xref: /qemu/hw/virtio/vhost.c (revision b4ab225c34944658ddff62e4ee8127c3838bcf1d)
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 "standard-headers/linux/vhost_types.h"
24 #include "exec/address-spaces.h"
25 #include "hw/virtio/virtio-bus.h"
26 #include "hw/virtio/virtio-access.h"
27 #include "migration/blocker.h"
28 #include "migration/qemu-file-types.h"
29 #include "sysemu/dma.h"
30 #include "sysemu/tcg.h"
31 #include "trace.h"
32 
33 /* enabled until disconnected backend stabilizes */
34 #define _VHOST_DEBUG 1
35 
36 #ifdef _VHOST_DEBUG
37 #define VHOST_OPS_DEBUG(fmt, ...) \
38     do { error_report(fmt ": %s (%d)", ## __VA_ARGS__, \
39                       strerror(errno), errno); } while (0)
40 #else
41 #define VHOST_OPS_DEBUG(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 = atomic_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 static int vhost_sync_dirty_bitmap(struct vhost_dev *dev,
110                                    MemoryRegionSection *section,
111                                    hwaddr first,
112                                    hwaddr last)
113 {
114     int i;
115     hwaddr start_addr;
116     hwaddr end_addr;
117 
118     if (!dev->log_enabled || !dev->started) {
119         return 0;
120     }
121     start_addr = section->offset_within_address_space;
122     end_addr = range_get_last(start_addr, int128_get64(section->size));
123     start_addr = MAX(first, start_addr);
124     end_addr = MIN(last, end_addr);
125 
126     for (i = 0; i < dev->mem->nregions; ++i) {
127         struct vhost_memory_region *reg = dev->mem->regions + i;
128         vhost_dev_sync_region(dev, section, start_addr, end_addr,
129                               reg->guest_phys_addr,
130                               range_get_last(reg->guest_phys_addr,
131                                              reg->memory_size));
132     }
133     for (i = 0; i < dev->nvqs; ++i) {
134         struct vhost_virtqueue *vq = dev->vqs + i;
135 
136         if (!vq->used_phys && !vq->used_size) {
137             continue;
138         }
139 
140         vhost_dev_sync_region(dev, section, start_addr, end_addr, vq->used_phys,
141                               range_get_last(vq->used_phys, vq->used_size));
142     }
143     return 0;
144 }
145 
146 static void vhost_log_sync(MemoryListener *listener,
147                           MemoryRegionSection *section)
148 {
149     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
150                                          memory_listener);
151     vhost_sync_dirty_bitmap(dev, section, 0x0, ~0x0ULL);
152 }
153 
154 static void vhost_log_sync_range(struct vhost_dev *dev,
155                                  hwaddr first, hwaddr last)
156 {
157     int i;
158     /* FIXME: this is N^2 in number of sections */
159     for (i = 0; i < dev->n_mem_sections; ++i) {
160         MemoryRegionSection *section = &dev->mem_sections[i];
161         vhost_sync_dirty_bitmap(dev, section, first, last);
162     }
163 }
164 
165 static uint64_t vhost_get_log_size(struct vhost_dev *dev)
166 {
167     uint64_t log_size = 0;
168     int i;
169     for (i = 0; i < dev->mem->nregions; ++i) {
170         struct vhost_memory_region *reg = dev->mem->regions + i;
171         uint64_t last = range_get_last(reg->guest_phys_addr,
172                                        reg->memory_size);
173         log_size = MAX(log_size, last / VHOST_LOG_CHUNK + 1);
174     }
175     for (i = 0; i < dev->nvqs; ++i) {
176         struct vhost_virtqueue *vq = dev->vqs + i;
177 
178         if (!vq->used_phys && !vq->used_size) {
179             continue;
180         }
181 
182         uint64_t last = vq->used_phys + vq->used_size - 1;
183         log_size = MAX(log_size, last / VHOST_LOG_CHUNK + 1);
184     }
185     return log_size;
186 }
187 
188 static struct vhost_log *vhost_log_alloc(uint64_t size, bool share)
189 {
190     Error *err = NULL;
191     struct vhost_log *log;
192     uint64_t logsize = size * sizeof(*(log->log));
193     int fd = -1;
194 
195     log = g_new0(struct vhost_log, 1);
196     if (share) {
197         log->log = qemu_memfd_alloc("vhost-log", logsize,
198                                     F_SEAL_GROW | F_SEAL_SHRINK | F_SEAL_SEAL,
199                                     &fd, &err);
200         if (err) {
201             error_report_err(err);
202             g_free(log);
203             return NULL;
204         }
205         memset(log->log, 0, logsize);
206     } else {
207         log->log = g_malloc0(logsize);
208     }
209 
210     log->size = size;
211     log->refcnt = 1;
212     log->fd = fd;
213 
214     return log;
215 }
216 
217 static struct vhost_log *vhost_log_get(uint64_t size, bool share)
218 {
219     struct vhost_log *log = share ? vhost_log_shm : vhost_log;
220 
221     if (!log || log->size != size) {
222         log = vhost_log_alloc(size, share);
223         if (share) {
224             vhost_log_shm = log;
225         } else {
226             vhost_log = log;
227         }
228     } else {
229         ++log->refcnt;
230     }
231 
232     return log;
233 }
234 
235 static void vhost_log_put(struct vhost_dev *dev, bool sync)
236 {
237     struct vhost_log *log = dev->log;
238 
239     if (!log) {
240         return;
241     }
242 
243     --log->refcnt;
244     if (log->refcnt == 0) {
245         /* Sync only the range covered by the old log */
246         if (dev->log_size && sync) {
247             vhost_log_sync_range(dev, 0, dev->log_size * VHOST_LOG_CHUNK - 1);
248         }
249 
250         if (vhost_log == log) {
251             g_free(log->log);
252             vhost_log = NULL;
253         } else if (vhost_log_shm == log) {
254             qemu_memfd_free(log->log, log->size * sizeof(*(log->log)),
255                             log->fd);
256             vhost_log_shm = NULL;
257         }
258 
259         g_free(log);
260     }
261 
262     dev->log = NULL;
263     dev->log_size = 0;
264 }
265 
266 static bool vhost_dev_log_is_shared(struct vhost_dev *dev)
267 {
268     return dev->vhost_ops->vhost_requires_shm_log &&
269            dev->vhost_ops->vhost_requires_shm_log(dev);
270 }
271 
272 static inline void vhost_dev_log_resize(struct vhost_dev *dev, uint64_t size)
273 {
274     struct vhost_log *log = vhost_log_get(size, vhost_dev_log_is_shared(dev));
275     uint64_t log_base = (uintptr_t)log->log;
276     int r;
277 
278     /* inform backend of log switching, this must be done before
279        releasing the current log, to ensure no logging is lost */
280     r = dev->vhost_ops->vhost_set_log_base(dev, log_base, log);
281     if (r < 0) {
282         VHOST_OPS_DEBUG("vhost_set_log_base failed");
283     }
284 
285     vhost_log_put(dev, true);
286     dev->log = log;
287     dev->log_size = size;
288 }
289 
290 static int vhost_dev_has_iommu(struct vhost_dev *dev)
291 {
292     VirtIODevice *vdev = dev->vdev;
293 
294     /*
295      * For vhost, VIRTIO_F_IOMMU_PLATFORM means the backend support
296      * incremental memory mapping API via IOTLB API. For platform that
297      * does not have IOMMU, there's no need to enable this feature
298      * which may cause unnecessary IOTLB miss/update trnasactions.
