xref: /openbmc/qemu/hw/virtio/vhost.c (revision da34e65c)
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 
366     --log->refcnt;
367     if (log->refcnt == 0) {
368         /* Sync only the range covered by the old log */
369         if (dev->log_size && sync) {
370             vhost_log_sync_range(dev, 0, dev->log_size * VHOST_LOG_CHUNK - 1);
371         }
372 
373         if (vhost_log == log) {
374             g_free(log->log);
375             vhost_log = NULL;
376         } else if (vhost_log_shm == log) {
377             qemu_memfd_free(log->log, log->size * sizeof(*(log->log)),
378                             log->fd);
379             vhost_log_shm = NULL;
380         }
381 
382         g_free(log);
383     }
384 }
385 
386 static bool vhost_dev_log_is_shared(struct vhost_dev *dev)
387 {
388     return dev->vhost_ops->vhost_requires_shm_log &&
389            dev->vhost_ops->vhost_requires_shm_log(dev);
390 }
391 
392 static inline void vhost_dev_log_resize(struct vhost_dev *dev, uint64_t size)
393 {
394     struct vhost_log *log = vhost_log_get(size, vhost_dev_log_is_shared(dev));
395     uint64_t log_base = (uintptr_t)log->log;
396     int r;
397 
398     /* inform backend of log switching, this must be done before
399        releasing the current log, to ensure no logging is lost */
400     r = dev->vhost_ops->vhost_set_log_base(dev, log_base, log);
401     assert(r >= 0);
402     vhost_log_put(dev, true);
403     dev->log = log;
404     dev->log_size = size;
405 }
406 
407 static int vhost_verify_ring_mappings(struct vhost_dev *dev,
408                                       uint64_t start_addr,
409                                       uint64_t size)
410 {
411     int i;
412     int r = 0;
413 
414     for (i = 0; !r && i < dev->nvqs; ++i) {
415         struct vhost_virtqueue *vq = dev->vqs + i;
416         hwaddr l;
417         void *p;
418 
419         if (!ranges_overlap(start_addr, size, vq->ring_phys, vq->ring_size)) {
420             continue;
421         }
422         l = vq->ring_size;
423         p = cpu_physical_memory_map(vq->ring_phys, &l, 1);
424         if (!p || l != vq->ring_size) {
425             fprintf(stderr, "Unable to map ring buffer for ring %d\n", i);
426             r = -ENOMEM;
427         }
428         if (p != vq->ring) {
429             fprintf(stderr, "Ring buffer relocated for ring %d\n", i);
430             r = -EBUSY;
431         }
432         cpu_physical_memory_unmap(p, l, 0, 0);
433     }
434     return r;
435 }
436 
437 static struct vhost_memory_region *vhost_dev_find_reg(struct vhost_dev *dev,
438 						      uint64_t start_addr,
439 						      uint64_t size)
440 {
441     int i, n = dev->mem->nregions;
442     for (i = 0; i < n; ++i) {
443         struct vhost_memory_region *reg = dev->mem->regions + i;
444         if (ranges_overlap(reg->guest_phys_addr, reg->memory_size,
445                            start_addr, size)) {
446             return reg;
447         }
448     }
449     return NULL;
450 }
451 
452 static bool vhost_dev_cmp_memory(struct vhost_dev *dev,
453                                  uint64_t start_addr,
454                                  uint64_t size,
455                                  uint64_t uaddr)
456 {
457     struct vhost_memory_region *reg = vhost_dev_find_reg(dev, start_addr, size);
458     uint64_t reglast;
459     uint64_t memlast;
460 
461     if (!reg) {
462         return true;
463     }
464 
465     reglast = range_get_last(reg->guest_phys_addr, reg->memory_size);
466     memlast = range_get_last(start_addr, size);
467 
468     /* Need to extend region? */
469     if (start_addr < reg->guest_phys_addr || memlast > reglast) {
470         return true;
471     }
472     /* userspace_addr changed? */
473     return uaddr != reg->userspace_addr + start_addr - reg->guest_phys_addr;
474 }
475 
476 static void vhost_set_memory(MemoryListener *listener,
477                              MemoryRegionSection *section,
478                              bool add)
479 {
480     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
