root / hw / vhost.c @ 2cae4119
History | View | Annotate | Download (29.3 kB)
| 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 <sys/ioctl.h> |
| 17 |
#include "vhost.h" |
| 18 |
#include "hw/hw.h" |
| 19 |
#include "range.h" |
| 20 |
#include <linux/vhost.h> |
| 21 |
#include "exec-memory.h" |
| 22 |
|
| 23 |
static void vhost_dev_sync_region(struct vhost_dev *dev, |
| 24 |
MemoryRegionSection *section, |
| 25 |
uint64_t mfirst, uint64_t mlast, |
| 26 |
uint64_t rfirst, uint64_t rlast) |
| 27 |
{
|
| 28 |
uint64_t start = MAX(mfirst, rfirst); |
| 29 |
uint64_t end = MIN(mlast, rlast); |
| 30 |
vhost_log_chunk_t *from = dev->log + start / VHOST_LOG_CHUNK; |
| 31 |
vhost_log_chunk_t *to = dev->log + end / VHOST_LOG_CHUNK + 1;
|
| 32 |
uint64_t addr = (start / VHOST_LOG_CHUNK) * VHOST_LOG_CHUNK; |
| 33 |
|
| 34 |
if (end < start) {
|
| 35 |
return;
|
| 36 |
} |
| 37 |
assert(end / VHOST_LOG_CHUNK < dev->log_size); |
| 38 |
assert(start / VHOST_LOG_CHUNK < dev->log_size); |
| 39 |
|
| 40 |
for (;from < to; ++from) {
|
| 41 |
vhost_log_chunk_t log; |
| 42 |
int bit;
|
| 43 |
/* We first check with non-atomic: much cheaper,
|
| 44 |
* and we expect non-dirty to be the common case. */
|
| 45 |
if (!*from) {
|
| 46 |
addr += VHOST_LOG_CHUNK; |
| 47 |
continue;
|
| 48 |
} |
| 49 |
/* Data must be read atomically. We don't really
|
| 50 |
* need the barrier semantics of __sync
|
| 51 |
* builtins, but it's easier to use them than
|
| 52 |
* roll our own. */
|
| 53 |
log = __sync_fetch_and_and(from, 0);
|
| 54 |
while ((bit = sizeof(log) > sizeof(int) ? |
| 55 |
ffsll(log) : ffs(log))) {
|
| 56 |
ram_addr_t ram_addr; |
| 57 |
bit -= 1;
|
| 58 |
ram_addr = section->offset_within_region + bit * VHOST_LOG_PAGE; |
| 59 |
memory_region_set_dirty(section->mr, ram_addr, VHOST_LOG_PAGE); |
| 60 |
log &= ~(0x1ull << bit);
|
| 61 |
} |
| 62 |
addr += VHOST_LOG_CHUNK; |
| 63 |
} |
| 64 |
} |
| 65 |
|
| 66 |
static int vhost_sync_dirty_bitmap(struct vhost_dev *dev, |
| 67 |
MemoryRegionSection *section, |
| 68 |
target_phys_addr_t start_addr, |
| 69 |
target_phys_addr_t end_addr) |
| 70 |
{
|
| 71 |
int i;
|
| 72 |
|
| 73 |
if (!dev->log_enabled || !dev->started) {
|
| 74 |
return 0; |
| 75 |
} |
| 76 |
for (i = 0; i < dev->mem->nregions; ++i) { |
| 77 |
struct vhost_memory_region *reg = dev->mem->regions + i;
|
| 78 |
vhost_dev_sync_region(dev, section, start_addr, end_addr, |
| 79 |
reg->guest_phys_addr, |
| 80 |
range_get_last(reg->guest_phys_addr, |
| 81 |
reg->memory_size)); |
| 82 |
} |
| 83 |
for (i = 0; i < dev->nvqs; ++i) { |
| 84 |
struct vhost_virtqueue *vq = dev->vqs + i;
|
| 85 |
vhost_dev_sync_region(dev, section, start_addr, end_addr, vq->used_phys, |
| 86 |
range_get_last(vq->used_phys, vq->used_size)); |
| 87 |
} |
| 88 |
return 0; |
| 89 |
} |
| 90 |
|
| 91 |
static void vhost_log_sync(MemoryListener *listener, |
| 92 |
MemoryRegionSection *section) |
| 93 |
{
|
| 94 |
struct vhost_dev *dev = container_of(listener, struct vhost_dev, |
| 95 |
memory_listener); |
| 96 |
target_phys_addr_t start_addr = section->offset_within_address_space; |
| 97 |
target_phys_addr_t end_addr = start_addr + section->size; |
| 98 |
|
| 99 |
vhost_sync_dirty_bitmap(dev, section, start_addr, end_addr); |
| 100 |
} |
| 101 |
|
| 102 |
/* Assign/unassign. Keep an unsorted array of non-overlapping
|
| 103 |
* memory regions in dev->mem. */
|
| 104 |
static void vhost_dev_unassign_memory(struct vhost_dev *dev, |
| 105 |
uint64_t start_addr, |
| 106 |
uint64_t size) |
| 107 |
{
|
| 108 |
int from, to, n = dev->mem->nregions;
|
| 109 |
/* Track overlapping/split regions for sanity checking. */
|
| 110 |
int overlap_start = 0, overlap_end = 0, overlap_middle = 0, split = 0; |
| 111 |
|
| 112 |
for (from = 0, to = 0; from < n; ++from, ++to) { |
| 113 |
struct vhost_memory_region *reg = dev->mem->regions + to;
|
| 114 |
uint64_t reglast; |
| 115 |
uint64_t memlast; |
| 116 |
uint64_t change; |
| 117 |
|
| 118 |
/* clone old region */
|
| 119 |
if (to != from) {
|
| 120 |
memcpy(reg, dev->mem->regions + from, sizeof *reg);
|
| 121 |
} |
| 122 |
|
| 123 |
/* No overlap is simple */
|
| 124 |
