hw/arm/raspi: fix CPRMAN base address
[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;
716 IOMMUMemoryRegion *iommu_mr;
717
718 if (!memory_region_is_iommu(section->mr)) {
719 return;
720 }
721
722 iommu_mr = IOMMU_MEMORY_REGION(section->mr);
723
724 iommu = g_malloc0(sizeof(*iommu));
725 end = int128_add(int128_make64(section->offset_within_region),
726 section->size);
727 end = int128_sub(end, int128_one());
728 iommu_idx = memory_region_iommu_attrs_to_index(iommu_mr,
729 MEMTXATTRS_UNSPECIFIED);
730 iommu_notifier_init(&iommu->n, vhost_iommu_unmap_notify,
731 IOMMU_NOTIFIER_UNMAP,
732 section->offset_within_region,
733 int128_get64(end),
734 iommu_idx);
735 iommu->mr = section->mr;
736 iommu->iommu_offset = section->offset_within_address_space -
737 section->offset_within_region;
738 iommu->hdev = dev;
739 memory_region_register_iommu_notifier(section->mr, &iommu->n,
740 &error_fatal);
741 QLIST_INSERT_HEAD(&dev->iommu_list, iommu, iommu_next);
742 /* TODO: can replay help performance here? */
743 }
744
745 static void vhost_iommu_region_del(MemoryListener *listener,
746 MemoryRegionSection *section)
747 {
748 struct vhost_dev *dev = container_of(listener, struct vhost_dev,
749 iommu_listener);
750 struct vhost_iommu *iommu;
751
752 if (!memory_region_is_iommu(section->mr)) {
753 return;
754 }
755
756 QLIST_FOREACH(iommu, &dev->iommu_list, iommu_next) {
757 if (iommu->mr == section->mr &&
758 iommu->n.start == section->offset_within_region) {
759 memory_region_unregister_iommu_notifier(iommu->mr,
760 &iommu->n);
761 QLIST_REMOVE(iommu, iommu_next);
762 g_free(iommu);
763 break;
764 }
765 }
766 }
767
768 static int vhost_virtqueue_set_addr(struct vhost_dev *dev,
769 struct vhost_virtqueue *vq,
770 unsigned idx, bool enable_log)
771 {
772 struct vhost_vring_addr addr;
773 int r;
774 memset(&addr, 0, sizeof(struct vhost_vring_addr));
775
776 if (dev->vhost_ops->vhost_vq_get_addr) {
777 r = dev->vhost_ops->vhost_vq_get_addr(dev, &addr, vq);
778 if (r < 0) {
779 VHOST_OPS_DEBUG("vhost_vq_get_addr failed");
780 return -errno;
781 }
782 } else {
783 addr.desc_user_addr = (uint64_t)(unsigned long)vq->desc;
784 addr.avail_user_addr = (uint64_t)(unsigned long)vq->avail;
785 addr.used_user_addr = (uint64_t)(unsigned long)vq->used;
786 }
787 addr.index = idx;
788 addr.log_guest_addr = vq->used_phys;
789 addr.flags = enable_log ? (1 << VHOST_VRING_F_LOG) : 0;
790 r = dev->vhost_ops->vhost_set_vring_addr(dev, &addr);
791 if (r < 0) {
792 VHOST_OPS_DEBUG("vhost_set_vring_addr failed");
793 return -errno;
794 }
795 return 0;
796 }
797
798 static int vhost_dev_set_features(struct vhost_dev *dev,
799 bool enable_log)
800 {
801 uint64_t features = dev->acked_features;
802 int r;
803 if (enable_log) {
804 features |= 0x1ULL << VHOST_F_LOG_ALL;
805 }
806 if (!vhost_dev_has_iommu(dev)) {
807 features &= ~(0x1ULL << VIRTIO_F_IOMMU_PLATFORM);
808 }
809 if (dev->vhost_ops->vhost_force_iommu) {
810 if (dev->vhost_ops->vhost_force_iommu(dev) == true) {
811 features |= 0x1ULL << VIRTIO_F_IOMMU_PLATFORM;
812 }
813 }
814 r = dev->vhost_ops->vhost_set_features(dev, features);
815 if (r < 0) {
816 VHOST_OPS_DEBUG("vhost_set_features failed");
817 goto out;
818 }
819 if (dev->vhost_ops->vhost_set_backend_cap) {
820 r = dev->vhost_ops->vhost_set_backend_cap(dev);
821 if (r < 0) {
822 VHOST_OPS_DEBUG("vhost_set_backend_cap failed");
823 goto out;
824 }
825 }
826
827 out:
828 return r < 0 ? -errno : 0;
829 }
830
831 static int vhost_dev_set_log(struct vhost_dev *dev, bool enable_log)
832 {
833 int r, i, idx;
834 hwaddr addr;
835
836 r = vhost_dev_set_features(dev, enable_log);
837 if (r < 0) {
838 goto err_features;
839 }
840 for (i = 0; i < dev->nvqs; ++i) {
841 idx = dev->vhost_ops->vhost_get_vq_index(dev, dev->vq_index + i);
842 addr = virtio_queue_get_desc_addr(dev->vdev, idx);
843 if (!addr) {
844 /*
845 * The queue might not be ready for start. If this
846 * is the case there is no reason to continue the process.
