xref: /openbmc/qemu/hw/virtio/vhost.c (revision 954a6c4f)
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 "qemu/atomic.h"
20 #include "qemu/range.h"
21 #include "qemu/error-report.h"
22 #include "qemu/memfd.h"
23 #include "standard-headers/linux/vhost_types.h"
24 #include "hw/virtio/virtio-bus.h"
25 #include "hw/virtio/virtio-access.h"
26 #include "migration/blocker.h"
27 #include "migration/qemu-file-types.h"
28 #include "sysemu/dma.h"
29 #include "trace.h"
30 
31 /* enabled until disconnected backend stabilizes */
32 #define _VHOST_DEBUG 1
33 
34 #ifdef _VHOST_DEBUG
35 #define VHOST_OPS_DEBUG(retval, fmt, ...) \
36     do { \
37         error_report(fmt ": %s (%d)", ## __VA_ARGS__, \
38                      strerror(-retval), -retval); \
39     } while (0)
40 #else
41 #define VHOST_OPS_DEBUG(retval, fmt, ...) \
42     do { } while (0)
43 #endif
44 
45 static struct vhost_log *vhost_log;
46 static struct vhost_log *vhost_log_shm;
47 
48 static unsigned int used_memslots;
49 static QLIST_HEAD(, vhost_dev) vhost_devices =
50     QLIST_HEAD_INITIALIZER(vhost_devices);
51 
52 bool vhost_has_free_slot(void)
53 {
54     unsigned int slots_limit = ~0U;
55     struct vhost_dev *hdev;
56 
57     QLIST_FOREACH(hdev, &vhost_devices, entry) {
58         unsigned int r = hdev->vhost_ops->vhost_backend_memslots_limit(hdev);
59         slots_limit = MIN(slots_limit, r);
60     }
61     return slots_limit > used_memslots;
62 }
63 
64 static void vhost_dev_sync_region(struct vhost_dev *dev,
65                                   MemoryRegionSection *section,
66                                   uint64_t mfirst, uint64_t mlast,
67                                   uint64_t rfirst, uint64_t rlast)
68 {
69     vhost_log_chunk_t *log = dev->log->log;
70 
71     uint64_t start = MAX(mfirst, rfirst);
72     uint64_t end = MIN(mlast, rlast);
73     vhost_log_chunk_t *from = log + start / VHOST_LOG_CHUNK;
74     vhost_log_chunk_t *to = log + end / VHOST_LOG_CHUNK + 1;
75     uint64_t addr = QEMU_ALIGN_DOWN(start, VHOST_LOG_CHUNK);
76 
77     if (end < start) {
78         return;
79     }
80     assert(end / VHOST_LOG_CHUNK < dev->log_size);
81     assert(start / VHOST_LOG_CHUNK < dev->log_size);
82 
83     for (;from < to; ++from) {
84         vhost_log_chunk_t log;
85         /* We first check with non-atomic: much cheaper,
86          * and we expect non-dirty to be the common case. */
87         if (!*from) {
88             addr += VHOST_LOG_CHUNK;
89             continue;
90         }
91         /* Data must be read atomically. We don't really need barrier semantics
92          * but it's easier to use atomic_* than roll our own. */
93         log = qatomic_xchg(from, 0);
94         while (log) {
95             int bit = ctzl(log);
96             hwaddr page_addr;
97             hwaddr section_offset;
98             hwaddr mr_offset;
99             page_addr = addr + bit * VHOST_LOG_PAGE;
100             section_offset = page_addr - section->offset_within_address_space;
101             mr_offset = section_offset + section->offset_within_region;
102             memory_region_set_dirty(section->mr, mr_offset, VHOST_LOG_PAGE);
103             log &= ~(0x1ull << bit);
104         }
105         addr += VHOST_LOG_CHUNK;
106     }
107 }
108 
109 static int vhost_sync_dirty_bitmap(struct vhost_dev *dev,
110                                    MemoryRegionSection *section,
111                                    hwaddr first,
112                                    hwaddr last)
113 {
114     int i;
115     hwaddr start_addr;
116     hwaddr end_addr;
117 
118     if (!dev->log_enabled || !dev->started) {
119         return 0;
120     }
121     start_addr = section->offset_within_address_space;
122     end_addr = range_get_last(start_addr, int128_get64(section->size));
123     start_addr = MAX(first, start_addr);
124     end_addr = MIN(last, end_addr);
125 
126     for (i = 0; i < dev->mem->nregions; ++i) {
127         struct vhost_memory_region *reg = dev->mem->regions + i;
128         vhost_dev_sync_region(dev, section, start_addr, end_addr,
129                               reg->guest_phys_addr,
130                               range_get_last(reg->guest_phys_addr,
131                                              reg->memory_size));
132     }
133     for (i = 0; i < dev->nvqs; ++i) {
134         struct vhost_virtqueue *vq = dev->vqs + i;
135 
136         if (!vq->used_phys && !vq->used_size) {
137             continue;
138         }
139 
140         vhost_dev_sync_region(dev, section, start_addr, end_addr, vq->used_phys,
141                               range_get_last(vq->used_phys, vq->used_size));
142     }
143     return 0;
144 }
145 
146 static void vhost_log_sync(MemoryListener *listener,
147                           MemoryRegionSection *section)
148 {
149     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
150                                          memory_listener);
151     vhost_sync_dirty_bitmap(dev, section, 0x0, ~0x0ULL);
152 }
153 
154 static void vhost_log_sync_range(struct vhost_dev *dev,
155                                  hwaddr first, hwaddr last)
156 {
157     int i;
158     /* FIXME: this is N^2 in number of sections */
159     for (i = 0; i < dev->n_mem_sections; ++i) {
160         MemoryRegionSection *section = &dev->mem_sections[i];
161         vhost_sync_dirty_bitmap(dev, section, first, last);
162     }
163 }
164 
165 static uint64_t vhost_get_log_size(struct vhost_dev *dev)
166 {
167     uint64_t log_size = 0;
168     int i;
169     for (i = 0; i < dev->mem->nregions; ++i) {
170         struct vhost_memory_region *reg = dev->mem->regions + i;
171         uint64_t last = range_get_last(reg->guest_phys_addr,
172                                        reg->memory_size);
173         log_size = MAX(log_size, last / VHOST_LOG_CHUNK + 1);
174     }
175     return log_size;
176 }
177 
178 static int vhost_set_backend_type(struct vhost_dev *dev,
179                                   VhostBackendType backend_type)
180 {
181     int r = 0;
182 
183     switch (backend_type) {
184 #ifdef CONFIG_VHOST_KERNEL
185     case VHOST_BACKEND_TYPE_KERNEL:
186         dev->vhost_ops = &kernel_ops;
187         break;
188 #endif
189 #ifdef CONFIG_VHOST_USER
190     case VHOST_BACKEND_TYPE_USER:
191         dev->vhost_ops = &user_ops;
192         break;
193 #endif
194 #ifdef CONFIG_VHOST_VDPA
195     case VHOST_BACKEND_TYPE_VDPA:
196         dev->vhost_ops = &vdpa_ops;
197         break;
198 #endif
199     default:
200         error_report("Unknown vhost backend type");
201         r = -1;
202     }
203 
204     return r;
205 }
206 
207 static struct vhost_log *vhost_log_alloc(uint64_t size, bool share)
208 {
209     Error *err = NULL;
210     struct vhost_log *log;
211     uint64_t logsize = size * sizeof(*(log->log));
212     int fd = -1;
213 
214     log = g_new0(struct vhost_log, 1);
215     if (share) {
216         log->log = qemu_memfd_alloc("vhost-log", logsize,
217                                     F_SEAL_GROW | F_SEAL_SHRINK | F_SEAL_SEAL,
218                                     &fd, &err);
219         if (err) {
220             error_report_err(err);
221             g_free(log);
222             return NULL;
223         }
224         memset(log->log, 0, logsize);
225     } else {
226         log->log = g_malloc0(logsize);
227     }
228 
229     log->size = size;
230     log->refcnt = 1;
231     log->fd = fd;
232 
233     return log;
234 }
235 
236 static struct vhost_log *vhost_log_get(uint64_t size, bool share)
237 {
238     struct vhost_log *log = share ? vhost_log_shm : vhost_log;
239 
240     if (!log || log->size != size) {
241         log = vhost_log_alloc(size, share);
242         if (share) {
243             vhost_log_shm = log;
244         } else {
245             vhost_log = log;
246         }
247     } else {
248         ++log->refcnt;
249     }
250 
251     return log;
252 }
253 
254 static void vhost_log_put(struct vhost_dev *dev, bool sync)
255 {
256     struct vhost_log *log = dev->log;
257 
258     if (!log) {
259         return;
260     }
261 
262     --log->refcnt;
263     if (log->refcnt == 0) {
264         /* Sync only the range covered by the old log */
265         if (dev->log_size && sync) {
266             vhost_log_sync_range(dev, 0, dev->log_size * VHOST_LOG_CHUNK - 1);
267         }
268 
269         if (vhost_log == log) {
270             g_free(log->log);
271             vhost_log = NULL;
272         } else if (vhost_log_shm == log) {
273             qemu_memfd_free(log->log, log->size * sizeof(*(log->log)),
274                             log->fd);
275             vhost_log_shm = NULL;
276         }
277 
278         g_free(log);
279     }
280 
281     dev->log = NULL;
282     dev->log_size = 0;
283 }
284 
285 static bool vhost_dev_log_is_shared(struct vhost_dev *dev)
286 {
287     return dev->vhost_ops->vhost_requires_shm_log &&
288            dev->vhost_ops->vhost_requires_shm_log(dev);
289 }
290 
291 static inline void vhost_dev_log_resize(struct vhost_dev *dev, uint64_t size)
292 {
293     struct vhost_log *log = vhost_log_get(size, vhost_dev_log_is_shared(dev));
294     uint64_t log_base = (uintptr_t)log->log;
295     int r;
296 
297     /* inform backend of log switching, this must be done before
298        releasing the current log, to ensure no logging is lost */
299     r = dev->vhost_ops->vhost_set_log_base(dev, log_base, log);
300     if (r < 0) {
301         VHOST_OPS_DEBUG(r, "vhost_set_log_base failed");
302     }
303 
304     vhost_log_put(dev, true);
305     dev->log = log;
306     dev->log_size = size;
307 }
308 
309 static bool vhost_dev_has_iommu(struct vhost_dev *dev)
310 {
311     VirtIODevice *vdev = dev->vdev;
312 
313     /*
314      * For vhost, VIRTIO_F_IOMMU_PLATFORM means the backend support
315      * incremental memory mapping API via IOTLB API. For platform that
316      * does not have IOMMU, there's no need to enable this feature
317      * which may cause unnecessary IOTLB miss/update transactions.
