xref: /openbmc/qemu/hw/virtio/vhost.c (revision 6016b7b4)
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 "hw/virtio/virtio-bus.h"
25 #include "hw/virtio/virtio-access.h"
26 #include "migration/blocker.h"
27 #include "migration/qemu-file-types.h"
28 #include "sysemu/dma.h"
29 #include "sysemu/tcg.h"
30 #include "trace.h"
31 
32 /* enabled until disconnected backend stabilizes */
33 #define _VHOST_DEBUG 1
34 
35 #ifdef _VHOST_DEBUG
36 #define VHOST_OPS_DEBUG(fmt, ...) \
37     do { error_report(fmt ": %s (%d)", ## __VA_ARGS__, \
38                       strerror(errno), errno); } while (0)
39 #else
40 #define VHOST_OPS_DEBUG(fmt, ...) \
41     do { } while (0)
42 #endif
43 
44 static struct vhost_log *vhost_log;
45 static struct vhost_log *vhost_log_shm;
46 
47 static unsigned int used_memslots;
48 static QLIST_HEAD(, vhost_dev) vhost_devices =
49     QLIST_HEAD_INITIALIZER(vhost_devices);
50 
51 bool vhost_has_free_slot(void)
52 {
53     unsigned int slots_limit = ~0U;
54     struct vhost_dev *hdev;
55 
56     QLIST_FOREACH(hdev, &vhost_devices, entry) {
57         unsigned int r = hdev->vhost_ops->vhost_backend_memslots_limit(hdev);
58         slots_limit = MIN(slots_limit, r);
59     }
60     return slots_limit > used_memslots;
61 }
62 
63 static void vhost_dev_sync_region(struct vhost_dev *dev,
64                                   MemoryRegionSection *section,
65                                   uint64_t mfirst, uint64_t mlast,
66                                   uint64_t rfirst, uint64_t rlast)
67 {
68     vhost_log_chunk_t *log = dev->log->log;
69 
70     uint64_t start = MAX(mfirst, rfirst);
71     uint64_t end = MIN(mlast, rlast);
72     vhost_log_chunk_t *from = log + start / VHOST_LOG_CHUNK;
73     vhost_log_chunk_t *to = log + end / VHOST_LOG_CHUNK + 1;
74     uint64_t addr = QEMU_ALIGN_DOWN(start, VHOST_LOG_CHUNK);
75 
76     if (end < start) {
77         return;
78     }
79     assert(end / VHOST_LOG_CHUNK < dev->log_size);
80     assert(start / VHOST_LOG_CHUNK < dev->log_size);
81 
82     for (;from < to; ++from) {
83         vhost_log_chunk_t log;
84         /* We first check with non-atomic: much cheaper,
85          * and we expect non-dirty to be the common case. */
86         if (!*from) {
87             addr += VHOST_LOG_CHUNK;
88             continue;
89         }
90         /* Data must be read atomically. We don't really need barrier semantics
91          * but it's easier to use atomic_* than roll our own. */
92         log = qatomic_xchg(from, 0);
93         while (log) {
94             int bit = ctzl(log);
95             hwaddr page_addr;
96             hwaddr section_offset;
97             hwaddr mr_offset;
98             page_addr = addr + bit * VHOST_LOG_PAGE;
99             section_offset = page_addr - section->offset_within_address_space;
100             mr_offset = section_offset + section->offset_within_region;
101             memory_region_set_dirty(section->mr, mr_offset, VHOST_LOG_PAGE);
102             log &= ~(0x1ull << bit);
103         }
104         addr += VHOST_LOG_CHUNK;
105     }
106 }
107 
108 static int vhost_sync_dirty_bitmap(struct vhost_dev *dev,
109                                    MemoryRegionSection *section,
110                                    hwaddr first,
111                                    hwaddr last)
112 {
113     int i;
114     hwaddr start_addr;
115     hwaddr end_addr;
116 
117     if (!dev->log_enabled || !dev->started) {
118         return 0;
119     }
120     start_addr = section->offset_within_address_space;
121     end_addr = range_get_last(start_addr, int128_get64(section->size));
122     start_addr = MAX(first, start_addr);
123     end_addr = MIN(last, end_addr);
124 
125     for (i = 0; i < dev->mem->nregions; ++i) {
126         struct vhost_memory_region *reg = dev->mem->regions + i;
127         vhost_dev_sync_region(dev, section, start_addr, end_addr,
128                               reg->guest_phys_addr,
129                               range_get_last(reg->guest_phys_addr,
130                                              reg->memory_size));
131     }
132     for (i = 0; i < dev->nvqs; ++i) {
133         struct vhost_virtqueue *vq = dev->vqs + i;
134 
135         if (!vq->used_phys && !vq->used_size) {
136             continue;
137         }
138 
139         vhost_dev_sync_region(dev, section, start_addr, end_addr, vq->used_phys,
140                               range_get_last(vq->used_phys, vq->used_size));
141     }
142     return 0;
143 }
144 
145 static void vhost_log_sync(MemoryListener *listener,
146                           MemoryRegionSection *section)
147 {
148     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
149                                          memory_listener);
150     vhost_sync_dirty_bitmap(dev, section, 0x0, ~0x0ULL);
151 }
152 
153 static void vhost_log_sync_range(struct vhost_dev *dev,
154                                  hwaddr first, hwaddr last)
155 {
156     int i;
157     /* FIXME: this is N^2 in number of sections */
158     for (i = 0; i < dev->n_mem_sections; ++i) {
159         MemoryRegionSection *section = &dev->mem_sections[i];
160         vhost_sync_dirty_bitmap(dev, section, first, last);
161     }
162 }
163 
164 static uint64_t vhost_get_log_size(struct vhost_dev *dev)
165 {
166     uint64_t log_size = 0;
167     int i;
168     for (i = 0; i < dev->mem->nregions; ++i) {
169         struct vhost_memory_region *reg = dev->mem->regions + i;
170         uint64_t last = range_get_last(reg->guest_phys_addr,
171                                        reg->memory_size);
172         log_size = MAX(log_size, last / VHOST_LOG_CHUNK + 1);
173     }
174     return log_size;
175 }
176 
177 static int vhost_set_backend_type(struct vhost_dev *dev,
178                                   VhostBackendType backend_type)
179 {
180     int r = 0;
181 
182     switch (backend_type) {
183 #ifdef CONFIG_VHOST_KERNEL
184     case VHOST_BACKEND_TYPE_KERNEL:
185         dev->vhost_ops = &kernel_ops;
186         break;
187 #endif
188 #ifdef CONFIG_VHOST_USER
189     case VHOST_BACKEND_TYPE_USER:
190         dev->vhost_ops = &user_ops;
191         break;
192 #endif
193 #ifdef CONFIG_VHOST_VDPA
194     case VHOST_BACKEND_TYPE_VDPA:
195         dev->vhost_ops = &vdpa_ops;
196         break;
197 #endif
198     default:
199         error_report("Unknown vhost backend type");
200         r = -1;
201     }
202 
203     return r;
204 }
205 
206 static struct vhost_log *vhost_log_alloc(uint64_t size, bool share)
207 {
208     Error *err = NULL;
209     struct vhost_log *log;
210     uint64_t logsize = size * sizeof(*(log->log));
211     int fd = -1;
212 
213     log = g_new0(struct vhost_log, 1);
214     if (share) {
215         log->log = qemu_memfd_alloc("vhost-log", logsize,
216                                     F_SEAL_GROW | F_SEAL_SHRINK | F_SEAL_SEAL,
217                                     &fd, &err);
218         if (err) {
219             error_report_err(err);
220             g_free(log);
221             return NULL;
222         }
223         memset(log->log, 0, logsize);
224     } else {
225         log->log = g_malloc0(logsize);
226     }
227 
228     log->size = size;
229     log->refcnt = 1;
230     log->fd = fd;
231 
232     return log;
233 }
234 
235 static struct vhost_log *vhost_log_get(uint64_t size, bool share)
236 {
237     struct vhost_log *log = share ? vhost_log_shm : vhost_log;
238 
239     if (!log || log->size != size) {
240         log = vhost_log_alloc(size, share);
241         if (share) {
242             vhost_log_shm = log;
243         } else {
244             vhost_log = log;
245         }
246     } else {
247         ++log->refcnt;
248     }
249 
250     return log;
251 }
252 
253 static void vhost_log_put(struct vhost_dev *dev, bool sync)
254 {
255     struct vhost_log *log = dev->log;
256 
257     if (!log) {
258         return;
259     }
260 
261     --log->refcnt;
262     if (log->refcnt == 0) {
263         /* Sync only the range covered by the old log */
264         if (dev->log_size && sync) {
265             vhost_log_sync_range(dev, 0, dev->log_size * VHOST_LOG_CHUNK - 1);
266         }
267 
268         if (vhost_log == log) {
269             g_free(log->log);
270             vhost_log = NULL;
271         } else if (vhost_log_shm == log) {
272             qemu_memfd_free(log->log, log->size * sizeof(*(log->log)),
273                             log->fd);
274             vhost_log_shm = NULL;
275         }
276 
277         g_free(log);
278     }
279 
280     dev->log = NULL;
281     dev->log_size = 0;
282 }
283 
284 static bool vhost_dev_log_is_shared(struct vhost_dev *dev)
285 {
286     return dev->vhost_ops->vhost_requires_shm_log &&
287            dev->vhost_ops->vhost_requires_shm_log(dev);
288 }
289 
290 static inline void vhost_dev_log_resize(struct vhost_dev *dev, uint64_t size)
291 {
292     struct vhost_log *log = vhost_log_get(size, vhost_dev_log_is_shared(dev));
293     uint64_t log_base = (uintptr_t)log->log;
294     int r;
295 
296     /* inform backend of log switching, this must be done before
297        releasing the current log, to ensure no logging is lost */
298     r = dev->vhost_ops->vhost_set_log_base(dev, log_base, log);
299     if (r < 0) {
300         VHOST_OPS_DEBUG("vhost_set_log_base failed");
301     }
302 
303     vhost_log_put(dev, true);
304     dev->log = log;
305     dev->log_size = size;
306 }
307 
308 static int vhost_dev_has_iommu(struct vhost_dev *dev)
309 {
310     VirtIODevice *vdev = dev->vdev;
311 
312     /*
313      * For vhost, VIRTIO_F_IOMMU_PLATFORM means the backend support
314      * incremental memory mapping API via IOTLB API. For platform that
315      * does not have IOMMU, there's no need to enable this feature
316      * which may cause unnecessary IOTLB miss/update transactions.