299      */
300     return vdev->dma_as != &address_space_memory &&
301            virtio_host_has_feature(vdev, VIRTIO_F_IOMMU_PLATFORM);
302 }
303 
304 static void *vhost_memory_map(struct vhost_dev *dev, hwaddr addr,
305                               hwaddr *plen, bool is_write)
306 {
307     if (!vhost_dev_has_iommu(dev)) {
308         return cpu_physical_memory_map(addr, plen, is_write);
309     } else {
310         return (void *)(uintptr_t)addr;
311     }
312 }
313 
314 static void vhost_memory_unmap(struct vhost_dev *dev, void *buffer,
315                                hwaddr len, int is_write,
316                                hwaddr access_len)
317 {
318     if (!vhost_dev_has_iommu(dev)) {
319         cpu_physical_memory_unmap(buffer, len, is_write, access_len);
320     }
321 }
322 
323 static int vhost_verify_ring_part_mapping(void *ring_hva,
324                                           uint64_t ring_gpa,
325                                           uint64_t ring_size,
326                                           void *reg_hva,
327                                           uint64_t reg_gpa,
328                                           uint64_t reg_size)
329 {
330     uint64_t hva_ring_offset;
331     uint64_t ring_last = range_get_last(ring_gpa, ring_size);
332     uint64_t reg_last = range_get_last(reg_gpa, reg_size);
333 
334     if (ring_last < reg_gpa || ring_gpa > reg_last) {
335         return 0;
336     }
337     /* check that whole ring's is mapped */
338     if (ring_last > reg_last) {
339         return -ENOMEM;
340     }
341     /* check that ring's MemoryRegion wasn't replaced */
342     hva_ring_offset = ring_gpa - reg_gpa;
343     if (ring_hva != reg_hva + hva_ring_offset) {
344         return -EBUSY;
345     }
346 
347     return 0;
348 }
349 
350 static int vhost_verify_ring_mappings(struct vhost_dev *dev,
351                                       void *reg_hva,
352                                       uint64_t reg_gpa,
353                                       uint64_t reg_size)
354 {
355     int i, j;
356     int r = 0;
357     const char *part_name[] = {
358         "descriptor table",
359         "available ring",
360         "used ring"
361     };
362 
363     if (vhost_dev_has_iommu(dev)) {
364         return 0;
365     }
366 
367     for (i = 0; i < dev->nvqs; ++i) {
368         struct vhost_virtqueue *vq = dev->vqs + i;
369 
370         if (vq->desc_phys == 0) {
371             continue;
372         }
373 
374         j = 0;
375         r = vhost_verify_ring_part_mapping(
376                 vq->desc, vq->desc_phys, vq->desc_size,
377                 reg_hva, reg_gpa, reg_size);
378         if (r) {
379             break;
380         }
381 
382         j++;
383         r = vhost_verify_ring_part_mapping(
384                 vq->avail, vq->avail_phys, vq->avail_size,
385                 reg_hva, reg_gpa, reg_size);
386         if (r) {
387             break;
388         }
389 
390         j++;
391         r = vhost_verify_ring_part_mapping(
392                 vq->used, vq->used_phys, vq->used_size,
393                 reg_hva, reg_gpa, reg_size);
394         if (r) {
395             break;
396         }
397     }
398 
399     if (r == -ENOMEM) {
400         error_report("Unable to map %s for ring %d", part_name[j], i);
401     } else if (r == -EBUSY) {
402         error_report("%s relocated for ring %d", part_name[j], i);
403     }
404     return r;
405 }
406 
407 /*
408  * vhost_section: identify sections needed for vhost access
409  *
410  * We only care about RAM sections here (where virtqueue and guest
411  * internals accessed by virtio might live). If we find one we still
412  * allow the backend to potentially filter it out of our list.
413  */
414 static bool vhost_section(struct vhost_dev *dev, MemoryRegionSection *section)
415 {
416     MemoryRegion *mr = section->mr;
417 
418     if (memory_region_is_ram(mr) && !memory_region_is_rom(mr)) {
419         uint8_t dirty_mask = memory_region_get_dirty_log_mask(mr);
420         uint8_t handled_dirty;
421 
422         /*
423          * Kernel based vhost doesn't handle any block which is doing
424          * dirty-tracking other than migration for which it has
425          * specific logging support. However for TCG the kernel never
426          * gets involved anyway so we can also ignore it's
427          * self-modiying code detection flags. However a vhost-user
428          * client could still confuse a TCG guest if it re-writes
429          * executable memory that has already been translated.
430          */
431         handled_dirty = (1 << DIRTY_MEMORY_MIGRATION) |
432             (1 << DIRTY_MEMORY_CODE);
433 
434         if (dirty_mask & ~handled_dirty) {
435             trace_vhost_reject_section(mr->name, 1);
436             return false;
437         }
438 
439         if (dev->vhost_ops->vhost_backend_mem_section_filter &&
440             !dev->vhost_ops->vhost_backend_mem_section_filter(dev, section)) {
441             trace_vhost_reject_section(mr->name, 2);
442             return false;
443         }
444 
445         trace_vhost_section(mr->name);
446         return true;
447     } else {
448         trace_vhost_reject_section(mr->name, 3);
449         return false;
450     }
451 }
452 
453 static void vhost_begin(MemoryListener *listener)
454 {
455     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
456                                          memory_listener);
457     dev->tmp_sections = NULL;
458     dev->n_tmp_sections = 0;
459 }
460 
461 static void vhost_commit(MemoryListener *listener)
462 {
463     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
464                                          memory_listener);
465     MemoryRegionSection *old_sections;
466     int n_old_sections;
467     uint64_t log_size;
468     size_t regions_size;
469     int r;
470     int i;
471     bool changed = false;
472 
473     /* Note we can be called before the device is started, but then
474      * starting the device calls set_mem_table, so we need to have
475      * built the data structures.
476      */
477     old_sections = dev->mem_sections;
478     n_old_sections = dev->n_mem_sections;
479     dev->mem_sections = dev->tmp_sections;
480     dev->n_mem_sections = dev->n_tmp_sections;
481 
482     if (dev->n_mem_sections != n_old_sections) {
483         changed = true;
484     } else {
485         /* Same size, lets check the contents */
486         for (int i = 0; i < n_old_sections; i++) {
487             if (!MemoryRegionSection_eq(&old_sections[i],
488                                         &dev->mem_sections[i])) {
489                 changed = true;
490                 break;
491             }
492         }
493     }
494 
495     trace_vhost_commit(dev->started, changed);
496     if (!changed) {
497         goto out;
498     }
499 
500     /* Rebuild the regions list from the new sections list */
501     regions_size = offsetof(struct vhost_memory, regions) +
502                        dev->n_mem_sections * sizeof dev->mem->regions[0];
503     dev->mem = g_realloc(dev->mem, regions_size);
504     dev->mem->nregions = dev->n_mem_sections;
505     used_memslots = dev->mem->nregions;
506     for (i = 0; i < dev->n_mem_sections; i++) {
507         struct vhost_memory_region *cur_vmr = dev->mem->regions + i;
508         struct MemoryRegionSection *mrs = dev->mem_sections + i;
509 
510         cur_vmr->guest_phys_addr = mrs->offset_within_address_space;
511         cur_vmr->memory_size     = int128_get64(mrs->size);
512         cur_vmr->userspace_addr  =
513             (uintptr_t)memory_region_get_ram_ptr(mrs->mr) +
514             mrs->offset_within_region;
515         cur_vmr->flags_padding   = 0;
516     }
517 
518     if (!dev->started) {
519         goto out;
520     }
521 
522     for (i = 0; i < dev->mem->nregions; i++) {
523         if (vhost_verify_ring_mappings(dev,
524                        (void *)(uintptr_t)dev->mem->regions[i].userspace_addr,
525                        dev->mem->regions[i].guest_phys_addr,
526                        dev->mem->regions[i].memory_size)) {
527             error_report("Verify ring failure on region %d", i);
528             abort();
529         }
530     }
531 
532     if (!dev->log_enabled) {
533         r = dev->vhost_ops->vhost_set_mem_table(dev, dev->mem);
534         if (r < 0) {
535             VHOST_OPS_DEBUG("vhost_set_mem_table failed");
536         }
537         goto out;
538     }
539     log_size = vhost_get_log_size(dev);
540     /* We allocate an extra 4K bytes to log,
541      * to reduce the * number of reallocations. */
542 #define VHOST_LOG_BUFFER (0x1000 / sizeof *dev->log)
543     /* To log more, must increase log size before table update. */
544     if (dev->log_size < log_size) {
545         vhost_dev_log_resize(dev, log_size + VHOST_LOG_BUFFER);
546     }
547     r = dev->vhost_ops->vhost_set_mem_table(dev, dev->mem);
548     if (r < 0) {
549         VHOST_OPS_DEBUG("vhost_set_mem_table failed");
550     }
551     /* To log less, can only decrease log size after table update. */
552     if (dev->log_size > log_size + VHOST_LOG_BUFFER) {
553         vhost_dev_log_resize(dev, log_size);
554     }
555 
556 out:
557     /* Deref the old list of sections, this must happen _after_ the
558      * vhost_set_mem_table to ensure the client isn't still using the
559      * section we're about to unref.