847 * The similar logic is used by the vhost_virtqueue_start()
848 * routine.
849 */
850 continue;
851 }
852 r = vhost_virtqueue_set_addr(dev, dev->vqs + i, idx,
853 enable_log);
854 if (r < 0) {
855 goto err_vq;
856 }
857 }
858 return 0;
859 err_vq:
860 for (; i >= 0; --i) {
861 idx = dev->vhost_ops->vhost_get_vq_index(dev, dev->vq_index + i);
862 vhost_virtqueue_set_addr(dev, dev->vqs + i, idx,
863 dev->log_enabled);
864 }
865 vhost_dev_set_features(dev, dev->log_enabled);
866 err_features:
867 return r;
868 }
869
870 static int vhost_migration_log(MemoryListener *listener, bool enable)
871 {
872 struct vhost_dev *dev = container_of(listener, struct vhost_dev,
873 memory_listener);
874 int r;
875 if (enable == dev->log_enabled) {
876 return 0;
877 }
878 if (!dev->started) {
879 dev->log_enabled = enable;
880 return 0;
881 }
882
883 r = 0;
884 if (!enable) {
885 r = vhost_dev_set_log(dev, false);
886 if (r < 0) {
887 goto check_dev_state;
888 }
889 vhost_log_put(dev, false);
890 } else {
891 vhost_dev_log_resize(dev, vhost_get_log_size(dev));
892 r = vhost_dev_set_log(dev, true);
893 if (r < 0) {
894 goto check_dev_state;
895 }
896 }
897
898 check_dev_state:
899 dev->log_enabled = enable;
900 /*
901 * vhost-user-* devices could change their state during log
902 * initialization due to disconnect. So check dev state after
903 * vhost communication.
904 */
905 if (!dev->started) {
906 /*
907 * Since device is in the stopped state, it is okay for
908 * migration. Return success.
909 */
910 r = 0;
911 }
912 if (r) {
913 /* An error is occured. */
914 dev->log_enabled = false;
915 }
916
917 return r;
918 }
919
920 static void vhost_log_global_start(MemoryListener *listener)
921 {
922 int r;
923
924 r = vhost_migration_log(listener, true);
925 if (r < 0) {
926 abort();
927 }
928 }
929
930 static void vhost_log_global_stop(MemoryListener *listener)
931 {
932 int r;
933
934 r = vhost_migration_log(listener, false);
935 if (r < 0) {
936 abort();
937 }
938 }
939
940 static void vhost_log_start(MemoryListener *listener,
941 MemoryRegionSection *section,
942 int old, int new)
943 {
944 /* FIXME: implement */
945 }
946
947 static void vhost_log_stop(MemoryListener *listener,
948 MemoryRegionSection *section,
949 int old, int new)
950 {
951 /* FIXME: implement */
952 }
953
954 /* The vhost driver natively knows how to handle the vrings of non
955 * cross-endian legacy devices and modern devices. Only legacy devices
956 * exposed to a bi-endian guest may require the vhost driver to use a
957 * specific endianness.
958 */
959 static inline bool vhost_needs_vring_endian(VirtIODevice *vdev)
960 {
961 if (virtio_vdev_has_feature(vdev, VIRTIO_F_VERSION_1)) {
962 return false;
963 }
964 #ifdef HOST_WORDS_BIGENDIAN
965 return vdev->device_endian == VIRTIO_DEVICE_ENDIAN_LITTLE;
966 #else
967 return vdev->device_endian == VIRTIO_DEVICE_ENDIAN_BIG;
968 #endif
969 }
970
971 static int vhost_virtqueue_set_vring_endian_legacy(struct vhost_dev *dev,
972 bool is_big_endian,
973 int vhost_vq_index)
974 {
975 struct vhost_vring_state s = {
976 .index = vhost_vq_index,
977 .num = is_big_endian
978 };
979
980 if (!dev->vhost_ops->vhost_set_vring_endian(dev, &s)) {
981 return 0;
982 }
983
984 VHOST_OPS_DEBUG("vhost_set_vring_endian failed");
985 if (errno == ENOTTY) {
986 error_report("vhost does not support cross-endian");
987 return -ENOSYS;
988 }
989
990 return -errno;
991 }
992
993 static int vhost_memory_region_lookup(struct vhost_dev *hdev,
994 uint64_t gpa, uint64_t *uaddr,
995 uint64_t *len)
996 {
997 int i;
998
999 for (i = 0; i < hdev->mem->nregions; i++) {
1000 struct vhost_memory_region *reg = hdev->mem->regions + i;
1001
1002 if (gpa >= reg->guest_phys_addr &&
1003 reg->guest_phys_addr + reg->memory_size > gpa) {
1004 *uaddr = reg->userspace_addr + gpa - reg->guest_phys_addr;
1005 *len = reg->guest_phys_addr + reg->memory_size - gpa;
1006 return 0;