317      */
318     return virtio_bus_device_iommu_enabled(vdev) &&
319            virtio_host_has_feature(vdev, VIRTIO_F_IOMMU_PLATFORM);
320 }
321 
322 static void *vhost_memory_map(struct vhost_dev *dev, hwaddr addr,
323                               hwaddr *plen, bool is_write)
324 {
325     if (!vhost_dev_has_iommu(dev)) {
326         return cpu_physical_memory_map(addr, plen, is_write);
327     } else {
328         return (void *)(uintptr_t)addr;
329     }
330 }
331 
332 static void vhost_memory_unmap(struct vhost_dev *dev, void *buffer,
333                                hwaddr len, int is_write,
334                                hwaddr access_len)
335 {
336     if (!vhost_dev_has_iommu(dev)) {
337         cpu_physical_memory_unmap(buffer, len, is_write, access_len);
338     }
339 }
340 
341 static int vhost_verify_ring_part_mapping(void *ring_hva,
342                                           uint64_t ring_gpa,
343                                           uint64_t ring_size,
344                                           void *reg_hva,
345                                           uint64_t reg_gpa,
346                                           uint64_t reg_size)
347 {
348     uint64_t hva_ring_offset;
349     uint64_t ring_last = range_get_last(ring_gpa, ring_size);
350     uint64_t reg_last = range_get_last(reg_gpa, reg_size);
351 
352     if (ring_last < reg_gpa || ring_gpa > reg_last) {
353         return 0;
354     }
355     /* check that whole ring's is mapped */
356     if (ring_last > reg_last) {
357         return -ENOMEM;
358     }
359     /* check that ring's MemoryRegion wasn't replaced */
360     hva_ring_offset = ring_gpa - reg_gpa;
361     if (ring_hva != reg_hva + hva_ring_offset) {
362         return -EBUSY;
363     }
364 
365     return 0;
366 }
367 
368 static int vhost_verify_ring_mappings(struct vhost_dev *dev,
369                                       void *reg_hva,
370                                       uint64_t reg_gpa,
371                                       uint64_t reg_size)
372 {
373     int i, j;
374     int r = 0;
375     const char *part_name[] = {
376         "descriptor table",
377         "available ring",
378         "used ring"
379     };
380 
381     if (vhost_dev_has_iommu(dev)) {
382         return 0;
383     }
384 
385     for (i = 0; i < dev->nvqs; ++i) {
386         struct vhost_virtqueue *vq = dev->vqs + i;
387 
388         if (vq->desc_phys == 0) {
389             continue;
390         }
391 
392         j = 0;
393         r = vhost_verify_ring_part_mapping(
394                 vq->desc, vq->desc_phys, vq->desc_size,
395                 reg_hva, reg_gpa, reg_size);
396         if (r) {
397             break;
398         }
399 
400         j++;
401         r = vhost_verify_ring_part_mapping(
402                 vq->avail, vq->avail_phys, vq->avail_size,
403                 reg_hva, reg_gpa, reg_size);
404         if (r) {
405             break;
406         }
407 
408         j++;
409         r = vhost_verify_ring_part_mapping(
410                 vq->used, vq->used_phys, vq->used_size,
411                 reg_hva, reg_gpa, reg_size);
412         if (r) {
413             break;
414         }
415     }
416 
417     if (r == -ENOMEM) {
418         error_report("Unable to map %s for ring %d", part_name[j], i);
419     } else if (r == -EBUSY) {
420         error_report("%s relocated for ring %d", part_name[j], i);
421     }
422     return r;
423 }
424 
425 /*
426  * vhost_section: identify sections needed for vhost access
427  *
428  * We only care about RAM sections here (where virtqueue and guest
429  * internals accessed by virtio might live). If we find one we still
430  * allow the backend to potentially filter it out of our list.
431  */
432 static bool vhost_section(struct vhost_dev *dev, MemoryRegionSection *section)
433 {
434     MemoryRegion *mr = section->mr;
435 
436     if (memory_region_is_ram(mr) && !memory_region_is_rom(mr)) {
437         uint8_t dirty_mask = memory_region_get_dirty_log_mask(mr);
438         uint8_t handled_dirty;
439 
440         /*
441          * Kernel based vhost doesn't handle any block which is doing
442          * dirty-tracking other than migration for which it has
443          * specific logging support. However for TCG the kernel never
444          * gets involved anyway so we can also ignore it's
445          * self-modiying code detection flags. However a vhost-user
446          * client could still confuse a TCG guest if it re-writes
447          * executable memory that has already been translated.
448          */
449         handled_dirty = (1 << DIRTY_MEMORY_MIGRATION) |
450             (1 << DIRTY_MEMORY_CODE);
451 
452         if (dirty_mask & ~handled_dirty) {
453             trace_vhost_reject_section(mr->name, 1);
454             return false;
455         }
456 
457         if (dev->vhost_ops->vhost_backend_mem_section_filter &&
458             !dev->vhost_ops->vhost_backend_mem_section_filter(dev, section)) {
459             trace_vhost_reject_section(mr->name, 2);
460             return false;
461         }
462 
463         trace_vhost_section(mr->name);
464         return true;
465     } else {
466         trace_vhost_reject_section(mr->name, 3);
467         return false;
468     }
469 }
470 
471 static void vhost_begin(MemoryListener *listener)
472 {
473     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
474                                          memory_listener);
475     dev->tmp_sections = NULL;
476     dev->n_tmp_sections = 0;
477 }
478 
479 static void vhost_commit(MemoryListener *listener)
480 {
481     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
482                                          memory_listener);
483     MemoryRegionSection *old_sections;
484     int n_old_sections;
485     uint64_t log_size;
486     size_t regions_size;
487     int r;
488     int i;
489     bool changed = false;
490 
491     /* Note we can be called before the device is started, but then
492      * starting the device calls set_mem_table, so we need to have
493      * built the data structures.
