xref: /openbmc/qemu/hw/virtio/vhost.c (revision b55d7d34)
1 /*
2  * vhost support
3  *
4  * Copyright Red Hat, Inc. 2010
5  *
6  * Authors:
7  *  Michael S. Tsirkin <mst@redhat.com>
8  *
9  * This work is licensed under the terms of the GNU GPL, version 2.  See
10  * the COPYING file in the top-level directory.
11  *
12  * Contributions after 2012-01-13 are licensed under the terms of the
13  * GNU GPL, version 2 or (at your option) any later version.
14  */
15 
16 #include "qemu/osdep.h"
17 #include "qapi/error.h"
18 #include "hw/virtio/vhost.h"
19 #include "hw/hw.h"
20 #include "qemu/atomic.h"
21 #include "qemu/range.h"
22 #include "qemu/error-report.h"
23 #include "qemu/memfd.h"
24 #include "standard-headers/linux/vhost_types.h"
25 #include "exec/address-spaces.h"
26 #include "hw/virtio/virtio-bus.h"
27 #include "hw/virtio/virtio-access.h"
28 #include "migration/blocker.h"
29 #include "sysemu/dma.h"
30 #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         changed = n_old_sections && memcmp(dev->mem_sections, old_sections,
455                          n_old_sections * sizeof(old_sections[0])) != 0;
456     }
457 
458     trace_vhost_commit(dev->started, changed);
459     if (!changed) {
460         goto out;
461     }
462 
463     /* Rebuild the regions list from the new sections list */
464     regions_size = offsetof(struct vhost_memory, regions) +
465                        dev->n_mem_sections * sizeof dev->mem->regions[0];
466     dev->mem = g_realloc(dev->mem, regions_size);
467     dev->mem->nregions = dev->n_mem_sections;
468     used_memslots = dev->mem->nregions;
469     for (i = 0; i < dev->n_mem_sections; i++) {
470         struct vhost_memory_region *cur_vmr = dev->mem->regions + i;
471         struct MemoryRegionSection *mrs = dev->mem_sections + i;
472 
473         cur_vmr->guest_phys_addr = mrs->offset_within_address_space;
474         cur_vmr->memory_size     = int128_get64(mrs->size);
475         cur_vmr->userspace_addr  =
476             (uintptr_t)memory_region_get_ram_ptr(mrs->mr) +
477             mrs->offset_within_region;
478         cur_vmr->flags_padding   = 0;
479     }
480 
481     if (!dev->started) {
482         goto out;
483     }
484 
485     for (i = 0; i < dev->mem->nregions; i++) {
486         if (vhost_verify_ring_mappings(dev,
487                        (void *)(uintptr_t)dev->mem->regions[i].userspace_addr,
488                        dev->mem->regions[i].guest_phys_addr,
489                        dev->mem->regions[i].memory_size)) {
490             error_report("Verify ring failure on region %d", i);
491             abort();
492         }
493     }
494 
495     if (!dev->log_enabled) {
496         r = dev->vhost_ops->vhost_set_mem_table(dev, dev->mem);
497         if (r < 0) {
498             VHOST_OPS_DEBUG("vhost_set_mem_table failed");
499         }
500         goto out;
501     }
502     log_size = vhost_get_log_size(dev);
503     /* We allocate an extra 4K bytes to log,
504      * to reduce the * number of reallocations. */
505 #define VHOST_LOG_BUFFER (0x1000 / sizeof *dev->log)
506     /* To log more, must increase log size before table update. */
507     if (dev->log_size < log_size) {
508         vhost_dev_log_resize(dev, log_size + VHOST_LOG_BUFFER);
509     }
510     r = dev->vhost_ops->vhost_set_mem_table(dev, dev->mem);
511     if (r < 0) {
512         VHOST_OPS_DEBUG("vhost_set_mem_table failed");
513     }
514     /* To log less, can only decrease log size after table update. */
515     if (dev->log_size > log_size + VHOST_LOG_BUFFER) {
516         vhost_dev_log_resize(dev, log_size);
517     }
518 
519 out:
520     /* Deref the old list of sections, this must happen _after_ the
521      * vhost_set_mem_table to ensure the client isn't still using the
522      * section we're about to unref.
523      */
524     while (n_old_sections--) {
525         memory_region_unref(old_sections[n_old_sections].mr);
526     }
527     g_free(old_sections);
528     return;
529 }
530 
531 /* Adds the section data to the tmp_section structure.
532  * It relies on the listener calling us in memory address order
533  * and for each region (via the _add and _nop methods) to
534  * join neighbours.
535  */
536 static void vhost_region_add_section(struct vhost_dev *dev,
537                                      MemoryRegionSection *section)
538 {
539     bool need_add = true;
540     uint64_t mrs_size = int128_get64(section->size);
541     uint64_t mrs_gpa = section->offset_within_address_space;
542     uintptr_t mrs_host = (uintptr_t)memory_region_get_ram_ptr(section->mr) +
543                          section->offset_within_region;
544     RAMBlock *mrs_rb = section->mr->ram_block;
545     size_t mrs_page = qemu_ram_pagesize(mrs_rb);
546 
547     trace_vhost_region_add_section(section->mr->name, mrs_gpa, mrs_size,
548                                    mrs_host);
549 
550     /* Round the section to it's page size */
551     /* First align the start down to a page boundary */
552     uint64_t alignage = mrs_host & (mrs_page - 1);
553     if (alignage) {
554         mrs_host -= alignage;
555         mrs_size += alignage;
556         mrs_gpa  -= alignage;
557     }
558     /* Now align the size up to a page boundary */
559     alignage = mrs_size & (mrs_page - 1);
560     if (alignage) {
561         mrs_size += mrs_page - alignage;
562     }
563     trace_vhost_region_add_section_aligned(section->mr->name, mrs_gpa, mrs_size,
564                                            mrs_host);
565 
566     if (dev->n_tmp_sections) {
567         /* Since we already have at least one section, lets see if
568          * this extends it; since we're scanning in order, we only
569          * have to look at the last one, and the FlatView that calls
570          * us shouldn't have overlaps.