481                                          memory_listener);
482     hwaddr start_addr = section->offset_within_address_space;
483     ram_addr_t size = int128_get64(section->size);
484     bool log_dirty =
485         memory_region_get_dirty_log_mask(section->mr) & ~(1 << DIRTY_MEMORY_MIGRATION);
486     int s = offsetof(struct vhost_memory, regions) +
487         (dev->mem->nregions + 1) * sizeof dev->mem->regions[0];
488     void *ram;
489 
490     dev->mem = g_realloc(dev->mem, s);
491 
492     if (log_dirty) {
493         add = false;
494     }
495 
496     assert(size);
497 
498     /* Optimize no-change case. At least cirrus_vga does this a lot at this time. */
499     ram = memory_region_get_ram_ptr(section->mr) + section->offset_within_region;
500     if (add) {
501         if (!vhost_dev_cmp_memory(dev, start_addr, size, (uintptr_t)ram)) {
502             /* Region exists with same address. Nothing to do. */
503             return;
504         }
505     } else {
506         if (!vhost_dev_find_reg(dev, start_addr, size)) {
507             /* Removing region that we don't access. Nothing to do. */
508             return;
509         }
510     }
511 
512     vhost_dev_unassign_memory(dev, start_addr, size);
513     if (add) {
514         /* Add given mapping, merging adjacent regions if any */
515         vhost_dev_assign_memory(dev, start_addr, size, (uintptr_t)ram);
516     } else {
517         /* Remove old mapping for this memory, if any. */
518         vhost_dev_unassign_memory(dev, start_addr, size);
519     }
520     dev->mem_changed_start_addr = MIN(dev->mem_changed_start_addr, start_addr);
521     dev->mem_changed_end_addr = MAX(dev->mem_changed_end_addr, start_addr + size - 1);
522     dev->memory_changed = true;
523     used_memslots = dev->mem->nregions;
524 }
525 
526 static bool vhost_section(MemoryRegionSection *section)
527 {
528     return memory_region_is_ram(section->mr);
529 }
530 
531 static void vhost_begin(MemoryListener *listener)
532 {
533     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
534                                          memory_listener);
535     dev->mem_changed_end_addr = 0;
536     dev->mem_changed_start_addr = -1;
537 }
538 
539 static void vhost_commit(MemoryListener *listener)
540 {
541     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
542                                          memory_listener);
543     hwaddr start_addr = 0;
544     ram_addr_t size = 0;
545     uint64_t log_size;
546     int r;
547 
548     if (!dev->memory_changed) {
549         return;
550     }
551     if (!dev->started) {
552         return;
553     }
554     if (dev->mem_changed_start_addr > dev->mem_changed_end_addr) {
555         return;
556     }
557 
558     if (dev->started) {
559         start_addr = dev->mem_changed_start_addr;
560         size = dev->mem_changed_end_addr - dev->mem_changed_start_addr + 1;
561 
562         r = vhost_verify_ring_mappings(dev, start_addr, size);
563         assert(r >= 0);
564     }
565 
566     if (!dev->log_enabled) {
567         r = dev->vhost_ops->vhost_set_mem_table(dev, dev->mem);
568         assert(r >= 0);
569         dev->memory_changed = false;
570         return;
571     }
572     log_size = vhost_get_log_size(dev);
573     /* We allocate an extra 4K bytes to log,
574      * to reduce the * number of reallocations. */
575 #define VHOST_LOG_BUFFER (0x1000 / sizeof *dev->log)
576     /* To log more, must increase log size before table update. */
577     if (dev->log_size < log_size) {
578         vhost_dev_log_resize(dev, log_size + VHOST_LOG_BUFFER);
579     }
580     r = dev->vhost_ops->vhost_set_mem_table(dev, dev->mem);
581     assert(r >= 0);
582     /* To log less, can only decrease log size after table update. */
583     if (dev->log_size > log_size + VHOST_LOG_BUFFER) {
584         vhost_dev_log_resize(dev, log_size);
585     }
586     dev->memory_changed = false;
587 }
588 
589 static void vhost_region_add(MemoryListener *listener,
590                              MemoryRegionSection *section)
591 {
592     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
