/src/qemu/hw/virtio/vhost.c
Line | Count | Source |
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 "qemu/log.h" |
24 | | #include "standard-headers/linux/vhost_types.h" |
25 | | #include "hw/virtio/virtio-bus.h" |
26 | | #include "hw/mem/memory-device.h" |
27 | | #include "migration/blocker.h" |
28 | | #include "migration/misc.h" |
29 | | #include "migration/qemu-file-types.h" |
30 | | #include "system/dma.h" |
31 | | #include "system/memory.h" |
32 | | #include "system/ramblock.h" |
33 | | #include "system/system.h" |
34 | | #include "trace.h" |
35 | | |
36 | | /* enabled until disconnected backend stabilizes */ |
37 | | #define _VHOST_DEBUG 1 |
38 | | |
39 | | #ifdef _VHOST_DEBUG |
40 | | #define VHOST_OPS_DEBUG(retval, fmt, ...) \ |
41 | 0 | do { \ |
42 | 0 | error_report(fmt ": %s (%d)", ## __VA_ARGS__, \ |
43 | 0 | strerror(-retval), -retval); \ |
44 | 0 | } while (0) |
45 | | #else |
46 | | #define VHOST_OPS_DEBUG(retval, fmt, ...) \ |
47 | | do { } while (0) |
48 | | #endif |
49 | | |
50 | | static struct vhost_log *vhost_log[VHOST_BACKEND_TYPE_MAX]; |
51 | | static struct vhost_log *vhost_log_shm[VHOST_BACKEND_TYPE_MAX]; |
52 | | static QLIST_HEAD(, vhost_dev) vhost_log_devs[VHOST_BACKEND_TYPE_MAX]; |
53 | | |
54 | | static QLIST_HEAD(, vhost_dev) vhost_devices = |
55 | | QLIST_HEAD_INITIALIZER(vhost_devices); |
56 | | |
57 | | /* |
58 | | * Release the device ownership if it was handed over to us by another |
59 | | * process, e.g. the CPR source. |
60 | | */ |
61 | | static void vhost_dev_try_release_owner(struct vhost_dev *hdev, |
62 | | const char *when) |
63 | 0 | { |
64 | 0 | int r; |
65 | |
|
66 | 0 | if (!hdev->owner || !hdev->owner_handed_over) { |
67 | 0 | return; |
68 | 0 | } |
69 | 0 | r = vhost_dev_reset_owner(hdev); |
70 | 0 | if (r < 0) { |
71 | 0 | error_report("vhost: cannot release device ownership on %s: %s", |
72 | 0 | when, strerror(-r)); |
73 | 0 | } |
74 | 0 | } |
75 | | |
76 | | /* |
77 | | * A CPR target takes the device ownership in post_load(). If the load |
78 | | * fails afterwards, release it before the source learns about the failure |
79 | | * and resumes, so that the source can take the device back. |
80 | | */ |
81 | | static int vhost_migration_notify(NotifierWithReturn *notifier, |
82 | | MigrationEvent *e, Error **errp) |
83 | 0 | { |
84 | 0 | struct vhost_dev *hdev = container_of(notifier, struct vhost_dev, |
85 | 0 | migration_notifier); |
86 | |
|
87 | 0 | if (e->type == MIG_EVENT_INCOMING_FAILED) { |
88 | 0 | vhost_dev_try_release_owner(hdev, "incoming migration failure"); |
89 | 0 | } |
90 | 0 | return 0; |
91 | 0 | } |
92 | | |
93 | | /* |
94 | | * Release the ownership of the vhost devices still held at exit(), so that |
95 | | * another process sharing the backend FDs (e.g. the CPR source) can take |
96 | | * them over. |
97 | | */ |
98 | | static void vhost_exit_notify(Notifier *notifier, void *data) |
99 | 0 | { |
100 | 0 | struct vhost_dev *hdev = container_of(notifier, struct vhost_dev, |
101 | 0 | exit_notifier); |
102 | |
|
103 | 0 | vhost_dev_try_release_owner(hdev, "exit"); |
104 | 0 | } |
105 | | |
106 | | unsigned int vhost_get_max_memslots(void) |
107 | 0 | { |
108 | 0 | unsigned int max = UINT_MAX; |
109 | 0 | struct vhost_dev *hdev; |
110 | |
|
111 | 0 | QLIST_FOREACH(hdev, &vhost_devices, entry) { |
112 | 0 | max = MIN(max, hdev->vhost_ops->vhost_memslots_limit(hdev)); |
113 | 0 | } |
114 | 0 | return max; |
115 | 0 | } |
116 | | |
117 | | unsigned int vhost_get_free_memslots(void) |
118 | 0 | { |
119 | 0 | unsigned int free = UINT_MAX; |
120 | 0 | struct vhost_dev *hdev; |
121 | |
|
122 | 0 | QLIST_FOREACH(hdev, &vhost_devices, entry) { |
123 | 0 | unsigned int r = hdev->vhost_ops->vhost_memslots_limit(hdev); |
124 | 0 | unsigned int cur_free = r - hdev->mem->nregions; |
125 | |
|
126 | 0 | if (unlikely(r < hdev->mem->nregions)) { |
127 | 0 | warn_report_once("used (%u) vhost backend memory slots exceed" |
128 | 0 | " the device limit (%u).", hdev->mem->nregions, r); |
129 | 0 | free = 0; |
130 | 0 | } else { |
131 | 0 | free = MIN(free, cur_free); |
132 | 0 | } |
133 | 0 | } |
134 | 0 | return free; |
135 | 0 | } |
136 | | |
137 | | static void vhost_dev_sync_region(struct vhost_dev *dev, |
138 | | MemoryRegionSection *section, |
139 | | uint64_t mfirst, uint64_t mlast, |
140 | | uint64_t rfirst, uint64_t rlast) |
141 | 0 | { |
142 | 0 | vhost_log_chunk_t *dev_log = dev->log->log; |
143 | |
|
144 | 0 | uint64_t start = MAX(mfirst, rfirst); |
145 | 0 | uint64_t end = MIN(mlast, rlast); |
146 | 0 | vhost_log_chunk_t *from = dev_log + start / VHOST_LOG_CHUNK; |
147 | 0 | vhost_log_chunk_t *to = dev_log + end / VHOST_LOG_CHUNK + 1; |
148 | 0 | uint64_t addr = QEMU_ALIGN_DOWN(start, VHOST_LOG_CHUNK); |
149 | |
|
150 | 0 | if (end < start) { |
151 | 0 | return; |
152 | 0 | } |
153 | 0 | assert(end / VHOST_LOG_CHUNK < dev->log_size); |
154 | 0 | assert(start / VHOST_LOG_CHUNK < dev->log_size); |
155 | |
|
156 | 0 | for (;from < to; ++from) { |
157 | 0 | vhost_log_chunk_t log; |
158 | | /* We first check with non-atomic: much cheaper, |
159 | | * and we expect non-dirty to be the common case. */ |
160 | 0 | if (!*from) { |
161 | 0 | addr += VHOST_LOG_CHUNK; |
162 | 0 | continue; |
163 | 0 | } |
164 | | /* Data must be read atomically. We don't really need barrier semantics |
165 | | * but it's easier to use atomic_* than roll our own. */ |
166 | 0 | log = qatomic_xchg(from, 0); |
167 | 0 | while (log) { |
168 | 0 | int bit = ctzl(log); |
169 | 0 | hwaddr page_addr; |
170 | 0 | hwaddr section_offset; |
171 | 0 | hwaddr mr_offset; |
172 | 0 | page_addr = addr + bit * VHOST_LOG_PAGE; |
173 | 0 | section_offset = page_addr - section->offset_within_address_space; |
174 | 0 | mr_offset = section_offset + section->offset_within_region; |
175 | 0 | memory_region_set_dirty(section->mr, mr_offset, VHOST_LOG_PAGE); |
176 | 0 | log &= ~(0x1ull << bit); |
177 | 0 | } |
178 | 0 | addr += VHOST_LOG_CHUNK; |
179 | 0 | } |
180 | 0 | } |
181 | | |
182 | | bool vhost_dev_has_iommu(struct vhost_dev *dev) |
183 | 0 | { |
184 | 0 | VirtIODevice *vdev = dev->vdev; |
185 | | |
186 | | /* |
187 | | * For vhost, VIRTIO_F_IOMMU_PLATFORM means the backend support |
188 | | * incremental memory mapping API via IOTLB API. For platform that |
189 | | * does not have IOMMU, there's no need to enable this feature |
190 | | * which may cause unnecessary IOTLB miss/update transactions. |
191 | | */ |
192 | 0 | if (vdev) { |
193 | 0 | return virtio_bus_device_iommu_enabled(vdev) && |
194 | 0 | virtio_host_has_feature(vdev, VIRTIO_F_IOMMU_PLATFORM); |
195 | 0 | } else { |
196 | 0 | return false; |
197 | 0 | } |
198 | 0 | } |
199 | | |
200 | | static inline bool vhost_dev_should_log(struct vhost_dev *dev) |
201 | 0 | { |
202 | 0 | assert(dev->vhost_ops); |
203 | 0 | assert(dev->vhost_ops->backend_type > VHOST_BACKEND_TYPE_NONE); |
204 | 0 | assert(dev->vhost_ops->backend_type < VHOST_BACKEND_TYPE_MAX); |
205 | |
|
206 | 0 | return dev == QLIST_FIRST(&vhost_log_devs[dev->vhost_ops->backend_type]); |
207 | 0 | } |
208 | | |
209 | | static inline void vhost_dev_elect_mem_logger(struct vhost_dev *hdev, bool add) |
210 | 0 | { |
211 | 0 | VhostBackendType backend_type; |
212 | |
|
213 | 0 | assert(hdev->vhost_ops); |
214 | |
|
215 | 0 | backend_type = hdev->vhost_ops->backend_type; |
216 | 0 | assert(backend_type > VHOST_BACKEND_TYPE_NONE); |
217 | 0 | assert(backend_type < VHOST_BACKEND_TYPE_MAX); |
218 | |
|
219 | 0 | if (add && !QLIST_IS_INSERTED(hdev, logdev_entry)) { |
220 | 0 | if (QLIST_EMPTY(&vhost_log_devs[backend_type])) { |
221 | 0 | QLIST_INSERT_HEAD(&vhost_log_devs[backend_type], |
222 | 0 | hdev, logdev_entry); |
223 | 0 | } else { |
224 | | /* |
225 | | * The first vhost_device in the list is selected as the shared |
226 | | * logger to scan memory sections. Put new entry next to the head |
227 | | * to avoid inadvertent change to the underlying logger device. |
228 | | * This is done in order to get better cache locality and to avoid |
229 | | * performance churn on the hot path for log scanning. Even when |
230 | | * new devices come and go quickly, it wouldn't end up changing |
231 | | * the active leading logger device at all. |
232 | | */ |
233 | 0 | QLIST_INSERT_AFTER(QLIST_FIRST(&vhost_log_devs[backend_type]), |
234 | 0 | hdev, logdev_entry); |
235 | 0 | } |
236 | 0 | } else if (!add && QLIST_IS_INSERTED(hdev, logdev_entry)) { |
237 | 0 | QLIST_REMOVE(hdev, logdev_entry); |
238 | 0 | } |
239 | 0 | } |
240 | | |
241 | | static int vhost_sync_dirty_bitmap(struct vhost_dev *dev, |
242 | | MemoryRegionSection *section, |
243 | | hwaddr first, |
244 | | hwaddr last) |
245 | 0 | { |
246 | 0 | int i; |
247 | 0 | hwaddr start_addr; |
248 | 0 | hwaddr end_addr; |
249 | |
|
250 | 0 | if (!dev->log_enabled || !dev->started) { |
251 | 0 | return 0; |
252 | 0 | } |
253 | 0 | start_addr = section->offset_within_address_space; |
254 | 0 | end_addr = range_get_last(start_addr, int128_get64(section->size)); |
255 | 0 | start_addr = MAX(first, start_addr); |
256 | 0 | end_addr = MIN(last, end_addr); |
257 | |
|
258 | 0 | if (vhost_dev_should_log(dev)) { |
259 | 0 | for (i = 0; i < dev->mem->nregions; ++i) { |
260 | 0 | struct vhost_memory_region *reg = dev->mem->regions + i; |
261 | 0 | vhost_dev_sync_region(dev, section, start_addr, end_addr, |
262 | 0 | reg->guest_phys_addr, |
263 | 0 | range_get_last(reg->guest_phys_addr, |
264 | 0 | reg->memory_size)); |
265 | 0 | } |
266 | 0 | } |
267 | 0 | for (i = 0; i < dev->nvqs; ++i) { |
268 | 0 | struct vhost_virtqueue *vq = dev->vqs + i; |
269 | |
|
270 | 0 | if (!vq->used_phys && !vq->used_size) { |
271 | 0 | continue; |
272 | 0 | } |
273 | | |
274 | 0 | if (vhost_dev_has_iommu(dev)) { |
275 | 0 | IOMMUTLBEntry iotlb; |
276 | 0 | hwaddr used_phys = vq->used_phys, used_size = vq->used_size; |
277 | 0 | hwaddr phys, s, offset; |
278 | |
|
279 | 0 | while (used_size) { |
280 | 0 | rcu_read_lock(); |
281 | 0 | iotlb = address_space_get_iotlb_entry(dev->vdev->dma_as, |
282 | 0 | used_phys, |
283 | 0 | true, |
284 | 0 | MEMTXATTRS_UNSPECIFIED); |
285 | 0 | rcu_read_unlock(); |
286 | |
|
287 | 0 | if (!iotlb.target_as) { |
288 | 0 | qemu_log_mask(LOG_GUEST_ERROR, "translation " |
289 | 0 | "failure for used_iova %"PRIx64"\n", |
290 | 0 | used_phys); |
291 | 0 | return -EINVAL; |
292 | 0 | } |
293 | | |
294 | 0 | offset = used_phys & iotlb.addr_mask; |
295 | 0 | phys = iotlb.translated_addr + offset; |
296 | | |
297 | | /* |
298 | | * Distance from start of used ring until last byte of |
299 | | * IOMMU page. |
300 | | */ |
301 | 0 | s = iotlb.addr_mask - offset; |
302 | | /* |
303 | | * Size of used ring, or of the part of it until end |
304 | | * of IOMMU page. To avoid zero result, do the adding |
305 | | * outside of MIN(). |
306 | | */ |
307 | 0 | s = MIN(s, used_size - 1) + 1; |
308 | |
|
309 | 0 | vhost_dev_sync_region(dev, section, start_addr, end_addr, phys, |
310 | 0 | range_get_last(phys, s)); |
311 | 0 | used_size -= s; |
312 | 0 | used_phys += s; |
313 | 0 | } |
314 | 0 | } else { |
315 | 0 | vhost_dev_sync_region(dev, section, start_addr, |
316 | 0 | end_addr, vq->used_phys, |
317 | 0 | range_get_last(vq->used_phys, vq->used_size)); |
318 | 0 | } |
319 | 0 | } |
320 | 0 | return 0; |
321 | 0 | } |
322 | | |
323 | | static void vhost_log_sync(MemoryListener *listener, |
324 | | MemoryRegionSection *section) |
325 | 0 | { |
326 | 0 | struct vhost_dev *dev = container_of(listener, struct vhost_dev, |
327 | 0 | memory_listener); |
328 | 0 | vhost_sync_dirty_bitmap(dev, section, 0x0, ~0x0ULL); |
329 | 0 | } |
330 | | |
331 | | static void vhost_log_sync_range(struct vhost_dev *dev, |
332 | | hwaddr first, hwaddr last) |
333 | 0 | { |
334 | 0 | int i; |
335 | | /* FIXME: this is N^2 in number of sections */ |
336 | 0 | for (i = 0; i < dev->n_mem_sections; ++i) { |
337 | 0 | MemoryRegionSection *section = &dev->mem_sections[i]; |
338 | 0 | vhost_sync_dirty_bitmap(dev, section, first, last); |
339 | 0 | } |
340 | 0 | } |
341 | | |
342 | | static uint64_t vhost_get_log_size(struct vhost_dev *dev) |
343 | 0 | { |
344 | 0 | uint64_t log_size = 0; |
345 | 0 | int i; |
346 | 0 | for (i = 0; i < dev->mem->nregions; ++i) { |
347 | 0 | struct vhost_memory_region *reg = dev->mem->regions + i; |
348 | 0 | uint64_t last = range_get_last(reg->guest_phys_addr, |
349 | 0 | reg->memory_size); |
350 | 0 | log_size = MAX(log_size, last / VHOST_LOG_CHUNK + 1); |