if (!ranges_overlap(reg->guest_phys_addr, reg->memory_size,
|
| 125 |
start_addr, size)) {
|
| 126 |
continue;
|
| 127 |
} |
| 128 |
|
| 129 |
/* Split only happens if supplied region
|
| 130 |
* is in the middle of an existing one. Thus it can not
|
| 131 |
* overlap with any other existing region. */
|
| 132 |
assert(!split); |
| 133 |
|
| 134 |
reglast = range_get_last(reg->guest_phys_addr, reg->memory_size); |
| 135 |
memlast = range_get_last(start_addr, size); |
| 136 |
|
| 137 |
/* Remove whole region */
|
| 138 |
if (start_addr <= reg->guest_phys_addr && memlast >= reglast) {
|
| 139 |
--dev->mem->nregions; |
| 140 |
--to; |
| 141 |
++overlap_middle; |
| 142 |
continue;
|
| 143 |
} |
| 144 |
|
| 145 |
/* Shrink region */
|
| 146 |
if (memlast >= reglast) {
|
| 147 |
reg->memory_size = start_addr - reg->guest_phys_addr; |
| 148 |
assert(reg->memory_size); |
| 149 |
assert(!overlap_end); |
| 150 |
++overlap_end; |
| 151 |
continue;
|
| 152 |
} |
| 153 |
|
| 154 |
/* Shift region */
|
| 155 |
if (start_addr <= reg->guest_phys_addr) {
|
| 156 |
change = memlast + 1 - reg->guest_phys_addr;
|
| 157 |
reg->memory_size -= change; |
| 158 |
reg->guest_phys_addr += change; |
| 159 |
reg->userspace_addr += change; |
| 160 |
assert(reg->memory_size); |
| 161 |
assert(!overlap_start); |
| 162 |
++overlap_start; |
| 163 |
continue;
|
| 164 |
} |
| 165 |
|
| 166 |
/* This only happens if supplied region
|
| 167 |
* is in the middle of an existing one. Thus it can not
|
| 168 |
* overlap with any other existing region. */
|
| 169 |
assert(!overlap_start); |
| 170 |
assert(!overlap_end); |
| 171 |
assert(!overlap_middle); |
| 172 |
/* Split region: shrink first part, shift second part. */
|
| 173 |
memcpy(dev->mem->regions + n, reg, sizeof *reg);
|
| 174 |
reg->memory_size = start_addr - reg->guest_phys_addr; |
| 175 |
assert(reg->memory_size); |
| 176 |
change = memlast + 1 - reg->guest_phys_addr;
|
| 177 |
reg = dev->mem->regions + n; |
| 178 |
reg->memory_size -= change; |
| 179 |
assert(reg->memory_size); |
| 180 |
reg->guest_phys_addr += change; |
| 181 |
reg->userspace_addr += change; |
| 182 |
/* Never add more than 1 region */
|
| 183 |
assert(dev->mem->nregions == n); |
| 184 |
++dev->mem->nregions; |
| 185 |
++split; |
| 186 |
} |
| 187 |
} |
| 188 |
|
| 189 |
/* Called after unassign, so no regions overlap the given range. */
|
| 190 |
static void vhost_dev_assign_memory(struct vhost_dev *dev, |
| 191 |
uint64_t start_addr, |
| 192 |
uint64_t size, |
| 193 |
uint64_t uaddr) |
| 194 |
{
|
| 195 |
int from, to;
|
| 196 |
struct vhost_memory_region *merged = NULL; |
| 197 |
for (from = 0, to = 0; from < dev->mem->nregions; ++from, ++to) { |
| 198 |
struct vhost_memory_region *reg = dev->mem->regions + to;
|
| 199 |
uint64_t prlast, urlast; |
| 200 |
uint64_t pmlast, umlast; |
| 201 |
uint64_t s, e, u; |
| 202 |
|
| 203 |
/* clone old region */
|
| 204 |
if (to != from) {
|
| 205 |
memcpy(reg, dev->mem->regions + from, sizeof *reg);
|
| 206 |
} |
| 207 |
prlast = range_get_last(reg->guest_phys_addr, reg->memory_size); |
| 208 |
pmlast = range_get_last(start_addr, size); |
| 209 |
urlast = range_get_last(reg->userspace_addr, reg->memory_size); |
| 210 |
umlast = range_get_last(uaddr, size); |
| 211 |
|
| 212 |
/* check for overlapping regions: should never happen. */
|
| 213 |
assert(prlast < start_addr || pmlast < reg->guest_phys_addr); |
| 214 |
/* Not an adjacent or overlapping region - do not merge. */
|
| 215 |
if ((prlast + 1 != start_addr || urlast + 1 != uaddr) && |
| 216 |
(pmlast + 1 != reg->guest_phys_addr ||
|
| 217 |
umlast + 1 != reg->userspace_addr)) {
|
| 218 |
continue;
|
| 219 |
} |
| 220 |
|
| 221 |
if (merged) {
|
| 222 |
--to; |
| 223 |
assert(to >= 0);
|
| 224 |
} else {
|
| 225 |
merged = reg; |
| 226 |
} |
| 227 |
u = MIN(uaddr, reg->userspace_addr); |
| 228 |
s = MIN(start_addr, reg->guest_phys_addr); |
| 229 |
e = MAX(pmlast, prlast); |
| 230 |
uaddr = merged->userspace_addr = u; |
| 231 |
start_addr = merged->guest_phys_addr = s; |
| 232 |
size = merged->memory_size = e - s + 1;
|
| 233 |
assert(merged->memory_size); |
| 234 |
} |
| 235 |
|
| 236 |
if (!merged) {
|
| 237 |
