| 1 | /* |
| 2 | * vfio PCI device over a UNIX socket. |
| 3 | * |
| 4 | * Copyright © 2018, 2021 Oracle and/or its affiliates. |
| 5 | * |
| 6 | * SPDX-License-Identifier: GPL-2.0-or-later |
| 7 | */ |
| 8 | |
| 9 | #include "qemu/osdep.h" |
| 10 | #include <sys/ioctl.h> |
| 11 | #include "qapi-visit-sockets.h" |
| 12 | #include "qemu/error-report.h" |
| 13 | |
| 14 | #include "hw/core/qdev-properties.h" |
| 15 | #include "hw/vfio/pci.h" |
| 16 | #include "hw/vfio-user/device.h" |
| 17 | #include "hw/vfio-user/proxy.h" |
| 18 | |
| 19 | #define TYPE_VFIO_USER_PCI "vfio-user-pci" |
| 20 | OBJECT_DECLARE_SIMPLE_TYPE(VFIOUserPCIDevice, VFIO_USER_PCI) |
| 21 | |
| 22 | struct VFIOUserPCIDevice { |
| 23 | VFIOPCIDevice parent_obj; |
| 24 | |
| 25 | SocketAddress *socket; |
| 26 | bool send_queued; /* all sends are queued */ |
| 27 | uint32_t wait_time; /* timeout for message replies */ |
| 28 | bool no_post; /* all region writes are sync */ |
| 29 | }; |
| 30 | |
| 31 | /* |
| 32 | * The server maintains the device's pending interrupts, |
| 33 | * via its MSIX table and PBA, so we treat these accesses |
| 34 | * like PCI config space and forward them. |
| 35 | */ |
| 36 | static uint64_t vfio_user_pba_read(void *opaque, hwaddr addr, |
| 37 | unsigned size) |
| 38 | { |
| 39 | VFIOPCIDevice *vdev = opaque; |
| 40 | VFIORegion *region = &vdev->bars[vdev->msix->pba_bar].region; |
| 41 | uint64_t data; |
| 42 | |
| 43 | /* server copy is what matters */ |
| 44 | data = vfio_region_read(region, addr + vdev->msix->pba_offset, size); |
| 45 | return data; |
| 46 | } |
| 47 | |
| 48 | static void vfio_user_pba_write(void *opaque, hwaddr addr, |
| 49 | uint64_t data, unsigned size) |
| 50 | { |
| 51 | /* dropped */ |
| 52 | } |
| 53 | |
| 54 | static const MemoryRegionOps vfio_user_pba_ops = { |
| 55 | .read = vfio_user_pba_read, |
| 56 | .write = vfio_user_pba_write, |
| 57 | .endianness = DEVICE_LITTLE_ENDIAN, |
| 58 | }; |
| 59 | |
| 60 | static void vfio_user_msix_setup(VFIOPCIDevice *vdev) |
| 61 | { |
| 62 | MemoryRegion *vfio_reg, *msix_reg, *pba_reg; |
| 63 | |
| 64 | pba_reg = g_new0(MemoryRegion, 1); |
| 65 | vdev->msix->pba_region = pba_reg; |
| 66 | |
| 67 | vfio_reg = vdev->bars[vdev->msix->pba_bar].mr; |
| 68 | msix_reg = &PCI_DEVICE(vdev)->msix_pba_mmio; |
| 69 | memory_region_init_io(pba_reg, OBJECT(vdev), &vfio_user_pba_ops, vdev, |
| 70 | "VFIO MSIX PBA", int128_get64(msix_reg->size)); |
| 71 | memory_region_add_subregion_overlap(vfio_reg, vdev->msix->pba_offset, |
| 72 | pba_reg, 1); |
| 73 | } |
| 74 | |
| 75 | static void vfio_user_msix_teardown(VFIOPCIDevice *vdev) |
| 76 | { |
| 77 | MemoryRegion *mr, *sub; |
| 78 | |
| 79 | mr = vdev->bars[vdev->msix->pba_bar].mr; |
| 80 | sub = vdev->msix->pba_region; |
