian_legacy(dev,
1295                                                 !virtio_is_big_endian(vdev),
1296                                                 vhost_vq_index);
1297     }
1298 
1299     vhost_memory_unmap(dev, vq->used, virtio_queue_get_used_size(vdev, idx),
1300                        1, virtio_queue_get_used_size(vdev, idx));
1301     vhost_memory_unmap(dev, vq->avail, virtio_queue_get_avail_size(vdev, idx),
1302                        0, virtio_queue_get_avail_size(vdev, idx));
1303     vhost_memory_unmap(dev, vq->desc, virtio_queue_get_desc_size(vdev, idx),
1304                        0, virtio_queue_get_desc_size(vdev, idx));
1305 }
1306 
1307 static int vhost_virtqueue_set_busyloop_timeout(struct vhost_dev *dev,
1308                                                 int n, uint32_t timeout)
1309 {
1310     int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, n);
1311     struct vhost_vring_state state = {
1312         .index = vhost_vq_index,
1313         .num = timeout,
1314     };
1315     int r;
1316 
1317     if (!dev->vhost_ops->vhost_set_vring_busyloop_timeout) {
1318         return -EINVAL;
1319     }
1320 
1321     r = dev->vhost_ops->vhost_set_vring_busyloop_timeout(dev, &state);
1322     if (r) {
1323         VHOST_OPS_DEBUG(r, "vhost_set_vring_busyloop_timeout failed");
1324         return r;
1325     }
1326 
1327     return 0;
1328 }
1329 
1330 static void vhost_virtqueue_error_notifier(EventNotifier *n)
1331 {
1332     struct vhost_virtqueue *vq = container_of(n, struct vhost_virtqueue,
1333                                               error_notifier);
1334     struct vhost_dev *dev = vq->dev;
1335     int index = vq - dev->vqs;
1336 
1337     if (event_notifier_test_and_clear(n) && dev->vdev) {
1338         VHOST_OPS_DEBUG(-EINVAL,  "vhost vring error in virtqueue %d",
1339                         dev->vq_index + index);
1340     }
1341 }
1342 
1343 static int vhost_virtqueue_init(struct vhost_dev *dev,
1344                                 struct vhost_virtqueue *vq, int n)
1345 {
1346     int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, n);
1347     struct vhost_vring_file file = {
1348         .index = vhost_vq_index,
1349     };
1350     int r = event_notifier_init(&vq->masked_notifier, 0);
1351     if (r < 0) {
1352         return r;
1353     }
1354 
1355     file.fd = event_notifier_get_wfd(&vq->masked_notifier);
1356     r = dev->vhost_ops->vhost_set_vring_call(dev, &file);
1357     if (r) {
1358         VHOST_OPS_DEBUG(r, "vhost_set_vring_call failed");
1359         goto fail_call;
1360     }
1361 
1362     vq->dev = dev;
1363 
1364     if (dev->vhost_ops->vhost_set_vring_err) {
1365         r = event_notifier_init(&vq->error_notifier, 0);
1366         if (r < 0) {
1367             goto fail_call;
1368         }
1369 
1370         file.fd = event_notifier_get_fd(&vq->error_notifier);
1371         r = dev->vhost_ops->vhost_set_vring_err(dev, &file);
1372         if (r) {
1373             VHOST_OPS_DEBUG(r, "vhost_set_vring_err failed");
1374             goto fail_err;
1375         }
1376 
1377         event_notifier_set_handler(&vq->error_notifier,
1378                                    vhost_virtqueue_error_notifier);
1379     }
1380 
1381     return 0;
1382 
1383 fail_err:
1384     event_notifier_cleanup(&vq->error_notifier);
1385 fail_call:
1386     event_notifier_cleanup(&vq->masked_notifier);
1387     return r;
1388 }
1389 
1390 static void vhost_virtqueue_cleanup(struct vhost_virtqueue *vq)
1391 {
1392     event_notifier_cleanup(&vq->masked_notifier);