318      */
319     if (vdev) {
320         return virtio_bus_device_iommu_enabled(vdev) &&
321             virtio_host_has_feature(vdev, VIRTIO_F_IOMMU_PLATFORM);
322     } else {
323         return false;
324     }
325 }
326 
327 static void *vhost_memory_map(struct vhost_dev *dev, hwaddr addr,
328                               hwaddr *plen, bool is_write)
329 {
330     if (!vhost_dev_has_iommu(dev)) {
331         return cpu_physical_memory_map(addr, plen, is_write);
332     } else {
333         return (void *)(uintptr_t)addr;
334     }
335 }
336 
337 static void vhost_memory_unmap(struct vhost_dev *dev, void *buffer,
338                                hwaddr len, int is_write,
339                                hwaddr access_len)
340 {
341     if (!vhost_dev_has_iommu(dev)) {
342         cpu_physical_memory_unmap(buffer, len, is_write, access_len);
343     }
344 }
345 
346 static int vhost_verify_ring_part_mapping(void *ring_hva,
347                                           uint64_t ring_gpa,
348                                           uint64_t ring_size,
349                                           void *reg_hva,
350                                           uint64_t reg_gpa,
351                                           uint64_t reg_size)
352 {
353     uint64_t hva_ring_offset;
354     uint64_t ring_last = range_get_last(ring_gpa, ring_size);
355     uint64_t reg_last = range_get_last(reg_gpa, reg_size);
356 
357     if (ring_last < reg_gpa || ring_gpa > reg_last) {
358         return 0;
359     }
360     /* check that whole ring's is mapped */
361     if (ring_last > reg_last) {
362         return -ENOMEM;
363     }
364     /* check that ring's MemoryRegion wasn't replaced */
365     hva_ring_offset = ring_gpa - reg_gpa;
366     if (ring_hva != reg_hva + hva_ring_offset) {
367         return -EBUSY;
368     }
369 
370     return 0;
371 }
372 
373 static int vhost_verify_ring_mappings(struct vhost_dev *dev,
374                                       void *reg_hva,
375                                       uint64_t reg_gpa,
376                                       uint64_t reg_size)
377 {
378     int i, j;
379     int r = 0;
380     const char *part_name[] = {
381         "descriptor table",
382         "available ring",
383         "used ring"
384     };
385 
386     if (vhost_dev_has_iommu(dev)) {
387         return 0;
388     }
389 
390     for (i = 0; i < dev->nvqs; ++i) {
391         struct vhost_virtqueue *vq = dev->vqs + i;
392 
393         if (vq->desc_phys == 0) {
394             continue;
395         }
396 
397         j = 0;
398         r = vhost_verify_ring_part_mapping(
399                 vq->desc, vq->desc_phys, vq->desc_size,
400                 reg_hva, reg_gpa, reg_size);
401         if (r) {
402             break;
403         }
404 
405         j++;
406         r = vhost_verify_ring_part_mapping(
407                 vq->avail, vq->avail_phys, vq->avail_size,
408                 reg_hva, reg_gpa, reg_size);
409         if (r) {
410             break;
411         }
412 
413         j++;
414         r = vhost_verify_ring_part_mapping(
415                 vq->used, vq->used_phys, vq->used_size,
416                 reg_hva, reg_gpa, reg_size);
417         if (r) {
418             break;
419         }
420     }
421 
422     if (r == -ENOMEM) {
423         error_report("Unable to map %s for ring %d", part_name[j], i);
424     } else if (r == -EBUSY) {
425         error_report("%s relocated for ring %d", part_name[j], i);
426     }
427     return r;
428 }
429 
430 /*
431  * vhost_section: identify sections needed for vhost access
432  *
433  * We only care about RAM sections here (where virtqueue and guest
434  * internals accessed by virtio might live). If we find one we still
435  * allow the backend to potentially filter it out of our list.
436  */
437 static bool vhost_section(struct vhost_dev *dev, MemoryRegionSection *section)
438 {
439     MemoryRegion *mr = section->mr;
440 
441     if (memory_region_is_ram(mr) && !memory_region_is_rom(mr)) {
442         uint8_t dirty_mask = memory_region_get_dirty_log_mask(mr);
443         uint8_t handled_dirty;
444 
445         /*
446          * Kernel based vhost doesn't handle any block which is doing
447          * dirty-tracking other than migration for which it has
448          * specific logging support. However for TCG the kernel never
449          * gets involved anyway so we can also ignore it's
450          * self-modiying code detection flags. However a vhost-user
451          * client could still confuse a TCG guest if it re-writes
452          * executable memory that has already been translated.
453          */
454         handled_dirty = (1 << DIRTY_MEMORY_MIGRATION) |
455             (1 << DIRTY_MEMORY_CODE);
456 
457         if (dirty_mask & ~handled_dirty) {
458             trace_vhost_reject_section(mr->name, 1);
459             return false;
460         }
461 
462         if (dev->vhost_ops->vhost_backend_mem_section_filter &&
463             !dev->vhost_ops->vhost_backend_mem_section_filter(dev, section)) {
464             trace_vhost_reject_section(mr->name, 2);
465             return false;
466         }
467 
468         trace_vhost_section(mr->name);
469         return true;
470     } else {
471         trace_vhost_reject_section(mr->name, 3);
472         return false;
473     }
474 }
475 
476 static void vhost_begin(MemoryListener *listener)
477 {
478     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
479                                          memory_listener);
480     dev->tmp_sections = NULL;
481     dev->n_tmp_sections = 0;
482 }
483 
484 static void vhost_commit(MemoryListener *listener)
485 {
486     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
487                                          memory_listener);
488     MemoryRegionSection *old_sections;
489     int n_old_sections;
490     uint64_t log_size;
491     size_t regions_size;
492     int r;
493     int i;
494     bool changed = false;
495 
496     /* Note we can be called before the device is started, but then
497      * starting the device calls set_mem_table, so we need to have
498      * built the data structures.
