xref: /openbmc/qemu/hw/virtio/vhost.c (revision 9121d02c)
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 <sys/ioctl.h>
17 #include "hw/virtio/vhost.h"
18 #include "hw/hw.h"
19 #include "qemu/atomic.h"
20 #include "qemu/range.h"
21 #include <linux/vhost.h>
22 #include "exec/address-spaces.h"
23 #include "hw/virtio/virtio-bus.h"
24 
25 static void vhost_dev_sync_region(struct vhost_dev *dev,
26                                   MemoryRegionSection *section,
27                                   uint64_t mfirst, uint64_t mlast,
28                                   uint64_t rfirst, uint64_t rlast)
29 {
30     uint64_t start = MAX(mfirst, rfirst);
31     uint64_t end = MIN(mlast, rlast);
32     vhost_log_chunk_t *from = dev->log + start / VHOST_LOG_CHUNK;
33     vhost_log_chunk_t *to = dev->log + end / VHOST_LOG_CHUNK + 1;
34     uint64_t addr = (start / VHOST_LOG_CHUNK) * VHOST_LOG_CHUNK;
35 
36     if (end < start) {
37         return;
38     }
39     assert(end / VHOST_LOG_CHUNK < dev->log_size);
40     assert(start / VHOST_LOG_CHUNK < dev->log_size);
41 
42     for (;from < to; ++from) {
43         vhost_log_chunk_t log;
44         int bit;
45         /* We first check with non-atomic: much cheaper,
46          * and we expect non-dirty to be the common case. */
47         if (!*from) {
48             addr += VHOST_LOG_CHUNK;
49             continue;
50         }
51         /* Data must be read atomically. We don't really need barrier semantics
52          * but it's easier to use atomic_* than roll our own. */
53         log = atomic_xchg(from, 0);
54         while ((bit = sizeof(log) > sizeof(int) ?
55                 ffsll(log) : ffs(log))) {
56             hwaddr page_addr;
57             hwaddr section_offset;
58             hwaddr mr_offset;
59             bit -= 1;
60             page_addr = addr + bit * VHOST_LOG_PAGE;
61             section_offset = page_addr - section->offset_within_address_space;
62             mr_offset = section_offset + section->offset_within_region;
63             memory_region_set_dirty(section->mr, mr_offset, VHOST_LOG_PAGE);
64             log &= ~(0x1ull << bit);
65         }
66         addr += VHOST_LOG_CHUNK;
67     }
68 }
69 
70 static int vhost_sync_dirty_bitmap(struct vhost_dev *dev,
71                                    MemoryRegionSection *section,
72                                    hwaddr first,
73                                    hwaddr last)
74 {
75     int i;
76     hwaddr start_addr;
77     hwaddr end_addr;
78 
79     if (!dev->log_enabled || !dev->started) {
80         return 0;
81     }
82     start_addr = section->offset_within_address_space;
83     end_addr = range_get_last(start_addr, int128_get64(section->size));
84     start_addr = MAX(first, start_addr);
85     end_addr = MIN(last, end_addr);
86 
87     for (i = 0; i < dev->mem->nregions; ++i) {
88         struct vhost_memory_region *reg = dev->mem->regions + i;
89         vhost_dev_sync_region(dev, section, start_addr, end_addr,
90                               reg->guest_phys_addr,
91                               range_get_last(reg->guest_phys_addr,
92                                              reg->memory_size));
93     }
94     for (i = 0; i < dev->nvqs; ++i) {
95         struct vhost_virtqueue *vq = dev->vqs + i;
96         vhost_dev_sync_region(dev, section, start_addr, end_addr, vq->used_phys,
97                               range_get_last(vq->used_phys, vq->used_size));
98     }
99     return 0;
100 }
101 
102 static void vhost_log_sync(MemoryListener *listener,
103                           MemoryRegionSection *section)
104 {
105     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
106                                          memory_listener);
107     vhost_sync_dirty_bitmap(dev, section, 0x0, ~0x0ULL);
108 }
109 
110 static void vhost_log_sync_range(struct vhost_dev *dev,
111                                  hwaddr first, hwaddr last)
112 {
113     int i;
114     /* FIXME: this is N^2 in number of sections */
115     for (i = 0; i < dev->n_mem_sections; ++i) {
116         MemoryRegionSection *section = &dev->mem_sections[i];
117         vhost_sync_dirty_bitmap(dev, section, first, last);
118     }
119 }
120 
121 /* Assign/unassign. Keep an unsorted array of non-overlapping
122  * memory regions in dev->mem. */
123 static void vhost_dev_unassign_memory(struct vhost_dev *dev,
124                                       uint64_t start_addr,
125                                       uint64_t size)
126 {
127     int from, to, n = dev->mem->nregions;
128     /* Track overlapping/split regions for sanity checking. */
129     int overlap_start = 0, overlap_end = 0, overlap_middle = 0, split = 0;
130 
