1 #include "kvm/virtio-pci.h" 2 3 #include "kvm/ioport.h" 4 #include "kvm/kvm.h" 5 #include "kvm/kvm-cpu.h" 6 #include "kvm/virtio-pci-dev.h" 7 #include "kvm/irq.h" 8 #include "kvm/virtio.h" 9 #include "kvm/ioeventfd.h" 10 #include "kvm/util.h" 11 12 #include <sys/ioctl.h> 13 #include <linux/virtio_pci.h> 14 #include <assert.h> 15 #include <string.h> 16 17 int virtio_pci__add_msix_route(struct virtio_pci *vpci, u32 vec) 18 { 19 int gsi; 20 struct msi_msg *msg; 21 22 if (vec == VIRTIO_MSI_NO_VECTOR) 23 return -EINVAL; 24 25 msg = &vpci->msix_table[vec].msg; 26 gsi = irq__add_msix_route(vpci->kvm, msg, vpci->dev_hdr.dev_num << 3); 27 /* 28 * We don't need IRQ routing if we can use 29 * MSI injection via the KVM_SIGNAL_MSI ioctl. 30 */ 31 if (gsi == -ENXIO && vpci->features & VIRTIO_PCI_F_SIGNAL_MSI) 32 return gsi; 33 34 if (gsi < 0) 35 die("failed to configure MSIs"); 36 37 return gsi; 38 } 39 40 static void virtio_pci__del_msix_route(struct virtio_pci *vpci, u32 gsi) 41 { 42 struct msi_msg msg = { 0 }; 43 44 irq__update_msix_route(vpci->kvm, gsi, &msg); 45 } 46 47 static void virtio_pci__ioevent_callback(struct kvm *kvm, void *param) 48 { 49 struct virtio_pci_ioevent_param *ioeventfd = param; 50 struct virtio_pci *vpci = ioeventfd->vdev->virtio; 51 52 ioeventfd->vdev->ops->notify_vq(kvm, vpci->dev, ioeventfd->vq); 53 } 54 55 int virtio_pci__init_ioeventfd(struct kvm *kvm, struct virtio_device *vdev, 56 u32 vq) 57 { 58 struct ioevent ioevent; 59 struct virtio_pci *vpci = vdev->virtio; 60 u32 mmio_addr = virtio_pci__mmio_addr(vpci); 61 u16 port_addr = virtio_pci__port_addr(vpci); 62 off_t offset = vpci->doorbell_offset; 63 int r, flags = 0; 64 int pio_fd, mmio_fd; 65 66 vpci->ioeventfds[vq] = (struct virtio_pci_ioevent_param) { 67 .vdev = vdev, 68 .vq = vq, 69 }; 70 71 ioevent = (struct ioevent) { 72 .fn = virtio_pci__ioevent_callback, 73 .fn_ptr = &vpci->ioeventfds[vq], 74 .datamatch = vq, 75 .fn_kvm = kvm, 76 }; 77 78 /* 79 * Vhost will poll the eventfd in host kernel side, otherwise we 80 * need to poll in userspace. 81 */ 82 if (!vdev->use_vhost) 83 flags |= IOEVENTFD_FLAG_USER_POLL; 84 85 /* ioport */ 86 ioevent.io_addr = port_addr + offset; 87 ioevent.io_len = sizeof(u16); 88 ioevent.fd = pio_fd = eventfd(0, 0); 89 r = ioeventfd__add_event(&ioevent, flags | IOEVENTFD_FLAG_PIO); 90 if (r) 91 return r; 92 93 /* mmio */ 94 ioevent.io_addr = mmio_addr + offset; 95 ioevent.io_len = sizeof(u16); 96 ioevent.fd = mmio_fd = eventfd(0, 0); 97 r = ioeventfd__add_event(&ioevent, flags); 98 if (r) 99 goto free_ioport_evt; 100 101 if (vdev->ops->notify_vq_eventfd) 102 vdev->ops->notify_vq_eventfd(kvm, vpci->dev, vq, 103 vdev->legacy ? pio_fd : mmio_fd); 104 return 0; 105 106 free_ioport_evt: 107 ioeventfd__del_event(port_addr + offset, vq); 108 return r; 109 } 110 111 int virtio_pci_init_vq(struct kvm *kvm, struct virtio_device *vdev, int vq) 112 { 113 int ret; 114 struct virtio_pci *vpci = vdev->virtio; 115 116 ret = virtio_pci__init_ioeventfd(kvm, vdev, vq); 117 if (ret) { 118 pr_err("couldn't add ioeventfd for vq %d: %d", vq, ret); 119 return ret; 120 } 121 return vdev->ops->init_vq(kvm, vpci->dev, vq); 