xref: /openbmc/qemu/hw/virtio/vhost.c (revision fc48ffc3)
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 <linux/vhost.h>
22 #include "exec/address-spaces.h"
23 #include "hw/virtio/virtio-bus.h"
24 #include "hw/virtio/virtio-access.h"
25 #include "migration/migration.h"
26 
27 static struct vhost_log *vhost_log;
28 
29 static void vhost_dev_sync_region(struct vhost_dev *dev,
30                                   MemoryRegionSection *section,
31                                   uint64_t mfirst, uint64_t mlast,
32                                   uint64_t rfirst, uint64_t rlast)
33 {
34     vhost_log_chunk_t *log = dev->log->log;
35 
36     uint64_t start = MAX(mfirst, rfirst);
37     uint64_t end = MIN(mlast, rlast);
38     vhost_log_chunk_t *from = log + start / VHOST_LOG_CHUNK;
39     vhost_log_chunk_t *to = log + end / VHOST_LOG_CHUNK + 1;
40     uint64_t addr = (start / VHOST_LOG_CHUNK) * VHOST_LOG_CHUNK;
41 
42     if (end < start) {
43         return;
44     }
45     assert(end / VHOST_LOG_CHUNK < dev->log_size);
46     assert(start / VHOST_LOG_CHUNK < dev->log_size);
47 
48     for (;from < to; ++from) {
49         vhost_log_chunk_t log;
50         /* We first check with non-atomic: much cheaper,
51          * and we expect non-dirty to be the common case. */
52         if (!*from) {
53             addr += VHOST_LOG_CHUNK;
54             continue;
55         }
56         /* Data must be read atomically. We don't really need barrier semantics
57          * but it's easier to use atomic_* than roll our own. */
58         log = atomic_xchg(from, 0);
59         while (log) {
60             int bit = ctzl(log);
61             hwaddr page_addr;
62             hwaddr section_offset;
63             hwaddr mr_offset;
64             page_addr = addr + bit * VHOST_LOG_PAGE;
65             section_offset = page_addr - section->offset_within_address_space;
66             mr_offset = section_offset + section->offset_within_region;
67             memory_region_set_dirty(section->mr, mr_offset, VHOST_LOG_PAGE);
68             log &= ~(0x1ull << bit);
69         }
70         addr += VHOST_LOG_CHUNK;
71     }
72 }
73 
74 static int vhost_sync_dirty_bitmap(struct vhost_dev *dev,
75                                    MemoryRegionSection *section,
76                                    hwaddr first,
77                                    hwaddr last)
78 {
79     int i;
80     hwaddr start_addr;
81     hwaddr end_addr;
82 
83     if (!dev->log_enabled || !dev->started) {
84         return 0;
85     }
86     start_addr = section->offset_within_address_space;
87     end_addr = range_get_last(start_addr, int128_get64(section->size));
88     start_addr = MAX(first, start_addr);
89     end_addr = MIN(last, end_addr);
90 
91     for (i = 0; i < dev->mem->nregions; ++i) {
92         struct vhost_memory_region *reg = dev->mem->regions + i;
93         vhost_dev_sync_region(dev, section, start_addr, end_addr,
94                               reg->guest_phys_addr,
95                               range_get_last(reg->guest_phys_addr,
96                                              reg->memory_size));
97     }
98     for (i = 0; i < dev->nvqs; ++i) {
99         struct vhost_virtqueue *vq = dev->vqs + i;
100         vhost_dev_sync_region(dev, section, start_addr, end_addr, vq->used_phys,
101                               range_get_last(vq->used_phys, vq->used_size));
102     }
103     return 0;
104 }
105 
106 static void vhost_log_sync(MemoryListener *listener,
107                           MemoryRegionSection *section)
108 {
109     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
110                                          memory_listener);
111     vhost_sync_dirty_bitmap(dev, section, 0x0, ~0x0ULL);
112 }
113 
114 static void vhost_log_sync_range(struct vhost_dev *dev,
115                                  hwaddr first, hwaddr last)
116 {
117     int i;
118     /* FIXME: this is N^2 in number of sections */
119     for (i = 0; i < dev->n_mem_sections; ++i) {
120         MemoryRegionSection *section = &dev->mem_sections[i];
121         vhost_sync_dirty_bitmap(dev, section, first, last);
122     }
123 }
124 
125 /* Assign/unassign. Keep an unsorted array of non-overlapping
126  * memory regions in dev->mem. */
127 static void vhost_dev_unassign_memory(struct vhost_dev *dev,
128                                       uint64_t start_addr,
129                                       uint64_t size)
130 {
131     int from, to, n = dev->mem->nregions;
132     /* Track overlapping/split regions for sanity checking. */
133     int overlap_start = 0, overlap_end = 0, overlap_middle = 0, split = 0;
134 
135     for (from = 0, to = 0; from < n; ++from, ++to) {
136         struct vhost_memory_region *reg = dev->mem->regions + to;
137         uint64_t reglast;
138         uint64_t memlast;
139         uint64_t change;
140 
141         /* clone old region */
142         if (to != from) {
143             memcpy(reg, dev->mem->regions + from, sizeof *reg);
144         }
145 
146         /* No overlap is simple */
147         if (!ranges_overlap(reg->guest_phys_addr, reg->memory_size,
148                             start_addr, size)) {
