xref: /openbmc/qemu/hw/virtio/vhost.c (revision 91bfcdb0)
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 "hw/virtio/vhost.h"
17 #include "hw/hw.h"
18 #include "qemu/atomic.h"
19 #include "qemu/range.h"
20 #include "qemu/error-report.h"
21 #include "qemu/memfd.h"
22 #include <linux/vhost.h>
23 #include "exec/address-spaces.h"
24 #include "hw/virtio/virtio-bus.h"
25 #include "hw/virtio/virtio-access.h"
26 #include "migration/migration.h"
27 
28 static struct vhost_log *vhost_log;
29 static struct vhost_log *vhost_log_shm;
30 
31 static unsigned int used_memslots;
32 static QLIST_HEAD(, vhost_dev) vhost_devices =
33     QLIST_HEAD_INITIALIZER(vhost_devices);
34 
35 bool vhost_has_free_slot(void)
36 {
37     unsigned int slots_limit = ~0U;
38     struct vhost_dev *hdev;
39 
40     QLIST_FOREACH(hdev, &vhost_devices, entry) {
41         unsigned int r = hdev->vhost_ops->vhost_backend_memslots_limit(hdev);
42         slots_limit = MIN(slots_limit, r);
43     }
44     return slots_limit > used_memslots;
45 }
46 
47 static void vhost_dev_sync_region(struct vhost_dev *dev,
48                                   MemoryRegionSection *section,
49                                   uint64_t mfirst, uint64_t mlast,
50                                   uint64_t rfirst, uint64_t rlast)
51 {
52     vhost_log_chunk_t *log = dev->log->log;
53 
54     uint64_t start = MAX(mfirst, rfirst);
55     uint64_t end = MIN(mlast, rlast);
56     vhost_log_chunk_t *from = log + start / VHOST_LOG_CHUNK;
57     vhost_log_chunk_t *to = log + end / VHOST_LOG_CHUNK + 1;
58     uint64_t addr = (start / VHOST_LOG_CHUNK) * VHOST_LOG_CHUNK;
59 
60     if (end < start) {
61         return;
62     }
63     assert(end / VHOST_LOG_CHUNK < dev->log_size);
64     assert(start / VHOST_LOG_CHUNK < dev->log_size);
65 
66     for (;from < to; ++from) {
67         vhost_log_chunk_t log;
68         /* We first check with non-atomic: much cheaper,
69          * and we expect non-dirty to be the common case. */
70         if (!*from) {
71             addr += VHOST_LOG_CHUNK;
72             continue;
73         }
74         /* Data must be read atomically. We don't really need barrier semantics
75          * but it's easier to use atomic_* than roll our own. */
76         log = atomic_xchg(from, 0);
77         while (log) {
78             int bit = ctzl(log);
79             hwaddr page_addr;
80             hwaddr section_offset;
81             hwaddr mr_offset;
82             page_addr = addr + bit * VHOST_LOG_PAGE;
83             section_offset = page_addr - section->offset_within_address_space;
84             mr_offset = section_offset + section->offset_within_region;
85             memory_region_set_dirty(section->mr, mr_offset, VHOST_LOG_PAGE);
86             log &= ~(0x1ull << bit);
87         }
88         addr += VHOST_LOG_CHUNK;
89     }
90 }
91 
92 static int vhost_sync_dirty_bitmap(struct vhost_dev *dev,
93                                    MemoryRegionSection *section,
94                                    hwaddr first,
95                                    hwaddr last)
96 {
97     int i;
98     hwaddr start_addr;
99     hwaddr end_addr;
100 
101     if (!dev->log_enabled || !dev->started) {
102         return 0;
103     }
104     start_addr = section->offset_within_address_space;
105     end_addr = range_get_last(start_addr, int128_get64(section->size));
106     start_addr = MAX(first, start_addr);
107     end_addr = MIN(last, end_addr);
108 
109     for (i = 0; i < dev->mem->nregions; ++i) {
110         struct vhost_memory_region *reg = dev->mem->regions + i;
111         vhost_dev_sync_region(dev, section, start_addr, end_addr,
112                               reg->guest_phys_addr,
113                               range_get_last(reg->guest_phys_addr,
114                                              reg->memory_size));
115     }
116     for (i = 0; i < dev->nvqs; ++i) {
117         struct vhost_virtqueue *vq = dev->vqs + i;
118         vhost_dev_sync_region(dev, section, start_addr, end_addr, vq->used_phys,
119                               range_get_last(vq->used_phys, vq->used_size));
120     }
121     return 0;
122 }
123 
124 static void vhost_log_sync(MemoryListener *listener,
125                           MemoryRegionSection *section)
126 {
127     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
128                                          memory_listener);
129     vhost_sync_dirty_bitmap(dev, section, 0x0, ~0x0ULL);
130 }
131 
132 static void vhost_log_sync_range(struct vhost_dev *dev,
133                                  hwaddr first, hwaddr last)
134 {
135     int i;
136     /* FIXME: this is N^2 in number of sections */
137     for (i = 0; i < dev->n_mem_sections; ++i) {
138         MemoryRegionSection *section = &dev->mem_sections[i];
139         vhost_sync_dirty_bitmap(dev, section, first, last);
140     }
141 }
142 
143 /* Assign/unassign. Keep an unsorted array of non-overlapping
144  * memory regions in dev->mem. */
145 static void vhost_dev_unassign_memory(struct vhost_dev *dev,
146                                       uint64_t start_addr,
147                                       uint64_t size)
148 {
149     int from, to, n = dev->mem->nregions;
150     /* Track overlapping/split regions for sanity checking. */
151     int overlap_start = 0, overlap_end = 0, overlap_middle = 0, split = 0;
152 
153     for (from = 0, to = 0; from < n; ++from, ++to) {
154         struct vhost_memory_region *reg = dev->mem->regions + to;
155         uint64_t reglast;
156         uint64_t memlast;