593                                          memory_listener);
594 
595     if (!vhost_section(section)) {
596         return;
597     }
598 
599     ++dev->n_mem_sections;
600     dev->mem_sections = g_renew(MemoryRegionSection, dev->mem_sections,
601                                 dev->n_mem_sections);
602     dev->mem_sections[dev->n_mem_sections - 1] = *section;
603     memory_region_ref(section->mr);
604     vhost_set_memory(listener, section, true);
605 }
606 
607 static void vhost_region_del(MemoryListener *listener,
608                              MemoryRegionSection *section)
609 {
610     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
611                                          memory_listener);
612     int i;
613 
614     if (!vhost_section(section)) {
615         return;
616     }
617 
618     vhost_set_memory(listener, section, false);
619     memory_region_unref(section->mr);
620     for (i = 0; i < dev->n_mem_sections; ++i) {
621         if (dev->mem_sections[i].offset_within_address_space
622             == section->offset_within_address_space) {
623             --dev->n_mem_sections;
624             memmove(&dev->mem_sections[i], &dev->mem_sections[i+1],
625                     (dev->n_mem_sections - i) * sizeof(*dev->mem_sections));
626             break;
627         }
628     }
629 }
630 
631 static void vhost_region_nop(MemoryListener *listener,
632                              MemoryRegionSection *section)
633 {
634 }
635 
636 static int vhost_virtqueue_set_addr(struct vhost_dev *dev,
637                                     struct vhost_virtqueue *vq,
638                                     unsigned idx, bool enable_log)
639 {
640     struct vhost_vring_addr addr = {
641         .index = idx,
642         .desc_user_addr = (uint64_t)(unsigned long)vq->desc,
643         .avail_user_addr = (uint64_t)(unsigned long)vq->avail,
644         .used_user_addr = (uint64_t)(unsigned long)vq->used,
645         .log_guest_addr = vq->used_phys,
646         .flags = enable_log ? (1 << VHOST_VRING_F_LOG) : 0,
647     };
648     int r = dev->vhost_ops->vhost_set_vring_addr(dev, &addr);
649     if (r < 0) {
650         return -errno;
651     }
652     return 0;
653 }
654 
655 static int vhost_dev_set_features(struct vhost_dev *dev, bool enable_log)
656 {
657     uint64_t features = dev->acked_features;
658     int r;
659     if (enable_log) {
660         features |= 0x1ULL << VHOST_F_LOG_ALL;
661     }
662     r = dev->vhost_ops->vhost_set_features(dev, features);
663     return r < 0 ? -errno : 0;
664 }
665 
666 static int vhost_dev_set_log(struct vhost_dev *dev, bool enable_log)
667 {
668     int r, t, i, idx;
669     r = vhost_dev_set_features(dev, enable_log);
670     if (r < 0) {
671         goto err_features;
672     }
673     for (i = 0; i < dev->nvqs; ++i) {
674         idx = dev->vhost_ops->vhost_get_vq_index(dev, dev->vq_index + i);
675         r = vhost_virtqueue_set_addr(dev, dev->vqs + i, idx,
676                                      enable_log);
677         if (r < 0) {
678             goto err_vq;
679         }
680     }
681     return 0;
682 err_vq:
683     for (; i >= 0; --i) {
684         idx = dev->vhost_ops->vhost_get_vq_index(dev, dev->vq_index + i);
685         t = vhost_virtqueue_set_addr(dev, dev->vqs + i, idx,
686                                      dev->log_enabled);
687         assert(t >= 0);
688     }
689     t = vhost_dev_set_features(dev, dev->log_enabled);
690     assert(t >= 0);
691 err_features:
692     return r;
693 }
694 
695 static int vhost_migration_log(MemoryListener *listener, int enable)
696 {
697     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
698                                          memory_listener);
699     int r;
700     if (!!enable == dev->log_enabled) {
701         return 0;
702     }
703     if (!dev->started) {
704         dev->log_enabled = enable;
705         return 0;
706     }
707     if (!enable) {
708         r = vhost_dev_set_log(dev, false);
709         if (r < 0) {
710             return r;
711         }