494      */
495     old_sections = dev->mem_sections;
496     n_old_sections = dev->n_mem_sections;
497     dev->mem_sections = dev->tmp_sections;
498     dev->n_mem_sections = dev->n_tmp_sections;
499 
500     if (dev->n_mem_sections != n_old_sections) {
501         changed = true;
502     } else {
503         /* Same size, lets check the contents */
504         for (int i = 0; i < n_old_sections; i++) {
505             if (!MemoryRegionSection_eq(&old_sections[i],
506                                         &dev->mem_sections[i])) {
507                 changed = true;
508                 break;
509             }
510         }
511     }
512 
513     trace_vhost_commit(dev->started, changed);
514     if (!changed) {
515         goto out;
516     }
517 
518     /* Rebuild the regions list from the new sections list */
519     regions_size = offsetof(struct vhost_memory, regions) +
520                        dev->n_mem_sections * sizeof dev->mem->regions[0];
521     dev->mem = g_realloc(dev->mem, regions_size);
522     dev->mem->nregions = dev->n_mem_sections;
523     used_memslots = dev->mem->nregions;
524     for (i = 0; i < dev->n_mem_sections; i++) {
525         struct vhost_memory_region *cur_vmr = dev->mem->regions + i;
526         struct MemoryRegionSection *mrs = dev->mem_sections + i;
527 
528         cur_vmr->guest_phys_addr = mrs->offset_within_address_space;
529         cur_vmr->memory_size     = int128_get64(mrs->size);
530         cur_vmr->userspace_addr  =
531             (uintptr_t)memory_region_get_ram_ptr(mrs->mr) +
532             mrs->offset_within_region;
533         cur_vmr->flags_padding   = 0;
534     }
535 
536     if (!dev->started) {
537         goto out;
538     }
539 
540     for (i = 0; i < dev->mem->nregions; i++) {
541         if (vhost_verify_ring_mappings(dev,
542                        (void *)(uintptr_t)dev->mem->regions[i].userspace_addr,
543                        dev->mem->regions[i].guest_phys_addr,
544                        dev->mem->regions[i].memory_size)) {
545             error_report("Verify ring failure on region %d", i);
546             abort();
547         }
548     }
549 
550     if (!dev->log_enabled) {
551         r = dev->vhost_ops->vhost_set_mem_table(dev, dev->mem);
552         if (r < 0) {
553             VHOST_OPS_DEBUG("vhost_set_mem_table failed");
554         }
555         goto out;
556     }
557     log_size = vhost_get_log_size(dev);
558     /* We allocate an extra 4K bytes to log,
559      * to reduce the * number of reallocations. */
560 #define VHOST_LOG_BUFFER (0x1000 / sizeof *dev->log)
561     /* To log more, must increase log size before table update. */
562     if (dev->log_size < log_size) {
563         vhost_dev_log_resize(dev, log_size + VHOST_LOG_BUFFER);
564     }
565     r = dev->vhost_ops->vhost_set_mem_table(dev, dev->mem);
566     if (r < 0) {
567         VHOST_OPS_DEBUG("vhost_set_mem_table failed");
568     }
569     /* To log less, can only decrease log size after table update. */
570     if (dev->log_size > log_size + VHOST_LOG_BUFFER) {
571         vhost_dev_log_resize(dev, log_size);
572     }
573 
574 out:
575     /* Deref the old list of sections, this must happen _after_ the
576      * vhost_set_mem_table to ensure the client isn't still using the
577      * section we're about to unref.
578      */
579     while (n_old_sections--) {
580         memory_region_unref(old_sections[n_old_sections].mr);
581     }
582     g_free(old_sections);
583     return;
584 }
585 
586 /* Adds the section data to the tmp_section structure.
587  * It relies on the listener calling us in memory address order
588  * and for each region (via the _add and _nop methods) to
589  * join neighbours.
590  */
591 static void vhost_region_add_section(struct vhost_dev *dev,
592                                      MemoryRegionSection *section)
593 {
594     bool need_add = true;
595     uint64_t mrs_size = int128_get64(section->size);
596     uint64_t mrs_gpa = section->offset_within_address_space;
597     uintptr_t mrs_host = (uintptr_t)memory_region_get_ram_ptr(section->mr) +
598                          section->offset_within_region;
599     RAMBlock *mrs_rb = section->mr->ram_block;
600 
601     trace_vhost_region_add_section(section->mr->name, mrs_gpa, mrs_size,
602                                    mrs_host);
603 
604     if (dev->vhost_ops->backend_type == VHOST_BACKEND_TYPE_USER) {
605         /* Round the section to it's page size */
606         /* First align the start down to a page boundary */
607         size_t mrs_page = qemu_ram_pagesize(mrs_rb);
608         uint64_t alignage = mrs_host & (mrs_page - 1);
609         if (alignage) {
610             mrs_host -= alignage;
611             mrs_size += alignage;
612             mrs_gpa  -= alignage;
613         }
614         /* Now align the size up to a page boundary */
615         alignage = mrs_size & (mrs_page - 1);
616         if (alignage) {
617             mrs_size += mrs_page - alignage;
618         }
619         trace_vhost_region_add_section_aligned(section->mr->name, mrs_gpa,
620                                                mrs_size, mrs_host);
621     }
622 
623     if (dev->n_tmp_sections) {
624         /* Since we already have at least one section, lets see if
625          * this extends it; since we're scanning in order, we only
626          * have to look at the last one, and the FlatView that calls
627          * us shouldn't have overlaps.
628          */
629         MemoryRegionSection *prev_sec = dev->tmp_sections +
630                                                (dev->n_tmp_sections - 1);
631         uint64_t prev_gpa_start = prev_sec->offset_within_address_space;
632         uint64_t prev_size = int128_get64(prev_sec->size);
633         uint64_t prev_gpa_end   = range_get_last(prev_gpa_start, prev_size);
634         uint64_t prev_host_start =
635                         (uintptr_t)memory_region_get_ram_ptr(prev_sec->mr) +
636                         prev_sec->offset_within_region;
637         uint64_t prev_host_end   = range_get_last(prev_host_start, prev_size);
638 
639         if (mrs_gpa <= (prev_gpa_end + 1)) {
640             /* OK, looks like overlapping/intersecting - it's possible that
641              * the rounding to page sizes has made them overlap, but they should
642              * match up in the same RAMBlock if they do.
643              */
644             if (mrs_gpa < prev_gpa_start) {
645                 error_report("%s:Section '%s' rounded to %"PRIx64
646                              " prior to previous '%s' %"PRIx64,
647                              __func__, section->mr->name, mrs_gpa,
648                              prev_sec->mr->name, prev_gpa_start);
649                 /* A way to cleanly fail here would be better */
650                 return;
651             }
652             /* Offset from the start of the previous GPA to this GPA */
653             size_t offset = mrs_gpa - prev_gpa_start;
654 
655             if (prev_host_start + offset == mrs_host &&
656                 section->mr == prev_sec->mr &&
657                 (!dev->vhost_ops->vhost_backend_can_merge ||
658                  dev->vhost_ops->vhost_backend_can_merge(dev,
659                     mrs_host, mrs_size,
660                     prev_host_start, prev_size))) {
661                 uint64_t max_end = MAX(prev_host_end, mrs_host + mrs_size);
662                 need_add = false;
663                 prev_sec->offset_within_address_space =
664                     MIN(prev_gpa_start, mrs_gpa);
665                 prev_sec->offset_within_region =
666                     MIN(prev_host_start, mrs_host) -
667                     (uintptr_t)memory_region_get_ram_ptr(prev_sec->mr);
668                 prev_sec->size = int128_make64(max_end - MIN(prev_host_start,
669                                                mrs_host));
670                 trace_vhost_region_add_section_merge(section->mr->name,
671                                         int128_get64(prev_sec->size),
672                                         prev_sec->offset_within_address_space,
673                                         prev_sec->offset_within_region);
674             } else {
675                 /* adjoining regions are fine, but overlapping ones with
676                  * different blocks/offsets shouldn't happen
677                  */
678                 if (mrs_gpa != prev_gpa_end + 1) {
679                     error_report("%s: Overlapping but not coherent sections "
680                                  "at %"PRIx64,
681                                  __func__, mrs_gpa);
682                     return;
683                 }
684             }
685         }
686     }
687 
688     if (need_add) {
689         ++dev->n_tmp_sections;
690         dev->tmp_sections = g_renew(MemoryRegionSection, dev->tmp_sections,
691                                     dev->n_tmp_sections);
692         dev->tmp_sections[dev->n_tmp_sections - 1] = *section;
693         /* The flatview isn't stable and we don't use it, making it NULL
694          * means we can memcmp the list.
