xref: /openbmc/qemu/hw/virtio/vhost.c (revision 75c5bb0b)
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 "exec/address-spaces.h"
25 #include "hw/virtio/virtio-bus.h"
26 #include "hw/virtio/virtio-access.h"
27 #include "migration/blocker.h"
28 #include "migration/qemu-file-types.h"
29 #include "sysemu/dma.h"
30 #include "trace.h"
31 
32 /* enabled until disconnected backend stabilizes */
33 #define _VHOST_DEBUG 1
34 
35 #ifdef _VHOST_DEBUG
36 #define VHOST_OPS_DEBUG(fmt, ...) \
37     do { error_report(fmt ": %s (%d)", ## __VA_ARGS__, \
38                       strerror(errno), errno); } while (0)
39 #else
40 #define VHOST_OPS_DEBUG(fmt, ...) \
41     do { } while (0)
42 #endif
43 
44 static struct vhost_log *vhost_log;
45 static struct vhost_log *vhost_log_shm;
46 
47 static unsigned int used_memslots;
48 static QLIST_HEAD(, vhost_dev) vhost_devices =
49     QLIST_HEAD_INITIALIZER(vhost_devices);
50 
51 bool vhost_has_free_slot(void)
52 {
53     unsigned int slots_limit = ~0U;
54     struct vhost_dev *hdev;
55 
56     QLIST_FOREACH(hdev, &vhost_devices, entry) {
57         unsigned int r = hdev->vhost_ops->vhost_backend_memslots_limit(hdev);
58         slots_limit = MIN(slots_limit, r);
59     }
60     return slots_limit > used_memslots;
61 }
62 
63 static void vhost_dev_sync_region(struct vhost_dev *dev,
64                                   MemoryRegionSection *section,
65                                   uint64_t mfirst, uint64_t mlast,
66                                   uint64_t rfirst, uint64_t rlast)
67 {
68     vhost_log_chunk_t *log = dev->log->log;
69 
70     uint64_t start = MAX(mfirst, rfirst);
71     uint64_t end = MIN(mlast, rlast);
72     vhost_log_chunk_t *from = log + start / VHOST_LOG_CHUNK;
73     vhost_log_chunk_t *to = log + end / VHOST_LOG_CHUNK + 1;
74     uint64_t addr = QEMU_ALIGN_DOWN(start, VHOST_LOG_CHUNK);
75 
76     if (end < start) {
77         return;
78     }
79     assert(end / VHOST_LOG_CHUNK < dev->log_size);
80     assert(start / VHOST_LOG_CHUNK < dev->log_size);
81 
82     for (;from < to; ++from) {
83         vhost_log_chunk_t log;
84         /* We first check with non-atomic: much cheaper,
85          * and we expect non-dirty to be the common case. */
86         if (!*from) {
87             addr += VHOST_LOG_CHUNK;
88             continue;
89         }
90         /* Data must be read atomically. We don't really need barrier semantics
91          * but it's easier to use atomic_* than roll our own. */
92         log = atomic_xchg(from, 0);
93         while (log) {
94             int bit = ctzl(log);
95             hwaddr page_addr;
96             hwaddr section_offset;
97             hwaddr mr_offset;
98             page_addr = addr + bit * VHOST_LOG_PAGE;
99             section_offset = page_addr - section->offset_within_address_space;
100             mr_offset = section_offset + section->offset_within_region;
101             memory_region_set_dirty(section->mr, mr_offset, VHOST_LOG_PAGE);
102             log &= ~(0x1ull << bit);
103         }
104         addr += VHOST_LOG_CHUNK;
105     }
106 }
107 
108 static int vhost_sync_dirty_bitmap(struct vhost_dev *dev,
109                                    MemoryRegionSection *section,
110                                    hwaddr first,
111                                    hwaddr last)
112 {
113     int i;
114     hwaddr start_addr;
115     hwaddr end_addr;
116 
117     if (!dev->log_enabled || !dev->started) {
118         return 0;
119     }
120     start_addr = section->offset_within_address_space;
121     end_addr = range_get_last(start_addr, int128_get64(section->size));
122     start_addr = MAX(first, start_addr);
123     end_addr = MIN(last, end_addr);
124 
125     for (i = 0; i < dev->mem->nregions; ++i) {
126         struct vhost_memory_region *reg = dev->mem->regions + i;
127         vhost_dev_sync_region(dev, section, start_addr, end_addr,
128                               reg->guest_phys_addr,
129                               range_get_last(reg->guest_phys_addr,
130                                              reg->memory_size));
131     }
132     for (i = 0; i < dev->nvqs; ++i) {
133         struct vhost_virtqueue *vq = dev->vqs + i;
134 
135         if (!vq->used_phys && !vq->used_size) {
136             continue;
137         }
138 
139         vhost_dev_sync_region(dev, section, start_addr, end_addr, vq->used_phys,
140                               range_get_last(vq->used_phys, vq->used_size));
141     }
142     return 0;
143 }
144 
145 static void vhost_log_sync(MemoryListener *listener,
146                           MemoryRegionSection *section)
147 {
148     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
149                                          memory_listener);
150     vhost_sync_dirty_bitmap(dev, section, 0x0, ~0x0ULL);
151 }
152 
153 static void vhost_log_sync_range(struct vhost_dev *dev,
154                                  hwaddr first, hwaddr last)
155 {
156     int i;
157     /* FIXME: this is N^2 in number of sections */
158     for (i = 0; i < dev->n_mem_sections; ++i) {
159         MemoryRegionSection *section = &dev->mem_sections[i];
160         vhost_sync_dirty_bitmap(dev, section, first, last);
161     }
162 }
163 
164 static uint64_t vhost_get_log_size(struct vhost_dev *dev)
165 {
166     uint64_t log_size = 0;
167     int i;
168     for (i = 0; i < dev->mem->nregions; ++i) {
169         struct vhost_memory_region *reg = dev->mem->regions + i;
170         uint64_t last = range_get_last(reg->guest_phys_addr,
171                                        reg->memory_size);
172         log_size = MAX(log_size, last / VHOST_LOG_CHUNK + 1);
173     }
174     for (i = 0; i < dev->nvqs; ++i) {
175         struct vhost_virtqueue *vq = dev->vqs + i;
176 
177         if (!vq->used_phys && !vq->used_size) {
178             continue;
179         }
180 
181         uint64_t last = vq->used_phys + vq->used_size - 1;
182         log_size = MAX(log_size, last / VHOST_LOG_CHUNK + 1);
183     }
184     return log_size;
185 }
186 
187 static struct vhost_log *vhost_log_alloc(uint64_t size, bool share)
188 {
189     Error *err = NULL;
190     struct vhost_log *log;
191     uint64_t logsize = size * sizeof(*(log->log));
192     int fd = -1;
193 
194     log = g_new0(struct vhost_log, 1);
195     if (share) {
196         log->log = qemu_memfd_alloc("vhost-log", logsize,
197                                     F_SEAL_GROW | F_SEAL_SHRINK | F_SEAL_SEAL,
198                                     &fd, &err);
199         if (err) {
200             error_report_err(err);
201             g_free(log);
202             return NULL;
203         }
204         memset(log->log, 0, logsize);
205     } else {
206         log->log = g_malloc0(logsize);
207     }
208 
209     log->size = size;
210     log->refcnt = 1;
211     log->fd = fd;
212 
213     return log;
214 }
215 
216 static struct vhost_log *vhost_log_get(uint64_t size, bool share)
217 {
218     struct vhost_log *log = share ? vhost_log_shm : vhost_log;
219 
220     if (!log || log->size != size) {
