xref: /openbmc/qemu/hw/virtio/vhost.c (revision c39f95dc)
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/blocker.h"
29 #include "sysemu/dma.h"
30 
31 /* enabled until disconnected backend stabilizes */
32 #define _VHOST_DEBUG 1
33 
34 #ifdef _VHOST_DEBUG
35 #define VHOST_OPS_DEBUG(fmt, ...) \
36     do { error_report(fmt ": %s (%d)", ## __VA_ARGS__, \
37                       strerror(errno), errno); } while (0)
38 #else
39 #define VHOST_OPS_DEBUG(fmt, ...) \
40     do { } while (0)
41 #endif
42 
43 static struct vhost_log *vhost_log;
44 static struct vhost_log *vhost_log_shm;
45 
46 static unsigned int used_memslots;
47 static QLIST_HEAD(, vhost_dev) vhost_devices =
48     QLIST_HEAD_INITIALIZER(vhost_devices);
49 
50 bool vhost_has_free_slot(void)
51 {
52     unsigned int slots_limit = ~0U;
53     struct vhost_dev *hdev;
54 
55     QLIST_FOREACH(hdev, &vhost_devices, entry) {
56         unsigned int r = hdev->vhost_ops->vhost_backend_memslots_limit(hdev);
57         slots_limit = MIN(slots_limit, r);
58     }
59     return slots_limit > used_memslots;
60 }
61 
62 static void vhost_dev_sync_region(struct vhost_dev *dev,
63                                   MemoryRegionSection *section,
64                                   uint64_t mfirst, uint64_t mlast,
65                                   uint64_t rfirst, uint64_t rlast)
66 {
67     vhost_log_chunk_t *log = dev->log->log;
68 
69     uint64_t start = MAX(mfirst, rfirst);
70     uint64_t end = MIN(mlast, rlast);
71     vhost_log_chunk_t *from = log + start / VHOST_LOG_CHUNK;
72     vhost_log_chunk_t *to = log + end / VHOST_LOG_CHUNK + 1;
73     uint64_t addr = QEMU_ALIGN_DOWN(start, VHOST_LOG_CHUNK);
74 
75     if (end < start) {
76         return;
77     }
78     assert(end / VHOST_LOG_CHUNK < dev->log_size);
79     assert(start / VHOST_LOG_CHUNK < dev->log_size);
80 
81     for (;from < to; ++from) {
82         vhost_log_chunk_t log;
83         /* We first check with non-atomic: much cheaper,
84          * and we expect non-dirty to be the common case. */
85         if (!*from) {
86             addr += VHOST_LOG_CHUNK;
87             continue;
88         }
89         /* Data must be read atomically. We don't really need barrier semantics
90          * but it's easier to use atomic_* than roll our own. */
91         log = atomic_xchg(from, 0);
92         while (log) {
93             int bit = ctzl(log);
94             hwaddr page_addr;
95             hwaddr section_offset;
96             hwaddr mr_offset;
97             page_addr = addr + bit * VHOST_LOG_PAGE;
98             section_offset = page_addr - section->offset_within_address_space;
99             mr_offset = section_offset + section->offset_within_region;
100             memory_region_set_dirty(section->mr, mr_offset, VHOST_LOG_PAGE);
101             log &= ~(0x1ull << bit);
102         }
103         addr += VHOST_LOG_CHUNK;
104     }
105 }
106 
107 static int vhost_sync_dirty_bitmap(struct vhost_dev *dev,
108                                    MemoryRegionSection *section,
109                                    hwaddr first,
110                                    hwaddr last)
111 {
112     int i;
113     hwaddr start_addr;
114     hwaddr end_addr;
115 
116     if (!dev->log_enabled || !dev->started) {
117         return 0;
118     }
119     start_addr = section->offset_within_address_space;
120     end_addr = range_get_last(start_addr, int128_get64(section->size));
121     start_addr = MAX(first, start_addr);
122     end_addr = MIN(last, end_addr);
123 
124     for (i = 0; i < dev->mem->nregions; ++i) {
125         struct vhost_memory_region *reg = dev->mem->regions + i;
126         vhost_dev_sync_region(dev, section, start_addr, end_addr,
127                               reg->guest_phys_addr,
128                               range_get_last(reg->guest_phys_addr,
129                                              reg->memory_size));
130     }
131     for (i = 0; i < dev->nvqs; ++i) {
132         struct vhost_virtqueue *vq = dev->vqs + i;
133         vhost_dev_sync_region(dev, section, start_addr, end_addr, vq->used_phys,
134                               range_get_last(vq->used_phys, vq->used_size));
135     }
136     return 0;
137 }
138 
139 static void vhost_log_sync(MemoryListener *listener,
140                           MemoryRegionSection *section)
141 {
142     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
143                                          memory_listener);
144     vhost_sync_dirty_bitmap(dev, section, 0x0, ~0x0ULL);
145 }
146 
147 static void vhost_log_sync_range(struct vhost_dev *dev,
148                                  hwaddr first, hwaddr last)
149 {
150     int i;
151     /* FIXME: this is N^2 in number of sections */
152     for (i = 0; i < dev->n_mem_sections; ++i) {
153         MemoryRegionSection *section = &dev->mem_sections[i];
154         vhost_sync_dirty_bitmap(dev, section, first, last);
155     }
156 }
157 
158 /* Assign/unassign. Keep an unsorted array of non-overlapping
159  * memory regions in dev->mem. */
160 static void vhost_dev_unassign_memory(struct vhost_dev *dev,
161                                       uint64_t start_addr,
162                                       uint64_t size)
163 {
164     int from, to, n = dev->mem->nregions;
165     /* Track overlapping/split regions for sanity checking. */
166     int overlap_start = 0, overlap_end = 0, overlap_middle = 0, split = 0;
167 
168     for (from = 0, to = 0; from < n; ++from, ++to) {
169         struct vhost_memory_region *reg = dev->mem->regions + to;
170         uint64_t reglast;
171         uint64_t memlast;
172         uint64_t change;
173 
174         /* clone old region */
175         if (to != from) {
176             memcpy(reg, dev->mem->regions + from, sizeof *reg);
177         }
178 
179         /* No overlap is simple */
180         if (!ranges_overlap(reg->guest_phys_addr, reg->memory_size,
181                             start_addr, size)) {
182             continue;
183         }
184 
185         /* Split 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(!split);
189 
190         reglast = range_get_last(reg->guest_phys_addr, reg->memory_size);
191         memlast = range_get_last(start_addr, size);
192 
193         /* Remove whole region */
194         if (start_addr <= reg->guest_phys_addr && memlast >= reglast) {
195             --dev->mem->nregions;
196             --to;
197             ++overlap_middle;
198             continue;
199         }
200 
201         /* Shrink region */
202         if (memlast >= reglast) {
203             reg->memory_size = start_addr - reg->guest_phys_addr;
204             assert(reg->memory_size);
205             assert(!overlap_end);
206             ++overlap_end;
207             continue;
208         }
209 
210         /* Shift region */
211         if (start_addr <= reg->guest_phys_addr) {
212             change = memlast + 1 - reg->guest_phys_addr;
213             reg->memory_size -= change;
214             reg->guest_phys_addr += change;
215             reg->userspace_addr += change;
216             assert(reg->memory_size);
217             assert(!overlap_start);
218             ++overlap_start;
219             continue;
220         }
221 
222         /* This only happens if supplied region
223          * is in the middle of an existing one. Thus it can not
224          * overlap with any other existing region. */
225         assert(!overlap_start);
226         assert(!overlap_end);
227         assert(!overlap_middle);