499      */
500     old_sections = dev->mem_sections;
501     n_old_sections = dev->n_mem_sections;
502     dev->mem_sections = dev->tmp_sections;
503     dev->n_mem_sections = dev->n_tmp_sections;
504 
505     if (dev->n_mem_sections != n_old_sections) {
506         changed = true;
507     } else {
508         /* Same size, lets check the contents */
509         for (int i = 0; i < n_old_sections; i++) {
510             if (!MemoryRegionSection_eq(&old_sections[i],
511                                         &dev->mem_sections[i])) {
512                 changed = true;
513                 break;
514             }
515         }
516     }
517 
518     trace_vhost_commit(dev->started, changed);
519     if (!changed) {
520         goto out;
521     }
522 
523     /* Rebuild the regions list from the new sections list */
524     regions_size = offsetof(struct vhost_memory, regions) +
525                        dev->n_mem_sections * sizeof dev->mem->regions[0];
526     dev->mem = g_realloc(dev->mem, regions_size);
527     dev->mem->nregions = dev->n_mem_sections;
528     used_memslots = dev->mem->nregions;
529     for (i = 0; i < dev->n_mem_sections; i++) {
530         struct vhost_memory_region *cur_vmr = dev->mem->regions + i;
531         struct MemoryRegionSection *mrs = dev->mem_sections + i;
532 
533         cur_vmr->guest_phys_addr = mrs->offset_within_address_space;
534         cur_vmr->memory_size     = int128_get64(mrs->size);
535         cur_vmr->userspace_addr  =
536             (uintptr_t)memory_region_get_ram_ptr(mrs->mr) +
537             mrs->offset_within_region;
538         cur_vmr->flags_padding   = 0;
539     }
540 
541     if (!dev->started) {
542         goto out;
543     }
544 
545     for (i = 0; i < dev->mem->nregions; i++) {
546         if (vhost_verify_ring_mappings(dev,
547                        (void *)(uintptr_t)dev->mem->regions[i].userspace_addr,
548                        dev->mem->regions[i].guest_phys_addr,
549                        dev->mem->regions[i].memory_size)) {
550             error_report("Verify ring failure on region %d", i);
551             abort();
552         }
553     }
554 
555     if (!dev->log_enabled) {
556         r = dev->vhost_ops->vhost_set_mem_table(dev, dev->mem);
557         if (r < 0) {
558             VHOST_OPS_DEBUG(r, "vhost_set_mem_table failed");
559         }
560         goto out;
561     }
562     log_size = vhost_get_log_size(dev);
563     /* We allocate an extra 4K bytes to log,
564      * to reduce the * number of reallocations. */
565 #define VHOST_LOG_BUFFER (0x1000 / sizeof *dev->log)
566     /* To log more, must increase log size before table update. */
567     if (dev->log_size < log_size) {
568         vhost_dev_log_resize(dev, log_size + VHOST_LOG_BUFFER);
569     }
570     r = dev->vhost_ops->vhost_set_mem_table(dev, dev->mem);
571     if (r < 0) {
572         VHOST_OPS_DEBUG(r, "vhost_set_mem_table failed");
573     }
574     /* To log less, can only decrease log size after table update. */
575     if (dev->log_size > log_size + VHOST_LOG_BUFFER) {
576         vhost_dev_log_resize(dev, log_size);
577     }
578 
579 out:
580     /* Deref the old list of sections, this must happen _after_ the
581      * vhost_set_mem_table to ensure the client isn't still using the
582      * section we're about to unref.
583      */
584     while (n_old_sections--) {
585         memory_region_unref(old_sections[n_old_sections].mr);
586     }
587     g_free(old_sections);
588     return;
589 }
590 
591 /* Adds the section data to the tmp_section structure.
592  * It relies on the listener calling us in memory address order
593  * and for each region (via the _add and _nop methods) to
594  * join neighbours.
595  */
596 static void vhost_region_add_section(struct vhost_dev *dev,
597                                      MemoryRegionSection *section)
598 {
599     bool need_add = true;
600     uint64_t mrs_size = int128_get64(section->size);
601     uint64_t mrs_gpa = section->offset_within_address_space;
602     uintptr_t mrs_host = (uintptr_t)memory_region_get_ram_ptr(section->mr) +
603                          section->offset_within_region;
604     RAMBlock *mrs_rb = section->mr->ram_block;
605 
606     trace_vhost_region_add_section(section->mr->name, mrs_gpa, mrs_size,
607                                    mrs_host);
608 
609     if (dev->vhost_ops->backend_type == VHOST_BACKEND_TYPE_USER) {
610         /* Round the section to it's page size */
611         /* First align the start down to a page boundary */
612         size_t mrs_page = qemu_ram_pagesize(mrs_rb);
613         uint64_t alignage = mrs_host & (mrs_page - 1);
614         if (alignage) {
615             mrs_host -= alignage;
616             mrs_size += alignage;
617             mrs_gpa  -= alignage;
618         }
619         /* Now align the size up to a page boundary */
620         alignage = mrs_size & (mrs_page - 1);
621         if (alignage) {
622             mrs_size += mrs_page - alignage;
623         }
624         trace_vhost_region_add_section_aligned(section->mr->name, mrs_gpa,
625                                                mrs_size, mrs_host);
626     }
627 
628     if (dev->n_tmp_sections) {
629         /* Since we already have at least one section, lets see if
630          * this extends it; since we're scanning in order, we only
631          * have to look at the last one, and the FlatView that calls
632          * us shouldn't have overlaps.
633          */
634         MemoryRegionSection *prev_sec = dev->tmp_sections +
635                                                (dev->n_tmp_sections - 1);
636         uint64_t prev_gpa_start = prev_sec->offset_within_address_space;
637         uint64_t prev_size = int128_get64(prev_sec->size);
638         uint64_t prev_gpa_end   = range_get_last(prev_gpa_start, prev_size);
639         uint64_t prev_host_start =
640                         (uintptr_t)memory_region_get_ram_ptr(prev_sec->mr) +
641                         prev_sec->offset_within_region;
642         uint64_t prev_host_end   = range_get_last(prev_host_start, prev_size);
643 
644         if (mrs_gpa <= (prev_gpa_end + 1)) {
645             /* OK, looks like overlapping/intersecting - it's possible that
646              * the rounding to page sizes has made them overlap, but they should
647              * match up in the same RAMBlock if they do.
648              */
649             if (mrs_gpa < prev_gpa_start) {
650                 error_report("%s:Section '%s' rounded to %"PRIx64
651                              " prior to previous '%s' %"PRIx64,
652                              __func__, section->mr->name, mrs_gpa,
653                              prev_sec->mr->name, prev_gpa_start);
654                 /* A way to cleanly fail here would be better */
655                 return;
656             }
657             /* Offset from the start of the previous GPA to this GPA */
658             size_t offset = mrs_gpa - prev_gpa_start;
659 
660             if (prev_host_start + offset == mrs_host &&
661                 section->mr == prev_sec->mr &&
662                 (!dev->vhost_ops->vhost_backend_can_merge ||
663                  dev->vhost_ops->vhost_backend_can_merge(dev,
664                     mrs_host, mrs_size,
665                     prev_host_start, prev_size))) {
666                 uint64_t max_end = MAX(prev_host_end, mrs_host + mrs_size);
667                 need_add = false;
668                 prev_sec->offset_within_address_space =
669                     MIN(prev_gpa_start, mrs_gpa);
670                 prev_sec->offset_within_region =
671                     MIN(prev_host_start, mrs_host) -
672                     (uintptr_t)memory_region_get_ram_ptr(prev_sec->mr);
673                 prev_sec->size = int128_make64(max_end - MIN(prev_host_start,
674                                                mrs_host));
675                 trace_vhost_region_add_section_merge(section->mr->name,
676                                         int128_get64(prev_sec->size),
677                                         prev_sec->offset_within_address_space,
678                                         prev_sec->offset_within_region);
679             } else {
680                 /* adjoining regions are fine, but overlapping ones with
681                  * different blocks/offsets shouldn't happen
682                  */
683                 if (mrs_gpa != prev_gpa_end + 1) {
684                     error_report("%s: Overlapping but not coherent sections "
685                                  "at %"PRIx64,
686                                  __func__, mrs_gpa);
687                     return;
688                 }
689             }
690         }
691     }
692 
693     if (need_add) {
694         ++dev->n_tmp_sections;
695         dev->tmp_sections = g_renew(MemoryRegionSection, dev->tmp_sections,
696                                     dev->n_tmp_sections);
697         dev->tmp_sections[dev->n_tmp_sections - 1] = *section;
698         /* The flatview isn't stable and we don't use it, making it NULL
699          * means we can memcmp the list.