1393     if (vq->dev->vhost_ops->vhost_set_vring_err) {
1394         event_notifier_set_handler(&vq->error_notifier, NULL);
1395         event_notifier_cleanup(&vq->error_notifier);
1396     }
1397 }
1398 
1399 int vhost_dev_init(struct vhost_dev *hdev, void *opaque,
1400                    VhostBackendType backend_type, uint32_t busyloop_timeout,
1401                    Error **errp)
1402 {
1403     uint64_t features;
1404     int i, r, n_initialized_vqs = 0;
1405 
1406     hdev->vdev = NULL;
1407     hdev->migration_blocker = NULL;
1408 
1409     r = vhost_set_backend_type(hdev, backend_type);
1410     assert(r >= 0);
1411 
1412     r = hdev->vhost_ops->vhost_backend_init(hdev, opaque, errp);
1413     if (r < 0) {
1414         goto fail;
1415     }
1416 
1417     r = hdev->vhost_ops->vhost_set_owner(hdev);
1418     if (r < 0) {
1419         error_setg_errno(errp, -r, "vhost_set_owner failed");
1420         goto fail;
1421     }
1422 
1423     r = hdev->vhost_ops->vhost_get_features(hdev, &features);
1424     if (r < 0) {
1425         error_setg_errno(errp, -r, "vhost_get_features failed");
1426         goto fail;
1427     }
1428 
1429     for (i = 0; i < hdev->nvqs; ++i, ++n_initialized_vqs) {
1430         r = vhost_virtqueue_init(hdev, hdev->vqs + i, hdev->vq_index + i);
1431         if (r < 0) {
1432             error_setg_errno(errp, -r, "Failed to initialize virtqueue %d", i);
1433             goto fail;
1434         }
1435     }
1436 
1437     if (busyloop_timeout) {
1438         for (i = 0; i < hdev->nvqs; ++i) {
1439             r = vhost_virtqueue_set_busyloop_timeout(hdev, hdev->vq_index + i,
1440                                                      busyloop_timeout);
1441             if (r < 0) {
1442                 error_setg_errno(errp, -r, "Failed to set busyloop timeout");
1443                 goto fail_busyloop;
1444             }
1445         }
1446     }
1447 
1448     hdev->features = features;
1449 
1450     hdev->memory_listener = (MemoryListener) {
1451         .name = "vhost",
1452         .begin = vhost_begin,
1453         .commit = vhost_commit,
1454         .region_add = vhost_region_addnop,
1455         .region_nop = vhost_region_addnop,
1456         .log_start = vhost_log_start,
1457         .log_stop = vhost_log_stop,
1458         .log_sync = vhost_log_sync,
1459         .log_global_start = vhost_log_global_start,
1460         .log_global_stop = vhost_log_global_stop,
1461         .priority = MEMORY_LISTENER_PRIORITY_DEV_BACKEND
1462     };
1463 
1464     hdev->iommu_listener = (MemoryListener) {
1465         .name = "vhost-iommu",
1466         .region_add = vhost_iommu_region_add,
1467         .region_del = vhost_iommu_region_del,
1468     };
1469 
1470     if (hdev->migration_blocker == NULL) {
1471         if (!(hdev->features & (0x1ULL << VHOST_F_LOG_ALL))) {
1472             error_setg(&hdev->migration_blocker,
1473                        "Migration disabled: vhost lacks VHOST_F_LOG_ALL feature.");
1474         } else if (vhost_dev_log_is_shared(hdev) && !qemu_memfd_alloc_check()) {
1475             error_setg(&hdev->migration_blocker,
1476                        "Migration disabled: failed to allocate shared memory");
1477         }
1478     }
1479 
1480     if (hdev->migration_blocker != NULL) {
1481         r = migrate_add_blocker(hdev->migration_blocker, errp);
1482         if (r < 0) {
1483             error_free(hdev->migration_blocker);
1484             goto fail_busyloop;
1485         }
1486     }
1487 
1488     hdev->mem = g_malloc0(offsetof(struct vhost_memory, regions));
