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 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 free_cpumask_var(vp_dev->msix_affinity_masks[i]); 258 } 259 260 if (vp_dev->msix_enabled) { 261 /* Disable the vector used for configuration */ 262 vp_dev->config_vector(vp_dev, VIRTIO_MSI_NO_VECTOR); 263 264 pci_free_irq_vectors(vp_dev->pci_dev); 265 vp_dev->msix_enabled = 0; 266 } 267 268 vp_dev->msix_vectors = 0; 269 vp_dev->msix_used_vectors = 0; 270 kfree(vp_dev->msix_names); 271 vp_dev->msix_names = NULL; 272 kfree(vp_dev->msix_affinity_masks); 273 vp_dev->msix_affinity_masks = NULL; 274 kfree(vp_dev->vqs); 275 vp_dev->vqs = NULL; 276 } 277 278 static int vp_find_vqs_msix(struct virtio_device *vdev, unsigned int nvqs, 279 struct virtqueue *vqs[], vq_callback_t *callbacks[], 280 const char * const names[], bool per_vq_vectors, 281 const bool *ctx, 282 struct irq_affinity *desc) 283 { 284 struct virtio_pci_device *vp_dev = to_vp_device(vdev); 285 u16 msix_vec; 286 int i, err, nvectors, allocated_vectors, queue_idx = 0; 287 288 vp_dev->vqs = kcalloc(nvqs, sizeof(*vp_dev->vqs), GFP_KERNEL); 289 if (!vp_dev->vqs) 290 return -ENOMEM; 291 292 if (per_vq_vectors) { 293 /* Best option: one for change interrupt, one per vq. */ 294 nvectors = 1; 295 for (i = 0; i < nvqs; ++i) 296 if (names[i] && callbacks[i]) 297 ++nvectors; 298 } else { 299 /* Second best: one for change, shared for all vqs. */ 300 nvectors = 2; 301 } 302 303 err = vp_request_msix_vectors(vdev, nvectors, per_vq_vectors, 304 per_vq_vectors ? desc : NULL); 305 if (err) 306 goto error_find; 307 308 vp_dev->per_vq_vectors = per_vq_vectors; 309 allocated_vectors = vp_dev->msix_used_vectors; 310 for (i = 0; i < nvqs; ++i) { 311 if (!names[i]) { 312 vqs[i] = NULL; 313 continue; 314 } 315 316 if (!callbacks[i]) 317 msix_vec = VIRTIO_MSI_NO_VECTOR; 318 else if (vp_dev->per_vq_vectors) 319 msix_vec = allocated_vectors++; 320 else 321 msix_vec = VP_MSIX_VQ_VECTOR; 322 vqs[i] = vp_setup_vq(vdev, queue_idx++, callbacks[i], names[i], 323 ctx ? ctx[i] : false, 324 msix_vec); 325 if (IS_ERR(vqs[i])) { 326 err = PTR_ERR(vqs[i]); 327 goto error_find; 328 } 329 330 if (!vp_dev->per_vq_vectors || msix_vec == VIRTIO_MSI_NO_VECTOR) 331 continue; 332 333 /* allocate per-vq irq if available and necessary */ 334 snprintf(vp_dev->msix_names[msix_vec], 335 sizeof *vp_dev->msix_names, 336 "%s-%s", 337 dev_name(&vp_dev->vdev.dev), names[i]); 338 err = request_irq(pci_irq_vector(vp_dev->pci_dev, msix_vec), 339 vring_interrupt, 0, 340 vp_dev->msix_names[msix_vec], 341 vqs[i]); 342 if (err) 343 goto error_find; 344 } 345 return 0; 346 347 error_find: 348 vp_del_vqs(vdev); 349 return err; 350 } 351 352 static int vp_find_vqs_intx(struct virtio_device *vdev, unsigned int nvqs, 353 struct virtqueue *vqs[], vq_callback_t *callbacks[], 354 const char * const names[], const bool *ctx) 355 { 356 struct virtio_pci_device *vp_dev = to_vp_device(vdev); 357 int i, err, queue_idx = 0; 358 359 vp_dev->vqs = kcalloc(nvqs, sizeof(*vp_dev->vqs), GFP_KERNEL); 360 if (!vp_dev->vqs) 361 return -ENOMEM; 362 363 err = request_irq(vp_dev->pci_dev->irq, vp_interrupt, IRQF_SHARED, 364 dev_name(&vdev->dev), vp_dev); 365 if (err) 366 goto out_del_vqs; 367 368 vp_dev->intx_enabled = 1; 369 vp_dev->per_vq_vectors = false; 370 for (i = 0; i < nvqs; ++i) { 371 if (!names[i]) { 372 vqs[i] = NULL; 373 continue; 374 } 375 