1007 }
1008 }
1009
1010 return -EFAULT;
1011 }
1012
1013 int vhost_device_iotlb_miss(struct vhost_dev *dev, uint64_t iova, int write)
1014 {
1015 IOMMUTLBEntry iotlb;
1016 uint64_t uaddr, len;
1017 int ret = -EFAULT;
1018
1019 RCU_READ_LOCK_GUARD();
1020
1021 trace_vhost_iotlb_miss(dev, 1);
1022
1023 iotlb = address_space_get_iotlb_entry(dev->vdev->dma_as,
1024 iova, write,
1025 MEMTXATTRS_UNSPECIFIED);
1026 if (iotlb.target_as != NULL) {
1027 ret = vhost_memory_region_lookup(dev, iotlb.translated_addr,
1028 &uaddr, &len);
1029 if (ret) {
1030 trace_vhost_iotlb_miss(dev, 3);
1031 error_report("Fail to lookup the translated address "
1032 "%"PRIx64, iotlb.translated_addr);
1033 goto out;
1034 }
1035
1036 len = MIN(iotlb.addr_mask + 1, len);
1037 iova = iova & ~iotlb.addr_mask;
1038
1039 ret = vhost_backend_update_device_iotlb(dev, iova, uaddr,
1040 len, iotlb.perm);
1041 if (ret) {
1042 trace_vhost_iotlb_miss(dev, 4);
1043 error_report("Fail to update device iotlb");
1044 goto out;
1045 }
1046 }
1047
1048 trace_vhost_iotlb_miss(dev, 2);
1049
1050 out:
1051 return ret;
1052 }
1053
1054 static int vhost_virtqueue_start(struct vhost_dev *dev,
1055 struct VirtIODevice *vdev,
1056 struct vhost_virtqueue *vq,
1057 unsigned idx)
1058 {
1059 BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev)));
1060 VirtioBusState *vbus = VIRTIO_BUS(qbus);
1061 VirtioBusClass *k = VIRTIO_BUS_GET_CLASS(vbus);
1062 hwaddr s, l, a;
1063 int r;
1064 int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, idx);
1065 struct vhost_vring_file file = {
1066 .index = vhost_vq_index
1067 };
1068 struct vhost_vring_state state = {
1069 .index = vhost_vq_index
1070 };
1071 struct VirtQueue *vvq = virtio_get_queue(vdev, idx);
1072
1073 a = virtio_queue_get_desc_addr(vdev, idx);
1074 if (a == 0) {
1075 /* Queue might not be ready for start */
1076 return 0;
1077 }
1078
1079 vq->num = state.num = virtio_queue_get_num(vdev, idx);
1080 r = dev->vhost_ops->vhost_set_vring_num(dev, &state);
1081 if (r) {
1082 VHOST_OPS_DEBUG("vhost_set_vring_num failed");
1083 return -errno;
1084 }
1085
1086 state.num = virtio_queue_get_last_avail_idx(vdev, idx);
1087 r = dev->vhost_ops->vhost_set_vring_base(dev, &state);
1088 if (r) {
1089 VHOST_OPS_DEBUG("vhost_set_vring_base failed");
1090 return -errno;
1091 }
1092
1093 if (vhost_needs_vring_endian(vdev)) {
1094 r = vhost_virtqueue_set_vring_endian_legacy(dev,
1095 virtio_is_big_endian(vdev),
1096 vhost_vq_index);
1097 if (r) {
1098 return -errno;
1099 }
1100 }
1101
1102 vq->desc_size = s = l = virtio_queue_get_desc_size(vdev, idx);
1103 vq->desc_phys = a;
1104 vq->desc = vhost_memory_map(dev, a, &l, false);
1105 if (!vq->desc || l != s) {
1106 r = -ENOMEM;
1107 goto fail_alloc_desc;
1108 }
1109 vq->avail_size = s = l = virtio_queue_get_avail_size(vdev, idx);
1110 vq->avail_phys = a = virtio_queue_get_avail_addr(vdev, idx);
1111 vq->avail = vhost_memory_map(dev, a, &l, false);
1112 if (!vq->avail || l != s) {
1113 r = -ENOMEM;
1114 goto fail_alloc_avail;
1115 }
1116 vq->used_size = s = l = virtio_queue_get_used_size(vdev, idx);
1117 vq->used_phys = a = virtio_queue_get_used_addr(vdev, idx);
1118 vq->used = vhost_memory_map(dev, a, &l, true);
1119 if (!vq->used || l != s) {
1120 r = -ENOMEM;
1121 goto fail_alloc_used;
1122 }
1123
1124 r = vhost_virtqueue_set_addr(dev, vq, vhost_vq_index, dev->log_enabled);
1125 if (r < 0) {
1126 r = -errno;
1127 goto fail_alloc;
1128 }
1129
1130 file.fd = event_notifier_get_fd(virtio_queue_get_host_notifier(vvq));
1131 r = dev->vhost_ops->vhost_set_vring_kick(dev, &file);
1132 if (r) {
1133 VHOST_OPS_DEBUG("vhost_set_vring_kick failed");
1134 r = -errno;
1135 goto fail_kick;
1136 }
1137
1138 /* Clear and discard previous events if any. */
1139 event_notifier_test_and_clear(&vq->masked_notifier);
1140
1141 /* Init vring in unmasked state, unless guest_notifier_mask
1142 * will do it later.