1489     hdev->n_mem_sections = 0;
1490     hdev->mem_sections = NULL;
1491     hdev->log = NULL;
1492     hdev->log_size = 0;
1493     hdev->log_enabled = false;
1494     hdev->started = false;
1495     memory_listener_register(&hdev->memory_listener, &address_space_memory);
1496     QLIST_INSERT_HEAD(&vhost_devices, hdev, entry);
1497 
1498     if (used_memslots > hdev->vhost_ops->vhost_backend_memslots_limit(hdev)) {
1499         error_setg(errp, "vhost backend memory slots limit is less"
1500                    " than current number of present memory slots");
1501         r = -EINVAL;
1502         goto fail_busyloop;
1503     }
1504 
1505     return 0;
1506 
1507 fail_busyloop:
1508     if (busyloop_timeout) {
1509         while (--i >= 0) {
1510             vhost_virtqueue_set_busyloop_timeout(hdev, hdev->vq_index + i, 0);
1511         }
1512     }
1513 fail:
1514     hdev->nvqs = n_initialized_vqs;
1515     vhost_dev_cleanup(hdev);
1516     return r;
1517 }
1518 
1519 void vhost_dev_cleanup(struct vhost_dev *hdev)
1520 {
1521     int i;
1522 
1523     trace_vhost_dev_cleanup(hdev);
1524 
1525     for (i = 0; i < hdev->nvqs; ++i) {
1526         vhost_virtqueue_cleanup(hdev->vqs + i);
1527     }
1528     if (hdev->mem) {
1529         /* those are only safe after successful init */
1530         memory_listener_unregister(&hdev->memory_listener);
1531         QLIST_REMOVE(hdev, entry);
1532     }
1533     if (hdev->migration_blocker) {
1534         migrate_del_blocker(hdev->migration_blocker);
1535         error_free(hdev->migration_blocker);
1536     }
1537     g_free(hdev->mem);
1538     g_free(hdev->mem_sections);
1539     if (hdev->vhost_ops) {
1540         hdev->vhost_ops->vhost_backend_cleanup(hdev);
1541     }
1542     assert(!hdev->log);
1543 
1544     memset(hdev, 0, sizeof(struct vhost_dev));
1545 }
1546 
1547 static void vhost_dev_disable_notifiers_nvqs(struct vhost_dev *hdev,
1548                                              VirtIODevice *vdev,
1549                                              unsigned int nvqs)
1550 {
1551     BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev)));
1552     int i, r;
1553 
1554     /*
1555      * Batch all the host notifiers in a single transaction to avoid
1556      * quadratic time complexity in address_space_update_ioeventfds().
1557      */
1558     memory_region_transaction_begin();
1559 
1560     for (i = 0; i < nvqs; ++i) {
1561         r = virtio_bus_set_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i,
1562                                          false);
1563         if (r < 0) {
1564             error_report("vhost VQ %d notifier cleanup failed: %d", i, -r);
1565         }
1566         assert(r >= 0);
1567     }
1568 
1569     /*
1570      * The transaction expects the ioeventfds to be open when it
1571      * commits. Do it now, before the cleanup loop.
1572      */
1573     memory_region_transaction_commit();
1574 
1575     for (i = 0; i < nvqs; ++i) {
1576         virtio_bus_cleanup_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i);
1577     }
1578     virtio_device_release_ioeventfd(vdev);
1579 }
1580 
1581 /* Stop processing guest IO notifications in qemu.
1582  * Start processing them in vhost in kernel.
1583  */
1584 int vhost_dev_enable_notifiers(struct vhost_dev *hdev, VirtIODevice *vdev)
1585 {
1586     BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev)));
1587     int i, r;
1588 
1589     /* We will pass the notifiers to the kernel, make sure that QEMU
1590      * doesn't interfere.