712         vhost_log_put(dev, false);
713         dev->log = NULL;
714         dev->log_size = 0;
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 TARGET_IS_BIENDIAN
771 #ifdef HOST_WORDS_BIGENDIAN
772     return vdev->device_endian == VIRTIO_DEVICE_ENDIAN_LITTLE;
773 #else
774     return vdev->device_endian == VIRTIO_DEVICE_ENDIAN_BIG;
775 #endif
776 #else
777     return false;
778 #endif
779 }
780 
781 static int vhost_virtqueue_set_vring_endian_legacy(struct vhost_dev *dev,
782                                                    bool is_big_endian,
783                                                    int vhost_vq_index)
784 {
785     struct vhost_vring_state s = {
786         .index = vhost_vq_index,
787         .num = is_big_endian
788     };
789 
790     if (!dev->vhost_ops->vhost_set_vring_endian(dev, &s)) {
791         return 0;
792     }
793 
794     if (errno == ENOTTY) {
795         error_report("vhost does not support cross-endian");
796         return -ENOSYS;
797     }
798 
799     return -errno;
800 }
801 
802 static int vhost_virtqueue_start(struct vhost_dev *dev,
803                                 struct VirtIODevice *vdev,
804                                 struct vhost_virtqueue *vq,
805                                 unsigned idx)
806 {
807     hwaddr s, l, a;
808     int r;
809     int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, idx);
810     struct vhost_vring_file file = {
811         .index = vhost_vq_index
812     };
813     struct vhost_vring_state state = {
814         .index = vhost_vq_index
815     };
816     struct VirtQueue *vvq = virtio_get_queue(vdev, idx);
817 
818 
819     vq->num = state.num = virtio_queue_get_num(vdev, idx);
820     r = dev->vhost_ops->vhost_set_vring_num(dev, &state);
821     if (r) {
822         return -errno;
823     }
824 
825     state.num = virtio_queue_get_last_avail_idx(vdev, idx);
826     r = dev->vhost_ops->vhost_set_vring_base(dev, &state);
827     if (r) {
828         return -errno;
829     }
830 
831     if (vhost_needs_vring_endian(vdev)) {
832         r = vhost_virtqueue_set_vring_endian_legacy(dev,
833                                                     virtio_is_big_endian(vdev),
834                                                     vhost_vq_index);
835         if (r) {
836             return -errno;
837         }
838     }
839 
840     s = l = virtio_queue_get_desc_size(vdev, idx);
841     a = virtio_queue_get_desc_addr(vdev, idx);
842     vq->desc = cpu_physical_memory_map(a, &l, 0);
843     if (!vq->desc || l != s) {
844         r = -ENOMEM;
845         goto fail_alloc_desc;
846     }
847     s = l = virtio_queue_get_avail_size(vdev, idx);
848     a = virtio_queue_get_avail_addr(vdev, idx);
849     vq->avail = cpu_physical_memory_map(a, &l, 0);
850     if (!vq->avail || l != s) {
851         r = -ENOMEM;
852         goto fail_alloc_avail;
853     }
854     vq->used_size = s = l = virtio_queue_get_used_size(vdev, idx);
855     vq->used_phys = a = virtio_queue_get_used_addr(vdev, idx);
856     vq->used = cpu_physical_memory_map(a, &l, 1);
857     if (!vq->used || l != s) {
858         r = -ENOMEM;
859         goto fail_alloc_used;
860     }
861 
862     vq->ring_size = s = l = virtio_queue_get_ring_size(vdev, idx);
863     vq->ring_phys = a = virtio_queue_get_ring_addr(vdev, idx);
864     vq->ring = cpu_physical_memory_map(a, &l, 1);
865     if (!vq->ring || l != s) {
866         r = -ENOMEM;
867         goto fail_alloc_ring;
868     }
869 
870     r = vhost_virtqueue_set_addr(dev, vq, vhost_vq_index, dev->log_enabled);
871     if (r < 0) {
872         r = -errno;
873         goto fail_alloc;
874     }
875 
876     file.fd = event_notifier_get_fd(virtio_queue_get_host_notifier(vvq));
877     r = dev->vhost_ops->vhost_set_vring_kick(dev, &file);
878     if (r) {
879         r = -errno;
880         goto fail_kick;
881     }
882 
883     /* Clear and discard previous events if any. */
884     event_notifier_test_and_clear(&vq->masked_notifier);
885 
886     /* Init vring in unmasked state, unless guest_notifier_mask
887      * will do it later.