149             continue;
150         }
151 
152         /* Split only happens if supplied region
153          * is in the middle of an existing one. Thus it can not
154          * overlap with any other existing region. */
155         assert(!split);
156 
157         reglast = range_get_last(reg->guest_phys_addr, reg->memory_size);
158         memlast = range_get_last(start_addr, size);
159 
160         /* Remove whole region */
161         if (start_addr <= reg->guest_phys_addr && memlast >= reglast) {
162             --dev->mem->nregions;
163             --to;
164             ++overlap_middle;
165             continue;
166         }
167 
168         /* Shrink region */
169         if (memlast >= reglast) {
170             reg->memory_size = start_addr - reg->guest_phys_addr;
171             assert(reg->memory_size);
172             assert(!overlap_end);
173             ++overlap_end;
174             continue;
175         }
176 
177         /* Shift region */
178         if (start_addr <= reg->guest_phys_addr) {
179             change = memlast + 1 - reg->guest_phys_addr;
180             reg->memory_size -= change;
181             reg->guest_phys_addr += change;
182             reg->userspace_addr += change;
183             assert(reg->memory_size);
184             assert(!overlap_start);
185             ++overlap_start;
186             continue;
187         }
188 
189         /* This only happens if supplied region
190          * is in the middle of an existing one. Thus it can not
191          * overlap with any other existing region. */
192         assert(!overlap_start);
193         assert(!overlap_end);
194         assert(!overlap_middle);
195         /* Split region: shrink first part, shift second part. */
196         memcpy(dev->mem->regions + n, reg, sizeof *reg);
197         reg->memory_size = start_addr - reg->guest_phys_addr;
198         assert(reg->memory_size);
199         change = memlast + 1 - reg->guest_phys_addr;
200         reg = dev->mem->regions + n;
201         reg->memory_size -= change;
202         assert(reg->memory_size);
203         reg->guest_phys_addr += change;
204         reg->userspace_addr += change;
205         /* Never add more than 1 region */
206         assert(dev->mem->nregions == n);
207         ++dev->mem->nregions;
208         ++split;
209     }
210 }
211 
212 /* Called after unassign, so no regions overlap the given range. */
213 static void vhost_dev_assign_memory(struct vhost_dev *dev,
214                                     uint64_t start_addr,
215                                     uint64_t size,
216                                     uint64_t uaddr)
217 {
218     int from, to;
219     struct vhost_memory_region *merged = NULL;
220     for (from = 0, to = 0; from < dev->mem->nregions; ++from, ++to) {
221         struct vhost_memory_region *reg = dev->mem->regions + to;
222         uint64_t prlast, urlast;
223         uint64_t pmlast, umlast;
224         uint64_t s, e, u;
225 
226         /* clone old region */
227         if (to != from) {
228             memcpy(reg, dev->mem->regions + from, sizeof *reg);
229         }
230         prlast = range_get_last(reg->guest_phys_addr, reg->memory_size);
231         pmlast = range_get_last(start_addr, size);
232         urlast = range_get_last(reg->userspace_addr, reg->memory_size);
233         umlast = range_get_last(uaddr, size);
234 
235         /* check for overlapping regions: should never happen. */
236         assert(prlast < start_addr || pmlast < reg->guest_phys_addr);
237         /* Not an adjacent or overlapping region - do not merge. */
238         if ((prlast + 1 != start_addr || urlast + 1 != uaddr) &&
239             (pmlast + 1 != reg->guest_phys_addr ||
240              umlast + 1 != reg->userspace_addr)) {
241             continue;
242         }
243 
244         if (merged) {
245             --to;
246             assert(to >= 0);
247         } else {
248             merged = reg;
249         }
250         u = MIN(uaddr, reg->userspace_addr);
251         s = MIN(start_addr, reg->guest_phys_addr);
252         e = MAX(pmlast, prlast);
253         uaddr = merged->userspace_addr = u;
254         start_addr = merged->guest_phys_addr = s;
255         size = merged->memory_size = e - s + 1;
256         assert(merged->memory_size);
257     }
258 
259     if (!merged) {
260         struct vhost_memory_region *reg = dev->mem->regions + to;
261         memset(reg, 0, sizeof *reg);
262         reg->memory_size = size;
263         assert(reg->memory_size);
264         reg->guest_phys_addr = start_addr;
265         reg->userspace_addr = uaddr;
266         ++to;
267     }
268     assert(to <= dev->mem->nregions + 1);
269     dev->mem->nregions = to;
270 }
271 
272 static uint64_t vhost_get_log_size(struct vhost_dev *dev)
273 {
274     uint64_t log_size = 0;
275     int i;
276     for (i = 0; i < dev->mem->nregions; ++i) {
277         struct vhost_memory_region *reg = dev->mem->regions + i;
278         uint64_t last = range_get_last(reg->guest_phys_addr,
279                                        reg->memory_size);
280         log_size = MAX(log_size, last / VHOST_LOG_CHUNK + 1);