1143 */
1144 if (!vdev->use_guest_notifier_mask) {
1145 /* TODO: check and handle errors. */
1146 vhost_virtqueue_mask(dev, vdev, idx, false);
1147 }
1148
1149 if (k->query_guest_notifiers &&
1150 k->query_guest_notifiers(qbus->parent) &&
1151 virtio_queue_vector(vdev, idx) == VIRTIO_NO_VECTOR) {
1152 file.fd = -1;
1153 r = dev->vhost_ops->vhost_set_vring_call(dev, &file);
1154 if (r) {
1155 goto fail_vector;
1156 }
1157 }
1158
1159 return 0;
1160
1161 fail_vector:
1162 fail_kick:
1163 fail_alloc:
1164 vhost_memory_unmap(dev, vq->used, virtio_queue_get_used_size(vdev, idx),
1165 0, 0);
1166 fail_alloc_used:
1167 vhost_memory_unmap(dev, vq->avail, virtio_queue_get_avail_size(vdev, idx),
1168 0, 0);
1169 fail_alloc_avail:
1170 vhost_memory_unmap(dev, vq->desc, virtio_queue_get_desc_size(vdev, idx),
1171 0, 0);
1172 fail_alloc_desc:
1173 return r;
1174 }
1175
1176 static void vhost_virtqueue_stop(struct vhost_dev *dev,
1177 struct VirtIODevice *vdev,
1178 struct vhost_virtqueue *vq,
1179 unsigned idx)
1180 {
1181 int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, idx);
1182 struct vhost_vring_state state = {
1183 .index = vhost_vq_index,
1184 };
1185 int r;
1186
1187 if (virtio_queue_get_desc_addr(vdev, idx) == 0) {
1188 /* Don't stop the virtqueue which might have not been started */
1189 return;
1190 }
1191
1192 r = dev->vhost_ops->vhost_get_vring_base(dev, &state);
1193 if (r < 0) {
1194 VHOST_OPS_DEBUG("vhost VQ %u ring restore failed: %d", idx, r);
1195 /* Connection to the backend is broken, so let's sync internal
1196 * last avail idx to the device used idx.
1197 */
1198 virtio_queue_restore_last_avail_idx(vdev, idx);
1199 } else {
1200 virtio_queue_set_last_avail_idx(vdev, idx, state.num);
1201 }
1202 virtio_queue_invalidate_signalled_used(vdev, idx);
1203 virtio_queue_update_used_idx(vdev, idx);
1204
1205 /* In the cross-endian case, we need to reset the vring endianness to
1206 * native as legacy devices expect so by default.
1207 */
1208 if (vhost_needs_vring_endian(vdev)) {
1209 vhost_virtqueue_set_vring_endian_legacy(dev,
1210 !virtio_is_big_endian(vdev),
1211 vhost_vq_index);
1212 }
1213
1214 vhost_memory_unmap(dev, vq->used, virtio_queue_get_used_size(vdev, idx),
1215 1, virtio_queue_get_used_size(vdev, idx));
1216 vhost_memory_unmap(dev, vq->avail, virtio_queue_get_avail_size(vdev, idx),
1217 0, virtio_queue_get_avail_size(vdev, idx));
1218 vhost_memory_unmap(dev, vq->desc, virtio_queue_get_desc_size(vdev, idx),
1219 0, virtio_queue_get_desc_size(vdev, idx));
1220 }
1221
1222 static void vhost_eventfd_add(MemoryListener *listener,
1223 MemoryRegionSection *section,
1224 bool match_data, uint64_t data, EventNotifier *e)
1225 {
1226 }
1227
1228 static void vhost_eventfd_del(MemoryListener *listener,
1229 MemoryRegionSection *section,
1230 bool match_data, uint64_t data, EventNotifier *e)
1231 {
1232 }
1233
1234 static int vhost_virtqueue_set_busyloop_timeout(struct vhost_dev *dev,
1235 int n, uint32_t timeout)
1236 {
1237 int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, n);
1238 struct vhost_vring_state state = {
1239 .index = vhost_vq_index,
1240 .num = timeout,
1241 };
1242 int r;
1243
1244 if (!dev->vhost_ops->vhost_set_vring_busyloop_timeout) {
1245 return -EINVAL;
1246 }
1247
1248 r = dev->vhost_ops->vhost_set_vring_busyloop_timeout(dev, &state);
1249 if (r) {
1250 VHOST_OPS_DEBUG("vhost_set_vring_busyloop_timeout failed");
1251 return r;
1252 }
1253
1254 return 0;
1255 }
1256
1257 static int vhost_virtqueue_init(struct vhost_dev *dev,
1258 struct vhost_virtqueue *vq, int n)