1591      */
1592     r = virtio_device_grab_ioeventfd(vdev);
1593     if (r < 0) {
1594         error_report("binding does not support host notifiers");
1595         return r;
1596     }
1597 
1598     /*
1599      * Batch all the host notifiers in a single transaction to avoid
1600      * quadratic time complexity in address_space_update_ioeventfds().
1601      */
1602     memory_region_transaction_begin();
1603 
1604     for (i = 0; i < hdev->nvqs; ++i) {
1605         r = virtio_bus_set_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i,
1606                                          true);
1607         if (r < 0) {
1608             error_report("vhost VQ %d notifier binding failed: %d", i, -r);
1609             memory_region_transaction_commit();
1610             vhost_dev_disable_notifiers_nvqs(hdev, vdev, i);
1611             return r;
1612         }
1613     }
1614 
1615     memory_region_transaction_commit();
1616 
1617     return 0;
1618 }
1619 
1620 /* Stop processing guest IO notifications in vhost.
1621  * Start processing them in qemu.
1622  * This might actually run the qemu handlers right away,
1623  * so virtio in qemu must be completely setup when this is called.
1624  */
1625 void vhost_dev_disable_notifiers(struct vhost_dev *hdev, VirtIODevice *vdev)
1626 {
1627     vhost_dev_disable_notifiers_nvqs(hdev, vdev, hdev->nvqs);
1628 }
1629 
1630 /* Test and clear event pending status.
1631  * Should be called after unmask to avoid losing events.
1632  */
1633 bool vhost_virtqueue_pending(struct vhost_dev *hdev, int n)
1634 {
1635     struct vhost_virtqueue *vq = hdev->vqs + n - hdev->vq_index;
1636     assert(n >= hdev->vq_index && n < hdev->vq_index + hdev->nvqs);
1637     return event_notifier_test_and_clear(&vq->masked_notifier);
1638 }
1639 
1640 /* Mask/unmask events from this vq. */
1641 void vhost_virtqueue_mask(struct vhost_dev *hdev, VirtIODevice *vdev, int n,
1642                          bool mask)
1643 {
1644     struct VirtQueue *vvq = virtio_get_queue(vdev, n);
1645     int r, index = n - hdev->vq_index;
1646     struct vhost_vring_file file;
1647 
1648     /* should only be called after backend is connected */
1649     assert(hdev->vhost_ops);
1650 
1651     if (mask) {
1652         assert(vdev->use_guest_notifier_mask);
1653         file.fd = event_notifier_get_wfd(&hdev->vqs[index].masked_notifier);
1654     } else {
1655         file.fd = event_notifier_get_wfd(virtio_queue_get_guest_notifier(vvq));
1656     }
1657 
1658     file.index = hdev->vhost_ops->vhost_get_vq_index(hdev, n);
1659     r = hdev->vhost_ops->vhost_set_vring_call(hdev, &file);
1660     if (r < 0) {
1661         error_report("vhost_set_vring_call failed %d", -r);
1662     }
1663 }
1664 
1665 bool vhost_config_pending(struct vhost_dev *hdev)
1666 {
1667     assert(hdev->vhost_ops);
1668     if ((hdev->started == false) ||
1669         (hdev->vhost_ops->vhost_set_config_call == NULL)) {
1670         return false;
1671     }
1672 
1673     EventNotifier *notifier =
1674         &hdev->vqs[VHOST_QUEUE_NUM_CONFIG_INR].masked_config_notifier;
1675     return event_notifier_test_and_clear(notifier);
1676 }
1677 
1678 void vhost_config_mask(struct vhost_dev *hdev, VirtIODevice *vdev, bool mask)
1679 {
1680     int fd;
1681     int r;
1682     EventNotifier *notifier =
1683         &hdev->vqs[VHOST_QUEUE_NUM_CONFIG_INR].masked_config_notifier;
1684     EventNotifier *config_notifier = &vdev->config_notifier;
1685     assert(hdev->vhost_ops);
1686 
1687     if ((hdev->started == false) ||
