vhost: switch to use IOTLB v2 format
[qemu.git] / hw / virtio / vhost.c
1 /*
2 * vhost support
3 *
4 * Copyright Red Hat, Inc. 2010
5 *
6 * Authors:
7 * Michael S. Tsirkin <mst@redhat.com>
8 *
9 * This work is licensed under the terms of the GNU GPL, version 2. See
10 * the COPYING file in the top-level directory.
11 *
12 * Contributions after 2012-01-13 are licensed under the terms of the
13 * GNU GPL, version 2 or (at your option) any later version.
14 */
15
16 #include "qemu/osdep.h"
17 #include "qapi/error.h"
18 #include "hw/virtio/vhost.h"
19 #include "qemu/atomic.h"
20 #include "qemu/range.h"
21 #include "qemu/error-report.h"
22 #include "qemu/memfd.h"
23 #include "standard-headers/linux/vhost_types.h"
24 #include "exec/address-spaces.h"
25 #include "hw/virtio/virtio-bus.h"
26 #include "hw/virtio/virtio-access.h"
27 #include "migration/blocker.h"
28 #include "migration/qemu-file-types.h"
29 #include "sysemu/dma.h"
30 #include "sysemu/tcg.h"
31 #include "trace.h"
32
33 /* enabled until disconnected backend stabilizes */
34 #define _VHOST_DEBUG 1
35
36 #ifdef _VHOST_DEBUG
37 #define VHOST_OPS_DEBUG(fmt, ...) \
38 do { error_report(fmt ": %s (%d)", ## __VA_ARGS__, \
39 strerror(errno), errno); } while (0)
40 #else
41 #define VHOST_OPS_DEBUG(fmt, ...) \
42 do { } while (0)
43 #endif
44
45 static struct vhost_log *vhost_log;
46 static struct vhost_log *vhost_log_shm;
47
48 static unsigned int used_memslots;
49 static QLIST_HEAD(, vhost_dev) vhost_devices =
50 QLIST_HEAD_INITIALIZER(vhost_devices);
51
52 bool vhost_has_free_slot(void)
53 {
54 unsigned int slots_limit = ~0U;
55 struct vhost_dev *hdev;
56
57 QLIST_FOREACH(hdev, &vhost_devices, entry) {
58 unsigned int r = hdev->vhost_ops->vhost_backend_memslots_limit(hdev);
59 slots_limit = MIN(slots_limit, r);
60 }
61 return slots_limit > used_memslots;
62 }
63
64 static void vhost_dev_sync_region(struct vhost_dev *dev,
65 MemoryRegionSection *section,
66 uint64_t mfirst, uint64_t mlast,
67 uint64_t rfirst, uint64_t rlast)
68 {
69 vhost_log_chunk_t *log = dev->log->log;
70
71 uint64_t start = MAX(mfirst, rfirst);
72 uint64_t end = MIN(mlast, rlast);
73 vhost_log_chunk_t *from = log + start / VHOST_LOG_CHUNK;
74 vhost_log_chunk_t *to = log + end / VHOST_LOG_CHUNK + 1;
75 uint64_t addr = QEMU_ALIGN_DOWN(start, VHOST_LOG_CHUNK);
76
77 if (end < start) {
78 return;
79 }
80 assert(end / VHOST_LOG_CHUNK < dev->log_size);
81 assert(start / VHOST_LOG_CHUNK < dev->log_size);
82
83 for (;from < to; ++from) {
84 vhost_log_chunk_t log;
85 /* We first check with non-atomic: much cheaper,
86 * and we expect non-dirty to be the common case. */
87 if (!*from) {
88 addr += VHOST_LOG_CHUNK;
89 continue;
90 }
91 /* Data must be read atomically. We don't really need barrier semantics
92 * but it's easier to use atomic_* than roll our own. */
93 log = qatomic_xchg(from, 0);
94 while (log) {
95 int bit = ctzl(log);
96 hwaddr page_addr;
97 hwaddr section_offset;
98 hwaddr mr_offset;
99 page_addr = addr + bit * VHOST_LOG_PAGE;
100 section_offset = page_addr - section->offset_within_address_space;
101 mr_offset = section_offset + section->offset_within_region;
102 memory_region_set_dirty(section->mr, mr_offset, VHOST_LOG_PAGE);
103 log &= ~(0x1ull << bit);
104 }
105 addr += VHOST_LOG_CHUNK;
106 }
107 }
108
109 static int vhost_sync_dirty_bitmap(struct vhost_dev *dev,
110 MemoryRegionSection *section,
111 hwaddr first,
112 hwaddr last)
113 {
114 int i;
115 hwaddr start_addr;
116 hwaddr end_addr;
117
118 if (!dev->log_enabled || !dev->started) {
119 return 0;
120 }
121 start_addr = section->offset_within_address_space;
122 end_addr = range_get_last(start_addr, int128_get64(section->size));
123 start_addr = MAX(first, start_addr);
124 end_addr = MIN(last, end_addr);
125
126 for (i = 0; i < dev->mem->nregions; ++i) {
127 struct vhost_memory_region *reg = dev->mem->regions + i;
128 vhost_dev_sync_region(dev, section, start_addr, end_addr,
129 reg->guest_phys_addr,
130 range_get_last(reg->guest_phys_addr,
131 reg->memory_size));
132 }
133 for (i = 0; i < dev->nvqs; ++i) {
134 struct vhost_virtqueue *vq = dev->vqs + i;
135
136 if (!vq->used_phys && !vq->used_size) {
137 continue;
138 }
139
140 vhost_dev_sync_region(dev, section, start_addr, end_addr, vq->used_phys,
141 range_get_last(vq->used_phys, vq->used_size));
142 }
143 return 0;
144 }
145
146 static void vhost_log_sync(MemoryListener *listener,
147 MemoryRegionSection *section)
148 {
149 struct vhost_dev *dev = container_of(listener, struct vhost_dev,
150 memory_listener);
151 vhost_sync_dirty_bitmap(dev, section, 0x0, ~0x0ULL);
152 }
153
154 static void vhost_log_sync_range(struct vhost_dev *dev,
155 hwaddr first, hwaddr last)
156 {
157 int i;
158 /* FIXME: this is N^2 in number of sections */
159 for (i = 0; i < dev->n_mem_sections; ++i) {
160 MemoryRegionSection *section = &dev->mem_sections[i];
161 vhost_sync_dirty_bitmap(dev, section, first, last);
162 }
163 }
164
165 static uint64_t vhost_get_log_size(struct vhost_dev *dev)
166 {
167 uint64_t log_size = 0;
168 int i;
169 for (i = 0; i < dev->mem->nregions; ++i) {
170 struct vhost_memory_region *reg = dev->mem->regions + i;
171 uint64_t last = range_get_last(reg->guest_phys_addr,
172 reg->memory_size);
173 log_size = MAX(log_size, last / VHOST_LOG_CHUNK + 1);
174 }
175 for (i = 0; i < dev->nvqs; ++i) {
176 struct vhost_virtqueue *vq = dev->vqs + i;
177
178 if (!vq->used_phys && !vq->used_size) {
179 continue;
180 }
181
182 uint64_t last = vq->used_phys + vq->used_size - 1;
183 log_size = MAX(log_size, last / VHOST_LOG_CHUNK + 1);
184 }
185 return log_size;
186 }
187
188 static struct vhost_log *vhost_log_alloc(uint64_t size, bool share)
189 {
190 Error *err = NULL;
191 struct vhost_log *log;
192 uint64_t logsize = size * sizeof(*(log->log));
193 int fd = -1;
194
195 log = g_new0(struct vhost_log, 1);
196 if (share) {
197 log->log = qemu_memfd_alloc("vhost-log", logsize,
198 F_SEAL_GROW | F_SEAL_SHRINK | F_SEAL_SEAL,
199 &fd, &err);
200 if (err) {
201 error_report_err(err);
202 g_free(log);
203 return NULL;
204 }
205 memset(log->log, 0, logsize);
206 } else {
207 log->log = g_malloc0(logsize);
208 }
209
210 log->size = size;
211 log->refcnt = 1;
212 log->fd = fd;
213
214 return log;
215 }
216
217 static struct vhost_log *vhost_log_get(uint64_t size, bool share)
218 {
219 struct vhost_log *log = share ? vhost_log_shm : vhost_log;
220
221 if (!log || log->size != size) {
222 log = vhost_log_alloc(size, share);
223 if (share) {
224 vhost_log_shm = log;
225 } else {
226 vhost_log = log;
227 }
228 } else {
229 ++log->refcnt;