228         /* Split region: shrink first part, shift second part. */
229         memcpy(dev->mem->regions + n, reg, sizeof *reg);
230         reg->memory_size = start_addr - reg->guest_phys_addr;
231         assert(reg->memory_size);
232         change = memlast + 1 - reg->guest_phys_addr;
233         reg = dev->mem->regions + n;
234         reg->memory_size -= change;
235         assert(reg->memory_size);
236         reg->guest_phys_addr += change;
237         reg->userspace_addr += change;
238         /* Never add more than 1 region */
239         assert(dev->mem->nregions == n);
240         ++dev->mem->nregions;
241         ++split;
242     }
243 }
244 
245 /* Called after unassign, so no regions overlap the given range. */
246 static void vhost_dev_assign_memory(struct vhost_dev *dev,
247                                     uint64_t start_addr,
248                                     uint64_t size,
249                                     uint64_t uaddr)
250 {
251     int from, to;
252     struct vhost_memory_region *merged = NULL;
253     for (from = 0, to = 0; from < dev->mem->nregions; ++from, ++to) {
254         struct vhost_memory_region *reg = dev->mem->regions + to;
255         uint64_t prlast, urlast;
256         uint64_t pmlast, umlast;
257         uint64_t s, e, u;
258 
259         /* clone old region */
260         if (to != from) {
261             memcpy(reg, dev->mem->regions + from, sizeof *reg);
262         }
263         prlast = range_get_last(reg->guest_phys_addr, reg->memory_size);
264         pmlast = range_get_last(start_addr, size);
265         urlast = range_get_last(reg->userspace_addr, reg->memory_size);
266         umlast = range_get_last(uaddr, size);
267 
268         /* check for overlapping regions: should never happen. */
269         assert(prlast < start_addr || pmlast < reg->guest_phys_addr);
270         /* Not an adjacent or overlapping region - do not merge. */
271         if ((prlast + 1 != start_addr || urlast + 1 != uaddr) &&
272             (pmlast + 1 != reg->guest_phys_addr ||
273              umlast + 1 != reg->userspace_addr)) {
274             continue;
275         }
276 
277         if (dev->vhost_ops->vhost_backend_can_merge &&
278             !dev->vhost_ops->vhost_backend_can_merge(dev, uaddr, size,
279                                                      reg->userspace_addr,
280                                                      reg->memory_size)) {
281             continue;
282         }
283 
284         if (merged) {
285             --to;
286             assert(to >= 0);
287         } else {
288             merged = reg;
289         }
290         u = MIN(uaddr, reg->userspace_addr);
291         s = MIN(start_addr, reg->guest_phys_addr);
292         e = MAX(pmlast, prlast);
293         uaddr = merged->userspace_addr = u;
294         start_addr = merged->guest_phys_addr = s;
295         size = merged->memory_size = e - s + 1;
296         assert(merged->memory_size);
297     }
298 
299     if (!merged) {
300         struct vhost_memory_region *reg = dev->mem->regions + to;
301         memset(reg, 0, sizeof *reg);
302         reg->memory_size = size;
303         assert(reg->memory_size);
304         reg->guest_phys_addr = start_addr;
305         reg->userspace_addr = uaddr;
306         ++to;
307     }
308     assert(to <= dev->mem->nregions + 1);
309     dev->mem->nregions = to;
310 }
311 
312 static uint64_t vhost_get_log_size(struct vhost_dev *dev)
313 {
314     uint64_t log_size = 0;
315     int i;
316     for (i = 0; i < dev->mem->nregions; ++i) {
317         struct vhost_memory_region *reg = dev->mem->regions + i;
318         uint64_t last = range_get_last(reg->guest_phys_addr,
319                                        reg->memory_size);
320         log_size = MAX(log_size, last / VHOST_LOG_CHUNK + 1);
321     }
322     for (i = 0; i < dev->nvqs; ++i) {
323         struct vhost_virtqueue *vq = dev->vqs + i;
324         uint64_t last = vq->used_phys + vq->used_size - 1;
325         log_size = MAX(log_size, last / VHOST_LOG_CHUNK + 1);
326     }
327     return log_size;
328 }
329 
330 static struct vhost_log *vhost_log_alloc(uint64_t size, bool share)
331 {
332     struct vhost_log *log;
333     uint64_t logsize = size * sizeof(*(log->log));
334     int fd = -1;
335 
336     log = g_new0(struct vhost_log, 1);
337     if (share) {
338         log->log = qemu_memfd_alloc("vhost-log", logsize,
339                                     F_SEAL_GROW | F_SEAL_SHRINK | F_SEAL_SEAL,
340                                     &fd);
341         memset(log->log, 0, logsize);
342     } else {
343         log->log = g_malloc0(logsize);
344     }
345 
346     log->size = size;
347     log->refcnt = 1;
348     log->fd = fd;
349 
350     return log;
351 }
352 
353 static struct vhost_log *vhost_log_get(uint64_t size, bool share)
354 {
355     struct vhost_log *log = share ? vhost_log_shm : vhost_log;
356 
357     if (!log || log->size != size) {
358         log = vhost_log_alloc(size, share);
359         if (share) {
360             vhost_log_shm = log;
361         } else {
362             vhost_log = log;
363         }
364     } else {
365         ++log->refcnt;
366     }
367 
368     return log;
369 }
370 
371 static void vhost_log_put(struct vhost_dev *dev, bool sync)
372 {
373     struct vhost_log *log = dev->log;
374 
375     if (!log) {
376         return;
377     }
378 
379     --log->refcnt;
380     if (log->refcnt == 0) {
381         /* Sync only the range covered by the old log */
382         if (dev->log_size && sync) {
383             vhost_log_sync_range(dev, 0, dev->log_size * VHOST_LOG_CHUNK - 1);
384         }
385 
386         if (vhost_log == log) {
387             g_free(log->log);
388             vhost_log = NULL;
389         } else if (vhost_log_shm == log) {
390             qemu_memfd_free(log->log, log->size * sizeof(*(log->log)),
391                             log->fd);
392             vhost_log_shm = NULL;
393         }
394 
395         g_free(log);
396     }
397 
398     dev->log = NULL;
399     dev->log_size = 0;
400 }
401 
402 static bool vhost_dev_log_is_shared(struct vhost_dev *dev)
403 {
404     return dev->vhost_ops->vhost_requires_shm_log &&
405            dev->vhost_ops->vhost_requires_shm_log(dev);
406 }
407 
408 static inline void vhost_dev_log_resize(struct vhost_dev *dev, uint64_t size)
409 {
410     struct vhost_log *log = vhost_log_get(size, vhost_dev_log_is_shared(dev));
411     uint64_t log_base = (uintptr_t)log->log;
412     int r;
413 
414     /* inform backend of log switching, this must be done before
415        releasing the current log, to ensure no logging is lost */
416     r = dev->vhost_ops->vhost_set_log_base(dev, log_base, log);
417     if (r < 0) {
418         VHOST_OPS_DEBUG("vhost_set_log_base failed");
419     }
420 
421     vhost_log_put(dev, true);
422     dev->log = log;
423     dev->log_size = size;
424 }
425 
426 static int vhost_dev_has_iommu(struct vhost_dev *dev)
427 {
428     VirtIODevice *vdev = dev->vdev;
429 
430     return virtio_host_has_feature(vdev, VIRTIO_F_IOMMU_PLATFORM);
431 }
432 
433 static void *vhost_memory_map(struct vhost_dev *dev, hwaddr addr,
434                               hwaddr *plen, int is_write)
435 {
436     if (!vhost_dev_has_iommu(dev)) {
437         return cpu_physical_memory_map(addr, plen, is_write);
438     } else {
439         return (void *)(uintptr_t)addr;
440     }
441 }
442 
443 static void vhost_memory_unmap(struct vhost_dev *dev, void *buffer,
444                                hwaddr len, int is_write,