157         uint64_t change;
158 
159         /* clone old region */
160         if (to != from) {
161             memcpy(reg, dev->mem->regions + from, sizeof *reg);
162         }
163 
164         /* No overlap is simple */
165         if (!ranges_overlap(reg->guest_phys_addr, reg->memory_size,
166                             start_addr, size)) {
167             continue;
168         }
169 
170         /* Split only happens if supplied region
171          * is in the middle of an existing one. Thus it can not
172          * overlap with any other existing region. */
173         assert(!split);
174 
175         reglast = range_get_last(reg->guest_phys_addr, reg->memory_size);
176         memlast = range_get_last(start_addr, size);
177 
178         /* Remove whole region */
179         if (start_addr <= reg->guest_phys_addr && memlast >= reglast) {
180             --dev->mem->nregions;
181             --to;
182             ++overlap_middle;
183             continue;
184         }
185 
186         /* Shrink region */
187         if (memlast >= reglast) {
188             reg->memory_size = start_addr - reg->guest_phys_addr;
189             assert(reg->memory_size);
190             assert(!overlap_end);
191             ++overlap_end;
192             continue;
193         }
194 
195         /* Shift region */
196         if (start_addr <= reg->guest_phys_addr) {
197             change = memlast + 1 - reg->guest_phys_addr;
198             reg->memory_size -= change;
199             reg->guest_phys_addr += change;
200             reg->userspace_addr += change;
201             assert(reg->memory_size);
202             assert(!overlap_start);
203             ++overlap_start;
204             continue;
205         }
206 
207         /* This only happens if supplied region
208          * is in the middle of an existing one. Thus it can not
209          * overlap with any other existing region. */
210         assert(!overlap_start);
211         assert(!overlap_end);
212         assert(!overlap_middle);
213         /* Split region: shrink first part, shift second part. */
214         memcpy(dev->mem->regions + n, reg, sizeof *reg);
215         reg->memory_size = start_addr - reg->guest_phys_addr;
216         assert(reg->memory_size);
217         change = memlast + 1 - reg->guest_phys_addr;
218         reg = dev->mem->regions + n;
219         reg->memory_size -= change;
220         assert(reg->memory_size);
221         reg->guest_phys_addr += change;
222         reg->userspace_addr += change;
223         /* Never add more than 1 region */
224         assert(dev->mem->nregions == n);
225         ++dev->mem->nregions;
226         ++split;
227     }
228 }
229 
230 /* Called after unassign, so no regions overlap the given range. */
231 static void vhost_dev_assign_memory(struct vhost_dev *dev,
232                                     uint64_t start_addr,
233                                     uint64_t size,
234                                     uint64_t uaddr)
235 {
236     int from, to;
237     struct vhost_memory_region *merged = NULL;
238     for (from = 0, to = 0; from < dev->mem->nregions; ++from, ++to) {
239         struct vhost_memory_region *reg = dev->mem->regions + to;
240         uint64_t prlast, urlast;
241         uint64_t pmlast, umlast;
242         uint64_t s, e, u;
243 
244         /* clone old region */
245         if (to != from) {
246             memcpy(reg, dev->mem->regions + from, sizeof *reg);
247         }
248         prlast = range_get_last(reg->guest_phys_addr, reg->memory_size);
249         pmlast = range_get_last(start_addr, size);
250         urlast = range_get_last(reg->userspace_addr, reg->memory_size);
251         umlast = range_get_last(uaddr, size);
252 
253         /* check for overlapping regions: should never happen. */
254         assert(prlast < start_addr || pmlast < reg->guest_phys_addr);
255         /* Not an adjacent or overlapping region - do not merge. */
256         if ((prlast + 1 != start_addr || urlast + 1 != uaddr) &&
257             (pmlast + 1 != reg->guest_phys_addr ||
258              umlast + 1 != reg->userspace_addr)) {
259             continue;
260         }
261 
262         if (merged) {
263             --to;
264             assert(to >= 0);
265         } else {
266             merged = reg;
267         }
268         u = MIN(uaddr, reg->userspace_addr);
269         s = MIN(start_addr, reg->guest_phys_addr);
270         e = MAX(pmlast, prlast);
271         uaddr = merged->userspace_addr = u;
272         start_addr = merged->guest_phys_addr = s;
273         size = merged->memory_size = e - s + 1;
274         assert(merged->memory_size);
275     }
276 
277     if (!merged) {
278         struct vhost_memory_region *reg = dev->mem->regions + to;
279         memset(reg, 0, sizeof *reg);
280         reg->memory_size = size;
281         assert(reg->memory_size);
282         reg->guest_phys_addr = start_addr;
283         reg->userspace_addr = uaddr;
284         ++to;
285     }
286     assert(to <= dev->mem->nregions + 1);
287     dev->mem->nregions = to;
288 }
289 
290 static uint64_t vhost_get_log_size(struct vhost_dev *dev)
291 {
292     uint64_t log_size = 0;
293     int i;
294     for (i = 0; i < dev->mem->nregions; ++i) {
295         struct vhost_memory_region *reg = dev->mem->regions + i;
296         uint64_t last = range_get_last(reg->guest_phys_addr,
297                                        reg->memory_size);
298         log_size = MAX(log_size, last / VHOST_LOG_CHUNK + 1);
299     }
300     for (i = 0; i < dev->nvqs; ++i) {
