xref: /openbmc/qemu/hw/virtio/vhost.c (revision 89854803)
1 /*
2  * vhost support
3  *
4  * Copyright Red Hat, Inc. 2010
5  *
6  * Authors:
7  *  Michael S. Tsirkin <mst@redhat.com>
8  *
9  * This work is licensed under the terms of the GNU GPL, version 2.  See
10  * the COPYING file in the top-level directory.
11  *
12  * Contributions after 2012-01-13 are licensed under the terms of the
13  * GNU GPL, version 2 or (at your option) any later version.
14  */
15 
16 #include "qemu/osdep.h"
17 #include "qapi/error.h"
18 #include "hw/virtio/vhost.h"
19 #include "hw/hw.h"
20 #include "qemu/atomic.h"
21 #include "qemu/range.h"
22 #include "qemu/error-report.h"
23 #include "qemu/memfd.h"
24 #include <linux/vhost.h>
25 #include "exec/address-spaces.h"
26 #include "hw/virtio/virtio-bus.h"
27 #include "hw/virtio/virtio-access.h"
28 #include "migration/blocker.h"
29 #include "sysemu/dma.h"
30 #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         vhost_dev_sync_region(dev, section, start_addr, end_addr, vq->used_phys,
135                               range_get_last(vq->used_phys, vq->used_size));
136     }
137     return 0;
138 }
139 
140 static void vhost_log_sync(MemoryListener *listener,
141                           MemoryRegionSection *section)
142 {
143     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
144                                          memory_listener);
145     vhost_sync_dirty_bitmap(dev, section, 0x0, ~0x0ULL);
146 }
147 
148 static void vhost_log_sync_range(struct vhost_dev *dev,
149                                  hwaddr first, hwaddr last)
150 {
151     int i;
152     /* FIXME: this is N^2 in number of sections */
153     for (i = 0; i < dev->n_mem_sections; ++i) {
154         MemoryRegionSection *section = &dev->mem_sections[i];
155         vhost_sync_dirty_bitmap(dev, section, first, last);
156     }
157 }
158 
159 static uint64_t vhost_get_log_size(struct vhost_dev *dev)
160 {
161     uint64_t log_size = 0;
162     int i;
163     for (i = 0; i < dev->mem->nregions; ++i) {
164         struct vhost_memory_region *reg = dev->mem->regions + i;
165         uint64_t last = range_get_last(reg->guest_phys_addr,
166                                        reg->memory_size);
167         log_size = MAX(log_size, last / VHOST_LOG_CHUNK + 1);
168     }
169     for (i = 0; i < dev->nvqs; ++i) {
170         struct vhost_virtqueue *vq = dev->vqs + i;
171         uint64_t last = vq->used_phys + vq->used_size - 1;
172         log_size = MAX(log_size, last / VHOST_LOG_CHUNK + 1);
173     }
174     return log_size;
175 }
176 
177 static struct vhost_log *vhost_log_alloc(uint64_t size, bool share)
178 {
179     Error *err = NULL;
180     struct vhost_log *log;
181     uint64_t logsize = size * sizeof(*(log->log));
182     int fd = -1;
183 
184     log = g_new0(struct vhost_log, 1);
185     if (share) {
186         log->log = qemu_memfd_alloc("vhost-log", logsize,
187                                     F_SEAL_GROW | F_SEAL_SHRINK | F_SEAL_SEAL,
188                                     &fd, &err);
189         if (err) {
190             error_report_err(err);
191             g_free(log);
192             return NULL;
193         }
194         memset(log->log, 0, logsize);
195     } else {
196         log->log = g_malloc0(logsize);
197     }
198 
199     log->size = size;
200     log->refcnt = 1;
201     log->fd = fd;
202 
203     return log;
204 }
205 
206 static struct vhost_log *vhost_log_get(uint64_t size, bool share)
207 {
208     struct vhost_log *log = share ? vhost_log_shm : vhost_log;
209 
210     if (!log || log->size != size) {
211         log = vhost_log_alloc(size, share);
212         if (share) {
213             vhost_log_shm = log;
214         } else {
215             vhost_log = log;
216         }
217     } else {
218         ++log->refcnt;
219     }
220 
221     return log;
222 }
223 
224 static void vhost_log_put(struct vhost_dev *dev, bool sync)
225 {
226     struct vhost_log *log = dev->log;
227 
228     if (!log) {
229         return;
230     }
231 
232     --log->refcnt;
233     if (log->refcnt == 0) {
234         /* Sync only the range covered by the old log */
235         if (dev->log_size && sync) {
236             vhost_log_sync_range(dev, 0, dev->log_size * VHOST_LOG_CHUNK - 1);
237         }
238 
239         if (vhost_log == log) {
240             g_free(log->log);
241             vhost_log = NULL;
242         } else if (vhost_log_shm == log) {
243             qemu_memfd_free(log->log, log->size * sizeof(*(log->log)),
244                             log->fd);
245             vhost_log_shm = NULL;
246         }
247 
248         g_free(log);
249     }
250 
251     dev->log = NULL;
252     dev->log_size = 0;
253 }
254 
255 static bool vhost_dev_log_is_shared(struct vhost_dev *dev)
256 {
257     return dev->vhost_ops->vhost_requires_shm_log &&
258            dev->vhost_ops->vhost_requires_shm_log(dev);
259 }
260 
261 static inline void vhost_dev_log_resize(struct vhost_dev *dev, uint64_t size)
262 {
263     struct vhost_log *log = vhost_log_get(size, vhost_dev_log_is_shared(dev));
264     uint64_t log_base = (uintptr_t)log->log;
265     int r;
266 
267     /* inform backend of log switching, this must be done before
268        releasing the current log, to ensure no logging is lost */
269     r = dev->vhost_ops->vhost_set_log_base(dev, log_base, log);
270     if (r < 0) {
271         VHOST_OPS_DEBUG("vhost_set_log_base failed");
272     }
273 
274     vhost_log_put(dev, true);
275     dev->log = log;
276     dev->log_size = size;
277 }
278 
279 static int vhost_dev_has_iommu(struct vhost_dev *dev)
280 {
281     VirtIODevice *vdev = dev->vdev;
282 
283     return virtio_host_has_feature(vdev, VIRTIO_F_IOMMU_PLATFORM);
284 }
285 
286 static void *vhost_memory_map(struct vhost_dev *dev, hwaddr addr,
287                               hwaddr *plen, int is_write)
288 {
289     if (!vhost_dev_has_iommu(dev)) {
290         return cpu_physical_memory_map(addr, plen, is_write);
291     } else {
292         return (void *)(uintptr_t)addr;
293     }
294 }
295 
296 static void vhost_memory_unmap(struct vhost_dev *dev, void *buffer,
297                                hwaddr len, int is_write,
298                                hwaddr access_len)
299 {
300     if (!vhost_dev_has_iommu(dev)) {
301         cpu_physical_memory_unmap(buffer, len, is_write, access_len);
302     }
303 }
304 
305 static int vhost_verify_ring_part_mapping(void *ring_hva,
306                                           uint64_t ring_gpa,
307                                           uint64_t ring_size,
308                                           void *reg_hva,
309                                           uint64_t reg_gpa,
310                                           uint64_t reg_size)
311 {
312     uint64_t hva_ring_offset;