131     for (from = 0, to = 0; from < n; ++from, ++to) {
132         struct vhost_memory_region *reg = dev->mem->regions + to;
133         uint64_t reglast;
134         uint64_t memlast;
135         uint64_t change;
136 
137         /* clone old region */
138         if (to != from) {
139             memcpy(reg, dev->mem->regions + from, sizeof *reg);
140         }
141 
142         /* No overlap is simple */
143         if (!ranges_overlap(reg->guest_phys_addr, reg->memory_size,
144                             start_addr, size)) {
145             continue;
146         }
147 
148         /* Split only happens if supplied region
149          * is in the middle of an existing one. Thus it can not
150          * overlap with any other existing region. */
151         assert(!split);
152 
153         reglast = range_get_last(reg->guest_phys_addr, reg->memory_size);
154         memlast = range_get_last(start_addr, size);
155 
156         /* Remove whole region */
157         if (start_addr <= reg->guest_phys_addr && memlast >= reglast) {
158             --dev->mem->nregions;
159             --to;
160             ++overlap_middle;
161             continue;
162         }
163 
164         /* Shrink region */
165         if (memlast >= reglast) {
166             reg->memory_size = start_addr - reg->guest_phys_addr;
167             assert(reg->memory_size);
168             assert(!overlap_end);
169             ++overlap_end;
170             continue;
171         }
172 
173         /* Shift region */
174         if (start_addr <= reg->guest_phys_addr) {
175             change = memlast + 1 - reg->guest_phys_addr;
176             reg->memory_size -= change;
177             reg->guest_phys_addr += change;
178             reg->userspace_addr += change;
179             assert(reg->memory_size);
180             assert(!overlap_start);
181             ++overlap_start;
182             continue;
183         }
184 
185         /* This only happens if supplied region
186          * is in the middle of an existing one. Thus it can not
187          * overlap with any other existing region. */
188         assert(!overlap_start);
189         assert(!overlap_end);
190         assert(!overlap_middle);
191         /* Split region: shrink first part, shift second part. */
192         memcpy(dev->mem->regions + n, reg, sizeof *reg);
193         reg->memory_size = start_addr - reg->guest_phys_addr;
194         assert(reg->memory_size);
195         change = memlast + 1 - reg->guest_phys_addr;
196         reg = dev->mem->regions + n;
197         reg->memory_size -= change;
198         assert(reg->memory_size);
199         reg->guest_phys_addr += change;
200         reg->userspace_addr += change;
201         /* Never add more than 1 region */
202         assert(dev->mem->nregions == n);
203         ++dev->mem->nregions;
204         ++split;
205     }
206 }
207 
208 /* Called after unassign, so no regions overlap the given range. */
209 static void vhost_dev_assign_memory(struct vhost_dev *dev,
210                                     uint64_t start_addr,
211                                     uint64_t size,
212                                     uint64_t uaddr)
213 {
214     int from, to;
215     struct vhost_memory_region *merged = NULL;
216     for (from = 0, to = 0; from < dev->mem->nregions; ++from, ++to) {
217         struct vhost_memory_region *reg = dev->mem->regions + to;
218         uint64_t prlast, urlast;
219         uint64_t pmlast, umlast;
220         uint64_t s, e, u;
221 
222         /* clone old region */
223         if (to != from) {
224             memcpy(reg, dev->mem->regions + from, sizeof *reg);
225         }
226         prlast = range_get_last(reg->guest_phys_addr, reg->memory_size);
227         pmlast = range_get_last(start_addr, size);
228         urlast = range_get_last(reg->userspace_addr, reg->memory_size);
229         umlast = range_get_last(uaddr, size);
230 
231         /* check for overlapping regions: should never happen. */
232         assert(prlast < start_addr || pmlast < reg->guest_phys_addr);
233         /* Not an adjacent or overlapping region - do not merge. */
234         if ((prlast + 1 != start_addr || urlast + 1 != uaddr) &&
235             (pmlast + 1 != reg->guest_phys_addr ||
236              umlast + 1 != reg->userspace_addr)) {
237             continue;
238         }
239 
240         if (merged) {
241             --to;
242             assert(to >= 0);
243         } else {
244             merged = reg;
245         }
246         u = MIN(uaddr, reg->userspace_addr);
247         s = MIN(start_addr, reg->guest_phys_addr);
248         e = MAX(pmlast, prlast);
249         uaddr = merged->userspace_addr = u;