445                                hwaddr access_len)
446 {
447     if (!vhost_dev_has_iommu(dev)) {
448         cpu_physical_memory_unmap(buffer, len, is_write, access_len);
449     }
450 }
451 
452 static int vhost_verify_ring_part_mapping(struct vhost_dev *dev,
453                                           void *part,
454                                           uint64_t part_addr,
455                                           uint64_t part_size,
456                                           uint64_t start_addr,
457                                           uint64_t size)
458 {
459     hwaddr l;
460     void *p;
461     int r = 0;
462 
463     if (!ranges_overlap(start_addr, size, part_addr, part_size)) {
464         return 0;
465     }
466     l = part_size;
467     p = vhost_memory_map(dev, part_addr, &l, 1);
468     if (!p || l != part_size) {
469         r = -ENOMEM;
470     }
471     if (p != part) {
472         r = -EBUSY;
473     }
474     vhost_memory_unmap(dev, p, l, 0, 0);
475     return r;
476 }
477 
478 static int vhost_verify_ring_mappings(struct vhost_dev *dev,
479                                       uint64_t start_addr,
480                                       uint64_t size)
481 {
482     int i, j;
483     int r = 0;
484     const char *part_name[] = {
485         "descriptor table",
486         "available ring",
487         "used ring"
488     };
489 
490     for (i = 0; i < dev->nvqs; ++i) {
491         struct vhost_virtqueue *vq = dev->vqs + i;
492 
493         j = 0;
494         r = vhost_verify_ring_part_mapping(dev, vq->desc, vq->desc_phys,
495                                            vq->desc_size, start_addr, size);
496         if (!r) {
497             break;
498         }
499 
500         j++;
501         r = vhost_verify_ring_part_mapping(dev, vq->avail, vq->avail_phys,
502                                            vq->avail_size, start_addr, size);
503         if (!r) {
504             break;
505         }
506 
507         j++;
508         r = vhost_verify_ring_part_mapping(dev, vq->used, vq->used_phys,
509                                            vq->used_size, start_addr, size);
510         if (!r) {
511             break;
512         }
513     }
514 
515     if (r == -ENOMEM) {
516         error_report("Unable to map %s for ring %d", part_name[j], i);
517     } else if (r == -EBUSY) {
518         error_report("%s relocated for ring %d", part_name[j], i);
519     }
520     return r;
521 }
522 
523 static struct vhost_memory_region *vhost_dev_find_reg(struct vhost_dev *dev,
524 						      uint64_t start_addr,
525 						      uint64_t size)
526 {
527     int i, n = dev->mem->nregions;
528     for (i = 0; i < n; ++i) {
529         struct vhost_memory_region *reg = dev->mem->regions + i;
530         if (ranges_overlap(reg->guest_phys_addr, reg->memory_size,
531                            start_addr, size)) {
532             return reg;
533         }
534     }
535     return NULL;
536 }
537 
538 static bool vhost_dev_cmp_memory(struct vhost_dev *dev,
539                                  uint64_t start_addr,
540                                  uint64_t size,
541                                  uint64_t uaddr)
542 {
543     struct vhost_memory_region *reg = vhost_dev_find_reg(dev, start_addr, size);
544     uint64_t reglast;
545     uint64_t memlast;
546 
547     if (!reg) {
548         return true;
549     }
550 
551     reglast = range_get_last(reg->guest_phys_addr, reg->memory_size);
552     memlast = range_get_last(start_addr, size);
553 
554     /* Need to extend region? */
555     if (start_addr < reg->guest_phys_addr || memlast > reglast) {
556         return true;
557     }
558     /* userspace_addr changed? */
559     return uaddr != reg->userspace_addr + start_addr - reg->guest_phys_addr;
560 }
561 
562 static void vhost_set_memory(MemoryListener *listener,
563                              MemoryRegionSection *section,
564                              bool add)
565 {
566     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
567                                          memory_listener);
568     hwaddr start_addr = section->offset_within_address_space;
569     ram_addr_t size = int128_get64(section->size);
570     bool log_dirty =
571         memory_region_get_dirty_log_mask(section->mr) & ~(1 << DIRTY_MEMORY_MIGRATION);
572     int s = offsetof(struct vhost_memory, regions) +
573         (dev->mem->nregions + 1) * sizeof dev->mem->regions[0];
574     void *ram;
575 
576     dev->mem = g_realloc(dev->mem, s);
577 
578     if (log_dirty) {
579         add = false;
580     }
581 
582     assert(size);
583 
584     /* Optimize no-change case. At least cirrus_vga does this a lot at this time. */
585     ram = memory_region_get_ram_ptr(section->mr) + section->offset_within_region;
586     if (add) {
587         if (!vhost_dev_cmp_memory(dev, start_addr, size, (uintptr_t)ram)) {
588             /* Region exists with same address. Nothing to do. */
589             return;
590         }
591     } else {
592         if (!vhost_dev_find_reg(dev, start_addr, size)) {
593             /* Removing region that we don't access. Nothing to do. */
594             return;
595         }
596     }
597 
598     vhost_dev_unassign_memory(dev, start_addr, size);
599     if (add) {
600         /* Add given mapping, merging adjacent regions if any */
601         vhost_dev_assign_memory(dev, start_addr, size, (uintptr_t)ram);
602     } else {
603         /* Remove old mapping for this memory, if any. */
604         vhost_dev_unassign_memory(dev, start_addr, size);
605     }
606     dev->mem_changed_start_addr = MIN(dev->mem_changed_start_addr, start_addr);
607     dev->mem_changed_end_addr = MAX(dev->mem_changed_end_addr, start_addr + size - 1);
608     dev->memory_changed = true;
609     used_memslots = dev->mem->nregions;
610 }
611 
612 static bool vhost_section(MemoryRegionSection *section)
613 {
614     return memory_region_is_ram(section->mr) &&
615         !memory_region_is_rom(section->mr);
616 }
617 
618 static void vhost_begin(MemoryListener *listener)
619 {
620     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
621                                          memory_listener);
622     dev->mem_changed_end_addr = 0;
623     dev->mem_changed_start_addr = -1;
624 }
625 
626 static void vhost_commit(MemoryListener *listener)
627 {
628     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
629                                          memory_listener);
630     hwaddr start_addr = 0;
631     ram_addr_t size = 0;
632     uint64_t log_size;
633     int r;
634 
635     if (!dev->memory_changed) {
636         return;
637     }
638     if (!dev->started) {
639         return;
640     }
641     if (dev->mem_changed_start_addr > dev->mem_changed_end_addr) {
642         return;
643     }
644 
645     if (dev->started) {
646         start_addr = dev->mem_changed_start_addr;
647         size = dev->mem_changed_end_addr - dev->mem_changed_start_addr + 1;
648 
649         r = vhost_verify_ring_mappings(dev, start_addr, size);
650         assert(r >= 0);
651     }
652 
653     if (!dev->log_enabled) {
654         r = dev->vhost_ops->vhost_set_mem_table(dev, dev->mem);
655         if (r < 0) {
656             VHOST_OPS_DEBUG("vhost_set_mem_table failed");
657         }
658         dev->memory_changed = false;
659         return;
660     }
661     log_size = vhost_get_log_size(dev);
662     /* We allocate an extra 4K bytes to log,
663      * to reduce the * number of reallocations. */