281     }
282     for (i = 0; i < dev->nvqs; ++i) {
283         struct vhost_virtqueue *vq = dev->vqs + i;
284         uint64_t last = vq->used_phys + vq->used_size - 1;
285         log_size = MAX(log_size, last / VHOST_LOG_CHUNK + 1);
286     }
287     return log_size;
288 }
289 static struct vhost_log *vhost_log_alloc(uint64_t size)
290 {
291     struct vhost_log *log = g_malloc0(sizeof *log + size * sizeof(*(log->log)));
292 
293     log->size = size;
294     log->refcnt = 1;
295 
296     return log;
297 }
298 
299 static struct vhost_log *vhost_log_get(uint64_t size)
300 {
301     if (!vhost_log || vhost_log->size != size) {
302         vhost_log = vhost_log_alloc(size);
303     } else {
304         ++vhost_log->refcnt;
305     }
306 
307     return vhost_log;
308 }
309 
310 static void vhost_log_put(struct vhost_dev *dev, bool sync)
311 {
312     struct vhost_log *log = dev->log;
313 
314     if (!log) {
315         return;
316     }
317 
318     --log->refcnt;
319     if (log->refcnt == 0) {
320         /* Sync only the range covered by the old log */
321         if (dev->log_size && sync) {
322             vhost_log_sync_range(dev, 0, dev->log_size * VHOST_LOG_CHUNK - 1);
323         }
324         if (vhost_log == log) {
325             vhost_log = NULL;
326         }
327         g_free(log);
328     }
329 }
330 
331 static inline void vhost_dev_log_resize(struct vhost_dev* dev, uint64_t size)
332 {
333     struct vhost_log *log = vhost_log_get(size);
334     uint64_t log_base = (uintptr_t)log->log;
335     int r;
336 
337     r = dev->vhost_ops->vhost_call(dev, VHOST_SET_LOG_BASE, &log_base);
338     assert(r >= 0);
339     vhost_log_put(dev, true);
340     dev->log = log;
341     dev->log_size = size;
342 }
343 
344 static int vhost_verify_ring_mappings(struct vhost_dev *dev,
345                                       uint64_t start_addr,
346                                       uint64_t size)
347 {
348     int i;
349     int r = 0;
350 
351     for (i = 0; !r && i < dev->nvqs; ++i) {
352         struct vhost_virtqueue *vq = dev->vqs + i;
353         hwaddr l;
354         void *p;
355 
356         if (!ranges_overlap(start_addr, size, vq->ring_phys, vq->ring_size)) {
357             continue;
358         }
359         l = vq->ring_size;
360         p = cpu_physical_memory_map(vq->ring_phys, &l, 1);
361         if (!p || l != vq->ring_size) {
362             fprintf(stderr, "Unable to map ring buffer for ring %d\n", i);
363             r = -ENOMEM;
364         }
365         if (p != vq->ring) {
366             fprintf(stderr, "Ring buffer relocated for ring %d\n", i);
367             r = -EBUSY;
368         }
369         cpu_physical_memory_unmap(p, l, 0, 0);
370     }
371     return r;
372 }
373 
374 static struct vhost_memory_region *vhost_dev_find_reg(struct vhost_dev *dev,
375 						      uint64_t start_addr,
376 						      uint64_t size)
377 {
378     int i, n = dev->mem->nregions;
379     for (i = 0; i < n; ++i) {
380         struct vhost_memory_region *reg = dev->mem->regions + i;
381         if (ranges_overlap(reg->guest_phys_addr, reg->memory_size,
382                            start_addr, size)) {
383             return reg;
384         }
385     }
386     return NULL;
387 }
388 
389 static bool vhost_dev_cmp_memory(struct vhost_dev *dev,
390                                  uint64_t start_addr,
391                                  uint64_t size,
392                                  uint64_t uaddr)
393 {
394     struct vhost_memory_region *reg = vhost_dev_find_reg(dev, start_addr, size);
395     uint64_t reglast;
396     uint64_t memlast;
397 
398     if (!reg) {
399         return true;
400     }
401 
402     reglast = range_get_last(reg->guest_phys_addr, reg->memory_size);
403     memlast = range_get_last(start_addr, size);
404 
405     /* Need to extend region? */
406     if (start_addr < reg->guest_phys_addr || memlast > reglast) {
407         return true;
408     }
409     /* userspace_addr changed? */
410     return uaddr != reg->userspace_addr + start_addr - reg->guest_phys_addr;
411 }
412 
413 static void vhost_set_memory(MemoryListener *listener,
414                              MemoryRegionSection *section,
415                              bool add)
416 {
417     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
418                                          memory_listener);
419     hwaddr start_addr = section->offset_within_address_space;
420     ram_addr_t size = int128_get64(section->size);
421     bool log_dirty =
422         memory_region_get_dirty_log_mask(section->mr) & ~(1 << DIRTY_MEMORY_MIGRATION);
423     int s = offsetof(struct vhost_memory, regions) +
424         (dev->mem->nregions + 1) * sizeof dev->mem->regions[0];
425     void *ram;
426 
427     dev->mem = g_realloc(dev->mem, s);
428 
429     if (log_dirty) {
430         add = false;
431     }
432 
433     assert(size);
434 
435     /* Optimize no-change case. At least cirrus_vga does this a lot at this time. */
436     ram = memory_region_get_ram_ptr(section->mr) + section->offset_within_region;
437     if (add) {
438         if (!vhost_dev_cmp_memory(dev, start_addr, size, (uintptr_t)ram)) {
