1 // SPDX-License-Identifier: GPL-2.0-or-later 2 /* 3 * Virtio PCI driver - common functionality for all device versions 4 * 5 * This module allows virtio devices to be used over a virtual PCI device. 6 * This can be used with QEMU based VMMs like KVM or Xen. 7 * 8 * Copyright IBM Corp. 2007 9 * Copyright Red Hat, Inc. 2014 10 * 11 * Authors: 12 * Anthony Liguori <aliguori@us.ibm.com> 13 * Rusty Russell <rusty@rustcorp.com.au> 14 * Michael S. Tsirkin <mst@redhat.com> 15 */ 16 17 #include "virtio_pci_common.h" 18 19 static bool force_legacy = false; 20 21 #if IS_ENABLED(CONFIG_VIRTIO_PCI_LEGACY) 22 module_param(force_legacy, bool, 0444); 23 MODULE_PARM_DESC(force_legacy, 24 "Force legacy mode for transitional virtio 1 devices"); 25 #endif 26 27 /* wait for pending irq handlers */ 28 void vp_synchronize_vectors(struct virtio_device *vdev) 29 { 30 struct virtio_pci_device *vp_dev = to_vp_device(vdev); 31 int i; 32 33 if (vp_dev->intx_enabled) 34 synchronize_irq(vp_dev->pci_dev->irq); 35 36 for (i = 0; i < vp_dev->msix_vectors; ++i) 37 synchronize_irq(pci_irq_vector(vp_dev->pci_dev, i)); 38 } 39 40 /* the notify function used when creating a virt queue */ 41 bool vp_notify(struct virtqueue *vq) 42 { 43 /* we write the queue's selector into the notification register to 44 * signal the other end */ 45 iowrite16(vq->index, (void __iomem *)vq->priv); 46 return true; 47 } 48 49 /* Handle a configuration change: Tell driver if it wants to know. */ 50 static irqreturn_t vp_config_changed(int irq, void *opaque) 51 { 52 struct virtio_pci_device *vp_dev = opaque; 53 54 virtio_config_changed(&vp_dev->vdev); 55 return IRQ_HANDLED; 56 } 57 58 /* Notify all virtqueues on an interrupt. */ 59 static irqreturn_t vp_vring_interrupt(int irq, void *opaque) 60 { 61 struct virtio_pci_device *vp_dev = opaque; 62 struct virtio_pci_vq_info *info; 63 irqreturn_t ret = IRQ_NONE; 64 unsigned long flags; 65 66 spin_lock_irqsave(&vp_dev->lock, flags); 67 list_for_each_entry(info, &vp_dev->virtqueues, node) { 68 if (vring_interrupt(irq, info->vq) == IRQ_HANDLED) 69 ret = IRQ_HANDLED; 70 } 71 spin_unlock_irqrestore(&vp_dev->lock, flags); 72 73 return ret; 74 } 75 76 /* A small wrapper to also acknowledge the interrupt when it's handled. 77 * I really need an EIO hook for the vring so I can ack the interrupt once we 78 * know that we'll be handling the IRQ but before we invoke the callback since 79 * the callback may notify the host which results in the host attempting to 80 * raise an interrupt that we would then mask once we acknowledged the 81 * interrupt. */ 82 static irqreturn_t vp_interrupt(int irq, void *opaque) 83 { 84 struct virtio_pci_device *vp_dev = opaque; 85 u8 isr; 86 87 /* reading the ISR has the effect of also clearing it so it's very 88 * important to save off the value. */ 89 isr = ioread8(vp_dev->isr); 90 91 /* It's definitely not us if the ISR was not high */ 92 if (!isr) 93 return IRQ_NONE; 94 95 /* Configuration change? Tell driver if it wants to know. */ 96 if (isr & VIRTIO_PCI_ISR_CONFIG) 97 vp_config_changed(irq, opaque); 