313     uint64_t ring_last = range_get_last(ring_gpa, ring_size);
314     uint64_t reg_last = range_get_last(reg_gpa, reg_size);
315 
316     if (ring_last < reg_gpa || ring_gpa > reg_last) {
317         return 0;
318     }
319     /* check that whole ring's is mapped */
320     if (ring_last > reg_last) {
321         return -ENOMEM;
322     }
323     /* check that ring's MemoryRegion wasn't replaced */
324     hva_ring_offset = ring_gpa - reg_gpa;
325     if (ring_hva != reg_hva + hva_ring_offset) {
326         return -EBUSY;
327     }
328 
329     return 0;
330 }
331 
332 static int vhost_verify_ring_mappings(struct vhost_dev *dev,
333                                       void *reg_hva,
334                                       uint64_t reg_gpa,
335                                       uint64_t reg_size)
336 {
337     int i, j;
338     int r = 0;
339     const char *part_name[] = {
340         "descriptor table",
341         "available ring",
342         "used ring"
343     };
344 
345     if (vhost_dev_has_iommu(dev)) {
346         return 0;
347     }
348 
349     for (i = 0; i < dev->nvqs; ++i) {
350         struct vhost_virtqueue *vq = dev->vqs + i;
351 
352         if (vq->desc_phys == 0) {
353             continue;
354         }
355 
356         j = 0;
357         r = vhost_verify_ring_part_mapping(
358                 vq->desc, vq->desc_phys, vq->desc_size,
359                 reg_hva, reg_gpa, reg_size);
360         if (r) {
361             break;
362         }
363 
364         j++;
365         r = vhost_verify_ring_part_mapping(
366                 vq->avail, vq->avail_phys, vq->avail_size,
367                 reg_hva, reg_gpa, reg_size);
368         if (r) {
369             break;
370         }
371 
372         j++;
373         r = vhost_verify_ring_part_mapping(
374                 vq->used, vq->used_phys, vq->used_size,
375                 reg_hva, reg_gpa, reg_size);
376         if (r) {
377             break;
378         }
379     }
380 
381     if (r == -ENOMEM) {
382         error_report("Unable to map %s for ring %d", part_name[j], i);
383     } else if (r == -EBUSY) {
384         error_report("%s relocated for ring %d", part_name[j], i);
385     }
386     return r;
387 }
388 
389 static bool vhost_section(MemoryRegionSection *section)
390 {
391     bool result;
392     bool log_dirty = memory_region_get_dirty_log_mask(section->mr) &
393                      ~(1 << DIRTY_MEMORY_MIGRATION);
394     result = memory_region_is_ram(section->mr) &&
395         !memory_region_is_rom(section->mr);
396 
397     /* Vhost doesn't handle any block which is doing dirty-tracking other
398      * than migration; this typically fires on VGA areas.
399      */
400     result &= !log_dirty;
401 
402     trace_vhost_section(section->mr->name, result);
403     return result;
404 }
405 
406 static void vhost_begin(MemoryListener *listener)
407 {
408     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
409                                          memory_listener);
410     dev->tmp_sections = NULL;
411     dev->n_tmp_sections = 0;
412 }
413 
414 static void vhost_commit(MemoryListener *listener)
415 {
416     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
417                                          memory_listener);
418     MemoryRegionSection *old_sections;
419     int n_old_sections;
420     uint64_t log_size;
421     size_t regions_size;
422     int r;
423     int i;
424     bool changed = false;
425 
426     /* Note we can be called before the device is started, but then
427      * starting the device calls set_mem_table, so we need to have
428      * built the data structures.
429      */
430     old_sections = dev->mem_sections;
431     n_old_sections = dev->n_mem_sections;
432     dev->mem_sections = dev->tmp_sections;
433     dev->n_mem_sections = dev->n_tmp_sections;
434 
435     if (dev->n_mem_sections != n_old_sections) {
436         changed = true;
437     } else {
438         /* Same size, lets check the contents */
439         changed = n_old_sections && memcmp(dev->mem_sections, old_sections,
440                          n_old_sections * sizeof(old_sections[0])) != 0;
441     }
442 
443     trace_vhost_commit(dev->started, changed);
444     if (!changed) {
445         goto out;
446     }
447 
448     /* Rebuild the regions list from the new sections list */
449     regions_size = offsetof(struct vhost_memory, regions) +
450                        dev->n_mem_sections * sizeof dev->mem->regions[0];
451     dev->mem = g_realloc(dev->mem, regions_size);
452     dev->mem->nregions = dev->n_mem_sections;
453     used_memslots = dev->mem->nregions;
454     for (i = 0; i < dev->n_mem_sections; i++) {
455         struct vhost_memory_region *cur_vmr = dev->mem->regions + i;
456         struct MemoryRegionSection *mrs = dev->mem_sections + i;
457 
458         cur_vmr->guest_phys_addr = mrs->offset_within_address_space;
459         cur_vmr->memory_size     = int128_get64(mrs->size);
460         cur_vmr->userspace_addr  =
461             (uintptr_t)memory_region_get_ram_ptr(mrs->mr) +
462             mrs->offset_within_region;
463         cur_vmr->flags_padding   = 0;
464     }
465 
466     if (!dev->started) {
467         goto out;
468     }
469 
470     for (i = 0; i < dev->mem->nregions; i++) {
471         if (vhost_verify_ring_mappings(dev,
472                        (void *)(uintptr_t)dev->mem->regions[i].userspace_addr,
473                        dev->mem->regions[i].guest_phys_addr,
474                        dev->mem->regions[i].memory_size)) {
475             error_report("Verify ring failure on region %d", i);
476             abort();
477         }
478     }
479 
480     if (!dev->log_enabled) {
481         r = dev->vhost_ops->vhost_set_mem_table(dev, dev->mem);
482         if (r < 0) {
483             VHOST_OPS_DEBUG("vhost_set_mem_table failed");
484         }
485         goto out;
486     }
487     log_size = vhost_get_log_size(dev);
488     /* We allocate an extra 4K bytes to log,
489      * to reduce the * number of reallocations. */
490 #define VHOST_LOG_BUFFER (0x1000 / sizeof *dev->log)
491     /* To log more, must increase log size before table update. */
492     if (dev->log_size < log_size) {
493         vhost_dev_log_resize(dev, log_size + VHOST_LOG_BUFFER);
494     }
495     r = dev->vhost_ops->vhost_set_mem_table(dev, dev->mem);
496     if (r < 0) {
497         VHOST_OPS_DEBUG("vhost_set_mem_table failed");
498     }
499     /* To log less, can only decrease log size after table update. */
500     if (dev->log_size > log_size + VHOST_LOG_BUFFER) {
501         vhost_dev_log_resize(dev, log_size);
502     }
503 
504 out:
505     /* Deref the old list of sections, this must happen _after_ the
506      * vhost_set_mem_table to ensure the client isn't still using the
507      * section we're about to unref.