695          */
696         dev->tmp_sections[dev->n_tmp_sections - 1].fv = NULL;
697         memory_region_ref(section->mr);
698     }
699 }
700 
701 /* Used for both add and nop callbacks */
702 static void vhost_region_addnop(MemoryListener *listener,
703                                 MemoryRegionSection *section)
704 {
705     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
706                                          memory_listener);
707 
708     if (!vhost_section(dev, section)) {
709         return;
710     }
711     vhost_region_add_section(dev, section);
712 }
713 
714 static void vhost_iommu_unmap_notify(IOMMUNotifier *n, IOMMUTLBEntry *iotlb)
715 {
716     struct vhost_iommu *iommu = container_of(n, struct vhost_iommu, n);
717     struct vhost_dev *hdev = iommu->hdev;
718     hwaddr iova = iotlb->iova + iommu->iommu_offset;
719 
720     if (vhost_backend_invalidate_device_iotlb(hdev, iova,
721                                               iotlb->addr_mask + 1)) {
722         error_report("Fail to invalidate device iotlb");
723     }
724 }
725 
726 static void vhost_iommu_region_add(MemoryListener *listener,
727                                    MemoryRegionSection *section)
728 {
729     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
730                                          iommu_listener);
731     struct vhost_iommu *iommu;
732     Int128 end;
733     int iommu_idx;
734     IOMMUMemoryRegion *iommu_mr;
735     int ret;
736 
737     if (!memory_region_is_iommu(section->mr)) {
738         return;
739     }
740 
741     iommu_mr = IOMMU_MEMORY_REGION(section->mr);
742 
743     iommu = g_malloc0(sizeof(*iommu));
744     end = int128_add(int128_make64(section->offset_within_region),
745                      section->size);
746     end = int128_sub(end, int128_one());
747     iommu_idx = memory_region_iommu_attrs_to_index(iommu_mr,
748                                                    MEMTXATTRS_UNSPECIFIED);
749     iommu_notifier_init(&iommu->n, vhost_iommu_unmap_notify,
750                         IOMMU_NOTIFIER_DEVIOTLB_UNMAP,
751                         section->offset_within_region,
752                         int128_get64(end),
753                         iommu_idx);
754     iommu->mr = section->mr;
755     iommu->iommu_offset = section->offset_within_address_space -
756                           section->offset_within_region;
757     iommu->hdev = dev;
758     ret = memory_region_register_iommu_notifier(section->mr, &iommu->n, NULL);
759     if (ret) {
760         /*
761          * Some vIOMMUs do not support dev-iotlb yet.  If so, try to use the
762          * UNMAP legacy message
763          */
764         iommu->n.notifier_flags = IOMMU_NOTIFIER_UNMAP;
765         memory_region_register_iommu_notifier(section->mr, &iommu->n,
766                                               &error_fatal);
767     }
768     QLIST_INSERT_HEAD(&dev->iommu_list, iommu, iommu_next);
769     /* TODO: can replay help performance here? */
770 }
771 
772 static void vhost_iommu_region_del(MemoryListener *listener,
773                                    MemoryRegionSection *section)
774 {
775     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
776                                          iommu_listener);
777     struct vhost_iommu *iommu;
778 
779     if (!memory_region_is_iommu(section->mr)) {
780         return;
781     }
782 
783     QLIST_FOREACH(iommu, &dev->iommu_list, iommu_next) {
784         if (iommu->mr == section->mr &&
785             iommu->n.start == section->offset_within_region) {
786             memory_region_unregister_iommu_notifier(iommu->mr,
787                                                     &iommu->n);
788             QLIST_REMOVE(iommu, iommu_next);
789             g_free(iommu);
790             break;
791         }
792     }
793 }
794 
795 static int vhost_virtqueue_set_addr(struct vhost_dev *dev,
796                                     struct vhost_virtqueue *vq,
797                                     unsigned idx, bool enable_log)
798 {
799     struct vhost_vring_addr addr;
800     int r;
801     memset(&addr, 0, sizeof(struct vhost_vring_addr));
802 
803     if (dev->vhost_ops->vhost_vq_get_addr) {
804         r = dev->vhost_ops->vhost_vq_get_addr(dev, &addr, vq);
805         if (r < 0) {
806             VHOST_OPS_DEBUG("vhost_vq_get_addr failed");
807             return -errno;
808         }
809     } else {
810         addr.desc_user_addr = (uint64_t)(unsigned long)vq->desc;
811         addr.avail_user_addr = (uint64_t)(unsigned long)vq->avail;
812         addr.used_user_addr = (uint64_t)(unsigned long)vq->used;
813     }
814     addr.index = idx;
815     addr.log_guest_addr = vq->used_phys;
816     addr.flags = enable_log ? (1 << VHOST_VRING_F_LOG) : 0;
817     r = dev->vhost_ops->vhost_set_vring_addr(dev, &addr);
818     if (r < 0) {
819         VHOST_OPS_DEBUG("vhost_set_vring_addr failed");
820         return -errno;
821     }
822     return 0;
823 }
824 
825 static int vhost_dev_set_features(struct vhost_dev *dev,
826                                   bool enable_log)
827 {
828     uint64_t features = dev->acked_features;
829     int r;
830     if (enable_log) {
831         features |= 0x1ULL << VHOST_F_LOG_ALL;
832     }
833     if (!vhost_dev_has_iommu(dev)) {
834         features &= ~(0x1ULL << VIRTIO_F_IOMMU_PLATFORM);
835     }
836     if (dev->vhost_ops->vhost_force_iommu) {
837         if (dev->vhost_ops->vhost_force_iommu(dev) == true) {
838             features |= 0x1ULL << VIRTIO_F_IOMMU_PLATFORM;
839        }
840     }
841     r = dev->vhost_ops->vhost_set_features(dev, features);
842     if (r < 0) {
843         VHOST_OPS_DEBUG("vhost_set_features failed");
844         goto out;
845     }
846     if (dev->vhost_ops->vhost_set_backend_cap) {
847         r = dev->vhost_ops->vhost_set_backend_cap(dev);
848         if (r < 0) {
849             VHOST_OPS_DEBUG("vhost_set_backend_cap failed");
850             goto out;
851         }
852     }
853 
854 out:
855     return r < 0 ? -errno : 0;
856 }
857 
858 static int vhost_dev_set_log(struct vhost_dev *dev, bool enable_log)
859 {
860     int r, i, idx;
861     hwaddr addr;
862 
863     r = vhost_dev_set_features(dev, enable_log);
864     if (r < 0) {
865         goto err_features;
866     }
867     for (i = 0; i < dev->nvqs; ++i) {
868         idx = dev->vhost_ops->vhost_get_vq_index(dev, dev->vq_index + i);
869         addr = virtio_queue_get_desc_addr(dev->vdev, idx);
870         if (!addr) {
871             /*
872              * The queue might not be ready for start. If this
873              * is the case there is no reason to continue the process.
874              * The similar logic is used by the vhost_virtqueue_start()
875              * routine.
876              */
877             continue;
878         }
879         r = vhost_virtqueue_set_addr(dev, dev->vqs + i, idx,
880                                      enable_log);
881         if (r < 0) {
882             goto err_vq;
883         }
884     }
885     return 0;
886 err_vq:
887     for (; i >= 0; --i) {
888         idx = dev->vhost_ops->vhost_get_vq_index(dev, dev->vq_index + i);
889         vhost_virtqueue_set_addr(dev, dev->vqs + i, idx,
890                                  dev->log_enabled);
891     }
892     vhost_dev_set_features(dev, dev->log_enabled);
893 err_features:
894     return r;
895 }
896 
897 static int vhost_migration_log(MemoryListener *listener, bool enable)
898 {
899     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
900                                          memory_listener);
901     int r;
902     if (enable == dev->log_enabled) {
903         return 0;
904     }
905     if (!dev->started) {
906         dev->log_enabled = enable;
907         return 0;
908     }
909 
910     r = 0;
911     if (!enable) {
912         r = vhost_dev_set_log(dev, false);
913         if (r < 0) {
914             goto check_dev_state;
915         }
916         vhost_log_put(dev, false);
917     } else {
918         vhost_dev_log_resize(dev, vhost_get_log_size(dev));
919         r = vhost_dev_set_log(dev, true);
920         if (r < 0) {
921             goto check_dev_state;
922         }
923     }
924 
925 check_dev_state:
926     dev->log_enabled = enable;
927     /*
928      * vhost-user-* devices could change their state during log
929      * initialization due to disconnect. So check dev state after
930      * vhost communication.
931      */
932     if (!dev->started) {
933         /*
934          * Since device is in the stopped state, it is okay for
935          * migration. Return success.