230 }
231
232 return log;
233 }
234
235 static void vhost_log_put(struct vhost_dev *dev, bool sync)
236 {
237 struct vhost_log *log = dev->log;
238
239 if (!log) {
240 return;
241 }
242
243 --log->refcnt;
244 if (log->refcnt == 0) {
245 /* Sync only the range covered by the old log */
246 if (dev->log_size && sync) {
247 vhost_log_sync_range(dev, 0, dev->log_size * VHOST_LOG_CHUNK - 1);
248 }
249
250 if (vhost_log == log) {
251 g_free(log->log);
252 vhost_log = NULL;
253 } else if (vhost_log_shm == log) {
254 qemu_memfd_free(log->log, log->size * sizeof(*(log->log)),
255 log->fd);
256 vhost_log_shm = NULL;
257 }
258
259 g_free(log);
260 }
261
262 dev->log = NULL;
263 dev->log_size = 0;
264 }
265
266 static bool vhost_dev_log_is_shared(struct vhost_dev *dev)
267 {
268 return dev->vhost_ops->vhost_requires_shm_log &&
269 dev->vhost_ops->vhost_requires_shm_log(dev);
270 }
271
272 static inline void vhost_dev_log_resize(struct vhost_dev *dev, uint64_t size)
273 {
274 struct vhost_log *log = vhost_log_get(size, vhost_dev_log_is_shared(dev));
275 uint64_t log_base = (uintptr_t)log->log;
276 int r;
277
278 /* inform backend of log switching, this must be done before
279 releasing the current log, to ensure no logging is lost */
280 r = dev->vhost_ops->vhost_set_log_base(dev, log_base, log);
281 if (r < 0) {
282 VHOST_OPS_DEBUG("vhost_set_log_base failed");
283 }
284
285 vhost_log_put(dev, true);
286 dev->log = log;
287 dev->log_size = size;
288 }
289
290 static int vhost_dev_has_iommu(struct vhost_dev *dev)
291 {
292 VirtIODevice *vdev = dev->vdev;
293
294 /*
295 * For vhost, VIRTIO_F_IOMMU_PLATFORM means the backend support
296 * incremental memory mapping API via IOTLB API. For platform that
297 * does not have IOMMU, there's no need to enable this feature
298 * which may cause unnecessary IOTLB miss/update trnasactions.
299 */
300 return vdev->dma_as != &address_space_memory &&
301 virtio_host_has_feature(vdev, VIRTIO_F_IOMMU_PLATFORM);
302 }
303
304 static void *vhost_memory_map(struct vhost_dev *dev, hwaddr addr,
305 hwaddr *plen, bool is_write)
306 {
307 if (!vhost_dev_has_iommu(dev)) {
308 return cpu_physical_memory_map(addr, plen, is_write);
309 } else {
310 return (void *)(uintptr_t)addr;
311 }
312 }
313
314 static void vhost_memory_unmap(struct vhost_dev *dev, void *buffer,
315 hwaddr len, int is_write,
316 hwaddr access_len)
317 {
318 if (!vhost_dev_has_iommu(dev)) {
319 cpu_physical_memory_unmap(buffer, len, is_write, access_len);
320 }
321 }
322
323 static int vhost_verify_ring_part_mapping(void *ring_hva,
324 uint64_t ring_gpa,
325 uint64_t ring_size,
326 void *reg_hva,
327 uint64_t reg_gpa,
328 uint64_t reg_size)
329 {
330 uint64_t hva_ring_offset;
331 uint64_t ring_last = range_get_last(ring_gpa, ring_size);
332 uint64_t reg_last = range_get_last(reg_gpa, reg_size);
333
334 if (ring_last < reg_gpa || ring_gpa > reg_last) {
335 return 0;
336 }
337 /* check that whole ring's is mapped */
338 if (ring_last > reg_last) {
339 return -ENOMEM;
340 }
341 /* check that ring's MemoryRegion wasn't replaced */
342 hva_ring_offset = ring_gpa - reg_gpa;
343 if (ring_hva != reg_hva + hva_ring_offset) {
344 return -EBUSY;
345 }
346
347 return 0;
348 }
349
350 static int vhost_verify_ring_mappings(struct vhost_dev *dev,
351 void *reg_hva,
352 uint64_t reg_gpa,
353 uint64_t reg_size)
354 {
355 int i, j;
356 int r = 0;
357 const char *part_name[] = {
358 "descriptor table",
359 "available ring",
360 "used ring"
361 };
362
363 if (vhost_dev_has_iommu(dev)) {
364 return 0;
365 }
366
367 for (i = 0; i < dev->nvqs; ++i) {
368 struct vhost_virtqueue *vq = dev->vqs + i;
369
370 if (vq->desc_phys == 0) {
371 continue;
372 }
373
374 j = 0;
375 r = vhost_verify_ring_part_mapping(
376 vq->desc, vq->desc_phys, vq->desc_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->avail, vq->avail_phys, vq->avail_size,
385 reg_hva, reg_gpa, reg_size);
386 if (r) {
387 break;
388 }
389
390 j++;
391 r = vhost_verify_ring_part_mapping(
392 vq->used, vq->used_phys, vq->used_size,
393 reg_hva, reg_gpa, reg_size);
394 if (r) {
395 break;
396 }
397 }
398
399 if (r == -ENOMEM) {
400 error_report("Unable to map %s for ring %d", part_name[j], i);
401 } else if (r == -EBUSY) {
402 error_report("%s relocated for ring %d", part_name[j], i);
403 }
404 return r;
405 }
406
407 /*
408 * vhost_section: identify sections needed for vhost access
409 *
410 * We only care about RAM sections here (where virtqueue and guest
411 * internals accessed by virtio might live). If we find one we still
412 * allow the backend to potentially filter it out of our list.
413 */
414 static bool vhost_section(struct vhost_dev *dev, MemoryRegionSection *section)
415 {
416 MemoryRegion *mr = section->mr;
417
418 if (memory_region_is_ram(mr) && !memory_region_is_rom(mr)) {
419 uint8_t dirty_mask = memory_region_get_dirty_log_mask(mr);
420 uint8_t handled_dirty;
421
422 /*
423 * Kernel based vhost doesn't handle any block which is doing
424 * dirty-tracking other than migration for which it has
425 * specific logging support. However for TCG the kernel never
426 * gets involved anyway so we can also ignore it's
427 * self-modiying code detection flags. However a vhost-user
428 * client could still confuse a TCG guest if it re-writes
429 * executable memory that has already been translated.
430 */
431 handled_dirty = (1 << DIRTY_MEMORY_MIGRATION) |
432 (1 << DIRTY_MEMORY_CODE);
433
434 if (dirty_mask & ~handled_dirty) {
435 trace_vhost_reject_section(mr->name, 1);
436 return false;
437 }
438
439 if (dev->vhost_ops->vhost_backend_mem_section_filter &&
440 !dev->vhost_ops->vhost_backend_mem_section_filter(dev, section)) {
441 trace_vhost_reject_section(mr->name, 2);
442 return false;
443 }
444
445 trace_vhost_section(mr->name);
446 return true;
447 } else {
448 trace_vhost_reject_section(mr->name, 3);
449 return false;
450 }
451 }
452
453 static void vhost_begin(MemoryListener *listener)
454 {
455 struct vhost_dev *dev = container_of(listener, struct vhost_dev,
456 memory_listener);
457 dev->tmp_sections = NULL;
458 dev->n_tmp_sections = 0;
459 }
460
461 static void vhost_commit(MemoryListener *listener)
462 {
463 struct vhost_dev *dev = container_of(listener, struct vhost_dev,
464 memory_listener);
465 MemoryRegionSection *old_sections;
466 int n_old_sections;
467 uint64_t log_size;
468 size_t regions_size;
469 int r;
470 int i;
471 bool changed = false;
472
473 /* Note we can be called before the device is started, but then
474 * starting the device calls set_mem_table, so we need to have
475 * built the data structures.