700          */
701         dev->tmp_sections[dev->n_tmp_sections - 1].fv = NULL;
702         memory_region_ref(section->mr);
703     }
704 }
705 
706 /* Used for both add and nop callbacks */
707 static void vhost_region_addnop(MemoryListener *listener,
708                                 MemoryRegionSection *section)
709 {
710     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
711                                          memory_listener);
712 
713     if (!vhost_section(dev, section)) {
714         return;
715     }
716     vhost_region_add_section(dev, section);
717 }
718 
719 static void vhost_iommu_unmap_notify(IOMMUNotifier *n, IOMMUTLBEntry *iotlb)
720 {
721     struct vhost_iommu *iommu = container_of(n, struct vhost_iommu, n);
722     struct vhost_dev *hdev = iommu->hdev;
723     hwaddr iova = iotlb->iova + iommu->iommu_offset;
724 
725     if (vhost_backend_invalidate_device_iotlb(hdev, iova,
726                                               iotlb->addr_mask + 1)) {
727         error_report("Fail to invalidate device iotlb");
728     }
729 }
730 
731 static void vhost_iommu_region_add(MemoryListener *listener,
732                                    MemoryRegionSection *section)
733 {
734     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
735                                          iommu_listener);
736     struct vhost_iommu *iommu;
737     Int128 end;
738     int iommu_idx;
739     IOMMUMemoryRegion *iommu_mr;
740     int ret;
741 
742     if (!memory_region_is_iommu(section->mr)) {
743         return;
744     }
745 
746     iommu_mr = IOMMU_MEMORY_REGION(section->mr);
747 
748     iommu = g_malloc0(sizeof(*iommu));
749     end = int128_add(int128_make64(section->offset_within_region),
750                      section->size);
751     end = int128_sub(end, int128_one());
752     iommu_idx = memory_region_iommu_attrs_to_index(iommu_mr,
753                                                    MEMTXATTRS_UNSPECIFIED);
754     iommu_notifier_init(&iommu->n, vhost_iommu_unmap_notify,
755                         IOMMU_NOTIFIER_DEVIOTLB_UNMAP,
756                         section->offset_within_region,
757                         int128_get64(end),
758                         iommu_idx);
759     iommu->mr = section->mr;
760     iommu->iommu_offset = section->offset_within_address_space -
761                           section->offset_within_region;
762     iommu->hdev = dev;
763     ret = memory_region_register_iommu_notifier(section->mr, &iommu->n, NULL);
764     if (ret) {
765         /*
766          * Some vIOMMUs do not support dev-iotlb yet.  If so, try to use the
767          * UNMAP legacy message
768          */
769         iommu->n.notifier_flags = IOMMU_NOTIFIER_UNMAP;
770         memory_region_register_iommu_notifier(section->mr, &iommu->n,
771                                               &error_fatal);
772     }
773     QLIST_INSERT_HEAD(&dev->iommu_list, iommu, iommu_next);
774     /* TODO: can replay help performance here? */
775 }
776 
777 static void vhost_iommu_region_del(MemoryListener *listener,
778                                    MemoryRegionSection *section)
779 {
780     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
781                                          iommu_listener);
782     struct vhost_iommu *iommu;
783 
784     if (!memory_region_is_iommu(section->mr)) {
785         return;
786     }
787 
788     QLIST_FOREACH(iommu, &dev->iommu_list, iommu_next) {
789         if (iommu->mr == section->mr &&
790             iommu->n.start == section->offset_within_region) {
791             memory_region_unregister_iommu_notifier(iommu->mr,
792                                                     &iommu->n);
793             QLIST_REMOVE(iommu, iommu_next);
794             g_free(iommu);
795             break;
796         }
797     }
798 }
799 
800 static int vhost_virtqueue_set_addr(struct vhost_dev *dev,
801                                     struct vhost_virtqueue *vq,
802                                     unsigned idx, bool enable_log)
803 {
804     struct vhost_vring_addr addr;
805     int r;
806     memset(&addr, 0, sizeof(struct vhost_vring_addr));
807 
808     if (dev->vhost_ops->vhost_vq_get_addr) {
809         r = dev->vhost_ops->vhost_vq_get_addr(dev, &addr, vq);
810         if (r < 0) {
811             VHOST_OPS_DEBUG(r, "vhost_vq_get_addr failed");
812             return r;
813         }
814     } else {
815         addr.desc_user_addr = (uint64_t)(unsigned long)vq->desc;
816         addr.avail_user_addr = (uint64_t)(unsigned long)vq->avail;
817         addr.used_user_addr = (uint64_t)(unsigned long)vq->used;
818     }
819     addr.index = idx;
820     addr.log_guest_addr = vq->used_phys;
821     addr.flags = enable_log ? (1 << VHOST_VRING_F_LOG) : 0;
822     r = dev->vhost_ops->vhost_set_vring_addr(dev, &addr);
823     if (r < 0) {
824         VHOST_OPS_DEBUG(r, "vhost_set_vring_addr failed");
825     }
826     return r;
827 }
828 
829 static int vhost_dev_set_features(struct vhost_dev *dev,
830                                   bool enable_log)
831 {
832     uint64_t features = dev->acked_features;
833     int r;
834     if (enable_log) {
835         features |= 0x1ULL << VHOST_F_LOG_ALL;
836     }
837     if (!vhost_dev_has_iommu(dev)) {
838         features &= ~(0x1ULL << VIRTIO_F_IOMMU_PLATFORM);
839     }
840     if (dev->vhost_ops->vhost_force_iommu) {
841         if (dev->vhost_ops->vhost_force_iommu(dev) == true) {
842             features |= 0x1ULL << VIRTIO_F_IOMMU_PLATFORM;
843        }
844     }
845     r = dev->vhost_ops->vhost_set_features(dev, features);
846     if (r < 0) {
847         VHOST_OPS_DEBUG(r, "vhost_set_features failed");
848         goto out;
849     }
850     if (dev->vhost_ops->vhost_set_backend_cap) {
851         r = dev->vhost_ops->vhost_set_backend_cap(dev);
852         if (r < 0) {
853             VHOST_OPS_DEBUG(r, "vhost_set_backend_cap failed");
854             goto out;
855         }
856     }
857 
858 out:
859     return r;
860 }
861 
862 static int vhost_dev_set_log(struct vhost_dev *dev, bool enable_log)
863 {
864     int r, i, idx;
865     hwaddr addr;
866 
867     r = vhost_dev_set_features(dev, enable_log);
868     if (r < 0) {
869         goto err_features;
870     }
871     for (i = 0; i < dev->nvqs; ++i) {
872         idx = dev->vhost_ops->vhost_get_vq_index(dev, dev->vq_index + i);
873         addr = virtio_queue_get_desc_addr(dev->vdev, idx);
874         if (!addr) {
875             /*
876              * The queue might not be ready for start. If this
877              * is the case there is no reason to continue the process.
878              * The similar logic is used by the vhost_virtqueue_start()
879              * routine.
880              */
881             continue;
882         }
883         r = vhost_virtqueue_set_addr(dev, dev->vqs + i, idx,
884                                      enable_log);
885         if (r < 0) {
886             goto err_vq;
887         }
888     }
889     return 0;
890 err_vq:
891     for (; i >= 0; --i) {
892         idx = dev->vhost_ops->vhost_get_vq_index(dev, dev->vq_index + i);
893         addr = virtio_queue_get_desc_addr(dev->vdev, idx);
894         if (!addr) {
895             continue;
896         }
897         vhost_virtqueue_set_addr(dev, dev->vqs + i, idx,
898                                  dev->log_enabled);
899     }
900     vhost_dev_set_features(dev, dev->log_enabled);
901 err_features:
902     return r;
903 }
904 
905 static int vhost_migration_log(MemoryListener *listener, bool enable)
906 {
907     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
908                                          memory_listener);
909     int r;
910     if (enable == dev->log_enabled) {
911         return 0;
912     }
913     if (!dev->started) {
914         dev->log_enabled = enable;
915         return 0;
916     }
917 
918     r = 0;
919     if (!enable) {
920         r = vhost_dev_set_log(dev, false);
921         if (r < 0) {
922             goto check_dev_state;
923         }
924         vhost_log_put(dev, false);
925     } else {
926         vhost_dev_log_resize(dev, vhost_get_log_size(dev));
927         r = vhost_dev_set_log(dev, true);
928         if (r < 0) {
929             goto check_dev_state;
930         }
931     }
932 
933 check_dev_state:
934     dev->log_enabled = enable;
935     /*
936      * vhost-user-* devices could change their state during log
937      * initialization due to disconnect. So check dev state after
938      * vhost communication.
939      */
940     if (!dev->started) {
941         /*
942          * Since device is in the stopped state, it is okay for
943          * migration. Return success.
944          */
945         r = 0;
946     }
947     if (r) {
948         /* An error occurred. */
949         dev->log_enabled = false;
950     }
951 
952     return r;
953 }
954 
955 static void vhost_log_global_start(MemoryListener *listener)
956 {
957     int r;
958 
959     r = vhost_migration_log(listener, true);
960     if (r < 0) {
961         abort();
962     }
963 }
964 
965 static void vhost_log_global_stop(MemoryListener *listener)
966 {
967     int r;
968 
969     r = vhost_migration_log(listener, false);
970     if (r < 0) {
971         abort();
972     }
973 }
974 
975 static void vhost_log_start(MemoryListener *listener,
976                             MemoryRegionSection *section,
977                             int old, int new)
978 {
979     /* FIXME: implement */
980 }
981 
982 static void vhost_log_stop(MemoryListener *listener,
983                            MemoryRegionSection *section,
984                            int old, int new)
985 {
986     /* FIXME: implement */
987 }
988 
989 /* The vhost driver natively knows how to handle the vrings of non
990  * cross-endian legacy devices and modern devices. Only legacy devices
991  * exposed to a bi-endian guest may require the vhost driver to use a
992  * specific endianness.