250         start_addr = merged->guest_phys_addr = s;
251         size = merged->memory_size = e - s + 1;
252         assert(merged->memory_size);
253     }
254 
255     if (!merged) {
256         struct vhost_memory_region *reg = dev->mem->regions + to;
257         memset(reg, 0, sizeof *reg);
258         reg->memory_size = size;
259         assert(reg->memory_size);
260         reg->guest_phys_addr = start_addr;
261         reg->userspace_addr = uaddr;
262         ++to;
263     }
264     assert(to <= dev->mem->nregions + 1);
265     dev->mem->nregions = to;
266 }
267 
268 static uint64_t vhost_get_log_size(struct vhost_dev *dev)
269 {
270     uint64_t log_size = 0;
271     int i;
272     for (i = 0; i < dev->mem->nregions; ++i) {
273         struct vhost_memory_region *reg = dev->mem->regions + i;
274         uint64_t last = range_get_last(reg->guest_phys_addr,
275                                        reg->memory_size);
276         log_size = MAX(log_size, last / VHOST_LOG_CHUNK + 1);
277     }
278     for (i = 0; i < dev->nvqs; ++i) {
279         struct vhost_virtqueue *vq = dev->vqs + i;
280         uint64_t last = vq->used_phys + vq->used_size - 1;
281         log_size = MAX(log_size, last / VHOST_LOG_CHUNK + 1);
282     }
283     return log_size;
284 }
285 
286 static inline void vhost_dev_log_resize(struct vhost_dev* dev, uint64_t size)
287 {
288     vhost_log_chunk_t *log;
289     uint64_t log_base;
290     int r;
291 
292     log = g_malloc0(size * sizeof *log);
293     log_base = (uint64_t)(unsigned long)log;
294     r = ioctl(dev->control, VHOST_SET_LOG_BASE, &log_base);
295     assert(r >= 0);
296     /* Sync only the range covered by the old log */
297     if (dev->log_size) {
298         vhost_log_sync_range(dev, 0, dev->log_size * VHOST_LOG_CHUNK - 1);
299     }
300     if (dev->log) {
301         g_free(dev->log);
302     }
303     dev->log = log;
304     dev->log_size = size;
305 }
306 
307 static int vhost_verify_ring_mappings(struct vhost_dev *dev,
308                                       uint64_t start_addr,
309                                       uint64_t size)
310 {
311     int i;
312     for (i = 0; i < dev->nvqs; ++i) {
313         struct vhost_virtqueue *vq = dev->vqs + i;
314         hwaddr l;
315         void *p;
316 
317         if (!ranges_overlap(start_addr, size, vq->ring_phys, vq->ring_size)) {
318             continue;
319         }
320         l = vq->ring_size;
321         p = cpu_physical_memory_map(vq->ring_phys, &l, 1);
322         if (!p || l != vq->ring_size) {
323             fprintf(stderr, "Unable to map ring buffer for ring %d\n", i);
324             return -ENOMEM;
325         }
326         if (p != vq->ring) {
327             fprintf(stderr, "Ring buffer relocated for ring %d\n", i);
328             return -EBUSY;
329         }
330         cpu_physical_memory_unmap(p, l, 0, 0);
331     }
332     return 0;
333 }
334 
335 static struct vhost_memory_region *vhost_dev_find_reg(struct vhost_dev *dev,
336 						      uint64_t start_addr,
337 						      uint64_t size)
338 {
339     int i, n = dev->mem->nregions;
340     for (i = 0; i < n; ++i) {
341         struct vhost_memory_region *reg = dev->mem->regions + i;
342         if (ranges_overlap(reg->guest_phys_addr, reg->memory_size,
343                            start_addr, size)) {
344             return reg;
345         }
346     }
347     return NULL;
348 }
349 
350 static bool vhost_dev_cmp_memory(struct vhost_dev *dev,
351                                  uint64_t start_addr,
352                                  uint64_t size,
353                                  uint64_t uaddr)
354 {
355     struct vhost_memory_region *reg = vhost_dev_find_reg(dev, start_addr, size);
356     uint64_t reglast;
357     uint64_t memlast;
358 
359     if (!reg) {
360         return true;
361     }
362 
363     reglast = range_get_last(reg->guest_phys_addr, reg->memory_size);
364     memlast = range_get_last(start_addr, size);
365 
366     /* Need to extend region? */
367     if (start_addr < reg->guest_phys_addr || memlast > reglast) {
368         return true;
369     }
370     /* userspace_addr changed? */
371     return uaddr != reg->userspace_addr + start_addr - reg->guest_phys_addr;
372 }
373 
374 static void vhost_set_memory(MemoryListener *listener,
375                              MemoryRegionSection *section,
376                              bool add)
377 {
378     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
379                                          memory_listener);
380     hwaddr start_addr = section->offset_within_address_space;
381     ram_addr_t size = int128_get64(section->size);
382     bool log_dirty = memory_region_is_logging(section->mr);