508      */
509     while (n_old_sections--) {
510         memory_region_unref(old_sections[n_old_sections].mr);
511     }
512     g_free(old_sections);
513     return;
514 }
515 
516 /* Adds the section data to the tmp_section structure.
517  * It relies on the listener calling us in memory address order
518  * and for each region (via the _add and _nop methods) to
519  * join neighbours.
520  */
521 static void vhost_region_add_section(struct vhost_dev *dev,
522                                      MemoryRegionSection *section)
523 {
524     bool need_add = true;
525     uint64_t mrs_size = int128_get64(section->size);
526     uint64_t mrs_gpa = section->offset_within_address_space;
527     uintptr_t mrs_host = (uintptr_t)memory_region_get_ram_ptr(section->mr) +
528                          section->offset_within_region;
529     RAMBlock *mrs_rb = section->mr->ram_block;
530     size_t mrs_page = qemu_ram_pagesize(mrs_rb);
531 
532     trace_vhost_region_add_section(section->mr->name, mrs_gpa, mrs_size,
533                                    mrs_host);
534 
535     /* Round the section to it's page size */
536     /* First align the start down to a page boundary */
537     uint64_t alignage = mrs_host & (mrs_page - 1);
538     if (alignage) {
539         mrs_host -= alignage;
540         mrs_size += alignage;
541         mrs_gpa  -= alignage;
542     }
543     /* Now align the size up to a page boundary */
544     alignage = mrs_size & (mrs_page - 1);
545     if (alignage) {
546         mrs_size += mrs_page - alignage;
547     }
548     trace_vhost_region_add_section_aligned(section->mr->name, mrs_gpa, mrs_size,
549                                            mrs_host);
550 
551     if (dev->n_tmp_sections) {
552         /* Since we already have at least one section, lets see if
553          * this extends it; since we're scanning in order, we only
554          * have to look at the last one, and the FlatView that calls
555          * us shouldn't have overlaps.
556          */
557         MemoryRegionSection *prev_sec = dev->tmp_sections +
558                                                (dev->n_tmp_sections - 1);
559         uint64_t prev_gpa_start = prev_sec->offset_within_address_space;
560         uint64_t prev_size = int128_get64(prev_sec->size);
561         uint64_t prev_gpa_end   = range_get_last(prev_gpa_start, prev_size);
562         uint64_t prev_host_start =
563                         (uintptr_t)memory_region_get_ram_ptr(prev_sec->mr) +
564                         prev_sec->offset_within_region;
565         uint64_t prev_host_end   = range_get_last(prev_host_start, prev_size);
566 
567         if (mrs_gpa <= (prev_gpa_end + 1)) {
568             /* OK, looks like overlapping/intersecting - it's possible that
569              * the rounding to page sizes has made them overlap, but they should
570              * match up in the same RAMBlock if they do.
571              */
572             if (mrs_gpa < prev_gpa_start) {
573                 error_report("%s:Section rounded to %"PRIx64
574                              " prior to previous %"PRIx64,
575                              __func__, mrs_gpa, prev_gpa_start);
576                 /* A way to cleanly fail here would be better */
577                 return;
578             }
579             /* Offset from the start of the previous GPA to this GPA */
580             size_t offset = mrs_gpa - prev_gpa_start;
581 
582             if (prev_host_start + offset == mrs_host &&
583                 section->mr == prev_sec->mr &&
584                 (!dev->vhost_ops->vhost_backend_can_merge ||
585                  dev->vhost_ops->vhost_backend_can_merge(dev,
586                     mrs_host, mrs_size,
587                     prev_host_start, prev_size))) {
588                 uint64_t max_end = MAX(prev_host_end, mrs_host + mrs_size);
589                 need_add = false;
590                 prev_sec->offset_within_address_space =
591                     MIN(prev_gpa_start, mrs_gpa);
592                 prev_sec->offset_within_region =
593                     MIN(prev_host_start, mrs_host) -
594                     (uintptr_t)memory_region_get_ram_ptr(prev_sec->mr);
595                 prev_sec->size = int128_make64(max_end - MIN(prev_host_start,
596                                                mrs_host));
597                 trace_vhost_region_add_section_merge(section->mr->name,
598                                         int128_get64(prev_sec->size),
599                                         prev_sec->offset_within_address_space,
600                                         prev_sec->offset_within_region);
601             } else {
602                 /* adjoining regions are fine, but overlapping ones with
603                  * different blocks/offsets shouldn't happen
604                  */
605                 if (mrs_gpa != prev_gpa_end + 1) {
606                     error_report("%s: Overlapping but not coherent sections "
607                                  "at %"PRIx64,
608                                  __func__, mrs_gpa);
609                     return;
610                 }
611             }
612         }
613     }
614 
615     if (need_add) {
616         ++dev->n_tmp_sections;
617         dev->tmp_sections = g_renew(MemoryRegionSection, dev->tmp_sections,
618                                     dev->n_tmp_sections);
619         dev->tmp_sections[dev->n_tmp_sections - 1] = *section;
620         /* The flatview isn't stable and we don't use it, making it NULL
621          * means we can memcmp the list.
