xref: /openbmc/qemu/hw/virtio/vhost.c (revision b3c818a4)
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 = qatomic_xchg(from, 0);
94         while (log) {
95             int bit = ctzl(log);
96             hwaddr page_addr;
97             hwaddr section_offset;
98             hwaddr mr_offset;
99             page_addr = addr + bit * VHOST_LOG_PAGE;
100             section_offset = page_addr - section->offset_within_address_space;
101             mr_offset = section_offset + section->offset_within_region;
102             memory_region_set_dirty(section->mr, mr_offset, VHOST_LOG_PAGE);
103             log &= ~(0x1ull << bit);
104         }
105         addr += VHOST_LOG_CHUNK;
106     }
107 }
108 
109 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     return log_size;
176 }
177 
178 static struct vhost_log *vhost_log_alloc(uint64_t size, bool share)
179 {
180     Error *err = NULL;
181     struct vhost_log *log;
182     uint64_t logsize = size * sizeof(*(log->log));
183     int fd = -1;
184 
185     log = g_new0(struct vhost_log, 1);
186     if (share) {
187         log->log = qemu_memfd_alloc("vhost-log", logsize,
188                                     F_SEAL_GROW | F_SEAL_SHRINK | F_SEAL_SEAL,
189                                     &fd, &err);
190         if (err) {
191             error_report_err(err);
192             g_free(log);
193             return NULL;
194         }
195         memset(log->log, 0, logsize);
196     } else {
197         log->log = g_malloc0(logsize);
198     }
199 
200     log->size = size;
201     log->refcnt = 1;
202     log->fd = fd;
203 
204     return log;
205 }
206 
207 static struct vhost_log *vhost_log_get(uint64_t size, bool share)
208 {
209     struct vhost_log *log = share ? vhost_log_shm : vhost_log;
210 
211     if (!log || log->size != size) {
212         log = vhost_log_alloc(size, share);
213         if (share) {
214             vhost_log_shm = log;
215         } else {
216             vhost_log = log;
217         }
218     } else {
219         ++log->refcnt;
220     }
221 
222     return log;
223 }
224 
225 static void vhost_log_put(struct vhost_dev *dev, bool sync)
226 {
227     struct vhost_log *log = dev->log;
228 
229     if (!log) {
230         return;
231     }
232 
233     --log->refcnt;
234     if (log->refcnt == 0) {
235         /* Sync only the range covered by the old log */
236         if (dev->log_size && sync) {
237             vhost_log_sync_range(dev, 0, dev->log_size * VHOST_LOG_CHUNK - 1);
238         }
239 
240         if (vhost_log == log) {
241             g_free(log->log);
242             vhost_log = NULL;
243         } else if (vhost_log_shm == log) {
244             qemu_memfd_free(log->log, log->size * sizeof(*(log->log)),
245                             log->fd);
246             vhost_log_shm = NULL;
247         }
248 
249         g_free(log);
250     }
251 
252     dev->log = NULL;
253     dev->log_size = 0;
254 }
255 
256 static bool vhost_dev_log_is_shared(struct vhost_dev *dev)
257 {
258     return dev->vhost_ops->vhost_requires_shm_log &&
259            dev->vhost_ops->vhost_requires_shm_log(dev);
260 }
261 
262 static inline void vhost_dev_log_resize(struct vhost_dev *dev, uint64_t size)
263 {
264     struct vhost_log *log = vhost_log_get(size, vhost_dev_log_is_shared(dev));
265     uint64_t log_base = (uintptr_t)log->log;
266     int r;
267 
268     /* inform backend of log switching, this must be done before
269        releasing the current log, to ensure no logging is lost */
270     r = dev->vhost_ops->vhost_set_log_base(dev, log_base, log);
271     if (r < 0) {
272         VHOST_OPS_DEBUG("vhost_set_log_base failed");
273     }
274 
275     vhost_log_put(dev, true);
276     dev->log = log;
277     dev->log_size = size;
278 }
279 
280 static int vhost_dev_has_iommu(struct vhost_dev *dev)
281 {
282     VirtIODevice *vdev = dev->vdev;
283 
284     /*
285      * For vhost, VIRTIO_F_IOMMU_PLATFORM means the backend support
286      * incremental memory mapping API via IOTLB API. For platform that
287      * does not have IOMMU, there's no need to enable this feature
288      * which may cause unnecessary IOTLB miss/update trnasactions.
289      */
290     return vdev->dma_as != &address_space_memory &&
291            virtio_host_has_feature(vdev, VIRTIO_F_IOMMU_PLATFORM);
292 }
293 
294 static void *vhost_memory_map(struct vhost_dev *dev, hwaddr addr,
295                               hwaddr *plen, bool is_write)
296 {
297     if (!vhost_dev_has_iommu(dev)) {
298         return cpu_physical_memory_map(addr, plen, is_write);
299     } else {
300         return (void *)(uintptr_t)addr;
301     }
302 }
303 
304 static void vhost_memory_unmap(struct vhost_dev *dev, void *buffer,
305                                hwaddr len, int is_write,
306                                hwaddr access_len)
307 {
308     if (!vhost_dev_has_iommu(dev)) {
309         cpu_physical_memory_unmap(buffer, len, is_write, access_len);
310     }
311 }
312 
313 static int vhost_verify_ring_part_mapping(void *ring_hva,
314                                           uint64_t ring_gpa,
315                                           uint64_t ring_size,
316                                           void *reg_hva,
317                                           uint64_t reg_gpa,
318                                           uint64_t reg_size)
319 {
320     uint64_t hva_ring_offset;
321     uint64_t ring_last = range_get_last(ring_gpa, ring_size);
322     uint64_t reg_last = range_get_last(reg_gpa, reg_size);
323 
324     if (ring_last < reg_gpa || ring_gpa > reg_last) {
325         return 0;
326     }
327     /* check that whole ring's is mapped */
328     if (ring_last > reg_last) {
329         return -ENOMEM;
330     }
331     /* check that ring's MemoryRegion wasn't replaced */
332     hva_ring_offset = ring_gpa - reg_gpa;
333     if (ring_hva != reg_hva + hva_ring_offset) {
334         return -EBUSY;
335     }
336 
337     return 0;
338 }
339 
340 static int vhost_verify_ring_mappings(struct vhost_dev *dev,
341                                       void *reg_hva,
342                                       uint64_t reg_gpa,
343                                       uint64_t reg_size)
344 {
345     int i, j;
346     int r = 0;
347     const char *part_name[] = {
348         "descriptor table",
349         "available ring",
350         "used ring"
351     };
352 
353     if (vhost_dev_has_iommu(dev)) {
354         return 0;
355     }
356 
357     for (i = 0; i < dev->nvqs; ++i) {
358         struct vhost_virtqueue *vq = dev->vqs + i;
359 
360         if (vq->desc_phys == 0) {
361             continue;
362         }
363 
364         j = 0;
365         r = vhost_verify_ring_part_mapping(
366                 vq->desc, vq->desc_phys, vq->desc_size,
367                 reg_hva, reg_gpa, reg_size);
368         if (r) {
369             break;
370         }
371 
372         j++;
373         r = vhost_verify_ring_part_mapping(
374                 vq->avail, vq->avail_phys, vq->avail_size,
375                 reg_hva, reg_gpa, reg_size);
376         if (r) {
377             break;
378         }
379 
380         j++;
381         r = vhost_verify_ring_part_mapping(
382                 vq->used, vq->used_phys, vq->used_size,
383                 reg_hva, reg_gpa, reg_size);
384         if (r) {
385             break;
386         }
387     }
388 
389     if (r == -ENOMEM) {
390         error_report("Unable to map %s for ring %d", part_name[j], i);
391     } else if (r == -EBUSY) {
392         error_report("%s relocated for ring %d", part_name[j], i);
393     }
394     return r;
395 }
396 
397 /*
398  * vhost_section: identify sections needed for vhost access
399  *
400  * We only care about RAM sections here (where virtqueue and guest
401  * internals accessed by virtio might live). If we find one we still
402  * allow the backend to potentially filter it out of our list.