993  */
994 static inline bool vhost_needs_vring_endian(VirtIODevice *vdev)
995 {
996     if (virtio_vdev_has_feature(vdev, VIRTIO_F_VERSION_1)) {
997         return false;
998     }
999 #if HOST_BIG_ENDIAN
1000     return vdev->device_endian == VIRTIO_DEVICE_ENDIAN_LITTLE;
1001 #else
1002     return vdev->device_endian == VIRTIO_DEVICE_ENDIAN_BIG;
1003 #endif
1004 }
1005 
1006 static int vhost_virtqueue_set_vring_endian_legacy(struct vhost_dev *dev,
1007                                                    bool is_big_endian,
1008                                                    int vhost_vq_index)
1009 {
1010     int r;
1011     struct vhost_vring_state s = {
1012         .index = vhost_vq_index,
1013         .num = is_big_endian
1014     };
1015 
1016     r = dev->vhost_ops->vhost_set_vring_endian(dev, &s);
1017     if (r < 0) {
1018         VHOST_OPS_DEBUG(r, "vhost_set_vring_endian failed");
1019     }
1020     return r;
1021 }
1022 
1023 static int vhost_memory_region_lookup(struct vhost_dev *hdev,
1024                                       uint64_t gpa, uint64_t *uaddr,
1025                                       uint64_t *len)
1026 {
1027     int i;
1028 
1029     for (i = 0; i < hdev->mem->nregions; i++) {
1030         struct vhost_memory_region *reg = hdev->mem->regions + i;
1031 
1032         if (gpa >= reg->guest_phys_addr &&
1033             reg->guest_phys_addr + reg->memory_size > gpa) {
1034             *uaddr = reg->userspace_addr + gpa - reg->guest_phys_addr;
1035             *len = reg->guest_phys_addr + reg->memory_size - gpa;
1036             return 0;
1037         }
1038     }
1039 
1040     return -EFAULT;
1041 }
1042 
1043 int vhost_device_iotlb_miss(struct vhost_dev *dev, uint64_t iova, int write)
1044 {
1045     IOMMUTLBEntry iotlb;
1046     uint64_t uaddr, len;
1047     int ret = -EFAULT;
1048 
1049     RCU_READ_LOCK_GUARD();
1050 
1051     trace_vhost_iotlb_miss(dev, 1);
1052 
1053     iotlb = address_space_get_iotlb_entry(dev->vdev->dma_as,
1054                                           iova, write,
1055                                           MEMTXATTRS_UNSPECIFIED);
1056     if (iotlb.target_as != NULL) {
1057         ret = vhost_memory_region_lookup(dev, iotlb.translated_addr,
1058                                          &uaddr, &len);
1059         if (ret) {
1060             trace_vhost_iotlb_miss(dev, 3);
1061             error_report("Fail to lookup the translated address "
1062                          "%"PRIx64, iotlb.translated_addr);
1063             goto out;
1064         }
1065 
1066         len = MIN(iotlb.addr_mask + 1, len);
1067         iova = iova & ~iotlb.addr_mask;
1068 
1069         ret = vhost_backend_update_device_iotlb(dev, iova, uaddr,
1070                                                 len, iotlb.perm);
1071         if (ret) {
1072             trace_vhost_iotlb_miss(dev, 4);
1073             error_report("Fail to update device iotlb");
1074             goto out;
1075         }
1076     }
1077 
1078     trace_vhost_iotlb_miss(dev, 2);
1079 
1080 out:
1081     return ret;
1082 }
1083 
1084 int vhost_virtqueue_start(struct vhost_dev *dev,
1085                           struct VirtIODevice *vdev,
1086                           struct vhost_virtqueue *vq,
1087                           unsigned idx)
1088 {
1089     BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev)));
1090     VirtioBusState *vbus = VIRTIO_BUS(qbus);
1091     VirtioBusClass *k = VIRTIO_BUS_GET_CLASS(vbus);
1092     hwaddr s, l, a;
1093     int r;
1094     int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, idx);
1095     struct vhost_vring_file file = {
1096         .index = vhost_vq_index
1097     };
1098     struct vhost_vring_state state = {
1099         .index = vhost_vq_index
1100     };
1101     struct VirtQueue *vvq = virtio_get_queue(vdev, idx);
1102 
1103     a = virtio_queue_get_desc_addr(vdev, idx);
1104     if (a == 0) {
1105         /* Queue might not be ready for start */
1106         return 0;
1107     }
1108 
1109     vq->num = state.num = virtio_queue_get_num(vdev, idx);
1110     r = dev->vhost_ops->vhost_set_vring_num(dev, &state);
1111     if (r) {
1112         VHOST_OPS_DEBUG(r, "vhost_set_vring_num failed");
1113         return r;
1114     }
1115 
1116     state.num = virtio_queue_get_last_avail_idx(vdev, idx);
1117     r = dev->vhost_ops->vhost_set_vring_base(dev, &state);
1118     if (r) {
1119         VHOST_OPS_DEBUG(r, "vhost_set_vring_base failed");
1120         return r;
1121     }
1122 
1123     if (vhost_needs_vring_endian(vdev)) {
1124         r = vhost_virtqueue_set_vring_endian_legacy(dev,
1125                                                     virtio_is_big_endian(vdev),
1126                                                     vhost_vq_index);
1127         if (r) {
1128             return r;
1129         }
1130     }
1131 
1132     vq->desc_size = s = l = virtio_queue_get_desc_size(vdev, idx);
1133     vq->desc_phys = a;
1134     vq->desc = vhost_memory_map(dev, a, &l, false);
1135     if (!vq->desc || l != s) {
1136         r = -ENOMEM;
1137         goto fail_alloc_desc;
1138     }
1139     vq->avail_size = s = l = virtio_queue_get_avail_size(vdev, idx);
1140     vq->avail_phys = a = virtio_queue_get_avail_addr(vdev, idx);
1141     vq->avail = vhost_memory_map(dev, a, &l, false);
1142     if (!vq->avail || l != s) {
1143         r = -ENOMEM;
1144         goto fail_alloc_avail;
1145     }
1146     vq->used_size = s = l = virtio_queue_get_used_size(vdev, idx);
1147     vq->used_phys = a = virtio_queue_get_used_addr(vdev, idx);
1148     vq->used = vhost_memory_map(dev, a, &l, true);
1149     if (!vq->used || l != s) {
1150         r = -ENOMEM;
1151         goto fail_alloc_used;
1152     }
1153 
1154     r = vhost_virtqueue_set_addr(dev, vq, vhost_vq_index, dev->log_enabled);
1155     if (r < 0) {
1156         goto fail_alloc;
1157     }
1158 
1159     file.fd = event_notifier_get_fd(virtio_queue_get_host_notifier(vvq));
1160     r = dev->vhost_ops->vhost_set_vring_kick(dev, &file);
1161     if (r) {
1162         VHOST_OPS_DEBUG(r, "vhost_set_vring_kick failed");
1163         goto fail_kick;
1164     }
1165 
1166     /* Clear and discard previous events if any. */
1167     event_notifier_test_and_clear(&vq->masked_notifier);
1168 
1169     /* Init vring in unmasked state, unless guest_notifier_mask
1170      * will do it later.
1171      */
1172     if (!vdev->use_guest_notifier_mask) {
1173         /* TODO: check and handle errors. */
1174         vhost_virtqueue_mask(dev, vdev, idx, false);
1175     }
1176 
1177     if (k->query_guest_notifiers &&
1178         k->query_guest_notifiers(qbus->parent) &&
1179         virtio_queue_vector(vdev, idx) == VIRTIO_NO_VECTOR) {
1180         file.fd = -1;
1181         r = dev->vhost_ops->vhost_set_vring_call(dev, &file);
1182         if (r) {
1183             goto fail_vector;
1184         }
1185     }
1186 
1187     return 0;
1188 
1189 fail_vector:
1190 fail_kick:
1191 fail_alloc:
1192     vhost_memory_unmap(dev, vq->used, virtio_queue_get_used_size(vdev, idx),
1193                        0, 0);
1194 fail_alloc_used:
1195     vhost_memory_unmap(dev, vq->avail, virtio_queue_get_avail_size(vdev, idx),
1196                        0, 0);
1197 fail_alloc_avail:
1198     vhost_memory_unmap(dev, vq->desc, virtio_queue_get_desc_size(vdev, idx),
1199                        0, 0);
1200 fail_alloc_desc:
1201     return r;
1202 }
1203 
1204 void vhost_virtqueue_stop(struct vhost_dev *dev,
1205                           struct VirtIODevice *vdev,
1206                           struct vhost_virtqueue *vq,
1207                           unsigned idx)
1208 {
1209     int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, idx);
1210     struct vhost_vring_state state = {
1211         .index = vhost_vq_index,
1212     };
1213     int r;
1214 
1215     if (virtio_queue_get_desc_addr(vdev, idx) == 0) {
1216         /* Don't stop the virtqueue which might have not been started */
1217         return;
1218     }
1219 
1220     r = dev->vhost_ops->vhost_get_vring_base(dev, &state);
1221     if (r < 0) {
1222         VHOST_OPS_DEBUG(r, "vhost VQ %u ring restore failed: %d", idx, r);
1223         /* Connection to the backend is broken, so let's sync internal
1224          * last avail idx to the device used idx.
1225          */
1226         virtio_queue_restore_last_avail_idx(vdev, idx);
1227     } else {
1228         virtio_queue_set_last_avail_idx(vdev, idx, state.num);
1229     }
1230     virtio_queue_invalidate_signalled_used(vdev, idx);
1231     virtio_queue_update_used_idx(vdev, idx);
1232 
1233     /* In the cross-endian case, we need to reset the vring endianness to
1234      * native as legacy devices expect so by default.