571          */
572         MemoryRegionSection *prev_sec = dev->tmp_sections +
573                                                (dev->n_tmp_sections - 1);
574         uint64_t prev_gpa_start = prev_sec->offset_within_address_space;
575         uint64_t prev_size = int128_get64(prev_sec->size);
576         uint64_t prev_gpa_end   = range_get_last(prev_gpa_start, prev_size);
577         uint64_t prev_host_start =
578                         (uintptr_t)memory_region_get_ram_ptr(prev_sec->mr) +
579                         prev_sec->offset_within_region;
580         uint64_t prev_host_end   = range_get_last(prev_host_start, prev_size);
581 
582         if (mrs_gpa <= (prev_gpa_end + 1)) {
583             /* OK, looks like overlapping/intersecting - it's possible that
584              * the rounding to page sizes has made them overlap, but they should
585              * match up in the same RAMBlock if they do.
586              */
587             if (mrs_gpa < prev_gpa_start) {
588                 error_report("%s:Section rounded to %"PRIx64
589                              " prior to previous %"PRIx64,
590                              __func__, mrs_gpa, prev_gpa_start);
591                 /* A way to cleanly fail here would be better */
592                 return;
593             }
594             /* Offset from the start of the previous GPA to this GPA */
595             size_t offset = mrs_gpa - prev_gpa_start;
596 
597             if (prev_host_start + offset == mrs_host &&
598                 section->mr == prev_sec->mr &&
599                 (!dev->vhost_ops->vhost_backend_can_merge ||
600                  dev->vhost_ops->vhost_backend_can_merge(dev,
601                     mrs_host, mrs_size,
602                     prev_host_start, prev_size))) {
603                 uint64_t max_end = MAX(prev_host_end, mrs_host + mrs_size);
604                 need_add = false;
605                 prev_sec->offset_within_address_space =
606                     MIN(prev_gpa_start, mrs_gpa);
607                 prev_sec->offset_within_region =
608                     MIN(prev_host_start, mrs_host) -
609                     (uintptr_t)memory_region_get_ram_ptr(prev_sec->mr);
610                 prev_sec->size = int128_make64(max_end - MIN(prev_host_start,
611                                                mrs_host));
612                 trace_vhost_region_add_section_merge(section->mr->name,
613                                         int128_get64(prev_sec->size),
614                                         prev_sec->offset_within_address_space,
615                                         prev_sec->offset_within_region);
616             } else {
617                 /* adjoining regions are fine, but overlapping ones with
618                  * different blocks/offsets shouldn't happen
619                  */
620                 if (mrs_gpa != prev_gpa_end + 1) {
621                     error_report("%s: Overlapping but not coherent sections "
622                                  "at %"PRIx64,
623                                  __func__, mrs_gpa);
624                     return;
625                 }
626             }
627         }
628     }
629 
630     if (need_add) {
631         ++dev->n_tmp_sections;
632         dev->tmp_sections = g_renew(MemoryRegionSection, dev->tmp_sections,
633                                     dev->n_tmp_sections);
634         dev->tmp_sections[dev->n_tmp_sections - 1] = *section;
635         /* The flatview isn't stable and we don't use it, making it NULL
636          * means we can memcmp the list.
637          */
638         dev->tmp_sections[dev->n_tmp_sections - 1].fv = NULL;
639         memory_region_ref(section->mr);
640     }
641 }
642 
643 /* Used for both add and nop callbacks */
644 static void vhost_region_addnop(MemoryListener *listener,
645                                 MemoryRegionSection *section)
646 {
647     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
648                                          memory_listener);
649 
650     if (!vhost_section(dev, section)) {
651         return;
652     }
653     vhost_region_add_section(dev, section);
654 }
655 
656 static void vhost_iommu_unmap_notify(IOMMUNotifier *n, IOMMUTLBEntry *iotlb)
657 {
658     struct vhost_iommu *iommu = container_of(n, struct vhost_iommu, n);
659     struct vhost_dev *hdev = iommu->hdev;
660     hwaddr iova = iotlb->iova + iommu->iommu_offset;
661 
662     if (vhost_backend_invalidate_device_iotlb(hdev, iova,
663                                               iotlb->addr_mask + 1)) {
664         error_report("Fail to invalidate device iotlb");
665     }
666 }
667 
668 static void vhost_iommu_region_add(MemoryListener *listener,
669                                    MemoryRegionSection *section)
670 {
671     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
672                                          iommu_listener);
673     struct vhost_iommu *iommu;
674     Int128 end;
675     int iommu_idx;
676     IOMMUMemoryRegion *iommu_mr;
677 
678     if (!memory_region_is_iommu(section->mr)) {
679         return;
680     }
681 
682     iommu_mr = IOMMU_MEMORY_REGION(section->mr);
683 
684     iommu = g_malloc0(sizeof(*iommu));
685     end = int128_add(int128_make64(section->offset_within_region),
686                      section->size);
687     end = int128_sub(end, int128_one());
688     iommu_idx = memory_region_iommu_attrs_to_index(iommu_mr,
689                                                    MEMTXATTRS_UNSPECIFIED);
