1 /* 2 * Virtio PCI driver - common functionality for all device versions 3 * 4 * This module allows virtio devices to be used over a virtual PCI device. 5 * This can be used with QEMU based VMMs like KVM or Xen. 6 * 7 * Copyright IBM Corp. 2007 8 * Copyright Red Hat, Inc. 2014 9 * 10 * Authors: 11 * Anthony Liguori <aliguori@us.ibm.com> 12 * Rusty Russell <rusty@rustcorp.com.au> 13 * Michael S. Tsirkin <mst@redhat.com> 14 * 15 * This work is licensed under the terms of the GNU GPL, version 2 or later. 16 * See the COPYING file in the top-level directory. 17 * 18 */ 19 20 #include "virtio_pci_common.h" 21 22 static bool force_legacy = false; 23 24 #if IS_ENABLED(CONFIG_VIRTIO_PCI_LEGACY) 25 module_param(force_legacy, bool, 0444); 26 MODULE_PARM_DESC(force_legacy, 27 "Force legacy mode for transitional virtio 1 devices"); 28 #endif 29 30 /* wait for pending irq handlers */ 31 void vp_synchronize_vectors(struct virtio_device *vdev) 32 { 33 struct virtio_pci_device *vp_dev = to_vp_device(vdev); 34 int i; 35 36 if (vp_dev->intx_enabled) 37 synchronize_irq(vp_dev->pci_dev->irq); 38 39 for (i = 0; i < vp_dev->msix_vectors; ++i) 40 synchronize_irq(pci_irq_vector(vp_dev->pci_dev, i)); 41 } 42 43 /* the notify function used when creating a virt queue */ 44 bool vp_notify(struct virtqueue *vq) 45 { 46 /* we write the queue's selector into the notification register to 47 * signal the other end */ 48 iowrite16(vq->index, (void __iomem *)vq->priv); 49 return true; 50 } 51 52 /* Handle a configuration change: Tell driver if it wants to know. */ 53 static irqreturn_t vp_config_changed(int irq, void *opaque) 54 { 55 struct virtio_pci_device *vp_dev = opaque; 56 57 virtio_config_changed(&vp_dev->vdev); 58 return IRQ_HANDLED; 59 } 60 61 /* Notify all virtqueues on an interrupt. */ 62 static irqreturn_t vp_vring_interrupt(int irq, void *opaque) 63 { 64 struct virtio_pci_device *vp_dev = opaque; 65 struct virtio_pci_vq_info *info; 66 irqreturn_t ret = IRQ_NONE; 67 unsigned long flags; 68 69 spin_lock_irqsave(&vp_dev->lock, flags); 70 list_for_each_entry(info, &vp_dev->virtqueues, node) { 71 if (vring_interrupt(irq, info->vq) == IRQ_HANDLED) 72 ret = IRQ_HANDLED; 73 } 74 spin_unlock_irqrestore(&vp_dev->lock, flags); 75 76 return ret; 77 } 78 79 /* A small wrapper to also acknowledge the interrupt when it's handled. 80 * I really need an EIO hook for the vring so I can ack the interrupt once we 81 * know that we'll be handling the IRQ but before we invoke the callback since 82 * the callback may notify the host which results in the host attempting to 83 * raise an interrupt that we would then mask once we acknowledged the 84 * interrupt. */ 85 static irqreturn_t vp_interrupt(int irq, void *opaque) 86 { 87 struct virtio_pci_device *vp_dev = opaque; 88 u8 isr; 89 90 /* reading the ISR has the effect of also clearing it so it's very 91 * important to save off the value. */ 92 isr = ioread8(vp_dev->isr); 93 94 /* It's definitely not us if