| 81 | memory_region_del_subregion(mr, sub); |
| 82 | |
| 83 | g_free(vdev->msix->pba_region); |
| 84 | vdev->msix->pba_region = NULL; |
| 85 | } |
| 86 | |
| 87 | static void vfio_user_dma_read(VFIOPCIDevice *vdev, VFIOUserDMARW *msg) |
| 88 | { |
| 89 | PCIDevice *pdev = PCI_DEVICE(vdev); |
| 90 | VFIOUserProxy *proxy = vdev->vbasedev.proxy; |
| 91 | VFIOUserDMARW *res; |
| 92 | MemTxResult r; |
| 93 | size_t size; |
| 94 | |
| 95 | if (msg->hdr.size < sizeof(*msg)) { |
| 96 | vfio_user_send_error(proxy, &msg->hdr, EINVAL); |
| 97 | return; |
| 98 | } |
| 99 | if (msg->count > proxy->max_xfer_size) { |
| 100 | vfio_user_send_error(proxy, &msg->hdr, E2BIG); |
| 101 | return; |
| 102 | } |
| 103 | |
| 104 | /* switch to our own message buffer */ |
| 105 | size = msg->count + sizeof(VFIOUserDMARW); |
| 106 | res = g_malloc0(size); |
| 107 | memcpy(res, msg, sizeof(*res)); |
| 108 | g_free(msg); |
| 109 | |
| 110 | r = pci_dma_read(pdev, res->offset, &res->data, res->count); |
| 111 | |
| 112 | /* |
| 113 | * pci_dma_read() doesn't support reporting short reads via the reply's |
| 114 | * count parameter; in this case, we'll reply with an error instead. |
| 115 | */ |
| 116 | switch (r) { |
| 117 | case MEMTX_OK: |
| 118 | if (res->hdr.flags & VFIO_USER_NO_REPLY) { |
| 119 | g_free(res); |
| 120 | return; |
| 121 | } |
| 122 | vfio_user_send_reply(proxy, &res->hdr, size); |
| 123 | break; |
| 124 | case MEMTX_ERROR: |
| 125 | vfio_user_send_error(proxy, &res->hdr, EFAULT); |
| 126 | break; |
| 127 | case MEMTX_DECODE_ERROR: |
| 128 | vfio_user_send_error(proxy, &res->hdr, ENODEV); |
| 129 | break; |
| 130 | case MEMTX_ACCESS_ERROR: |
| 131 | vfio_user_send_error(proxy, &res->hdr, EPERM); |
| 132 | break; |
| 133 | default: |
| 134 | error_printf("vfio_user_dma_read unknown error %d\n", r); |
| 135 | vfio_user_send_error(vdev->vbasedev.proxy, &res->hdr, EINVAL); |
| 136 | } |
| 137 | } |
| 138 | |
| 139 | static void vfio_user_dma_write(VFIOPCIDevice *vdev, VFIOUserDMARW *msg) |
| 140 | { |
| 141 | PCIDevice *pdev = PCI_DEVICE(vdev); |
| 142 | VFIOUserProxy *proxy = vdev->vbasedev.proxy; |
| 143 | VFIOUserDMARW *res; |
| 144 | MemTxResult r; |
| 145 | |
| 146 | if (msg->hdr.size < sizeof(*msg)) { |
| 147 | vfio_user_send_error(proxy, &msg->hdr, EINVAL); |
| 148 | return; |
| 149 | } |
| 150 | /* make sure transfer count isn't larger than the message data */ |
| 151 | if (msg->count > msg->hdr.size - sizeof(*msg)) { |
| 152 | vfio_user_send_error(proxy, &msg->hdr, E2BIG); |
| 153 | return; |
| 154 | } |
| 155 | |
| 156 | r = pci_dma_write(pdev, msg->offset, &msg->data, msg->count); |
| 157 | |
| 158 | res = g_malloc0(sizeof(*res)); |
| 159 | memcpy(res, msg, sizeof(*res)); |
| 160 | g_free(msg); |
| 161 | |
| 162 | /* |