98 99 return vp_vring_interrupt(irq, opaque); 100 } 101 102 static int vp_request_msix_vectors(struct virtio_device *vdev, int nvectors, 103 bool per_vq_vectors, struct irq_affinity *desc) 104 { 105 struct virtio_pci_device *vp_dev = to_vp_device(vdev); 106 const char *name = dev_name(&vp_dev->vdev.dev); 107 unsigned int flags = PCI_IRQ_MSIX; 108 unsigned int i, v; 109 int err = -ENOMEM; 110 111 vp_dev->msix_vectors = nvectors; 112 113 vp_dev->msix_names = kmalloc_array(nvectors, 114 sizeof(*vp_dev->msix_names), 115 GFP_KERNEL); 116 if (!vp_dev->msix_names) 117 goto error; 118 vp_dev->msix_affinity_masks 119 = kcalloc(nvectors, sizeof(*vp_dev->msix_affinity_masks), 120 GFP_KERNEL); 121 if (!vp_dev->msix_affinity_masks) 122 goto error; 123 for (i = 0; i < nvectors; ++i) 124 if (!alloc_cpumask_var(&vp_dev->msix_affinity_masks[i], 125 GFP_KERNEL)) 126 goto error; 127 128 if (desc) { 129 flags |= PCI_IRQ_AFFINITY; 130 desc->pre_vectors++; /* virtio config vector */ 131 } 132 133 err = pci_alloc_irq_vectors_affinity(vp_dev->pci_dev, nvectors, 134 nvectors, flags, desc); 135 if (err < 0) 136 goto error; 137 vp_dev->msix_enabled = 1; 138 139 /* Set the vector used for configuration */ 140 v = vp_dev->msix_used_vectors; 141 snprintf(vp_dev->msix_names[v], sizeof *vp_dev->msix_names, 142 "%s-config", name); 143 err = request_irq(pci_irq_vector(vp_dev->pci_dev, v), 144 vp_config_changed, 0, vp_dev->msix_names[v], 145 vp_dev); 146 if (err) 147 goto error; 148 ++vp_dev->msix_used_vectors; 149 150 v = vp_dev->config_vector(vp_dev, v); 151 /* Verify we had enough resources to assign the vector */ 152 if (v == VIRTIO_MSI_NO_VECTOR) { 153 err = -EBUSY; 154 goto error; 155 } 156 157 if (!per_vq_vectors) { 158 /* Shared vector for all VQs */ 159 v = vp_dev->msix_used_vectors; 160 snprintf(vp_dev->msix_names[v], sizeof *vp_dev->msix_names, 161 "%s-virtqueues", name); 162 err = request_irq(pci_irq_vector(vp_dev->pci_dev, v), 163 vp_vring_interrupt, 0, vp_dev->msix_names[v], 164 vp_dev); 165 if (err) 166 goto error; 167 ++vp_dev->msix_used_vectors; 168 } 169 return 0; 170 error: 171 return err; 172 } 173 174 static struct virtqueue *vp_setup_vq(struct virtio_device *vdev, unsigned int index, 175 void (*callback)(struct virtqueue *vq), 176 const char *name, 177 bool ctx, 178 u16 msix_vec) 179 { 180 struct virtio_pci_device *vp_dev = to_vp_device(vdev); 181 struct virtio_pci_vq_info *info = kmalloc(sizeof *info, GFP_KERNEL); 182 struct virtqueue *vq; 183 unsigned long flags; 184 185 /* fill out our structure that represents an active queue */ 186 if (!info) 187 return ERR_PTR(-ENOMEM); 188 189 vq = vp_dev->setup_vq(vp_dev, info, index, callback, name, ctx, 190 msix_vec); 191 if (IS_ERR(vq)) 192 goto out_info; 193 194 info->vq = vq; 195 if (callback) { 196 spin_lock_irqsave(&vp_dev->lock, flags); 197 list_add(&info->node, &vp_dev->virtqueues); 198 spin_unlock_irqrestore(&vp_dev->lock, flags); 199 } else { 200 INIT_LIST_HEAD(&info->node); 201 } 202 203 vp_dev->vqs[index] = info; 204 return vq; 205 206 out_info: 207 