struct vhost_memory_region *reg = dev->mem->regions + to;
|
| 238 |
memset(reg, 0, sizeof *reg); |
| 239 |
reg->memory_size = size; |
| 240 |
assert(reg->memory_size); |
| 241 |
reg->guest_phys_addr = start_addr; |
| 242 |
reg->userspace_addr = uaddr; |
| 243 |
++to; |
| 244 |
} |
| 245 |
assert(to <= dev->mem->nregions + 1);
|
| 246 |
dev->mem->nregions = to; |
| 247 |
} |
| 248 |
|
| 249 |
static uint64_t vhost_get_log_size(struct vhost_dev *dev) |
| 250 |
{
|
| 251 |
uint64_t log_size = 0;
|
| 252 |
int i;
|
| 253 |
for (i = 0; i < dev->mem->nregions; ++i) { |
| 254 |
struct vhost_memory_region *reg = dev->mem->regions + i;
|
| 255 |
uint64_t last = range_get_last(reg->guest_phys_addr, |
| 256 |
reg->memory_size); |
| 257 |
log_size = MAX(log_size, last / VHOST_LOG_CHUNK + 1);
|
| 258 |
} |
| 259 |
for (i = 0; i < dev->nvqs; ++i) { |
| 260 |
struct vhost_virtqueue *vq = dev->vqs + i;
|
| 261 |
uint64_t last = vq->used_phys + vq->used_size - 1;
|
| 262 |
log_size = MAX(log_size, last / VHOST_LOG_CHUNK + 1);
|
| 263 |
} |
| 264 |
return log_size;
|
| 265 |
} |
| 266 |
|
| 267 |
static inline void vhost_dev_log_resize(struct vhost_dev* dev, uint64_t size) |
| 268 |
{
|
| 269 |
vhost_log_chunk_t *log; |
| 270 |
uint64_t log_base; |
| 271 |
int r, i;
|
| 272 |
if (size) {
|
| 273 |
log = g_malloc0(size * sizeof *log);
|
| 274 |
} else {
|
| 275 |
log = NULL;
|
| 276 |
} |
| 277 |
log_base = (uint64_t)(unsigned long)log; |
| 278 |
r = ioctl(dev->control, VHOST_SET_LOG_BASE, &log_base); |
| 279 |
assert(r >= 0);
|
| 280 |
for (i = 0; i < dev->n_mem_sections; ++i) { |
| 281 |
/* Sync only the range covered by the old log */
|
| 282 |
vhost_sync_dirty_bitmap(dev, &dev->mem_sections[i], 0,
|
| 283 |
dev->log_size * VHOST_LOG_CHUNK - 1);
|
| 284 |
} |
| 285 |
if (dev->log) {
|
| 286 |
g_free(dev->log); |
| 287 |
} |
| 288 |
dev->log = log; |
| 289 |
dev->log_size = size; |
| 290 |
} |
| 291 |
|
| 292 |
static int vhost_verify_ring_mappings(struct vhost_dev *dev, |
| 293 |
uint64_t start_addr, |
| 294 |
uint64_t size) |
| 295 |
{
|
| 296 |
int i;
|
| 297 |
for (i = 0; i < dev->nvqs; ++i) { |
| 298 |
struct vhost_virtqueue *vq = dev->vqs + i;
|
| 299 |
target_phys_addr_t l; |
| 300 |
void *p;
|
| 301 |
|
| 302 |
if (!ranges_overlap(start_addr, size, vq->ring_phys, vq->ring_size)) {
|
| 303 |
continue;
|
| 304 |
} |
| 305 |
l = vq->ring_size; |
| 306 |
p = cpu_physical_memory_map(vq->ring_phys, &l, 1);
|
| 307 |
if (!p || l != vq->ring_size) {
|
| 308 |
fprintf(stderr, "Unable to map ring buffer for ring %d\n", i);
|
| 309 |
return -ENOMEM;
|
| 310 |
} |
| 311 |
if (p != vq->ring) {
|
| 312 |
fprintf(stderr, "Ring buffer relocated for ring %d\n", i);
|
| 313 |
return -EBUSY;
|
| 314 |
} |
| 315 |
cpu_physical_memory_unmap(p, l, 0, 0); |
| 316 |
} |
| 317 |
return 0; |
| 318 |
} |
| 319 |
|
| 320 |
static struct vhost_memory_region *vhost_dev_find_reg(struct vhost_dev *dev, |
| 321 |
uint64_t start_addr, |
| 322 |
uint64_t size) |
| 323 |
{
|
| 324 |
int i, n = dev->mem->nregions;
|
| 325 |
for (i = 0; i < n; ++i) { |
| 326 |
struct vhost_memory_region *reg = dev->mem->regions + i;
|
| 327 |
if (ranges_overlap(reg->guest_phys_addr, reg->memory_size,
|
| 328 |
start_addr, size)) {
|
| 329 |
return reg;
|
| 330 |
} |
| 331 |
} |
| 332 |
return NULL; |
| 333 |
} |
| 334 |
|
| 335 |
static bool vhost_dev_cmp_memory(struct vhost_dev *dev, |
| 336 |
uint64_t start_addr, |
| 337 |
uint64_t size, |
| 338 |
uint64_t uaddr) |
| 339 |
{
|
| 340 |
struct vhost_memory_region *reg = vhost_dev_find_reg(dev, start_addr, size);
|
| 341 |
uint64_t reglast; |
| 342 |
uint64_t memlast; |
| 343 |
|
| 344 |
if (!reg) {
|
| 345 |
return true; |
| 346 |
} |
| 347 |
|
| 348 |
reglast = range_get_last(reg->guest_phys_addr, reg->memory_size); |
| 349 |
memlast = range_get_last(start_addr, size); |
| 350 |
|
| 351 |
/* Need to extend region? */
|
| 352 |
if (start_addr < reg->guest_phys_addr || memlast > reglast) {
|
| 353 |
return true; |
| 354 |
} |
| 355 |
/* userspace_addr changed? */
|
| 356 |
return uaddr != reg->userspace_addr + start_addr - reg->guest_phys_addr;
|
| 357 |
} |
| 358 |
|
| 359 |