476 */
477 old_sections = dev->mem_sections;
478 n_old_sections = dev->n_mem_sections;
479 dev->mem_sections = dev->tmp_sections;
480 dev->n_mem_sections = dev->n_tmp_sections;
481
482 if (dev->n_mem_sections != n_old_sections) {
483 changed = true;
484 } else {
485 /* Same size, lets check the contents */
486 for (int i = 0; i < n_old_sections; i++) {
487 if (!MemoryRegionSection_eq(&old_sections[i],
488 &dev->mem_sections[i])) {
489 changed = true;
490 break;
491 }
492 }
493 }
494
495 trace_vhost_commit(dev->started, changed);
496 if (!changed) {
497 goto out;
498 }
499
500 /* Rebuild the regions list from the new sections list */
501 regions_size = offsetof(struct vhost_memory, regions) +
502 dev->n_mem_sections * sizeof dev->mem->regions[0];
503 dev->mem = g_realloc(dev->mem, regions_size);
504 dev->mem->nregions = dev->n_mem_sections;
505 used_memslots = dev->mem->nregions;
506 for (i = 0; i < dev->n_mem_sections; i++) {
507 struct vhost_memory_region *cur_vmr = dev->mem->regions + i;
508 struct MemoryRegionSection *mrs = dev->mem_sections + i;
509
510 cur_vmr->guest_phys_addr = mrs->offset_within_address_space;
511 cur_vmr->memory_size = int128_get64(mrs->size);
512 cur_vmr->userspace_addr =
513 (uintptr_t)memory_region_get_ram_ptr(mrs->mr) +
514 mrs->offset_within_region;
515 cur_vmr->flags_padding = 0;
516 }
517
518 if (!dev->started) {
519 goto out;
520 }
521
522 for (i = 0; i < dev->mem->nregions; i++) {
523 if (vhost_verify_ring_mappings(dev,
524 (void *)(uintptr_t)dev->mem->regions[i].userspace_addr,
525 dev->mem->regions[i].guest_phys_addr,
526 dev->mem->regions[i].memory_size)) {
527 error_report("Verify ring failure on region %d", i);
528 abort();
529 }
530 }
531
532 if (!dev->log_enabled) {
533 r = dev->vhost_ops->vhost_set_mem_table(dev, dev->mem);
534 if (r < 0) {
535 VHOST_OPS_DEBUG("vhost_set_mem_table failed");
536 }
537 goto out;
538 }
539 log_size = vhost_get_log_size(dev);
540 /* We allocate an extra 4K bytes to log,
541 * to reduce the * number of reallocations. */
542 #define VHOST_LOG_BUFFER (0x1000 / sizeof *dev->log)
543 /* To log more, must increase log size before table update. */
544 if (dev->log_size < log_size) {
545 vhost_dev_log_resize(dev, log_size + VHOST_LOG_BUFFER);
546 }
547 r = dev->vhost_ops->vhost_set_mem_table(dev, dev->mem);
548 if (r < 0) {
549 VHOST_OPS_DEBUG("vhost_set_mem_table failed");
550 }
551 /* To log less, can only decrease log size after table update. */
552 if (dev->log_size > log_size + VHOST_LOG_BUFFER) {
553 vhost_dev_log_resize(dev, log_size);
554 }
555
556 out:
557 /* Deref the old list of sections, this must happen _after_ the
558 * vhost_set_mem_table to ensure the client isn't still using the
559 * section we're about to unref.
560 */
561 while (n_old_sections--) {
562 memory_region_unref(old_sections[n_old_sections].mr);
563 }
564 g_free(old_sections);
565 return;
566 }
567
568 /* Adds the section data to the tmp_section structure.
569 * It relies on the listener calling us in memory address order
570 * and for each region (via the _add and _nop methods) to
571 * join neighbours.
572 */
573 static void vhost_region_add_section(struct vhost_dev *dev,
574 MemoryRegionSection *section)
575 {
576 bool need_add = true;
577 uint64_t mrs_size = int128_get64(section->size);
578 uint64_t mrs_gpa = section->offset_within_address_space;
579 uintptr_t mrs_host = (uintptr_t)memory_region_get_ram_ptr(section->mr) +
580 section->offset_within_region;
581 RAMBlock *mrs_rb = section->mr->ram_block;
582
583 trace_vhost_region_add_section(section->mr->name, mrs_gpa, mrs_size,
584 mrs_host);
585
586 if (dev->vhost_ops->backend_type == VHOST_BACKEND_TYPE_USER) {
587 /* Round the section to it's page size */
588 /* First align the start down to a page boundary */
589 size_t mrs_page = qemu_ram_pagesize(mrs_rb);
590 uint64_t alignage = mrs_host & (mrs_page - 1);
591 if (alignage) {
592 mrs_host -= alignage;
593 mrs_size += alignage;
594 mrs_gpa -= alignage;
595 }
596 /* Now align the size up to a page boundary */
597 alignage = mrs_size & (mrs_page - 1);
598 if (alignage) {
599 mrs_size += mrs_page - alignage;
600 }
601 trace_vhost_region_add_section_aligned(section->mr->name, mrs_gpa,
602 mrs_size, mrs_host);
603 }
604
605 if (dev->n_tmp_sections) {
606 /* Since we already have at least one section, lets see if
607 * this extends it; since we're scanning in order, we only
608 * have to look at the last one, and the FlatView that calls
609 * us shouldn't have overlaps.
610 */
611 MemoryRegionSection *prev_sec = dev->tmp_sections +
612 (dev->n_tmp_sections - 1);
613 uint64_t prev_gpa_start = prev_sec->offset_within_address_space;
614 uint64_t prev_size = int128_get64(prev_sec->size);
615 uint64_t prev_gpa_end = range_get_last(prev_gpa_start, prev_size);
616 uint64_t prev_host_start =
617 (uintptr_t)memory_region_get_ram_ptr(prev_sec->mr) +
618 prev_sec->offset_within_region;
619 uint64_t prev_host_end = range_get_last(prev_host_start, prev_size);
620
621 if (mrs_gpa <= (prev_gpa_end + 1)) {
622 /* OK, looks like overlapping/intersecting - it's possible that
623 * the rounding to page sizes has made them overlap, but they should
624 * match up in the same RAMBlock if they do.
625 */
626 if (mrs_gpa < prev_gpa_start) {
627 error_report("%s:Section '%s' rounded to %"PRIx64
628 " prior to previous '%s' %"PRIx64,
629 __func__, section->mr->name, mrs_gpa,
630 prev_sec->mr->name, prev_gpa_start);
631 /* A way to cleanly fail here would be better */
632 return;
633 }
634 /* Offset from the start of the previous GPA to this GPA */
635 size_t offset = mrs_gpa - prev_gpa_start;
636
637 if (prev_host_start + offset == mrs_host &&
638 section->mr == prev_sec->mr &&
639 (!dev->vhost_ops->vhost_backend_can_merge ||
640 dev->vhost_ops->vhost_backend_can_merge(dev,
641 mrs_host, mrs_size,
642 prev_host_start, prev_size))) {
643 uint64_t max_end = MAX(prev_host_end, mrs_host + mrs_size);
644 need_add = false;
645 prev_sec->offset_within_address_space =
646 MIN(prev_gpa_start, mrs_gpa);
647 prev_sec->offset_within_region =
648 MIN(prev_host_start, mrs_host) -
649 (uintptr_t)memory_region_get_ram_ptr(prev_sec->mr);
650 prev_sec->size = int128_make64(max_end - MIN(prev_host_start,
651 mrs_host));
652 trace_vhost_region_add_section_merge(section->mr->name,
653 int128_get64(prev_sec->size),
654 prev_sec->offset_within_address_space,
655 prev_sec->offset_within_region);
656 } else {
657 /* adjoining regions are fine, but overlapping ones with
658 * different blocks/offsets shouldn't happen
659 */
660 if (mrs_gpa != prev_gpa_end + 1) {
661 error_report("%s: Overlapping but not coherent sections "
662 "at %"PRIx64,
663 __func__, mrs_gpa);
664 return;
665 }
666 }
667 }
668 }
669
670 if (need_add) {
671 ++dev->n_tmp_sections;
672 dev->tmp_sections = g_renew(MemoryRegionSection, dev->tmp_sections,
673 dev->n_tmp_sections);
674 dev->tmp_sections[dev->n_tmp_sections - 1] = *section;
675 /* The flatview isn't stable and we don't use it, making it NULL
676 * means we can memcmp the list.