kfree(info); 208 return vq; 209 } 210 211 static void vp_del_vq(struct virtqueue *vq) 212 { 213 struct virtio_pci_device *vp_dev = to_vp_device(vq->vdev); 214 struct virtio_pci_vq_info *info = vp_dev->vqs[vq->index]; 215 unsigned long flags; 216 217 /* 218 * If it fails during re-enable reset vq. This way we won't rejoin 219 * info->node to the queue. Prevent unexpected irqs. 220 */ 221 if (!vq->reset) { 222 spin_lock_irqsave(&vp_dev->lock, flags); 223 list_del(&info->node); 224 spin_unlock_irqrestore(&vp_dev->lock, flags); 225 } 226 227 vp_dev->del_vq(info); 228 kfree(info); 229 } 230 231 /* the config->del_vqs() implementation */ 232 void vp_del_vqs(struct virtio_device *vdev) 233 { 234 struct virtio_pci_device *vp_dev = to_vp_device(vdev); 235 struct virtqueue *vq, *n; 236 int i; 237 238 list_for_each_entry_safe(vq, n, &vdev->vqs, list) { 239 if (vp_dev->per_vq_vectors) { 240 int v = vp_dev->vqs[vq->index]->msix_vector; 241 242 if (v != VIRTIO_MSI_NO_VECTOR) { 243 int irq = pci_irq_vector(vp_dev->pci_dev, v); 244 245 irq_set_affinity_hint(irq, NULL); 246 free_irq(irq, vq); 247 } 248 } 249 vp_del_vq(vq); 250 } 251 vp_dev->per_vq_vectors = false; 252 253 if (vp_dev->intx_enabled) { 254 free_irq(vp_dev->pci_dev->irq, vp_dev); 255 vp_dev->intx_enabled = 0; 256 } 257 258 for (i = 0; i < vp_dev->msix_used_vectors; ++i) 259 free_irq(pci_irq_vector(vp_dev->pci_dev, i), vp_dev); 260 261 if (vp_dev->msix_affinity_masks) { 262 for (i = 0; i < vp_dev->msix_vectors; i++) 263 free_cpumask_var(vp_dev->msix_affinity_masks[i]); 264 } 265 266 if (vp_dev->msix_enabled) { 267 /* Disable the vector used for configuration */ 268 vp_dev->config_vector(vp_dev, VIRTIO_MSI_NO_VECTOR); 269 270 pci_free_irq_vectors(vp_dev->pci_dev); 271 vp_dev->msix_enabled = 0; 272 } 273 274 vp_dev->msix_vectors = 0; 275 vp_dev->msix_used_vectors = 0; 276 kfree(vp_dev->msix_names); 277 vp_dev->msix_names = NULL; 278 kfree(vp_dev->msix_affinity_masks); 279 vp_dev->msix_affinity_masks = NULL; 280 kfree(vp_dev->vqs); 281 vp_dev->vqs = NULL; 282 } 283 284 static int vp_find_vqs_msix(struct virtio_device *vdev, unsigned int nvqs, 285 struct virtqueue *vqs[], vq_callback_t *callbacks[], 286 const char * const names[], bool per_vq_vectors, 287 const bool *ctx, 288 struct irq_affinity *desc) 289 { 290 struct virtio_pci_device *vp_dev = to_vp_device(vdev); 291 u16 msix_vec; 292 int i, err, nvectors, allocated_vectors, queue_idx = 0; 293 294 vp_dev->vqs = kcalloc(nvqs, sizeof(*vp_dev->vqs), GFP_KERNEL); 295 if (!vp_dev->vqs) 296 return -ENOMEM; 297 298 if (per_vq_vectors) { 299 /* Best option: one for change interrupt, one per vq. */ 300 nvectors = 1; 301 for (i = 0; i < nvqs; ++i) 302 if (names[i] && callbacks[i]) 303 ++nvectors; 304 } else { 305 /* Second best: one for change, shared for all vqs. */ 306 nvectors = 2; 307 } 308 309 err = vp_request_msix_vectors(vdev, nvectors, per_vq_vectors, 310 per_vq_vectors ? desc : NULL); 311 if (err) 312 goto error_find; 313 314 vp_dev->per_vq_vectors = per_vq_vectors; 315 allocated_vectors = vp_dev->msix_used_vectors; 