122 } 123 124 void virtio_pci_exit_vq(struct kvm *kvm, struct virtio_device *vdev, int vq) 125 { 126 struct virtio_pci *vpci = vdev->virtio; 127 u32 mmio_addr = virtio_pci__mmio_addr(vpci); 128 u16 port_addr = virtio_pci__port_addr(vpci); 129 off_t offset = vpci->doorbell_offset; 130 131 virtio_pci__del_msix_route(vpci, vpci->gsis[vq]); 132 vpci->gsis[vq] = 0; 133 vpci->vq_vector[vq] = VIRTIO_MSI_NO_VECTOR; 134 ioeventfd__del_event(mmio_addr + offset, vq); 135 ioeventfd__del_event(port_addr + offset, vq); 136 virtio_exit_vq(kvm, vdev, vpci->dev, vq); 137 } 138 139 static void update_msix_map(struct virtio_pci *vpci, 140 struct msix_table *msix_entry, u32 vecnum) 141 { 142 u32 gsi, i; 143 144 /* Find the GSI number used for that vector */ 145 if (vecnum == vpci->config_vector) { 146 gsi = vpci->config_gsi; 147 } else { 148 for (i = 0; i < VIRTIO_PCI_MAX_VQ; i++) 149 if (vpci->vq_vector[i] == vecnum) 150 break; 151 if (i == VIRTIO_PCI_MAX_VQ) 152 return; 153 gsi = vpci->gsis[i]; 154 } 155 156 if (gsi == 0) 157 return; 158 159 msix_entry = &msix_entry[vecnum]; 160 irq__update_msix_route(vpci->kvm, gsi, &msix_entry->msg); 161 } 162 163 static void virtio_pci__msix_mmio_callback(struct kvm_cpu *vcpu, 164 u64 addr, u8 *data, u32 len, 165 u8 is_write, void *ptr) 166 { 167 struct virtio_device *vdev = ptr; 168 struct virtio_pci *vpci = vdev->virtio; 169 struct msix_table *table; 170 u32 msix_io_addr = virtio_pci__msix_io_addr(vpci); 171 u32 pba_offset; 172 int vecnum; 173 size_t offset; 174 175 BUILD_BUG_ON(VIRTIO_NR_MSIX > (sizeof(vpci->msix_pba) * 8)); 176 177 pba_offset = vpci->pci_hdr.msix.pba_offset & ~PCI_MSIX_TABLE_BIR; 178 if (addr >= msix_io_addr + pba_offset) { 179 /* Read access to PBA */ 180 if (is_write) 181 return; 182 offset = addr - (msix_io_addr + pba_offset); 183 if ((offset + len) > sizeof (vpci->msix_pba)) 184 return; 185 memcpy(data, (void *)&vpci->msix_pba + offset, len); 186 return; 187 } 188 189 table = vpci->msix_table; 190 offset = addr - msix_io_addr; 191 192 vecnum = offset / sizeof(struct msix_table); 193 offset = offset % sizeof(struct msix_table); 194 195 if (!is_write) { 196 memcpy(data, (void *)&table[vecnum] + offset, len); 197 return; 198 } 199 200 memcpy((void *)&table[vecnum] + offset, data, len); 201 202 /* Did we just update the address or payload? */ 203 if (offset < offsetof(struct msix_table, ctrl)) 204 update_msix_map(vpci, table, vecnum); 205 } 206 207 static void virtio_pci__signal_msi(struct kvm *kvm, struct virtio_pci *vpci, 208 int vec) 209 { 210 struct kvm_msi msi = { 211 .address_lo = vpci->msix_table[vec].msg.address_lo, 212 .address_hi = vpci->msix_table[vec].msg.address_hi, 213 .data = vpci->msix_table[vec].msg.data, 214 }; 215 216 if (kvm->msix_needs_devid) { 217 msi.flags = KVM_MSI_VALID_DEVID; 218 msi.devid = vpci->dev_hdr.dev_num << 3; 219 } 220 221 irq__signal_msi(kvm, &msi); 222 } 223 224 int virtio_pci__signal_vq(struct kvm *kvm, struct virtio_device *vdev, u32 vq) 225 { 226 struct virtio_pci *vpci = vdev->virtio; 227 int tbl = vpci->vq_vector[vq]; 228 229 if (virtio_pci__msix_enabled(vpci) && tbl != VIRTIO_MSI_NO_VECTOR) { 