439             /* Region exists with same address. Nothing to do. */
440             return;
441         }
442     } else {
443         if (!vhost_dev_find_reg(dev, start_addr, size)) {
444             /* Removing region that we don't access. Nothing to do. */
445             return;
446         }
447     }
448 
449     vhost_dev_unassign_memory(dev, start_addr, size);
450     if (add) {
451         /* Add given mapping, merging adjacent regions if any */
452         vhost_dev_assign_memory(dev, start_addr, size, (uintptr_t)ram);
453     } else {
454         /* Remove old mapping for this memory, if any. */
455         vhost_dev_unassign_memory(dev, start_addr, size);
456     }
457     dev->mem_changed_start_addr = MIN(dev->mem_changed_start_addr, start_addr);
458     dev->mem_changed_end_addr = MAX(dev->mem_changed_end_addr, start_addr + size - 1);
459     dev->memory_changed = true;
460 }
461 
462 static bool vhost_section(MemoryRegionSection *section)
463 {
464     return memory_region_is_ram(section->mr);
465 }
466 
467 static void vhost_begin(MemoryListener *listener)
468 {
469     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
470                                          memory_listener);
471     dev->mem_changed_end_addr = 0;
472     dev->mem_changed_start_addr = -1;
473 }
474 
475 static void vhost_commit(MemoryListener *listener)
476 {
477     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
478                                          memory_listener);
479     hwaddr start_addr = 0;
480     ram_addr_t size = 0;
481     uint64_t log_size;
482     int r;
483 
484     if (!dev->memory_changed) {
485         return;
486     }
487     if (!dev->started) {
488         return;
489     }
490     if (dev->mem_changed_start_addr > dev->mem_changed_end_addr) {
491         return;
492     }
493 
494     if (dev->started) {
495         start_addr = dev->mem_changed_start_addr;
496         size = dev->mem_changed_end_addr - dev->mem_changed_start_addr + 1;
497 
498         r = vhost_verify_ring_mappings(dev, start_addr, size);
499         assert(r >= 0);
500     }
501 
502     if (!dev->log_enabled) {
503         r = dev->vhost_ops->vhost_call(dev, VHOST_SET_MEM_TABLE, dev->mem);
504         assert(r >= 0);
505         dev->memory_changed = false;
506         return;
507     }
508     log_size = vhost_get_log_size(dev);
509     /* We allocate an extra 4K bytes to log,
510      * to reduce the * number of reallocations. */
511 #define VHOST_LOG_BUFFER (0x1000 / sizeof *dev->log)
512     /* To log more, must increase log size before table update. */
513     if (dev->log_size < log_size) {
514         vhost_dev_log_resize(dev, log_size + VHOST_LOG_BUFFER);
515     }
516     r = dev->vhost_ops->vhost_call(dev, VHOST_SET_MEM_TABLE, dev->mem);
517     assert(r >= 0);
518     /* To log less, can only decrease log size after table update. */
519     if (dev->log_size > log_size + VHOST_LOG_BUFFER) {
520         vhost_dev_log_resize(dev, log_size);
521     }
522     dev->memory_changed = false;
523 }
524 
525 static void vhost_region_add(MemoryListener *listener,
526                              MemoryRegionSection *section)
527 {
528     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
529                                          memory_listener);
530 
531     if (!vhost_section(section)) {
532         return;
533     }
534 
535     ++dev->n_mem_sections;
536     dev->mem_sections = g_renew(MemoryRegionSection, dev->mem_sections,
537                                 dev->n_mem_sections);
538     dev->mem_sections[dev->n_mem_sections - 1] = *section;
539     memory_region_ref(section->mr);
540     vhost_set_memory(listener, section, true);
541 }
542 
543 static void vhost_region_del(MemoryListener *listener,
544                              MemoryRegionSection *section)
545 {
546     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
547                                          memory_listener);
548     int i;
549 
550     if (!vhost_section(section)) {
551         return;
552     }
553 
554     vhost_set_memory(listener, section, false);
555     memory_region_unref(section->mr);
556     for (i = 0; i < dev->n_mem_sections; ++i) {
557         if (dev->mem_sections[i].offset_within_address_space
558             == section->offset_within_address_space) {
559             --dev->n_mem_sections;
560             memmove(&dev->mem_sections[i], &dev->mem_sections[i+1],
561                     (dev->n_mem_sections - i) * sizeof(*dev->mem_sections));
562             break;
563         }
564     }
565 }
566 
567 static void vhost_region_nop(MemoryListener *listener,
568                              MemoryRegionSection *section)
569 {
570 }
571 
572 static int vhost_virtqueue_set_addr(struct vhost_dev *dev,
573                                     struct vhost_virtqueue *vq,
574                                     unsigned idx, bool enable_log)
575 {
576     struct vhost_vring_addr addr = {
577         .index = idx,
578         .desc_user_addr = (uint64_t)(unsigned long)vq->desc,
579         .avail_user_addr = (uint64_t)(unsigned long)vq->avail,