690     iommu_notifier_init(&iommu->n, vhost_iommu_unmap_notify,
691                         IOMMU_NOTIFIER_UNMAP,
692                         section->offset_within_region,
693                         int128_get64(end),
694                         iommu_idx);
695     iommu->mr = section->mr;
696     iommu->iommu_offset = section->offset_within_address_space -
697                           section->offset_within_region;
698     iommu->hdev = dev;
699     memory_region_register_iommu_notifier(section->mr, &iommu->n);
700     QLIST_INSERT_HEAD(&dev->iommu_list, iommu, iommu_next);
701     /* TODO: can replay help performance here? */
702 }
703 
704 static void vhost_iommu_region_del(MemoryListener *listener,
705                                    MemoryRegionSection *section)
706 {
707     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
708                                          iommu_listener);
709     struct vhost_iommu *iommu;
710 
711     if (!memory_region_is_iommu(section->mr)) {
712         return;
713     }
714 
715     QLIST_FOREACH(iommu, &dev->iommu_list, iommu_next) {
716         if (iommu->mr == section->mr &&
717             iommu->n.start == section->offset_within_region) {
718             memory_region_unregister_iommu_notifier(iommu->mr,
719                                                     &iommu->n);
720             QLIST_REMOVE(iommu, iommu_next);
721             g_free(iommu);
722             break;
723         }
724     }
725 }
726 
727 static int vhost_virtqueue_set_addr(struct vhost_dev *dev,
728                                     struct vhost_virtqueue *vq,
729                                     unsigned idx, bool enable_log)
730 {
731     struct vhost_vring_addr addr = {
732         .index = idx,
733         .desc_user_addr = (uint64_t)(unsigned long)vq->desc,
734         .avail_user_addr = (uint64_t)(unsigned long)vq->avail,
735         .used_user_addr = (uint64_t)(unsigned long)vq->used,
736         .log_guest_addr = vq->used_phys,
737         .flags = enable_log ? (1 << VHOST_VRING_F_LOG) : 0,
738     };
739     int r = dev->vhost_ops->vhost_set_vring_addr(dev, &addr);
740     if (r < 0) {
741         VHOST_OPS_DEBUG("vhost_set_vring_addr failed");
742         return -errno;
743     }
744     return 0;
745 }
746 
747 static int vhost_dev_set_features(struct vhost_dev *dev,
748                                   bool enable_log)
749 {
750     uint64_t features = dev->acked_features;
751     int r;
752     if (enable_log) {
753         features |= 0x1ULL << VHOST_F_LOG_ALL;
754     }
755     r = dev->vhost_ops->vhost_set_features(dev, features);
756     if (r < 0) {
757         VHOST_OPS_DEBUG("vhost_set_features failed");
758     }
759     return r < 0 ? -errno : 0;
760 }
761 
762 static int vhost_dev_set_log(struct vhost_dev *dev, bool enable_log)
763 {
764     int r, i, idx;
765     r = vhost_dev_set_features(dev, enable_log);
766     if (r < 0) {
767         goto err_features;
768     }
769     for (i = 0; i < dev->nvqs; ++i) {
770         idx = dev->vhost_ops->vhost_get_vq_index(dev, dev->vq_index + i);
771         r = vhost_virtqueue_set_addr(dev, dev->vqs + i, idx,
772                                      enable_log);
773         if (r < 0) {
774             goto err_vq;
775         }
776     }
777     return 0;
778 err_vq:
779     for (; i >= 0; --i) {
780         idx = dev->vhost_ops->vhost_get_vq_index(dev, dev->vq_index + i);
781         vhost_virtqueue_set_addr(dev, dev->vqs + i, idx,
782                                  dev->log_enabled);
783     }
784     vhost_dev_set_features(dev, dev->log_enabled);
785 err_features:
786     return r;
787 }
788 
789 static int vhost_migration_log(MemoryListener *listener, int enable)
790 {
791     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
792                                          memory_listener);
793     int r;
794     if (!!enable == dev->log_enabled) {
795         return 0;
796     }
797     if (!dev->started) {
798         dev->log_enabled = enable;
799         return 0;
800     }
801     if (!enable) {
802         r = vhost_dev_set_log(dev, false);
803         if (r < 0) {
804             return r;
805         }
806         vhost_log_put(dev, false);
807     } else {
808         vhost_dev_log_resize(dev, vhost_get_log_size(dev));
809         r = vhost_dev_set_log(dev, true);
810         if (r < 0) {
811             return r;
812         }
813     }
814     dev->log_enabled = enable;
815     return 0;
816 }
817 
818 static void vhost_log_global_start(MemoryListener *listener)
819 {
820     int r;
821 
822     r = vhost_migration_log(listener, true);
823     if (r < 0) {
824         abort();
825     }
826 }
827 
828 static void vhost_log_global_stop(MemoryListener *listener)
829 {
830     int r;
831 
832     r = vhost_migration_log(listener, false);
833     if (r < 0) {
834         abort();
835     }
836 }
837 
838 static void vhost_log_start(MemoryListener *listener,
839                             MemoryRegionSection *section,
840                             int old, int new)
841 {
842     /* FIXME: implement */
843 }
844 
845 static void vhost_log_stop(MemoryListener *listener,
846                            MemoryRegionSection *section,
847                            int old, int new)
848 {
849     /* FIXME: implement */
850 }
851 
852 /* The vhost driver natively knows how to handle the vrings of non
853  * cross-endian legacy devices and modern devices. Only legacy devices
854  * exposed to a bi-endian guest may require the vhost driver to use a
855  * specific endianness.