221         log = vhost_log_alloc(size, share);
222         if (share) {
223             vhost_log_shm = log;
224         } else {
225             vhost_log = log;
226         }
227     } else {
228         ++log->refcnt;
229     }
230 
231     return log;
232 }
233 
234 static void vhost_log_put(struct vhost_dev *dev, bool sync)
235 {
236     struct vhost_log *log = dev->log;
237 
238     if (!log) {
239         return;
240     }
241 
242     --log->refcnt;
243     if (log->refcnt == 0) {
244         /* Sync only the range covered by the old log */
245         if (dev->log_size && sync) {
246             vhost_log_sync_range(dev, 0, dev->log_size * VHOST_LOG_CHUNK - 1);
247         }
248 
249         if (vhost_log == log) {
250             g_free(log->log);
251             vhost_log = NULL;
252         } else if (vhost_log_shm == log) {
253             qemu_memfd_free(log->log, log->size * sizeof(*(log->log)),
254                             log->fd);
255             vhost_log_shm = NULL;
256         }
257 
258         g_free(log);
259     }
260 
261     dev->log = NULL;
262     dev->log_size = 0;
263 }
264 
265 static bool vhost_dev_log_is_shared(struct vhost_dev *dev)
266 {
267     return dev->vhost_ops->vhost_requires_shm_log &&
268            dev->vhost_ops->vhost_requires_shm_log(dev);
269 }
270 
271 static inline void vhost_dev_log_resize(struct vhost_dev *dev, uint64_t size)
272 {
273     struct vhost_log *log = vhost_log_get(size, vhost_dev_log_is_shared(dev));
274     uint64_t log_base = (uintptr_t)log->log;
275     int r;
276 
277     /* inform backend of log switching, this must be done before
278        releasing the current log, to ensure no logging is lost */
279     r = dev->vhost_ops->vhost_set_log_base(dev, log_base, log);
280     if (r < 0) {
281         VHOST_OPS_DEBUG("vhost_set_log_base failed");
282     }
283 
284     vhost_log_put(dev, true);
285     dev->log = log;
286     dev->log_size = size;
287 }
288 
289 static int vhost_dev_has_iommu(struct vhost_dev *dev)
290 {
291     VirtIODevice *vdev = dev->vdev;
292 
293     return virtio_host_has_feature(vdev, VIRTIO_F_IOMMU_PLATFORM);
294 }
295 
296 static void *vhost_memory_map(struct vhost_dev *dev, hwaddr addr,
297                               hwaddr *plen, int is_write)
298 {
299     if (!vhost_dev_has_iommu(dev)) {
300         return cpu_physical_memory_map(addr, plen, is_write);
301     } else {
302         return (void *)(uintptr_t)addr;
303     }
304 }
305 
306 static void vhost_memory_unmap(struct vhost_dev *dev, void *buffer,
307                                hwaddr len, int is_write,
308                                hwaddr access_len)
309 {
310     if (!vhost_dev_has_iommu(dev)) {
311         cpu_physical_memory_unmap(buffer, len, is_write, access_len);
312     }
313 }
314 
315 static int vhost_verify_ring_part_mapping(void *ring_hva,
316                                           uint64_t ring_gpa,
317                                           uint64_t ring_size,
318                                           void *reg_hva,
319                                           uint64_t reg_gpa,
320                                           uint64_t reg_size)
321 {
322     uint64_t hva_ring_offset;
323     uint64_t ring_last = range_get_last(ring_gpa, ring_size);
324     uint64_t reg_last = range_get_last(reg_gpa, reg_size);
325 
326     if (ring_last < reg_gpa || ring_gpa > reg_last) {
327         return 0;
328     }
329     /* check that whole ring's is mapped */
330     if (ring_last > reg_last) {
331         return -ENOMEM;
332     }
333     /* check that ring's MemoryRegion wasn't replaced */
334     hva_ring_offset = ring_gpa - reg_gpa;
335     if (ring_hva != reg_hva + hva_ring_offset) {
336         return -EBUSY;
337     }
338 
339     return 0;
340 }
341 
342 static int vhost_verify_ring_mappings(struct vhost_dev *dev,
343                                       void *reg_hva,
344                                       uint64_t reg_gpa,
345                                       uint64_t reg_size)
346 {
347     int i, j;
348     int r = 0;
349     const char *part_name[] = {
350         "descriptor table",
351         "available ring",
352         "used ring"
353     };
354 
355     if (vhost_dev_has_iommu(dev)) {
356         return 0;
357     }
358 
359     for (i = 0; i < dev->nvqs; ++i) {
360         struct vhost_virtqueue *vq = dev->vqs + i;
361 
362         if (vq->desc_phys == 0) {
363             continue;
364         }
365 
366         j = 0;
367         r = vhost_verify_ring_part_mapping(
368                 vq->desc, vq->desc_phys, vq->desc_size,
369                 reg_hva, reg_gpa, reg_size);
370         if (r) {
371             break;
372         }
373 
374         j++;
375         r = vhost_verify_ring_part_mapping(
376                 vq->avail, vq->avail_phys, vq->avail_size,
377                 reg_hva, reg_gpa, reg_size);
378         if (r) {
379             break;
380         }
381 
382         j++;
383         r = vhost_verify_ring_part_mapping(
384                 vq->used, vq->used_phys, vq->used_size,
385                 reg_hva, reg_gpa, reg_size);
386         if (r) {
387             break;
388         }
389     }
390 
391     if (r == -ENOMEM) {
392         error_report("Unable to map %s for ring %d", part_name[j], i);
393     } else if (r == -EBUSY) {
394         error_report("%s relocated for ring %d", part_name[j], i);
395     }
396     return r;
397 }
398 
399 static bool vhost_section(struct vhost_dev *dev, MemoryRegionSection *section)
400 {
401     bool result;
402     bool log_dirty = memory_region_get_dirty_log_mask(section->mr) &
403                      ~(1 << DIRTY_MEMORY_MIGRATION);
404     result = memory_region_is_ram(section->mr) &&
405         !memory_region_is_rom(section->mr);
406 
407     /* Vhost doesn't handle any block which is doing dirty-tracking other
408      * than migration; this typically fires on VGA areas.
409      */
410     result &= !log_dirty;
411 
412     if (result && dev->vhost_ops->vhost_backend_mem_section_filter) {
413         result &=
414             dev->vhost_ops->vhost_backend_mem_section_filter(dev, section);
415     }
416 
417     trace_vhost_section(section->mr->name, result);
418     return result;
419 }
420 
421 static void vhost_begin(MemoryListener *listener)
422 {
423     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
424                                          memory_listener);
425     dev->tmp_sections = NULL;
426     dev->n_tmp_sections = 0;
427 }
428 
429 static void vhost_commit(MemoryListener *listener)
430 {
431     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
432                                          memory_listener);
433     MemoryRegionSection *old_sections;
434     int n_old_sections;
435     uint64_t log_size;
436     size_t regions_size;
437     int r;
438     int i;
439     bool changed = false;
440 
441     /* Note we can be called before the device is started, but then
442      * starting the device calls set_mem_table, so we need to have
443      * built the data structures.