580         .used_user_addr = (uint64_t)(unsigned long)vq->used,
581         .log_guest_addr = vq->used_phys,
582         .flags = enable_log ? (1 << VHOST_VRING_F_LOG) : 0,
583     };
584     int r = dev->vhost_ops->vhost_call(dev, VHOST_SET_VRING_ADDR, &addr);
585     if (r < 0) {
586         return -errno;
587     }
588     return 0;
589 }
590 
591 static int vhost_dev_set_features(struct vhost_dev *dev, bool enable_log)
592 {
593     uint64_t features = dev->acked_features;
594     int r;
595     if (enable_log) {
596         features |= 0x1ULL << VHOST_F_LOG_ALL;
597     }
598     r = dev->vhost_ops->vhost_call(dev, VHOST_SET_FEATURES, &features);
599     return r < 0 ? -errno : 0;
600 }
601 
602 static int vhost_dev_set_log(struct vhost_dev *dev, bool enable_log)
603 {
604     int r, t, i;
605     r = vhost_dev_set_features(dev, enable_log);
606     if (r < 0) {
607         goto err_features;
608     }
609     for (i = 0; i < dev->nvqs; ++i) {
610         r = vhost_virtqueue_set_addr(dev, dev->vqs + i, i,
611                                      enable_log);
612         if (r < 0) {
613             goto err_vq;
614         }
615     }
616     return 0;
617 err_vq:
618     for (; i >= 0; --i) {
619         t = vhost_virtqueue_set_addr(dev, dev->vqs + i, i,
620                                      dev->log_enabled);
621         assert(t >= 0);
622     }
623     t = vhost_dev_set_features(dev, dev->log_enabled);
624     assert(t >= 0);
625 err_features:
626     return r;
627 }
628 
629 static int vhost_migration_log(MemoryListener *listener, int enable)
630 {
631     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
632                                          memory_listener);
633     int r;
634     if (!!enable == dev->log_enabled) {
635         return 0;
636     }
637     if (!dev->started) {
638         dev->log_enabled = enable;
639         return 0;
640     }
641     if (!enable) {
642         r = vhost_dev_set_log(dev, false);
643         if (r < 0) {
644             return r;
645         }
646         vhost_log_put(dev, false);
647         dev->log = NULL;
648         dev->log_size = 0;
649     } else {
650         vhost_dev_log_resize(dev, vhost_get_log_size(dev));
651         r = vhost_dev_set_log(dev, true);
652         if (r < 0) {
653             return r;
654         }
655     }
656     dev->log_enabled = enable;
657     return 0;
658 }
659 
660 static void vhost_log_global_start(MemoryListener *listener)
661 {
662     int r;
663 
664     r = vhost_migration_log(listener, true);
665     if (r < 0) {
666         abort();
667     }
668 }
669 
670 static void vhost_log_global_stop(MemoryListener *listener)
671 {
672     int r;
673 
674     r = vhost_migration_log(listener, false);
675     if (r < 0) {
676         abort();
677     }
678 }
679 
680 static void vhost_log_start(MemoryListener *listener,
681                             MemoryRegionSection *section,
682                             int old, int new)
683 {
684     /* FIXME: implement */
685 }
686 
687 static void vhost_log_stop(MemoryListener *listener,
688                            MemoryRegionSection *section,
689                            int old, int new)
690 {
691     /* FIXME: implement */
692 }
693 
694 static int vhost_virtqueue_set_vring_endian_legacy(struct vhost_dev *dev,
695                                                    bool is_big_endian,
696                                                    int vhost_vq_index)
697 {
698     struct vhost_vring_state s = {
699         .index = vhost_vq_index,
700         .num = is_big_endian
701     };
702 
703     if (!dev->vhost_ops->vhost_call(dev, VHOST_SET_VRING_ENDIAN, &s)) {
704         return 0;
705     }
706 
707     if (errno == ENOTTY) {
708         error_report("vhost does not support cross-endian");
709         return -ENOSYS;
710     }
711 
712     return -errno;
713 }
714 
715 static int vhost_virtqueue_start(struct vhost_dev *dev,
716                                 struct VirtIODevice *vdev,
717                                 struct vhost_virtqueue *vq,
718                                 unsigned idx)
719 {
720     hwaddr s, l, a;
721     int r;
722     int vhost_vq_index = idx - dev->vq_index;
723     struct vhost_vring_file file = {
724         .index = vhost_vq_index
725     };
726     struct vhost_vring_state state = {
727         .index = vhost_vq_index
728     };
729     struct VirtQueue *vvq = virtio_get_queue(vdev, idx);
730 
731     assert(idx >= dev->vq_index && idx < dev->vq_index + dev->nvqs);
732 
733     vq->num = state.num = virtio_queue_get_num(vdev, idx);
734     r = dev->vhost_ops->vhost_call(dev, VHOST_SET_VRING_NUM, &state);
735     if (r) {
736         return -errno;
737     }
738 
739     state.num = virtio_queue_get_last_avail_idx(vdev, idx);
740     r = dev->vhost_ops->vhost_call(dev, VHOST_SET_VRING_BASE, &state);
741     if (r) {
742         return -errno;
743     }
744 
745     if (!virtio_has_feature(vdev, VIRTIO_F_VERSION_1) &&
746         virtio_legacy_is_cross_endian(vdev)) {
747         r = vhost_virtqueue_set_vring_endian_legacy(dev,
748                                                     virtio_is_big_endian(vdev),