316 for (i = 0; i < nvqs; ++i) { 317 if (!names[i]) { 318 vqs[i] = NULL; 319 continue; 320 } 321 322 if (!callbacks[i]) 323 msix_vec = VIRTIO_MSI_NO_VECTOR; 324 else if (vp_dev->per_vq_vectors) 325 msix_vec = allocated_vectors++; 326 else 327 msix_vec = VP_MSIX_VQ_VECTOR; 328 vqs[i] = vp_setup_vq(vdev, queue_idx++, callbacks[i], names[i], 329 ctx ? ctx[i] : false, 330 msix_vec); 331 if (IS_ERR(vqs[i])) { 332 err = PTR_ERR(vqs[i]); 333 goto error_find; 334 } 335 336 if (!vp_dev->per_vq_vectors || msix_vec == VIRTIO_MSI_NO_VECTOR) 337 continue; 338 339 /* allocate per-vq irq if available and necessary */ 340 snprintf(vp_dev->msix_names[msix_vec], 341 sizeof *vp_dev->msix_names, 342 "%s-%s", 343 dev_name(&vp_dev->vdev.dev), names[i]); 344 err = request_irq(pci_irq_vector(vp_dev->pci_dev, msix_vec), 345 vring_interrupt, 0, 346 vp_dev->msix_names[msix_vec], 347 vqs[i]); 348 if (err) { 349 vp_del_vq(vqs[i]); 350 goto error_find; 351 } 352 } 353 return 0; 354 355 error_find: 356 vp_del_vqs(vdev); 357 return err; 358 } 359 360 static int vp_find_vqs_intx(struct virtio_device *vdev, unsigned int nvqs, 361 struct virtqueue *vqs[], vq_callback_t *callbacks[], 362 const char * const names[], const bool *ctx) 363 { 364 struct virtio_pci_device *vp_dev = to_vp_device(vdev); 365 int i, err, queue_idx = 0; 366 367 vp_dev->vqs = kcalloc(nvqs, sizeof(*vp_dev->vqs), GFP_KERNEL); 368 if (!vp_dev->vqs) 369 return -ENOMEM; 370 371 err = request_irq(vp_dev->pci_dev->irq, vp_interrupt, IRQF_SHARED, 372 dev_name(&vdev->dev), vp_dev); 373 if (err) 374 goto out_del_vqs; 375 376 vp_dev->intx_enabled = 1; 377 vp_dev->per_vq_vectors = false; 378 for (i = 0; i < nvqs; ++i) { 379 if (!names[i]) { 380 vqs[i] = NULL; 381 continue; 382 } 383 vqs[i] = vp_setup_vq(vdev, queue_idx++, callbacks[i], names[i], 384 ctx ? ctx[i] : false, 385 VIRTIO_MSI_NO_VECTOR); 386 if (IS_ERR(vqs[i])) { 387 err = PTR_ERR(vqs[i]); 388 goto out_del_vqs; 389 } 390 } 391 392 return 0; 393 out_del_vqs: 394 vp_del_vqs(vdev); 395 return err; 396 } 397 398 /* the config->find_vqs() implementation */ 399 int vp_find_vqs(struct virtio_device *vdev, unsigned int nvqs, 400 struct virtqueue *vqs[], vq_callback_t *callbacks[], 401 const char * const names[], const bool *ctx, 402 struct irq_affinity *desc) 403 { 404 int err; 405 406 /* Try MSI-X with one vector per queue. */ 407 err = vp_find_vqs_msix(vdev, nvqs, vqs, callbacks, names, true, ctx, desc); 408 if (!err) 409 return 0; 410 /* Fallback: MSI-X with one vector for config, one shared for queues. */ 411 err = vp_find_vqs_msix(vdev, nvqs, vqs, callbacks, names, false, ctx, desc); 412 if (!err) 413 return 0; 414 /* Is there an interrupt? If not give up. */ 415 if (!