301         struct vhost_virtqueue *vq = dev->vqs + i;
302         uint64_t last = vq->used_phys + vq->used_size - 1;
303         log_size = MAX(log_size, last / VHOST_LOG_CHUNK + 1);
304     }
305     return log_size;
306 }
307 
308 static struct vhost_log *vhost_log_alloc(uint64_t size, bool share)
309 {
310     struct vhost_log *log;
311     uint64_t logsize = size * sizeof(*(log->log));
312     int fd = -1;
313 
314     log = g_new0(struct vhost_log, 1);
315     if (share) {
316         log->log = qemu_memfd_alloc("vhost-log", logsize,
317                                     F_SEAL_GROW | F_SEAL_SHRINK | F_SEAL_SEAL,
318                                     &fd);
319         memset(log->log, 0, logsize);
320     } else {
321         log->log = g_malloc0(logsize);
322     }
323 
324     log->size = size;
325     log->refcnt = 1;
326     log->fd = fd;
327 
328     return log;
329 }
330 
331 static struct vhost_log *vhost_log_get(uint64_t size, bool share)
332 {
333     struct vhost_log *log = share ? vhost_log_shm : vhost_log;
334 
335     if (!log || log->size != size) {
336         log = vhost_log_alloc(size, share);
337         if (share) {
338             vhost_log_shm = log;
339         } else {
340             vhost_log = log;
341         }
342     } else {
343         ++log->refcnt;
344     }
345 
346     return log;
347 }
348 
349 static void vhost_log_put(struct vhost_dev *dev, bool sync)
350 {
351     struct vhost_log *log = dev->log;
352 
353     if (!log) {
354         return;
355     }
356 
357     --log->refcnt;
358     if (log->refcnt == 0) {
359         /* Sync only the range covered by the old log */
360         if (dev->log_size && sync) {
361             vhost_log_sync_range(dev, 0, dev->log_size * VHOST_LOG_CHUNK - 1);
362         }
363 
364         if (vhost_log == log) {
365             g_free(log->log);
366             vhost_log = NULL;
367         } else if (vhost_log_shm == log) {
368             qemu_memfd_free(log->log, log->size * sizeof(*(log->log)),
369                             log->fd);
370             vhost_log_shm = NULL;
371         }
372 
373         g_free(log);
374     }
375 }
376 
377 static bool vhost_dev_log_is_shared(struct vhost_dev *dev)
378 {
379     return dev->vhost_ops->vhost_requires_shm_log &&
380            dev->vhost_ops->vhost_requires_shm_log(dev);
381 }
382 
383 static inline void vhost_dev_log_resize(struct vhost_dev *dev, uint64_t size)
384 {
385     struct vhost_log *log = vhost_log_get(size, vhost_dev_log_is_shared(dev));
386     uint64_t log_base = (uintptr_t)log->log;
387     int r;
388 
389     /* inform backend of log switching, this must be done before
390        releasing the current log, to ensure no logging is lost */
391     r = dev->vhost_ops->vhost_set_log_base(dev, log_base, log);
392     assert(r >= 0);
393     vhost_log_put(dev, true);
394     dev->log = log;
395     dev->log_size = size;
396 }
397 
398 static int vhost_verify_ring_mappings(struct vhost_dev *dev,
399                                       uint64_t start_addr,
400                                       uint64_t size)
401 {
402     int i;
403     int r = 0;
404 
405     for (i = 0; !r && i < dev->nvqs; ++i) {
406         struct vhost_virtqueue *vq = dev->vqs + i;
407         hwaddr l;
408         void *p;
409 
410         if (!ranges_overlap(start_addr, size, vq->ring_phys, vq->ring_size)) {
411             continue;
412         }
413         l = vq->ring_size;
414         p = cpu_physical_memory_map(vq->ring_phys, &l, 1);
415         if (!p || l != vq->ring_size) {
416             fprintf(stderr, "Unable to map ring buffer for ring %d\n", i);
417             r = -ENOMEM;
418         }
419         if (p != vq->ring) {
420             fprintf(stderr, "Ring buffer relocated for ring %d\n", i);
421             r = -EBUSY;
422         }
423         cpu_physical_memory_unmap(p, l, 0, 0);
424     }
425     return r;
426 }
427 
428 static struct vhost_memory_region *vhost_dev_find_reg(struct vhost_dev *dev,
429 						      uint64_t start_addr,
430 						      uint64_t size)
431 {
432     int i, n = dev->mem->nregions;
433     for (i = 0; i < n; ++i) {
434         struct vhost_memory_region *reg = dev->mem->regions + i;
435         if (ranges_overlap(reg->guest_phys_addr, reg->memory_size,
436                            start_addr, size)) {
437             return reg;
438         }
439     }
440     return NULL;
441 }
442 
443 static bool vhost_dev_cmp_memory(struct vhost_dev *dev,
444                                  uint64_t start_addr,
445                                  uint64_t size,
446                                  uint64_t uaddr)
447 {
448     struct vhost_memory_region *reg = vhost_dev_find_reg(dev, start_addr, size);
449     uint64_t reglast;
450     uint64_t memlast;
451 
452     if (!reg) {
453         return true;
454     }
455 
456     reglast = range_get_last(reg->guest_phys_addr, reg->memory_size);
457     memlast = range_get_last(start_addr, size);
458 
459     /* Need to extend region? */
460     if (start_addr < reg->guest_phys_addr || memlast > reglast) {
461         return true;
462     }
463     /* userspace_addr changed? */
464     return uaddr != reg->userspace_addr + start_addr - reg->guest_phys_addr;
465 }
466 
467 static void vhost_set_memory(MemoryListener *listener,
468                              MemoryRegionSection *section,
469                              bool add)
470 {
471     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
472                                          memory_listener);
473     hwaddr start_addr = section->offset_within_address_space;
