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 u32 size, 178 bool ctx, 179 u16 msix_vec) 180 { 181 struct virtio_pci_device *vp_dev = to_vp_device(vdev); 182 struct virtio_pci_vq_info *info = kmalloc(sizeof *info, GFP_KERNEL); 183 struct virtqueue *vq; 184 unsigned long flags; 185 186 /* fill out our structure that represents an active queue */ 187 if (!info) 188 return ERR_PTR(-ENOMEM); 189 190 vq = vp_dev->setup_vq(vp_dev, info, index, callback, name, size, ctx, 191 msix_vec); 192 if (IS_ERR(vq)) 193 goto out_info; 194 195 info->vq = vq; 196 if (callback) { 197 spin_lock_irqsave(&vp_dev->lock, flags); 198 list_add(&info->node, &vp_dev->virtqueues); 199 spin_unlock_irqrestore(&vp_dev->lock, flags); 200 } else { 201 INIT_LIST_HEAD(&info->node); 202 } 203 204 vp_dev->vqs[index] = info; 205 return vq; 206 207 out_info: 208 kfree(info); 209 return vq; 210 } 211 212 static void vp_del_vq(struct virtqueue *vq) 213 { 214 struct virtio_pci_device *vp_dev = to_vp_device(vq->vdev); 215 struct virtio_pci_vq_info *info = vp_dev->vqs[vq->index]; 216 unsigned long flags; 217 218 /* 219 * If it fails during re-enable reset vq. This way we won't rejoin 220 * info->node to the queue. Prevent unexpected irqs. 221 */ 222 if (!vq->reset) { 223 spin_lock_irqsave(&vp_dev->lock, flags); 224 list_del(&info->node); 225 spin_unlock_irqrestore(&vp_dev->lock, flags); 226 } 227 228 vp_dev->del_vq(info); 229 kfree(info); 230 } 231 232 /* the config->del_vqs() implementation */ 233 void vp_del_vqs(struct virtio_device *vdev) 234 { 235 struct virtio_pci_device *vp_dev = to_vp_device(vdev); 236 struct virtqueue *vq, *n; 237 int i; 238 239 list_for_each_entry_safe(vq, n, &vdev->vqs, list) { 240 if (vp_dev->per_vq_vectors) { 241 int v = vp_dev->vqs[vq->index]->msix_vector; 242 243 if (v != VIRTIO_MSI_NO_VECTOR) { 244 int irq = pci_irq_vector(vp_dev->pci_dev, v); 245 246 irq_set_affinity_hint(irq, NULL); 247 free_irq(irq, vq); 248 } 249 } 250 vp_del_vq(vq); 251 } 252 vp_dev->per_vq_vectors = false; 253 254 if (vp_dev->intx_enabled) { 255 free_irq(vp_dev->pci_dev->irq, vp_dev); 256 vp_dev->intx_enabled = 0; 257 } 258 259 for (i = 0; i < vp_dev->msix_used_vectors; ++i) 260 free_irq(pci_irq_vector(vp_dev->pci_dev, i), vp_dev); 261 262 if (vp_dev->msix_affinity_masks) { 263 for (i = 0; i < vp_dev->msix_vectors; i++) 264 free_cpumask_var(vp_dev->msix_affinity_masks[i]); 265 } 266 267 if (vp_dev->msix_enabled) { 268 /* Disable the vector used for configuration */ 269 vp_dev->config_vector(vp_dev, VIRTIO_MSI_NO_VECTOR); 270 271 pci_free_irq_vectors(vp_dev->pci_dev); 272 vp_dev->msix_enabled = 0; 273 } 274 275 vp_dev->msix_vectors = 0; 276 vp_dev->msix_used_vectors = 0; 277 kfree(vp_dev->msix_names); 278 vp_dev->msix_names = NULL; 279 kfree(vp_dev->msix_affinity_masks); 280 vp_dev->msix_affinity_masks = NULL; 281 kfree(vp_dev->vqs); 282 vp_dev->vqs = NULL; 283 } 284 285 static int vp_find_vqs_msix(struct virtio_device *vdev, unsigned int nvqs, 286 struct virtqueue *vqs[], vq_callback_t *callbacks[], 287 const char * const names[], u32 sizes[], bool per_vq_vectors, 288 const bool *ctx, 289 struct