383     int s = offsetof(struct vhost_memory, regions) +
384         (dev->mem->nregions + 1) * sizeof dev->mem->regions[0];
385     void *ram;
386 
387     dev->mem = g_realloc(dev->mem, s);
388 
389     if (log_dirty) {
390         add = false;
391     }
392 
393     assert(size);
394 
395     /* Optimize no-change case. At least cirrus_vga does this a lot at this time. */
396     ram = memory_region_get_ram_ptr(section->mr) + section->offset_within_region;
397     if (add) {
398         if (!vhost_dev_cmp_memory(dev, start_addr, size, (uintptr_t)ram)) {
399             /* Region exists with same address. Nothing to do. */
400             return;
401         }
402     } else {
403         if (!vhost_dev_find_reg(dev, start_addr, size)) {
404             /* Removing region that we don't access. Nothing to do. */
405             return;
406         }
407     }
408 
409     vhost_dev_unassign_memory(dev, start_addr, size);
410     if (add) {
411         /* Add given mapping, merging adjacent regions if any */
412         vhost_dev_assign_memory(dev, start_addr, size, (uintptr_t)ram);
413     } else {
414         /* Remove old mapping for this memory, if any. */
415         vhost_dev_unassign_memory(dev, start_addr, size);
416     }
417     dev->mem_changed_start_addr = MIN(dev->mem_changed_start_addr, start_addr);
418     dev->mem_changed_end_addr = MAX(dev->mem_changed_end_addr, start_addr + size - 1);
419     dev->memory_changed = true;
420 }
421 
422 static bool vhost_section(MemoryRegionSection *section)
423 {
424     return memory_region_is_ram(section->mr);
425 }
426 
427 static void vhost_begin(MemoryListener *listener)
428 {
429     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
430                                          memory_listener);
431     dev->mem_changed_end_addr = 0;
432     dev->mem_changed_start_addr = -1;
433 }
434 
435 static void vhost_commit(MemoryListener *listener)
436 {
437     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
438                                          memory_listener);
439     hwaddr start_addr = 0;
440     ram_addr_t size = 0;
441     uint64_t log_size;
442     int r;
443 
444     if (!dev->memory_changed) {
445         return;
446     }
447     if (!dev->started) {
448         return;
449     }
450     if (dev->mem_changed_start_addr > dev->mem_changed_end_addr) {
451         return;
452     }
453 
454     if (dev->started) {
455         start_addr = dev->mem_changed_start_addr;
456         size = dev->mem_changed_end_addr - dev->mem_changed_start_addr + 1;
457 
458         r = vhost_verify_ring_mappings(dev, start_addr, size);
459         assert(r >= 0);
460     }
461 
462     if (!dev->log_enabled) {
463         r = ioctl(dev->control, VHOST_SET_MEM_TABLE, dev->mem);
464         assert(r >= 0);
465         dev->memory_changed = false;
466         return;
467     }
468     log_size = vhost_get_log_size(dev);
469     /* We allocate an extra 4K bytes to log,
470      * to reduce the * number of reallocations. */
471 #define VHOST_LOG_BUFFER (0x1000 / sizeof *dev->log)
472     /* To log more, must increase log size before table update. */
473     if (dev->log_size < log_size) {
474         vhost_dev_log_resize(dev, log_size + VHOST_LOG_BUFFER);
475     }
476     r = ioctl(dev->control, VHOST_SET_MEM_TABLE, dev->mem);
477     assert(r >= 0);
478     /* To log less, can only decrease log size after table update. */
479     if (dev->log_size > log_size + VHOST_LOG_BUFFER) {
480         vhost_dev_log_resize(dev, log_size);
481     }
482     dev->memory_changed = false;
483 }
484 
485 static void vhost_region_add(MemoryListener *listener,
486                              MemoryRegionSection *section)
487 {
488     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
489                                          memory_listener);
490 
491     if (!vhost_section(section)) {
492         return;
493     }
494 
495     ++dev->n_mem_sections;
496     dev->mem_sections = g_renew(MemoryRegionSection, dev->mem_sections,
497                                 dev->n_mem_sections);
498     dev->mem_sections[dev->n_mem_sections - 1] = *section;
499     memory_region_ref(section->mr);
500     vhost_set_memory(listener, section, true);
501 }
502 
503 static void vhost_region_del(MemoryListener *listener,
504                              MemoryRegionSection *section)
505 {
506     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
507                                          memory_listener);
508     int i;
509 
510     if (!vhost_section(section)) {
511         return;
512     }
513 
514     vhost_set_memory(listener, section, false);
515     memory_region_unref(section->mr);