403  */
404 static bool vhost_section(struct vhost_dev *dev, MemoryRegionSection *section)
405 {
406     MemoryRegion *mr = section->mr;
407 
408     if (memory_region_is_ram(mr) && !memory_region_is_rom(mr)) {
409         uint8_t dirty_mask = memory_region_get_dirty_log_mask(mr);
410         uint8_t handled_dirty;
411 
412         /*
413          * Kernel based vhost doesn't handle any block which is doing
414          * dirty-tracking other than migration for which it has
415          * specific logging support. However for TCG the kernel never
416          * gets involved anyway so we can also ignore it's
417          * self-modiying code detection flags. However a vhost-user
418          * client could still confuse a TCG guest if it re-writes
419          * executable memory that has already been translated.
420          */
421         handled_dirty = (1 << DIRTY_MEMORY_MIGRATION) |
422             (1 << DIRTY_MEMORY_CODE);
423 
424         if (dirty_mask & ~handled_dirty) {
425             trace_vhost_reject_section(mr->name, 1);
426             return false;
427         }
428 
429         if (dev->vhost_ops->vhost_backend_mem_section_filter &&
430             !dev->vhost_ops->vhost_backend_mem_section_filter(dev, section)) {
431             trace_vhost_reject_section(mr->name, 2);
432             return false;
433         }
434 
435         trace_vhost_section(mr->name);
436         return true;
437     } else {
438         trace_vhost_reject_section(mr->name, 3);
439         return false;
440     }
441 }
442 
443 static void vhost_begin(MemoryListener *listener)
444 {
445     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
446                                          memory_listener);
447     dev->tmp_sections = NULL;
448     dev->n_tmp_sections = 0;
449 }
450 
451 static void vhost_commit(MemoryListener *listener)
452 {
453     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
454                                          memory_listener);
455     MemoryRegionSection *old_sections;
456     int n_old_sections;
457     uint64_t log_size;
458     size_t regions_size;
459     int r;
460     int i;
461     bool changed = false;
462 
463     /* Note we can be called before the device is started, but then
464      * starting the device calls set_mem_table, so we need to have
465      * built the data structures.
466      */
467     old_sections = dev->mem_sections;
468     n_old_sections = dev->n_mem_sections;
469     dev->mem_sections = dev->tmp_sections;
470     dev->n_mem_sections = dev->n_tmp_sections;
471 
472     if (dev->n_mem_sections != n_old_sections) {
473         changed = true;
474     } else {
475         /* Same size, lets check the contents */
476         for (int i = 0; i < n_old_sections; i++) {
477             if (!MemoryRegionSection_eq(&old_sections[i],
478                                         &dev->mem_sections[i])) {
479                 changed = true;
480                 break;
481             }
482         }
483     }
484 
485     trace_vhost_commit(dev->started, changed);
486     if (!changed) {
487         goto out;
488     }
489 
490     /* Rebuild the regions list from the new sections list */
491     regions_size = offsetof(struct vhost_memory, regions) +
492                        dev->n_mem_sections * sizeof dev->mem->regions[0];
493     dev->mem = g_realloc(dev->mem, regions_size);
494     dev->mem->nregions = dev->n_mem_sections;
495     used_memslots = dev->mem->nregions;
496     for (i = 0; i < dev->n_mem_sections; i++) {
497         struct vhost_memory_region *cur_vmr = dev->mem->regions + i;
498         struct MemoryRegionSection *mrs = dev->mem_sections + i;
499 
500         cur_vmr->guest_phys_addr = mrs->offset_within_address_space;
501         cur_vmr->memory_size     = int128_get64(mrs->size);
502         cur_vmr->userspace_addr  =
503             (uintptr_t)memory_region_get_ram_ptr(mrs->mr) +
504             mrs->offset_within_region;
505         cur_vmr->flags_padding   = 0;
506     }
507 
508     if (!dev->started) {
509         goto out;
510     }
511 
512     for (i = 0; i < dev->mem->nregions; i++) {
513         if (vhost_verify_ring_mappings(dev,
514                        (void *)(uintptr_t)dev->mem->regions[i].userspace_addr,
515                        dev->mem->regions[i].guest_phys_addr,
516                        dev->mem->regions[i].memory_size)) {
517             error_report("Verify ring failure on region %d", i);
518             abort();
519         }
520     }
521 
522     if (!dev->log_enabled) {
523         r = dev->vhost_ops->vhost_set_mem_table(dev, dev->mem);
524         if (r < 0) {
525             VHOST_OPS_DEBUG("vhost_set_mem_table failed");
526         }
527         goto out;
528     }
529     log_size = vhost_get_log_size(dev);
530     /* We allocate an extra 4K bytes to log,
531      * to reduce the * number of reallocations. */
532 #define VHOST_LOG_BUFFER (0x1000 / sizeof *dev->log)
533     /* To log more, must increase log size before table update. */
534     if (dev->log_size < log_size) {
535         vhost_dev_log_resize(dev, log_size + VHOST_LOG_BUFFER);
536     }
537     r = dev->vhost_ops->vhost_set_mem_table(dev, dev->mem);
538     if (r < 0) {
539         VHOST_OPS_DEBUG("vhost_set_mem_table failed");
540     }
541     /* To log less, can only decrease log size after table update. */
542     if (dev->log_size > log_size + VHOST_LOG_BUFFER) {
543         vhost_dev_log_resize(dev, log_size);
544     }
545 
546 out:
547     /* Deref the old list of sections, this must happen _after_ the
548      * vhost_set_mem_table to ensure the client isn't still using the
549      * section we're about to unref.
550      */
551     while (n_old_sections--) {
552         memory_region_unref(old_sections[n_old_sections].mr);
553     }
554     g_free(old_sections);
555     return;
556 }
557 
558 /* Adds the section data to the tmp_section structure.
559  * It relies on the listener calling us in memory address order
560  * and for each region (via the _add and _nop methods) to
561  * join neighbours.
562  */
563 static void vhost_region_add_section(struct vhost_dev *dev,
564                                      MemoryRegionSection *section)
565 {
566     bool need_add = true;
567     uint64_t mrs_size = int128_get64(section->size);
568     uint64_t mrs_gpa = section->offset_within_address_space;
569     uintptr_t mrs_host = (uintptr_t)memory_region_get_ram_ptr(section->mr) +
570                          section->offset_within_region;
571     RAMBlock *mrs_rb = section->mr->ram_block;
572 
573     trace_vhost_region_add_section(section->mr->name, mrs_gpa, mrs_size,
574                                    mrs_host);
575 
576     if (dev->vhost_ops->backend_type == VHOST_BACKEND_TYPE_USER) {
577         /* Round the section to it's page size */
578         /* First align the start down to a page boundary */
579         size_t mrs_page = qemu_ram_pagesize(mrs_rb);
580         uint64_t alignage = mrs_host & (mrs_page - 1);
581         if (alignage) {
582             mrs_host -= alignage;
583             mrs_size += alignage;
584             mrs_gpa  -= alignage;
585         }
586         /* Now align the size up to a page boundary */
587         alignage = mrs_size & (mrs_page - 1);
588         if (alignage) {
589             mrs_size += mrs_page - alignage;
590         }
591         trace_vhost_region_add_section_aligned(section->mr->name, mrs_gpa,
592                                                mrs_size, mrs_host);
593     }
594 
595     if (dev->n_tmp_sections) {
596         /* Since we already have at least one section, lets see if
597          * this extends it; since we're scanning in order, we only
598          * have to look at the last one, and the FlatView that calls
599          * us shouldn't have overlaps.