474     ram_addr_t size = int128_get64(section->size);
475     bool log_dirty =
476         memory_region_get_dirty_log_mask(section->mr) & ~(1 << DIRTY_MEMORY_MIGRATION);
477     int s = offsetof(struct vhost_memory, regions) +
478         (dev->mem->nregions + 1) * sizeof dev->mem->regions[0];
479     void *ram;
480 
481     dev->mem = g_realloc(dev->mem, s);
482 
483     if (log_dirty) {
484         add = false;
485     }
486 
487     assert(size);
488 
489     /* Optimize no-change case. At least cirrus_vga does this a lot at this time. */
490     ram = memory_region_get_ram_ptr(section->mr) + section->offset_within_region;
491     if (add) {
492         if (!vhost_dev_cmp_memory(dev, start_addr, size, (uintptr_t)ram)) {
493             /* Region exists with same address. Nothing to do. */
494             return;
495         }
496     } else {
497         if (!vhost_dev_find_reg(dev, start_addr, size)) {
498             /* Removing region that we don't access. Nothing to do. */
499             return;
500         }
501     }
502 
503     vhost_dev_unassign_memory(dev, start_addr, size);
504     if (add) {
505         /* Add given mapping, merging adjacent regions if any */
506         vhost_dev_assign_memory(dev, start_addr, size, (uintptr_t)ram);
507     } else {
508         /* Remove old mapping for this memory, if any. */
509         vhost_dev_unassign_memory(dev, start_addr, size);
510     }
511     dev->mem_changed_start_addr = MIN(dev->mem_changed_start_addr, start_addr);
512     dev->mem_changed_end_addr = MAX(dev->mem_changed_end_addr, start_addr + size - 1);
513     dev->memory_changed = true;
514     used_memslots = dev->mem->nregions;
515 }
516 
517 static bool vhost_section(MemoryRegionSection *section)
518 {
519     return memory_region_is_ram(section->mr);
520 }
521 
522 static void vhost_begin(MemoryListener *listener)
523 {
524     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
525                                          memory_listener);
526     dev->mem_changed_end_addr = 0;
527     dev->mem_changed_start_addr = -1;
528 }
529 
530 static void vhost_commit(MemoryListener *listener)
531 {
532     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
533                                          memory_listener);
534     hwaddr start_addr = 0;
535     ram_addr_t size = 0;
536     uint64_t log_size;
537     int r;
538 
539     if (!dev->memory_changed) {
540         return;
541     }
542     if (!dev->started) {
543         return;
544     }
545     if (dev->mem_changed_start_addr > dev->mem_changed_end_addr) {
546         return;
547     }
548 
549     if (dev->started) {
550         start_addr = dev->mem_changed_start_addr;
551         size = dev->mem_changed_end_addr - dev->mem_changed_start_addr + 1;
552 
553         r = vhost_verify_ring_mappings(dev, start_addr, size);
554         assert(r >= 0);
555     }
556 
557     if (!dev->log_enabled) {
558         r = dev->vhost_ops->vhost_set_mem_table(dev, dev->mem);
559         assert(r >= 0);
560         dev->memory_changed = false;
561         return;
562     }
563     log_size = vhost_get_log_size(dev);
564     /* We allocate an extra 4K bytes to log,
565      * to reduce the * number of reallocations. */
566 #define VHOST_LOG_BUFFER (0x1000 / sizeof *dev->log)
567     /* To log more, must increase log size before table update. */
568     if (dev->log_size < log_size) {
569         vhost_dev_log_resize(dev, log_size + VHOST_LOG_BUFFER);
570     }
571     r = dev->vhost_ops->vhost_set_mem_table(dev, dev->mem);
572     assert(r >= 0);
573     /* To log less, can only decrease log size after table update. */
574     if (dev->log_size > log_size + VHOST_LOG_BUFFER) {
575         vhost_dev_log_resize(dev, log_size);
576     }
577     dev->memory_changed = false;
578 }
579 
580 static void vhost_region_add(MemoryListener *listener,
581                              MemoryRegionSection *section)
582 {
583     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
584                                          memory_listener);
585 
586     if (!vhost_section(section)) {
587         return;
588     }
589 
590     ++dev->n_mem_sections;
591     dev->mem_sections = g_renew(MemoryRegionSection, dev->mem_sections,
592                                 dev->n_mem_sections);
593     dev->mem_sections[dev->n_mem_sections - 1] = *section;
594     memory_region_ref(section->mr);
595     vhost_set_memory(listener, section, true);
596 }
597 
598 static void vhost_region_del(MemoryListener *listener,
599                              MemoryRegionSection *section)
600 {
601     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
602                                          memory_listener);
603     int i;
604 
605     if (!vhost_section(section)) {
606         return;
607     }
608 
609     vhost_set_memory(listener, section, false);
610     memory_region_unref(section->mr);
611     for (i = 0; i < dev->n_mem_sections; ++i) {
612         if (dev->mem_sections[i].offset_within_address_space
613             == section->offset_within_address_space) {
614             --dev->n_mem_sections;
615             memmove(&dev->mem_sections[i], &dev->mem_sections[i+1],
616                     (dev->n_mem_sections - i) * sizeof(*dev->mem_sections));
617             break;
618         }
619     }
620 }
621 
622 static void vhost_region_nop(MemoryListener *listener,
623                              MemoryRegionSection *section)
624 {
625 }
626 
627 static int vhost_virtqueue_set_addr(struct vhost_dev *dev,