1688         (hdev->vhost_ops->vhost_set_config_call == NULL)) {
1689         return;
1690     }
1691     if (mask) {
1692         assert(vdev->use_guest_notifier_mask);
1693         fd = event_notifier_get_fd(notifier);
1694     } else {
1695         fd = event_notifier_get_fd(config_notifier);
1696     }
1697     r = hdev->vhost_ops->vhost_set_config_call(hdev, fd);
1698     if (r < 0) {
1699         error_report("vhost_set_config_call failed %d", -r);
1700     }
1701 }
1702 
1703 static void vhost_stop_config_intr(struct vhost_dev *dev)
1704 {
1705     int fd = -1;
1706     assert(dev->vhost_ops);
1707     if (dev->vhost_ops->vhost_set_config_call) {
1708         dev->vhost_ops->vhost_set_config_call(dev, fd);
1709     }
1710 }
1711 
1712 static void vhost_start_config_intr(struct vhost_dev *dev)
1713 {
1714     int r;
1715 
1716     assert(dev->vhost_ops);
1717     int fd = event_notifier_get_fd(&dev->vdev->config_notifier);
1718     if (dev->vhost_ops->vhost_set_config_call) {
1719         r = dev->vhost_ops->vhost_set_config_call(dev, fd);
1720         if (!r) {
1721             event_notifier_set(&dev->vdev->config_notifier);
1722         }
1723     }
1724 }
1725 
1726 uint64_t vhost_get_features(struct vhost_dev *hdev, const int *feature_bits,
1727                             uint64_t features)
1728 {
1729     const int *bit = feature_bits;
1730     while (*bit != VHOST_INVALID_FEATURE_BIT) {
1731         uint64_t bit_mask = (1ULL << *bit);
1732         if (!(hdev->features & bit_mask)) {
1733             features &= ~bit_mask;
1734         }
1735         bit++;
1736     }
1737     return features;
1738 }
1739 
1740 void vhost_ack_features(struct vhost_dev *hdev, const int *feature_bits,
1741                         uint64_t features)
1742 {
1743     const int *bit = feature_bits;
1744     while (*bit != VHOST_INVALID_FEATURE_BIT) {
1745         uint64_t bit_mask = (1ULL << *bit);
1746         if (features & bit_mask) {
1747             hdev->acked_features |= bit_mask;
1748         }
1749         bit++;
1750     }
1751 }
1752 
1753 int vhost_dev_get_config(struct vhost_dev *hdev, uint8_t *config,
1754                          uint32_t config_len, Error **errp)
1755 {
1756     assert(hdev->vhost_ops);
1757 
1758     if (hdev->vhost_ops->vhost_get_config) {
1759         return hdev->vhost_ops->vhost_get_config(hdev, config, config_len,
1760                                                  errp);
1761     }
1762 
1763     error_setg(errp, "vhost_get_config not implemented");
1764     return -ENOSYS;
1765 }
1766 
1767 int vhost_dev_set_config(struct vhost_dev *hdev, const uint8_t *data,
1768                          uint32_t offset, uint32_t size, uint32_t flags)
1769 {
1770     assert(hdev->vhost_ops);
1771 
1772     if (hdev->vhost_ops->vhost_set_config) {
1773         return hdev->vhost_ops->vhost_set_config(hdev, data, offset,
1774                                                  size, flags);
1775     }
1776 
1777     return -ENOSYS;
1778 }
1779 
1780 void vhost_dev_set_config_notifier(struct vhost_dev *hdev,
1781                                    const VhostDevConfigOps *ops)
1782 {
1783     hdev->config_ops = ops;
1784 }
1785 
1786 void vhost_dev_free_inflight(struct vhost_inflight *inflight)
1787 {
1788     if (inflight && inflight->addr) {
1789         qemu_memfd_free(inflight->addr, inflight->size, inflight->fd);
1790         inflight->addr = NULL;
1791         inflight->fd = -1;
1792     }
1793 }
1794 
1795 static int vhost_dev_resize_inflight(struct vhost_inflight *inflight,