(to_vp_device(vdev)->pci_dev->irq)) 416 return err; 417 /* Finally fall back to regular interrupts. */ 418 return vp_find_vqs_intx(vdev, nvqs, vqs, callbacks, names, ctx); 419 } 420 421 const char *vp_bus_name(struct virtio_device *vdev) 422 { 423 struct virtio_pci_device *vp_dev = to_vp_device(vdev); 424 425 return pci_name(vp_dev->pci_dev); 426 } 427 428 /* Setup the affinity for a virtqueue: 429 * - force the affinity for per vq vector 430 * - OR over all affinities for shared MSI 431 * - ignore the affinity request if we're using INTX 432 */ 433 int vp_set_vq_affinity(struct virtqueue *vq, const struct cpumask *cpu_mask) 434 { 435 struct virtio_device *vdev = vq->vdev; 436 struct virtio_pci_device *vp_dev = to_vp_device(vdev); 437 struct virtio_pci_vq_info *info = vp_dev->vqs[vq->index]; 438 struct cpumask *mask; 439 unsigned int irq; 440 441 if (!vq->callback) 442 return -EINVAL; 443 444 if (vp_dev->msix_enabled) { 445 mask = vp_dev->msix_affinity_masks[info->msix_vector]; 446 irq = pci_irq_vector(vp_dev->pci_dev, info->msix_vector); 447 if (!cpu_mask) 448 irq_set_affinity_hint(irq, NULL); 449 else { 450 cpumask_copy(mask, cpu_mask); 451 irq_set_affinity_hint(irq, mask); 452 } 453 } 454 return 0; 455 } 456 457 const struct cpumask *vp_get_vq_affinity(struct virtio_device *vdev, int index) 458 { 459 struct virtio_pci_device *vp_dev = to_vp_device(vdev); 460 461 if (!vp_dev->per_vq_vectors || 462 vp_dev->vqs[index]->msix_vector == VIRTIO_MSI_NO_VECTOR) 463 return NULL; 464 465 return pci_irq_get_affinity(vp_dev->pci_dev, 466 vp_dev->vqs[index]->msix_vector); 467 } 468 469 #ifdef CONFIG_PM_SLEEP 470 static int virtio_pci_freeze(struct device *dev) 471 { 472 struct pci_dev *pci_dev = to_pci_dev(dev); 473 struct virtio_pci_device *vp_dev = pci_get_drvdata(pci_dev); 474 int ret; 475 476 ret = virtio_device_freeze(&vp_dev->vdev); 477 478 if (!ret) 479 pci_disable_device(pci_dev); 480 return ret; 481 } 482 483 static int virtio_pci_restore(struct device *dev) 484 { 485 struct pci_dev *pci_dev = to_pci_dev(dev); 486 struct virtio_pci_device *vp_dev = pci_get_drvdata(pci_dev); 487 int ret; 488 489 ret = pci_enable_device(pci_dev); 490 if (ret) 491 return ret; 492 493 pci_set_master(pci_dev); 494 return virtio_device_restore(&vp_dev->vdev); 495 } 496 497 static const struct dev_pm_ops virtio_pci_pm_ops = { 498 SET_SYSTEM_SLEEP_PM_OPS(virtio_pci_freeze, virtio_pci_restore) 499 }; 500 #endif 501 502 503 /* Qumranet donated their vendor ID for devices 0x1000 thru 0x10FF. */ 504 static const struct pci_device_id virtio_pci_id_table[] = { 505 { PCI_DEVICE(PCI_VENDOR_ID_REDHAT_QUMRANET, PCI_ANY_ID) }, 506 { 0 } 507 }; 508 509 MODULE_DEVICE_TABLE(pci, virtio_pci_id_table); 510 511 static void virtio_pci_release_dev(struct device *_d) 512 { 513 struct virtio_device *vdev = dev_to_virtio(_d); 514 struct virtio_pci_device *vp_dev = to_vp_device(vdev); 515 516 /* As struct device is a kobject, it's not safe to 517 * free the memory (including the reference counter itself) 518 * until it's release callback. */ 519 kfree(vp_dev); 520 } 521 522 static int virtio_pci_probe(struct pci_dev *pci_dev, 523 const struct pci_device_id *id) 524 { 525 struct virtio_pci_device *vp_dev, *reg_dev = NULL; 526 int rc; 527 528 /* allocate our structure and fill it out */ 529 vp_dev = kzalloc(sizeof(struct virtio_pci_device), GFP_KERNEL); 530 if (!vp_dev) 531 return -ENOMEM; 532 