444      */
445     old_sections = dev->mem_sections;
446     n_old_sections = dev->n_mem_sections;
447     dev->mem_sections = dev->tmp_sections;
448     dev->n_mem_sections = dev->n_tmp_sections;
449 
450     if (dev->n_mem_sections != n_old_sections) {
451         changed = true;
452     } else {
453         /* Same size, lets check the contents */
454         for (int i = 0; i < n_old_sections; i++) {
455             if (!MemoryRegionSection_eq(&old_sections[i],
456                                         &dev->mem_sections[i])) {
457                 changed = true;
458                 break;
459             }
460         }
461     }
462 
463     trace_vhost_commit(dev->started, changed);
464     if (!changed) {
465         goto out;
466     }
467 
468     /* Rebuild the regions list from the new sections list */
469     regions_size = offsetof(struct vhost_memory, regions) +
470                        dev->n_mem_sections * sizeof dev->mem->regions[0];
471     dev->mem = g_realloc(dev->mem, regions_size);
472     dev->mem->nregions = dev->n_mem_sections;
473     used_memslots = dev->mem->nregions;
474     for (i = 0; i < dev->n_mem_sections; i++) {
475         struct vhost_memory_region *cur_vmr = dev->mem->regions + i;
476         struct MemoryRegionSection *mrs = dev->mem_sections + i;
477 
478         cur_vmr->guest_phys_addr = mrs->offset_within_address_space;
479         cur_vmr->memory_size     = int128_get64(mrs->size);
480         cur_vmr->userspace_addr  =
481             (uintptr_t)memory_region_get_ram_ptr(mrs->mr) +
482             mrs->offset_within_region;
483         cur_vmr->flags_padding   = 0;
484     }
485 
486     if (!dev->started) {
487         goto out;
488     }
489 
490     for (i = 0; i < dev->mem->nregions; i++) {
491         if (vhost_verify_ring_mappings(dev,
492                        (void *)(uintptr_t)dev->mem->regions[i].userspace_addr,
493                        dev->mem->regions[i].guest_phys_addr,
494                        dev->mem->regions[i].memory_size)) {
495             error_report("Verify ring failure on region %d", i);
496             abort();
497         }
498     }
499 
500     if (!dev->log_enabled) {
501         r = dev->vhost_ops->vhost_set_mem_table(dev, dev->mem);
502         if (r < 0) {
503             VHOST_OPS_DEBUG("vhost_set_mem_table failed");
504         }
505         goto out;
506     }
507     log_size = vhost_get_log_size(dev);
508     /* We allocate an extra 4K bytes to log,
509      * to reduce the * number of reallocations. */
510 #define VHOST_LOG_BUFFER (0x1000 / sizeof *dev->log)
511     /* To log more, must increase log size before table update. */
512     if (dev->log_size < log_size) {
513         vhost_dev_log_resize(dev, log_size + VHOST_LOG_BUFFER);
514     }
515     r = dev->vhost_ops->vhost_set_mem_table(dev, dev->mem);
516     if (r < 0) {
517         VHOST_OPS_DEBUG("vhost_set_mem_table failed");
518     }
519     /* To log less, can only decrease log size after table update. */
520     if (dev->log_size > log_size + VHOST_LOG_BUFFER) {
521         vhost_dev_log_resize(dev, log_size);
522     }
523 
524 out:
525     /* Deref the old list of sections, this must happen _after_ the
526      * vhost_set_mem_table to ensure the client isn't still using the
527      * section we're about to unref.
528      */
529     while (n_old_sections--) {
530         memory_region_unref(old_sections[n_old_sections].mr);
531     }
532     g_free(old_sections);
533     return;
534 }
535 
536 /* Adds the section data to the tmp_section structure.
537  * It relies on the listener calling us in memory address order
538  * and for each region (via the _add and _nop methods) to
539  * join neighbours.
540  */
541 static void vhost_region_add_section(struct vhost_dev *dev,
542                                      MemoryRegionSection *section)
543 {
544     bool need_add = true;
545     uint64_t mrs_size = int128_get64(section->size);
546     uint64_t mrs_gpa = section->offset_within_address_space;
547     uintptr_t mrs_host = (uintptr_t)memory_region_get_ram_ptr(section->mr) +
548                          section->offset_within_region;
549     RAMBlock *mrs_rb = section->mr->ram_block;
550     size_t mrs_page = qemu_ram_pagesize(mrs_rb);
551 
552     trace_vhost_region_add_section(section->mr->name, mrs_gpa, mrs_size,
553                                    mrs_host);
554 
555     /* Round the section to it's page size */
556     /* First align the start down to a page boundary */
557     uint64_t alignage = mrs_host & (mrs_page - 1);
558     if (alignage) {
559         mrs_host -= alignage;
560         mrs_size += alignage;
561         mrs_gpa  -= alignage;
562     }
563     /* Now align the size up to a page boundary */
564     alignage = mrs_size & (mrs_page - 1);
565     if (alignage) {
566         mrs_size += mrs_page - alignage;
567     }
568     trace_vhost_region_add_section_aligned(section->mr->name, mrs_gpa, mrs_size,
569                                            mrs_host);
570 
571     if (dev->n_tmp_sections) {
572         /* Since we already have at least one section, lets see if
573          * this extends it; since we're scanning in order, we only
574          * have to look at the last one, and the FlatView that calls
575          * us shouldn't have overlaps.
576          */
577         MemoryRegionSection *prev_sec = dev->tmp_sections +
578                                                (dev->n_tmp_sections - 1);
579         uint64_t prev_gpa_start = prev_sec->offset_within_address_space;
580         uint64_t prev_size = int128_get64(prev_sec->size);
581         uint64_t prev_gpa_end   = range_get_last(prev_gpa_start, prev_size);
582         uint64_t prev_host_start =
583                         (uintptr_t)memory_region_get_ram_ptr(prev_sec->mr) +
584                         prev_sec->offset_within_region;
585         uint64_t prev_host_end   = range_get_last(prev_host_start, prev_size);
586 
587         if (mrs_gpa <= (prev_gpa_end + 1)) {
588             /* OK, looks like overlapping/intersecting - it's possible that
589              * the rounding to page sizes has made them overlap, but they should
590              * match up in the same RAMBlock if they do.
591              */
592             if (mrs_gpa < prev_gpa_start) {
593                 error_report("%s:Section rounded to %"PRIx64
594                              " prior to previous %"PRIx64,
595                              __func__, mrs_gpa, prev_gpa_start);
596                 /* A way to cleanly fail here would be better */
597                 return;
598             }
599             /* Offset from the start of the previous GPA to this GPA */
600             size_t offset = mrs_gpa - prev_gpa_start;
601 
602             if (prev_host_start + offset == mrs_host &&
603                 section->mr == prev_sec->mr &&
604                 (!dev->vhost_ops->vhost_backend_can_merge ||
605                  dev->vhost_ops->vhost_backend_can_merge(dev,
606                     mrs_host, mrs_size,
607                     prev_host_start, prev_size))) {
608                 uint64_t max_end = MAX(prev_host_end, mrs_host + mrs_size);
609                 need_add = false;
610                 prev_sec->offset_within_address_space =
611                     MIN(prev_gpa_start, mrs_gpa);
612                 prev_sec->offset_within_region =
613                     MIN(prev_host_start, mrs_host) -
614                     (uintptr_t)memory_region_get_ram_ptr(prev_sec->mr);
615                 prev_sec->size = int128_make64(max_end - MIN(prev_host_start,
616                                                mrs_host));
617                 trace_vhost_region_add_section_merge(section->mr->name,
618                                         int128_get64(prev_sec->size),
619                                         prev_sec->offset_within_address_space,
620                                         prev_sec->offset_within_region);
621             } else {
622                 /* adjoining regions are fine, but overlapping ones with
623                  * different blocks/offsets shouldn't happen
624                  */
625                 if (mrs_gpa != prev_gpa_end + 1) {
626                     error_report("%s: Overlapping but not coherent sections "
627                                  "at %"PRIx64,
628                                  __func__, mrs_gpa);
629                     return;
630                 }
631             }
632         }
633     }
634 
635     if (need_add) {
636         ++dev->n_tmp_sections;
637         dev->tmp_sections = g_renew(MemoryRegionSection, dev->tmp_sections,
638                                     dev->n_tmp_sections);
639         dev->tmp_sections[dev->n_tmp_sections - 1] = *section;
640         /* The flatview isn't stable and we don't use it, making it NULL
641          * means we can memcmp the list.