1796                                      uint64_t new_size)
1797 {
1798     Error *err = NULL;
1799     int fd = -1;
1800     void *addr = qemu_memfd_alloc("vhost-inflight", new_size,
1801                                   F_SEAL_GROW | F_SEAL_SHRINK | F_SEAL_SEAL,
1802                                   &fd, &err);
1803 
1804     if (err) {
1805         error_report_err(err);
1806         return -ENOMEM;
1807     }
1808 
1809     vhost_dev_free_inflight(inflight);
1810     inflight->offset = 0;
1811     inflight->addr = addr;
1812     inflight->fd = fd;
1813     inflight->size = new_size;
1814 
1815     return 0;
1816 }
1817 
1818 void vhost_dev_save_inflight(struct vhost_inflight *inflight, QEMUFile *f)
1819 {
1820     if (inflight->addr) {
1821         qemu_put_be64(f, inflight->size);
1822         qemu_put_be16(f, inflight->queue_size);
1823         qemu_put_buffer(f, inflight->addr, inflight->size);
1824     } else {
1825         qemu_put_be64(f, 0);
1826     }
1827 }
1828 
1829 int vhost_dev_load_inflight(struct vhost_inflight *inflight, QEMUFile *f)
1830 {
1831     uint64_t size;
1832 
1833     size = qemu_get_be64(f);
1834     if (!size) {
1835         return 0;
1836     }
1837 
1838     if (inflight->size != size) {
1839         int ret = vhost_dev_resize_inflight(inflight, size);
1840         if (ret < 0) {
1841             return ret;
1842         }
1843     }
1844     inflight->queue_size = qemu_get_be16(f);
1845 
1846     qemu_get_buffer(f, inflight->addr, size);
1847 
1848     return 0;
1849 }
1850 
1851 int vhost_dev_prepare_inflight(struct vhost_dev *hdev, VirtIODevice *vdev)
1852 {
1853     int r;
1854 
1855     if (hdev->vhost_ops->vhost_get_inflight_fd == NULL ||
1856         hdev->vhost_ops->vhost_set_inflight_fd == NULL) {
1857         return 0;
1858     }
1859 
1860     hdev->vdev = vdev;
1861 
1862     r = vhost_dev_set_features(hdev, hdev->log_enabled);
1863     if (r < 0) {
1864         VHOST_OPS_DEBUG(r, "vhost_dev_prepare_inflight failed");
1865         return r;
1866     }
1867 
1868     return 0;
1869 }
1870 
1871 int vhost_dev_set_inflight(struct vhost_dev *dev,
1872                            struct vhost_inflight *inflight)
1873 {
1874     int r;
1875 
1876     if (dev->vhost_ops->vhost_set_inflight_fd && inflight->addr) {
1877         r = dev->vhost_ops->vhost_set_inflight_fd(dev, inflight);
1878         if (r) {
1879             VHOST_OPS_DEBUG(r, "vhost_set_inflight_fd failed");
1880             return r;
1881         }
1882     }
1883 
1884     return 0;
1885 }
1886 
1887 int vhost_dev_get_inflight(struct vhost_dev *dev, uint16_t queue_size,
1888                            struct vhost_inflight *inflight)
1889 {
1890     int r;
1891 
1892     if (dev->vhost_ops->vhost_get_inflight_fd) {
1893         r = dev->vhost_ops->vhost_get_inflight_fd(dev, queue_size, inflight);
1894         if (r) {
1895             VHOST_OPS_DEBUG(r, "vhost_get_inflight_fd failed");
1896             return r;
1897         }
1898     }
1899 
1900     return 0;
1901 }
1902 
1903 static int vhost_dev_set_vring_enable(struct vhost_dev *hdev, int enable)
1904 {
1905     if (!hdev->vhost_ops->vhost_set_vring_enable) {
1906         return 0;
1907     }
1908 
1909     /*
1910      * For vhost-user devices, if VHOST_USER_F_PROTOCOL_FEATURES has not
1911      * been negotiated, the rings start directly in the enabled state, and
1912      * .vhost_set_vring_enable callback will fail since
1913      * VHOST_USER_SET_VRING_ENABLE is not supported.