749                                                     vhost_vq_index);
750         if (r) {
751             return -errno;
752         }
753     }
754 
755     s = l = virtio_queue_get_desc_size(vdev, idx);
756     a = virtio_queue_get_desc_addr(vdev, idx);
757     vq->desc = cpu_physical_memory_map(a, &l, 0);
758     if (!vq->desc || l != s) {
759         r = -ENOMEM;
760         goto fail_alloc_desc;
761     }
762     s = l = virtio_queue_get_avail_size(vdev, idx);
763     a = virtio_queue_get_avail_addr(vdev, idx);
764     vq->avail = cpu_physical_memory_map(a, &l, 0);
765     if (!vq->avail || l != s) {
766         r = -ENOMEM;
767         goto fail_alloc_avail;
768     }
769     vq->used_size = s = l = virtio_queue_get_used_size(vdev, idx);
770     vq->used_phys = a = virtio_queue_get_used_addr(vdev, idx);
771     vq->used = cpu_physical_memory_map(a, &l, 1);
772     if (!vq->used || l != s) {
773         r = -ENOMEM;
774         goto fail_alloc_used;
775     }
776 
777     vq->ring_size = s = l = virtio_queue_get_ring_size(vdev, idx);
778     vq->ring_phys = a = virtio_queue_get_ring_addr(vdev, idx);
779     vq->ring = cpu_physical_memory_map(a, &l, 1);
780     if (!vq->ring || l != s) {
781         r = -ENOMEM;
782         goto fail_alloc_ring;
783     }
784 
785     r = vhost_virtqueue_set_addr(dev, vq, vhost_vq_index, dev->log_enabled);
786     if (r < 0) {
787         r = -errno;
788         goto fail_alloc;
789     }
790 
791     file.fd = event_notifier_get_fd(virtio_queue_get_host_notifier(vvq));
792     r = dev->vhost_ops->vhost_call(dev, VHOST_SET_VRING_KICK, &file);
793     if (r) {
794         r = -errno;
795         goto fail_kick;
796     }
797 
798     /* Clear and discard previous events if any. */
799     event_notifier_test_and_clear(&vq->masked_notifier);
800 
801     return 0;
802 
803 fail_kick:
804 fail_alloc:
805     cpu_physical_memory_unmap(vq->ring, virtio_queue_get_ring_size(vdev, idx),
806                               0, 0);
807 fail_alloc_ring:
808     cpu_physical_memory_unmap(vq->used, virtio_queue_get_used_size(vdev, idx),
809                               0, 0);
810 fail_alloc_used:
811     cpu_physical_memory_unmap(vq->avail, virtio_queue_get_avail_size(vdev, idx),
812                               0, 0);
813 fail_alloc_avail:
814     cpu_physical_memory_unmap(vq->desc, virtio_queue_get_desc_size(vdev, idx),
815                               0, 0);
816 fail_alloc_desc:
817     return r;
818 }
819 
820 static void vhost_virtqueue_stop(struct vhost_dev *dev,
821                                     struct VirtIODevice *vdev,
822                                     struct vhost_virtqueue *vq,
823                                     unsigned idx)
824 {
825     int vhost_vq_index = idx - dev->vq_index;
826     struct vhost_vring_state state = {
827         .index = vhost_vq_index,
828     };
829     int r;
830     assert(idx >= dev->vq_index && idx < dev->vq_index + dev->nvqs);
831     r = dev->vhost_ops->vhost_call(dev, VHOST_GET_VRING_BASE, &state);
832     if (r < 0) {
833         fprintf(stderr, "vhost VQ %d ring restore failed: %d\n", idx, r);
834         fflush(stderr);
835     }
836     virtio_queue_set_last_avail_idx(vdev, idx, state.num);
837     virtio_queue_invalidate_signalled_used(vdev, idx);
838 
839     /* In the cross-endian case, we need to reset the vring endianness to
840      * native as legacy devices expect so by default.
841      */
842     if (!virtio_has_feature(vdev, VIRTIO_F_VERSION_1) &&
843         virtio_legacy_is_cross_endian(vdev)) {
844         r = vhost_virtqueue_set_vring_endian_legacy(dev,
845                                                     !virtio_is_big_endian(vdev),
846                                                     vhost_vq_index);
847         if (r < 0) {
848             error_report("failed to reset vring endianness");
849         }
850     }
851 
852     assert (r >= 0);
853     cpu_physical_memory_unmap(vq->ring, virtio_queue_get_ring_size(vdev, idx),
854                               0, virtio_queue_get_ring_size(vdev, idx));
855     cpu_physical_memory_unmap(vq->used, virtio_queue_get_used_size(vdev, idx),
856                               1, virtio_queue_get_used_size(vdev, idx));
857     cpu_physical_memory_unmap(vq->avail, virtio_queue_get_avail_size(vdev, idx),
858                               0, virtio_queue_get_avail_size(vdev, idx));
859     cpu_physical_memory_unmap(vq->desc, virtio_queue_get_desc_size(vdev, idx),
860                               0, virtio_queue_get_desc_size(vdev, idx));
861 }
862 
863 static void vhost_eventfd_add(MemoryListener *listener,
864                               MemoryRegionSection *section,
865                               bool match_data, uint64_t data, EventNotifier *e)
866 {
867 }
868 
869 static void vhost_eventfd_del(MemoryListener *listener,
870                               MemoryRegionSection *section,
871                               bool match_data, uint64_t data, EventNotifier *e)
872 {
873 }
874 
875 static int vhost_virtqueue_init(struct vhost_dev *dev,
876                                 struct vhost_virtqueue *vq, int n)