560      */
561     while (n_old_sections--) {
562         memory_region_unref(old_sections[n_old_sections].mr);
563     }
564     g_free(old_sections);
565     return;
566 }
567 
568 /* Adds the section data to the tmp_section structure.
569  * It relies on the listener calling us in memory address order
570  * and for each region (via the _add and _nop methods) to
571  * join neighbours.
572  */
573 static void vhost_region_add_section(struct vhost_dev *dev,
574                                      MemoryRegionSection *section)
575 {
576     bool need_add = true;
577     uint64_t mrs_size = int128_get64(section->size);
578     uint64_t mrs_gpa = section->offset_within_address_space;
579     uintptr_t mrs_host = (uintptr_t)memory_region_get_ram_ptr(section->mr) +
580                          section->offset_within_region;
581     RAMBlock *mrs_rb = section->mr->ram_block;
582 
583     trace_vhost_region_add_section(section->mr->name, mrs_gpa, mrs_size,
584                                    mrs_host);
585 
586     if (dev->vhost_ops->backend_type == VHOST_BACKEND_TYPE_USER) {
587         /* Round the section to it's page size */
588         /* First align the start down to a page boundary */
589         size_t mrs_page = qemu_ram_pagesize(mrs_rb);
590         uint64_t alignage = mrs_host & (mrs_page - 1);
591         if (alignage) {
592             mrs_host -= alignage;
593             mrs_size += alignage;
594             mrs_gpa  -= alignage;
595         }
596         /* Now align the size up to a page boundary */
597         alignage = mrs_size & (mrs_page - 1);
598         if (alignage) {
599             mrs_size += mrs_page - alignage;
600         }
601         trace_vhost_region_add_section_aligned(section->mr->name, mrs_gpa,
602                                                mrs_size, mrs_host);
603     }
604 
605     if (dev->n_tmp_sections) {
606         /* Since we already have at least one section, lets see if
607          * this extends it; since we're scanning in order, we only
608          * have to look at the last one, and the FlatView that calls
609          * us shouldn't have overlaps.
610          */
611         MemoryRegionSection *prev_sec = dev->tmp_sections +
612                                                (dev->n_tmp_sections - 1);
613         uint64_t prev_gpa_start = prev_sec->offset_within_address_space;
614         uint64_t prev_size = int128_get64(prev_sec->size);
615         uint64_t prev_gpa_end   = range_get_last(prev_gpa_start, prev_size);
616         uint64_t prev_host_start =
617                         (uintptr_t)memory_region_get_ram_ptr(prev_sec->mr) +
618                         prev_sec->offset_within_region;
619         uint64_t prev_host_end   = range_get_last(prev_host_start, prev_size);
620 
621         if (mrs_gpa <= (prev_gpa_end + 1)) {
622             /* OK, looks like overlapping/intersecting - it's possible that
623              * the rounding to page sizes has made them overlap, but they should
624              * match up in the same RAMBlock if they do.
625              */
626             if (mrs_gpa < prev_gpa_start) {
627                 error_report("%s:Section '%s' rounded to %"PRIx64
628                              " prior to previous '%s' %"PRIx64,
629                              __func__, section->mr->name, mrs_gpa,
630                              prev_sec->mr->name, prev_gpa_start);
631                 /* A way to cleanly fail here would be better */
632                 return;
633             }
634             /* Offset from the start of the previous GPA to this GPA */
635             size_t offset = mrs_gpa - prev_gpa_start;
636 
637             if (prev_host_start + offset == mrs_host &&
638                 section->mr == prev_sec->mr &&
639                 (!dev->vhost_ops->vhost_backend_can_merge ||
640                  dev->vhost_ops->vhost_backend_can_merge(dev,
641                     mrs_host, mrs_size,
642                     prev_host_start, prev_size))) {
643                 uint64_t max_end = MAX(prev_host_end, mrs_host + mrs_size);
644                 need_add = false;
645                 prev_sec->offset_within_address_space =
646                     MIN(prev_gpa_start, mrs_gpa);
647                 prev_sec->offset_within_region =
648                     MIN(prev_host_start, mrs_host) -
649                     (uintptr_t)memory_region_get_ram_ptr(prev_sec->mr);
650                 prev_sec->size = int128_make64(max_end - MIN(prev_host_start,
651                                                mrs_host));
652                 trace_vhost_region_add_section_merge(section->mr->name,
653                                         int128_get64(prev_sec->size),
654                                         prev_sec->offset_within_address_space,
655                                         prev_sec->offset_within_region);
656             } else {
657                 /* adjoining regions are fine, but overlapping ones with
658                  * different blocks/offsets shouldn't happen
659                  */
660                 if (mrs_gpa != prev_gpa_end + 1) {
661                     error_report("%s: Overlapping but not coherent sections "
662                                  "at %"PRIx64,
663                                  __func__, mrs_gpa);
664                     return;
665                 }
666             }
667         }
668     }
669 
670     if (need_add) {
671         ++dev->n_tmp_sections;
672         dev->tmp_sections = g_renew(MemoryRegionSection, dev->tmp_sections,
673                                     dev->n_tmp_sections);
674         dev->tmp_sections[dev->n_tmp_sections - 1] = *section;
675         /* The flatview isn't stable and we don't use it, making it NULL
676          * means we can memcmp the list.