642          */
643         dev->tmp_sections[dev->n_tmp_sections - 1].fv = NULL;
644         memory_region_ref(section->mr);
645     }
646 }
647 
648 /* Used for both add and nop callbacks */
649 static void vhost_region_addnop(MemoryListener *listener,
650                                 MemoryRegionSection *section)
651 {
652     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
653                                          memory_listener);
654 
655     if (!vhost_section(dev, section)) {
656         return;
657     }
658     vhost_region_add_section(dev, section);
659 }
660 
661 static void vhost_iommu_unmap_notify(IOMMUNotifier *n, IOMMUTLBEntry *iotlb)
662 {
663     struct vhost_iommu *iommu = container_of(n, struct vhost_iommu, n);
664     struct vhost_dev *hdev = iommu->hdev;
665     hwaddr iova = iotlb->iova + iommu->iommu_offset;
666 
667     if (vhost_backend_invalidate_device_iotlb(hdev, iova,
668                                               iotlb->addr_mask + 1)) {
669         error_report("Fail to invalidate device iotlb");
670     }
671 }
672 
673 static void vhost_iommu_region_add(MemoryListener *listener,
674                                    MemoryRegionSection *section)
675 {
676     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
677                                          iommu_listener);
678     struct vhost_iommu *iommu;
679     Int128 end;
680     int iommu_idx, ret;
681     IOMMUMemoryRegion *iommu_mr;
682     Error *err = NULL;
683 
684     if (!memory_region_is_iommu(section->mr)) {
685         return;
686     }
687 
688     iommu_mr = IOMMU_MEMORY_REGION(section->mr);
689 
690     iommu = g_malloc0(sizeof(*iommu));
691     end = int128_add(int128_make64(section->offset_within_region),
692                      section->size);
693     end = int128_sub(end, int128_one());
694     iommu_idx = memory_region_iommu_attrs_to_index(iommu_mr,
695                                                    MEMTXATTRS_UNSPECIFIED);
696     iommu_notifier_init(&iommu->n, vhost_iommu_unmap_notify,
697                         IOMMU_NOTIFIER_UNMAP,
698                         section->offset_within_region,
699                         int128_get64(end),
700                         iommu_idx);
701     iommu->mr = section->mr;
702     iommu->iommu_offset = section->offset_within_address_space -
703                           section->offset_within_region;
704     iommu->hdev = dev;
705     ret = memory_region_register_iommu_notifier(section->mr, &iommu->n, &err);
706     if (ret) {
707         error_report_err(err);
708         exit(1);
709     }
710     QLIST_INSERT_HEAD(&dev->iommu_list, iommu, iommu_next);
711     /* TODO: can replay help performance here? */
712 }
713 
714 static void vhost_iommu_region_del(MemoryListener *listener,
715                                    MemoryRegionSection *section)
716 {
717     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
718                                          iommu_listener);
719     struct vhost_iommu *iommu;
720 
721     if (!memory_region_is_iommu(section->mr)) {
722         return;
723     }
724 
725     QLIST_FOREACH(iommu, &dev->iommu_list, iommu_next) {
726         if (iommu->mr == section->mr &&
727             iommu->n.start == section->offset_within_region) {
728             memory_region_unregister_iommu_notifier(iommu->mr,
729                                                     &iommu->n);
730             QLIST_REMOVE(iommu, iommu_next);
731             g_free(iommu);
732             break;
733         }
734     }
735 }
736 
737 static int vhost_virtqueue_set_addr(struct vhost_dev *dev,
738                                     struct vhost_virtqueue *vq,
739                                     unsigned idx, bool enable_log)
740 {
741     struct vhost_vring_addr addr = {
742         .index = idx,
743         .desc_user_addr = (uint64_t)(unsigned long)vq->desc,
744         .avail_user_addr = (uint64_t)(unsigned long)vq->avail,
745         .used_user_addr = (uint64_t)(unsigned long)vq->used,
746         .log_guest_addr = vq->used_phys,
747         .flags = enable_log ? (1 << VHOST_VRING_F_LOG) : 0,
748     };
749     int r = dev->vhost_ops->vhost_set_vring_addr(dev, &addr);
750     if (r < 0) {
751         VHOST_OPS_DEBUG("vhost_set_vring_addr failed");
752         return -errno;
753     }
754     return 0;
755 }
756 
757 static int vhost_dev_set_features(struct vhost_dev *dev,
758                                   bool enable_log)
759 {
760     uint64_t features = dev->acked_features;
761     int r;
762     if (enable_log) {
763         features |= 0x1ULL << VHOST_F_LOG_ALL;
764     }
765     r = dev->vhost_ops->vhost_set_features(dev, features);
766     if (r < 0) {
767         VHOST_OPS_DEBUG("vhost_set_features failed");
768     }
769     return r < 0 ? -errno : 0;
770 }
771 
772 static int vhost_dev_set_log(struct vhost_dev *dev, bool enable_log)
773 {
774     int r, i, idx;
775     r = vhost_dev_set_features(dev, enable_log);
776     if (r < 0) {
777         goto err_features;
778     }
779     for (i = 0; i < dev->nvqs; ++i) {
780         idx = dev->vhost_ops->vhost_get_vq_index(dev, dev->vq_index + i);
781         r = vhost_virtqueue_set_addr(dev, dev->vqs + i, idx,
782                                      enable_log);
783         if (r < 0) {
784             goto err_vq;
785         }
786     }
787     return 0;
788 err_vq:
789     for (; i >= 0; --i) {
790         idx = dev->vhost_ops->vhost_get_vq_index(dev, dev->vq_index + i);
791         vhost_virtqueue_set_addr(dev, dev->vqs + i, idx,
792                                  dev->log_enabled);
793     }
794     vhost_dev_set_features(dev, dev->log_enabled);
795 err_features:
796     return r;
797 }
798 
799 static int vhost_migration_log(MemoryListener *listener, int enable)
800 {
801     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
802                                          memory_listener);
803     int r;
804     if (!!enable == dev->log_enabled) {
805         return 0;
806     }
807     if (!dev->started) {
808         dev->log_enabled = enable;
809         return 0;
810     }
811     if (!enable) {
812         r = vhost_dev_set_log(dev, false);
813         if (r < 0) {
814             return r;
815         }
816         vhost_log_put(dev, false);
817     } else {
818         vhost_dev_log_resize(dev, vhost_get_log_size(dev));
819         r = vhost_dev_set_log(dev, true);
820         if (r < 0) {
821             return r;
822         }
823     }
824     dev->log_enabled = enable;
825     return 0;
826 }
827 
828 static void vhost_log_global_start(MemoryListener *listener)
829 {
830     int r;
831 
832     r = vhost_migration_log(listener, true);
833     if (r < 0) {
834         abort();
835     }
836 }
837 
838 static void vhost_log_global_stop(MemoryListener *listener)
839 {
840     int r;
841 
842     r = vhost_migration_log(listener, false);
843     if (r < 0) {
844         abort();
845     }
846 }
847 
848 static void vhost_log_start(MemoryListener *listener,
849                             MemoryRegionSection *section,
850                             int old, int new)
851 {
852     /* FIXME: implement */
853 }
854 
855 static void vhost_log_stop(MemoryListener *listener,
856                            MemoryRegionSection *section,
857                            int old, int new)
858 {
859     /* FIXME: implement */
860 }
861 
862 /* The vhost driver natively knows how to handle the vrings of non
863  * cross-endian legacy devices and modern devices. Only legacy devices
864  * exposed to a bi-endian guest may require the vhost driver to use a
865  * specific endianness.
