xref: /qemu/hw/virtio/vhost.c (revision 5be5f9be724d43fbec3d6f955a43bc64a62d55cc)
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 "hw/hw.h"
20 #include "qemu/atomic.h"
21 #include "qemu/range.h"
22 #include "qemu/error-report.h"
23 #include "qemu/memfd.h"
24 #include <linux/vhost.h>
25 #include "exec/address-spaces.h"
26 #include "hw/virtio/virtio-bus.h"
27 #include "hw/virtio/virtio-access.h"
28 #include "migration/migration.h"
29 
30 static struct vhost_log *vhost_log;
31 static struct vhost_log *vhost_log_shm;
32 
33 static unsigned int used_memslots;
34 static QLIST_HEAD(, vhost_dev) vhost_devices =
35     QLIST_HEAD_INITIALIZER(vhost_devices);
36 
37 bool vhost_has_free_slot(void)
38 {
39     unsigned int slots_limit = ~0U;
40     struct vhost_dev *hdev;
41 
42     QLIST_FOREACH(hdev, &vhost_devices, entry) {
43         unsigned int r = hdev->vhost_ops->vhost_backend_memslots_limit(hdev);
44         slots_limit = MIN(slots_limit, r);
45     }
46     return slots_limit > used_memslots;
47 }
48 
49 static void vhost_dev_sync_region(struct vhost_dev *dev,
50                                   MemoryRegionSection *section,
51                                   uint64_t mfirst, uint64_t mlast,
52                                   uint64_t rfirst, uint64_t rlast)
53 {
54     vhost_log_chunk_t *log = dev->log->log;
55 
56     uint64_t start = MAX(mfirst, rfirst);
57     uint64_t end = MIN(mlast, rlast);
58     vhost_log_chunk_t *from = log + start / VHOST_LOG_CHUNK;
59     vhost_log_chunk_t *to = log + end / VHOST_LOG_CHUNK + 1;
60     uint64_t addr = (start / VHOST_LOG_CHUNK) * VHOST_LOG_CHUNK;
61 
62     if (end < start) {
63         return;
64     }
65     assert(end / VHOST_LOG_CHUNK < dev->log_size);
66     assert(start / VHOST_LOG_CHUNK < dev->log_size);
67 
68     for (;from < to; ++from) {
69         vhost_log_chunk_t log;
70         /* We first check with non-atomic: much cheaper,
71          * and we expect non-dirty to be the common case. */
72         if (!*from) {
73             addr += VHOST_LOG_CHUNK;
74             continue;
75         }
76         /* Data must be read atomically. We don't really need barrier semantics
77          * but it's easier to use atomic_* than roll our own. */
78         log = atomic_xchg(from, 0);
79         while (log) {
80             int bit = ctzl(log);
81             hwaddr page_addr;
82             hwaddr section_offset;
83             hwaddr mr_offset;
84             page_addr = addr + bit * VHOST_LOG_PAGE;
85             section_offset = page_addr - section->offset_within_address_space;
86             mr_offset = section_offset + section->offset_within_region;
87             memory_region_set_dirty(section->mr, mr_offset, VHOST_LOG_PAGE);
88             log &= ~(0x1ull << bit);
89         }
90         addr += VHOST_LOG_CHUNK;
91     }
92 }
93 
94 static int vhost_sync_dirty_bitmap(struct vhost_dev *dev,
95                                    MemoryRegionSection *section,
96                                    hwaddr first,
97                                    hwaddr last)
98 {
99     int i;
100     hwaddr start_addr;
101     hwaddr end_addr;
102 
103     if (!dev->log_enabled || !dev->started) {
104         return 0;
105     }
106     start_addr = section->offset_within_address_space;
107     end_addr = range_get_last(start_addr, int128_get64(section->size));
108     start_addr = MAX(first, start_addr);
109     end_addr = MIN(last, end_addr);
110 
111     for (i = 0; i < dev->mem->nregions; ++i) {
112         struct vhost_memory_region *reg = dev->mem->regions + i;
113         vhost_dev_sync_region(dev, section, start_addr, end_addr,
114                               reg->guest_phys_addr,
115                               range_get_last(reg->guest_phys_addr,
116                                              reg->memory_size));
117     }
118     for (i = 0; i < dev->nvqs; ++i) {
119         struct vhost_virtqueue *vq = dev->vqs + i;
120         vhost_dev_sync_region(dev, section, start_addr, end_addr, vq->used_phys,
121                               range_get_last(vq->used_phys, vq->used_size));
122     }
123     return 0;
124 }
125 
126 static void vhost_log_sync(MemoryListener *listener,
127                           MemoryRegionSection *section)
128 {
129     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
130                                          memory_listener);
131     vhost_sync_dirty_bitmap(dev, section, 0x0, ~0x0ULL);
132 }
133 
134 static void vhost_log_sync_range(struct vhost_dev *dev,
135                                  hwaddr first, hwaddr last)
136 {
137     int i;
138     /* FIXME: this is N^2 in number of sections */
139     for (i = 0; i < dev->n_mem_sections; ++i) {
140         MemoryRegionSection *section = &dev->mem_sections[i];
141         vhost_sync_dirty_bitmap(dev, section, first, last);
142     }
143 }
144 
145 /* Assign/unassign. Keep an unsorted array of non-overlapping
146  * memory regions in dev->mem. */
147 static void vhost_dev_unassign_memory(struct vhost_dev *dev,
148                                       uint64_t start_addr,
149                                       uint64_t size)
150 {
151     int from, to, n = dev->mem->nregions;
152     /* Track overlapping/split regions for sanity checking. */
153     int overlap_start = 0, overlap_end = 0, overlap_middle = 0, split = 0;
154 
155     for (from = 0, to = 0; from < n; ++from, ++to) {
156         struct vhost_memory_region *reg = dev->mem->regions + to;
157         uint64_t reglast;
158         uint64_t memlast;
159         uint64_t change;
160 
161         /* clone old region */
162         if (to != from) {
163             memcpy(reg, dev->mem->regions + from, sizeof *reg);
164         }
165 
166         /* No overlap is simple */
167         if (!ranges_overlap(reg->guest_phys_addr, reg->memory_size,
168                             start_addr, size)) {
169             continue;
170         }
171 
172         /* Split only happens if supplied region
173          * is in the middle of an existing one. Thus it can not
174          * overlap with any other existing region. */