static void vhost_set_memory(MemoryListener *listener, |
| 360 |
MemoryRegionSection *section, |
| 361 |
bool add)
|
| 362 |
{
|
| 363 |
struct vhost_dev *dev = container_of(listener, struct vhost_dev, |
| 364 |
memory_listener); |
| 365 |
target_phys_addr_t start_addr = section->offset_within_address_space; |
| 366 |
ram_addr_t size = section->size; |
| 367 |
bool log_dirty = memory_region_is_logging(section->mr);
|
| 368 |
int s = offsetof(struct vhost_memory, regions) + |
| 369 |
(dev->mem->nregions + 1) * sizeof dev->mem->regions[0]; |
| 370 |
uint64_t log_size; |
| 371 |
int r;
|
| 372 |
void *ram;
|
| 373 |
|
| 374 |
dev->mem = g_realloc(dev->mem, s); |
| 375 |
|
| 376 |
if (log_dirty) {
|
| 377 |
add = false;
|
| 378 |
} |
| 379 |
|
| 380 |
assert(size); |
| 381 |
|
| 382 |
/* Optimize no-change case. At least cirrus_vga does this a lot at this time. */
|
| 383 |
ram = memory_region_get_ram_ptr(section->mr) + section->offset_within_region; |
| 384 |
if (add) {
|
| 385 |
if (!vhost_dev_cmp_memory(dev, start_addr, size, (uintptr_t)ram)) {
|
| 386 |
/* Region exists with same address. Nothing to do. */
|
| 387 |
return;
|
| 388 |
} |
| 389 |
} else {
|
| 390 |
if (!vhost_dev_find_reg(dev, start_addr, size)) {
|
| 391 |
/* Removing region that we don't access. Nothing to do. */
|
| 392 |
return;
|
| 393 |
} |
| 394 |
} |
| 395 |
|
| 396 |
vhost_dev_unassign_memory(dev, start_addr, size); |
| 397 |
if (add) {
|
| 398 |
/* Add given mapping, merging adjacent regions if any */
|
| 399 |
vhost_dev_assign_memory(dev, start_addr, size, (uintptr_t)ram); |
| 400 |
} else {
|
| 401 |
/* Remove old mapping for this memory, if any. */
|
| 402 |
vhost_dev_unassign_memory(dev, start_addr, size); |
| 403 |
} |
| 404 |
|
| 405 |
if (!dev->started) {
|
| 406 |
return;
|
| 407 |
} |
| 408 |
|
| 409 |
if (dev->started) {
|
| 410 |
r = vhost_verify_ring_mappings(dev, start_addr, size); |
| 411 |
assert(r >= 0);
|
| 412 |
} |
| 413 |
|
| 414 |
if (!dev->log_enabled) {
|
| 415 |
r = ioctl(dev->control, VHOST_SET_MEM_TABLE, dev->mem); |
| 416 |
assert(r >= 0);
|
| 417 |
return;
|
| 418 |
} |
| 419 |
log_size = vhost_get_log_size(dev); |
| 420 |
/* We allocate an extra 4K bytes to log,
|
| 421 |
* to reduce the * number of reallocations. */
|
| 422 |
#define VHOST_LOG_BUFFER (0x1000 / sizeof *dev->log) |
| 423 |
/* To log more, must increase log size before table update. */
|
| 424 |
if (dev->log_size < log_size) {
|
| 425 |
vhost_dev_log_resize(dev, log_size + VHOST_LOG_BUFFER); |
| 426 |
} |
| 427 |
r = ioctl(dev->control, VHOST_SET_MEM_TABLE, dev->mem); |
| 428 |
assert(r >= 0);
|
| 429 |
/* To log less, can only decrease log size after table update. */
|
| 430 |
if (dev->log_size > log_size + VHOST_LOG_BUFFER) {
|
| 431 |
vhost_dev_log_resize(dev, log_size); |
| 432 |
} |
| 433 |
} |
| 434 |
|
| 435 |
static bool vhost_section(MemoryRegionSection *section) |
| 436 |
{
|
| 437 |
return section->address_space == get_system_memory()
|
| 438 |
&& memory_region_is_ram(section->mr); |
| 439 |
} |
| 440 |
|
| 441 |
static void vhost_begin(MemoryListener *listener) |
| 442 |
{
|
| 443 |
} |
| 444 |
|
| 445 |
static void vhost_commit(MemoryListener *listener) |
| 446 |
{
|
| 447 |
} |
| 448 |
|
| 449 |
static void vhost_region_add(MemoryListener *listener, |
| 450 |
MemoryRegionSection *section) |
| 451 |
{
|
| 452 |
struct vhost_dev *dev = container_of(listener, struct vhost_dev, |
| 453 |
memory_listener); |
| 454 |
|
| 455 |
if (!vhost_section(section)) {
|
| 456 |
return;
|
| 457 |
} |
| 458 |
|
| 459 |
++dev->n_mem_sections; |
| 460 |
dev->mem_sections = g_renew(MemoryRegionSection, dev->mem_sections, |
| 461 |
dev->n_mem_sections); |
| 462 |
dev->mem_sections[dev->n_mem_sections - 1] = *section;
|
| 463 |
vhost_set_memory(listener, section, true);
|
| 464 |
} |
| 465 |
|
| 466 |
static void vhost_region_del(MemoryListener *listener, |
| 467 |
MemoryRegionSection *section) |
| 468 |
{
|
| 469 |
struct vhost_dev *dev = container_of(listener, struct vhost_dev, |
| 470 |
memory_listener); |
| 471 |
int i;
|
| 472 |
|
| 473 |
if (!vhost_section(section)) {
|
| 474 |
return;
|
| 475 |
} |
| 476 |
|
| 477 |