| 163 | * pci_dma_write() doesn't support reporting short writes via the reply's |
| 164 | * count parameter; in this case, we'll reply with an error instead. |
| 165 | */ |
| 166 | switch (r) { |
| 167 | case MEMTX_OK: |
| 168 | if (res->hdr.flags & VFIO_USER_NO_REPLY) { |
| 169 | g_free(res); |
| 170 | return; |
| 171 | } |
| 172 | |
| 173 | vfio_user_send_reply(proxy, &res->hdr, sizeof(*res)); |
| 174 | break; |
| 175 | case MEMTX_ERROR: |
| 176 | vfio_user_send_error(proxy, &res->hdr, EFAULT); |
| 177 | break; |
| 178 | case MEMTX_DECODE_ERROR: |
| 179 | vfio_user_send_error(proxy, &res->hdr, ENODEV); |
| 180 | break; |
| 181 | case MEMTX_ACCESS_ERROR: |
| 182 | vfio_user_send_error(proxy, &res->hdr, EPERM); |
| 183 | break; |
| 184 | default: |
| 185 | error_printf("vfio_user_dma_write unknown error %d\n", r); |
| 186 | vfio_user_send_error(vdev->vbasedev.proxy, &res->hdr, EINVAL); |
| 187 | } |
| 188 | } |
| 189 | |
| 190 | /* |
| 191 | * Incoming request message callback. |
| 192 | * |
| 193 | * Runs off main loop, so BQL held. |
| 194 | */ |
| 195 | static void vfio_user_pci_process_req(void *opaque, VFIOUserMsg *msg) |
| 196 | { |
| 197 | VFIOPCIDevice *vdev = opaque; |
| 198 | VFIOUserHdr *hdr = msg->hdr; |
| 199 | |
| 200 | /* no incoming PCI requests pass FDs */ |
| 201 | if (msg->fds != NULL) { |
| 202 | vfio_user_send_error(vdev->vbasedev.proxy, hdr, EINVAL); |
| 203 | vfio_user_putfds(msg); |
| 204 | return; |
| 205 | } |
| 206 | |
| 207 | switch (hdr->command) { |
| 208 | case VFIO_USER_DMA_READ: |
| 209 | vfio_user_dma_read(vdev, (VFIOUserDMARW *)hdr); |
| 210 | break; |
| 211 | case VFIO_USER_DMA_WRITE: |
| 212 | vfio_user_dma_write(vdev, (VFIOUserDMARW *)hdr); |
| 213 | break; |
| 214 | default: |
| 215 | error_printf("vfio_user_pci_process_req unknown cmd %d\n", |
| 216 | hdr->command); |
| 217 | vfio_user_send_error(vdev->vbasedev.proxy, hdr, ENOSYS); |
| 218 | } |
| 219 | } |
| 220 | |
| 221 | /* |
| 222 | * Emulated devices don't use host hot reset |
| 223 | */ |
| 224 | static void vfio_user_compute_needs_reset(VFIODevice *vbasedev) |
| 225 | { |
| 226 | vbasedev->needs_reset = false; |
| 227 | } |
| 228 | |
| 229 | static Object *vfio_user_pci_get_object(VFIODevice *vbasedev) |
| 230 | { |
| 231 | VFIOUserPCIDevice *vdev = VFIO_USER_PCI(container_of(vbasedev, |
| 232 | VFIOPCIDevice, |
| 233 | vbasedev)); |
| 234 | |
| 235 | return OBJECT(vdev); |
| 236 | } |
| 237 | |
| 238 | static VFIODeviceOps vfio_user_pci_ops = { |
| 239 | .vfio_compute_needs_reset = vfio_user_compute_needs_reset, |
| 240 | .vfio_eoi = vfio_pci_intx_eoi, |
| 241 | .vfio_get_object = vfio_user_pci_get_object, |
| 242 | /* No live migration support yet. */ |
| 243 | .vfio_save_config = NULL, |
| 244 | .vfio_load_config = NULL, |