877 {
878     struct vhost_vring_file file = {
879         .index = n,
880     };
881     int r = event_notifier_init(&vq->masked_notifier, 0);
882     if (r < 0) {
883         return r;
884     }
885 
886     file.fd = event_notifier_get_fd(&vq->masked_notifier);
887     r = dev->vhost_ops->vhost_call(dev, VHOST_SET_VRING_CALL, &file);
888     if (r) {
889         r = -errno;
890         goto fail_call;
891     }
892     return 0;
893 fail_call:
894     event_notifier_cleanup(&vq->masked_notifier);
895     return r;
896 }
897 
898 static void vhost_virtqueue_cleanup(struct vhost_virtqueue *vq)
899 {
900     event_notifier_cleanup(&vq->masked_notifier);
901 }
902 
903 int vhost_dev_init(struct vhost_dev *hdev, void *opaque,
904                    VhostBackendType backend_type)
905 {
906     uint64_t features;
907     int i, r;
908 
909     if (vhost_set_backend_type(hdev, backend_type) < 0) {
910         close((uintptr_t)opaque);
911         return -1;
912     }
913 
914     if (hdev->vhost_ops->vhost_backend_init(hdev, opaque) < 0) {
915         close((uintptr_t)opaque);
916         return -errno;
917     }
918 
919     r = hdev->vhost_ops->vhost_call(hdev, VHOST_SET_OWNER, NULL);
920     if (r < 0) {
921         goto fail;
922     }
923 
924     r = hdev->vhost_ops->vhost_call(hdev, VHOST_GET_FEATURES, &features);
925     if (r < 0) {
926         goto fail;
927     }
928 
929     for (i = 0; i < hdev->nvqs; ++i) {
930         r = vhost_virtqueue_init(hdev, hdev->vqs + i, i);
931         if (r < 0) {
932             goto fail_vq;
933         }
934     }
935     hdev->features = features;
936 
937     hdev->memory_listener = (MemoryListener) {
938         .begin = vhost_begin,
939         .commit = vhost_commit,
940         .region_add = vhost_region_add,
941         .region_del = vhost_region_del,
942         .region_nop = vhost_region_nop,
943         .log_start = vhost_log_start,
944         .log_stop = vhost_log_stop,
945         .log_sync = vhost_log_sync,
946         .log_global_start = vhost_log_global_start,
947         .log_global_stop = vhost_log_global_stop,
948         .eventfd_add = vhost_eventfd_add,
949         .eventfd_del = vhost_eventfd_del,
950         .priority = 10
951     };
952     hdev->migration_blocker = NULL;
953     if (!(hdev->features & (0x1ULL << VHOST_F_LOG_ALL))) {
954         error_setg(&hdev->migration_blocker,
955                    "Migration disabled: vhost lacks VHOST_F_LOG_ALL feature.");
956         migrate_add_blocker(hdev->migration_blocker);
957     }
958     hdev->mem = g_malloc0(offsetof(struct vhost_memory, regions));
959     hdev->n_mem_sections = 0;
960     hdev->mem_sections = NULL;
961     hdev->log = NULL;
962     hdev->log_size = 0;
963     hdev->log_enabled = false;
964     hdev->started = false;
965     hdev->memory_changed = false;
966     memory_listener_register(&hdev->memory_listener, &address_space_memory);
967     return 0;
968 fail_vq:
969     while (--i >= 0) {
970         vhost_virtqueue_cleanup(hdev->vqs + i);
971     }
972 fail:
973     r = -errno;
974     hdev->vhost_ops->vhost_backend_cleanup(hdev);
975     return r;
976 }
977 
978 void vhost_dev_cleanup(struct vhost_dev *hdev)
979 {
980     int i;
981     for (i = 0; i < hdev->nvqs; ++i) {
982         vhost_virtqueue_cleanup(hdev->vqs + i);
983     }
984     memory_listener_unregister(&hdev->memory_listener);
985     if (hdev->migration_blocker) {
986         migrate_del_blocker(hdev->migration_blocker);
987         error_free(hdev->migration_blocker);
988     }
989     g_free(hdev->mem);
990     g_free(hdev->mem_sections);
991     hdev->vhost_ops->vhost_backend_cleanup(hdev);
992 }
993 
994 /* Stop processing guest IO notifications in qemu.
995  * Start processing them in vhost in kernel.
996  */
997 int vhost_dev_enable_notifiers(struct vhost_dev *hdev, VirtIODevice *vdev)
998 {
999     BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev)));
1000     VirtioBusState *vbus = VIRTIO_BUS(qbus);
1001     VirtioBusClass *k = VIRTIO_BUS_GET_CLASS(vbus);
1002     int i, r;
1003     if (!k->set_host_notifier) {
1004         fprintf(stderr, "binding does not support host notifiers\n");
1005         r = -ENOSYS;
1006         goto fail;
1007     }
1008 
1009     for (i = 0; i < hdev->nvqs; ++i) {
1010         r = k->set_host_notifier(qbus->parent, hdev->vq_index + i, true);
1011         if (r < 0) {
1012             fprintf(stderr, "vhost VQ %d notifier binding failed: %d\n", i, -r);
1013             goto fail_vq;
1014         }
1015     }
1016 
1017     return 0;
1018 fail_vq:
1019     while (--i >= 0) {
1020         r = k->set_host_notifier(qbus->parent, hdev->vq_index + i, false);
1021         if (r < 0) {
1022             fprintf(stderr, "vhost VQ %d notifier cleanup error: %d\n", i, -r);
1023             fflush(stderr);
1024         }
1025         assert (r >= 0);
1026     }
1027 fail:
1028     return r;
1029 }
1030 
1031 /* Stop processing guest IO notifications in vhost.
1032  * Start processing them in qemu.
1033  * This might actually run the qemu handlers right away,
1034  * so virtio in qemu must be completely setup when this is called.