664 #define VHOST_LOG_BUFFER (0x1000 / sizeof *dev->log)
665     /* To log more, must increase log size before table update. */
666     if (dev->log_size < log_size) {
667         vhost_dev_log_resize(dev, log_size + VHOST_LOG_BUFFER);
668     }
669     r = dev->vhost_ops->vhost_set_mem_table(dev, dev->mem);
670     if (r < 0) {
671         VHOST_OPS_DEBUG("vhost_set_mem_table failed");
672     }
673     /* To log less, can only decrease log size after table update. */
674     if (dev->log_size > log_size + VHOST_LOG_BUFFER) {
675         vhost_dev_log_resize(dev, log_size);
676     }
677     dev->memory_changed = false;
678 }
679 
680 static void vhost_region_add(MemoryListener *listener,
681                              MemoryRegionSection *section)
682 {
683     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
684                                          memory_listener);
685 
686     if (!vhost_section(section)) {
687         return;
688     }
689 
690     ++dev->n_mem_sections;
691     dev->mem_sections = g_renew(MemoryRegionSection, dev->mem_sections,
692                                 dev->n_mem_sections);
693     dev->mem_sections[dev->n_mem_sections - 1] = *section;
694     memory_region_ref(section->mr);
695     vhost_set_memory(listener, section, true);
696 }
697 
698 static void vhost_region_del(MemoryListener *listener,
699                              MemoryRegionSection *section)
700 {
701     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
702                                          memory_listener);
703     int i;
704 
705     if (!vhost_section(section)) {
706         return;
707     }
708 
709     vhost_set_memory(listener, section, false);
710     memory_region_unref(section->mr);
711     for (i = 0; i < dev->n_mem_sections; ++i) {
712         if (dev->mem_sections[i].offset_within_address_space
713             == section->offset_within_address_space) {
714             --dev->n_mem_sections;
715             memmove(&dev->mem_sections[i], &dev->mem_sections[i+1],
716                     (dev->n_mem_sections - i) * sizeof(*dev->mem_sections));
717             break;
718         }
719     }
720 }
721 
722 static void vhost_iommu_unmap_notify(IOMMUNotifier *n, IOMMUTLBEntry *iotlb)
723 {
724     struct vhost_iommu *iommu = container_of(n, struct vhost_iommu, n);
725     struct vhost_dev *hdev = iommu->hdev;
726     hwaddr iova = iotlb->iova + iommu->iommu_offset;
727 
728     if (vhost_backend_invalidate_device_iotlb(hdev, iova,
729                                               iotlb->addr_mask + 1)) {
730         error_report("Fail to invalidate device iotlb");
731     }
732 }
733 
734 static void vhost_iommu_region_add(MemoryListener *listener,
735                                    MemoryRegionSection *section)
736 {
737     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
738                                          iommu_listener);
739     struct vhost_iommu *iommu;
740     Int128 end;
741 
742     if (!memory_region_is_iommu(section->mr)) {
743         return;
744     }
745 
746     iommu = g_malloc0(sizeof(*iommu));
747     end = int128_add(int128_make64(section->offset_within_region),
748                      section->size);
749     end = int128_sub(end, int128_one());
750     iommu_notifier_init(&iommu->n, vhost_iommu_unmap_notify,
751                         IOMMU_NOTIFIER_UNMAP,
752                         section->offset_within_region,
753                         int128_get64(end));
754     iommu->mr = section->mr;
755     iommu->iommu_offset = section->offset_within_address_space -
756                           section->offset_within_region;
757     iommu->hdev = dev;
758     memory_region_register_iommu_notifier(section->mr, &iommu->n);
759     QLIST_INSERT_HEAD(&dev->iommu_list, iommu, iommu_next);
760     /* TODO: can replay help performance here? */
761 }
762 
763 static void vhost_iommu_region_del(MemoryListener *listener,
764                                    MemoryRegionSection *section)
765 {
766     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
767                                          iommu_listener);
768     struct vhost_iommu *iommu;
769 
770     if (!memory_region_is_iommu(section->mr)) {
771         return;
772     }
773 
774     QLIST_FOREACH(iommu, &dev->iommu_list, iommu_next) {
775         if (iommu->mr == section->mr &&
776             iommu->n.start == section->offset_within_region) {
777             memory_region_unregister_iommu_notifier(iommu->mr,
778                                                     &iommu->n);
779             QLIST_REMOVE(iommu, iommu_next);
780             g_free(iommu);
781             break;
782         }
783     }
784 }
785 
786 static void vhost_region_nop(MemoryListener *listener,
787                              MemoryRegionSection *section)
788 {
789 }
790 
791 static int vhost_virtqueue_set_addr(struct vhost_dev *dev,
792                                     struct vhost_virtqueue *vq,
793                                     unsigned idx, bool enable_log)
794 {
795     struct vhost_vring_addr addr = {
796         .index = idx,
797         .desc_user_addr = (uint64_t)(unsigned long)vq->desc,
798         .avail_user_addr = (uint64_t)(unsigned long)vq->avail,
799         .used_user_addr = (uint64_t)(unsigned long)vq->used,
800         .log_guest_addr = vq->used_phys,
801         .flags = enable_log ? (1 << VHOST_VRING_F_LOG) : 0,
802     };
803     int r = dev->vhost_ops->vhost_set_vring_addr(dev, &addr);
804     if (r < 0) {
805         VHOST_OPS_DEBUG("vhost_set_vring_addr failed");
806         return -errno;
807     }
808     return 0;
809 }
810 
811 static int vhost_dev_set_features(struct vhost_dev *dev,
812                                   bool enable_log)
813 {
814     uint64_t features = dev->acked_features;
815     int r;
816     if (enable_log) {
817         features |= 0x1ULL << VHOST_F_LOG_ALL;
818     }
819     r = dev->vhost_ops->vhost_set_features(dev, features);
820     if (r < 0) {
821         VHOST_OPS_DEBUG("vhost_set_features failed");
822     }
823     return r < 0 ? -errno : 0;
824 }
825 
826 static int vhost_dev_set_log(struct vhost_dev *dev, bool enable_log)
827 {
828     int r, i, idx;
829     r = vhost_dev_set_features(dev, enable_log);
830     if (r < 0) {
831         goto err_features;
832     }
833     for (i = 0; i < dev->nvqs; ++i) {
834         idx = dev->vhost_ops->vhost_get_vq_index(dev, dev->vq_index + i);
835         r = vhost_virtqueue_set_addr(dev, dev->vqs + i, idx,
836                                      enable_log);
837         if (r < 0) {
838             goto err_vq;
839         }
840     }
841     return 0;
842 err_vq:
843     for (; i >= 0; --i) {
844         idx = dev->vhost_ops->vhost_get_vq_index(dev, dev->vq_index + i);
845         vhost_virtqueue_set_addr(dev, dev->vqs + i, idx,
846                                  dev->log_enabled);
847     }
848     vhost_dev_set_features(dev, dev->log_enabled);
849 err_features:
850     return r;
851 }
852 
853 static int vhost_migration_log(MemoryListener *listener, int enable)
854 {
855     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
856                                          memory_listener);
857     int r;
858     if (!!enable == dev->log_enabled) {
859         return 0;
860     }
861     if (!dev->started) {
862         dev->log_enabled = enable;
863         return 0;
864     }
865     if (!enable) {
866         r = vhost_dev_set_log(dev, false);
867         if (r < 0) {
868             return r;
869         }