677 */
678 dev->tmp_sections[dev->n_tmp_sections - 1].fv = NULL;
679 memory_region_ref(section->mr);
680 }
681 }
682
683 /* Used for both add and nop callbacks */
684 static void vhost_region_addnop(MemoryListener *listener,
685 MemoryRegionSection *section)
686 {
687 struct vhost_dev *dev = container_of(listener, struct vhost_dev,
688 memory_listener);
689
690 if (!vhost_section(dev, section)) {
691 return;
692 }
693 vhost_region_add_section(dev, section);
694 }
695
696 static void vhost_iommu_unmap_notify(IOMMUNotifier *n, IOMMUTLBEntry *iotlb)
697 {
698 struct vhost_iommu *iommu = container_of(n, struct vhost_iommu, n);
699 struct vhost_dev *hdev = iommu->hdev;
700 hwaddr iova = iotlb->iova + iommu->iommu_offset;
701
702 if (vhost_backend_invalidate_device_iotlb(hdev, iova,
703 iotlb->addr_mask + 1)) {
704 error_report("Fail to invalidate device iotlb");
705 }
706 }
707
708 static void vhost_iommu_region_add(MemoryListener *listener,
709 MemoryRegionSection *section)
710 {
711 struct vhost_dev *dev = container_of(listener, struct vhost_dev,
712 iommu_listener);
713 struct vhost_iommu *iommu;
714 Int128 end;
715 int iommu_idx, ret;
716 IOMMUMemoryRegion *iommu_mr;
717 Error *err = NULL;
718
719 if (!memory_region_is_iommu(section->mr)) {
720 return;
721 }
722
723 iommu_mr = IOMMU_MEMORY_REGION(section->mr);
724
725 iommu = g_malloc0(sizeof(*iommu));
726 end = int128_add(int128_make64(section->offset_within_region),
727 section->size);
728 end = int128_sub(end, int128_one());
729 iommu_idx = memory_region_iommu_attrs_to_index(iommu_mr,
730 MEMTXATTRS_UNSPECIFIED);
731 iommu_notifier_init(&iommu->n, vhost_iommu_unmap_notify,
732 IOMMU_NOTIFIER_UNMAP,
733 section->offset_within_region,
734 int128_get64(end),
735 iommu_idx);
736 iommu->mr = section->mr;
737 iommu->iommu_offset = section->offset_within_address_space -
738 section->offset_within_region;
739 iommu->hdev = dev;
740 ret = memory_region_register_iommu_notifier(section->mr, &iommu->n, &err);
741 if (ret) {
742 error_report_err(err);
743 exit(1);
744 }
745 QLIST_INSERT_HEAD(&dev->iommu_list, iommu, iommu_next);
746 /* TODO: can replay help performance here? */
747 }
748
749 static void vhost_iommu_region_del(MemoryListener *listener,
750 MemoryRegionSection *section)
751 {
752 struct vhost_dev *dev = container_of(listener, struct vhost_dev,
753 iommu_listener);
754 struct vhost_iommu *iommu;
755
756 if (!memory_region_is_iommu(section->mr)) {
757 return;
758 }
759
760 QLIST_FOREACH(iommu, &dev->iommu_list, iommu_next) {
761 if (iommu->mr == section->mr &&
762 iommu->n.start == section->offset_within_region) {
763 memory_region_unregister_iommu_notifier(iommu->mr,
764 &iommu->n);
765 QLIST_REMOVE(iommu, iommu_next);
766 g_free(iommu);
767 break;
768 }
769 }
770 }
771
772 static int vhost_virtqueue_set_addr(struct vhost_dev *dev,
773 struct vhost_virtqueue *vq,
774 unsigned idx, bool enable_log)
775 {
776 struct vhost_vring_addr addr;
777 int r;
778 memset(&addr, 0, sizeof(struct vhost_vring_addr));
779
780 if (dev->vhost_ops->vhost_vq_get_addr) {
781 r = dev->vhost_ops->vhost_vq_get_addr(dev, &addr, vq);
782 if (r < 0) {
783 VHOST_OPS_DEBUG("vhost_vq_get_addr failed");
784 return -errno;
785 }
786 } else {
787 addr.desc_user_addr = (uint64_t)(unsigned long)vq->desc;
788 addr.avail_user_addr = (uint64_t)(unsigned long)vq->avail;
789 addr.used_user_addr = (uint64_t)(unsigned long)vq->used;
790 }
791 addr.index = idx;
792 addr.log_guest_addr = vq->used_phys;
793 addr.flags = enable_log ? (1 << VHOST_VRING_F_LOG) : 0;
794 r = dev->vhost_ops->vhost_set_vring_addr(dev, &addr);
795 if (r < 0) {
796 VHOST_OPS_DEBUG("vhost_set_vring_addr failed");
797 return -errno;
798 }
799 return 0;
800 }
801
802 static int vhost_dev_set_features(struct vhost_dev *dev,
803 bool enable_log)
804 {
805 uint64_t features = dev->acked_features;
806 int r;
807 if (enable_log) {
808 features |= 0x1ULL << VHOST_F_LOG_ALL;
809 }
810 if (!vhost_dev_has_iommu(dev)) {
811 features &= ~(0x1ULL << VIRTIO_F_IOMMU_PLATFORM);
812 }
813 if (dev->vhost_ops->vhost_force_iommu) {
814 if (dev->vhost_ops->vhost_force_iommu(dev) == true) {
815 features |= 0x1ULL << VIRTIO_F_IOMMU_PLATFORM;
816 }
817 }
818 r = dev->vhost_ops->vhost_set_features(dev, features);
819 if (r < 0) {
820 VHOST_OPS_DEBUG("vhost_set_features failed");
821 goto out;
822 }
823 if (dev->vhost_ops->vhost_set_backend_cap) {
824 r = dev->vhost_ops->vhost_set_backend_cap(dev);
825 if (r < 0) {
826 VHOST_OPS_DEBUG("vhost_set_backend_cap failed");
827 goto out;
828 }
829 }
830
831 out:
832 return r < 0 ? -errno : 0;
833 }
834
835 static int vhost_dev_set_log(struct vhost_dev *dev, bool enable_log)
836 {
837 int r, i, idx;
838 r = vhost_dev_set_features(dev, enable_log);
839 if (r < 0) {
840 goto err_features;
841 }
842 for (i = 0; i < dev->nvqs; ++i) {
843 idx = dev->vhost_ops->vhost_get_vq_index(dev, dev->vq_index + i);
844 r = vhost_virtqueue_set_addr(dev, dev->vqs + i, idx,
845 enable_log);
846 if (r < 0) {
847 goto err_vq;
848 }
849 }
850 return 0;
851 err_vq:
852 for (; i >= 0; --i) {
853 idx = dev->vhost_ops->vhost_get_vq_index(dev, dev->vq_index + i);
854 vhost_virtqueue_set_addr(dev, dev->vqs + i, idx,
855 dev->log_enabled);
856 }
857 vhost_dev_set_features(dev, dev->log_enabled);
858 err_features:
859 return r;
860 }
861
862 static int vhost_migration_log(MemoryListener *listener, bool enable)
863 {
864 struct vhost_dev *dev = container_of(listener, struct vhost_dev,
865 memory_listener);
866 int r;
867 if (enable == dev->log_enabled) {
868 return 0;
869 }
870 if (!dev->started) {
871 dev->log_enabled = enable;
872 return 0;
873 }
874 if (!enable) {
875 r = vhost_dev_set_log(dev, false);
876 if (r < 0) {
877 return r;
878 }
879 vhost_log_put(dev, false);
880 } else {
881 vhost_dev_log_resize(dev, vhost_get_log_size(dev));
882 r = vhost_dev_set_log(dev, true);
883 if (r < 0) {
884 return r;
885 }
886 }
887 dev->log_enabled = enable;
888 return 0;
889 }
890
891 static void vhost_log_global_start(MemoryListener *listener)
892 {
893 int r;
894
895 r = vhost_migration_log(listener, true);
896 if (r < 0) {
897 abort();
898 }
899 }
900
901 static void vhost_log_global_stop(MemoryListener *listener)
902 {
903 int r;
904
905 r = vhost_migration_log(listener, false);
906 if (r < 0) {
907 abort();
908 }
909 }
910
911 static void vhost_log_start(MemoryListener *listener,
912 MemoryRegionSection *section,
913 int old, int new)
914 {
915 /* FIXME: implement */
916 }
917
918 static void vhost_log_stop(MemoryListener *listener,
919 MemoryRegionSection *section,
920 int old, int new)
921 {
922 /* FIXME: implement */
923 }
924
925 /* The vhost driver natively knows how to handle the vrings of non
926 * cross-endian legacy devices and modern devices. Only legacy devices
927 * exposed to a bi-endian guest may require the vhost driver to use a
928 * specific endianness.