1259 {
1260 int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, n);
1261 struct vhost_vring_file file = {
1262 .index = vhost_vq_index,
1263 };
1264 int r = event_notifier_init(&vq->masked_notifier, 0);
1265 if (r < 0) {
1266 return r;
1267 }
1268
1269 file.fd = event_notifier_get_fd(&vq->masked_notifier);
1270 r = dev->vhost_ops->vhost_set_vring_call(dev, &file);
1271 if (r) {
1272 VHOST_OPS_DEBUG("vhost_set_vring_call failed");
1273 r = -errno;
1274 goto fail_call;
1275 }
1276
1277 vq->dev = dev;
1278
1279 return 0;
1280 fail_call:
1281 event_notifier_cleanup(&vq->masked_notifier);
1282 return r;
1283 }
1284
1285 static void vhost_virtqueue_cleanup(struct vhost_virtqueue *vq)
1286 {
1287 event_notifier_cleanup(&vq->masked_notifier);
1288 }
1289
1290 int vhost_dev_init(struct vhost_dev *hdev, void *opaque,
1291 VhostBackendType backend_type, uint32_t busyloop_timeout)
1292 {
1293 uint64_t features;
1294 int i, r, n_initialized_vqs = 0;
1295 Error *local_err = NULL;
1296
1297 hdev->vdev = NULL;
1298 hdev->migration_blocker = NULL;
1299
1300 r = vhost_set_backend_type(hdev, backend_type);
1301 assert(r >= 0);
1302
1303 r = hdev->vhost_ops->vhost_backend_init(hdev, opaque);
1304 if (r < 0) {
1305 goto fail;
1306 }
1307
1308 r = hdev->vhost_ops->vhost_set_owner(hdev);
1309 if (r < 0) {
1310 VHOST_OPS_DEBUG("vhost_set_owner failed");
1311 goto fail;
1312 }
1313
1314 r = hdev->vhost_ops->vhost_get_features(hdev, &features);
1315 if (r < 0) {
1316 VHOST_OPS_DEBUG("vhost_get_features failed");
1317 goto fail;
1318 }
1319
1320 for (i = 0; i < hdev->nvqs; ++i, ++n_initialized_vqs) {
1321 r = vhost_virtqueue_init(hdev, hdev->vqs + i, hdev->vq_index + i);
1322 if (r < 0) {
1323 goto fail;
1324 }
1325 }
1326
1327 if (busyloop_timeout) {
1328 for (i = 0; i < hdev->nvqs; ++i) {
1329 r = vhost_virtqueue_set_busyloop_timeout(hdev, hdev->vq_index + i,
1330 busyloop_timeout);
1331 if (r < 0) {
1332 goto fail_busyloop;
1333 }
1334 }
1335 }
1336
1337 hdev->features = features;
1338
1339 hdev->memory_listener = (MemoryListener) {
1340 .begin = vhost_begin,
1341 .commit = vhost_commit,
1342 .region_add = vhost_region_addnop,
1343 .region_nop = vhost_region_addnop,
1344 .log_start = vhost_log_start,
1345 .log_stop = vhost_log_stop,
1346 .log_sync = vhost_log_sync,
1347 .log_global_start = vhost_log_global_start,
1348 .log_global_stop = vhost_log_global_stop,
1349 .eventfd_add = vhost_eventfd_add,
1350 .eventfd_del = vhost_eventfd_del,
1351 .priority = 10
1352 };
1353
1354 hdev->iommu_listener = (MemoryListener) {
1355 .region_add = vhost_iommu_region_add,
1356 .region_del = vhost_iommu_region_del,
1357 };
1358
1359 if (hdev->migration_blocker == NULL) {
1360 if (!(hdev->features & (0x1ULL << VHOST_F_LOG_ALL))) {
1361 error_setg(&hdev->migration_blocker,
1362 "Migration disabled: vhost lacks VHOST_F_LOG_ALL feature.");
1363 } else if (vhost_dev_log_is_shared(hdev) && !qemu_memfd_alloc_check()) {
1364 error_setg(&hdev->migration_blocker,
1365 "Migration disabled: failed to allocate shared memory");
1366 }
1367 }
1368
1369 if (hdev->migration_blocker != NULL) {
1370 r = migrate_add_blocker(hdev->migration_blocker, &local_err);
1371 if (local_err) {
1372 error_report_err(local_err);
1373 error_free(hdev->migration_blocker);
1374 goto fail_busyloop;
1375 }
1376 }
1377
1378 hdev->mem = g_malloc0(offsetof(struct vhost_memory, regions));
1379 hdev->n_mem_sections = 0;
1380 hdev->mem_sections = NULL;
1381 hdev->log = NULL;
1382 hdev->log_size = 0;
1383 hdev->log_enabled = false;
1384 hdev->started = false;