888      */
889     if (!vdev->use_guest_notifier_mask) {
890         /* TODO: check and handle errors. */
891         vhost_virtqueue_mask(dev, vdev, idx, false);
892     }
893 
894     return 0;
895 
896 fail_kick:
897 fail_alloc:
898     cpu_physical_memory_unmap(vq->ring, virtio_queue_get_ring_size(vdev, idx),
899                               0, 0);
900 fail_alloc_ring:
901     cpu_physical_memory_unmap(vq->used, virtio_queue_get_used_size(vdev, idx),
902                               0, 0);
903 fail_alloc_used:
904     cpu_physical_memory_unmap(vq->avail, virtio_queue_get_avail_size(vdev, idx),
905                               0, 0);
906 fail_alloc_avail:
907     cpu_physical_memory_unmap(vq->desc, virtio_queue_get_desc_size(vdev, idx),
908                               0, 0);
909 fail_alloc_desc:
910     return r;
911 }
912 
913 static void vhost_virtqueue_stop(struct vhost_dev *dev,
914                                     struct VirtIODevice *vdev,
915                                     struct vhost_virtqueue *vq,
916                                     unsigned idx)
917 {
918     int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, idx);
919     struct vhost_vring_state state = {
920         .index = vhost_vq_index,
921     };
922     int r;
923 
924     r = dev->vhost_ops->vhost_get_vring_base(dev, &state);
925     if (r < 0) {
926         fprintf(stderr, "vhost VQ %d ring restore failed: %d\n", idx, r);
927         fflush(stderr);
928     }
929     virtio_queue_set_last_avail_idx(vdev, idx, state.num);
930     virtio_queue_invalidate_signalled_used(vdev, idx);
931 
932     /* In the cross-endian case, we need to reset the vring endianness to
933      * native as legacy devices expect so by default.
934      */
935     if (vhost_needs_vring_endian(vdev)) {
936         r = vhost_virtqueue_set_vring_endian_legacy(dev,
937                                                     !virtio_is_big_endian(vdev),
938                                                     vhost_vq_index);
939         if (r < 0) {
940             error_report("failed to reset vring endianness");
941         }
942     }
943 
944     assert (r >= 0);
945     cpu_physical_memory_unmap(vq->ring, virtio_queue_get_ring_size(vdev, idx),
946                               0, virtio_queue_get_ring_size(vdev, idx));
947     cpu_physical_memory_unmap(vq->used, virtio_queue_get_used_size(vdev, idx),
948                               1, virtio_queue_get_used_size(vdev, idx));
949     cpu_physical_memory_unmap(vq->avail, virtio_queue_get_avail_size(vdev, idx),
950                               0, virtio_queue_get_avail_size(vdev, idx));
951     cpu_physical_memory_unmap(vq->desc, virtio_queue_get_desc_size(vdev, idx),
952                               0, virtio_queue_get_desc_size(vdev, idx));
953 }
954 
955 static void vhost_eventfd_add(MemoryListener *listener,
956                               MemoryRegionSection *section,
957                               bool match_data, uint64_t data, EventNotifier *e)
958 {
959 }
960 
961 static void vhost_eventfd_del(MemoryListener *listener,
962                               MemoryRegionSection *section,
963                               bool match_data, uint64_t data, EventNotifier *e)
964 {
965 }
966 
967 static int vhost_virtqueue_init(struct vhost_dev *dev,
968                                 struct vhost_virtqueue *vq, int n)
969 {
970     int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, n);
971     struct vhost_vring_file file = {
972         .index = vhost_vq_index,
973     };
974     int r = event_notifier_init(&vq->masked_notifier, 0);
975     if (r < 0) {
976         return r;
977     }
978 
979     file.fd = event_notifier_get_fd(&vq->masked_notifier);
980     r = dev->vhost_ops->vhost_set_vring_call(dev, &file);
981     if (r) {
982         r = -errno;
983         goto fail_call;
984     }
985     return 0;
986 fail_call:
987     event_notifier_cleanup(&vq->masked_notifier);
988     return r;
989 }
990 