856  */
857 static inline bool vhost_needs_vring_endian(VirtIODevice *vdev)
858 {
859     if (virtio_vdev_has_feature(vdev, VIRTIO_F_VERSION_1)) {
860         return false;
861     }
862 #ifdef HOST_WORDS_BIGENDIAN
863     return vdev->device_endian == VIRTIO_DEVICE_ENDIAN_LITTLE;
864 #else
865     return vdev->device_endian == VIRTIO_DEVICE_ENDIAN_BIG;
866 #endif
867 }
868 
869 static int vhost_virtqueue_set_vring_endian_legacy(struct vhost_dev *dev,
870                                                    bool is_big_endian,
871                                                    int vhost_vq_index)
872 {
873     struct vhost_vring_state s = {
874         .index = vhost_vq_index,
875         .num = is_big_endian
876     };
877 
878     if (!dev->vhost_ops->vhost_set_vring_endian(dev, &s)) {
879         return 0;
880     }
881 
882     VHOST_OPS_DEBUG("vhost_set_vring_endian failed");
883     if (errno == ENOTTY) {
884         error_report("vhost does not support cross-endian");
885         return -ENOSYS;
886     }
887 
888     return -errno;
889 }
890 
891 static int vhost_memory_region_lookup(struct vhost_dev *hdev,
892                                       uint64_t gpa, uint64_t *uaddr,
893                                       uint64_t *len)
894 {
895     int i;
896 
897     for (i = 0; i < hdev->mem->nregions; i++) {
898         struct vhost_memory_region *reg = hdev->mem->regions + i;
899 
900         if (gpa >= reg->guest_phys_addr &&
901             reg->guest_phys_addr + reg->memory_size > gpa) {
902             *uaddr = reg->userspace_addr + gpa - reg->guest_phys_addr;
903             *len = reg->guest_phys_addr + reg->memory_size - gpa;
904             return 0;
905         }
906     }
907 
908     return -EFAULT;
909 }
910 
911 int vhost_device_iotlb_miss(struct vhost_dev *dev, uint64_t iova, int write)
912 {
913     IOMMUTLBEntry iotlb;
914     uint64_t uaddr, len;
915     int ret = -EFAULT;
916 
917     rcu_read_lock();
918 
919     trace_vhost_iotlb_miss(dev, 1);
920 
921     iotlb = address_space_get_iotlb_entry(dev->vdev->dma_as,
922                                           iova, write,
923                                           MEMTXATTRS_UNSPECIFIED);
924     if (iotlb.target_as != NULL) {
925         ret = vhost_memory_region_lookup(dev, iotlb.translated_addr,
926                                          &uaddr, &len);
927         if (ret) {
928             trace_vhost_iotlb_miss(dev, 3);
929             error_report("Fail to lookup the translated address "
930                          "%"PRIx64, iotlb.translated_addr);
931             goto out;
932         }
933 
934         len = MIN(iotlb.addr_mask + 1, len);
935         iova = iova & ~iotlb.addr_mask;
936 
937         ret = vhost_backend_update_device_iotlb(dev, iova, uaddr,
938                                                 len, iotlb.perm);
939         if (ret) {
940             trace_vhost_iotlb_miss(dev, 4);
941             error_report("Fail to update device iotlb");
942             goto out;
943         }
944     }
945 
946     trace_vhost_iotlb_miss(dev, 2);
947 
948 out:
949     rcu_read_unlock();
950 
951     return ret;
952 }
953 
954 static int vhost_virtqueue_start(struct vhost_dev *dev,
955                                 struct VirtIODevice *vdev,
956                                 struct vhost_virtqueue *vq,
957                                 unsigned idx)
958 {
959     BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev)));
960     VirtioBusState *vbus = VIRTIO_BUS(qbus);
961     VirtioBusClass *k = VIRTIO_BUS_GET_CLASS(vbus);
962     hwaddr s, l, a;
963     int r;
964     int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, idx);
965     struct vhost_vring_file file = {
966         .index = vhost_vq_index
967     };
968     struct vhost_vring_state state = {
969         .index = vhost_vq_index
970     };
971     struct VirtQueue *vvq = virtio_get_queue(vdev, idx);
972 
973     a = virtio_queue_get_desc_addr(vdev, idx);
974     if (a == 0) {
975         /* Queue might not be ready for start */
976         return 0;
977     }
978 
979     vq->num = state.num = virtio_queue_get_num(vdev, idx);
980     r = dev->vhost_ops->vhost_set_vring_num(dev, &state);
981     if (r) {
982         VHOST_OPS_DEBUG("vhost_set_vring_num failed");
983         return -errno;
984     }
985 
986     state.num = virtio_queue_get_last_avail_idx(vdev, idx);
987     r = dev->vhost_ops->vhost_set_vring_base(dev, &state);
988     if (r) {
989         VHOST_OPS_DEBUG("vhost_set_vring_base failed");
990         return -errno;
991     }
992 
993     if (vhost_needs_vring_endian(vdev)) {
994         r = vhost_virtqueue_set_vring_endian_legacy(dev,
995                                                     virtio_is_big_endian(vdev),
996                                                     vhost_vq_index);
997         if (r) {
998             return -errno;
999         }
1000     }
1001 
1002     vq->desc_size = s = l = virtio_queue_get_desc_size(vdev, idx);
1003     vq->desc_phys = a;
1004     vq->desc = vhost_memory_map(dev, a, &l, 0);
1005     if (!vq->desc || l != s) {
1006         r = -ENOMEM;
1007         goto fail_alloc_desc;
1008     }
1009     vq->avail_size = s = l = virtio_queue_get_avail_size(vdev, idx);
1010     vq->avail_phys = a = virtio_queue_get_avail_addr(vdev, idx);
1011     vq->avail = vhost_memory_map(dev, a, &l, 0);
1012     if (!vq->avail || l != s) {
1013         r = -ENOMEM;
1014         goto fail_alloc_avail;
1015     }
1016     vq->used_size = s = l = virtio_queue_get_used_size(vdev, idx);
1017     vq->used_phys = a = virtio_queue_get_used_addr(vdev, idx);
1018     vq->used = vhost_memory_map(dev, a, &l, 1);
1019     if (!vq->used || l != s) {
1020         r = -ENOMEM;
1021         goto fail_alloc_used;
1022     }
1023 
1024     r = vhost_virtqueue_set_addr(dev, vq, vhost_vq_index, dev->log_enabled);
1025     if (r < 0) {
1026         r = -errno;
1027         goto fail_alloc;
1028     }
1029 
1030     file.fd = event_notifier_get_fd(virtio_queue_get_host_notifier(vvq));
1031     r = dev->vhost_ops->vhost_set_vring_kick(dev, &file);
1032     if (r) {
1033         VHOST_OPS_DEBUG("vhost_set_vring_kick failed");
1034         r = -errno;
1035         goto fail_kick;
1036     }
1037 
1038     /* Clear and discard previous events if any. */
1039     event_notifier_test_and_clear(&vq->masked_notifier);
1040 
1041     /* Init vring in unmasked state, unless guest_notifier_mask
1042      * will do it later.