1385 memory_listener_register(&hdev->memory_listener, &address_space_memory);
1386 QLIST_INSERT_HEAD(&vhost_devices, hdev, entry);
1387
1388 if (used_memslots > hdev->vhost_ops->vhost_backend_memslots_limit(hdev)) {
1389 error_report("vhost backend memory slots limit is less"
1390 " than current number of present memory slots");
1391 r = -1;
1392 if (busyloop_timeout) {
1393 goto fail_busyloop;
1394 } else {
1395 goto fail;
1396 }
1397 }
1398
1399 return 0;
1400
1401 fail_busyloop:
1402 while (--i >= 0) {
1403 vhost_virtqueue_set_busyloop_timeout(hdev, hdev->vq_index + i, 0);
1404 }
1405 fail:
1406 hdev->nvqs = n_initialized_vqs;
1407 vhost_dev_cleanup(hdev);
1408 return r;
1409 }
1410
1411 void vhost_dev_cleanup(struct vhost_dev *hdev)
1412 {
1413 int i;
1414
1415 for (i = 0; i < hdev->nvqs; ++i) {
1416 vhost_virtqueue_cleanup(hdev->vqs + i);
1417 }
1418 if (hdev->mem) {
1419 /* those are only safe after successful init */
1420 memory_listener_unregister(&hdev->memory_listener);
1421 QLIST_REMOVE(hdev, entry);
1422 }
1423 if (hdev->migration_blocker) {
1424 migrate_del_blocker(hdev->migration_blocker);
1425 error_free(hdev->migration_blocker);
1426 }
1427 g_free(hdev->mem);
1428 g_free(hdev->mem_sections);
1429 if (hdev->vhost_ops) {
1430 hdev->vhost_ops->vhost_backend_cleanup(hdev);
1431 }
1432 assert(!hdev->log);
1433
1434 memset(hdev, 0, sizeof(struct vhost_dev));
1435 }
1436
1437 /* Stop processing guest IO notifications in qemu.
1438 * Start processing them in vhost in kernel.
1439 */
1440 int vhost_dev_enable_notifiers(struct vhost_dev *hdev, VirtIODevice *vdev)
1441 {
1442 BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev)));
1443 int i, r, e;
1444
1445 /* We will pass the notifiers to the kernel, make sure that QEMU
1446 * doesn't interfere.
1447 */
1448 r = virtio_device_grab_ioeventfd(vdev);
1449 if (r < 0) {
1450 error_report("binding does not support host notifiers");
1451 goto fail;
1452 }
1453
1454 for (i = 0; i < hdev->nvqs; ++i) {
1455 r = virtio_bus_set_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i,
1456 true);
1457 if (r < 0) {
1458 error_report("vhost VQ %d notifier binding failed: %d", i, -r);
1459 goto fail_vq;
1460 }
1461 }
1462
1463 return 0;
1464 fail_vq:
1465 while (--i >= 0) {
1466 e = virtio_bus_set_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i,
1467 false);
1468 if (e < 0) {
1469 error_report("vhost VQ %d notifier cleanup error: %d", i, -r);
1470 }
1471 assert (e >= 0);
1472 virtio_bus_cleanup_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i);
1473 }
1474 virtio_device_release_ioeventfd(vdev);
1475 fail:
1476 return r;
1477 }
1478
1479 /* Stop processing guest IO notifications in vhost.
1480 * Start processing them in qemu.
1481 * This might actually run the qemu handlers right away,
1482 * so virtio in qemu must be completely setup when this is called.
1483 */
1484 void vhost_dev_disable_notifiers(struct vhost_dev *hdev, VirtIODevice *vdev)
1485 {
1486 BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev)));
1487 int i, r;
1488
1489 for (i = 0; i < hdev->nvqs; ++i) {
1490 r = virtio_bus_set_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i,
1491 false);
1492 if (r < 0) {
1493 error_report("vhost VQ %d notifier cleanup failed: %d", i, -r);
1494 }
1495 assert (r >= 0);
1496 virtio_bus_cleanup_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i);
1497 }
1498 virtio_device_release_ioeventfd(vdev);
1499 }
1500
1501 /* Test and clear event pending status.
1502 * Should be called after unmask to avoid losing events.