533 pci_set_drvdata(pci_dev, vp_dev); 534 vp_dev->vdev.dev.parent = &pci_dev->dev; 535 vp_dev->vdev.dev.release = virtio_pci_release_dev; 536 vp_dev->pci_dev = pci_dev; 537 INIT_LIST_HEAD(&vp_dev->virtqueues); 538 spin_lock_init(&vp_dev->lock); 539 540 /* enable the device */ 541 rc = pci_enable_device(pci_dev); 542 if (rc) 543 goto err_enable_device; 544 545 if (force_legacy) { 546 rc = virtio_pci_legacy_probe(vp_dev); 547 /* Also try modern mode if we can't map BAR0 (no IO space). */ 548 if (rc == -ENODEV || rc == -ENOMEM) 549 rc = virtio_pci_modern_probe(vp_dev); 550 if (rc) 551 goto err_probe; 552 } else { 553 rc = virtio_pci_modern_probe(vp_dev); 554 if (rc == -ENODEV) 555 rc = virtio_pci_legacy_probe(vp_dev); 556 if (rc) 557 goto err_probe; 558 } 559 560 pci_set_master(pci_dev); 561 562 rc = register_virtio_device(&vp_dev->vdev); 563 reg_dev = vp_dev; 564 if (rc) 565 goto err_register; 566 567 return 0; 568 569 err_register: 570 if (vp_dev->is_legacy) 571 virtio_pci_legacy_remove(vp_dev); 572 else 573 virtio_pci_modern_remove(vp_dev); 574 err_probe: 575 pci_disable_device(pci_dev); 576 err_enable_device: 577 if (reg_dev) 578 put_device(&vp_dev->vdev.dev); 579 else 580 kfree(vp_dev); 581 return rc; 582 } 583 584 static void virtio_pci_remove(struct pci_dev *pci_dev) 585 { 586 struct virtio_pci_device *vp_dev = pci_get_drvdata(pci_dev); 587 struct device *dev = get_device(&vp_dev->vdev.dev); 588 589 /* 590 * Device is marked broken on surprise removal so that virtio upper 591 * layers can abort any ongoing operation. 592 */ 593 if (!pci_device_is_present(pci_dev)) 594 virtio_break_device(&vp_dev->vdev); 595 596 pci_disable_sriov(pci_dev); 597 598 unregister_virtio_device(&vp_dev->vdev); 599 600 if (vp_dev->is_legacy) 601 virtio_pci_legacy_remove(vp_dev); 602 else 603 virtio_pci_modern_remove(vp_dev); 604 605 pci_disable_device(pci_dev); 606 put_device(dev); 607 } 608 609 static int virtio_pci_sriov_configure(struct pci_dev *pci_dev, int num_vfs) 610 { 611 struct virtio_pci_device *vp_dev = pci_get_drvdata(pci_dev); 612 struct virtio_device *vdev = &vp_dev->vdev; 613 int ret; 614 615 if (!(vdev->config->get_status(vdev) & VIRTIO_CONFIG_S_DRIVER_OK)) 616 return -EBUSY; 617 618 if (!__virtio_test_bit(vdev, VIRTIO_F_SR_IOV)) 619 return -EINVAL; 620 621 if (pci_vfs_assigned(pci_dev)) 622 return -EPERM; 623 624 if (num_vfs == 0) { 625 pci_disable_sriov(pci_dev); 626 return 0; 627 } 628 629 ret = pci_enable_sriov(pci_dev, num_vfs); 630 if (ret < 0) 631 return ret; 632 633 return num_vfs; 634 } 635 636 static struct pci_driver virtio_pci_driver = { 637 .name = "virtio-pci", 638 .id_table = virtio_pci_id_table, 639 .probe = virtio_pci_probe, 640 .remove = virtio_pci_remove, 641 #ifdef CONFIG_PM_SLEEP 642 .driver.pm = &virtio_pci_pm_ops, 643 #endif 644 .sriov_configure = virtio_pci_sriov_configure, 645 }; 646 647 module_pci_driver(virtio_pci_driver); 648 649 MODULE_AUTHOR("Anthony Liguori <aliguori@us.ibm.com>"); 650 MODULE_DESCRIPTION("virtio-pci"); 651 MODULE_LICENSE("GPL"); 652 MODULE_VERSION("1"); 653