| 245 | }; |
| 246 | |
| 247 | static void vfio_user_pci_realize(PCIDevice *pdev, Error **errp) |
| 248 | { |
| 249 | ERRP_GUARD(); |
| 250 | VFIOUserPCIDevice *udev = VFIO_USER_PCI(pdev); |
| 251 | VFIOPCIDevice *vdev = VFIO_PCI_DEVICE(pdev); |
| 252 | VFIODevice *vbasedev = &vdev->vbasedev; |
| 253 | const char *sock_name; |
| 254 | |
| 255 | AddressSpace *as; |
| 256 | SocketAddress addr; |
| 257 | VFIOUserProxy *proxy; |
| 258 | |
| 259 | if (!udev->socket) { |
| 260 | error_setg(errp, "No socket specified"); |
| 261 | error_append_hint(errp, "e.g. -device '{" |
| 262 | "\"driver\":\"vfio-user-pci\", " |
| 263 | "\"socket\": {\"path\": \"/tmp/vfio-user.sock\", " |
| 264 | "\"type\": \"unix\"}'" |
| 265 | "}'\n"); |
| 266 | return; |
| 267 | } |
| 268 | |
| 269 | sock_name = udev->socket->u.q_unix.path; |
| 270 | |
| 271 | vbasedev->name = g_strdup_printf("vfio-user:%s", sock_name); |
| 272 | |
| 273 | memset(&addr, 0, sizeof(addr)); |
| 274 | addr.type = SOCKET_ADDRESS_TYPE_UNIX; |
| 275 | addr.u.q_unix.path = (char *)sock_name; |
| 276 | proxy = vfio_user_connect_dev(&addr, errp); |
| 277 | if (!proxy) { |
| 278 | return; |
| 279 | } |
| 280 | vbasedev->proxy = proxy; |
| 281 | vfio_user_set_handler(vbasedev, vfio_user_pci_process_req, vdev); |
| 282 | |
| 283 | if (udev->send_queued) { |
| 284 | proxy->flags |= VFIO_PROXY_FORCE_QUEUED; |
| 285 | } |
| 286 | |
| 287 | if (udev->no_post) { |
| 288 | proxy->flags |= VFIO_PROXY_NO_POST; |
| 289 | } |
| 290 | |
| 291 | /* user specified or 5 sec default */ |
| 292 | proxy->wait_time = udev->wait_time; |
| 293 | |
| 294 | if (!vfio_user_validate_version(proxy, errp)) { |
| 295 | goto error; |
| 296 | } |
| 297 | |
| 298 | /* |
| 299 | * Use socket-based device I/O instead of vfio kernel driver. |
| 300 | */ |
| 301 | vbasedev->io_ops = &vfio_user_device_io_ops_sock; |
| 302 | |
| 303 | /* |
| 304 | * vfio-user devices are effectively mdevs (don't use a host iommu). |
| 305 | */ |
| 306 | vbasedev->mdev = true; |
| 307 | |
| 308 | /* |
| 309 | * Enable per-region fds. |
| 310 | */ |
| 311 | vbasedev->use_region_fds = true; |
| 312 | |
| 313 | as = pci_device_iommu_address_space(pdev); |
| 314 | if (!vfio_device_attach_by_iommu_type(TYPE_VFIO_IOMMU_USER, |
| 315 | vbasedev->name, vbasedev, |
| 316 | as, errp)) { |
| 317 | goto error; |
| 318 | } |
| 319 | |
| 320 | if (!vfio_pci_populate_device(vdev, errp)) { |
| 321 | goto error; |
| 322 | } |
| 323 | |
| 324 | if (!vfio_pci_config_setup(vdev, errp)) { |
| 325 | goto error; |
| 326 | } |
| 327 | |
| 328 | /* |
| 329 | * vfio_pci_config_setup will have registered the device's BARs |
| 330 | * and setup any MSIX BARs, so errors after it succeeds must |
| 331 | * use out_teardown |
| 332 | */ |
| 333 | |
