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 goto error_find; 350 } 351 return 0; 352 353 error_find: 354 vp_del_vqs(vdev); 355 return err; 356 } 357 358 static int vp_find_vqs_intx(struct virtio_device *vdev, unsigned int nvqs, 359 struct virtqueue *vqs[], vq_callback_t *callbacks[], 360 const char * const names[], const bool *ctx) 361 { 362 struct virtio_pci_device *vp_dev = to_vp_device(vdev); 363 int i, err, queue_idx = 0; 364 365 vp_dev->vqs = kcalloc(nvqs, sizeof(*vp_dev->vqs), GFP_KERNEL); 366 if (!vp_dev->vqs) 367 return -ENOMEM; 368 369 err = request_irq(vp_dev->pci_dev->irq, vp_interrupt, IRQF_SHARED, 370 dev_name(&vdev->dev), vp_dev); 371 if (err) 372 goto out_del_vqs; 373 374 vp_dev->intx_enabled = 1; 375 vp_dev->per_vq_vectors = false; 376 for (i = 0; i < nvqs; ++i) { 377 if (!names[i]) { 378 vqs[i] = NULL; 379 continue; 380 } 381 vqs[i] = vp_setup_vq(vdev, queue_idx++, callbacks[i], names[i], 382 ctx ? ctx[i] : false, 383 VIRTIO_MSI_NO_VECTOR); 384 if (IS_ERR(vqs[i])) { 385 err = PTR_ERR(vqs[i]); 386 goto out_del_vqs; 387 } 388 } 389 390 return 0; 391 out_del_vqs: 392 vp_del_vqs(vdev); 393 return err; 394 } 395 396 /* the config->find_vqs() implementation */ 397 int vp_find_vqs(struct virtio_device *vdev, unsigned int nvqs, 398 struct virtqueue *vqs[], vq_callback_t *callbacks[], 399 const char * const names[], const bool *ctx, 400 struct irq_affinity *desc) 401 { 402 int err; 403 404 /* Try MSI-X with one vector per queue. */ 405 err = vp_find_vqs_msix(vdev, nvqs, vqs, callbacks, names, true, ctx, desc); 406 if (!err) 407 return 0; 408 /* Fallback: MSI-X with one vector for config, one shared for queues. */ 409 err = vp_find_vqs_msix(vdev, nvqs, vqs, callbacks, names, false, ctx, desc); 410 if (!err) 411 return 0; 412 /* Finally fall back to regular interrupts. */ 413 return vp_find_vqs_intx(vdev, nvqs, vqs, callbacks, names, ctx); 414 } 415 416 const char *vp_bus_name(struct virtio_device *vdev) 417 { 418 struct virtio_pci_device *vp_dev = to_vp_device(vdev); 419 420 return pci_name(vp_dev->pci_dev); 421 } 422 423 /* Setup the affinity for a virtqueue: 424 * - force the affinity for per vq vector 425 * - OR over all affinities for shared MSI 426 * - ignore the affinity request if we're using INTX 427 */ 428 int vp_set_vq_affinity(struct virtqueue *vq, const struct cpumask *cpu_mask) 429 { 430 struct virtio_device *vdev = vq->vdev; 431 struct virtio_pci_device *vp_dev = to_vp_device(vdev); 432 struct virtio_pci_vq_info *info = vp_dev->vqs[vq->index]; 433 struct cpumask *mask; 434 unsigned int irq; 435 436 if (!vq->callback) 437 return -EINVAL; 438 439 if (vp_dev->msix_enabled) { 440 mask = vp_dev->msix_affinity_masks[info->msix_vector]; 441 irq = pci_irq_vector(vp_dev->pci_dev, info->msix_vector); 442 if (!cpu_mask) 443 irq_set_affinity_hint(irq, NULL); 444 else { 445 cpumask_copy(mask, cpu_mask); 446 irq_set_affinity_hint(irq, mask); 447 } 448 } 449 return 0; 450 } 451 452 const struct cpumask *vp_get_vq_affinity(struct virtio_device *vdev, int index) 453 { 454 struct virtio_pci_device *vp_dev = to_vp_device(vdev); 455 456 if (!vp_dev->per_vq_vectors || 457 vp_dev->vqs[index]->msix_vector == VIRTIO_MSI_NO_VECTOR) 458 return NULL; 459 460 return pci_irq_get_affinity(vp_dev->pci_dev, 461 vp_dev->vqs[index]->msix_vector); 462 } 463 464 #ifdef CONFIG_PM_SLEEP 465 static int virtio_pci_freeze(struct device *dev) 466 { 467 struct pci_dev *pci_dev = to_pci_dev(dev); 468 struct virtio_pci_device *vp_dev = pci_get_drvdata(pci_dev); 469 int ret; 470 471 ret = virtio_device_freeze(&vp_dev->vdev); 472 473 if (!ret) 474 pci_disable_device(pci_dev); 475 return ret; 476 } 477 478 static int virtio_pci_restore(struct device *dev) 479 { 480 struct pci_dev *pci_dev = to_pci_dev(dev); 481 struct virtio_pci_device *vp_dev = pci_get_drvdata(pci_dev); 482 int ret; 483 484 ret = pci_enable_device(pci_dev); 485 if (ret) 486 return ret; 487 488 pci_set_master(pci_dev); 489 return virtio_device_restore(&vp_dev->vdev); 490 } 491 492 static const struct dev_pm_ops virtio_pci_pm_ops = { 493 SET_SYSTEM_SLEEP_PM_OPS(virtio_pci_freeze, virtio_pci_restore) 