870         vhost_log_put(dev, false);
871     } else {
872         vhost_dev_log_resize(dev, vhost_get_log_size(dev));
873         r = vhost_dev_set_log(dev, true);
874         if (r < 0) {
875             return r;
876         }
877     }
878     dev->log_enabled = enable;
879     return 0;
880 }
881 
882 static void vhost_log_global_start(MemoryListener *listener)
883 {
884     int r;
885 
886     r = vhost_migration_log(listener, true);
887     if (r < 0) {
888         abort();
889     }
890 }
891 
892 static void vhost_log_global_stop(MemoryListener *listener)
893 {
894     int r;
895 
896     r = vhost_migration_log(listener, false);
897     if (r < 0) {
898         abort();
899     }
900 }
901 
902 static void vhost_log_start(MemoryListener *listener,
903                             MemoryRegionSection *section,
904                             int old, int new)
905 {
906     /* FIXME: implement */
907 }
908 
909 static void vhost_log_stop(MemoryListener *listener,
910                            MemoryRegionSection *section,
911                            int old, int new)
912 {
913     /* FIXME: implement */
914 }
915 
916 /* The vhost driver natively knows how to handle the vrings of non
917  * cross-endian legacy devices and modern devices. Only legacy devices
918  * exposed to a bi-endian guest may require the vhost driver to use a
919  * specific endianness.
920  */
921 static inline bool vhost_needs_vring_endian(VirtIODevice *vdev)
922 {
923     if (virtio_vdev_has_feature(vdev, VIRTIO_F_VERSION_1)) {
924         return false;
925     }
926 #ifdef HOST_WORDS_BIGENDIAN
927     return vdev->device_endian == VIRTIO_DEVICE_ENDIAN_LITTLE;
928 #else
929     return vdev->device_endian == VIRTIO_DEVICE_ENDIAN_BIG;
930 #endif
931 }
932 
933 static int vhost_virtqueue_set_vring_endian_legacy(struct vhost_dev *dev,
934                                                    bool is_big_endian,
935                                                    int vhost_vq_index)
936 {
937     struct vhost_vring_state s = {
938         .index = vhost_vq_index,
939         .num = is_big_endian
940     };
941 
942     if (!dev->vhost_ops->vhost_set_vring_endian(dev, &s)) {
943         return 0;
944     }
945 
946     VHOST_OPS_DEBUG("vhost_set_vring_endian failed");
947     if (errno == ENOTTY) {
948         error_report("vhost does not support cross-endian");
949         return -ENOSYS;
950     }
951 
952     return -errno;
953 }
954 
955 static int vhost_memory_region_lookup(struct vhost_dev *hdev,
956                                       uint64_t gpa, uint64_t *uaddr,
957                                       uint64_t *len)
958 {
959     int i;
960 
961     for (i = 0; i < hdev->mem->nregions; i++) {
962         struct vhost_memory_region *reg = hdev->mem->regions + i;
963 
964         if (gpa >= reg->guest_phys_addr &&
965             reg->guest_phys_addr + reg->memory_size > gpa) {
966             *uaddr = reg->userspace_addr + gpa - reg->guest_phys_addr;
967             *len = reg->guest_phys_addr + reg->memory_size - gpa;
968             return 0;
969         }
970     }
971 
972     return -EFAULT;
973 }
974 
975 int vhost_device_iotlb_miss(struct vhost_dev *dev, uint64_t iova, int write)
976 {
977     IOMMUTLBEntry iotlb;
978     uint64_t uaddr, len;
979     int ret = -EFAULT;
980 
981     rcu_read_lock();
982 
983     iotlb = address_space_get_iotlb_entry(dev->vdev->dma_as,
984                                           iova, write);
985     if (iotlb.target_as != NULL) {
986         ret = vhost_memory_region_lookup(dev, iotlb.translated_addr,
987                                          &uaddr, &len);
988         if (ret) {
989             error_report("Fail to lookup the translated address "
990                          "%"PRIx64, iotlb.translated_addr);
991             goto out;
992         }
993 
994         len = MIN(iotlb.addr_mask + 1, len);
995         iova = iova & ~iotlb.addr_mask;
996 
997         ret = vhost_backend_update_device_iotlb(dev, iova, uaddr,
998                                                 len, iotlb.perm);
999         if (ret) {
1000             error_report("Fail to update device iotlb");
1001             goto out;
1002         }
1003     }
1004 out:
1005     rcu_read_unlock();
1006 
1007     return ret;
1008 }
1009 
1010 static int vhost_virtqueue_start(struct vhost_dev *dev,
1011                                 struct VirtIODevice *vdev,
1012                                 struct vhost_virtqueue *vq,
1013                                 unsigned idx)
1014 {
1015     BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev)));
1016     VirtioBusState *vbus = VIRTIO_BUS(qbus);
1017     VirtioBusClass *k = VIRTIO_BUS_GET_CLASS(vbus);
1018     hwaddr s, l, a;
1019     int r;
1020     int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, idx);
1021     struct vhost_vring_file file = {
1022         .index = vhost_vq_index
1023     };
1024     struct vhost_vring_state state = {
1025         .index = vhost_vq_index
1026     };
1027     struct VirtQueue *vvq = virtio_get_queue(vdev, idx);
1028 
1029 
1030     vq->num = state.num = virtio_queue_get_num(vdev, idx);
1031     r = dev->vhost_ops->vhost_set_vring_num(dev, &state);
1032     if (r) {
1033         VHOST_OPS_DEBUG("vhost_set_vring_num failed");
1034         return -errno;
1035     }
1036 
1037     state.num = virtio_queue_get_last_avail_idx(vdev, idx);
1038     r = dev->vhost_ops->vhost_set_vring_base(dev, &state);
1039     if (r) {
1040         VHOST_OPS_DEBUG("vhost_set_vring_base failed");
1041         return -errno;
1042     }
1043 
1044     if (vhost_needs_vring_endian(vdev)) {
1045         r = vhost_virtqueue_set_vring_endian_legacy(dev,
1046                                                     virtio_is_big_endian(vdev),
1047                                                     vhost_vq_index);
1048         if (r) {
1049             return -errno;
1050         }
1051     }
1052 
1053     vq->desc_size = s = l = virtio_queue_get_desc_size(vdev, idx);
1054     vq->desc_phys = a = virtio_queue_get_desc_addr(vdev, idx);
1055     vq->desc = vhost_memory_map(dev, a, &l, 0);
1056     if (!vq->desc || l != s) {
1057         r = -ENOMEM;
1058         goto fail_alloc_desc;
1059     }
1060     vq->avail_size = s = l = virtio_queue_get_avail_size(vdev, idx);
1061     vq->avail_phys = a = virtio_queue_get_avail_addr(vdev, idx);
1062     vq->avail = vhost_memory_map(dev, a, &l, 0);
1063     if (!vq->avail || l != s) {
1064         r = -ENOMEM;
1065         goto fail_alloc_avail;
1066     }
1067     vq->used_size = s = l = virtio_queue_get_used_size(vdev, idx);
1068     vq->used_phys = a = virtio_queue_get_used_addr(vdev, idx);
1069     vq->used = vhost_memory_map(dev, a, &l, 1);
1070     if (!vq->used || l != s) {
1071         r = -ENOMEM;
1072         goto fail_alloc_used;
1073     }
1074 
1075     r = vhost_virtqueue_set_addr(dev, vq, vhost_vq_index, dev->log_enabled);
1076     if (r < 0) {
1077         r = -errno;
1078         goto fail_alloc;
1079     }
1080 
1081     file.fd = event_notifier_get_fd(virtio_queue_get_host_notifier(vvq));
1082     r = dev->vhost_ops->vhost_set_vring_kick(dev, &file);
1083     if (r) {
1084         VHOST_OPS_DEBUG("vhost_set_vring_kick failed");
1085         r = -errno;
1086         goto fail_kick;
1087     }
1088 
1089     /* Clear and discard previous events if any. */
1090     event_notifier_test_and_clear(&vq->masked_notifier);
1091 
1092     /* Init vring in unmasked state, unless guest_notifier_mask
1093      * will do it later.