1043      */
1044     if (!vdev->use_guest_notifier_mask) {
1045         /* TODO: check and handle errors. */
1046         vhost_virtqueue_mask(dev, vdev, idx, false);
1047     }
1048 
1049     if (k->query_guest_notifiers &&
1050         k->query_guest_notifiers(qbus->parent) &&
1051         virtio_queue_vector(vdev, idx) == VIRTIO_NO_VECTOR) {
1052         file.fd = -1;
1053         r = dev->vhost_ops->vhost_set_vring_call(dev, &file);
1054         if (r) {
1055             goto fail_vector;
1056         }
1057     }
1058 
1059     return 0;
1060 
1061 fail_vector:
1062 fail_kick:
1063 fail_alloc:
1064     vhost_memory_unmap(dev, vq->used, virtio_queue_get_used_size(vdev, idx),
1065                        0, 0);
1066 fail_alloc_used:
1067     vhost_memory_unmap(dev, vq->avail, virtio_queue_get_avail_size(vdev, idx),
1068                        0, 0);
1069 fail_alloc_avail:
1070     vhost_memory_unmap(dev, vq->desc, virtio_queue_get_desc_size(vdev, idx),
1071                        0, 0);
1072 fail_alloc_desc:
1073     return r;
1074 }
1075 
1076 static void vhost_virtqueue_stop(struct vhost_dev *dev,
1077                                     struct VirtIODevice *vdev,
1078                                     struct vhost_virtqueue *vq,
1079                                     unsigned idx)
1080 {
1081     int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, idx);
1082     struct vhost_vring_state state = {
1083         .index = vhost_vq_index,
1084     };
1085     int r;
1086 
1087     if (virtio_queue_get_desc_addr(vdev, idx) == 0) {
1088         /* Don't stop the virtqueue which might have not been started */
1089         return;
1090     }
1091 
1092     r = dev->vhost_ops->vhost_get_vring_base(dev, &state);
1093     if (r < 0) {
1094         VHOST_OPS_DEBUG("vhost VQ %u ring restore failed: %d", idx, r);
1095         /* Connection to the backend is broken, so let's sync internal
1096          * last avail idx to the device used idx.
1097          */
1098         virtio_queue_restore_last_avail_idx(vdev, idx);
1099     } else {
1100         virtio_queue_set_last_avail_idx(vdev, idx, state.num);
1101     }
1102     virtio_queue_invalidate_signalled_used(vdev, idx);
1103     virtio_queue_update_used_idx(vdev, idx);
1104 
1105     /* In the cross-endian case, we need to reset the vring endianness to
1106      * native as legacy devices expect so by default.
1107      */
1108     if (vhost_needs_vring_endian(vdev)) {
1109         vhost_virtqueue_set_vring_endian_legacy(dev,
1110                                                 !virtio_is_big_endian(vdev),
1111                                                 vhost_vq_index);
1112     }
1113 
1114     vhost_memory_unmap(dev, vq->used, virtio_queue_get_used_size(vdev, idx),
1115                        1, virtio_queue_get_used_size(vdev, idx));
1116     vhost_memory_unmap(dev, vq->avail, virtio_queue_get_avail_size(vdev, idx),
1117                        0, virtio_queue_get_avail_size(vdev, idx));
1118     vhost_memory_unmap(dev, vq->desc, virtio_queue_get_desc_size(vdev, idx),
1119                        0, virtio_queue_get_desc_size(vdev, idx));
1120 }
1121 
1122 static void vhost_eventfd_add(MemoryListener *listener,
1123                               MemoryRegionSection *section,
1124                               bool match_data, uint64_t data, EventNotifier *e)
1125 {
1126 }
1127 
1128 static void vhost_eventfd_del(MemoryListener *listener,
1129                               MemoryRegionSection *section,
1130                               bool match_data, uint64_t data, EventNotifier *e)
1131 {
1132 }
1133 
1134 static int vhost_virtqueue_set_busyloop_timeout(struct vhost_dev *dev,
1135                                                 int n, uint32_t timeout)
1136 {
1137     int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, n);
1138     struct vhost_vring_state state = {
1139         .index = vhost_vq_index,
1140         .num = timeout,
1141     };
1142     int r;
1143 
1144     if (!dev->vhost_ops->vhost_set_vring_busyloop_timeout) {
1145         return -EINVAL;
1146     }
1147 
1148     r = dev->vhost_ops->vhost_set_vring_busyloop_timeout(dev, &state);
1149     if (r) {
1150         VHOST_OPS_DEBUG("vhost_set_vring_busyloop_timeout failed");
1151         return r;
1152     }
1153 
1154     return 0;
1155 }
1156 
1157 static int vhost_virtqueue_init(struct vhost_dev *dev,
1158                                 struct vhost_virtqueue *vq, int n)
1159 {
1160     int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, n);
1161     struct vhost_vring_file file = {
1162         .index = vhost_vq_index,
1163     };
1164     int r = event_notifier_init(&vq->masked_notifier, 0);
1165     if (r < 0) {
1166         return r;
1167     }
1168 
1169     file.fd = event_notifier_get_fd(&vq->masked_notifier);
1170     r = dev->vhost_ops->vhost_set_vring_call(dev, &file);
1171     if (r) {
1172         VHOST_OPS_DEBUG("vhost_set_vring_call failed");
1173         r = -errno;
1174         goto fail_call;
1175     }
1176 
1177     vq->dev = dev;
1178 
1179     return 0;
1180 fail_call:
1181     event_notifier_cleanup(&vq->masked_notifier);
1182     return r;
1183 }
1184 
1185 static void vhost_virtqueue_cleanup(struct vhost_virtqueue *vq)
1186 {
1187     event_notifier_cleanup(&vq->masked_notifier);
1188 }
1189 
1190 int vhost_dev_init(struct vhost_dev *hdev, void *opaque,
1191                    VhostBackendType backend_type, uint32_t busyloop_timeout)
1192 {
1193     uint64_t features;
1194     int i, r, n_initialized_vqs = 0;
1195     Error *local_err = NULL;
1196 
1197     hdev->vdev = NULL;
1198     hdev->migration_blocker = NULL;
1199 
1200     r = vhost_set_backend_type(hdev, backend_type);
1201     assert(r >= 0);
1202 
1203     r = hdev->vhost_ops->vhost_backend_init(hdev, opaque);
1204     if (r < 0) {
1205         goto fail;
1206     }
1207 
1208     r = hdev->vhost_ops->vhost_set_owner(hdev);
1209     if (r < 0) {