628                                     struct vhost_virtqueue *vq,
629                                     unsigned idx, bool enable_log)
630 {
631     struct vhost_vring_addr addr = {
632         .index = idx,
633         .desc_user_addr = (uint64_t)(unsigned long)vq->desc,
634         .avail_user_addr = (uint64_t)(unsigned long)vq->avail,
635         .used_user_addr = (uint64_t)(unsigned long)vq->used,
636         .log_guest_addr = vq->used_phys,
637         .flags = enable_log ? (1 << VHOST_VRING_F_LOG) : 0,
638     };
639     int r = dev->vhost_ops->vhost_set_vring_addr(dev, &addr);
640     if (r < 0) {
641         return -errno;
642     }
643     return 0;
644 }
645 
646 static int vhost_dev_set_features(struct vhost_dev *dev, bool enable_log)
647 {
648     uint64_t features = dev->acked_features;
649     int r;
650     if (enable_log) {
651         features |= 0x1ULL << VHOST_F_LOG_ALL;
652     }
653     r = dev->vhost_ops->vhost_set_features(dev, features);
654     return r < 0 ? -errno : 0;
655 }
656 
657 static int vhost_dev_set_log(struct vhost_dev *dev, bool enable_log)
658 {
659     int r, t, i, idx;
660     r = vhost_dev_set_features(dev, enable_log);
661     if (r < 0) {
662         goto err_features;
663     }
664     for (i = 0; i < dev->nvqs; ++i) {
665         idx = dev->vhost_ops->vhost_get_vq_index(dev, dev->vq_index + i);
666         r = vhost_virtqueue_set_addr(dev, dev->vqs + i, idx,
667                                      enable_log);
668         if (r < 0) {
669             goto err_vq;
670         }
671     }
672     return 0;
673 err_vq:
674     for (; i >= 0; --i) {
675         idx = dev->vhost_ops->vhost_get_vq_index(dev, dev->vq_index + i);
676         t = vhost_virtqueue_set_addr(dev, dev->vqs + i, idx,
677                                      dev->log_enabled);
678         assert(t >= 0);
679     }
680     t = vhost_dev_set_features(dev, dev->log_enabled);
681     assert(t >= 0);
682 err_features:
683     return r;
684 }
685 
686 static int vhost_migration_log(MemoryListener *listener, int enable)
687 {
688     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
689                                          memory_listener);
690     int r;
691     if (!!enable == dev->log_enabled) {
692         return 0;
693     }
694     if (!dev->started) {
695         dev->log_enabled = enable;
696         return 0;
697     }
698     if (!enable) {
699         r = vhost_dev_set_log(dev, false);
700         if (r < 0) {
701             return r;
702         }
703         vhost_log_put(dev, false);
704         dev->log = NULL;
705         dev->log_size = 0;
706     } else {
707         vhost_dev_log_resize(dev, vhost_get_log_size(dev));
708         r = vhost_dev_set_log(dev, true);
709         if (r < 0) {
710             return r;
711         }
712     }
713     dev->log_enabled = enable;
714     return 0;
715 }
716 
717 static void vhost_log_global_start(MemoryListener *listener)
718 {
719     int r;
720 
721     r = vhost_migration_log(listener, true);
722     if (r < 0) {
723         abort();
724     }
725 }
726 
727 static void vhost_log_global_stop(MemoryListener *listener)
728 {
729     int r;
730 
731     r = vhost_migration_log(listener, false);
732     if (r < 0) {
733         abort();
734     }
735 }
736 
737 static void vhost_log_start(MemoryListener *listener,
738                             MemoryRegionSection *section,
739                             int old, int new)
740 {
741     /* FIXME: implement */
742 }
743 
744 static void vhost_log_stop(MemoryListener *listener,
745                            MemoryRegionSection *section,
746                            int old, int new)
747 {
748     /* FIXME: implement */
749 }
750 
751 static int vhost_virtqueue_set_vring_endian_legacy(struct vhost_dev *dev,
752                                                    bool is_big_endian,
753                                                    int vhost_vq_index)
754 {
755     struct vhost_vring_state s = {
756         .index = vhost_vq_index,
757         .num = is_big_endian
758     };
759 
760     if (!dev->vhost_ops->vhost_set_vring_endian(dev, &s)) {
761         return 0;
762     }
763 
764     if (errno == ENOTTY) {
765         error_report("vhost does not support cross-endian");
766         return -ENOSYS;
767     }
768 
769     return -errno;
770 }
771 
772 static int vhost_virtqueue_start(struct vhost_dev *dev,
773                                 struct VirtIODevice *vdev,
774                                 struct vhost_virtqueue *vq,
775                                 unsigned idx)
776 {
777     hwaddr s, l, a;
778     int r;
779     int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, idx);
780     struct vhost_vring_file file = {
781         .index = vhost_vq_index
782     };
783     struct vhost_vring_state state = {
784         .index = vhost_vq_index
785     };
786     struct VirtQueue *vvq = virtio_get_queue(vdev, idx);
787 
788 
789     vq->num = state.num = virtio_queue_get_num(vdev, idx);
790     r = dev->vhost_ops->vhost_set_vring_num(dev, &state);
791     if (r) {
792         return -errno;
793     }
794 
795     state.num = virtio_queue_get_last_avail_idx(vdev, idx);
796     r = dev->vhost_ops->vhost_set_vring_base(dev, &state);
797     if (r) {
798         return -errno;
799     }
800 
801     if (!virtio_vdev_has_feature(vdev, VIRTIO_F_VERSION_1) &&
802         virtio_legacy_is_cross_endian(vdev)) {
803         r = vhost_virtqueue_set_vring_endian_legacy(dev,
804                                                     virtio_is_big_endian(vdev),