vqs[i] = vp_setup_vq(vdev, queue_idx++, callbacks[i], names[i], 376 ctx ? ctx[i] : false, 377 VIRTIO_MSI_NO_VECTOR); 378 if (IS_ERR(vqs[i])) { 379 err = PTR_ERR(vqs[i]); 380 goto out_del_vqs; 381 } 382 } 383 384 return 0; 385 out_del_vqs: 386 vp_del_vqs(vdev); 387 return err; 388 } 389 390 /* the config->find_vqs() implementation */ 391 int vp_find_vqs(struct virtio_device *vdev, unsigned int nvqs, 392 struct virtqueue *vqs[], vq_callback_t *callbacks[], 393 const char * const names[], const bool *ctx, 394 struct irq_affinity *desc) 395 { 396 int err; 397 398 /* Try MSI-X with one vector per queue. */ 399 err = vp_find_vqs_msix(vdev, nvqs, vqs, callbacks, names, true, ctx, desc); 400 if (!err) 401 return 0; 402 /* Fallback: MSI-X with one vector for config, one shared for queues. */ 403 err = vp_find_vqs_msix(vdev, nvqs, vqs, callbacks, names, false, ctx, desc); 404 if (!err) 405 return 0; 406 /* Finally fall back to regular interrupts. */ 407 return vp_find_vqs_intx(vdev, nvqs, vqs, callbacks, names, ctx); 408 } 409 410 const char *vp_bus_name(struct virtio_device *vdev) 411 { 412 struct virtio_pci_device *vp_dev = to_vp_device(vdev); 413 414 return pci_name(vp_dev->pci_dev); 415 } 416 417 /* Setup the affinity for a virtqueue: 418 * - force the affinity for per vq vector 419 * - OR over all affinities for shared MSI 420 * - ignore the affinity request if we're using INTX 421 */ 422 int vp_set_vq_affinity(struct virtqueue *vq, const struct cpumask *cpu_mask) 423 { 424 struct virtio_device *vdev = vq->vdev; 425 struct virtio_pci_device *vp_dev = to_vp_device(vdev); 426 struct virtio_pci_vq_info *info = vp_dev->vqs[vq->index]; 427 struct cpumask *mask; 428 unsigned int irq; 429 430 if (!vq->callback) 431 return -EINVAL; 432 433 if (vp_dev->msix_enabled) { 434 mask = vp_dev->msix_affinity_masks[info->msix_vector]; 435 irq = pci_irq_vector(vp_dev->pci_dev, info->msix_vector); 436 if (!cpu_mask) 437 irq_set_affinity_hint(irq, NULL); 438 else { 439 cpumask_copy(mask, cpu_mask); 440 irq_set_affinity_hint(irq, mask); 441 } 442 } 443 return 0; 444 } 445 446 const struct cpumask *vp_get_vq_affinity(struct virtio_device *vdev, int index) 447 { 448 struct virtio_pci_device *vp_dev = to_vp_device(vdev); 449 450 if (!vp_dev->per_vq_vectors || 451 vp_dev->vqs[index]->msix_vector == VIRTIO_MSI_NO_VECTOR) 452 return NULL; 453 454 return pci_irq_get_affinity(vp_dev->pci_dev, 455 vp_dev->vqs[index]->msix_vector); 456 } 457 458 #ifdef CONFIG_PM_SLEEP 459 static int virtio_pci_freeze(struct device *dev) 460 { 461 struct pci_dev *pci_dev = to_pci_dev(dev); 462 struct virtio_pci_device *vp_dev = pci_get_drvdata(pci_dev); 463 int ret; 464 465 ret = virtio_device_freeze(&vp_dev->vdev); 466 467 if (!ret) 468 pci_disable_device(pci_dev); 469 return ret; 470 } 471 472 static int virtio_pci_restore(struct device *dev) 473 { 474 struct pci_dev *pci_dev = to_pci_dev(dev); 475 struct virtio_pci_device *vp_dev = pci_get_drvdata(pci_dev); 476 int ret; 477 478 ret = pci_enable_device(pci_dev); 479 if (ret) 480 return ret; 481 482 pci_set_master(pci_dev); 483 return virtio_device_restore(&vp_dev->vdev); 484 } 485 486 static const struct dev_pm_ops virtio_pci_pm_ops = { 487 SET_SYSTEM_SLEEP_PM_OPS(virtio_pci_freeze, virtio_pci_restore) 488 }; 489 #endif 490 491 492 /* Qumranet donated their vendor ID for devices 0x1000 thru 0x10FF. */ 493 static const struct pci_device_id virtio_pci_id_table[] = { 494 { PCI_DEVICE(PCI_VENDOR_ID_REDHAT_QUMRANET, PCI_ANY_ID) }, 495 { 0 } 496 }; 497 498 MODULE_DEVICE_TABLE(pci, virtio_pci_id_table); 499 500 static void virtio_pci_release_dev(struct device *_d) 501 { 502 struct virtio_device *vdev = dev_to_virtio(_d); 503 struct virtio_pci_device *vp_dev = to_vp_device(vdev); 504 505 /* As struct device is a kobject, it's not safe to 506 * free the memory (including the reference counter itself) 507 * until it's release callback. */ 508 kfree(vp_dev); 509 } 510 511 static int virtio_pci_probe(struct pci_dev *pci_dev, 512 const struct pci_device_id *id) 513 { 514 struct virtio_pci_device *vp_dev, *reg_dev = NULL; 515 int rc; 516 517 /* allocate our structure and fill it out */ 518 vp_dev = kzalloc(sizeof(struct virtio_pci_device), GFP_KERNEL); 519 if (!vp_dev) 520 return -ENOMEM; 521 522 pci_set_drvdata(pci_dev, vp_dev); 523 vp_dev->vdev.dev.parent = &pci_dev->dev; 524 vp_dev->vdev.dev.release = virtio_pci_release_dev; 525 vp_dev->pci_dev = pci_dev; 526 INIT_LIST_HEAD(&vp_dev->virtqueues); 527 spin_lock_init(&vp_dev->lock); 528 529 /* enable the device */ 530 rc = pci_enable_device(pci_dev); 531 if (rc) 532 goto err_enable_device; 533 534 if (force_legacy) { 535 rc = virtio_pci_legacy_probe(vp_dev); 536 /* Also try modern mode if we can't map BAR0 (no IO space). */ 537 if (rc == -ENODEV || rc == -ENOMEM) 538 rc = virtio_pci_modern_probe(vp_dev); 539 if (rc) 540 goto err_probe; 541 } else { 542 rc = virtio_pci_modern_probe(vp_dev); 543 if (rc == -ENODEV) 544 rc = virtio_pci_legacy_probe(vp_dev); 545 if (rc) 546 goto err_probe; 547 } 548 549 pci_set_master(pci_dev); 550 551 vp_dev->is_legacy = vp_dev->ldev.ioaddr ? true : false; 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->is_legacy) 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 /* 581 * Device is marked broken on surprise removal so that virtio upper 582 * layers can abort any ongoing operation. 583 */ 584 if (!pci_device_is_present(pci_dev)) 585 virtio_break_device(&vp_dev->vdev); 586 587 pci_disable_sriov(pci_dev); 588 589 unregister_virtio_device(&vp_dev->vdev); 590 591 if (vp_dev->is_legacy) 592 virtio_pci_legacy_remove(vp_dev); 593 else 594 virtio_pci_modern_remove(vp_dev); 595 596 pci_disable_device(pci_dev); 597 put_device(dev); 598 } 599 600 static int virtio_pci_sriov_configure(struct pci_dev *pci_dev, int num_vfs) 601 { 602 struct virtio_pci_device *vp_dev = pci_get_drvdata(pci_dev); 603 struct virtio_device *vdev = &vp_dev->vdev; 604 int ret; 605 606 if (!(vdev->config->get_status(vdev) & VIRTIO_CONFIG_S_DRIVER_OK)) 607 return -EBUSY; 608 609 if (!__virtio_test_bit(vdev, VIRTIO_F_SR_IOV)) 610 return -EINVAL; 611 612 if (pci_vfs_assigned(pci_dev)) 613 return -EPERM; 614 615 if (num_vfs == 0) { 616 pci_disable_sriov(pci_dev); 617 return 0; 618 } 619 620 ret = pci_enable_sriov(pci_dev, num_vfs); 621 if (ret < 0) 622 return ret; 623 624 return num_vfs; 625 } 626 627 static struct pci_driver virtio_pci_driver = { 628 .name = "virtio-pci", 629 .id_table = virtio_pci_id_table, 630 .probe = virtio_pci_probe, 631 .remove = virtio_pci_remove, 632 #ifdef CONFIG_PM_SLEEP 633 .driver.pm = &virtio_pci_pm_ops, 634 #endif 635 .sriov_configure = virtio_pci_sriov_configure, 636 }; 637 638 module_pci_driver(virtio_pci_driver); 639 640 MODULE_AUTHOR("Anthony Liguori <aliguori@us.ibm.com>"); 641 MODULE_DESCRIPTION("virtio-pci"); 642 MODULE_LICENSE("GPL"); 643 MODULE_VERSION("1"); 644