677          */
678         dev->tmp_sections[dev->n_tmp_sections - 1].fv = NULL;
679         memory_region_ref(section->mr);
680     }
681 }
682 
683 /* Used for both add and nop callbacks */
684 static void vhost_region_addnop(MemoryListener *listener,
685                                 MemoryRegionSection *section)
686 {
687     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
688                                          memory_listener);
689 
690     if (!vhost_section(dev, section)) {
691         return;
692     }
693     vhost_region_add_section(dev, section);
694 }
695 
696 static void vhost_iommu_unmap_notify(IOMMUNotifier *n, IOMMUTLBEntry *iotlb)
697 {
698     struct vhost_iommu *iommu = container_of(n, struct vhost_iommu, n);
699     struct vhost_dev *hdev = iommu->hdev;
700     hwaddr iova = iotlb->iova + iommu->iommu_offset;
701 
702     if (vhost_backend_invalidate_device_iotlb(hdev, iova,
703                                               iotlb->addr_mask + 1)) {
704         error_report("Fail to invalidate device iotlb");
705     }
706 }
707 
708 static void vhost_iommu_region_add(MemoryListener *listener,
709                                    MemoryRegionSection *section)
710 {
711     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
712                                          iommu_listener);
713     struct vhost_iommu *iommu;
714     Int128 end;
715     int iommu_idx, ret;
716     IOMMUMemoryRegion *iommu_mr;
717     Error *err = NULL;
718 
719     if (!memory_region_is_iommu(section->mr)) {
720         return;
721     }
722 
723     iommu_mr = IOMMU_MEMORY_REGION(section->mr);
724 
725     iommu = g_malloc0(sizeof(*iommu));
726     end = int128_add(int128_make64(section->offset_within_region),
727                      section->size);
728     end = int128_sub(end, int128_one());
729     iommu_idx = memory_region_iommu_attrs_to_index(iommu_mr,
730                                                    MEMTXATTRS_UNSPECIFIED);
731     iommu_notifier_init(&iommu->n, vhost_iommu_unmap_notify,
732                         IOMMU_NOTIFIER_UNMAP,
733                         section->offset_within_region,
734                         int128_get64(end),
735                         iommu_idx);
736     iommu->mr = section->mr;
737     iommu->iommu_offset = section->offset_within_address_space -
738                           section->offset_within_region;
739     iommu->hdev = dev;
740     ret = memory_region_register_iommu_notifier(section->mr, &iommu->n, &err);
741     if (ret) {
742         error_report_err(err);
743         exit(1);
744     }
745     QLIST_INSERT_HEAD(&dev->iommu_list, iommu, iommu_next);
746     /* TODO: can replay help performance here? */
747 }
748 
749 static void vhost_iommu_region_del(MemoryListener *listener,
750                                    MemoryRegionSection *section)
751 {
752     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
753                                          iommu_listener);
754     struct vhost_iommu *iommu;
755 
756     if (!memory_region_is_iommu(section->mr)) {
757         return;
758     }
759 
760     QLIST_FOREACH(iommu, &dev->iommu_list, iommu_next) {
761         if (iommu->mr == section->mr &&
762             iommu->n.start == section->offset_within_region) {
763             memory_region_unregister_iommu_notifier(iommu->mr,
764                                                     &iommu->n);
765             QLIST_REMOVE(iommu, iommu_next);
766             g_free(iommu);
767             break;
768         }
769     }
770 }
771 
772 static int vhost_virtqueue_set_addr(struct vhost_dev *dev,
773                                     struct vhost_virtqueue *vq,
774                                     unsigned idx, bool enable_log)
775 {
776     struct vhost_vring_addr addr;
777     int r;
778     memset(&addr, 0, sizeof(struct vhost_vring_addr));
779 
780     if (dev->vhost_ops->vhost_vq_get_addr) {
781         r = dev->vhost_ops->vhost_vq_get_addr(dev, &addr, vq);
782         if (r < 0) {
783             VHOST_OPS_DEBUG("vhost_vq_get_addr failed");
784             return -errno;
785         }
786     } else {
787         addr.desc_user_addr = (uint64_t)(unsigned long)vq->desc;
788         addr.avail_user_addr = (uint64_t)(unsigned long)vq->avail;
789         addr.used_user_addr = (uint64_t)(unsigned long)vq->used;
790     }
791     addr.index = idx;
792     addr.log_guest_addr = vq->used_phys;
793     addr.flags = enable_log ? (1 << VHOST_VRING_F_LOG) : 0;
794     r = dev->vhost_ops->vhost_set_vring_addr(dev, &addr);
795     if (r < 0) {
796         VHOST_OPS_DEBUG("vhost_set_vring_addr failed");
797         return -errno;
798     }
799     return 0;
800 }
801 
802 static int vhost_dev_set_features(struct vhost_dev *dev,
803                                   bool enable_log)
804 {
805     uint64_t features = dev->acked_features;
806     int r;
807     if (enable_log) {
808         features |= 0x1ULL << VHOST_F_LOG_ALL;
809     }
810     if (!vhost_dev_has_iommu(dev)) {
811         features &= ~(0x1ULL << VIRTIO_F_IOMMU_PLATFORM);
812     }
813     r = dev->vhost_ops->vhost_set_features(dev, features);
814     if (r < 0) {
815         VHOST_OPS_DEBUG("vhost_set_features failed");
816     }
817     return r < 0 ? -errno : 0;
818 }
819 
820 static int vhost_dev_set_log(struct vhost_dev *dev, bool enable_log)
821 {
822     int r, i, idx;
823     r = vhost_dev_set_features(dev, enable_log);
824     if (r < 0) {
825         goto err_features;
826     }
827     for (i = 0; i < dev->nvqs; ++i) {
828         idx = dev->vhost_ops->vhost_get_vq_index(dev, dev->vq_index + i);
829         r = vhost_virtqueue_set_addr(dev, dev->vqs + i, idx,
830                                      enable_log);
831         if (r < 0) {
832             goto err_vq;
833         }
834     }
835     return 0;
836 err_vq:
837     for (; i >= 0; --i) {
838         idx = dev->vhost_ops->vhost_get_vq_index(dev, dev->vq_index + i);
839         vhost_virtqueue_set_addr(dev, dev->vqs + i, idx,
840                                  dev->log_enabled);
841     }
842     vhost_dev_set_features(dev, dev->log_enabled);
843 err_features:
844     return r;
845 }
846 
847 static int vhost_migration_log(MemoryListener *listener, bool enable)
848 {
849     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
850                                          memory_listener);
851     int r;
852     if (enable == dev->log_enabled) {
853         return 0;
854     }
855     if (!dev->started) {
856         dev->log_enabled = enable;
857         return 0;
858     }
859     if (!enable) {
860         r = vhost_dev_set_log(dev, false);
861         if (r < 0) {
862             return r;
863         }
864         vhost_log_put(dev, false);
865     } else {
866         vhost_dev_log_resize(dev, vhost_get_log_size(dev));
867         r = vhost_dev_set_log(dev, true);
868         if (r < 0) {
869             return r;
870         }
871     }
872     dev->log_enabled = enable;
873     return 0;
874 }
875 
876 static void vhost_log_global_start(MemoryListener *listener)
877 {
878     int r;
879 
880     r = vhost_migration_log(listener, true);
881     if (r < 0) {
882         abort();
883     }
884 }
885 
886 static void vhost_log_global_stop(MemoryListener *listener)
887 {
888     int r;
889 
890     r = vhost_migration_log(listener, false);
891     if (r < 0) {
892         abort();
893     }
894 }
895 
896 static void vhost_log_start(MemoryListener *listener,
897                             MemoryRegionSection *section,
898                             int old, int new)
899 {
900     /* FIXME: implement */
901 }
902 
903 static void vhost_log_stop(MemoryListener *listener,
904                            MemoryRegionSection *section,
905                            int old, int new)
906 {
907     /* FIXME: implement */
908 }
909 
910 /* The vhost driver natively knows how to handle the vrings of non
911  * cross-endian legacy devices and modern devices. Only legacy devices
912  * exposed to a bi-endian guest may require the vhost driver to use a
913  * specific endianness.