1914      */
1915     if (hdev->vhost_ops->backend_type == VHOST_BACKEND_TYPE_USER &&
1916         !virtio_has_feature(hdev->backend_features,
1917                             VHOST_USER_F_PROTOCOL_FEATURES)) {
1918         return 0;
1919     }
1920 
1921     return hdev->vhost_ops->vhost_set_vring_enable(hdev, enable);
1922 }
1923 
1924 /* Host notifiers must be enabled at this point. */
1925 int vhost_dev_start(struct vhost_dev *hdev, VirtIODevice *vdev, bool vrings)
1926 {
1927     int i, r;
1928 
1929     /* should only be called after backend is connected */
1930     assert(hdev->vhost_ops);
1931 
1932     trace_vhost_dev_start(hdev, vdev->name, vrings);
1933 
1934     vdev->vhost_started = true;
1935     hdev->started = true;
1936     hdev->vdev = vdev;
1937 
1938     r = vhost_dev_set_features(hdev, hdev->log_enabled);
1939     if (r < 0) {
1940         goto fail_features;
1941     }
1942 
1943     if (vhost_dev_has_iommu(hdev)) {
1944         memory_listener_register(&hdev->iommu_listener, vdev->dma_as);
1945     }
1946 
1947     r = hdev->vhost_ops->vhost_set_mem_table(hdev, hdev->mem);
1948     if (r < 0) {
1949         VHOST_OPS_DEBUG(r, "vhost_set_mem_table failed");
1950         goto fail_mem;
1951     }
1952     for (i = 0; i < hdev->nvqs; ++i) {
1953         r = vhost_virtqueue_start(hdev,
1954                                   vdev,
1955                                   hdev->vqs + i,
1956                                   hdev->vq_index + i);
1957         if (r < 0) {
1958             goto fail_vq;
1959         }
1960     }
1961 
1962     r = event_notifier_init(
1963         &hdev->vqs[VHOST_QUEUE_NUM_CONFIG_INR].masked_config_notifier, 0);
1964     if (r < 0) {
1965         VHOST_OPS_DEBUG(r, "event_notifier_init failed");
1966         goto fail_vq;
1967     }
1968     event_notifier_test_and_clear(
1969         &hdev->vqs[VHOST_QUEUE_NUM_CONFIG_INR].masked_config_notifier);
1970     if (!vdev->use_guest_notifier_mask) {
1971         vhost_config_mask(hdev, vdev, true);
1972     }
1973     if (hdev->log_enabled) {
1974         uint64_t log_base;
1975 
1976         hdev->log_size = vhost_get_log_size(hdev);
1977         hdev->log = vhost_log_get(hdev->log_size,
1978                                   vhost_dev_log_is_shared(hdev));
1979         log_base = (uintptr_t)hdev->log->log;
1980         r = hdev->vhost_ops->vhost_set_log_base(hdev,
1981                                                 hdev->log_size ? log_base : 0,
1982                                                 hdev->log);
1983         if (r < 0) {
1984             VHOST_OPS_DEBUG(r, "vhost_set_log_base failed");
1985             goto fail_log;
1986         }
1987     }
1988     if (vrings) {
1989         r = vhost_dev_set_vring_enable(hdev, true);
1990         if (r) {
1991             goto fail_log;
1992         }
1993     }
1994     if (hdev->vhost_ops->vhost_dev_start) {
1995         r = hdev->vhost_ops->vhost_dev_start(hdev, true);
1996         if (r) {
1997             goto fail_start;
1998         }
1999     }
2000     if (vhost_dev_has_iommu(hdev) &&