516     for (i = 0; i < dev->n_mem_sections; ++i) {
517         if (dev->mem_sections[i].offset_within_address_space
518             == section->offset_within_address_space) {
519             --dev->n_mem_sections;
520             memmove(&dev->mem_sections[i], &dev->mem_sections[i+1],
521                     (dev->n_mem_sections - i) * sizeof(*dev->mem_sections));
522             break;
523         }
524     }
525 }
526 
527 static void vhost_region_nop(MemoryListener *listener,
528                              MemoryRegionSection *section)
529 {
530 }
531 
532 static int vhost_virtqueue_set_addr(struct vhost_dev *dev,
533                                     struct vhost_virtqueue *vq,
534                                     unsigned idx, bool enable_log)
535 {
536     struct vhost_vring_addr addr = {
537         .index = idx,
538         .desc_user_addr = (uint64_t)(unsigned long)vq->desc,
539         .avail_user_addr = (uint64_t)(unsigned long)vq->avail,
540         .used_user_addr = (uint64_t)(unsigned long)vq->used,
541         .log_guest_addr = vq->used_phys,
542         .flags = enable_log ? (1 << VHOST_VRING_F_LOG) : 0,
543     };
544     int r = ioctl(dev->control, VHOST_SET_VRING_ADDR, &addr);
545     if (r < 0) {
546         return -errno;
547     }
548     return 0;
549 }
550 
551 static int vhost_dev_set_features(struct vhost_dev *dev, bool enable_log)
552 {
553     uint64_t features = dev->acked_features;
554     int r;
555     if (enable_log) {
556         features |= 0x1 << VHOST_F_LOG_ALL;
557     }
558     r = ioctl(dev->control, VHOST_SET_FEATURES, &features);
559     return r < 0 ? -errno : 0;
560 }
561 
562 static int vhost_dev_set_log(struct vhost_dev *dev, bool enable_log)
563 {
564     int r, t, i;
565     r = vhost_dev_set_features(dev, enable_log);
566     if (r < 0) {
567         goto err_features;
568     }
569     for (i = 0; i < dev->nvqs; ++i) {
570         r = vhost_virtqueue_set_addr(dev, dev->vqs + i, i,
571                                      enable_log);
572         if (r < 0) {
573             goto err_vq;
574         }
575     }
576     return 0;
577 err_vq:
578     for (; i >= 0; --i) {
579         t = vhost_virtqueue_set_addr(dev, dev->vqs + i, i,
580                                      dev->log_enabled);
581         assert(t >= 0);
582     }
583     t = vhost_dev_set_features(dev, dev->log_enabled);
584     assert(t >= 0);
585 err_features:
586     return r;
587 }
588 
589 static int vhost_migration_log(MemoryListener *listener, int enable)
590 {
591     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
592                                          memory_listener);
593     int r;
594     if (!!enable == dev->log_enabled) {
595         return 0;
596     }
597     if (!dev->started) {
598         dev->log_enabled = enable;
599         return 0;
600     }
601     if (!enable) {
602         r = vhost_dev_set_log(dev, false);
603         if (r < 0) {
604             return r;
605         }
606         if (dev->log) {
607             g_free(dev->log);
608         }
609         dev->log = NULL;
610         dev->log_size = 0;
611     } else {
612         vhost_dev_log_resize(dev, vhost_get_log_size(dev));
613         r = vhost_dev_set_log(dev, true);
614         if (r < 0) {
615             return r;
616         }
617     }
618     dev->log_enabled = enable;
619     return 0;
620 }
621 
622 static void vhost_log_global_start(MemoryListener *listener)
623 {
624     int r;
625 
626     r = vhost_migration_log(listener, true);
627     if (r < 0) {
628         abort();
629     }
630 }
631 
632 static void vhost_log_global_stop(MemoryListener *listener)
633 {
634     int r;
635 
636     r = vhost_migration_log(listener, false);
637     if (r < 0) {
638         abort();
639     }
640 }
641 
642 static void vhost_log_start(MemoryListener *listener,
643                             MemoryRegionSection *section)
644 {
645     /* FIXME: implement */
646 }
647 
648 static void vhost_log_stop(MemoryListener *listener,
649                            MemoryRegionSection *section)
650 {
651     /* FIXME: implement */
652 }
653 
654 static int vhost_virtqueue_start(struct vhost_dev *dev,
655                                 struct VirtIODevice *vdev,
656                                 struct vhost_virtqueue *vq,
657                                 unsigned idx)
658 {
659     hwaddr s, l, a;
660     int r;
661     int vhost_vq_index = idx - dev->vq_index;
662     struct vhost_vring_file file = {
663         .index = vhost_vq_index
664     };
665     struct vhost_vring_state state = {
666         .index = vhost_vq_index
667     };
668     struct VirtQueue *vvq = virtio_get_queue(vdev, idx);
669 
670     assert(idx >= dev->vq_index && idx < dev->vq_index + dev->nvqs);