914  */
915 static inline bool vhost_needs_vring_endian(VirtIODevice *vdev)
916 {
917     if (virtio_vdev_has_feature(vdev, VIRTIO_F_VERSION_1)) {
918         return false;
919     }
920 #ifdef HOST_WORDS_BIGENDIAN
921     return vdev->device_endian == VIRTIO_DEVICE_ENDIAN_LITTLE;
922 #else
923     return vdev->device_endian == VIRTIO_DEVICE_ENDIAN_BIG;
924 #endif
925 }
926 
927 static int vhost_virtqueue_set_vring_endian_legacy(struct vhost_dev *dev,
928                                                    bool is_big_endian,
929                                                    int vhost_vq_index)
930 {
931     struct vhost_vring_state s = {
932         .index = vhost_vq_index,
933         .num = is_big_endian
934     };
935 
936     if (!dev->vhost_ops->vhost_set_vring_endian(dev, &s)) {
937         return 0;
938     }
939 
940     VHOST_OPS_DEBUG("vhost_set_vring_endian failed");
941     if (errno == ENOTTY) {
942         error_report("vhost does not support cross-endian");
943         return -ENOSYS;
944     }
945 
946     return -errno;
947 }
948 
949 static int vhost_memory_region_lookup(struct vhost_dev *hdev,
950                                       uint64_t gpa, uint64_t *uaddr,
951                                       uint64_t *len)
952 {
953     int i;
954 
955     for (i = 0; i < hdev->mem->nregions; i++) {
956         struct vhost_memory_region *reg = hdev->mem->regions + i;
957 
958         if (gpa >= reg->guest_phys_addr &&
959             reg->guest_phys_addr + reg->memory_size > gpa) {
960             *uaddr = reg->userspace_addr + gpa - reg->guest_phys_addr;
961             *len = reg->guest_phys_addr + reg->memory_size - gpa;
962             return 0;
963         }
964     }
965 
966     return -EFAULT;
967 }
968 
969 int vhost_device_iotlb_miss(struct vhost_dev *dev, uint64_t iova, int write)
970 {
971     IOMMUTLBEntry iotlb;
972     uint64_t uaddr, len;
973     int ret = -EFAULT;
974 
975     RCU_READ_LOCK_GUARD();
976 
977     trace_vhost_iotlb_miss(dev, 1);
978 
979     iotlb = address_space_get_iotlb_entry(dev->vdev->dma_as,
980                                           iova, write,
981                                           MEMTXATTRS_UNSPECIFIED);
982     if (iotlb.target_as != NULL) {
983         ret = vhost_memory_region_lookup(dev, iotlb.translated_addr,
984                                          &uaddr, &len);
985         if (ret) {
986             trace_vhost_iotlb_miss(dev, 3);
987             error_report("Fail to lookup the translated address "
988                          "%"PRIx64, iotlb.translated_addr);
989             goto out;
990         }
991 
992         len = MIN(iotlb.addr_mask + 1, len);
993         iova = iova & ~iotlb.addr_mask;
994 
995         ret = vhost_backend_update_device_iotlb(dev, iova, uaddr,
996                                                 len, iotlb.perm);
997         if (ret) {
998             trace_vhost_iotlb_miss(dev, 4);
999             error_report("Fail to update device iotlb");
1000             goto out;
1001         }
1002     }
1003 
1004     trace_vhost_iotlb_miss(dev, 2);
1005 
1006 out:
1007     return ret;
1008 }
1009 
1010 static int vhost_virtqueue_start(struct vhost_dev *dev,
1011                                 struct VirtIODevice *vdev,
1012                                 struct vhost_virtqueue *vq,
1013                                 unsigned idx)
1014 {
1015     BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev)));
1016     VirtioBusState *vbus = VIRTIO_BUS(qbus);
1017     VirtioBusClass *k = VIRTIO_BUS_GET_CLASS(vbus);
1018     hwaddr s, l, a;
1019     int r;
1020     int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, idx);
1021     struct vhost_vring_file file = {
1022         .index = vhost_vq_index
1023     };
1024     struct vhost_vring_state state = {
1025         .index = vhost_vq_index
1026     };
1027     struct VirtQueue *vvq = virtio_get_queue(vdev, idx);
1028 
1029     a = virtio_queue_get_desc_addr(vdev, idx);
1030     if (a == 0) {
1031         /* Queue might not be ready for start */
1032         return 0;
1033     }
1034 
1035     vq->num = state.num = virtio_queue_get_num(vdev, idx);
1036     r = dev->vhost_ops->vhost_set_vring_num(dev, &state);
1037     if (r) {
1038         VHOST_OPS_DEBUG("vhost_set_vring_num failed");
1039         return -errno;
1040     }
1041 
1042     state.num = virtio_queue_get_last_avail_idx(vdev, idx);
1043     r = dev->vhost_ops->vhost_set_vring_base(dev, &state);
1044     if (r) {
1045         VHOST_OPS_DEBUG("vhost_set_vring_base failed");
1046         return -errno;
1047     }
1048 
1049     if (vhost_needs_vring_endian(vdev)) {
1050         r = vhost_virtqueue_set_vring_endian_legacy(dev,
1051                                                     virtio_is_big_endian(vdev),
1052                                                     vhost_vq_index);
1053         if (r) {
1054             return -errno;
1055         }
1056     }
1057 
1058     vq->desc_size = s = l = virtio_queue_get_desc_size(vdev, idx);
1059     vq->desc_phys = a;
1060     vq->desc = vhost_memory_map(dev, a, &l, false);
1061     if (!vq->desc || l != s) {
1062         r = -ENOMEM;
1063         goto fail_alloc_desc;
1064     }
1065     vq->avail_size = s = l = virtio_queue_get_avail_size(vdev, idx);
1066     vq->avail_phys = a = virtio_queue_get_avail_addr(vdev, idx);
1067     vq->avail = vhost_memory_map(dev, a, &l, false);
1068     if (!vq->avail || l != s) {
1069         r = -ENOMEM;
1070         goto fail_alloc_avail;
1071     }
1072     vq->used_size = s = l = virtio_queue_get_used_size(vdev, idx);
1073     vq->used_phys = a = virtio_queue_get_used_addr(vdev, idx);
1074     vq->used = vhost_memory_map(dev, a, &l, true);
1075     if (!vq->used || l != s) {
1076         r = -ENOMEM;
1077         goto fail_alloc_used;
1078     }
1079 
1080     r = vhost_virtqueue_set_addr(dev, vq, vhost_vq_index, dev->log_enabled);
1081     if (r < 0) {
1082         r = -errno;
1083         goto fail_alloc;
1084     }
1085 
1086     file.fd = event_notifier_get_fd(virtio_queue_get_host_notifier(vvq));
1087     r = dev->vhost_ops->vhost_set_vring_kick(dev, &file);
1088     if (r) {
1089         VHOST_OPS_DEBUG("vhost_set_vring_kick failed");
1090         r = -errno;
1091         goto fail_kick;
1092     }
1093 
1094     /* Clear and discard previous events if any. */
1095     event_notifier_test_and_clear(&vq->masked_notifier);
1096 
1097     /* Init vring in unmasked state, unless guest_notifier_mask
1098      * will do it later.
1099      */
1100     if (!vdev->use_guest_notifier_mask) {
1101         /* TODO: check and handle errors. */
1102         vhost_virtqueue_mask(dev, vdev, idx, false);
1103     }
1104 
1105     if (k->query_guest_notifiers &&
1106         k->query_guest_notifiers(qbus->parent) &&
1107         virtio_queue_vector(vdev, idx) == VIRTIO_NO_VECTOR) {
1108         file.fd = -1;
1109         r = dev->vhost_ops->vhost_set_vring_call(dev, &file);
1110         if (r) {
1111             goto fail_vector;
1112         }
1113     }
1114 
1115     return 0;
1116 
1117 fail_vector:
1118 fail_kick:
1119 fail_alloc:
1120     vhost_memory_unmap(dev, vq->used, virtio_queue_get_used_size(vdev, idx),
1121                        0, 0);
1122 fail_alloc_used:
1123     vhost_memory_unmap(dev, vq->avail, virtio_queue_get_avail_size(vdev, idx),
1124                        0, 0);
1125 fail_alloc_avail:
1126     vhost_memory_unmap(dev, vq->desc, virtio_queue_get_desc_size(vdev, idx),
1127                        0, 0);
1128 fail_alloc_desc:
1129     return r;
1130 }
1131 
1132 static void vhost_virtqueue_stop(struct vhost_dev *dev,
1133                                     struct VirtIODevice *vdev,
1134                                     struct vhost_virtqueue *vq,
1135                                     unsigned idx)
1136 {
1137     int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, idx);
1138     struct vhost_vring_state state = {
1139         .index = vhost_vq_index,
1140     };
1141     int r;
1142 
1143     if (virtio_queue_get_desc_addr(vdev, idx) == 0) {
1144         /* Don't stop the virtqueue which might have not been started */
1145         return;
1146     }
1147 
1148     r = dev->vhost_ops->vhost_get_vring_base(dev, &state);
1149     if (r < 0) {
1150         VHOST_OPS_DEBUG("vhost VQ %u ring restore failed: %d", idx, r);
1151         /* Connection to the backend is broken, so let's sync internal
1152          * last avail idx to the device used idx.