1094      */
1095     if (!vdev->use_guest_notifier_mask) {
1096         /* TODO: check and handle errors. */
1097         vhost_virtqueue_mask(dev, vdev, idx, false);
1098     }
1099 
1100     if (k->query_guest_notifiers &&
1101         k->query_guest_notifiers(qbus->parent) &&
1102         virtio_queue_vector(vdev, idx) == VIRTIO_NO_VECTOR) {
1103         file.fd = -1;
1104         r = dev->vhost_ops->vhost_set_vring_call(dev, &file);
1105         if (r) {
1106             goto fail_vector;
1107         }
1108     }
1109 
1110     return 0;
1111 
1112 fail_vector:
1113 fail_kick:
1114 fail_alloc:
1115     vhost_memory_unmap(dev, vq->used, virtio_queue_get_used_size(vdev, idx),
1116                        0, 0);
1117 fail_alloc_used:
1118     vhost_memory_unmap(dev, vq->avail, virtio_queue_get_avail_size(vdev, idx),
1119                        0, 0);
1120 fail_alloc_avail:
1121     vhost_memory_unmap(dev, vq->desc, virtio_queue_get_desc_size(vdev, idx),
1122                        0, 0);
1123 fail_alloc_desc:
1124     return r;
1125 }
1126 
1127 static void vhost_virtqueue_stop(struct vhost_dev *dev,
1128                                     struct VirtIODevice *vdev,
1129                                     struct vhost_virtqueue *vq,
1130                                     unsigned idx)
1131 {
1132     int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, idx);
1133     struct vhost_vring_state state = {
1134         .index = vhost_vq_index,
1135     };
1136     int r;
1137 
1138     r = dev->vhost_ops->vhost_get_vring_base(dev, &state);
1139     if (r < 0) {
1140         VHOST_OPS_DEBUG("vhost VQ %d ring restore failed: %d", idx, r);
1141     } else {
1142         virtio_queue_set_last_avail_idx(vdev, idx, state.num);
1143     }
1144     virtio_queue_invalidate_signalled_used(vdev, idx);
1145     virtio_queue_update_used_idx(vdev, idx);
1146 
1147     /* In the cross-endian case, we need to reset the vring endianness to
1148      * native as legacy devices expect so by default.
1149      */
1150     if (vhost_needs_vring_endian(vdev)) {
1151         vhost_virtqueue_set_vring_endian_legacy(dev,
1152                                                 !virtio_is_big_endian(vdev),
1153                                                 vhost_vq_index);
1154     }
1155 
1156     vhost_memory_unmap(dev, vq->used, virtio_queue_get_used_size(vdev, idx),
1157                        1, virtio_queue_get_used_size(vdev, idx));
1158     vhost_memory_unmap(dev, vq->avail, virtio_queue_get_avail_size(vdev, idx),
1159                        0, virtio_queue_get_avail_size(vdev, idx));
1160     vhost_memory_unmap(dev, vq->desc, virtio_queue_get_desc_size(vdev, idx),
1161                        0, virtio_queue_get_desc_size(vdev, idx));
1162 }
1163 
1164 static void vhost_eventfd_add(MemoryListener *listener,
1165                               MemoryRegionSection *section,
1166                               bool match_data, uint64_t data, EventNotifier *e)
1167 {
1168 }
1169 
1170 static void vhost_eventfd_del(MemoryListener *listener,
1171                               MemoryRegionSection *section,
1172                               bool match_data, uint64_t data, EventNotifier *e)
1173 {
1174 }
1175 
1176 static int vhost_virtqueue_set_busyloop_timeout(struct vhost_dev *dev,
1177                                                 int n, uint32_t timeout)
1178 {
1179     int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, n);
1180     struct vhost_vring_state state = {
1181         .index = vhost_vq_index,
1182         .num = timeout,
1183     };
1184     int r;
1185 
1186     if (!dev->vhost_ops->vhost_set_vring_busyloop_timeout) {
1187         return -EINVAL;
1188     }
1189 
1190     r = dev->vhost_ops->vhost_set_vring_busyloop_timeout(dev, &state);
1191     if (r) {
1192         VHOST_OPS_DEBUG("vhost_set_vring_busyloop_timeout failed");
1193         return r;
1194     }
1195 
1196     return 0;
1197 }
1198 
1199 static int vhost_virtqueue_init(struct vhost_dev *dev,
1200                                 struct vhost_virtqueue *vq, int n)
1201 {
1202     int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, n);
1203     struct vhost_vring_file file = {
1204         .index = vhost_vq_index,
1205     };
1206     int r = event_notifier_init(&vq->masked_notifier, 0);
1207     if (r < 0) {
1208         return r;
1209     }
1210 
1211     file.fd = event_notifier_get_fd(&vq->masked_notifier);
1212     r = dev->vhost_ops->vhost_set_vring_call(dev, &file);
1213     if (r) {
1214         VHOST_OPS_DEBUG("vhost_set_vring_call failed");
1215         r = -errno;
1216         goto fail_call;
1217     }
1218 
1219     vq->dev = dev;
1220 
1221     return 0;
1222 fail_call:
1223     event_notifier_cleanup(&vq->masked_notifier);
1224     return r;
1225 }
1226 
1227 static void vhost_virtqueue_cleanup(struct vhost_virtqueue *vq)
1228 {
1229     event_notifier_cleanup(&vq->masked_notifier);
1230 }
1231 
1232 int vhost_dev_init(struct vhost_dev *hdev, void *opaque,
1233                    VhostBackendType backend_type, uint32_t busyloop_timeout)
1234 {
1235     uint64_t features;
1236     int i, r, n_initialized_vqs = 0;
1237     Error *local_err = NULL;
1238 
1239     hdev->vdev = NULL;
1240     hdev->migration_blocker = NULL;
1241 
1242     r = vhost_set_backend_type(hdev, backend_type);
1243     assert(r >= 0);
1244 
1245     r = hdev->vhost_ops->vhost_backend_init(hdev, opaque);
1246     if (r < 0) {
1247         goto fail;
1248     }
1249 
1250     if (used_memslots > hdev->vhost_ops->vhost_backend_memslots_limit(hdev)) {
1251         error_report("vhost backend memory slots limit is less"
1252                 " than current number of present memory slots");
1253         r = -1;
1254         goto fail;
1255     }
1256 
1257     r = hdev->vhost_ops->vhost_set_owner(hdev);
1258     if (r < 0) {
1259         VHOST_OPS_DEBUG("vhost_set_owner failed");