230 if (vpci->pci_hdr.msix.ctrl & cpu_to_le16(PCI_MSIX_FLAGS_MASKALL) || 231 vpci->msix_table[tbl].ctrl & cpu_to_le16(PCI_MSIX_ENTRY_CTRL_MASKBIT)) { 232 233 vpci->msix_pba |= 1 << tbl; 234 return 0; 235 } 236 237 if (vpci->features & VIRTIO_PCI_F_SIGNAL_MSI) 238 virtio_pci__signal_msi(kvm, vpci, vpci->vq_vector[vq]); 239 else 240 kvm__irq_trigger(kvm, vpci->gsis[vq]); 241 } else { 242 vpci->isr = VIRTIO_IRQ_HIGH; 243 kvm__irq_line(kvm, vpci->legacy_irq_line, VIRTIO_IRQ_HIGH); 244 } 245 return 0; 246 } 247 248 int virtio_pci__signal_config(struct kvm *kvm, struct virtio_device *vdev) 249 { 250 struct virtio_pci *vpci = vdev->virtio; 251 int tbl = vpci->config_vector; 252 253 if (virtio_pci__msix_enabled(vpci) && tbl != VIRTIO_MSI_NO_VECTOR) { 254 if (vpci->pci_hdr.msix.ctrl & cpu_to_le16(PCI_MSIX_FLAGS_MASKALL) || 255 vpci->msix_table[tbl].ctrl & cpu_to_le16(PCI_MSIX_ENTRY_CTRL_MASKBIT)) { 256 257 vpci->msix_pba |= 1 << tbl; 258 return 0; 259 } 260 261 if (vpci->features & VIRTIO_PCI_F_SIGNAL_MSI) 262 virtio_pci__signal_msi(kvm, vpci, tbl); 263 else 264 kvm__irq_trigger(kvm, vpci->config_gsi); 265 } else { 266 vpci->isr = VIRTIO_PCI_ISR_CONFIG; 267 kvm__irq_trigger(kvm, vpci->legacy_irq_line); 268 } 269 270 return 0; 271 } 272 273 static int virtio_pci__bar_activate(struct kvm *kvm, 274 struct pci_device_header *pci_hdr, 275 int bar_num, void *data) 276 { 277 struct virtio_device *vdev = data; 278 u32 bar_addr, bar_size; 279 int r = -EINVAL; 280 281 assert(bar_num <= 2); 282 283 bar_addr = pci__bar_address(pci_hdr, bar_num); 284 bar_size = pci__bar_size(pci_hdr, bar_num); 285 286 switch (bar_num) { 287 case 0: 288 r = kvm__register_pio(kvm, bar_addr, bar_size, 289 virtio_pci_legacy__io_mmio_callback, 290 vdev); 291 break; 292 case 1: 293 r = kvm__register_mmio(kvm, bar_addr, bar_size, false, 294 virtio_pci_legacy__io_mmio_callback, 295 vdev); 296 break; 297 case 2: 298 r = kvm__register_mmio(kvm, bar_addr, bar_size, false, 299 virtio_pci__msix_mmio_callback, vdev); 300 break; 301 } 302 303 return r; 304 } 305 306 static int virtio_pci__bar_deactivate(struct kvm *kvm, 307 struct pci_device_header *pci_hdr, 308 int bar_num, void *data) 309 { 310 u32 bar_addr; 311 bool success; 312 int r = -EINVAL; 313 314 assert(bar_num <= 2); 315 316 bar_addr = pci__bar_address(pci_hdr, bar_num); 317 318 switch (bar_num) { 319 case 0: 320 r = kvm__deregister_pio(kvm, bar_addr); 321 break; 322 case 1: 323 case 2: 324 success = kvm__deregister_mmio(kvm, bar_addr); 325 /* kvm__deregister_mmio fails when the region is not found. */ 326 r = (success ? 0 : -ENOENT); 327 break; 328 } 329 330 return r; 331 } 332 333 int virtio_pci__init(struct kvm *kvm, void *dev, struct virtio_device *vdev, 334 int device_id, int subsys_id, int class) 335 { 336 struct virtio_pci *vpci = vdev->virtio; 337 u32 mmio_addr, msix_io_block; 338 u16 port_addr; 339 int r; 340 341 vpci->kvm = kvm; 342 vpci->dev = dev; 343 344 BUILD_BUG_ON(!is_power_of_two(PCI_IO_SIZE)); 345 346 port_addr = pci_get_io_port_block(PCI_IO_SIZE); 347 mmio_addr = pci_get_mmio_block(PCI_IO_SIZE); 348 msix_io_block = pci_get_mmio_block(VIRTIO_MSIX_BAR_SIZE); 