1153          */
1154         virtio_queue_restore_last_avail_idx(vdev, idx);
1155     } else {
1156         virtio_queue_set_last_avail_idx(vdev, idx, state.num);
1157     }
1158     virtio_queue_invalidate_signalled_used(vdev, idx);
1159     virtio_queue_update_used_idx(vdev, idx);
1160 
1161     /* In the cross-endian case, we need to reset the vring endianness to
1162      * native as legacy devices expect so by default.
1163      */
1164     if (vhost_needs_vring_endian(vdev)) {
1165         vhost_virtqueue_set_vring_endian_legacy(dev,
1166                                                 !virtio_is_big_endian(vdev),
1167                                                 vhost_vq_index);
1168     }
1169 
1170     vhost_memory_unmap(dev, vq->used, virtio_queue_get_used_size(vdev, idx),
1171                        1, virtio_queue_get_used_size(vdev, idx));
1172     vhost_memory_unmap(dev, vq->avail, virtio_queue_get_avail_size(vdev, idx),
1173                        0, virtio_queue_get_avail_size(vdev, idx));
1174     vhost_memory_unmap(dev, vq->desc, virtio_queue_get_desc_size(vdev, idx),
1175                        0, virtio_queue_get_desc_size(vdev, idx));
1176 }
1177 
1178 static void vhost_eventfd_add(MemoryListener *listener,
1179                               MemoryRegionSection *section,
1180                               bool match_data, uint64_t data, EventNotifier *e)
1181 {
1182 }
1183 
1184 static void vhost_eventfd_del(MemoryListener *listener,
1185                               MemoryRegionSection *section,
1186                               bool match_data, uint64_t data, EventNotifier *e)
1187 {
1188 }
1189 
1190 static int vhost_virtqueue_set_busyloop_timeout(struct vhost_dev *dev,
1191                                                 int n, uint32_t timeout)
1192 {
1193     int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, n);
1194     struct vhost_vring_state state = {
1195         .index = vhost_vq_index,
1196         .num = timeout,
1197     };
1198     int r;
1199 
1200     if (!dev->vhost_ops->vhost_set_vring_busyloop_timeout) {
1201         return -EINVAL;
1202     }
1203 
1204     r = dev->vhost_ops->vhost_set_vring_busyloop_timeout(dev, &state);
1205     if (r) {
1206         VHOST_OPS_DEBUG("vhost_set_vring_busyloop_timeout failed");
1207         return r;
1208     }
1209 
1210     return 0;
1211 }
1212 
1213 static int vhost_virtqueue_init(struct vhost_dev *dev,
1214                                 struct vhost_virtqueue *vq, int n)
1215 {
1216     int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, n);
1217     struct vhost_vring_file file = {
1218         .index = vhost_vq_index,
1219     };
1220     int r = event_notifier_init(&vq->masked_notifier, 0);
1221     if (r < 0) {
1222         return r;
1223     }
1224 
1225     file.fd = event_notifier_get_fd(&vq->masked_notifier);
1226     r = dev->vhost_ops->vhost_set_vring_call(dev, &file);
1227     if (r) {
1228         VHOST_OPS_DEBUG("vhost_set_vring_call failed");
1229         r = -errno;
1230         goto fail_call;
1231     }
1232 
1233     vq->dev = dev;
1234 
1235     return 0;
1236 fail_call:
1237     event_notifier_cleanup(&vq->masked_notifier);
1238     return r;
1239 }
1240 
1241 static void vhost_virtqueue_cleanup(struct vhost_virtqueue *vq)
1242 {
1243     event_notifier_cleanup(&vq->masked_notifier);
1244 }
1245 
1246 int vhost_dev_init(struct vhost_dev *hdev, void *opaque,
1247                    VhostBackendType backend_type, uint32_t busyloop_timeout)
1248 {
1249     uint64_t features;
1250     int i, r, n_initialized_vqs = 0;
1251     Error *local_err = NULL;
1252 
1253     hdev->vdev = NULL;
1254     hdev->migration_blocker = NULL;
1255 
1256     r = vhost_set_backend_type(hdev, backend_type);
1257     assert(r >= 0);
1258 
1259     r = hdev->vhost_ops->vhost_backend_init(hdev, opaque);
1260     if (r < 0) {
1261         goto fail;
1262     }
1263 
1264     r = hdev->vhost_ops->vhost_set_owner(hdev);
1265     if (r < 0) {
1266         VHOST_OPS_DEBUG("vhost_set_owner failed");
1267         goto fail;
1268     }
1269 
1270     r = hdev->vhost_ops->vhost_get_features(hdev, &features);
1271     if (r < 0) {
1272         VHOST_OPS_DEBUG("vhost_get_features failed");
1273         goto fail;
1274     }
1275 
1276     for (i = 0; i < hdev->nvqs; ++i, ++n_initialized_vqs) {
1277         r = vhost_virtqueue_init(hdev, hdev->vqs + i, hdev->vq_index + i);
1278         if (r < 0) {
1279             goto fail;
1280         }
1281     }
1282 
1283     if (busyloop_timeout) {
1284         for (i = 0; i < hdev->nvqs; ++i) {
1285             r = vhost_virtqueue_set_busyloop_timeout(hdev, hdev->vq_index + i,
1286                                                      busyloop_timeout);
1287             if (r < 0) {
1288                 goto fail_busyloop;
1289             }
1290         }
1291     }
1292 
1293     hdev->features = features;
1294 
1295     hdev->memory_listener = (MemoryListener) {
1296         .begin = vhost_begin,
1297         .commit = vhost_commit,
1298         .region_add = vhost_region_addnop,
1299         .region_nop = vhost_region_addnop,
1300         .log_start = vhost_log_start,
1301         .log_stop = vhost_log_stop,
1302         .log_sync = vhost_log_sync,
1303         .log_global_start = vhost_log_global_start,
1304         .log_global_stop = vhost_log_global_stop,
1305         .eventfd_add = vhost_eventfd_add,
1306         .eventfd_del = vhost_eventfd_del,
1307         .priority = 10
1308     };
1309 
1310     hdev->iommu_listener = (MemoryListener) {
1311         .region_add = vhost_iommu_region_add,
1312         .region_del = vhost_iommu_region_del,
1313     };
1314 
1315     if (hdev->migration_blocker == NULL) {
1316         if (!(hdev->features & (0x1ULL << VHOST_F_LOG_ALL))) {
1317             error_setg(&hdev->migration_blocker,
1318                        "Migration disabled: vhost lacks VHOST_F_LOG_ALL feature.");
1319         } else if (vhost_dev_log_is_shared(hdev) && !qemu_memfd_alloc_check()) {
1320             error_setg(&hdev->migration_blocker,
1321                        "Migration disabled: failed to allocate shared memory");
1322         }
1323     }
1324 
1325     if (hdev->migration_blocker != NULL) {
1326         r = migrate_add_blocker(hdev->migration_blocker, &local_err);
1327         if (local_err) {
1328             error_report_err(local_err);
1329             error_free(hdev->migration_blocker);
1330             goto fail_busyloop;
1331         }
1332     }
1333 
1334     hdev->mem = g_malloc0(offsetof(struct vhost_memory, regions));
1335     hdev->n_mem_sections = 0;
1336     hdev->mem_sections = NULL;
1337     hdev->log = NULL;
1338     hdev->log_size = 0;
1339     hdev->log_enabled = false;
1340     hdev->started = false;
1341     memory_listener_register(&hdev->memory_listener, &address_space_memory);
1342     QLIST_INSERT_HEAD(&vhost_devices, hdev, entry);
1343 
1344     if (used_memslots > hdev->vhost_ops->vhost_backend_memslots_limit(hdev)) {
1345         error_report("vhost backend memory slots limit is less"
1346                 " than current number of present memory slots");
1347         r = -1;
1348         if (busyloop_timeout) {
1349             goto fail_busyloop;
1350         } else {
1351             goto fail;
1352         }
1353     }
1354 
1355     return 0;
1356 
1357 fail_busyloop:
1358     while (--i >= 0) {
1359         vhost_virtqueue_set_busyloop_timeout(hdev, hdev->vq_index + i, 0);
1360     }
1361 fail:
1362     hdev->nvqs = n_initialized_vqs;
1363     vhost_dev_cleanup(hdev);
1364     return r;
1365 }
1366 
1367 void vhost_dev_cleanup(struct vhost_dev *hdev)
1368 {
1369     int i;
1370 
1371     for (i = 0; i < hdev->nvqs; ++i) {
1372         vhost_virtqueue_cleanup(hdev->vqs + i);
1373     }
1374     if (hdev->mem) {
1375         /* those are only safe after successful init */
1376         memory_listener_unregister(&hdev->memory_listener);
1377         QLIST_REMOVE(hdev, entry);
1378     }
1379     if (hdev->migration_blocker) {
1380         migrate_del_blocker(hdev->migration_blocker);
1381         error_free(hdev->migration_blocker);
1382     }
1383     g_free(hdev->mem);
1384     g_free(hdev->mem_sections);
1385     if (hdev->vhost_ops) {
1386         hdev->vhost_ops->vhost_backend_cleanup(hdev);
1387     }
1388     assert(!hdev->log);
1389 
1390     memset(hdev, 0, sizeof(struct vhost_dev));
1391 }
1392 
1393 /* Stop processing guest IO notifications in qemu.