866  */
867 static inline bool vhost_needs_vring_endian(VirtIODevice *vdev)
868 {
869     if (virtio_vdev_has_feature(vdev, VIRTIO_F_VERSION_1)) {
870         return false;
871     }
872 #ifdef HOST_WORDS_BIGENDIAN
873     return vdev->device_endian == VIRTIO_DEVICE_ENDIAN_LITTLE;
874 #else
875     return vdev->device_endian == VIRTIO_DEVICE_ENDIAN_BIG;
876 #endif
877 }
878 
879 static int vhost_virtqueue_set_vring_endian_legacy(struct vhost_dev *dev,
880                                                    bool is_big_endian,
881                                                    int vhost_vq_index)
882 {
883     struct vhost_vring_state s = {
884         .index = vhost_vq_index,
885         .num = is_big_endian
886     };
887 
888     if (!dev->vhost_ops->vhost_set_vring_endian(dev, &s)) {
889         return 0;
890     }
891 
892     VHOST_OPS_DEBUG("vhost_set_vring_endian failed");
893     if (errno == ENOTTY) {
894         error_report("vhost does not support cross-endian");
895         return -ENOSYS;
896     }
897 
898     return -errno;
899 }
900 
901 static int vhost_memory_region_lookup(struct vhost_dev *hdev,
902                                       uint64_t gpa, uint64_t *uaddr,
903                                       uint64_t *len)
904 {
905     int i;
906 
907     for (i = 0; i < hdev->mem->nregions; i++) {
908         struct vhost_memory_region *reg = hdev->mem->regions + i;
909 
910         if (gpa >= reg->guest_phys_addr &&
911             reg->guest_phys_addr + reg->memory_size > gpa) {
912             *uaddr = reg->userspace_addr + gpa - reg->guest_phys_addr;
913             *len = reg->guest_phys_addr + reg->memory_size - gpa;
914             return 0;
915         }
916     }
917 
918     return -EFAULT;
919 }
920 
921 int vhost_device_iotlb_miss(struct vhost_dev *dev, uint64_t iova, int write)
922 {
923     IOMMUTLBEntry iotlb;
924     uint64_t uaddr, len;
925     int ret = -EFAULT;
926 
927     RCU_READ_LOCK_GUARD();
928 
929     trace_vhost_iotlb_miss(dev, 1);
930 
931     iotlb = address_space_get_iotlb_entry(dev->vdev->dma_as,
932                                           iova, write,
933                                           MEMTXATTRS_UNSPECIFIED);
934     if (iotlb.target_as != NULL) {
935         ret = vhost_memory_region_lookup(dev, iotlb.translated_addr,
936                                          &uaddr, &len);
937         if (ret) {
938             trace_vhost_iotlb_miss(dev, 3);
939             error_report("Fail to lookup the translated address "
940                          "%"PRIx64, iotlb.translated_addr);
941             goto out;
942         }
943 
944         len = MIN(iotlb.addr_mask + 1, len);
945         iova = iova & ~iotlb.addr_mask;
946 
947         ret = vhost_backend_update_device_iotlb(dev, iova, uaddr,
948                                                 len, iotlb.perm);
949         if (ret) {
950             trace_vhost_iotlb_miss(dev, 4);
951             error_report("Fail to update device iotlb");
952             goto out;
953         }
954     }
955 
956     trace_vhost_iotlb_miss(dev, 2);
957 
958 out:
959     return ret;
960 }
961 
962 static int vhost_virtqueue_start(struct vhost_dev *dev,
963                                 struct VirtIODevice *vdev,
964                                 struct vhost_virtqueue *vq,
965                                 unsigned idx)
966 {
967     BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev)));
968     VirtioBusState *vbus = VIRTIO_BUS(qbus);
969     VirtioBusClass *k = VIRTIO_BUS_GET_CLASS(vbus);
970     hwaddr s, l, a;
971     int r;
972     int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, idx);
973     struct vhost_vring_file file = {
974         .index = vhost_vq_index
975     };
976     struct vhost_vring_state state = {
977         .index = vhost_vq_index
978     };
979     struct VirtQueue *vvq = virtio_get_queue(vdev, idx);
980 
981     a = virtio_queue_get_desc_addr(vdev, idx);
982     if (a == 0) {
983         /* Queue might not be ready for start */
984         return 0;
985     }
986 
987     vq->num = state.num = virtio_queue_get_num(vdev, idx);
988     r = dev->vhost_ops->vhost_set_vring_num(dev, &state);
989     if (r) {
990         VHOST_OPS_DEBUG("vhost_set_vring_num failed");
991         return -errno;
992     }
993 
994     state.num = virtio_queue_get_last_avail_idx(vdev, idx);
995     r = dev->vhost_ops->vhost_set_vring_base(dev, &state);
996     if (r) {
997         VHOST_OPS_DEBUG("vhost_set_vring_base failed");
998         return -errno;
999     }
1000 
1001     if (vhost_needs_vring_endian(vdev)) {
1002         r = vhost_virtqueue_set_vring_endian_legacy(dev,
1003                                                     virtio_is_big_endian(vdev),
1004                                                     vhost_vq_index);
1005         if (r) {
1006             return -errno;
1007         }
1008     }
1009 
1010     vq->desc_size = s = l = virtio_queue_get_desc_size(vdev, idx);
1011     vq->desc_phys = a;
1012     vq->desc = vhost_memory_map(dev, a, &l, 0);
1013     if (!vq->desc || l != s) {
1014         r = -ENOMEM;
1015         goto fail_alloc_desc;
1016     }
1017     vq->avail_size = s = l = virtio_queue_get_avail_size(vdev, idx);
1018     vq->avail_phys = a = virtio_queue_get_avail_addr(vdev, idx);
1019     vq->avail = vhost_memory_map(dev, a, &l, 0);
1020     if (!vq->avail || l != s) {
1021         r = -ENOMEM;
1022         goto fail_alloc_avail;
1023     }
1024     vq->used_size = s = l = virtio_queue_get_used_size(vdev, idx);
1025     vq->used_phys = a = virtio_queue_get_used_addr(vdev, idx);
1026     vq->used = vhost_memory_map(dev, a, &l, 1);
1027     if (!vq->used || l != s) {
1028         r = -ENOMEM;
1029         goto fail_alloc_used;
1030     }
1031 
1032     r = vhost_virtqueue_set_addr(dev, vq, vhost_vq_index, dev->log_enabled);
1033     if (r < 0) {
1034         r = -errno;
1035         goto fail_alloc;
1036     }
1037 
1038     file.fd = event_notifier_get_fd(virtio_queue_get_host_notifier(vvq));
1039     r = dev->vhost_ops->vhost_set_vring_kick(dev, &file);
1040     if (r) {
1041         VHOST_OPS_DEBUG("vhost_set_vring_kick failed");
1042         r = -errno;
1043         goto fail_kick;
1044     }
1045 
1046     /* Clear and discard previous events if any. */
1047     event_notifier_test_and_clear(&vq->masked_notifier);
1048 
1049     /* Init vring in unmasked state, unless guest_notifier_mask
1050      * will do it later.
1051      */
1052     if (!vdev->use_guest_notifier_mask) {
1053         /* TODO: check and handle errors. */
1054         vhost_virtqueue_mask(dev, vdev, idx, false);
1055     }
1056 
1057     if (k->query_guest_notifiers &&
1058         k->query_guest_notifiers(qbus->parent) &&
1059         virtio_queue_vector(vdev, idx) == VIRTIO_NO_VECTOR) {
1060         file.fd = -1;
1061         r = dev->vhost_ops->vhost_set_vring_call(dev, &file);
1062         if (r) {
1063             goto fail_vector;
1064         }
1065     }
1066 
1067     return 0;
1068 
1069 fail_vector:
1070 fail_kick:
1071 fail_alloc:
1072     vhost_memory_unmap(dev, vq->used, virtio_queue_get_used_size(vdev, idx),
1073                        0, 0);
1074 fail_alloc_used:
1075     vhost_memory_unmap(dev, vq->avail, virtio_queue_get_avail_size(vdev, idx),
1076                        0, 0);
1077 fail_alloc_avail:
1078     vhost_memory_unmap(dev, vq->desc, virtio_queue_get_desc_size(vdev, idx),
1079                        0, 0);
1080 fail_alloc_desc:
1081     return r;
1082 }
1083 
1084 static void vhost_virtqueue_stop(struct vhost_dev *dev,
1085                                     struct VirtIODevice *vdev,
1086                                     struct vhost_virtqueue *vq,
1087                                     unsigned idx)
1088 {
1089     int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, idx);
1090     struct vhost_vring_state state = {
1091         .index = vhost_vq_index,
1092     };
1093     int r;
1094 
1095     if (virtio_queue_get_desc_addr(vdev, idx) == 0) {
1096         /* Don't stop the virtqueue which might have not been started */
1097         return;
1098     }
1099 
1100     r = dev->vhost_ops->vhost_get_vring_base(dev, &state);
1101     if (r < 0) {
1102         VHOST_OPS_DEBUG("vhost VQ %u ring restore failed: %d", idx, r);
1103         /* Connection to the backend is broken, so let's sync internal
1104          * last avail idx to the device used idx.