1260         goto fail;
1261     }
1262 
1263     r = hdev->vhost_ops->vhost_get_features(hdev, &features);
1264     if (r < 0) {
1265         VHOST_OPS_DEBUG("vhost_get_features failed");
1266         goto fail;
1267     }
1268 
1269     for (i = 0; i < hdev->nvqs; ++i, ++n_initialized_vqs) {
1270         r = vhost_virtqueue_init(hdev, hdev->vqs + i, hdev->vq_index + i);
1271         if (r < 0) {
1272             goto fail;
1273         }
1274     }
1275 
1276     if (busyloop_timeout) {
1277         for (i = 0; i < hdev->nvqs; ++i) {
1278             r = vhost_virtqueue_set_busyloop_timeout(hdev, hdev->vq_index + i,
1279                                                      busyloop_timeout);
1280             if (r < 0) {
1281                 goto fail_busyloop;
1282             }
1283         }
1284     }
1285 
1286     hdev->features = features;
1287 
1288     hdev->memory_listener = (MemoryListener) {
1289         .begin = vhost_begin,
1290         .commit = vhost_commit,
1291         .region_add = vhost_region_add,
1292         .region_del = vhost_region_del,
1293         .region_nop = vhost_region_nop,
1294         .log_start = vhost_log_start,
1295         .log_stop = vhost_log_stop,
1296         .log_sync = vhost_log_sync,
1297         .log_global_start = vhost_log_global_start,
1298         .log_global_stop = vhost_log_global_stop,
1299         .eventfd_add = vhost_eventfd_add,
1300         .eventfd_del = vhost_eventfd_del,
1301         .priority = 10
1302     };
1303 
1304     hdev->iommu_listener = (MemoryListener) {
1305         .region_add = vhost_iommu_region_add,
1306         .region_del = vhost_iommu_region_del,
1307     };
1308 
1309     if (hdev->migration_blocker == NULL) {
1310         if (!(hdev->features & (0x1ULL << VHOST_F_LOG_ALL))) {
1311             error_setg(&hdev->migration_blocker,
1312                        "Migration disabled: vhost lacks VHOST_F_LOG_ALL feature.");
1313         } else if (vhost_dev_log_is_shared(hdev) && !qemu_memfd_check()) {
1314             error_setg(&hdev->migration_blocker,
1315                        "Migration disabled: failed to allocate shared memory");
1316         }
1317     }
1318 
1319     if (hdev->migration_blocker != NULL) {
1320         r = migrate_add_blocker(hdev->migration_blocker, &local_err);
1321         if (local_err) {
1322             error_report_err(local_err);
1323             error_free(hdev->migration_blocker);
1324             goto fail_busyloop;
1325         }
1326     }
1327 
1328     hdev->mem = g_malloc0(offsetof(struct vhost_memory, regions));
1329     hdev->n_mem_sections = 0;
1330     hdev->mem_sections = NULL;
1331     hdev->log = NULL;
1332     hdev->log_size = 0;
1333     hdev->log_enabled = false;
1334     hdev->started = false;
1335     hdev->memory_changed = false;
1336     memory_listener_register(&hdev->memory_listener, &address_space_memory);
1337     QLIST_INSERT_HEAD(&vhost_devices, hdev, entry);
1338     return 0;
1339 
1340 fail_busyloop:
1341     while (--i >= 0) {
1342         vhost_virtqueue_set_busyloop_timeout(hdev, hdev->vq_index + i, 0);
1343     }
1344 fail:
1345     hdev->nvqs = n_initialized_vqs;
1346     vhost_dev_cleanup(hdev);
1347     return r;
1348 }
1349 
1350 void vhost_dev_cleanup(struct vhost_dev *hdev)
1351 {
1352     int i;
1353 
1354     for (i = 0; i < hdev->nvqs; ++i) {
1355         vhost_virtqueue_cleanup(hdev->vqs + i);
1356     }
1357     if (hdev->mem) {
1358         /* those are only safe after successful init */
1359         memory_listener_unregister(&hdev->memory_listener);
1360         for (i = 0; i < hdev->n_mem_sections; ++i) {
1361             MemoryRegionSection *section = &hdev->mem_sections[i];
1362             memory_region_unref(section->mr);
1363         }
1364         QLIST_REMOVE(hdev, entry);
1365     }
1366     if (hdev->migration_blocker) {
1367         migrate_del_blocker(hdev->migration_blocker);
1368         error_free(hdev->migration_blocker);
1369     }
1370     g_free(hdev->mem);
1371     g_free(hdev->mem_sections);
1372     if (hdev->vhost_ops) {
1373         hdev->vhost_ops->vhost_backend_cleanup(hdev);
1374     }
1375     assert(!hdev->log);
1376 
1377     memset(hdev, 0, sizeof(struct vhost_dev));
1378 }
1379 
1380 /* Stop processing guest IO notifications in qemu.
1381  * Start processing them in vhost in kernel.
1382  */
1383 int vhost_dev_enable_notifiers(struct vhost_dev *hdev, VirtIODevice *vdev)
1384 {
1385     BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev)));
1386     int i, r, e;
1387 
1388     /* We will pass the notifiers to the kernel, make sure that QEMU
1389      * doesn't interfere.
1390      */
1391     r = virtio_device_grab_ioeventfd(vdev);
1392     if (r < 0) {
1393         error_report("binding does not support host notifiers");
1394         goto fail;
1395     }
1396 
1397     for (i = 0; i < hdev->nvqs; ++i) {
1398         r = virtio_bus_set_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i,
1399                                          true);
1400         if (r < 0) {
1401             error_report("vhost VQ %d notifier binding failed: %d", i, -r);
1402             goto fail_vq;
1403         }
1404     }
1405 
1406     return 0;
1407 fail_vq:
1408     while (--i >= 0) {
1409         e = virtio_bus_set_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i,
1410                                          false);
1411         if (e < 0) {
1412             error_report("vhost VQ %d notifier cleanup error: %d", i, -r);
1413         }
1414         assert (e >= 0);
1415     }
1416     virtio_device_release_ioeventfd(vdev);
1417 fail:
1418     return r;
1419 }
1420 
1421 /* Stop processing guest IO notifications in vhost.
1422  * Start processing them in qemu.
1423  * This might actually run the qemu handlers right away,
1424  * so virtio in qemu must be completely setup when this is called.