vhost_set_memory(listener, section, false);
|
| 478 |
for (i = 0; i < dev->n_mem_sections; ++i) { |
| 479 |
if (dev->mem_sections[i].offset_within_address_space
|
| 480 |
== section->offset_within_address_space) {
|
| 481 |
--dev->n_mem_sections; |
| 482 |
memmove(&dev->mem_sections[i], &dev->mem_sections[i+1],
|
| 483 |
(dev->n_mem_sections - i) * sizeof(*dev->mem_sections));
|
| 484 |
break;
|
| 485 |
} |
| 486 |
} |
| 487 |
} |
| 488 |
|
| 489 |
static void vhost_region_nop(MemoryListener *listener, |
| 490 |
MemoryRegionSection *section) |
| 491 |
{
|
| 492 |
} |
| 493 |
|
| 494 |
static int vhost_virtqueue_set_addr(struct vhost_dev *dev, |
| 495 |
struct vhost_virtqueue *vq,
|
| 496 |
unsigned idx, bool enable_log) |
| 497 |
{
|
| 498 |
struct vhost_vring_addr addr = {
|
| 499 |
.index = idx, |
| 500 |
.desc_user_addr = (uint64_t)(unsigned long)vq->desc, |
| 501 |
.avail_user_addr = (uint64_t)(unsigned long)vq->avail, |
| 502 |
.used_user_addr = (uint64_t)(unsigned long)vq->used, |
| 503 |
.log_guest_addr = vq->used_phys, |
| 504 |
.flags = enable_log ? (1 << VHOST_VRING_F_LOG) : 0, |
| 505 |
}; |
| 506 |
int r = ioctl(dev->control, VHOST_SET_VRING_ADDR, &addr);
|
| 507 |
if (r < 0) { |
| 508 |
return -errno;
|
| 509 |
} |
| 510 |
return 0; |
| 511 |
} |
| 512 |
|
| 513 |
static int vhost_dev_set_features(struct vhost_dev *dev, bool enable_log) |
| 514 |
{
|
| 515 |
uint64_t features = dev->acked_features; |
| 516 |
int r;
|
| 517 |
if (enable_log) {
|
| 518 |
features |= 0x1 << VHOST_F_LOG_ALL;
|
| 519 |
} |
| 520 |
r = ioctl(dev->control, VHOST_SET_FEATURES, &features); |
| 521 |
return r < 0 ? -errno : 0; |
| 522 |
} |
| 523 |
|
| 524 |
static int vhost_dev_set_log(struct vhost_dev *dev, bool enable_log) |
| 525 |
{
|
| 526 |
int r, t, i;
|
| 527 |
r = vhost_dev_set_features(dev, enable_log); |
| 528 |
if (r < 0) { |
| 529 |
goto err_features;
|
| 530 |
} |
| 531 |
for (i = 0; i < dev->nvqs; ++i) { |
| 532 |
r = vhost_virtqueue_set_addr(dev, dev->vqs + i, i, |
| 533 |
enable_log); |
| 534 |
if (r < 0) { |
| 535 |
goto err_vq;
|
| 536 |
} |
| 537 |
} |
| 538 |
return 0; |
| 539 |
err_vq:
|
| 540 |
for (; i >= 0; --i) { |
| 541 |
t = vhost_virtqueue_set_addr(dev, dev->vqs + i, i, |
| 542 |
dev->log_enabled); |
| 543 |
assert(t >= 0);
|
| 544 |
} |
| 545 |
t = vhost_dev_set_features(dev, dev->log_enabled); |
| 546 |
assert(t >= 0);
|
| 547 |
err_features:
|
| 548 |
return r;
|
| 549 |
} |
| 550 |
|
| 551 |
static int vhost_migration_log(MemoryListener *listener, int enable) |
| 552 |
{
|
| 553 |
struct vhost_dev *dev = container_of(listener, struct vhost_dev, |
| 554 |
memory_listener); |
| 555 |
int r;
|
| 556 |
if (!!enable == dev->log_enabled) {
|
| 557 |
return 0; |
| 558 |
} |
| 559 |
if (!dev->started) {
|
| 560 |
dev->log_enabled = enable; |
| 561 |
return 0; |
| 562 |
} |
| 563 |
if (!enable) {
|
| 564 |
r = vhost_dev_set_log(dev, false);
|
| 565 |
if (r < 0) { |
| 566 |
return r;
|
| 567 |
} |
| 568 |
if (dev->log) {
|
| 569 |
g_free(dev->log); |
| 570 |
} |
| 571 |
dev->log = NULL;
|
| 572 |
dev->log_size = 0;
|
| 573 |
} else {
|
| 574 |
vhost_dev_log_resize(dev, vhost_get_log_size(dev)); |
| 575 |
r = vhost_dev_set_log(dev, true);
|
| 576 |
if (r < 0) { |
| 577 |
return r;
|
| 578 |
} |
| 579 |
} |
| 580 |
dev->log_enabled = enable; |
| 581 |
return 0; |
| 582 |
} |
| 583 |
|
| 584 |
static void vhost_log_global_start(MemoryListener *listener) |
| 585 |
{
|
| 586 |
int r;
|
| 587 |
|
| 588 |
r = vhost_migration_log(listener, true);
|
| 589 |
if (r < 0) { |
| 590 |
abort(); |
| 591 |
} |
| 592 |
} |
| 593 |
|
| 594 |
static void vhost_log_global_stop(MemoryListener *listener) |
| 595 |
{
|
| 596 |
int r;
|
| 597 |
|
| 598 |
r = vhost_migration_log(listener, false);
|
| 599 |
if (r < 0) { |
| 600 |
abort(); |
| 601 |
} |
| 602 |
} |
| 603 |
|
| 604 |
static void vhost_log_start(MemoryListener *listener, |
| 605 |
MemoryRegionSection *section) |
| 606 |
{
|
| 607 |
/* FIXME: implement */
|
| 608 |
} |
| 609 |
|
| 610 |
static void vhost_log_stop(MemoryListener *listener, |
| 611 |
MemoryRegionSection *section) |
| 612 |
{
|
| 613 |
/* FIXME: implement */
|
| 614 |
} |