175         assert(!split);
176 
177         reglast = range_get_last(reg->guest_phys_addr, reg->memory_size);
178         memlast = range_get_last(start_addr, size);
179 
180         /* Remove whole region */
181         if (start_addr <= reg->guest_phys_addr && memlast >= reglast) {
182             --dev->mem->nregions;
183             --to;
184             ++overlap_middle;
185             continue;
186         }
187 
188         /* Shrink region */
189         if (memlast >= reglast) {
190             reg->memory_size = start_addr - reg->guest_phys_addr;
191             assert(reg->memory_size);
192             assert(!overlap_end);
193             ++overlap_end;
194             continue;
195         }
196 
197         /* Shift region */
198         if (start_addr <= reg->guest_phys_addr) {
199             change = memlast + 1 - reg->guest_phys_addr;
200             reg->memory_size -= change;
201             reg->guest_phys_addr += change;
202             reg->userspace_addr += change;
203             assert(reg->memory_size);
204             assert(!overlap_start);
205             ++overlap_start;
206             continue;
207         }
208 
209         /* This only happens if supplied region
210          * is in the middle of an existing one. Thus it can not
211          * overlap with any other existing region. */
212         assert(!overlap_start);
213         assert(!overlap_end);
214         assert(!overlap_middle);
215         /* Split region: shrink first part, shift second part. */
216         memcpy(dev->mem->regions + n, reg, sizeof *reg);
217         reg->memory_size = start_addr - reg->guest_phys_addr;
218         assert(reg->memory_size);
219         change = memlast + 1 - reg->guest_phys_addr;
220         reg = dev->mem->regions + n;
221         reg->memory_size -= change;
222         assert(reg->memory_size);
223         reg->guest_phys_addr += change;
224         reg->userspace_addr += change;
225         /* Never add more than 1 region */
226         assert(dev->mem->nregions == n);
227         ++dev->mem->nregions;
228         ++split;
229     }
230 }
231 
232 /* Called after unassign, so no regions overlap the given range. */
233 static void vhost_dev_assign_memory(struct vhost_dev *dev,
234                                     uint64_t start_addr,
235                                     uint64_t size,
236                                     uint64_t uaddr)
237 {
238     int from, to;
239     struct vhost_memory_region *merged = NULL;
240     for (from = 0, to = 0; from < dev->mem->nregions; ++from, ++to) {
241         struct vhost_memory_region *reg = dev->mem->regions + to;
242         uint64_t prlast, urlast;
243         uint64_t pmlast, umlast;
244         uint64_t s, e, u;
245 
246         /* clone old region */
247         if (to != from) {
248             memcpy(reg, dev->mem->regions + from, sizeof *reg);
249         }
250         prlast = range_get_last(reg->guest_phys_addr, reg->memory_size);
251         pmlast = range_get_last(start_addr, size);
252         urlast = range_get_last(reg->userspace_addr, reg->memory_size);
253         umlast = range_get_last(uaddr, size);
254 
255         /* check for overlapping regions: should never happen. */
256         assert(prlast < start_addr || pmlast < reg->guest_phys_addr);
257         /* Not an adjacent or overlapping region - do not merge. */
258         if ((prlast + 1 != start_addr || urlast + 1 != uaddr) &&
259             (pmlast + 1 != reg->guest_phys_addr ||
260              umlast + 1 != reg->userspace_addr)) {
261             continue;
262         }
263 
264         if (dev->vhost_ops->vhost_backend_can_merge &&
265             !dev->vhost_ops->vhost_backend_can_merge(dev, uaddr, size,
266                                                      reg->userspace_addr,
267                                                      reg->memory_size)) {
268             continue;
269         }
270 
271         if (merged) {
272             --to;
273             assert(to >= 0);
274         } else {
275             merged = reg;
276         }
277         u = MIN(uaddr, reg->userspace_addr);
278         s = MIN(start_addr, reg->guest_phys_addr);
279         e = MAX(pmlast, prlast);
280         uaddr = merged->userspace_addr = u;
281         start_addr = merged->guest_phys_addr = s;
282         size = merged->memory_size = e - s + 1;
283         assert(merged->memory_size);
284     }
285 
286     if (!merged) {
287         struct vhost_memory_region *reg = dev->mem->regions + to;
288         memset(reg, 0, sizeof *reg);
289         reg->memory_size = size;
290         assert(reg->memory_size);
291         reg->guest_phys_addr = start_addr;
292         reg->userspace_addr = uaddr;
293         ++to;
294     }
295     assert(to <= dev->mem->nregions + 1);
296     dev->mem->nregions = to;
297 }
298 
299 static uint64_t vhost_get_log_size(struct vhost_dev *dev)
300 {
301     uint64_t log_size = 0;
302     int i;
303     for (i = 0; i < dev->mem->nregions; ++i) {
304         struct vhost_memory_region *reg = dev->mem->regions + i;
305         uint64_t last = range_get_last(reg->guest_phys_addr,
306                                        reg->memory_size);
307         log_size = MAX(log_size, last / VHOST_LOG_CHUNK + 1);
308     }
309     for (i = 0; i < dev->nvqs; ++i) {
310         struct vhost_virtqueue *vq = dev->vqs + i;
311         uint64_t last = vq->used_phys + vq->used_size - 1;
312         log_size = MAX(log_size, last / VHOST_LOG_CHUNK + 1);
313     }
314     return log_size;
315 }
316 
317 static struct vhost_log *vhost_log_alloc(uint64_t size, bool share)
318 {
319     struct vhost_log *log;
320     uint64_t logsize = size * sizeof(*(log->log));
321     int fd = -1;
322 
323     log = g_new0(struct vhost_log, 1);
324     if (share) {
325         log->log = qemu_memfd_alloc("vhost-log", logsize,
326                                     F_SEAL_GROW | F_SEAL_SHRINK | F_SEAL_SEAL,
327                                     &fd);
328         memset(log->log, 0, logsize);
329     } else {
330         log->log = g_malloc0(logsize);
331     }
332 
333     log->size = size;