991 static void vhost_virtqueue_cleanup(struct vhost_virtqueue *vq)
992 {
993     event_notifier_cleanup(&vq->masked_notifier);
994 }
995 
996 int vhost_dev_init(struct vhost_dev *hdev, void *opaque,
997                    VhostBackendType backend_type)
998 {
999     uint64_t features;
1000     int i, r;
1001 
1002     hdev->migration_blocker = NULL;
1003 
1004     if (vhost_set_backend_type(hdev, backend_type) < 0) {
1005         close((uintptr_t)opaque);
1006         return -1;
1007     }
1008 
1009     if (hdev->vhost_ops->vhost_backend_init(hdev, opaque) < 0) {
1010         close((uintptr_t)opaque);
1011         return -errno;
1012     }
1013 
1014     if (used_memslots > hdev->vhost_ops->vhost_backend_memslots_limit(hdev)) {
1015         fprintf(stderr, "vhost backend memory slots limit is less"
1016                 " than current number of present memory slots\n");
1017         close((uintptr_t)opaque);
1018         return -1;
1019     }
1020     QLIST_INSERT_HEAD(&vhost_devices, hdev, entry);
1021 
1022     r = hdev->vhost_ops->vhost_set_owner(hdev);
1023     if (r < 0) {
1024         goto fail;
1025     }
1026 
1027     r = hdev->vhost_ops->vhost_get_features(hdev, &features);
1028     if (r < 0) {
1029         goto fail;
1030     }
1031 
1032     for (i = 0; i < hdev->nvqs; ++i) {
1033         r = vhost_virtqueue_init(hdev, hdev->vqs + i, hdev->vq_index + i);
1034         if (r < 0) {
1035             goto fail_vq;
1036         }
1037     }
1038     hdev->features = features;
1039 
1040     hdev->memory_listener = (MemoryListener) {
1041         .begin = vhost_begin,
1042         .commit = vhost_commit,
1043         .region_add = vhost_region_add,
1044         .region_del = vhost_region_del,
1045         .region_nop = vhost_region_nop,
1046         .log_start = vhost_log_start,
1047         .log_stop = vhost_log_stop,
1048         .log_sync = vhost_log_sync,
1049         .log_global_start = vhost_log_global_start,
1050         .log_global_stop = vhost_log_global_stop,
1051         .eventfd_add = vhost_eventfd_add,
1052         .eventfd_del = vhost_eventfd_del,
1053         .priority = 10
1054     };
1055 
1056     if (hdev->migration_blocker == NULL) {
1057         if (!(hdev->features & (0x1ULL << VHOST_F_LOG_ALL))) {
1058             error_setg(&hdev->migration_blocker,
1059                        "Migration disabled: vhost lacks VHOST_F_LOG_ALL feature.");
1060         } else if (!qemu_memfd_check()) {
1061             error_setg(&hdev->migration_blocker,
1062                        "Migration disabled: failed to allocate shared memory");
1063         }
1064     }
1065 
1066     if (hdev->migration_blocker != NULL) {
1067         migrate_add_blocker(hdev->migration_blocker);
1068     }
1069 
1070     hdev->mem = g_malloc0(offsetof(struct vhost_memory, regions));
1071     hdev->n_mem_sections = 0;
1072     hdev->mem_sections = NULL;
1073     hdev->log = NULL;
1074     hdev->log_size = 0;
1075     hdev->log_enabled = false;
1076     hdev->started = false;
1077     hdev->memory_changed = false;
1078     memory_listener_register(&hdev->memory_listener, &address_space_memory);
1079     return 0;
1080 fail_vq:
1081     while (--i >= 0) {
1082         vhost_virtqueue_cleanup(hdev->vqs + i);
1083     }
1084 fail:
1085     r = -errno;
1086     hdev->vhost_ops->vhost_backend_cleanup(hdev);
1087     QLIST_REMOVE(hdev, entry);
1088     return r;
1089 }
1090 
1091 void vhost_dev_cleanup(struct vhost_dev *hdev)
1092 {
1093     int i;
1094     for (i = 0; i < hdev->nvqs; ++i) {
1095         vhost_virtqueue_cleanup(hdev->vqs + i);
1096     }
1097     memory_listener_unregister(&hdev->memory_listener);
1098     if (hdev->migration_blocker) {
1099         migrate_del_blocker(hdev->migration_blocker);
1100         error_free(hdev->migration_blocker);
1101     }
1102     g_free(hdev->mem);
1103     g_free(hdev->mem_sections);
1104     hdev->vhost_ops->vhost_backend_cleanup(hdev);
1105     QLIST_REMOVE(hdev, entry);
1106 }
1107 
1108 /* Stop processing guest IO notifications in qemu.