1105          */
1106         virtio_queue_restore_last_avail_idx(vdev, idx);
1107     } else {
1108         virtio_queue_set_last_avail_idx(vdev, idx, state.num);
1109     }
1110     virtio_queue_invalidate_signalled_used(vdev, idx);
1111     virtio_queue_update_used_idx(vdev, idx);
1112 
1113     /* In the cross-endian case, we need to reset the vring endianness to
1114      * native as legacy devices expect so by default.
1115      */
1116     if (vhost_needs_vring_endian(vdev)) {
1117         vhost_virtqueue_set_vring_endian_legacy(dev,
1118                                                 !virtio_is_big_endian(vdev),
1119                                                 vhost_vq_index);
1120     }
1121 
1122     vhost_memory_unmap(dev, vq->used, virtio_queue_get_used_size(vdev, idx),
1123                        1, virtio_queue_get_used_size(vdev, idx));
1124     vhost_memory_unmap(dev, vq->avail, virtio_queue_get_avail_size(vdev, idx),
1125                        0, virtio_queue_get_avail_size(vdev, idx));
1126     vhost_memory_unmap(dev, vq->desc, virtio_queue_get_desc_size(vdev, idx),
1127                        0, virtio_queue_get_desc_size(vdev, idx));
1128 }
1129 
1130 static void vhost_eventfd_add(MemoryListener *listener,
1131                               MemoryRegionSection *section,
1132                               bool match_data, uint64_t data, EventNotifier *e)
1133 {
1134 }
1135 
1136 static void vhost_eventfd_del(MemoryListener *listener,
1137                               MemoryRegionSection *section,
1138                               bool match_data, uint64_t data, EventNotifier *e)
1139 {
1140 }
1141 
1142 static int vhost_virtqueue_set_busyloop_timeout(struct vhost_dev *dev,
1143                                                 int n, uint32_t timeout)
1144 {
1145     int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, n);
1146     struct vhost_vring_state state = {
1147         .index = vhost_vq_index,
1148         .num = timeout,
1149     };
1150     int r;
1151 
1152     if (!dev->vhost_ops->vhost_set_vring_busyloop_timeout) {
1153         return -EINVAL;
1154     }
1155 
1156     r = dev->vhost_ops->vhost_set_vring_busyloop_timeout(dev, &state);
1157     if (r) {
1158         VHOST_OPS_DEBUG("vhost_set_vring_busyloop_timeout failed");
1159         return r;
1160     }
1161 
1162     return 0;
1163 }
1164 
1165 static int vhost_virtqueue_init(struct vhost_dev *dev,
1166                                 struct vhost_virtqueue *vq, int n)
1167 {
1168     int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, n);
1169     struct vhost_vring_file file = {
1170         .index = vhost_vq_index,
1171     };
1172     int r = event_notifier_init(&vq->masked_notifier, 0);
1173     if (r < 0) {
1174         return r;
1175     }
1176 
1177     file.fd = event_notifier_get_fd(&vq->masked_notifier);
1178     r = dev->vhost_ops->vhost_set_vring_call(dev, &file);
1179     if (r) {
1180         VHOST_OPS_DEBUG("vhost_set_vring_call failed");
1181         r = -errno;
1182         goto fail_call;
1183     }
1184 
1185     vq->dev = dev;
1186 
1187     return 0;
1188 fail_call:
1189     event_notifier_cleanup(&vq->masked_notifier);
1190     return r;
1191 }
1192 
1193 static void vhost_virtqueue_cleanup(struct vhost_virtqueue *vq)
1194 {
1195     event_notifier_cleanup(&vq->masked_notifier);
1196 }
1197 
1198 int vhost_dev_init(struct vhost_dev *hdev, void *opaque,
1199                    VhostBackendType backend_type, uint32_t busyloop_timeout)
1200 {
1201     uint64_t features;
1202     int i, r, n_initialized_vqs = 0;
1203     Error *local_err = NULL;
1204 
1205     hdev->vdev = NULL;
1206     hdev->migration_blocker = NULL;
1207 
1208     r = vhost_set_backend_type(hdev, backend_type);
1209     assert(r >= 0);
1210 
1211     r = hdev->vhost_ops->vhost_backend_init(hdev, opaque);
1212     if (r < 0) {
1213         goto fail;
1214     }
1215 
1216     r = hdev->vhost_ops->vhost_set_owner(hdev);
1217     if (r < 0) {
1218         VHOST_OPS_DEBUG("vhost_set_owner failed");
1219         goto fail;
1220     }
1221 
1222     r = hdev->vhost_ops->vhost_get_features(hdev, &features);
1223     if (r < 0) {
1224         VHOST_OPS_DEBUG("vhost_get_features failed");
1225         goto fail;
1226     }
1227 
1228     for (i = 0; i < hdev->nvqs; ++i, ++n_initialized_vqs) {
1229         r = vhost_virtqueue_init(hdev, hdev->vqs + i, hdev->vq_index + i);
1230         if (r < 0) {
1231             goto fail;
1232         }
1233     }
1234 
1235     if (busyloop_timeout) {
1236         for (i = 0; i < hdev->nvqs; ++i) {
1237             r = vhost_virtqueue_set_busyloop_timeout(hdev, hdev->vq_index + i,
1238                                                      busyloop_timeout);
1239             if (r < 0) {
1240                 goto fail_busyloop;
1241             }
1242         }
1243     }
1244 
1245     hdev->features = features;
1246 
1247     hdev->memory_listener = (MemoryListener) {
1248         .begin = vhost_begin,
1249         .commit = vhost_commit,
1250         .region_add = vhost_region_addnop,
1251         .region_nop = vhost_region_addnop,
1252         .log_start = vhost_log_start,
1253         .log_stop = vhost_log_stop,
1254         .log_sync = vhost_log_sync,
1255         .log_global_start = vhost_log_global_start,
1256         .log_global_stop = vhost_log_global_stop,
1257         .eventfd_add = vhost_eventfd_add,
1258         .eventfd_del = vhost_eventfd_del,
1259         .priority = 10
1260     };
1261 
1262     hdev->iommu_listener = (MemoryListener) {
1263         .region_add = vhost_iommu_region_add,
1264         .region_del = vhost_iommu_region_del,
1265     };
1266 
1267     if (hdev->migration_blocker == NULL) {
1268         if (!(hdev->features & (0x1ULL << VHOST_F_LOG_ALL))) {
1269             error_setg(&hdev->migration_blocker,
1270                        "Migration disabled: vhost lacks VHOST_F_LOG_ALL feature.");
1271         } else if (vhost_dev_log_is_shared(hdev) && !qemu_memfd_alloc_check()) {
1272             error_setg(&hdev->migration_blocker,
1273                        "Migration disabled: failed to allocate shared memory");
1274         }
1275     }
1276 
1277     if (hdev->migration_blocker != NULL) {
1278         r = migrate_add_blocker(hdev->migration_blocker, &local_err);
1279         if (local_err) {
1280             error_report_err(local_err);
1281             error_free(hdev->migration_blocker);
1282             goto fail_busyloop;
1283         }
1284     }
1285 
1286     hdev->mem = g_malloc0(offsetof(struct vhost_memory, regions));
1287     hdev->n_mem_sections = 0;
1288     hdev->mem_sections = NULL;
1289     hdev->log = NULL;
1290     hdev->log_size = 0;
1291     hdev->log_enabled = false;
1292     hdev->started = false;
1293     memory_listener_register(&hdev->memory_listener, &address_space_memory);