334     log->refcnt = 1;
335     log->fd = fd;
336 
337     return log;
338 }
339 
340 static struct vhost_log *vhost_log_get(uint64_t size, bool share)
341 {
342     struct vhost_log *log = share ? vhost_log_shm : vhost_log;
343 
344     if (!log || log->size != size) {
345         log = vhost_log_alloc(size, share);
346         if (share) {
347             vhost_log_shm = log;
348         } else {
349             vhost_log = log;
350         }
351     } else {
352         ++log->refcnt;
353     }
354 
355     return log;
356 }
357 
358 static void vhost_log_put(struct vhost_dev *dev, bool sync)
359 {
360     struct vhost_log *log = dev->log;
361 
362     if (!log) {
363         return;
364     }
365     dev->log = NULL;
366     dev->log_size = 0;
367 
368     --log->refcnt;
369     if (log->refcnt == 0) {
370         /* Sync only the range covered by the old log */
371         if (dev->log_size && sync) {
372             vhost_log_sync_range(dev, 0, dev->log_size * VHOST_LOG_CHUNK - 1);
373         }
374 
375         if (vhost_log == log) {
376             g_free(log->log);
377             vhost_log = NULL;
378         } else if (vhost_log_shm == log) {
379             qemu_memfd_free(log->log, log->size * sizeof(*(log->log)),
380                             log->fd);
381             vhost_log_shm = NULL;
382         }
383 
384         g_free(log);
385     }
386 }
387 
388 static bool vhost_dev_log_is_shared(struct vhost_dev *dev)
389 {
390     return dev->vhost_ops->vhost_requires_shm_log &&
391            dev->vhost_ops->vhost_requires_shm_log(dev);
392 }
393 
394 static inline void vhost_dev_log_resize(struct vhost_dev *dev, uint64_t size)
395 {
396     struct vhost_log *log = vhost_log_get(size, vhost_dev_log_is_shared(dev));
397     uint64_t log_base = (uintptr_t)log->log;
398     int r;
399 
400     /* inform backend of log switching, this must be done before
401        releasing the current log, to ensure no logging is lost */
402     r = dev->vhost_ops->vhost_set_log_base(dev, log_base, log);
403     assert(r >= 0);
404     vhost_log_put(dev, true);
405     dev->log = log;
406     dev->log_size = size;
407 }
408 
409 static int vhost_verify_ring_mappings(struct vhost_dev *dev,
410                                       uint64_t start_addr,
411                                       uint64_t size)
412 {
413     int i;
414     int r = 0;
415 
416     for (i = 0; !r && i < dev->nvqs; ++i) {
417         struct vhost_virtqueue *vq = dev->vqs + i;
418         hwaddr l;
419         void *p;
420 
421         if (!ranges_overlap(start_addr, size, vq->ring_phys, vq->ring_size)) {
422             continue;
423         }
424         l = vq->ring_size;
425         p = cpu_physical_memory_map(vq->ring_phys, &l, 1);
426         if (!p || l != vq->ring_size) {
427             fprintf(stderr, "Unable to map ring buffer for ring %d\n", i);
428             r = -ENOMEM;
429         }
430         if (p != vq->ring) {
431             fprintf(stderr, "Ring buffer relocated for ring %d\n", i);
432             r = -EBUSY;
433         }
434         cpu_physical_memory_unmap(p, l, 0, 0);
435     }
436     return r;
437 }
438 
439 static struct vhost_memory_region *vhost_dev_find_reg(struct vhost_dev *dev,
440 						      uint64_t start_addr,
441 						      uint64_t size)
442 {
443     int i, n = dev->mem->nregions;
444     for (i = 0; i < n; ++i) {
445         struct vhost_memory_region *reg = dev->mem->regions + i;
446         if (ranges_overlap(reg->guest_phys_addr, reg->memory_size,
447                            start_addr, size)) {
448             return reg;
449         }
450     }
451     return NULL;
452 }
453 
454 static bool vhost_dev_cmp_memory(struct vhost_dev *dev,
455                                  uint64_t start_addr,
456                                  uint64_t size,
457                                  uint64_t uaddr)
458 {
459     struct vhost_memory_region *reg = vhost_dev_find_reg(dev, start_addr, size);
460     uint64_t reglast;
461     uint64_t memlast;
462 
463     if (!reg) {
464         return true;
465     }
466 
467     reglast = range_get_last(reg->guest_phys_addr, reg->memory_size);
468     memlast = range_get_last(start_addr, size);
469 
470     /* Need to extend region? */
471     if (start_addr < reg->guest_phys_addr || memlast > reglast) {
472         return true;
473     }
474     /* userspace_addr changed? */
475     return uaddr != reg->userspace_addr + start_addr - reg->guest_phys_addr;
476 }
477 
478 static void vhost_set_memory(MemoryListener *listener,
479                              MemoryRegionSection *section,
480                              bool add)
481 {
482     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
483                                          memory_listener);
484     hwaddr start_addr = section->offset_within_address_space;
485     ram_addr_t size = int128_get64(section->size);
486     bool log_dirty =
487         memory_region_get_dirty_log_mask(section->mr) & ~(1 << DIRTY_MEMORY_MIGRATION);
488     int s = offsetof(struct vhost_memory, regions) +
489         (dev->mem->nregions + 1) * sizeof dev->mem->regions[0];
490     void *ram;
491 
492     dev->mem = g_realloc(dev->mem, s);
493 
494     if (log_dirty) {
495         add = false;
496     }
497 
498     assert(size);
499 
500     /* Optimize no-change case. At least cirrus_vga does this a lot at this time. */
501     ram = memory_region_get_ram_ptr(section->mr) + section->offset_within_region;
502     if (add) {
503         if (!vhost_dev_cmp_memory(dev, start_addr, size, (uintptr_t)ram)) {
504             /* Region exists with same address. Nothing to do. */
505             return;
506         }
507     } else {
508         if (!vhost_dev_find_reg(dev, start_addr, size)) {
509             /* Removing region that we don't access. Nothing to do. */
510             return;
511         }
512     }
513 
514     vhost_dev_unassign_memory(dev, start_addr, size);
515     if (add) {
516         /* Add given mapping, merging adjacent regions if any */
517         vhost_dev_assign_memory(dev, start_addr, size, (uintptr_t)ram);