1109  * Start processing them in vhost in kernel.
1110  */
1111 int vhost_dev_enable_notifiers(struct vhost_dev *hdev, VirtIODevice *vdev)
1112 {
1113     BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev)));
1114     VirtioBusState *vbus = VIRTIO_BUS(qbus);
1115     VirtioBusClass *k = VIRTIO_BUS_GET_CLASS(vbus);
1116     int i, r, e;
1117     if (!k->set_host_notifier) {
1118         fprintf(stderr, "binding does not support host notifiers\n");
1119         r = -ENOSYS;
1120         goto fail;
1121     }
1122 
1123     for (i = 0; i < hdev->nvqs; ++i) {
1124         r = k->set_host_notifier(qbus->parent, hdev->vq_index + i, true);
1125         if (r < 0) {
1126             fprintf(stderr, "vhost VQ %d notifier binding failed: %d\n", i, -r);
1127             goto fail_vq;
1128         }
1129     }
1130 
1131     return 0;
1132 fail_vq:
1133     while (--i >= 0) {
1134         e = k->set_host_notifier(qbus->parent, hdev->vq_index + i, false);
1135         if (e < 0) {
1136             fprintf(stderr, "vhost VQ %d notifier cleanup error: %d\n", i, -r);
1137             fflush(stderr);
1138         }
1139         assert (e >= 0);
1140     }
1141 fail:
1142     return r;
1143 }
1144 
1145 /* Stop processing guest IO notifications in vhost.
1146  * Start processing them in qemu.
1147  * This might actually run the qemu handlers right away,
1148  * so virtio in qemu must be completely setup when this is called.
1149  */
1150 void vhost_dev_disable_notifiers(struct vhost_dev *hdev, VirtIODevice *vdev)
1151 {
1152     BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev)));
1153     VirtioBusState *vbus = VIRTIO_BUS(qbus);
1154     VirtioBusClass *k = VIRTIO_BUS_GET_CLASS(vbus);
1155     int i, r;
1156 
1157     for (i = 0; i < hdev->nvqs; ++i) {
1158         r = k->set_host_notifier(qbus->parent, hdev->vq_index + i, false);
1159         if (r < 0) {
1160             fprintf(stderr, "vhost VQ %d notifier cleanup failed: %d\n", i, -r);
1161             fflush(stderr);
1162         }
1163         assert (r >= 0);
1164     }
1165 }
1166 
1167 /* Test and clear event pending status.
1168  * Should be called after unmask to avoid losing events.