1294     QLIST_INSERT_HEAD(&vhost_devices, hdev, entry);
1295 
1296     if (used_memslots > hdev->vhost_ops->vhost_backend_memslots_limit(hdev)) {
1297         error_report("vhost backend memory slots limit is less"
1298                 " than current number of present memory slots");
1299         r = -1;
1300         if (busyloop_timeout) {
1301             goto fail_busyloop;
1302         } else {
1303             goto fail;
1304         }
1305     }
1306 
1307     return 0;
1308 
1309 fail_busyloop:
1310     while (--i >= 0) {
1311         vhost_virtqueue_set_busyloop_timeout(hdev, hdev->vq_index + i, 0);
1312     }
1313 fail:
1314     hdev->nvqs = n_initialized_vqs;
1315     vhost_dev_cleanup(hdev);
1316     return r;
1317 }
1318 
1319 void vhost_dev_cleanup(struct vhost_dev *hdev)
1320 {
1321     int i;
1322 
1323     for (i = 0; i < hdev->nvqs; ++i) {
1324         vhost_virtqueue_cleanup(hdev->vqs + i);
1325     }
1326     if (hdev->mem) {
1327         /* those are only safe after successful init */
1328         memory_listener_unregister(&hdev->memory_listener);
1329         QLIST_REMOVE(hdev, entry);
1330     }
1331     if (hdev->migration_blocker) {
1332         migrate_del_blocker(hdev->migration_blocker);
1333         error_free(hdev->migration_blocker);
1334     }
1335     g_free(hdev->mem);
1336     g_free(hdev->mem_sections);
1337     if (hdev->vhost_ops) {
1338         hdev->vhost_ops->vhost_backend_cleanup(hdev);
1339     }
1340     assert(!hdev->log);
1341 
1342     memset(hdev, 0, sizeof(struct vhost_dev));
1343 }
1344 
1345 /* Stop processing guest IO notifications in qemu.
1346  * Start processing them in vhost in kernel.
1347  */
1348 int vhost_dev_enable_notifiers(struct vhost_dev *hdev, VirtIODevice *vdev)
1349 {
1350     BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev)));
1351     int i, r, e;
1352 
1353     /* We will pass the notifiers to the kernel, make sure that QEMU
1354      * doesn't interfere.
1355      */
1356     r = virtio_device_grab_ioeventfd(vdev);
1357     if (r < 0) {
1358         error_report("binding does not support host notifiers");
1359         goto fail;
1360     }
1361 
1362     for (i = 0; i < hdev->nvqs; ++i) {
1363         r = virtio_bus_set_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i,
1364                                          true);
1365         if (r < 0) {
1366             error_report("vhost VQ %d notifier binding failed: %d", i, -r);
1367             goto fail_vq;
1368         }
1369     }
1370 
1371     return 0;
1372 fail_vq:
1373     while (--i >= 0) {
1374         e = virtio_bus_set_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i,
1375                                          false);
1376         if (e < 0) {
1377             error_report("vhost VQ %d notifier cleanup error: %d", i, -r);
1378         }
1379         assert (e >= 0);
1380         virtio_bus_cleanup_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i);
1381     }
1382     virtio_device_release_ioeventfd(vdev);
1383 fail:
1384     return r;
1385 }
1386 
1387 /* Stop processing guest IO notifications in vhost.
1388  * Start processing them in qemu.
1389  * This might actually run the qemu handlers right away,
1390  * so virtio in qemu must be completely setup when this is called.
1391  */
1392 void vhost_dev_disable_notifiers(struct vhost_dev *hdev, VirtIODevice *vdev)
1393 {
1394     BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev)));
1395     int i, r;
1396 
1397     for (i = 0; i < hdev->nvqs; ++i) {
1398         r = virtio_bus_set_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i,
1399                                          false);
1400         if (r < 0) {
1401             error_report("vhost VQ %d notifier cleanup failed: %d", i, -r);
1402         }
1403         assert (r >= 0);
1404         virtio_bus_cleanup_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i);
1405     }
1406     virtio_device_release_ioeventfd(vdev);
1407 }
1408 
1409 /* Test and clear event pending status.
1410  * Should be called after unmask to avoid losing events.
1411  */
1412 bool vhost_virtqueue_pending(struct vhost_dev *hdev, int n)
1413 {
1414     struct vhost_virtqueue *vq = hdev->vqs + n - hdev->vq_index;
1415     assert(n >= hdev->vq_index && n < hdev->vq_index + hdev->nvqs);
1416     return event_notifier_test_and_clear(&vq->masked_notifier);
1417 }
1418 
1419 /* Mask/unmask events from this vq. */
1420 void vhost_virtqueue_mask(struct vhost_dev *hdev, VirtIODevice *vdev, int n,
1421                          bool mask)
1422 {
1423     struct VirtQueue *vvq = virtio_get_queue(vdev, n);
1424     int r, index = n - hdev->vq_index;
1425     struct vhost_vring_file file;
1426 
1427     /* should only be called after backend is connected */
1428     assert(hdev->vhost_ops);
1429 
1430     if (mask) {
1431         assert(vdev->use_guest_notifier_mask);
1432         file.fd = event_notifier_get_fd(&hdev->vqs[index].masked_notifier);
1433     } else {
1434         file.fd = event_notifier_get_fd(virtio_queue_get_guest_notifier(vvq));
1435     }
1436 
1437     file.index = hdev->vhost_ops->vhost_get_vq_index(hdev, n);
1438     r = hdev->vhost_ops->vhost_set_vring_call(hdev, &file);
1439     if (r < 0) {
1440         VHOST_OPS_DEBUG("vhost_set_vring_call failed");
1441     }
1442 }
1443 
1444 uint64_t vhost_get_features(struct vhost_dev *hdev, const int *feature_bits,
1445                             uint64_t features)
1446 {
1447     const int *bit = feature_bits;
1448     while (*bit != VHOST_INVALID_FEATURE_BIT) {
1449         uint64_t bit_mask = (1ULL << *bit);
1450         if (!(hdev->features & bit_mask)) {
1451             features &= ~bit_mask;
1452         }
1453         bit++;
1454     }
1455     return features;
1456 }
1457 
1458 void vhost_ack_features(struct vhost_dev *hdev, const int *feature_bits,
1459                         uint64_t features)
1460 {
1461     const int *bit = feature_bits;
1462     while (*bit != VHOST_INVALID_FEATURE_BIT) {
1463         uint64_t bit_mask = (1ULL << *bit);
1464         if (features & bit_mask) {
1465             hdev->acked_features |= bit_mask;
1466         }
1467         bit++;
1468     }
1469 }
1470 
1471 int vhost_dev_get_config(struct vhost_dev *hdev, uint8_t *config,
1472                          uint32_t config_len)
1473 {
1474     assert(hdev->vhost_ops);
1475 
1476     if (hdev->vhost_ops->vhost_get_config) {
1477         return hdev->vhost_ops->vhost_get_config(hdev, config, config_len);
1478     }
1479 
1480     return -1;
1481 }
1482 
1483 int vhost_dev_set_config(struct vhost_dev *hdev, const uint8_t *data,
1484                          uint32_t offset, uint32_t size, uint32_t flags)
1485 {
1486     assert(hdev->vhost_ops);
1487 
1488     if (hdev->vhost_ops->vhost_set_config) {
1489         return hdev->vhost_ops->vhost_set_config(hdev, data, offset,