351 | 0 | } |
352 | 0 | return log_size; |
353 | 0 | } |
354 | | |
355 | | static int vhost_set_backend_type(struct vhost_dev *dev, |
356 | | VhostBackendType backend_type) |
357 | 0 | { |
358 | 0 | int r = 0; |
359 | |
|
360 | 0 | switch (backend_type) { |
361 | 0 | #ifdef CONFIG_VHOST_KERNEL |
362 | 0 | case VHOST_BACKEND_TYPE_KERNEL: |
363 | 0 | dev->vhost_ops = &kernel_ops; |
364 | 0 | break; |
365 | 0 | #endif |
366 | 0 | #ifdef CONFIG_VHOST_USER |
367 | 0 | case VHOST_BACKEND_TYPE_USER: |
368 | 0 | dev->vhost_ops = &user_ops; |
369 | 0 | break; |
370 | 0 | #endif |
371 | 0 | #ifdef CONFIG_VHOST_VDPA |
372 | 0 | case VHOST_BACKEND_TYPE_VDPA: |
373 | 0 | dev->vhost_ops = &vdpa_ops; |
374 | 0 | break; |
375 | 0 | #endif |
376 | 0 | default: |
377 | 0 | error_report("Unknown vhost backend type"); |
378 | 0 | r = -1; |
379 | 0 | } |
380 | | |
381 | 0 | if (r == 0) { |
382 | 0 | assert(dev->vhost_ops->backend_type == backend_type); |
383 | 0 | } |
384 | |
|
385 | 0 | return r; |
386 | 0 | } |
387 | | |
388 | | static struct vhost_log *vhost_log_alloc(uint64_t size, bool share) |
389 | 0 | { |
390 | 0 | Error *err = NULL; |
391 | 0 | struct vhost_log *log; |
392 | 0 | uint64_t logsize = size * sizeof(*(log->log)); |
393 | 0 | int fd = -1; |
394 | |
|
395 | 0 | log = g_new0(struct vhost_log, 1); |
396 | 0 | if (share) { |
397 | 0 | log->log = qemu_memfd_alloc("vhost-log", logsize, |
398 | 0 | F_SEAL_GROW | F_SEAL_SHRINK | F_SEAL_SEAL, |
399 | 0 | &fd, &err); |
400 | 0 | if (err) { |
401 | 0 | error_report_err(err); |
402 | 0 | g_free(log); |
403 | 0 | return NULL; |
404 | 0 | } |
405 | 0 | memset(log->log, 0, logsize); |
406 | 0 | } else { |
407 | 0 | log->log = g_malloc0(logsize); |
408 | 0 | } |
409 | | |
410 | 0 | log->size = size; |
411 | 0 | log->refcnt = 1; |
412 | 0 | log->fd = fd; |
413 | |
|
414 | 0 | return log; |
415 | 0 | } |
416 | | |
417 | | static struct vhost_log *vhost_log_get(VhostBackendType backend_type, |
418 | | uint64_t size, bool share) |
419 | 0 | { |
420 | 0 | struct vhost_log *log; |
421 | |
|
422 | 0 | assert(backend_type > VHOST_BACKEND_TYPE_NONE); |
423 | 0 | assert(backend_type < VHOST_BACKEND_TYPE_MAX); |
424 | |
|
425 | 0 | log = share ? vhost_log_shm[backend_type] : vhost_log[backend_type]; |
426 | |
|
427 | 0 | if (!log || log->size != size) { |
428 | 0 | log = vhost_log_alloc(size, share); |
429 | 0 | if (share) { |
430 | 0 | vhost_log_shm[backend_type] = log; |
431 | 0 | } else { |
432 | 0 | vhost_log[backend_type] = log; |
433 | 0 | } |
434 | 0 | } else { |
435 | 0 | ++log->refcnt; |
436 | 0 | } |
437 | |
|
438 | 0 | return log; |
439 | 0 | } |
440 | | |
441 | | static void vhost_log_put(struct vhost_dev *dev, bool sync) |
442 | 0 | { |
443 | 0 | struct vhost_log *log = dev->log; |
444 | 0 | VhostBackendType backend_type; |
445 | |
|
446 | 0 | if (!log) { |
447 | 0 | return; |
448 | 0 | } |
449 | | |
450 | 0 | assert(dev->vhost_ops); |
451 | 0 | backend_type = dev->vhost_ops->backend_type; |
452 | |
|
453 | 0 | if (backend_type == VHOST_BACKEND_TYPE_NONE || |
454 | 0 | backend_type >= VHOST_BACKEND_TYPE_MAX) { |
455 | 0 | return; |
456 | 0 | } |
457 | | |
458 | 0 | --log->refcnt; |
459 | 0 | if (log->refcnt == 0) { |
460 | | /* Sync only the range covered by the old log */ |
461 | 0 | if (dev->log_size && sync) { |
462 | 0 | vhost_log_sync_range(dev, 0, dev->log_size * VHOST_LOG_CHUNK - 1); |
463 | 0 | } |
464 | |
|
465 | 0 | if (vhost_log[backend_type] == log) { |
466 | 0 | g_free(log->log); |
467 | 0 | vhost_log[backend_type] = NULL; |
468 | 0 | } else if (vhost_log_shm[backend_type] == log) { |
469 | 0 | qemu_memfd_free(log->log, log->size * sizeof(*(log->log)), |
470 | 0 | log->fd); |
471 | 0 | vhost_log_shm[backend_type] = NULL; |
472 | 0 | } |
473 | |
|
474 | 0 | g_free(log); |
475 | 0 | } |
476 | |
|
477 | 0 | vhost_dev_elect_mem_logger(dev, false); |
478 | 0 | dev->log = NULL; |
479 | 0 | dev->log_size = 0; |
480 | 0 | } |
481 | | |
482 | | static bool vhost_dev_log_is_shared(struct vhost_dev *dev) |
483 | 0 | { |
484 | 0 | return dev->vhost_ops->vhost_requires_shm_log && |
485 | 0 | dev->vhost_ops->vhost_requires_shm_log(dev); |
486 | 0 | } |
487 | | |
488 | | static inline void vhost_dev_log_resize(struct vhost_dev *dev, uint64_t size) |
489 | 0 | { |
490 | 0 | struct vhost_log *log = vhost_log_get(dev->vhost_ops->backend_type, |
491 | 0 | size, vhost_dev_log_is_shared(dev)); |
492 | 0 | uint64_t log_base = (uintptr_t)log->log; |
493 | 0 | int r; |
494 | | |
495 | | /* inform backend of log switching, this must be done before |
496 | | releasing the current log, to ensure no logging is lost */ |
497 | 0 | r = dev->vhost_ops->vhost_set_log_base(dev, log_base, log); |
498 | 0 | if (r < 0) { |
499 | 0 | VHOST_OPS_DEBUG(r, "vhost_set_log_base failed"); |
500 | 0 | } |
501 | |
|
502 | 0 | vhost_log_put(dev, true); |
503 | 0 | dev->log = log; |
504 | 0 | dev->log_size = size; |
505 | 0 | } |
506 | | |
507 | | static void *vhost_memory_map(struct vhost_dev *dev, hwaddr addr, |
508 | | hwaddr len, bool is_write) |
509 | 0 | { |
510 | 0 | hwaddr mapped_len = len; |
511 | 0 | void *res = address_space_map(dev->vdev->dma_as, addr, &mapped_len, |
512 | 0 | is_write, MEMTXATTRS_UNSPECIFIED); |
513 | 0 | if (!res) { |
514 | 0 | return NULL; |
515 | 0 | } |
516 | 0 | if (len != mapped_len) { |
517 | 0 | address_space_unmap(dev->vdev->dma_as, res, mapped_len, 0, 0); |
518 | 0 | return NULL; |
519 | 0 | } |
520 | 0 | return res; |
521 | 0 | } |
522 | | |
523 | | static void vhost_memory_unmap(struct vhost_dev *dev, void **buffer, |
524 | | hwaddr len, int is_write, |
525 | | hwaddr access_len) |
526 | 0 | { |
527 | 0 | if (!*buffer) { |
528 | 0 | return; |
529 | 0 | } |
530 | | |
531 | 0 | address_space_unmap(dev->vdev->dma_as, *buffer, len, is_write, |
532 | 0 | access_len); |
533 | 0 | *buffer = NULL; |
534 | 0 | } |
535 | | |
536 | | static void vhost_vrings_unmap(struct vhost_dev *dev, |
537 | | struct vhost_virtqueue *vq, bool touched) |
538 | 0 | { |
539 | 0 | if (vhost_dev_has_iommu(dev)) { |
540 | 0 | return; |
541 | 0 | } |
542 | | |
543 | 0 | vhost_memory_unmap(dev, &vq->used_user, vq->used_size, touched, |
544 | 0 | touched ? vq->used_size : 0); |
545 | 0 | vhost_memory_unmap(dev, &vq->avail_user, vq->avail_size, 0, |
546 | 0 | touched ? vq->avail_size : 0); |
547 | 0 | vhost_memory_unmap(dev, &vq->desc_user, vq->desc_size, 0, |
548 | 0 | touched ? vq->desc_size : 0); |
549 | 0 | } |
550 | | |
551 | | static int vhost_vrings_map(struct vhost_dev *dev, |
552 | | struct VirtIODevice *vdev, |
553 | | struct vhost_virtqueue *vq, |
554 | | unsigned idx) |
555 | 0 | { |
556 | 0 | vq->desc_size = virtio_queue_get_desc_size(vdev, idx); |
557 | 0 | vq->desc_phys = virtio_queue_get_desc_addr(vdev, idx); |
558 | 0 | vq->desc_user = NULL; |
559 | 0 | vq->avail_size = virtio_queue_get_avail_size(vdev, idx); |
560 | 0 | vq->avail_phys = virtio_queue_get_avail_addr(vdev, idx); |
561 | 0 | vq->avail_user = NULL; |
562 | 0 | vq->used_size = virtio_queue_get_used_size(vdev, idx); |
563 | 0 | vq->used_phys = virtio_queue_get_used_addr(vdev, idx); |
564 | 0 | vq->used_user = NULL; |
565 | |
|
566 | 0 | if (vq->desc_phys == 0) { |
567 | | /* Queue might not be ready for start */ |
568 | 0 | return 0; |
569 | 0 | } |
570 | | |
571 | 0 | if (vhost_dev_has_iommu(dev)) { |
572 | 0 | return 1; |
573 | 0 | } |
574 | | |
575 | 0 | vq->desc_user = vhost_memory_map(dev, vq->desc_phys, vq->desc_size, false); |
576 | 0 | if (!vq->desc_user) { |
577 | 0 | goto fail; |
578 | 0 | } |
579 | 0 | vq->avail_user = vhost_memory_map(dev, vq->avail_phys, vq->avail_size, |
580 | 0 | false); |
581 | 0 | if (!vq->avail_user) { |
582 | 0 | goto fail; |
583 | 0 | } |
584 | 0 | vq->used_user = vhost_memory_map(dev, vq->used_phys, vq->used_size, true); |
585 | 0 | if (!vq->used_user) { |
586 | 0 | goto fail; |
587 | 0 | } |
588 | | |
589 | 0 | return 1; |
590 | | |
591 | 0 | fail: |
592 | 0 | vhost_vrings_unmap(dev, vq, false); |
593 | 0 | return -ENOMEM; |
594 | 0 | } |
595 | | |
596 | | static int vhost_verify_ring_part_mapping(void *ring_hva, |
597 | | uint64_t ring_gpa, |
598 | | uint64_t ring_size, |
599 | | void *reg_hva, |
600 | | uint64_t reg_gpa, |
601 | | uint64_t reg_size) |
602 | 0 | { |
603 | 0 | uint64_t hva_ring_offset; |
604 | 0 | uint64_t ring_last = range_get_last(ring_gpa, ring_size); |
605 | 0 | uint64_t reg_last = range_get_last(reg_gpa, reg_size); |
606 | |
|
607 | 0 | if (ring_last < reg_gpa || ring_gpa > reg_last) { |
608 | 0 | return 0; |
609 | 0 | } |
610 | | /* check that whole ring's is mapped */ |
611 | 0 | if (ring_last > reg_last) { |
612 | 0 | return -ENOMEM; |
613 | 0 | } |
614 | | /* check that ring's MemoryRegion wasn't replaced */ |
615 | 0 | hva_ring_offset = ring_gpa - reg_gpa; |
616 | 0 | if (ring_hva != reg_hva + hva_ring_offset) { |
617 | 0 | return -EBUSY; |
618 | 0 | } |
619 | | |
620 | 0 | return 0; |
621 | 0 | } |
622 | | |
623 | | static int vhost_verify_ring_mappings(struct vhost_dev *dev, |
624 | | void *reg_hva, |
625 | | uint64_t reg_gpa, |
626 | | uint64_t reg_size) |
627 | 0 | { |
628 | 0 | int i, j; |
629 | 0 | int r = 0; |
630 | 0 | const char *part_name[] = { |
631 | 0 | "descriptor table", |
632 | 0 | "available ring", |
633 | 0 | "used ring" |
634 | 0 | }; |
635 | |
|
636 | 0 | if (vhost_dev_has_iommu(dev)) { |
637 | 0 | return 0; |
638 | 0 | } |
639 | | |
640 | 0 | for (i = 0; i < dev->nvqs; ++i) { |
641 | 0 | struct vhost_virtqueue *vq = dev->vqs + i; |
642 | |
|
643 | 0 | if (vq->desc_phys == 0) { |
644 | 0 | continue; |
645 | 0 | } |
646 | | |
647 | 0 | j = 0; |
648 | 0 | r = vhost_verify_ring_part_mapping( |
649 | 0 | vq->desc_user, vq->desc_phys, vq->desc_size, |
650 | 0 | reg_hva, reg_gpa, reg_size); |
651 | 0 | if (r) { |
652 | 0 | break; |
653 | 0 | } |
654 | | |
655 | 0 | j++; |
656 | 0 | r = vhost_verify_ring_part_mapping( |
657 | 0 | vq->avail_user, vq->avail_phys, vq->avail_size, |
658 | 0 | reg_hva, reg_gpa, reg_size); |
659 | 0 | if (r) { |
660 | 0 | break; |
661 | 0 | } |
662 | | |
663 | 0 | j++; |
664 | 0 | r = vhost_verify_ring_part_mapping( |
665 | 0 | vq->used_user, vq->used_phys, vq->used_size, |
666 | 0 | reg_hva, reg_gpa, reg_size); |
667 | 0 | if (r) { |
668 | 0 | break; |
669 | 0 | } |
670 | 0 | } |
671 | |
|
672 | 0 | if (r == -ENOMEM) { |
673 | 0 | error_report("Unable to map %s for ring %d", part_name[j], i); |
674 | 0 | } else if (r == -EBUSY) { |
675 | 0 | error_report("%s relocated for ring %d", part_name[j], i); |
676 | 0 | } |
677 | 0 | return r; |
678 | 0 | } |
679 | | |
680 | | /* |
681 | | * vhost_section: identify sections needed for vhost access |
682 | | * |
683 | | * We only care about RAM sections here (where virtqueue and guest |
684 | | * internals accessed by virtio might live). |
685 | | */ |
686 | | static bool vhost_section(struct vhost_dev *dev, MemoryRegionSection *section) |
687 | 0 | { |
688 | 0 | MemoryRegion *mr = section->mr; |
689 | |
|
690 | 0 | if (memory_region_is_ram(mr) && !memory_region_is_rom(mr)) { |
691 | 0 | uint8_t dirty_mask = memory_region_get_dirty_log_mask(mr); |
692 | 0 | uint8_t handled_dirty; |
693 | | |
694 | | /* |
695 | | * Kernel based vhost doesn't handle any block which is doing |
696 | | * dirty-tracking other than migration for which it has |
697 | | * specific logging support. However for TCG the kernel never |
698 | | * gets involved anyway so we can also ignore it's |
699 | | * self-modiying code detection flags. However a vhost-user |
700 | | * client could still confuse a TCG guest if it re-writes |
701 | | * executable memory that has already been translated. |
702 | | */ |
703 | 0 | handled_dirty = (1 << DIRTY_MEMORY_MIGRATION) | |
704 | 0 | (1 << DIRTY_MEMORY_CODE); |
705 | |
|
706 | 0 | if (dirty_mask & ~handled_dirty) { |
707 | 0 | trace_vhost_reject_section(mr->name, 1); |
708 | 0 | return false; |
709 | 0 | } |
710 | | |
711 | | /* |
712 | | * Some backends (like vhost-user) can only handle memory regions |
713 | | * that have an fd (can be mapped into a different process). Filter |
714 | | * the ones without an fd out, if requested. Also make sure that |
715 | | * this region is mapped as shared so that the vhost backend can |
716 | | * observe modifications to this region, otherwise we consider it |
717 | | * private. |
718 | | */ |
719 | 0 | if ((memory_region_get_fd(section->mr) < 0 || |
720 | 0 | !qemu_ram_is_shared(section->mr->ram_block)) && |
