xref: /openbmc/qemu/hw/virtio/vhost.c (revision 2d7fedeb)
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 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_init(struct vhost_dev *dev,
964                                 struct vhost_virtqueue *vq, int n)
965 {
966     int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, n);
967     struct vhost_vring_file file = {
968         .index = vhost_vq_index,
969     };
970     int r = event_notifier_init(&vq->masked_notifier, 0);
971     if (r < 0) {
972         return r;
973     }
974 
975     file.fd = event_notifier_get_fd(&vq->masked_notifier);
976     r = dev->vhost_ops->vhost_set_vring_call(dev, &file);
977     if (r) {
978         r = -errno;
979         goto fail_call;
980     }
981     return 0;
982 fail_call:
983     event_notifier_cleanup(&vq->masked_notifier);
984     return r;
985 }
986 
987 static void vhost_virtqueue_cleanup(struct vhost_virtqueue *vq)
988 {
989     event_notifier_cleanup(&vq->masked_notifier);
990 }
991 
992 int vhost_dev_init(struct vhost_dev *hdev, void *opaque,
993                    VhostBackendType backend_type)
994 {
995     uint64_t features;
996     int i, r;
997 
998     hdev->migration_blocker = NULL;
999 
1000     if (vhost_set_backend_type(hdev, backend_type) < 0) {
1001         close((uintptr_t)opaque);
1002         return -1;
1003     }
1004 
1005     if (hdev->vhost_ops->vhost_backend_init(hdev, opaque) < 0) {
1006         close((uintptr_t)opaque);
1007         return -errno;
1008     }
1009 
1010     if (used_memslots > hdev->vhost_ops->vhost_backend_memslots_limit(hdev)) {
1011         fprintf(stderr, "vhost backend memory slots limit is less"
1012                 " than current number of present memory slots\n");
1013         close((uintptr_t)opaque);
1014         return -1;
1015     }
1016     QLIST_INSERT_HEAD(&vhost_devices, hdev, entry);
1017 
1018     r = hdev->vhost_ops->vhost_set_owner(hdev);
1019     if (r < 0) {
1020         goto fail;
1021     }
1022 
1023     r = hdev->vhost_ops->vhost_get_features(hdev, &features);
1024     if (r < 0) {
1025         goto fail;
1026     }
1027 
1028     for (i = 0; i < hdev->nvqs; ++i) {
1029         r = vhost_virtqueue_init(hdev, hdev->vqs + i, hdev->vq_index + i);
1030         if (r < 0) {
1031             goto fail_vq;
1032         }
1033     }
1034     hdev->features = features;
1035 
1036     hdev->memory_listener = (MemoryListener) {
1037         .begin = vhost_begin,
1038         .commit = vhost_commit,
1039         .region_add = vhost_region_add,
1040         .region_del = vhost_region_del,
1041         .region_nop = vhost_region_nop,
1042         .log_start = vhost_log_start,
1043         .log_stop = vhost_log_stop,
1044         .log_sync = vhost_log_sync,
1045         .log_global_start = vhost_log_global_start,
1046         .log_global_stop = vhost_log_global_stop,
1047         .eventfd_add = vhost_eventfd_add,
1048         .eventfd_del = vhost_eventfd_del,
1049         .priority = 10
1050     };
1051 
1052     if (hdev->migration_blocker == NULL) {
1053         if (!(hdev->features & (0x1ULL << VHOST_F_LOG_ALL))) {
1054             error_setg(&hdev->migration_blocker,
1055                        "Migration disabled: vhost lacks VHOST_F_LOG_ALL feature.");
1056         } else if (!qemu_memfd_check()) {
1057             error_setg(&hdev->migration_blocker,
1058                        "Migration disabled: failed to allocate shared memory");
1059         }
1060     }
1061 
1062     if (hdev->migration_blocker != NULL) {
1063         migrate_add_blocker(hdev->migration_blocker);
1064     }
1065 
1066     hdev->mem = g_malloc0(offsetof(struct vhost_memory, regions));
1067     hdev->n_mem_sections = 0;
1068     hdev->mem_sections = NULL;
1069     hdev->log = NULL;
1070     hdev->log_size = 0;
1071     hdev->log_enabled = false;
1072     hdev->started = false;
1073     hdev->memory_changed = false;
1074     memory_listener_register(&hdev->memory_listener, &address_space_memory);
1075     return 0;
1076 fail_vq:
1077     while (--i >= 0) {
1078         vhost_virtqueue_cleanup(hdev->vqs + i);
1079     }
1080 fail:
1081     r = -errno;
1082     hdev->vhost_ops->vhost_backend_cleanup(hdev);
1083     QLIST_REMOVE(hdev, entry);
1084     return r;
1085 }
1086 
1087 void vhost_dev_cleanup(struct vhost_dev *hdev)
1088 {
1089     int i;
1090     for (i = 0; i < hdev->nvqs; ++i) {
1091         vhost_virtqueue_cleanup(hdev->vqs + i);
1092     }
1093     memory_listener_unregister(&hdev->memory_listener);
1094     if (hdev->migration_blocker) {
1095         migrate_del_blocker(hdev->migration_blocker);
1096         error_free(hdev->migration_blocker);
1097     }
1098     g_free(hdev->mem);
1099     g_free(hdev->mem_sections);
1100     hdev->vhost_ops->vhost_backend_cleanup(hdev);
1101     QLIST_REMOVE(hdev, entry);
1102 }
1103 
1104 /* Stop processing guest IO notifications in qemu.