518     } else {
519         /* Remove old mapping for this memory, if any. */
520         vhost_dev_unassign_memory(dev, start_addr, size);
521     }
522     dev->mem_changed_start_addr = MIN(dev->mem_changed_start_addr, start_addr);
523     dev->mem_changed_end_addr = MAX(dev->mem_changed_end_addr, start_addr + size - 1);
524     dev->memory_changed = true;
525     used_memslots = dev->mem->nregions;
526 }
527 
528 static bool vhost_section(MemoryRegionSection *section)
529 {
530     return memory_region_is_ram(section->mr);
531 }
532 
533 static void vhost_begin(MemoryListener *listener)
534 {
535     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
536                                          memory_listener);
537     dev->mem_changed_end_addr = 0;
538     dev->mem_changed_start_addr = -1;
539 }
540 
541 static void vhost_commit(MemoryListener *listener)
542 {
543     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
544                                          memory_listener);
545     hwaddr start_addr = 0;
546     ram_addr_t size = 0;
547     uint64_t log_size;
548     int r;
549 
550     if (!dev->memory_changed) {
551         return;
552     }
553     if (!dev->started) {
554         return;
555     }
556     if (dev->mem_changed_start_addr > dev->mem_changed_end_addr) {
557         return;
558     }
559 
560     if (dev->started) {
561         start_addr = dev->mem_changed_start_addr;
562         size = dev->mem_changed_end_addr - dev->mem_changed_start_addr + 1;
563 
564         r = vhost_verify_ring_mappings(dev, start_addr, size);
565         assert(r >= 0);
566     }
567 
568     if (!dev->log_enabled) {
569         r = dev->vhost_ops->vhost_set_mem_table(dev, dev->mem);
570         assert(r >= 0);
571         dev->memory_changed = false;
572         return;
573     }
574     log_size = vhost_get_log_size(dev);
575     /* We allocate an extra 4K bytes to log,
576      * to reduce the * number of reallocations. */
577 #define VHOST_LOG_BUFFER (0x1000 / sizeof *dev->log)
578     /* To log more, must increase log size before table update. */
579     if (dev->log_size < log_size) {
580         vhost_dev_log_resize(dev, log_size + VHOST_LOG_BUFFER);
581     }
582     r = dev->vhost_ops->vhost_set_mem_table(dev, dev->mem);
583     assert(r >= 0);
584     /* To log less, can only decrease log size after table update. */
585     if (dev->log_size > log_size + VHOST_LOG_BUFFER) {
586         vhost_dev_log_resize(dev, log_size);
587     }
588     dev->memory_changed = false;
589 }
590 
591 static void vhost_region_add(MemoryListener *listener,
592                              MemoryRegionSection *section)
593 {
594     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
595                                          memory_listener);
596 
597     if (!vhost_section(section)) {
598         return;
599     }
600 
601     ++dev->n_mem_sections;
602     dev->mem_sections = g_renew(MemoryRegionSection, dev->mem_sections,
603                                 dev->n_mem_sections);
604     dev->mem_sections[dev->n_mem_sections - 1] = *section;
605     memory_region_ref(section->mr);
606     vhost_set_memory(listener, section, true);
607 }
608 
609 static void vhost_region_del(MemoryListener *listener,
610                              MemoryRegionSection *section)
611 {
612     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
613                                          memory_listener);
614     int i;
615 
616     if (!vhost_section(section)) {
617         return;
618     }
619 
620     vhost_set_memory(listener, section, false);
621     memory_region_unref(section->mr);
622     for (i = 0; i < dev->n_mem_sections; ++i) {
623         if (dev->mem_sections[i].offset_within_address_space
624             == section->offset_within_address_space) {
625             --dev->n_mem_sections;
626             memmove(&dev->mem_sections[i], &dev->mem_sections[i+1],
627                     (dev->n_mem_sections - i) * sizeof(*dev->mem_sections));
628             break;
629         }
630     }
631 }
632 
633 static void vhost_region_nop(MemoryListener *listener,
634                              MemoryRegionSection *section)
635 {
636 }
637 
638 static int vhost_virtqueue_set_addr(struct vhost_dev *dev,
639                                     struct vhost_virtqueue *vq,
640                                     unsigned idx, bool enable_log)
641 {
642     struct vhost_vring_addr addr = {
643         .index = idx,
644         .desc_user_addr = (uint64_t)(unsigned long)vq->desc,
645         .avail_user_addr = (uint64_t)(unsigned long)vq->avail,
646         .used_user_addr = (uint64_t)(unsigned long)vq->used,
647         .log_guest_addr = vq->used_phys,
648         .flags = enable_log ? (1 << VHOST_VRING_F_LOG) : 0,
649     };
650     int r = dev->vhost_ops->vhost_set_vring_addr(dev, &addr);
651     if (r < 0) {
652         return -errno;
653     }
654     return 0;
655 }
656 
657 static int vhost_dev_set_features(struct vhost_dev *dev, bool enable_log)
658 {
659     uint64_t features = dev->acked_features;
660     int r;
661     if (enable_log) {
662         features |= 0x1ULL << VHOST_F_LOG_ALL;
663     }
664     r = dev->vhost_ops->vhost_set_features(dev, features);
665     return r < 0 ? -errno : 0;
666 }
667 
668 static int vhost_dev_set_log(struct vhost_dev *dev, bool enable_log)
669 {
670     int r, t, i, idx;
671     r = vhost_dev_set_features(dev, enable_log);
672     if (r < 0) {
673         goto err_features;
674     }
675     for (i = 0; i < dev->nvqs; ++i) {
676         idx = dev->vhost_ops->vhost_get_vq_index(dev, dev->vq_index + i);
677         r = vhost_virtqueue_set_addr(dev, dev->vqs + i, idx,
678                                      enable_log);
679         if (r < 0) {
680             goto err_vq;
681         }
682     }
683     return 0;
684 err_vq:
685     for (; i >= 0; --i) {
686         idx = dev->vhost_ops->vhost_get_vq_index(dev, dev->vq_index + i);
687         t = vhost_virtqueue_set_addr(dev, dev->vqs + i, idx,
688                                      dev->log_enabled);
689         assert(t >= 0);
