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 flags = PCI_IRQ_MSIX; 108 unsigned 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 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 spin_lock_irqsave(&vp_dev->lock, flags); 218 list_del(&info->node); 219 spin_unlock_irqrestore(&vp_dev->lock, flags); 220 221 vp_dev->del_vq(info); 222 kfree(info); 223 } 224 225 /* the config->del_vqs() implementation */ 226 void vp_del_vqs(struct virtio_device *vdev) 227 { 228 struct virtio_pci_device *vp_dev = to_vp_device(vdev); 229 struct virtqueue *vq, *n; 230 int i; 231 232 list_for_each_entry_safe(vq, n, &vdev->vqs, list) { 233 if (vp_dev->per_vq_vectors) { 234 int v = vp_dev->vqs[vq->index]->msix_vector; 235 236 if (v != VIRTIO_MSI_NO_VECTOR) { 237 int irq = pci_irq_vector(vp_dev->pci_dev, v); 238 239 irq_set_affinity_hint(irq, NULL); 240 free_irq(irq, vq); 241 } 242 } 243 vp_del_vq(vq); 244 } 245 vp_dev->per_vq_vectors = false; 246 247 if (vp_dev->intx_enabled) { 248 free_irq(vp_dev->pci_dev->irq, vp_dev); 249 vp_dev->intx_enabled = 0; 250 } 251 252 for (i = 0; i < vp_dev->msix_used_vectors; ++i) 253 free_irq(pci_irq_vector(vp_dev->pci_dev, i), vp_dev); 254 255 if (vp_dev->msix_affinity_masks) { 256 for (i = 0; i < vp_dev->msix_vectors; i++) 257 if (vp_dev->msix_affinity_masks[i]) 258 free_cpumask_var(vp_dev->msix_affinity_masks[i]); 259 } 260 261 if (vp_dev->msix_enabled) { 262 /* Disable the vector used for configuration */ 263 vp_dev->config_vector(vp_dev, VIRTIO_MSI_NO_VECTOR); 264 265 pci_free_irq_vectors(vp_dev->pci_dev); 266 vp_dev->msix_enabled = 0; 267 } 268 269 vp_dev->msix_vectors = 0; 270 vp_dev->msix_used_vectors = 0; 271 kfree(vp_dev->msix_names); 272 vp_dev->msix_names = NULL; 273 kfree(vp_dev->msix_affinity_masks); 274 vp_dev->msix_affinity_masks = NULL; 275 kfree(vp_dev->vqs); 276 vp_dev->vqs = NULL; 277 } 278 279 static int vp_find_vqs_msix(struct virtio_device *vdev, unsigned nvqs, 280 struct virtqueue *vqs[], vq_callback_t *callbacks[], 281 const char * const names[], bool per_vq_vectors, 282 const bool *ctx, 283 struct irq_affinity *desc) 284 { 285 struct virtio_pci_device *vp_dev = to_vp_device(vdev); 286 u16 msix_vec; 287 int i, err, nvectors, allocated_vectors, queue_idx = 0; 288 289 vp_dev->vqs = kcalloc(nvqs, sizeof(*vp_dev->vqs), GFP_KERNEL); 290 if (!vp_dev->vqs) 291 return -ENOMEM; 292 293 if (per_vq_vectors) { 294 /* Best option: one for change interrupt, one per vq. */ 295 nvectors = 1; 296 for (i = 0; i < nvqs; ++i) 297 if (names[i] && callbacks[i]) 298 ++nvectors; 299 } else { 300 /* Second best: one for change, shared for all vqs. */ 301 nvectors = 2; 302 } 303 304 err = vp_request_msix_vectors(vdev, nvectors, per_vq_vectors, 305 per_vq_vectors ? desc : NULL); 306 if (err) 307 goto error_find; 308 309 vp_dev->per_vq_vectors = per_vq_vectors; 310 allocated_vectors = vp_dev->msix_used_vectors; 311 for (i = 0; i < nvqs; ++i) { 312 if (!names[i]) { 313 vqs[i] = NULL; 314 continue; 315 } 316 317 if (!callbacks[i]) 318 msix_vec = VIRTIO_MSI_NO_VECTOR; 319 else if (vp_dev->per_vq_vectors) 320 msix_vec = allocated_vectors++; 321 else 322 msix_vec = VP_MSIX_VQ_VECTOR; 323 vqs[i] = vp_setup_vq(vdev, queue_idx++, callbacks[i], names[i], 324 ctx ? ctx[i] : false, 325 msix_vec); 326 if (IS_ERR(vqs[i])) { 327 err = PTR_ERR(vqs[i]); 328 goto error_find; 329 } 330 331 if (!vp_dev->per_vq_vectors || msix_vec == VIRTIO_MSI_NO_VECTOR) 332 continue; 333 334 /* allocate per-vq irq if available and necessary */ 335 snprintf(vp_dev->msix_names[msix_vec], 336 sizeof *vp_dev->msix_names, 337 "%s-%s", 338 dev_name(&vp_dev->vdev.dev), names[i]); 339 err = request_irq(pci_irq_vector(vp_dev->pci_dev, msix_vec), 340 vring_interrupt, 0, 341 vp_dev->msix_names[msix_vec], 342 vqs[i]); 343 if (err) 344 goto error_find; 345 } 346 return 0; 347 348 error_find: 349 vp_del_vqs(vdev); 350 return err; 351 } 352 353 static int vp_find_vqs_intx(struct virtio_device *vdev, unsigned nvqs, 354 struct virtqueue *vqs[], vq_callback_t *callbacks[], 355 const char * const names[], const bool *ctx) 356 { 357 struct virtio_pci_device *vp_dev = to_vp_device(vdev); 358 int i, err, queue_idx = 0; 359 360 vp_dev->vqs = kcalloc(nvqs, sizeof(*vp_dev->vqs), GFP_KERNEL); 361 if (!vp_dev->vqs) 362 return -ENOMEM; 363 364 err = request_irq(vp_dev->pci_dev->irq, vp_interrupt, IRQF_SHARED, 365 dev_name(&vdev->dev), vp_dev); 366 if (err) 367 goto out_del_vqs; 368 369 vp_dev->intx_enabled = 1; 370 vp_dev->per_vq_vectors = false; 371 for (i = 0; i < nvqs; ++i) { 372 if (!names[i]) { 373 vqs[i] = NULL; 374 continue; 375 } 376 vqs[i] = vp_setup_vq(vdev, queue_idx++, callbacks[i], names[i], 377 ctx ? ctx[i] : false, 378 VIRTIO_MSI_NO_VECTOR); 379 if (IS_ERR(vqs[i])) { 380 err = PTR_ERR(vqs[i]); 381 goto out_del_vqs; 382 } 383 } 384 385 return 0; 386 out_del_vqs: 387 vp_del_vqs(vdev); 388 return err; 389 } 390 391 /* the config->find_vqs() implementation */ 392 int vp_find_vqs(struct virtio_device *vdev, unsigned nvqs, 393 struct virtqueue *vqs[], vq_callback_t *callbacks[], 394 const char * const names[], const bool *ctx, 395 struct irq_affinity *desc) 396 { 397 int err; 398 399 /* Try MSI-X with one vector per queue. */ 400 err = vp_find_vqs_msix(vdev, nvqs, vqs, callbacks, names, true, ctx, desc); 401 if (!err) 402 return 0; 403 /* Fallback: MSI-X with one vector for config, one shared for queues. */ 404 err = vp_find_vqs_msix(vdev, nvqs, vqs, callbacks, names, false, ctx, desc); 405 if (!err) 406 return 0; 407 /* Finally fall back to regular interrupts. */ 408 return vp_find_vqs_intx(vdev, nvqs, vqs, callbacks, names, ctx); 409 } 410 411 const char *vp_bus_name(struct virtio_device *vdev) 412 { 413 struct virtio_pci_device *vp_dev = to_vp_device(vdev); 414 415 return pci_name(vp_dev->pci_dev); 416 } 417 418 /* Setup the affinity for a virtqueue: 419 * - force the affinity for per vq vector 420 * - OR over all affinities for shared MSI 421 * - ignore the