| 615 |
|
| 616 |
static int vhost_virtqueue_init(struct vhost_dev *dev, |
| 617 |
struct VirtIODevice *vdev,
|
| 618 |
struct vhost_virtqueue *vq,
|
| 619 |
unsigned idx)
|
| 620 |
{
|
| 621 |
target_phys_addr_t s, l, a; |
| 622 |
int r;
|
| 623 |
struct vhost_vring_file file = {
|
| 624 |
.index = idx, |
| 625 |
}; |
| 626 |
struct vhost_vring_state state = {
|
| 627 |
.index = idx, |
| 628 |
}; |
| 629 |
struct VirtQueue *vvq = virtio_get_queue(vdev, idx);
|
| 630 |
|
| 631 |
vq->num = state.num = virtio_queue_get_num(vdev, idx); |
| 632 |
r = ioctl(dev->control, VHOST_SET_VRING_NUM, &state); |
| 633 |
if (r) {
|
| 634 |
return -errno;
|
| 635 |
} |
| 636 |
|
| 637 |
state.num = virtio_queue_get_last_avail_idx(vdev, idx); |
| 638 |
r = ioctl(dev->control, VHOST_SET_VRING_BASE, &state); |
| 639 |
if (r) {
|
| 640 |
return -errno;
|
| 641 |
} |
| 642 |
|
| 643 |
s = l = virtio_queue_get_desc_size(vdev, idx); |
| 644 |
a = virtio_queue_get_desc_addr(vdev, idx); |
| 645 |
vq->desc = cpu_physical_memory_map(a, &l, 0);
|
| 646 |
if (!vq->desc || l != s) {
|
| 647 |
r = -ENOMEM; |
| 648 |
goto fail_alloc_desc;
|
| 649 |
} |
| 650 |
s = l = virtio_queue_get_avail_size(vdev, idx); |
| 651 |
a = virtio_queue_get_avail_addr(vdev, idx); |
| 652 |
vq->avail = cpu_physical_memory_map(a, &l, 0);
|
| 653 |
if (!vq->avail || l != s) {
|
| 654 |
r = -ENOMEM; |
| 655 |
goto fail_alloc_avail;
|
| 656 |
} |
| 657 |
vq->used_size = s = l = virtio_queue_get_used_size(vdev, idx); |
| 658 |
vq->used_phys = a = virtio_queue_get_used_addr(vdev, idx); |
| 659 |
vq->used = cpu_physical_memory_map(a, &l, 1);
|
| 660 |
if (!vq->used || l != s) {
|
| 661 |
r = -ENOMEM; |
| 662 |
goto fail_alloc_used;
|
| 663 |
} |
| 664 |
|
| 665 |
vq->ring_size = s = l = virtio_queue_get_ring_size(vdev, idx); |
| 666 |
vq->ring_phys = a = virtio_queue_get_ring_addr(vdev, idx); |
| 667 |
vq->ring = cpu_physical_memory_map(a, &l, 1);
|
| 668 |
if (!vq->ring || l != s) {
|
| 669 |
r = -ENOMEM; |
| 670 |
goto fail_alloc_ring;
|
| 671 |
} |
| 672 |
|
| 673 |
r = vhost_virtqueue_set_addr(dev, vq, idx, dev->log_enabled); |
| 674 |
if (r < 0) { |
| 675 |
r = -errno; |
| 676 |
goto fail_alloc;
|
| 677 |
} |
| 678 |
file.fd = event_notifier_get_fd(virtio_queue_get_host_notifier(vvq)); |
| 679 |
r = ioctl(dev->control, VHOST_SET_VRING_KICK, &file); |
| 680 |
if (r) {
|
| 681 |
r = -errno; |
| 682 |
goto fail_kick;
|
| 683 |
} |
| 684 |
|
| 685 |
file.fd = event_notifier_get_fd(virtio_queue_get_guest_notifier(vvq)); |
| 686 |
r = ioctl(dev->control, VHOST_SET_VRING_CALL, &file); |
| 687 |
if (r) {
|
| 688 |
r = -errno; |
| 689 |
goto fail_call;
|
| 690 |
} |
| 691 |
|
| 692 |
return 0; |
| 693 |
|
| 694 |
fail_call:
|
| 695 |
fail_kick:
|
| 696 |
fail_alloc:
|
| 697 |
cpu_physical_memory_unmap(vq->ring, virtio_queue_get_ring_size(vdev, idx), |
| 698 |
0, 0); |
| 699 |
fail_alloc_ring:
|
| 700 |
cpu_physical_memory_unmap(vq->used, virtio_queue_get_used_size(vdev, idx), |
| 701 |
0, 0); |
| 702 |
fail_alloc_used:
|
| 703 |
cpu_physical_memory_unmap(vq->avail, virtio_queue_get_avail_size(vdev, idx), |
| 704 |
0, 0); |
| 705 |
fail_alloc_avail:
|
| 706 |
cpu_physical_memory_unmap(vq->desc, virtio_queue_get_desc_size(vdev, idx), |
| 707 |
0, 0); |
| 708 |
fail_alloc_desc:
|
| 709 |
return r;
|
| 710 |
} |
| 711 |
|
| 712 |
static void vhost_virtqueue_cleanup(struct vhost_dev *dev, |
| 713 |
struct VirtIODevice *vdev,
|
| 714 |
struct vhost_virtqueue *vq,
|
| 715 |
unsigned idx)
|
| 716 |
{
|
| 717 |
struct vhost_vring_state state = {
|
| 718 |
.index = idx, |
| 719 |
}; |
| 720 |
int r;
|
| 721 |
r = ioctl(dev->control, VHOST_GET_VRING_BASE, &state); |
| 722 |
if (r < 0) { |
| 723 |
fprintf(stderr, "vhost VQ %d ring restore failed: %d\n", idx, r);
|
| 724 |
fflush(stderr); |
| 725 |
} |
| 726 |
virtio_queue_set_last_avail_idx(vdev, idx, state.num); |
| 727 |
assert (r >= 0);
|
| 728 |
cpu_physical_memory_unmap(vq->ring, virtio_queue_get_ring_size(vdev, idx), |
| 729 |
0, virtio_queue_get_ring_size(vdev, idx));
|
| 730 |
cpu_physical_memory_unmap(vq->used, virtio_queue_get_used_size(vdev, idx), |
| 731 |
1, virtio_queue_get_used_size(vdev, idx));