irq_affinity *desc) 290 { 291 struct virtio_pci_device *vp_dev = to_vp_device(vdev); 292 u16 msix_vec; 293 int i, err, nvectors, allocated_vectors, queue_idx = 0; 294 295 vp_dev->vqs = kcalloc(nvqs, sizeof(*vp_dev->vqs), GFP_KERNEL); 296 if (!vp_dev->vqs) 297 return -ENOMEM; 298 299 if (per_vq_vectors) { 300 /* Best option: one for change interrupt, one per vq. */ 301 nvectors = 1; 302 for (i = 0; i < nvqs; ++i) 303 if (names[i] && callbacks[i]) 304 ++nvectors; 305 } else { 306 /* Second best: one for change, shared for all vqs. */ 307 nvectors = 2; 308 } 309 310 err = vp_request_msix_vectors(vdev, nvectors, per_vq_vectors, 311 per_vq_vectors ? desc : NULL); 312 if (err) 313 goto error_find; 314 315 vp_dev->per_vq_vectors = per_vq_vectors; 316 allocated_vectors = vp_dev->msix_used_vectors; 317 for (i = 0; i < nvqs; ++i) { 318 if (!names[i]) { 319 vqs[i] = NULL; 320 continue; 321 } 322 323 if (!callbacks[i]) 324 msix_vec = VIRTIO_MSI_NO_VECTOR; 325 else if (vp_dev->per_vq_vectors) 326 msix_vec = allocated_vectors++; 327 else 328 msix_vec = VP_MSIX_VQ_VECTOR; 329 vqs[i] = vp_setup_vq(vdev, queue_idx++, callbacks[i], names[i], 330 sizes ? sizes[i] : 0, 331 ctx ? ctx[i] : false, msix_vec); 332 if (IS_ERR(vqs[i])) { 333 err = PTR_ERR(vqs[i]); 334 goto error_find; 335 } 336 337 if (!vp_dev->per_vq_vectors || msix_vec == VIRTIO_MSI_NO_VECTOR) 338 continue; 339 340 /* allocate per-vq irq if available and necessary */ 341 snprintf(vp_dev->msix_names[msix_vec], 342 sizeof *vp_dev->msix_names, 343 "%s-%s", 344 dev_name(&vp_dev->vdev.dev), names[i]); 345 err = request_irq(pci_irq_vector(vp_dev->pci_dev, msix_vec), 346 vring_interrupt, 0, 347 vp_dev->msix_names[msix_vec], 348 vqs[i]); 349 if (err) 350 goto error_find; 351 } 352 return 0; 353 354 error_find: 355 vp_del_vqs(vdev); 356 return err; 357 } 358 359 static int vp_find_vqs_intx(struct virtio_device *vdev, unsigned int nvqs, 360 struct virtqueue *vqs[], vq_callback_t *callbacks[], 361 const char * const names[], u32 sizes[], const bool *ctx) 362 { 363 struct virtio_pci_device *vp_dev = to_vp_device(vdev); 364 int i, err, queue_idx = 0; 365 366 vp_dev->vqs = kcalloc(nvqs, sizeof(*vp_dev->vqs), GFP_KERNEL); 367 if (!vp_dev->vqs) 368 return -ENOMEM; 369 370 err = request_irq(vp_dev->pci_dev->irq, vp_interrupt, IRQF_SHARED, 371 dev_name(&vdev->dev), vp_dev); 372 if (err) 373 goto out_del_vqs; 374 375 vp_dev->intx_enabled = 1; 376 vp_dev->per_vq_vectors = false; 377 for (i = 0; i < nvqs; ++i) { 378 if (!names[i]) { 379 vqs[i] = NULL; 380 continue; 381 } 382 vqs[i] = vp_setup_vq(vdev, queue_idx++, callbacks[i], names[i], 383 sizes ? sizes[i] : 0, 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[], u32 sizes[], 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, sizes, 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, sizes, false, ctx, desc); 412 if (!err) 413 return 0; 414 /* Finally fall back to regular interrupts. */ 415 return vp_find_vqs_intx(vdev, nvqs, vqs, callbacks, names, sizes, ctx); 416 } 417 418 const char *vp_bus_name(struct virtio_device *vdev) 419 { 420 struct virtio_pci_device *vp_dev = to_vp_device(vdev); 421 422 return pci_name(vp_dev->pci_dev); 423 } 424 425 /* Setup the affinity for a virtqueue: 426 * - force the affinity for per vq vector 427 * - OR over all affinities for shared MSI 428 * - ignore the affinity request if we're using INTX 429 */ 430 int vp_set_vq_affinity(struct virtqueue *vq, const struct cpumask *cpu_mask) 431 { 432 struct virtio_device *vdev = vq->vdev; 433 struct virtio_pci_device *vp_dev = to_vp_device(vdev); 434 struct virtio_pci_vq_info *info = vp_dev->vqs[vq->index]; 435 struct cpumask *mask; 436 unsigned int irq; 437 438 if (!vq->callback) 439 return -EINVAL; 440 441 if (vp_dev->msix_enabled) { 442 mask = vp_dev->msix_affinity_masks[info->msix_vector]; 443 irq = pci_irq_vector(vp_dev->pci_dev, info->msix_vector); 444 if (!cpu_mask) 445 irq_set_affinity_hint(irq, NULL); 446 else { 447 cpumask_copy(mask, cpu_mask); 448 irq_set_affinity_hint(irq, mask); 449 } 450 } 451 return 0; 452 } 453 454 const struct cpumask *vp_get_vq_affinity(struct virtio_device *vdev, int index) 455 { 456 struct virtio_pci_device *vp_dev = to_vp_device(vdev); 457 458 if (!vp_dev->per_vq_vectors || 459 vp_dev->vqs[index]->msix_vector == VIRTIO_MSI_NO_VECTOR) 460 return NULL; 461 462 return pci_irq_get_affinity(vp_dev->pci_dev, 463 vp_dev->vqs[index]->msix_vector); 464 } 465 466 #ifdef CONFIG_PM_SLEEP 467 static int virtio_pci_freeze(struct device *dev) 468 { 469 struct pci_dev *pci_dev = to_pci_dev(dev); 470 struct virtio_pci_device *vp_dev = pci_get_drvdata(pci_dev); 471 int ret; 472 473 ret = virtio_device_freeze(&vp_dev->vdev); 474 475 if (!ret) 476 pci_disable_device(pci_dev); 477 return ret; 478 } 479 480 static int virtio_pci_restore(struct device *dev) 481 { 482 struct pci_dev *pci_dev = to_pci_dev(dev); 483 struct virtio_pci_device *vp_dev = pci_get_drvdata(pci_dev); 484 int ret; 485 486 ret = pci_enable_device(pci_dev); 487 if (ret) 488 return ret; 489 490 pci_set_master(pci_dev); 491 return virtio_device_restore(&vp_dev->vdev); 492 } 493 494 static const struct dev_pm_ops virtio_pci_pm_ops = { 495 SET_SYSTEM_SLEEP_PM_OPS(virtio_pci_freeze, virtio_pci_restore) 496 }; 497 #endif 498 499 500 /* Qumranet donated their vendor ID for devices 0x1000 thru 0x10FF. */ 501 static const struct pci_device_id virtio_pci_id_table[] = { 502 { PCI_DEVICE(PCI_VENDOR_ID_REDHAT_QUMRANET, PCI_ANY_ID) }, 503 { 0 } 504 }; 505 506 MODULE_DEVICE_TABLE(pci, virtio_pci_id_table); 507 508 static void virtio_pci_release_dev(struct device *_d) 509 { 510 struct virtio_device *vdev = dev_to_virtio(_d); 511 struct virtio_pci_device *vp_dev = to_vp_device(vdev); 512 513 /* As struct device is a kobject, it's not safe to 514 * free the memory (including the reference counter itself) 515 * until it's release callback. */ 516 kfree(vp_dev); 517 } 518 519 static int virtio_pci_probe(struct pci_dev *pci_dev, 520 const struct pci_device_id *id) 521 { 522 struct virtio_pci_device *vp_dev, *reg_dev = NULL; 523 int rc; 524 525 /* allocate our structure and fill it out */ 526 vp_dev = kzalloc(sizeof(struct virtio_pci_device), GFP_KERNEL); 527 if (!vp_dev) 528 return -ENOMEM; 529 