1503 */
1504 bool vhost_virtqueue_pending(struct vhost_dev *hdev, int n)
1505 {
1506 struct vhost_virtqueue *vq = hdev->vqs + n - hdev->vq_index;
1507 assert(n >= hdev->vq_index && n < hdev->vq_index + hdev->nvqs);
1508 return event_notifier_test_and_clear(&vq->masked_notifier);
1509 }
1510
1511 /* Mask/unmask events from this vq. */
1512 void vhost_virtqueue_mask(struct vhost_dev *hdev, VirtIODevice *vdev, int n,
1513 bool mask)
1514 {
1515 struct VirtQueue *vvq = virtio_get_queue(vdev, n);
1516 int r, index = n - hdev->vq_index;
1517 struct vhost_vring_file file;
1518
1519 /* should only be called after backend is connected */
1520 assert(hdev->vhost_ops);
1521
1522 if (mask) {
1523 assert(vdev->use_guest_notifier_mask);
1524 file.fd = event_notifier_get_fd(&hdev->vqs[index].masked_notifier);
1525 } else {
1526 file.fd = event_notifier_get_fd(virtio_queue_get_guest_notifier(vvq));
1527 }
1528
1529 file.index = hdev->vhost_ops->vhost_get_vq_index(hdev, n);
1530 r = hdev->vhost_ops->vhost_set_vring_call(hdev, &file);
1531 if (r < 0) {
1532 VHOST_OPS_DEBUG("vhost_set_vring_call failed");
1533 }
1534 }
1535
1536 uint64_t vhost_get_features(struct vhost_dev *hdev, const int *feature_bits,
1537 uint64_t features)
1538 {
1539 const int *bit = feature_bits;
1540 while (*bit != VHOST_INVALID_FEATURE_BIT) {
1541 uint64_t bit_mask = (1ULL << *bit);
1542 if (!(hdev->features & bit_mask)) {
1543 features &= ~bit_mask;
1544 }
1545 bit++;
1546 }
1547 return features;
1548 }
1549
1550 void vhost_ack_features(struct vhost_dev *hdev, const int *feature_bits,
1551 uint64_t features)
1552 {
1553 const int *bit = feature_bits;
1554 while (*bit != VHOST_INVALID_FEATURE_BIT) {
1555 uint64_t bit_mask = (1ULL << *bit);
1556 if (features & bit_mask) {
1557 hdev->acked_features |= bit_mask;
1558 }
1559 bit++;
1560 }
1561 }
1562
1563 int vhost_dev_get_config(struct vhost_dev *hdev, uint8_t *config,
1564 uint32_t config_len)
1565 {
1566 assert(hdev->vhost_ops);
1567
1568 if (hdev->vhost_ops->vhost_get_config) {
1569 return hdev->vhost_ops->vhost_get_config(hdev, config, config_len);
1570 }
1571
1572 return -1;
1573 }
1574
1575 int vhost_dev_set_config(struct vhost_dev *hdev, const uint8_t *data,
1576 uint32_t offset, uint32_t size, uint32_t flags)
1577 {
1578 assert(hdev->vhost_ops);
1579
1580 if (hdev->vhost_ops->vhost_set_config) {
1581 return hdev->vhost_ops->vhost_set_config(hdev, data, offset,
1582 size, flags);
1583 }
1584
1585 return -1;
1586 }
1587
1588 void vhost_dev_set_config_notifier(struct vhost_dev *hdev,
1589 const VhostDevConfigOps *ops)
1590 {
1591 hdev->config_ops = ops;
1592 }
1593
1594 void vhost_dev_free_inflight(struct vhost_inflight *inflight)
1595 {
1596 if (inflight && inflight->addr) {
1597 qemu_memfd_free(inflight->addr, inflight->size, inflight->fd);
1598 inflight->addr = NULL;
1599 inflight->fd = -1;
1600 }
1601 }
1602
1603 static int vhost_dev_resize_inflight(struct vhost_inflight *inflight,
1604 uint64_t new_size)
1605 {
1606 Error *err = NULL;
1607 int fd = -1;
1608 void *addr = qemu_memfd_alloc("vhost-inflight", new_size,
1609 F_SEAL_GROW | F_SEAL_SHRINK | F_SEAL_SEAL,
1610 &fd, &err);
1611
1612 if (err) {
1613 error_report_err(err);
1614 return -1;
1615 }
1616
1617 vhost_dev_free_inflight(inflight);
1618 inflight->offset = 0;
1619 inflight->addr = addr;
1620 inflight->fd = fd;
1621 inflight->size = new_size;
1622
1623 return 0;
1624 }
1625
1626 void vhost_dev_save_inflight(struct vhost_inflight *inflight, QEMUFile *f)
1627 {
1628 if (inflight->addr) {
1629 qemu_put_be64(f, inflight->size);
1630 qemu_put_be16(f, inflight->queue_size);
1631 qemu_put_buffer(f, inflight->addr, inflight->size);
1632 } else {
1633 qemu_put_be64(f, 0);
1634 }
1635 }
1636
1637 int vhost_dev_load_inflight(struct vhost_inflight *inflight, QEMUFile *f)
1638 {
1639 uint64_t size;
1640
1641 size = qemu_get_be64(f);
1642 if (!size) {
1643 return 0;
1644 }
1645
1646 if (inflight->size != size) {
1647 if (vhost_dev_resize_inflight(inflight, size)) {
1648 return -1;
1649 }
1650 }
1651 inflight->queue_size = qemu_get_be16(f);
1652
1653 qemu_get_buffer(f, inflight->addr, size);