690     }
691     t = vhost_dev_set_features(dev, dev->log_enabled);
692     assert(t >= 0);
693 err_features:
694     return r;
695 }
696 
697 static int vhost_migration_log(MemoryListener *listener, int enable)
698 {
699     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
700                                          memory_listener);
701     int r;
702     if (!!enable == dev->log_enabled) {
703         return 0;
704     }
705     if (!dev->started) {
706         dev->log_enabled = enable;
707         return 0;
708     }
709     if (!enable) {
710         r = vhost_dev_set_log(dev, false);
711         if (r < 0) {
712             return r;
713         }
714         vhost_log_put(dev, false);
715     } else {
716         vhost_dev_log_resize(dev, vhost_get_log_size(dev));
717         r = vhost_dev_set_log(dev, true);
718         if (r < 0) {
719             return r;
720         }
721     }
722     dev->log_enabled = enable;
723     return 0;
724 }
725 
726 static void vhost_log_global_start(MemoryListener *listener)
727 {
728     int r;
729 
730     r = vhost_migration_log(listener, true);
731     if (r < 0) {
732         abort();
733     }
734 }
735 
736 static void vhost_log_global_stop(MemoryListener *listener)
737 {
738     int r;
739 
740     r = vhost_migration_log(listener, false);
741     if (r < 0) {
742         abort();
743     }
744 }
745 
746 static void vhost_log_start(MemoryListener *listener,
747                             MemoryRegionSection *section,
748                             int old, int new)
749 {
750     /* FIXME: implement */
751 }
752 
753 static void vhost_log_stop(MemoryListener *listener,
754                            MemoryRegionSection *section,
755                            int old, int new)
756 {
757     /* FIXME: implement */
758 }
759 
760 /* The vhost driver natively knows how to handle the vrings of non
761  * cross-endian legacy devices and modern devices. Only legacy devices
762  * exposed to a bi-endian guest may require the vhost driver to use a
763  * specific endianness.
764  */
765 static inline bool vhost_needs_vring_endian(VirtIODevice *vdev)
766 {
767     if (virtio_vdev_has_feature(vdev, VIRTIO_F_VERSION_1)) {
768         return false;
769     }
770 #ifdef HOST_WORDS_BIGENDIAN
771     return vdev->device_endian == VIRTIO_DEVICE_ENDIAN_LITTLE;
772 #else
773     return vdev->device_endian == VIRTIO_DEVICE_ENDIAN_BIG;
774 #endif
775 }
776 
777 static int vhost_virtqueue_set_vring_endian_legacy(struct vhost_dev *dev,
778                                                    bool is_big_endian,
779                                                    int vhost_vq_index)
780 {
781     struct vhost_vring_state s = {
782         .index = vhost_vq_index,
783         .num = is_big_endian
784     };
785 
786     if (!dev->vhost_ops->vhost_set_vring_endian(dev, &s)) {
787         return 0;
788     }
789 
790     if (errno == ENOTTY) {
791         error_report("vhost does not support cross-endian");
792         return -ENOSYS;
793     }
794 
795     return -errno;
796 }
797 
798 static int vhost_virtqueue_start(struct vhost_dev *dev,
799                                 struct VirtIODevice *vdev,
800                                 struct vhost_virtqueue *vq,
801                                 unsigned idx)
802 {
803     hwaddr s, l, a;
804     int r;
805     int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, idx);
806     struct vhost_vring_file file = {
807         .index = vhost_vq_index
808     };
809     struct vhost_vring_state state = {
810         .index = vhost_vq_index
811     };
812     struct VirtQueue *vvq = virtio_get_queue(vdev, idx);
813 
814 
815     vq->num = state.num = virtio_queue_get_num(vdev, idx);
816     r = dev->vhost_ops->vhost_set_vring_num(dev, &state);
817     if (r) {
818         return -errno;
819     }
820 
821     state.num = virtio_queue_get_last_avail_idx(vdev, idx);
822     r = dev->vhost_ops->vhost_set_vring_base(dev, &state);
823     if (r) {
824         return -errno;
825     }
826 
827     if (vhost_needs_vring_endian(vdev)) {
828         r = vhost_virtqueue_set_vring_endian_legacy(dev,
829                                                     virtio_is_big_endian(vdev),
830                                                     vhost_vq_index);
831         if (r) {
832             return -errno;
833         }
834     }
835 
836     s = l = virtio_queue_get_desc_size(vdev, idx);
837     a = virtio_queue_get_desc_addr(vdev, idx);
838     vq->desc = cpu_physical_memory_map(a, &l, 0);
839     if (!vq->desc || l != s) {
840         r = -ENOMEM;
841         goto fail_alloc_desc;
842     }
843     s = l = virtio_queue_get_avail_size(vdev, idx);
844     a = virtio_queue_get_avail_addr(vdev, idx);
845     vq->avail = cpu_physical_memory_map(a, &l, 0);
846     if (!vq->avail || l != s) {
847         r = -ENOMEM;
848         goto fail_alloc_avail;
849     }
850     vq->used_size = s = l = virtio_queue_get_used_size(vdev, idx);
851     vq->used_phys = a = virtio_queue_get_used_addr(vdev, idx);
852     vq->used = cpu_physical_memory_map(a, &l, 1);
853     if (!vq->used || l != s) {
854         r = -ENOMEM;
855         goto fail_alloc_used;
856     }
857 
858     vq->ring_size = s = l = virtio_queue_get_ring_size(vdev, idx);
859     vq->ring_phys = a = virtio_queue_get_ring_addr(vdev, idx);
860     vq->ring = cpu_physical_memory_map(a, &l, 1);
861     if (!vq->ring || l != s) {
862         r = -ENOMEM;
863         goto fail_alloc_ring;
864     }
865 
866     r = vhost_virtqueue_set_addr(dev, vq, vhost_vq_index, dev->log_enabled);
867     if (r < 0) {
868         r = -errno;
869         goto fail_alloc;
870     }
871 
872     file.fd = event_notifier_get_fd(virtio_queue_get_host_notifier(vvq));
873     r = dev->vhost_ops->vhost_set_vring_kick(dev, &file);
874     if (r) {
875         r = -errno;
876         goto fail_kick;
877     }
878 
879     /* Clear and discard previous events if any. */
880     event_notifier_test_and_clear(&vq->masked_notifier);
881 
882     /* Init vring in unmasked state, unless guest_notifier_mask
883      * will do it later.