936          */
937         r = 0;
938     }
939     if (r) {
940         /* An error occurred. */
941         dev->log_enabled = false;
942     }
943 
944     return r;
945 }
946 
947 static void vhost_log_global_start(MemoryListener *listener)
948 {
949     int r;
950 
951     r = vhost_migration_log(listener, true);
952     if (r < 0) {
953         abort();
954     }
955 }
956 
957 static void vhost_log_global_stop(MemoryListener *listener)
958 {
959     int r;
960 
961     r = vhost_migration_log(listener, false);
962     if (r < 0) {
963         abort();
964     }
965 }
966 
967 static void vhost_log_start(MemoryListener *listener,
968                             MemoryRegionSection *section,
969                             int old, int new)
970 {
971     /* FIXME: implement */
972 }
973 
974 static void vhost_log_stop(MemoryListener *listener,
975                            MemoryRegionSection *section,
976                            int old, int new)
977 {
978     /* FIXME: implement */
979 }
980 
981 /* The vhost driver natively knows how to handle the vrings of non
982  * cross-endian legacy devices and modern devices. Only legacy devices
983  * exposed to a bi-endian guest may require the vhost driver to use a
984  * specific endianness.
985  */
986 static inline bool vhost_needs_vring_endian(VirtIODevice *vdev)
987 {
988     if (virtio_vdev_has_feature(vdev, VIRTIO_F_VERSION_1)) {
989         return false;
990     }
991 #ifdef HOST_WORDS_BIGENDIAN
992     return vdev->device_endian == VIRTIO_DEVICE_ENDIAN_LITTLE;
993 #else
994     return vdev->device_endian == VIRTIO_DEVICE_ENDIAN_BIG;
995 #endif
996 }
997 
998 static int vhost_virtqueue_set_vring_endian_legacy(struct vhost_dev *dev,
999                                                    bool is_big_endian,
1000                                                    int vhost_vq_index)
1001 {
1002     struct vhost_vring_state s = {
1003         .index = vhost_vq_index,
1004         .num = is_big_endian
1005     };
1006 
1007     if (!dev->vhost_ops->vhost_set_vring_endian(dev, &s)) {
1008         return 0;
1009     }
1010 
1011     VHOST_OPS_DEBUG("vhost_set_vring_endian failed");
1012     if (errno == ENOTTY) {
1013         error_report("vhost does not support cross-endian");
1014         return -ENOSYS;
1015     }
1016 
1017     return -errno;
1018 }
1019 
1020 static int vhost_memory_region_lookup(struct vhost_dev *hdev,
1021                                       uint64_t gpa, uint64_t *uaddr,
1022                                       uint64_t *len)
1023 {
1024     int i;
1025 
1026     for (i = 0; i < hdev->mem->nregions; i++) {
1027         struct vhost_memory_region *reg = hdev->mem->regions + i;
1028 
1029         if (gpa >= reg->guest_phys_addr &&
1030             reg->guest_phys_addr + reg->memory_size > gpa) {
1031             *uaddr = reg->userspace_addr + gpa - reg->guest_phys_addr;
1032             *len = reg->guest_phys_addr + reg->memory_size - gpa;
1033             return 0;
1034         }
1035     }
1036 
1037     return -EFAULT;
1038 }
1039 
1040 int vhost_device_iotlb_miss(struct vhost_dev *dev, uint64_t iova, int write)
1041 {
1042     IOMMUTLBEntry iotlb;
1043     uint64_t uaddr, len;
1044     int ret = -EFAULT;
1045 
1046     RCU_READ_LOCK_GUARD();
1047 
1048     trace_vhost_iotlb_miss(dev, 1);
1049 
1050     iotlb = address_space_get_iotlb_entry(dev->vdev->dma_as,
1051                                           iova, write,
1052                                           MEMTXATTRS_UNSPECIFIED);
1053     if (iotlb.target_as != NULL) {
1054         ret = vhost_memory_region_lookup(dev, iotlb.translated_addr,
1055                                          &uaddr, &len);
1056         if (ret) {
1057             trace_vhost_iotlb_miss(dev, 3);
1058             error_report("Fail to lookup the translated address "
1059                          "%"PRIx64, iotlb.translated_addr);
1060             goto out;
1061         }
1062 
1063         len = MIN(iotlb.addr_mask + 1, len);
1064         iova = iova & ~iotlb.addr_mask;
1065 
1066         ret = vhost_backend_update_device_iotlb(dev, iova, uaddr,
1067                                                 len, iotlb.perm);
1068         if (ret) {
1069             trace_vhost_iotlb_miss(dev, 4);
1070             error_report("Fail to update device iotlb");
1071             goto out;
1072         }
1073     }
1074 
1075     trace_vhost_iotlb_miss(dev, 2);
1076 
1077 out:
1078     return ret;
1079 }
1080 
1081 static int vhost_virtqueue_start(struct vhost_dev *dev,
1082                                 struct VirtIODevice *vdev,
1083                                 struct vhost_virtqueue *vq,
1084                                 unsigned idx)
1085 {
1086     BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev)));
1087     VirtioBusState *vbus = VIRTIO_BUS(qbus);
1088     VirtioBusClass *k = VIRTIO_BUS_GET_CLASS(vbus);
1089     hwaddr s, l, a;
1090     int r;
1091     int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, idx);
1092     struct vhost_vring_file file = {
1093         .index = vhost_vq_index
1094     };
1095     struct vhost_vring_state state = {
1096         .index = vhost_vq_index
1097     };
1098     struct VirtQueue *vvq = virtio_get_queue(vdev, idx);
1099 
1100     a = virtio_queue_get_desc_addr(vdev, idx);
1101     if (a == 0) {
1102         /* Queue might not be ready for start */
1103         return 0;
1104     }
1105 
1106     vq->num = state.num = virtio_queue_get_num(vdev, idx);
1107     r = dev->vhost_ops->vhost_set_vring_num(dev, &state);
1108     if (r) {
1109         VHOST_OPS_DEBUG("vhost_set_vring_num failed");
1110         return -errno;
1111     }
1112 
1113     state.num = virtio_queue_get_last_avail_idx(vdev, idx);
1114     r = dev->vhost_ops->vhost_set_vring_base(dev, &state);
1115     if (r) {
1116         VHOST_OPS_DEBUG("vhost_set_vring_base failed");
1117         return -errno;
1118     }
1119 
1120     if (vhost_needs_vring_endian(vdev)) {
1121         r = vhost_virtqueue_set_vring_endian_legacy(dev,
1122                                                     virtio_is_big_endian(vdev),
1123                                                     vhost_vq_index);
1124         if (r) {
1125             return -errno;
1126         }
1127     }
1128 
1129     vq->desc_size = s = l = virtio_queue_get_desc_size(vdev, idx);
1130     vq->desc_phys = a;
1131     vq->desc = vhost_memory_map(dev, a, &l, false);
1132     if (!vq->desc || l != s) {
1133         r = -ENOMEM;
1134         goto fail_alloc_desc;
1135     }
1136     vq->avail_size = s = l = virtio_queue_get_avail_size(vdev, idx);
1137     vq->avail_phys = a = virtio_queue_get_avail_addr(vdev, idx);
1138     vq->avail = vhost_memory_map(dev, a, &l, false);
1139     if (!vq->avail || l != s) {
1140         r = -ENOMEM;
1141         goto fail_alloc_avail;
1142     }
1143     vq->used_size = s = l = virtio_queue_get_used_size(vdev, idx);
1144     vq->used_phys = a = virtio_queue_get_used_addr(vdev, idx);
1145     vq->used = vhost_memory_map(dev, a, &l, true);
1146     if (!vq->used || l != s) {
1147         r = -ENOMEM;
1148         goto fail_alloc_used;
1149     }
1150 
1151     r = vhost_virtqueue_set_addr(dev, vq, vhost_vq_index, dev->log_enabled);
1152     if (r < 0) {
1153         r = -errno;
1154         goto fail_alloc;
1155     }
1156 
1157     file.fd = event_notifier_get_fd(virtio_queue_get_host_notifier(vvq));
1158     r = dev->vhost_ops->vhost_set_vring_kick(dev, &file);
1159     if (r) {
1160         VHOST_OPS_DEBUG("vhost_set_vring_kick failed");
1161         r = -errno;
1162         goto fail_kick;
1163     }
1164 
1165     /* Clear and discard previous events if any. */
1166     event_notifier_test_and_clear(&vq->masked_notifier);
1167 
1168     /* Init vring in unmasked state, unless guest_notifier_mask
1169      * will do it later.