721 | 0 | dev->vhost_ops->vhost_no_private_memslots && |
722 | 0 | dev->vhost_ops->vhost_no_private_memslots(dev)) { |
723 | 0 | trace_vhost_reject_section(mr->name, 2); |
724 | 0 | return false; |
725 | 0 | } |
726 | | |
727 | 0 | trace_vhost_section(mr->name); |
728 | 0 | return true; |
729 | 0 | } else { |
730 | 0 | trace_vhost_reject_section(mr->name, 3); |
731 | 0 | return false; |
732 | 0 | } |
733 | 0 | } |
734 | | |
735 | | static void vhost_begin(MemoryListener *listener) |
736 | 0 | { |
737 | 0 | struct vhost_dev *dev = container_of(listener, struct vhost_dev, |
738 | 0 | memory_listener); |
739 | 0 | dev->tmp_sections = NULL; |
740 | 0 | dev->n_tmp_sections = 0; |
741 | 0 | } |
742 | | |
743 | | static void vhost_commit(MemoryListener *listener) |
744 | 0 | { |
745 | 0 | struct vhost_dev *dev = container_of(listener, struct vhost_dev, |
746 | 0 | memory_listener); |
747 | 0 | MemoryRegionSection *old_sections; |
748 | 0 | int n_old_sections; |
749 | 0 | uint64_t log_size; |
750 | 0 | size_t regions_size; |
751 | 0 | int r; |
752 | 0 | int i; |
753 | 0 | bool changed = false; |
754 | | |
755 | | /* Note we can be called before the device is started, but then |
756 | | * starting the device calls set_mem_table, so we need to have |
757 | | * built the data structures. |
758 | | */ |
759 | 0 | old_sections = dev->mem_sections; |
760 | 0 | n_old_sections = dev->n_mem_sections; |
761 | 0 | dev->mem_sections = dev->tmp_sections; |
762 | 0 | dev->n_mem_sections = dev->n_tmp_sections; |
763 | |
|
764 | 0 | if (dev->n_mem_sections != n_old_sections) { |
765 | 0 | changed = true; |
766 | 0 | } else { |
767 | | /* Same size, lets check the contents */ |
768 | 0 | for (i = 0; i < n_old_sections; i++) { |
769 | 0 | if (!MemoryRegionSection_eq(&old_sections[i], |
770 | 0 | &dev->mem_sections[i])) { |
771 | 0 | changed = true; |
772 | 0 | break; |
773 | 0 | } |
774 | 0 | } |
775 | 0 | } |
776 | |
|
777 | 0 | trace_vhost_commit(dev->started, changed); |
778 | 0 | if (!changed) { |
779 | 0 | goto out; |
780 | 0 | } |
781 | | |
782 | | /* Rebuild the regions list from the new sections list */ |
783 | 0 | regions_size = offsetof(struct vhost_memory, regions) + |
784 | 0 | dev->n_mem_sections * sizeof dev->mem->regions[0]; |
785 | 0 | dev->mem = g_realloc(dev->mem, regions_size); |
786 | 0 | dev->mem->nregions = dev->n_mem_sections; |
787 | |
|
788 | 0 | for (i = 0; i < dev->n_mem_sections; i++) { |
789 | 0 | struct vhost_memory_region *cur_vmr = dev->mem->regions + i; |
790 | 0 | struct MemoryRegionSection *mrs = dev->mem_sections + i; |
791 | |
|
792 | 0 | cur_vmr->guest_phys_addr = mrs->offset_within_address_space; |
793 | 0 | cur_vmr->memory_size = int128_get64(mrs->size); |
794 | 0 | cur_vmr->userspace_addr = |
795 | 0 | (uintptr_t)memory_region_get_ram_ptr(mrs->mr) + |
796 | 0 | mrs->offset_within_region; |
797 | 0 | cur_vmr->flags_padding = 0; |
798 | 0 | } |
799 | |
|
800 | 0 | if (!dev->started) { |
801 | 0 | goto out; |
802 | 0 | } |
803 | | |
804 | 0 | for (i = 0; i < dev->mem->nregions; i++) { |
805 | 0 | if (vhost_verify_ring_mappings(dev, |
806 | 0 | (void *)(uintptr_t)dev->mem->regions[i].userspace_addr, |
807 | 0 | dev->mem->regions[i].guest_phys_addr, |
808 | 0 | dev->mem->regions[i].memory_size)) { |
809 | 0 | virtio_error(dev->vdev, |
810 | 0 | "Verify ring failure on region %d", i); |
811 | 0 | goto out; |
812 | 0 | } |
813 | 0 | } |
814 | | |
815 | 0 | if (!dev->log_enabled) { |
816 | 0 | r = dev->vhost_ops->vhost_set_mem_table(dev, dev->mem); |
817 | 0 | if (r < 0) { |
818 | 0 | VHOST_OPS_DEBUG(r, "vhost_set_mem_table failed"); |
819 | 0 | } |
820 | 0 | goto out; |
821 | 0 | } |
822 | 0 | log_size = vhost_get_log_size(dev); |
823 | | /* We allocate an extra 4K bytes to log, |
824 | | * to reduce the * number of reallocations. */ |
825 | 0 | #define VHOST_LOG_BUFFER (0x1000 / sizeof *dev->log) |
826 | | /* To log more, must increase log size before table update. */ |
827 | 0 | if (dev->log_size < log_size) { |
828 | 0 | vhost_dev_log_resize(dev, log_size + VHOST_LOG_BUFFER); |
829 | 0 | } |
830 | 0 | r = dev->vhost_ops->vhost_set_mem_table(dev, dev->mem); |
831 | 0 | if (r < 0) { |
832 | 0 | VHOST_OPS_DEBUG(r, "vhost_set_mem_table failed"); |
833 | 0 | } |
834 | | /* To log less, can only decrease log size after table update. */ |
835 | 0 | if (dev->log_size > log_size + VHOST_LOG_BUFFER) { |
836 | 0 | vhost_dev_log_resize(dev, log_size); |
837 | 0 | } |
838 | |
|
839 | 0 | out: |
840 | | /* Deref the old list of sections, this must happen _after_ the |
841 | | * vhost_set_mem_table to ensure the client isn't still using the |
842 | | * section we're about to unref. |
843 | | */ |
844 | 0 | while (n_old_sections--) { |
845 | 0 | memory_region_unref(old_sections[n_old_sections].mr); |
846 | 0 | } |
847 | 0 | g_free(old_sections); |
848 | 0 | } |
849 | | |
850 | | /* Adds the section data to the tmp_section structure. |
851 | | * It relies on the listener calling us in memory address order |
852 | | * and for each region (via the _add and _nop methods) to |
853 | | * join neighbours. |
854 | | */ |
855 | | static void vhost_region_add_section(struct vhost_dev *dev, |
856 | | MemoryRegionSection *section) |
857 | 0 | { |
858 | 0 | bool need_add = true; |
859 | 0 | uint64_t mrs_size = int128_get64(section->size); |
860 | 0 | uint64_t mrs_gpa = section->offset_within_address_space; |
861 | 0 | uintptr_t mrs_host = (uintptr_t)memory_region_get_ram_ptr(section->mr) + |
862 | 0 | section->offset_within_region; |
863 | 0 | RAMBlock *mrs_rb = section->mr->ram_block; |
864 | |
|
865 | 0 | trace_vhost_region_add_section(section->mr->name, mrs_gpa, mrs_size, |
866 | 0 | mrs_host); |
867 | |
|
868 | 0 | if (dev->vhost_ops->backend_type == VHOST_BACKEND_TYPE_USER) { |
869 | | /* Round the section to it's page size */ |
870 | | /* First align the start down to a page boundary */ |
871 | 0 | size_t mrs_page = qemu_ram_pagesize(mrs_rb); |
872 | 0 | uint64_t alignage = mrs_host & (mrs_page - 1); |
873 | 0 | if (alignage) { |
874 | 0 | mrs_host -= alignage; |
875 | 0 | mrs_size += alignage; |
876 | 0 | mrs_gpa -= alignage; |
877 | 0 | } |
878 | | /* Now align the size up to a page boundary */ |
879 | 0 | alignage = mrs_size & (mrs_page - 1); |
880 | 0 | if (alignage) { |
881 | 0 | mrs_size += mrs_page - alignage; |
882 | 0 | } |
883 | 0 | trace_vhost_region_add_section_aligned(section->mr->name, mrs_gpa, |
884 | 0 | mrs_size, mrs_host); |
885 | 0 | } |
886 | |
|
887 | 0 | if (dev->n_tmp_sections && !section->unmergeable) { |
888 | | /* Since we already have at least one section, lets see if |
889 | | * this extends it; since we're scanning in order, we only |
890 | | * have to look at the last one, and the FlatView that calls |
891 | | * us shouldn't have overlaps. |
892 | | */ |
893 | 0 | MemoryRegionSection *prev_sec = dev->tmp_sections + |
894 | 0 | (dev->n_tmp_sections - 1); |
895 | 0 | uint64_t prev_gpa_start = prev_sec->offset_within_address_space; |
896 | 0 | uint64_t prev_size = int128_get64(prev_sec->size); |
897 | 0 | uint64_t prev_gpa_end = range_get_last(prev_gpa_start, prev_size); |
898 | 0 | uint64_t prev_host_start = |
899 | 0 | (uintptr_t)memory_region_get_ram_ptr(prev_sec->mr) + |
900 | 0 | prev_sec->offset_within_region; |
901 | 0 | uint64_t prev_host_end = range_get_last(prev_host_start, prev_size); |
902 | |
|
903 | 0 | if (mrs_gpa <= (prev_gpa_end + 1)) { |
904 | | /* OK, looks like overlapping/intersecting - it's possible that |
905 | | * the rounding to page sizes has made them overlap, but they should |
906 | | * match up in the same RAMBlock if they do. |
907 | | */ |
908 | 0 | if (mrs_gpa < prev_gpa_start) { |
909 | 0 | error_report("%s:Section '%s' rounded to %"PRIx64 |
910 | 0 | " prior to previous '%s' %"PRIx64, |
911 | 0 | __func__, section->mr->name, mrs_gpa, |
912 | 0 | prev_sec->mr->name, prev_gpa_start); |
913 | | /* A way to cleanly fail here would be better */ |
914 | 0 | return; |
915 | 0 | } |
916 | | /* Offset from the start of the previous GPA to this GPA */ |
917 | 0 | size_t offset = mrs_gpa - prev_gpa_start; |
918 | |
|
919 | 0 | if (prev_host_start + offset == mrs_host && |
920 | 0 | section->mr == prev_sec->mr && !prev_sec->unmergeable) { |
921 | 0 | uint64_t max_end = MAX(prev_host_end, mrs_host + mrs_size); |
922 | 0 | need_add = false; |
923 | 0 | prev_sec->offset_within_address_space = |
924 | 0 | MIN(prev_gpa_start, mrs_gpa); |
925 | 0 | prev_sec->offset_within_region = |
926 | 0 | MIN(prev_host_start, mrs_host) - |
927 | 0 | (uintptr_t)memory_region_get_ram_ptr(prev_sec->mr); |
928 | 0 | prev_sec->size = int128_make64(max_end - MIN(prev_host_start, |
929 | 0 | mrs_host)); |
930 | 0 | trace_vhost_region_add_section_merge(section->mr->name, |
931 | 0 | int128_get64(prev_sec->size), |
932 | 0 | prev_sec->offset_within_address_space, |
933 | 0 | prev_sec->offset_within_region); |
934 | 0 | } else { |
935 | | /* adjoining regions are fine, but overlapping ones with |
936 | | * different blocks/offsets shouldn't happen |
937 | | */ |
938 | 0 | if (mrs_gpa != prev_gpa_end + 1) { |
939 | 0 | error_report("%s: Overlapping but not coherent sections " |
940 | 0 | "at %"PRIx64, |
941 | 0 | __func__, mrs_gpa); |
942 | 0 | return; |
943 | 0 | } |
944 | 0 | } |
945 | 0 | } |
946 | 0 | } |
947 | | |
948 | 0 | if (need_add) { |
949 | 0 | ++dev->n_tmp_sections; |
950 | 0 | dev->tmp_sections = g_renew(MemoryRegionSection, dev->tmp_sections, |
951 | 0 | dev->n_tmp_sections); |
952 | 0 | dev->tmp_sections[dev->n_tmp_sections - 1] = *section; |
953 | | /* The flatview isn't stable and we don't use it, making it NULL |
954 | | * means we can memcmp the list. |
955 | | */ |
956 | 0 | dev->tmp_sections[dev->n_tmp_sections - 1].fv = NULL; |
957 | 0 | memory_region_ref(section->mr); |
958 | 0 | } |
959 | 0 | } |
960 | | |
961 | | /* Used for both add and nop callbacks */ |
962 | | static void vhost_region_addnop(MemoryListener *listener, |
963 | | MemoryRegionSection *section) |
964 | 0 | { |
965 | 0 | struct vhost_dev *dev = container_of(listener, struct vhost_dev, |
966 | 0 | memory_listener); |
967 | |
|
968 | 0 | if (!vhost_section(dev, section)) { |
969 | 0 | return; |
970 | 0 | } |
971 | 0 | vhost_region_add_section(dev, section); |
972 | 0 | } |
973 | | |
974 | | static int vhost_update_device_iotlb(struct vhost_dev *dev, |
975 | | uint64_t iova, |
976 | | uint64_t paddr, |
977 | | uint64_t uaddr, |
978 | | uint64_t len, |
979 | | IOMMUAccessFlags perm) |
980 | 0 | { |
981 | 0 | bool phys = dev->vhost_ops->vhost_phys_iotlb_msg && |
982 | 0 | dev->vhost_ops->vhost_phys_iotlb_msg(dev); |
983 | 0 | struct vhost_iotlb_msg imsg; |
984 | |
|
985 | 0 | imsg.iova = iova; |
986 | 0 | imsg.uaddr = phys ? paddr : uaddr; |
987 | 0 | imsg.size = len; |
988 | 0 | imsg.type = VHOST_IOTLB_UPDATE; |
989 | |
|
990 | 0 | switch (perm) { |
991 | 0 | case IOMMU_RO: |
992 | 0 | imsg.perm = VHOST_ACCESS_RO; |
993 | 0 | break; |
994 | 0 | case IOMMU_WO: |
995 | 0 | imsg.perm = VHOST_ACCESS_WO; |
996 | 0 | break; |
997 | 0 | case IOMMU_RW: |
998 | 0 | imsg.perm = VHOST_ACCESS_RW; |
999 | 0 | break; |
1000 | 0 | default: |
1001 | 0 | return -EINVAL; |
1002 | 0 | } |
1003 | | |
1004 | 0 | if (dev->vhost_ops && dev->vhost_ops->vhost_send_device_iotlb_msg) { |
1005 | 0 | return dev->vhost_ops->vhost_send_device_iotlb_msg(dev, &imsg); |
1006 | 0 | } |
1007 | | |
1008 | 0 | return -ENODEV; |
1009 | 0 | } |
1010 | | |
1011 | | static int vhost_invalidate_device_iotlb(struct vhost_dev *dev, |
1012 | | uint64_t iova, uint64_t len) |
1013 | 0 | { |
1014 | 0 | struct vhost_iotlb_msg imsg; |
1015 | |
|
1016 | 0 | imsg.iova = iova; |
1017 | 0 | imsg.size = len; |
1018 | 0 | imsg.type = VHOST_IOTLB_INVALIDATE; |
1019 | |
|
1020 | 0 | if (dev->vhost_ops && dev->vhost_ops->vhost_send_device_iotlb_msg) { |
1021 | 0 | return dev->vhost_ops->vhost_send_device_iotlb_msg(dev, &imsg); |
1022 | 0 | } |
1023 | | |
1024 | 0 | return -ENODEV; |
1025 | 0 | } |
1026 | | |
1027 | | int vhost_handle_iotlb_msg(struct vhost_dev *dev, struct vhost_iotlb_msg *imsg) |
1028 | 0 | { |
1029 | 0 | int ret = 0; |
1030 | |
|
1031 | 0 | if (unlikely(!dev->vdev)) { |
1032 | 0 | error_report("Unexpected IOTLB message when virtio device is stopped"); |
1033 | 0 | return -EINVAL; |
1034 | 0 | } |
1035 | | |
1036 | 0 | switch (imsg->type) { |
1037 | 0 | case VHOST_IOTLB_MISS: |
1038 | 0 | ret = vhost_device_iotlb_miss(dev, imsg->iova, |