600          */
601         MemoryRegionSection *prev_sec = dev->tmp_sections +
602                                                (dev->n_tmp_sections - 1);
603         uint64_t prev_gpa_start = prev_sec->offset_within_address_space;
604         uint64_t prev_size = int128_get64(prev_sec->size);
605         uint64_t prev_gpa_end   = range_get_last(prev_gpa_start, prev_size);
606         uint64_t prev_host_start =
607                         (uintptr_t)memory_region_get_ram_ptr(prev_sec->mr) +
608                         prev_sec->offset_within_region;
609         uint64_t prev_host_end   = range_get_last(prev_host_start, prev_size);
610 
611         if (mrs_gpa <= (prev_gpa_end + 1)) {
612             /* OK, looks like overlapping/intersecting - it's possible that
613              * the rounding to page sizes has made them overlap, but they should
614              * match up in the same RAMBlock if they do.
615              */
616             if (mrs_gpa < prev_gpa_start) {
617                 error_report("%s:Section '%s' rounded to %"PRIx64
618                              " prior to previous '%s' %"PRIx64,
619                              __func__, section->mr->name, mrs_gpa,
620                              prev_sec->mr->name, prev_gpa_start);
621                 /* A way to cleanly fail here would be better */
622                 return;
623             }
624             /* Offset from the start of the previous GPA to this GPA */
625             size_t offset = mrs_gpa - prev_gpa_start;
626 
627             if (prev_host_start + offset == mrs_host &&
628                 section->mr == prev_sec->mr &&
629                 (!dev->vhost_ops->vhost_backend_can_merge ||
630                  dev->vhost_ops->vhost_backend_can_merge(dev,
631                     mrs_host, mrs_size,
632                     prev_host_start, prev_size))) {
633                 uint64_t max_end = MAX(prev_host_end, mrs_host + mrs_size);
634                 need_add = false;
635                 prev_sec->offset_within_address_space =
636                     MIN(prev_gpa_start, mrs_gpa);
637                 prev_sec->offset_within_region =
638                     MIN(prev_host_start, mrs_host) -
639                     (uintptr_t)memory_region_get_ram_ptr(prev_sec->mr);
640                 prev_sec->size = int128_make64(max_end - MIN(prev_host_start,
641                                                mrs_host));
642                 trace_vhost_region_add_section_merge(section->mr->name,
643                                         int128_get64(prev_sec->size),
644                                         prev_sec->offset_within_address_space,
645                                         prev_sec->offset_within_region);
646             } else {
647                 /* adjoining regions are fine, but overlapping ones with
648                  * different blocks/offsets shouldn't happen
649                  */
650                 if (mrs_gpa != prev_gpa_end + 1) {
651                     error_report("%s: Overlapping but not coherent sections "
652                                  "at %"PRIx64,
653                                  __func__, mrs_gpa);
654                     return;
655                 }
656             }
657         }
658     }
659 
660     if (need_add) {
661         ++dev->n_tmp_sections;
662         dev->tmp_sections = g_renew(MemoryRegionSection, dev->tmp_sections,
663                                     dev->n_tmp_sections);
664         dev->tmp_sections[dev->n_tmp_sections - 1] = *section;
665         /* The flatview isn't stable and we don't use it, making it NULL
666          * means we can memcmp the list.
667          */
668         dev->tmp_sections[dev->n_tmp_sections - 1].fv = NULL;
669         memory_region_ref(section->mr);
670     }
671 }
672 
673 /* Used for both add and nop callbacks */
674 static void vhost_region_addnop(MemoryListener *listener,
675                                 MemoryRegionSection *section)
676 {
677     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
678                                          memory_listener);
679 
680     if (!vhost_section(dev, section)) {
681         return;
682     }
683     vhost_region_add_section(dev, section);
684 }
685 
686 static void vhost_iommu_unmap_notify(IOMMUNotifier *n, IOMMUTLBEntry *iotlb)
687 {
688     struct vhost_iommu *iommu = container_of(n, struct vhost_iommu, n);
689     struct vhost_dev *hdev = iommu->hdev;
690     hwaddr iova = iotlb->iova + iommu->iommu_offset;
691 
692     if (vhost_backend_invalidate_device_iotlb(hdev, iova,
693                                               iotlb->addr_mask + 1)) {
694         error_report("Fail to invalidate device iotlb");
695     }
696 }
697 
698 static void vhost_iommu_region_add(MemoryListener *listener,
699                                    MemoryRegionSection *section)
700 {
701     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
702                                          iommu_listener);
703     struct vhost_iommu *iommu;
704     Int128 end;
705     int iommu_idx;
706     IOMMUMemoryRegion *iommu_mr;
707 
708     if (!memory_region_is_iommu(section->mr)) {
709         return;
710     }
711 
712     iommu_mr = IOMMU_MEMORY_REGION(section->mr);
713 
714     iommu = g_malloc0(sizeof(*iommu));
715     end = int128_add(int128_make64(section->offset_within_region),
716                      section->size);
717     end = int128_sub(end, int128_one());
718     iommu_idx = memory_region_iommu_attrs_to_index(iommu_mr,
719                                                    MEMTXATTRS_UNSPECIFIED);
720     iommu_notifier_init(&iommu->n, vhost_iommu_unmap_notify,
721                         IOMMU_NOTIFIER_DEVIOTLB_UNMAP,
722                         section->offset_within_region,
723                         int128_get64(end),
724                         iommu_idx);
725     iommu->mr = section->mr;
726     iommu->iommu_offset = section->offset_within_address_space -
727                           section->offset_within_region;
728     iommu->hdev = dev;
729     memory_region_register_iommu_notifier(section->mr, &iommu->n,
730                                           &error_fatal);
731     QLIST_INSERT_HEAD(&dev->iommu_list, iommu, iommu_next);
732     /* TODO: can replay help performance here? */
733 }
734 
735 static void vhost_iommu_region_del(MemoryListener *listener,
736                                    MemoryRegionSection *section)
737 {
738     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
739                                          iommu_listener);
740     struct vhost_iommu *iommu;
741 
742     if (!memory_region_is_iommu(section->mr)) {
743         return;
744     }
745 
746     QLIST_FOREACH(iommu, &dev->iommu_list, iommu_next) {
747         if (iommu->mr == section->mr &&
748             iommu->n.start == section->offset_within_region) {
749             memory_region_unregister_iommu_notifier(iommu->mr,
750                                                     &iommu->n);
751             QLIST_REMOVE(iommu, iommu_next);
752             g_free(iommu);
753             break;
754         }
755     }
756 }
757 
758 static int vhost_virtqueue_set_addr(struct vhost_dev *dev,
759                                     struct vhost_virtqueue *vq,
760                                     unsigned idx, bool enable_log)
761 {
762     struct vhost_vring_addr addr;
763     int r;
764     memset(&addr, 0, sizeof(struct vhost_vring_addr));
765 
766     if (dev->vhost_ops->vhost_vq_get_addr) {
767         r = dev->vhost_ops->vhost_vq_get_addr(dev, &addr, vq);
768         if (r < 0) {
769             VHOST_OPS_DEBUG("vhost_vq_get_addr failed");
770             return -errno;
771         }
772     } else {
773         addr.desc_user_addr = (uint64_t)(unsigned long)vq->desc;
774         addr.avail_user_addr = (uint64_t)(unsigned long)vq->avail;
775         addr.used_user_addr = (uint64_t)(unsigned long)vq->used;
776     }
777     addr.index = idx;
778     addr.log_guest_addr = vq->used_phys;
779     addr.flags = enable_log ? (1 << VHOST_VRING_F_LOG) : 0;
780     r = dev->vhost_ops->vhost_set_vring_addr(dev, &addr);
781     if (r < 0) {
782         VHOST_OPS_DEBUG("vhost_set_vring_addr failed");
783         return -errno;
784     }
785     return 0;
786 }
787 
788 static int vhost_dev_set_features(struct vhost_dev *dev,
789                                   bool enable_log)
790 {
791     uint64_t features = dev->acked_features;
792     int r;
793     if (enable_log) {
794         features |= 0x1ULL << VHOST_F_LOG_ALL;
795     }
796     if (!vhost_dev_has_iommu(dev)) {
797         features &= ~(0x1ULL << VIRTIO_F_IOMMU_PLATFORM);
798     }
799     if (dev->vhost_ops->vhost_force_iommu) {
800         if (dev->vhost_ops->vhost_force_iommu(dev) == true) {
801             features |= 0x1ULL << VIRTIO_F_IOMMU_PLATFORM;
802        }
803     }
804     r = dev->vhost_ops->vhost_set_features(dev, features);
805     if (r < 0) {
806         VHOST_OPS_DEBUG("vhost_set_features failed");
807         goto out;
808     }
809     if (dev->vhost_ops->vhost_set_backend_cap) {
810         r = dev->vhost_ops->vhost_set_backend_cap(dev);
811         if (r < 0) {
812             VHOST_OPS_DEBUG("vhost_set_backend_cap failed");
813             goto out;
814         }
815     }
816 
817 out:
818     return r < 0 ? -errno : 0;
819 }
820 
821 static int vhost_dev_set_log(struct vhost_dev *dev, bool enable_log)
822 {
823     int r, i, idx;
824     hwaddr addr;
825 
826     r = vhost_dev_set_features(dev, enable_log);
827     if (r < 0) {
828         goto err_features;
829     }
830     for (i = 0; i < dev->nvqs; ++i) {
831         idx = dev->vhost_ops->vhost_get_vq_index(dev, dev->vq_index + i);
832         addr = virtio_queue_get_desc_addr(dev->vdev, idx);
833         if (!addr) {
834             /*
835              * The queue might not be ready for start. If this
836              * is the case there is no reason to continue the process.