1105  * Start processing them in vhost in kernel.
1106  */
1107 int vhost_dev_enable_notifiers(struct vhost_dev *hdev, VirtIODevice *vdev)
1108 {
1109     BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev)));
1110     VirtioBusState *vbus = VIRTIO_BUS(qbus);
1111     VirtioBusClass *k = VIRTIO_BUS_GET_CLASS(vbus);
1112     int i, r, e;
1113     if (!k->set_host_notifier) {
1114         fprintf(stderr, "binding does not support host notifiers\n");
1115         r = -ENOSYS;
1116         goto fail;
1117     }
1118 
1119     for (i = 0; i < hdev->nvqs; ++i) {
1120         r = k->set_host_notifier(qbus->parent, hdev->vq_index + i, true);
1121         if (r < 0) {
1122             fprintf(stderr, "vhost VQ %d notifier binding failed: %d\n", i, -r);
1123             goto fail_vq;
1124         }
1125     }
1126 
1127     return 0;
1128 fail_vq:
1129     while (--i >= 0) {
1130         e = k->set_host_notifier(qbus->parent, hdev->vq_index + i, false);
1131         if (e < 0) {
1132             fprintf(stderr, "vhost VQ %d notifier cleanup error: %d\n", i, -r);
1133             fflush(stderr);
1134         }
1135         assert (e >= 0);
1136     }
1137 fail:
1138     return r;
1139 }
1140 
1141 /* Stop processing guest IO notifications in vhost.
1142  * Start processing them in qemu.
1143  * This might actually run the qemu handlers right away,
1144  * so virtio in qemu must be completely setup when this is called.
1145  */
1146 void vhost_dev_disable_notifiers(struct vhost_dev *hdev, VirtIODevice *vdev)
1147 {
1148     BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev)));
1149     VirtioBusState *vbus = VIRTIO_BUS(qbus);
1150     VirtioBusClass *k = VIRTIO_BUS_GET_CLASS(vbus);
1151     int i, r;
1152 
1153     for (i = 0; i < hdev->nvqs; ++i) {
1154         r = k->set_host_notifier(qbus->parent, hdev->vq_index + i, false);
1155         if (r < 0) {
1156             fprintf(stderr, "vhost VQ %d notifier cleanup failed: %d\n", i, -r);
1157             fflush(stderr);
1158         }
1159         assert (r >= 0);
1160     }
1161 }
1162 
1163 /* Test and clear event pending status.
1164  * Should be called after unmask to avoid losing events.