1039 | 0 | imsg->perm != VHOST_ACCESS_RO); |
1040 | 0 | break; |
1041 | 0 | case VHOST_IOTLB_ACCESS_FAIL: |
1042 | | /* FIXME: report device iotlb error */ |
1043 | 0 | error_report("Access failure IOTLB message type not supported"); |
1044 | 0 | ret = -ENOTSUP; |
1045 | 0 | break; |
1046 | 0 | case VHOST_IOTLB_UPDATE: |
1047 | 0 | case VHOST_IOTLB_INVALIDATE: |
1048 | 0 | default: |
1049 | 0 | error_report("Unexpected IOTLB message type"); |
1050 | 0 | ret = -EINVAL; |
1051 | 0 | break; |
1052 | 0 | } |
1053 | | |
1054 | 0 | return ret; |
1055 | 0 | } |
1056 | | |
1057 | | static void vhost_iommu_unmap_notify(IOMMUNotifier *n, IOMMUTLBEntry *iotlb) |
1058 | 0 | { |
1059 | 0 | struct vhost_iommu *iommu = container_of(n, struct vhost_iommu, n); |
1060 | 0 | struct vhost_dev *hdev = iommu->hdev; |
1061 | 0 | hwaddr iova = iotlb->iova + iommu->iommu_offset; |
1062 | |
|
1063 | 0 | if (vhost_invalidate_device_iotlb(hdev, iova, iotlb->addr_mask + 1)) { |
1064 | 0 | error_report("Fail to invalidate device iotlb"); |
1065 | 0 | } |
1066 | 0 | } |
1067 | | |
1068 | | static void vhost_iommu_region_add(MemoryListener *listener, |
1069 | | MemoryRegionSection *section) |
1070 | 0 | { |
1071 | 0 | struct vhost_dev *dev = container_of(listener, struct vhost_dev, |
1072 | 0 | iommu_listener); |
1073 | 0 | struct vhost_iommu *iommu; |
1074 | 0 | Int128 end; |
1075 | 0 | int iommu_idx; |
1076 | 0 | IOMMUMemoryRegion *iommu_mr; |
1077 | |
|
1078 | 0 | if (!memory_region_is_iommu(section->mr)) { |
1079 | 0 | return; |
1080 | 0 | } |
1081 | | |
1082 | 0 | iommu_mr = IOMMU_MEMORY_REGION(section->mr); |
1083 | |
|
1084 | 0 | iommu = g_malloc0(sizeof(*iommu)); |
1085 | 0 | end = int128_add(int128_make64(section->offset_within_region), |
1086 | 0 | section->size); |
1087 | 0 | end = int128_sub(end, int128_one()); |
1088 | 0 | iommu_idx = memory_region_iommu_attrs_to_index(iommu_mr, |
1089 | 0 | MEMTXATTRS_UNSPECIFIED); |
1090 | 0 | iommu_notifier_init(&iommu->n, vhost_iommu_unmap_notify, |
1091 | 0 | dev->vdev->device_iotlb_enabled ? |
1092 | 0 | IOMMU_NOTIFIER_DEVIOTLB_UNMAP : |
1093 | 0 | IOMMU_NOTIFIER_UNMAP, |
1094 | 0 | section->offset_within_region, |
1095 | 0 | int128_get64(end), |
1096 | 0 | iommu_idx); |
1097 | 0 | iommu->mr = section->mr; |
1098 | 0 | iommu->iommu_offset = section->offset_within_address_space - |
1099 | 0 | section->offset_within_region; |
1100 | 0 | iommu->hdev = dev; |
1101 | 0 | memory_region_register_iommu_notifier(section->mr, &iommu->n, |
1102 | 0 | &error_fatal); |
1103 | 0 | QLIST_INSERT_HEAD(&dev->iommu_list, iommu, iommu_next); |
1104 | | /* TODO: can replay help performance here? */ |
1105 | 0 | } |
1106 | | |
1107 | | static void vhost_iommu_region_del(MemoryListener *listener, |
1108 | | MemoryRegionSection *section) |
1109 | 0 | { |
1110 | 0 | struct vhost_dev *dev = container_of(listener, struct vhost_dev, |
1111 | 0 | iommu_listener); |
1112 | 0 | struct vhost_iommu *iommu; |
1113 | |
|
1114 | 0 | if (!memory_region_is_iommu(section->mr)) { |
1115 | 0 | return; |
1116 | 0 | } |
1117 | | |
1118 | 0 | QLIST_FOREACH(iommu, &dev->iommu_list, iommu_next) { |
1119 | 0 | if (iommu->mr == section->mr && |
1120 | 0 | iommu->n.start == section->offset_within_region) { |
1121 | 0 | memory_region_unregister_iommu_notifier(iommu->mr, |
1122 | 0 | &iommu->n); |
1123 | 0 | QLIST_REMOVE(iommu, iommu_next); |
1124 | 0 | g_free(iommu); |
1125 | 0 | break; |
1126 | 0 | } |
1127 | 0 | } |
1128 | 0 | } |
1129 | | |
1130 | | void vhost_toggle_device_iotlb(VirtIODevice *vdev) |
1131 | 0 | { |
1132 | 0 | VirtioDeviceClass *vdc = VIRTIO_DEVICE_GET_CLASS(vdev); |
1133 | 0 | struct vhost_dev *dev; |
1134 | 0 | struct vhost_iommu *iommu; |
1135 | |
|
1136 | 0 | if (vdev->vhost_started) { |
1137 | 0 | dev = vdc->get_vhost(vdev); |
1138 | 0 | } else { |
1139 | 0 | return; |
1140 | 0 | } |
1141 | | |
1142 | 0 | QLIST_FOREACH(iommu, &dev->iommu_list, iommu_next) { |
1143 | 0 | memory_region_unregister_iommu_notifier(iommu->mr, &iommu->n); |
1144 | 0 | iommu->n.notifier_flags = vdev->device_iotlb_enabled ? |
1145 | 0 | IOMMU_NOTIFIER_DEVIOTLB_UNMAP : IOMMU_NOTIFIER_UNMAP; |
1146 | 0 | memory_region_register_iommu_notifier(iommu->mr, &iommu->n, |
1147 | 0 | &error_fatal); |
1148 | 0 | } |
1149 | 0 | } |
1150 | | |
1151 | | static int vhost_virtqueue_set_addr(struct vhost_dev *dev, |
1152 | | struct vhost_virtqueue *vq, |
1153 | | unsigned idx, bool enable_log) |
1154 | 0 | { |
1155 | 0 | bool phys = dev->vhost_ops->vhost_phys_vring_addr && |
1156 | 0 | dev->vhost_ops->vhost_phys_vring_addr(dev); |
1157 | 0 | struct vhost_vring_addr addr; |
1158 | 0 | int r; |
1159 | 0 | memset(&addr, 0, sizeof(struct vhost_vring_addr)); |
1160 | |
|
1161 | 0 | if (phys || vhost_dev_has_iommu(dev)) { |
1162 | 0 | addr.desc_user_addr = (uint64_t)(unsigned long)vq->desc_phys; |
1163 | 0 | addr.avail_user_addr = (uint64_t)(unsigned long)vq->avail_phys; |
1164 | 0 | addr.used_user_addr = (uint64_t)(unsigned long)vq->used_phys; |
1165 | 0 | } else { |
1166 | 0 | addr.desc_user_addr = (uint64_t)(unsigned long)vq->desc_user; |
1167 | 0 | addr.avail_user_addr = (uint64_t)(unsigned long)vq->avail_user; |
1168 | 0 | addr.used_user_addr = (uint64_t)(unsigned long)vq->used_user; |
1169 | 0 | } |
1170 | 0 | addr.index = idx; |
1171 | 0 | addr.log_guest_addr = vq->used_phys; |
1172 | 0 | addr.flags = enable_log ? (1 << VHOST_VRING_F_LOG) : 0; |
1173 | 0 | r = dev->vhost_ops->vhost_set_vring_addr(dev, &addr); |
1174 | 0 | if (r < 0) { |
1175 | 0 | VHOST_OPS_DEBUG(r, "vhost_set_vring_addr failed"); |
1176 | 0 | } |
1177 | 0 | return r; |
1178 | 0 | } |
1179 | | |
1180 | | static int vhost_dev_set_features(struct vhost_dev *dev, |
1181 | | bool enable_log) |
1182 | 0 | { |
1183 | 0 | uint64_t features[VIRTIO_FEATURES_NU64S]; |
1184 | 0 | int r; |
1185 | |
|
1186 | 0 | virtio_features_copy(features, dev->acked_features_ex); |
1187 | 0 | if (enable_log) { |
1188 | 0 | virtio_add_feature_ex(features, VHOST_F_LOG_ALL); |
1189 | 0 | } |
1190 | 0 | if (!vhost_dev_has_iommu(dev)) { |
1191 | 0 | virtio_clear_feature_ex(features, VIRTIO_F_IOMMU_PLATFORM); |
1192 | 0 | } |
1193 | 0 | if (dev->vhost_ops->vhost_force_iommu) { |
1194 | 0 | if (dev->vhost_ops->vhost_force_iommu(dev) == true) { |
1195 | 0 | virtio_add_feature_ex(features, VIRTIO_F_IOMMU_PLATFORM); |
1196 | 0 | } |
1197 | 0 | } |
1198 | |
|
1199 | 0 | if (virtio_features_use_ex(features) && |
1200 | 0 | !dev->vhost_ops->vhost_set_features_ex) { |
1201 | 0 | r = -EINVAL; |
1202 | 0 | VHOST_OPS_DEBUG(r, "extended features without device support"); |
1203 | 0 | goto out; |
1204 | 0 | } |
1205 | | |
1206 | 0 | if (dev->vhost_ops->vhost_set_features_ex) { |
1207 | 0 | r = dev->vhost_ops->vhost_set_features_ex(dev, features); |
1208 | 0 | } else { |
1209 | 0 | r = dev->vhost_ops->vhost_set_features(dev, features[0]); |
1210 | 0 | } |
1211 | 0 | if (r < 0) { |
1212 | 0 | VHOST_OPS_DEBUG(r, "vhost_set_features failed"); |
1213 | 0 | goto out; |
1214 | 0 | } |
1215 | 0 | if (dev->vhost_ops->vhost_set_backend_cap) { |
1216 | 0 | r = dev->vhost_ops->vhost_set_backend_cap(dev); |
1217 | 0 | if (r < 0) { |
1218 | 0 | VHOST_OPS_DEBUG(r, "vhost_set_backend_cap failed"); |
1219 | 0 | goto out; |
1220 | 0 | } |
1221 | 0 | } |
1222 | | |
1223 | 0 | out: |
1224 | 0 | return r; |
1225 | 0 | } |
1226 | | |
1227 | | static int vhost_dev_set_log(struct vhost_dev *dev, bool enable_log) |
1228 | 0 | { |
1229 | 0 | int r, i, idx; |
1230 | 0 | hwaddr addr; |
1231 | |
|
1232 | 0 | r = vhost_dev_set_features(dev, enable_log); |
1233 | 0 | if (r < 0) { |
1234 | 0 | goto err_features; |
1235 | 0 | } |
1236 | 0 | for (i = 0; i < dev->nvqs; ++i) { |
1237 | 0 | idx = dev->vhost_ops->vhost_get_vq_index(dev, dev->vq_index + i); |
1238 | 0 | addr = virtio_queue_get_desc_addr(dev->vdev, idx); |
1239 | 0 | if (!addr) { |
1240 | | /* |
1241 | | * The queue might not be ready for start. If this |
1242 | | * is the case there is no reason to continue the process. |
1243 | | * The similar logic is used by the vhost_virtqueue_start() |
1244 | | * routine. |
1245 | | */ |
1246 | 0 | continue; |
1247 | 0 | } |
1248 | 0 | r = vhost_virtqueue_set_addr(dev, dev->vqs + i, idx, |
1249 | 0 | enable_log); |
1250 | 0 | if (r < 0) { |
1251 | 0 | goto err_vq; |
1252 | 0 | } |
1253 | 0 | } |
1254 | | |
1255 | | /* |
1256 | | * At log start we select our vhost_device logger that will scan the |
1257 | | * memory sections and skip for the others. This is possible because |
1258 | | * the log is shared amongst all vhost devices for a given type of |
1259 | | * backend. |
1260 | | */ |
1261 | 0 | vhost_dev_elect_mem_logger(dev, enable_log); |
1262 | |
|
1263 | 0 | return 0; |
1264 | 0 | err_vq: |
1265 | 0 | for (; i >= 0; --i) { |
1266 | 0 | idx = dev->vhost_ops->vhost_get_vq_index(dev, dev->vq_index + i); |
1267 | 0 | addr = virtio_queue_get_desc_addr(dev->vdev, idx); |
1268 | 0 | if (!addr) { |
1269 | 0 | continue; |
1270 | 0 | } |
1271 | 0 | vhost_virtqueue_set_addr(dev, dev->vqs + i, idx, |
1272 | 0 | dev->log_enabled); |
1273 | 0 | } |
1274 | 0 | vhost_dev_set_features(dev, dev->log_enabled); |
1275 | 0 | err_features: |
1276 | 0 | return r; |
1277 | 0 | } |
1278 | | |
1279 | | static int vhost_migration_log(MemoryListener *listener, bool enable) |
1280 | 0 | { |
1281 | 0 | struct vhost_dev *dev = container_of(listener, struct vhost_dev, |
1282 | 0 | memory_listener); |
1283 | 0 | int r; |
1284 | 0 | if (enable == dev->log_enabled) { |
1285 | 0 | return 0; |
1286 | 0 | } |
1287 | 0 | if (!dev->started) { |
1288 | 0 | dev->log_enabled = enable; |
1289 | 0 | return 0; |
1290 | 0 | } |
1291 | | |
1292 | 0 | r = 0; |
1293 | 0 | if (!enable) { |
1294 | 0 | r = vhost_dev_set_log(dev, false); |
1295 | 0 | if (r < 0) { |
1296 | 0 | goto check_dev_state; |
1297 | 0 | } |
1298 | 0 | vhost_log_put(dev, false); |
1299 | 0 | } else { |
1300 | 0 | vhost_dev_log_resize(dev, vhost_get_log_size(dev)); |
1301 | 0 | r = vhost_dev_set_log(dev, true); |
1302 | 0 | if (r < 0) { |
1303 | 0 | goto check_dev_state; |
1304 | 0 | } |
1305 | 0 | } |
1306 | | |
1307 | 0 | check_dev_state: |
1308 | 0 | dev->log_enabled = enable; |
1309 | | /* |
1310 | | * vhost-user-* devices could change their state during log |
1311 | | * initialization due to disconnect. So check dev state after |
1312 | | * vhost communication. |
1313 | | */ |
1314 | 0 | if (!dev->started) { |
1315 | | /* |
1316 | | * Since device is in the stopped state, it is okay for |
1317 | | * migration. Return success. |
1318 | | */ |
1319 | 0 | r = 0; |
1320 | 0 | } |
1321 | 0 | if (r) { |
1322 | | /* An error occurred. */ |
1323 | 0 | dev->log_enabled = false; |
1324 | 0 | } |
1325 | |
|
1326 | 0 | return r; |
1327 | 0 | } |
1328 | | |
1329 | | static bool vhost_log_global_start(MemoryListener *listener, Error **errp) |
1330 | 0 | { |
1331 | 0 | int r; |
1332 | |
|
1333 | 0 | r = vhost_migration_log(listener, true); |
1334 | 0 | if (r < 0) { |
1335 | 0 | error_setg_errno(errp, -r, "vhost: Failed to start logging"); |
1336 | 0 | return false; |
1337 | 0 | } |
1338 | 0 | return true; |
1339 | 0 | } |
1340 | | |
1341 | | static void vhost_log_global_stop(MemoryListener *listener) |
1342 | 0 | { |
1343 | 0 | int r; |
1344 | |
|
1345 | 0 | r = vhost_migration_log(listener, false); |
1346 | 0 | if (r < 0) { |
1347 | | /* Not fatal, so report it, but take no further action */ |
1348 | 0 | warn_report("vhost: Failed to stop logging"); |
1349 | 0 | } |
1350 | 0 | } |
1351 | | |
1352 | | static void vhost_log_start(MemoryListener *listener, |
1353 | | MemoryRegionSection *section, |
1354 | | int old, int new) |
1355 | 0 | { |
1356 | | /* FIXME: implement */ |
1357 | 0 | } |
1358 | | |
1359 | | static void vhost_log_stop(MemoryListener *listener, |
1360 | | MemoryRegionSection *section, |
1361 | | int old, int new) |
1362 | 0 | { |
1363 | | /* FIXME: implement */ |
1364 | 0 | } |
1365 | | |
1366 | | /* The vhost driver natively knows how to handle the vrings of non |
1367 | | * cross-endian legacy devices and modern devices. Only legacy devices |
1368 | | * exposed to a bi-endian guest may require the vhost driver to use a |
1369 | | * specific endianness. |
1370 | | */ |
1371 | | static inline bool vhost_needs_vring_endian(VirtIODevice *vdev) |
1372 | 0 | { |
1373 | 0 | if (virtio_vdev_is_legacy(vdev)) { |
1374 | 0 | return vdev->device_endian == (HOST_BIG_ENDIAN |
1375 | 0 | ? VIRTIO_DEVICE_ENDIAN_LITTLE |