622          */
623         dev->tmp_sections[dev->n_tmp_sections - 1].fv = NULL;
624         memory_region_ref(section->mr);
625     }
626 }
627 
628 /* Used for both add and nop callbacks */
629 static void vhost_region_addnop(MemoryListener *listener,
630                                 MemoryRegionSection *section)
631 {
632     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
633                                          memory_listener);
634 
635     if (!vhost_section(section)) {
636         return;
637     }
638     vhost_region_add_section(dev, section);
639 }
640 
641 static void vhost_iommu_unmap_notify(IOMMUNotifier *n, IOMMUTLBEntry *iotlb)
642 {
643     struct vhost_iommu *iommu = container_of(n, struct vhost_iommu, n);
644     struct vhost_dev *hdev = iommu->hdev;
645     hwaddr iova = iotlb->iova + iommu->iommu_offset;
646 
647     if (vhost_backend_invalidate_device_iotlb(hdev, iova,
648                                               iotlb->addr_mask + 1)) {
649         error_report("Fail to invalidate device iotlb");
650     }
651 }
652 
653 static void vhost_iommu_region_add(MemoryListener *listener,
654                                    MemoryRegionSection *section)
655 {
656     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
657                                          iommu_listener);
658     struct vhost_iommu *iommu;
659     Int128 end;
660 
661     if (!memory_region_is_iommu(section->mr)) {
662         return;
663     }
664 
665     iommu = g_malloc0(sizeof(*iommu));
666     end = int128_add(int128_make64(section->offset_within_region),
667                      section->size);
668     end = int128_sub(end, int128_one());
669     iommu_notifier_init(&iommu->n, vhost_iommu_unmap_notify,
670                         IOMMU_NOTIFIER_UNMAP,
671                         section->offset_within_region,
672                         int128_get64(end));
673     iommu->mr = section->mr;
674     iommu->iommu_offset = section->offset_within_address_space -
675                           section->offset_within_region;
676     iommu->hdev = dev;
677     memory_region_register_iommu_notifier(section->mr, &iommu->n);
678     QLIST_INSERT_HEAD(&dev->iommu_list, iommu, iommu_next);
679     /* TODO: can replay help performance here? */
680 }
681 
682 static void vhost_iommu_region_del(MemoryListener *listener,
683                                    MemoryRegionSection *section)
684 {
685     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
686                                          iommu_listener);
687     struct vhost_iommu *iommu;
688 
689     if (!memory_region_is_iommu(section->mr)) {
690         return;
691     }
692 
693     QLIST_FOREACH(iommu, &dev->iommu_list, iommu_next) {
694         if (iommu->mr == section->mr &&
695             iommu->n.start == section->offset_within_region) {
696             memory_region_unregister_iommu_notifier(iommu->mr,
697                                                     &iommu->n);
698             QLIST_REMOVE(iommu, iommu_next);
699             g_free(iommu);
700             break;
701         }
702     }
703 }
704 
705 static int vhost_virtqueue_set_addr(struct vhost_dev *dev,
706                                     struct vhost_virtqueue *vq,
707                                     unsigned idx, bool enable_log)
708 {
709     struct vhost_vring_addr addr = {
710         .index = idx,
711         .desc_user_addr = (uint64_t)(unsigned long)vq->desc,
712         .avail_user_addr = (uint64_t)(unsigned long)vq->avail,
713         .used_user_addr = (uint64_t)(unsigned long)vq->used,
714         .log_guest_addr = vq->used_phys,
715         .flags = enable_log ? (1 << VHOST_VRING_F_LOG) : 0,
716     };
717     int r = dev->vhost_ops->vhost_set_vring_addr(dev, &addr);
718     if (r < 0) {
719         VHOST_OPS_DEBUG("vhost_set_vring_addr failed");
720         return -errno;
721     }
722     return 0;
723 }
724 
725 static int vhost_dev_set_features(struct vhost_dev *dev,
726                                   bool enable_log)
727 {
728     uint64_t features = dev->acked_features;
729     int r;
730     if (enable_log) {
731         features |= 0x1ULL << VHOST_F_LOG_ALL;
732     }
733     r = dev->vhost_ops->vhost_set_features(dev, features);
734     if (r < 0) {
735         VHOST_OPS_DEBUG("vhost_set_features failed");
736     }
737     return r < 0 ? -errno : 0;
738 }
739 
740 static int vhost_dev_set_log(struct vhost_dev *dev, bool enable_log)
741 {
742     int r, i, idx;
743     r = vhost_dev_set_features(dev, enable_log);
744     if (r < 0) {
745         goto err_features;
746     }
747     for (i = 0; i < dev->nvqs; ++i) {
748         idx = dev->vhost_ops->vhost_get_vq_index(dev, dev->vq_index + i);
749         r = vhost_virtqueue_set_addr(dev, dev->vqs + i, idx,
750                                      enable_log);
751         if (r < 0) {
752             goto err_vq;
753         }
754     }
755     return 0;
756 err_vq:
757     for (; i >= 0; --i) {
758         idx = dev->vhost_ops->vhost_get_vq_index(dev, dev->vq_index + i);
759         vhost_virtqueue_set_addr(dev, dev->vqs + i, idx,
760                                  dev->log_enabled);
761     }
762     vhost_dev_set_features(dev, dev->log_enabled);
763 err_features:
764     return r;
765 }
766 
767 static int vhost_migration_log(MemoryListener *listener, int enable)
768 {
769     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
770                                          memory_listener);
771     int r;
772     if (!!enable == dev->log_enabled) {
773         return 0;
774     }
775     if (!dev->started) {
776         dev->log_enabled = enable;
777         return 0;
778     }
779     if (!enable) {
780         r = vhost_dev_set_log(dev, false);
781         if (r < 0) {
782             return r;
783         }
784         vhost_log_put(dev, false);
785     } else {
786         vhost_dev_log_resize(dev, vhost_get_log_size(dev));
787         r = vhost_dev_set_log(dev, true);
788         if (r < 0) {
789             return r;
790         }
791     }
792     dev->log_enabled = enable;
793     return 0;
794 }
795 
796 static void vhost_log_global_start(MemoryListener *listener)
797 {
798     int r;
799 
800     r = vhost_migration_log(listener, true);
801     if (r < 0) {
802         abort();
803     }
804 }
805 
806 static void vhost_log_global_stop(MemoryListener *listener)
807 {
808     int r;
809 
810     r = vhost_migration_log(listener, false);
811     if (r < 0) {
812         abort();
813     }
814 }
815 
816 static void vhost_log_start(MemoryListener *listener,
817                             MemoryRegionSection *section,
818                             int old, int new)
819 {
820     /* FIXME: implement */
821 }
822 
823 static void vhost_log_stop(MemoryListener *listener,
824                            MemoryRegionSection *section,
825                            int old, int new)
826 {
827     /* FIXME: implement */
828 }
829 
830 /* The vhost driver natively knows how to handle the vrings of non
831  * cross-endian legacy devices and modern devices. Only legacy devices
832  * exposed to a bi-endian guest may require the vhost driver to use a
833  * specific endianness.