929 */
930 static inline bool vhost_needs_vring_endian(VirtIODevice *vdev)
931 {
932 if (virtio_vdev_has_feature(vdev, VIRTIO_F_VERSION_1)) {
933 return false;
934 }
935 #ifdef HOST_WORDS_BIGENDIAN
936 return vdev->device_endian == VIRTIO_DEVICE_ENDIAN_LITTLE;
937 #else
938 return vdev->device_endian == VIRTIO_DEVICE_ENDIAN_BIG;
939 #endif
940 }
941
942 static int vhost_virtqueue_set_vring_endian_legacy(struct vhost_dev *dev,
943 bool is_big_endian,
944 int vhost_vq_index)
945 {
946 struct vhost_vring_state s = {
947 .index = vhost_vq_index,
948 .num = is_big_endian
949 };
950
951 if (!dev->vhost_ops->vhost_set_vring_endian(dev, &s)) {
952 return 0;
953 }
954
955 VHOST_OPS_DEBUG("vhost_set_vring_endian failed");
956 if (errno == ENOTTY) {
957 error_report("vhost does not support cross-endian");
958 return -ENOSYS;
959 }
960
961 return -errno;
962 }
963
964 static int vhost_memory_region_lookup(struct vhost_dev *hdev,
965 uint64_t gpa, uint64_t *uaddr,
966 uint64_t *len)
967 {
968 int i;
969
970 for (i = 0; i < hdev->mem->nregions; i++) {
971 struct vhost_memory_region *reg = hdev->mem->regions + i;
972
973 if (gpa >= reg->guest_phys_addr &&
974 reg->guest_phys_addr + reg->memory_size > gpa) {
975 *uaddr = reg->userspace_addr + gpa - reg->guest_phys_addr;
976 *len = reg->guest_phys_addr + reg->memory_size - gpa;
977 return 0;
978 }
979 }
980
981 return -EFAULT;
982 }
983
984 int vhost_device_iotlb_miss(struct vhost_dev *dev, uint64_t iova, int write)
985 {
986 IOMMUTLBEntry iotlb;
987 uint64_t uaddr, len;
988 int ret = -EFAULT;
989
990 RCU_READ_LOCK_GUARD();
991
992 trace_vhost_iotlb_miss(dev, 1);
993
994 iotlb = address_space_get_iotlb_entry(dev->vdev->dma_as,
995 iova, write,
996 MEMTXATTRS_UNSPECIFIED);
997 if (iotlb.target_as != NULL) {
998 ret = vhost_memory_region_lookup(dev, iotlb.translated_addr,
999 &uaddr, &len);
1000 if (ret) {
1001 trace_vhost_iotlb_miss(dev, 3);
1002 error_report("Fail to lookup the translated address "
1003 "%"PRIx64, iotlb.translated_addr);
1004 goto out;
1005 }
1006
1007 len = MIN(iotlb.addr_mask + 1, len);
1008 iova = iova & ~iotlb.addr_mask;
1009
1010 ret = vhost_backend_update_device_iotlb(dev, iova, uaddr,
1011 len, iotlb.perm);
1012 if (ret) {
1013 trace_vhost_iotlb_miss(dev, 4);
1014 error_report("Fail to update device iotlb");
1015 goto out;
1016 }
1017 }
1018
1019 trace_vhost_iotlb_miss(dev, 2);
1020
1021 out:
1022 return ret;
1023 }
1024
1025 static int vhost_virtqueue_start(struct vhost_dev *dev,
1026 struct VirtIODevice *vdev,
1027 struct vhost_virtqueue *vq,
1028 unsigned idx)
1029 {
1030 BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev)));
1031 VirtioBusState *vbus = VIRTIO_BUS(qbus);
1032 VirtioBusClass *k = VIRTIO_BUS_GET_CLASS(vbus);
1033 hwaddr s, l, a;
1034 int r;
1035 int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, idx);
1036 struct vhost_vring_file file = {
1037 .index = vhost_vq_index
1038 };
1039 struct vhost_vring_state state = {
1040 .index = vhost_vq_index
1041 };
1042 struct VirtQueue *vvq = virtio_get_queue(vdev, idx);
1043
1044 a = virtio_queue_get_desc_addr(vdev, idx);
1045 if (a == 0) {
1046 /* Queue might not be ready for start */
1047 return 0;
1048 }
1049
1050 vq->num = state.num = virtio_queue_get_num(vdev, idx);
1051 r = dev->vhost_ops->vhost_set_vring_num(dev, &state);
1052 if (r) {
1053 VHOST_OPS_DEBUG("vhost_set_vring_num failed");
1054 return -errno;
1055 }
1056
1057 state.num = virtio_queue_get_last_avail_idx(vdev, idx);
1058 r = dev->vhost_ops->vhost_set_vring_base(dev, &state);
1059 if (r) {
1060 VHOST_OPS_DEBUG("vhost_set_vring_base failed");
1061 return -errno;
1062 }
1063
1064 if (vhost_needs_vring_endian(vdev)) {
1065 r = vhost_virtqueue_set_vring_endian_legacy(dev,
1066 virtio_is_big_endian(vdev),
1067 vhost_vq_index);
1068 if (r) {
1069 return -errno;
1070 }
1071 }
1072
1073 vq->desc_size = s = l = virtio_queue_get_desc_size(vdev, idx);
1074 vq->desc_phys = a;
1075 vq->desc = vhost_memory_map(dev, a, &l, false);
1076 if (!vq->desc || l != s) {
1077 r = -ENOMEM;
1078 goto fail_alloc_desc;
1079 }
1080 vq->avail_size = s = l = virtio_queue_get_avail_size(vdev, idx);
1081 vq->avail_phys = a = virtio_queue_get_avail_addr(vdev, idx);
1082 vq->avail = vhost_memory_map(dev, a, &l, false);
1083 if (!vq->avail || l != s) {
1084 r = -ENOMEM;
1085 goto fail_alloc_avail;
1086 }
1087 vq->used_size = s = l = virtio_queue_get_used_size(vdev, idx);
1088 vq->used_phys = a = virtio_queue_get_used_addr(vdev, idx);
1089 vq->used = vhost_memory_map(dev, a, &l, true);
1090 if (!vq->used || l != s) {
1091 r = -ENOMEM;
1092 goto fail_alloc_used;
1093 }
1094
1095 r = vhost_virtqueue_set_addr(dev, vq, vhost_vq_index, dev->log_enabled);
1096 if (r < 0) {
1097 r = -errno;
1098 goto fail_alloc;
1099 }
1100
1101 file.fd = event_notifier_get_fd(virtio_queue_get_host_notifier(vvq));
1102 r = dev->vhost_ops->vhost_set_vring_kick(dev, &file);
1103 if (r) {
1104 VHOST_OPS_DEBUG("vhost_set_vring_kick failed");
1105 r = -errno;
1106 goto fail_kick;
1107 }
1108
1109 /* Clear and discard previous events if any. */
1110 event_notifier_test_and_clear(&vq->masked_notifier);
1111
1112 /* Init vring in unmasked state, unless guest_notifier_mask
1113 * will do it later.