1376 | 0 | : VIRTIO_DEVICE_ENDIAN_BIG); |
1377 | 0 | } |
1378 | 0 | return false; |
1379 | 0 | } |
1380 | | |
1381 | | static int vhost_virtqueue_set_vring_endian_legacy(struct vhost_dev *dev, |
1382 | | bool is_big_endian, |
1383 | | int vhost_vq_index) |
1384 | 0 | { |
1385 | 0 | int r; |
1386 | 0 | struct vhost_vring_state s = { |
1387 | 0 | .index = vhost_vq_index, |
1388 | 0 | .num = is_big_endian |
1389 | 0 | }; |
1390 | |
|
1391 | 0 | r = dev->vhost_ops->vhost_set_vring_endian(dev, &s); |
1392 | 0 | if (r < 0) { |
1393 | 0 | VHOST_OPS_DEBUG(r, "vhost_set_vring_endian failed"); |
1394 | 0 | } |
1395 | 0 | return r; |
1396 | 0 | } |
1397 | | |
1398 | | static int vhost_memory_region_lookup(struct vhost_dev *hdev, |
1399 | | uint64_t gpa, uint64_t *uaddr, |
1400 | | uint64_t *len) |
1401 | 0 | { |
1402 | 0 | int i; |
1403 | |
|
1404 | 0 | for (i = 0; i < hdev->mem->nregions; i++) { |
1405 | 0 | struct vhost_memory_region *reg = hdev->mem->regions + i; |
1406 | |
|
1407 | 0 | if (gpa >= reg->guest_phys_addr && |
1408 | 0 | reg->guest_phys_addr + reg->memory_size > gpa) { |
1409 | 0 | *uaddr = reg->userspace_addr + gpa - reg->guest_phys_addr; |
1410 | 0 | *len = reg->guest_phys_addr + reg->memory_size - gpa; |
1411 | 0 | return 0; |
1412 | 0 | } |
1413 | 0 | } |
1414 | | |
1415 | 0 | return -EFAULT; |
1416 | 0 | } |
1417 | | |
1418 | | int vhost_device_iotlb_miss(struct vhost_dev *dev, uint64_t iova, int write) |
1419 | 0 | { |
1420 | 0 | IOMMUTLBEntry iotlb; |
1421 | 0 | uint64_t uaddr, len; |
1422 | 0 | int ret = -EFAULT; |
1423 | |
|
1424 | 0 | RCU_READ_LOCK_GUARD(); |
1425 | |
|
1426 | 0 | trace_vhost_iotlb_miss(dev, 1); |
1427 | |
|
1428 | 0 | iotlb = address_space_get_iotlb_entry(dev->vdev->dma_as, |
1429 | 0 | iova, write, |
1430 | 0 | MEMTXATTRS_UNSPECIFIED); |
1431 | 0 | if (iotlb.target_as != NULL) { |
1432 | 0 | ret = vhost_memory_region_lookup(dev, iotlb.translated_addr, |
1433 | 0 | &uaddr, &len); |
1434 | 0 | if (ret) { |
1435 | 0 | trace_vhost_iotlb_miss(dev, 3); |
1436 | 0 | error_report("Fail to lookup the translated address " |
1437 | 0 | "%"PRIx64, iotlb.translated_addr); |
1438 | 0 | goto out; |
1439 | 0 | } |
1440 | | |
1441 | 0 | len = MIN(iotlb.addr_mask + 1, len); |
1442 | 0 | iova = iova & ~iotlb.addr_mask; |
1443 | |
|
1444 | 0 | ret = vhost_update_device_iotlb(dev, iova, iotlb.translated_addr, uaddr, |
1445 | 0 | len, iotlb.perm); |
1446 | 0 | if (ret) { |
1447 | 0 | trace_vhost_iotlb_miss(dev, 4); |
1448 | 0 | error_report("Fail to update device iotlb"); |
1449 | 0 | goto out; |
1450 | 0 | } |
1451 | 0 | } |
1452 | | |
1453 | 0 | trace_vhost_iotlb_miss(dev, 2); |
1454 | |
|
1455 | 0 | out: |
1456 | 0 | return ret; |
1457 | 0 | } |
1458 | | |
1459 | | int vhost_virtqueue_start(struct vhost_dev *dev, |
1460 | | struct VirtIODevice *vdev, |
1461 | | struct vhost_virtqueue *vq, |
1462 | | unsigned idx) |
1463 | 0 | { |
1464 | 0 | BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev))); |
1465 | 0 | VirtioBusState *vbus = VIRTIO_BUS(qbus); |
1466 | 0 | VirtioBusClass *k = VIRTIO_BUS_GET_CLASS(vbus); |
1467 | 0 | int r; |
1468 | 0 | int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, idx); |
1469 | 0 | struct vhost_vring_file file = { |
1470 | 0 | .index = vhost_vq_index |
1471 | 0 | }; |
1472 | 0 | struct vhost_vring_state state = { |
1473 | 0 | .index = vhost_vq_index |
1474 | 0 | }; |
1475 | 0 | struct VirtQueue *vvq = virtio_get_queue(vdev, idx); |
1476 | |
|
1477 | 0 | trace_vhost_virtqueue_start_in(dev, vdev->name, idx); |
1478 | |
|
1479 | 0 | r = vhost_vrings_map(dev, vdev, vq, idx); |
1480 | 0 | if (r <= 0) { |
1481 | 0 | return r; |
1482 | 0 | } |
1483 | | |
1484 | 0 | vq->num = state.num = virtio_queue_get_num(vdev, idx); |
1485 | 0 | r = dev->vhost_ops->vhost_set_vring_num(dev, &state); |
1486 | 0 | if (r) { |
1487 | 0 | VHOST_OPS_DEBUG(r, "vhost_set_vring_num failed"); |
1488 | 0 | goto fail; |
1489 | 0 | } |
1490 | | |
1491 | 0 | state.num = virtio_queue_get_last_avail_idx(vdev, idx); |
1492 | 0 | r = dev->vhost_ops->vhost_set_vring_base(dev, &state); |
1493 | 0 | if (r) { |
1494 | 0 | VHOST_OPS_DEBUG(r, "vhost_set_vring_base failed"); |
1495 | 0 | goto fail; |
1496 | 0 | } |
1497 | | |
1498 | 0 | if (vhost_needs_vring_endian(vdev)) { |
1499 | 0 | r = vhost_virtqueue_set_vring_endian_legacy( |
1500 | 0 | dev, virtio_vdev_is_big_endian(vdev), vhost_vq_index); |
1501 | 0 | if (r) { |
1502 | 0 | goto fail; |
1503 | 0 | } |
1504 | 0 | } |
1505 | | |
1506 | 0 | r = vhost_virtqueue_set_addr(dev, vq, vhost_vq_index, dev->log_enabled); |
1507 | 0 | if (r < 0) { |
1508 | 0 | goto fail; |
1509 | 0 | } |
1510 | | |
1511 | 0 | file.fd = event_notifier_get_fd(virtio_queue_get_host_notifier(vvq)); |
1512 | 0 | r = dev->vhost_ops->vhost_set_vring_kick(dev, &file); |
1513 | 0 | if (r) { |
1514 | 0 | VHOST_OPS_DEBUG(r, "vhost_set_vring_kick failed"); |
1515 | 0 | goto fail; |
1516 | 0 | } |
1517 | | |
1518 | | /* Clear and discard previous events if any. */ |
1519 | 0 | event_notifier_test_and_clear(&vq->masked_notifier); |
1520 | | |
1521 | | /* Init vring in unmasked state, unless guest_notifier_mask |
1522 | | * will do it later. |
1523 | | */ |
1524 | 0 | if (!vdev->use_guest_notifier_mask) { |
1525 | 0 | r = vhost_virtqueue_mask(dev, vdev, idx, false); |
1526 | 0 | if (r < 0) { |
1527 | 0 | goto fail; |
1528 | 0 | } |
1529 | 0 | } |
1530 | | |
1531 | 0 | if (k->query_guest_notifiers && |
1532 | 0 | k->query_guest_notifiers(qbus->parent) && |
1533 | 0 | virtio_queue_vector(vdev, idx) == VIRTIO_NO_VECTOR) { |
1534 | 0 | file.fd = -1; |
1535 | 0 | r = dev->vhost_ops->vhost_set_vring_call(dev, &file); |
1536 | 0 | if (r) { |
1537 | 0 | goto fail; |
1538 | 0 | } |
1539 | 0 | } |
1540 | | |
1541 | 0 | trace_vhost_virtqueue_start_out(dev, vdev->name, idx); |
1542 | |
|
1543 | 0 | return 0; |
1544 | | |
1545 | 0 | fail: |
1546 | 0 | vhost_vrings_unmap(dev, vq, false); |
1547 | 0 | return r; |
1548 | 0 | } |
1549 | | |
1550 | | static int do_vhost_virtqueue_stop(struct vhost_dev *dev, |
1551 | | struct VirtIODevice *vdev, |
1552 | | struct vhost_virtqueue *vq, |
1553 | | unsigned idx, bool force, |
1554 | | bool skip_drain) |
1555 | 0 | { |
1556 | 0 | int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, idx); |
1557 | 0 | struct vhost_vring_state state = { |
1558 | 0 | .index = vhost_vq_index, |
1559 | 0 | }; |
1560 | 0 | int r = 0; |
1561 | |
|
1562 | 0 | trace_vhost_virtqueue_stop_in(dev, vdev->name, idx); |
1563 | |
|
1564 | 0 | if (virtio_queue_get_desc_addr(vdev, idx) == 0) { |
1565 | | /* Don't stop the virtqueue which might have not been started */ |
1566 | 0 | return 0; |
1567 | 0 | } |
1568 | | |
1569 | 0 | if (!force) { |
1570 | 0 | if (!skip_drain) { |
1571 | 0 | r = dev->vhost_ops->vhost_get_vring_base(dev, &state); |
1572 | 0 | } else { |
1573 | 0 | r = dev->vhost_ops->vhost_get_vring_base_skip_drain(dev, &state); |
1574 | 0 | } |
1575 | 0 | if (r < 0) { |
1576 | 0 | VHOST_OPS_DEBUG(r, "vhost VQ %u ring restore failed: %d", idx, r); |
1577 | 0 | } |
1578 | 0 | } |
1579 | |
|
1580 | 0 | if (r < 0 || force) { |
1581 | | /* Connection to the backend is broken, so let's sync internal |
1582 | | * last avail idx to the device used idx. |
1583 | | */ |
1584 | 0 | virtio_queue_restore_last_avail_idx(vdev, idx); |
1585 | 0 | } else { |
1586 | 0 | virtio_queue_set_last_avail_idx(vdev, idx, state.num); |
1587 | 0 | } |
1588 | 0 | virtio_queue_invalidate_signalled_used(vdev, idx); |
1589 | 0 | virtio_queue_update_used_idx(vdev, idx); |
1590 | | |
1591 | | /* In the cross-endian case, we need to reset the vring endianness to |
1592 | | * native as legacy devices expect so by default. |
1593 | | */ |
1594 | 0 | if (vhost_needs_vring_endian(vdev)) { |
1595 | 0 | vhost_virtqueue_set_vring_endian_legacy(dev, |
1596 | 0 | !virtio_vdev_is_big_endian(vdev), |
1597 | 0 | vhost_vq_index); |
1598 | 0 | } |
1599 | |
|
1600 | 0 | vhost_vrings_unmap(dev, vq, true); |
1601 | |
|
1602 | 0 | trace_vhost_virtqueue_stop_out(dev, vdev->name, idx); |
1603 | 0 | return r; |
1604 | 0 | } |
1605 | | |
1606 | | int vhost_virtqueue_stop(struct vhost_dev *dev, |
1607 | | struct VirtIODevice *vdev, |
1608 | | struct vhost_virtqueue *vq, |
1609 | | unsigned idx, |
1610 | | bool skip_drain) |
1611 | 0 | { |
1612 | 0 | return do_vhost_virtqueue_stop(dev, vdev, vq, idx, false, skip_drain); |
1613 | 0 | } |
1614 | | |
1615 | | static int vhost_virtqueue_set_busyloop_timeout(struct vhost_dev *dev, |
1616 | | int n, uint32_t timeout) |
1617 | 0 | { |
1618 | 0 | int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, n); |
1619 | 0 | struct vhost_vring_state state = { |
1620 | 0 | .index = vhost_vq_index, |
1621 | 0 | .num = timeout, |
1622 | 0 | }; |
1623 | 0 | int r; |
1624 | |
|
1625 | 0 | if (!dev->vhost_ops->vhost_set_vring_busyloop_timeout) { |
1626 | 0 | return -EINVAL; |
1627 | 0 | } |
1628 | | |
1629 | 0 | r = dev->vhost_ops->vhost_set_vring_busyloop_timeout(dev, &state); |
1630 | 0 | if (r) { |
1631 | 0 | VHOST_OPS_DEBUG(r, "vhost_set_vring_busyloop_timeout failed"); |
1632 | 0 | return r; |
1633 | 0 | } |
1634 | | |
1635 | 0 | return 0; |
1636 | 0 | } |
1637 | | |
1638 | | static void vhost_virtqueue_error_notifier(EventNotifier *n) |
1639 | 0 | { |
1640 | 0 | struct vhost_virtqueue *vq = container_of(n, struct vhost_virtqueue, |
1641 | 0 | error_notifier); |
1642 | 0 | struct vhost_dev *dev = vq->dev; |
1643 | 0 | int index = vq - dev->vqs; |
1644 | |
|
1645 | 0 | if (event_notifier_test_and_clear(n) && dev->vdev) { |
1646 | 0 | VHOST_OPS_DEBUG(-EINVAL, "vhost vring error in virtqueue %d", |
1647 | 0 | dev->vq_index + index); |
1648 | 0 | } |
1649 | 0 | } |
1650 | | |
1651 | | static int vhost_virtqueue_init(struct vhost_dev *dev, |
1652 | | struct vhost_virtqueue *vq, int n, |
1653 | | bool busyloop_timeout) |
1654 | 0 | { |
1655 | 0 | int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, n); |
1656 | 0 | struct vhost_vring_file file = { |
1657 | 0 | .index = vhost_vq_index, |
1658 | 0 | }; |
1659 | 0 | int r = event_notifier_init(&vq->masked_notifier, 0); |
1660 | 0 | if (r < 0) { |
1661 | 0 | return r; |
1662 | 0 | } |
1663 | | |
1664 | 0 | file.fd = event_notifier_get_wfd(&vq->masked_notifier); |
1665 | 0 | r = dev->vhost_ops->vhost_set_vring_call(dev, &file); |
1666 | 0 | if (r) { |
1667 | 0 | VHOST_OPS_DEBUG(r, "vhost_set_vring_call failed"); |
1668 | 0 | goto fail_call; |
1669 | 0 | } |
1670 | | |
1671 | 0 | vq->dev = dev; |
1672 | |
|
1673 | 0 | if (dev->vhost_ops->vhost_set_vring_err) { |
1674 | 0 | r = event_notifier_init(&vq->error_notifier, 0); |
1675 | 0 | if (r < 0) { |
1676 | 0 | goto fail_call; |
1677 | 0 | } |
1678 | | |
1679 | 0 | file.fd = event_notifier_get_fd(&vq->error_notifier); |
1680 | 0 | r = dev->vhost_ops->vhost_set_vring_err(dev, &file); |
1681 | 0 | if (r) { |
1682 | 0 | VHOST_OPS_DEBUG(r, "vhost_set_vring_err failed"); |
1683 | 0 | goto fail_err; |
1684 | 0 | } |
1685 | | |
1686 | 0 | event_notifier_set_handler(&vq->error_notifier, |
1687 | 0 | vhost_virtqueue_error_notifier); |
1688 | 0 | } |
1689 | | |
1690 | 0 | if (busyloop_timeout) { |
1691 | 0 | r = vhost_virtqueue_set_busyloop_timeout(dev, n, busyloop_timeout); |
1692 | 0 | if (r < 0) { |
1693 | 0 | VHOST_OPS_DEBUG(r, "Failed to set busyloop timeout"); |
1694 | 0 | goto fail_err; |
1695 | 0 | } |
1696 | 0 | } |
1697 | | |
1698 | 0 | return 0; |
1699 | | |
1700 | 0 | fail_err: |
1701 | 0 | event_notifier_cleanup(&vq->error_notifier); |
1702 | 0 | fail_call: |
1703 | 0 | event_notifier_cleanup(&vq->masked_notifier); |
1704 | 0 | return r; |
1705 | 0 | } |
1706 | | |
1707 | | static void vhost_virtqueue_cleanup(struct vhost_virtqueue *vq) |
1708 | 0 | { |
1709 | 0 | event_notifier_cleanup(&vq->masked_notifier); |
1710 | 0 | if (vq->dev && vq->dev->vhost_ops->vhost_set_vring_err) { |
1711 | 0 | event_notifier_set_handler(&vq->error_notifier, NULL); |
1712 | 0 | event_notifier_cleanup(&vq->error_notifier); |
1713 | 0 | } |
1714 | 0 | } |
1715 | | |
1716 | | static int vhost_dev_init_features(struct vhost_dev *hdev) |
1717 | 0 | { |
1718 | 0 | uint64_t features64; |
1719 | 0 | int r; |
1720 | |
|
1721 | 0 | if (hdev->vhost_ops->vhost_get_features_ex) { |
1722 | 0 | return hdev->vhost_ops->vhost_get_features_ex(hdev, hdev->_features_ex); |
1723 | 0 | } |
1724 | | |
1725 | 0 | r = hdev->vhost_ops->vhost_get_features(hdev, &features64); |
1726 | 0 | virtio_features_from_u64(hdev->_features_ex, features64); |