834  */
835 static inline bool vhost_needs_vring_endian(VirtIODevice *vdev)
836 {
837     if (virtio_vdev_has_feature(vdev, VIRTIO_F_VERSION_1)) {
838         return false;
839     }
840 #ifdef HOST_WORDS_BIGENDIAN
841     return vdev->device_endian == VIRTIO_DEVICE_ENDIAN_LITTLE;
842 #else
843     return vdev->device_endian == VIRTIO_DEVICE_ENDIAN_BIG;
844 #endif
845 }
846 
847 static int vhost_virtqueue_set_vring_endian_legacy(struct vhost_dev *dev,
848                                                    bool is_big_endian,
849                                                    int vhost_vq_index)
850 {
851     struct vhost_vring_state s = {
852         .index = vhost_vq_index,
853         .num = is_big_endian
854     };
855 
856     if (!dev->vhost_ops->vhost_set_vring_endian(dev, &s)) {
857         return 0;
858     }
859 
860     VHOST_OPS_DEBUG("vhost_set_vring_endian failed");
861     if (errno == ENOTTY) {
862         error_report("vhost does not support cross-endian");
863         return -ENOSYS;
864     }
865 
866     return -errno;
867 }
868 
869 static int vhost_memory_region_lookup(struct vhost_dev *hdev,
870                                       uint64_t gpa, uint64_t *uaddr,
871                                       uint64_t *len)
872 {
873     int i;
874 
875     for (i = 0; i < hdev->mem->nregions; i++) {
876         struct vhost_memory_region *reg = hdev->mem->regions + i;
877 
878         if (gpa >= reg->guest_phys_addr &&
879             reg->guest_phys_addr + reg->memory_size > gpa) {
880             *uaddr = reg->userspace_addr + gpa - reg->guest_phys_addr;
881             *len = reg->guest_phys_addr + reg->memory_size - gpa;
882             return 0;
883         }
884     }
885 
886     return -EFAULT;
887 }
888 
889 int vhost_device_iotlb_miss(struct vhost_dev *dev, uint64_t iova, int write)
890 {
891     IOMMUTLBEntry iotlb;
892     uint64_t uaddr, len;
893     int ret = -EFAULT;
894 
895     rcu_read_lock();
896 
897     iotlb = address_space_get_iotlb_entry(dev->vdev->dma_as,
898                                           iova, write);
899     if (iotlb.target_as != NULL) {
900         ret = vhost_memory_region_lookup(dev, iotlb.translated_addr,
901                                          &uaddr, &len);
902         if (ret) {
903             error_report("Fail to lookup the translated address "
904                          "%"PRIx64, iotlb.translated_addr);
905             goto out;
906         }
907 
908         len = MIN(iotlb.addr_mask + 1, len);
909         iova = iova & ~iotlb.addr_mask;
910 
911         ret = vhost_backend_update_device_iotlb(dev, iova, uaddr,
912                                                 len, iotlb.perm);
913         if (ret) {
914             error_report("Fail to update device iotlb");
915             goto out;
916         }
917     }
918 out:
919     rcu_read_unlock();
920 
921     return ret;
922 }
923 
924 static int vhost_virtqueue_start(struct vhost_dev *dev,
925                                 struct VirtIODevice *vdev,
926                                 struct vhost_virtqueue *vq,
927                                 unsigned idx)
928 {
929     BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev)));
930     VirtioBusState *vbus = VIRTIO_BUS(qbus);
931     VirtioBusClass *k = VIRTIO_BUS_GET_CLASS(vbus);
932     hwaddr s, l, a;
933     int r;
934     int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, idx);
935     struct vhost_vring_file file = {
936         .index = vhost_vq_index
937     };
938     struct vhost_vring_state state = {
939         .index = vhost_vq_index
940     };
941     struct VirtQueue *vvq = virtio_get_queue(vdev, idx);
942 
943     a = virtio_queue_get_desc_addr(vdev, idx);
944     if (a == 0) {
945         /* Queue might not be ready for start */
946         return 0;
947     }
948 
949     vq->num = state.num = virtio_queue_get_num(vdev, idx);
950     r = dev->vhost_ops->vhost_set_vring_num(dev, &state);
951     if (r) {
952         VHOST_OPS_DEBUG("vhost_set_vring_num failed");
953         return -errno;
954     }
955 
956     state.num = virtio_queue_get_last_avail_idx(vdev, idx);
957     r = dev->vhost_ops->vhost_set_vring_base(dev, &state);
958     if (r) {
959         VHOST_OPS_DEBUG("vhost_set_vring_base failed");
960         return -errno;
961     }
962 
963     if (vhost_needs_vring_endian(vdev)) {
964         r = vhost_virtqueue_set_vring_endian_legacy(dev,
965                                                     virtio_is_big_endian(vdev),
966                                                     vhost_vq_index);
967         if (r) {
968             return -errno;
969         }
970     }
971 
972     vq->desc_size = s = l = virtio_queue_get_desc_size(vdev, idx);
973     vq->desc_phys = a;
974     vq->desc = vhost_memory_map(dev, a, &l, 0);
975     if (!vq->desc || l != s) {
976         r = -ENOMEM;
977         goto fail_alloc_desc;
978     }
979     vq->avail_size = s = l = virtio_queue_get_avail_size(vdev, idx);
980     vq->avail_phys = a = virtio_queue_get_avail_addr(vdev, idx);
981     vq->avail = vhost_memory_map(dev, a, &l, 0);
982     if (!vq->avail || l != s) {
983         r = -ENOMEM;
984         goto fail_alloc_avail;
985     }
986     vq->used_size = s = l = virtio_queue_get_used_size(vdev, idx);
987     vq->used_phys = a = virtio_queue_get_used_addr(vdev, idx);
988     vq->used = vhost_memory_map(dev, a, &l, 1);
989     if (!vq->used || l != s) {
990         r = -ENOMEM;
991         goto fail_alloc_used;
992     }
993 
994     r = vhost_virtqueue_set_addr(dev, vq, vhost_vq_index, dev->log_enabled);
995     if (r < 0) {
996         r = -errno;
997         goto fail_alloc;
998     }
999 
1000     file.fd = event_notifier_get_fd(virtio_queue_get_host_notifier(vvq));
1001     r = dev->vhost_ops->vhost_set_vring_kick(dev, &file);
1002     if (r) {
1003         VHOST_OPS_DEBUG("vhost_set_vring_kick failed");
1004         r = -errno;
1005         goto fail_kick;
1006     }
1007 
1008     /* Clear and discard previous events if any. */
1009     event_notifier_test_and_clear(&vq->masked_notifier);
1010 
1011     /* Init vring in unmasked state, unless guest_notifier_mask
1012      * will do it later.
1013      */
1014     if (!vdev->use_guest_notifier_mask) {
1015         /* TODO: check and handle errors. */
1016         vhost_virtqueue_mask(dev, vdev, idx, false);
1017     }
1018 
1019     if (k->query_guest_notifiers &&
1020         k->query_guest_notifiers(qbus->parent) &&
1021         virtio_queue_vector(vdev, idx) == VIRTIO_NO_VECTOR) {
1022         file.fd = -1;
1023         r = dev->vhost_ops->vhost_set_vring_call(dev, &file);
1024         if (r) {
1025             goto fail_vector;
1026         }
1027     }
1028 
1029     return 0;
1030 
1031 fail_vector:
1032 fail_kick:
1033 fail_alloc:
1034     vhost_memory_unmap(dev, vq->used, virtio_queue_get_used_size(vdev, idx),
1035                        0, 0);
1036 fail_alloc_used:
1037     vhost_memory_unmap(dev, vq->avail, virtio_queue_get_avail_size(vdev, idx),
1038                        0, 0);
1039 fail_alloc_avail:
1040     vhost_memory_unmap(dev, vq->desc, virtio_queue_get_desc_size(vdev, idx),
1041                        0, 0);
1042 fail_alloc_desc:
1043     return r;
1044 }
1045 
1046 static void vhost_virtqueue_stop(struct vhost_dev *dev,
1047                                     struct VirtIODevice *vdev,
1048                                     struct vhost_virtqueue *vq,
1049                                     unsigned idx)
1050 {
1051     int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, idx);
1052     struct vhost_vring_state state = {
1053         .index = vhost_vq_index,
1054     };
1055     int r;
1056     int a;
1057 
1058     a = virtio_queue_get_desc_addr(vdev, idx);
1059     if (a == 0) {
1060         /* Don't stop the virtqueue which might have not been started */
1061         return;
1062     }
1063 
1064     r = dev->vhost_ops->vhost_get_vring_base(dev, &state);
1065     if (r < 0) {
1066         VHOST_OPS_DEBUG("vhost VQ %d ring restore failed: %d", idx, r);
1067         /* Connection to the backend is broken, so let's sync internal
1068          * last avail idx to the device used idx.