805                                                     vhost_vq_index);
806         if (r) {
807             return -errno;
808         }
809     }
810 
811     s = l = virtio_queue_get_desc_size(vdev, idx);
812     a = virtio_queue_get_desc_addr(vdev, idx);
813     vq->desc = cpu_physical_memory_map(a, &l, 0);
814     if (!vq->desc || l != s) {
815         r = -ENOMEM;
816         goto fail_alloc_desc;
817     }
818     s = l = virtio_queue_get_avail_size(vdev, idx);
819     a = virtio_queue_get_avail_addr(vdev, idx);
820     vq->avail = cpu_physical_memory_map(a, &l, 0);
821     if (!vq->avail || l != s) {
822         r = -ENOMEM;
823         goto fail_alloc_avail;
824     }
825     vq->used_size = s = l = virtio_queue_get_used_size(vdev, idx);
826     vq->used_phys = a = virtio_queue_get_used_addr(vdev, idx);
827     vq->used = cpu_physical_memory_map(a, &l, 1);
828     if (!vq->used || l != s) {
829         r = -ENOMEM;
830         goto fail_alloc_used;
831     }
832 
833     vq->ring_size = s = l = virtio_queue_get_ring_size(vdev, idx);
834     vq->ring_phys = a = virtio_queue_get_ring_addr(vdev, idx);
835     vq->ring = cpu_physical_memory_map(a, &l, 1);
836     if (!vq->ring || l != s) {
837         r = -ENOMEM;
838         goto fail_alloc_ring;
839     }
840 
841     r = vhost_virtqueue_set_addr(dev, vq, vhost_vq_index, dev->log_enabled);
842     if (r < 0) {
843         r = -errno;
844         goto fail_alloc;
845     }
846 
847     file.fd = event_notifier_get_fd(virtio_queue_get_host_notifier(vvq));
848     r = dev->vhost_ops->vhost_set_vring_kick(dev, &file);
849     if (r) {
850         r = -errno;
851         goto fail_kick;
852     }
853 
854     /* Clear and discard previous events if any. */
855     event_notifier_test_and_clear(&vq->masked_notifier);
856 
857     return 0;
858 
859 fail_kick:
860 fail_alloc:
861     cpu_physical_memory_unmap(vq->ring, virtio_queue_get_ring_size(vdev, idx),
862                               0, 0);
863 fail_alloc_ring:
864     cpu_physical_memory_unmap(vq->used, virtio_queue_get_used_size(vdev, idx),
865                               0, 0);
866 fail_alloc_used:
867     cpu_physical_memory_unmap(vq->avail, virtio_queue_get_avail_size(vdev, idx),
868                               0, 0);
869 fail_alloc_avail:
870     cpu_physical_memory_unmap(vq->desc, virtio_queue_get_desc_size(vdev, idx),
871                               0, 0);
872 fail_alloc_desc:
873     return r;
874 }
875 
876 static void vhost_virtqueue_stop(struct vhost_dev *dev,
877                                     struct VirtIODevice *vdev,
878                                     struct vhost_virtqueue *vq,
879                                     unsigned idx)
880 {
881     int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, idx);
882     struct vhost_vring_state state = {
883         .index = vhost_vq_index,
884     };
885     int r;
886 
887     r = dev->vhost_ops->vhost_get_vring_base(dev, &state);
888     if (r < 0) {
889         fprintf(stderr, "vhost VQ %d ring restore failed: %d\n", idx, r);
890         fflush(stderr);
891     }
892     virtio_queue_set_last_avail_idx(vdev, idx, state.num);
893     virtio_queue_invalidate_signalled_used(vdev, idx);
894 
895     /* In the cross-endian case, we need to reset the vring endianness to
896      * native as legacy devices expect so by default.
897      */
898     if (!virtio_vdev_has_feature(vdev, VIRTIO_F_VERSION_1) &&
899         virtio_legacy_is_cross_endian(vdev)) {
900         r = vhost_virtqueue_set_vring_endian_legacy(dev,
901                                                     !virtio_is_big_endian(vdev),
902                                                     vhost_vq_index);
903         if (r < 0) {
904             error_report("failed to reset vring endianness");
905         }
906     }
907 
908     assert (r >= 0);
909     cpu_physical_memory_unmap(vq->ring, virtio_queue_get_ring_size(vdev, idx),
910                               0, virtio_queue_get_ring_size(vdev, idx));
911     cpu_physical_memory_unmap(vq->used, virtio_queue_get_used_size(vdev, idx),
912                               1, virtio_queue_get_used_size(vdev, idx));
913     cpu_physical_memory_unmap(vq->avail, virtio_queue_get_avail_size(vdev, idx),
914                               0, virtio_queue_get_avail_size(vdev, idx));
915     cpu_physical_memory_unmap(vq->desc, virtio_queue_get_desc_size(vdev, idx),
916                               0, virtio_queue_get_desc_size(vdev, idx));
917 }
918 
919 static void vhost_eventfd_add(MemoryListener *listener,
920                               MemoryRegionSection *section,
921                               bool match_data, uint64_t data, EventNotifier *e)
922 {
923 }
924 
925 static void vhost_eventfd_del(MemoryListener *listener,
926                               MemoryRegionSection *section,
927                               bool match_data, uint64_t data, EventNotifier *e)
928 {
929 }
930 
931 static int vhost_virtqueue_init(struct vhost_dev *dev,
932                                 struct vhost_virtqueue *vq, int n)
933 {
934     int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, n);
935     struct vhost_vring_file file = {
936         .index = vhost_vq_index,
937     };
938     int r = event_notifier_init(&vq->masked_notifier, 0);
939     if (r < 0) {
940         return r;
941     }
942 
943     file.fd = event_notifier_get_fd(&vq->masked_notifier);
944     r = dev->vhost_ops->vhost_set_vring_call(dev, &file);
945     if (r) {
946         r = -errno;
947         goto fail_call;