1394  * Start processing them in vhost in kernel.
1395  */
1396 int vhost_dev_enable_notifiers(struct vhost_dev *hdev, VirtIODevice *vdev)
1397 {
1398     BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev)));
1399     int i, r, e;
1400 
1401     /* We will pass the notifiers to the kernel, make sure that QEMU
1402      * doesn't interfere.
1403      */
1404     r = virtio_device_grab_ioeventfd(vdev);
1405     if (r < 0) {
1406         error_report("binding does not support host notifiers");
1407         goto fail;
1408     }
1409 
1410     for (i = 0; i < hdev->nvqs; ++i) {
1411         r = virtio_bus_set_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i,
1412                                          true);
1413         if (r < 0) {
1414             error_report("vhost VQ %d notifier binding failed: %d", i, -r);
1415             goto fail_vq;
1416         }
1417     }
1418 
1419     return 0;
1420 fail_vq:
1421     while (--i >= 0) {
1422         e = virtio_bus_set_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i,
1423                                          false);
1424         if (e < 0) {
1425             error_report("vhost VQ %d notifier cleanup error: %d", i, -r);
1426         }
1427         assert (e >= 0);
1428         virtio_bus_cleanup_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i);
1429     }
1430     virtio_device_release_ioeventfd(vdev);
1431 fail:
1432     return r;
1433 }
1434 
1435 /* Stop processing guest IO notifications in vhost.
1436  * Start processing them in qemu.
1437  * This might actually run the qemu handlers right away,
1438  * so virtio in qemu must be completely setup when this is called.
1439  */
1440 void vhost_dev_disable_notifiers(struct vhost_dev *hdev, VirtIODevice *vdev)
1441 {
1442     BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev)));
1443     int i, r;
1444 
1445     for (i = 0; i < hdev->nvqs; ++i) {
1446         r = virtio_bus_set_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i,
1447                                          false);
1448         if (r < 0) {
1449             error_report("vhost VQ %d notifier cleanup failed: %d", i, -r);
1450         }
1451         assert (r >= 0);
1452         virtio_bus_cleanup_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i);
1453     }
1454     virtio_device_release_ioeventfd(vdev);
1455 }
1456 
1457 /* Test and clear event pending status.
1458  * Should be called after unmask to avoid losing events.
1459  */
1460 bool vhost_virtqueue_pending(struct vhost_dev *hdev, int n)
1461 {
1462     struct vhost_virtqueue *vq = hdev->vqs + n - hdev->vq_index;
1463     assert(n >= hdev->vq_index && n < hdev->vq_index + hdev->nvqs);
1464     return event_notifier_test_and_clear(&vq->masked_notifier);
1465 }
1466 
1467 /* Mask/unmask events from this vq. */
1468 void vhost_virtqueue_mask(struct vhost_dev *hdev, VirtIODevice *vdev, int n,
1469                          bool mask)
1470 {
1471     struct VirtQueue *vvq = virtio_get_queue(vdev, n);
1472     int r, index = n - hdev->vq_index;
1473     struct vhost_vring_file file;
1474 
1475     /* should only be called after backend is connected */
1476     assert(hdev->vhost_ops);
1477 
1478     if (mask) {
1479         assert(vdev->use_guest_notifier_mask);
1480         file.fd = event_notifier_get_fd(&hdev->vqs[index].masked_notifier);
1481     } else {
1482         file.fd = event_notifier_get_fd(virtio_queue_get_guest_notifier(vvq));
1483     }
1484 
1485     file.index = hdev->vhost_ops->vhost_get_vq_index(hdev, n);
1486     r = hdev->vhost_ops->vhost_set_vring_call(hdev, &file);
1487     if (r < 0) {
1488         VHOST_OPS_DEBUG("vhost_set_vring_call failed");
1489     }
1490 }
1491 
1492 uint64_t vhost_get_features(struct vhost_dev *hdev, const int *feature_bits,
1493                             uint64_t features)
1494 {
1495     const int *bit = feature_bits;
1496     while (*bit != VHOST_INVALID_FEATURE_BIT) {
1497         uint64_t bit_mask = (1ULL << *bit);
1498         if (!(hdev->features & bit_mask)) {
1499             features &= ~bit_mask;
1500         }
1501         bit++;
1502     }
1503     return features;
1504 }
1505 
1506 void vhost_ack_features(struct vhost_dev *hdev, const int *feature_bits,
1507                         uint64_t features)
1508 {
1509     const int *bit = feature_bits;
1510     while (*bit != VHOST_INVALID_FEATURE_BIT) {
1511         uint64_t bit_mask = (1ULL << *bit);
1512         if (features & bit_mask) {
1513             hdev->acked_features |= bit_mask;
1514         }
1515         bit++;
1516     }
1517 }
1518 
1519 int vhost_dev_get_config(struct vhost_dev *hdev, uint8_t *config,
1520                          uint32_t config_len)
1521 {
1522     assert(hdev->vhost_ops);
1523 
1524     if (hdev->vhost_ops->vhost_get_config) {
1525         return hdev->vhost_ops->vhost_get_config(hdev, config, config_len);
1526     }
1527 
1528     return -1;
1529 }
1530 
1531 int vhost_dev_set_config(struct vhost_dev *hdev, const uint8_t *data,
1532                          uint32_t offset, uint32_t size, uint32_t flags)
1533 {
1534     assert(hdev->vhost_ops);
1535 
1536     if (hdev->vhost_ops->vhost_set_config) {
1537         return hdev->vhost_ops->vhost_set_config(hdev, data, offset,
1538                                                  size, flags);
1539     }
1540 
1541     return -1;
1542 }
1543 
1544 void vhost_dev_set_config_notifier(struct vhost_dev *hdev,
1545                                    const VhostDevConfigOps *ops)
1546 {
1547     hdev->config_ops = ops;
1548 }
1549 
1550 void vhost_dev_free_inflight(struct vhost_inflight *inflight)
1551 {
1552     if (inflight && inflight->addr) {
1553         qemu_memfd_free(inflight->addr, inflight->size, inflight->fd);
1554         inflight->addr = NULL;
1555         inflight->fd = -1;
1556     }
1557 }
1558 
1559 static int vhost_dev_resize_inflight(struct vhost_inflight *inflight,
1560                                      uint64_t new_size)
1561 {
1562     Error *err = NULL;
1563     int fd = -1;
1564     void *addr = qemu_memfd_alloc("vhost-inflight", new_size,
1565                                   F_SEAL_GROW | F_SEAL_SHRINK | F_SEAL_SEAL,
1566                                   &fd, &err);
1567 
1568     if (err) {
1569         error_report_err(err);
1570         return -1;