1235      */
1236     if (vhost_needs_vring_endian(vdev)) {
1237         vhost_virtqueue_set_vring_endian_legacy(dev,
1238                                                 !virtio_is_big_endian(vdev),
1239                                                 vhost_vq_index);
1240     }
1241 
1242     vhost_memory_unmap(dev, vq->used, virtio_queue_get_used_size(vdev, idx),
1243                        1, virtio_queue_get_used_size(vdev, idx));
1244     vhost_memory_unmap(dev, vq->avail, virtio_queue_get_avail_size(vdev, idx),
1245                        0, virtio_queue_get_avail_size(vdev, idx));
1246     vhost_memory_unmap(dev, vq->desc, virtio_queue_get_desc_size(vdev, idx),
1247                        0, virtio_queue_get_desc_size(vdev, idx));
1248 }
1249 
1250 static void vhost_eventfd_add(MemoryListener *listener,
1251                               MemoryRegionSection *section,
1252                               bool match_data, uint64_t data, EventNotifier *e)
1253 {
1254 }
1255 
1256 static void vhost_eventfd_del(MemoryListener *listener,
1257                               MemoryRegionSection *section,
1258                               bool match_data, uint64_t data, EventNotifier *e)
1259 {
1260 }
1261 
1262 static int vhost_virtqueue_set_busyloop_timeout(struct vhost_dev *dev,
1263                                                 int n, uint32_t timeout)
1264 {
1265     int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, n);
1266     struct vhost_vring_state state = {
1267         .index = vhost_vq_index,
1268         .num = timeout,
1269     };
1270     int r;
1271 
1272     if (!dev->vhost_ops->vhost_set_vring_busyloop_timeout) {
1273         return -EINVAL;
1274     }
1275 
1276     r = dev->vhost_ops->vhost_set_vring_busyloop_timeout(dev, &state);
1277     if (r) {
1278         VHOST_OPS_DEBUG(r, "vhost_set_vring_busyloop_timeout failed");
1279         return r;
1280     }
1281 
1282     return 0;
1283 }
1284 
1285 static void vhost_virtqueue_error_notifier(EventNotifier *n)
1286 {
1287     struct vhost_virtqueue *vq = container_of(n, struct vhost_virtqueue,
1288                                               error_notifier);
1289     struct vhost_dev *dev = vq->dev;
1290     int index = vq - dev->vqs;
1291 
1292     if (event_notifier_test_and_clear(n) && dev->vdev) {
1293         VHOST_OPS_DEBUG(-EINVAL,  "vhost vring error in virtqueue %d",
1294                         dev->vq_index + index);
1295     }
1296 }
1297 
1298 static int vhost_virtqueue_init(struct vhost_dev *dev,
1299                                 struct vhost_virtqueue *vq, int n)
1300 {
1301     int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, n);
1302     struct vhost_vring_file file = {
1303         .index = vhost_vq_index,
1304     };
1305     int r = event_notifier_init(&vq->masked_notifier, 0);
1306     if (r < 0) {
1307         return r;
1308     }
1309 
1310     file.fd = event_notifier_get_wfd(&vq->masked_notifier);
1311     r = dev->vhost_ops->vhost_set_vring_call(dev, &file);
1312     if (r) {
1313         VHOST_OPS_DEBUG(r, "vhost_set_vring_call failed");
1314         goto fail_call;
1315     }
1316 
1317     vq->dev = dev;
1318 
1319     if (dev->vhost_ops->vhost_set_vring_err) {
1320         r = event_notifier_init(&vq->error_notifier, 0);
1321         if (r < 0) {
1322             goto fail_call;
1323         }
1324 
1325         file.fd = event_notifier_get_fd(&vq->error_notifier);
1326         r = dev->vhost_ops->vhost_set_vring_err(dev, &file);
1327         if (r) {
1328             VHOST_OPS_DEBUG(r, "vhost_set_vring_err failed");
1329             goto fail_err;
1330         }
1331 
1332         event_notifier_set_handler(&vq->error_notifier,
1333                                    vhost_virtqueue_error_notifier);
1334     }
1335 
1336     return 0;
1337 
1338 fail_err:
1339     event_notifier_cleanup(&vq->error_notifier);
1340 fail_call:
1341     event_notifier_cleanup(&vq->masked_notifier);
1342     return r;
1343 }
1344 
1345 static void vhost_virtqueue_cleanup(struct vhost_virtqueue *vq)
1346 {
1347     event_notifier_cleanup(&vq->masked_notifier);
1348     if (vq->dev->vhost_ops->vhost_set_vring_err) {
1349         event_notifier_set_handler(&vq->error_notifier, NULL);
1350         event_notifier_cleanup(&vq->error_notifier);
1351     }
1352 }
1353 
1354 int vhost_dev_init(struct vhost_dev *hdev, void *opaque,
1355                    VhostBackendType backend_type, uint32_t busyloop_timeout,
1356                    Error **errp)
1357 {
1358     uint64_t features;
1359     int i, r, n_initialized_vqs = 0;
1360 
1361     hdev->vdev = NULL;
1362     hdev->migration_blocker = NULL;
1363 
1364     r = vhost_set_backend_type(hdev, backend_type);
1365     assert(r >= 0);
1366 
1367     r = hdev->vhost_ops->vhost_backend_init(hdev, opaque, errp);
1368     if (r < 0) {
1369         goto fail;
1370     }
1371 
1372     r = hdev->vhost_ops->vhost_set_owner(hdev);
1373     if (r < 0) {
1374         error_setg_errno(errp, -r, "vhost_set_owner failed");
1375         goto fail;
1376     }
1377 
1378     r = hdev->vhost_ops->vhost_get_features(hdev, &features);
1379     if (r < 0) {
1380         error_setg_errno(errp, -r, "vhost_get_features failed");
1381         goto fail;
1382     }
1383 
1384     for (i = 0; i < hdev->nvqs; ++i, ++n_initialized_vqs) {
1385         r = vhost_virtqueue_init(hdev, hdev->vqs + i, hdev->vq_index + i);
1386         if (r < 0) {
1387             error_setg_errno(errp, -r, "Failed to initialize virtqueue %d", i);
1388             goto fail;
1389         }
1390     }
1391 
1392     if (busyloop_timeout) {
1393         for (i = 0; i < hdev->nvqs; ++i) {
1394             r = vhost_virtqueue_set_busyloop_timeout(hdev, hdev->vq_index + i,
1395                                                      busyloop_timeout);
1396             if (r < 0) {
1397                 error_setg_errno(errp, -r, "Failed to set busyloop timeout");
1398                 goto fail_busyloop;
1399             }
1400         }
1401     }
1402 
1403     hdev->features = features;
1404 
1405     hdev->memory_listener = (MemoryListener) {
1406         .name = "vhost",
1407         .begin = vhost_begin,
1408         .commit = vhost_commit,
1409         .region_add = vhost_region_addnop,
1410         .region_nop = vhost_region_addnop,
1411         .log_start = vhost_log_start,
1412         .log_stop = vhost_log_stop,
1413         .log_sync = vhost_log_sync,
1414         .log_global_start = vhost_log_global_start,
1415         .log_global_stop = vhost_log_global_stop,
1416         .eventfd_add = vhost_eventfd_add,
1417         .eventfd_del = vhost_eventfd_del,
1418         .priority = 10
1419     };
1420 
1421     hdev->iommu_listener = (MemoryListener) {
1422         .name = "vhost-iommu",
1423         .region_add = vhost_iommu_region_add,
1424         .region_del = vhost_iommu_region_del,
1425     };
1426 
1427     if (hdev->migration_blocker == NULL) {
1428         if (!(hdev->features & (0x1ULL << VHOST_F_LOG_ALL))) {
1429             error_setg(&hdev->migration_blocker,
1430                        "Migration disabled: vhost lacks VHOST_F_LOG_ALL feature.");
1431         } else if (vhost_dev_log_is_shared(hdev) && !qemu_memfd_alloc_check()) {
1432             error_setg(&hdev->migration_blocker,
1433                        "Migration disabled: failed to allocate shared memory");
1434         }
1435     }
1436 
1437     if (hdev->migration_blocker != NULL) {
1438         r = migrate_add_blocker(hdev->migration_blocker, errp);
1439         if (r < 0) {
1440             error_free(hdev->migration_blocker);
1441             goto fail_busyloop;
1442         }
1443     }
1444 
1445     hdev->mem = g_malloc0(offsetof(struct vhost_memory, regions));
1446     hdev->n_mem_sections = 0;
1447     hdev->mem_sections = NULL;
1448     hdev->log = NULL;
1449     hdev->log_size = 0;
1450     hdev->log_enabled = false;
1451     hdev->started = false;
1452     memory_listener_register(&hdev->memory_listener, &address_space_memory);
1453     QLIST_INSERT_HEAD(&vhost_devices, hdev, entry);
1454 
1455     if (used_memslots > hdev->vhost_ops->vhost_backend_memslots_limit(hdev)) {
1456         error_setg(errp, "vhost backend memory slots limit is less"
1457                    " than current number of present memory slots");
1458         r = -EINVAL;
1459         goto fail_busyloop;
1460     }
1461 
1462     return 0;
1463 
1464 fail_busyloop:
1465     if (busyloop_timeout) {
1466         while (--i >= 0) {
1467             vhost_virtqueue_set_busyloop_timeout(hdev, hdev->vq_index + i, 0);
1468         }
1469     }
1470 fail:
1471     hdev->nvqs = n_initialized_vqs;
1472     vhost_dev_cleanup(hdev);
1473     return r;
1474 }
1475 
1476 void vhost_dev_cleanup(struct vhost_dev *hdev)
1477 {
1478     int i;
1479 
1480     trace_vhost_dev_cleanup(hdev);
1481 
1482     for (i = 0; i < hdev->nvqs; ++i) {
1483         vhost_virtqueue_cleanup(hdev->vqs + i);
1484     }
1485     if (hdev->mem) {
1486         /* those are only safe after successful init */
1487         memory_listener_unregister(&hdev->memory_listener);
1488         QLIST_REMOVE(hdev, entry);
1489     }
1490     if (hdev->migration_blocker) {
1491         migrate_del_blocker(hdev->migration_blocker);
1492         error_free(hdev->migration_blocker);
1493     }
1494     g_free(hdev->mem);
1495     g_free(hdev->mem_sections);
1496     if (hdev->vhost_ops) {
1497         hdev->vhost_ops->vhost_backend_cleanup(hdev);
1498     }
1499     assert(!hdev->log);
1500 
1501     memset(hdev, 0, sizeof(struct vhost_dev));
1502 }
1503 
1504 /* Stop processing guest IO notifications in qemu.