1170      */
1171     if (!vdev->use_guest_notifier_mask) {
1172         /* TODO: check and handle errors. */
1173         vhost_virtqueue_mask(dev, vdev, idx, false);
1174     }
1175 
1176     if (k->query_guest_notifiers &&
1177         k->query_guest_notifiers(qbus->parent) &&
1178         virtio_queue_vector(vdev, idx) == VIRTIO_NO_VECTOR) {
1179         file.fd = -1;
1180         r = dev->vhost_ops->vhost_set_vring_call(dev, &file);
1181         if (r) {
1182             goto fail_vector;
1183         }
1184     }
1185 
1186     return 0;
1187 
1188 fail_vector:
1189 fail_kick:
1190 fail_alloc:
1191     vhost_memory_unmap(dev, vq->used, virtio_queue_get_used_size(vdev, idx),
1192                        0, 0);
1193 fail_alloc_used:
1194     vhost_memory_unmap(dev, vq->avail, virtio_queue_get_avail_size(vdev, idx),
1195                        0, 0);
1196 fail_alloc_avail:
1197     vhost_memory_unmap(dev, vq->desc, virtio_queue_get_desc_size(vdev, idx),
1198                        0, 0);
1199 fail_alloc_desc:
1200     return r;
1201 }
1202 
1203 static void vhost_virtqueue_stop(struct vhost_dev *dev,
1204                                     struct VirtIODevice *vdev,
1205                                     struct vhost_virtqueue *vq,
1206                                     unsigned idx)
1207 {
1208     int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, idx);
1209     struct vhost_vring_state state = {
1210         .index = vhost_vq_index,
1211     };
1212     int r;
1213 
1214     if (virtio_queue_get_desc_addr(vdev, idx) == 0) {
1215         /* Don't stop the virtqueue which might have not been started */
1216         return;
1217     }
1218 
1219     r = dev->vhost_ops->vhost_get_vring_base(dev, &state);
1220     if (r < 0) {
1221         VHOST_OPS_DEBUG("vhost VQ %u ring restore failed: %d", idx, r);
1222         /* Connection to the backend is broken, so let's sync internal
1223          * last avail idx to the device used idx.
1224          */
1225         virtio_queue_restore_last_avail_idx(vdev, idx);
1226     } else {
1227         virtio_queue_set_last_avail_idx(vdev, idx, state.num);
1228     }
1229     virtio_queue_invalidate_signalled_used(vdev, idx);
1230     virtio_queue_update_used_idx(vdev, idx);
1231 
1232     /* In the cross-endian case, we need to reset the vring endianness to
1233      * native as legacy devices expect so by default.
1234      */
1235     if (vhost_needs_vring_endian(vdev)) {
1236         vhost_virtqueue_set_vring_endian_legacy(dev,
1237                                                 !virtio_is_big_endian(vdev),
1238                                                 vhost_vq_index);
1239     }
1240 
1241     vhost_memory_unmap(dev, vq->used, virtio_queue_get_used_size(vdev, idx),
1242                        1, virtio_queue_get_used_size(vdev, idx));
1243     vhost_memory_unmap(dev, vq->avail, virtio_queue_get_avail_size(vdev, idx),
1244                        0, virtio_queue_get_avail_size(vdev, idx));
1245     vhost_memory_unmap(dev, vq->desc, virtio_queue_get_desc_size(vdev, idx),
1246                        0, virtio_queue_get_desc_size(vdev, idx));
1247 }
1248 
1249 static void vhost_eventfd_add(MemoryListener *listener,
1250                               MemoryRegionSection *section,
1251                               bool match_data, uint64_t data, EventNotifier *e)
1252 {
1253 }
1254 
1255 static void vhost_eventfd_del(MemoryListener *listener,
1256                               MemoryRegionSection *section,
1257                               bool match_data, uint64_t data, EventNotifier *e)
1258 {
1259 }
1260 
1261 static int vhost_virtqueue_set_busyloop_timeout(struct vhost_dev *dev,
1262                                                 int n, uint32_t timeout)
1263 {
1264     int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, n);
1265     struct vhost_vring_state state = {
1266         .index = vhost_vq_index,
1267         .num = timeout,
1268     };
1269     int r;
1270 
1271     if (!dev->vhost_ops->vhost_set_vring_busyloop_timeout) {
1272         return -EINVAL;
1273     }
1274 
1275     r = dev->vhost_ops->vhost_set_vring_busyloop_timeout(dev, &state);
1276     if (r) {
1277         VHOST_OPS_DEBUG("vhost_set_vring_busyloop_timeout failed");
1278         return r;
1279     }
1280 
1281     return 0;
1282 }
1283 
1284 static int vhost_virtqueue_init(struct vhost_dev *dev,
1285                                 struct vhost_virtqueue *vq, int n)
1286 {
1287     int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, n);
1288     struct vhost_vring_file file = {
1289         .index = vhost_vq_index,
1290     };
1291     int r = event_notifier_init(&vq->masked_notifier, 0);
1292     if (r < 0) {
1293         return r;
1294     }
1295 
1296     file.fd = event_notifier_get_fd(&vq->masked_notifier);
1297     r = dev->vhost_ops->vhost_set_vring_call(dev, &file);
1298     if (r) {
1299         VHOST_OPS_DEBUG("vhost_set_vring_call failed");
1300         r = -errno;
1301         goto fail_call;
1302     }
1303 
1304     vq->dev = dev;
1305 
1306     return 0;
1307 fail_call:
1308     event_notifier_cleanup(&vq->masked_notifier);
1309     return r;
1310 }
1311 
1312 static void vhost_virtqueue_cleanup(struct vhost_virtqueue *vq)
1313 {
1314     event_notifier_cleanup(&vq->masked_notifier);
1315 }
1316 
1317 int vhost_dev_init(struct vhost_dev *hdev, void *opaque,
1318                    VhostBackendType backend_type, uint32_t busyloop_timeout,
1319                    Error **errp)
1320 {
1321     uint64_t features;
1322     int i, r, n_initialized_vqs = 0;
1323 
1324     hdev->vdev = NULL;
1325     hdev->migration_blocker = NULL;
1326 
1327     r = vhost_set_backend_type(hdev, backend_type);
1328     assert(r >= 0);
1329 
1330     r = hdev->vhost_ops->vhost_backend_init(hdev, opaque, errp);
1331     if (r < 0) {
1332         goto fail;
1333     }
1334 
1335     r = hdev->vhost_ops->vhost_set_owner(hdev);
1336     if (r < 0) {
1337         error_setg_errno(errp, -r, "vhost_set_owner failed");
1338         goto fail;
1339     }
1340 
1341     r = hdev->vhost_ops->vhost_get_features(hdev, &features);
1342     if (r < 0) {
1343         error_setg_errno(errp, -r, "vhost_get_features failed");
1344         goto fail;
1345     }
1346 
1347     for (i = 0; i < hdev->nvqs; ++i, ++n_initialized_vqs) {
1348         r = vhost_virtqueue_init(hdev, hdev->vqs + i, hdev->vq_index + i);
1349         if (r < 0) {
1350             error_setg_errno(errp, -r, "Failed to initialize virtqueue %d", i);
1351             goto fail;
1352         }
1353     }
1354 
1355     if (busyloop_timeout) {
1356         for (i = 0; i < hdev->nvqs; ++i) {
1357             r = vhost_virtqueue_set_busyloop_timeout(hdev, hdev->vq_index + i,
1358                                                      busyloop_timeout);
1359             if (r < 0) {
1360                 error_setg_errno(errp, -r, "Failed to set busyloop timeout");
1361                 goto fail_busyloop;
1362             }
1363         }
1364     }
1365 
1366     hdev->features = features;
1367 
1368     hdev->memory_listener = (MemoryListener) {
1369         .name = "vhost",
1370         .begin = vhost_begin,
1371         .commit = vhost_commit,
1372         .region_add = vhost_region_addnop,
1373         .region_nop = vhost_region_addnop,
1374         .log_start = vhost_log_start,
1375         .log_stop = vhost_log_stop,
1376         .log_sync = vhost_log_sync,
1377         .log_global_start = vhost_log_global_start,
1378         .log_global_stop = vhost_log_global_stop,
1379         .eventfd_add = vhost_eventfd_add,
1380         .eventfd_del = vhost_eventfd_del,
1381         .priority = 10
1382     };
1383 
1384     hdev->iommu_listener = (MemoryListener) {
1385         .name = "vhost-iommu",
1386         .region_add = vhost_iommu_region_add,
1387         .region_del = vhost_iommu_region_del,
1388     };
1389 
1390     if (hdev->migration_blocker == NULL) {