affinity request if we're using INTX 422 */ 423 int vp_set_vq_affinity(struct virtqueue *vq, const struct cpumask *cpu_mask) 424 { 425 struct virtio_device *vdev = vq->vdev; 426 struct virtio_pci_device *vp_dev = to_vp_device(vdev); 427 struct virtio_pci_vq_info *info = vp_dev->vqs[vq->index]; 428 struct cpumask *mask; 429 unsigned int irq; 430 431 if (!vq->callback) 432 return -EINVAL; 433 434 if (vp_dev->msix_enabled) { 435 mask = vp_dev->msix_affinity_masks[info->msix_vector]; 436 irq = pci_irq_vector(vp_dev->pci_dev, info->msix_vector); 437 if (!cpu_mask) 438 irq_set_affinity_hint(irq, NULL); 439 else { 440 cpumask_copy(mask, cpu_mask); 441 irq_set_affinity_hint(irq, mask); 442 } 443 } 444 return 0; 445 } 446 447 const struct cpumask *vp_get_vq_affinity(struct virtio_device *vdev, int index) 448 { 449 struct virtio_pci_device *vp_dev = to_vp_device(vdev); 450 451 if (!vp_dev->per_vq_vectors || 452 vp_dev->vqs[index]->msix_vector == VIRTIO_MSI_NO_VECTOR) 453 return NULL; 454 455 return pci_irq_get_affinity(vp_dev->pci_dev, 456 vp_dev->vqs[index]->msix_vector); 457 } 458 459 #ifdef CONFIG_PM_SLEEP 460 static int virtio_pci_freeze(struct device *dev) 461 { 462 struct pci_dev *pci_dev = to_pci_dev(dev); 463 struct virtio_pci_device *vp_dev = pci_get_drvdata(pci_dev); 464 int ret; 465 466 ret = virtio_device_freeze(&vp_dev->vdev); 467 468 if (!ret) 469 pci_disable_device(pci_dev); 470 return ret; 471 } 472 473 static int virtio_pci_restore(struct device *dev) 474 { 475 struct pci_dev *pci_dev = to_pci_dev(dev); 476 struct virtio_pci_device *vp_dev = pci_get_drvdata(pci_dev); 477 int ret; 478 479 ret = pci_enable_device(pci_dev); 480 if (ret) 481 return ret; 482 483 pci_set_master(pci_dev); 484 return virtio_device_restore(&vp_dev->vdev); 485 } 486 487 static const struct dev_pm_ops virtio_pci_pm_ops = { 488 SET_SYSTEM_SLEEP_PM_OPS(virtio_pci_freeze, virtio_pci_restore) 489 }; 490 #endif 491 492 493 /* Qumranet donated their vendor ID for devices 0x1000 thru 0x10FF. */ 494 static const struct pci_device_id virtio_pci_id_table[] = { 495 { PCI_DEVICE(PCI_VENDOR_ID_REDHAT_QUMRANET, PCI_ANY_ID) }, 496 { 0 } 497 }; 498 499 MODULE_DEVICE_TABLE(pci, virtio_pci_id_table); 500 501 static void virtio_pci_release_dev(struct device *_d) 502 { 503 struct virtio_device *vdev = dev_to_virtio(_d); 504 struct virtio_pci_device *vp_dev = to_vp_device(vdev); 505 506 /* As struct device is a kobject, it's not safe to 507 * free the memory (including the reference counter itself) 508 * until it's release callback. */ 509 kfree(vp_dev); 510 } 511 512 static int virtio_pci_probe(struct pci_dev *pci_dev, 513 const struct pci_device_id *id) 514 { 515 struct virtio_pci_device *vp_dev, *reg_dev = NULL; 516 int rc; 517 518 /* allocate our structure and fill it out */ 519 vp_dev = kzalloc(sizeof(struct virtio_pci_device), GFP_KERNEL); 520 if (!vp_dev) 521 return -ENOMEM; 522 523 pci_set_drvdata(pci_dev, vp_dev); 524 vp_dev->vdev.dev.parent = &pci_dev->dev; 