1505  * Start processing them in vhost in kernel.
1506  */
1507 int vhost_dev_enable_notifiers(struct vhost_dev *hdev, VirtIODevice *vdev)
1508 {
1509     BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev)));
1510     int i, r, e;
1511 
1512     /* We will pass the notifiers to the kernel, make sure that QEMU
1513      * doesn't interfere.
1514      */
1515     r = virtio_device_grab_ioeventfd(vdev);
1516     if (r < 0) {
1517         error_report("binding does not support host notifiers");
1518         goto fail;
1519     }
1520 
1521     for (i = 0; i < hdev->nvqs; ++i) {
1522         r = virtio_bus_set_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i,
1523                                          true);
1524         if (r < 0) {
1525             error_report("vhost VQ %d notifier binding failed: %d", i, -r);
1526             goto fail_vq;
1527         }
1528     }
1529 
1530     return 0;
1531 fail_vq:
1532     while (--i >= 0) {
1533         e = virtio_bus_set_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i,
1534                                          false);
1535         if (e < 0) {
1536             error_report("vhost VQ %d notifier cleanup error: %d", i, -r);
1537         }
1538         assert (e >= 0);
1539         virtio_bus_cleanup_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i);
1540     }
1541     virtio_device_release_ioeventfd(vdev);
1542 fail:
1543     return r;
1544 }
1545 
1546 /* Stop processing guest IO notifications in vhost.
1547  * Start processing them in qemu.
1548  * This might actually run the qemu handlers right away,
1549  * so virtio in qemu must be completely setup when this is called.
1550  */
1551 void vhost_dev_disable_notifiers(struct vhost_dev *hdev, VirtIODevice *vdev)
1552 {
1553     BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev)));
1554     int i, r;
1555 
1556     for (i = 0; i < hdev->nvqs; ++i) {
1557         r = virtio_bus_set_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i,
1558                                          false);
1559         if (r < 0) {
1560             error_report("vhost VQ %d notifier cleanup failed: %d", i, -r);
1561         }
1562         assert (r >= 0);
1563         virtio_bus_cleanup_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i);
1564     }
1565     virtio_device_release_ioeventfd(vdev);
1566 }
1567 
1568 /* Test and clear event pending status.
1569  * Should be called after unmask to avoid losing events.
1570  */
1571 bool vhost_virtqueue_pending(struct vhost_dev *hdev, int n)
1572 {
1573     struct vhost_virtqueue *vq = hdev->vqs + n - hdev->vq_index;
1574     assert(n >= hdev->vq_index && n < hdev->vq_index + hdev->nvqs);
1575     return event_notifier_test_and_clear(&vq->masked_notifier);
1576 }
1577 
1578 /* Mask/unmask events from this vq. */
1579 void vhost_virtqueue_mask(struct vhost_dev *hdev, VirtIODevice *vdev, int n,
1580                          bool mask)
1581 {
1582     struct VirtQueue *vvq = virtio_get_queue(vdev, n);
1583     int r, index = n - hdev->vq_index;
1584     struct vhost_vring_file file;
1585 
1586     /* should only be called after backend is connected */
1587     assert(hdev->vhost_ops);
1588 
1589     if (mask) {
1590         assert(vdev->use_guest_notifier_mask);
1591         file.fd = event_notifier_get_wfd(&hdev->vqs[index].masked_notifier);
1592     } else {
1593         file.fd = event_notifier_get_wfd(virtio_queue_get_guest_notifier(vvq));
1594     }
1595 
1596     file.index = hdev->vhost_ops->vhost_get_vq_index(hdev, n);
1597     r = hdev->vhost_ops->vhost_set_vring_call(hdev, &file);
1598     if (r < 0) {
1599         VHOST_OPS_DEBUG(r, "vhost_set_vring_call failed");
1600     }
1601 }
1602 
1603 uint64_t vhost_get_features(struct vhost_dev *hdev, const int *feature_bits,
1604                             uint64_t features)
1605 {
1606     const int *bit = feature_bits;
1607     while (*bit != VHOST_INVALID_FEATURE_BIT) {
1608         uint64_t bit_mask = (1ULL << *bit);
1609         if (!(hdev->features & bit_mask)) {
1610             features &= ~bit_mask;
1611         }
1612         bit++;
1613     }
1614     return features;
1615 }
1616 
1617 void vhost_ack_features(struct vhost_dev *hdev, const int *feature_bits,
1618                         uint64_t features)
1619 {
1620     const int *bit = feature_bits;
1621     while (*bit != VHOST_INVALID_FEATURE_BIT) {
1622         uint64_t bit_mask = (1ULL << *bit);
1623         if (features & bit_mask) {
1624             hdev->acked_features |= bit_mask;
1625         }
1626         bit++;
1627     }
1628 }
1629 
1630 int vhost_dev_get_config(struct vhost_dev *hdev, uint8_t *config,
1631                          uint32_t config_len, Error **errp)
1632 {
1633     assert(hdev->vhost_ops);
1634 
1635     if (hdev->vhost_ops->vhost_get_config) {
1636         return hdev->vhost_ops->vhost_get_config(hdev, config, config_len,
1637                                                  errp);
1638     }
1639 
1640     error_setg(errp, "vhost_get_config not implemented");
1641     return -ENOSYS;
1642 }
1643 
1644 int vhost_dev_set_config(struct vhost_dev *hdev, const uint8_t *data,
1645                          uint32_t offset, uint32_t size, uint32_t flags)
1646 {
1647     assert(hdev->vhost_ops);
1648 
1649     if (hdev->vhost_ops->vhost_set_config) {
1650         return hdev->vhost_ops->vhost_set_config(hdev, data, offset,
1651                                                  size, flags);
1652     }
1653 
1654     return -ENOSYS;
1655 }
1656 
1657 void vhost_dev_set_config_notifier(struct vhost_dev *hdev,
1658                                    const VhostDevConfigOps *ops)
1659 {
1660     hdev->config_ops = ops;
1661 }
1662 
1663 void vhost_dev_free_inflight(struct vhost_inflight *inflight)
1664 {
1665     if (inflight && inflight->addr) {
1666         qemu_memfd_free(inflight->addr, inflight->size, inflight->fd);
1667         inflight->addr = NULL;
1668         inflight->fd = -1;
1669     }
1670 }
1671 
1672 static int vhost_dev_resize_inflight(struct vhost_inflight *inflight,
1673                                      uint64_t new_size)
1674 {
1675     Error *err = NULL;
1676     int fd = -1;
1677     void *addr = qemu_memfd_alloc("vhost-inflight", new_size,
1678                                   F_SEAL_GROW | F_SEAL_SHRINK | F_SEAL_SEAL,
1679                                   &fd, &err);
1680 
1681     if (err) {
1682         error_report_err(err);
1683         return -ENOMEM;
1684     }
1685 
1686     vhost_dev_free_inflight(inflight);
1687     inflight->offset = 0;
1688     inflight->addr = addr;
1689     inflight->fd = fd;
1690     inflight->size = new_size;
1691 
1692     return 0;
1693 }
1694 
1695 void vhost_dev_save_inflight(struct vhost_inflight *inflight, QEMUFile *f)
1696 {
1697     if (inflight->addr) {
1698         qemu_put_be64(f, inflight->size);
1699         qemu_put_be16(f, inflight->queue_size);
1700         qemu_put_buffer(f, inflight->addr, inflight->size);
1701     } else {
1702         qemu_put_be64(f, 0);
1703     }
1704 }
1705 
1706 int vhost_dev_load_inflight(struct vhost_inflight *inflight, QEMUFile *f)
1707 {
1708     uint64_t size;
1709 
1710     size = qemu_get_be64(f);
1711     if (!size) {
1712         return 0;
1713     }
1714 
1715     if (inflight->size != size) {
1716         int ret = vhost_dev_resize_inflight(inflight, size);