| 334 | if (!vfio_pci_add_capabilities(vdev, errp)) { |
| 335 | goto out_teardown; |
| 336 | } |
| 337 | |
| 338 | if (vdev->msix != NULL) { |
| 339 | vfio_user_msix_setup(vdev); |
| 340 | } |
| 341 | |
| 342 | if (!vfio_pci_interrupt_setup(vdev, errp)) { |
| 343 | goto out_teardown; |
| 344 | } |
| 345 | |
| 346 | vfio_pci_register_err_notifier(vdev); |
| 347 | vfio_pci_register_req_notifier(vdev); |
| 348 | |
| 349 | return; |
| 350 | |
| 351 | out_teardown: |
| 352 | vfio_pci_teardown_msi(vdev); |
| 353 | vfio_pci_bars_exit(vdev); |
| 354 | error: |
| 355 | error_prepend(errp, VFIO_MSG_PREFIX, vdev->vbasedev.name); |
| 356 | vfio_pci_put_device(vdev); |
| 357 | } |
| 358 | |
| 359 | static void vfio_user_pci_init(Object *obj) |
| 360 | { |
| 361 | PCIDevice *pci_dev = PCI_DEVICE(obj); |
| 362 | VFIOPCIDevice *vdev = VFIO_PCI_DEVICE(obj); |
| 363 | VFIODevice *vbasedev = &vdev->vbasedev; |
| 364 | |
| 365 | device_add_bootindex_property(obj, &vdev->bootindex, |
| 366 | "bootindex", NULL, |
| 367 | &pci_dev->qdev); |
| 368 | vdev->host.domain = ~0U; |
| 369 | vdev->host.bus = ~0U; |
| 370 | vdev->host.slot = ~0U; |
| 371 | vdev->host.function = ~0U; |
| 372 | |
| 373 | vfio_device_init(vbasedev, VFIO_DEVICE_TYPE_PCI, &vfio_user_pci_ops, |
| 374 | DEVICE(vdev), false); |
| 375 | |
| 376 | vdev->nv_gpudirect_clique = 0xFF; |
| 377 | |
| 378 | /* |
| 379 | * QEMU_PCI_CAP_EXPRESS initialization does not depend on QEMU command |
| 380 | * line, therefore, no need to wait to realize like other devices. |
| 381 | */ |
| 382 | pci_dev->cap_present |= QEMU_PCI_CAP_EXPRESS; |
| 383 | } |
| 384 | |
| 385 | static void vfio_user_pci_finalize(Object *obj) |
| 386 | { |
| 387 | VFIOPCIDevice *vdev = VFIO_PCI_DEVICE(obj); |
| 388 | VFIODevice *vbasedev = &vdev->vbasedev; |
| 389 | |
| 390 | if (vdev->msix != NULL) { |
| 391 | vfio_user_msix_teardown(vdev); |
| 392 | } |
| 393 | |
| 394 | vfio_pci_put_device(vdev); |
| 395 | |
| 396 | if (vbasedev->proxy != NULL) { |
| 397 | vfio_user_disconnect(vbasedev->proxy); |
| 398 | } |
| 399 | } |
| 400 | |
| 401 | static void vfio_user_pci_reset(DeviceState *dev) |
| 402 | { |
| 403 | VFIOPCIDevice *vdev = VFIO_PCI_DEVICE(dev); |
| 404 | VFIODevice *vbasedev = &vdev->vbasedev; |
| 405 | |
| 406 | vfio_pci_pre_reset(vdev); |
| 407 | |
| 408 | if (vbasedev->reset_works) { |
| 409 | vfio_user_device_reset(vbasedev->proxy); |
| 410 | } |
| 411 | |
| 412 | vfio_pci_post_reset(vdev); |
| 413 | } |
| 414 | |
| 415 | static const Property vfio_user_pci_properties[] = { |
| 416 | DEFINE_PROP_UINT32("x-pci-vendor-id", VFIOPCIDevice, |
| 417 | vendor_id, PCI_ANY_ID), |
| 418 | DEFINE_PROP_UINT32("x-pci-device-id", VFIOPCIDevice, |
| 419 | device_id, PCI_ANY_ID), |
| 420 | DEFINE_PROP_UINT32("x-pci-sub-vendor-id", VFIOPCIDevice, |