1169  */
1170 bool vhost_virtqueue_pending(struct vhost_dev *hdev, int n)
1171 {
1172     struct vhost_virtqueue *vq = hdev->vqs + n - hdev->vq_index;
1173     assert(n >= hdev->vq_index && n < hdev->vq_index + hdev->nvqs);
1174     return event_notifier_test_and_clear(&vq->masked_notifier);
1175 }
1176 
1177 /* Mask/unmask events from this vq. */
1178 void vhost_virtqueue_mask(struct vhost_dev *hdev, VirtIODevice *vdev, int n,
1179                          bool mask)
1180 {
1181     struct VirtQueue *vvq = virtio_get_queue(vdev, n);
1182     int r, index = n - hdev->vq_index;
1183     struct vhost_vring_file file;
1184 
1185     if (mask) {
1186         assert(vdev->use_guest_notifier_mask);
1187         file.fd = event_notifier_get_fd(&hdev->vqs[index].masked_notifier);
1188     } else {
1189         file.fd = event_notifier_get_fd(virtio_queue_get_guest_notifier(vvq));
1190     }
1191 
1192     file.index = hdev->vhost_ops->vhost_get_vq_index(hdev, n);
1193     r = hdev->vhost_ops->vhost_set_vring_call(hdev, &file);
1194     assert(r >= 0);
1195 }
1196 
1197 uint64_t vhost_get_features(struct vhost_dev *hdev, const int *feature_bits,
1198                             uint64_t features)
1199 {
1200     const int *bit = feature_bits;
1201     while (*bit != VHOST_INVALID_FEATURE_BIT) {
1202         uint64_t bit_mask = (1ULL << *bit);
1203         if (!(hdev->features & bit_mask)) {
1204             features &= ~bit_mask;
1205         }
1206         bit++;
1207     }
1208     return features;
1209 }
1210 
1211 void vhost_ack_features(struct vhost_dev *hdev, const int *feature_bits,
1212                         uint64_t features)
1213 {
1214     const int *bit = feature_bits;
1215     while (*bit != VHOST_INVALID_FEATURE_BIT) {
1216         uint64_t bit_mask = (1ULL << *bit);
1217         if (features & bit_mask) {
1218             hdev->acked_features |= bit_mask;
1219         }
1220         bit++;
1221     }
1222 }
1223 
1224 /* Host notifiers must be enabled at this point. */
1225 int vhost_dev_start(struct vhost_dev *hdev, VirtIODevice *vdev)
1226 {
1227     int i, r;
1228 
1229     hdev->started = true;
1230 
1231     r = vhost_dev_set_features(hdev, hdev->log_enabled);
1232     if (r < 0) {
1233         goto fail_features;
1234     }
1235     r = hdev->vhost_ops->vhost_set_mem_table(hdev, hdev->mem);
1236     if (r < 0) {
1237         r = -errno;
1238         goto fail_mem;
1239     }
1240     for (i = 0; i < hdev->nvqs; ++i) {
1241         r = vhost_virtqueue_start(hdev,
1242                                   vdev,
1243                                   hdev->vqs + i,
1244                                   hdev->vq_index + i);
1245         if (r < 0) {
1246             goto fail_vq;
1247         }
1248     }
1249 
1250     if (hdev->log_enabled) {
1251         uint64_t log_base;
1252 
1253         hdev->log_size = vhost_get_log_size(hdev);
1254         hdev->log = vhost_log_get(hdev->log_size,
1255                                   vhost_dev_log_is_shared(hdev));
1256         log_base = (uintptr_t)hdev->log->log;
1257         r = hdev->vhost_ops->vhost_set_log_base(hdev,
1258                                                 hdev->log_size ? log_base : 0,
1259                                                 hdev->log);
1260         if (r < 0) {
1261             r = -errno;
1262             goto fail_log;
1263         }
1264     }
1265 
1266     return 0;
1267 fail_log:
1268     vhost_log_put(hdev, false);
1269 fail_vq:
1270     while (--i >= 0) {
1271         vhost_virtqueue_stop(hdev,
1272                              vdev,
1273                              hdev->vqs + i,
1274                              hdev->vq_index + i);
1275     }
1276     i = hdev->nvqs;
1277 fail_mem:
1278 fail_features:
1279 
1280     hdev->started = false;
1281     return r;
1282 }
1283 
1284 /* Host notifiers must be enabled at this point. */
1285 void vhost_dev_stop(struct vhost_dev *hdev, VirtIODevice *vdev)
1286 {
1287     int i;
1288 
1289     for (i = 0; i < hdev->nvqs; ++i) {
1290         vhost_virtqueue_stop(hdev,
1291                              vdev,
1292                              hdev->vqs + i,
1293                              hdev->vq_index + i);
1294     }
1295 
1296     vhost_log_put(hdev, true);
1297     hdev->started = false;
1298     hdev->log = NULL;
1299     hdev->log_size = 0;
1300 }
1301 
1302