948     }
949     return 0;
950 fail_call:
951     event_notifier_cleanup(&vq->masked_notifier);
952     return r;
953 }
954 
955 static void vhost_virtqueue_cleanup(struct vhost_virtqueue *vq)
956 {
957     event_notifier_cleanup(&vq->masked_notifier);
958 }
959 
960 int vhost_dev_init(struct vhost_dev *hdev, void *opaque,
961                    VhostBackendType backend_type)
962 {
963     uint64_t features;
964     int i, r;
965 
966     hdev->migration_blocker = NULL;
967 
968     if (vhost_set_backend_type(hdev, backend_type) < 0) {
969         close((uintptr_t)opaque);
970         return -1;
971     }
972 
973     if (hdev->vhost_ops->vhost_backend_init(hdev, opaque) < 0) {
974         close((uintptr_t)opaque);
975         return -errno;
976     }
977 
978     if (used_memslots > hdev->vhost_ops->vhost_backend_memslots_limit(hdev)) {
979         fprintf(stderr, "vhost backend memory slots limit is less"
980                 " than current number of present memory slots\n");
981         close((uintptr_t)opaque);
982         return -1;
983     }
984     QLIST_INSERT_HEAD(&vhost_devices, hdev, entry);
985 
986     r = hdev->vhost_ops->vhost_set_owner(hdev);
987     if (r < 0) {
988         goto fail;
989     }
990 
991     r = hdev->vhost_ops->vhost_get_features(hdev, &features);
992     if (r < 0) {
993         goto fail;
994     }
995 
996     for (i = 0; i < hdev->nvqs; ++i) {
997         r = vhost_virtqueue_init(hdev, hdev->vqs + i, hdev->vq_index + i);
998         if (r < 0) {
999             goto fail_vq;
1000         }
1001     }
1002     hdev->features = features;
1003 
1004     hdev->memory_listener = (MemoryListener) {
1005         .begin = vhost_begin,
1006         .commit = vhost_commit,
1007         .region_add = vhost_region_add,
1008         .region_del = vhost_region_del,
1009         .region_nop = vhost_region_nop,
1010         .log_start = vhost_log_start,
1011         .log_stop = vhost_log_stop,
1012         .log_sync = vhost_log_sync,
1013         .log_global_start = vhost_log_global_start,
1014         .log_global_stop = vhost_log_global_stop,
1015         .eventfd_add = vhost_eventfd_add,
1016         .eventfd_del = vhost_eventfd_del,
1017         .priority = 10
1018     };
1019 
1020     if (hdev->migration_blocker == NULL) {
1021         if (!(hdev->features & (0x1ULL << VHOST_F_LOG_ALL))) {
1022             error_setg(&hdev->migration_blocker,
1023                        "Migration disabled: vhost lacks VHOST_F_LOG_ALL feature.");
1024         } else if (!qemu_memfd_check()) {
1025             error_setg(&hdev->migration_blocker,
1026                        "Migration disabled: failed to allocate shared memory");
1027         }
1028     }
1029 
1030     if (hdev->migration_blocker != NULL) {
1031         migrate_add_blocker(hdev->migration_blocker);
1032     }
1033 
1034     hdev->mem = g_malloc0(offsetof(struct vhost_memory, regions));
1035     hdev->n_mem_sections = 0;
1036     hdev->mem_sections = NULL;
1037     hdev->log = NULL;
1038     hdev->log_size = 0;
1039     hdev->log_enabled = false;
1040     hdev->started = false;
1041     hdev->memory_changed = false;
1042     memory_listener_register(&hdev->memory_listener, &address_space_memory);
1043     return 0;
1044 fail_vq:
1045     while (--i >= 0) {
1046         vhost_virtqueue_cleanup(hdev->vqs + i);
1047     }
1048 fail:
1049     r = -errno;
1050     hdev->vhost_ops->vhost_backend_cleanup(hdev);
1051     QLIST_REMOVE(hdev, entry);
1052     return r;
1053 }
1054 
1055 void vhost_dev_cleanup(struct vhost_dev *hdev)
1056 {
1057     int i;
1058     for (i = 0; i < hdev->nvqs; ++i) {
1059         vhost_virtqueue_cleanup(hdev->vqs + i);
1060     }
1061     memory_listener_unregister(&hdev->memory_listener);
1062     if (hdev->migration_blocker) {
1063         migrate_del_blocker(hdev->migration_blocker);
1064         error_free(hdev->migration_blocker);
1065     }
1066     g_free(hdev->mem);
1067     g_free(hdev->mem_sections);
1068     hdev->vhost_ops->vhost_backend_cleanup(hdev);
1069     QLIST_REMOVE(hdev, entry);
1070 }
1071 
1072 /* Stop processing guest IO notifications in qemu.
1073  * Start processing them in vhost in kernel.
1074  */
1075 int vhost_dev_enable_notifiers(struct vhost_dev *hdev, VirtIODevice *vdev)
1076 {
1077     BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev)));
1078     VirtioBusState *vbus = VIRTIO_BUS(qbus);
1079     VirtioBusClass *k = VIRTIO_BUS_GET_CLASS(vbus);
1080     int i, r, e;
1081     if (!k->set_host_notifier) {
1082         fprintf(stderr, "binding does not support host notifiers\n");
1083         r = -ENOSYS;
1084         goto fail;
1085     }
1086 
1087     for (i = 0; i < hdev->nvqs; ++i) {
1088         r = k->set_host_notifier(qbus->parent, hdev->vq_index + i, true);
1089         if (r < 0) {
1090             fprintf(stderr, "vhost VQ %d notifier binding failed: %d\n", i, -r);
1091             goto fail_vq;
1092         }
1093     }
1094 
1095     return 0;
1096 fail_vq:
1097     while (--i >= 0) {
1098         e = k->set_host_notifier(qbus->parent, hdev->vq_index + i, false);
1099         if (e < 0) {
1100             fprintf(stderr, "vhost VQ %d notifier cleanup error: %d\n", i, -r);
1101             fflush(stderr);
1102         }
1103         assert (e >= 0);
1104     }
1105 fail:
1106     return r;
1107 }
1108 
1109 /* Stop processing guest IO notifications in vhost.
1110  * Start processing them in qemu.
1111  * This might actually run the qemu handlers right away,
1112  * so virtio in qemu must be completely setup when this is called.