| 421 | sub_vendor_id, PCI_ANY_ID), |
| 422 | DEFINE_PROP_UINT32("x-pci-sub-device-id", VFIOPCIDevice, |
| 423 | sub_device_id, PCI_ANY_ID), |
| 424 | DEFINE_PROP_UINT32("x-pci-class-code", VFIOPCIDevice, |
| 425 | class_code, PCI_ANY_ID), |
| 426 | DEFINE_PROP_BOOL("x-send-queued", VFIOUserPCIDevice, send_queued, false), |
| 427 | DEFINE_PROP_UINT32("x-msg-timeout", VFIOUserPCIDevice, wait_time, 5000), |
| 428 | DEFINE_PROP_BOOL("x-no-posted-writes", VFIOUserPCIDevice, no_post, false), |
| 429 | }; |
| 430 | |
| 431 | static void vfio_user_pci_set_socket(Object *obj, Visitor *v, const char *name, |
| 432 | void *opaque, Error **errp) |
| 433 | { |
| 434 | VFIOUserPCIDevice *udev = VFIO_USER_PCI(obj); |
| 435 | bool success; |
| 436 | |
| 437 | if (VFIO_PCI_DEVICE(udev)->vbasedev.proxy) { |
| 438 | error_setg(errp, "Proxy is connected"); |
| 439 | return; |
| 440 | } |
| 441 | |
| 442 | qapi_free_SocketAddress(udev->socket); |
| 443 | |
| 444 | udev->socket = NULL; |
| 445 | |
| 446 | success = visit_type_SocketAddress(v, name, &udev->socket, errp); |
| 447 | |
| 448 | if (!success) { |
| 449 | return; |
| 450 | } |
| 451 | |
| 452 | if (udev->socket->type != SOCKET_ADDRESS_TYPE_UNIX) { |
| 453 | error_setg(errp, "Unsupported socket type %s", |
| 454 | SocketAddressType_str(udev->socket->type)); |
| 455 | qapi_free_SocketAddress(udev->socket); |
| 456 | udev->socket = NULL; |
| 457 | return; |
| 458 | } |
| 459 | } |
| 460 | |
| 461 | static void vfio_user_pci_class_init(ObjectClass *klass, const void *data) |
| 462 | { |
| 463 | DeviceClass *dc = DEVICE_CLASS(klass); |
| 464 | PCIDeviceClass *pdc = PCI_DEVICE_CLASS(klass); |
| 465 | |
| 466 | device_class_set_legacy_reset(dc, vfio_user_pci_reset); |
| 467 | device_class_set_props(dc, vfio_user_pci_properties); |
| 468 | |
| 469 | object_class_property_add(klass, "socket", "SocketAddress", NULL, |
| 470 | vfio_user_pci_set_socket, NULL, NULL); |
| 471 | object_class_property_set_description(klass, "socket", |
| 472 | "SocketAddress (UNIX sockets only)"); |
| 473 | |
| 474 | dc->desc = "VFIO over socket PCI device assignment"; |
| 475 | pdc->realize = vfio_user_pci_realize; |
| 476 | } |
| 477 | |
| 478 | static const TypeInfo vfio_user_pci_info = { |
| 479 | .name = TYPE_VFIO_USER_PCI, |
| 480 | .parent = TYPE_VFIO_PCI_DEVICE, |
| 481 | .instance_size = sizeof(VFIOUserPCIDevice), |
| 482 | .class_init = vfio_user_pci_class_init, |
| 483 | .instance_init = vfio_user_pci_init, |
| 484 | .instance_finalize = vfio_user_pci_finalize, |
| 485 | }; |
| 486 | |
| 487 | static void register_vfio_user_dev_type(void) |
| 488 | { |
| 489 | type_register_static(&vfio_user_pci_info); |
| 490 | } |
| 491 | |
| 492 | type_init(register_vfio_user_dev_type) |