671 
672     vq->num = state.num = virtio_queue_get_num(vdev, idx);
673     r = ioctl(dev->control, VHOST_SET_VRING_NUM, &state);
674     if (r) {
675         return -errno;
676     }
677 
678     state.num = virtio_queue_get_last_avail_idx(vdev, idx);
679     r = ioctl(dev->control, VHOST_SET_VRING_BASE, &state);
680     if (r) {
681         return -errno;
682     }
683 
684     s = l = virtio_queue_get_desc_size(vdev, idx);
685     a = virtio_queue_get_desc_addr(vdev, idx);
686     vq->desc = cpu_physical_memory_map(a, &l, 0);
687     if (!vq->desc || l != s) {
688         r = -ENOMEM;
689         goto fail_alloc_desc;
690     }
691     s = l = virtio_queue_get_avail_size(vdev, idx);
692     a = virtio_queue_get_avail_addr(vdev, idx);
693     vq->avail = cpu_physical_memory_map(a, &l, 0);
694     if (!vq->avail || l != s) {
695         r = -ENOMEM;
696         goto fail_alloc_avail;
697     }
698     vq->used_size = s = l = virtio_queue_get_used_size(vdev, idx);
699     vq->used_phys = a = virtio_queue_get_used_addr(vdev, idx);
700     vq->used = cpu_physical_memory_map(a, &l, 1);
701     if (!vq->used || l != s) {
702         r = -ENOMEM;
703         goto fail_alloc_used;
704     }
705 
706     vq->ring_size = s = l = virtio_queue_get_ring_size(vdev, idx);
707     vq->ring_phys = a = virtio_queue_get_ring_addr(vdev, idx);
708     vq->ring = cpu_physical_memory_map(a, &l, 1);
709     if (!vq->ring || l != s) {
710         r = -ENOMEM;
711         goto fail_alloc_ring;
712     }
713 
714     r = vhost_virtqueue_set_addr(dev, vq, vhost_vq_index, dev->log_enabled);
715     if (r < 0) {
716         r = -errno;
717         goto fail_alloc;
718     }
719 
720     file.fd = event_notifier_get_fd(virtio_queue_get_host_notifier(vvq));
721     r = ioctl(dev->control, VHOST_SET_VRING_KICK, &file);
722     if (r) {
723         r = -errno;
724         goto fail_kick;
725     }
726 
727     /* Clear and discard previous events if any. */
728     event_notifier_test_and_clear(&vq->masked_notifier);
729 
730     return 0;
731 
732 fail_kick:
733 fail_alloc:
734     cpu_physical_memory_unmap(vq->ring, virtio_queue_get_ring_size(vdev, idx),
735                               0, 0);
736 fail_alloc_ring:
737     cpu_physical_memory_unmap(vq->used, virtio_queue_get_used_size(vdev, idx),
738                               0, 0);
739 fail_alloc_used:
740     cpu_physical_memory_unmap(vq->avail, virtio_queue_get_avail_size(vdev, idx),
741                               0, 0);
742 fail_alloc_avail:
743     cpu_physical_memory_unmap(vq->desc, virtio_queue_get_desc_size(vdev, idx),
744                               0, 0);
745 fail_alloc_desc:
746     return r;
747 }
748 
749 static void vhost_virtqueue_stop(struct vhost_dev *dev,
750                                     struct VirtIODevice *vdev,
751                                     struct vhost_virtqueue *vq,
752                                     unsigned idx)
753 {
754     struct vhost_vring_state state = {
755         .index = idx - dev->vq_index
756     };
757     int r;
758     assert(idx >= dev->vq_index && idx < dev->vq_index + dev->nvqs);
759     r = ioctl(dev->control, VHOST_GET_VRING_BASE, &state);
760     if (r < 0) {
761         fprintf(stderr, "vhost VQ %d ring restore failed: %d\n", idx, r);
762         fflush(stderr);
763     }
764     virtio_queue_set_last_avail_idx(vdev, idx, state.num);
765     assert (r >= 0);
766     cpu_physical_memory_unmap(vq->ring, virtio_queue_get_ring_size(vdev, idx),
767                               0, virtio_queue_get_ring_size(vdev, idx));
768     cpu_physical_memory_unmap(vq->used, virtio_queue_get_used_size(vdev, idx),
769                               1, virtio_queue_get_used_size(vdev, idx));
770     cpu_physical_memory_unmap(vq->avail, virtio_queue_get_avail_size(vdev, idx),
771                               0, virtio_queue_get_avail_size(vdev, idx));
772     cpu_physical_memory_unmap(vq->desc, virtio_queue_get_desc_size(vdev, idx),
773                               0, virtio_queue_get_desc_size(vdev, idx));
774 }
775 
776 static void vhost_eventfd_add(MemoryListener *listener,
777                               MemoryRegionSection *section,
778                               bool match_data, uint64_t data, EventNotifier *e)
779 {
780 }
781 
782 static void vhost_eventfd_del(MemoryListener *listener,
783                               MemoryRegionSection *section,
784                               bool match_data, uint64_t data, EventNotifier *e)
785 {
786 }
787 
788 static int vhost_virtqueue_init(struct vhost_dev *dev,
789                                 struct vhost_virtqueue *vq, int n)
790 {
791     struct vhost_vring_file file = {