884      */
885     if (!vdev->use_guest_notifier_mask) {
886         /* TODO: check and handle errors. */
887         vhost_virtqueue_mask(dev, vdev, idx, false);
888     }
889 
890     return 0;
891 
892 fail_kick:
893 fail_alloc:
894     cpu_physical_memory_unmap(vq->ring, virtio_queue_get_ring_size(vdev, idx),
895                               0, 0);
896 fail_alloc_ring:
897     cpu_physical_memory_unmap(vq->used, virtio_queue_get_used_size(vdev, idx),
898                               0, 0);
899 fail_alloc_used:
900     cpu_physical_memory_unmap(vq->avail, virtio_queue_get_avail_size(vdev, idx),
901                               0, 0);
902 fail_alloc_avail:
903     cpu_physical_memory_unmap(vq->desc, virtio_queue_get_desc_size(vdev, idx),
904                               0, 0);
905 fail_alloc_desc:
906     return r;
907 }
908 
909 static void vhost_virtqueue_stop(struct vhost_dev *dev,
910                                     struct VirtIODevice *vdev,
911                                     struct vhost_virtqueue *vq,
912                                     unsigned idx)
913 {
914     int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, idx);
915     struct vhost_vring_state state = {
916         .index = vhost_vq_index,
917     };
918     int r;
919 
920     r = dev->vhost_ops->vhost_get_vring_base(dev, &state);
921     if (r < 0) {
922         fprintf(stderr, "vhost VQ %d ring restore failed: %d\n", idx, r);
923         fflush(stderr);
924     }
925     virtio_queue_set_last_avail_idx(vdev, idx, state.num);
926     virtio_queue_invalidate_signalled_used(vdev, idx);
927 
928     /* In the cross-endian case, we need to reset the vring endianness to
929      * native as legacy devices expect so by default.
930      */
931     if (vhost_needs_vring_endian(vdev)) {
932         r = vhost_virtqueue_set_vring_endian_legacy(dev,
933                                                     !virtio_is_big_endian(vdev),
934                                                     vhost_vq_index);
935         if (r < 0) {
936             error_report("failed to reset vring endianness");
937         }
938     }
939 
940     assert (r >= 0);
941     cpu_physical_memory_unmap(vq->ring, virtio_queue_get_ring_size(vdev, idx),
942                               0, virtio_queue_get_ring_size(vdev, idx));
943     cpu_physical_memory_unmap(vq->used, virtio_queue_get_used_size(vdev, idx),
944                               1, virtio_queue_get_used_size(vdev, idx));
945     cpu_physical_memory_unmap(vq->avail, virtio_queue_get_avail_size(vdev, idx),
946                               0, virtio_queue_get_avail_size(vdev, idx));
947     cpu_physical_memory_unmap(vq->desc, virtio_queue_get_desc_size(vdev, idx),
948                               0, virtio_queue_get_desc_size(vdev, idx));
949 }
950 
951 static void vhost_eventfd_add(MemoryListener *listener,
952                               MemoryRegionSection *section,
953                               bool match_data, uint64_t data, EventNotifier *e)
954 {
955 }
956 
957 static void vhost_eventfd_del(MemoryListener *listener,
958                               MemoryRegionSection *section,
959                               bool match_data, uint64_t data, EventNotifier *e)
960 {
961 }
962 
963 static int vhost_virtqueue_set_busyloop_timeout(struct vhost_dev *dev,
964                                                 int n, uint32_t timeout)
965 {
966     int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, n);
967     struct vhost_vring_state state = {
968         .index = vhost_vq_index,
969         .num = timeout,
970     };
971     int r;
972 
973     if (!dev->vhost_ops->vhost_set_vring_busyloop_timeout) {
974         return -EINVAL;
975     }
976 
977     r = dev->vhost_ops->vhost_set_vring_busyloop_timeout(dev, &state);
978     if (r) {
979         return r;
980     }
981 
982     return 0;
983 }
984 
985 static int vhost_virtqueue_init(struct vhost_dev *dev,
986                                 struct vhost_virtqueue *vq, int n)
987 {
988     int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, n);
989     struct vhost_vring_file file = {
990         .index = vhost_vq_index,
991     };
992     int r = event_notifier_init(&vq->masked_notifier, 0);
993     if (r < 0) {
994         return r;
995     }
996 
997     file.fd = event_notifier_get_fd(&vq->masked_notifier);
998     r = dev->vhost_ops->vhost_set_vring_call(dev, &file);
999     if (r) {
1000         r = -errno;
1001         goto fail_call;
1002     }
1003     return 0;
1004 fail_call:
1005     event_notifier_cleanup(&vq->masked_notifier);
1006     return r;
1007 }
1008 
1009 static void vhost_virtqueue_cleanup(struct vhost_virtqueue *vq)
1010 {
1011     event_notifier_cleanup(&vq->masked_notifier);
1012 }
1013 
1014 int vhost_dev_init(struct vhost_dev *hdev, void *opaque,
1015                    VhostBackendType backend_type, uint32_t busyloop_timeout)
1016 {
1017     uint64_t features;
1018     int i, r;
1019 
1020     hdev->migration_blocker = NULL;
1021 
1022     r = vhost_set_backend_type(hdev, backend_type);
1023     assert(r >= 0);
1024 
1025     r = hdev->vhost_ops->vhost_backend_init(hdev, opaque);
1026     if (r < 0) {
1027         goto fail;
1028     }
1029 
1030     if (used_memslots > hdev->vhost_ops->vhost_backend_memslots_limit(hdev)) {