the ISR was not high */ 95 if (!isr) 96 return IRQ_NONE; 97 98 /* Configuration change? Tell driver if it wants to know. */ 99 if (isr & VIRTIO_PCI_ISR_CONFIG) 100 vp_config_changed(irq, opaque); 101 102 return vp_vring_interrupt(irq, opaque); 103 } 104 105 static int vp_request_msix_vectors(struct virtio_device *vdev, int nvectors, 106 bool per_vq_vectors, struct irq_affinity *desc) 107 { 108 struct virtio_pci_device *vp_dev = to_vp_device(vdev); 109 const char *name = dev_name(&vp_dev->vdev.dev); 110 unsigned flags = PCI_IRQ_MSIX; 111 unsigned i, v; 112 int err = -ENOMEM; 113 114 vp_dev->msix_vectors = nvectors; 115 116 vp_dev->msix_names = kmalloc_array(nvectors, 117 sizeof(*vp_dev->msix_names), 118 GFP_KERNEL); 119 if (!vp_dev->msix_names) 120 goto error; 121 vp_dev->msix_affinity_masks 122 = kcalloc(nvectors, sizeof(*vp_dev->msix_affinity_masks), 123 GFP_KERNEL); 124 if (!vp_dev->msix_affinity_masks) 125 goto error; 126 for (i = 0; i < nvectors; ++i) 127 if (!alloc_cpumask_var(&vp_dev->msix_affinity_masks[i], 128 GFP_KERNEL)) 129 goto error; 130 131 if (desc) { 132 flags |= PCI_IRQ_AFFINITY; 133 desc->pre_vectors++; /* virtio config vector */ 134 } 135 136 err = pci_alloc_irq_vectors_affinity(vp_dev->pci_dev, nvectors, 137 nvectors, flags, desc); 138 if (err < 0) 139 goto error; 140 vp_dev->msix_enabled = 1; 141 142 /* Set the vector used for configuration */ 143 v = vp_dev->msix_used_vectors; 144 snprintf(vp_dev->msix_names[v], sizeof *vp_dev->msix_names, 145 "%s-config", name); 146 err = request_irq(pci_irq_vector(vp_dev->pci_dev, v), 147 vp_config_changed, 0, vp_dev->msix_names[v], 148 vp_dev); 149 if (err) 150 goto error; 151 ++vp_dev->msix_used_vectors; 152 153 v = vp_dev->config_vector(vp_dev, v); 154 /* Verify we had enough resources to assign the vector */ 155 if (v == VIRTIO_MSI_NO_VECTOR) { 156 err = -EBUSY; 157 goto error; 158 } 159 160 if (!per_vq_vectors) { 161 /* Shared vector for all VQs */ 162 v = vp_dev->msix_used_vectors; 163 snprintf(vp_dev->msix_names[v], sizeof *vp_dev->msix_names, 164 "%s-virtqueues", name); 165 err = request_irq(pci_irq_vector(vp_dev->pci_dev, v), 166 vp_vring_interrupt, 0, vp_dev->msix_names[v], 167 vp_dev); 168 if (err) 169 goto error; 170 ++vp_dev->msix_used_vectors; 171 } 172 return 0; 173 error: 174 return err; 175 } 176 177 static struct virtqueue *vp_setup_vq(struct virtio_device *vdev, unsigned index, 178 void (*callback)(struct virtqueue *vq), 179 const char *name, 180 bool ctx, 181 u16 msix_vec) 182 { 183 struct virtio_pci_device *vp_dev = to_vp_device(vdev); 184 struct virtio_pci_vq_info *info = kmalloc(sizeof *info, GFP_KERNEL); 185 struct virtqueue *vq; 186 unsigned long flags; 187 188 /* fill out our structure that represents an active queue */ 189 if (!info) 190 return ERR_PTR(-ENOMEM); 191 192 vq = vp_dev->setup_vq(vp_dev, info, index, callback, name, ctx, 193 msix_vec); 194 if (IS_ERR(vq)) 195 goto out_info; 196 197 info->vq = vq; 198 if (callback) { 199 