837              * The similar logic is used by the vhost_virtqueue_start()
838              * routine.
839              */
840             continue;
841         }
842         r = vhost_virtqueue_set_addr(dev, dev->vqs + i, idx,
843                                      enable_log);
844         if (r < 0) {
845             goto err_vq;
846         }
847     }
848     return 0;
849 err_vq:
850     for (; i >= 0; --i) {
851         idx = dev->vhost_ops->vhost_get_vq_index(dev, dev->vq_index + i);
852         vhost_virtqueue_set_addr(dev, dev->vqs + i, idx,
853                                  dev->log_enabled);
854     }
855     vhost_dev_set_features(dev, dev->log_enabled);
856 err_features:
857     return r;
858 }
859 
860 static int vhost_migration_log(MemoryListener *listener, bool enable)
861 {
862     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
863                                          memory_listener);
864     int r;
865     if (enable == dev->log_enabled) {
866         return 0;
867     }
868     if (!dev->started) {
869         dev->log_enabled = enable;
870         return 0;
871     }
872 
873     r = 0;
874     if (!enable) {
875         r = vhost_dev_set_log(dev, false);
876         if (r < 0) {
877             goto check_dev_state;
878         }
879         vhost_log_put(dev, false);
880     } else {
881         vhost_dev_log_resize(dev, vhost_get_log_size(dev));
882         r = vhost_dev_set_log(dev, true);
883         if (r < 0) {
884             goto check_dev_state;
885         }
886     }
887 
888 check_dev_state:
889     dev->log_enabled = enable;
890     /*
891      * vhost-user-* devices could change their state during log
892      * initialization due to disconnect. So check dev state after
893      * vhost communication.
894      */
895     if (!dev->started) {
896         /*
897          * Since device is in the stopped state, it is okay for
898          * migration. Return success.
899          */
900         r = 0;
901     }
902     if (r) {
903         /* An error is occured. */
904         dev->log_enabled = false;
905     }
906 
907     return r;
908 }
909 
910 static void vhost_log_global_start(MemoryListener *listener)
911 {
912     int r;
913 
914     r = vhost_migration_log(listener, true);
915     if (r < 0) {
916         abort();
917     }
918 }
919 
920 static void vhost_log_global_stop(MemoryListener *listener)
921 {
922     int r;
923 
924     r = vhost_migration_log(listener, false);
925     if (r < 0) {
926         abort();
927     }
928 }
929 
930 static void vhost_log_start(MemoryListener *listener,
931                             MemoryRegionSection *section,
932                             int old, int new)
933 {
934     /* FIXME: implement */
935 }
936 
937 static void vhost_log_stop(MemoryListener *listener,
938                            MemoryRegionSection *section,
939                            int old, int new)
940 {
941     /* FIXME: implement */
942 }
943 
944 /* The vhost driver natively knows how to handle the vrings of non
945  * cross-endian legacy devices and modern devices. Only legacy devices
946  * exposed to a bi-endian guest may require the vhost driver to use a
947  * specific endianness.
948  */
949 static inline bool vhost_needs_vring_endian(VirtIODevice *vdev)
950 {
951     if (virtio_vdev_has_feature(vdev, VIRTIO_F_VERSION_1)) {
952         return false;
953     }
954 #ifdef HOST_WORDS_BIGENDIAN
955     return vdev->device_endian == VIRTIO_DEVICE_ENDIAN_LITTLE;
956 #else
957     return vdev->device_endian == VIRTIO_DEVICE_ENDIAN_BIG;
958 #endif
959 }
960 
961 static int vhost_virtqueue_set_vring_endian_legacy(struct vhost_dev *dev,
962                                                    bool is_big_endian,
963                                                    int vhost_vq_index)
964 {
965     struct vhost_vring_state s = {
966         .index = vhost_vq_index,
967         .num = is_big_endian
968     };
969 
970     if (!dev->vhost_ops->vhost_set_vring_endian(dev, &s)) {
971         return 0;
972     }
973 
974     VHOST_OPS_DEBUG("vhost_set_vring_endian failed");
975     if (errno == ENOTTY) {
976         error_report("vhost does not support cross-endian");
977         return -ENOSYS;
978     }
979 
980     return -errno;
981 }
982 
983 static int vhost_memory_region_lookup(struct vhost_dev *hdev,
984                                       uint64_t gpa, uint64_t *uaddr,
985                                       uint64_t *len)
986 {
987     int i;
988 
989     for (i = 0; i < hdev->mem->nregions; i++) {
990         struct vhost_memory_region *reg = hdev->mem->regions + i;
991 
992         if (gpa >= reg->guest_phys_addr &&
993             reg->guest_phys_addr + reg->memory_size > gpa) {
994             *uaddr = reg->userspace_addr + gpa - reg->guest_phys_addr;
995             *len = reg->guest_phys_addr + reg->memory_size - gpa;
996             return 0;
997         }
998     }
999 
1000     return -EFAULT;
1001 }
1002 
1003 int vhost_device_iotlb_miss(struct vhost_dev *dev, uint64_t iova, int write)
1004 {
1005     IOMMUTLBEntry iotlb;
1006     uint64_t uaddr, len;
1007     int ret = -EFAULT;
1008 
1009     RCU_READ_LOCK_GUARD();
1010 
1011     trace_vhost_iotlb_miss(dev, 1);
1012 
1013     iotlb = address_space_get_iotlb_entry(dev->vdev->dma_as,
1014                                           iova, write,
1015                                           MEMTXATTRS_UNSPECIFIED);
1016     if (iotlb.target_as != NULL) {
1017         ret = vhost_memory_region_lookup(dev, iotlb.translated_addr,
1018                                          &uaddr, &len);
1019         if (ret) {
1020             trace_vhost_iotlb_miss(dev, 3);
1021             error_report("Fail to lookup the translated address "
1022                          "%"PRIx64, iotlb.translated_addr);
1023             goto out;
1024         }
1025 
1026         len = MIN(iotlb.addr_mask + 1, len);
1027         iova = iova & ~iotlb.addr_mask;
1028 
1029         ret = vhost_backend_update_device_iotlb(dev, iova, uaddr,
1030                                                 len, iotlb.perm);