494 }; 495 #endif 496 497 498 /* Qumranet donated their vendor ID for devices 0x1000 thru 0x10FF. */ 499 static const struct pci_device_id virtio_pci_id_table[] = { 500 { PCI_DEVICE(PCI_VENDOR_ID_REDHAT_QUMRANET, PCI_ANY_ID) }, 501 { 0 } 502 }; 503 504 MODULE_DEVICE_TABLE(pci, virtio_pci_id_table); 505 506 static void virtio_pci_release_dev(struct device *_d) 507 { 508 struct virtio_device *vdev = dev_to_virtio(_d); 509 struct virtio_pci_device *vp_dev = to_vp_device(vdev); 510 511 /* As struct device is a kobject, it's not safe to 512 * free the memory (including the reference counter itself) 513 * until it's release callback. */ 514 kfree(vp_dev); 515 } 516 517 static int virtio_pci_probe(struct pci_dev *pci_dev, 518 const struct pci_device_id *id) 519 { 520 struct virtio_pci_device *vp_dev, *reg_dev = NULL; 521 int rc; 522 523 /* allocate our structure and fill it out */ 524 vp_dev = kzalloc(sizeof(struct virtio_pci_device), GFP_KERNEL); 525 if (!vp_dev) 526 return -ENOMEM; 527 528 pci_set_drvdata(pci_dev, vp_dev); 529 vp_dev->vdev.dev.parent = &pci_dev->dev; 530 vp_dev->vdev.dev.release = virtio_pci_release_dev; 531 vp_dev->pci_dev = pci_dev; 532 INIT_LIST_HEAD(&vp_dev->virtqueues); 533 spin_lock_init(&vp_dev->lock); 534 535 /* enable the device */ 536 rc = pci_enable_device(pci_dev); 537 if (rc) 538 goto err_enable_device; 539 540 if (force_legacy) { 541 rc = virtio_pci_legacy_probe(vp_dev); 542 /* Also try modern mode if we can't map BAR0 (no IO space). */ 543 if (rc == -ENODEV || rc == -ENOMEM) 544 rc = virtio_pci_modern_probe(vp_dev); 545 if (rc) 546 goto err_probe; 547 } else { 548 rc = virtio_pci_modern_probe(vp_dev); 549 if (rc == -ENODEV) 550 rc = virtio_pci_legacy_probe(vp_dev); 551 if (rc) 552 goto err_probe; 553 } 554 555 pci_set_master(pci_dev); 556 557 vp_dev->is_legacy = vp_dev->ldev.ioaddr ? true : false; 558 559 rc = register_virtio_device(&vp_dev->vdev); 560 reg_dev = vp_dev; 561 if (rc) 562 goto err_register; 563 564 return 0; 565 566 err_register: 567 if (vp_dev->is_legacy) 568 virtio_pci_legacy_remove(vp_dev); 569 else 570 virtio_pci_modern_remove(vp_dev); 571 err_probe: 572 pci_disable_device(pci_dev); 573 err_enable_device: 574 if (reg_dev) 575 put_device(&vp_dev->vdev.dev); 576 else 577 kfree(vp_dev); 578 return rc; 579 } 580 581 static void virtio_pci_remove(struct pci_dev *pci_dev) 582 { 583 struct virtio_pci_device *vp_dev = pci_get_drvdata(pci_dev); 584 struct device *dev = get_device(&vp_dev->vdev.dev); 585 586 /* 587 * Device is marked broken on surprise removal so that virtio upper 588 * layers can abort any ongoing operation. 589 */ 590 if (!pci_device_is_present(pci_dev)) 591 virtio_break_device(&vp_dev->vdev); 592 593 pci_disable_sriov(pci_dev); 594 595 unregister_virtio_device(&vp_dev->vdev); 596 597 if (vp_dev->is_legacy) 598 virtio_pci_legacy_remove(vp_dev); 599 else 600 virtio_pci_modern_remove(vp_dev); 601 602 pci_disable_device(pci_dev); 603 put_device(dev); 604 } 605 606 static int virtio_pci_sriov_configure(struct pci_dev *pci_dev, int num_vfs) 607 { 608 struct virtio_pci_device *vp_dev = pci_get_drvdata(pci_dev); 609 struct virtio_device *vdev = &vp_dev->vdev; 610 int ret; 611 612 if (!(vdev->config->get_status(vdev) & VIRTIO_CONFIG_S_DRIVER_OK)) 613 return -EBUSY; 614 615 if (!__virtio_test_bit(vdev, VIRTIO_F_SR_IOV)) 616 return -EINVAL; 617 618 if (pci_vfs_assigned(pci_dev)) 619 return -EPERM; 620 621 if (num_vfs == 0) { 622 pci_disable_sriov(pci_dev); 623 return 0; 624 } 625 626 ret = pci_enable_sriov(pci_dev, num_vfs); 627 if (ret < 0) 628 return ret; 629 630 return num_vfs; 631 } 632 633 static struct pci_driver virtio_pci_driver = { 634 .name = "virtio-pci", 635 .id_table = virtio_pci_id_table, 636 .probe = virtio_pci_probe, 637 .remove = virtio_pci_remove, 638 #ifdef CONFIG_PM_SLEEP 639 .driver.pm = &virtio_pci_pm_ops, 640 #endif 641 .sriov_configure = virtio_pci_sriov_configure, 642 }; 643 644 module_pci_driver(virtio_pci_driver); 645 646 MODULE_AUTHOR("Anthony Liguori <aliguori@us.ibm.com>"); 647 MODULE_DESCRIPTION("virtio-pci"); 648 MODULE_LICENSE("GPL"); 649 MODULE_VERSION("1"); 650