1114 */
1115 if (!vdev->use_guest_notifier_mask) {
1116 /* TODO: check and handle errors. */
1117 vhost_virtqueue_mask(dev, vdev, idx, false);
1118 }
1119
1120 if (k->query_guest_notifiers &&
1121 k->query_guest_notifiers(qbus->parent) &&
1122 virtio_queue_vector(vdev, idx) == VIRTIO_NO_VECTOR) {
1123 file.fd = -1;
1124 r = dev->vhost_ops->vhost_set_vring_call(dev, &file);
1125 if (r) {
1126 goto fail_vector;
1127 }
1128 }
1129
1130 return 0;
1131
1132 fail_vector:
1133 fail_kick:
1134 fail_alloc:
1135 vhost_memory_unmap(dev, vq->used, virtio_queue_get_used_size(vdev, idx),
1136 0, 0);
1137 fail_alloc_used:
1138 vhost_memory_unmap(dev, vq->avail, virtio_queue_get_avail_size(vdev, idx),
1139 0, 0);
1140 fail_alloc_avail:
1141 vhost_memory_unmap(dev, vq->desc, virtio_queue_get_desc_size(vdev, idx),
1142 0, 0);
1143 fail_alloc_desc:
1144 return r;
1145 }
1146
1147 static void vhost_virtqueue_stop(struct vhost_dev *dev,
1148 struct VirtIODevice *vdev,
1149 struct vhost_virtqueue *vq,
1150 unsigned idx)
1151 {
1152 int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, idx);
1153 struct vhost_vring_state state = {
1154 .index = vhost_vq_index,
1155 };
1156 int r;
1157
1158 if (virtio_queue_get_desc_addr(vdev, idx) == 0) {
1159 /* Don't stop the virtqueue which might have not been started */
1160 return;
1161 }
1162
1163 r = dev->vhost_ops->vhost_get_vring_base(dev, &state);
1164 if (r < 0) {
1165 VHOST_OPS_DEBUG("vhost VQ %u ring restore failed: %d", idx, r);
1166 /* Connection to the backend is broken, so let's sync internal
1167 * last avail idx to the device used idx.
1168 */
1169 virtio_queue_restore_last_avail_idx(vdev, idx);
1170 } else {
1171 virtio_queue_set_last_avail_idx(vdev, idx, state.num);
1172 }
1173 virtio_queue_invalidate_signalled_used(vdev, idx);
1174 virtio_queue_update_used_idx(vdev, idx);
1175
1176 /* In the cross-endian case, we need to reset the vring endianness to
1177 * native as legacy devices expect so by default.
1178 */
1179 if (vhost_needs_vring_endian(vdev)) {
1180 vhost_virtqueue_set_vring_endian_legacy(dev,
1181 !virtio_is_big_endian(vdev),
1182 vhost_vq_index);
1183 }
1184
1185 vhost_memory_unmap(dev, vq->used, virtio_queue_get_used_size(vdev, idx),
1186 1, virtio_queue_get_used_size(vdev, idx));
1187 vhost_memory_unmap(dev, vq->avail, virtio_queue_get_avail_size(vdev, idx),
1188 0, virtio_queue_get_avail_size(vdev, idx));
1189 vhost_memory_unmap(dev, vq->desc, virtio_queue_get_desc_size(vdev, idx),
1190 0, virtio_queue_get_desc_size(vdev, idx));
1191 }
1192
1193 static void vhost_eventfd_add(MemoryListener *listener,
1194 MemoryRegionSection *section,
1195 bool match_data, uint64_t data, EventNotifier *e)
1196 {
1197 }
1198
1199 static void vhost_eventfd_del(MemoryListener *listener,
1200 MemoryRegionSection *section,
1201 bool match_data, uint64_t data, EventNotifier *e)
1202 {
1203 }
1204
1205 static int vhost_virtqueue_set_busyloop_timeout(struct vhost_dev *dev,
1206 int n, uint32_t timeout)
1207 {
1208 int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, n);
1209 struct vhost_vring_state state = {
1210 .index = vhost_vq_index,
1211 .num = timeout,
1212 };
1213 int r;
1214
1215 if (!dev->vhost_ops->vhost_set_vring_busyloop_timeout) {
1216 return -EINVAL;
1217 }
1218
1219 r = dev->vhost_ops->vhost_set_vring_busyloop_timeout(dev, &state);
1220 if (r) {
1221 VHOST_OPS_DEBUG("vhost_set_vring_busyloop_timeout failed");
1222 return r;
1223 }
1224
1225 return 0;
1226 }
1227
1228 static int vhost_virtqueue_init(struct vhost_dev *dev,
1229 struct vhost_virtqueue *vq, int n)
1230 {
1231 int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, n);
1232 struct vhost_vring_file file = {
1233 .index = vhost_vq_index,
1234 };
1235 int r = event_notifier_init(&vq->masked_notifier, 0);
1236 if (r < 0) {
1237 return r;
1238 }
1239
1240 file.fd = event_notifier_get_fd(&vq->masked_notifier);
1241 r = dev->vhost_ops->vhost_set_vring_call(dev, &file);
1242 if (r) {
1243 VHOST_OPS_DEBUG("vhost_set_vring_call failed");
1244 r = -errno;
1245 goto fail_call;
1246 }
1247
1248 vq->dev = dev;
1249
1250 return 0;
1251 fail_call:
1252 event_notifier_cleanup(&vq->masked_notifier);
1253 return r;
1254 }
1255
1256 static void vhost_virtqueue_cleanup(struct vhost_virtqueue *vq)
1257 {
1258 event_notifier_cleanup(&vq->masked_notifier);
1259 }
1260
1261 int vhost_dev_init(struct vhost_dev *hdev, void *opaque,
1262 VhostBackendType backend_type, uint32_t busyloop_timeout)
1263 {
1264 uint64_t features;
1265 int i, r, n_initialized_vqs = 0;
1266 Error *local_err = NULL;
1267
1268 hdev->vdev = NULL;
1269 hdev->migration_blocker = NULL;
1270
1271 r = vhost_set_backend_type(hdev, backend_type);
1272 assert(r >= 0);
1273
1274 r = hdev->vhost_ops->vhost_backend_init(hdev, opaque);
1275 if (r < 0) {
1276 goto fail;
1277 }
1278
1279 r = hdev->vhost_ops->vhost_set_owner(hdev);
1280 if (r < 0) {
1281 VHOST_OPS_DEBUG("vhost_set_owner failed");
1282 goto fail;
1283 }
1284
1285 r = hdev->vhost_ops->vhost_get_features(hdev, &features);
1286 if (r < 0) {
1287 VHOST_OPS_DEBUG("vhost_get_features failed");
1288 goto fail;
1289 }
1290
1291 for (i = 0; i < hdev->nvqs; ++i, ++n_initialized_vqs) {
1292 r = vhost_virtqueue_init(hdev, hdev->vqs + i, hdev->vq_index + i);
1293 if (r < 0) {
1294 goto fail;
1295 }
1296 }
1297
1298 if (busyloop_timeout) {
1299 for (i = 0; i < hdev->nvqs; ++i) {
1300 r = vhost_virtqueue_set_busyloop_timeout(hdev, hdev->vq_index + i,
1301 busyloop_timeout);
1302 if (r < 0) {
1303 goto fail_busyloop;
1304 }
1305 }
1306 }
1307
1308 hdev->features = features;
1309
1310 hdev->memory_listener = (MemoryListener) {
1311 .begin = vhost_begin,
1312 .commit = vhost_commit,
1313 .region_add = vhost_region_addnop,
1314 .region_nop = vhost_region_addnop,
1315 .log_start = vhost_log_start,
1316 .log_stop = vhost_log_stop,
1317 .log_sync = vhost_log_sync,
1318 .log_global_start = vhost_log_global_start,
1319 .log_global_stop = vhost_log_global_stop,
1320 .eventfd_add = vhost_eventfd_add,
1321 .eventfd_del = vhost_eventfd_del,
1322 .priority = 10
1323 };
1324
1325 hdev->iommu_listener = (MemoryListener) {
1326 .region_add = vhost_iommu_region_add,
1327 .region_del = vhost_iommu_region_del,
1328 };
1329
1330 if (hdev->migration_blocker == NULL) {
1331 if (!(hdev->features & (0x1ULL << VHOST_F_LOG_ALL))) {