1069          */
1070         virtio_queue_restore_last_avail_idx(vdev, idx);
1071     } else {
1072         virtio_queue_set_last_avail_idx(vdev, idx, state.num);
1073     }
1074     virtio_queue_invalidate_signalled_used(vdev, idx);
1075     virtio_queue_update_used_idx(vdev, idx);
1076 
1077     /* In the cross-endian case, we need to reset the vring endianness to
1078      * native as legacy devices expect so by default.
1079      */
1080     if (vhost_needs_vring_endian(vdev)) {
1081         vhost_virtqueue_set_vring_endian_legacy(dev,
1082                                                 !virtio_is_big_endian(vdev),
1083                                                 vhost_vq_index);
1084     }
1085 
1086     vhost_memory_unmap(dev, vq->used, virtio_queue_get_used_size(vdev, idx),
1087                        1, virtio_queue_get_used_size(vdev, idx));
1088     vhost_memory_unmap(dev, vq->avail, virtio_queue_get_avail_size(vdev, idx),
1089                        0, virtio_queue_get_avail_size(vdev, idx));
1090     vhost_memory_unmap(dev, vq->desc, virtio_queue_get_desc_size(vdev, idx),
1091                        0, virtio_queue_get_desc_size(vdev, idx));
1092 }
1093 
1094 static void vhost_eventfd_add(MemoryListener *listener,
1095                               MemoryRegionSection *section,
1096                               bool match_data, uint64_t data, EventNotifier *e)
1097 {
1098 }
1099 
1100 static void vhost_eventfd_del(MemoryListener *listener,
1101                               MemoryRegionSection *section,
1102                               bool match_data, uint64_t data, EventNotifier *e)
1103 {
1104 }
1105 
1106 static int vhost_virtqueue_set_busyloop_timeout(struct vhost_dev *dev,
1107                                                 int n, uint32_t timeout)
1108 {
1109     int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, n);
1110     struct vhost_vring_state state = {
1111         .index = vhost_vq_index,
1112         .num = timeout,
1113     };
1114     int r;
1115 
1116     if (!dev->vhost_ops->vhost_set_vring_busyloop_timeout) {
1117         return -EINVAL;
1118     }
1119 
1120     r = dev->vhost_ops->vhost_set_vring_busyloop_timeout(dev, &state);
1121     if (r) {
1122         VHOST_OPS_DEBUG("vhost_set_vring_busyloop_timeout failed");
1123         return r;
1124     }
1125 
1126     return 0;
1127 }
1128 
1129 static int vhost_virtqueue_init(struct vhost_dev *dev,
1130                                 struct vhost_virtqueue *vq, int n)
1131 {
1132     int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, n);
1133     struct vhost_vring_file file = {
1134         .index = vhost_vq_index,
1135     };
1136     int r = event_notifier_init(&vq->masked_notifier, 0);
1137     if (r < 0) {
1138         return r;
1139     }
1140 
1141     file.fd = event_notifier_get_fd(&vq->masked_notifier);
1142     r = dev->vhost_ops->vhost_set_vring_call(dev, &file);
1143     if (r) {
1144         VHOST_OPS_DEBUG("vhost_set_vring_call failed");
1145         r = -errno;
1146         goto fail_call;
1147     }
1148 
1149     vq->dev = dev;
1150 
1151     return 0;
1152 fail_call:
1153     event_notifier_cleanup(&vq->masked_notifier);
1154     return r;
1155 }
1156 
1157 static void vhost_virtqueue_cleanup(struct vhost_virtqueue *vq)
1158 {
1159     event_notifier_cleanup(&vq->masked_notifier);
1160 }
1161 
1162 int vhost_dev_init(struct vhost_dev *hdev, void *opaque,
1163                    VhostBackendType backend_type, uint32_t busyloop_timeout)
1164 {
1165     uint64_t features;
1166     int i, r, n_initialized_vqs = 0;
1167     Error *local_err = NULL;
1168 
1169     hdev->vdev = NULL;
1170     hdev->migration_blocker = NULL;
1171 
1172     r = vhost_set_backend_type(hdev, backend_type);
1173     assert(r >= 0);
1174 
1175     r = hdev->vhost_ops->vhost_backend_init(hdev, opaque);
1176     if (r < 0) {
1177         goto fail;
1178     }
1179 
1180     r = hdev->vhost_ops->vhost_set_owner(hdev);
1181     if (r < 0) {
1182         VHOST_OPS_DEBUG("vhost_set_owner failed");
1183         goto fail;
1184     }
1185 
1186     r = hdev->vhost_ops->vhost_get_features(hdev, &features);
1187     if (r < 0) {
1188         VHOST_OPS_DEBUG("vhost_get_features failed");
1189         goto fail;
1190     }
1191 
1192     for (i = 0; i < hdev->nvqs; ++i, ++n_initialized_vqs) {
1193         r = vhost_virtqueue_init(hdev, hdev->vqs + i, hdev->vq_index + i);
1194         if (r < 0) {
1195             goto fail;
1196         }
1197     }
1198 
1199     if (busyloop_timeout) {
1200         for (i = 0; i < hdev->nvqs; ++i) {
1201             r = vhost_virtqueue_set_busyloop_timeout(hdev, hdev->vq_index + i,
1202                                                      busyloop_timeout);
1203             if (r < 0) {
1204                 goto fail_busyloop;
1205             }
1206         }
1207     }
1208 
1209     hdev->features = features;
1210 
1211     hdev->memory_listener = (MemoryListener) {
1212         .begin = vhost_begin,
1213         .commit = vhost_commit,
1214         .region_add = vhost_region_addnop,
1215         .region_nop = vhost_region_addnop,
1216         .log_start = vhost_log_start,
1217         .log_stop = vhost_log_stop,
1218         .log_sync = vhost_log_sync,
1219         .log_global_start = vhost_log_global_start,
1220         .log_global_stop = vhost_log_global_stop,
1221         .eventfd_add = vhost_eventfd_add,
1222         .eventfd_del = vhost_eventfd_del,