525 vp_dev->vdev.dev.release = virtio_pci_release_dev; 526 vp_dev->pci_dev = pci_dev; 527 INIT_LIST_HEAD(&vp_dev->virtqueues); 528 spin_lock_init(&vp_dev->lock); 529 530 /* enable the device */ 531 rc = pci_enable_device(pci_dev); 532 if (rc) 533 goto err_enable_device; 534 535 if (force_legacy) { 536 rc = virtio_pci_legacy_probe(vp_dev); 537 /* Also try modern mode if we can't map BAR0 (no IO space). */ 538 if (rc == -ENODEV || rc == -ENOMEM) 539 rc = virtio_pci_modern_probe(vp_dev); 540 if (rc) 541 goto err_probe; 542 } else { 543 rc = virtio_pci_modern_probe(vp_dev); 544 if (rc == -ENODEV) 545 rc = virtio_pci_legacy_probe(vp_dev); 546 if (rc) 547 goto err_probe; 548 } 549 550 pci_set_master(pci_dev); 551 552 rc = register_virtio_device(&vp_dev->vdev); 553 reg_dev = vp_dev; 554 if (rc) 555 goto err_register; 556 557 return 0; 558 559 err_register: 560 if (vp_dev->ioaddr) 561 virtio_pci_legacy_remove(vp_dev); 562 else 563 virtio_pci_modern_remove(vp_dev); 564 err_probe: 565 pci_disable_device(pci_dev); 566 err_enable_device: 567 if (reg_dev) 568 put_device(&vp_dev->vdev.dev); 569 else 570 kfree(vp_dev); 571 return rc; 572 } 573 574 static void virtio_pci_remove(struct pci_dev *pci_dev) 575 { 576 struct virtio_pci_device *vp_dev = pci_get_drvdata(pci_dev); 577 struct device *dev = get_device(&vp_dev->vdev.dev); 578 579 /* 580 * Device is marked broken on surprise removal so that virtio upper 581 * layers can abort any ongoing operation. 582 */ 583 if (!pci_device_is_present(pci_dev)) 584 virtio_break_device(&vp_dev->vdev); 585 586 pci_disable_sriov(pci_dev); 587 588 unregister_virtio_device(&vp_dev->vdev); 589 590 if (vp_dev->ioaddr) 591 virtio_pci_legacy_remove(vp_dev); 592 else 593 virtio_pci_modern_remove(vp_dev); 594 595 pci_disable_device(pci_dev); 596 put_device(dev); 597 } 598 599 static int virtio_pci_sriov_configure(struct pci_dev *pci_dev, int num_vfs) 600 { 601 struct virtio_pci_device *vp_dev = pci_get_drvdata(pci_dev); 602 struct virtio_device *vdev = &vp_dev->vdev; 603 int ret; 604 605 if (!(vdev->config->get_status(vdev) & VIRTIO_CONFIG_S_DRIVER_OK)) 606 return -EBUSY; 607 608 if (!__virtio_test_bit(vdev, VIRTIO_F_SR_IOV)) 609 return -EINVAL; 610 611 if (pci_vfs_assigned(pci_dev)) 612 return -EPERM; 613 614 if (num_vfs == 0) { 615 pci_disable_sriov(pci_dev); 616 return 0; 617 } 618 619 ret = pci_enable_sriov(pci_dev, num_vfs); 620 if (ret < 0) 621 return ret; 622 623 return num_vfs; 624 } 625 626 static struct pci_driver virtio_pci_driver = { 627 .name = "virtio-pci", 628 .id_table = virtio_pci_id_table, 629 .probe = virtio_pci_probe, 630 .remove = virtio_pci_remove, 631 #ifdef CONFIG_PM_SLEEP 632 .driver.pm = &virtio_pci_pm_ops, 633 #endif 634 .sriov_configure = virtio_pci_sriov_configure, 635 }; 636 637 module_pci_driver(virtio_pci_driver); 638 639 MODULE_AUTHOR("Anthony Liguori <aliguori@us.ibm.com>"); 640 MODULE_DESCRIPTION("virtio-pci"); 641 MODULE_LICENSE("GPL"); 642 MODULE_VERSION("1"); 643