792         .index = n,
793     };
794     int r = event_notifier_init(&vq->masked_notifier, 0);
795     if (r < 0) {
796         return r;
797     }
798 
799     file.fd = event_notifier_get_fd(&vq->masked_notifier);
800     r = ioctl(dev->control, VHOST_SET_VRING_CALL, &file);
801     if (r) {
802         r = -errno;
803         goto fail_call;
804     }
805     return 0;
806 fail_call:
807     event_notifier_cleanup(&vq->masked_notifier);
808     return r;
809 }
810 
811 static void vhost_virtqueue_cleanup(struct vhost_virtqueue *vq)
812 {
813     event_notifier_cleanup(&vq->masked_notifier);
814 }
815 
816 int vhost_dev_init(struct vhost_dev *hdev, int devfd, const char *devpath,
817                    bool force)
818 {
819     uint64_t features;
820     int i, r;
821     if (devfd >= 0) {
822         hdev->control = devfd;
823     } else {
824         hdev->control = open(devpath, O_RDWR);
825         if (hdev->control < 0) {
826             return -errno;
827         }
828     }
829     r = ioctl(hdev->control, VHOST_SET_OWNER, NULL);
830     if (r < 0) {
831         goto fail;
832     }
833 
834     r = ioctl(hdev->control, VHOST_GET_FEATURES, &features);
835     if (r < 0) {
836         goto fail;
837     }
838 
839     for (i = 0; i < hdev->nvqs; ++i) {
840         r = vhost_virtqueue_init(hdev, hdev->vqs + i, i);
841         if (r < 0) {
842             goto fail_vq;
843         }
844     }
845     hdev->features = features;
846 
847     hdev->memory_listener = (MemoryListener) {
848         .begin = vhost_begin,
849         .commit = vhost_commit,
850         .region_add = vhost_region_add,
851         .region_del = vhost_region_del,
852         .region_nop = vhost_region_nop,
853         .log_start = vhost_log_start,
854         .log_stop = vhost_log_stop,
855         .log_sync = vhost_log_sync,
856         .log_global_start = vhost_log_global_start,
857         .log_global_stop = vhost_log_global_stop,
858         .eventfd_add = vhost_eventfd_add,
859         .eventfd_del = vhost_eventfd_del,
860         .priority = 10
861     };
862     hdev->mem = g_malloc0(offsetof(struct vhost_memory, regions));
863     hdev->n_mem_sections = 0;
864     hdev->mem_sections = NULL;
865     hdev->log = NULL;
866     hdev->log_size = 0;
867     hdev->log_enabled = false;
868     hdev->started = false;
869     hdev->memory_changed = false;
870     memory_listener_register(&hdev->memory_listener, &address_space_memory);
871     hdev->force = force;
872     return 0;
873 fail_vq:
874     while (--i >= 0) {
875         vhost_virtqueue_cleanup(hdev->vqs + i);
876     }
877 fail:
878     r = -errno;
879     close(hdev->control);
880     return r;
881 }
882 
883 void vhost_dev_cleanup(struct vhost_dev *hdev)
884 {
885     int i;
886     for (i = 0; i < hdev->nvqs; ++i) {
887         vhost_virtqueue_cleanup(hdev->vqs + i);
888     }
889     memory_listener_unregister(&hdev->memory_listener);
890     g_free(hdev->mem);
891     g_free(hdev->mem_sections);
892     close(hdev->control);
893 }
894 
895 bool vhost_dev_query(struct vhost_dev *hdev, VirtIODevice *vdev)
896 {
897     BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev)));
898     VirtioBusState *vbus = VIRTIO_BUS(qbus);
899     VirtioBusClass *k = VIRTIO_BUS_GET_CLASS(vbus);
900 
901     return !k->query_guest_notifiers ||
902            k->query_guest_notifiers(qbus->parent) ||
903            hdev->force;
904 }
905 
906 /* Stop processing guest IO notifications in qemu.
907  * Start processing them in vhost in kernel.
908  */
909 int vhost_dev_enable_notifiers(struct vhost_dev *hdev, VirtIODevice *vdev)
910 {
911     BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev)));
912     VirtioBusState *vbus = VIRTIO_BUS(qbus);
913     VirtioBusClass *k = VIRTIO_BUS_GET_CLASS(vbus);
914     int i, r;
915     if (!k->set_host_notifier) {
916         fprintf(stderr, "binding does not support host notifiers\n");
917         r = -ENOSYS;
918         goto fail;
919     }
920 
921     for (i = 0; i < hdev->nvqs; ++i) {
922         r = k->set_host_notifier(qbus->parent, hdev->vq_index + i, true);
923         if (r < 0) {
924             fprintf(stderr, "vhost VQ %d notifier binding failed: %d\n", i, -r);
925             goto fail_vq;
926         }
927     }
928 
929     return 0;
930 fail_vq:
931     while (--i >= 0) {
932         r = k->set_host_notifier(qbus->parent, hdev->vq_index + i, false);
933         if (r < 0) {
934             fprintf(stderr, "vhost VQ %d notifier cleanup error: %d\n", i, -r);
935             fflush(stderr);
936         }
937         assert (r >= 0);
938     }
939 fail:
940     return r;
941 }
942 
943 /* Stop processing guest IO notifications in vhost.