1490                                                  size, flags);
1491     }
1492 
1493     return -1;
1494 }
1495 
1496 void vhost_dev_set_config_notifier(struct vhost_dev *hdev,
1497                                    const VhostDevConfigOps *ops)
1498 {
1499     hdev->config_ops = ops;
1500 }
1501 
1502 void vhost_dev_free_inflight(struct vhost_inflight *inflight)
1503 {
1504     if (inflight->addr) {
1505         qemu_memfd_free(inflight->addr, inflight->size, inflight->fd);
1506         inflight->addr = NULL;
1507         inflight->fd = -1;
1508     }
1509 }
1510 
1511 static int vhost_dev_resize_inflight(struct vhost_inflight *inflight,
1512                                      uint64_t new_size)
1513 {
1514     Error *err = NULL;
1515     int fd = -1;
1516     void *addr = qemu_memfd_alloc("vhost-inflight", new_size,
1517                                   F_SEAL_GROW | F_SEAL_SHRINK | F_SEAL_SEAL,
1518                                   &fd, &err);
1519 
1520     if (err) {
1521         error_report_err(err);
1522         return -1;
1523     }
1524 
1525     vhost_dev_free_inflight(inflight);
1526     inflight->offset = 0;
1527     inflight->addr = addr;
1528     inflight->fd = fd;
1529     inflight->size = new_size;
1530 
1531     return 0;
1532 }
1533 
1534 void vhost_dev_save_inflight(struct vhost_inflight *inflight, QEMUFile *f)
1535 {
1536     if (inflight->addr) {
1537         qemu_put_be64(f, inflight->size);
1538         qemu_put_be16(f, inflight->queue_size);
1539         qemu_put_buffer(f, inflight->addr, inflight->size);
1540     } else {
1541         qemu_put_be64(f, 0);
1542     }
1543 }
1544 
1545 int vhost_dev_load_inflight(struct vhost_inflight *inflight, QEMUFile *f)
1546 {
1547     uint64_t size;
1548 
1549     size = qemu_get_be64(f);
1550     if (!size) {
1551         return 0;
1552     }
1553 
1554     if (inflight->size != size) {
1555         if (vhost_dev_resize_inflight(inflight, size)) {
1556             return -1;
1557         }
1558     }
1559     inflight->queue_size = qemu_get_be16(f);
1560 
1561     qemu_get_buffer(f, inflight->addr, size);
1562 
1563     return 0;
1564 }
1565 
1566 int vhost_dev_set_inflight(struct vhost_dev *dev,
1567                            struct vhost_inflight *inflight)
1568 {
1569     int r;
1570 
1571     if (dev->vhost_ops->vhost_set_inflight_fd && inflight->addr) {
1572         r = dev->vhost_ops->vhost_set_inflight_fd(dev, inflight);
1573         if (r) {
1574             VHOST_OPS_DEBUG("vhost_set_inflight_fd failed");
1575             return -errno;
1576         }
1577     }
1578 
1579     return 0;
1580 }
1581 
1582 int vhost_dev_get_inflight(struct vhost_dev *dev, uint16_t queue_size,
1583                            struct vhost_inflight *inflight)
1584 {
1585     int r;
1586 
1587     if (dev->vhost_ops->vhost_get_inflight_fd) {
1588         r = dev->vhost_ops->vhost_get_inflight_fd(dev, queue_size, inflight);
1589         if (r) {
1590             VHOST_OPS_DEBUG("vhost_get_inflight_fd failed");
1591             return -errno;
1592         }
1593     }
1594 
1595     return 0;
1596 }
1597 
1598 /* Host notifiers must be enabled at this point. */
1599 int vhost_dev_start(struct vhost_dev *hdev, VirtIODevice *vdev)
1600 {
1601     int i, r;
1602 
1603     /* should only be called after backend is connected */
1604     assert(hdev->vhost_ops);
1605 
1606     hdev->started = true;
1607     hdev->vdev = vdev;
1608 
1609     r = vhost_dev_set_features(hdev, hdev->log_enabled);
1610     if (r < 0) {
1611         goto fail_features;
1612     }
1613 
1614     if (vhost_dev_has_iommu(hdev)) {
1615         memory_listener_register(&hdev->iommu_listener, vdev->dma_as);
1616     }
1617 
1618     r = hdev->vhost_ops->vhost_set_mem_table(hdev, hdev->mem);
1619     if (r < 0) {
1620         VHOST_OPS_DEBUG("vhost_set_mem_table failed");
1621         r = -errno;
1622         goto fail_mem;
1623     }
1624     for (i = 0; i < hdev->nvqs; ++i) {
1625         r = vhost_virtqueue_start(hdev,
1626                                   vdev,
1627                                   hdev->vqs + i,
1628                                   hdev->vq_index + i);
1629         if (r < 0) {
1630             goto fail_vq;
1631         }
1632     }
1633 
1634     if (hdev->log_enabled) {
1635         uint64_t log_base;
1636 
1637         hdev->log_size = vhost_get_log_size(hdev);
1638         hdev->log = vhost_log_get(hdev->log_size,
1639                                   vhost_dev_log_is_shared(hdev));
1640         log_base = (uintptr_t)hdev->log->log;
1641         r = hdev->vhost_ops->vhost_set_log_base(hdev,
1642                                                 hdev->log_size ? log_base : 0,
1643                                                 hdev->log);
1644         if (r < 0) {
1645             VHOST_OPS_DEBUG("vhost_set_log_base failed");
1646             r = -errno;
1647             goto fail_log;
1648         }
1649     }
1650 
1651     if (vhost_dev_has_iommu(hdev)) {
1652         hdev->vhost_ops->vhost_set_iotlb_callback(hdev, true);
1653 
1654         /* Update used ring information for IOTLB to work correctly,
1655          * vhost-kernel code requires for this.*/
1656         for (i = 0; i < hdev->nvqs; ++i) {
1657             struct vhost_virtqueue *vq = hdev->vqs + i;
1658             vhost_device_iotlb_miss(hdev, vq->used_phys, true);
1659         }
1660     }
1661     return 0;
1662 fail_log:
1663     vhost_log_put(hdev, false);
1664 fail_vq:
1665     while (--i >= 0) {
1666         vhost_virtqueue_stop(hdev,
1667                              vdev,
1668                              hdev->vqs + i,
1669                              hdev->vq_index + i);
1670     }
1671 
1672 fail_mem:
1673 fail_features:
1674 
1675     hdev->started = false;
1676     return r;
1677 }
1678 
1679 /* Host notifiers must be enabled at this point. */
1680 void vhost_dev_stop(struct vhost_dev *hdev, VirtIODevice *vdev)
1681 {
1682     int i;
1683 
1684     /* should only be called after backend is connected */
1685     assert(hdev->vhost_ops);
1686 
1687     for (i = 0; i < hdev->nvqs; ++i) {
1688         vhost_virtqueue_stop(hdev,
1689                              vdev,
1690                              hdev->vqs + i,
1691                              hdev->vq_index + i);
1692     }
1693 
1694     if (vhost_dev_has_iommu(hdev)) {
1695         hdev->vhost_ops->vhost_set_iotlb_callback(hdev, false);
1696         memory_listener_unregister(&hdev->iommu_listener);
1697     }
1698     vhost_log_put(hdev, true);
1699     hdev->started = false;
1700     hdev->vdev = NULL;
1701 }
1702 
1703 int vhost_net_set_backend(struct vhost_dev *hdev,
1704                           struct vhost_vring_file *file)
1705 {
1706     if (hdev->vhost_ops->vhost_net_set_backend) {
1707         return hdev->vhost_ops->vhost_net_set_backend(hdev, file);
1708     }
1709 
1710     return -1;
1711 }
1712