1031         fprintf(stderr, "vhost backend memory slots limit is less"
1032                 " than current number of present memory slots\n");
1033         r = -1;
1034         goto fail;
1035     }
1036 
1037     r = hdev->vhost_ops->vhost_set_owner(hdev);
1038     if (r < 0) {
1039         goto fail;
1040     }
1041 
1042     r = hdev->vhost_ops->vhost_get_features(hdev, &features);
1043     if (r < 0) {
1044         goto fail;
1045     }
1046 
1047     for (i = 0; i < hdev->nvqs; ++i) {
1048         r = vhost_virtqueue_init(hdev, hdev->vqs + i, hdev->vq_index + i);
1049         if (r < 0) {
1050             goto fail_vq;
1051         }
1052     }
1053 
1054     if (busyloop_timeout) {
1055         for (i = 0; i < hdev->nvqs; ++i) {
1056             r = vhost_virtqueue_set_busyloop_timeout(hdev, hdev->vq_index + i,
1057                                                      busyloop_timeout);
1058             if (r < 0) {
1059                 goto fail_busyloop;
1060             }
1061         }
1062     }
1063 
1064     hdev->features = features;
1065 
1066     hdev->memory_listener = (MemoryListener) {
1067         .begin = vhost_begin,
1068         .commit = vhost_commit,
1069         .region_add = vhost_region_add,
1070         .region_del = vhost_region_del,
1071         .region_nop = vhost_region_nop,
1072         .log_start = vhost_log_start,
1073         .log_stop = vhost_log_stop,
1074         .log_sync = vhost_log_sync,
1075         .log_global_start = vhost_log_global_start,
1076         .log_global_stop = vhost_log_global_stop,
1077         .eventfd_add = vhost_eventfd_add,
1078         .eventfd_del = vhost_eventfd_del,
1079         .priority = 10
1080     };
1081 
1082     if (hdev->migration_blocker == NULL) {
1083         if (!(hdev->features & (0x1ULL << VHOST_F_LOG_ALL))) {
1084             error_setg(&hdev->migration_blocker,
1085                        "Migration disabled: vhost lacks VHOST_F_LOG_ALL feature.");
1086         } else if (!qemu_memfd_check()) {
1087             error_setg(&hdev->migration_blocker,
1088                        "Migration disabled: failed to allocate shared memory");
1089         }
1090     }
1091 
1092     if (hdev->migration_blocker != NULL) {
1093         migrate_add_blocker(hdev->migration_blocker);
1094     }
1095 
1096     hdev->mem = g_malloc0(offsetof(struct vhost_memory, regions));
1097     hdev->n_mem_sections = 0;
1098     hdev->mem_sections = NULL;
1099     hdev->log = NULL;
1100     hdev->log_size = 0;
1101     hdev->log_enabled = false;
1102     hdev->started = false;
1103     hdev->memory_changed = false;
1104     memory_listener_register(&hdev->memory_listener, &address_space_memory);
1105     QLIST_INSERT_HEAD(&vhost_devices, hdev, entry);
1106     return 0;
1107 fail_busyloop:
1108     while (--i >= 0) {
1109         vhost_virtqueue_set_busyloop_timeout(hdev, hdev->vq_index + i, 0);
1110     }
1111     i = hdev->nvqs;
1112 fail_vq:
1113     while (--i >= 0) {
1114         vhost_virtqueue_cleanup(hdev->vqs + i);
1115     }
1116 fail:
1117     r = -errno;
1118     hdev->vhost_ops->vhost_backend_cleanup(hdev);
1119     QLIST_REMOVE(hdev, entry);
1120     return r;
1121 }
1122 
1123 void vhost_dev_cleanup(struct vhost_dev *hdev)
1124 {
1125     int i;
1126     for (i = 0; i < hdev->nvqs; ++i) {
1127         vhost_virtqueue_cleanup(hdev->vqs + i);
1128     }
1129     if (hdev->mem) {
1130         /* those are only safe after successful init */
1131         memory_listener_unregister(&hdev->memory_listener);
1132         QLIST_REMOVE(hdev, entry);
1133     }
1134     if (hdev->migration_blocker) {
1135         migrate_del_blocker(hdev->migration_blocker);
1136         error_free(hdev->migration_blocker);
1137     }
1138     g_free(hdev->mem);
1139     g_free(hdev->mem_sections);
1140     hdev->vhost_ops->vhost_backend_cleanup(hdev);
1141     assert(!hdev->log);
1142 }
1143 
1144 /* Stop processing guest IO notifications in qemu.
1145  * Start processing them in vhost in kernel.
1146  */
1147 int vhost_dev_enable_notifiers(struct vhost_dev *hdev, VirtIODevice *vdev)
1148 {
1149     BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev)));
1150     VirtioBusState *vbus = VIRTIO_BUS(qbus);
1151     VirtioBusClass *k = VIRTIO_BUS_GET_CLASS(vbus);
1152     int i, r, e;
1153     if (!k->ioeventfd_started) {
1154         fprintf(stderr, "binding does not support host notifiers\n");
1155         r = -ENOSYS;
1156         goto fail;
1157     }
1158 
1159     for (i = 0; i < hdev->nvqs; ++i) {
1160         r = virtio_bus_set_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i,
1161                                          true);
1162         if (r < 0) {
1163             fprintf(stderr, "vhost VQ %d notifier binding failed: %d\n", i, -r);
1164             goto fail_vq;
1165         }
1166     }
1167 
1168     return 0;
1169 fail_vq:
1170     while (--i >= 0) {
1171         e = virtio_bus_set_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i,
1172                                          false);
1173         if (e < 0) {
1174             fprintf(stderr, "vhost VQ %d notifier cleanup error: %d\n", i, -r);
1175             fflush(stderr);
1176         }
1177         assert (e >= 0);
1178     }
1179 fail:
1180     return r;
1181 }
1182 
1183 /* Stop processing guest IO notifications in vhost.