1165  */
1166 bool vhost_virtqueue_pending(struct vhost_dev *hdev, int n)
1167 {
1168     struct vhost_virtqueue *vq = hdev->vqs + n - hdev->vq_index;
1169     assert(n >= hdev->vq_index && n < hdev->vq_index + hdev->nvqs);
1170     return event_notifier_test_and_clear(&vq->masked_notifier);
1171 }
1172 
1173 /* Mask/unmask events from this vq. */
1174 void vhost_virtqueue_mask(struct vhost_dev *hdev, VirtIODevice *vdev, int n,
1175                          bool mask)
1176 {
1177     struct VirtQueue *vvq = virtio_get_queue(vdev, n);
1178     int r, index = n - hdev->vq_index;
1179     struct vhost_vring_file file;
1180 
1181     if (mask) {
1182         assert(vdev->use_guest_notifier_mask);
1183         file.fd = event_notifier_get_fd(&hdev->vqs[index].masked_notifier);
1184     } else {
1185         file.fd = event_notifier_get_fd(virtio_queue_get_guest_notifier(vvq));
1186     }
1187 
1188     file.index = hdev->vhost_ops->vhost_get_vq_index(hdev, n);
1189     r = hdev->vhost_ops->vhost_set_vring_call(hdev, &file);
1190     assert(r >= 0);
1191 }
1192 
1193 uint64_t vhost_get_features(struct vhost_dev *hdev, const int *feature_bits,
1194                             uint64_t features)
1195 {
1196     const int *bit = feature_bits;
1197     while (*bit != VHOST_INVALID_FEATURE_BIT) {
1198         uint64_t bit_mask = (1ULL << *bit);
1199         if (!(hdev->features & bit_mask)) {
1200             features &= ~bit_mask;
1201         }
1202         bit++;
1203     }
1204     return features;
1205 }
1206 
1207 void vhost_ack_features(struct vhost_dev *hdev, const int *feature_bits,
1208                         uint64_t features)
1209 {
1210     const int *bit = feature_bits;
1211     while (*bit != VHOST_INVALID_FEATURE_BIT) {
1212         uint64_t bit_mask = (1ULL << *bit);
1213         if (features & bit_mask) {
1214             hdev->acked_features |= bit_mask;
1215         }
1216         bit++;
1217     }
1218 }
1219 
1220 /* Host notifiers must be enabled at this point. */
1221 int vhost_dev_start(struct vhost_dev *hdev, VirtIODevice *vdev)
1222 {
1223     int i, r;
1224 
1225     hdev->started = true;
1226 
1227     r = vhost_dev_set_features(hdev, hdev->log_enabled);
1228     if (r < 0) {
1229         goto fail_features;
1230     }
1231     r = hdev->vhost_ops->vhost_set_mem_table(hdev, hdev->mem);
1232     if (r < 0) {
1233         r = -errno;
1234         goto fail_mem;
1235     }
1236     for (i = 0; i < hdev->nvqs; ++i) {
1237         r = vhost_virtqueue_start(hdev,
1238                                   vdev,
1239                                   hdev->vqs + i,
1240                                   hdev->vq_index + i);
1241         if (r < 0) {
1242             goto fail_vq;
1243         }
1244     }
1245 
1246     if (hdev->log_enabled) {
1247         uint64_t log_base;
1248 
1249         hdev->log_size = vhost_get_log_size(hdev);
1250         hdev->log = vhost_log_get(hdev->log_size,
1251                                   vhost_dev_log_is_shared(hdev));
1252         log_base = (uintptr_t)hdev->log->log;
1253         r = hdev->vhost_ops->vhost_set_log_base(hdev,
1254                                                 hdev->log_size ? log_base : 0,
1255                                                 hdev->log);
1256         if (r < 0) {
1257             r = -errno;
1258             goto fail_log;
1259         }
1260     }
1261 
1262     return 0;
1263 fail_log:
1264     vhost_log_put(hdev, false);
1265 fail_vq:
1266     while (--i >= 0) {
1267         vhost_virtqueue_stop(hdev,
1268                              vdev,
1269                              hdev->vqs + i,
1270                              hdev->vq_index + i);
1271     }
1272     i = hdev->nvqs;
1273 fail_mem:
1274 fail_features:
1275 
1276     hdev->started = false;
1277     return r;
1278 }
1279 
1280 /* Host notifiers must be enabled at this point. */
1281 void vhost_dev_stop(struct vhost_dev *hdev, VirtIODevice *vdev)
1282 {
1283     int i;
1284 
1285     for (i = 0; i < hdev->nvqs; ++i) {
1286         vhost_virtqueue_stop(hdev,
1287                              vdev,
1288                              hdev->vqs + i,
1289                              hdev->vq_index + i);
1290     }
1291 
1292     vhost_log_put(hdev, true);
1293     hdev->started = false;
1294     hdev->log = NULL;
1295     hdev->log_size = 0;
1296 }
1297 
1298