1031         if (ret) {
1032             trace_vhost_iotlb_miss(dev, 4);
1033             error_report("Fail to update device iotlb");
1034             goto out;
1035         }
1036     }
1037 
1038     trace_vhost_iotlb_miss(dev, 2);
1039 
1040 out:
1041     return ret;
1042 }
1043 
1044 static int vhost_virtqueue_start(struct vhost_dev *dev,
1045                                 struct VirtIODevice *vdev,
1046                                 struct vhost_virtqueue *vq,
1047                                 unsigned idx)
1048 {
1049     BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev)));
1050     VirtioBusState *vbus = VIRTIO_BUS(qbus);
1051     VirtioBusClass *k = VIRTIO_BUS_GET_CLASS(vbus);
1052     hwaddr s, l, a;
1053     int r;
1054     int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, idx);
1055     struct vhost_vring_file file = {
1056         .index = vhost_vq_index
1057     };
1058     struct vhost_vring_state state = {
1059         .index = vhost_vq_index
1060     };
1061     struct VirtQueue *vvq = virtio_get_queue(vdev, idx);
1062 
1063     a = virtio_queue_get_desc_addr(vdev, idx);
1064     if (a == 0) {
1065         /* Queue might not be ready for start */
1066         return 0;
1067     }
1068 
1069     vq->num = state.num = virtio_queue_get_num(vdev, idx);
1070     r = dev->vhost_ops->vhost_set_vring_num(dev, &state);
1071     if (r) {
1072         VHOST_OPS_DEBUG("vhost_set_vring_num failed");
1073         return -errno;
1074     }
1075 
1076     state.num = virtio_queue_get_last_avail_idx(vdev, idx);
1077     r = dev->vhost_ops->vhost_set_vring_base(dev, &state);
1078     if (r) {
1079         VHOST_OPS_DEBUG("vhost_set_vring_base failed");
1080         return -errno;
1081     }
1082 
1083     if (vhost_needs_vring_endian(vdev)) {
1084         r = vhost_virtqueue_set_vring_endian_legacy(dev,
1085                                                     virtio_is_big_endian(vdev),
1086                                                     vhost_vq_index);
1087         if (r) {
1088             return -errno;
1089         }
1090     }
1091 
1092     vq->desc_size = s = l = virtio_queue_get_desc_size(vdev, idx);
1093     vq->desc_phys = a;
1094     vq->desc = vhost_memory_map(dev, a, &l, false);
1095     if (!vq->desc || l != s) {
1096         r = -ENOMEM;
1097         goto fail_alloc_desc;
1098     }
1099     vq->avail_size = s = l = virtio_queue_get_avail_size(vdev, idx);
1100     vq->avail_phys = a = virtio_queue_get_avail_addr(vdev, idx);
1101     vq->avail = vhost_memory_map(dev, a, &l, false);
1102     if (!vq->avail || l != s) {
1103         r = -ENOMEM;
1104         goto fail_alloc_avail;
1105     }
1106     vq->used_size = s = l = virtio_queue_get_used_size(vdev, idx);
1107     vq->used_phys = a = virtio_queue_get_used_addr(vdev, idx);
1108     vq->used = vhost_memory_map(dev, a, &l, true);
1109     if (!vq->used || l != s) {
1110         r = -ENOMEM;
1111         goto fail_alloc_used;
1112     }
1113 
1114     r = vhost_virtqueue_set_addr(dev, vq, vhost_vq_index, dev->log_enabled);
1115     if (r < 0) {
1116         r = -errno;
1117         goto fail_alloc;
1118     }
1119 
1120     file.fd = event_notifier_get_fd(virtio_queue_get_host_notifier(vvq));
1121     r = dev->vhost_ops->vhost_set_vring_kick(dev, &file);
1122     if (r) {
1123         VHOST_OPS_DEBUG("vhost_set_vring_kick failed");
1124         r = -errno;
1125         goto fail_kick;
1126     }
1127 
1128     /* Clear and discard previous events if any. */
1129     event_notifier_test_and_clear(&vq->masked_notifier);
1130 
1131     /* Init vring in unmasked state, unless guest_notifier_mask
1132      * will do it later.
1133      */
1134     if (!vdev->use_guest_notifier_mask) {
1135         /* TODO: check and handle errors. */
1136         vhost_virtqueue_mask(dev, vdev, idx, false);
1137     }
1138 
1139     if (k->query_guest_notifiers &&
1140         k->query_guest_notifiers(qbus->parent) &&
1141         virtio_queue_vector(vdev, idx) == VIRTIO_NO_VECTOR) {
1142         file.fd = -1;
1143         r = dev->vhost_ops->vhost_set_vring_call(dev, &file);
1144         if (r) {
1145             goto fail_vector;
1146         }
1147     }
1148 
1149     return 0;
1150 
1151 fail_vector:
1152 fail_kick:
1153 fail_alloc:
1154     vhost_memory_unmap(dev, vq->used, virtio_queue_get_used_size(vdev, idx),
1155                        0, 0);
1156 fail_alloc_used:
1157     vhost_memory_unmap(dev, vq->avail, virtio_queue_get_avail_size(vdev, idx),
1158                        0, 0);
1159 fail_alloc_avail:
1160     vhost_memory_unmap(dev, vq->desc, virtio_queue_get_desc_size(vdev, idx),
1161                        0, 0);
1162 fail_alloc_desc:
1163     return r;
1164 }
1165 
1166 static void vhost_virtqueue_stop(struct vhost_dev *dev,
1167                                     struct VirtIODevice *vdev,
1168                                     struct vhost_virtqueue *vq,
1169                                     unsigned idx)
1170 {
1171     int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, idx);
1172     struct vhost_vring_state state = {
1173         .index = vhost_vq_index,
1174     };
1175     int r;
1176 
1177     if (virtio_queue_get_desc_addr(vdev, idx) == 0) {
1178         /* Don't stop the virtqueue which might have not been started */
1179         return;
1180     }
1181 
1182     r = dev->vhost_ops->vhost_get_vring_base(dev, &state);
1183     if (r < 0) {
1184         VHOST_OPS_DEBUG("vhost VQ %u ring restore failed: %d", idx, r);
1185         /* Connection to the backend is broken, so let's sync internal
1186          * last avail idx to the device used idx.
1187          */
1188         virtio_queue_restore_last_avail_idx(vdev, idx);
1189     } else {
1190         virtio_queue_set_last_avail_idx(vdev, idx, state.num);
1191     }
1192     virtio_queue_invalidate_signalled_used(vdev, idx);
1193     virtio_queue_update_used_idx(vdev, idx);
1194 
1195     /* In the cross-endian case, we need to reset the vring endianness to
1196      * native as legacy devices expect so by default.