|
| 732 |
cpu_physical_memory_unmap(vq->avail, virtio_queue_get_avail_size(vdev, idx), |
| 733 |
0, virtio_queue_get_avail_size(vdev, idx));
|
| 734 |
cpu_physical_memory_unmap(vq->desc, virtio_queue_get_desc_size(vdev, idx), |
| 735 |
0, virtio_queue_get_desc_size(vdev, idx));
|
| 736 |
} |
| 737 |
|
| 738 |
static void vhost_eventfd_add(MemoryListener *listener, |
| 739 |
MemoryRegionSection *section, |
| 740 |
bool match_data, uint64_t data, EventNotifier *e)
|
| 741 |
{
|
| 742 |
} |
| 743 |
|
| 744 |
static void vhost_eventfd_del(MemoryListener *listener, |
| 745 |
MemoryRegionSection *section, |
| 746 |
bool match_data, uint64_t data, EventNotifier *e)
|
| 747 |
{
|
| 748 |
} |
| 749 |
|
| 750 |
int vhost_dev_init(struct vhost_dev *hdev, int devfd, const char *devpath, |
| 751 |
bool force)
|
| 752 |
{
|
| 753 |
uint64_t features; |
| 754 |
int r;
|
| 755 |
if (devfd >= 0) { |
| 756 |
hdev->control = devfd; |
| 757 |
} else {
|
| 758 |
hdev->control = open(devpath, O_RDWR); |
| 759 |
if (hdev->control < 0) { |
| 760 |
return -errno;
|
| 761 |
} |
| 762 |
} |
| 763 |
r = ioctl(hdev->control, VHOST_SET_OWNER, NULL);
|
| 764 |
if (r < 0) { |
| 765 |
goto fail;
|
| 766 |
} |
| 767 |
|
| 768 |
r = ioctl(hdev->control, VHOST_GET_FEATURES, &features); |
| 769 |
if (r < 0) { |
| 770 |
goto fail;
|
| 771 |
} |
| 772 |
hdev->features = features; |
| 773 |
|
| 774 |
hdev->memory_listener = (MemoryListener) {
|
| 775 |
.begin = vhost_begin, |
| 776 |
.commit = vhost_commit, |
| 777 |
.region_add = vhost_region_add, |
| 778 |
.region_del = vhost_region_del, |
| 779 |
.region_nop = vhost_region_nop, |
| 780 |
.log_start = vhost_log_start, |
| 781 |
.log_stop = vhost_log_stop, |
| 782 |
.log_sync = vhost_log_sync, |
| 783 |
.log_global_start = vhost_log_global_start, |
| 784 |
.log_global_stop = vhost_log_global_stop, |
| 785 |
.eventfd_add = vhost_eventfd_add, |
| 786 |
.eventfd_del = vhost_eventfd_del, |
| 787 |
.priority = 10
|
| 788 |
}; |
| 789 |
hdev->mem = g_malloc0(offsetof(struct vhost_memory, regions));
|
| 790 |
hdev->n_mem_sections = 0;
|
| 791 |
hdev->mem_sections = NULL;
|
| 792 |
hdev->log = NULL;
|
| 793 |
hdev->log_size = 0;
|
| 794 |
hdev->log_enabled = false;
|
| 795 |
hdev->started = false;
|
| 796 |
memory_listener_register(&hdev->memory_listener, NULL);
|
| 797 |
hdev->force = force; |
| 798 |
return 0; |
| 799 |
fail:
|
| 800 |
r = -errno; |
| 801 |
close(hdev->control); |
| 802 |
return r;
|
| 803 |
} |
| 804 |
|
| 805 |
void vhost_dev_cleanup(struct vhost_dev *hdev) |
| 806 |
{
|
| 807 |
memory_listener_unregister(&hdev->memory_listener); |
| 808 |
g_free(hdev->mem); |
| 809 |
g_free(hdev->mem_sections); |
| 810 |
close(hdev->control); |
| 811 |
} |
| 812 |
|
| 813 |
bool vhost_dev_query(struct vhost_dev *hdev, VirtIODevice *vdev) |
| 814 |
{
|
| 815 |
return !vdev->binding->query_guest_notifiers ||
|
| 816 |
vdev->binding->query_guest_notifiers(vdev->binding_opaque) || |
| 817 |
hdev->force; |
| 818 |
} |
| 819 |
|
| 820 |
/* Stop processing guest IO notifications in qemu.
|
| 821 |
* Start processing them in vhost in kernel.
|
| 822 |
*/
|
| 823 |
int vhost_dev_enable_notifiers(struct vhost_dev *hdev, VirtIODevice *vdev) |
| 824 |
{
|
| 825 |
int i, r;
|
| 826 |
if (!vdev->binding->set_host_notifier) {
|
| 827 |
fprintf(stderr, "binding does not support host notifiers\n");
|
| 828 |
r = -ENOSYS; |
| 829 |
goto fail;
|
| 830 |
} |
| 831 |
|
| 832 |
for (i = 0; i < hdev->nvqs; ++i) { |
| 833 |
r = vdev->binding->set_host_notifier(vdev->binding_opaque, i, true);
|
| 834 |
if (r < 0) { |
| 835 |
fprintf(stderr, "vhost VQ %d notifier binding failed: %d\n", i, -r);
|
| 836 |
goto fail_vq;
|
| 837 |
} |
| 838 |
} |
| 839 |
|
| 840 |
return 0; |
| 841 |
fail_vq:
|
| 842 |
while (--i >= 0) { |
| 843 |
r = vdev->binding->set_host_notifier(vdev->binding_opaque, i, false);
|
| 844 |
if (r < 0) { |
| 845 |
fprintf(stderr, "vhost VQ %d notifier cleanup error: %d\n", i, -r);
|
| 846 |
fflush(stderr); |
| 847 |
} |
| 848 |
assert (r >= 0);
|
| 849 |
} |
| 850 |
fail:
|
| 851 |
return r;
|
| 852 |
} |
| 853 |
|
| 854 |
/* Stop processing guest IO notifications in vhost.