1425  */
1426 void vhost_dev_disable_notifiers(struct vhost_dev *hdev, VirtIODevice *vdev)
1427 {
1428     BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev)));
1429     int i, r;
1430 
1431     for (i = 0; i < hdev->nvqs; ++i) {
1432         r = virtio_bus_set_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i,
1433                                          false);
1434         if (r < 0) {
1435             error_report("vhost VQ %d notifier cleanup failed: %d", i, -r);
1436         }
1437         assert (r >= 0);
1438     }
1439     virtio_device_release_ioeventfd(vdev);
1440 }
1441 
1442 /* Test and clear event pending status.
1443  * Should be called after unmask to avoid losing events.
1444  */
1445 bool vhost_virtqueue_pending(struct vhost_dev *hdev, int n)
1446 {
1447     struct vhost_virtqueue *vq = hdev->vqs + n - hdev->vq_index;
1448     assert(n >= hdev->vq_index && n < hdev->vq_index + hdev->nvqs);
1449     return event_notifier_test_and_clear(&vq->masked_notifier);
1450 }
1451 
1452 /* Mask/unmask events from this vq. */
1453 void vhost_virtqueue_mask(struct vhost_dev *hdev, VirtIODevice *vdev, int n,
1454                          bool mask)
1455 {
1456     struct VirtQueue *vvq = virtio_get_queue(vdev, n);
1457     int r, index = n - hdev->vq_index;
1458     struct vhost_vring_file file;
1459 
1460     /* should only be called after backend is connected */
1461     assert(hdev->vhost_ops);
1462 
1463     if (mask) {
1464         assert(vdev->use_guest_notifier_mask);
1465         file.fd = event_notifier_get_fd(&hdev->vqs[index].masked_notifier);
1466     } else {
1467         file.fd = event_notifier_get_fd(virtio_queue_get_guest_notifier(vvq));
1468     }
1469 
1470     file.index = hdev->vhost_ops->vhost_get_vq_index(hdev, n);
1471     r = hdev->vhost_ops->vhost_set_vring_call(hdev, &file);
1472     if (r < 0) {
1473         VHOST_OPS_DEBUG("vhost_set_vring_call failed");
1474     }
1475 }
1476 
1477 uint64_t vhost_get_features(struct vhost_dev *hdev, const int *feature_bits,
1478                             uint64_t features)
1479 {
1480     const int *bit = feature_bits;
1481     while (*bit != VHOST_INVALID_FEATURE_BIT) {
1482         uint64_t bit_mask = (1ULL << *bit);
1483         if (!(hdev->features & bit_mask)) {
1484             features &= ~bit_mask;
1485         }
1486         bit++;
1487     }
1488     return features;
1489 }
1490 
1491 void vhost_ack_features(struct vhost_dev *hdev, const int *feature_bits,
1492                         uint64_t features)
1493 {
1494     const int *bit = feature_bits;
1495     while (*bit != VHOST_INVALID_FEATURE_BIT) {
1496         uint64_t bit_mask = (1ULL << *bit);
1497         if (features & bit_mask) {
1498             hdev->acked_features |= bit_mask;
1499         }
1500         bit++;
1501     }
1502 }
1503 
1504 /* Host notifiers must be enabled at this point. */
1505 int vhost_dev_start(struct vhost_dev *hdev, VirtIODevice *vdev)
1506 {
1507     int i, r;
1508 
1509     /* should only be called after backend is connected */
1510     assert(hdev->vhost_ops);
1511 
1512     hdev->started = true;
1513     hdev->vdev = vdev;
1514 
1515     r = vhost_dev_set_features(hdev, hdev->log_enabled);
1516     if (r < 0) {
1517         goto fail_features;
1518     }
1519 
1520     if (vhost_dev_has_iommu(hdev)) {
1521         memory_listener_register(&hdev->iommu_listener, vdev->dma_as);
1522     }
1523 
1524     r = hdev->vhost_ops->vhost_set_mem_table(hdev, hdev->mem);
1525     if (r < 0) {
1526         VHOST_OPS_DEBUG("vhost_set_mem_table failed");
1527         r = -errno;
1528         goto fail_mem;
1529     }
1530     for (i = 0; i < hdev->nvqs; ++i) {
1531         r = vhost_virtqueue_start(hdev,
1532                                   vdev,
1533                                   hdev->vqs + i,
1534                                   hdev->vq_index + i);
1535         if (r < 0) {
1536             goto fail_vq;
1537         }
1538     }
1539 
1540     if (hdev->log_enabled) {
1541         uint64_t log_base;
1542 
1543         hdev->log_size = vhost_get_log_size(hdev);
1544         hdev->log = vhost_log_get(hdev->log_size,
1545                                   vhost_dev_log_is_shared(hdev));
1546         log_base = (uintptr_t)hdev->log->log;
1547         r = hdev->vhost_ops->vhost_set_log_base(hdev,
1548                                                 hdev->log_size ? log_base : 0,
1549                                                 hdev->log);
1550         if (r < 0) {
1551             VHOST_OPS_DEBUG("vhost_set_log_base failed");
1552             r = -errno;
1553             goto fail_log;
1554         }
1555     }
1556 
1557     if (vhost_dev_has_iommu(hdev)) {
1558         hdev->vhost_ops->vhost_set_iotlb_callback(hdev, true);
1559 
1560         /* Update used ring information for IOTLB to work correctly,
1561          * vhost-kernel code requires for this.*/
1562         for (i = 0; i < hdev->nvqs; ++i) {
1563             struct vhost_virtqueue *vq = hdev->vqs + i;
1564             vhost_device_iotlb_miss(hdev, vq->used_phys, true);
1565         }
1566     }
1567     return 0;
1568 fail_log:
1569     vhost_log_put(hdev, false);
1570 fail_vq:
1571     while (--i >= 0) {
1572         vhost_virtqueue_stop(hdev,
1573                              vdev,
1574                              hdev->vqs + i,
1575                              hdev->vq_index + i);
1576     }
1577     i = hdev->nvqs;
1578 
1579 fail_mem:
1580 fail_features:
1581 
1582     hdev->started = false;
1583     return r;
1584 }
1585 
1586 /* Host notifiers must be enabled at this point. */
1587 void vhost_dev_stop(struct vhost_dev *hdev, VirtIODevice *vdev)
1588 {
1589     int i;
1590 
1591     /* should only be called after backend is connected */
1592     assert(hdev->vhost_ops);
1593 
1594     for (i = 0; i < hdev->nvqs; ++i) {
1595         vhost_virtqueue_stop(hdev,
1596                              vdev,
1597                              hdev->vqs + i,
1598                              hdev->vq_index + i);
1599     }
1600 
1601     if (vhost_dev_has_iommu(hdev)) {
1602         hdev->vhost_ops->vhost_set_iotlb_callback(hdev, false);
1603         memory_listener_unregister(&hdev->iommu_listener);
1604     }
1605     vhost_log_put(hdev, true);
1606     hdev->started = false;
1607     hdev->vdev = NULL;
1608 }
1609 
1610 int vhost_net_set_backend(struct vhost_dev *hdev,
1611                           struct vhost_vring_file *file)
1612 {
1613     if (hdev->vhost_ops->vhost_net_set_backend) {
1614         return hdev->vhost_ops->vhost_net_set_backend(hdev, file);
1615     }
1616 
1617     return -1;
1618 }
1619