1210         VHOST_OPS_DEBUG("vhost_set_owner failed");
1211         goto fail;
1212     }
1213 
1214     r = hdev->vhost_ops->vhost_get_features(hdev, &features);
1215     if (r < 0) {
1216         VHOST_OPS_DEBUG("vhost_get_features failed");
1217         goto fail;
1218     }
1219 
1220     for (i = 0; i < hdev->nvqs; ++i, ++n_initialized_vqs) {
1221         r = vhost_virtqueue_init(hdev, hdev->vqs + i, hdev->vq_index + i);
1222         if (r < 0) {
1223             goto fail;
1224         }
1225     }
1226 
1227     if (busyloop_timeout) {
1228         for (i = 0; i < hdev->nvqs; ++i) {
1229             r = vhost_virtqueue_set_busyloop_timeout(hdev, hdev->vq_index + i,
1230                                                      busyloop_timeout);
1231             if (r < 0) {
1232                 goto fail_busyloop;
1233             }
1234         }
1235     }
1236 
1237     hdev->features = features;
1238 
1239     hdev->memory_listener = (MemoryListener) {
1240         .begin = vhost_begin,
1241         .commit = vhost_commit,
1242         .region_add = vhost_region_addnop,
1243         .region_nop = vhost_region_addnop,
1244         .log_start = vhost_log_start,
1245         .log_stop = vhost_log_stop,
1246         .log_sync = vhost_log_sync,
1247         .log_global_start = vhost_log_global_start,
1248         .log_global_stop = vhost_log_global_stop,
1249         .eventfd_add = vhost_eventfd_add,
1250         .eventfd_del = vhost_eventfd_del,
1251         .priority = 10
1252     };
1253 
1254     hdev->iommu_listener = (MemoryListener) {
1255         .region_add = vhost_iommu_region_add,
1256         .region_del = vhost_iommu_region_del,
1257     };
1258 
1259     if (hdev->migration_blocker == NULL) {
1260         if (!(hdev->features & (0x1ULL << VHOST_F_LOG_ALL))) {
1261             error_setg(&hdev->migration_blocker,
1262                        "Migration disabled: vhost lacks VHOST_F_LOG_ALL feature.");
1263         } else if (vhost_dev_log_is_shared(hdev) && !qemu_memfd_alloc_check()) {
1264             error_setg(&hdev->migration_blocker,
1265                        "Migration disabled: failed to allocate shared memory");
1266         }
1267     }
1268 
1269     if (hdev->migration_blocker != NULL) {
1270         r = migrate_add_blocker(hdev->migration_blocker, &local_err);
1271         if (local_err) {
1272             error_report_err(local_err);
1273             error_free(hdev->migration_blocker);
1274             goto fail_busyloop;
1275         }
1276     }
1277 
1278     hdev->mem = g_malloc0(offsetof(struct vhost_memory, regions));
1279     hdev->n_mem_sections = 0;
1280     hdev->mem_sections = NULL;
1281     hdev->log = NULL;
1282     hdev->log_size = 0;
1283     hdev->log_enabled = false;
1284     hdev->started = false;
1285     memory_listener_register(&hdev->memory_listener, &address_space_memory);
1286     QLIST_INSERT_HEAD(&vhost_devices, hdev, entry);
1287 
1288     if (used_memslots > hdev->vhost_ops->vhost_backend_memslots_limit(hdev)) {
1289         error_report("vhost backend memory slots limit is less"
1290                 " than current number of present memory slots");
1291         r = -1;
1292         if (busyloop_timeout) {
1293             goto fail_busyloop;
1294         } else {
1295             goto fail;
1296         }
1297     }
1298 
1299     return 0;
1300 
1301 fail_busyloop:
1302     while (--i >= 0) {
1303         vhost_virtqueue_set_busyloop_timeout(hdev, hdev->vq_index + i, 0);
1304     }
1305 fail:
1306     hdev->nvqs = n_initialized_vqs;
1307     vhost_dev_cleanup(hdev);
1308     return r;
1309 }
1310 
1311 void vhost_dev_cleanup(struct vhost_dev *hdev)
1312 {
1313     int i;
1314 
1315     for (i = 0; i < hdev->nvqs; ++i) {
1316         vhost_virtqueue_cleanup(hdev->vqs + i);
1317     }
1318     if (hdev->mem) {
1319         /* those are only safe after successful init */
1320         memory_listener_unregister(&hdev->memory_listener);
1321         QLIST_REMOVE(hdev, entry);
1322     }
1323     if (hdev->migration_blocker) {
1324         migrate_del_blocker(hdev->migration_blocker);
1325         error_free(hdev->migration_blocker);
1326     }
1327     g_free(hdev->mem);
1328     g_free(hdev->mem_sections);
1329     if (hdev->vhost_ops) {
1330         hdev->vhost_ops->vhost_backend_cleanup(hdev);
1331     }
1332     assert(!hdev->log);
1333 
1334     memset(hdev, 0, sizeof(struct vhost_dev));
1335 }
1336 
1337 /* Stop processing guest IO notifications in qemu.
1338  * Start processing them in vhost in kernel.
1339  */
1340 int vhost_dev_enable_notifiers(struct vhost_dev *hdev, VirtIODevice *vdev)
1341 {
1342     BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev)));
1343     int i, r, e;
1344 
1345     /* We will pass the notifiers to the kernel, make sure that QEMU
1346      * doesn't interfere.
1347      */
1348     r = virtio_device_grab_ioeventfd(vdev);
1349     if (r < 0) {
1350         error_report("binding does not support host notifiers");
1351         goto fail;
1352     }
1353 
1354     for (i = 0; i < hdev->nvqs; ++i) {
1355         r = virtio_bus_set_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i,
1356                                          true);
1357         if (r < 0) {
1358             error_report("vhost VQ %d notifier binding failed: %d", i, -r);
1359             goto fail_vq;
1360         }
1361     }
1362 
1363     return 0;
1364 fail_vq:
1365     while (--i >= 0) {
1366         e = virtio_bus_set_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i,
1367                                          false);
1368         if (e < 0) {
1369             error_report("vhost VQ %d notifier cleanup error: %d", i, -r);
1370         }
1371         assert (e >= 0);
1372         virtio_bus_cleanup_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i);
1373     }
1374     virtio_device_release_ioeventfd(vdev);
1375 fail:
1376     return r;
1377 }
1378 
1379 /* Stop processing guest IO notifications in vhost.