spin_lock_irqsave(&vp_dev->lock, flags); 200 list_add(&info->node, &vp_dev->virtqueues); 201 spin_unlock_irqrestore(&vp_dev->lock, flags); 202 } else { 203 INIT_LIST_HEAD(&info->node); 204 } 205 206 vp_dev->vqs[index] = info; 207 return vq; 208 209 out_info: 210 kfree(info); 211 return vq; 212 } 213 214 static void vp_del_vq(struct virtqueue *vq) 215 { 216 struct virtio_pci_device *vp_dev = to_vp_device(vq->vdev); 217 struct virtio_pci_vq_info *info = vp_dev->vqs[vq->index]; 218 unsigned long flags; 219 220 spin_lock_irqsave(&vp_dev->lock, flags); 221 list_del(&info->node); 222 spin_unlock_irqrestore(&vp_dev->lock, flags); 223 224 vp_dev->del_vq(info); 225 kfree(info); 226 } 227 228 /* the config->del_vqs() implementation */ 229 void vp_del_vqs(struct virtio_device *vdev) 230 { 231 struct virtio_pci_device *vp_dev = to_vp_device(vdev); 232 struct virtqueue *vq, *n; 233 int i; 234 235 list_for_each_entry_safe(vq, n, &vdev->vqs, list) { 236 if (vp_dev->per_vq_vectors) { 237 int v = vp_dev->vqs[vq->index]->msix_vector; 238 239 if (v != VIRTIO_MSI_NO_VECTOR) { 240 int irq = pci_irq_vector(vp_dev->pci_dev, v); 241 242 irq_set_affinity_hint(irq, NULL); 243 free_irq(irq, vq); 244 } 245 } 246 vp_del_vq(vq); 247 } 248 vp_dev->per_vq_vectors = false; 249 250 if (vp_dev->intx_enabled) { 251 free_irq(vp_dev->pci_dev->irq, vp_dev); 252 vp_dev->intx_enabled = 0; 253 } 254 255 for (i = 0; i < vp_dev->msix_used_vectors; ++i) 256 free_irq(pci_irq_vector(vp_dev->pci_dev, i), vp_dev); 257 258 for (i = 0; i < vp_dev->msix_vectors; i++) 259 if (vp_dev->msix_affinity_masks[i]) 260 free_cpumask_var(vp_dev->msix_affinity_masks[i]); 261 262 if (vp_dev->msix_enabled) { 263 /* Disable the vector used for configuration */ 264 vp_dev->config_vector(vp_dev, VIRTIO_MSI_NO_VECTOR); 265 266 pci_free_irq_vectors(vp_dev->pci_dev); 267 vp_dev->msix_enabled = 0; 268 } 269 270 vp_dev->msix_vectors = 0; 271 vp_dev->msix_used_vectors = 0; 272 kfree(vp_dev->msix_names); 273 vp_dev->msix_names = NULL; 274 kfree(vp_dev->msix_affinity_masks); 275 vp_dev->msix_affinity_masks = NULL; 276 kfree(vp_dev->vqs); 277 vp_dev->vqs = NULL; 278 } 279 280 static int vp_find_vqs_msix(struct virtio_device *vdev, unsigned nvqs, 281 struct virtqueue *vqs[], vq_callback_t *callbacks[], 282 const char * const names[], bool per_vq_vectors, 283 const bool *ctx, 284 struct irq_affinity *desc) 285 { 286 struct virtio_pci_device *vp_dev = to_vp_device(vdev); 287 u16 msix_vec; 288 int i, err, nvectors, allocated_vectors; 289 290 vp_dev->vqs = kcalloc(nvqs, sizeof(*vp_dev->vqs), GFP_KERNEL); 291 if (!vp_dev->vqs) 292 return -ENOMEM; 293 294 if (per_vq_vectors) { 295 /* Best option: one for change interrupt, one per vq. */ 296 nvectors = 1; 297 for (i = 0; i < nvqs; ++i) 298 if (callbacks[i]) 299 ++nvectors; 300 } else { 301 /* Second best: one for change, shared for all vqs. */ 302 nvectors = 2; 303 } 304 305 err = vp_request_msix_vectors(vdev, nvectors, per_vq_vectors, 306 per_vq_vectors ? desc : NULL); 307 if (err) 308 goto error_find; 309 310 vp_dev->per_vq_vectors = per_vq_vectors; 311 allocated_vectors = vp_dev->msix_used_vectors; 312 for (i = 0; i < nvqs; ++i) { 313 if (!names[i]) { 314 vqs[i] = NULL; 315 continue; 316 } 317 318 if (!callbacks[i]) 319 msix_vec = VIRTIO_MSI_NO_VECTOR; 320 else if (vp_dev->per_vq_vectors) 321 msix_vec = allocated_vectors++; 322 else 323 msix_vec = VP_MSIX_VQ_VECTOR; 324 vqs[i] = vp_setup_vq(vdev, i, callbacks[i], names[i], 325 ctx ? ctx[i] : false, 326 msix_vec); 327 if (IS_ERR(vqs[i])) { 328 err = PTR_ERR(vqs[i]); 329 goto error_find; 330 } 331 332 if (!vp_dev->per_vq_vectors || msix_vec == VIRTIO_MSI_NO_VECTOR) 333 continue; 334 335 /* allocate per-vq irq if available and necessary */ 336 snprintf(vp_dev->msix_names[msix_vec], 337 sizeof *vp_dev->msix_names, 338 "%s-%s", 339 dev_name(&vp_dev->vdev.dev), names[i]); 340 err = request_irq(pci_irq_vector(vp_dev->pci_dev, msix_vec), 341 vring_interrupt, 0, 342 vp_dev->msix_names[msix_vec], 343 vqs[i]); 344 if (err) 345 goto error_find; 346 } 347 return 0; 348 349 error_find: 350 vp_del_vqs(vdev); 351 return err; 352 } 353 354 static int vp_find_vqs_intx(struct virtio_device *vdev, unsigned nvqs, 355 struct virtqueue *vqs[], vq_callback_t *callbacks[], 356 const char * const names[], const bool *ctx) 357 { 358 struct virtio_pci_device *vp_dev = to_vp_device(vdev); 359 int i, err; 360 361 vp_dev->vqs = kcalloc(nvqs, sizeof(*vp_dev->vqs), GFP_KERNEL); 362 if (!vp_dev->vqs) 363 return -ENOMEM; 364 365 err = request_irq(vp_dev->pci_dev->irq, vp_interrupt, IRQF_SHARED, 366 dev_name(&vdev->dev), vp_dev); 367 if (err) 368 goto out_del_vqs; 369 370 vp_dev->intx_enabled = 1; 371 vp_dev->per_vq_vectors = false; 372 for (i = 0; i < nvqs; ++i) { 373 if (!names[i]) { 374 vqs[i] = NULL; 375 continue; 376 } 377 vqs[i] = vp_setup_vq(vdev, i, callbacks[i], names[i], 378 ctx ? ctx[i] : false, 379 VIRTIO_MSI_NO_VECTOR); 380 if (IS_ERR(vqs[i])) { 381 err = PTR_ERR(vqs[i]); 382 goto out_del_vqs; 383 } 384 } 385 386 return 0; 387 out_del_vqs: 388 vp_del_vqs(vdev); 389 return err; 390 } 391 392 /* the config->find_vqs() implementation */ 393 int vp_find_vqs(struct virtio_device *vdev, unsigned nvqs, 394 struct virtqueue *vqs[], vq_callback_t *callbacks[], 395 const char * const names[], const bool *ctx, 396 struct irq_affinity *desc) 397 { 398 int err; 399 400 /* Try MSI-X with one vector per queue. */ 401 err = vp_find_vqs_msix(vdev, nvqs, vqs, callbacks, names, true, ctx, desc); 402 if (!err) 403 return 0; 404 /* Fallback: MSI-X with one vector for config, one shared for queues. */ 405 err = vp_find_vqs_msix(vdev, nvqs, vqs, callbacks, names, false, ctx, desc); 406 if (!err) 407 return 0; 408 /* Finally fall back to regular interrupts. */ 409 return vp_find_vqs_intx(vdev, nvqs, vqs, callbacks, names, ctx); 410 } 411 412 const char *vp_bus_name(struct virtio_device *vdev) 413 { 414 struct virtio_pci_device *vp_dev = to_vp_device(vdev); 