1223         .priority = 10
1224     };
1225 
1226     hdev->iommu_listener = (MemoryListener) {
1227         .region_add = vhost_iommu_region_add,
1228         .region_del = vhost_iommu_region_del,
1229     };
1230 
1231     if (hdev->migration_blocker == NULL) {
1232         if (!(hdev->features & (0x1ULL << VHOST_F_LOG_ALL))) {
1233             error_setg(&hdev->migration_blocker,
1234                        "Migration disabled: vhost lacks VHOST_F_LOG_ALL feature.");
1235         } else if (vhost_dev_log_is_shared(hdev) && !qemu_memfd_alloc_check()) {
1236             error_setg(&hdev->migration_blocker,
1237                        "Migration disabled: failed to allocate shared memory");
1238         }
1239     }
1240 
1241     if (hdev->migration_blocker != NULL) {
1242         r = migrate_add_blocker(hdev->migration_blocker, &local_err);
1243         if (local_err) {
1244             error_report_err(local_err);
1245             error_free(hdev->migration_blocker);
1246             goto fail_busyloop;
1247         }
1248     }
1249 
1250     hdev->mem = g_malloc0(offsetof(struct vhost_memory, regions));
1251     hdev->n_mem_sections = 0;
1252     hdev->mem_sections = NULL;
1253     hdev->log = NULL;
1254     hdev->log_size = 0;
1255     hdev->log_enabled = false;
1256     hdev->started = false;
1257     memory_listener_register(&hdev->memory_listener, &address_space_memory);
1258     QLIST_INSERT_HEAD(&vhost_devices, hdev, entry);
1259 
1260     if (used_memslots > hdev->vhost_ops->vhost_backend_memslots_limit(hdev)) {
1261         error_report("vhost backend memory slots limit is less"
1262                 " than current number of present memory slots");
1263         r = -1;
1264         if (busyloop_timeout) {
1265             goto fail_busyloop;
1266         } else {
1267             goto fail;
1268         }
1269     }
1270 
1271     return 0;
1272 
1273 fail_busyloop:
1274     while (--i >= 0) {
1275         vhost_virtqueue_set_busyloop_timeout(hdev, hdev->vq_index + i, 0);
1276     }
1277 fail:
1278     hdev->nvqs = n_initialized_vqs;
1279     vhost_dev_cleanup(hdev);
1280     return r;
1281 }
1282 
1283 void vhost_dev_cleanup(struct vhost_dev *hdev)
1284 {
1285     int i;
1286 
1287     for (i = 0; i < hdev->nvqs; ++i) {
1288         vhost_virtqueue_cleanup(hdev->vqs + i);
1289     }
1290     if (hdev->mem) {
1291         /* those are only safe after successful init */
1292         memory_listener_unregister(&hdev->memory_listener);
1293         QLIST_REMOVE(hdev, entry);
1294     }
1295     if (hdev->migration_blocker) {
1296         migrate_del_blocker(hdev->migration_blocker);
1297         error_free(hdev->migration_blocker);
1298     }
1299     g_free(hdev->mem);
1300     g_free(hdev->mem_sections);
1301     if (hdev->vhost_ops) {
1302         hdev->vhost_ops->vhost_backend_cleanup(hdev);
1303     }
1304     assert(!hdev->log);
1305 
1306     memset(hdev, 0, sizeof(struct vhost_dev));
1307 }
1308 
1309 /* Stop processing guest IO notifications in qemu.
1310  * Start processing them in vhost in kernel.
1311  */
1312 int vhost_dev_enable_notifiers(struct vhost_dev *hdev, VirtIODevice *vdev)
1313 {
1314     BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev)));
1315     int i, r, e;
1316 
1317     /* We will pass the notifiers to the kernel, make sure that QEMU
1318      * doesn't interfere.
1319      */
1320     r = virtio_device_grab_ioeventfd(vdev);
1321     if (r < 0) {
1322         error_report("binding does not support host notifiers");
1323         goto fail;
1324     }
1325 
1326     for (i = 0; i < hdev->nvqs; ++i) {
1327         r = virtio_bus_set_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i,
1328                                          true);
1329         if (r < 0) {
1330             error_report("vhost VQ %d notifier binding failed: %d", i, -r);
1331             goto fail_vq;
1332         }
1333     }
1334 
1335     return 0;
1336 fail_vq:
1337     while (--i >= 0) {
1338         e = virtio_bus_set_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i,
1339                                          false);
1340         if (e < 0) {
1341             error_report("vhost VQ %d notifier cleanup error: %d", i, -r);
1342         }
1343         assert (e >= 0);
1344         virtio_bus_cleanup_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i);
1345     }
1346     virtio_device_release_ioeventfd(vdev);
1347 fail:
1348     return r;
1349 }
1350 
1351 /* Stop processing guest IO notifications in vhost.
1352  * Start processing them in qemu.
1353  * This might actually run the qemu handlers right away,
1354  * so virtio in qemu must be completely setup when this is called.
1355  */
1356 void vhost_dev_disable_notifiers(struct vhost_dev *hdev, VirtIODevice *vdev)
1357 {
1358     BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev)));
1359     int i, r;
1360 
1361     for (i = 0; i < hdev->nvqs; ++i) {
1362         r = virtio_bus_set_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i,
1363                                          false);
1364         if (r < 0) {
1365             error_report("vhost VQ %d notifier cleanup failed: %d", i, -r);
1366         }
1367         assert (r >= 0);
1368         virtio_bus_cleanup_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i);
1369     }
1370     virtio_device_release_ioeventfd(vdev);
1371 }
1372 
1373 /* Test and clear event pending status.
1374  * Should be called after unmask to avoid losing events.