1380  * Start processing them in qemu.
1381  * This might actually run the qemu handlers right away,
1382  * so virtio in qemu must be completely setup when this is called.
1383  */
1384 void vhost_dev_disable_notifiers(struct vhost_dev *hdev, VirtIODevice *vdev)
1385 {
1386     BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev)));
1387     int i, r;
1388 
1389     for (i = 0; i < hdev->nvqs; ++i) {
1390         r = virtio_bus_set_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i,
1391                                          false);
1392         if (r < 0) {
1393             error_report("vhost VQ %d notifier cleanup failed: %d", i, -r);
1394         }
1395         assert (r >= 0);
1396         virtio_bus_cleanup_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i);
1397     }
1398     virtio_device_release_ioeventfd(vdev);
1399 }
1400 
1401 /* Test and clear event pending status.
1402  * Should be called after unmask to avoid losing events.
1403  */
1404 bool vhost_virtqueue_pending(struct vhost_dev *hdev, int n)
1405 {
1406     struct vhost_virtqueue *vq = hdev->vqs + n - hdev->vq_index;
1407     assert(n >= hdev->vq_index && n < hdev->vq_index + hdev->nvqs);
1408     return event_notifier_test_and_clear(&vq->masked_notifier);
1409 }
1410 
1411 /* Mask/unmask events from this vq. */
1412 void vhost_virtqueue_mask(struct vhost_dev *hdev, VirtIODevice *vdev, int n,
1413                          bool mask)
1414 {
1415     struct VirtQueue *vvq = virtio_get_queue(vdev, n);
1416     int r, index = n - hdev->vq_index;
1417     struct vhost_vring_file file;
1418 
1419     /* should only be called after backend is connected */
1420     assert(hdev->vhost_ops);
1421 
1422     if (mask) {
1423         assert(vdev->use_guest_notifier_mask);
1424         file.fd = event_notifier_get_fd(&hdev->vqs[index].masked_notifier);
1425     } else {
1426         file.fd = event_notifier_get_fd(virtio_queue_get_guest_notifier(vvq));
1427     }
1428 
1429     file.index = hdev->vhost_ops->vhost_get_vq_index(hdev, n);
1430     r = hdev->vhost_ops->vhost_set_vring_call(hdev, &file);
1431     if (r < 0) {
1432         VHOST_OPS_DEBUG("vhost_set_vring_call failed");
1433     }
1434 }
1435 
1436 uint64_t vhost_get_features(struct vhost_dev *hdev, const int *feature_bits,
1437                             uint64_t features)
1438 {
1439     const int *bit = feature_bits;
1440     while (*bit != VHOST_INVALID_FEATURE_BIT) {
1441         uint64_t bit_mask = (1ULL << *bit);
1442         if (!(hdev->features & bit_mask)) {
1443             features &= ~bit_mask;
1444         }
1445         bit++;
1446     }
1447     return features;
1448 }
1449 
1450 void vhost_ack_features(struct vhost_dev *hdev, const int *feature_bits,
1451                         uint64_t features)
1452 {
1453     const int *bit = feature_bits;
1454     while (*bit != VHOST_INVALID_FEATURE_BIT) {
1455         uint64_t bit_mask = (1ULL << *bit);
1456         if (features & bit_mask) {
1457             hdev->acked_features |= bit_mask;
1458         }
1459         bit++;
1460     }
1461 }
1462 
1463 int vhost_dev_get_config(struct vhost_dev *hdev, uint8_t *config,
1464                          uint32_t config_len)
1465 {
1466     assert(hdev->vhost_ops);
1467 
1468     if (hdev->vhost_ops->vhost_get_config) {
1469         return hdev->vhost_ops->vhost_get_config(hdev, config, config_len);
1470     }
1471 
1472     return -1;
1473 }
1474 
1475 int vhost_dev_set_config(struct vhost_dev *hdev, const uint8_t *data,
1476                          uint32_t offset, uint32_t size, uint32_t flags)
1477 {
1478     assert(hdev->vhost_ops);
1479 
1480     if (hdev->vhost_ops->vhost_set_config) {
1481         return hdev->vhost_ops->vhost_set_config(hdev, data, offset,
1482                                                  size, flags);
1483     }
1484 
1485     return -1;
1486 }
1487 
1488 void vhost_dev_set_config_notifier(struct vhost_dev *hdev,
1489                                    const VhostDevConfigOps *ops)
1490 {
1491     hdev->config_ops = ops;
1492 }
1493 
1494 void vhost_dev_free_inflight(struct vhost_inflight *inflight)
1495 {
1496     if (inflight->addr) {
1497         qemu_memfd_free(inflight->addr, inflight->size, inflight->fd);
1498         inflight->addr = NULL;
1499         inflight->fd = -1;
1500     }
1501 }
1502 
1503 static int vhost_dev_resize_inflight(struct vhost_inflight *inflight,
1504                                      uint64_t new_size)
1505 {
1506     Error *err = NULL;
1507     int fd = -1;
1508     void *addr = qemu_memfd_alloc("vhost-inflight", new_size,
1509                                   F_SEAL_GROW | F_SEAL_SHRINK | F_SEAL_SEAL,
1510                                   &fd, &err);
1511 
1512     if (err) {
1513         error_report_err(err);
1514         return -1;
1515     }
1516 
1517     vhost_dev_free_inflight(inflight);
1518     inflight->offset = 0;
1519     inflight->addr = addr;
1520     inflight->fd = fd;
1521     inflight->size = new_size;
1522 
1523     return 0;
1524 }
1525 
1526 void vhost_dev_save_inflight(struct vhost_inflight *inflight, QEMUFile *f)
1527 {
1528     if (inflight->addr) {
1529         qemu_put_be64(f, inflight->size);
1530         qemu_put_be16(f, inflight->queue_size);
1531         qemu_put_buffer(f, inflight->addr, inflight->size);
1532     } else {