1197      */
1198     if (vhost_needs_vring_endian(vdev)) {
1199         vhost_virtqueue_set_vring_endian_legacy(dev,
1200                                                 !virtio_is_big_endian(vdev),
1201                                                 vhost_vq_index);
1202     }
1203 
1204     vhost_memory_unmap(dev, vq->used, virtio_queue_get_used_size(vdev, idx),
1205                        1, virtio_queue_get_used_size(vdev, idx));
1206     vhost_memory_unmap(dev, vq->avail, virtio_queue_get_avail_size(vdev, idx),
1207                        0, virtio_queue_get_avail_size(vdev, idx));
1208     vhost_memory_unmap(dev, vq->desc, virtio_queue_get_desc_size(vdev, idx),
1209                        0, virtio_queue_get_desc_size(vdev, idx));
1210 }
1211 
1212 static void vhost_eventfd_add(MemoryListener *listener,
1213                               MemoryRegionSection *section,
1214                               bool match_data, uint64_t data, EventNotifier *e)
1215 {
1216 }
1217 
1218 static void vhost_eventfd_del(MemoryListener *listener,
1219                               MemoryRegionSection *section,
1220                               bool match_data, uint64_t data, EventNotifier *e)
1221 {
1222 }
1223 
1224 static int vhost_virtqueue_set_busyloop_timeout(struct vhost_dev *dev,
1225                                                 int n, uint32_t timeout)
1226 {
1227     int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, n);
1228     struct vhost_vring_state state = {
1229         .index = vhost_vq_index,
1230         .num = timeout,
1231     };
1232     int r;
1233 
1234     if (!dev->vhost_ops->vhost_set_vring_busyloop_timeout) {
1235         return -EINVAL;
1236     }
1237 
1238     r = dev->vhost_ops->vhost_set_vring_busyloop_timeout(dev, &state);
1239     if (r) {
1240         VHOST_OPS_DEBUG("vhost_set_vring_busyloop_timeout failed");
1241         return r;
1242     }
1243 
1244     return 0;
1245 }
1246 
1247 static int vhost_virtqueue_init(struct vhost_dev *dev,
1248                                 struct vhost_virtqueue *vq, int n)
1249 {
1250     int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, n);
1251     struct vhost_vring_file file = {
1252         .index = vhost_vq_index,
1253     };
1254     int r = event_notifier_init(&vq->masked_notifier, 0);
1255     if (r < 0) {
1256         return r;
1257     }
1258 
1259     file.fd = event_notifier_get_fd(&vq->masked_notifier);
1260     r = dev->vhost_ops->vhost_set_vring_call(dev, &file);
1261     if (r) {
1262         VHOST_OPS_DEBUG("vhost_set_vring_call failed");
1263         r = -errno;
1264         goto fail_call;
1265     }
1266 
1267     vq->dev = dev;
1268 
1269     return 0;
1270 fail_call:
1271     event_notifier_cleanup(&vq->masked_notifier);
1272     return r;
1273 }
1274 
1275 static void vhost_virtqueue_cleanup(struct vhost_virtqueue *vq)
1276 {
1277     event_notifier_cleanup(&vq->masked_notifier);
1278 }
1279 
1280 int vhost_dev_init(struct vhost_dev *hdev, void *opaque,
1281                    VhostBackendType backend_type, uint32_t busyloop_timeout)
1282 {
1283     uint64_t features;
1284     int i, r, n_initialized_vqs = 0;
1285     Error *local_err = NULL;
1286 
1287     hdev->vdev = NULL;
1288     hdev->migration_blocker = NULL;
1289 
1290     r = vhost_set_backend_type(hdev, backend_type);
1291     assert(r >= 0);
1292 
1293     r = hdev->vhost_ops->vhost_backend_init(hdev, opaque);
1294     if (r < 0) {
1295         goto fail;
1296     }
1297 
1298     r = hdev->vhost_ops->vhost_set_owner(hdev);
1299     if (r < 0) {
1300         VHOST_OPS_DEBUG("vhost_set_owner failed");
1301         goto fail;
1302     }
1303 
1304     r = hdev->vhost_ops->vhost_get_features(hdev, &features);
1305     if (r < 0) {
1306         VHOST_OPS_DEBUG("vhost_get_features failed");
1307         goto fail;
1308     }
1309 
1310     for (i = 0; i < hdev->nvqs; ++i, ++n_initialized_vqs) {
1311         r = vhost_virtqueue_init(hdev, hdev->vqs + i, hdev->vq_index + i);
1312         if (r < 0) {
1313             goto fail;
1314         }
1315     }
1316 
1317     if (busyloop_timeout) {
1318         for (i = 0; i < hdev->nvqs; ++i) {
1319             r = vhost_virtqueue_set_busyloop_timeout(hdev, hdev->vq_index + i,
1320                                                      busyloop_timeout);
1321             if (r < 0) {
1322                 goto fail_busyloop;
1323             }
1324         }
1325     }
1326 
1327     hdev->features = features;
1328 
1329     hdev->memory_listener = (MemoryListener) {
1330         .begin = vhost_begin,
1331         .commit = vhost_commit,
1332         .region_add = vhost_region_addnop,
1333         .region_nop = vhost_region_addnop,
1334         .log_start = vhost_log_start,
1335         .log_stop = vhost_log_stop,
1336         .log_sync = vhost_log_sync,
1337         .log_global_start = vhost_log_global_start,
1338         .log_global_stop = vhost_log_global_stop,
1339         .eventfd_add = vhost_eventfd_add,
1340         .eventfd_del = vhost_eventfd_del,
1341         .priority = 10
1342     };
1343 
1344     hdev->iommu_listener = (MemoryListener) {
1345         .region_add = vhost_iommu_region_add,
1346         .region_del = vhost_iommu_region_del,
1347     };
1348 
1349     if (hdev->migration_blocker == NULL) {
1350         if (!(hdev->features & (0x1ULL << VHOST_F_LOG_ALL))) {
1351             error_setg(&hdev->migration_blocker,
1352                        "Migration disabled: vhost lacks VHOST_F_LOG_ALL feature.");
1353         } else if (vhost_dev_log_is_shared(hdev) && !qemu_memfd_alloc_check()) {
1354             error_setg(&hdev->migration_blocker,
1355                        "Migration disabled: failed to allocate shared memory");
1356         }
1357     }
1358 
1359     if (hdev->migration_blocker != NULL) {
1360         r = migrate_add_blocker(hdev->migration_blocker, &local_err);
1361         if (local_err) {
1362             error_report_err(local_err);
1363             error_free(hdev->migration_blocker);
1364             goto fail_busyloop;
1365         }
1366     }
1367 
1368     hdev->mem = g_malloc0(offsetof(struct vhost_memory, regions));
1369     hdev->n_mem_sections = 0;
1370     hdev->mem_sections = NULL;
1371     hdev->log = NULL;
1372     hdev->log_size = 0;
1373     hdev->log_enabled = false;
1374     hdev->started = false;
1375     memory_listener_register(&hdev->memory_listener, &address_space_memory);
1376     QLIST_INSERT_HEAD(&vhost_devices, hdev, entry);
1377 
1378     if (used_memslots > hdev->vhost_ops->vhost_backend_memslots_limit(hdev)) {
1379         error_report("vhost backend memory slots limit is less"
1380                 " than current number of present memory slots");
1381         r = -1;
1382         if (busyloop_timeout) {
1383             goto fail_busyloop;
1384         } else {
1385             goto fail;
1386         }
1387     }
1388 
1389     return 0;
1390 
1391 fail_busyloop:
1392     while (--i >= 0) {
1393         vhost_virtqueue_set_busyloop_timeout(hdev, hdev->vq_index + i, 0);
1394     }
1395 fail:
1396     hdev->nvqs = n_initialized_vqs;
1397     vhost_dev_cleanup(hdev);
1398     return r;
1399 }
1400 
1401 void vhost_dev_cleanup(struct vhost_dev *hdev)
1402 {
1403     int i;
1404 
1405     for (i = 0; i < hdev->nvqs; ++i) {
1406         vhost_virtqueue_cleanup(hdev->vqs + i);
1407     }
1408     if (hdev->mem) {
1409         /* those are only safe after successful init */
1410         memory_listener_unregister(&hdev->memory_listener);
1411         QLIST_REMOVE(hdev, entry);
1412     }
1413     if (hdev->migration_blocker) {
1414         migrate_del_blocker(hdev->migration_blocker);
1415         error_free(hdev->migration_blocker);
1416     }
1417     g_free(hdev->mem);
1418     g_free(hdev->mem_sections);
1419     if (hdev->vhost_ops) {
1420         hdev->vhost_ops->vhost_backend_cleanup(hdev);
1421     }
1422     assert(!hdev->log);
1423 
1424     memset(hdev, 0, sizeof(struct vhost_dev));
1425 }
1426 
1427 /* Stop processing guest IO notifications in qemu.