530 pci_set_drvdata(pci_dev, vp_dev); 531 vp_dev->vdev.dev.parent = &pci_dev->dev; 532 vp_dev->vdev.dev.release = virtio_pci_release_dev; 533 vp_dev->pci_dev = pci_dev; 534 INIT_LIST_HEAD(&vp_dev->virtqueues); 535 spin_lock_init(&vp_dev->lock); 536 537 /* enable the device */ 538 rc = pci_enable_device(pci_dev); 539 if (rc) 540 goto err_enable_device; 541 542 if (force_legacy) { 543 rc = virtio_pci_legacy_probe(vp_dev); 544 /* Also try modern mode if we can't map BAR0 (no IO space). */ 545 if (rc == -ENODEV || rc == -ENOMEM) 546 rc = virtio_pci_modern_probe(vp_dev); 547 if (rc) 548 goto err_probe; 549 } else { 550 rc = virtio_pci_modern_probe(vp_dev); 551 if (rc == -ENODEV) 552 rc = virtio_pci_legacy_probe(vp_dev); 553 if (rc) 554 goto err_probe; 555 } 556 557 pci_set_master(pci_dev); 558 559 vp_dev->is_legacy = vp_dev->ldev.ioaddr ? true : false; 560 561 rc = register_virtio_device(&vp_dev->vdev); 562 reg_dev = vp_dev; 563 if (rc) 564 goto err_register; 565 566 return 0; 567 568 err_register: 569 if (vp_dev->is_legacy) 570 virtio_pci_legacy_remove(vp_dev); 571 else 572 virtio_pci_modern_remove(vp_dev); 573 err_probe: 574 pci_disable_device(pci_dev); 575 err_enable_device: 576 if (reg_dev) 577 put_device(&vp_dev->vdev.dev); 578 else 579 kfree(vp_dev); 580 return rc; 581 } 582 583 static void virtio_pci_remove(struct pci_dev *pci_dev) 584 { 585 struct virtio_pci_device *vp_dev = pci_get_drvdata(pci_dev); 586 struct device *dev = get_device(&vp_dev->vdev.dev); 587 588 /* 589 * Device is marked broken on surprise removal so that virtio upper 590 * layers can abort any ongoing operation. 591 */ 592 if (!pci_device_is_present(pci_dev)) 593 virtio_break_device(&vp_dev->vdev); 594 595 pci_disable_sriov(pci_dev); 596 597 unregister_virtio_device(&vp_dev->vdev); 598 599 if (vp_dev->is_legacy) 600 virtio_pci_legacy_remove(vp_dev); 601 else 602 virtio_pci_modern_remove(vp_dev); 603 604 pci_disable_device(pci_dev); 605 put_device(dev); 606 } 607 608 static int virtio_pci_sriov_configure(struct pci_dev *pci_dev, int num_vfs) 609 { 610 struct virtio_pci_device *vp_dev = pci_get_drvdata(pci_dev); 611 struct virtio_device *vdev = &vp_dev->vdev; 612 int ret; 613 614 if (!(vdev->config->get_status(vdev) & VIRTIO_CONFIG_S_DRIVER_OK)) 615 return -EBUSY; 616 617 if (!__virtio_test_bit(vdev, VIRTIO_F_SR_IOV)) 618 return -EINVAL; 619 620 if (pci_vfs_assigned(pci_dev)) 621 return -EPERM; 622 623 if (num_vfs == 0) { 624 pci_disable_sriov(pci_dev); 625 return 0; 626 } 627 628 ret = pci_enable_sriov(pci_dev, num_vfs); 629 if (ret < 0) 630 return ret; 631 632 return num_vfs; 633 } 634 635 static struct pci_driver virtio_pci_driver = { 636 .name = "virtio-pci", 637 .id_table = virtio_pci_id_table, 638 .probe = virtio_pci_probe, 639 .remove = virtio_pci_remove, 640 #ifdef CONFIG_PM_SLEEP 641 .driver.pm = &virtio_pci_pm_ops, 642 #endif 643 .sriov_configure = virtio_pci_sriov_configure, 644 }; 645 646 module_pci_driver(virtio_pci_driver); 647 648 MODULE_AUTHOR("Anthony Liguori <aliguori@us.ibm.com>"); 649 MODULE_DESCRIPTION("virtio-pci"); 650 MODULE_LICENSE("GPL"); 651 MODULE_VERSION("1"); 652