1654
1655 return 0;
1656 }
1657
1658 int vhost_dev_set_inflight(struct vhost_dev *dev,
1659 struct vhost_inflight *inflight)
1660 {
1661 int r;
1662
1663 if (dev->vhost_ops->vhost_set_inflight_fd && inflight->addr) {
1664 r = dev->vhost_ops->vhost_set_inflight_fd(dev, inflight);
1665 if (r) {
1666 VHOST_OPS_DEBUG("vhost_set_inflight_fd failed");
1667 return -errno;
1668 }
1669 }
1670
1671 return 0;
1672 }
1673
1674 int vhost_dev_get_inflight(struct vhost_dev *dev, uint16_t queue_size,
1675 struct vhost_inflight *inflight)
1676 {
1677 int r;
1678
1679 if (dev->vhost_ops->vhost_get_inflight_fd) {
1680 r = dev->vhost_ops->vhost_get_inflight_fd(dev, queue_size, inflight);
1681 if (r) {
1682 VHOST_OPS_DEBUG("vhost_get_inflight_fd failed");
1683 return -errno;
1684 }
1685 }
1686
1687 return 0;
1688 }
1689
1690 /* Host notifiers must be enabled at this point. */
1691 int vhost_dev_start(struct vhost_dev *hdev, VirtIODevice *vdev)
1692 {
1693 int i, r;
1694
1695 /* should only be called after backend is connected */
1696 assert(hdev->vhost_ops);
1697
1698 hdev->started = true;
1699 hdev->vdev = vdev;
1700
1701 r = vhost_dev_set_features(hdev, hdev->log_enabled);
1702 if (r < 0) {
1703 goto fail_features;
1704 }
1705
1706 if (vhost_dev_has_iommu(hdev)) {
1707 memory_listener_register(&hdev->iommu_listener, vdev->dma_as);
1708 }
1709
1710 r = hdev->vhost_ops->vhost_set_mem_table(hdev, hdev->mem);
1711 if (r < 0) {
1712 VHOST_OPS_DEBUG("vhost_set_mem_table failed");
1713 r = -errno;
1714 goto fail_mem;
1715 }
1716 for (i = 0; i < hdev->nvqs; ++i) {
1717 r = vhost_virtqueue_start(hdev,
1718 vdev,
1719 hdev->vqs + i,
1720 hdev->vq_index + i);
1721 if (r < 0) {
1722 goto fail_vq;
1723 }
1724 }
1725
1726 if (hdev->log_enabled) {
1727 uint64_t log_base;
1728
1729 hdev->log_size = vhost_get_log_size(hdev);
1730 hdev->log = vhost_log_get(hdev->log_size,
1731 vhost_dev_log_is_shared(hdev));
1732 log_base = (uintptr_t)hdev->log->log;
1733 r = hdev->vhost_ops->vhost_set_log_base(hdev,
1734 hdev->log_size ? log_base : 0,
1735 hdev->log);
1736 if (r < 0) {
1737 VHOST_OPS_DEBUG("vhost_set_log_base failed");
1738 r = -errno;
1739 goto fail_log;
1740 }
1741 }
1742 if (hdev->vhost_ops->vhost_dev_start) {
1743 r = hdev->vhost_ops->vhost_dev_start(hdev, true);
1744 if (r) {
1745 goto fail_log;
1746 }
1747 }
1748 if (vhost_dev_has_iommu(hdev) &&
1749 hdev->vhost_ops->vhost_set_iotlb_callback) {
1750 hdev->vhost_ops->vhost_set_iotlb_callback(hdev, true);
1751
1752 /* Update used ring information for IOTLB to work correctly,
1753 * vhost-kernel code requires for this.*/
1754 for (i = 0; i < hdev->nvqs; ++i) {
1755 struct vhost_virtqueue *vq = hdev->vqs + i;
1756 vhost_device_iotlb_miss(hdev, vq->used_phys, true);
1757 }
1758 }
1759 return 0;
1760 fail_log:
1761 vhost_log_put(hdev, false);
1762 fail_vq:
1763 while (--i >= 0) {
1764 vhost_virtqueue_stop(hdev,
1765 vdev,
1766 hdev->vqs + i,
1767 hdev->vq_index + i);
1768 }
1769
1770 fail_mem:
1771 fail_features:
1772
1773 hdev->started = false;
1774 return r;
1775 }
1776
1777 /* Host notifiers must be enabled at this point. */
1778 void vhost_dev_stop(struct vhost_dev *hdev, VirtIODevice *vdev)
1779 {
1780 int i;
1781
1782 /* should only be called after backend is connected */
1783 assert(hdev->vhost_ops);
1784
1785 if (hdev->vhost_ops->vhost_dev_start) {
1786 hdev->vhost_ops->vhost_dev_start(hdev, false);
1787 }
1788 for (i = 0; i < hdev->nvqs; ++i) {
1789 vhost_virtqueue_stop(hdev,
1790 vdev,
1791 hdev->vqs + i,
1792 hdev->vq_index + i);
1793 }
1794
1795 if (vhost_dev_has_iommu(hdev)) {
1796 if (hdev->vhost_ops->vhost_set_iotlb_callback) {
1797 hdev->vhost_ops->vhost_set_iotlb_callback(hdev, false);
1798 }
1799 memory_listener_unregister(&hdev->iommu_listener);
1800 }
1801 vhost_log_put(hdev, true);
1802 hdev->started = false;
1803 hdev->vdev = NULL;
1804 }
1805
1806 int vhost_net_set_backend(struct vhost_dev *hdev,
1807 struct vhost_vring_file *file)
1808 {
1809 if (hdev->vhost_ops->vhost_net_set_backend) {
1810 return hdev->vhost_ops->vhost_net_set_backend(hdev, file);
1811 }
1812
1813 return -1;
1814 }