1428  * Start processing them in vhost in kernel.
1429  */
1430 int vhost_dev_enable_notifiers(struct vhost_dev *hdev, VirtIODevice *vdev)
1431 {
1432     BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev)));
1433     int i, r, e;
1434 
1435     /* We will pass the notifiers to the kernel, make sure that QEMU
1436      * doesn't interfere.
1437      */
1438     r = virtio_device_grab_ioeventfd(vdev);
1439     if (r < 0) {
1440         error_report("binding does not support host notifiers");
1441         goto fail;
1442     }
1443 
1444     for (i = 0; i < hdev->nvqs; ++i) {
1445         r = virtio_bus_set_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i,
1446                                          true);
1447         if (r < 0) {
1448             error_report("vhost VQ %d notifier binding failed: %d", i, -r);
1449             goto fail_vq;
1450         }
1451     }
1452 
1453     return 0;
1454 fail_vq:
1455     while (--i >= 0) {
1456         e = virtio_bus_set_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i,
1457                                          false);
1458         if (e < 0) {
1459             error_report("vhost VQ %d notifier cleanup error: %d", i, -r);
1460         }
1461         assert (e >= 0);
1462         virtio_bus_cleanup_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i);
1463     }
1464     virtio_device_release_ioeventfd(vdev);
1465 fail:
1466     return r;
1467 }
1468 
1469 /* Stop processing guest IO notifications in vhost.
1470  * Start processing them in qemu.
1471  * This might actually run the qemu handlers right away,
1472  * so virtio in qemu must be completely setup when this is called.
1473  */
1474 void vhost_dev_disable_notifiers(struct vhost_dev *hdev, VirtIODevice *vdev)
1475 {
1476     BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev)));
1477     int i, r;
1478 
1479     for (i = 0; i < hdev->nvqs; ++i) {
1480         r = virtio_bus_set_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i,
1481                                          false);
1482         if (r < 0) {
1483             error_report("vhost VQ %d notifier cleanup failed: %d", i, -r);
1484         }
1485         assert (r >= 0);
1486         virtio_bus_cleanup_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i);
1487     }
1488     virtio_device_release_ioeventfd(vdev);
1489 }
1490 
1491 /* Test and clear event pending status.
1492  * Should be called after unmask to avoid losing events.
1493  */
1494 bool vhost_virtqueue_pending(struct vhost_dev *hdev, int n)
1495 {
1496     struct vhost_virtqueue *vq = hdev->vqs + n - hdev->vq_index;
1497     assert(n >= hdev->vq_index && n < hdev->vq_index + hdev->nvqs);
1498     return event_notifier_test_and_clear(&vq->masked_notifier);
1499 }
1500 
1501 /* Mask/unmask events from this vq. */
1502 void vhost_virtqueue_mask(struct vhost_dev *hdev, VirtIODevice *vdev, int n,
1503                          bool mask)
1504 {
1505     struct VirtQueue *vvq = virtio_get_queue(vdev, n);
1506     int r, index = n - hdev->vq_index;
1507     struct vhost_vring_file file;
1508 
1509     /* should only be called after backend is connected */
1510     assert(hdev->vhost_ops);
1511 
1512     if (mask) {
1513         assert(vdev->use_guest_notifier_mask);
1514         file.fd = event_notifier_get_fd(&hdev->vqs[index].masked_notifier);
1515     } else {
1516         file.fd = event_notifier_get_fd(virtio_queue_get_guest_notifier(vvq));
1517     }
1518 
1519     file.index = hdev->vhost_ops->vhost_get_vq_index(hdev, n);
1520     r = hdev->vhost_ops->vhost_set_vring_call(hdev, &file);
1521     if (r < 0) {
1522         VHOST_OPS_DEBUG("vhost_set_vring_call failed");
1523     }
1524 }
1525 
1526 uint64_t vhost_get_features(struct vhost_dev *hdev, const int *feature_bits,
1527                             uint64_t features)
1528 {
1529     const int *bit = feature_bits;
1530     while (*bit != VHOST_INVALID_FEATURE_BIT) {
1531         uint64_t bit_mask = (1ULL << *bit);
1532         if (!(hdev->features & bit_mask)) {
1533             features &= ~bit_mask;
1534         }
1535         bit++;
1536     }
1537     return features;
1538 }
1539 
1540 void vhost_ack_features(struct vhost_dev *hdev, const int *feature_bits,
1541                         uint64_t features)
1542 {
1543     const int *bit = feature_bits;
1544     while (*bit != VHOST_INVALID_FEATURE_BIT) {
1545         uint64_t bit_mask = (1ULL << *bit);
1546         if (features & bit_mask) {
1547             hdev->acked_features |= bit_mask;
1548         }
1549         bit++;
1550     }
1551 }
1552 
1553 int vhost_dev_get_config(struct vhost_dev *hdev, uint8_t *config,
1554                          uint32_t config_len)
1555 {
1556     assert(hdev->vhost_ops);
1557 
1558     if (hdev->vhost_ops->vhost_get_config) {
1559         return hdev->vhost_ops->vhost_get_config(hdev, config, config_len);
1560     }
1561 
1562     return -1;
1563 }
1564 
1565 int vhost_dev_set_config(struct vhost_dev *hdev, const uint8_t *data,
1566                          uint32_t offset, uint32_t size, uint32_t flags)
1567 {
1568     assert(hdev->vhost_ops);
1569 
1570     if (hdev->vhost_ops->vhost_set_config) {
1571         return hdev->vhost_ops->vhost_set_config(hdev, data, offset,
1572                                                  size, flags);
1573     }
1574 
1575     return -1;
1576 }
1577 
1578 void vhost_dev_set_config_notifier(struct vhost_dev *hdev,
1579                                    const VhostDevConfigOps *ops)
1580 {
1581     hdev->config_ops = ops;
1582 }
1583 
1584 void vhost_dev_free_inflight(struct vhost_inflight *inflight)
1585 {
1586     if (inflight && inflight->addr) {
1587         qemu_memfd_free(inflight->addr, inflight->size, inflight->fd);
1588         inflight->addr = NULL;
1589         inflight->fd = -1;
1590     }
1591 }
1592 
1593 static int vhost_dev_resize_inflight(struct vhost_inflight *inflight,
1594                                      uint64_t new_size)
1595 {
1596     Error *err = NULL;
1597     int fd = -1;
1598     void *addr = qemu_memfd_alloc("vhost-inflight", new_size,
1599                                   F_SEAL_GROW | F_SEAL_SHRINK | F_SEAL_SEAL,
1600                                   &fd, &err);
1601 
1602     if (err) {
1603         error_report_err(err);
1604         return -1;
1605     }
1606 
1607     vhost_dev_free_inflight(inflight);
1608     inflight->offset = 0;
1609     inflight->addr = addr;
1610     inflight->fd = fd;
1611     inflight->size = new_size;
1612 
1613     return 0;
1614 }
1615 