1717         if (ret < 0) {
1718             return ret;
1719         }
1720     }
1721     inflight->queue_size = qemu_get_be16(f);
1722 
1723     qemu_get_buffer(f, inflight->addr, size);
1724 
1725     return 0;
1726 }
1727 
1728 int vhost_dev_prepare_inflight(struct vhost_dev *hdev, VirtIODevice *vdev)
1729 {
1730     int r;
1731 
1732     if (hdev->vhost_ops->vhost_get_inflight_fd == NULL ||
1733         hdev->vhost_ops->vhost_set_inflight_fd == NULL) {
1734         return 0;
1735     }
1736 
1737     hdev->vdev = vdev;
1738 
1739     r = vhost_dev_set_features(hdev, hdev->log_enabled);
1740     if (r < 0) {
1741         VHOST_OPS_DEBUG(r, "vhost_dev_prepare_inflight failed");
1742         return r;
1743     }
1744 
1745     return 0;
1746 }
1747 
1748 int vhost_dev_set_inflight(struct vhost_dev *dev,
1749                            struct vhost_inflight *inflight)
1750 {
1751     int r;
1752 
1753     if (dev->vhost_ops->vhost_set_inflight_fd && inflight->addr) {
1754         r = dev->vhost_ops->vhost_set_inflight_fd(dev, inflight);
1755         if (r) {
1756             VHOST_OPS_DEBUG(r, "vhost_set_inflight_fd failed");
1757             return r;
1758         }
1759     }
1760 
1761     return 0;
1762 }
1763 
1764 int vhost_dev_get_inflight(struct vhost_dev *dev, uint16_t queue_size,
1765                            struct vhost_inflight *inflight)
1766 {
1767     int r;
1768 
1769     if (dev->vhost_ops->vhost_get_inflight_fd) {
1770         r = dev->vhost_ops->vhost_get_inflight_fd(dev, queue_size, inflight);
1771         if (r) {
1772             VHOST_OPS_DEBUG(r, "vhost_get_inflight_fd failed");
1773             return r;
1774         }
1775     }
1776 
1777     return 0;
1778 }
1779 
1780 static int vhost_dev_set_vring_enable(struct vhost_dev *hdev, int enable)
1781 {
1782     if (!hdev->vhost_ops->vhost_set_vring_enable) {
1783         return 0;
1784     }
1785 
1786     /*
1787      * For vhost-user devices, if VHOST_USER_F_PROTOCOL_FEATURES has not
1788      * been negotiated, the rings start directly in the enabled state, and
1789      * .vhost_set_vring_enable callback will fail since
1790      * VHOST_USER_SET_VRING_ENABLE is not supported.
1791      */
1792     if (hdev->vhost_ops->backend_type == VHOST_BACKEND_TYPE_USER &&
1793         !virtio_has_feature(hdev->backend_features,
1794                             VHOST_USER_F_PROTOCOL_FEATURES)) {
1795         return 0;
1796     }
1797 
1798     return hdev->vhost_ops->vhost_set_vring_enable(hdev, enable);
1799 }
1800 
1801 /* Host notifiers must be enabled at this point. */
1802 int vhost_dev_start(struct vhost_dev *hdev, VirtIODevice *vdev, bool vrings)
1803 {
1804     int i, r;
1805 
1806     /* should only be called after backend is connected */
1807     assert(hdev->vhost_ops);
1808 
1809     trace_vhost_dev_start(hdev, vdev->name, vrings);
1810 
1811     vdev->vhost_started = true;
1812     hdev->started = true;
1813     hdev->vdev = vdev;
1814 
1815     r = vhost_dev_set_features(hdev, hdev->log_enabled);
1816     if (r < 0) {
1817         goto fail_features;
1818     }
1819 
1820     if (vhost_dev_has_iommu(hdev)) {
1821         memory_listener_register(&hdev->iommu_listener, vdev->dma_as);
1822     }
1823 
1824     r = hdev->vhost_ops->vhost_set_mem_table(hdev, hdev->mem);
1825     if (r < 0) {
1826         VHOST_OPS_DEBUG(r, "vhost_set_mem_table failed");
1827         goto fail_mem;
1828     }
1829     for (i = 0; i < hdev->nvqs; ++i) {
1830         r = vhost_virtqueue_start(hdev,
1831                                   vdev,
1832                                   hdev->vqs + i,
1833                                   hdev->vq_index + i);
1834         if (r < 0) {
1835             goto fail_vq;
1836         }
1837     }
1838 
1839     if (hdev->log_enabled) {
1840         uint64_t log_base;
1841 
1842         hdev->log_size = vhost_get_log_size(hdev);
1843         hdev->log = vhost_log_get(hdev->log_size,
1844                                   vhost_dev_log_is_shared(hdev));
1845         log_base = (uintptr_t)hdev->log->log;
1846         r = hdev->vhost_ops->vhost_set_log_base(hdev,
1847                                                 hdev->log_size ? log_base : 0,
1848                                                 hdev->log);
1849         if (r < 0) {
1850             VHOST_OPS_DEBUG(r, "vhost_set_log_base failed");
1851             goto fail_log;
1852         }
1853     }
1854     if (vrings) {
1855         r = vhost_dev_set_vring_enable(hdev, true);
1856         if (r) {
1857             goto fail_log;
1858         }
1859     }
1860     if (hdev->vhost_ops->vhost_dev_start) {
1861         r = hdev->vhost_ops->vhost_dev_start(hdev, true);
1862         if (r) {
1863             goto fail_start;
1864         }
1865     }
1866     if (vhost_dev_has_iommu(hdev) &&
1867         hdev->vhost_ops->vhost_set_iotlb_callback) {
1868             hdev->vhost_ops->vhost_set_iotlb_callback(hdev, true);
1869 
1870         /* Update used ring information for IOTLB to work correctly,
1871          * vhost-kernel code requires for this.*/
1872         for (i = 0; i < hdev->nvqs; ++i) {
1873             struct vhost_virtqueue *vq = hdev->vqs + i;
1874             vhost_device_iotlb_miss(hdev, vq->used_phys, true);
1875         }
1876     }
1877     return 0;
1878 fail_start:
1879     if (vrings) {
1880         vhost_dev_set_vring_enable(hdev, false);
1881     }
1882 fail_log:
1883     vhost_log_put(hdev, false);
1884 fail_vq:
1885     while (--i >= 0) {
1886         vhost_virtqueue_stop(hdev,
1887                              vdev,
1888                              hdev->vqs + i,
1889                              hdev->vq_index + i);
1890     }
1891 
1892 fail_mem:
1893 fail_features:
1894     vdev->vhost_started = false;
1895     hdev->started = false;
1896     return r;
1897 }
1898 
1899 /* Host notifiers must be enabled at this point. */
1900 void vhost_dev_stop(struct vhost_dev *hdev, VirtIODevice *vdev, bool vrings)
1901 {
1902     int i;
1903 
1904     /* should only be called after backend is connected */
1905     assert(hdev->vhost_ops);
1906 
1907     trace_vhost_dev_stop(hdev, vdev->name, vrings);
1908 
1909     if (hdev->vhost_ops->vhost_dev_start) {
1910         hdev->vhost_ops->vhost_dev_start(hdev, false);
1911     }
1912     if (vrings) {
1913         vhost_dev_set_vring_enable(hdev, false);
1914     }
1915     for (i = 0; i < hdev->nvqs; ++i) {
1916         vhost_virtqueue_stop(hdev,
1917                              vdev,
1918                              hdev->vqs + i,
1919                              hdev->vq_index + i);
1920     }
1921 
1922     if (vhost_dev_has_iommu(hdev)) {
1923         if (hdev->vhost_ops->vhost_set_iotlb_callback) {
1924             hdev->vhost_ops->vhost_set_iotlb_callback(hdev, false);
1925         }
1926         memory_listener_unregister(&hdev->iommu_listener);
1927     }
1928     vhost_log_put(hdev, true);
1929     hdev->started = false;
1930     vdev->vhost_started = false;
1931     hdev->vdev = NULL;
1932 }
1933 
1934 int vhost_net_set_backend(struct vhost_dev *hdev,
1935                           struct vhost_vring_file *file)
1936 {
1937     if (hdev->vhost_ops->vhost_net_set_backend) {
1938         return hdev->vhost_ops->vhost_net_set_backend(hdev, file);
1939     }
1940 
1941     return -ENOSYS;
1942 }
1943