1727 | 0 | return r; |
1728 | 0 | } |
1729 | | |
1730 | | int vhost_dev_init_backend(struct vhost_dev *hdev, void *opaque, |
1731 | | VhostBackendType backend_type, Error **errp) |
1732 | 0 | { |
1733 | 0 | int r; |
1734 | |
|
1735 | 0 | r = vhost_set_backend_type(hdev, backend_type); |
1736 | 0 | assert(r >= 0); |
1737 | |
|
1738 | 0 | r = hdev->vhost_ops->vhost_init(hdev, opaque, errp); |
1739 | 0 | if (r < 0) { |
1740 | 0 | return r; |
1741 | 0 | } |
1742 | | |
1743 | 0 | r = vhost_dev_init_features(hdev); |
1744 | 0 | if (r < 0) { |
1745 | 0 | error_setg_errno(errp, -r, "vhost_init_features failed"); |
1746 | 0 | return r; |
1747 | 0 | } |
1748 | | |
1749 | 0 | return 0; |
1750 | 0 | } |
1751 | | |
1752 | | int vhost_dev_set_owner(struct vhost_dev *hdev) |
1753 | 0 | { |
1754 | 0 | int r; |
1755 | |
|
1756 | 0 | assert(hdev->vhost_ops); |
1757 | 0 | if (!hdev->vhost_ops->vhost_set_owner) { |
1758 | 0 | return -ENOSYS; |
1759 | 0 | } |
1760 | 0 | r = hdev->vhost_ops->vhost_set_owner(hdev); |
1761 | 0 | if (!r) { |
1762 | 0 | hdev->owner = true; |
1763 | 0 | } |
1764 | 0 | return r; |
1765 | 0 | } |
1766 | | |
1767 | | int vhost_dev_reset_owner(struct vhost_dev *hdev) |
1768 | 0 | { |
1769 | 0 | int r; |
1770 | |
|
1771 | 0 | assert(hdev->vhost_ops); |
1772 | 0 | if (!hdev->vhost_ops->vhost_reset_owner) { |
1773 | 0 | return -ENOSYS; |
1774 | 0 | } |
1775 | 0 | r = hdev->vhost_ops->vhost_reset_owner(hdev); |
1776 | 0 | if (!r) { |
1777 | 0 | hdev->owner = false; |
1778 | 0 | } |
1779 | 0 | return r; |
1780 | 0 | } |
1781 | | |
1782 | | int vhost_dev_init(struct vhost_dev *hdev, void *opaque, |
1783 | | VhostBackendType backend_type, uint32_t busyloop_timeout, |
1784 | | Error **errp) |
1785 | 0 | { |
1786 | 0 | unsigned int used, reserved, limit; |
1787 | 0 | int i, r, n_initialized_vqs = 0; |
1788 | |
|
1789 | 0 | trace_vhost_dev_init_in(hdev); |
1790 | |
|
1791 | 0 | hdev->vdev = NULL; |
1792 | 0 | hdev->migration_blocker = NULL; |
1793 | |
|
1794 | 0 | r = vhost_dev_init_backend(hdev, opaque, backend_type, errp); |
1795 | 0 | if (r < 0) { |
1796 | 0 | goto fail; |
1797 | 0 | } |
1798 | | |
1799 | 0 | r = vhost_dev_set_owner(hdev); |
1800 | 0 | if (r < 0) { |
1801 | 0 | error_setg_errno(errp, -r, "vhost_set_owner failed"); |
1802 | 0 | goto fail; |
1803 | 0 | } |
1804 | | |
1805 | 0 | limit = hdev->vhost_ops->vhost_memslots_limit(hdev); |
1806 | 0 | if (limit < MEMORY_DEVICES_SAFE_MAX_MEMSLOTS && |
1807 | 0 | memory_devices_memslot_auto_decision_active()) { |
1808 | 0 | error_setg(errp, "some memory device (like virtio-mem)" |
1809 | 0 | " decided how many memory slots to use based on the overall" |
1810 | 0 | " number of memory slots; this vhost backend would further" |
1811 | 0 | " restricts the overall number of memory slots"); |
1812 | 0 | error_append_hint(errp, "Try plugging this vhost backend before" |
1813 | 0 | " plugging such memory devices.\n"); |
1814 | 0 | r = -EINVAL; |
1815 | 0 | goto fail; |
1816 | 0 | } |
1817 | | |
1818 | 0 | for (i = 0; i < hdev->nvqs; ++i, ++n_initialized_vqs) { |
1819 | 0 | r = vhost_virtqueue_init(hdev, hdev->vqs + i, hdev->vq_index + i, |
1820 | 0 | busyloop_timeout); |
1821 | 0 | if (r < 0) { |
1822 | 0 | error_setg_errno(errp, -r, "Failed to initialize virtqueue %d", i); |
1823 | 0 | goto fail; |
1824 | 0 | } |
1825 | 0 | } |
1826 | | |
1827 | 0 | hdev->memory_listener = (MemoryListener) { |
1828 | 0 | .name = "vhost", |
1829 | 0 | .begin = vhost_begin, |
1830 | 0 | .commit = vhost_commit, |
1831 | 0 | .region_add = vhost_region_addnop, |
1832 | 0 | .region_nop = vhost_region_addnop, |
1833 | 0 | .log_start = vhost_log_start, |
1834 | 0 | .log_stop = vhost_log_stop, |
1835 | 0 | .log_sync = vhost_log_sync, |
1836 | 0 | .log_global_start = vhost_log_global_start, |
1837 | 0 | .log_global_stop = vhost_log_global_stop, |
1838 | 0 | .priority = MEMORY_LISTENER_PRIORITY_DEV_BACKEND |
1839 | 0 | }; |
1840 | |
|
1841 | 0 | hdev->iommu_listener = (MemoryListener) { |
1842 | 0 | .name = "vhost-iommu", |
1843 | 0 | .region_add = vhost_iommu_region_add, |
1844 | 0 | .region_del = vhost_iommu_region_del, |
1845 | 0 | }; |
1846 | |
|
1847 | 0 | if (hdev->migration_blocker == NULL) { |
1848 | 0 | if (!vhost_dev_has_feature_ex(hdev, VHOST_F_LOG_ALL)) { |
1849 | 0 | error_setg(&hdev->migration_blocker, |
1850 | 0 | "Migration disabled: vhost lacks VHOST_F_LOG_ALL feature."); |
1851 | 0 | } else if (vhost_dev_log_is_shared(hdev) && !qemu_memfd_alloc_check()) { |
1852 | 0 | error_setg(&hdev->migration_blocker, |
1853 | 0 | "Migration disabled: failed to allocate shared memory"); |
1854 | 0 | } |
1855 | 0 | } |
1856 | |
|
1857 | 0 | if (hdev->migration_blocker != NULL) { |
1858 | 0 | r = migrate_add_blocker_normal(&hdev->migration_blocker, errp); |
1859 | 0 | if (r < 0) { |
1860 | 0 | goto fail; |
1861 | 0 | } |
1862 | 0 | } |
1863 | | |
1864 | 0 | hdev->mem = g_malloc0(offsetof(struct vhost_memory, regions)); |
1865 | 0 | hdev->n_mem_sections = 0; |
1866 | 0 | hdev->mem_sections = NULL; |
1867 | 0 | hdev->log = NULL; |
1868 | 0 | hdev->log_size = 0; |
1869 | 0 | hdev->log_enabled = false; |
1870 | 0 | hdev->started = false; |
1871 | 0 | memory_listener_register(&hdev->memory_listener, &address_space_memory); |
1872 | 0 | hdev->exit_notifier.notify = vhost_exit_notify; |
1873 | 0 | qemu_add_exit_notifier(&hdev->exit_notifier); |
1874 | 0 | migration_add_notifier_modes(&hdev->migration_notifier, |
1875 | 0 | vhost_migration_notify, |
1876 | 0 | BIT(MIG_MODE_CPR_TRANSFER) | |
1877 | 0 | BIT(MIG_MODE_CPR_EXEC)); |
1878 | 0 | QLIST_INSERT_HEAD(&vhost_devices, hdev, entry); |
1879 | | |
1880 | | /* |
1881 | | * The listener we registered properly setup the number of required |
1882 | | * memslots in vhost_commit(). |
1883 | | */ |
1884 | 0 | used = hdev->mem->nregions; |
1885 | | |
1886 | | /* |
1887 | | * We assume that all reserved memslots actually require a real memslot |
1888 | | * in our vhost backend. This might not be true, for example, if the |
1889 | | * memslot would be ROM. If ever relevant, we can optimize for that -- |
1890 | | * but we'll need additional information about the reservations. |
1891 | | */ |
1892 | 0 | reserved = memory_devices_get_reserved_memslots(); |
1893 | 0 | if (used + reserved > limit) { |
1894 | 0 | error_setg(errp, "vhost backend memory slots limit (%d) is less" |
1895 | 0 | " than current number of used (%d) and reserved (%d)" |
1896 | 0 | " memory slots for memory devices.", limit, used, reserved); |
1897 | 0 | r = -EINVAL; |
1898 | 0 | goto fail; |
1899 | 0 | } |
1900 | | |
1901 | 0 | hdev->initialized = true; |
1902 | |
|
1903 | 0 | trace_vhost_dev_init_out(hdev); |
1904 | |
|
1905 | 0 | return 0; |
1906 | | |
1907 | 0 | fail: |
1908 | 0 | hdev->nvqs = n_initialized_vqs; |
1909 | 0 | vhost_dev_cleanup(hdev); |
1910 | 0 | return r; |
1911 | 0 | } |
1912 | | |
1913 | | void vhost_dev_cleanup(struct vhost_dev *hdev) |
1914 | 0 | { |
1915 | 0 | int i; |
1916 | |
|
1917 | 0 | trace_vhost_dev_cleanup(hdev); |
1918 | |
|
1919 | 0 | for (i = 0; i < hdev->nvqs; ++i) { |
1920 | 0 | vhost_virtqueue_cleanup(hdev->vqs + i); |
1921 | 0 | } |
1922 | 0 | if (hdev->mem) { |
1923 | | /* those are only safe after successful init */ |
1924 | 0 | memory_listener_unregister(&hdev->memory_listener); |
1925 | 0 | QLIST_REMOVE(hdev, entry); |
1926 | 0 | } |
1927 | 0 | if (hdev->exit_notifier.notify) { |
1928 | 0 | qemu_remove_exit_notifier(&hdev->exit_notifier); |
1929 | 0 | hdev->exit_notifier.notify = NULL; |
1930 | 0 | } |
1931 | 0 | migration_remove_notifier(&hdev->migration_notifier); |
1932 | 0 | migrate_del_blocker(&hdev->migration_blocker); |
1933 | 0 | g_free(hdev->mem); |
1934 | 0 | g_free(hdev->mem_sections); |
1935 | 0 | if (hdev->vhost_ops) { |
1936 | | /* |
1937 | | * Release the device ownership if we hold it. Needed e.g. for |
1938 | | * failing CPR target, so the source can re-acquire it. |
1939 | | */ |
1940 | 0 | vhost_dev_try_release_owner(hdev, "cleanup"); |
1941 | 0 | hdev->vhost_ops->vhost_cleanup(hdev); |
1942 | 0 | } |
1943 | 0 | assert(!hdev->log); |
1944 | |
|
1945 | 0 | memset(hdev, 0, sizeof(struct vhost_dev)); |
1946 | 0 | } |
1947 | | |
1948 | | void vhost_dev_disable_notifiers_nvqs(struct vhost_dev *hdev, |
1949 | | VirtIODevice *vdev, |
1950 | | unsigned int nvqs) |
1951 | 0 | { |
1952 | 0 | BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev))); |
1953 | 0 | int i, r; |
1954 | | |
1955 | | /* |
1956 | | * Batch all the host notifiers in a single transaction to avoid |
1957 | | * quadratic time complexity in address_space_update_ioeventfds(). |
1958 | | */ |
1959 | 0 | memory_region_transaction_begin(); |
1960 | |
|
1961 | 0 | for (i = 0; i < nvqs; ++i) { |
1962 | 0 | r = virtio_bus_set_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i, |
1963 | 0 | false); |
1964 | 0 | if (r < 0) { |
1965 | 0 | error_report("vhost VQ %d notifier cleanup failed: %d", i, -r); |
1966 | 0 | } |
1967 | 0 | assert(r >= 0); |
1968 | 0 | } |
1969 | | |
1970 | | /* |
1971 | | * The transaction expects the ioeventfds to be open when it |
1972 | | * commits. Do it now, before the cleanup loop. |
1973 | | */ |
1974 | 0 | memory_region_transaction_commit(); |
1975 | |
|
1976 | 0 | for (i = 0; i < nvqs; ++i) { |
1977 | 0 | virtio_bus_cleanup_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i); |
1978 | 0 | } |
1979 | 0 | virtio_device_release_ioeventfd(vdev); |
1980 | 0 | } |
1981 | | |
1982 | | /* Stop processing guest IO notifications in qemu. |
1983 | | * Start processing them in vhost in kernel. |
1984 | | */ |
1985 | | int vhost_dev_enable_notifiers(struct vhost_dev *hdev, VirtIODevice *vdev) |
1986 | 0 | { |
1987 | 0 | BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev))); |
1988 | 0 | int i, r; |
1989 | | |
1990 | | /* We will pass the notifiers to the kernel, make sure that QEMU |
1991 | | * doesn't interfere. |
1992 | | */ |
1993 | 0 | r = virtio_device_grab_ioeventfd(vdev); |
1994 | 0 | if (r < 0) { |
1995 | 0 | error_report("binding does not support host notifiers"); |
1996 | 0 | return r; |
1997 | 0 | } |
1998 | | |
1999 | | /* |
2000 | | * Batch all the host notifiers in a single transaction to avoid |
2001 | | * quadratic time complexity in address_space_update_ioeventfds(). |
2002 | | */ |
2003 | 0 | memory_region_transaction_begin(); |
2004 | |
|
2005 | 0 | for (i = 0; i < hdev->nvqs; ++i) { |
2006 | 0 | r = virtio_bus_set_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i, |
2007 | 0 | true); |
2008 | 0 | if (r < 0) { |
2009 | 0 | error_report("vhost VQ %d notifier binding failed: %d", i, -r); |
2010 | 0 | memory_region_transaction_commit(); |
2011 | 0 | vhost_dev_disable_notifiers_nvqs(hdev, vdev, i); |
2012 | 0 | return r; |
2013 | 0 | } |
2014 | 0 | } |
2015 | | |
2016 | 0 | memory_region_transaction_commit(); |
2017 | |
|
2018 | 0 | return 0; |
2019 | 0 | } |
2020 | | |
2021 | | /* Stop processing guest IO notifications in vhost. |
2022 | | * Start processing them in qemu. |
2023 | | * This might actually run the qemu handlers right away, |
2024 | | * so virtio in qemu must be completely setup when this is called. |
2025 | | */ |
2026 | | void vhost_dev_disable_notifiers(struct vhost_dev *hdev, VirtIODevice *vdev) |
2027 | 0 | { |
2028 | 0 | vhost_dev_disable_notifiers_nvqs(hdev, vdev, hdev->nvqs); |
2029 | 0 | } |
2030 | | |
2031 | | /* Test and clear event pending status. |
2032 | | * Should be called after unmask to avoid losing events. |
2033 | | */ |
2034 | | bool vhost_virtqueue_pending(struct vhost_dev *hdev, int n) |
2035 | 0 | { |
2036 | 0 | struct vhost_virtqueue *vq = hdev->vqs + n - hdev->vq_index; |
2037 | 0 | assert(n >= hdev->vq_index && n < hdev->vq_index + hdev->nvqs); |
2038 | 0 | return event_notifier_test_and_clear(&vq->masked_notifier); |
2039 | 0 | } |
2040 | | |
2041 | | /* Mask/unmask events from this vq. */ |
2042 | | int vhost_virtqueue_mask(struct vhost_dev *hdev, VirtIODevice *vdev, int n, |
2043 | | bool mask) |
2044 | 0 | { |
2045 | 0 | struct VirtQueue *vvq = virtio_get_queue(vdev, n); |
2046 | 0 | int r, index = n - hdev->vq_index; |
2047 | 0 | struct vhost_vring_file file; |
2048 | | |
2049 | | /* should only be called after backend is connected */ |
2050 | 0 | assert(hdev->vhost_ops); |
2051 | |
|
2052 | 0 | if (mask) { |
2053 | 0 | assert(vdev->use_guest_notifier_mask); |
2054 | 0 | file.fd = event_notifier_get_wfd(&hdev->vqs[index].masked_notifier); |
2055 | 0 | } else { |