1332 error_setg(&hdev->migration_blocker,
1333 "Migration disabled: vhost lacks VHOST_F_LOG_ALL feature.");
1334 } else if (vhost_dev_log_is_shared(hdev) && !qemu_memfd_alloc_check()) {
1335 error_setg(&hdev->migration_blocker,
1336 "Migration disabled: failed to allocate shared memory");
1337 }
1338 }
1339
1340 if (hdev->migration_blocker != NULL) {
1341 r = migrate_add_blocker(hdev->migration_blocker, &local_err);
1342 if (local_err) {
1343 error_report_err(local_err);
1344 error_free(hdev->migration_blocker);
1345 goto fail_busyloop;
1346 }
1347 }
1348
1349 hdev->mem = g_malloc0(offsetof(struct vhost_memory, regions));
1350 hdev->n_mem_sections = 0;
1351 hdev->mem_sections = NULL;
1352 hdev->log = NULL;
1353 hdev->log_size = 0;
1354 hdev->log_enabled = false;
1355 hdev->started = false;
1356 memory_listener_register(&hdev->memory_listener, &address_space_memory);
1357 QLIST_INSERT_HEAD(&vhost_devices, hdev, entry);
1358
1359 if (used_memslots > hdev->vhost_ops->vhost_backend_memslots_limit(hdev)) {
1360 error_report("vhost backend memory slots limit is less"
1361 " than current number of present memory slots");
1362 r = -1;
1363 if (busyloop_timeout) {
1364 goto fail_busyloop;
1365 } else {
1366 goto fail;
1367 }
1368 }
1369
1370 return 0;
1371
1372 fail_busyloop:
1373 while (--i >= 0) {
1374 vhost_virtqueue_set_busyloop_timeout(hdev, hdev->vq_index + i, 0);
1375 }
1376 fail:
1377 hdev->nvqs = n_initialized_vqs;
1378 vhost_dev_cleanup(hdev);
1379 return r;
1380 }
1381
1382 void vhost_dev_cleanup(struct vhost_dev *hdev)
1383 {
1384 int i;
1385
1386 for (i = 0; i < hdev->nvqs; ++i) {
1387 vhost_virtqueue_cleanup(hdev->vqs + i);
1388 }
1389 if (hdev->mem) {
1390 /* those are only safe after successful init */
1391 memory_listener_unregister(&hdev->memory_listener);
1392 QLIST_REMOVE(hdev, entry);
1393 }
1394 if (hdev->migration_blocker) {
1395 migrate_del_blocker(hdev->migration_blocker);
1396 error_free(hdev->migration_blocker);
1397 }
1398 g_free(hdev->mem);
1399 g_free(hdev->mem_sections);
1400 if (hdev->vhost_ops) {
1401 hdev->vhost_ops->vhost_backend_cleanup(hdev);
1402 }
1403 assert(!hdev->log);
1404
1405 memset(hdev, 0, sizeof(struct vhost_dev));
1406 }
1407
1408 /* Stop processing guest IO notifications in qemu.
1409 * Start processing them in vhost in kernel.
1410 */
1411 int vhost_dev_enable_notifiers(struct vhost_dev *hdev, VirtIODevice *vdev)
1412 {
1413 BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev)));
1414 int i, r, e;
1415
1416 /* We will pass the notifiers to the kernel, make sure that QEMU
1417 * doesn't interfere.
1418 */
1419 r = virtio_device_grab_ioeventfd(vdev);
1420 if (r < 0) {
1421 error_report("binding does not support host notifiers");
1422 goto fail;
1423 }
1424
1425 for (i = 0; i < hdev->nvqs; ++i) {
1426 r = virtio_bus_set_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i,
1427 true);
1428 if (r < 0) {
1429 error_report("vhost VQ %d notifier binding failed: %d", i, -r);
1430 goto fail_vq;
1431 }
1432 }
1433
1434 return 0;
1435 fail_vq:
1436 while (--i >= 0) {
1437 e = virtio_bus_set_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i,
1438 false);
1439 if (e < 0) {
1440 error_report("vhost VQ %d notifier cleanup error: %d", i, -r);
1441 }
1442 assert (e >= 0);
1443 virtio_bus_cleanup_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i);
1444 }
1445 virtio_device_release_ioeventfd(vdev);
1446 fail:
1447 return r;
1448 }
1449
1450 /* Stop processing guest IO notifications in vhost.
1451 * Start processing them in qemu.
1452 * This might actually run the qemu handlers right away,
1453 * so virtio in qemu must be completely setup when this is called.
1454 */
1455 void vhost_dev_disable_notifiers(struct vhost_dev *hdev, VirtIODevice *vdev)
1456 {
1457 BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev)));
1458 int i, r;
1459
1460 for (i = 0; i < hdev->nvqs; ++i) {
1461 r = virtio_bus_set_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i,
1462 false);
1463 if (r < 0) {
1464 error_report("vhost VQ %d notifier cleanup failed: %d", i, -r);
1465 }
1466 assert (r >= 0);
1467 virtio_bus_cleanup_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i);
1468 }
1469 virtio_device_release_ioeventfd(vdev);
1470 }
1471
1472 /* Test and clear event pending status.
1473 * Should be called after unmask to avoid losing events.
1474 */
1475 bool vhost_virtqueue_pending(struct vhost_dev *hdev, int n)
1476 {
1477 struct vhost_virtqueue *vq = hdev->vqs + n - hdev->vq_index;
1478 assert(n >= hdev->vq_index && n < hdev->vq_index + hdev->nvqs);
1479 return event_notifier_test_and_clear(&vq->masked_notifier);
1480 }
1481
1482 /* Mask/unmask events from this vq. */
1483 void vhost_virtqueue_mask(struct vhost_dev *hdev, VirtIODevice *vdev, int n,
1484 bool mask)
1485 {
1486 struct VirtQueue *vvq = virtio_get_queue(vdev, n);
1487 int r, index = n - hdev->vq_index;
1488 struct vhost_vring_file file;
1489
1490 /* should only be called after backend is connected */
1491 assert(hdev->vhost_ops);
1492
1493 if (mask) {
1494 assert(vdev->use_guest_notifier_mask);
1495 file.fd = event_notifier_get_fd(&hdev->vqs[index].masked_notifier);
1496 } else {
1497 file.fd = event_notifier_get_fd(virtio_queue_get_guest_notifier(vvq));
1498 }
1499
1500 file.index = hdev->vhost_ops->vhost_get_vq_index(hdev, n);
1501 r = hdev->vhost_ops->vhost_set_vring_call(hdev, &file);
1502 if (r < 0) {
1503 VHOST_OPS_DEBUG("vhost_set_vring_call failed");
1504 }
1505 }
1506
1507 uint64_t vhost_get_features(struct vhost_dev *hdev, const int *feature_bits,
1508 uint64_t features)
1509 {
1510 const int *bit = feature_bits;
1511 while (*bit != VHOST_INVALID_FEATURE_BIT) {
1512 uint64_t bit_mask = (1ULL << *bit);
1513 if (!(hdev->features & bit_mask)) {
1514 features &= ~bit_mask;
1515 }
1516 bit++;
1517 }
1518 return features;
1519 }
1520
1521 void vhost_ack_features(struct vhost_dev *hdev, const int *feature_bits,
1522 uint64_t features)
1523 {
1524 const int *bit = feature_bits;
1525 while (*bit != VHOST_INVALID_FEATURE_BIT) {
1526 uint64_t bit_mask = (1ULL << *bit);
1527 if (features & bit_mask) {
1528 hdev->acked_features |= bit_mask;
1529 }
1530 bit++;
1531 }
1532 }
1533
1534 int vhost_dev_get_config(struct vhost_dev *hdev, uint8_t *config,
1535 uint32_t config_len)