2001         hdev->vhost_ops->vhost_set_iotlb_callback) {
2002             hdev->vhost_ops->vhost_set_iotlb_callback(hdev, true);
2003 
2004         /* Update used ring information for IOTLB to work correctly,
2005          * vhost-kernel code requires for this.*/
2006         for (i = 0; i < hdev->nvqs; ++i) {
2007             struct vhost_virtqueue *vq = hdev->vqs + i;
2008             vhost_device_iotlb_miss(hdev, vq->used_phys, true);
2009         }
2010     }
2011     vhost_start_config_intr(hdev);
2012     return 0;
2013 fail_start:
2014     if (vrings) {
2015         vhost_dev_set_vring_enable(hdev, false);
2016     }
2017 fail_log:
2018     vhost_log_put(hdev, false);
2019 fail_vq:
2020     while (--i >= 0) {
2021         vhost_virtqueue_stop(hdev,
2022                              vdev,
2023                              hdev->vqs + i,
2024                              hdev->vq_index + i);
2025     }
2026 
2027 fail_mem:
2028     if (vhost_dev_has_iommu(hdev)) {
2029         memory_listener_unregister(&hdev->iommu_listener);
2030     }
2031 fail_features:
2032     vdev->vhost_started = false;
2033     hdev->started = false;
2034     return r;
2035 }
2036 
2037 /* Host notifiers must be enabled at this point. */
2038 void vhost_dev_stop(struct vhost_dev *hdev, VirtIODevice *vdev, bool vrings)
2039 {
2040     int i;
2041 
2042     /* should only be called after backend is connected */
2043     assert(hdev->vhost_ops);
2044     event_notifier_test_and_clear(
2045         &hdev->vqs[VHOST_QUEUE_NUM_CONFIG_INR].masked_config_notifier);
2046     event_notifier_test_and_clear(&vdev->config_notifier);
2047     event_notifier_cleanup(
2048         &hdev->vqs[VHOST_QUEUE_NUM_CONFIG_INR].masked_config_notifier);
2049 
2050     trace_vhost_dev_stop(hdev, vdev->name, vrings);
2051 
2052     if (hdev->vhost_ops->vhost_dev_start) {
2053         hdev->vhost_ops->vhost_dev_start(hdev, false);
2054     }
2055     if (vrings) {
2056         vhost_dev_set_vring_enable(hdev, false);
2057     }
2058     for (i = 0; i < hdev->nvqs; ++i) {
2059         vhost_virtqueue_stop(hdev,
2060                              vdev,
2061                              hdev->vqs + i,
2062                              hdev->vq_index + i);
2063     }
2064     if (hdev->vhost_ops->vhost_reset_status) {
2065         hdev->vhost_ops->vhost_reset_status(hdev);
2066     }
2067 
2068     if (vhost_dev_has_iommu(hdev)) {
2069         if (hdev->vhost_ops->vhost_set_iotlb_callback) {
2070             hdev->vhost_ops->vhost_set_iotlb_callback(hdev, false);
2071         }
2072         memory_listener_unregister(&hdev->iommu_listener);
2073     }
2074     vhost_stop_config_intr(hdev);
2075     vhost_log_put(hdev, true);
2076     hdev->started = false;
2077     vdev->vhost_started = false;
2078     hdev->vdev = NULL;
2079 }
2080 
2081 int vhost_net_set_backend(struct vhost_dev *hdev,
2082                           struct vhost_vring_file *file)
2083 {
2084     if (hdev->vhost_ops->vhost_net_set_backend) {
2085         return hdev->vhost_ops->vhost_net_set_backend(hdev, file);
2086     }
2087 
2088     return -ENOSYS;
2089 }
2090