1391         if (!(hdev->features & (0x1ULL << VHOST_F_LOG_ALL))) {
1392             error_setg(&hdev->migration_blocker,
1393                        "Migration disabled: vhost lacks VHOST_F_LOG_ALL feature.");
1394         } else if (vhost_dev_log_is_shared(hdev) && !qemu_memfd_alloc_check()) {
1395             error_setg(&hdev->migration_blocker,
1396                        "Migration disabled: failed to allocate shared memory");
1397         }
1398     }
1399 
1400     if (hdev->migration_blocker != NULL) {
1401         r = migrate_add_blocker(hdev->migration_blocker, errp);
1402         if (r < 0) {
1403             error_free(hdev->migration_blocker);
1404             goto fail_busyloop;
1405         }
1406     }
1407 
1408     hdev->mem = g_malloc0(offsetof(struct vhost_memory, regions));
1409     hdev->n_mem_sections = 0;
1410     hdev->mem_sections = NULL;
1411     hdev->log = NULL;
1412     hdev->log_size = 0;
1413     hdev->log_enabled = false;
1414     hdev->started = false;
1415     memory_listener_register(&hdev->memory_listener, &address_space_memory);
1416     QLIST_INSERT_HEAD(&vhost_devices, hdev, entry);
1417 
1418     if (used_memslots > hdev->vhost_ops->vhost_backend_memslots_limit(hdev)) {
1419         error_setg(errp, "vhost backend memory slots limit is less"
1420                    " than current number of present memory slots");
1421         r = -EINVAL;
1422         goto fail_busyloop;
1423     }
1424 
1425     return 0;
1426 
1427 fail_busyloop:
1428     if (busyloop_timeout) {
1429         while (--i >= 0) {
1430             vhost_virtqueue_set_busyloop_timeout(hdev, hdev->vq_index + i, 0);
1431         }
1432     }
1433 fail:
1434     hdev->nvqs = n_initialized_vqs;
1435     vhost_dev_cleanup(hdev);
1436     return r;
1437 }
1438 
1439 void vhost_dev_cleanup(struct vhost_dev *hdev)
1440 {
1441     int i;
1442 
1443     for (i = 0; i < hdev->nvqs; ++i) {
1444         vhost_virtqueue_cleanup(hdev->vqs + i);
1445     }
1446     if (hdev->mem) {
1447         /* those are only safe after successful init */
1448         memory_listener_unregister(&hdev->memory_listener);
1449         QLIST_REMOVE(hdev, entry);
1450     }
1451     if (hdev->migration_blocker) {
1452         migrate_del_blocker(hdev->migration_blocker);
1453         error_free(hdev->migration_blocker);
1454     }
1455     g_free(hdev->mem);
1456     g_free(hdev->mem_sections);
1457     if (hdev->vhost_ops) {
1458         hdev->vhost_ops->vhost_backend_cleanup(hdev);
1459     }
1460     assert(!hdev->log);
1461 
1462     memset(hdev, 0, sizeof(struct vhost_dev));
1463 }
1464 
1465 /* Stop processing guest IO notifications in qemu.
1466  * Start processing them in vhost in kernel.
1467  */
1468 int vhost_dev_enable_notifiers(struct vhost_dev *hdev, VirtIODevice *vdev)
1469 {
1470     BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev)));
1471     int i, r, e;
1472 
1473     /* We will pass the notifiers to the kernel, make sure that QEMU
1474      * doesn't interfere.
1475      */
1476     r = virtio_device_grab_ioeventfd(vdev);
1477     if (r < 0) {
1478         error_report("binding does not support host notifiers");
1479         goto fail;
1480     }
1481 
1482     for (i = 0; i < hdev->nvqs; ++i) {
1483         r = virtio_bus_set_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i,
1484                                          true);
1485         if (r < 0) {
1486             error_report("vhost VQ %d notifier binding failed: %d", i, -r);
1487             goto fail_vq;
1488         }
1489     }
1490 
1491     return 0;
1492 fail_vq:
1493     while (--i >= 0) {
1494         e = virtio_bus_set_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i,
1495                                          false);
1496         if (e < 0) {
1497             error_report("vhost VQ %d notifier cleanup error: %d", i, -r);
1498         }
1499         assert (e >= 0);
1500         virtio_bus_cleanup_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i);
1501     }
1502     virtio_device_release_ioeventfd(vdev);
1503 fail:
1504     return r;
1505 }
1506 
1507 /* Stop processing guest IO notifications in vhost.
1508  * Start processing them in qemu.
1509  * This might actually run the qemu handlers right away,
1510  * so virtio in qemu must be completely setup when this is called.
1511  */
1512 void vhost_dev_disable_notifiers(struct vhost_dev *hdev, VirtIODevice *vdev)
1513 {
1514     BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev)));
1515     int i, r;
1516 
1517     for (i = 0; i < hdev->nvqs; ++i) {
1518         r = virtio_bus_set_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i,
1519                                          false);
1520         if (r < 0) {
1521             error_report("vhost VQ %d notifier cleanup failed: %d", i, -r);
1522         }
1523         assert (r >= 0);
1524         virtio_bus_cleanup_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i);
1525     }
1526     virtio_device_release_ioeventfd(vdev);
1527 }
1528 
1529 /* Test and clear event pending status.
1530  * Should be called after unmask to avoid losing events.
1531  */
1532 bool vhost_virtqueue_pending(struct vhost_dev *hdev, int n)
1533 {
1534     struct vhost_virtqueue *vq = hdev->vqs + n - hdev->vq_index;
1535     assert(n >= hdev->vq_index && n < hdev->vq_index + hdev->nvqs);
1536     return event_notifier_test_and_clear(&vq->masked_notifier);
1537 }
1538 
1539 /* Mask/unmask events from this vq. */
1540 void vhost_virtqueue_mask(struct vhost_dev *hdev, VirtIODevice *vdev, int n,
1541                          bool mask)
1542 {
1543     struct VirtQueue *vvq = virtio_get_queue(vdev, n);
1544     int r, index = n - hdev->vq_index;
1545     struct vhost_vring_file file;
1546 
1547     /* should only be called after backend is connected */
1548     assert(hdev->vhost_ops);
1549 
1550     if (mask) {
1551         assert(vdev->use_guest_notifier_mask);
1552         file.fd = event_notifier_get_fd(&hdev->vqs[index].masked_notifier);
1553     } else {
1554         file.fd = event_notifier_get_fd(virtio_queue_get_guest_notifier(vvq));
1555     }
1556 
1557     file.index = hdev->vhost_ops->vhost_get_vq_index(hdev, n);
1558     r = hdev->vhost_ops->vhost_set_vring_call(hdev, &file);
1559     if (r < 0) {
1560         VHOST_OPS_DEBUG("vhost_set_vring_call failed");
1561     }
1562 }
1563 
1564 uint64_t vhost_get_features(struct vhost_dev *hdev, const int *feature_bits,
1565                             uint64_t features)
1566 {
1567     const int *bit = feature_bits;
1568     while (*bit != VHOST_INVALID_FEATURE_BIT) {
1569         uint64_t bit_mask = (1ULL << *bit);
1570         if (!(hdev->features & bit_mask)) {
1571             features &= ~bit_mask;
1572         }
1573         bit++;
1574     }
1575     return features;
1576 }
1577 
1578 void vhost_ack_features(struct vhost_dev *hdev, const int *feature_bits,
1579                         uint64_t features)
1580 {
1581     const int *bit = feature_bits;
1582     while (*bit != VHOST_INVALID_FEATURE_BIT) {
1583         uint64_t bit_mask = (1ULL << *bit);
1584         if (features & bit_mask) {
1585             hdev->acked_features |= bit_mask;
1586         }
1587         bit++;
1588     }
1589 }
1590 
1591 int vhost_dev_get_config(struct vhost_dev *hdev, uint8_t *config,
1592                          uint32_t config_len, Error **errp)
1593 {
1594     assert(hdev->vhost_ops);
1595 
1596     if (hdev->vhost_ops->vhost_get_config) {
1597         return hdev->vhost_ops->vhost_get_config(hdev, config, config_len,
1598                                                  errp);
1599     }
1600 
1601     error_setg(errp, "vhost_get_config not implemented");
1602     return -ENOTSUP;
1603 }
1604 
1605 int vhost_dev_set_config(struct vhost_dev *hdev, const uint8_t *data,
1606                          uint32_t offset, uint32_t size, uint32_t flags)
1607 {
1608     assert(hdev->vhost_ops);
1609 
1610     if (hdev->vhost_ops->vhost_set_config) {