1035  */
1036 void vhost_dev_disable_notifiers(struct vhost_dev *hdev, VirtIODevice *vdev)
1037 {
1038     BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev)));
1039     VirtioBusState *vbus = VIRTIO_BUS(qbus);
1040     VirtioBusClass *k = VIRTIO_BUS_GET_CLASS(vbus);
1041     int i, r;
1042 
1043     for (i = 0; i < hdev->nvqs; ++i) {
1044         r = k->set_host_notifier(qbus->parent, hdev->vq_index + i, false);
1045         if (r < 0) {
1046             fprintf(stderr, "vhost VQ %d notifier cleanup failed: %d\n", i, -r);
1047             fflush(stderr);
1048         }
1049         assert (r >= 0);
1050     }
1051 }
1052 
1053 /* Test and clear event pending status.
1054  * Should be called after unmask to avoid losing events.
1055  */
1056 bool vhost_virtqueue_pending(struct vhost_dev *hdev, int n)
1057 {
1058     struct vhost_virtqueue *vq = hdev->vqs + n - hdev->vq_index;
1059     assert(n >= hdev->vq_index && n < hdev->vq_index + hdev->nvqs);
1060     return event_notifier_test_and_clear(&vq->masked_notifier);
1061 }
1062 
1063 /* Mask/unmask events from this vq. */
1064 void vhost_virtqueue_mask(struct vhost_dev *hdev, VirtIODevice *vdev, int n,
1065                          bool mask)
1066 {
1067     struct VirtQueue *vvq = virtio_get_queue(vdev, n);
1068     int r, index = n - hdev->vq_index;
1069 
1070     assert(n >= hdev->vq_index && n < hdev->vq_index + hdev->nvqs);
1071 
1072     struct vhost_vring_file file = {
1073         .index = index
1074     };
1075     if (mask) {
1076         file.fd = event_notifier_get_fd(&hdev->vqs[index].masked_notifier);
1077     } else {
1078         file.fd = event_notifier_get_fd(virtio_queue_get_guest_notifier(vvq));
1079     }
1080     r = hdev->vhost_ops->vhost_call(hdev, VHOST_SET_VRING_CALL, &file);
1081     assert(r >= 0);
1082 }
1083 
1084 uint64_t vhost_get_features(struct vhost_dev *hdev, const int *feature_bits,
1085                             uint64_t features)
1086 {
1087     const int *bit = feature_bits;
1088     while (*bit != VHOST_INVALID_FEATURE_BIT) {
1089         uint64_t bit_mask = (1ULL << *bit);
1090         if (!(hdev->features & bit_mask)) {
1091             features &= ~bit_mask;
1092         }
1093         bit++;
1094     }
1095     return features;
1096 }
1097 
1098 void vhost_ack_features(struct vhost_dev *hdev, const int *feature_bits,
1099                         uint64_t features)
1100 {
1101     const int *bit = feature_bits;
1102     while (*bit != VHOST_INVALID_FEATURE_BIT) {
1103         uint64_t bit_mask = (1ULL << *bit);
1104         if (features & bit_mask) {
1105             hdev->acked_features |= bit_mask;
1106         }
1107         bit++;
1108     }
1109 }
1110 
1111 /* Host notifiers must be enabled at this point. */
1112 int vhost_dev_start(struct vhost_dev *hdev, VirtIODevice *vdev)
1113 {
1114     int i, r;
1115 
1116     hdev->started = true;
1117 
1118     r = vhost_dev_set_features(hdev, hdev->log_enabled);
1119     if (r < 0) {
1120         goto fail_features;
1121     }
1122     r = hdev->vhost_ops->vhost_call(hdev, VHOST_SET_MEM_TABLE, hdev->mem);
1123     if (r < 0) {
1124         r = -errno;
1125         goto fail_mem;
1126     }
1127     for (i = 0; i < hdev->nvqs; ++i) {
1128         r = vhost_virtqueue_start(hdev,
1129                                   vdev,
1130                                   hdev->vqs + i,
1131                                   hdev->vq_index + i);
1132         if (r < 0) {
1133             goto fail_vq;
1134         }
1135     }
1136 
1137     if (hdev->log_enabled) {
1138         uint64_t log_base;
1139 
1140         hdev->log_size = vhost_get_log_size(hdev);
1141         hdev->log = vhost_log_get(hdev->log_size);
1142         log_base = (uintptr_t)hdev->log->log;
1143         r = hdev->vhost_ops->vhost_call(hdev, VHOST_SET_LOG_BASE,
1144                                         hdev->log_size ? &log_base : NULL);
1145         if (r < 0) {
1146             r = -errno;
1147             goto fail_log;
1148         }
1149     }
1150 
1151     return 0;
1152 fail_log:
1153     vhost_log_put(hdev, false);
1154 fail_vq:
1155     while (--i >= 0) {
1156         vhost_virtqueue_stop(hdev,
1157                              vdev,
1158                              hdev->vqs + i,
1159                              hdev->vq_index + i);
1160     }
1161     i = hdev->nvqs;
1162 fail_mem:
1163 fail_features:
1164 
1165     hdev->started = false;
1166     return r;
1167 }
1168 
1169 /* Host notifiers must be enabled at this point. */
1170 void vhost_dev_stop(struct vhost_dev *hdev, VirtIODevice *vdev)
1171 {
1172     int i;
1173 
1174     for (i = 0; i < hdev->nvqs; ++i) {
1175         vhost_virtqueue_stop(hdev,
1176                              vdev,
1177                              hdev->vqs + i,
1178                              hdev->vq_index + i);
1179     }
1180 
1181     vhost_log_put(hdev, true);
1182     hdev->started = false;
1183     hdev->log = NULL;
1184     hdev->log_size = 0;
1185 }
1186 
1187