944  * Start processing them in qemu.
945  * This might actually run the qemu handlers right away,
946  * so virtio in qemu must be completely setup when this is called.
947  */
948 void vhost_dev_disable_notifiers(struct vhost_dev *hdev, VirtIODevice *vdev)
949 {
950     BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev)));
951     VirtioBusState *vbus = VIRTIO_BUS(qbus);
952     VirtioBusClass *k = VIRTIO_BUS_GET_CLASS(vbus);
953     int i, r;
954 
955     for (i = 0; i < hdev->nvqs; ++i) {
956         r = k->set_host_notifier(qbus->parent, hdev->vq_index + i, false);
957         if (r < 0) {
958             fprintf(stderr, "vhost VQ %d notifier cleanup failed: %d\n", i, -r);
959             fflush(stderr);
960         }
961         assert (r >= 0);
962     }
963 }
964 
965 /* Test and clear event pending status.
966  * Should be called after unmask to avoid losing events.
967  */
968 bool vhost_virtqueue_pending(struct vhost_dev *hdev, int n)
969 {
970     struct vhost_virtqueue *vq = hdev->vqs + n - hdev->vq_index;
971     assert(hdev->started);
972     assert(n >= hdev->vq_index && n < hdev->vq_index + hdev->nvqs);
973     return event_notifier_test_and_clear(&vq->masked_notifier);
974 }
975 
976 /* Mask/unmask events from this vq. */
977 void vhost_virtqueue_mask(struct vhost_dev *hdev, VirtIODevice *vdev, int n,
978                          bool mask)
979 {
980     struct VirtQueue *vvq = virtio_get_queue(vdev, n);
981     int r, index = n - hdev->vq_index;
982 
983     assert(hdev->started);
984     assert(n >= hdev->vq_index && n < hdev->vq_index + hdev->nvqs);
985 
986     struct vhost_vring_file file = {
987         .index = index
988     };
989     if (mask) {
990         file.fd = event_notifier_get_fd(&hdev->vqs[index].masked_notifier);
991     } else {
992         file.fd = event_notifier_get_fd(virtio_queue_get_guest_notifier(vvq));
993     }
994     r = ioctl(hdev->control, VHOST_SET_VRING_CALL, &file);
995     assert(r >= 0);
996 }
997 
998 /* Host notifiers must be enabled at this point. */
999 int vhost_dev_start(struct vhost_dev *hdev, VirtIODevice *vdev)
1000 {
1001     int i, r;
1002 
1003     hdev->started = true;
1004 
1005     r = vhost_dev_set_features(hdev, hdev->log_enabled);
1006     if (r < 0) {
1007         goto fail_features;
1008     }
1009     r = ioctl(hdev->control, VHOST_SET_MEM_TABLE, hdev->mem);
1010     if (r < 0) {
1011         r = -errno;
1012         goto fail_mem;
1013     }
1014     for (i = 0; i < hdev->nvqs; ++i) {
1015         r = vhost_virtqueue_start(hdev,
1016                                   vdev,
1017                                   hdev->vqs + i,
1018                                   hdev->vq_index + i);
1019         if (r < 0) {
1020             goto fail_vq;
1021         }
1022     }
1023 
1024     if (hdev->log_enabled) {
1025         hdev->log_size = vhost_get_log_size(hdev);
1026         hdev->log = hdev->log_size ?
1027             g_malloc0(hdev->log_size * sizeof *hdev->log) : NULL;
1028         r = ioctl(hdev->control, VHOST_SET_LOG_BASE,
1029                   (uint64_t)(unsigned long)hdev->log);
1030         if (r < 0) {
1031             r = -errno;
1032             goto fail_log;
1033         }
1034     }
1035 
1036     return 0;
1037 fail_log:
1038 fail_vq:
1039     while (--i >= 0) {
1040         vhost_virtqueue_stop(hdev,
1041                              vdev,
1042                              hdev->vqs + i,
1043                              hdev->vq_index + i);
1044     }
1045     i = hdev->nvqs;
1046 fail_mem:
1047 fail_features:
1048 
1049     hdev->started = false;
1050     return r;
1051 }
1052 
1053 /* Host notifiers must be enabled at this point. */
1054 void vhost_dev_stop(struct vhost_dev *hdev, VirtIODevice *vdev)
1055 {
1056     int i;
1057 
1058     for (i = 0; i < hdev->nvqs; ++i) {
1059         vhost_virtqueue_stop(hdev,
1060                              vdev,
1061                              hdev->vqs + i,
1062                              hdev->vq_index + i);
1063     }
1064     vhost_log_sync_range(hdev, 0, ~0x0ull);
1065 
1066     hdev->started = false;
1067     g_free(hdev->log);
1068     hdev->log = NULL;
1069     hdev->log_size = 0;
1070 }
1071 
1072