1611         return hdev->vhost_ops->vhost_set_config(hdev, data, offset,
1612                                                  size, flags);
1613     }
1614 
1615     return -1;
1616 }
1617 
1618 void vhost_dev_set_config_notifier(struct vhost_dev *hdev,
1619                                    const VhostDevConfigOps *ops)
1620 {
1621     hdev->config_ops = ops;
1622 }
1623 
1624 void vhost_dev_free_inflight(struct vhost_inflight *inflight)
1625 {
1626     if (inflight && inflight->addr) {
1627         qemu_memfd_free(inflight->addr, inflight->size, inflight->fd);
1628         inflight->addr = NULL;
1629         inflight->fd = -1;
1630     }
1631 }
1632 
1633 static int vhost_dev_resize_inflight(struct vhost_inflight *inflight,
1634                                      uint64_t new_size)
1635 {
1636     Error *err = NULL;
1637     int fd = -1;
1638     void *addr = qemu_memfd_alloc("vhost-inflight", new_size,
1639                                   F_SEAL_GROW | F_SEAL_SHRINK | F_SEAL_SEAL,
1640                                   &fd, &err);
1641 
1642     if (err) {
1643         error_report_err(err);
1644         return -1;
1645     }
1646 
1647     vhost_dev_free_inflight(inflight);
1648     inflight->offset = 0;
1649     inflight->addr = addr;
1650     inflight->fd = fd;
1651     inflight->size = new_size;
1652 
1653     return 0;
1654 }
1655 
1656 void vhost_dev_save_inflight(struct vhost_inflight *inflight, QEMUFile *f)
1657 {
1658     if (inflight->addr) {
1659         qemu_put_be64(f, inflight->size);
1660         qemu_put_be16(f, inflight->queue_size);
1661         qemu_put_buffer(f, inflight->addr, inflight->size);
1662     } else {
1663         qemu_put_be64(f, 0);
1664     }
1665 }
1666 
1667 int vhost_dev_load_inflight(struct vhost_inflight *inflight, QEMUFile *f)
1668 {
1669     uint64_t size;
1670 
1671     size = qemu_get_be64(f);
1672     if (!size) {
1673         return 0;
1674     }
1675 
1676     if (inflight->size != size) {
1677         if (vhost_dev_resize_inflight(inflight, size)) {
1678             return -1;
1679         }
1680     }
1681     inflight->queue_size = qemu_get_be16(f);
1682 
1683     qemu_get_buffer(f, inflight->addr, size);
1684 
1685     return 0;
1686 }
1687 
1688 int vhost_dev_prepare_inflight(struct vhost_dev *hdev, VirtIODevice *vdev)
1689 {
1690     int r;
1691 
1692     if (hdev->vhost_ops->vhost_get_inflight_fd == NULL ||
1693         hdev->vhost_ops->vhost_set_inflight_fd == NULL) {
1694         return 0;
1695     }
1696 
1697     hdev->vdev = vdev;
1698 
1699     r = vhost_dev_set_features(hdev, hdev->log_enabled);
1700     if (r < 0) {
1701         VHOST_OPS_DEBUG("vhost_dev_prepare_inflight failed");
1702         return r;
1703     }
1704 
1705     return 0;
1706 }
1707 
1708 int vhost_dev_set_inflight(struct vhost_dev *dev,
1709                            struct vhost_inflight *inflight)
1710 {
1711     int r;
1712 
1713     if (dev->vhost_ops->vhost_set_inflight_fd && inflight->addr) {
1714         r = dev->vhost_ops->vhost_set_inflight_fd(dev, inflight);
1715         if (r) {
1716             VHOST_OPS_DEBUG("vhost_set_inflight_fd failed");
1717             return -errno;
1718         }
1719     }
1720 
1721     return 0;
1722 }
1723 
1724 int vhost_dev_get_inflight(struct vhost_dev *dev, uint16_t queue_size,
1725                            struct vhost_inflight *inflight)
1726 {
1727     int r;
1728 
1729     if (dev->vhost_ops->vhost_get_inflight_fd) {
1730         r = dev->vhost_ops->vhost_get_inflight_fd(dev, queue_size, inflight);
1731         if (r) {
1732             VHOST_OPS_DEBUG("vhost_get_inflight_fd failed");
1733             return -errno;
1734         }
1735     }
1736 
1737     return 0;
1738 }
1739 
1740 /* Host notifiers must be enabled at this point. */
1741 int vhost_dev_start(struct vhost_dev *hdev, VirtIODevice *vdev)
1742 {
1743     int i, r;
1744 
1745     /* should only be called after backend is connected */
1746     assert(hdev->vhost_ops);
1747 
1748     hdev->started = true;
1749     hdev->vdev = vdev;
1750 
1751     r = vhost_dev_set_features(hdev, hdev->log_enabled);
1752     if (r < 0) {
1753         goto fail_features;
1754     }
1755 
1756     if (vhost_dev_has_iommu(hdev)) {
1757         memory_listener_register(&hdev->iommu_listener, vdev->dma_as);
1758     }
1759 
1760     r = hdev->vhost_ops->vhost_set_mem_table(hdev, hdev->mem);
1761     if (r < 0) {
1762         VHOST_OPS_DEBUG("vhost_set_mem_table failed");
1763         r = -errno;
1764         goto fail_mem;
1765     }
1766     for (i = 0; i < hdev->nvqs; ++i) {
1767         r = vhost_virtqueue_start(hdev,
1768                                   vdev,
1769                                   hdev->vqs + i,
1770                                   hdev->vq_index + i);
1771         if (r < 0) {
1772             goto fail_vq;
1773         }
1774     }
1775 
1776     if (hdev->log_enabled) {
1777         uint64_t log_base;
1778 
1779         hdev->log_size = vhost_get_log_size(hdev);
1780         hdev->log = vhost_log_get(hdev->log_size,
1781                                   vhost_dev_log_is_shared(hdev));
1782         log_base = (uintptr_t)hdev->log->log;
1783         r = hdev->vhost_ops->vhost_set_log_base(hdev,
1784                                                 hdev->log_size ? log_base : 0,
1785                                                 hdev->log);
1786         if (r < 0) {
1787             VHOST_OPS_DEBUG("vhost_set_log_base failed");
1788             r = -errno;
1789             goto fail_log;
1790         }
1791     }
1792     if (hdev->vhost_ops->vhost_dev_start) {
1793         r = hdev->vhost_ops->vhost_dev_start(hdev, true);
1794         if (r) {
1795             goto fail_log;
1796         }
1797     }
1798     if (vhost_dev_has_iommu(hdev) &&
1799         hdev->vhost_ops->vhost_set_iotlb_callback) {
1800             hdev->vhost_ops->vhost_set_iotlb_callback(hdev, true);
1801 
1802         /* Update used ring information for IOTLB to work correctly,
1803          * vhost-kernel code requires for this.*/
1804         for (i = 0; i < hdev->nvqs; ++i) {
1805             struct vhost_virtqueue *vq = hdev->vqs + i;
1806             vhost_device_iotlb_miss(hdev, vq->used_phys, true);
1807         }
1808     }
1809     return 0;
1810 fail_log:
1811     vhost_log_put(hdev, false);
1812 fail_vq:
1813     while (--i >= 0) {
1814         vhost_virtqueue_stop(hdev,
1815                              vdev,
1816                              hdev->vqs + i,
1817                              hdev->vq_index + i);
1818     }
1819 
1820 fail_mem:
1821 fail_features:
1822 
1823     hdev->started = false;
1824     return r;
1825 }
1826 
1827 /* Host notifiers must be enabled at this point. */
1828 void vhost_dev_stop(struct vhost_dev *hdev, VirtIODevice *vdev)
1829 {
1830     int i;
1831 
1832     /* should only be called after backend is connected */
1833     assert(hdev->vhost_ops);
1834 
1835     if (hdev->vhost_ops->vhost_dev_start) {
1836         hdev->vhost_ops->vhost_dev_start(hdev, false);
1837     }
1838     for (i = 0; i < hdev->nvqs; ++i) {
1839         vhost_virtqueue_stop(hdev,
1840                              vdev,
1841                              hdev->vqs + i,
1842                              hdev->vq_index + i);
1843     }
1844 
1845     if (vhost_dev_has_iommu(hdev)) {
1846         if (hdev->vhost_ops->vhost_set_iotlb_callback) {
1847             hdev->vhost_ops->vhost_set_iotlb_callback(hdev, false);
1848         }
1849         memory_listener_unregister(&hdev->iommu_listener);
1850     }
1851     vhost_log_put(hdev, true);
1852     hdev->started = false;
1853     hdev->vdev = NULL;
1854 }
1855 
1856 int vhost_net_set_backend(struct vhost_dev *hdev,
1857                           struct vhost_vring_file *file)
1858 {
1859     if (hdev->vhost_ops->vhost_net_set_backend) {
1860         return hdev->vhost_ops->vhost_net_set_backend(hdev, file);
1861     }
1862 
1863     return -1;
1864 }
1865