1616 void vhost_dev_save_inflight(struct vhost_inflight *inflight, QEMUFile *f)
1617 {
1618     if (inflight->addr) {
1619         qemu_put_be64(f, inflight->size);
1620         qemu_put_be16(f, inflight->queue_size);
1621         qemu_put_buffer(f, inflight->addr, inflight->size);
1622     } else {
1623         qemu_put_be64(f, 0);
1624     }
1625 }
1626 
1627 int vhost_dev_load_inflight(struct vhost_inflight *inflight, QEMUFile *f)
1628 {
1629     uint64_t size;
1630 
1631     size = qemu_get_be64(f);
1632     if (!size) {
1633         return 0;
1634     }
1635 
1636     if (inflight->size != size) {
1637         if (vhost_dev_resize_inflight(inflight, size)) {
1638             return -1;
1639         }
1640     }
1641     inflight->queue_size = qemu_get_be16(f);
1642 
1643     qemu_get_buffer(f, inflight->addr, size);
1644 
1645     return 0;
1646 }
1647 
1648 int vhost_dev_prepare_inflight(struct vhost_dev *hdev, VirtIODevice *vdev)
1649 {
1650     int r;
1651 
1652     if (hdev->vhost_ops->vhost_get_inflight_fd == NULL ||
1653         hdev->vhost_ops->vhost_set_inflight_fd == NULL) {
1654         return 0;
1655     }
1656 
1657     hdev->vdev = vdev;
1658 
1659     r = vhost_dev_set_features(hdev, hdev->log_enabled);
1660     if (r < 0) {
1661         VHOST_OPS_DEBUG("vhost_dev_prepare_inflight failed");
1662         return r;
1663     }
1664 
1665     return 0;
1666 }
1667 
1668 int vhost_dev_set_inflight(struct vhost_dev *dev,
1669                            struct vhost_inflight *inflight)
1670 {
1671     int r;
1672 
1673     if (dev->vhost_ops->vhost_set_inflight_fd && inflight->addr) {
1674         r = dev->vhost_ops->vhost_set_inflight_fd(dev, inflight);
1675         if (r) {
1676             VHOST_OPS_DEBUG("vhost_set_inflight_fd failed");
1677             return -errno;
1678         }
1679     }
1680 
1681     return 0;
1682 }
1683 
1684 int vhost_dev_get_inflight(struct vhost_dev *dev, uint16_t queue_size,
1685                            struct vhost_inflight *inflight)
1686 {
1687     int r;
1688 
1689     if (dev->vhost_ops->vhost_get_inflight_fd) {
1690         r = dev->vhost_ops->vhost_get_inflight_fd(dev, queue_size, inflight);
1691         if (r) {
1692             VHOST_OPS_DEBUG("vhost_get_inflight_fd failed");
1693             return -errno;
1694         }
1695     }
1696 
1697     return 0;
1698 }
1699 
1700 /* Host notifiers must be enabled at this point. */
1701 int vhost_dev_start(struct vhost_dev *hdev, VirtIODevice *vdev)
1702 {
1703     int i, r;
1704 
1705     /* should only be called after backend is connected */
1706     assert(hdev->vhost_ops);
1707 
1708     hdev->started = true;
1709     hdev->vdev = vdev;
1710 
1711     r = vhost_dev_set_features(hdev, hdev->log_enabled);
1712     if (r < 0) {
1713         goto fail_features;
1714     }
1715 
1716     if (vhost_dev_has_iommu(hdev)) {
1717         memory_listener_register(&hdev->iommu_listener, vdev->dma_as);
1718     }
1719 
1720     r = hdev->vhost_ops->vhost_set_mem_table(hdev, hdev->mem);
1721     if (r < 0) {
1722         VHOST_OPS_DEBUG("vhost_set_mem_table failed");
1723         r = -errno;
1724         goto fail_mem;
1725     }
1726     for (i = 0; i < hdev->nvqs; ++i) {
1727         r = vhost_virtqueue_start(hdev,
1728                                   vdev,
1729                                   hdev->vqs + i,
1730                                   hdev->vq_index + i);
1731         if (r < 0) {
1732             goto fail_vq;
1733         }
1734     }
1735 
1736     if (hdev->log_enabled) {
1737         uint64_t log_base;
1738 
1739         hdev->log_size = vhost_get_log_size(hdev);
1740         hdev->log = vhost_log_get(hdev->log_size,
1741                                   vhost_dev_log_is_shared(hdev));
1742         log_base = (uintptr_t)hdev->log->log;
1743         r = hdev->vhost_ops->vhost_set_log_base(hdev,
1744                                                 hdev->log_size ? log_base : 0,
1745                                                 hdev->log);
1746         if (r < 0) {
1747             VHOST_OPS_DEBUG("vhost_set_log_base failed");
1748             r = -errno;
1749             goto fail_log;
1750         }
1751     }
1752     if (hdev->vhost_ops->vhost_dev_start) {
1753         r = hdev->vhost_ops->vhost_dev_start(hdev, true);
1754         if (r) {
1755             goto fail_log;
1756         }
1757     }
1758     if (vhost_dev_has_iommu(hdev) &&
1759         hdev->vhost_ops->vhost_set_iotlb_callback) {
1760             hdev->vhost_ops->vhost_set_iotlb_callback(hdev, true);
1761 
1762         /* Update used ring information for IOTLB to work correctly,
1763          * vhost-kernel code requires for this.*/
1764         for (i = 0; i < hdev->nvqs; ++i) {
1765             struct vhost_virtqueue *vq = hdev->vqs + i;
1766             vhost_device_iotlb_miss(hdev, vq->used_phys, true);
1767         }
1768     }
1769     return 0;
1770 fail_log:
1771     vhost_log_put(hdev, false);
1772 fail_vq:
1773     while (--i >= 0) {
1774         vhost_virtqueue_stop(hdev,
1775                              vdev,
1776                              hdev->vqs + i,
1777                              hdev->vq_index + i);
1778     }
1779 
1780 fail_mem:
1781 fail_features:
1782 
1783     hdev->started = false;
1784     return r;
1785 }
1786 
1787 /* Host notifiers must be enabled at this point. */
1788 void vhost_dev_stop(struct vhost_dev *hdev, VirtIODevice *vdev)
1789 {
1790     int i;
1791 
1792     /* should only be called after backend is connected */
1793     assert(hdev->vhost_ops);
1794 
1795     if (hdev->vhost_ops->vhost_dev_start) {
1796         hdev->vhost_ops->vhost_dev_start(hdev, false);
1797     }
1798     for (i = 0; i < hdev->nvqs; ++i) {
1799         vhost_virtqueue_stop(hdev,
1800                              vdev,
1801                              hdev->vqs + i,
1802                              hdev->vq_index + i);
1803     }
1804 
1805     if (vhost_dev_has_iommu(hdev)) {
1806         if (hdev->vhost_ops->vhost_set_iotlb_callback) {
1807             hdev->vhost_ops->vhost_set_iotlb_callback(hdev, false);
1808         }
1809         memory_listener_unregister(&hdev->iommu_listener);
1810     }
1811     vhost_log_put(hdev, true);
1812     hdev->started = false;
1813     hdev->vdev = NULL;
1814 }
1815 
1816 int vhost_net_set_backend(struct vhost_dev *hdev,
1817                           struct vhost_vring_file *file)
1818 {
1819     if (hdev->vhost_ops->vhost_net_set_backend) {
1820         return hdev->vhost_ops->vhost_net_set_backend(hdev, file);
1821     }
1822 
1823     return -1;
1824 }
1825