415 416 return pci_name(vp_dev->pci_dev); 417 } 418 419 /* Setup the affinity for a virtqueue: 420 * - force the affinity for per vq vector 421 * - OR over all affinities for shared MSI 422 * - ignore the affinity request if we're using INTX 423 */ 424 int vp_set_vq_affinity(struct virtqueue *vq, const struct cpumask *cpu_mask) 425 { 426 struct virtio_device *vdev = vq->vdev; 427 struct virtio_pci_device *vp_dev = to_vp_device(vdev); 428 struct virtio_pci_vq_info *info = vp_dev->vqs[vq->index]; 429 struct cpumask *mask; 430 unsigned int irq; 431 432 if (!vq->callback) 433 return -EINVAL; 434 435 if (vp_dev->msix_enabled) { 436 mask = vp_dev->msix_affinity_masks[info->msix_vector]; 437 irq = pci_irq_vector(vp_dev->pci_dev, info->msix_vector); 438 if (!cpu_mask) 439 irq_set_affinity_hint(irq, NULL); 440 else { 441 cpumask_copy(mask, cpu_mask); 442 irq_set_affinity_hint(irq, mask); 443 } 444 } 445 return 0; 446 } 447 448 const struct cpumask *vp_get_vq_affinity(struct virtio_device *vdev, int index) 449 { 450 struct virtio_pci_device *vp_dev = to_vp_device(vdev); 451 452 if (!vp_dev->per_vq_vectors || 453 vp_dev->vqs[index]->msix_vector == VIRTIO_MSI_NO_VECTOR) 454 return NULL; 455 456 return pci_irq_get_affinity(vp_dev->pci_dev, 457 vp_dev->vqs[index]->msix_vector); 458 } 459 460 #ifdef CONFIG_PM_SLEEP 461 static int virtio_pci_freeze(struct device *dev) 462 { 463 struct pci_dev *pci_dev = to_pci_dev(dev); 464 struct virtio_pci_device *vp_dev = pci_get_drvdata(pci_dev); 465 int ret; 466 467 ret = virtio_device_freeze(&vp_dev->vdev); 468 469 if (!ret) 470 pci_disable_device(pci_dev); 471 return ret; 472 } 473 474 static int virtio_pci_restore(struct device *dev) 475 { 476 struct pci_dev *pci_dev = to_pci_dev(dev); 477 struct virtio_pci_device *vp_dev = pci_get_drvdata(pci_dev); 478 int ret; 479 480 ret = pci_enable_device(pci_dev); 481 if (ret) 482 return ret; 483 484 pci_set_master(pci_dev); 485 return virtio_device_restore(&vp_dev->vdev); 486 } 487 488 static const struct dev_pm_ops virtio_pci_pm_ops = { 489 SET_SYSTEM_SLEEP_PM_OPS(virtio_pci_freeze, virtio_pci_restore) 490 }; 491 #endif 492 493 494 /* Qumranet donated their vendor ID for devices 0x1000 thru 0x10FF. */ 495 static const struct pci_device_id virtio_pci_id_table[] = { 496 { PCI_DEVICE(PCI_VENDOR_ID_REDHAT_QUMRANET, PCI_ANY_ID) }, 497 { 0 } 498 }; 499 500 MODULE_DEVICE_TABLE(pci, virtio_pci_id_table); 501 502 static void virtio_pci_release_dev(struct device *_d) 503 { 504 struct virtio_device *vdev = dev_to_virtio(_d); 505 struct virtio_pci_device *vp_dev = to_vp_device(vdev); 506 507 /* As struct device is a kobject, it's not safe to 508 * free the memory (including the reference counter itself) 509 * until it's release callback. */ 510 kfree(vp_dev); 511 } 512 513 static int virtio_pci_probe(struct pci_dev *pci_dev, 514 const struct pci_device_id *id) 515 { 516 struct virtio_pci_device *vp_dev, *reg_dev = NULL; 517 int rc; 518 519 /* allocate our