349 350 vpci->doorbell_offset = VIRTIO_PCI_QUEUE_NOTIFY; 351 352 vpci->pci_hdr = (struct pci_device_header) { 353 .vendor_id = cpu_to_le16(PCI_VENDOR_ID_REDHAT_QUMRANET), 354 .device_id = cpu_to_le16(device_id), 355 .command = PCI_COMMAND_IO | PCI_COMMAND_MEMORY, 356 .header_type = PCI_HEADER_TYPE_NORMAL, 357 .revision_id = 0, 358 .class[0] = class & 0xff, 359 .class[1] = (class >> 8) & 0xff, 360 .class[2] = (class >> 16) & 0xff, 361 .subsys_vendor_id = cpu_to_le16(PCI_SUBSYSTEM_VENDOR_ID_REDHAT_QUMRANET), 362 .subsys_id = cpu_to_le16(subsys_id), 363 .bar[0] = cpu_to_le32(port_addr 364 | PCI_BASE_ADDRESS_SPACE_IO), 365 .bar[1] = cpu_to_le32(mmio_addr 366 | PCI_BASE_ADDRESS_SPACE_MEMORY), 367 .bar[2] = cpu_to_le32(msix_io_block 368 | PCI_BASE_ADDRESS_SPACE_MEMORY), 369 .status = cpu_to_le16(PCI_STATUS_CAP_LIST), 370 .capabilities = (void *)&vpci->pci_hdr.msix - (void *)&vpci->pci_hdr, 371 .bar_size[0] = cpu_to_le32(PCI_IO_SIZE), 372 .bar_size[1] = cpu_to_le32(PCI_IO_SIZE), 373 .bar_size[2] = cpu_to_le32(VIRTIO_MSIX_BAR_SIZE), 374 }; 375 376 r = pci__register_bar_regions(kvm, &vpci->pci_hdr, 377 virtio_pci__bar_activate, 378 virtio_pci__bar_deactivate, vdev); 379 if (r < 0) 380 return r; 381 382 vpci->dev_hdr = (struct device_header) { 383 .bus_type = DEVICE_BUS_PCI, 384 .data = &vpci->pci_hdr, 385 }; 386 387 vpci->pci_hdr.msix.cap = PCI_CAP_ID_MSIX; 388 vpci->pci_hdr.msix.next = 0; 389 /* 390 * We at most have VIRTIO_NR_MSIX entries (VIRTIO_PCI_MAX_VQ 391 * entries for virt queue, VIRTIO_PCI_MAX_CONFIG entries for 392 * config). 393 * 394 * To quote the PCI spec: 395 * 396 * System software reads this field to determine the 397 * MSI-X Table Size N, which is encoded as N-1. 398 * For example, a returned value of "00000000011" 399 * indicates a table size of 4. 400 */ 401 vpci->pci_hdr.msix.ctrl = cpu_to_le16(VIRTIO_NR_MSIX - 1); 402 403 /* Both table and PBA are mapped to the same BAR (2) */ 404 vpci->pci_hdr.msix.table_offset = cpu_to_le32(2); 405 vpci->pci_hdr.msix.pba_offset = cpu_to_le32(2 | VIRTIO_MSIX_TABLE_SIZE); 406 vpci->config_vector = 0; 407 408 if (irq__can_signal_msi(kvm)) 409 vpci->features |= VIRTIO_PCI_F_SIGNAL_MSI; 410 411 vpci->legacy_irq_line = pci__assign_irq(&vpci->pci_hdr); 412 413 r = device__register(&vpci->dev_hdr); 414 if (r < 0) 415 return r; 416 417 return 0; 418 } 419 420 int virtio_pci__reset(struct kvm *kvm, struct virtio_device *vdev) 421 { 422 unsigned int vq; 423 struct virtio_pci *vpci = vdev->virtio; 424 425 virtio_pci__del_msix_route(vpci, vpci->config_gsi); 426 vpci->config_gsi = 0; 427 vpci->config_vector = VIRTIO_MSI_NO_VECTOR; 428 429 for (vq = 0; vq < vdev->ops->get_vq_count(kvm, vpci->dev); vq++) 430 virtio_pci_exit_vq(kvm, vdev, vq); 431 432 return 0; 433 } 434 435 int virtio_pci__exit(struct kvm *kvm, struct virtio_device *vdev) 436 { 437 struct virtio_pci *vpci = vdev->virtio; 438 439 virtio_pci__reset(kvm, vdev); 440 kvm__deregister_mmio(kvm, virtio_pci__mmio_addr(vpci)); 441 kvm__deregister_mmio(kvm, virtio_pci__msix_io_addr(vpci)); 442 kvm__deregister_pio(kvm, virtio_pci__port_addr(vpci)); 443 444 return 0; 445 } 446