1533         qemu_put_be64(f, 0);
1534     }
1535 }
1536 
1537 int vhost_dev_load_inflight(struct vhost_inflight *inflight, QEMUFile *f)
1538 {
1539     uint64_t size;
1540 
1541     size = qemu_get_be64(f);
1542     if (!size) {
1543         return 0;
1544     }
1545 
1546     if (inflight->size != size) {
1547         if (vhost_dev_resize_inflight(inflight, size)) {
1548             return -1;
1549         }
1550     }
1551     inflight->queue_size = qemu_get_be16(f);
1552 
1553     qemu_get_buffer(f, inflight->addr, size);
1554 
1555     return 0;
1556 }
1557 
1558 int vhost_dev_set_inflight(struct vhost_dev *dev,
1559                            struct vhost_inflight *inflight)
1560 {
1561     int r;
1562 
1563     if (dev->vhost_ops->vhost_set_inflight_fd && inflight->addr) {
1564         r = dev->vhost_ops->vhost_set_inflight_fd(dev, inflight);
1565         if (r) {
1566             VHOST_OPS_DEBUG("vhost_set_inflight_fd failed");
1567             return -errno;
1568         }
1569     }
1570 
1571     return 0;
1572 }
1573 
1574 int vhost_dev_get_inflight(struct vhost_dev *dev, uint16_t queue_size,
1575                            struct vhost_inflight *inflight)
1576 {
1577     int r;
1578 
1579     if (dev->vhost_ops->vhost_get_inflight_fd) {
1580         r = dev->vhost_ops->vhost_get_inflight_fd(dev, queue_size, inflight);
1581         if (r) {
1582             VHOST_OPS_DEBUG("vhost_get_inflight_fd failed");
1583             return -errno;
1584         }
1585     }
1586 
1587     return 0;
1588 }
1589 
1590 /* Host notifiers must be enabled at this point. */
1591 int vhost_dev_start(struct vhost_dev *hdev, VirtIODevice *vdev)
1592 {
1593     int i, r;
1594 
1595     /* should only be called after backend is connected */
1596     assert(hdev->vhost_ops);
1597 
1598     hdev->started = true;
1599     hdev->vdev = vdev;
1600 
1601     r = vhost_dev_set_features(hdev, hdev->log_enabled);
1602     if (r < 0) {
1603         goto fail_features;
1604     }
1605 
1606     if (vhost_dev_has_iommu(hdev)) {
1607         memory_listener_register(&hdev->iommu_listener, vdev->dma_as);
1608     }
1609 
1610     r = hdev->vhost_ops->vhost_set_mem_table(hdev, hdev->mem);
1611     if (r < 0) {
1612         VHOST_OPS_DEBUG("vhost_set_mem_table failed");
1613         r = -errno;
1614         goto fail_mem;
1615     }
1616     for (i = 0; i < hdev->nvqs; ++i) {
1617         r = vhost_virtqueue_start(hdev,
1618                                   vdev,
1619                                   hdev->vqs + i,
1620                                   hdev->vq_index + i);
1621         if (r < 0) {
1622             goto fail_vq;
1623         }
1624     }
1625 
1626     if (hdev->log_enabled) {
1627         uint64_t log_base;
1628 
1629         hdev->log_size = vhost_get_log_size(hdev);
1630         hdev->log = vhost_log_get(hdev->log_size,
1631                                   vhost_dev_log_is_shared(hdev));
1632         log_base = (uintptr_t)hdev->log->log;
1633         r = hdev->vhost_ops->vhost_set_log_base(hdev,
1634                                                 hdev->log_size ? log_base : 0,
1635                                                 hdev->log);
1636         if (r < 0) {
1637             VHOST_OPS_DEBUG("vhost_set_log_base failed");
1638             r = -errno;
1639             goto fail_log;
1640         }
1641     }
1642 
1643     if (vhost_dev_has_iommu(hdev)) {
1644         hdev->vhost_ops->vhost_set_iotlb_callback(hdev, true);
1645 
1646         /* Update used ring information for IOTLB to work correctly,
1647          * vhost-kernel code requires for this.*/
1648         for (i = 0; i < hdev->nvqs; ++i) {
1649             struct vhost_virtqueue *vq = hdev->vqs + i;
1650             vhost_device_iotlb_miss(hdev, vq->used_phys, true);
1651         }
1652     }
1653     return 0;
1654 fail_log:
1655     vhost_log_put(hdev, false);
1656 fail_vq:
1657     while (--i >= 0) {
1658         vhost_virtqueue_stop(hdev,
1659                              vdev,
1660                              hdev->vqs + i,
1661                              hdev->vq_index + i);
1662     }
1663 
1664 fail_mem:
1665 fail_features:
1666 
1667     hdev->started = false;
1668     return r;
1669 }
1670 
1671 /* Host notifiers must be enabled at this point. */
1672 void vhost_dev_stop(struct vhost_dev *hdev, VirtIODevice *vdev)
1673 {
1674     int i;
1675 
1676     /* should only be called after backend is connected */
1677     assert(hdev->vhost_ops);
1678 
1679     for (i = 0; i < hdev->nvqs; ++i) {
1680         vhost_virtqueue_stop(hdev,
1681                              vdev,
1682                              hdev->vqs + i,
1683                              hdev->vq_index + i);
1684     }
1685 
1686     if (vhost_dev_has_iommu(hdev)) {
1687         hdev->vhost_ops->vhost_set_iotlb_callback(hdev, false);
1688         memory_listener_unregister(&hdev->iommu_listener);
1689     }
1690     vhost_log_put(hdev, true);
1691     hdev->started = false;
1692     hdev->vdev = NULL;
1693 }
1694 
1695 int vhost_net_set_backend(struct vhost_dev *hdev,
1696                           struct vhost_vring_file *file)
1697 {
1698     if (hdev->vhost_ops->vhost_net_set_backend) {
1699         return hdev->vhost_ops->vhost_net_set_backend(hdev, file);
1700     }
1701 
1702     return -1;
1703 }
1704