1113  */
1114 void vhost_dev_disable_notifiers(struct vhost_dev *hdev, VirtIODevice *vdev)
1115 {
1116     BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev)));
1117     VirtioBusState *vbus = VIRTIO_BUS(qbus);
1118     VirtioBusClass *k = VIRTIO_BUS_GET_CLASS(vbus);
1119     int i, r;
1120 
1121     for (i = 0; i < hdev->nvqs; ++i) {
1122         r = k->set_host_notifier(qbus->parent, hdev->vq_index + i, false);
1123         if (r < 0) {
1124             fprintf(stderr, "vhost VQ %d notifier cleanup failed: %d\n", i, -r);
1125             fflush(stderr);
1126         }
1127         assert (r >= 0);
1128     }
1129 }
1130 
1131 /* Test and clear event pending status.
1132  * Should be called after unmask to avoid losing events.
1133  */
1134 bool vhost_virtqueue_pending(struct vhost_dev *hdev, int n)
1135 {
1136     struct vhost_virtqueue *vq = hdev->vqs + n - hdev->vq_index;
1137     assert(n >= hdev->vq_index && n < hdev->vq_index + hdev->nvqs);
1138     return event_notifier_test_and_clear(&vq->masked_notifier);
1139 }
1140 
1141 /* Mask/unmask events from this vq. */
1142 void vhost_virtqueue_mask(struct vhost_dev *hdev, VirtIODevice *vdev, int n,
1143                          bool mask)
1144 {
1145     struct VirtQueue *vvq = virtio_get_queue(vdev, n);
1146     int r, index = n - hdev->vq_index;
1147     struct vhost_vring_file file;
1148 
1149     if (mask) {
1150         file.fd = event_notifier_get_fd(&hdev->vqs[index].masked_notifier);
1151     } else {
1152         file.fd = event_notifier_get_fd(virtio_queue_get_guest_notifier(vvq));
1153     }
1154 
1155     file.index = hdev->vhost_ops->vhost_get_vq_index(hdev, n);
1156     r = hdev->vhost_ops->vhost_set_vring_call(hdev, &file);
1157     assert(r >= 0);
1158 }
1159 
1160 uint64_t vhost_get_features(struct vhost_dev *hdev, const int *feature_bits,
1161                             uint64_t features)
1162 {
1163     const int *bit = feature_bits;
1164     while (*bit != VHOST_INVALID_FEATURE_BIT) {
1165         uint64_t bit_mask = (1ULL << *bit);
1166         if (!(hdev->features & bit_mask)) {
1167             features &= ~bit_mask;
1168         }
1169         bit++;
1170     }
1171     return features;
1172 }
1173 
1174 void vhost_ack_features(struct vhost_dev *hdev, const int *feature_bits,
1175                         uint64_t features)
1176 {
1177     const int *bit = feature_bits;
1178     while (*bit != VHOST_INVALID_FEATURE_BIT) {
1179         uint64_t bit_mask = (1ULL << *bit);
1180         if (features & bit_mask) {
1181             hdev->acked_features |= bit_mask;
1182         }
1183         bit++;
1184     }
1185 }
1186 
1187 /* Host notifiers must be enabled at this point. */
1188 int vhost_dev_start(struct vhost_dev *hdev, VirtIODevice *vdev)
1189 {
1190     int i, r;
1191 
1192     hdev->started = true;
1193 
1194     r = vhost_dev_set_features(hdev, hdev->log_enabled);
1195     if (r < 0) {
1196         goto fail_features;
1197     }
1198     r = hdev->vhost_ops->vhost_set_mem_table(hdev, hdev->mem);
1199     if (r < 0) {
1200         r = -errno;
1201         goto fail_mem;
1202     }
1203     for (i = 0; i < hdev->nvqs; ++i) {
1204         r = vhost_virtqueue_start(hdev,
1205                                   vdev,
1206                                   hdev->vqs + i,
1207                                   hdev->vq_index + i);
1208         if (r < 0) {
1209             goto fail_vq;
1210         }
1211     }
1212 
1213     if (hdev->log_enabled) {
1214         uint64_t log_base;
1215 
1216         hdev->log_size = vhost_get_log_size(hdev);
1217         hdev->log = vhost_log_get(hdev->log_size,
1218                                   vhost_dev_log_is_shared(hdev));
1219         log_base = (uintptr_t)hdev->log->log;
1220         r = hdev->vhost_ops->vhost_set_log_base(hdev,
1221                                                 hdev->log_size ? log_base : 0,
1222                                                 hdev->log);
1223         if (r < 0) {
1224             r = -errno;
1225             goto fail_log;
1226         }
1227     }
1228 
1229     return 0;
1230 fail_log:
1231     vhost_log_put(hdev, false);
1232 fail_vq:
1233     while (--i >= 0) {
1234         vhost_virtqueue_stop(hdev,
1235                              vdev,
1236                              hdev->vqs + i,
1237                              hdev->vq_index + i);
1238     }
1239     i = hdev->nvqs;
1240 fail_mem:
1241 fail_features:
1242 
1243     hdev->started = false;
1244     return r;
1245 }
1246 
1247 /* Host notifiers must be enabled at this point. */
1248 void vhost_dev_stop(struct vhost_dev *hdev, VirtIODevice *vdev)
1249 {
1250     int i;
1251 
1252     for (i = 0; i < hdev->nvqs; ++i) {
1253         vhost_virtqueue_stop(hdev,
1254                              vdev,
1255                              hdev->vqs + i,
1256                              hdev->vq_index + i);
1257     }
1258 
1259     vhost_log_put(hdev, true);
1260     hdev->started = false;
1261     hdev->log = NULL;
1262     hdev->log_size = 0;
1263 }
1264 
1265