2056 | 0 | file.fd = event_notifier_get_wfd(virtio_queue_get_guest_notifier(vvq)); |
2057 | 0 | } |
2058 | |
|
2059 | 0 | file.index = hdev->vhost_ops->vhost_get_vq_index(hdev, n); |
2060 | 0 | r = hdev->vhost_ops->vhost_set_vring_call(hdev, &file); |
2061 | 0 | if (r < 0) { |
2062 | 0 | error_report("vhost_set_vring_call failed %d", -r); |
2063 | 0 | return r; |
2064 | 0 | } |
2065 | 0 | return 0; |
2066 | 0 | } |
2067 | | |
2068 | | bool vhost_config_pending(struct vhost_dev *hdev) |
2069 | 0 | { |
2070 | 0 | assert(hdev->vhost_ops); |
2071 | 0 | if ((hdev->started == false) || |
2072 | 0 | (hdev->vhost_ops->vhost_set_config_call == NULL)) { |
2073 | 0 | return false; |
2074 | 0 | } |
2075 | | |
2076 | 0 | EventNotifier *notifier = |
2077 | 0 | &hdev->vqs[VHOST_QUEUE_NUM_CONFIG_INR].masked_config_notifier; |
2078 | 0 | return event_notifier_test_and_clear(notifier); |
2079 | 0 | } |
2080 | | |
2081 | | void vhost_config_mask(struct vhost_dev *hdev, VirtIODevice *vdev, bool mask) |
2082 | 0 | { |
2083 | 0 | int fd; |
2084 | 0 | int r; |
2085 | 0 | EventNotifier *notifier = |
2086 | 0 | &hdev->vqs[VHOST_QUEUE_NUM_CONFIG_INR].masked_config_notifier; |
2087 | 0 | EventNotifier *config_notifier = virtio_config_get_guest_notifier(vdev); |
2088 | 0 | assert(hdev->vhost_ops); |
2089 | |
|
2090 | 0 | if ((hdev->started == false) || |
2091 | 0 | (hdev->vhost_ops->vhost_set_config_call == NULL)) { |
2092 | 0 | return; |
2093 | 0 | } |
2094 | 0 | if (mask) { |
2095 | 0 | assert(vdev->use_guest_notifier_mask); |
2096 | 0 | fd = event_notifier_get_fd(notifier); |
2097 | 0 | } else { |
2098 | 0 | fd = event_notifier_get_fd(config_notifier); |
2099 | 0 | } |
2100 | 0 | r = hdev->vhost_ops->vhost_set_config_call(hdev, fd); |
2101 | 0 | if (r < 0) { |
2102 | 0 | error_report("vhost_set_config_call failed %d", -r); |
2103 | 0 | } |
2104 | 0 | } |
2105 | | |
2106 | | static void vhost_stop_config_intr(struct vhost_dev *dev) |
2107 | 0 | { |
2108 | 0 | int fd = -1; |
2109 | 0 | assert(dev->vhost_ops); |
2110 | 0 | if (dev->vhost_ops->vhost_set_config_call) { |
2111 | 0 | dev->vhost_ops->vhost_set_config_call(dev, fd); |
2112 | 0 | } |
2113 | 0 | } |
2114 | | |
2115 | | static void vhost_start_config_intr(struct vhost_dev *dev) |
2116 | 0 | { |
2117 | 0 | int r; |
2118 | 0 | EventNotifier *config_notifier = |
2119 | 0 | virtio_config_get_guest_notifier(dev->vdev); |
2120 | |
|
2121 | 0 | assert(dev->vhost_ops); |
2122 | 0 | int fd = event_notifier_get_fd(config_notifier); |
2123 | 0 | if (dev->vhost_ops->vhost_set_config_call) { |
2124 | 0 | r = dev->vhost_ops->vhost_set_config_call(dev, fd); |
2125 | 0 | if (!r) { |
2126 | 0 | event_notifier_set(config_notifier); |
2127 | 0 | } |
2128 | 0 | } |
2129 | 0 | } |
2130 | | |
2131 | | void vhost_get_features_ex(struct vhost_dev *hdev, |
2132 | | const int *feature_bits, |
2133 | | uint64_t *features) |
2134 | 0 | { |
2135 | 0 | const int *bit = feature_bits; |
2136 | |
|
2137 | 0 | while (*bit != VHOST_INVALID_FEATURE_BIT) { |
2138 | 0 | if (!vhost_dev_has_feature_ex(hdev, *bit)) { |
2139 | 0 | virtio_clear_feature_ex(features, *bit); |
2140 | 0 | } |
2141 | 0 | bit++; |
2142 | 0 | } |
2143 | 0 | } |
2144 | | |
2145 | | static bool vhost_inflight_buffer_pre_load(void *opaque, Error **errp) |
2146 | 0 | { |
2147 | 0 | struct vhost_inflight *inflight = opaque; |
2148 | |
|
2149 | 0 | int fd = -1; |
2150 | 0 | void *addr = qemu_memfd_alloc("vhost-inflight", inflight->size, |
2151 | 0 | F_SEAL_GROW | F_SEAL_SHRINK | F_SEAL_SEAL, |
2152 | 0 | &fd, errp); |
2153 | 0 | if (!addr) { |
2154 | 0 | return false; |
2155 | 0 | } |
2156 | | |
2157 | 0 | inflight->offset = 0; |
2158 | 0 | inflight->addr = addr; |
2159 | 0 | inflight->fd = fd; |
2160 | |
|
2161 | 0 | return true; |
2162 | 0 | } |
2163 | | |
2164 | | const VMStateDescription vmstate_vhost_inflight_region_buffer = { |
2165 | | .name = "vhost-inflight-region/buffer", |
2166 | | .pre_load_errp = vhost_inflight_buffer_pre_load, |
2167 | | .fields = (const VMStateField[]) { |
2168 | | VMSTATE_VBUFFER_UINT64(addr, struct vhost_inflight, 0, NULL, size), |
2169 | | VMSTATE_END_OF_LIST() |
2170 | | } |
2171 | | }; |
2172 | | |
2173 | | static bool vhost_inflight_region_post_load(void *opaque, |
2174 | | int version_id, |
2175 | | Error **errp) |
2176 | 0 | { |
2177 | 0 | struct vhost_inflight *inflight = opaque; |
2178 | |
|
2179 | 0 | if (inflight->addr == NULL) { |
2180 | 0 | error_setg(errp, "inflight buffer subsection has not been loaded"); |
2181 | 0 | return false; |
2182 | 0 | } |
2183 | | |
2184 | 0 | return true; |
2185 | 0 | } |
2186 | | |
2187 | | const VMStateDescription vmstate_vhost_inflight_region = { |
2188 | | .name = "vhost-inflight-region", |
2189 | | .post_load_errp = vhost_inflight_region_post_load, |
2190 | | .fields = (const VMStateField[]) { |
2191 | | VMSTATE_UINT64(size, struct vhost_inflight), |
2192 | | VMSTATE_UINT16(queue_size, struct vhost_inflight), |
2193 | | VMSTATE_END_OF_LIST() |
2194 | | }, |
2195 | | .subsections = (const VMStateDescription * const []) { |
2196 | | &vmstate_vhost_inflight_region_buffer, |
2197 | | NULL |
2198 | | } |
2199 | | }; |
2200 | | |
2201 | | void vhost_ack_features_ex(struct vhost_dev *hdev, const int *feature_bits, |
2202 | | const uint64_t *features) |
2203 | 0 | { |
2204 | 0 | const int *bit = feature_bits; |
2205 | 0 | while (*bit != VHOST_INVALID_FEATURE_BIT) { |
2206 | 0 | if (virtio_has_feature_ex(features, *bit)) { |
2207 | 0 | virtio_add_feature_ex(hdev->acked_features_ex, *bit); |
2208 | 0 | } |
2209 | 0 | bit++; |
2210 | 0 | } |
2211 | 0 | } |
2212 | | |
2213 | | int vhost_dev_get_config(struct vhost_dev *hdev, uint8_t *config, |
2214 | | uint32_t config_len, Error **errp) |
2215 | 0 | { |
2216 | 0 | assert(hdev->vhost_ops); |
2217 | |
|
2218 | 0 | if (hdev->vhost_ops->vhost_get_config) { |
2219 | 0 | return hdev->vhost_ops->vhost_get_config(hdev, config, config_len, |
2220 | 0 | errp); |
2221 | 0 | } |
2222 | | |
2223 | 0 | error_setg(errp, "vhost_get_config not implemented"); |
2224 | 0 | return -ENOSYS; |
2225 | 0 | } |
2226 | | |
2227 | | int vhost_dev_set_config(struct vhost_dev *hdev, const uint8_t *data, |
2228 | | uint32_t offset, uint32_t size, uint32_t flags) |
2229 | 0 | { |
2230 | 0 | assert(hdev->vhost_ops); |
2231 | |
|
2232 | 0 | if (hdev->vhost_ops->vhost_set_config) { |
2233 | 0 | return hdev->vhost_ops->vhost_set_config(hdev, data, offset, |
2234 | 0 | size, flags); |
2235 | 0 | } |
2236 | | |
2237 | 0 | return -ENOSYS; |
2238 | 0 | } |
2239 | | |
2240 | | void vhost_dev_set_config_notifier(struct vhost_dev *hdev, |
2241 | | const VhostDevConfigOps *ops) |
2242 | 0 | { |
2243 | 0 | hdev->config_ops = ops; |
2244 | 0 | } |
2245 | | |
2246 | | void vhost_dev_free_inflight(struct vhost_inflight *inflight) |
2247 | 0 | { |
2248 | 0 | if (inflight && inflight->addr) { |
2249 | 0 | qemu_memfd_free(inflight->addr, inflight->size, inflight->fd); |
2250 | 0 | inflight->addr = NULL; |
2251 | 0 | inflight->fd = -1; |
2252 | 0 | } |
2253 | 0 | } |
2254 | | |
2255 | | int vhost_dev_prepare_inflight(struct vhost_dev *hdev, VirtIODevice *vdev) |
2256 | 0 | { |
2257 | 0 | int r; |
2258 | |
|
2259 | 0 | if (hdev->vhost_ops->vhost_get_inflight_fd == NULL || |
2260 | 0 | hdev->vhost_ops->vhost_set_inflight_fd == NULL) { |
2261 | 0 | return 0; |
2262 | 0 | } |
2263 | | |
2264 | 0 | hdev->vdev = vdev; |
2265 | |
|
2266 | 0 | r = vhost_dev_set_features(hdev, hdev->log_enabled); |
2267 | 0 | if (r < 0) { |
2268 | 0 | VHOST_OPS_DEBUG(r, "vhost_dev_prepare_inflight failed"); |
2269 | 0 | return r; |
2270 | 0 | } |
2271 | | |
2272 | 0 | return 0; |
2273 | 0 | } |
2274 | | |
2275 | | int vhost_dev_set_inflight(struct vhost_dev *dev, |
2276 | | struct vhost_inflight *inflight) |
2277 | 0 | { |
2278 | 0 | int r; |
2279 | |
|
2280 | 0 | if (dev->vhost_ops->vhost_set_inflight_fd && inflight->addr) { |
2281 | 0 | r = dev->vhost_ops->vhost_set_inflight_fd(dev, inflight); |
2282 | 0 | if (r) { |
2283 | 0 | VHOST_OPS_DEBUG(r, "vhost_set_inflight_fd failed"); |
2284 | 0 | return r; |
2285 | 0 | } |
2286 | 0 | } |
2287 | | |
2288 | 0 | return 0; |
2289 | 0 | } |
2290 | | |
2291 | | int vhost_dev_get_inflight(struct vhost_dev *dev, uint16_t queue_size, |
2292 | | struct vhost_inflight *inflight) |
2293 | 0 | { |
2294 | 0 | int r; |
2295 | |
|
2296 | 0 | if (dev->vhost_ops->vhost_get_inflight_fd) { |
2297 | 0 | r = dev->vhost_ops->vhost_get_inflight_fd(dev, queue_size, inflight); |
2298 | 0 | if (r) { |
2299 | 0 | VHOST_OPS_DEBUG(r, "vhost_get_inflight_fd failed"); |
2300 | 0 | return r; |
2301 | 0 | } |
2302 | 0 | } |
2303 | | |
2304 | 0 | return 0; |
2305 | 0 | } |
2306 | | |
2307 | | /* |
2308 | | * Host notifiers must be enabled at this point. |
2309 | | * |
2310 | | * If @vrings is true, this function will enable all vrings before starting the |
2311 | | * device. If it is false, the vring initialization is left to be done by the |
2312 | | * caller. |
2313 | | */ |
2314 | | int vhost_dev_start(struct vhost_dev *hdev, VirtIODevice *vdev, bool vrings) |
2315 | 0 | { |
2316 | 0 | int i, r; |
2317 | | |
2318 | | /* should only be called after backend is connected */ |
2319 | 0 | assert(hdev->vhost_ops); |
2320 | |
|
2321 | 0 | trace_vhost_dev_start_in(hdev, vdev->name, vrings); |
2322 | |
|
2323 | 0 | vdev->vhost_started = true; |
2324 | 0 | hdev->started = true; |
2325 | 0 | hdev->vdev = vdev; |
2326 | |
|
2327 | 0 | r = vhost_dev_set_features(hdev, hdev->log_enabled); |
2328 | 0 | if (r < 0) { |
2329 | 0 | goto fail_features; |
2330 | 0 | } |
2331 | | |
2332 | 0 | if (vhost_dev_has_iommu(hdev)) { |
2333 | 0 | memory_listener_register(&hdev->iommu_listener, vdev->dma_as); |
2334 | 0 | } |
2335 | |
|
2336 | 0 | r = hdev->vhost_ops->vhost_set_mem_table(hdev, hdev->mem); |
2337 | 0 | if (r < 0) { |
2338 | 0 | VHOST_OPS_DEBUG(r, "vhost_set_mem_table failed"); |
2339 | 0 | goto fail_mem; |
2340 | 0 | } |
2341 | 0 | for (i = 0; i < hdev->nvqs; ++i) { |
2342 | 0 | r = vhost_virtqueue_start(hdev, |
2343 | 0 | vdev, |
2344 | 0 | hdev->vqs + i, |
2345 | 0 | hdev->vq_index + i); |
2346 | 0 | if (r < 0) { |
2347 | 0 | goto fail_vq; |
2348 | 0 | } |
2349 | 0 | } |
2350 | | |
2351 | 0 | r = event_notifier_init( |
2352 | 0 | &hdev->vqs[VHOST_QUEUE_NUM_CONFIG_INR].masked_config_notifier, 0); |
2353 | 0 | if (r < 0) { |
2354 | 0 | VHOST_OPS_DEBUG(r, "event_notifier_init failed"); |
2355 | 0 | goto fail_vq; |
2356 | 0 | } |
2357 | 0 | event_notifier_test_and_clear( |
2358 | 0 | &hdev->vqs[VHOST_QUEUE_NUM_CONFIG_INR].masked_config_notifier); |
2359 | 0 | if (!vdev->use_guest_notifier_mask) { |
2360 | 0 | vhost_config_mask(hdev, vdev, true); |
2361 | 0 | } |
2362 | 0 | if (hdev->log_enabled) { |
2363 | 0 | uint64_t log_base; |
2364 | |
|
2365 | 0 | hdev->log_size = vhost_get_log_size(hdev); |
2366 | 0 | hdev->log = vhost_log_get(hdev->vhost_ops->backend_type, |
2367 | 0 | hdev->log_size, |
2368 | 0 | vhost_dev_log_is_shared(hdev)); |
2369 | 0 | log_base = (uintptr_t)hdev->log->log; |
2370 | 0 | r = hdev->vhost_ops->vhost_set_log_base(hdev, |
2371 | 0 | hdev->log_size ? log_base : 0, |
2372 | 0 | hdev->log); |
2373 | 0 | if (r < 0) { |
2374 | 0 | VHOST_OPS_DEBUG(r, "vhost_set_log_base failed"); |
2375 | 0 | goto fail_log; |
2376 | 0 | } |
2377 | 0 | vhost_dev_elect_mem_logger(hdev, true); |
2378 | 0 | } |
2379 | 0 | if (vrings) { |
2380 | 0 | r = vhost_dev_set_vring_enable(hdev, true); |
2381 | 0 | if (r) { |
2382 | 0 | goto fail_log; |
2383 | 0 | } |
2384 | 0 | } |
2385 | 0 | if (hdev->vhost_ops->vhost_dev_start) { |
2386 | 0 | r = hdev->vhost_ops->vhost_dev_start(hdev, true); |
2387 | 0 | if (r) { |
2388 | 0 | goto fail_start; |
2389 | 0 | } |
2390 | 0 | } |
2391 | 0 | if (vhost_dev_has_iommu(hdev) && |
2392 | 0 | hdev->vhost_ops->vhost_set_iotlb_callback) { |
2393 | 0 | hdev->vhost_ops->vhost_set_iotlb_callback(hdev, true); |
2394 | | |
2395 | | /* Update used ring information for IOTLB to work correctly, |
2396 | | * vhost-kernel code requires for this.*/ |
2397 | 0 | for (i = 0; i < hdev->nvqs; ++i) { |
2398 | 0 | struct vhost_virtqueue *vq = hdev->vqs + i; |
2399 | 0 | r = vhost_device_iotlb_miss(hdev, vq->used_phys, true); |
2400 | 0 | if (r) { |
2401 | 0 | goto fail_iotlb; |
2402 | 0 | } |
2403 | 0 | } |
2404 | 0 | } |
2405 | 0 | vhost_start_config_intr(hdev); |
2406 | |
|
2407 | 0 | trace_vhost_dev_start_out(hdev, vdev->name); |
2408 | 0 | return 0; |
2409 | 0 | fail_iotlb: |