structure and fill it out */ 520 vp_dev = kzalloc(sizeof(struct virtio_pci_device), GFP_KERNEL); 521 if (!vp_dev) 522 return -ENOMEM; 523 524 pci_set_drvdata(pci_dev, vp_dev); 525 vp_dev->vdev.dev.parent = &pci_dev->dev; 526 vp_dev->vdev.dev.release = virtio_pci_release_dev; 527 vp_dev->pci_dev = pci_dev; 528 INIT_LIST_HEAD(&vp_dev->virtqueues); 529 spin_lock_init(&vp_dev->lock); 530 531 /* enable the device */ 532 rc = pci_enable_device(pci_dev); 533 if (rc) 534 goto err_enable_device; 535 536 if (force_legacy) { 537 rc = virtio_pci_legacy_probe(vp_dev); 538 /* Also try modern mode if we can't map BAR0 (no IO space). */ 539 if (rc == -ENODEV || rc == -ENOMEM) 540 rc = virtio_pci_modern_probe(vp_dev); 541 if (rc) 542 goto err_probe; 543 } else { 544 rc = virtio_pci_modern_probe(vp_dev); 545 if (rc == -ENODEV) 546 rc = virtio_pci_legacy_probe(vp_dev); 547 if (rc) 548 goto err_probe; 549 } 550 551 pci_set_master(pci_dev); 552 553 rc = register_virtio_device(&vp_dev->vdev); 554 reg_dev = vp_dev; 555 if (rc) 556 goto err_register; 557 558 return 0; 559 560 err_register: 561 if (vp_dev->ioaddr) 562 virtio_pci_legacy_remove(vp_dev); 563 else 564 virtio_pci_modern_remove(vp_dev); 565 err_probe: 566 pci_disable_device(pci_dev); 567 err_enable_device: 568 if (reg_dev) 569 put_device(&vp_dev->vdev.dev); 570 else 571 kfree(vp_dev); 572 return rc; 573 } 574 575 static void virtio_pci_remove(struct pci_dev *pci_dev) 576 { 577 struct virtio_pci_device *vp_dev = pci_get_drvdata(pci_dev); 578 struct device *dev = get_device(&vp_dev->vdev.dev); 579 580 pci_disable_sriov(pci_dev); 581 582 unregister_virtio_device(&vp_dev->vdev); 583 584 if (vp_dev->ioaddr) 585 virtio_pci_legacy_remove(vp_dev); 586 else 587 virtio_pci_modern_remove(vp_dev); 588 589 pci_disable_device(pci_dev); 590 put_device(dev); 591 } 592 593 static int virtio_pci_sriov_configure(struct pci_dev *pci_dev, int num_vfs) 594 { 595 struct virtio_pci_device *vp_dev = pci_get_drvdata(pci_dev); 596 struct virtio_device *vdev = &vp_dev->vdev; 597 int ret; 598 599 if (!(vdev->config->get_status(vdev) & VIRTIO_CONFIG_S_DRIVER_OK)) 600 return -EBUSY; 601 602 if (!__virtio_test_bit(vdev, VIRTIO_F_SR_IOV)) 603 return -EINVAL; 604 605 if (pci_vfs_assigned(pci_dev)) 606 return -EPERM; 607 608 if (num_vfs == 0) { 609 pci_disable_sriov(pci_dev); 610 return 0; 611 } 612 613 ret = pci_enable_sriov(pci_dev, num_vfs); 614 if (ret < 0) 615 return ret; 616 617 return num_vfs; 618 } 619 620 static struct pci_driver virtio_pci_driver = { 621 .name = "virtio-pci", 622 .id_table = virtio_pci_id_table, 623 .probe = virtio_pci_probe, 624 .remove = virtio_pci_remove, 625 #ifdef CONFIG_PM_SLEEP 626 .driver.pm = &virtio_pci_pm_ops, 627 #endif 628 .sriov_configure = virtio_pci_sriov_configure, 629 }; 630 631 module_pci_driver(virtio_pci_driver); 632 633 MODULE_AUTHOR("Anthony Liguori <aliguori@us.ibm.com>"); 634 MODULE_DESCRIPTION("virtio-pci"); 635 MODULE_LICENSE("GPL"); 636 MODULE_VERSION("1"); 637