1184  * Start processing them in qemu.
1185  * This might actually run the qemu handlers right away,
1186  * so virtio in qemu must be completely setup when this is called.
1187  */
1188 void vhost_dev_disable_notifiers(struct vhost_dev *hdev, VirtIODevice *vdev)
1189 {
1190     BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev)));
1191     int i, r;
1192 
1193     for (i = 0; i < hdev->nvqs; ++i) {
1194         r = virtio_bus_set_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i,
1195                                          false);
1196         if (r < 0) {
1197             fprintf(stderr, "vhost VQ %d notifier cleanup failed: %d\n", i, -r);
1198             fflush(stderr);
1199         }
1200         assert (r >= 0);
1201     }
1202 }
1203 
1204 /* Test and clear event pending status.
1205  * Should be called after unmask to avoid losing events.
1206  */
1207 bool vhost_virtqueue_pending(struct vhost_dev *hdev, int n)
1208 {
1209     struct vhost_virtqueue *vq = hdev->vqs + n - hdev->vq_index;
1210     assert(n >= hdev->vq_index && n < hdev->vq_index + hdev->nvqs);
1211     return event_notifier_test_and_clear(&vq->masked_notifier);
1212 }
1213 
1214 /* Mask/unmask events from this vq. */
1215 void vhost_virtqueue_mask(struct vhost_dev *hdev, VirtIODevice *vdev, int n,
1216                          bool mask)
1217 {
1218     struct VirtQueue *vvq = virtio_get_queue(vdev, n);
1219     int r, index = n - hdev->vq_index;
1220     struct vhost_vring_file file;
1221 
1222     if (mask) {
1223         assert(vdev->use_guest_notifier_mask);
1224         file.fd = event_notifier_get_fd(&hdev->vqs[index].masked_notifier);
1225     } else {
1226         file.fd = event_notifier_get_fd(virtio_queue_get_guest_notifier(vvq));
1227     }
1228 
1229     file.index = hdev->vhost_ops->vhost_get_vq_index(hdev, n);
1230     r = hdev->vhost_ops->vhost_set_vring_call(hdev, &file);
1231     assert(r >= 0);
1232 }
1233 
1234 uint64_t vhost_get_features(struct vhost_dev *hdev, const int *feature_bits,
1235                             uint64_t features)
1236 {
1237     const int *bit = feature_bits;
1238     while (*bit != VHOST_INVALID_FEATURE_BIT) {
1239         uint64_t bit_mask = (1ULL << *bit);
1240         if (!(hdev->features & bit_mask)) {
1241             features &= ~bit_mask;
1242         }
1243         bit++;
1244     }
1245     return features;
1246 }
1247 
1248 void vhost_ack_features(struct vhost_dev *hdev, const int *feature_bits,
1249                         uint64_t features)
1250 {
1251     const int *bit = feature_bits;
1252     while (*bit != VHOST_INVALID_FEATURE_BIT) {
1253         uint64_t bit_mask = (1ULL << *bit);
1254         if (features & bit_mask) {
1255             hdev->acked_features |= bit_mask;
1256         }
1257         bit++;
1258     }
1259 }
1260 
1261 /* Host notifiers must be enabled at this point. */
1262 int vhost_dev_start(struct vhost_dev *hdev, VirtIODevice *vdev)
1263 {
1264     int i, r;
1265 
1266     hdev->started = true;
1267 
1268     r = vhost_dev_set_features(hdev, hdev->log_enabled);
1269     if (r < 0) {
1270         goto fail_features;
1271     }
1272     r = hdev->vhost_ops->vhost_set_mem_table(hdev, hdev->mem);
1273     if (r < 0) {
1274         r = -errno;
1275         goto fail_mem;
1276     }
1277     for (i = 0; i < hdev->nvqs; ++i) {
1278         r = vhost_virtqueue_start(hdev,
1279                                   vdev,
1280                                   hdev->vqs + i,
1281                                   hdev->vq_index + i);
1282         if (r < 0) {
1283             goto fail_vq;
1284         }
1285     }
1286 
1287     if (hdev->log_enabled) {
1288         uint64_t log_base;
1289 
1290         hdev->log_size = vhost_get_log_size(hdev);
1291         hdev->log = vhost_log_get(hdev->log_size,
1292                                   vhost_dev_log_is_shared(hdev));
1293         log_base = (uintptr_t)hdev->log->log;
1294         r = hdev->vhost_ops->vhost_set_log_base(hdev,
1295                                                 hdev->log_size ? log_base : 0,
1296                                                 hdev->log);
1297         if (r < 0) {
1298             r = -errno;
1299             goto fail_log;
1300         }
1301     }
1302 
1303     return 0;
1304 fail_log:
1305     vhost_log_put(hdev, false);
1306 fail_vq:
1307     while (--i >= 0) {
1308         vhost_virtqueue_stop(hdev,
1309                              vdev,
1310                              hdev->vqs + i,
1311                              hdev->vq_index + i);
1312     }
1313     i = hdev->nvqs;
1314 fail_mem:
1315 fail_features:
1316 
1317     hdev->started = false;
1318     return r;
1319 }
1320 
1321 /* Host notifiers must be enabled at this point. */
1322 void vhost_dev_stop(struct vhost_dev *hdev, VirtIODevice *vdev)
1323 {
1324     int i;
1325 
1326     for (i = 0; i < hdev->nvqs; ++i) {
1327         vhost_virtqueue_stop(hdev,
1328                              vdev,
1329                              hdev->vqs + i,
1330                              hdev->vq_index + i);
1331     }
1332 
1333     vhost_log_put(hdev, true);
1334     hdev->started = false;
1335 }
1336