1536 {
1537 assert(hdev->vhost_ops);
1538
1539 if (hdev->vhost_ops->vhost_get_config) {
1540 return hdev->vhost_ops->vhost_get_config(hdev, config, config_len);
1541 }
1542
1543 return -1;
1544 }
1545
1546 int vhost_dev_set_config(struct vhost_dev *hdev, const uint8_t *data,
1547 uint32_t offset, uint32_t size, uint32_t flags)
1548 {
1549 assert(hdev->vhost_ops);
1550
1551 if (hdev->vhost_ops->vhost_set_config) {
1552 return hdev->vhost_ops->vhost_set_config(hdev, data, offset,
1553 size, flags);
1554 }
1555
1556 return -1;
1557 }
1558
1559 void vhost_dev_set_config_notifier(struct vhost_dev *hdev,
1560 const VhostDevConfigOps *ops)
1561 {
1562 hdev->config_ops = ops;
1563 }
1564
1565 void vhost_dev_free_inflight(struct vhost_inflight *inflight)
1566 {
1567 if (inflight && inflight->addr) {
1568 qemu_memfd_free(inflight->addr, inflight->size, inflight->fd);
1569 inflight->addr = NULL;
1570 inflight->fd = -1;
1571 }
1572 }
1573
1574 static int vhost_dev_resize_inflight(struct vhost_inflight *inflight,
1575 uint64_t new_size)
1576 {
1577 Error *err = NULL;
1578 int fd = -1;
1579 void *addr = qemu_memfd_alloc("vhost-inflight", new_size,
1580 F_SEAL_GROW | F_SEAL_SHRINK | F_SEAL_SEAL,
1581 &fd, &err);
1582
1583 if (err) {
1584 error_report_err(err);
1585 return -1;
1586 }
1587
1588 vhost_dev_free_inflight(inflight);
1589 inflight->offset = 0;
1590 inflight->addr = addr;
1591 inflight->fd = fd;
1592 inflight->size = new_size;
1593
1594 return 0;
1595 }
1596
1597 void vhost_dev_save_inflight(struct vhost_inflight *inflight, QEMUFile *f)
1598 {
1599 if (inflight->addr) {
1600 qemu_put_be64(f, inflight->size);
1601 qemu_put_be16(f, inflight->queue_size);
1602 qemu_put_buffer(f, inflight->addr, inflight->size);
1603 } else {
1604 qemu_put_be64(f, 0);
1605 }
1606 }
1607
1608 int vhost_dev_load_inflight(struct vhost_inflight *inflight, QEMUFile *f)
1609 {
1610 uint64_t size;
1611
1612 size = qemu_get_be64(f);
1613 if (!size) {
1614 return 0;
1615 }
1616
1617 if (inflight->size != size) {
1618 if (vhost_dev_resize_inflight(inflight, size)) {
1619 return -1;
1620 }
1621 }
1622 inflight->queue_size = qemu_get_be16(f);
1623
1624 qemu_get_buffer(f, inflight->addr, size);
1625
1626 return 0;
1627 }
1628
1629 int vhost_dev_set_inflight(struct vhost_dev *dev,
1630 struct vhost_inflight *inflight)
1631 {
1632 int r;
1633
1634 if (dev->vhost_ops->vhost_set_inflight_fd && inflight->addr) {
1635 r = dev->vhost_ops->vhost_set_inflight_fd(dev, inflight);
1636 if (r) {
1637 VHOST_OPS_DEBUG("vhost_set_inflight_fd failed");
1638 return -errno;
1639 }
1640 }
1641
1642 return 0;
1643 }
1644
1645 int vhost_dev_get_inflight(struct vhost_dev *dev, uint16_t queue_size,
1646 struct vhost_inflight *inflight)
1647 {
1648 int r;
1649
1650 if (dev->vhost_ops->vhost_get_inflight_fd) {
1651 r = dev->vhost_ops->vhost_get_inflight_fd(dev, queue_size, inflight);
1652 if (r) {
1653 VHOST_OPS_DEBUG("vhost_get_inflight_fd failed");
1654 return -errno;
1655 }
1656 }
1657
1658 return 0;
1659 }
1660
1661 /* Host notifiers must be enabled at this point. */
1662 int vhost_dev_start(struct vhost_dev *hdev, VirtIODevice *vdev)
1663 {
1664 int i, r;
1665
1666 /* should only be called after backend is connected */
1667 assert(hdev->vhost_ops);
1668
1669 hdev->started = true;
1670 hdev->vdev = vdev;
1671
1672 r = vhost_dev_set_features(hdev, hdev->log_enabled);
1673 if (r < 0) {
1674 goto fail_features;
1675 }
1676
1677 if (vhost_dev_has_iommu(hdev)) {
1678 memory_listener_register(&hdev->iommu_listener, vdev->dma_as);
1679 }
1680
1681 r = hdev->vhost_ops->vhost_set_mem_table(hdev, hdev->mem);
1682 if (r < 0) {
1683 VHOST_OPS_DEBUG("vhost_set_mem_table failed");
1684 r = -errno;
1685 goto fail_mem;
1686 }
1687 for (i = 0; i < hdev->nvqs; ++i) {
1688 r = vhost_virtqueue_start(hdev,
1689 vdev,
1690 hdev->vqs + i,
1691 hdev->vq_index + i);
1692 if (r < 0) {
1693 goto fail_vq;
1694 }
1695 }
1696
1697 if (hdev->log_enabled) {
1698 uint64_t log_base;
1699
1700 hdev->log_size = vhost_get_log_size(hdev);
1701 hdev->log = vhost_log_get(hdev->log_size,
1702 vhost_dev_log_is_shared(hdev));
1703 log_base = (uintptr_t)hdev->log->log;
1704 r = hdev->vhost_ops->vhost_set_log_base(hdev,
1705 hdev->log_size ? log_base : 0,
1706 hdev->log);
1707 if (r < 0) {
1708 VHOST_OPS_DEBUG("vhost_set_log_base failed");
1709 r = -errno;
1710 goto fail_log;
1711 }
1712 }
1713 if (hdev->vhost_ops->vhost_dev_start) {
1714 r = hdev->vhost_ops->vhost_dev_start(hdev, true);
1715 if (r) {
1716 goto fail_log;
1717 }
1718 }
1719 if (vhost_dev_has_iommu(hdev) &&
1720 hdev->vhost_ops->vhost_set_iotlb_callback) {
1721 hdev->vhost_ops->vhost_set_iotlb_callback(hdev, true);
1722
1723 /* Update used ring information for IOTLB to work correctly,
1724 * vhost-kernel code requires for this.*/
1725 for (i = 0; i < hdev->nvqs; ++i) {
1726 struct vhost_virtqueue *vq = hdev->vqs + i;
1727 vhost_device_iotlb_miss(hdev, vq->used_phys, true);
1728 }
1729 }
1730 return 0;
1731 fail_log:
1732 vhost_log_put(hdev, false);
1733 fail_vq:
1734 while (--i >= 0) {
1735 vhost_virtqueue_stop(hdev,
1736 vdev,
1737 hdev->vqs + i,
1738 hdev->vq_index + i);
1739 }
1740
1741 fail_mem:
1742 fail_features:
1743
1744 hdev->started = false;
1745 return r;
1746 }
1747
1748 /* Host notifiers must be enabled at this point. */
1749 void vhost_dev_stop(struct vhost_dev *hdev, VirtIODevice *vdev)
1750 {
1751 int i;
1752
1753 /* should only be called after backend is connected */
1754 assert(hdev->vhost_ops);
1755
1756 if (hdev->vhost_ops->vhost_dev_start) {
1757 hdev->vhost_ops->vhost_dev_start(hdev, false);
1758 }
1759 for (i = 0; i < hdev->nvqs; ++i) {
1760 vhost_virtqueue_stop(hdev,
1761 vdev,
1762 hdev->vqs + i,
1763 hdev->vq_index + i);
1764 }
1765
1766 if (vhost_dev_has_iommu(hdev)) {
1767 if (hdev->vhost_ops->vhost_set_iotlb_callback) {
1768 hdev->vhost_ops->vhost_set_iotlb_callback(hdev, false);
1769 }
1770 memory_listener_unregister(&hdev->iommu_listener);
1771 }
1772 vhost_log_put(hdev, true);
1773 hdev->started = false;
1774 hdev->vdev = NULL;
1775 }
1776
1777 int vhost_net_set_backend(struct vhost_dev *hdev,
1778 struct vhost_vring_file *file)
1779 {
1780 if (hdev->vhost_ops->vhost_net_set_backend) {
1781 return hdev->vhost_ops->vhost_net_set_backend(hdev, file);
1782 }
1783
1784 return -1;
1785 }