|
| 855 |
* Start processing them in qemu.
|
| 856 |
* This might actually run the qemu handlers right away,
|
| 857 |
* so virtio in qemu must be completely setup when this is called.
|
| 858 |
*/
|
| 859 |
void vhost_dev_disable_notifiers(struct vhost_dev *hdev, VirtIODevice *vdev) |
| 860 |
{
|
| 861 |
int i, r;
|
| 862 |
|
| 863 |
for (i = 0; i < hdev->nvqs; ++i) { |
| 864 |
r = vdev->binding->set_host_notifier(vdev->binding_opaque, i, false);
|
| 865 |
if (r < 0) { |
| 866 |
fprintf(stderr, "vhost VQ %d notifier cleanup failed: %d\n", i, -r);
|
| 867 |
fflush(stderr); |
| 868 |
} |
| 869 |
assert (r >= 0);
|
| 870 |
} |
| 871 |
} |
| 872 |
|
| 873 |
/* Host notifiers must be enabled at this point. */
|
| 874 |
int vhost_dev_start(struct vhost_dev *hdev, VirtIODevice *vdev) |
| 875 |
{
|
| 876 |
int i, r;
|
| 877 |
if (!vdev->binding->set_guest_notifiers) {
|
| 878 |
fprintf(stderr, "binding does not support guest notifiers\n");
|
| 879 |
r = -ENOSYS; |
| 880 |
goto fail;
|
| 881 |
} |
| 882 |
|
| 883 |
r = vdev->binding->set_guest_notifiers(vdev->binding_opaque, true);
|
| 884 |
if (r < 0) { |
| 885 |
fprintf(stderr, "Error binding guest notifier: %d\n", -r);
|
| 886 |
goto fail_notifiers;
|
| 887 |
} |
| 888 |
|
| 889 |
r = vhost_dev_set_features(hdev, hdev->log_enabled); |
| 890 |
if (r < 0) { |
| 891 |
goto fail_features;
|
| 892 |
} |
| 893 |
r = ioctl(hdev->control, VHOST_SET_MEM_TABLE, hdev->mem); |
| 894 |
if (r < 0) { |
| 895 |
r = -errno; |
| 896 |
goto fail_mem;
|
| 897 |
} |
| 898 |
for (i = 0; i < hdev->nvqs; ++i) { |
| 899 |
r = vhost_virtqueue_init(hdev, |
| 900 |
vdev, |
| 901 |
hdev->vqs + i, |
| 902 |
i); |
| 903 |
if (r < 0) { |
| 904 |
goto fail_vq;
|
| 905 |
} |
| 906 |
} |
| 907 |
|
| 908 |
if (hdev->log_enabled) {
|
| 909 |
hdev->log_size = vhost_get_log_size(hdev); |
| 910 |
hdev->log = hdev->log_size ? |
| 911 |
g_malloc0(hdev->log_size * sizeof *hdev->log) : NULL; |
| 912 |
r = ioctl(hdev->control, VHOST_SET_LOG_BASE, |
| 913 |
(uint64_t)(unsigned long)hdev->log); |
| 914 |
if (r < 0) { |
| 915 |
r = -errno; |
| 916 |
goto fail_log;
|
| 917 |
} |
| 918 |
} |
| 919 |
|
| 920 |
hdev->started = true;
|
| 921 |
|
| 922 |
return 0; |
| 923 |
fail_log:
|
| 924 |
fail_vq:
|
| 925 |
while (--i >= 0) { |
| 926 |
vhost_virtqueue_cleanup(hdev, |
| 927 |
vdev, |
| 928 |
hdev->vqs + i, |
| 929 |
i); |
| 930 |
} |
| 931 |
fail_mem:
|
| 932 |
fail_features:
|
| 933 |
vdev->binding->set_guest_notifiers(vdev->binding_opaque, false);
|
| 934 |
fail_notifiers:
|
| 935 |
fail:
|
| 936 |
return r;
|
| 937 |
} |
| 938 |
|
| 939 |
/* Host notifiers must be enabled at this point. */
|
| 940 |
void vhost_dev_stop(struct vhost_dev *hdev, VirtIODevice *vdev) |
| 941 |
{
|
| 942 |
int i, r;
|
| 943 |
|
| 944 |
for (i = 0; i < hdev->nvqs; ++i) { |
| 945 |
vhost_virtqueue_cleanup(hdev, |
| 946 |
vdev, |
| 947 |
hdev->vqs + i, |
| 948 |
i); |
| 949 |
} |
| 950 |
for (i = 0; i < hdev->n_mem_sections; ++i) { |
| 951 |
vhost_sync_dirty_bitmap(hdev, &hdev->mem_sections[i], |
| 952 |
0, (target_phys_addr_t)~0x0ull); |
| 953 |
} |
| 954 |
r = vdev->binding->set_guest_notifiers(vdev->binding_opaque, false);
|
| 955 |
if (r < 0) { |
| 956 |
fprintf(stderr, "vhost guest notifier cleanup failed: %d\n", r);
|
| 957 |
fflush(stderr); |
| 958 |
} |
| 959 |
assert (r >= 0);
|
| 960 |
|
| 961 |
hdev->started = false;
|
| 962 |
g_free(hdev->log); |
| 963 |
hdev->log = NULL;
|
| 964 |
hdev->log_size = 0;
|
| 965 |
} |