1375  */
1376 bool vhost_virtqueue_pending(struct vhost_dev *hdev, int n)
1377 {
1378     struct vhost_virtqueue *vq = hdev->vqs + n - hdev->vq_index;
1379     assert(n >= hdev->vq_index && n < hdev->vq_index + hdev->nvqs);
1380     return event_notifier_test_and_clear(&vq->masked_notifier);
1381 }
1382 
1383 /* Mask/unmask events from this vq. */
1384 void vhost_virtqueue_mask(struct vhost_dev *hdev, VirtIODevice *vdev, int n,
1385                          bool mask)
1386 {
1387     struct VirtQueue *vvq = virtio_get_queue(vdev, n);
1388     int r, index = n - hdev->vq_index;
1389     struct vhost_vring_file file;
1390 
1391     /* should only be called after backend is connected */
1392     assert(hdev->vhost_ops);
1393 
1394     if (mask) {
1395         assert(vdev->use_guest_notifier_mask);
1396         file.fd = event_notifier_get_fd(&hdev->vqs[index].masked_notifier);
1397     } else {
1398         file.fd = event_notifier_get_fd(virtio_queue_get_guest_notifier(vvq));
1399     }
1400 
1401     file.index = hdev->vhost_ops->vhost_get_vq_index(hdev, n);
1402     r = hdev->vhost_ops->vhost_set_vring_call(hdev, &file);
1403     if (r < 0) {
1404         VHOST_OPS_DEBUG("vhost_set_vring_call failed");
1405     }
1406 }
1407 
1408 uint64_t vhost_get_features(struct vhost_dev *hdev, const int *feature_bits,
1409                             uint64_t features)
1410 {
1411     const int *bit = feature_bits;
1412     while (*bit != VHOST_INVALID_FEATURE_BIT) {
1413         uint64_t bit_mask = (1ULL << *bit);
1414         if (!(hdev->features & bit_mask)) {
1415             features &= ~bit_mask;
1416         }
1417         bit++;
1418     }
1419     return features;
1420 }
1421 
1422 void vhost_ack_features(struct vhost_dev *hdev, const int *feature_bits,
1423                         uint64_t features)
1424 {
1425     const int *bit = feature_bits;
1426     while (*bit != VHOST_INVALID_FEATURE_BIT) {
1427         uint64_t bit_mask = (1ULL << *bit);
1428         if (features & bit_mask) {
1429             hdev->acked_features |= bit_mask;
1430         }
1431         bit++;
1432     }
1433 }
1434 
1435 int vhost_dev_get_config(struct vhost_dev *hdev, uint8_t *config,
1436                          uint32_t config_len)
1437 {
1438     assert(hdev->vhost_ops);
1439 
1440     if (hdev->vhost_ops->vhost_get_config) {
1441         return hdev->vhost_ops->vhost_get_config(hdev, config, config_len);
1442     }
1443 
1444     return -1;
1445 }
1446 
1447 int vhost_dev_set_config(struct vhost_dev *hdev, const uint8_t *data,
1448                          uint32_t offset, uint32_t size, uint32_t flags)
1449 {
1450     assert(hdev->vhost_ops);
1451 
1452     if (hdev->vhost_ops->vhost_set_config) {
1453         return hdev->vhost_ops->vhost_set_config(hdev, data, offset,
1454                                                  size, flags);
1455     }
1456 
1457     return -1;
1458 }
1459 
1460 void vhost_dev_set_config_notifier(struct vhost_dev *hdev,
1461                                    const VhostDevConfigOps *ops)
1462 {
1463     hdev->config_ops = ops;
1464 }
1465 
1466 /* Host notifiers must be enabled at this point. */
1467 int vhost_dev_start(struct vhost_dev *hdev, VirtIODevice *vdev)
1468 {
1469     int i, r;
1470 
1471     /* should only be called after backend is connected */
1472     assert(hdev->vhost_ops);
1473 
1474     hdev->started = true;
1475     hdev->vdev = vdev;
1476 
1477     r = vhost_dev_set_features(hdev, hdev->log_enabled);
1478     if (r < 0) {
1479         goto fail_features;
1480     }
1481 
1482     if (vhost_dev_has_iommu(hdev)) {
1483         memory_listener_register(&hdev->iommu_listener, vdev->dma_as);
1484     }
1485 
1486     r = hdev->vhost_ops->vhost_set_mem_table(hdev, hdev->mem);
1487     if (r < 0) {
1488         VHOST_OPS_DEBUG("vhost_set_mem_table failed");
1489         r = -errno;
1490         goto fail_mem;
1491     }
1492     for (i = 0; i < hdev->nvqs; ++i) {
1493         r = vhost_virtqueue_start(hdev,
1494                                   vdev,
1495                                   hdev->vqs + i,
1496                                   hdev->vq_index + i);
1497         if (r < 0) {
1498             goto fail_vq;
1499         }
1500     }
1501 
1502     if (hdev->log_enabled) {
1503         uint64_t log_base;
1504 
1505         hdev->log_size = vhost_get_log_size(hdev);
1506         hdev->log = vhost_log_get(hdev->log_size,
1507                                   vhost_dev_log_is_shared(hdev));
1508         log_base = (uintptr_t)hdev->log->log;
1509         r = hdev->vhost_ops->vhost_set_log_base(hdev,
1510                                                 hdev->log_size ? log_base : 0,
1511                                                 hdev->log);
1512         if (r < 0) {
1513             VHOST_OPS_DEBUG("vhost_set_log_base failed");
1514             r = -errno;
1515             goto fail_log;
1516         }
1517     }
1518 
1519     if (vhost_dev_has_iommu(hdev)) {
1520         hdev->vhost_ops->vhost_set_iotlb_callback(hdev, true);
1521 
1522         /* Update used ring information for IOTLB to work correctly,
1523          * vhost-kernel code requires for this.*/
1524         for (i = 0; i < hdev->nvqs; ++i) {
1525             struct vhost_virtqueue *vq = hdev->vqs + i;
1526             vhost_device_iotlb_miss(hdev, vq->used_phys, true);
1527         }
1528     }
1529     return 0;
1530 fail_log:
1531     vhost_log_put(hdev, false);
1532 fail_vq:
1533     while (--i >= 0) {
1534         vhost_virtqueue_stop(hdev,
1535                              vdev,
1536                              hdev->vqs + i,
1537                              hdev->vq_index + i);
1538     }
1539     i = hdev->nvqs;
1540 
1541 fail_mem:
1542 fail_features:
1543 
1544     hdev->started = false;
1545     return r;
1546 }
1547 
1548 /* Host notifiers must be enabled at this point. */
1549 void vhost_dev_stop(struct vhost_dev *hdev, VirtIODevice *vdev)
1550 {
1551     int i;
1552 
1553     /* should only be called after backend is connected */
1554     assert(hdev->vhost_ops);
1555 
1556     for (i = 0; i < hdev->nvqs; ++i) {
1557         vhost_virtqueue_stop(hdev,
1558                              vdev,
1559                              hdev->vqs + i,
1560                              hdev->vq_index + i);
1561     }
1562 
1563     if (vhost_dev_has_iommu(hdev)) {
1564         hdev->vhost_ops->vhost_set_iotlb_callback(hdev, false);
1565         memory_listener_unregister(&hdev->iommu_listener);
1566     }
1567     vhost_log_put(hdev, true);
1568     hdev->started = false;
1569     hdev->vdev = NULL;
1570 }
1571 
1572 int vhost_net_set_backend(struct vhost_dev *hdev,
1573                           struct vhost_vring_file *file)
1574 {
1575     if (hdev->vhost_ops->vhost_net_set_backend) {
1576         return hdev->vhost_ops->vhost_net_set_backend(hdev, file);
1577     }
1578 
1579     return -1;
1580 }
1581