2410 | 0 | if (vhost_dev_has_iommu(hdev) && |
2411 | 0 | hdev->vhost_ops->vhost_set_iotlb_callback) { |
2412 | 0 | hdev->vhost_ops->vhost_set_iotlb_callback(hdev, false); |
2413 | 0 | } |
2414 | 0 | if (hdev->vhost_ops->vhost_dev_start) { |
2415 | 0 | hdev->vhost_ops->vhost_dev_start(hdev, false); |
2416 | 0 | } |
2417 | 0 | fail_start: |
2418 | 0 | if (vrings) { |
2419 | 0 | vhost_dev_set_vring_enable(hdev, false); |
2420 | 0 | } |
2421 | 0 | fail_log: |
2422 | 0 | vhost_log_put(hdev, false); |
2423 | 0 | fail_vq: |
2424 | 0 | while (--i >= 0) { |
2425 | 0 | vhost_virtqueue_stop(hdev, |
2426 | 0 | vdev, |
2427 | 0 | hdev->vqs + i, |
2428 | 0 | hdev->vq_index + i, |
2429 | 0 | false); |
2430 | 0 | } |
2431 | |
|
2432 | 0 | fail_mem: |
2433 | 0 | if (vhost_dev_has_iommu(hdev)) { |
2434 | 0 | memory_listener_unregister(&hdev->iommu_listener); |
2435 | 0 | } |
2436 | 0 | fail_features: |
2437 | 0 | vdev->vhost_started = false; |
2438 | 0 | hdev->started = false; |
2439 | 0 | return r; |
2440 | 0 | } |
2441 | | |
2442 | | /* Host notifiers must be enabled at this point. */ |
2443 | | static int do_vhost_dev_stop(struct vhost_dev *hdev, VirtIODevice *vdev, |
2444 | | bool vrings, bool force, bool skip_drain) |
2445 | 0 | { |
2446 | 0 | int i; |
2447 | 0 | int rc = 0; |
2448 | 0 | EventNotifier *config_notifier = virtio_config_get_guest_notifier(vdev); |
2449 | | |
2450 | | /* should only be called after backend is connected */ |
2451 | 0 | assert(hdev->vhost_ops); |
2452 | 0 | event_notifier_test_and_clear( |
2453 | 0 | &hdev->vqs[VHOST_QUEUE_NUM_CONFIG_INR].masked_config_notifier); |
2454 | 0 | event_notifier_test_and_clear(config_notifier); |
2455 | 0 | event_notifier_cleanup( |
2456 | 0 | &hdev->vqs[VHOST_QUEUE_NUM_CONFIG_INR].masked_config_notifier); |
2457 | |
|
2458 | 0 | trace_vhost_dev_stop_in(hdev, vdev->name, vrings); |
2459 | |
|
2460 | 0 | if (hdev->vhost_ops->vhost_dev_start) { |
2461 | 0 | hdev->vhost_ops->vhost_dev_start(hdev, false); |
2462 | 0 | } |
2463 | 0 | if (vrings) { |
2464 | 0 | vhost_dev_set_vring_enable(hdev, false); |
2465 | 0 | } |
2466 | 0 | for (i = 0; i < hdev->nvqs; ++i) { |
2467 | 0 | rc |= do_vhost_virtqueue_stop(hdev, |
2468 | 0 | vdev, |
2469 | 0 | hdev->vqs + i, |
2470 | 0 | hdev->vq_index + i, |
2471 | 0 | force, |
2472 | 0 | skip_drain); |
2473 | 0 | } |
2474 | 0 | if (hdev->vhost_ops->vhost_reset_status) { |
2475 | 0 | hdev->vhost_ops->vhost_reset_status(hdev); |
2476 | 0 | } |
2477 | |
|
2478 | 0 | if (vhost_dev_has_iommu(hdev)) { |
2479 | 0 | if (hdev->vhost_ops->vhost_set_iotlb_callback) { |
2480 | 0 | hdev->vhost_ops->vhost_set_iotlb_callback(hdev, false); |
2481 | 0 | } |
2482 | 0 | memory_listener_unregister(&hdev->iommu_listener); |
2483 | 0 | } |
2484 | 0 | vhost_stop_config_intr(hdev); |
2485 | 0 | vhost_log_put(hdev, true); |
2486 | 0 | hdev->started = false; |
2487 | 0 | vdev->vhost_started = false; |
2488 | 0 | hdev->vdev = NULL; |
2489 | |
|
2490 | 0 | trace_vhost_dev_stop_out(hdev, vdev->name); |
2491 | 0 | return rc; |
2492 | 0 | } |
2493 | | |
2494 | | int vhost_dev_stop(struct vhost_dev *hdev, VirtIODevice *vdev, bool vrings, |
2495 | | bool skip_drain) |
2496 | 0 | { |
2497 | 0 | return do_vhost_dev_stop(hdev, vdev, vrings, false, skip_drain); |
2498 | 0 | } |
2499 | | |
2500 | | int vhost_dev_force_stop(struct vhost_dev *hdev, VirtIODevice *vdev, |
2501 | | bool vrings) |
2502 | 0 | { |
2503 | 0 | return do_vhost_dev_stop(hdev, vdev, vrings, true, false); |
2504 | 0 | } |
2505 | | |
2506 | | int vhost_net_set_backend(struct vhost_dev *hdev, |
2507 | | struct vhost_vring_file *file) |
2508 | 0 | { |
2509 | 0 | if (hdev->vhost_ops->vhost_net_set_backend) { |
2510 | 0 | return hdev->vhost_ops->vhost_net_set_backend(hdev, file); |
2511 | 0 | } |
2512 | | |
2513 | 0 | return -ENOSYS; |
2514 | 0 | } |
2515 | | |
2516 | | int vhost_reset_device(struct vhost_dev *hdev) |
2517 | 0 | { |
2518 | 0 | if (hdev->vhost_ops->vhost_reset_device) { |
2519 | 0 | return hdev->vhost_ops->vhost_reset_device(hdev); |
2520 | 0 | } |
2521 | | |
2522 | 0 | return -ENOSYS; |
2523 | 0 | } |
2524 | | |
2525 | | bool vhost_supports_device_state(struct vhost_dev *dev) |
2526 | 0 | { |
2527 | 0 | if (dev->vhost_ops->vhost_supports_device_state) { |
2528 | 0 | return dev->vhost_ops->vhost_supports_device_state(dev); |
2529 | 0 | } |
2530 | | |
2531 | 0 | return false; |
2532 | 0 | } |
2533 | | |
2534 | | int vhost_set_device_state_fd(struct vhost_dev *dev, |
2535 | | VhostDeviceStateDirection direction, |
2536 | | VhostDeviceStatePhase phase, |
2537 | | int fd, |
2538 | | int *reply_fd, |
2539 | | Error **errp) |
2540 | 0 | { |
2541 | 0 | if (dev->vhost_ops->vhost_set_device_state_fd) { |
2542 | 0 | return dev->vhost_ops->vhost_set_device_state_fd(dev, direction, phase, |
2543 | 0 | fd, reply_fd, errp); |
2544 | 0 | } |
2545 | | |
2546 | 0 | error_setg(errp, |
2547 | 0 | "vhost transport does not support migration state transfer"); |
2548 | 0 | return -ENOSYS; |
2549 | 0 | } |
2550 | | |
2551 | | int vhost_check_device_state(struct vhost_dev *dev, Error **errp) |
2552 | 0 | { |
2553 | 0 | if (dev->vhost_ops->vhost_check_device_state) { |
2554 | 0 | return dev->vhost_ops->vhost_check_device_state(dev, errp); |
2555 | 0 | } |
2556 | | |
2557 | 0 | error_setg(errp, |
2558 | 0 | "vhost transport does not support migration state transfer"); |
2559 | 0 | return -ENOSYS; |
2560 | 0 | } |
2561 | | |
2562 | | int vhost_save_backend_state(struct vhost_dev *dev, QEMUFile *f, Error **errp) |
2563 | 0 | { |
2564 | 0 | ERRP_GUARD(); |
2565 | | /* Maximum chunk size in which to transfer the state */ |
2566 | 0 | const size_t chunk_size = 1 * 1024 * 1024; |
2567 | 0 | g_autofree void *transfer_buf = NULL; |
2568 | 0 | g_autoptr(GError) g_err = NULL; |
2569 | 0 | int pipe_fds[2], read_fd = -1, write_fd = -1, reply_fd = -1; |
2570 | 0 | int ret; |
2571 | | |
2572 | | /* [0] for reading (our end), [1] for writing (back-end's end) */ |
2573 | 0 | if (!g_unix_open_pipe(pipe_fds, FD_CLOEXEC, &g_err)) { |
2574 | 0 | error_setg(errp, "Failed to set up state transfer pipe: %s", |
2575 | 0 | g_err->message); |
2576 | 0 | ret = -EINVAL; |
2577 | 0 | goto fail; |
2578 | 0 | } |
2579 | | |
2580 | 0 | read_fd = pipe_fds[0]; |
2581 | 0 | write_fd = pipe_fds[1]; |
2582 | | |
2583 | | /* |
2584 | | * VHOST_TRANSFER_STATE_PHASE_STOPPED means the device must be stopped. |
2585 | | * Ideally, it is suspended, but SUSPEND/RESUME currently do not exist for |
2586 | | * vhost-user, so just check that it is stopped at all. |
2587 | | */ |
2588 | 0 | assert(!dev->started); |
2589 | | |
2590 | | /* Transfer ownership of write_fd to the back-end */ |
2591 | 0 | ret = vhost_set_device_state_fd(dev, |
2592 | 0 | VHOST_TRANSFER_STATE_DIRECTION_SAVE, |
2593 | 0 | VHOST_TRANSFER_STATE_PHASE_STOPPED, |
2594 | 0 | write_fd, |
2595 | 0 | &reply_fd, |
2596 | 0 | errp); |
2597 | 0 | if (ret < 0) { |
2598 | 0 | error_prepend(errp, "Failed to initiate state transfer: "); |
2599 | 0 | goto fail; |
2600 | 0 | } |
2601 | | |
2602 | | /* If the back-end wishes to use a different pipe, switch over */ |
2603 | 0 | if (reply_fd >= 0) { |
2604 | 0 | close(read_fd); |
2605 | 0 | read_fd = reply_fd; |
2606 | 0 | } |
2607 | |
|
2608 | 0 | transfer_buf = g_malloc(chunk_size); |
2609 | |
|
2610 | 0 | while (true) { |
2611 | 0 | ssize_t read_ret; |
2612 | |
|
2613 | 0 | read_ret = RETRY_ON_EINTR(read(read_fd, transfer_buf, chunk_size)); |
2614 | 0 | if (read_ret < 0) { |
2615 | 0 | ret = -errno; |
2616 | 0 | error_setg_errno(errp, -ret, "Failed to receive state"); |
2617 | 0 | goto fail; |
2618 | 0 | } |
2619 | | |
2620 | 0 | assert(read_ret <= chunk_size); |
2621 | 0 | qemu_put_be32(f, read_ret); |
2622 | |
|
2623 | 0 | if (read_ret == 0) { |
2624 | | /* EOF */ |
2625 | 0 | break; |
2626 | 0 | } |
2627 | | |
2628 | 0 | qemu_put_buffer(f, transfer_buf, read_ret); |
2629 | 0 | } |
2630 | | |
2631 | | /* |
2632 | | * Back-end will not really care, but be clean and close our end of the pipe |
2633 | | * before inquiring the back-end about whether transfer was successful |
2634 | | */ |
2635 | 0 | close(read_fd); |
2636 | 0 | read_fd = -1; |
2637 | | |
2638 | | /* Also, verify that the device is still stopped */ |
2639 | 0 | assert(!dev->started); |
2640 | |
|
2641 | 0 | ret = vhost_check_device_state(dev, errp); |
2642 | 0 | if (ret < 0) { |
2643 | 0 | goto fail; |
2644 | 0 | } |
2645 | | |
2646 | 0 | ret = 0; |
2647 | 0 | fail: |
2648 | 0 | if (read_fd >= 0) { |
2649 | 0 | close(read_fd); |
2650 | 0 | } |
2651 | |
|
2652 | 0 | return ret; |
2653 | 0 | } |
2654 | | |
2655 | | int vhost_load_backend_state(struct vhost_dev *dev, QEMUFile *f, Error **errp) |
2656 | 0 | { |
2657 | 0 | ERRP_GUARD(); |
2658 | 0 | size_t transfer_buf_size = 0; |
2659 | 0 | g_autofree void *transfer_buf = NULL; |
2660 | 0 | g_autoptr(GError) g_err = NULL; |
2661 | 0 | int pipe_fds[2], read_fd = -1, write_fd = -1, reply_fd = -1; |
2662 | 0 | int ret; |
2663 | | |
2664 | | /* [0] for reading (back-end's end), [1] for writing (our end) */ |
2665 | 0 | if (!g_unix_open_pipe(pipe_fds, FD_CLOEXEC, &g_err)) { |
2666 | 0 | error_setg(errp, "Failed to set up state transfer pipe: %s", |
2667 | 0 | g_err->message); |
2668 | 0 | ret = -EINVAL; |
2669 | 0 | goto fail; |
2670 | 0 | } |
2671 | | |
2672 | 0 | read_fd = pipe_fds[0]; |
2673 | 0 | write_fd = pipe_fds[1]; |
2674 | | |
2675 | | /* |
2676 | | * VHOST_TRANSFER_STATE_PHASE_STOPPED means the device must be stopped. |
2677 | | * Ideally, it is suspended, but SUSPEND/RESUME currently do not exist for |
2678 | | * vhost-user, so just check that it is stopped at all. |
2679 | | */ |
2680 | 0 | assert(!dev->started); |
2681 | | |
2682 | | /* Transfer ownership of read_fd to the back-end */ |
2683 | 0 | ret = vhost_set_device_state_fd(dev, |
2684 | 0 | VHOST_TRANSFER_STATE_DIRECTION_LOAD, |
2685 | 0 | VHOST_TRANSFER_STATE_PHASE_STOPPED, |
2686 | 0 | read_fd, |
2687 | 0 | &reply_fd, |
2688 | 0 | errp); |
2689 | 0 | if (ret < 0) { |
2690 | 0 | error_prepend(errp, "Failed to initiate state transfer: "); |
2691 | 0 | goto fail; |
2692 | 0 | } |
2693 | | |
2694 | | /* If the back-end wishes to use a different pipe, switch over */ |
2695 | 0 | if (reply_fd >= 0) { |
2696 | 0 | close(write_fd); |
2697 | 0 | write_fd = reply_fd; |
2698 | 0 | } |
2699 | |
|
2700 | 0 | while (true) { |
2701 | 0 | size_t this_chunk_size = qemu_get_be32(f); |
2702 | 0 | ssize_t write_ret; |
2703 | 0 | const uint8_t *transfer_pointer; |
2704 | |
|
2705 | 0 | if (this_chunk_size == 0) { |
2706 | | /* End of state */ |
2707 | 0 | break; |
2708 | 0 | } |
2709 | | |
2710 | 0 | if (transfer_buf_size < this_chunk_size) { |
2711 | 0 | transfer_buf = g_realloc(transfer_buf, this_chunk_size); |
2712 | 0 | transfer_buf_size = this_chunk_size; |
2713 | 0 | } |
2714 | |
|
2715 | 0 | if (qemu_get_buffer(f, transfer_buf, this_chunk_size) < |
2716 | 0 | this_chunk_size) |
2717 | 0 | { |
2718 | 0 | error_setg(errp, "Failed to read state"); |
2719 | 0 | ret = -EINVAL; |
2720 | 0 | goto fail; |
2721 | 0 | } |
2722 | | |
2723 | 0 | transfer_pointer = transfer_buf; |
2724 | 0 | while (this_chunk_size > 0) { |
2725 | 0 | write_ret = RETRY_ON_EINTR( |
2726 | 0 | write(write_fd, transfer_pointer, this_chunk_size) |
2727 | 0 | ); |
2728 | 0 | if (write_ret < 0) { |
2729 | 0 | ret = -errno; |
2730 | 0 | error_setg_errno(errp, -ret, "Failed to send state"); |
2731 | 0 | goto fail; |
2732 | 0 | } else if (write_ret == 0) { |
2733 | 0 | error_setg(errp, "Failed to send state: Connection is closed"); |
2734 | 0 | ret = -ECONNRESET; |
2735 | 0 | goto fail; |
2736 | 0 | } |
2737 | | |
2738 | 0 | assert(write_ret <= this_chunk_size); |
2739 | 0 | this_chunk_size -= write_ret; |
2740 | 0 | transfer_pointer += write_ret; |
2741 | 0 | } |
2742 | 0 | } |
2743 | | |
2744 | | /* |
2745 | | * Close our end, thus ending transfer, before inquiring the back-end about |
2746 | | * whether transfer was successful |
2747 | | */ |
2748 | 0 | close(write_fd); |
2749 | 0 | write_fd = -1; |
2750 | | |
2751 | | /* Also, verify that the device is still stopped */ |
2752 | 0 | assert(!dev->started); |
2753 | |
|
2754 | 0 | ret = vhost_check_device_state(dev, errp); |
2755 | 0 | if (ret < 0) { |
2756 | 0 | goto fail; |
2757 | 0 | } |
2758 | | |
2759 | 0 | ret = 0; |
2760 | 0 | fail: |
2761 | 0 | if (write_fd >= 0) { |
2762 | 0 | close(write_fd); |
2763 | 0 | } |
2764 | |
|
2765 | 0 | return ret; |
2766 | 0 | } |