1571     }
1572 
1573     vhost_dev_free_inflight(inflight);
1574     inflight->offset = 0;
1575     inflight->addr = addr;
1576     inflight->fd = fd;
1577     inflight->size = new_size;
1578 
1579     return 0;
1580 }
1581 
1582 void vhost_dev_save_inflight(struct vhost_inflight *inflight, QEMUFile *f)
1583 {
1584     if (inflight->addr) {
1585         qemu_put_be64(f, inflight->size);
1586         qemu_put_be16(f, inflight->queue_size);
1587         qemu_put_buffer(f, inflight->addr, inflight->size);
1588     } else {
1589         qemu_put_be64(f, 0);
1590     }
1591 }
1592 
1593 int vhost_dev_load_inflight(struct vhost_inflight *inflight, QEMUFile *f)
1594 {
1595     uint64_t size;
1596 
1597     size = qemu_get_be64(f);
1598     if (!size) {
1599         return 0;
1600     }
1601 
1602     if (inflight->size != size) {
1603         if (vhost_dev_resize_inflight(inflight, size)) {
1604             return -1;
1605         }
1606     }
1607     inflight->queue_size = qemu_get_be16(f);
1608 
1609     qemu_get_buffer(f, inflight->addr, size);
1610 
1611     return 0;
1612 }
1613 
1614 int vhost_dev_set_inflight(struct vhost_dev *dev,
1615                            struct vhost_inflight *inflight)
1616 {
1617     int r;
1618 
1619     if (dev->vhost_ops->vhost_set_inflight_fd && inflight->addr) {
1620         r = dev->vhost_ops->vhost_set_inflight_fd(dev, inflight);
1621         if (r) {
1622             VHOST_OPS_DEBUG("vhost_set_inflight_fd failed");
1623             return -errno;
1624         }
1625     }
1626 
1627     return 0;
1628 }
1629 
1630 int vhost_dev_get_inflight(struct vhost_dev *dev, uint16_t queue_size,
1631                            struct vhost_inflight *inflight)
1632 {
1633     int r;
1634 
1635     if (dev->vhost_ops->vhost_get_inflight_fd) {
1636         r = dev->vhost_ops->vhost_get_inflight_fd(dev, queue_size, inflight);
1637         if (r) {
1638             VHOST_OPS_DEBUG("vhost_get_inflight_fd failed");
1639             return -errno;
1640         }
1641     }
1642 
1643     return 0;
1644 }
1645 
1646 /* Host notifiers must be enabled at this point. */
1647 int vhost_dev_start(struct vhost_dev *hdev, VirtIODevice *vdev)
1648 {
1649     int i, r;
1650 
1651     /* should only be called after backend is connected */
1652     assert(hdev->vhost_ops);
1653 
1654     hdev->started = true;
1655     hdev->vdev = vdev;
1656 
1657     r = vhost_dev_set_features(hdev, hdev->log_enabled);
1658     if (r < 0) {
1659         goto fail_features;
1660     }
1661 
1662     if (vhost_dev_has_iommu(hdev)) {
1663         memory_listener_register(&hdev->iommu_listener, vdev->dma_as);
1664     }
1665 
1666     r = hdev->vhost_ops->vhost_set_mem_table(hdev, hdev->mem);
1667     if (r < 0) {
1668         VHOST_OPS_DEBUG("vhost_set_mem_table failed");
1669         r = -errno;
1670         goto fail_mem;
1671     }
1672     for (i = 0; i < hdev->nvqs; ++i) {
1673         r = vhost_virtqueue_start(hdev,
1674                                   vdev,
1675                                   hdev->vqs + i,
1676                                   hdev->vq_index + i);
1677         if (r < 0) {
1678             goto fail_vq;
1679         }
1680     }
1681 
1682     if (hdev->log_enabled) {
1683         uint64_t log_base;
1684 
1685         hdev->log_size = vhost_get_log_size(hdev);
1686         hdev->log = vhost_log_get(hdev->log_size,
1687                                   vhost_dev_log_is_shared(hdev));
1688         log_base = (uintptr_t)hdev->log->log;
1689         r = hdev->vhost_ops->vhost_set_log_base(hdev,
1690                                                 hdev->log_size ? log_base : 0,
1691                                                 hdev->log);
1692         if (r < 0) {
1693             VHOST_OPS_DEBUG("vhost_set_log_base failed");
1694             r = -errno;
1695             goto fail_log;
1696         }
1697     }
1698     if (hdev->vhost_ops->vhost_dev_start) {
1699         r = hdev->vhost_ops->vhost_dev_start(hdev, true);
1700         if (r) {
1701             goto fail_log;
1702         }
1703     }
1704     if (vhost_dev_has_iommu(hdev) &&
1705         hdev->vhost_ops->vhost_set_iotlb_callback) {
1706             hdev->vhost_ops->vhost_set_iotlb_callback(hdev, true);
1707 
1708         /* Update used ring information for IOTLB to work correctly,
1709          * vhost-kernel code requires for this.*/
1710         for (i = 0; i < hdev->nvqs; ++i) {
1711             struct vhost_virtqueue *vq = hdev->vqs + i;
1712             vhost_device_iotlb_miss(hdev, vq->used_phys, true);
1713         }
1714     }
1715     return 0;
1716 fail_log:
1717     vhost_log_put(hdev, false);
1718 fail_vq:
1719     while (--i >= 0) {
1720         vhost_virtqueue_stop(hdev,
1721                              vdev,
1722                              hdev->vqs + i,
1723                              hdev->vq_index + i);
1724     }
1725 
1726 fail_mem:
1727 fail_features:
1728 
1729     hdev->started = false;
1730     return r;
1731 }
1732 
1733 /* Host notifiers must be enabled at this point. */
1734 void vhost_dev_stop(struct vhost_dev *hdev, VirtIODevice *vdev)
1735 {
1736     int i;
1737 
1738     /* should only be called after backend is connected */
1739     assert(hdev->vhost_ops);
1740 
1741     if (hdev->vhost_ops->vhost_dev_start) {
1742         hdev->vhost_ops->vhost_dev_start(hdev, false);
1743     }
1744     for (i = 0; i < hdev->nvqs; ++i) {
1745         vhost_virtqueue_stop(hdev,
1746                              vdev,
1747                              hdev->vqs + i,
1748                              hdev->vq_index + i);
1749     }
1750 
1751     if (vhost_dev_has_iommu(hdev)) {
1752         if (hdev->vhost_ops->vhost_set_iotlb_callback) {
1753             hdev->vhost_ops->vhost_set_iotlb_callback(hdev, false);
1754         }
1755         memory_listener_unregister(&hdev->iommu_listener);
1756     }
1757     vhost_log_put(hdev, true);
1758     hdev->started = false;
1759     hdev->vdev = NULL;
1760 }
1761 
1762 int vhost_net_set_backend(struct vhost_dev *hdev,
1763                           struct vhost_vring_file *file)
1764 {
1765     if (hdev->vhost_ops->vhost_net_set_backend) {
1766         return hdev->vhost_ops->vhost_net_set_backend(hdev, file);
1767     }
1768 
1769     return -1;
1770 }
1771